fs/btrfs/file.c

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