fs/btrfs/file.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/fs.h>
  20 #include <linux/pagemap.h>
  21 #include <linux/highmem.h>
  22 #include <linux/time.h>
  23 #include <linux/init.h>
  24 #include <linux/string.h>
  25 #include <linux/backing-dev.h>
  26 #include <linux/mpage.h>
  27 #include <linux/aio.h>
  28 #include <linux/falloc.h>
  29 #include <linux/swap.h>
  30 #include <linux/writeback.h>
  31 #include <linux/statfs.h>
  32 #include <linux/compat.h>
  33 #include <linux/slab.h>
  34 #include <linux/btrfs.h>
  35 #include "ctree.h"
  36 #include "disk-io.h"
  37 #include "transaction.h"
  38 #include "btrfs_inode.h"
  39 #include "print-tree.h"
  40 #include "tree-log.h"
  41 #include "locking.h"
  42 #include "compat.h"
  43 #include "volumes.h"
  44
  45 static struct kmem_cache *btrfs_inode_defrag_cachep;
  46 /*
  47  * when auto defrag is enabled we
  48  * queue up these defrag structs to remember which
  49  * inodes need defragging passes
  50  */
  51 struct inode_defrag {
  52         struct rb_node rb_node;
  53         /* objectid */
  54         u64 ino;
  55         /*
  56          * transid where the defrag was added, we search for
  57          * extents newer than this
  58          */
  59         u64 transid;
  60
  61         /* root objectid */
  62         u64 root;
  63
  64         /* last offset we were able to defrag */
  65         u64 last_offset;
  66
  67         /* if we've wrapped around back to zero once already */
  68         int cycled;
  69 };
  70
  71 static int __compare_inode_defrag(struct inode_defrag *defrag1,
  72                                   struct inode_defrag *defrag2)
  73 {
  74         if (defrag1->root > defrag2->root)
  75                 return 1;
  76         else if (defrag1->root < defrag2->root)
  77                 return -1;
  78         else if (defrag1->ino > defrag2->ino)
  79                 return 1;
  80         else if (defrag1->ino < defrag2->ino)
  81                 return -1;
  82         else
  83                 return 0;
  84 }
  85
  86 /* pop a record for an inode into the defrag tree.  The lock
  87  * must be held already
  88  *
  89  * If you're inserting a record for an older transid than an
  90  * existing record, the transid already in the tree is lowered
  91  *
  92  * If an existing record is found the defrag item you
  93  * pass in is freed
  94  */
  95 static int __btrfs_add_inode_defrag(struct inode *inode,
  96                                     struct inode_defrag *defrag)
  97 {
  98         struct btrfs_root *root = BTRFS_I(inode)->root;
  99         struct inode_defrag *entry;
 100         struct rb_node **p;
 101         struct rb_node *parent = NULL;
 102         int ret;
 103
 104         p = &root->fs_info->defrag_inodes.rb_node;
 105         while (*p) {
 106                 parent = *p;
 107                 entry = rb_entry(parent, struct inode_defrag, rb_node);
 108
 109                 ret = __compare_inode_defrag(defrag, entry);
 110                 if (ret < 0)
 111                         p = &parent->rb_left;
 112                 else if (ret > 0)
 113                         p = &parent->rb_right;
 114                 else {
 115                         /* if we're reinserting an entry for
 116                          * an old defrag run, make sure to
 117                          * lower the transid of our existing record
 118                          */
 119                         if (defrag->transid < entry->transid)
 120                                 entry->transid = defrag->transid;
 121                         if (defrag->last_offset > entry->last_offset)
 122                                 entry->last_offset = defrag->last_offset;
 123                         return -EEXIST;
 124                 }
 125         }
 126         set_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
 127         rb_link_node(&defrag->rb_node, parent, p);
 128         rb_insert_color(&defrag->rb_node, &root->fs_info->defrag_inodes);
 129         return 0;
 130 }
 131
 132 static inline int __need_auto_defrag(struct btrfs_root *root)
 133 {
 134         if (!btrfs_test_opt(root, AUTO_DEFRAG))
 135                 return 0;
 136
 137         if (btrfs_fs_closing(root->fs_info))
 138                 return 0;
 139
 140         return 1;
 141 }
 142
 143 /*
 144  * insert a defrag record for this inode if auto defrag is
 145  * enabled
 146  */
 147 int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
 148                            struct inode *inode)
 149 {
 150         struct btrfs_root *root = BTRFS_I(inode)->root;
 151         struct inode_defrag *defrag;
 152         u64 transid;
 153         int ret;
 154
 155         if (!__need_auto_defrag(root))
 156                 return 0;
 157
 158         if (test_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags))
 159                 return 0;
 160
 161         if (trans)
 162                 transid = trans->transid;
 163         else
 164                 transid = BTRFS_I(inode)->root->last_trans;
 165
 166         defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
 167         if (!defrag)
 168                 return -ENOMEM;
 169
 170         defrag->ino = btrfs_ino(inode);
 171         defrag->transid = transid;
 172         defrag->root = root->root_key.objectid;
 173
 174         spin_lock(&root->fs_info->defrag_inodes_lock);
 175         if (!test_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags)) {
 176                 /*
 177                  * If we set IN_DEFRAG flag and evict the inode from memory,
 178                  * and then re-read this inode, this new inode doesn't have
 179                  * IN_DEFRAG flag. At the case, we may find the existed defrag.
 180                  */
 181                 ret = __btrfs_add_inode_defrag(inode, defrag);
 182                 if (ret)
 183                         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 184         } else {
 185                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 186         }
 187         spin_unlock(&root->fs_info->defrag_inodes_lock);
 188         return 0;
 189 }
 190
 191 /*
 192  * Requeue the defrag object. If there is a defrag object that points to
 193  * the same inode in the tree, we will merge them together (by
 194  * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
 195  */
 196 static void btrfs_requeue_inode_defrag(struct inode *inode,
 197                                        struct inode_defrag *defrag)
 198 {
 199         struct btrfs_root *root = BTRFS_I(inode)->root;
 200         int ret;
 201
 202         if (!__need_auto_defrag(root))
 203                 goto out;
 204
 205         /*
 206          * Here we don't check the IN_DEFRAG flag, because we need merge
 207          * them together.
 208          */
 209         spin_lock(&root->fs_info->defrag_inodes_lock);
 210         ret = __btrfs_add_inode_defrag(inode, defrag);
 211         spin_unlock(&root->fs_info->defrag_inodes_lock);
 212         if (ret)
 213                 goto out;
 214         return;
 215 out:
 216         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 217 }
 218
 219 /*
 220  * pick the defragable inode that we want, if it doesn't exist, we will get
 221  * the next one.
 222  */
 223 static struct inode_defrag *
 224 btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
 225 {
 226         struct inode_defrag *entry = NULL;
 227         struct inode_defrag tmp;
 228         struct rb_node *p;
 229         struct rb_node *parent = NULL;
 230         int ret;
 231
 232         tmp.ino = ino;
 233         tmp.root = root;
 234
 235         spin_lock(&fs_info->defrag_inodes_lock);
 236         p = fs_info->defrag_inodes.rb_node;
 237         while (p) {
 238                 parent = p;
 239                 entry = rb_entry(parent, struct inode_defrag, rb_node);
 240
 241                 ret = __compare_inode_defrag(&tmp, entry);
 242                 if (ret < 0)
 243                         p = parent->rb_left;
 244                 else if (ret > 0)
 245                         p = parent->rb_right;
 246                 else
 247                         goto out;
 248         }
 249
 250         if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
 251                 parent = rb_next(parent);
 252                 if (parent)
 253                         entry = rb_entry(parent, struct inode_defrag, rb_node);
 254                 else
 255                         entry = NULL;
 256         }
 257 out:
 258         if (entry)
 259                 rb_erase(parent, &fs_info->defrag_inodes);
 260         spin_unlock(&fs_info->defrag_inodes_lock);
 261         return entry;
 262 }
 263
 264 void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
 265 {
 266         struct inode_defrag *defrag;
 267         struct rb_node *node;
 268
 269         spin_lock(&fs_info->defrag_inodes_lock);
 270         node = rb_first(&fs_info->defrag_inodes);
 271         while (node) {
 272                 rb_erase(node, &fs_info->defrag_inodes);
 273                 defrag = rb_entry(node, struct inode_defrag, rb_node);
 274                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 275
 276                 if (need_resched()) {
 277                         spin_unlock(&fs_info->defrag_inodes_lock);
 278                         cond_resched();
 279                         spin_lock(&fs_info->defrag_inodes_lock);
 280                 }
 281
 282                 node = rb_first(&fs_info->defrag_inodes);
 283         }
 284         spin_unlock(&fs_info->defrag_inodes_lock);
 285 }
 286
 287 #define BTRFS_DEFRAG_BATCH      1024
 288
 289 static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
 290                                     struct inode_defrag *defrag)
 291 {
 292         struct btrfs_root *inode_root;
 293         struct inode *inode;
 294         struct btrfs_key key;
 295         struct btrfs_ioctl_defrag_range_args range;
 296         int num_defrag;
 297         int index;
 298         int ret;
 299
 300         /* get the inode */
 301         key.objectid = defrag->root;
 302         btrfs_set_key_type(&key, BTRFS_ROOT_ITEM_KEY);
 303         key.offset = (u64)-1;
 304
 305         index = srcu_read_lock(&fs_info->subvol_srcu);
 306
 307         inode_root = btrfs_read_fs_root_no_name(fs_info, &key);
 308         if (IS_ERR(inode_root)) {
 309                 ret = PTR_ERR(inode_root);
 310                 goto cleanup;
 311         }
 312
 313         if (btrfs_root_refs(&inode_root->root_item) == 0) {
 314                 ret = -ENOENT;
 315                 goto cleanup;
 316         }
 317
 318         key.objectid = defrag->ino;
 319         btrfs_set_key_type(&key, BTRFS_INODE_ITEM_KEY);
 320         key.offset = 0;
 321         inode = btrfs_iget(fs_info->sb, &key, inode_root, NULL);
 322         if (IS_ERR(inode)) {
 323                 ret = PTR_ERR(inode);
 324                 goto cleanup;
 325         }
 326         srcu_read_unlock(&fs_info->subvol_srcu, index);
 327
 328         /* do a chunk of defrag */
 329         clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
 330         memset(&range, 0, sizeof(range));
 331         range.len = (u64)-1;
 332         range.start = defrag->last_offset;
 333
 334         sb_start_write(fs_info->sb);
 335         num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
 336                                        BTRFS_DEFRAG_BATCH);
 337         sb_end_write(fs_info->sb);
 338         /*
 339          * if we filled the whole defrag batch, there
 340          * must be more work to do.  Queue this defrag
 341          * again
 342          */
 343         if (num_defrag == BTRFS_DEFRAG_BATCH) {
 344                 defrag->last_offset = range.start;
 345                 btrfs_requeue_inode_defrag(inode, defrag);
 346         } else if (defrag->last_offset && !defrag->cycled) {
 347                 /*
 348                  * we didn't fill our defrag batch, but
 349                  * we didn't start at zero.  Make sure we loop
 350                  * around to the start of the file.
 351                  */
 352                 defrag->last_offset = 0;
 353                 defrag->cycled = 1;
 354                 btrfs_requeue_inode_defrag(inode, defrag);
 355         } else {
 356                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 357         }
 358
 359         iput(inode);
 360         return 0;
 361 cleanup:
 362         srcu_read_unlock(&fs_info->subvol_srcu, index);
 363         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 364         return ret;
 365 }
 366
 367 /*
 368  * run through the list of inodes in the FS that need
 369  * defragging
 370  */
 371 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
 372 {
 373         struct inode_defrag *defrag;
 374         u64 first_ino = 0;
 375         u64 root_objectid = 0;
 376
 377         atomic_inc(&fs_info->defrag_running);
 378         while(1) {
 379                 /* Pause the auto defragger. */
 380                 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
 381                              &fs_info->fs_state))
 382                         break;
 383
 384                 if (!__need_auto_defrag(fs_info->tree_root))
 385                         break;
 386
 387                 /* find an inode to defrag */
 388                 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
 389                                                  first_ino);
 390                 if (!defrag) {
 391                         if (root_objectid || first_ino) {
 392                                 root_objectid = 0;
 393                                 first_ino = 0;
 394                                 continue;
 395                         } else {
 396                                 break;
 397                         }
 398                 }
 399
 400                 first_ino = defrag->ino + 1;
 401                 root_objectid = defrag->root;
 402
 403                 __btrfs_run_defrag_inode(fs_info, defrag);
 404         }
 405         atomic_dec(&fs_info->defrag_running);
 406
 407         /*
 408          * during unmount, we use the transaction_wait queue to
 409          * wait for the defragger to stop
 410          */
 411         wake_up(&fs_info->transaction_wait);
 412         return 0;
 413 }
 414
 415 /* simple helper to fault in pages and copy.  This should go away
 416  * and be replaced with calls into generic code.
 417  */
 418 static noinline int btrfs_copy_from_user(loff_t pos, int num_pages,
 419                                          size_t write_bytes,
 420                                          struct page **prepared_pages,
 421                                          struct iov_iter *i)
 422 {
 423         size_t copied = 0;
 424         size_t total_copied = 0;
 425         int pg = 0;
 426         int offset = pos & (PAGE_CACHE_SIZE - 1);
 427
 428         while (write_bytes > 0) {
 429                 size_t count = min_t(size_t,
 430                                      PAGE_CACHE_SIZE - offset, write_bytes);
 431                 struct page *page = prepared_pages[pg];
 432                 /*
 433                  * Copy data from userspace to the current page
 434                  *
 435                  * Disable pagefault to avoid recursive lock since
 436                  * the pages are already locked
 437                  */
 438                 pagefault_disable();
 439                 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
 440                 pagefault_enable();
 441
 442                 /* Flush processor's dcache for this page */
 443                 flush_dcache_page(page);
 444
 445                 /*
 446                  * if we get a partial write, we can end up with
 447                  * partially up to date pages.  These add
 448                  * a lot of complexity, so make sure they don't
 449                  * happen by forcing this copy to be retried.
 450                  *
 451                  * The rest of the btrfs_file_write code will fall
 452                  * back to page at a time copies after we return 0.
 453                  */
 454                 if (!PageUptodate(page) && copied < count)
 455                         copied = 0;
 456
 457                 iov_iter_advance(i, copied);
 458                 write_bytes -= copied;
 459                 total_copied += copied;
 460
 461                 /* Return to btrfs_file_aio_write to fault page */
 462                 if (unlikely(copied == 0))
 463                         break;
 464
 465                 if (unlikely(copied < PAGE_CACHE_SIZE - offset)) {
 466                         offset += copied;
 467                 } else {
 468                         pg++;
 469                         offset = 0;
 470                 }
 471         }
 472         return total_copied;
 473 }
 474
 475 /*
 476  * unlocks pages after btrfs_file_write is done with them
 477  */
 478 static void btrfs_drop_pages(struct page **pages, size_t num_pages)
 479 {
 480         size_t i;
 481         for (i = 0; i < num_pages; i++) {
 482                 /* page checked is some magic around finding pages that
 483                  * have been modified without going through btrfs_set_page_dirty
 484                  * clear it here
 485                  */
 486                 ClearPageChecked(pages[i]);
 487                 unlock_page(pages[i]);
 488                 mark_page_accessed(pages[i]);
 489                 page_cache_release(pages[i]);
 490         }
 491 }
 492
 493 /*
 494  * after copy_from_user, pages need to be dirtied and we need to make
 495  * sure holes are created between the current EOF and the start of
 496  * any next extents (if required).
 497  *
 498  * this also makes the decision about creating an inline extent vs
 499  * doing real data extents, marking pages dirty and delalloc as required.
 500  */
 501 int btrfs_dirty_pages(struct btrfs_root *root, struct inode *inode,
 502                              struct page **pages, size_t num_pages,
 503                              loff_t pos, size_t write_bytes,
 504                              struct extent_state **cached)
 505 {
 506         int err = 0;
 507         int i;
 508         u64 num_bytes;
 509         u64 start_pos;
 510         u64 end_of_last_block;
 511         u64 end_pos = pos + write_bytes;
 512         loff_t isize = i_size_read(inode);
 513
 514         start_pos = pos & ~((u64)root->sectorsize - 1);
 515         num_bytes = ALIGN(write_bytes + pos - start_pos, root->sectorsize);
 516
 517         end_of_last_block = start_pos + num_bytes - 1;
 518         err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
 519                                         cached);
 520         if (err)
 521                 return err;
 522
 523         for (i = 0; i < num_pages; i++) {
 524                 struct page *p = pages[i];
 525                 SetPageUptodate(p);
 526                 ClearPageChecked(p);
 527                 set_page_dirty(p);
 528         }
 529
 530         /*
 531          * we've only changed i_size in ram, and we haven't updated
 532          * the disk i_size.  There is no need to log the inode
 533          * at this time.
 534          */
 535         if (end_pos > isize)
 536                 i_size_write(inode, end_pos);
 537         return 0;
 538 }
 539
 540 /*
 541  * this drops all the extents in the cache that intersect the range
 542  * [start, end].  Existing extents are split as required.
 543  */
 544 void btrfs_drop_extent_cache(struct inode *inode, u64 start, u64 end,
 545                              int skip_pinned)
 546 {
 547         struct extent_map *em;
 548         struct extent_map *split = NULL;
 549         struct extent_map *split2 = NULL;
 550         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 551         u64 len = end - start + 1;
 552         u64 gen;
 553         int ret;
 554         int testend = 1;
 555         unsigned long flags;
 556         int compressed = 0;
 557         bool modified;
 558
 559         WARN_ON(end < start);
 560         if (end == (u64)-1) {
 561                 len = (u64)-1;
 562                 testend = 0;
 563         }
 564         while (1) {
 565                 int no_splits = 0;
 566
 567                 modified = false;
 568                 if (!split)
 569                         split = alloc_extent_map();
 570                 if (!split2)
 571                         split2 = alloc_extent_map();
 572                 if (!split || !split2)
 573                         no_splits = 1;
 574
 575                 write_lock(&em_tree->lock);
 576                 em = lookup_extent_mapping(em_tree, start, len);
 577                 if (!em) {
 578                         write_unlock(&em_tree->lock);
 579                         break;
 580                 }
 581                 flags = em->flags;
 582                 gen = em->generation;
 583                 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
 584                         if (testend && em->start + em->len >= start + len) {
 585                                 free_extent_map(em);
 586                                 write_unlock(&em_tree->lock);
 587                                 break;
 588                         }
 589                         start = em->start + em->len;
 590                         if (testend)
 591                                 len = start + len - (em->start + em->len);
 592                         free_extent_map(em);
 593                         write_unlock(&em_tree->lock);
 594                         continue;
 595                 }
 596                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 597                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
 598                 clear_bit(EXTENT_FLAG_LOGGING, &flags);
 599                 modified = !list_empty(&em->list);
 600                 remove_extent_mapping(em_tree, em);
 601                 if (no_splits)
 602                         goto next;
 603
 604                 if (em->start < start) {
 605                         split->start = em->start;
 606                         split->len = start - em->start;
 607
 608                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 609                                 split->orig_start = em->orig_start;
 610                                 split->block_start = em->block_start;
 611
 612                                 if (compressed)
 613                                         split->block_len = em->block_len;
 614                                 else
 615                                         split->block_len = split->len;
 616                                 split->orig_block_len = max(split->block_len,
 617                                                 em->orig_block_len);
 618                                 split->ram_bytes = em->ram_bytes;
 619                         } else {
 620                                 split->orig_start = split->start;
 621                                 split->block_len = 0;
 622                                 split->block_start = em->block_start;
 623                                 split->orig_block_len = 0;
 624                                 split->ram_bytes = split->len;
 625                         }
 626
 627                         split->generation = gen;
 628                         split->bdev = em->bdev;
 629                         split->flags = flags;
 630                         split->compress_type = em->compress_type;
 631                         ret = add_extent_mapping(em_tree, split, modified);
 632                         BUG_ON(ret); /* Logic error */
 633                         free_extent_map(split);
 634                         split = split2;
 635                         split2 = NULL;
 636                 }
 637                 if (testend && em->start + em->len > start + len) {
 638                         u64 diff = start + len - em->start;
 639
 640                         split->start = start + len;
 641                         split->len = em->start + em->len - (start + len);
 642                         split->bdev = em->bdev;
 643                         split->flags = flags;
 644                         split->compress_type = em->compress_type;
 645                         split->generation = gen;
 646
 647                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 648                                 split->orig_block_len = max(em->block_len,
 649                                                     em->orig_block_len);
 650
 651                                 split->ram_bytes = em->ram_bytes;
 652                                 if (compressed) {
 653                                         split->block_len = em->block_len;
 654                                         split->block_start = em->block_start;
 655                                         split->orig_start = em->orig_start;
 656                                 } else {
 657                                         split->block_len = split->len;
 658                                         split->block_start = em->block_start
 659                                                 + diff;
 660                                         split->orig_start = em->orig_start;
 661                                 }
 662                         } else {
 663                                 split->ram_bytes = split->len;
 664                                 split->orig_start = split->start;
 665                                 split->block_len = 0;
 666                                 split->block_start = em->block_start;
 667                                 split->orig_block_len = 0;
 668                         }
 669
 670                         ret = add_extent_mapping(em_tree, split, modified);
 671                         BUG_ON(ret); /* Logic error */
 672                         free_extent_map(split);
 673                         split = NULL;
 674                 }
 675 next:
 676                 write_unlock(&em_tree->lock);
 677
 678                 /* once for us */
 679                 free_extent_map(em);
 680                 /* once for the tree*/
 681                 free_extent_map(em);
 682         }
 683         if (split)
 684                 free_extent_map(split);
 685         if (split2)
 686                 free_extent_map(split2);
 687 }
 688
 689 /*
 690  * this is very complex, but the basic idea is to drop all extents
 691  * in the range start - end.  hint_block is filled in with a block number
 692  * that would be a good hint to the block allocator for this file.
 693  *
 694  * If an extent intersects the range but is not entirely inside the range
 695  * it is either truncated or split.  Anything entirely inside the range
 696  * is deleted from the tree.
 697  */
 698 int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
 699                          struct btrfs_root *root, struct inode *inode,
 700                          struct btrfs_path *path, u64 start, u64 end,
 701                          u64 *drop_end, int drop_cache)
 702 {
 703         struct extent_buffer *leaf;
 704         struct btrfs_file_extent_item *fi;
 705         struct btrfs_key key;
 706         struct btrfs_key new_key;
 707         u64 ino = btrfs_ino(inode);
 708         u64 search_start = start;
 709         u64 disk_bytenr = 0;
 710         u64 num_bytes = 0;
 711         u64 extent_offset = 0;
 712         u64 extent_end = 0;
 713         int del_nr = 0;
 714         int del_slot = 0;
 715         int extent_type;
 716         int recow;
 717         int ret;
 718         int modify_tree = -1;
 719         int update_refs = (root->ref_cows || root == root->fs_info->tree_root);
 720         int found = 0;
 721
 722         if (drop_cache)
 723                 btrfs_drop_extent_cache(inode, start, end - 1, 0);
 724
 725         if (start >= BTRFS_I(inode)->disk_i_size)
 726                 modify_tree = 0;
 727
 728         while (1) {
 729                 recow = 0;
 730                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
 731                                                search_start, modify_tree);
 732                 if (ret < 0)
 733                         break;
 734                 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
 735                         leaf = path->nodes[0];
 736                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
 737                         if (key.objectid == ino &&
 738                             key.type == BTRFS_EXTENT_DATA_KEY)
 739                                 path->slots[0]--;
 740                 }
 741                 ret = 0;
 742 next_slot:
 743                 leaf = path->nodes[0];
 744                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 745                         BUG_ON(del_nr > 0);
 746                         ret = btrfs_next_leaf(root, path);
 747                         if (ret < 0)
 748                                 break;
 749                         if (ret > 0) {
 750                                 ret = 0;
 751                                 break;
 752                         }
 753                         leaf = path->nodes[0];
 754                         recow = 1;
 755                 }
 756
 757                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 758                 if (key.objectid > ino ||
 759                     key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
 760                         break;
 761
 762                 fi = btrfs_item_ptr(leaf, path->slots[0],
 763                                     struct btrfs_file_extent_item);
 764                 extent_type = btrfs_file_extent_type(leaf, fi);
 765
 766                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
 767                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
 768                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 769                         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 770                         extent_offset = btrfs_file_extent_offset(leaf, fi);
 771                         extent_end = key.offset +
 772                                 btrfs_file_extent_num_bytes(leaf, fi);
 773                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 774                         extent_end = key.offset +
 775                                 btrfs_file_extent_inline_len(leaf, fi);
 776                 } else {
 777                         WARN_ON(1);
 778                         extent_end = search_start;
 779                 }
 780
 781                 if (extent_end <= search_start) {
 782                         path->slots[0]++;
 783                         goto next_slot;
 784                 }
 785
 786                 found = 1;
 787                 search_start = max(key.offset, start);
 788                 if (recow || !modify_tree) {
 789                         modify_tree = -1;
 790                         btrfs_release_path(path);
 791                         continue;
 792                 }
 793
 794                 /*
 795                  *     | - range to drop - |
 796                  *  | -------- extent -------- |
 797                  */
 798                 if (start > key.offset && end < extent_end) {
 799                         BUG_ON(del_nr > 0);
 800                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
 801
 802                         memcpy(&new_key, &key, sizeof(new_key));
 803                         new_key.offset = start;
 804                         ret = btrfs_duplicate_item(trans, root, path,
 805                                                    &new_key);
 806                         if (ret == -EAGAIN) {
 807                                 btrfs_release_path(path);
 808                                 continue;
 809                         }
 810                         if (ret < 0)
 811                                 break;
 812
 813                         leaf = path->nodes[0];
 814                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 815                                             struct btrfs_file_extent_item);
 816                         btrfs_set_file_extent_num_bytes(leaf, fi,
 817                                                         start - key.offset);
 818
 819                         fi = btrfs_item_ptr(leaf, path->slots[0],
 820                                             struct btrfs_file_extent_item);
 821
 822                         extent_offset += start - key.offset;
 823                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 824                         btrfs_set_file_extent_num_bytes(leaf, fi,
 825                                                         extent_end - start);
 826                         btrfs_mark_buffer_dirty(leaf);
 827
 828                         if (update_refs && disk_bytenr > 0) {
 829                                 ret = btrfs_inc_extent_ref(trans, root,
 830                                                 disk_bytenr, num_bytes, 0,
 831                                                 root->root_key.objectid,
 832                                                 new_key.objectid,
 833                                                 start - extent_offset, 0);
 834                                 BUG_ON(ret); /* -ENOMEM */
 835                         }
 836                         key.offset = start;
 837                 }
 838                 /*
 839                  *  | ---- range to drop ----- |
 840                  *      | -------- extent -------- |
 841                  */
 842                 if (start <= key.offset && end < extent_end) {
 843                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
 844
 845                         memcpy(&new_key, &key, sizeof(new_key));
 846                         new_key.offset = end;
 847                         btrfs_set_item_key_safe(root, path, &new_key);
 848
 849                         extent_offset += end - key.offset;
 850                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 851                         btrfs_set_file_extent_num_bytes(leaf, fi,
 852                                                         extent_end - end);
 853                         btrfs_mark_buffer_dirty(leaf);
 854                         if (update_refs && disk_bytenr > 0)
 855                                 inode_sub_bytes(inode, end - key.offset);
 856                         break;
 857                 }
 858
 859                 search_start = extent_end;
 860                 /*
 861                  *       | ---- range to drop ----- |
 862                  *  | -------- extent -------- |
 863                  */
 864                 if (start > key.offset && end >= extent_end) {
 865                         BUG_ON(del_nr > 0);
 866                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
 867
 868                         btrfs_set_file_extent_num_bytes(leaf, fi,
 869                                                         start - key.offset);
 870                         btrfs_mark_buffer_dirty(leaf);
 871                         if (update_refs && disk_bytenr > 0)
 872                                 inode_sub_bytes(inode, extent_end - start);
 873                         if (end == extent_end)
 874                                 break;
 875
 876                         path->slots[0]++;
 877                         goto next_slot;
 878                 }
 879
 880                 /*
 881                  *  | ---- range to drop ----- |
 882                  *    | ------ extent ------ |
 883                  */
 884                 if (start <= key.offset && end >= extent_end) {
 885                         if (del_nr == 0) {
 886                                 del_slot = path->slots[0];
 887                                 del_nr = 1;
 888                         } else {
 889                                 BUG_ON(del_slot + del_nr != path->slots[0]);
 890                                 del_nr++;
 891                         }
 892
 893                         if (update_refs &&
 894                             extent_type == BTRFS_FILE_EXTENT_INLINE) {
 895                                 inode_sub_bytes(inode,
 896                                                 extent_end - key.offset);
 897                                 extent_end = ALIGN(extent_end,
 898                                                    root->sectorsize);
 899                         } else if (update_refs && disk_bytenr > 0) {
 900                                 ret = btrfs_free_extent(trans, root,
 901                                                 disk_bytenr, num_bytes, 0,
 902                                                 root->root_key.objectid,
 903                                                 key.objectid, key.offset -
 904                                                 extent_offset, 0);
 905                                 BUG_ON(ret); /* -ENOMEM */
 906                                 inode_sub_bytes(inode,
 907                                                 extent_end - key.offset);
 908                         }
 909
 910                         if (end == extent_end)
 911                                 break;
 912
 913                         if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
 914                                 path->slots[0]++;
 915                                 goto next_slot;
 916                         }
 917
 918                         ret = btrfs_del_items(trans, root, path, del_slot,
 919                                               del_nr);
 920                         if (ret) {
 921                                 btrfs_abort_transaction(trans, root, ret);
 922                                 break;
 923                         }
 924
 925                         del_nr = 0;
 926                         del_slot = 0;
 927
 928                         btrfs_release_path(path);
 929                         continue;
 930                 }
 931
 932                 BUG_ON(1);
 933         }
 934
 935         if (!ret && del_nr > 0) {
 936                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
 937                 if (ret)
 938                         btrfs_abort_transaction(trans, root, ret);
 939         }
 940
 941         if (drop_end)
 942                 *drop_end = found ? min(end, extent_end) : end;
 943         btrfs_release_path(path);
 944         return ret;
 945 }
 946
 947 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
 948                        struct btrfs_root *root, struct inode *inode, u64 start,
 949                        u64 end, int drop_cache)
 950 {
 951         struct btrfs_path *path;
 952         int ret;
 953
 954         path = btrfs_alloc_path();
 955         if (!path)
 956                 return -ENOMEM;
 957         ret = __btrfs_drop_extents(trans, root, inode, path, start, end, NULL,
 958                                    drop_cache);
 959         btrfs_free_path(path);
 960         return ret;
 961 }
 962
 963 static int extent_mergeable(struct extent_buffer *leaf, int slot,
 964                             u64 objectid, u64 bytenr, u64 orig_offset,
 965                             u64 *start, u64 *end)
 966 {
 967         struct btrfs_file_extent_item *fi;
 968         struct btrfs_key key;
 969         u64 extent_end;
 970
 971         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
 972                 return 0;
 973
 974         btrfs_item_key_to_cpu(leaf, &key, slot);
 975         if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
 976                 return 0;
 977
 978         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
 979         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
 980             btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
 981             btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
 982             btrfs_file_extent_compression(leaf, fi) ||
 983             btrfs_file_extent_encryption(leaf, fi) ||
 984             btrfs_file_extent_other_encoding(leaf, fi))
 985                 return 0;
 986
 987         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
 988         if ((*start && *start != key.offset) || (*end && *end != extent_end))
 989                 return 0;
 990
 991         *start = key.offset;
 992         *end = extent_end;
 993         return 1;
 994 }
 995
 996 /*
 997  * Mark extent in the range start - end as written.
 998  *
 999  * This changes extent type from 'pre-allocated' to 'regular'. If only
1000  * part of extent is marked as written, the extent will be split into
1001  * two or three.
1002  */
1003 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1004                               struct inode *inode, u64 start, u64 end)
1005 {
1006         struct btrfs_root *root = BTRFS_I(inode)->root;
1007         struct extent_buffer *leaf;
1008         struct btrfs_path *path;
1009         struct btrfs_file_extent_item *fi;
1010         struct btrfs_key key;
1011         struct btrfs_key new_key;
1012         u64 bytenr;
1013         u64 num_bytes;
1014         u64 extent_end;
1015         u64 orig_offset;
1016         u64 other_start;
1017         u64 other_end;
1018         u64 split;
1019         int del_nr = 0;
1020         int del_slot = 0;
1021         int recow;
1022         int ret;
1023         u64 ino = btrfs_ino(inode);
1024
1025         path = btrfs_alloc_path();
1026         if (!path)
1027                 return -ENOMEM;
1028 again:
1029         recow = 0;
1030         split = start;
1031         key.objectid = ino;
1032         key.type = BTRFS_EXTENT_DATA_KEY;
1033         key.offset = split;
1034
1035         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1036         if (ret < 0)
1037                 goto out;
1038         if (ret > 0 && path->slots[0] > 0)
1039                 path->slots[0]--;
1040
1041         leaf = path->nodes[0];
1042         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1043         BUG_ON(key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY);
1044         fi = btrfs_item_ptr(leaf, path->slots[0],
1045                             struct btrfs_file_extent_item);
1046         BUG_ON(btrfs_file_extent_type(leaf, fi) !=
1047                BTRFS_FILE_EXTENT_PREALLOC);
1048         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1049         BUG_ON(key.offset > start || extent_end < end);
1050
1051         bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1052         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1053         orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1054         memcpy(&new_key, &key, sizeof(new_key));
1055
1056         if (start == key.offset && end < extent_end) {
1057                 other_start = 0;
1058                 other_end = start;
1059                 if (extent_mergeable(leaf, path->slots[0] - 1,
1060                                      ino, bytenr, orig_offset,
1061                                      &other_start, &other_end)) {
1062                         new_key.offset = end;
1063                         btrfs_set_item_key_safe(root, path, &new_key);
1064                         fi = btrfs_item_ptr(leaf, path->slots[0],
1065                                             struct btrfs_file_extent_item);
1066                         btrfs_set_file_extent_generation(leaf, fi,
1067                                                          trans->transid);
1068                         btrfs_set_file_extent_num_bytes(leaf, fi,
1069                                                         extent_end - end);
1070                         btrfs_set_file_extent_offset(leaf, fi,
1071                                                      end - orig_offset);
1072                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1073                                             struct btrfs_file_extent_item);
1074                         btrfs_set_file_extent_generation(leaf, fi,
1075                                                          trans->transid);
1076                         btrfs_set_file_extent_num_bytes(leaf, fi,
1077                                                         end - other_start);
1078                         btrfs_mark_buffer_dirty(leaf);
1079                         goto out;
1080                 }
1081         }
1082
1083         if (start > key.offset && end == extent_end) {
1084                 other_start = end;
1085                 other_end = 0;
1086                 if (extent_mergeable(leaf, path->slots[0] + 1,
1087                                      ino, bytenr, orig_offset,
1088                                      &other_start, &other_end)) {
1089                         fi = btrfs_item_ptr(leaf, path->slots[0],
1090                                             struct btrfs_file_extent_item);
1091                         btrfs_set_file_extent_num_bytes(leaf, fi,
1092                                                         start - key.offset);
1093                         btrfs_set_file_extent_generation(leaf, fi,
1094                                                          trans->transid);
1095                         path->slots[0]++;
1096                         new_key.offset = start;
1097                         btrfs_set_item_key_safe(root, path, &new_key);
1098
1099                         fi = btrfs_item_ptr(leaf, path->slots[0],
1100                                             struct btrfs_file_extent_item);
1101                         btrfs_set_file_extent_generation(leaf, fi,
1102                                                          trans->transid);
1103                         btrfs_set_file_extent_num_bytes(leaf, fi,
1104                                                         other_end - start);
1105                         btrfs_set_file_extent_offset(leaf, fi,
1106                                                      start - orig_offset);
1107                         btrfs_mark_buffer_dirty(leaf);
1108                         goto out;
1109                 }
1110         }
1111
1112         while (start > key.offset || end < extent_end) {
1113                 if (key.offset == start)
1114                         split = end;
1115
1116                 new_key.offset = split;
1117                 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1118                 if (ret == -EAGAIN) {
1119                         btrfs_release_path(path);
1120                         goto again;
1121                 }
1122                 if (ret < 0) {
1123                         btrfs_abort_transaction(trans, root, ret);
1124                         goto out;
1125                 }
1126
1127                 leaf = path->nodes[0];
1128                 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1129                                     struct btrfs_file_extent_item);
1130                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1131                 btrfs_set_file_extent_num_bytes(leaf, fi,
1132                                                 split - key.offset);
1133
1134                 fi = btrfs_item_ptr(leaf, path->slots[0],
1135                                     struct btrfs_file_extent_item);
1136
1137                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1138                 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1139                 btrfs_set_file_extent_num_bytes(leaf, fi,
1140                                                 extent_end - split);
1141                 btrfs_mark_buffer_dirty(leaf);
1142
1143                 ret = btrfs_inc_extent_ref(trans, root, bytenr, num_bytes, 0,
1144                                            root->root_key.objectid,
1145                                            ino, orig_offset, 0);
1146                 BUG_ON(ret); /* -ENOMEM */
1147
1148                 if (split == start) {
1149                         key.offset = start;
1150                 } else {
1151                         BUG_ON(start != key.offset);
1152                         path->slots[0]--;
1153                         extent_end = end;
1154                 }
1155                 recow = 1;
1156         }
1157
1158         other_start = end;
1159         other_end = 0;
1160         if (extent_mergeable(leaf, path->slots[0] + 1,
1161                              ino, bytenr, orig_offset,
1162                              &other_start, &other_end)) {
1163                 if (recow) {
1164                         btrfs_release_path(path);
1165                         goto again;
1166                 }
1167                 extent_end = other_end;
1168                 del_slot = path->slots[0] + 1;
1169                 del_nr++;
1170                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1171                                         0, root->root_key.objectid,
1172                                         ino, orig_offset, 0);
1173                 BUG_ON(ret); /* -ENOMEM */
1174         }
1175         other_start = 0;
1176         other_end = start;
1177         if (extent_mergeable(leaf, path->slots[0] - 1,
1178                              ino, bytenr, orig_offset,
1179                              &other_start, &other_end)) {
1180                 if (recow) {
1181                         btrfs_release_path(path);
1182                         goto again;
1183                 }
1184                 key.offset = other_start;
1185                 del_slot = path->slots[0];
1186                 del_nr++;
1187                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1188                                         0, root->root_key.objectid,
1189                                         ino, orig_offset, 0);
1190                 BUG_ON(ret); /* -ENOMEM */
1191         }
1192         if (del_nr == 0) {
1193                 fi = btrfs_item_ptr(leaf, path->slots[0],
1194                            struct btrfs_file_extent_item);
1195                 btrfs_set_file_extent_type(leaf, fi,
1196                                            BTRFS_FILE_EXTENT_REG);
1197                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1198                 btrfs_mark_buffer_dirty(leaf);
1199         } else {
1200                 fi = btrfs_item_ptr(leaf, del_slot - 1,
1201                            struct btrfs_file_extent_item);
1202                 btrfs_set_file_extent_type(leaf, fi,
1203                                            BTRFS_FILE_EXTENT_REG);
1204                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1205                 btrfs_set_file_extent_num_bytes(leaf, fi,
1206                                                 extent_end - key.offset);
1207                 btrfs_mark_buffer_dirty(leaf);
1208
1209                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1210                 if (ret < 0) {
1211                         btrfs_abort_transaction(trans, root, ret);
1212                         goto out;
1213                 }
1214         }
1215 out:
1216         btrfs_free_path(path);
1217         return 0;
1218 }
1219
1220 /*
1221  * on error we return an unlocked page and the error value
1222  * on success we return a locked page and 0
1223  */
1224 static int prepare_uptodate_page(struct page *page, u64 pos,
1225                                  bool force_uptodate)
1226 {
1227         int ret = 0;
1228
1229         if (((pos & (PAGE_CACHE_SIZE - 1)) || force_uptodate) &&
1230             !PageUptodate(page)) {
1231                 ret = btrfs_readpage(NULL, page);
1232                 if (ret)
1233                         return ret;
1234                 lock_page(page);
1235                 if (!PageUptodate(page)) {
1236                         unlock_page(page);
1237                         return -EIO;
1238                 }
1239         }
1240         return 0;
1241 }
1242
1243 /*
1244  * this gets pages into the page cache and locks them down, it also properly
1245  * waits for data=ordered extents to finish before allowing the pages to be
1246  * modified.
1247  */
1248 static noinline int prepare_pages(struct btrfs_root *root, struct file *file,
1249                          struct page **pages, size_t num_pages,
1250                          loff_t pos, unsigned long first_index,
1251                          size_t write_bytes, bool force_uptodate)
1252 {
1253         struct extent_state *cached_state = NULL;
1254         int i;
1255         unsigned long index = pos >> PAGE_CACHE_SHIFT;
1256         struct inode *inode = file_inode(file);
1257         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1258         int err = 0;
1259         int faili = 0;
1260         u64 start_pos;
1261         u64 last_pos;
1262
1263         start_pos = pos & ~((u64)root->sectorsize - 1);
1264         last_pos = ((u64)index + num_pages) << PAGE_CACHE_SHIFT;
1265
1266 again:
1267         for (i = 0; i < num_pages; i++) {
1268                 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1269                                                mask | __GFP_WRITE);
1270                 if (!pages[i]) {
1271                         faili = i - 1;
1272                         err = -ENOMEM;
1273                         goto fail;
1274                 }
1275
1276                 if (i == 0)
1277                         err = prepare_uptodate_page(pages[i], pos,
1278                                                     force_uptodate);
1279                 if (i == num_pages - 1)
1280                         err = prepare_uptodate_page(pages[i],
1281                                                     pos + write_bytes, false);
1282                 if (err) {
1283                         page_cache_release(pages[i]);
1284                         faili = i - 1;
1285                         goto fail;
1286                 }
1287                 wait_on_page_writeback(pages[i]);
1288         }
1289         err = 0;
1290         if (start_pos < inode->i_size) {
1291                 struct btrfs_ordered_extent *ordered;
1292                 lock_extent_bits(&BTRFS_I(inode)->io_tree,
1293                                  start_pos, last_pos - 1, 0, &cached_state);
1294                 ordered = btrfs_lookup_first_ordered_extent(inode,
1295                                                             last_pos - 1);
1296                 if (ordered &&
1297                     ordered->file_offset + ordered->len > start_pos &&
1298                     ordered->file_offset < last_pos) {
1299                         btrfs_put_ordered_extent(ordered);
1300                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1301                                              start_pos, last_pos - 1,
1302                                              &cached_state, GFP_NOFS);
1303                         for (i = 0; i < num_pages; i++) {
1304                                 unlock_page(pages[i]);
1305                                 page_cache_release(pages[i]);
1306                         }
1307                         btrfs_wait_ordered_range(inode, start_pos,
1308                                                  last_pos - start_pos);
1309                         goto again;
1310                 }
1311                 if (ordered)
1312                         btrfs_put_ordered_extent(ordered);
1313
1314                 clear_extent_bit(&BTRFS_I(inode)->io_tree, start_pos,
1315                                   last_pos - 1, EXTENT_DIRTY | EXTENT_DELALLOC |
1316                                   EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
1317                                   0, 0, &cached_state, GFP_NOFS);
1318                 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1319                                      start_pos, last_pos - 1, &cached_state,
1320                                      GFP_NOFS);
1321         }
1322         for (i = 0; i < num_pages; i++) {
1323                 if (clear_page_dirty_for_io(pages[i]))
1324                         account_page_redirty(pages[i]);
1325                 set_page_extent_mapped(pages[i]);
1326                 WARN_ON(!PageLocked(pages[i]));
1327         }
1328         return 0;
1329 fail:
1330         while (faili >= 0) {
1331                 unlock_page(pages[faili]);
1332                 page_cache_release(pages[faili]);
1333                 faili--;
1334         }
1335         return err;
1336
1337 }
1338
1339 static noinline int check_can_nocow(struct inode *inode, loff_t pos,
1340                                     size_t *write_bytes)
1341 {
1342         struct btrfs_root *root = BTRFS_I(inode)->root;
1343         struct btrfs_ordered_extent *ordered;
1344         u64 lockstart, lockend;
1345         u64 num_bytes;
1346         int ret;
1347
1348         lockstart = round_down(pos, root->sectorsize);
1349         lockend = lockstart + round_up(*write_bytes, root->sectorsize) - 1;
1350
1351         while (1) {
1352                 lock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend);
1353                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
1354                                                      lockend - lockstart + 1);
1355                 if (!ordered) {
1356                         break;
1357                 }
1358                 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend);
1359                 btrfs_start_ordered_extent(inode, ordered, 1);
1360                 btrfs_put_ordered_extent(ordered);
1361         }
1362
1363         num_bytes = lockend - lockstart + 1;
1364         ret = can_nocow_extent(inode, lockstart, &num_bytes, NULL, NULL, NULL);
1365         if (ret <= 0) {
1366                 ret = 0;
1367         } else {
1368                 clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
1369                                  EXTENT_DIRTY | EXTENT_DELALLOC |
1370                                  EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 0, 0,
1371                                  NULL, GFP_NOFS);
1372                 *write_bytes = min_t(size_t, *write_bytes, num_bytes);
1373         }
1374
1375         unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend);
1376
1377         return ret;
1378 }
1379
1380 static noinline ssize_t __btrfs_buffered_write(struct file *file,
1381                                                struct iov_iter *i,
1382                                                loff_t pos)
1383 {
1384         struct inode *inode = file_inode(file);
1385         struct btrfs_root *root = BTRFS_I(inode)->root;
1386         struct page **pages = NULL;
1387         u64 release_bytes = 0;
1388         unsigned long first_index;
1389         size_t num_written = 0;
1390         int nrptrs;
1391         int ret = 0;
1392         bool only_release_metadata = false;
1393         bool force_page_uptodate = false;
1394
1395         nrptrs = min((iov_iter_count(i) + PAGE_CACHE_SIZE - 1) /
1396                      PAGE_CACHE_SIZE, PAGE_CACHE_SIZE /
1397                      (sizeof(struct page *)));
1398         nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1399         nrptrs = max(nrptrs, 8);
1400         pages = kmalloc(nrptrs * sizeof(struct page *), GFP_KERNEL);
1401         if (!pages)
1402                 return -ENOMEM;
1403
1404         first_index = pos >> PAGE_CACHE_SHIFT;
1405
1406         while (iov_iter_count(i) > 0) {
1407                 size_t offset = pos & (PAGE_CACHE_SIZE - 1);
1408                 size_t write_bytes = min(iov_iter_count(i),
1409                                          nrptrs * (size_t)PAGE_CACHE_SIZE -
1410                                          offset);
1411                 size_t num_pages = (write_bytes + offset +
1412                                     PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1413                 size_t reserve_bytes;
1414                 size_t dirty_pages;
1415                 size_t copied;
1416
1417                 WARN_ON(num_pages > nrptrs);
1418
1419                 /*
1420                  * Fault pages before locking them in prepare_pages
1421                  * to avoid recursive lock
1422                  */
1423                 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1424                         ret = -EFAULT;
1425                         break;
1426                 }
1427
1428                 reserve_bytes = num_pages << PAGE_CACHE_SHIFT;
1429                 ret = btrfs_check_data_free_space(inode, reserve_bytes);
1430                 if (ret == -ENOSPC &&
1431                     (BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1432                                               BTRFS_INODE_PREALLOC))) {
1433                         ret = check_can_nocow(inode, pos, &write_bytes);
1434                         if (ret > 0) {
1435                                 only_release_metadata = true;
1436                                 /*
1437                                  * our prealloc extent may be smaller than
1438                                  * write_bytes, so scale down.
1439                                  */
1440                                 num_pages = (write_bytes + offset +
1441                                              PAGE_CACHE_SIZE - 1) >>
1442                                         PAGE_CACHE_SHIFT;
1443                                 reserve_bytes = num_pages << PAGE_CACHE_SHIFT;
1444                                 ret = 0;
1445                         } else {
1446                                 ret = -ENOSPC;
1447                         }
1448                 }
1449
1450                 if (ret)
1451                         break;
1452
1453                 ret = btrfs_delalloc_reserve_metadata(inode, reserve_bytes);
1454                 if (ret) {
1455                         if (!only_release_metadata)
1456                                 btrfs_free_reserved_data_space(inode,
1457                                                                reserve_bytes);
1458                         break;
1459                 }
1460
1461                 release_bytes = reserve_bytes;
1462
1463                 /*
1464                  * This is going to setup the pages array with the number of
1465                  * pages we want, so we don't really need to worry about the
1466                  * contents of pages from loop to loop
1467                  */
1468                 ret = prepare_pages(root, file, pages, num_pages,
1469                                     pos, first_index, write_bytes,
1470                                     force_page_uptodate);
1471                 if (ret)
1472                         break;
1473
1474                 copied = btrfs_copy_from_user(pos, num_pages,
1475                                            write_bytes, pages, i);
1476
1477                 /*
1478                  * if we have trouble faulting in the pages, fall
1479                  * back to one page at a time
1480                  */
1481                 if (copied < write_bytes)
1482                         nrptrs = 1;
1483
1484                 if (copied == 0) {
1485                         force_page_uptodate = true;
1486                         dirty_pages = 0;
1487                 } else {
1488                         force_page_uptodate = false;
1489                         dirty_pages = (copied + offset +
1490                                        PAGE_CACHE_SIZE - 1) >>
1491                                        PAGE_CACHE_SHIFT;
1492                 }
1493
1494                 /*
1495                  * If we had a short copy we need to release the excess delaloc
1496                  * bytes we reserved.  We need to increment outstanding_extents
1497                  * because btrfs_delalloc_release_space will decrement it, but
1498                  * we still have an outstanding extent for the chunk we actually
1499                  * managed to copy.
1500                  */
1501                 if (num_pages > dirty_pages) {
1502                         release_bytes = (num_pages - dirty_pages) <<
1503                                 PAGE_CACHE_SHIFT;
1504                         if (copied > 0) {
1505                                 spin_lock(&BTRFS_I(inode)->lock);
1506                                 BTRFS_I(inode)->outstanding_extents++;
1507                                 spin_unlock(&BTRFS_I(inode)->lock);
1508                         }
1509                         if (only_release_metadata)
1510                                 btrfs_delalloc_release_metadata(inode,
1511                                                                 release_bytes);
1512                         else
1513                                 btrfs_delalloc_release_space(inode,
1514                                                              release_bytes);
1515                 }
1516
1517                 release_bytes = dirty_pages << PAGE_CACHE_SHIFT;
1518                 if (copied > 0) {
1519                         ret = btrfs_dirty_pages(root, inode, pages,
1520                                                 dirty_pages, pos, copied,
1521                                                 NULL);
1522                         if (ret) {
1523                                 btrfs_drop_pages(pages, num_pages);
1524                                 break;
1525                         }
1526                 }
1527
1528                 release_bytes = 0;
1529                 btrfs_drop_pages(pages, num_pages);
1530
1531                 if (only_release_metadata && copied > 0) {
1532                         u64 lockstart = round_down(pos, root->sectorsize);
1533                         u64 lockend = lockstart +
1534                                 (dirty_pages << PAGE_CACHE_SHIFT) - 1;
1535
1536                         set_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
1537                                        lockend, EXTENT_NORESERVE, NULL,
1538                                        NULL, GFP_NOFS);
1539                         only_release_metadata = false;
1540                 }
1541
1542                 cond_resched();
1543
1544                 balance_dirty_pages_ratelimited(inode->i_mapping);
1545                 if (dirty_pages < (root->leafsize >> PAGE_CACHE_SHIFT) + 1)
1546                         btrfs_btree_balance_dirty(root);
1547
1548                 pos += copied;
1549                 num_written += copied;
1550         }
1551
1552         kfree(pages);
1553
1554         if (release_bytes) {
1555                 if (only_release_metadata)
1556                         btrfs_delalloc_release_metadata(inode, release_bytes);
1557                 else
1558                         btrfs_delalloc_release_space(inode, release_bytes);
1559         }
1560
1561         return num_written ? num_written : ret;
1562 }
1563
1564 static ssize_t __btrfs_direct_write(struct kiocb *iocb,
1565                                     const struct iovec *iov,
1566                                     unsigned long nr_segs, loff_t pos,
1567                                     loff_t *ppos, size_t count, size_t ocount)
1568 {
1569         struct file *file = iocb->ki_filp;
1570         struct iov_iter i;
1571         ssize_t written;
1572         ssize_t written_buffered;
1573         loff_t endbyte;
1574         int err;
1575
1576         written = generic_file_direct_write(iocb, iov, &nr_segs, pos, ppos,
1577                                             count, ocount);
1578
1579         if (written < 0 || written == count)
1580                 return written;
1581
1582         pos += written;
1583         count -= written;
1584         iov_iter_init(&i, iov, nr_segs, count, written);
1585         written_buffered = __btrfs_buffered_write(file, &i, pos);
1586         if (written_buffered < 0) {
1587                 err = written_buffered;
1588                 goto out;
1589         }
1590         endbyte = pos + written_buffered - 1;
1591         err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
1592         if (err)
1593                 goto out;
1594         written += written_buffered;
1595         *ppos = pos + written_buffered;
1596         invalidate_mapping_pages(file->f_mapping, pos >> PAGE_CACHE_SHIFT,
1597                                  endbyte >> PAGE_CACHE_SHIFT);
1598 out:
1599         return written ? written : err;
1600 }
1601
1602 static void update_time_for_write(struct inode *inode)
1603 {
1604         struct timespec now;
1605
1606         if (IS_NOCMTIME(inode))
1607                 return;
1608
1609         now = current_fs_time(inode->i_sb);
1610         if (!timespec_equal(&inode->i_mtime, &now))
1611                 inode->i_mtime = now;
1612
1613         if (!timespec_equal(&inode->i_ctime, &now))
1614                 inode->i_ctime = now;
1615
1616         if (IS_I_VERSION(inode))
1617                 inode_inc_iversion(inode);
1618 }
1619
1620 static ssize_t btrfs_file_aio_write(struct kiocb *iocb,
1621                                     const struct iovec *iov,
1622                                     unsigned long nr_segs, loff_t pos)
1623 {
1624         struct file *file = iocb->ki_filp;
1625         struct inode *inode = file_inode(file);
1626         struct btrfs_root *root = BTRFS_I(inode)->root;
1627         loff_t *ppos = &iocb->ki_pos;
1628         u64 start_pos;
1629         ssize_t num_written = 0;
1630         ssize_t err = 0;
1631         size_t count, ocount;
1632         bool sync = (file->f_flags & O_DSYNC) || IS_SYNC(file->f_mapping->host);
1633
1634         mutex_lock(&inode->i_mutex);
1635
1636         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
1637         if (err) {
1638                 mutex_unlock(&inode->i_mutex);
1639                 goto out;
1640         }
1641         count = ocount;
1642
1643         current->backing_dev_info = inode->i_mapping->backing_dev_info;
1644         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
1645         if (err) {
1646                 mutex_unlock(&inode->i_mutex);
1647                 goto out;
1648         }
1649
1650         if (count == 0) {
1651                 mutex_unlock(&inode->i_mutex);
1652                 goto out;
1653         }
1654
1655         err = file_remove_suid(file);
1656         if (err) {
1657                 mutex_unlock(&inode->i_mutex);
1658                 goto out;
1659         }
1660
1661         /*
1662          * If BTRFS flips readonly due to some impossible error
1663          * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1664          * although we have opened a file as writable, we have
1665          * to stop this write operation to ensure FS consistency.
1666          */
1667         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state)) {
1668                 mutex_unlock(&inode->i_mutex);
1669                 err = -EROFS;
1670                 goto out;
1671         }
1672
1673         /*
1674          * We reserve space for updating the inode when we reserve space for the
1675          * extent we are going to write, so we will enospc out there.  We don't
1676          * need to start yet another transaction to update the inode as we will
1677          * update the inode when we finish writing whatever data we write.
1678          */
1679         update_time_for_write(inode);
1680
1681         start_pos = round_down(pos, root->sectorsize);
1682         if (start_pos > i_size_read(inode)) {
1683                 err = btrfs_cont_expand(inode, i_size_read(inode), start_pos);
1684                 if (err) {
1685                         mutex_unlock(&inode->i_mutex);
1686                         goto out;
1687                 }
1688         }
1689
1690         if (sync)
1691                 atomic_inc(&BTRFS_I(inode)->sync_writers);
1692
1693         if (unlikely(file->f_flags & O_DIRECT)) {
1694                 num_written = __btrfs_direct_write(iocb, iov, nr_segs,
1695                                                    pos, ppos, count, ocount);
1696         } else {
1697                 struct iov_iter i;
1698
1699                 iov_iter_init(&i, iov, nr_segs, count, num_written);
1700
1701                 num_written = __btrfs_buffered_write(file, &i, pos);
1702                 if (num_written > 0)
1703                         *ppos = pos + num_written;
1704         }
1705
1706         mutex_unlock(&inode->i_mutex);
1707
1708         /*
1709          * we want to make sure fsync finds this change
1710          * but we haven't joined a transaction running right now.
1711          *
1712          * Later on, someone is sure to update the inode and get the
1713          * real transid recorded.
1714          *
1715          * We set last_trans now to the fs_info generation + 1,
1716          * this will either be one more than the running transaction
1717          * or the generation used for the next transaction if there isn't
1718          * one running right now.
1719          *
1720          * We also have to set last_sub_trans to the current log transid,
1721          * otherwise subsequent syncs to a file that's been synced in this
1722          * transaction will appear to have already occured.
1723          */
1724         BTRFS_I(inode)->last_trans = root->fs_info->generation + 1;
1725         BTRFS_I(inode)->last_sub_trans = root->log_transid;
1726         if (num_written > 0 || num_written == -EIOCBQUEUED) {
1727                 err = generic_write_sync(file, pos, num_written);
1728                 if (err < 0 && num_written > 0)
1729                         num_written = err;
1730         }
1731
1732         if (sync)
1733                 atomic_dec(&BTRFS_I(inode)->sync_writers);
1734 out:
1735         current->backing_dev_info = NULL;
1736         return num_written ? num_written : err;
1737 }
1738
1739 int btrfs_release_file(struct inode *inode, struct file *filp)
1740 {
1741         /*
1742          * ordered_data_close is set by settattr when we are about to truncate
1743          * a file from a non-zero size to a zero size.  This tries to
1744          * flush down new bytes that may have been written if the
1745          * application were using truncate to replace a file in place.
1746          */
1747         if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
1748                                &BTRFS_I(inode)->runtime_flags)) {
1749                 struct btrfs_trans_handle *trans;
1750                 struct btrfs_root *root = BTRFS_I(inode)->root;
1751
1752                 /*
1753                  * We need to block on a committing transaction to keep us from
1754                  * throwing a ordered operation on to the list and causing
1755                  * something like sync to deadlock trying to flush out this
1756                  * inode.
1757                  */
1758                 trans = btrfs_start_transaction(root, 0);
1759                 if (IS_ERR(trans))
1760                         return PTR_ERR(trans);
1761                 btrfs_add_ordered_operation(trans, BTRFS_I(inode)->root, inode);
1762                 btrfs_end_transaction(trans, root);
1763                 if (inode->i_size > BTRFS_ORDERED_OPERATIONS_FLUSH_LIMIT)
1764                         filemap_flush(inode->i_mapping);
1765         }
1766         if (filp->private_data)
1767                 btrfs_ioctl_trans_end(filp);
1768         return 0;
1769 }
1770
1771 /*
1772  * fsync call for both files and directories.  This logs the inode into
1773  * the tree log instead of forcing full commits whenever possible.
1774  *
1775  * It needs to call filemap_fdatawait so that all ordered extent updates are
1776  * in the metadata btree are up to date for copying to the log.
1777  *
1778  * It drops the inode mutex before doing the tree log commit.  This is an
1779  * important optimization for directories because holding the mutex prevents
1780  * new operations on the dir while we write to disk.
1781  */
1782 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
1783 {
1784         struct dentry *dentry = file->f_path.dentry;
1785         struct inode *inode = dentry->d_inode;
1786         struct btrfs_root *root = BTRFS_I(inode)->root;
1787         int ret = 0;
1788         struct btrfs_trans_handle *trans;
1789         bool full_sync = 0;
1790
1791         trace_btrfs_sync_file(file, datasync);
1792
1793         /*
1794          * We write the dirty pages in the range and wait until they complete
1795          * out of the ->i_mutex. If so, we can flush the dirty pages by
1796          * multi-task, and make the performance up.  See
1797          * btrfs_wait_ordered_range for an explanation of the ASYNC check.
1798          */
1799         atomic_inc(&BTRFS_I(inode)->sync_writers);
1800         ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
1801         if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1802                              &BTRFS_I(inode)->runtime_flags))
1803                 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
1804         atomic_dec(&BTRFS_I(inode)->sync_writers);
1805         if (ret)
1806                 return ret;
1807
1808         mutex_lock(&inode->i_mutex);
1809
1810         /*
1811          * We flush the dirty pages again to avoid some dirty pages in the
1812          * range being left.
1813          */
1814         atomic_inc(&root->log_batch);
1815         full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1816                              &BTRFS_I(inode)->runtime_flags);
1817         if (full_sync)
1818                 btrfs_wait_ordered_range(inode, start, end - start + 1);
1819         atomic_inc(&root->log_batch);
1820
1821         /*
1822          * check the transaction that last modified this inode
1823          * and see if its already been committed
1824          */
1825         if (!BTRFS_I(inode)->last_trans) {
1826                 mutex_unlock(&inode->i_mutex);
1827                 goto out;
1828         }
1829
1830         /*
1831          * if the last transaction that changed this file was before
1832          * the current transaction, we can bail out now without any
1833          * syncing
1834          */
1835         smp_mb();
1836         if (btrfs_inode_in_log(inode, root->fs_info->generation) ||
1837             BTRFS_I(inode)->last_trans <=
1838             root->fs_info->last_trans_committed) {
1839                 BTRFS_I(inode)->last_trans = 0;
1840
1841                 /*
1842                  * We'v had everything committed since the last time we were
1843                  * modified so clear this flag in case it was set for whatever
1844                  * reason, it's no longer relevant.
1845                  */
1846                 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1847                           &BTRFS_I(inode)->runtime_flags);
1848                 mutex_unlock(&inode->i_mutex);
1849                 goto out;
1850         }
1851
1852         /*
1853          * ok we haven't committed the transaction yet, lets do a commit
1854          */
1855         if (file->private_data)
1856                 btrfs_ioctl_trans_end(file);
1857
1858         trans = btrfs_start_transaction(root, 0);
1859         if (IS_ERR(trans)) {
1860                 ret = PTR_ERR(trans);
1861                 mutex_unlock(&inode->i_mutex);
1862                 goto out;
1863         }
1864
1865         ret = btrfs_log_dentry_safe(trans, root, dentry);
1866         if (ret < 0) {
1867                 mutex_unlock(&inode->i_mutex);
1868                 goto out;
1869         }
1870
1871         /* we've logged all the items and now have a consistent
1872          * version of the file in the log.  It is possible that
1873          * someone will come in and modify the file, but that's
1874          * fine because the log is consistent on disk, and we
1875          * have references to all of the file's extents
1876          *
1877          * It is possible that someone will come in and log the
1878          * file again, but that will end up using the synchronization
1879          * inside btrfs_sync_log to keep things safe.
1880          */
1881         mutex_unlock(&inode->i_mutex);
1882
1883         if (ret != BTRFS_NO_LOG_SYNC) {
1884                 if (ret > 0) {
1885                         /*
1886                          * If we didn't already wait for ordered extents we need
1887                          * to do that now.
1888                          */
1889                         if (!full_sync)
1890                                 btrfs_wait_ordered_range(inode, start,
1891                                                          end - start + 1);
1892                         ret = btrfs_commit_transaction(trans, root);
1893                 } else {
1894                         ret = btrfs_sync_log(trans, root);
1895                         if (ret == 0) {
1896                                 ret = btrfs_end_transaction(trans, root);
1897                         } else {
1898                                 if (!full_sync)
1899                                         btrfs_wait_ordered_range(inode, start,
1900                                                                  end -
1901                                                                  start + 1);
1902                                 ret = btrfs_commit_transaction(trans, root);
1903                         }
1904                 }
1905         } else {
1906                 ret = btrfs_end_transaction(trans, root);
1907         }
1908 out:
1909         return ret > 0 ? -EIO : ret;
1910 }
1911
1912 static const struct vm_operations_struct btrfs_file_vm_ops = {
1913         .fault          = filemap_fault,
1914         .page_mkwrite   = btrfs_page_mkwrite,
1915         .remap_pages    = generic_file_remap_pages,
1916 };
1917
1918 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
1919 {
1920         struct address_space *mapping = filp->f_mapping;
1921
1922         if (!mapping->a_ops->readpage)
1923                 return -ENOEXEC;
1924
1925         file_accessed(filp);
1926         vma->vm_ops = &btrfs_file_vm_ops;
1927
1928         return 0;
1929 }
1930
1931 static int hole_mergeable(struct inode *inode, struct extent_buffer *leaf,
1932                           int slot, u64 start, u64 end)
1933 {
1934         struct btrfs_file_extent_item *fi;
1935         struct btrfs_key key;
1936
1937         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1938                 return 0;
1939
1940         btrfs_item_key_to_cpu(leaf, &key, slot);
1941         if (key.objectid != btrfs_ino(inode) ||
1942             key.type != BTRFS_EXTENT_DATA_KEY)
1943                 return 0;
1944
1945         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1946
1947         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
1948                 return 0;
1949
1950         if (btrfs_file_extent_disk_bytenr(leaf, fi))
1951                 return 0;
1952
1953         if (key.offset == end)
1954                 return 1;
1955         if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
1956                 return 1;
1957         return 0;
1958 }
1959
1960 static int fill_holes(struct btrfs_trans_handle *trans, struct inode *inode,
1961                       struct btrfs_path *path, u64 offset, u64 end)
1962 {
1963         struct btrfs_root *root = BTRFS_I(inode)->root;
1964         struct extent_buffer *leaf;
1965         struct btrfs_file_extent_item *fi;
1966         struct extent_map *hole_em;
1967         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
1968         struct btrfs_key key;
1969         int ret;
1970
1971         key.objectid = btrfs_ino(inode);
1972         key.type = BTRFS_EXTENT_DATA_KEY;
1973         key.offset = offset;
1974
1975
1976         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
1977         if (ret < 0)
1978                 return ret;
1979         BUG_ON(!ret);
1980
1981         leaf = path->nodes[0];
1982         if (hole_mergeable(inode, leaf, path->slots[0]-1, offset, end)) {
1983                 u64 num_bytes;
1984
1985                 path->slots[0]--;
1986                 fi = btrfs_item_ptr(leaf, path->slots[0],
1987                                     struct btrfs_file_extent_item);
1988                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
1989                         end - offset;
1990                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
1991                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
1992                 btrfs_set_file_extent_offset(leaf, fi, 0);
1993                 btrfs_mark_buffer_dirty(leaf);
1994                 goto out;
1995         }
1996
1997         if (hole_mergeable(inode, leaf, path->slots[0]+1, offset, end)) {
1998                 u64 num_bytes;
1999
2000                 path->slots[0]++;
2001                 key.offset = offset;
2002                 btrfs_set_item_key_safe(root, path, &key);
2003                 fi = btrfs_item_ptr(leaf, path->slots[0],
2004                                     struct btrfs_file_extent_item);
2005                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2006                         offset;
2007                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2008                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2009                 btrfs_set_file_extent_offset(leaf, fi, 0);
2010                 btrfs_mark_buffer_dirty(leaf);
2011                 goto out;
2012         }
2013         btrfs_release_path(path);
2014
2015         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode), offset,
2016                                        0, 0, end - offset, 0, end - offset,
2017                                        0, 0, 0);
2018         if (ret)
2019                 return ret;
2020
2021 out:
2022         btrfs_release_path(path);
2023
2024         hole_em = alloc_extent_map();
2025         if (!hole_em) {
2026                 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2027                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2028                         &BTRFS_I(inode)->runtime_flags);
2029         } else {
2030                 hole_em->start = offset;
2031                 hole_em->len = end - offset;
2032                 hole_em->ram_bytes = hole_em->len;
2033                 hole_em->orig_start = offset;
2034
2035                 hole_em->block_start = EXTENT_MAP_HOLE;
2036                 hole_em->block_len = 0;
2037                 hole_em->orig_block_len = 0;
2038                 hole_em->bdev = root->fs_info->fs_devices->latest_bdev;
2039                 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2040                 hole_em->generation = trans->transid;
2041
2042                 do {
2043                         btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2044                         write_lock(&em_tree->lock);
2045                         ret = add_extent_mapping(em_tree, hole_em, 1);
2046                         write_unlock(&em_tree->lock);
2047                 } while (ret == -EEXIST);
2048                 free_extent_map(hole_em);
2049                 if (ret)
2050                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2051                                 &BTRFS_I(inode)->runtime_flags);
2052         }
2053
2054         return 0;
2055 }
2056
2057 static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2058 {
2059         struct btrfs_root *root = BTRFS_I(inode)->root;
2060         struct extent_state *cached_state = NULL;
2061         struct btrfs_path *path;
2062         struct btrfs_block_rsv *rsv;
2063         struct btrfs_trans_handle *trans;
2064         u64 lockstart = round_up(offset, BTRFS_I(inode)->root->sectorsize);
2065         u64 lockend = round_down(offset + len,
2066                                  BTRFS_I(inode)->root->sectorsize) - 1;
2067         u64 cur_offset = lockstart;
2068         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
2069         u64 drop_end;
2070         int ret = 0;
2071         int err = 0;
2072         bool same_page = ((offset >> PAGE_CACHE_SHIFT) ==
2073                           ((offset + len - 1) >> PAGE_CACHE_SHIFT));
2074
2075         btrfs_wait_ordered_range(inode, offset, len);
2076
2077         mutex_lock(&inode->i_mutex);
2078         /*
2079          * We needn't truncate any page which is beyond the end of the file
2080          * because we are sure there is no data there.
2081          */
2082         /*
2083          * Only do this if we are in the same page and we aren't doing the
2084          * entire page.
2085          */
2086         if (same_page && len < PAGE_CACHE_SIZE) {
2087                 if (offset < round_up(inode->i_size, PAGE_CACHE_SIZE))
2088                         ret = btrfs_truncate_page(inode, offset, len, 0);
2089                 mutex_unlock(&inode->i_mutex);
2090                 return ret;
2091         }
2092
2093         /* zero back part of the first page */
2094         if (offset < round_up(inode->i_size, PAGE_CACHE_SIZE)) {
2095                 ret = btrfs_truncate_page(inode, offset, 0, 0);
2096                 if (ret) {
2097                         mutex_unlock(&inode->i_mutex);
2098                         return ret;
2099                 }
2100         }
2101
2102         /* zero the front end of the last page */
2103         if (offset + len < round_up(inode->i_size, PAGE_CACHE_SIZE)) {
2104                 ret = btrfs_truncate_page(inode, offset + len, 0, 1);
2105                 if (ret) {
2106                         mutex_unlock(&inode->i_mutex);
2107                         return ret;
2108                 }
2109         }
2110
2111         if (lockend < lockstart) {
2112                 mutex_unlock(&inode->i_mutex);
2113                 return 0;
2114         }
2115
2116         while (1) {
2117                 struct btrfs_ordered_extent *ordered;
2118
2119                 truncate_pagecache_range(inode, lockstart, lockend);
2120
2121                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2122                                  0, &cached_state);
2123                 ordered = btrfs_lookup_first_ordered_extent(inode, lockend);
2124
2125                 /*
2126                  * We need to make sure we have no ordered extents in this range
2127                  * and nobody raced in and read a page in this range, if we did
2128                  * we need to try again.
2129                  */
2130                 if ((!ordered ||
2131                     (ordered->file_offset + ordered->len < lockstart ||
2132                      ordered->file_offset > lockend)) &&
2133                      !test_range_bit(&BTRFS_I(inode)->io_tree, lockstart,
2134                                      lockend, EXTENT_UPTODATE, 0,
2135                                      cached_state)) {
2136                         if (ordered)
2137                                 btrfs_put_ordered_extent(ordered);
2138                         break;
2139                 }
2140                 if (ordered)
2141                         btrfs_put_ordered_extent(ordered);
2142                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2143                                      lockend, &cached_state, GFP_NOFS);
2144                 btrfs_wait_ordered_range(inode, lockstart,
2145                                          lockend - lockstart + 1);
2146         }
2147
2148         path = btrfs_alloc_path();
2149         if (!path) {
2150                 ret = -ENOMEM;
2151                 goto out;
2152         }
2153
2154         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
2155         if (!rsv) {
2156                 ret = -ENOMEM;
2157                 goto out_free;
2158         }
2159         rsv->size = btrfs_calc_trunc_metadata_size(root, 1);
2160         rsv->failfast = 1;
2161
2162         /*
2163          * 1 - update the inode
2164          * 1 - removing the extents in the range
2165          * 1 - adding the hole extent
2166          */
2167         trans = btrfs_start_transaction(root, 3);
2168         if (IS_ERR(trans)) {
2169                 err = PTR_ERR(trans);
2170                 goto out_free;
2171         }
2172
2173         ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv, rsv,
2174                                       min_size);
2175         BUG_ON(ret);
2176         trans->block_rsv = rsv;
2177
2178         while (cur_offset < lockend) {
2179                 ret = __btrfs_drop_extents(trans, root, inode, path,
2180                                            cur_offset, lockend + 1,
2181                                            &drop_end, 1);
2182                 if (ret != -ENOSPC)
2183                         break;
2184
2185                 trans->block_rsv = &root->fs_info->trans_block_rsv;
2186
2187                 ret = fill_holes(trans, inode, path, cur_offset, drop_end);
2188                 if (ret) {
2189                         err = ret;
2190                         break;
2191                 }
2192
2193                 cur_offset = drop_end;
2194
2195                 ret = btrfs_update_inode(trans, root, inode);
2196                 if (ret) {
2197                         err = ret;
2198                         break;
2199                 }
2200
2201                 btrfs_end_transaction(trans, root);
2202                 btrfs_btree_balance_dirty(root);
2203
2204                 trans = btrfs_start_transaction(root, 3);
2205                 if (IS_ERR(trans)) {
2206                         ret = PTR_ERR(trans);
2207                         trans = NULL;
2208                         break;
2209                 }
2210
2211                 ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv,
2212                                               rsv, min_size);
2213                 BUG_ON(ret);    /* shouldn't happen */
2214                 trans->block_rsv = rsv;
2215         }
2216
2217         if (ret) {
2218                 err = ret;
2219                 goto out_trans;
2220         }
2221
2222         trans->block_rsv = &root->fs_info->trans_block_rsv;
2223         ret = fill_holes(trans, inode, path, cur_offset, drop_end);
2224         if (ret) {
2225                 err = ret;
2226                 goto out_trans;
2227         }
2228
2229 out_trans:
2230         if (!trans)
2231                 goto out_free;
2232
2233         inode_inc_iversion(inode);
2234         inode->i_mtime = inode->i_ctime = CURRENT_TIME;
2235
2236         trans->block_rsv = &root->fs_info->trans_block_rsv;
2237         ret = btrfs_update_inode(trans, root, inode);
2238         btrfs_end_transaction(trans, root);
2239         btrfs_btree_balance_dirty(root);
2240 out_free:
2241         btrfs_free_path(path);
2242         btrfs_free_block_rsv(root, rsv);
2243 out:
2244         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2245                              &cached_state, GFP_NOFS);
2246         mutex_unlock(&inode->i_mutex);
2247         if (ret && !err)
2248                 err = ret;
2249         return err;
2250 }
2251
2252 static long btrfs_fallocate(struct file *file, int mode,
2253                             loff_t offset, loff_t len)
2254 {
2255         struct inode *inode = file_inode(file);
2256         struct extent_state *cached_state = NULL;
2257         struct btrfs_root *root = BTRFS_I(inode)->root;
2258         u64 cur_offset;
2259         u64 last_byte;
2260         u64 alloc_start;
2261         u64 alloc_end;
2262         u64 alloc_hint = 0;
2263         u64 locked_end;
2264         struct extent_map *em;
2265         int blocksize = BTRFS_I(inode)->root->sectorsize;
2266         int ret;
2267
2268         alloc_start = round_down(offset, blocksize);
2269         alloc_end = round_up(offset + len, blocksize);
2270
2271         /* Make sure we aren't being give some crap mode */
2272         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2273                 return -EOPNOTSUPP;
2274
2275         if (mode & FALLOC_FL_PUNCH_HOLE)
2276                 return btrfs_punch_hole(inode, offset, len);
2277
2278         /*
2279          * Make sure we have enough space before we do the
2280          * allocation.
2281          */
2282         ret = btrfs_check_data_free_space(inode, alloc_end - alloc_start);
2283         if (ret)
2284                 return ret;
2285         if (root->fs_info->quota_enabled) {
2286                 ret = btrfs_qgroup_reserve(root, alloc_end - alloc_start);
2287                 if (ret)
2288                         goto out_reserve_fail;
2289         }
2290
2291         mutex_lock(&inode->i_mutex);
2292         ret = inode_newsize_ok(inode, alloc_end);
2293         if (ret)
2294                 goto out;
2295
2296         if (alloc_start > inode->i_size) {
2297                 ret = btrfs_cont_expand(inode, i_size_read(inode),
2298                                         alloc_start);
2299                 if (ret)
2300                         goto out;
2301         } else {
2302                 /*
2303                  * If we are fallocating from the end of the file onward we
2304                  * need to zero out the end of the page if i_size lands in the
2305                  * middle of a page.
2306                  */
2307                 ret = btrfs_truncate_page(inode, inode->i_size, 0, 0);
2308                 if (ret)
2309                         goto out;
2310         }
2311
2312         /*
2313          * wait for ordered IO before we have any locks.  We'll loop again
2314          * below with the locks held.
2315          */
2316         btrfs_wait_ordered_range(inode, alloc_start, alloc_end - alloc_start);
2317
2318         locked_end = alloc_end - 1;
2319         while (1) {
2320                 struct btrfs_ordered_extent *ordered;
2321
2322                 /* the extent lock is ordered inside the running
2323                  * transaction
2324                  */
2325                 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
2326                                  locked_end, 0, &cached_state);
2327                 ordered = btrfs_lookup_first_ordered_extent(inode,
2328                                                             alloc_end - 1);
2329                 if (ordered &&
2330                     ordered->file_offset + ordered->len > alloc_start &&
2331                     ordered->file_offset < alloc_end) {
2332                         btrfs_put_ordered_extent(ordered);
2333                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
2334                                              alloc_start, locked_end,
2335                                              &cached_state, GFP_NOFS);
2336                         /*
2337                          * we can't wait on the range with the transaction
2338                          * running or with the extent lock held
2339                          */
2340                         btrfs_wait_ordered_range(inode, alloc_start,
2341                                                  alloc_end - alloc_start);
2342                 } else {
2343                         if (ordered)
2344                                 btrfs_put_ordered_extent(ordered);
2345                         break;
2346                 }
2347         }
2348
2349         cur_offset = alloc_start;
2350         while (1) {
2351                 u64 actual_end;
2352
2353                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
2354                                       alloc_end - cur_offset, 0);
2355                 if (IS_ERR_OR_NULL(em)) {
2356                         if (!em)
2357                                 ret = -ENOMEM;
2358                         else
2359                                 ret = PTR_ERR(em);
2360                         break;
2361                 }
2362                 last_byte = min(extent_map_end(em), alloc_end);
2363                 actual_end = min_t(u64, extent_map_end(em), offset + len);
2364                 last_byte = ALIGN(last_byte, blocksize);
2365
2366                 if (em->block_start == EXTENT_MAP_HOLE ||
2367                     (cur_offset >= inode->i_size &&
2368                      !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
2369                         ret = btrfs_prealloc_file_range(inode, mode, cur_offset,
2370                                                         last_byte - cur_offset,
2371                                                         1 << inode->i_blkbits,
2372                                                         offset + len,
2373                                                         &alloc_hint);
2374
2375                         if (ret < 0) {
2376                                 free_extent_map(em);
2377                                 break;
2378                         }
2379                 } else if (actual_end > inode->i_size &&
2380                            !(mode & FALLOC_FL_KEEP_SIZE)) {
2381                         /*
2382                          * We didn't need to allocate any more space, but we
2383                          * still extended the size of the file so we need to
2384                          * update i_size.
2385                          */
2386                         inode->i_ctime = CURRENT_TIME;
2387                         i_size_write(inode, actual_end);
2388                         btrfs_ordered_update_i_size(inode, actual_end, NULL);
2389                 }
2390                 free_extent_map(em);
2391
2392                 cur_offset = last_byte;
2393                 if (cur_offset >= alloc_end) {
2394                         ret = 0;
2395                         break;
2396                 }
2397         }
2398         unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
2399                              &cached_state, GFP_NOFS);
2400 out:
2401         mutex_unlock(&inode->i_mutex);
2402         if (root->fs_info->quota_enabled)
2403                 btrfs_qgroup_free(root, alloc_end - alloc_start);
2404 out_reserve_fail:
2405         /* Let go of our reservation. */
2406         btrfs_free_reserved_data_space(inode, alloc_end - alloc_start);
2407         return ret;
2408 }
2409
2410 static int find_desired_extent(struct inode *inode, loff_t *offset, int whence)
2411 {
2412         struct btrfs_root *root = BTRFS_I(inode)->root;
2413         struct extent_map *em;
2414         struct extent_state *cached_state = NULL;
2415         u64 lockstart = *offset;
2416         u64 lockend = i_size_read(inode);
2417         u64 start = *offset;
2418         u64 orig_start = *offset;
2419         u64 len = i_size_read(inode);
2420         u64 last_end = 0;
2421         int ret = 0;
2422
2423         lockend = max_t(u64, root->sectorsize, lockend);
2424         if (lockend <= lockstart)
2425                 lockend = lockstart + root->sectorsize;
2426
2427         lockend--;
2428         len = lockend - lockstart + 1;
2429
2430         len = max_t(u64, len, root->sectorsize);
2431         if (inode->i_size == 0)
2432                 return -ENXIO;
2433
2434         lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend, 0,
2435                          &cached_state);
2436
2437         /*
2438          * Delalloc is such a pain.  If we have a hole and we have pending
2439          * delalloc for a portion of the hole we will get back a hole that
2440          * exists for the entire range since it hasn't been actually written
2441          * yet.  So to take care of this case we need to look for an extent just
2442          * before the position we want in case there is outstanding delalloc
2443          * going on here.
2444          */
2445         if (whence == SEEK_HOLE && start != 0) {
2446                 if (start <= root->sectorsize)
2447                         em = btrfs_get_extent_fiemap(inode, NULL, 0, 0,
2448                                                      root->sectorsize, 0);
2449                 else
2450                         em = btrfs_get_extent_fiemap(inode, NULL, 0,
2451                                                      start - root->sectorsize,
2452                                                      root->sectorsize, 0);
2453                 if (IS_ERR(em)) {
2454                         ret = PTR_ERR(em);
2455                         goto out;
2456                 }
2457                 last_end = em->start + em->len;
2458                 if (em->block_start == EXTENT_MAP_DELALLOC)
2459                         last_end = min_t(u64, last_end, inode->i_size);
2460                 free_extent_map(em);
2461         }
2462
2463         while (1) {
2464                 em = btrfs_get_extent_fiemap(inode, NULL, 0, start, len, 0);
2465                 if (IS_ERR(em)) {
2466                         ret = PTR_ERR(em);
2467                         break;
2468                 }
2469
2470                 if (em->block_start == EXTENT_MAP_HOLE) {
2471                         if (test_bit(EXTENT_FLAG_VACANCY, &em->flags)) {
2472                                 if (last_end <= orig_start) {
2473                                         free_extent_map(em);
2474                                         ret = -ENXIO;
2475                                         break;
2476                                 }
2477                         }
2478
2479                         if (whence == SEEK_HOLE) {
2480                                 *offset = start;
2481                                 free_extent_map(em);
2482                                 break;
2483                         }
2484                 } else {
2485                         if (whence == SEEK_DATA) {
2486                                 if (em->block_start == EXTENT_MAP_DELALLOC) {
2487                                         if (start >= inode->i_size) {
2488                                                 free_extent_map(em);
2489                                                 ret = -ENXIO;
2490                                                 break;
2491                                         }
2492                                 }
2493
2494                                 if (!test_bit(EXTENT_FLAG_PREALLOC,
2495                                               &em->flags)) {
2496                                         *offset = start;
2497                                         free_extent_map(em);
2498                                         break;
2499                                 }
2500                         }
2501                 }
2502
2503                 start = em->start + em->len;
2504                 last_end = em->start + em->len;
2505
2506                 if (em->block_start == EXTENT_MAP_DELALLOC)
2507                         last_end = min_t(u64, last_end, inode->i_size);
2508
2509                 if (test_bit(EXTENT_FLAG_VACANCY, &em->flags)) {
2510                         free_extent_map(em);
2511                         ret = -ENXIO;
2512                         break;
2513                 }
2514                 free_extent_map(em);
2515                 cond_resched();
2516         }
2517         if (!ret)
2518                 *offset = min(*offset, inode->i_size);
2519 out:
2520         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2521                              &cached_state, GFP_NOFS);
2522         return ret;
2523 }
2524
2525 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
2526 {
2527         struct inode *inode = file->f_mapping->host;
2528         int ret;
2529
2530         mutex_lock(&inode->i_mutex);
2531         switch (whence) {
2532         case SEEK_END:
2533         case SEEK_CUR:
2534                 offset = generic_file_llseek(file, offset, whence);
2535                 goto out;
2536         case SEEK_DATA:
2537         case SEEK_HOLE:
2538                 if (offset >= i_size_read(inode)) {
2539                         mutex_unlock(&inode->i_mutex);
2540                         return -ENXIO;
2541                 }
2542
2543                 ret = find_desired_extent(inode, &offset, whence);
2544                 if (ret) {
2545                         mutex_unlock(&inode->i_mutex);
2546                         return ret;
2547                 }
2548         }
2549
2550         offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
2551 out:
2552         mutex_unlock(&inode->i_mutex);
2553         return offset;
2554 }
2555
2556 const struct file_operations btrfs_file_operations = {
2557         .llseek         = btrfs_file_llseek,
2558         .read           = do_sync_read,
2559         .write          = do_sync_write,
2560         .aio_read       = generic_file_aio_read,
2561         .splice_read    = generic_file_splice_read,
2562         .aio_write      = btrfs_file_aio_write,
2563         .mmap           = btrfs_file_mmap,
2564         .open           = generic_file_open,
2565         .release        = btrfs_release_file,
2566         .fsync          = btrfs_sync_file,
2567         .fallocate      = btrfs_fallocate,
2568         .unlocked_ioctl = btrfs_ioctl,
2569 #ifdef CONFIG_COMPAT
2570         .compat_ioctl   = btrfs_ioctl,
2571 #endif
2572 };
2573
2574 void btrfs_auto_defrag_exit(void)
2575 {
2576         if (btrfs_inode_defrag_cachep)
2577                 kmem_cache_destroy(btrfs_inode_defrag_cachep);
2578 }
2579
2580 int btrfs_auto_defrag_init(void)
2581 {
2582         btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
2583                                         sizeof(struct inode_defrag), 0,
2584                                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD,
2585                                         NULL);
2586         if (!btrfs_inode_defrag_cachep)
2587                 return -ENOMEM;
2588
2589         return 0;
2590 }