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;
 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         update_refs = (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
 728                        root == root->fs_info->tree_root);
 729         while (1) {
 730                 recow = 0;
 731                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
 732                                                search_start, modify_tree);
 733                 if (ret < 0)
 734                         break;
 735                 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
 736                         leaf = path->nodes[0];
 737                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
 738                         if (key.objectid == ino &&
 739                             key.type == BTRFS_EXTENT_DATA_KEY)
 740                                 path->slots[0]--;
 741                 }
 742                 ret = 0;
 743                 leafs_visited++;
 744 next_slot:
 745                 leaf = path->nodes[0];
 746                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 747                         BUG_ON(del_nr > 0);
 748                         ret = btrfs_next_leaf(root, path);
 749                         if (ret < 0)
 750                                 break;
 751                         if (ret > 0) {
 752                                 ret = 0;
 753                                 break;
 754                         }
 755                         leafs_visited++;
 756                         leaf = path->nodes[0];
 757                         recow = 1;
 758                 }
 759
 760                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 761                 if (key.objectid > ino ||
 762                     key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
 763                         break;
 764
 765                 fi = btrfs_item_ptr(leaf, path->slots[0],
 766                                     struct btrfs_file_extent_item);
 767                 extent_type = btrfs_file_extent_type(leaf, fi);
 768
 769                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
 770                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
 771                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 772                         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 773                         extent_offset = btrfs_file_extent_offset(leaf, fi);
 774                         extent_end = key.offset +
 775                                 btrfs_file_extent_num_bytes(leaf, fi);
 776                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 777                         extent_end = key.offset +
 778                                 btrfs_file_extent_inline_len(leaf,
 779                                                      path->slots[0], fi);
 780                 } else {
 781                         WARN_ON(1);
 782                         extent_end = search_start;
 783                 }
 784
 785                 /*
 786                  * Don't skip extent items representing 0 byte lengths. They
 787                  * used to be created (bug) if while punching holes we hit
 788                  * -ENOSPC condition. So if we find one here, just ensure we
 789                  * delete it, otherwise we would insert a new file extent item
 790                  * with the same key (offset) as that 0 bytes length file
 791                  * extent item in the call to setup_items_for_insert() later
 792                  * in this function.
 793                  */
 794                 if (extent_end == key.offset && extent_end >= search_start)
 795                         goto delete_extent_item;
 796
 797                 if (extent_end <= search_start) {
 798                         path->slots[0]++;
 799                         goto next_slot;
 800                 }
 801
 802                 found = 1;
 803                 search_start = max(key.offset, start);
 804                 if (recow || !modify_tree) {
 805                         modify_tree = -1;
 806                         btrfs_release_path(path);
 807                         continue;
 808                 }
 809
 810                 /*
 811                  *     | - range to drop - |
 812                  *  | -------- extent -------- |
 813                  */
 814                 if (start > key.offset && end < extent_end) {
 815                         BUG_ON(del_nr > 0);
 816                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 817                                 ret = -EOPNOTSUPP;
 818                                 break;
 819                         }
 820
 821                         memcpy(&new_key, &key, sizeof(new_key));
 822                         new_key.offset = start;
 823                         ret = btrfs_duplicate_item(trans, root, path,
 824                                                    &new_key);
 825                         if (ret == -EAGAIN) {
 826                                 btrfs_release_path(path);
 827                                 continue;
 828                         }
 829                         if (ret < 0)
 830                                 break;
 831
 832                         leaf = path->nodes[0];
 833                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 834                                             struct btrfs_file_extent_item);
 835                         btrfs_set_file_extent_num_bytes(leaf, fi,
 836                                                         start - key.offset);
 837
 838                         fi = btrfs_item_ptr(leaf, path->slots[0],
 839                                             struct btrfs_file_extent_item);
 840
 841                         extent_offset += start - key.offset;
 842                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 843                         btrfs_set_file_extent_num_bytes(leaf, fi,
 844                                                         extent_end - start);
 845                         btrfs_mark_buffer_dirty(leaf);
 846
 847                         if (update_refs && disk_bytenr > 0) {
 848                                 ret = btrfs_inc_extent_ref(trans, root,
 849                                                 disk_bytenr, num_bytes, 0,
 850                                                 root->root_key.objectid,
 851                                                 new_key.objectid,
 852                                                 start - extent_offset, 0);
 853                                 BUG_ON(ret); /* -ENOMEM */
 854                         }
 855                         key.offset = start;
 856                 }
 857                 /*
 858                  *  | ---- range to drop ----- |
 859                  *      | -------- extent -------- |
 860                  */
 861                 if (start <= key.offset && end < extent_end) {
 862                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 863                                 ret = -EOPNOTSUPP;
 864                                 break;
 865                         }
 866
 867                         memcpy(&new_key, &key, sizeof(new_key));
 868                         new_key.offset = end;
 869                         btrfs_set_item_key_safe(root, path, &new_key);
 870
 871                         extent_offset += end - key.offset;
 872                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 873                         btrfs_set_file_extent_num_bytes(leaf, fi,
 874                                                         extent_end - end);
 875                         btrfs_mark_buffer_dirty(leaf);
 876                         if (update_refs && disk_bytenr > 0)
 877                                 inode_sub_bytes(inode, end - key.offset);
 878                         break;
 879                 }
 880
 881                 search_start = extent_end;
 882                 /*
 883                  *       | ---- range to drop ----- |
 884                  *  | -------- extent -------- |
 885                  */
 886                 if (start > key.offset && end >= extent_end) {
 887                         BUG_ON(del_nr > 0);
 888                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 889                                 ret = -EOPNOTSUPP;
 890                                 break;
 891                         }
 892
 893                         btrfs_set_file_extent_num_bytes(leaf, fi,
 894                                                         start - key.offset);
 895                         btrfs_mark_buffer_dirty(leaf);
 896                         if (update_refs && disk_bytenr > 0)
 897                                 inode_sub_bytes(inode, extent_end - start);
 898                         if (end == extent_end)
 899                                 break;
 900
 901                         path->slots[0]++;
 902                         goto next_slot;
 903                 }
 904
 905                 /*
 906                  *  | ---- range to drop ----- |
 907                  *    | ------ extent ------ |
 908                  */
 909                 if (start <= key.offset && end >= extent_end) {
 910 delete_extent_item:
 911                         if (del_nr == 0) {
 912                                 del_slot = path->slots[0];
 913                                 del_nr = 1;
 914                         } else {
 915                                 BUG_ON(del_slot + del_nr != path->slots[0]);
 916                                 del_nr++;
 917                         }
 918
 919                         if (update_refs &&
 920                             extent_type == BTRFS_FILE_EXTENT_INLINE) {
 921                                 inode_sub_bytes(inode,
 922                                                 extent_end - key.offset);
 923                                 extent_end = ALIGN(extent_end,
 924                                                    root->sectorsize);
 925                         } else if (update_refs && disk_bytenr > 0) {
 926                                 ret = btrfs_free_extent(trans, root,
 927                                                 disk_bytenr, num_bytes, 0,
 928                                                 root->root_key.objectid,
 929                                                 key.objectid, key.offset -
 930                                                 extent_offset, 0);
 931                                 BUG_ON(ret); /* -ENOMEM */
 932                                 inode_sub_bytes(inode,
 933                                                 extent_end - key.offset);
 934                         }
 935
 936                         if (end == extent_end)
 937                                 break;
 938
 939                         if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
 940                                 path->slots[0]++;
 941                                 goto next_slot;
 942                         }
 943
 944                         ret = btrfs_del_items(trans, root, path, del_slot,
 945                                               del_nr);
 946                         if (ret) {
 947                                 btrfs_abort_transaction(trans, root, ret);
 948                                 break;
 949                         }
 950
 951                         del_nr = 0;
 952                         del_slot = 0;
 953
 954                         btrfs_release_path(path);
 955                         continue;
 956                 }
 957
 958                 BUG_ON(1);
 959         }
 960
 961         if (!ret && del_nr > 0) {
 962                 /*
 963                  * Set path->slots[0] to first slot, so that after the delete
 964                  * if items are move off from our leaf to its immediate left or
 965                  * right neighbor leafs, we end up with a correct and adjusted
 966                  * path->slots[0] for our insertion (if replace_extent != 0).
 967                  */
 968                 path->slots[0] = del_slot;
 969                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
 970                 if (ret)
 971                         btrfs_abort_transaction(trans, root, ret);
 972         }
 973
 974         leaf = path->nodes[0];
 975         /*
 976          * If btrfs_del_items() was called, it might have deleted a leaf, in
 977          * which case it unlocked our path, so check path->locks[0] matches a
 978          * write lock.
 979          */
 980         if (!ret && replace_extent && leafs_visited == 1 &&
 981             (path->locks[0] == BTRFS_WRITE_LOCK_BLOCKING ||
 982              path->locks[0] == BTRFS_WRITE_LOCK) &&
 983             btrfs_leaf_free_space(root, leaf) >=
 984             sizeof(struct btrfs_item) + extent_item_size) {
 985
 986                 key.objectid = ino;
 987                 key.type = BTRFS_EXTENT_DATA_KEY;
 988                 key.offset = start;
 989                 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
 990                         struct btrfs_key slot_key;
 991
 992                         btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
 993                         if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
 994                                 path->slots[0]++;
 995                 }
 996                 setup_items_for_insert(root, path, &key,
 997                                        &extent_item_size,
 998                                        extent_item_size,
 999                                        sizeof(struct btrfs_item) +
1000                                        extent_item_size, 1);
1001                 *key_inserted = 1;
1002         }
1003
1004         if (!replace_extent || !(*key_inserted))
1005                 btrfs_release_path(path);
1006         if (drop_end)
1007                 *drop_end = found ? min(end, extent_end) : end;
1008         return ret;
1009 }
1010
1011 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
1012                        struct btrfs_root *root, struct inode *inode, u64 start,
1013                        u64 end, int drop_cache)
1014 {
1015         struct btrfs_path *path;
1016         int ret;
1017
1018         path = btrfs_alloc_path();
1019         if (!path)
1020                 return -ENOMEM;
1021         ret = __btrfs_drop_extents(trans, root, inode, path, start, end, NULL,
1022                                    drop_cache, 0, 0, NULL);
1023         btrfs_free_path(path);
1024         return ret;
1025 }
1026
1027 static int extent_mergeable(struct extent_buffer *leaf, int slot,
1028                             u64 objectid, u64 bytenr, u64 orig_offset,
1029                             u64 *start, u64 *end)
1030 {
1031         struct btrfs_file_extent_item *fi;
1032         struct btrfs_key key;
1033         u64 extent_end;
1034
1035         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1036                 return 0;
1037
1038         btrfs_item_key_to_cpu(leaf, &key, slot);
1039         if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1040                 return 0;
1041
1042         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1043         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1044             btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1045             btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1046             btrfs_file_extent_compression(leaf, fi) ||
1047             btrfs_file_extent_encryption(leaf, fi) ||
1048             btrfs_file_extent_other_encoding(leaf, fi))
1049                 return 0;
1050
1051         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1052         if ((*start && *start != key.offset) || (*end && *end != extent_end))
1053                 return 0;
1054
1055         *start = key.offset;
1056         *end = extent_end;
1057         return 1;
1058 }
1059
1060 /*
1061  * Mark extent in the range start - end as written.
1062  *
1063  * This changes extent type from 'pre-allocated' to 'regular'. If only
1064  * part of extent is marked as written, the extent will be split into
1065  * two or three.
1066  */
1067 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1068                               struct inode *inode, u64 start, u64 end)
1069 {
1070         struct btrfs_root *root = BTRFS_I(inode)->root;
1071         struct extent_buffer *leaf;
1072         struct btrfs_path *path;
1073         struct btrfs_file_extent_item *fi;
1074         struct btrfs_key key;
1075         struct btrfs_key new_key;
1076         u64 bytenr;
1077         u64 num_bytes;
1078         u64 extent_end;
1079         u64 orig_offset;
1080         u64 other_start;
1081         u64 other_end;
1082         u64 split;
1083         int del_nr = 0;
1084         int del_slot = 0;
1085         int recow;
1086         int ret;
1087         u64 ino = btrfs_ino(inode);
1088
1089         path = btrfs_alloc_path();
1090         if (!path)
1091                 return -ENOMEM;
1092 again:
1093         recow = 0;
1094         split = start;
1095         key.objectid = ino;
1096         key.type = BTRFS_EXTENT_DATA_KEY;
1097         key.offset = split;
1098
1099         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1100         if (ret < 0)
1101                 goto out;
1102         if (ret > 0 && path->slots[0] > 0)
1103                 path->slots[0]--;
1104
1105         leaf = path->nodes[0];
1106         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1107         BUG_ON(key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY);
1108         fi = btrfs_item_ptr(leaf, path->slots[0],
1109                             struct btrfs_file_extent_item);
1110         BUG_ON(btrfs_file_extent_type(leaf, fi) !=
1111                BTRFS_FILE_EXTENT_PREALLOC);
1112         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1113         BUG_ON(key.offset > start || extent_end < end);
1114
1115         bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1116         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1117         orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1118         memcpy(&new_key, &key, sizeof(new_key));
1119
1120         if (start == key.offset && end < extent_end) {
1121                 other_start = 0;
1122                 other_end = start;
1123                 if (extent_mergeable(leaf, path->slots[0] - 1,
1124                                      ino, bytenr, orig_offset,
1125                                      &other_start, &other_end)) {
1126                         new_key.offset = end;
1127                         btrfs_set_item_key_safe(root, path, &new_key);
1128                         fi = btrfs_item_ptr(leaf, path->slots[0],
1129                                             struct btrfs_file_extent_item);
1130                         btrfs_set_file_extent_generation(leaf, fi,
1131                                                          trans->transid);
1132                         btrfs_set_file_extent_num_bytes(leaf, fi,
1133                                                         extent_end - end);
1134                         btrfs_set_file_extent_offset(leaf, fi,
1135                                                      end - orig_offset);
1136                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1137                                             struct btrfs_file_extent_item);
1138                         btrfs_set_file_extent_generation(leaf, fi,
1139                                                          trans->transid);
1140                         btrfs_set_file_extent_num_bytes(leaf, fi,
1141                                                         end - other_start);
1142                         btrfs_mark_buffer_dirty(leaf);
1143                         goto out;
1144                 }
1145         }
1146
1147         if (start > key.offset && end == extent_end) {
1148                 other_start = end;
1149                 other_end = 0;
1150                 if (extent_mergeable(leaf, path->slots[0] + 1,
1151                                      ino, bytenr, orig_offset,
1152                                      &other_start, &other_end)) {
1153                         fi = btrfs_item_ptr(leaf, path->slots[0],
1154                                             struct btrfs_file_extent_item);
1155                         btrfs_set_file_extent_num_bytes(leaf, fi,
1156                                                         start - key.offset);
1157                         btrfs_set_file_extent_generation(leaf, fi,
1158                                                          trans->transid);
1159                         path->slots[0]++;
1160                         new_key.offset = start;
1161                         btrfs_set_item_key_safe(root, path, &new_key);
1162
1163                         fi = btrfs_item_ptr(leaf, path->slots[0],
1164                                             struct btrfs_file_extent_item);
1165                         btrfs_set_file_extent_generation(leaf, fi,
1166                                                          trans->transid);
1167                         btrfs_set_file_extent_num_bytes(leaf, fi,
1168                                                         other_end - start);
1169                         btrfs_set_file_extent_offset(leaf, fi,
1170                                                      start - orig_offset);
1171                         btrfs_mark_buffer_dirty(leaf);
1172                         goto out;
1173                 }
1174         }
1175
1176         while (start > key.offset || end < extent_end) {
1177                 if (key.offset == start)
1178                         split = end;
1179
1180                 new_key.offset = split;
1181                 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1182                 if (ret == -EAGAIN) {
1183                         btrfs_release_path(path);
1184                         goto again;
1185                 }
1186                 if (ret < 0) {
1187                         btrfs_abort_transaction(trans, root, ret);
1188                         goto out;
1189                 }
1190
1191                 leaf = path->nodes[0];
1192                 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1193                                     struct btrfs_file_extent_item);
1194                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1195                 btrfs_set_file_extent_num_bytes(leaf, fi,
1196                                                 split - key.offset);
1197
1198                 fi = btrfs_item_ptr(leaf, path->slots[0],
1199                                     struct btrfs_file_extent_item);
1200
1201                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1202                 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1203                 btrfs_set_file_extent_num_bytes(leaf, fi,
1204                                                 extent_end - split);
1205                 btrfs_mark_buffer_dirty(leaf);
1206
1207                 ret = btrfs_inc_extent_ref(trans, root, bytenr, num_bytes, 0,
1208                                            root->root_key.objectid,
1209                                            ino, orig_offset, 0);
1210                 BUG_ON(ret); /* -ENOMEM */
1211
1212                 if (split == start) {
1213                         key.offset = start;
1214                 } else {
1215                         BUG_ON(start != key.offset);
1216                         path->slots[0]--;
1217                         extent_end = end;
1218                 }
1219                 recow = 1;
1220         }
1221
1222         other_start = end;
1223         other_end = 0;
1224         if (extent_mergeable(leaf, path->slots[0] + 1,
1225                              ino, bytenr, orig_offset,
1226                              &other_start, &other_end)) {
1227                 if (recow) {
1228                         btrfs_release_path(path);
1229                         goto again;
1230                 }
1231                 extent_end = other_end;
1232                 del_slot = path->slots[0] + 1;
1233                 del_nr++;
1234                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1235                                         0, root->root_key.objectid,
1236                                         ino, orig_offset, 0);
1237                 BUG_ON(ret); /* -ENOMEM */
1238         }
1239         other_start = 0;
1240         other_end = start;
1241         if (extent_mergeable(leaf, path->slots[0] - 1,
1242                              ino, bytenr, orig_offset,
1243                              &other_start, &other_end)) {
1244                 if (recow) {
1245                         btrfs_release_path(path);
1246                         goto again;
1247                 }
1248                 key.offset = other_start;
1249                 del_slot = path->slots[0];
1250                 del_nr++;
1251                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1252                                         0, root->root_key.objectid,
1253                                         ino, orig_offset, 0);
1254                 BUG_ON(ret); /* -ENOMEM */
1255         }
1256         if (del_nr == 0) {
1257                 fi = btrfs_item_ptr(leaf, path->slots[0],
1258                            struct btrfs_file_extent_item);
1259                 btrfs_set_file_extent_type(leaf, fi,
1260                                            BTRFS_FILE_EXTENT_REG);
1261                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1262                 btrfs_mark_buffer_dirty(leaf);
1263         } else {
1264                 fi = btrfs_item_ptr(leaf, del_slot - 1,
1265                            struct btrfs_file_extent_item);
1266                 btrfs_set_file_extent_type(leaf, fi,
1267                                            BTRFS_FILE_EXTENT_REG);
1268                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1269                 btrfs_set_file_extent_num_bytes(leaf, fi,
1270                                                 extent_end - key.offset);
1271                 btrfs_mark_buffer_dirty(leaf);
1272
1273                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1274                 if (ret < 0) {
1275                         btrfs_abort_transaction(trans, root, ret);
1276                         goto out;
1277                 }
1278         }
1279 out:
1280         btrfs_free_path(path);
1281         return 0;
1282 }
1283
1284 /*
1285  * on error we return an unlocked page and the error value
1286  * on success we return a locked page and 0
1287  */
1288 static int prepare_uptodate_page(struct page *page, u64 pos,
1289                                  bool force_uptodate)
1290 {
1291         int ret = 0;
1292
1293         if (((pos & (PAGE_CACHE_SIZE - 1)) || force_uptodate) &&
1294             !PageUptodate(page)) {
1295                 ret = btrfs_readpage(NULL, page);
1296                 if (ret)
1297                         return ret;
1298                 lock_page(page);
1299                 if (!PageUptodate(page)) {
1300                         unlock_page(page);
1301                         return -EIO;
1302                 }
1303         }
1304         return 0;
1305 }
1306
1307 /*
1308  * this just gets pages into the page cache and locks them down.
1309  */
1310 static noinline int prepare_pages(struct inode *inode, struct page **pages,
1311                                   size_t num_pages, loff_t pos,
1312                                   size_t write_bytes, bool force_uptodate)
1313 {
1314         int i;
1315         unsigned long index = pos >> PAGE_CACHE_SHIFT;
1316         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1317         int err = 0;
1318         int faili;
1319
1320         for (i = 0; i < num_pages; i++) {
1321                 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1322                                                mask | __GFP_WRITE);
1323                 if (!pages[i]) {
1324                         faili = i - 1;
1325                         err = -ENOMEM;
1326                         goto fail;
1327                 }
1328
1329                 if (i == 0)
1330                         err = prepare_uptodate_page(pages[i], pos,
1331                                                     force_uptodate);
1332                 if (i == num_pages - 1)
1333                         err = prepare_uptodate_page(pages[i],
1334                                                     pos + write_bytes, false);
1335                 if (err) {
1336                         page_cache_release(pages[i]);
1337                         faili = i - 1;
1338                         goto fail;
1339                 }
1340                 wait_on_page_writeback(pages[i]);
1341         }
1342
1343         return 0;
1344 fail:
1345         while (faili >= 0) {
1346                 unlock_page(pages[faili]);
1347                 page_cache_release(pages[faili]);
1348                 faili--;
1349         }
1350         return err;
1351
1352 }
1353
1354 /*
1355  * This function locks the extent and properly waits for data=ordered extents
1356  * to finish before allowing the pages to be modified if need.
1357  *
1358  * The return value:
1359  * 1 - the extent is locked
1360  * 0 - the extent is not locked, and everything is OK
1361  * -EAGAIN - need re-prepare the pages
1362  * the other < 0 number - Something wrong happens
1363  */
1364 static noinline int
1365 lock_and_cleanup_extent_if_need(struct inode *inode, struct page **pages,
1366                                 size_t num_pages, loff_t pos,
1367                                 u64 *lockstart, u64 *lockend,
1368                                 struct extent_state **cached_state)
1369 {
1370         u64 start_pos;
1371         u64 last_pos;
1372         int i;
1373         int ret = 0;
1374
1375         start_pos = pos & ~((u64)PAGE_CACHE_SIZE - 1);
1376         last_pos = start_pos + ((u64)num_pages << PAGE_CACHE_SHIFT) - 1;
1377
1378         if (start_pos < inode->i_size) {
1379                 struct btrfs_ordered_extent *ordered;
1380                 lock_extent_bits(&BTRFS_I(inode)->io_tree,
1381                                  start_pos, last_pos, 0, cached_state);
1382                 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1383                                                      last_pos - start_pos + 1);
1384                 if (ordered &&
1385                     ordered->file_offset + ordered->len > start_pos &&
1386                     ordered->file_offset <= last_pos) {
1387                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1388                                              start_pos, last_pos,
1389                                              cached_state, GFP_NOFS);
1390                         for (i = 0; i < num_pages; i++) {
1391                                 unlock_page(pages[i]);
1392                                 page_cache_release(pages[i]);
1393                         }
1394                         btrfs_start_ordered_extent(inode, ordered, 1);
1395                         btrfs_put_ordered_extent(ordered);
1396                         return -EAGAIN;
1397                 }
1398                 if (ordered)
1399                         btrfs_put_ordered_extent(ordered);
1400
1401                 clear_extent_bit(&BTRFS_I(inode)->io_tree, start_pos,
1402                                   last_pos, EXTENT_DIRTY | EXTENT_DELALLOC |
1403                                   EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
1404                                   0, 0, cached_state, GFP_NOFS);
1405                 *lockstart = start_pos;
1406                 *lockend = last_pos;
1407                 ret = 1;
1408         }
1409
1410         for (i = 0; i < num_pages; i++) {
1411                 if (clear_page_dirty_for_io(pages[i]))
1412                         account_page_redirty(pages[i]);
1413                 set_page_extent_mapped(pages[i]);
1414                 WARN_ON(!PageLocked(pages[i]));
1415         }
1416
1417         return ret;
1418 }
1419
1420 static noinline int check_can_nocow(struct inode *inode, loff_t pos,
1421                                     size_t *write_bytes)
1422 {
1423         struct btrfs_root *root = BTRFS_I(inode)->root;
1424         struct btrfs_ordered_extent *ordered;
1425         u64 lockstart, lockend;
1426         u64 num_bytes;
1427         int ret;
1428
1429         ret = btrfs_start_nocow_write(root);
1430         if (!ret)
1431                 return -ENOSPC;
1432
1433         lockstart = round_down(pos, root->sectorsize);
1434         lockend = round_up(pos + *write_bytes, root->sectorsize) - 1;
1435
1436         while (1) {
1437                 lock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend);
1438                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
1439                                                      lockend - lockstart + 1);
1440                 if (!ordered) {
1441                         break;
1442                 }
1443                 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend);
1444                 btrfs_start_ordered_extent(inode, ordered, 1);
1445                 btrfs_put_ordered_extent(ordered);
1446         }
1447
1448         num_bytes = lockend - lockstart + 1;
1449         ret = can_nocow_extent(inode, lockstart, &num_bytes, NULL, NULL, NULL);
1450         if (ret <= 0) {
1451                 ret = 0;
1452                 btrfs_end_nocow_write(root);
1453         } else {
1454                 *write_bytes = min_t(size_t, *write_bytes ,
1455                                      num_bytes - pos + lockstart);
1456         }
1457
1458         unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend);
1459
1460         return ret;
1461 }
1462
1463 static noinline ssize_t __btrfs_buffered_write(struct file *file,
1464                                                struct iov_iter *i,
1465                                                loff_t pos)
1466 {
1467         struct inode *inode = file_inode(file);
1468         struct btrfs_root *root = BTRFS_I(inode)->root;
1469         struct page **pages = NULL;
1470         struct extent_state *cached_state = NULL;
1471         u64 release_bytes = 0;
1472         u64 lockstart;
1473         u64 lockend;
1474         unsigned long first_index;
1475         size_t num_written = 0;
1476         int nrptrs;
1477         int ret = 0;
1478         bool only_release_metadata = false;
1479         bool force_page_uptodate = false;
1480         bool need_unlock;
1481
1482         nrptrs = min((iov_iter_count(i) + PAGE_CACHE_SIZE - 1) /
1483                      PAGE_CACHE_SIZE, PAGE_CACHE_SIZE /
1484                      (sizeof(struct page *)));
1485         nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1486         nrptrs = max(nrptrs, 8);
1487         pages = kmalloc(nrptrs * sizeof(struct page *), GFP_KERNEL);
1488         if (!pages)
1489                 return -ENOMEM;
1490
1491         first_index = pos >> PAGE_CACHE_SHIFT;
1492
1493         while (iov_iter_count(i) > 0) {
1494                 size_t offset = pos & (PAGE_CACHE_SIZE - 1);
1495                 size_t write_bytes = min(iov_iter_count(i),
1496                                          nrptrs * (size_t)PAGE_CACHE_SIZE -
1497                                          offset);
1498                 size_t num_pages = (write_bytes + offset +
1499                                     PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1500                 size_t reserve_bytes;
1501                 size_t dirty_pages;
1502                 size_t copied;
1503
1504                 WARN_ON(num_pages > nrptrs);
1505
1506                 /*
1507                  * Fault pages before locking them in prepare_pages
1508                  * to avoid recursive lock
1509                  */
1510                 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1511                         ret = -EFAULT;
1512                         break;
1513                 }
1514
1515                 reserve_bytes = num_pages << PAGE_CACHE_SHIFT;
1516                 ret = btrfs_check_data_free_space(inode, reserve_bytes);
1517                 if (ret == -ENOSPC &&
1518                     (BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1519                                               BTRFS_INODE_PREALLOC))) {
1520                         ret = check_can_nocow(inode, pos, &write_bytes);
1521                         if (ret > 0) {
1522                                 only_release_metadata = true;
1523                                 /*
1524                                  * our prealloc extent may be smaller than
1525                                  * write_bytes, so scale down.
1526                                  */
1527                                 num_pages = (write_bytes + offset +
1528                                              PAGE_CACHE_SIZE - 1) >>
1529                                         PAGE_CACHE_SHIFT;
1530                                 reserve_bytes = num_pages << PAGE_CACHE_SHIFT;
1531                                 ret = 0;
1532                         } else {
1533                                 ret = -ENOSPC;
1534                         }
1535                 }
1536
1537                 if (ret)
1538                         break;
1539
1540                 ret = btrfs_delalloc_reserve_metadata(inode, reserve_bytes);
1541                 if (ret) {
1542                         if (!only_release_metadata)
1543                                 btrfs_free_reserved_data_space(inode,
1544                                                                reserve_bytes);
1545                         else
1546                                 btrfs_end_nocow_write(root);
1547                         break;
1548                 }
1549
1550                 release_bytes = reserve_bytes;
1551                 need_unlock = false;
1552 again:
1553                 /*
1554                  * This is going to setup the pages array with the number of
1555                  * pages we want, so we don't really need to worry about the
1556                  * contents of pages from loop to loop
1557                  */
1558                 ret = prepare_pages(inode, pages, num_pages,
1559                                     pos, write_bytes,
1560                                     force_page_uptodate);
1561                 if (ret)
1562                         break;
1563
1564                 ret = lock_and_cleanup_extent_if_need(inode, pages, num_pages,
1565                                                       pos, &lockstart, &lockend,
1566                                                       &cached_state);
1567                 if (ret < 0) {
1568                         if (ret == -EAGAIN)
1569                                 goto again;
1570                         break;
1571                 } else if (ret > 0) {
1572                         need_unlock = true;
1573                         ret = 0;
1574                 }
1575
1576                 copied = btrfs_copy_from_user(pos, num_pages,
1577                                            write_bytes, pages, i);
1578
1579                 /*
1580                  * if we have trouble faulting in the pages, fall
1581                  * back to one page at a time
1582                  */
1583                 if (copied < write_bytes)
1584                         nrptrs = 1;
1585
1586                 if (copied == 0) {
1587                         force_page_uptodate = true;
1588                         dirty_pages = 0;
1589                 } else {
1590                         force_page_uptodate = false;
1591                         dirty_pages = (copied + offset +
1592                                        PAGE_CACHE_SIZE - 1) >>
1593                                        PAGE_CACHE_SHIFT;
1594                 }
1595
1596                 /*
1597                  * If we had a short copy we need to release the excess delaloc
1598                  * bytes we reserved.  We need to increment outstanding_extents
1599                  * because btrfs_delalloc_release_space will decrement it, but
1600                  * we still have an outstanding extent for the chunk we actually
1601                  * managed to copy.
1602                  */
1603                 if (num_pages > dirty_pages) {
1604                         release_bytes = (num_pages - dirty_pages) <<
1605                                 PAGE_CACHE_SHIFT;
1606                         if (copied > 0) {
1607                                 spin_lock(&BTRFS_I(inode)->lock);
1608                                 BTRFS_I(inode)->outstanding_extents++;
1609                                 spin_unlock(&BTRFS_I(inode)->lock);
1610                         }
1611                         if (only_release_metadata)
1612                                 btrfs_delalloc_release_metadata(inode,
1613                                                                 release_bytes);
1614                         else
1615                                 btrfs_delalloc_release_space(inode,
1616                                                              release_bytes);
1617                 }
1618
1619                 release_bytes = dirty_pages << PAGE_CACHE_SHIFT;
1620
1621                 if (copied > 0)
1622                         ret = btrfs_dirty_pages(root, inode, pages,
1623                                                 dirty_pages, pos, copied,
1624                                                 NULL);
1625                 if (need_unlock)
1626                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1627                                              lockstart, lockend, &cached_state,
1628                                              GFP_NOFS);
1629                 if (ret) {
1630                         btrfs_drop_pages(pages, num_pages);
1631                         break;
1632                 }
1633
1634                 release_bytes = 0;
1635                 if (only_release_metadata)
1636                         btrfs_end_nocow_write(root);
1637
1638                 if (only_release_metadata && copied > 0) {
1639                         u64 lockstart = round_down(pos, root->sectorsize);
1640                         u64 lockend = lockstart +
1641                                 (dirty_pages << PAGE_CACHE_SHIFT) - 1;
1642
1643                         set_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
1644                                        lockend, EXTENT_NORESERVE, NULL,
1645                                        NULL, GFP_NOFS);
1646                         only_release_metadata = false;
1647                 }
1648
1649                 btrfs_drop_pages(pages, num_pages);
1650
1651                 cond_resched();
1652
1653                 balance_dirty_pages_ratelimited(inode->i_mapping);
1654                 if (dirty_pages < (root->leafsize >> PAGE_CACHE_SHIFT) + 1)
1655                         btrfs_btree_balance_dirty(root);
1656
1657                 pos += copied;
1658                 num_written += copied;
1659         }
1660
1661         kfree(pages);
1662
1663         if (release_bytes) {
1664                 if (only_release_metadata) {
1665                         btrfs_end_nocow_write(root);
1666                         btrfs_delalloc_release_metadata(inode, release_bytes);
1667                 } else {
1668                         btrfs_delalloc_release_space(inode, release_bytes);
1669                 }
1670         }
1671
1672         return num_written ? num_written : ret;
1673 }
1674
1675 static ssize_t __btrfs_direct_write(struct kiocb *iocb,
1676                                     const struct iovec *iov,
1677                                     unsigned long nr_segs, loff_t pos,
1678                                     size_t count, size_t ocount)
1679 {
1680         struct file *file = iocb->ki_filp;
1681         struct iov_iter i;
1682         ssize_t written;
1683         ssize_t written_buffered;
1684         loff_t endbyte;
1685         int err;
1686
1687         written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
1688                                             count, ocount);
1689
1690         if (written < 0 || written == count)
1691                 return written;
1692
1693         pos += written;
1694         count -= written;
1695         iov_iter_init(&i, iov, nr_segs, count, written);
1696         written_buffered = __btrfs_buffered_write(file, &i, pos);
1697         if (written_buffered < 0) {
1698                 err = written_buffered;
1699                 goto out;
1700         }
1701         endbyte = pos + written_buffered - 1;
1702         err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
1703         if (err)
1704                 goto out;
1705         written += written_buffered;
1706         iocb->ki_pos = pos + written_buffered;
1707         invalidate_mapping_pages(file->f_mapping, pos >> PAGE_CACHE_SHIFT,
1708                                  endbyte >> PAGE_CACHE_SHIFT);
1709 out:
1710         return written ? written : err;
1711 }
1712
1713 static void update_time_for_write(struct inode *inode)
1714 {
1715         struct timespec now;
1716
1717         if (IS_NOCMTIME(inode))
1718                 return;
1719
1720         now = current_fs_time(inode->i_sb);
1721         if (!timespec_equal(&inode->i_mtime, &now))
1722                 inode->i_mtime = now;
1723
1724         if (!timespec_equal(&inode->i_ctime, &now))
1725                 inode->i_ctime = now;
1726
1727         if (IS_I_VERSION(inode))
1728                 inode_inc_iversion(inode);
1729 }
1730
1731 static ssize_t btrfs_file_aio_write(struct kiocb *iocb,
1732                                     const struct iovec *iov,
1733                                     unsigned long nr_segs, loff_t pos)
1734 {
1735         struct file *file = iocb->ki_filp;
1736         struct inode *inode = file_inode(file);
1737         struct btrfs_root *root = BTRFS_I(inode)->root;
1738         u64 start_pos;
1739         u64 end_pos;
1740         ssize_t num_written = 0;
1741         ssize_t err = 0;
1742         size_t count, ocount;
1743         bool sync = (file->f_flags & O_DSYNC) || IS_SYNC(file->f_mapping->host);
1744
1745         mutex_lock(&inode->i_mutex);
1746
1747         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
1748         if (err) {
1749                 mutex_unlock(&inode->i_mutex);
1750                 goto out;
1751         }
1752         count = ocount;
1753
1754         current->backing_dev_info = inode->i_mapping->backing_dev_info;
1755         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
1756         if (err) {
1757                 mutex_unlock(&inode->i_mutex);
1758                 goto out;
1759         }
1760
1761         if (count == 0) {
1762                 mutex_unlock(&inode->i_mutex);
1763                 goto out;
1764         }
1765
1766         err = file_remove_suid(file);
1767         if (err) {
1768                 mutex_unlock(&inode->i_mutex);
1769                 goto out;
1770         }
1771
1772         /*
1773          * If BTRFS flips readonly due to some impossible error
1774          * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1775          * although we have opened a file as writable, we have
1776          * to stop this write operation to ensure FS consistency.
1777          */
1778         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state)) {
1779                 mutex_unlock(&inode->i_mutex);
1780                 err = -EROFS;
1781                 goto out;
1782         }
1783
1784         /*
1785          * We reserve space for updating the inode when we reserve space for the
1786          * extent we are going to write, so we will enospc out there.  We don't
1787          * need to start yet another transaction to update the inode as we will
1788          * update the inode when we finish writing whatever data we write.
1789          */
1790         update_time_for_write(inode);
1791
1792         start_pos = round_down(pos, root->sectorsize);
1793         if (start_pos > i_size_read(inode)) {
1794                 /* Expand hole size to cover write data, preventing empty gap */
1795                 end_pos = round_up(pos + count, root->sectorsize);
1796                 err = btrfs_cont_expand(inode, i_size_read(inode), end_pos);
1797                 if (err) {
1798                         mutex_unlock(&inode->i_mutex);
1799                         goto out;
1800                 }
1801         }
1802
1803         if (sync)
1804                 atomic_inc(&BTRFS_I(inode)->sync_writers);
1805
1806         if (unlikely(file->f_flags & O_DIRECT)) {
1807                 num_written = __btrfs_direct_write(iocb, iov, nr_segs,
1808                                                    pos, count, ocount);
1809         } else {
1810                 struct iov_iter i;
1811
1812                 iov_iter_init(&i, iov, nr_segs, count, num_written);
1813
1814                 num_written = __btrfs_buffered_write(file, &i, pos);
1815                 if (num_written > 0)
1816                         iocb->ki_pos = pos + num_written;
1817         }
1818
1819         mutex_unlock(&inode->i_mutex);
1820
1821         /*
1822          * we want to make sure fsync finds this change
1823          * but we haven't joined a transaction running right now.
1824          *
1825          * Later on, someone is sure to update the inode and get the
1826          * real transid recorded.
1827          *
1828          * We set last_trans now to the fs_info generation + 1,
1829          * this will either be one more than the running transaction
1830          * or the generation used for the next transaction if there isn't
1831          * one running right now.
1832          *
1833          * We also have to set last_sub_trans to the current log transid,
1834          * otherwise subsequent syncs to a file that's been synced in this
1835          * transaction will appear to have already occured.
1836          */
1837         BTRFS_I(inode)->last_trans = root->fs_info->generation + 1;
1838         BTRFS_I(inode)->last_sub_trans = root->log_transid;
1839         if (num_written > 0) {
1840                 err = generic_write_sync(file, pos, num_written);
1841                 if (err < 0)
1842                         num_written = err;
1843         }
1844
1845         if (sync)
1846                 atomic_dec(&BTRFS_I(inode)->sync_writers);
1847 out:
1848         current->backing_dev_info = NULL;
1849         return num_written ? num_written : err;
1850 }
1851
1852 int btrfs_release_file(struct inode *inode, struct file *filp)
1853 {
1854         /*
1855          * ordered_data_close is set by settattr when we are about to truncate
1856          * a file from a non-zero size to a zero size.  This tries to
1857          * flush down new bytes that may have been written if the
1858          * application were using truncate to replace a file in place.
1859          */
1860         if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
1861                                &BTRFS_I(inode)->runtime_flags)) {
1862                 struct btrfs_trans_handle *trans;
1863                 struct btrfs_root *root = BTRFS_I(inode)->root;
1864
1865                 /*
1866                  * We need to block on a committing transaction to keep us from
1867                  * throwing a ordered operation on to the list and causing
1868                  * something like sync to deadlock trying to flush out this
1869                  * inode.
1870                  */
1871                 trans = btrfs_start_transaction(root, 0);
1872                 if (IS_ERR(trans))
1873                         return PTR_ERR(trans);
1874                 btrfs_add_ordered_operation(trans, BTRFS_I(inode)->root, inode);
1875                 btrfs_end_transaction(trans, root);
1876                 if (inode->i_size > BTRFS_ORDERED_OPERATIONS_FLUSH_LIMIT)
1877                         filemap_flush(inode->i_mapping);
1878         }
1879         if (filp->private_data)
1880                 btrfs_ioctl_trans_end(filp);
1881         return 0;
1882 }
1883
1884 /*
1885  * fsync call for both files and directories.  This logs the inode into
1886  * the tree log instead of forcing full commits whenever possible.
1887  *
1888  * It needs to call filemap_fdatawait so that all ordered extent updates are
1889  * in the metadata btree are up to date for copying to the log.
1890  *
1891  * It drops the inode mutex before doing the tree log commit.  This is an
1892  * important optimization for directories because holding the mutex prevents
1893  * new operations on the dir while we write to disk.
1894  */
1895 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
1896 {
1897         struct dentry *dentry = file->f_path.dentry;
1898         struct inode *inode = dentry->d_inode;
1899         struct btrfs_root *root = BTRFS_I(inode)->root;
1900         struct btrfs_trans_handle *trans;
1901         struct btrfs_log_ctx ctx;
1902         int ret = 0;
1903         bool full_sync = 0;
1904
1905         trace_btrfs_sync_file(file, datasync);
1906
1907         /*
1908          * We write the dirty pages in the range and wait until they complete
1909          * out of the ->i_mutex. If so, we can flush the dirty pages by
1910          * multi-task, and make the performance up.  See
1911          * btrfs_wait_ordered_range for an explanation of the ASYNC check.
1912          */
1913         atomic_inc(&BTRFS_I(inode)->sync_writers);
1914         ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
1915         if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1916                              &BTRFS_I(inode)->runtime_flags))
1917                 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
1918         atomic_dec(&BTRFS_I(inode)->sync_writers);
1919         if (ret)
1920                 return ret;
1921
1922         mutex_lock(&inode->i_mutex);
1923
1924         /*
1925          * We flush the dirty pages again to avoid some dirty pages in the
1926          * range being left.
1927          */
1928         atomic_inc(&root->log_batch);
1929         full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1930                              &BTRFS_I(inode)->runtime_flags);
1931         if (full_sync) {
1932                 ret = btrfs_wait_ordered_range(inode, start, end - start + 1);
1933                 if (ret) {
1934                         mutex_unlock(&inode->i_mutex);
1935                         goto out;
1936                 }
1937         }
1938         atomic_inc(&root->log_batch);
1939
1940         /*
1941          * check the transaction that last modified this inode
1942          * and see if its already been committed
1943          */
1944         if (!BTRFS_I(inode)->last_trans) {
1945                 mutex_unlock(&inode->i_mutex);
1946                 goto out;
1947         }
1948
1949         /*
1950          * if the last transaction that changed this file was before
1951          * the current transaction, we can bail out now without any
1952          * syncing
1953          */
1954         smp_mb();
1955         if (btrfs_inode_in_log(inode, root->fs_info->generation) ||
1956             BTRFS_I(inode)->last_trans <=
1957             root->fs_info->last_trans_committed) {
1958                 BTRFS_I(inode)->last_trans = 0;
1959
1960                 /*
1961                  * We'v had everything committed since the last time we were
1962                  * modified so clear this flag in case it was set for whatever
1963                  * reason, it's no longer relevant.
1964                  */
1965                 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1966                           &BTRFS_I(inode)->runtime_flags);
1967                 mutex_unlock(&inode->i_mutex);
1968                 goto out;
1969         }
1970
1971         /*
1972          * ok we haven't committed the transaction yet, lets do a commit
1973          */
1974         if (file->private_data)
1975                 btrfs_ioctl_trans_end(file);
1976
1977         /*
1978          * We use start here because we will need to wait on the IO to complete
1979          * in btrfs_sync_log, which could require joining a transaction (for
1980          * example checking cross references in the nocow path).  If we use join
1981          * here we could get into a situation where we're waiting on IO to
1982          * happen that is blocked on a transaction trying to commit.  With start
1983          * we inc the extwriter counter, so we wait for all extwriters to exit
1984          * before we start blocking join'ers.  This comment is to keep somebody
1985          * from thinking they are super smart and changing this to
1986          * btrfs_join_transaction *cough*Josef*cough*.
1987          */
1988         trans = btrfs_start_transaction(root, 0);
1989         if (IS_ERR(trans)) {
1990                 ret = PTR_ERR(trans);
1991                 mutex_unlock(&inode->i_mutex);
1992                 goto out;
1993         }
1994         trans->sync = true;
1995
1996         btrfs_init_log_ctx(&ctx);
1997
1998         ret = btrfs_log_dentry_safe(trans, root, dentry, &ctx);
1999         if (ret < 0) {
2000                 /* Fallthrough and commit/free transaction. */
2001                 ret = 1;
2002         }
2003
2004         /* we've logged all the items and now have a consistent
2005          * version of the file in the log.  It is possible that
2006          * someone will come in and modify the file, but that's
2007          * fine because the log is consistent on disk, and we
2008          * have references to all of the file's extents
2009          *
2010          * It is possible that someone will come in and log the
2011          * file again, but that will end up using the synchronization
2012          * inside btrfs_sync_log to keep things safe.
2013          */
2014         mutex_unlock(&inode->i_mutex);
2015
2016         if (ret != BTRFS_NO_LOG_SYNC) {
2017                 if (!ret) {
2018                         ret = btrfs_sync_log(trans, root, &ctx);
2019                         if (!ret) {
2020                                 ret = btrfs_end_transaction(trans, root);
2021                                 goto out;
2022                         }
2023                 }
2024                 if (!full_sync) {
2025                         ret = btrfs_wait_ordered_range(inode, start,
2026                                                        end - start + 1);
2027                         if (ret)
2028                                 goto out;
2029                 }
2030                 ret = btrfs_commit_transaction(trans, root);
2031         } else {
2032                 ret = btrfs_end_transaction(trans, root);
2033         }
2034 out:
2035         return ret > 0 ? -EIO : ret;
2036 }
2037
2038 static const struct vm_operations_struct btrfs_file_vm_ops = {
2039         .fault          = filemap_fault,
2040         .map_pages      = filemap_map_pages,
2041         .page_mkwrite   = btrfs_page_mkwrite,
2042         .remap_pages    = generic_file_remap_pages,
2043 };
2044
2045 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
2046 {
2047         struct address_space *mapping = filp->f_mapping;
2048
2049         if (!mapping->a_ops->readpage)
2050                 return -ENOEXEC;
2051
2052         file_accessed(filp);
2053         vma->vm_ops = &btrfs_file_vm_ops;
2054
2055         return 0;
2056 }
2057
2058 static int hole_mergeable(struct inode *inode, struct extent_buffer *leaf,
2059                           int slot, u64 start, u64 end)
2060 {
2061         struct btrfs_file_extent_item *fi;
2062         struct btrfs_key key;
2063
2064         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2065                 return 0;
2066
2067         btrfs_item_key_to_cpu(leaf, &key, slot);
2068         if (key.objectid != btrfs_ino(inode) ||
2069             key.type != BTRFS_EXTENT_DATA_KEY)
2070                 return 0;
2071
2072         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2073
2074         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2075                 return 0;
2076
2077         if (btrfs_file_extent_disk_bytenr(leaf, fi))
2078                 return 0;
2079
2080         if (key.offset == end)
2081                 return 1;
2082         if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2083                 return 1;
2084         return 0;
2085 }
2086
2087 static int fill_holes(struct btrfs_trans_handle *trans, struct inode *inode,
2088                       struct btrfs_path *path, u64 offset, u64 end)
2089 {
2090         struct btrfs_root *root = BTRFS_I(inode)->root;
2091         struct extent_buffer *leaf;
2092         struct btrfs_file_extent_item *fi;
2093         struct extent_map *hole_em;
2094         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2095         struct btrfs_key key;
2096         int ret;
2097
2098         if (btrfs_fs_incompat(root->fs_info, NO_HOLES))
2099                 goto out;
2100
2101         key.objectid = btrfs_ino(inode);
2102         key.type = BTRFS_EXTENT_DATA_KEY;
2103         key.offset = offset;
2104
2105         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2106         if (ret < 0)
2107                 return ret;
2108         BUG_ON(!ret);
2109
2110         leaf = path->nodes[0];
2111         if (hole_mergeable(inode, leaf, path->slots[0]-1, offset, end)) {
2112                 u64 num_bytes;
2113
2114                 path->slots[0]--;
2115                 fi = btrfs_item_ptr(leaf, path->slots[0],
2116                                     struct btrfs_file_extent_item);
2117                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2118                         end - offset;
2119                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2120                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2121                 btrfs_set_file_extent_offset(leaf, fi, 0);
2122                 btrfs_mark_buffer_dirty(leaf);
2123                 goto out;
2124         }
2125
2126         if (hole_mergeable(inode, leaf, path->slots[0]+1, offset, end)) {
2127                 u64 num_bytes;
2128
2129                 path->slots[0]++;
2130                 key.offset = offset;
2131                 btrfs_set_item_key_safe(root, path, &key);
2132                 fi = btrfs_item_ptr(leaf, path->slots[0],
2133                                     struct btrfs_file_extent_item);
2134                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2135                         offset;
2136                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2137                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2138                 btrfs_set_file_extent_offset(leaf, fi, 0);
2139                 btrfs_mark_buffer_dirty(leaf);
2140                 goto out;
2141         }
2142         btrfs_release_path(path);
2143
2144         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode), offset,
2145                                        0, 0, end - offset, 0, end - offset,
2146                                        0, 0, 0);
2147         if (ret)
2148                 return ret;
2149
2150 out:
2151         btrfs_release_path(path);
2152
2153         hole_em = alloc_extent_map();
2154         if (!hole_em) {
2155                 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2156                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2157                         &BTRFS_I(inode)->runtime_flags);
2158         } else {
2159                 hole_em->start = offset;
2160                 hole_em->len = end - offset;
2161                 hole_em->ram_bytes = hole_em->len;
2162                 hole_em->orig_start = offset;
2163
2164                 hole_em->block_start = EXTENT_MAP_HOLE;
2165                 hole_em->block_len = 0;
2166                 hole_em->orig_block_len = 0;
2167                 hole_em->bdev = root->fs_info->fs_devices->latest_bdev;
2168                 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2169                 hole_em->generation = trans->transid;
2170
2171                 do {
2172                         btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2173                         write_lock(&em_tree->lock);
2174                         ret = add_extent_mapping(em_tree, hole_em, 1);
2175                         write_unlock(&em_tree->lock);
2176                 } while (ret == -EEXIST);
2177                 free_extent_map(hole_em);
2178                 if (ret)
2179                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2180                                 &BTRFS_I(inode)->runtime_flags);
2181         }
2182
2183         return 0;
2184 }
2185
2186 static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2187 {
2188         struct btrfs_root *root = BTRFS_I(inode)->root;
2189         struct extent_state *cached_state = NULL;
2190         struct btrfs_path *path;
2191         struct btrfs_block_rsv *rsv;
2192         struct btrfs_trans_handle *trans;
2193         u64 lockstart = round_up(offset, BTRFS_I(inode)->root->sectorsize);
2194         u64 lockend = round_down(offset + len,
2195                                  BTRFS_I(inode)->root->sectorsize) - 1;
2196         u64 cur_offset = lockstart;
2197         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
2198         u64 drop_end;
2199         int ret = 0;
2200         int err = 0;
2201         int rsv_count;
2202         bool same_page = ((offset >> PAGE_CACHE_SHIFT) ==
2203                           ((offset + len - 1) >> PAGE_CACHE_SHIFT));
2204         bool no_holes = btrfs_fs_incompat(root->fs_info, NO_HOLES);
2205         u64 ino_size;
2206
2207         ret = btrfs_wait_ordered_range(inode, offset, len);
2208         if (ret)
2209                 return ret;
2210
2211         mutex_lock(&inode->i_mutex);
2212         ino_size = round_up(inode->i_size, PAGE_CACHE_SIZE);
2213         /*
2214          * We needn't truncate any page which is beyond the end of the file
2215          * because we are sure there is no data there.
2216          */
2217         /*
2218          * Only do this if we are in the same page and we aren't doing the
2219          * entire page.
2220          */
2221         if (same_page && len < PAGE_CACHE_SIZE) {
2222                 if (offset < ino_size)
2223                         ret = btrfs_truncate_page(inode, offset, len, 0);
2224                 mutex_unlock(&inode->i_mutex);
2225                 return ret;
2226         }
2227
2228         /* zero back part of the first page */
2229         if (offset < ino_size) {
2230                 ret = btrfs_truncate_page(inode, offset, 0, 0);
2231                 if (ret) {
2232                         mutex_unlock(&inode->i_mutex);
2233                         return ret;
2234                 }
2235         }
2236
2237         /* zero the front end of the last page */
2238         if (offset + len < ino_size) {
2239                 ret = btrfs_truncate_page(inode, offset + len, 0, 1);
2240                 if (ret) {
2241                         mutex_unlock(&inode->i_mutex);
2242                         return ret;
2243                 }
2244         }
2245
2246         if (lockend < lockstart) {
2247                 mutex_unlock(&inode->i_mutex);
2248                 return 0;
2249         }
2250
2251         while (1) {
2252                 struct btrfs_ordered_extent *ordered;
2253
2254                 truncate_pagecache_range(inode, lockstart, lockend);
2255
2256                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2257                                  0, &cached_state);
2258                 ordered = btrfs_lookup_first_ordered_extent(inode, lockend);
2259
2260                 /*
2261                  * We need to make sure we have no ordered extents in this range
2262                  * and nobody raced in and read a page in this range, if we did
2263                  * we need to try again.
2264                  */
2265                 if ((!ordered ||
2266                     (ordered->file_offset + ordered->len <= lockstart ||
2267                      ordered->file_offset > lockend)) &&
2268                      !test_range_bit(&BTRFS_I(inode)->io_tree, lockstart,
2269                                      lockend, EXTENT_UPTODATE, 0,
2270                                      cached_state)) {
2271                         if (ordered)
2272                                 btrfs_put_ordered_extent(ordered);
2273                         break;
2274                 }
2275                 if (ordered)
2276                         btrfs_put_ordered_extent(ordered);
2277                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2278                                      lockend, &cached_state, GFP_NOFS);
2279                 ret = btrfs_wait_ordered_range(inode, lockstart,
2280                                                lockend - lockstart + 1);
2281                 if (ret) {
2282                         mutex_unlock(&inode->i_mutex);
2283                         return ret;
2284                 }
2285         }
2286
2287         path = btrfs_alloc_path();
2288         if (!path) {
2289                 ret = -ENOMEM;
2290                 goto out;
2291         }
2292
2293         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
2294         if (!rsv) {
2295                 ret = -ENOMEM;
2296                 goto out_free;
2297         }
2298         rsv->size = btrfs_calc_trunc_metadata_size(root, 1);
2299         rsv->failfast = 1;
2300
2301         /*
2302          * 1 - update the inode
2303          * 1 - removing the extents in the range
2304          * 1 - adding the hole extent if no_holes isn't set
2305          */
2306         rsv_count = no_holes ? 2 : 3;
2307         trans = btrfs_start_transaction(root, rsv_count);
2308         if (IS_ERR(trans)) {
2309                 err = PTR_ERR(trans);
2310                 goto out_free;
2311         }
2312
2313         ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv, rsv,
2314                                       min_size);
2315         BUG_ON(ret);
2316         trans->block_rsv = rsv;
2317
2318         while (cur_offset < lockend) {
2319                 ret = __btrfs_drop_extents(trans, root, inode, path,
2320                                            cur_offset, lockend + 1,
2321                                            &drop_end, 1, 0, 0, NULL);
2322                 if (ret != -ENOSPC)
2323                         break;
2324
2325                 trans->block_rsv = &root->fs_info->trans_block_rsv;
2326
2327                 if (cur_offset < ino_size) {
2328                         ret = fill_holes(trans, inode, path, cur_offset,
2329                                          drop_end);
2330                         if (ret) {
2331                                 err = ret;
2332                                 break;
2333                         }
2334                 }
2335
2336                 cur_offset = drop_end;
2337
2338                 ret = btrfs_update_inode(trans, root, inode);
2339                 if (ret) {
2340                         err = ret;
2341                         break;
2342                 }
2343
2344                 btrfs_end_transaction(trans, root);
2345                 btrfs_btree_balance_dirty(root);
2346
2347                 trans = btrfs_start_transaction(root, rsv_count);
2348                 if (IS_ERR(trans)) {
2349                         ret = PTR_ERR(trans);
2350                         trans = NULL;
2351                         break;
2352                 }
2353
2354                 ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv,
2355                                               rsv, min_size);
2356                 BUG_ON(ret);    /* shouldn't happen */
2357                 trans->block_rsv = rsv;
2358         }
2359
2360         if (ret) {
2361                 err = ret;
2362                 goto out_trans;
2363         }
2364
2365         trans->block_rsv = &root->fs_info->trans_block_rsv;
2366         /*
2367          * Don't insert file hole extent item if it's for a range beyond eof
2368          * (because it's useless) or if it represents a 0 bytes range (when
2369          * cur_offset == drop_end).
2370          */
2371         if (cur_offset < ino_size && cur_offset < drop_end) {
2372                 ret = fill_holes(trans, inode, path, cur_offset, drop_end);
2373                 if (ret) {
2374                         err = ret;
2375                         goto out_trans;
2376                 }
2377         }
2378
2379 out_trans:
2380         if (!trans)
2381                 goto out_free;
2382
2383         inode_inc_iversion(inode);
2384         inode->i_mtime = inode->i_ctime = CURRENT_TIME;
2385
2386         trans->block_rsv = &root->fs_info->trans_block_rsv;
2387         ret = btrfs_update_inode(trans, root, inode);
2388         btrfs_end_transaction(trans, root);
2389         btrfs_btree_balance_dirty(root);
2390 out_free:
2391         btrfs_free_path(path);
2392         btrfs_free_block_rsv(root, rsv);
2393 out:
2394         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2395                              &cached_state, GFP_NOFS);
2396         mutex_unlock(&inode->i_mutex);
2397         if (ret && !err)
2398                 err = ret;
2399         return err;
2400 }
2401
2402 static long btrfs_fallocate(struct file *file, int mode,
2403                             loff_t offset, loff_t len)
2404 {
2405         struct inode *inode = file_inode(file);
2406         struct extent_state *cached_state = NULL;
2407         struct btrfs_root *root = BTRFS_I(inode)->root;
2408         u64 cur_offset;
2409         u64 last_byte;
2410         u64 alloc_start;
2411         u64 alloc_end;
2412         u64 alloc_hint = 0;
2413         u64 locked_end;
2414         struct extent_map *em;
2415         int blocksize = BTRFS_I(inode)->root->sectorsize;
2416         int ret;
2417
2418         alloc_start = round_down(offset, blocksize);
2419         alloc_end = round_up(offset + len, blocksize);
2420
2421         /* Make sure we aren't being give some crap mode */
2422         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2423                 return -EOPNOTSUPP;
2424
2425         if (mode & FALLOC_FL_PUNCH_HOLE)
2426                 return btrfs_punch_hole(inode, offset, len);
2427
2428         /*
2429          * Make sure we have enough space before we do the
2430          * allocation.
2431          */
2432         ret = btrfs_check_data_free_space(inode, alloc_end - alloc_start);
2433         if (ret)
2434                 return ret;
2435         if (root->fs_info->quota_enabled) {
2436                 ret = btrfs_qgroup_reserve(root, alloc_end - alloc_start);
2437                 if (ret)
2438                         goto out_reserve_fail;
2439         }
2440
2441         mutex_lock(&inode->i_mutex);
2442         ret = inode_newsize_ok(inode, alloc_end);
2443         if (ret)
2444                 goto out;
2445
2446         if (alloc_start > inode->i_size) {
2447                 ret = btrfs_cont_expand(inode, i_size_read(inode),
2448                                         alloc_start);
2449                 if (ret)
2450                         goto out;
2451         } else {
2452                 /*
2453                  * If we are fallocating from the end of the file onward we
2454                  * need to zero out the end of the page if i_size lands in the
2455                  * middle of a page.
2456                  */
2457                 ret = btrfs_truncate_page(inode, inode->i_size, 0, 0);
2458                 if (ret)
2459                         goto out;
2460         }
2461
2462         /*
2463          * wait for ordered IO before we have any locks.  We'll loop again
2464          * below with the locks held.
2465          */
2466         ret = btrfs_wait_ordered_range(inode, alloc_start,
2467                                        alloc_end - alloc_start);
2468         if (ret)
2469                 goto out;
2470
2471         locked_end = alloc_end - 1;
2472         while (1) {
2473                 struct btrfs_ordered_extent *ordered;
2474
2475                 /* the extent lock is ordered inside the running
2476                  * transaction
2477                  */
2478                 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
2479                                  locked_end, 0, &cached_state);
2480                 ordered = btrfs_lookup_first_ordered_extent(inode,
2481                                                             alloc_end - 1);
2482                 if (ordered &&
2483                     ordered->file_offset + ordered->len > alloc_start &&
2484                     ordered->file_offset < alloc_end) {
2485                         btrfs_put_ordered_extent(ordered);
2486                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
2487                                              alloc_start, locked_end,
2488                                              &cached_state, GFP_NOFS);
2489                         /*
2490                          * we can't wait on the range with the transaction
2491                          * running or with the extent lock held
2492                          */
2493                         ret = btrfs_wait_ordered_range(inode, alloc_start,
2494                                                        alloc_end - alloc_start);
2495                         if (ret)
2496                                 goto out;
2497                 } else {
2498                         if (ordered)
2499                                 btrfs_put_ordered_extent(ordered);
2500                         break;
2501                 }
2502         }
2503
2504         cur_offset = alloc_start;
2505         while (1) {
2506                 u64 actual_end;
2507
2508                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
2509                                       alloc_end - cur_offset, 0);
2510                 if (IS_ERR_OR_NULL(em)) {
2511                         if (!em)
2512                                 ret = -ENOMEM;
2513                         else
2514                                 ret = PTR_ERR(em);
2515                         break;
2516                 }
2517                 last_byte = min(extent_map_end(em), alloc_end);
2518                 actual_end = min_t(u64, extent_map_end(em), offset + len);
2519                 last_byte = ALIGN(last_byte, blocksize);
2520
2521                 if (em->block_start == EXTENT_MAP_HOLE ||
2522                     (cur_offset >= inode->i_size &&
2523                      !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
2524                         ret = btrfs_prealloc_file_range(inode, mode, cur_offset,
2525                                                         last_byte - cur_offset,
2526                                                         1 << inode->i_blkbits,
2527                                                         offset + len,
2528                                                         &alloc_hint);
2529
2530                         if (ret < 0) {
2531                                 free_extent_map(em);
2532                                 break;
2533                         }
2534                 } else if (actual_end > inode->i_size &&
2535                            !(mode & FALLOC_FL_KEEP_SIZE)) {
2536                         /*
2537                          * We didn't need to allocate any more space, but we
2538                          * still extended the size of the file so we need to
2539                          * update i_size.
2540                          */
2541                         inode->i_ctime = CURRENT_TIME;
2542                         i_size_write(inode, actual_end);
2543                         btrfs_ordered_update_i_size(inode, actual_end, NULL);
2544                 }
2545                 free_extent_map(em);
2546
2547                 cur_offset = last_byte;
2548                 if (cur_offset >= alloc_end) {
2549                         ret = 0;
2550                         break;
2551                 }
2552         }
2553         unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
2554                              &cached_state, GFP_NOFS);
2555 out:
2556         mutex_unlock(&inode->i_mutex);
2557         if (root->fs_info->quota_enabled)
2558                 btrfs_qgroup_free(root, alloc_end - alloc_start);
2559 out_reserve_fail:
2560         /* Let go of our reservation. */
2561         btrfs_free_reserved_data_space(inode, alloc_end - alloc_start);
2562         return ret;
2563 }
2564
2565 static int find_desired_extent(struct inode *inode, loff_t *offset, int whence)
2566 {
2567         struct btrfs_root *root = BTRFS_I(inode)->root;
2568         struct extent_map *em = NULL;
2569         struct extent_state *cached_state = NULL;
2570         u64 lockstart = *offset;
2571         u64 lockend = i_size_read(inode);
2572         u64 start = *offset;
2573         u64 len = i_size_read(inode);
2574         int ret = 0;
2575
2576         lockend = max_t(u64, root->sectorsize, lockend);
2577         if (lockend <= lockstart)
2578                 lockend = lockstart + root->sectorsize;
2579
2580         lockend--;
2581         len = lockend - lockstart + 1;
2582
2583         len = max_t(u64, len, root->sectorsize);
2584         if (inode->i_size == 0)
2585                 return -ENXIO;
2586
2587         lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend, 0,
2588                          &cached_state);
2589
2590         while (start < inode->i_size) {
2591                 em = btrfs_get_extent_fiemap(inode, NULL, 0, start, len, 0);
2592                 if (IS_ERR(em)) {
2593                         ret = PTR_ERR(em);
2594                         em = NULL;
2595                         break;
2596                 }
2597
2598                 if (whence == SEEK_HOLE &&
2599                     (em->block_start == EXTENT_MAP_HOLE ||
2600                      test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
2601                         break;
2602                 else if (whence == SEEK_DATA &&
2603                            (em->block_start != EXTENT_MAP_HOLE &&
2604                             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
2605                         break;
2606
2607                 start = em->start + em->len;
2608                 free_extent_map(em);
2609                 em = NULL;
2610                 cond_resched();
2611         }
2612         free_extent_map(em);
2613         if (!ret) {
2614                 if (whence == SEEK_DATA && start >= inode->i_size)
2615                         ret = -ENXIO;
2616                 else
2617                         *offset = min_t(loff_t, start, inode->i_size);
2618         }
2619         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2620                              &cached_state, GFP_NOFS);
2621         return ret;
2622 }
2623
2624 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
2625 {
2626         struct inode *inode = file->f_mapping->host;
2627         int ret;
2628
2629         mutex_lock(&inode->i_mutex);
2630         switch (whence) {
2631         case SEEK_END:
2632         case SEEK_CUR:
2633                 offset = generic_file_llseek(file, offset, whence);
2634                 goto out;
2635         case SEEK_DATA:
2636         case SEEK_HOLE:
2637                 if (offset >= i_size_read(inode)) {
2638                         mutex_unlock(&inode->i_mutex);
2639                         return -ENXIO;
2640                 }
2641
2642                 ret = find_desired_extent(inode, &offset, whence);
2643                 if (ret) {
2644                         mutex_unlock(&inode->i_mutex);
2645                         return ret;
2646                 }
2647         }
2648
2649         offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
2650 out:
2651         mutex_unlock(&inode->i_mutex);
2652         return offset;
2653 }
2654
2655 const struct file_operations btrfs_file_operations = {
2656         .llseek         = btrfs_file_llseek,
2657         .read           = do_sync_read,
2658         .write          = do_sync_write,
2659         .aio_read       = generic_file_aio_read,
2660         .splice_read    = generic_file_splice_read,
2661         .aio_write      = btrfs_file_aio_write,
2662         .mmap           = btrfs_file_mmap,
2663         .open           = generic_file_open,
2664         .release        = btrfs_release_file,
2665         .fsync          = btrfs_sync_file,
2666         .fallocate      = btrfs_fallocate,
2667         .unlocked_ioctl = btrfs_ioctl,
2668 #ifdef CONFIG_COMPAT
2669         .compat_ioctl   = btrfs_ioctl,
2670 #endif
2671 };
2672
2673 void btrfs_auto_defrag_exit(void)
2674 {
2675         if (btrfs_inode_defrag_cachep)
2676                 kmem_cache_destroy(btrfs_inode_defrag_cachep);
2677 }
2678
2679 int btrfs_auto_defrag_init(void)
2680 {
2681         btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
2682                                         sizeof(struct inode_defrag), 0,
2683                                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD,
2684                                         NULL);
2685         if (!btrfs_inode_defrag_cachep)
2686                 return -ENOMEM;
2687
2688         return 0;
2689 }