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 <linux/fsverity.h>
  20 #include "ctree.h"
  21 #include "disk-io.h"
  22 #include "transaction.h"
  23 #include "btrfs_inode.h"
  24 #include "print-tree.h"
  25 #include "tree-log.h"
  26 #include "locking.h"
  27 #include "volumes.h"
  28 #include "qgroup.h"
  29 #include "compression.h"
  30 #include "delalloc-space.h"
  31 #include "reflink.h"
  32 #include "subpage.h"
  33
  34 static struct kmem_cache *btrfs_inode_defrag_cachep;
  35 /*
  36  * when auto defrag is enabled we
  37  * queue up these defrag structs to remember which
  38  * inodes need defragging passes
  39  */
  40 struct inode_defrag {
  41         struct rb_node rb_node;
  42         /* objectid */
  43         u64 ino;
  44         /*
  45          * transid where the defrag was added, we search for
  46          * extents newer than this
  47          */
  48         u64 transid;
  49
  50         /* root objectid */
  51         u64 root;
  52
  53         /* last offset we were able to defrag */
  54         u64 last_offset;
  55
  56         /* if we've wrapped around back to zero once already */
  57         int cycled;
  58 };
  59
  60 static int __compare_inode_defrag(struct inode_defrag *defrag1,
  61                                   struct inode_defrag *defrag2)
  62 {
  63         if (defrag1->root > defrag2->root)
  64                 return 1;
  65         else if (defrag1->root < defrag2->root)
  66                 return -1;
  67         else if (defrag1->ino > defrag2->ino)
  68                 return 1;
  69         else if (defrag1->ino < defrag2->ino)
  70                 return -1;
  71         else
  72                 return 0;
  73 }
  74
  75 /* pop a record for an inode into the defrag tree.  The lock
  76  * must be held already
  77  *
  78  * If you're inserting a record for an older transid than an
  79  * existing record, the transid already in the tree is lowered
  80  *
  81  * If an existing record is found the defrag item you
  82  * pass in is freed
  83  */
  84 static int __btrfs_add_inode_defrag(struct btrfs_inode *inode,
  85                                     struct inode_defrag *defrag)
  86 {
  87         struct btrfs_fs_info *fs_info = inode->root->fs_info;
  88         struct inode_defrag *entry;
  89         struct rb_node **p;
  90         struct rb_node *parent = NULL;
  91         int ret;
  92
  93         p = &fs_info->defrag_inodes.rb_node;
  94         while (*p) {
  95                 parent = *p;
  96                 entry = rb_entry(parent, struct inode_defrag, rb_node);
  97
  98                 ret = __compare_inode_defrag(defrag, entry);
  99                 if (ret < 0)
 100                         p = &parent->rb_left;
 101                 else if (ret > 0)
 102                         p = &parent->rb_right;
 103                 else {
 104                         /* if we're reinserting an entry for
 105                          * an old defrag run, make sure to
 106                          * lower the transid of our existing record
 107                          */
 108                         if (defrag->transid < entry->transid)
 109                                 entry->transid = defrag->transid;
 110                         if (defrag->last_offset > entry->last_offset)
 111                                 entry->last_offset = defrag->last_offset;
 112                         return -EEXIST;
 113                 }
 114         }
 115         set_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags);
 116         rb_link_node(&defrag->rb_node, parent, p);
 117         rb_insert_color(&defrag->rb_node, &fs_info->defrag_inodes);
 118         return 0;
 119 }
 120
 121 static inline int __need_auto_defrag(struct btrfs_fs_info *fs_info)
 122 {
 123         if (!btrfs_test_opt(fs_info, AUTO_DEFRAG))
 124                 return 0;
 125
 126         if (btrfs_fs_closing(fs_info))
 127                 return 0;
 128
 129         return 1;
 130 }
 131
 132 /*
 133  * insert a defrag record for this inode if auto defrag is
 134  * enabled
 135  */
 136 int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
 137                            struct btrfs_inode *inode)
 138 {
 139         struct btrfs_root *root = inode->root;
 140         struct btrfs_fs_info *fs_info = root->fs_info;
 141         struct inode_defrag *defrag;
 142         u64 transid;
 143         int ret;
 144
 145         if (!__need_auto_defrag(fs_info))
 146                 return 0;
 147
 148         if (test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags))
 149                 return 0;
 150
 151         if (trans)
 152                 transid = trans->transid;
 153         else
 154                 transid = inode->root->last_trans;
 155
 156         defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
 157         if (!defrag)
 158                 return -ENOMEM;
 159
 160         defrag->ino = btrfs_ino(inode);
 161         defrag->transid = transid;
 162         defrag->root = root->root_key.objectid;
 163
 164         spin_lock(&fs_info->defrag_inodes_lock);
 165         if (!test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags)) {
 166                 /*
 167                  * If we set IN_DEFRAG flag and evict the inode from memory,
 168                  * and then re-read this inode, this new inode doesn't have
 169                  * IN_DEFRAG flag. At the case, we may find the existed defrag.
 170                  */
 171                 ret = __btrfs_add_inode_defrag(inode, defrag);
 172                 if (ret)
 173                         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 174         } else {
 175                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 176         }
 177         spin_unlock(&fs_info->defrag_inodes_lock);
 178         return 0;
 179 }
 180
 181 /*
 182  * Requeue the defrag object. If there is a defrag object that points to
 183  * the same inode in the tree, we will merge them together (by
 184  * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
 185  */
 186 static void btrfs_requeue_inode_defrag(struct btrfs_inode *inode,
 187                                        struct inode_defrag *defrag)
 188 {
 189         struct btrfs_fs_info *fs_info = inode->root->fs_info;
 190         int ret;
 191
 192         if (!__need_auto_defrag(fs_info))
 193                 goto out;
 194
 195         /*
 196          * Here we don't check the IN_DEFRAG flag, because we need merge
 197          * them together.
 198          */
 199         spin_lock(&fs_info->defrag_inodes_lock);
 200         ret = __btrfs_add_inode_defrag(inode, defrag);
 201         spin_unlock(&fs_info->defrag_inodes_lock);
 202         if (ret)
 203                 goto out;
 204         return;
 205 out:
 206         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 207 }
 208
 209 /*
 210  * pick the defragable inode that we want, if it doesn't exist, we will get
 211  * the next one.
 212  */
 213 static struct inode_defrag *
 214 btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
 215 {
 216         struct inode_defrag *entry = NULL;
 217         struct inode_defrag tmp;
 218         struct rb_node *p;
 219         struct rb_node *parent = NULL;
 220         int ret;
 221
 222         tmp.ino = ino;
 223         tmp.root = root;
 224
 225         spin_lock(&fs_info->defrag_inodes_lock);
 226         p = fs_info->defrag_inodes.rb_node;
 227         while (p) {
 228                 parent = p;
 229                 entry = rb_entry(parent, struct inode_defrag, rb_node);
 230
 231                 ret = __compare_inode_defrag(&tmp, entry);
 232                 if (ret < 0)
 233                         p = parent->rb_left;
 234                 else if (ret > 0)
 235                         p = parent->rb_right;
 236                 else
 237                         goto out;
 238         }
 239
 240         if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
 241                 parent = rb_next(parent);
 242                 if (parent)
 243                         entry = rb_entry(parent, struct inode_defrag, rb_node);
 244                 else
 245                         entry = NULL;
 246         }
 247 out:
 248         if (entry)
 249                 rb_erase(parent, &fs_info->defrag_inodes);
 250         spin_unlock(&fs_info->defrag_inodes_lock);
 251         return entry;
 252 }
 253
 254 void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
 255 {
 256         struct inode_defrag *defrag;
 257         struct rb_node *node;
 258
 259         spin_lock(&fs_info->defrag_inodes_lock);
 260         node = rb_first(&fs_info->defrag_inodes);
 261         while (node) {
 262                 rb_erase(node, &fs_info->defrag_inodes);
 263                 defrag = rb_entry(node, struct inode_defrag, rb_node);
 264                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 265
 266                 cond_resched_lock(&fs_info->defrag_inodes_lock);
 267
 268                 node = rb_first(&fs_info->defrag_inodes);
 269         }
 270         spin_unlock(&fs_info->defrag_inodes_lock);
 271 }
 272
 273 #define BTRFS_DEFRAG_BATCH      1024
 274
 275 static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
 276                                     struct inode_defrag *defrag)
 277 {
 278         struct btrfs_root *inode_root;
 279         struct inode *inode;
 280         struct btrfs_ioctl_defrag_range_args range;
 281         int num_defrag;
 282         int ret;
 283
 284         /* get the inode */
 285         inode_root = btrfs_get_fs_root(fs_info, defrag->root, true);
 286         if (IS_ERR(inode_root)) {
 287                 ret = PTR_ERR(inode_root);
 288                 goto cleanup;
 289         }
 290
 291         inode = btrfs_iget(fs_info->sb, defrag->ino, inode_root);
 292         btrfs_put_root(inode_root);
 293         if (IS_ERR(inode)) {
 294                 ret = PTR_ERR(inode);
 295                 goto cleanup;
 296         }
 297
 298         /* do a chunk of defrag */
 299         clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
 300         memset(&range, 0, sizeof(range));
 301         range.len = (u64)-1;
 302         range.start = defrag->last_offset;
 303
 304         sb_start_write(fs_info->sb);
 305         num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
 306                                        BTRFS_DEFRAG_BATCH);
 307         sb_end_write(fs_info->sb);
 308         /*
 309          * if we filled the whole defrag batch, there
 310          * must be more work to do.  Queue this defrag
 311          * again
 312          */
 313         if (num_defrag == BTRFS_DEFRAG_BATCH) {
 314                 defrag->last_offset = range.start;
 315                 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
 316         } else if (defrag->last_offset && !defrag->cycled) {
 317                 /*
 318                  * we didn't fill our defrag batch, but
 319                  * we didn't start at zero.  Make sure we loop
 320                  * around to the start of the file.
 321                  */
 322                 defrag->last_offset = 0;
 323                 defrag->cycled = 1;
 324                 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
 325         } else {
 326                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 327         }
 328
 329         iput(inode);
 330         return 0;
 331 cleanup:
 332         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 333         return ret;
 334 }
 335
 336 /*
 337  * run through the list of inodes in the FS that need
 338  * defragging
 339  */
 340 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
 341 {
 342         struct inode_defrag *defrag;
 343         u64 first_ino = 0;
 344         u64 root_objectid = 0;
 345
 346         atomic_inc(&fs_info->defrag_running);
 347         while (1) {
 348                 /* Pause the auto defragger. */
 349                 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
 350                              &fs_info->fs_state))
 351                         break;
 352
 353                 if (!__need_auto_defrag(fs_info))
 354                         break;
 355
 356                 /* find an inode to defrag */
 357                 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
 358                                                  first_ino);
 359                 if (!defrag) {
 360                         if (root_objectid || first_ino) {
 361                                 root_objectid = 0;
 362                                 first_ino = 0;
 363                                 continue;
 364                         } else {
 365                                 break;
 366                         }
 367                 }
 368
 369                 first_ino = defrag->ino + 1;
 370                 root_objectid = defrag->root;
 371
 372                 __btrfs_run_defrag_inode(fs_info, defrag);
 373         }
 374         atomic_dec(&fs_info->defrag_running);
 375
 376         /*
 377          * during unmount, we use the transaction_wait queue to
 378          * wait for the defragger to stop
 379          */
 380         wake_up(&fs_info->transaction_wait);
 381         return 0;
 382 }
 383
 384 /* simple helper to fault in pages and copy.  This should go away
 385  * and be replaced with calls into generic code.
 386  */
 387 static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
 388                                          struct page **prepared_pages,
 389                                          struct iov_iter *i)
 390 {
 391         size_t copied = 0;
 392         size_t total_copied = 0;
 393         int pg = 0;
 394         int offset = offset_in_page(pos);
 395
 396         while (write_bytes > 0) {
 397                 size_t count = min_t(size_t,
 398                                      PAGE_SIZE - offset, write_bytes);
 399                 struct page *page = prepared_pages[pg];
 400                 /*
 401                  * Copy data from userspace to the current page
 402                  */
 403                 copied = copy_page_from_iter_atomic(page, offset, count, i);
 404
 405                 /* Flush processor's dcache for this page */
 406                 flush_dcache_page(page);
 407
 408                 /*
 409                  * if we get a partial write, we can end up with
 410                  * partially up to date pages.  These add
 411                  * a lot of complexity, so make sure they don't
 412                  * happen by forcing this copy to be retried.
 413                  *
 414                  * The rest of the btrfs_file_write code will fall
 415                  * back to page at a time copies after we return 0.
 416                  */
 417                 if (unlikely(copied < count)) {
 418                         if (!PageUptodate(page)) {
 419                                 iov_iter_revert(i, copied);
 420                                 copied = 0;
 421                         }
 422                         if (!copied)
 423                                 break;
 424                 }
 425
 426                 write_bytes -= copied;
 427                 total_copied += copied;
 428                 offset += copied;
 429                 if (offset == PAGE_SIZE) {
 430                         pg++;
 431                         offset = 0;
 432                 }
 433         }
 434         return total_copied;
 435 }
 436
 437 /*
 438  * unlocks pages after btrfs_file_write is done with them
 439  */
 440 static void btrfs_drop_pages(struct page **pages, size_t num_pages)
 441 {
 442         size_t i;
 443         for (i = 0; i < num_pages; i++) {
 444                 /* page checked is some magic around finding pages that
 445                  * have been modified without going through btrfs_set_page_dirty
 446                  * clear it here. There should be no need to mark the pages
 447                  * accessed as prepare_pages should have marked them accessed
 448                  * in prepare_pages via find_or_create_page()
 449                  */
 450                 ClearPageChecked(pages[i]);
 451                 unlock_page(pages[i]);
 452                 put_page(pages[i]);
 453         }
 454 }
 455
 456 /*
 457  * After btrfs_copy_from_user(), update the following things for delalloc:
 458  * - Mark newly dirtied pages as DELALLOC in the io tree.
 459  *   Used to advise which range is to be written back.
 460  * - Mark modified pages as Uptodate/Dirty and not needing COW fixup
 461  * - Update inode size for past EOF write
 462  */
 463 int btrfs_dirty_pages(struct btrfs_inode *inode, struct page **pages,
 464                       size_t num_pages, loff_t pos, size_t write_bytes,
 465                       struct extent_state **cached, bool noreserve)
 466 {
 467         struct btrfs_fs_info *fs_info = inode->root->fs_info;
 468         int err = 0;
 469         int i;
 470         u64 num_bytes;
 471         u64 start_pos;
 472         u64 end_of_last_block;
 473         u64 end_pos = pos + write_bytes;
 474         loff_t isize = i_size_read(&inode->vfs_inode);
 475         unsigned int extra_bits = 0;
 476
 477         if (write_bytes == 0)
 478                 return 0;
 479
 480         if (noreserve)
 481                 extra_bits |= EXTENT_NORESERVE;
 482
 483         start_pos = round_down(pos, fs_info->sectorsize);
 484         num_bytes = round_up(write_bytes + pos - start_pos,
 485                              fs_info->sectorsize);
 486         ASSERT(num_bytes <= U32_MAX);
 487
 488         end_of_last_block = start_pos + num_bytes - 1;
 489
 490         /*
 491          * The pages may have already been dirty, clear out old accounting so
 492          * we can set things up properly
 493          */
 494         clear_extent_bit(&inode->io_tree, start_pos, end_of_last_block,
 495                          EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
 496                          0, 0, cached);
 497
 498         err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
 499                                         extra_bits, cached);
 500         if (err)
 501                 return err;
 502
 503         for (i = 0; i < num_pages; i++) {
 504                 struct page *p = pages[i];
 505
 506                 btrfs_page_clamp_set_uptodate(fs_info, p, start_pos, num_bytes);
 507                 ClearPageChecked(p);
 508                 btrfs_page_clamp_set_dirty(fs_info, p, start_pos, num_bytes);
 509         }
 510
 511         /*
 512          * we've only changed i_size in ram, and we haven't updated
 513          * the disk i_size.  There is no need to log the inode
 514          * at this time.
 515          */
 516         if (end_pos > isize)
 517                 i_size_write(&inode->vfs_inode, end_pos);
 518         return 0;
 519 }
 520
 521 /*
 522  * this drops all the extents in the cache that intersect the range
 523  * [start, end].  Existing extents are split as required.
 524  */
 525 void btrfs_drop_extent_cache(struct btrfs_inode *inode, u64 start, u64 end,
 526                              int skip_pinned)
 527 {
 528         struct extent_map *em;
 529         struct extent_map *split = NULL;
 530         struct extent_map *split2 = NULL;
 531         struct extent_map_tree *em_tree = &inode->extent_tree;
 532         u64 len = end - start + 1;
 533         u64 gen;
 534         int ret;
 535         int testend = 1;
 536         unsigned long flags;
 537         int compressed = 0;
 538         bool modified;
 539
 540         WARN_ON(end < start);
 541         if (end == (u64)-1) {
 542                 len = (u64)-1;
 543                 testend = 0;
 544         }
 545         while (1) {
 546                 int no_splits = 0;
 547
 548                 modified = false;
 549                 if (!split)
 550                         split = alloc_extent_map();
 551                 if (!split2)
 552                         split2 = alloc_extent_map();
 553                 if (!split || !split2)
 554                         no_splits = 1;
 555
 556                 write_lock(&em_tree->lock);
 557                 em = lookup_extent_mapping(em_tree, start, len);
 558                 if (!em) {
 559                         write_unlock(&em_tree->lock);
 560                         break;
 561                 }
 562                 flags = em->flags;
 563                 gen = em->generation;
 564                 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
 565                         if (testend && em->start + em->len >= start + len) {
 566                                 free_extent_map(em);
 567                                 write_unlock(&em_tree->lock);
 568                                 break;
 569                         }
 570                         start = em->start + em->len;
 571                         if (testend)
 572                                 len = start + len - (em->start + em->len);
 573                         free_extent_map(em);
 574                         write_unlock(&em_tree->lock);
 575                         continue;
 576                 }
 577                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 578                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
 579                 clear_bit(EXTENT_FLAG_LOGGING, &flags);
 580                 modified = !list_empty(&em->list);
 581                 if (no_splits)
 582                         goto next;
 583
 584                 if (em->start < start) {
 585                         split->start = em->start;
 586                         split->len = start - em->start;
 587
 588                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 589                                 split->orig_start = em->orig_start;
 590                                 split->block_start = em->block_start;
 591
 592                                 if (compressed)
 593                                         split->block_len = em->block_len;
 594                                 else
 595                                         split->block_len = split->len;
 596                                 split->orig_block_len = max(split->block_len,
 597                                                 em->orig_block_len);
 598                                 split->ram_bytes = em->ram_bytes;
 599                         } else {
 600                                 split->orig_start = split->start;
 601                                 split->block_len = 0;
 602                                 split->block_start = em->block_start;
 603                                 split->orig_block_len = 0;
 604                                 split->ram_bytes = split->len;
 605                         }
 606
 607                         split->generation = gen;
 608                         split->flags = flags;
 609                         split->compress_type = em->compress_type;
 610                         replace_extent_mapping(em_tree, em, split, modified);
 611                         free_extent_map(split);
 612                         split = split2;
 613                         split2 = NULL;
 614                 }
 615                 if (testend && em->start + em->len > start + len) {
 616                         u64 diff = start + len - em->start;
 617
 618                         split->start = start + len;
 619                         split->len = em->start + em->len - (start + len);
 620                         split->flags = flags;
 621                         split->compress_type = em->compress_type;
 622                         split->generation = gen;
 623
 624                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 625                                 split->orig_block_len = max(em->block_len,
 626                                                     em->orig_block_len);
 627
 628                                 split->ram_bytes = em->ram_bytes;
 629                                 if (compressed) {
 630                                         split->block_len = em->block_len;
 631                                         split->block_start = em->block_start;
 632                                         split->orig_start = em->orig_start;
 633                                 } else {
 634                                         split->block_len = split->len;
 635                                         split->block_start = em->block_start
 636                                                 + diff;
 637                                         split->orig_start = em->orig_start;
 638                                 }
 639                         } else {
 640                                 split->ram_bytes = split->len;
 641                                 split->orig_start = split->start;
 642                                 split->block_len = 0;
 643                                 split->block_start = em->block_start;
 644                                 split->orig_block_len = 0;
 645                         }
 646
 647                         if (extent_map_in_tree(em)) {
 648                                 replace_extent_mapping(em_tree, em, split,
 649                                                        modified);
 650                         } else {
 651                                 ret = add_extent_mapping(em_tree, split,
 652                                                          modified);
 653                                 ASSERT(ret == 0); /* Logic error */
 654                         }
 655                         free_extent_map(split);
 656                         split = NULL;
 657                 }
 658 next:
 659                 if (extent_map_in_tree(em))
 660                         remove_extent_mapping(em_tree, em);
 661                 write_unlock(&em_tree->lock);
 662
 663                 /* once for us */
 664                 free_extent_map(em);
 665                 /* once for the tree*/
 666                 free_extent_map(em);
 667         }
 668         if (split)
 669                 free_extent_map(split);
 670         if (split2)
 671                 free_extent_map(split2);
 672 }
 673
 674 /*
 675  * this is very complex, but the basic idea is to drop all extents
 676  * in the range start - end.  hint_block is filled in with a block number
 677  * that would be a good hint to the block allocator for this file.
 678  *
 679  * If an extent intersects the range but is not entirely inside the range
 680  * it is either truncated or split.  Anything entirely inside the range
 681  * is deleted from the tree.
 682  *
 683  * Note: the VFS' inode number of bytes is not updated, it's up to the caller
 684  * to deal with that. We set the field 'bytes_found' of the arguments structure
 685  * with the number of allocated bytes found in the target range, so that the
 686  * caller can update the inode's number of bytes in an atomic way when
 687  * replacing extents in a range to avoid races with stat(2).
 688  */
 689 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
 690                        struct btrfs_root *root, struct btrfs_inode *inode,
 691                        struct btrfs_drop_extents_args *args)
 692 {
 693         struct btrfs_fs_info *fs_info = root->fs_info;
 694         struct extent_buffer *leaf;
 695         struct btrfs_file_extent_item *fi;
 696         struct btrfs_ref ref = { 0 };
 697         struct btrfs_key key;
 698         struct btrfs_key new_key;
 699         u64 ino = btrfs_ino(inode);
 700         u64 search_start = args->start;
 701         u64 disk_bytenr = 0;
 702         u64 num_bytes = 0;
 703         u64 extent_offset = 0;
 704         u64 extent_end = 0;
 705         u64 last_end = args->start;
 706         int del_nr = 0;
 707         int del_slot = 0;
 708         int extent_type;
 709         int recow;
 710         int ret;
 711         int modify_tree = -1;
 712         int update_refs;
 713         int found = 0;
 714         int leafs_visited = 0;
 715         struct btrfs_path *path = args->path;
 716
 717         args->bytes_found = 0;
 718         args->extent_inserted = false;
 719
 720         /* Must always have a path if ->replace_extent is true */
 721         ASSERT(!(args->replace_extent && !args->path));
 722
 723         if (!path) {
 724                 path = btrfs_alloc_path();
 725                 if (!path) {
 726                         ret = -ENOMEM;
 727                         goto out;
 728                 }
 729         }
 730
 731         if (args->drop_cache)
 732                 btrfs_drop_extent_cache(inode, args->start, args->end - 1, 0);
 733
 734         if (args->start >= inode->disk_i_size && !args->replace_extent)
 735                 modify_tree = 0;
 736
 737         update_refs = (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID);
 738         while (1) {
 739                 recow = 0;
 740                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
 741                                                search_start, modify_tree);
 742                 if (ret < 0)
 743                         break;
 744                 if (ret > 0 && path->slots[0] > 0 && search_start == args->start) {
 745                         leaf = path->nodes[0];
 746                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
 747                         if (key.objectid == ino &&
 748                             key.type == BTRFS_EXTENT_DATA_KEY)
 749                                 path->slots[0]--;
 750                 }
 751                 ret = 0;
 752                 leafs_visited++;
 753 next_slot:
 754                 leaf = path->nodes[0];
 755                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 756                         BUG_ON(del_nr > 0);
 757                         ret = btrfs_next_leaf(root, path);
 758                         if (ret < 0)
 759                                 break;
 760                         if (ret > 0) {
 761                                 ret = 0;
 762                                 break;
 763                         }
 764                         leafs_visited++;
 765                         leaf = path->nodes[0];
 766                         recow = 1;
 767                 }
 768
 769                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 770
 771                 if (key.objectid > ino)
 772                         break;
 773                 if (WARN_ON_ONCE(key.objectid < ino) ||
 774                     key.type < BTRFS_EXTENT_DATA_KEY) {
 775                         ASSERT(del_nr == 0);
 776                         path->slots[0]++;
 777                         goto next_slot;
 778                 }
 779                 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= args->end)
 780                         break;
 781
 782                 fi = btrfs_item_ptr(leaf, path->slots[0],
 783                                     struct btrfs_file_extent_item);
 784                 extent_type = btrfs_file_extent_type(leaf, fi);
 785
 786                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
 787                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
 788                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 789                         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 790                         extent_offset = btrfs_file_extent_offset(leaf, fi);
 791                         extent_end = key.offset +
 792                                 btrfs_file_extent_num_bytes(leaf, fi);
 793                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 794                         extent_end = key.offset +
 795                                 btrfs_file_extent_ram_bytes(leaf, fi);
 796                 } else {
 797                         /* can't happen */
 798                         BUG();
 799                 }
 800
 801                 /*
 802                  * Don't skip extent items representing 0 byte lengths. They
 803                  * used to be created (bug) if while punching holes we hit
 804                  * -ENOSPC condition. So if we find one here, just ensure we
 805                  * delete it, otherwise we would insert a new file extent item
 806                  * with the same key (offset) as that 0 bytes length file
 807                  * extent item in the call to setup_items_for_insert() later
 808                  * in this function.
 809                  */
 810                 if (extent_end == key.offset && extent_end >= search_start) {
 811                         last_end = extent_end;
 812                         goto delete_extent_item;
 813                 }
 814
 815                 if (extent_end <= search_start) {
 816                         path->slots[0]++;
 817                         goto next_slot;
 818                 }
 819
 820                 found = 1;
 821                 search_start = max(key.offset, args->start);
 822                 if (recow || !modify_tree) {
 823                         modify_tree = -1;
 824                         btrfs_release_path(path);
 825                         continue;
 826                 }
 827
 828                 /*
 829                  *     | - range to drop - |
 830                  *  | -------- extent -------- |
 831                  */
 832                 if (args->start > key.offset && args->end < extent_end) {
 833                         BUG_ON(del_nr > 0);
 834                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 835                                 ret = -EOPNOTSUPP;
 836                                 break;
 837                         }
 838
 839                         memcpy(&new_key, &key, sizeof(new_key));
 840                         new_key.offset = args->start;
 841                         ret = btrfs_duplicate_item(trans, root, path,
 842                                                    &new_key);
 843                         if (ret == -EAGAIN) {
 844                                 btrfs_release_path(path);
 845                                 continue;
 846                         }
 847                         if (ret < 0)
 848                                 break;
 849
 850                         leaf = path->nodes[0];
 851                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 852                                             struct btrfs_file_extent_item);
 853                         btrfs_set_file_extent_num_bytes(leaf, fi,
 854                                                         args->start - key.offset);
 855
 856                         fi = btrfs_item_ptr(leaf, path->slots[0],
 857                                             struct btrfs_file_extent_item);
 858
 859                         extent_offset += args->start - key.offset;
 860                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 861                         btrfs_set_file_extent_num_bytes(leaf, fi,
 862                                                         extent_end - args->start);
 863                         btrfs_mark_buffer_dirty(leaf);
 864
 865                         if (update_refs && disk_bytenr > 0) {
 866                                 btrfs_init_generic_ref(&ref,
 867                                                 BTRFS_ADD_DELAYED_REF,
 868                                                 disk_bytenr, num_bytes, 0);
 869                                 btrfs_init_data_ref(&ref,
 870                                                 root->root_key.objectid,
 871                                                 new_key.objectid,
 872                                                 args->start - extent_offset,
 873                                                 0, false);
 874                                 ret = btrfs_inc_extent_ref(trans, &ref);
 875                                 BUG_ON(ret); /* -ENOMEM */
 876                         }
 877                         key.offset = args->start;
 878                 }
 879                 /*
 880                  * From here on out we will have actually dropped something, so
 881                  * last_end can be updated.
 882                  */
 883                 last_end = extent_end;
 884
 885                 /*
 886                  *  | ---- range to drop ----- |
 887                  *      | -------- extent -------- |
 888                  */
 889                 if (args->start <= key.offset && args->end < extent_end) {
 890                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 891                                 ret = -EOPNOTSUPP;
 892                                 break;
 893                         }
 894
 895                         memcpy(&new_key, &key, sizeof(new_key));
 896                         new_key.offset = args->end;
 897                         btrfs_set_item_key_safe(fs_info, path, &new_key);
 898
 899                         extent_offset += args->end - key.offset;
 900                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 901                         btrfs_set_file_extent_num_bytes(leaf, fi,
 902                                                         extent_end - args->end);
 903                         btrfs_mark_buffer_dirty(leaf);
 904                         if (update_refs && disk_bytenr > 0)
 905                                 args->bytes_found += args->end - key.offset;
 906                         break;
 907                 }
 908
 909                 search_start = extent_end;
 910                 /*
 911                  *       | ---- range to drop ----- |
 912                  *  | -------- extent -------- |
 913                  */
 914                 if (args->start > key.offset && args->end >= extent_end) {
 915                         BUG_ON(del_nr > 0);
 916                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 917                                 ret = -EOPNOTSUPP;
 918                                 break;
 919                         }
 920
 921                         btrfs_set_file_extent_num_bytes(leaf, fi,
 922                                                         args->start - key.offset);
 923                         btrfs_mark_buffer_dirty(leaf);
 924                         if (update_refs && disk_bytenr > 0)
 925                                 args->bytes_found += extent_end - args->start;
 926                         if (args->end == extent_end)
 927                                 break;
 928
 929                         path->slots[0]++;
 930                         goto next_slot;
 931                 }
 932
 933                 /*
 934                  *  | ---- range to drop ----- |
 935                  *    | ------ extent ------ |
 936                  */
 937                 if (args->start <= key.offset && args->end >= extent_end) {
 938 delete_extent_item:
 939                         if (del_nr == 0) {
 940                                 del_slot = path->slots[0];
 941                                 del_nr = 1;
 942                         } else {
 943                                 BUG_ON(del_slot + del_nr != path->slots[0]);
 944                                 del_nr++;
 945                         }
 946
 947                         if (update_refs &&
 948                             extent_type == BTRFS_FILE_EXTENT_INLINE) {
 949                                 args->bytes_found += extent_end - key.offset;
 950                                 extent_end = ALIGN(extent_end,
 951                                                    fs_info->sectorsize);
 952                         } else if (update_refs && disk_bytenr > 0) {
 953                                 btrfs_init_generic_ref(&ref,
 954                                                 BTRFS_DROP_DELAYED_REF,
 955                                                 disk_bytenr, num_bytes, 0);
 956                                 btrfs_init_data_ref(&ref,
 957                                                 root->root_key.objectid,
 958                                                 key.objectid,
 959                                                 key.offset - extent_offset, 0,
 960                                                 false);
 961                                 ret = btrfs_free_extent(trans, &ref);
 962                                 BUG_ON(ret); /* -ENOMEM */
 963                                 args->bytes_found += extent_end - key.offset;
 964                         }
 965
 966                         if (args->end == extent_end)
 967                                 break;
 968
 969                         if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
 970                                 path->slots[0]++;
 971                                 goto next_slot;
 972                         }
 973
 974                         ret = btrfs_del_items(trans, root, path, del_slot,
 975                                               del_nr);
 976                         if (ret) {
 977                                 btrfs_abort_transaction(trans, ret);
 978                                 break;
 979                         }
 980
 981                         del_nr = 0;
 982                         del_slot = 0;
 983
 984                         btrfs_release_path(path);
 985                         continue;
 986                 }
 987
 988                 BUG();
 989         }
 990
 991         if (!ret && del_nr > 0) {
 992                 /*
 993                  * Set path->slots[0] to first slot, so that after the delete
 994                  * if items are move off from our leaf to its immediate left or
 995                  * right neighbor leafs, we end up with a correct and adjusted
 996                  * path->slots[0] for our insertion (if args->replace_extent).
 997                  */
 998                 path->slots[0] = del_slot;
 999                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1000                 if (ret)
1001                         btrfs_abort_transaction(trans, ret);
1002         }
1003
1004         leaf = path->nodes[0];
1005         /*
1006          * If btrfs_del_items() was called, it might have deleted a leaf, in
1007          * which case it unlocked our path, so check path->locks[0] matches a
1008          * write lock.
1009          */
1010         if (!ret && args->replace_extent && leafs_visited == 1 &&
1011             path->locks[0] == BTRFS_WRITE_LOCK &&
1012             btrfs_leaf_free_space(leaf) >=
1013             sizeof(struct btrfs_item) + args->extent_item_size) {
1014
1015                 key.objectid = ino;
1016                 key.type = BTRFS_EXTENT_DATA_KEY;
1017                 key.offset = args->start;
1018                 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
1019                         struct btrfs_key slot_key;
1020
1021                         btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
1022                         if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
1023                                 path->slots[0]++;
1024                 }
1025                 setup_items_for_insert(root, path, &key,
1026                                        &args->extent_item_size, 1);
1027                 args->extent_inserted = true;
1028         }
1029
1030         if (!args->path)
1031                 btrfs_free_path(path);
1032         else if (!args->extent_inserted)
1033                 btrfs_release_path(path);
1034 out:
1035         args->drop_end = found ? min(args->end, last_end) : args->end;
1036
1037         return ret;
1038 }
1039
1040 static int extent_mergeable(struct extent_buffer *leaf, int slot,
1041                             u64 objectid, u64 bytenr, u64 orig_offset,
1042                             u64 *start, u64 *end)
1043 {
1044         struct btrfs_file_extent_item *fi;
1045         struct btrfs_key key;
1046         u64 extent_end;
1047
1048         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1049                 return 0;
1050
1051         btrfs_item_key_to_cpu(leaf, &key, slot);
1052         if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1053                 return 0;
1054
1055         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1056         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1057             btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1058             btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1059             btrfs_file_extent_compression(leaf, fi) ||
1060             btrfs_file_extent_encryption(leaf, fi) ||
1061             btrfs_file_extent_other_encoding(leaf, fi))
1062                 return 0;
1063
1064         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1065         if ((*start && *start != key.offset) || (*end && *end != extent_end))
1066                 return 0;
1067
1068         *start = key.offset;
1069         *end = extent_end;
1070         return 1;
1071 }
1072
1073 /*
1074  * Mark extent in the range start - end as written.
1075  *
1076  * This changes extent type from 'pre-allocated' to 'regular'. If only
1077  * part of extent is marked as written, the extent will be split into
1078  * two or three.
1079  */
1080 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1081                               struct btrfs_inode *inode, u64 start, u64 end)
1082 {
1083         struct btrfs_fs_info *fs_info = trans->fs_info;
1084         struct btrfs_root *root = inode->root;
1085         struct extent_buffer *leaf;
1086         struct btrfs_path *path;
1087         struct btrfs_file_extent_item *fi;
1088         struct btrfs_ref ref = { 0 };
1089         struct btrfs_key key;
1090         struct btrfs_key new_key;
1091         u64 bytenr;
1092         u64 num_bytes;
1093         u64 extent_end;
1094         u64 orig_offset;
1095         u64 other_start;
1096         u64 other_end;
1097         u64 split;
1098         int del_nr = 0;
1099         int del_slot = 0;
1100         int recow;
1101         int ret = 0;
1102         u64 ino = btrfs_ino(inode);
1103
1104         path = btrfs_alloc_path();
1105         if (!path)
1106                 return -ENOMEM;
1107 again:
1108         recow = 0;
1109         split = start;
1110         key.objectid = ino;
1111         key.type = BTRFS_EXTENT_DATA_KEY;
1112         key.offset = split;
1113
1114         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1115         if (ret < 0)
1116                 goto out;
1117         if (ret > 0 && path->slots[0] > 0)
1118                 path->slots[0]--;
1119
1120         leaf = path->nodes[0];
1121         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1122         if (key.objectid != ino ||
1123             key.type != BTRFS_EXTENT_DATA_KEY) {
1124                 ret = -EINVAL;
1125                 btrfs_abort_transaction(trans, ret);
1126                 goto out;
1127         }
1128         fi = btrfs_item_ptr(leaf, path->slots[0],
1129                             struct btrfs_file_extent_item);
1130         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
1131                 ret = -EINVAL;
1132                 btrfs_abort_transaction(trans, ret);
1133                 goto out;
1134         }
1135         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1136         if (key.offset > start || extent_end < end) {
1137                 ret = -EINVAL;
1138                 btrfs_abort_transaction(trans, ret);
1139                 goto out;
1140         }
1141
1142         bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1143         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1144         orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1145         memcpy(&new_key, &key, sizeof(new_key));
1146
1147         if (start == key.offset && end < extent_end) {
1148                 other_start = 0;
1149                 other_end = start;
1150                 if (extent_mergeable(leaf, path->slots[0] - 1,
1151                                      ino, bytenr, orig_offset,
1152                                      &other_start, &other_end)) {
1153                         new_key.offset = end;
1154                         btrfs_set_item_key_safe(fs_info, path, &new_key);
1155                         fi = btrfs_item_ptr(leaf, path->slots[0],
1156                                             struct btrfs_file_extent_item);
1157                         btrfs_set_file_extent_generation(leaf, fi,
1158                                                          trans->transid);
1159                         btrfs_set_file_extent_num_bytes(leaf, fi,
1160                                                         extent_end - end);
1161                         btrfs_set_file_extent_offset(leaf, fi,
1162                                                      end - orig_offset);
1163                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1164                                             struct btrfs_file_extent_item);
1165                         btrfs_set_file_extent_generation(leaf, fi,
1166                                                          trans->transid);
1167                         btrfs_set_file_extent_num_bytes(leaf, fi,
1168                                                         end - other_start);
1169                         btrfs_mark_buffer_dirty(leaf);
1170                         goto out;
1171                 }
1172         }
1173
1174         if (start > key.offset && end == extent_end) {
1175                 other_start = end;
1176                 other_end = 0;
1177                 if (extent_mergeable(leaf, path->slots[0] + 1,
1178                                      ino, bytenr, orig_offset,
1179                                      &other_start, &other_end)) {
1180                         fi = btrfs_item_ptr(leaf, path->slots[0],
1181                                             struct btrfs_file_extent_item);
1182                         btrfs_set_file_extent_num_bytes(leaf, fi,
1183                                                         start - key.offset);
1184                         btrfs_set_file_extent_generation(leaf, fi,
1185                                                          trans->transid);
1186                         path->slots[0]++;
1187                         new_key.offset = start;
1188                         btrfs_set_item_key_safe(fs_info, path, &new_key);
1189
1190                         fi = btrfs_item_ptr(leaf, path->slots[0],
1191                                             struct btrfs_file_extent_item);
1192                         btrfs_set_file_extent_generation(leaf, fi,
1193                                                          trans->transid);
1194                         btrfs_set_file_extent_num_bytes(leaf, fi,
1195                                                         other_end - start);
1196                         btrfs_set_file_extent_offset(leaf, fi,
1197                                                      start - orig_offset);
1198                         btrfs_mark_buffer_dirty(leaf);
1199                         goto out;
1200                 }
1201         }
1202
1203         while (start > key.offset || end < extent_end) {
1204                 if (key.offset == start)
1205                         split = end;
1206
1207                 new_key.offset = split;
1208                 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1209                 if (ret == -EAGAIN) {
1210                         btrfs_release_path(path);
1211                         goto again;
1212                 }
1213                 if (ret < 0) {
1214                         btrfs_abort_transaction(trans, ret);
1215                         goto out;
1216                 }
1217
1218                 leaf = path->nodes[0];
1219                 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1220                                     struct btrfs_file_extent_item);
1221                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1222                 btrfs_set_file_extent_num_bytes(leaf, fi,
1223                                                 split - key.offset);
1224
1225                 fi = btrfs_item_ptr(leaf, path->slots[0],
1226                                     struct btrfs_file_extent_item);
1227
1228                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1229                 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1230                 btrfs_set_file_extent_num_bytes(leaf, fi,
1231                                                 extent_end - split);
1232                 btrfs_mark_buffer_dirty(leaf);
1233
1234                 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, bytenr,
1235                                        num_bytes, 0);
1236                 btrfs_init_data_ref(&ref, root->root_key.objectid, ino,
1237                                     orig_offset, 0, false);
1238                 ret = btrfs_inc_extent_ref(trans, &ref);
1239                 if (ret) {
1240                         btrfs_abort_transaction(trans, ret);
1241                         goto out;
1242                 }
1243
1244                 if (split == start) {
1245                         key.offset = start;
1246                 } else {
1247                         if (start != key.offset) {
1248                                 ret = -EINVAL;
1249                                 btrfs_abort_transaction(trans, ret);
1250                                 goto out;
1251                         }
1252                         path->slots[0]--;
1253                         extent_end = end;
1254                 }
1255                 recow = 1;
1256         }
1257
1258         other_start = end;
1259         other_end = 0;
1260         btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
1261                                num_bytes, 0);
1262         btrfs_init_data_ref(&ref, root->root_key.objectid, ino, orig_offset,
1263                             0, false);
1264         if (extent_mergeable(leaf, path->slots[0] + 1,
1265                              ino, bytenr, orig_offset,
1266                              &other_start, &other_end)) {
1267                 if (recow) {
1268                         btrfs_release_path(path);
1269                         goto again;
1270                 }
1271                 extent_end = other_end;
1272                 del_slot = path->slots[0] + 1;
1273                 del_nr++;
1274                 ret = btrfs_free_extent(trans, &ref);
1275                 if (ret) {
1276                         btrfs_abort_transaction(trans, ret);
1277                         goto out;
1278                 }
1279         }
1280         other_start = 0;
1281         other_end = start;
1282         if (extent_mergeable(leaf, path->slots[0] - 1,
1283                              ino, bytenr, orig_offset,
1284                              &other_start, &other_end)) {
1285                 if (recow) {
1286                         btrfs_release_path(path);
1287                         goto again;
1288                 }
1289                 key.offset = other_start;
1290                 del_slot = path->slots[0];
1291                 del_nr++;
1292                 ret = btrfs_free_extent(trans, &ref);
1293                 if (ret) {
1294                         btrfs_abort_transaction(trans, ret);
1295                         goto out;
1296                 }
1297         }
1298         if (del_nr == 0) {
1299                 fi = btrfs_item_ptr(leaf, path->slots[0],
1300                            struct btrfs_file_extent_item);
1301                 btrfs_set_file_extent_type(leaf, fi,
1302                                            BTRFS_FILE_EXTENT_REG);
1303                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1304                 btrfs_mark_buffer_dirty(leaf);
1305         } else {
1306                 fi = btrfs_item_ptr(leaf, del_slot - 1,
1307                            struct btrfs_file_extent_item);
1308                 btrfs_set_file_extent_type(leaf, fi,
1309                                            BTRFS_FILE_EXTENT_REG);
1310                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1311                 btrfs_set_file_extent_num_bytes(leaf, fi,
1312                                                 extent_end - key.offset);
1313                 btrfs_mark_buffer_dirty(leaf);
1314
1315                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1316                 if (ret < 0) {
1317                         btrfs_abort_transaction(trans, ret);
1318                         goto out;
1319                 }
1320         }
1321 out:
1322         btrfs_free_path(path);
1323         return ret;
1324 }
1325
1326 /*
1327  * on error we return an unlocked page and the error value
1328  * on success we return a locked page and 0
1329  */
1330 static int prepare_uptodate_page(struct inode *inode,
1331                                  struct page *page, u64 pos,
1332                                  bool force_uptodate)
1333 {
1334         int ret = 0;
1335
1336         if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1337             !PageUptodate(page)) {
1338                 ret = btrfs_readpage(NULL, page);
1339                 if (ret)
1340                         return ret;
1341                 lock_page(page);
1342                 if (!PageUptodate(page)) {
1343                         unlock_page(page);
1344                         return -EIO;
1345                 }
1346
1347                 /*
1348                  * Since btrfs_readpage() will unlock the page before it
1349                  * returns, there is a window where btrfs_releasepage() can be
1350                  * called to release the page.  Here we check both inode
1351                  * mapping and PagePrivate() to make sure the page was not
1352                  * released.
1353                  *
1354                  * The private flag check is essential for subpage as we need
1355                  * to store extra bitmap using page->private.
1356                  */
1357                 if (page->mapping != inode->i_mapping || !PagePrivate(page)) {
1358                         unlock_page(page);
1359                         return -EAGAIN;
1360                 }
1361         }
1362         return 0;
1363 }
1364
1365 /*
1366  * this just gets pages into the page cache and locks them down.
1367  */
1368 static noinline int prepare_pages(struct inode *inode, struct page **pages,
1369                                   size_t num_pages, loff_t pos,
1370                                   size_t write_bytes, bool force_uptodate)
1371 {
1372         int i;
1373         unsigned long index = pos >> PAGE_SHIFT;
1374         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1375         int err = 0;
1376         int faili;
1377
1378         for (i = 0; i < num_pages; i++) {
1379 again:
1380                 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1381                                                mask | __GFP_WRITE);
1382                 if (!pages[i]) {
1383                         faili = i - 1;
1384                         err = -ENOMEM;
1385                         goto fail;
1386                 }
1387
1388                 err = set_page_extent_mapped(pages[i]);
1389                 if (err < 0) {
1390                         faili = i;
1391                         goto fail;
1392                 }
1393
1394                 if (i == 0)
1395                         err = prepare_uptodate_page(inode, pages[i], pos,
1396                                                     force_uptodate);
1397                 if (!err && i == num_pages - 1)
1398                         err = prepare_uptodate_page(inode, pages[i],
1399                                                     pos + write_bytes, false);
1400                 if (err) {
1401                         put_page(pages[i]);
1402                         if (err == -EAGAIN) {
1403                                 err = 0;
1404                                 goto again;
1405                         }
1406                         faili = i - 1;
1407                         goto fail;
1408                 }
1409                 wait_on_page_writeback(pages[i]);
1410         }
1411
1412         return 0;
1413 fail:
1414         while (faili >= 0) {
1415                 unlock_page(pages[faili]);
1416                 put_page(pages[faili]);
1417                 faili--;
1418         }
1419         return err;
1420
1421 }
1422
1423 /*
1424  * This function locks the extent and properly waits for data=ordered extents
1425  * to finish before allowing the pages to be modified if need.
1426  *
1427  * The return value:
1428  * 1 - the extent is locked
1429  * 0 - the extent is not locked, and everything is OK
1430  * -EAGAIN - need re-prepare the pages
1431  * the other < 0 number - Something wrong happens
1432  */
1433 static noinline int
1434 lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct page **pages,
1435                                 size_t num_pages, loff_t pos,
1436                                 size_t write_bytes,
1437                                 u64 *lockstart, u64 *lockend,
1438                                 struct extent_state **cached_state)
1439 {
1440         struct btrfs_fs_info *fs_info = inode->root->fs_info;
1441         u64 start_pos;
1442         u64 last_pos;
1443         int i;
1444         int ret = 0;
1445
1446         start_pos = round_down(pos, fs_info->sectorsize);
1447         last_pos = round_up(pos + write_bytes, fs_info->sectorsize) - 1;
1448
1449         if (start_pos < inode->vfs_inode.i_size) {
1450                 struct btrfs_ordered_extent *ordered;
1451
1452                 lock_extent_bits(&inode->io_tree, start_pos, last_pos,
1453                                 cached_state);
1454                 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1455                                                      last_pos - start_pos + 1);
1456                 if (ordered &&
1457                     ordered->file_offset + ordered->num_bytes > start_pos &&
1458                     ordered->file_offset <= last_pos) {
1459                         unlock_extent_cached(&inode->io_tree, start_pos,
1460                                         last_pos, cached_state);
1461                         for (i = 0; i < num_pages; i++) {
1462                                 unlock_page(pages[i]);
1463                                 put_page(pages[i]);
1464                         }
1465                         btrfs_start_ordered_extent(ordered, 1);
1466                         btrfs_put_ordered_extent(ordered);
1467                         return -EAGAIN;
1468                 }
1469                 if (ordered)
1470                         btrfs_put_ordered_extent(ordered);
1471
1472                 *lockstart = start_pos;
1473                 *lockend = last_pos;
1474                 ret = 1;
1475         }
1476
1477         /*
1478          * We should be called after prepare_pages() which should have locked
1479          * all pages in the range.
1480          */
1481         for (i = 0; i < num_pages; i++)
1482                 WARN_ON(!PageLocked(pages[i]));
1483
1484         return ret;
1485 }
1486
1487 static int check_can_nocow(struct btrfs_inode *inode, loff_t pos,
1488                            size_t *write_bytes, bool nowait)
1489 {
1490         struct btrfs_fs_info *fs_info = inode->root->fs_info;
1491         struct btrfs_root *root = inode->root;
1492         u64 lockstart, lockend;
1493         u64 num_bytes;
1494         int ret;
1495
1496         if (!(inode->flags & (BTRFS_INODE_NODATACOW | BTRFS_INODE_PREALLOC)))
1497                 return 0;
1498
1499         if (!nowait && !btrfs_drew_try_write_lock(&root->snapshot_lock))
1500                 return -EAGAIN;
1501
1502         lockstart = round_down(pos, fs_info->sectorsize);
1503         lockend = round_up(pos + *write_bytes,
1504                            fs_info->sectorsize) - 1;
1505         num_bytes = lockend - lockstart + 1;
1506
1507         if (nowait) {
1508                 struct btrfs_ordered_extent *ordered;
1509
1510                 if (!try_lock_extent(&inode->io_tree, lockstart, lockend))
1511                         return -EAGAIN;
1512
1513                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
1514                                                      num_bytes);
1515                 if (ordered) {
1516                         btrfs_put_ordered_extent(ordered);
1517                         ret = -EAGAIN;
1518                         goto out_unlock;
1519                 }
1520         } else {
1521                 btrfs_lock_and_flush_ordered_range(inode, lockstart,
1522                                                    lockend, NULL);
1523         }
1524
1525         ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1526                         NULL, NULL, NULL, false);
1527         if (ret <= 0) {
1528                 ret = 0;
1529                 if (!nowait)
1530                         btrfs_drew_write_unlock(&root->snapshot_lock);
1531         } else {
1532                 *write_bytes = min_t(size_t, *write_bytes ,
1533                                      num_bytes - pos + lockstart);
1534         }
1535 out_unlock:
1536         unlock_extent(&inode->io_tree, lockstart, lockend);
1537
1538         return ret;
1539 }
1540
1541 static int check_nocow_nolock(struct btrfs_inode *inode, loff_t pos,
1542                               size_t *write_bytes)
1543 {
1544         return check_can_nocow(inode, pos, write_bytes, true);
1545 }
1546
1547 /*
1548  * Check if we can do nocow write into the range [@pos, @pos + @write_bytes)
1549  *
1550  * @pos:         File offset
1551  * @write_bytes: The length to write, will be updated to the nocow writeable
1552  *               range
1553  *
1554  * This function will flush ordered extents in the range to ensure proper
1555  * nocow checks.
1556  *
1557  * Return:
1558  * >0           and update @write_bytes if we can do nocow write
1559  *  0           if we can't do nocow write
1560  * -EAGAIN      if we can't get the needed lock or there are ordered extents
1561  *              for * (nowait == true) case
1562  * <0           if other error happened
1563  *
1564  * NOTE: Callers need to release the lock by btrfs_check_nocow_unlock().
1565  */
1566 int btrfs_check_nocow_lock(struct btrfs_inode *inode, loff_t pos,
1567                            size_t *write_bytes)
1568 {
1569         return check_can_nocow(inode, pos, write_bytes, false);
1570 }
1571
1572 void btrfs_check_nocow_unlock(struct btrfs_inode *inode)
1573 {
1574         btrfs_drew_write_unlock(&inode->root->snapshot_lock);
1575 }
1576
1577 static void update_time_for_write(struct inode *inode)
1578 {
1579         struct timespec64 now;
1580
1581         if (IS_NOCMTIME(inode))
1582                 return;
1583
1584         now = current_time(inode);
1585         if (!timespec64_equal(&inode->i_mtime, &now))
1586                 inode->i_mtime = now;
1587
1588         if (!timespec64_equal(&inode->i_ctime, &now))
1589                 inode->i_ctime = now;
1590
1591         if (IS_I_VERSION(inode))
1592                 inode_inc_iversion(inode);
1593 }
1594
1595 static int btrfs_write_check(struct kiocb *iocb, struct iov_iter *from,
1596                              size_t count)
1597 {
1598         struct file *file = iocb->ki_filp;
1599         struct inode *inode = file_inode(file);
1600         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1601         loff_t pos = iocb->ki_pos;
1602         int ret;
1603         loff_t oldsize;
1604         loff_t start_pos;
1605
1606         if (iocb->ki_flags & IOCB_NOWAIT) {
1607                 size_t nocow_bytes = count;
1608
1609                 /* We will allocate space in case nodatacow is not set, so bail */
1610                 if (check_nocow_nolock(BTRFS_I(inode), pos, &nocow_bytes) <= 0)
1611                         return -EAGAIN;
1612                 /*
1613                  * There are holes in the range or parts of the range that must
1614                  * be COWed (shared extents, RO block groups, etc), so just bail
1615                  * out.
1616                  */
1617                 if (nocow_bytes < count)
1618                         return -EAGAIN;
1619         }
1620
1621         current->backing_dev_info = inode_to_bdi(inode);
1622         ret = file_remove_privs(file);
1623         if (ret)
1624                 return ret;
1625
1626         /*
1627          * We reserve space for updating the inode when we reserve space for the
1628          * extent we are going to write, so we will enospc out there.  We don't
1629          * need to start yet another transaction to update the inode as we will
1630          * update the inode when we finish writing whatever data we write.
1631          */
1632         update_time_for_write(inode);
1633
1634         start_pos = round_down(pos, fs_info->sectorsize);
1635         oldsize = i_size_read(inode);
1636         if (start_pos > oldsize) {
1637                 /* Expand hole size to cover write data, preventing empty gap */
1638                 loff_t end_pos = round_up(pos + count, fs_info->sectorsize);
1639
1640                 ret = btrfs_cont_expand(BTRFS_I(inode), oldsize, end_pos);
1641                 if (ret) {
1642                         current->backing_dev_info = NULL;
1643                         return ret;
1644                 }
1645         }
1646
1647         return 0;
1648 }
1649
1650 static noinline ssize_t btrfs_buffered_write(struct kiocb *iocb,
1651                                                struct iov_iter *i)
1652 {
1653         struct file *file = iocb->ki_filp;
1654         loff_t pos;
1655         struct inode *inode = file_inode(file);
1656         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1657         struct page **pages = NULL;
1658         struct extent_changeset *data_reserved = NULL;
1659         u64 release_bytes = 0;
1660         u64 lockstart;
1661         u64 lockend;
1662         size_t num_written = 0;
1663         int nrptrs;
1664         ssize_t ret;
1665         bool only_release_metadata = false;
1666         bool force_page_uptodate = false;
1667         loff_t old_isize = i_size_read(inode);
1668         unsigned int ilock_flags = 0;
1669
1670         if (iocb->ki_flags & IOCB_NOWAIT)
1671                 ilock_flags |= BTRFS_ILOCK_TRY;
1672
1673         ret = btrfs_inode_lock(inode, ilock_flags);
1674         if (ret < 0)
1675                 return ret;
1676
1677         ret = generic_write_checks(iocb, i);
1678         if (ret <= 0)
1679                 goto out;
1680
1681         ret = btrfs_write_check(iocb, i, ret);
1682         if (ret < 0)
1683                 goto out;
1684
1685         pos = iocb->ki_pos;
1686         nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1687                         PAGE_SIZE / (sizeof(struct page *)));
1688         nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1689         nrptrs = max(nrptrs, 8);
1690         pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1691         if (!pages) {
1692                 ret = -ENOMEM;
1693                 goto out;
1694         }
1695
1696         while (iov_iter_count(i) > 0) {
1697                 struct extent_state *cached_state = NULL;
1698                 size_t offset = offset_in_page(pos);
1699                 size_t sector_offset;
1700                 size_t write_bytes = min(iov_iter_count(i),
1701                                          nrptrs * (size_t)PAGE_SIZE -
1702                                          offset);
1703                 size_t num_pages;
1704                 size_t reserve_bytes;
1705                 size_t dirty_pages;
1706                 size_t copied;
1707                 size_t dirty_sectors;
1708                 size_t num_sectors;
1709                 int extents_locked;
1710
1711                 /*
1712                  * Fault pages before locking them in prepare_pages
1713                  * to avoid recursive lock
1714                  */
1715                 if (unlikely(fault_in_iov_iter_readable(i, write_bytes))) {
1716                         ret = -EFAULT;
1717                         break;
1718                 }
1719
1720                 only_release_metadata = false;
1721                 sector_offset = pos & (fs_info->sectorsize - 1);
1722
1723                 extent_changeset_release(data_reserved);
1724                 ret = btrfs_check_data_free_space(BTRFS_I(inode),
1725                                                   &data_reserved, pos,
1726                                                   write_bytes);
1727                 if (ret < 0) {
1728                         /*
1729                          * If we don't have to COW at the offset, reserve
1730                          * metadata only. write_bytes may get smaller than
1731                          * requested here.
1732                          */
1733                         if (btrfs_check_nocow_lock(BTRFS_I(inode), pos,
1734                                                    &write_bytes) > 0)
1735                                 only_release_metadata = true;
1736                         else
1737                                 break;
1738                 }
1739
1740                 num_pages = DIV_ROUND_UP(write_bytes + offset, PAGE_SIZE);
1741                 WARN_ON(num_pages > nrptrs);
1742                 reserve_bytes = round_up(write_bytes + sector_offset,
1743                                          fs_info->sectorsize);
1744                 WARN_ON(reserve_bytes == 0);
1745                 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1746                                 reserve_bytes);
1747                 if (ret) {
1748                         if (!only_release_metadata)
1749                                 btrfs_free_reserved_data_space(BTRFS_I(inode),
1750                                                 data_reserved, pos,
1751                                                 write_bytes);
1752                         else
1753                                 btrfs_check_nocow_unlock(BTRFS_I(inode));
1754                         break;
1755                 }
1756
1757                 release_bytes = reserve_bytes;
1758 again:
1759                 /*
1760                  * This is going to setup the pages array with the number of
1761                  * pages we want, so we don't really need to worry about the
1762                  * contents of pages from loop to loop
1763                  */
1764                 ret = prepare_pages(inode, pages, num_pages,
1765                                     pos, write_bytes,
1766                                     force_page_uptodate);
1767                 if (ret) {
1768                         btrfs_delalloc_release_extents(BTRFS_I(inode),
1769                                                        reserve_bytes);
1770                         break;
1771                 }
1772
1773                 extents_locked = lock_and_cleanup_extent_if_need(
1774                                 BTRFS_I(inode), pages,
1775                                 num_pages, pos, write_bytes, &lockstart,
1776                                 &lockend, &cached_state);
1777                 if (extents_locked < 0) {
1778                         if (extents_locked == -EAGAIN)
1779                                 goto again;
1780                         btrfs_delalloc_release_extents(BTRFS_I(inode),
1781                                                        reserve_bytes);
1782                         ret = extents_locked;
1783                         break;
1784                 }
1785
1786                 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1787
1788                 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1789                 dirty_sectors = round_up(copied + sector_offset,
1790                                         fs_info->sectorsize);
1791                 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1792
1793                 /*
1794                  * if we have trouble faulting in the pages, fall
1795                  * back to one page at a time
1796                  */
1797                 if (copied < write_bytes)
1798                         nrptrs = 1;
1799
1800                 if (copied == 0) {
1801                         force_page_uptodate = true;
1802                         dirty_sectors = 0;
1803                         dirty_pages = 0;
1804                 } else {
1805                         force_page_uptodate = false;
1806                         dirty_pages = DIV_ROUND_UP(copied + offset,
1807                                                    PAGE_SIZE);
1808                 }
1809
1810                 if (num_sectors > dirty_sectors) {
1811                         /* release everything except the sectors we dirtied */
1812                         release_bytes -= dirty_sectors << fs_info->sectorsize_bits;
1813                         if (only_release_metadata) {
1814                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1815                                                         release_bytes, true);
1816                         } else {
1817                                 u64 __pos;
1818
1819                                 __pos = round_down(pos,
1820                                                    fs_info->sectorsize) +
1821                                         (dirty_pages << PAGE_SHIFT);
1822                                 btrfs_delalloc_release_space(BTRFS_I(inode),
1823                                                 data_reserved, __pos,
1824                                                 release_bytes, true);
1825                         }
1826                 }
1827
1828                 release_bytes = round_up(copied + sector_offset,
1829                                         fs_info->sectorsize);
1830
1831                 ret = btrfs_dirty_pages(BTRFS_I(inode), pages,
1832                                         dirty_pages, pos, copied,
1833                                         &cached_state, only_release_metadata);
1834
1835                 /*
1836                  * If we have not locked the extent range, because the range's
1837                  * start offset is >= i_size, we might still have a non-NULL
1838                  * cached extent state, acquired while marking the extent range
1839                  * as delalloc through btrfs_dirty_pages(). Therefore free any
1840                  * possible cached extent state to avoid a memory leak.
1841                  */
1842                 if (extents_locked)
1843                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1844                                              lockstart, lockend, &cached_state);
1845                 else
1846                         free_extent_state(cached_state);
1847
1848                 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1849                 if (ret) {
1850                         btrfs_drop_pages(pages, num_pages);
1851                         break;
1852                 }
1853
1854                 release_bytes = 0;
1855                 if (only_release_metadata)
1856                         btrfs_check_nocow_unlock(BTRFS_I(inode));
1857
1858                 btrfs_drop_pages(pages, num_pages);
1859
1860                 cond_resched();
1861
1862                 balance_dirty_pages_ratelimited(inode->i_mapping);
1863
1864                 pos += copied;
1865                 num_written += copied;
1866         }
1867
1868         kfree(pages);
1869
1870         if (release_bytes) {
1871                 if (only_release_metadata) {
1872                         btrfs_check_nocow_unlock(BTRFS_I(inode));
1873                         btrfs_delalloc_release_metadata(BTRFS_I(inode),
1874                                         release_bytes, true);
1875                 } else {
1876                         btrfs_delalloc_release_space(BTRFS_I(inode),
1877                                         data_reserved,
1878                                         round_down(pos, fs_info->sectorsize),
1879                                         release_bytes, true);
1880                 }
1881         }
1882
1883         extent_changeset_free(data_reserved);
1884         if (num_written > 0) {
1885                 pagecache_isize_extended(inode, old_isize, iocb->ki_pos);
1886                 iocb->ki_pos += num_written;
1887         }
1888 out:
1889         btrfs_inode_unlock(inode, ilock_flags);
1890         return num_written ? num_written : ret;
1891 }
1892
1893 static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
1894                                const struct iov_iter *iter, loff_t offset)
1895 {
1896         const u32 blocksize_mask = fs_info->sectorsize - 1;
1897
1898         if (offset & blocksize_mask)
1899                 return -EINVAL;
1900
1901         if (iov_iter_alignment(iter) & blocksize_mask)
1902                 return -EINVAL;
1903
1904         return 0;
1905 }
1906
1907 static ssize_t btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1908 {
1909         const bool is_sync_write = (iocb->ki_flags & IOCB_DSYNC);
1910         struct file *file = iocb->ki_filp;
1911         struct inode *inode = file_inode(file);
1912         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1913         loff_t pos;
1914         ssize_t written = 0;
1915         ssize_t written_buffered;
1916         size_t prev_left = 0;
1917         loff_t endbyte;
1918         ssize_t err;
1919         unsigned int ilock_flags = 0;
1920
1921         if (iocb->ki_flags & IOCB_NOWAIT)
1922                 ilock_flags |= BTRFS_ILOCK_TRY;
1923
1924         /* If the write DIO is within EOF, use a shared lock */
1925         if (iocb->ki_pos + iov_iter_count(from) <= i_size_read(inode))
1926                 ilock_flags |= BTRFS_ILOCK_SHARED;
1927
1928 relock:
1929         err = btrfs_inode_lock(inode, ilock_flags);
1930         if (err < 0)
1931                 return err;
1932
1933         err = generic_write_checks(iocb, from);
1934         if (err <= 0) {
1935                 btrfs_inode_unlock(inode, ilock_flags);
1936                 return err;
1937         }
1938
1939         err = btrfs_write_check(iocb, from, err);
1940         if (err < 0) {
1941                 btrfs_inode_unlock(inode, ilock_flags);
1942                 goto out;
1943         }
1944
1945         pos = iocb->ki_pos;
1946         /*
1947          * Re-check since file size may have changed just before taking the
1948          * lock or pos may have changed because of O_APPEND in generic_write_check()
1949          */
1950         if ((ilock_flags & BTRFS_ILOCK_SHARED) &&
1951             pos + iov_iter_count(from) > i_size_read(inode)) {
1952                 btrfs_inode_unlock(inode, ilock_flags);
1953                 ilock_flags &= ~BTRFS_ILOCK_SHARED;
1954                 goto relock;
1955         }
1956
1957         if (check_direct_IO(fs_info, from, pos)) {
1958                 btrfs_inode_unlock(inode, ilock_flags);
1959                 goto buffered;
1960         }
1961
1962         /*
1963          * We remove IOCB_DSYNC so that we don't deadlock when iomap_dio_rw()
1964          * calls generic_write_sync() (through iomap_dio_complete()), because
1965          * that results in calling fsync (btrfs_sync_file()) which will try to
1966          * lock the inode in exclusive/write mode.
1967          */
1968         if (is_sync_write)
1969                 iocb->ki_flags &= ~IOCB_DSYNC;
1970
1971         /*
1972          * The iov_iter can be mapped to the same file range we are writing to.
1973          * If that's the case, then we will deadlock in the iomap code, because
1974          * it first calls our callback btrfs_dio_iomap_begin(), which will create
1975          * an ordered extent, and after that it will fault in the pages that the
1976          * iov_iter refers to. During the fault in we end up in the readahead
1977          * pages code (starting at btrfs_readahead()), which will lock the range,
1978          * find that ordered extent and then wait for it to complete (at
1979          * btrfs_lock_and_flush_ordered_range()), resulting in a deadlock since
1980          * obviously the ordered extent can never complete as we didn't submit
1981          * yet the respective bio(s). This always happens when the buffer is
1982          * memory mapped to the same file range, since the iomap DIO code always
1983          * invalidates pages in the target file range (after starting and waiting
1984          * for any writeback).
1985          *
1986          * So here we disable page faults in the iov_iter and then retry if we
1987          * got -EFAULT, faulting in the pages before the retry.
1988          */
1989 again:
1990         from->nofault = true;
1991         err = iomap_dio_rw(iocb, from, &btrfs_dio_iomap_ops, &btrfs_dio_ops,
1992                            IOMAP_DIO_PARTIAL, written);
1993         from->nofault = false;
1994
1995         /* No increment (+=) because iomap returns a cumulative value. */
1996         if (err > 0)
1997                 written = err;
1998
1999         if (iov_iter_count(from) > 0 && (err == -EFAULT || err > 0)) {
2000                 const size_t left = iov_iter_count(from);
2001                 /*
2002                  * We have more data left to write. Try to fault in as many as
2003                  * possible of the remainder pages and retry. We do this without
2004                  * releasing and locking again the inode, to prevent races with
2005                  * truncate.
2006                  *
2007                  * Also, in case the iov refers to pages in the file range of the
2008                  * file we want to write to (due to a mmap), we could enter an
2009                  * infinite loop if we retry after faulting the pages in, since
2010                  * iomap will invalidate any pages in the range early on, before
2011                  * it tries to fault in the pages of the iov. So we keep track of
2012                  * how much was left of iov in the previous EFAULT and fallback
2013                  * to buffered IO in case we haven't made any progress.
2014                  */
2015                 if (left == prev_left) {
2016                         err = -ENOTBLK;
2017                 } else {
2018                         fault_in_iov_iter_readable(from, left);
2019                         prev_left = left;
2020                         goto again;
2021                 }
2022         }
2023
2024         btrfs_inode_unlock(inode, ilock_flags);
2025
2026         /*
2027          * Add back IOCB_DSYNC. Our caller, btrfs_file_write_iter(), will do
2028          * the fsync (call generic_write_sync()).
2029          */
2030         if (is_sync_write)
2031                 iocb->ki_flags |= IOCB_DSYNC;
2032
2033         /* If 'err' is -ENOTBLK then it means we must fallback to buffered IO. */
2034         if ((err < 0 && err != -ENOTBLK) || !iov_iter_count(from))
2035                 goto out;
2036
2037 buffered:
2038         pos = iocb->ki_pos;
2039         written_buffered = btrfs_buffered_write(iocb, from);
2040         if (written_buffered < 0) {
2041                 err = written_buffered;
2042                 goto out;
2043         }
2044         /*
2045          * Ensure all data is persisted. We want the next direct IO read to be
2046          * able to read what was just written.
2047          */
2048         endbyte = pos + written_buffered - 1;
2049         err = btrfs_fdatawrite_range(inode, pos, endbyte);
2050         if (err)
2051                 goto out;
2052         err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
2053         if (err)
2054                 goto out;
2055         written += written_buffered;
2056         iocb->ki_pos = pos + written_buffered;
2057         invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
2058                                  endbyte >> PAGE_SHIFT);
2059 out:
2060         return err < 0 ? err : written;
2061 }
2062
2063 static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
2064                                     struct iov_iter *from)
2065 {
2066         struct file *file = iocb->ki_filp;
2067         struct btrfs_inode *inode = BTRFS_I(file_inode(file));
2068         ssize_t num_written = 0;
2069         const bool sync = iocb->ki_flags & IOCB_DSYNC;
2070
2071         /*
2072          * If the fs flips readonly due to some impossible error, although we
2073          * have opened a file as writable, we have to stop this write operation
2074          * to ensure consistency.
2075          */
2076         if (test_bit(BTRFS_FS_STATE_ERROR, &inode->root->fs_info->fs_state))
2077                 return -EROFS;
2078
2079         if (!(iocb->ki_flags & IOCB_DIRECT) &&
2080             (iocb->ki_flags & IOCB_NOWAIT))
2081                 return -EOPNOTSUPP;
2082
2083         if (sync)
2084                 atomic_inc(&inode->sync_writers);
2085
2086         if (iocb->ki_flags & IOCB_DIRECT)
2087                 num_written = btrfs_direct_write(iocb, from);
2088         else
2089                 num_written = btrfs_buffered_write(iocb, from);
2090
2091         btrfs_set_inode_last_sub_trans(inode);
2092
2093         if (num_written > 0)
2094                 num_written = generic_write_sync(iocb, num_written);
2095
2096         if (sync)
2097                 atomic_dec(&inode->sync_writers);
2098
2099         current->backing_dev_info = NULL;
2100         return num_written;
2101 }
2102
2103 int btrfs_release_file(struct inode *inode, struct file *filp)
2104 {
2105         struct btrfs_file_private *private = filp->private_data;
2106
2107         if (private && private->filldir_buf)
2108                 kfree(private->filldir_buf);
2109         kfree(private);
2110         filp->private_data = NULL;
2111
2112         /*
2113          * Set by setattr when we are about to truncate a file from a non-zero
2114          * size to a zero size.  This tries to flush down new bytes that may
2115          * have been written if the application were using truncate to replace
2116          * a file in place.
2117          */
2118         if (test_and_clear_bit(BTRFS_INODE_FLUSH_ON_CLOSE,
2119                                &BTRFS_I(inode)->runtime_flags))
2120                         filemap_flush(inode->i_mapping);
2121         return 0;
2122 }
2123
2124 static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
2125 {
2126         int ret;
2127         struct blk_plug plug;
2128
2129         /*
2130          * This is only called in fsync, which would do synchronous writes, so
2131          * a plug can merge adjacent IOs as much as possible.  Esp. in case of
2132          * multiple disks using raid profile, a large IO can be split to
2133          * several segments of stripe length (currently 64K).
2134          */
2135         blk_start_plug(&plug);
2136         atomic_inc(&BTRFS_I(inode)->sync_writers);
2137         ret = btrfs_fdatawrite_range(inode, start, end);
2138         atomic_dec(&BTRFS_I(inode)->sync_writers);
2139         blk_finish_plug(&plug);
2140
2141         return ret;
2142 }
2143
2144 static inline bool skip_inode_logging(const struct btrfs_log_ctx *ctx)
2145 {
2146         struct btrfs_inode *inode = BTRFS_I(ctx->inode);
2147         struct btrfs_fs_info *fs_info = inode->root->fs_info;
2148
2149         if (btrfs_inode_in_log(inode, fs_info->generation) &&
2150             list_empty(&ctx->ordered_extents))
2151                 return true;
2152
2153         /*
2154          * If we are doing a fast fsync we can not bail out if the inode's
2155          * last_trans is <= then the last committed transaction, because we only
2156          * update the last_trans of the inode during ordered extent completion,
2157          * and for a fast fsync we don't wait for that, we only wait for the
2158          * writeback to complete.
2159          */
2160         if (inode->last_trans <= fs_info->last_trans_committed &&
2161             (test_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags) ||
2162              list_empty(&ctx->ordered_extents)))
2163                 return true;
2164
2165         return false;
2166 }
2167
2168 /*
2169  * fsync call for both files and directories.  This logs the inode into
2170  * the tree log instead of forcing full commits whenever possible.
2171  *
2172  * It needs to call filemap_fdatawait so that all ordered extent updates are
2173  * in the metadata btree are up to date for copying to the log.
2174  *
2175  * It drops the inode mutex before doing the tree log commit.  This is an
2176  * important optimization for directories because holding the mutex prevents
2177  * new operations on the dir while we write to disk.
2178  */
2179 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
2180 {
2181         struct dentry *dentry = file_dentry(file);
2182         struct inode *inode = d_inode(dentry);
2183         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2184         struct btrfs_root *root = BTRFS_I(inode)->root;
2185         struct btrfs_trans_handle *trans;
2186         struct btrfs_log_ctx ctx;
2187         int ret = 0, err;
2188         u64 len;
2189         bool full_sync;
2190
2191         trace_btrfs_sync_file(file, datasync);
2192
2193         btrfs_init_log_ctx(&ctx, inode);
2194
2195         /*
2196          * Always set the range to a full range, otherwise we can get into
2197          * several problems, from missing file extent items to represent holes
2198          * when not using the NO_HOLES feature, to log tree corruption due to
2199          * races between hole detection during logging and completion of ordered
2200          * extents outside the range, to missing checksums due to ordered extents
2201          * for which we flushed only a subset of their pages.
2202          */
2203         start = 0;
2204         end = LLONG_MAX;
2205         len = (u64)LLONG_MAX + 1;
2206
2207         /*
2208          * We write the dirty pages in the range and wait until they complete
2209          * out of the ->i_mutex. If so, we can flush the dirty pages by
2210          * multi-task, and make the performance up.  See
2211          * btrfs_wait_ordered_range for an explanation of the ASYNC check.
2212          */
2213         ret = start_ordered_ops(inode, start, end);
2214         if (ret)
2215                 goto out;
2216
2217         btrfs_inode_lock(inode, BTRFS_ILOCK_MMAP);
2218
2219         atomic_inc(&root->log_batch);
2220
2221         /*
2222          * Always check for the full sync flag while holding the inode's lock,
2223          * to avoid races with other tasks. The flag must be either set all the
2224          * time during logging or always off all the time while logging.
2225          */
2226         full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2227                              &BTRFS_I(inode)->runtime_flags);
2228
2229         /*
2230          * Before we acquired the inode's lock and the mmap lock, someone may
2231          * have dirtied more pages in the target range. We need to make sure
2232          * that writeback for any such pages does not start while we are logging
2233          * the inode, because if it does, any of the following might happen when
2234          * we are not doing a full inode sync:
2235          *
2236          * 1) We log an extent after its writeback finishes but before its
2237          *    checksums are added to the csum tree, leading to -EIO errors
2238          *    when attempting to read the extent after a log replay.
2239          *
2240          * 2) We can end up logging an extent before its writeback finishes.
2241          *    Therefore after the log replay we will have a file extent item
2242          *    pointing to an unwritten extent (and no data checksums as well).
2243          *
2244          * So trigger writeback for any eventual new dirty pages and then we
2245          * wait for all ordered extents to complete below.
2246          */
2247         ret = start_ordered_ops(inode, start, end);
2248         if (ret) {
2249                 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
2250                 goto out;
2251         }
2252
2253         /*
2254          * We have to do this here to avoid the priority inversion of waiting on
2255          * IO of a lower priority task while holding a transaction open.
2256          *
2257          * For a full fsync we wait for the ordered extents to complete while
2258          * for a fast fsync we wait just for writeback to complete, and then
2259          * attach the ordered extents to the transaction so that a transaction
2260          * commit waits for their completion, to avoid data loss if we fsync,
2261          * the current transaction commits before the ordered extents complete
2262          * and a power failure happens right after that.
2263          *
2264          * For zoned filesystem, if a write IO uses a ZONE_APPEND command, the
2265          * logical address recorded in the ordered extent may change. We need
2266          * to wait for the IO to stabilize the logical address.
2267          */
2268         if (full_sync || btrfs_is_zoned(fs_info)) {
2269                 ret = btrfs_wait_ordered_range(inode, start, len);
2270         } else {
2271                 /*
2272                  * Get our ordered extents as soon as possible to avoid doing
2273                  * checksum lookups in the csum tree, and use instead the
2274                  * checksums attached to the ordered extents.
2275                  */
2276                 btrfs_get_ordered_extents_for_logging(BTRFS_I(inode),
2277                                                       &ctx.ordered_extents);
2278                 ret = filemap_fdatawait_range(inode->i_mapping, start, end);
2279         }
2280
2281         if (ret)
2282                 goto out_release_extents;
2283
2284         atomic_inc(&root->log_batch);
2285
2286         smp_mb();
2287         if (skip_inode_logging(&ctx)) {
2288                 /*
2289                  * We've had everything committed since the last time we were
2290                  * modified so clear this flag in case it was set for whatever
2291                  * reason, it's no longer relevant.
2292                  */
2293                 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2294                           &BTRFS_I(inode)->runtime_flags);
2295                 /*
2296                  * An ordered extent might have started before and completed
2297                  * already with io errors, in which case the inode was not
2298                  * updated and we end up here. So check the inode's mapping
2299                  * for any errors that might have happened since we last
2300                  * checked called fsync.
2301                  */
2302                 ret = filemap_check_wb_err(inode->i_mapping, file->f_wb_err);
2303                 goto out_release_extents;
2304         }
2305
2306         /*
2307          * We use start here because we will need to wait on the IO to complete
2308          * in btrfs_sync_log, which could require joining a transaction (for
2309          * example checking cross references in the nocow path).  If we use join
2310          * here we could get into a situation where we're waiting on IO to
2311          * happen that is blocked on a transaction trying to commit.  With start
2312          * we inc the extwriter counter, so we wait for all extwriters to exit
2313          * before we start blocking joiners.  This comment is to keep somebody
2314          * from thinking they are super smart and changing this to
2315          * btrfs_join_transaction *cough*Josef*cough*.
2316          */
2317         trans = btrfs_start_transaction(root, 0);
2318         if (IS_ERR(trans)) {
2319                 ret = PTR_ERR(trans);
2320                 goto out_release_extents;
2321         }
2322         trans->in_fsync = true;
2323
2324         ret = btrfs_log_dentry_safe(trans, dentry, &ctx);
2325         btrfs_release_log_ctx_extents(&ctx);
2326         if (ret < 0) {
2327                 /* Fallthrough and commit/free transaction. */
2328                 ret = 1;
2329         }
2330
2331         /* we've logged all the items and now have a consistent
2332          * version of the file in the log.  It is possible that
2333          * someone will come in and modify the file, but that's
2334          * fine because the log is consistent on disk, and we
2335          * have references to all of the file's extents
2336          *
2337          * It is possible that someone will come in and log the
2338          * file again, but that will end up using the synchronization
2339          * inside btrfs_sync_log to keep things safe.
2340          */
2341         btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
2342
2343         if (ret == BTRFS_NO_LOG_SYNC) {
2344                 ret = btrfs_end_transaction(trans);
2345                 goto out;
2346         }
2347
2348         /* We successfully logged the inode, attempt to sync the log. */
2349         if (!ret) {
2350                 ret = btrfs_sync_log(trans, root, &ctx);
2351                 if (!ret) {
2352                         ret = btrfs_end_transaction(trans);
2353                         goto out;
2354                 }
2355         }
2356
2357         /*
2358          * At this point we need to commit the transaction because we had
2359          * btrfs_need_log_full_commit() or some other error.
2360          *
2361          * If we didn't do a full sync we have to stop the trans handle, wait on
2362          * the ordered extents, start it again and commit the transaction.  If
2363          * we attempt to wait on the ordered extents here we could deadlock with
2364          * something like fallocate() that is holding the extent lock trying to
2365          * start a transaction while some other thread is trying to commit the
2366          * transaction while we (fsync) are currently holding the transaction
2367          * open.
2368          */
2369         if (!full_sync) {
2370                 ret = btrfs_end_transaction(trans);
2371                 if (ret)
2372                         goto out;
2373                 ret = btrfs_wait_ordered_range(inode, start, len);
2374                 if (ret)
2375                         goto out;
2376
2377                 /*
2378                  * This is safe to use here because we're only interested in
2379                  * making sure the transaction that had the ordered extents is
2380                  * committed.  We aren't waiting on anything past this point,
2381                  * we're purely getting the transaction and committing it.
2382                  */
2383                 trans = btrfs_attach_transaction_barrier(root);
2384                 if (IS_ERR(trans)) {
2385                         ret = PTR_ERR(trans);
2386
2387                         /*
2388                          * We committed the transaction and there's no currently
2389                          * running transaction, this means everything we care
2390                          * about made it to disk and we are done.
2391                          */
2392                         if (ret == -ENOENT)
2393                                 ret = 0;
2394                         goto out;
2395                 }
2396         }
2397
2398         ret = btrfs_commit_transaction(trans);
2399 out:
2400         ASSERT(list_empty(&ctx.list));
2401         err = file_check_and_advance_wb_err(file);
2402         if (!ret)
2403                 ret = err;
2404         return ret > 0 ? -EIO : ret;
2405
2406 out_release_extents:
2407         btrfs_release_log_ctx_extents(&ctx);
2408         btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
2409         goto out;
2410 }
2411
2412 static const struct vm_operations_struct btrfs_file_vm_ops = {
2413         .fault          = filemap_fault,
2414         .map_pages      = filemap_map_pages,
2415         .page_mkwrite   = btrfs_page_mkwrite,
2416 };
2417
2418 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
2419 {
2420         struct address_space *mapping = filp->f_mapping;
2421
2422         if (!mapping->a_ops->readpage)
2423                 return -ENOEXEC;
2424
2425         file_accessed(filp);
2426         vma->vm_ops = &btrfs_file_vm_ops;
2427
2428         return 0;
2429 }
2430
2431 static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2432                           int slot, u64 start, u64 end)
2433 {
2434         struct btrfs_file_extent_item *fi;
2435         struct btrfs_key key;
2436
2437         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2438                 return 0;
2439
2440         btrfs_item_key_to_cpu(leaf, &key, slot);
2441         if (key.objectid != btrfs_ino(inode) ||
2442             key.type != BTRFS_EXTENT_DATA_KEY)
2443                 return 0;
2444
2445         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2446
2447         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2448                 return 0;
2449
2450         if (btrfs_file_extent_disk_bytenr(leaf, fi))
2451                 return 0;
2452
2453         if (key.offset == end)
2454                 return 1;
2455         if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2456                 return 1;
2457         return 0;
2458 }
2459
2460 static int fill_holes(struct btrfs_trans_handle *trans,
2461                 struct btrfs_inode *inode,
2462                 struct btrfs_path *path, u64 offset, u64 end)
2463 {
2464         struct btrfs_fs_info *fs_info = trans->fs_info;
2465         struct btrfs_root *root = inode->root;
2466         struct extent_buffer *leaf;
2467         struct btrfs_file_extent_item *fi;
2468         struct extent_map *hole_em;
2469         struct extent_map_tree *em_tree = &inode->extent_tree;
2470         struct btrfs_key key;
2471         int ret;
2472
2473         if (btrfs_fs_incompat(fs_info, NO_HOLES))
2474                 goto out;
2475
2476         key.objectid = btrfs_ino(inode);
2477         key.type = BTRFS_EXTENT_DATA_KEY;
2478         key.offset = offset;
2479
2480         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2481         if (ret <= 0) {
2482                 /*
2483                  * We should have dropped this offset, so if we find it then
2484                  * something has gone horribly wrong.
2485                  */
2486                 if (ret == 0)
2487                         ret = -EINVAL;
2488                 return ret;
2489         }
2490
2491         leaf = path->nodes[0];
2492         if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2493                 u64 num_bytes;
2494
2495                 path->slots[0]--;
2496                 fi = btrfs_item_ptr(leaf, path->slots[0],
2497                                     struct btrfs_file_extent_item);
2498                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2499                         end - offset;
2500                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2501                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2502                 btrfs_set_file_extent_offset(leaf, fi, 0);
2503                 btrfs_mark_buffer_dirty(leaf);
2504                 goto out;
2505         }
2506
2507         if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2508                 u64 num_bytes;
2509
2510                 key.offset = offset;
2511                 btrfs_set_item_key_safe(fs_info, path, &key);
2512                 fi = btrfs_item_ptr(leaf, path->slots[0],
2513                                     struct btrfs_file_extent_item);
2514                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2515                         offset;
2516                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2517                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2518                 btrfs_set_file_extent_offset(leaf, fi, 0);
2519                 btrfs_mark_buffer_dirty(leaf);
2520                 goto out;
2521         }
2522         btrfs_release_path(path);
2523
2524         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode),
2525                         offset, 0, 0, end - offset, 0, end - offset, 0, 0, 0);
2526         if (ret)
2527                 return ret;
2528
2529 out:
2530         btrfs_release_path(path);
2531
2532         hole_em = alloc_extent_map();
2533         if (!hole_em) {
2534                 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2535                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
2536         } else {
2537                 hole_em->start = offset;
2538                 hole_em->len = end - offset;
2539                 hole_em->ram_bytes = hole_em->len;
2540                 hole_em->orig_start = offset;
2541
2542                 hole_em->block_start = EXTENT_MAP_HOLE;
2543                 hole_em->block_len = 0;
2544                 hole_em->orig_block_len = 0;
2545                 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2546                 hole_em->generation = trans->transid;
2547
2548                 do {
2549                         btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2550                         write_lock(&em_tree->lock);
2551                         ret = add_extent_mapping(em_tree, hole_em, 1);
2552                         write_unlock(&em_tree->lock);
2553                 } while (ret == -EEXIST);
2554                 free_extent_map(hole_em);
2555                 if (ret)
2556                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2557                                         &inode->runtime_flags);
2558         }
2559
2560         return 0;
2561 }
2562
2563 /*
2564  * Find a hole extent on given inode and change start/len to the end of hole
2565  * extent.(hole/vacuum extent whose em->start <= start &&
2566  *         em->start + em->len > start)
2567  * When a hole extent is found, return 1 and modify start/len.
2568  */
2569 static int find_first_non_hole(struct btrfs_inode *inode, u64 *start, u64 *len)
2570 {
2571         struct btrfs_fs_info *fs_info = inode->root->fs_info;
2572         struct extent_map *em;
2573         int ret = 0;
2574
2575         em = btrfs_get_extent(inode, NULL, 0,
2576                               round_down(*start, fs_info->sectorsize),
2577                               round_up(*len, fs_info->sectorsize));
2578         if (IS_ERR(em))
2579                 return PTR_ERR(em);
2580
2581         /* Hole or vacuum extent(only exists in no-hole mode) */
2582         if (em->block_start == EXTENT_MAP_HOLE) {
2583                 ret = 1;
2584                 *len = em->start + em->len > *start + *len ?
2585                        0 : *start + *len - em->start - em->len;
2586                 *start = em->start + em->len;
2587         }
2588         free_extent_map(em);
2589         return ret;
2590 }
2591
2592 static int btrfs_punch_hole_lock_range(struct inode *inode,
2593                                        const u64 lockstart,
2594                                        const u64 lockend,
2595                                        struct extent_state **cached_state)
2596 {
2597         /*
2598          * For subpage case, if the range is not at page boundary, we could
2599          * have pages at the leading/tailing part of the range.
2600          * This could lead to dead loop since filemap_range_has_page()
2601          * will always return true.
2602          * So here we need to do extra page alignment for
2603          * filemap_range_has_page().
2604          */
2605         const u64 page_lockstart = round_up(lockstart, PAGE_SIZE);
2606         const u64 page_lockend = round_down(lockend + 1, PAGE_SIZE) - 1;
2607
2608         while (1) {
2609                 struct btrfs_ordered_extent *ordered;
2610                 int ret;
2611
2612                 truncate_pagecache_range(inode, lockstart, lockend);
2613
2614                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2615                                  cached_state);
2616                 ordered = btrfs_lookup_first_ordered_extent(BTRFS_I(inode),
2617                                                             lockend);
2618
2619                 /*
2620                  * We need to make sure we have no ordered extents in this range
2621                  * and nobody raced in and read a page in this range, if we did
2622                  * we need to try again.
2623                  */
2624                 if ((!ordered ||
2625                     (ordered->file_offset + ordered->num_bytes <= lockstart ||
2626                      ordered->file_offset > lockend)) &&
2627                      !filemap_range_has_page(inode->i_mapping,
2628                                              page_lockstart, page_lockend)) {
2629                         if (ordered)
2630                                 btrfs_put_ordered_extent(ordered);
2631                         break;
2632                 }
2633                 if (ordered)
2634                         btrfs_put_ordered_extent(ordered);
2635                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2636                                      lockend, cached_state);
2637                 ret = btrfs_wait_ordered_range(inode, lockstart,
2638                                                lockend - lockstart + 1);
2639                 if (ret)
2640                         return ret;
2641         }
2642         return 0;
2643 }
2644
2645 static int btrfs_insert_replace_extent(struct btrfs_trans_handle *trans,
2646                                      struct btrfs_inode *inode,
2647                                      struct btrfs_path *path,
2648                                      struct btrfs_replace_extent_info *extent_info,
2649                                      const u64 replace_len,
2650                                      const u64 bytes_to_drop)
2651 {
2652         struct btrfs_fs_info *fs_info = trans->fs_info;
2653         struct btrfs_root *root = inode->root;
2654         struct btrfs_file_extent_item *extent;
2655         struct extent_buffer *leaf;
2656         struct btrfs_key key;
2657         int slot;
2658         struct btrfs_ref ref = { 0 };
2659         int ret;
2660
2661         if (replace_len == 0)
2662                 return 0;
2663
2664         if (extent_info->disk_offset == 0 &&
2665             btrfs_fs_incompat(fs_info, NO_HOLES)) {
2666                 btrfs_update_inode_bytes(inode, 0, bytes_to_drop);
2667                 return 0;
2668         }
2669
2670         key.objectid = btrfs_ino(inode);
2671         key.type = BTRFS_EXTENT_DATA_KEY;
2672         key.offset = extent_info->file_offset;
2673         ret = btrfs_insert_empty_item(trans, root, path, &key,
2674                                       sizeof(struct btrfs_file_extent_item));
2675         if (ret)
2676                 return ret;
2677         leaf = path->nodes[0];
2678         slot = path->slots[0];
2679         write_extent_buffer(leaf, extent_info->extent_buf,
2680                             btrfs_item_ptr_offset(leaf, slot),
2681                             sizeof(struct btrfs_file_extent_item));
2682         extent = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2683         ASSERT(btrfs_file_extent_type(leaf, extent) != BTRFS_FILE_EXTENT_INLINE);
2684         btrfs_set_file_extent_offset(leaf, extent, extent_info->data_offset);
2685         btrfs_set_file_extent_num_bytes(leaf, extent, replace_len);
2686         if (extent_info->is_new_extent)
2687                 btrfs_set_file_extent_generation(leaf, extent, trans->transid);
2688         btrfs_mark_buffer_dirty(leaf);
2689         btrfs_release_path(path);
2690
2691         ret = btrfs_inode_set_file_extent_range(inode, extent_info->file_offset,
2692                                                 replace_len);
2693         if (ret)
2694                 return ret;
2695
2696         /* If it's a hole, nothing more needs to be done. */
2697         if (extent_info->disk_offset == 0) {
2698                 btrfs_update_inode_bytes(inode, 0, bytes_to_drop);
2699                 return 0;
2700         }
2701
2702         btrfs_update_inode_bytes(inode, replace_len, bytes_to_drop);
2703
2704         if (extent_info->is_new_extent && extent_info->insertions == 0) {
2705                 key.objectid = extent_info->disk_offset;
2706                 key.type = BTRFS_EXTENT_ITEM_KEY;
2707                 key.offset = extent_info->disk_len;
2708                 ret = btrfs_alloc_reserved_file_extent(trans, root,
2709                                                        btrfs_ino(inode),
2710                                                        extent_info->file_offset,
2711                                                        extent_info->qgroup_reserved,
2712                                                        &key);
2713         } else {
2714                 u64 ref_offset;
2715
2716                 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2717                                        extent_info->disk_offset,
2718                                        extent_info->disk_len, 0);
2719                 ref_offset = extent_info->file_offset - extent_info->data_offset;
2720                 btrfs_init_data_ref(&ref, root->root_key.objectid,
2721                                     btrfs_ino(inode), ref_offset, 0, false);
2722                 ret = btrfs_inc_extent_ref(trans, &ref);
2723         }
2724
2725         extent_info->insertions++;
2726
2727         return ret;
2728 }
2729
2730 /*
2731  * The respective range must have been previously locked, as well as the inode.
2732  * The end offset is inclusive (last byte of the range).
2733  * @extent_info is NULL for fallocate's hole punching and non-NULL when replacing
2734  * the file range with an extent.
2735  * When not punching a hole, we don't want to end up in a state where we dropped
2736  * extents without inserting a new one, so we must abort the transaction to avoid
2737  * a corruption.
2738  */
2739 int btrfs_replace_file_extents(struct btrfs_inode *inode,
2740                                struct btrfs_path *path, const u64 start,
2741                                const u64 end,
2742                                struct btrfs_replace_extent_info *extent_info,
2743                                struct btrfs_trans_handle **trans_out)
2744 {
2745         struct btrfs_drop_extents_args drop_args = { 0 };
2746         struct btrfs_root *root = inode->root;
2747         struct btrfs_fs_info *fs_info = root->fs_info;
2748         u64 min_size = btrfs_calc_insert_metadata_size(fs_info, 1);
2749         u64 ino_size = round_up(inode->vfs_inode.i_size, fs_info->sectorsize);
2750         struct btrfs_trans_handle *trans = NULL;
2751         struct btrfs_block_rsv *rsv;
2752         unsigned int rsv_count;
2753         u64 cur_offset;
2754         u64 len = end - start;
2755         int ret = 0;
2756
2757         if (end <= start)
2758                 return -EINVAL;
2759
2760         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2761         if (!rsv) {
2762                 ret = -ENOMEM;
2763                 goto out;
2764         }
2765         rsv->size = btrfs_calc_insert_metadata_size(fs_info, 1);
2766         rsv->failfast = 1;
2767
2768         /*
2769          * 1 - update the inode
2770          * 1 - removing the extents in the range
2771          * 1 - adding the hole extent if no_holes isn't set or if we are
2772          *     replacing the range with a new extent
2773          */
2774         if (!btrfs_fs_incompat(fs_info, NO_HOLES) || extent_info)
2775                 rsv_count = 3;
2776         else
2777                 rsv_count = 2;
2778
2779         trans = btrfs_start_transaction(root, rsv_count);
2780         if (IS_ERR(trans)) {
2781                 ret = PTR_ERR(trans);
2782                 trans = NULL;
2783                 goto out_free;
2784         }
2785
2786         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2787                                       min_size, false);
2788         BUG_ON(ret);
2789         trans->block_rsv = rsv;
2790
2791         cur_offset = start;
2792         drop_args.path = path;
2793         drop_args.end = end + 1;
2794         drop_args.drop_cache = true;
2795         while (cur_offset < end) {
2796                 drop_args.start = cur_offset;
2797                 ret = btrfs_drop_extents(trans, root, inode, &drop_args);
2798                 /* If we are punching a hole decrement the inode's byte count */
2799                 if (!extent_info)
2800                         btrfs_update_inode_bytes(inode, 0,
2801                                                  drop_args.bytes_found);
2802                 if (ret != -ENOSPC) {
2803                         /*
2804                          * The only time we don't want to abort is if we are
2805                          * attempting to clone a partial inline extent, in which
2806                          * case we'll get EOPNOTSUPP.  However if we aren't
2807                          * clone we need to abort no matter what, because if we
2808                          * got EOPNOTSUPP via prealloc then we messed up and
2809                          * need to abort.
2810                          */
2811                         if (ret &&
2812                             (ret != -EOPNOTSUPP ||
2813                              (extent_info && extent_info->is_new_extent)))
2814                                 btrfs_abort_transaction(trans, ret);
2815                         break;
2816                 }
2817
2818                 trans->block_rsv = &fs_info->trans_block_rsv;
2819
2820                 if (!extent_info && cur_offset < drop_args.drop_end &&
2821                     cur_offset < ino_size) {
2822                         ret = fill_holes(trans, inode, path, cur_offset,
2823                                          drop_args.drop_end);
2824                         if (ret) {
2825                                 /*
2826                                  * If we failed then we didn't insert our hole
2827                                  * entries for the area we dropped, so now the
2828                                  * fs is corrupted, so we must abort the
2829                                  * transaction.
2830                                  */
2831                                 btrfs_abort_transaction(trans, ret);
2832                                 break;
2833                         }
2834                 } else if (!extent_info && cur_offset < drop_args.drop_end) {
2835                         /*
2836                          * We are past the i_size here, but since we didn't
2837                          * insert holes we need to clear the mapped area so we
2838                          * know to not set disk_i_size in this area until a new
2839                          * file extent is inserted here.
2840                          */
2841                         ret = btrfs_inode_clear_file_extent_range(inode,
2842                                         cur_offset,
2843                                         drop_args.drop_end - cur_offset);
2844                         if (ret) {
2845                                 /*
2846                                  * We couldn't clear our area, so we could
2847                                  * presumably adjust up and corrupt the fs, so
2848                                  * we need to abort.
2849                                  */
2850                                 btrfs_abort_transaction(trans, ret);
2851                                 break;
2852                         }
2853                 }
2854
2855                 if (extent_info &&
2856                     drop_args.drop_end > extent_info->file_offset) {
2857                         u64 replace_len = drop_args.drop_end -
2858                                           extent_info->file_offset;
2859
2860                         ret = btrfs_insert_replace_extent(trans, inode, path,
2861                                         extent_info, replace_len,
2862                                         drop_args.bytes_found);
2863                         if (ret) {
2864                                 btrfs_abort_transaction(trans, ret);
2865                                 break;
2866                         }
2867                         extent_info->data_len -= replace_len;
2868                         extent_info->data_offset += replace_len;
2869                         extent_info->file_offset += replace_len;
2870                 }
2871
2872                 ret = btrfs_update_inode(trans, root, inode);
2873                 if (ret)
2874                         break;
2875
2876                 btrfs_end_transaction(trans);
2877                 btrfs_btree_balance_dirty(fs_info);
2878
2879                 trans = btrfs_start_transaction(root, rsv_count);
2880                 if (IS_ERR(trans)) {
2881                         ret = PTR_ERR(trans);
2882                         trans = NULL;
2883                         break;
2884                 }
2885
2886                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2887                                               rsv, min_size, false);
2888                 BUG_ON(ret);    /* shouldn't happen */
2889                 trans->block_rsv = rsv;
2890
2891                 cur_offset = drop_args.drop_end;
2892                 len = end - cur_offset;
2893                 if (!extent_info && len) {
2894                         ret = find_first_non_hole(inode, &cur_offset, &len);
2895                         if (unlikely(ret < 0))
2896                                 break;
2897                         if (ret && !len) {
2898                                 ret = 0;
2899                                 break;
2900                         }
2901                 }
2902         }
2903
2904         /*
2905          * If we were cloning, force the next fsync to be a full one since we
2906          * we replaced (or just dropped in the case of cloning holes when
2907          * NO_HOLES is enabled) file extent items and did not setup new extent
2908          * maps for the replacement extents (or holes).
2909          */
2910         if (extent_info && !extent_info->is_new_extent)
2911                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
2912
2913         if (ret)
2914                 goto out_trans;
2915
2916         trans->block_rsv = &fs_info->trans_block_rsv;
2917         /*
2918          * If we are using the NO_HOLES feature we might have had already an
2919          * hole that overlaps a part of the region [lockstart, lockend] and
2920          * ends at (or beyond) lockend. Since we have no file extent items to
2921          * represent holes, drop_end can be less than lockend and so we must
2922          * make sure we have an extent map representing the existing hole (the
2923          * call to __btrfs_drop_extents() might have dropped the existing extent
2924          * map representing the existing hole), otherwise the fast fsync path
2925          * will not record the existence of the hole region
2926          * [existing_hole_start, lockend].
2927          */
2928         if (drop_args.drop_end <= end)
2929                 drop_args.drop_end = end + 1;
2930         /*
2931          * Don't insert file hole extent item if it's for a range beyond eof
2932          * (because it's useless) or if it represents a 0 bytes range (when
2933          * cur_offset == drop_end).
2934          */
2935         if (!extent_info && cur_offset < ino_size &&
2936             cur_offset < drop_args.drop_end) {
2937                 ret = fill_holes(trans, inode, path, cur_offset,
2938                                  drop_args.drop_end);
2939                 if (ret) {
2940                         /* Same comment as above. */
2941                         btrfs_abort_transaction(trans, ret);
2942                         goto out_trans;
2943                 }
2944         } else if (!extent_info && cur_offset < drop_args.drop_end) {
2945                 /* See the comment in the loop above for the reasoning here. */
2946                 ret = btrfs_inode_clear_file_extent_range(inode, cur_offset,
2947                                         drop_args.drop_end - cur_offset);
2948                 if (ret) {
2949                         btrfs_abort_transaction(trans, ret);
2950                         goto out_trans;
2951                 }
2952
2953         }
2954         if (extent_info) {
2955                 ret = btrfs_insert_replace_extent(trans, inode, path,
2956                                 extent_info, extent_info->data_len,
2957                                 drop_args.bytes_found);
2958                 if (ret) {
2959                         btrfs_abort_transaction(trans, ret);
2960                         goto out_trans;
2961                 }
2962         }
2963
2964 out_trans:
2965         if (!trans)
2966                 goto out_free;
2967
2968         trans->block_rsv = &fs_info->trans_block_rsv;
2969         if (ret)
2970                 btrfs_end_transaction(trans);
2971         else
2972                 *trans_out = trans;
2973 out_free:
2974         btrfs_free_block_rsv(fs_info, rsv);
2975 out:
2976         return ret;
2977 }
2978
2979 static int btrfs_punch_hole(struct file *file, loff_t offset, loff_t len)
2980 {
2981         struct inode *inode = file_inode(file);
2982         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2983         struct btrfs_root *root = BTRFS_I(inode)->root;
2984         struct extent_state *cached_state = NULL;
2985         struct btrfs_path *path;
2986         struct btrfs_trans_handle *trans = NULL;
2987         u64 lockstart;
2988         u64 lockend;
2989         u64 tail_start;
2990         u64 tail_len;
2991         u64 orig_start = offset;
2992         int ret = 0;
2993         bool same_block;
2994         u64 ino_size;
2995         bool truncated_block = false;
2996         bool updated_inode = false;
2997
2998         ret = btrfs_wait_ordered_range(inode, offset, len);
2999         if (ret)
3000                 return ret;
3001
3002         btrfs_inode_lock(inode, BTRFS_ILOCK_MMAP);
3003         ino_size = round_up(inode->i_size, fs_info->sectorsize);
3004         ret = find_first_non_hole(BTRFS_I(inode), &offset, &len);
3005         if (ret < 0)
3006                 goto out_only_mutex;
3007         if (ret && !len) {
3008                 /* Already in a large hole */
3009                 ret = 0;
3010                 goto out_only_mutex;
3011         }
3012
3013         ret = file_modified(file);
3014         if (ret)
3015                 goto out_only_mutex;
3016
3017         lockstart = round_up(offset, btrfs_inode_sectorsize(BTRFS_I(inode)));
3018         lockend = round_down(offset + len,
3019                              btrfs_inode_sectorsize(BTRFS_I(inode))) - 1;
3020         same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
3021                 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
3022         /*
3023          * We needn't truncate any block which is beyond the end of the file
3024          * because we are sure there is no data there.
3025          */
3026         /*
3027          * Only do this if we are in the same block and we aren't doing the
3028          * entire block.
3029          */
3030         if (same_block && len < fs_info->sectorsize) {
3031                 if (offset < ino_size) {
3032                         truncated_block = true;
3033                         ret = btrfs_truncate_block(BTRFS_I(inode), offset, len,
3034                                                    0);
3035                 } else {
3036                         ret = 0;
3037                 }
3038                 goto out_only_mutex;
3039         }
3040
3041         /* zero back part of the first block */
3042         if (offset < ino_size) {
3043                 truncated_block = true;
3044                 ret = btrfs_truncate_block(BTRFS_I(inode), offset, 0, 0);
3045                 if (ret) {
3046                         btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
3047                         return ret;
3048                 }
3049         }
3050
3051         /* Check the aligned pages after the first unaligned page,
3052          * if offset != orig_start, which means the first unaligned page
3053          * including several following pages are already in holes,
3054          * the extra check can be skipped */
3055         if (offset == orig_start) {
3056                 /* after truncate page, check hole again */
3057                 len = offset + len - lockstart;
3058                 offset = lockstart;
3059                 ret = find_first_non_hole(BTRFS_I(inode), &offset, &len);
3060                 if (ret < 0)
3061                         goto out_only_mutex;
3062                 if (ret && !len) {
3063                         ret = 0;
3064                         goto out_only_mutex;
3065                 }
3066                 lockstart = offset;
3067         }
3068
3069         /* Check the tail unaligned part is in a hole */
3070         tail_start = lockend + 1;
3071         tail_len = offset + len - tail_start;
3072         if (tail_len) {
3073                 ret = find_first_non_hole(BTRFS_I(inode), &tail_start, &tail_len);
3074                 if (unlikely(ret < 0))
3075                         goto out_only_mutex;
3076                 if (!ret) {
3077                         /* zero the front end of the last page */
3078                         if (tail_start + tail_len < ino_size) {
3079                                 truncated_block = true;
3080                                 ret = btrfs_truncate_block(BTRFS_I(inode),
3081                                                         tail_start + tail_len,
3082                                                         0, 1);
3083                                 if (ret)
3084                                         goto out_only_mutex;
3085                         }
3086                 }
3087         }
3088
3089         if (lockend < lockstart) {
3090                 ret = 0;
3091                 goto out_only_mutex;
3092         }
3093
3094         ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
3095                                           &cached_state);
3096         if (ret)
3097                 goto out_only_mutex;
3098
3099         path = btrfs_alloc_path();
3100         if (!path) {
3101                 ret = -ENOMEM;
3102                 goto out;
3103         }
3104
3105         ret = btrfs_replace_file_extents(BTRFS_I(inode), path, lockstart,
3106                                          lockend, NULL, &trans);
3107         btrfs_free_path(path);
3108         if (ret)
3109                 goto out;
3110
3111         ASSERT(trans != NULL);
3112         inode_inc_iversion(inode);
3113         inode->i_mtime = inode->i_ctime = current_time(inode);
3114         ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
3115         updated_inode = true;
3116         btrfs_end_transaction(trans);
3117         btrfs_btree_balance_dirty(fs_info);
3118 out:
3119         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3120                              &cached_state);
3121 out_only_mutex:
3122         if (!updated_inode && truncated_block && !ret) {
3123                 /*
3124                  * If we only end up zeroing part of a page, we still need to
3125                  * update the inode item, so that all the time fields are
3126                  * updated as well as the necessary btrfs inode in memory fields
3127                  * for detecting, at fsync time, if the inode isn't yet in the
3128                  * log tree or it's there but not up to date.
3129                  */
3130                 struct timespec64 now = current_time(inode);
3131
3132                 inode_inc_iversion(inode);
3133                 inode->i_mtime = now;
3134                 inode->i_ctime = now;
3135                 trans = btrfs_start_transaction(root, 1);
3136                 if (IS_ERR(trans)) {
3137                         ret = PTR_ERR(trans);
3138                 } else {
3139                         int ret2;
3140
3141                         ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
3142                         ret2 = btrfs_end_transaction(trans);
3143                         if (!ret)
3144                                 ret = ret2;
3145                 }
3146         }
3147         btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
3148         return ret;
3149 }
3150
3151 /* Helper structure to record which range is already reserved */
3152 struct falloc_range {
3153         struct list_head list;
3154         u64 start;
3155         u64 len;
3156 };
3157
3158 /*
3159  * Helper function to add falloc range
3160  *
3161  * Caller should have locked the larger range of extent containing
3162  * [start, len)
3163  */
3164 static int add_falloc_range(struct list_head *head, u64 start, u64 len)
3165 {
3166         struct falloc_range *range = NULL;
3167
3168         if (!list_empty(head)) {
3169                 /*
3170                  * As fallocate iterates by bytenr order, we only need to check
3171                  * the last range.
3172                  */
3173                 range = list_last_entry(head, struct falloc_range, list);
3174                 if (range->start + range->len == start) {
3175                         range->len += len;
3176                         return 0;
3177                 }
3178         }
3179
3180         range = kmalloc(sizeof(*range), GFP_KERNEL);
3181         if (!range)
3182                 return -ENOMEM;
3183         range->start = start;
3184         range->len = len;
3185         list_add_tail(&range->list, head);
3186         return 0;
3187 }
3188
3189 static int btrfs_fallocate_update_isize(struct inode *inode,
3190                                         const u64 end,
3191                                         const int mode)
3192 {
3193         struct btrfs_trans_handle *trans;
3194         struct btrfs_root *root = BTRFS_I(inode)->root;
3195         int ret;
3196         int ret2;
3197
3198         if (mode & FALLOC_FL_KEEP_SIZE || end <= i_size_read(inode))
3199                 return 0;
3200
3201         trans = btrfs_start_transaction(root, 1);
3202         if (IS_ERR(trans))
3203                 return PTR_ERR(trans);
3204
3205         inode->i_ctime = current_time(inode);
3206         i_size_write(inode, end);
3207         btrfs_inode_safe_disk_i_size_write(BTRFS_I(inode), 0);
3208         ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
3209         ret2 = btrfs_end_transaction(trans);
3210
3211         return ret ? ret : ret2;
3212 }
3213
3214 enum {
3215         RANGE_BOUNDARY_WRITTEN_EXTENT,
3216         RANGE_BOUNDARY_PREALLOC_EXTENT,
3217         RANGE_BOUNDARY_HOLE,
3218 };
3219
3220 static int btrfs_zero_range_check_range_boundary(struct btrfs_inode *inode,
3221                                                  u64 offset)
3222 {
3223         const u64 sectorsize = btrfs_inode_sectorsize(inode);
3224         struct extent_map *em;
3225         int ret;
3226
3227         offset = round_down(offset, sectorsize);
3228         em = btrfs_get_extent(inode, NULL, 0, offset, sectorsize);
3229         if (IS_ERR(em))
3230                 return PTR_ERR(em);
3231
3232         if (em->block_start == EXTENT_MAP_HOLE)
3233                 ret = RANGE_BOUNDARY_HOLE;
3234         else if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
3235                 ret = RANGE_BOUNDARY_PREALLOC_EXTENT;
3236         else
3237                 ret = RANGE_BOUNDARY_WRITTEN_EXTENT;
3238
3239         free_extent_map(em);
3240         return ret;
3241 }
3242
3243 static int btrfs_zero_range(struct inode *inode,
3244                             loff_t offset,
3245                             loff_t len,
3246                             const int mode)
3247 {
3248         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
3249         struct extent_map *em;
3250         struct extent_changeset *data_reserved = NULL;
3251         int ret;
3252         u64 alloc_hint = 0;
3253         const u64 sectorsize = btrfs_inode_sectorsize(BTRFS_I(inode));
3254         u64 alloc_start = round_down(offset, sectorsize);
3255         u64 alloc_end = round_up(offset + len, sectorsize);
3256         u64 bytes_to_reserve = 0;
3257         bool space_reserved = false;
3258
3259         inode_dio_wait(inode);
3260
3261         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
3262                               alloc_end - alloc_start);
3263         if (IS_ERR(em)) {
3264                 ret = PTR_ERR(em);
3265                 goto out;
3266         }
3267
3268         /*
3269          * Avoid hole punching and extent allocation for some cases. More cases
3270          * could be considered, but these are unlikely common and we keep things
3271          * as simple as possible for now. Also, intentionally, if the target
3272          * range contains one or more prealloc extents together with regular
3273          * extents and holes, we drop all the existing extents and allocate a
3274          * new prealloc extent, so that we get a larger contiguous disk extent.
3275          */
3276         if (em->start <= alloc_start &&
3277             test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3278                 const u64 em_end = em->start + em->len;
3279
3280                 if (em_end >= offset + len) {
3281                         /*
3282                          * The whole range is already a prealloc extent,
3283                          * do nothing except updating the inode's i_size if
3284                          * needed.
3285                          */
3286                         free_extent_map(em);
3287                         ret = btrfs_fallocate_update_isize(inode, offset + len,
3288                                                            mode);
3289                         goto out;
3290                 }
3291                 /*
3292                  * Part of the range is already a prealloc extent, so operate
3293                  * only on the remaining part of the range.
3294                  */
3295                 alloc_start = em_end;
3296                 ASSERT(IS_ALIGNED(alloc_start, sectorsize));
3297                 len = offset + len - alloc_start;
3298                 offset = alloc_start;
3299                 alloc_hint = em->block_start + em->len;
3300         }
3301         free_extent_map(em);
3302
3303         if (BTRFS_BYTES_TO_BLKS(fs_info, offset) ==
3304             BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1)) {
3305                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
3306                                       sectorsize);
3307                 if (IS_ERR(em)) {
3308                         ret = PTR_ERR(em);
3309                         goto out;
3310                 }
3311
3312                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3313                         free_extent_map(em);
3314                         ret = btrfs_fallocate_update_isize(inode, offset + len,
3315                                                            mode);
3316                         goto out;
3317                 }
3318                 if (len < sectorsize && em->block_start != EXTENT_MAP_HOLE) {
3319                         free_extent_map(em);
3320                         ret = btrfs_truncate_block(BTRFS_I(inode), offset, len,
3321                                                    0);
3322                         if (!ret)
3323                                 ret = btrfs_fallocate_update_isize(inode,
3324                                                                    offset + len,
3325                                                                    mode);
3326                         return ret;
3327                 }
3328                 free_extent_map(em);
3329                 alloc_start = round_down(offset, sectorsize);
3330                 alloc_end = alloc_start + sectorsize;
3331                 goto reserve_space;
3332         }
3333
3334         alloc_start = round_up(offset, sectorsize);
3335         alloc_end = round_down(offset + len, sectorsize);
3336
3337         /*
3338          * For unaligned ranges, check the pages at the boundaries, they might
3339          * map to an extent, in which case we need to partially zero them, or
3340          * they might map to a hole, in which case we need our allocation range
3341          * to cover them.
3342          */
3343         if (!IS_ALIGNED(offset, sectorsize)) {
3344                 ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
3345                                                             offset);
3346                 if (ret < 0)
3347                         goto out;
3348                 if (ret == RANGE_BOUNDARY_HOLE) {
3349                         alloc_start = round_down(offset, sectorsize);
3350                         ret = 0;
3351                 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3352                         ret = btrfs_truncate_block(BTRFS_I(inode), offset, 0, 0);
3353                         if (ret)
3354                                 goto out;
3355                 } else {
3356                         ret = 0;
3357                 }
3358         }
3359
3360         if (!IS_ALIGNED(offset + len, sectorsize)) {
3361                 ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
3362                                                             offset + len);
3363                 if (ret < 0)
3364                         goto out;
3365                 if (ret == RANGE_BOUNDARY_HOLE) {
3366                         alloc_end = round_up(offset + len, sectorsize);
3367                         ret = 0;
3368                 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3369                         ret = btrfs_truncate_block(BTRFS_I(inode), offset + len,
3370                                                    0, 1);
3371                         if (ret)
3372                                 goto out;
3373                 } else {
3374                         ret = 0;
3375                 }
3376         }
3377
3378 reserve_space:
3379         if (alloc_start < alloc_end) {
3380                 struct extent_state *cached_state = NULL;
3381                 const u64 lockstart = alloc_start;
3382                 const u64 lockend = alloc_end - 1;
3383
3384                 bytes_to_reserve = alloc_end - alloc_start;
3385                 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3386                                                       bytes_to_reserve);
3387                 if (ret < 0)
3388                         goto out;
3389                 space_reserved = true;
3390                 ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
3391                                                   &cached_state);
3392                 if (ret)
3393                         goto out;
3394                 ret = btrfs_qgroup_reserve_data(BTRFS_I(inode), &data_reserved,
3395                                                 alloc_start, bytes_to_reserve);
3396                 if (ret) {
3397                         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
3398                                              lockend, &cached_state);
3399                         goto out;
3400                 }
3401                 ret = btrfs_prealloc_file_range(inode, mode, alloc_start,
3402                                                 alloc_end - alloc_start,
3403                                                 i_blocksize(inode),
3404                                                 offset + len, &alloc_hint);
3405                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
3406                                      lockend, &cached_state);
3407                 /* btrfs_prealloc_file_range releases reserved space on error */
3408                 if (ret) {
3409                         space_reserved = false;
3410                         goto out;
3411                 }
3412         }
3413         ret = btrfs_fallocate_update_isize(inode, offset + len, mode);
3414  out:
3415         if (ret && space_reserved)
3416                 btrfs_free_reserved_data_space(BTRFS_I(inode), data_reserved,
3417                                                alloc_start, bytes_to_reserve);
3418         extent_changeset_free(data_reserved);
3419
3420         return ret;
3421 }
3422
3423 static long btrfs_fallocate(struct file *file, int mode,
3424                             loff_t offset, loff_t len)
3425 {
3426         struct inode *inode = file_inode(file);
3427         struct extent_state *cached_state = NULL;
3428         struct extent_changeset *data_reserved = NULL;
3429         struct falloc_range *range;
3430         struct falloc_range *tmp;
3431         struct list_head reserve_list;
3432         u64 cur_offset;
3433         u64 last_byte;
3434         u64 alloc_start;
3435         u64 alloc_end;
3436         u64 alloc_hint = 0;
3437         u64 locked_end;
3438         u64 actual_end = 0;
3439         struct extent_map *em;
3440         int blocksize = btrfs_inode_sectorsize(BTRFS_I(inode));
3441         int ret;
3442
3443         /* Do not allow fallocate in ZONED mode */
3444         if (btrfs_is_zoned(btrfs_sb(inode->i_sb)))
3445                 return -EOPNOTSUPP;
3446
3447         alloc_start = round_down(offset, blocksize);
3448         alloc_end = round_up(offset + len, blocksize);
3449         cur_offset = alloc_start;
3450
3451         /* Make sure we aren't being give some crap mode */
3452         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3453                      FALLOC_FL_ZERO_RANGE))
3454                 return -EOPNOTSUPP;
3455
3456         if (mode & FALLOC_FL_PUNCH_HOLE)
3457                 return btrfs_punch_hole(file, offset, len);
3458
3459         /*
3460          * Only trigger disk allocation, don't trigger qgroup reserve
3461          *
3462          * For qgroup space, it will be checked later.
3463          */
3464         if (!(mode & FALLOC_FL_ZERO_RANGE)) {
3465                 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3466                                                       alloc_end - alloc_start);
3467                 if (ret < 0)
3468                         return ret;
3469         }
3470
3471         btrfs_inode_lock(inode, BTRFS_ILOCK_MMAP);
3472
3473         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
3474                 ret = inode_newsize_ok(inode, offset + len);
3475                 if (ret)
3476                         goto out;
3477         }
3478
3479         ret = file_modified(file);
3480         if (ret)
3481                 goto out;
3482
3483         /*
3484          * TODO: Move these two operations after we have checked
3485          * accurate reserved space, or fallocate can still fail but
3486          * with page truncated or size expanded.
3487          *
3488          * But that's a minor problem and won't do much harm BTW.
3489          */
3490         if (alloc_start > inode->i_size) {
3491                 ret = btrfs_cont_expand(BTRFS_I(inode), i_size_read(inode),
3492                                         alloc_start);
3493                 if (ret)
3494                         goto out;
3495         } else if (offset + len > inode->i_size) {
3496                 /*
3497                  * If we are fallocating from the end of the file onward we
3498                  * need to zero out the end of the block if i_size lands in the
3499                  * middle of a block.
3500                  */
3501                 ret = btrfs_truncate_block(BTRFS_I(inode), inode->i_size, 0, 0);
3502                 if (ret)
3503                         goto out;
3504         }
3505
3506         /*
3507          * wait for ordered IO before we have any locks.  We'll loop again
3508          * below with the locks held.
3509          */
3510         ret = btrfs_wait_ordered_range(inode, alloc_start,
3511                                        alloc_end - alloc_start);
3512         if (ret)
3513                 goto out;
3514
3515         if (mode & FALLOC_FL_ZERO_RANGE) {
3516                 ret = btrfs_zero_range(inode, offset, len, mode);
3517                 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
3518                 return ret;
3519         }
3520
3521         locked_end = alloc_end - 1;
3522         while (1) {
3523                 struct btrfs_ordered_extent *ordered;
3524
3525                 /* the extent lock is ordered inside the running
3526                  * transaction
3527                  */
3528                 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
3529                                  locked_end, &cached_state);
3530                 ordered = btrfs_lookup_first_ordered_extent(BTRFS_I(inode),
3531                                                             locked_end);
3532
3533                 if (ordered &&
3534                     ordered->file_offset + ordered->num_bytes > alloc_start &&
3535                     ordered->file_offset < alloc_end) {
3536                         btrfs_put_ordered_extent(ordered);
3537                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
3538                                              alloc_start, locked_end,
3539                                              &cached_state);
3540                         /*
3541                          * we can't wait on the range with the transaction
3542                          * running or with the extent lock held
3543                          */
3544                         ret = btrfs_wait_ordered_range(inode, alloc_start,
3545                                                        alloc_end - alloc_start);
3546                         if (ret)
3547                                 goto out;
3548                 } else {
3549                         if (ordered)
3550                                 btrfs_put_ordered_extent(ordered);
3551                         break;
3552                 }
3553         }
3554
3555         /* First, check if we exceed the qgroup limit */
3556         INIT_LIST_HEAD(&reserve_list);
3557         while (cur_offset < alloc_end) {
3558                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
3559                                       alloc_end - cur_offset);
3560                 if (IS_ERR(em)) {
3561                         ret = PTR_ERR(em);
3562                         break;
3563                 }
3564                 last_byte = min(extent_map_end(em), alloc_end);
3565                 actual_end = min_t(u64, extent_map_end(em), offset + len);
3566                 last_byte = ALIGN(last_byte, blocksize);
3567                 if (em->block_start == EXTENT_MAP_HOLE ||
3568                     (cur_offset >= inode->i_size &&
3569                      !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
3570                         ret = add_falloc_range(&reserve_list, cur_offset,
3571                                                last_byte - cur_offset);
3572                         if (ret < 0) {
3573                                 free_extent_map(em);
3574                                 break;
3575                         }
3576                         ret = btrfs_qgroup_reserve_data(BTRFS_I(inode),
3577                                         &data_reserved, cur_offset,
3578                                         last_byte - cur_offset);
3579                         if (ret < 0) {
3580                                 cur_offset = last_byte;
3581                                 free_extent_map(em);
3582                                 break;
3583                         }
3584                 } else {
3585                         /*
3586                          * Do not need to reserve unwritten extent for this
3587                          * range, free reserved data space first, otherwise
3588                          * it'll result in false ENOSPC error.
3589                          */
3590                         btrfs_free_reserved_data_space(BTRFS_I(inode),
3591                                 data_reserved, cur_offset,
3592                                 last_byte - cur_offset);
3593                 }
3594                 free_extent_map(em);
3595                 cur_offset = last_byte;
3596         }
3597
3598         /*
3599          * If ret is still 0, means we're OK to fallocate.
3600          * Or just cleanup the list and exit.
3601          */
3602         list_for_each_entry_safe(range, tmp, &reserve_list, list) {
3603                 if (!ret)
3604                         ret = btrfs_prealloc_file_range(inode, mode,
3605                                         range->start,
3606                                         range->len, i_blocksize(inode),
3607                                         offset + len, &alloc_hint);
3608                 else
3609                         btrfs_free_reserved_data_space(BTRFS_I(inode),
3610                                         data_reserved, range->start,
3611                                         range->len);
3612                 list_del(&range->list);
3613                 kfree(range);
3614         }
3615         if (ret < 0)
3616                 goto out_unlock;
3617
3618         /*
3619          * We didn't need to allocate any more space, but we still extended the
3620          * size of the file so we need to update i_size and the inode item.
3621          */
3622         ret = btrfs_fallocate_update_isize(inode, actual_end, mode);
3623 out_unlock:
3624         unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3625                              &cached_state);
3626 out:
3627         btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
3628         /* Let go of our reservation. */
3629         if (ret != 0 && !(mode & FALLOC_FL_ZERO_RANGE))
3630                 btrfs_free_reserved_data_space(BTRFS_I(inode), data_reserved,
3631                                 cur_offset, alloc_end - cur_offset);
3632         extent_changeset_free(data_reserved);
3633         return ret;
3634 }
3635
3636 static loff_t find_desired_extent(struct btrfs_inode *inode, loff_t offset,
3637                                   int whence)
3638 {
3639         struct btrfs_fs_info *fs_info = inode->root->fs_info;
3640         struct extent_map *em = NULL;
3641         struct extent_state *cached_state = NULL;
3642         loff_t i_size = inode->vfs_inode.i_size;
3643         u64 lockstart;
3644         u64 lockend;
3645         u64 start;
3646         u64 len;
3647         int ret = 0;
3648
3649         if (i_size == 0 || offset >= i_size)
3650                 return -ENXIO;
3651
3652         /*
3653          * offset can be negative, in this case we start finding DATA/HOLE from
3654          * the very start of the file.
3655          */
3656         start = max_t(loff_t, 0, offset);
3657
3658         lockstart = round_down(start, fs_info->sectorsize);
3659         lockend = round_up(i_size, fs_info->sectorsize);
3660         if (lockend <= lockstart)
3661                 lockend = lockstart + fs_info->sectorsize;
3662         lockend--;
3663         len = lockend - lockstart + 1;
3664
3665         lock_extent_bits(&inode->io_tree, lockstart, lockend, &cached_state);
3666
3667         while (start < i_size) {
3668                 em = btrfs_get_extent_fiemap(inode, start, len);
3669                 if (IS_ERR(em)) {
3670                         ret = PTR_ERR(em);
3671                         em = NULL;
3672                         break;
3673                 }
3674
3675                 if (whence == SEEK_HOLE &&
3676                     (em->block_start == EXTENT_MAP_HOLE ||
3677                      test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3678                         break;
3679                 else if (whence == SEEK_DATA &&
3680                            (em->block_start != EXTENT_MAP_HOLE &&
3681                             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3682                         break;
3683
3684                 start = em->start + em->len;
3685                 free_extent_map(em);
3686                 em = NULL;
3687                 cond_resched();
3688         }
3689         free_extent_map(em);
3690         unlock_extent_cached(&inode->io_tree, lockstart, lockend,
3691                              &cached_state);
3692         if (ret) {
3693                 offset = ret;
3694         } else {
3695                 if (whence == SEEK_DATA && start >= i_size)
3696                         offset = -ENXIO;
3697                 else
3698                         offset = min_t(loff_t, start, i_size);
3699         }
3700
3701         return offset;
3702 }
3703
3704 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3705 {
3706         struct inode *inode = file->f_mapping->host;
3707
3708         switch (whence) {
3709         default:
3710                 return generic_file_llseek(file, offset, whence);
3711         case SEEK_DATA:
3712         case SEEK_HOLE:
3713                 btrfs_inode_lock(inode, BTRFS_ILOCK_SHARED);
3714                 offset = find_desired_extent(BTRFS_I(inode), offset, whence);
3715                 btrfs_inode_unlock(inode, BTRFS_ILOCK_SHARED);
3716                 break;
3717         }
3718
3719         if (offset < 0)
3720                 return offset;
3721
3722         return vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3723 }
3724
3725 static int btrfs_file_open(struct inode *inode, struct file *filp)
3726 {
3727         int ret;
3728
3729         filp->f_mode |= FMODE_NOWAIT | FMODE_BUF_RASYNC;
3730
3731         ret = fsverity_file_open(inode, filp);
3732         if (ret)
3733                 return ret;
3734         return generic_file_open(inode, filp);
3735 }
3736
3737 static int check_direct_read(struct btrfs_fs_info *fs_info,
3738                              const struct iov_iter *iter, loff_t offset)
3739 {
3740         int ret;
3741         int i, seg;
3742
3743         ret = check_direct_IO(fs_info, iter, offset);
3744         if (ret < 0)
3745                 return ret;
3746
3747         if (!iter_is_iovec(iter))
3748                 return 0;
3749
3750         for (seg = 0; seg < iter->nr_segs; seg++)
3751                 for (i = seg + 1; i < iter->nr_segs; i++)
3752                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
3753                                 return -EINVAL;
3754         return 0;
3755 }
3756
3757 static ssize_t btrfs_direct_read(struct kiocb *iocb, struct iov_iter *to)
3758 {
3759         struct inode *inode = file_inode(iocb->ki_filp);
3760         size_t prev_left = 0;
3761         ssize_t read = 0;
3762         ssize_t ret;
3763
3764         if (fsverity_active(inode))
3765                 return 0;
3766
3767         if (check_direct_read(btrfs_sb(inode->i_sb), to, iocb->ki_pos))
3768                 return 0;
3769
3770         btrfs_inode_lock(inode, BTRFS_ILOCK_SHARED);
3771 again:
3772         /*
3773          * This is similar to what we do for direct IO writes, see the comment
3774          * at btrfs_direct_write(), but we also disable page faults in addition
3775          * to disabling them only at the iov_iter level. This is because when
3776          * reading from a hole or prealloc extent, iomap calls iov_iter_zero(),
3777          * which can still trigger page fault ins despite having set ->nofault
3778          * to true of our 'to' iov_iter.
3779          *
3780          * The difference to direct IO writes is that we deadlock when trying
3781          * to lock the extent range in the inode's tree during he page reads
3782          * triggered by the fault in (while for writes it is due to waiting for
3783          * our own ordered extent). This is because for direct IO reads,
3784          * btrfs_dio_iomap_begin() returns with the extent range locked, which
3785          * is only unlocked in the endio callback (end_bio_extent_readpage()).
3786          */
3787         pagefault_disable();
3788         to->nofault = true;
3789         ret = iomap_dio_rw(iocb, to, &btrfs_dio_iomap_ops, &btrfs_dio_ops,
3790                            IOMAP_DIO_PARTIAL, read);
3791         to->nofault = false;
3792         pagefault_enable();
3793
3794         /* No increment (+=) because iomap returns a cumulative value. */
3795         if (ret > 0)
3796                 read = ret;
3797
3798         if (iov_iter_count(to) > 0 && (ret == -EFAULT || ret > 0)) {
3799                 const size_t left = iov_iter_count(to);
3800
3801                 if (left == prev_left) {
3802                         /*
3803                          * We didn't make any progress since the last attempt,
3804                          * fallback to a buffered read for the remainder of the
3805                          * range. This is just to avoid any possibility of looping
3806                          * for too long.
3807                          */
3808                         ret = read;
3809                 } else {
3810                         /*
3811                          * We made some progress since the last retry or this is
3812                          * the first time we are retrying. Fault in as many pages
3813                          * as possible and retry.
3814                          */
3815                         fault_in_iov_iter_writeable(to, left);
3816                         prev_left = left;
3817                         goto again;
3818                 }
3819         }
3820         btrfs_inode_unlock(inode, BTRFS_ILOCK_SHARED);
3821         return ret < 0 ? ret : read;
3822 }
3823
3824 static ssize_t btrfs_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
3825 {
3826         ssize_t ret = 0;
3827
3828         if (iocb->ki_flags & IOCB_DIRECT) {
3829                 ret = btrfs_direct_read(iocb, to);
3830                 if (ret < 0 || !iov_iter_count(to) ||
3831                     iocb->ki_pos >= i_size_read(file_inode(iocb->ki_filp)))
3832                         return ret;
3833         }
3834
3835         return filemap_read(iocb, to, ret);
3836 }
3837
3838 const struct file_operations btrfs_file_operations = {
3839         .llseek         = btrfs_file_llseek,
3840         .read_iter      = btrfs_file_read_iter,
3841         .splice_read    = generic_file_splice_read,
3842         .write_iter     = btrfs_file_write_iter,
3843         .splice_write   = iter_file_splice_write,
3844         .mmap           = btrfs_file_mmap,
3845         .open           = btrfs_file_open,
3846         .release        = btrfs_release_file,
3847         .fsync          = btrfs_sync_file,
3848         .fallocate      = btrfs_fallocate,
3849         .unlocked_ioctl = btrfs_ioctl,
3850 #ifdef CONFIG_COMPAT
3851         .compat_ioctl   = btrfs_compat_ioctl,
3852 #endif
3853         .remap_file_range = btrfs_remap_file_range,
3854 };
3855
3856 void __cold btrfs_auto_defrag_exit(void)
3857 {
3858         kmem_cache_destroy(btrfs_inode_defrag_cachep);
3859 }
3860
3861 int __init btrfs_auto_defrag_init(void)
3862 {
3863         btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
3864                                         sizeof(struct inode_defrag), 0,
3865                                         SLAB_MEM_SPREAD,
3866                                         NULL);
3867         if (!btrfs_inode_defrag_cachep)
3868                 return -ENOMEM;
3869
3870         return 0;
3871 }
3872
3873 int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3874 {
3875         int ret;
3876
3877         /*
3878          * So with compression we will find and lock a dirty page and clear the
3879          * first one as dirty, setup an async extent, and immediately return
3880          * with the entire range locked but with nobody actually marked with
3881          * writeback.  So we can't just filemap_write_and_wait_range() and
3882          * expect it to work since it will just kick off a thread to do the
3883          * actual work.  So we need to call filemap_fdatawrite_range _again_
3884          * since it will wait on the page lock, which won't be unlocked until
3885          * after the pages have been marked as writeback and so we're good to go
3886          * from there.  We have to do this otherwise we'll miss the ordered
3887          * extents and that results in badness.  Please Josef, do not think you
3888          * know better and pull this out at some point in the future, it is
3889          * right and you are wrong.
3890          */
3891         ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3892         if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3893                              &BTRFS_I(inode)->runtime_flags))
3894                 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3895
3896         return ret;
3897 }