fs/btrfs/file.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/fs.h>
   7 #include <linux/pagemap.h>
   8 #include <linux/time.h>
   9 #include <linux/init.h>
  10 #include <linux/string.h>
  11 #include <linux/backing-dev.h>
  12 #include <linux/falloc.h>
  13 #include <linux/writeback.h>
  14 #include <linux/compat.h>
  15 #include <linux/slab.h>
  16 #include <linux/btrfs.h>
  17 #include <linux/uio.h>
  18 #include <linux/iversion.h>
  19 #include "ctree.h"
  20 #include "disk-io.h"
  21 #include "transaction.h"
  22 #include "btrfs_inode.h"
  23 #include "print-tree.h"
  24 #include "tree-log.h"
  25 #include "locking.h"
  26 #include "volumes.h"
  27 #include "qgroup.h"
  28 #include "compression.h"
  29 #include "delalloc-space.h"
  30 #include "reflink.h"
  31
  32 static struct kmem_cache *btrfs_inode_defrag_cachep;
  33 /*
  34  * when auto defrag is enabled we
  35  * queue up these defrag structs to remember which
  36  * inodes need defragging passes
  37  */
  38 struct inode_defrag {
  39         struct rb_node rb_node;
  40         /* objectid */
  41         u64 ino;
  42         /*
  43          * transid where the defrag was added, we search for
  44          * extents newer than this
  45          */
  46         u64 transid;
  47
  48         /* root objectid */
  49         u64 root;
  50
  51         /* last offset we were able to defrag */
  52         u64 last_offset;
  53
  54         /* if we've wrapped around back to zero once already */
  55         int cycled;
  56 };
  57
  58 static int __compare_inode_defrag(struct inode_defrag *defrag1,
  59                                   struct inode_defrag *defrag2)
  60 {
  61         if (defrag1->root > defrag2->root)
  62                 return 1;
  63         else if (defrag1->root < defrag2->root)
  64                 return -1;
  65         else if (defrag1->ino > defrag2->ino)
  66                 return 1;
  67         else if (defrag1->ino < defrag2->ino)
  68                 return -1;
  69         else
  70                 return 0;
  71 }
  72
  73 /* pop a record for an inode into the defrag tree.  The lock
  74  * must be held already
  75  *
  76  * If you're inserting a record for an older transid than an
  77  * existing record, the transid already in the tree is lowered
  78  *
  79  * If an existing record is found the defrag item you
  80  * pass in is freed
  81  */
  82 static int __btrfs_add_inode_defrag(struct btrfs_inode *inode,
  83                                     struct inode_defrag *defrag)
  84 {
  85         struct btrfs_fs_info *fs_info = inode->root->fs_info;
  86         struct inode_defrag *entry;
  87         struct rb_node **p;
  88         struct rb_node *parent = NULL;
  89         int ret;
  90
  91         p = &fs_info->defrag_inodes.rb_node;
  92         while (*p) {
  93                 parent = *p;
  94                 entry = rb_entry(parent, struct inode_defrag, rb_node);
  95
  96                 ret = __compare_inode_defrag(defrag, entry);
  97                 if (ret < 0)
  98                         p = &parent->rb_left;
  99                 else if (ret > 0)
 100                         p = &parent->rb_right;
 101                 else {
 102                         /* if we're reinserting an entry for
 103                          * an old defrag run, make sure to
 104                          * lower the transid of our existing record
 105                          */
 106                         if (defrag->transid < entry->transid)
 107                                 entry->transid = defrag->transid;
 108                         if (defrag->last_offset > entry->last_offset)
 109                                 entry->last_offset = defrag->last_offset;
 110                         return -EEXIST;
 111                 }
 112         }
 113         set_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags);
 114         rb_link_node(&defrag->rb_node, parent, p);
 115         rb_insert_color(&defrag->rb_node, &fs_info->defrag_inodes);
 116         return 0;
 117 }
 118
 119 static inline int __need_auto_defrag(struct btrfs_fs_info *fs_info)
 120 {
 121         if (!btrfs_test_opt(fs_info, AUTO_DEFRAG))
 122                 return 0;
 123
 124         if (btrfs_fs_closing(fs_info))
 125                 return 0;
 126
 127         return 1;
 128 }
 129
 130 /*
 131  * insert a defrag record for this inode if auto defrag is
 132  * enabled
 133  */
 134 int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
 135                            struct btrfs_inode *inode)
 136 {
 137         struct btrfs_root *root = inode->root;
 138         struct btrfs_fs_info *fs_info = root->fs_info;
 139         struct inode_defrag *defrag;
 140         u64 transid;
 141         int ret;
 142
 143         if (!__need_auto_defrag(fs_info))
 144                 return 0;
 145
 146         if (test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags))
 147                 return 0;
 148
 149         if (trans)
 150                 transid = trans->transid;
 151         else
 152                 transid = inode->root->last_trans;
 153
 154         defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
 155         if (!defrag)
 156                 return -ENOMEM;
 157
 158         defrag->ino = btrfs_ino(inode);
 159         defrag->transid = transid;
 160         defrag->root = root->root_key.objectid;
 161
 162         spin_lock(&fs_info->defrag_inodes_lock);
 163         if (!test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags)) {
 164                 /*
 165                  * If we set IN_DEFRAG flag and evict the inode from memory,
 166                  * and then re-read this inode, this new inode doesn't have
 167                  * IN_DEFRAG flag. At the case, we may find the existed defrag.
 168                  */
 169                 ret = __btrfs_add_inode_defrag(inode, defrag);
 170                 if (ret)
 171                         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 172         } else {
 173                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 174         }
 175         spin_unlock(&fs_info->defrag_inodes_lock);
 176         return 0;
 177 }
 178
 179 /*
 180  * Requeue the defrag object. If there is a defrag object that points to
 181  * the same inode in the tree, we will merge them together (by
 182  * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
 183  */
 184 static void btrfs_requeue_inode_defrag(struct btrfs_inode *inode,
 185                                        struct inode_defrag *defrag)
 186 {
 187         struct btrfs_fs_info *fs_info = inode->root->fs_info;
 188         int ret;
 189
 190         if (!__need_auto_defrag(fs_info))
 191                 goto out;
 192
 193         /*
 194          * Here we don't check the IN_DEFRAG flag, because we need merge
 195          * them together.
 196          */
 197         spin_lock(&fs_info->defrag_inodes_lock);
 198         ret = __btrfs_add_inode_defrag(inode, defrag);
 199         spin_unlock(&fs_info->defrag_inodes_lock);
 200         if (ret)
 201                 goto out;
 202         return;
 203 out:
 204         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 205 }
 206
 207 /*
 208  * pick the defragable inode that we want, if it doesn't exist, we will get
 209  * the next one.
 210  */
 211 static struct inode_defrag *
 212 btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
 213 {
 214         struct inode_defrag *entry = NULL;
 215         struct inode_defrag tmp;
 216         struct rb_node *p;
 217         struct rb_node *parent = NULL;
 218         int ret;
 219
 220         tmp.ino = ino;
 221         tmp.root = root;
 222
 223         spin_lock(&fs_info->defrag_inodes_lock);
 224         p = fs_info->defrag_inodes.rb_node;
 225         while (p) {
 226                 parent = p;
 227                 entry = rb_entry(parent, struct inode_defrag, rb_node);
 228
 229                 ret = __compare_inode_defrag(&tmp, entry);
 230                 if (ret < 0)
 231                         p = parent->rb_left;
 232                 else if (ret > 0)
 233                         p = parent->rb_right;
 234                 else
 235                         goto out;
 236         }
 237
 238         if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
 239                 parent = rb_next(parent);
 240                 if (parent)
 241                         entry = rb_entry(parent, struct inode_defrag, rb_node);
 242                 else
 243                         entry = NULL;
 244         }
 245 out:
 246         if (entry)
 247                 rb_erase(parent, &fs_info->defrag_inodes);
 248         spin_unlock(&fs_info->defrag_inodes_lock);
 249         return entry;
 250 }
 251
 252 void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
 253 {
 254         struct inode_defrag *defrag;
 255         struct rb_node *node;
 256
 257         spin_lock(&fs_info->defrag_inodes_lock);
 258         node = rb_first(&fs_info->defrag_inodes);
 259         while (node) {
 260                 rb_erase(node, &fs_info->defrag_inodes);
 261                 defrag = rb_entry(node, struct inode_defrag, rb_node);
 262                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 263
 264                 cond_resched_lock(&fs_info->defrag_inodes_lock);
 265
 266                 node = rb_first(&fs_info->defrag_inodes);
 267         }
 268         spin_unlock(&fs_info->defrag_inodes_lock);
 269 }
 270
 271 #define BTRFS_DEFRAG_BATCH      1024
 272
 273 static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
 274                                     struct inode_defrag *defrag)
 275 {
 276         struct btrfs_root *inode_root;
 277         struct inode *inode;
 278         struct btrfs_ioctl_defrag_range_args range;
 279         int num_defrag;
 280         int ret;
 281
 282         /* get the inode */
 283         inode_root = btrfs_get_fs_root(fs_info, defrag->root, true);
 284         if (IS_ERR(inode_root)) {
 285                 ret = PTR_ERR(inode_root);
 286                 goto cleanup;
 287         }
 288
 289         inode = btrfs_iget(fs_info->sb, defrag->ino, inode_root);
 290         btrfs_put_root(inode_root);
 291         if (IS_ERR(inode)) {
 292                 ret = PTR_ERR(inode);
 293                 goto cleanup;
 294         }
 295
 296         /* do a chunk of defrag */
 297         clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
 298         memset(&range, 0, sizeof(range));
 299         range.len = (u64)-1;
 300         range.start = defrag->last_offset;
 301
 302         sb_start_write(fs_info->sb);
 303         num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
 304                                        BTRFS_DEFRAG_BATCH);
 305         sb_end_write(fs_info->sb);
 306         /*
 307          * if we filled the whole defrag batch, there
 308          * must be more work to do.  Queue this defrag
 309          * again
 310          */
 311         if (num_defrag == BTRFS_DEFRAG_BATCH) {
 312                 defrag->last_offset = range.start;
 313                 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
 314         } else if (defrag->last_offset && !defrag->cycled) {
 315                 /*
 316                  * we didn't fill our defrag batch, but
 317                  * we didn't start at zero.  Make sure we loop
 318                  * around to the start of the file.
 319                  */
 320                 defrag->last_offset = 0;
 321                 defrag->cycled = 1;
 322                 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
 323         } else {
 324                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 325         }
 326
 327         iput(inode);
 328         return 0;
 329 cleanup:
 330         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 331         return ret;
 332 }
 333
 334 /*
 335  * run through the list of inodes in the FS that need
 336  * defragging
 337  */
 338 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
 339 {
 340         struct inode_defrag *defrag;
 341         u64 first_ino = 0;
 342         u64 root_objectid = 0;
 343
 344         atomic_inc(&fs_info->defrag_running);
 345         while (1) {
 346                 /* Pause the auto defragger. */
 347                 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
 348                              &fs_info->fs_state))
 349                         break;
 350
 351                 if (!__need_auto_defrag(fs_info))
 352                         break;
 353
 354                 /* find an inode to defrag */
 355                 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
 356                                                  first_ino);
 357                 if (!defrag) {
 358                         if (root_objectid || first_ino) {
 359                                 root_objectid = 0;
 360                                 first_ino = 0;
 361                                 continue;
 362                         } else {
 363                                 break;
 364                         }
 365                 }
 366
 367                 first_ino = defrag->ino + 1;
 368                 root_objectid = defrag->root;
 369
 370                 __btrfs_run_defrag_inode(fs_info, defrag);
 371         }
 372         atomic_dec(&fs_info->defrag_running);
 373
 374         /*
 375          * during unmount, we use the transaction_wait queue to
 376          * wait for the defragger to stop
 377          */
 378         wake_up(&fs_info->transaction_wait);
 379         return 0;
 380 }
 381
 382 /* simple helper to fault in pages and copy.  This should go away
 383  * and be replaced with calls into generic code.
 384  */
 385 static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
 386                                          struct page **prepared_pages,
 387                                          struct iov_iter *i)
 388 {
 389         size_t copied = 0;
 390         size_t total_copied = 0;
 391         int pg = 0;
 392         int offset = offset_in_page(pos);
 393
 394         while (write_bytes > 0) {
 395                 size_t count = min_t(size_t,
 396                                      PAGE_SIZE - offset, write_bytes);
 397                 struct page *page = prepared_pages[pg];
 398                 /*
 399                  * Copy data from userspace to the current page
 400                  */
 401                 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
 402
 403                 /* Flush processor's dcache for this page */
 404                 flush_dcache_page(page);
 405
 406                 /*
 407                  * if we get a partial write, we can end up with
 408                  * partially up to date pages.  These add
 409                  * a lot of complexity, so make sure they don't
 410                  * happen by forcing this copy to be retried.
 411                  *
 412                  * The rest of the btrfs_file_write code will fall
 413                  * back to page at a time copies after we return 0.
 414                  */
 415                 if (!PageUptodate(page) && copied < count)
 416                         copied = 0;
 417
 418                 iov_iter_advance(i, copied);
 419                 write_bytes -= copied;
 420                 total_copied += copied;
 421
 422                 /* Return to btrfs_file_write_iter to fault page */
 423                 if (unlikely(copied == 0))
 424                         break;
 425
 426                 if (copied < PAGE_SIZE - offset) {
 427                         offset += copied;
 428                 } else {
 429                         pg++;
 430                         offset = 0;
 431                 }
 432         }
 433         return total_copied;
 434 }
 435
 436 /*
 437  * unlocks pages after btrfs_file_write is done with them
 438  */
 439 static void btrfs_drop_pages(struct page **pages, size_t num_pages)
 440 {
 441         size_t i;
 442         for (i = 0; i < num_pages; i++) {
 443                 /* page checked is some magic around finding pages that
 444                  * have been modified without going through btrfs_set_page_dirty
 445                  * clear it here. There should be no need to mark the pages
 446                  * accessed as prepare_pages should have marked them accessed
 447                  * in prepare_pages via find_or_create_page()
 448                  */
 449                 ClearPageChecked(pages[i]);
 450                 unlock_page(pages[i]);
 451                 put_page(pages[i]);
 452         }
 453 }
 454
 455 /*
 456  * After btrfs_copy_from_user(), update the following things for delalloc:
 457  * - Mark newly dirtied pages as DELALLOC in the io tree.
 458  *   Used to advise which range is to be written back.
 459  * - Mark modified pages as Uptodate/Dirty and not needing COW fixup
 460  * - Update inode size for past EOF write
 461  */
 462 int btrfs_dirty_pages(struct btrfs_inode *inode, struct page **pages,
 463                       size_t num_pages, loff_t pos, size_t write_bytes,
 464                       struct extent_state **cached, bool noreserve)
 465 {
 466         struct btrfs_fs_info *fs_info = inode->root->fs_info;
 467         int err = 0;
 468         int i;
 469         u64 num_bytes;
 470         u64 start_pos;
 471         u64 end_of_last_block;
 472         u64 end_pos = pos + write_bytes;
 473         loff_t isize = i_size_read(&inode->vfs_inode);
 474         unsigned int extra_bits = 0;
 475
 476         if (write_bytes == 0)
 477                 return 0;
 478
 479         if (noreserve)
 480                 extra_bits |= EXTENT_NORESERVE;
 481
 482         start_pos = round_down(pos, fs_info->sectorsize);
 483         num_bytes = round_up(write_bytes + pos - start_pos,
 484                              fs_info->sectorsize);
 485
 486         end_of_last_block = start_pos + num_bytes - 1;
 487
 488         /*
 489          * The pages may have already been dirty, clear out old accounting so
 490          * we can set things up properly
 491          */
 492         clear_extent_bit(&inode->io_tree, start_pos, end_of_last_block,
 493                          EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
 494                          0, 0, cached);
 495
 496         err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
 497                                         extra_bits, cached);
 498         if (err)
 499                 return err;
 500
 501         for (i = 0; i < num_pages; i++) {
 502                 struct page *p = pages[i];
 503                 SetPageUptodate(p);
 504                 ClearPageChecked(p);
 505                 set_page_dirty(p);
 506         }
 507
 508         /*
 509          * we've only changed i_size in ram, and we haven't updated
 510          * the disk i_size.  There is no need to log the inode
 511          * at this time.
 512          */
 513         if (end_pos > isize)
 514                 i_size_write(&inode->vfs_inode, end_pos);
 515         return 0;
 516 }
 517
 518 /*
 519  * this drops all the extents in the cache that intersect the range
 520  * [start, end].  Existing extents are split as required.
 521  */
 522 void btrfs_drop_extent_cache(struct btrfs_inode *inode, u64 start, u64 end,
 523                              int skip_pinned)
 524 {
 525         struct extent_map *em;
 526         struct extent_map *split = NULL;
 527         struct extent_map *split2 = NULL;
 528         struct extent_map_tree *em_tree = &inode->extent_tree;
 529         u64 len = end - start + 1;
 530         u64 gen;
 531         int ret;
 532         int testend = 1;
 533         unsigned long flags;
 534         int compressed = 0;
 535         bool modified;
 536
 537         WARN_ON(end < start);
 538         if (end == (u64)-1) {
 539                 len = (u64)-1;
 540                 testend = 0;
 541         }
 542         while (1) {
 543                 int no_splits = 0;
 544
 545                 modified = false;
 546                 if (!split)
 547                         split = alloc_extent_map();
 548                 if (!split2)
 549                         split2 = alloc_extent_map();
 550                 if (!split || !split2)
 551                         no_splits = 1;
 552
 553                 write_lock(&em_tree->lock);
 554                 em = lookup_extent_mapping(em_tree, start, len);
 555                 if (!em) {
 556                         write_unlock(&em_tree->lock);
 557                         break;
 558                 }
 559                 flags = em->flags;
 560                 gen = em->generation;
 561                 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
 562                         if (testend && em->start + em->len >= start + len) {
 563                                 free_extent_map(em);
 564                                 write_unlock(&em_tree->lock);
 565                                 break;
 566                         }
 567                         start = em->start + em->len;
 568                         if (testend)
 569                                 len = start + len - (em->start + em->len);
 570                         free_extent_map(em);
 571                         write_unlock(&em_tree->lock);
 572                         continue;
 573                 }
 574                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 575                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
 576                 clear_bit(EXTENT_FLAG_LOGGING, &flags);
 577                 modified = !list_empty(&em->list);
 578                 if (no_splits)
 579                         goto next;
 580
 581                 if (em->start < start) {
 582                         split->start = em->start;
 583                         split->len = start - em->start;
 584
 585                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 586                                 split->orig_start = em->orig_start;
 587                                 split->block_start = em->block_start;
 588
 589                                 if (compressed)
 590                                         split->block_len = em->block_len;
 591                                 else
 592                                         split->block_len = split->len;
 593                                 split->orig_block_len = max(split->block_len,
 594                                                 em->orig_block_len);
 595                                 split->ram_bytes = em->ram_bytes;
 596                         } else {
 597                                 split->orig_start = split->start;
 598                                 split->block_len = 0;
 599                                 split->block_start = em->block_start;
 600                                 split->orig_block_len = 0;
 601                                 split->ram_bytes = split->len;
 602                         }
 603
 604                         split->generation = gen;
 605                         split->flags = flags;
 606                         split->compress_type = em->compress_type;
 607                         replace_extent_mapping(em_tree, em, split, modified);
 608                         free_extent_map(split);
 609                         split = split2;
 610                         split2 = NULL;
 611                 }
 612                 if (testend && em->start + em->len > start + len) {
 613                         u64 diff = start + len - em->start;
 614
 615                         split->start = start + len;
 616                         split->len = em->start + em->len - (start + len);
 617                         split->flags = flags;
 618                         split->compress_type = em->compress_type;
 619                         split->generation = gen;
 620
 621                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 622                                 split->orig_block_len = max(em->block_len,
 623                                                     em->orig_block_len);
 624
 625                                 split->ram_bytes = em->ram_bytes;
 626                                 if (compressed) {
 627                                         split->block_len = em->block_len;
 628                                         split->block_start = em->block_start;
 629                                         split->orig_start = em->orig_start;
 630                                 } else {
 631                                         split->block_len = split->len;
 632                                         split->block_start = em->block_start
 633                                                 + diff;
 634                                         split->orig_start = em->orig_start;
 635                                 }
 636                         } else {
 637                                 split->ram_bytes = split->len;
 638                                 split->orig_start = split->start;
 639                                 split->block_len = 0;
 640                                 split->block_start = em->block_start;
 641                                 split->orig_block_len = 0;
 642                         }
 643
 644                         if (extent_map_in_tree(em)) {
 645                                 replace_extent_mapping(em_tree, em, split,
 646                                                        modified);
 647                         } else {
 648                                 ret = add_extent_mapping(em_tree, split,
 649                                                          modified);
 650                                 ASSERT(ret == 0); /* Logic error */
 651                         }
 652                         free_extent_map(split);
 653                         split = NULL;
 654                 }
 655 next:
 656                 if (extent_map_in_tree(em))
 657                         remove_extent_mapping(em_tree, em);
 658                 write_unlock(&em_tree->lock);
 659
 660                 /* once for us */
 661                 free_extent_map(em);
 662                 /* once for the tree*/
 663                 free_extent_map(em);
 664         }
 665         if (split)
 666                 free_extent_map(split);
 667         if (split2)
 668                 free_extent_map(split2);
 669 }
 670
 671 /*
 672  * this is very complex, but the basic idea is to drop all extents
 673  * in the range start - end.  hint_block is filled in with a block number
 674  * that would be a good hint to the block allocator for this file.
 675  *
 676  * If an extent intersects the range but is not entirely inside the range
 677  * it is either truncated or split.  Anything entirely inside the range
 678  * is deleted from the tree.
 679  *
 680  * Note: the VFS' inode number of bytes is not updated, it's up to the caller
 681  * to deal with that. We set the field 'bytes_found' of the arguments structure
 682  * with the number of allocated bytes found in the target range, so that the
 683  * caller can update the inode's number of bytes in an atomic way when
 684  * replacing extents in a range to avoid races with stat(2).
 685  */
 686 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
 687                        struct btrfs_root *root, struct btrfs_inode *inode,
 688                        struct btrfs_drop_extents_args *args)
 689 {
 690         struct btrfs_fs_info *fs_info = root->fs_info;
 691         struct extent_buffer *leaf;
 692         struct btrfs_file_extent_item *fi;
 693         struct btrfs_ref ref = { 0 };
 694         struct btrfs_key key;
 695         struct btrfs_key new_key;
 696         u64 ino = btrfs_ino(inode);
 697         u64 search_start = args->start;
 698         u64 disk_bytenr = 0;
 699         u64 num_bytes = 0;
 700         u64 extent_offset = 0;
 701         u64 extent_end = 0;
 702         u64 last_end = args->start;
 703         int del_nr = 0;
 704         int del_slot = 0;
 705         int extent_type;
 706         int recow;
 707         int ret;
 708         int modify_tree = -1;
 709         int update_refs;
 710         int found = 0;
 711         int leafs_visited = 0;
 712         struct btrfs_path *path = args->path;
 713
 714         args->bytes_found = 0;
 715         args->extent_inserted = false;
 716
 717         /* Must always have a path if ->replace_extent is true */
 718         ASSERT(!(args->replace_extent && !args->path));
 719
 720         if (!path) {
 721                 path = btrfs_alloc_path();
 722                 if (!path) {
 723                         ret = -ENOMEM;
 724                         goto out;
 725                 }
 726         }
 727
 728         if (args->drop_cache)
 729                 btrfs_drop_extent_cache(inode, args->start, args->end - 1, 0);
 730
 731         if (args->start >= inode->disk_i_size && !args->replace_extent)
 732                 modify_tree = 0;
 733
 734         update_refs = (test_bit(BTRFS_ROOT_SHAREABLE, &root->state) ||
 735                        root == fs_info->tree_root);
 736         while (1) {
 737                 recow = 0;
 738                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
 739                                                search_start, modify_tree);
 740                 if (ret < 0)
 741                         break;
 742                 if (ret > 0 && path->slots[0] > 0 && search_start == args->start) {
 743                         leaf = path->nodes[0];
 744                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
 745                         if (key.objectid == ino &&
 746                             key.type == BTRFS_EXTENT_DATA_KEY)
 747                                 path->slots[0]--;
 748                 }
 749                 ret = 0;
 750                 leafs_visited++;
 751 next_slot:
 752                 leaf = path->nodes[0];
 753                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 754                         BUG_ON(del_nr > 0);
 755                         ret = btrfs_next_leaf(root, path);
 756                         if (ret < 0)
 757                                 break;
 758                         if (ret > 0) {
 759                                 ret = 0;
 760                                 break;
 761                         }
 762                         leafs_visited++;
 763                         leaf = path->nodes[0];
 764                         recow = 1;
 765                 }
 766
 767                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 768
 769                 if (key.objectid > ino)
 770                         break;
 771                 if (WARN_ON_ONCE(key.objectid < ino) ||
 772                     key.type < BTRFS_EXTENT_DATA_KEY) {
 773                         ASSERT(del_nr == 0);
 774                         path->slots[0]++;
 775                         goto next_slot;
 776                 }
 777                 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= args->end)
 778                         break;
 779
 780                 fi = btrfs_item_ptr(leaf, path->slots[0],
 781                                     struct btrfs_file_extent_item);
 782                 extent_type = btrfs_file_extent_type(leaf, fi);
 783
 784                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
 785                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
 786                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 787                         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 788                         extent_offset = btrfs_file_extent_offset(leaf, fi);
 789                         extent_end = key.offset +
 790                                 btrfs_file_extent_num_bytes(leaf, fi);
 791                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 792                         extent_end = key.offset +
 793                                 btrfs_file_extent_ram_bytes(leaf, fi);
 794                 } else {
 795                         /* can't happen */
 796                         BUG();
 797                 }
 798
 799                 /*
 800                  * Don't skip extent items representing 0 byte lengths. They
 801                  * used to be created (bug) if while punching holes we hit
 802                  * -ENOSPC condition. So if we find one here, just ensure we
 803                  * delete it, otherwise we would insert a new file extent item
 804                  * with the same key (offset) as that 0 bytes length file
 805                  * extent item in the call to setup_items_for_insert() later
 806                  * in this function.
 807                  */
 808                 if (extent_end == key.offset && extent_end >= search_start) {
 809                         last_end = extent_end;
 810                         goto delete_extent_item;
 811                 }
 812
 813                 if (extent_end <= search_start) {
 814                         path->slots[0]++;
 815                         goto next_slot;
 816                 }
 817
 818                 found = 1;
 819                 search_start = max(key.offset, args->start);
 820                 if (recow || !modify_tree) {
 821                         modify_tree = -1;
 822                         btrfs_release_path(path);
 823                         continue;
 824                 }
 825
 826                 /*
 827                  *     | - range to drop - |
 828                  *  | -------- extent -------- |
 829                  */
 830                 if (args->start > key.offset && args->end < extent_end) {
 831                         BUG_ON(del_nr > 0);
 832                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 833                                 ret = -EOPNOTSUPP;
 834                                 break;
 835                         }
 836
 837                         memcpy(&new_key, &key, sizeof(new_key));
 838                         new_key.offset = args->start;
 839                         ret = btrfs_duplicate_item(trans, root, path,
 840                                                    &new_key);
 841                         if (ret == -EAGAIN) {
 842                                 btrfs_release_path(path);
 843                                 continue;
 844                         }
 845                         if (ret < 0)
 846                                 break;
 847
 848                         leaf = path->nodes[0];
 849                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 850                                             struct btrfs_file_extent_item);
 851                         btrfs_set_file_extent_num_bytes(leaf, fi,
 852                                                         args->start - key.offset);
 853
 854                         fi = btrfs_item_ptr(leaf, path->slots[0],
 855                                             struct btrfs_file_extent_item);
 856
 857                         extent_offset += args->start - key.offset;
 858                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 859                         btrfs_set_file_extent_num_bytes(leaf, fi,
 860                                                         extent_end - args->start);
 861                         btrfs_mark_buffer_dirty(leaf);
 862
 863                         if (update_refs && disk_bytenr > 0) {
 864                                 btrfs_init_generic_ref(&ref,
 865                                                 BTRFS_ADD_DELAYED_REF,
 866                                                 disk_bytenr, num_bytes, 0);
 867                                 btrfs_init_data_ref(&ref,
 868                                                 root->root_key.objectid,
 869                                                 new_key.objectid,
 870                                                 args->start - extent_offset);
 871                                 ret = btrfs_inc_extent_ref(trans, &ref);
 872                                 BUG_ON(ret); /* -ENOMEM */
 873                         }
 874                         key.offset = args->start;
 875                 }
 876                 /*
 877                  * From here on out we will have actually dropped something, so
 878                  * last_end can be updated.
 879                  */
 880                 last_end = extent_end;
 881
 882                 /*
 883                  *  | ---- range to drop ----- |
 884                  *      | -------- extent -------- |
 885                  */
 886                 if (args->start <= key.offset && args->end < extent_end) {
 887                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 888                                 ret = -EOPNOTSUPP;
 889                                 break;
 890                         }
 891
 892                         memcpy(&new_key, &key, sizeof(new_key));
 893                         new_key.offset = args->end;
 894                         btrfs_set_item_key_safe(fs_info, path, &new_key);
 895
 896                         extent_offset += args->end - key.offset;
 897                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 898                         btrfs_set_file_extent_num_bytes(leaf, fi,
 899                                                         extent_end - args->end);
 900                         btrfs_mark_buffer_dirty(leaf);
 901                         if (update_refs && disk_bytenr > 0)
 902                                 args->bytes_found += args->end - key.offset;
 903                         break;
 904                 }
 905
 906                 search_start = extent_end;
 907                 /*
 908                  *       | ---- range to drop ----- |
 909                  *  | -------- extent -------- |
 910                  */
 911                 if (args->start > key.offset && args->end >= extent_end) {
 912                         BUG_ON(del_nr > 0);
 913                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 914                                 ret = -EOPNOTSUPP;
 915                                 break;
 916                         }
 917
 918                         btrfs_set_file_extent_num_bytes(leaf, fi,
 919                                                         args->start - key.offset);
 920                         btrfs_mark_buffer_dirty(leaf);
 921                         if (update_refs && disk_bytenr > 0)
 922                                 args->bytes_found += extent_end - args->start;
 923                         if (args->end == extent_end)
 924                                 break;
 925
 926                         path->slots[0]++;
 927                         goto next_slot;
 928                 }
 929
 930                 /*
 931                  *  | ---- range to drop ----- |
 932                  *    | ------ extent ------ |
 933                  */
 934                 if (args->start <= key.offset && args->end >= extent_end) {
 935 delete_extent_item:
 936                         if (del_nr == 0) {
 937                                 del_slot = path->slots[0];
 938                                 del_nr = 1;
 939                         } else {
 940                                 BUG_ON(del_slot + del_nr != path->slots[0]);
 941                                 del_nr++;
 942                         }
 943
 944                         if (update_refs &&
 945                             extent_type == BTRFS_FILE_EXTENT_INLINE) {
 946                                 args->bytes_found += extent_end - key.offset;
 947                                 extent_end = ALIGN(extent_end,
 948                                                    fs_info->sectorsize);
 949                         } else if (update_refs && disk_bytenr > 0) {
 950                                 btrfs_init_generic_ref(&ref,
 951                                                 BTRFS_DROP_DELAYED_REF,
 952                                                 disk_bytenr, num_bytes, 0);
 953                                 btrfs_init_data_ref(&ref,
 954                                                 root->root_key.objectid,
 955                                                 key.objectid,
 956                                                 key.offset - extent_offset);
 957                                 ret = btrfs_free_extent(trans, &ref);
 958                                 BUG_ON(ret); /* -ENOMEM */
 959                                 args->bytes_found += extent_end - key.offset;
 960                         }
 961
 962                         if (args->end == extent_end)
 963                                 break;
 964
 965                         if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
 966                                 path->slots[0]++;
 967                                 goto next_slot;
 968                         }
 969
 970                         ret = btrfs_del_items(trans, root, path, del_slot,
 971                                               del_nr);
 972                         if (ret) {
 973                                 btrfs_abort_transaction(trans, ret);
 974                                 break;
 975                         }
 976
 977                         del_nr = 0;
 978                         del_slot = 0;
 979
 980                         btrfs_release_path(path);
 981                         continue;
 982                 }
 983
 984                 BUG();
 985         }
 986
 987         if (!ret && del_nr > 0) {
 988                 /*
 989                  * Set path->slots[0] to first slot, so that after the delete
 990                  * if items are move off from our leaf to its immediate left or
 991                  * right neighbor leafs, we end up with a correct and adjusted
 992                  * path->slots[0] for our insertion (if args->replace_extent).
 993                  */
 994                 path->slots[0] = del_slot;
 995                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
 996                 if (ret)
 997                         btrfs_abort_transaction(trans, ret);
 998         }
 999
1000         leaf = path->nodes[0];
1001         /*
1002          * If btrfs_del_items() was called, it might have deleted a leaf, in
1003          * which case it unlocked our path, so check path->locks[0] matches a
1004          * write lock.
1005          */
1006         if (!ret && args->replace_extent && leafs_visited == 1 &&
1007             path->locks[0] == BTRFS_WRITE_LOCK &&
1008             btrfs_leaf_free_space(leaf) >=
1009             sizeof(struct btrfs_item) + args->extent_item_size) {
1010
1011                 key.objectid = ino;
1012                 key.type = BTRFS_EXTENT_DATA_KEY;
1013                 key.offset = args->start;
1014                 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
1015                         struct btrfs_key slot_key;
1016
1017                         btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
1018                         if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
1019                                 path->slots[0]++;
1020                 }
1021                 setup_items_for_insert(root, path, &key,
1022                                        &args->extent_item_size, 1);
1023                 args->extent_inserted = true;
1024         }
1025
1026         if (!args->path)
1027                 btrfs_free_path(path);
1028         else if (!args->extent_inserted)
1029                 btrfs_release_path(path);
1030 out:
1031         args->drop_end = found ? min(args->end, last_end) : args->end;
1032
1033         return ret;
1034 }
1035
1036 static int extent_mergeable(struct extent_buffer *leaf, int slot,
1037                             u64 objectid, u64 bytenr, u64 orig_offset,
1038                             u64 *start, u64 *end)
1039 {
1040         struct btrfs_file_extent_item *fi;
1041         struct btrfs_key key;
1042         u64 extent_end;
1043
1044         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1045                 return 0;
1046
1047         btrfs_item_key_to_cpu(leaf, &key, slot);
1048         if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1049                 return 0;
1050
1051         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1052         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1053             btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1054             btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1055             btrfs_file_extent_compression(leaf, fi) ||
1056             btrfs_file_extent_encryption(leaf, fi) ||
1057             btrfs_file_extent_other_encoding(leaf, fi))
1058                 return 0;
1059
1060         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1061         if ((*start && *start != key.offset) || (*end && *end != extent_end))
1062                 return 0;
1063
1064         *start = key.offset;
1065         *end = extent_end;
1066         return 1;
1067 }
1068
1069 /*
1070  * Mark extent in the range start - end as written.
1071  *
1072  * This changes extent type from 'pre-allocated' to 'regular'. If only
1073  * part of extent is marked as written, the extent will be split into
1074  * two or three.
1075  */
1076 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1077                               struct btrfs_inode *inode, u64 start, u64 end)
1078 {
1079         struct btrfs_fs_info *fs_info = trans->fs_info;
1080         struct btrfs_root *root = inode->root;
1081         struct extent_buffer *leaf;
1082         struct btrfs_path *path;
1083         struct btrfs_file_extent_item *fi;
1084         struct btrfs_ref ref = { 0 };
1085         struct btrfs_key key;
1086         struct btrfs_key new_key;
1087         u64 bytenr;
1088         u64 num_bytes;
1089         u64 extent_end;
1090         u64 orig_offset;
1091         u64 other_start;
1092         u64 other_end;
1093         u64 split;
1094         int del_nr = 0;
1095         int del_slot = 0;
1096         int recow;
1097         int ret;
1098         u64 ino = btrfs_ino(inode);
1099
1100         path = btrfs_alloc_path();
1101         if (!path)
1102                 return -ENOMEM;
1103 again:
1104         recow = 0;
1105         split = start;
1106         key.objectid = ino;
1107         key.type = BTRFS_EXTENT_DATA_KEY;
1108         key.offset = split;
1109
1110         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1111         if (ret < 0)
1112                 goto out;
1113         if (ret > 0 && path->slots[0] > 0)
1114                 path->slots[0]--;
1115
1116         leaf = path->nodes[0];
1117         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1118         if (key.objectid != ino ||
1119             key.type != BTRFS_EXTENT_DATA_KEY) {
1120                 ret = -EINVAL;
1121                 btrfs_abort_transaction(trans, ret);
1122                 goto out;
1123         }
1124         fi = btrfs_item_ptr(leaf, path->slots[0],
1125                             struct btrfs_file_extent_item);
1126         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
1127                 ret = -EINVAL;
1128                 btrfs_abort_transaction(trans, ret);
1129                 goto out;
1130         }
1131         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1132         if (key.offset > start || extent_end < end) {
1133                 ret = -EINVAL;
1134                 btrfs_abort_transaction(trans, ret);
1135                 goto out;
1136         }
1137
1138         bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1139         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1140         orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1141         memcpy(&new_key, &key, sizeof(new_key));
1142
1143         if (start == key.offset && end < extent_end) {
1144                 other_start = 0;
1145                 other_end = start;
1146                 if (extent_mergeable(leaf, path->slots[0] - 1,
1147                                      ino, bytenr, orig_offset,
1148                                      &other_start, &other_end)) {
1149                         new_key.offset = end;
1150                         btrfs_set_item_key_safe(fs_info, path, &new_key);
1151                         fi = btrfs_item_ptr(leaf, path->slots[0],
1152                                             struct btrfs_file_extent_item);
1153                         btrfs_set_file_extent_generation(leaf, fi,
1154                                                          trans->transid);
1155                         btrfs_set_file_extent_num_bytes(leaf, fi,
1156                                                         extent_end - end);
1157                         btrfs_set_file_extent_offset(leaf, fi,
1158                                                      end - orig_offset);
1159                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1160                                             struct btrfs_file_extent_item);
1161                         btrfs_set_file_extent_generation(leaf, fi,
1162                                                          trans->transid);
1163                         btrfs_set_file_extent_num_bytes(leaf, fi,
1164                                                         end - other_start);
1165                         btrfs_mark_buffer_dirty(leaf);
1166                         goto out;
1167                 }
1168         }
1169
1170         if (start > key.offset && end == extent_end) {
1171                 other_start = end;
1172                 other_end = 0;
1173                 if (extent_mergeable(leaf, path->slots[0] + 1,
1174                                      ino, bytenr, orig_offset,
1175                                      &other_start, &other_end)) {
1176                         fi = btrfs_item_ptr(leaf, path->slots[0],
1177                                             struct btrfs_file_extent_item);
1178                         btrfs_set_file_extent_num_bytes(leaf, fi,
1179                                                         start - key.offset);
1180                         btrfs_set_file_extent_generation(leaf, fi,
1181                                                          trans->transid);
1182                         path->slots[0]++;
1183                         new_key.offset = start;
1184                         btrfs_set_item_key_safe(fs_info, path, &new_key);
1185
1186                         fi = btrfs_item_ptr(leaf, path->slots[0],
1187                                             struct btrfs_file_extent_item);
1188                         btrfs_set_file_extent_generation(leaf, fi,
1189                                                          trans->transid);
1190                         btrfs_set_file_extent_num_bytes(leaf, fi,
1191                                                         other_end - start);
1192                         btrfs_set_file_extent_offset(leaf, fi,
1193                                                      start - orig_offset);
1194                         btrfs_mark_buffer_dirty(leaf);
1195                         goto out;
1196                 }
1197         }
1198
1199         while (start > key.offset || end < extent_end) {
1200                 if (key.offset == start)
1201                         split = end;
1202
1203                 new_key.offset = split;
1204                 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1205                 if (ret == -EAGAIN) {
1206                         btrfs_release_path(path);
1207                         goto again;
1208                 }
1209                 if (ret < 0) {
1210                         btrfs_abort_transaction(trans, ret);
1211                         goto out;
1212                 }
1213
1214                 leaf = path->nodes[0];
1215                 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1216                                     struct btrfs_file_extent_item);
1217                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1218                 btrfs_set_file_extent_num_bytes(leaf, fi,
1219                                                 split - key.offset);
1220
1221                 fi = btrfs_item_ptr(leaf, path->slots[0],
1222                                     struct btrfs_file_extent_item);
1223
1224                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1225                 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1226                 btrfs_set_file_extent_num_bytes(leaf, fi,
1227                                                 extent_end - split);
1228                 btrfs_mark_buffer_dirty(leaf);
1229
1230                 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, bytenr,
1231                                        num_bytes, 0);
1232                 btrfs_init_data_ref(&ref, root->root_key.objectid, ino,
1233                                     orig_offset);
1234                 ret = btrfs_inc_extent_ref(trans, &ref);
1235                 if (ret) {
1236                         btrfs_abort_transaction(trans, ret);
1237                         goto out;
1238                 }
1239
1240                 if (split == start) {
1241                         key.offset = start;
1242                 } else {
1243                         if (start != key.offset) {
1244                                 ret = -EINVAL;
1245                                 btrfs_abort_transaction(trans, ret);
1246                                 goto out;
1247                         }
1248                         path->slots[0]--;
1249                         extent_end = end;
1250                 }
1251                 recow = 1;
1252         }
1253
1254         other_start = end;
1255         other_end = 0;
1256         btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
1257                                num_bytes, 0);
1258         btrfs_init_data_ref(&ref, root->root_key.objectid, ino, orig_offset);
1259         if (extent_mergeable(leaf, path->slots[0] + 1,
1260                              ino, bytenr, orig_offset,
1261                              &other_start, &other_end)) {
1262                 if (recow) {
1263                         btrfs_release_path(path);
1264                         goto again;
1265                 }
1266                 extent_end = other_end;
1267                 del_slot = path->slots[0] + 1;
1268                 del_nr++;
1269                 ret = btrfs_free_extent(trans, &ref);
1270                 if (ret) {
1271                         btrfs_abort_transaction(trans, ret);
1272                         goto out;
1273                 }
1274         }
1275         other_start = 0;
1276         other_end = start;
1277         if (extent_mergeable(leaf, path->slots[0] - 1,
1278                              ino, bytenr, orig_offset,
1279                              &other_start, &other_end)) {
1280                 if (recow) {
1281                         btrfs_release_path(path);
1282                         goto again;
1283                 }
1284                 key.offset = other_start;
1285                 del_slot = path->slots[0];
1286                 del_nr++;
1287                 ret = btrfs_free_extent(trans, &ref);
1288                 if (ret) {
1289                         btrfs_abort_transaction(trans, ret);
1290                         goto out;
1291                 }
1292         }
1293         if (del_nr == 0) {
1294                 fi = btrfs_item_ptr(leaf, path->slots[0],
1295                            struct btrfs_file_extent_item);
1296                 btrfs_set_file_extent_type(leaf, fi,
1297                                            BTRFS_FILE_EXTENT_REG);
1298                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1299                 btrfs_mark_buffer_dirty(leaf);
1300         } else {
1301                 fi = btrfs_item_ptr(leaf, del_slot - 1,
1302                            struct btrfs_file_extent_item);
1303                 btrfs_set_file_extent_type(leaf, fi,
1304                                            BTRFS_FILE_EXTENT_REG);
1305                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1306                 btrfs_set_file_extent_num_bytes(leaf, fi,
1307                                                 extent_end - key.offset);
1308                 btrfs_mark_buffer_dirty(leaf);
1309
1310                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1311                 if (ret < 0) {
1312                         btrfs_abort_transaction(trans, ret);
1313                         goto out;
1314                 }
1315         }
1316 out:
1317         btrfs_free_path(path);
1318         return 0;
1319 }
1320
1321 /*
1322  * on error we return an unlocked page and the error value
1323  * on success we return a locked page and 0
1324  */
1325 static int prepare_uptodate_page(struct inode *inode,
1326                                  struct page *page, u64 pos,
1327                                  bool force_uptodate)
1328 {
1329         int ret = 0;
1330
1331         if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1332             !PageUptodate(page)) {
1333                 ret = btrfs_readpage(NULL, page);
1334                 if (ret)
1335                         return ret;
1336                 lock_page(page);
1337                 if (!PageUptodate(page)) {
1338                         unlock_page(page);
1339                         return -EIO;
1340                 }
1341                 if (page->mapping != inode->i_mapping) {
1342                         unlock_page(page);
1343                         return -EAGAIN;
1344                 }
1345         }
1346         return 0;
1347 }
1348
1349 /*
1350  * this just gets pages into the page cache and locks them down.
1351  */
1352 static noinline int prepare_pages(struct inode *inode, struct page **pages,
1353                                   size_t num_pages, loff_t pos,
1354                                   size_t write_bytes, bool force_uptodate)
1355 {
1356         int i;
1357         unsigned long index = pos >> PAGE_SHIFT;
1358         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1359         int err = 0;
1360         int faili;
1361
1362         for (i = 0; i < num_pages; i++) {
1363 again:
1364                 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1365                                                mask | __GFP_WRITE);
1366                 if (!pages[i]) {
1367                         faili = i - 1;
1368                         err = -ENOMEM;
1369                         goto fail;
1370                 }
1371
1372                 err = set_page_extent_mapped(pages[i]);
1373                 if (err < 0) {
1374                         faili = i;
1375                         goto fail;
1376                 }
1377
1378                 if (i == 0)
1379                         err = prepare_uptodate_page(inode, pages[i], pos,
1380                                                     force_uptodate);
1381                 if (!err && i == num_pages - 1)
1382                         err = prepare_uptodate_page(inode, pages[i],
1383                                                     pos + write_bytes, false);
1384                 if (err) {
1385                         put_page(pages[i]);
1386                         if (err == -EAGAIN) {
1387                                 err = 0;
1388                                 goto again;
1389                         }
1390                         faili = i - 1;
1391                         goto fail;
1392                 }
1393                 wait_on_page_writeback(pages[i]);
1394         }
1395
1396         return 0;
1397 fail:
1398         while (faili >= 0) {
1399                 unlock_page(pages[faili]);
1400                 put_page(pages[faili]);
1401                 faili--;
1402         }
1403         return err;
1404
1405 }
1406
1407 /*
1408  * This function locks the extent and properly waits for data=ordered extents
1409  * to finish before allowing the pages to be modified if need.
1410  *
1411  * The return value:
1412  * 1 - the extent is locked
1413  * 0 - the extent is not locked, and everything is OK
1414  * -EAGAIN - need re-prepare the pages
1415  * the other < 0 number - Something wrong happens
1416  */
1417 static noinline int
1418 lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct page **pages,
1419                                 size_t num_pages, loff_t pos,
1420                                 size_t write_bytes,
1421                                 u64 *lockstart, u64 *lockend,
1422                                 struct extent_state **cached_state)
1423 {
1424         struct btrfs_fs_info *fs_info = inode->root->fs_info;
1425         u64 start_pos;
1426         u64 last_pos;
1427         int i;
1428         int ret = 0;
1429
1430         start_pos = round_down(pos, fs_info->sectorsize);
1431         last_pos = round_up(pos + write_bytes, fs_info->sectorsize) - 1;
1432
1433         if (start_pos < inode->vfs_inode.i_size) {
1434                 struct btrfs_ordered_extent *ordered;
1435
1436                 lock_extent_bits(&inode->io_tree, start_pos, last_pos,
1437                                 cached_state);
1438                 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1439                                                      last_pos - start_pos + 1);
1440                 if (ordered &&
1441                     ordered->file_offset + ordered->num_bytes > start_pos &&
1442                     ordered->file_offset <= last_pos) {
1443                         unlock_extent_cached(&inode->io_tree, start_pos,
1444                                         last_pos, cached_state);
1445                         for (i = 0; i < num_pages; i++) {
1446                                 unlock_page(pages[i]);
1447                                 put_page(pages[i]);
1448                         }
1449                         btrfs_start_ordered_extent(ordered, 1);
1450                         btrfs_put_ordered_extent(ordered);
1451                         return -EAGAIN;
1452                 }
1453                 if (ordered)
1454                         btrfs_put_ordered_extent(ordered);
1455
1456                 *lockstart = start_pos;
1457                 *lockend = last_pos;
1458                 ret = 1;
1459         }
1460
1461         /*
1462          * We should be called after prepare_pages() which should have locked
1463          * all pages in the range.
1464          */
1465         for (i = 0; i < num_pages; i++)
1466                 WARN_ON(!PageLocked(pages[i]));
1467
1468         return ret;
1469 }
1470
1471 static int check_can_nocow(struct btrfs_inode *inode, loff_t pos,
1472                            size_t *write_bytes, bool nowait)
1473 {
1474         struct btrfs_fs_info *fs_info = inode->root->fs_info;
1475         struct btrfs_root *root = inode->root;
1476         u64 lockstart, lockend;
1477         u64 num_bytes;
1478         int ret;
1479
1480         if (!(inode->flags & (BTRFS_INODE_NODATACOW | BTRFS_INODE_PREALLOC)))
1481                 return 0;
1482
1483         if (!nowait && !btrfs_drew_try_write_lock(&root->snapshot_lock))
1484                 return -EAGAIN;
1485
1486         lockstart = round_down(pos, fs_info->sectorsize);
1487         lockend = round_up(pos + *write_bytes,
1488                            fs_info->sectorsize) - 1;
1489         num_bytes = lockend - lockstart + 1;
1490
1491         if (nowait) {
1492                 struct btrfs_ordered_extent *ordered;
1493
1494                 if (!try_lock_extent(&inode->io_tree, lockstart, lockend))
1495                         return -EAGAIN;
1496
1497                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
1498                                                      num_bytes);
1499                 if (ordered) {
1500                         btrfs_put_ordered_extent(ordered);
1501                         ret = -EAGAIN;
1502                         goto out_unlock;
1503                 }
1504         } else {
1505                 btrfs_lock_and_flush_ordered_range(inode, lockstart,
1506                                                    lockend, NULL);
1507         }
1508
1509         ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1510                         NULL, NULL, NULL, false);
1511         if (ret <= 0) {
1512                 ret = 0;
1513                 if (!nowait)
1514                         btrfs_drew_write_unlock(&root->snapshot_lock);
1515         } else {
1516                 *write_bytes = min_t(size_t, *write_bytes ,
1517                                      num_bytes - pos + lockstart);
1518         }
1519 out_unlock:
1520         unlock_extent(&inode->io_tree, lockstart, lockend);
1521
1522         return ret;
1523 }
1524
1525 static int check_nocow_nolock(struct btrfs_inode *inode, loff_t pos,
1526                               size_t *write_bytes)
1527 {
1528         return check_can_nocow(inode, pos, write_bytes, true);
1529 }
1530
1531 /*
1532  * Check if we can do nocow write into the range [@pos, @pos + @write_bytes)
1533  *
1534  * @pos:         File offset
1535  * @write_bytes: The length to write, will be updated to the nocow writeable
1536  *               range
1537  *
1538  * This function will flush ordered extents in the range to ensure proper
1539  * nocow checks.
1540  *
1541  * Return:
1542  * >0           and update @write_bytes if we can do nocow write
1543  *  0           if we can't do nocow write
1544  * -EAGAIN      if we can't get the needed lock or there are ordered extents
1545  *              for * (nowait == true) case
1546  * <0           if other error happened
1547  *
1548  * NOTE: Callers need to release the lock by btrfs_check_nocow_unlock().
1549  */
1550 int btrfs_check_nocow_lock(struct btrfs_inode *inode, loff_t pos,
1551                            size_t *write_bytes)
1552 {
1553         return check_can_nocow(inode, pos, write_bytes, false);
1554 }
1555
1556 void btrfs_check_nocow_unlock(struct btrfs_inode *inode)
1557 {
1558         btrfs_drew_write_unlock(&inode->root->snapshot_lock);
1559 }
1560
1561 static void update_time_for_write(struct inode *inode)
1562 {
1563         struct timespec64 now;
1564
1565         if (IS_NOCMTIME(inode))
1566                 return;
1567
1568         now = current_time(inode);
1569         if (!timespec64_equal(&inode->i_mtime, &now))
1570                 inode->i_mtime = now;
1571
1572         if (!timespec64_equal(&inode->i_ctime, &now))
1573                 inode->i_ctime = now;
1574
1575         if (IS_I_VERSION(inode))
1576                 inode_inc_iversion(inode);
1577 }
1578
1579 static int btrfs_write_check(struct kiocb *iocb, struct iov_iter *from,
1580                              size_t count)
1581 {
1582         struct file *file = iocb->ki_filp;
1583         struct inode *inode = file_inode(file);
1584         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1585         loff_t pos = iocb->ki_pos;
1586         int ret;
1587         loff_t oldsize;
1588         loff_t start_pos;
1589
1590         if (iocb->ki_flags & IOCB_NOWAIT) {
1591                 size_t nocow_bytes = count;
1592
1593                 /* We will allocate space in case nodatacow is not set, so bail */
1594                 if (check_nocow_nolock(BTRFS_I(inode), pos, &nocow_bytes) <= 0)
1595                         return -EAGAIN;
1596                 /*
1597                  * There are holes in the range or parts of the range that must
1598                  * be COWed (shared extents, RO block groups, etc), so just bail
1599                  * out.
1600                  */
1601                 if (nocow_bytes < count)
1602                         return -EAGAIN;
1603         }
1604
1605         current->backing_dev_info = inode_to_bdi(inode);
1606         ret = file_remove_privs(file);
1607         if (ret)
1608                 return ret;
1609
1610         /*
1611          * We reserve space for updating the inode when we reserve space for the
1612          * extent we are going to write, so we will enospc out there.  We don't
1613          * need to start yet another transaction to update the inode as we will
1614          * update the inode when we finish writing whatever data we write.
1615          */
1616         update_time_for_write(inode);
1617
1618         start_pos = round_down(pos, fs_info->sectorsize);
1619         oldsize = i_size_read(inode);
1620         if (start_pos > oldsize) {
1621                 /* Expand hole size to cover write data, preventing empty gap */
1622                 loff_t end_pos = round_up(pos + count, fs_info->sectorsize);
1623
1624                 ret = btrfs_cont_expand(BTRFS_I(inode), oldsize, end_pos);
1625                 if (ret) {
1626                         current->backing_dev_info = NULL;
1627                         return ret;
1628                 }
1629         }
1630
1631         return 0;
1632 }
1633
1634 static noinline ssize_t btrfs_buffered_write(struct kiocb *iocb,
1635                                                struct iov_iter *i)
1636 {
1637         struct file *file = iocb->ki_filp;
1638         loff_t pos;
1639         struct inode *inode = file_inode(file);
1640         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1641         struct page **pages = NULL;
1642         struct extent_changeset *data_reserved = NULL;
1643         u64 release_bytes = 0;
1644         u64 lockstart;
1645         u64 lockend;
1646         size_t num_written = 0;
1647         int nrptrs;
1648         ssize_t ret;
1649         bool only_release_metadata = false;
1650         bool force_page_uptodate = false;
1651         loff_t old_isize = i_size_read(inode);
1652         unsigned int ilock_flags = 0;
1653
1654         if (iocb->ki_flags & IOCB_NOWAIT)
1655                 ilock_flags |= BTRFS_ILOCK_TRY;
1656
1657         ret = btrfs_inode_lock(inode, ilock_flags);
1658         if (ret < 0)
1659                 return ret;
1660
1661         ret = generic_write_checks(iocb, i);
1662         if (ret <= 0)
1663                 goto out;
1664
1665         ret = btrfs_write_check(iocb, i, ret);
1666         if (ret < 0)
1667                 goto out;
1668
1669         pos = iocb->ki_pos;
1670         nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1671                         PAGE_SIZE / (sizeof(struct page *)));
1672         nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1673         nrptrs = max(nrptrs, 8);
1674         pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1675         if (!pages) {
1676                 ret = -ENOMEM;
1677                 goto out;
1678         }
1679
1680         while (iov_iter_count(i) > 0) {
1681                 struct extent_state *cached_state = NULL;
1682                 size_t offset = offset_in_page(pos);
1683                 size_t sector_offset;
1684                 size_t write_bytes = min(iov_iter_count(i),
1685                                          nrptrs * (size_t)PAGE_SIZE -
1686                                          offset);
1687                 size_t num_pages;
1688                 size_t reserve_bytes;
1689                 size_t dirty_pages;
1690                 size_t copied;
1691                 size_t dirty_sectors;
1692                 size_t num_sectors;
1693                 int extents_locked;
1694
1695                 /*
1696                  * Fault pages before locking them in prepare_pages
1697                  * to avoid recursive lock
1698                  */
1699                 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1700                         ret = -EFAULT;
1701                         break;
1702                 }
1703
1704                 only_release_metadata = false;
1705                 sector_offset = pos & (fs_info->sectorsize - 1);
1706
1707                 extent_changeset_release(data_reserved);
1708                 ret = btrfs_check_data_free_space(BTRFS_I(inode),
1709                                                   &data_reserved, pos,
1710                                                   write_bytes);
1711                 if (ret < 0) {
1712                         /*
1713                          * If we don't have to COW at the offset, reserve
1714                          * metadata only. write_bytes may get smaller than
1715                          * requested here.
1716                          */
1717                         if (btrfs_check_nocow_lock(BTRFS_I(inode), pos,
1718                                                    &write_bytes) > 0)
1719                                 only_release_metadata = true;
1720                         else
1721                                 break;
1722                 }
1723
1724                 num_pages = DIV_ROUND_UP(write_bytes + offset, PAGE_SIZE);
1725                 WARN_ON(num_pages > nrptrs);
1726                 reserve_bytes = round_up(write_bytes + sector_offset,
1727                                          fs_info->sectorsize);
1728                 WARN_ON(reserve_bytes == 0);
1729                 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1730                                 reserve_bytes);
1731                 if (ret) {
1732                         if (!only_release_metadata)
1733                                 btrfs_free_reserved_data_space(BTRFS_I(inode),
1734                                                 data_reserved, pos,
1735                                                 write_bytes);
1736                         else
1737                                 btrfs_check_nocow_unlock(BTRFS_I(inode));
1738                         break;
1739                 }
1740
1741                 release_bytes = reserve_bytes;
1742 again:
1743                 /*
1744                  * This is going to setup the pages array with the number of
1745                  * pages we want, so we don't really need to worry about the
1746                  * contents of pages from loop to loop
1747                  */
1748                 ret = prepare_pages(inode, pages, num_pages,
1749                                     pos, write_bytes,
1750                                     force_page_uptodate);
1751                 if (ret) {
1752                         btrfs_delalloc_release_extents(BTRFS_I(inode),
1753                                                        reserve_bytes);
1754                         break;
1755                 }
1756
1757                 extents_locked = lock_and_cleanup_extent_if_need(
1758                                 BTRFS_I(inode), pages,
1759                                 num_pages, pos, write_bytes, &lockstart,
1760                                 &lockend, &cached_state);
1761                 if (extents_locked < 0) {
1762                         if (extents_locked == -EAGAIN)
1763                                 goto again;
1764                         btrfs_delalloc_release_extents(BTRFS_I(inode),
1765                                                        reserve_bytes);
1766                         ret = extents_locked;
1767                         break;
1768                 }
1769
1770                 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1771
1772                 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1773                 dirty_sectors = round_up(copied + sector_offset,
1774                                         fs_info->sectorsize);
1775                 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1776
1777                 /*
1778                  * if we have trouble faulting in the pages, fall
1779                  * back to one page at a time
1780                  */
1781                 if (copied < write_bytes)
1782                         nrptrs = 1;
1783
1784                 if (copied == 0) {
1785                         force_page_uptodate = true;
1786                         dirty_sectors = 0;
1787                         dirty_pages = 0;
1788                 } else {
1789                         force_page_uptodate = false;
1790                         dirty_pages = DIV_ROUND_UP(copied + offset,
1791                                                    PAGE_SIZE);
1792                 }
1793
1794                 if (num_sectors > dirty_sectors) {
1795                         /* release everything except the sectors we dirtied */
1796                         release_bytes -= dirty_sectors << fs_info->sectorsize_bits;
1797                         if (only_release_metadata) {
1798                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1799                                                         release_bytes, true);
1800                         } else {
1801                                 u64 __pos;
1802
1803                                 __pos = round_down(pos,
1804                                                    fs_info->sectorsize) +
1805                                         (dirty_pages << PAGE_SHIFT);
1806                                 btrfs_delalloc_release_space(BTRFS_I(inode),
1807                                                 data_reserved, __pos,
1808                                                 release_bytes, true);
1809                         }
1810                 }
1811
1812                 release_bytes = round_up(copied + sector_offset,
1813                                         fs_info->sectorsize);
1814
1815                 ret = btrfs_dirty_pages(BTRFS_I(inode), pages,
1816                                         dirty_pages, pos, copied,
1817                                         &cached_state, only_release_metadata);
1818
1819                 /*
1820                  * If we have not locked the extent range, because the range's
1821                  * start offset is >= i_size, we might still have a non-NULL
1822                  * cached extent state, acquired while marking the extent range
1823                  * as delalloc through btrfs_dirty_pages(). Therefore free any
1824                  * possible cached extent state to avoid a memory leak.
1825                  */
1826                 if (extents_locked)
1827                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1828                                              lockstart, lockend, &cached_state);
1829                 else
1830                         free_extent_state(cached_state);
1831
1832                 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1833                 if (ret) {
1834                         btrfs_drop_pages(pages, num_pages);
1835                         break;
1836                 }
1837
1838                 release_bytes = 0;
1839                 if (only_release_metadata)
1840                         btrfs_check_nocow_unlock(BTRFS_I(inode));
1841
1842                 btrfs_drop_pages(pages, num_pages);
1843
1844                 cond_resched();
1845
1846                 balance_dirty_pages_ratelimited(inode->i_mapping);
1847
1848                 pos += copied;
1849                 num_written += copied;
1850         }
1851
1852         kfree(pages);
1853
1854         if (release_bytes) {
1855                 if (only_release_metadata) {
1856                         btrfs_check_nocow_unlock(BTRFS_I(inode));
1857                         btrfs_delalloc_release_metadata(BTRFS_I(inode),
1858                                         release_bytes, true);
1859                 } else {
1860                         btrfs_delalloc_release_space(BTRFS_I(inode),
1861                                         data_reserved,
1862                                         round_down(pos, fs_info->sectorsize),
1863                                         release_bytes, true);
1864                 }
1865         }
1866
1867         extent_changeset_free(data_reserved);
1868         if (num_written > 0) {
1869                 pagecache_isize_extended(inode, old_isize, iocb->ki_pos);
1870                 iocb->ki_pos += num_written;
1871         }
1872 out:
1873         btrfs_inode_unlock(inode, ilock_flags);
1874         return num_written ? num_written : ret;
1875 }
1876
1877 static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
1878                                const struct iov_iter *iter, loff_t offset)
1879 {
1880         const u32 blocksize_mask = fs_info->sectorsize - 1;
1881
1882         if (offset & blocksize_mask)
1883                 return -EINVAL;
1884
1885         if (iov_iter_alignment(iter) & blocksize_mask)
1886                 return -EINVAL;
1887
1888         return 0;
1889 }
1890
1891 static ssize_t btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1892 {
1893         struct file *file = iocb->ki_filp;
1894         struct inode *inode = file_inode(file);
1895         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1896         loff_t pos;
1897         ssize_t written = 0;
1898         ssize_t written_buffered;
1899         loff_t endbyte;
1900         ssize_t err;
1901         unsigned int ilock_flags = 0;
1902         struct iomap_dio *dio = NULL;
1903
1904         if (iocb->ki_flags & IOCB_NOWAIT)
1905                 ilock_flags |= BTRFS_ILOCK_TRY;
1906
1907         /* If the write DIO is within EOF, use a shared lock */
1908         if (iocb->ki_pos + iov_iter_count(from) <= i_size_read(inode))
1909                 ilock_flags |= BTRFS_ILOCK_SHARED;
1910
1911 relock:
1912         err = btrfs_inode_lock(inode, ilock_flags);
1913         if (err < 0)
1914                 return err;
1915
1916         err = generic_write_checks(iocb, from);
1917         if (err <= 0) {
1918                 btrfs_inode_unlock(inode, ilock_flags);
1919                 return err;
1920         }
1921
1922         err = btrfs_write_check(iocb, from, err);
1923         if (err < 0) {
1924                 btrfs_inode_unlock(inode, ilock_flags);
1925                 goto out;
1926         }
1927
1928         pos = iocb->ki_pos;
1929         /*
1930          * Re-check since file size may have changed just before taking the
1931          * lock or pos may have changed because of O_APPEND in generic_write_check()
1932          */
1933         if ((ilock_flags & BTRFS_ILOCK_SHARED) &&
1934             pos + iov_iter_count(from) > i_size_read(inode)) {
1935                 btrfs_inode_unlock(inode, ilock_flags);
1936                 ilock_flags &= ~BTRFS_ILOCK_SHARED;
1937                 goto relock;
1938         }
1939
1940         if (check_direct_IO(fs_info, from, pos)) {
1941                 btrfs_inode_unlock(inode, ilock_flags);
1942                 goto buffered;
1943         }
1944
1945         dio = __iomap_dio_rw(iocb, from, &btrfs_dio_iomap_ops,
1946                              &btrfs_dio_ops, is_sync_kiocb(iocb));
1947
1948         btrfs_inode_unlock(inode, ilock_flags);
1949
1950         if (IS_ERR_OR_NULL(dio)) {
1951                 err = PTR_ERR_OR_ZERO(dio);
1952                 if (err < 0 && err != -ENOTBLK)
1953                         goto out;
1954         } else {
1955                 written = iomap_dio_complete(dio);
1956         }
1957
1958         if (written < 0 || !iov_iter_count(from)) {
1959                 err = written;
1960                 goto out;
1961         }
1962
1963 buffered:
1964         pos = iocb->ki_pos;
1965         written_buffered = btrfs_buffered_write(iocb, from);
1966         if (written_buffered < 0) {
1967                 err = written_buffered;
1968                 goto out;
1969         }
1970         /*
1971          * Ensure all data is persisted. We want the next direct IO read to be
1972          * able to read what was just written.
1973          */
1974         endbyte = pos + written_buffered - 1;
1975         err = btrfs_fdatawrite_range(inode, pos, endbyte);
1976         if (err)
1977                 goto out;
1978         err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1979         if (err)
1980                 goto out;
1981         written += written_buffered;
1982         iocb->ki_pos = pos + written_buffered;
1983         invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1984                                  endbyte >> PAGE_SHIFT);
1985 out:
1986         return written ? written : err;
1987 }
1988
1989 static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
1990                                     struct iov_iter *from)
1991 {
1992         struct file *file = iocb->ki_filp;
1993         struct btrfs_inode *inode = BTRFS_I(file_inode(file));
1994         ssize_t num_written = 0;
1995         const bool sync = iocb->ki_flags & IOCB_DSYNC;
1996
1997         /*
1998          * If the fs flips readonly due to some impossible error, although we
1999          * have opened a file as writable, we have to stop this write operation
2000          * to ensure consistency.
2001          */
2002         if (test_bit(BTRFS_FS_STATE_ERROR, &inode->root->fs_info->fs_state))
2003                 return -EROFS;
2004
2005         if (!(iocb->ki_flags & IOCB_DIRECT) &&
2006             (iocb->ki_flags & IOCB_NOWAIT))
2007                 return -EOPNOTSUPP;
2008
2009         if (sync)
2010                 atomic_inc(&inode->sync_writers);
2011
2012         if (iocb->ki_flags & IOCB_DIRECT)
2013                 num_written = btrfs_direct_write(iocb, from);
2014         else
2015                 num_written = btrfs_buffered_write(iocb, from);
2016
2017         /*
2018          * We also have to set last_sub_trans to the current log transid,
2019          * otherwise subsequent syncs to a file that's been synced in this
2020          * transaction will appear to have already occurred.
2021          */
2022         spin_lock(&inode->lock);
2023         inode->last_sub_trans = inode->root->log_transid;
2024         spin_unlock(&inode->lock);
2025         if (num_written > 0)
2026                 num_written = generic_write_sync(iocb, num_written);
2027
2028         if (sync)
2029                 atomic_dec(&inode->sync_writers);
2030
2031         current->backing_dev_info = NULL;
2032         return num_written;
2033 }
2034
2035 int btrfs_release_file(struct inode *inode, struct file *filp)
2036 {
2037         struct btrfs_file_private *private = filp->private_data;
2038
2039         if (private && private->filldir_buf)
2040                 kfree(private->filldir_buf);
2041         kfree(private);
2042         filp->private_data = NULL;
2043
2044         /*
2045          * Set by setattr when we are about to truncate a file from a non-zero
2046          * size to a zero size.  This tries to flush down new bytes that may
2047          * have been written if the application were using truncate to replace
2048          * a file in place.
2049          */
2050         if (test_and_clear_bit(BTRFS_INODE_FLUSH_ON_CLOSE,
2051                                &BTRFS_I(inode)->runtime_flags))
2052                         filemap_flush(inode->i_mapping);
2053         return 0;
2054 }
2055
2056 static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
2057 {
2058         int ret;
2059         struct blk_plug plug;
2060
2061         /*
2062          * This is only called in fsync, which would do synchronous writes, so
2063          * a plug can merge adjacent IOs as much as possible.  Esp. in case of
2064          * multiple disks using raid profile, a large IO can be split to
2065          * several segments of stripe length (currently 64K).
2066          */
2067         blk_start_plug(&plug);
2068         atomic_inc(&BTRFS_I(inode)->sync_writers);
2069         ret = btrfs_fdatawrite_range(inode, start, end);
2070         atomic_dec(&BTRFS_I(inode)->sync_writers);
2071         blk_finish_plug(&plug);
2072
2073         return ret;
2074 }
2075
2076 /*
2077  * fsync call for both files and directories.  This logs the inode into
2078  * the tree log instead of forcing full commits whenever possible.
2079  *
2080  * It needs to call filemap_fdatawait so that all ordered extent updates are
2081  * in the metadata btree are up to date for copying to the log.
2082  *
2083  * It drops the inode mutex before doing the tree log commit.  This is an
2084  * important optimization for directories because holding the mutex prevents
2085  * new operations on the dir while we write to disk.
2086  */
2087 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
2088 {
2089         struct dentry *dentry = file_dentry(file);
2090         struct inode *inode = d_inode(dentry);
2091         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2092         struct btrfs_root *root = BTRFS_I(inode)->root;
2093         struct btrfs_trans_handle *trans;
2094         struct btrfs_log_ctx ctx;
2095         int ret = 0, err;
2096         u64 len;
2097         bool full_sync;
2098
2099         trace_btrfs_sync_file(file, datasync);
2100
2101         btrfs_init_log_ctx(&ctx, inode);
2102
2103         /*
2104          * Always set the range to a full range, otherwise we can get into
2105          * several problems, from missing file extent items to represent holes
2106          * when not using the NO_HOLES feature, to log tree corruption due to
2107          * races between hole detection during logging and completion of ordered
2108          * extents outside the range, to missing checksums due to ordered extents
2109          * for which we flushed only a subset of their pages.
2110          */
2111         start = 0;
2112         end = LLONG_MAX;
2113         len = (u64)LLONG_MAX + 1;
2114
2115         /*
2116          * We write the dirty pages in the range and wait until they complete
2117          * out of the ->i_mutex. If so, we can flush the dirty pages by
2118          * multi-task, and make the performance up.  See
2119          * btrfs_wait_ordered_range for an explanation of the ASYNC check.
2120          */
2121         ret = start_ordered_ops(inode, start, end);
2122         if (ret)
2123                 goto out;
2124
2125         inode_lock(inode);
2126
2127         atomic_inc(&root->log_batch);
2128
2129         /*
2130          * Always check for the full sync flag while holding the inode's lock,
2131          * to avoid races with other tasks. The flag must be either set all the
2132          * time during logging or always off all the time while logging.
2133          */
2134         full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2135                              &BTRFS_I(inode)->runtime_flags);
2136
2137         /*
2138          * Before we acquired the inode's lock, someone may have dirtied more
2139          * pages in the target range. We need to make sure that writeback for
2140          * any such pages does not start while we are logging the inode, because
2141          * if it does, any of the following might happen when we are not doing a
2142          * full inode sync:
2143          *
2144          * 1) We log an extent after its writeback finishes but before its
2145          *    checksums are added to the csum tree, leading to -EIO errors
2146          *    when attempting to read the extent after a log replay.
2147          *
2148          * 2) We can end up logging an extent before its writeback finishes.
2149          *    Therefore after the log replay we will have a file extent item
2150          *    pointing to an unwritten extent (and no data checksums as well).
2151          *
2152          * So trigger writeback for any eventual new dirty pages and then we
2153          * wait for all ordered extents to complete below.
2154          */
2155         ret = start_ordered_ops(inode, start, end);
2156         if (ret) {
2157                 inode_unlock(inode);
2158                 goto out;
2159         }
2160
2161         /*
2162          * We have to do this here to avoid the priority inversion of waiting on
2163          * IO of a lower priority task while holding a transaction open.
2164          *
2165          * For a full fsync we wait for the ordered extents to complete while
2166          * for a fast fsync we wait just for writeback to complete, and then
2167          * attach the ordered extents to the transaction so that a transaction
2168          * commit waits for their completion, to avoid data loss if we fsync,
2169          * the current transaction commits before the ordered extents complete
2170          * and a power failure happens right after that.
2171          */
2172         if (full_sync) {
2173                 ret = btrfs_wait_ordered_range(inode, start, len);
2174         } else {
2175                 /*
2176                  * Get our ordered extents as soon as possible to avoid doing
2177                  * checksum lookups in the csum tree, and use instead the
2178                  * checksums attached to the ordered extents.
2179                  */
2180                 btrfs_get_ordered_extents_for_logging(BTRFS_I(inode),
2181                                                       &ctx.ordered_extents);
2182                 ret = filemap_fdatawait_range(inode->i_mapping, start, end);
2183         }
2184
2185         if (ret)
2186                 goto out_release_extents;
2187
2188         atomic_inc(&root->log_batch);
2189
2190         /*
2191          * If we are doing a fast fsync we can not bail out if the inode's
2192          * last_trans is <= then the last committed transaction, because we only
2193          * update the last_trans of the inode during ordered extent completion,
2194          * and for a fast fsync we don't wait for that, we only wait for the
2195          * writeback to complete.
2196          */
2197         smp_mb();
2198         if (btrfs_inode_in_log(BTRFS_I(inode), fs_info->generation) ||
2199             (BTRFS_I(inode)->last_trans <= fs_info->last_trans_committed &&
2200              (full_sync || list_empty(&ctx.ordered_extents)))) {
2201                 /*
2202                  * We've had everything committed since the last time we were
2203                  * modified so clear this flag in case it was set for whatever
2204                  * reason, it's no longer relevant.
2205                  */
2206                 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2207                           &BTRFS_I(inode)->runtime_flags);
2208                 /*
2209                  * An ordered extent might have started before and completed
2210                  * already with io errors, in which case the inode was not
2211                  * updated and we end up here. So check the inode's mapping
2212                  * for any errors that might have happened since we last
2213                  * checked called fsync.
2214                  */
2215                 ret = filemap_check_wb_err(inode->i_mapping, file->f_wb_err);
2216                 goto out_release_extents;
2217         }
2218
2219         /*
2220          * We use start here because we will need to wait on the IO to complete
2221          * in btrfs_sync_log, which could require joining a transaction (for
2222          * example checking cross references in the nocow path).  If we use join
2223          * here we could get into a situation where we're waiting on IO to
2224          * happen that is blocked on a transaction trying to commit.  With start
2225          * we inc the extwriter counter, so we wait for all extwriters to exit
2226          * before we start blocking joiners.  This comment is to keep somebody
2227          * from thinking they are super smart and changing this to
2228          * btrfs_join_transaction *cough*Josef*cough*.
2229          */
2230         trans = btrfs_start_transaction(root, 0);
2231         if (IS_ERR(trans)) {
2232                 ret = PTR_ERR(trans);
2233                 goto out_release_extents;
2234         }
2235         trans->in_fsync = true;
2236
2237         ret = btrfs_log_dentry_safe(trans, dentry, &ctx);
2238         btrfs_release_log_ctx_extents(&ctx);
2239         if (ret < 0) {
2240                 /* Fallthrough and commit/free transaction. */
2241                 ret = 1;
2242         }
2243
2244         /* we've logged all the items and now have a consistent
2245          * version of the file in the log.  It is possible that
2246          * someone will come in and modify the file, but that's
2247          * fine because the log is consistent on disk, and we
2248          * have references to all of the file's extents
2249          *
2250          * It is possible that someone will come in and log the
2251          * file again, but that will end up using the synchronization
2252          * inside btrfs_sync_log to keep things safe.
2253          */
2254         inode_unlock(inode);
2255
2256         if (ret != BTRFS_NO_LOG_SYNC) {
2257                 if (!ret) {
2258                         ret = btrfs_sync_log(trans, root, &ctx);
2259                         if (!ret) {
2260                                 ret = btrfs_end_transaction(trans);
2261                                 goto out;
2262                         }
2263                 }
2264                 if (!full_sync) {
2265                         ret = btrfs_wait_ordered_range(inode, start, len);
2266                         if (ret) {
2267                                 btrfs_end_transaction(trans);
2268                                 goto out;
2269                         }
2270                 }
2271                 ret = btrfs_commit_transaction(trans);
2272         } else {
2273                 ret = btrfs_end_transaction(trans);
2274         }
2275 out:
2276         ASSERT(list_empty(&ctx.list));
2277         err = file_check_and_advance_wb_err(file);
2278         if (!ret)
2279                 ret = err;
2280         return ret > 0 ? -EIO : ret;
2281
2282 out_release_extents:
2283         btrfs_release_log_ctx_extents(&ctx);
2284         inode_unlock(inode);
2285         goto out;
2286 }
2287
2288 static const struct vm_operations_struct btrfs_file_vm_ops = {
2289         .fault          = filemap_fault,
2290         .map_pages      = filemap_map_pages,
2291         .page_mkwrite   = btrfs_page_mkwrite,
2292 };
2293
2294 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
2295 {
2296         struct address_space *mapping = filp->f_mapping;
2297
2298         if (!mapping->a_ops->readpage)
2299                 return -ENOEXEC;
2300
2301         file_accessed(filp);
2302         vma->vm_ops = &btrfs_file_vm_ops;
2303
2304         return 0;
2305 }
2306
2307 static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2308                           int slot, u64 start, u64 end)
2309 {
2310         struct btrfs_file_extent_item *fi;
2311         struct btrfs_key key;
2312
2313         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2314                 return 0;
2315
2316         btrfs_item_key_to_cpu(leaf, &key, slot);
2317         if (key.objectid != btrfs_ino(inode) ||
2318             key.type != BTRFS_EXTENT_DATA_KEY)
2319                 return 0;
2320
2321         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2322
2323         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2324                 return 0;
2325
2326         if (btrfs_file_extent_disk_bytenr(leaf, fi))
2327                 return 0;
2328
2329         if (key.offset == end)
2330                 return 1;
2331         if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2332                 return 1;
2333         return 0;
2334 }
2335
2336 static int fill_holes(struct btrfs_trans_handle *trans,
2337                 struct btrfs_inode *inode,
2338                 struct btrfs_path *path, u64 offset, u64 end)
2339 {
2340         struct btrfs_fs_info *fs_info = trans->fs_info;
2341         struct btrfs_root *root = inode->root;
2342         struct extent_buffer *leaf;
2343         struct btrfs_file_extent_item *fi;
2344         struct extent_map *hole_em;
2345         struct extent_map_tree *em_tree = &inode->extent_tree;
2346         struct btrfs_key key;
2347         int ret;
2348
2349         if (btrfs_fs_incompat(fs_info, NO_HOLES))
2350                 goto out;
2351
2352         key.objectid = btrfs_ino(inode);
2353         key.type = BTRFS_EXTENT_DATA_KEY;
2354         key.offset = offset;
2355
2356         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2357         if (ret <= 0) {
2358                 /*
2359                  * We should have dropped this offset, so if we find it then
2360                  * something has gone horribly wrong.
2361                  */
2362                 if (ret == 0)
2363                         ret = -EINVAL;
2364                 return ret;
2365         }
2366
2367         leaf = path->nodes[0];
2368         if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2369                 u64 num_bytes;
2370
2371                 path->slots[0]--;
2372                 fi = btrfs_item_ptr(leaf, path->slots[0],
2373                                     struct btrfs_file_extent_item);
2374                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2375                         end - offset;
2376                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2377                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2378                 btrfs_set_file_extent_offset(leaf, fi, 0);
2379                 btrfs_mark_buffer_dirty(leaf);
2380                 goto out;
2381         }
2382
2383         if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2384                 u64 num_bytes;
2385
2386                 key.offset = offset;
2387                 btrfs_set_item_key_safe(fs_info, path, &key);
2388                 fi = btrfs_item_ptr(leaf, path->slots[0],
2389                                     struct btrfs_file_extent_item);
2390                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2391                         offset;
2392                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2393                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2394                 btrfs_set_file_extent_offset(leaf, fi, 0);
2395                 btrfs_mark_buffer_dirty(leaf);
2396                 goto out;
2397         }
2398         btrfs_release_path(path);
2399
2400         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode),
2401                         offset, 0, 0, end - offset, 0, end - offset, 0, 0, 0);
2402         if (ret)
2403                 return ret;
2404
2405 out:
2406         btrfs_release_path(path);
2407
2408         hole_em = alloc_extent_map();
2409         if (!hole_em) {
2410                 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2411                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
2412         } else {
2413                 hole_em->start = offset;
2414                 hole_em->len = end - offset;
2415                 hole_em->ram_bytes = hole_em->len;
2416                 hole_em->orig_start = offset;
2417
2418                 hole_em->block_start = EXTENT_MAP_HOLE;
2419                 hole_em->block_len = 0;
2420                 hole_em->orig_block_len = 0;
2421                 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2422                 hole_em->generation = trans->transid;
2423
2424                 do {
2425                         btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2426                         write_lock(&em_tree->lock);
2427                         ret = add_extent_mapping(em_tree, hole_em, 1);
2428                         write_unlock(&em_tree->lock);
2429                 } while (ret == -EEXIST);
2430                 free_extent_map(hole_em);
2431                 if (ret)
2432                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2433                                         &inode->runtime_flags);
2434         }
2435
2436         return 0;
2437 }
2438
2439 /*
2440  * Find a hole extent on given inode and change start/len to the end of hole
2441  * extent.(hole/vacuum extent whose em->start <= start &&
2442  *         em->start + em->len > start)
2443  * When a hole extent is found, return 1 and modify start/len.
2444  */
2445 static int find_first_non_hole(struct btrfs_inode *inode, u64 *start, u64 *len)
2446 {
2447         struct btrfs_fs_info *fs_info = inode->root->fs_info;
2448         struct extent_map *em;
2449         int ret = 0;
2450
2451         em = btrfs_get_extent(inode, NULL, 0,
2452                               round_down(*start, fs_info->sectorsize),
2453                               round_up(*len, fs_info->sectorsize));
2454         if (IS_ERR(em))
2455                 return PTR_ERR(em);
2456
2457         /* Hole or vacuum extent(only exists in no-hole mode) */
2458         if (em->block_start == EXTENT_MAP_HOLE) {
2459                 ret = 1;
2460                 *len = em->start + em->len > *start + *len ?
2461                        0 : *start + *len - em->start - em->len;
2462                 *start = em->start + em->len;
2463         }
2464         free_extent_map(em);
2465         return ret;
2466 }
2467
2468 static int btrfs_punch_hole_lock_range(struct inode *inode,
2469                                        const u64 lockstart,
2470                                        const u64 lockend,
2471                                        struct extent_state **cached_state)
2472 {
2473         while (1) {
2474                 struct btrfs_ordered_extent *ordered;
2475                 int ret;
2476
2477                 truncate_pagecache_range(inode, lockstart, lockend);
2478
2479                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2480                                  cached_state);
2481                 ordered = btrfs_lookup_first_ordered_extent(BTRFS_I(inode),
2482                                                             lockend);
2483
2484                 /*
2485                  * We need to make sure we have no ordered extents in this range
2486                  * and nobody raced in and read a page in this range, if we did
2487                  * we need to try again.
2488                  */
2489                 if ((!ordered ||
2490                     (ordered->file_offset + ordered->num_bytes <= lockstart ||
2491                      ordered->file_offset > lockend)) &&
2492                      !filemap_range_has_page(inode->i_mapping,
2493                                              lockstart, lockend)) {
2494                         if (ordered)
2495                                 btrfs_put_ordered_extent(ordered);
2496                         break;
2497                 }
2498                 if (ordered)
2499                         btrfs_put_ordered_extent(ordered);
2500                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2501                                      lockend, cached_state);
2502                 ret = btrfs_wait_ordered_range(inode, lockstart,
2503                                                lockend - lockstart + 1);
2504                 if (ret)
2505                         return ret;
2506         }
2507         return 0;
2508 }
2509
2510 static int btrfs_insert_replace_extent(struct btrfs_trans_handle *trans,
2511                                      struct btrfs_inode *inode,
2512                                      struct btrfs_path *path,
2513                                      struct btrfs_replace_extent_info *extent_info,
2514                                      const u64 replace_len,
2515                                      const u64 bytes_to_drop)
2516 {
2517         struct btrfs_fs_info *fs_info = trans->fs_info;
2518         struct btrfs_root *root = inode->root;
2519         struct btrfs_file_extent_item *extent;
2520         struct extent_buffer *leaf;
2521         struct btrfs_key key;
2522         int slot;
2523         struct btrfs_ref ref = { 0 };
2524         int ret;
2525
2526         if (replace_len == 0)
2527                 return 0;
2528
2529         if (extent_info->disk_offset == 0 &&
2530             btrfs_fs_incompat(fs_info, NO_HOLES)) {
2531                 btrfs_update_inode_bytes(inode, 0, bytes_to_drop);
2532                 return 0;
2533         }
2534
2535         key.objectid = btrfs_ino(inode);
2536         key.type = BTRFS_EXTENT_DATA_KEY;
2537         key.offset = extent_info->file_offset;
2538         ret = btrfs_insert_empty_item(trans, root, path, &key,
2539                                       sizeof(struct btrfs_file_extent_item));
2540         if (ret)
2541                 return ret;
2542         leaf = path->nodes[0];
2543         slot = path->slots[0];
2544         write_extent_buffer(leaf, extent_info->extent_buf,
2545                             btrfs_item_ptr_offset(leaf, slot),
2546                             sizeof(struct btrfs_file_extent_item));
2547         extent = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2548         ASSERT(btrfs_file_extent_type(leaf, extent) != BTRFS_FILE_EXTENT_INLINE);
2549         btrfs_set_file_extent_offset(leaf, extent, extent_info->data_offset);
2550         btrfs_set_file_extent_num_bytes(leaf, extent, replace_len);
2551         if (extent_info->is_new_extent)
2552                 btrfs_set_file_extent_generation(leaf, extent, trans->transid);
2553         btrfs_mark_buffer_dirty(leaf);
2554         btrfs_release_path(path);
2555
2556         ret = btrfs_inode_set_file_extent_range(inode, extent_info->file_offset,
2557                                                 replace_len);
2558         if (ret)
2559                 return ret;
2560
2561         /* If it's a hole, nothing more needs to be done. */
2562         if (extent_info->disk_offset == 0) {
2563                 btrfs_update_inode_bytes(inode, 0, bytes_to_drop);
2564                 return 0;
2565         }
2566
2567         btrfs_update_inode_bytes(inode, replace_len, bytes_to_drop);
2568
2569         if (extent_info->is_new_extent && extent_info->insertions == 0) {
2570                 key.objectid = extent_info->disk_offset;
2571                 key.type = BTRFS_EXTENT_ITEM_KEY;
2572                 key.offset = extent_info->disk_len;
2573                 ret = btrfs_alloc_reserved_file_extent(trans, root,
2574                                                        btrfs_ino(inode),
2575                                                        extent_info->file_offset,
2576                                                        extent_info->qgroup_reserved,
2577                                                        &key);
2578         } else {
2579                 u64 ref_offset;
2580
2581                 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2582                                        extent_info->disk_offset,
2583                                        extent_info->disk_len, 0);
2584                 ref_offset = extent_info->file_offset - extent_info->data_offset;
2585                 btrfs_init_data_ref(&ref, root->root_key.objectid,
2586                                     btrfs_ino(inode), ref_offset);
2587                 ret = btrfs_inc_extent_ref(trans, &ref);
2588         }
2589
2590         extent_info->insertions++;
2591
2592         return ret;
2593 }
2594
2595 /*
2596  * The respective range must have been previously locked, as well as the inode.
2597  * The end offset is inclusive (last byte of the range).
2598  * @extent_info is NULL for fallocate's hole punching and non-NULL when replacing
2599  * the file range with an extent.
2600  * When not punching a hole, we don't want to end up in a state where we dropped
2601  * extents without inserting a new one, so we must abort the transaction to avoid
2602  * a corruption.
2603  */
2604 int btrfs_replace_file_extents(struct inode *inode, struct btrfs_path *path,
2605                            const u64 start, const u64 end,
2606                            struct btrfs_replace_extent_info *extent_info,
2607                            struct btrfs_trans_handle **trans_out)
2608 {
2609         struct btrfs_drop_extents_args drop_args = { 0 };
2610         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2611         u64 min_size = btrfs_calc_insert_metadata_size(fs_info, 1);
2612         u64 ino_size = round_up(inode->i_size, fs_info->sectorsize);
2613         struct btrfs_root *root = BTRFS_I(inode)->root;
2614         struct btrfs_trans_handle *trans = NULL;
2615         struct btrfs_block_rsv *rsv;
2616         unsigned int rsv_count;
2617         u64 cur_offset;
2618         u64 len = end - start;
2619         int ret = 0;
2620
2621         if (end <= start)
2622                 return -EINVAL;
2623
2624         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2625         if (!rsv) {
2626                 ret = -ENOMEM;
2627                 goto out;
2628         }
2629         rsv->size = btrfs_calc_insert_metadata_size(fs_info, 1);
2630         rsv->failfast = 1;
2631
2632         /*
2633          * 1 - update the inode
2634          * 1 - removing the extents in the range
2635          * 1 - adding the hole extent if no_holes isn't set or if we are
2636          *     replacing the range with a new extent
2637          */
2638         if (!btrfs_fs_incompat(fs_info, NO_HOLES) || extent_info)
2639                 rsv_count = 3;
2640         else
2641                 rsv_count = 2;
2642
2643         trans = btrfs_start_transaction(root, rsv_count);
2644         if (IS_ERR(trans)) {
2645                 ret = PTR_ERR(trans);
2646                 trans = NULL;
2647                 goto out_free;
2648         }
2649
2650         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2651                                       min_size, false);
2652         BUG_ON(ret);
2653         trans->block_rsv = rsv;
2654
2655         cur_offset = start;
2656         drop_args.path = path;
2657         drop_args.end = end + 1;
2658         drop_args.drop_cache = true;
2659         while (cur_offset < end) {
2660                 drop_args.start = cur_offset;
2661                 ret = btrfs_drop_extents(trans, root, BTRFS_I(inode), &drop_args);
2662                 /* If we are punching a hole decrement the inode's byte count */
2663                 if (!extent_info)
2664                         btrfs_update_inode_bytes(BTRFS_I(inode), 0,
2665                                                  drop_args.bytes_found);
2666                 if (ret != -ENOSPC) {
2667                         /*
2668                          * When cloning we want to avoid transaction aborts when
2669                          * nothing was done and we are attempting to clone parts
2670                          * of inline extents, in such cases -EOPNOTSUPP is
2671                          * returned by __btrfs_drop_extents() without having
2672                          * changed anything in the file.
2673                          */
2674                         if (extent_info && !extent_info->is_new_extent &&
2675                             ret && ret != -EOPNOTSUPP)
2676                                 btrfs_abort_transaction(trans, ret);
2677                         break;
2678                 }
2679
2680                 trans->block_rsv = &fs_info->trans_block_rsv;
2681
2682                 if (!extent_info && cur_offset < drop_args.drop_end &&
2683                     cur_offset < ino_size) {
2684                         ret = fill_holes(trans, BTRFS_I(inode), path,
2685                                          cur_offset, drop_args.drop_end);
2686                         if (ret) {
2687                                 /*
2688                                  * If we failed then we didn't insert our hole
2689                                  * entries for the area we dropped, so now the
2690                                  * fs is corrupted, so we must abort the
2691                                  * transaction.
2692                                  */
2693                                 btrfs_abort_transaction(trans, ret);
2694                                 break;
2695                         }
2696                 } else if (!extent_info && cur_offset < drop_args.drop_end) {
2697                         /*
2698                          * We are past the i_size here, but since we didn't
2699                          * insert holes we need to clear the mapped area so we
2700                          * know to not set disk_i_size in this area until a new
2701                          * file extent is inserted here.
2702                          */
2703                         ret = btrfs_inode_clear_file_extent_range(BTRFS_I(inode),
2704                                         cur_offset,
2705                                         drop_args.drop_end - cur_offset);
2706                         if (ret) {
2707                                 /*
2708                                  * We couldn't clear our area, so we could
2709                                  * presumably adjust up and corrupt the fs, so
2710                                  * we need to abort.
2711                                  */
2712                                 btrfs_abort_transaction(trans, ret);
2713                                 break;
2714                         }
2715                 }
2716
2717                 if (extent_info &&
2718                     drop_args.drop_end > extent_info->file_offset) {
2719                         u64 replace_len = drop_args.drop_end -
2720                                           extent_info->file_offset;
2721
2722                         ret = btrfs_insert_replace_extent(trans, BTRFS_I(inode),
2723                                         path, extent_info, replace_len,
2724                                         drop_args.bytes_found);
2725                         if (ret) {
2726                                 btrfs_abort_transaction(trans, ret);
2727                                 break;
2728                         }
2729                         extent_info->data_len -= replace_len;
2730                         extent_info->data_offset += replace_len;
2731                         extent_info->file_offset += replace_len;
2732                 }
2733
2734                 cur_offset = drop_args.drop_end;
2735
2736                 ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
2737                 if (ret)
2738                         break;
2739
2740                 btrfs_end_transaction(trans);
2741                 btrfs_btree_balance_dirty(fs_info);
2742
2743                 trans = btrfs_start_transaction(root, rsv_count);
2744                 if (IS_ERR(trans)) {
2745                         ret = PTR_ERR(trans);
2746                         trans = NULL;
2747                         break;
2748                 }
2749
2750                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2751                                               rsv, min_size, false);
2752                 BUG_ON(ret);    /* shouldn't happen */
2753                 trans->block_rsv = rsv;
2754
2755                 if (!extent_info) {
2756                         ret = find_first_non_hole(BTRFS_I(inode), &cur_offset,
2757                                                   &len);
2758                         if (unlikely(ret < 0))
2759                                 break;
2760                         if (ret && !len) {
2761                                 ret = 0;
2762                                 break;
2763                         }
2764                 }
2765         }
2766
2767         /*
2768          * If we were cloning, force the next fsync to be a full one since we
2769          * we replaced (or just dropped in the case of cloning holes when
2770          * NO_HOLES is enabled) extents and extent maps.
2771          * This is for the sake of simplicity, and cloning into files larger
2772          * than 16Mb would force the full fsync any way (when
2773          * try_release_extent_mapping() is invoked during page cache truncation.
2774          */
2775         if (extent_info && !extent_info->is_new_extent)
2776                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2777                         &BTRFS_I(inode)->runtime_flags);
2778
2779         if (ret)
2780                 goto out_trans;
2781
2782         trans->block_rsv = &fs_info->trans_block_rsv;
2783         /*
2784          * If we are using the NO_HOLES feature we might have had already an
2785          * hole that overlaps a part of the region [lockstart, lockend] and
2786          * ends at (or beyond) lockend. Since we have no file extent items to
2787          * represent holes, drop_end can be less than lockend and so we must
2788          * make sure we have an extent map representing the existing hole (the
2789          * call to __btrfs_drop_extents() might have dropped the existing extent
2790          * map representing the existing hole), otherwise the fast fsync path
2791          * will not record the existence of the hole region
2792          * [existing_hole_start, lockend].
2793          */
2794         if (drop_args.drop_end <= end)
2795                 drop_args.drop_end = end + 1;
2796         /*
2797          * Don't insert file hole extent item if it's for a range beyond eof
2798          * (because it's useless) or if it represents a 0 bytes range (when
2799          * cur_offset == drop_end).
2800          */
2801         if (!extent_info && cur_offset < ino_size &&
2802             cur_offset < drop_args.drop_end) {
2803                 ret = fill_holes(trans, BTRFS_I(inode), path,
2804                                  cur_offset, drop_args.drop_end);
2805                 if (ret) {
2806                         /* Same comment as above. */
2807                         btrfs_abort_transaction(trans, ret);
2808                         goto out_trans;
2809                 }
2810         } else if (!extent_info && cur_offset < drop_args.drop_end) {
2811                 /* See the comment in the loop above for the reasoning here. */
2812                 ret = btrfs_inode_clear_file_extent_range(BTRFS_I(inode),
2813                                 cur_offset, drop_args.drop_end - cur_offset);
2814                 if (ret) {
2815                         btrfs_abort_transaction(trans, ret);
2816                         goto out_trans;
2817                 }
2818
2819         }
2820         if (extent_info) {
2821                 ret = btrfs_insert_replace_extent(trans, BTRFS_I(inode), path,
2822                                 extent_info, extent_info->data_len,
2823                                 drop_args.bytes_found);
2824                 if (ret) {
2825                         btrfs_abort_transaction(trans, ret);
2826                         goto out_trans;
2827                 }
2828         }
2829
2830 out_trans:
2831         if (!trans)
2832                 goto out_free;
2833
2834         trans->block_rsv = &fs_info->trans_block_rsv;
2835         if (ret)
2836                 btrfs_end_transaction(trans);
2837         else
2838                 *trans_out = trans;
2839 out_free:
2840         btrfs_free_block_rsv(fs_info, rsv);
2841 out:
2842         return ret;
2843 }
2844
2845 static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2846 {
2847         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2848         struct btrfs_root *root = BTRFS_I(inode)->root;
2849         struct extent_state *cached_state = NULL;
2850         struct btrfs_path *path;
2851         struct btrfs_trans_handle *trans = NULL;
2852         u64 lockstart;
2853         u64 lockend;
2854         u64 tail_start;
2855         u64 tail_len;
2856         u64 orig_start = offset;
2857         int ret = 0;
2858         bool same_block;
2859         u64 ino_size;
2860         bool truncated_block = false;
2861         bool updated_inode = false;
2862
2863         ret = btrfs_wait_ordered_range(inode, offset, len);
2864         if (ret)
2865                 return ret;
2866
2867         inode_lock(inode);
2868         ino_size = round_up(inode->i_size, fs_info->sectorsize);
2869         ret = find_first_non_hole(BTRFS_I(inode), &offset, &len);
2870         if (ret < 0)
2871                 goto out_only_mutex;
2872         if (ret && !len) {
2873                 /* Already in a large hole */
2874                 ret = 0;
2875                 goto out_only_mutex;
2876         }
2877
2878         lockstart = round_up(offset, btrfs_inode_sectorsize(BTRFS_I(inode)));
2879         lockend = round_down(offset + len,
2880                              btrfs_inode_sectorsize(BTRFS_I(inode))) - 1;
2881         same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2882                 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2883         /*
2884          * We needn't truncate any block which is beyond the end of the file
2885          * because we are sure there is no data there.
2886          */
2887         /*
2888          * Only do this if we are in the same block and we aren't doing the
2889          * entire block.
2890          */
2891         if (same_block && len < fs_info->sectorsize) {
2892                 if (offset < ino_size) {
2893                         truncated_block = true;
2894                         ret = btrfs_truncate_block(BTRFS_I(inode), offset, len,
2895                                                    0);
2896                 } else {
2897                         ret = 0;
2898                 }
2899                 goto out_only_mutex;
2900         }
2901
2902         /* zero back part of the first block */
2903         if (offset < ino_size) {
2904                 truncated_block = true;
2905                 ret = btrfs_truncate_block(BTRFS_I(inode), offset, 0, 0);
2906                 if (ret) {
2907                         inode_unlock(inode);
2908                         return ret;
2909                 }
2910         }
2911
2912         /* Check the aligned pages after the first unaligned page,
2913          * if offset != orig_start, which means the first unaligned page
2914          * including several following pages are already in holes,
2915          * the extra check can be skipped */
2916         if (offset == orig_start) {
2917                 /* after truncate page, check hole again */
2918                 len = offset + len - lockstart;
2919                 offset = lockstart;
2920                 ret = find_first_non_hole(BTRFS_I(inode), &offset, &len);
2921                 if (ret < 0)
2922                         goto out_only_mutex;
2923                 if (ret && !len) {
2924                         ret = 0;
2925                         goto out_only_mutex;
2926                 }
2927                 lockstart = offset;
2928         }
2929
2930         /* Check the tail unaligned part is in a hole */
2931         tail_start = lockend + 1;
2932         tail_len = offset + len - tail_start;
2933         if (tail_len) {
2934                 ret = find_first_non_hole(BTRFS_I(inode), &tail_start, &tail_len);
2935                 if (unlikely(ret < 0))
2936                         goto out_only_mutex;
2937                 if (!ret) {
2938                         /* zero the front end of the last page */
2939                         if (tail_start + tail_len < ino_size) {
2940                                 truncated_block = true;
2941                                 ret = btrfs_truncate_block(BTRFS_I(inode),
2942                                                         tail_start + tail_len,
2943                                                         0, 1);
2944                                 if (ret)
2945                                         goto out_only_mutex;
2946                         }
2947                 }
2948         }
2949
2950         if (lockend < lockstart) {
2951                 ret = 0;
2952                 goto out_only_mutex;
2953         }
2954
2955         ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
2956                                           &cached_state);
2957         if (ret)
2958                 goto out_only_mutex;
2959
2960         path = btrfs_alloc_path();
2961         if (!path) {
2962                 ret = -ENOMEM;
2963                 goto out;
2964         }
2965
2966         ret = btrfs_replace_file_extents(inode, path, lockstart, lockend, NULL,
2967                                      &trans);
2968         btrfs_free_path(path);
2969         if (ret)
2970                 goto out;
2971
2972         ASSERT(trans != NULL);
2973         inode_inc_iversion(inode);
2974         inode->i_mtime = inode->i_ctime = current_time(inode);
2975         ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
2976         updated_inode = true;
2977         btrfs_end_transaction(trans);
2978         btrfs_btree_balance_dirty(fs_info);
2979 out:
2980         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2981                              &cached_state);
2982 out_only_mutex:
2983         if (!updated_inode && truncated_block && !ret) {
2984                 /*
2985                  * If we only end up zeroing part of a page, we still need to
2986                  * update the inode item, so that all the time fields are
2987                  * updated as well as the necessary btrfs inode in memory fields
2988                  * for detecting, at fsync time, if the inode isn't yet in the
2989                  * log tree or it's there but not up to date.
2990                  */
2991                 struct timespec64 now = current_time(inode);
2992
2993                 inode_inc_iversion(inode);
2994                 inode->i_mtime = now;
2995                 inode->i_ctime = now;
2996                 trans = btrfs_start_transaction(root, 1);
2997                 if (IS_ERR(trans)) {
2998                         ret = PTR_ERR(trans);
2999                 } else {
3000                         int ret2;
3001
3002                         ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
3003                         ret2 = btrfs_end_transaction(trans);
3004                         if (!ret)
3005                                 ret = ret2;
3006                 }
3007         }
3008         inode_unlock(inode);
3009         return ret;
3010 }
3011
3012 /* Helper structure to record which range is already reserved */
3013 struct falloc_range {
3014         struct list_head list;
3015         u64 start;
3016         u64 len;
3017 };
3018
3019 /*
3020  * Helper function to add falloc range
3021  *
3022  * Caller should have locked the larger range of extent containing
3023  * [start, len)
3024  */
3025 static int add_falloc_range(struct list_head *head, u64 start, u64 len)
3026 {
3027         struct falloc_range *prev = NULL;
3028         struct falloc_range *range = NULL;
3029
3030         if (list_empty(head))
3031                 goto insert;
3032
3033         /*
3034          * As fallocate iterate by bytenr order, we only need to check
3035          * the last range.
3036          */
3037         prev = list_entry(head->prev, struct falloc_range, list);
3038         if (prev->start + prev->len == start) {
3039                 prev->len += len;
3040                 return 0;
3041         }
3042 insert:
3043         range = kmalloc(sizeof(*range), GFP_KERNEL);
3044         if (!range)
3045                 return -ENOMEM;
3046         range->start = start;
3047         range->len = len;
3048         list_add_tail(&range->list, head);
3049         return 0;
3050 }
3051
3052 static int btrfs_fallocate_update_isize(struct inode *inode,
3053                                         const u64 end,
3054                                         const int mode)
3055 {
3056         struct btrfs_trans_handle *trans;
3057         struct btrfs_root *root = BTRFS_I(inode)->root;
3058         int ret;
3059         int ret2;
3060
3061         if (mode & FALLOC_FL_KEEP_SIZE || end <= i_size_read(inode))
3062                 return 0;
3063
3064         trans = btrfs_start_transaction(root, 1);
3065         if (IS_ERR(trans))
3066                 return PTR_ERR(trans);
3067
3068         inode->i_ctime = current_time(inode);
3069         i_size_write(inode, end);
3070         btrfs_inode_safe_disk_i_size_write(BTRFS_I(inode), 0);
3071         ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
3072         ret2 = btrfs_end_transaction(trans);
3073
3074         return ret ? ret : ret2;
3075 }
3076
3077 enum {
3078         RANGE_BOUNDARY_WRITTEN_EXTENT,
3079         RANGE_BOUNDARY_PREALLOC_EXTENT,
3080         RANGE_BOUNDARY_HOLE,
3081 };
3082
3083 static int btrfs_zero_range_check_range_boundary(struct btrfs_inode *inode,
3084                                                  u64 offset)
3085 {
3086         const u64 sectorsize = btrfs_inode_sectorsize(inode);
3087         struct extent_map *em;
3088         int ret;
3089
3090         offset = round_down(offset, sectorsize);
3091         em = btrfs_get_extent(inode, NULL, 0, offset, sectorsize);
3092         if (IS_ERR(em))
3093                 return PTR_ERR(em);
3094
3095         if (em->block_start == EXTENT_MAP_HOLE)
3096                 ret = RANGE_BOUNDARY_HOLE;
3097         else if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
3098                 ret = RANGE_BOUNDARY_PREALLOC_EXTENT;
3099         else
3100                 ret = RANGE_BOUNDARY_WRITTEN_EXTENT;
3101
3102         free_extent_map(em);
3103         return ret;
3104 }
3105
3106 static int btrfs_zero_range(struct inode *inode,
3107                             loff_t offset,
3108                             loff_t len,
3109                             const int mode)
3110 {
3111         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
3112         struct extent_map *em;
3113         struct extent_changeset *data_reserved = NULL;
3114         int ret;
3115         u64 alloc_hint = 0;
3116         const u64 sectorsize = btrfs_inode_sectorsize(BTRFS_I(inode));
3117         u64 alloc_start = round_down(offset, sectorsize);
3118         u64 alloc_end = round_up(offset + len, sectorsize);
3119         u64 bytes_to_reserve = 0;
3120         bool space_reserved = false;
3121
3122         inode_dio_wait(inode);
3123
3124         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
3125                               alloc_end - alloc_start);
3126         if (IS_ERR(em)) {
3127                 ret = PTR_ERR(em);
3128                 goto out;
3129         }
3130
3131         /*
3132          * Avoid hole punching and extent allocation for some cases. More cases
3133          * could be considered, but these are unlikely common and we keep things
3134          * as simple as possible for now. Also, intentionally, if the target
3135          * range contains one or more prealloc extents together with regular
3136          * extents and holes, we drop all the existing extents and allocate a
3137          * new prealloc extent, so that we get a larger contiguous disk extent.
3138          */
3139         if (em->start <= alloc_start &&
3140             test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3141                 const u64 em_end = em->start + em->len;
3142
3143                 if (em_end >= offset + len) {
3144                         /*
3145                          * The whole range is already a prealloc extent,
3146                          * do nothing except updating the inode's i_size if
3147                          * needed.
3148                          */
3149                         free_extent_map(em);
3150                         ret = btrfs_fallocate_update_isize(inode, offset + len,
3151                                                            mode);
3152                         goto out;
3153                 }
3154                 /*
3155                  * Part of the range is already a prealloc extent, so operate
3156                  * only on the remaining part of the range.
3157                  */
3158                 alloc_start = em_end;
3159                 ASSERT(IS_ALIGNED(alloc_start, sectorsize));
3160                 len = offset + len - alloc_start;
3161                 offset = alloc_start;
3162                 alloc_hint = em->block_start + em->len;
3163         }
3164         free_extent_map(em);
3165
3166         if (BTRFS_BYTES_TO_BLKS(fs_info, offset) ==
3167             BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1)) {
3168                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
3169                                       sectorsize);
3170                 if (IS_ERR(em)) {
3171                         ret = PTR_ERR(em);
3172                         goto out;
3173                 }
3174
3175                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3176                         free_extent_map(em);
3177                         ret = btrfs_fallocate_update_isize(inode, offset + len,
3178                                                            mode);
3179                         goto out;
3180                 }
3181                 if (len < sectorsize && em->block_start != EXTENT_MAP_HOLE) {
3182                         free_extent_map(em);
3183                         ret = btrfs_truncate_block(BTRFS_I(inode), offset, len,
3184                                                    0);
3185                         if (!ret)
3186                                 ret = btrfs_fallocate_update_isize(inode,
3187                                                                    offset + len,
3188                                                                    mode);
3189                         return ret;
3190                 }
3191                 free_extent_map(em);
3192                 alloc_start = round_down(offset, sectorsize);
3193                 alloc_end = alloc_start + sectorsize;
3194                 goto reserve_space;
3195         }
3196
3197         alloc_start = round_up(offset, sectorsize);
3198         alloc_end = round_down(offset + len, sectorsize);
3199
3200         /*
3201          * For unaligned ranges, check the pages at the boundaries, they might
3202          * map to an extent, in which case we need to partially zero them, or
3203          * they might map to a hole, in which case we need our allocation range
3204          * to cover them.
3205          */
3206         if (!IS_ALIGNED(offset, sectorsize)) {
3207                 ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
3208                                                             offset);
3209                 if (ret < 0)
3210                         goto out;
3211                 if (ret == RANGE_BOUNDARY_HOLE) {
3212                         alloc_start = round_down(offset, sectorsize);
3213                         ret = 0;
3214                 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3215                         ret = btrfs_truncate_block(BTRFS_I(inode), offset, 0, 0);
3216                         if (ret)
3217                                 goto out;
3218                 } else {
3219                         ret = 0;
3220                 }
3221         }
3222
3223         if (!IS_ALIGNED(offset + len, sectorsize)) {
3224                 ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
3225                                                             offset + len);
3226                 if (ret < 0)
3227                         goto out;
3228                 if (ret == RANGE_BOUNDARY_HOLE) {
3229                         alloc_end = round_up(offset + len, sectorsize);
3230                         ret = 0;
3231                 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3232                         ret = btrfs_truncate_block(BTRFS_I(inode), offset + len,
3233                                                    0, 1);
3234                         if (ret)
3235                                 goto out;
3236                 } else {
3237                         ret = 0;
3238                 }
3239         }
3240
3241 reserve_space:
3242         if (alloc_start < alloc_end) {
3243                 struct extent_state *cached_state = NULL;
3244                 const u64 lockstart = alloc_start;
3245                 const u64 lockend = alloc_end - 1;
3246
3247                 bytes_to_reserve = alloc_end - alloc_start;
3248                 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3249                                                       bytes_to_reserve);
3250                 if (ret < 0)
3251                         goto out;
3252                 space_reserved = true;
3253                 ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
3254                                                   &cached_state);
3255                 if (ret)
3256                         goto out;
3257                 ret = btrfs_qgroup_reserve_data(BTRFS_I(inode), &data_reserved,
3258                                                 alloc_start, bytes_to_reserve);
3259                 if (ret)
3260                         goto out;
3261                 ret = btrfs_prealloc_file_range(inode, mode, alloc_start,
3262                                                 alloc_end - alloc_start,
3263                                                 i_blocksize(inode),
3264                                                 offset + len, &alloc_hint);
3265                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
3266                                      lockend, &cached_state);
3267                 /* btrfs_prealloc_file_range releases reserved space on error */
3268                 if (ret) {
3269                         space_reserved = false;
3270                         goto out;
3271                 }
3272         }
3273         ret = btrfs_fallocate_update_isize(inode, offset + len, mode);
3274  out:
3275         if (ret && space_reserved)
3276                 btrfs_free_reserved_data_space(BTRFS_I(inode), data_reserved,
3277                                                alloc_start, bytes_to_reserve);
3278         extent_changeset_free(data_reserved);
3279
3280         return ret;
3281 }
3282
3283 static long btrfs_fallocate(struct file *file, int mode,
3284                             loff_t offset, loff_t len)
3285 {
3286         struct inode *inode = file_inode(file);
3287         struct extent_state *cached_state = NULL;
3288         struct extent_changeset *data_reserved = NULL;
3289         struct falloc_range *range;
3290         struct falloc_range *tmp;
3291         struct list_head reserve_list;
3292         u64 cur_offset;
3293         u64 last_byte;
3294         u64 alloc_start;
3295         u64 alloc_end;
3296         u64 alloc_hint = 0;
3297         u64 locked_end;
3298         u64 actual_end = 0;
3299         struct extent_map *em;
3300         int blocksize = btrfs_inode_sectorsize(BTRFS_I(inode));
3301         int ret;
3302
3303         /* Do not allow fallocate in ZONED mode */
3304         if (btrfs_is_zoned(btrfs_sb(inode->i_sb)))
3305                 return -EOPNOTSUPP;
3306
3307         alloc_start = round_down(offset, blocksize);
3308         alloc_end = round_up(offset + len, blocksize);
3309         cur_offset = alloc_start;
3310
3311         /* Make sure we aren't being give some crap mode */
3312         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3313                      FALLOC_FL_ZERO_RANGE))
3314                 return -EOPNOTSUPP;
3315
3316         if (mode & FALLOC_FL_PUNCH_HOLE)
3317                 return btrfs_punch_hole(inode, offset, len);
3318
3319         /*
3320          * Only trigger disk allocation, don't trigger qgroup reserve
3321          *
3322          * For qgroup space, it will be checked later.
3323          */
3324         if (!(mode & FALLOC_FL_ZERO_RANGE)) {
3325                 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3326                                                       alloc_end - alloc_start);
3327                 if (ret < 0)
3328                         return ret;
3329         }
3330
3331         btrfs_inode_lock(inode, 0);
3332
3333         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
3334                 ret = inode_newsize_ok(inode, offset + len);
3335                 if (ret)
3336                         goto out;
3337         }
3338
3339         /*
3340          * TODO: Move these two operations after we have checked
3341          * accurate reserved space, or fallocate can still fail but
3342          * with page truncated or size expanded.
3343          *
3344          * But that's a minor problem and won't do much harm BTW.
3345          */
3346         if (alloc_start > inode->i_size) {
3347                 ret = btrfs_cont_expand(BTRFS_I(inode), i_size_read(inode),
3348                                         alloc_start);
3349                 if (ret)
3350                         goto out;
3351         } else if (offset + len > inode->i_size) {
3352                 /*
3353                  * If we are fallocating from the end of the file onward we
3354                  * need to zero out the end of the block if i_size lands in the
3355                  * middle of a block.
3356                  */
3357                 ret = btrfs_truncate_block(BTRFS_I(inode), inode->i_size, 0, 0);
3358                 if (ret)
3359                         goto out;
3360         }
3361
3362         /*
3363          * wait for ordered IO before we have any locks.  We'll loop again
3364          * below with the locks held.
3365          */
3366         ret = btrfs_wait_ordered_range(inode, alloc_start,
3367                                        alloc_end - alloc_start);
3368         if (ret)
3369                 goto out;
3370
3371         if (mode & FALLOC_FL_ZERO_RANGE) {
3372                 ret = btrfs_zero_range(inode, offset, len, mode);
3373                 inode_unlock(inode);
3374                 return ret;
3375         }
3376
3377         locked_end = alloc_end - 1;
3378         while (1) {
3379                 struct btrfs_ordered_extent *ordered;
3380
3381                 /* the extent lock is ordered inside the running
3382                  * transaction
3383                  */
3384                 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
3385                                  locked_end, &cached_state);
3386                 ordered = btrfs_lookup_first_ordered_extent(BTRFS_I(inode),
3387                                                             locked_end);
3388
3389                 if (ordered &&
3390                     ordered->file_offset + ordered->num_bytes > alloc_start &&
3391                     ordered->file_offset < alloc_end) {
3392                         btrfs_put_ordered_extent(ordered);
3393                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
3394                                              alloc_start, locked_end,
3395                                              &cached_state);
3396                         /*
3397                          * we can't wait on the range with the transaction
3398                          * running or with the extent lock held
3399                          */
3400                         ret = btrfs_wait_ordered_range(inode, alloc_start,
3401                                                        alloc_end - alloc_start);
3402                         if (ret)
3403                                 goto out;
3404                 } else {
3405                         if (ordered)
3406                                 btrfs_put_ordered_extent(ordered);
3407                         break;
3408                 }
3409         }
3410
3411         /* First, check if we exceed the qgroup limit */
3412         INIT_LIST_HEAD(&reserve_list);
3413         while (cur_offset < alloc_end) {
3414                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
3415                                       alloc_end - cur_offset);
3416                 if (IS_ERR(em)) {
3417                         ret = PTR_ERR(em);
3418                         break;
3419                 }
3420                 last_byte = min(extent_map_end(em), alloc_end);
3421                 actual_end = min_t(u64, extent_map_end(em), offset + len);
3422                 last_byte = ALIGN(last_byte, blocksize);
3423                 if (em->block_start == EXTENT_MAP_HOLE ||
3424                     (cur_offset >= inode->i_size &&
3425                      !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
3426                         ret = add_falloc_range(&reserve_list, cur_offset,
3427                                                last_byte - cur_offset);
3428                         if (ret < 0) {
3429                                 free_extent_map(em);
3430                                 break;
3431                         }
3432                         ret = btrfs_qgroup_reserve_data(BTRFS_I(inode),
3433                                         &data_reserved, cur_offset,
3434                                         last_byte - cur_offset);
3435                         if (ret < 0) {
3436                                 cur_offset = last_byte;
3437                                 free_extent_map(em);
3438                                 break;
3439                         }
3440                 } else {
3441                         /*
3442                          * Do not need to reserve unwritten extent for this
3443                          * range, free reserved data space first, otherwise
3444                          * it'll result in false ENOSPC error.
3445                          */
3446                         btrfs_free_reserved_data_space(BTRFS_I(inode),
3447                                 data_reserved, cur_offset,
3448                                 last_byte - cur_offset);
3449                 }
3450                 free_extent_map(em);
3451                 cur_offset = last_byte;
3452         }
3453
3454         /*
3455          * If ret is still 0, means we're OK to fallocate.
3456          * Or just cleanup the list and exit.
3457          */
3458         list_for_each_entry_safe(range, tmp, &reserve_list, list) {
3459                 if (!ret)
3460                         ret = btrfs_prealloc_file_range(inode, mode,
3461                                         range->start,
3462                                         range->len, i_blocksize(inode),
3463                                         offset + len, &alloc_hint);
3464                 else
3465                         btrfs_free_reserved_data_space(BTRFS_I(inode),
3466                                         data_reserved, range->start,
3467                                         range->len);
3468                 list_del(&range->list);
3469                 kfree(range);
3470         }
3471         if (ret < 0)
3472                 goto out_unlock;
3473
3474         /*
3475          * We didn't need to allocate any more space, but we still extended the
3476          * size of the file so we need to update i_size and the inode item.
3477          */
3478         ret = btrfs_fallocate_update_isize(inode, actual_end, mode);
3479 out_unlock:
3480         unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3481                              &cached_state);
3482 out:
3483         inode_unlock(inode);
3484         /* Let go of our reservation. */
3485         if (ret != 0 && !(mode & FALLOC_FL_ZERO_RANGE))
3486                 btrfs_free_reserved_data_space(BTRFS_I(inode), data_reserved,
3487                                 cur_offset, alloc_end - cur_offset);
3488         extent_changeset_free(data_reserved);
3489         return ret;
3490 }
3491
3492 static loff_t find_desired_extent(struct inode *inode, loff_t offset,
3493                                   int whence)
3494 {
3495         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3496         struct extent_map *em = NULL;
3497         struct extent_state *cached_state = NULL;
3498         loff_t i_size = inode->i_size;
3499         u64 lockstart;
3500         u64 lockend;
3501         u64 start;
3502         u64 len;
3503         int ret = 0;
3504
3505         if (i_size == 0 || offset >= i_size)
3506                 return -ENXIO;
3507
3508         /*
3509          * offset can be negative, in this case we start finding DATA/HOLE from
3510          * the very start of the file.
3511          */
3512         start = max_t(loff_t, 0, offset);
3513
3514         lockstart = round_down(start, fs_info->sectorsize);
3515         lockend = round_up(i_size, fs_info->sectorsize);
3516         if (lockend <= lockstart)
3517                 lockend = lockstart + fs_info->sectorsize;
3518         lockend--;
3519         len = lockend - lockstart + 1;
3520
3521         lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3522                          &cached_state);
3523
3524         while (start < i_size) {
3525                 em = btrfs_get_extent_fiemap(BTRFS_I(inode), start, len);
3526                 if (IS_ERR(em)) {
3527                         ret = PTR_ERR(em);
3528                         em = NULL;
3529                         break;
3530                 }
3531
3532                 if (whence == SEEK_HOLE &&
3533                     (em->block_start == EXTENT_MAP_HOLE ||
3534                      test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3535                         break;
3536                 else if (whence == SEEK_DATA &&
3537                            (em->block_start != EXTENT_MAP_HOLE &&
3538                             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3539                         break;
3540
3541                 start = em->start + em->len;
3542                 free_extent_map(em);
3543                 em = NULL;
3544                 cond_resched();
3545         }
3546         free_extent_map(em);
3547         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3548                              &cached_state);
3549         if (ret) {
3550                 offset = ret;
3551         } else {
3552                 if (whence == SEEK_DATA && start >= i_size)
3553                         offset = -ENXIO;
3554                 else
3555                         offset = min_t(loff_t, start, i_size);
3556         }
3557
3558         return offset;
3559 }
3560
3561 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3562 {
3563         struct inode *inode = file->f_mapping->host;
3564
3565         switch (whence) {
3566         default:
3567                 return generic_file_llseek(file, offset, whence);
3568         case SEEK_DATA:
3569         case SEEK_HOLE:
3570                 btrfs_inode_lock(inode, BTRFS_ILOCK_SHARED);
3571                 offset = find_desired_extent(inode, offset, whence);
3572                 btrfs_inode_unlock(inode, BTRFS_ILOCK_SHARED);
3573                 break;
3574         }
3575
3576         if (offset < 0)
3577                 return offset;
3578
3579         return vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3580 }
3581
3582 static int btrfs_file_open(struct inode *inode, struct file *filp)
3583 {
3584         filp->f_mode |= FMODE_NOWAIT | FMODE_BUF_RASYNC;
3585         return generic_file_open(inode, filp);
3586 }
3587
3588 static int check_direct_read(struct btrfs_fs_info *fs_info,
3589                              const struct iov_iter *iter, loff_t offset)
3590 {
3591         int ret;
3592         int i, seg;
3593
3594         ret = check_direct_IO(fs_info, iter, offset);
3595         if (ret < 0)
3596                 return ret;
3597
3598         if (!iter_is_iovec(iter))
3599                 return 0;
3600
3601         for (seg = 0; seg < iter->nr_segs; seg++)
3602                 for (i = seg + 1; i < iter->nr_segs; i++)
3603                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
3604                                 return -EINVAL;
3605         return 0;
3606 }
3607
3608 static ssize_t btrfs_direct_read(struct kiocb *iocb, struct iov_iter *to)
3609 {
3610         struct inode *inode = file_inode(iocb->ki_filp);
3611         ssize_t ret;
3612
3613         if (check_direct_read(btrfs_sb(inode->i_sb), to, iocb->ki_pos))
3614                 return 0;
3615
3616         btrfs_inode_lock(inode, BTRFS_ILOCK_SHARED);
3617         ret = iomap_dio_rw(iocb, to, &btrfs_dio_iomap_ops, &btrfs_dio_ops,
3618                            is_sync_kiocb(iocb));
3619         btrfs_inode_unlock(inode, BTRFS_ILOCK_SHARED);
3620         return ret;
3621 }
3622
3623 static ssize_t btrfs_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
3624 {
3625         ssize_t ret = 0;
3626
3627         if (iocb->ki_flags & IOCB_DIRECT) {
3628                 ret = btrfs_direct_read(iocb, to);
3629                 if (ret < 0 || !iov_iter_count(to) ||
3630                     iocb->ki_pos >= i_size_read(file_inode(iocb->ki_filp)))
3631                         return ret;
3632         }
3633
3634         return generic_file_buffered_read(iocb, to, ret);
3635 }
3636
3637 const struct file_operations btrfs_file_operations = {
3638         .llseek         = btrfs_file_llseek,
3639         .read_iter      = btrfs_file_read_iter,
3640         .splice_read    = generic_file_splice_read,
3641         .write_iter     = btrfs_file_write_iter,
3642         .splice_write   = iter_file_splice_write,
3643         .mmap           = btrfs_file_mmap,
3644         .open           = btrfs_file_open,
3645         .release        = btrfs_release_file,
3646         .fsync          = btrfs_sync_file,
3647         .fallocate      = btrfs_fallocate,
3648         .unlocked_ioctl = btrfs_ioctl,
3649 #ifdef CONFIG_COMPAT
3650         .compat_ioctl   = btrfs_compat_ioctl,
3651 #endif
3652         .remap_file_range = btrfs_remap_file_range,
3653 };
3654
3655 void __cold btrfs_auto_defrag_exit(void)
3656 {
3657         kmem_cache_destroy(btrfs_inode_defrag_cachep);
3658 }
3659
3660 int __init btrfs_auto_defrag_init(void)
3661 {
3662         btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
3663                                         sizeof(struct inode_defrag), 0,
3664                                         SLAB_MEM_SPREAD,
3665                                         NULL);
3666         if (!btrfs_inode_defrag_cachep)
3667                 return -ENOMEM;
3668
3669         return 0;
3670 }
3671
3672 int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3673 {
3674         int ret;
3675
3676         /*
3677          * So with compression we will find and lock a dirty page and clear the
3678          * first one as dirty, setup an async extent, and immediately return
3679          * with the entire range locked but with nobody actually marked with
3680          * writeback.  So we can't just filemap_write_and_wait_range() and
3681          * expect it to work since it will just kick off a thread to do the
3682          * actual work.  So we need to call filemap_fdatawrite_range _again_
3683          * since it will wait on the page lock, which won't be unlocked until
3684          * after the pages have been marked as writeback and so we're good to go
3685          * from there.  We have to do this otherwise we'll miss the ordered
3686          * extents and that results in badness.  Please Josef, do not think you
3687          * know better and pull this out at some point in the future, it is
3688          * right and you are wrong.
3689          */
3690         ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3691         if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3692                              &BTRFS_I(inode)->runtime_flags))
3693                 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3694
3695         return ret;
3696 }