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_write_iter 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->start < start) {
 600                         split->start = em->start;
 601                         split->len = start - em->start;
 602
 603                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 604                                 split->orig_start = em->orig_start;
 605                                 split->block_start = em->block_start;
 606
 607                                 if (compressed)
 608                                         split->block_len = em->block_len;
 609                                 else
 610                                         split->block_len = split->len;
 611                                 split->orig_block_len = max(split->block_len,
 612                                                 em->orig_block_len);
 613                                 split->ram_bytes = em->ram_bytes;
 614                         } else {
 615                                 split->orig_start = split->start;
 616                                 split->block_len = 0;
 617                                 split->block_start = em->block_start;
 618                                 split->orig_block_len = 0;
 619                                 split->ram_bytes = split->len;
 620                         }
 621
 622                         split->generation = gen;
 623                         split->bdev = em->bdev;
 624                         split->flags = flags;
 625                         split->compress_type = em->compress_type;
 626                         ret = add_extent_mapping(em_tree, split, modified);
 627                         BUG_ON(ret); /* Logic error */
 628                         free_extent_map(split);
 629                         split = split2;
 630                         split2 = NULL;
 631                 }
 632                 if (testend && em->start + em->len > start + len) {
 633                         u64 diff = start + len - em->start;
 634
 635                         split->start = start + len;
 636                         split->len = em->start + em->len - (start + len);
 637                         split->bdev = em->bdev;
 638                         split->flags = flags;
 639                         split->compress_type = em->compress_type;
 640                         split->generation = gen;
 641
 642                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 643                                 split->orig_block_len = max(em->block_len,
 644                                                     em->orig_block_len);
 645
 646                                 split->ram_bytes = em->ram_bytes;
 647                                 if (compressed) {
 648                                         split->block_len = em->block_len;
 649                                         split->block_start = em->block_start;
 650                                         split->orig_start = em->orig_start;
 651                                 } else {
 652                                         split->block_len = split->len;
 653                                         split->block_start = em->block_start
 654                                                 + diff;
 655                                         split->orig_start = em->orig_start;
 656                                 }
 657                         } else {
 658                                 split->ram_bytes = split->len;
 659                                 split->orig_start = split->start;
 660                                 split->block_len = 0;
 661                                 split->block_start = em->block_start;
 662                                 split->orig_block_len = 0;
 663                         }
 664
 665                         ret = add_extent_mapping(em_tree, split, modified);
 666                         BUG_ON(ret); /* Logic error */
 667                         free_extent_map(split);
 668                         split = NULL;
 669                 }
 670 next:
 671                 write_unlock(&em_tree->lock);
 672
 673                 /* once for us */
 674                 free_extent_map(em);
 675                 /* once for the tree*/
 676                 free_extent_map(em);
 677         }
 678         if (split)
 679                 free_extent_map(split);
 680         if (split2)
 681                 free_extent_map(split2);
 682 }
 683
 684 /*
 685  * this is very complex, but the basic idea is to drop all extents
 686  * in the range start - end.  hint_block is filled in with a block number
 687  * that would be a good hint to the block allocator for this file.
 688  *
 689  * If an extent intersects the range but is not entirely inside the range
 690  * it is either truncated or split.  Anything entirely inside the range
 691  * is deleted from the tree.
 692  */
 693 int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
 694                          struct btrfs_root *root, struct inode *inode,
 695                          struct btrfs_path *path, u64 start, u64 end,
 696                          u64 *drop_end, int drop_cache)
 697 {
 698         struct extent_buffer *leaf;
 699         struct btrfs_file_extent_item *fi;
 700         struct btrfs_key key;
 701         struct btrfs_key new_key;
 702         u64 ino = btrfs_ino(inode);
 703         u64 search_start = start;
 704         u64 disk_bytenr = 0;
 705         u64 num_bytes = 0;
 706         u64 extent_offset = 0;
 707         u64 extent_end = 0;
 708         int del_nr = 0;
 709         int del_slot = 0;
 710         int extent_type;
 711         int recow;
 712         int ret;
 713         int modify_tree = -1;
 714         int update_refs = (root->ref_cows || root == root->fs_info->tree_root);
 715         int found = 0;
 716
 717         if (drop_cache)
 718                 btrfs_drop_extent_cache(inode, start, end - 1, 0);
 719
 720         if (start >= BTRFS_I(inode)->disk_i_size)
 721                 modify_tree = 0;
 722
 723         while (1) {
 724                 recow = 0;
 725                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
 726                                                search_start, modify_tree);
 727                 if (ret < 0)
 728                         break;
 729                 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
 730                         leaf = path->nodes[0];
 731                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
 732                         if (key.objectid == ino &&
 733                             key.type == BTRFS_EXTENT_DATA_KEY)
 734                                 path->slots[0]--;
 735                 }
 736                 ret = 0;
 737 next_slot:
 738                 leaf = path->nodes[0];
 739                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 740                         BUG_ON(del_nr > 0);
 741                         ret = btrfs_next_leaf(root, path);
 742                         if (ret < 0)
 743                                 break;
 744                         if (ret > 0) {
 745                                 ret = 0;
 746                                 break;
 747                         }
 748                         leaf = path->nodes[0];
 749                         recow = 1;
 750                 }
 751
 752                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 753                 if (key.objectid > ino ||
 754                     key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
 755                         break;
 756
 757                 fi = btrfs_item_ptr(leaf, path->slots[0],
 758                                     struct btrfs_file_extent_item);
 759                 extent_type = btrfs_file_extent_type(leaf, fi);
 760
 761                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
 762                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
 763                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 764                         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 765                         extent_offset = btrfs_file_extent_offset(leaf, fi);
 766                         extent_end = key.offset +
 767                                 btrfs_file_extent_num_bytes(leaf, fi);
 768                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 769                         extent_end = key.offset +
 770                                 btrfs_file_extent_inline_len(leaf, fi);
 771                 } else {
 772                         WARN_ON(1);
 773                         extent_end = search_start;
 774                 }
 775
 776                 if (extent_end <= search_start) {
 777                         path->slots[0]++;
 778                         goto next_slot;
 779                 }
 780
 781                 found = 1;
 782                 search_start = max(key.offset, start);
 783                 if (recow || !modify_tree) {
 784                         modify_tree = -1;
 785                         btrfs_release_path(path);
 786                         continue;
 787                 }
 788
 789                 /*
 790                  *     | - range to drop - |
 791                  *  | -------- extent -------- |
 792                  */
 793                 if (start > key.offset && end < extent_end) {
 794                         BUG_ON(del_nr > 0);
 795                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
 796
 797                         memcpy(&new_key, &key, sizeof(new_key));
 798                         new_key.offset = start;
 799                         ret = btrfs_duplicate_item(trans, root, path,
 800                                                    &new_key);
 801                         if (ret == -EAGAIN) {
 802                                 btrfs_release_path(path);
 803                                 continue;
 804                         }
 805                         if (ret < 0)
 806                                 break;
 807
 808                         leaf = path->nodes[0];
 809                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 810                                             struct btrfs_file_extent_item);
 811                         btrfs_set_file_extent_num_bytes(leaf, fi,
 812                                                         start - key.offset);
 813
 814                         fi = btrfs_item_ptr(leaf, path->slots[0],
 815                                             struct btrfs_file_extent_item);
 816
 817                         extent_offset += start - key.offset;
 818                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 819                         btrfs_set_file_extent_num_bytes(leaf, fi,
 820                                                         extent_end - start);
 821                         btrfs_mark_buffer_dirty(leaf);
 822
 823                         if (update_refs && disk_bytenr > 0) {
 824                                 ret = btrfs_inc_extent_ref(trans, root,
 825                                                 disk_bytenr, num_bytes, 0,
 826                                                 root->root_key.objectid,
 827                                                 new_key.objectid,
 828                                                 start - extent_offset, 0);
 829                                 BUG_ON(ret); /* -ENOMEM */
 830                         }
 831                         key.offset = start;
 832                 }
 833                 /*
 834                  *  | ---- range to drop ----- |
 835                  *      | -------- extent -------- |
 836                  */
 837                 if (start <= key.offset && end < extent_end) {
 838                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
 839
 840                         memcpy(&new_key, &key, sizeof(new_key));
 841                         new_key.offset = end;
 842                         btrfs_set_item_key_safe(root, path, &new_key);
 843
 844                         extent_offset += end - key.offset;
 845                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 846                         btrfs_set_file_extent_num_bytes(leaf, fi,
 847                                                         extent_end - end);
 848                         btrfs_mark_buffer_dirty(leaf);
 849                         if (update_refs && disk_bytenr > 0)
 850                                 inode_sub_bytes(inode, end - key.offset);
 851                         break;
 852                 }
 853
 854                 search_start = extent_end;
 855                 /*
 856                  *       | ---- range to drop ----- |
 857                  *  | -------- extent -------- |
 858                  */
 859                 if (start > key.offset && end >= extent_end) {
 860                         BUG_ON(del_nr > 0);
 861                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
 862
 863                         btrfs_set_file_extent_num_bytes(leaf, fi,
 864                                                         start - key.offset);
 865                         btrfs_mark_buffer_dirty(leaf);
 866                         if (update_refs && disk_bytenr > 0)
 867                                 inode_sub_bytes(inode, extent_end - start);
 868                         if (end == extent_end)
 869                                 break;
 870
 871                         path->slots[0]++;
 872                         goto next_slot;
 873                 }
 874
 875                 /*
 876                  *  | ---- range to drop ----- |
 877                  *    | ------ extent ------ |
 878                  */
 879                 if (start <= key.offset && end >= extent_end) {
 880                         if (del_nr == 0) {
 881                                 del_slot = path->slots[0];
 882                                 del_nr = 1;
 883                         } else {
 884                                 BUG_ON(del_slot + del_nr != path->slots[0]);
 885                                 del_nr++;
 886                         }
 887
 888                         if (update_refs &&
 889                             extent_type == BTRFS_FILE_EXTENT_INLINE) {
 890                                 inode_sub_bytes(inode,
 891                                                 extent_end - key.offset);
 892                                 extent_end = ALIGN(extent_end,
 893                                                    root->sectorsize);
 894                         } else if (update_refs && disk_bytenr > 0) {
 895                                 ret = btrfs_free_extent(trans, root,
 896                                                 disk_bytenr, num_bytes, 0,
 897                                                 root->root_key.objectid,
 898                                                 key.objectid, key.offset -
 899                                                 extent_offset, 0);
 900                                 BUG_ON(ret); /* -ENOMEM */
 901                                 inode_sub_bytes(inode,
 902                                                 extent_end - key.offset);
 903                         }
 904
 905                         if (end == extent_end)
 906                                 break;
 907
 908                         if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
 909                                 path->slots[0]++;
 910                                 goto next_slot;
 911                         }
 912
 913                         ret = btrfs_del_items(trans, root, path, del_slot,
 914                                               del_nr);
 915                         if (ret) {
 916                                 btrfs_abort_transaction(trans, root, ret);
 917                                 break;
 918                         }
 919
 920                         del_nr = 0;
 921                         del_slot = 0;
 922
 923                         btrfs_release_path(path);
 924                         continue;
 925                 }
 926
 927                 BUG_ON(1);
 928         }
 929
 930         if (!ret && del_nr > 0) {
 931                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
 932                 if (ret)
 933                         btrfs_abort_transaction(trans, root, ret);
 934         }
 935
 936         if (drop_end)
 937                 *drop_end = found ? min(end, extent_end) : end;
 938         btrfs_release_path(path);
 939         return ret;
 940 }
 941
 942 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
 943                        struct btrfs_root *root, struct inode *inode, u64 start,
 944                        u64 end, int drop_cache)
 945 {
 946         struct btrfs_path *path;
 947         int ret;
 948
 949         path = btrfs_alloc_path();
 950         if (!path)
 951                 return -ENOMEM;
 952         ret = __btrfs_drop_extents(trans, root, inode, path, start, end, NULL,
 953                                    drop_cache);
 954         btrfs_free_path(path);
 955         return ret;
 956 }
 957
 958 static int extent_mergeable(struct extent_buffer *leaf, int slot,
 959                             u64 objectid, u64 bytenr, u64 orig_offset,
 960                             u64 *start, u64 *end)
 961 {
 962         struct btrfs_file_extent_item *fi;
 963         struct btrfs_key key;
 964         u64 extent_end;
 965
 966         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
 967                 return 0;
 968
 969         btrfs_item_key_to_cpu(leaf, &key, slot);
 970         if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
 971                 return 0;
 972
 973         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
 974         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
 975             btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
 976             btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
 977             btrfs_file_extent_compression(leaf, fi) ||
 978             btrfs_file_extent_encryption(leaf, fi) ||
 979             btrfs_file_extent_other_encoding(leaf, fi))
 980                 return 0;
 981
 982         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
 983         if ((*start && *start != key.offset) || (*end && *end != extent_end))
 984                 return 0;
 985
 986         *start = key.offset;
 987         *end = extent_end;
 988         return 1;
 989 }
 990
 991 /*
 992  * Mark extent in the range start - end as written.
 993  *
 994  * This changes extent type from 'pre-allocated' to 'regular'. If only
 995  * part of extent is marked as written, the extent will be split into
 996  * two or three.
 997  */
 998 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
 999                               struct inode *inode, u64 start, u64 end)
1000 {
1001         struct btrfs_root *root = BTRFS_I(inode)->root;
1002         struct extent_buffer *leaf;
1003         struct btrfs_path *path;
1004         struct btrfs_file_extent_item *fi;
1005         struct btrfs_key key;
1006         struct btrfs_key new_key;
1007         u64 bytenr;
1008         u64 num_bytes;
1009         u64 extent_end;
1010         u64 orig_offset;
1011         u64 other_start;
1012         u64 other_end;
1013         u64 split;
1014         int del_nr = 0;
1015         int del_slot = 0;
1016         int recow;
1017         int ret;
1018         u64 ino = btrfs_ino(inode);
1019
1020         path = btrfs_alloc_path();
1021         if (!path)
1022                 return -ENOMEM;
1023 again:
1024         recow = 0;
1025         split = start;
1026         key.objectid = ino;
1027         key.type = BTRFS_EXTENT_DATA_KEY;
1028         key.offset = split;
1029
1030         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1031         if (ret < 0)
1032                 goto out;
1033         if (ret > 0 && path->slots[0] > 0)
1034                 path->slots[0]--;
1035
1036         leaf = path->nodes[0];
1037         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1038         BUG_ON(key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY);
1039         fi = btrfs_item_ptr(leaf, path->slots[0],
1040                             struct btrfs_file_extent_item);
1041         BUG_ON(btrfs_file_extent_type(leaf, fi) !=
1042                BTRFS_FILE_EXTENT_PREALLOC);
1043         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1044         BUG_ON(key.offset > start || extent_end < end);
1045
1046         bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1047         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1048         orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1049         memcpy(&new_key, &key, sizeof(new_key));
1050
1051         if (start == key.offset && end < extent_end) {
1052                 other_start = 0;
1053                 other_end = start;
1054                 if (extent_mergeable(leaf, path->slots[0] - 1,
1055                                      ino, bytenr, orig_offset,
1056                                      &other_start, &other_end)) {
1057                         new_key.offset = end;
1058                         btrfs_set_item_key_safe(root, path, &new_key);
1059                         fi = btrfs_item_ptr(leaf, path->slots[0],
1060                                             struct btrfs_file_extent_item);
1061                         btrfs_set_file_extent_generation(leaf, fi,
1062                                                          trans->transid);
1063                         btrfs_set_file_extent_num_bytes(leaf, fi,
1064                                                         extent_end - end);
1065                         btrfs_set_file_extent_offset(leaf, fi,
1066                                                      end - orig_offset);
1067                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1068                                             struct btrfs_file_extent_item);
1069                         btrfs_set_file_extent_generation(leaf, fi,
1070                                                          trans->transid);
1071                         btrfs_set_file_extent_num_bytes(leaf, fi,
1072                                                         end - other_start);
1073                         btrfs_mark_buffer_dirty(leaf);
1074                         goto out;
1075                 }
1076         }
1077
1078         if (start > key.offset && end == extent_end) {
1079                 other_start = end;
1080                 other_end = 0;
1081                 if (extent_mergeable(leaf, path->slots[0] + 1,
1082                                      ino, bytenr, orig_offset,
1083                                      &other_start, &other_end)) {
1084                         fi = btrfs_item_ptr(leaf, path->slots[0],
1085                                             struct btrfs_file_extent_item);
1086                         btrfs_set_file_extent_num_bytes(leaf, fi,
1087                                                         start - key.offset);
1088                         btrfs_set_file_extent_generation(leaf, fi,
1089                                                          trans->transid);
1090                         path->slots[0]++;
1091                         new_key.offset = start;
1092                         btrfs_set_item_key_safe(root, path, &new_key);
1093
1094                         fi = btrfs_item_ptr(leaf, path->slots[0],
1095                                             struct btrfs_file_extent_item);
1096                         btrfs_set_file_extent_generation(leaf, fi,
1097                                                          trans->transid);
1098                         btrfs_set_file_extent_num_bytes(leaf, fi,
1099                                                         other_end - start);
1100                         btrfs_set_file_extent_offset(leaf, fi,
1101                                                      start - orig_offset);
1102                         btrfs_mark_buffer_dirty(leaf);
1103                         goto out;
1104                 }
1105         }
1106
1107         while (start > key.offset || end < extent_end) {
1108                 if (key.offset == start)
1109                         split = end;
1110
1111                 new_key.offset = split;
1112                 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1113                 if (ret == -EAGAIN) {
1114                         btrfs_release_path(path);
1115                         goto again;
1116                 }
1117                 if (ret < 0) {
1118                         btrfs_abort_transaction(trans, root, ret);
1119                         goto out;
1120                 }
1121
1122                 leaf = path->nodes[0];
1123                 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1124                                     struct btrfs_file_extent_item);
1125                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1126                 btrfs_set_file_extent_num_bytes(leaf, fi,
1127                                                 split - key.offset);
1128
1129                 fi = btrfs_item_ptr(leaf, path->slots[0],
1130                                     struct btrfs_file_extent_item);
1131
1132                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1133                 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1134                 btrfs_set_file_extent_num_bytes(leaf, fi,
1135                                                 extent_end - split);
1136                 btrfs_mark_buffer_dirty(leaf);
1137
1138                 ret = btrfs_inc_extent_ref(trans, root, bytenr, num_bytes, 0,
1139                                            root->root_key.objectid,
1140                                            ino, orig_offset, 0);
1141                 BUG_ON(ret); /* -ENOMEM */
1142
1143                 if (split == start) {
1144                         key.offset = start;
1145                 } else {
1146                         BUG_ON(start != key.offset);
1147                         path->slots[0]--;
1148                         extent_end = end;
1149                 }
1150                 recow = 1;
1151         }
1152
1153         other_start = end;
1154         other_end = 0;
1155         if (extent_mergeable(leaf, path->slots[0] + 1,
1156                              ino, bytenr, orig_offset,
1157                              &other_start, &other_end)) {
1158                 if (recow) {
1159                         btrfs_release_path(path);
1160                         goto again;
1161                 }
1162                 extent_end = other_end;
1163                 del_slot = path->slots[0] + 1;
1164                 del_nr++;
1165                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1166                                         0, root->root_key.objectid,
1167                                         ino, orig_offset, 0);
1168                 BUG_ON(ret); /* -ENOMEM */
1169         }
1170         other_start = 0;
1171         other_end = start;
1172         if (extent_mergeable(leaf, path->slots[0] - 1,
1173                              ino, bytenr, orig_offset,
1174                              &other_start, &other_end)) {
1175                 if (recow) {
1176                         btrfs_release_path(path);
1177                         goto again;
1178                 }
1179                 key.offset = other_start;
1180                 del_slot = path->slots[0];
1181                 del_nr++;
1182                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1183                                         0, root->root_key.objectid,
1184                                         ino, orig_offset, 0);
1185                 BUG_ON(ret); /* -ENOMEM */
1186         }
1187         if (del_nr == 0) {
1188                 fi = btrfs_item_ptr(leaf, path->slots[0],
1189                            struct btrfs_file_extent_item);
1190                 btrfs_set_file_extent_type(leaf, fi,
1191                                            BTRFS_FILE_EXTENT_REG);
1192                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1193                 btrfs_mark_buffer_dirty(leaf);
1194         } else {
1195                 fi = btrfs_item_ptr(leaf, del_slot - 1,
1196                            struct btrfs_file_extent_item);
1197                 btrfs_set_file_extent_type(leaf, fi,
1198                                            BTRFS_FILE_EXTENT_REG);
1199                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1200                 btrfs_set_file_extent_num_bytes(leaf, fi,
1201                                                 extent_end - key.offset);
1202                 btrfs_mark_buffer_dirty(leaf);
1203
1204                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1205                 if (ret < 0) {
1206                         btrfs_abort_transaction(trans, root, ret);
1207                         goto out;
1208                 }
1209         }
1210 out:
1211         btrfs_free_path(path);
1212         return 0;
1213 }
1214
1215 /*
1216  * on error we return an unlocked page and the error value
1217  * on success we return a locked page and 0
1218  */
1219 static int prepare_uptodate_page(struct page *page, u64 pos,
1220                                  bool force_uptodate)
1221 {
1222         int ret = 0;
1223
1224         if (((pos & (PAGE_CACHE_SIZE - 1)) || force_uptodate) &&
1225             !PageUptodate(page)) {
1226                 ret = btrfs_readpage(NULL, page);
1227                 if (ret)
1228                         return ret;
1229                 lock_page(page);
1230                 if (!PageUptodate(page)) {
1231                         unlock_page(page);
1232                         return -EIO;
1233                 }
1234         }
1235         return 0;
1236 }
1237
1238 /*
1239  * this gets pages into the page cache and locks them down, it also properly
1240  * waits for data=ordered extents to finish before allowing the pages to be
1241  * modified.
1242  */
1243 static noinline int prepare_pages(struct btrfs_root *root, struct file *file,
1244                          struct page **pages, size_t num_pages,
1245                          loff_t pos, unsigned long first_index,
1246                          size_t write_bytes, bool force_uptodate)
1247 {
1248         struct extent_state *cached_state = NULL;
1249         int i;
1250         unsigned long index = pos >> PAGE_CACHE_SHIFT;
1251         struct inode *inode = file_inode(file);
1252         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1253         int err = 0;
1254         int faili = 0;
1255         u64 start_pos;
1256         u64 last_pos;
1257
1258         start_pos = pos & ~((u64)root->sectorsize - 1);
1259         last_pos = ((u64)index + num_pages) << PAGE_CACHE_SHIFT;
1260
1261 again:
1262         for (i = 0; i < num_pages; i++) {
1263                 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1264                                                mask | __GFP_WRITE);
1265                 if (!pages[i]) {
1266                         faili = i - 1;
1267                         err = -ENOMEM;
1268                         goto fail;
1269                 }
1270
1271                 if (i == 0)
1272                         err = prepare_uptodate_page(pages[i], pos,
1273                                                     force_uptodate);
1274                 if (i == num_pages - 1)
1275                         err = prepare_uptodate_page(pages[i],
1276                                                     pos + write_bytes, false);
1277                 if (err) {
1278                         page_cache_release(pages[i]);
1279                         faili = i - 1;
1280                         goto fail;
1281                 }
1282                 wait_on_page_writeback(pages[i]);
1283         }
1284         err = 0;
1285         if (start_pos < inode->i_size) {
1286                 struct btrfs_ordered_extent *ordered;
1287                 lock_extent_bits(&BTRFS_I(inode)->io_tree,
1288                                  start_pos, last_pos - 1, 0, &cached_state);
1289                 ordered = btrfs_lookup_first_ordered_extent(inode,
1290                                                             last_pos - 1);
1291                 if (ordered &&
1292                     ordered->file_offset + ordered->len > start_pos &&
1293                     ordered->file_offset < last_pos) {
1294                         btrfs_put_ordered_extent(ordered);
1295                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1296                                              start_pos, last_pos - 1,
1297                                              &cached_state, GFP_NOFS);
1298                         for (i = 0; i < num_pages; i++) {
1299                                 unlock_page(pages[i]);
1300                                 page_cache_release(pages[i]);
1301                         }
1302                         btrfs_wait_ordered_range(inode, start_pos,
1303                                                  last_pos - start_pos);
1304                         goto again;
1305                 }
1306                 if (ordered)
1307                         btrfs_put_ordered_extent(ordered);
1308
1309                 clear_extent_bit(&BTRFS_I(inode)->io_tree, start_pos,
1310                                   last_pos - 1, EXTENT_DIRTY | EXTENT_DELALLOC |
1311                                   EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
1312                                   0, 0, &cached_state, GFP_NOFS);
1313                 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1314                                      start_pos, last_pos - 1, &cached_state,
1315                                      GFP_NOFS);
1316         }
1317         for (i = 0; i < num_pages; i++) {
1318                 if (clear_page_dirty_for_io(pages[i]))
1319                         account_page_redirty(pages[i]);
1320                 set_page_extent_mapped(pages[i]);
1321                 WARN_ON(!PageLocked(pages[i]));
1322         }
1323         return 0;
1324 fail:
1325         while (faili >= 0) {
1326                 unlock_page(pages[faili]);
1327                 page_cache_release(pages[faili]);
1328                 faili--;
1329         }
1330         return err;
1331
1332 }
1333
1334 static noinline int check_can_nocow(struct inode *inode, loff_t pos,
1335                                     size_t *write_bytes)
1336 {
1337         struct btrfs_root *root = BTRFS_I(inode)->root;
1338         struct btrfs_ordered_extent *ordered;
1339         u64 lockstart, lockend;
1340         u64 num_bytes;
1341         int ret;
1342
1343         lockstart = round_down(pos, root->sectorsize);
1344         lockend = lockstart + round_up(*write_bytes, root->sectorsize) - 1;
1345
1346         while (1) {
1347                 lock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend);
1348                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
1349                                                      lockend - lockstart + 1);
1350                 if (!ordered) {
1351                         break;
1352                 }
1353                 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend);
1354                 btrfs_start_ordered_extent(inode, ordered, 1);
1355                 btrfs_put_ordered_extent(ordered);
1356         }
1357
1358         num_bytes = lockend - lockstart + 1;
1359         ret = can_nocow_extent(inode, lockstart, &num_bytes, NULL, NULL, NULL);
1360         if (ret <= 0) {
1361                 ret = 0;
1362         } else {
1363                 clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
1364                                  EXTENT_DIRTY | EXTENT_DELALLOC |
1365                                  EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 0, 0,
1366                                  NULL, GFP_NOFS);
1367                 *write_bytes = min_t(size_t, *write_bytes, num_bytes);
1368         }
1369
1370         unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend);
1371
1372         return ret;
1373 }
1374
1375 static noinline ssize_t __btrfs_buffered_write(struct file *file,
1376                                                struct iov_iter *i,
1377                                                loff_t pos)
1378 {
1379         struct inode *inode = file_inode(file);
1380         struct btrfs_root *root = BTRFS_I(inode)->root;
1381         struct page **pages = NULL;
1382         u64 release_bytes = 0;
1383         unsigned long first_index;
1384         size_t num_written = 0;
1385         int nrptrs;
1386         int ret = 0;
1387         bool only_release_metadata = false;
1388         bool force_page_uptodate = false;
1389
1390         nrptrs = min((iov_iter_count(i) + PAGE_CACHE_SIZE - 1) /
1391                      PAGE_CACHE_SIZE, PAGE_CACHE_SIZE /
1392                      (sizeof(struct page *)));
1393         nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1394         nrptrs = max(nrptrs, 8);
1395         pages = kmalloc(nrptrs * sizeof(struct page *), GFP_KERNEL);
1396         if (!pages)
1397                 return -ENOMEM;
1398
1399         first_index = pos >> PAGE_CACHE_SHIFT;
1400
1401         while (iov_iter_count(i) > 0) {
1402                 size_t offset = pos & (PAGE_CACHE_SIZE - 1);
1403                 size_t write_bytes = min(iov_iter_count(i),
1404                                          nrptrs * (size_t)PAGE_CACHE_SIZE -
1405                                          offset);
1406                 size_t num_pages = (write_bytes + offset +
1407                                     PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1408                 size_t reserve_bytes;
1409                 size_t dirty_pages;
1410                 size_t copied;
1411
1412                 WARN_ON(num_pages > nrptrs);
1413
1414                 /*
1415                  * Fault pages before locking them in prepare_pages
1416                  * to avoid recursive lock
1417                  */
1418                 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1419                         ret = -EFAULT;
1420                         break;
1421                 }
1422
1423                 reserve_bytes = num_pages << PAGE_CACHE_SHIFT;
1424                 ret = btrfs_check_data_free_space(inode, reserve_bytes);
1425                 if (ret == -ENOSPC &&
1426                     (BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1427                                               BTRFS_INODE_PREALLOC))) {
1428                         ret = check_can_nocow(inode, pos, &write_bytes);
1429                         if (ret > 0) {
1430                                 only_release_metadata = true;
1431                                 /*
1432                                  * our prealloc extent may be smaller than
1433                                  * write_bytes, so scale down.
1434                                  */
1435                                 num_pages = (write_bytes + offset +
1436                                              PAGE_CACHE_SIZE - 1) >>
1437                                         PAGE_CACHE_SHIFT;
1438                                 reserve_bytes = num_pages << PAGE_CACHE_SHIFT;
1439                                 ret = 0;
1440                         } else {
1441                                 ret = -ENOSPC;
1442                         }
1443                 }
1444
1445                 if (ret)
1446                         break;
1447
1448                 ret = btrfs_delalloc_reserve_metadata(inode, reserve_bytes);
1449                 if (ret) {
1450                         if (!only_release_metadata)
1451                                 btrfs_free_reserved_data_space(inode,
1452                                                                reserve_bytes);
1453                         break;
1454                 }
1455
1456                 release_bytes = reserve_bytes;
1457
1458                 /*
1459                  * This is going to setup the pages array with the number of
1460                  * pages we want, so we don't really need to worry about the
1461                  * contents of pages from loop to loop
1462                  */
1463                 ret = prepare_pages(root, file, pages, num_pages,
1464                                     pos, first_index, write_bytes,
1465                                     force_page_uptodate);
1466                 if (ret)
1467                         break;
1468
1469                 copied = btrfs_copy_from_user(pos, num_pages,
1470                                            write_bytes, pages, i);
1471
1472                 /*
1473                  * if we have trouble faulting in the pages, fall
1474                  * back to one page at a time
1475                  */
1476                 if (copied < write_bytes)
1477                         nrptrs = 1;
1478
1479                 if (copied == 0) {
1480                         force_page_uptodate = true;
1481                         dirty_pages = 0;
1482                 } else {
1483                         force_page_uptodate = false;
1484                         dirty_pages = (copied + offset +
1485                                        PAGE_CACHE_SIZE - 1) >>
1486                                        PAGE_CACHE_SHIFT;
1487                 }
1488
1489                 /*
1490                  * If we had a short copy we need to release the excess delaloc
1491                  * bytes we reserved.  We need to increment outstanding_extents
1492                  * because btrfs_delalloc_release_space will decrement it, but
1493                  * we still have an outstanding extent for the chunk we actually
1494                  * managed to copy.
1495                  */
1496                 if (num_pages > dirty_pages) {
1497                         release_bytes = (num_pages - dirty_pages) <<
1498                                 PAGE_CACHE_SHIFT;
1499                         if (copied > 0) {
1500                                 spin_lock(&BTRFS_I(inode)->lock);
1501                                 BTRFS_I(inode)->outstanding_extents++;
1502                                 spin_unlock(&BTRFS_I(inode)->lock);
1503                         }
1504                         if (only_release_metadata)
1505                                 btrfs_delalloc_release_metadata(inode,
1506                                                                 release_bytes);
1507                         else
1508                                 btrfs_delalloc_release_space(inode,
1509                                                              release_bytes);
1510                 }
1511
1512                 release_bytes = dirty_pages << PAGE_CACHE_SHIFT;
1513                 if (copied > 0) {
1514                         ret = btrfs_dirty_pages(root, inode, pages,
1515                                                 dirty_pages, pos, copied,
1516                                                 NULL);
1517                         if (ret) {
1518                                 btrfs_drop_pages(pages, num_pages);
1519                                 break;
1520                         }
1521                 }
1522
1523                 release_bytes = 0;
1524                 btrfs_drop_pages(pages, num_pages);
1525
1526                 if (only_release_metadata && copied > 0) {
1527                         u64 lockstart = round_down(pos, root->sectorsize);
1528                         u64 lockend = lockstart +
1529                                 (dirty_pages << PAGE_CACHE_SHIFT) - 1;
1530
1531                         set_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
1532                                        lockend, EXTENT_NORESERVE, NULL,
1533                                        NULL, GFP_NOFS);
1534                         only_release_metadata = false;
1535                 }
1536
1537                 cond_resched();
1538
1539                 balance_dirty_pages_ratelimited(inode->i_mapping);
1540                 if (dirty_pages < (root->leafsize >> PAGE_CACHE_SHIFT) + 1)
1541                         btrfs_btree_balance_dirty(root);
1542
1543                 pos += copied;
1544                 num_written += copied;
1545         }
1546
1547         kfree(pages);
1548
1549         if (release_bytes) {
1550                 if (only_release_metadata)
1551                         btrfs_delalloc_release_metadata(inode, release_bytes);
1552                 else
1553                         btrfs_delalloc_release_space(inode, release_bytes);
1554         }
1555
1556         return num_written ? num_written : ret;
1557 }
1558
1559 static ssize_t __btrfs_direct_write(struct kiocb *iocb,
1560                                      struct iov_iter *iter, loff_t pos,
1561                                     loff_t *ppos, size_t count)
1562 {
1563         struct file *file = iocb->ki_filp;
1564         ssize_t written;
1565         ssize_t written_buffered;
1566         loff_t endbyte;
1567         int err;
1568
1569         written = generic_file_direct_write_iter(iocb, iter, pos, ppos, count);
1570
1571         if (written < 0 || written == count)
1572                 return written;
1573
1574         pos += written;
1575         count -= written;
1576         written_buffered = __btrfs_buffered_write(file, iter, pos);
1577         if (written_buffered < 0) {
1578                 err = written_buffered;
1579                 goto out;
1580         }
1581         endbyte = pos + written_buffered - 1;
1582         err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
1583         if (err)
1584                 goto out;
1585         written += written_buffered;
1586         *ppos = pos + written_buffered;
1587         invalidate_mapping_pages(file->f_mapping, pos >> PAGE_CACHE_SHIFT,
1588                                  endbyte >> PAGE_CACHE_SHIFT);
1589 out:
1590         return written ? written : err;
1591 }
1592
1593 static void update_time_for_write(struct inode *inode)
1594 {
1595         struct timespec now;
1596
1597         if (IS_NOCMTIME(inode))
1598                 return;
1599
1600         now = current_fs_time(inode->i_sb);
1601         if (!timespec_equal(&inode->i_mtime, &now))
1602                 inode->i_mtime = now;
1603
1604         if (!timespec_equal(&inode->i_ctime, &now))
1605                 inode->i_ctime = now;
1606
1607         if (IS_I_VERSION(inode))
1608                 inode_inc_iversion(inode);
1609 }
1610
1611 static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
1612                                      struct iov_iter *iter, loff_t pos)
1613 {
1614         struct file *file = iocb->ki_filp;
1615         struct inode *inode = file_inode(file);
1616         struct btrfs_root *root = BTRFS_I(inode)->root;
1617         loff_t *ppos = &iocb->ki_pos;
1618         u64 start_pos;
1619         ssize_t num_written = 0;
1620         ssize_t err = 0;
1621         size_t count;
1622         bool sync = (file->f_flags & O_DSYNC) || IS_SYNC(file->f_mapping->host);
1623
1624         mutex_lock(&inode->i_mutex);
1625
1626         count = iov_iter_count(iter);
1627
1628         current->backing_dev_info = inode->i_mapping->backing_dev_info;
1629         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
1630         if (err) {
1631                 mutex_unlock(&inode->i_mutex);
1632                 goto out;
1633         }
1634
1635         if (count == 0) {
1636                 mutex_unlock(&inode->i_mutex);
1637                 goto out;
1638         }
1639
1640         err = file_remove_suid(file);
1641         if (err) {
1642                 mutex_unlock(&inode->i_mutex);
1643                 goto out;
1644         }
1645
1646         /*
1647          * If BTRFS flips readonly due to some impossible error
1648          * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1649          * although we have opened a file as writable, we have
1650          * to stop this write operation to ensure FS consistency.
1651          */
1652         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state)) {
1653                 mutex_unlock(&inode->i_mutex);
1654                 err = -EROFS;
1655                 goto out;
1656         }
1657
1658         /*
1659          * We reserve space for updating the inode when we reserve space for the
1660          * extent we are going to write, so we will enospc out there.  We don't
1661          * need to start yet another transaction to update the inode as we will
1662          * update the inode when we finish writing whatever data we write.
1663          */
1664         update_time_for_write(inode);
1665
1666         start_pos = round_down(pos, root->sectorsize);
1667         if (start_pos > i_size_read(inode)) {
1668                 err = btrfs_cont_expand(inode, i_size_read(inode), start_pos);
1669                 if (err) {
1670                         mutex_unlock(&inode->i_mutex);
1671                         goto out;
1672                 }
1673         }
1674
1675         if (sync)
1676                 atomic_inc(&BTRFS_I(inode)->sync_writers);
1677
1678         if (unlikely(file->f_flags & O_DIRECT)) {
1679                 num_written = __btrfs_direct_write(iocb, iter, pos, ppos,
1680                                                    count);
1681         } else {
1682                 num_written = __btrfs_buffered_write(file, iter, pos);
1683                 if (num_written > 0)
1684                         *ppos = pos + num_written;
1685         }
1686
1687         mutex_unlock(&inode->i_mutex);
1688
1689         /*
1690          * we want to make sure fsync finds this change
1691          * but we haven't joined a transaction running right now.
1692          *
1693          * Later on, someone is sure to update the inode and get the
1694          * real transid recorded.
1695          *
1696          * We set last_trans now to the fs_info generation + 1,
1697          * this will either be one more than the running transaction
1698          * or the generation used for the next transaction if there isn't
1699          * one running right now.
1700          *
1701          * We also have to set last_sub_trans to the current log transid,
1702          * otherwise subsequent syncs to a file that's been synced in this
1703          * transaction will appear to have already occured.
1704          */
1705         BTRFS_I(inode)->last_trans = root->fs_info->generation + 1;
1706         BTRFS_I(inode)->last_sub_trans = root->log_transid;
1707         if (num_written > 0) {
1708                 err = generic_write_sync(file, pos, num_written);
1709                 if (err < 0 && num_written > 0)
1710                         num_written = err;
1711         }
1712
1713         if (sync)
1714                 atomic_dec(&BTRFS_I(inode)->sync_writers);
1715 out:
1716         current->backing_dev_info = NULL;
1717         return num_written ? num_written : err;
1718 }
1719
1720 int btrfs_release_file(struct inode *inode, struct file *filp)
1721 {
1722         /*
1723          * ordered_data_close is set by settattr when we are about to truncate
1724          * a file from a non-zero size to a zero size.  This tries to
1725          * flush down new bytes that may have been written if the
1726          * application were using truncate to replace a file in place.
1727          */
1728         if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
1729                                &BTRFS_I(inode)->runtime_flags)) {
1730                 struct btrfs_trans_handle *trans;
1731                 struct btrfs_root *root = BTRFS_I(inode)->root;
1732
1733                 /*
1734                  * We need to block on a committing transaction to keep us from
1735                  * throwing a ordered operation on to the list and causing
1736                  * something like sync to deadlock trying to flush out this
1737                  * inode.
1738                  */
1739                 trans = btrfs_start_transaction(root, 0);
1740                 if (IS_ERR(trans))
1741                         return PTR_ERR(trans);
1742                 btrfs_add_ordered_operation(trans, BTRFS_I(inode)->root, inode);
1743                 btrfs_end_transaction(trans, root);
1744                 if (inode->i_size > BTRFS_ORDERED_OPERATIONS_FLUSH_LIMIT)
1745                         filemap_flush(inode->i_mapping);
1746         }
1747         if (filp->private_data)
1748                 btrfs_ioctl_trans_end(filp);
1749         return 0;
1750 }
1751
1752 /*
1753  * fsync call for both files and directories.  This logs the inode into
1754  * the tree log instead of forcing full commits whenever possible.
1755  *
1756  * It needs to call filemap_fdatawait so that all ordered extent updates are
1757  * in the metadata btree are up to date for copying to the log.
1758  *
1759  * It drops the inode mutex before doing the tree log commit.  This is an
1760  * important optimization for directories because holding the mutex prevents
1761  * new operations on the dir while we write to disk.
1762  */
1763 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
1764 {
1765         struct dentry *dentry = file->f_path.dentry;
1766         struct inode *inode = dentry->d_inode;
1767         struct btrfs_root *root = BTRFS_I(inode)->root;
1768         int ret = 0;
1769         struct btrfs_trans_handle *trans;
1770         bool full_sync = 0;
1771
1772         trace_btrfs_sync_file(file, datasync);
1773
1774         /*
1775          * We write the dirty pages in the range and wait until they complete
1776          * out of the ->i_mutex. If so, we can flush the dirty pages by
1777          * multi-task, and make the performance up.  See
1778          * btrfs_wait_ordered_range for an explanation of the ASYNC check.
1779          */
1780         atomic_inc(&BTRFS_I(inode)->sync_writers);
1781         ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
1782         if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1783                              &BTRFS_I(inode)->runtime_flags))
1784                 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
1785         atomic_dec(&BTRFS_I(inode)->sync_writers);
1786         if (ret)
1787                 return ret;
1788
1789         mutex_lock(&inode->i_mutex);
1790
1791         /*
1792          * We flush the dirty pages again to avoid some dirty pages in the
1793          * range being left.
1794          */
1795         atomic_inc(&root->log_batch);
1796         full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1797                              &BTRFS_I(inode)->runtime_flags);
1798         if (full_sync)
1799                 btrfs_wait_ordered_range(inode, start, end - start + 1);
1800         atomic_inc(&root->log_batch);
1801
1802         /*
1803          * check the transaction that last modified this inode
1804          * and see if its already been committed
1805          */
1806         if (!BTRFS_I(inode)->last_trans) {
1807                 mutex_unlock(&inode->i_mutex);
1808                 goto out;
1809         }
1810
1811         /*
1812          * if the last transaction that changed this file was before
1813          * the current transaction, we can bail out now without any
1814          * syncing
1815          */
1816         smp_mb();
1817         if (btrfs_inode_in_log(inode, root->fs_info->generation) ||
1818             BTRFS_I(inode)->last_trans <=
1819             root->fs_info->last_trans_committed) {
1820                 BTRFS_I(inode)->last_trans = 0;
1821
1822                 /*
1823                  * We'v had everything committed since the last time we were
1824                  * modified so clear this flag in case it was set for whatever
1825                  * reason, it's no longer relevant.
1826                  */
1827                 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1828                           &BTRFS_I(inode)->runtime_flags);
1829                 mutex_unlock(&inode->i_mutex);
1830                 goto out;
1831         }
1832
1833         /*
1834          * ok we haven't committed the transaction yet, lets do a commit
1835          */
1836         if (file->private_data)
1837                 btrfs_ioctl_trans_end(file);
1838
1839         trans = btrfs_start_transaction(root, 0);
1840         if (IS_ERR(trans)) {
1841                 ret = PTR_ERR(trans);
1842                 mutex_unlock(&inode->i_mutex);
1843                 goto out;
1844         }
1845
1846         ret = btrfs_log_dentry_safe(trans, root, dentry);
1847         if (ret < 0) {
1848                 /* Fallthrough and commit/free transaction. */
1849                 ret = 1;
1850         }
1851
1852         /* we've logged all the items and now have a consistent
1853          * version of the file in the log.  It is possible that
1854          * someone will come in and modify the file, but that's
1855          * fine because the log is consistent on disk, and we
1856          * have references to all of the file's extents
1857          *
1858          * It is possible that someone will come in and log the
1859          * file again, but that will end up using the synchronization
1860          * inside btrfs_sync_log to keep things safe.
1861          */
1862         mutex_unlock(&inode->i_mutex);
1863
1864         if (ret != BTRFS_NO_LOG_SYNC) {
1865                 if (ret > 0) {
1866                         /*
1867                          * If we didn't already wait for ordered extents we need
1868                          * to do that now.
1869                          */
1870                         if (!full_sync)
1871                                 btrfs_wait_ordered_range(inode, start,
1872                                                          end - start + 1);
1873                         ret = btrfs_commit_transaction(trans, root);
1874                 } else {
1875                         ret = btrfs_sync_log(trans, root);
1876                         if (ret == 0) {
1877                                 ret = btrfs_end_transaction(trans, root);
1878                         } else {
1879                                 if (!full_sync)
1880                                         btrfs_wait_ordered_range(inode, start,
1881                                                                  end -
1882                                                                  start + 1);
1883                                 ret = btrfs_commit_transaction(trans, root);
1884                         }
1885                 }
1886         } else {
1887                 ret = btrfs_end_transaction(trans, root);
1888         }
1889 out:
1890         return ret > 0 ? -EIO : ret;
1891 }
1892
1893 static const struct vm_operations_struct btrfs_file_vm_ops = {
1894         .fault          = filemap_fault,
1895         .page_mkwrite   = btrfs_page_mkwrite,
1896         .remap_pages    = generic_file_remap_pages,
1897 };
1898
1899 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
1900 {
1901         struct address_space *mapping = filp->f_mapping;
1902
1903         if (!mapping->a_ops->readpage)
1904                 return -ENOEXEC;
1905
1906         file_accessed(filp);
1907         vma->vm_ops = &btrfs_file_vm_ops;
1908
1909         return 0;
1910 }
1911
1912 static int hole_mergeable(struct inode *inode, struct extent_buffer *leaf,
1913                           int slot, u64 start, u64 end)
1914 {
1915         struct btrfs_file_extent_item *fi;
1916         struct btrfs_key key;
1917
1918         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1919                 return 0;
1920
1921         btrfs_item_key_to_cpu(leaf, &key, slot);
1922         if (key.objectid != btrfs_ino(inode) ||
1923             key.type != BTRFS_EXTENT_DATA_KEY)
1924                 return 0;
1925
1926         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1927
1928         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
1929                 return 0;
1930
1931         if (btrfs_file_extent_disk_bytenr(leaf, fi))
1932                 return 0;
1933
1934         if (key.offset == end)
1935                 return 1;
1936         if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
1937                 return 1;
1938         return 0;
1939 }
1940
1941 static int fill_holes(struct btrfs_trans_handle *trans, struct inode *inode,
1942                       struct btrfs_path *path, u64 offset, u64 end)
1943 {
1944         struct btrfs_root *root = BTRFS_I(inode)->root;
1945         struct extent_buffer *leaf;
1946         struct btrfs_file_extent_item *fi;
1947         struct extent_map *hole_em;
1948         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
1949         struct btrfs_key key;
1950         int ret;
1951
1952         key.objectid = btrfs_ino(inode);
1953         key.type = BTRFS_EXTENT_DATA_KEY;
1954         key.offset = offset;
1955
1956
1957         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
1958         if (ret < 0)
1959                 return ret;
1960         BUG_ON(!ret);
1961
1962         leaf = path->nodes[0];
1963         if (hole_mergeable(inode, leaf, path->slots[0]-1, offset, end)) {
1964                 u64 num_bytes;
1965
1966                 path->slots[0]--;
1967                 fi = btrfs_item_ptr(leaf, path->slots[0],
1968                                     struct btrfs_file_extent_item);
1969                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
1970                         end - offset;
1971                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
1972                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
1973                 btrfs_set_file_extent_offset(leaf, fi, 0);
1974                 btrfs_mark_buffer_dirty(leaf);
1975                 goto out;
1976         }
1977
1978         if (hole_mergeable(inode, leaf, path->slots[0]+1, offset, end)) {
1979                 u64 num_bytes;
1980
1981                 path->slots[0]++;
1982                 key.offset = offset;
1983                 btrfs_set_item_key_safe(root, path, &key);
1984                 fi = btrfs_item_ptr(leaf, path->slots[0],
1985                                     struct btrfs_file_extent_item);
1986                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
1987                         offset;
1988                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
1989                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
1990                 btrfs_set_file_extent_offset(leaf, fi, 0);
1991                 btrfs_mark_buffer_dirty(leaf);
1992                 goto out;
1993         }
1994         btrfs_release_path(path);
1995
1996         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode), offset,
1997                                        0, 0, end - offset, 0, end - offset,
1998                                        0, 0, 0);
1999         if (ret)
2000                 return ret;
2001
2002 out:
2003         btrfs_release_path(path);
2004
2005         hole_em = alloc_extent_map();
2006         if (!hole_em) {
2007                 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2008                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2009                         &BTRFS_I(inode)->runtime_flags);
2010         } else {
2011                 hole_em->start = offset;
2012                 hole_em->len = end - offset;
2013                 hole_em->ram_bytes = hole_em->len;
2014                 hole_em->orig_start = offset;
2015
2016                 hole_em->block_start = EXTENT_MAP_HOLE;
2017                 hole_em->block_len = 0;
2018                 hole_em->orig_block_len = 0;
2019                 hole_em->bdev = root->fs_info->fs_devices->latest_bdev;
2020                 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2021                 hole_em->generation = trans->transid;
2022
2023                 do {
2024                         btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2025                         write_lock(&em_tree->lock);
2026                         ret = add_extent_mapping(em_tree, hole_em, 1);
2027                         write_unlock(&em_tree->lock);
2028                 } while (ret == -EEXIST);
2029                 free_extent_map(hole_em);
2030                 if (ret)
2031                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2032                                 &BTRFS_I(inode)->runtime_flags);
2033         }
2034
2035         return 0;
2036 }
2037
2038 static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2039 {
2040         struct btrfs_root *root = BTRFS_I(inode)->root;
2041         struct extent_state *cached_state = NULL;
2042         struct btrfs_path *path;
2043         struct btrfs_block_rsv *rsv;
2044         struct btrfs_trans_handle *trans;
2045         u64 lockstart = round_up(offset, BTRFS_I(inode)->root->sectorsize);
2046         u64 lockend = round_down(offset + len,
2047                                  BTRFS_I(inode)->root->sectorsize) - 1;
2048         u64 cur_offset = lockstart;
2049         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
2050         u64 drop_end;
2051         int ret = 0;
2052         int err = 0;
2053         bool same_page = ((offset >> PAGE_CACHE_SHIFT) ==
2054                           ((offset + len - 1) >> PAGE_CACHE_SHIFT));
2055
2056         btrfs_wait_ordered_range(inode, offset, len);
2057
2058         mutex_lock(&inode->i_mutex);
2059         /*
2060          * We needn't truncate any page which is beyond the end of the file
2061          * because we are sure there is no data there.
2062          */
2063         /*
2064          * Only do this if we are in the same page and we aren't doing the
2065          * entire page.
2066          */
2067         if (same_page && len < PAGE_CACHE_SIZE) {
2068                 if (offset < round_up(inode->i_size, PAGE_CACHE_SIZE))
2069                         ret = btrfs_truncate_page(inode, offset, len, 0);
2070                 mutex_unlock(&inode->i_mutex);
2071                 return ret;
2072         }
2073
2074         /* zero back part of the first page */
2075         if (offset < round_up(inode->i_size, PAGE_CACHE_SIZE)) {
2076                 ret = btrfs_truncate_page(inode, offset, 0, 0);
2077                 if (ret) {
2078                         mutex_unlock(&inode->i_mutex);
2079                         return ret;
2080                 }
2081         }
2082
2083         /* zero the front end of the last page */
2084         if (offset + len < round_up(inode->i_size, PAGE_CACHE_SIZE)) {
2085                 ret = btrfs_truncate_page(inode, offset + len, 0, 1);
2086                 if (ret) {
2087                         mutex_unlock(&inode->i_mutex);
2088                         return ret;
2089                 }
2090         }
2091
2092         if (lockend < lockstart) {
2093                 mutex_unlock(&inode->i_mutex);
2094                 return 0;
2095         }
2096
2097         while (1) {
2098                 struct btrfs_ordered_extent *ordered;
2099
2100                 truncate_pagecache_range(inode, lockstart, lockend);
2101
2102                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2103                                  0, &cached_state);
2104                 ordered = btrfs_lookup_first_ordered_extent(inode, lockend);
2105
2106                 /*
2107                  * We need to make sure we have no ordered extents in this range
2108                  * and nobody raced in and read a page in this range, if we did
2109                  * we need to try again.
2110                  */
2111                 if ((!ordered ||
2112                     (ordered->file_offset + ordered->len < lockstart ||
2113                      ordered->file_offset > lockend)) &&
2114                      !test_range_bit(&BTRFS_I(inode)->io_tree, lockstart,
2115                                      lockend, EXTENT_UPTODATE, 0,
2116                                      cached_state)) {
2117                         if (ordered)
2118                                 btrfs_put_ordered_extent(ordered);
2119                         break;
2120                 }
2121                 if (ordered)
2122                         btrfs_put_ordered_extent(ordered);
2123                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2124                                      lockend, &cached_state, GFP_NOFS);
2125                 btrfs_wait_ordered_range(inode, lockstart,
2126                                          lockend - lockstart + 1);
2127         }
2128
2129         path = btrfs_alloc_path();
2130         if (!path) {
2131                 ret = -ENOMEM;
2132                 goto out;
2133         }
2134
2135         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
2136         if (!rsv) {
2137                 ret = -ENOMEM;
2138                 goto out_free;
2139         }
2140         rsv->size = btrfs_calc_trunc_metadata_size(root, 1);
2141         rsv->failfast = 1;
2142
2143         /*
2144          * 1 - update the inode
2145          * 1 - removing the extents in the range
2146          * 1 - adding the hole extent
2147          */
2148         trans = btrfs_start_transaction(root, 3);
2149         if (IS_ERR(trans)) {
2150                 err = PTR_ERR(trans);
2151                 goto out_free;
2152         }
2153
2154         ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv, rsv,
2155                                       min_size);
2156         BUG_ON(ret);
2157         trans->block_rsv = rsv;
2158
2159         while (cur_offset < lockend) {
2160                 ret = __btrfs_drop_extents(trans, root, inode, path,
2161                                            cur_offset, lockend + 1,
2162                                            &drop_end, 1);
2163                 if (ret != -ENOSPC)
2164                         break;
2165
2166                 trans->block_rsv = &root->fs_info->trans_block_rsv;
2167
2168                 ret = fill_holes(trans, inode, path, cur_offset, drop_end);
2169                 if (ret) {
2170                         err = ret;
2171                         break;
2172                 }
2173
2174                 cur_offset = drop_end;
2175
2176                 ret = btrfs_update_inode(trans, root, inode);
2177                 if (ret) {
2178                         err = ret;
2179                         break;
2180                 }
2181
2182                 btrfs_end_transaction(trans, root);
2183                 btrfs_btree_balance_dirty(root);
2184
2185                 trans = btrfs_start_transaction(root, 3);
2186                 if (IS_ERR(trans)) {
2187                         ret = PTR_ERR(trans);
2188                         trans = NULL;
2189                         break;
2190                 }
2191
2192                 ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv,
2193                                               rsv, min_size);
2194                 BUG_ON(ret);    /* shouldn't happen */
2195                 trans->block_rsv = rsv;
2196         }
2197
2198         if (ret) {
2199                 err = ret;
2200                 goto out_trans;
2201         }
2202
2203         trans->block_rsv = &root->fs_info->trans_block_rsv;
2204         ret = fill_holes(trans, inode, path, cur_offset, drop_end);
2205         if (ret) {
2206                 err = ret;
2207                 goto out_trans;
2208         }
2209
2210 out_trans:
2211         if (!trans)
2212                 goto out_free;
2213
2214         inode_inc_iversion(inode);
2215         inode->i_mtime = inode->i_ctime = CURRENT_TIME;
2216
2217         trans->block_rsv = &root->fs_info->trans_block_rsv;
2218         ret = btrfs_update_inode(trans, root, inode);
2219         btrfs_end_transaction(trans, root);
2220         btrfs_btree_balance_dirty(root);
2221 out_free:
2222         btrfs_free_path(path);
2223         btrfs_free_block_rsv(root, rsv);
2224 out:
2225         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2226                              &cached_state, GFP_NOFS);
2227         mutex_unlock(&inode->i_mutex);
2228         if (ret && !err)
2229                 err = ret;
2230         return err;
2231 }
2232
2233 static long btrfs_fallocate(struct file *file, int mode,
2234                             loff_t offset, loff_t len)
2235 {
2236         struct inode *inode = file_inode(file);
2237         struct extent_state *cached_state = NULL;
2238         struct btrfs_root *root = BTRFS_I(inode)->root;
2239         u64 cur_offset;
2240         u64 last_byte;
2241         u64 alloc_start;
2242         u64 alloc_end;
2243         u64 alloc_hint = 0;
2244         u64 locked_end;
2245         struct extent_map *em;
2246         int blocksize = BTRFS_I(inode)->root->sectorsize;
2247         int ret;
2248
2249         alloc_start = round_down(offset, blocksize);
2250         alloc_end = round_up(offset + len, blocksize);
2251
2252         /* Make sure we aren't being give some crap mode */
2253         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2254                 return -EOPNOTSUPP;
2255
2256         if (mode & FALLOC_FL_PUNCH_HOLE)
2257                 return btrfs_punch_hole(inode, offset, len);
2258
2259         /*
2260          * Make sure we have enough space before we do the
2261          * allocation.
2262          */
2263         ret = btrfs_check_data_free_space(inode, alloc_end - alloc_start);
2264         if (ret)
2265                 return ret;
2266         if (root->fs_info->quota_enabled) {
2267                 ret = btrfs_qgroup_reserve(root, alloc_end - alloc_start);
2268                 if (ret)
2269                         goto out_reserve_fail;
2270         }
2271
2272         mutex_lock(&inode->i_mutex);
2273         ret = inode_newsize_ok(inode, alloc_end);
2274         if (ret)
2275                 goto out;
2276
2277         if (alloc_start > inode->i_size) {
2278                 ret = btrfs_cont_expand(inode, i_size_read(inode),
2279                                         alloc_start);
2280                 if (ret)
2281                         goto out;
2282         } else {
2283                 /*
2284                  * If we are fallocating from the end of the file onward we
2285                  * need to zero out the end of the page if i_size lands in the
2286                  * middle of a page.
2287                  */
2288                 ret = btrfs_truncate_page(inode, inode->i_size, 0, 0);
2289                 if (ret)
2290                         goto out;
2291         }
2292
2293         /*
2294          * wait for ordered IO before we have any locks.  We'll loop again
2295          * below with the locks held.
2296          */
2297         btrfs_wait_ordered_range(inode, alloc_start, alloc_end - alloc_start);
2298
2299         locked_end = alloc_end - 1;
2300         while (1) {
2301                 struct btrfs_ordered_extent *ordered;
2302
2303                 /* the extent lock is ordered inside the running
2304                  * transaction
2305                  */
2306                 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
2307                                  locked_end, 0, &cached_state);
2308                 ordered = btrfs_lookup_first_ordered_extent(inode,
2309                                                             alloc_end - 1);
2310                 if (ordered &&
2311                     ordered->file_offset + ordered->len > alloc_start &&
2312                     ordered->file_offset < alloc_end) {
2313                         btrfs_put_ordered_extent(ordered);
2314                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
2315                                              alloc_start, locked_end,
2316                                              &cached_state, GFP_NOFS);
2317                         /*
2318                          * we can't wait on the range with the transaction
2319                          * running or with the extent lock held
2320                          */
2321                         btrfs_wait_ordered_range(inode, alloc_start,
2322                                                  alloc_end - alloc_start);
2323                 } else {
2324                         if (ordered)
2325                                 btrfs_put_ordered_extent(ordered);
2326                         break;
2327                 }
2328         }
2329
2330         cur_offset = alloc_start;
2331         while (1) {
2332                 u64 actual_end;
2333
2334                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
2335                                       alloc_end - cur_offset, 0);
2336                 if (IS_ERR_OR_NULL(em)) {
2337                         if (!em)
2338                                 ret = -ENOMEM;
2339                         else
2340                                 ret = PTR_ERR(em);
2341                         break;
2342                 }
2343                 last_byte = min(extent_map_end(em), alloc_end);
2344                 actual_end = min_t(u64, extent_map_end(em), offset + len);
2345                 last_byte = ALIGN(last_byte, blocksize);
2346
2347                 if (em->block_start == EXTENT_MAP_HOLE ||
2348                     (cur_offset >= inode->i_size &&
2349                      !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
2350                         ret = btrfs_prealloc_file_range(inode, mode, cur_offset,
2351                                                         last_byte - cur_offset,
2352                                                         1 << inode->i_blkbits,
2353                                                         offset + len,
2354                                                         &alloc_hint);
2355
2356                         if (ret < 0) {
2357                                 free_extent_map(em);
2358                                 break;
2359                         }
2360                 } else if (actual_end > inode->i_size &&
2361                            !(mode & FALLOC_FL_KEEP_SIZE)) {
2362                         /*
2363                          * We didn't need to allocate any more space, but we
2364                          * still extended the size of the file so we need to
2365                          * update i_size.
2366                          */
2367                         inode->i_ctime = CURRENT_TIME;
2368                         i_size_write(inode, actual_end);
2369                         btrfs_ordered_update_i_size(inode, actual_end, NULL);
2370                 }
2371                 free_extent_map(em);
2372
2373                 cur_offset = last_byte;
2374                 if (cur_offset >= alloc_end) {
2375                         ret = 0;
2376                         break;
2377                 }
2378         }
2379         unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
2380                              &cached_state, GFP_NOFS);
2381 out:
2382         mutex_unlock(&inode->i_mutex);
2383         if (root->fs_info->quota_enabled)
2384                 btrfs_qgroup_free(root, alloc_end - alloc_start);
2385 out_reserve_fail:
2386         /* Let go of our reservation. */
2387         btrfs_free_reserved_data_space(inode, alloc_end - alloc_start);
2388         return ret;
2389 }
2390
2391 static int find_desired_extent(struct inode *inode, loff_t *offset, int whence)
2392 {
2393         struct btrfs_root *root = BTRFS_I(inode)->root;
2394         struct extent_map *em;
2395         struct extent_state *cached_state = NULL;
2396         u64 lockstart = *offset;
2397         u64 lockend = i_size_read(inode);
2398         u64 start = *offset;
2399         u64 orig_start = *offset;
2400         u64 len = i_size_read(inode);
2401         u64 last_end = 0;
2402         int ret = 0;
2403
2404         lockend = max_t(u64, root->sectorsize, lockend);
2405         if (lockend <= lockstart)
2406                 lockend = lockstart + root->sectorsize;
2407
2408         lockend--;
2409         len = lockend - lockstart + 1;
2410
2411         len = max_t(u64, len, root->sectorsize);
2412         if (inode->i_size == 0)
2413                 return -ENXIO;
2414
2415         lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend, 0,
2416                          &cached_state);
2417
2418         /*
2419          * Delalloc is such a pain.  If we have a hole and we have pending
2420          * delalloc for a portion of the hole we will get back a hole that
2421          * exists for the entire range since it hasn't been actually written
2422          * yet.  So to take care of this case we need to look for an extent just
2423          * before the position we want in case there is outstanding delalloc
2424          * going on here.
2425          */
2426         if (whence == SEEK_HOLE && start != 0) {
2427                 if (start <= root->sectorsize)
2428                         em = btrfs_get_extent_fiemap(inode, NULL, 0, 0,
2429                                                      root->sectorsize, 0);
2430                 else
2431                         em = btrfs_get_extent_fiemap(inode, NULL, 0,
2432                                                      start - root->sectorsize,
2433                                                      root->sectorsize, 0);
2434                 if (IS_ERR(em)) {
2435                         ret = PTR_ERR(em);
2436                         goto out;
2437                 }
2438                 last_end = em->start + em->len;
2439                 if (em->block_start == EXTENT_MAP_DELALLOC)
2440                         last_end = min_t(u64, last_end, inode->i_size);
2441                 free_extent_map(em);
2442         }
2443
2444         while (1) {
2445                 em = btrfs_get_extent_fiemap(inode, NULL, 0, start, len, 0);
2446                 if (IS_ERR(em)) {
2447                         ret = PTR_ERR(em);
2448                         break;
2449                 }
2450
2451                 if (em->block_start == EXTENT_MAP_HOLE) {
2452                         if (test_bit(EXTENT_FLAG_VACANCY, &em->flags)) {
2453                                 if (last_end <= orig_start) {
2454                                         free_extent_map(em);
2455                                         ret = -ENXIO;
2456                                         break;
2457                                 }
2458                         }
2459
2460                         if (whence == SEEK_HOLE) {
2461                                 *offset = start;
2462                                 free_extent_map(em);
2463                                 break;
2464                         }
2465                 } else {
2466                         if (whence == SEEK_DATA) {
2467                                 if (em->block_start == EXTENT_MAP_DELALLOC) {
2468                                         if (start >= inode->i_size) {
2469                                                 free_extent_map(em);
2470                                                 ret = -ENXIO;
2471                                                 break;
2472                                         }
2473                                 }
2474
2475                                 if (!test_bit(EXTENT_FLAG_PREALLOC,
2476                                               &em->flags)) {
2477                                         *offset = start;
2478                                         free_extent_map(em);
2479                                         break;
2480                                 }
2481                         }
2482                 }
2483
2484                 start = em->start + em->len;
2485                 last_end = em->start + em->len;
2486
2487                 if (em->block_start == EXTENT_MAP_DELALLOC)
2488                         last_end = min_t(u64, last_end, inode->i_size);
2489
2490                 if (test_bit(EXTENT_FLAG_VACANCY, &em->flags)) {
2491                         free_extent_map(em);
2492                         ret = -ENXIO;
2493                         break;
2494                 }
2495                 free_extent_map(em);
2496                 cond_resched();
2497         }
2498         if (!ret)
2499                 *offset = min(*offset, inode->i_size);
2500 out:
2501         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2502                              &cached_state, GFP_NOFS);
2503         return ret;
2504 }
2505
2506 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
2507 {
2508         struct inode *inode = file->f_mapping->host;
2509         int ret;
2510
2511         mutex_lock(&inode->i_mutex);
2512         switch (whence) {
2513         case SEEK_END:
2514         case SEEK_CUR:
2515                 offset = generic_file_llseek(file, offset, whence);
2516                 goto out;
2517         case SEEK_DATA:
2518         case SEEK_HOLE:
2519                 if (offset >= i_size_read(inode)) {
2520                         mutex_unlock(&inode->i_mutex);
2521                         return -ENXIO;
2522                 }
2523
2524                 ret = find_desired_extent(inode, &offset, whence);
2525                 if (ret) {
2526                         mutex_unlock(&inode->i_mutex);
2527                         return ret;
2528                 }
2529         }
2530
2531         offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
2532 out:
2533         mutex_unlock(&inode->i_mutex);
2534         return offset;
2535 }
2536
2537 const struct file_operations btrfs_file_operations = {
2538         .llseek         = btrfs_file_llseek,
2539         .read           = do_sync_read,
2540         .write          = do_sync_write,
2541         .splice_read    = generic_file_splice_read,
2542         .read_iter      = generic_file_read_iter,
2543         .write_iter     = btrfs_file_write_iter,
2544         .mmap           = btrfs_file_mmap,
2545         .open           = generic_file_open,
2546         .release        = btrfs_release_file,
2547         .fsync          = btrfs_sync_file,
2548         .fallocate      = btrfs_fallocate,
2549         .unlocked_ioctl = btrfs_ioctl,
2550 #ifdef CONFIG_COMPAT
2551         .compat_ioctl   = btrfs_ioctl,
2552 #endif
2553 };
2554
2555 void btrfs_auto_defrag_exit(void)
2556 {
2557         if (btrfs_inode_defrag_cachep)
2558                 kmem_cache_destroy(btrfs_inode_defrag_cachep);
2559 }
2560
2561 int btrfs_auto_defrag_init(void)
2562 {
2563         btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
2564                                         sizeof(struct inode_defrag), 0,
2565                                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD,
2566                                         NULL);
2567         if (!btrfs_inode_defrag_cachep)
2568                 return -ENOMEM;
2569
2570         return 0;
2571 }