fs/btrfs/inode.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/kernel.h>
  20 #include <linux/bio.h>
  21 #include <linux/buffer_head.h>
  22 #include <linux/file.h>
  23 #include <linux/fs.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/highmem.h>
  26 #include <linux/time.h>
  27 #include <linux/init.h>
  28 #include <linux/string.h>
  29 #include <linux/backing-dev.h>
  30 #include <linux/mpage.h>
  31 #include <linux/swap.h>
  32 #include <linux/writeback.h>
  33 #include <linux/compat.h>
  34 #include <linux/bit_spinlock.h>
  35 #include <linux/xattr.h>
  36 #include <linux/posix_acl.h>
  37 #include <linux/falloc.h>
  38 #include <linux/slab.h>
  39 #include <linux/ratelimit.h>
  40 #include <linux/mount.h>
  41 #include <linux/btrfs.h>
  42 #include <linux/blkdev.h>
  43 #include <linux/posix_acl_xattr.h>
  44 #include <linux/uio.h>
  45 #include <linux/magic.h>
  46 #include <linux/iversion.h>
  47 #include "ctree.h"
  48 #include "disk-io.h"
  49 #include "transaction.h"
  50 #include "btrfs_inode.h"
  51 #include "print-tree.h"
  52 #include "ordered-data.h"
  53 #include "xattr.h"
  54 #include "tree-log.h"
  55 #include "volumes.h"
  56 #include "compression.h"
  57 #include "locking.h"
  58 #include "free-space-cache.h"
  59 #include "inode-map.h"
  60 #include "backref.h"
  61 #include "hash.h"
  62 #include "props.h"
  63 #include "qgroup.h"
  64 #include "dedupe.h"
  65
  66 struct btrfs_iget_args {
  67         struct btrfs_key *location;
  68         struct btrfs_root *root;
  69 };
  70
  71 struct btrfs_dio_data {
  72         u64 reserve;
  73         u64 unsubmitted_oe_range_start;
  74         u64 unsubmitted_oe_range_end;
  75         int overwrite;
  76 };
  77
  78 static const struct inode_operations btrfs_dir_inode_operations;
  79 static const struct inode_operations btrfs_symlink_inode_operations;
  80 static const struct inode_operations btrfs_dir_ro_inode_operations;
  81 static const struct inode_operations btrfs_special_inode_operations;
  82 static const struct inode_operations btrfs_file_inode_operations;
  83 static const struct address_space_operations btrfs_aops;
  84 static const struct address_space_operations btrfs_symlink_aops;
  85 static const struct file_operations btrfs_dir_file_operations;
  86 static const struct extent_io_ops btrfs_extent_io_ops;
  87
  88 static struct kmem_cache *btrfs_inode_cachep;
  89 struct kmem_cache *btrfs_trans_handle_cachep;
  90 struct kmem_cache *btrfs_path_cachep;
  91 struct kmem_cache *btrfs_free_space_cachep;
  92
  93 #define S_SHIFT 12
  94 static const unsigned char btrfs_type_by_mode[S_IFMT >> S_SHIFT] = {
  95         [S_IFREG >> S_SHIFT]    = BTRFS_FT_REG_FILE,
  96         [S_IFDIR >> S_SHIFT]    = BTRFS_FT_DIR,
  97         [S_IFCHR >> S_SHIFT]    = BTRFS_FT_CHRDEV,
  98         [S_IFBLK >> S_SHIFT]    = BTRFS_FT_BLKDEV,
  99         [S_IFIFO >> S_SHIFT]    = BTRFS_FT_FIFO,
 100         [S_IFSOCK >> S_SHIFT]   = BTRFS_FT_SOCK,
 101         [S_IFLNK >> S_SHIFT]    = BTRFS_FT_SYMLINK,
 102 };
 103
 104 static int btrfs_setsize(struct inode *inode, struct iattr *attr);
 105 static int btrfs_truncate(struct inode *inode);
 106 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent);
 107 static noinline int cow_file_range(struct inode *inode,
 108                                    struct page *locked_page,
 109                                    u64 start, u64 end, u64 delalloc_end,
 110                                    int *page_started, unsigned long *nr_written,
 111                                    int unlock, struct btrfs_dedupe_hash *hash);
 112 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
 113                                        u64 orig_start, u64 block_start,
 114                                        u64 block_len, u64 orig_block_len,
 115                                        u64 ram_bytes, int compress_type,
 116                                        int type);
 117
 118 static void __endio_write_update_ordered(struct inode *inode,
 119                                          const u64 offset, const u64 bytes,
 120                                          const bool uptodate);
 121
 122 /*
 123  * Cleanup all submitted ordered extents in specified range to handle errors
 124  * from the fill_dellaloc() callback.
 125  *
 126  * NOTE: caller must ensure that when an error happens, it can not call
 127  * extent_clear_unlock_delalloc() to clear both the bits EXTENT_DO_ACCOUNTING
 128  * and EXTENT_DELALLOC simultaneously, because that causes the reserved metadata
 129  * to be released, which we want to happen only when finishing the ordered
 130  * extent (btrfs_finish_ordered_io()). Also note that the caller of the
 131  * fill_delalloc() callback already does proper cleanup for the first page of
 132  * the range, that is, it invokes the callback writepage_end_io_hook() for the
 133  * range of the first page.
 134  */
 135 static inline void btrfs_cleanup_ordered_extents(struct inode *inode,
 136                                                  const u64 offset,
 137                                                  const u64 bytes)
 138 {
 139         unsigned long index = offset >> PAGE_SHIFT;
 140         unsigned long end_index = (offset + bytes - 1) >> PAGE_SHIFT;
 141         struct page *page;
 142
 143         while (index <= end_index) {
 144                 page = find_get_page(inode->i_mapping, index);
 145                 index++;
 146                 if (!page)
 147                         continue;
 148                 ClearPagePrivate2(page);
 149                 put_page(page);
 150         }
 151         return __endio_write_update_ordered(inode, offset + PAGE_SIZE,
 152                                             bytes - PAGE_SIZE, false);
 153 }
 154
 155 static int btrfs_dirty_inode(struct inode *inode);
 156
 157 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 158 void btrfs_test_inode_set_ops(struct inode *inode)
 159 {
 160         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
 161 }
 162 #endif
 163
 164 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 165                                      struct inode *inode,  struct inode *dir,
 166                                      const struct qstr *qstr)
 167 {
 168         int err;
 169
 170         err = btrfs_init_acl(trans, inode, dir);
 171         if (!err)
 172                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 173         return err;
 174 }
 175
 176 /*
 177  * this does all the hard work for inserting an inline extent into
 178  * the btree.  The caller should have done a btrfs_drop_extents so that
 179  * no overlapping inline items exist in the btree
 180  */
 181 static int insert_inline_extent(struct btrfs_trans_handle *trans,
 182                                 struct btrfs_path *path, int extent_inserted,
 183                                 struct btrfs_root *root, struct inode *inode,
 184                                 u64 start, size_t size, size_t compressed_size,
 185                                 int compress_type,
 186                                 struct page **compressed_pages)
 187 {
 188         struct extent_buffer *leaf;
 189         struct page *page = NULL;
 190         char *kaddr;
 191         unsigned long ptr;
 192         struct btrfs_file_extent_item *ei;
 193         int ret;
 194         size_t cur_size = size;
 195         unsigned long offset;
 196
 197         if (compressed_size && compressed_pages)
 198                 cur_size = compressed_size;
 199
 200         inode_add_bytes(inode, size);
 201
 202         if (!extent_inserted) {
 203                 struct btrfs_key key;
 204                 size_t datasize;
 205
 206                 key.objectid = btrfs_ino(BTRFS_I(inode));
 207                 key.offset = start;
 208                 key.type = BTRFS_EXTENT_DATA_KEY;
 209
 210                 datasize = btrfs_file_extent_calc_inline_size(cur_size);
 211                 path->leave_spinning = 1;
 212                 ret = btrfs_insert_empty_item(trans, root, path, &key,
 213                                               datasize);
 214                 if (ret)
 215                         goto fail;
 216         }
 217         leaf = path->nodes[0];
 218         ei = btrfs_item_ptr(leaf, path->slots[0],
 219                             struct btrfs_file_extent_item);
 220         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 221         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 222         btrfs_set_file_extent_encryption(leaf, ei, 0);
 223         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 224         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 225         ptr = btrfs_file_extent_inline_start(ei);
 226
 227         if (compress_type != BTRFS_COMPRESS_NONE) {
 228                 struct page *cpage;
 229                 int i = 0;
 230                 while (compressed_size > 0) {
 231                         cpage = compressed_pages[i];
 232                         cur_size = min_t(unsigned long, compressed_size,
 233                                        PAGE_SIZE);
 234
 235                         kaddr = kmap_atomic(cpage);
 236                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 237                         kunmap_atomic(kaddr);
 238
 239                         i++;
 240                         ptr += cur_size;
 241                         compressed_size -= cur_size;
 242                 }
 243                 btrfs_set_file_extent_compression(leaf, ei,
 244                                                   compress_type);
 245         } else {
 246                 page = find_get_page(inode->i_mapping,
 247                                      start >> PAGE_SHIFT);
 248                 btrfs_set_file_extent_compression(leaf, ei, 0);
 249                 kaddr = kmap_atomic(page);
 250                 offset = start & (PAGE_SIZE - 1);
 251                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 252                 kunmap_atomic(kaddr);
 253                 put_page(page);
 254         }
 255         btrfs_mark_buffer_dirty(leaf);
 256         btrfs_release_path(path);
 257
 258         /*
 259          * we're an inline extent, so nobody can
 260          * extend the file past i_size without locking
 261          * a page we already have locked.
 262          *
 263          * We must do any isize and inode updates
 264          * before we unlock the pages.  Otherwise we
 265          * could end up racing with unlink.
 266          */
 267         BTRFS_I(inode)->disk_i_size = inode->i_size;
 268         ret = btrfs_update_inode(trans, root, inode);
 269
 270 fail:
 271         return ret;
 272 }
 273
 274
 275 /*
 276  * conditionally insert an inline extent into the file.  This
 277  * does the checks required to make sure the data is small enough
 278  * to fit as an inline extent.
 279  */
 280 static noinline int cow_file_range_inline(struct btrfs_root *root,
 281                                           struct inode *inode, u64 start,
 282                                           u64 end, size_t compressed_size,
 283                                           int compress_type,
 284                                           struct page **compressed_pages)
 285 {
 286         struct btrfs_fs_info *fs_info = root->fs_info;
 287         struct btrfs_trans_handle *trans;
 288         u64 isize = i_size_read(inode);
 289         u64 actual_end = min(end + 1, isize);
 290         u64 inline_len = actual_end - start;
 291         u64 aligned_end = ALIGN(end, fs_info->sectorsize);
 292         u64 data_len = inline_len;
 293         int ret;
 294         struct btrfs_path *path;
 295         int extent_inserted = 0;
 296         u32 extent_item_size;
 297
 298         if (compressed_size)
 299                 data_len = compressed_size;
 300
 301         if (start > 0 ||
 302             actual_end > fs_info->sectorsize ||
 303             data_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info) ||
 304             (!compressed_size &&
 305             (actual_end & (fs_info->sectorsize - 1)) == 0) ||
 306             end + 1 < isize ||
 307             data_len > fs_info->max_inline) {
 308                 return 1;
 309         }
 310
 311         path = btrfs_alloc_path();
 312         if (!path)
 313                 return -ENOMEM;
 314
 315         trans = btrfs_join_transaction(root);
 316         if (IS_ERR(trans)) {
 317                 btrfs_free_path(path);
 318                 return PTR_ERR(trans);
 319         }
 320         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
 321
 322         if (compressed_size && compressed_pages)
 323                 extent_item_size = btrfs_file_extent_calc_inline_size(
 324                    compressed_size);
 325         else
 326                 extent_item_size = btrfs_file_extent_calc_inline_size(
 327                     inline_len);
 328
 329         ret = __btrfs_drop_extents(trans, root, inode, path,
 330                                    start, aligned_end, NULL,
 331                                    1, 1, extent_item_size, &extent_inserted);
 332         if (ret) {
 333                 btrfs_abort_transaction(trans, ret);
 334                 goto out;
 335         }
 336
 337         if (isize > actual_end)
 338                 inline_len = min_t(u64, isize, actual_end);
 339         ret = insert_inline_extent(trans, path, extent_inserted,
 340                                    root, inode, start,
 341                                    inline_len, compressed_size,
 342                                    compress_type, compressed_pages);
 343         if (ret && ret != -ENOSPC) {
 344                 btrfs_abort_transaction(trans, ret);
 345                 goto out;
 346         } else if (ret == -ENOSPC) {
 347                 ret = 1;
 348                 goto out;
 349         }
 350
 351         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 352         btrfs_drop_extent_cache(BTRFS_I(inode), start, aligned_end - 1, 0);
 353 out:
 354         /*
 355          * Don't forget to free the reserved space, as for inlined extent
 356          * it won't count as data extent, free them directly here.
 357          * And at reserve time, it's always aligned to page size, so
 358          * just free one page here.
 359          */
 360         btrfs_qgroup_free_data(inode, NULL, 0, PAGE_SIZE);
 361         btrfs_free_path(path);
 362         btrfs_end_transaction(trans);
 363         return ret;
 364 }
 365
 366 struct async_extent {
 367         u64 start;
 368         u64 ram_size;
 369         u64 compressed_size;
 370         struct page **pages;
 371         unsigned long nr_pages;
 372         int compress_type;
 373         struct list_head list;
 374 };
 375
 376 struct async_cow {
 377         struct inode *inode;
 378         struct btrfs_root *root;
 379         struct page *locked_page;
 380         u64 start;
 381         u64 end;
 382         unsigned int write_flags;
 383         struct list_head extents;
 384         struct btrfs_work work;
 385 };
 386
 387 static noinline int add_async_extent(struct async_cow *cow,
 388                                      u64 start, u64 ram_size,
 389                                      u64 compressed_size,
 390                                      struct page **pages,
 391                                      unsigned long nr_pages,
 392                                      int compress_type)
 393 {
 394         struct async_extent *async_extent;
 395
 396         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 397         BUG_ON(!async_extent); /* -ENOMEM */
 398         async_extent->start = start;
 399         async_extent->ram_size = ram_size;
 400         async_extent->compressed_size = compressed_size;
 401         async_extent->pages = pages;
 402         async_extent->nr_pages = nr_pages;
 403         async_extent->compress_type = compress_type;
 404         list_add_tail(&async_extent->list, &cow->extents);
 405         return 0;
 406 }
 407
 408 static inline int inode_need_compress(struct inode *inode, u64 start, u64 end)
 409 {
 410         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 411
 412         /* force compress */
 413         if (btrfs_test_opt(fs_info, FORCE_COMPRESS))
 414                 return 1;
 415         /* defrag ioctl */
 416         if (BTRFS_I(inode)->defrag_compress)
 417                 return 1;
 418         /* bad compression ratios */
 419         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
 420                 return 0;
 421         if (btrfs_test_opt(fs_info, COMPRESS) ||
 422             BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS ||
 423             BTRFS_I(inode)->prop_compress)
 424                 return btrfs_compress_heuristic(inode, start, end);
 425         return 0;
 426 }
 427
 428 static inline void inode_should_defrag(struct btrfs_inode *inode,
 429                 u64 start, u64 end, u64 num_bytes, u64 small_write)
 430 {
 431         /* If this is a small write inside eof, kick off a defrag */
 432         if (num_bytes < small_write &&
 433             (start > 0 || end + 1 < inode->disk_i_size))
 434                 btrfs_add_inode_defrag(NULL, inode);
 435 }
 436
 437 /*
 438  * we create compressed extents in two phases.  The first
 439  * phase compresses a range of pages that have already been
 440  * locked (both pages and state bits are locked).
 441  *
 442  * This is done inside an ordered work queue, and the compression
 443  * is spread across many cpus.  The actual IO submission is step
 444  * two, and the ordered work queue takes care of making sure that
 445  * happens in the same order things were put onto the queue by
 446  * writepages and friends.
 447  *
 448  * If this code finds it can't get good compression, it puts an
 449  * entry onto the work queue to write the uncompressed bytes.  This
 450  * makes sure that both compressed inodes and uncompressed inodes
 451  * are written in the same order that the flusher thread sent them
 452  * down.
 453  */
 454 static noinline void compress_file_range(struct inode *inode,
 455                                         struct page *locked_page,
 456                                         u64 start, u64 end,
 457                                         struct async_cow *async_cow,
 458                                         int *num_added)
 459 {
 460         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 461         struct btrfs_root *root = BTRFS_I(inode)->root;
 462         u64 blocksize = fs_info->sectorsize;
 463         u64 actual_end;
 464         u64 isize = i_size_read(inode);
 465         int ret = 0;
 466         struct page **pages = NULL;
 467         unsigned long nr_pages;
 468         unsigned long total_compressed = 0;
 469         unsigned long total_in = 0;
 470         int i;
 471         int will_compress;
 472         int compress_type = fs_info->compress_type;
 473         int redirty = 0;
 474
 475         inode_should_defrag(BTRFS_I(inode), start, end, end - start + 1,
 476                         SZ_16K);
 477
 478         actual_end = min_t(u64, isize, end + 1);
 479 again:
 480         will_compress = 0;
 481         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 482         BUILD_BUG_ON((BTRFS_MAX_COMPRESSED % PAGE_SIZE) != 0);
 483         nr_pages = min_t(unsigned long, nr_pages,
 484                         BTRFS_MAX_COMPRESSED / PAGE_SIZE);
 485
 486         /*
 487          * we don't want to send crud past the end of i_size through
 488          * compression, that's just a waste of CPU time.  So, if the
 489          * end of the file is before the start of our current
 490          * requested range of bytes, we bail out to the uncompressed
 491          * cleanup code that can deal with all of this.
 492          *
 493          * It isn't really the fastest way to fix things, but this is a
 494          * very uncommon corner.
 495          */
 496         if (actual_end <= start)
 497                 goto cleanup_and_bail_uncompressed;
 498
 499         total_compressed = actual_end - start;
 500
 501         /*
 502          * skip compression for a small file range(<=blocksize) that
 503          * isn't an inline extent, since it doesn't save disk space at all.
 504          */
 505         if (total_compressed <= blocksize &&
 506            (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 507                 goto cleanup_and_bail_uncompressed;
 508
 509         total_compressed = min_t(unsigned long, total_compressed,
 510                         BTRFS_MAX_UNCOMPRESSED);
 511         total_in = 0;
 512         ret = 0;
 513
 514         /*
 515          * we do compression for mount -o compress and when the
 516          * inode has not been flagged as nocompress.  This flag can
 517          * change at any time if we discover bad compression ratios.
 518          */
 519         if (inode_need_compress(inode, start, end)) {
 520                 WARN_ON(pages);
 521                 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS);
 522                 if (!pages) {
 523                         /* just bail out to the uncompressed code */
 524                         goto cont;
 525                 }
 526
 527                 if (BTRFS_I(inode)->defrag_compress)
 528                         compress_type = BTRFS_I(inode)->defrag_compress;
 529                 else if (BTRFS_I(inode)->prop_compress)
 530                         compress_type = BTRFS_I(inode)->prop_compress;
 531
 532                 /*
 533                  * we need to call clear_page_dirty_for_io on each
 534                  * page in the range.  Otherwise applications with the file
 535                  * mmap'd can wander in and change the page contents while
 536                  * we are compressing them.
 537                  *
 538                  * If the compression fails for any reason, we set the pages
 539                  * dirty again later on.
 540                  *
 541                  * Note that the remaining part is redirtied, the start pointer
 542                  * has moved, the end is the original one.
 543                  */
 544                 if (!redirty) {
 545                         extent_range_clear_dirty_for_io(inode, start, end);
 546                         redirty = 1;
 547                 }
 548
 549                 /* Compression level is applied here and only here */
 550                 ret = btrfs_compress_pages(
 551                         compress_type | (fs_info->compress_level << 4),
 552                                            inode->i_mapping, start,
 553                                            pages,
 554                                            &nr_pages,
 555                                            &total_in,
 556                                            &total_compressed);
 557
 558                 if (!ret) {
 559                         unsigned long offset = total_compressed &
 560                                 (PAGE_SIZE - 1);
 561                         struct page *page = pages[nr_pages - 1];
 562                         char *kaddr;
 563
 564                         /* zero the tail end of the last page, we might be
 565                          * sending it down to disk
 566                          */
 567                         if (offset) {
 568                                 kaddr = kmap_atomic(page);
 569                                 memset(kaddr + offset, 0,
 570                                        PAGE_SIZE - offset);
 571                                 kunmap_atomic(kaddr);
 572                         }
 573                         will_compress = 1;
 574                 }
 575         }
 576 cont:
 577         if (start == 0) {
 578                 /* lets try to make an inline extent */
 579                 if (ret || total_in < actual_end) {
 580                         /* we didn't compress the entire range, try
 581                          * to make an uncompressed inline extent.
 582                          */
 583                         ret = cow_file_range_inline(root, inode, start, end,
 584                                             0, BTRFS_COMPRESS_NONE, NULL);
 585                 } else {
 586                         /* try making a compressed inline extent */
 587                         ret = cow_file_range_inline(root, inode, start, end,
 588                                                     total_compressed,
 589                                                     compress_type, pages);
 590                 }
 591                 if (ret <= 0) {
 592                         unsigned long clear_flags = EXTENT_DELALLOC |
 593                                 EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
 594                                 EXTENT_DO_ACCOUNTING;
 595                         unsigned long page_error_op;
 596
 597                         page_error_op = ret < 0 ? PAGE_SET_ERROR : 0;
 598
 599                         /*
 600                          * inline extent creation worked or returned error,
 601                          * we don't need to create any more async work items.
 602                          * Unlock and free up our temp pages.
 603                          *
 604                          * We use DO_ACCOUNTING here because we need the
 605                          * delalloc_release_metadata to be done _after_ we drop
 606                          * our outstanding extent for clearing delalloc for this
 607                          * range.
 608                          */
 609                         extent_clear_unlock_delalloc(inode, start, end, end,
 610                                                      NULL, clear_flags,
 611                                                      PAGE_UNLOCK |
 612                                                      PAGE_CLEAR_DIRTY |
 613                                                      PAGE_SET_WRITEBACK |
 614                                                      page_error_op |
 615                                                      PAGE_END_WRITEBACK);
 616                         goto free_pages_out;
 617                 }
 618         }
 619
 620         if (will_compress) {
 621                 /*
 622                  * we aren't doing an inline extent round the compressed size
 623                  * up to a block size boundary so the allocator does sane
 624                  * things
 625                  */
 626                 total_compressed = ALIGN(total_compressed, blocksize);
 627
 628                 /*
 629                  * one last check to make sure the compression is really a
 630                  * win, compare the page count read with the blocks on disk,
 631                  * compression must free at least one sector size
 632                  */
 633                 total_in = ALIGN(total_in, PAGE_SIZE);
 634                 if (total_compressed + blocksize <= total_in) {
 635                         *num_added += 1;
 636
 637                         /*
 638                          * The async work queues will take care of doing actual
 639                          * allocation on disk for these compressed pages, and
 640                          * will submit them to the elevator.
 641                          */
 642                         add_async_extent(async_cow, start, total_in,
 643                                         total_compressed, pages, nr_pages,
 644                                         compress_type);
 645
 646                         if (start + total_in < end) {
 647                                 start += total_in;
 648                                 pages = NULL;
 649                                 cond_resched();
 650                                 goto again;
 651                         }
 652                         return;
 653                 }
 654         }
 655         if (pages) {
 656                 /*
 657                  * the compression code ran but failed to make things smaller,
 658                  * free any pages it allocated and our page pointer array
 659                  */
 660                 for (i = 0; i < nr_pages; i++) {
 661                         WARN_ON(pages[i]->mapping);
 662                         put_page(pages[i]);
 663                 }
 664                 kfree(pages);
 665                 pages = NULL;
 666                 total_compressed = 0;
 667                 nr_pages = 0;
 668
 669                 /* flag the file so we don't compress in the future */
 670                 if (!btrfs_test_opt(fs_info, FORCE_COMPRESS) &&
 671                     !(BTRFS_I(inode)->prop_compress)) {
 672                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 673                 }
 674         }
 675 cleanup_and_bail_uncompressed:
 676         /*
 677          * No compression, but we still need to write the pages in the file
 678          * we've been given so far.  redirty the locked page if it corresponds
 679          * to our extent and set things up for the async work queue to run
 680          * cow_file_range to do the normal delalloc dance.
 681          */
 682         if (page_offset(locked_page) >= start &&
 683             page_offset(locked_page) <= end)
 684                 __set_page_dirty_nobuffers(locked_page);
 685                 /* unlocked later on in the async handlers */
 686
 687         if (redirty)
 688                 extent_range_redirty_for_io(inode, start, end);
 689         add_async_extent(async_cow, start, end - start + 1, 0, NULL, 0,
 690                          BTRFS_COMPRESS_NONE);
 691         *num_added += 1;
 692
 693         return;
 694
 695 free_pages_out:
 696         for (i = 0; i < nr_pages; i++) {
 697                 WARN_ON(pages[i]->mapping);
 698                 put_page(pages[i]);
 699         }
 700         kfree(pages);
 701 }
 702
 703 static void free_async_extent_pages(struct async_extent *async_extent)
 704 {
 705         int i;
 706
 707         if (!async_extent->pages)
 708                 return;
 709
 710         for (i = 0; i < async_extent->nr_pages; i++) {
 711                 WARN_ON(async_extent->pages[i]->mapping);
 712                 put_page(async_extent->pages[i]);
 713         }
 714         kfree(async_extent->pages);
 715         async_extent->nr_pages = 0;
 716         async_extent->pages = NULL;
 717 }
 718
 719 /*
 720  * phase two of compressed writeback.  This is the ordered portion
 721  * of the code, which only gets called in the order the work was
 722  * queued.  We walk all the async extents created by compress_file_range
 723  * and send them down to the disk.
 724  */
 725 static noinline void submit_compressed_extents(struct inode *inode,
 726                                               struct async_cow *async_cow)
 727 {
 728         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 729         struct async_extent *async_extent;
 730         u64 alloc_hint = 0;
 731         struct btrfs_key ins;
 732         struct extent_map *em;
 733         struct btrfs_root *root = BTRFS_I(inode)->root;
 734         struct extent_io_tree *io_tree;
 735         int ret = 0;
 736
 737 again:
 738         while (!list_empty(&async_cow->extents)) {
 739                 async_extent = list_entry(async_cow->extents.next,
 740                                           struct async_extent, list);
 741                 list_del(&async_extent->list);
 742
 743                 io_tree = &BTRFS_I(inode)->io_tree;
 744
 745 retry:
 746                 /* did the compression code fall back to uncompressed IO? */
 747                 if (!async_extent->pages) {
 748                         int page_started = 0;
 749                         unsigned long nr_written = 0;
 750
 751                         lock_extent(io_tree, async_extent->start,
 752                                          async_extent->start +
 753                                          async_extent->ram_size - 1);
 754
 755                         /* allocate blocks */
 756                         ret = cow_file_range(inode, async_cow->locked_page,
 757                                              async_extent->start,
 758                                              async_extent->start +
 759                                              async_extent->ram_size - 1,
 760                                              async_extent->start +
 761                                              async_extent->ram_size - 1,
 762                                              &page_started, &nr_written, 0,
 763                                              NULL);
 764
 765                         /* JDM XXX */
 766
 767                         /*
 768                          * if page_started, cow_file_range inserted an
 769                          * inline extent and took care of all the unlocking
 770                          * and IO for us.  Otherwise, we need to submit
 771                          * all those pages down to the drive.
 772                          */
 773                         if (!page_started && !ret)
 774                                 extent_write_locked_range(inode,
 775                                                   async_extent->start,
 776                                                   async_extent->start +
 777                                                   async_extent->ram_size - 1,
 778                                                   WB_SYNC_ALL);
 779                         else if (ret)
 780                                 unlock_page(async_cow->locked_page);
 781                         kfree(async_extent);
 782                         cond_resched();
 783                         continue;
 784                 }
 785
 786                 lock_extent(io_tree, async_extent->start,
 787                             async_extent->start + async_extent->ram_size - 1);
 788
 789                 ret = btrfs_reserve_extent(root, async_extent->ram_size,
 790                                            async_extent->compressed_size,
 791                                            async_extent->compressed_size,
 792                                            0, alloc_hint, &ins, 1, 1);
 793                 if (ret) {
 794                         free_async_extent_pages(async_extent);
 795
 796                         if (ret == -ENOSPC) {
 797                                 unlock_extent(io_tree, async_extent->start,
 798                                               async_extent->start +
 799                                               async_extent->ram_size - 1);
 800
 801                                 /*
 802                                  * we need to redirty the pages if we decide to
 803                                  * fallback to uncompressed IO, otherwise we
 804                                  * will not submit these pages down to lower
 805                                  * layers.
 806                                  */
 807                                 extent_range_redirty_for_io(inode,
 808                                                 async_extent->start,
 809                                                 async_extent->start +
 810                                                 async_extent->ram_size - 1);
 811
 812                                 goto retry;
 813                         }
 814                         goto out_free;
 815                 }
 816                 /*
 817                  * here we're doing allocation and writeback of the
 818                  * compressed pages
 819                  */
 820                 em = create_io_em(inode, async_extent->start,
 821                                   async_extent->ram_size, /* len */
 822                                   async_extent->start, /* orig_start */
 823                                   ins.objectid, /* block_start */
 824                                   ins.offset, /* block_len */
 825                                   ins.offset, /* orig_block_len */
 826                                   async_extent->ram_size, /* ram_bytes */
 827                                   async_extent->compress_type,
 828                                   BTRFS_ORDERED_COMPRESSED);
 829                 if (IS_ERR(em))
 830                         /* ret value is not necessary due to void function */
 831                         goto out_free_reserve;
 832                 free_extent_map(em);
 833
 834                 ret = btrfs_add_ordered_extent_compress(inode,
 835                                                 async_extent->start,
 836                                                 ins.objectid,
 837                                                 async_extent->ram_size,
 838                                                 ins.offset,
 839                                                 BTRFS_ORDERED_COMPRESSED,
 840                                                 async_extent->compress_type);
 841                 if (ret) {
 842                         btrfs_drop_extent_cache(BTRFS_I(inode),
 843                                                 async_extent->start,
 844                                                 async_extent->start +
 845                                                 async_extent->ram_size - 1, 0);
 846                         goto out_free_reserve;
 847                 }
 848                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 849
 850                 /*
 851                  * clear dirty, set writeback and unlock the pages.
 852                  */
 853                 extent_clear_unlock_delalloc(inode, async_extent->start,
 854                                 async_extent->start +
 855                                 async_extent->ram_size - 1,
 856                                 async_extent->start +
 857                                 async_extent->ram_size - 1,
 858                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
 859                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 860                                 PAGE_SET_WRITEBACK);
 861                 if (btrfs_submit_compressed_write(inode,
 862                                     async_extent->start,
 863                                     async_extent->ram_size,
 864                                     ins.objectid,
 865                                     ins.offset, async_extent->pages,
 866                                     async_extent->nr_pages,
 867                                     async_cow->write_flags)) {
 868                         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
 869                         struct page *p = async_extent->pages[0];
 870                         const u64 start = async_extent->start;
 871                         const u64 end = start + async_extent->ram_size - 1;
 872
 873                         p->mapping = inode->i_mapping;
 874                         tree->ops->writepage_end_io_hook(p, start, end,
 875                                                          NULL, 0);
 876                         p->mapping = NULL;
 877                         extent_clear_unlock_delalloc(inode, start, end, end,
 878                                                      NULL, 0,
 879                                                      PAGE_END_WRITEBACK |
 880                                                      PAGE_SET_ERROR);
 881                         free_async_extent_pages(async_extent);
 882                 }
 883                 alloc_hint = ins.objectid + ins.offset;
 884                 kfree(async_extent);
 885                 cond_resched();
 886         }
 887         return;
 888 out_free_reserve:
 889         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 890         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
 891 out_free:
 892         extent_clear_unlock_delalloc(inode, async_extent->start,
 893                                      async_extent->start +
 894                                      async_extent->ram_size - 1,
 895                                      async_extent->start +
 896                                      async_extent->ram_size - 1,
 897                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
 898                                      EXTENT_DELALLOC_NEW |
 899                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
 900                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 901                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
 902                                      PAGE_SET_ERROR);
 903         free_async_extent_pages(async_extent);
 904         kfree(async_extent);
 905         goto again;
 906 }
 907
 908 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 909                                       u64 num_bytes)
 910 {
 911         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 912         struct extent_map *em;
 913         u64 alloc_hint = 0;
 914
 915         read_lock(&em_tree->lock);
 916         em = search_extent_mapping(em_tree, start, num_bytes);
 917         if (em) {
 918                 /*
 919                  * if block start isn't an actual block number then find the
 920                  * first block in this inode and use that as a hint.  If that
 921                  * block is also bogus then just don't worry about it.
 922                  */
 923                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 924                         free_extent_map(em);
 925                         em = search_extent_mapping(em_tree, 0, 0);
 926                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 927                                 alloc_hint = em->block_start;
 928                         if (em)
 929                                 free_extent_map(em);
 930                 } else {
 931                         alloc_hint = em->block_start;
 932                         free_extent_map(em);
 933                 }
 934         }
 935         read_unlock(&em_tree->lock);
 936
 937         return alloc_hint;
 938 }
 939
 940 /*
 941  * when extent_io.c finds a delayed allocation range in the file,
 942  * the call backs end up in this code.  The basic idea is to
 943  * allocate extents on disk for the range, and create ordered data structs
 944  * in ram to track those extents.
 945  *
 946  * locked_page is the page that writepage had locked already.  We use
 947  * it to make sure we don't do extra locks or unlocks.
 948  *
 949  * *page_started is set to one if we unlock locked_page and do everything
 950  * required to start IO on it.  It may be clean and already done with
 951  * IO when we return.
 952  */
 953 static noinline int cow_file_range(struct inode *inode,
 954                                    struct page *locked_page,
 955                                    u64 start, u64 end, u64 delalloc_end,
 956                                    int *page_started, unsigned long *nr_written,
 957                                    int unlock, struct btrfs_dedupe_hash *hash)
 958 {
 959         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 960         struct btrfs_root *root = BTRFS_I(inode)->root;
 961         u64 alloc_hint = 0;
 962         u64 num_bytes;
 963         unsigned long ram_size;
 964         u64 disk_num_bytes;
 965         u64 cur_alloc_size = 0;
 966         u64 blocksize = fs_info->sectorsize;
 967         struct btrfs_key ins;
 968         struct extent_map *em;
 969         unsigned clear_bits;
 970         unsigned long page_ops;
 971         bool extent_reserved = false;
 972         int ret = 0;
 973
 974         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
 975                 WARN_ON_ONCE(1);
 976                 ret = -EINVAL;
 977                 goto out_unlock;
 978         }
 979
 980         num_bytes = ALIGN(end - start + 1, blocksize);
 981         num_bytes = max(blocksize,  num_bytes);
 982         disk_num_bytes = num_bytes;
 983
 984         inode_should_defrag(BTRFS_I(inode), start, end, num_bytes, SZ_64K);
 985
 986         if (start == 0) {
 987                 /* lets try to make an inline extent */
 988                 ret = cow_file_range_inline(root, inode, start, end, 0,
 989                                         BTRFS_COMPRESS_NONE, NULL);
 990                 if (ret == 0) {
 991                         /*
 992                          * We use DO_ACCOUNTING here because we need the
 993                          * delalloc_release_metadata to be run _after_ we drop
 994                          * our outstanding extent for clearing delalloc for this
 995                          * range.
 996                          */
 997                         extent_clear_unlock_delalloc(inode, start, end,
 998                                      delalloc_end, NULL,
 999                                      EXTENT_LOCKED | EXTENT_DELALLOC |
1000                                      EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
1001                                      EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1002                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
1003                                      PAGE_END_WRITEBACK);
1004                         *nr_written = *nr_written +
1005                              (end - start + PAGE_SIZE) / PAGE_SIZE;
1006                         *page_started = 1;
1007                         goto out;
1008                 } else if (ret < 0) {
1009                         goto out_unlock;
1010                 }
1011         }
1012
1013         BUG_ON(disk_num_bytes >
1014                btrfs_super_total_bytes(fs_info->super_copy));
1015
1016         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
1017         btrfs_drop_extent_cache(BTRFS_I(inode), start,
1018                         start + num_bytes - 1, 0);
1019
1020         while (disk_num_bytes > 0) {
1021                 cur_alloc_size = disk_num_bytes;
1022                 ret = btrfs_reserve_extent(root, cur_alloc_size, cur_alloc_size,
1023                                            fs_info->sectorsize, 0, alloc_hint,
1024                                            &ins, 1, 1);
1025                 if (ret < 0)
1026                         goto out_unlock;
1027                 cur_alloc_size = ins.offset;
1028                 extent_reserved = true;
1029
1030                 ram_size = ins.offset;
1031                 em = create_io_em(inode, start, ins.offset, /* len */
1032                                   start, /* orig_start */
1033                                   ins.objectid, /* block_start */
1034                                   ins.offset, /* block_len */
1035                                   ins.offset, /* orig_block_len */
1036                                   ram_size, /* ram_bytes */
1037                                   BTRFS_COMPRESS_NONE, /* compress_type */
1038                                   BTRFS_ORDERED_REGULAR /* type */);
1039                 if (IS_ERR(em))
1040                         goto out_reserve;
1041                 free_extent_map(em);
1042
1043                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
1044                                                ram_size, cur_alloc_size, 0);
1045                 if (ret)
1046                         goto out_drop_extent_cache;
1047
1048                 if (root->root_key.objectid ==
1049                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1050                         ret = btrfs_reloc_clone_csums(inode, start,
1051                                                       cur_alloc_size);
1052                         /*
1053                          * Only drop cache here, and process as normal.
1054                          *
1055                          * We must not allow extent_clear_unlock_delalloc()
1056                          * at out_unlock label to free meta of this ordered
1057                          * extent, as its meta should be freed by
1058                          * btrfs_finish_ordered_io().
1059                          *
1060                          * So we must continue until @start is increased to
1061                          * skip current ordered extent.
1062                          */
1063                         if (ret)
1064                                 btrfs_drop_extent_cache(BTRFS_I(inode), start,
1065                                                 start + ram_size - 1, 0);
1066                 }
1067
1068                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1069
1070                 /* we're not doing compressed IO, don't unlock the first
1071                  * page (which the caller expects to stay locked), don't
1072                  * clear any dirty bits and don't set any writeback bits
1073                  *
1074                  * Do set the Private2 bit so we know this page was properly
1075                  * setup for writepage
1076                  */
1077                 page_ops = unlock ? PAGE_UNLOCK : 0;
1078                 page_ops |= PAGE_SET_PRIVATE2;
1079
1080                 extent_clear_unlock_delalloc(inode, start,
1081                                              start + ram_size - 1,
1082                                              delalloc_end, locked_page,
1083                                              EXTENT_LOCKED | EXTENT_DELALLOC,
1084                                              page_ops);
1085                 if (disk_num_bytes < cur_alloc_size)
1086                         disk_num_bytes = 0;
1087                 else
1088                         disk_num_bytes -= cur_alloc_size;
1089                 num_bytes -= cur_alloc_size;
1090                 alloc_hint = ins.objectid + ins.offset;
1091                 start += cur_alloc_size;
1092                 extent_reserved = false;
1093
1094                 /*
1095                  * btrfs_reloc_clone_csums() error, since start is increased
1096                  * extent_clear_unlock_delalloc() at out_unlock label won't
1097                  * free metadata of current ordered extent, we're OK to exit.
1098                  */
1099                 if (ret)
1100                         goto out_unlock;
1101         }
1102 out:
1103         return ret;
1104
1105 out_drop_extent_cache:
1106         btrfs_drop_extent_cache(BTRFS_I(inode), start, start + ram_size - 1, 0);
1107 out_reserve:
1108         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1109         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
1110 out_unlock:
1111         clear_bits = EXTENT_LOCKED | EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
1112                 EXTENT_DEFRAG | EXTENT_CLEAR_META_RESV;
1113         page_ops = PAGE_UNLOCK | PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
1114                 PAGE_END_WRITEBACK;
1115         /*
1116          * If we reserved an extent for our delalloc range (or a subrange) and
1117          * failed to create the respective ordered extent, then it means that
1118          * when we reserved the extent we decremented the extent's size from
1119          * the data space_info's bytes_may_use counter and incremented the
1120          * space_info's bytes_reserved counter by the same amount. We must make
1121          * sure extent_clear_unlock_delalloc() does not try to decrement again
1122          * the data space_info's bytes_may_use counter, therefore we do not pass
1123          * it the flag EXTENT_CLEAR_DATA_RESV.
1124          */
1125         if (extent_reserved) {
1126                 extent_clear_unlock_delalloc(inode, start,
1127                                              start + cur_alloc_size,
1128                                              start + cur_alloc_size,
1129                                              locked_page,
1130                                              clear_bits,
1131                                              page_ops);
1132                 start += cur_alloc_size;
1133                 if (start >= end)
1134                         goto out;
1135         }
1136         extent_clear_unlock_delalloc(inode, start, end, delalloc_end,
1137                                      locked_page,
1138                                      clear_bits | EXTENT_CLEAR_DATA_RESV,
1139                                      page_ops);
1140         goto out;
1141 }
1142
1143 /*
1144  * work queue call back to started compression on a file and pages
1145  */
1146 static noinline void async_cow_start(struct btrfs_work *work)
1147 {
1148         struct async_cow *async_cow;
1149         int num_added = 0;
1150         async_cow = container_of(work, struct async_cow, work);
1151
1152         compress_file_range(async_cow->inode, async_cow->locked_page,
1153                             async_cow->start, async_cow->end, async_cow,
1154                             &num_added);
1155         if (num_added == 0) {
1156                 btrfs_add_delayed_iput(async_cow->inode);
1157                 async_cow->inode = NULL;
1158         }
1159 }
1160
1161 /*
1162  * work queue call back to submit previously compressed pages
1163  */
1164 static noinline void async_cow_submit(struct btrfs_work *work)
1165 {
1166         struct btrfs_fs_info *fs_info;
1167         struct async_cow *async_cow;
1168         struct btrfs_root *root;
1169         unsigned long nr_pages;
1170
1171         async_cow = container_of(work, struct async_cow, work);
1172
1173         root = async_cow->root;
1174         fs_info = root->fs_info;
1175         nr_pages = (async_cow->end - async_cow->start + PAGE_SIZE) >>
1176                 PAGE_SHIFT;
1177
1178         /*
1179          * atomic_sub_return implies a barrier for waitqueue_active
1180          */
1181         if (atomic_sub_return(nr_pages, &fs_info->async_delalloc_pages) <
1182             5 * SZ_1M &&
1183             waitqueue_active(&fs_info->async_submit_wait))
1184                 wake_up(&fs_info->async_submit_wait);
1185
1186         if (async_cow->inode)
1187                 submit_compressed_extents(async_cow->inode, async_cow);
1188 }
1189
1190 static noinline void async_cow_free(struct btrfs_work *work)
1191 {
1192         struct async_cow *async_cow;
1193         async_cow = container_of(work, struct async_cow, work);
1194         if (async_cow->inode)
1195                 btrfs_add_delayed_iput(async_cow->inode);
1196         kfree(async_cow);
1197 }
1198
1199 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
1200                                 u64 start, u64 end, int *page_started,
1201                                 unsigned long *nr_written,
1202                                 unsigned int write_flags)
1203 {
1204         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1205         struct async_cow *async_cow;
1206         struct btrfs_root *root = BTRFS_I(inode)->root;
1207         unsigned long nr_pages;
1208         u64 cur_end;
1209
1210         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
1211                          1, 0, NULL);
1212         while (start < end) {
1213                 async_cow = kmalloc(sizeof(*async_cow), GFP_NOFS);
1214                 BUG_ON(!async_cow); /* -ENOMEM */
1215                 async_cow->inode = igrab(inode);
1216                 async_cow->root = root;
1217                 async_cow->locked_page = locked_page;
1218                 async_cow->start = start;
1219                 async_cow->write_flags = write_flags;
1220
1221                 if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1222                     !btrfs_test_opt(fs_info, FORCE_COMPRESS))
1223                         cur_end = end;
1224                 else
1225                         cur_end = min(end, start + SZ_512K - 1);
1226
1227                 async_cow->end = cur_end;
1228                 INIT_LIST_HEAD(&async_cow->extents);
1229
1230                 btrfs_init_work(&async_cow->work,
1231                                 btrfs_delalloc_helper,
1232                                 async_cow_start, async_cow_submit,
1233                                 async_cow_free);
1234
1235                 nr_pages = (cur_end - start + PAGE_SIZE) >>
1236                         PAGE_SHIFT;
1237                 atomic_add(nr_pages, &fs_info->async_delalloc_pages);
1238
1239                 btrfs_queue_work(fs_info->delalloc_workers, &async_cow->work);
1240
1241                 *nr_written += nr_pages;
1242                 start = cur_end + 1;
1243         }
1244         *page_started = 1;
1245         return 0;
1246 }
1247
1248 static noinline int csum_exist_in_range(struct btrfs_fs_info *fs_info,
1249                                         u64 bytenr, u64 num_bytes)
1250 {
1251         int ret;
1252         struct btrfs_ordered_sum *sums;
1253         LIST_HEAD(list);
1254
1255         ret = btrfs_lookup_csums_range(fs_info->csum_root, bytenr,
1256                                        bytenr + num_bytes - 1, &list, 0);
1257         if (ret == 0 && list_empty(&list))
1258                 return 0;
1259
1260         while (!list_empty(&list)) {
1261                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1262                 list_del(&sums->list);
1263                 kfree(sums);
1264         }
1265         return 1;
1266 }
1267
1268 /*
1269  * when nowcow writeback call back.  This checks for snapshots or COW copies
1270  * of the extents that exist in the file, and COWs the file as required.
1271  *
1272  * If no cow copies or snapshots exist, we write directly to the existing
1273  * blocks on disk
1274  */
1275 static noinline int run_delalloc_nocow(struct inode *inode,
1276                                        struct page *locked_page,
1277                               u64 start, u64 end, int *page_started, int force,
1278                               unsigned long *nr_written)
1279 {
1280         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1281         struct btrfs_root *root = BTRFS_I(inode)->root;
1282         struct extent_buffer *leaf;
1283         struct btrfs_path *path;
1284         struct btrfs_file_extent_item *fi;
1285         struct btrfs_key found_key;
1286         struct extent_map *em;
1287         u64 cow_start;
1288         u64 cur_offset;
1289         u64 extent_end;
1290         u64 extent_offset;
1291         u64 disk_bytenr;
1292         u64 num_bytes;
1293         u64 disk_num_bytes;
1294         u64 ram_bytes;
1295         int extent_type;
1296         int ret, err;
1297         int type;
1298         int nocow;
1299         int check_prev = 1;
1300         bool nolock;
1301         u64 ino = btrfs_ino(BTRFS_I(inode));
1302
1303         path = btrfs_alloc_path();
1304         if (!path) {
1305                 extent_clear_unlock_delalloc(inode, start, end, end,
1306                                              locked_page,
1307                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1308                                              EXTENT_DO_ACCOUNTING |
1309                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1310                                              PAGE_CLEAR_DIRTY |
1311                                              PAGE_SET_WRITEBACK |
1312                                              PAGE_END_WRITEBACK);
1313                 return -ENOMEM;
1314         }
1315
1316         nolock = btrfs_is_free_space_inode(BTRFS_I(inode));
1317
1318         cow_start = (u64)-1;
1319         cur_offset = start;
1320         while (1) {
1321                 ret = btrfs_lookup_file_extent(NULL, root, path, ino,
1322                                                cur_offset, 0);
1323                 if (ret < 0)
1324                         goto error;
1325                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1326                         leaf = path->nodes[0];
1327                         btrfs_item_key_to_cpu(leaf, &found_key,
1328                                               path->slots[0] - 1);
1329                         if (found_key.objectid == ino &&
1330                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1331                                 path->slots[0]--;
1332                 }
1333                 check_prev = 0;
1334 next_slot:
1335                 leaf = path->nodes[0];
1336                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1337                         ret = btrfs_next_leaf(root, path);
1338                         if (ret < 0) {
1339                                 if (cow_start != (u64)-1)
1340                                         cur_offset = cow_start;
1341                                 goto error;
1342                         }
1343                         if (ret > 0)
1344                                 break;
1345                         leaf = path->nodes[0];
1346                 }
1347
1348                 nocow = 0;
1349                 disk_bytenr = 0;
1350                 num_bytes = 0;
1351                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1352
1353                 if (found_key.objectid > ino)
1354                         break;
1355                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1356                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1357                         path->slots[0]++;
1358                         goto next_slot;
1359                 }
1360                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1361                     found_key.offset > end)
1362                         break;
1363
1364                 if (found_key.offset > cur_offset) {
1365                         extent_end = found_key.offset;
1366                         extent_type = 0;
1367                         goto out_check;
1368                 }
1369
1370                 fi = btrfs_item_ptr(leaf, path->slots[0],
1371                                     struct btrfs_file_extent_item);
1372                 extent_type = btrfs_file_extent_type(leaf, fi);
1373
1374                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1375                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1376                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1377                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1378                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1379                         extent_end = found_key.offset +
1380                                 btrfs_file_extent_num_bytes(leaf, fi);
1381                         disk_num_bytes =
1382                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1383                         if (extent_end <= start) {
1384                                 path->slots[0]++;
1385                                 goto next_slot;
1386                         }
1387                         if (disk_bytenr == 0)
1388                                 goto out_check;
1389                         if (btrfs_file_extent_compression(leaf, fi) ||
1390                             btrfs_file_extent_encryption(leaf, fi) ||
1391                             btrfs_file_extent_other_encoding(leaf, fi))
1392                                 goto out_check;
1393                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1394                                 goto out_check;
1395                         if (btrfs_extent_readonly(fs_info, disk_bytenr))
1396                                 goto out_check;
1397                         if (btrfs_cross_ref_exist(root, ino,
1398                                                   found_key.offset -
1399                                                   extent_offset, disk_bytenr))
1400                                 goto out_check;
1401                         disk_bytenr += extent_offset;
1402                         disk_bytenr += cur_offset - found_key.offset;
1403                         num_bytes = min(end + 1, extent_end) - cur_offset;
1404                         /*
1405                          * if there are pending snapshots for this root,
1406                          * we fall into common COW way.
1407                          */
1408                         if (!nolock) {
1409                                 err = btrfs_start_write_no_snapshotting(root);
1410                                 if (!err)
1411                                         goto out_check;
1412                         }
1413                         /*
1414                          * force cow if csum exists in the range.
1415                          * this ensure that csum for a given extent are
1416                          * either valid or do not exist.
1417                          */
1418                         if (csum_exist_in_range(fs_info, disk_bytenr,
1419                                                 num_bytes)) {
1420                                 if (!nolock)
1421                                         btrfs_end_write_no_snapshotting(root);
1422                                 goto out_check;
1423                         }
1424                         if (!btrfs_inc_nocow_writers(fs_info, disk_bytenr)) {
1425                                 if (!nolock)
1426                                         btrfs_end_write_no_snapshotting(root);
1427                                 goto out_check;
1428                         }
1429                         nocow = 1;
1430                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1431                         extent_end = found_key.offset +
1432                                 btrfs_file_extent_inline_len(leaf,
1433                                                      path->slots[0], fi);
1434                         extent_end = ALIGN(extent_end,
1435                                            fs_info->sectorsize);
1436                 } else {
1437                         BUG_ON(1);
1438                 }
1439 out_check:
1440                 if (extent_end <= start) {
1441                         path->slots[0]++;
1442                         if (!nolock && nocow)
1443                                 btrfs_end_write_no_snapshotting(root);
1444                         if (nocow)
1445                                 btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1446                         goto next_slot;
1447                 }
1448                 if (!nocow) {
1449                         if (cow_start == (u64)-1)
1450                                 cow_start = cur_offset;
1451                         cur_offset = extent_end;
1452                         if (cur_offset > end)
1453                                 break;
1454                         path->slots[0]++;
1455                         goto next_slot;
1456                 }
1457
1458                 btrfs_release_path(path);
1459                 if (cow_start != (u64)-1) {
1460                         ret = cow_file_range(inode, locked_page,
1461                                              cow_start, found_key.offset - 1,
1462                                              end, page_started, nr_written, 1,
1463                                              NULL);
1464                         if (ret) {
1465                                 if (!nolock && nocow)
1466                                         btrfs_end_write_no_snapshotting(root);
1467                                 if (nocow)
1468                                         btrfs_dec_nocow_writers(fs_info,
1469                                                                 disk_bytenr);
1470                                 goto error;
1471                         }
1472                         cow_start = (u64)-1;
1473                 }
1474
1475                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1476                         u64 orig_start = found_key.offset - extent_offset;
1477
1478                         em = create_io_em(inode, cur_offset, num_bytes,
1479                                           orig_start,
1480                                           disk_bytenr, /* block_start */
1481                                           num_bytes, /* block_len */
1482                                           disk_num_bytes, /* orig_block_len */
1483                                           ram_bytes, BTRFS_COMPRESS_NONE,
1484                                           BTRFS_ORDERED_PREALLOC);
1485                         if (IS_ERR(em)) {
1486                                 if (!nolock && nocow)
1487                                         btrfs_end_write_no_snapshotting(root);
1488                                 if (nocow)
1489                                         btrfs_dec_nocow_writers(fs_info,
1490                                                                 disk_bytenr);
1491                                 ret = PTR_ERR(em);
1492                                 goto error;
1493                         }
1494                         free_extent_map(em);
1495                 }
1496
1497                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1498                         type = BTRFS_ORDERED_PREALLOC;
1499                 } else {
1500                         type = BTRFS_ORDERED_NOCOW;
1501                 }
1502
1503                 ret = btrfs_add_ordered_extent(inode, cur_offset, disk_bytenr,
1504                                                num_bytes, num_bytes, type);
1505                 if (nocow)
1506                         btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1507                 BUG_ON(ret); /* -ENOMEM */
1508
1509                 if (root->root_key.objectid ==
1510                     BTRFS_DATA_RELOC_TREE_OBJECTID)
1511                         /*
1512                          * Error handled later, as we must prevent
1513                          * extent_clear_unlock_delalloc() in error handler
1514                          * from freeing metadata of created ordered extent.
1515                          */
1516                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1517                                                       num_bytes);
1518
1519                 extent_clear_unlock_delalloc(inode, cur_offset,
1520                                              cur_offset + num_bytes - 1, end,
1521                                              locked_page, EXTENT_LOCKED |
1522                                              EXTENT_DELALLOC |
1523                                              EXTENT_CLEAR_DATA_RESV,
1524                                              PAGE_UNLOCK | PAGE_SET_PRIVATE2);
1525
1526                 if (!nolock && nocow)
1527                         btrfs_end_write_no_snapshotting(root);
1528                 cur_offset = extent_end;
1529
1530                 /*
1531                  * btrfs_reloc_clone_csums() error, now we're OK to call error
1532                  * handler, as metadata for created ordered extent will only
1533                  * be freed by btrfs_finish_ordered_io().
1534                  */
1535                 if (ret)
1536                         goto error;
1537                 if (cur_offset > end)
1538                         break;
1539         }
1540         btrfs_release_path(path);
1541
1542         if (cur_offset <= end && cow_start == (u64)-1) {
1543                 cow_start = cur_offset;
1544                 cur_offset = end;
1545         }
1546
1547         if (cow_start != (u64)-1) {
1548                 ret = cow_file_range(inode, locked_page, cow_start, end, end,
1549                                      page_started, nr_written, 1, NULL);
1550                 if (ret)
1551                         goto error;
1552         }
1553
1554 error:
1555         if (ret && cur_offset < end)
1556                 extent_clear_unlock_delalloc(inode, cur_offset, end, end,
1557                                              locked_page, EXTENT_LOCKED |
1558                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1559                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1560                                              PAGE_CLEAR_DIRTY |
1561                                              PAGE_SET_WRITEBACK |
1562                                              PAGE_END_WRITEBACK);
1563         btrfs_free_path(path);
1564         return ret;
1565 }
1566
1567 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1568 {
1569
1570         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1571             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1572                 return 0;
1573
1574         /*
1575          * @defrag_bytes is a hint value, no spinlock held here,
1576          * if is not zero, it means the file is defragging.
1577          * Force cow if given extent needs to be defragged.
1578          */
1579         if (BTRFS_I(inode)->defrag_bytes &&
1580             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1581                            EXTENT_DEFRAG, 0, NULL))
1582                 return 1;
1583
1584         return 0;
1585 }
1586
1587 /*
1588  * extent_io.c call back to do delayed allocation processing
1589  */
1590 static int run_delalloc_range(void *private_data, struct page *locked_page,
1591                               u64 start, u64 end, int *page_started,
1592                               unsigned long *nr_written,
1593                               struct writeback_control *wbc)
1594 {
1595         struct inode *inode = private_data;
1596         int ret;
1597         int force_cow = need_force_cow(inode, start, end);
1598         unsigned int write_flags = wbc_to_write_flags(wbc);
1599
1600         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1601                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1602                                          page_started, 1, nr_written);
1603         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1604                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1605                                          page_started, 0, nr_written);
1606         } else if (!inode_need_compress(inode, start, end)) {
1607                 ret = cow_file_range(inode, locked_page, start, end, end,
1608                                       page_started, nr_written, 1, NULL);
1609         } else {
1610                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1611                         &BTRFS_I(inode)->runtime_flags);
1612                 ret = cow_file_range_async(inode, locked_page, start, end,
1613                                            page_started, nr_written,
1614                                            write_flags);
1615         }
1616         if (ret)
1617                 btrfs_cleanup_ordered_extents(inode, start, end - start + 1);
1618         return ret;
1619 }
1620
1621 static void btrfs_split_extent_hook(void *private_data,
1622                                     struct extent_state *orig, u64 split)
1623 {
1624         struct inode *inode = private_data;
1625         u64 size;
1626
1627         /* not delalloc, ignore it */
1628         if (!(orig->state & EXTENT_DELALLOC))
1629                 return;
1630
1631         size = orig->end - orig->start + 1;
1632         if (size > BTRFS_MAX_EXTENT_SIZE) {
1633                 u32 num_extents;
1634                 u64 new_size;
1635
1636                 /*
1637                  * See the explanation in btrfs_merge_extent_hook, the same
1638                  * applies here, just in reverse.
1639                  */
1640                 new_size = orig->end - split + 1;
1641                 num_extents = count_max_extents(new_size);
1642                 new_size = split - orig->start;
1643                 num_extents += count_max_extents(new_size);
1644                 if (count_max_extents(size) >= num_extents)
1645                         return;
1646         }
1647
1648         spin_lock(&BTRFS_I(inode)->lock);
1649         btrfs_mod_outstanding_extents(BTRFS_I(inode), 1);
1650         spin_unlock(&BTRFS_I(inode)->lock);
1651 }
1652
1653 /*
1654  * extent_io.c merge_extent_hook, used to track merged delayed allocation
1655  * extents so we can keep track of new extents that are just merged onto old
1656  * extents, such as when we are doing sequential writes, so we can properly
1657  * account for the metadata space we'll need.
1658  */
1659 static void btrfs_merge_extent_hook(void *private_data,
1660                                     struct extent_state *new,
1661                                     struct extent_state *other)
1662 {
1663         struct inode *inode = private_data;
1664         u64 new_size, old_size;
1665         u32 num_extents;
1666
1667         /* not delalloc, ignore it */
1668         if (!(other->state & EXTENT_DELALLOC))
1669                 return;
1670
1671         if (new->start > other->start)
1672                 new_size = new->end - other->start + 1;
1673         else
1674                 new_size = other->end - new->start + 1;
1675
1676         /* we're not bigger than the max, unreserve the space and go */
1677         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1678                 spin_lock(&BTRFS_I(inode)->lock);
1679                 btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1680                 spin_unlock(&BTRFS_I(inode)->lock);
1681                 return;
1682         }
1683
1684         /*
1685          * We have to add up either side to figure out how many extents were
1686          * accounted for before we merged into one big extent.  If the number of
1687          * extents we accounted for is <= the amount we need for the new range
1688          * then we can return, otherwise drop.  Think of it like this
1689          *
1690          * [ 4k][MAX_SIZE]
1691          *
1692          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1693          * need 2 outstanding extents, on one side we have 1 and the other side
1694          * we have 1 so they are == and we can return.  But in this case
1695          *
1696          * [MAX_SIZE+4k][MAX_SIZE+4k]
1697          *
1698          * Each range on their own accounts for 2 extents, but merged together
1699          * they are only 3 extents worth of accounting, so we need to drop in
1700          * this case.
1701          */
1702         old_size = other->end - other->start + 1;
1703         num_extents = count_max_extents(old_size);
1704         old_size = new->end - new->start + 1;
1705         num_extents += count_max_extents(old_size);
1706         if (count_max_extents(new_size) >= num_extents)
1707                 return;
1708
1709         spin_lock(&BTRFS_I(inode)->lock);
1710         btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1711         spin_unlock(&BTRFS_I(inode)->lock);
1712 }
1713
1714 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1715                                       struct inode *inode)
1716 {
1717         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1718
1719         spin_lock(&root->delalloc_lock);
1720         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1721                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1722                               &root->delalloc_inodes);
1723                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1724                         &BTRFS_I(inode)->runtime_flags);
1725                 root->nr_delalloc_inodes++;
1726                 if (root->nr_delalloc_inodes == 1) {
1727                         spin_lock(&fs_info->delalloc_root_lock);
1728                         BUG_ON(!list_empty(&root->delalloc_root));
1729                         list_add_tail(&root->delalloc_root,
1730                                       &fs_info->delalloc_roots);
1731                         spin_unlock(&fs_info->delalloc_root_lock);
1732                 }
1733         }
1734         spin_unlock(&root->delalloc_lock);
1735 }
1736
1737 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1738                                      struct btrfs_inode *inode)
1739 {
1740         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1741
1742         spin_lock(&root->delalloc_lock);
1743         if (!list_empty(&inode->delalloc_inodes)) {
1744                 list_del_init(&inode->delalloc_inodes);
1745                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1746                           &inode->runtime_flags);
1747                 root->nr_delalloc_inodes--;
1748                 if (!root->nr_delalloc_inodes) {
1749                         spin_lock(&fs_info->delalloc_root_lock);
1750                         BUG_ON(list_empty(&root->delalloc_root));
1751                         list_del_init(&root->delalloc_root);
1752                         spin_unlock(&fs_info->delalloc_root_lock);
1753                 }
1754         }
1755         spin_unlock(&root->delalloc_lock);
1756 }
1757
1758 /*
1759  * extent_io.c set_bit_hook, used to track delayed allocation
1760  * bytes in this file, and to maintain the list of inodes that
1761  * have pending delalloc work to be done.
1762  */
1763 static void btrfs_set_bit_hook(void *private_data,
1764                                struct extent_state *state, unsigned *bits)
1765 {
1766         struct inode *inode = private_data;
1767
1768         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1769
1770         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
1771                 WARN_ON(1);
1772         /*
1773          * set_bit and clear bit hooks normally require _irqsave/restore
1774          * but in this case, we are only testing for the DELALLOC
1775          * bit, which is only set or cleared with irqs on
1776          */
1777         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1778                 struct btrfs_root *root = BTRFS_I(inode)->root;
1779                 u64 len = state->end + 1 - state->start;
1780                 u32 num_extents = count_max_extents(len);
1781                 bool do_list = !btrfs_is_free_space_inode(BTRFS_I(inode));
1782
1783                 spin_lock(&BTRFS_I(inode)->lock);
1784                 btrfs_mod_outstanding_extents(BTRFS_I(inode), num_extents);
1785                 spin_unlock(&BTRFS_I(inode)->lock);
1786
1787                 /* For sanity tests */
1788                 if (btrfs_is_testing(fs_info))
1789                         return;
1790
1791                 percpu_counter_add_batch(&fs_info->delalloc_bytes, len,
1792                                          fs_info->delalloc_batch);
1793                 spin_lock(&BTRFS_I(inode)->lock);
1794                 BTRFS_I(inode)->delalloc_bytes += len;
1795                 if (*bits & EXTENT_DEFRAG)
1796                         BTRFS_I(inode)->defrag_bytes += len;
1797                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1798                                          &BTRFS_I(inode)->runtime_flags))
1799                         btrfs_add_delalloc_inodes(root, inode);
1800                 spin_unlock(&BTRFS_I(inode)->lock);
1801         }
1802
1803         if (!(state->state & EXTENT_DELALLOC_NEW) &&
1804             (*bits & EXTENT_DELALLOC_NEW)) {
1805                 spin_lock(&BTRFS_I(inode)->lock);
1806                 BTRFS_I(inode)->new_delalloc_bytes += state->end + 1 -
1807                         state->start;
1808                 spin_unlock(&BTRFS_I(inode)->lock);
1809         }
1810 }
1811
1812 /*
1813  * extent_io.c clear_bit_hook, see set_bit_hook for why
1814  */
1815 static void btrfs_clear_bit_hook(void *private_data,
1816                                  struct extent_state *state,
1817                                  unsigned *bits)
1818 {
1819         struct btrfs_inode *inode = BTRFS_I((struct inode *)private_data);
1820         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1821         u64 len = state->end + 1 - state->start;
1822         u32 num_extents = count_max_extents(len);
1823
1824         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG)) {
1825                 spin_lock(&inode->lock);
1826                 inode->defrag_bytes -= len;
1827                 spin_unlock(&inode->lock);
1828         }
1829
1830         /*
1831          * set_bit and clear bit hooks normally require _irqsave/restore
1832          * but in this case, we are only testing for the DELALLOC
1833          * bit, which is only set or cleared with irqs on
1834          */
1835         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1836                 struct btrfs_root *root = inode->root;
1837                 bool do_list = !btrfs_is_free_space_inode(inode);
1838
1839                 spin_lock(&inode->lock);
1840                 btrfs_mod_outstanding_extents(inode, -num_extents);
1841                 spin_unlock(&inode->lock);
1842
1843                 /*
1844                  * We don't reserve metadata space for space cache inodes so we
1845                  * don't need to call dellalloc_release_metadata if there is an
1846                  * error.
1847                  */
1848                 if (*bits & EXTENT_CLEAR_META_RESV &&
1849                     root != fs_info->tree_root)
1850                         btrfs_delalloc_release_metadata(inode, len);
1851
1852                 /* For sanity tests. */
1853                 if (btrfs_is_testing(fs_info))
1854                         return;
1855
1856                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID &&
1857                     do_list && !(state->state & EXTENT_NORESERVE) &&
1858                     (*bits & EXTENT_CLEAR_DATA_RESV))
1859                         btrfs_free_reserved_data_space_noquota(
1860                                         &inode->vfs_inode,
1861                                         state->start, len);
1862
1863                 percpu_counter_add_batch(&fs_info->delalloc_bytes, -len,
1864                                          fs_info->delalloc_batch);
1865                 spin_lock(&inode->lock);
1866                 inode->delalloc_bytes -= len;
1867                 if (do_list && inode->delalloc_bytes == 0 &&
1868                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1869                                         &inode->runtime_flags))
1870                         btrfs_del_delalloc_inode(root, inode);
1871                 spin_unlock(&inode->lock);
1872         }
1873
1874         if ((state->state & EXTENT_DELALLOC_NEW) &&
1875             (*bits & EXTENT_DELALLOC_NEW)) {
1876                 spin_lock(&inode->lock);
1877                 ASSERT(inode->new_delalloc_bytes >= len);
1878                 inode->new_delalloc_bytes -= len;
1879                 spin_unlock(&inode->lock);
1880         }
1881 }
1882
1883 /*
1884  * extent_io.c merge_bio_hook, this must check the chunk tree to make sure
1885  * we don't create bios that span stripes or chunks
1886  *
1887  * return 1 if page cannot be merged to bio
1888  * return 0 if page can be merged to bio
1889  * return error otherwise
1890  */
1891 int btrfs_merge_bio_hook(struct page *page, unsigned long offset,
1892                          size_t size, struct bio *bio,
1893                          unsigned long bio_flags)
1894 {
1895         struct inode *inode = page->mapping->host;
1896         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1897         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
1898         u64 length = 0;
1899         u64 map_length;
1900         int ret;
1901
1902         if (bio_flags & EXTENT_BIO_COMPRESSED)
1903                 return 0;
1904
1905         length = bio->bi_iter.bi_size;
1906         map_length = length;
1907         ret = btrfs_map_block(fs_info, btrfs_op(bio), logical, &map_length,
1908                               NULL, 0);
1909         if (ret < 0)
1910                 return ret;
1911         if (map_length < length + size)
1912                 return 1;
1913         return 0;
1914 }
1915
1916 /*
1917  * in order to insert checksums into the metadata in large chunks,
1918  * we wait until bio submission time.   All the pages in the bio are
1919  * checksummed and sums are attached onto the ordered extent record.
1920  *
1921  * At IO completion time the cums attached on the ordered extent record
1922  * are inserted into the btree
1923  */
1924 static blk_status_t __btrfs_submit_bio_start(void *private_data, struct bio *bio,
1925                                     int mirror_num, unsigned long bio_flags,
1926                                     u64 bio_offset)
1927 {
1928         struct inode *inode = private_data;
1929         blk_status_t ret = 0;
1930
1931         ret = btrfs_csum_one_bio(inode, bio, 0, 0);
1932         BUG_ON(ret); /* -ENOMEM */
1933         return 0;
1934 }
1935
1936 /*
1937  * in order to insert checksums into the metadata in large chunks,
1938  * we wait until bio submission time.   All the pages in the bio are
1939  * checksummed and sums are attached onto the ordered extent record.
1940  *
1941  * At IO completion time the cums attached on the ordered extent record
1942  * are inserted into the btree
1943  */
1944 static blk_status_t __btrfs_submit_bio_done(void *private_data, struct bio *bio,
1945                           int mirror_num, unsigned long bio_flags,
1946                           u64 bio_offset)
1947 {
1948         struct inode *inode = private_data;
1949         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1950         blk_status_t ret;
1951
1952         ret = btrfs_map_bio(fs_info, bio, mirror_num, 1);
1953         if (ret) {
1954                 bio->bi_status = ret;
1955                 bio_endio(bio);
1956         }
1957         return ret;
1958 }
1959
1960 /*
1961  * extent_io.c submission hook. This does the right thing for csum calculation
1962  * on write, or reading the csums from the tree before a read.
1963  *
1964  * Rules about async/sync submit,
1965  * a) read:                             sync submit
1966  *
1967  * b) write without checksum:           sync submit
1968  *
1969  * c) write with checksum:
1970  *    c-1) if bio is issued by fsync:   sync submit
1971  *         (sync_writers != 0)
1972  *
1973  *    c-2) if root is reloc root:       sync submit
1974  *         (only in case of buffered IO)
1975  *
1976  *    c-3) otherwise:                   async submit
1977  */
1978 static blk_status_t btrfs_submit_bio_hook(void *private_data, struct bio *bio,
1979                                  int mirror_num, unsigned long bio_flags,
1980                                  u64 bio_offset)
1981 {
1982         struct inode *inode = private_data;
1983         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1984         struct btrfs_root *root = BTRFS_I(inode)->root;
1985         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
1986         blk_status_t ret = 0;
1987         int skip_sum;
1988         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
1989
1990         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
1991
1992         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
1993                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
1994
1995         if (bio_op(bio) != REQ_OP_WRITE) {
1996                 ret = btrfs_bio_wq_end_io(fs_info, bio, metadata);
1997                 if (ret)
1998                         goto out;
1999
2000                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
2001                         ret = btrfs_submit_compressed_read(inode, bio,
2002                                                            mirror_num,
2003                                                            bio_flags);
2004                         goto out;
2005                 } else if (!skip_sum) {
2006                         ret = btrfs_lookup_bio_sums(inode, bio, NULL);
2007                         if (ret)
2008                                 goto out;
2009                 }
2010                 goto mapit;
2011         } else if (async && !skip_sum) {
2012                 /* csum items have already been cloned */
2013                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2014                         goto mapit;
2015                 /* we're doing a write, do the async checksumming */
2016                 ret = btrfs_wq_submit_bio(fs_info, bio, mirror_num, bio_flags,
2017                                           bio_offset, inode,
2018                                           __btrfs_submit_bio_start,
2019                                           __btrfs_submit_bio_done);
2020                 goto out;
2021         } else if (!skip_sum) {
2022                 ret = btrfs_csum_one_bio(inode, bio, 0, 0);
2023                 if (ret)
2024                         goto out;
2025         }
2026
2027 mapit:
2028         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
2029
2030 out:
2031         if (ret) {
2032                 bio->bi_status = ret;
2033                 bio_endio(bio);
2034         }
2035         return ret;
2036 }
2037
2038 /*
2039  * given a list of ordered sums record them in the inode.  This happens
2040  * at IO completion time based on sums calculated at bio submission time.
2041  */
2042 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
2043                              struct inode *inode, struct list_head *list)
2044 {
2045         struct btrfs_ordered_sum *sum;
2046
2047         list_for_each_entry(sum, list, list) {
2048                 trans->adding_csums = true;
2049                 btrfs_csum_file_blocks(trans,
2050                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
2051                 trans->adding_csums = false;
2052         }
2053         return 0;
2054 }
2055
2056 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
2057                               unsigned int extra_bits,
2058                               struct extent_state **cached_state, int dedupe)
2059 {
2060         WARN_ON((end & (PAGE_SIZE - 1)) == 0);
2061         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
2062                                    extra_bits, cached_state);
2063 }
2064
2065 /* see btrfs_writepage_start_hook for details on why this is required */
2066 struct btrfs_writepage_fixup {
2067         struct page *page;
2068         struct btrfs_work work;
2069 };
2070
2071 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
2072 {
2073         struct btrfs_writepage_fixup *fixup;
2074         struct btrfs_ordered_extent *ordered;
2075         struct extent_state *cached_state = NULL;
2076         struct extent_changeset *data_reserved = NULL;
2077         struct page *page;
2078         struct inode *inode;
2079         u64 page_start;
2080         u64 page_end;
2081         int ret;
2082
2083         fixup = container_of(work, struct btrfs_writepage_fixup, work);
2084         page = fixup->page;
2085 again:
2086         lock_page(page);
2087         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
2088                 ClearPageChecked(page);
2089                 goto out_page;
2090         }
2091
2092         inode = page->mapping->host;
2093         page_start = page_offset(page);
2094         page_end = page_offset(page) + PAGE_SIZE - 1;
2095
2096         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end,
2097                          &cached_state);
2098
2099         /* already ordered? We're done */
2100         if (PagePrivate2(page))
2101                 goto out;
2102
2103         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
2104                                         PAGE_SIZE);
2105         if (ordered) {
2106                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2107                                      page_end, &cached_state);
2108                 unlock_page(page);
2109                 btrfs_start_ordered_extent(inode, ordered, 1);
2110                 btrfs_put_ordered_extent(ordered);
2111                 goto again;
2112         }
2113
2114         ret = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
2115                                            PAGE_SIZE);
2116         if (ret) {
2117                 mapping_set_error(page->mapping, ret);
2118                 end_extent_writepage(page, ret, page_start, page_end);
2119                 ClearPageChecked(page);
2120                 goto out;
2121          }
2122
2123         ret = btrfs_set_extent_delalloc(inode, page_start, page_end, 0,
2124                                         &cached_state, 0);
2125         if (ret) {
2126                 mapping_set_error(page->mapping, ret);
2127                 end_extent_writepage(page, ret, page_start, page_end);
2128                 ClearPageChecked(page);
2129                 goto out;
2130         }
2131
2132         ClearPageChecked(page);
2133         set_page_dirty(page);
2134         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
2135 out:
2136         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2137                              &cached_state);
2138 out_page:
2139         unlock_page(page);
2140         put_page(page);
2141         kfree(fixup);
2142         extent_changeset_free(data_reserved);
2143 }
2144
2145 /*
2146  * There are a few paths in the higher layers of the kernel that directly
2147  * set the page dirty bit without asking the filesystem if it is a
2148  * good idea.  This causes problems because we want to make sure COW
2149  * properly happens and the data=ordered rules are followed.
2150  *
2151  * In our case any range that doesn't have the ORDERED bit set
2152  * hasn't been properly setup for IO.  We kick off an async process
2153  * to fix it up.  The async helper will wait for ordered extents, set
2154  * the delalloc bit and make it safe to write the page.
2155  */
2156 static int btrfs_writepage_start_hook(struct page *page, u64 start, u64 end)
2157 {
2158         struct inode *inode = page->mapping->host;
2159         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2160         struct btrfs_writepage_fixup *fixup;
2161
2162         /* this page is properly in the ordered list */
2163         if (TestClearPagePrivate2(page))
2164                 return 0;
2165
2166         if (PageChecked(page))
2167                 return -EAGAIN;
2168
2169         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2170         if (!fixup)
2171                 return -EAGAIN;
2172
2173         SetPageChecked(page);
2174         get_page(page);
2175         btrfs_init_work(&fixup->work, btrfs_fixup_helper,
2176                         btrfs_writepage_fixup_worker, NULL, NULL);
2177         fixup->page = page;
2178         btrfs_queue_work(fs_info->fixup_workers, &fixup->work);
2179         return -EBUSY;
2180 }
2181
2182 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2183                                        struct inode *inode, u64 file_pos,
2184                                        u64 disk_bytenr, u64 disk_num_bytes,
2185                                        u64 num_bytes, u64 ram_bytes,
2186                                        u8 compression, u8 encryption,
2187                                        u16 other_encoding, int extent_type)
2188 {
2189         struct btrfs_root *root = BTRFS_I(inode)->root;
2190         struct btrfs_file_extent_item *fi;
2191         struct btrfs_path *path;
2192         struct extent_buffer *leaf;
2193         struct btrfs_key ins;
2194         u64 qg_released;
2195         int extent_inserted = 0;
2196         int ret;
2197
2198         path = btrfs_alloc_path();
2199         if (!path)
2200                 return -ENOMEM;
2201
2202         /*
2203          * we may be replacing one extent in the tree with another.
2204          * The new extent is pinned in the extent map, and we don't want
2205          * to drop it from the cache until it is completely in the btree.
2206          *
2207          * So, tell btrfs_drop_extents to leave this extent in the cache.
2208          * the caller is expected to unpin it and allow it to be merged
2209          * with the others.
2210          */
2211         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2212                                    file_pos + num_bytes, NULL, 0,
2213                                    1, sizeof(*fi), &extent_inserted);
2214         if (ret)
2215                 goto out;
2216
2217         if (!extent_inserted) {
2218                 ins.objectid = btrfs_ino(BTRFS_I(inode));
2219                 ins.offset = file_pos;
2220                 ins.type = BTRFS_EXTENT_DATA_KEY;
2221
2222                 path->leave_spinning = 1;
2223                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2224                                               sizeof(*fi));
2225                 if (ret)
2226                         goto out;
2227         }
2228         leaf = path->nodes[0];
2229         fi = btrfs_item_ptr(leaf, path->slots[0],
2230                             struct btrfs_file_extent_item);
2231         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2232         btrfs_set_file_extent_type(leaf, fi, extent_type);
2233         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2234         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2235         btrfs_set_file_extent_offset(leaf, fi, 0);
2236         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2237         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2238         btrfs_set_file_extent_compression(leaf, fi, compression);
2239         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2240         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2241
2242         btrfs_mark_buffer_dirty(leaf);
2243         btrfs_release_path(path);
2244
2245         inode_add_bytes(inode, num_bytes);
2246
2247         ins.objectid = disk_bytenr;
2248         ins.offset = disk_num_bytes;
2249         ins.type = BTRFS_EXTENT_ITEM_KEY;
2250
2251         /*
2252          * Release the reserved range from inode dirty range map, as it is
2253          * already moved into delayed_ref_head
2254          */
2255         ret = btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2256         if (ret < 0)
2257                 goto out;
2258         qg_released = ret;
2259         ret = btrfs_alloc_reserved_file_extent(trans, root,
2260                                                btrfs_ino(BTRFS_I(inode)),
2261                                                file_pos, qg_released, &ins);
2262 out:
2263         btrfs_free_path(path);
2264
2265         return ret;
2266 }
2267
2268 /* snapshot-aware defrag */
2269 struct sa_defrag_extent_backref {
2270         struct rb_node node;
2271         struct old_sa_defrag_extent *old;
2272         u64 root_id;
2273         u64 inum;
2274         u64 file_pos;
2275         u64 extent_offset;
2276         u64 num_bytes;
2277         u64 generation;
2278 };
2279
2280 struct old_sa_defrag_extent {
2281         struct list_head list;
2282         struct new_sa_defrag_extent *new;
2283
2284         u64 extent_offset;
2285         u64 bytenr;
2286         u64 offset;
2287         u64 len;
2288         int count;
2289 };
2290
2291 struct new_sa_defrag_extent {
2292         struct rb_root root;
2293         struct list_head head;
2294         struct btrfs_path *path;
2295         struct inode *inode;
2296         u64 file_pos;
2297         u64 len;
2298         u64 bytenr;
2299         u64 disk_len;
2300         u8 compress_type;
2301 };
2302
2303 static int backref_comp(struct sa_defrag_extent_backref *b1,
2304                         struct sa_defrag_extent_backref *b2)
2305 {
2306         if (b1->root_id < b2->root_id)
2307                 return -1;
2308         else if (b1->root_id > b2->root_id)
2309                 return 1;
2310
2311         if (b1->inum < b2->inum)
2312                 return -1;
2313         else if (b1->inum > b2->inum)
2314                 return 1;
2315
2316         if (b1->file_pos < b2->file_pos)
2317                 return -1;
2318         else if (b1->file_pos > b2->file_pos)
2319                 return 1;
2320
2321         /*
2322          * [------------------------------] ===> (a range of space)
2323          *     |<--->|   |<---->| =============> (fs/file tree A)
2324          * |<---------------------------->| ===> (fs/file tree B)
2325          *
2326          * A range of space can refer to two file extents in one tree while
2327          * refer to only one file extent in another tree.
2328          *
2329          * So we may process a disk offset more than one time(two extents in A)
2330          * and locate at the same extent(one extent in B), then insert two same
2331          * backrefs(both refer to the extent in B).
2332          */
2333         return 0;
2334 }
2335
2336 static void backref_insert(struct rb_root *root,
2337                            struct sa_defrag_extent_backref *backref)
2338 {
2339         struct rb_node **p = &root->rb_node;
2340         struct rb_node *parent = NULL;
2341         struct sa_defrag_extent_backref *entry;
2342         int ret;
2343
2344         while (*p) {
2345                 parent = *p;
2346                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
2347
2348                 ret = backref_comp(backref, entry);
2349                 if (ret < 0)
2350                         p = &(*p)->rb_left;
2351                 else
2352                         p = &(*p)->rb_right;
2353         }
2354
2355         rb_link_node(&backref->node, parent, p);
2356         rb_insert_color(&backref->node, root);
2357 }
2358
2359 /*
2360  * Note the backref might has changed, and in this case we just return 0.
2361  */
2362 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
2363                                        void *ctx)
2364 {
2365         struct btrfs_file_extent_item *extent;
2366         struct old_sa_defrag_extent *old = ctx;
2367         struct new_sa_defrag_extent *new = old->new;
2368         struct btrfs_path *path = new->path;
2369         struct btrfs_key key;
2370         struct btrfs_root *root;
2371         struct sa_defrag_extent_backref *backref;
2372         struct extent_buffer *leaf;
2373         struct inode *inode = new->inode;
2374         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2375         int slot;
2376         int ret;
2377         u64 extent_offset;
2378         u64 num_bytes;
2379
2380         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2381             inum == btrfs_ino(BTRFS_I(inode)))
2382                 return 0;
2383
2384         key.objectid = root_id;
2385         key.type = BTRFS_ROOT_ITEM_KEY;
2386         key.offset = (u64)-1;
2387
2388         root = btrfs_read_fs_root_no_name(fs_info, &key);
2389         if (IS_ERR(root)) {
2390                 if (PTR_ERR(root) == -ENOENT)
2391                         return 0;
2392                 WARN_ON(1);
2393                 btrfs_debug(fs_info, "inum=%llu, offset=%llu, root_id=%llu",
2394                          inum, offset, root_id);
2395                 return PTR_ERR(root);
2396         }
2397
2398         key.objectid = inum;
2399         key.type = BTRFS_EXTENT_DATA_KEY;
2400         if (offset > (u64)-1 << 32)
2401                 key.offset = 0;
2402         else
2403                 key.offset = offset;
2404
2405         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2406         if (WARN_ON(ret < 0))
2407                 return ret;
2408         ret = 0;
2409
2410         while (1) {
2411                 cond_resched();
2412
2413                 leaf = path->nodes[0];
2414                 slot = path->slots[0];
2415
2416                 if (slot >= btrfs_header_nritems(leaf)) {
2417                         ret = btrfs_next_leaf(root, path);
2418                         if (ret < 0) {
2419                                 goto out;
2420                         } else if (ret > 0) {
2421                                 ret = 0;
2422                                 goto out;
2423                         }
2424                         continue;
2425                 }
2426
2427                 path->slots[0]++;
2428
2429                 btrfs_item_key_to_cpu(leaf, &key, slot);
2430
2431                 if (key.objectid > inum)
2432                         goto out;
2433
2434                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2435                         continue;
2436
2437                 extent = btrfs_item_ptr(leaf, slot,
2438                                         struct btrfs_file_extent_item);
2439
2440                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2441                         continue;
2442
2443                 /*
2444                  * 'offset' refers to the exact key.offset,
2445                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2446                  * (key.offset - extent_offset).
2447                  */
2448                 if (key.offset != offset)
2449                         continue;
2450
2451                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2452                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2453
2454                 if (extent_offset >= old->extent_offset + old->offset +
2455                     old->len || extent_offset + num_bytes <=
2456                     old->extent_offset + old->offset)
2457                         continue;
2458                 break;
2459         }
2460
2461         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2462         if (!backref) {
2463                 ret = -ENOENT;
2464                 goto out;
2465         }
2466
2467         backref->root_id = root_id;
2468         backref->inum = inum;
2469         backref->file_pos = offset;
2470         backref->num_bytes = num_bytes;
2471         backref->extent_offset = extent_offset;
2472         backref->generation = btrfs_file_extent_generation(leaf, extent);
2473         backref->old = old;
2474         backref_insert(&new->root, backref);
2475         old->count++;
2476 out:
2477         btrfs_release_path(path);
2478         WARN_ON(ret);
2479         return ret;
2480 }
2481
2482 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2483                                    struct new_sa_defrag_extent *new)
2484 {
2485         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2486         struct old_sa_defrag_extent *old, *tmp;
2487         int ret;
2488
2489         new->path = path;
2490
2491         list_for_each_entry_safe(old, tmp, &new->head, list) {
2492                 ret = iterate_inodes_from_logical(old->bytenr +
2493                                                   old->extent_offset, fs_info,
2494                                                   path, record_one_backref,
2495                                                   old, false);
2496                 if (ret < 0 && ret != -ENOENT)
2497                         return false;
2498
2499                 /* no backref to be processed for this extent */
2500                 if (!old->count) {
2501                         list_del(&old->list);
2502                         kfree(old);
2503                 }
2504         }
2505
2506         if (list_empty(&new->head))
2507                 return false;
2508
2509         return true;
2510 }
2511
2512 static int relink_is_mergable(struct extent_buffer *leaf,
2513                               struct btrfs_file_extent_item *fi,
2514                               struct new_sa_defrag_extent *new)
2515 {
2516         if (btrfs_file_extent_disk_bytenr(leaf, fi) != new->bytenr)
2517                 return 0;
2518
2519         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2520                 return 0;
2521
2522         if (btrfs_file_extent_compression(leaf, fi) != new->compress_type)
2523                 return 0;
2524
2525         if (btrfs_file_extent_encryption(leaf, fi) ||
2526             btrfs_file_extent_other_encoding(leaf, fi))
2527                 return 0;
2528
2529         return 1;
2530 }
2531
2532 /*
2533  * Note the backref might has changed, and in this case we just return 0.
2534  */
2535 static noinline int relink_extent_backref(struct btrfs_path *path,
2536                                  struct sa_defrag_extent_backref *prev,
2537                                  struct sa_defrag_extent_backref *backref)
2538 {
2539         struct btrfs_file_extent_item *extent;
2540         struct btrfs_file_extent_item *item;
2541         struct btrfs_ordered_extent *ordered;
2542         struct btrfs_trans_handle *trans;
2543         struct btrfs_root *root;
2544         struct btrfs_key key;
2545         struct extent_buffer *leaf;
2546         struct old_sa_defrag_extent *old = backref->old;
2547         struct new_sa_defrag_extent *new = old->new;
2548         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2549         struct inode *inode;
2550         struct extent_state *cached = NULL;
2551         int ret = 0;
2552         u64 start;
2553         u64 len;
2554         u64 lock_start;
2555         u64 lock_end;
2556         bool merge = false;
2557         int index;
2558
2559         if (prev && prev->root_id == backref->root_id &&
2560             prev->inum == backref->inum &&
2561             prev->file_pos + prev->num_bytes == backref->file_pos)
2562                 merge = true;
2563
2564         /* step 1: get root */
2565         key.objectid = backref->root_id;
2566         key.type = BTRFS_ROOT_ITEM_KEY;
2567         key.offset = (u64)-1;
2568
2569         index = srcu_read_lock(&fs_info->subvol_srcu);
2570
2571         root = btrfs_read_fs_root_no_name(fs_info, &key);
2572         if (IS_ERR(root)) {
2573                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2574                 if (PTR_ERR(root) == -ENOENT)
2575                         return 0;
2576                 return PTR_ERR(root);
2577         }
2578
2579         if (btrfs_root_readonly(root)) {
2580                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2581                 return 0;
2582         }
2583
2584         /* step 2: get inode */
2585         key.objectid = backref->inum;
2586         key.type = BTRFS_INODE_ITEM_KEY;
2587         key.offset = 0;
2588
2589         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
2590         if (IS_ERR(inode)) {
2591                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2592                 return 0;
2593         }
2594
2595         srcu_read_unlock(&fs_info->subvol_srcu, index);
2596
2597         /* step 3: relink backref */
2598         lock_start = backref->file_pos;
2599         lock_end = backref->file_pos + backref->num_bytes - 1;
2600         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2601                          &cached);
2602
2603         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2604         if (ordered) {
2605                 btrfs_put_ordered_extent(ordered);
2606                 goto out_unlock;
2607         }
2608
2609         trans = btrfs_join_transaction(root);
2610         if (IS_ERR(trans)) {
2611                 ret = PTR_ERR(trans);
2612                 goto out_unlock;
2613         }
2614
2615         key.objectid = backref->inum;
2616         key.type = BTRFS_EXTENT_DATA_KEY;
2617         key.offset = backref->file_pos;
2618
2619         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2620         if (ret < 0) {
2621                 goto out_free_path;
2622         } else if (ret > 0) {
2623                 ret = 0;
2624                 goto out_free_path;
2625         }
2626
2627         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2628                                 struct btrfs_file_extent_item);
2629
2630         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2631             backref->generation)
2632                 goto out_free_path;
2633
2634         btrfs_release_path(path);
2635
2636         start = backref->file_pos;
2637         if (backref->extent_offset < old->extent_offset + old->offset)
2638                 start += old->extent_offset + old->offset -
2639                          backref->extent_offset;
2640
2641         len = min(backref->extent_offset + backref->num_bytes,
2642                   old->extent_offset + old->offset + old->len);
2643         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2644
2645         ret = btrfs_drop_extents(trans, root, inode, start,
2646                                  start + len, 1);
2647         if (ret)
2648                 goto out_free_path;
2649 again:
2650         key.objectid = btrfs_ino(BTRFS_I(inode));
2651         key.type = BTRFS_EXTENT_DATA_KEY;
2652         key.offset = start;
2653
2654         path->leave_spinning = 1;
2655         if (merge) {
2656                 struct btrfs_file_extent_item *fi;
2657                 u64 extent_len;
2658                 struct btrfs_key found_key;
2659
2660                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2661                 if (ret < 0)
2662                         goto out_free_path;
2663
2664                 path->slots[0]--;
2665                 leaf = path->nodes[0];
2666                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2667
2668                 fi = btrfs_item_ptr(leaf, path->slots[0],
2669                                     struct btrfs_file_extent_item);
2670                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2671
2672                 if (extent_len + found_key.offset == start &&
2673                     relink_is_mergable(leaf, fi, new)) {
2674                         btrfs_set_file_extent_num_bytes(leaf, fi,
2675                                                         extent_len + len);
2676                         btrfs_mark_buffer_dirty(leaf);
2677                         inode_add_bytes(inode, len);
2678
2679                         ret = 1;
2680                         goto out_free_path;
2681                 } else {
2682                         merge = false;
2683                         btrfs_release_path(path);
2684                         goto again;
2685                 }
2686         }
2687
2688         ret = btrfs_insert_empty_item(trans, root, path, &key,
2689                                         sizeof(*extent));
2690         if (ret) {
2691                 btrfs_abort_transaction(trans, ret);
2692                 goto out_free_path;
2693         }
2694
2695         leaf = path->nodes[0];
2696         item = btrfs_item_ptr(leaf, path->slots[0],
2697                                 struct btrfs_file_extent_item);
2698         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2699         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2700         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2701         btrfs_set_file_extent_num_bytes(leaf, item, len);
2702         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2703         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2704         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2705         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2706         btrfs_set_file_extent_encryption(leaf, item, 0);
2707         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2708
2709         btrfs_mark_buffer_dirty(leaf);
2710         inode_add_bytes(inode, len);
2711         btrfs_release_path(path);
2712
2713         ret = btrfs_inc_extent_ref(trans, root, new->bytenr,
2714                         new->disk_len, 0,
2715                         backref->root_id, backref->inum,
2716                         new->file_pos); /* start - extent_offset */
2717         if (ret) {
2718                 btrfs_abort_transaction(trans, ret);
2719                 goto out_free_path;
2720         }
2721
2722         ret = 1;
2723 out_free_path:
2724         btrfs_release_path(path);
2725         path->leave_spinning = 0;
2726         btrfs_end_transaction(trans);
2727 out_unlock:
2728         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2729                              &cached);
2730         iput(inode);
2731         return ret;
2732 }
2733
2734 static void free_sa_defrag_extent(struct new_sa_defrag_extent *new)
2735 {
2736         struct old_sa_defrag_extent *old, *tmp;
2737
2738         if (!new)
2739                 return;
2740
2741         list_for_each_entry_safe(old, tmp, &new->head, list) {
2742                 kfree(old);
2743         }
2744         kfree(new);
2745 }
2746
2747 static void relink_file_extents(struct new_sa_defrag_extent *new)
2748 {
2749         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2750         struct btrfs_path *path;
2751         struct sa_defrag_extent_backref *backref;
2752         struct sa_defrag_extent_backref *prev = NULL;
2753         struct inode *inode;
2754         struct btrfs_root *root;
2755         struct rb_node *node;
2756         int ret;
2757
2758         inode = new->inode;
2759         root = BTRFS_I(inode)->root;
2760
2761         path = btrfs_alloc_path();
2762         if (!path)
2763                 return;
2764
2765         if (!record_extent_backrefs(path, new)) {
2766                 btrfs_free_path(path);
2767                 goto out;
2768         }
2769         btrfs_release_path(path);
2770
2771         while (1) {
2772                 node = rb_first(&new->root);
2773                 if (!node)
2774                         break;
2775                 rb_erase(node, &new->root);
2776
2777                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2778
2779                 ret = relink_extent_backref(path, prev, backref);
2780                 WARN_ON(ret < 0);
2781
2782                 kfree(prev);
2783
2784                 if (ret == 1)
2785                         prev = backref;
2786                 else
2787                         prev = NULL;
2788                 cond_resched();
2789         }
2790         kfree(prev);
2791
2792         btrfs_free_path(path);
2793 out:
2794         free_sa_defrag_extent(new);
2795
2796         atomic_dec(&fs_info->defrag_running);
2797         wake_up(&fs_info->transaction_wait);
2798 }
2799
2800 static struct new_sa_defrag_extent *
2801 record_old_file_extents(struct inode *inode,
2802                         struct btrfs_ordered_extent *ordered)
2803 {
2804         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2805         struct btrfs_root *root = BTRFS_I(inode)->root;
2806         struct btrfs_path *path;
2807         struct btrfs_key key;
2808         struct old_sa_defrag_extent *old;
2809         struct new_sa_defrag_extent *new;
2810         int ret;
2811
2812         new = kmalloc(sizeof(*new), GFP_NOFS);
2813         if (!new)
2814                 return NULL;
2815
2816         new->inode = inode;
2817         new->file_pos = ordered->file_offset;
2818         new->len = ordered->len;
2819         new->bytenr = ordered->start;
2820         new->disk_len = ordered->disk_len;
2821         new->compress_type = ordered->compress_type;
2822         new->root = RB_ROOT;
2823         INIT_LIST_HEAD(&new->head);
2824
2825         path = btrfs_alloc_path();
2826         if (!path)
2827                 goto out_kfree;
2828
2829         key.objectid = btrfs_ino(BTRFS_I(inode));
2830         key.type = BTRFS_EXTENT_DATA_KEY;
2831         key.offset = new->file_pos;
2832
2833         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2834         if (ret < 0)
2835                 goto out_free_path;
2836         if (ret > 0 && path->slots[0] > 0)
2837                 path->slots[0]--;
2838
2839         /* find out all the old extents for the file range */
2840         while (1) {
2841                 struct btrfs_file_extent_item *extent;
2842                 struct extent_buffer *l;
2843                 int slot;
2844                 u64 num_bytes;
2845                 u64 offset;
2846                 u64 end;
2847                 u64 disk_bytenr;
2848                 u64 extent_offset;
2849
2850                 l = path->nodes[0];
2851                 slot = path->slots[0];
2852
2853                 if (slot >= btrfs_header_nritems(l)) {
2854                         ret = btrfs_next_leaf(root, path);
2855                         if (ret < 0)
2856                                 goto out_free_path;
2857                         else if (ret > 0)
2858                                 break;
2859                         continue;
2860                 }
2861
2862                 btrfs_item_key_to_cpu(l, &key, slot);
2863
2864                 if (key.objectid != btrfs_ino(BTRFS_I(inode)))
2865                         break;
2866                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2867                         break;
2868                 if (key.offset >= new->file_pos + new->len)
2869                         break;
2870
2871                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2872
2873                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
2874                 if (key.offset + num_bytes < new->file_pos)
2875                         goto next;
2876
2877                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
2878                 if (!disk_bytenr)
2879                         goto next;
2880
2881                 extent_offset = btrfs_file_extent_offset(l, extent);
2882
2883                 old = kmalloc(sizeof(*old), GFP_NOFS);
2884                 if (!old)
2885                         goto out_free_path;
2886
2887                 offset = max(new->file_pos, key.offset);
2888                 end = min(new->file_pos + new->len, key.offset + num_bytes);
2889
2890                 old->bytenr = disk_bytenr;
2891                 old->extent_offset = extent_offset;
2892                 old->offset = offset - key.offset;
2893                 old->len = end - offset;
2894                 old->new = new;
2895                 old->count = 0;
2896                 list_add_tail(&old->list, &new->head);
2897 next:
2898                 path->slots[0]++;
2899                 cond_resched();
2900         }
2901
2902         btrfs_free_path(path);
2903         atomic_inc(&fs_info->defrag_running);
2904
2905         return new;
2906
2907 out_free_path:
2908         btrfs_free_path(path);
2909 out_kfree:
2910         free_sa_defrag_extent(new);
2911         return NULL;
2912 }
2913
2914 static void btrfs_release_delalloc_bytes(struct btrfs_fs_info *fs_info,
2915                                          u64 start, u64 len)
2916 {
2917         struct btrfs_block_group_cache *cache;
2918
2919         cache = btrfs_lookup_block_group(fs_info, start);
2920         ASSERT(cache);
2921
2922         spin_lock(&cache->lock);
2923         cache->delalloc_bytes -= len;
2924         spin_unlock(&cache->lock);
2925
2926         btrfs_put_block_group(cache);
2927 }
2928
2929 /* as ordered data IO finishes, this gets called so we can finish
2930  * an ordered extent if the range of bytes in the file it covers are
2931  * fully written.
2932  */
2933 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
2934 {
2935         struct inode *inode = ordered_extent->inode;
2936         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2937         struct btrfs_root *root = BTRFS_I(inode)->root;
2938         struct btrfs_trans_handle *trans = NULL;
2939         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2940         struct extent_state *cached_state = NULL;
2941         struct new_sa_defrag_extent *new = NULL;
2942         int compress_type = 0;
2943         int ret = 0;
2944         u64 logical_len = ordered_extent->len;
2945         bool nolock;
2946         bool truncated = false;
2947         bool range_locked = false;
2948         bool clear_new_delalloc_bytes = false;
2949
2950         if (!test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
2951             !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags) &&
2952             !test_bit(BTRFS_ORDERED_DIRECT, &ordered_extent->flags))
2953                 clear_new_delalloc_bytes = true;
2954
2955         nolock = btrfs_is_free_space_inode(BTRFS_I(inode));
2956
2957         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
2958                 ret = -EIO;
2959                 goto out;
2960         }
2961
2962         btrfs_free_io_failure_record(BTRFS_I(inode),
2963                         ordered_extent->file_offset,
2964                         ordered_extent->file_offset +
2965                         ordered_extent->len - 1);
2966
2967         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
2968                 truncated = true;
2969                 logical_len = ordered_extent->truncated_len;
2970                 /* Truncated the entire extent, don't bother adding */
2971                 if (!logical_len)
2972                         goto out;
2973         }
2974
2975         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
2976                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
2977
2978                 /*
2979                  * For mwrite(mmap + memset to write) case, we still reserve
2980                  * space for NOCOW range.
2981                  * As NOCOW won't cause a new delayed ref, just free the space
2982                  */
2983                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
2984                                        ordered_extent->len);
2985                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2986                 if (nolock)
2987                         trans = btrfs_join_transaction_nolock(root);
2988                 else
2989                         trans = btrfs_join_transaction(root);
2990                 if (IS_ERR(trans)) {
2991                         ret = PTR_ERR(trans);
2992                         trans = NULL;
2993                         goto out;
2994                 }
2995                 trans->block_rsv = &BTRFS_I(inode)->block_rsv;
2996                 ret = btrfs_update_inode_fallback(trans, root, inode);
2997                 if (ret) /* -ENOMEM or corruption */
2998                         btrfs_abort_transaction(trans, ret);
2999                 goto out;
3000         }
3001
3002         range_locked = true;
3003         lock_extent_bits(io_tree, ordered_extent->file_offset,
3004                          ordered_extent->file_offset + ordered_extent->len - 1,
3005                          &cached_state);
3006
3007         ret = test_range_bit(io_tree, ordered_extent->file_offset,
3008                         ordered_extent->file_offset + ordered_extent->len - 1,
3009                         EXTENT_DEFRAG, 0, cached_state);
3010         if (ret) {
3011                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
3012                 if (0 && last_snapshot >= BTRFS_I(inode)->generation)
3013                         /* the inode is shared */
3014                         new = record_old_file_extents(inode, ordered_extent);
3015
3016                 clear_extent_bit(io_tree, ordered_extent->file_offset,
3017                         ordered_extent->file_offset + ordered_extent->len - 1,
3018                         EXTENT_DEFRAG, 0, 0, &cached_state);
3019         }
3020
3021         if (nolock)
3022                 trans = btrfs_join_transaction_nolock(root);
3023         else
3024                 trans = btrfs_join_transaction(root);
3025         if (IS_ERR(trans)) {
3026                 ret = PTR_ERR(trans);
3027                 trans = NULL;
3028                 goto out;
3029         }
3030
3031         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
3032
3033         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
3034                 compress_type = ordered_extent->compress_type;
3035         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
3036                 BUG_ON(compress_type);
3037                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
3038                                        ordered_extent->len);
3039                 ret = btrfs_mark_extent_written(trans, BTRFS_I(inode),
3040                                                 ordered_extent->file_offset,
3041                                                 ordered_extent->file_offset +
3042                                                 logical_len);
3043         } else {
3044                 BUG_ON(root == fs_info->tree_root);
3045                 ret = insert_reserved_file_extent(trans, inode,
3046                                                 ordered_extent->file_offset,
3047                                                 ordered_extent->start,
3048                                                 ordered_extent->disk_len,
3049                                                 logical_len, logical_len,
3050                                                 compress_type, 0, 0,
3051                                                 BTRFS_FILE_EXTENT_REG);
3052                 if (!ret)
3053                         btrfs_release_delalloc_bytes(fs_info,
3054                                                      ordered_extent->start,
3055                                                      ordered_extent->disk_len);
3056         }
3057         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
3058                            ordered_extent->file_offset, ordered_extent->len,
3059                            trans->transid);
3060         if (ret < 0) {
3061                 btrfs_abort_transaction(trans, ret);
3062                 goto out;
3063         }
3064
3065         add_pending_csums(trans, inode, &ordered_extent->list);
3066
3067         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
3068         ret = btrfs_update_inode_fallback(trans, root, inode);
3069         if (ret) { /* -ENOMEM or corruption */
3070                 btrfs_abort_transaction(trans, ret);
3071                 goto out;
3072         }
3073         ret = 0;
3074 out:
3075         if (range_locked || clear_new_delalloc_bytes) {
3076                 unsigned int clear_bits = 0;
3077
3078                 if (range_locked)
3079                         clear_bits |= EXTENT_LOCKED;
3080                 if (clear_new_delalloc_bytes)
3081                         clear_bits |= EXTENT_DELALLOC_NEW;
3082                 clear_extent_bit(&BTRFS_I(inode)->io_tree,
3083                                  ordered_extent->file_offset,
3084                                  ordered_extent->file_offset +
3085                                  ordered_extent->len - 1,
3086                                  clear_bits,
3087                                  (clear_bits & EXTENT_LOCKED) ? 1 : 0,
3088                                  0, &cached_state);
3089         }
3090
3091         if (trans)
3092                 btrfs_end_transaction(trans);
3093
3094         if (ret || truncated) {
3095                 u64 start, end;
3096
3097                 if (truncated)
3098                         start = ordered_extent->file_offset + logical_len;
3099                 else
3100                         start = ordered_extent->file_offset;
3101                 end = ordered_extent->file_offset + ordered_extent->len - 1;
3102                 clear_extent_uptodate(io_tree, start, end, NULL);
3103
3104                 /* Drop the cache for the part of the extent we didn't write. */
3105                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0);
3106
3107                 /*
3108                  * If the ordered extent had an IOERR or something else went
3109                  * wrong we need to return the space for this ordered extent
3110                  * back to the allocator.  We only free the extent in the
3111                  * truncated case if we didn't write out the extent at all.
3112                  */
3113                 if ((ret || !logical_len) &&
3114                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
3115                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags))
3116                         btrfs_free_reserved_extent(fs_info,
3117                                                    ordered_extent->start,
3118                                                    ordered_extent->disk_len, 1);
3119         }
3120
3121
3122         /*
3123          * This needs to be done to make sure anybody waiting knows we are done
3124          * updating everything for this ordered extent.
3125          */
3126         btrfs_remove_ordered_extent(inode, ordered_extent);
3127
3128         /* for snapshot-aware defrag */
3129         if (new) {
3130                 if (ret) {
3131                         free_sa_defrag_extent(new);
3132                         atomic_dec(&fs_info->defrag_running);
3133                 } else {
3134                         relink_file_extents(new);
3135                 }
3136         }
3137
3138         /* once for us */
3139         btrfs_put_ordered_extent(ordered_extent);
3140         /* once for the tree */
3141         btrfs_put_ordered_extent(ordered_extent);
3142
3143         return ret;
3144 }
3145
3146 static void finish_ordered_fn(struct btrfs_work *work)
3147 {
3148         struct btrfs_ordered_extent *ordered_extent;
3149         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
3150         btrfs_finish_ordered_io(ordered_extent);
3151 }
3152
3153 static void btrfs_writepage_end_io_hook(struct page *page, u64 start, u64 end,
3154                                 struct extent_state *state, int uptodate)
3155 {
3156         struct inode *inode = page->mapping->host;
3157         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3158         struct btrfs_ordered_extent *ordered_extent = NULL;
3159         struct btrfs_workqueue *wq;
3160         btrfs_work_func_t func;
3161
3162         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
3163
3164         ClearPagePrivate2(page);
3165         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
3166                                             end - start + 1, uptodate))
3167                 return;
3168
3169         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
3170                 wq = fs_info->endio_freespace_worker;
3171                 func = btrfs_freespace_write_helper;
3172         } else {
3173                 wq = fs_info->endio_write_workers;
3174                 func = btrfs_endio_write_helper;
3175         }
3176
3177         btrfs_init_work(&ordered_extent->work, func, finish_ordered_fn, NULL,
3178                         NULL);
3179         btrfs_queue_work(wq, &ordered_extent->work);
3180 }
3181
3182 static int __readpage_endio_check(struct inode *inode,
3183                                   struct btrfs_io_bio *io_bio,
3184                                   int icsum, struct page *page,
3185                                   int pgoff, u64 start, size_t len)
3186 {
3187         char *kaddr;
3188         u32 csum_expected;
3189         u32 csum = ~(u32)0;
3190
3191         csum_expected = *(((u32 *)io_bio->csum) + icsum);
3192
3193         kaddr = kmap_atomic(page);
3194         csum = btrfs_csum_data(kaddr + pgoff, csum,  len);
3195         btrfs_csum_final(csum, (u8 *)&csum);
3196         if (csum != csum_expected)
3197                 goto zeroit;
3198
3199         kunmap_atomic(kaddr);
3200         return 0;
3201 zeroit:
3202         btrfs_print_data_csum_error(BTRFS_I(inode), start, csum, csum_expected,
3203                                     io_bio->mirror_num);
3204         memset(kaddr + pgoff, 1, len);
3205         flush_dcache_page(page);
3206         kunmap_atomic(kaddr);
3207         return -EIO;
3208 }
3209
3210 /*
3211  * when reads are done, we need to check csums to verify the data is correct
3212  * if there's a match, we allow the bio to finish.  If not, the code in
3213  * extent_io.c will try to find good copies for us.
3214  */
3215 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
3216                                       u64 phy_offset, struct page *page,
3217                                       u64 start, u64 end, int mirror)
3218 {
3219         size_t offset = start - page_offset(page);
3220         struct inode *inode = page->mapping->host;
3221         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3222         struct btrfs_root *root = BTRFS_I(inode)->root;
3223
3224         if (PageChecked(page)) {
3225                 ClearPageChecked(page);
3226                 return 0;
3227         }
3228
3229         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
3230                 return 0;
3231
3232         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
3233             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
3234                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM);
3235                 return 0;
3236         }
3237
3238         phy_offset >>= inode->i_sb->s_blocksize_bits;
3239         return __readpage_endio_check(inode, io_bio, phy_offset, page, offset,
3240                                       start, (size_t)(end - start + 1));
3241 }
3242
3243 void btrfs_add_delayed_iput(struct inode *inode)
3244 {
3245         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3246         struct btrfs_inode *binode = BTRFS_I(inode);
3247
3248         if (atomic_add_unless(&inode->i_count, -1, 1))
3249                 return;
3250
3251         spin_lock(&fs_info->delayed_iput_lock);
3252         if (binode->delayed_iput_count == 0) {
3253                 ASSERT(list_empty(&binode->delayed_iput));
3254                 list_add_tail(&binode->delayed_iput, &fs_info->delayed_iputs);
3255         } else {
3256                 binode->delayed_iput_count++;
3257         }
3258         spin_unlock(&fs_info->delayed_iput_lock);
3259 }
3260
3261 void btrfs_run_delayed_iputs(struct btrfs_fs_info *fs_info)
3262 {
3263
3264         spin_lock(&fs_info->delayed_iput_lock);
3265         while (!list_empty(&fs_info->delayed_iputs)) {
3266                 struct btrfs_inode *inode;
3267
3268                 inode = list_first_entry(&fs_info->delayed_iputs,
3269                                 struct btrfs_inode, delayed_iput);
3270                 if (inode->delayed_iput_count) {
3271                         inode->delayed_iput_count--;
3272                         list_move_tail(&inode->delayed_iput,
3273                                         &fs_info->delayed_iputs);
3274                 } else {
3275                         list_del_init(&inode->delayed_iput);
3276                 }
3277                 spin_unlock(&fs_info->delayed_iput_lock);
3278                 iput(&inode->vfs_inode);
3279                 spin_lock(&fs_info->delayed_iput_lock);
3280         }
3281         spin_unlock(&fs_info->delayed_iput_lock);
3282 }
3283
3284 /*
3285  * This is called in transaction commit time. If there are no orphan
3286  * files in the subvolume, it removes orphan item and frees block_rsv
3287  * structure.
3288  */
3289 void btrfs_orphan_commit_root(struct btrfs_trans_handle *trans,
3290                               struct btrfs_root *root)
3291 {
3292         struct btrfs_fs_info *fs_info = root->fs_info;
3293         struct btrfs_block_rsv *block_rsv;
3294         int ret;
3295
3296         if (atomic_read(&root->orphan_inodes) ||
3297             root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE)
3298                 return;
3299
3300         spin_lock(&root->orphan_lock);
3301         if (atomic_read(&root->orphan_inodes)) {
3302                 spin_unlock(&root->orphan_lock);
3303                 return;
3304         }
3305
3306         if (root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE) {
3307                 spin_unlock(&root->orphan_lock);
3308                 return;
3309         }
3310
3311         block_rsv = root->orphan_block_rsv;
3312         root->orphan_block_rsv = NULL;
3313         spin_unlock(&root->orphan_lock);
3314
3315         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state) &&
3316             btrfs_root_refs(&root->root_item) > 0) {
3317                 ret = btrfs_del_orphan_item(trans, fs_info->tree_root,
3318                                             root->root_key.objectid);
3319                 if (ret)
3320                         btrfs_abort_transaction(trans, ret);
3321                 else
3322                         clear_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED,
3323                                   &root->state);
3324         }
3325
3326         if (block_rsv) {
3327                 WARN_ON(block_rsv->size > 0);
3328                 btrfs_free_block_rsv(fs_info, block_rsv);
3329         }
3330 }
3331
3332 /*
3333  * This creates an orphan entry for the given inode in case something goes
3334  * wrong in the middle of an unlink/truncate.
3335  *
3336  * NOTE: caller of this function should reserve 5 units of metadata for
3337  *       this function.
3338  */
3339 int btrfs_orphan_add(struct btrfs_trans_handle *trans,
3340                 struct btrfs_inode *inode)
3341 {
3342         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
3343         struct btrfs_root *root = inode->root;
3344         struct btrfs_block_rsv *block_rsv = NULL;
3345         int reserve = 0;
3346         int insert = 0;
3347         int ret;
3348
3349         if (!root->orphan_block_rsv) {
3350                 block_rsv = btrfs_alloc_block_rsv(fs_info,
3351                                                   BTRFS_BLOCK_RSV_TEMP);
3352                 if (!block_rsv)
3353                         return -ENOMEM;
3354         }
3355
3356         spin_lock(&root->orphan_lock);
3357         if (!root->orphan_block_rsv) {
3358                 root->orphan_block_rsv = block_rsv;
3359         } else if (block_rsv) {
3360                 btrfs_free_block_rsv(fs_info, block_rsv);
3361                 block_rsv = NULL;
3362         }
3363
3364         if (!test_and_set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3365                               &inode->runtime_flags)) {
3366 #if 0
3367                 /*
3368                  * For proper ENOSPC handling, we should do orphan
3369                  * cleanup when mounting. But this introduces backward
3370                  * compatibility issue.
3371                  */
3372                 if (!xchg(&root->orphan_item_inserted, 1))
3373                         insert = 2;
3374                 else
3375                         insert = 1;
3376 #endif
3377                 insert = 1;
3378                 atomic_inc(&root->orphan_inodes);
3379         }
3380
3381         if (!test_and_set_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3382                               &inode->runtime_flags))
3383                 reserve = 1;
3384         spin_unlock(&root->orphan_lock);
3385
3386         /* grab metadata reservation from transaction handle */
3387         if (reserve) {
3388                 ret = btrfs_orphan_reserve_metadata(trans, inode);
3389                 ASSERT(!ret);
3390                 if (ret) {
3391                         /*
3392                          * dec doesn't need spin_lock as ->orphan_block_rsv
3393                          * would be released only if ->orphan_inodes is
3394                          * zero.
3395                          */
3396                         atomic_dec(&root->orphan_inodes);
3397                         clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3398                                   &inode->runtime_flags);
3399                         if (insert)
3400                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3401                                           &inode->runtime_flags);
3402                         return ret;
3403                 }
3404         }
3405
3406         /* insert an orphan item to track this unlinked/truncated file */
3407         if (insert >= 1) {
3408                 ret = btrfs_insert_orphan_item(trans, root, btrfs_ino(inode));
3409                 if (ret) {
3410                         if (reserve) {
3411                                 clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3412                                           &inode->runtime_flags);
3413                                 btrfs_orphan_release_metadata(inode);
3414                         }
3415                         /*
3416                          * btrfs_orphan_commit_root may race with us and set
3417                          * ->orphan_block_rsv to zero, in order to avoid that,
3418                          * decrease ->orphan_inodes after everything is done.
3419                          */
3420                         atomic_dec(&root->orphan_inodes);
3421                         if (ret != -EEXIST) {
3422                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3423                                           &inode->runtime_flags);
3424                                 btrfs_abort_transaction(trans, ret);
3425                                 return ret;
3426                         }
3427                 }
3428                 ret = 0;
3429         }
3430
3431         /* insert an orphan item to track subvolume contains orphan files */
3432         if (insert >= 2) {
3433                 ret = btrfs_insert_orphan_item(trans, fs_info->tree_root,
3434                                                root->root_key.objectid);
3435                 if (ret && ret != -EEXIST) {
3436                         btrfs_abort_transaction(trans, ret);
3437                         return ret;
3438                 }
3439         }
3440         return 0;
3441 }
3442
3443 /*
3444  * We have done the truncate/delete so we can go ahead and remove the orphan
3445  * item for this particular inode.
3446  */
3447 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3448                             struct btrfs_inode *inode)
3449 {
3450         struct btrfs_root *root = inode->root;
3451         int delete_item = 0;
3452         int ret = 0;
3453
3454         if (test_and_clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3455                                &inode->runtime_flags))
3456                 delete_item = 1;
3457
3458         if (delete_item && trans)
3459                 ret = btrfs_del_orphan_item(trans, root, btrfs_ino(inode));
3460
3461         if (test_and_clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3462                                &inode->runtime_flags))
3463                 btrfs_orphan_release_metadata(inode);
3464
3465         /*
3466          * btrfs_orphan_commit_root may race with us and set ->orphan_block_rsv
3467          * to zero, in order to avoid that, decrease ->orphan_inodes after
3468          * everything is done.
3469          */
3470         if (delete_item)
3471                 atomic_dec(&root->orphan_inodes);
3472
3473         return ret;
3474 }
3475
3476 /*
3477  * this cleans up any orphans that may be left on the list from the last use
3478  * of this root.
3479  */
3480 int btrfs_orphan_cleanup(struct btrfs_root *root)
3481 {
3482         struct btrfs_fs_info *fs_info = root->fs_info;
3483         struct btrfs_path *path;
3484         struct extent_buffer *leaf;
3485         struct btrfs_key key, found_key;
3486         struct btrfs_trans_handle *trans;
3487         struct inode *inode;
3488         u64 last_objectid = 0;
3489         int ret = 0, nr_unlink = 0, nr_truncate = 0;
3490
3491         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3492                 return 0;
3493
3494         path = btrfs_alloc_path();
3495         if (!path) {
3496                 ret = -ENOMEM;
3497                 goto out;
3498         }
3499         path->reada = READA_BACK;
3500
3501         key.objectid = BTRFS_ORPHAN_OBJECTID;
3502         key.type = BTRFS_ORPHAN_ITEM_KEY;
3503         key.offset = (u64)-1;
3504
3505         while (1) {
3506                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3507                 if (ret < 0)
3508                         goto out;
3509
3510                 /*
3511                  * if ret == 0 means we found what we were searching for, which
3512                  * is weird, but possible, so only screw with path if we didn't
3513                  * find the key and see if we have stuff that matches
3514                  */
3515                 if (ret > 0) {
3516                         ret = 0;
3517                         if (path->slots[0] == 0)
3518                                 break;
3519                         path->slots[0]--;
3520                 }
3521
3522                 /* pull out the item */
3523                 leaf = path->nodes[0];
3524                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3525
3526                 /* make sure the item matches what we want */
3527                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3528                         break;
3529                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3530                         break;
3531
3532                 /* release the path since we're done with it */
3533                 btrfs_release_path(path);
3534
3535                 /*
3536                  * this is where we are basically btrfs_lookup, without the
3537                  * crossing root thing.  we store the inode number in the
3538                  * offset of the orphan item.
3539                  */
3540
3541                 if (found_key.offset == last_objectid) {
3542                         btrfs_err(fs_info,
3543                                   "Error removing orphan entry, stopping orphan cleanup");
3544                         ret = -EINVAL;
3545                         goto out;
3546                 }
3547
3548                 last_objectid = found_key.offset;
3549
3550                 found_key.objectid = found_key.offset;
3551                 found_key.type = BTRFS_INODE_ITEM_KEY;
3552                 found_key.offset = 0;
3553                 inode = btrfs_iget(fs_info->sb, &found_key, root, NULL);
3554                 ret = PTR_ERR_OR_ZERO(inode);
3555                 if (ret && ret != -ENOENT)
3556                         goto out;
3557
3558                 if (ret == -ENOENT && root == fs_info->tree_root) {
3559                         struct btrfs_root *dead_root;
3560                         struct btrfs_fs_info *fs_info = root->fs_info;
3561                         int is_dead_root = 0;
3562
3563                         /*
3564                          * this is an orphan in the tree root. Currently these
3565                          * could come from 2 sources:
3566                          *  a) a snapshot deletion in progress
3567                          *  b) a free space cache inode
3568                          * We need to distinguish those two, as the snapshot
3569                          * orphan must not get deleted.
3570                          * find_dead_roots already ran before us, so if this
3571                          * is a snapshot deletion, we should find the root
3572                          * in the dead_roots list
3573                          */
3574                         spin_lock(&fs_info->trans_lock);
3575                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3576                                             root_list) {
3577                                 if (dead_root->root_key.objectid ==
3578                                     found_key.objectid) {
3579                                         is_dead_root = 1;
3580                                         break;
3581                                 }
3582                         }
3583                         spin_unlock(&fs_info->trans_lock);
3584                         if (is_dead_root) {
3585                                 /* prevent this orphan from being found again */
3586                                 key.offset = found_key.objectid - 1;
3587                                 continue;
3588                         }
3589                 }
3590                 /*
3591                  * Inode is already gone but the orphan item is still there,
3592                  * kill the orphan item.
3593                  */
3594                 if (ret == -ENOENT) {
3595                         trans = btrfs_start_transaction(root, 1);
3596                         if (IS_ERR(trans)) {
3597                                 ret = PTR_ERR(trans);
3598                                 goto out;
3599                         }
3600                         btrfs_debug(fs_info, "auto deleting %Lu",
3601                                     found_key.objectid);
3602                         ret = btrfs_del_orphan_item(trans, root,
3603                                                     found_key.objectid);
3604                         btrfs_end_transaction(trans);
3605                         if (ret)
3606                                 goto out;
3607                         continue;
3608                 }
3609
3610                 /*
3611                  * add this inode to the orphan list so btrfs_orphan_del does
3612                  * the proper thing when we hit it
3613                  */
3614                 set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3615                         &BTRFS_I(inode)->runtime_flags);
3616                 atomic_inc(&root->orphan_inodes);
3617
3618                 /* if we have links, this was a truncate, lets do that */
3619                 if (inode->i_nlink) {
3620                         if (WARN_ON(!S_ISREG(inode->i_mode))) {
3621                                 iput(inode);
3622                                 continue;
3623                         }
3624                         nr_truncate++;
3625
3626                         /* 1 for the orphan item deletion. */
3627                         trans = btrfs_start_transaction(root, 1);
3628                         if (IS_ERR(trans)) {
3629                                 iput(inode);
3630                                 ret = PTR_ERR(trans);
3631                                 goto out;
3632                         }
3633                         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
3634                         btrfs_end_transaction(trans);
3635                         if (ret) {
3636                                 iput(inode);
3637                                 goto out;
3638                         }
3639
3640                         ret = btrfs_truncate(inode);
3641                         if (ret)
3642                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
3643                 } else {
3644                         nr_unlink++;
3645                 }
3646
3647                 /* this will do delete_inode and everything for us */
3648                 iput(inode);
3649                 if (ret)
3650                         goto out;
3651         }
3652         /* release the path since we're done with it */
3653         btrfs_release_path(path);
3654
3655         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3656
3657         if (root->orphan_block_rsv)
3658                 btrfs_block_rsv_release(fs_info, root->orphan_block_rsv,
3659                                         (u64)-1);
3660
3661         if (root->orphan_block_rsv ||
3662             test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3663                 trans = btrfs_join_transaction(root);
3664                 if (!IS_ERR(trans))
3665                         btrfs_end_transaction(trans);
3666         }
3667
3668         if (nr_unlink)
3669                 btrfs_debug(fs_info, "unlinked %d orphans", nr_unlink);
3670         if (nr_truncate)
3671                 btrfs_debug(fs_info, "truncated %d orphans", nr_truncate);
3672
3673 out:
3674         if (ret)
3675                 btrfs_err(fs_info, "could not do orphan cleanup %d", ret);
3676         btrfs_free_path(path);
3677         return ret;
3678 }
3679
3680 /*
3681  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3682  * don't find any xattrs, we know there can't be any acls.
3683  *
3684  * slot is the slot the inode is in, objectid is the objectid of the inode
3685  */
3686 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3687                                           int slot, u64 objectid,
3688                                           int *first_xattr_slot)
3689 {
3690         u32 nritems = btrfs_header_nritems(leaf);
3691         struct btrfs_key found_key;
3692         static u64 xattr_access = 0;
3693         static u64 xattr_default = 0;
3694         int scanned = 0;
3695
3696         if (!xattr_access) {
3697                 xattr_access = btrfs_name_hash(XATTR_NAME_POSIX_ACL_ACCESS,
3698                                         strlen(XATTR_NAME_POSIX_ACL_ACCESS));
3699                 xattr_default = btrfs_name_hash(XATTR_NAME_POSIX_ACL_DEFAULT,
3700                                         strlen(XATTR_NAME_POSIX_ACL_DEFAULT));
3701         }
3702
3703         slot++;
3704         *first_xattr_slot = -1;
3705         while (slot < nritems) {
3706                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3707
3708                 /* we found a different objectid, there must not be acls */
3709                 if (found_key.objectid != objectid)
3710                         return 0;
3711
3712                 /* we found an xattr, assume we've got an acl */
3713                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3714                         if (*first_xattr_slot == -1)
3715                                 *first_xattr_slot = slot;
3716                         if (found_key.offset == xattr_access ||
3717                             found_key.offset == xattr_default)
3718                                 return 1;
3719                 }
3720
3721                 /*
3722                  * we found a key greater than an xattr key, there can't
3723                  * be any acls later on
3724                  */
3725                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3726                         return 0;
3727
3728                 slot++;
3729                 scanned++;
3730
3731                 /*
3732                  * it goes inode, inode backrefs, xattrs, extents,
3733                  * so if there are a ton of hard links to an inode there can
3734                  * be a lot of backrefs.  Don't waste time searching too hard,
3735                  * this is just an optimization
3736                  */
3737                 if (scanned >= 8)
3738                         break;
3739         }
3740         /* we hit the end of the leaf before we found an xattr or
3741          * something larger than an xattr.  We have to assume the inode
3742          * has acls
3743          */
3744         if (*first_xattr_slot == -1)
3745                 *first_xattr_slot = slot;
3746         return 1;
3747 }
3748
3749 /*
3750  * read an inode from the btree into the in-memory inode
3751  */
3752 static int btrfs_read_locked_inode(struct inode *inode)
3753 {
3754         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3755         struct btrfs_path *path;
3756         struct extent_buffer *leaf;
3757         struct btrfs_inode_item *inode_item;
3758         struct btrfs_root *root = BTRFS_I(inode)->root;
3759         struct btrfs_key location;
3760         unsigned long ptr;
3761         int maybe_acls;
3762         u32 rdev;
3763         int ret;
3764         bool filled = false;
3765         int first_xattr_slot;
3766
3767         ret = btrfs_fill_inode(inode, &rdev);
3768         if (!ret)
3769                 filled = true;
3770
3771         path = btrfs_alloc_path();
3772         if (!path) {
3773                 ret = -ENOMEM;
3774                 goto make_bad;
3775         }
3776
3777         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3778
3779         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3780         if (ret) {
3781                 if (ret > 0)
3782                         ret = -ENOENT;
3783                 goto make_bad;
3784         }
3785
3786         leaf = path->nodes[0];
3787
3788         if (filled)
3789                 goto cache_index;
3790
3791         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3792                                     struct btrfs_inode_item);
3793         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3794         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3795         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3796         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3797         btrfs_i_size_write(BTRFS_I(inode), btrfs_inode_size(leaf, inode_item));
3798
3799         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3800         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3801
3802         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3803         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3804
3805         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3806         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3807
3808         BTRFS_I(inode)->i_otime.tv_sec =
3809                 btrfs_timespec_sec(leaf, &inode_item->otime);
3810         BTRFS_I(inode)->i_otime.tv_nsec =
3811                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3812
3813         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3814         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3815         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3816
3817         inode_set_iversion_queried(inode,
3818                                    btrfs_inode_sequence(leaf, inode_item));
3819         inode->i_generation = BTRFS_I(inode)->generation;
3820         inode->i_rdev = 0;
3821         rdev = btrfs_inode_rdev(leaf, inode_item);
3822
3823         BTRFS_I(inode)->index_cnt = (u64)-1;
3824         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3825
3826 cache_index:
3827         /*
3828          * If we were modified in the current generation and evicted from memory
3829          * and then re-read we need to do a full sync since we don't have any
3830          * idea about which extents were modified before we were evicted from
3831          * cache.
3832          *
3833          * This is required for both inode re-read from disk and delayed inode
3834          * in delayed_nodes_tree.
3835          */
3836         if (BTRFS_I(inode)->last_trans == fs_info->generation)
3837                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3838                         &BTRFS_I(inode)->runtime_flags);
3839
3840         /*
3841          * We don't persist the id of the transaction where an unlink operation
3842          * against the inode was last made. So here we assume the inode might
3843          * have been evicted, and therefore the exact value of last_unlink_trans
3844          * lost, and set it to last_trans to avoid metadata inconsistencies
3845          * between the inode and its parent if the inode is fsync'ed and the log
3846          * replayed. For example, in the scenario:
3847          *
3848          * touch mydir/foo
3849          * ln mydir/foo mydir/bar
3850          * sync
3851          * unlink mydir/bar
3852          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3853          * xfs_io -c fsync mydir/foo
3854          * <power failure>
3855          * mount fs, triggers fsync log replay
3856          *
3857          * We must make sure that when we fsync our inode foo we also log its
3858          * parent inode, otherwise after log replay the parent still has the
3859          * dentry with the "bar" name but our inode foo has a link count of 1
3860          * and doesn't have an inode ref with the name "bar" anymore.
3861          *
3862          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3863          * but it guarantees correctness at the expense of occasional full
3864          * transaction commits on fsync if our inode is a directory, or if our
3865          * inode is not a directory, logging its parent unnecessarily.
3866          */
3867         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3868
3869         path->slots[0]++;
3870         if (inode->i_nlink != 1 ||
3871             path->slots[0] >= btrfs_header_nritems(leaf))
3872                 goto cache_acl;
3873
3874         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3875         if (location.objectid != btrfs_ino(BTRFS_I(inode)))
3876                 goto cache_acl;
3877
3878         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3879         if (location.type == BTRFS_INODE_REF_KEY) {
3880                 struct btrfs_inode_ref *ref;
3881
3882                 ref = (struct btrfs_inode_ref *)ptr;
3883                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3884         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3885                 struct btrfs_inode_extref *extref;
3886
3887                 extref = (struct btrfs_inode_extref *)ptr;
3888                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3889                                                                      extref);
3890         }
3891 cache_acl:
3892         /*
3893          * try to precache a NULL acl entry for files that don't have
3894          * any xattrs or acls
3895          */
3896         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3897                         btrfs_ino(BTRFS_I(inode)), &first_xattr_slot);
3898         if (first_xattr_slot != -1) {
3899                 path->slots[0] = first_xattr_slot;
3900                 ret = btrfs_load_inode_props(inode, path);
3901                 if (ret)
3902                         btrfs_err(fs_info,
3903                                   "error loading props for ino %llu (root %llu): %d",
3904                                   btrfs_ino(BTRFS_I(inode)),
3905                                   root->root_key.objectid, ret);
3906         }
3907         btrfs_free_path(path);
3908
3909         if (!maybe_acls)
3910                 cache_no_acl(inode);
3911
3912         switch (inode->i_mode & S_IFMT) {
3913         case S_IFREG:
3914                 inode->i_mapping->a_ops = &btrfs_aops;
3915                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3916                 inode->i_fop = &btrfs_file_operations;
3917                 inode->i_op = &btrfs_file_inode_operations;
3918                 break;
3919         case S_IFDIR:
3920                 inode->i_fop = &btrfs_dir_file_operations;
3921                 inode->i_op = &btrfs_dir_inode_operations;
3922                 break;
3923         case S_IFLNK:
3924                 inode->i_op = &btrfs_symlink_inode_operations;
3925                 inode_nohighmem(inode);
3926                 inode->i_mapping->a_ops = &btrfs_symlink_aops;
3927                 break;
3928         default:
3929                 inode->i_op = &btrfs_special_inode_operations;
3930                 init_special_inode(inode, inode->i_mode, rdev);
3931                 break;
3932         }
3933
3934         btrfs_update_iflags(inode);
3935         return 0;
3936
3937 make_bad:
3938         btrfs_free_path(path);
3939         make_bad_inode(inode);
3940         return ret;
3941 }
3942
3943 /*
3944  * given a leaf and an inode, copy the inode fields into the leaf
3945  */
3946 static void fill_inode_item(struct btrfs_trans_handle *trans,
3947                             struct extent_buffer *leaf,
3948                             struct btrfs_inode_item *item,
3949                             struct inode *inode)
3950 {
3951         struct btrfs_map_token token;
3952
3953         btrfs_init_map_token(&token);
3954
3955         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3956         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3957         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3958                                    &token);
3959         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3960         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3961
3962         btrfs_set_token_timespec_sec(leaf, &item->atime,
3963                                      inode->i_atime.tv_sec, &token);
3964         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3965                                       inode->i_atime.tv_nsec, &token);
3966
3967         btrfs_set_token_timespec_sec(leaf, &item->mtime,
3968                                      inode->i_mtime.tv_sec, &token);
3969         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
3970                                       inode->i_mtime.tv_nsec, &token);
3971
3972         btrfs_set_token_timespec_sec(leaf, &item->ctime,
3973                                      inode->i_ctime.tv_sec, &token);
3974         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
3975                                       inode->i_ctime.tv_nsec, &token);
3976
3977         btrfs_set_token_timespec_sec(leaf, &item->otime,
3978                                      BTRFS_I(inode)->i_otime.tv_sec, &token);
3979         btrfs_set_token_timespec_nsec(leaf, &item->otime,
3980                                       BTRFS_I(inode)->i_otime.tv_nsec, &token);
3981
3982         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3983                                      &token);
3984         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
3985                                          &token);
3986         btrfs_set_token_inode_sequence(leaf, item, inode_peek_iversion(inode),
3987                                        &token);
3988         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3989         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3990         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3991         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3992 }
3993
3994 /*
3995  * copy everything in the in-memory inode into the btree.
3996  */
3997 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
3998                                 struct btrfs_root *root, struct inode *inode)
3999 {
4000         struct btrfs_inode_item *inode_item;
4001         struct btrfs_path *path;
4002         struct extent_buffer *leaf;
4003         int ret;
4004
4005         path = btrfs_alloc_path();
4006         if (!path)
4007                 return -ENOMEM;
4008
4009         path->leave_spinning = 1;
4010         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
4011                                  1);
4012         if (ret) {
4013                 if (ret > 0)
4014                         ret = -ENOENT;
4015                 goto failed;
4016         }
4017
4018         leaf = path->nodes[0];
4019         inode_item = btrfs_item_ptr(leaf, path->slots[0],
4020                                     struct btrfs_inode_item);
4021
4022         fill_inode_item(trans, leaf, inode_item, inode);
4023         btrfs_mark_buffer_dirty(leaf);
4024         btrfs_set_inode_last_trans(trans, inode);
4025         ret = 0;
4026 failed:
4027         btrfs_free_path(path);
4028         return ret;
4029 }
4030
4031 /*
4032  * copy everything in the in-memory inode into the btree.
4033  */
4034 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
4035                                 struct btrfs_root *root, struct inode *inode)
4036 {
4037         struct btrfs_fs_info *fs_info = root->fs_info;
4038         int ret;
4039
4040         /*
4041          * If the inode is a free space inode, we can deadlock during commit
4042          * if we put it into the delayed code.
4043          *
4044          * The data relocation inode should also be directly updated
4045          * without delay
4046          */
4047         if (!btrfs_is_free_space_inode(BTRFS_I(inode))
4048             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
4049             && !test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags)) {
4050                 btrfs_update_root_times(trans, root);
4051
4052                 ret = btrfs_delayed_update_inode(trans, root, inode);
4053                 if (!ret)
4054                         btrfs_set_inode_last_trans(trans, inode);
4055                 return ret;
4056         }
4057
4058         return btrfs_update_inode_item(trans, root, inode);
4059 }
4060
4061 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
4062                                          struct btrfs_root *root,
4063                                          struct inode *inode)
4064 {
4065         int ret;
4066
4067         ret = btrfs_update_inode(trans, root, inode);
4068         if (ret == -ENOSPC)
4069                 return btrfs_update_inode_item(trans, root, inode);
4070         return ret;
4071 }
4072
4073 /*
4074  * unlink helper that gets used here in inode.c and in the tree logging
4075  * recovery code.  It remove a link in a directory with a given name, and
4076  * also drops the back refs in the inode to the directory
4077  */
4078 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4079                                 struct btrfs_root *root,
4080                                 struct btrfs_inode *dir,
4081                                 struct btrfs_inode *inode,
4082                                 const char *name, int name_len)
4083 {
4084         struct btrfs_fs_info *fs_info = root->fs_info;
4085         struct btrfs_path *path;
4086         int ret = 0;
4087         struct extent_buffer *leaf;
4088         struct btrfs_dir_item *di;
4089         struct btrfs_key key;
4090         u64 index;
4091         u64 ino = btrfs_ino(inode);
4092         u64 dir_ino = btrfs_ino(dir);
4093
4094         path = btrfs_alloc_path();
4095         if (!path) {
4096                 ret = -ENOMEM;
4097                 goto out;
4098         }
4099
4100         path->leave_spinning = 1;
4101         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4102                                     name, name_len, -1);
4103         if (IS_ERR(di)) {
4104                 ret = PTR_ERR(di);
4105                 goto err;
4106         }
4107         if (!di) {
4108                 ret = -ENOENT;
4109                 goto err;
4110         }
4111         leaf = path->nodes[0];
4112         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4113         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4114         if (ret)
4115                 goto err;
4116         btrfs_release_path(path);
4117
4118         /*
4119          * If we don't have dir index, we have to get it by looking up
4120          * the inode ref, since we get the inode ref, remove it directly,
4121          * it is unnecessary to do delayed deletion.
4122          *
4123          * But if we have dir index, needn't search inode ref to get it.
4124          * Since the inode ref is close to the inode item, it is better
4125          * that we delay to delete it, and just do this deletion when
4126          * we update the inode item.
4127          */
4128         if (inode->dir_index) {
4129                 ret = btrfs_delayed_delete_inode_ref(inode);
4130                 if (!ret) {
4131                         index = inode->dir_index;
4132                         goto skip_backref;
4133                 }
4134         }
4135
4136         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
4137                                   dir_ino, &index);
4138         if (ret) {
4139                 btrfs_info(fs_info,
4140                         "failed to delete reference to %.*s, inode %llu parent %llu",
4141                         name_len, name, ino, dir_ino);
4142                 btrfs_abort_transaction(trans, ret);
4143                 goto err;
4144         }
4145 skip_backref:
4146         ret = btrfs_delete_delayed_dir_index(trans, fs_info, dir, index);
4147         if (ret) {
4148                 btrfs_abort_transaction(trans, ret);
4149                 goto err;
4150         }
4151
4152         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len, inode,
4153                         dir_ino);
4154         if (ret != 0 && ret != -ENOENT) {
4155                 btrfs_abort_transaction(trans, ret);
4156                 goto err;
4157         }
4158
4159         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len, dir,
4160                         index);
4161         if (ret == -ENOENT)
4162                 ret = 0;
4163         else if (ret)
4164                 btrfs_abort_transaction(trans, ret);
4165 err:
4166         btrfs_free_path(path);
4167         if (ret)
4168                 goto out;
4169
4170         btrfs_i_size_write(dir, dir->vfs_inode.i_size - name_len * 2);
4171         inode_inc_iversion(&inode->vfs_inode);
4172         inode_inc_iversion(&dir->vfs_inode);
4173         inode->vfs_inode.i_ctime = dir->vfs_inode.i_mtime =
4174                 dir->vfs_inode.i_ctime = current_time(&inode->vfs_inode);
4175         ret = btrfs_update_inode(trans, root, &dir->vfs_inode);
4176 out:
4177         return ret;
4178 }
4179
4180 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4181                        struct btrfs_root *root,
4182                        struct btrfs_inode *dir, struct btrfs_inode *inode,
4183                        const char *name, int name_len)
4184 {
4185         int ret;
4186         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
4187         if (!ret) {
4188                 drop_nlink(&inode->vfs_inode);
4189                 ret = btrfs_update_inode(trans, root, &inode->vfs_inode);
4190         }
4191         return ret;
4192 }
4193
4194 /*
4195  * helper to start transaction for unlink and rmdir.
4196  *
4197  * unlink and rmdir are special in btrfs, they do not always free space, so
4198  * if we cannot make our reservations the normal way try and see if there is
4199  * plenty of slack room in the global reserve to migrate, otherwise we cannot
4200  * allow the unlink to occur.
4201  */
4202 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
4203 {
4204         struct btrfs_root *root = BTRFS_I(dir)->root;
4205
4206         /*
4207          * 1 for the possible orphan item
4208          * 1 for the dir item
4209          * 1 for the dir index
4210          * 1 for the inode ref
4211          * 1 for the inode
4212          */
4213         return btrfs_start_transaction_fallback_global_rsv(root, 5, 5);
4214 }
4215
4216 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
4217 {
4218         struct btrfs_root *root = BTRFS_I(dir)->root;
4219         struct btrfs_trans_handle *trans;
4220         struct inode *inode = d_inode(dentry);
4221         int ret;
4222
4223         trans = __unlink_start_trans(dir);
4224         if (IS_ERR(trans))
4225                 return PTR_ERR(trans);
4226
4227         btrfs_record_unlink_dir(trans, BTRFS_I(dir), BTRFS_I(d_inode(dentry)),
4228                         0);
4229
4230         ret = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4231                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4232                         dentry->d_name.len);
4233         if (ret)
4234                 goto out;
4235
4236         if (inode->i_nlink == 0) {
4237                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
4238                 if (ret)
4239                         goto out;
4240         }
4241
4242 out:
4243         btrfs_end_transaction(trans);
4244         btrfs_btree_balance_dirty(root->fs_info);
4245         return ret;
4246 }
4247
4248 int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
4249                         struct btrfs_root *root,
4250                         struct inode *dir, u64 objectid,
4251                         const char *name, int name_len)
4252 {
4253         struct btrfs_fs_info *fs_info = root->fs_info;
4254         struct btrfs_path *path;
4255         struct extent_buffer *leaf;
4256         struct btrfs_dir_item *di;
4257         struct btrfs_key key;
4258         u64 index;
4259         int ret;
4260         u64 dir_ino = btrfs_ino(BTRFS_I(dir));
4261
4262         path = btrfs_alloc_path();
4263         if (!path)
4264                 return -ENOMEM;
4265
4266         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4267                                    name, name_len, -1);
4268         if (IS_ERR_OR_NULL(di)) {
4269                 if (!di)
4270                         ret = -ENOENT;
4271                 else
4272                         ret = PTR_ERR(di);
4273                 goto out;
4274         }
4275
4276         leaf = path->nodes[0];
4277         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4278         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
4279         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4280         if (ret) {
4281                 btrfs_abort_transaction(trans, ret);
4282                 goto out;
4283         }
4284         btrfs_release_path(path);
4285
4286         ret = btrfs_del_root_ref(trans, fs_info, objectid,
4287                                  root->root_key.objectid, dir_ino,
4288                                  &index, name, name_len);
4289         if (ret < 0) {
4290                 if (ret != -ENOENT) {
4291                         btrfs_abort_transaction(trans, ret);
4292                         goto out;
4293                 }
4294                 di = btrfs_search_dir_index_item(root, path, dir_ino,
4295                                                  name, name_len);
4296                 if (IS_ERR_OR_NULL(di)) {
4297                         if (!di)
4298                                 ret = -ENOENT;
4299                         else
4300                                 ret = PTR_ERR(di);
4301                         btrfs_abort_transaction(trans, ret);
4302                         goto out;
4303                 }
4304
4305                 leaf = path->nodes[0];
4306                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4307                 btrfs_release_path(path);
4308                 index = key.offset;
4309         }
4310         btrfs_release_path(path);
4311
4312         ret = btrfs_delete_delayed_dir_index(trans, fs_info, BTRFS_I(dir), index);
4313         if (ret) {
4314                 btrfs_abort_transaction(trans, ret);
4315                 goto out;
4316         }
4317
4318         btrfs_i_size_write(BTRFS_I(dir), dir->i_size - name_len * 2);
4319         inode_inc_iversion(dir);
4320         dir->i_mtime = dir->i_ctime = current_time(dir);
4321         ret = btrfs_update_inode_fallback(trans, root, dir);
4322         if (ret)
4323                 btrfs_abort_transaction(trans, ret);
4324 out:
4325         btrfs_free_path(path);
4326         return ret;
4327 }
4328
4329 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4330 {
4331         struct inode *inode = d_inode(dentry);
4332         int err = 0;
4333         struct btrfs_root *root = BTRFS_I(dir)->root;
4334         struct btrfs_trans_handle *trans;
4335         u64 last_unlink_trans;
4336
4337         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4338                 return -ENOTEMPTY;
4339         if (btrfs_ino(BTRFS_I(inode)) == BTRFS_FIRST_FREE_OBJECTID)
4340                 return -EPERM;
4341
4342         trans = __unlink_start_trans(dir);
4343         if (IS_ERR(trans))
4344                 return PTR_ERR(trans);
4345
4346         if (unlikely(btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4347                 err = btrfs_unlink_subvol(trans, root, dir,
4348                                           BTRFS_I(inode)->location.objectid,
4349                                           dentry->d_name.name,
4350                                           dentry->d_name.len);
4351                 goto out;
4352         }
4353
4354         err = btrfs_orphan_add(trans, BTRFS_I(inode));
4355         if (err)
4356                 goto out;
4357
4358         last_unlink_trans = BTRFS_I(inode)->last_unlink_trans;
4359
4360         /* now the directory is empty */
4361         err = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4362                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4363                         dentry->d_name.len);
4364         if (!err) {
4365                 btrfs_i_size_write(BTRFS_I(inode), 0);
4366                 /*
4367                  * Propagate the last_unlink_trans value of the deleted dir to
4368                  * its parent directory. This is to prevent an unrecoverable
4369                  * log tree in the case we do something like this:
4370                  * 1) create dir foo
4371                  * 2) create snapshot under dir foo
4372                  * 3) delete the snapshot
4373                  * 4) rmdir foo
4374                  * 5) mkdir foo
4375                  * 6) fsync foo or some file inside foo
4376                  */
4377                 if (last_unlink_trans >= trans->transid)
4378                         BTRFS_I(dir)->last_unlink_trans = last_unlink_trans;
4379         }
4380 out:
4381         btrfs_end_transaction(trans);
4382         btrfs_btree_balance_dirty(root->fs_info);
4383
4384         return err;
4385 }
4386
4387 static int truncate_space_check(struct btrfs_trans_handle *trans,
4388                                 struct btrfs_root *root,
4389                                 u64 bytes_deleted)
4390 {
4391         struct btrfs_fs_info *fs_info = root->fs_info;
4392         int ret;
4393
4394         /*
4395          * This is only used to apply pressure to the enospc system, we don't
4396          * intend to use this reservation at all.
4397          */
4398         bytes_deleted = btrfs_csum_bytes_to_leaves(fs_info, bytes_deleted);
4399         bytes_deleted *= fs_info->nodesize;
4400         ret = btrfs_block_rsv_add(root, &fs_info->trans_block_rsv,
4401                                   bytes_deleted, BTRFS_RESERVE_NO_FLUSH);
4402         if (!ret) {
4403                 trace_btrfs_space_reservation(fs_info, "transaction",
4404                                               trans->transid,
4405                                               bytes_deleted, 1);
4406                 trans->bytes_reserved += bytes_deleted;
4407         }
4408         return ret;
4409
4410 }
4411
4412 /*
4413  * Return this if we need to call truncate_block for the last bit of the
4414  * truncate.
4415  */
4416 #define NEED_TRUNCATE_BLOCK 1
4417
4418 /*
4419  * this can truncate away extent items, csum items and directory items.
4420  * It starts at a high offset and removes keys until it can't find
4421  * any higher than new_size
4422  *
4423  * csum items that cross the new i_size are truncated to the new size
4424  * as well.
4425  *
4426  * min_type is the minimum key type to truncate down to.  If set to 0, this
4427  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4428  */
4429 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4430                                struct btrfs_root *root,
4431                                struct inode *inode,
4432                                u64 new_size, u32 min_type)
4433 {
4434         struct btrfs_fs_info *fs_info = root->fs_info;
4435         struct btrfs_path *path;
4436         struct extent_buffer *leaf;
4437         struct btrfs_file_extent_item *fi;
4438         struct btrfs_key key;
4439         struct btrfs_key found_key;
4440         u64 extent_start = 0;
4441         u64 extent_num_bytes = 0;
4442         u64 extent_offset = 0;
4443         u64 item_end = 0;
4444         u64 last_size = new_size;
4445         u32 found_type = (u8)-1;
4446         int found_extent;
4447         int del_item;
4448         int pending_del_nr = 0;
4449         int pending_del_slot = 0;
4450         int extent_type = -1;
4451         int ret;
4452         int err = 0;
4453         u64 ino = btrfs_ino(BTRFS_I(inode));
4454         u64 bytes_deleted = 0;
4455         bool be_nice = false;
4456         bool should_throttle = false;
4457         bool should_end = false;
4458
4459         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4460
4461         /*
4462          * for non-free space inodes and ref cows, we want to back off from
4463          * time to time
4464          */
4465         if (!btrfs_is_free_space_inode(BTRFS_I(inode)) &&
4466             test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4467                 be_nice = true;
4468
4469         path = btrfs_alloc_path();
4470         if (!path)
4471                 return -ENOMEM;
4472         path->reada = READA_BACK;
4473
4474         /*
4475          * We want to drop from the next block forward in case this new size is
4476          * not block aligned since we will be keeping the last block of the
4477          * extent just the way it is.
4478          */
4479         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4480             root == fs_info->tree_root)
4481                 btrfs_drop_extent_cache(BTRFS_I(inode), ALIGN(new_size,
4482                                         fs_info->sectorsize),
4483                                         (u64)-1, 0);
4484
4485         /*
4486          * This function is also used to drop the items in the log tree before
4487          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4488          * it is used to drop the loged items. So we shouldn't kill the delayed
4489          * items.
4490          */
4491         if (min_type == 0 && root == BTRFS_I(inode)->root)
4492                 btrfs_kill_delayed_inode_items(BTRFS_I(inode));
4493
4494         key.objectid = ino;
4495         key.offset = (u64)-1;
4496         key.type = (u8)-1;
4497
4498 search_again:
4499         /*
4500          * with a 16K leaf size and 128MB extents, you can actually queue
4501          * up a huge file in a single leaf.  Most of the time that
4502          * bytes_deleted is > 0, it will be huge by the time we get here
4503          */
4504         if (be_nice && bytes_deleted > SZ_32M) {
4505                 if (btrfs_should_end_transaction(trans)) {
4506                         err = -EAGAIN;
4507                         goto error;
4508                 }
4509         }
4510
4511
4512         path->leave_spinning = 1;
4513         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4514         if (ret < 0) {
4515                 err = ret;
4516                 goto out;
4517         }
4518
4519         if (ret > 0) {
4520                 /* there are no items in the tree for us to truncate, we're
4521                  * done
4522                  */
4523                 if (path->slots[0] == 0)
4524                         goto out;
4525                 path->slots[0]--;
4526         }
4527
4528         while (1) {
4529                 fi = NULL;
4530                 leaf = path->nodes[0];
4531                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4532                 found_type = found_key.type;
4533
4534                 if (found_key.objectid != ino)
4535                         break;
4536
4537                 if (found_type < min_type)
4538                         break;
4539
4540                 item_end = found_key.offset;
4541                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4542                         fi = btrfs_item_ptr(leaf, path->slots[0],
4543                                             struct btrfs_file_extent_item);
4544                         extent_type = btrfs_file_extent_type(leaf, fi);
4545                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4546                                 item_end +=
4547                                     btrfs_file_extent_num_bytes(leaf, fi);
4548
4549                                 trace_btrfs_truncate_show_fi_regular(
4550                                         BTRFS_I(inode), leaf, fi,
4551                                         found_key.offset);
4552                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4553                                 item_end += btrfs_file_extent_inline_len(leaf,
4554                                                          path->slots[0], fi);
4555
4556                                 trace_btrfs_truncate_show_fi_inline(
4557                                         BTRFS_I(inode), leaf, fi, path->slots[0],
4558                                         found_key.offset);
4559                         }
4560                         item_end--;
4561                 }
4562                 if (found_type > min_type) {
4563                         del_item = 1;
4564                 } else {
4565                         if (item_end < new_size)
4566                                 break;
4567                         if (found_key.offset >= new_size)
4568                                 del_item = 1;
4569                         else
4570                                 del_item = 0;
4571                 }
4572                 found_extent = 0;
4573                 /* FIXME, shrink the extent if the ref count is only 1 */
4574                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4575                         goto delete;
4576
4577                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4578                         u64 num_dec;
4579                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4580                         if (!del_item) {
4581                                 u64 orig_num_bytes =
4582                                         btrfs_file_extent_num_bytes(leaf, fi);
4583                                 extent_num_bytes = ALIGN(new_size -
4584                                                 found_key.offset,
4585                                                 fs_info->sectorsize);
4586                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4587                                                          extent_num_bytes);
4588                                 num_dec = (orig_num_bytes -
4589                                            extent_num_bytes);
4590                                 if (test_bit(BTRFS_ROOT_REF_COWS,
4591                                              &root->state) &&
4592                                     extent_start != 0)
4593                                         inode_sub_bytes(inode, num_dec);
4594                                 btrfs_mark_buffer_dirty(leaf);
4595                         } else {
4596                                 extent_num_bytes =
4597                                         btrfs_file_extent_disk_num_bytes(leaf,
4598                                                                          fi);
4599                                 extent_offset = found_key.offset -
4600                                         btrfs_file_extent_offset(leaf, fi);
4601
4602                                 /* FIXME blocksize != 4096 */
4603                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4604                                 if (extent_start != 0) {
4605                                         found_extent = 1;
4606                                         if (test_bit(BTRFS_ROOT_REF_COWS,
4607                                                      &root->state))
4608                                                 inode_sub_bytes(inode, num_dec);
4609                                 }
4610                         }
4611                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4612                         /*
4613                          * we can't truncate inline items that have had
4614                          * special encodings
4615                          */
4616                         if (!del_item &&
4617                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4618                             btrfs_file_extent_other_encoding(leaf, fi) == 0 &&
4619                             btrfs_file_extent_compression(leaf, fi) == 0) {
4620                                 u32 size = (u32)(new_size - found_key.offset);
4621
4622                                 btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4623                                 size = btrfs_file_extent_calc_inline_size(size);
4624                                 btrfs_truncate_item(root->fs_info, path, size, 1);
4625                         } else if (!del_item) {
4626                                 /*
4627                                  * We have to bail so the last_size is set to
4628                                  * just before this extent.
4629                                  */
4630                                 err = NEED_TRUNCATE_BLOCK;
4631                                 break;
4632                         }
4633
4634                         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4635                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4636                 }
4637 delete:
4638                 if (del_item)
4639                         last_size = found_key.offset;
4640                 else
4641                         last_size = new_size;
4642                 if (del_item) {
4643                         if (!pending_del_nr) {
4644                                 /* no pending yet, add ourselves */
4645                                 pending_del_slot = path->slots[0];
4646                                 pending_del_nr = 1;
4647                         } else if (pending_del_nr &&
4648                                    path->slots[0] + 1 == pending_del_slot) {
4649                                 /* hop on the pending chunk */
4650                                 pending_del_nr++;
4651                                 pending_del_slot = path->slots[0];
4652                         } else {
4653                                 BUG();
4654                         }
4655                 } else {
4656                         break;
4657                 }
4658                 should_throttle = false;
4659
4660                 if (found_extent &&
4661                     (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4662                      root == fs_info->tree_root)) {
4663                         btrfs_set_path_blocking(path);
4664                         bytes_deleted += extent_num_bytes;
4665                         ret = btrfs_free_extent(trans, root, extent_start,
4666                                                 extent_num_bytes, 0,
4667                                                 btrfs_header_owner(leaf),
4668                                                 ino, extent_offset);
4669                         BUG_ON(ret);
4670                         if (btrfs_should_throttle_delayed_refs(trans, fs_info))
4671                                 btrfs_async_run_delayed_refs(fs_info,
4672                                         trans->delayed_ref_updates * 2,
4673                                         trans->transid, 0);
4674                         if (be_nice) {
4675                                 if (truncate_space_check(trans, root,
4676                                                          extent_num_bytes)) {
4677                                         should_end = true;
4678                                 }
4679                                 if (btrfs_should_throttle_delayed_refs(trans,
4680                                                                        fs_info))
4681                                         should_throttle = true;
4682                         }
4683                 }
4684
4685                 if (found_type == BTRFS_INODE_ITEM_KEY)
4686                         break;
4687
4688                 if (path->slots[0] == 0 ||
4689                     path->slots[0] != pending_del_slot ||
4690                     should_throttle || should_end) {
4691                         if (pending_del_nr) {
4692                                 ret = btrfs_del_items(trans, root, path,
4693                                                 pending_del_slot,
4694                                                 pending_del_nr);
4695                                 if (ret) {
4696                                         btrfs_abort_transaction(trans, ret);
4697                                         goto error;
4698                                 }
4699                                 pending_del_nr = 0;
4700                         }
4701                         btrfs_release_path(path);
4702                         if (should_throttle) {
4703                                 unsigned long updates = trans->delayed_ref_updates;
4704                                 if (updates) {
4705                                         trans->delayed_ref_updates = 0;
4706                                         ret = btrfs_run_delayed_refs(trans,
4707                                                                    fs_info,
4708                                                                    updates * 2);
4709                                         if (ret && !err)
4710                                                 err = ret;
4711                                 }
4712                         }
4713                         /*
4714                          * if we failed to refill our space rsv, bail out
4715                          * and let the transaction restart
4716                          */
4717                         if (should_end) {
4718                                 err = -EAGAIN;
4719                                 goto error;
4720                         }
4721                         goto search_again;
4722                 } else {
4723                         path->slots[0]--;
4724                 }
4725         }
4726 out:
4727         if (pending_del_nr) {
4728                 ret = btrfs_del_items(trans, root, path, pending_del_slot,
4729                                       pending_del_nr);
4730                 if (ret)
4731                         btrfs_abort_transaction(trans, ret);
4732         }
4733 error:
4734         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4735                 ASSERT(last_size >= new_size);
4736                 if (!err && last_size > new_size)
4737                         last_size = new_size;
4738                 btrfs_ordered_update_i_size(inode, last_size, NULL);
4739         }
4740
4741         btrfs_free_path(path);
4742
4743         if (be_nice && bytes_deleted > SZ_32M) {
4744                 unsigned long updates = trans->delayed_ref_updates;
4745                 if (updates) {
4746                         trans->delayed_ref_updates = 0;
4747                         ret = btrfs_run_delayed_refs(trans, fs_info,
4748                                                      updates * 2);
4749                         if (ret && !err)
4750                                 err = ret;
4751                 }
4752         }
4753         return err;
4754 }
4755
4756 /*
4757  * btrfs_truncate_block - read, zero a chunk and write a block
4758  * @inode - inode that we're zeroing
4759  * @from - the offset to start zeroing
4760  * @len - the length to zero, 0 to zero the entire range respective to the
4761  *      offset
4762  * @front - zero up to the offset instead of from the offset on
4763  *
4764  * This will find the block for the "from" offset and cow the block and zero the
4765  * part we want to zero.  This is used with truncate and hole punching.
4766  */
4767 int btrfs_truncate_block(struct inode *inode, loff_t from, loff_t len,
4768                         int front)
4769 {
4770         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4771         struct address_space *mapping = inode->i_mapping;
4772         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4773         struct btrfs_ordered_extent *ordered;
4774         struct extent_state *cached_state = NULL;
4775         struct extent_changeset *data_reserved = NULL;
4776         char *kaddr;
4777         u32 blocksize = fs_info->sectorsize;
4778         pgoff_t index = from >> PAGE_SHIFT;
4779         unsigned offset = from & (blocksize - 1);
4780         struct page *page;
4781         gfp_t mask = btrfs_alloc_write_mask(mapping);
4782         int ret = 0;
4783         u64 block_start;
4784         u64 block_end;
4785
4786         if (IS_ALIGNED(offset, blocksize) &&
4787             (!len || IS_ALIGNED(len, blocksize)))
4788                 goto out;
4789
4790         block_start = round_down(from, blocksize);
4791         block_end = block_start + blocksize - 1;
4792
4793         ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
4794                                            block_start, blocksize);
4795         if (ret)
4796                 goto out;
4797
4798 again:
4799         page = find_or_create_page(mapping, index, mask);
4800         if (!page) {
4801                 btrfs_delalloc_release_space(inode, data_reserved,
4802                                              block_start, blocksize);
4803                 btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize);
4804                 ret = -ENOMEM;
4805                 goto out;
4806         }
4807
4808         if (!PageUptodate(page)) {
4809                 ret = btrfs_readpage(NULL, page);
4810                 lock_page(page);
4811                 if (page->mapping != mapping) {
4812                         unlock_page(page);
4813                         put_page(page);
4814                         goto again;
4815                 }
4816                 if (!PageUptodate(page)) {
4817                         ret = -EIO;
4818                         goto out_unlock;
4819                 }
4820         }
4821         wait_on_page_writeback(page);
4822
4823         lock_extent_bits(io_tree, block_start, block_end, &cached_state);
4824         set_page_extent_mapped(page);
4825
4826         ordered = btrfs_lookup_ordered_extent(inode, block_start);
4827         if (ordered) {
4828                 unlock_extent_cached(io_tree, block_start, block_end,
4829                                      &cached_state);
4830                 unlock_page(page);
4831                 put_page(page);
4832                 btrfs_start_ordered_extent(inode, ordered, 1);
4833                 btrfs_put_ordered_extent(ordered);
4834                 goto again;
4835         }
4836
4837         clear_extent_bit(&BTRFS_I(inode)->io_tree, block_start, block_end,
4838                           EXTENT_DIRTY | EXTENT_DELALLOC |
4839                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4840                           0, 0, &cached_state);
4841
4842         ret = btrfs_set_extent_delalloc(inode, block_start, block_end, 0,
4843                                         &cached_state, 0);
4844         if (ret) {
4845                 unlock_extent_cached(io_tree, block_start, block_end,
4846                                      &cached_state);
4847                 goto out_unlock;
4848         }
4849
4850         if (offset != blocksize) {
4851                 if (!len)
4852                         len = blocksize - offset;
4853                 kaddr = kmap(page);
4854                 if (front)
4855                         memset(kaddr + (block_start - page_offset(page)),
4856                                 0, offset);
4857                 else
4858                         memset(kaddr + (block_start - page_offset(page)) +  offset,
4859                                 0, len);
4860                 flush_dcache_page(page);
4861                 kunmap(page);
4862         }
4863         ClearPageChecked(page);
4864         set_page_dirty(page);
4865         unlock_extent_cached(io_tree, block_start, block_end, &cached_state);
4866
4867 out_unlock:
4868         if (ret)
4869                 btrfs_delalloc_release_space(inode, data_reserved, block_start,
4870                                              blocksize);
4871         btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize);
4872         unlock_page(page);
4873         put_page(page);
4874 out:
4875         extent_changeset_free(data_reserved);
4876         return ret;
4877 }
4878
4879 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
4880                              u64 offset, u64 len)
4881 {
4882         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4883         struct btrfs_trans_handle *trans;
4884         int ret;
4885
4886         /*
4887          * Still need to make sure the inode looks like it's been updated so
4888          * that any holes get logged if we fsync.
4889          */
4890         if (btrfs_fs_incompat(fs_info, NO_HOLES)) {
4891                 BTRFS_I(inode)->last_trans = fs_info->generation;
4892                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
4893                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
4894                 return 0;
4895         }
4896
4897         /*
4898          * 1 - for the one we're dropping
4899          * 1 - for the one we're adding
4900          * 1 - for updating the inode.
4901          */
4902         trans = btrfs_start_transaction(root, 3);
4903         if (IS_ERR(trans))
4904                 return PTR_ERR(trans);
4905
4906         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
4907         if (ret) {
4908                 btrfs_abort_transaction(trans, ret);
4909                 btrfs_end_transaction(trans);
4910                 return ret;
4911         }
4912
4913         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(BTRFS_I(inode)),
4914                         offset, 0, 0, len, 0, len, 0, 0, 0);
4915         if (ret)
4916                 btrfs_abort_transaction(trans, ret);
4917         else
4918                 btrfs_update_inode(trans, root, inode);
4919         btrfs_end_transaction(trans);
4920         return ret;
4921 }
4922
4923 /*
4924  * This function puts in dummy file extents for the area we're creating a hole
4925  * for.  So if we are truncating this file to a larger size we need to insert
4926  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
4927  * the range between oldsize and size
4928  */
4929 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
4930 {
4931         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4932         struct btrfs_root *root = BTRFS_I(inode)->root;
4933         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4934         struct extent_map *em = NULL;
4935         struct extent_state *cached_state = NULL;
4936         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
4937         u64 hole_start = ALIGN(oldsize, fs_info->sectorsize);
4938         u64 block_end = ALIGN(size, fs_info->sectorsize);
4939         u64 last_byte;
4940         u64 cur_offset;
4941         u64 hole_size;
4942         int err = 0;
4943
4944         /*
4945          * If our size started in the middle of a block we need to zero out the
4946          * rest of the block before we expand the i_size, otherwise we could
4947          * expose stale data.
4948          */
4949         err = btrfs_truncate_block(inode, oldsize, 0, 0);
4950         if (err)
4951                 return err;
4952
4953         if (size <= hole_start)
4954                 return 0;
4955
4956         while (1) {
4957                 struct btrfs_ordered_extent *ordered;
4958
4959                 lock_extent_bits(io_tree, hole_start, block_end - 1,
4960                                  &cached_state);
4961                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), hole_start,
4962                                                      block_end - hole_start);
4963                 if (!ordered)
4964                         break;
4965                 unlock_extent_cached(io_tree, hole_start, block_end - 1,
4966                                      &cached_state);
4967                 btrfs_start_ordered_extent(inode, ordered, 1);
4968                 btrfs_put_ordered_extent(ordered);
4969         }
4970
4971         cur_offset = hole_start;
4972         while (1) {
4973                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
4974                                 block_end - cur_offset, 0);
4975                 if (IS_ERR(em)) {
4976                         err = PTR_ERR(em);
4977                         em = NULL;
4978                         break;
4979                 }
4980                 last_byte = min(extent_map_end(em), block_end);
4981                 last_byte = ALIGN(last_byte, fs_info->sectorsize);
4982                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
4983                         struct extent_map *hole_em;
4984                         hole_size = last_byte - cur_offset;
4985
4986                         err = maybe_insert_hole(root, inode, cur_offset,
4987                                                 hole_size);
4988                         if (err)
4989                                 break;
4990                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
4991                                                 cur_offset + hole_size - 1, 0);
4992                         hole_em = alloc_extent_map();
4993                         if (!hole_em) {
4994                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
4995                                         &BTRFS_I(inode)->runtime_flags);
4996                                 goto next;
4997                         }
4998                         hole_em->start = cur_offset;
4999                         hole_em->len = hole_size;
5000                         hole_em->orig_start = cur_offset;
5001
5002                         hole_em->block_start = EXTENT_MAP_HOLE;
5003                         hole_em->block_len = 0;
5004                         hole_em->orig_block_len = 0;
5005                         hole_em->ram_bytes = hole_size;
5006                         hole_em->bdev = fs_info->fs_devices->latest_bdev;
5007                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
5008                         hole_em->generation = fs_info->generation;
5009
5010                         while (1) {
5011                                 write_lock(&em_tree->lock);
5012                                 err = add_extent_mapping(em_tree, hole_em, 1);
5013                                 write_unlock(&em_tree->lock);
5014                                 if (err != -EEXIST)
5015                                         break;
5016                                 btrfs_drop_extent_cache(BTRFS_I(inode),
5017                                                         cur_offset,
5018                                                         cur_offset +
5019                                                         hole_size - 1, 0);
5020                         }
5021                         free_extent_map(hole_em);
5022                 }
5023 next:
5024                 free_extent_map(em);
5025                 em = NULL;
5026                 cur_offset = last_byte;
5027                 if (cur_offset >= block_end)
5028                         break;
5029         }
5030         free_extent_map(em);
5031         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state);
5032         return err;
5033 }
5034
5035 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
5036 {
5037         struct btrfs_root *root = BTRFS_I(inode)->root;
5038         struct btrfs_trans_handle *trans;
5039         loff_t oldsize = i_size_read(inode);
5040         loff_t newsize = attr->ia_size;
5041         int mask = attr->ia_valid;
5042         int ret;
5043
5044         /*
5045          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
5046          * special case where we need to update the times despite not having
5047          * these flags set.  For all other operations the VFS set these flags
5048          * explicitly if it wants a timestamp update.
5049          */
5050         if (newsize != oldsize) {
5051                 inode_inc_iversion(inode);
5052                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
5053                         inode->i_ctime = inode->i_mtime =
5054                                 current_time(inode);
5055         }
5056
5057         if (newsize > oldsize) {
5058                 /*
5059                  * Don't do an expanding truncate while snapshotting is ongoing.
5060                  * This is to ensure the snapshot captures a fully consistent
5061                  * state of this file - if the snapshot captures this expanding
5062                  * truncation, it must capture all writes that happened before
5063                  * this truncation.
5064                  */
5065                 btrfs_wait_for_snapshot_creation(root);
5066                 ret = btrfs_cont_expand(inode, oldsize, newsize);
5067                 if (ret) {
5068                         btrfs_end_write_no_snapshotting(root);
5069                         return ret;
5070                 }
5071
5072                 trans = btrfs_start_transaction(root, 1);
5073                 if (IS_ERR(trans)) {
5074                         btrfs_end_write_no_snapshotting(root);
5075                         return PTR_ERR(trans);
5076                 }
5077
5078                 i_size_write(inode, newsize);
5079                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
5080                 pagecache_isize_extended(inode, oldsize, newsize);
5081                 ret = btrfs_update_inode(trans, root, inode);
5082                 btrfs_end_write_no_snapshotting(root);
5083                 btrfs_end_transaction(trans);
5084         } else {
5085
5086                 /*
5087                  * We're truncating a file that used to have good data down to
5088                  * zero. Make sure it gets into the ordered flush list so that
5089                  * any new writes get down to disk quickly.
5090                  */
5091                 if (newsize == 0)
5092                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
5093                                 &BTRFS_I(inode)->runtime_flags);
5094
5095                 /*
5096                  * 1 for the orphan item we're going to add
5097                  * 1 for the orphan item deletion.
5098                  */
5099                 trans = btrfs_start_transaction(root, 2);
5100                 if (IS_ERR(trans))
5101                         return PTR_ERR(trans);
5102
5103                 /*
5104                  * We need to do this in case we fail at _any_ point during the
5105                  * actual truncate.  Once we do the truncate_setsize we could
5106                  * invalidate pages which forces any outstanding ordered io to
5107                  * be instantly completed which will give us extents that need
5108                  * to be truncated.  If we fail to get an orphan inode down we
5109                  * could have left over extents that were never meant to live,
5110                  * so we need to guarantee from this point on that everything
5111                  * will be consistent.
5112                  */
5113                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
5114                 btrfs_end_transaction(trans);
5115                 if (ret)
5116                         return ret;
5117
5118                 /* we don't support swapfiles, so vmtruncate shouldn't fail */
5119                 truncate_setsize(inode, newsize);
5120
5121                 /* Disable nonlocked read DIO to avoid the end less truncate */
5122                 btrfs_inode_block_unlocked_dio(BTRFS_I(inode));
5123                 inode_dio_wait(inode);
5124                 btrfs_inode_resume_unlocked_dio(BTRFS_I(inode));
5125
5126                 ret = btrfs_truncate(inode);
5127                 if (ret && inode->i_nlink) {
5128                         int err;
5129
5130                         /* To get a stable disk_i_size */
5131                         err = btrfs_wait_ordered_range(inode, 0, (u64)-1);
5132                         if (err) {
5133                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5134                                 return err;
5135                         }
5136
5137                         /*
5138                          * failed to truncate, disk_i_size is only adjusted down
5139                          * as we remove extents, so it should represent the true
5140                          * size of the inode, so reset the in memory size and
5141                          * delete our orphan entry.
5142                          */
5143                         trans = btrfs_join_transaction(root);
5144                         if (IS_ERR(trans)) {
5145                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5146                                 return ret;
5147                         }
5148                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
5149                         err = btrfs_orphan_del(trans, BTRFS_I(inode));
5150                         if (err)
5151                                 btrfs_abort_transaction(trans, err);
5152                         btrfs_end_transaction(trans);
5153                 }
5154         }
5155
5156         return ret;
5157 }
5158
5159 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
5160 {
5161         struct inode *inode = d_inode(dentry);
5162         struct btrfs_root *root = BTRFS_I(inode)->root;
5163         int err;
5164
5165         if (btrfs_root_readonly(root))
5166                 return -EROFS;
5167
5168         err = setattr_prepare(dentry, attr);
5169         if (err)
5170                 return err;
5171
5172         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
5173                 err = btrfs_setsize(inode, attr);
5174                 if (err)
5175                         return err;
5176         }
5177
5178         if (attr->ia_valid) {
5179                 setattr_copy(inode, attr);
5180                 inode_inc_iversion(inode);
5181                 err = btrfs_dirty_inode(inode);
5182
5183                 if (!err && attr->ia_valid & ATTR_MODE)
5184                         err = posix_acl_chmod(inode, inode->i_mode);
5185         }
5186
5187         return err;
5188 }
5189
5190 /*
5191  * While truncating the inode pages during eviction, we get the VFS calling
5192  * btrfs_invalidatepage() against each page of the inode. This is slow because
5193  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
5194  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
5195  * extent_state structures over and over, wasting lots of time.
5196  *
5197  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
5198  * those expensive operations on a per page basis and do only the ordered io
5199  * finishing, while we release here the extent_map and extent_state structures,
5200  * without the excessive merging and splitting.
5201  */
5202 static void evict_inode_truncate_pages(struct inode *inode)
5203 {
5204         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5205         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
5206         struct rb_node *node;
5207
5208         ASSERT(inode->i_state & I_FREEING);
5209         truncate_inode_pages_final(&inode->i_data);
5210
5211         write_lock(&map_tree->lock);
5212         while (!RB_EMPTY_ROOT(&map_tree->map)) {
5213                 struct extent_map *em;
5214
5215                 node = rb_first(&map_tree->map);
5216                 em = rb_entry(node, struct extent_map, rb_node);
5217                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
5218                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
5219                 remove_extent_mapping(map_tree, em);
5220                 free_extent_map(em);
5221                 if (need_resched()) {
5222                         write_unlock(&map_tree->lock);
5223                         cond_resched();
5224                         write_lock(&map_tree->lock);
5225                 }
5226         }
5227         write_unlock(&map_tree->lock);
5228
5229         /*
5230          * Keep looping until we have no more ranges in the io tree.
5231          * We can have ongoing bios started by readpages (called from readahead)
5232          * that have their endio callback (extent_io.c:end_bio_extent_readpage)
5233          * still in progress (unlocked the pages in the bio but did not yet
5234          * unlocked the ranges in the io tree). Therefore this means some
5235          * ranges can still be locked and eviction started because before
5236          * submitting those bios, which are executed by a separate task (work
5237          * queue kthread), inode references (inode->i_count) were not taken
5238          * (which would be dropped in the end io callback of each bio).
5239          * Therefore here we effectively end up waiting for those bios and
5240          * anyone else holding locked ranges without having bumped the inode's
5241          * reference count - if we don't do it, when they access the inode's
5242          * io_tree to unlock a range it may be too late, leading to an
5243          * use-after-free issue.
5244          */
5245         spin_lock(&io_tree->lock);
5246         while (!RB_EMPTY_ROOT(&io_tree->state)) {
5247                 struct extent_state *state;
5248                 struct extent_state *cached_state = NULL;
5249                 u64 start;
5250                 u64 end;
5251
5252                 node = rb_first(&io_tree->state);
5253                 state = rb_entry(node, struct extent_state, rb_node);
5254                 start = state->start;
5255                 end = state->end;
5256                 spin_unlock(&io_tree->lock);
5257
5258                 lock_extent_bits(io_tree, start, end, &cached_state);
5259
5260                 /*
5261                  * If still has DELALLOC flag, the extent didn't reach disk,
5262                  * and its reserved space won't be freed by delayed_ref.
5263                  * So we need to free its reserved space here.
5264                  * (Refer to comment in btrfs_invalidatepage, case 2)
5265                  *
5266                  * Note, end is the bytenr of last byte, so we need + 1 here.
5267                  */
5268                 if (state->state & EXTENT_DELALLOC)
5269                         btrfs_qgroup_free_data(inode, NULL, start, end - start + 1);
5270
5271                 clear_extent_bit(io_tree, start, end,
5272                                  EXTENT_LOCKED | EXTENT_DIRTY |
5273                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
5274                                  EXTENT_DEFRAG, 1, 1, &cached_state);
5275
5276                 cond_resched();
5277                 spin_lock(&io_tree->lock);
5278         }
5279         spin_unlock(&io_tree->lock);
5280 }
5281
5282 void btrfs_evict_inode(struct inode *inode)
5283 {
5284         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5285         struct btrfs_trans_handle *trans;
5286         struct btrfs_root *root = BTRFS_I(inode)->root;
5287         struct btrfs_block_rsv *rsv, *global_rsv;
5288         int steal_from_global = 0;
5289         u64 min_size;
5290         int ret;
5291
5292         trace_btrfs_inode_evict(inode);
5293
5294         if (!root) {
5295                 clear_inode(inode);
5296                 return;
5297         }
5298
5299         min_size = btrfs_calc_trunc_metadata_size(fs_info, 1);
5300
5301         evict_inode_truncate_pages(inode);
5302
5303         if (inode->i_nlink &&
5304             ((btrfs_root_refs(&root->root_item) != 0 &&
5305               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5306              btrfs_is_free_space_inode(BTRFS_I(inode))))
5307                 goto no_delete;
5308
5309         if (is_bad_inode(inode)) {
5310                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5311                 goto no_delete;
5312         }
5313         /* do we really want it for ->i_nlink > 0 and zero btrfs_root_refs? */
5314         if (!special_file(inode->i_mode))
5315                 btrfs_wait_ordered_range(inode, 0, (u64)-1);
5316
5317         btrfs_free_io_failure_record(BTRFS_I(inode), 0, (u64)-1);
5318
5319         if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags)) {
5320                 BUG_ON(test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
5321                                  &BTRFS_I(inode)->runtime_flags));
5322                 goto no_delete;
5323         }
5324
5325         if (inode->i_nlink > 0) {
5326                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5327                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5328                 goto no_delete;
5329         }
5330
5331         ret = btrfs_commit_inode_delayed_inode(BTRFS_I(inode));
5332         if (ret) {
5333                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5334                 goto no_delete;
5335         }
5336
5337         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
5338         if (!rsv) {
5339                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5340                 goto no_delete;
5341         }
5342         rsv->size = min_size;
5343         rsv->failfast = 1;
5344         global_rsv = &fs_info->global_block_rsv;
5345
5346         btrfs_i_size_write(BTRFS_I(inode), 0);
5347
5348         /*
5349          * This is a bit simpler than btrfs_truncate since we've already
5350          * reserved our space for our orphan item in the unlink, so we just
5351          * need to reserve some slack space in case we add bytes and update
5352          * inode item when doing the truncate.
5353          */
5354         while (1) {
5355                 ret = btrfs_block_rsv_refill(root, rsv, min_size,
5356                                              BTRFS_RESERVE_FLUSH_LIMIT);
5357
5358                 /*
5359                  * Try and steal from the global reserve since we will
5360                  * likely not use this space anyway, we want to try as
5361                  * hard as possible to get this to work.
5362                  */
5363                 if (ret)
5364                         steal_from_global++;
5365                 else
5366                         steal_from_global = 0;
5367                 ret = 0;
5368
5369                 /*
5370                  * steal_from_global == 0: we reserved stuff, hooray!
5371                  * steal_from_global == 1: we didn't reserve stuff, boo!
5372                  * steal_from_global == 2: we've committed, still not a lot of
5373                  * room but maybe we'll have room in the global reserve this
5374                  * time.
5375                  * steal_from_global == 3: abandon all hope!
5376                  */
5377                 if (steal_from_global > 2) {
5378                         btrfs_warn(fs_info,
5379                                    "Could not get space for a delete, will truncate on mount %d",
5380                                    ret);
5381                         btrfs_orphan_del(NULL, BTRFS_I(inode));
5382                         btrfs_free_block_rsv(fs_info, rsv);
5383                         goto no_delete;
5384                 }
5385
5386                 trans = btrfs_join_transaction(root);
5387                 if (IS_ERR(trans)) {
5388                         btrfs_orphan_del(NULL, BTRFS_I(inode));
5389                         btrfs_free_block_rsv(fs_info, rsv);
5390                         goto no_delete;
5391                 }
5392
5393                 /*
5394                  * We can't just steal from the global reserve, we need to make
5395                  * sure there is room to do it, if not we need to commit and try
5396                  * again.
5397                  */
5398                 if (steal_from_global) {
5399                         if (!btrfs_check_space_for_delayed_refs(trans, fs_info))
5400                                 ret = btrfs_block_rsv_migrate(global_rsv, rsv,
5401                                                               min_size, 0);
5402                         else
5403                                 ret = -ENOSPC;
5404                 }
5405
5406                 /*
5407                  * Couldn't steal from the global reserve, we have too much
5408                  * pending stuff built up, commit the transaction and try it
5409                  * again.
5410                  */
5411                 if (ret) {
5412                         ret = btrfs_commit_transaction(trans);
5413                         if (ret) {
5414                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5415                                 btrfs_free_block_rsv(fs_info, rsv);
5416                                 goto no_delete;
5417                         }
5418                         continue;
5419                 } else {
5420                         steal_from_global = 0;
5421                 }
5422
5423                 trans->block_rsv = rsv;
5424
5425                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5426                 if (ret != -ENOSPC && ret != -EAGAIN)
5427                         break;
5428
5429                 trans->block_rsv = &fs_info->trans_block_rsv;
5430                 btrfs_end_transaction(trans);
5431                 trans = NULL;
5432                 btrfs_btree_balance_dirty(fs_info);
5433         }
5434
5435         btrfs_free_block_rsv(fs_info, rsv);
5436
5437         /*
5438          * Errors here aren't a big deal, it just means we leave orphan items
5439          * in the tree.  They will be cleaned up on the next mount.
5440          */
5441         if (ret == 0) {
5442                 trans->block_rsv = root->orphan_block_rsv;
5443                 btrfs_orphan_del(trans, BTRFS_I(inode));
5444         } else {
5445                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5446         }
5447
5448         trans->block_rsv = &fs_info->trans_block_rsv;
5449         if (!(root == fs_info->tree_root ||
5450               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5451                 btrfs_return_ino(root, btrfs_ino(BTRFS_I(inode)));
5452
5453         btrfs_end_transaction(trans);
5454         btrfs_btree_balance_dirty(fs_info);
5455 no_delete:
5456         btrfs_remove_delayed_node(BTRFS_I(inode));
5457         clear_inode(inode);
5458 }
5459
5460 /*
5461  * this returns the key found in the dir entry in the location pointer.
5462  * If no dir entries were found, location->objectid is 0.
5463  */
5464 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5465                                struct btrfs_key *location)
5466 {
5467         const char *name = dentry->d_name.name;
5468         int namelen = dentry->d_name.len;
5469         struct btrfs_dir_item *di;
5470         struct btrfs_path *path;
5471         struct btrfs_root *root = BTRFS_I(dir)->root;
5472         int ret = 0;
5473
5474         path = btrfs_alloc_path();
5475         if (!path)
5476                 return -ENOMEM;
5477
5478         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(BTRFS_I(dir)),
5479                         name, namelen, 0);
5480         if (IS_ERR(di))
5481                 ret = PTR_ERR(di);
5482
5483         if (IS_ERR_OR_NULL(di))
5484                 goto out_err;
5485
5486         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5487         if (location->type != BTRFS_INODE_ITEM_KEY &&
5488             location->type != BTRFS_ROOT_ITEM_KEY) {
5489                 btrfs_warn(root->fs_info,
5490 "%s gets something invalid in DIR_ITEM (name %s, directory ino %llu, location(%llu %u %llu))",
5491                            __func__, name, btrfs_ino(BTRFS_I(dir)),
5492                            location->objectid, location->type, location->offset);
5493                 goto out_err;
5494         }
5495 out:
5496         btrfs_free_path(path);
5497         return ret;
5498 out_err:
5499         location->objectid = 0;
5500         goto out;
5501 }
5502
5503 /*
5504  * when we hit a tree root in a directory, the btrfs part of the inode
5505  * needs to be changed to reflect the root directory of the tree root.  This
5506  * is kind of like crossing a mount point.
5507  */
5508 static int fixup_tree_root_location(struct btrfs_fs_info *fs_info,
5509                                     struct inode *dir,
5510                                     struct dentry *dentry,
5511                                     struct btrfs_key *location,
5512                                     struct btrfs_root **sub_root)
5513 {
5514         struct btrfs_path *path;
5515         struct btrfs_root *new_root;
5516         struct btrfs_root_ref *ref;
5517         struct extent_buffer *leaf;
5518         struct btrfs_key key;
5519         int ret;
5520         int err = 0;
5521
5522         path = btrfs_alloc_path();
5523         if (!path) {
5524                 err = -ENOMEM;
5525                 goto out;
5526         }
5527
5528         err = -ENOENT;
5529         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5530         key.type = BTRFS_ROOT_REF_KEY;
5531         key.offset = location->objectid;
5532
5533         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
5534         if (ret) {
5535                 if (ret < 0)
5536                         err = ret;
5537                 goto out;
5538         }
5539
5540         leaf = path->nodes[0];
5541         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5542         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(BTRFS_I(dir)) ||
5543             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5544                 goto out;
5545
5546         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5547                                    (unsigned long)(ref + 1),
5548                                    dentry->d_name.len);
5549         if (ret)
5550                 goto out;
5551
5552         btrfs_release_path(path);
5553
5554         new_root = btrfs_read_fs_root_no_name(fs_info, location);
5555         if (IS_ERR(new_root)) {
5556                 err = PTR_ERR(new_root);
5557                 goto out;
5558         }
5559
5560         *sub_root = new_root;
5561         location->objectid = btrfs_root_dirid(&new_root->root_item);
5562         location->type = BTRFS_INODE_ITEM_KEY;
5563         location->offset = 0;
5564         err = 0;
5565 out:
5566         btrfs_free_path(path);
5567         return err;
5568 }
5569
5570 static void inode_tree_add(struct inode *inode)
5571 {
5572         struct btrfs_root *root = BTRFS_I(inode)->root;
5573         struct btrfs_inode *entry;
5574         struct rb_node **p;
5575         struct rb_node *parent;
5576         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5577         u64 ino = btrfs_ino(BTRFS_I(inode));
5578
5579         if (inode_unhashed(inode))
5580                 return;
5581         parent = NULL;
5582         spin_lock(&root->inode_lock);
5583         p = &root->inode_tree.rb_node;
5584         while (*p) {
5585                 parent = *p;
5586                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5587
5588                 if (ino < btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5589                         p = &parent->rb_left;
5590                 else if (ino > btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5591                         p = &parent->rb_right;
5592                 else {
5593                         WARN_ON(!(entry->vfs_inode.i_state &
5594                                   (I_WILL_FREE | I_FREEING)));
5595                         rb_replace_node(parent, new, &root->inode_tree);
5596                         RB_CLEAR_NODE(parent);
5597                         spin_unlock(&root->inode_lock);
5598                         return;
5599                 }
5600         }
5601         rb_link_node(new, parent, p);
5602         rb_insert_color(new, &root->inode_tree);
5603         spin_unlock(&root->inode_lock);
5604 }
5605
5606 static void inode_tree_del(struct inode *inode)
5607 {
5608         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5609         struct btrfs_root *root = BTRFS_I(inode)->root;
5610         int empty = 0;
5611
5612         spin_lock(&root->inode_lock);
5613         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5614                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5615                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5616                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5617         }
5618         spin_unlock(&root->inode_lock);
5619
5620         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5621                 synchronize_srcu(&fs_info->subvol_srcu);
5622                 spin_lock(&root->inode_lock);
5623                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5624                 spin_unlock(&root->inode_lock);
5625                 if (empty)
5626                         btrfs_add_dead_root(root);
5627         }
5628 }
5629
5630 void btrfs_invalidate_inodes(struct btrfs_root *root)
5631 {
5632         struct btrfs_fs_info *fs_info = root->fs_info;
5633         struct rb_node *node;
5634         struct rb_node *prev;
5635         struct btrfs_inode *entry;
5636         struct inode *inode;
5637         u64 objectid = 0;
5638
5639         if (!test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
5640                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
5641
5642         spin_lock(&root->inode_lock);
5643 again:
5644         node = root->inode_tree.rb_node;
5645         prev = NULL;
5646         while (node) {
5647                 prev = node;
5648                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5649
5650                 if (objectid < btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5651                         node = node->rb_left;
5652                 else if (objectid > btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5653                         node = node->rb_right;
5654                 else
5655                         break;
5656         }
5657         if (!node) {
5658                 while (prev) {
5659                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
5660                         if (objectid <= btrfs_ino(BTRFS_I(&entry->vfs_inode))) {
5661                                 node = prev;
5662                                 break;
5663                         }
5664                         prev = rb_next(prev);
5665                 }
5666         }
5667         while (node) {
5668                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5669                 objectid = btrfs_ino(BTRFS_I(&entry->vfs_inode)) + 1;
5670                 inode = igrab(&entry->vfs_inode);
5671                 if (inode) {
5672                         spin_unlock(&root->inode_lock);
5673                         if (atomic_read(&inode->i_count) > 1)
5674                                 d_prune_aliases(inode);
5675                         /*
5676                          * btrfs_drop_inode will have it removed from
5677                          * the inode cache when its usage count
5678                          * hits zero.
5679                          */
5680                         iput(inode);
5681                         cond_resched();
5682                         spin_lock(&root->inode_lock);
5683                         goto again;
5684                 }
5685
5686                 if (cond_resched_lock(&root->inode_lock))
5687                         goto again;
5688
5689                 node = rb_next(node);
5690         }
5691         spin_unlock(&root->inode_lock);
5692 }
5693
5694 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5695 {
5696         struct btrfs_iget_args *args = p;
5697         inode->i_ino = args->location->objectid;
5698         memcpy(&BTRFS_I(inode)->location, args->location,
5699                sizeof(*args->location));
5700         BTRFS_I(inode)->root = args->root;
5701         return 0;
5702 }
5703
5704 static int btrfs_find_actor(struct inode *inode, void *opaque)
5705 {
5706         struct btrfs_iget_args *args = opaque;
5707         return args->location->objectid == BTRFS_I(inode)->location.objectid &&
5708                 args->root == BTRFS_I(inode)->root;
5709 }
5710
5711 static struct inode *btrfs_iget_locked(struct super_block *s,
5712                                        struct btrfs_key *location,
5713                                        struct btrfs_root *root)
5714 {
5715         struct inode *inode;
5716         struct btrfs_iget_args args;
5717         unsigned long hashval = btrfs_inode_hash(location->objectid, root);
5718
5719         args.location = location;
5720         args.root = root;
5721
5722         inode = iget5_locked(s, hashval, btrfs_find_actor,
5723                              btrfs_init_locked_inode,
5724                              (void *)&args);
5725         return inode;
5726 }
5727
5728 /* Get an inode object given its location and corresponding root.
5729  * Returns in *is_new if the inode was read from disk
5730  */
5731 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
5732                          struct btrfs_root *root, int *new)
5733 {
5734         struct inode *inode;
5735
5736         inode = btrfs_iget_locked(s, location, root);
5737         if (!inode)
5738                 return ERR_PTR(-ENOMEM);
5739
5740         if (inode->i_state & I_NEW) {
5741                 int ret;
5742
5743                 ret = btrfs_read_locked_inode(inode);
5744                 if (!is_bad_inode(inode)) {
5745                         inode_tree_add(inode);
5746                         unlock_new_inode(inode);
5747                         if (new)
5748                                 *new = 1;
5749                 } else {
5750                         unlock_new_inode(inode);
5751                         iput(inode);
5752                         ASSERT(ret < 0);
5753                         inode = ERR_PTR(ret < 0 ? ret : -ESTALE);
5754                 }
5755         }
5756
5757         return inode;
5758 }
5759
5760 static struct inode *new_simple_dir(struct super_block *s,
5761                                     struct btrfs_key *key,
5762                                     struct btrfs_root *root)
5763 {
5764         struct inode *inode = new_inode(s);
5765
5766         if (!inode)
5767                 return ERR_PTR(-ENOMEM);
5768
5769         BTRFS_I(inode)->root = root;
5770         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5771         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5772
5773         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5774         inode->i_op = &btrfs_dir_ro_inode_operations;
5775         inode->i_opflags &= ~IOP_XATTR;
5776         inode->i_fop = &simple_dir_operations;
5777         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5778         inode->i_mtime = current_time(inode);
5779         inode->i_atime = inode->i_mtime;
5780         inode->i_ctime = inode->i_mtime;
5781         BTRFS_I(inode)->i_otime = inode->i_mtime;
5782
5783         return inode;
5784 }
5785
5786 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5787 {
5788         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
5789         struct inode *inode;
5790         struct btrfs_root *root = BTRFS_I(dir)->root;
5791         struct btrfs_root *sub_root = root;
5792         struct btrfs_key location;
5793         int index;
5794         int ret = 0;
5795
5796         if (dentry->d_name.len > BTRFS_NAME_LEN)
5797                 return ERR_PTR(-ENAMETOOLONG);
5798
5799         ret = btrfs_inode_by_name(dir, dentry, &location);
5800         if (ret < 0)
5801                 return ERR_PTR(ret);
5802
5803         if (location.objectid == 0)
5804                 return ERR_PTR(-ENOENT);
5805
5806         if (location.type == BTRFS_INODE_ITEM_KEY) {
5807                 inode = btrfs_iget(dir->i_sb, &location, root, NULL);
5808                 return inode;
5809         }
5810
5811         index = srcu_read_lock(&fs_info->subvol_srcu);
5812         ret = fixup_tree_root_location(fs_info, dir, dentry,
5813                                        &location, &sub_root);
5814         if (ret < 0) {
5815                 if (ret != -ENOENT)
5816                         inode = ERR_PTR(ret);
5817                 else
5818                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5819         } else {
5820                 inode = btrfs_iget(dir->i_sb, &location, sub_root, NULL);
5821         }
5822         srcu_read_unlock(&fs_info->subvol_srcu, index);
5823
5824         if (!IS_ERR(inode) && root != sub_root) {
5825                 down_read(&fs_info->cleanup_work_sem);
5826                 if (!sb_rdonly(inode->i_sb))
5827                         ret = btrfs_orphan_cleanup(sub_root);
5828                 up_read(&fs_info->cleanup_work_sem);
5829                 if (ret) {
5830                         iput(inode);
5831                         inode = ERR_PTR(ret);
5832                 }
5833         }
5834
5835         return inode;
5836 }
5837
5838 static int btrfs_dentry_delete(const struct dentry *dentry)
5839 {
5840         struct btrfs_root *root;
5841         struct inode *inode = d_inode(dentry);
5842
5843         if (!inode && !IS_ROOT(dentry))
5844                 inode = d_inode(dentry->d_parent);
5845
5846         if (inode) {
5847                 root = BTRFS_I(inode)->root;
5848                 if (btrfs_root_refs(&root->root_item) == 0)
5849                         return 1;
5850
5851                 if (btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5852                         return 1;
5853         }
5854         return 0;
5855 }
5856
5857 static void btrfs_dentry_release(struct dentry *dentry)
5858 {
5859         kfree(dentry->d_fsdata);
5860 }
5861
5862 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5863                                    unsigned int flags)
5864 {
5865         struct inode *inode;
5866
5867         inode = btrfs_lookup_dentry(dir, dentry);
5868         if (IS_ERR(inode)) {
5869                 if (PTR_ERR(inode) == -ENOENT)
5870                         inode = NULL;
5871                 else
5872                         return ERR_CAST(inode);
5873         }
5874
5875         return d_splice_alias(inode, dentry);
5876 }
5877
5878 unsigned char btrfs_filetype_table[] = {
5879         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
5880 };
5881
5882 /*
5883  * All this infrastructure exists because dir_emit can fault, and we are holding
5884  * the tree lock when doing readdir.  For now just allocate a buffer and copy
5885  * our information into that, and then dir_emit from the buffer.  This is
5886  * similar to what NFS does, only we don't keep the buffer around in pagecache
5887  * because I'm afraid I'll mess that up.  Long term we need to make filldir do
5888  * copy_to_user_inatomic so we don't have to worry about page faulting under the
5889  * tree lock.
5890  */
5891 static int btrfs_opendir(struct inode *inode, struct file *file)
5892 {
5893         struct btrfs_file_private *private;
5894
5895         private = kzalloc(sizeof(struct btrfs_file_private), GFP_KERNEL);
5896         if (!private)
5897                 return -ENOMEM;
5898         private->filldir_buf = kzalloc(PAGE_SIZE, GFP_KERNEL);
5899         if (!private->filldir_buf) {
5900                 kfree(private);
5901                 return -ENOMEM;
5902         }
5903         file->private_data = private;
5904         return 0;
5905 }
5906
5907 struct dir_entry {
5908         u64 ino;
5909         u64 offset;
5910         unsigned type;
5911         int name_len;
5912 };
5913
5914 static int btrfs_filldir(void *addr, int entries, struct dir_context *ctx)
5915 {
5916         while (entries--) {
5917                 struct dir_entry *entry = addr;
5918                 char *name = (char *)(entry + 1);
5919
5920                 ctx->pos = entry->offset;
5921                 if (!dir_emit(ctx, name, entry->name_len, entry->ino,
5922                               entry->type))
5923                         return 1;
5924                 addr += sizeof(struct dir_entry) + entry->name_len;
5925                 ctx->pos++;
5926         }
5927         return 0;
5928 }
5929
5930 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
5931 {
5932         struct inode *inode = file_inode(file);
5933         struct btrfs_root *root = BTRFS_I(inode)->root;
5934         struct btrfs_file_private *private = file->private_data;
5935         struct btrfs_dir_item *di;
5936         struct btrfs_key key;
5937         struct btrfs_key found_key;
5938         struct btrfs_path *path;
5939         void *addr;
5940         struct list_head ins_list;
5941         struct list_head del_list;
5942         int ret;
5943         struct extent_buffer *leaf;
5944         int slot;
5945         char *name_ptr;
5946         int name_len;
5947         int entries = 0;
5948         int total_len = 0;
5949         bool put = false;
5950         struct btrfs_key location;
5951
5952         if (!dir_emit_dots(file, ctx))
5953                 return 0;
5954
5955         path = btrfs_alloc_path();
5956         if (!path)
5957                 return -ENOMEM;
5958
5959         addr = private->filldir_buf;
5960         path->reada = READA_FORWARD;
5961
5962         INIT_LIST_HEAD(&ins_list);
5963         INIT_LIST_HEAD(&del_list);
5964         put = btrfs_readdir_get_delayed_items(inode, &ins_list, &del_list);
5965
5966 again:
5967         key.type = BTRFS_DIR_INDEX_KEY;
5968         key.offset = ctx->pos;
5969         key.objectid = btrfs_ino(BTRFS_I(inode));
5970
5971         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5972         if (ret < 0)
5973                 goto err;
5974
5975         while (1) {
5976                 struct dir_entry *entry;
5977
5978                 leaf = path->nodes[0];
5979                 slot = path->slots[0];
5980                 if (slot >= btrfs_header_nritems(leaf)) {
5981                         ret = btrfs_next_leaf(root, path);
5982                         if (ret < 0)
5983                                 goto err;
5984                         else if (ret > 0)
5985                                 break;
5986                         continue;
5987                 }
5988
5989                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
5990
5991                 if (found_key.objectid != key.objectid)
5992                         break;
5993                 if (found_key.type != BTRFS_DIR_INDEX_KEY)
5994                         break;
5995                 if (found_key.offset < ctx->pos)
5996                         goto next;
5997                 if (btrfs_should_delete_dir_index(&del_list, found_key.offset))
5998                         goto next;
5999                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
6000                 name_len = btrfs_dir_name_len(leaf, di);
6001                 if ((total_len + sizeof(struct dir_entry) + name_len) >=
6002                     PAGE_SIZE) {
6003                         btrfs_release_path(path);
6004                         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
6005                         if (ret)
6006                                 goto nopos;
6007                         addr = private->filldir_buf;
6008                         entries = 0;
6009                         total_len = 0;
6010                         goto again;
6011                 }
6012
6013                 entry = addr;
6014                 entry->name_len = name_len;
6015                 name_ptr = (char *)(entry + 1);
6016                 read_extent_buffer(leaf, name_ptr, (unsigned long)(di + 1),
6017                                    name_len);
6018                 entry->type = btrfs_filetype_table[btrfs_dir_type(leaf, di)];
6019                 btrfs_dir_item_key_to_cpu(leaf, di, &location);
6020                 entry->ino = location.objectid;
6021                 entry->offset = found_key.offset;
6022                 entries++;
6023                 addr += sizeof(struct dir_entry) + name_len;
6024                 total_len += sizeof(struct dir_entry) + name_len;
6025 next:
6026                 path->slots[0]++;
6027         }
6028         btrfs_release_path(path);
6029
6030         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
6031         if (ret)
6032                 goto nopos;
6033
6034         ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list);
6035         if (ret)
6036                 goto nopos;
6037
6038         /*
6039          * Stop new entries from being returned after we return the last
6040          * entry.
6041          *
6042          * New directory entries are assigned a strictly increasing
6043          * offset.  This means that new entries created during readdir
6044          * are *guaranteed* to be seen in the future by that readdir.
6045          * This has broken buggy programs which operate on names as
6046          * they're returned by readdir.  Until we re-use freed offsets
6047          * we have this hack to stop new entries from being returned
6048          * under the assumption that they'll never reach this huge
6049          * offset.
6050          *
6051          * This is being careful not to overflow 32bit loff_t unless the
6052          * last entry requires it because doing so has broken 32bit apps
6053          * in the past.
6054          */
6055         if (ctx->pos >= INT_MAX)
6056                 ctx->pos = LLONG_MAX;
6057         else
6058                 ctx->pos = INT_MAX;
6059 nopos:
6060         ret = 0;
6061 err:
6062         if (put)
6063                 btrfs_readdir_put_delayed_items(inode, &ins_list, &del_list);
6064         btrfs_free_path(path);
6065         return ret;
6066 }
6067
6068 int btrfs_write_inode(struct inode *inode, struct writeback_control *wbc)
6069 {
6070         struct btrfs_root *root = BTRFS_I(inode)->root;
6071         struct btrfs_trans_handle *trans;
6072         int ret = 0;
6073         bool nolock = false;
6074
6075         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
6076                 return 0;
6077
6078         if (btrfs_fs_closing(root->fs_info) &&
6079                         btrfs_is_free_space_inode(BTRFS_I(inode)))
6080                 nolock = true;
6081
6082         if (wbc->sync_mode == WB_SYNC_ALL) {
6083                 if (nolock)
6084                         trans = btrfs_join_transaction_nolock(root);
6085                 else
6086                         trans = btrfs_join_transaction(root);
6087                 if (IS_ERR(trans))
6088                         return PTR_ERR(trans);
6089                 ret = btrfs_commit_transaction(trans);
6090         }
6091         return ret;
6092 }
6093
6094 /*
6095  * This is somewhat expensive, updating the tree every time the
6096  * inode changes.  But, it is most likely to find the inode in cache.
6097  * FIXME, needs more benchmarking...there are no reasons other than performance
6098  * to keep or drop this code.
6099  */
6100 static int btrfs_dirty_inode(struct inode *inode)
6101 {
6102         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6103         struct btrfs_root *root = BTRFS_I(inode)->root;
6104         struct btrfs_trans_handle *trans;
6105         int ret;
6106
6107         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
6108                 return 0;
6109
6110         trans = btrfs_join_transaction(root);
6111         if (IS_ERR(trans))
6112                 return PTR_ERR(trans);
6113
6114         ret = btrfs_update_inode(trans, root, inode);
6115         if (ret && ret == -ENOSPC) {
6116                 /* whoops, lets try again with the full transaction */
6117                 btrfs_end_transaction(trans);
6118                 trans = btrfs_start_transaction(root, 1);
6119                 if (IS_ERR(trans))
6120                         return PTR_ERR(trans);
6121
6122                 ret = btrfs_update_inode(trans, root, inode);
6123         }
6124         btrfs_end_transaction(trans);
6125         if (BTRFS_I(inode)->delayed_node)
6126                 btrfs_balance_delayed_items(fs_info);
6127
6128         return ret;
6129 }
6130
6131 /*
6132  * This is a copy of file_update_time.  We need this so we can return error on
6133  * ENOSPC for updating the inode in the case of file write and mmap writes.
6134  */
6135 static int btrfs_update_time(struct inode *inode, struct timespec *now,
6136                              int flags)
6137 {
6138         struct btrfs_root *root = BTRFS_I(inode)->root;
6139         bool dirty = flags & ~S_VERSION;
6140
6141         if (btrfs_root_readonly(root))
6142                 return -EROFS;
6143
6144         if (flags & S_VERSION)
6145                 dirty |= inode_maybe_inc_iversion(inode, dirty);
6146         if (flags & S_CTIME)
6147                 inode->i_ctime = *now;
6148         if (flags & S_MTIME)
6149                 inode->i_mtime = *now;
6150         if (flags & S_ATIME)
6151                 inode->i_atime = *now;
6152         return dirty ? btrfs_dirty_inode(inode) : 0;
6153 }
6154
6155 /*
6156  * find the highest existing sequence number in a directory
6157  * and then set the in-memory index_cnt variable to reflect
6158  * free sequence numbers
6159  */
6160 static int btrfs_set_inode_index_count(struct btrfs_inode *inode)
6161 {
6162         struct btrfs_root *root = inode->root;
6163         struct btrfs_key key, found_key;
6164         struct btrfs_path *path;
6165         struct extent_buffer *leaf;
6166         int ret;
6167
6168         key.objectid = btrfs_ino(inode);
6169         key.type = BTRFS_DIR_INDEX_KEY;
6170         key.offset = (u64)-1;
6171
6172         path = btrfs_alloc_path();
6173         if (!path)
6174                 return -ENOMEM;
6175
6176         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6177         if (ret < 0)
6178                 goto out;
6179         /* FIXME: we should be able to handle this */
6180         if (ret == 0)
6181                 goto out;
6182         ret = 0;
6183
6184         /*
6185          * MAGIC NUMBER EXPLANATION:
6186          * since we search a directory based on f_pos we have to start at 2
6187          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
6188          * else has to start at 2
6189          */
6190         if (path->slots[0] == 0) {
6191                 inode->index_cnt = 2;
6192                 goto out;
6193         }
6194
6195         path->slots[0]--;
6196
6197         leaf = path->nodes[0];
6198         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6199
6200         if (found_key.objectid != btrfs_ino(inode) ||
6201             found_key.type != BTRFS_DIR_INDEX_KEY) {
6202                 inode->index_cnt = 2;
6203                 goto out;
6204         }
6205
6206         inode->index_cnt = found_key.offset + 1;
6207 out:
6208         btrfs_free_path(path);
6209         return ret;
6210 }
6211
6212 /*
6213  * helper to find a free sequence number in a given directory.  This current
6214  * code is very simple, later versions will do smarter things in the btree
6215  */
6216 int btrfs_set_inode_index(struct btrfs_inode *dir, u64 *index)
6217 {
6218         int ret = 0;
6219
6220         if (dir->index_cnt == (u64)-1) {
6221                 ret = btrfs_inode_delayed_dir_index_count(dir);
6222                 if (ret) {
6223                         ret = btrfs_set_inode_index_count(dir);
6224                         if (ret)
6225                                 return ret;
6226                 }
6227         }
6228
6229         *index = dir->index_cnt;
6230         dir->index_cnt++;
6231
6232         return ret;
6233 }
6234
6235 static int btrfs_insert_inode_locked(struct inode *inode)
6236 {
6237         struct btrfs_iget_args args;
6238         args.location = &BTRFS_I(inode)->location;
6239         args.root = BTRFS_I(inode)->root;
6240
6241         return insert_inode_locked4(inode,
6242                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
6243                    btrfs_find_actor, &args);
6244 }
6245
6246 /*
6247  * Inherit flags from the parent inode.
6248  *
6249  * Currently only the compression flags and the cow flags are inherited.
6250  */
6251 static void btrfs_inherit_iflags(struct inode *inode, struct inode *dir)
6252 {
6253         unsigned int flags;
6254
6255         if (!dir)
6256                 return;
6257
6258         flags = BTRFS_I(dir)->flags;
6259
6260         if (flags & BTRFS_INODE_NOCOMPRESS) {
6261                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_COMPRESS;
6262                 BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
6263         } else if (flags & BTRFS_INODE_COMPRESS) {
6264                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_NOCOMPRESS;
6265                 BTRFS_I(inode)->flags |= BTRFS_INODE_COMPRESS;
6266         }
6267
6268         if (flags & BTRFS_INODE_NODATACOW) {
6269                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW;
6270                 if (S_ISREG(inode->i_mode))
6271                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6272         }
6273
6274         btrfs_update_iflags(inode);
6275 }
6276
6277 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
6278                                      struct btrfs_root *root,
6279                                      struct inode *dir,
6280                                      const char *name, int name_len,
6281                                      u64 ref_objectid, u64 objectid,
6282                                      umode_t mode, u64 *index)
6283 {
6284         struct btrfs_fs_info *fs_info = root->fs_info;
6285         struct inode *inode;
6286         struct btrfs_inode_item *inode_item;
6287         struct btrfs_key *location;
6288         struct btrfs_path *path;
6289         struct btrfs_inode_ref *ref;
6290         struct btrfs_key key[2];
6291         u32 sizes[2];
6292         int nitems = name ? 2 : 1;
6293         unsigned long ptr;
6294         int ret;
6295
6296         path = btrfs_alloc_path();
6297         if (!path)
6298                 return ERR_PTR(-ENOMEM);
6299
6300         inode = new_inode(fs_info->sb);
6301         if (!inode) {
6302                 btrfs_free_path(path);
6303                 return ERR_PTR(-ENOMEM);
6304         }
6305
6306         /*
6307          * O_TMPFILE, set link count to 0, so that after this point,
6308          * we fill in an inode item with the correct link count.
6309          */
6310         if (!name)
6311                 set_nlink(inode, 0);
6312
6313         /*
6314          * we have to initialize this early, so we can reclaim the inode
6315          * number if we fail afterwards in this function.
6316          */
6317         inode->i_ino = objectid;
6318
6319         if (dir && name) {
6320                 trace_btrfs_inode_request(dir);
6321
6322                 ret = btrfs_set_inode_index(BTRFS_I(dir), index);
6323                 if (ret) {
6324                         btrfs_free_path(path);
6325                         iput(inode);
6326                         return ERR_PTR(ret);
6327                 }
6328         } else if (dir) {
6329                 *index = 0;
6330         }
6331         /*
6332          * index_cnt is ignored for everything but a dir,
6333          * btrfs_set_inode_index_count has an explanation for the magic
6334          * number
6335          */
6336         BTRFS_I(inode)->index_cnt = 2;
6337         BTRFS_I(inode)->dir_index = *index;
6338         BTRFS_I(inode)->root = root;
6339         BTRFS_I(inode)->generation = trans->transid;
6340         inode->i_generation = BTRFS_I(inode)->generation;
6341
6342         /*
6343          * We could have gotten an inode number from somebody who was fsynced
6344          * and then removed in this same transaction, so let's just set full
6345          * sync since it will be a full sync anyway and this will blow away the
6346          * old info in the log.
6347          */
6348         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
6349
6350         key[0].objectid = objectid;
6351         key[0].type = BTRFS_INODE_ITEM_KEY;
6352         key[0].offset = 0;
6353
6354         sizes[0] = sizeof(struct btrfs_inode_item);
6355
6356         if (name) {
6357                 /*
6358                  * Start new inodes with an inode_ref. This is slightly more
6359                  * efficient for small numbers of hard links since they will
6360                  * be packed into one item. Extended refs will kick in if we
6361                  * add more hard links than can fit in the ref item.
6362                  */
6363                 key[1].objectid = objectid;
6364                 key[1].type = BTRFS_INODE_REF_KEY;
6365                 key[1].offset = ref_objectid;
6366
6367                 sizes[1] = name_len + sizeof(*ref);
6368         }
6369
6370         location = &BTRFS_I(inode)->location;
6371         location->objectid = objectid;
6372         location->offset = 0;
6373         location->type = BTRFS_INODE_ITEM_KEY;
6374
6375         ret = btrfs_insert_inode_locked(inode);
6376         if (ret < 0)
6377                 goto fail;
6378
6379         path->leave_spinning = 1;
6380         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
6381         if (ret != 0)
6382                 goto fail_unlock;
6383
6384         inode_init_owner(inode, dir, mode);
6385         inode_set_bytes(inode, 0);
6386
6387         inode->i_mtime = current_time(inode);
6388         inode->i_atime = inode->i_mtime;
6389         inode->i_ctime = inode->i_mtime;
6390         BTRFS_I(inode)->i_otime = inode->i_mtime;
6391
6392         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
6393                                   struct btrfs_inode_item);
6394         memzero_extent_buffer(path->nodes[0], (unsigned long)inode_item,
6395                              sizeof(*inode_item));
6396         fill_inode_item(trans, path->nodes[0], inode_item, inode);
6397
6398         if (name) {
6399                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
6400                                      struct btrfs_inode_ref);
6401                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6402                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6403                 ptr = (unsigned long)(ref + 1);
6404                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6405         }
6406
6407         btrfs_mark_buffer_dirty(path->nodes[0]);
6408         btrfs_free_path(path);
6409
6410         btrfs_inherit_iflags(inode, dir);
6411
6412         if (S_ISREG(mode)) {
6413                 if (btrfs_test_opt(fs_info, NODATASUM))
6414                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6415                 if (btrfs_test_opt(fs_info, NODATACOW))
6416                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6417                                 BTRFS_INODE_NODATASUM;
6418         }
6419
6420         inode_tree_add(inode);
6421
6422         trace_btrfs_inode_new(inode);
6423         btrfs_set_inode_last_trans(trans, inode);
6424
6425         btrfs_update_root_times(trans, root);
6426
6427         ret = btrfs_inode_inherit_props(trans, inode, dir);
6428         if (ret)
6429                 btrfs_err(fs_info,
6430                           "error inheriting props for ino %llu (root %llu): %d",
6431                         btrfs_ino(BTRFS_I(inode)), root->root_key.objectid, ret);
6432
6433         return inode;
6434
6435 fail_unlock:
6436         unlock_new_inode(inode);
6437 fail:
6438         if (dir && name)
6439                 BTRFS_I(dir)->index_cnt--;
6440         btrfs_free_path(path);
6441         iput(inode);
6442         return ERR_PTR(ret);
6443 }
6444
6445 static inline u8 btrfs_inode_type(struct inode *inode)
6446 {
6447         return btrfs_type_by_mode[(inode->i_mode & S_IFMT) >> S_SHIFT];
6448 }
6449
6450 /*
6451  * utility function to add 'inode' into 'parent_inode' with
6452  * a give name and a given sequence number.
6453  * if 'add_backref' is true, also insert a backref from the
6454  * inode to the parent directory.
6455  */
6456 int btrfs_add_link(struct btrfs_trans_handle *trans,
6457                    struct btrfs_inode *parent_inode, struct btrfs_inode *inode,
6458                    const char *name, int name_len, int add_backref, u64 index)
6459 {
6460         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6461         int ret = 0;
6462         struct btrfs_key key;
6463         struct btrfs_root *root = parent_inode->root;
6464         u64 ino = btrfs_ino(inode);
6465         u64 parent_ino = btrfs_ino(parent_inode);
6466
6467         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6468                 memcpy(&key, &inode->root->root_key, sizeof(key));
6469         } else {
6470                 key.objectid = ino;
6471                 key.type = BTRFS_INODE_ITEM_KEY;
6472                 key.offset = 0;
6473         }
6474
6475         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6476                 ret = btrfs_add_root_ref(trans, fs_info, key.objectid,
6477                                          root->root_key.objectid, parent_ino,
6478                                          index, name, name_len);
6479         } else if (add_backref) {
6480                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6481                                              parent_ino, index);
6482         }
6483
6484         /* Nothing to clean up yet */
6485         if (ret)
6486                 return ret;
6487
6488         ret = btrfs_insert_dir_item(trans, root, name, name_len,
6489                                     parent_inode, &key,
6490                                     btrfs_inode_type(&inode->vfs_inode), index);
6491         if (ret == -EEXIST || ret == -EOVERFLOW)
6492                 goto fail_dir_item;
6493         else if (ret) {
6494                 btrfs_abort_transaction(trans, ret);
6495                 return ret;
6496         }
6497
6498         btrfs_i_size_write(parent_inode, parent_inode->vfs_inode.i_size +
6499                            name_len * 2);
6500         inode_inc_iversion(&parent_inode->vfs_inode);
6501         parent_inode->vfs_inode.i_mtime = parent_inode->vfs_inode.i_ctime =
6502                 current_time(&parent_inode->vfs_inode);
6503         ret = btrfs_update_inode(trans, root, &parent_inode->vfs_inode);
6504         if (ret)
6505                 btrfs_abort_transaction(trans, ret);
6506         return ret;
6507
6508 fail_dir_item:
6509         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6510                 u64 local_index;
6511                 int err;
6512                 err = btrfs_del_root_ref(trans, fs_info, key.objectid,
6513                                          root->root_key.objectid, parent_ino,
6514                                          &local_index, name, name_len);
6515
6516         } else if (add_backref) {
6517                 u64 local_index;
6518                 int err;
6519
6520                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6521                                           ino, parent_ino, &local_index);
6522         }
6523         return ret;
6524 }
6525
6526 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6527                             struct btrfs_inode *dir, struct dentry *dentry,
6528                             struct btrfs_inode *inode, int backref, u64 index)
6529 {
6530         int err = btrfs_add_link(trans, dir, inode,
6531                                  dentry->d_name.name, dentry->d_name.len,
6532                                  backref, index);
6533         if (err > 0)
6534                 err = -EEXIST;
6535         return err;
6536 }
6537
6538 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6539                         umode_t mode, dev_t rdev)
6540 {
6541         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6542         struct btrfs_trans_handle *trans;
6543         struct btrfs_root *root = BTRFS_I(dir)->root;
6544         struct inode *inode = NULL;
6545         int err;
6546         int drop_inode = 0;
6547         u64 objectid;
6548         u64 index = 0;
6549
6550         /*
6551          * 2 for inode item and ref
6552          * 2 for dir items
6553          * 1 for xattr if selinux is on
6554          */
6555         trans = btrfs_start_transaction(root, 5);
6556         if (IS_ERR(trans))
6557                 return PTR_ERR(trans);
6558
6559         err = btrfs_find_free_ino(root, &objectid);
6560         if (err)
6561                 goto out_unlock;
6562
6563         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6564                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6565                         mode, &index);
6566         if (IS_ERR(inode)) {
6567                 err = PTR_ERR(inode);
6568                 goto out_unlock;
6569         }
6570
6571         /*
6572         * If the active LSM wants to access the inode during
6573         * d_instantiate it needs these. Smack checks to see
6574         * if the filesystem supports xattrs by looking at the
6575         * ops vector.
6576         */
6577         inode->i_op = &btrfs_special_inode_operations;
6578         init_special_inode(inode, inode->i_mode, rdev);
6579
6580         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6581         if (err)
6582                 goto out_unlock_inode;
6583
6584         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6585                         0, index);
6586         if (err) {
6587                 goto out_unlock_inode;
6588         } else {
6589                 btrfs_update_inode(trans, root, inode);
6590                 unlock_new_inode(inode);
6591                 d_instantiate(dentry, inode);
6592         }
6593
6594 out_unlock:
6595         btrfs_end_transaction(trans);
6596         btrfs_btree_balance_dirty(fs_info);
6597         if (drop_inode) {
6598                 inode_dec_link_count(inode);
6599                 iput(inode);
6600         }
6601         return err;
6602
6603 out_unlock_inode:
6604         drop_inode = 1;
6605         unlock_new_inode(inode);
6606         goto out_unlock;
6607
6608 }
6609
6610 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6611                         umode_t mode, bool excl)
6612 {
6613         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6614         struct btrfs_trans_handle *trans;
6615         struct btrfs_root *root = BTRFS_I(dir)->root;
6616         struct inode *inode = NULL;
6617         int drop_inode_on_err = 0;
6618         int err;
6619         u64 objectid;
6620         u64 index = 0;
6621
6622         /*
6623          * 2 for inode item and ref
6624          * 2 for dir items
6625          * 1 for xattr if selinux is on
6626          */
6627         trans = btrfs_start_transaction(root, 5);
6628         if (IS_ERR(trans))
6629                 return PTR_ERR(trans);
6630
6631         err = btrfs_find_free_ino(root, &objectid);
6632         if (err)
6633                 goto out_unlock;
6634
6635         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6636                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6637                         mode, &index);
6638         if (IS_ERR(inode)) {
6639                 err = PTR_ERR(inode);
6640                 goto out_unlock;
6641         }
6642         drop_inode_on_err = 1;
6643         /*
6644         * If the active LSM wants to access the inode during
6645         * d_instantiate it needs these. Smack checks to see
6646         * if the filesystem supports xattrs by looking at the
6647         * ops vector.
6648         */
6649         inode->i_fop = &btrfs_file_operations;
6650         inode->i_op = &btrfs_file_inode_operations;
6651         inode->i_mapping->a_ops = &btrfs_aops;
6652
6653         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6654         if (err)
6655                 goto out_unlock_inode;
6656
6657         err = btrfs_update_inode(trans, root, inode);
6658         if (err)
6659                 goto out_unlock_inode;
6660
6661         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6662                         0, index);
6663         if (err)
6664                 goto out_unlock_inode;
6665
6666         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6667         unlock_new_inode(inode);
6668         d_instantiate(dentry, inode);
6669
6670 out_unlock:
6671         btrfs_end_transaction(trans);
6672         if (err && drop_inode_on_err) {
6673                 inode_dec_link_count(inode);
6674                 iput(inode);
6675         }
6676         btrfs_btree_balance_dirty(fs_info);
6677         return err;
6678
6679 out_unlock_inode:
6680         unlock_new_inode(inode);
6681         goto out_unlock;
6682
6683 }
6684
6685 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6686                       struct dentry *dentry)
6687 {
6688         struct btrfs_trans_handle *trans = NULL;
6689         struct btrfs_root *root = BTRFS_I(dir)->root;
6690         struct inode *inode = d_inode(old_dentry);
6691         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6692         u64 index;
6693         int err;
6694         int drop_inode = 0;
6695
6696         /* do not allow sys_link's with other subvols of the same device */
6697         if (root->objectid != BTRFS_I(inode)->root->objectid)
6698                 return -EXDEV;
6699
6700         if (inode->i_nlink >= BTRFS_LINK_MAX)
6701                 return -EMLINK;
6702
6703         err = btrfs_set_inode_index(BTRFS_I(dir), &index);
6704         if (err)
6705                 goto fail;
6706
6707         /*
6708          * 2 items for inode and inode ref
6709          * 2 items for dir items
6710          * 1 item for parent inode
6711          */
6712         trans = btrfs_start_transaction(root, 5);
6713         if (IS_ERR(trans)) {
6714                 err = PTR_ERR(trans);
6715                 trans = NULL;
6716                 goto fail;
6717         }
6718
6719         /* There are several dir indexes for this inode, clear the cache. */
6720         BTRFS_I(inode)->dir_index = 0ULL;
6721         inc_nlink(inode);
6722         inode_inc_iversion(inode);
6723         inode->i_ctime = current_time(inode);
6724         ihold(inode);
6725         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6726
6727         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6728                         1, index);
6729
6730         if (err) {
6731                 drop_inode = 1;
6732         } else {
6733                 struct dentry *parent = dentry->d_parent;
6734                 err = btrfs_update_inode(trans, root, inode);
6735                 if (err)
6736                         goto fail;
6737                 if (inode->i_nlink == 1) {
6738                         /*
6739                          * If new hard link count is 1, it's a file created
6740                          * with open(2) O_TMPFILE flag.
6741                          */
6742                         err = btrfs_orphan_del(trans, BTRFS_I(inode));
6743                         if (err)
6744                                 goto fail;
6745                 }
6746                 d_instantiate(dentry, inode);
6747                 btrfs_log_new_name(trans, BTRFS_I(inode), NULL, parent);
6748         }
6749
6750 fail:
6751         if (trans)
6752                 btrfs_end_transaction(trans);
6753         if (drop_inode) {
6754                 inode_dec_link_count(inode);
6755                 iput(inode);
6756         }
6757         btrfs_btree_balance_dirty(fs_info);
6758         return err;
6759 }
6760
6761 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6762 {
6763         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6764         struct inode *inode = NULL;
6765         struct btrfs_trans_handle *trans;
6766         struct btrfs_root *root = BTRFS_I(dir)->root;
6767         int err = 0;
6768         int drop_on_err = 0;
6769         u64 objectid = 0;
6770         u64 index = 0;
6771
6772         /*
6773          * 2 items for inode and ref
6774          * 2 items for dir items
6775          * 1 for xattr if selinux is on
6776          */
6777         trans = btrfs_start_transaction(root, 5);
6778         if (IS_ERR(trans))
6779                 return PTR_ERR(trans);
6780
6781         err = btrfs_find_free_ino(root, &objectid);
6782         if (err)
6783                 goto out_fail;
6784
6785         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6786                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6787                         S_IFDIR | mode, &index);
6788         if (IS_ERR(inode)) {
6789                 err = PTR_ERR(inode);
6790                 goto out_fail;
6791         }
6792
6793         drop_on_err = 1;
6794         /* these must be set before we unlock the inode */
6795         inode->i_op = &btrfs_dir_inode_operations;
6796         inode->i_fop = &btrfs_dir_file_operations;
6797
6798         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6799         if (err)
6800                 goto out_fail_inode;
6801
6802         btrfs_i_size_write(BTRFS_I(inode), 0);
6803         err = btrfs_update_inode(trans, root, inode);
6804         if (err)
6805                 goto out_fail_inode;
6806
6807         err = btrfs_add_link(trans, BTRFS_I(dir), BTRFS_I(inode),
6808                         dentry->d_name.name,
6809                         dentry->d_name.len, 0, index);
6810         if (err)
6811                 goto out_fail_inode;
6812
6813         d_instantiate(dentry, inode);
6814         /*
6815          * mkdir is special.  We're unlocking after we call d_instantiate
6816          * to avoid a race with nfsd calling d_instantiate.
6817          */
6818         unlock_new_inode(inode);
6819         drop_on_err = 0;
6820
6821 out_fail:
6822         btrfs_end_transaction(trans);
6823         if (drop_on_err) {
6824                 inode_dec_link_count(inode);
6825                 iput(inode);
6826         }
6827         btrfs_btree_balance_dirty(fs_info);
6828         return err;
6829
6830 out_fail_inode:
6831         unlock_new_inode(inode);
6832         goto out_fail;
6833 }
6834
6835 static noinline int uncompress_inline(struct btrfs_path *path,
6836                                       struct page *page,
6837                                       size_t pg_offset, u64 extent_offset,
6838                                       struct btrfs_file_extent_item *item)
6839 {
6840         int ret;
6841         struct extent_buffer *leaf = path->nodes[0];
6842         char *tmp;
6843         size_t max_size;
6844         unsigned long inline_size;
6845         unsigned long ptr;
6846         int compress_type;
6847
6848         WARN_ON(pg_offset != 0);
6849         compress_type = btrfs_file_extent_compression(leaf, item);
6850         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6851         inline_size = btrfs_file_extent_inline_item_len(leaf,
6852                                         btrfs_item_nr(path->slots[0]));
6853         tmp = kmalloc(inline_size, GFP_NOFS);
6854         if (!tmp)
6855                 return -ENOMEM;
6856         ptr = btrfs_file_extent_inline_start(item);
6857
6858         read_extent_buffer(leaf, tmp, ptr, inline_size);
6859
6860         max_size = min_t(unsigned long, PAGE_SIZE, max_size);
6861         ret = btrfs_decompress(compress_type, tmp, page,
6862                                extent_offset, inline_size, max_size);
6863
6864         /*
6865          * decompression code contains a memset to fill in any space between the end
6866          * of the uncompressed data and the end of max_size in case the decompressed
6867          * data ends up shorter than ram_bytes.  That doesn't cover the hole between
6868          * the end of an inline extent and the beginning of the next block, so we
6869          * cover that region here.
6870          */
6871
6872         if (max_size + pg_offset < PAGE_SIZE) {
6873                 char *map = kmap(page);
6874                 memset(map + pg_offset + max_size, 0, PAGE_SIZE - max_size - pg_offset);
6875                 kunmap(page);
6876         }
6877         kfree(tmp);
6878         return ret;
6879 }
6880
6881 /*
6882  * a bit scary, this does extent mapping from logical file offset to the disk.
6883  * the ugly parts come from merging extents from the disk with the in-ram
6884  * representation.  This gets more complex because of the data=ordered code,
6885  * where the in-ram extents might be locked pending data=ordered completion.
6886  *
6887  * This also copies inline extents directly into the page.
6888  */
6889 struct extent_map *btrfs_get_extent(struct btrfs_inode *inode,
6890                 struct page *page,
6891             size_t pg_offset, u64 start, u64 len,
6892                 int create)
6893 {
6894         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6895         int ret;
6896         int err = 0;
6897         u64 extent_start = 0;
6898         u64 extent_end = 0;
6899         u64 objectid = btrfs_ino(inode);
6900         u32 found_type;
6901         struct btrfs_path *path = NULL;
6902         struct btrfs_root *root = inode->root;
6903         struct btrfs_file_extent_item *item;
6904         struct extent_buffer *leaf;
6905         struct btrfs_key found_key;
6906         struct extent_map *em = NULL;
6907         struct extent_map_tree *em_tree = &inode->extent_tree;
6908         struct extent_io_tree *io_tree = &inode->io_tree;
6909         const bool new_inline = !page || create;
6910
6911         read_lock(&em_tree->lock);
6912         em = lookup_extent_mapping(em_tree, start, len);
6913         if (em)
6914                 em->bdev = fs_info->fs_devices->latest_bdev;
6915         read_unlock(&em_tree->lock);
6916
6917         if (em) {
6918                 if (em->start > start || em->start + em->len <= start)
6919                         free_extent_map(em);
6920                 else if (em->block_start == EXTENT_MAP_INLINE && page)
6921                         free_extent_map(em);
6922                 else
6923                         goto out;
6924         }
6925         em = alloc_extent_map();
6926         if (!em) {
6927                 err = -ENOMEM;
6928                 goto out;
6929         }
6930         em->bdev = fs_info->fs_devices->latest_bdev;
6931         em->start = EXTENT_MAP_HOLE;
6932         em->orig_start = EXTENT_MAP_HOLE;
6933         em->len = (u64)-1;
6934         em->block_len = (u64)-1;
6935
6936         if (!path) {
6937                 path = btrfs_alloc_path();
6938                 if (!path) {
6939                         err = -ENOMEM;
6940                         goto out;
6941                 }
6942                 /*
6943                  * Chances are we'll be called again, so go ahead and do
6944                  * readahead
6945                  */
6946                 path->reada = READA_FORWARD;
6947         }
6948
6949         ret = btrfs_lookup_file_extent(NULL, root, path, objectid, start, 0);
6950         if (ret < 0) {
6951                 err = ret;
6952                 goto out;
6953         }
6954
6955         if (ret != 0) {
6956                 if (path->slots[0] == 0)
6957                         goto not_found;
6958                 path->slots[0]--;
6959         }
6960
6961         leaf = path->nodes[0];
6962         item = btrfs_item_ptr(leaf, path->slots[0],
6963                               struct btrfs_file_extent_item);
6964         /* are we inside the extent that was found? */
6965         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6966         found_type = found_key.type;
6967         if (found_key.objectid != objectid ||
6968             found_type != BTRFS_EXTENT_DATA_KEY) {
6969                 /*
6970                  * If we backup past the first extent we want to move forward
6971                  * and see if there is an extent in front of us, otherwise we'll
6972                  * say there is a hole for our whole search range which can
6973                  * cause problems.
6974                  */
6975                 extent_end = start;
6976                 goto next;
6977         }
6978
6979         found_type = btrfs_file_extent_type(leaf, item);
6980         extent_start = found_key.offset;
6981         if (found_type == BTRFS_FILE_EXTENT_REG ||
6982             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6983                 extent_end = extent_start +
6984                        btrfs_file_extent_num_bytes(leaf, item);
6985
6986                 trace_btrfs_get_extent_show_fi_regular(inode, leaf, item,
6987                                                        extent_start);
6988         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6989                 size_t size;
6990                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
6991                 extent_end = ALIGN(extent_start + size,
6992                                    fs_info->sectorsize);
6993
6994                 trace_btrfs_get_extent_show_fi_inline(inode, leaf, item,
6995                                                       path->slots[0],
6996                                                       extent_start);
6997         }
6998 next:
6999         if (start >= extent_end) {
7000                 path->slots[0]++;
7001                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
7002                         ret = btrfs_next_leaf(root, path);
7003                         if (ret < 0) {
7004                                 err = ret;
7005                                 goto out;
7006                         }
7007                         if (ret > 0)
7008                                 goto not_found;
7009                         leaf = path->nodes[0];
7010                 }
7011                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7012                 if (found_key.objectid != objectid ||
7013                     found_key.type != BTRFS_EXTENT_DATA_KEY)
7014                         goto not_found;
7015                 if (start + len <= found_key.offset)
7016                         goto not_found;
7017                 if (start > found_key.offset)
7018                         goto next;
7019                 em->start = start;
7020                 em->orig_start = start;
7021                 em->len = found_key.offset - start;
7022                 goto not_found_em;
7023         }
7024
7025         btrfs_extent_item_to_extent_map(inode, path, item,
7026                         new_inline, em);
7027
7028         if (found_type == BTRFS_FILE_EXTENT_REG ||
7029             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7030                 goto insert;
7031         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
7032                 unsigned long ptr;
7033                 char *map;
7034                 size_t size;
7035                 size_t extent_offset;
7036                 size_t copy_size;
7037
7038                 if (new_inline)
7039                         goto out;
7040
7041                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
7042                 extent_offset = page_offset(page) + pg_offset - extent_start;
7043                 copy_size = min_t(u64, PAGE_SIZE - pg_offset,
7044                                   size - extent_offset);
7045                 em->start = extent_start + extent_offset;
7046                 em->len = ALIGN(copy_size, fs_info->sectorsize);
7047                 em->orig_block_len = em->len;
7048                 em->orig_start = em->start;
7049                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
7050                 if (!PageUptodate(page)) {
7051                         if (btrfs_file_extent_compression(leaf, item) !=
7052                             BTRFS_COMPRESS_NONE) {
7053                                 ret = uncompress_inline(path, page, pg_offset,
7054                                                         extent_offset, item);
7055                                 if (ret) {
7056                                         err = ret;
7057                                         goto out;
7058                                 }
7059                         } else {
7060                                 map = kmap(page);
7061                                 read_extent_buffer(leaf, map + pg_offset, ptr,
7062                                                    copy_size);
7063                                 if (pg_offset + copy_size < PAGE_SIZE) {
7064                                         memset(map + pg_offset + copy_size, 0,
7065                                                PAGE_SIZE - pg_offset -
7066                                                copy_size);
7067                                 }
7068                                 kunmap(page);
7069                         }
7070                         flush_dcache_page(page);
7071                 }
7072                 set_extent_uptodate(io_tree, em->start,
7073                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
7074                 goto insert;
7075         }
7076 not_found:
7077         em->start = start;
7078         em->orig_start = start;
7079         em->len = len;
7080 not_found_em:
7081         em->block_start = EXTENT_MAP_HOLE;
7082 insert:
7083         btrfs_release_path(path);
7084         if (em->start > start || extent_map_end(em) <= start) {
7085                 btrfs_err(fs_info,
7086                           "bad extent! em: [%llu %llu] passed [%llu %llu]",
7087                           em->start, em->len, start, len);
7088                 err = -EIO;
7089                 goto out;
7090         }
7091
7092         err = 0;
7093         write_lock(&em_tree->lock);
7094         err = btrfs_add_extent_mapping(em_tree, &em, start, len);
7095         write_unlock(&em_tree->lock);
7096 out:
7097
7098         trace_btrfs_get_extent(root, inode, em);
7099
7100         btrfs_free_path(path);
7101         if (err) {
7102                 free_extent_map(em);
7103                 return ERR_PTR(err);
7104         }
7105         BUG_ON(!em); /* Error is always set */
7106         return em;
7107 }
7108
7109 struct extent_map *btrfs_get_extent_fiemap(struct btrfs_inode *inode,
7110                 struct page *page,
7111                 size_t pg_offset, u64 start, u64 len,
7112                 int create)
7113 {
7114         struct extent_map *em;
7115         struct extent_map *hole_em = NULL;
7116         u64 range_start = start;
7117         u64 end;
7118         u64 found;
7119         u64 found_end;
7120         int err = 0;
7121
7122         em = btrfs_get_extent(inode, page, pg_offset, start, len, create);
7123         if (IS_ERR(em))
7124                 return em;
7125         /*
7126          * If our em maps to:
7127          * - a hole or
7128          * - a pre-alloc extent,
7129          * there might actually be delalloc bytes behind it.
7130          */
7131         if (em->block_start != EXTENT_MAP_HOLE &&
7132             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7133                 return em;
7134         else
7135                 hole_em = em;
7136
7137         /* check to see if we've wrapped (len == -1 or similar) */
7138         end = start + len;
7139         if (end < start)
7140                 end = (u64)-1;
7141         else
7142                 end -= 1;
7143
7144         em = NULL;
7145
7146         /* ok, we didn't find anything, lets look for delalloc */
7147         found = count_range_bits(&inode->io_tree, &range_start,
7148                                  end, len, EXTENT_DELALLOC, 1);
7149         found_end = range_start + found;
7150         if (found_end < range_start)
7151                 found_end = (u64)-1;
7152
7153         /*
7154          * we didn't find anything useful, return
7155          * the original results from get_extent()
7156          */
7157         if (range_start > end || found_end <= start) {
7158                 em = hole_em;
7159                 hole_em = NULL;
7160                 goto out;
7161         }
7162
7163         /* adjust the range_start to make sure it doesn't
7164          * go backwards from the start they passed in
7165          */
7166         range_start = max(start, range_start);
7167         found = found_end - range_start;
7168
7169         if (found > 0) {
7170                 u64 hole_start = start;
7171                 u64 hole_len = len;
7172
7173                 em = alloc_extent_map();
7174                 if (!em) {
7175                         err = -ENOMEM;
7176                         goto out;
7177                 }
7178                 /*
7179                  * when btrfs_get_extent can't find anything it
7180                  * returns one huge hole
7181                  *
7182                  * make sure what it found really fits our range, and
7183                  * adjust to make sure it is based on the start from
7184                  * the caller
7185                  */
7186                 if (hole_em) {
7187                         u64 calc_end = extent_map_end(hole_em);
7188
7189                         if (calc_end <= start || (hole_em->start > end)) {
7190                                 free_extent_map(hole_em);
7191                                 hole_em = NULL;
7192                         } else {
7193                                 hole_start = max(hole_em->start, start);
7194                                 hole_len = calc_end - hole_start;
7195                         }
7196                 }
7197                 em->bdev = NULL;
7198                 if (hole_em && range_start > hole_start) {
7199                         /* our hole starts before our delalloc, so we
7200                          * have to return just the parts of the hole
7201                          * that go until  the delalloc starts
7202                          */
7203                         em->len = min(hole_len,
7204                                       range_start - hole_start);
7205                         em->start = hole_start;
7206                         em->orig_start = hole_start;
7207                         /*
7208                          * don't adjust block start at all,
7209                          * it is fixed at EXTENT_MAP_HOLE
7210                          */
7211                         em->block_start = hole_em->block_start;
7212                         em->block_len = hole_len;
7213                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
7214                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
7215                 } else {
7216                         em->start = range_start;
7217                         em->len = found;
7218                         em->orig_start = range_start;
7219                         em->block_start = EXTENT_MAP_DELALLOC;
7220                         em->block_len = found;
7221                 }
7222         } else {
7223                 return hole_em;
7224         }
7225 out:
7226
7227         free_extent_map(hole_em);
7228         if (err) {
7229                 free_extent_map(em);
7230                 return ERR_PTR(err);
7231         }
7232         return em;
7233 }
7234
7235 static struct extent_map *btrfs_create_dio_extent(struct inode *inode,
7236                                                   const u64 start,
7237                                                   const u64 len,
7238                                                   const u64 orig_start,
7239                                                   const u64 block_start,
7240                                                   const u64 block_len,
7241                                                   const u64 orig_block_len,
7242                                                   const u64 ram_bytes,
7243                                                   const int type)
7244 {
7245         struct extent_map *em = NULL;
7246         int ret;
7247
7248         if (type != BTRFS_ORDERED_NOCOW) {
7249                 em = create_io_em(inode, start, len, orig_start,
7250                                   block_start, block_len, orig_block_len,
7251                                   ram_bytes,
7252                                   BTRFS_COMPRESS_NONE, /* compress_type */
7253                                   type);
7254                 if (IS_ERR(em))
7255                         goto out;
7256         }
7257         ret = btrfs_add_ordered_extent_dio(inode, start, block_start,
7258                                            len, block_len, type);
7259         if (ret) {
7260                 if (em) {
7261                         free_extent_map(em);
7262                         btrfs_drop_extent_cache(BTRFS_I(inode), start,
7263                                                 start + len - 1, 0);
7264                 }
7265                 em = ERR_PTR(ret);
7266         }
7267  out:
7268
7269         return em;
7270 }
7271
7272 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
7273                                                   u64 start, u64 len)
7274 {
7275         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7276         struct btrfs_root *root = BTRFS_I(inode)->root;
7277         struct extent_map *em;
7278         struct btrfs_key ins;
7279         u64 alloc_hint;
7280         int ret;
7281
7282         alloc_hint = get_extent_allocation_hint(inode, start, len);
7283         ret = btrfs_reserve_extent(root, len, len, fs_info->sectorsize,
7284                                    0, alloc_hint, &ins, 1, 1);
7285         if (ret)
7286                 return ERR_PTR(ret);
7287
7288         em = btrfs_create_dio_extent(inode, start, ins.offset, start,
7289                                      ins.objectid, ins.offset, ins.offset,
7290                                      ins.offset, BTRFS_ORDERED_REGULAR);
7291         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
7292         if (IS_ERR(em))
7293                 btrfs_free_reserved_extent(fs_info, ins.objectid,
7294                                            ins.offset, 1);
7295
7296         return em;
7297 }
7298
7299 /*
7300  * returns 1 when the nocow is safe, < 1 on error, 0 if the
7301  * block must be cow'd
7302  */
7303 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
7304                               u64 *orig_start, u64 *orig_block_len,
7305                               u64 *ram_bytes)
7306 {
7307         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7308         struct btrfs_path *path;
7309         int ret;
7310         struct extent_buffer *leaf;
7311         struct btrfs_root *root = BTRFS_I(inode)->root;
7312         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7313         struct btrfs_file_extent_item *fi;
7314         struct btrfs_key key;
7315         u64 disk_bytenr;
7316         u64 backref_offset;
7317         u64 extent_end;
7318         u64 num_bytes;
7319         int slot;
7320         int found_type;
7321         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
7322
7323         path = btrfs_alloc_path();
7324         if (!path)
7325                 return -ENOMEM;
7326
7327         ret = btrfs_lookup_file_extent(NULL, root, path,
7328                         btrfs_ino(BTRFS_I(inode)), offset, 0);
7329         if (ret < 0)
7330                 goto out;
7331
7332         slot = path->slots[0];
7333         if (ret == 1) {
7334                 if (slot == 0) {
7335                         /* can't find the item, must cow */
7336                         ret = 0;
7337                         goto out;
7338                 }
7339                 slot--;
7340         }
7341         ret = 0;
7342         leaf = path->nodes[0];
7343         btrfs_item_key_to_cpu(leaf, &key, slot);
7344         if (key.objectid != btrfs_ino(BTRFS_I(inode)) ||
7345             key.type != BTRFS_EXTENT_DATA_KEY) {
7346                 /* not our file or wrong item type, must cow */
7347                 goto out;
7348         }
7349
7350         if (key.offset > offset) {
7351                 /* Wrong offset, must cow */
7352                 goto out;
7353         }
7354
7355         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
7356         found_type = btrfs_file_extent_type(leaf, fi);
7357         if (found_type != BTRFS_FILE_EXTENT_REG &&
7358             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
7359                 /* not a regular extent, must cow */
7360                 goto out;
7361         }
7362
7363         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
7364                 goto out;
7365
7366         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
7367         if (extent_end <= offset)
7368                 goto out;
7369
7370         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7371         if (disk_bytenr == 0)
7372                 goto out;
7373
7374         if (btrfs_file_extent_compression(leaf, fi) ||
7375             btrfs_file_extent_encryption(leaf, fi) ||
7376             btrfs_file_extent_other_encoding(leaf, fi))
7377                 goto out;
7378
7379         backref_offset = btrfs_file_extent_offset(leaf, fi);
7380
7381         if (orig_start) {
7382                 *orig_start = key.offset - backref_offset;
7383                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
7384                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
7385         }
7386
7387         if (btrfs_extent_readonly(fs_info, disk_bytenr))
7388                 goto out;
7389
7390         num_bytes = min(offset + *len, extent_end) - offset;
7391         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7392                 u64 range_end;
7393
7394                 range_end = round_up(offset + num_bytes,
7395                                      root->fs_info->sectorsize) - 1;
7396                 ret = test_range_bit(io_tree, offset, range_end,
7397                                      EXTENT_DELALLOC, 0, NULL);
7398                 if (ret) {
7399                         ret = -EAGAIN;
7400                         goto out;
7401                 }
7402         }
7403
7404         btrfs_release_path(path);
7405
7406         /*
7407          * look for other files referencing this extent, if we
7408          * find any we must cow
7409          */
7410
7411         ret = btrfs_cross_ref_exist(root, btrfs_ino(BTRFS_I(inode)),
7412                                     key.offset - backref_offset, disk_bytenr);
7413         if (ret) {
7414                 ret = 0;
7415                 goto out;
7416         }
7417
7418         /*
7419          * adjust disk_bytenr and num_bytes to cover just the bytes
7420          * in this extent we are about to write.  If there
7421          * are any csums in that range we have to cow in order
7422          * to keep the csums correct
7423          */
7424         disk_bytenr += backref_offset;
7425         disk_bytenr += offset - key.offset;
7426         if (csum_exist_in_range(fs_info, disk_bytenr, num_bytes))
7427                 goto out;
7428         /*
7429          * all of the above have passed, it is safe to overwrite this extent
7430          * without cow
7431          */
7432         *len = num_bytes;
7433         ret = 1;
7434 out:
7435         btrfs_free_path(path);
7436         return ret;
7437 }
7438
7439 bool btrfs_page_exists_in_range(struct inode *inode, loff_t start, loff_t end)
7440 {
7441         struct radix_tree_root *root = &inode->i_mapping->page_tree;
7442         bool found = false;
7443         void **pagep = NULL;
7444         struct page *page = NULL;
7445         unsigned long start_idx;
7446         unsigned long end_idx;
7447
7448         start_idx = start >> PAGE_SHIFT;
7449
7450         /*
7451          * end is the last byte in the last page.  end == start is legal
7452          */
7453         end_idx = end >> PAGE_SHIFT;
7454
7455         rcu_read_lock();
7456
7457         /* Most of the code in this while loop is lifted from
7458          * find_get_page.  It's been modified to begin searching from a
7459          * page and return just the first page found in that range.  If the
7460          * found idx is less than or equal to the end idx then we know that
7461          * a page exists.  If no pages are found or if those pages are
7462          * outside of the range then we're fine (yay!) */
7463         while (page == NULL &&
7464                radix_tree_gang_lookup_slot(root, &pagep, NULL, start_idx, 1)) {
7465                 page = radix_tree_deref_slot(pagep);
7466                 if (unlikely(!page))
7467                         break;
7468
7469                 if (radix_tree_exception(page)) {
7470                         if (radix_tree_deref_retry(page)) {
7471                                 page = NULL;
7472                                 continue;
7473                         }
7474                         /*
7475                          * Otherwise, shmem/tmpfs must be storing a swap entry
7476                          * here as an exceptional entry: so return it without
7477                          * attempting to raise page count.
7478                          */
7479                         page = NULL;
7480                         break; /* TODO: Is this relevant for this use case? */
7481                 }
7482
7483                 if (!page_cache_get_speculative(page)) {
7484                         page = NULL;
7485                         continue;
7486                 }
7487
7488                 /*
7489                  * Has the page moved?
7490                  * This is part of the lockless pagecache protocol. See
7491                  * include/linux/pagemap.h for details.
7492                  */
7493                 if (unlikely(page != *pagep)) {
7494                         put_page(page);
7495                         page = NULL;
7496                 }
7497         }
7498
7499         if (page) {
7500                 if (page->index <= end_idx)
7501                         found = true;
7502                 put_page(page);
7503         }
7504
7505         rcu_read_unlock();
7506         return found;
7507 }
7508
7509 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7510                               struct extent_state **cached_state, int writing)
7511 {
7512         struct btrfs_ordered_extent *ordered;
7513         int ret = 0;
7514
7515         while (1) {
7516                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7517                                  cached_state);
7518                 /*
7519                  * We're concerned with the entire range that we're going to be
7520                  * doing DIO to, so we need to make sure there's no ordered
7521                  * extents in this range.
7522                  */
7523                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), lockstart,
7524                                                      lockend - lockstart + 1);
7525
7526                 /*
7527                  * We need to make sure there are no buffered pages in this
7528                  * range either, we could have raced between the invalidate in
7529                  * generic_file_direct_write and locking the extent.  The
7530                  * invalidate needs to happen so that reads after a write do not
7531                  * get stale data.
7532                  */
7533                 if (!ordered &&
7534                     (!writing ||
7535                      !btrfs_page_exists_in_range(inode, lockstart, lockend)))
7536                         break;
7537
7538                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7539                                      cached_state);
7540
7541                 if (ordered) {
7542                         /*
7543                          * If we are doing a DIO read and the ordered extent we
7544                          * found is for a buffered write, we can not wait for it
7545                          * to complete and retry, because if we do so we can
7546                          * deadlock with concurrent buffered writes on page
7547                          * locks. This happens only if our DIO read covers more
7548                          * than one extent map, if at this point has already
7549                          * created an ordered extent for a previous extent map
7550                          * and locked its range in the inode's io tree, and a
7551                          * concurrent write against that previous extent map's
7552                          * range and this range started (we unlock the ranges
7553                          * in the io tree only when the bios complete and
7554                          * buffered writes always lock pages before attempting
7555                          * to lock range in the io tree).
7556                          */
7557                         if (writing ||
7558                             test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags))
7559                                 btrfs_start_ordered_extent(inode, ordered, 1);
7560                         else
7561                                 ret = -ENOTBLK;
7562                         btrfs_put_ordered_extent(ordered);
7563                 } else {
7564                         /*
7565                          * We could trigger writeback for this range (and wait
7566                          * for it to complete) and then invalidate the pages for
7567                          * this range (through invalidate_inode_pages2_range()),
7568                          * but that can lead us to a deadlock with a concurrent
7569                          * call to readpages() (a buffered read or a defrag call
7570                          * triggered a readahead) on a page lock due to an
7571                          * ordered dio extent we created before but did not have
7572                          * yet a corresponding bio submitted (whence it can not
7573                          * complete), which makes readpages() wait for that
7574                          * ordered extent to complete while holding a lock on
7575                          * that page.
7576                          */
7577                         ret = -ENOTBLK;
7578                 }
7579
7580                 if (ret)
7581                         break;
7582
7583                 cond_resched();
7584         }
7585
7586         return ret;
7587 }
7588
7589 /* The callers of this must take lock_extent() */
7590 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
7591                                        u64 orig_start, u64 block_start,
7592                                        u64 block_len, u64 orig_block_len,
7593                                        u64 ram_bytes, int compress_type,
7594                                        int type)
7595 {
7596         struct extent_map_tree *em_tree;
7597         struct extent_map *em;
7598         struct btrfs_root *root = BTRFS_I(inode)->root;
7599         int ret;
7600
7601         ASSERT(type == BTRFS_ORDERED_PREALLOC ||
7602                type == BTRFS_ORDERED_COMPRESSED ||
7603                type == BTRFS_ORDERED_NOCOW ||
7604                type == BTRFS_ORDERED_REGULAR);
7605
7606         em_tree = &BTRFS_I(inode)->extent_tree;
7607         em = alloc_extent_map();
7608         if (!em)
7609                 return ERR_PTR(-ENOMEM);
7610
7611         em->start = start;
7612         em->orig_start = orig_start;
7613         em->len = len;
7614         em->block_len = block_len;
7615         em->block_start = block_start;
7616         em->bdev = root->fs_info->fs_devices->latest_bdev;
7617         em->orig_block_len = orig_block_len;
7618         em->ram_bytes = ram_bytes;
7619         em->generation = -1;
7620         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7621         if (type == BTRFS_ORDERED_PREALLOC) {
7622                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7623         } else if (type == BTRFS_ORDERED_COMPRESSED) {
7624                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
7625                 em->compress_type = compress_type;
7626         }
7627
7628         do {
7629                 btrfs_drop_extent_cache(BTRFS_I(inode), em->start,
7630                                 em->start + em->len - 1, 0);
7631                 write_lock(&em_tree->lock);
7632                 ret = add_extent_mapping(em_tree, em, 1);
7633                 write_unlock(&em_tree->lock);
7634                 /*
7635                  * The caller has taken lock_extent(), who could race with us
7636                  * to add em?
7637                  */
7638         } while (ret == -EEXIST);
7639
7640         if (ret) {
7641                 free_extent_map(em);
7642                 return ERR_PTR(ret);
7643         }
7644
7645         /* em got 2 refs now, callers needs to do free_extent_map once. */
7646         return em;
7647 }
7648
7649 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7650                                    struct buffer_head *bh_result, int create)
7651 {
7652         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7653         struct extent_map *em;
7654         struct extent_state *cached_state = NULL;
7655         struct btrfs_dio_data *dio_data = NULL;
7656         u64 start = iblock << inode->i_blkbits;
7657         u64 lockstart, lockend;
7658         u64 len = bh_result->b_size;
7659         int unlock_bits = EXTENT_LOCKED;
7660         int ret = 0;
7661
7662         if (create)
7663                 unlock_bits |= EXTENT_DIRTY;
7664         else
7665                 len = min_t(u64, len, fs_info->sectorsize);
7666
7667         lockstart = start;
7668         lockend = start + len - 1;
7669
7670         if (current->journal_info) {
7671                 /*
7672                  * Need to pull our outstanding extents and set journal_info to NULL so
7673                  * that anything that needs to check if there's a transaction doesn't get
7674                  * confused.
7675                  */
7676                 dio_data = current->journal_info;
7677                 current->journal_info = NULL;
7678         }
7679
7680         /*
7681          * If this errors out it's because we couldn't invalidate pagecache for
7682          * this range and we need to fallback to buffered.
7683          */
7684         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7685                                create)) {
7686                 ret = -ENOTBLK;
7687                 goto err;
7688         }
7689
7690         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, start, len, 0);
7691         if (IS_ERR(em)) {
7692                 ret = PTR_ERR(em);
7693                 goto unlock_err;
7694         }
7695
7696         /*
7697          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7698          * io.  INLINE is special, and we could probably kludge it in here, but
7699          * it's still buffered so for safety lets just fall back to the generic
7700          * buffered path.
7701          *
7702          * For COMPRESSED we _have_ to read the entire extent in so we can
7703          * decompress it, so there will be buffering required no matter what we
7704          * do, so go ahead and fallback to buffered.
7705          *
7706          * We return -ENOTBLK because that's what makes DIO go ahead and go back
7707          * to buffered IO.  Don't blame me, this is the price we pay for using
7708          * the generic code.
7709          */
7710         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7711             em->block_start == EXTENT_MAP_INLINE) {
7712                 free_extent_map(em);
7713                 ret = -ENOTBLK;
7714                 goto unlock_err;
7715         }
7716
7717         /* Just a good old fashioned hole, return */
7718         if (!create && (em->block_start == EXTENT_MAP_HOLE ||
7719                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
7720                 free_extent_map(em);
7721                 goto unlock_err;
7722         }
7723
7724         /*
7725          * We don't allocate a new extent in the following cases
7726          *
7727          * 1) The inode is marked as NODATACOW.  In this case we'll just use the
7728          * existing extent.
7729          * 2) The extent is marked as PREALLOC.  We're good to go here and can
7730          * just use the extent.
7731          *
7732          */
7733         if (!create) {
7734                 len = min(len, em->len - (start - em->start));
7735                 lockstart = start + len;
7736                 goto unlock;
7737         }
7738
7739         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7740             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7741              em->block_start != EXTENT_MAP_HOLE)) {
7742                 int type;
7743                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7744
7745                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7746                         type = BTRFS_ORDERED_PREALLOC;
7747                 else
7748                         type = BTRFS_ORDERED_NOCOW;
7749                 len = min(len, em->len - (start - em->start));
7750                 block_start = em->block_start + (start - em->start);
7751
7752                 if (can_nocow_extent(inode, start, &len, &orig_start,
7753                                      &orig_block_len, &ram_bytes) == 1 &&
7754                     btrfs_inc_nocow_writers(fs_info, block_start)) {
7755                         struct extent_map *em2;
7756
7757                         em2 = btrfs_create_dio_extent(inode, start, len,
7758                                                       orig_start, block_start,
7759                                                       len, orig_block_len,
7760                                                       ram_bytes, type);
7761                         btrfs_dec_nocow_writers(fs_info, block_start);
7762                         if (type == BTRFS_ORDERED_PREALLOC) {
7763                                 free_extent_map(em);
7764                                 em = em2;
7765                         }
7766                         if (em2 && IS_ERR(em2)) {
7767                                 ret = PTR_ERR(em2);
7768                                 goto unlock_err;
7769                         }
7770                         /*
7771                          * For inode marked NODATACOW or extent marked PREALLOC,
7772                          * use the existing or preallocated extent, so does not
7773                          * need to adjust btrfs_space_info's bytes_may_use.
7774                          */
7775                         btrfs_free_reserved_data_space_noquota(inode,
7776                                         start, len);
7777                         goto unlock;
7778                 }
7779         }
7780
7781         /*
7782          * this will cow the extent, reset the len in case we changed
7783          * it above
7784          */
7785         len = bh_result->b_size;
7786         free_extent_map(em);
7787         em = btrfs_new_extent_direct(inode, start, len);
7788         if (IS_ERR(em)) {
7789                 ret = PTR_ERR(em);
7790                 goto unlock_err;
7791         }
7792         len = min(len, em->len - (start - em->start));
7793 unlock:
7794         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7795                 inode->i_blkbits;
7796         bh_result->b_size = len;
7797         bh_result->b_bdev = em->bdev;
7798         set_buffer_mapped(bh_result);
7799         if (create) {
7800                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7801                         set_buffer_new(bh_result);
7802
7803                 /*
7804                  * Need to update the i_size under the extent lock so buffered
7805                  * readers will get the updated i_size when we unlock.
7806                  */
7807                 if (!dio_data->overwrite && start + len > i_size_read(inode))
7808                         i_size_write(inode, start + len);
7809
7810                 WARN_ON(dio_data->reserve < len);
7811                 dio_data->reserve -= len;
7812                 dio_data->unsubmitted_oe_range_end = start + len;
7813                 current->journal_info = dio_data;
7814         }
7815
7816         /*
7817          * In the case of write we need to clear and unlock the entire range,
7818          * in the case of read we need to unlock only the end area that we
7819          * aren't using if there is any left over space.
7820          */
7821         if (lockstart < lockend) {
7822                 clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
7823                                  lockend, unlock_bits, 1, 0,
7824                                  &cached_state);
7825         } else {
7826                 free_extent_state(cached_state);
7827         }
7828
7829         free_extent_map(em);
7830
7831         return 0;
7832
7833 unlock_err:
7834         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7835                          unlock_bits, 1, 0, &cached_state);
7836 err:
7837         if (dio_data)
7838                 current->journal_info = dio_data;
7839         return ret;
7840 }
7841
7842 static inline blk_status_t submit_dio_repair_bio(struct inode *inode,
7843                                                  struct bio *bio,
7844                                                  int mirror_num)
7845 {
7846         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7847         blk_status_t ret;
7848
7849         BUG_ON(bio_op(bio) == REQ_OP_WRITE);
7850
7851         ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DIO_REPAIR);
7852         if (ret)
7853                 return ret;
7854
7855         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
7856
7857         return ret;
7858 }
7859
7860 static int btrfs_check_dio_repairable(struct inode *inode,
7861                                       struct bio *failed_bio,
7862                                       struct io_failure_record *failrec,
7863                                       int failed_mirror)
7864 {
7865         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7866         int num_copies;
7867
7868         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
7869         if (num_copies == 1) {
7870                 /*
7871                  * we only have a single copy of the data, so don't bother with
7872                  * all the retry and error correction code that follows. no
7873                  * matter what the error is, it is very likely to persist.
7874                  */
7875                 btrfs_debug(fs_info,
7876                         "Check DIO Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
7877                         num_copies, failrec->this_mirror, failed_mirror);
7878                 return 0;
7879         }
7880
7881         failrec->failed_mirror = failed_mirror;
7882         failrec->this_mirror++;
7883         if (failrec->this_mirror == failed_mirror)
7884                 failrec->this_mirror++;
7885
7886         if (failrec->this_mirror > num_copies) {
7887                 btrfs_debug(fs_info,
7888                         "Check DIO Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
7889                         num_copies, failrec->this_mirror, failed_mirror);
7890                 return 0;
7891         }
7892
7893         return 1;
7894 }
7895
7896 static blk_status_t dio_read_error(struct inode *inode, struct bio *failed_bio,
7897                                    struct page *page, unsigned int pgoff,
7898                                    u64 start, u64 end, int failed_mirror,
7899                                    bio_end_io_t *repair_endio, void *repair_arg)
7900 {
7901         struct io_failure_record *failrec;
7902         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7903         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
7904         struct bio *bio;
7905         int isector;
7906         unsigned int read_mode = 0;
7907         int segs;
7908         int ret;
7909         blk_status_t status;
7910         struct bio_vec bvec;
7911
7912         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
7913
7914         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
7915         if (ret)
7916                 return errno_to_blk_status(ret);
7917
7918         ret = btrfs_check_dio_repairable(inode, failed_bio, failrec,
7919                                          failed_mirror);
7920         if (!ret) {
7921                 free_io_failure(failure_tree, io_tree, failrec);
7922                 return BLK_STS_IOERR;
7923         }
7924
7925         segs = bio_segments(failed_bio);
7926         bio_get_first_bvec(failed_bio, &bvec);
7927         if (segs > 1 ||
7928             (bvec.bv_len > btrfs_inode_sectorsize(inode)))
7929                 read_mode |= REQ_FAILFAST_DEV;
7930
7931         isector = start - btrfs_io_bio(failed_bio)->logical;
7932         isector >>= inode->i_sb->s_blocksize_bits;
7933         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
7934                                 pgoff, isector, repair_endio, repair_arg);
7935         bio_set_op_attrs(bio, REQ_OP_READ, read_mode);
7936
7937         btrfs_debug(BTRFS_I(inode)->root->fs_info,
7938                     "repair DIO read error: submitting new dio read[%#x] to this_mirror=%d, in_validation=%d",
7939                     read_mode, failrec->this_mirror, failrec->in_validation);
7940
7941         status = submit_dio_repair_bio(inode, bio, failrec->this_mirror);
7942         if (status) {
7943                 free_io_failure(failure_tree, io_tree, failrec);
7944                 bio_put(bio);
7945         }
7946
7947         return status;
7948 }
7949
7950 struct btrfs_retry_complete {
7951         struct completion done;
7952         struct inode *inode;
7953         u64 start;
7954         int uptodate;
7955 };
7956
7957 static void btrfs_retry_endio_nocsum(struct bio *bio)
7958 {
7959         struct btrfs_retry_complete *done = bio->bi_private;
7960         struct inode *inode = done->inode;
7961         struct bio_vec *bvec;
7962         struct extent_io_tree *io_tree, *failure_tree;
7963         int i;
7964
7965         if (bio->bi_status)
7966                 goto end;
7967
7968         ASSERT(bio->bi_vcnt == 1);
7969         io_tree = &BTRFS_I(inode)->io_tree;
7970         failure_tree = &BTRFS_I(inode)->io_failure_tree;
7971         ASSERT(bio_first_bvec_all(bio)->bv_len == btrfs_inode_sectorsize(inode));
7972
7973         done->uptodate = 1;
7974         ASSERT(!bio_flagged(bio, BIO_CLONED));
7975         bio_for_each_segment_all(bvec, bio, i)
7976                 clean_io_failure(BTRFS_I(inode)->root->fs_info, failure_tree,
7977                                  io_tree, done->start, bvec->bv_page,
7978                                  btrfs_ino(BTRFS_I(inode)), 0);
7979 end:
7980         complete(&done->done);
7981         bio_put(bio);
7982 }
7983
7984 static blk_status_t __btrfs_correct_data_nocsum(struct inode *inode,
7985                                                 struct btrfs_io_bio *io_bio)
7986 {
7987         struct btrfs_fs_info *fs_info;
7988         struct bio_vec bvec;
7989         struct bvec_iter iter;
7990         struct btrfs_retry_complete done;
7991         u64 start;
7992         unsigned int pgoff;
7993         u32 sectorsize;
7994         int nr_sectors;
7995         blk_status_t ret;
7996         blk_status_t err = BLK_STS_OK;
7997
7998         fs_info = BTRFS_I(inode)->root->fs_info;
7999         sectorsize = fs_info->sectorsize;
8000
8001         start = io_bio->logical;
8002         done.inode = inode;
8003         io_bio->bio.bi_iter = io_bio->iter;
8004
8005         bio_for_each_segment(bvec, &io_bio->bio, iter) {
8006                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
8007                 pgoff = bvec.bv_offset;
8008
8009 next_block_or_try_again:
8010                 done.uptodate = 0;
8011                 done.start = start;
8012                 init_completion(&done.done);
8013
8014                 ret = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
8015                                 pgoff, start, start + sectorsize - 1,
8016                                 io_bio->mirror_num,
8017                                 btrfs_retry_endio_nocsum, &done);
8018                 if (ret) {
8019                         err = ret;
8020                         goto next;
8021                 }
8022
8023                 wait_for_completion_io(&done.done);
8024
8025                 if (!done.uptodate) {
8026                         /* We might have another mirror, so try again */
8027                         goto next_block_or_try_again;
8028                 }
8029
8030 next:
8031                 start += sectorsize;
8032
8033                 nr_sectors--;
8034                 if (nr_sectors) {
8035                         pgoff += sectorsize;
8036                         ASSERT(pgoff < PAGE_SIZE);
8037                         goto next_block_or_try_again;
8038                 }
8039         }
8040
8041         return err;
8042 }
8043
8044 static void btrfs_retry_endio(struct bio *bio)
8045 {
8046         struct btrfs_retry_complete *done = bio->bi_private;
8047         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8048         struct extent_io_tree *io_tree, *failure_tree;
8049         struct inode *inode = done->inode;
8050         struct bio_vec *bvec;
8051         int uptodate;
8052         int ret;
8053         int i;
8054
8055         if (bio->bi_status)
8056                 goto end;
8057
8058         uptodate = 1;
8059
8060         ASSERT(bio->bi_vcnt == 1);
8061         ASSERT(bio_first_bvec_all(bio)->bv_len == btrfs_inode_sectorsize(done->inode));
8062
8063         io_tree = &BTRFS_I(inode)->io_tree;
8064         failure_tree = &BTRFS_I(inode)->io_failure_tree;
8065
8066         ASSERT(!bio_flagged(bio, BIO_CLONED));
8067         bio_for_each_segment_all(bvec, bio, i) {
8068                 ret = __readpage_endio_check(inode, io_bio, i, bvec->bv_page,
8069                                              bvec->bv_offset, done->start,
8070                                              bvec->bv_len);
8071                 if (!ret)
8072                         clean_io_failure(BTRFS_I(inode)->root->fs_info,
8073                                          failure_tree, io_tree, done->start,
8074                                          bvec->bv_page,
8075                                          btrfs_ino(BTRFS_I(inode)),
8076                                          bvec->bv_offset);
8077                 else
8078                         uptodate = 0;
8079         }
8080
8081         done->uptodate = uptodate;
8082 end:
8083         complete(&done->done);
8084         bio_put(bio);
8085 }
8086
8087 static blk_status_t __btrfs_subio_endio_read(struct inode *inode,
8088                 struct btrfs_io_bio *io_bio, blk_status_t err)
8089 {
8090         struct btrfs_fs_info *fs_info;
8091         struct bio_vec bvec;
8092         struct bvec_iter iter;
8093         struct btrfs_retry_complete done;
8094         u64 start;
8095         u64 offset = 0;
8096         u32 sectorsize;
8097         int nr_sectors;
8098         unsigned int pgoff;
8099         int csum_pos;
8100         bool uptodate = (err == 0);
8101         int ret;
8102         blk_status_t status;
8103
8104         fs_info = BTRFS_I(inode)->root->fs_info;
8105         sectorsize = fs_info->sectorsize;
8106
8107         err = BLK_STS_OK;
8108         start = io_bio->logical;
8109         done.inode = inode;
8110         io_bio->bio.bi_iter = io_bio->iter;
8111
8112         bio_for_each_segment(bvec, &io_bio->bio, iter) {
8113                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
8114
8115                 pgoff = bvec.bv_offset;
8116 next_block:
8117                 if (uptodate) {
8118                         csum_pos = BTRFS_BYTES_TO_BLKS(fs_info, offset);
8119                         ret = __readpage_endio_check(inode, io_bio, csum_pos,
8120                                         bvec.bv_page, pgoff, start, sectorsize);
8121                         if (likely(!ret))
8122                                 goto next;
8123                 }
8124 try_again:
8125                 done.uptodate = 0;
8126                 done.start = start;
8127                 init_completion(&done.done);
8128
8129                 status = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
8130                                         pgoff, start, start + sectorsize - 1,
8131                                         io_bio->mirror_num, btrfs_retry_endio,
8132                                         &done);
8133                 if (status) {
8134                         err = status;
8135                         goto next;
8136                 }
8137
8138                 wait_for_completion_io(&done.done);
8139
8140                 if (!done.uptodate) {
8141                         /* We might have another mirror, so try again */
8142                         goto try_again;
8143                 }
8144 next:
8145                 offset += sectorsize;
8146                 start += sectorsize;
8147
8148                 ASSERT(nr_sectors);
8149
8150                 nr_sectors--;
8151                 if (nr_sectors) {
8152                         pgoff += sectorsize;
8153                         ASSERT(pgoff < PAGE_SIZE);
8154                         goto next_block;
8155                 }
8156         }
8157
8158         return err;
8159 }
8160
8161 static blk_status_t btrfs_subio_endio_read(struct inode *inode,
8162                 struct btrfs_io_bio *io_bio, blk_status_t err)
8163 {
8164         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8165
8166         if (skip_csum) {
8167                 if (unlikely(err))
8168                         return __btrfs_correct_data_nocsum(inode, io_bio);
8169                 else
8170                         return BLK_STS_OK;
8171         } else {
8172                 return __btrfs_subio_endio_read(inode, io_bio, err);
8173         }
8174 }
8175
8176 static void btrfs_endio_direct_read(struct bio *bio)
8177 {
8178         struct btrfs_dio_private *dip = bio->bi_private;
8179         struct inode *inode = dip->inode;
8180         struct bio *dio_bio;
8181         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8182         blk_status_t err = bio->bi_status;
8183
8184         if (dip->flags & BTRFS_DIO_ORIG_BIO_SUBMITTED)
8185                 err = btrfs_subio_endio_read(inode, io_bio, err);
8186
8187         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
8188                       dip->logical_offset + dip->bytes - 1);
8189         dio_bio = dip->dio_bio;
8190
8191         kfree(dip);
8192
8193         dio_bio->bi_status = err;
8194         dio_end_io(dio_bio);
8195
8196         if (io_bio->end_io)
8197                 io_bio->end_io(io_bio, blk_status_to_errno(err));
8198         bio_put(bio);
8199 }
8200
8201 static void __endio_write_update_ordered(struct inode *inode,
8202                                          const u64 offset, const u64 bytes,
8203                                          const bool uptodate)
8204 {
8205         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8206         struct btrfs_ordered_extent *ordered = NULL;
8207         struct btrfs_workqueue *wq;
8208         btrfs_work_func_t func;
8209         u64 ordered_offset = offset;
8210         u64 ordered_bytes = bytes;
8211         u64 last_offset;
8212         int ret;
8213
8214         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
8215                 wq = fs_info->endio_freespace_worker;
8216                 func = btrfs_freespace_write_helper;
8217         } else {
8218                 wq = fs_info->endio_write_workers;
8219                 func = btrfs_endio_write_helper;
8220         }
8221
8222 again:
8223         last_offset = ordered_offset;
8224         ret = btrfs_dec_test_first_ordered_pending(inode, &ordered,
8225                                                    &ordered_offset,
8226                                                    ordered_bytes,
8227                                                    uptodate);
8228         if (!ret)
8229                 goto out_test;
8230
8231         btrfs_init_work(&ordered->work, func, finish_ordered_fn, NULL, NULL);
8232         btrfs_queue_work(wq, &ordered->work);
8233 out_test:
8234         /*
8235          * If btrfs_dec_test_ordered_pending does not find any ordered extent
8236          * in the range, we can exit.
8237          */
8238         if (ordered_offset == last_offset)
8239                 return;
8240         /*
8241          * our bio might span multiple ordered extents.  If we haven't
8242          * completed the accounting for the whole dio, go back and try again
8243          */
8244         if (ordered_offset < offset + bytes) {
8245                 ordered_bytes = offset + bytes - ordered_offset;
8246                 ordered = NULL;
8247                 goto again;
8248         }
8249 }
8250
8251 static void btrfs_endio_direct_write(struct bio *bio)
8252 {
8253         struct btrfs_dio_private *dip = bio->bi_private;
8254         struct bio *dio_bio = dip->dio_bio;
8255
8256         __endio_write_update_ordered(dip->inode, dip->logical_offset,
8257                                      dip->bytes, !bio->bi_status);
8258
8259         kfree(dip);
8260
8261         dio_bio->bi_status = bio->bi_status;
8262         dio_end_io(dio_bio);
8263         bio_put(bio);
8264 }
8265
8266 static blk_status_t __btrfs_submit_bio_start_direct_io(void *private_data,
8267                                     struct bio *bio, int mirror_num,
8268                                     unsigned long bio_flags, u64 offset)
8269 {
8270         struct inode *inode = private_data;
8271         blk_status_t ret;
8272         ret = btrfs_csum_one_bio(inode, bio, offset, 1);
8273         BUG_ON(ret); /* -ENOMEM */
8274         return 0;
8275 }
8276
8277 static void btrfs_end_dio_bio(struct bio *bio)
8278 {
8279         struct btrfs_dio_private *dip = bio->bi_private;
8280         blk_status_t err = bio->bi_status;
8281
8282         if (err)
8283                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
8284                            "direct IO failed ino %llu rw %d,%u sector %#Lx len %u err no %d",
8285                            btrfs_ino(BTRFS_I(dip->inode)), bio_op(bio),
8286                            bio->bi_opf,
8287                            (unsigned long long)bio->bi_iter.bi_sector,
8288                            bio->bi_iter.bi_size, err);
8289
8290         if (dip->subio_endio)
8291                 err = dip->subio_endio(dip->inode, btrfs_io_bio(bio), err);
8292
8293         if (err) {
8294                 dip->errors = 1;
8295
8296                 /*
8297                  * before atomic variable goto zero, we must make sure
8298                  * dip->errors is perceived to be set.
8299                  */
8300                 smp_mb__before_atomic();
8301         }
8302
8303         /* if there are more bios still pending for this dio, just exit */
8304         if (!atomic_dec_and_test(&dip->pending_bios))
8305                 goto out;
8306
8307         if (dip->errors) {
8308                 bio_io_error(dip->orig_bio);
8309         } else {
8310                 dip->dio_bio->bi_status = BLK_STS_OK;
8311                 bio_endio(dip->orig_bio);
8312         }
8313 out:
8314         bio_put(bio);
8315 }
8316
8317 static inline blk_status_t btrfs_lookup_and_bind_dio_csum(struct inode *inode,
8318                                                  struct btrfs_dio_private *dip,
8319                                                  struct bio *bio,
8320                                                  u64 file_offset)
8321 {
8322         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8323         struct btrfs_io_bio *orig_io_bio = btrfs_io_bio(dip->orig_bio);
8324         blk_status_t ret;
8325
8326         /*
8327          * We load all the csum data we need when we submit
8328          * the first bio to reduce the csum tree search and
8329          * contention.
8330          */
8331         if (dip->logical_offset == file_offset) {
8332                 ret = btrfs_lookup_bio_sums_dio(inode, dip->orig_bio,
8333                                                 file_offset);
8334                 if (ret)
8335                         return ret;
8336         }
8337
8338         if (bio == dip->orig_bio)
8339                 return 0;
8340
8341         file_offset -= dip->logical_offset;
8342         file_offset >>= inode->i_sb->s_blocksize_bits;
8343         io_bio->csum = (u8 *)(((u32 *)orig_io_bio->csum) + file_offset);
8344
8345         return 0;
8346 }
8347
8348 static inline blk_status_t
8349 __btrfs_submit_dio_bio(struct bio *bio, struct inode *inode, u64 file_offset,
8350                        int async_submit)
8351 {
8352         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8353         struct btrfs_dio_private *dip = bio->bi_private;
8354         bool write = bio_op(bio) == REQ_OP_WRITE;
8355         blk_status_t ret;
8356
8357         /* Check btrfs_submit_bio_hook() for rules about async submit. */
8358         if (async_submit)
8359                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
8360
8361         if (!write) {
8362                 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
8363                 if (ret)
8364                         goto err;
8365         }
8366
8367         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
8368                 goto map;
8369
8370         if (write && async_submit) {
8371                 ret = btrfs_wq_submit_bio(fs_info, bio, 0, 0,
8372                                           file_offset, inode,
8373                                           __btrfs_submit_bio_start_direct_io,
8374                                           __btrfs_submit_bio_done);
8375                 goto err;
8376         } else if (write) {
8377                 /*
8378                  * If we aren't doing async submit, calculate the csum of the
8379                  * bio now.
8380                  */
8381                 ret = btrfs_csum_one_bio(inode, bio, file_offset, 1);
8382                 if (ret)
8383                         goto err;
8384         } else {
8385                 ret = btrfs_lookup_and_bind_dio_csum(inode, dip, bio,
8386                                                      file_offset);
8387                 if (ret)
8388                         goto err;
8389         }
8390 map:
8391         ret = btrfs_map_bio(fs_info, bio, 0, 0);
8392 err:
8393         return ret;
8394 }
8395
8396 static int btrfs_submit_direct_hook(struct btrfs_dio_private *dip)
8397 {
8398         struct inode *inode = dip->inode;
8399         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8400         struct bio *bio;
8401         struct bio *orig_bio = dip->orig_bio;
8402         u64 start_sector = orig_bio->bi_iter.bi_sector;
8403         u64 file_offset = dip->logical_offset;
8404         u64 map_length;
8405         int async_submit = 0;
8406         u64 submit_len;
8407         int clone_offset = 0;
8408         int clone_len;
8409         int ret;
8410         blk_status_t status;
8411
8412         map_length = orig_bio->bi_iter.bi_size;
8413         submit_len = map_length;
8414         ret = btrfs_map_block(fs_info, btrfs_op(orig_bio), start_sector << 9,
8415                               &map_length, NULL, 0);
8416         if (ret)
8417                 return -EIO;
8418
8419         if (map_length >= submit_len) {
8420                 bio = orig_bio;
8421                 dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
8422                 goto submit;
8423         }
8424
8425         /* async crcs make it difficult to collect full stripe writes. */
8426         if (btrfs_data_alloc_profile(fs_info) & BTRFS_BLOCK_GROUP_RAID56_MASK)
8427                 async_submit = 0;
8428         else
8429                 async_submit = 1;
8430
8431         /* bio split */
8432         ASSERT(map_length <= INT_MAX);
8433         atomic_inc(&dip->pending_bios);
8434         do {
8435                 clone_len = min_t(int, submit_len, map_length);
8436
8437                 /*
8438                  * This will never fail as it's passing GPF_NOFS and
8439                  * the allocation is backed by btrfs_bioset.
8440                  */
8441                 bio = btrfs_bio_clone_partial(orig_bio, clone_offset,
8442                                               clone_len);
8443                 bio->bi_private = dip;
8444                 bio->bi_end_io = btrfs_end_dio_bio;
8445                 btrfs_io_bio(bio)->logical = file_offset;
8446
8447                 ASSERT(submit_len >= clone_len);
8448                 submit_len -= clone_len;
8449                 if (submit_len == 0)
8450                         break;
8451
8452                 /*
8453                  * Increase the count before we submit the bio so we know
8454                  * the end IO handler won't happen before we increase the
8455                  * count. Otherwise, the dip might get freed before we're
8456                  * done setting it up.
8457                  */
8458                 atomic_inc(&dip->pending_bios);
8459
8460                 status = __btrfs_submit_dio_bio(bio, inode, file_offset,
8461                                                 async_submit);
8462                 if (status) {
8463                         bio_put(bio);
8464                         atomic_dec(&dip->pending_bios);
8465                         goto out_err;
8466                 }
8467
8468                 clone_offset += clone_len;
8469                 start_sector += clone_len >> 9;
8470                 file_offset += clone_len;
8471
8472                 map_length = submit_len;
8473                 ret = btrfs_map_block(fs_info, btrfs_op(orig_bio),
8474                                       start_sector << 9, &map_length, NULL, 0);
8475                 if (ret)
8476                         goto out_err;
8477         } while (submit_len > 0);
8478
8479 submit:
8480         status = __btrfs_submit_dio_bio(bio, inode, file_offset, async_submit);
8481         if (!status)
8482                 return 0;
8483
8484         bio_put(bio);
8485 out_err:
8486         dip->errors = 1;
8487         /*
8488          * before atomic variable goto zero, we must
8489          * make sure dip->errors is perceived to be set.
8490          */
8491         smp_mb__before_atomic();
8492         if (atomic_dec_and_test(&dip->pending_bios))
8493                 bio_io_error(dip->orig_bio);
8494
8495         /* bio_end_io() will handle error, so we needn't return it */
8496         return 0;
8497 }
8498
8499 static void btrfs_submit_direct(struct bio *dio_bio, struct inode *inode,
8500                                 loff_t file_offset)
8501 {
8502         struct btrfs_dio_private *dip = NULL;
8503         struct bio *bio = NULL;
8504         struct btrfs_io_bio *io_bio;
8505         bool write = (bio_op(dio_bio) == REQ_OP_WRITE);
8506         int ret = 0;
8507
8508         bio = btrfs_bio_clone(dio_bio);
8509
8510         dip = kzalloc(sizeof(*dip), GFP_NOFS);
8511         if (!dip) {
8512                 ret = -ENOMEM;
8513                 goto free_ordered;
8514         }
8515
8516         dip->private = dio_bio->bi_private;
8517         dip->inode = inode;
8518         dip->logical_offset = file_offset;
8519         dip->bytes = dio_bio->bi_iter.bi_size;
8520         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
8521         bio->bi_private = dip;
8522         dip->orig_bio = bio;
8523         dip->dio_bio = dio_bio;
8524         atomic_set(&dip->pending_bios, 0);
8525         io_bio = btrfs_io_bio(bio);
8526         io_bio->logical = file_offset;
8527
8528         if (write) {
8529                 bio->bi_end_io = btrfs_endio_direct_write;
8530         } else {
8531                 bio->bi_end_io = btrfs_endio_direct_read;
8532                 dip->subio_endio = btrfs_subio_endio_read;
8533         }
8534
8535         /*
8536          * Reset the range for unsubmitted ordered extents (to a 0 length range)
8537          * even if we fail to submit a bio, because in such case we do the
8538          * corresponding error handling below and it must not be done a second
8539          * time by btrfs_direct_IO().
8540          */
8541         if (write) {
8542                 struct btrfs_dio_data *dio_data = current->journal_info;
8543
8544                 dio_data->unsubmitted_oe_range_end = dip->logical_offset +
8545                         dip->bytes;
8546                 dio_data->unsubmitted_oe_range_start =
8547                         dio_data->unsubmitted_oe_range_end;
8548         }
8549
8550         ret = btrfs_submit_direct_hook(dip);
8551         if (!ret)
8552                 return;
8553
8554         if (io_bio->end_io)
8555                 io_bio->end_io(io_bio, ret);
8556
8557 free_ordered:
8558         /*
8559          * If we arrived here it means either we failed to submit the dip
8560          * or we either failed to clone the dio_bio or failed to allocate the
8561          * dip. If we cloned the dio_bio and allocated the dip, we can just
8562          * call bio_endio against our io_bio so that we get proper resource
8563          * cleanup if we fail to submit the dip, otherwise, we must do the
8564          * same as btrfs_endio_direct_[write|read] because we can't call these
8565          * callbacks - they require an allocated dip and a clone of dio_bio.
8566          */
8567         if (bio && dip) {
8568                 bio_io_error(bio);
8569                 /*
8570                  * The end io callbacks free our dip, do the final put on bio
8571                  * and all the cleanup and final put for dio_bio (through
8572                  * dio_end_io()).
8573                  */
8574                 dip = NULL;
8575                 bio = NULL;
8576         } else {
8577                 if (write)
8578                         __endio_write_update_ordered(inode,
8579                                                 file_offset,
8580                                                 dio_bio->bi_iter.bi_size,
8581                                                 false);
8582                 else
8583                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
8584                               file_offset + dio_bio->bi_iter.bi_size - 1);
8585
8586                 dio_bio->bi_status = BLK_STS_IOERR;
8587                 /*
8588                  * Releases and cleans up our dio_bio, no need to bio_put()
8589                  * nor bio_endio()/bio_io_error() against dio_bio.
8590                  */
8591                 dio_end_io(dio_bio);
8592         }
8593         if (bio)
8594                 bio_put(bio);
8595         kfree(dip);
8596 }
8597
8598 static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
8599                                const struct iov_iter *iter, loff_t offset)
8600 {
8601         int seg;
8602         int i;
8603         unsigned int blocksize_mask = fs_info->sectorsize - 1;
8604         ssize_t retval = -EINVAL;
8605
8606         if (offset & blocksize_mask)
8607                 goto out;
8608
8609         if (iov_iter_alignment(iter) & blocksize_mask)
8610                 goto out;
8611
8612         /* If this is a write we don't need to check anymore */
8613         if (iov_iter_rw(iter) != READ || !iter_is_iovec(iter))
8614                 return 0;
8615         /*
8616          * Check to make sure we don't have duplicate iov_base's in this
8617          * iovec, if so return EINVAL, otherwise we'll get csum errors
8618          * when reading back.
8619          */
8620         for (seg = 0; seg < iter->nr_segs; seg++) {
8621                 for (i = seg + 1; i < iter->nr_segs; i++) {
8622                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
8623                                 goto out;
8624                 }
8625         }
8626         retval = 0;
8627 out:
8628         return retval;
8629 }
8630
8631 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
8632 {
8633         struct file *file = iocb->ki_filp;
8634         struct inode *inode = file->f_mapping->host;
8635         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8636         struct btrfs_dio_data dio_data = { 0 };
8637         struct extent_changeset *data_reserved = NULL;
8638         loff_t offset = iocb->ki_pos;
8639         size_t count = 0;
8640         int flags = 0;
8641         bool wakeup = true;
8642         bool relock = false;
8643         ssize_t ret;
8644
8645         if (check_direct_IO(fs_info, iter, offset))
8646                 return 0;
8647
8648         inode_dio_begin(inode);
8649
8650         /*
8651          * The generic stuff only does filemap_write_and_wait_range, which
8652          * isn't enough if we've written compressed pages to this area, so
8653          * we need to flush the dirty pages again to make absolutely sure
8654          * that any outstanding dirty pages are on disk.
8655          */
8656         count = iov_iter_count(iter);
8657         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8658                      &BTRFS_I(inode)->runtime_flags))
8659                 filemap_fdatawrite_range(inode->i_mapping, offset,
8660                                          offset + count - 1);
8661
8662         if (iov_iter_rw(iter) == WRITE) {
8663                 /*
8664                  * If the write DIO is beyond the EOF, we need update
8665                  * the isize, but it is protected by i_mutex. So we can
8666                  * not unlock the i_mutex at this case.
8667                  */
8668                 if (offset + count <= inode->i_size) {
8669                         dio_data.overwrite = 1;
8670                         inode_unlock(inode);
8671                         relock = true;
8672                 } else if (iocb->ki_flags & IOCB_NOWAIT) {
8673                         ret = -EAGAIN;
8674                         goto out;
8675                 }
8676                 ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
8677                                                    offset, count);
8678                 if (ret)
8679                         goto out;
8680
8681                 /*
8682                  * We need to know how many extents we reserved so that we can
8683                  * do the accounting properly if we go over the number we
8684                  * originally calculated.  Abuse current->journal_info for this.
8685                  */
8686                 dio_data.reserve = round_up(count,
8687                                             fs_info->sectorsize);
8688                 dio_data.unsubmitted_oe_range_start = (u64)offset;
8689                 dio_data.unsubmitted_oe_range_end = (u64)offset;
8690                 current->journal_info = &dio_data;
8691                 down_read(&BTRFS_I(inode)->dio_sem);
8692         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
8693                                      &BTRFS_I(inode)->runtime_flags)) {
8694                 inode_dio_end(inode);
8695                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
8696                 wakeup = false;
8697         }
8698
8699         ret = __blockdev_direct_IO(iocb, inode,
8700                                    fs_info->fs_devices->latest_bdev,
8701                                    iter, btrfs_get_blocks_direct, NULL,
8702                                    btrfs_submit_direct, flags);
8703         if (iov_iter_rw(iter) == WRITE) {
8704                 up_read(&BTRFS_I(inode)->dio_sem);
8705                 current->journal_info = NULL;
8706                 if (ret < 0 && ret != -EIOCBQUEUED) {
8707                         if (dio_data.reserve)
8708                                 btrfs_delalloc_release_space(inode, data_reserved,
8709                                         offset, dio_data.reserve);
8710                         /*
8711                          * On error we might have left some ordered extents
8712                          * without submitting corresponding bios for them, so
8713                          * cleanup them up to avoid other tasks getting them
8714                          * and waiting for them to complete forever.
8715                          */
8716                         if (dio_data.unsubmitted_oe_range_start <
8717                             dio_data.unsubmitted_oe_range_end)
8718                                 __endio_write_update_ordered(inode,
8719                                         dio_data.unsubmitted_oe_range_start,
8720                                         dio_data.unsubmitted_oe_range_end -
8721                                         dio_data.unsubmitted_oe_range_start,
8722                                         false);
8723                 } else if (ret >= 0 && (size_t)ret < count)
8724                         btrfs_delalloc_release_space(inode, data_reserved,
8725                                         offset, count - (size_t)ret);
8726                 btrfs_delalloc_release_extents(BTRFS_I(inode), count);
8727         }
8728 out:
8729         if (wakeup)
8730                 inode_dio_end(inode);
8731         if (relock)
8732                 inode_lock(inode);
8733
8734         extent_changeset_free(data_reserved);
8735         return ret;
8736 }
8737
8738 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
8739
8740 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
8741                 __u64 start, __u64 len)
8742 {
8743         int     ret;
8744
8745         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
8746         if (ret)
8747                 return ret;
8748
8749         return extent_fiemap(inode, fieinfo, start, len);
8750 }
8751
8752 int btrfs_readpage(struct file *file, struct page *page)
8753 {
8754         struct extent_io_tree *tree;
8755         tree = &BTRFS_I(page->mapping->host)->io_tree;
8756         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
8757 }
8758
8759 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
8760 {
8761         struct inode *inode = page->mapping->host;
8762         int ret;
8763
8764         if (current->flags & PF_MEMALLOC) {
8765                 redirty_page_for_writepage(wbc, page);
8766                 unlock_page(page);
8767                 return 0;
8768         }
8769
8770         /*
8771          * If we are under memory pressure we will call this directly from the
8772          * VM, we need to make sure we have the inode referenced for the ordered
8773          * extent.  If not just return like we didn't do anything.
8774          */
8775         if (!igrab(inode)) {
8776                 redirty_page_for_writepage(wbc, page);
8777                 return AOP_WRITEPAGE_ACTIVATE;
8778         }
8779         ret = extent_write_full_page(page, wbc);
8780         btrfs_add_delayed_iput(inode);
8781         return ret;
8782 }
8783
8784 static int btrfs_writepages(struct address_space *mapping,
8785                             struct writeback_control *wbc)
8786 {
8787         struct extent_io_tree *tree;
8788
8789         tree = &BTRFS_I(mapping->host)->io_tree;
8790         return extent_writepages(tree, mapping, wbc);
8791 }
8792
8793 static int
8794 btrfs_readpages(struct file *file, struct address_space *mapping,
8795                 struct list_head *pages, unsigned nr_pages)
8796 {
8797         struct extent_io_tree *tree;
8798         tree = &BTRFS_I(mapping->host)->io_tree;
8799         return extent_readpages(tree, mapping, pages, nr_pages);
8800 }
8801 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8802 {
8803         struct extent_io_tree *tree;
8804         struct extent_map_tree *map;
8805         int ret;
8806
8807         tree = &BTRFS_I(page->mapping->host)->io_tree;
8808         map = &BTRFS_I(page->mapping->host)->extent_tree;
8809         ret = try_release_extent_mapping(map, tree, page, gfp_flags);
8810         if (ret == 1) {
8811                 ClearPagePrivate(page);
8812                 set_page_private(page, 0);
8813                 put_page(page);
8814         }
8815         return ret;
8816 }
8817
8818 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8819 {
8820         if (PageWriteback(page) || PageDirty(page))
8821                 return 0;
8822         return __btrfs_releasepage(page, gfp_flags);
8823 }
8824
8825 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
8826                                  unsigned int length)
8827 {
8828         struct inode *inode = page->mapping->host;
8829         struct extent_io_tree *tree;
8830         struct btrfs_ordered_extent *ordered;
8831         struct extent_state *cached_state = NULL;
8832         u64 page_start = page_offset(page);
8833         u64 page_end = page_start + PAGE_SIZE - 1;
8834         u64 start;
8835         u64 end;
8836         int inode_evicting = inode->i_state & I_FREEING;
8837
8838         /*
8839          * we have the page locked, so new writeback can't start,
8840          * and the dirty bit won't be cleared while we are here.
8841          *
8842          * Wait for IO on this page so that we can safely clear
8843          * the PagePrivate2 bit and do ordered accounting
8844          */
8845         wait_on_page_writeback(page);
8846
8847         tree = &BTRFS_I(inode)->io_tree;
8848         if (offset) {
8849                 btrfs_releasepage(page, GFP_NOFS);
8850                 return;
8851         }
8852
8853         if (!inode_evicting)
8854                 lock_extent_bits(tree, page_start, page_end, &cached_state);
8855 again:
8856         start = page_start;
8857         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
8858                                         page_end - start + 1);
8859         if (ordered) {
8860                 end = min(page_end, ordered->file_offset + ordered->len - 1);
8861                 /*
8862                  * IO on this page will never be started, so we need
8863                  * to account for any ordered extents now
8864                  */
8865                 if (!inode_evicting)
8866                         clear_extent_bit(tree, start, end,
8867                                          EXTENT_DIRTY | EXTENT_DELALLOC |
8868                                          EXTENT_DELALLOC_NEW |
8869                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
8870                                          EXTENT_DEFRAG, 1, 0, &cached_state);
8871                 /*
8872                  * whoever cleared the private bit is responsible
8873                  * for the finish_ordered_io
8874                  */
8875                 if (TestClearPagePrivate2(page)) {
8876                         struct btrfs_ordered_inode_tree *tree;
8877                         u64 new_len;
8878
8879                         tree = &BTRFS_I(inode)->ordered_tree;
8880
8881                         spin_lock_irq(&tree->lock);
8882                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
8883                         new_len = start - ordered->file_offset;
8884                         if (new_len < ordered->truncated_len)
8885                                 ordered->truncated_len = new_len;
8886                         spin_unlock_irq(&tree->lock);
8887
8888                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
8889                                                            start,
8890                                                            end - start + 1, 1))
8891                                 btrfs_finish_ordered_io(ordered);
8892                 }
8893                 btrfs_put_ordered_extent(ordered);
8894                 if (!inode_evicting) {
8895                         cached_state = NULL;
8896                         lock_extent_bits(tree, start, end,
8897                                          &cached_state);
8898                 }
8899
8900                 start = end + 1;
8901                 if (start < page_end)
8902                         goto again;
8903         }
8904
8905         /*
8906          * Qgroup reserved space handler
8907          * Page here will be either
8908          * 1) Already written to disk
8909          *    In this case, its reserved space is released from data rsv map
8910          *    and will be freed by delayed_ref handler finally.
8911          *    So even we call qgroup_free_data(), it won't decrease reserved
8912          *    space.
8913          * 2) Not written to disk
8914          *    This means the reserved space should be freed here. However,
8915          *    if a truncate invalidates the page (by clearing PageDirty)
8916          *    and the page is accounted for while allocating extent
8917          *    in btrfs_check_data_free_space() we let delayed_ref to
8918          *    free the entire extent.
8919          */
8920         if (PageDirty(page))
8921                 btrfs_qgroup_free_data(inode, NULL, page_start, PAGE_SIZE);
8922         if (!inode_evicting) {
8923                 clear_extent_bit(tree, page_start, page_end,
8924                                  EXTENT_LOCKED | EXTENT_DIRTY |
8925                                  EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
8926                                  EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 1, 1,
8927                                  &cached_state);
8928
8929                 __btrfs_releasepage(page, GFP_NOFS);
8930         }
8931
8932         ClearPageChecked(page);
8933         if (PagePrivate(page)) {
8934                 ClearPagePrivate(page);
8935                 set_page_private(page, 0);
8936                 put_page(page);
8937         }
8938 }
8939
8940 /*
8941  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
8942  * called from a page fault handler when a page is first dirtied. Hence we must
8943  * be careful to check for EOF conditions here. We set the page up correctly
8944  * for a written page which means we get ENOSPC checking when writing into
8945  * holes and correct delalloc and unwritten extent mapping on filesystems that
8946  * support these features.
8947  *
8948  * We are not allowed to take the i_mutex here so we have to play games to
8949  * protect against truncate races as the page could now be beyond EOF.  Because
8950  * vmtruncate() writes the inode size before removing pages, once we have the
8951  * page lock we can determine safely if the page is beyond EOF. If it is not
8952  * beyond EOF, then the page is guaranteed safe against truncation until we
8953  * unlock the page.
8954  */
8955 int btrfs_page_mkwrite(struct vm_fault *vmf)
8956 {
8957         struct page *page = vmf->page;
8958         struct inode *inode = file_inode(vmf->vma->vm_file);
8959         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8960         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8961         struct btrfs_ordered_extent *ordered;
8962         struct extent_state *cached_state = NULL;
8963         struct extent_changeset *data_reserved = NULL;
8964         char *kaddr;
8965         unsigned long zero_start;
8966         loff_t size;
8967         int ret;
8968         int reserved = 0;
8969         u64 reserved_space;
8970         u64 page_start;
8971         u64 page_end;
8972         u64 end;
8973
8974         reserved_space = PAGE_SIZE;
8975
8976         sb_start_pagefault(inode->i_sb);
8977         page_start = page_offset(page);
8978         page_end = page_start + PAGE_SIZE - 1;
8979         end = page_end;
8980
8981         /*
8982          * Reserving delalloc space after obtaining the page lock can lead to
8983          * deadlock. For example, if a dirty page is locked by this function
8984          * and the call to btrfs_delalloc_reserve_space() ends up triggering
8985          * dirty page write out, then the btrfs_writepage() function could
8986          * end up waiting indefinitely to get a lock on the page currently
8987          * being processed by btrfs_page_mkwrite() function.
8988          */
8989         ret = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
8990                                            reserved_space);
8991         if (!ret) {
8992                 ret = file_update_time(vmf->vma->vm_file);
8993                 reserved = 1;
8994         }
8995         if (ret) {
8996                 if (ret == -ENOMEM)
8997                         ret = VM_FAULT_OOM;
8998                 else /* -ENOSPC, -EIO, etc */
8999                         ret = VM_FAULT_SIGBUS;
9000                 if (reserved)
9001                         goto out;
9002                 goto out_noreserve;
9003         }
9004
9005         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
9006 again:
9007         lock_page(page);
9008         size = i_size_read(inode);
9009
9010         if ((page->mapping != inode->i_mapping) ||
9011             (page_start >= size)) {
9012                 /* page got truncated out from underneath us */
9013                 goto out_unlock;
9014         }
9015         wait_on_page_writeback(page);
9016
9017         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
9018         set_page_extent_mapped(page);
9019
9020         /*
9021          * we can't set the delalloc bits if there are pending ordered
9022          * extents.  Drop our locks and wait for them to finish
9023          */
9024         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
9025                         PAGE_SIZE);
9026         if (ordered) {
9027                 unlock_extent_cached(io_tree, page_start, page_end,
9028                                      &cached_state);
9029                 unlock_page(page);
9030                 btrfs_start_ordered_extent(inode, ordered, 1);
9031                 btrfs_put_ordered_extent(ordered);
9032                 goto again;
9033         }
9034
9035         if (page->index == ((size - 1) >> PAGE_SHIFT)) {
9036                 reserved_space = round_up(size - page_start,
9037                                           fs_info->sectorsize);
9038                 if (reserved_space < PAGE_SIZE) {
9039                         end = page_start + reserved_space - 1;
9040                         btrfs_delalloc_release_space(inode, data_reserved,
9041                                         page_start, PAGE_SIZE - reserved_space);
9042                 }
9043         }
9044
9045         /*
9046          * page_mkwrite gets called when the page is firstly dirtied after it's
9047          * faulted in, but write(2) could also dirty a page and set delalloc
9048          * bits, thus in this case for space account reason, we still need to
9049          * clear any delalloc bits within this page range since we have to
9050          * reserve data&meta space before lock_page() (see above comments).
9051          */
9052         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, end,
9053                           EXTENT_DIRTY | EXTENT_DELALLOC |
9054                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
9055                           0, 0, &cached_state);
9056
9057         ret = btrfs_set_extent_delalloc(inode, page_start, end, 0,
9058                                         &cached_state, 0);
9059         if (ret) {
9060                 unlock_extent_cached(io_tree, page_start, page_end,
9061                                      &cached_state);
9062                 ret = VM_FAULT_SIGBUS;
9063                 goto out_unlock;
9064         }
9065         ret = 0;
9066
9067         /* page is wholly or partially inside EOF */
9068         if (page_start + PAGE_SIZE > size)
9069                 zero_start = size & ~PAGE_MASK;
9070         else
9071                 zero_start = PAGE_SIZE;
9072
9073         if (zero_start != PAGE_SIZE) {
9074                 kaddr = kmap(page);
9075                 memset(kaddr + zero_start, 0, PAGE_SIZE - zero_start);
9076                 flush_dcache_page(page);
9077                 kunmap(page);
9078         }
9079         ClearPageChecked(page);
9080         set_page_dirty(page);
9081         SetPageUptodate(page);
9082
9083         BTRFS_I(inode)->last_trans = fs_info->generation;
9084         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
9085         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
9086
9087         unlock_extent_cached(io_tree, page_start, page_end, &cached_state);
9088
9089 out_unlock:
9090         if (!ret) {
9091                 btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
9092                 sb_end_pagefault(inode->i_sb);
9093                 extent_changeset_free(data_reserved);
9094                 return VM_FAULT_LOCKED;
9095         }
9096         unlock_page(page);
9097 out:
9098         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
9099         btrfs_delalloc_release_space(inode, data_reserved, page_start,
9100                                      reserved_space);
9101 out_noreserve:
9102         sb_end_pagefault(inode->i_sb);
9103         extent_changeset_free(data_reserved);
9104         return ret;
9105 }
9106
9107 static int btrfs_truncate(struct inode *inode)
9108 {
9109         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9110         struct btrfs_root *root = BTRFS_I(inode)->root;
9111         struct btrfs_block_rsv *rsv;
9112         int ret = 0;
9113         int err = 0;
9114         struct btrfs_trans_handle *trans;
9115         u64 mask = fs_info->sectorsize - 1;
9116         u64 min_size = btrfs_calc_trunc_metadata_size(fs_info, 1);
9117
9118         ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
9119                                        (u64)-1);
9120         if (ret)
9121                 return ret;
9122
9123         /*
9124          * Yes ladies and gentlemen, this is indeed ugly.  The fact is we have
9125          * 3 things going on here
9126          *
9127          * 1) We need to reserve space for our orphan item and the space to
9128          * delete our orphan item.  Lord knows we don't want to have a dangling
9129          * orphan item because we didn't reserve space to remove it.
9130          *
9131          * 2) We need to reserve space to update our inode.
9132          *
9133          * 3) We need to have something to cache all the space that is going to
9134          * be free'd up by the truncate operation, but also have some slack
9135          * space reserved in case it uses space during the truncate (thank you
9136          * very much snapshotting).
9137          *
9138          * And we need these to all be separate.  The fact is we can use a lot of
9139          * space doing the truncate, and we have no earthly idea how much space
9140          * we will use, so we need the truncate reservation to be separate so it
9141          * doesn't end up using space reserved for updating the inode or
9142          * removing the orphan item.  We also need to be able to stop the
9143          * transaction and start a new one, which means we need to be able to
9144          * update the inode several times, and we have no idea of knowing how
9145          * many times that will be, so we can't just reserve 1 item for the
9146          * entirety of the operation, so that has to be done separately as well.
9147          * Then there is the orphan item, which does indeed need to be held on
9148          * to for the whole operation, and we need nobody to touch this reserved
9149          * space except the orphan code.
9150          *
9151          * So that leaves us with
9152          *
9153          * 1) root->orphan_block_rsv - for the orphan deletion.
9154          * 2) rsv - for the truncate reservation, which we will steal from the
9155          * transaction reservation.
9156          * 3) fs_info->trans_block_rsv - this will have 1 items worth left for
9157          * updating the inode.
9158          */
9159         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
9160         if (!rsv)
9161                 return -ENOMEM;
9162         rsv->size = min_size;
9163         rsv->failfast = 1;
9164
9165         /*
9166          * 1 for the truncate slack space
9167          * 1 for updating the inode.
9168          */
9169         trans = btrfs_start_transaction(root, 2);
9170         if (IS_ERR(trans)) {
9171                 err = PTR_ERR(trans);
9172                 goto out;
9173         }
9174
9175         /* Migrate the slack space for the truncate to our reserve */
9176         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
9177                                       min_size, 0);
9178         BUG_ON(ret);
9179
9180         /*
9181          * So if we truncate and then write and fsync we normally would just
9182          * write the extents that changed, which is a problem if we need to
9183          * first truncate that entire inode.  So set this flag so we write out
9184          * all of the extents in the inode to the sync log so we're completely
9185          * safe.
9186          */
9187         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
9188         trans->block_rsv = rsv;
9189
9190         while (1) {
9191                 ret = btrfs_truncate_inode_items(trans, root, inode,
9192                                                  inode->i_size,
9193                                                  BTRFS_EXTENT_DATA_KEY);
9194                 trans->block_rsv = &fs_info->trans_block_rsv;
9195                 if (ret != -ENOSPC && ret != -EAGAIN) {
9196                         err = ret;
9197                         break;
9198                 }
9199
9200                 ret = btrfs_update_inode(trans, root, inode);
9201                 if (ret) {
9202                         err = ret;
9203                         break;
9204                 }
9205
9206                 btrfs_end_transaction(trans);
9207                 btrfs_btree_balance_dirty(fs_info);
9208
9209                 trans = btrfs_start_transaction(root, 2);
9210                 if (IS_ERR(trans)) {
9211                         ret = err = PTR_ERR(trans);
9212                         trans = NULL;
9213                         break;
9214                 }
9215
9216                 btrfs_block_rsv_release(fs_info, rsv, -1);
9217                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
9218                                               rsv, min_size, 0);
9219                 BUG_ON(ret);    /* shouldn't happen */
9220                 trans->block_rsv = rsv;
9221         }
9222
9223         /*
9224          * We can't call btrfs_truncate_block inside a trans handle as we could
9225          * deadlock with freeze, if we got NEED_TRUNCATE_BLOCK then we know
9226          * we've truncated everything except the last little bit, and can do
9227          * btrfs_truncate_block and then update the disk_i_size.
9228          */
9229         if (ret == NEED_TRUNCATE_BLOCK) {
9230                 btrfs_end_transaction(trans);
9231                 btrfs_btree_balance_dirty(fs_info);
9232
9233                 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
9234                 if (ret)
9235                         goto out;
9236                 trans = btrfs_start_transaction(root, 1);
9237                 if (IS_ERR(trans)) {
9238                         ret = PTR_ERR(trans);
9239                         goto out;
9240                 }
9241                 btrfs_ordered_update_i_size(inode, inode->i_size, NULL);
9242         }
9243
9244         if (ret == 0 && inode->i_nlink > 0) {
9245                 trans->block_rsv = root->orphan_block_rsv;
9246                 ret = btrfs_orphan_del(trans, BTRFS_I(inode));
9247                 if (ret)
9248                         err = ret;
9249         }
9250
9251         if (trans) {
9252                 trans->block_rsv = &fs_info->trans_block_rsv;
9253                 ret = btrfs_update_inode(trans, root, inode);
9254                 if (ret && !err)
9255                         err = ret;
9256
9257                 ret = btrfs_end_transaction(trans);
9258                 btrfs_btree_balance_dirty(fs_info);
9259         }
9260 out:
9261         btrfs_free_block_rsv(fs_info, rsv);
9262
9263         if (ret && !err)
9264                 err = ret;
9265
9266         return err;
9267 }
9268
9269 /*
9270  * create a new subvolume directory/inode (helper for the ioctl).
9271  */
9272 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
9273                              struct btrfs_root *new_root,
9274                              struct btrfs_root *parent_root,
9275                              u64 new_dirid)
9276 {
9277         struct inode *inode;
9278         int err;
9279         u64 index = 0;
9280
9281         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
9282                                 new_dirid, new_dirid,
9283                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
9284                                 &index);
9285         if (IS_ERR(inode))
9286                 return PTR_ERR(inode);
9287         inode->i_op = &btrfs_dir_inode_operations;
9288         inode->i_fop = &btrfs_dir_file_operations;
9289
9290         set_nlink(inode, 1);
9291         btrfs_i_size_write(BTRFS_I(inode), 0);
9292         unlock_new_inode(inode);
9293
9294         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
9295         if (err)
9296                 btrfs_err(new_root->fs_info,
9297                           "error inheriting subvolume %llu properties: %d",
9298                           new_root->root_key.objectid, err);
9299
9300         err = btrfs_update_inode(trans, new_root, inode);
9301
9302         iput(inode);
9303         return err;
9304 }
9305
9306 struct inode *btrfs_alloc_inode(struct super_block *sb)
9307 {
9308         struct btrfs_fs_info *fs_info = btrfs_sb(sb);
9309         struct btrfs_inode *ei;
9310         struct inode *inode;
9311
9312         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_KERNEL);
9313         if (!ei)
9314                 return NULL;
9315
9316         ei->root = NULL;
9317         ei->generation = 0;
9318         ei->last_trans = 0;
9319         ei->last_sub_trans = 0;
9320         ei->logged_trans = 0;
9321         ei->delalloc_bytes = 0;
9322         ei->new_delalloc_bytes = 0;
9323         ei->defrag_bytes = 0;
9324         ei->disk_i_size = 0;
9325         ei->flags = 0;
9326         ei->csum_bytes = 0;
9327         ei->index_cnt = (u64)-1;
9328         ei->dir_index = 0;
9329         ei->last_unlink_trans = 0;
9330         ei->last_log_commit = 0;
9331         ei->delayed_iput_count = 0;
9332
9333         spin_lock_init(&ei->lock);
9334         ei->outstanding_extents = 0;
9335         if (sb->s_magic != BTRFS_TEST_MAGIC)
9336                 btrfs_init_metadata_block_rsv(fs_info, &ei->block_rsv,
9337                                               BTRFS_BLOCK_RSV_DELALLOC);
9338         ei->runtime_flags = 0;
9339         ei->prop_compress = BTRFS_COMPRESS_NONE;
9340         ei->defrag_compress = BTRFS_COMPRESS_NONE;
9341
9342         ei->delayed_node = NULL;
9343
9344         ei->i_otime.tv_sec = 0;
9345         ei->i_otime.tv_nsec = 0;
9346
9347         inode = &ei->vfs_inode;
9348         extent_map_tree_init(&ei->extent_tree);
9349         extent_io_tree_init(&ei->io_tree, inode);
9350         extent_io_tree_init(&ei->io_failure_tree, inode);
9351         ei->io_tree.track_uptodate = 1;
9352         ei->io_failure_tree.track_uptodate = 1;
9353         atomic_set(&ei->sync_writers, 0);
9354         mutex_init(&ei->log_mutex);
9355         mutex_init(&ei->delalloc_mutex);
9356         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
9357         INIT_LIST_HEAD(&ei->delalloc_inodes);
9358         INIT_LIST_HEAD(&ei->delayed_iput);
9359         RB_CLEAR_NODE(&ei->rb_node);
9360         init_rwsem(&ei->dio_sem);
9361
9362         return inode;
9363 }
9364
9365 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
9366 void btrfs_test_destroy_inode(struct inode *inode)
9367 {
9368         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9369         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9370 }
9371 #endif
9372
9373 static void btrfs_i_callback(struct rcu_head *head)
9374 {
9375         struct inode *inode = container_of(head, struct inode, i_rcu);
9376         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9377 }
9378
9379 void btrfs_destroy_inode(struct inode *inode)
9380 {
9381         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9382         struct btrfs_ordered_extent *ordered;
9383         struct btrfs_root *root = BTRFS_I(inode)->root;
9384
9385         WARN_ON(!hlist_empty(&inode->i_dentry));
9386         WARN_ON(inode->i_data.nrpages);
9387         WARN_ON(BTRFS_I(inode)->block_rsv.reserved);
9388         WARN_ON(BTRFS_I(inode)->block_rsv.size);
9389         WARN_ON(BTRFS_I(inode)->outstanding_extents);
9390         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
9391         WARN_ON(BTRFS_I(inode)->new_delalloc_bytes);
9392         WARN_ON(BTRFS_I(inode)->csum_bytes);
9393         WARN_ON(BTRFS_I(inode)->defrag_bytes);
9394
9395         /*
9396          * This can happen where we create an inode, but somebody else also
9397          * created the same inode and we need to destroy the one we already
9398          * created.
9399          */
9400         if (!root)
9401                 goto free;
9402
9403         if (test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
9404                      &BTRFS_I(inode)->runtime_flags)) {
9405                 btrfs_info(fs_info, "inode %llu still on the orphan list",
9406                            btrfs_ino(BTRFS_I(inode)));
9407                 atomic_dec(&root->orphan_inodes);
9408         }
9409
9410         while (1) {
9411                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
9412                 if (!ordered)
9413                         break;
9414                 else {
9415                         btrfs_err(fs_info,
9416                                   "found ordered extent %llu %llu on inode cleanup",
9417                                   ordered->file_offset, ordered->len);
9418                         btrfs_remove_ordered_extent(inode, ordered);
9419                         btrfs_put_ordered_extent(ordered);
9420                         btrfs_put_ordered_extent(ordered);
9421                 }
9422         }
9423         btrfs_qgroup_check_reserved_leak(inode);
9424         inode_tree_del(inode);
9425         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9426 free:
9427         call_rcu(&inode->i_rcu, btrfs_i_callback);
9428 }
9429
9430 int btrfs_drop_inode(struct inode *inode)
9431 {
9432         struct btrfs_root *root = BTRFS_I(inode)->root;
9433
9434         if (root == NULL)
9435                 return 1;
9436
9437         /* the snap/subvol tree is on deleting */
9438         if (btrfs_root_refs(&root->root_item) == 0)
9439                 return 1;
9440         else
9441                 return generic_drop_inode(inode);
9442 }
9443
9444 static void init_once(void *foo)
9445 {
9446         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
9447
9448         inode_init_once(&ei->vfs_inode);
9449 }
9450
9451 void btrfs_destroy_cachep(void)
9452 {
9453         /*
9454          * Make sure all delayed rcu free inodes are flushed before we
9455          * destroy cache.
9456          */
9457         rcu_barrier();
9458         kmem_cache_destroy(btrfs_inode_cachep);
9459         kmem_cache_destroy(btrfs_trans_handle_cachep);
9460         kmem_cache_destroy(btrfs_path_cachep);
9461         kmem_cache_destroy(btrfs_free_space_cachep);
9462 }
9463
9464 int __init btrfs_init_cachep(void)
9465 {
9466         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
9467                         sizeof(struct btrfs_inode), 0,
9468                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD | SLAB_ACCOUNT,
9469                         init_once);
9470         if (!btrfs_inode_cachep)
9471                 goto fail;
9472
9473         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
9474                         sizeof(struct btrfs_trans_handle), 0,
9475                         SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
9476         if (!btrfs_trans_handle_cachep)
9477                 goto fail;
9478
9479         btrfs_path_cachep = kmem_cache_create("btrfs_path",
9480                         sizeof(struct btrfs_path), 0,
9481                         SLAB_MEM_SPREAD, NULL);
9482         if (!btrfs_path_cachep)
9483                 goto fail;
9484
9485         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
9486                         sizeof(struct btrfs_free_space), 0,
9487                         SLAB_MEM_SPREAD, NULL);
9488         if (!btrfs_free_space_cachep)
9489                 goto fail;
9490
9491         return 0;
9492 fail:
9493         btrfs_destroy_cachep();
9494         return -ENOMEM;
9495 }
9496
9497 static int btrfs_getattr(const struct path *path, struct kstat *stat,
9498                          u32 request_mask, unsigned int flags)
9499 {
9500         u64 delalloc_bytes;
9501         struct inode *inode = d_inode(path->dentry);
9502         u32 blocksize = inode->i_sb->s_blocksize;
9503         u32 bi_flags = BTRFS_I(inode)->flags;
9504
9505         stat->result_mask |= STATX_BTIME;
9506         stat->btime.tv_sec = BTRFS_I(inode)->i_otime.tv_sec;
9507         stat->btime.tv_nsec = BTRFS_I(inode)->i_otime.tv_nsec;
9508         if (bi_flags & BTRFS_INODE_APPEND)
9509                 stat->attributes |= STATX_ATTR_APPEND;
9510         if (bi_flags & BTRFS_INODE_COMPRESS)
9511                 stat->attributes |= STATX_ATTR_COMPRESSED;
9512         if (bi_flags & BTRFS_INODE_IMMUTABLE)
9513                 stat->attributes |= STATX_ATTR_IMMUTABLE;
9514         if (bi_flags & BTRFS_INODE_NODUMP)
9515                 stat->attributes |= STATX_ATTR_NODUMP;
9516
9517         stat->attributes_mask |= (STATX_ATTR_APPEND |
9518                                   STATX_ATTR_COMPRESSED |
9519                                   STATX_ATTR_IMMUTABLE |
9520                                   STATX_ATTR_NODUMP);
9521
9522         generic_fillattr(inode, stat);
9523         stat->dev = BTRFS_I(inode)->root->anon_dev;
9524
9525         spin_lock(&BTRFS_I(inode)->lock);
9526         delalloc_bytes = BTRFS_I(inode)->new_delalloc_bytes;
9527         spin_unlock(&BTRFS_I(inode)->lock);
9528         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
9529                         ALIGN(delalloc_bytes, blocksize)) >> 9;
9530         return 0;
9531 }
9532
9533 static int btrfs_rename_exchange(struct inode *old_dir,
9534                               struct dentry *old_dentry,
9535                               struct inode *new_dir,
9536                               struct dentry *new_dentry)
9537 {
9538         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9539         struct btrfs_trans_handle *trans;
9540         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9541         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9542         struct inode *new_inode = new_dentry->d_inode;
9543         struct inode *old_inode = old_dentry->d_inode;
9544         struct timespec ctime = current_time(old_inode);
9545         struct dentry *parent;
9546         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9547         u64 new_ino = btrfs_ino(BTRFS_I(new_inode));
9548         u64 old_idx = 0;
9549         u64 new_idx = 0;
9550         u64 root_objectid;
9551         int ret;
9552         bool root_log_pinned = false;
9553         bool dest_log_pinned = false;
9554
9555         /* we only allow rename subvolume link between subvolumes */
9556         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9557                 return -EXDEV;
9558
9559         /* close the race window with snapshot create/destroy ioctl */
9560         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9561                 down_read(&fs_info->subvol_sem);
9562         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9563                 down_read(&fs_info->subvol_sem);
9564
9565         /*
9566          * We want to reserve the absolute worst case amount of items.  So if
9567          * both inodes are subvols and we need to unlink them then that would
9568          * require 4 item modifications, but if they are both normal inodes it
9569          * would require 5 item modifications, so we'll assume their normal
9570          * inodes.  So 5 * 2 is 10, plus 2 for the new links, so 12 total items
9571          * should cover the worst case number of items we'll modify.
9572          */
9573         trans = btrfs_start_transaction(root, 12);
9574         if (IS_ERR(trans)) {
9575                 ret = PTR_ERR(trans);
9576                 goto out_notrans;
9577         }
9578
9579         /*
9580          * We need to find a free sequence number both in the source and
9581          * in the destination directory for the exchange.
9582          */
9583         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &old_idx);
9584         if (ret)
9585                 goto out_fail;
9586         ret = btrfs_set_inode_index(BTRFS_I(old_dir), &new_idx);
9587         if (ret)
9588                 goto out_fail;
9589
9590         BTRFS_I(old_inode)->dir_index = 0ULL;
9591         BTRFS_I(new_inode)->dir_index = 0ULL;
9592
9593         /* Reference for the source. */
9594         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9595                 /* force full log commit if subvolume involved. */
9596                 btrfs_set_log_full_commit(fs_info, trans);
9597         } else {
9598                 btrfs_pin_log_trans(root);
9599                 root_log_pinned = true;
9600                 ret = btrfs_insert_inode_ref(trans, dest,
9601                                              new_dentry->d_name.name,
9602                                              new_dentry->d_name.len,
9603                                              old_ino,
9604                                              btrfs_ino(BTRFS_I(new_dir)),
9605                                              old_idx);
9606                 if (ret)
9607                         goto out_fail;
9608         }
9609
9610         /* And now for the dest. */
9611         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9612                 /* force full log commit if subvolume involved. */
9613                 btrfs_set_log_full_commit(fs_info, trans);
9614         } else {
9615                 btrfs_pin_log_trans(dest);
9616                 dest_log_pinned = true;
9617                 ret = btrfs_insert_inode_ref(trans, root,
9618                                              old_dentry->d_name.name,
9619                                              old_dentry->d_name.len,
9620                                              new_ino,
9621                                              btrfs_ino(BTRFS_I(old_dir)),
9622                                              new_idx);
9623                 if (ret)
9624                         goto out_fail;
9625         }
9626
9627         /* Update inode version and ctime/mtime. */
9628         inode_inc_iversion(old_dir);
9629         inode_inc_iversion(new_dir);
9630         inode_inc_iversion(old_inode);
9631         inode_inc_iversion(new_inode);
9632         old_dir->i_ctime = old_dir->i_mtime = ctime;
9633         new_dir->i_ctime = new_dir->i_mtime = ctime;
9634         old_inode->i_ctime = ctime;
9635         new_inode->i_ctime = ctime;
9636
9637         if (old_dentry->d_parent != new_dentry->d_parent) {
9638                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9639                                 BTRFS_I(old_inode), 1);
9640                 btrfs_record_unlink_dir(trans, BTRFS_I(new_dir),
9641                                 BTRFS_I(new_inode), 1);
9642         }
9643
9644         /* src is a subvolume */
9645         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9646                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9647                 ret = btrfs_unlink_subvol(trans, root, old_dir,
9648                                           root_objectid,
9649                                           old_dentry->d_name.name,
9650                                           old_dentry->d_name.len);
9651         } else { /* src is an inode */
9652                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9653                                            BTRFS_I(old_dentry->d_inode),
9654                                            old_dentry->d_name.name,
9655                                            old_dentry->d_name.len);
9656                 if (!ret)
9657                         ret = btrfs_update_inode(trans, root, old_inode);
9658         }
9659         if (ret) {
9660                 btrfs_abort_transaction(trans, ret);
9661                 goto out_fail;
9662         }
9663
9664         /* dest is a subvolume */
9665         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9666                 root_objectid = BTRFS_I(new_inode)->root->root_key.objectid;
9667                 ret = btrfs_unlink_subvol(trans, dest, new_dir,
9668                                           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                 btrfs_log_new_name(trans, BTRFS_I(old_inode), BTRFS_I(old_dir),
9708                                 parent);
9709                 btrfs_end_log_trans(root);
9710                 root_log_pinned = false;
9711         }
9712         if (dest_log_pinned) {
9713                 parent = old_dentry->d_parent;
9714                 btrfs_log_new_name(trans, BTRFS_I(new_inode), BTRFS_I(new_dir),
9715                                 parent);
9716                 btrfs_end_log_trans(dest);
9717                 dest_log_pinned = false;
9718         }
9719 out_fail:
9720         /*
9721          * If we have pinned a log and an error happened, we unpin tasks
9722          * trying to sync the log and force them to fallback to a transaction
9723          * commit if the log currently contains any of the inodes involved in
9724          * this rename operation (to ensure we do not persist a log with an
9725          * inconsistent state for any of these inodes or leading to any
9726          * inconsistencies when replayed). If the transaction was aborted, the
9727          * abortion reason is propagated to userspace when attempting to commit
9728          * the transaction. If the log does not contain any of these inodes, we
9729          * allow the tasks to sync it.
9730          */
9731         if (ret && (root_log_pinned || dest_log_pinned)) {
9732                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
9733                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
9734                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
9735                     (new_inode &&
9736                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
9737                         btrfs_set_log_full_commit(fs_info, trans);
9738
9739                 if (root_log_pinned) {
9740                         btrfs_end_log_trans(root);
9741                         root_log_pinned = false;
9742                 }
9743                 if (dest_log_pinned) {
9744                         btrfs_end_log_trans(dest);
9745                         dest_log_pinned = false;
9746                 }
9747         }
9748         ret = btrfs_end_transaction(trans);
9749 out_notrans:
9750         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9751                 up_read(&fs_info->subvol_sem);
9752         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9753                 up_read(&fs_info->subvol_sem);
9754
9755         return ret;
9756 }
9757
9758 static int btrfs_whiteout_for_rename(struct btrfs_trans_handle *trans,
9759                                      struct btrfs_root *root,
9760                                      struct inode *dir,
9761                                      struct dentry *dentry)
9762 {
9763         int ret;
9764         struct inode *inode;
9765         u64 objectid;
9766         u64 index;
9767
9768         ret = btrfs_find_free_ino(root, &objectid);
9769         if (ret)
9770                 return ret;
9771
9772         inode = btrfs_new_inode(trans, root, dir,
9773                                 dentry->d_name.name,
9774                                 dentry->d_name.len,
9775                                 btrfs_ino(BTRFS_I(dir)),
9776                                 objectid,
9777                                 S_IFCHR | WHITEOUT_MODE,
9778                                 &index);
9779
9780         if (IS_ERR(inode)) {
9781                 ret = PTR_ERR(inode);
9782                 return ret;
9783         }
9784
9785         inode->i_op = &btrfs_special_inode_operations;
9786         init_special_inode(inode, inode->i_mode,
9787                 WHITEOUT_DEV);
9788
9789         ret = btrfs_init_inode_security(trans, inode, dir,
9790                                 &dentry->d_name);
9791         if (ret)
9792                 goto out;
9793
9794         ret = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
9795                                 BTRFS_I(inode), 0, index);
9796         if (ret)
9797                 goto out;
9798
9799         ret = btrfs_update_inode(trans, root, inode);
9800 out:
9801         unlock_new_inode(inode);
9802         if (ret)
9803                 inode_dec_link_count(inode);
9804         iput(inode);
9805
9806         return ret;
9807 }
9808
9809 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
9810                            struct inode *new_dir, struct dentry *new_dentry,
9811                            unsigned int flags)
9812 {
9813         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9814         struct btrfs_trans_handle *trans;
9815         unsigned int trans_num_items;
9816         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9817         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9818         struct inode *new_inode = d_inode(new_dentry);
9819         struct inode *old_inode = d_inode(old_dentry);
9820         u64 index = 0;
9821         u64 root_objectid;
9822         int ret;
9823         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9824         bool log_pinned = false;
9825
9826         if (btrfs_ino(BTRFS_I(new_dir)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
9827                 return -EPERM;
9828
9829         /* we only allow rename subvolume link between subvolumes */
9830         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9831                 return -EXDEV;
9832
9833         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
9834             (new_inode && btrfs_ino(BTRFS_I(new_inode)) == BTRFS_FIRST_FREE_OBJECTID))
9835                 return -ENOTEMPTY;
9836
9837         if (S_ISDIR(old_inode->i_mode) && new_inode &&
9838             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
9839                 return -ENOTEMPTY;
9840
9841
9842         /* check for collisions, even if the  name isn't there */
9843         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
9844                              new_dentry->d_name.name,
9845                              new_dentry->d_name.len);
9846
9847         if (ret) {
9848                 if (ret == -EEXIST) {
9849                         /* we shouldn't get
9850                          * eexist without a new_inode */
9851                         if (WARN_ON(!new_inode)) {
9852                                 return ret;
9853                         }
9854                 } else {
9855                         /* maybe -EOVERFLOW */
9856                         return ret;
9857                 }
9858         }
9859         ret = 0;
9860
9861         /*
9862          * we're using rename to replace one file with another.  Start IO on it
9863          * now so  we don't add too much work to the end of the transaction
9864          */
9865         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
9866                 filemap_flush(old_inode->i_mapping);
9867
9868         /* close the racy window with snapshot create/destroy ioctl */
9869         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9870                 down_read(&fs_info->subvol_sem);
9871         /*
9872          * We want to reserve the absolute worst case amount of items.  So if
9873          * both inodes are subvols and we need to unlink them then that would
9874          * require 4 item modifications, but if they are both normal inodes it
9875          * would require 5 item modifications, so we'll assume they are normal
9876          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
9877          * should cover the worst case number of items we'll modify.
9878          * If our rename has the whiteout flag, we need more 5 units for the
9879          * new inode (1 inode item, 1 inode ref, 2 dir items and 1 xattr item
9880          * when selinux is enabled).
9881          */
9882         trans_num_items = 11;
9883         if (flags & RENAME_WHITEOUT)
9884                 trans_num_items += 5;
9885         trans = btrfs_start_transaction(root, trans_num_items);
9886         if (IS_ERR(trans)) {
9887                 ret = PTR_ERR(trans);
9888                 goto out_notrans;
9889         }
9890
9891         if (dest != root)
9892                 btrfs_record_root_in_trans(trans, dest);
9893
9894         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &index);
9895         if (ret)
9896                 goto out_fail;
9897
9898         BTRFS_I(old_inode)->dir_index = 0ULL;
9899         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9900                 /* force full log commit if subvolume involved. */
9901                 btrfs_set_log_full_commit(fs_info, trans);
9902         } else {
9903                 btrfs_pin_log_trans(root);
9904                 log_pinned = true;
9905                 ret = btrfs_insert_inode_ref(trans, dest,
9906                                              new_dentry->d_name.name,
9907                                              new_dentry->d_name.len,
9908                                              old_ino,
9909                                              btrfs_ino(BTRFS_I(new_dir)), index);
9910                 if (ret)
9911                         goto out_fail;
9912         }
9913
9914         inode_inc_iversion(old_dir);
9915         inode_inc_iversion(new_dir);
9916         inode_inc_iversion(old_inode);
9917         old_dir->i_ctime = old_dir->i_mtime =
9918         new_dir->i_ctime = new_dir->i_mtime =
9919         old_inode->i_ctime = current_time(old_dir);
9920
9921         if (old_dentry->d_parent != new_dentry->d_parent)
9922                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9923                                 BTRFS_I(old_inode), 1);
9924
9925         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9926                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9927                 ret = btrfs_unlink_subvol(trans, root, old_dir, root_objectid,
9928                                         old_dentry->d_name.name,
9929                                         old_dentry->d_name.len);
9930         } else {
9931                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9932                                         BTRFS_I(d_inode(old_dentry)),
9933                                         old_dentry->d_name.name,
9934                                         old_dentry->d_name.len);
9935                 if (!ret)
9936                         ret = btrfs_update_inode(trans, root, old_inode);
9937         }
9938         if (ret) {
9939                 btrfs_abort_transaction(trans, ret);
9940                 goto out_fail;
9941         }
9942
9943         if (new_inode) {
9944                 inode_inc_iversion(new_inode);
9945                 new_inode->i_ctime = current_time(new_inode);
9946                 if (unlikely(btrfs_ino(BTRFS_I(new_inode)) ==
9947                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
9948                         root_objectid = BTRFS_I(new_inode)->location.objectid;
9949                         ret = btrfs_unlink_subvol(trans, dest, new_dir,
9950                                                 root_objectid,
9951                                                 new_dentry->d_name.name,
9952                                                 new_dentry->d_name.len);
9953                         BUG_ON(new_inode->i_nlink == 0);
9954                 } else {
9955                         ret = btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
9956                                                  BTRFS_I(d_inode(new_dentry)),
9957                                                  new_dentry->d_name.name,
9958                                                  new_dentry->d_name.len);
9959                 }
9960                 if (!ret && new_inode->i_nlink == 0)
9961                         ret = btrfs_orphan_add(trans,
9962                                         BTRFS_I(d_inode(new_dentry)));
9963                 if (ret) {
9964                         btrfs_abort_transaction(trans, ret);
9965                         goto out_fail;
9966                 }
9967         }
9968
9969         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
9970                              new_dentry->d_name.name,
9971                              new_dentry->d_name.len, 0, index);
9972         if (ret) {
9973                 btrfs_abort_transaction(trans, ret);
9974                 goto out_fail;
9975         }
9976
9977         if (old_inode->i_nlink == 1)
9978                 BTRFS_I(old_inode)->dir_index = index;
9979
9980         if (log_pinned) {
9981                 struct dentry *parent = new_dentry->d_parent;
9982
9983                 btrfs_log_new_name(trans, BTRFS_I(old_inode), BTRFS_I(old_dir),
9984                                 parent);
9985                 btrfs_end_log_trans(root);
9986                 log_pinned = false;
9987         }
9988
9989         if (flags & RENAME_WHITEOUT) {
9990                 ret = btrfs_whiteout_for_rename(trans, root, old_dir,
9991                                                 old_dentry);
9992
9993                 if (ret) {
9994                         btrfs_abort_transaction(trans, ret);
9995                         goto out_fail;
9996                 }
9997         }
9998 out_fail:
9999         /*
10000          * If we have pinned the log and an error happened, we unpin tasks
10001          * trying to sync the log and force them to fallback to a transaction
10002          * commit if the log currently contains any of the inodes involved in
10003          * this rename operation (to ensure we do not persist a log with an
10004          * inconsistent state for any of these inodes or leading to any
10005          * inconsistencies when replayed). If the transaction was aborted, the
10006          * abortion reason is propagated to userspace when attempting to commit
10007          * the transaction. If the log does not contain any of these inodes, we
10008          * allow the tasks to sync it.
10009          */
10010         if (ret && log_pinned) {
10011                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
10012                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
10013                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
10014                     (new_inode &&
10015                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
10016                         btrfs_set_log_full_commit(fs_info, trans);
10017
10018                 btrfs_end_log_trans(root);
10019                 log_pinned = false;
10020         }
10021         btrfs_end_transaction(trans);
10022 out_notrans:
10023         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
10024                 up_read(&fs_info->subvol_sem);
10025
10026         return ret;
10027 }
10028
10029 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
10030                          struct inode *new_dir, struct dentry *new_dentry,
10031                          unsigned int flags)
10032 {
10033         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
10034                 return -EINVAL;
10035
10036         if (flags & RENAME_EXCHANGE)
10037                 return btrfs_rename_exchange(old_dir, old_dentry, new_dir,
10038                                           new_dentry);
10039
10040         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry, flags);
10041 }
10042
10043 static void btrfs_run_delalloc_work(struct btrfs_work *work)
10044 {
10045         struct btrfs_delalloc_work *delalloc_work;
10046         struct inode *inode;
10047
10048         delalloc_work = container_of(work, struct btrfs_delalloc_work,
10049                                      work);
10050         inode = delalloc_work->inode;
10051         filemap_flush(inode->i_mapping);
10052         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
10053                                 &BTRFS_I(inode)->runtime_flags))
10054                 filemap_flush(inode->i_mapping);
10055
10056         if (delalloc_work->delay_iput)
10057                 btrfs_add_delayed_iput(inode);
10058         else
10059                 iput(inode);
10060         complete(&delalloc_work->completion);
10061 }
10062
10063 struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode,
10064                                                     int delay_iput)
10065 {
10066         struct btrfs_delalloc_work *work;
10067
10068         work = kmalloc(sizeof(*work), GFP_NOFS);
10069         if (!work)
10070                 return NULL;
10071
10072         init_completion(&work->completion);
10073         INIT_LIST_HEAD(&work->list);
10074         work->inode = inode;
10075         work->delay_iput = delay_iput;
10076         WARN_ON_ONCE(!inode);
10077         btrfs_init_work(&work->work, btrfs_flush_delalloc_helper,
10078                         btrfs_run_delalloc_work, NULL, NULL);
10079
10080         return work;
10081 }
10082
10083 void btrfs_wait_and_free_delalloc_work(struct btrfs_delalloc_work *work)
10084 {
10085         wait_for_completion(&work->completion);
10086         kfree(work);
10087 }
10088
10089 /*
10090  * some fairly slow code that needs optimization. This walks the list
10091  * of all the inodes with pending delalloc and forces them to disk.
10092  */
10093 static int __start_delalloc_inodes(struct btrfs_root *root, int delay_iput,
10094                                    int nr)
10095 {
10096         struct btrfs_inode *binode;
10097         struct inode *inode;
10098         struct btrfs_delalloc_work *work, *next;
10099         struct list_head works;
10100         struct list_head splice;
10101         int ret = 0;
10102
10103         INIT_LIST_HEAD(&works);
10104         INIT_LIST_HEAD(&splice);
10105
10106         mutex_lock(&root->delalloc_mutex);
10107         spin_lock(&root->delalloc_lock);
10108         list_splice_init(&root->delalloc_inodes, &splice);
10109         while (!list_empty(&splice)) {
10110                 binode = list_entry(splice.next, struct btrfs_inode,
10111                                     delalloc_inodes);
10112
10113                 list_move_tail(&binode->delalloc_inodes,
10114                                &root->delalloc_inodes);
10115                 inode = igrab(&binode->vfs_inode);
10116                 if (!inode) {
10117                         cond_resched_lock(&root->delalloc_lock);
10118                         continue;
10119                 }
10120                 spin_unlock(&root->delalloc_lock);
10121
10122                 work = btrfs_alloc_delalloc_work(inode, delay_iput);
10123                 if (!work) {
10124                         if (delay_iput)
10125                                 btrfs_add_delayed_iput(inode);
10126                         else
10127                                 iput(inode);
10128                         ret = -ENOMEM;
10129                         goto out;
10130                 }
10131                 list_add_tail(&work->list, &works);
10132                 btrfs_queue_work(root->fs_info->flush_workers,
10133                                  &work->work);
10134                 ret++;
10135                 if (nr != -1 && ret >= nr)
10136                         goto out;
10137                 cond_resched();
10138                 spin_lock(&root->delalloc_lock);
10139         }
10140         spin_unlock(&root->delalloc_lock);
10141
10142 out:
10143         list_for_each_entry_safe(work, next, &works, list) {
10144                 list_del_init(&work->list);
10145                 btrfs_wait_and_free_delalloc_work(work);
10146         }
10147
10148         if (!list_empty_careful(&splice)) {
10149                 spin_lock(&root->delalloc_lock);
10150                 list_splice_tail(&splice, &root->delalloc_inodes);
10151                 spin_unlock(&root->delalloc_lock);
10152         }
10153         mutex_unlock(&root->delalloc_mutex);
10154         return ret;
10155 }
10156
10157 int btrfs_start_delalloc_inodes(struct btrfs_root *root, int delay_iput)
10158 {
10159         struct btrfs_fs_info *fs_info = root->fs_info;
10160         int ret;
10161
10162         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10163                 return -EROFS;
10164
10165         ret = __start_delalloc_inodes(root, delay_iput, -1);
10166         if (ret > 0)
10167                 ret = 0;
10168         return ret;
10169 }
10170
10171 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int delay_iput,
10172                                int nr)
10173 {
10174         struct btrfs_root *root;
10175         struct list_head splice;
10176         int ret;
10177
10178         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10179                 return -EROFS;
10180
10181         INIT_LIST_HEAD(&splice);
10182
10183         mutex_lock(&fs_info->delalloc_root_mutex);
10184         spin_lock(&fs_info->delalloc_root_lock);
10185         list_splice_init(&fs_info->delalloc_roots, &splice);
10186         while (!list_empty(&splice) && nr) {
10187                 root = list_first_entry(&splice, struct btrfs_root,
10188                                         delalloc_root);
10189                 root = btrfs_grab_fs_root(root);
10190                 BUG_ON(!root);
10191                 list_move_tail(&root->delalloc_root,
10192                                &fs_info->delalloc_roots);
10193                 spin_unlock(&fs_info->delalloc_root_lock);
10194
10195                 ret = __start_delalloc_inodes(root, delay_iput, nr);
10196                 btrfs_put_fs_root(root);
10197                 if (ret < 0)
10198                         goto out;
10199
10200                 if (nr != -1) {
10201                         nr -= ret;
10202                         WARN_ON(nr < 0);
10203                 }
10204                 spin_lock(&fs_info->delalloc_root_lock);
10205         }
10206         spin_unlock(&fs_info->delalloc_root_lock);
10207
10208         ret = 0;
10209 out:
10210         if (!list_empty_careful(&splice)) {
10211                 spin_lock(&fs_info->delalloc_root_lock);
10212                 list_splice_tail(&splice, &fs_info->delalloc_roots);
10213                 spin_unlock(&fs_info->delalloc_root_lock);
10214         }
10215         mutex_unlock(&fs_info->delalloc_root_mutex);
10216         return ret;
10217 }
10218
10219 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
10220                          const char *symname)
10221 {
10222         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10223         struct btrfs_trans_handle *trans;
10224         struct btrfs_root *root = BTRFS_I(dir)->root;
10225         struct btrfs_path *path;
10226         struct btrfs_key key;
10227         struct inode *inode = NULL;
10228         int err;
10229         int drop_inode = 0;
10230         u64 objectid;
10231         u64 index = 0;
10232         int name_len;
10233         int datasize;
10234         unsigned long ptr;
10235         struct btrfs_file_extent_item *ei;
10236         struct extent_buffer *leaf;
10237
10238         name_len = strlen(symname);
10239         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info))
10240                 return -ENAMETOOLONG;
10241
10242         /*
10243          * 2 items for inode item and ref
10244          * 2 items for dir items
10245          * 1 item for updating parent inode item
10246          * 1 item for the inline extent item
10247          * 1 item for xattr if selinux is on
10248          */
10249         trans = btrfs_start_transaction(root, 7);
10250         if (IS_ERR(trans))
10251                 return PTR_ERR(trans);
10252
10253         err = btrfs_find_free_ino(root, &objectid);
10254         if (err)
10255                 goto out_unlock;
10256
10257         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
10258                                 dentry->d_name.len, btrfs_ino(BTRFS_I(dir)),
10259                                 objectid, S_IFLNK|S_IRWXUGO, &index);
10260         if (IS_ERR(inode)) {
10261                 err = PTR_ERR(inode);
10262                 goto out_unlock;
10263         }
10264
10265         /*
10266         * If the active LSM wants to access the inode during
10267         * d_instantiate it needs these. Smack checks to see
10268         * if the filesystem supports xattrs by looking at the
10269         * ops vector.
10270         */
10271         inode->i_fop = &btrfs_file_operations;
10272         inode->i_op = &btrfs_file_inode_operations;
10273         inode->i_mapping->a_ops = &btrfs_aops;
10274         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10275
10276         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
10277         if (err)
10278                 goto out_unlock_inode;
10279
10280         path = btrfs_alloc_path();
10281         if (!path) {
10282                 err = -ENOMEM;
10283                 goto out_unlock_inode;
10284         }
10285         key.objectid = btrfs_ino(BTRFS_I(inode));
10286         key.offset = 0;
10287         key.type = BTRFS_EXTENT_DATA_KEY;
10288         datasize = btrfs_file_extent_calc_inline_size(name_len);
10289         err = btrfs_insert_empty_item(trans, root, path, &key,
10290                                       datasize);
10291         if (err) {
10292                 btrfs_free_path(path);
10293                 goto out_unlock_inode;
10294         }
10295         leaf = path->nodes[0];
10296         ei = btrfs_item_ptr(leaf, path->slots[0],
10297                             struct btrfs_file_extent_item);
10298         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
10299         btrfs_set_file_extent_type(leaf, ei,
10300                                    BTRFS_FILE_EXTENT_INLINE);
10301         btrfs_set_file_extent_encryption(leaf, ei, 0);
10302         btrfs_set_file_extent_compression(leaf, ei, 0);
10303         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
10304         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
10305
10306         ptr = btrfs_file_extent_inline_start(ei);
10307         write_extent_buffer(leaf, symname, ptr, name_len);
10308         btrfs_mark_buffer_dirty(leaf);
10309         btrfs_free_path(path);
10310
10311         inode->i_op = &btrfs_symlink_inode_operations;
10312         inode_nohighmem(inode);
10313         inode->i_mapping->a_ops = &btrfs_symlink_aops;
10314         inode_set_bytes(inode, name_len);
10315         btrfs_i_size_write(BTRFS_I(inode), name_len);
10316         err = btrfs_update_inode(trans, root, inode);
10317         /*
10318          * Last step, add directory indexes for our symlink inode. This is the
10319          * last step to avoid extra cleanup of these indexes if an error happens
10320          * elsewhere above.
10321          */
10322         if (!err)
10323                 err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
10324                                 BTRFS_I(inode), 0, index);
10325         if (err) {
10326                 drop_inode = 1;
10327                 goto out_unlock_inode;
10328         }
10329
10330         unlock_new_inode(inode);
10331         d_instantiate(dentry, inode);
10332
10333 out_unlock:
10334         btrfs_end_transaction(trans);
10335         if (drop_inode) {
10336                 inode_dec_link_count(inode);
10337                 iput(inode);
10338         }
10339         btrfs_btree_balance_dirty(fs_info);
10340         return err;
10341
10342 out_unlock_inode:
10343         drop_inode = 1;
10344         unlock_new_inode(inode);
10345         goto out_unlock;
10346 }
10347
10348 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
10349                                        u64 start, u64 num_bytes, u64 min_size,
10350                                        loff_t actual_len, u64 *alloc_hint,
10351                                        struct btrfs_trans_handle *trans)
10352 {
10353         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
10354         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
10355         struct extent_map *em;
10356         struct btrfs_root *root = BTRFS_I(inode)->root;
10357         struct btrfs_key ins;
10358         u64 cur_offset = start;
10359         u64 i_size;
10360         u64 cur_bytes;
10361         u64 last_alloc = (u64)-1;
10362         int ret = 0;
10363         bool own_trans = true;
10364         u64 end = start + num_bytes - 1;
10365
10366         if (trans)
10367                 own_trans = false;
10368         while (num_bytes > 0) {
10369                 if (own_trans) {
10370                         trans = btrfs_start_transaction(root, 3);
10371                         if (IS_ERR(trans)) {
10372                                 ret = PTR_ERR(trans);
10373                                 break;
10374                         }
10375                 }
10376
10377                 cur_bytes = min_t(u64, num_bytes, SZ_256M);
10378                 cur_bytes = max(cur_bytes, min_size);
10379                 /*
10380                  * If we are severely fragmented we could end up with really
10381                  * small allocations, so if the allocator is returning small
10382                  * chunks lets make its job easier by only searching for those
10383                  * sized chunks.
10384                  */
10385                 cur_bytes = min(cur_bytes, last_alloc);
10386                 ret = btrfs_reserve_extent(root, cur_bytes, cur_bytes,
10387                                 min_size, 0, *alloc_hint, &ins, 1, 0);
10388                 if (ret) {
10389                         if (own_trans)
10390                                 btrfs_end_transaction(trans);
10391                         break;
10392                 }
10393                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
10394
10395                 last_alloc = ins.offset;
10396                 ret = insert_reserved_file_extent(trans, inode,
10397                                                   cur_offset, ins.objectid,
10398                                                   ins.offset, ins.offset,
10399                                                   ins.offset, 0, 0, 0,
10400                                                   BTRFS_FILE_EXTENT_PREALLOC);
10401                 if (ret) {
10402                         btrfs_free_reserved_extent(fs_info, ins.objectid,
10403                                                    ins.offset, 0);
10404                         btrfs_abort_transaction(trans, ret);
10405                         if (own_trans)
10406                                 btrfs_end_transaction(trans);
10407                         break;
10408                 }
10409
10410                 btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10411                                         cur_offset + ins.offset -1, 0);
10412
10413                 em = alloc_extent_map();
10414                 if (!em) {
10415                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
10416                                 &BTRFS_I(inode)->runtime_flags);
10417                         goto next;
10418                 }
10419
10420                 em->start = cur_offset;
10421                 em->orig_start = cur_offset;
10422                 em->len = ins.offset;
10423                 em->block_start = ins.objectid;
10424                 em->block_len = ins.offset;
10425                 em->orig_block_len = ins.offset;
10426                 em->ram_bytes = ins.offset;
10427                 em->bdev = fs_info->fs_devices->latest_bdev;
10428                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
10429                 em->generation = trans->transid;
10430
10431                 while (1) {
10432                         write_lock(&em_tree->lock);
10433                         ret = add_extent_mapping(em_tree, em, 1);
10434                         write_unlock(&em_tree->lock);
10435                         if (ret != -EEXIST)
10436                                 break;
10437                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10438                                                 cur_offset + ins.offset - 1,
10439                                                 0);
10440                 }
10441                 free_extent_map(em);
10442 next:
10443                 num_bytes -= ins.offset;
10444                 cur_offset += ins.offset;
10445                 *alloc_hint = ins.objectid + ins.offset;
10446
10447                 inode_inc_iversion(inode);
10448                 inode->i_ctime = current_time(inode);
10449                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
10450                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
10451                     (actual_len > inode->i_size) &&
10452                     (cur_offset > inode->i_size)) {
10453                         if (cur_offset > actual_len)
10454                                 i_size = actual_len;
10455                         else
10456                                 i_size = cur_offset;
10457                         i_size_write(inode, i_size);
10458                         btrfs_ordered_update_i_size(inode, i_size, NULL);
10459                 }
10460
10461                 ret = btrfs_update_inode(trans, root, inode);
10462
10463                 if (ret) {
10464                         btrfs_abort_transaction(trans, ret);
10465                         if (own_trans)
10466                                 btrfs_end_transaction(trans);
10467                         break;
10468                 }
10469
10470                 if (own_trans)
10471                         btrfs_end_transaction(trans);
10472         }
10473         if (cur_offset < end)
10474                 btrfs_free_reserved_data_space(inode, NULL, cur_offset,
10475                         end - cur_offset + 1);
10476         return ret;
10477 }
10478
10479 int btrfs_prealloc_file_range(struct inode *inode, int mode,
10480                               u64 start, u64 num_bytes, u64 min_size,
10481                               loff_t actual_len, u64 *alloc_hint)
10482 {
10483         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10484                                            min_size, actual_len, alloc_hint,
10485                                            NULL);
10486 }
10487
10488 int btrfs_prealloc_file_range_trans(struct inode *inode,
10489                                     struct btrfs_trans_handle *trans, int mode,
10490                                     u64 start, u64 num_bytes, u64 min_size,
10491                                     loff_t actual_len, u64 *alloc_hint)
10492 {
10493         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10494                                            min_size, actual_len, alloc_hint, trans);
10495 }
10496
10497 static int btrfs_set_page_dirty(struct page *page)
10498 {
10499         return __set_page_dirty_nobuffers(page);
10500 }
10501
10502 static int btrfs_permission(struct inode *inode, int mask)
10503 {
10504         struct btrfs_root *root = BTRFS_I(inode)->root;
10505         umode_t mode = inode->i_mode;
10506
10507         if (mask & MAY_WRITE &&
10508             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
10509                 if (btrfs_root_readonly(root))
10510                         return -EROFS;
10511                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
10512                         return -EACCES;
10513         }
10514         return generic_permission(inode, mask);
10515 }
10516
10517 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
10518 {
10519         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10520         struct btrfs_trans_handle *trans;
10521         struct btrfs_root *root = BTRFS_I(dir)->root;
10522         struct inode *inode = NULL;
10523         u64 objectid;
10524         u64 index;
10525         int ret = 0;
10526
10527         /*
10528          * 5 units required for adding orphan entry
10529          */
10530         trans = btrfs_start_transaction(root, 5);
10531         if (IS_ERR(trans))
10532                 return PTR_ERR(trans);
10533
10534         ret = btrfs_find_free_ino(root, &objectid);
10535         if (ret)
10536                 goto out;
10537
10538         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
10539                         btrfs_ino(BTRFS_I(dir)), objectid, mode, &index);
10540         if (IS_ERR(inode)) {
10541                 ret = PTR_ERR(inode);
10542                 inode = NULL;
10543                 goto out;
10544         }
10545
10546         inode->i_fop = &btrfs_file_operations;
10547         inode->i_op = &btrfs_file_inode_operations;
10548
10549         inode->i_mapping->a_ops = &btrfs_aops;
10550         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10551
10552         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
10553         if (ret)
10554                 goto out_inode;
10555
10556         ret = btrfs_update_inode(trans, root, inode);
10557         if (ret)
10558                 goto out_inode;
10559         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10560         if (ret)
10561                 goto out_inode;
10562
10563         /*
10564          * We set number of links to 0 in btrfs_new_inode(), and here we set
10565          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
10566          * through:
10567          *
10568          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
10569          */
10570         set_nlink(inode, 1);
10571         unlock_new_inode(inode);
10572         d_tmpfile(dentry, inode);
10573         mark_inode_dirty(inode);
10574
10575 out:
10576         btrfs_end_transaction(trans);
10577         if (ret)
10578                 iput(inode);
10579         btrfs_btree_balance_dirty(fs_info);
10580         return ret;
10581
10582 out_inode:
10583         unlock_new_inode(inode);
10584         goto out;
10585
10586 }
10587
10588 __attribute__((const))
10589 static int btrfs_readpage_io_failed_hook(struct page *page, int failed_mirror)
10590 {
10591         return -EAGAIN;
10592 }
10593
10594 static struct btrfs_fs_info *iotree_fs_info(void *private_data)
10595 {
10596         struct inode *inode = private_data;
10597         return btrfs_sb(inode->i_sb);
10598 }
10599
10600 static void btrfs_check_extent_io_range(void *private_data, const char *caller,
10601                                         u64 start, u64 end)
10602 {
10603         struct inode *inode = private_data;
10604         u64 isize;
10605
10606         isize = i_size_read(inode);
10607         if (end >= PAGE_SIZE && (end % 2) == 0 && end != isize - 1) {
10608                 btrfs_debug_rl(BTRFS_I(inode)->root->fs_info,
10609                     "%s: ino %llu isize %llu odd range [%llu,%llu]",
10610                         caller, btrfs_ino(BTRFS_I(inode)), isize, start, end);
10611         }
10612 }
10613
10614 void btrfs_set_range_writeback(void *private_data, u64 start, u64 end)
10615 {
10616         struct inode *inode = private_data;
10617         unsigned long index = start >> PAGE_SHIFT;
10618         unsigned long end_index = end >> PAGE_SHIFT;
10619         struct page *page;
10620
10621         while (index <= end_index) {
10622                 page = find_get_page(inode->i_mapping, index);
10623                 ASSERT(page); /* Pages should be in the extent_io_tree */
10624                 set_page_writeback(page);
10625                 put_page(page);
10626                 index++;
10627         }
10628 }
10629
10630 static const struct inode_operations btrfs_dir_inode_operations = {
10631         .getattr        = btrfs_getattr,
10632         .lookup         = btrfs_lookup,
10633         .create         = btrfs_create,
10634         .unlink         = btrfs_unlink,
10635         .link           = btrfs_link,
10636         .mkdir          = btrfs_mkdir,
10637         .rmdir          = btrfs_rmdir,
10638         .rename         = btrfs_rename2,
10639         .symlink        = btrfs_symlink,
10640         .setattr        = btrfs_setattr,
10641         .mknod          = btrfs_mknod,
10642         .listxattr      = btrfs_listxattr,
10643         .permission     = btrfs_permission,
10644         .get_acl        = btrfs_get_acl,
10645         .set_acl        = btrfs_set_acl,
10646         .update_time    = btrfs_update_time,
10647         .tmpfile        = btrfs_tmpfile,
10648 };
10649 static const struct inode_operations btrfs_dir_ro_inode_operations = {
10650         .lookup         = btrfs_lookup,
10651         .permission     = btrfs_permission,
10652         .update_time    = btrfs_update_time,
10653 };
10654
10655 static const struct file_operations btrfs_dir_file_operations = {
10656         .llseek         = generic_file_llseek,
10657         .read           = generic_read_dir,
10658         .iterate_shared = btrfs_real_readdir,
10659         .open           = btrfs_opendir,
10660         .unlocked_ioctl = btrfs_ioctl,
10661 #ifdef CONFIG_COMPAT
10662         .compat_ioctl   = btrfs_compat_ioctl,
10663 #endif
10664         .release        = btrfs_release_file,
10665         .fsync          = btrfs_sync_file,
10666 };
10667
10668 static const struct extent_io_ops btrfs_extent_io_ops = {
10669         /* mandatory callbacks */
10670         .submit_bio_hook = btrfs_submit_bio_hook,
10671         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
10672         .merge_bio_hook = btrfs_merge_bio_hook,
10673         .readpage_io_failed_hook = btrfs_readpage_io_failed_hook,
10674         .tree_fs_info = iotree_fs_info,
10675         .set_range_writeback = btrfs_set_range_writeback,
10676
10677         /* optional callbacks */
10678         .fill_delalloc = run_delalloc_range,
10679         .writepage_end_io_hook = btrfs_writepage_end_io_hook,
10680         .writepage_start_hook = btrfs_writepage_start_hook,
10681         .set_bit_hook = btrfs_set_bit_hook,
10682         .clear_bit_hook = btrfs_clear_bit_hook,
10683         .merge_extent_hook = btrfs_merge_extent_hook,
10684         .split_extent_hook = btrfs_split_extent_hook,
10685         .check_extent_io_range = btrfs_check_extent_io_range,
10686 };
10687
10688 /*
10689  * btrfs doesn't support the bmap operation because swapfiles
10690  * use bmap to make a mapping of extents in the file.  They assume
10691  * these extents won't change over the life of the file and they
10692  * use the bmap result to do IO directly to the drive.
10693  *
10694  * the btrfs bmap call would return logical addresses that aren't
10695  * suitable for IO and they also will change frequently as COW
10696  * operations happen.  So, swapfile + btrfs == corruption.
10697  *
10698  * For now we're avoiding this by dropping bmap.
10699  */
10700 static const struct address_space_operations btrfs_aops = {
10701         .readpage       = btrfs_readpage,
10702         .writepage      = btrfs_writepage,
10703         .writepages     = btrfs_writepages,
10704         .readpages      = btrfs_readpages,
10705         .direct_IO      = btrfs_direct_IO,
10706         .invalidatepage = btrfs_invalidatepage,
10707         .releasepage    = btrfs_releasepage,
10708         .set_page_dirty = btrfs_set_page_dirty,
10709         .error_remove_page = generic_error_remove_page,
10710 };
10711
10712 static const struct address_space_operations btrfs_symlink_aops = {
10713         .readpage       = btrfs_readpage,
10714         .writepage      = btrfs_writepage,
10715         .invalidatepage = btrfs_invalidatepage,
10716         .releasepage    = btrfs_releasepage,
10717 };
10718
10719 static const struct inode_operations btrfs_file_inode_operations = {
10720         .getattr        = btrfs_getattr,
10721         .setattr        = btrfs_setattr,
10722         .listxattr      = btrfs_listxattr,
10723         .permission     = btrfs_permission,
10724         .fiemap         = btrfs_fiemap,
10725         .get_acl        = btrfs_get_acl,
10726         .set_acl        = btrfs_set_acl,
10727         .update_time    = btrfs_update_time,
10728 };
10729 static const struct inode_operations btrfs_special_inode_operations = {
10730         .getattr        = btrfs_getattr,
10731         .setattr        = btrfs_setattr,
10732         .permission     = btrfs_permission,
10733         .listxattr      = btrfs_listxattr,
10734         .get_acl        = btrfs_get_acl,
10735         .set_acl        = btrfs_set_acl,
10736         .update_time    = btrfs_update_time,
10737 };
10738 static const struct inode_operations btrfs_symlink_inode_operations = {
10739         .get_link       = page_get_link,
10740         .getattr        = btrfs_getattr,
10741         .setattr        = btrfs_setattr,
10742         .permission     = btrfs_permission,
10743         .listxattr      = btrfs_listxattr,
10744         .update_time    = btrfs_update_time,
10745 };
10746
10747 const struct dentry_operations btrfs_dentry_operations = {
10748         .d_delete       = btrfs_dentry_delete,
10749         .d_release      = btrfs_dentry_release,
10750 };