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 "volumes.h"
  43
  44 static struct kmem_cache *btrfs_inode_defrag_cachep;
  45 /*
  46  * when auto defrag is enabled we
  47  * queue up these defrag structs to remember which
  48  * inodes need defragging passes
  49  */
  50 struct inode_defrag {
  51         struct rb_node rb_node;
  52         /* objectid */
  53         u64 ino;
  54         /*
  55          * transid where the defrag was added, we search for
  56          * extents newer than this
  57          */
  58         u64 transid;
  59
  60         /* root objectid */
  61         u64 root;
  62
  63         /* last offset we were able to defrag */
  64         u64 last_offset;
  65
  66         /* if we've wrapped around back to zero once already */
  67         int cycled;
  68 };
  69
  70 static int __compare_inode_defrag(struct inode_defrag *defrag1,
  71                                   struct inode_defrag *defrag2)
  72 {
  73         if (defrag1->root > defrag2->root)
  74                 return 1;
  75         else if (defrag1->root < defrag2->root)
  76                 return -1;
  77         else if (defrag1->ino > defrag2->ino)
  78                 return 1;
  79         else if (defrag1->ino < defrag2->ino)
  80                 return -1;
  81         else
  82                 return 0;
  83 }
  84
  85 /* pop a record for an inode into the defrag tree.  The lock
  86  * must be held already
  87  *
  88  * If you're inserting a record for an older transid than an
  89  * existing record, the transid already in the tree is lowered
  90  *
  91  * If an existing record is found the defrag item you
  92  * pass in is freed
  93  */
  94 static int __btrfs_add_inode_defrag(struct inode *inode,
  95                                     struct inode_defrag *defrag)
  96 {
  97         struct btrfs_root *root = BTRFS_I(inode)->root;
  98         struct inode_defrag *entry;
  99         struct rb_node **p;
 100         struct rb_node *parent = NULL;
 101         int ret;
 102
 103         p = &root->fs_info->defrag_inodes.rb_node;
 104         while (*p) {
 105                 parent = *p;
 106                 entry = rb_entry(parent, struct inode_defrag, rb_node);
 107
 108                 ret = __compare_inode_defrag(defrag, entry);
 109                 if (ret < 0)
 110                         p = &parent->rb_left;
 111                 else if (ret > 0)
 112                         p = &parent->rb_right;
 113                 else {
 114                         /* if we're reinserting an entry for
 115                          * an old defrag run, make sure to
 116                          * lower the transid of our existing record
 117                          */
 118                         if (defrag->transid < entry->transid)
 119                                 entry->transid = defrag->transid;
 120                         if (defrag->last_offset > entry->last_offset)
 121                                 entry->last_offset = defrag->last_offset;
 122                         return -EEXIST;
 123                 }
 124         }
 125         set_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
 126         rb_link_node(&defrag->rb_node, parent, p);
 127         rb_insert_color(&defrag->rb_node, &root->fs_info->defrag_inodes);
 128         return 0;
 129 }
 130
 131 static inline int __need_auto_defrag(struct btrfs_root *root)
 132 {
 133         if (!btrfs_test_opt(root, AUTO_DEFRAG))
 134                 return 0;
 135
 136         if (btrfs_fs_closing(root->fs_info))
 137                 return 0;
 138
 139         return 1;
 140 }
 141
 142 /*
 143  * insert a defrag record for this inode if auto defrag is
 144  * enabled
 145  */
 146 int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
 147                            struct inode *inode)
 148 {
 149         struct btrfs_root *root = BTRFS_I(inode)->root;
 150         struct inode_defrag *defrag;
 151         u64 transid;
 152         int ret;
 153
 154         if (!__need_auto_defrag(root))
 155                 return 0;
 156
 157         if (test_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags))
 158                 return 0;
 159
 160         if (trans)
 161                 transid = trans->transid;
 162         else
 163                 transid = BTRFS_I(inode)->root->last_trans;
 164
 165         defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
 166         if (!defrag)
 167                 return -ENOMEM;
 168
 169         defrag->ino = btrfs_ino(inode);
 170         defrag->transid = transid;
 171         defrag->root = root->root_key.objectid;
 172
 173         spin_lock(&root->fs_info->defrag_inodes_lock);
 174         if (!test_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags)) {
 175                 /*
 176                  * If we set IN_DEFRAG flag and evict the inode from memory,
 177                  * and then re-read this inode, this new inode doesn't have
 178                  * IN_DEFRAG flag. At the case, we may find the existed defrag.
 179                  */
 180                 ret = __btrfs_add_inode_defrag(inode, defrag);
 181                 if (ret)
 182                         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 183         } else {
 184                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 185         }
 186         spin_unlock(&root->fs_info->defrag_inodes_lock);
 187         return 0;
 188 }
 189
 190 /*
 191  * Requeue the defrag object. If there is a defrag object that points to
 192  * the same inode in the tree, we will merge them together (by
 193  * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
 194  */
 195 static void btrfs_requeue_inode_defrag(struct inode *inode,
 196                                        struct inode_defrag *defrag)
 197 {
 198         struct btrfs_root *root = BTRFS_I(inode)->root;
 199         int ret;
 200
 201         if (!__need_auto_defrag(root))
 202                 goto out;
 203
 204         /*
 205          * Here we don't check the IN_DEFRAG flag, because we need merge
 206          * them together.
 207          */
 208         spin_lock(&root->fs_info->defrag_inodes_lock);
 209         ret = __btrfs_add_inode_defrag(inode, defrag);
 210         spin_unlock(&root->fs_info->defrag_inodes_lock);
 211         if (ret)
 212                 goto out;
 213         return;
 214 out:
 215         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 216 }
 217
 218 /*
 219  * pick the defragable inode that we want, if it doesn't exist, we will get
 220  * the next one.
 221  */
 222 static struct inode_defrag *
 223 btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
 224 {
 225         struct inode_defrag *entry = NULL;
 226         struct inode_defrag tmp;
 227         struct rb_node *p;
 228         struct rb_node *parent = NULL;
 229         int ret;
 230
 231         tmp.ino = ino;
 232         tmp.root = root;
 233
 234         spin_lock(&fs_info->defrag_inodes_lock);
 235         p = fs_info->defrag_inodes.rb_node;
 236         while (p) {
 237                 parent = p;
 238                 entry = rb_entry(parent, struct inode_defrag, rb_node);
 239
 240                 ret = __compare_inode_defrag(&tmp, entry);
 241                 if (ret < 0)
 242                         p = parent->rb_left;
 243                 else if (ret > 0)
 244                         p = parent->rb_right;
 245                 else
 246                         goto out;
 247         }
 248
 249         if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
 250                 parent = rb_next(parent);
 251                 if (parent)
 252                         entry = rb_entry(parent, struct inode_defrag, rb_node);
 253                 else
 254                         entry = NULL;
 255         }
 256 out:
 257         if (entry)
 258                 rb_erase(parent, &fs_info->defrag_inodes);
 259         spin_unlock(&fs_info->defrag_inodes_lock);
 260         return entry;
 261 }
 262
 263 void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
 264 {
 265         struct inode_defrag *defrag;
 266         struct rb_node *node;
 267
 268         spin_lock(&fs_info->defrag_inodes_lock);
 269         node = rb_first(&fs_info->defrag_inodes);
 270         while (node) {
 271                 rb_erase(node, &fs_info->defrag_inodes);
 272                 defrag = rb_entry(node, struct inode_defrag, rb_node);
 273                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 274
 275                 if (need_resched()) {
 276                         spin_unlock(&fs_info->defrag_inodes_lock);
 277                         cond_resched();
 278                         spin_lock(&fs_info->defrag_inodes_lock);
 279                 }
 280
 281                 node = rb_first(&fs_info->defrag_inodes);
 282         }
 283         spin_unlock(&fs_info->defrag_inodes_lock);
 284 }
 285
 286 #define BTRFS_DEFRAG_BATCH      1024
 287
 288 static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
 289                                     struct inode_defrag *defrag)
 290 {
 291         struct btrfs_root *inode_root;
 292         struct inode *inode;
 293         struct btrfs_key key;
 294         struct btrfs_ioctl_defrag_range_args range;
 295         int num_defrag;
 296         int index;
 297         int ret;
 298
 299         /* get the inode */
 300         key.objectid = defrag->root;
 301         btrfs_set_key_type(&key, BTRFS_ROOT_ITEM_KEY);
 302         key.offset = (u64)-1;
 303
 304         index = srcu_read_lock(&fs_info->subvol_srcu);
 305
 306         inode_root = btrfs_read_fs_root_no_name(fs_info, &key);
 307         if (IS_ERR(inode_root)) {
 308                 ret = PTR_ERR(inode_root);
 309                 goto cleanup;
 310         }
 311
 312         key.objectid = defrag->ino;
 313         btrfs_set_key_type(&key, BTRFS_INODE_ITEM_KEY);
 314         key.offset = 0;
 315         inode = btrfs_iget(fs_info->sb, &key, inode_root, NULL);
 316         if (IS_ERR(inode)) {
 317                 ret = PTR_ERR(inode);
 318                 goto cleanup;
 319         }
 320         srcu_read_unlock(&fs_info->subvol_srcu, index);
 321
 322         /* do a chunk of defrag */
 323         clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
 324         memset(&range, 0, sizeof(range));
 325         range.len = (u64)-1;
 326         range.start = defrag->last_offset;
 327
 328         sb_start_write(fs_info->sb);
 329         num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
 330                                        BTRFS_DEFRAG_BATCH);
 331         sb_end_write(fs_info->sb);
 332         /*
 333          * if we filled the whole defrag batch, there
 334          * must be more work to do.  Queue this defrag
 335          * again
 336          */
 337         if (num_defrag == BTRFS_DEFRAG_BATCH) {
 338                 defrag->last_offset = range.start;
 339                 btrfs_requeue_inode_defrag(inode, defrag);
 340         } else if (defrag->last_offset && !defrag->cycled) {
 341                 /*
 342                  * we didn't fill our defrag batch, but
 343                  * we didn't start at zero.  Make sure we loop
 344                  * around to the start of the file.
 345                  */
 346                 defrag->last_offset = 0;
 347                 defrag->cycled = 1;
 348                 btrfs_requeue_inode_defrag(inode, defrag);
 349         } else {
 350                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 351         }
 352
 353         iput(inode);
 354         return 0;
 355 cleanup:
 356         srcu_read_unlock(&fs_info->subvol_srcu, index);
 357         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 358         return ret;
 359 }
 360
 361 /*
 362  * run through the list of inodes in the FS that need
 363  * defragging
 364  */
 365 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
 366 {
 367         struct inode_defrag *defrag;
 368         u64 first_ino = 0;
 369         u64 root_objectid = 0;
 370
 371         atomic_inc(&fs_info->defrag_running);
 372         while (1) {
 373                 /* Pause the auto defragger. */
 374                 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
 375                              &fs_info->fs_state))
 376                         break;
 377
 378                 if (!__need_auto_defrag(fs_info->tree_root))
 379                         break;
 380
 381                 /* find an inode to defrag */
 382                 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
 383                                                  first_ino);
 384                 if (!defrag) {
 385                         if (root_objectid || first_ino) {
 386                                 root_objectid = 0;
 387                                 first_ino = 0;
 388                                 continue;
 389                         } else {
 390                                 break;
 391                         }
 392                 }
 393
 394                 first_ino = defrag->ino + 1;
 395                 root_objectid = defrag->root;
 396
 397                 __btrfs_run_defrag_inode(fs_info, defrag);
 398         }
 399         atomic_dec(&fs_info->defrag_running);
 400
 401         /*
 402          * during unmount, we use the transaction_wait queue to
 403          * wait for the defragger to stop
 404          */
 405         wake_up(&fs_info->transaction_wait);
 406         return 0;
 407 }
 408
 409 /* simple helper to fault in pages and copy.  This should go away
 410  * and be replaced with calls into generic code.
 411  */
 412 static noinline int btrfs_copy_from_user(loff_t pos, int num_pages,
 413                                          size_t write_bytes,
 414                                          struct page **prepared_pages,
 415                                          struct iov_iter *i)
 416 {
 417         size_t copied = 0;
 418         size_t total_copied = 0;
 419         int pg = 0;
 420         int offset = pos & (PAGE_CACHE_SIZE - 1);
 421
 422         while (write_bytes > 0) {
 423                 size_t count = min_t(size_t,
 424                                      PAGE_CACHE_SIZE - offset, write_bytes);
 425                 struct page *page = prepared_pages[pg];
 426                 /*
 427                  * Copy data from userspace to the current page
 428                  */
 429                 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
 430
 431                 /* Flush processor's dcache for this page */
 432                 flush_dcache_page(page);
 433
 434                 /*
 435                  * if we get a partial write, we can end up with
 436                  * partially up to date pages.  These add
 437                  * a lot of complexity, so make sure they don't
 438                  * happen by forcing this copy to be retried.
 439                  *
 440                  * The rest of the btrfs_file_write code will fall
 441                  * back to page at a time copies after we return 0.
 442                  */
 443                 if (!PageUptodate(page) && copied < count)
 444                         copied = 0;
 445
 446                 iov_iter_advance(i, copied);
 447                 write_bytes -= copied;
 448                 total_copied += copied;
 449
 450                 /* Return to btrfs_file_aio_write to fault page */
 451                 if (unlikely(copied == 0))
 452                         break;
 453
 454                 if (unlikely(copied < PAGE_CACHE_SIZE - offset)) {
 455                         offset += copied;
 456                 } else {
 457                         pg++;
 458                         offset = 0;
 459                 }
 460         }
 461         return total_copied;
 462 }
 463
 464 /*
 465  * unlocks pages after btrfs_file_write is done with them
 466  */
 467 static void btrfs_drop_pages(struct page **pages, size_t num_pages)
 468 {
 469         size_t i;
 470         for (i = 0; i < num_pages; i++) {
 471                 /* page checked is some magic around finding pages that
 472                  * have been modified without going through btrfs_set_page_dirty
 473                  * clear it here
 474                  */
 475                 ClearPageChecked(pages[i]);
 476                 unlock_page(pages[i]);
 477                 mark_page_accessed(pages[i]);
 478                 page_cache_release(pages[i]);
 479         }
 480 }
 481
 482 /*
 483  * after copy_from_user, pages need to be dirtied and we need to make
 484  * sure holes are created between the current EOF and the start of
 485  * any next extents (if required).
 486  *
 487  * this also makes the decision about creating an inline extent vs
 488  * doing real data extents, marking pages dirty and delalloc as required.
 489  */
 490 int btrfs_dirty_pages(struct btrfs_root *root, struct inode *inode,
 491                              struct page **pages, size_t num_pages,
 492                              loff_t pos, size_t write_bytes,
 493                              struct extent_state **cached)
 494 {
 495         int err = 0;
 496         int i;
 497         u64 num_bytes;
 498         u64 start_pos;
 499         u64 end_of_last_block;
 500         u64 end_pos = pos + write_bytes;
 501         loff_t isize = i_size_read(inode);
 502
 503         start_pos = pos & ~((u64)root->sectorsize - 1);
 504         num_bytes = ALIGN(write_bytes + pos - start_pos, root->sectorsize);
 505
 506         end_of_last_block = start_pos + num_bytes - 1;
 507         err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
 508                                         cached);
 509         if (err)
 510                 return err;
 511
 512         for (i = 0; i < num_pages; i++) {
 513                 struct page *p = pages[i];
 514                 SetPageUptodate(p);
 515                 ClearPageChecked(p);
 516                 set_page_dirty(p);
 517         }
 518
 519         /*
 520          * we've only changed i_size in ram, and we haven't updated
 521          * the disk i_size.  There is no need to log the inode
 522          * at this time.
 523          */
 524         if (end_pos > isize)
 525                 i_size_write(inode, end_pos);
 526         return 0;
 527 }
 528
 529 /*
 530  * this drops all the extents in the cache that intersect the range
 531  * [start, end].  Existing extents are split as required.
 532  */
 533 void btrfs_drop_extent_cache(struct inode *inode, u64 start, u64 end,
 534                              int skip_pinned)
 535 {
 536         struct extent_map *em;
 537         struct extent_map *split = NULL;
 538         struct extent_map *split2 = NULL;
 539         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 540         u64 len = end - start + 1;
 541         u64 gen;
 542         int ret;
 543         int testend = 1;
 544         unsigned long flags;
 545         int compressed = 0;
 546         bool modified;
 547
 548         WARN_ON(end < start);
 549         if (end == (u64)-1) {
 550                 len = (u64)-1;
 551                 testend = 0;
 552         }
 553         while (1) {
 554                 int no_splits = 0;
 555
 556                 modified = false;
 557                 if (!split)
 558                         split = alloc_extent_map();
 559                 if (!split2)
 560                         split2 = alloc_extent_map();
 561                 if (!split || !split2)
 562                         no_splits = 1;
 563
 564                 write_lock(&em_tree->lock);
 565                 em = lookup_extent_mapping(em_tree, start, len);
 566                 if (!em) {
 567                         write_unlock(&em_tree->lock);
 568                         break;
 569                 }
 570                 flags = em->flags;
 571                 gen = em->generation;
 572                 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
 573                         if (testend && em->start + em->len >= start + len) {
 574                                 free_extent_map(em);
 575                                 write_unlock(&em_tree->lock);
 576                                 break;
 577                         }
 578                         start = em->start + em->len;
 579                         if (testend)
 580                                 len = start + len - (em->start + em->len);
 581                         free_extent_map(em);
 582                         write_unlock(&em_tree->lock);
 583                         continue;
 584                 }
 585                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 586                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
 587                 clear_bit(EXTENT_FLAG_LOGGING, &flags);
 588                 modified = !list_empty(&em->list);
 589                 if (no_splits)
 590                         goto next;
 591
 592                 if (em->start < start) {
 593                         split->start = em->start;
 594                         split->len = start - em->start;
 595
 596                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 597                                 split->orig_start = em->orig_start;
 598                                 split->block_start = em->block_start;
 599
 600                                 if (compressed)
 601                                         split->block_len = em->block_len;
 602                                 else
 603                                         split->block_len = split->len;
 604                                 split->orig_block_len = max(split->block_len,
 605                                                 em->orig_block_len);
 606                                 split->ram_bytes = em->ram_bytes;
 607                         } else {
 608                                 split->orig_start = split->start;
 609                                 split->block_len = 0;
 610                                 split->block_start = em->block_start;
 611                                 split->orig_block_len = 0;
 612                                 split->ram_bytes = split->len;
 613                         }
 614
 615                         split->generation = gen;
 616                         split->bdev = em->bdev;
 617                         split->flags = flags;
 618                         split->compress_type = em->compress_type;
 619                         replace_extent_mapping(em_tree, em, split, modified);
 620                         free_extent_map(split);
 621                         split = split2;
 622                         split2 = NULL;
 623                 }
 624                 if (testend && em->start + em->len > start + len) {
 625                         u64 diff = start + len - em->start;
 626
 627                         split->start = start + len;
 628                         split->len = em->start + em->len - (start + len);
 629                         split->bdev = em->bdev;
 630                         split->flags = flags;
 631                         split->compress_type = em->compress_type;
 632                         split->generation = gen;
 633
 634                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 635                                 split->orig_block_len = max(em->block_len,
 636                                                     em->orig_block_len);
 637
 638                                 split->ram_bytes = em->ram_bytes;
 639                                 if (compressed) {
 640                                         split->block_len = em->block_len;
 641                                         split->block_start = em->block_start;
 642                                         split->orig_start = em->orig_start;
 643                                 } else {
 644                                         split->block_len = split->len;
 645                                         split->block_start = em->block_start
 646                                                 + diff;
 647                                         split->orig_start = em->orig_start;
 648                                 }
 649                         } else {
 650                                 split->ram_bytes = split->len;
 651                                 split->orig_start = split->start;
 652                                 split->block_len = 0;
 653                                 split->block_start = em->block_start;
 654                                 split->orig_block_len = 0;
 655                         }
 656
 657                         if (extent_map_in_tree(em)) {
 658                                 replace_extent_mapping(em_tree, em, split,
 659                                                        modified);
 660                         } else {
 661                                 ret = add_extent_mapping(em_tree, split,
 662                                                          modified);
 663                                 ASSERT(ret == 0); /* Logic error */
 664                         }
 665                         free_extent_map(split);
 666                         split = NULL;
 667                 }
 668 next:
 669                 if (extent_map_in_tree(em))
 670                         remove_extent_mapping(em_tree, em);
 671                 write_unlock(&em_tree->lock);
 672
 673                 /* once for us */
 674                 free_extent_map(em);
 675                 /* once for the tree*/
 676                 free_extent_map(em);
 677         }
 678         if (split)
 679                 free_extent_map(split);
 680         if (split2)
 681                 free_extent_map(split2);
 682 }
 683
 684 /*
 685  * this is very complex, but the basic idea is to drop all extents
 686  * in the range start - end.  hint_block is filled in with a block number
 687  * that would be a good hint to the block allocator for this file.
 688  *
 689  * If an extent intersects the range but is not entirely inside the range
 690  * it is either truncated or split.  Anything entirely inside the range
 691  * is deleted from the tree.
 692  */
 693 int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
 694                          struct btrfs_root *root, struct inode *inode,
 695                          struct btrfs_path *path, u64 start, u64 end,
 696                          u64 *drop_end, int drop_cache,
 697                          int replace_extent,
 698                          u32 extent_item_size,
 699                          int *key_inserted)
 700 {
 701         struct extent_buffer *leaf;
 702         struct btrfs_file_extent_item *fi;
 703         struct btrfs_key key;
 704         struct btrfs_key new_key;
 705         u64 ino = btrfs_ino(inode);
 706         u64 search_start = start;
 707         u64 disk_bytenr = 0;
 708         u64 num_bytes = 0;
 709         u64 extent_offset = 0;
 710         u64 extent_end = 0;
 711         int del_nr = 0;
 712         int del_slot = 0;
 713         int extent_type;
 714         int recow;
 715         int ret;
 716         int modify_tree = -1;
 717         int update_refs = (root->ref_cows || root == root->fs_info->tree_root);
 718         int found = 0;
 719         int leafs_visited = 0;
 720
 721         if (drop_cache)
 722                 btrfs_drop_extent_cache(inode, start, end - 1, 0);
 723
 724         if (start >= BTRFS_I(inode)->disk_i_size && !replace_extent)
 725                 modify_tree = 0;
 726
 727         while (1) {
 728                 recow = 0;
 729                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
 730                                                search_start, modify_tree);
 731                 if (ret < 0)
 732                         break;
 733                 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
 734                         leaf = path->nodes[0];
 735                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
 736                         if (key.objectid == ino &&
 737                             key.type == BTRFS_EXTENT_DATA_KEY)
 738                                 path->slots[0]--;
 739                 }
 740                 ret = 0;
 741                 leafs_visited++;
 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                         leafs_visited++;
 754                         leaf = path->nodes[0];
 755                         recow = 1;
 756                 }
 757
 758                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 759                 if (key.objectid > ino ||
 760                     key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
 761                         break;
 762
 763                 fi = btrfs_item_ptr(leaf, path->slots[0],
 764                                     struct btrfs_file_extent_item);
 765                 extent_type = btrfs_file_extent_type(leaf, fi);
 766
 767                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
 768                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
 769                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 770                         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 771                         extent_offset = btrfs_file_extent_offset(leaf, fi);
 772                         extent_end = key.offset +
 773                                 btrfs_file_extent_num_bytes(leaf, fi);
 774                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 775                         extent_end = key.offset +
 776                                 btrfs_file_extent_inline_len(leaf,
 777                                                      path->slots[0], fi);
 778                 } else {
 779                         WARN_ON(1);
 780                         extent_end = search_start;
 781                 }
 782
 783                 if (extent_end <= search_start) {
 784                         path->slots[0]++;
 785                         goto next_slot;
 786                 }
 787
 788                 found = 1;
 789                 search_start = max(key.offset, start);
 790                 if (recow || !modify_tree) {
 791                         modify_tree = -1;
 792                         btrfs_release_path(path);
 793                         continue;
 794                 }
 795
 796                 /*
 797                  *     | - range to drop - |
 798                  *  | -------- extent -------- |
 799                  */
 800                 if (start > key.offset && end < extent_end) {
 801                         BUG_ON(del_nr > 0);
 802                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 803                                 ret = -EOPNOTSUPP;
 804                                 break;
 805                         }
 806
 807                         memcpy(&new_key, &key, sizeof(new_key));
 808                         new_key.offset = start;
 809                         ret = btrfs_duplicate_item(trans, root, path,
 810                                                    &new_key);
 811                         if (ret == -EAGAIN) {
 812                                 btrfs_release_path(path);
 813                                 continue;
 814                         }
 815                         if (ret < 0)
 816                                 break;
 817
 818                         leaf = path->nodes[0];
 819                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 820                                             struct btrfs_file_extent_item);
 821                         btrfs_set_file_extent_num_bytes(leaf, fi,
 822                                                         start - key.offset);
 823
 824                         fi = btrfs_item_ptr(leaf, path->slots[0],
 825                                             struct btrfs_file_extent_item);
 826
 827                         extent_offset += start - key.offset;
 828                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 829                         btrfs_set_file_extent_num_bytes(leaf, fi,
 830                                                         extent_end - start);
 831                         btrfs_mark_buffer_dirty(leaf);
 832
 833                         if (update_refs && disk_bytenr > 0) {
 834                                 ret = btrfs_inc_extent_ref(trans, root,
 835                                                 disk_bytenr, num_bytes, 0,
 836                                                 root->root_key.objectid,
 837                                                 new_key.objectid,
 838                                                 start - extent_offset, 0);
 839                                 BUG_ON(ret); /* -ENOMEM */
 840                         }
 841                         key.offset = start;
 842                 }
 843                 /*
 844                  *  | ---- range to drop ----- |
 845                  *      | -------- extent -------- |
 846                  */
 847                 if (start <= key.offset && end < extent_end) {
 848                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 849                                 ret = -EOPNOTSUPP;
 850                                 break;
 851                         }
 852
 853                         memcpy(&new_key, &key, sizeof(new_key));
 854                         new_key.offset = end;
 855                         btrfs_set_item_key_safe(root, path, &new_key);
 856
 857                         extent_offset += end - key.offset;
 858                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 859                         btrfs_set_file_extent_num_bytes(leaf, fi,
 860                                                         extent_end - end);
 861                         btrfs_mark_buffer_dirty(leaf);
 862                         if (update_refs && disk_bytenr > 0)
 863                                 inode_sub_bytes(inode, end - key.offset);
 864                         break;
 865                 }
 866
 867                 search_start = extent_end;
 868                 /*
 869                  *       | ---- range to drop ----- |
 870                  *  | -------- extent -------- |
 871                  */
 872                 if (start > key.offset && end >= extent_end) {
 873                         BUG_ON(del_nr > 0);
 874                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 875                                 ret = -EOPNOTSUPP;
 876                                 break;
 877                         }
 878
 879                         btrfs_set_file_extent_num_bytes(leaf, fi,
 880                                                         start - key.offset);
 881                         btrfs_mark_buffer_dirty(leaf);
 882                         if (update_refs && disk_bytenr > 0)
 883                                 inode_sub_bytes(inode, extent_end - start);
 884                         if (end == extent_end)
 885                                 break;
 886
 887                         path->slots[0]++;
 888                         goto next_slot;
 889                 }
 890
 891                 /*
 892                  *  | ---- range to drop ----- |
 893                  *    | ------ extent ------ |
 894                  */
 895                 if (start <= key.offset && end >= extent_end) {
 896                         if (del_nr == 0) {
 897                                 del_slot = path->slots[0];
 898                                 del_nr = 1;
 899                         } else {
 900                                 BUG_ON(del_slot + del_nr != path->slots[0]);
 901                                 del_nr++;
 902                         }
 903
 904                         if (update_refs &&
 905                             extent_type == BTRFS_FILE_EXTENT_INLINE) {
 906                                 inode_sub_bytes(inode,
 907                                                 extent_end - key.offset);
 908                                 extent_end = ALIGN(extent_end,
 909                                                    root->sectorsize);
 910                         } else if (update_refs && disk_bytenr > 0) {
 911                                 ret = btrfs_free_extent(trans, root,
 912                                                 disk_bytenr, num_bytes, 0,
 913                                                 root->root_key.objectid,
 914                                                 key.objectid, key.offset -
 915                                                 extent_offset, 0);
 916                                 BUG_ON(ret); /* -ENOMEM */
 917                                 inode_sub_bytes(inode,
 918                                                 extent_end - key.offset);
 919                         }
 920
 921                         if (end == extent_end)
 922                                 break;
 923
 924                         if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
 925                                 path->slots[0]++;
 926                                 goto next_slot;
 927                         }
 928
 929                         ret = btrfs_del_items(trans, root, path, del_slot,
 930                                               del_nr);
 931                         if (ret) {
 932                                 btrfs_abort_transaction(trans, root, ret);
 933                                 break;
 934                         }
 935
 936                         del_nr = 0;
 937                         del_slot = 0;
 938
 939                         btrfs_release_path(path);
 940                         continue;
 941                 }
 942
 943                 BUG_ON(1);
 944         }
 945
 946         if (!ret && del_nr > 0) {
 947                 /*
 948                  * Set path->slots[0] to first slot, so that after the delete
 949                  * if items are move off from our leaf to its immediate left or
 950                  * right neighbor leafs, we end up with a correct and adjusted
 951                  * path->slots[0] for our insertion (if replace_extent != 0).
 952                  */
 953                 path->slots[0] = del_slot;
 954                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
 955                 if (ret)
 956                         btrfs_abort_transaction(trans, root, ret);
 957         }
 958
 959         leaf = path->nodes[0];
 960         /*
 961          * If btrfs_del_items() was called, it might have deleted a leaf, in
 962          * which case it unlocked our path, so check path->locks[0] matches a
 963          * write lock.
 964          */
 965         if (!ret && replace_extent && leafs_visited == 1 &&
 966             (path->locks[0] == BTRFS_WRITE_LOCK_BLOCKING ||
 967              path->locks[0] == BTRFS_WRITE_LOCK) &&
 968             btrfs_leaf_free_space(root, leaf) >=
 969             sizeof(struct btrfs_item) + extent_item_size) {
 970
 971                 key.objectid = ino;
 972                 key.type = BTRFS_EXTENT_DATA_KEY;
 973                 key.offset = start;
 974                 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
 975                         struct btrfs_key slot_key;
 976
 977                         btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
 978                         if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
 979                                 path->slots[0]++;
 980                 }
 981                 setup_items_for_insert(root, path, &key,
 982                                        &extent_item_size,
 983                                        extent_item_size,
 984                                        sizeof(struct btrfs_item) +
 985                                        extent_item_size, 1);
 986                 *key_inserted = 1;
 987         }
 988
 989         if (!replace_extent || !(*key_inserted))
 990                 btrfs_release_path(path);
 991         if (drop_end)
 992                 *drop_end = found ? min(end, extent_end) : end;
 993         return ret;
 994 }
 995
 996 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
 997                        struct btrfs_root *root, struct inode *inode, u64 start,
 998                        u64 end, int drop_cache)
 999 {
1000         struct btrfs_path *path;
1001         int ret;
1002
1003         path = btrfs_alloc_path();
1004         if (!path)
1005                 return -ENOMEM;
1006         ret = __btrfs_drop_extents(trans, root, inode, path, start, end, NULL,
1007                                    drop_cache, 0, 0, NULL);
1008         btrfs_free_path(path);
1009         return ret;
1010 }
1011
1012 static int extent_mergeable(struct extent_buffer *leaf, int slot,
1013                             u64 objectid, u64 bytenr, u64 orig_offset,
1014                             u64 *start, u64 *end)
1015 {
1016         struct btrfs_file_extent_item *fi;
1017         struct btrfs_key key;
1018         u64 extent_end;
1019
1020         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1021                 return 0;
1022
1023         btrfs_item_key_to_cpu(leaf, &key, slot);
1024         if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1025                 return 0;
1026
1027         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1028         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1029             btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1030             btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1031             btrfs_file_extent_compression(leaf, fi) ||
1032             btrfs_file_extent_encryption(leaf, fi) ||
1033             btrfs_file_extent_other_encoding(leaf, fi))
1034                 return 0;
1035
1036         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1037         if ((*start && *start != key.offset) || (*end && *end != extent_end))
1038                 return 0;
1039
1040         *start = key.offset;
1041         *end = extent_end;
1042         return 1;
1043 }
1044
1045 /*
1046  * Mark extent in the range start - end as written.
1047  *
1048  * This changes extent type from 'pre-allocated' to 'regular'. If only
1049  * part of extent is marked as written, the extent will be split into
1050  * two or three.
1051  */
1052 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1053                               struct inode *inode, u64 start, u64 end)
1054 {
1055         struct btrfs_root *root = BTRFS_I(inode)->root;
1056         struct extent_buffer *leaf;
1057         struct btrfs_path *path;
1058         struct btrfs_file_extent_item *fi;
1059         struct btrfs_key key;
1060         struct btrfs_key new_key;
1061         u64 bytenr;
1062         u64 num_bytes;
1063         u64 extent_end;
1064         u64 orig_offset;
1065         u64 other_start;
1066         u64 other_end;
1067         u64 split;
1068         int del_nr = 0;
1069         int del_slot = 0;
1070         int recow;
1071         int ret;
1072         u64 ino = btrfs_ino(inode);
1073
1074         path = btrfs_alloc_path();
1075         if (!path)
1076                 return -ENOMEM;
1077 again:
1078         recow = 0;
1079         split = start;
1080         key.objectid = ino;
1081         key.type = BTRFS_EXTENT_DATA_KEY;
1082         key.offset = split;
1083
1084         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1085         if (ret < 0)
1086                 goto out;
1087         if (ret > 0 && path->slots[0] > 0)
1088                 path->slots[0]--;
1089
1090         leaf = path->nodes[0];
1091         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1092         BUG_ON(key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY);
1093         fi = btrfs_item_ptr(leaf, path->slots[0],
1094                             struct btrfs_file_extent_item);
1095         BUG_ON(btrfs_file_extent_type(leaf, fi) !=
1096                BTRFS_FILE_EXTENT_PREALLOC);
1097         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1098         BUG_ON(key.offset > start || extent_end < end);
1099
1100         bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1101         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1102         orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1103         memcpy(&new_key, &key, sizeof(new_key));
1104
1105         if (start == key.offset && end < extent_end) {
1106                 other_start = 0;
1107                 other_end = start;
1108                 if (extent_mergeable(leaf, path->slots[0] - 1,
1109                                      ino, bytenr, orig_offset,
1110                                      &other_start, &other_end)) {
1111                         new_key.offset = end;
1112                         btrfs_set_item_key_safe(root, path, &new_key);
1113                         fi = btrfs_item_ptr(leaf, path->slots[0],
1114                                             struct btrfs_file_extent_item);
1115                         btrfs_set_file_extent_generation(leaf, fi,
1116                                                          trans->transid);
1117                         btrfs_set_file_extent_num_bytes(leaf, fi,
1118                                                         extent_end - end);
1119                         btrfs_set_file_extent_offset(leaf, fi,
1120                                                      end - orig_offset);
1121                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1122                                             struct btrfs_file_extent_item);
1123                         btrfs_set_file_extent_generation(leaf, fi,
1124                                                          trans->transid);
1125                         btrfs_set_file_extent_num_bytes(leaf, fi,
1126                                                         end - other_start);
1127                         btrfs_mark_buffer_dirty(leaf);
1128                         goto out;
1129                 }
1130         }
1131
1132         if (start > key.offset && end == extent_end) {
1133                 other_start = end;
1134                 other_end = 0;
1135                 if (extent_mergeable(leaf, path->slots[0] + 1,
1136                                      ino, bytenr, orig_offset,
1137                                      &other_start, &other_end)) {
1138                         fi = btrfs_item_ptr(leaf, path->slots[0],
1139                                             struct btrfs_file_extent_item);
1140                         btrfs_set_file_extent_num_bytes(leaf, fi,
1141                                                         start - key.offset);
1142                         btrfs_set_file_extent_generation(leaf, fi,
1143                                                          trans->transid);
1144                         path->slots[0]++;
1145                         new_key.offset = start;
1146                         btrfs_set_item_key_safe(root, path, &new_key);
1147
1148                         fi = btrfs_item_ptr(leaf, path->slots[0],
1149                                             struct btrfs_file_extent_item);
1150                         btrfs_set_file_extent_generation(leaf, fi,
1151                                                          trans->transid);
1152                         btrfs_set_file_extent_num_bytes(leaf, fi,
1153                                                         other_end - start);
1154                         btrfs_set_file_extent_offset(leaf, fi,
1155                                                      start - orig_offset);
1156                         btrfs_mark_buffer_dirty(leaf);
1157                         goto out;
1158                 }
1159         }
1160
1161         while (start > key.offset || end < extent_end) {
1162                 if (key.offset == start)
1163                         split = end;
1164
1165                 new_key.offset = split;
1166                 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1167                 if (ret == -EAGAIN) {
1168                         btrfs_release_path(path);
1169                         goto again;
1170                 }
1171                 if (ret < 0) {
1172                         btrfs_abort_transaction(trans, root, ret);
1173                         goto out;
1174                 }
1175
1176                 leaf = path->nodes[0];
1177                 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1178                                     struct btrfs_file_extent_item);
1179                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1180                 btrfs_set_file_extent_num_bytes(leaf, fi,
1181                                                 split - key.offset);
1182
1183                 fi = btrfs_item_ptr(leaf, path->slots[0],
1184                                     struct btrfs_file_extent_item);
1185
1186                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1187                 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1188                 btrfs_set_file_extent_num_bytes(leaf, fi,
1189                                                 extent_end - split);
1190                 btrfs_mark_buffer_dirty(leaf);
1191
1192                 ret = btrfs_inc_extent_ref(trans, root, bytenr, num_bytes, 0,
1193                                            root->root_key.objectid,
1194                                            ino, orig_offset, 0);
1195                 BUG_ON(ret); /* -ENOMEM */
1196
1197                 if (split == start) {
1198                         key.offset = start;
1199                 } else {
1200                         BUG_ON(start != key.offset);
1201                         path->slots[0]--;
1202                         extent_end = end;
1203                 }
1204                 recow = 1;
1205         }
1206
1207         other_start = end;
1208         other_end = 0;
1209         if (extent_mergeable(leaf, path->slots[0] + 1,
1210                              ino, bytenr, orig_offset,
1211                              &other_start, &other_end)) {
1212                 if (recow) {
1213                         btrfs_release_path(path);
1214                         goto again;
1215                 }
1216                 extent_end = other_end;
1217                 del_slot = path->slots[0] + 1;
1218                 del_nr++;
1219                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1220                                         0, root->root_key.objectid,
1221                                         ino, orig_offset, 0);
1222                 BUG_ON(ret); /* -ENOMEM */
1223         }
1224         other_start = 0;
1225         other_end = start;
1226         if (extent_mergeable(leaf, path->slots[0] - 1,
1227                              ino, bytenr, orig_offset,
1228                              &other_start, &other_end)) {
1229                 if (recow) {
1230                         btrfs_release_path(path);
1231                         goto again;
1232                 }
1233                 key.offset = other_start;
1234                 del_slot = path->slots[0];
1235                 del_nr++;
1236                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1237                                         0, root->root_key.objectid,
1238                                         ino, orig_offset, 0);
1239                 BUG_ON(ret); /* -ENOMEM */
1240         }
1241         if (del_nr == 0) {
1242                 fi = btrfs_item_ptr(leaf, path->slots[0],
1243                            struct btrfs_file_extent_item);
1244                 btrfs_set_file_extent_type(leaf, fi,
1245                                            BTRFS_FILE_EXTENT_REG);
1246                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1247                 btrfs_mark_buffer_dirty(leaf);
1248         } else {
1249                 fi = btrfs_item_ptr(leaf, del_slot - 1,
1250                            struct btrfs_file_extent_item);
1251                 btrfs_set_file_extent_type(leaf, fi,
1252                                            BTRFS_FILE_EXTENT_REG);
1253                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1254                 btrfs_set_file_extent_num_bytes(leaf, fi,
1255                                                 extent_end - key.offset);
1256                 btrfs_mark_buffer_dirty(leaf);
1257
1258                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1259                 if (ret < 0) {
1260                         btrfs_abort_transaction(trans, root, ret);
1261                         goto out;
1262                 }
1263         }
1264 out:
1265         btrfs_free_path(path);
1266         return 0;
1267 }
1268
1269 /*
1270  * on error we return an unlocked page and the error value
1271  * on success we return a locked page and 0
1272  */
1273 static int prepare_uptodate_page(struct page *page, u64 pos,
1274                                  bool force_uptodate)
1275 {
1276         int ret = 0;
1277
1278         if (((pos & (PAGE_CACHE_SIZE - 1)) || force_uptodate) &&
1279             !PageUptodate(page)) {
1280                 ret = btrfs_readpage(NULL, page);
1281                 if (ret)
1282                         return ret;
1283                 lock_page(page);
1284                 if (!PageUptodate(page)) {
1285                         unlock_page(page);
1286                         return -EIO;
1287                 }
1288         }
1289         return 0;
1290 }
1291
1292 /*
1293  * this just gets pages into the page cache and locks them down.
1294  */
1295 static noinline int prepare_pages(struct inode *inode, struct page **pages,
1296                                   size_t num_pages, loff_t pos,
1297                                   size_t write_bytes, bool force_uptodate)
1298 {
1299         int i;
1300         unsigned long index = pos >> PAGE_CACHE_SHIFT;
1301         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1302         int err = 0;
1303         int faili;
1304
1305         for (i = 0; i < num_pages; i++) {
1306                 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1307                                                mask | __GFP_WRITE);
1308                 if (!pages[i]) {
1309                         faili = i - 1;
1310                         err = -ENOMEM;
1311                         goto fail;
1312                 }
1313
1314                 if (i == 0)
1315                         err = prepare_uptodate_page(pages[i], pos,
1316                                                     force_uptodate);
1317                 if (i == num_pages - 1)
1318                         err = prepare_uptodate_page(pages[i],
1319                                                     pos + write_bytes, false);
1320                 if (err) {
1321                         page_cache_release(pages[i]);
1322                         faili = i - 1;
1323                         goto fail;
1324                 }
1325                 wait_on_page_writeback(pages[i]);
1326         }
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 /*
1340  * This function locks the extent and properly waits for data=ordered extents
1341  * to finish before allowing the pages to be modified if need.
1342  *
1343  * The return value:
1344  * 1 - the extent is locked
1345  * 0 - the extent is not locked, and everything is OK
1346  * -EAGAIN - need re-prepare the pages
1347  * the other < 0 number - Something wrong happens
1348  */
1349 static noinline int
1350 lock_and_cleanup_extent_if_need(struct inode *inode, struct page **pages,
1351                                 size_t num_pages, loff_t pos,
1352                                 u64 *lockstart, u64 *lockend,
1353                                 struct extent_state **cached_state)
1354 {
1355         u64 start_pos;
1356         u64 last_pos;
1357         int i;
1358         int ret = 0;
1359
1360         start_pos = pos & ~((u64)PAGE_CACHE_SIZE - 1);
1361         last_pos = start_pos + ((u64)num_pages << PAGE_CACHE_SHIFT) - 1;
1362
1363         if (start_pos < inode->i_size) {
1364                 struct btrfs_ordered_extent *ordered;
1365                 lock_extent_bits(&BTRFS_I(inode)->io_tree,
1366                                  start_pos, last_pos, 0, cached_state);
1367                 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1368                                                      last_pos - start_pos + 1);
1369                 if (ordered &&
1370                     ordered->file_offset + ordered->len > start_pos &&
1371                     ordered->file_offset <= last_pos) {
1372                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1373                                              start_pos, last_pos,
1374                                              cached_state, GFP_NOFS);
1375                         for (i = 0; i < num_pages; i++) {
1376                                 unlock_page(pages[i]);
1377                                 page_cache_release(pages[i]);
1378                         }
1379                         btrfs_start_ordered_extent(inode, ordered, 1);
1380                         btrfs_put_ordered_extent(ordered);
1381                         return -EAGAIN;
1382                 }
1383                 if (ordered)
1384                         btrfs_put_ordered_extent(ordered);
1385
1386                 clear_extent_bit(&BTRFS_I(inode)->io_tree, start_pos,
1387                                   last_pos, EXTENT_DIRTY | EXTENT_DELALLOC |
1388                                   EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
1389                                   0, 0, cached_state, GFP_NOFS);
1390                 *lockstart = start_pos;
1391                 *lockend = last_pos;
1392                 ret = 1;
1393         }
1394
1395         for (i = 0; i < num_pages; i++) {
1396                 if (clear_page_dirty_for_io(pages[i]))
1397                         account_page_redirty(pages[i]);
1398                 set_page_extent_mapped(pages[i]);
1399                 WARN_ON(!PageLocked(pages[i]));
1400         }
1401
1402         return ret;
1403 }
1404
1405 static noinline int check_can_nocow(struct inode *inode, loff_t pos,
1406                                     size_t *write_bytes)
1407 {
1408         struct btrfs_root *root = BTRFS_I(inode)->root;
1409         struct btrfs_ordered_extent *ordered;
1410         u64 lockstart, lockend;
1411         u64 num_bytes;
1412         int ret;
1413
1414         ret = btrfs_start_nocow_write(root);
1415         if (!ret)
1416                 return -ENOSPC;
1417
1418         lockstart = round_down(pos, root->sectorsize);
1419         lockend = round_up(pos + *write_bytes, root->sectorsize) - 1;
1420
1421         while (1) {
1422                 lock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend);
1423                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
1424                                                      lockend - lockstart + 1);
1425                 if (!ordered) {
1426                         break;
1427                 }
1428                 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend);
1429                 btrfs_start_ordered_extent(inode, ordered, 1);
1430                 btrfs_put_ordered_extent(ordered);
1431         }
1432
1433         num_bytes = lockend - lockstart + 1;
1434         ret = can_nocow_extent(inode, lockstart, &num_bytes, NULL, NULL, NULL);
1435         if (ret <= 0) {
1436                 ret = 0;
1437                 btrfs_end_nocow_write(root);
1438         } else {
1439                 *write_bytes = min_t(size_t, *write_bytes ,
1440                                      num_bytes - pos + lockstart);
1441         }
1442
1443         unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend);
1444
1445         return ret;
1446 }
1447
1448 static noinline ssize_t __btrfs_buffered_write(struct file *file,
1449                                                struct iov_iter *i,
1450                                                loff_t pos)
1451 {
1452         struct inode *inode = file_inode(file);
1453         struct btrfs_root *root = BTRFS_I(inode)->root;
1454         struct page **pages = NULL;
1455         struct extent_state *cached_state = NULL;
1456         u64 release_bytes = 0;
1457         u64 lockstart;
1458         u64 lockend;
1459         unsigned long first_index;
1460         size_t num_written = 0;
1461         int nrptrs;
1462         int ret = 0;
1463         bool only_release_metadata = false;
1464         bool force_page_uptodate = false;
1465         bool need_unlock;
1466
1467         nrptrs = min((iov_iter_count(i) + PAGE_CACHE_SIZE - 1) /
1468                      PAGE_CACHE_SIZE, PAGE_CACHE_SIZE /
1469                      (sizeof(struct page *)));
1470         nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1471         nrptrs = max(nrptrs, 8);
1472         pages = kmalloc(nrptrs * sizeof(struct page *), GFP_KERNEL);
1473         if (!pages)
1474                 return -ENOMEM;
1475
1476         first_index = pos >> PAGE_CACHE_SHIFT;
1477
1478         while (iov_iter_count(i) > 0) {
1479                 size_t offset = pos & (PAGE_CACHE_SIZE - 1);
1480                 size_t write_bytes = min(iov_iter_count(i),
1481                                          nrptrs * (size_t)PAGE_CACHE_SIZE -
1482                                          offset);
1483                 size_t num_pages = (write_bytes + offset +
1484                                     PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1485                 size_t reserve_bytes;
1486                 size_t dirty_pages;
1487                 size_t copied;
1488
1489                 WARN_ON(num_pages > nrptrs);
1490
1491                 /*
1492                  * Fault pages before locking them in prepare_pages
1493                  * to avoid recursive lock
1494                  */
1495                 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1496                         ret = -EFAULT;
1497                         break;
1498                 }
1499
1500                 reserve_bytes = num_pages << PAGE_CACHE_SHIFT;
1501                 ret = btrfs_check_data_free_space(inode, reserve_bytes);
1502                 if (ret == -ENOSPC &&
1503                     (BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1504                                               BTRFS_INODE_PREALLOC))) {
1505                         ret = check_can_nocow(inode, pos, &write_bytes);
1506                         if (ret > 0) {
1507                                 only_release_metadata = true;
1508                                 /*
1509                                  * our prealloc extent may be smaller than
1510                                  * write_bytes, so scale down.
1511                                  */
1512                                 num_pages = (write_bytes + offset +
1513                                              PAGE_CACHE_SIZE - 1) >>
1514                                         PAGE_CACHE_SHIFT;
1515                                 reserve_bytes = num_pages << PAGE_CACHE_SHIFT;
1516                                 ret = 0;
1517                         } else {
1518                                 ret = -ENOSPC;
1519                         }
1520                 }
1521
1522                 if (ret)
1523                         break;
1524
1525                 ret = btrfs_delalloc_reserve_metadata(inode, reserve_bytes);
1526                 if (ret) {
1527                         if (!only_release_metadata)
1528                                 btrfs_free_reserved_data_space(inode,
1529                                                                reserve_bytes);
1530                         else
1531                                 btrfs_end_nocow_write(root);
1532                         break;
1533                 }
1534
1535                 release_bytes = reserve_bytes;
1536                 need_unlock = false;
1537 again:
1538                 /*
1539                  * This is going to setup the pages array with the number of
1540                  * pages we want, so we don't really need to worry about the
1541                  * contents of pages from loop to loop
1542                  */
1543                 ret = prepare_pages(inode, pages, num_pages,
1544                                     pos, write_bytes,
1545                                     force_page_uptodate);
1546                 if (ret)
1547                         break;
1548
1549                 ret = lock_and_cleanup_extent_if_need(inode, pages, num_pages,
1550                                                       pos, &lockstart, &lockend,
1551                                                       &cached_state);
1552                 if (ret < 0) {
1553                         if (ret == -EAGAIN)
1554                                 goto again;
1555                         break;
1556                 } else if (ret > 0) {
1557                         need_unlock = true;
1558                         ret = 0;
1559                 }
1560
1561                 copied = btrfs_copy_from_user(pos, num_pages,
1562                                            write_bytes, pages, i);
1563
1564                 /*
1565                  * if we have trouble faulting in the pages, fall
1566                  * back to one page at a time
1567                  */
1568                 if (copied < write_bytes)
1569                         nrptrs = 1;
1570
1571                 if (copied == 0) {
1572                         force_page_uptodate = true;
1573                         dirty_pages = 0;
1574                 } else {
1575                         force_page_uptodate = false;
1576                         dirty_pages = (copied + offset +
1577                                        PAGE_CACHE_SIZE - 1) >>
1578                                        PAGE_CACHE_SHIFT;
1579                 }
1580
1581                 /*
1582                  * If we had a short copy we need to release the excess delaloc
1583                  * bytes we reserved.  We need to increment outstanding_extents
1584                  * because btrfs_delalloc_release_space will decrement it, but
1585                  * we still have an outstanding extent for the chunk we actually
1586                  * managed to copy.
1587                  */
1588                 if (num_pages > dirty_pages) {
1589                         release_bytes = (num_pages - dirty_pages) <<
1590                                 PAGE_CACHE_SHIFT;
1591                         if (copied > 0) {
1592                                 spin_lock(&BTRFS_I(inode)->lock);
1593                                 BTRFS_I(inode)->outstanding_extents++;
1594                                 spin_unlock(&BTRFS_I(inode)->lock);
1595                         }
1596                         if (only_release_metadata)
1597                                 btrfs_delalloc_release_metadata(inode,
1598                                                                 release_bytes);
1599                         else
1600                                 btrfs_delalloc_release_space(inode,
1601                                                              release_bytes);
1602                 }
1603
1604                 release_bytes = dirty_pages << PAGE_CACHE_SHIFT;
1605
1606                 if (copied > 0)
1607                         ret = btrfs_dirty_pages(root, inode, pages,
1608                                                 dirty_pages, pos, copied,
1609                                                 NULL);
1610                 if (need_unlock)
1611                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1612                                              lockstart, lockend, &cached_state,
1613                                              GFP_NOFS);
1614                 if (ret) {
1615                         btrfs_drop_pages(pages, num_pages);
1616                         break;
1617                 }
1618
1619                 release_bytes = 0;
1620                 if (only_release_metadata)
1621                         btrfs_end_nocow_write(root);
1622
1623                 if (only_release_metadata && copied > 0) {
1624                         u64 lockstart = round_down(pos, root->sectorsize);
1625                         u64 lockend = lockstart +
1626                                 (dirty_pages << PAGE_CACHE_SHIFT) - 1;
1627
1628                         set_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
1629                                        lockend, EXTENT_NORESERVE, NULL,
1630                                        NULL, GFP_NOFS);
1631                         only_release_metadata = false;
1632                 }
1633
1634                 btrfs_drop_pages(pages, num_pages);
1635
1636                 cond_resched();
1637
1638                 balance_dirty_pages_ratelimited(inode->i_mapping);
1639                 if (dirty_pages < (root->leafsize >> PAGE_CACHE_SHIFT) + 1)
1640                         btrfs_btree_balance_dirty(root);
1641
1642                 pos += copied;
1643                 num_written += copied;
1644         }
1645
1646         kfree(pages);
1647
1648         if (release_bytes) {
1649                 if (only_release_metadata) {
1650                         btrfs_end_nocow_write(root);
1651                         btrfs_delalloc_release_metadata(inode, release_bytes);
1652                 } else {
1653                         btrfs_delalloc_release_space(inode, release_bytes);
1654                 }
1655         }
1656
1657         return num_written ? num_written : ret;
1658 }
1659
1660 static ssize_t __btrfs_direct_write(struct kiocb *iocb,
1661                                     const struct iovec *iov,
1662                                     unsigned long nr_segs, loff_t pos,
1663                                     size_t count, size_t ocount)
1664 {
1665         struct file *file = iocb->ki_filp;
1666         struct iov_iter i;
1667         ssize_t written;
1668         ssize_t written_buffered;
1669         loff_t endbyte;
1670         int err;
1671
1672         written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
1673                                             count, ocount);
1674
1675         if (written < 0 || written == count)
1676                 return written;
1677
1678         pos += written;
1679         count -= written;
1680         iov_iter_init(&i, iov, nr_segs, count, written);
1681         written_buffered = __btrfs_buffered_write(file, &i, pos);
1682         if (written_buffered < 0) {
1683                 err = written_buffered;
1684                 goto out;
1685         }
1686         endbyte = pos + written_buffered - 1;
1687         err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
1688         if (err)
1689                 goto out;
1690         written += written_buffered;
1691         iocb->ki_pos = pos + written_buffered;
1692         invalidate_mapping_pages(file->f_mapping, pos >> PAGE_CACHE_SHIFT,
1693                                  endbyte >> PAGE_CACHE_SHIFT);
1694 out:
1695         return written ? written : err;
1696 }
1697
1698 static void update_time_for_write(struct inode *inode)
1699 {
1700         struct timespec now;
1701
1702         if (IS_NOCMTIME(inode))
1703                 return;
1704
1705         now = current_fs_time(inode->i_sb);
1706         if (!timespec_equal(&inode->i_mtime, &now))
1707                 inode->i_mtime = now;
1708
1709         if (!timespec_equal(&inode->i_ctime, &now))
1710                 inode->i_ctime = now;
1711
1712         if (IS_I_VERSION(inode))
1713                 inode_inc_iversion(inode);
1714 }
1715
1716 static ssize_t btrfs_file_aio_write(struct kiocb *iocb,
1717                                     const struct iovec *iov,
1718                                     unsigned long nr_segs, loff_t pos)
1719 {
1720         struct file *file = iocb->ki_filp;
1721         struct inode *inode = file_inode(file);
1722         struct btrfs_root *root = BTRFS_I(inode)->root;
1723         u64 start_pos;
1724         u64 end_pos;
1725         ssize_t num_written = 0;
1726         ssize_t err = 0;
1727         size_t count, ocount;
1728         bool sync = (file->f_flags & O_DSYNC) || IS_SYNC(file->f_mapping->host);
1729
1730         mutex_lock(&inode->i_mutex);
1731
1732         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
1733         if (err) {
1734                 mutex_unlock(&inode->i_mutex);
1735                 goto out;
1736         }
1737         count = ocount;
1738
1739         current->backing_dev_info = inode->i_mapping->backing_dev_info;
1740         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
1741         if (err) {
1742                 mutex_unlock(&inode->i_mutex);
1743                 goto out;
1744         }
1745
1746         if (count == 0) {
1747                 mutex_unlock(&inode->i_mutex);
1748                 goto out;
1749         }
1750
1751         err = file_remove_suid(file);
1752         if (err) {
1753                 mutex_unlock(&inode->i_mutex);
1754                 goto out;
1755         }
1756
1757         /*
1758          * If BTRFS flips readonly due to some impossible error
1759          * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1760          * although we have opened a file as writable, we have
1761          * to stop this write operation to ensure FS consistency.
1762          */
1763         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state)) {
1764                 mutex_unlock(&inode->i_mutex);
1765                 err = -EROFS;
1766                 goto out;
1767         }
1768
1769         /*
1770          * We reserve space for updating the inode when we reserve space for the
1771          * extent we are going to write, so we will enospc out there.  We don't
1772          * need to start yet another transaction to update the inode as we will
1773          * update the inode when we finish writing whatever data we write.
1774          */
1775         update_time_for_write(inode);
1776
1777         start_pos = round_down(pos, root->sectorsize);
1778         if (start_pos > i_size_read(inode)) {
1779                 /* Expand hole size to cover write data, preventing empty gap */
1780                 end_pos = round_up(pos + count, root->sectorsize);
1781                 err = btrfs_cont_expand(inode, i_size_read(inode), end_pos);
1782                 if (err) {
1783                         mutex_unlock(&inode->i_mutex);
1784                         goto out;
1785                 }
1786         }
1787
1788         if (sync)
1789                 atomic_inc(&BTRFS_I(inode)->sync_writers);
1790
1791         if (unlikely(file->f_flags & O_DIRECT)) {
1792                 num_written = __btrfs_direct_write(iocb, iov, nr_segs,
1793                                                    pos, count, ocount);
1794         } else {
1795                 struct iov_iter i;
1796
1797                 iov_iter_init(&i, iov, nr_segs, count, num_written);
1798
1799                 num_written = __btrfs_buffered_write(file, &i, pos);
1800                 if (num_written > 0)
1801                         iocb->ki_pos = pos + num_written;
1802         }
1803
1804         mutex_unlock(&inode->i_mutex);
1805
1806         /*
1807          * we want to make sure fsync finds this change
1808          * but we haven't joined a transaction running right now.
1809          *
1810          * Later on, someone is sure to update the inode and get the
1811          * real transid recorded.
1812          *
1813          * We set last_trans now to the fs_info generation + 1,
1814          * this will either be one more than the running transaction
1815          * or the generation used for the next transaction if there isn't
1816          * one running right now.
1817          *
1818          * We also have to set last_sub_trans to the current log transid,
1819          * otherwise subsequent syncs to a file that's been synced in this
1820          * transaction will appear to have already occured.
1821          */
1822         BTRFS_I(inode)->last_trans = root->fs_info->generation + 1;
1823         BTRFS_I(inode)->last_sub_trans = root->log_transid;
1824         if (num_written > 0) {
1825                 err = generic_write_sync(file, pos, num_written);
1826                 if (err < 0)
1827                         num_written = err;
1828         }
1829
1830         if (sync)
1831                 atomic_dec(&BTRFS_I(inode)->sync_writers);
1832 out:
1833         current->backing_dev_info = NULL;
1834         return num_written ? num_written : err;
1835 }
1836
1837 int btrfs_release_file(struct inode *inode, struct file *filp)
1838 {
1839         /*
1840          * ordered_data_close is set by settattr when we are about to truncate
1841          * a file from a non-zero size to a zero size.  This tries to
1842          * flush down new bytes that may have been written if the
1843          * application were using truncate to replace a file in place.
1844          */
1845         if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
1846                                &BTRFS_I(inode)->runtime_flags)) {
1847                 struct btrfs_trans_handle *trans;
1848                 struct btrfs_root *root = BTRFS_I(inode)->root;
1849
1850                 /*
1851                  * We need to block on a committing transaction to keep us from
1852                  * throwing a ordered operation on to the list and causing
1853                  * something like sync to deadlock trying to flush out this
1854                  * inode.
1855                  */
1856                 trans = btrfs_start_transaction(root, 0);
1857                 if (IS_ERR(trans))
1858                         return PTR_ERR(trans);
1859                 btrfs_add_ordered_operation(trans, BTRFS_I(inode)->root, inode);
1860                 btrfs_end_transaction(trans, root);
1861                 if (inode->i_size > BTRFS_ORDERED_OPERATIONS_FLUSH_LIMIT)
1862                         filemap_flush(inode->i_mapping);
1863         }
1864         if (filp->private_data)
1865                 btrfs_ioctl_trans_end(filp);
1866         return 0;
1867 }
1868
1869 /*
1870  * fsync call for both files and directories.  This logs the inode into
1871  * the tree log instead of forcing full commits whenever possible.
1872  *
1873  * It needs to call filemap_fdatawait so that all ordered extent updates are
1874  * in the metadata btree are up to date for copying to the log.
1875  *
1876  * It drops the inode mutex before doing the tree log commit.  This is an
1877  * important optimization for directories because holding the mutex prevents
1878  * new operations on the dir while we write to disk.
1879  */
1880 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
1881 {
1882         struct dentry *dentry = file->f_path.dentry;
1883         struct inode *inode = dentry->d_inode;
1884         struct btrfs_root *root = BTRFS_I(inode)->root;
1885         struct btrfs_trans_handle *trans;
1886         struct btrfs_log_ctx ctx;
1887         int ret = 0;
1888         bool full_sync = 0;
1889
1890         trace_btrfs_sync_file(file, datasync);
1891
1892         /*
1893          * We write the dirty pages in the range and wait until they complete
1894          * out of the ->i_mutex. If so, we can flush the dirty pages by
1895          * multi-task, and make the performance up.  See
1896          * btrfs_wait_ordered_range for an explanation of the ASYNC check.
1897          */
1898         atomic_inc(&BTRFS_I(inode)->sync_writers);
1899         ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
1900         if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1901                              &BTRFS_I(inode)->runtime_flags))
1902                 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
1903         atomic_dec(&BTRFS_I(inode)->sync_writers);
1904         if (ret)
1905                 return ret;
1906
1907         mutex_lock(&inode->i_mutex);
1908
1909         /*
1910          * We flush the dirty pages again to avoid some dirty pages in the
1911          * range being left.
1912          */
1913         atomic_inc(&root->log_batch);
1914         full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1915                              &BTRFS_I(inode)->runtime_flags);
1916         if (full_sync) {
1917                 ret = btrfs_wait_ordered_range(inode, start, end - start + 1);
1918                 if (ret) {
1919                         mutex_unlock(&inode->i_mutex);
1920                         goto out;
1921                 }
1922         }
1923         atomic_inc(&root->log_batch);
1924
1925         /*
1926          * check the transaction that last modified this inode
1927          * and see if its already been committed
1928          */
1929         if (!BTRFS_I(inode)->last_trans) {
1930                 mutex_unlock(&inode->i_mutex);
1931                 goto out;
1932         }
1933
1934         /*
1935          * if the last transaction that changed this file was before
1936          * the current transaction, we can bail out now without any
1937          * syncing
1938          */
1939         smp_mb();
1940         if (btrfs_inode_in_log(inode, root->fs_info->generation) ||
1941             BTRFS_I(inode)->last_trans <=
1942             root->fs_info->last_trans_committed) {
1943                 BTRFS_I(inode)->last_trans = 0;
1944
1945                 /*
1946                  * We'v had everything committed since the last time we were
1947                  * modified so clear this flag in case it was set for whatever
1948                  * reason, it's no longer relevant.
1949                  */
1950                 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1951                           &BTRFS_I(inode)->runtime_flags);
1952                 mutex_unlock(&inode->i_mutex);
1953                 goto out;
1954         }
1955
1956         /*
1957          * ok we haven't committed the transaction yet, lets do a commit
1958          */
1959         if (file->private_data)
1960                 btrfs_ioctl_trans_end(file);
1961
1962         /*
1963          * We use start here because we will need to wait on the IO to complete
1964          * in btrfs_sync_log, which could require joining a transaction (for
1965          * example checking cross references in the nocow path).  If we use join
1966          * here we could get into a situation where we're waiting on IO to
1967          * happen that is blocked on a transaction trying to commit.  With start
1968          * we inc the extwriter counter, so we wait for all extwriters to exit
1969          * before we start blocking join'ers.  This comment is to keep somebody
1970          * from thinking they are super smart and changing this to
1971          * btrfs_join_transaction *cough*Josef*cough*.
1972          */
1973         trans = btrfs_start_transaction(root, 0);
1974         if (IS_ERR(trans)) {
1975                 ret = PTR_ERR(trans);
1976                 mutex_unlock(&inode->i_mutex);
1977                 goto out;
1978         }
1979         trans->sync = true;
1980
1981         btrfs_init_log_ctx(&ctx);
1982
1983         ret = btrfs_log_dentry_safe(trans, root, dentry, &ctx);
1984         if (ret < 0) {
1985                 /* Fallthrough and commit/free transaction. */
1986                 ret = 1;
1987         }
1988
1989         /* we've logged all the items and now have a consistent
1990          * version of the file in the log.  It is possible that
1991          * someone will come in and modify the file, but that's
1992          * fine because the log is consistent on disk, and we
1993          * have references to all of the file's extents
1994          *
1995          * It is possible that someone will come in and log the
1996          * file again, but that will end up using the synchronization
1997          * inside btrfs_sync_log to keep things safe.
1998          */
1999         mutex_unlock(&inode->i_mutex);
2000
2001         if (ret != BTRFS_NO_LOG_SYNC) {
2002                 if (!ret) {
2003                         ret = btrfs_sync_log(trans, root, &ctx);
2004                         if (!ret) {
2005                                 ret = btrfs_end_transaction(trans, root);
2006                                 goto out;
2007                         }
2008                 }
2009                 if (!full_sync) {
2010                         ret = btrfs_wait_ordered_range(inode, start,
2011                                                        end - start + 1);
2012                         if (ret)
2013                                 goto out;
2014                 }
2015                 ret = btrfs_commit_transaction(trans, root);
2016         } else {
2017                 ret = btrfs_end_transaction(trans, root);
2018         }
2019 out:
2020         return ret > 0 ? -EIO : ret;
2021 }
2022
2023 static const struct vm_operations_struct btrfs_file_vm_ops = {
2024         .fault          = filemap_fault,
2025         .map_pages      = filemap_map_pages,
2026         .page_mkwrite   = btrfs_page_mkwrite,
2027         .remap_pages    = generic_file_remap_pages,
2028 };
2029
2030 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
2031 {
2032         struct address_space *mapping = filp->f_mapping;
2033
2034         if (!mapping->a_ops->readpage)
2035                 return -ENOEXEC;
2036
2037         file_accessed(filp);
2038         vma->vm_ops = &btrfs_file_vm_ops;
2039
2040         return 0;
2041 }
2042
2043 static int hole_mergeable(struct inode *inode, struct extent_buffer *leaf,
2044                           int slot, u64 start, u64 end)
2045 {
2046         struct btrfs_file_extent_item *fi;
2047         struct btrfs_key key;
2048
2049         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2050                 return 0;
2051
2052         btrfs_item_key_to_cpu(leaf, &key, slot);
2053         if (key.objectid != btrfs_ino(inode) ||
2054             key.type != BTRFS_EXTENT_DATA_KEY)
2055                 return 0;
2056
2057         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2058
2059         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2060                 return 0;
2061
2062         if (btrfs_file_extent_disk_bytenr(leaf, fi))
2063                 return 0;
2064
2065         if (key.offset == end)
2066                 return 1;
2067         if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2068                 return 1;
2069         return 0;
2070 }
2071
2072 static int fill_holes(struct btrfs_trans_handle *trans, struct inode *inode,
2073                       struct btrfs_path *path, u64 offset, u64 end)
2074 {
2075         struct btrfs_root *root = BTRFS_I(inode)->root;
2076         struct extent_buffer *leaf;
2077         struct btrfs_file_extent_item *fi;
2078         struct extent_map *hole_em;
2079         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2080         struct btrfs_key key;
2081         int ret;
2082
2083         if (btrfs_fs_incompat(root->fs_info, NO_HOLES))
2084                 goto out;
2085
2086         key.objectid = btrfs_ino(inode);
2087         key.type = BTRFS_EXTENT_DATA_KEY;
2088         key.offset = offset;
2089
2090         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2091         if (ret < 0)
2092                 return ret;
2093         BUG_ON(!ret);
2094
2095         leaf = path->nodes[0];
2096         if (hole_mergeable(inode, leaf, path->slots[0]-1, offset, end)) {
2097                 u64 num_bytes;
2098
2099                 path->slots[0]--;
2100                 fi = btrfs_item_ptr(leaf, path->slots[0],
2101                                     struct btrfs_file_extent_item);
2102                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2103                         end - offset;
2104                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2105                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2106                 btrfs_set_file_extent_offset(leaf, fi, 0);
2107                 btrfs_mark_buffer_dirty(leaf);
2108                 goto out;
2109         }
2110
2111         if (hole_mergeable(inode, leaf, path->slots[0]+1, offset, end)) {
2112                 u64 num_bytes;
2113
2114                 path->slots[0]++;
2115                 key.offset = offset;
2116                 btrfs_set_item_key_safe(root, path, &key);
2117                 fi = btrfs_item_ptr(leaf, path->slots[0],
2118                                     struct btrfs_file_extent_item);
2119                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2120                         offset;
2121                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2122                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2123                 btrfs_set_file_extent_offset(leaf, fi, 0);
2124                 btrfs_mark_buffer_dirty(leaf);
2125                 goto out;
2126         }
2127         btrfs_release_path(path);
2128
2129         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode), offset,
2130                                        0, 0, end - offset, 0, end - offset,
2131                                        0, 0, 0);
2132         if (ret)
2133                 return ret;
2134
2135 out:
2136         btrfs_release_path(path);
2137
2138         hole_em = alloc_extent_map();
2139         if (!hole_em) {
2140                 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2141                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2142                         &BTRFS_I(inode)->runtime_flags);
2143         } else {
2144                 hole_em->start = offset;
2145                 hole_em->len = end - offset;
2146                 hole_em->ram_bytes = hole_em->len;
2147                 hole_em->orig_start = offset;
2148
2149                 hole_em->block_start = EXTENT_MAP_HOLE;
2150                 hole_em->block_len = 0;
2151                 hole_em->orig_block_len = 0;
2152                 hole_em->bdev = root->fs_info->fs_devices->latest_bdev;
2153                 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2154                 hole_em->generation = trans->transid;
2155
2156                 do {
2157                         btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2158                         write_lock(&em_tree->lock);
2159                         ret = add_extent_mapping(em_tree, hole_em, 1);
2160                         write_unlock(&em_tree->lock);
2161                 } while (ret == -EEXIST);
2162                 free_extent_map(hole_em);
2163                 if (ret)
2164                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2165                                 &BTRFS_I(inode)->runtime_flags);
2166         }
2167
2168         return 0;
2169 }
2170
2171 static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2172 {
2173         struct btrfs_root *root = BTRFS_I(inode)->root;
2174         struct extent_state *cached_state = NULL;
2175         struct btrfs_path *path;
2176         struct btrfs_block_rsv *rsv;
2177         struct btrfs_trans_handle *trans;
2178         u64 lockstart = round_up(offset, BTRFS_I(inode)->root->sectorsize);
2179         u64 lockend = round_down(offset + len,
2180                                  BTRFS_I(inode)->root->sectorsize) - 1;
2181         u64 cur_offset = lockstart;
2182         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
2183         u64 drop_end;
2184         int ret = 0;
2185         int err = 0;
2186         int rsv_count;
2187         bool same_page = ((offset >> PAGE_CACHE_SHIFT) ==
2188                           ((offset + len - 1) >> PAGE_CACHE_SHIFT));
2189         bool no_holes = btrfs_fs_incompat(root->fs_info, NO_HOLES);
2190         u64 ino_size = round_up(inode->i_size, PAGE_CACHE_SIZE);
2191
2192         ret = btrfs_wait_ordered_range(inode, offset, len);
2193         if (ret)
2194                 return ret;
2195
2196         mutex_lock(&inode->i_mutex);
2197         /*
2198          * We needn't truncate any page which is beyond the end of the file
2199          * because we are sure there is no data there.
2200          */
2201         /*
2202          * Only do this if we are in the same page and we aren't doing the
2203          * entire page.
2204          */
2205         if (same_page && len < PAGE_CACHE_SIZE) {
2206                 if (offset < ino_size)
2207                         ret = btrfs_truncate_page(inode, offset, len, 0);
2208                 mutex_unlock(&inode->i_mutex);
2209                 return ret;
2210         }
2211
2212         /* zero back part of the first page */
2213         if (offset < ino_size) {
2214                 ret = btrfs_truncate_page(inode, offset, 0, 0);
2215                 if (ret) {
2216                         mutex_unlock(&inode->i_mutex);
2217                         return ret;
2218                 }
2219         }
2220
2221         /* zero the front end of the last page */
2222         if (offset + len < ino_size) {
2223                 ret = btrfs_truncate_page(inode, offset + len, 0, 1);
2224                 if (ret) {
2225                         mutex_unlock(&inode->i_mutex);
2226                         return ret;
2227                 }
2228         }
2229
2230         if (lockend < lockstart) {
2231                 mutex_unlock(&inode->i_mutex);
2232                 return 0;
2233         }
2234
2235         while (1) {
2236                 struct btrfs_ordered_extent *ordered;
2237
2238                 truncate_pagecache_range(inode, lockstart, lockend);
2239
2240                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2241                                  0, &cached_state);
2242                 ordered = btrfs_lookup_first_ordered_extent(inode, lockend);
2243
2244                 /*
2245                  * We need to make sure we have no ordered extents in this range
2246                  * and nobody raced in and read a page in this range, if we did
2247                  * we need to try again.
2248                  */
2249                 if ((!ordered ||
2250                     (ordered->file_offset + ordered->len <= lockstart ||
2251                      ordered->file_offset > lockend)) &&
2252                      !test_range_bit(&BTRFS_I(inode)->io_tree, lockstart,
2253                                      lockend, EXTENT_UPTODATE, 0,
2254                                      cached_state)) {
2255                         if (ordered)
2256                                 btrfs_put_ordered_extent(ordered);
2257                         break;
2258                 }
2259                 if (ordered)
2260                         btrfs_put_ordered_extent(ordered);
2261                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2262                                      lockend, &cached_state, GFP_NOFS);
2263                 ret = btrfs_wait_ordered_range(inode, lockstart,
2264                                                lockend - lockstart + 1);
2265                 if (ret) {
2266                         mutex_unlock(&inode->i_mutex);
2267                         return ret;
2268                 }
2269         }
2270
2271         path = btrfs_alloc_path();
2272         if (!path) {
2273                 ret = -ENOMEM;
2274                 goto out;
2275         }
2276
2277         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
2278         if (!rsv) {
2279                 ret = -ENOMEM;
2280                 goto out_free;
2281         }
2282         rsv->size = btrfs_calc_trunc_metadata_size(root, 1);
2283         rsv->failfast = 1;
2284
2285         /*
2286          * 1 - update the inode
2287          * 1 - removing the extents in the range
2288          * 1 - adding the hole extent if no_holes isn't set
2289          */
2290         rsv_count = no_holes ? 2 : 3;
2291         trans = btrfs_start_transaction(root, rsv_count);
2292         if (IS_ERR(trans)) {
2293                 err = PTR_ERR(trans);
2294                 goto out_free;
2295         }
2296
2297         ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv, rsv,
2298                                       min_size);
2299         BUG_ON(ret);
2300         trans->block_rsv = rsv;
2301
2302         while (cur_offset < lockend) {
2303                 ret = __btrfs_drop_extents(trans, root, inode, path,
2304                                            cur_offset, lockend + 1,
2305                                            &drop_end, 1, 0, 0, NULL);
2306                 if (ret != -ENOSPC)
2307                         break;
2308
2309                 trans->block_rsv = &root->fs_info->trans_block_rsv;
2310
2311                 if (cur_offset < ino_size) {
2312                         ret = fill_holes(trans, inode, path, cur_offset,
2313                                          drop_end);
2314                         if (ret) {
2315                                 err = ret;
2316                                 break;
2317                         }
2318                 }
2319
2320                 cur_offset = drop_end;
2321
2322                 ret = btrfs_update_inode(trans, root, inode);
2323                 if (ret) {
2324                         err = ret;
2325                         break;
2326                 }
2327
2328                 btrfs_end_transaction(trans, root);
2329                 btrfs_btree_balance_dirty(root);
2330
2331                 trans = btrfs_start_transaction(root, rsv_count);
2332                 if (IS_ERR(trans)) {
2333                         ret = PTR_ERR(trans);
2334                         trans = NULL;
2335                         break;
2336                 }
2337
2338                 ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv,
2339                                               rsv, min_size);
2340                 BUG_ON(ret);    /* shouldn't happen */
2341                 trans->block_rsv = rsv;
2342         }
2343
2344         if (ret) {
2345                 err = ret;
2346                 goto out_trans;
2347         }
2348
2349         trans->block_rsv = &root->fs_info->trans_block_rsv;
2350         if (cur_offset < ino_size) {
2351                 ret = fill_holes(trans, inode, path, cur_offset, drop_end);
2352                 if (ret) {
2353                         err = ret;
2354                         goto out_trans;
2355                 }
2356         }
2357
2358 out_trans:
2359         if (!trans)
2360                 goto out_free;
2361
2362         inode_inc_iversion(inode);
2363         inode->i_mtime = inode->i_ctime = CURRENT_TIME;
2364
2365         trans->block_rsv = &root->fs_info->trans_block_rsv;
2366         ret = btrfs_update_inode(trans, root, inode);
2367         btrfs_end_transaction(trans, root);
2368         btrfs_btree_balance_dirty(root);
2369 out_free:
2370         btrfs_free_path(path);
2371         btrfs_free_block_rsv(root, rsv);
2372 out:
2373         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2374                              &cached_state, GFP_NOFS);
2375         mutex_unlock(&inode->i_mutex);
2376         if (ret && !err)
2377                 err = ret;
2378         return err;
2379 }
2380
2381 static long btrfs_fallocate(struct file *file, int mode,
2382                             loff_t offset, loff_t len)
2383 {
2384         struct inode *inode = file_inode(file);
2385         struct extent_state *cached_state = NULL;
2386         struct btrfs_root *root = BTRFS_I(inode)->root;
2387         u64 cur_offset;
2388         u64 last_byte;
2389         u64 alloc_start;
2390         u64 alloc_end;
2391         u64 alloc_hint = 0;
2392         u64 locked_end;
2393         struct extent_map *em;
2394         int blocksize = BTRFS_I(inode)->root->sectorsize;
2395         int ret;
2396
2397         alloc_start = round_down(offset, blocksize);
2398         alloc_end = round_up(offset + len, blocksize);
2399
2400         /* Make sure we aren't being give some crap mode */
2401         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2402                 return -EOPNOTSUPP;
2403
2404         if (mode & FALLOC_FL_PUNCH_HOLE)
2405                 return btrfs_punch_hole(inode, offset, len);
2406
2407         /*
2408          * Make sure we have enough space before we do the
2409          * allocation.
2410          */
2411         ret = btrfs_check_data_free_space(inode, alloc_end - alloc_start);
2412         if (ret)
2413                 return ret;
2414         if (root->fs_info->quota_enabled) {
2415                 ret = btrfs_qgroup_reserve(root, alloc_end - alloc_start);
2416                 if (ret)
2417                         goto out_reserve_fail;
2418         }
2419
2420         mutex_lock(&inode->i_mutex);
2421         ret = inode_newsize_ok(inode, alloc_end);
2422         if (ret)
2423                 goto out;
2424
2425         if (alloc_start > inode->i_size) {
2426                 ret = btrfs_cont_expand(inode, i_size_read(inode),
2427                                         alloc_start);
2428                 if (ret)
2429                         goto out;
2430         } else {
2431                 /*
2432                  * If we are fallocating from the end of the file onward we
2433                  * need to zero out the end of the page if i_size lands in the
2434                  * middle of a page.
2435                  */
2436                 ret = btrfs_truncate_page(inode, inode->i_size, 0, 0);
2437                 if (ret)
2438                         goto out;
2439         }
2440
2441         /*
2442          * wait for ordered IO before we have any locks.  We'll loop again
2443          * below with the locks held.
2444          */
2445         ret = btrfs_wait_ordered_range(inode, alloc_start,
2446                                        alloc_end - alloc_start);
2447         if (ret)
2448                 goto out;
2449
2450         locked_end = alloc_end - 1;
2451         while (1) {
2452                 struct btrfs_ordered_extent *ordered;
2453
2454                 /* the extent lock is ordered inside the running
2455                  * transaction
2456                  */
2457                 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
2458                                  locked_end, 0, &cached_state);
2459                 ordered = btrfs_lookup_first_ordered_extent(inode,
2460                                                             alloc_end - 1);
2461                 if (ordered &&
2462                     ordered->file_offset + ordered->len > alloc_start &&
2463                     ordered->file_offset < alloc_end) {
2464                         btrfs_put_ordered_extent(ordered);
2465                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
2466                                              alloc_start, locked_end,
2467                                              &cached_state, GFP_NOFS);
2468                         /*
2469                          * we can't wait on the range with the transaction
2470                          * running or with the extent lock held
2471                          */
2472                         ret = btrfs_wait_ordered_range(inode, alloc_start,
2473                                                        alloc_end - alloc_start);
2474                         if (ret)
2475                                 goto out;
2476                 } else {
2477                         if (ordered)
2478                                 btrfs_put_ordered_extent(ordered);
2479                         break;
2480                 }
2481         }
2482
2483         cur_offset = alloc_start;
2484         while (1) {
2485                 u64 actual_end;
2486
2487                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
2488                                       alloc_end - cur_offset, 0);
2489                 if (IS_ERR_OR_NULL(em)) {
2490                         if (!em)
2491                                 ret = -ENOMEM;
2492                         else
2493                                 ret = PTR_ERR(em);
2494                         break;
2495                 }
2496                 last_byte = min(extent_map_end(em), alloc_end);
2497                 actual_end = min_t(u64, extent_map_end(em), offset + len);
2498                 last_byte = ALIGN(last_byte, blocksize);
2499
2500                 if (em->block_start == EXTENT_MAP_HOLE ||
2501                     (cur_offset >= inode->i_size &&
2502                      !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
2503                         ret = btrfs_prealloc_file_range(inode, mode, cur_offset,
2504                                                         last_byte - cur_offset,
2505                                                         1 << inode->i_blkbits,
2506                                                         offset + len,
2507                                                         &alloc_hint);
2508
2509                         if (ret < 0) {
2510                                 free_extent_map(em);
2511                                 break;
2512                         }
2513                 } else if (actual_end > inode->i_size &&
2514                            !(mode & FALLOC_FL_KEEP_SIZE)) {
2515                         /*
2516                          * We didn't need to allocate any more space, but we
2517                          * still extended the size of the file so we need to
2518                          * update i_size.
2519                          */
2520                         inode->i_ctime = CURRENT_TIME;
2521                         i_size_write(inode, actual_end);
2522                         btrfs_ordered_update_i_size(inode, actual_end, NULL);
2523                 }
2524                 free_extent_map(em);
2525
2526                 cur_offset = last_byte;
2527                 if (cur_offset >= alloc_end) {
2528                         ret = 0;
2529                         break;
2530                 }
2531         }
2532         unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
2533                              &cached_state, GFP_NOFS);
2534 out:
2535         mutex_unlock(&inode->i_mutex);
2536         if (root->fs_info->quota_enabled)
2537                 btrfs_qgroup_free(root, alloc_end - alloc_start);
2538 out_reserve_fail:
2539         /* Let go of our reservation. */
2540         btrfs_free_reserved_data_space(inode, alloc_end - alloc_start);
2541         return ret;
2542 }
2543
2544 static int find_desired_extent(struct inode *inode, loff_t *offset, int whence)
2545 {
2546         struct btrfs_root *root = BTRFS_I(inode)->root;
2547         struct extent_map *em = NULL;
2548         struct extent_state *cached_state = NULL;
2549         u64 lockstart = *offset;
2550         u64 lockend = i_size_read(inode);
2551         u64 start = *offset;
2552         u64 len = i_size_read(inode);
2553         int ret = 0;
2554
2555         lockend = max_t(u64, root->sectorsize, lockend);
2556         if (lockend <= lockstart)
2557                 lockend = lockstart + root->sectorsize;
2558
2559         lockend--;
2560         len = lockend - lockstart + 1;
2561
2562         len = max_t(u64, len, root->sectorsize);
2563         if (inode->i_size == 0)
2564                 return -ENXIO;
2565
2566         lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend, 0,
2567                          &cached_state);
2568
2569         while (start < inode->i_size) {
2570                 em = btrfs_get_extent_fiemap(inode, NULL, 0, start, len, 0);
2571                 if (IS_ERR(em)) {
2572                         ret = PTR_ERR(em);
2573                         em = NULL;
2574                         break;
2575                 }
2576
2577                 if (whence == SEEK_HOLE &&
2578                     (em->block_start == EXTENT_MAP_HOLE ||
2579                      test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
2580                         break;
2581                 else if (whence == SEEK_DATA &&
2582                            (em->block_start != EXTENT_MAP_HOLE &&
2583                             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
2584                         break;
2585
2586                 start = em->start + em->len;
2587                 free_extent_map(em);
2588                 em = NULL;
2589                 cond_resched();
2590         }
2591         free_extent_map(em);
2592         if (!ret) {
2593                 if (whence == SEEK_DATA && start >= inode->i_size)
2594                         ret = -ENXIO;
2595                 else
2596                         *offset = min_t(loff_t, start, inode->i_size);
2597         }
2598         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2599                              &cached_state, GFP_NOFS);
2600         return ret;
2601 }
2602
2603 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
2604 {
2605         struct inode *inode = file->f_mapping->host;
2606         int ret;
2607
2608         mutex_lock(&inode->i_mutex);
2609         switch (whence) {
2610         case SEEK_END:
2611         case SEEK_CUR:
2612                 offset = generic_file_llseek(file, offset, whence);
2613                 goto out;
2614         case SEEK_DATA:
2615         case SEEK_HOLE:
2616                 if (offset >= i_size_read(inode)) {
2617                         mutex_unlock(&inode->i_mutex);
2618                         return -ENXIO;
2619                 }
2620
2621                 ret = find_desired_extent(inode, &offset, whence);
2622                 if (ret) {
2623                         mutex_unlock(&inode->i_mutex);
2624                         return ret;
2625                 }
2626         }
2627
2628         offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
2629 out:
2630         mutex_unlock(&inode->i_mutex);
2631         return offset;
2632 }
2633
2634 const struct file_operations btrfs_file_operations = {
2635         .llseek         = btrfs_file_llseek,
2636         .read           = do_sync_read,
2637         .write          = do_sync_write,
2638         .aio_read       = generic_file_aio_read,
2639         .splice_read    = generic_file_splice_read,
2640         .aio_write      = btrfs_file_aio_write,
2641         .mmap           = btrfs_file_mmap,
2642         .open           = generic_file_open,
2643         .release        = btrfs_release_file,
2644         .fsync          = btrfs_sync_file,
2645         .fallocate      = btrfs_fallocate,
2646         .unlocked_ioctl = btrfs_ioctl,
2647 #ifdef CONFIG_COMPAT
2648         .compat_ioctl   = btrfs_ioctl,
2649 #endif
2650 };
2651
2652 void btrfs_auto_defrag_exit(void)
2653 {
2654         if (btrfs_inode_defrag_cachep)
2655                 kmem_cache_destroy(btrfs_inode_defrag_cachep);
2656 }
2657
2658 int btrfs_auto_defrag_init(void)
2659 {
2660         btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
2661                                         sizeof(struct inode_defrag), 0,
2662                                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD,
2663                                         NULL);
2664         if (!btrfs_inode_defrag_cachep)
2665                 return -ENOMEM;
2666
2667         return 0;
2668 }