fs/btrfs/inode.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/kernel.h>
   7 #include <linux/bio.h>
   8 #include <linux/buffer_head.h>
   9 #include <linux/file.h>
  10 #include <linux/fs.h>
  11 #include <linux/pagemap.h>
  12 #include <linux/highmem.h>
  13 #include <linux/time.h>
  14 #include <linux/init.h>
  15 #include <linux/string.h>
  16 #include <linux/backing-dev.h>
  17 #include <linux/writeback.h>
  18 #include <linux/compat.h>
  19 #include <linux/xattr.h>
  20 #include <linux/posix_acl.h>
  21 #include <linux/falloc.h>
  22 #include <linux/slab.h>
  23 #include <linux/ratelimit.h>
  24 #include <linux/btrfs.h>
  25 #include <linux/blkdev.h>
  26 #include <linux/posix_acl_xattr.h>
  27 #include <linux/uio.h>
  28 #include <linux/magic.h>
  29 #include <linux/iversion.h>
  30 #include <linux/swap.h>
  31 #include <linux/sched/mm.h>
  32 #include <asm/unaligned.h>
  33 #include "misc.h"
  34 #include "ctree.h"
  35 #include "disk-io.h"
  36 #include "transaction.h"
  37 #include "btrfs_inode.h"
  38 #include "print-tree.h"
  39 #include "ordered-data.h"
  40 #include "xattr.h"
  41 #include "tree-log.h"
  42 #include "volumes.h"
  43 #include "compression.h"
  44 #include "locking.h"
  45 #include "free-space-cache.h"
  46 #include "inode-map.h"
  47 #include "backref.h"
  48 #include "props.h"
  49 #include "qgroup.h"
  50 #include "delalloc-space.h"
  51 #include "block-group.h"
  52
  53 struct btrfs_iget_args {
  54         struct btrfs_key *location;
  55         struct btrfs_root *root;
  56 };
  57
  58 struct btrfs_dio_data {
  59         u64 reserve;
  60         u64 unsubmitted_oe_range_start;
  61         u64 unsubmitted_oe_range_end;
  62         int overwrite;
  63 };
  64
  65 static const struct inode_operations btrfs_dir_inode_operations;
  66 static const struct inode_operations btrfs_symlink_inode_operations;
  67 static const struct inode_operations btrfs_dir_ro_inode_operations;
  68 static const struct inode_operations btrfs_special_inode_operations;
  69 static const struct inode_operations btrfs_file_inode_operations;
  70 static const struct address_space_operations btrfs_aops;
  71 static const struct file_operations btrfs_dir_file_operations;
  72 static const struct extent_io_ops btrfs_extent_io_ops;
  73
  74 static struct kmem_cache *btrfs_inode_cachep;
  75 struct kmem_cache *btrfs_trans_handle_cachep;
  76 struct kmem_cache *btrfs_path_cachep;
  77 struct kmem_cache *btrfs_free_space_cachep;
  78 struct kmem_cache *btrfs_free_space_bitmap_cachep;
  79
  80 static int btrfs_setsize(struct inode *inode, struct iattr *attr);
  81 static int btrfs_truncate(struct inode *inode, bool skip_writeback);
  82 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent);
  83 static noinline int cow_file_range(struct inode *inode,
  84                                    struct page *locked_page,
  85                                    u64 start, u64 end, int *page_started,
  86                                    unsigned long *nr_written, int unlock);
  87 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
  88                                        u64 orig_start, u64 block_start,
  89                                        u64 block_len, u64 orig_block_len,
  90                                        u64 ram_bytes, int compress_type,
  91                                        int type);
  92
  93 static void __endio_write_update_ordered(struct inode *inode,
  94                                          const u64 offset, const u64 bytes,
  95                                          const bool uptodate);
  96
  97 /*
  98  * Cleanup all submitted ordered extents in specified range to handle errors
  99  * from the btrfs_run_delalloc_range() callback.
 100  *
 101  * NOTE: caller must ensure that when an error happens, it can not call
 102  * extent_clear_unlock_delalloc() to clear both the bits EXTENT_DO_ACCOUNTING
 103  * and EXTENT_DELALLOC simultaneously, because that causes the reserved metadata
 104  * to be released, which we want to happen only when finishing the ordered
 105  * extent (btrfs_finish_ordered_io()).
 106  */
 107 static inline void btrfs_cleanup_ordered_extents(struct inode *inode,
 108                                                  struct page *locked_page,
 109                                                  u64 offset, u64 bytes)
 110 {
 111         unsigned long index = offset >> PAGE_SHIFT;
 112         unsigned long end_index = (offset + bytes - 1) >> PAGE_SHIFT;
 113         u64 page_start = page_offset(locked_page);
 114         u64 page_end = page_start + PAGE_SIZE - 1;
 115
 116         struct page *page;
 117
 118         while (index <= end_index) {
 119                 page = find_get_page(inode->i_mapping, index);
 120                 index++;
 121                 if (!page)
 122                         continue;
 123                 ClearPagePrivate2(page);
 124                 put_page(page);
 125         }
 126
 127         /*
 128          * In case this page belongs to the delalloc range being instantiated
 129          * then skip it, since the first page of a range is going to be
 130          * properly cleaned up by the caller of run_delalloc_range
 131          */
 132         if (page_start >= offset && page_end <= (offset + bytes - 1)) {
 133                 offset += PAGE_SIZE;
 134                 bytes -= PAGE_SIZE;
 135         }
 136
 137         return __endio_write_update_ordered(inode, offset, bytes, false);
 138 }
 139
 140 static int btrfs_dirty_inode(struct inode *inode);
 141
 142 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 143 void btrfs_test_inode_set_ops(struct inode *inode)
 144 {
 145         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
 146 }
 147 #endif
 148
 149 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 150                                      struct inode *inode,  struct inode *dir,
 151                                      const struct qstr *qstr)
 152 {
 153         int err;
 154
 155         err = btrfs_init_acl(trans, inode, dir);
 156         if (!err)
 157                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 158         return err;
 159 }
 160
 161 /*
 162  * this does all the hard work for inserting an inline extent into
 163  * the btree.  The caller should have done a btrfs_drop_extents so that
 164  * no overlapping inline items exist in the btree
 165  */
 166 static int insert_inline_extent(struct btrfs_trans_handle *trans,
 167                                 struct btrfs_path *path, int extent_inserted,
 168                                 struct btrfs_root *root, struct inode *inode,
 169                                 u64 start, size_t size, size_t compressed_size,
 170                                 int compress_type,
 171                                 struct page **compressed_pages)
 172 {
 173         struct extent_buffer *leaf;
 174         struct page *page = NULL;
 175         char *kaddr;
 176         unsigned long ptr;
 177         struct btrfs_file_extent_item *ei;
 178         int ret;
 179         size_t cur_size = size;
 180         unsigned long offset;
 181
 182         ASSERT((compressed_size > 0 && compressed_pages) ||
 183                (compressed_size == 0 && !compressed_pages));
 184
 185         if (compressed_size && compressed_pages)
 186                 cur_size = compressed_size;
 187
 188         inode_add_bytes(inode, size);
 189
 190         if (!extent_inserted) {
 191                 struct btrfs_key key;
 192                 size_t datasize;
 193
 194                 key.objectid = btrfs_ino(BTRFS_I(inode));
 195                 key.offset = start;
 196                 key.type = BTRFS_EXTENT_DATA_KEY;
 197
 198                 datasize = btrfs_file_extent_calc_inline_size(cur_size);
 199                 path->leave_spinning = 1;
 200                 ret = btrfs_insert_empty_item(trans, root, path, &key,
 201                                               datasize);
 202                 if (ret)
 203                         goto fail;
 204         }
 205         leaf = path->nodes[0];
 206         ei = btrfs_item_ptr(leaf, path->slots[0],
 207                             struct btrfs_file_extent_item);
 208         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 209         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 210         btrfs_set_file_extent_encryption(leaf, ei, 0);
 211         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 212         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 213         ptr = btrfs_file_extent_inline_start(ei);
 214
 215         if (compress_type != BTRFS_COMPRESS_NONE) {
 216                 struct page *cpage;
 217                 int i = 0;
 218                 while (compressed_size > 0) {
 219                         cpage = compressed_pages[i];
 220                         cur_size = min_t(unsigned long, compressed_size,
 221                                        PAGE_SIZE);
 222
 223                         kaddr = kmap_atomic(cpage);
 224                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 225                         kunmap_atomic(kaddr);
 226
 227                         i++;
 228                         ptr += cur_size;
 229                         compressed_size -= cur_size;
 230                 }
 231                 btrfs_set_file_extent_compression(leaf, ei,
 232                                                   compress_type);
 233         } else {
 234                 page = find_get_page(inode->i_mapping,
 235                                      start >> PAGE_SHIFT);
 236                 btrfs_set_file_extent_compression(leaf, ei, 0);
 237                 kaddr = kmap_atomic(page);
 238                 offset = offset_in_page(start);
 239                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 240                 kunmap_atomic(kaddr);
 241                 put_page(page);
 242         }
 243         btrfs_mark_buffer_dirty(leaf);
 244         btrfs_release_path(path);
 245
 246         /*
 247          * we're an inline extent, so nobody can
 248          * extend the file past i_size without locking
 249          * a page we already have locked.
 250          *
 251          * We must do any isize and inode updates
 252          * before we unlock the pages.  Otherwise we
 253          * could end up racing with unlink.
 254          */
 255         BTRFS_I(inode)->disk_i_size = inode->i_size;
 256         ret = btrfs_update_inode(trans, root, inode);
 257
 258 fail:
 259         return ret;
 260 }
 261
 262
 263 /*
 264  * conditionally insert an inline extent into the file.  This
 265  * does the checks required to make sure the data is small enough
 266  * to fit as an inline extent.
 267  */
 268 static noinline int cow_file_range_inline(struct inode *inode, u64 start,
 269                                           u64 end, size_t compressed_size,
 270                                           int compress_type,
 271                                           struct page **compressed_pages)
 272 {
 273         struct btrfs_root *root = BTRFS_I(inode)->root;
 274         struct btrfs_fs_info *fs_info = root->fs_info;
 275         struct btrfs_trans_handle *trans;
 276         u64 isize = i_size_read(inode);
 277         u64 actual_end = min(end + 1, isize);
 278         u64 inline_len = actual_end - start;
 279         u64 aligned_end = ALIGN(end, fs_info->sectorsize);
 280         u64 data_len = inline_len;
 281         int ret;
 282         struct btrfs_path *path;
 283         int extent_inserted = 0;
 284         u32 extent_item_size;
 285
 286         if (compressed_size)
 287                 data_len = compressed_size;
 288
 289         if (start > 0 ||
 290             actual_end > fs_info->sectorsize ||
 291             data_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info) ||
 292             (!compressed_size &&
 293             (actual_end & (fs_info->sectorsize - 1)) == 0) ||
 294             end + 1 < isize ||
 295             data_len > fs_info->max_inline) {
 296                 return 1;
 297         }
 298
 299         path = btrfs_alloc_path();
 300         if (!path)
 301                 return -ENOMEM;
 302
 303         trans = btrfs_join_transaction(root);
 304         if (IS_ERR(trans)) {
 305                 btrfs_free_path(path);
 306                 return PTR_ERR(trans);
 307         }
 308         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
 309
 310         if (compressed_size && compressed_pages)
 311                 extent_item_size = btrfs_file_extent_calc_inline_size(
 312                    compressed_size);
 313         else
 314                 extent_item_size = btrfs_file_extent_calc_inline_size(
 315                     inline_len);
 316
 317         ret = __btrfs_drop_extents(trans, root, inode, path,
 318                                    start, aligned_end, NULL,
 319                                    1, 1, extent_item_size, &extent_inserted);
 320         if (ret) {
 321                 btrfs_abort_transaction(trans, ret);
 322                 goto out;
 323         }
 324
 325         if (isize > actual_end)
 326                 inline_len = min_t(u64, isize, actual_end);
 327         ret = insert_inline_extent(trans, path, extent_inserted,
 328                                    root, inode, start,
 329                                    inline_len, compressed_size,
 330                                    compress_type, compressed_pages);
 331         if (ret && ret != -ENOSPC) {
 332                 btrfs_abort_transaction(trans, ret);
 333                 goto out;
 334         } else if (ret == -ENOSPC) {
 335                 ret = 1;
 336                 goto out;
 337         }
 338
 339         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 340         btrfs_drop_extent_cache(BTRFS_I(inode), start, aligned_end - 1, 0);
 341 out:
 342         /*
 343          * Don't forget to free the reserved space, as for inlined extent
 344          * it won't count as data extent, free them directly here.
 345          * And at reserve time, it's always aligned to page size, so
 346          * just free one page here.
 347          */
 348         btrfs_qgroup_free_data(inode, NULL, 0, PAGE_SIZE);
 349         btrfs_free_path(path);
 350         btrfs_end_transaction(trans);
 351         return ret;
 352 }
 353
 354 struct async_extent {
 355         u64 start;
 356         u64 ram_size;
 357         u64 compressed_size;
 358         struct page **pages;
 359         unsigned long nr_pages;
 360         int compress_type;
 361         struct list_head list;
 362 };
 363
 364 struct async_chunk {
 365         struct inode *inode;
 366         struct page *locked_page;
 367         u64 start;
 368         u64 end;
 369         unsigned int write_flags;
 370         struct list_head extents;
 371         struct cgroup_subsys_state *blkcg_css;
 372         struct btrfs_work work;
 373         atomic_t *pending;
 374 };
 375
 376 struct async_cow {
 377         /* Number of chunks in flight; must be first in the structure */
 378         atomic_t num_chunks;
 379         struct async_chunk chunks[];
 380 };
 381
 382 static noinline int add_async_extent(struct async_chunk *cow,
 383                                      u64 start, u64 ram_size,
 384                                      u64 compressed_size,
 385                                      struct page **pages,
 386                                      unsigned long nr_pages,
 387                                      int compress_type)
 388 {
 389         struct async_extent *async_extent;
 390
 391         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 392         BUG_ON(!async_extent); /* -ENOMEM */
 393         async_extent->start = start;
 394         async_extent->ram_size = ram_size;
 395         async_extent->compressed_size = compressed_size;
 396         async_extent->pages = pages;
 397         async_extent->nr_pages = nr_pages;
 398         async_extent->compress_type = compress_type;
 399         list_add_tail(&async_extent->list, &cow->extents);
 400         return 0;
 401 }
 402
 403 /*
 404  * Check if the inode has flags compatible with compression
 405  */
 406 static inline bool inode_can_compress(struct inode *inode)
 407 {
 408         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW ||
 409             BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
 410                 return false;
 411         return true;
 412 }
 413
 414 /*
 415  * Check if the inode needs to be submitted to compression, based on mount
 416  * options, defragmentation, properties or heuristics.
 417  */
 418 static inline int inode_need_compress(struct inode *inode, u64 start, u64 end)
 419 {
 420         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 421
 422         if (!inode_can_compress(inode)) {
 423                 WARN(IS_ENABLED(CONFIG_BTRFS_DEBUG),
 424                         KERN_ERR "BTRFS: unexpected compression for ino %llu\n",
 425                         btrfs_ino(BTRFS_I(inode)));
 426                 return 0;
 427         }
 428         /* force compress */
 429         if (btrfs_test_opt(fs_info, FORCE_COMPRESS))
 430                 return 1;
 431         /* defrag ioctl */
 432         if (BTRFS_I(inode)->defrag_compress)
 433                 return 1;
 434         /* bad compression ratios */
 435         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
 436                 return 0;
 437         if (btrfs_test_opt(fs_info, COMPRESS) ||
 438             BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS ||
 439             BTRFS_I(inode)->prop_compress)
 440                 return btrfs_compress_heuristic(inode, start, end);
 441         return 0;
 442 }
 443
 444 static inline void inode_should_defrag(struct btrfs_inode *inode,
 445                 u64 start, u64 end, u64 num_bytes, u64 small_write)
 446 {
 447         /* If this is a small write inside eof, kick off a defrag */
 448         if (num_bytes < small_write &&
 449             (start > 0 || end + 1 < inode->disk_i_size))
 450                 btrfs_add_inode_defrag(NULL, inode);
 451 }
 452
 453 /*
 454  * we create compressed extents in two phases.  The first
 455  * phase compresses a range of pages that have already been
 456  * locked (both pages and state bits are locked).
 457  *
 458  * This is done inside an ordered work queue, and the compression
 459  * is spread across many cpus.  The actual IO submission is step
 460  * two, and the ordered work queue takes care of making sure that
 461  * happens in the same order things were put onto the queue by
 462  * writepages and friends.
 463  *
 464  * If this code finds it can't get good compression, it puts an
 465  * entry onto the work queue to write the uncompressed bytes.  This
 466  * makes sure that both compressed inodes and uncompressed inodes
 467  * are written in the same order that the flusher thread sent them
 468  * down.
 469  */
 470 static noinline int compress_file_range(struct async_chunk *async_chunk)
 471 {
 472         struct inode *inode = async_chunk->inode;
 473         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 474         u64 blocksize = fs_info->sectorsize;
 475         u64 start = async_chunk->start;
 476         u64 end = async_chunk->end;
 477         u64 actual_end;
 478         u64 i_size;
 479         int ret = 0;
 480         struct page **pages = NULL;
 481         unsigned long nr_pages;
 482         unsigned long total_compressed = 0;
 483         unsigned long total_in = 0;
 484         int i;
 485         int will_compress;
 486         int compress_type = fs_info->compress_type;
 487         int compressed_extents = 0;
 488         int redirty = 0;
 489
 490         inode_should_defrag(BTRFS_I(inode), start, end, end - start + 1,
 491                         SZ_16K);
 492
 493         /*
 494          * We need to save i_size before now because it could change in between
 495          * us evaluating the size and assigning it.  This is because we lock and
 496          * unlock the page in truncate and fallocate, and then modify the i_size
 497          * later on.
 498          *
 499          * The barriers are to emulate READ_ONCE, remove that once i_size_read
 500          * does that for us.
 501          */
 502         barrier();
 503         i_size = i_size_read(inode);
 504         barrier();
 505         actual_end = min_t(u64, i_size, end + 1);
 506 again:
 507         will_compress = 0;
 508         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 509         BUILD_BUG_ON((BTRFS_MAX_COMPRESSED % PAGE_SIZE) != 0);
 510         nr_pages = min_t(unsigned long, nr_pages,
 511                         BTRFS_MAX_COMPRESSED / PAGE_SIZE);
 512
 513         /*
 514          * we don't want to send crud past the end of i_size through
 515          * compression, that's just a waste of CPU time.  So, if the
 516          * end of the file is before the start of our current
 517          * requested range of bytes, we bail out to the uncompressed
 518          * cleanup code that can deal with all of this.
 519          *
 520          * It isn't really the fastest way to fix things, but this is a
 521          * very uncommon corner.
 522          */
 523         if (actual_end <= start)
 524                 goto cleanup_and_bail_uncompressed;
 525
 526         total_compressed = actual_end - start;
 527
 528         /*
 529          * skip compression for a small file range(<=blocksize) that
 530          * isn't an inline extent, since it doesn't save disk space at all.
 531          */
 532         if (total_compressed <= blocksize &&
 533            (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 534                 goto cleanup_and_bail_uncompressed;
 535
 536         total_compressed = min_t(unsigned long, total_compressed,
 537                         BTRFS_MAX_UNCOMPRESSED);
 538         total_in = 0;
 539         ret = 0;
 540
 541         /*
 542          * we do compression for mount -o compress and when the
 543          * inode has not been flagged as nocompress.  This flag can
 544          * change at any time if we discover bad compression ratios.
 545          */
 546         if (inode_need_compress(inode, start, end)) {
 547                 WARN_ON(pages);
 548                 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS);
 549                 if (!pages) {
 550                         /* just bail out to the uncompressed code */
 551                         nr_pages = 0;
 552                         goto cont;
 553                 }
 554
 555                 if (BTRFS_I(inode)->defrag_compress)
 556                         compress_type = BTRFS_I(inode)->defrag_compress;
 557                 else if (BTRFS_I(inode)->prop_compress)
 558                         compress_type = BTRFS_I(inode)->prop_compress;
 559
 560                 /*
 561                  * we need to call clear_page_dirty_for_io on each
 562                  * page in the range.  Otherwise applications with the file
 563                  * mmap'd can wander in and change the page contents while
 564                  * we are compressing them.
 565                  *
 566                  * If the compression fails for any reason, we set the pages
 567                  * dirty again later on.
 568                  *
 569                  * Note that the remaining part is redirtied, the start pointer
 570                  * has moved, the end is the original one.
 571                  */
 572                 if (!redirty) {
 573                         extent_range_clear_dirty_for_io(inode, start, end);
 574                         redirty = 1;
 575                 }
 576
 577                 /* Compression level is applied here and only here */
 578                 ret = btrfs_compress_pages(
 579                         compress_type | (fs_info->compress_level << 4),
 580                                            inode->i_mapping, start,
 581                                            pages,
 582                                            &nr_pages,
 583                                            &total_in,
 584                                            &total_compressed);
 585
 586                 if (!ret) {
 587                         unsigned long offset = offset_in_page(total_compressed);
 588                         struct page *page = pages[nr_pages - 1];
 589                         char *kaddr;
 590
 591                         /* zero the tail end of the last page, we might be
 592                          * sending it down to disk
 593                          */
 594                         if (offset) {
 595                                 kaddr = kmap_atomic(page);
 596                                 memset(kaddr + offset, 0,
 597                                        PAGE_SIZE - offset);
 598                                 kunmap_atomic(kaddr);
 599                         }
 600                         will_compress = 1;
 601                 }
 602         }
 603 cont:
 604         if (start == 0) {
 605                 /* lets try to make an inline extent */
 606                 if (ret || total_in < actual_end) {
 607                         /* we didn't compress the entire range, try
 608                          * to make an uncompressed inline extent.
 609                          */
 610                         ret = cow_file_range_inline(inode, start, end, 0,
 611                                                     BTRFS_COMPRESS_NONE, NULL);
 612                 } else {
 613                         /* try making a compressed inline extent */
 614                         ret = cow_file_range_inline(inode, start, end,
 615                                                     total_compressed,
 616                                                     compress_type, pages);
 617                 }
 618                 if (ret <= 0) {
 619                         unsigned long clear_flags = EXTENT_DELALLOC |
 620                                 EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
 621                                 EXTENT_DO_ACCOUNTING;
 622                         unsigned long page_error_op;
 623
 624                         page_error_op = ret < 0 ? PAGE_SET_ERROR : 0;
 625
 626                         /*
 627                          * inline extent creation worked or returned error,
 628                          * we don't need to create any more async work items.
 629                          * Unlock and free up our temp pages.
 630                          *
 631                          * We use DO_ACCOUNTING here because we need the
 632                          * delalloc_release_metadata to be done _after_ we drop
 633                          * our outstanding extent for clearing delalloc for this
 634                          * range.
 635                          */
 636                         extent_clear_unlock_delalloc(inode, start, end, NULL,
 637                                                      clear_flags,
 638                                                      PAGE_UNLOCK |
 639                                                      PAGE_CLEAR_DIRTY |
 640                                                      PAGE_SET_WRITEBACK |
 641                                                      page_error_op |
 642                                                      PAGE_END_WRITEBACK);
 643
 644                         for (i = 0; i < nr_pages; i++) {
 645                                 WARN_ON(pages[i]->mapping);
 646                                 put_page(pages[i]);
 647                         }
 648                         kfree(pages);
 649
 650                         return 0;
 651                 }
 652         }
 653
 654         if (will_compress) {
 655                 /*
 656                  * we aren't doing an inline extent round the compressed size
 657                  * up to a block size boundary so the allocator does sane
 658                  * things
 659                  */
 660                 total_compressed = ALIGN(total_compressed, blocksize);
 661
 662                 /*
 663                  * one last check to make sure the compression is really a
 664                  * win, compare the page count read with the blocks on disk,
 665                  * compression must free at least one sector size
 666                  */
 667                 total_in = ALIGN(total_in, PAGE_SIZE);
 668                 if (total_compressed + blocksize <= total_in) {
 669                         compressed_extents++;
 670
 671                         /*
 672                          * The async work queues will take care of doing actual
 673                          * allocation on disk for these compressed pages, and
 674                          * will submit them to the elevator.
 675                          */
 676                         add_async_extent(async_chunk, start, total_in,
 677                                         total_compressed, pages, nr_pages,
 678                                         compress_type);
 679
 680                         if (start + total_in < end) {
 681                                 start += total_in;
 682                                 pages = NULL;
 683                                 cond_resched();
 684                                 goto again;
 685                         }
 686                         return compressed_extents;
 687                 }
 688         }
 689         if (pages) {
 690                 /*
 691                  * the compression code ran but failed to make things smaller,
 692                  * free any pages it allocated and our page pointer array
 693                  */
 694                 for (i = 0; i < nr_pages; i++) {
 695                         WARN_ON(pages[i]->mapping);
 696                         put_page(pages[i]);
 697                 }
 698                 kfree(pages);
 699                 pages = NULL;
 700                 total_compressed = 0;
 701                 nr_pages = 0;
 702
 703                 /* flag the file so we don't compress in the future */
 704                 if (!btrfs_test_opt(fs_info, FORCE_COMPRESS) &&
 705                     !(BTRFS_I(inode)->prop_compress)) {
 706                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 707                 }
 708         }
 709 cleanup_and_bail_uncompressed:
 710         /*
 711          * No compression, but we still need to write the pages in the file
 712          * we've been given so far.  redirty the locked page if it corresponds
 713          * to our extent and set things up for the async work queue to run
 714          * cow_file_range to do the normal delalloc dance.
 715          */
 716         if (async_chunk->locked_page &&
 717             (page_offset(async_chunk->locked_page) >= start &&
 718              page_offset(async_chunk->locked_page)) <= end) {
 719                 __set_page_dirty_nobuffers(async_chunk->locked_page);
 720                 /* unlocked later on in the async handlers */
 721         }
 722
 723         if (redirty)
 724                 extent_range_redirty_for_io(inode, start, end);
 725         add_async_extent(async_chunk, start, end - start + 1, 0, NULL, 0,
 726                          BTRFS_COMPRESS_NONE);
 727         compressed_extents++;
 728
 729         return compressed_extents;
 730 }
 731
 732 static void free_async_extent_pages(struct async_extent *async_extent)
 733 {
 734         int i;
 735
 736         if (!async_extent->pages)
 737                 return;
 738
 739         for (i = 0; i < async_extent->nr_pages; i++) {
 740                 WARN_ON(async_extent->pages[i]->mapping);
 741                 put_page(async_extent->pages[i]);
 742         }
 743         kfree(async_extent->pages);
 744         async_extent->nr_pages = 0;
 745         async_extent->pages = NULL;
 746 }
 747
 748 /*
 749  * phase two of compressed writeback.  This is the ordered portion
 750  * of the code, which only gets called in the order the work was
 751  * queued.  We walk all the async extents created by compress_file_range
 752  * and send them down to the disk.
 753  */
 754 static noinline void submit_compressed_extents(struct async_chunk *async_chunk)
 755 {
 756         struct inode *inode = async_chunk->inode;
 757         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 758         struct async_extent *async_extent;
 759         u64 alloc_hint = 0;
 760         struct btrfs_key ins;
 761         struct extent_map *em;
 762         struct btrfs_root *root = BTRFS_I(inode)->root;
 763         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
 764         int ret = 0;
 765
 766 again:
 767         while (!list_empty(&async_chunk->extents)) {
 768                 async_extent = list_entry(async_chunk->extents.next,
 769                                           struct async_extent, list);
 770                 list_del(&async_extent->list);
 771
 772 retry:
 773                 lock_extent(io_tree, async_extent->start,
 774                             async_extent->start + async_extent->ram_size - 1);
 775                 /* did the compression code fall back to uncompressed IO? */
 776                 if (!async_extent->pages) {
 777                         int page_started = 0;
 778                         unsigned long nr_written = 0;
 779
 780                         /* allocate blocks */
 781                         ret = cow_file_range(inode, async_chunk->locked_page,
 782                                              async_extent->start,
 783                                              async_extent->start +
 784                                              async_extent->ram_size - 1,
 785                                              &page_started, &nr_written, 0);
 786
 787                         /* JDM XXX */
 788
 789                         /*
 790                          * if page_started, cow_file_range inserted an
 791                          * inline extent and took care of all the unlocking
 792                          * and IO for us.  Otherwise, we need to submit
 793                          * all those pages down to the drive.
 794                          */
 795                         if (!page_started && !ret)
 796                                 extent_write_locked_range(inode,
 797                                                   async_extent->start,
 798                                                   async_extent->start +
 799                                                   async_extent->ram_size - 1,
 800                                                   WB_SYNC_ALL);
 801                         else if (ret && async_chunk->locked_page)
 802                                 unlock_page(async_chunk->locked_page);
 803                         kfree(async_extent);
 804                         cond_resched();
 805                         continue;
 806                 }
 807
 808                 ret = btrfs_reserve_extent(root, async_extent->ram_size,
 809                                            async_extent->compressed_size,
 810                                            async_extent->compressed_size,
 811                                            0, alloc_hint, &ins, 1, 1);
 812                 if (ret) {
 813                         free_async_extent_pages(async_extent);
 814
 815                         if (ret == -ENOSPC) {
 816                                 unlock_extent(io_tree, async_extent->start,
 817                                               async_extent->start +
 818                                               async_extent->ram_size - 1);
 819
 820                                 /*
 821                                  * we need to redirty the pages if we decide to
 822                                  * fallback to uncompressed IO, otherwise we
 823                                  * will not submit these pages down to lower
 824                                  * layers.
 825                                  */
 826                                 extent_range_redirty_for_io(inode,
 827                                                 async_extent->start,
 828                                                 async_extent->start +
 829                                                 async_extent->ram_size - 1);
 830
 831                                 goto retry;
 832                         }
 833                         goto out_free;
 834                 }
 835                 /*
 836                  * here we're doing allocation and writeback of the
 837                  * compressed pages
 838                  */
 839                 em = create_io_em(inode, async_extent->start,
 840                                   async_extent->ram_size, /* len */
 841                                   async_extent->start, /* orig_start */
 842                                   ins.objectid, /* block_start */
 843                                   ins.offset, /* block_len */
 844                                   ins.offset, /* orig_block_len */
 845                                   async_extent->ram_size, /* ram_bytes */
 846                                   async_extent->compress_type,
 847                                   BTRFS_ORDERED_COMPRESSED);
 848                 if (IS_ERR(em))
 849                         /* ret value is not necessary due to void function */
 850                         goto out_free_reserve;
 851                 free_extent_map(em);
 852
 853                 ret = btrfs_add_ordered_extent_compress(inode,
 854                                                 async_extent->start,
 855                                                 ins.objectid,
 856                                                 async_extent->ram_size,
 857                                                 ins.offset,
 858                                                 BTRFS_ORDERED_COMPRESSED,
 859                                                 async_extent->compress_type);
 860                 if (ret) {
 861                         btrfs_drop_extent_cache(BTRFS_I(inode),
 862                                                 async_extent->start,
 863                                                 async_extent->start +
 864                                                 async_extent->ram_size - 1, 0);
 865                         goto out_free_reserve;
 866                 }
 867                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 868
 869                 /*
 870                  * clear dirty, set writeback and unlock the pages.
 871                  */
 872                 extent_clear_unlock_delalloc(inode, async_extent->start,
 873                                 async_extent->start +
 874                                 async_extent->ram_size - 1,
 875                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
 876                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 877                                 PAGE_SET_WRITEBACK);
 878                 if (btrfs_submit_compressed_write(inode,
 879                                     async_extent->start,
 880                                     async_extent->ram_size,
 881                                     ins.objectid,
 882                                     ins.offset, async_extent->pages,
 883                                     async_extent->nr_pages,
 884                                     async_chunk->write_flags,
 885                                     async_chunk->blkcg_css)) {
 886                         struct page *p = async_extent->pages[0];
 887                         const u64 start = async_extent->start;
 888                         const u64 end = start + async_extent->ram_size - 1;
 889
 890                         p->mapping = inode->i_mapping;
 891                         btrfs_writepage_endio_finish_ordered(p, start, end, 0);
 892
 893                         p->mapping = NULL;
 894                         extent_clear_unlock_delalloc(inode, start, end,
 895                                                      NULL, 0,
 896                                                      PAGE_END_WRITEBACK |
 897                                                      PAGE_SET_ERROR);
 898                         free_async_extent_pages(async_extent);
 899                 }
 900                 alloc_hint = ins.objectid + ins.offset;
 901                 kfree(async_extent);
 902                 cond_resched();
 903         }
 904         return;
 905 out_free_reserve:
 906         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 907         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
 908 out_free:
 909         extent_clear_unlock_delalloc(inode, async_extent->start,
 910                                      async_extent->start +
 911                                      async_extent->ram_size - 1,
 912                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
 913                                      EXTENT_DELALLOC_NEW |
 914                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
 915                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 916                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
 917                                      PAGE_SET_ERROR);
 918         free_async_extent_pages(async_extent);
 919         kfree(async_extent);
 920         goto again;
 921 }
 922
 923 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 924                                       u64 num_bytes)
 925 {
 926         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 927         struct extent_map *em;
 928         u64 alloc_hint = 0;
 929
 930         read_lock(&em_tree->lock);
 931         em = search_extent_mapping(em_tree, start, num_bytes);
 932         if (em) {
 933                 /*
 934                  * if block start isn't an actual block number then find the
 935                  * first block in this inode and use that as a hint.  If that
 936                  * block is also bogus then just don't worry about it.
 937                  */
 938                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 939                         free_extent_map(em);
 940                         em = search_extent_mapping(em_tree, 0, 0);
 941                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 942                                 alloc_hint = em->block_start;
 943                         if (em)
 944                                 free_extent_map(em);
 945                 } else {
 946                         alloc_hint = em->block_start;
 947                         free_extent_map(em);
 948                 }
 949         }
 950         read_unlock(&em_tree->lock);
 951
 952         return alloc_hint;
 953 }
 954
 955 /*
 956  * when extent_io.c finds a delayed allocation range in the file,
 957  * the call backs end up in this code.  The basic idea is to
 958  * allocate extents on disk for the range, and create ordered data structs
 959  * in ram to track those extents.
 960  *
 961  * locked_page is the page that writepage had locked already.  We use
 962  * it to make sure we don't do extra locks or unlocks.
 963  *
 964  * *page_started is set to one if we unlock locked_page and do everything
 965  * required to start IO on it.  It may be clean and already done with
 966  * IO when we return.
 967  */
 968 static noinline int cow_file_range(struct inode *inode,
 969                                    struct page *locked_page,
 970                                    u64 start, u64 end, int *page_started,
 971                                    unsigned long *nr_written, int unlock)
 972 {
 973         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 974         struct btrfs_root *root = BTRFS_I(inode)->root;
 975         u64 alloc_hint = 0;
 976         u64 num_bytes;
 977         unsigned long ram_size;
 978         u64 cur_alloc_size = 0;
 979         u64 blocksize = fs_info->sectorsize;
 980         struct btrfs_key ins;
 981         struct extent_map *em;
 982         unsigned clear_bits;
 983         unsigned long page_ops;
 984         bool extent_reserved = false;
 985         int ret = 0;
 986
 987         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
 988                 WARN_ON_ONCE(1);
 989                 ret = -EINVAL;
 990                 goto out_unlock;
 991         }
 992
 993         num_bytes = ALIGN(end - start + 1, blocksize);
 994         num_bytes = max(blocksize,  num_bytes);
 995         ASSERT(num_bytes <= btrfs_super_total_bytes(fs_info->super_copy));
 996
 997         inode_should_defrag(BTRFS_I(inode), start, end, num_bytes, SZ_64K);
 998
 999         if (start == 0) {
1000                 /* lets try to make an inline extent */
1001                 ret = cow_file_range_inline(inode, start, end, 0,
1002                                             BTRFS_COMPRESS_NONE, NULL);
1003                 if (ret == 0) {
1004                         /*
1005                          * We use DO_ACCOUNTING here because we need the
1006                          * delalloc_release_metadata to be run _after_ we drop
1007                          * our outstanding extent for clearing delalloc for this
1008                          * range.
1009                          */
1010                         extent_clear_unlock_delalloc(inode, start, end, NULL,
1011                                      EXTENT_LOCKED | EXTENT_DELALLOC |
1012                                      EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
1013                                      EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1014                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
1015                                      PAGE_END_WRITEBACK);
1016                         *nr_written = *nr_written +
1017                              (end - start + PAGE_SIZE) / PAGE_SIZE;
1018                         *page_started = 1;
1019                         goto out;
1020                 } else if (ret < 0) {
1021                         goto out_unlock;
1022                 }
1023         }
1024
1025         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
1026         btrfs_drop_extent_cache(BTRFS_I(inode), start,
1027                         start + num_bytes - 1, 0);
1028
1029         while (num_bytes > 0) {
1030                 cur_alloc_size = num_bytes;
1031                 ret = btrfs_reserve_extent(root, cur_alloc_size, cur_alloc_size,
1032                                            fs_info->sectorsize, 0, alloc_hint,
1033                                            &ins, 1, 1);
1034                 if (ret < 0)
1035                         goto out_unlock;
1036                 cur_alloc_size = ins.offset;
1037                 extent_reserved = true;
1038
1039                 ram_size = ins.offset;
1040                 em = create_io_em(inode, start, ins.offset, /* len */
1041                                   start, /* orig_start */
1042                                   ins.objectid, /* block_start */
1043                                   ins.offset, /* block_len */
1044                                   ins.offset, /* orig_block_len */
1045                                   ram_size, /* ram_bytes */
1046                                   BTRFS_COMPRESS_NONE, /* compress_type */
1047                                   BTRFS_ORDERED_REGULAR /* type */);
1048                 if (IS_ERR(em)) {
1049                         ret = PTR_ERR(em);
1050                         goto out_reserve;
1051                 }
1052                 free_extent_map(em);
1053
1054                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
1055                                                ram_size, cur_alloc_size, 0);
1056                 if (ret)
1057                         goto out_drop_extent_cache;
1058
1059                 if (root->root_key.objectid ==
1060                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1061                         ret = btrfs_reloc_clone_csums(inode, start,
1062                                                       cur_alloc_size);
1063                         /*
1064                          * Only drop cache here, and process as normal.
1065                          *
1066                          * We must not allow extent_clear_unlock_delalloc()
1067                          * at out_unlock label to free meta of this ordered
1068                          * extent, as its meta should be freed by
1069                          * btrfs_finish_ordered_io().
1070                          *
1071                          * So we must continue until @start is increased to
1072                          * skip current ordered extent.
1073                          */
1074                         if (ret)
1075                                 btrfs_drop_extent_cache(BTRFS_I(inode), start,
1076                                                 start + ram_size - 1, 0);
1077                 }
1078
1079                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1080
1081                 /* we're not doing compressed IO, don't unlock the first
1082                  * page (which the caller expects to stay locked), don't
1083                  * clear any dirty bits and don't set any writeback bits
1084                  *
1085                  * Do set the Private2 bit so we know this page was properly
1086                  * setup for writepage
1087                  */
1088                 page_ops = unlock ? PAGE_UNLOCK : 0;
1089                 page_ops |= PAGE_SET_PRIVATE2;
1090
1091                 extent_clear_unlock_delalloc(inode, start,
1092                                              start + ram_size - 1,
1093                                              locked_page,
1094                                              EXTENT_LOCKED | EXTENT_DELALLOC,
1095                                              page_ops);
1096                 if (num_bytes < cur_alloc_size)
1097                         num_bytes = 0;
1098                 else
1099                         num_bytes -= cur_alloc_size;
1100                 alloc_hint = ins.objectid + ins.offset;
1101                 start += cur_alloc_size;
1102                 extent_reserved = false;
1103
1104                 /*
1105                  * btrfs_reloc_clone_csums() error, since start is increased
1106                  * extent_clear_unlock_delalloc() at out_unlock label won't
1107                  * free metadata of current ordered extent, we're OK to exit.
1108                  */
1109                 if (ret)
1110                         goto out_unlock;
1111         }
1112 out:
1113         return ret;
1114
1115 out_drop_extent_cache:
1116         btrfs_drop_extent_cache(BTRFS_I(inode), start, start + ram_size - 1, 0);
1117 out_reserve:
1118         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1119         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
1120 out_unlock:
1121         clear_bits = EXTENT_LOCKED | EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
1122                 EXTENT_DEFRAG | EXTENT_CLEAR_META_RESV;
1123         page_ops = PAGE_UNLOCK | PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
1124                 PAGE_END_WRITEBACK;
1125         /*
1126          * If we reserved an extent for our delalloc range (or a subrange) and
1127          * failed to create the respective ordered extent, then it means that
1128          * when we reserved the extent we decremented the extent's size from
1129          * the data space_info's bytes_may_use counter and incremented the
1130          * space_info's bytes_reserved counter by the same amount. We must make
1131          * sure extent_clear_unlock_delalloc() does not try to decrement again
1132          * the data space_info's bytes_may_use counter, therefore we do not pass
1133          * it the flag EXTENT_CLEAR_DATA_RESV.
1134          */
1135         if (extent_reserved) {
1136                 extent_clear_unlock_delalloc(inode, start,
1137                                              start + cur_alloc_size,
1138                                              locked_page,
1139                                              clear_bits,
1140                                              page_ops);
1141                 start += cur_alloc_size;
1142                 if (start >= end)
1143                         goto out;
1144         }
1145         extent_clear_unlock_delalloc(inode, start, end, locked_page,
1146                                      clear_bits | EXTENT_CLEAR_DATA_RESV,
1147                                      page_ops);
1148         goto out;
1149 }
1150
1151 /*
1152  * work queue call back to started compression on a file and pages
1153  */
1154 static noinline void async_cow_start(struct btrfs_work *work)
1155 {
1156         struct async_chunk *async_chunk;
1157         int compressed_extents;
1158
1159         async_chunk = container_of(work, struct async_chunk, work);
1160
1161         compressed_extents = compress_file_range(async_chunk);
1162         if (compressed_extents == 0) {
1163                 btrfs_add_delayed_iput(async_chunk->inode);
1164                 async_chunk->inode = NULL;
1165         }
1166 }
1167
1168 /*
1169  * work queue call back to submit previously compressed pages
1170  */
1171 static noinline void async_cow_submit(struct btrfs_work *work)
1172 {
1173         struct async_chunk *async_chunk = container_of(work, struct async_chunk,
1174                                                      work);
1175         struct btrfs_fs_info *fs_info = btrfs_work_owner(work);
1176         unsigned long nr_pages;
1177
1178         nr_pages = (async_chunk->end - async_chunk->start + PAGE_SIZE) >>
1179                 PAGE_SHIFT;
1180
1181         /* atomic_sub_return implies a barrier */
1182         if (atomic_sub_return(nr_pages, &fs_info->async_delalloc_pages) <
1183             5 * SZ_1M)
1184                 cond_wake_up_nomb(&fs_info->async_submit_wait);
1185
1186         /*
1187          * ->inode could be NULL if async_chunk_start has failed to compress,
1188          * in which case we don't have anything to submit, yet we need to
1189          * always adjust ->async_delalloc_pages as its paired with the init
1190          * happening in cow_file_range_async
1191          */
1192         if (async_chunk->inode)
1193                 submit_compressed_extents(async_chunk);
1194 }
1195
1196 static noinline void async_cow_free(struct btrfs_work *work)
1197 {
1198         struct async_chunk *async_chunk;
1199
1200         async_chunk = container_of(work, struct async_chunk, work);
1201         if (async_chunk->inode)
1202                 btrfs_add_delayed_iput(async_chunk->inode);
1203         if (async_chunk->blkcg_css)
1204                 css_put(async_chunk->blkcg_css);
1205         /*
1206          * Since the pointer to 'pending' is at the beginning of the array of
1207          * async_chunk's, freeing it ensures the whole array has been freed.
1208          */
1209         if (atomic_dec_and_test(async_chunk->pending))
1210                 kvfree(async_chunk->pending);
1211 }
1212
1213 static int cow_file_range_async(struct inode *inode,
1214                                 struct writeback_control *wbc,
1215                                 struct page *locked_page,
1216                                 u64 start, u64 end, int *page_started,
1217                                 unsigned long *nr_written)
1218 {
1219         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1220         struct cgroup_subsys_state *blkcg_css = wbc_blkcg_css(wbc);
1221         struct async_cow *ctx;
1222         struct async_chunk *async_chunk;
1223         unsigned long nr_pages;
1224         u64 cur_end;
1225         u64 num_chunks = DIV_ROUND_UP(end - start, SZ_512K);
1226         int i;
1227         bool should_compress;
1228         unsigned nofs_flag;
1229         const unsigned int write_flags = wbc_to_write_flags(wbc);
1230
1231         unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
1232
1233         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1234             !btrfs_test_opt(fs_info, FORCE_COMPRESS)) {
1235                 num_chunks = 1;
1236                 should_compress = false;
1237         } else {
1238                 should_compress = true;
1239         }
1240
1241         nofs_flag = memalloc_nofs_save();
1242         ctx = kvmalloc(struct_size(ctx, chunks, num_chunks), GFP_KERNEL);
1243         memalloc_nofs_restore(nofs_flag);
1244
1245         if (!ctx) {
1246                 unsigned clear_bits = EXTENT_LOCKED | EXTENT_DELALLOC |
1247                         EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
1248                         EXTENT_DO_ACCOUNTING;
1249                 unsigned long page_ops = PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
1250                         PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
1251                         PAGE_SET_ERROR;
1252
1253                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1254                                              clear_bits, page_ops);
1255                 return -ENOMEM;
1256         }
1257
1258         async_chunk = ctx->chunks;
1259         atomic_set(&ctx->num_chunks, num_chunks);
1260
1261         for (i = 0; i < num_chunks; i++) {
1262                 if (should_compress)
1263                         cur_end = min(end, start + SZ_512K - 1);
1264                 else
1265                         cur_end = end;
1266
1267                 /*
1268                  * igrab is called higher up in the call chain, take only the
1269                  * lightweight reference for the callback lifetime
1270                  */
1271                 ihold(inode);
1272                 async_chunk[i].pending = &ctx->num_chunks;
1273                 async_chunk[i].inode = inode;
1274                 async_chunk[i].start = start;
1275                 async_chunk[i].end = cur_end;
1276                 async_chunk[i].write_flags = write_flags;
1277                 INIT_LIST_HEAD(&async_chunk[i].extents);
1278
1279                 /*
1280                  * The locked_page comes all the way from writepage and its
1281                  * the original page we were actually given.  As we spread
1282                  * this large delalloc region across multiple async_chunk
1283                  * structs, only the first struct needs a pointer to locked_page
1284                  *
1285                  * This way we don't need racey decisions about who is supposed
1286                  * to unlock it.
1287                  */
1288                 if (locked_page) {
1289                         /*
1290                          * Depending on the compressibility, the pages might or
1291                          * might not go through async.  We want all of them to
1292                          * be accounted against wbc once.  Let's do it here
1293                          * before the paths diverge.  wbc accounting is used
1294                          * only for foreign writeback detection and doesn't
1295                          * need full accuracy.  Just account the whole thing
1296                          * against the first page.
1297                          */
1298                         wbc_account_cgroup_owner(wbc, locked_page,
1299                                                  cur_end - start);
1300                         async_chunk[i].locked_page = locked_page;
1301                         locked_page = NULL;
1302                 } else {
1303                         async_chunk[i].locked_page = NULL;
1304                 }
1305
1306                 if (blkcg_css != blkcg_root_css) {
1307                         css_get(blkcg_css);
1308                         async_chunk[i].blkcg_css = blkcg_css;
1309                 } else {
1310                         async_chunk[i].blkcg_css = NULL;
1311                 }
1312
1313                 btrfs_init_work(&async_chunk[i].work, async_cow_start,
1314                                 async_cow_submit, async_cow_free);
1315
1316                 nr_pages = DIV_ROUND_UP(cur_end - start, PAGE_SIZE);
1317                 atomic_add(nr_pages, &fs_info->async_delalloc_pages);
1318
1319                 btrfs_queue_work(fs_info->delalloc_workers, &async_chunk[i].work);
1320
1321                 *nr_written += nr_pages;
1322                 start = cur_end + 1;
1323         }
1324         *page_started = 1;
1325         return 0;
1326 }
1327
1328 static noinline int csum_exist_in_range(struct btrfs_fs_info *fs_info,
1329                                         u64 bytenr, u64 num_bytes)
1330 {
1331         int ret;
1332         struct btrfs_ordered_sum *sums;
1333         LIST_HEAD(list);
1334
1335         ret = btrfs_lookup_csums_range(fs_info->csum_root, bytenr,
1336                                        bytenr + num_bytes - 1, &list, 0);
1337         if (ret == 0 && list_empty(&list))
1338                 return 0;
1339
1340         while (!list_empty(&list)) {
1341                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1342                 list_del(&sums->list);
1343                 kfree(sums);
1344         }
1345         if (ret < 0)
1346                 return ret;
1347         return 1;
1348 }
1349
1350 /*
1351  * when nowcow writeback call back.  This checks for snapshots or COW copies
1352  * of the extents that exist in the file, and COWs the file as required.
1353  *
1354  * If no cow copies or snapshots exist, we write directly to the existing
1355  * blocks on disk
1356  */
1357 static noinline int run_delalloc_nocow(struct inode *inode,
1358                                        struct page *locked_page,
1359                                        const u64 start, const u64 end,
1360                                        int *page_started, int force,
1361                                        unsigned long *nr_written)
1362 {
1363         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1364         struct btrfs_root *root = BTRFS_I(inode)->root;
1365         struct btrfs_path *path;
1366         u64 cow_start = (u64)-1;
1367         u64 cur_offset = start;
1368         int ret;
1369         bool check_prev = true;
1370         const bool freespace_inode = btrfs_is_free_space_inode(BTRFS_I(inode));
1371         u64 ino = btrfs_ino(BTRFS_I(inode));
1372         bool nocow = false;
1373         u64 disk_bytenr = 0;
1374
1375         path = btrfs_alloc_path();
1376         if (!path) {
1377                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1378                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1379                                              EXTENT_DO_ACCOUNTING |
1380                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1381                                              PAGE_CLEAR_DIRTY |
1382                                              PAGE_SET_WRITEBACK |
1383                                              PAGE_END_WRITEBACK);
1384                 return -ENOMEM;
1385         }
1386
1387         while (1) {
1388                 struct btrfs_key found_key;
1389                 struct btrfs_file_extent_item *fi;
1390                 struct extent_buffer *leaf;
1391                 u64 extent_end;
1392                 u64 extent_offset;
1393                 u64 num_bytes = 0;
1394                 u64 disk_num_bytes;
1395                 u64 ram_bytes;
1396                 int extent_type;
1397
1398                 nocow = false;
1399
1400                 ret = btrfs_lookup_file_extent(NULL, root, path, ino,
1401                                                cur_offset, 0);
1402                 if (ret < 0)
1403                         goto error;
1404
1405                 /*
1406                  * If there is no extent for our range when doing the initial
1407                  * search, then go back to the previous slot as it will be the
1408                  * one containing the search offset
1409                  */
1410                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1411                         leaf = path->nodes[0];
1412                         btrfs_item_key_to_cpu(leaf, &found_key,
1413                                               path->slots[0] - 1);
1414                         if (found_key.objectid == ino &&
1415                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1416                                 path->slots[0]--;
1417                 }
1418                 check_prev = false;
1419 next_slot:
1420                 /* Go to next leaf if we have exhausted the current one */
1421                 leaf = path->nodes[0];
1422                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1423                         ret = btrfs_next_leaf(root, path);
1424                         if (ret < 0) {
1425                                 if (cow_start != (u64)-1)
1426                                         cur_offset = cow_start;
1427                                 goto error;
1428                         }
1429                         if (ret > 0)
1430                                 break;
1431                         leaf = path->nodes[0];
1432                 }
1433
1434                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1435
1436                 /* Didn't find anything for our INO */
1437                 if (found_key.objectid > ino)
1438                         break;
1439                 /*
1440                  * Keep searching until we find an EXTENT_ITEM or there are no
1441                  * more extents for this inode
1442                  */
1443                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1444                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1445                         path->slots[0]++;
1446                         goto next_slot;
1447                 }
1448
1449                 /* Found key is not EXTENT_DATA_KEY or starts after req range */
1450                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1451                     found_key.offset > end)
1452                         break;
1453
1454                 /*
1455                  * If the found extent starts after requested offset, then
1456                  * adjust extent_end to be right before this extent begins
1457                  */
1458                 if (found_key.offset > cur_offset) {
1459                         extent_end = found_key.offset;
1460                         extent_type = 0;
1461                         goto out_check;
1462                 }
1463
1464                 /*
1465                  * Found extent which begins before our range and potentially
1466                  * intersect it
1467                  */
1468                 fi = btrfs_item_ptr(leaf, path->slots[0],
1469                                     struct btrfs_file_extent_item);
1470                 extent_type = btrfs_file_extent_type(leaf, fi);
1471
1472                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1473                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1474                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1475                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1476                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1477                         extent_end = found_key.offset +
1478                                 btrfs_file_extent_num_bytes(leaf, fi);
1479                         disk_num_bytes =
1480                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1481                         /*
1482                          * If extent we got ends before our range starts, skip
1483                          * to next extent
1484                          */
1485                         if (extent_end <= start) {
1486                                 path->slots[0]++;
1487                                 goto next_slot;
1488                         }
1489                         /* Skip holes */
1490                         if (disk_bytenr == 0)
1491                                 goto out_check;
1492                         /* Skip compressed/encrypted/encoded extents */
1493                         if (btrfs_file_extent_compression(leaf, fi) ||
1494                             btrfs_file_extent_encryption(leaf, fi) ||
1495                             btrfs_file_extent_other_encoding(leaf, fi))
1496                                 goto out_check;
1497                         /*
1498                          * If extent is created before the last volume's snapshot
1499                          * this implies the extent is shared, hence we can't do
1500                          * nocow. This is the same check as in
1501                          * btrfs_cross_ref_exist but without calling
1502                          * btrfs_search_slot.
1503                          */
1504                         if (!freespace_inode &&
1505                             btrfs_file_extent_generation(leaf, fi) <=
1506                             btrfs_root_last_snapshot(&root->root_item))
1507                                 goto out_check;
1508                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1509                                 goto out_check;
1510                         /* If extent is RO, we must COW it */
1511                         if (btrfs_extent_readonly(fs_info, disk_bytenr))
1512                                 goto out_check;
1513                         ret = btrfs_cross_ref_exist(root, ino,
1514                                                     found_key.offset -
1515                                                     extent_offset, disk_bytenr);
1516                         if (ret) {
1517                                 /*
1518                                  * ret could be -EIO if the above fails to read
1519                                  * metadata.
1520                                  */
1521                                 if (ret < 0) {
1522                                         if (cow_start != (u64)-1)
1523                                                 cur_offset = cow_start;
1524                                         goto error;
1525                                 }
1526
1527                                 WARN_ON_ONCE(freespace_inode);
1528                                 goto out_check;
1529                         }
1530                         disk_bytenr += extent_offset;
1531                         disk_bytenr += cur_offset - found_key.offset;
1532                         num_bytes = min(end + 1, extent_end) - cur_offset;
1533                         /*
1534                          * If there are pending snapshots for this root, we
1535                          * fall into common COW way
1536                          */
1537                         if (!freespace_inode && atomic_read(&root->snapshot_force_cow))
1538                                 goto out_check;
1539                         /*
1540                          * force cow if csum exists in the range.
1541                          * this ensure that csum for a given extent are
1542                          * either valid or do not exist.
1543                          */
1544                         ret = csum_exist_in_range(fs_info, disk_bytenr,
1545                                                   num_bytes);
1546                         if (ret) {
1547                                 /*
1548                                  * ret could be -EIO if the above fails to read
1549                                  * metadata.
1550                                  */
1551                                 if (ret < 0) {
1552                                         if (cow_start != (u64)-1)
1553                                                 cur_offset = cow_start;
1554                                         goto error;
1555                                 }
1556                                 WARN_ON_ONCE(freespace_inode);
1557                                 goto out_check;
1558                         }
1559                         if (!btrfs_inc_nocow_writers(fs_info, disk_bytenr))
1560                                 goto out_check;
1561                         nocow = true;
1562                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1563                         extent_end = found_key.offset + ram_bytes;
1564                         extent_end = ALIGN(extent_end, fs_info->sectorsize);
1565                         /* Skip extents outside of our requested range */
1566                         if (extent_end <= start) {
1567                                 path->slots[0]++;
1568                                 goto next_slot;
1569                         }
1570                 } else {
1571                         /* If this triggers then we have a memory corruption */
1572                         BUG();
1573                 }
1574 out_check:
1575                 /*
1576                  * If nocow is false then record the beginning of the range
1577                  * that needs to be COWed
1578                  */
1579                 if (!nocow) {
1580                         if (cow_start == (u64)-1)
1581                                 cow_start = cur_offset;
1582                         cur_offset = extent_end;
1583                         if (cur_offset > end)
1584                                 break;
1585                         path->slots[0]++;
1586                         goto next_slot;
1587                 }
1588
1589                 btrfs_release_path(path);
1590
1591                 /*
1592                  * COW range from cow_start to found_key.offset - 1. As the key
1593                  * will contain the beginning of the first extent that can be
1594                  * NOCOW, following one which needs to be COW'ed
1595                  */
1596                 if (cow_start != (u64)-1) {
1597                         ret = cow_file_range(inode, locked_page,
1598                                              cow_start, found_key.offset - 1,
1599                                              page_started, nr_written, 1);
1600                         if (ret) {
1601                                 if (nocow)
1602                                         btrfs_dec_nocow_writers(fs_info,
1603                                                                 disk_bytenr);
1604                                 goto error;
1605                         }
1606                         cow_start = (u64)-1;
1607                 }
1608
1609                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1610                         u64 orig_start = found_key.offset - extent_offset;
1611                         struct extent_map *em;
1612
1613                         em = create_io_em(inode, cur_offset, num_bytes,
1614                                           orig_start,
1615                                           disk_bytenr, /* block_start */
1616                                           num_bytes, /* block_len */
1617                                           disk_num_bytes, /* orig_block_len */
1618                                           ram_bytes, BTRFS_COMPRESS_NONE,
1619                                           BTRFS_ORDERED_PREALLOC);
1620                         if (IS_ERR(em)) {
1621                                 if (nocow)
1622                                         btrfs_dec_nocow_writers(fs_info,
1623                                                                 disk_bytenr);
1624                                 ret = PTR_ERR(em);
1625                                 goto error;
1626                         }
1627                         free_extent_map(em);
1628                         ret = btrfs_add_ordered_extent(inode, cur_offset,
1629                                                        disk_bytenr, num_bytes,
1630                                                        num_bytes,
1631                                                        BTRFS_ORDERED_PREALLOC);
1632                         if (ret) {
1633                                 btrfs_drop_extent_cache(BTRFS_I(inode),
1634                                                         cur_offset,
1635                                                         cur_offset + num_bytes - 1,
1636                                                         0);
1637                                 goto error;
1638                         }
1639                 } else {
1640                         ret = btrfs_add_ordered_extent(inode, cur_offset,
1641                                                        disk_bytenr, num_bytes,
1642                                                        num_bytes,
1643                                                        BTRFS_ORDERED_NOCOW);
1644                         if (ret)
1645                                 goto error;
1646                 }
1647
1648                 if (nocow)
1649                         btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1650                 nocow = false;
1651
1652                 if (root->root_key.objectid ==
1653                     BTRFS_DATA_RELOC_TREE_OBJECTID)
1654                         /*
1655                          * Error handled later, as we must prevent
1656                          * extent_clear_unlock_delalloc() in error handler
1657                          * from freeing metadata of created ordered extent.
1658                          */
1659                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1660                                                       num_bytes);
1661
1662                 extent_clear_unlock_delalloc(inode, cur_offset,
1663                                              cur_offset + num_bytes - 1,
1664                                              locked_page, EXTENT_LOCKED |
1665                                              EXTENT_DELALLOC |
1666                                              EXTENT_CLEAR_DATA_RESV,
1667                                              PAGE_UNLOCK | PAGE_SET_PRIVATE2);
1668
1669                 cur_offset = extent_end;
1670
1671                 /*
1672                  * btrfs_reloc_clone_csums() error, now we're OK to call error
1673                  * handler, as metadata for created ordered extent will only
1674                  * be freed by btrfs_finish_ordered_io().
1675                  */
1676                 if (ret)
1677                         goto error;
1678                 if (cur_offset > end)
1679                         break;
1680         }
1681         btrfs_release_path(path);
1682
1683         if (cur_offset <= end && cow_start == (u64)-1)
1684                 cow_start = cur_offset;
1685
1686         if (cow_start != (u64)-1) {
1687                 cur_offset = end;
1688                 ret = cow_file_range(inode, locked_page, cow_start, end,
1689                                      page_started, nr_written, 1);
1690                 if (ret)
1691                         goto error;
1692         }
1693
1694 error:
1695         if (nocow)
1696                 btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1697
1698         if (ret && cur_offset < end)
1699                 extent_clear_unlock_delalloc(inode, cur_offset, end,
1700                                              locked_page, EXTENT_LOCKED |
1701                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1702                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1703                                              PAGE_CLEAR_DIRTY |
1704                                              PAGE_SET_WRITEBACK |
1705                                              PAGE_END_WRITEBACK);
1706         btrfs_free_path(path);
1707         return ret;
1708 }
1709
1710 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1711 {
1712
1713         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1714             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1715                 return 0;
1716
1717         /*
1718          * @defrag_bytes is a hint value, no spinlock held here,
1719          * if is not zero, it means the file is defragging.
1720          * Force cow if given extent needs to be defragged.
1721          */
1722         if (BTRFS_I(inode)->defrag_bytes &&
1723             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1724                            EXTENT_DEFRAG, 0, NULL))
1725                 return 1;
1726
1727         return 0;
1728 }
1729
1730 /*
1731  * Function to process delayed allocation (create CoW) for ranges which are
1732  * being touched for the first time.
1733  */
1734 int btrfs_run_delalloc_range(struct inode *inode, struct page *locked_page,
1735                 u64 start, u64 end, int *page_started, unsigned long *nr_written,
1736                 struct writeback_control *wbc)
1737 {
1738         int ret;
1739         int force_cow = need_force_cow(inode, start, end);
1740
1741         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1742                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1743                                          page_started, 1, nr_written);
1744         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1745                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1746                                          page_started, 0, nr_written);
1747         } else if (!inode_can_compress(inode) ||
1748                    !inode_need_compress(inode, start, end)) {
1749                 ret = cow_file_range(inode, locked_page, start, end,
1750                                       page_started, nr_written, 1);
1751         } else {
1752                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1753                         &BTRFS_I(inode)->runtime_flags);
1754                 ret = cow_file_range_async(inode, wbc, locked_page, start, end,
1755                                            page_started, nr_written);
1756         }
1757         if (ret)
1758                 btrfs_cleanup_ordered_extents(inode, locked_page, start,
1759                                               end - start + 1);
1760         return ret;
1761 }
1762
1763 void btrfs_split_delalloc_extent(struct inode *inode,
1764                                  struct extent_state *orig, u64 split)
1765 {
1766         u64 size;
1767
1768         /* not delalloc, ignore it */
1769         if (!(orig->state & EXTENT_DELALLOC))
1770                 return;
1771
1772         size = orig->end - orig->start + 1;
1773         if (size > BTRFS_MAX_EXTENT_SIZE) {
1774                 u32 num_extents;
1775                 u64 new_size;
1776
1777                 /*
1778                  * See the explanation in btrfs_merge_delalloc_extent, the same
1779                  * applies here, just in reverse.
1780                  */
1781                 new_size = orig->end - split + 1;
1782                 num_extents = count_max_extents(new_size);
1783                 new_size = split - orig->start;
1784                 num_extents += count_max_extents(new_size);
1785                 if (count_max_extents(size) >= num_extents)
1786                         return;
1787         }
1788
1789         spin_lock(&BTRFS_I(inode)->lock);
1790         btrfs_mod_outstanding_extents(BTRFS_I(inode), 1);
1791         spin_unlock(&BTRFS_I(inode)->lock);
1792 }
1793
1794 /*
1795  * Handle merged delayed allocation extents so we can keep track of new extents
1796  * that are just merged onto old extents, such as when we are doing sequential
1797  * writes, so we can properly account for the metadata space we'll need.
1798  */
1799 void btrfs_merge_delalloc_extent(struct inode *inode, struct extent_state *new,
1800                                  struct extent_state *other)
1801 {
1802         u64 new_size, old_size;
1803         u32 num_extents;
1804
1805         /* not delalloc, ignore it */
1806         if (!(other->state & EXTENT_DELALLOC))
1807                 return;
1808
1809         if (new->start > other->start)
1810                 new_size = new->end - other->start + 1;
1811         else
1812                 new_size = other->end - new->start + 1;
1813
1814         /* we're not bigger than the max, unreserve the space and go */
1815         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1816                 spin_lock(&BTRFS_I(inode)->lock);
1817                 btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1818                 spin_unlock(&BTRFS_I(inode)->lock);
1819                 return;
1820         }
1821
1822         /*
1823          * We have to add up either side to figure out how many extents were
1824          * accounted for before we merged into one big extent.  If the number of
1825          * extents we accounted for is <= the amount we need for the new range
1826          * then we can return, otherwise drop.  Think of it like this
1827          *
1828          * [ 4k][MAX_SIZE]
1829          *
1830          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1831          * need 2 outstanding extents, on one side we have 1 and the other side
1832          * we have 1 so they are == and we can return.  But in this case
1833          *
1834          * [MAX_SIZE+4k][MAX_SIZE+4k]
1835          *
1836          * Each range on their own accounts for 2 extents, but merged together
1837          * they are only 3 extents worth of accounting, so we need to drop in
1838          * this case.
1839          */
1840         old_size = other->end - other->start + 1;
1841         num_extents = count_max_extents(old_size);
1842         old_size = new->end - new->start + 1;
1843         num_extents += count_max_extents(old_size);
1844         if (count_max_extents(new_size) >= num_extents)
1845                 return;
1846
1847         spin_lock(&BTRFS_I(inode)->lock);
1848         btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1849         spin_unlock(&BTRFS_I(inode)->lock);
1850 }
1851
1852 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1853                                       struct inode *inode)
1854 {
1855         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1856
1857         spin_lock(&root->delalloc_lock);
1858         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1859                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1860                               &root->delalloc_inodes);
1861                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1862                         &BTRFS_I(inode)->runtime_flags);
1863                 root->nr_delalloc_inodes++;
1864                 if (root->nr_delalloc_inodes == 1) {
1865                         spin_lock(&fs_info->delalloc_root_lock);
1866                         BUG_ON(!list_empty(&root->delalloc_root));
1867                         list_add_tail(&root->delalloc_root,
1868                                       &fs_info->delalloc_roots);
1869                         spin_unlock(&fs_info->delalloc_root_lock);
1870                 }
1871         }
1872         spin_unlock(&root->delalloc_lock);
1873 }
1874
1875
1876 void __btrfs_del_delalloc_inode(struct btrfs_root *root,
1877                                 struct btrfs_inode *inode)
1878 {
1879         struct btrfs_fs_info *fs_info = root->fs_info;
1880
1881         if (!list_empty(&inode->delalloc_inodes)) {
1882                 list_del_init(&inode->delalloc_inodes);
1883                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1884                           &inode->runtime_flags);
1885                 root->nr_delalloc_inodes--;
1886                 if (!root->nr_delalloc_inodes) {
1887                         ASSERT(list_empty(&root->delalloc_inodes));
1888                         spin_lock(&fs_info->delalloc_root_lock);
1889                         BUG_ON(list_empty(&root->delalloc_root));
1890                         list_del_init(&root->delalloc_root);
1891                         spin_unlock(&fs_info->delalloc_root_lock);
1892                 }
1893         }
1894 }
1895
1896 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1897                                      struct btrfs_inode *inode)
1898 {
1899         spin_lock(&root->delalloc_lock);
1900         __btrfs_del_delalloc_inode(root, inode);
1901         spin_unlock(&root->delalloc_lock);
1902 }
1903
1904 /*
1905  * Properly track delayed allocation bytes in the inode and to maintain the
1906  * list of inodes that have pending delalloc work to be done.
1907  */
1908 void btrfs_set_delalloc_extent(struct inode *inode, struct extent_state *state,
1909                                unsigned *bits)
1910 {
1911         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1912
1913         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
1914                 WARN_ON(1);
1915         /*
1916          * set_bit and clear bit hooks normally require _irqsave/restore
1917          * but in this case, we are only testing for the DELALLOC
1918          * bit, which is only set or cleared with irqs on
1919          */
1920         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1921                 struct btrfs_root *root = BTRFS_I(inode)->root;
1922                 u64 len = state->end + 1 - state->start;
1923                 u32 num_extents = count_max_extents(len);
1924                 bool do_list = !btrfs_is_free_space_inode(BTRFS_I(inode));
1925
1926                 spin_lock(&BTRFS_I(inode)->lock);
1927                 btrfs_mod_outstanding_extents(BTRFS_I(inode), num_extents);
1928                 spin_unlock(&BTRFS_I(inode)->lock);
1929
1930                 /* For sanity tests */
1931                 if (btrfs_is_testing(fs_info))
1932                         return;
1933
1934                 percpu_counter_add_batch(&fs_info->delalloc_bytes, len,
1935                                          fs_info->delalloc_batch);
1936                 spin_lock(&BTRFS_I(inode)->lock);
1937                 BTRFS_I(inode)->delalloc_bytes += len;
1938                 if (*bits & EXTENT_DEFRAG)
1939                         BTRFS_I(inode)->defrag_bytes += len;
1940                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1941                                          &BTRFS_I(inode)->runtime_flags))
1942                         btrfs_add_delalloc_inodes(root, inode);
1943                 spin_unlock(&BTRFS_I(inode)->lock);
1944         }
1945
1946         if (!(state->state & EXTENT_DELALLOC_NEW) &&
1947             (*bits & EXTENT_DELALLOC_NEW)) {
1948                 spin_lock(&BTRFS_I(inode)->lock);
1949                 BTRFS_I(inode)->new_delalloc_bytes += state->end + 1 -
1950                         state->start;
1951                 spin_unlock(&BTRFS_I(inode)->lock);
1952         }
1953 }
1954
1955 /*
1956  * Once a range is no longer delalloc this function ensures that proper
1957  * accounting happens.
1958  */
1959 void btrfs_clear_delalloc_extent(struct inode *vfs_inode,
1960                                  struct extent_state *state, unsigned *bits)
1961 {
1962         struct btrfs_inode *inode = BTRFS_I(vfs_inode);
1963         struct btrfs_fs_info *fs_info = btrfs_sb(vfs_inode->i_sb);
1964         u64 len = state->end + 1 - state->start;
1965         u32 num_extents = count_max_extents(len);
1966
1967         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG)) {
1968                 spin_lock(&inode->lock);
1969                 inode->defrag_bytes -= len;
1970                 spin_unlock(&inode->lock);
1971         }
1972
1973         /*
1974          * set_bit and clear bit hooks normally require _irqsave/restore
1975          * but in this case, we are only testing for the DELALLOC
1976          * bit, which is only set or cleared with irqs on
1977          */
1978         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1979                 struct btrfs_root *root = inode->root;
1980                 bool do_list = !btrfs_is_free_space_inode(inode);
1981
1982                 spin_lock(&inode->lock);
1983                 btrfs_mod_outstanding_extents(inode, -num_extents);
1984                 spin_unlock(&inode->lock);
1985
1986                 /*
1987                  * We don't reserve metadata space for space cache inodes so we
1988                  * don't need to call delalloc_release_metadata if there is an
1989                  * error.
1990                  */
1991                 if (*bits & EXTENT_CLEAR_META_RESV &&
1992                     root != fs_info->tree_root)
1993                         btrfs_delalloc_release_metadata(inode, len, false);
1994
1995                 /* For sanity tests. */
1996                 if (btrfs_is_testing(fs_info))
1997                         return;
1998
1999                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID &&
2000                     do_list && !(state->state & EXTENT_NORESERVE) &&
2001                     (*bits & EXTENT_CLEAR_DATA_RESV))
2002                         btrfs_free_reserved_data_space_noquota(
2003                                         &inode->vfs_inode,
2004                                         state->start, len);
2005
2006                 percpu_counter_add_batch(&fs_info->delalloc_bytes, -len,
2007                                          fs_info->delalloc_batch);
2008                 spin_lock(&inode->lock);
2009                 inode->delalloc_bytes -= len;
2010                 if (do_list && inode->delalloc_bytes == 0 &&
2011                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
2012                                         &inode->runtime_flags))
2013                         btrfs_del_delalloc_inode(root, inode);
2014                 spin_unlock(&inode->lock);
2015         }
2016
2017         if ((state->state & EXTENT_DELALLOC_NEW) &&
2018             (*bits & EXTENT_DELALLOC_NEW)) {
2019                 spin_lock(&inode->lock);
2020                 ASSERT(inode->new_delalloc_bytes >= len);
2021                 inode->new_delalloc_bytes -= len;
2022                 spin_unlock(&inode->lock);
2023         }
2024 }
2025
2026 /*
2027  * btrfs_bio_fits_in_stripe - Checks whether the size of the given bio will fit
2028  * in a chunk's stripe. This function ensures that bios do not span a
2029  * stripe/chunk
2030  *
2031  * @page - The page we are about to add to the bio
2032  * @size - size we want to add to the bio
2033  * @bio - bio we want to ensure is smaller than a stripe
2034  * @bio_flags - flags of the bio
2035  *
2036  * return 1 if page cannot be added to the bio
2037  * return 0 if page can be added to the bio
2038  * return error otherwise
2039  */
2040 int btrfs_bio_fits_in_stripe(struct page *page, size_t size, struct bio *bio,
2041                              unsigned long bio_flags)
2042 {
2043         struct inode *inode = page->mapping->host;
2044         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2045         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
2046         u64 length = 0;
2047         u64 map_length;
2048         int ret;
2049         struct btrfs_io_geometry geom;
2050
2051         if (bio_flags & EXTENT_BIO_COMPRESSED)
2052                 return 0;
2053
2054         length = bio->bi_iter.bi_size;
2055         map_length = length;
2056         ret = btrfs_get_io_geometry(fs_info, btrfs_op(bio), logical, map_length,
2057                                     &geom);
2058         if (ret < 0)
2059                 return ret;
2060
2061         if (geom.len < length + size)
2062                 return 1;
2063         return 0;
2064 }
2065
2066 /*
2067  * in order to insert checksums into the metadata in large chunks,
2068  * we wait until bio submission time.   All the pages in the bio are
2069  * checksummed and sums are attached onto the ordered extent record.
2070  *
2071  * At IO completion time the cums attached on the ordered extent record
2072  * are inserted into the btree
2073  */
2074 static blk_status_t btrfs_submit_bio_start(void *private_data, struct bio *bio,
2075                                     u64 bio_offset)
2076 {
2077         struct inode *inode = private_data;
2078         blk_status_t ret = 0;
2079
2080         ret = btrfs_csum_one_bio(inode, bio, 0, 0);
2081         BUG_ON(ret); /* -ENOMEM */
2082         return 0;
2083 }
2084
2085 /*
2086  * extent_io.c submission hook. This does the right thing for csum calculation
2087  * on write, or reading the csums from the tree before a read.
2088  *
2089  * Rules about async/sync submit,
2090  * a) read:                             sync submit
2091  *
2092  * b) write without checksum:           sync submit
2093  *
2094  * c) write with checksum:
2095  *    c-1) if bio is issued by fsync:   sync submit
2096  *         (sync_writers != 0)
2097  *
2098  *    c-2) if root is reloc root:       sync submit
2099  *         (only in case of buffered IO)
2100  *
2101  *    c-3) otherwise:                   async submit
2102  */
2103 static blk_status_t btrfs_submit_bio_hook(struct inode *inode, struct bio *bio,
2104                                           int mirror_num,
2105                                           unsigned long bio_flags)
2106
2107 {
2108         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2109         struct btrfs_root *root = BTRFS_I(inode)->root;
2110         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
2111         blk_status_t ret = 0;
2112         int skip_sum;
2113         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
2114
2115         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
2116
2117         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
2118                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
2119
2120         if (bio_op(bio) != REQ_OP_WRITE) {
2121                 ret = btrfs_bio_wq_end_io(fs_info, bio, metadata);
2122                 if (ret)
2123                         goto out;
2124
2125                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
2126                         ret = btrfs_submit_compressed_read(inode, bio,
2127                                                            mirror_num,
2128                                                            bio_flags);
2129                         goto out;
2130                 } else if (!skip_sum) {
2131                         ret = btrfs_lookup_bio_sums(inode, bio, NULL);
2132                         if (ret)
2133                                 goto out;
2134                 }
2135                 goto mapit;
2136         } else if (async && !skip_sum) {
2137                 /* csum items have already been cloned */
2138                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2139                         goto mapit;
2140                 /* we're doing a write, do the async checksumming */
2141                 ret = btrfs_wq_submit_bio(fs_info, bio, mirror_num, bio_flags,
2142                                           0, inode, btrfs_submit_bio_start);
2143                 goto out;
2144         } else if (!skip_sum) {
2145                 ret = btrfs_csum_one_bio(inode, bio, 0, 0);
2146                 if (ret)
2147                         goto out;
2148         }
2149
2150 mapit:
2151         ret = btrfs_map_bio(fs_info, bio, mirror_num);
2152
2153 out:
2154         if (ret) {
2155                 bio->bi_status = ret;
2156                 bio_endio(bio);
2157         }
2158         return ret;
2159 }
2160
2161 /*
2162  * given a list of ordered sums record them in the inode.  This happens
2163  * at IO completion time based on sums calculated at bio submission time.
2164  */
2165 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
2166                              struct inode *inode, struct list_head *list)
2167 {
2168         struct btrfs_ordered_sum *sum;
2169         int ret;
2170
2171         list_for_each_entry(sum, list, list) {
2172                 trans->adding_csums = true;
2173                 ret = btrfs_csum_file_blocks(trans,
2174                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
2175                 trans->adding_csums = false;
2176                 if (ret)
2177                         return ret;
2178         }
2179         return 0;
2180 }
2181
2182 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
2183                               unsigned int extra_bits,
2184                               struct extent_state **cached_state)
2185 {
2186         WARN_ON(PAGE_ALIGNED(end));
2187         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
2188                                    extra_bits, cached_state);
2189 }
2190
2191 /* see btrfs_writepage_start_hook for details on why this is required */
2192 struct btrfs_writepage_fixup {
2193         struct page *page;
2194         struct btrfs_work work;
2195 };
2196
2197 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
2198 {
2199         struct btrfs_writepage_fixup *fixup;
2200         struct btrfs_ordered_extent *ordered;
2201         struct extent_state *cached_state = NULL;
2202         struct extent_changeset *data_reserved = NULL;
2203         struct page *page;
2204         struct inode *inode;
2205         u64 page_start;
2206         u64 page_end;
2207         int ret;
2208
2209         fixup = container_of(work, struct btrfs_writepage_fixup, work);
2210         page = fixup->page;
2211 again:
2212         lock_page(page);
2213         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
2214                 ClearPageChecked(page);
2215                 goto out_page;
2216         }
2217
2218         inode = page->mapping->host;
2219         page_start = page_offset(page);
2220         page_end = page_offset(page) + PAGE_SIZE - 1;
2221
2222         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end,
2223                          &cached_state);
2224
2225         /* already ordered? We're done */
2226         if (PagePrivate2(page))
2227                 goto out;
2228
2229         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
2230                                         PAGE_SIZE);
2231         if (ordered) {
2232                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2233                                      page_end, &cached_state);
2234                 unlock_page(page);
2235                 btrfs_start_ordered_extent(inode, ordered, 1);
2236                 btrfs_put_ordered_extent(ordered);
2237                 goto again;
2238         }
2239
2240         ret = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
2241                                            PAGE_SIZE);
2242         if (ret) {
2243                 mapping_set_error(page->mapping, ret);
2244                 end_extent_writepage(page, ret, page_start, page_end);
2245                 ClearPageChecked(page);
2246                 goto out;
2247          }
2248
2249         ret = btrfs_set_extent_delalloc(inode, page_start, page_end, 0,
2250                                         &cached_state);
2251         if (ret) {
2252                 mapping_set_error(page->mapping, ret);
2253                 end_extent_writepage(page, ret, page_start, page_end);
2254                 ClearPageChecked(page);
2255                 goto out_reserved;
2256         }
2257
2258         ClearPageChecked(page);
2259         set_page_dirty(page);
2260 out_reserved:
2261         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
2262         if (ret)
2263                 btrfs_delalloc_release_space(inode, data_reserved, page_start,
2264                                              PAGE_SIZE, true);
2265 out:
2266         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2267                              &cached_state);
2268 out_page:
2269         unlock_page(page);
2270         put_page(page);
2271         kfree(fixup);
2272         extent_changeset_free(data_reserved);
2273 }
2274
2275 /*
2276  * There are a few paths in the higher layers of the kernel that directly
2277  * set the page dirty bit without asking the filesystem if it is a
2278  * good idea.  This causes problems because we want to make sure COW
2279  * properly happens and the data=ordered rules are followed.
2280  *
2281  * In our case any range that doesn't have the ORDERED bit set
2282  * hasn't been properly setup for IO.  We kick off an async process
2283  * to fix it up.  The async helper will wait for ordered extents, set
2284  * the delalloc bit and make it safe to write the page.
2285  */
2286 int btrfs_writepage_cow_fixup(struct page *page, u64 start, u64 end)
2287 {
2288         struct inode *inode = page->mapping->host;
2289         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2290         struct btrfs_writepage_fixup *fixup;
2291
2292         /* this page is properly in the ordered list */
2293         if (TestClearPagePrivate2(page))
2294                 return 0;
2295
2296         if (PageChecked(page))
2297                 return -EAGAIN;
2298
2299         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2300         if (!fixup)
2301                 return -EAGAIN;
2302
2303         SetPageChecked(page);
2304         get_page(page);
2305         btrfs_init_work(&fixup->work, btrfs_writepage_fixup_worker, NULL, NULL);
2306         fixup->page = page;
2307         btrfs_queue_work(fs_info->fixup_workers, &fixup->work);
2308         return -EBUSY;
2309 }
2310
2311 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2312                                        struct inode *inode, u64 file_pos,
2313                                        u64 disk_bytenr, u64 disk_num_bytes,
2314                                        u64 num_bytes, u64 ram_bytes,
2315                                        u8 compression, u8 encryption,
2316                                        u16 other_encoding, int extent_type)
2317 {
2318         struct btrfs_root *root = BTRFS_I(inode)->root;
2319         struct btrfs_file_extent_item *fi;
2320         struct btrfs_path *path;
2321         struct extent_buffer *leaf;
2322         struct btrfs_key ins;
2323         u64 qg_released;
2324         int extent_inserted = 0;
2325         int ret;
2326
2327         path = btrfs_alloc_path();
2328         if (!path)
2329                 return -ENOMEM;
2330
2331         /*
2332          * we may be replacing one extent in the tree with another.
2333          * The new extent is pinned in the extent map, and we don't want
2334          * to drop it from the cache until it is completely in the btree.
2335          *
2336          * So, tell btrfs_drop_extents to leave this extent in the cache.
2337          * the caller is expected to unpin it and allow it to be merged
2338          * with the others.
2339          */
2340         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2341                                    file_pos + num_bytes, NULL, 0,
2342                                    1, sizeof(*fi), &extent_inserted);
2343         if (ret)
2344                 goto out;
2345
2346         if (!extent_inserted) {
2347                 ins.objectid = btrfs_ino(BTRFS_I(inode));
2348                 ins.offset = file_pos;
2349                 ins.type = BTRFS_EXTENT_DATA_KEY;
2350
2351                 path->leave_spinning = 1;
2352                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2353                                               sizeof(*fi));
2354                 if (ret)
2355                         goto out;
2356         }
2357         leaf = path->nodes[0];
2358         fi = btrfs_item_ptr(leaf, path->slots[0],
2359                             struct btrfs_file_extent_item);
2360         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2361         btrfs_set_file_extent_type(leaf, fi, extent_type);
2362         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2363         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2364         btrfs_set_file_extent_offset(leaf, fi, 0);
2365         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2366         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2367         btrfs_set_file_extent_compression(leaf, fi, compression);
2368         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2369         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2370
2371         btrfs_mark_buffer_dirty(leaf);
2372         btrfs_release_path(path);
2373
2374         inode_add_bytes(inode, num_bytes);
2375
2376         ins.objectid = disk_bytenr;
2377         ins.offset = disk_num_bytes;
2378         ins.type = BTRFS_EXTENT_ITEM_KEY;
2379
2380         /*
2381          * Release the reserved range from inode dirty range map, as it is
2382          * already moved into delayed_ref_head
2383          */
2384         ret = btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2385         if (ret < 0)
2386                 goto out;
2387         qg_released = ret;
2388         ret = btrfs_alloc_reserved_file_extent(trans, root,
2389                                                btrfs_ino(BTRFS_I(inode)),
2390                                                file_pos, qg_released, &ins);
2391 out:
2392         btrfs_free_path(path);
2393
2394         return ret;
2395 }
2396
2397 /* snapshot-aware defrag */
2398 struct sa_defrag_extent_backref {
2399         struct rb_node node;
2400         struct old_sa_defrag_extent *old;
2401         u64 root_id;
2402         u64 inum;
2403         u64 file_pos;
2404         u64 extent_offset;
2405         u64 num_bytes;
2406         u64 generation;
2407 };
2408
2409 struct old_sa_defrag_extent {
2410         struct list_head list;
2411         struct new_sa_defrag_extent *new;
2412
2413         u64 extent_offset;
2414         u64 bytenr;
2415         u64 offset;
2416         u64 len;
2417         int count;
2418 };
2419
2420 struct new_sa_defrag_extent {
2421         struct rb_root root;
2422         struct list_head head;
2423         struct btrfs_path *path;
2424         struct inode *inode;
2425         u64 file_pos;
2426         u64 len;
2427         u64 bytenr;
2428         u64 disk_len;
2429         u8 compress_type;
2430 };
2431
2432 static int backref_comp(struct sa_defrag_extent_backref *b1,
2433                         struct sa_defrag_extent_backref *b2)
2434 {
2435         if (b1->root_id < b2->root_id)
2436                 return -1;
2437         else if (b1->root_id > b2->root_id)
2438                 return 1;
2439
2440         if (b1->inum < b2->inum)
2441                 return -1;
2442         else if (b1->inum > b2->inum)
2443                 return 1;
2444
2445         if (b1->file_pos < b2->file_pos)
2446                 return -1;
2447         else if (b1->file_pos > b2->file_pos)
2448                 return 1;
2449
2450         /*
2451          * [------------------------------] ===> (a range of space)
2452          *     |<--->|   |<---->| =============> (fs/file tree A)
2453          * |<---------------------------->| ===> (fs/file tree B)
2454          *
2455          * A range of space can refer to two file extents in one tree while
2456          * refer to only one file extent in another tree.
2457          *
2458          * So we may process a disk offset more than one time(two extents in A)
2459          * and locate at the same extent(one extent in B), then insert two same
2460          * backrefs(both refer to the extent in B).
2461          */
2462         return 0;
2463 }
2464
2465 static void backref_insert(struct rb_root *root,
2466                            struct sa_defrag_extent_backref *backref)
2467 {
2468         struct rb_node **p = &root->rb_node;
2469         struct rb_node *parent = NULL;
2470         struct sa_defrag_extent_backref *entry;
2471         int ret;
2472
2473         while (*p) {
2474                 parent = *p;
2475                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
2476
2477                 ret = backref_comp(backref, entry);
2478                 if (ret < 0)
2479                         p = &(*p)->rb_left;
2480                 else
2481                         p = &(*p)->rb_right;
2482         }
2483
2484         rb_link_node(&backref->node, parent, p);
2485         rb_insert_color(&backref->node, root);
2486 }
2487
2488 /*
2489  * Note the backref might has changed, and in this case we just return 0.
2490  */
2491 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
2492                                        void *ctx)
2493 {
2494         struct btrfs_file_extent_item *extent;
2495         struct old_sa_defrag_extent *old = ctx;
2496         struct new_sa_defrag_extent *new = old->new;
2497         struct btrfs_path *path = new->path;
2498         struct btrfs_key key;
2499         struct btrfs_root *root;
2500         struct sa_defrag_extent_backref *backref;
2501         struct extent_buffer *leaf;
2502         struct inode *inode = new->inode;
2503         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2504         int slot;
2505         int ret;
2506         u64 extent_offset;
2507         u64 num_bytes;
2508
2509         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2510             inum == btrfs_ino(BTRFS_I(inode)))
2511                 return 0;
2512
2513         key.objectid = root_id;
2514         key.type = BTRFS_ROOT_ITEM_KEY;
2515         key.offset = (u64)-1;
2516
2517         root = btrfs_read_fs_root_no_name(fs_info, &key);
2518         if (IS_ERR(root)) {
2519                 if (PTR_ERR(root) == -ENOENT)
2520                         return 0;
2521                 WARN_ON(1);
2522                 btrfs_debug(fs_info, "inum=%llu, offset=%llu, root_id=%llu",
2523                          inum, offset, root_id);
2524                 return PTR_ERR(root);
2525         }
2526
2527         key.objectid = inum;
2528         key.type = BTRFS_EXTENT_DATA_KEY;
2529         if (offset > (u64)-1 << 32)
2530                 key.offset = 0;
2531         else
2532                 key.offset = offset;
2533
2534         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2535         if (WARN_ON(ret < 0))
2536                 return ret;
2537         ret = 0;
2538
2539         while (1) {
2540                 cond_resched();
2541
2542                 leaf = path->nodes[0];
2543                 slot = path->slots[0];
2544
2545                 if (slot >= btrfs_header_nritems(leaf)) {
2546                         ret = btrfs_next_leaf(root, path);
2547                         if (ret < 0) {
2548                                 goto out;
2549                         } else if (ret > 0) {
2550                                 ret = 0;
2551                                 goto out;
2552                         }
2553                         continue;
2554                 }
2555
2556                 path->slots[0]++;
2557
2558                 btrfs_item_key_to_cpu(leaf, &key, slot);
2559
2560                 if (key.objectid > inum)
2561                         goto out;
2562
2563                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2564                         continue;
2565
2566                 extent = btrfs_item_ptr(leaf, slot,
2567                                         struct btrfs_file_extent_item);
2568
2569                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2570                         continue;
2571
2572                 /*
2573                  * 'offset' refers to the exact key.offset,
2574                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2575                  * (key.offset - extent_offset).
2576                  */
2577                 if (key.offset != offset)
2578                         continue;
2579
2580                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2581                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2582
2583                 if (extent_offset >= old->extent_offset + old->offset +
2584                     old->len || extent_offset + num_bytes <=
2585                     old->extent_offset + old->offset)
2586                         continue;
2587                 break;
2588         }
2589
2590         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2591         if (!backref) {
2592                 ret = -ENOENT;
2593                 goto out;
2594         }
2595
2596         backref->root_id = root_id;
2597         backref->inum = inum;
2598         backref->file_pos = offset;
2599         backref->num_bytes = num_bytes;
2600         backref->extent_offset = extent_offset;
2601         backref->generation = btrfs_file_extent_generation(leaf, extent);
2602         backref->old = old;
2603         backref_insert(&new->root, backref);
2604         old->count++;
2605 out:
2606         btrfs_release_path(path);
2607         WARN_ON(ret);
2608         return ret;
2609 }
2610
2611 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2612                                    struct new_sa_defrag_extent *new)
2613 {
2614         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2615         struct old_sa_defrag_extent *old, *tmp;
2616         int ret;
2617
2618         new->path = path;
2619
2620         list_for_each_entry_safe(old, tmp, &new->head, list) {
2621                 ret = iterate_inodes_from_logical(old->bytenr +
2622                                                   old->extent_offset, fs_info,
2623                                                   path, record_one_backref,
2624                                                   old, false);
2625                 if (ret < 0 && ret != -ENOENT)
2626                         return false;
2627
2628                 /* no backref to be processed for this extent */
2629                 if (!old->count) {
2630                         list_del(&old->list);
2631                         kfree(old);
2632                 }
2633         }
2634
2635         if (list_empty(&new->head))
2636                 return false;
2637
2638         return true;
2639 }
2640
2641 static int relink_is_mergable(struct extent_buffer *leaf,
2642                               struct btrfs_file_extent_item *fi,
2643                               struct new_sa_defrag_extent *new)
2644 {
2645         if (btrfs_file_extent_disk_bytenr(leaf, fi) != new->bytenr)
2646                 return 0;
2647
2648         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2649                 return 0;
2650
2651         if (btrfs_file_extent_compression(leaf, fi) != new->compress_type)
2652                 return 0;
2653
2654         if (btrfs_file_extent_encryption(leaf, fi) ||
2655             btrfs_file_extent_other_encoding(leaf, fi))
2656                 return 0;
2657
2658         return 1;
2659 }
2660
2661 /*
2662  * Note the backref might has changed, and in this case we just return 0.
2663  */
2664 static noinline int relink_extent_backref(struct btrfs_path *path,
2665                                  struct sa_defrag_extent_backref *prev,
2666                                  struct sa_defrag_extent_backref *backref)
2667 {
2668         struct btrfs_file_extent_item *extent;
2669         struct btrfs_file_extent_item *item;
2670         struct btrfs_ordered_extent *ordered;
2671         struct btrfs_trans_handle *trans;
2672         struct btrfs_ref ref = { 0 };
2673         struct btrfs_root *root;
2674         struct btrfs_key key;
2675         struct extent_buffer *leaf;
2676         struct old_sa_defrag_extent *old = backref->old;
2677         struct new_sa_defrag_extent *new = old->new;
2678         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2679         struct inode *inode;
2680         struct extent_state *cached = NULL;
2681         int ret = 0;
2682         u64 start;
2683         u64 len;
2684         u64 lock_start;
2685         u64 lock_end;
2686         bool merge = false;
2687         int index;
2688
2689         if (prev && prev->root_id == backref->root_id &&
2690             prev->inum == backref->inum &&
2691             prev->file_pos + prev->num_bytes == backref->file_pos)
2692                 merge = true;
2693
2694         /* step 1: get root */
2695         key.objectid = backref->root_id;
2696         key.type = BTRFS_ROOT_ITEM_KEY;
2697         key.offset = (u64)-1;
2698
2699         index = srcu_read_lock(&fs_info->subvol_srcu);
2700
2701         root = btrfs_read_fs_root_no_name(fs_info, &key);
2702         if (IS_ERR(root)) {
2703                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2704                 if (PTR_ERR(root) == -ENOENT)
2705                         return 0;
2706                 return PTR_ERR(root);
2707         }
2708
2709         if (btrfs_root_readonly(root)) {
2710                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2711                 return 0;
2712         }
2713
2714         /* step 2: get inode */
2715         key.objectid = backref->inum;
2716         key.type = BTRFS_INODE_ITEM_KEY;
2717         key.offset = 0;
2718
2719         inode = btrfs_iget(fs_info->sb, &key, root);
2720         if (IS_ERR(inode)) {
2721                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2722                 return 0;
2723         }
2724
2725         srcu_read_unlock(&fs_info->subvol_srcu, index);
2726
2727         /* step 3: relink backref */
2728         lock_start = backref->file_pos;
2729         lock_end = backref->file_pos + backref->num_bytes - 1;
2730         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2731                          &cached);
2732
2733         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2734         if (ordered) {
2735                 btrfs_put_ordered_extent(ordered);
2736                 goto out_unlock;
2737         }
2738
2739         trans = btrfs_join_transaction(root);
2740         if (IS_ERR(trans)) {
2741                 ret = PTR_ERR(trans);
2742                 goto out_unlock;
2743         }
2744
2745         key.objectid = backref->inum;
2746         key.type = BTRFS_EXTENT_DATA_KEY;
2747         key.offset = backref->file_pos;
2748
2749         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2750         if (ret < 0) {
2751                 goto out_free_path;
2752         } else if (ret > 0) {
2753                 ret = 0;
2754                 goto out_free_path;
2755         }
2756
2757         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2758                                 struct btrfs_file_extent_item);
2759
2760         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2761             backref->generation)
2762                 goto out_free_path;
2763
2764         btrfs_release_path(path);
2765
2766         start = backref->file_pos;
2767         if (backref->extent_offset < old->extent_offset + old->offset)
2768                 start += old->extent_offset + old->offset -
2769                          backref->extent_offset;
2770
2771         len = min(backref->extent_offset + backref->num_bytes,
2772                   old->extent_offset + old->offset + old->len);
2773         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2774
2775         ret = btrfs_drop_extents(trans, root, inode, start,
2776                                  start + len, 1);
2777         if (ret)
2778                 goto out_free_path;
2779 again:
2780         key.objectid = btrfs_ino(BTRFS_I(inode));
2781         key.type = BTRFS_EXTENT_DATA_KEY;
2782         key.offset = start;
2783
2784         path->leave_spinning = 1;
2785         if (merge) {
2786                 struct btrfs_file_extent_item *fi;
2787                 u64 extent_len;
2788                 struct btrfs_key found_key;
2789
2790                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2791                 if (ret < 0)
2792                         goto out_free_path;
2793
2794                 path->slots[0]--;
2795                 leaf = path->nodes[0];
2796                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2797
2798                 fi = btrfs_item_ptr(leaf, path->slots[0],
2799                                     struct btrfs_file_extent_item);
2800                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2801
2802                 if (extent_len + found_key.offset == start &&
2803                     relink_is_mergable(leaf, fi, new)) {
2804                         btrfs_set_file_extent_num_bytes(leaf, fi,
2805                                                         extent_len + len);
2806                         btrfs_mark_buffer_dirty(leaf);
2807                         inode_add_bytes(inode, len);
2808
2809                         ret = 1;
2810                         goto out_free_path;
2811                 } else {
2812                         merge = false;
2813                         btrfs_release_path(path);
2814                         goto again;
2815                 }
2816         }
2817
2818         ret = btrfs_insert_empty_item(trans, root, path, &key,
2819                                         sizeof(*extent));
2820         if (ret) {
2821                 btrfs_abort_transaction(trans, ret);
2822                 goto out_free_path;
2823         }
2824
2825         leaf = path->nodes[0];
2826         item = btrfs_item_ptr(leaf, path->slots[0],
2827                                 struct btrfs_file_extent_item);
2828         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2829         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2830         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2831         btrfs_set_file_extent_num_bytes(leaf, item, len);
2832         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2833         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2834         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2835         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2836         btrfs_set_file_extent_encryption(leaf, item, 0);
2837         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2838
2839         btrfs_mark_buffer_dirty(leaf);
2840         inode_add_bytes(inode, len);
2841         btrfs_release_path(path);
2842
2843         btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new->bytenr,
2844                                new->disk_len, 0);
2845         btrfs_init_data_ref(&ref, backref->root_id, backref->inum,
2846                             new->file_pos);  /* start - extent_offset */
2847         ret = btrfs_inc_extent_ref(trans, &ref);
2848         if (ret) {
2849                 btrfs_abort_transaction(trans, ret);
2850                 goto out_free_path;
2851         }
2852
2853         ret = 1;
2854 out_free_path:
2855         btrfs_release_path(path);
2856         path->leave_spinning = 0;
2857         btrfs_end_transaction(trans);
2858 out_unlock:
2859         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2860                              &cached);
2861         iput(inode);
2862         return ret;
2863 }
2864
2865 static void free_sa_defrag_extent(struct new_sa_defrag_extent *new)
2866 {
2867         struct old_sa_defrag_extent *old, *tmp;
2868
2869         if (!new)
2870                 return;
2871
2872         list_for_each_entry_safe(old, tmp, &new->head, list) {
2873                 kfree(old);
2874         }
2875         kfree(new);
2876 }
2877
2878 static void relink_file_extents(struct new_sa_defrag_extent *new)
2879 {
2880         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2881         struct btrfs_path *path;
2882         struct sa_defrag_extent_backref *backref;
2883         struct sa_defrag_extent_backref *prev = NULL;
2884         struct rb_node *node;
2885         int ret;
2886
2887         path = btrfs_alloc_path();
2888         if (!path)
2889                 return;
2890
2891         if (!record_extent_backrefs(path, new)) {
2892                 btrfs_free_path(path);
2893                 goto out;
2894         }
2895         btrfs_release_path(path);
2896
2897         while (1) {
2898                 node = rb_first(&new->root);
2899                 if (!node)
2900                         break;
2901                 rb_erase(node, &new->root);
2902
2903                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2904
2905                 ret = relink_extent_backref(path, prev, backref);
2906                 WARN_ON(ret < 0);
2907
2908                 kfree(prev);
2909
2910                 if (ret == 1)
2911                         prev = backref;
2912                 else
2913                         prev = NULL;
2914                 cond_resched();
2915         }
2916         kfree(prev);
2917
2918         btrfs_free_path(path);
2919 out:
2920         free_sa_defrag_extent(new);
2921
2922         atomic_dec(&fs_info->defrag_running);
2923         wake_up(&fs_info->transaction_wait);
2924 }
2925
2926 static struct new_sa_defrag_extent *
2927 record_old_file_extents(struct inode *inode,
2928                         struct btrfs_ordered_extent *ordered)
2929 {
2930         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2931         struct btrfs_root *root = BTRFS_I(inode)->root;
2932         struct btrfs_path *path;
2933         struct btrfs_key key;
2934         struct old_sa_defrag_extent *old;
2935         struct new_sa_defrag_extent *new;
2936         int ret;
2937
2938         new = kmalloc(sizeof(*new), GFP_NOFS);
2939         if (!new)
2940                 return NULL;
2941
2942         new->inode = inode;
2943         new->file_pos = ordered->file_offset;
2944         new->len = ordered->len;
2945         new->bytenr = ordered->start;
2946         new->disk_len = ordered->disk_len;
2947         new->compress_type = ordered->compress_type;
2948         new->root = RB_ROOT;
2949         INIT_LIST_HEAD(&new->head);
2950
2951         path = btrfs_alloc_path();
2952         if (!path)
2953                 goto out_kfree;
2954
2955         key.objectid = btrfs_ino(BTRFS_I(inode));
2956         key.type = BTRFS_EXTENT_DATA_KEY;
2957         key.offset = new->file_pos;
2958
2959         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2960         if (ret < 0)
2961                 goto out_free_path;
2962         if (ret > 0 && path->slots[0] > 0)
2963                 path->slots[0]--;
2964
2965         /* find out all the old extents for the file range */
2966         while (1) {
2967                 struct btrfs_file_extent_item *extent;
2968                 struct extent_buffer *l;
2969                 int slot;
2970                 u64 num_bytes;
2971                 u64 offset;
2972                 u64 end;
2973                 u64 disk_bytenr;
2974                 u64 extent_offset;
2975
2976                 l = path->nodes[0];
2977                 slot = path->slots[0];
2978
2979                 if (slot >= btrfs_header_nritems(l)) {
2980                         ret = btrfs_next_leaf(root, path);
2981                         if (ret < 0)
2982                                 goto out_free_path;
2983                         else if (ret > 0)
2984                                 break;
2985                         continue;
2986                 }
2987
2988                 btrfs_item_key_to_cpu(l, &key, slot);
2989
2990                 if (key.objectid != btrfs_ino(BTRFS_I(inode)))
2991                         break;
2992                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2993                         break;
2994                 if (key.offset >= new->file_pos + new->len)
2995                         break;
2996
2997                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2998
2999                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
3000                 if (key.offset + num_bytes < new->file_pos)
3001                         goto next;
3002
3003                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
3004                 if (!disk_bytenr)
3005                         goto next;
3006
3007                 extent_offset = btrfs_file_extent_offset(l, extent);
3008
3009                 old = kmalloc(sizeof(*old), GFP_NOFS);
3010                 if (!old)
3011                         goto out_free_path;
3012
3013                 offset = max(new->file_pos, key.offset);
3014                 end = min(new->file_pos + new->len, key.offset + num_bytes);
3015
3016                 old->bytenr = disk_bytenr;
3017                 old->extent_offset = extent_offset;
3018                 old->offset = offset - key.offset;
3019                 old->len = end - offset;
3020                 old->new = new;
3021                 old->count = 0;
3022                 list_add_tail(&old->list, &new->head);
3023 next:
3024                 path->slots[0]++;
3025                 cond_resched();
3026         }
3027
3028         btrfs_free_path(path);
3029         atomic_inc(&fs_info->defrag_running);
3030
3031         return new;
3032
3033 out_free_path:
3034         btrfs_free_path(path);
3035 out_kfree:
3036         free_sa_defrag_extent(new);
3037         return NULL;
3038 }
3039
3040 static void btrfs_release_delalloc_bytes(struct btrfs_fs_info *fs_info,
3041                                          u64 start, u64 len)
3042 {
3043         struct btrfs_block_group *cache;
3044
3045         cache = btrfs_lookup_block_group(fs_info, start);
3046         ASSERT(cache);
3047
3048         spin_lock(&cache->lock);
3049         cache->delalloc_bytes -= len;
3050         spin_unlock(&cache->lock);
3051
3052         btrfs_put_block_group(cache);
3053 }
3054
3055 /* as ordered data IO finishes, this gets called so we can finish
3056  * an ordered extent if the range of bytes in the file it covers are
3057  * fully written.
3058  */
3059 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
3060 {
3061         struct inode *inode = ordered_extent->inode;
3062         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3063         struct btrfs_root *root = BTRFS_I(inode)->root;
3064         struct btrfs_trans_handle *trans = NULL;
3065         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3066         struct extent_state *cached_state = NULL;
3067         struct new_sa_defrag_extent *new = NULL;
3068         int compress_type = 0;
3069         int ret = 0;
3070         u64 logical_len = ordered_extent->len;
3071         bool freespace_inode;
3072         bool truncated = false;
3073         bool range_locked = false;
3074         bool clear_new_delalloc_bytes = false;
3075         bool clear_reserved_extent = true;
3076
3077         if (!test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
3078             !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags) &&
3079             !test_bit(BTRFS_ORDERED_DIRECT, &ordered_extent->flags))
3080                 clear_new_delalloc_bytes = true;
3081
3082         freespace_inode = btrfs_is_free_space_inode(BTRFS_I(inode));
3083
3084         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
3085                 ret = -EIO;
3086                 goto out;
3087         }
3088
3089         btrfs_free_io_failure_record(BTRFS_I(inode),
3090                         ordered_extent->file_offset,
3091                         ordered_extent->file_offset +
3092                         ordered_extent->len - 1);
3093
3094         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
3095                 truncated = true;
3096                 logical_len = ordered_extent->truncated_len;
3097                 /* Truncated the entire extent, don't bother adding */
3098                 if (!logical_len)
3099                         goto out;
3100         }
3101
3102         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
3103                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
3104
3105                 /*
3106                  * For mwrite(mmap + memset to write) case, we still reserve
3107                  * space for NOCOW range.
3108                  * As NOCOW won't cause a new delayed ref, just free the space
3109                  */
3110                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
3111                                        ordered_extent->len);
3112                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
3113                 if (freespace_inode)
3114                         trans = btrfs_join_transaction_spacecache(root);
3115                 else
3116                         trans = btrfs_join_transaction(root);
3117                 if (IS_ERR(trans)) {
3118                         ret = PTR_ERR(trans);
3119                         trans = NULL;
3120                         goto out;
3121                 }
3122                 trans->block_rsv = &BTRFS_I(inode)->block_rsv;
3123                 ret = btrfs_update_inode_fallback(trans, root, inode);
3124                 if (ret) /* -ENOMEM or corruption */
3125                         btrfs_abort_transaction(trans, ret);
3126                 goto out;
3127         }
3128
3129         range_locked = true;
3130         lock_extent_bits(io_tree, ordered_extent->file_offset,
3131                          ordered_extent->file_offset + ordered_extent->len - 1,
3132                          &cached_state);
3133
3134         ret = test_range_bit(io_tree, ordered_extent->file_offset,
3135                         ordered_extent->file_offset + ordered_extent->len - 1,
3136                         EXTENT_DEFRAG, 0, cached_state);
3137         if (ret) {
3138                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
3139                 if (0 && last_snapshot >= BTRFS_I(inode)->generation)
3140                         /* the inode is shared */
3141                         new = record_old_file_extents(inode, ordered_extent);
3142
3143                 clear_extent_bit(io_tree, ordered_extent->file_offset,
3144                         ordered_extent->file_offset + ordered_extent->len - 1,
3145                         EXTENT_DEFRAG, 0, 0, &cached_state);
3146         }
3147
3148         if (freespace_inode)
3149                 trans = btrfs_join_transaction_spacecache(root);
3150         else
3151                 trans = btrfs_join_transaction(root);
3152         if (IS_ERR(trans)) {
3153                 ret = PTR_ERR(trans);
3154                 trans = NULL;
3155                 goto out;
3156         }
3157
3158         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
3159
3160         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
3161                 compress_type = ordered_extent->compress_type;
3162         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
3163                 BUG_ON(compress_type);
3164                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
3165                                        ordered_extent->len);
3166                 ret = btrfs_mark_extent_written(trans, BTRFS_I(inode),
3167                                                 ordered_extent->file_offset,
3168                                                 ordered_extent->file_offset +
3169                                                 logical_len);
3170         } else {
3171                 BUG_ON(root == fs_info->tree_root);
3172                 ret = insert_reserved_file_extent(trans, inode,
3173                                                 ordered_extent->file_offset,
3174                                                 ordered_extent->start,
3175                                                 ordered_extent->disk_len,
3176                                                 logical_len, logical_len,
3177                                                 compress_type, 0, 0,
3178                                                 BTRFS_FILE_EXTENT_REG);
3179                 if (!ret) {
3180                         clear_reserved_extent = false;
3181                         btrfs_release_delalloc_bytes(fs_info,
3182                                                      ordered_extent->start,
3183                                                      ordered_extent->disk_len);
3184                 }
3185         }
3186         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
3187                            ordered_extent->file_offset, ordered_extent->len,
3188                            trans->transid);
3189         if (ret < 0) {
3190                 btrfs_abort_transaction(trans, ret);
3191                 goto out;
3192         }
3193
3194         ret = add_pending_csums(trans, inode, &ordered_extent->list);
3195         if (ret) {
3196                 btrfs_abort_transaction(trans, ret);
3197                 goto out;
3198         }
3199
3200         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
3201         ret = btrfs_update_inode_fallback(trans, root, inode);
3202         if (ret) { /* -ENOMEM or corruption */
3203                 btrfs_abort_transaction(trans, ret);
3204                 goto out;
3205         }
3206         ret = 0;
3207 out:
3208         if (range_locked || clear_new_delalloc_bytes) {
3209                 unsigned int clear_bits = 0;
3210
3211                 if (range_locked)
3212                         clear_bits |= EXTENT_LOCKED;
3213                 if (clear_new_delalloc_bytes)
3214                         clear_bits |= EXTENT_DELALLOC_NEW;
3215                 clear_extent_bit(&BTRFS_I(inode)->io_tree,
3216                                  ordered_extent->file_offset,
3217                                  ordered_extent->file_offset +
3218                                  ordered_extent->len - 1,
3219                                  clear_bits,
3220                                  (clear_bits & EXTENT_LOCKED) ? 1 : 0,
3221                                  0, &cached_state);
3222         }
3223
3224         if (trans)
3225                 btrfs_end_transaction(trans);
3226
3227         if (ret || truncated) {
3228                 u64 start, end;
3229
3230                 if (truncated)
3231                         start = ordered_extent->file_offset + logical_len;
3232                 else
3233                         start = ordered_extent->file_offset;
3234                 end = ordered_extent->file_offset + ordered_extent->len - 1;
3235                 clear_extent_uptodate(io_tree, start, end, NULL);
3236
3237                 /* Drop the cache for the part of the extent we didn't write. */
3238                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0);
3239
3240                 /*
3241                  * If the ordered extent had an IOERR or something else went
3242                  * wrong we need to return the space for this ordered extent
3243                  * back to the allocator.  We only free the extent in the
3244                  * truncated case if we didn't write out the extent at all.
3245                  *
3246                  * If we made it past insert_reserved_file_extent before we
3247                  * errored out then we don't need to do this as the accounting
3248                  * has already been done.
3249                  */
3250                 if ((ret || !logical_len) &&
3251                     clear_reserved_extent &&
3252                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
3253                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags))
3254                         btrfs_free_reserved_extent(fs_info,
3255                                                    ordered_extent->start,
3256                                                    ordered_extent->disk_len, 1);
3257         }
3258
3259
3260         /*
3261          * This needs to be done to make sure anybody waiting knows we are done
3262          * updating everything for this ordered extent.
3263          */
3264         btrfs_remove_ordered_extent(inode, ordered_extent);
3265
3266         /* for snapshot-aware defrag */
3267         if (new) {
3268                 if (ret) {
3269                         free_sa_defrag_extent(new);
3270                         atomic_dec(&fs_info->defrag_running);
3271                 } else {
3272                         relink_file_extents(new);
3273                 }
3274         }
3275
3276         /* once for us */
3277         btrfs_put_ordered_extent(ordered_extent);
3278         /* once for the tree */
3279         btrfs_put_ordered_extent(ordered_extent);
3280
3281         return ret;
3282 }
3283
3284 static void finish_ordered_fn(struct btrfs_work *work)
3285 {
3286         struct btrfs_ordered_extent *ordered_extent;
3287         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
3288         btrfs_finish_ordered_io(ordered_extent);
3289 }
3290
3291 void btrfs_writepage_endio_finish_ordered(struct page *page, u64 start,
3292                                           u64 end, int uptodate)
3293 {
3294         struct inode *inode = page->mapping->host;
3295         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3296         struct btrfs_ordered_extent *ordered_extent = NULL;
3297         struct btrfs_workqueue *wq;
3298
3299         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
3300
3301         ClearPagePrivate2(page);
3302         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
3303                                             end - start + 1, uptodate))
3304                 return;
3305
3306         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
3307                 wq = fs_info->endio_freespace_worker;
3308         else
3309                 wq = fs_info->endio_write_workers;
3310
3311         btrfs_init_work(&ordered_extent->work, finish_ordered_fn, NULL, NULL);
3312         btrfs_queue_work(wq, &ordered_extent->work);
3313 }
3314
3315 static int __readpage_endio_check(struct inode *inode,
3316                                   struct btrfs_io_bio *io_bio,
3317                                   int icsum, struct page *page,
3318                                   int pgoff, u64 start, size_t len)
3319 {
3320         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3321         SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
3322         char *kaddr;
3323         u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
3324         u8 *csum_expected;
3325         u8 csum[BTRFS_CSUM_SIZE];
3326
3327         csum_expected = ((u8 *)io_bio->csum) + icsum * csum_size;
3328
3329         kaddr = kmap_atomic(page);
3330         shash->tfm = fs_info->csum_shash;
3331
3332         crypto_shash_init(shash);
3333         crypto_shash_update(shash, kaddr + pgoff, len);
3334         crypto_shash_final(shash, csum);
3335
3336         if (memcmp(csum, csum_expected, csum_size))
3337                 goto zeroit;
3338
3339         kunmap_atomic(kaddr);
3340         return 0;
3341 zeroit:
3342         btrfs_print_data_csum_error(BTRFS_I(inode), start, csum, csum_expected,
3343                                     io_bio->mirror_num);
3344         memset(kaddr + pgoff, 1, len);
3345         flush_dcache_page(page);
3346         kunmap_atomic(kaddr);
3347         return -EIO;
3348 }
3349
3350 /*
3351  * when reads are done, we need to check csums to verify the data is correct
3352  * if there's a match, we allow the bio to finish.  If not, the code in
3353  * extent_io.c will try to find good copies for us.
3354  */
3355 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
3356                                       u64 phy_offset, struct page *page,
3357                                       u64 start, u64 end, int mirror)
3358 {
3359         size_t offset = start - page_offset(page);
3360         struct inode *inode = page->mapping->host;
3361         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3362         struct btrfs_root *root = BTRFS_I(inode)->root;
3363
3364         if (PageChecked(page)) {
3365                 ClearPageChecked(page);
3366                 return 0;
3367         }
3368
3369         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
3370                 return 0;
3371
3372         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
3373             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
3374                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM);
3375                 return 0;
3376         }
3377
3378         phy_offset >>= inode->i_sb->s_blocksize_bits;
3379         return __readpage_endio_check(inode, io_bio, phy_offset, page, offset,
3380                                       start, (size_t)(end - start + 1));
3381 }
3382
3383 /*
3384  * btrfs_add_delayed_iput - perform a delayed iput on @inode
3385  *
3386  * @inode: The inode we want to perform iput on
3387  *
3388  * This function uses the generic vfs_inode::i_count to track whether we should
3389  * just decrement it (in case it's > 1) or if this is the last iput then link
3390  * the inode to the delayed iput machinery. Delayed iputs are processed at
3391  * transaction commit time/superblock commit/cleaner kthread.
3392  */
3393 void btrfs_add_delayed_iput(struct inode *inode)
3394 {
3395         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3396         struct btrfs_inode *binode = BTRFS_I(inode);
3397
3398         if (atomic_add_unless(&inode->i_count, -1, 1))
3399                 return;
3400
3401         atomic_inc(&fs_info->nr_delayed_iputs);
3402         spin_lock(&fs_info->delayed_iput_lock);
3403         ASSERT(list_empty(&binode->delayed_iput));
3404         list_add_tail(&binode->delayed_iput, &fs_info->delayed_iputs);
3405         spin_unlock(&fs_info->delayed_iput_lock);
3406         if (!test_bit(BTRFS_FS_CLEANER_RUNNING, &fs_info->flags))
3407                 wake_up_process(fs_info->cleaner_kthread);
3408 }
3409
3410 static void run_delayed_iput_locked(struct btrfs_fs_info *fs_info,
3411                                     struct btrfs_inode *inode)
3412 {
3413         list_del_init(&inode->delayed_iput);
3414         spin_unlock(&fs_info->delayed_iput_lock);
3415         iput(&inode->vfs_inode);
3416         if (atomic_dec_and_test(&fs_info->nr_delayed_iputs))
3417                 wake_up(&fs_info->delayed_iputs_wait);
3418         spin_lock(&fs_info->delayed_iput_lock);
3419 }
3420
3421 static void btrfs_run_delayed_iput(struct btrfs_fs_info *fs_info,
3422                                    struct btrfs_inode *inode)
3423 {
3424         if (!list_empty(&inode->delayed_iput)) {
3425                 spin_lock(&fs_info->delayed_iput_lock);
3426                 if (!list_empty(&inode->delayed_iput))
3427                         run_delayed_iput_locked(fs_info, inode);
3428                 spin_unlock(&fs_info->delayed_iput_lock);
3429         }
3430 }
3431
3432 void btrfs_run_delayed_iputs(struct btrfs_fs_info *fs_info)
3433 {
3434
3435         spin_lock(&fs_info->delayed_iput_lock);
3436         while (!list_empty(&fs_info->delayed_iputs)) {
3437                 struct btrfs_inode *inode;
3438
3439                 inode = list_first_entry(&fs_info->delayed_iputs,
3440                                 struct btrfs_inode, delayed_iput);
3441                 run_delayed_iput_locked(fs_info, inode);
3442         }
3443         spin_unlock(&fs_info->delayed_iput_lock);
3444 }
3445
3446 /**
3447  * btrfs_wait_on_delayed_iputs - wait on the delayed iputs to be done running
3448  * @fs_info - the fs_info for this fs
3449  * @return - EINTR if we were killed, 0 if nothing's pending
3450  *
3451  * This will wait on any delayed iputs that are currently running with KILLABLE
3452  * set.  Once they are all done running we will return, unless we are killed in
3453  * which case we return EINTR. This helps in user operations like fallocate etc
3454  * that might get blocked on the iputs.
3455  */
3456 int btrfs_wait_on_delayed_iputs(struct btrfs_fs_info *fs_info)
3457 {
3458         int ret = wait_event_killable(fs_info->delayed_iputs_wait,
3459                         atomic_read(&fs_info->nr_delayed_iputs) == 0);
3460         if (ret)
3461                 return -EINTR;
3462         return 0;
3463 }
3464
3465 /*
3466  * This creates an orphan entry for the given inode in case something goes wrong
3467  * in the middle of an unlink.
3468  */
3469 int btrfs_orphan_add(struct btrfs_trans_handle *trans,
3470                      struct btrfs_inode *inode)
3471 {
3472         int ret;
3473
3474         ret = btrfs_insert_orphan_item(trans, inode->root, btrfs_ino(inode));
3475         if (ret && ret != -EEXIST) {
3476                 btrfs_abort_transaction(trans, ret);
3477                 return ret;
3478         }
3479
3480         return 0;
3481 }
3482
3483 /*
3484  * We have done the delete so we can go ahead and remove the orphan item for
3485  * this particular inode.
3486  */
3487 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3488                             struct btrfs_inode *inode)
3489 {
3490         return btrfs_del_orphan_item(trans, inode->root, btrfs_ino(inode));
3491 }
3492
3493 /*
3494  * this cleans up any orphans that may be left on the list from the last use
3495  * of this root.
3496  */
3497 int btrfs_orphan_cleanup(struct btrfs_root *root)
3498 {
3499         struct btrfs_fs_info *fs_info = root->fs_info;
3500         struct btrfs_path *path;
3501         struct extent_buffer *leaf;
3502         struct btrfs_key key, found_key;
3503         struct btrfs_trans_handle *trans;
3504         struct inode *inode;
3505         u64 last_objectid = 0;
3506         int ret = 0, nr_unlink = 0;
3507
3508         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3509                 return 0;
3510
3511         path = btrfs_alloc_path();
3512         if (!path) {
3513                 ret = -ENOMEM;
3514                 goto out;
3515         }
3516         path->reada = READA_BACK;
3517
3518         key.objectid = BTRFS_ORPHAN_OBJECTID;
3519         key.type = BTRFS_ORPHAN_ITEM_KEY;
3520         key.offset = (u64)-1;
3521
3522         while (1) {
3523                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3524                 if (ret < 0)
3525                         goto out;
3526
3527                 /*
3528                  * if ret == 0 means we found what we were searching for, which
3529                  * is weird, but possible, so only screw with path if we didn't
3530                  * find the key and see if we have stuff that matches
3531                  */
3532                 if (ret > 0) {
3533                         ret = 0;
3534                         if (path->slots[0] == 0)
3535                                 break;
3536                         path->slots[0]--;
3537                 }
3538
3539                 /* pull out the item */
3540                 leaf = path->nodes[0];
3541                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3542
3543                 /* make sure the item matches what we want */
3544                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3545                         break;
3546                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3547                         break;
3548
3549                 /* release the path since we're done with it */
3550                 btrfs_release_path(path);
3551
3552                 /*
3553                  * this is where we are basically btrfs_lookup, without the
3554                  * crossing root thing.  we store the inode number in the
3555                  * offset of the orphan item.
3556                  */
3557
3558                 if (found_key.offset == last_objectid) {
3559                         btrfs_err(fs_info,
3560                                   "Error removing orphan entry, stopping orphan cleanup");
3561                         ret = -EINVAL;
3562                         goto out;
3563                 }
3564
3565                 last_objectid = found_key.offset;
3566
3567                 found_key.objectid = found_key.offset;
3568                 found_key.type = BTRFS_INODE_ITEM_KEY;
3569                 found_key.offset = 0;
3570                 inode = btrfs_iget(fs_info->sb, &found_key, root);
3571                 ret = PTR_ERR_OR_ZERO(inode);
3572                 if (ret && ret != -ENOENT)
3573                         goto out;
3574
3575                 if (ret == -ENOENT && root == fs_info->tree_root) {
3576                         struct btrfs_root *dead_root;
3577                         struct btrfs_fs_info *fs_info = root->fs_info;
3578                         int is_dead_root = 0;
3579
3580                         /*
3581                          * this is an orphan in the tree root. Currently these
3582                          * could come from 2 sources:
3583                          *  a) a snapshot deletion in progress
3584                          *  b) a free space cache inode
3585                          * We need to distinguish those two, as the snapshot
3586                          * orphan must not get deleted.
3587                          * find_dead_roots already ran before us, so if this
3588                          * is a snapshot deletion, we should find the root
3589                          * in the dead_roots list
3590                          */
3591                         spin_lock(&fs_info->trans_lock);
3592                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3593                                             root_list) {
3594                                 if (dead_root->root_key.objectid ==
3595                                     found_key.objectid) {
3596                                         is_dead_root = 1;
3597                                         break;
3598                                 }
3599                         }
3600                         spin_unlock(&fs_info->trans_lock);
3601                         if (is_dead_root) {
3602                                 /* prevent this orphan from being found again */
3603                                 key.offset = found_key.objectid - 1;
3604                                 continue;
3605                         }
3606
3607                 }
3608
3609                 /*
3610                  * If we have an inode with links, there are a couple of
3611                  * possibilities. Old kernels (before v3.12) used to create an
3612                  * orphan item for truncate indicating that there were possibly
3613                  * extent items past i_size that needed to be deleted. In v3.12,
3614                  * truncate was changed to update i_size in sync with the extent
3615                  * items, but the (useless) orphan item was still created. Since
3616                  * v4.18, we don't create the orphan item for truncate at all.
3617                  *
3618                  * So, this item could mean that we need to do a truncate, but
3619                  * only if this filesystem was last used on a pre-v3.12 kernel
3620                  * and was not cleanly unmounted. The odds of that are quite
3621                  * slim, and it's a pain to do the truncate now, so just delete
3622                  * the orphan item.
3623                  *
3624                  * It's also possible that this orphan item was supposed to be
3625                  * deleted but wasn't. The inode number may have been reused,
3626                  * but either way, we can delete the orphan item.
3627                  */
3628                 if (ret == -ENOENT || inode->i_nlink) {
3629                         if (!ret)
3630                                 iput(inode);
3631                         trans = btrfs_start_transaction(root, 1);
3632                         if (IS_ERR(trans)) {
3633                                 ret = PTR_ERR(trans);
3634                                 goto out;
3635                         }
3636                         btrfs_debug(fs_info, "auto deleting %Lu",
3637                                     found_key.objectid);
3638                         ret = btrfs_del_orphan_item(trans, root,
3639                                                     found_key.objectid);
3640                         btrfs_end_transaction(trans);
3641                         if (ret)
3642                                 goto out;
3643                         continue;
3644                 }
3645
3646                 nr_unlink++;
3647
3648                 /* this will do delete_inode and everything for us */
3649                 iput(inode);
3650         }
3651         /* release the path since we're done with it */
3652         btrfs_release_path(path);
3653
3654         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3655
3656         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3657                 trans = btrfs_join_transaction(root);
3658                 if (!IS_ERR(trans))
3659                         btrfs_end_transaction(trans);
3660         }
3661
3662         if (nr_unlink)
3663                 btrfs_debug(fs_info, "unlinked %d orphans", nr_unlink);
3664
3665 out:
3666         if (ret)
3667                 btrfs_err(fs_info, "could not do orphan cleanup %d", ret);
3668         btrfs_free_path(path);
3669         return ret;
3670 }
3671
3672 /*
3673  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3674  * don't find any xattrs, we know there can't be any acls.
3675  *
3676  * slot is the slot the inode is in, objectid is the objectid of the inode
3677  */
3678 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3679                                           int slot, u64 objectid,
3680                                           int *first_xattr_slot)
3681 {
3682         u32 nritems = btrfs_header_nritems(leaf);
3683         struct btrfs_key found_key;
3684         static u64 xattr_access = 0;
3685         static u64 xattr_default = 0;
3686         int scanned = 0;
3687
3688         if (!xattr_access) {
3689                 xattr_access = btrfs_name_hash(XATTR_NAME_POSIX_ACL_ACCESS,
3690                                         strlen(XATTR_NAME_POSIX_ACL_ACCESS));
3691                 xattr_default = btrfs_name_hash(XATTR_NAME_POSIX_ACL_DEFAULT,
3692                                         strlen(XATTR_NAME_POSIX_ACL_DEFAULT));
3693         }
3694
3695         slot++;
3696         *first_xattr_slot = -1;
3697         while (slot < nritems) {
3698                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3699
3700                 /* we found a different objectid, there must not be acls */
3701                 if (found_key.objectid != objectid)
3702                         return 0;
3703
3704                 /* we found an xattr, assume we've got an acl */
3705                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3706                         if (*first_xattr_slot == -1)
3707                                 *first_xattr_slot = slot;
3708                         if (found_key.offset == xattr_access ||
3709                             found_key.offset == xattr_default)
3710                                 return 1;
3711                 }
3712
3713                 /*
3714                  * we found a key greater than an xattr key, there can't
3715                  * be any acls later on
3716                  */
3717                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3718                         return 0;
3719
3720                 slot++;
3721                 scanned++;
3722
3723                 /*
3724                  * it goes inode, inode backrefs, xattrs, extents,
3725                  * so if there are a ton of hard links to an inode there can
3726                  * be a lot of backrefs.  Don't waste time searching too hard,
3727                  * this is just an optimization
3728                  */
3729                 if (scanned >= 8)
3730                         break;
3731         }
3732         /* we hit the end of the leaf before we found an xattr or
3733          * something larger than an xattr.  We have to assume the inode
3734          * has acls
3735          */
3736         if (*first_xattr_slot == -1)
3737                 *first_xattr_slot = slot;
3738         return 1;
3739 }
3740
3741 /*
3742  * read an inode from the btree into the in-memory inode
3743  */
3744 static int btrfs_read_locked_inode(struct inode *inode,
3745                                    struct btrfs_path *in_path)
3746 {
3747         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3748         struct btrfs_path *path = in_path;
3749         struct extent_buffer *leaf;
3750         struct btrfs_inode_item *inode_item;
3751         struct btrfs_root *root = BTRFS_I(inode)->root;
3752         struct btrfs_key location;
3753         unsigned long ptr;
3754         int maybe_acls;
3755         u32 rdev;
3756         int ret;
3757         bool filled = false;
3758         int first_xattr_slot;
3759
3760         ret = btrfs_fill_inode(inode, &rdev);
3761         if (!ret)
3762                 filled = true;
3763
3764         if (!path) {
3765                 path = btrfs_alloc_path();
3766                 if (!path)
3767                         return -ENOMEM;
3768         }
3769
3770         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3771
3772         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3773         if (ret) {
3774                 if (path != in_path)
3775                         btrfs_free_path(path);
3776                 return ret;
3777         }
3778
3779         leaf = path->nodes[0];
3780
3781         if (filled)
3782                 goto cache_index;
3783
3784         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3785                                     struct btrfs_inode_item);
3786         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3787         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3788         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3789         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3790         btrfs_i_size_write(BTRFS_I(inode), btrfs_inode_size(leaf, inode_item));
3791
3792         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3793         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3794
3795         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3796         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3797
3798         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3799         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3800
3801         BTRFS_I(inode)->i_otime.tv_sec =
3802                 btrfs_timespec_sec(leaf, &inode_item->otime);
3803         BTRFS_I(inode)->i_otime.tv_nsec =
3804                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3805
3806         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3807         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3808         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3809
3810         inode_set_iversion_queried(inode,
3811                                    btrfs_inode_sequence(leaf, inode_item));
3812         inode->i_generation = BTRFS_I(inode)->generation;
3813         inode->i_rdev = 0;
3814         rdev = btrfs_inode_rdev(leaf, inode_item);
3815
3816         BTRFS_I(inode)->index_cnt = (u64)-1;
3817         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3818
3819 cache_index:
3820         /*
3821          * If we were modified in the current generation and evicted from memory
3822          * and then re-read we need to do a full sync since we don't have any
3823          * idea about which extents were modified before we were evicted from
3824          * cache.
3825          *
3826          * This is required for both inode re-read from disk and delayed inode
3827          * in delayed_nodes_tree.
3828          */
3829         if (BTRFS_I(inode)->last_trans == fs_info->generation)
3830                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3831                         &BTRFS_I(inode)->runtime_flags);
3832
3833         /*
3834          * We don't persist the id of the transaction where an unlink operation
3835          * against the inode was last made. So here we assume the inode might
3836          * have been evicted, and therefore the exact value of last_unlink_trans
3837          * lost, and set it to last_trans to avoid metadata inconsistencies
3838          * between the inode and its parent if the inode is fsync'ed and the log
3839          * replayed. For example, in the scenario:
3840          *
3841          * touch mydir/foo
3842          * ln mydir/foo mydir/bar
3843          * sync
3844          * unlink mydir/bar
3845          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3846          * xfs_io -c fsync mydir/foo
3847          * <power failure>
3848          * mount fs, triggers fsync log replay
3849          *
3850          * We must make sure that when we fsync our inode foo we also log its
3851          * parent inode, otherwise after log replay the parent still has the
3852          * dentry with the "bar" name but our inode foo has a link count of 1
3853          * and doesn't have an inode ref with the name "bar" anymore.
3854          *
3855          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3856          * but it guarantees correctness at the expense of occasional full
3857          * transaction commits on fsync if our inode is a directory, or if our
3858          * inode is not a directory, logging its parent unnecessarily.
3859          */
3860         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3861
3862         path->slots[0]++;
3863         if (inode->i_nlink != 1 ||
3864             path->slots[0] >= btrfs_header_nritems(leaf))
3865                 goto cache_acl;
3866
3867         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3868         if (location.objectid != btrfs_ino(BTRFS_I(inode)))
3869                 goto cache_acl;
3870
3871         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3872         if (location.type == BTRFS_INODE_REF_KEY) {
3873                 struct btrfs_inode_ref *ref;
3874
3875                 ref = (struct btrfs_inode_ref *)ptr;
3876                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3877         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3878                 struct btrfs_inode_extref *extref;
3879
3880                 extref = (struct btrfs_inode_extref *)ptr;
3881                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3882                                                                      extref);
3883         }
3884 cache_acl:
3885         /*
3886          * try to precache a NULL acl entry for files that don't have
3887          * any xattrs or acls
3888          */
3889         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3890                         btrfs_ino(BTRFS_I(inode)), &first_xattr_slot);
3891         if (first_xattr_slot != -1) {
3892                 path->slots[0] = first_xattr_slot;
3893                 ret = btrfs_load_inode_props(inode, path);
3894                 if (ret)
3895                         btrfs_err(fs_info,
3896                                   "error loading props for ino %llu (root %llu): %d",
3897                                   btrfs_ino(BTRFS_I(inode)),
3898                                   root->root_key.objectid, ret);
3899         }
3900         if (path != in_path)
3901                 btrfs_free_path(path);
3902
3903         if (!maybe_acls)
3904                 cache_no_acl(inode);
3905
3906         switch (inode->i_mode & S_IFMT) {
3907         case S_IFREG:
3908                 inode->i_mapping->a_ops = &btrfs_aops;
3909                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3910                 inode->i_fop = &btrfs_file_operations;
3911                 inode->i_op = &btrfs_file_inode_operations;
3912                 break;
3913         case S_IFDIR:
3914                 inode->i_fop = &btrfs_dir_file_operations;
3915                 inode->i_op = &btrfs_dir_inode_operations;
3916                 break;
3917         case S_IFLNK:
3918                 inode->i_op = &btrfs_symlink_inode_operations;
3919                 inode_nohighmem(inode);
3920                 inode->i_mapping->a_ops = &btrfs_aops;
3921                 break;
3922         default:
3923                 inode->i_op = &btrfs_special_inode_operations;
3924                 init_special_inode(inode, inode->i_mode, rdev);
3925                 break;
3926         }
3927
3928         btrfs_sync_inode_flags_to_i_flags(inode);
3929         return 0;
3930 }
3931
3932 /*
3933  * given a leaf and an inode, copy the inode fields into the leaf
3934  */
3935 static void fill_inode_item(struct btrfs_trans_handle *trans,
3936                             struct extent_buffer *leaf,
3937                             struct btrfs_inode_item *item,
3938                             struct inode *inode)
3939 {
3940         struct btrfs_map_token token;
3941
3942         btrfs_init_map_token(&token, leaf);
3943
3944         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3945         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3946         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3947                                    &token);
3948         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3949         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3950
3951         btrfs_set_token_timespec_sec(leaf, &item->atime,
3952                                      inode->i_atime.tv_sec, &token);
3953         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3954                                       inode->i_atime.tv_nsec, &token);
3955
3956         btrfs_set_token_timespec_sec(leaf, &item->mtime,
3957                                      inode->i_mtime.tv_sec, &token);
3958         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
3959                                       inode->i_mtime.tv_nsec, &token);
3960
3961         btrfs_set_token_timespec_sec(leaf, &item->ctime,
3962                                      inode->i_ctime.tv_sec, &token);
3963         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
3964                                       inode->i_ctime.tv_nsec, &token);
3965
3966         btrfs_set_token_timespec_sec(leaf, &item->otime,
3967                                      BTRFS_I(inode)->i_otime.tv_sec, &token);
3968         btrfs_set_token_timespec_nsec(leaf, &item->otime,
3969                                       BTRFS_I(inode)->i_otime.tv_nsec, &token);
3970
3971         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3972                                      &token);
3973         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
3974                                          &token);
3975         btrfs_set_token_inode_sequence(leaf, item, inode_peek_iversion(inode),
3976                                        &token);
3977         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3978         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3979         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3980         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3981 }
3982
3983 /*
3984  * copy everything in the in-memory inode into the btree.
3985  */
3986 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
3987                                 struct btrfs_root *root, struct inode *inode)
3988 {
3989         struct btrfs_inode_item *inode_item;
3990         struct btrfs_path *path;
3991         struct extent_buffer *leaf;
3992         int ret;
3993
3994         path = btrfs_alloc_path();
3995         if (!path)
3996                 return -ENOMEM;
3997
3998         path->leave_spinning = 1;
3999         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
4000                                  1);
4001         if (ret) {
4002                 if (ret > 0)
4003                         ret = -ENOENT;
4004                 goto failed;
4005         }
4006
4007         leaf = path->nodes[0];
4008         inode_item = btrfs_item_ptr(leaf, path->slots[0],
4009                                     struct btrfs_inode_item);
4010
4011         fill_inode_item(trans, leaf, inode_item, inode);
4012         btrfs_mark_buffer_dirty(leaf);
4013         btrfs_set_inode_last_trans(trans, inode);
4014         ret = 0;
4015 failed:
4016         btrfs_free_path(path);
4017         return ret;
4018 }
4019
4020 /*
4021  * copy everything in the in-memory inode into the btree.
4022  */
4023 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
4024                                 struct btrfs_root *root, struct inode *inode)
4025 {
4026         struct btrfs_fs_info *fs_info = root->fs_info;
4027         int ret;
4028
4029         /*
4030          * If the inode is a free space inode, we can deadlock during commit
4031          * if we put it into the delayed code.
4032          *
4033          * The data relocation inode should also be directly updated
4034          * without delay
4035          */
4036         if (!btrfs_is_free_space_inode(BTRFS_I(inode))
4037             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
4038             && !test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags)) {
4039                 btrfs_update_root_times(trans, root);
4040
4041                 ret = btrfs_delayed_update_inode(trans, root, inode);
4042                 if (!ret)
4043                         btrfs_set_inode_last_trans(trans, inode);
4044                 return ret;
4045         }
4046
4047         return btrfs_update_inode_item(trans, root, inode);
4048 }
4049
4050 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
4051                                          struct btrfs_root *root,
4052                                          struct inode *inode)
4053 {
4054         int ret;
4055
4056         ret = btrfs_update_inode(trans, root, inode);
4057         if (ret == -ENOSPC)
4058                 return btrfs_update_inode_item(trans, root, inode);
4059         return ret;
4060 }
4061
4062 /*
4063  * unlink helper that gets used here in inode.c and in the tree logging
4064  * recovery code.  It remove a link in a directory with a given name, and
4065  * also drops the back refs in the inode to the directory
4066  */
4067 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4068                                 struct btrfs_root *root,
4069                                 struct btrfs_inode *dir,
4070                                 struct btrfs_inode *inode,
4071                                 const char *name, int name_len)
4072 {
4073         struct btrfs_fs_info *fs_info = root->fs_info;
4074         struct btrfs_path *path;
4075         int ret = 0;
4076         struct btrfs_dir_item *di;
4077         u64 index;
4078         u64 ino = btrfs_ino(inode);
4079         u64 dir_ino = btrfs_ino(dir);
4080
4081         path = btrfs_alloc_path();
4082         if (!path) {
4083                 ret = -ENOMEM;
4084                 goto out;
4085         }
4086
4087         path->leave_spinning = 1;
4088         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4089                                     name, name_len, -1);
4090         if (IS_ERR_OR_NULL(di)) {
4091                 ret = di ? PTR_ERR(di) : -ENOENT;
4092                 goto err;
4093         }
4094         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4095         if (ret)
4096                 goto err;
4097         btrfs_release_path(path);
4098
4099         /*
4100          * If we don't have dir index, we have to get it by looking up
4101          * the inode ref, since we get the inode ref, remove it directly,
4102          * it is unnecessary to do delayed deletion.
4103          *
4104          * But if we have dir index, needn't search inode ref to get it.
4105          * Since the inode ref is close to the inode item, it is better
4106          * that we delay to delete it, and just do this deletion when
4107          * we update the inode item.
4108          */
4109         if (inode->dir_index) {
4110                 ret = btrfs_delayed_delete_inode_ref(inode);
4111                 if (!ret) {
4112                         index = inode->dir_index;
4113                         goto skip_backref;
4114                 }
4115         }
4116
4117         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
4118                                   dir_ino, &index);
4119         if (ret) {
4120                 btrfs_info(fs_info,
4121                         "failed to delete reference to %.*s, inode %llu parent %llu",
4122                         name_len, name, ino, dir_ino);
4123                 btrfs_abort_transaction(trans, ret);
4124                 goto err;
4125         }
4126 skip_backref:
4127         ret = btrfs_delete_delayed_dir_index(trans, dir, index);
4128         if (ret) {
4129                 btrfs_abort_transaction(trans, ret);
4130                 goto err;
4131         }
4132
4133         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len, inode,
4134                         dir_ino);
4135         if (ret != 0 && ret != -ENOENT) {
4136                 btrfs_abort_transaction(trans, ret);
4137                 goto err;
4138         }
4139
4140         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len, dir,
4141                         index);
4142         if (ret == -ENOENT)
4143                 ret = 0;
4144         else if (ret)
4145                 btrfs_abort_transaction(trans, ret);
4146
4147         /*
4148          * If we have a pending delayed iput we could end up with the final iput
4149          * being run in btrfs-cleaner context.  If we have enough of these built
4150          * up we can end up burning a lot of time in btrfs-cleaner without any
4151          * way to throttle the unlinks.  Since we're currently holding a ref on
4152          * the inode we can run the delayed iput here without any issues as the
4153          * final iput won't be done until after we drop the ref we're currently
4154          * holding.
4155          */
4156         btrfs_run_delayed_iput(fs_info, inode);
4157 err:
4158         btrfs_free_path(path);
4159         if (ret)
4160                 goto out;
4161
4162         btrfs_i_size_write(dir, dir->vfs_inode.i_size - name_len * 2);
4163         inode_inc_iversion(&inode->vfs_inode);
4164         inode_inc_iversion(&dir->vfs_inode);
4165         inode->vfs_inode.i_ctime = dir->vfs_inode.i_mtime =
4166                 dir->vfs_inode.i_ctime = current_time(&inode->vfs_inode);
4167         ret = btrfs_update_inode(trans, root, &dir->vfs_inode);
4168 out:
4169         return ret;
4170 }
4171
4172 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4173                        struct btrfs_root *root,
4174                        struct btrfs_inode *dir, struct btrfs_inode *inode,
4175                        const char *name, int name_len)
4176 {
4177         int ret;
4178         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
4179         if (!ret) {
4180                 drop_nlink(&inode->vfs_inode);
4181                 ret = btrfs_update_inode(trans, root, &inode->vfs_inode);
4182         }
4183         return ret;
4184 }
4185
4186 /*
4187  * helper to start transaction for unlink and rmdir.
4188  *
4189  * unlink and rmdir are special in btrfs, they do not always free space, so
4190  * if we cannot make our reservations the normal way try and see if there is
4191  * plenty of slack room in the global reserve to migrate, otherwise we cannot
4192  * allow the unlink to occur.
4193  */
4194 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
4195 {
4196         struct btrfs_root *root = BTRFS_I(dir)->root;
4197
4198         /*
4199          * 1 for the possible orphan item
4200          * 1 for the dir item
4201          * 1 for the dir index
4202          * 1 for the inode ref
4203          * 1 for the inode
4204          */
4205         return btrfs_start_transaction_fallback_global_rsv(root, 5, 5);
4206 }
4207
4208 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
4209 {
4210         struct btrfs_root *root = BTRFS_I(dir)->root;
4211         struct btrfs_trans_handle *trans;
4212         struct inode *inode = d_inode(dentry);
4213         int ret;
4214
4215         trans = __unlink_start_trans(dir);
4216         if (IS_ERR(trans))
4217                 return PTR_ERR(trans);
4218
4219         btrfs_record_unlink_dir(trans, BTRFS_I(dir), BTRFS_I(d_inode(dentry)),
4220                         0);
4221
4222         ret = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4223                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4224                         dentry->d_name.len);
4225         if (ret)
4226                 goto out;
4227
4228         if (inode->i_nlink == 0) {
4229                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
4230                 if (ret)
4231                         goto out;
4232         }
4233
4234 out:
4235         btrfs_end_transaction(trans);
4236         btrfs_btree_balance_dirty(root->fs_info);
4237         return ret;
4238 }
4239
4240 static int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
4241                                struct inode *dir, u64 objectid,
4242                                const char *name, int name_len)
4243 {
4244         struct btrfs_root *root = BTRFS_I(dir)->root;
4245         struct btrfs_path *path;
4246         struct extent_buffer *leaf;
4247         struct btrfs_dir_item *di;
4248         struct btrfs_key key;
4249         u64 index;
4250         int ret;
4251         u64 dir_ino = btrfs_ino(BTRFS_I(dir));
4252
4253         path = btrfs_alloc_path();
4254         if (!path)
4255                 return -ENOMEM;
4256
4257         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4258                                    name, name_len, -1);
4259         if (IS_ERR_OR_NULL(di)) {
4260                 ret = di ? PTR_ERR(di) : -ENOENT;
4261                 goto out;
4262         }
4263
4264         leaf = path->nodes[0];
4265         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4266         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
4267         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4268         if (ret) {
4269                 btrfs_abort_transaction(trans, ret);
4270                 goto out;
4271         }
4272         btrfs_release_path(path);
4273
4274         ret = btrfs_del_root_ref(trans, objectid, root->root_key.objectid,
4275                                  dir_ino, &index, name, name_len);
4276         if (ret < 0) {
4277                 if (ret != -ENOENT) {
4278                         btrfs_abort_transaction(trans, ret);
4279                         goto out;
4280                 }
4281                 di = btrfs_search_dir_index_item(root, path, dir_ino,
4282                                                  name, name_len);
4283                 if (IS_ERR_OR_NULL(di)) {
4284                         if (!di)
4285                                 ret = -ENOENT;
4286                         else
4287                                 ret = PTR_ERR(di);
4288                         btrfs_abort_transaction(trans, ret);
4289                         goto out;
4290                 }
4291
4292                 leaf = path->nodes[0];
4293                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4294                 index = key.offset;
4295         }
4296         btrfs_release_path(path);
4297
4298         ret = btrfs_delete_delayed_dir_index(trans, BTRFS_I(dir), index);
4299         if (ret) {
4300                 btrfs_abort_transaction(trans, ret);
4301                 goto out;
4302         }
4303
4304         btrfs_i_size_write(BTRFS_I(dir), dir->i_size - name_len * 2);
4305         inode_inc_iversion(dir);
4306         dir->i_mtime = dir->i_ctime = current_time(dir);
4307         ret = btrfs_update_inode_fallback(trans, root, dir);
4308         if (ret)
4309                 btrfs_abort_transaction(trans, ret);
4310 out:
4311         btrfs_free_path(path);
4312         return ret;
4313 }
4314
4315 /*
4316  * Helper to check if the subvolume references other subvolumes or if it's
4317  * default.
4318  */
4319 static noinline int may_destroy_subvol(struct btrfs_root *root)
4320 {
4321         struct btrfs_fs_info *fs_info = root->fs_info;
4322         struct btrfs_path *path;
4323         struct btrfs_dir_item *di;
4324         struct btrfs_key key;
4325         u64 dir_id;
4326         int ret;
4327
4328         path = btrfs_alloc_path();
4329         if (!path)
4330                 return -ENOMEM;
4331
4332         /* Make sure this root isn't set as the default subvol */
4333         dir_id = btrfs_super_root_dir(fs_info->super_copy);
4334         di = btrfs_lookup_dir_item(NULL, fs_info->tree_root, path,
4335                                    dir_id, "default", 7, 0);
4336         if (di && !IS_ERR(di)) {
4337                 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &key);
4338                 if (key.objectid == root->root_key.objectid) {
4339                         ret = -EPERM;
4340                         btrfs_err(fs_info,
4341                                   "deleting default subvolume %llu is not allowed",
4342                                   key.objectid);
4343                         goto out;
4344                 }
4345                 btrfs_release_path(path);
4346         }
4347
4348         key.objectid = root->root_key.objectid;
4349         key.type = BTRFS_ROOT_REF_KEY;
4350         key.offset = (u64)-1;
4351
4352         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
4353         if (ret < 0)
4354                 goto out;
4355         BUG_ON(ret == 0);
4356
4357         ret = 0;
4358         if (path->slots[0] > 0) {
4359                 path->slots[0]--;
4360                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
4361                 if (key.objectid == root->root_key.objectid &&
4362                     key.type == BTRFS_ROOT_REF_KEY)
4363                         ret = -ENOTEMPTY;
4364         }
4365 out:
4366         btrfs_free_path(path);
4367         return ret;
4368 }
4369
4370 /* Delete all dentries for inodes belonging to the root */
4371 static void btrfs_prune_dentries(struct btrfs_root *root)
4372 {
4373         struct btrfs_fs_info *fs_info = root->fs_info;
4374         struct rb_node *node;
4375         struct rb_node *prev;
4376         struct btrfs_inode *entry;
4377         struct inode *inode;
4378         u64 objectid = 0;
4379
4380         if (!test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
4381                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
4382
4383         spin_lock(&root->inode_lock);
4384 again:
4385         node = root->inode_tree.rb_node;
4386         prev = NULL;
4387         while (node) {
4388                 prev = node;
4389                 entry = rb_entry(node, struct btrfs_inode, rb_node);
4390
4391                 if (objectid < btrfs_ino(entry))
4392                         node = node->rb_left;
4393                 else if (objectid > btrfs_ino(entry))
4394                         node = node->rb_right;
4395                 else
4396                         break;
4397         }
4398         if (!node) {
4399                 while (prev) {
4400                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
4401                         if (objectid <= btrfs_ino(entry)) {
4402                                 node = prev;
4403                                 break;
4404                         }
4405                         prev = rb_next(prev);
4406                 }
4407         }
4408         while (node) {
4409                 entry = rb_entry(node, struct btrfs_inode, rb_node);
4410                 objectid = btrfs_ino(entry) + 1;
4411                 inode = igrab(&entry->vfs_inode);
4412                 if (inode) {
4413                         spin_unlock(&root->inode_lock);
4414                         if (atomic_read(&inode->i_count) > 1)
4415                                 d_prune_aliases(inode);
4416                         /*
4417                          * btrfs_drop_inode will have it removed from the inode
4418                          * cache when its usage count hits zero.
4419                          */
4420                         iput(inode);
4421                         cond_resched();
4422                         spin_lock(&root->inode_lock);
4423                         goto again;
4424                 }
4425
4426                 if (cond_resched_lock(&root->inode_lock))
4427                         goto again;
4428
4429                 node = rb_next(node);
4430         }
4431         spin_unlock(&root->inode_lock);
4432 }
4433
4434 int btrfs_delete_subvolume(struct inode *dir, struct dentry *dentry)
4435 {
4436         struct btrfs_fs_info *fs_info = btrfs_sb(dentry->d_sb);
4437         struct btrfs_root *root = BTRFS_I(dir)->root;
4438         struct inode *inode = d_inode(dentry);
4439         struct btrfs_root *dest = BTRFS_I(inode)->root;
4440         struct btrfs_trans_handle *trans;
4441         struct btrfs_block_rsv block_rsv;
4442         u64 root_flags;
4443         int ret;
4444         int err;
4445
4446         /*
4447          * Don't allow to delete a subvolume with send in progress. This is
4448          * inside the inode lock so the error handling that has to drop the bit
4449          * again is not run concurrently.
4450          */
4451         spin_lock(&dest->root_item_lock);
4452         if (dest->send_in_progress) {
4453                 spin_unlock(&dest->root_item_lock);
4454                 btrfs_warn(fs_info,
4455                            "attempt to delete subvolume %llu during send",
4456                            dest->root_key.objectid);
4457                 return -EPERM;
4458         }
4459         root_flags = btrfs_root_flags(&dest->root_item);
4460         btrfs_set_root_flags(&dest->root_item,
4461                              root_flags | BTRFS_ROOT_SUBVOL_DEAD);
4462         spin_unlock(&dest->root_item_lock);
4463
4464         down_write(&fs_info->subvol_sem);
4465
4466         err = may_destroy_subvol(dest);
4467         if (err)
4468                 goto out_up_write;
4469
4470         btrfs_init_block_rsv(&block_rsv, BTRFS_BLOCK_RSV_TEMP);
4471         /*
4472          * One for dir inode,
4473          * two for dir entries,
4474          * two for root ref/backref.
4475          */
4476         err = btrfs_subvolume_reserve_metadata(root, &block_rsv, 5, true);
4477         if (err)
4478                 goto out_up_write;
4479
4480         trans = btrfs_start_transaction(root, 0);
4481         if (IS_ERR(trans)) {
4482                 err = PTR_ERR(trans);
4483                 goto out_release;
4484         }
4485         trans->block_rsv = &block_rsv;
4486         trans->bytes_reserved = block_rsv.size;
4487
4488         btrfs_record_snapshot_destroy(trans, BTRFS_I(dir));
4489
4490         ret = btrfs_unlink_subvol(trans, dir, dest->root_key.objectid,
4491                                   dentry->d_name.name, dentry->d_name.len);
4492         if (ret) {
4493                 err = ret;
4494                 btrfs_abort_transaction(trans, ret);
4495                 goto out_end_trans;
4496         }
4497
4498         btrfs_record_root_in_trans(trans, dest);
4499
4500         memset(&dest->root_item.drop_progress, 0,
4501                 sizeof(dest->root_item.drop_progress));
4502         dest->root_item.drop_level = 0;
4503         btrfs_set_root_refs(&dest->root_item, 0);
4504
4505         if (!test_and_set_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &dest->state)) {
4506                 ret = btrfs_insert_orphan_item(trans,
4507                                         fs_info->tree_root,
4508                                         dest->root_key.objectid);
4509                 if (ret) {
4510                         btrfs_abort_transaction(trans, ret);
4511                         err = ret;
4512                         goto out_end_trans;
4513                 }
4514         }
4515
4516         ret = btrfs_uuid_tree_remove(trans, dest->root_item.uuid,
4517                                   BTRFS_UUID_KEY_SUBVOL,
4518                                   dest->root_key.objectid);
4519         if (ret && ret != -ENOENT) {
4520                 btrfs_abort_transaction(trans, ret);
4521                 err = ret;
4522                 goto out_end_trans;
4523         }
4524         if (!btrfs_is_empty_uuid(dest->root_item.received_uuid)) {
4525                 ret = btrfs_uuid_tree_remove(trans,
4526                                           dest->root_item.received_uuid,
4527                                           BTRFS_UUID_KEY_RECEIVED_SUBVOL,
4528                                           dest->root_key.objectid);
4529                 if (ret && ret != -ENOENT) {
4530                         btrfs_abort_transaction(trans, ret);
4531                         err = ret;
4532                         goto out_end_trans;
4533                 }
4534         }
4535
4536 out_end_trans:
4537         trans->block_rsv = NULL;
4538         trans->bytes_reserved = 0;
4539         ret = btrfs_end_transaction(trans);
4540         if (ret && !err)
4541                 err = ret;
4542         inode->i_flags |= S_DEAD;
4543 out_release:
4544         btrfs_subvolume_release_metadata(fs_info, &block_rsv);
4545 out_up_write:
4546         up_write(&fs_info->subvol_sem);
4547         if (err) {
4548                 spin_lock(&dest->root_item_lock);
4549                 root_flags = btrfs_root_flags(&dest->root_item);
4550                 btrfs_set_root_flags(&dest->root_item,
4551                                 root_flags & ~BTRFS_ROOT_SUBVOL_DEAD);
4552                 spin_unlock(&dest->root_item_lock);
4553         } else {
4554                 d_invalidate(dentry);
4555                 btrfs_prune_dentries(dest);
4556                 ASSERT(dest->send_in_progress == 0);
4557
4558                 /* the last ref */
4559                 if (dest->ino_cache_inode) {
4560                         iput(dest->ino_cache_inode);
4561                         dest->ino_cache_inode = NULL;
4562                 }
4563         }
4564
4565         return err;
4566 }
4567
4568 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4569 {
4570         struct inode *inode = d_inode(dentry);
4571         int err = 0;
4572         struct btrfs_root *root = BTRFS_I(dir)->root;
4573         struct btrfs_trans_handle *trans;
4574         u64 last_unlink_trans;
4575
4576         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4577                 return -ENOTEMPTY;
4578         if (btrfs_ino(BTRFS_I(inode)) == BTRFS_FIRST_FREE_OBJECTID)
4579                 return btrfs_delete_subvolume(dir, dentry);
4580
4581         trans = __unlink_start_trans(dir);
4582         if (IS_ERR(trans))
4583                 return PTR_ERR(trans);
4584
4585         if (unlikely(btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4586                 err = btrfs_unlink_subvol(trans, dir,
4587                                           BTRFS_I(inode)->location.objectid,
4588                                           dentry->d_name.name,
4589                                           dentry->d_name.len);
4590                 goto out;
4591         }
4592
4593         err = btrfs_orphan_add(trans, BTRFS_I(inode));
4594         if (err)
4595                 goto out;
4596
4597         last_unlink_trans = BTRFS_I(inode)->last_unlink_trans;
4598
4599         /* now the directory is empty */
4600         err = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4601                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4602                         dentry->d_name.len);
4603         if (!err) {
4604                 btrfs_i_size_write(BTRFS_I(inode), 0);
4605                 /*
4606                  * Propagate the last_unlink_trans value of the deleted dir to
4607                  * its parent directory. This is to prevent an unrecoverable
4608                  * log tree in the case we do something like this:
4609                  * 1) create dir foo
4610                  * 2) create snapshot under dir foo
4611                  * 3) delete the snapshot
4612                  * 4) rmdir foo
4613                  * 5) mkdir foo
4614                  * 6) fsync foo or some file inside foo
4615                  */
4616                 if (last_unlink_trans >= trans->transid)
4617                         BTRFS_I(dir)->last_unlink_trans = last_unlink_trans;
4618         }
4619 out:
4620         btrfs_end_transaction(trans);
4621         btrfs_btree_balance_dirty(root->fs_info);
4622
4623         return err;
4624 }
4625
4626 /*
4627  * Return this if we need to call truncate_block for the last bit of the
4628  * truncate.
4629  */
4630 #define NEED_TRUNCATE_BLOCK 1
4631
4632 /*
4633  * this can truncate away extent items, csum items and directory items.
4634  * It starts at a high offset and removes keys until it can't find
4635  * any higher than new_size
4636  *
4637  * csum items that cross the new i_size are truncated to the new size
4638  * as well.
4639  *
4640  * min_type is the minimum key type to truncate down to.  If set to 0, this
4641  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4642  */
4643 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4644                                struct btrfs_root *root,
4645                                struct inode *inode,
4646                                u64 new_size, u32 min_type)
4647 {
4648         struct btrfs_fs_info *fs_info = root->fs_info;
4649         struct btrfs_path *path;
4650         struct extent_buffer *leaf;
4651         struct btrfs_file_extent_item *fi;
4652         struct btrfs_key key;
4653         struct btrfs_key found_key;
4654         u64 extent_start = 0;
4655         u64 extent_num_bytes = 0;
4656         u64 extent_offset = 0;
4657         u64 item_end = 0;
4658         u64 last_size = new_size;
4659         u32 found_type = (u8)-1;
4660         int found_extent;
4661         int del_item;
4662         int pending_del_nr = 0;
4663         int pending_del_slot = 0;
4664         int extent_type = -1;
4665         int ret;
4666         u64 ino = btrfs_ino(BTRFS_I(inode));
4667         u64 bytes_deleted = 0;
4668         bool be_nice = false;
4669         bool should_throttle = false;
4670
4671         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4672
4673         /*
4674          * for non-free space inodes and ref cows, we want to back off from
4675          * time to time
4676          */
4677         if (!btrfs_is_free_space_inode(BTRFS_I(inode)) &&
4678             test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4679                 be_nice = true;
4680
4681         path = btrfs_alloc_path();
4682         if (!path)
4683                 return -ENOMEM;
4684         path->reada = READA_BACK;
4685
4686         /*
4687          * We want to drop from the next block forward in case this new size is
4688          * not block aligned since we will be keeping the last block of the
4689          * extent just the way it is.
4690          */
4691         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4692             root == fs_info->tree_root)
4693                 btrfs_drop_extent_cache(BTRFS_I(inode), ALIGN(new_size,
4694                                         fs_info->sectorsize),
4695                                         (u64)-1, 0);
4696
4697         /*
4698          * This function is also used to drop the items in the log tree before
4699          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4700          * it is used to drop the logged items. So we shouldn't kill the delayed
4701          * items.
4702          */
4703         if (min_type == 0 && root == BTRFS_I(inode)->root)
4704                 btrfs_kill_delayed_inode_items(BTRFS_I(inode));
4705
4706         key.objectid = ino;
4707         key.offset = (u64)-1;
4708         key.type = (u8)-1;
4709
4710 search_again:
4711         /*
4712          * with a 16K leaf size and 128MB extents, you can actually queue
4713          * up a huge file in a single leaf.  Most of the time that
4714          * bytes_deleted is > 0, it will be huge by the time we get here
4715          */
4716         if (be_nice && bytes_deleted > SZ_32M &&
4717             btrfs_should_end_transaction(trans)) {
4718                 ret = -EAGAIN;
4719                 goto out;
4720         }
4721
4722         path->leave_spinning = 1;
4723         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4724         if (ret < 0)
4725                 goto out;
4726
4727         if (ret > 0) {
4728                 ret = 0;
4729                 /* there are no items in the tree for us to truncate, we're
4730                  * done
4731                  */
4732                 if (path->slots[0] == 0)
4733                         goto out;
4734                 path->slots[0]--;
4735         }
4736
4737         while (1) {
4738                 fi = NULL;
4739                 leaf = path->nodes[0];
4740                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4741                 found_type = found_key.type;
4742
4743                 if (found_key.objectid != ino)
4744                         break;
4745
4746                 if (found_type < min_type)
4747                         break;
4748
4749                 item_end = found_key.offset;
4750                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4751                         fi = btrfs_item_ptr(leaf, path->slots[0],
4752                                             struct btrfs_file_extent_item);
4753                         extent_type = btrfs_file_extent_type(leaf, fi);
4754                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4755                                 item_end +=
4756                                     btrfs_file_extent_num_bytes(leaf, fi);
4757
4758                                 trace_btrfs_truncate_show_fi_regular(
4759                                         BTRFS_I(inode), leaf, fi,
4760                                         found_key.offset);
4761                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4762                                 item_end += btrfs_file_extent_ram_bytes(leaf,
4763                                                                         fi);
4764
4765                                 trace_btrfs_truncate_show_fi_inline(
4766                                         BTRFS_I(inode), leaf, fi, path->slots[0],
4767                                         found_key.offset);
4768                         }
4769                         item_end--;
4770                 }
4771                 if (found_type > min_type) {
4772                         del_item = 1;
4773                 } else {
4774                         if (item_end < new_size)
4775                                 break;
4776                         if (found_key.offset >= new_size)
4777                                 del_item = 1;
4778                         else
4779                                 del_item = 0;
4780                 }
4781                 found_extent = 0;
4782                 /* FIXME, shrink the extent if the ref count is only 1 */
4783                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4784                         goto delete;
4785
4786                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4787                         u64 num_dec;
4788                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4789                         if (!del_item) {
4790                                 u64 orig_num_bytes =
4791                                         btrfs_file_extent_num_bytes(leaf, fi);
4792                                 extent_num_bytes = ALIGN(new_size -
4793                                                 found_key.offset,
4794                                                 fs_info->sectorsize);
4795                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4796                                                          extent_num_bytes);
4797                                 num_dec = (orig_num_bytes -
4798                                            extent_num_bytes);
4799                                 if (test_bit(BTRFS_ROOT_REF_COWS,
4800                                              &root->state) &&
4801                                     extent_start != 0)
4802                                         inode_sub_bytes(inode, num_dec);
4803                                 btrfs_mark_buffer_dirty(leaf);
4804                         } else {
4805                                 extent_num_bytes =
4806                                         btrfs_file_extent_disk_num_bytes(leaf,
4807                                                                          fi);
4808                                 extent_offset = found_key.offset -
4809                                         btrfs_file_extent_offset(leaf, fi);
4810
4811                                 /* FIXME blocksize != 4096 */
4812                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4813                                 if (extent_start != 0) {
4814                                         found_extent = 1;
4815                                         if (test_bit(BTRFS_ROOT_REF_COWS,
4816                                                      &root->state))
4817                                                 inode_sub_bytes(inode, num_dec);
4818                                 }
4819                         }
4820                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4821                         /*
4822                          * we can't truncate inline items that have had
4823                          * special encodings
4824                          */
4825                         if (!del_item &&
4826                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4827                             btrfs_file_extent_other_encoding(leaf, fi) == 0 &&
4828                             btrfs_file_extent_compression(leaf, fi) == 0) {
4829                                 u32 size = (u32)(new_size - found_key.offset);
4830
4831                                 btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4832                                 size = btrfs_file_extent_calc_inline_size(size);
4833                                 btrfs_truncate_item(path, size, 1);
4834                         } else if (!del_item) {
4835                                 /*
4836                                  * We have to bail so the last_size is set to
4837                                  * just before this extent.
4838                                  */
4839                                 ret = NEED_TRUNCATE_BLOCK;
4840                                 break;
4841                         }
4842
4843                         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4844                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4845                 }
4846 delete:
4847                 if (del_item)
4848                         last_size = found_key.offset;
4849                 else
4850                         last_size = new_size;
4851                 if (del_item) {
4852                         if (!pending_del_nr) {
4853                                 /* no pending yet, add ourselves */
4854                                 pending_del_slot = path->slots[0];
4855                                 pending_del_nr = 1;
4856                         } else if (pending_del_nr &&
4857                                    path->slots[0] + 1 == pending_del_slot) {
4858                                 /* hop on the pending chunk */
4859                                 pending_del_nr++;
4860                                 pending_del_slot = path->slots[0];
4861                         } else {
4862                                 BUG();
4863                         }
4864                 } else {
4865                         break;
4866                 }
4867                 should_throttle = false;
4868
4869                 if (found_extent &&
4870                     (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4871                      root == fs_info->tree_root)) {
4872                         struct btrfs_ref ref = { 0 };
4873
4874                         btrfs_set_path_blocking(path);
4875                         bytes_deleted += extent_num_bytes;
4876
4877                         btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF,
4878                                         extent_start, extent_num_bytes, 0);
4879                         ref.real_root = root->root_key.objectid;
4880                         btrfs_init_data_ref(&ref, btrfs_header_owner(leaf),
4881                                         ino, extent_offset);
4882                         ret = btrfs_free_extent(trans, &ref);
4883                         if (ret) {
4884                                 btrfs_abort_transaction(trans, ret);
4885                                 break;
4886                         }
4887                         if (be_nice) {
4888                                 if (btrfs_should_throttle_delayed_refs(trans))
4889                                         should_throttle = true;
4890                         }
4891                 }
4892
4893                 if (found_type == BTRFS_INODE_ITEM_KEY)
4894                         break;
4895
4896                 if (path->slots[0] == 0 ||
4897                     path->slots[0] != pending_del_slot ||
4898                     should_throttle) {
4899                         if (pending_del_nr) {
4900                                 ret = btrfs_del_items(trans, root, path,
4901                                                 pending_del_slot,
4902                                                 pending_del_nr);
4903                                 if (ret) {
4904                                         btrfs_abort_transaction(trans, ret);
4905                                         break;
4906                                 }
4907                                 pending_del_nr = 0;
4908                         }
4909                         btrfs_release_path(path);
4910
4911                         /*
4912                          * We can generate a lot of delayed refs, so we need to
4913                          * throttle every once and a while and make sure we're
4914                          * adding enough space to keep up with the work we are
4915                          * generating.  Since we hold a transaction here we
4916                          * can't flush, and we don't want to FLUSH_LIMIT because
4917                          * we could have generated too many delayed refs to
4918                          * actually allocate, so just bail if we're short and
4919                          * let the normal reservation dance happen higher up.
4920                          */
4921                         if (should_throttle) {
4922                                 ret = btrfs_delayed_refs_rsv_refill(fs_info,
4923                                                         BTRFS_RESERVE_NO_FLUSH);
4924                                 if (ret) {
4925                                         ret = -EAGAIN;
4926                                         break;
4927                                 }
4928                         }
4929                         goto search_again;
4930                 } else {
4931                         path->slots[0]--;
4932                 }
4933         }
4934 out:
4935         if (ret >= 0 && pending_del_nr) {
4936                 int err;
4937
4938                 err = btrfs_del_items(trans, root, path, pending_del_slot,
4939                                       pending_del_nr);
4940                 if (err) {
4941                         btrfs_abort_transaction(trans, err);
4942                         ret = err;
4943                 }
4944         }
4945         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4946                 ASSERT(last_size >= new_size);
4947                 if (!ret && last_size > new_size)
4948                         last_size = new_size;
4949                 btrfs_ordered_update_i_size(inode, last_size, NULL);
4950         }
4951
4952         btrfs_free_path(path);
4953         return ret;
4954 }
4955
4956 /*
4957  * btrfs_truncate_block - read, zero a chunk and write a block
4958  * @inode - inode that we're zeroing
4959  * @from - the offset to start zeroing
4960  * @len - the length to zero, 0 to zero the entire range respective to the
4961  *      offset
4962  * @front - zero up to the offset instead of from the offset on
4963  *
4964  * This will find the block for the "from" offset and cow the block and zero the
4965  * part we want to zero.  This is used with truncate and hole punching.
4966  */
4967 int btrfs_truncate_block(struct inode *inode, loff_t from, loff_t len,
4968                         int front)
4969 {
4970         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4971         struct address_space *mapping = inode->i_mapping;
4972         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4973         struct btrfs_ordered_extent *ordered;
4974         struct extent_state *cached_state = NULL;
4975         struct extent_changeset *data_reserved = NULL;
4976         char *kaddr;
4977         u32 blocksize = fs_info->sectorsize;
4978         pgoff_t index = from >> PAGE_SHIFT;
4979         unsigned offset = from & (blocksize - 1);
4980         struct page *page;
4981         gfp_t mask = btrfs_alloc_write_mask(mapping);
4982         int ret = 0;
4983         u64 block_start;
4984         u64 block_end;
4985
4986         if (IS_ALIGNED(offset, blocksize) &&
4987             (!len || IS_ALIGNED(len, blocksize)))
4988                 goto out;
4989
4990         block_start = round_down(from, blocksize);
4991         block_end = block_start + blocksize - 1;
4992
4993         ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
4994                                            block_start, blocksize);
4995         if (ret)
4996                 goto out;
4997
4998 again:
4999         page = find_or_create_page(mapping, index, mask);
5000         if (!page) {
5001                 btrfs_delalloc_release_space(inode, data_reserved,
5002                                              block_start, blocksize, true);
5003                 btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize);
5004                 ret = -ENOMEM;
5005                 goto out;
5006         }
5007
5008         if (!PageUptodate(page)) {
5009                 ret = btrfs_readpage(NULL, page);
5010                 lock_page(page);
5011                 if (page->mapping != mapping) {
5012                         unlock_page(page);
5013                         put_page(page);
5014                         goto again;
5015                 }
5016                 if (!PageUptodate(page)) {
5017                         ret = -EIO;
5018                         goto out_unlock;
5019                 }
5020         }
5021         wait_on_page_writeback(page);
5022
5023         lock_extent_bits(io_tree, block_start, block_end, &cached_state);
5024         set_page_extent_mapped(page);
5025
5026         ordered = btrfs_lookup_ordered_extent(inode, block_start);
5027         if (ordered) {
5028                 unlock_extent_cached(io_tree, block_start, block_end,
5029                                      &cached_state);
5030                 unlock_page(page);
5031                 put_page(page);
5032                 btrfs_start_ordered_extent(inode, ordered, 1);
5033                 btrfs_put_ordered_extent(ordered);
5034                 goto again;
5035         }
5036
5037         clear_extent_bit(&BTRFS_I(inode)->io_tree, block_start, block_end,
5038                          EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
5039                          0, 0, &cached_state);
5040
5041         ret = btrfs_set_extent_delalloc(inode, block_start, block_end, 0,
5042                                         &cached_state);
5043         if (ret) {
5044                 unlock_extent_cached(io_tree, block_start, block_end,
5045                                      &cached_state);
5046                 goto out_unlock;
5047         }
5048
5049         if (offset != blocksize) {
5050                 if (!len)
5051                         len = blocksize - offset;
5052                 kaddr = kmap(page);
5053                 if (front)
5054                         memset(kaddr + (block_start - page_offset(page)),
5055                                 0, offset);
5056                 else
5057                         memset(kaddr + (block_start - page_offset(page)) +  offset,
5058                                 0, len);
5059                 flush_dcache_page(page);
5060                 kunmap(page);
5061         }
5062         ClearPageChecked(page);
5063         set_page_dirty(page);
5064         unlock_extent_cached(io_tree, block_start, block_end, &cached_state);
5065
5066 out_unlock:
5067         if (ret)
5068                 btrfs_delalloc_release_space(inode, data_reserved, block_start,
5069                                              blocksize, true);
5070         btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize);
5071         unlock_page(page);
5072         put_page(page);
5073 out:
5074         extent_changeset_free(data_reserved);
5075         return ret;
5076 }
5077
5078 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
5079                              u64 offset, u64 len)
5080 {
5081         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5082         struct btrfs_trans_handle *trans;
5083         int ret;
5084
5085         /*
5086          * Still need to make sure the inode looks like it's been updated so
5087          * that any holes get logged if we fsync.
5088          */
5089         if (btrfs_fs_incompat(fs_info, NO_HOLES)) {
5090                 BTRFS_I(inode)->last_trans = fs_info->generation;
5091                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
5092                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
5093                 return 0;
5094         }
5095
5096         /*
5097          * 1 - for the one we're dropping
5098          * 1 - for the one we're adding
5099          * 1 - for updating the inode.
5100          */
5101         trans = btrfs_start_transaction(root, 3);
5102         if (IS_ERR(trans))
5103                 return PTR_ERR(trans);
5104
5105         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
5106         if (ret) {
5107                 btrfs_abort_transaction(trans, ret);
5108                 btrfs_end_transaction(trans);
5109                 return ret;
5110         }
5111
5112         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(BTRFS_I(inode)),
5113                         offset, 0, 0, len, 0, len, 0, 0, 0);
5114         if (ret)
5115                 btrfs_abort_transaction(trans, ret);
5116         else
5117                 btrfs_update_inode(trans, root, inode);
5118         btrfs_end_transaction(trans);
5119         return ret;
5120 }
5121
5122 /*
5123  * This function puts in dummy file extents for the area we're creating a hole
5124  * for.  So if we are truncating this file to a larger size we need to insert
5125  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
5126  * the range between oldsize and size
5127  */
5128 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
5129 {
5130         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5131         struct btrfs_root *root = BTRFS_I(inode)->root;
5132         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5133         struct extent_map *em = NULL;
5134         struct extent_state *cached_state = NULL;
5135         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
5136         u64 hole_start = ALIGN(oldsize, fs_info->sectorsize);
5137         u64 block_end = ALIGN(size, fs_info->sectorsize);
5138         u64 last_byte;
5139         u64 cur_offset;
5140         u64 hole_size;
5141         int err = 0;
5142
5143         /*
5144          * If our size started in the middle of a block we need to zero out the
5145          * rest of the block before we expand the i_size, otherwise we could
5146          * expose stale data.
5147          */
5148         err = btrfs_truncate_block(inode, oldsize, 0, 0);
5149         if (err)
5150                 return err;
5151
5152         if (size <= hole_start)
5153                 return 0;
5154
5155         btrfs_lock_and_flush_ordered_range(io_tree, BTRFS_I(inode), hole_start,
5156                                            block_end - 1, &cached_state);
5157         cur_offset = hole_start;
5158         while (1) {
5159                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
5160                                 block_end - cur_offset, 0);
5161                 if (IS_ERR(em)) {
5162                         err = PTR_ERR(em);
5163                         em = NULL;
5164                         break;
5165                 }
5166                 last_byte = min(extent_map_end(em), block_end);
5167                 last_byte = ALIGN(last_byte, fs_info->sectorsize);
5168                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
5169                         struct extent_map *hole_em;
5170                         hole_size = last_byte - cur_offset;
5171
5172                         err = maybe_insert_hole(root, inode, cur_offset,
5173                                                 hole_size);
5174                         if (err)
5175                                 break;
5176                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
5177                                                 cur_offset + hole_size - 1, 0);
5178                         hole_em = alloc_extent_map();
5179                         if (!hole_em) {
5180                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
5181                                         &BTRFS_I(inode)->runtime_flags);
5182                                 goto next;
5183                         }
5184                         hole_em->start = cur_offset;
5185                         hole_em->len = hole_size;
5186                         hole_em->orig_start = cur_offset;
5187
5188                         hole_em->block_start = EXTENT_MAP_HOLE;
5189                         hole_em->block_len = 0;
5190                         hole_em->orig_block_len = 0;
5191                         hole_em->ram_bytes = hole_size;
5192                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
5193                         hole_em->generation = fs_info->generation;
5194
5195                         while (1) {
5196                                 write_lock(&em_tree->lock);
5197                                 err = add_extent_mapping(em_tree, hole_em, 1);
5198                                 write_unlock(&em_tree->lock);
5199                                 if (err != -EEXIST)
5200                                         break;
5201                                 btrfs_drop_extent_cache(BTRFS_I(inode),
5202                                                         cur_offset,
5203                                                         cur_offset +
5204                                                         hole_size - 1, 0);
5205                         }
5206                         free_extent_map(hole_em);
5207                 }
5208 next:
5209                 free_extent_map(em);
5210                 em = NULL;
5211                 cur_offset = last_byte;
5212                 if (cur_offset >= block_end)
5213                         break;
5214         }
5215         free_extent_map(em);
5216         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state);
5217         return err;
5218 }
5219
5220 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
5221 {
5222         struct btrfs_root *root = BTRFS_I(inode)->root;
5223         struct btrfs_trans_handle *trans;
5224         loff_t oldsize = i_size_read(inode);
5225         loff_t newsize = attr->ia_size;
5226         int mask = attr->ia_valid;
5227         int ret;
5228
5229         /*
5230          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
5231          * special case where we need to update the times despite not having
5232          * these flags set.  For all other operations the VFS set these flags
5233          * explicitly if it wants a timestamp update.
5234          */
5235         if (newsize != oldsize) {
5236                 inode_inc_iversion(inode);
5237                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
5238                         inode->i_ctime = inode->i_mtime =
5239                                 current_time(inode);
5240         }
5241
5242         if (newsize > oldsize) {
5243                 /*
5244                  * Don't do an expanding truncate while snapshotting is ongoing.
5245                  * This is to ensure the snapshot captures a fully consistent
5246                  * state of this file - if the snapshot captures this expanding
5247                  * truncation, it must capture all writes that happened before
5248                  * this truncation.
5249                  */
5250                 btrfs_wait_for_snapshot_creation(root);
5251                 ret = btrfs_cont_expand(inode, oldsize, newsize);
5252                 if (ret) {
5253                         btrfs_end_write_no_snapshotting(root);
5254                         return ret;
5255                 }
5256
5257                 trans = btrfs_start_transaction(root, 1);
5258                 if (IS_ERR(trans)) {
5259                         btrfs_end_write_no_snapshotting(root);
5260                         return PTR_ERR(trans);
5261                 }
5262
5263                 i_size_write(inode, newsize);
5264                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
5265                 pagecache_isize_extended(inode, oldsize, newsize);
5266                 ret = btrfs_update_inode(trans, root, inode);
5267                 btrfs_end_write_no_snapshotting(root);
5268                 btrfs_end_transaction(trans);
5269         } else {
5270
5271                 /*
5272                  * We're truncating a file that used to have good data down to
5273                  * zero. Make sure it gets into the ordered flush list so that
5274                  * any new writes get down to disk quickly.
5275                  */
5276                 if (newsize == 0)
5277                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
5278                                 &BTRFS_I(inode)->runtime_flags);
5279
5280                 truncate_setsize(inode, newsize);
5281
5282                 /* Disable nonlocked read DIO to avoid the endless truncate */
5283                 btrfs_inode_block_unlocked_dio(BTRFS_I(inode));
5284                 inode_dio_wait(inode);
5285                 btrfs_inode_resume_unlocked_dio(BTRFS_I(inode));
5286
5287                 ret = btrfs_truncate(inode, newsize == oldsize);
5288                 if (ret && inode->i_nlink) {
5289                         int err;
5290
5291                         /*
5292                          * Truncate failed, so fix up the in-memory size. We
5293                          * adjusted disk_i_size down as we removed extents, so
5294                          * wait for disk_i_size to be stable and then update the
5295                          * in-memory size to match.
5296                          */
5297                         err = btrfs_wait_ordered_range(inode, 0, (u64)-1);
5298                         if (err)
5299                                 return err;
5300                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
5301                 }
5302         }
5303
5304         return ret;
5305 }
5306
5307 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
5308 {
5309         struct inode *inode = d_inode(dentry);
5310         struct btrfs_root *root = BTRFS_I(inode)->root;
5311         int err;
5312
5313         if (btrfs_root_readonly(root))
5314                 return -EROFS;
5315
5316         err = setattr_prepare(dentry, attr);
5317         if (err)
5318                 return err;
5319
5320         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
5321                 err = btrfs_setsize(inode, attr);
5322                 if (err)
5323                         return err;
5324         }
5325
5326         if (attr->ia_valid) {
5327                 setattr_copy(inode, attr);
5328                 inode_inc_iversion(inode);
5329                 err = btrfs_dirty_inode(inode);
5330
5331                 if (!err && attr->ia_valid & ATTR_MODE)
5332                         err = posix_acl_chmod(inode, inode->i_mode);
5333         }
5334
5335         return err;
5336 }
5337
5338 /*
5339  * While truncating the inode pages during eviction, we get the VFS calling
5340  * btrfs_invalidatepage() against each page of the inode. This is slow because
5341  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
5342  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
5343  * extent_state structures over and over, wasting lots of time.
5344  *
5345  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
5346  * those expensive operations on a per page basis and do only the ordered io
5347  * finishing, while we release here the extent_map and extent_state structures,
5348  * without the excessive merging and splitting.
5349  */
5350 static void evict_inode_truncate_pages(struct inode *inode)
5351 {
5352         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5353         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
5354         struct rb_node *node;
5355
5356         ASSERT(inode->i_state & I_FREEING);
5357         truncate_inode_pages_final(&inode->i_data);
5358
5359         write_lock(&map_tree->lock);
5360         while (!RB_EMPTY_ROOT(&map_tree->map.rb_root)) {
5361                 struct extent_map *em;
5362
5363                 node = rb_first_cached(&map_tree->map);
5364                 em = rb_entry(node, struct extent_map, rb_node);
5365                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
5366                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
5367                 remove_extent_mapping(map_tree, em);
5368                 free_extent_map(em);
5369                 if (need_resched()) {
5370                         write_unlock(&map_tree->lock);
5371                         cond_resched();
5372                         write_lock(&map_tree->lock);
5373                 }
5374         }
5375         write_unlock(&map_tree->lock);
5376
5377         /*
5378          * Keep looping until we have no more ranges in the io tree.
5379          * We can have ongoing bios started by readpages (called from readahead)
5380          * that have their endio callback (extent_io.c:end_bio_extent_readpage)
5381          * still in progress (unlocked the pages in the bio but did not yet
5382          * unlocked the ranges in the io tree). Therefore this means some
5383          * ranges can still be locked and eviction started because before
5384          * submitting those bios, which are executed by a separate task (work
5385          * queue kthread), inode references (inode->i_count) were not taken
5386          * (which would be dropped in the end io callback of each bio).
5387          * Therefore here we effectively end up waiting for those bios and
5388          * anyone else holding locked ranges without having bumped the inode's
5389          * reference count - if we don't do it, when they access the inode's
5390          * io_tree to unlock a range it may be too late, leading to an
5391          * use-after-free issue.
5392          */
5393         spin_lock(&io_tree->lock);
5394         while (!RB_EMPTY_ROOT(&io_tree->state)) {
5395                 struct extent_state *state;
5396                 struct extent_state *cached_state = NULL;
5397                 u64 start;
5398                 u64 end;
5399                 unsigned state_flags;
5400
5401                 node = rb_first(&io_tree->state);
5402                 state = rb_entry(node, struct extent_state, rb_node);
5403                 start = state->start;
5404                 end = state->end;
5405                 state_flags = state->state;
5406                 spin_unlock(&io_tree->lock);
5407
5408                 lock_extent_bits(io_tree, start, end, &cached_state);
5409
5410                 /*
5411                  * If still has DELALLOC flag, the extent didn't reach disk,
5412                  * and its reserved space won't be freed by delayed_ref.
5413                  * So we need to free its reserved space here.
5414                  * (Refer to comment in btrfs_invalidatepage, case 2)
5415                  *
5416                  * Note, end is the bytenr of last byte, so we need + 1 here.
5417                  */
5418                 if (state_flags & EXTENT_DELALLOC)
5419                         btrfs_qgroup_free_data(inode, NULL, start, end - start + 1);
5420
5421                 clear_extent_bit(io_tree, start, end,
5422                                  EXTENT_LOCKED | EXTENT_DELALLOC |
5423                                  EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 1, 1,
5424                                  &cached_state);
5425
5426                 cond_resched();
5427                 spin_lock(&io_tree->lock);
5428         }
5429         spin_unlock(&io_tree->lock);
5430 }
5431
5432 static struct btrfs_trans_handle *evict_refill_and_join(struct btrfs_root *root,
5433                                                         struct btrfs_block_rsv *rsv)
5434 {
5435         struct btrfs_fs_info *fs_info = root->fs_info;
5436         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5437         struct btrfs_trans_handle *trans;
5438         u64 delayed_refs_extra = btrfs_calc_insert_metadata_size(fs_info, 1);
5439         int ret;
5440
5441         /*
5442          * Eviction should be taking place at some place safe because of our
5443          * delayed iputs.  However the normal flushing code will run delayed
5444          * iputs, so we cannot use FLUSH_ALL otherwise we'll deadlock.
5445          *
5446          * We reserve the delayed_refs_extra here again because we can't use
5447          * btrfs_start_transaction(root, 0) for the same deadlocky reason as
5448          * above.  We reserve our extra bit here because we generate a ton of
5449          * delayed refs activity by truncating.
5450          *
5451          * If we cannot make our reservation we'll attempt to steal from the
5452          * global reserve, because we really want to be able to free up space.
5453          */
5454         ret = btrfs_block_rsv_refill(root, rsv, rsv->size + delayed_refs_extra,
5455                                      BTRFS_RESERVE_FLUSH_EVICT);
5456         if (ret) {
5457                 /*
5458                  * Try to steal from the global reserve if there is space for
5459                  * it.
5460                  */
5461                 if (btrfs_check_space_for_delayed_refs(fs_info) ||
5462                     btrfs_block_rsv_migrate(global_rsv, rsv, rsv->size, 0)) {
5463                         btrfs_warn(fs_info,
5464                                    "could not allocate space for delete; will truncate on mount");
5465                         return ERR_PTR(-ENOSPC);
5466                 }
5467                 delayed_refs_extra = 0;
5468         }
5469
5470         trans = btrfs_join_transaction(root);
5471         if (IS_ERR(trans))
5472                 return trans;
5473
5474         if (delayed_refs_extra) {
5475                 trans->block_rsv = &fs_info->trans_block_rsv;
5476                 trans->bytes_reserved = delayed_refs_extra;
5477                 btrfs_block_rsv_migrate(rsv, trans->block_rsv,
5478                                         delayed_refs_extra, 1);
5479         }
5480         return trans;
5481 }
5482
5483 void btrfs_evict_inode(struct inode *inode)
5484 {
5485         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5486         struct btrfs_trans_handle *trans;
5487         struct btrfs_root *root = BTRFS_I(inode)->root;
5488         struct btrfs_block_rsv *rsv;
5489         int ret;
5490
5491         trace_btrfs_inode_evict(inode);
5492
5493         if (!root) {
5494                 clear_inode(inode);
5495                 return;
5496         }
5497
5498         evict_inode_truncate_pages(inode);
5499
5500         if (inode->i_nlink &&
5501             ((btrfs_root_refs(&root->root_item) != 0 &&
5502               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5503              btrfs_is_free_space_inode(BTRFS_I(inode))))
5504                 goto no_delete;
5505
5506         if (is_bad_inode(inode))
5507                 goto no_delete;
5508
5509         btrfs_free_io_failure_record(BTRFS_I(inode), 0, (u64)-1);
5510
5511         if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
5512                 goto no_delete;
5513
5514         if (inode->i_nlink > 0) {
5515                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5516                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5517                 goto no_delete;
5518         }
5519
5520         ret = btrfs_commit_inode_delayed_inode(BTRFS_I(inode));
5521         if (ret)
5522                 goto no_delete;
5523
5524         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
5525         if (!rsv)
5526                 goto no_delete;
5527         rsv->size = btrfs_calc_metadata_size(fs_info, 1);
5528         rsv->failfast = 1;
5529
5530         btrfs_i_size_write(BTRFS_I(inode), 0);
5531
5532         while (1) {
5533                 trans = evict_refill_and_join(root, rsv);
5534                 if (IS_ERR(trans))
5535                         goto free_rsv;
5536
5537                 trans->block_rsv = rsv;
5538
5539                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5540                 trans->block_rsv = &fs_info->trans_block_rsv;
5541                 btrfs_end_transaction(trans);
5542                 btrfs_btree_balance_dirty(fs_info);
5543                 if (ret && ret != -ENOSPC && ret != -EAGAIN)
5544                         goto free_rsv;
5545                 else if (!ret)
5546                         break;
5547         }
5548
5549         /*
5550          * Errors here aren't a big deal, it just means we leave orphan items in
5551          * the tree. They will be cleaned up on the next mount. If the inode
5552          * number gets reused, cleanup deletes the orphan item without doing
5553          * anything, and unlink reuses the existing orphan item.
5554          *
5555          * If it turns out that we are dropping too many of these, we might want
5556          * to add a mechanism for retrying these after a commit.
5557          */
5558         trans = evict_refill_and_join(root, rsv);
5559         if (!IS_ERR(trans)) {
5560                 trans->block_rsv = rsv;
5561                 btrfs_orphan_del(trans, BTRFS_I(inode));
5562                 trans->block_rsv = &fs_info->trans_block_rsv;
5563                 btrfs_end_transaction(trans);
5564         }
5565
5566         if (!(root == fs_info->tree_root ||
5567               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5568                 btrfs_return_ino(root, btrfs_ino(BTRFS_I(inode)));
5569
5570 free_rsv:
5571         btrfs_free_block_rsv(fs_info, rsv);
5572 no_delete:
5573         /*
5574          * If we didn't successfully delete, the orphan item will still be in
5575          * the tree and we'll retry on the next mount. Again, we might also want
5576          * to retry these periodically in the future.
5577          */
5578         btrfs_remove_delayed_node(BTRFS_I(inode));
5579         clear_inode(inode);
5580 }
5581
5582 /*
5583  * Return the key found in the dir entry in the location pointer, fill @type
5584  * with BTRFS_FT_*, and return 0.
5585  *
5586  * If no dir entries were found, returns -ENOENT.
5587  * If found a corrupted location in dir entry, returns -EUCLEAN.
5588  */
5589 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5590                                struct btrfs_key *location, u8 *type)
5591 {
5592         const char *name = dentry->d_name.name;
5593         int namelen = dentry->d_name.len;
5594         struct btrfs_dir_item *di;
5595         struct btrfs_path *path;
5596         struct btrfs_root *root = BTRFS_I(dir)->root;
5597         int ret = 0;
5598
5599         path = btrfs_alloc_path();
5600         if (!path)
5601                 return -ENOMEM;
5602
5603         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(BTRFS_I(dir)),
5604                         name, namelen, 0);
5605         if (IS_ERR_OR_NULL(di)) {
5606                 ret = di ? PTR_ERR(di) : -ENOENT;
5607                 goto out;
5608         }
5609
5610         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5611         if (location->type != BTRFS_INODE_ITEM_KEY &&
5612             location->type != BTRFS_ROOT_ITEM_KEY) {
5613                 ret = -EUCLEAN;
5614                 btrfs_warn(root->fs_info,
5615 "%s gets something invalid in DIR_ITEM (name %s, directory ino %llu, location(%llu %u %llu))",
5616                            __func__, name, btrfs_ino(BTRFS_I(dir)),
5617                            location->objectid, location->type, location->offset);
5618         }
5619         if (!ret)
5620                 *type = btrfs_dir_type(path->nodes[0], di);
5621 out:
5622         btrfs_free_path(path);
5623         return ret;
5624 }
5625
5626 /*
5627  * when we hit a tree root in a directory, the btrfs part of the inode
5628  * needs to be changed to reflect the root directory of the tree root.  This
5629  * is kind of like crossing a mount point.
5630  */
5631 static int fixup_tree_root_location(struct btrfs_fs_info *fs_info,
5632                                     struct inode *dir,
5633                                     struct dentry *dentry,
5634                                     struct btrfs_key *location,
5635                                     struct btrfs_root **sub_root)
5636 {
5637         struct btrfs_path *path;
5638         struct btrfs_root *new_root;
5639         struct btrfs_root_ref *ref;
5640         struct extent_buffer *leaf;
5641         struct btrfs_key key;
5642         int ret;
5643         int err = 0;
5644
5645         path = btrfs_alloc_path();
5646         if (!path) {
5647                 err = -ENOMEM;
5648                 goto out;
5649         }
5650
5651         err = -ENOENT;
5652         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5653         key.type = BTRFS_ROOT_REF_KEY;
5654         key.offset = location->objectid;
5655
5656         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
5657         if (ret) {
5658                 if (ret < 0)
5659                         err = ret;
5660                 goto out;
5661         }
5662
5663         leaf = path->nodes[0];
5664         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5665         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(BTRFS_I(dir)) ||
5666             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5667                 goto out;
5668
5669         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5670                                    (unsigned long)(ref + 1),
5671                                    dentry->d_name.len);
5672         if (ret)
5673                 goto out;
5674
5675         btrfs_release_path(path);
5676
5677         new_root = btrfs_read_fs_root_no_name(fs_info, location);
5678         if (IS_ERR(new_root)) {
5679                 err = PTR_ERR(new_root);
5680                 goto out;
5681         }
5682
5683         *sub_root = new_root;
5684         location->objectid = btrfs_root_dirid(&new_root->root_item);
5685         location->type = BTRFS_INODE_ITEM_KEY;
5686         location->offset = 0;
5687         err = 0;
5688 out:
5689         btrfs_free_path(path);
5690         return err;
5691 }
5692
5693 static void inode_tree_add(struct inode *inode)
5694 {
5695         struct btrfs_root *root = BTRFS_I(inode)->root;
5696         struct btrfs_inode *entry;
5697         struct rb_node **p;
5698         struct rb_node *parent;
5699         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5700         u64 ino = btrfs_ino(BTRFS_I(inode));
5701
5702         if (inode_unhashed(inode))
5703                 return;
5704         parent = NULL;
5705         spin_lock(&root->inode_lock);
5706         p = &root->inode_tree.rb_node;
5707         while (*p) {
5708                 parent = *p;
5709                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5710
5711                 if (ino < btrfs_ino(entry))
5712                         p = &parent->rb_left;
5713                 else if (ino > btrfs_ino(entry))
5714                         p = &parent->rb_right;
5715                 else {
5716                         WARN_ON(!(entry->vfs_inode.i_state &
5717                                   (I_WILL_FREE | I_FREEING)));
5718                         rb_replace_node(parent, new, &root->inode_tree);
5719                         RB_CLEAR_NODE(parent);
5720                         spin_unlock(&root->inode_lock);
5721                         return;
5722                 }
5723         }
5724         rb_link_node(new, parent, p);
5725         rb_insert_color(new, &root->inode_tree);
5726         spin_unlock(&root->inode_lock);
5727 }
5728
5729 static void inode_tree_del(struct inode *inode)
5730 {
5731         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5732         struct btrfs_root *root = BTRFS_I(inode)->root;
5733         int empty = 0;
5734
5735         spin_lock(&root->inode_lock);
5736         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5737                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5738                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5739                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5740         }
5741         spin_unlock(&root->inode_lock);
5742
5743         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5744                 synchronize_srcu(&fs_info->subvol_srcu);
5745                 spin_lock(&root->inode_lock);
5746                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5747                 spin_unlock(&root->inode_lock);
5748                 if (empty)
5749                         btrfs_add_dead_root(root);
5750         }
5751 }
5752
5753
5754 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5755 {
5756         struct btrfs_iget_args *args = p;
5757         inode->i_ino = args->location->objectid;
5758         memcpy(&BTRFS_I(inode)->location, args->location,
5759                sizeof(*args->location));
5760         BTRFS_I(inode)->root = args->root;
5761         return 0;
5762 }
5763
5764 static int btrfs_find_actor(struct inode *inode, void *opaque)
5765 {
5766         struct btrfs_iget_args *args = opaque;
5767         return args->location->objectid == BTRFS_I(inode)->location.objectid &&
5768                 args->root == BTRFS_I(inode)->root;
5769 }
5770
5771 static struct inode *btrfs_iget_locked(struct super_block *s,
5772                                        struct btrfs_key *location,
5773                                        struct btrfs_root *root)
5774 {
5775         struct inode *inode;
5776         struct btrfs_iget_args args;
5777         unsigned long hashval = btrfs_inode_hash(location->objectid, root);
5778
5779         args.location = location;
5780         args.root = root;
5781
5782         inode = iget5_locked(s, hashval, btrfs_find_actor,
5783                              btrfs_init_locked_inode,
5784                              (void *)&args);
5785         return inode;
5786 }
5787
5788 /*
5789  * Get an inode object given its location and corresponding root.
5790  * Path can be preallocated to prevent recursing back to iget through
5791  * allocator. NULL is also valid but may require an additional allocation
5792  * later.
5793  */
5794 struct inode *btrfs_iget_path(struct super_block *s, struct btrfs_key *location,
5795                               struct btrfs_root *root, struct btrfs_path *path)
5796 {
5797         struct inode *inode;
5798
5799         inode = btrfs_iget_locked(s, location, root);
5800         if (!inode)
5801                 return ERR_PTR(-ENOMEM);
5802
5803         if (inode->i_state & I_NEW) {
5804                 int ret;
5805
5806                 ret = btrfs_read_locked_inode(inode, path);
5807                 if (!ret) {
5808                         inode_tree_add(inode);
5809                         unlock_new_inode(inode);
5810                 } else {
5811                         iget_failed(inode);
5812                         /*
5813                          * ret > 0 can come from btrfs_search_slot called by
5814                          * btrfs_read_locked_inode, this means the inode item
5815                          * was not found.
5816                          */
5817                         if (ret > 0)
5818                                 ret = -ENOENT;
5819                         inode = ERR_PTR(ret);
5820                 }
5821         }
5822
5823         return inode;
5824 }
5825
5826 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
5827                          struct btrfs_root *root)
5828 {
5829         return btrfs_iget_path(s, location, root, NULL);
5830 }
5831
5832 static struct inode *new_simple_dir(struct super_block *s,
5833                                     struct btrfs_key *key,
5834                                     struct btrfs_root *root)
5835 {
5836         struct inode *inode = new_inode(s);
5837
5838         if (!inode)
5839                 return ERR_PTR(-ENOMEM);
5840
5841         BTRFS_I(inode)->root = root;
5842         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5843         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5844
5845         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5846         inode->i_op = &btrfs_dir_ro_inode_operations;
5847         inode->i_opflags &= ~IOP_XATTR;
5848         inode->i_fop = &simple_dir_operations;
5849         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5850         inode->i_mtime = current_time(inode);
5851         inode->i_atime = inode->i_mtime;
5852         inode->i_ctime = inode->i_mtime;
5853         BTRFS_I(inode)->i_otime = inode->i_mtime;
5854
5855         return inode;
5856 }
5857
5858 static inline u8 btrfs_inode_type(struct inode *inode)
5859 {
5860         /*
5861          * Compile-time asserts that generic FT_* types still match
5862          * BTRFS_FT_* types
5863          */
5864         BUILD_BUG_ON(BTRFS_FT_UNKNOWN != FT_UNKNOWN);
5865         BUILD_BUG_ON(BTRFS_FT_REG_FILE != FT_REG_FILE);
5866         BUILD_BUG_ON(BTRFS_FT_DIR != FT_DIR);
5867         BUILD_BUG_ON(BTRFS_FT_CHRDEV != FT_CHRDEV);
5868         BUILD_BUG_ON(BTRFS_FT_BLKDEV != FT_BLKDEV);
5869         BUILD_BUG_ON(BTRFS_FT_FIFO != FT_FIFO);
5870         BUILD_BUG_ON(BTRFS_FT_SOCK != FT_SOCK);
5871         BUILD_BUG_ON(BTRFS_FT_SYMLINK != FT_SYMLINK);
5872
5873         return fs_umode_to_ftype(inode->i_mode);
5874 }
5875
5876 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5877 {
5878         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
5879         struct inode *inode;
5880         struct btrfs_root *root = BTRFS_I(dir)->root;
5881         struct btrfs_root *sub_root = root;
5882         struct btrfs_key location;
5883         u8 di_type = 0;
5884         int index;
5885         int ret = 0;
5886
5887         if (dentry->d_name.len > BTRFS_NAME_LEN)
5888                 return ERR_PTR(-ENAMETOOLONG);
5889
5890         ret = btrfs_inode_by_name(dir, dentry, &location, &di_type);
5891         if (ret < 0)
5892                 return ERR_PTR(ret);
5893
5894         if (location.type == BTRFS_INODE_ITEM_KEY) {
5895                 inode = btrfs_iget(dir->i_sb, &location, root);
5896                 if (IS_ERR(inode))
5897                         return inode;
5898
5899                 /* Do extra check against inode mode with di_type */
5900                 if (btrfs_inode_type(inode) != di_type) {
5901                         btrfs_crit(fs_info,
5902 "inode mode mismatch with dir: inode mode=0%o btrfs type=%u dir type=%u",
5903                                   inode->i_mode, btrfs_inode_type(inode),
5904                                   di_type);
5905                         iput(inode);
5906                         return ERR_PTR(-EUCLEAN);
5907                 }
5908                 return inode;
5909         }
5910
5911         index = srcu_read_lock(&fs_info->subvol_srcu);
5912         ret = fixup_tree_root_location(fs_info, dir, dentry,
5913                                        &location, &sub_root);
5914         if (ret < 0) {
5915                 if (ret != -ENOENT)
5916                         inode = ERR_PTR(ret);
5917                 else
5918                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5919         } else {
5920                 inode = btrfs_iget(dir->i_sb, &location, sub_root);
5921         }
5922         srcu_read_unlock(&fs_info->subvol_srcu, index);
5923
5924         if (!IS_ERR(inode) && root != sub_root) {
5925                 down_read(&fs_info->cleanup_work_sem);
5926                 if (!sb_rdonly(inode->i_sb))
5927                         ret = btrfs_orphan_cleanup(sub_root);
5928                 up_read(&fs_info->cleanup_work_sem);
5929                 if (ret) {
5930                         iput(inode);
5931                         inode = ERR_PTR(ret);
5932                 }
5933         }
5934
5935         return inode;
5936 }
5937
5938 static int btrfs_dentry_delete(const struct dentry *dentry)
5939 {
5940         struct btrfs_root *root;
5941         struct inode *inode = d_inode(dentry);
5942
5943         if (!inode && !IS_ROOT(dentry))
5944                 inode = d_inode(dentry->d_parent);
5945
5946         if (inode) {
5947                 root = BTRFS_I(inode)->root;
5948                 if (btrfs_root_refs(&root->root_item) == 0)
5949                         return 1;
5950
5951                 if (btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5952                         return 1;
5953         }
5954         return 0;
5955 }
5956
5957 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5958                                    unsigned int flags)
5959 {
5960         struct inode *inode = btrfs_lookup_dentry(dir, dentry);
5961
5962         if (inode == ERR_PTR(-ENOENT))
5963                 inode = NULL;
5964         return d_splice_alias(inode, dentry);
5965 }
5966
5967 /*
5968  * All this infrastructure exists because dir_emit can fault, and we are holding
5969  * the tree lock when doing readdir.  For now just allocate a buffer and copy
5970  * our information into that, and then dir_emit from the buffer.  This is
5971  * similar to what NFS does, only we don't keep the buffer around in pagecache
5972  * because I'm afraid I'll mess that up.  Long term we need to make filldir do
5973  * copy_to_user_inatomic so we don't have to worry about page faulting under the
5974  * tree lock.
5975  */
5976 static int btrfs_opendir(struct inode *inode, struct file *file)
5977 {
5978         struct btrfs_file_private *private;
5979
5980         private = kzalloc(sizeof(struct btrfs_file_private), GFP_KERNEL);
5981         if (!private)
5982                 return -ENOMEM;
5983         private->filldir_buf = kzalloc(PAGE_SIZE, GFP_KERNEL);
5984         if (!private->filldir_buf) {
5985                 kfree(private);
5986                 return -ENOMEM;
5987         }
5988         file->private_data = private;
5989         return 0;
5990 }
5991
5992 struct dir_entry {
5993         u64 ino;
5994         u64 offset;
5995         unsigned type;
5996         int name_len;
5997 };
5998
5999 static int btrfs_filldir(void *addr, int entries, struct dir_context *ctx)
6000 {
6001         while (entries--) {
6002                 struct dir_entry *entry = addr;
6003                 char *name = (char *)(entry + 1);
6004
6005                 ctx->pos = get_unaligned(&entry->offset);
6006                 if (!dir_emit(ctx, name, get_unaligned(&entry->name_len),
6007                                          get_unaligned(&entry->ino),
6008                                          get_unaligned(&entry->type)))
6009                         return 1;
6010                 addr += sizeof(struct dir_entry) +
6011                         get_unaligned(&entry->name_len);
6012                 ctx->pos++;
6013         }
6014         return 0;
6015 }
6016
6017 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
6018 {
6019         struct inode *inode = file_inode(file);
6020         struct btrfs_root *root = BTRFS_I(inode)->root;
6021         struct btrfs_file_private *private = file->private_data;
6022         struct btrfs_dir_item *di;
6023         struct btrfs_key key;
6024         struct btrfs_key found_key;
6025         struct btrfs_path *path;
6026         void *addr;
6027         struct list_head ins_list;
6028         struct list_head del_list;
6029         int ret;
6030         struct extent_buffer *leaf;
6031         int slot;
6032         char *name_ptr;
6033         int name_len;
6034         int entries = 0;
6035         int total_len = 0;
6036         bool put = false;
6037         struct btrfs_key location;
6038
6039         if (!dir_emit_dots(file, ctx))
6040                 return 0;
6041
6042         path = btrfs_alloc_path();
6043         if (!path)
6044                 return -ENOMEM;
6045
6046         addr = private->filldir_buf;
6047         path->reada = READA_FORWARD;
6048
6049         INIT_LIST_HEAD(&ins_list);
6050         INIT_LIST_HEAD(&del_list);
6051         put = btrfs_readdir_get_delayed_items(inode, &ins_list, &del_list);
6052
6053 again:
6054         key.type = BTRFS_DIR_INDEX_KEY;
6055         key.offset = ctx->pos;
6056         key.objectid = btrfs_ino(BTRFS_I(inode));
6057
6058         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6059         if (ret < 0)
6060                 goto err;
6061
6062         while (1) {
6063                 struct dir_entry *entry;
6064
6065                 leaf = path->nodes[0];
6066                 slot = path->slots[0];
6067                 if (slot >= btrfs_header_nritems(leaf)) {
6068                         ret = btrfs_next_leaf(root, path);
6069                         if (ret < 0)
6070                                 goto err;
6071                         else if (ret > 0)
6072                                 break;
6073                         continue;
6074                 }
6075
6076                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
6077
6078                 if (found_key.objectid != key.objectid)
6079                         break;
6080                 if (found_key.type != BTRFS_DIR_INDEX_KEY)
6081                         break;
6082                 if (found_key.offset < ctx->pos)
6083                         goto next;
6084                 if (btrfs_should_delete_dir_index(&del_list, found_key.offset))
6085                         goto next;
6086                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
6087                 name_len = btrfs_dir_name_len(leaf, di);
6088                 if ((total_len + sizeof(struct dir_entry) + name_len) >=
6089                     PAGE_SIZE) {
6090                         btrfs_release_path(path);
6091                         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
6092                         if (ret)
6093                                 goto nopos;
6094                         addr = private->filldir_buf;
6095                         entries = 0;
6096                         total_len = 0;
6097                         goto again;
6098                 }
6099
6100                 entry = addr;
6101                 put_unaligned(name_len, &entry->name_len);
6102                 name_ptr = (char *)(entry + 1);
6103                 read_extent_buffer(leaf, name_ptr, (unsigned long)(di + 1),
6104                                    name_len);
6105                 put_unaligned(fs_ftype_to_dtype(btrfs_dir_type(leaf, di)),
6106                                 &entry->type);
6107                 btrfs_dir_item_key_to_cpu(leaf, di, &location);
6108                 put_unaligned(location.objectid, &entry->ino);
6109                 put_unaligned(found_key.offset, &entry->offset);
6110                 entries++;
6111                 addr += sizeof(struct dir_entry) + name_len;
6112                 total_len += sizeof(struct dir_entry) + name_len;
6113 next:
6114                 path->slots[0]++;
6115         }
6116         btrfs_release_path(path);
6117
6118         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
6119         if (ret)
6120                 goto nopos;
6121
6122         ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list);
6123         if (ret)
6124                 goto nopos;
6125
6126         /*
6127          * Stop new entries from being returned after we return the last
6128          * entry.
6129          *
6130          * New directory entries are assigned a strictly increasing
6131          * offset.  This means that new entries created during readdir
6132          * are *guaranteed* to be seen in the future by that readdir.
6133          * This has broken buggy programs which operate on names as
6134          * they're returned by readdir.  Until we re-use freed offsets
6135          * we have this hack to stop new entries from being returned
6136          * under the assumption that they'll never reach this huge
6137          * offset.
6138          *
6139          * This is being careful not to overflow 32bit loff_t unless the
6140          * last entry requires it because doing so has broken 32bit apps
6141          * in the past.
6142          */
6143         if (ctx->pos >= INT_MAX)
6144                 ctx->pos = LLONG_MAX;
6145         else
6146                 ctx->pos = INT_MAX;
6147 nopos:
6148         ret = 0;
6149 err:
6150         if (put)
6151                 btrfs_readdir_put_delayed_items(inode, &ins_list, &del_list);
6152         btrfs_free_path(path);
6153         return ret;
6154 }
6155
6156 /*
6157  * This is somewhat expensive, updating the tree every time the
6158  * inode changes.  But, it is most likely to find the inode in cache.
6159  * FIXME, needs more benchmarking...there are no reasons other than performance
6160  * to keep or drop this code.
6161  */
6162 static int btrfs_dirty_inode(struct inode *inode)
6163 {
6164         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6165         struct btrfs_root *root = BTRFS_I(inode)->root;
6166         struct btrfs_trans_handle *trans;
6167         int ret;
6168
6169         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
6170                 return 0;
6171
6172         trans = btrfs_join_transaction(root);
6173         if (IS_ERR(trans))
6174                 return PTR_ERR(trans);
6175
6176         ret = btrfs_update_inode(trans, root, inode);
6177         if (ret && ret == -ENOSPC) {
6178                 /* whoops, lets try again with the full transaction */
6179                 btrfs_end_transaction(trans);
6180                 trans = btrfs_start_transaction(root, 1);
6181                 if (IS_ERR(trans))
6182                         return PTR_ERR(trans);
6183
6184                 ret = btrfs_update_inode(trans, root, inode);
6185         }
6186         btrfs_end_transaction(trans);
6187         if (BTRFS_I(inode)->delayed_node)
6188                 btrfs_balance_delayed_items(fs_info);
6189
6190         return ret;
6191 }
6192
6193 /*
6194  * This is a copy of file_update_time.  We need this so we can return error on
6195  * ENOSPC for updating the inode in the case of file write and mmap writes.
6196  */
6197 static int btrfs_update_time(struct inode *inode, struct timespec64 *now,
6198                              int flags)
6199 {
6200         struct btrfs_root *root = BTRFS_I(inode)->root;
6201         bool dirty = flags & ~S_VERSION;
6202
6203         if (btrfs_root_readonly(root))
6204                 return -EROFS;
6205
6206         if (flags & S_VERSION)
6207                 dirty |= inode_maybe_inc_iversion(inode, dirty);
6208         if (flags & S_CTIME)
6209                 inode->i_ctime = *now;
6210         if (flags & S_MTIME)
6211                 inode->i_mtime = *now;
6212         if (flags & S_ATIME)
6213                 inode->i_atime = *now;
6214         return dirty ? btrfs_dirty_inode(inode) : 0;
6215 }
6216
6217 /*
6218  * find the highest existing sequence number in a directory
6219  * and then set the in-memory index_cnt variable to reflect
6220  * free sequence numbers
6221  */
6222 static int btrfs_set_inode_index_count(struct btrfs_inode *inode)
6223 {
6224         struct btrfs_root *root = inode->root;
6225         struct btrfs_key key, found_key;
6226         struct btrfs_path *path;
6227         struct extent_buffer *leaf;
6228         int ret;
6229
6230         key.objectid = btrfs_ino(inode);
6231         key.type = BTRFS_DIR_INDEX_KEY;
6232         key.offset = (u64)-1;
6233
6234         path = btrfs_alloc_path();
6235         if (!path)
6236                 return -ENOMEM;
6237
6238         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6239         if (ret < 0)
6240                 goto out;
6241         /* FIXME: we should be able to handle this */
6242         if (ret == 0)
6243                 goto out;
6244         ret = 0;
6245
6246         /*
6247          * MAGIC NUMBER EXPLANATION:
6248          * since we search a directory based on f_pos we have to start at 2
6249          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
6250          * else has to start at 2
6251          */
6252         if (path->slots[0] == 0) {
6253                 inode->index_cnt = 2;
6254                 goto out;
6255         }
6256
6257         path->slots[0]--;
6258
6259         leaf = path->nodes[0];
6260         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6261
6262         if (found_key.objectid != btrfs_ino(inode) ||
6263             found_key.type != BTRFS_DIR_INDEX_KEY) {
6264                 inode->index_cnt = 2;
6265                 goto out;
6266         }
6267
6268         inode->index_cnt = found_key.offset + 1;
6269 out:
6270         btrfs_free_path(path);
6271         return ret;
6272 }
6273
6274 /*
6275  * helper to find a free sequence number in a given directory.  This current
6276  * code is very simple, later versions will do smarter things in the btree
6277  */
6278 int btrfs_set_inode_index(struct btrfs_inode *dir, u64 *index)
6279 {
6280         int ret = 0;
6281
6282         if (dir->index_cnt == (u64)-1) {
6283                 ret = btrfs_inode_delayed_dir_index_count(dir);
6284                 if (ret) {
6285                         ret = btrfs_set_inode_index_count(dir);
6286                         if (ret)
6287                                 return ret;
6288                 }
6289         }
6290
6291         *index = dir->index_cnt;
6292         dir->index_cnt++;
6293
6294         return ret;
6295 }
6296
6297 static int btrfs_insert_inode_locked(struct inode *inode)
6298 {
6299         struct btrfs_iget_args args;
6300         args.location = &BTRFS_I(inode)->location;
6301         args.root = BTRFS_I(inode)->root;
6302
6303         return insert_inode_locked4(inode,
6304                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
6305                    btrfs_find_actor, &args);
6306 }
6307
6308 /*
6309  * Inherit flags from the parent inode.
6310  *
6311  * Currently only the compression flags and the cow flags are inherited.
6312  */
6313 static void btrfs_inherit_iflags(struct inode *inode, struct inode *dir)
6314 {
6315         unsigned int flags;
6316
6317         if (!dir)
6318                 return;
6319
6320         flags = BTRFS_I(dir)->flags;
6321
6322         if (flags & BTRFS_INODE_NOCOMPRESS) {
6323                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_COMPRESS;
6324                 BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
6325         } else if (flags & BTRFS_INODE_COMPRESS) {
6326                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_NOCOMPRESS;
6327                 BTRFS_I(inode)->flags |= BTRFS_INODE_COMPRESS;
6328         }
6329
6330         if (flags & BTRFS_INODE_NODATACOW) {
6331                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW;
6332                 if (S_ISREG(inode->i_mode))
6333                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6334         }
6335
6336         btrfs_sync_inode_flags_to_i_flags(inode);
6337 }
6338
6339 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
6340                                      struct btrfs_root *root,
6341                                      struct inode *dir,
6342                                      const char *name, int name_len,
6343                                      u64 ref_objectid, u64 objectid,
6344                                      umode_t mode, u64 *index)
6345 {
6346         struct btrfs_fs_info *fs_info = root->fs_info;
6347         struct inode *inode;
6348         struct btrfs_inode_item *inode_item;
6349         struct btrfs_key *location;
6350         struct btrfs_path *path;
6351         struct btrfs_inode_ref *ref;
6352         struct btrfs_key key[2];
6353         u32 sizes[2];
6354         int nitems = name ? 2 : 1;
6355         unsigned long ptr;
6356         unsigned int nofs_flag;
6357         int ret;
6358
6359         path = btrfs_alloc_path();
6360         if (!path)
6361                 return ERR_PTR(-ENOMEM);
6362
6363         nofs_flag = memalloc_nofs_save();
6364         inode = new_inode(fs_info->sb);
6365         memalloc_nofs_restore(nofs_flag);
6366         if (!inode) {
6367                 btrfs_free_path(path);
6368                 return ERR_PTR(-ENOMEM);
6369         }
6370
6371         /*
6372          * O_TMPFILE, set link count to 0, so that after this point,
6373          * we fill in an inode item with the correct link count.
6374          */
6375         if (!name)
6376                 set_nlink(inode, 0);
6377
6378         /*
6379          * we have to initialize this early, so we can reclaim the inode
6380          * number if we fail afterwards in this function.
6381          */
6382         inode->i_ino = objectid;
6383
6384         if (dir && name) {
6385                 trace_btrfs_inode_request(dir);
6386
6387                 ret = btrfs_set_inode_index(BTRFS_I(dir), index);
6388                 if (ret) {
6389                         btrfs_free_path(path);
6390                         iput(inode);
6391                         return ERR_PTR(ret);
6392                 }
6393         } else if (dir) {
6394                 *index = 0;
6395         }
6396         /*
6397          * index_cnt is ignored for everything but a dir,
6398          * btrfs_set_inode_index_count has an explanation for the magic
6399          * number
6400          */
6401         BTRFS_I(inode)->index_cnt = 2;
6402         BTRFS_I(inode)->dir_index = *index;
6403         BTRFS_I(inode)->root = root;
6404         BTRFS_I(inode)->generation = trans->transid;
6405         inode->i_generation = BTRFS_I(inode)->generation;
6406
6407         /*
6408          * We could have gotten an inode number from somebody who was fsynced
6409          * and then removed in this same transaction, so let's just set full
6410          * sync since it will be a full sync anyway and this will blow away the
6411          * old info in the log.
6412          */
6413         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
6414
6415         key[0].objectid = objectid;
6416         key[0].type = BTRFS_INODE_ITEM_KEY;
6417         key[0].offset = 0;
6418
6419         sizes[0] = sizeof(struct btrfs_inode_item);
6420
6421         if (name) {
6422                 /*
6423                  * Start new inodes with an inode_ref. This is slightly more
6424                  * efficient for small numbers of hard links since they will
6425                  * be packed into one item. Extended refs will kick in if we
6426                  * add more hard links than can fit in the ref item.
6427                  */
6428                 key[1].objectid = objectid;
6429                 key[1].type = BTRFS_INODE_REF_KEY;
6430                 key[1].offset = ref_objectid;
6431
6432                 sizes[1] = name_len + sizeof(*ref);
6433         }
6434
6435         location = &BTRFS_I(inode)->location;
6436         location->objectid = objectid;
6437         location->offset = 0;
6438         location->type = BTRFS_INODE_ITEM_KEY;
6439
6440         ret = btrfs_insert_inode_locked(inode);
6441         if (ret < 0) {
6442                 iput(inode);
6443                 goto fail;
6444         }
6445
6446         path->leave_spinning = 1;
6447         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
6448         if (ret != 0)
6449                 goto fail_unlock;
6450
6451         inode_init_owner(inode, dir, mode);
6452         inode_set_bytes(inode, 0);
6453
6454         inode->i_mtime = current_time(inode);
6455         inode->i_atime = inode->i_mtime;
6456         inode->i_ctime = inode->i_mtime;
6457         BTRFS_I(inode)->i_otime = inode->i_mtime;
6458
6459         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
6460                                   struct btrfs_inode_item);
6461         memzero_extent_buffer(path->nodes[0], (unsigned long)inode_item,
6462                              sizeof(*inode_item));
6463         fill_inode_item(trans, path->nodes[0], inode_item, inode);
6464
6465         if (name) {
6466                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
6467                                      struct btrfs_inode_ref);
6468                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6469                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6470                 ptr = (unsigned long)(ref + 1);
6471                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6472         }
6473
6474         btrfs_mark_buffer_dirty(path->nodes[0]);
6475         btrfs_free_path(path);
6476
6477         btrfs_inherit_iflags(inode, dir);
6478
6479         if (S_ISREG(mode)) {
6480                 if (btrfs_test_opt(fs_info, NODATASUM))
6481                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6482                 if (btrfs_test_opt(fs_info, NODATACOW))
6483                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6484                                 BTRFS_INODE_NODATASUM;
6485         }
6486
6487         inode_tree_add(inode);
6488
6489         trace_btrfs_inode_new(inode);
6490         btrfs_set_inode_last_trans(trans, inode);
6491
6492         btrfs_update_root_times(trans, root);
6493
6494         ret = btrfs_inode_inherit_props(trans, inode, dir);
6495         if (ret)
6496                 btrfs_err(fs_info,
6497                           "error inheriting props for ino %llu (root %llu): %d",
6498                         btrfs_ino(BTRFS_I(inode)), root->root_key.objectid, ret);
6499
6500         return inode;
6501
6502 fail_unlock:
6503         discard_new_inode(inode);
6504 fail:
6505         if (dir && name)
6506                 BTRFS_I(dir)->index_cnt--;
6507         btrfs_free_path(path);
6508         return ERR_PTR(ret);
6509 }
6510
6511 /*
6512  * utility function to add 'inode' into 'parent_inode' with
6513  * a give name and a given sequence number.
6514  * if 'add_backref' is true, also insert a backref from the
6515  * inode to the parent directory.
6516  */
6517 int btrfs_add_link(struct btrfs_trans_handle *trans,
6518                    struct btrfs_inode *parent_inode, struct btrfs_inode *inode,
6519                    const char *name, int name_len, int add_backref, u64 index)
6520 {
6521         int ret = 0;
6522         struct btrfs_key key;
6523         struct btrfs_root *root = parent_inode->root;
6524         u64 ino = btrfs_ino(inode);
6525         u64 parent_ino = btrfs_ino(parent_inode);
6526
6527         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6528                 memcpy(&key, &inode->root->root_key, sizeof(key));
6529         } else {
6530                 key.objectid = ino;
6531                 key.type = BTRFS_INODE_ITEM_KEY;
6532                 key.offset = 0;
6533         }
6534
6535         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6536                 ret = btrfs_add_root_ref(trans, key.objectid,
6537                                          root->root_key.objectid, parent_ino,
6538                                          index, name, name_len);
6539         } else if (add_backref) {
6540                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6541                                              parent_ino, index);
6542         }
6543
6544         /* Nothing to clean up yet */
6545         if (ret)
6546                 return ret;
6547
6548         ret = btrfs_insert_dir_item(trans, name, name_len, parent_inode, &key,
6549                                     btrfs_inode_type(&inode->vfs_inode), index);
6550         if (ret == -EEXIST || ret == -EOVERFLOW)
6551                 goto fail_dir_item;
6552         else if (ret) {
6553                 btrfs_abort_transaction(trans, ret);
6554                 return ret;
6555         }
6556
6557         btrfs_i_size_write(parent_inode, parent_inode->vfs_inode.i_size +
6558                            name_len * 2);
6559         inode_inc_iversion(&parent_inode->vfs_inode);
6560         /*
6561          * If we are replaying a log tree, we do not want to update the mtime
6562          * and ctime of the parent directory with the current time, since the
6563          * log replay procedure is responsible for setting them to their correct
6564          * values (the ones it had when the fsync was done).
6565          */
6566         if (!test_bit(BTRFS_FS_LOG_RECOVERING, &root->fs_info->flags)) {
6567                 struct timespec64 now = current_time(&parent_inode->vfs_inode);
6568
6569                 parent_inode->vfs_inode.i_mtime = now;
6570                 parent_inode->vfs_inode.i_ctime = now;
6571         }
6572         ret = btrfs_update_inode(trans, root, &parent_inode->vfs_inode);
6573         if (ret)
6574                 btrfs_abort_transaction(trans, ret);
6575         return ret;
6576
6577 fail_dir_item:
6578         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6579                 u64 local_index;
6580                 int err;
6581                 err = btrfs_del_root_ref(trans, key.objectid,
6582                                          root->root_key.objectid, parent_ino,
6583                                          &local_index, name, name_len);
6584                 if (err)
6585                         btrfs_abort_transaction(trans, err);
6586         } else if (add_backref) {
6587                 u64 local_index;
6588                 int err;
6589
6590                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6591                                           ino, parent_ino, &local_index);
6592                 if (err)
6593                         btrfs_abort_transaction(trans, err);
6594         }
6595
6596         /* Return the original error code */
6597         return ret;
6598 }
6599
6600 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6601                             struct btrfs_inode *dir, struct dentry *dentry,
6602                             struct btrfs_inode *inode, int backref, u64 index)
6603 {
6604         int err = btrfs_add_link(trans, dir, inode,
6605                                  dentry->d_name.name, dentry->d_name.len,
6606                                  backref, index);
6607         if (err > 0)
6608                 err = -EEXIST;
6609         return err;
6610 }
6611
6612 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6613                         umode_t mode, dev_t rdev)
6614 {
6615         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6616         struct btrfs_trans_handle *trans;
6617         struct btrfs_root *root = BTRFS_I(dir)->root;
6618         struct inode *inode = NULL;
6619         int err;
6620         u64 objectid;
6621         u64 index = 0;
6622
6623         /*
6624          * 2 for inode item and ref
6625          * 2 for dir items
6626          * 1 for xattr if selinux is on
6627          */
6628         trans = btrfs_start_transaction(root, 5);
6629         if (IS_ERR(trans))
6630                 return PTR_ERR(trans);
6631
6632         err = btrfs_find_free_ino(root, &objectid);
6633         if (err)
6634                 goto out_unlock;
6635
6636         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6637                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6638                         mode, &index);
6639         if (IS_ERR(inode)) {
6640                 err = PTR_ERR(inode);
6641                 inode = NULL;
6642                 goto out_unlock;
6643         }
6644
6645         /*
6646         * If the active LSM wants to access the inode during
6647         * d_instantiate it needs these. Smack checks to see
6648         * if the filesystem supports xattrs by looking at the
6649         * ops vector.
6650         */
6651         inode->i_op = &btrfs_special_inode_operations;
6652         init_special_inode(inode, inode->i_mode, rdev);
6653
6654         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6655         if (err)
6656                 goto out_unlock;
6657
6658         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6659                         0, index);
6660         if (err)
6661                 goto out_unlock;
6662
6663         btrfs_update_inode(trans, root, inode);
6664         d_instantiate_new(dentry, inode);
6665
6666 out_unlock:
6667         btrfs_end_transaction(trans);
6668         btrfs_btree_balance_dirty(fs_info);
6669         if (err && inode) {
6670                 inode_dec_link_count(inode);
6671                 discard_new_inode(inode);
6672         }
6673         return err;
6674 }
6675
6676 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6677                         umode_t mode, bool excl)
6678 {
6679         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6680         struct btrfs_trans_handle *trans;
6681         struct btrfs_root *root = BTRFS_I(dir)->root;
6682         struct inode *inode = NULL;
6683         int err;
6684         u64 objectid;
6685         u64 index = 0;
6686
6687         /*
6688          * 2 for inode item and ref
6689          * 2 for dir items
6690          * 1 for xattr if selinux is on
6691          */
6692         trans = btrfs_start_transaction(root, 5);
6693         if (IS_ERR(trans))
6694                 return PTR_ERR(trans);
6695
6696         err = btrfs_find_free_ino(root, &objectid);
6697         if (err)
6698                 goto out_unlock;
6699
6700         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6701                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6702                         mode, &index);
6703         if (IS_ERR(inode)) {
6704                 err = PTR_ERR(inode);
6705                 inode = NULL;
6706                 goto out_unlock;
6707         }
6708         /*
6709         * If the active LSM wants to access the inode during
6710         * d_instantiate it needs these. Smack checks to see
6711         * if the filesystem supports xattrs by looking at the
6712         * ops vector.
6713         */
6714         inode->i_fop = &btrfs_file_operations;
6715         inode->i_op = &btrfs_file_inode_operations;
6716         inode->i_mapping->a_ops = &btrfs_aops;
6717
6718         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6719         if (err)
6720                 goto out_unlock;
6721
6722         err = btrfs_update_inode(trans, root, inode);
6723         if (err)
6724                 goto out_unlock;
6725
6726         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6727                         0, index);
6728         if (err)
6729                 goto out_unlock;
6730
6731         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6732         d_instantiate_new(dentry, inode);
6733
6734 out_unlock:
6735         btrfs_end_transaction(trans);
6736         if (err && inode) {
6737                 inode_dec_link_count(inode);
6738                 discard_new_inode(inode);
6739         }
6740         btrfs_btree_balance_dirty(fs_info);
6741         return err;
6742 }
6743
6744 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6745                       struct dentry *dentry)
6746 {
6747         struct btrfs_trans_handle *trans = NULL;
6748         struct btrfs_root *root = BTRFS_I(dir)->root;
6749         struct inode *inode = d_inode(old_dentry);
6750         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6751         u64 index;
6752         int err;
6753         int drop_inode = 0;
6754
6755         /* do not allow sys_link's with other subvols of the same device */
6756         if (root->root_key.objectid != BTRFS_I(inode)->root->root_key.objectid)
6757                 return -EXDEV;
6758
6759         if (inode->i_nlink >= BTRFS_LINK_MAX)
6760                 return -EMLINK;
6761
6762         err = btrfs_set_inode_index(BTRFS_I(dir), &index);
6763         if (err)
6764                 goto fail;
6765
6766         /*
6767          * 2 items for inode and inode ref
6768          * 2 items for dir items
6769          * 1 item for parent inode
6770          * 1 item for orphan item deletion if O_TMPFILE
6771          */
6772         trans = btrfs_start_transaction(root, inode->i_nlink ? 5 : 6);
6773         if (IS_ERR(trans)) {
6774                 err = PTR_ERR(trans);
6775                 trans = NULL;
6776                 goto fail;
6777         }
6778
6779         /* There are several dir indexes for this inode, clear the cache. */
6780         BTRFS_I(inode)->dir_index = 0ULL;
6781         inc_nlink(inode);
6782         inode_inc_iversion(inode);
6783         inode->i_ctime = current_time(inode);
6784         ihold(inode);
6785         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6786
6787         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6788                         1, index);
6789
6790         if (err) {
6791                 drop_inode = 1;
6792         } else {
6793                 struct dentry *parent = dentry->d_parent;
6794                 int ret;
6795
6796                 err = btrfs_update_inode(trans, root, inode);
6797                 if (err)
6798                         goto fail;
6799                 if (inode->i_nlink == 1) {
6800                         /*
6801                          * If new hard link count is 1, it's a file created
6802                          * with open(2) O_TMPFILE flag.
6803                          */
6804                         err = btrfs_orphan_del(trans, BTRFS_I(inode));
6805                         if (err)
6806                                 goto fail;
6807                 }
6808                 d_instantiate(dentry, inode);
6809                 ret = btrfs_log_new_name(trans, BTRFS_I(inode), NULL, parent,
6810                                          true, NULL);
6811                 if (ret == BTRFS_NEED_TRANS_COMMIT) {
6812                         err = btrfs_commit_transaction(trans);
6813                         trans = NULL;
6814                 }
6815         }
6816
6817 fail:
6818         if (trans)
6819                 btrfs_end_transaction(trans);
6820         if (drop_inode) {
6821                 inode_dec_link_count(inode);
6822                 iput(inode);
6823         }
6824         btrfs_btree_balance_dirty(fs_info);
6825         return err;
6826 }
6827
6828 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6829 {
6830         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6831         struct inode *inode = NULL;
6832         struct btrfs_trans_handle *trans;
6833         struct btrfs_root *root = BTRFS_I(dir)->root;
6834         int err = 0;
6835         u64 objectid = 0;
6836         u64 index = 0;
6837
6838         /*
6839          * 2 items for inode and ref
6840          * 2 items for dir items
6841          * 1 for xattr if selinux is on
6842          */
6843         trans = btrfs_start_transaction(root, 5);
6844         if (IS_ERR(trans))
6845                 return PTR_ERR(trans);
6846
6847         err = btrfs_find_free_ino(root, &objectid);
6848         if (err)
6849                 goto out_fail;
6850
6851         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6852                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6853                         S_IFDIR | mode, &index);
6854         if (IS_ERR(inode)) {
6855                 err = PTR_ERR(inode);
6856                 inode = NULL;
6857                 goto out_fail;
6858         }
6859
6860         /* these must be set before we unlock the inode */
6861         inode->i_op = &btrfs_dir_inode_operations;
6862         inode->i_fop = &btrfs_dir_file_operations;
6863
6864         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6865         if (err)
6866                 goto out_fail;
6867
6868         btrfs_i_size_write(BTRFS_I(inode), 0);
6869         err = btrfs_update_inode(trans, root, inode);
6870         if (err)
6871                 goto out_fail;
6872
6873         err = btrfs_add_link(trans, BTRFS_I(dir), BTRFS_I(inode),
6874                         dentry->d_name.name,
6875                         dentry->d_name.len, 0, index);
6876         if (err)
6877                 goto out_fail;
6878
6879         d_instantiate_new(dentry, inode);
6880
6881 out_fail:
6882         btrfs_end_transaction(trans);
6883         if (err && inode) {
6884                 inode_dec_link_count(inode);
6885                 discard_new_inode(inode);
6886         }
6887         btrfs_btree_balance_dirty(fs_info);
6888         return err;
6889 }
6890
6891 static noinline int uncompress_inline(struct btrfs_path *path,
6892                                       struct page *page,
6893                                       size_t pg_offset, u64 extent_offset,
6894                                       struct btrfs_file_extent_item *item)
6895 {
6896         int ret;
6897         struct extent_buffer *leaf = path->nodes[0];
6898         char *tmp;
6899         size_t max_size;
6900         unsigned long inline_size;
6901         unsigned long ptr;
6902         int compress_type;
6903
6904         WARN_ON(pg_offset != 0);
6905         compress_type = btrfs_file_extent_compression(leaf, item);
6906         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6907         inline_size = btrfs_file_extent_inline_item_len(leaf,
6908                                         btrfs_item_nr(path->slots[0]));
6909         tmp = kmalloc(inline_size, GFP_NOFS);
6910         if (!tmp)
6911                 return -ENOMEM;
6912         ptr = btrfs_file_extent_inline_start(item);
6913
6914         read_extent_buffer(leaf, tmp, ptr, inline_size);
6915
6916         max_size = min_t(unsigned long, PAGE_SIZE, max_size);
6917         ret = btrfs_decompress(compress_type, tmp, page,
6918                                extent_offset, inline_size, max_size);
6919
6920         /*
6921          * decompression code contains a memset to fill in any space between the end
6922          * of the uncompressed data and the end of max_size in case the decompressed
6923          * data ends up shorter than ram_bytes.  That doesn't cover the hole between
6924          * the end of an inline extent and the beginning of the next block, so we
6925          * cover that region here.
6926          */
6927
6928         if (max_size + pg_offset < PAGE_SIZE) {
6929                 char *map = kmap(page);
6930                 memset(map + pg_offset + max_size, 0, PAGE_SIZE - max_size - pg_offset);
6931                 kunmap(page);
6932         }
6933         kfree(tmp);
6934         return ret;
6935 }
6936
6937 /*
6938  * a bit scary, this does extent mapping from logical file offset to the disk.
6939  * the ugly parts come from merging extents from the disk with the in-ram
6940  * representation.  This gets more complex because of the data=ordered code,
6941  * where the in-ram extents might be locked pending data=ordered completion.
6942  *
6943  * This also copies inline extents directly into the page.
6944  */
6945 struct extent_map *btrfs_get_extent(struct btrfs_inode *inode,
6946                                     struct page *page,
6947                                     size_t pg_offset, u64 start, u64 len,
6948                                     int create)
6949 {
6950         struct btrfs_fs_info *fs_info = inode->root->fs_info;
6951         int ret;
6952         int err = 0;
6953         u64 extent_start = 0;
6954         u64 extent_end = 0;
6955         u64 objectid = btrfs_ino(inode);
6956         int extent_type = -1;
6957         struct btrfs_path *path = NULL;
6958         struct btrfs_root *root = inode->root;
6959         struct btrfs_file_extent_item *item;
6960         struct extent_buffer *leaf;
6961         struct btrfs_key found_key;
6962         struct extent_map *em = NULL;
6963         struct extent_map_tree *em_tree = &inode->extent_tree;
6964         struct extent_io_tree *io_tree = &inode->io_tree;
6965         const bool new_inline = !page || create;
6966
6967         read_lock(&em_tree->lock);
6968         em = lookup_extent_mapping(em_tree, start, len);
6969         read_unlock(&em_tree->lock);
6970
6971         if (em) {
6972                 if (em->start > start || em->start + em->len <= start)
6973                         free_extent_map(em);
6974                 else if (em->block_start == EXTENT_MAP_INLINE && page)
6975                         free_extent_map(em);
6976                 else
6977                         goto out;
6978         }
6979         em = alloc_extent_map();
6980         if (!em) {
6981                 err = -ENOMEM;
6982                 goto out;
6983         }
6984         em->start = EXTENT_MAP_HOLE;
6985         em->orig_start = EXTENT_MAP_HOLE;
6986         em->len = (u64)-1;
6987         em->block_len = (u64)-1;
6988
6989         path = btrfs_alloc_path();
6990         if (!path) {
6991                 err = -ENOMEM;
6992                 goto out;
6993         }
6994
6995         /* Chances are we'll be called again, so go ahead and do readahead */
6996         path->reada = READA_FORWARD;
6997
6998         /*
6999          * Unless we're going to uncompress the inline extent, no sleep would
7000          * happen.
7001          */
7002         path->leave_spinning = 1;
7003
7004         ret = btrfs_lookup_file_extent(NULL, root, path, objectid, start, 0);
7005         if (ret < 0) {
7006                 err = ret;
7007                 goto out;
7008         } else if (ret > 0) {
7009                 if (path->slots[0] == 0)
7010                         goto not_found;
7011                 path->slots[0]--;
7012         }
7013
7014         leaf = path->nodes[0];
7015         item = btrfs_item_ptr(leaf, path->slots[0],
7016                               struct btrfs_file_extent_item);
7017         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7018         if (found_key.objectid != objectid ||
7019             found_key.type != BTRFS_EXTENT_DATA_KEY) {
7020                 /*
7021                  * If we backup past the first extent we want to move forward
7022                  * and see if there is an extent in front of us, otherwise we'll
7023                  * say there is a hole for our whole search range which can
7024                  * cause problems.
7025                  */
7026                 extent_end = start;
7027                 goto next;
7028         }
7029
7030         extent_type = btrfs_file_extent_type(leaf, item);
7031         extent_start = found_key.offset;
7032         if (extent_type == BTRFS_FILE_EXTENT_REG ||
7033             extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
7034                 /* Only regular file could have regular/prealloc extent */
7035                 if (!S_ISREG(inode->vfs_inode.i_mode)) {
7036                         ret = -EUCLEAN;
7037                         btrfs_crit(fs_info,
7038                 "regular/prealloc extent found for non-regular inode %llu",
7039                                    btrfs_ino(inode));
7040                         goto out;
7041                 }
7042                 extent_end = extent_start +
7043                        btrfs_file_extent_num_bytes(leaf, item);
7044
7045                 trace_btrfs_get_extent_show_fi_regular(inode, leaf, item,
7046                                                        extent_start);
7047         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
7048                 size_t size;
7049
7050                 size = btrfs_file_extent_ram_bytes(leaf, item);
7051                 extent_end = ALIGN(extent_start + size,
7052                                    fs_info->sectorsize);
7053
7054                 trace_btrfs_get_extent_show_fi_inline(inode, leaf, item,
7055                                                       path->slots[0],
7056                                                       extent_start);
7057         }
7058 next:
7059         if (start >= extent_end) {
7060                 path->slots[0]++;
7061                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
7062                         ret = btrfs_next_leaf(root, path);
7063                         if (ret < 0) {
7064                                 err = ret;
7065                                 goto out;
7066                         } else if (ret > 0) {
7067                                 goto not_found;
7068                         }
7069                         leaf = path->nodes[0];
7070                 }
7071                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7072                 if (found_key.objectid != objectid ||
7073                     found_key.type != BTRFS_EXTENT_DATA_KEY)
7074                         goto not_found;
7075                 if (start + len <= found_key.offset)
7076                         goto not_found;
7077                 if (start > found_key.offset)
7078                         goto next;
7079
7080                 /* New extent overlaps with existing one */
7081                 em->start = start;
7082                 em->orig_start = start;
7083                 em->len = found_key.offset - start;
7084                 em->block_start = EXTENT_MAP_HOLE;
7085                 goto insert;
7086         }
7087
7088         btrfs_extent_item_to_extent_map(inode, path, item,
7089                         new_inline, em);
7090
7091         if (extent_type == BTRFS_FILE_EXTENT_REG ||
7092             extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
7093                 goto insert;
7094         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
7095                 unsigned long ptr;
7096                 char *map;
7097                 size_t size;
7098                 size_t extent_offset;
7099                 size_t copy_size;
7100
7101                 if (new_inline)
7102                         goto out;
7103
7104                 size = btrfs_file_extent_ram_bytes(leaf, item);
7105                 extent_offset = page_offset(page) + pg_offset - extent_start;
7106                 copy_size = min_t(u64, PAGE_SIZE - pg_offset,
7107                                   size - extent_offset);
7108                 em->start = extent_start + extent_offset;
7109                 em->len = ALIGN(copy_size, fs_info->sectorsize);
7110                 em->orig_block_len = em->len;
7111                 em->orig_start = em->start;
7112                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
7113
7114                 btrfs_set_path_blocking(path);
7115                 if (!PageUptodate(page)) {
7116                         if (btrfs_file_extent_compression(leaf, item) !=
7117                             BTRFS_COMPRESS_NONE) {
7118                                 ret = uncompress_inline(path, page, pg_offset,
7119                                                         extent_offset, item);
7120                                 if (ret) {
7121                                         err = ret;
7122                                         goto out;
7123                                 }
7124                         } else {
7125                                 map = kmap(page);
7126                                 read_extent_buffer(leaf, map + pg_offset, ptr,
7127                                                    copy_size);
7128                                 if (pg_offset + copy_size < PAGE_SIZE) {
7129                                         memset(map + pg_offset + copy_size, 0,
7130                                                PAGE_SIZE - pg_offset -
7131                                                copy_size);
7132                                 }
7133                                 kunmap(page);
7134                         }
7135                         flush_dcache_page(page);
7136                 }
7137                 set_extent_uptodate(io_tree, em->start,
7138                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
7139                 goto insert;
7140         }
7141 not_found:
7142         em->start = start;
7143         em->orig_start = start;
7144         em->len = len;
7145         em->block_start = EXTENT_MAP_HOLE;
7146 insert:
7147         btrfs_release_path(path);
7148         if (em->start > start || extent_map_end(em) <= start) {
7149                 btrfs_err(fs_info,
7150                           "bad extent! em: [%llu %llu] passed [%llu %llu]",
7151                           em->start, em->len, start, len);
7152                 err = -EIO;
7153                 goto out;
7154         }
7155
7156         err = 0;
7157         write_lock(&em_tree->lock);
7158         err = btrfs_add_extent_mapping(fs_info, em_tree, &em, start, len);
7159         write_unlock(&em_tree->lock);
7160 out:
7161         btrfs_free_path(path);
7162
7163         trace_btrfs_get_extent(root, inode, em);
7164
7165         if (err) {
7166                 free_extent_map(em);
7167                 return ERR_PTR(err);
7168         }
7169         BUG_ON(!em); /* Error is always set */
7170         return em;
7171 }
7172
7173 struct extent_map *btrfs_get_extent_fiemap(struct btrfs_inode *inode,
7174                                            u64 start, u64 len)
7175 {
7176         struct extent_map *em;
7177         struct extent_map *hole_em = NULL;
7178         u64 delalloc_start = start;
7179         u64 end;
7180         u64 delalloc_len;
7181         u64 delalloc_end;
7182         int err = 0;
7183
7184         em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
7185         if (IS_ERR(em))
7186                 return em;
7187         /*
7188          * If our em maps to:
7189          * - a hole or
7190          * - a pre-alloc extent,
7191          * there might actually be delalloc bytes behind it.
7192          */
7193         if (em->block_start != EXTENT_MAP_HOLE &&
7194             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7195                 return em;
7196         else
7197                 hole_em = em;
7198
7199         /* check to see if we've wrapped (len == -1 or similar) */
7200         end = start + len;
7201         if (end < start)
7202                 end = (u64)-1;
7203         else
7204                 end -= 1;
7205
7206         em = NULL;
7207
7208         /* ok, we didn't find anything, lets look for delalloc */
7209         delalloc_len = count_range_bits(&inode->io_tree, &delalloc_start,
7210                                  end, len, EXTENT_DELALLOC, 1);
7211         delalloc_end = delalloc_start + delalloc_len;
7212         if (delalloc_end < delalloc_start)
7213                 delalloc_end = (u64)-1;
7214
7215         /*
7216          * We didn't find anything useful, return the original results from
7217          * get_extent()
7218          */
7219         if (delalloc_start > end || delalloc_end <= start) {
7220                 em = hole_em;
7221                 hole_em = NULL;
7222                 goto out;
7223         }
7224
7225         /*
7226          * Adjust the delalloc_start to make sure it doesn't go backwards from
7227          * the start they passed in
7228          */
7229         delalloc_start = max(start, delalloc_start);
7230         delalloc_len = delalloc_end - delalloc_start;
7231
7232         if (delalloc_len > 0) {
7233                 u64 hole_start;
7234                 u64 hole_len;
7235                 const u64 hole_end = extent_map_end(hole_em);
7236
7237                 em = alloc_extent_map();
7238                 if (!em) {
7239                         err = -ENOMEM;
7240                         goto out;
7241                 }
7242
7243                 ASSERT(hole_em);
7244                 /*
7245                  * When btrfs_get_extent can't find anything it returns one
7246                  * huge hole
7247                  *
7248                  * Make sure what it found really fits our range, and adjust to
7249                  * make sure it is based on the start from the caller
7250                  */
7251                 if (hole_end <= start || hole_em->start > end) {
7252                        free_extent_map(hole_em);
7253                        hole_em = NULL;
7254                 } else {
7255                        hole_start = max(hole_em->start, start);
7256                        hole_len = hole_end - hole_start;
7257                 }
7258
7259                 if (hole_em && delalloc_start > hole_start) {
7260                         /*
7261                          * Our hole starts before our delalloc, so we have to
7262                          * return just the parts of the hole that go until the
7263                          * delalloc starts
7264                          */
7265                         em->len = min(hole_len, delalloc_start - hole_start);
7266                         em->start = hole_start;
7267                         em->orig_start = hole_start;
7268                         /*
7269                          * Don't adjust block start at all, it is fixed at
7270                          * EXTENT_MAP_HOLE
7271                          */
7272                         em->block_start = hole_em->block_start;
7273                         em->block_len = hole_len;
7274                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
7275                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
7276                 } else {
7277                         /*
7278                          * Hole is out of passed range or it starts after
7279                          * delalloc range
7280                          */
7281                         em->start = delalloc_start;
7282                         em->len = delalloc_len;
7283                         em->orig_start = delalloc_start;
7284                         em->block_start = EXTENT_MAP_DELALLOC;
7285                         em->block_len = delalloc_len;
7286                 }
7287         } else {
7288                 return hole_em;
7289         }
7290 out:
7291
7292         free_extent_map(hole_em);
7293         if (err) {
7294                 free_extent_map(em);
7295                 return ERR_PTR(err);
7296         }
7297         return em;
7298 }
7299
7300 static struct extent_map *btrfs_create_dio_extent(struct inode *inode,
7301                                                   const u64 start,
7302                                                   const u64 len,
7303                                                   const u64 orig_start,
7304                                                   const u64 block_start,
7305                                                   const u64 block_len,
7306                                                   const u64 orig_block_len,
7307                                                   const u64 ram_bytes,
7308                                                   const int type)
7309 {
7310         struct extent_map *em = NULL;
7311         int ret;
7312
7313         if (type != BTRFS_ORDERED_NOCOW) {
7314                 em = create_io_em(inode, start, len, orig_start,
7315                                   block_start, block_len, orig_block_len,
7316                                   ram_bytes,
7317                                   BTRFS_COMPRESS_NONE, /* compress_type */
7318                                   type);
7319                 if (IS_ERR(em))
7320                         goto out;
7321         }
7322         ret = btrfs_add_ordered_extent_dio(inode, start, block_start,
7323                                            len, block_len, type);
7324         if (ret) {
7325                 if (em) {
7326                         free_extent_map(em);
7327                         btrfs_drop_extent_cache(BTRFS_I(inode), start,
7328                                                 start + len - 1, 0);
7329                 }
7330                 em = ERR_PTR(ret);
7331         }
7332  out:
7333
7334         return em;
7335 }
7336
7337 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
7338                                                   u64 start, u64 len)
7339 {
7340         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7341         struct btrfs_root *root = BTRFS_I(inode)->root;
7342         struct extent_map *em;
7343         struct btrfs_key ins;
7344         u64 alloc_hint;
7345         int ret;
7346
7347         alloc_hint = get_extent_allocation_hint(inode, start, len);
7348         ret = btrfs_reserve_extent(root, len, len, fs_info->sectorsize,
7349                                    0, alloc_hint, &ins, 1, 1);
7350         if (ret)
7351                 return ERR_PTR(ret);
7352
7353         em = btrfs_create_dio_extent(inode, start, ins.offset, start,
7354                                      ins.objectid, ins.offset, ins.offset,
7355                                      ins.offset, BTRFS_ORDERED_REGULAR);
7356         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
7357         if (IS_ERR(em))
7358                 btrfs_free_reserved_extent(fs_info, ins.objectid,
7359                                            ins.offset, 1);
7360
7361         return em;
7362 }
7363
7364 /*
7365  * returns 1 when the nocow is safe, < 1 on error, 0 if the
7366  * block must be cow'd
7367  */
7368 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
7369                               u64 *orig_start, u64 *orig_block_len,
7370                               u64 *ram_bytes)
7371 {
7372         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7373         struct btrfs_path *path;
7374         int ret;
7375         struct extent_buffer *leaf;
7376         struct btrfs_root *root = BTRFS_I(inode)->root;
7377         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7378         struct btrfs_file_extent_item *fi;
7379         struct btrfs_key key;
7380         u64 disk_bytenr;
7381         u64 backref_offset;
7382         u64 extent_end;
7383         u64 num_bytes;
7384         int slot;
7385         int found_type;
7386         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
7387
7388         path = btrfs_alloc_path();
7389         if (!path)
7390                 return -ENOMEM;
7391
7392         ret = btrfs_lookup_file_extent(NULL, root, path,
7393                         btrfs_ino(BTRFS_I(inode)), offset, 0);
7394         if (ret < 0)
7395                 goto out;
7396
7397         slot = path->slots[0];
7398         if (ret == 1) {
7399                 if (slot == 0) {
7400                         /* can't find the item, must cow */
7401                         ret = 0;
7402                         goto out;
7403                 }
7404                 slot--;
7405         }
7406         ret = 0;
7407         leaf = path->nodes[0];
7408         btrfs_item_key_to_cpu(leaf, &key, slot);
7409         if (key.objectid != btrfs_ino(BTRFS_I(inode)) ||
7410             key.type != BTRFS_EXTENT_DATA_KEY) {
7411                 /* not our file or wrong item type, must cow */
7412                 goto out;
7413         }
7414
7415         if (key.offset > offset) {
7416                 /* Wrong offset, must cow */
7417                 goto out;
7418         }
7419
7420         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
7421         found_type = btrfs_file_extent_type(leaf, fi);
7422         if (found_type != BTRFS_FILE_EXTENT_REG &&
7423             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
7424                 /* not a regular extent, must cow */
7425                 goto out;
7426         }
7427
7428         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
7429                 goto out;
7430
7431         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
7432         if (extent_end <= offset)
7433                 goto out;
7434
7435         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7436         if (disk_bytenr == 0)
7437                 goto out;
7438
7439         if (btrfs_file_extent_compression(leaf, fi) ||
7440             btrfs_file_extent_encryption(leaf, fi) ||
7441             btrfs_file_extent_other_encoding(leaf, fi))
7442                 goto out;
7443
7444         /*
7445          * Do the same check as in btrfs_cross_ref_exist but without the
7446          * unnecessary search.
7447          */
7448         if (btrfs_file_extent_generation(leaf, fi) <=
7449             btrfs_root_last_snapshot(&root->root_item))
7450                 goto out;
7451
7452         backref_offset = btrfs_file_extent_offset(leaf, fi);
7453
7454         if (orig_start) {
7455                 *orig_start = key.offset - backref_offset;
7456                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
7457                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
7458         }
7459
7460         if (btrfs_extent_readonly(fs_info, disk_bytenr))
7461                 goto out;
7462
7463         num_bytes = min(offset + *len, extent_end) - offset;
7464         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7465                 u64 range_end;
7466
7467                 range_end = round_up(offset + num_bytes,
7468                                      root->fs_info->sectorsize) - 1;
7469                 ret = test_range_bit(io_tree, offset, range_end,
7470                                      EXTENT_DELALLOC, 0, NULL);
7471                 if (ret) {
7472                         ret = -EAGAIN;
7473                         goto out;
7474                 }
7475         }
7476
7477         btrfs_release_path(path);
7478
7479         /*
7480          * look for other files referencing this extent, if we
7481          * find any we must cow
7482          */
7483
7484         ret = btrfs_cross_ref_exist(root, btrfs_ino(BTRFS_I(inode)),
7485                                     key.offset - backref_offset, disk_bytenr);
7486         if (ret) {
7487                 ret = 0;
7488                 goto out;
7489         }
7490
7491         /*
7492          * adjust disk_bytenr and num_bytes to cover just the bytes
7493          * in this extent we are about to write.  If there
7494          * are any csums in that range we have to cow in order
7495          * to keep the csums correct
7496          */
7497         disk_bytenr += backref_offset;
7498         disk_bytenr += offset - key.offset;
7499         if (csum_exist_in_range(fs_info, disk_bytenr, num_bytes))
7500                 goto out;
7501         /*
7502          * all of the above have passed, it is safe to overwrite this extent
7503          * without cow
7504          */
7505         *len = num_bytes;
7506         ret = 1;
7507 out:
7508         btrfs_free_path(path);
7509         return ret;
7510 }
7511
7512 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7513                               struct extent_state **cached_state, int writing)
7514 {
7515         struct btrfs_ordered_extent *ordered;
7516         int ret = 0;
7517
7518         while (1) {
7519                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7520                                  cached_state);
7521                 /*
7522                  * We're concerned with the entire range that we're going to be
7523                  * doing DIO to, so we need to make sure there's no ordered
7524                  * extents in this range.
7525                  */
7526                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), lockstart,
7527                                                      lockend - lockstart + 1);
7528
7529                 /*
7530                  * We need to make sure there are no buffered pages in this
7531                  * range either, we could have raced between the invalidate in
7532                  * generic_file_direct_write and locking the extent.  The
7533                  * invalidate needs to happen so that reads after a write do not
7534                  * get stale data.
7535                  */
7536                 if (!ordered &&
7537                     (!writing || !filemap_range_has_page(inode->i_mapping,
7538                                                          lockstart, lockend)))
7539                         break;
7540
7541                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7542                                      cached_state);
7543
7544                 if (ordered) {
7545                         /*
7546                          * If we are doing a DIO read and the ordered extent we
7547                          * found is for a buffered write, we can not wait for it
7548                          * to complete and retry, because if we do so we can
7549                          * deadlock with concurrent buffered writes on page
7550                          * locks. This happens only if our DIO read covers more
7551                          * than one extent map, if at this point has already
7552                          * created an ordered extent for a previous extent map
7553                          * and locked its range in the inode's io tree, and a
7554                          * concurrent write against that previous extent map's
7555                          * range and this range started (we unlock the ranges
7556                          * in the io tree only when the bios complete and
7557                          * buffered writes always lock pages before attempting
7558                          * to lock range in the io tree).
7559                          */
7560                         if (writing ||
7561                             test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags))
7562                                 btrfs_start_ordered_extent(inode, ordered, 1);
7563                         else
7564                                 ret = -ENOTBLK;
7565                         btrfs_put_ordered_extent(ordered);
7566                 } else {
7567                         /*
7568                          * We could trigger writeback for this range (and wait
7569                          * for it to complete) and then invalidate the pages for
7570                          * this range (through invalidate_inode_pages2_range()),
7571                          * but that can lead us to a deadlock with a concurrent
7572                          * call to readpages() (a buffered read or a defrag call
7573                          * triggered a readahead) on a page lock due to an
7574                          * ordered dio extent we created before but did not have
7575                          * yet a corresponding bio submitted (whence it can not
7576                          * complete), which makes readpages() wait for that
7577                          * ordered extent to complete while holding a lock on
7578                          * that page.
7579                          */
7580                         ret = -ENOTBLK;
7581                 }
7582
7583                 if (ret)
7584                         break;
7585
7586                 cond_resched();
7587         }
7588
7589         return ret;
7590 }
7591
7592 /* The callers of this must take lock_extent() */
7593 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
7594                                        u64 orig_start, u64 block_start,
7595                                        u64 block_len, u64 orig_block_len,
7596                                        u64 ram_bytes, int compress_type,
7597                                        int type)
7598 {
7599         struct extent_map_tree *em_tree;
7600         struct extent_map *em;
7601         int ret;
7602
7603         ASSERT(type == BTRFS_ORDERED_PREALLOC ||
7604                type == BTRFS_ORDERED_COMPRESSED ||
7605                type == BTRFS_ORDERED_NOCOW ||
7606                type == BTRFS_ORDERED_REGULAR);
7607
7608         em_tree = &BTRFS_I(inode)->extent_tree;
7609         em = alloc_extent_map();
7610         if (!em)
7611                 return ERR_PTR(-ENOMEM);
7612
7613         em->start = start;
7614         em->orig_start = orig_start;
7615         em->len = len;
7616         em->block_len = block_len;
7617         em->block_start = block_start;
7618         em->orig_block_len = orig_block_len;
7619         em->ram_bytes = ram_bytes;
7620         em->generation = -1;
7621         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7622         if (type == BTRFS_ORDERED_PREALLOC) {
7623                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7624         } else if (type == BTRFS_ORDERED_COMPRESSED) {
7625                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
7626                 em->compress_type = compress_type;
7627         }
7628
7629         do {
7630                 btrfs_drop_extent_cache(BTRFS_I(inode), em->start,
7631                                 em->start + em->len - 1, 0);
7632                 write_lock(&em_tree->lock);
7633                 ret = add_extent_mapping(em_tree, em, 1);
7634                 write_unlock(&em_tree->lock);
7635                 /*
7636                  * The caller has taken lock_extent(), who could race with us
7637                  * to add em?
7638                  */
7639         } while (ret == -EEXIST);
7640
7641         if (ret) {
7642                 free_extent_map(em);
7643                 return ERR_PTR(ret);
7644         }
7645
7646         /* em got 2 refs now, callers needs to do free_extent_map once. */
7647         return em;
7648 }
7649
7650
7651 static int btrfs_get_blocks_direct_read(struct extent_map *em,
7652                                         struct buffer_head *bh_result,
7653                                         struct inode *inode,
7654                                         u64 start, u64 len)
7655 {
7656         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7657
7658         if (em->block_start == EXTENT_MAP_HOLE ||
7659                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7660                 return -ENOENT;
7661
7662         len = min(len, em->len - (start - em->start));
7663
7664         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7665                 inode->i_blkbits;
7666         bh_result->b_size = len;
7667         bh_result->b_bdev = fs_info->fs_devices->latest_bdev;
7668         set_buffer_mapped(bh_result);
7669
7670         return 0;
7671 }
7672
7673 static int btrfs_get_blocks_direct_write(struct extent_map **map,
7674                                          struct buffer_head *bh_result,
7675                                          struct inode *inode,
7676                                          struct btrfs_dio_data *dio_data,
7677                                          u64 start, u64 len)
7678 {
7679         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7680         struct extent_map *em = *map;
7681         int ret = 0;
7682
7683         /*
7684          * We don't allocate a new extent in the following cases
7685          *
7686          * 1) The inode is marked as NODATACOW. In this case we'll just use the
7687          * existing extent.
7688          * 2) The extent is marked as PREALLOC. We're good to go here and can
7689          * just use the extent.
7690          *
7691          */
7692         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7693             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7694              em->block_start != EXTENT_MAP_HOLE)) {
7695                 int type;
7696                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7697
7698                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7699                         type = BTRFS_ORDERED_PREALLOC;
7700                 else
7701                         type = BTRFS_ORDERED_NOCOW;
7702                 len = min(len, em->len - (start - em->start));
7703                 block_start = em->block_start + (start - em->start);
7704
7705                 if (can_nocow_extent(inode, start, &len, &orig_start,
7706                                      &orig_block_len, &ram_bytes) == 1 &&
7707                     btrfs_inc_nocow_writers(fs_info, block_start)) {
7708                         struct extent_map *em2;
7709
7710                         em2 = btrfs_create_dio_extent(inode, start, len,
7711                                                       orig_start, block_start,
7712                                                       len, orig_block_len,
7713                                                       ram_bytes, type);
7714                         btrfs_dec_nocow_writers(fs_info, block_start);
7715                         if (type == BTRFS_ORDERED_PREALLOC) {
7716                                 free_extent_map(em);
7717                                 *map = em = em2;
7718                         }
7719
7720                         if (em2 && IS_ERR(em2)) {
7721                                 ret = PTR_ERR(em2);
7722                                 goto out;
7723                         }
7724                         /*
7725                          * For inode marked NODATACOW or extent marked PREALLOC,
7726                          * use the existing or preallocated extent, so does not
7727                          * need to adjust btrfs_space_info's bytes_may_use.
7728                          */
7729                         btrfs_free_reserved_data_space_noquota(inode, start,
7730                                                                len);
7731                         goto skip_cow;
7732                 }
7733         }
7734
7735         /* this will cow the extent */
7736         len = bh_result->b_size;
7737         free_extent_map(em);
7738         *map = em = btrfs_new_extent_direct(inode, start, len);
7739         if (IS_ERR(em)) {
7740                 ret = PTR_ERR(em);
7741                 goto out;
7742         }
7743
7744         len = min(len, em->len - (start - em->start));
7745
7746 skip_cow:
7747         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7748                 inode->i_blkbits;
7749         bh_result->b_size = len;
7750         bh_result->b_bdev = fs_info->fs_devices->latest_bdev;
7751         set_buffer_mapped(bh_result);
7752
7753         if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7754                 set_buffer_new(bh_result);
7755
7756         /*
7757          * Need to update the i_size under the extent lock so buffered
7758          * readers will get the updated i_size when we unlock.
7759          */
7760         if (!dio_data->overwrite && start + len > i_size_read(inode))
7761                 i_size_write(inode, start + len);
7762
7763         WARN_ON(dio_data->reserve < len);
7764         dio_data->reserve -= len;
7765         dio_data->unsubmitted_oe_range_end = start + len;
7766         current->journal_info = dio_data;
7767 out:
7768         return ret;
7769 }
7770
7771 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7772                                    struct buffer_head *bh_result, int create)
7773 {
7774         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7775         struct extent_map *em;
7776         struct extent_state *cached_state = NULL;
7777         struct btrfs_dio_data *dio_data = NULL;
7778         u64 start = iblock << inode->i_blkbits;
7779         u64 lockstart, lockend;
7780         u64 len = bh_result->b_size;
7781         int ret = 0;
7782
7783         if (!create)
7784                 len = min_t(u64, len, fs_info->sectorsize);
7785
7786         lockstart = start;
7787         lockend = start + len - 1;
7788
7789         if (current->journal_info) {
7790                 /*
7791                  * Need to pull our outstanding extents and set journal_info to NULL so
7792                  * that anything that needs to check if there's a transaction doesn't get
7793                  * confused.
7794                  */
7795                 dio_data = current->journal_info;
7796                 current->journal_info = NULL;
7797         }
7798
7799         /*
7800          * If this errors out it's because we couldn't invalidate pagecache for
7801          * this range and we need to fallback to buffered.
7802          */
7803         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7804                                create)) {
7805                 ret = -ENOTBLK;
7806                 goto err;
7807         }
7808
7809         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, start, len, 0);
7810         if (IS_ERR(em)) {
7811                 ret = PTR_ERR(em);
7812                 goto unlock_err;
7813         }
7814
7815         /*
7816          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7817          * io.  INLINE is special, and we could probably kludge it in here, but
7818          * it's still buffered so for safety lets just fall back to the generic
7819          * buffered path.
7820          *
7821          * For COMPRESSED we _have_ to read the entire extent in so we can
7822          * decompress it, so there will be buffering required no matter what we
7823          * do, so go ahead and fallback to buffered.
7824          *
7825          * We return -ENOTBLK because that's what makes DIO go ahead and go back
7826          * to buffered IO.  Don't blame me, this is the price we pay for using
7827          * the generic code.
7828          */
7829         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7830             em->block_start == EXTENT_MAP_INLINE) {
7831                 free_extent_map(em);
7832                 ret = -ENOTBLK;
7833                 goto unlock_err;
7834         }
7835
7836         if (create) {
7837                 ret = btrfs_get_blocks_direct_write(&em, bh_result, inode,
7838                                                     dio_data, start, len);
7839                 if (ret < 0)
7840                         goto unlock_err;
7841
7842                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
7843                                      lockend, &cached_state);
7844         } else {
7845                 ret = btrfs_get_blocks_direct_read(em, bh_result, inode,
7846                                                    start, len);
7847                 /* Can be negative only if we read from a hole */
7848                 if (ret < 0) {
7849                         ret = 0;
7850                         free_extent_map(em);
7851                         goto unlock_err;
7852                 }
7853                 /*
7854                  * We need to unlock only the end area that we aren't using.
7855                  * The rest is going to be unlocked by the endio routine.
7856                  */
7857                 lockstart = start + bh_result->b_size;
7858                 if (lockstart < lockend) {
7859                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
7860                                              lockstart, lockend, &cached_state);
7861                 } else {
7862                         free_extent_state(cached_state);
7863                 }
7864         }
7865
7866         free_extent_map(em);
7867
7868         return 0;
7869
7870 unlock_err:
7871         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7872                              &cached_state);
7873 err:
7874         if (dio_data)
7875                 current->journal_info = dio_data;
7876         return ret;
7877 }
7878
7879 static inline blk_status_t submit_dio_repair_bio(struct inode *inode,
7880                                                  struct bio *bio,
7881                                                  int mirror_num)
7882 {
7883         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7884         blk_status_t ret;
7885
7886         BUG_ON(bio_op(bio) == REQ_OP_WRITE);
7887
7888         ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DIO_REPAIR);
7889         if (ret)
7890                 return ret;
7891
7892         ret = btrfs_map_bio(fs_info, bio, mirror_num);
7893
7894         return ret;
7895 }
7896
7897 static int btrfs_check_dio_repairable(struct inode *inode,
7898                                       struct bio *failed_bio,
7899                                       struct io_failure_record *failrec,
7900                                       int failed_mirror)
7901 {
7902         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7903         int num_copies;
7904
7905         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
7906         if (num_copies == 1) {
7907                 /*
7908                  * we only have a single copy of the data, so don't bother with
7909                  * all the retry and error correction code that follows. no
7910                  * matter what the error is, it is very likely to persist.
7911                  */
7912                 btrfs_debug(fs_info,
7913                         "Check DIO Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
7914                         num_copies, failrec->this_mirror, failed_mirror);
7915                 return 0;
7916         }
7917
7918         failrec->failed_mirror = failed_mirror;
7919         failrec->this_mirror++;
7920         if (failrec->this_mirror == failed_mirror)
7921                 failrec->this_mirror++;
7922
7923         if (failrec->this_mirror > num_copies) {
7924                 btrfs_debug(fs_info,
7925                         "Check DIO Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
7926                         num_copies, failrec->this_mirror, failed_mirror);
7927                 return 0;
7928         }
7929
7930         return 1;
7931 }
7932
7933 static blk_status_t dio_read_error(struct inode *inode, struct bio *failed_bio,
7934                                    struct page *page, unsigned int pgoff,
7935                                    u64 start, u64 end, int failed_mirror,
7936                                    bio_end_io_t *repair_endio, void *repair_arg)
7937 {
7938         struct io_failure_record *failrec;
7939         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7940         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
7941         struct bio *bio;
7942         int isector;
7943         unsigned int read_mode = 0;
7944         int segs;
7945         int ret;
7946         blk_status_t status;
7947         struct bio_vec bvec;
7948
7949         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
7950
7951         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
7952         if (ret)
7953                 return errno_to_blk_status(ret);
7954
7955         ret = btrfs_check_dio_repairable(inode, failed_bio, failrec,
7956                                          failed_mirror);
7957         if (!ret) {
7958                 free_io_failure(failure_tree, io_tree, failrec);
7959                 return BLK_STS_IOERR;
7960         }
7961
7962         segs = bio_segments(failed_bio);
7963         bio_get_first_bvec(failed_bio, &bvec);
7964         if (segs > 1 ||
7965             (bvec.bv_len > btrfs_inode_sectorsize(inode)))
7966                 read_mode |= REQ_FAILFAST_DEV;
7967
7968         isector = start - btrfs_io_bio(failed_bio)->logical;
7969         isector >>= inode->i_sb->s_blocksize_bits;
7970         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
7971                                 pgoff, isector, repair_endio, repair_arg);
7972         bio->bi_opf = REQ_OP_READ | read_mode;
7973
7974         btrfs_debug(BTRFS_I(inode)->root->fs_info,
7975                     "repair DIO read error: submitting new dio read[%#x] to this_mirror=%d, in_validation=%d",
7976                     read_mode, failrec->this_mirror, failrec->in_validation);
7977
7978         status = submit_dio_repair_bio(inode, bio, failrec->this_mirror);
7979         if (status) {
7980                 free_io_failure(failure_tree, io_tree, failrec);
7981                 bio_put(bio);
7982         }
7983
7984         return status;
7985 }
7986
7987 struct btrfs_retry_complete {
7988         struct completion done;
7989         struct inode *inode;
7990         u64 start;
7991         int uptodate;
7992 };
7993
7994 static void btrfs_retry_endio_nocsum(struct bio *bio)
7995 {
7996         struct btrfs_retry_complete *done = bio->bi_private;
7997         struct inode *inode = done->inode;
7998         struct bio_vec *bvec;
7999         struct extent_io_tree *io_tree, *failure_tree;
8000         struct bvec_iter_all iter_all;
8001
8002         if (bio->bi_status)
8003                 goto end;
8004
8005         ASSERT(bio->bi_vcnt == 1);
8006         io_tree = &BTRFS_I(inode)->io_tree;
8007         failure_tree = &BTRFS_I(inode)->io_failure_tree;
8008         ASSERT(bio_first_bvec_all(bio)->bv_len == btrfs_inode_sectorsize(inode));
8009
8010         done->uptodate = 1;
8011         ASSERT(!bio_flagged(bio, BIO_CLONED));
8012         bio_for_each_segment_all(bvec, bio, iter_all)
8013                 clean_io_failure(BTRFS_I(inode)->root->fs_info, failure_tree,
8014                                  io_tree, done->start, bvec->bv_page,
8015                                  btrfs_ino(BTRFS_I(inode)), 0);
8016 end:
8017         complete(&done->done);
8018         bio_put(bio);
8019 }
8020
8021 static blk_status_t __btrfs_correct_data_nocsum(struct inode *inode,
8022                                                 struct btrfs_io_bio *io_bio)
8023 {
8024         struct btrfs_fs_info *fs_info;
8025         struct bio_vec bvec;
8026         struct bvec_iter iter;
8027         struct btrfs_retry_complete done;
8028         u64 start;
8029         unsigned int pgoff;
8030         u32 sectorsize;
8031         int nr_sectors;
8032         blk_status_t ret;
8033         blk_status_t err = BLK_STS_OK;
8034
8035         fs_info = BTRFS_I(inode)->root->fs_info;
8036         sectorsize = fs_info->sectorsize;
8037
8038         start = io_bio->logical;
8039         done.inode = inode;
8040         io_bio->bio.bi_iter = io_bio->iter;
8041
8042         bio_for_each_segment(bvec, &io_bio->bio, iter) {
8043                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
8044                 pgoff = bvec.bv_offset;
8045
8046 next_block_or_try_again:
8047                 done.uptodate = 0;
8048                 done.start = start;
8049                 init_completion(&done.done);
8050
8051                 ret = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
8052                                 pgoff, start, start + sectorsize - 1,
8053                                 io_bio->mirror_num,
8054                                 btrfs_retry_endio_nocsum, &done);
8055                 if (ret) {
8056                         err = ret;
8057                         goto next;
8058                 }
8059
8060                 wait_for_completion_io(&done.done);
8061
8062                 if (!done.uptodate) {
8063                         /* We might have another mirror, so try again */
8064                         goto next_block_or_try_again;
8065                 }
8066
8067 next:
8068                 start += sectorsize;
8069
8070                 nr_sectors--;
8071                 if (nr_sectors) {
8072                         pgoff += sectorsize;
8073                         ASSERT(pgoff < PAGE_SIZE);
8074                         goto next_block_or_try_again;
8075                 }
8076         }
8077
8078         return err;
8079 }
8080
8081 static void btrfs_retry_endio(struct bio *bio)
8082 {
8083         struct btrfs_retry_complete *done = bio->bi_private;
8084         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8085         struct extent_io_tree *io_tree, *failure_tree;
8086         struct inode *inode = done->inode;
8087         struct bio_vec *bvec;
8088         int uptodate;
8089         int ret;
8090         int i = 0;
8091         struct bvec_iter_all iter_all;
8092
8093         if (bio->bi_status)
8094                 goto end;
8095
8096         uptodate = 1;
8097
8098         ASSERT(bio->bi_vcnt == 1);
8099         ASSERT(bio_first_bvec_all(bio)->bv_len == btrfs_inode_sectorsize(done->inode));
8100
8101         io_tree = &BTRFS_I(inode)->io_tree;
8102         failure_tree = &BTRFS_I(inode)->io_failure_tree;
8103
8104         ASSERT(!bio_flagged(bio, BIO_CLONED));
8105         bio_for_each_segment_all(bvec, bio, iter_all) {
8106                 ret = __readpage_endio_check(inode, io_bio, i, bvec->bv_page,
8107                                              bvec->bv_offset, done->start,
8108                                              bvec->bv_len);
8109                 if (!ret)
8110                         clean_io_failure(BTRFS_I(inode)->root->fs_info,
8111                                          failure_tree, io_tree, done->start,
8112                                          bvec->bv_page,
8113                                          btrfs_ino(BTRFS_I(inode)),
8114                                          bvec->bv_offset);
8115                 else
8116                         uptodate = 0;
8117                 i++;
8118         }
8119
8120         done->uptodate = uptodate;
8121 end:
8122         complete(&done->done);
8123         bio_put(bio);
8124 }
8125
8126 static blk_status_t __btrfs_subio_endio_read(struct inode *inode,
8127                 struct btrfs_io_bio *io_bio, blk_status_t err)
8128 {
8129         struct btrfs_fs_info *fs_info;
8130         struct bio_vec bvec;
8131         struct bvec_iter iter;
8132         struct btrfs_retry_complete done;
8133         u64 start;
8134         u64 offset = 0;
8135         u32 sectorsize;
8136         int nr_sectors;
8137         unsigned int pgoff;
8138         int csum_pos;
8139         bool uptodate = (err == 0);
8140         int ret;
8141         blk_status_t status;
8142
8143         fs_info = BTRFS_I(inode)->root->fs_info;
8144         sectorsize = fs_info->sectorsize;
8145
8146         err = BLK_STS_OK;
8147         start = io_bio->logical;
8148         done.inode = inode;
8149         io_bio->bio.bi_iter = io_bio->iter;
8150
8151         bio_for_each_segment(bvec, &io_bio->bio, iter) {
8152                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
8153
8154                 pgoff = bvec.bv_offset;
8155 next_block:
8156                 if (uptodate) {
8157                         csum_pos = BTRFS_BYTES_TO_BLKS(fs_info, offset);
8158                         ret = __readpage_endio_check(inode, io_bio, csum_pos,
8159                                         bvec.bv_page, pgoff, start, sectorsize);
8160                         if (likely(!ret))
8161                                 goto next;
8162                 }
8163 try_again:
8164                 done.uptodate = 0;
8165                 done.start = start;
8166                 init_completion(&done.done);
8167
8168                 status = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
8169                                         pgoff, start, start + sectorsize - 1,
8170                                         io_bio->mirror_num, btrfs_retry_endio,
8171                                         &done);
8172                 if (status) {
8173                         err = status;
8174                         goto next;
8175                 }
8176
8177                 wait_for_completion_io(&done.done);
8178
8179                 if (!done.uptodate) {
8180                         /* We might have another mirror, so try again */
8181                         goto try_again;
8182                 }
8183 next:
8184                 offset += sectorsize;
8185                 start += sectorsize;
8186
8187                 ASSERT(nr_sectors);
8188
8189                 nr_sectors--;
8190                 if (nr_sectors) {
8191                         pgoff += sectorsize;
8192                         ASSERT(pgoff < PAGE_SIZE);
8193                         goto next_block;
8194                 }
8195         }
8196
8197         return err;
8198 }
8199
8200 static blk_status_t btrfs_subio_endio_read(struct inode *inode,
8201                 struct btrfs_io_bio *io_bio, blk_status_t err)
8202 {
8203         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8204
8205         if (skip_csum) {
8206                 if (unlikely(err))
8207                         return __btrfs_correct_data_nocsum(inode, io_bio);
8208                 else
8209                         return BLK_STS_OK;
8210         } else {
8211                 return __btrfs_subio_endio_read(inode, io_bio, err);
8212         }
8213 }
8214
8215 static void btrfs_endio_direct_read(struct bio *bio)
8216 {
8217         struct btrfs_dio_private *dip = bio->bi_private;
8218         struct inode *inode = dip->inode;
8219         struct bio *dio_bio;
8220         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8221         blk_status_t err = bio->bi_status;
8222
8223         if (dip->flags & BTRFS_DIO_ORIG_BIO_SUBMITTED)
8224                 err = btrfs_subio_endio_read(inode, io_bio, err);
8225
8226         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
8227                       dip->logical_offset + dip->bytes - 1);
8228         dio_bio = dip->dio_bio;
8229
8230         kfree(dip);
8231
8232         dio_bio->bi_status = err;
8233         dio_end_io(dio_bio);
8234         btrfs_io_bio_free_csum(io_bio);
8235         bio_put(bio);
8236 }
8237
8238 static void __endio_write_update_ordered(struct inode *inode,
8239                                          const u64 offset, const u64 bytes,
8240                                          const bool uptodate)
8241 {
8242         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8243         struct btrfs_ordered_extent *ordered = NULL;
8244         struct btrfs_workqueue *wq;
8245         u64 ordered_offset = offset;
8246         u64 ordered_bytes = bytes;
8247         u64 last_offset;
8248
8249         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
8250                 wq = fs_info->endio_freespace_worker;
8251         else
8252                 wq = fs_info->endio_write_workers;
8253
8254         while (ordered_offset < offset + bytes) {
8255                 last_offset = ordered_offset;
8256                 if (btrfs_dec_test_first_ordered_pending(inode, &ordered,
8257                                                            &ordered_offset,
8258                                                            ordered_bytes,
8259                                                            uptodate)) {
8260                         btrfs_init_work(&ordered->work, finish_ordered_fn, NULL,
8261                                         NULL);
8262                         btrfs_queue_work(wq, &ordered->work);
8263                 }
8264                 /*
8265                  * If btrfs_dec_test_ordered_pending does not find any ordered
8266                  * extent in the range, we can exit.
8267                  */
8268                 if (ordered_offset == last_offset)
8269                         return;
8270                 /*
8271                  * Our bio might span multiple ordered extents. In this case
8272                  * we keep going until we have accounted the whole dio.
8273                  */
8274                 if (ordered_offset < offset + bytes) {
8275                         ordered_bytes = offset + bytes - ordered_offset;
8276                         ordered = NULL;
8277                 }
8278         }
8279 }
8280
8281 static void btrfs_endio_direct_write(struct bio *bio)
8282 {
8283         struct btrfs_dio_private *dip = bio->bi_private;
8284         struct bio *dio_bio = dip->dio_bio;
8285
8286         __endio_write_update_ordered(dip->inode, dip->logical_offset,
8287                                      dip->bytes, !bio->bi_status);
8288
8289         kfree(dip);
8290
8291         dio_bio->bi_status = bio->bi_status;
8292         dio_end_io(dio_bio);
8293         bio_put(bio);
8294 }
8295
8296 static blk_status_t btrfs_submit_bio_start_direct_io(void *private_data,
8297                                     struct bio *bio, u64 offset)
8298 {
8299         struct inode *inode = private_data;
8300         blk_status_t ret;
8301         ret = btrfs_csum_one_bio(inode, bio, offset, 1);
8302         BUG_ON(ret); /* -ENOMEM */
8303         return 0;
8304 }
8305
8306 static void btrfs_end_dio_bio(struct bio *bio)
8307 {
8308         struct btrfs_dio_private *dip = bio->bi_private;
8309         blk_status_t err = bio->bi_status;
8310
8311         if (err)
8312                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
8313                            "direct IO failed ino %llu rw %d,%u sector %#Lx len %u err no %d",
8314                            btrfs_ino(BTRFS_I(dip->inode)), bio_op(bio),
8315                            bio->bi_opf,
8316                            (unsigned long long)bio->bi_iter.bi_sector,
8317                            bio->bi_iter.bi_size, err);
8318
8319         if (dip->subio_endio)
8320                 err = dip->subio_endio(dip->inode, btrfs_io_bio(bio), err);
8321
8322         if (err) {
8323                 /*
8324                  * We want to perceive the errors flag being set before
8325                  * decrementing the reference count. We don't need a barrier
8326                  * since atomic operations with a return value are fully
8327                  * ordered as per atomic_t.txt
8328                  */
8329                 dip->errors = 1;
8330         }
8331
8332         /* if there are more bios still pending for this dio, just exit */
8333         if (!atomic_dec_and_test(&dip->pending_bios))
8334                 goto out;
8335
8336         if (dip->errors) {
8337                 bio_io_error(dip->orig_bio);
8338         } else {
8339                 dip->dio_bio->bi_status = BLK_STS_OK;
8340                 bio_endio(dip->orig_bio);
8341         }
8342 out:
8343         bio_put(bio);
8344 }
8345
8346 static inline blk_status_t btrfs_lookup_and_bind_dio_csum(struct inode *inode,
8347                                                  struct btrfs_dio_private *dip,
8348                                                  struct bio *bio,
8349                                                  u64 file_offset)
8350 {
8351         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8352         struct btrfs_io_bio *orig_io_bio = btrfs_io_bio(dip->orig_bio);
8353         blk_status_t ret;
8354
8355         /*
8356          * We load all the csum data we need when we submit
8357          * the first bio to reduce the csum tree search and
8358          * contention.
8359          */
8360         if (dip->logical_offset == file_offset) {
8361                 ret = btrfs_lookup_bio_sums_dio(inode, dip->orig_bio,
8362                                                 file_offset);
8363                 if (ret)
8364                         return ret;
8365         }
8366
8367         if (bio == dip->orig_bio)
8368                 return 0;
8369
8370         file_offset -= dip->logical_offset;
8371         file_offset >>= inode->i_sb->s_blocksize_bits;
8372         io_bio->csum = (u8 *)(((u32 *)orig_io_bio->csum) + file_offset);
8373
8374         return 0;
8375 }
8376
8377 static inline blk_status_t btrfs_submit_dio_bio(struct bio *bio,
8378                 struct inode *inode, u64 file_offset, int async_submit)
8379 {
8380         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8381         struct btrfs_dio_private *dip = bio->bi_private;
8382         bool write = bio_op(bio) == REQ_OP_WRITE;
8383         blk_status_t ret;
8384
8385         /* Check btrfs_submit_bio_hook() for rules about async submit. */
8386         if (async_submit)
8387                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
8388
8389         if (!write) {
8390                 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
8391                 if (ret)
8392                         goto err;
8393         }
8394
8395         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
8396                 goto map;
8397
8398         if (write && async_submit) {
8399                 ret = btrfs_wq_submit_bio(fs_info, bio, 0, 0,
8400                                           file_offset, inode,
8401                                           btrfs_submit_bio_start_direct_io);
8402                 goto err;
8403         } else if (write) {
8404                 /*
8405                  * If we aren't doing async submit, calculate the csum of the
8406                  * bio now.
8407                  */
8408                 ret = btrfs_csum_one_bio(inode, bio, file_offset, 1);
8409                 if (ret)
8410                         goto err;
8411         } else {
8412                 ret = btrfs_lookup_and_bind_dio_csum(inode, dip, bio,
8413                                                      file_offset);
8414                 if (ret)
8415                         goto err;
8416         }
8417 map:
8418         ret = btrfs_map_bio(fs_info, bio, 0);
8419 err:
8420         return ret;
8421 }
8422
8423 static int btrfs_submit_direct_hook(struct btrfs_dio_private *dip)
8424 {
8425         struct inode *inode = dip->inode;
8426         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8427         struct bio *bio;
8428         struct bio *orig_bio = dip->orig_bio;
8429         u64 start_sector = orig_bio->bi_iter.bi_sector;
8430         u64 file_offset = dip->logical_offset;
8431         int async_submit = 0;
8432         u64 submit_len;
8433         int clone_offset = 0;
8434         int clone_len;
8435         int ret;
8436         blk_status_t status;
8437         struct btrfs_io_geometry geom;
8438
8439         submit_len = orig_bio->bi_iter.bi_size;
8440         ret = btrfs_get_io_geometry(fs_info, btrfs_op(orig_bio),
8441                                     start_sector << 9, submit_len, &geom);
8442         if (ret)
8443                 return -EIO;
8444
8445         if (geom.len >= submit_len) {
8446                 bio = orig_bio;
8447                 dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
8448                 goto submit;
8449         }
8450
8451         /* async crcs make it difficult to collect full stripe writes. */
8452         if (btrfs_data_alloc_profile(fs_info) & BTRFS_BLOCK_GROUP_RAID56_MASK)
8453                 async_submit = 0;
8454         else
8455                 async_submit = 1;
8456
8457         /* bio split */
8458         ASSERT(geom.len <= INT_MAX);
8459         atomic_inc(&dip->pending_bios);
8460         do {
8461                 clone_len = min_t(int, submit_len, geom.len);
8462
8463                 /*
8464                  * This will never fail as it's passing GPF_NOFS and
8465                  * the allocation is backed by btrfs_bioset.
8466                  */
8467                 bio = btrfs_bio_clone_partial(orig_bio, clone_offset,
8468                                               clone_len);
8469                 bio->bi_private = dip;
8470                 bio->bi_end_io = btrfs_end_dio_bio;
8471                 btrfs_io_bio(bio)->logical = file_offset;
8472
8473                 ASSERT(submit_len >= clone_len);
8474                 submit_len -= clone_len;
8475                 if (submit_len == 0)
8476                         break;
8477
8478                 /*
8479                  * Increase the count before we submit the bio so we know
8480                  * the end IO handler won't happen before we increase the
8481                  * count. Otherwise, the dip might get freed before we're
8482                  * done setting it up.
8483                  */
8484                 atomic_inc(&dip->pending_bios);
8485
8486                 status = btrfs_submit_dio_bio(bio, inode, file_offset,
8487                                                 async_submit);
8488                 if (status) {
8489                         bio_put(bio);
8490                         atomic_dec(&dip->pending_bios);
8491                         goto out_err;
8492                 }
8493
8494                 clone_offset += clone_len;
8495                 start_sector += clone_len >> 9;
8496                 file_offset += clone_len;
8497
8498                 ret = btrfs_get_io_geometry(fs_info, btrfs_op(orig_bio),
8499                                       start_sector << 9, submit_len, &geom);
8500                 if (ret)
8501                         goto out_err;
8502         } while (submit_len > 0);
8503
8504 submit:
8505         status = btrfs_submit_dio_bio(bio, inode, file_offset, async_submit);
8506         if (!status)
8507                 return 0;
8508
8509         bio_put(bio);
8510 out_err:
8511         dip->errors = 1;
8512         /*
8513          * Before atomic variable goto zero, we must  make sure dip->errors is
8514          * perceived to be set. This ordering is ensured by the fact that an
8515          * atomic operations with a return value are fully ordered as per
8516          * atomic_t.txt
8517          */
8518         if (atomic_dec_and_test(&dip->pending_bios))
8519                 bio_io_error(dip->orig_bio);
8520
8521         /* bio_end_io() will handle error, so we needn't return it */
8522         return 0;
8523 }
8524
8525 static void btrfs_submit_direct(struct bio *dio_bio, struct inode *inode,
8526                                 loff_t file_offset)
8527 {
8528         struct btrfs_dio_private *dip = NULL;
8529         struct bio *bio = NULL;
8530         struct btrfs_io_bio *io_bio;
8531         bool write = (bio_op(dio_bio) == REQ_OP_WRITE);
8532         int ret = 0;
8533
8534         bio = btrfs_bio_clone(dio_bio);
8535
8536         dip = kzalloc(sizeof(*dip), GFP_NOFS);
8537         if (!dip) {
8538                 ret = -ENOMEM;
8539                 goto free_ordered;
8540         }
8541
8542         dip->private = dio_bio->bi_private;
8543         dip->inode = inode;
8544         dip->logical_offset = file_offset;
8545         dip->bytes = dio_bio->bi_iter.bi_size;
8546         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
8547         bio->bi_private = dip;
8548         dip->orig_bio = bio;
8549         dip->dio_bio = dio_bio;
8550         atomic_set(&dip->pending_bios, 0);
8551         io_bio = btrfs_io_bio(bio);
8552         io_bio->logical = file_offset;
8553
8554         if (write) {
8555                 bio->bi_end_io = btrfs_endio_direct_write;
8556         } else {
8557                 bio->bi_end_io = btrfs_endio_direct_read;
8558                 dip->subio_endio = btrfs_subio_endio_read;
8559         }
8560
8561         /*
8562          * Reset the range for unsubmitted ordered extents (to a 0 length range)
8563          * even if we fail to submit a bio, because in such case we do the
8564          * corresponding error handling below and it must not be done a second
8565          * time by btrfs_direct_IO().
8566          */
8567         if (write) {
8568                 struct btrfs_dio_data *dio_data = current->journal_info;
8569
8570                 dio_data->unsubmitted_oe_range_end = dip->logical_offset +
8571                         dip->bytes;
8572                 dio_data->unsubmitted_oe_range_start =
8573                         dio_data->unsubmitted_oe_range_end;
8574         }
8575
8576         ret = btrfs_submit_direct_hook(dip);
8577         if (!ret)
8578                 return;
8579
8580         btrfs_io_bio_free_csum(io_bio);
8581
8582 free_ordered:
8583         /*
8584          * If we arrived here it means either we failed to submit the dip
8585          * or we either failed to clone the dio_bio or failed to allocate the
8586          * dip. If we cloned the dio_bio and allocated the dip, we can just
8587          * call bio_endio against our io_bio so that we get proper resource
8588          * cleanup if we fail to submit the dip, otherwise, we must do the
8589          * same as btrfs_endio_direct_[write|read] because we can't call these
8590          * callbacks - they require an allocated dip and a clone of dio_bio.
8591          */
8592         if (bio && dip) {
8593                 bio_io_error(bio);
8594                 /*
8595                  * The end io callbacks free our dip, do the final put on bio
8596                  * and all the cleanup and final put for dio_bio (through
8597                  * dio_end_io()).
8598                  */
8599                 dip = NULL;
8600                 bio = NULL;
8601         } else {
8602                 if (write)
8603                         __endio_write_update_ordered(inode,
8604                                                 file_offset,
8605                                                 dio_bio->bi_iter.bi_size,
8606                                                 false);
8607                 else
8608                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
8609                               file_offset + dio_bio->bi_iter.bi_size - 1);
8610
8611                 dio_bio->bi_status = BLK_STS_IOERR;
8612                 /*
8613                  * Releases and cleans up our dio_bio, no need to bio_put()
8614                  * nor bio_endio()/bio_io_error() against dio_bio.
8615                  */
8616                 dio_end_io(dio_bio);
8617         }
8618         if (bio)
8619                 bio_put(bio);
8620         kfree(dip);
8621 }
8622
8623 static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
8624                                const struct iov_iter *iter, loff_t offset)
8625 {
8626         int seg;
8627         int i;
8628         unsigned int blocksize_mask = fs_info->sectorsize - 1;
8629         ssize_t retval = -EINVAL;
8630
8631         if (offset & blocksize_mask)
8632                 goto out;
8633
8634         if (iov_iter_alignment(iter) & blocksize_mask)
8635                 goto out;
8636
8637         /* If this is a write we don't need to check anymore */
8638         if (iov_iter_rw(iter) != READ || !iter_is_iovec(iter))
8639                 return 0;
8640         /*
8641          * Check to make sure we don't have duplicate iov_base's in this
8642          * iovec, if so return EINVAL, otherwise we'll get csum errors
8643          * when reading back.
8644          */
8645         for (seg = 0; seg < iter->nr_segs; seg++) {
8646                 for (i = seg + 1; i < iter->nr_segs; i++) {
8647                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
8648                                 goto out;
8649                 }
8650         }
8651         retval = 0;
8652 out:
8653         return retval;
8654 }
8655
8656 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
8657 {
8658         struct file *file = iocb->ki_filp;
8659         struct inode *inode = file->f_mapping->host;
8660         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8661         struct btrfs_dio_data dio_data = { 0 };
8662         struct extent_changeset *data_reserved = NULL;
8663         loff_t offset = iocb->ki_pos;
8664         size_t count = 0;
8665         int flags = 0;
8666         bool wakeup = true;
8667         bool relock = false;
8668         ssize_t ret;
8669
8670         if (check_direct_IO(fs_info, iter, offset))
8671                 return 0;
8672
8673         inode_dio_begin(inode);
8674
8675         /*
8676          * The generic stuff only does filemap_write_and_wait_range, which
8677          * isn't enough if we've written compressed pages to this area, so
8678          * we need to flush the dirty pages again to make absolutely sure
8679          * that any outstanding dirty pages are on disk.
8680          */
8681         count = iov_iter_count(iter);
8682         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8683                      &BTRFS_I(inode)->runtime_flags))
8684                 filemap_fdatawrite_range(inode->i_mapping, offset,
8685                                          offset + count - 1);
8686
8687         if (iov_iter_rw(iter) == WRITE) {
8688                 /*
8689                  * If the write DIO is beyond the EOF, we need update
8690                  * the isize, but it is protected by i_mutex. So we can
8691                  * not unlock the i_mutex at this case.
8692                  */
8693                 if (offset + count <= inode->i_size) {
8694                         dio_data.overwrite = 1;
8695                         inode_unlock(inode);
8696                         relock = true;
8697                 } else if (iocb->ki_flags & IOCB_NOWAIT) {
8698                         ret = -EAGAIN;
8699                         goto out;
8700                 }
8701                 ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
8702                                                    offset, count);
8703                 if (ret)
8704                         goto out;
8705
8706                 /*
8707                  * We need to know how many extents we reserved so that we can
8708                  * do the accounting properly if we go over the number we
8709                  * originally calculated.  Abuse current->journal_info for this.
8710                  */
8711                 dio_data.reserve = round_up(count,
8712                                             fs_info->sectorsize);
8713                 dio_data.unsubmitted_oe_range_start = (u64)offset;
8714                 dio_data.unsubmitted_oe_range_end = (u64)offset;
8715                 current->journal_info = &dio_data;
8716                 down_read(&BTRFS_I(inode)->dio_sem);
8717         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
8718                                      &BTRFS_I(inode)->runtime_flags)) {
8719                 inode_dio_end(inode);
8720                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
8721                 wakeup = false;
8722         }
8723
8724         ret = __blockdev_direct_IO(iocb, inode,
8725                                    fs_info->fs_devices->latest_bdev,
8726                                    iter, btrfs_get_blocks_direct, NULL,
8727                                    btrfs_submit_direct, flags);
8728         if (iov_iter_rw(iter) == WRITE) {
8729                 up_read(&BTRFS_I(inode)->dio_sem);
8730                 current->journal_info = NULL;
8731                 if (ret < 0 && ret != -EIOCBQUEUED) {
8732                         if (dio_data.reserve)
8733                                 btrfs_delalloc_release_space(inode, data_reserved,
8734                                         offset, dio_data.reserve, true);
8735                         /*
8736                          * On error we might have left some ordered extents
8737                          * without submitting corresponding bios for them, so
8738                          * cleanup them up to avoid other tasks getting them
8739                          * and waiting for them to complete forever.
8740                          */
8741                         if (dio_data.unsubmitted_oe_range_start <
8742                             dio_data.unsubmitted_oe_range_end)
8743                                 __endio_write_update_ordered(inode,
8744                                         dio_data.unsubmitted_oe_range_start,
8745                                         dio_data.unsubmitted_oe_range_end -
8746                                         dio_data.unsubmitted_oe_range_start,
8747                                         false);
8748                 } else if (ret >= 0 && (size_t)ret < count)
8749                         btrfs_delalloc_release_space(inode, data_reserved,
8750                                         offset, count - (size_t)ret, true);
8751                 btrfs_delalloc_release_extents(BTRFS_I(inode), count);
8752         }
8753 out:
8754         if (wakeup)
8755                 inode_dio_end(inode);
8756         if (relock)
8757                 inode_lock(inode);
8758
8759         extent_changeset_free(data_reserved);
8760         return ret;
8761 }
8762
8763 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
8764
8765 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
8766                 __u64 start, __u64 len)
8767 {
8768         int     ret;
8769
8770         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
8771         if (ret)
8772                 return ret;
8773
8774         return extent_fiemap(inode, fieinfo, start, len);
8775 }
8776
8777 int btrfs_readpage(struct file *file, struct page *page)
8778 {
8779         struct extent_io_tree *tree;
8780         tree = &BTRFS_I(page->mapping->host)->io_tree;
8781         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
8782 }
8783
8784 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
8785 {
8786         struct inode *inode = page->mapping->host;
8787         int ret;
8788
8789         if (current->flags & PF_MEMALLOC) {
8790                 redirty_page_for_writepage(wbc, page);
8791                 unlock_page(page);
8792                 return 0;
8793         }
8794
8795         /*
8796          * If we are under memory pressure we will call this directly from the
8797          * VM, we need to make sure we have the inode referenced for the ordered
8798          * extent.  If not just return like we didn't do anything.
8799          */
8800         if (!igrab(inode)) {
8801                 redirty_page_for_writepage(wbc, page);
8802                 return AOP_WRITEPAGE_ACTIVATE;
8803         }
8804         ret = extent_write_full_page(page, wbc);
8805         btrfs_add_delayed_iput(inode);
8806         return ret;
8807 }
8808
8809 static int btrfs_writepages(struct address_space *mapping,
8810                             struct writeback_control *wbc)
8811 {
8812         return extent_writepages(mapping, wbc);
8813 }
8814
8815 static int
8816 btrfs_readpages(struct file *file, struct address_space *mapping,
8817                 struct list_head *pages, unsigned nr_pages)
8818 {
8819         return extent_readpages(mapping, pages, nr_pages);
8820 }
8821
8822 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8823 {
8824         int ret = try_release_extent_mapping(page, gfp_flags);
8825         if (ret == 1) {
8826                 ClearPagePrivate(page);
8827                 set_page_private(page, 0);
8828                 put_page(page);
8829         }
8830         return ret;
8831 }
8832
8833 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8834 {
8835         if (PageWriteback(page) || PageDirty(page))
8836                 return 0;
8837         return __btrfs_releasepage(page, gfp_flags);
8838 }
8839
8840 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
8841                                  unsigned int length)
8842 {
8843         struct inode *inode = page->mapping->host;
8844         struct extent_io_tree *tree;
8845         struct btrfs_ordered_extent *ordered;
8846         struct extent_state *cached_state = NULL;
8847         u64 page_start = page_offset(page);
8848         u64 page_end = page_start + PAGE_SIZE - 1;
8849         u64 start;
8850         u64 end;
8851         int inode_evicting = inode->i_state & I_FREEING;
8852
8853         /*
8854          * we have the page locked, so new writeback can't start,
8855          * and the dirty bit won't be cleared while we are here.
8856          *
8857          * Wait for IO on this page so that we can safely clear
8858          * the PagePrivate2 bit and do ordered accounting
8859          */
8860         wait_on_page_writeback(page);
8861
8862         tree = &BTRFS_I(inode)->io_tree;
8863         if (offset) {
8864                 btrfs_releasepage(page, GFP_NOFS);
8865                 return;
8866         }
8867
8868         if (!inode_evicting)
8869                 lock_extent_bits(tree, page_start, page_end, &cached_state);
8870 again:
8871         start = page_start;
8872         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
8873                                         page_end - start + 1);
8874         if (ordered) {
8875                 end = min(page_end, ordered->file_offset + ordered->len - 1);
8876                 /*
8877                  * IO on this page will never be started, so we need
8878                  * to account for any ordered extents now
8879                  */
8880                 if (!inode_evicting)
8881                         clear_extent_bit(tree, start, end,
8882                                          EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
8883                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
8884                                          EXTENT_DEFRAG, 1, 0, &cached_state);
8885                 /*
8886                  * whoever cleared the private bit is responsible
8887                  * for the finish_ordered_io
8888                  */
8889                 if (TestClearPagePrivate2(page)) {
8890                         struct btrfs_ordered_inode_tree *tree;
8891                         u64 new_len;
8892
8893                         tree = &BTRFS_I(inode)->ordered_tree;
8894
8895                         spin_lock_irq(&tree->lock);
8896                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
8897                         new_len = start - ordered->file_offset;
8898                         if (new_len < ordered->truncated_len)
8899                                 ordered->truncated_len = new_len;
8900                         spin_unlock_irq(&tree->lock);
8901
8902                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
8903                                                            start,
8904                                                            end - start + 1, 1))
8905                                 btrfs_finish_ordered_io(ordered);
8906                 }
8907                 btrfs_put_ordered_extent(ordered);
8908                 if (!inode_evicting) {
8909                         cached_state = NULL;
8910                         lock_extent_bits(tree, start, end,
8911                                          &cached_state);
8912                 }
8913
8914                 start = end + 1;
8915                 if (start < page_end)
8916                         goto again;
8917         }
8918
8919         /*
8920          * Qgroup reserved space handler
8921          * Page here will be either
8922          * 1) Already written to disk
8923          *    In this case, its reserved space is released from data rsv map
8924          *    and will be freed by delayed_ref handler finally.
8925          *    So even we call qgroup_free_data(), it won't decrease reserved
8926          *    space.
8927          * 2) Not written to disk
8928          *    This means the reserved space should be freed here. However,
8929          *    if a truncate invalidates the page (by clearing PageDirty)
8930          *    and the page is accounted for while allocating extent
8931          *    in btrfs_check_data_free_space() we let delayed_ref to
8932          *    free the entire extent.
8933          */
8934         if (PageDirty(page))
8935                 btrfs_qgroup_free_data(inode, NULL, page_start, PAGE_SIZE);
8936         if (!inode_evicting) {
8937                 clear_extent_bit(tree, page_start, page_end, EXTENT_LOCKED |
8938                                  EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
8939                                  EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 1, 1,
8940                                  &cached_state);
8941
8942                 __btrfs_releasepage(page, GFP_NOFS);
8943         }
8944
8945         ClearPageChecked(page);
8946         if (PagePrivate(page)) {
8947                 ClearPagePrivate(page);
8948                 set_page_private(page, 0);
8949                 put_page(page);
8950         }
8951 }
8952
8953 /*
8954  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
8955  * called from a page fault handler when a page is first dirtied. Hence we must
8956  * be careful to check for EOF conditions here. We set the page up correctly
8957  * for a written page which means we get ENOSPC checking when writing into
8958  * holes and correct delalloc and unwritten extent mapping on filesystems that
8959  * support these features.
8960  *
8961  * We are not allowed to take the i_mutex here so we have to play games to
8962  * protect against truncate races as the page could now be beyond EOF.  Because
8963  * truncate_setsize() writes the inode size before removing pages, once we have
8964  * the page lock we can determine safely if the page is beyond EOF. If it is not
8965  * beyond EOF, then the page is guaranteed safe against truncation until we
8966  * unlock the page.
8967  */
8968 vm_fault_t btrfs_page_mkwrite(struct vm_fault *vmf)
8969 {
8970         struct page *page = vmf->page;
8971         struct inode *inode = file_inode(vmf->vma->vm_file);
8972         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8973         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8974         struct btrfs_ordered_extent *ordered;
8975         struct extent_state *cached_state = NULL;
8976         struct extent_changeset *data_reserved = NULL;
8977         char *kaddr;
8978         unsigned long zero_start;
8979         loff_t size;
8980         vm_fault_t ret;
8981         int ret2;
8982         int reserved = 0;
8983         u64 reserved_space;
8984         u64 page_start;
8985         u64 page_end;
8986         u64 end;
8987
8988         reserved_space = PAGE_SIZE;
8989
8990         sb_start_pagefault(inode->i_sb);
8991         page_start = page_offset(page);
8992         page_end = page_start + PAGE_SIZE - 1;
8993         end = page_end;
8994
8995         /*
8996          * Reserving delalloc space after obtaining the page lock can lead to
8997          * deadlock. For example, if a dirty page is locked by this function
8998          * and the call to btrfs_delalloc_reserve_space() ends up triggering
8999          * dirty page write out, then the btrfs_writepage() function could
9000          * end up waiting indefinitely to get a lock on the page currently
9001          * being processed by btrfs_page_mkwrite() function.
9002          */
9003         ret2 = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
9004                                            reserved_space);
9005         if (!ret2) {
9006                 ret2 = file_update_time(vmf->vma->vm_file);
9007                 reserved = 1;
9008         }
9009         if (ret2) {
9010                 ret = vmf_error(ret2);
9011                 if (reserved)
9012                         goto out;
9013                 goto out_noreserve;
9014         }
9015
9016         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
9017 again:
9018         lock_page(page);
9019         size = i_size_read(inode);
9020
9021         if ((page->mapping != inode->i_mapping) ||
9022             (page_start >= size)) {
9023                 /* page got truncated out from underneath us */
9024                 goto out_unlock;
9025         }
9026         wait_on_page_writeback(page);
9027
9028         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
9029         set_page_extent_mapped(page);
9030
9031         /*
9032          * we can't set the delalloc bits if there are pending ordered
9033          * extents.  Drop our locks and wait for them to finish
9034          */
9035         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
9036                         PAGE_SIZE);
9037         if (ordered) {
9038                 unlock_extent_cached(io_tree, page_start, page_end,
9039                                      &cached_state);
9040                 unlock_page(page);
9041                 btrfs_start_ordered_extent(inode, ordered, 1);
9042                 btrfs_put_ordered_extent(ordered);
9043                 goto again;
9044         }
9045
9046         if (page->index == ((size - 1) >> PAGE_SHIFT)) {
9047                 reserved_space = round_up(size - page_start,
9048                                           fs_info->sectorsize);
9049                 if (reserved_space < PAGE_SIZE) {
9050                         end = page_start + reserved_space - 1;
9051                         btrfs_delalloc_release_space(inode, data_reserved,
9052                                         page_start, PAGE_SIZE - reserved_space,
9053                                         true);
9054                 }
9055         }
9056
9057         /*
9058          * page_mkwrite gets called when the page is firstly dirtied after it's
9059          * faulted in, but write(2) could also dirty a page and set delalloc
9060          * bits, thus in this case for space account reason, we still need to
9061          * clear any delalloc bits within this page range since we have to
9062          * reserve data&meta space before lock_page() (see above comments).
9063          */
9064         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, end,
9065                           EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
9066                           EXTENT_DEFRAG, 0, 0, &cached_state);
9067
9068         ret2 = btrfs_set_extent_delalloc(inode, page_start, end, 0,
9069                                         &cached_state);
9070         if (ret2) {
9071                 unlock_extent_cached(io_tree, page_start, page_end,
9072                                      &cached_state);
9073                 ret = VM_FAULT_SIGBUS;
9074                 goto out_unlock;
9075         }
9076         ret2 = 0;
9077
9078         /* page is wholly or partially inside EOF */
9079         if (page_start + PAGE_SIZE > size)
9080                 zero_start = offset_in_page(size);
9081         else
9082                 zero_start = PAGE_SIZE;
9083
9084         if (zero_start != PAGE_SIZE) {
9085                 kaddr = kmap(page);
9086                 memset(kaddr + zero_start, 0, PAGE_SIZE - zero_start);
9087                 flush_dcache_page(page);
9088                 kunmap(page);
9089         }
9090         ClearPageChecked(page);
9091         set_page_dirty(page);
9092         SetPageUptodate(page);
9093
9094         BTRFS_I(inode)->last_trans = fs_info->generation;
9095         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
9096         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
9097
9098         unlock_extent_cached(io_tree, page_start, page_end, &cached_state);
9099
9100         if (!ret2) {
9101                 btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
9102                 sb_end_pagefault(inode->i_sb);
9103                 extent_changeset_free(data_reserved);
9104                 return VM_FAULT_LOCKED;
9105         }
9106
9107 out_unlock:
9108         unlock_page(page);
9109 out:
9110         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
9111         btrfs_delalloc_release_space(inode, data_reserved, page_start,
9112                                      reserved_space, (ret != 0));
9113 out_noreserve:
9114         sb_end_pagefault(inode->i_sb);
9115         extent_changeset_free(data_reserved);
9116         return ret;
9117 }
9118
9119 static int btrfs_truncate(struct inode *inode, bool skip_writeback)
9120 {
9121         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9122         struct btrfs_root *root = BTRFS_I(inode)->root;
9123         struct btrfs_block_rsv *rsv;
9124         int ret;
9125         struct btrfs_trans_handle *trans;
9126         u64 mask = fs_info->sectorsize - 1;
9127         u64 min_size = btrfs_calc_metadata_size(fs_info, 1);
9128
9129         if (!skip_writeback) {
9130                 ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
9131                                                (u64)-1);
9132                 if (ret)
9133                         return ret;
9134         }
9135
9136         /*
9137          * Yes ladies and gentlemen, this is indeed ugly.  We have a couple of
9138          * things going on here:
9139          *
9140          * 1) We need to reserve space to update our inode.
9141          *
9142          * 2) We need to have something to cache all the space that is going to
9143          * be free'd up by the truncate operation, but also have some slack
9144          * space reserved in case it uses space during the truncate (thank you
9145          * very much snapshotting).
9146          *
9147          * And we need these to be separate.  The fact is we can use a lot of
9148          * space doing the truncate, and we have no earthly idea how much space
9149          * we will use, so we need the truncate reservation to be separate so it
9150          * doesn't end up using space reserved for updating the inode.  We also
9151          * need to be able to stop the transaction and start a new one, which
9152          * means we need to be able to update the inode several times, and we
9153          * have no idea of knowing how many times that will be, so we can't just
9154          * reserve 1 item for the entirety of the operation, so that has to be
9155          * done separately as well.
9156          *
9157          * So that leaves us with
9158          *
9159          * 1) rsv - for the truncate reservation, which we will steal from the
9160          * transaction reservation.
9161          * 2) fs_info->trans_block_rsv - this will have 1 items worth left for
9162          * updating the inode.
9163          */
9164         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
9165         if (!rsv)
9166                 return -ENOMEM;
9167         rsv->size = min_size;
9168         rsv->failfast = 1;
9169
9170         /*
9171          * 1 for the truncate slack space
9172          * 1 for updating the inode.
9173          */
9174         trans = btrfs_start_transaction(root, 2);
9175         if (IS_ERR(trans)) {
9176                 ret = PTR_ERR(trans);
9177                 goto out;
9178         }
9179
9180         /* Migrate the slack space for the truncate to our reserve */
9181         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
9182                                       min_size, false);
9183         BUG_ON(ret);
9184
9185         /*
9186          * So if we truncate and then write and fsync we normally would just
9187          * write the extents that changed, which is a problem if we need to
9188          * first truncate that entire inode.  So set this flag so we write out
9189          * all of the extents in the inode to the sync log so we're completely
9190          * safe.
9191          */
9192         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
9193         trans->block_rsv = rsv;
9194
9195         while (1) {
9196                 ret = btrfs_truncate_inode_items(trans, root, inode,
9197                                                  inode->i_size,
9198                                                  BTRFS_EXTENT_DATA_KEY);
9199                 trans->block_rsv = &fs_info->trans_block_rsv;
9200                 if (ret != -ENOSPC && ret != -EAGAIN)
9201                         break;
9202
9203                 ret = btrfs_update_inode(trans, root, inode);
9204                 if (ret)
9205                         break;
9206
9207                 btrfs_end_transaction(trans);
9208                 btrfs_btree_balance_dirty(fs_info);
9209
9210                 trans = btrfs_start_transaction(root, 2);
9211                 if (IS_ERR(trans)) {
9212                         ret = PTR_ERR(trans);
9213                         trans = NULL;
9214                         break;
9215                 }
9216
9217                 btrfs_block_rsv_release(fs_info, rsv, -1);
9218                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
9219                                               rsv, min_size, false);
9220                 BUG_ON(ret);    /* shouldn't happen */
9221                 trans->block_rsv = rsv;
9222         }
9223
9224         /*
9225          * We can't call btrfs_truncate_block inside a trans handle as we could
9226          * deadlock with freeze, if we got NEED_TRUNCATE_BLOCK then we know
9227          * we've truncated everything except the last little bit, and can do
9228          * btrfs_truncate_block and then update the disk_i_size.
9229          */
9230         if (ret == NEED_TRUNCATE_BLOCK) {
9231                 btrfs_end_transaction(trans);
9232                 btrfs_btree_balance_dirty(fs_info);
9233
9234                 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
9235                 if (ret)
9236                         goto out;
9237                 trans = btrfs_start_transaction(root, 1);
9238                 if (IS_ERR(trans)) {
9239                         ret = PTR_ERR(trans);
9240                         goto out;
9241                 }
9242                 btrfs_ordered_update_i_size(inode, inode->i_size, NULL);
9243         }
9244
9245         if (trans) {
9246                 int ret2;
9247
9248                 trans->block_rsv = &fs_info->trans_block_rsv;
9249                 ret2 = btrfs_update_inode(trans, root, inode);
9250                 if (ret2 && !ret)
9251                         ret = ret2;
9252
9253                 ret2 = btrfs_end_transaction(trans);
9254                 if (ret2 && !ret)
9255                         ret = ret2;
9256                 btrfs_btree_balance_dirty(fs_info);
9257         }
9258 out:
9259         btrfs_free_block_rsv(fs_info, rsv);
9260
9261         return ret;
9262 }
9263
9264 /*
9265  * create a new subvolume directory/inode (helper for the ioctl).
9266  */
9267 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
9268                              struct btrfs_root *new_root,
9269                              struct btrfs_root *parent_root,
9270                              u64 new_dirid)
9271 {
9272         struct inode *inode;
9273         int err;
9274         u64 index = 0;
9275
9276         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
9277                                 new_dirid, new_dirid,
9278                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
9279                                 &index);
9280         if (IS_ERR(inode))
9281                 return PTR_ERR(inode);
9282         inode->i_op = &btrfs_dir_inode_operations;
9283         inode->i_fop = &btrfs_dir_file_operations;
9284
9285         set_nlink(inode, 1);
9286         btrfs_i_size_write(BTRFS_I(inode), 0);
9287         unlock_new_inode(inode);
9288
9289         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
9290         if (err)
9291                 btrfs_err(new_root->fs_info,
9292                           "error inheriting subvolume %llu properties: %d",
9293                           new_root->root_key.objectid, err);
9294
9295         err = btrfs_update_inode(trans, new_root, inode);
9296
9297         iput(inode);
9298         return err;
9299 }
9300
9301 struct inode *btrfs_alloc_inode(struct super_block *sb)
9302 {
9303         struct btrfs_fs_info *fs_info = btrfs_sb(sb);
9304         struct btrfs_inode *ei;
9305         struct inode *inode;
9306
9307         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_KERNEL);
9308         if (!ei)
9309                 return NULL;
9310
9311         ei->root = NULL;
9312         ei->generation = 0;
9313         ei->last_trans = 0;
9314         ei->last_sub_trans = 0;
9315         ei->logged_trans = 0;
9316         ei->delalloc_bytes = 0;
9317         ei->new_delalloc_bytes = 0;
9318         ei->defrag_bytes = 0;
9319         ei->disk_i_size = 0;
9320         ei->flags = 0;
9321         ei->csum_bytes = 0;
9322         ei->index_cnt = (u64)-1;
9323         ei->dir_index = 0;
9324         ei->last_unlink_trans = 0;
9325         ei->last_log_commit = 0;
9326
9327         spin_lock_init(&ei->lock);
9328         ei->outstanding_extents = 0;
9329         if (sb->s_magic != BTRFS_TEST_MAGIC)
9330                 btrfs_init_metadata_block_rsv(fs_info, &ei->block_rsv,
9331                                               BTRFS_BLOCK_RSV_DELALLOC);
9332         ei->runtime_flags = 0;
9333         ei->prop_compress = BTRFS_COMPRESS_NONE;
9334         ei->defrag_compress = BTRFS_COMPRESS_NONE;
9335
9336         ei->delayed_node = NULL;
9337
9338         ei->i_otime.tv_sec = 0;
9339         ei->i_otime.tv_nsec = 0;
9340
9341         inode = &ei->vfs_inode;
9342         extent_map_tree_init(&ei->extent_tree);
9343         extent_io_tree_init(fs_info, &ei->io_tree, IO_TREE_INODE_IO, inode);
9344         extent_io_tree_init(fs_info, &ei->io_failure_tree,
9345                             IO_TREE_INODE_IO_FAILURE, inode);
9346         ei->io_tree.track_uptodate = true;
9347         ei->io_failure_tree.track_uptodate = true;
9348         atomic_set(&ei->sync_writers, 0);
9349         mutex_init(&ei->log_mutex);
9350         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
9351         INIT_LIST_HEAD(&ei->delalloc_inodes);
9352         INIT_LIST_HEAD(&ei->delayed_iput);
9353         RB_CLEAR_NODE(&ei->rb_node);
9354         init_rwsem(&ei->dio_sem);
9355
9356         return inode;
9357 }
9358
9359 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
9360 void btrfs_test_destroy_inode(struct inode *inode)
9361 {
9362         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9363         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9364 }
9365 #endif
9366
9367 void btrfs_free_inode(struct inode *inode)
9368 {
9369         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9370 }
9371
9372 void btrfs_destroy_inode(struct inode *inode)
9373 {
9374         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9375         struct btrfs_ordered_extent *ordered;
9376         struct btrfs_root *root = BTRFS_I(inode)->root;
9377
9378         WARN_ON(!hlist_empty(&inode->i_dentry));
9379         WARN_ON(inode->i_data.nrpages);
9380         WARN_ON(BTRFS_I(inode)->block_rsv.reserved);
9381         WARN_ON(BTRFS_I(inode)->block_rsv.size);
9382         WARN_ON(BTRFS_I(inode)->outstanding_extents);
9383         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
9384         WARN_ON(BTRFS_I(inode)->new_delalloc_bytes);
9385         WARN_ON(BTRFS_I(inode)->csum_bytes);
9386         WARN_ON(BTRFS_I(inode)->defrag_bytes);
9387
9388         /*
9389          * This can happen where we create an inode, but somebody else also
9390          * created the same inode and we need to destroy the one we already
9391          * created.
9392          */
9393         if (!root)
9394                 return;
9395
9396         while (1) {
9397                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
9398                 if (!ordered)
9399                         break;
9400                 else {
9401                         btrfs_err(fs_info,
9402                                   "found ordered extent %llu %llu on inode cleanup",
9403                                   ordered->file_offset, ordered->len);
9404                         btrfs_remove_ordered_extent(inode, ordered);
9405                         btrfs_put_ordered_extent(ordered);
9406                         btrfs_put_ordered_extent(ordered);
9407                 }
9408         }
9409         btrfs_qgroup_check_reserved_leak(inode);
9410         inode_tree_del(inode);
9411         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9412 }
9413
9414 int btrfs_drop_inode(struct inode *inode)
9415 {
9416         struct btrfs_root *root = BTRFS_I(inode)->root;
9417
9418         if (root == NULL)
9419                 return 1;
9420
9421         /* the snap/subvol tree is on deleting */
9422         if (btrfs_root_refs(&root->root_item) == 0)
9423                 return 1;
9424         else
9425                 return generic_drop_inode(inode);
9426 }
9427
9428 static void init_once(void *foo)
9429 {
9430         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
9431
9432         inode_init_once(&ei->vfs_inode);
9433 }
9434
9435 void __cold btrfs_destroy_cachep(void)
9436 {
9437         /*
9438          * Make sure all delayed rcu free inodes are flushed before we
9439          * destroy cache.
9440          */
9441         rcu_barrier();
9442         kmem_cache_destroy(btrfs_inode_cachep);
9443         kmem_cache_destroy(btrfs_trans_handle_cachep);
9444         kmem_cache_destroy(btrfs_path_cachep);
9445         kmem_cache_destroy(btrfs_free_space_cachep);
9446         kmem_cache_destroy(btrfs_free_space_bitmap_cachep);
9447 }
9448
9449 int __init btrfs_init_cachep(void)
9450 {
9451         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
9452                         sizeof(struct btrfs_inode), 0,
9453                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD | SLAB_ACCOUNT,
9454                         init_once);
9455         if (!btrfs_inode_cachep)
9456                 goto fail;
9457
9458         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
9459                         sizeof(struct btrfs_trans_handle), 0,
9460                         SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
9461         if (!btrfs_trans_handle_cachep)
9462                 goto fail;
9463
9464         btrfs_path_cachep = kmem_cache_create("btrfs_path",
9465                         sizeof(struct btrfs_path), 0,
9466                         SLAB_MEM_SPREAD, NULL);
9467         if (!btrfs_path_cachep)
9468                 goto fail;
9469
9470         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
9471                         sizeof(struct btrfs_free_space), 0,
9472                         SLAB_MEM_SPREAD, NULL);
9473         if (!btrfs_free_space_cachep)
9474                 goto fail;
9475
9476         btrfs_free_space_bitmap_cachep = kmem_cache_create("btrfs_free_space_bitmap",
9477                                                         PAGE_SIZE, PAGE_SIZE,
9478                                                         SLAB_RED_ZONE, NULL);
9479         if (!btrfs_free_space_bitmap_cachep)
9480                 goto fail;
9481
9482         return 0;
9483 fail:
9484         btrfs_destroy_cachep();
9485         return -ENOMEM;
9486 }
9487
9488 static int btrfs_getattr(const struct path *path, struct kstat *stat,
9489                          u32 request_mask, unsigned int flags)
9490 {
9491         u64 delalloc_bytes;
9492         struct inode *inode = d_inode(path->dentry);
9493         u32 blocksize = inode->i_sb->s_blocksize;
9494         u32 bi_flags = BTRFS_I(inode)->flags;
9495
9496         stat->result_mask |= STATX_BTIME;
9497         stat->btime.tv_sec = BTRFS_I(inode)->i_otime.tv_sec;
9498         stat->btime.tv_nsec = BTRFS_I(inode)->i_otime.tv_nsec;
9499         if (bi_flags & BTRFS_INODE_APPEND)
9500                 stat->attributes |= STATX_ATTR_APPEND;
9501         if (bi_flags & BTRFS_INODE_COMPRESS)
9502                 stat->attributes |= STATX_ATTR_COMPRESSED;
9503         if (bi_flags & BTRFS_INODE_IMMUTABLE)
9504                 stat->attributes |= STATX_ATTR_IMMUTABLE;
9505         if (bi_flags & BTRFS_INODE_NODUMP)
9506                 stat->attributes |= STATX_ATTR_NODUMP;
9507
9508         stat->attributes_mask |= (STATX_ATTR_APPEND |
9509                                   STATX_ATTR_COMPRESSED |
9510                                   STATX_ATTR_IMMUTABLE |
9511                                   STATX_ATTR_NODUMP);
9512
9513         generic_fillattr(inode, stat);
9514         stat->dev = BTRFS_I(inode)->root->anon_dev;
9515
9516         spin_lock(&BTRFS_I(inode)->lock);
9517         delalloc_bytes = BTRFS_I(inode)->new_delalloc_bytes;
9518         spin_unlock(&BTRFS_I(inode)->lock);
9519         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
9520                         ALIGN(delalloc_bytes, blocksize)) >> 9;
9521         return 0;
9522 }
9523
9524 static int btrfs_rename_exchange(struct inode *old_dir,
9525                               struct dentry *old_dentry,
9526                               struct inode *new_dir,
9527                               struct dentry *new_dentry)
9528 {
9529         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9530         struct btrfs_trans_handle *trans;
9531         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9532         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9533         struct inode *new_inode = new_dentry->d_inode;
9534         struct inode *old_inode = old_dentry->d_inode;
9535         struct timespec64 ctime = current_time(old_inode);
9536         struct dentry *parent;
9537         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9538         u64 new_ino = btrfs_ino(BTRFS_I(new_inode));
9539         u64 old_idx = 0;
9540         u64 new_idx = 0;
9541         u64 root_objectid;
9542         int ret;
9543         bool root_log_pinned = false;
9544         bool dest_log_pinned = false;
9545         struct btrfs_log_ctx ctx_root;
9546         struct btrfs_log_ctx ctx_dest;
9547         bool sync_log_root = false;
9548         bool sync_log_dest = false;
9549         bool commit_transaction = false;
9550
9551         /* we only allow rename subvolume link between subvolumes */
9552         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9553                 return -EXDEV;
9554
9555         btrfs_init_log_ctx(&ctx_root, old_inode);
9556         btrfs_init_log_ctx(&ctx_dest, new_inode);
9557
9558         /* close the race window with snapshot create/destroy ioctl */
9559         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9560                 down_read(&fs_info->subvol_sem);
9561         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9562                 down_read(&fs_info->subvol_sem);
9563
9564         /*
9565          * We want to reserve the absolute worst case amount of items.  So if
9566          * both inodes are subvols and we need to unlink them then that would
9567          * require 4 item modifications, but if they are both normal inodes it
9568          * would require 5 item modifications, so we'll assume their normal
9569          * inodes.  So 5 * 2 is 10, plus 2 for the new links, so 12 total items
9570          * should cover the worst case number of items we'll modify.
9571          */
9572         trans = btrfs_start_transaction(root, 12);
9573         if (IS_ERR(trans)) {
9574                 ret = PTR_ERR(trans);
9575                 goto out_notrans;
9576         }
9577
9578         if (dest != root)
9579                 btrfs_record_root_in_trans(trans, dest);
9580
9581         /*
9582          * We need to find a free sequence number both in the source and
9583          * in the destination directory for the exchange.
9584          */
9585         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &old_idx);
9586         if (ret)
9587                 goto out_fail;
9588         ret = btrfs_set_inode_index(BTRFS_I(old_dir), &new_idx);
9589         if (ret)
9590                 goto out_fail;
9591
9592         BTRFS_I(old_inode)->dir_index = 0ULL;
9593         BTRFS_I(new_inode)->dir_index = 0ULL;
9594
9595         /* Reference for the source. */
9596         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9597                 /* force full log commit if subvolume involved. */
9598                 btrfs_set_log_full_commit(trans);
9599         } else {
9600                 btrfs_pin_log_trans(root);
9601                 root_log_pinned = true;
9602                 ret = btrfs_insert_inode_ref(trans, dest,
9603                                              new_dentry->d_name.name,
9604                                              new_dentry->d_name.len,
9605                                              old_ino,
9606                                              btrfs_ino(BTRFS_I(new_dir)),
9607                                              old_idx);
9608                 if (ret)
9609                         goto out_fail;
9610         }
9611
9612         /* And now for the dest. */
9613         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9614                 /* force full log commit if subvolume involved. */
9615                 btrfs_set_log_full_commit(trans);
9616         } else {
9617                 btrfs_pin_log_trans(dest);
9618                 dest_log_pinned = true;
9619                 ret = btrfs_insert_inode_ref(trans, root,
9620                                              old_dentry->d_name.name,
9621                                              old_dentry->d_name.len,
9622                                              new_ino,
9623                                              btrfs_ino(BTRFS_I(old_dir)),
9624                                              new_idx);
9625                 if (ret)
9626                         goto out_fail;
9627         }
9628
9629         /* Update inode version and ctime/mtime. */
9630         inode_inc_iversion(old_dir);
9631         inode_inc_iversion(new_dir);
9632         inode_inc_iversion(old_inode);
9633         inode_inc_iversion(new_inode);
9634         old_dir->i_ctime = old_dir->i_mtime = ctime;
9635         new_dir->i_ctime = new_dir->i_mtime = ctime;
9636         old_inode->i_ctime = ctime;
9637         new_inode->i_ctime = ctime;
9638
9639         if (old_dentry->d_parent != new_dentry->d_parent) {
9640                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9641                                 BTRFS_I(old_inode), 1);
9642                 btrfs_record_unlink_dir(trans, BTRFS_I(new_dir),
9643                                 BTRFS_I(new_inode), 1);
9644         }
9645
9646         /* src is a subvolume */
9647         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9648                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9649                 ret = btrfs_unlink_subvol(trans, old_dir, root_objectid,
9650                                           old_dentry->d_name.name,
9651                                           old_dentry->d_name.len);
9652         } else { /* src is an inode */
9653                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9654                                            BTRFS_I(old_dentry->d_inode),
9655                                            old_dentry->d_name.name,
9656                                            old_dentry->d_name.len);
9657                 if (!ret)
9658                         ret = btrfs_update_inode(trans, root, old_inode);
9659         }
9660         if (ret) {
9661                 btrfs_abort_transaction(trans, ret);
9662                 goto out_fail;
9663         }
9664
9665         /* dest is a subvolume */
9666         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9667                 root_objectid = BTRFS_I(new_inode)->root->root_key.objectid;
9668                 ret = btrfs_unlink_subvol(trans, new_dir, root_objectid,
9669                                           new_dentry->d_name.name,
9670                                           new_dentry->d_name.len);
9671         } else { /* dest is an inode */
9672                 ret = __btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
9673                                            BTRFS_I(new_dentry->d_inode),
9674                                            new_dentry->d_name.name,
9675                                            new_dentry->d_name.len);
9676                 if (!ret)
9677                         ret = btrfs_update_inode(trans, dest, new_inode);
9678         }
9679         if (ret) {
9680                 btrfs_abort_transaction(trans, ret);
9681                 goto out_fail;
9682         }
9683
9684         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
9685                              new_dentry->d_name.name,
9686                              new_dentry->d_name.len, 0, old_idx);
9687         if (ret) {
9688                 btrfs_abort_transaction(trans, ret);
9689                 goto out_fail;
9690         }
9691
9692         ret = btrfs_add_link(trans, BTRFS_I(old_dir), BTRFS_I(new_inode),
9693                              old_dentry->d_name.name,
9694                              old_dentry->d_name.len, 0, new_idx);
9695         if (ret) {
9696                 btrfs_abort_transaction(trans, ret);
9697                 goto out_fail;
9698         }
9699
9700         if (old_inode->i_nlink == 1)
9701                 BTRFS_I(old_inode)->dir_index = old_idx;
9702         if (new_inode->i_nlink == 1)
9703                 BTRFS_I(new_inode)->dir_index = new_idx;
9704
9705         if (root_log_pinned) {
9706                 parent = new_dentry->d_parent;
9707                 ret = btrfs_log_new_name(trans, BTRFS_I(old_inode),
9708                                          BTRFS_I(old_dir), parent,
9709                                          false, &ctx_root);
9710                 if (ret == BTRFS_NEED_LOG_SYNC)
9711                         sync_log_root = true;
9712                 else if (ret == BTRFS_NEED_TRANS_COMMIT)
9713                         commit_transaction = true;
9714                 ret = 0;
9715                 btrfs_end_log_trans(root);
9716                 root_log_pinned = false;
9717         }
9718         if (dest_log_pinned) {
9719                 if (!commit_transaction) {
9720                         parent = old_dentry->d_parent;
9721                         ret = btrfs_log_new_name(trans, BTRFS_I(new_inode),
9722                                                  BTRFS_I(new_dir), parent,
9723                                                  false, &ctx_dest);
9724                         if (ret == BTRFS_NEED_LOG_SYNC)
9725                                 sync_log_dest = true;
9726                         else if (ret == BTRFS_NEED_TRANS_COMMIT)
9727                                 commit_transaction = true;
9728                         ret = 0;
9729                 }
9730                 btrfs_end_log_trans(dest);
9731                 dest_log_pinned = false;
9732         }
9733 out_fail:
9734         /*
9735          * If we have pinned a log and an error happened, we unpin tasks
9736          * trying to sync the log and force them to fallback to a transaction
9737          * commit if the log currently contains any of the inodes involved in
9738          * this rename operation (to ensure we do not persist a log with an
9739          * inconsistent state for any of these inodes or leading to any
9740          * inconsistencies when replayed). If the transaction was aborted, the
9741          * abortion reason is propagated to userspace when attempting to commit
9742          * the transaction. If the log does not contain any of these inodes, we
9743          * allow the tasks to sync it.
9744          */
9745         if (ret && (root_log_pinned || dest_log_pinned)) {
9746                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
9747                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
9748                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
9749                     (new_inode &&
9750                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
9751                         btrfs_set_log_full_commit(trans);
9752
9753                 if (root_log_pinned) {
9754                         btrfs_end_log_trans(root);
9755                         root_log_pinned = false;
9756                 }
9757                 if (dest_log_pinned) {
9758                         btrfs_end_log_trans(dest);
9759                         dest_log_pinned = false;
9760                 }
9761         }
9762         if (!ret && sync_log_root && !commit_transaction) {
9763                 ret = btrfs_sync_log(trans, BTRFS_I(old_inode)->root,
9764                                      &ctx_root);
9765                 if (ret)
9766                         commit_transaction = true;
9767         }
9768         if (!ret && sync_log_dest && !commit_transaction) {
9769                 ret = btrfs_sync_log(trans, BTRFS_I(new_inode)->root,
9770                                      &ctx_dest);
9771                 if (ret)
9772                         commit_transaction = true;
9773         }
9774         if (commit_transaction) {
9775                 /*
9776                  * We may have set commit_transaction when logging the new name
9777                  * in the destination root, in which case we left the source
9778                  * root context in the list of log contextes. So make sure we
9779                  * remove it to avoid invalid memory accesses, since the context
9780                  * was allocated in our stack frame.
9781                  */
9782                 if (sync_log_root) {
9783                         mutex_lock(&root->log_mutex);
9784                         list_del_init(&ctx_root.list);
9785                         mutex_unlock(&root->log_mutex);
9786                 }
9787                 ret = btrfs_commit_transaction(trans);
9788         } else {
9789                 int ret2;
9790
9791                 ret2 = btrfs_end_transaction(trans);
9792                 ret = ret ? ret : ret2;
9793         }
9794 out_notrans:
9795         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9796                 up_read(&fs_info->subvol_sem);
9797         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9798                 up_read(&fs_info->subvol_sem);
9799
9800         ASSERT(list_empty(&ctx_root.list));
9801         ASSERT(list_empty(&ctx_dest.list));
9802
9803         return ret;
9804 }
9805
9806 static int btrfs_whiteout_for_rename(struct btrfs_trans_handle *trans,
9807                                      struct btrfs_root *root,
9808                                      struct inode *dir,
9809                                      struct dentry *dentry)
9810 {
9811         int ret;
9812         struct inode *inode;
9813         u64 objectid;
9814         u64 index;
9815
9816         ret = btrfs_find_free_ino(root, &objectid);
9817         if (ret)
9818                 return ret;
9819
9820         inode = btrfs_new_inode(trans, root, dir,
9821                                 dentry->d_name.name,
9822                                 dentry->d_name.len,
9823                                 btrfs_ino(BTRFS_I(dir)),
9824                                 objectid,
9825                                 S_IFCHR | WHITEOUT_MODE,
9826                                 &index);
9827
9828         if (IS_ERR(inode)) {
9829                 ret = PTR_ERR(inode);
9830                 return ret;
9831         }
9832
9833         inode->i_op = &btrfs_special_inode_operations;
9834         init_special_inode(inode, inode->i_mode,
9835                 WHITEOUT_DEV);
9836
9837         ret = btrfs_init_inode_security(trans, inode, dir,
9838                                 &dentry->d_name);
9839         if (ret)
9840                 goto out;
9841
9842         ret = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
9843                                 BTRFS_I(inode), 0, index);
9844         if (ret)
9845                 goto out;
9846
9847         ret = btrfs_update_inode(trans, root, inode);
9848 out:
9849         unlock_new_inode(inode);
9850         if (ret)
9851                 inode_dec_link_count(inode);
9852         iput(inode);
9853
9854         return ret;
9855 }
9856
9857 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
9858                            struct inode *new_dir, struct dentry *new_dentry,
9859                            unsigned int flags)
9860 {
9861         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9862         struct btrfs_trans_handle *trans;
9863         unsigned int trans_num_items;
9864         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9865         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9866         struct inode *new_inode = d_inode(new_dentry);
9867         struct inode *old_inode = d_inode(old_dentry);
9868         u64 index = 0;
9869         u64 root_objectid;
9870         int ret;
9871         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9872         bool log_pinned = false;
9873         struct btrfs_log_ctx ctx;
9874         bool sync_log = false;
9875         bool commit_transaction = false;
9876
9877         if (btrfs_ino(BTRFS_I(new_dir)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
9878                 return -EPERM;
9879
9880         /* we only allow rename subvolume link between subvolumes */
9881         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9882                 return -EXDEV;
9883
9884         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
9885             (new_inode && btrfs_ino(BTRFS_I(new_inode)) == BTRFS_FIRST_FREE_OBJECTID))
9886                 return -ENOTEMPTY;
9887
9888         if (S_ISDIR(old_inode->i_mode) && new_inode &&
9889             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
9890                 return -ENOTEMPTY;
9891
9892
9893         /* check for collisions, even if the  name isn't there */
9894         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
9895                              new_dentry->d_name.name,
9896                              new_dentry->d_name.len);
9897
9898         if (ret) {
9899                 if (ret == -EEXIST) {
9900                         /* we shouldn't get
9901                          * eexist without a new_inode */
9902                         if (WARN_ON(!new_inode)) {
9903                                 return ret;
9904                         }
9905                 } else {
9906                         /* maybe -EOVERFLOW */
9907                         return ret;
9908                 }
9909         }
9910         ret = 0;
9911
9912         /*
9913          * we're using rename to replace one file with another.  Start IO on it
9914          * now so  we don't add too much work to the end of the transaction
9915          */
9916         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
9917                 filemap_flush(old_inode->i_mapping);
9918
9919         /* close the racy window with snapshot create/destroy ioctl */
9920         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9921                 down_read(&fs_info->subvol_sem);
9922         /*
9923          * We want to reserve the absolute worst case amount of items.  So if
9924          * both inodes are subvols and we need to unlink them then that would
9925          * require 4 item modifications, but if they are both normal inodes it
9926          * would require 5 item modifications, so we'll assume they are normal
9927          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
9928          * should cover the worst case number of items we'll modify.
9929          * If our rename has the whiteout flag, we need more 5 units for the
9930          * new inode (1 inode item, 1 inode ref, 2 dir items and 1 xattr item
9931          * when selinux is enabled).
9932          */
9933         trans_num_items = 11;
9934         if (flags & RENAME_WHITEOUT)
9935                 trans_num_items += 5;
9936         trans = btrfs_start_transaction(root, trans_num_items);
9937         if (IS_ERR(trans)) {
9938                 ret = PTR_ERR(trans);
9939                 goto out_notrans;
9940         }
9941
9942         if (dest != root)
9943                 btrfs_record_root_in_trans(trans, dest);
9944
9945         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &index);
9946         if (ret)
9947                 goto out_fail;
9948
9949         BTRFS_I(old_inode)->dir_index = 0ULL;
9950         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9951                 /* force full log commit if subvolume involved. */
9952                 btrfs_set_log_full_commit(trans);
9953         } else {
9954                 btrfs_pin_log_trans(root);
9955                 log_pinned = true;
9956                 ret = btrfs_insert_inode_ref(trans, dest,
9957                                              new_dentry->d_name.name,
9958                                              new_dentry->d_name.len,
9959                                              old_ino,
9960                                              btrfs_ino(BTRFS_I(new_dir)), index);
9961                 if (ret)
9962                         goto out_fail;
9963         }
9964
9965         inode_inc_iversion(old_dir);
9966         inode_inc_iversion(new_dir);
9967         inode_inc_iversion(old_inode);
9968         old_dir->i_ctime = old_dir->i_mtime =
9969         new_dir->i_ctime = new_dir->i_mtime =
9970         old_inode->i_ctime = current_time(old_dir);
9971
9972         if (old_dentry->d_parent != new_dentry->d_parent)
9973                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9974                                 BTRFS_I(old_inode), 1);
9975
9976         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9977                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9978                 ret = btrfs_unlink_subvol(trans, old_dir, root_objectid,
9979                                         old_dentry->d_name.name,
9980                                         old_dentry->d_name.len);
9981         } else {
9982                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9983                                         BTRFS_I(d_inode(old_dentry)),
9984                                         old_dentry->d_name.name,
9985                                         old_dentry->d_name.len);
9986                 if (!ret)
9987                         ret = btrfs_update_inode(trans, root, old_inode);
9988         }
9989         if (ret) {
9990                 btrfs_abort_transaction(trans, ret);
9991                 goto out_fail;
9992         }
9993
9994         if (new_inode) {
9995                 inode_inc_iversion(new_inode);
9996                 new_inode->i_ctime = current_time(new_inode);
9997                 if (unlikely(btrfs_ino(BTRFS_I(new_inode)) ==
9998                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
9999                         root_objectid = BTRFS_I(new_inode)->location.objectid;
10000                         ret = btrfs_unlink_subvol(trans, new_dir, root_objectid,
10001                                                 new_dentry->d_name.name,
10002                                                 new_dentry->d_name.len);
10003                         BUG_ON(new_inode->i_nlink == 0);
10004                 } else {
10005                         ret = btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
10006                                                  BTRFS_I(d_inode(new_dentry)),
10007                                                  new_dentry->d_name.name,
10008                                                  new_dentry->d_name.len);
10009                 }
10010                 if (!ret && new_inode->i_nlink == 0)
10011                         ret = btrfs_orphan_add(trans,
10012                                         BTRFS_I(d_inode(new_dentry)));
10013                 if (ret) {
10014                         btrfs_abort_transaction(trans, ret);
10015                         goto out_fail;
10016                 }
10017         }
10018
10019         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
10020                              new_dentry->d_name.name,
10021                              new_dentry->d_name.len, 0, index);
10022         if (ret) {
10023                 btrfs_abort_transaction(trans, ret);
10024                 goto out_fail;
10025         }
10026
10027         if (old_inode->i_nlink == 1)
10028                 BTRFS_I(old_inode)->dir_index = index;
10029
10030         if (log_pinned) {
10031                 struct dentry *parent = new_dentry->d_parent;
10032
10033                 btrfs_init_log_ctx(&ctx, old_inode);
10034                 ret = btrfs_log_new_name(trans, BTRFS_I(old_inode),
10035                                          BTRFS_I(old_dir), parent,
10036                                          false, &ctx);
10037                 if (ret == BTRFS_NEED_LOG_SYNC)
10038                         sync_log = true;
10039                 else if (ret == BTRFS_NEED_TRANS_COMMIT)
10040                         commit_transaction = true;
10041                 ret = 0;
10042                 btrfs_end_log_trans(root);
10043                 log_pinned = false;
10044         }
10045
10046         if (flags & RENAME_WHITEOUT) {
10047                 ret = btrfs_whiteout_for_rename(trans, root, old_dir,
10048                                                 old_dentry);
10049
10050                 if (ret) {
10051                         btrfs_abort_transaction(trans, ret);
10052                         goto out_fail;
10053                 }
10054         }
10055 out_fail:
10056         /*
10057          * If we have pinned the log and an error happened, we unpin tasks
10058          * trying to sync the log and force them to fallback to a transaction
10059          * commit if the log currently contains any of the inodes involved in
10060          * this rename operation (to ensure we do not persist a log with an
10061          * inconsistent state for any of these inodes or leading to any
10062          * inconsistencies when replayed). If the transaction was aborted, the
10063          * abortion reason is propagated to userspace when attempting to commit
10064          * the transaction. If the log does not contain any of these inodes, we
10065          * allow the tasks to sync it.
10066          */
10067         if (ret && log_pinned) {
10068                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
10069                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
10070                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
10071                     (new_inode &&
10072                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
10073                         btrfs_set_log_full_commit(trans);
10074
10075                 btrfs_end_log_trans(root);
10076                 log_pinned = false;
10077         }
10078         if (!ret && sync_log) {
10079                 ret = btrfs_sync_log(trans, BTRFS_I(old_inode)->root, &ctx);
10080                 if (ret)
10081                         commit_transaction = true;
10082         }
10083         if (commit_transaction) {
10084                 ret = btrfs_commit_transaction(trans);
10085         } else {
10086                 int ret2;
10087
10088                 ret2 = btrfs_end_transaction(trans);
10089                 ret = ret ? ret : ret2;
10090         }
10091 out_notrans:
10092         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
10093                 up_read(&fs_info->subvol_sem);
10094
10095         return ret;
10096 }
10097
10098 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
10099                          struct inode *new_dir, struct dentry *new_dentry,
10100                          unsigned int flags)
10101 {
10102         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
10103                 return -EINVAL;
10104
10105         if (flags & RENAME_EXCHANGE)
10106                 return btrfs_rename_exchange(old_dir, old_dentry, new_dir,
10107                                           new_dentry);
10108
10109         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry, flags);
10110 }
10111
10112 struct btrfs_delalloc_work {
10113         struct inode *inode;
10114         struct completion completion;
10115         struct list_head list;
10116         struct btrfs_work work;
10117 };
10118
10119 static void btrfs_run_delalloc_work(struct btrfs_work *work)
10120 {
10121         struct btrfs_delalloc_work *delalloc_work;
10122         struct inode *inode;
10123
10124         delalloc_work = container_of(work, struct btrfs_delalloc_work,
10125                                      work);
10126         inode = delalloc_work->inode;
10127         filemap_flush(inode->i_mapping);
10128         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
10129                                 &BTRFS_I(inode)->runtime_flags))
10130                 filemap_flush(inode->i_mapping);
10131
10132         iput(inode);
10133         complete(&delalloc_work->completion);
10134 }
10135
10136 static struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode)
10137 {
10138         struct btrfs_delalloc_work *work;
10139
10140         work = kmalloc(sizeof(*work), GFP_NOFS);
10141         if (!work)
10142                 return NULL;
10143
10144         init_completion(&work->completion);
10145         INIT_LIST_HEAD(&work->list);
10146         work->inode = inode;
10147         btrfs_init_work(&work->work, btrfs_run_delalloc_work, NULL, NULL);
10148
10149         return work;
10150 }
10151
10152 /*
10153  * some fairly slow code that needs optimization. This walks the list
10154  * of all the inodes with pending delalloc and forces them to disk.
10155  */
10156 static int start_delalloc_inodes(struct btrfs_root *root, int nr, bool snapshot)
10157 {
10158         struct btrfs_inode *binode;
10159         struct inode *inode;
10160         struct btrfs_delalloc_work *work, *next;
10161         struct list_head works;
10162         struct list_head splice;
10163         int ret = 0;
10164
10165         INIT_LIST_HEAD(&works);
10166         INIT_LIST_HEAD(&splice);
10167
10168         mutex_lock(&root->delalloc_mutex);
10169         spin_lock(&root->delalloc_lock);
10170         list_splice_init(&root->delalloc_inodes, &splice);
10171         while (!list_empty(&splice)) {
10172                 binode = list_entry(splice.next, struct btrfs_inode,
10173                                     delalloc_inodes);
10174
10175                 list_move_tail(&binode->delalloc_inodes,
10176                                &root->delalloc_inodes);
10177                 inode = igrab(&binode->vfs_inode);
10178                 if (!inode) {
10179                         cond_resched_lock(&root->delalloc_lock);
10180                         continue;
10181                 }
10182                 spin_unlock(&root->delalloc_lock);
10183
10184                 if (snapshot)
10185                         set_bit(BTRFS_INODE_SNAPSHOT_FLUSH,
10186                                 &binode->runtime_flags);
10187                 work = btrfs_alloc_delalloc_work(inode);
10188                 if (!work) {
10189                         iput(inode);
10190                         ret = -ENOMEM;
10191                         goto out;
10192                 }
10193                 list_add_tail(&work->list, &works);
10194                 btrfs_queue_work(root->fs_info->flush_workers,
10195                                  &work->work);
10196                 ret++;
10197                 if (nr != -1 && ret >= nr)
10198                         goto out;
10199                 cond_resched();
10200                 spin_lock(&root->delalloc_lock);
10201         }
10202         spin_unlock(&root->delalloc_lock);
10203
10204 out:
10205         list_for_each_entry_safe(work, next, &works, list) {
10206                 list_del_init(&work->list);
10207                 wait_for_completion(&work->completion);
10208                 kfree(work);
10209         }
10210
10211         if (!list_empty(&splice)) {
10212                 spin_lock(&root->delalloc_lock);
10213                 list_splice_tail(&splice, &root->delalloc_inodes);
10214                 spin_unlock(&root->delalloc_lock);
10215         }
10216         mutex_unlock(&root->delalloc_mutex);
10217         return ret;
10218 }
10219
10220 int btrfs_start_delalloc_snapshot(struct btrfs_root *root)
10221 {
10222         struct btrfs_fs_info *fs_info = root->fs_info;
10223         int ret;
10224
10225         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10226                 return -EROFS;
10227
10228         ret = start_delalloc_inodes(root, -1, true);
10229         if (ret > 0)
10230                 ret = 0;
10231         return ret;
10232 }
10233
10234 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int nr)
10235 {
10236         struct btrfs_root *root;
10237         struct list_head splice;
10238         int ret;
10239
10240         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10241                 return -EROFS;
10242
10243         INIT_LIST_HEAD(&splice);
10244
10245         mutex_lock(&fs_info->delalloc_root_mutex);
10246         spin_lock(&fs_info->delalloc_root_lock);
10247         list_splice_init(&fs_info->delalloc_roots, &splice);
10248         while (!list_empty(&splice) && nr) {
10249                 root = list_first_entry(&splice, struct btrfs_root,
10250                                         delalloc_root);
10251                 root = btrfs_grab_fs_root(root);
10252                 BUG_ON(!root);
10253                 list_move_tail(&root->delalloc_root,
10254                                &fs_info->delalloc_roots);
10255                 spin_unlock(&fs_info->delalloc_root_lock);
10256
10257                 ret = start_delalloc_inodes(root, nr, false);
10258                 btrfs_put_fs_root(root);
10259                 if (ret < 0)
10260                         goto out;
10261
10262                 if (nr != -1) {
10263                         nr -= ret;
10264                         WARN_ON(nr < 0);
10265                 }
10266                 spin_lock(&fs_info->delalloc_root_lock);
10267         }
10268         spin_unlock(&fs_info->delalloc_root_lock);
10269
10270         ret = 0;
10271 out:
10272         if (!list_empty(&splice)) {
10273                 spin_lock(&fs_info->delalloc_root_lock);
10274                 list_splice_tail(&splice, &fs_info->delalloc_roots);
10275                 spin_unlock(&fs_info->delalloc_root_lock);
10276         }
10277         mutex_unlock(&fs_info->delalloc_root_mutex);
10278         return ret;
10279 }
10280
10281 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
10282                          const char *symname)
10283 {
10284         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10285         struct btrfs_trans_handle *trans;
10286         struct btrfs_root *root = BTRFS_I(dir)->root;
10287         struct btrfs_path *path;
10288         struct btrfs_key key;
10289         struct inode *inode = NULL;
10290         int err;
10291         u64 objectid;
10292         u64 index = 0;
10293         int name_len;
10294         int datasize;
10295         unsigned long ptr;
10296         struct btrfs_file_extent_item *ei;
10297         struct extent_buffer *leaf;
10298
10299         name_len = strlen(symname);
10300         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info))
10301                 return -ENAMETOOLONG;
10302
10303         /*
10304          * 2 items for inode item and ref
10305          * 2 items for dir items
10306          * 1 item for updating parent inode item
10307          * 1 item for the inline extent item
10308          * 1 item for xattr if selinux is on
10309          */
10310         trans = btrfs_start_transaction(root, 7);
10311         if (IS_ERR(trans))
10312                 return PTR_ERR(trans);
10313
10314         err = btrfs_find_free_ino(root, &objectid);
10315         if (err)
10316                 goto out_unlock;
10317
10318         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
10319                                 dentry->d_name.len, btrfs_ino(BTRFS_I(dir)),
10320                                 objectid, S_IFLNK|S_IRWXUGO, &index);
10321         if (IS_ERR(inode)) {
10322                 err = PTR_ERR(inode);
10323                 inode = NULL;
10324                 goto out_unlock;
10325         }
10326
10327         /*
10328         * If the active LSM wants to access the inode during
10329         * d_instantiate it needs these. Smack checks to see
10330         * if the filesystem supports xattrs by looking at the
10331         * ops vector.
10332         */
10333         inode->i_fop = &btrfs_file_operations;
10334         inode->i_op = &btrfs_file_inode_operations;
10335         inode->i_mapping->a_ops = &btrfs_aops;
10336         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10337
10338         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
10339         if (err)
10340                 goto out_unlock;
10341
10342         path = btrfs_alloc_path();
10343         if (!path) {
10344                 err = -ENOMEM;
10345                 goto out_unlock;
10346         }
10347         key.objectid = btrfs_ino(BTRFS_I(inode));
10348         key.offset = 0;
10349         key.type = BTRFS_EXTENT_DATA_KEY;
10350         datasize = btrfs_file_extent_calc_inline_size(name_len);
10351         err = btrfs_insert_empty_item(trans, root, path, &key,
10352                                       datasize);
10353         if (err) {
10354                 btrfs_free_path(path);
10355                 goto out_unlock;
10356         }
10357         leaf = path->nodes[0];
10358         ei = btrfs_item_ptr(leaf, path->slots[0],
10359                             struct btrfs_file_extent_item);
10360         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
10361         btrfs_set_file_extent_type(leaf, ei,
10362                                    BTRFS_FILE_EXTENT_INLINE);
10363         btrfs_set_file_extent_encryption(leaf, ei, 0);
10364         btrfs_set_file_extent_compression(leaf, ei, 0);
10365         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
10366         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
10367
10368         ptr = btrfs_file_extent_inline_start(ei);
10369         write_extent_buffer(leaf, symname, ptr, name_len);
10370         btrfs_mark_buffer_dirty(leaf);
10371         btrfs_free_path(path);
10372
10373         inode->i_op = &btrfs_symlink_inode_operations;
10374         inode_nohighmem(inode);
10375         inode_set_bytes(inode, name_len);
10376         btrfs_i_size_write(BTRFS_I(inode), name_len);
10377         err = btrfs_update_inode(trans, root, inode);
10378         /*
10379          * Last step, add directory indexes for our symlink inode. This is the
10380          * last step to avoid extra cleanup of these indexes if an error happens
10381          * elsewhere above.
10382          */
10383         if (!err)
10384                 err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
10385                                 BTRFS_I(inode), 0, index);
10386         if (err)
10387                 goto out_unlock;
10388
10389         d_instantiate_new(dentry, inode);
10390
10391 out_unlock:
10392         btrfs_end_transaction(trans);
10393         if (err && inode) {
10394                 inode_dec_link_count(inode);
10395                 discard_new_inode(inode);
10396         }
10397         btrfs_btree_balance_dirty(fs_info);
10398         return err;
10399 }
10400
10401 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
10402                                        u64 start, u64 num_bytes, u64 min_size,
10403                                        loff_t actual_len, u64 *alloc_hint,
10404                                        struct btrfs_trans_handle *trans)
10405 {
10406         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
10407         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
10408         struct extent_map *em;
10409         struct btrfs_root *root = BTRFS_I(inode)->root;
10410         struct btrfs_key ins;
10411         u64 cur_offset = start;
10412         u64 i_size;
10413         u64 cur_bytes;
10414         u64 last_alloc = (u64)-1;
10415         int ret = 0;
10416         bool own_trans = true;
10417         u64 end = start + num_bytes - 1;
10418
10419         if (trans)
10420                 own_trans = false;
10421         while (num_bytes > 0) {
10422                 if (own_trans) {
10423                         trans = btrfs_start_transaction(root, 3);
10424                         if (IS_ERR(trans)) {
10425                                 ret = PTR_ERR(trans);
10426                                 break;
10427                         }
10428                 }
10429
10430                 cur_bytes = min_t(u64, num_bytes, SZ_256M);
10431                 cur_bytes = max(cur_bytes, min_size);
10432                 /*
10433                  * If we are severely fragmented we could end up with really
10434                  * small allocations, so if the allocator is returning small
10435                  * chunks lets make its job easier by only searching for those
10436                  * sized chunks.
10437                  */
10438                 cur_bytes = min(cur_bytes, last_alloc);
10439                 ret = btrfs_reserve_extent(root, cur_bytes, cur_bytes,
10440                                 min_size, 0, *alloc_hint, &ins, 1, 0);
10441                 if (ret) {
10442                         if (own_trans)
10443                                 btrfs_end_transaction(trans);
10444                         break;
10445                 }
10446                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
10447
10448                 last_alloc = ins.offset;
10449                 ret = insert_reserved_file_extent(trans, inode,
10450                                                   cur_offset, ins.objectid,
10451                                                   ins.offset, ins.offset,
10452                                                   ins.offset, 0, 0, 0,
10453                                                   BTRFS_FILE_EXTENT_PREALLOC);
10454                 if (ret) {
10455                         btrfs_free_reserved_extent(fs_info, ins.objectid,
10456                                                    ins.offset, 0);
10457                         btrfs_abort_transaction(trans, ret);
10458                         if (own_trans)
10459                                 btrfs_end_transaction(trans);
10460                         break;
10461                 }
10462
10463                 btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10464                                         cur_offset + ins.offset -1, 0);
10465
10466                 em = alloc_extent_map();
10467                 if (!em) {
10468                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
10469                                 &BTRFS_I(inode)->runtime_flags);
10470                         goto next;
10471                 }
10472
10473                 em->start = cur_offset;
10474                 em->orig_start = cur_offset;
10475                 em->len = ins.offset;
10476                 em->block_start = ins.objectid;
10477                 em->block_len = ins.offset;
10478                 em->orig_block_len = ins.offset;
10479                 em->ram_bytes = ins.offset;
10480                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
10481                 em->generation = trans->transid;
10482
10483                 while (1) {
10484                         write_lock(&em_tree->lock);
10485                         ret = add_extent_mapping(em_tree, em, 1);
10486                         write_unlock(&em_tree->lock);
10487                         if (ret != -EEXIST)
10488                                 break;
10489                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10490                                                 cur_offset + ins.offset - 1,
10491                                                 0);
10492                 }
10493                 free_extent_map(em);
10494 next:
10495                 num_bytes -= ins.offset;
10496                 cur_offset += ins.offset;
10497                 *alloc_hint = ins.objectid + ins.offset;
10498
10499                 inode_inc_iversion(inode);
10500                 inode->i_ctime = current_time(inode);
10501                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
10502                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
10503                     (actual_len > inode->i_size) &&
10504                     (cur_offset > inode->i_size)) {
10505                         if (cur_offset > actual_len)
10506                                 i_size = actual_len;
10507                         else
10508                                 i_size = cur_offset;
10509                         i_size_write(inode, i_size);
10510                         btrfs_ordered_update_i_size(inode, i_size, NULL);
10511                 }
10512
10513                 ret = btrfs_update_inode(trans, root, inode);
10514
10515                 if (ret) {
10516                         btrfs_abort_transaction(trans, ret);
10517                         if (own_trans)
10518                                 btrfs_end_transaction(trans);
10519                         break;
10520                 }
10521
10522                 if (own_trans)
10523                         btrfs_end_transaction(trans);
10524         }
10525         if (cur_offset < end)
10526                 btrfs_free_reserved_data_space(inode, NULL, cur_offset,
10527                         end - cur_offset + 1);
10528         return ret;
10529 }
10530
10531 int btrfs_prealloc_file_range(struct inode *inode, int mode,
10532                               u64 start, u64 num_bytes, u64 min_size,
10533                               loff_t actual_len, u64 *alloc_hint)
10534 {
10535         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10536                                            min_size, actual_len, alloc_hint,
10537                                            NULL);
10538 }
10539
10540 int btrfs_prealloc_file_range_trans(struct inode *inode,
10541                                     struct btrfs_trans_handle *trans, int mode,
10542                                     u64 start, u64 num_bytes, u64 min_size,
10543                                     loff_t actual_len, u64 *alloc_hint)
10544 {
10545         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10546                                            min_size, actual_len, alloc_hint, trans);
10547 }
10548
10549 static int btrfs_set_page_dirty(struct page *page)
10550 {
10551         return __set_page_dirty_nobuffers(page);
10552 }
10553
10554 static int btrfs_permission(struct inode *inode, int mask)
10555 {
10556         struct btrfs_root *root = BTRFS_I(inode)->root;
10557         umode_t mode = inode->i_mode;
10558
10559         if (mask & MAY_WRITE &&
10560             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
10561                 if (btrfs_root_readonly(root))
10562                         return -EROFS;
10563                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
10564                         return -EACCES;
10565         }
10566         return generic_permission(inode, mask);
10567 }
10568
10569 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
10570 {
10571         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10572         struct btrfs_trans_handle *trans;
10573         struct btrfs_root *root = BTRFS_I(dir)->root;
10574         struct inode *inode = NULL;
10575         u64 objectid;
10576         u64 index;
10577         int ret = 0;
10578
10579         /*
10580          * 5 units required for adding orphan entry
10581          */
10582         trans = btrfs_start_transaction(root, 5);
10583         if (IS_ERR(trans))
10584                 return PTR_ERR(trans);
10585
10586         ret = btrfs_find_free_ino(root, &objectid);
10587         if (ret)
10588                 goto out;
10589
10590         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
10591                         btrfs_ino(BTRFS_I(dir)), objectid, mode, &index);
10592         if (IS_ERR(inode)) {
10593                 ret = PTR_ERR(inode);
10594                 inode = NULL;
10595                 goto out;
10596         }
10597
10598         inode->i_fop = &btrfs_file_operations;
10599         inode->i_op = &btrfs_file_inode_operations;
10600
10601         inode->i_mapping->a_ops = &btrfs_aops;
10602         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10603
10604         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
10605         if (ret)
10606                 goto out;
10607
10608         ret = btrfs_update_inode(trans, root, inode);
10609         if (ret)
10610                 goto out;
10611         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10612         if (ret)
10613                 goto out;
10614
10615         /*
10616          * We set number of links to 0 in btrfs_new_inode(), and here we set
10617          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
10618          * through:
10619          *
10620          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
10621          */
10622         set_nlink(inode, 1);
10623         d_tmpfile(dentry, inode);
10624         unlock_new_inode(inode);
10625         mark_inode_dirty(inode);
10626 out:
10627         btrfs_end_transaction(trans);
10628         if (ret && inode)
10629                 discard_new_inode(inode);
10630         btrfs_btree_balance_dirty(fs_info);
10631         return ret;
10632 }
10633
10634 void btrfs_set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
10635 {
10636         struct inode *inode = tree->private_data;
10637         unsigned long index = start >> PAGE_SHIFT;
10638         unsigned long end_index = end >> PAGE_SHIFT;
10639         struct page *page;
10640
10641         while (index <= end_index) {
10642                 page = find_get_page(inode->i_mapping, index);
10643                 ASSERT(page); /* Pages should be in the extent_io_tree */
10644                 set_page_writeback(page);
10645                 put_page(page);
10646                 index++;
10647         }
10648 }
10649
10650 #ifdef CONFIG_SWAP
10651 /*
10652  * Add an entry indicating a block group or device which is pinned by a
10653  * swapfile. Returns 0 on success, 1 if there is already an entry for it, or a
10654  * negative errno on failure.
10655  */
10656 static int btrfs_add_swapfile_pin(struct inode *inode, void *ptr,
10657                                   bool is_block_group)
10658 {
10659         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
10660         struct btrfs_swapfile_pin *sp, *entry;
10661         struct rb_node **p;
10662         struct rb_node *parent = NULL;
10663
10664         sp = kmalloc(sizeof(*sp), GFP_NOFS);
10665         if (!sp)
10666                 return -ENOMEM;
10667         sp->ptr = ptr;
10668         sp->inode = inode;
10669         sp->is_block_group = is_block_group;
10670
10671         spin_lock(&fs_info->swapfile_pins_lock);
10672         p = &fs_info->swapfile_pins.rb_node;
10673         while (*p) {
10674                 parent = *p;
10675                 entry = rb_entry(parent, struct btrfs_swapfile_pin, node);
10676                 if (sp->ptr < entry->ptr ||
10677                     (sp->ptr == entry->ptr && sp->inode < entry->inode)) {
10678                         p = &(*p)->rb_left;
10679                 } else if (sp->ptr > entry->ptr ||
10680                            (sp->ptr == entry->ptr && sp->inode > entry->inode)) {
10681                         p = &(*p)->rb_right;
10682                 } else {
10683                         spin_unlock(&fs_info->swapfile_pins_lock);
10684                         kfree(sp);
10685                         return 1;
10686                 }
10687         }
10688         rb_link_node(&sp->node, parent, p);
10689         rb_insert_color(&sp->node, &fs_info->swapfile_pins);
10690         spin_unlock(&fs_info->swapfile_pins_lock);
10691         return 0;
10692 }
10693
10694 /* Free all of the entries pinned by this swapfile. */
10695 static void btrfs_free_swapfile_pins(struct inode *inode)
10696 {
10697         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
10698         struct btrfs_swapfile_pin *sp;
10699         struct rb_node *node, *next;
10700
10701         spin_lock(&fs_info->swapfile_pins_lock);
10702         node = rb_first(&fs_info->swapfile_pins);
10703         while (node) {
10704                 next = rb_next(node);
10705                 sp = rb_entry(node, struct btrfs_swapfile_pin, node);
10706                 if (sp->inode == inode) {
10707                         rb_erase(&sp->node, &fs_info->swapfile_pins);
10708                         if (sp->is_block_group)
10709                                 btrfs_put_block_group(sp->ptr);
10710                         kfree(sp);
10711                 }
10712                 node = next;
10713         }
10714         spin_unlock(&fs_info->swapfile_pins_lock);
10715 }
10716
10717 struct btrfs_swap_info {
10718         u64 start;
10719         u64 block_start;
10720         u64 block_len;
10721         u64 lowest_ppage;
10722         u64 highest_ppage;
10723         unsigned long nr_pages;
10724         int nr_extents;
10725 };
10726
10727 static int btrfs_add_swap_extent(struct swap_info_struct *sis,
10728                                  struct btrfs_swap_info *bsi)
10729 {
10730         unsigned long nr_pages;
10731         u64 first_ppage, first_ppage_reported, next_ppage;
10732         int ret;
10733
10734         first_ppage = ALIGN(bsi->block_start, PAGE_SIZE) >> PAGE_SHIFT;
10735         next_ppage = ALIGN_DOWN(bsi->block_start + bsi->block_len,
10736                                 PAGE_SIZE) >> PAGE_SHIFT;
10737
10738         if (first_ppage >= next_ppage)
10739                 return 0;
10740         nr_pages = next_ppage - first_ppage;
10741
10742         first_ppage_reported = first_ppage;
10743         if (bsi->start == 0)
10744                 first_ppage_reported++;
10745         if (bsi->lowest_ppage > first_ppage_reported)
10746                 bsi->lowest_ppage = first_ppage_reported;
10747         if (bsi->highest_ppage < (next_ppage - 1))
10748                 bsi->highest_ppage = next_ppage - 1;
10749
10750         ret = add_swap_extent(sis, bsi->nr_pages, nr_pages, first_ppage);
10751         if (ret < 0)
10752                 return ret;
10753         bsi->nr_extents += ret;
10754         bsi->nr_pages += nr_pages;
10755         return 0;
10756 }
10757
10758 static void btrfs_swap_deactivate(struct file *file)
10759 {
10760         struct inode *inode = file_inode(file);
10761
10762         btrfs_free_swapfile_pins(inode);
10763         atomic_dec(&BTRFS_I(inode)->root->nr_swapfiles);
10764 }
10765
10766 static int btrfs_swap_activate(struct swap_info_struct *sis, struct file *file,
10767                                sector_t *span)
10768 {
10769         struct inode *inode = file_inode(file);
10770         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
10771         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
10772         struct extent_state *cached_state = NULL;
10773         struct extent_map *em = NULL;
10774         struct btrfs_device *device = NULL;
10775         struct btrfs_swap_info bsi = {
10776                 .lowest_ppage = (sector_t)-1ULL,
10777         };
10778         int ret = 0;
10779         u64 isize;
10780         u64 start;
10781
10782         /*
10783          * If the swap file was just created, make sure delalloc is done. If the
10784          * file changes again after this, the user is doing something stupid and
10785          * we don't really care.
10786          */
10787         ret = btrfs_wait_ordered_range(inode, 0, (u64)-1);
10788         if (ret)
10789                 return ret;
10790
10791         /*
10792          * The inode is locked, so these flags won't change after we check them.
10793          */
10794         if (BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS) {
10795                 btrfs_warn(fs_info, "swapfile must not be compressed");
10796                 return -EINVAL;
10797         }
10798         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW)) {
10799                 btrfs_warn(fs_info, "swapfile must not be copy-on-write");
10800                 return -EINVAL;
10801         }
10802         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
10803                 btrfs_warn(fs_info, "swapfile must not be checksummed");
10804                 return -EINVAL;
10805         }
10806
10807         /*
10808          * Balance or device remove/replace/resize can move stuff around from
10809          * under us. The EXCL_OP flag makes sure they aren't running/won't run
10810          * concurrently while we are mapping the swap extents, and
10811          * fs_info->swapfile_pins prevents them from running while the swap file
10812          * is active and moving the extents. Note that this also prevents a
10813          * concurrent device add which isn't actually necessary, but it's not
10814          * really worth the trouble to allow it.
10815          */
10816         if (test_and_set_bit(BTRFS_FS_EXCL_OP, &fs_info->flags)) {
10817                 btrfs_warn(fs_info,
10818            "cannot activate swapfile while exclusive operation is running");
10819                 return -EBUSY;
10820         }
10821         /*
10822          * Snapshots can create extents which require COW even if NODATACOW is
10823          * set. We use this counter to prevent snapshots. We must increment it
10824          * before walking the extents because we don't want a concurrent
10825          * snapshot to run after we've already checked the extents.
10826          */
10827         atomic_inc(&BTRFS_I(inode)->root->nr_swapfiles);
10828
10829         isize = ALIGN_DOWN(inode->i_size, fs_info->sectorsize);
10830
10831         lock_extent_bits(io_tree, 0, isize - 1, &cached_state);
10832         start = 0;
10833         while (start < isize) {
10834                 u64 logical_block_start, physical_block_start;
10835                 struct btrfs_block_group *bg;
10836                 u64 len = isize - start;
10837
10838                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, start, len, 0);
10839                 if (IS_ERR(em)) {
10840                         ret = PTR_ERR(em);
10841                         goto out;
10842                 }
10843
10844                 if (em->block_start == EXTENT_MAP_HOLE) {
10845                         btrfs_warn(fs_info, "swapfile must not have holes");
10846                         ret = -EINVAL;
10847                         goto out;
10848                 }
10849                 if (em->block_start == EXTENT_MAP_INLINE) {
10850                         /*
10851                          * It's unlikely we'll ever actually find ourselves
10852                          * here, as a file small enough to fit inline won't be
10853                          * big enough to store more than the swap header, but in
10854                          * case something changes in the future, let's catch it
10855                          * here rather than later.
10856                          */
10857                         btrfs_warn(fs_info, "swapfile must not be inline");
10858                         ret = -EINVAL;
10859                         goto out;
10860                 }
10861                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
10862                         btrfs_warn(fs_info, "swapfile must not be compressed");
10863                         ret = -EINVAL;
10864                         goto out;
10865                 }
10866
10867                 logical_block_start = em->block_start + (start - em->start);
10868                 len = min(len, em->len - (start - em->start));
10869                 free_extent_map(em);
10870                 em = NULL;
10871
10872                 ret = can_nocow_extent(inode, start, &len, NULL, NULL, NULL);
10873                 if (ret < 0) {
10874                         goto out;
10875                 } else if (ret) {
10876                         ret = 0;
10877                 } else {
10878                         btrfs_warn(fs_info,
10879                                    "swapfile must not be copy-on-write");
10880                         ret = -EINVAL;
10881                         goto out;
10882                 }
10883
10884                 em = btrfs_get_chunk_map(fs_info, logical_block_start, len);
10885                 if (IS_ERR(em)) {
10886                         ret = PTR_ERR(em);
10887                         goto out;
10888                 }
10889
10890                 if (em->map_lookup->type & BTRFS_BLOCK_GROUP_PROFILE_MASK) {
10891                         btrfs_warn(fs_info,
10892                                    "swapfile must have single data profile");
10893                         ret = -EINVAL;
10894                         goto out;
10895                 }
10896
10897                 if (device == NULL) {
10898                         device = em->map_lookup->stripes[0].dev;
10899                         ret = btrfs_add_swapfile_pin(inode, device, false);
10900                         if (ret == 1)
10901                                 ret = 0;
10902                         else if (ret)
10903                                 goto out;
10904                 } else if (device != em->map_lookup->stripes[0].dev) {
10905                         btrfs_warn(fs_info, "swapfile must be on one device");
10906                         ret = -EINVAL;
10907                         goto out;
10908                 }
10909
10910                 physical_block_start = (em->map_lookup->stripes[0].physical +
10911                                         (logical_block_start - em->start));
10912                 len = min(len, em->len - (logical_block_start - em->start));
10913                 free_extent_map(em);
10914                 em = NULL;
10915
10916                 bg = btrfs_lookup_block_group(fs_info, logical_block_start);
10917                 if (!bg) {
10918                         btrfs_warn(fs_info,
10919                            "could not find block group containing swapfile");
10920                         ret = -EINVAL;
10921                         goto out;
10922                 }
10923
10924                 ret = btrfs_add_swapfile_pin(inode, bg, true);
10925                 if (ret) {
10926                         btrfs_put_block_group(bg);
10927                         if (ret == 1)
10928                                 ret = 0;
10929                         else
10930                                 goto out;
10931                 }
10932
10933                 if (bsi.block_len &&
10934                     bsi.block_start + bsi.block_len == physical_block_start) {
10935                         bsi.block_len += len;
10936                 } else {
10937                         if (bsi.block_len) {
10938                                 ret = btrfs_add_swap_extent(sis, &bsi);
10939                                 if (ret)
10940                                         goto out;
10941                         }
10942                         bsi.start = start;
10943                         bsi.block_start = physical_block_start;
10944                         bsi.block_len = len;
10945                 }
10946
10947                 start += len;
10948         }
10949
10950         if (bsi.block_len)
10951                 ret = btrfs_add_swap_extent(sis, &bsi);
10952
10953 out:
10954         if (!IS_ERR_OR_NULL(em))
10955                 free_extent_map(em);
10956
10957         unlock_extent_cached(io_tree, 0, isize - 1, &cached_state);
10958
10959         if (ret)
10960                 btrfs_swap_deactivate(file);
10961
10962         clear_bit(BTRFS_FS_EXCL_OP, &fs_info->flags);
10963
10964         if (ret)
10965                 return ret;
10966
10967         if (device)
10968                 sis->bdev = device->bdev;
10969         *span = bsi.highest_ppage - bsi.lowest_ppage + 1;
10970         sis->max = bsi.nr_pages;
10971         sis->pages = bsi.nr_pages - 1;
10972         sis->highest_bit = bsi.nr_pages - 1;
10973         return bsi.nr_extents;
10974 }
10975 #else
10976 static void btrfs_swap_deactivate(struct file *file)
10977 {
10978 }
10979
10980 static int btrfs_swap_activate(struct swap_info_struct *sis, struct file *file,
10981                                sector_t *span)
10982 {
10983         return -EOPNOTSUPP;
10984 }
10985 #endif
10986
10987 static const struct inode_operations btrfs_dir_inode_operations = {
10988         .getattr        = btrfs_getattr,
10989         .lookup         = btrfs_lookup,
10990         .create         = btrfs_create,
10991         .unlink         = btrfs_unlink,
10992         .link           = btrfs_link,
10993         .mkdir          = btrfs_mkdir,
10994         .rmdir          = btrfs_rmdir,
10995         .rename         = btrfs_rename2,
10996         .symlink        = btrfs_symlink,
10997         .setattr        = btrfs_setattr,
10998         .mknod          = btrfs_mknod,
10999         .listxattr      = btrfs_listxattr,
11000         .permission     = btrfs_permission,
11001         .get_acl        = btrfs_get_acl,
11002         .set_acl        = btrfs_set_acl,
11003         .update_time    = btrfs_update_time,
11004         .tmpfile        = btrfs_tmpfile,
11005 };
11006 static const struct inode_operations btrfs_dir_ro_inode_operations = {
11007         .lookup         = btrfs_lookup,
11008         .permission     = btrfs_permission,
11009         .update_time    = btrfs_update_time,
11010 };
11011
11012 static const struct file_operations btrfs_dir_file_operations = {
11013         .llseek         = generic_file_llseek,
11014         .read           = generic_read_dir,
11015         .iterate_shared = btrfs_real_readdir,
11016         .open           = btrfs_opendir,
11017         .unlocked_ioctl = btrfs_ioctl,
11018 #ifdef CONFIG_COMPAT
11019         .compat_ioctl   = btrfs_compat_ioctl,
11020 #endif
11021         .release        = btrfs_release_file,
11022         .fsync          = btrfs_sync_file,
11023 };
11024
11025 static const struct extent_io_ops btrfs_extent_io_ops = {
11026         /* mandatory callbacks */
11027         .submit_bio_hook = btrfs_submit_bio_hook,
11028         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
11029 };
11030
11031 /*
11032  * btrfs doesn't support the bmap operation because swapfiles
11033  * use bmap to make a mapping of extents in the file.  They assume
11034  * these extents won't change over the life of the file and they
11035  * use the bmap result to do IO directly to the drive.
11036  *
11037  * the btrfs bmap call would return logical addresses that aren't
11038  * suitable for IO and they also will change frequently as COW
11039  * operations happen.  So, swapfile + btrfs == corruption.
11040  *
11041  * For now we're avoiding this by dropping bmap.
11042  */
11043 static const struct address_space_operations btrfs_aops = {
11044         .readpage       = btrfs_readpage,
11045         .writepage      = btrfs_writepage,
11046         .writepages     = btrfs_writepages,
11047         .readpages      = btrfs_readpages,
11048         .direct_IO      = btrfs_direct_IO,
11049         .invalidatepage = btrfs_invalidatepage,
11050         .releasepage    = btrfs_releasepage,
11051         .set_page_dirty = btrfs_set_page_dirty,
11052         .error_remove_page = generic_error_remove_page,
11053         .swap_activate  = btrfs_swap_activate,
11054         .swap_deactivate = btrfs_swap_deactivate,
11055 };
11056
11057 static const struct inode_operations btrfs_file_inode_operations = {
11058         .getattr        = btrfs_getattr,
11059         .setattr        = btrfs_setattr,
11060         .listxattr      = btrfs_listxattr,
11061         .permission     = btrfs_permission,
11062         .fiemap         = btrfs_fiemap,
11063         .get_acl        = btrfs_get_acl,
11064         .set_acl        = btrfs_set_acl,
11065         .update_time    = btrfs_update_time,
11066 };
11067 static const struct inode_operations btrfs_special_inode_operations = {
11068         .getattr        = btrfs_getattr,
11069         .setattr        = btrfs_setattr,
11070         .permission     = btrfs_permission,
11071         .listxattr      = btrfs_listxattr,
11072         .get_acl        = btrfs_get_acl,
11073         .set_acl        = btrfs_set_acl,
11074         .update_time    = btrfs_update_time,
11075 };
11076 static const struct inode_operations btrfs_symlink_inode_operations = {
11077         .get_link       = page_get_link,
11078         .getattr        = btrfs_getattr,
11079         .setattr        = btrfs_setattr,
11080         .permission     = btrfs_permission,
11081         .listxattr      = btrfs_listxattr,
11082         .update_time    = btrfs_update_time,
11083 };
11084
11085 const struct dentry_operations btrfs_dentry_operations = {
11086         .d_delete       = btrfs_dentry_delete,
11087 };