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/falloc.h>
  28 #include <linux/swap.h>
  29 #include <linux/writeback.h>
  30 #include <linux/compat.h>
  31 #include <linux/slab.h>
  32 #include <linux/btrfs.h>
  33 #include <linux/uio.h>
  34 #include "ctree.h"
  35 #include "disk-io.h"
  36 #include "transaction.h"
  37 #include "btrfs_inode.h"
  38 #include "print-tree.h"
  39 #include "tree-log.h"
  40 #include "locking.h"
  41 #include "volumes.h"
  42 #include "qgroup.h"
  43 #include "compression.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 btrfs_inode *inode,
  96                                     struct inode_defrag *defrag)
  97 {
  98         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
  99         struct inode_defrag *entry;
 100         struct rb_node **p;
 101         struct rb_node *parent = NULL;
 102         int ret;
 103
 104         p = &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, &inode->runtime_flags);
 127         rb_link_node(&defrag->rb_node, parent, p);
 128         rb_insert_color(&defrag->rb_node, &fs_info->defrag_inodes);
 129         return 0;
 130 }
 131
 132 static inline int __need_auto_defrag(struct btrfs_fs_info *fs_info)
 133 {
 134         if (!btrfs_test_opt(fs_info, AUTO_DEFRAG))
 135                 return 0;
 136
 137         if (btrfs_fs_closing(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 btrfs_inode *inode)
 149 {
 150         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
 151         struct btrfs_root *root = inode->root;
 152         struct inode_defrag *defrag;
 153         u64 transid;
 154         int ret;
 155
 156         if (!__need_auto_defrag(fs_info))
 157                 return 0;
 158
 159         if (test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags))
 160                 return 0;
 161
 162         if (trans)
 163                 transid = trans->transid;
 164         else
 165                 transid = inode->root->last_trans;
 166
 167         defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
 168         if (!defrag)
 169                 return -ENOMEM;
 170
 171         defrag->ino = btrfs_ino(inode);
 172         defrag->transid = transid;
 173         defrag->root = root->root_key.objectid;
 174
 175         spin_lock(&fs_info->defrag_inodes_lock);
 176         if (!test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags)) {
 177                 /*
 178                  * If we set IN_DEFRAG flag and evict the inode from memory,
 179                  * and then re-read this inode, this new inode doesn't have
 180                  * IN_DEFRAG flag. At the case, we may find the existed defrag.
 181                  */
 182                 ret = __btrfs_add_inode_defrag(inode, defrag);
 183                 if (ret)
 184                         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 185         } else {
 186                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 187         }
 188         spin_unlock(&fs_info->defrag_inodes_lock);
 189         return 0;
 190 }
 191
 192 /*
 193  * Requeue the defrag object. If there is a defrag object that points to
 194  * the same inode in the tree, we will merge them together (by
 195  * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
 196  */
 197 static void btrfs_requeue_inode_defrag(struct btrfs_inode *inode,
 198                                        struct inode_defrag *defrag)
 199 {
 200         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
 201         int ret;
 202
 203         if (!__need_auto_defrag(fs_info))
 204                 goto out;
 205
 206         /*
 207          * Here we don't check the IN_DEFRAG flag, because we need merge
 208          * them together.
 209          */
 210         spin_lock(&fs_info->defrag_inodes_lock);
 211         ret = __btrfs_add_inode_defrag(inode, defrag);
 212         spin_unlock(&fs_info->defrag_inodes_lock);
 213         if (ret)
 214                 goto out;
 215         return;
 216 out:
 217         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 218 }
 219
 220 /*
 221  * pick the defragable inode that we want, if it doesn't exist, we will get
 222  * the next one.
 223  */
 224 static struct inode_defrag *
 225 btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
 226 {
 227         struct inode_defrag *entry = NULL;
 228         struct inode_defrag tmp;
 229         struct rb_node *p;
 230         struct rb_node *parent = NULL;
 231         int ret;
 232
 233         tmp.ino = ino;
 234         tmp.root = root;
 235
 236         spin_lock(&fs_info->defrag_inodes_lock);
 237         p = fs_info->defrag_inodes.rb_node;
 238         while (p) {
 239                 parent = p;
 240                 entry = rb_entry(parent, struct inode_defrag, rb_node);
 241
 242                 ret = __compare_inode_defrag(&tmp, entry);
 243                 if (ret < 0)
 244                         p = parent->rb_left;
 245                 else if (ret > 0)
 246                         p = parent->rb_right;
 247                 else
 248                         goto out;
 249         }
 250
 251         if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
 252                 parent = rb_next(parent);
 253                 if (parent)
 254                         entry = rb_entry(parent, struct inode_defrag, rb_node);
 255                 else
 256                         entry = NULL;
 257         }
 258 out:
 259         if (entry)
 260                 rb_erase(parent, &fs_info->defrag_inodes);
 261         spin_unlock(&fs_info->defrag_inodes_lock);
 262         return entry;
 263 }
 264
 265 void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
 266 {
 267         struct inode_defrag *defrag;
 268         struct rb_node *node;
 269
 270         spin_lock(&fs_info->defrag_inodes_lock);
 271         node = rb_first(&fs_info->defrag_inodes);
 272         while (node) {
 273                 rb_erase(node, &fs_info->defrag_inodes);
 274                 defrag = rb_entry(node, struct inode_defrag, rb_node);
 275                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 276
 277                 cond_resched_lock(&fs_info->defrag_inodes_lock);
 278
 279                 node = rb_first(&fs_info->defrag_inodes);
 280         }
 281         spin_unlock(&fs_info->defrag_inodes_lock);
 282 }
 283
 284 #define BTRFS_DEFRAG_BATCH      1024
 285
 286 static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
 287                                     struct inode_defrag *defrag)
 288 {
 289         struct btrfs_root *inode_root;
 290         struct inode *inode;
 291         struct btrfs_key key;
 292         struct btrfs_ioctl_defrag_range_args range;
 293         int num_defrag;
 294         int index;
 295         int ret;
 296
 297         /* get the inode */
 298         key.objectid = defrag->root;
 299         key.type = BTRFS_ROOT_ITEM_KEY;
 300         key.offset = (u64)-1;
 301
 302         index = srcu_read_lock(&fs_info->subvol_srcu);
 303
 304         inode_root = btrfs_read_fs_root_no_name(fs_info, &key);
 305         if (IS_ERR(inode_root)) {
 306                 ret = PTR_ERR(inode_root);
 307                 goto cleanup;
 308         }
 309
 310         key.objectid = defrag->ino;
 311         key.type = BTRFS_INODE_ITEM_KEY;
 312         key.offset = 0;
 313         inode = btrfs_iget(fs_info->sb, &key, inode_root, NULL);
 314         if (IS_ERR(inode)) {
 315                 ret = PTR_ERR(inode);
 316                 goto cleanup;
 317         }
 318         srcu_read_unlock(&fs_info->subvol_srcu, index);
 319
 320         /* do a chunk of defrag */
 321         clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
 322         memset(&range, 0, sizeof(range));
 323         range.len = (u64)-1;
 324         range.start = defrag->last_offset;
 325
 326         sb_start_write(fs_info->sb);
 327         num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
 328                                        BTRFS_DEFRAG_BATCH);
 329         sb_end_write(fs_info->sb);
 330         /*
 331          * if we filled the whole defrag batch, there
 332          * must be more work to do.  Queue this defrag
 333          * again
 334          */
 335         if (num_defrag == BTRFS_DEFRAG_BATCH) {
 336                 defrag->last_offset = range.start;
 337                 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
 338         } else if (defrag->last_offset && !defrag->cycled) {
 339                 /*
 340                  * we didn't fill our defrag batch, but
 341                  * we didn't start at zero.  Make sure we loop
 342                  * around to the start of the file.
 343                  */
 344                 defrag->last_offset = 0;
 345                 defrag->cycled = 1;
 346                 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
 347         } else {
 348                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 349         }
 350
 351         iput(inode);
 352         return 0;
 353 cleanup:
 354         srcu_read_unlock(&fs_info->subvol_srcu, index);
 355         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 356         return ret;
 357 }
 358
 359 /*
 360  * run through the list of inodes in the FS that need
 361  * defragging
 362  */
 363 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
 364 {
 365         struct inode_defrag *defrag;
 366         u64 first_ino = 0;
 367         u64 root_objectid = 0;
 368
 369         atomic_inc(&fs_info->defrag_running);
 370         while (1) {
 371                 /* Pause the auto defragger. */
 372                 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
 373                              &fs_info->fs_state))
 374                         break;
 375
 376                 if (!__need_auto_defrag(fs_info))
 377                         break;
 378
 379                 /* find an inode to defrag */
 380                 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
 381                                                  first_ino);
 382                 if (!defrag) {
 383                         if (root_objectid || first_ino) {
 384                                 root_objectid = 0;
 385                                 first_ino = 0;
 386                                 continue;
 387                         } else {
 388                                 break;
 389                         }
 390                 }
 391
 392                 first_ino = defrag->ino + 1;
 393                 root_objectid = defrag->root;
 394
 395                 __btrfs_run_defrag_inode(fs_info, defrag);
 396         }
 397         atomic_dec(&fs_info->defrag_running);
 398
 399         /*
 400          * during unmount, we use the transaction_wait queue to
 401          * wait for the defragger to stop
 402          */
 403         wake_up(&fs_info->transaction_wait);
 404         return 0;
 405 }
 406
 407 /* simple helper to fault in pages and copy.  This should go away
 408  * and be replaced with calls into generic code.
 409  */
 410 static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
 411                                          struct page **prepared_pages,
 412                                          struct iov_iter *i)
 413 {
 414         size_t copied = 0;
 415         size_t total_copied = 0;
 416         int pg = 0;
 417         int offset = pos & (PAGE_SIZE - 1);
 418
 419         while (write_bytes > 0) {
 420                 size_t count = min_t(size_t,
 421                                      PAGE_SIZE - offset, write_bytes);
 422                 struct page *page = prepared_pages[pg];
 423                 /*
 424                  * Copy data from userspace to the current page
 425                  */
 426                 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
 427
 428                 /* Flush processor's dcache for this page */
 429                 flush_dcache_page(page);
 430
 431                 /*
 432                  * if we get a partial write, we can end up with
 433                  * partially up to date pages.  These add
 434                  * a lot of complexity, so make sure they don't
 435                  * happen by forcing this copy to be retried.
 436                  *
 437                  * The rest of the btrfs_file_write code will fall
 438                  * back to page at a time copies after we return 0.
 439                  */
 440                 if (!PageUptodate(page) && copied < count)
 441                         copied = 0;
 442
 443                 iov_iter_advance(i, copied);
 444                 write_bytes -= copied;
 445                 total_copied += copied;
 446
 447                 /* Return to btrfs_file_write_iter to fault page */
 448                 if (unlikely(copied == 0))
 449                         break;
 450
 451                 if (copied < PAGE_SIZE - offset) {
 452                         offset += copied;
 453                 } else {
 454                         pg++;
 455                         offset = 0;
 456                 }
 457         }
 458         return total_copied;
 459 }
 460
 461 /*
 462  * unlocks pages after btrfs_file_write is done with them
 463  */
 464 static void btrfs_drop_pages(struct page **pages, size_t num_pages)
 465 {
 466         size_t i;
 467         for (i = 0; i < num_pages; i++) {
 468                 /* page checked is some magic around finding pages that
 469                  * have been modified without going through btrfs_set_page_dirty
 470                  * clear it here. There should be no need to mark the pages
 471                  * accessed as prepare_pages should have marked them accessed
 472                  * in prepare_pages via find_or_create_page()
 473                  */
 474                 ClearPageChecked(pages[i]);
 475                 unlock_page(pages[i]);
 476                 put_page(pages[i]);
 477         }
 478 }
 479
 480 /*
 481  * after copy_from_user, pages need to be dirtied and we need to make
 482  * sure holes are created between the current EOF and the start of
 483  * any next extents (if required).
 484  *
 485  * this also makes the decision about creating an inline extent vs
 486  * doing real data extents, marking pages dirty and delalloc as required.
 487  */
 488 int btrfs_dirty_pages(struct inode *inode, struct page **pages,
 489                       size_t num_pages, loff_t pos, size_t write_bytes,
 490                       struct extent_state **cached)
 491 {
 492         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 493         int err = 0;
 494         int i;
 495         u64 num_bytes;
 496         u64 start_pos;
 497         u64 end_of_last_block;
 498         u64 end_pos = pos + write_bytes;
 499         loff_t isize = i_size_read(inode);
 500
 501         start_pos = pos & ~((u64) fs_info->sectorsize - 1);
 502         num_bytes = round_up(write_bytes + pos - start_pos,
 503                              fs_info->sectorsize);
 504
 505         end_of_last_block = start_pos + num_bytes - 1;
 506         err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
 507                                         cached, 0);
 508         if (err)
 509                 return err;
 510
 511         for (i = 0; i < num_pages; i++) {
 512                 struct page *p = pages[i];
 513                 SetPageUptodate(p);
 514                 ClearPageChecked(p);
 515                 set_page_dirty(p);
 516         }
 517
 518         /*
 519          * we've only changed i_size in ram, and we haven't updated
 520          * the disk i_size.  There is no need to log the inode
 521          * at this time.
 522          */
 523         if (end_pos > isize)
 524                 i_size_write(inode, end_pos);
 525         return 0;
 526 }
 527
 528 /*
 529  * this drops all the extents in the cache that intersect the range
 530  * [start, end].  Existing extents are split as required.
 531  */
 532 void btrfs_drop_extent_cache(struct btrfs_inode *inode, u64 start, u64 end,
 533                              int skip_pinned)
 534 {
 535         struct extent_map *em;
 536         struct extent_map *split = NULL;
 537         struct extent_map *split2 = NULL;
 538         struct extent_map_tree *em_tree = &inode->extent_tree;
 539         u64 len = end - start + 1;
 540         u64 gen;
 541         int ret;
 542         int testend = 1;
 543         unsigned long flags;
 544         int compressed = 0;
 545         bool modified;
 546
 547         WARN_ON(end < start);
 548         if (end == (u64)-1) {
 549                 len = (u64)-1;
 550                 testend = 0;
 551         }
 552         while (1) {
 553                 int no_splits = 0;
 554
 555                 modified = false;
 556                 if (!split)
 557                         split = alloc_extent_map();
 558                 if (!split2)
 559                         split2 = alloc_extent_map();
 560                 if (!split || !split2)
 561                         no_splits = 1;
 562
 563                 write_lock(&em_tree->lock);
 564                 em = lookup_extent_mapping(em_tree, start, len);
 565                 if (!em) {
 566                         write_unlock(&em_tree->lock);
 567                         break;
 568                 }
 569                 flags = em->flags;
 570                 gen = em->generation;
 571                 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
 572                         if (testend && em->start + em->len >= start + len) {
 573                                 free_extent_map(em);
 574                                 write_unlock(&em_tree->lock);
 575                                 break;
 576                         }
 577                         start = em->start + em->len;
 578                         if (testend)
 579                                 len = start + len - (em->start + em->len);
 580                         free_extent_map(em);
 581                         write_unlock(&em_tree->lock);
 582                         continue;
 583                 }
 584                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 585                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
 586                 clear_bit(EXTENT_FLAG_LOGGING, &flags);
 587                 modified = !list_empty(&em->list);
 588                 if (no_splits)
 589                         goto next;
 590
 591                 if (em->start < start) {
 592                         split->start = em->start;
 593                         split->len = start - em->start;
 594
 595                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 596                                 split->orig_start = em->orig_start;
 597                                 split->block_start = em->block_start;
 598
 599                                 if (compressed)
 600                                         split->block_len = em->block_len;
 601                                 else
 602                                         split->block_len = split->len;
 603                                 split->orig_block_len = max(split->block_len,
 604                                                 em->orig_block_len);
 605                                 split->ram_bytes = em->ram_bytes;
 606                         } else {
 607                                 split->orig_start = split->start;
 608                                 split->block_len = 0;
 609                                 split->block_start = em->block_start;
 610                                 split->orig_block_len = 0;
 611                                 split->ram_bytes = split->len;
 612                         }
 613
 614                         split->generation = gen;
 615                         split->bdev = em->bdev;
 616                         split->flags = flags;
 617                         split->compress_type = em->compress_type;
 618                         replace_extent_mapping(em_tree, em, split, modified);
 619                         free_extent_map(split);
 620                         split = split2;
 621                         split2 = NULL;
 622                 }
 623                 if (testend && em->start + em->len > start + len) {
 624                         u64 diff = start + len - em->start;
 625
 626                         split->start = start + len;
 627                         split->len = em->start + em->len - (start + len);
 628                         split->bdev = em->bdev;
 629                         split->flags = flags;
 630                         split->compress_type = em->compress_type;
 631                         split->generation = gen;
 632
 633                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 634                                 split->orig_block_len = max(em->block_len,
 635                                                     em->orig_block_len);
 636
 637                                 split->ram_bytes = em->ram_bytes;
 638                                 if (compressed) {
 639                                         split->block_len = em->block_len;
 640                                         split->block_start = em->block_start;
 641                                         split->orig_start = em->orig_start;
 642                                 } else {
 643                                         split->block_len = split->len;
 644                                         split->block_start = em->block_start
 645                                                 + diff;
 646                                         split->orig_start = em->orig_start;
 647                                 }
 648                         } else {
 649                                 split->ram_bytes = split->len;
 650                                 split->orig_start = split->start;
 651                                 split->block_len = 0;
 652                                 split->block_start = em->block_start;
 653                                 split->orig_block_len = 0;
 654                         }
 655
 656                         if (extent_map_in_tree(em)) {
 657                                 replace_extent_mapping(em_tree, em, split,
 658                                                        modified);
 659                         } else {
 660                                 ret = add_extent_mapping(em_tree, split,
 661                                                          modified);
 662                                 ASSERT(ret == 0); /* Logic error */
 663                         }
 664                         free_extent_map(split);
 665                         split = NULL;
 666                 }
 667 next:
 668                 if (extent_map_in_tree(em))
 669                         remove_extent_mapping(em_tree, em);
 670                 write_unlock(&em_tree->lock);
 671
 672                 /* once for us */
 673                 free_extent_map(em);
 674                 /* once for the tree*/
 675                 free_extent_map(em);
 676         }
 677         if (split)
 678                 free_extent_map(split);
 679         if (split2)
 680                 free_extent_map(split2);
 681 }
 682
 683 /*
 684  * this is very complex, but the basic idea is to drop all extents
 685  * in the range start - end.  hint_block is filled in with a block number
 686  * that would be a good hint to the block allocator for this file.
 687  *
 688  * If an extent intersects the range but is not entirely inside the range
 689  * it is either truncated or split.  Anything entirely inside the range
 690  * is deleted from the tree.
 691  */
 692 int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
 693                          struct btrfs_root *root, struct inode *inode,
 694                          struct btrfs_path *path, u64 start, u64 end,
 695                          u64 *drop_end, int drop_cache,
 696                          int replace_extent,
 697                          u32 extent_item_size,
 698                          int *key_inserted)
 699 {
 700         struct btrfs_fs_info *fs_info = root->fs_info;
 701         struct extent_buffer *leaf;
 702         struct btrfs_file_extent_item *fi;
 703         struct btrfs_key key;
 704         struct btrfs_key new_key;
 705         u64 ino = btrfs_ino(BTRFS_I(inode));
 706         u64 search_start = start;
 707         u64 disk_bytenr = 0;
 708         u64 num_bytes = 0;
 709         u64 extent_offset = 0;
 710         u64 extent_end = 0;
 711         u64 last_end = start;
 712         int del_nr = 0;
 713         int del_slot = 0;
 714         int extent_type;
 715         int recow;
 716         int ret;
 717         int modify_tree = -1;
 718         int update_refs;
 719         int found = 0;
 720         int leafs_visited = 0;
 721
 722         if (drop_cache)
 723                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end - 1, 0);
 724
 725         if (start >= BTRFS_I(inode)->disk_i_size && !replace_extent)
 726                 modify_tree = 0;
 727
 728         update_refs = (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
 729                        root == fs_info->tree_root);
 730         while (1) {
 731                 recow = 0;
 732                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
 733                                                search_start, modify_tree);
 734                 if (ret < 0)
 735                         break;
 736                 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
 737                         leaf = path->nodes[0];
 738                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
 739                         if (key.objectid == ino &&
 740                             key.type == BTRFS_EXTENT_DATA_KEY)
 741                                 path->slots[0]--;
 742                 }
 743                 ret = 0;
 744                 leafs_visited++;
 745 next_slot:
 746                 leaf = path->nodes[0];
 747                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 748                         BUG_ON(del_nr > 0);
 749                         ret = btrfs_next_leaf(root, path);
 750                         if (ret < 0)
 751                                 break;
 752                         if (ret > 0) {
 753                                 ret = 0;
 754                                 break;
 755                         }
 756                         leafs_visited++;
 757                         leaf = path->nodes[0];
 758                         recow = 1;
 759                 }
 760
 761                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 762
 763                 if (key.objectid > ino)
 764                         break;
 765                 if (WARN_ON_ONCE(key.objectid < ino) ||
 766                     key.type < BTRFS_EXTENT_DATA_KEY) {
 767                         ASSERT(del_nr == 0);
 768                         path->slots[0]++;
 769                         goto next_slot;
 770                 }
 771                 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
 772                         break;
 773
 774                 fi = btrfs_item_ptr(leaf, path->slots[0],
 775                                     struct btrfs_file_extent_item);
 776                 extent_type = btrfs_file_extent_type(leaf, fi);
 777
 778                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
 779                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
 780                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 781                         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 782                         extent_offset = btrfs_file_extent_offset(leaf, fi);
 783                         extent_end = key.offset +
 784                                 btrfs_file_extent_num_bytes(leaf, fi);
 785                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 786                         extent_end = key.offset +
 787                                 btrfs_file_extent_inline_len(leaf,
 788                                                      path->slots[0], fi);
 789                 } else {
 790                         /* can't happen */
 791                         BUG();
 792                 }
 793
 794                 /*
 795                  * Don't skip extent items representing 0 byte lengths. They
 796                  * used to be created (bug) if while punching holes we hit
 797                  * -ENOSPC condition. So if we find one here, just ensure we
 798                  * delete it, otherwise we would insert a new file extent item
 799                  * with the same key (offset) as that 0 bytes length file
 800                  * extent item in the call to setup_items_for_insert() later
 801                  * in this function.
 802                  */
 803                 if (extent_end == key.offset && extent_end >= search_start) {
 804                         last_end = extent_end;
 805                         goto delete_extent_item;
 806                 }
 807
 808                 if (extent_end <= search_start) {
 809                         path->slots[0]++;
 810                         goto next_slot;
 811                 }
 812
 813                 found = 1;
 814                 search_start = max(key.offset, start);
 815                 if (recow || !modify_tree) {
 816                         modify_tree = -1;
 817                         btrfs_release_path(path);
 818                         continue;
 819                 }
 820
 821                 /*
 822                  *     | - range to drop - |
 823                  *  | -------- extent -------- |
 824                  */
 825                 if (start > key.offset && end < extent_end) {
 826                         BUG_ON(del_nr > 0);
 827                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 828                                 ret = -EOPNOTSUPP;
 829                                 break;
 830                         }
 831
 832                         memcpy(&new_key, &key, sizeof(new_key));
 833                         new_key.offset = start;
 834                         ret = btrfs_duplicate_item(trans, root, path,
 835                                                    &new_key);
 836                         if (ret == -EAGAIN) {
 837                                 btrfs_release_path(path);
 838                                 continue;
 839                         }
 840                         if (ret < 0)
 841                                 break;
 842
 843                         leaf = path->nodes[0];
 844                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 845                                             struct btrfs_file_extent_item);
 846                         btrfs_set_file_extent_num_bytes(leaf, fi,
 847                                                         start - key.offset);
 848
 849                         fi = btrfs_item_ptr(leaf, path->slots[0],
 850                                             struct btrfs_file_extent_item);
 851
 852                         extent_offset += start - key.offset;
 853                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 854                         btrfs_set_file_extent_num_bytes(leaf, fi,
 855                                                         extent_end - start);
 856                         btrfs_mark_buffer_dirty(leaf);
 857
 858                         if (update_refs && disk_bytenr > 0) {
 859                                 ret = btrfs_inc_extent_ref(trans, fs_info,
 860                                                 disk_bytenr, num_bytes, 0,
 861                                                 root->root_key.objectid,
 862                                                 new_key.objectid,
 863                                                 start - extent_offset);
 864                                 BUG_ON(ret); /* -ENOMEM */
 865                         }
 866                         key.offset = start;
 867                 }
 868                 /*
 869                  * From here on out we will have actually dropped something, so
 870                  * last_end can be updated.
 871                  */
 872                 last_end = extent_end;
 873
 874                 /*
 875                  *  | ---- range to drop ----- |
 876                  *      | -------- extent -------- |
 877                  */
 878                 if (start <= key.offset && end < extent_end) {
 879                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 880                                 ret = -EOPNOTSUPP;
 881                                 break;
 882                         }
 883
 884                         memcpy(&new_key, &key, sizeof(new_key));
 885                         new_key.offset = end;
 886                         btrfs_set_item_key_safe(fs_info, path, &new_key);
 887
 888                         extent_offset += end - key.offset;
 889                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 890                         btrfs_set_file_extent_num_bytes(leaf, fi,
 891                                                         extent_end - end);
 892                         btrfs_mark_buffer_dirty(leaf);
 893                         if (update_refs && disk_bytenr > 0)
 894                                 inode_sub_bytes(inode, end - key.offset);
 895                         break;
 896                 }
 897
 898                 search_start = extent_end;
 899                 /*
 900                  *       | ---- range to drop ----- |
 901                  *  | -------- extent -------- |
 902                  */
 903                 if (start > key.offset && end >= extent_end) {
 904                         BUG_ON(del_nr > 0);
 905                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 906                                 ret = -EOPNOTSUPP;
 907                                 break;
 908                         }
 909
 910                         btrfs_set_file_extent_num_bytes(leaf, fi,
 911                                                         start - key.offset);
 912                         btrfs_mark_buffer_dirty(leaf);
 913                         if (update_refs && disk_bytenr > 0)
 914                                 inode_sub_bytes(inode, extent_end - start);
 915                         if (end == extent_end)
 916                                 break;
 917
 918                         path->slots[0]++;
 919                         goto next_slot;
 920                 }
 921
 922                 /*
 923                  *  | ---- range to drop ----- |
 924                  *    | ------ extent ------ |
 925                  */
 926                 if (start <= key.offset && end >= extent_end) {
 927 delete_extent_item:
 928                         if (del_nr == 0) {
 929                                 del_slot = path->slots[0];
 930                                 del_nr = 1;
 931                         } else {
 932                                 BUG_ON(del_slot + del_nr != path->slots[0]);
 933                                 del_nr++;
 934                         }
 935
 936                         if (update_refs &&
 937                             extent_type == BTRFS_FILE_EXTENT_INLINE) {
 938                                 inode_sub_bytes(inode,
 939                                                 extent_end - key.offset);
 940                                 extent_end = ALIGN(extent_end,
 941                                                    fs_info->sectorsize);
 942                         } else if (update_refs && disk_bytenr > 0) {
 943                                 ret = btrfs_free_extent(trans, fs_info,
 944                                                 disk_bytenr, num_bytes, 0,
 945                                                 root->root_key.objectid,
 946                                                 key.objectid, key.offset -
 947                                                 extent_offset);
 948                                 BUG_ON(ret); /* -ENOMEM */
 949                                 inode_sub_bytes(inode,
 950                                                 extent_end - key.offset);
 951                         }
 952
 953                         if (end == extent_end)
 954                                 break;
 955
 956                         if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
 957                                 path->slots[0]++;
 958                                 goto next_slot;
 959                         }
 960
 961                         ret = btrfs_del_items(trans, root, path, del_slot,
 962                                               del_nr);
 963                         if (ret) {
 964                                 btrfs_abort_transaction(trans, ret);
 965                                 break;
 966                         }
 967
 968                         del_nr = 0;
 969                         del_slot = 0;
 970
 971                         btrfs_release_path(path);
 972                         continue;
 973                 }
 974
 975                 BUG_ON(1);
 976         }
 977
 978         if (!ret && del_nr > 0) {
 979                 /*
 980                  * Set path->slots[0] to first slot, so that after the delete
 981                  * if items are move off from our leaf to its immediate left or
 982                  * right neighbor leafs, we end up with a correct and adjusted
 983                  * path->slots[0] for our insertion (if replace_extent != 0).
 984                  */
 985                 path->slots[0] = del_slot;
 986                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
 987                 if (ret)
 988                         btrfs_abort_transaction(trans, ret);
 989         }
 990
 991         leaf = path->nodes[0];
 992         /*
 993          * If btrfs_del_items() was called, it might have deleted a leaf, in
 994          * which case it unlocked our path, so check path->locks[0] matches a
 995          * write lock.
 996          */
 997         if (!ret && replace_extent && leafs_visited == 1 &&
 998             (path->locks[0] == BTRFS_WRITE_LOCK_BLOCKING ||
 999              path->locks[0] == BTRFS_WRITE_LOCK) &&
1000             btrfs_leaf_free_space(fs_info, leaf) >=
1001             sizeof(struct btrfs_item) + extent_item_size) {
1002
1003                 key.objectid = ino;
1004                 key.type = BTRFS_EXTENT_DATA_KEY;
1005                 key.offset = start;
1006                 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
1007                         struct btrfs_key slot_key;
1008
1009                         btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
1010                         if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
1011                                 path->slots[0]++;
1012                 }
1013                 setup_items_for_insert(root, path, &key,
1014                                        &extent_item_size,
1015                                        extent_item_size,
1016                                        sizeof(struct btrfs_item) +
1017                                        extent_item_size, 1);
1018                 *key_inserted = 1;
1019         }
1020
1021         if (!replace_extent || !(*key_inserted))
1022                 btrfs_release_path(path);
1023         if (drop_end)
1024                 *drop_end = found ? min(end, last_end) : end;
1025         return ret;
1026 }
1027
1028 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
1029                        struct btrfs_root *root, struct inode *inode, u64 start,
1030                        u64 end, int drop_cache)
1031 {
1032         struct btrfs_path *path;
1033         int ret;
1034
1035         path = btrfs_alloc_path();
1036         if (!path)
1037                 return -ENOMEM;
1038         ret = __btrfs_drop_extents(trans, root, inode, path, start, end, NULL,
1039                                    drop_cache, 0, 0, NULL);
1040         btrfs_free_path(path);
1041         return ret;
1042 }
1043
1044 static int extent_mergeable(struct extent_buffer *leaf, int slot,
1045                             u64 objectid, u64 bytenr, u64 orig_offset,
1046                             u64 *start, u64 *end)
1047 {
1048         struct btrfs_file_extent_item *fi;
1049         struct btrfs_key key;
1050         u64 extent_end;
1051
1052         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1053                 return 0;
1054
1055         btrfs_item_key_to_cpu(leaf, &key, slot);
1056         if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1057                 return 0;
1058
1059         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1060         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1061             btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1062             btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1063             btrfs_file_extent_compression(leaf, fi) ||
1064             btrfs_file_extent_encryption(leaf, fi) ||
1065             btrfs_file_extent_other_encoding(leaf, fi))
1066                 return 0;
1067
1068         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1069         if ((*start && *start != key.offset) || (*end && *end != extent_end))
1070                 return 0;
1071
1072         *start = key.offset;
1073         *end = extent_end;
1074         return 1;
1075 }
1076
1077 /*
1078  * Mark extent in the range start - end as written.
1079  *
1080  * This changes extent type from 'pre-allocated' to 'regular'. If only
1081  * part of extent is marked as written, the extent will be split into
1082  * two or three.
1083  */
1084 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1085                               struct btrfs_inode *inode, u64 start, u64 end)
1086 {
1087         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1088         struct btrfs_root *root = inode->root;
1089         struct extent_buffer *leaf;
1090         struct btrfs_path *path;
1091         struct btrfs_file_extent_item *fi;
1092         struct btrfs_key key;
1093         struct btrfs_key new_key;
1094         u64 bytenr;
1095         u64 num_bytes;
1096         u64 extent_end;
1097         u64 orig_offset;
1098         u64 other_start;
1099         u64 other_end;
1100         u64 split;
1101         int del_nr = 0;
1102         int del_slot = 0;
1103         int recow;
1104         int ret;
1105         u64 ino = btrfs_ino(inode);
1106
1107         path = btrfs_alloc_path();
1108         if (!path)
1109                 return -ENOMEM;
1110 again:
1111         recow = 0;
1112         split = start;
1113         key.objectid = ino;
1114         key.type = BTRFS_EXTENT_DATA_KEY;
1115         key.offset = split;
1116
1117         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1118         if (ret < 0)
1119                 goto out;
1120         if (ret > 0 && path->slots[0] > 0)
1121                 path->slots[0]--;
1122
1123         leaf = path->nodes[0];
1124         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1125         if (key.objectid != ino ||
1126             key.type != BTRFS_EXTENT_DATA_KEY) {
1127                 ret = -EINVAL;
1128                 btrfs_abort_transaction(trans, ret);
1129                 goto out;
1130         }
1131         fi = btrfs_item_ptr(leaf, path->slots[0],
1132                             struct btrfs_file_extent_item);
1133         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
1134                 ret = -EINVAL;
1135                 btrfs_abort_transaction(trans, ret);
1136                 goto out;
1137         }
1138         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1139         if (key.offset > start || extent_end < end) {
1140                 ret = -EINVAL;
1141                 btrfs_abort_transaction(trans, ret);
1142                 goto out;
1143         }
1144
1145         bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1146         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1147         orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1148         memcpy(&new_key, &key, sizeof(new_key));
1149
1150         if (start == key.offset && end < extent_end) {
1151                 other_start = 0;
1152                 other_end = start;
1153                 if (extent_mergeable(leaf, path->slots[0] - 1,
1154                                      ino, bytenr, orig_offset,
1155                                      &other_start, &other_end)) {
1156                         new_key.offset = end;
1157                         btrfs_set_item_key_safe(fs_info, path, &new_key);
1158                         fi = btrfs_item_ptr(leaf, path->slots[0],
1159                                             struct btrfs_file_extent_item);
1160                         btrfs_set_file_extent_generation(leaf, fi,
1161                                                          trans->transid);
1162                         btrfs_set_file_extent_num_bytes(leaf, fi,
1163                                                         extent_end - end);
1164                         btrfs_set_file_extent_offset(leaf, fi,
1165                                                      end - orig_offset);
1166                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1167                                             struct btrfs_file_extent_item);
1168                         btrfs_set_file_extent_generation(leaf, fi,
1169                                                          trans->transid);
1170                         btrfs_set_file_extent_num_bytes(leaf, fi,
1171                                                         end - other_start);
1172                         btrfs_mark_buffer_dirty(leaf);
1173                         goto out;
1174                 }
1175         }
1176
1177         if (start > key.offset && end == extent_end) {
1178                 other_start = end;
1179                 other_end = 0;
1180                 if (extent_mergeable(leaf, path->slots[0] + 1,
1181                                      ino, bytenr, orig_offset,
1182                                      &other_start, &other_end)) {
1183                         fi = btrfs_item_ptr(leaf, path->slots[0],
1184                                             struct btrfs_file_extent_item);
1185                         btrfs_set_file_extent_num_bytes(leaf, fi,
1186                                                         start - key.offset);
1187                         btrfs_set_file_extent_generation(leaf, fi,
1188                                                          trans->transid);
1189                         path->slots[0]++;
1190                         new_key.offset = start;
1191                         btrfs_set_item_key_safe(fs_info, path, &new_key);
1192
1193                         fi = btrfs_item_ptr(leaf, path->slots[0],
1194                                             struct btrfs_file_extent_item);
1195                         btrfs_set_file_extent_generation(leaf, fi,
1196                                                          trans->transid);
1197                         btrfs_set_file_extent_num_bytes(leaf, fi,
1198                                                         other_end - start);
1199                         btrfs_set_file_extent_offset(leaf, fi,
1200                                                      start - orig_offset);
1201                         btrfs_mark_buffer_dirty(leaf);
1202                         goto out;
1203                 }
1204         }
1205
1206         while (start > key.offset || end < extent_end) {
1207                 if (key.offset == start)
1208                         split = end;
1209
1210                 new_key.offset = split;
1211                 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1212                 if (ret == -EAGAIN) {
1213                         btrfs_release_path(path);
1214                         goto again;
1215                 }
1216                 if (ret < 0) {
1217                         btrfs_abort_transaction(trans, ret);
1218                         goto out;
1219                 }
1220
1221                 leaf = path->nodes[0];
1222                 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1223                                     struct btrfs_file_extent_item);
1224                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1225                 btrfs_set_file_extent_num_bytes(leaf, fi,
1226                                                 split - key.offset);
1227
1228                 fi = btrfs_item_ptr(leaf, path->slots[0],
1229                                     struct btrfs_file_extent_item);
1230
1231                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1232                 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1233                 btrfs_set_file_extent_num_bytes(leaf, fi,
1234                                                 extent_end - split);
1235                 btrfs_mark_buffer_dirty(leaf);
1236
1237                 ret = btrfs_inc_extent_ref(trans, fs_info, bytenr, num_bytes,
1238                                            0, root->root_key.objectid,
1239                                            ino, orig_offset);
1240                 if (ret) {
1241                         btrfs_abort_transaction(trans, ret);
1242                         goto out;
1243                 }
1244
1245                 if (split == start) {
1246                         key.offset = start;
1247                 } else {
1248                         if (start != key.offset) {
1249                                 ret = -EINVAL;
1250                                 btrfs_abort_transaction(trans, ret);
1251                                 goto out;
1252                         }
1253                         path->slots[0]--;
1254                         extent_end = end;
1255                 }
1256                 recow = 1;
1257         }
1258
1259         other_start = end;
1260         other_end = 0;
1261         if (extent_mergeable(leaf, path->slots[0] + 1,
1262                              ino, bytenr, orig_offset,
1263                              &other_start, &other_end)) {
1264                 if (recow) {
1265                         btrfs_release_path(path);
1266                         goto again;
1267                 }
1268                 extent_end = other_end;
1269                 del_slot = path->slots[0] + 1;
1270                 del_nr++;
1271                 ret = btrfs_free_extent(trans, fs_info, bytenr, num_bytes,
1272                                         0, root->root_key.objectid,
1273                                         ino, orig_offset);
1274                 if (ret) {
1275                         btrfs_abort_transaction(trans, ret);
1276                         goto out;
1277                 }
1278         }
1279         other_start = 0;
1280         other_end = start;
1281         if (extent_mergeable(leaf, path->slots[0] - 1,
1282                              ino, bytenr, orig_offset,
1283                              &other_start, &other_end)) {
1284                 if (recow) {
1285                         btrfs_release_path(path);
1286                         goto again;
1287                 }
1288                 key.offset = other_start;
1289                 del_slot = path->slots[0];
1290                 del_nr++;
1291                 ret = btrfs_free_extent(trans, fs_info, bytenr, num_bytes,
1292                                         0, root->root_key.objectid,
1293                                         ino, orig_offset);
1294                 if (ret) {
1295                         btrfs_abort_transaction(trans, ret);
1296                         goto out;
1297                 }
1298         }
1299         if (del_nr == 0) {
1300                 fi = btrfs_item_ptr(leaf, path->slots[0],
1301                            struct btrfs_file_extent_item);
1302                 btrfs_set_file_extent_type(leaf, fi,
1303                                            BTRFS_FILE_EXTENT_REG);
1304                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1305                 btrfs_mark_buffer_dirty(leaf);
1306         } else {
1307                 fi = btrfs_item_ptr(leaf, del_slot - 1,
1308                            struct btrfs_file_extent_item);
1309                 btrfs_set_file_extent_type(leaf, fi,
1310                                            BTRFS_FILE_EXTENT_REG);
1311                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1312                 btrfs_set_file_extent_num_bytes(leaf, fi,
1313                                                 extent_end - key.offset);
1314                 btrfs_mark_buffer_dirty(leaf);
1315
1316                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1317                 if (ret < 0) {
1318                         btrfs_abort_transaction(trans, ret);
1319                         goto out;
1320                 }
1321         }
1322 out:
1323         btrfs_free_path(path);
1324         return 0;
1325 }
1326
1327 /*
1328  * on error we return an unlocked page and the error value
1329  * on success we return a locked page and 0
1330  */
1331 static int prepare_uptodate_page(struct inode *inode,
1332                                  struct page *page, u64 pos,
1333                                  bool force_uptodate)
1334 {
1335         int ret = 0;
1336
1337         if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1338             !PageUptodate(page)) {
1339                 ret = btrfs_readpage(NULL, page);
1340                 if (ret)
1341                         return ret;
1342                 lock_page(page);
1343                 if (!PageUptodate(page)) {
1344                         unlock_page(page);
1345                         return -EIO;
1346                 }
1347                 if (page->mapping != inode->i_mapping) {
1348                         unlock_page(page);
1349                         return -EAGAIN;
1350                 }
1351         }
1352         return 0;
1353 }
1354
1355 /*
1356  * this just gets pages into the page cache and locks them down.
1357  */
1358 static noinline int prepare_pages(struct inode *inode, struct page **pages,
1359                                   size_t num_pages, loff_t pos,
1360                                   size_t write_bytes, bool force_uptodate)
1361 {
1362         int i;
1363         unsigned long index = pos >> PAGE_SHIFT;
1364         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1365         int err = 0;
1366         int faili;
1367
1368         for (i = 0; i < num_pages; i++) {
1369 again:
1370                 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1371                                                mask | __GFP_WRITE);
1372                 if (!pages[i]) {
1373                         faili = i - 1;
1374                         err = -ENOMEM;
1375                         goto fail;
1376                 }
1377
1378                 if (i == 0)
1379                         err = prepare_uptodate_page(inode, pages[i], pos,
1380                                                     force_uptodate);
1381                 if (!err && i == num_pages - 1)
1382                         err = prepare_uptodate_page(inode, pages[i],
1383                                                     pos + write_bytes, false);
1384                 if (err) {
1385                         put_page(pages[i]);
1386                         if (err == -EAGAIN) {
1387                                 err = 0;
1388                                 goto again;
1389                         }
1390                         faili = i - 1;
1391                         goto fail;
1392                 }
1393                 wait_on_page_writeback(pages[i]);
1394         }
1395
1396         return 0;
1397 fail:
1398         while (faili >= 0) {
1399                 unlock_page(pages[faili]);
1400                 put_page(pages[faili]);
1401                 faili--;
1402         }
1403         return err;
1404
1405 }
1406
1407 /*
1408  * This function locks the extent and properly waits for data=ordered extents
1409  * to finish before allowing the pages to be modified if need.
1410  *
1411  * The return value:
1412  * 1 - the extent is locked
1413  * 0 - the extent is not locked, and everything is OK
1414  * -EAGAIN - need re-prepare the pages
1415  * the other < 0 number - Something wrong happens
1416  */
1417 static noinline int
1418 lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct page **pages,
1419                                 size_t num_pages, loff_t pos,
1420                                 size_t write_bytes,
1421                                 u64 *lockstart, u64 *lockend,
1422                                 struct extent_state **cached_state)
1423 {
1424         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1425         u64 start_pos;
1426         u64 last_pos;
1427         int i;
1428         int ret = 0;
1429
1430         start_pos = round_down(pos, fs_info->sectorsize);
1431         last_pos = start_pos
1432                 + round_up(pos + write_bytes - start_pos,
1433                            fs_info->sectorsize) - 1;
1434
1435         if (start_pos < inode->vfs_inode.i_size) {
1436                 struct btrfs_ordered_extent *ordered;
1437                 lock_extent_bits(&inode->io_tree, start_pos, last_pos,
1438                                 cached_state);
1439                 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1440                                                      last_pos - start_pos + 1);
1441                 if (ordered &&
1442                     ordered->file_offset + ordered->len > start_pos &&
1443                     ordered->file_offset <= last_pos) {
1444                         unlock_extent_cached(&inode->io_tree, start_pos,
1445                                         last_pos, cached_state, GFP_NOFS);
1446                         for (i = 0; i < num_pages; i++) {
1447                                 unlock_page(pages[i]);
1448                                 put_page(pages[i]);
1449                         }
1450                         btrfs_start_ordered_extent(&inode->vfs_inode,
1451                                         ordered, 1);
1452                         btrfs_put_ordered_extent(ordered);
1453                         return -EAGAIN;
1454                 }
1455                 if (ordered)
1456                         btrfs_put_ordered_extent(ordered);
1457
1458                 clear_extent_bit(&inode->io_tree, start_pos,
1459                                   last_pos, EXTENT_DIRTY | EXTENT_DELALLOC |
1460                                   EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
1461                                   0, 0, cached_state, GFP_NOFS);
1462                 *lockstart = start_pos;
1463                 *lockend = last_pos;
1464                 ret = 1;
1465         }
1466
1467         for (i = 0; i < num_pages; i++) {
1468                 if (clear_page_dirty_for_io(pages[i]))
1469                         account_page_redirty(pages[i]);
1470                 set_page_extent_mapped(pages[i]);
1471                 WARN_ON(!PageLocked(pages[i]));
1472         }
1473
1474         return ret;
1475 }
1476
1477 static noinline int check_can_nocow(struct btrfs_inode *inode, loff_t pos,
1478                                     size_t *write_bytes)
1479 {
1480         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1481         struct btrfs_root *root = inode->root;
1482         struct btrfs_ordered_extent *ordered;
1483         u64 lockstart, lockend;
1484         u64 num_bytes;
1485         int ret;
1486
1487         ret = btrfs_start_write_no_snapshoting(root);
1488         if (!ret)
1489                 return -ENOSPC;
1490
1491         lockstart = round_down(pos, fs_info->sectorsize);
1492         lockend = round_up(pos + *write_bytes,
1493                            fs_info->sectorsize) - 1;
1494
1495         while (1) {
1496                 lock_extent(&inode->io_tree, lockstart, lockend);
1497                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
1498                                                      lockend - lockstart + 1);
1499                 if (!ordered) {
1500                         break;
1501                 }
1502                 unlock_extent(&inode->io_tree, lockstart, lockend);
1503                 btrfs_start_ordered_extent(&inode->vfs_inode, ordered, 1);
1504                 btrfs_put_ordered_extent(ordered);
1505         }
1506
1507         num_bytes = lockend - lockstart + 1;
1508         ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1509                         NULL, NULL, NULL);
1510         if (ret <= 0) {
1511                 ret = 0;
1512                 btrfs_end_write_no_snapshoting(root);
1513         } else {
1514                 *write_bytes = min_t(size_t, *write_bytes ,
1515                                      num_bytes - pos + lockstart);
1516         }
1517
1518         unlock_extent(&inode->io_tree, lockstart, lockend);
1519
1520         return ret;
1521 }
1522
1523 static noinline ssize_t __btrfs_buffered_write(struct file *file,
1524                                                struct iov_iter *i,
1525                                                loff_t pos)
1526 {
1527         struct inode *inode = file_inode(file);
1528         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1529         struct btrfs_root *root = BTRFS_I(inode)->root;
1530         struct page **pages = NULL;
1531         struct extent_state *cached_state = NULL;
1532         u64 release_bytes = 0;
1533         u64 lockstart;
1534         u64 lockend;
1535         size_t num_written = 0;
1536         int nrptrs;
1537         int ret = 0;
1538         bool only_release_metadata = false;
1539         bool force_page_uptodate = false;
1540         bool need_unlock;
1541
1542         nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1543                         PAGE_SIZE / (sizeof(struct page *)));
1544         nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1545         nrptrs = max(nrptrs, 8);
1546         pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1547         if (!pages)
1548                 return -ENOMEM;
1549
1550         while (iov_iter_count(i) > 0) {
1551                 size_t offset = pos & (PAGE_SIZE - 1);
1552                 size_t sector_offset;
1553                 size_t write_bytes = min(iov_iter_count(i),
1554                                          nrptrs * (size_t)PAGE_SIZE -
1555                                          offset);
1556                 size_t num_pages = DIV_ROUND_UP(write_bytes + offset,
1557                                                 PAGE_SIZE);
1558                 size_t reserve_bytes;
1559                 size_t dirty_pages;
1560                 size_t copied;
1561                 size_t dirty_sectors;
1562                 size_t num_sectors;
1563
1564                 WARN_ON(num_pages > nrptrs);
1565
1566                 /*
1567                  * Fault pages before locking them in prepare_pages
1568                  * to avoid recursive lock
1569                  */
1570                 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1571                         ret = -EFAULT;
1572                         break;
1573                 }
1574
1575                 sector_offset = pos & (fs_info->sectorsize - 1);
1576                 reserve_bytes = round_up(write_bytes + sector_offset,
1577                                 fs_info->sectorsize);
1578
1579                 ret = btrfs_check_data_free_space(inode, pos, write_bytes);
1580                 if (ret < 0) {
1581                         if ((BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1582                                                       BTRFS_INODE_PREALLOC)) &&
1583                             check_can_nocow(BTRFS_I(inode), pos,
1584                                         &write_bytes) > 0) {
1585                                 /*
1586                                  * For nodata cow case, no need to reserve
1587                                  * data space.
1588                                  */
1589                                 only_release_metadata = true;
1590                                 /*
1591                                  * our prealloc extent may be smaller than
1592                                  * write_bytes, so scale down.
1593                                  */
1594                                 num_pages = DIV_ROUND_UP(write_bytes + offset,
1595                                                          PAGE_SIZE);
1596                                 reserve_bytes = round_up(write_bytes +
1597                                                          sector_offset,
1598                                                          fs_info->sectorsize);
1599                         } else {
1600                                 break;
1601                         }
1602                 }
1603
1604                 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1605                                 reserve_bytes);
1606                 if (ret) {
1607                         if (!only_release_metadata)
1608                                 btrfs_free_reserved_data_space(inode, pos,
1609                                                                write_bytes);
1610                         else
1611                                 btrfs_end_write_no_snapshoting(root);
1612                         break;
1613                 }
1614
1615                 release_bytes = reserve_bytes;
1616                 need_unlock = false;
1617 again:
1618                 /*
1619                  * This is going to setup the pages array with the number of
1620                  * pages we want, so we don't really need to worry about the
1621                  * contents of pages from loop to loop
1622                  */
1623                 ret = prepare_pages(inode, pages, num_pages,
1624                                     pos, write_bytes,
1625                                     force_page_uptodate);
1626                 if (ret)
1627                         break;
1628
1629                 ret = lock_and_cleanup_extent_if_need(BTRFS_I(inode), pages,
1630                                 num_pages, pos, write_bytes, &lockstart,
1631                                 &lockend, &cached_state);
1632                 if (ret < 0) {
1633                         if (ret == -EAGAIN)
1634                                 goto again;
1635                         break;
1636                 } else if (ret > 0) {
1637                         need_unlock = true;
1638                         ret = 0;
1639                 }
1640
1641                 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1642
1643                 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1644                 dirty_sectors = round_up(copied + sector_offset,
1645                                         fs_info->sectorsize);
1646                 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1647
1648                 /*
1649                  * if we have trouble faulting in the pages, fall
1650                  * back to one page at a time
1651                  */
1652                 if (copied < write_bytes)
1653                         nrptrs = 1;
1654
1655                 if (copied == 0) {
1656                         force_page_uptodate = true;
1657                         dirty_sectors = 0;
1658                         dirty_pages = 0;
1659                 } else {
1660                         force_page_uptodate = false;
1661                         dirty_pages = DIV_ROUND_UP(copied + offset,
1662                                                    PAGE_SIZE);
1663                 }
1664
1665                 /*
1666                  * If we had a short copy we need to release the excess delaloc
1667                  * bytes we reserved.  We need to increment outstanding_extents
1668                  * because btrfs_delalloc_release_space and
1669                  * btrfs_delalloc_release_metadata will decrement it, but
1670                  * we still have an outstanding extent for the chunk we actually
1671                  * managed to copy.
1672                  */
1673                 if (num_sectors > dirty_sectors) {
1674                         /* release everything except the sectors we dirtied */
1675                         release_bytes -= dirty_sectors <<
1676                                                 fs_info->sb->s_blocksize_bits;
1677                         if (copied > 0) {
1678                                 spin_lock(&BTRFS_I(inode)->lock);
1679                                 BTRFS_I(inode)->outstanding_extents++;
1680                                 spin_unlock(&BTRFS_I(inode)->lock);
1681                         }
1682                         if (only_release_metadata) {
1683                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1684                                                                 release_bytes);
1685                         } else {
1686                                 u64 __pos;
1687
1688                                 __pos = round_down(pos,
1689                                                    fs_info->sectorsize) +
1690                                         (dirty_pages << PAGE_SHIFT);
1691                                 btrfs_delalloc_release_space(inode, __pos,
1692                                                              release_bytes);
1693                         }
1694                 }
1695
1696                 release_bytes = round_up(copied + sector_offset,
1697                                         fs_info->sectorsize);
1698
1699                 if (copied > 0)
1700                         ret = btrfs_dirty_pages(inode, pages, dirty_pages,
1701                                                 pos, copied, NULL);
1702                 if (need_unlock)
1703                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1704                                              lockstart, lockend, &cached_state,
1705                                              GFP_NOFS);
1706                 if (ret) {
1707                         btrfs_drop_pages(pages, num_pages);
1708                         break;
1709                 }
1710
1711                 release_bytes = 0;
1712                 if (only_release_metadata)
1713                         btrfs_end_write_no_snapshoting(root);
1714
1715                 if (only_release_metadata && copied > 0) {
1716                         lockstart = round_down(pos,
1717                                                fs_info->sectorsize);
1718                         lockend = round_up(pos + copied,
1719                                            fs_info->sectorsize) - 1;
1720
1721                         set_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
1722                                        lockend, EXTENT_NORESERVE, NULL,
1723                                        NULL, GFP_NOFS);
1724                         only_release_metadata = false;
1725                 }
1726
1727                 btrfs_drop_pages(pages, num_pages);
1728
1729                 cond_resched();
1730
1731                 balance_dirty_pages_ratelimited(inode->i_mapping);
1732                 if (dirty_pages < (fs_info->nodesize >> PAGE_SHIFT) + 1)
1733                         btrfs_btree_balance_dirty(fs_info);
1734
1735                 pos += copied;
1736                 num_written += copied;
1737         }
1738
1739         kfree(pages);
1740
1741         if (release_bytes) {
1742                 if (only_release_metadata) {
1743                         btrfs_end_write_no_snapshoting(root);
1744                         btrfs_delalloc_release_metadata(BTRFS_I(inode),
1745                                         release_bytes);
1746                 } else {
1747                         btrfs_delalloc_release_space(inode,
1748                                                 round_down(pos, fs_info->sectorsize),
1749                                                 release_bytes);
1750                 }
1751         }
1752
1753         return num_written ? num_written : ret;
1754 }
1755
1756 static ssize_t __btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1757 {
1758         struct file *file = iocb->ki_filp;
1759         struct inode *inode = file_inode(file);
1760         loff_t pos = iocb->ki_pos;
1761         ssize_t written;
1762         ssize_t written_buffered;
1763         loff_t endbyte;
1764         int err;
1765
1766         written = generic_file_direct_write(iocb, from);
1767
1768         if (written < 0 || !iov_iter_count(from))
1769                 return written;
1770
1771         pos += written;
1772         written_buffered = __btrfs_buffered_write(file, from, pos);
1773         if (written_buffered < 0) {
1774                 err = written_buffered;
1775                 goto out;
1776         }
1777         /*
1778          * Ensure all data is persisted. We want the next direct IO read to be
1779          * able to read what was just written.
1780          */
1781         endbyte = pos + written_buffered - 1;
1782         err = btrfs_fdatawrite_range(inode, pos, endbyte);
1783         if (err)
1784                 goto out;
1785         err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1786         if (err)
1787                 goto out;
1788         written += written_buffered;
1789         iocb->ki_pos = pos + written_buffered;
1790         invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1791                                  endbyte >> PAGE_SHIFT);
1792 out:
1793         return written ? written : err;
1794 }
1795
1796 static void update_time_for_write(struct inode *inode)
1797 {
1798         struct timespec now;
1799
1800         if (IS_NOCMTIME(inode))
1801                 return;
1802
1803         now = current_time(inode);
1804         if (!timespec_equal(&inode->i_mtime, &now))
1805                 inode->i_mtime = now;
1806
1807         if (!timespec_equal(&inode->i_ctime, &now))
1808                 inode->i_ctime = now;
1809
1810         if (IS_I_VERSION(inode))
1811                 inode_inc_iversion(inode);
1812 }
1813
1814 static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
1815                                     struct iov_iter *from)
1816 {
1817         struct file *file = iocb->ki_filp;
1818         struct inode *inode = file_inode(file);
1819         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1820         struct btrfs_root *root = BTRFS_I(inode)->root;
1821         u64 start_pos;
1822         u64 end_pos;
1823         ssize_t num_written = 0;
1824         bool sync = (file->f_flags & O_DSYNC) || IS_SYNC(file->f_mapping->host);
1825         ssize_t err;
1826         loff_t pos;
1827         size_t count;
1828         loff_t oldsize;
1829         int clean_page = 0;
1830
1831         inode_lock(inode);
1832         err = generic_write_checks(iocb, from);
1833         if (err <= 0) {
1834                 inode_unlock(inode);
1835                 return err;
1836         }
1837
1838         current->backing_dev_info = inode_to_bdi(inode);
1839         err = file_remove_privs(file);
1840         if (err) {
1841                 inode_unlock(inode);
1842                 goto out;
1843         }
1844
1845         /*
1846          * If BTRFS flips readonly due to some impossible error
1847          * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1848          * although we have opened a file as writable, we have
1849          * to stop this write operation to ensure FS consistency.
1850          */
1851         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
1852                 inode_unlock(inode);
1853                 err = -EROFS;
1854                 goto out;
1855         }
1856
1857         /*
1858          * We reserve space for updating the inode when we reserve space for the
1859          * extent we are going to write, so we will enospc out there.  We don't
1860          * need to start yet another transaction to update the inode as we will
1861          * update the inode when we finish writing whatever data we write.
1862          */
1863         update_time_for_write(inode);
1864
1865         pos = iocb->ki_pos;
1866         count = iov_iter_count(from);
1867         start_pos = round_down(pos, fs_info->sectorsize);
1868         oldsize = i_size_read(inode);
1869         if (start_pos > oldsize) {
1870                 /* Expand hole size to cover write data, preventing empty gap */
1871                 end_pos = round_up(pos + count,
1872                                    fs_info->sectorsize);
1873                 err = btrfs_cont_expand(inode, oldsize, end_pos);
1874                 if (err) {
1875                         inode_unlock(inode);
1876                         goto out;
1877                 }
1878                 if (start_pos > round_up(oldsize, fs_info->sectorsize))
1879                         clean_page = 1;
1880         }
1881
1882         if (sync)
1883                 atomic_inc(&BTRFS_I(inode)->sync_writers);
1884
1885         if (iocb->ki_flags & IOCB_DIRECT) {
1886                 num_written = __btrfs_direct_write(iocb, from);
1887         } else {
1888                 num_written = __btrfs_buffered_write(file, from, pos);
1889                 if (num_written > 0)
1890                         iocb->ki_pos = pos + num_written;
1891                 if (clean_page)
1892                         pagecache_isize_extended(inode, oldsize,
1893                                                 i_size_read(inode));
1894         }
1895
1896         inode_unlock(inode);
1897
1898         /*
1899          * We also have to set last_sub_trans to the current log transid,
1900          * otherwise subsequent syncs to a file that's been synced in this
1901          * transaction will appear to have already occurred.
1902          */
1903         spin_lock(&BTRFS_I(inode)->lock);
1904         BTRFS_I(inode)->last_sub_trans = root->log_transid;
1905         spin_unlock(&BTRFS_I(inode)->lock);
1906         if (num_written > 0)
1907                 num_written = generic_write_sync(iocb, num_written);
1908
1909         if (sync)
1910                 atomic_dec(&BTRFS_I(inode)->sync_writers);
1911 out:
1912         current->backing_dev_info = NULL;
1913         return num_written ? num_written : err;
1914 }
1915
1916 int btrfs_release_file(struct inode *inode, struct file *filp)
1917 {
1918         if (filp->private_data)
1919                 btrfs_ioctl_trans_end(filp);
1920         /*
1921          * ordered_data_close is set by settattr when we are about to truncate
1922          * a file from a non-zero size to a zero size.  This tries to
1923          * flush down new bytes that may have been written if the
1924          * application were using truncate to replace a file in place.
1925          */
1926         if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
1927                                &BTRFS_I(inode)->runtime_flags))
1928                         filemap_flush(inode->i_mapping);
1929         return 0;
1930 }
1931
1932 static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
1933 {
1934         int ret;
1935
1936         atomic_inc(&BTRFS_I(inode)->sync_writers);
1937         ret = btrfs_fdatawrite_range(inode, start, end);
1938         atomic_dec(&BTRFS_I(inode)->sync_writers);
1939
1940         return ret;
1941 }
1942
1943 /*
1944  * fsync call for both files and directories.  This logs the inode into
1945  * the tree log instead of forcing full commits whenever possible.
1946  *
1947  * It needs to call filemap_fdatawait so that all ordered extent updates are
1948  * in the metadata btree are up to date for copying to the log.
1949  *
1950  * It drops the inode mutex before doing the tree log commit.  This is an
1951  * important optimization for directories because holding the mutex prevents
1952  * new operations on the dir while we write to disk.
1953  */
1954 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
1955 {
1956         struct dentry *dentry = file_dentry(file);
1957         struct inode *inode = d_inode(dentry);
1958         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1959         struct btrfs_root *root = BTRFS_I(inode)->root;
1960         struct btrfs_trans_handle *trans;
1961         struct btrfs_log_ctx ctx;
1962         int ret = 0;
1963         bool full_sync = 0;
1964         u64 len;
1965
1966         /*
1967          * The range length can be represented by u64, we have to do the typecasts
1968          * to avoid signed overflow if it's [0, LLONG_MAX] eg. from fsync()
1969          */
1970         len = (u64)end - (u64)start + 1;
1971         trace_btrfs_sync_file(file, datasync);
1972
1973         /*
1974          * We write the dirty pages in the range and wait until they complete
1975          * out of the ->i_mutex. If so, we can flush the dirty pages by
1976          * multi-task, and make the performance up.  See
1977          * btrfs_wait_ordered_range for an explanation of the ASYNC check.
1978          */
1979         ret = start_ordered_ops(inode, start, end);
1980         if (ret)
1981                 return ret;
1982
1983         inode_lock(inode);
1984         atomic_inc(&root->log_batch);
1985         full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1986                              &BTRFS_I(inode)->runtime_flags);
1987         /*
1988          * We might have have had more pages made dirty after calling
1989          * start_ordered_ops and before acquiring the inode's i_mutex.
1990          */
1991         if (full_sync) {
1992                 /*
1993                  * For a full sync, we need to make sure any ordered operations
1994                  * start and finish before we start logging the inode, so that
1995                  * all extents are persisted and the respective file extent
1996                  * items are in the fs/subvol btree.
1997                  */
1998                 ret = btrfs_wait_ordered_range(inode, start, len);
1999         } else {
2000                 /*
2001                  * Start any new ordered operations before starting to log the
2002                  * inode. We will wait for them to finish in btrfs_sync_log().
2003                  *
2004                  * Right before acquiring the inode's mutex, we might have new
2005                  * writes dirtying pages, which won't immediately start the
2006                  * respective ordered operations - that is done through the
2007                  * fill_delalloc callbacks invoked from the writepage and
2008                  * writepages address space operations. So make sure we start
2009                  * all ordered operations before starting to log our inode. Not
2010                  * doing this means that while logging the inode, writeback
2011                  * could start and invoke writepage/writepages, which would call
2012                  * the fill_delalloc callbacks (cow_file_range,
2013                  * submit_compressed_extents). These callbacks add first an
2014                  * extent map to the modified list of extents and then create
2015                  * the respective ordered operation, which means in
2016                  * tree-log.c:btrfs_log_inode() we might capture all existing
2017                  * ordered operations (with btrfs_get_logged_extents()) before
2018                  * the fill_delalloc callback adds its ordered operation, and by
2019                  * the time we visit the modified list of extent maps (with
2020                  * btrfs_log_changed_extents()), we see and process the extent
2021                  * map they created. We then use the extent map to construct a
2022                  * file extent item for logging without waiting for the
2023                  * respective ordered operation to finish - this file extent
2024                  * item points to a disk location that might not have yet been
2025                  * written to, containing random data - so after a crash a log
2026                  * replay will make our inode have file extent items that point
2027                  * to disk locations containing invalid data, as we returned
2028                  * success to userspace without waiting for the respective
2029                  * ordered operation to finish, because it wasn't captured by
2030                  * btrfs_get_logged_extents().
2031                  */
2032                 ret = start_ordered_ops(inode, start, end);
2033         }
2034         if (ret) {
2035                 inode_unlock(inode);
2036                 goto out;
2037         }
2038         atomic_inc(&root->log_batch);
2039
2040         /*
2041          * If the last transaction that changed this file was before the current
2042          * transaction and we have the full sync flag set in our inode, we can
2043          * bail out now without any syncing.
2044          *
2045          * Note that we can't bail out if the full sync flag isn't set. This is
2046          * because when the full sync flag is set we start all ordered extents
2047          * and wait for them to fully complete - when they complete they update
2048          * the inode's last_trans field through:
2049          *
2050          *     btrfs_finish_ordered_io() ->
2051          *         btrfs_update_inode_fallback() ->
2052          *             btrfs_update_inode() ->
2053          *                 btrfs_set_inode_last_trans()
2054          *
2055          * So we are sure that last_trans is up to date and can do this check to
2056          * bail out safely. For the fast path, when the full sync flag is not
2057          * set in our inode, we can not do it because we start only our ordered
2058          * extents and don't wait for them to complete (that is when
2059          * btrfs_finish_ordered_io runs), so here at this point their last_trans
2060          * value might be less than or equals to fs_info->last_trans_committed,
2061          * and setting a speculative last_trans for an inode when a buffered
2062          * write is made (such as fs_info->generation + 1 for example) would not
2063          * be reliable since after setting the value and before fsync is called
2064          * any number of transactions can start and commit (transaction kthread
2065          * commits the current transaction periodically), and a transaction
2066          * commit does not start nor waits for ordered extents to complete.
2067          */
2068         smp_mb();
2069         if (btrfs_inode_in_log(BTRFS_I(inode), fs_info->generation) ||
2070             (full_sync && BTRFS_I(inode)->last_trans <=
2071              fs_info->last_trans_committed) ||
2072             (!btrfs_have_ordered_extents_in_range(inode, start, len) &&
2073              BTRFS_I(inode)->last_trans
2074              <= fs_info->last_trans_committed)) {
2075                 /*
2076                  * We've had everything committed since the last time we were
2077                  * modified so clear this flag in case it was set for whatever
2078                  * reason, it's no longer relevant.
2079                  */
2080                 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2081                           &BTRFS_I(inode)->runtime_flags);
2082                 /*
2083                  * An ordered extent might have started before and completed
2084                  * already with io errors, in which case the inode was not
2085                  * updated and we end up here. So check the inode's mapping
2086                  * flags for any errors that might have happened while doing
2087                  * writeback of file data.
2088                  */
2089                 ret = filemap_check_errors(inode->i_mapping);
2090                 inode_unlock(inode);
2091                 goto out;
2092         }
2093
2094         /*
2095          * ok we haven't committed the transaction yet, lets do a commit
2096          */
2097         if (file->private_data)
2098                 btrfs_ioctl_trans_end(file);
2099
2100         /*
2101          * We use start here because we will need to wait on the IO to complete
2102          * in btrfs_sync_log, which could require joining a transaction (for
2103          * example checking cross references in the nocow path).  If we use join
2104          * here we could get into a situation where we're waiting on IO to
2105          * happen that is blocked on a transaction trying to commit.  With start
2106          * we inc the extwriter counter, so we wait for all extwriters to exit
2107          * before we start blocking join'ers.  This comment is to keep somebody
2108          * from thinking they are super smart and changing this to
2109          * btrfs_join_transaction *cough*Josef*cough*.
2110          */
2111         trans = btrfs_start_transaction(root, 0);
2112         if (IS_ERR(trans)) {
2113                 ret = PTR_ERR(trans);
2114                 inode_unlock(inode);
2115                 goto out;
2116         }
2117         trans->sync = true;
2118
2119         btrfs_init_log_ctx(&ctx, inode);
2120
2121         ret = btrfs_log_dentry_safe(trans, root, dentry, start, end, &ctx);
2122         if (ret < 0) {
2123                 /* Fallthrough and commit/free transaction. */
2124                 ret = 1;
2125         }
2126
2127         /* we've logged all the items and now have a consistent
2128          * version of the file in the log.  It is possible that
2129          * someone will come in and modify the file, but that's
2130          * fine because the log is consistent on disk, and we
2131          * have references to all of the file's extents
2132          *
2133          * It is possible that someone will come in and log the
2134          * file again, but that will end up using the synchronization
2135          * inside btrfs_sync_log to keep things safe.
2136          */
2137         inode_unlock(inode);
2138
2139         /*
2140          * If any of the ordered extents had an error, just return it to user
2141          * space, so that the application knows some writes didn't succeed and
2142          * can take proper action (retry for e.g.). Blindly committing the
2143          * transaction in this case, would fool userspace that everything was
2144          * successful. And we also want to make sure our log doesn't contain
2145          * file extent items pointing to extents that weren't fully written to -
2146          * just like in the non fast fsync path, where we check for the ordered
2147          * operation's error flag before writing to the log tree and return -EIO
2148          * if any of them had this flag set (btrfs_wait_ordered_range) -
2149          * therefore we need to check for errors in the ordered operations,
2150          * which are indicated by ctx.io_err.
2151          */
2152         if (ctx.io_err) {
2153                 btrfs_end_transaction(trans);
2154                 ret = ctx.io_err;
2155                 goto out;
2156         }
2157
2158         if (ret != BTRFS_NO_LOG_SYNC) {
2159                 if (!ret) {
2160                         ret = btrfs_sync_log(trans, root, &ctx);
2161                         if (!ret) {
2162                                 ret = btrfs_end_transaction(trans);
2163                                 goto out;
2164                         }
2165                 }
2166                 if (!full_sync) {
2167                         ret = btrfs_wait_ordered_range(inode, start, len);
2168                         if (ret) {
2169                                 btrfs_end_transaction(trans);
2170                                 goto out;
2171                         }
2172                 }
2173                 ret = btrfs_commit_transaction(trans);
2174         } else {
2175                 ret = btrfs_end_transaction(trans);
2176         }
2177 out:
2178         return ret > 0 ? -EIO : ret;
2179 }
2180
2181 static const struct vm_operations_struct btrfs_file_vm_ops = {
2182         .fault          = filemap_fault,
2183         .map_pages      = filemap_map_pages,
2184         .page_mkwrite   = btrfs_page_mkwrite,
2185 };
2186
2187 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
2188 {
2189         struct address_space *mapping = filp->f_mapping;
2190
2191         if (!mapping->a_ops->readpage)
2192                 return -ENOEXEC;
2193
2194         file_accessed(filp);
2195         vma->vm_ops = &btrfs_file_vm_ops;
2196
2197         return 0;
2198 }
2199
2200 static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2201                           int slot, u64 start, u64 end)
2202 {
2203         struct btrfs_file_extent_item *fi;
2204         struct btrfs_key key;
2205
2206         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2207                 return 0;
2208
2209         btrfs_item_key_to_cpu(leaf, &key, slot);
2210         if (key.objectid != btrfs_ino(inode) ||
2211             key.type != BTRFS_EXTENT_DATA_KEY)
2212                 return 0;
2213
2214         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2215
2216         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2217                 return 0;
2218
2219         if (btrfs_file_extent_disk_bytenr(leaf, fi))
2220                 return 0;
2221
2222         if (key.offset == end)
2223                 return 1;
2224         if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2225                 return 1;
2226         return 0;
2227 }
2228
2229 static int fill_holes(struct btrfs_trans_handle *trans,
2230                 struct btrfs_inode *inode,
2231                 struct btrfs_path *path, u64 offset, u64 end)
2232 {
2233         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
2234         struct btrfs_root *root = inode->root;
2235         struct extent_buffer *leaf;
2236         struct btrfs_file_extent_item *fi;
2237         struct extent_map *hole_em;
2238         struct extent_map_tree *em_tree = &inode->extent_tree;
2239         struct btrfs_key key;
2240         int ret;
2241
2242         if (btrfs_fs_incompat(fs_info, NO_HOLES))
2243                 goto out;
2244
2245         key.objectid = btrfs_ino(inode);
2246         key.type = BTRFS_EXTENT_DATA_KEY;
2247         key.offset = offset;
2248
2249         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2250         if (ret <= 0) {
2251                 /*
2252                  * We should have dropped this offset, so if we find it then
2253                  * something has gone horribly wrong.
2254                  */
2255                 if (ret == 0)
2256                         ret = -EINVAL;
2257                 return ret;
2258         }
2259
2260         leaf = path->nodes[0];
2261         if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2262                 u64 num_bytes;
2263
2264                 path->slots[0]--;
2265                 fi = btrfs_item_ptr(leaf, path->slots[0],
2266                                     struct btrfs_file_extent_item);
2267                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2268                         end - offset;
2269                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2270                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2271                 btrfs_set_file_extent_offset(leaf, fi, 0);
2272                 btrfs_mark_buffer_dirty(leaf);
2273                 goto out;
2274         }
2275
2276         if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2277                 u64 num_bytes;
2278
2279                 key.offset = offset;
2280                 btrfs_set_item_key_safe(fs_info, path, &key);
2281                 fi = btrfs_item_ptr(leaf, path->slots[0],
2282                                     struct btrfs_file_extent_item);
2283                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2284                         offset;
2285                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2286                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2287                 btrfs_set_file_extent_offset(leaf, fi, 0);
2288                 btrfs_mark_buffer_dirty(leaf);
2289                 goto out;
2290         }
2291         btrfs_release_path(path);
2292
2293         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode),
2294                         offset, 0, 0, end - offset, 0, end - offset, 0, 0, 0);
2295         if (ret)
2296                 return ret;
2297
2298 out:
2299         btrfs_release_path(path);
2300
2301         hole_em = alloc_extent_map();
2302         if (!hole_em) {
2303                 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2304                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
2305         } else {
2306                 hole_em->start = offset;
2307                 hole_em->len = end - offset;
2308                 hole_em->ram_bytes = hole_em->len;
2309                 hole_em->orig_start = offset;
2310
2311                 hole_em->block_start = EXTENT_MAP_HOLE;
2312                 hole_em->block_len = 0;
2313                 hole_em->orig_block_len = 0;
2314                 hole_em->bdev = fs_info->fs_devices->latest_bdev;
2315                 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2316                 hole_em->generation = trans->transid;
2317
2318                 do {
2319                         btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2320                         write_lock(&em_tree->lock);
2321                         ret = add_extent_mapping(em_tree, hole_em, 1);
2322                         write_unlock(&em_tree->lock);
2323                 } while (ret == -EEXIST);
2324                 free_extent_map(hole_em);
2325                 if (ret)
2326                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2327                                         &inode->runtime_flags);
2328         }
2329
2330         return 0;
2331 }
2332
2333 /*
2334  * Find a hole extent on given inode and change start/len to the end of hole
2335  * extent.(hole/vacuum extent whose em->start <= start &&
2336  *         em->start + em->len > start)
2337  * When a hole extent is found, return 1 and modify start/len.
2338  */
2339 static int find_first_non_hole(struct inode *inode, u64 *start, u64 *len)
2340 {
2341         struct extent_map *em;
2342         int ret = 0;
2343
2344         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, *start, *len, 0);
2345         if (IS_ERR(em))
2346                 return PTR_ERR(em);
2347
2348         /* Hole or vacuum extent(only exists in no-hole mode) */
2349         if (em->block_start == EXTENT_MAP_HOLE) {
2350                 ret = 1;
2351                 *len = em->start + em->len > *start + *len ?
2352                        0 : *start + *len - em->start - em->len;
2353                 *start = em->start + em->len;
2354         }
2355         free_extent_map(em);
2356         return ret;
2357 }
2358
2359 static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2360 {
2361         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2362         struct btrfs_root *root = BTRFS_I(inode)->root;
2363         struct extent_state *cached_state = NULL;
2364         struct btrfs_path *path;
2365         struct btrfs_block_rsv *rsv;
2366         struct btrfs_trans_handle *trans;
2367         u64 lockstart;
2368         u64 lockend;
2369         u64 tail_start;
2370         u64 tail_len;
2371         u64 orig_start = offset;
2372         u64 cur_offset;
2373         u64 min_size = btrfs_calc_trans_metadata_size(fs_info, 1);
2374         u64 drop_end;
2375         int ret = 0;
2376         int err = 0;
2377         unsigned int rsv_count;
2378         bool same_block;
2379         bool no_holes = btrfs_fs_incompat(fs_info, NO_HOLES);
2380         u64 ino_size;
2381         bool truncated_block = false;
2382         bool updated_inode = false;
2383
2384         ret = btrfs_wait_ordered_range(inode, offset, len);
2385         if (ret)
2386                 return ret;
2387
2388         inode_lock(inode);
2389         ino_size = round_up(inode->i_size, fs_info->sectorsize);
2390         ret = find_first_non_hole(inode, &offset, &len);
2391         if (ret < 0)
2392                 goto out_only_mutex;
2393         if (ret && !len) {
2394                 /* Already in a large hole */
2395                 ret = 0;
2396                 goto out_only_mutex;
2397         }
2398
2399         lockstart = round_up(offset, btrfs_inode_sectorsize(inode));
2400         lockend = round_down(offset + len,
2401                              btrfs_inode_sectorsize(inode)) - 1;
2402         same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2403                 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2404         /*
2405          * We needn't truncate any block which is beyond the end of the file
2406          * because we are sure there is no data there.
2407          */
2408         /*
2409          * Only do this if we are in the same block and we aren't doing the
2410          * entire block.
2411          */
2412         if (same_block && len < fs_info->sectorsize) {
2413                 if (offset < ino_size) {
2414                         truncated_block = true;
2415                         ret = btrfs_truncate_block(inode, offset, len, 0);
2416                 } else {
2417                         ret = 0;
2418                 }
2419                 goto out_only_mutex;
2420         }
2421
2422         /* zero back part of the first block */
2423         if (offset < ino_size) {
2424                 truncated_block = true;
2425                 ret = btrfs_truncate_block(inode, offset, 0, 0);
2426                 if (ret) {
2427                         inode_unlock(inode);
2428                         return ret;
2429                 }
2430         }
2431
2432         /* Check the aligned pages after the first unaligned page,
2433          * if offset != orig_start, which means the first unaligned page
2434          * including several following pages are already in holes,
2435          * the extra check can be skipped */
2436         if (offset == orig_start) {
2437                 /* after truncate page, check hole again */
2438                 len = offset + len - lockstart;
2439                 offset = lockstart;
2440                 ret = find_first_non_hole(inode, &offset, &len);
2441                 if (ret < 0)
2442                         goto out_only_mutex;
2443                 if (ret && !len) {
2444                         ret = 0;
2445                         goto out_only_mutex;
2446                 }
2447                 lockstart = offset;
2448         }
2449
2450         /* Check the tail unaligned part is in a hole */
2451         tail_start = lockend + 1;
2452         tail_len = offset + len - tail_start;
2453         if (tail_len) {
2454                 ret = find_first_non_hole(inode, &tail_start, &tail_len);
2455                 if (unlikely(ret < 0))
2456                         goto out_only_mutex;
2457                 if (!ret) {
2458                         /* zero the front end of the last page */
2459                         if (tail_start + tail_len < ino_size) {
2460                                 truncated_block = true;
2461                                 ret = btrfs_truncate_block(inode,
2462                                                         tail_start + tail_len,
2463                                                         0, 1);
2464                                 if (ret)
2465                                         goto out_only_mutex;
2466                         }
2467                 }
2468         }
2469
2470         if (lockend < lockstart) {
2471                 ret = 0;
2472                 goto out_only_mutex;
2473         }
2474
2475         while (1) {
2476                 struct btrfs_ordered_extent *ordered;
2477
2478                 truncate_pagecache_range(inode, lockstart, lockend);
2479
2480                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2481                                  &cached_state);
2482                 ordered = btrfs_lookup_first_ordered_extent(inode, lockend);
2483
2484                 /*
2485                  * We need to make sure we have no ordered extents in this range
2486                  * and nobody raced in and read a page in this range, if we did
2487                  * we need to try again.
2488                  */
2489                 if ((!ordered ||
2490                     (ordered->file_offset + ordered->len <= lockstart ||
2491                      ordered->file_offset > lockend)) &&
2492                      !btrfs_page_exists_in_range(inode, lockstart, lockend)) {
2493                         if (ordered)
2494                                 btrfs_put_ordered_extent(ordered);
2495                         break;
2496                 }
2497                 if (ordered)
2498                         btrfs_put_ordered_extent(ordered);
2499                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2500                                      lockend, &cached_state, GFP_NOFS);
2501                 ret = btrfs_wait_ordered_range(inode, lockstart,
2502                                                lockend - lockstart + 1);
2503                 if (ret) {
2504                         inode_unlock(inode);
2505                         return ret;
2506                 }
2507         }
2508
2509         path = btrfs_alloc_path();
2510         if (!path) {
2511                 ret = -ENOMEM;
2512                 goto out;
2513         }
2514
2515         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2516         if (!rsv) {
2517                 ret = -ENOMEM;
2518                 goto out_free;
2519         }
2520         rsv->size = btrfs_calc_trans_metadata_size(fs_info, 1);
2521         rsv->failfast = 1;
2522
2523         /*
2524          * 1 - update the inode
2525          * 1 - removing the extents in the range
2526          * 1 - adding the hole extent if no_holes isn't set
2527          */
2528         rsv_count = no_holes ? 2 : 3;
2529         trans = btrfs_start_transaction(root, rsv_count);
2530         if (IS_ERR(trans)) {
2531                 err = PTR_ERR(trans);
2532                 goto out_free;
2533         }
2534
2535         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2536                                       min_size, 0);
2537         BUG_ON(ret);
2538         trans->block_rsv = rsv;
2539
2540         cur_offset = lockstart;
2541         len = lockend - cur_offset;
2542         while (cur_offset < lockend) {
2543                 ret = __btrfs_drop_extents(trans, root, inode, path,
2544                                            cur_offset, lockend + 1,
2545                                            &drop_end, 1, 0, 0, NULL);
2546                 if (ret != -ENOSPC)
2547                         break;
2548
2549                 trans->block_rsv = &fs_info->trans_block_rsv;
2550
2551                 if (cur_offset < drop_end && cur_offset < ino_size) {
2552                         ret = fill_holes(trans, BTRFS_I(inode), path,
2553                                         cur_offset, drop_end);
2554                         if (ret) {
2555                                 /*
2556                                  * If we failed then we didn't insert our hole
2557                                  * entries for the area we dropped, so now the
2558                                  * fs is corrupted, so we must abort the
2559                                  * transaction.
2560                                  */
2561                                 btrfs_abort_transaction(trans, ret);
2562                                 err = ret;
2563                                 break;
2564                         }
2565                 }
2566
2567                 cur_offset = drop_end;
2568
2569                 ret = btrfs_update_inode(trans, root, inode);
2570                 if (ret) {
2571                         err = ret;
2572                         break;
2573                 }
2574
2575                 btrfs_end_transaction(trans);
2576                 btrfs_btree_balance_dirty(fs_info);
2577
2578                 trans = btrfs_start_transaction(root, rsv_count);
2579                 if (IS_ERR(trans)) {
2580                         ret = PTR_ERR(trans);
2581                         trans = NULL;
2582                         break;
2583                 }
2584
2585                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2586                                               rsv, min_size, 0);
2587                 BUG_ON(ret);    /* shouldn't happen */
2588                 trans->block_rsv = rsv;
2589
2590                 ret = find_first_non_hole(inode, &cur_offset, &len);
2591                 if (unlikely(ret < 0))
2592                         break;
2593                 if (ret && !len) {
2594                         ret = 0;
2595                         break;
2596                 }
2597         }
2598
2599         if (ret) {
2600                 err = ret;
2601                 goto out_trans;
2602         }
2603
2604         trans->block_rsv = &fs_info->trans_block_rsv;
2605         /*
2606          * If we are using the NO_HOLES feature we might have had already an
2607          * hole that overlaps a part of the region [lockstart, lockend] and
2608          * ends at (or beyond) lockend. Since we have no file extent items to
2609          * represent holes, drop_end can be less than lockend and so we must
2610          * make sure we have an extent map representing the existing hole (the
2611          * call to __btrfs_drop_extents() might have dropped the existing extent
2612          * map representing the existing hole), otherwise the fast fsync path
2613          * will not record the existence of the hole region
2614          * [existing_hole_start, lockend].
2615          */
2616         if (drop_end <= lockend)
2617                 drop_end = lockend + 1;
2618         /*
2619          * Don't insert file hole extent item if it's for a range beyond eof
2620          * (because it's useless) or if it represents a 0 bytes range (when
2621          * cur_offset == drop_end).
2622          */
2623         if (cur_offset < ino_size && cur_offset < drop_end) {
2624                 ret = fill_holes(trans, BTRFS_I(inode), path,
2625                                 cur_offset, drop_end);
2626                 if (ret) {
2627                         /* Same comment as above. */
2628                         btrfs_abort_transaction(trans, ret);
2629                         err = ret;
2630                         goto out_trans;
2631                 }
2632         }
2633
2634 out_trans:
2635         if (!trans)
2636                 goto out_free;
2637
2638         inode_inc_iversion(inode);
2639         inode->i_mtime = inode->i_ctime = current_time(inode);
2640
2641         trans->block_rsv = &fs_info->trans_block_rsv;
2642         ret = btrfs_update_inode(trans, root, inode);
2643         updated_inode = true;
2644         btrfs_end_transaction(trans);
2645         btrfs_btree_balance_dirty(fs_info);
2646 out_free:
2647         btrfs_free_path(path);
2648         btrfs_free_block_rsv(fs_info, rsv);
2649 out:
2650         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2651                              &cached_state, GFP_NOFS);
2652 out_only_mutex:
2653         if (!updated_inode && truncated_block && !ret && !err) {
2654                 /*
2655                  * If we only end up zeroing part of a page, we still need to
2656                  * update the inode item, so that all the time fields are
2657                  * updated as well as the necessary btrfs inode in memory fields
2658                  * for detecting, at fsync time, if the inode isn't yet in the
2659                  * log tree or it's there but not up to date.
2660                  */
2661                 trans = btrfs_start_transaction(root, 1);
2662                 if (IS_ERR(trans)) {
2663                         err = PTR_ERR(trans);
2664                 } else {
2665                         err = btrfs_update_inode(trans, root, inode);
2666                         ret = btrfs_end_transaction(trans);
2667                 }
2668         }
2669         inode_unlock(inode);
2670         if (ret && !err)
2671                 err = ret;
2672         return err;
2673 }
2674
2675 /* Helper structure to record which range is already reserved */
2676 struct falloc_range {
2677         struct list_head list;
2678         u64 start;
2679         u64 len;
2680 };
2681
2682 /*
2683  * Helper function to add falloc range
2684  *
2685  * Caller should have locked the larger range of extent containing
2686  * [start, len)
2687  */
2688 static int add_falloc_range(struct list_head *head, u64 start, u64 len)
2689 {
2690         struct falloc_range *prev = NULL;
2691         struct falloc_range *range = NULL;
2692
2693         if (list_empty(head))
2694                 goto insert;
2695
2696         /*
2697          * As fallocate iterate by bytenr order, we only need to check
2698          * the last range.
2699          */
2700         prev = list_entry(head->prev, struct falloc_range, list);
2701         if (prev->start + prev->len == start) {
2702                 prev->len += len;
2703                 return 0;
2704         }
2705 insert:
2706         range = kmalloc(sizeof(*range), GFP_KERNEL);
2707         if (!range)
2708                 return -ENOMEM;
2709         range->start = start;
2710         range->len = len;
2711         list_add_tail(&range->list, head);
2712         return 0;
2713 }
2714
2715 static long btrfs_fallocate(struct file *file, int mode,
2716                             loff_t offset, loff_t len)
2717 {
2718         struct inode *inode = file_inode(file);
2719         struct extent_state *cached_state = NULL;
2720         struct falloc_range *range;
2721         struct falloc_range *tmp;
2722         struct list_head reserve_list;
2723         u64 cur_offset;
2724         u64 last_byte;
2725         u64 alloc_start;
2726         u64 alloc_end;
2727         u64 alloc_hint = 0;
2728         u64 locked_end;
2729         u64 actual_end = 0;
2730         struct extent_map *em;
2731         int blocksize = btrfs_inode_sectorsize(inode);
2732         int ret;
2733
2734         alloc_start = round_down(offset, blocksize);
2735         alloc_end = round_up(offset + len, blocksize);
2736         cur_offset = alloc_start;
2737
2738         /* Make sure we aren't being give some crap mode */
2739         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2740                 return -EOPNOTSUPP;
2741
2742         if (mode & FALLOC_FL_PUNCH_HOLE)
2743                 return btrfs_punch_hole(inode, offset, len);
2744
2745         /*
2746          * Only trigger disk allocation, don't trigger qgroup reserve
2747          *
2748          * For qgroup space, it will be checked later.
2749          */
2750         ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
2751                         alloc_end - alloc_start);
2752         if (ret < 0)
2753                 return ret;
2754
2755         inode_lock(inode);
2756
2757         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
2758                 ret = inode_newsize_ok(inode, offset + len);
2759                 if (ret)
2760                         goto out;
2761         }
2762
2763         /*
2764          * TODO: Move these two operations after we have checked
2765          * accurate reserved space, or fallocate can still fail but
2766          * with page truncated or size expanded.
2767          *
2768          * But that's a minor problem and won't do much harm BTW.
2769          */
2770         if (alloc_start > inode->i_size) {
2771                 ret = btrfs_cont_expand(inode, i_size_read(inode),
2772                                         alloc_start);
2773                 if (ret)
2774                         goto out;
2775         } else if (offset + len > inode->i_size) {
2776                 /*
2777                  * If we are fallocating from the end of the file onward we
2778                  * need to zero out the end of the block if i_size lands in the
2779                  * middle of a block.
2780                  */
2781                 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
2782                 if (ret)
2783                         goto out;
2784         }
2785
2786         /*
2787          * wait for ordered IO before we have any locks.  We'll loop again
2788          * below with the locks held.
2789          */
2790         ret = btrfs_wait_ordered_range(inode, alloc_start,
2791                                        alloc_end - alloc_start);
2792         if (ret)
2793                 goto out;
2794
2795         locked_end = alloc_end - 1;
2796         while (1) {
2797                 struct btrfs_ordered_extent *ordered;
2798
2799                 /* the extent lock is ordered inside the running
2800                  * transaction
2801                  */
2802                 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
2803                                  locked_end, &cached_state);
2804                 ordered = btrfs_lookup_first_ordered_extent(inode,
2805                                                             alloc_end - 1);
2806                 if (ordered &&
2807                     ordered->file_offset + ordered->len > alloc_start &&
2808                     ordered->file_offset < alloc_end) {
2809                         btrfs_put_ordered_extent(ordered);
2810                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
2811                                              alloc_start, locked_end,
2812                                              &cached_state, GFP_KERNEL);
2813                         /*
2814                          * we can't wait on the range with the transaction
2815                          * running or with the extent lock held
2816                          */
2817                         ret = btrfs_wait_ordered_range(inode, alloc_start,
2818                                                        alloc_end - alloc_start);
2819                         if (ret)
2820                                 goto out;
2821                 } else {
2822                         if (ordered)
2823                                 btrfs_put_ordered_extent(ordered);
2824                         break;
2825                 }
2826         }
2827
2828         /* First, check if we exceed the qgroup limit */
2829         INIT_LIST_HEAD(&reserve_list);
2830         while (1) {
2831                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
2832                                       alloc_end - cur_offset, 0);
2833                 if (IS_ERR(em)) {
2834                         ret = PTR_ERR(em);
2835                         break;
2836                 }
2837                 last_byte = min(extent_map_end(em), alloc_end);
2838                 actual_end = min_t(u64, extent_map_end(em), offset + len);
2839                 last_byte = ALIGN(last_byte, blocksize);
2840                 if (em->block_start == EXTENT_MAP_HOLE ||
2841                     (cur_offset >= inode->i_size &&
2842                      !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
2843                         ret = add_falloc_range(&reserve_list, cur_offset,
2844                                                last_byte - cur_offset);
2845                         if (ret < 0) {
2846                                 free_extent_map(em);
2847                                 break;
2848                         }
2849                         ret = btrfs_qgroup_reserve_data(inode, cur_offset,
2850                                         last_byte - cur_offset);
2851                         if (ret < 0)
2852                                 break;
2853                 } else {
2854                         /*
2855                          * Do not need to reserve unwritten extent for this
2856                          * range, free reserved data space first, otherwise
2857                          * it'll result in false ENOSPC error.
2858                          */
2859                         btrfs_free_reserved_data_space(inode, cur_offset,
2860                                 last_byte - cur_offset);
2861                 }
2862                 free_extent_map(em);
2863                 cur_offset = last_byte;
2864                 if (cur_offset >= alloc_end)
2865                         break;
2866         }
2867
2868         /*
2869          * If ret is still 0, means we're OK to fallocate.
2870          * Or just cleanup the list and exit.
2871          */
2872         list_for_each_entry_safe(range, tmp, &reserve_list, list) {
2873                 if (!ret)
2874                         ret = btrfs_prealloc_file_range(inode, mode,
2875                                         range->start,
2876                                         range->len, i_blocksize(inode),
2877                                         offset + len, &alloc_hint);
2878                 else
2879                         btrfs_free_reserved_data_space(inode, range->start,
2880                                                        range->len);
2881                 list_del(&range->list);
2882                 kfree(range);
2883         }
2884         if (ret < 0)
2885                 goto out_unlock;
2886
2887         if (actual_end > inode->i_size &&
2888             !(mode & FALLOC_FL_KEEP_SIZE)) {
2889                 struct btrfs_trans_handle *trans;
2890                 struct btrfs_root *root = BTRFS_I(inode)->root;
2891
2892                 /*
2893                  * We didn't need to allocate any more space, but we
2894                  * still extended the size of the file so we need to
2895                  * update i_size and the inode item.
2896                  */
2897                 trans = btrfs_start_transaction(root, 1);
2898                 if (IS_ERR(trans)) {
2899                         ret = PTR_ERR(trans);
2900                 } else {
2901                         inode->i_ctime = current_time(inode);
2902                         i_size_write(inode, actual_end);
2903                         btrfs_ordered_update_i_size(inode, actual_end, NULL);
2904                         ret = btrfs_update_inode(trans, root, inode);
2905                         if (ret)
2906                                 btrfs_end_transaction(trans);
2907                         else
2908                                 ret = btrfs_end_transaction(trans);
2909                 }
2910         }
2911 out_unlock:
2912         unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
2913                              &cached_state, GFP_KERNEL);
2914 out:
2915         inode_unlock(inode);
2916         /* Let go of our reservation. */
2917         if (ret != 0)
2918                 btrfs_free_reserved_data_space(inode, alloc_start,
2919                                        alloc_end - cur_offset);
2920         return ret;
2921 }
2922
2923 static int find_desired_extent(struct inode *inode, loff_t *offset, int whence)
2924 {
2925         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2926         struct extent_map *em = NULL;
2927         struct extent_state *cached_state = NULL;
2928         u64 lockstart;
2929         u64 lockend;
2930         u64 start;
2931         u64 len;
2932         int ret = 0;
2933
2934         if (inode->i_size == 0)
2935                 return -ENXIO;
2936
2937         /*
2938          * *offset can be negative, in this case we start finding DATA/HOLE from
2939          * the very start of the file.
2940          */
2941         start = max_t(loff_t, 0, *offset);
2942
2943         lockstart = round_down(start, fs_info->sectorsize);
2944         lockend = round_up(i_size_read(inode),
2945                            fs_info->sectorsize);
2946         if (lockend <= lockstart)
2947                 lockend = lockstart + fs_info->sectorsize;
2948         lockend--;
2949         len = lockend - lockstart + 1;
2950
2951         lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2952                          &cached_state);
2953
2954         while (start < inode->i_size) {
2955                 em = btrfs_get_extent_fiemap(BTRFS_I(inode), NULL, 0,
2956                                 start, len, 0);
2957                 if (IS_ERR(em)) {
2958                         ret = PTR_ERR(em);
2959                         em = NULL;
2960                         break;
2961                 }
2962
2963                 if (whence == SEEK_HOLE &&
2964                     (em->block_start == EXTENT_MAP_HOLE ||
2965                      test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
2966                         break;
2967                 else if (whence == SEEK_DATA &&
2968                            (em->block_start != EXTENT_MAP_HOLE &&
2969                             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
2970                         break;
2971
2972                 start = em->start + em->len;
2973                 free_extent_map(em);
2974                 em = NULL;
2975                 cond_resched();
2976         }
2977         free_extent_map(em);
2978         if (!ret) {
2979                 if (whence == SEEK_DATA && start >= inode->i_size)
2980                         ret = -ENXIO;
2981                 else
2982                         *offset = min_t(loff_t, start, inode->i_size);
2983         }
2984         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2985                              &cached_state, GFP_NOFS);
2986         return ret;
2987 }
2988
2989 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
2990 {
2991         struct inode *inode = file->f_mapping->host;
2992         int ret;
2993
2994         inode_lock(inode);
2995         switch (whence) {
2996         case SEEK_END:
2997         case SEEK_CUR:
2998                 offset = generic_file_llseek(file, offset, whence);
2999                 goto out;
3000         case SEEK_DATA:
3001         case SEEK_HOLE:
3002                 if (offset >= i_size_read(inode)) {
3003                         inode_unlock(inode);
3004                         return -ENXIO;
3005                 }
3006
3007                 ret = find_desired_extent(inode, &offset, whence);
3008                 if (ret) {
3009                         inode_unlock(inode);
3010                         return ret;
3011                 }
3012         }
3013
3014         offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3015 out:
3016         inode_unlock(inode);
3017         return offset;
3018 }
3019
3020 const struct file_operations btrfs_file_operations = {
3021         .llseek         = btrfs_file_llseek,
3022         .read_iter      = generic_file_read_iter,
3023         .splice_read    = generic_file_splice_read,
3024         .write_iter     = btrfs_file_write_iter,
3025         .mmap           = btrfs_file_mmap,
3026         .open           = generic_file_open,
3027         .release        = btrfs_release_file,
3028         .fsync          = btrfs_sync_file,
3029         .fallocate      = btrfs_fallocate,
3030         .unlocked_ioctl = btrfs_ioctl,
3031 #ifdef CONFIG_COMPAT
3032         .compat_ioctl   = btrfs_compat_ioctl,
3033 #endif
3034         .clone_file_range = btrfs_clone_file_range,
3035         .dedupe_file_range = btrfs_dedupe_file_range,
3036 };
3037
3038 void btrfs_auto_defrag_exit(void)
3039 {
3040         kmem_cache_destroy(btrfs_inode_defrag_cachep);
3041 }
3042
3043 int btrfs_auto_defrag_init(void)
3044 {
3045         btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
3046                                         sizeof(struct inode_defrag), 0,
3047                                         SLAB_MEM_SPREAD,
3048                                         NULL);
3049         if (!btrfs_inode_defrag_cachep)
3050                 return -ENOMEM;
3051
3052         return 0;
3053 }
3054
3055 int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3056 {
3057         int ret;
3058
3059         /*
3060          * So with compression we will find and lock a dirty page and clear the
3061          * first one as dirty, setup an async extent, and immediately return
3062          * with the entire range locked but with nobody actually marked with
3063          * writeback.  So we can't just filemap_write_and_wait_range() and
3064          * expect it to work since it will just kick off a thread to do the
3065          * actual work.  So we need to call filemap_fdatawrite_range _again_
3066          * since it will wait on the page lock, which won't be unlocked until
3067          * after the pages have been marked as writeback and so we're good to go
3068          * from there.  We have to do this otherwise we'll miss the ordered
3069          * extents and that results in badness.  Please Josef, do not think you
3070          * know better and pull this out at some point in the future, it is
3071          * right and you are wrong.
3072          */
3073         ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3074         if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3075                              &BTRFS_I(inode)->runtime_flags))
3076                 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3077
3078         return ret;
3079 }