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