fs/btrfs/free-space-cache.c

   1 /*
   2  * Copyright (C) 2008 Red Hat.  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/pagemap.h>
  20 #include <linux/sched.h>
  21 #include <linux/slab.h>
  22 #include <linux/math64.h>
  23 #include <linux/ratelimit.h>
  24 #include "ctree.h"
  25 #include "free-space-cache.h"
  26 #include "transaction.h"
  27 #include "disk-io.h"
  28 #include "extent_io.h"
  29 #include "inode-map.h"
  30 #include "volumes.h"
  31
  32 #define BITS_PER_BITMAP         (PAGE_SIZE * 8UL)
  33 #define MAX_CACHE_BYTES_PER_GIG SZ_32K
  34
  35 struct btrfs_trim_range {
  36         u64 start;
  37         u64 bytes;
  38         struct list_head list;
  39 };
  40
  41 static int link_free_space(struct btrfs_free_space_ctl *ctl,
  42                            struct btrfs_free_space *info);
  43 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
  44                               struct btrfs_free_space *info);
  45
  46 static struct inode *__lookup_free_space_inode(struct btrfs_root *root,
  47                                                struct btrfs_path *path,
  48                                                u64 offset)
  49 {
  50         struct btrfs_key key;
  51         struct btrfs_key location;
  52         struct btrfs_disk_key disk_key;
  53         struct btrfs_free_space_header *header;
  54         struct extent_buffer *leaf;
  55         struct inode *inode = NULL;
  56         int ret;
  57
  58         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
  59         key.offset = offset;
  60         key.type = 0;
  61
  62         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
  63         if (ret < 0)
  64                 return ERR_PTR(ret);
  65         if (ret > 0) {
  66                 btrfs_release_path(path);
  67                 return ERR_PTR(-ENOENT);
  68         }
  69
  70         leaf = path->nodes[0];
  71         header = btrfs_item_ptr(leaf, path->slots[0],
  72                                 struct btrfs_free_space_header);
  73         btrfs_free_space_key(leaf, header, &disk_key);
  74         btrfs_disk_key_to_cpu(&location, &disk_key);
  75         btrfs_release_path(path);
  76
  77         inode = btrfs_iget(root->fs_info->sb, &location, root, NULL);
  78         if (IS_ERR(inode))
  79                 return inode;
  80         if (is_bad_inode(inode)) {
  81                 iput(inode);
  82                 return ERR_PTR(-ENOENT);
  83         }
  84
  85         mapping_set_gfp_mask(inode->i_mapping,
  86                         mapping_gfp_constraint(inode->i_mapping,
  87                         ~(__GFP_FS | __GFP_HIGHMEM)));
  88
  89         return inode;
  90 }
  91
  92 struct inode *lookup_free_space_inode(struct btrfs_root *root,
  93                                       struct btrfs_block_group_cache
  94                                       *block_group, struct btrfs_path *path)
  95 {
  96         struct inode *inode = NULL;
  97         u32 flags = BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
  98
  99         spin_lock(&block_group->lock);
 100         if (block_group->inode)
 101                 inode = igrab(block_group->inode);
 102         spin_unlock(&block_group->lock);
 103         if (inode)
 104                 return inode;
 105
 106         inode = __lookup_free_space_inode(root, path,
 107                                           block_group->key.objectid);
 108         if (IS_ERR(inode))
 109                 return inode;
 110
 111         spin_lock(&block_group->lock);
 112         if (!((BTRFS_I(inode)->flags & flags) == flags)) {
 113                 btrfs_info(root->fs_info,
 114                         "Old style space inode found, converting.");
 115                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM |
 116                         BTRFS_INODE_NODATACOW;
 117                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 118         }
 119
 120         if (!block_group->iref) {
 121                 block_group->inode = igrab(inode);
 122                 block_group->iref = 1;
 123         }
 124         spin_unlock(&block_group->lock);
 125
 126         return inode;
 127 }
 128
 129 static int __create_free_space_inode(struct btrfs_root *root,
 130                                      struct btrfs_trans_handle *trans,
 131                                      struct btrfs_path *path,
 132                                      u64 ino, u64 offset)
 133 {
 134         struct btrfs_key key;
 135         struct btrfs_disk_key disk_key;
 136         struct btrfs_free_space_header *header;
 137         struct btrfs_inode_item *inode_item;
 138         struct extent_buffer *leaf;
 139         u64 flags = BTRFS_INODE_NOCOMPRESS | BTRFS_INODE_PREALLOC;
 140         int ret;
 141
 142         ret = btrfs_insert_empty_inode(trans, root, path, ino);
 143         if (ret)
 144                 return ret;
 145
 146         /* We inline crc's for the free disk space cache */
 147         if (ino != BTRFS_FREE_INO_OBJECTID)
 148                 flags |= BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
 149
 150         leaf = path->nodes[0];
 151         inode_item = btrfs_item_ptr(leaf, path->slots[0],
 152                                     struct btrfs_inode_item);
 153         btrfs_item_key(leaf, &disk_key, path->slots[0]);
 154         memzero_extent_buffer(leaf, (unsigned long)inode_item,
 155                              sizeof(*inode_item));
 156         btrfs_set_inode_generation(leaf, inode_item, trans->transid);
 157         btrfs_set_inode_size(leaf, inode_item, 0);
 158         btrfs_set_inode_nbytes(leaf, inode_item, 0);
 159         btrfs_set_inode_uid(leaf, inode_item, 0);
 160         btrfs_set_inode_gid(leaf, inode_item, 0);
 161         btrfs_set_inode_mode(leaf, inode_item, S_IFREG | 0600);
 162         btrfs_set_inode_flags(leaf, inode_item, flags);
 163         btrfs_set_inode_nlink(leaf, inode_item, 1);
 164         btrfs_set_inode_transid(leaf, inode_item, trans->transid);
 165         btrfs_set_inode_block_group(leaf, inode_item, offset);
 166         btrfs_mark_buffer_dirty(leaf);
 167         btrfs_release_path(path);
 168
 169         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 170         key.offset = offset;
 171         key.type = 0;
 172         ret = btrfs_insert_empty_item(trans, root, path, &key,
 173                                       sizeof(struct btrfs_free_space_header));
 174         if (ret < 0) {
 175                 btrfs_release_path(path);
 176                 return ret;
 177         }
 178
 179         leaf = path->nodes[0];
 180         header = btrfs_item_ptr(leaf, path->slots[0],
 181                                 struct btrfs_free_space_header);
 182         memzero_extent_buffer(leaf, (unsigned long)header, sizeof(*header));
 183         btrfs_set_free_space_key(leaf, header, &disk_key);
 184         btrfs_mark_buffer_dirty(leaf);
 185         btrfs_release_path(path);
 186
 187         return 0;
 188 }
 189
 190 int create_free_space_inode(struct btrfs_root *root,
 191                             struct btrfs_trans_handle *trans,
 192                             struct btrfs_block_group_cache *block_group,
 193                             struct btrfs_path *path)
 194 {
 195         int ret;
 196         u64 ino;
 197
 198         ret = btrfs_find_free_objectid(root, &ino);
 199         if (ret < 0)
 200                 return ret;
 201
 202         return __create_free_space_inode(root, trans, path, ino,
 203                                          block_group->key.objectid);
 204 }
 205
 206 int btrfs_check_trunc_cache_free_space(struct btrfs_root *root,
 207                                        struct btrfs_block_rsv *rsv)
 208 {
 209         u64 needed_bytes;
 210         int ret;
 211
 212         /* 1 for slack space, 1 for updating the inode */
 213         needed_bytes = btrfs_calc_trunc_metadata_size(root, 1) +
 214                 btrfs_calc_trans_metadata_size(root, 1);
 215
 216         spin_lock(&rsv->lock);
 217         if (rsv->reserved < needed_bytes)
 218                 ret = -ENOSPC;
 219         else
 220                 ret = 0;
 221         spin_unlock(&rsv->lock);
 222         return ret;
 223 }
 224
 225 int btrfs_truncate_free_space_cache(struct btrfs_root *root,
 226                                     struct btrfs_trans_handle *trans,
 227                                     struct btrfs_block_group_cache *block_group,
 228                                     struct inode *inode)
 229 {
 230         int ret = 0;
 231         struct btrfs_path *path = btrfs_alloc_path();
 232         bool locked = false;
 233
 234         if (!path) {
 235                 ret = -ENOMEM;
 236                 goto fail;
 237         }
 238
 239         if (block_group) {
 240                 locked = true;
 241                 mutex_lock(&trans->transaction->cache_write_mutex);
 242                 if (!list_empty(&block_group->io_list)) {
 243                         list_del_init(&block_group->io_list);
 244
 245                         btrfs_wait_cache_io(root, trans, block_group,
 246                                             &block_group->io_ctl, path,
 247                                             block_group->key.objectid);
 248                         btrfs_put_block_group(block_group);
 249                 }
 250
 251                 /*
 252                  * now that we've truncated the cache away, its no longer
 253                  * setup or written
 254                  */
 255                 spin_lock(&block_group->lock);
 256                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 257                 spin_unlock(&block_group->lock);
 258         }
 259         btrfs_free_path(path);
 260
 261         btrfs_i_size_write(inode, 0);
 262         truncate_pagecache(inode, 0);
 263
 264         /*
 265          * We don't need an orphan item because truncating the free space cache
 266          * will never be split across transactions.
 267          * We don't need to check for -EAGAIN because we're a free space
 268          * cache inode
 269          */
 270         ret = btrfs_truncate_inode_items(trans, root, inode,
 271                                          0, BTRFS_EXTENT_DATA_KEY);
 272         if (ret)
 273                 goto fail;
 274
 275         ret = btrfs_update_inode(trans, root, inode);
 276
 277 fail:
 278         if (locked)
 279                 mutex_unlock(&trans->transaction->cache_write_mutex);
 280         if (ret)
 281                 btrfs_abort_transaction(trans, ret);
 282
 283         return ret;
 284 }
 285
 286 static int readahead_cache(struct inode *inode)
 287 {
 288         struct file_ra_state *ra;
 289         unsigned long last_index;
 290
 291         ra = kzalloc(sizeof(*ra), GFP_NOFS);
 292         if (!ra)
 293                 return -ENOMEM;
 294
 295         file_ra_state_init(ra, inode->i_mapping);
 296         last_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
 297
 298         page_cache_sync_readahead(inode->i_mapping, ra, NULL, 0, last_index);
 299
 300         kfree(ra);
 301
 302         return 0;
 303 }
 304
 305 static int io_ctl_init(struct btrfs_io_ctl *io_ctl, struct inode *inode,
 306                        struct btrfs_root *root, int write)
 307 {
 308         int num_pages;
 309         int check_crcs = 0;
 310
 311         num_pages = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
 312
 313         if (btrfs_ino(inode) != BTRFS_FREE_INO_OBJECTID)
 314                 check_crcs = 1;
 315
 316         /* Make sure we can fit our crcs into the first page */
 317         if (write && check_crcs &&
 318             (num_pages * sizeof(u32)) >= PAGE_SIZE)
 319                 return -ENOSPC;
 320
 321         memset(io_ctl, 0, sizeof(struct btrfs_io_ctl));
 322
 323         io_ctl->pages = kcalloc(num_pages, sizeof(struct page *), GFP_NOFS);
 324         if (!io_ctl->pages)
 325                 return -ENOMEM;
 326
 327         io_ctl->num_pages = num_pages;
 328         io_ctl->root = root;
 329         io_ctl->check_crcs = check_crcs;
 330         io_ctl->inode = inode;
 331
 332         return 0;
 333 }
 334
 335 static void io_ctl_free(struct btrfs_io_ctl *io_ctl)
 336 {
 337         kfree(io_ctl->pages);
 338         io_ctl->pages = NULL;
 339 }
 340
 341 static void io_ctl_unmap_page(struct btrfs_io_ctl *io_ctl)
 342 {
 343         if (io_ctl->cur) {
 344                 io_ctl->cur = NULL;
 345                 io_ctl->orig = NULL;
 346         }
 347 }
 348
 349 static void io_ctl_map_page(struct btrfs_io_ctl *io_ctl, int clear)
 350 {
 351         ASSERT(io_ctl->index < io_ctl->num_pages);
 352         io_ctl->page = io_ctl->pages[io_ctl->index++];
 353         io_ctl->cur = page_address(io_ctl->page);
 354         io_ctl->orig = io_ctl->cur;
 355         io_ctl->size = PAGE_SIZE;
 356         if (clear)
 357                 memset(io_ctl->cur, 0, PAGE_SIZE);
 358 }
 359
 360 static void io_ctl_drop_pages(struct btrfs_io_ctl *io_ctl)
 361 {
 362         int i;
 363
 364         io_ctl_unmap_page(io_ctl);
 365
 366         for (i = 0; i < io_ctl->num_pages; i++) {
 367                 if (io_ctl->pages[i]) {
 368                         ClearPageChecked(io_ctl->pages[i]);
 369                         unlock_page(io_ctl->pages[i]);
 370                         put_page(io_ctl->pages[i]);
 371                 }
 372         }
 373 }
 374
 375 static int io_ctl_prepare_pages(struct btrfs_io_ctl *io_ctl, struct inode *inode,
 376                                 int uptodate)
 377 {
 378         struct page *page;
 379         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
 380         int i;
 381
 382         for (i = 0; i < io_ctl->num_pages; i++) {
 383                 page = find_or_create_page(inode->i_mapping, i, mask);
 384                 if (!page) {
 385                         io_ctl_drop_pages(io_ctl);
 386                         return -ENOMEM;
 387                 }
 388                 io_ctl->pages[i] = page;
 389                 if (uptodate && !PageUptodate(page)) {
 390                         btrfs_readpage(NULL, page);
 391                         lock_page(page);
 392                         if (!PageUptodate(page)) {
 393                                 btrfs_err(BTRFS_I(inode)->root->fs_info,
 394                                            "error reading free space cache");
 395                                 io_ctl_drop_pages(io_ctl);
 396                                 return -EIO;
 397                         }
 398                 }
 399         }
 400
 401         for (i = 0; i < io_ctl->num_pages; i++) {
 402                 clear_page_dirty_for_io(io_ctl->pages[i]);
 403                 set_page_extent_mapped(io_ctl->pages[i]);
 404         }
 405
 406         return 0;
 407 }
 408
 409 static void io_ctl_set_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
 410 {
 411         __le64 *val;
 412
 413         io_ctl_map_page(io_ctl, 1);
 414
 415         /*
 416          * Skip the csum areas.  If we don't check crcs then we just have a
 417          * 64bit chunk at the front of the first page.
 418          */
 419         if (io_ctl->check_crcs) {
 420                 io_ctl->cur += (sizeof(u32) * io_ctl->num_pages);
 421                 io_ctl->size -= sizeof(u64) + (sizeof(u32) * io_ctl->num_pages);
 422         } else {
 423                 io_ctl->cur += sizeof(u64);
 424                 io_ctl->size -= sizeof(u64) * 2;
 425         }
 426
 427         val = io_ctl->cur;
 428         *val = cpu_to_le64(generation);
 429         io_ctl->cur += sizeof(u64);
 430 }
 431
 432 static int io_ctl_check_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
 433 {
 434         __le64 *gen;
 435
 436         /*
 437          * Skip the crc area.  If we don't check crcs then we just have a 64bit
 438          * chunk at the front of the first page.
 439          */
 440         if (io_ctl->check_crcs) {
 441                 io_ctl->cur += sizeof(u32) * io_ctl->num_pages;
 442                 io_ctl->size -= sizeof(u64) +
 443                         (sizeof(u32) * io_ctl->num_pages);
 444         } else {
 445                 io_ctl->cur += sizeof(u64);
 446                 io_ctl->size -= sizeof(u64) * 2;
 447         }
 448
 449         gen = io_ctl->cur;
 450         if (le64_to_cpu(*gen) != generation) {
 451                 btrfs_err_rl(io_ctl->root->fs_info,
 452                         "space cache generation (%llu) does not match inode (%llu)",
 453                                 *gen, generation);
 454                 io_ctl_unmap_page(io_ctl);
 455                 return -EIO;
 456         }
 457         io_ctl->cur += sizeof(u64);
 458         return 0;
 459 }
 460
 461 static void io_ctl_set_crc(struct btrfs_io_ctl *io_ctl, int index)
 462 {
 463         u32 *tmp;
 464         u32 crc = ~(u32)0;
 465         unsigned offset = 0;
 466
 467         if (!io_ctl->check_crcs) {
 468                 io_ctl_unmap_page(io_ctl);
 469                 return;
 470         }
 471
 472         if (index == 0)
 473                 offset = sizeof(u32) * io_ctl->num_pages;
 474
 475         crc = btrfs_csum_data(io_ctl->orig + offset, crc,
 476                               PAGE_SIZE - offset);
 477         btrfs_csum_final(crc, (u8 *)&crc);
 478         io_ctl_unmap_page(io_ctl);
 479         tmp = page_address(io_ctl->pages[0]);
 480         tmp += index;
 481         *tmp = crc;
 482 }
 483
 484 static int io_ctl_check_crc(struct btrfs_io_ctl *io_ctl, int index)
 485 {
 486         u32 *tmp, val;
 487         u32 crc = ~(u32)0;
 488         unsigned offset = 0;
 489
 490         if (!io_ctl->check_crcs) {
 491                 io_ctl_map_page(io_ctl, 0);
 492                 return 0;
 493         }
 494
 495         if (index == 0)
 496                 offset = sizeof(u32) * io_ctl->num_pages;
 497
 498         tmp = page_address(io_ctl->pages[0]);
 499         tmp += index;
 500         val = *tmp;
 501
 502         io_ctl_map_page(io_ctl, 0);
 503         crc = btrfs_csum_data(io_ctl->orig + offset, crc,
 504                               PAGE_SIZE - offset);
 505         btrfs_csum_final(crc, (u8 *)&crc);
 506         if (val != crc) {
 507                 btrfs_err_rl(io_ctl->root->fs_info,
 508                         "csum mismatch on free space cache");
 509                 io_ctl_unmap_page(io_ctl);
 510                 return -EIO;
 511         }
 512
 513         return 0;
 514 }
 515
 516 static int io_ctl_add_entry(struct btrfs_io_ctl *io_ctl, u64 offset, u64 bytes,
 517                             void *bitmap)
 518 {
 519         struct btrfs_free_space_entry *entry;
 520
 521         if (!io_ctl->cur)
 522                 return -ENOSPC;
 523
 524         entry = io_ctl->cur;
 525         entry->offset = cpu_to_le64(offset);
 526         entry->bytes = cpu_to_le64(bytes);
 527         entry->type = (bitmap) ? BTRFS_FREE_SPACE_BITMAP :
 528                 BTRFS_FREE_SPACE_EXTENT;
 529         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
 530         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
 531
 532         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
 533                 return 0;
 534
 535         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 536
 537         /* No more pages to map */
 538         if (io_ctl->index >= io_ctl->num_pages)
 539                 return 0;
 540
 541         /* map the next page */
 542         io_ctl_map_page(io_ctl, 1);
 543         return 0;
 544 }
 545
 546 static int io_ctl_add_bitmap(struct btrfs_io_ctl *io_ctl, void *bitmap)
 547 {
 548         if (!io_ctl->cur)
 549                 return -ENOSPC;
 550
 551         /*
 552          * If we aren't at the start of the current page, unmap this one and
 553          * map the next one if there is any left.
 554          */
 555         if (io_ctl->cur != io_ctl->orig) {
 556                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 557                 if (io_ctl->index >= io_ctl->num_pages)
 558                         return -ENOSPC;
 559                 io_ctl_map_page(io_ctl, 0);
 560         }
 561
 562         memcpy(io_ctl->cur, bitmap, PAGE_SIZE);
 563         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 564         if (io_ctl->index < io_ctl->num_pages)
 565                 io_ctl_map_page(io_ctl, 0);
 566         return 0;
 567 }
 568
 569 static void io_ctl_zero_remaining_pages(struct btrfs_io_ctl *io_ctl)
 570 {
 571         /*
 572          * If we're not on the boundary we know we've modified the page and we
 573          * need to crc the page.
 574          */
 575         if (io_ctl->cur != io_ctl->orig)
 576                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 577         else
 578                 io_ctl_unmap_page(io_ctl);
 579
 580         while (io_ctl->index < io_ctl->num_pages) {
 581                 io_ctl_map_page(io_ctl, 1);
 582                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 583         }
 584 }
 585
 586 static int io_ctl_read_entry(struct btrfs_io_ctl *io_ctl,
 587                             struct btrfs_free_space *entry, u8 *type)
 588 {
 589         struct btrfs_free_space_entry *e;
 590         int ret;
 591
 592         if (!io_ctl->cur) {
 593                 ret = io_ctl_check_crc(io_ctl, io_ctl->index);
 594                 if (ret)
 595                         return ret;
 596         }
 597
 598         e = io_ctl->cur;
 599         entry->offset = le64_to_cpu(e->offset);
 600         entry->bytes = le64_to_cpu(e->bytes);
 601         *type = e->type;
 602         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
 603         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
 604
 605         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
 606                 return 0;
 607
 608         io_ctl_unmap_page(io_ctl);
 609
 610         return 0;
 611 }
 612
 613 static int io_ctl_read_bitmap(struct btrfs_io_ctl *io_ctl,
 614                               struct btrfs_free_space *entry)
 615 {
 616         int ret;
 617
 618         ret = io_ctl_check_crc(io_ctl, io_ctl->index);
 619         if (ret)
 620                 return ret;
 621
 622         memcpy(entry->bitmap, io_ctl->cur, PAGE_SIZE);
 623         io_ctl_unmap_page(io_ctl);
 624
 625         return 0;
 626 }
 627
 628 /*
 629  * Since we attach pinned extents after the fact we can have contiguous sections
 630  * of free space that are split up in entries.  This poses a problem with the
 631  * tree logging stuff since it could have allocated across what appears to be 2
 632  * entries since we would have merged the entries when adding the pinned extents
 633  * back to the free space cache.  So run through the space cache that we just
 634  * loaded and merge contiguous entries.  This will make the log replay stuff not
 635  * blow up and it will make for nicer allocator behavior.
 636  */
 637 static void merge_space_tree(struct btrfs_free_space_ctl *ctl)
 638 {
 639         struct btrfs_free_space *e, *prev = NULL;
 640         struct rb_node *n;
 641
 642 again:
 643         spin_lock(&ctl->tree_lock);
 644         for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
 645                 e = rb_entry(n, struct btrfs_free_space, offset_index);
 646                 if (!prev)
 647                         goto next;
 648                 if (e->bitmap || prev->bitmap)
 649                         goto next;
 650                 if (prev->offset + prev->bytes == e->offset) {
 651                         unlink_free_space(ctl, prev);
 652                         unlink_free_space(ctl, e);
 653                         prev->bytes += e->bytes;
 654                         kmem_cache_free(btrfs_free_space_cachep, e);
 655                         link_free_space(ctl, prev);
 656                         prev = NULL;
 657                         spin_unlock(&ctl->tree_lock);
 658                         goto again;
 659                 }
 660 next:
 661                 prev = e;
 662         }
 663         spin_unlock(&ctl->tree_lock);
 664 }
 665
 666 static int __load_free_space_cache(struct btrfs_root *root, struct inode *inode,
 667                                    struct btrfs_free_space_ctl *ctl,
 668                                    struct btrfs_path *path, u64 offset)
 669 {
 670         struct btrfs_free_space_header *header;
 671         struct extent_buffer *leaf;
 672         struct btrfs_io_ctl io_ctl;
 673         struct btrfs_key key;
 674         struct btrfs_free_space *e, *n;
 675         LIST_HEAD(bitmaps);
 676         u64 num_entries;
 677         u64 num_bitmaps;
 678         u64 generation;
 679         u8 type;
 680         int ret = 0;
 681
 682         /* Nothing in the space cache, goodbye */
 683         if (!i_size_read(inode))
 684                 return 0;
 685
 686         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 687         key.offset = offset;
 688         key.type = 0;
 689
 690         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
 691         if (ret < 0)
 692                 return 0;
 693         else if (ret > 0) {
 694                 btrfs_release_path(path);
 695                 return 0;
 696         }
 697
 698         ret = -1;
 699
 700         leaf = path->nodes[0];
 701         header = btrfs_item_ptr(leaf, path->slots[0],
 702                                 struct btrfs_free_space_header);
 703         num_entries = btrfs_free_space_entries(leaf, header);
 704         num_bitmaps = btrfs_free_space_bitmaps(leaf, header);
 705         generation = btrfs_free_space_generation(leaf, header);
 706         btrfs_release_path(path);
 707
 708         if (!BTRFS_I(inode)->generation) {
 709                 btrfs_info(root->fs_info,
 710                            "The free space cache file (%llu) is invalid. skip it\n",
 711                            offset);
 712                 return 0;
 713         }
 714
 715         if (BTRFS_I(inode)->generation != generation) {
 716                 btrfs_err(root->fs_info,
 717                         "free space inode generation (%llu) did not match free space cache generation (%llu)",
 718                         BTRFS_I(inode)->generation, generation);
 719                 return 0;
 720         }
 721
 722         if (!num_entries)
 723                 return 0;
 724
 725         ret = io_ctl_init(&io_ctl, inode, root, 0);
 726         if (ret)
 727                 return ret;
 728
 729         ret = readahead_cache(inode);
 730         if (ret)
 731                 goto out;
 732
 733         ret = io_ctl_prepare_pages(&io_ctl, inode, 1);
 734         if (ret)
 735                 goto out;
 736
 737         ret = io_ctl_check_crc(&io_ctl, 0);
 738         if (ret)
 739                 goto free_cache;
 740
 741         ret = io_ctl_check_generation(&io_ctl, generation);
 742         if (ret)
 743                 goto free_cache;
 744
 745         while (num_entries) {
 746                 e = kmem_cache_zalloc(btrfs_free_space_cachep,
 747                                       GFP_NOFS);
 748                 if (!e)
 749                         goto free_cache;
 750
 751                 ret = io_ctl_read_entry(&io_ctl, e, &type);
 752                 if (ret) {
 753                         kmem_cache_free(btrfs_free_space_cachep, e);
 754                         goto free_cache;
 755                 }
 756
 757                 if (!e->bytes) {
 758                         kmem_cache_free(btrfs_free_space_cachep, e);
 759                         goto free_cache;
 760                 }
 761
 762                 if (type == BTRFS_FREE_SPACE_EXTENT) {
 763                         spin_lock(&ctl->tree_lock);
 764                         ret = link_free_space(ctl, e);
 765                         spin_unlock(&ctl->tree_lock);
 766                         if (ret) {
 767                                 btrfs_err(root->fs_info,
 768                                         "Duplicate entries in free space cache, dumping");
 769                                 kmem_cache_free(btrfs_free_space_cachep, e);
 770                                 goto free_cache;
 771                         }
 772                 } else {
 773                         ASSERT(num_bitmaps);
 774                         num_bitmaps--;
 775                         e->bitmap = kzalloc(PAGE_SIZE, GFP_NOFS);
 776                         if (!e->bitmap) {
 777                                 kmem_cache_free(
 778                                         btrfs_free_space_cachep, e);
 779                                 goto free_cache;
 780                         }
 781                         spin_lock(&ctl->tree_lock);
 782                         ret = link_free_space(ctl, e);
 783                         ctl->total_bitmaps++;
 784                         ctl->op->recalc_thresholds(ctl);
 785                         spin_unlock(&ctl->tree_lock);
 786                         if (ret) {
 787                                 btrfs_err(root->fs_info,
 788                                         "Duplicate entries in free space cache, dumping");
 789                                 kmem_cache_free(btrfs_free_space_cachep, e);
 790                                 goto free_cache;
 791                         }
 792                         list_add_tail(&e->list, &bitmaps);
 793                 }
 794
 795                 num_entries--;
 796         }
 797
 798         io_ctl_unmap_page(&io_ctl);
 799
 800         /*
 801          * We add the bitmaps at the end of the entries in order that
 802          * the bitmap entries are added to the cache.
 803          */
 804         list_for_each_entry_safe(e, n, &bitmaps, list) {
 805                 list_del_init(&e->list);
 806                 ret = io_ctl_read_bitmap(&io_ctl, e);
 807                 if (ret)
 808                         goto free_cache;
 809         }
 810
 811         io_ctl_drop_pages(&io_ctl);
 812         merge_space_tree(ctl);
 813         ret = 1;
 814 out:
 815         io_ctl_free(&io_ctl);
 816         return ret;
 817 free_cache:
 818         io_ctl_drop_pages(&io_ctl);
 819         __btrfs_remove_free_space_cache(ctl);
 820         goto out;
 821 }
 822
 823 int load_free_space_cache(struct btrfs_fs_info *fs_info,
 824                           struct btrfs_block_group_cache *block_group)
 825 {
 826         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
 827         struct btrfs_root *root = fs_info->tree_root;
 828         struct inode *inode;
 829         struct btrfs_path *path;
 830         int ret = 0;
 831         bool matched;
 832         u64 used = btrfs_block_group_used(&block_group->item);
 833
 834         /*
 835          * If this block group has been marked to be cleared for one reason or
 836          * another then we can't trust the on disk cache, so just return.
 837          */
 838         spin_lock(&block_group->lock);
 839         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
 840                 spin_unlock(&block_group->lock);
 841                 return 0;
 842         }
 843         spin_unlock(&block_group->lock);
 844
 845         path = btrfs_alloc_path();
 846         if (!path)
 847                 return 0;
 848         path->search_commit_root = 1;
 849         path->skip_locking = 1;
 850
 851         inode = lookup_free_space_inode(root, block_group, path);
 852         if (IS_ERR(inode)) {
 853                 btrfs_free_path(path);
 854                 return 0;
 855         }
 856
 857         /* We may have converted the inode and made the cache invalid. */
 858         spin_lock(&block_group->lock);
 859         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
 860                 spin_unlock(&block_group->lock);
 861                 btrfs_free_path(path);
 862                 goto out;
 863         }
 864         spin_unlock(&block_group->lock);
 865
 866         ret = __load_free_space_cache(fs_info->tree_root, inode, ctl,
 867                                       path, block_group->key.objectid);
 868         btrfs_free_path(path);
 869         if (ret <= 0)
 870                 goto out;
 871
 872         spin_lock(&ctl->tree_lock);
 873         matched = (ctl->free_space == (block_group->key.offset - used -
 874                                        block_group->bytes_super));
 875         spin_unlock(&ctl->tree_lock);
 876
 877         if (!matched) {
 878                 __btrfs_remove_free_space_cache(ctl);
 879                 btrfs_warn(fs_info,
 880                            "block group %llu has wrong amount of free space",
 881                            block_group->key.objectid);
 882                 ret = -1;
 883         }
 884 out:
 885         if (ret < 0) {
 886                 /* This cache is bogus, make sure it gets cleared */
 887                 spin_lock(&block_group->lock);
 888                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 889                 spin_unlock(&block_group->lock);
 890                 ret = 0;
 891
 892                 btrfs_warn(fs_info,
 893                            "failed to load free space cache for block group %llu, rebuilding it now",
 894                            block_group->key.objectid);
 895         }
 896
 897         iput(inode);
 898         return ret;
 899 }
 900
 901 static noinline_for_stack
 902 int write_cache_extent_entries(struct btrfs_io_ctl *io_ctl,
 903                               struct btrfs_free_space_ctl *ctl,
 904                               struct btrfs_block_group_cache *block_group,
 905                               int *entries, int *bitmaps,
 906                               struct list_head *bitmap_list)
 907 {
 908         int ret;
 909         struct btrfs_free_cluster *cluster = NULL;
 910         struct btrfs_free_cluster *cluster_locked = NULL;
 911         struct rb_node *node = rb_first(&ctl->free_space_offset);
 912         struct btrfs_trim_range *trim_entry;
 913
 914         /* Get the cluster for this block_group if it exists */
 915         if (block_group && !list_empty(&block_group->cluster_list)) {
 916                 cluster = list_entry(block_group->cluster_list.next,
 917                                      struct btrfs_free_cluster,
 918                                      block_group_list);
 919         }
 920
 921         if (!node && cluster) {
 922                 cluster_locked = cluster;
 923                 spin_lock(&cluster_locked->lock);
 924                 node = rb_first(&cluster->root);
 925                 cluster = NULL;
 926         }
 927
 928         /* Write out the extent entries */
 929         while (node) {
 930                 struct btrfs_free_space *e;
 931
 932                 e = rb_entry(node, struct btrfs_free_space, offset_index);
 933                 *entries += 1;
 934
 935                 ret = io_ctl_add_entry(io_ctl, e->offset, e->bytes,
 936                                        e->bitmap);
 937                 if (ret)
 938                         goto fail;
 939
 940                 if (e->bitmap) {
 941                         list_add_tail(&e->list, bitmap_list);
 942                         *bitmaps += 1;
 943                 }
 944                 node = rb_next(node);
 945                 if (!node && cluster) {
 946                         node = rb_first(&cluster->root);
 947                         cluster_locked = cluster;
 948                         spin_lock(&cluster_locked->lock);
 949                         cluster = NULL;
 950                 }
 951         }
 952         if (cluster_locked) {
 953                 spin_unlock(&cluster_locked->lock);
 954                 cluster_locked = NULL;
 955         }
 956
 957         /*
 958          * Make sure we don't miss any range that was removed from our rbtree
 959          * because trimming is running. Otherwise after a umount+mount (or crash
 960          * after committing the transaction) we would leak free space and get
 961          * an inconsistent free space cache report from fsck.
 962          */
 963         list_for_each_entry(trim_entry, &ctl->trimming_ranges, list) {
 964                 ret = io_ctl_add_entry(io_ctl, trim_entry->start,
 965                                        trim_entry->bytes, NULL);
 966                 if (ret)
 967                         goto fail;
 968                 *entries += 1;
 969         }
 970
 971         return 0;
 972 fail:
 973         if (cluster_locked)
 974                 spin_unlock(&cluster_locked->lock);
 975         return -ENOSPC;
 976 }
 977
 978 static noinline_for_stack int
 979 update_cache_item(struct btrfs_trans_handle *trans,
 980                   struct btrfs_root *root,
 981                   struct inode *inode,
 982                   struct btrfs_path *path, u64 offset,
 983                   int entries, int bitmaps)
 984 {
 985         struct btrfs_key key;
 986         struct btrfs_free_space_header *header;
 987         struct extent_buffer *leaf;
 988         int ret;
 989
 990         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 991         key.offset = offset;
 992         key.type = 0;
 993
 994         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
 995         if (ret < 0) {
 996                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
 997                                  EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0, NULL,
 998                                  GFP_NOFS);
 999                 goto fail;
1000         }
1001         leaf = path->nodes[0];
1002         if (ret > 0) {
1003                 struct btrfs_key found_key;
1004                 ASSERT(path->slots[0]);
1005                 path->slots[0]--;
1006                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1007                 if (found_key.objectid != BTRFS_FREE_SPACE_OBJECTID ||
1008                     found_key.offset != offset) {
1009                         clear_extent_bit(&BTRFS_I(inode)->io_tree, 0,
1010                                          inode->i_size - 1,
1011                                          EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0,
1012                                          NULL, GFP_NOFS);
1013                         btrfs_release_path(path);
1014                         goto fail;
1015                 }
1016         }
1017
1018         BTRFS_I(inode)->generation = trans->transid;
1019         header = btrfs_item_ptr(leaf, path->slots[0],
1020                                 struct btrfs_free_space_header);
1021         btrfs_set_free_space_entries(leaf, header, entries);
1022         btrfs_set_free_space_bitmaps(leaf, header, bitmaps);
1023         btrfs_set_free_space_generation(leaf, header, trans->transid);
1024         btrfs_mark_buffer_dirty(leaf);
1025         btrfs_release_path(path);
1026
1027         return 0;
1028
1029 fail:
1030         return -1;
1031 }
1032
1033 static noinline_for_stack int
1034 write_pinned_extent_entries(struct btrfs_root *root,
1035                             struct btrfs_block_group_cache *block_group,
1036                             struct btrfs_io_ctl *io_ctl,
1037                             int *entries)
1038 {
1039         u64 start, extent_start, extent_end, len;
1040         struct extent_io_tree *unpin = NULL;
1041         int ret;
1042
1043         if (!block_group)
1044                 return 0;
1045
1046         /*
1047          * We want to add any pinned extents to our free space cache
1048          * so we don't leak the space
1049          *
1050          * We shouldn't have switched the pinned extents yet so this is the
1051          * right one
1052          */
1053         unpin = root->fs_info->pinned_extents;
1054
1055         start = block_group->key.objectid;
1056
1057         while (start < block_group->key.objectid + block_group->key.offset) {
1058                 ret = find_first_extent_bit(unpin, start,
1059                                             &extent_start, &extent_end,
1060                                             EXTENT_DIRTY, NULL);
1061                 if (ret)
1062                         return 0;
1063
1064                 /* This pinned extent is out of our range */
1065                 if (extent_start >= block_group->key.objectid +
1066                     block_group->key.offset)
1067                         return 0;
1068
1069                 extent_start = max(extent_start, start);
1070                 extent_end = min(block_group->key.objectid +
1071                                  block_group->key.offset, extent_end + 1);
1072                 len = extent_end - extent_start;
1073
1074                 *entries += 1;
1075                 ret = io_ctl_add_entry(io_ctl, extent_start, len, NULL);
1076                 if (ret)
1077                         return -ENOSPC;
1078
1079                 start = extent_end;
1080         }
1081
1082         return 0;
1083 }
1084
1085 static noinline_for_stack int
1086 write_bitmap_entries(struct btrfs_io_ctl *io_ctl, struct list_head *bitmap_list)
1087 {
1088         struct btrfs_free_space *entry, *next;
1089         int ret;
1090
1091         /* Write out the bitmaps */
1092         list_for_each_entry_safe(entry, next, bitmap_list, list) {
1093                 ret = io_ctl_add_bitmap(io_ctl, entry->bitmap);
1094                 if (ret)
1095                         return -ENOSPC;
1096                 list_del_init(&entry->list);
1097         }
1098
1099         return 0;
1100 }
1101
1102 static int flush_dirty_cache(struct inode *inode)
1103 {
1104         int ret;
1105
1106         ret = btrfs_wait_ordered_range(inode, 0, (u64)-1);
1107         if (ret)
1108                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1109                                  EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0, NULL,
1110                                  GFP_NOFS);
1111
1112         return ret;
1113 }
1114
1115 static void noinline_for_stack
1116 cleanup_bitmap_list(struct list_head *bitmap_list)
1117 {
1118         struct btrfs_free_space *entry, *next;
1119
1120         list_for_each_entry_safe(entry, next, bitmap_list, list)
1121                 list_del_init(&entry->list);
1122 }
1123
1124 static void noinline_for_stack
1125 cleanup_write_cache_enospc(struct inode *inode,
1126                            struct btrfs_io_ctl *io_ctl,
1127                            struct extent_state **cached_state,
1128                            struct list_head *bitmap_list)
1129 {
1130         io_ctl_drop_pages(io_ctl);
1131         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1132                              i_size_read(inode) - 1, cached_state,
1133                              GFP_NOFS);
1134 }
1135
1136 int btrfs_wait_cache_io(struct btrfs_root *root,
1137                         struct btrfs_trans_handle *trans,
1138                         struct btrfs_block_group_cache *block_group,
1139                         struct btrfs_io_ctl *io_ctl,
1140                         struct btrfs_path *path, u64 offset)
1141 {
1142         int ret;
1143         struct inode *inode = io_ctl->inode;
1144
1145         if (!inode)
1146                 return 0;
1147
1148         if (block_group)
1149                 root = root->fs_info->tree_root;
1150
1151         /* Flush the dirty pages in the cache file. */
1152         ret = flush_dirty_cache(inode);
1153         if (ret)
1154                 goto out;
1155
1156         /* Update the cache item to tell everyone this cache file is valid. */
1157         ret = update_cache_item(trans, root, inode, path, offset,
1158                                 io_ctl->entries, io_ctl->bitmaps);
1159 out:
1160         io_ctl_free(io_ctl);
1161         if (ret) {
1162                 invalidate_inode_pages2(inode->i_mapping);
1163                 BTRFS_I(inode)->generation = 0;
1164                 if (block_group) {
1165 #ifdef DEBUG
1166                         btrfs_err(root->fs_info,
1167                                 "failed to write free space cache for block group %llu",
1168                                 block_group->key.objectid);
1169 #endif
1170                 }
1171         }
1172         btrfs_update_inode(trans, root, inode);
1173
1174         if (block_group) {
1175                 /* the dirty list is protected by the dirty_bgs_lock */
1176                 spin_lock(&trans->transaction->dirty_bgs_lock);
1177
1178                 /* the disk_cache_state is protected by the block group lock */
1179                 spin_lock(&block_group->lock);
1180
1181                 /*
1182                  * only mark this as written if we didn't get put back on
1183                  * the dirty list while waiting for IO.   Otherwise our
1184                  * cache state won't be right, and we won't get written again
1185                  */
1186                 if (!ret && list_empty(&block_group->dirty_list))
1187                         block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1188                 else if (ret)
1189                         block_group->disk_cache_state = BTRFS_DC_ERROR;
1190
1191                 spin_unlock(&block_group->lock);
1192                 spin_unlock(&trans->transaction->dirty_bgs_lock);
1193                 io_ctl->inode = NULL;
1194                 iput(inode);
1195         }
1196
1197         return ret;
1198
1199 }
1200
1201 /**
1202  * __btrfs_write_out_cache - write out cached info to an inode
1203  * @root - the root the inode belongs to
1204  * @ctl - the free space cache we are going to write out
1205  * @block_group - the block_group for this cache if it belongs to a block_group
1206  * @trans - the trans handle
1207  * @path - the path to use
1208  * @offset - the offset for the key we'll insert
1209  *
1210  * This function writes out a free space cache struct to disk for quick recovery
1211  * on mount.  This will return 0 if it was successful in writing the cache out,
1212  * or an errno if it was not.
1213  */
1214 static int __btrfs_write_out_cache(struct btrfs_root *root, struct inode *inode,
1215                                    struct btrfs_free_space_ctl *ctl,
1216                                    struct btrfs_block_group_cache *block_group,
1217                                    struct btrfs_io_ctl *io_ctl,
1218                                    struct btrfs_trans_handle *trans,
1219                                    struct btrfs_path *path, u64 offset)
1220 {
1221         struct extent_state *cached_state = NULL;
1222         LIST_HEAD(bitmap_list);
1223         int entries = 0;
1224         int bitmaps = 0;
1225         int ret;
1226         int must_iput = 0;
1227
1228         if (!i_size_read(inode))
1229                 return -EIO;
1230
1231         WARN_ON(io_ctl->pages);
1232         ret = io_ctl_init(io_ctl, inode, root, 1);
1233         if (ret)
1234                 return ret;
1235
1236         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA)) {
1237                 down_write(&block_group->data_rwsem);
1238                 spin_lock(&block_group->lock);
1239                 if (block_group->delalloc_bytes) {
1240                         block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1241                         spin_unlock(&block_group->lock);
1242                         up_write(&block_group->data_rwsem);
1243                         BTRFS_I(inode)->generation = 0;
1244                         ret = 0;
1245                         must_iput = 1;
1246                         goto out;
1247                 }
1248                 spin_unlock(&block_group->lock);
1249         }
1250
1251         /* Lock all pages first so we can lock the extent safely. */
1252         ret = io_ctl_prepare_pages(io_ctl, inode, 0);
1253         if (ret)
1254                 goto out;
1255
1256         lock_extent_bits(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
1257                          &cached_state);
1258
1259         io_ctl_set_generation(io_ctl, trans->transid);
1260
1261         mutex_lock(&ctl->cache_writeout_mutex);
1262         /* Write out the extent entries in the free space cache */
1263         spin_lock(&ctl->tree_lock);
1264         ret = write_cache_extent_entries(io_ctl, ctl,
1265                                          block_group, &entries, &bitmaps,
1266                                          &bitmap_list);
1267         if (ret)
1268                 goto out_nospc_locked;
1269
1270         /*
1271          * Some spaces that are freed in the current transaction are pinned,
1272          * they will be added into free space cache after the transaction is
1273          * committed, we shouldn't lose them.
1274          *
1275          * If this changes while we are working we'll get added back to
1276          * the dirty list and redo it.  No locking needed
1277          */
1278         ret = write_pinned_extent_entries(root, block_group, io_ctl, &entries);
1279         if (ret)
1280                 goto out_nospc_locked;
1281
1282         /*
1283          * At last, we write out all the bitmaps and keep cache_writeout_mutex
1284          * locked while doing it because a concurrent trim can be manipulating
1285          * or freeing the bitmap.
1286          */
1287         ret = write_bitmap_entries(io_ctl, &bitmap_list);
1288         spin_unlock(&ctl->tree_lock);
1289         mutex_unlock(&ctl->cache_writeout_mutex);
1290         if (ret)
1291                 goto out_nospc;
1292
1293         /* Zero out the rest of the pages just to make sure */
1294         io_ctl_zero_remaining_pages(io_ctl);
1295
1296         /* Everything is written out, now we dirty the pages in the file. */
1297         ret = btrfs_dirty_pages(root, inode, io_ctl->pages, io_ctl->num_pages,
1298                                 0, i_size_read(inode), &cached_state);
1299         if (ret)
1300                 goto out_nospc;
1301
1302         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1303                 up_write(&block_group->data_rwsem);
1304         /*
1305          * Release the pages and unlock the extent, we will flush
1306          * them out later
1307          */
1308         io_ctl_drop_pages(io_ctl);
1309
1310         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1311                              i_size_read(inode) - 1, &cached_state, GFP_NOFS);
1312
1313         /*
1314          * at this point the pages are under IO and we're happy,
1315          * The caller is responsible for waiting on them and updating the
1316          * the cache and the inode
1317          */
1318         io_ctl->entries = entries;
1319         io_ctl->bitmaps = bitmaps;
1320
1321         ret = btrfs_fdatawrite_range(inode, 0, (u64)-1);
1322         if (ret)
1323                 goto out;
1324
1325         return 0;
1326
1327 out:
1328         io_ctl->inode = NULL;
1329         io_ctl_free(io_ctl);
1330         if (ret) {
1331                 invalidate_inode_pages2(inode->i_mapping);
1332                 BTRFS_I(inode)->generation = 0;
1333         }
1334         btrfs_update_inode(trans, root, inode);
1335         if (must_iput)
1336                 iput(inode);
1337         return ret;
1338
1339 out_nospc_locked:
1340         cleanup_bitmap_list(&bitmap_list);
1341         spin_unlock(&ctl->tree_lock);
1342         mutex_unlock(&ctl->cache_writeout_mutex);
1343
1344 out_nospc:
1345         cleanup_write_cache_enospc(inode, io_ctl, &cached_state, &bitmap_list);
1346
1347         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1348                 up_write(&block_group->data_rwsem);
1349
1350         goto out;
1351 }
1352
1353 int btrfs_write_out_cache(struct btrfs_root *root,
1354                           struct btrfs_trans_handle *trans,
1355                           struct btrfs_block_group_cache *block_group,
1356                           struct btrfs_path *path)
1357 {
1358         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1359         struct inode *inode;
1360         int ret = 0;
1361
1362         root = root->fs_info->tree_root;
1363
1364         spin_lock(&block_group->lock);
1365         if (block_group->disk_cache_state < BTRFS_DC_SETUP) {
1366                 spin_unlock(&block_group->lock);
1367                 return 0;
1368         }
1369         spin_unlock(&block_group->lock);
1370
1371         inode = lookup_free_space_inode(root, block_group, path);
1372         if (IS_ERR(inode))
1373                 return 0;
1374
1375         ret = __btrfs_write_out_cache(root, inode, ctl, block_group,
1376                                       &block_group->io_ctl, trans,
1377                                       path, block_group->key.objectid);
1378         if (ret) {
1379 #ifdef DEBUG
1380                 btrfs_err(root->fs_info,
1381                         "failed to write free space cache for block group %llu",
1382                         block_group->key.objectid);
1383 #endif
1384                 spin_lock(&block_group->lock);
1385                 block_group->disk_cache_state = BTRFS_DC_ERROR;
1386                 spin_unlock(&block_group->lock);
1387
1388                 block_group->io_ctl.inode = NULL;
1389                 iput(inode);
1390         }
1391
1392         /*
1393          * if ret == 0 the caller is expected to call btrfs_wait_cache_io
1394          * to wait for IO and put the inode
1395          */
1396
1397         return ret;
1398 }
1399
1400 static inline unsigned long offset_to_bit(u64 bitmap_start, u32 unit,
1401                                           u64 offset)
1402 {
1403         ASSERT(offset >= bitmap_start);
1404         offset -= bitmap_start;
1405         return (unsigned long)(div_u64(offset, unit));
1406 }
1407
1408 static inline unsigned long bytes_to_bits(u64 bytes, u32 unit)
1409 {
1410         return (unsigned long)(div_u64(bytes, unit));
1411 }
1412
1413 static inline u64 offset_to_bitmap(struct btrfs_free_space_ctl *ctl,
1414                                    u64 offset)
1415 {
1416         u64 bitmap_start;
1417         u64 bytes_per_bitmap;
1418
1419         bytes_per_bitmap = BITS_PER_BITMAP * ctl->unit;
1420         bitmap_start = offset - ctl->start;
1421         bitmap_start = div64_u64(bitmap_start, bytes_per_bitmap);
1422         bitmap_start *= bytes_per_bitmap;
1423         bitmap_start += ctl->start;
1424
1425         return bitmap_start;
1426 }
1427
1428 static int tree_insert_offset(struct rb_root *root, u64 offset,
1429                               struct rb_node *node, int bitmap)
1430 {
1431         struct rb_node **p = &root->rb_node;
1432         struct rb_node *parent = NULL;
1433         struct btrfs_free_space *info;
1434
1435         while (*p) {
1436                 parent = *p;
1437                 info = rb_entry(parent, struct btrfs_free_space, offset_index);
1438
1439                 if (offset < info->offset) {
1440                         p = &(*p)->rb_left;
1441                 } else if (offset > info->offset) {
1442                         p = &(*p)->rb_right;
1443                 } else {
1444                         /*
1445                          * we could have a bitmap entry and an extent entry
1446                          * share the same offset.  If this is the case, we want
1447                          * the extent entry to always be found first if we do a
1448                          * linear search through the tree, since we want to have
1449                          * the quickest allocation time, and allocating from an
1450                          * extent is faster than allocating from a bitmap.  So
1451                          * if we're inserting a bitmap and we find an entry at
1452                          * this offset, we want to go right, or after this entry
1453                          * logically.  If we are inserting an extent and we've
1454                          * found a bitmap, we want to go left, or before
1455                          * logically.
1456                          */
1457                         if (bitmap) {
1458                                 if (info->bitmap) {
1459                                         WARN_ON_ONCE(1);
1460                                         return -EEXIST;
1461                                 }
1462                                 p = &(*p)->rb_right;
1463                         } else {
1464                                 if (!info->bitmap) {
1465                                         WARN_ON_ONCE(1);
1466                                         return -EEXIST;
1467                                 }
1468                                 p = &(*p)->rb_left;
1469                         }
1470                 }
1471         }
1472
1473         rb_link_node(node, parent, p);
1474         rb_insert_color(node, root);
1475
1476         return 0;
1477 }
1478
1479 /*
1480  * searches the tree for the given offset.
1481  *
1482  * fuzzy - If this is set, then we are trying to make an allocation, and we just
1483  * want a section that has at least bytes size and comes at or after the given
1484  * offset.
1485  */
1486 static struct btrfs_free_space *
1487 tree_search_offset(struct btrfs_free_space_ctl *ctl,
1488                    u64 offset, int bitmap_only, int fuzzy)
1489 {
1490         struct rb_node *n = ctl->free_space_offset.rb_node;
1491         struct btrfs_free_space *entry, *prev = NULL;
1492
1493         /* find entry that is closest to the 'offset' */
1494         while (1) {
1495                 if (!n) {
1496                         entry = NULL;
1497                         break;
1498                 }
1499
1500                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1501                 prev = entry;
1502
1503                 if (offset < entry->offset)
1504                         n = n->rb_left;
1505                 else if (offset > entry->offset)
1506                         n = n->rb_right;
1507                 else
1508                         break;
1509         }
1510
1511         if (bitmap_only) {
1512                 if (!entry)
1513                         return NULL;
1514                 if (entry->bitmap)
1515                         return entry;
1516
1517                 /*
1518                  * bitmap entry and extent entry may share same offset,
1519                  * in that case, bitmap entry comes after extent entry.
1520                  */
1521                 n = rb_next(n);
1522                 if (!n)
1523                         return NULL;
1524                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1525                 if (entry->offset != offset)
1526                         return NULL;
1527
1528                 WARN_ON(!entry->bitmap);
1529                 return entry;
1530         } else if (entry) {
1531                 if (entry->bitmap) {
1532                         /*
1533                          * if previous extent entry covers the offset,
1534                          * we should return it instead of the bitmap entry
1535                          */
1536                         n = rb_prev(&entry->offset_index);
1537                         if (n) {
1538                                 prev = rb_entry(n, struct btrfs_free_space,
1539                                                 offset_index);
1540                                 if (!prev->bitmap &&
1541                                     prev->offset + prev->bytes > offset)
1542                                         entry = prev;
1543                         }
1544                 }
1545                 return entry;
1546         }
1547
1548         if (!prev)
1549                 return NULL;
1550
1551         /* find last entry before the 'offset' */
1552         entry = prev;
1553         if (entry->offset > offset) {
1554                 n = rb_prev(&entry->offset_index);
1555                 if (n) {
1556                         entry = rb_entry(n, struct btrfs_free_space,
1557                                         offset_index);
1558                         ASSERT(entry->offset <= offset);
1559                 } else {
1560                         if (fuzzy)
1561                                 return entry;
1562                         else
1563                                 return NULL;
1564                 }
1565         }
1566
1567         if (entry->bitmap) {
1568                 n = rb_prev(&entry->offset_index);
1569                 if (n) {
1570                         prev = rb_entry(n, struct btrfs_free_space,
1571                                         offset_index);
1572                         if (!prev->bitmap &&
1573                             prev->offset + prev->bytes > offset)
1574                                 return prev;
1575                 }
1576                 if (entry->offset + BITS_PER_BITMAP * ctl->unit > offset)
1577                         return entry;
1578         } else if (entry->offset + entry->bytes > offset)
1579                 return entry;
1580
1581         if (!fuzzy)
1582                 return NULL;
1583
1584         while (1) {
1585                 if (entry->bitmap) {
1586                         if (entry->offset + BITS_PER_BITMAP *
1587                             ctl->unit > offset)
1588                                 break;
1589                 } else {
1590                         if (entry->offset + entry->bytes > offset)
1591                                 break;
1592                 }
1593
1594                 n = rb_next(&entry->offset_index);
1595                 if (!n)
1596                         return NULL;
1597                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1598         }
1599         return entry;
1600 }
1601
1602 static inline void
1603 __unlink_free_space(struct btrfs_free_space_ctl *ctl,
1604                     struct btrfs_free_space *info)
1605 {
1606         rb_erase(&info->offset_index, &ctl->free_space_offset);
1607         ctl->free_extents--;
1608 }
1609
1610 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
1611                               struct btrfs_free_space *info)
1612 {
1613         __unlink_free_space(ctl, info);
1614         ctl->free_space -= info->bytes;
1615 }
1616
1617 static int link_free_space(struct btrfs_free_space_ctl *ctl,
1618                            struct btrfs_free_space *info)
1619 {
1620         int ret = 0;
1621
1622         ASSERT(info->bytes || info->bitmap);
1623         ret = tree_insert_offset(&ctl->free_space_offset, info->offset,
1624                                  &info->offset_index, (info->bitmap != NULL));
1625         if (ret)
1626                 return ret;
1627
1628         ctl->free_space += info->bytes;
1629         ctl->free_extents++;
1630         return ret;
1631 }
1632
1633 static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
1634 {
1635         struct btrfs_block_group_cache *block_group = ctl->private;
1636         u64 max_bytes;
1637         u64 bitmap_bytes;
1638         u64 extent_bytes;
1639         u64 size = block_group->key.offset;
1640         u64 bytes_per_bg = BITS_PER_BITMAP * ctl->unit;
1641         u64 max_bitmaps = div64_u64(size + bytes_per_bg - 1, bytes_per_bg);
1642
1643         max_bitmaps = max_t(u64, max_bitmaps, 1);
1644
1645         ASSERT(ctl->total_bitmaps <= max_bitmaps);
1646
1647         /*
1648          * The goal is to keep the total amount of memory used per 1gb of space
1649          * at or below 32k, so we need to adjust how much memory we allow to be
1650          * used by extent based free space tracking
1651          */
1652         if (size < SZ_1G)
1653                 max_bytes = MAX_CACHE_BYTES_PER_GIG;
1654         else
1655                 max_bytes = MAX_CACHE_BYTES_PER_GIG * div_u64(size, SZ_1G);
1656
1657         /*
1658          * we want to account for 1 more bitmap than what we have so we can make
1659          * sure we don't go over our overall goal of MAX_CACHE_BYTES_PER_GIG as
1660          * we add more bitmaps.
1661          */
1662         bitmap_bytes = (ctl->total_bitmaps + 1) * ctl->unit;
1663
1664         if (bitmap_bytes >= max_bytes) {
1665                 ctl->extents_thresh = 0;
1666                 return;
1667         }
1668
1669         /*
1670          * we want the extent entry threshold to always be at most 1/2 the max
1671          * bytes we can have, or whatever is less than that.
1672          */
1673         extent_bytes = max_bytes - bitmap_bytes;
1674         extent_bytes = min_t(u64, extent_bytes, max_bytes >> 1);
1675
1676         ctl->extents_thresh =
1677                 div_u64(extent_bytes, sizeof(struct btrfs_free_space));
1678 }
1679
1680 static inline void __bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1681                                        struct btrfs_free_space *info,
1682                                        u64 offset, u64 bytes)
1683 {
1684         unsigned long start, count;
1685
1686         start = offset_to_bit(info->offset, ctl->unit, offset);
1687         count = bytes_to_bits(bytes, ctl->unit);
1688         ASSERT(start + count <= BITS_PER_BITMAP);
1689
1690         bitmap_clear(info->bitmap, start, count);
1691
1692         info->bytes -= bytes;
1693 }
1694
1695 static void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1696                               struct btrfs_free_space *info, u64 offset,
1697                               u64 bytes)
1698 {
1699         __bitmap_clear_bits(ctl, info, offset, bytes);
1700         ctl->free_space -= bytes;
1701 }
1702
1703 static void bitmap_set_bits(struct btrfs_free_space_ctl *ctl,
1704                             struct btrfs_free_space *info, u64 offset,
1705                             u64 bytes)
1706 {
1707         unsigned long start, count;
1708
1709         start = offset_to_bit(info->offset, ctl->unit, offset);
1710         count = bytes_to_bits(bytes, ctl->unit);
1711         ASSERT(start + count <= BITS_PER_BITMAP);
1712
1713         bitmap_set(info->bitmap, start, count);
1714
1715         info->bytes += bytes;
1716         ctl->free_space += bytes;
1717 }
1718
1719 /*
1720  * If we can not find suitable extent, we will use bytes to record
1721  * the size of the max extent.
1722  */
1723 static int search_bitmap(struct btrfs_free_space_ctl *ctl,
1724                          struct btrfs_free_space *bitmap_info, u64 *offset,
1725                          u64 *bytes, bool for_alloc)
1726 {
1727         unsigned long found_bits = 0;
1728         unsigned long max_bits = 0;
1729         unsigned long bits, i;
1730         unsigned long next_zero;
1731         unsigned long extent_bits;
1732
1733         /*
1734          * Skip searching the bitmap if we don't have a contiguous section that
1735          * is large enough for this allocation.
1736          */
1737         if (for_alloc &&
1738             bitmap_info->max_extent_size &&
1739             bitmap_info->max_extent_size < *bytes) {
1740                 *bytes = bitmap_info->max_extent_size;
1741                 return -1;
1742         }
1743
1744         i = offset_to_bit(bitmap_info->offset, ctl->unit,
1745                           max_t(u64, *offset, bitmap_info->offset));
1746         bits = bytes_to_bits(*bytes, ctl->unit);
1747
1748         for_each_set_bit_from(i, bitmap_info->bitmap, BITS_PER_BITMAP) {
1749                 if (for_alloc && bits == 1) {
1750                         found_bits = 1;
1751                         break;
1752                 }
1753                 next_zero = find_next_zero_bit(bitmap_info->bitmap,
1754                                                BITS_PER_BITMAP, i);
1755                 extent_bits = next_zero - i;
1756                 if (extent_bits >= bits) {
1757                         found_bits = extent_bits;
1758                         break;
1759                 } else if (extent_bits > max_bits) {
1760                         max_bits = extent_bits;
1761                 }
1762                 i = next_zero;
1763         }
1764
1765         if (found_bits) {
1766                 *offset = (u64)(i * ctl->unit) + bitmap_info->offset;
1767                 *bytes = (u64)(found_bits) * ctl->unit;
1768                 return 0;
1769         }
1770
1771         *bytes = (u64)(max_bits) * ctl->unit;
1772         bitmap_info->max_extent_size = *bytes;
1773         return -1;
1774 }
1775
1776 /* Cache the size of the max extent in bytes */
1777 static struct btrfs_free_space *
1778 find_free_space(struct btrfs_free_space_ctl *ctl, u64 *offset, u64 *bytes,
1779                 unsigned long align, u64 *max_extent_size)
1780 {
1781         struct btrfs_free_space *entry;
1782         struct rb_node *node;
1783         u64 tmp;
1784         u64 align_off;
1785         int ret;
1786
1787         if (!ctl->free_space_offset.rb_node)
1788                 goto out;
1789
1790         entry = tree_search_offset(ctl, offset_to_bitmap(ctl, *offset), 0, 1);
1791         if (!entry)
1792                 goto out;
1793
1794         for (node = &entry->offset_index; node; node = rb_next(node)) {
1795                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1796                 if (entry->bytes < *bytes) {
1797                         if (entry->bytes > *max_extent_size)
1798                                 *max_extent_size = entry->bytes;
1799                         continue;
1800                 }
1801
1802                 /* make sure the space returned is big enough
1803                  * to match our requested alignment
1804                  */
1805                 if (*bytes >= align) {
1806                         tmp = entry->offset - ctl->start + align - 1;
1807                         tmp = div64_u64(tmp, align);
1808                         tmp = tmp * align + ctl->start;
1809                         align_off = tmp - entry->offset;
1810                 } else {
1811                         align_off = 0;
1812                         tmp = entry->offset;
1813                 }
1814
1815                 if (entry->bytes < *bytes + align_off) {
1816                         if (entry->bytes > *max_extent_size)
1817                                 *max_extent_size = entry->bytes;
1818                         continue;
1819                 }
1820
1821                 if (entry->bitmap) {
1822                         u64 size = *bytes;
1823
1824                         ret = search_bitmap(ctl, entry, &tmp, &size, true);
1825                         if (!ret) {
1826                                 *offset = tmp;
1827                                 *bytes = size;
1828                                 return entry;
1829                         } else if (size > *max_extent_size) {
1830                                 *max_extent_size = size;
1831                         }
1832                         continue;
1833                 }
1834
1835                 *offset = tmp;
1836                 *bytes = entry->bytes - align_off;
1837                 return entry;
1838         }
1839 out:
1840         return NULL;
1841 }
1842
1843 static void add_new_bitmap(struct btrfs_free_space_ctl *ctl,
1844                            struct btrfs_free_space *info, u64 offset)
1845 {
1846         info->offset = offset_to_bitmap(ctl, offset);
1847         info->bytes = 0;
1848         INIT_LIST_HEAD(&info->list);
1849         link_free_space(ctl, info);
1850         ctl->total_bitmaps++;
1851
1852         ctl->op->recalc_thresholds(ctl);
1853 }
1854
1855 static void free_bitmap(struct btrfs_free_space_ctl *ctl,
1856                         struct btrfs_free_space *bitmap_info)
1857 {
1858         unlink_free_space(ctl, bitmap_info);
1859         kfree(bitmap_info->bitmap);
1860         kmem_cache_free(btrfs_free_space_cachep, bitmap_info);
1861         ctl->total_bitmaps--;
1862         ctl->op->recalc_thresholds(ctl);
1863 }
1864
1865 static noinline int remove_from_bitmap(struct btrfs_free_space_ctl *ctl,
1866                               struct btrfs_free_space *bitmap_info,
1867                               u64 *offset, u64 *bytes)
1868 {
1869         u64 end;
1870         u64 search_start, search_bytes;
1871         int ret;
1872
1873 again:
1874         end = bitmap_info->offset + (u64)(BITS_PER_BITMAP * ctl->unit) - 1;
1875
1876         /*
1877          * We need to search for bits in this bitmap.  We could only cover some
1878          * of the extent in this bitmap thanks to how we add space, so we need
1879          * to search for as much as it as we can and clear that amount, and then
1880          * go searching for the next bit.
1881          */
1882         search_start = *offset;
1883         search_bytes = ctl->unit;
1884         search_bytes = min(search_bytes, end - search_start + 1);
1885         ret = search_bitmap(ctl, bitmap_info, &search_start, &search_bytes,
1886                             false);
1887         if (ret < 0 || search_start != *offset)
1888                 return -EINVAL;
1889
1890         /* We may have found more bits than what we need */
1891         search_bytes = min(search_bytes, *bytes);
1892
1893         /* Cannot clear past the end of the bitmap */
1894         search_bytes = min(search_bytes, end - search_start + 1);
1895
1896         bitmap_clear_bits(ctl, bitmap_info, search_start, search_bytes);
1897         *offset += search_bytes;
1898         *bytes -= search_bytes;
1899
1900         if (*bytes) {
1901                 struct rb_node *next = rb_next(&bitmap_info->offset_index);
1902                 if (!bitmap_info->bytes)
1903                         free_bitmap(ctl, bitmap_info);
1904
1905                 /*
1906                  * no entry after this bitmap, but we still have bytes to
1907                  * remove, so something has gone wrong.
1908                  */
1909                 if (!next)
1910                         return -EINVAL;
1911
1912                 bitmap_info = rb_entry(next, struct btrfs_free_space,
1913                                        offset_index);
1914
1915                 /*
1916                  * if the next entry isn't a bitmap we need to return to let the
1917                  * extent stuff do its work.
1918                  */
1919                 if (!bitmap_info->bitmap)
1920                         return -EAGAIN;
1921
1922                 /*
1923                  * Ok the next item is a bitmap, but it may not actually hold
1924                  * the information for the rest of this free space stuff, so
1925                  * look for it, and if we don't find it return so we can try
1926                  * everything over again.
1927                  */
1928                 search_start = *offset;
1929                 search_bytes = ctl->unit;
1930                 ret = search_bitmap(ctl, bitmap_info, &search_start,
1931                                     &search_bytes, false);
1932                 if (ret < 0 || search_start != *offset)
1933                         return -EAGAIN;
1934
1935                 goto again;
1936         } else if (!bitmap_info->bytes)
1937                 free_bitmap(ctl, bitmap_info);
1938
1939         return 0;
1940 }
1941
1942 static u64 add_bytes_to_bitmap(struct btrfs_free_space_ctl *ctl,
1943                                struct btrfs_free_space *info, u64 offset,
1944                                u64 bytes)
1945 {
1946         u64 bytes_to_set = 0;
1947         u64 end;
1948
1949         end = info->offset + (u64)(BITS_PER_BITMAP * ctl->unit);
1950
1951         bytes_to_set = min(end - offset, bytes);
1952
1953         bitmap_set_bits(ctl, info, offset, bytes_to_set);
1954
1955         /*
1956          * We set some bytes, we have no idea what the max extent size is
1957          * anymore.
1958          */
1959         info->max_extent_size = 0;
1960
1961         return bytes_to_set;
1962
1963 }
1964
1965 static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
1966                       struct btrfs_free_space *info)
1967 {
1968         struct btrfs_block_group_cache *block_group = ctl->private;
1969         bool forced = false;
1970
1971 #ifdef CONFIG_BTRFS_DEBUG
1972         if (btrfs_should_fragment_free_space(block_group->fs_info->extent_root,
1973                                              block_group))
1974                 forced = true;
1975 #endif
1976
1977         /*
1978          * If we are below the extents threshold then we can add this as an
1979          * extent, and don't have to deal with the bitmap
1980          */
1981         if (!forced && ctl->free_extents < ctl->extents_thresh) {
1982                 /*
1983                  * If this block group has some small extents we don't want to
1984                  * use up all of our free slots in the cache with them, we want
1985                  * to reserve them to larger extents, however if we have plenty
1986                  * of cache left then go ahead an dadd them, no sense in adding
1987                  * the overhead of a bitmap if we don't have to.
1988                  */
1989                 if (info->bytes <= block_group->sectorsize * 4) {
1990                         if (ctl->free_extents * 2 <= ctl->extents_thresh)
1991                                 return false;
1992                 } else {
1993                         return false;
1994                 }
1995         }
1996
1997         /*
1998          * The original block groups from mkfs can be really small, like 8
1999          * megabytes, so don't bother with a bitmap for those entries.  However
2000          * some block groups can be smaller than what a bitmap would cover but
2001          * are still large enough that they could overflow the 32k memory limit,
2002          * so allow those block groups to still be allowed to have a bitmap
2003          * entry.
2004          */
2005         if (((BITS_PER_BITMAP * ctl->unit) >> 1) > block_group->key.offset)
2006                 return false;
2007
2008         return true;
2009 }
2010
2011 static const struct btrfs_free_space_op free_space_op = {
2012         .recalc_thresholds      = recalculate_thresholds,
2013         .use_bitmap             = use_bitmap,
2014 };
2015
2016 static int insert_into_bitmap(struct btrfs_free_space_ctl *ctl,
2017                               struct btrfs_free_space *info)
2018 {
2019         struct btrfs_free_space *bitmap_info;
2020         struct btrfs_block_group_cache *block_group = NULL;
2021         int added = 0;
2022         u64 bytes, offset, bytes_added;
2023         int ret;
2024
2025         bytes = info->bytes;
2026         offset = info->offset;
2027
2028         if (!ctl->op->use_bitmap(ctl, info))
2029                 return 0;
2030
2031         if (ctl->op == &free_space_op)
2032                 block_group = ctl->private;
2033 again:
2034         /*
2035          * Since we link bitmaps right into the cluster we need to see if we
2036          * have a cluster here, and if so and it has our bitmap we need to add
2037          * the free space to that bitmap.
2038          */
2039         if (block_group && !list_empty(&block_group->cluster_list)) {
2040                 struct btrfs_free_cluster *cluster;
2041                 struct rb_node *node;
2042                 struct btrfs_free_space *entry;
2043
2044                 cluster = list_entry(block_group->cluster_list.next,
2045                                      struct btrfs_free_cluster,
2046                                      block_group_list);
2047                 spin_lock(&cluster->lock);
2048                 node = rb_first(&cluster->root);
2049                 if (!node) {
2050                         spin_unlock(&cluster->lock);
2051                         goto no_cluster_bitmap;
2052                 }
2053
2054                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2055                 if (!entry->bitmap) {
2056                         spin_unlock(&cluster->lock);
2057                         goto no_cluster_bitmap;
2058                 }
2059
2060                 if (entry->offset == offset_to_bitmap(ctl, offset)) {
2061                         bytes_added = add_bytes_to_bitmap(ctl, entry,
2062                                                           offset, bytes);
2063                         bytes -= bytes_added;
2064                         offset += bytes_added;
2065                 }
2066                 spin_unlock(&cluster->lock);
2067                 if (!bytes) {
2068                         ret = 1;
2069                         goto out;
2070                 }
2071         }
2072
2073 no_cluster_bitmap:
2074         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2075                                          1, 0);
2076         if (!bitmap_info) {
2077                 ASSERT(added == 0);
2078                 goto new_bitmap;
2079         }
2080
2081         bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes);
2082         bytes -= bytes_added;
2083         offset += bytes_added;
2084         added = 0;
2085
2086         if (!bytes) {
2087                 ret = 1;
2088                 goto out;
2089         } else
2090                 goto again;
2091
2092 new_bitmap:
2093         if (info && info->bitmap) {
2094                 add_new_bitmap(ctl, info, offset);
2095                 added = 1;
2096                 info = NULL;
2097                 goto again;
2098         } else {
2099                 spin_unlock(&ctl->tree_lock);
2100
2101                 /* no pre-allocated info, allocate a new one */
2102                 if (!info) {
2103                         info = kmem_cache_zalloc(btrfs_free_space_cachep,
2104                                                  GFP_NOFS);
2105                         if (!info) {
2106                                 spin_lock(&ctl->tree_lock);
2107                                 ret = -ENOMEM;
2108                                 goto out;
2109                         }
2110                 }
2111
2112                 /* allocate the bitmap */
2113                 info->bitmap = kzalloc(PAGE_SIZE, GFP_NOFS);
2114                 spin_lock(&ctl->tree_lock);
2115                 if (!info->bitmap) {
2116                         ret = -ENOMEM;
2117                         goto out;
2118                 }
2119                 goto again;
2120         }
2121
2122 out:
2123         if (info) {
2124                 if (info->bitmap)
2125                         kfree(info->bitmap);
2126                 kmem_cache_free(btrfs_free_space_cachep, info);
2127         }
2128
2129         return ret;
2130 }
2131
2132 static bool try_merge_free_space(struct btrfs_free_space_ctl *ctl,
2133                           struct btrfs_free_space *info, bool update_stat)
2134 {
2135         struct btrfs_free_space *left_info;
2136         struct btrfs_free_space *right_info;
2137         bool merged = false;
2138         u64 offset = info->offset;
2139         u64 bytes = info->bytes;
2140
2141         /*
2142          * first we want to see if there is free space adjacent to the range we
2143          * are adding, if there is remove that struct and add a new one to
2144          * cover the entire range
2145          */
2146         right_info = tree_search_offset(ctl, offset + bytes, 0, 0);
2147         if (right_info && rb_prev(&right_info->offset_index))
2148                 left_info = rb_entry(rb_prev(&right_info->offset_index),
2149                                      struct btrfs_free_space, offset_index);
2150         else
2151                 left_info = tree_search_offset(ctl, offset - 1, 0, 0);
2152
2153         if (right_info && !right_info->bitmap) {
2154                 if (update_stat)
2155                         unlink_free_space(ctl, right_info);
2156                 else
2157                         __unlink_free_space(ctl, right_info);
2158                 info->bytes += right_info->bytes;
2159                 kmem_cache_free(btrfs_free_space_cachep, right_info);
2160                 merged = true;
2161         }
2162
2163         if (left_info && !left_info->bitmap &&
2164             left_info->offset + left_info->bytes == offset) {
2165                 if (update_stat)
2166                         unlink_free_space(ctl, left_info);
2167                 else
2168                         __unlink_free_space(ctl, left_info);
2169                 info->offset = left_info->offset;
2170                 info->bytes += left_info->bytes;
2171                 kmem_cache_free(btrfs_free_space_cachep, left_info);
2172                 merged = true;
2173         }
2174
2175         return merged;
2176 }
2177
2178 static bool steal_from_bitmap_to_end(struct btrfs_free_space_ctl *ctl,
2179                                      struct btrfs_free_space *info,
2180                                      bool update_stat)
2181 {
2182         struct btrfs_free_space *bitmap;
2183         unsigned long i;
2184         unsigned long j;
2185         const u64 end = info->offset + info->bytes;
2186         const u64 bitmap_offset = offset_to_bitmap(ctl, end);
2187         u64 bytes;
2188
2189         bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2190         if (!bitmap)
2191                 return false;
2192
2193         i = offset_to_bit(bitmap->offset, ctl->unit, end);
2194         j = find_next_zero_bit(bitmap->bitmap, BITS_PER_BITMAP, i);
2195         if (j == i)
2196                 return false;
2197         bytes = (j - i) * ctl->unit;
2198         info->bytes += bytes;
2199
2200         if (update_stat)
2201                 bitmap_clear_bits(ctl, bitmap, end, bytes);
2202         else
2203                 __bitmap_clear_bits(ctl, bitmap, end, bytes);
2204
2205         if (!bitmap->bytes)
2206                 free_bitmap(ctl, bitmap);
2207
2208         return true;
2209 }
2210
2211 static bool steal_from_bitmap_to_front(struct btrfs_free_space_ctl *ctl,
2212                                        struct btrfs_free_space *info,
2213                                        bool update_stat)
2214 {
2215         struct btrfs_free_space *bitmap;
2216         u64 bitmap_offset;
2217         unsigned long i;
2218         unsigned long j;
2219         unsigned long prev_j;
2220         u64 bytes;
2221
2222         bitmap_offset = offset_to_bitmap(ctl, info->offset);
2223         /* If we're on a boundary, try the previous logical bitmap. */
2224         if (bitmap_offset == info->offset) {
2225                 if (info->offset == 0)
2226                         return false;
2227                 bitmap_offset = offset_to_bitmap(ctl, info->offset - 1);
2228         }
2229
2230         bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2231         if (!bitmap)
2232                 return false;
2233
2234         i = offset_to_bit(bitmap->offset, ctl->unit, info->offset) - 1;
2235         j = 0;
2236         prev_j = (unsigned long)-1;
2237         for_each_clear_bit_from(j, bitmap->bitmap, BITS_PER_BITMAP) {
2238                 if (j > i)
2239                         break;
2240                 prev_j = j;
2241         }
2242         if (prev_j == i)
2243                 return false;
2244
2245         if (prev_j == (unsigned long)-1)
2246                 bytes = (i + 1) * ctl->unit;
2247         else
2248                 bytes = (i - prev_j) * ctl->unit;
2249
2250         info->offset -= bytes;
2251         info->bytes += bytes;
2252
2253         if (update_stat)
2254                 bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2255         else
2256                 __bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2257
2258         if (!bitmap->bytes)
2259                 free_bitmap(ctl, bitmap);
2260
2261         return true;
2262 }
2263
2264 /*
2265  * We prefer always to allocate from extent entries, both for clustered and
2266  * non-clustered allocation requests. So when attempting to add a new extent
2267  * entry, try to see if there's adjacent free space in bitmap entries, and if
2268  * there is, migrate that space from the bitmaps to the extent.
2269  * Like this we get better chances of satisfying space allocation requests
2270  * because we attempt to satisfy them based on a single cache entry, and never
2271  * on 2 or more entries - even if the entries represent a contiguous free space
2272  * region (e.g. 1 extent entry + 1 bitmap entry starting where the extent entry
2273  * ends).
2274  */
2275 static void steal_from_bitmap(struct btrfs_free_space_ctl *ctl,
2276                               struct btrfs_free_space *info,
2277                               bool update_stat)
2278 {
2279         /*
2280          * Only work with disconnected entries, as we can change their offset,
2281          * and must be extent entries.
2282          */
2283         ASSERT(!info->bitmap);
2284         ASSERT(RB_EMPTY_NODE(&info->offset_index));
2285
2286         if (ctl->total_bitmaps > 0) {
2287                 bool stole_end;
2288                 bool stole_front = false;
2289
2290                 stole_end = steal_from_bitmap_to_end(ctl, info, update_stat);
2291                 if (ctl->total_bitmaps > 0)
2292                         stole_front = steal_from_bitmap_to_front(ctl, info,
2293                                                                  update_stat);
2294
2295                 if (stole_end || stole_front)
2296                         try_merge_free_space(ctl, info, update_stat);
2297         }
2298 }
2299
2300 int __btrfs_add_free_space(struct btrfs_fs_info *fs_info,
2301                            struct btrfs_free_space_ctl *ctl,
2302                            u64 offset, u64 bytes)
2303 {
2304         struct btrfs_free_space *info;
2305         int ret = 0;
2306
2307         info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
2308         if (!info)
2309                 return -ENOMEM;
2310
2311         info->offset = offset;
2312         info->bytes = bytes;
2313         RB_CLEAR_NODE(&info->offset_index);
2314
2315         spin_lock(&ctl->tree_lock);
2316
2317         if (try_merge_free_space(ctl, info, true))
2318                 goto link;
2319
2320         /*
2321          * There was no extent directly to the left or right of this new
2322          * extent then we know we're going to have to allocate a new extent, so
2323          * before we do that see if we need to drop this into a bitmap
2324          */
2325         ret = insert_into_bitmap(ctl, info);
2326         if (ret < 0) {
2327                 goto out;
2328         } else if (ret) {
2329                 ret = 0;
2330                 goto out;
2331         }
2332 link:
2333         /*
2334          * Only steal free space from adjacent bitmaps if we're sure we're not
2335          * going to add the new free space to existing bitmap entries - because
2336          * that would mean unnecessary work that would be reverted. Therefore
2337          * attempt to steal space from bitmaps if we're adding an extent entry.
2338          */
2339         steal_from_bitmap(ctl, info, true);
2340
2341         ret = link_free_space(ctl, info);
2342         if (ret)
2343                 kmem_cache_free(btrfs_free_space_cachep, info);
2344 out:
2345         spin_unlock(&ctl->tree_lock);
2346
2347         if (ret) {
2348                 btrfs_crit(fs_info, "unable to add free space :%d", ret);
2349                 ASSERT(ret != -EEXIST);
2350         }
2351
2352         return ret;
2353 }
2354
2355 int btrfs_remove_free_space(struct btrfs_block_group_cache *block_group,
2356                             u64 offset, u64 bytes)
2357 {
2358         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2359         struct btrfs_free_space *info;
2360         int ret;
2361         bool re_search = false;
2362
2363         spin_lock(&ctl->tree_lock);
2364
2365 again:
2366         ret = 0;
2367         if (!bytes)
2368                 goto out_lock;
2369
2370         info = tree_search_offset(ctl, offset, 0, 0);
2371         if (!info) {
2372                 /*
2373                  * oops didn't find an extent that matched the space we wanted
2374                  * to remove, look for a bitmap instead
2375                  */
2376                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2377                                           1, 0);
2378                 if (!info) {
2379                         /*
2380                          * If we found a partial bit of our free space in a
2381                          * bitmap but then couldn't find the other part this may
2382                          * be a problem, so WARN about it.
2383                          */
2384                         WARN_ON(re_search);
2385                         goto out_lock;
2386                 }
2387         }
2388
2389         re_search = false;
2390         if (!info->bitmap) {
2391                 unlink_free_space(ctl, info);
2392                 if (offset == info->offset) {
2393                         u64 to_free = min(bytes, info->bytes);
2394
2395                         info->bytes -= to_free;
2396                         info->offset += to_free;
2397                         if (info->bytes) {
2398                                 ret = link_free_space(ctl, info);
2399                                 WARN_ON(ret);
2400                         } else {
2401                                 kmem_cache_free(btrfs_free_space_cachep, info);
2402                         }
2403
2404                         offset += to_free;
2405                         bytes -= to_free;
2406                         goto again;
2407                 } else {
2408                         u64 old_end = info->bytes + info->offset;
2409
2410                         info->bytes = offset - info->offset;
2411                         ret = link_free_space(ctl, info);
2412                         WARN_ON(ret);
2413                         if (ret)
2414                                 goto out_lock;
2415
2416                         /* Not enough bytes in this entry to satisfy us */
2417                         if (old_end < offset + bytes) {
2418                                 bytes -= old_end - offset;
2419                                 offset = old_end;
2420                                 goto again;
2421                         } else if (old_end == offset + bytes) {
2422                                 /* all done */
2423                                 goto out_lock;
2424                         }
2425                         spin_unlock(&ctl->tree_lock);
2426
2427                         ret = btrfs_add_free_space(block_group, offset + bytes,
2428                                                    old_end - (offset + bytes));
2429                         WARN_ON(ret);
2430                         goto out;
2431                 }
2432         }
2433
2434         ret = remove_from_bitmap(ctl, info, &offset, &bytes);
2435         if (ret == -EAGAIN) {
2436                 re_search = true;
2437                 goto again;
2438         }
2439 out_lock:
2440         spin_unlock(&ctl->tree_lock);
2441 out:
2442         return ret;
2443 }
2444
2445 void btrfs_dump_free_space(struct btrfs_block_group_cache *block_group,
2446                            u64 bytes)
2447 {
2448         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2449         struct btrfs_free_space *info;
2450         struct rb_node *n;
2451         int count = 0;
2452
2453         for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
2454                 info = rb_entry(n, struct btrfs_free_space, offset_index);
2455                 if (info->bytes >= bytes && !block_group->ro)
2456                         count++;
2457                 btrfs_crit(block_group->fs_info,
2458                            "entry offset %llu, bytes %llu, bitmap %s",
2459                            info->offset, info->bytes,
2460                        (info->bitmap) ? "yes" : "no");
2461         }
2462         btrfs_info(block_group->fs_info, "block group has cluster?: %s",
2463                list_empty(&block_group->cluster_list) ? "no" : "yes");
2464         btrfs_info(block_group->fs_info,
2465                    "%d blocks of free space at or bigger than bytes is", count);
2466 }
2467
2468 void btrfs_init_free_space_ctl(struct btrfs_block_group_cache *block_group)
2469 {
2470         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2471
2472         spin_lock_init(&ctl->tree_lock);
2473         ctl->unit = block_group->sectorsize;
2474         ctl->start = block_group->key.objectid;
2475         ctl->private = block_group;
2476         ctl->op = &free_space_op;
2477         INIT_LIST_HEAD(&ctl->trimming_ranges);
2478         mutex_init(&ctl->cache_writeout_mutex);
2479
2480         /*
2481          * we only want to have 32k of ram per block group for keeping
2482          * track of free space, and if we pass 1/2 of that we want to
2483          * start converting things over to using bitmaps
2484          */
2485         ctl->extents_thresh = (SZ_32K / 2) / sizeof(struct btrfs_free_space);
2486 }
2487
2488 /*
2489  * for a given cluster, put all of its extents back into the free
2490  * space cache.  If the block group passed doesn't match the block group
2491  * pointed to by the cluster, someone else raced in and freed the
2492  * cluster already.  In that case, we just return without changing anything
2493  */
2494 static int
2495 __btrfs_return_cluster_to_free_space(
2496                              struct btrfs_block_group_cache *block_group,
2497                              struct btrfs_free_cluster *cluster)
2498 {
2499         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2500         struct btrfs_free_space *entry;
2501         struct rb_node *node;
2502
2503         spin_lock(&cluster->lock);
2504         if (cluster->block_group != block_group)
2505                 goto out;
2506
2507         cluster->block_group = NULL;
2508         cluster->window_start = 0;
2509         list_del_init(&cluster->block_group_list);
2510
2511         node = rb_first(&cluster->root);
2512         while (node) {
2513                 bool bitmap;
2514
2515                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2516                 node = rb_next(&entry->offset_index);
2517                 rb_erase(&entry->offset_index, &cluster->root);
2518                 RB_CLEAR_NODE(&entry->offset_index);
2519
2520                 bitmap = (entry->bitmap != NULL);
2521                 if (!bitmap) {
2522                         try_merge_free_space(ctl, entry, false);
2523                         steal_from_bitmap(ctl, entry, false);
2524                 }
2525                 tree_insert_offset(&ctl->free_space_offset,
2526                                    entry->offset, &entry->offset_index, bitmap);
2527         }
2528         cluster->root = RB_ROOT;
2529
2530 out:
2531         spin_unlock(&cluster->lock);
2532         btrfs_put_block_group(block_group);
2533         return 0;
2534 }
2535
2536 static void __btrfs_remove_free_space_cache_locked(
2537                                 struct btrfs_free_space_ctl *ctl)
2538 {
2539         struct btrfs_free_space *info;
2540         struct rb_node *node;
2541
2542         while ((node = rb_last(&ctl->free_space_offset)) != NULL) {
2543                 info = rb_entry(node, struct btrfs_free_space, offset_index);
2544                 if (!info->bitmap) {
2545                         unlink_free_space(ctl, info);
2546                         kmem_cache_free(btrfs_free_space_cachep, info);
2547                 } else {
2548                         free_bitmap(ctl, info);
2549                 }
2550
2551                 cond_resched_lock(&ctl->tree_lock);
2552         }
2553 }
2554
2555 void __btrfs_remove_free_space_cache(struct btrfs_free_space_ctl *ctl)
2556 {
2557         spin_lock(&ctl->tree_lock);
2558         __btrfs_remove_free_space_cache_locked(ctl);
2559         spin_unlock(&ctl->tree_lock);
2560 }
2561
2562 void btrfs_remove_free_space_cache(struct btrfs_block_group_cache *block_group)
2563 {
2564         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2565         struct btrfs_free_cluster *cluster;
2566         struct list_head *head;
2567
2568         spin_lock(&ctl->tree_lock);
2569         while ((head = block_group->cluster_list.next) !=
2570                &block_group->cluster_list) {
2571                 cluster = list_entry(head, struct btrfs_free_cluster,
2572                                      block_group_list);
2573
2574                 WARN_ON(cluster->block_group != block_group);
2575                 __btrfs_return_cluster_to_free_space(block_group, cluster);
2576
2577                 cond_resched_lock(&ctl->tree_lock);
2578         }
2579         __btrfs_remove_free_space_cache_locked(ctl);
2580         spin_unlock(&ctl->tree_lock);
2581
2582 }
2583
2584 u64 btrfs_find_space_for_alloc(struct btrfs_block_group_cache *block_group,
2585                                u64 offset, u64 bytes, u64 empty_size,
2586                                u64 *max_extent_size)
2587 {
2588         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2589         struct btrfs_free_space *entry = NULL;
2590         u64 bytes_search = bytes + empty_size;
2591         u64 ret = 0;
2592         u64 align_gap = 0;
2593         u64 align_gap_len = 0;
2594
2595         spin_lock(&ctl->tree_lock);
2596         entry = find_free_space(ctl, &offset, &bytes_search,
2597                                 block_group->full_stripe_len, max_extent_size);
2598         if (!entry)
2599                 goto out;
2600
2601         ret = offset;
2602         if (entry->bitmap) {
2603                 bitmap_clear_bits(ctl, entry, offset, bytes);
2604                 if (!entry->bytes)
2605                         free_bitmap(ctl, entry);
2606         } else {
2607                 unlink_free_space(ctl, entry);
2608                 align_gap_len = offset - entry->offset;
2609                 align_gap = entry->offset;
2610
2611                 entry->offset = offset + bytes;
2612                 WARN_ON(entry->bytes < bytes + align_gap_len);
2613
2614                 entry->bytes -= bytes + align_gap_len;
2615                 if (!entry->bytes)
2616                         kmem_cache_free(btrfs_free_space_cachep, entry);
2617                 else
2618                         link_free_space(ctl, entry);
2619         }
2620 out:
2621         spin_unlock(&ctl->tree_lock);
2622
2623         if (align_gap_len)
2624                 __btrfs_add_free_space(block_group->fs_info, ctl,
2625                                        align_gap, align_gap_len);
2626         return ret;
2627 }
2628
2629 /*
2630  * given a cluster, put all of its extents back into the free space
2631  * cache.  If a block group is passed, this function will only free
2632  * a cluster that belongs to the passed block group.
2633  *
2634  * Otherwise, it'll get a reference on the block group pointed to by the
2635  * cluster and remove the cluster from it.
2636  */
2637 int btrfs_return_cluster_to_free_space(
2638                                struct btrfs_block_group_cache *block_group,
2639                                struct btrfs_free_cluster *cluster)
2640 {
2641         struct btrfs_free_space_ctl *ctl;
2642         int ret;
2643
2644         /* first, get a safe pointer to the block group */
2645         spin_lock(&cluster->lock);
2646         if (!block_group) {
2647                 block_group = cluster->block_group;
2648                 if (!block_group) {
2649                         spin_unlock(&cluster->lock);
2650                         return 0;
2651                 }
2652         } else if (cluster->block_group != block_group) {
2653                 /* someone else has already freed it don't redo their work */
2654                 spin_unlock(&cluster->lock);
2655                 return 0;
2656         }
2657         atomic_inc(&block_group->count);
2658         spin_unlock(&cluster->lock);
2659
2660         ctl = block_group->free_space_ctl;
2661
2662         /* now return any extents the cluster had on it */
2663         spin_lock(&ctl->tree_lock);
2664         ret = __btrfs_return_cluster_to_free_space(block_group, cluster);
2665         spin_unlock(&ctl->tree_lock);
2666
2667         /* finally drop our ref */
2668         btrfs_put_block_group(block_group);
2669         return ret;
2670 }
2671
2672 static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group_cache *block_group,
2673                                    struct btrfs_free_cluster *cluster,
2674                                    struct btrfs_free_space *entry,
2675                                    u64 bytes, u64 min_start,
2676                                    u64 *max_extent_size)
2677 {
2678         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2679         int err;
2680         u64 search_start = cluster->window_start;
2681         u64 search_bytes = bytes;
2682         u64 ret = 0;
2683
2684         search_start = min_start;
2685         search_bytes = bytes;
2686
2687         err = search_bitmap(ctl, entry, &search_start, &search_bytes, true);
2688         if (err) {
2689                 if (search_bytes > *max_extent_size)
2690                         *max_extent_size = search_bytes;
2691                 return 0;
2692         }
2693
2694         ret = search_start;
2695         __bitmap_clear_bits(ctl, entry, ret, bytes);
2696
2697         return ret;
2698 }
2699
2700 /*
2701  * given a cluster, try to allocate 'bytes' from it, returns 0
2702  * if it couldn't find anything suitably large, or a logical disk offset
2703  * if things worked out
2704  */
2705 u64 btrfs_alloc_from_cluster(struct btrfs_block_group_cache *block_group,
2706                              struct btrfs_free_cluster *cluster, u64 bytes,
2707                              u64 min_start, u64 *max_extent_size)
2708 {
2709         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2710         struct btrfs_free_space *entry = NULL;
2711         struct rb_node *node;
2712         u64 ret = 0;
2713
2714         spin_lock(&cluster->lock);
2715         if (bytes > cluster->max_size)
2716                 goto out;
2717
2718         if (cluster->block_group != block_group)
2719                 goto out;
2720
2721         node = rb_first(&cluster->root);
2722         if (!node)
2723                 goto out;
2724
2725         entry = rb_entry(node, struct btrfs_free_space, offset_index);
2726         while (1) {
2727                 if (entry->bytes < bytes && entry->bytes > *max_extent_size)
2728                         *max_extent_size = entry->bytes;
2729
2730                 if (entry->bytes < bytes ||
2731                     (!entry->bitmap && entry->offset < min_start)) {
2732                         node = rb_next(&entry->offset_index);
2733                         if (!node)
2734                                 break;
2735                         entry = rb_entry(node, struct btrfs_free_space,
2736                                          offset_index);
2737                         continue;
2738                 }
2739
2740                 if (entry->bitmap) {
2741                         ret = btrfs_alloc_from_bitmap(block_group,
2742                                                       cluster, entry, bytes,
2743                                                       cluster->window_start,
2744                                                       max_extent_size);
2745                         if (ret == 0) {
2746                                 node = rb_next(&entry->offset_index);
2747                                 if (!node)
2748                                         break;
2749                                 entry = rb_entry(node, struct btrfs_free_space,
2750                                                  offset_index);
2751                                 continue;
2752                         }
2753                         cluster->window_start += bytes;
2754                 } else {
2755                         ret = entry->offset;
2756
2757                         entry->offset += bytes;
2758                         entry->bytes -= bytes;
2759                 }
2760
2761                 if (entry->bytes == 0)
2762                         rb_erase(&entry->offset_index, &cluster->root);
2763                 break;
2764         }
2765 out:
2766         spin_unlock(&cluster->lock);
2767
2768         if (!ret)
2769                 return 0;
2770
2771         spin_lock(&ctl->tree_lock);
2772
2773         ctl->free_space -= bytes;
2774         if (entry->bytes == 0) {
2775                 ctl->free_extents--;
2776                 if (entry->bitmap) {
2777                         kfree(entry->bitmap);
2778                         ctl->total_bitmaps--;
2779                         ctl->op->recalc_thresholds(ctl);
2780                 }
2781                 kmem_cache_free(btrfs_free_space_cachep, entry);
2782         }
2783
2784         spin_unlock(&ctl->tree_lock);
2785
2786         return ret;
2787 }
2788
2789 static int btrfs_bitmap_cluster(struct btrfs_block_group_cache *block_group,
2790                                 struct btrfs_free_space *entry,
2791                                 struct btrfs_free_cluster *cluster,
2792                                 u64 offset, u64 bytes,
2793                                 u64 cont1_bytes, u64 min_bytes)
2794 {
2795         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2796         unsigned long next_zero;
2797         unsigned long i;
2798         unsigned long want_bits;
2799         unsigned long min_bits;
2800         unsigned long found_bits;
2801         unsigned long max_bits = 0;
2802         unsigned long start = 0;
2803         unsigned long total_found = 0;
2804         int ret;
2805
2806         i = offset_to_bit(entry->offset, ctl->unit,
2807                           max_t(u64, offset, entry->offset));
2808         want_bits = bytes_to_bits(bytes, ctl->unit);
2809         min_bits = bytes_to_bits(min_bytes, ctl->unit);
2810
2811         /*
2812          * Don't bother looking for a cluster in this bitmap if it's heavily
2813          * fragmented.
2814          */
2815         if (entry->max_extent_size &&
2816             entry->max_extent_size < cont1_bytes)
2817                 return -ENOSPC;
2818 again:
2819         found_bits = 0;
2820         for_each_set_bit_from(i, entry->bitmap, BITS_PER_BITMAP) {
2821                 next_zero = find_next_zero_bit(entry->bitmap,
2822                                                BITS_PER_BITMAP, i);
2823                 if (next_zero - i >= min_bits) {
2824                         found_bits = next_zero - i;
2825                         if (found_bits > max_bits)
2826                                 max_bits = found_bits;
2827                         break;
2828                 }
2829                 if (next_zero - i > max_bits)
2830                         max_bits = next_zero - i;
2831                 i = next_zero;
2832         }
2833
2834         if (!found_bits) {
2835                 entry->max_extent_size = (u64)max_bits * ctl->unit;
2836                 return -ENOSPC;
2837         }
2838
2839         if (!total_found) {
2840                 start = i;
2841                 cluster->max_size = 0;
2842         }
2843
2844         total_found += found_bits;
2845
2846         if (cluster->max_size < found_bits * ctl->unit)
2847                 cluster->max_size = found_bits * ctl->unit;
2848
2849         if (total_found < want_bits || cluster->max_size < cont1_bytes) {
2850                 i = next_zero + 1;
2851                 goto again;
2852         }
2853
2854         cluster->window_start = start * ctl->unit + entry->offset;
2855         rb_erase(&entry->offset_index, &ctl->free_space_offset);
2856         ret = tree_insert_offset(&cluster->root, entry->offset,
2857                                  &entry->offset_index, 1);
2858         ASSERT(!ret); /* -EEXIST; Logic error */
2859
2860         trace_btrfs_setup_cluster(block_group, cluster,
2861                                   total_found * ctl->unit, 1);
2862         return 0;
2863 }
2864
2865 /*
2866  * This searches the block group for just extents to fill the cluster with.
2867  * Try to find a cluster with at least bytes total bytes, at least one
2868  * extent of cont1_bytes, and other clusters of at least min_bytes.
2869  */
2870 static noinline int
2871 setup_cluster_no_bitmap(struct btrfs_block_group_cache *block_group,
2872                         struct btrfs_free_cluster *cluster,
2873                         struct list_head *bitmaps, u64 offset, u64 bytes,
2874                         u64 cont1_bytes, u64 min_bytes)
2875 {
2876         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2877         struct btrfs_free_space *first = NULL;
2878         struct btrfs_free_space *entry = NULL;
2879         struct btrfs_free_space *last;
2880         struct rb_node *node;
2881         u64 window_free;
2882         u64 max_extent;
2883         u64 total_size = 0;
2884
2885         entry = tree_search_offset(ctl, offset, 0, 1);
2886         if (!entry)
2887                 return -ENOSPC;
2888
2889         /*
2890          * We don't want bitmaps, so just move along until we find a normal
2891          * extent entry.
2892          */
2893         while (entry->bitmap || entry->bytes < min_bytes) {
2894                 if (entry->bitmap && list_empty(&entry->list))
2895                         list_add_tail(&entry->list, bitmaps);
2896                 node = rb_next(&entry->offset_index);
2897                 if (!node)
2898                         return -ENOSPC;
2899                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2900         }
2901
2902         window_free = entry->bytes;
2903         max_extent = entry->bytes;
2904         first = entry;
2905         last = entry;
2906
2907         for (node = rb_next(&entry->offset_index); node;
2908              node = rb_next(&entry->offset_index)) {
2909                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2910
2911                 if (entry->bitmap) {
2912                         if (list_empty(&entry->list))
2913                                 list_add_tail(&entry->list, bitmaps);
2914                         continue;
2915                 }
2916
2917                 if (entry->bytes < min_bytes)
2918                         continue;
2919
2920                 last = entry;
2921                 window_free += entry->bytes;
2922                 if (entry->bytes > max_extent)
2923                         max_extent = entry->bytes;
2924         }
2925
2926         if (window_free < bytes || max_extent < cont1_bytes)
2927                 return -ENOSPC;
2928
2929         cluster->window_start = first->offset;
2930
2931         node = &first->offset_index;
2932
2933         /*
2934          * now we've found our entries, pull them out of the free space
2935          * cache and put them into the cluster rbtree
2936          */
2937         do {
2938                 int ret;
2939
2940                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2941                 node = rb_next(&entry->offset_index);
2942                 if (entry->bitmap || entry->bytes < min_bytes)
2943                         continue;
2944
2945                 rb_erase(&entry->offset_index, &ctl->free_space_offset);
2946                 ret = tree_insert_offset(&cluster->root, entry->offset,
2947                                          &entry->offset_index, 0);
2948                 total_size += entry->bytes;
2949                 ASSERT(!ret); /* -EEXIST; Logic error */
2950         } while (node && entry != last);
2951
2952         cluster->max_size = max_extent;
2953         trace_btrfs_setup_cluster(block_group, cluster, total_size, 0);
2954         return 0;
2955 }
2956
2957 /*
2958  * This specifically looks for bitmaps that may work in the cluster, we assume
2959  * that we have already failed to find extents that will work.
2960  */
2961 static noinline int
2962 setup_cluster_bitmap(struct btrfs_block_group_cache *block_group,
2963                      struct btrfs_free_cluster *cluster,
2964                      struct list_head *bitmaps, u64 offset, u64 bytes,
2965                      u64 cont1_bytes, u64 min_bytes)
2966 {
2967         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2968         struct btrfs_free_space *entry = NULL;
2969         int ret = -ENOSPC;
2970         u64 bitmap_offset = offset_to_bitmap(ctl, offset);
2971
2972         if (ctl->total_bitmaps == 0)
2973                 return -ENOSPC;
2974
2975         /*
2976          * The bitmap that covers offset won't be in the list unless offset
2977          * is just its start offset.
2978          */
2979         if (!list_empty(bitmaps))
2980                 entry = list_first_entry(bitmaps, struct btrfs_free_space, list);
2981
2982         if (!entry || entry->offset != bitmap_offset) {
2983                 entry = tree_search_offset(ctl, bitmap_offset, 1, 0);
2984                 if (entry && list_empty(&entry->list))
2985                         list_add(&entry->list, bitmaps);
2986         }
2987
2988         list_for_each_entry(entry, bitmaps, list) {
2989                 if (entry->bytes < bytes)
2990                         continue;
2991                 ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
2992                                            bytes, cont1_bytes, min_bytes);
2993                 if (!ret)
2994                         return 0;
2995         }
2996
2997         /*
2998          * The bitmaps list has all the bitmaps that record free space
2999          * starting after offset, so no more search is required.
3000          */
3001         return -ENOSPC;
3002 }
3003
3004 /*
3005  * here we try to find a cluster of blocks in a block group.  The goal
3006  * is to find at least bytes+empty_size.
3007  * We might not find them all in one contiguous area.
3008  *
3009  * returns zero and sets up cluster if things worked out, otherwise
3010  * it returns -enospc
3011  */
3012 int btrfs_find_space_cluster(struct btrfs_root *root,
3013                              struct btrfs_block_group_cache *block_group,
3014                              struct btrfs_free_cluster *cluster,
3015                              u64 offset, u64 bytes, u64 empty_size)
3016 {
3017         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3018         struct btrfs_free_space *entry, *tmp;
3019         LIST_HEAD(bitmaps);
3020         u64 min_bytes;
3021         u64 cont1_bytes;
3022         int ret;
3023
3024         /*
3025          * Choose the minimum extent size we'll require for this
3026          * cluster.  For SSD_SPREAD, don't allow any fragmentation.
3027          * For metadata, allow allocates with smaller extents.  For
3028          * data, keep it dense.
3029          */
3030         if (btrfs_test_opt(root->fs_info, SSD_SPREAD)) {
3031                 cont1_bytes = min_bytes = bytes + empty_size;
3032         } else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
3033                 cont1_bytes = bytes;
3034                 min_bytes = block_group->sectorsize;
3035         } else {
3036                 cont1_bytes = max(bytes, (bytes + empty_size) >> 2);
3037                 min_bytes = block_group->sectorsize;
3038         }
3039
3040         spin_lock(&ctl->tree_lock);
3041
3042         /*
3043          * If we know we don't have enough space to make a cluster don't even
3044          * bother doing all the work to try and find one.
3045          */
3046         if (ctl->free_space < bytes) {
3047                 spin_unlock(&ctl->tree_lock);
3048                 return -ENOSPC;
3049         }
3050
3051         spin_lock(&cluster->lock);
3052
3053         /* someone already found a cluster, hooray */
3054         if (cluster->block_group) {
3055                 ret = 0;
3056                 goto out;
3057         }
3058
3059         trace_btrfs_find_cluster(block_group, offset, bytes, empty_size,
3060                                  min_bytes);
3061
3062         ret = setup_cluster_no_bitmap(block_group, cluster, &bitmaps, offset,
3063                                       bytes + empty_size,
3064                                       cont1_bytes, min_bytes);
3065         if (ret)
3066                 ret = setup_cluster_bitmap(block_group, cluster, &bitmaps,
3067                                            offset, bytes + empty_size,
3068                                            cont1_bytes, min_bytes);
3069
3070         /* Clear our temporary list */
3071         list_for_each_entry_safe(entry, tmp, &bitmaps, list)
3072                 list_del_init(&entry->list);
3073
3074         if (!ret) {
3075                 atomic_inc(&block_group->count);
3076                 list_add_tail(&cluster->block_group_list,
3077                               &block_group->cluster_list);
3078                 cluster->block_group = block_group;
3079         } else {
3080                 trace_btrfs_failed_cluster_setup(block_group);
3081         }
3082 out:
3083         spin_unlock(&cluster->lock);
3084         spin_unlock(&ctl->tree_lock);
3085
3086         return ret;
3087 }
3088
3089 /*
3090  * simple code to zero out a cluster
3091  */
3092 void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
3093 {
3094         spin_lock_init(&cluster->lock);
3095         spin_lock_init(&cluster->refill_lock);
3096         cluster->root = RB_ROOT;
3097         cluster->max_size = 0;
3098         cluster->fragmented = false;
3099         INIT_LIST_HEAD(&cluster->block_group_list);
3100         cluster->block_group = NULL;
3101 }
3102
3103 static int do_trimming(struct btrfs_block_group_cache *block_group,
3104                        u64 *total_trimmed, u64 start, u64 bytes,
3105                        u64 reserved_start, u64 reserved_bytes,
3106                        struct btrfs_trim_range *trim_entry)
3107 {
3108         struct btrfs_space_info *space_info = block_group->space_info;
3109         struct btrfs_fs_info *fs_info = block_group->fs_info;
3110         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3111         int ret;
3112         int update = 0;
3113         u64 trimmed = 0;
3114
3115         spin_lock(&space_info->lock);
3116         spin_lock(&block_group->lock);
3117         if (!block_group->ro) {
3118                 block_group->reserved += reserved_bytes;
3119                 space_info->bytes_reserved += reserved_bytes;
3120                 update = 1;
3121         }
3122         spin_unlock(&block_group->lock);
3123         spin_unlock(&space_info->lock);
3124
3125         ret = btrfs_discard_extent(fs_info->extent_root,
3126                                    start, bytes, &trimmed);
3127         if (!ret)
3128                 *total_trimmed += trimmed;
3129
3130         mutex_lock(&ctl->cache_writeout_mutex);
3131         btrfs_add_free_space(block_group, reserved_start, reserved_bytes);
3132         list_del(&trim_entry->list);
3133         mutex_unlock(&ctl->cache_writeout_mutex);
3134
3135         if (update) {
3136                 spin_lock(&space_info->lock);
3137                 spin_lock(&block_group->lock);
3138                 if (block_group->ro)
3139                         space_info->bytes_readonly += reserved_bytes;
3140                 block_group->reserved -= reserved_bytes;
3141                 space_info->bytes_reserved -= reserved_bytes;
3142                 spin_unlock(&space_info->lock);
3143                 spin_unlock(&block_group->lock);
3144         }
3145
3146         return ret;
3147 }
3148
3149 static int trim_no_bitmap(struct btrfs_block_group_cache *block_group,
3150                           u64 *total_trimmed, u64 start, u64 end, u64 minlen)
3151 {
3152         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3153         struct btrfs_free_space *entry;
3154         struct rb_node *node;
3155         int ret = 0;
3156         u64 extent_start;
3157         u64 extent_bytes;
3158         u64 bytes;
3159
3160         while (start < end) {
3161                 struct btrfs_trim_range trim_entry;
3162
3163                 mutex_lock(&ctl->cache_writeout_mutex);
3164                 spin_lock(&ctl->tree_lock);
3165
3166                 if (ctl->free_space < minlen) {
3167                         spin_unlock(&ctl->tree_lock);
3168                         mutex_unlock(&ctl->cache_writeout_mutex);
3169                         break;
3170                 }
3171
3172                 entry = tree_search_offset(ctl, start, 0, 1);
3173                 if (!entry) {
3174                         spin_unlock(&ctl->tree_lock);
3175                         mutex_unlock(&ctl->cache_writeout_mutex);
3176                         break;
3177                 }
3178
3179                 /* skip bitmaps */
3180                 while (entry->bitmap) {
3181                         node = rb_next(&entry->offset_index);
3182                         if (!node) {
3183                                 spin_unlock(&ctl->tree_lock);
3184                                 mutex_unlock(&ctl->cache_writeout_mutex);
3185                                 goto out;
3186                         }
3187                         entry = rb_entry(node, struct btrfs_free_space,
3188                                          offset_index);
3189                 }
3190
3191                 if (entry->offset >= end) {
3192                         spin_unlock(&ctl->tree_lock);
3193                         mutex_unlock(&ctl->cache_writeout_mutex);
3194                         break;
3195                 }
3196
3197                 extent_start = entry->offset;
3198                 extent_bytes = entry->bytes;
3199                 start = max(start, extent_start);
3200                 bytes = min(extent_start + extent_bytes, end) - start;
3201                 if (bytes < minlen) {
3202                         spin_unlock(&ctl->tree_lock);
3203                         mutex_unlock(&ctl->cache_writeout_mutex);
3204                         goto next;
3205                 }
3206
3207                 unlink_free_space(ctl, entry);
3208                 kmem_cache_free(btrfs_free_space_cachep, entry);
3209
3210                 spin_unlock(&ctl->tree_lock);
3211                 trim_entry.start = extent_start;
3212                 trim_entry.bytes = extent_bytes;
3213                 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3214                 mutex_unlock(&ctl->cache_writeout_mutex);
3215
3216                 ret = do_trimming(block_group, total_trimmed, start, bytes,
3217                                   extent_start, extent_bytes, &trim_entry);
3218                 if (ret)
3219                         break;
3220 next:
3221                 start += bytes;
3222
3223                 if (fatal_signal_pending(current)) {
3224                         ret = -ERESTARTSYS;
3225                         break;
3226                 }
3227
3228                 cond_resched();
3229         }
3230 out:
3231         return ret;
3232 }
3233
3234 static int trim_bitmaps(struct btrfs_block_group_cache *block_group,
3235                         u64 *total_trimmed, u64 start, u64 end, u64 minlen)
3236 {
3237         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3238         struct btrfs_free_space *entry;
3239         int ret = 0;
3240         int ret2;
3241         u64 bytes;
3242         u64 offset = offset_to_bitmap(ctl, start);
3243
3244         while (offset < end) {
3245                 bool next_bitmap = false;
3246                 struct btrfs_trim_range trim_entry;
3247
3248                 mutex_lock(&ctl->cache_writeout_mutex);
3249                 spin_lock(&ctl->tree_lock);
3250
3251                 if (ctl->free_space < minlen) {
3252                         spin_unlock(&ctl->tree_lock);
3253                         mutex_unlock(&ctl->cache_writeout_mutex);
3254                         break;
3255                 }
3256
3257                 entry = tree_search_offset(ctl, offset, 1, 0);
3258                 if (!entry) {
3259                         spin_unlock(&ctl->tree_lock);
3260                         mutex_unlock(&ctl->cache_writeout_mutex);
3261                         next_bitmap = true;
3262                         goto next;
3263                 }
3264
3265                 bytes = minlen;
3266                 ret2 = search_bitmap(ctl, entry, &start, &bytes, false);
3267                 if (ret2 || start >= end) {
3268                         spin_unlock(&ctl->tree_lock);
3269                         mutex_unlock(&ctl->cache_writeout_mutex);
3270                         next_bitmap = true;
3271                         goto next;
3272                 }
3273
3274                 bytes = min(bytes, end - start);
3275                 if (bytes < minlen) {
3276                         spin_unlock(&ctl->tree_lock);
3277                         mutex_unlock(&ctl->cache_writeout_mutex);
3278                         goto next;
3279                 }
3280
3281                 bitmap_clear_bits(ctl, entry, start, bytes);
3282                 if (entry->bytes == 0)
3283                         free_bitmap(ctl, entry);
3284
3285                 spin_unlock(&ctl->tree_lock);
3286                 trim_entry.start = start;
3287                 trim_entry.bytes = bytes;
3288                 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3289                 mutex_unlock(&ctl->cache_writeout_mutex);
3290
3291                 ret = do_trimming(block_group, total_trimmed, start, bytes,
3292                                   start, bytes, &trim_entry);
3293                 if (ret)
3294                         break;
3295 next:
3296                 if (next_bitmap) {
3297                         offset += BITS_PER_BITMAP * ctl->unit;
3298                 } else {
3299                         start += bytes;
3300                         if (start >= offset + BITS_PER_BITMAP * ctl->unit)
3301                                 offset += BITS_PER_BITMAP * ctl->unit;
3302                 }
3303
3304                 if (fatal_signal_pending(current)) {
3305                         ret = -ERESTARTSYS;
3306                         break;
3307                 }
3308
3309                 cond_resched();
3310         }
3311
3312         return ret;
3313 }
3314
3315 void btrfs_get_block_group_trimming(struct btrfs_block_group_cache *cache)
3316 {
3317         atomic_inc(&cache->trimming);
3318 }
3319
3320 void btrfs_put_block_group_trimming(struct btrfs_block_group_cache *block_group)
3321 {
3322         struct extent_map_tree *em_tree;
3323         struct extent_map *em;
3324         bool cleanup;
3325
3326         spin_lock(&block_group->lock);
3327         cleanup = (atomic_dec_and_test(&block_group->trimming) &&
3328                    block_group->removed);
3329         spin_unlock(&block_group->lock);
3330
3331         if (cleanup) {
3332                 lock_chunks(block_group->fs_info->chunk_root);
3333                 em_tree = &block_group->fs_info->mapping_tree.map_tree;
3334                 write_lock(&em_tree->lock);
3335                 em = lookup_extent_mapping(em_tree, block_group->key.objectid,
3336                                            1);
3337                 BUG_ON(!em); /* logic error, can't happen */
3338                 /*
3339                  * remove_extent_mapping() will delete us from the pinned_chunks
3340                  * list, which is protected by the chunk mutex.
3341                  */
3342                 remove_extent_mapping(em_tree, em);
3343                 write_unlock(&em_tree->lock);
3344                 unlock_chunks(block_group->fs_info->chunk_root);
3345
3346                 /* once for us and once for the tree */
3347                 free_extent_map(em);
3348                 free_extent_map(em);
3349
3350                 /*
3351                  * We've left one free space entry and other tasks trimming
3352                  * this block group have left 1 entry each one. Free them.
3353                  */
3354                 __btrfs_remove_free_space_cache(block_group->free_space_ctl);
3355         }
3356 }
3357
3358 int btrfs_trim_block_group(struct btrfs_block_group_cache *block_group,
3359                            u64 *trimmed, u64 start, u64 end, u64 minlen)
3360 {
3361         int ret;
3362
3363         *trimmed = 0;
3364
3365         spin_lock(&block_group->lock);
3366         if (block_group->removed) {
3367                 spin_unlock(&block_group->lock);
3368                 return 0;
3369         }
3370         btrfs_get_block_group_trimming(block_group);
3371         spin_unlock(&block_group->lock);
3372
3373         ret = trim_no_bitmap(block_group, trimmed, start, end, minlen);
3374         if (ret)
3375                 goto out;
3376
3377         ret = trim_bitmaps(block_group, trimmed, start, end, minlen);
3378 out:
3379         btrfs_put_block_group_trimming(block_group);
3380         return ret;
3381 }
3382
3383 /*
3384  * Find the left-most item in the cache tree, and then return the
3385  * smallest inode number in the item.
3386  *
3387  * Note: the returned inode number may not be the smallest one in
3388  * the tree, if the left-most item is a bitmap.
3389  */
3390 u64 btrfs_find_ino_for_alloc(struct btrfs_root *fs_root)
3391 {
3392         struct btrfs_free_space_ctl *ctl = fs_root->free_ino_ctl;
3393         struct btrfs_free_space *entry = NULL;
3394         u64 ino = 0;
3395
3396         spin_lock(&ctl->tree_lock);
3397
3398         if (RB_EMPTY_ROOT(&ctl->free_space_offset))
3399                 goto out;
3400
3401         entry = rb_entry(rb_first(&ctl->free_space_offset),
3402                          struct btrfs_free_space, offset_index);
3403
3404         if (!entry->bitmap) {
3405                 ino = entry->offset;
3406
3407                 unlink_free_space(ctl, entry);
3408                 entry->offset++;
3409                 entry->bytes--;
3410                 if (!entry->bytes)
3411                         kmem_cache_free(btrfs_free_space_cachep, entry);
3412                 else
3413                         link_free_space(ctl, entry);
3414         } else {
3415                 u64 offset = 0;
3416                 u64 count = 1;
3417                 int ret;
3418
3419                 ret = search_bitmap(ctl, entry, &offset, &count, true);
3420                 /* Logic error; Should be empty if it can't find anything */
3421                 ASSERT(!ret);
3422
3423                 ino = offset;
3424                 bitmap_clear_bits(ctl, entry, offset, 1);
3425                 if (entry->bytes == 0)
3426                         free_bitmap(ctl, entry);
3427         }
3428 out:
3429         spin_unlock(&ctl->tree_lock);
3430
3431         return ino;
3432 }
3433
3434 struct inode *lookup_free_ino_inode(struct btrfs_root *root,
3435                                     struct btrfs_path *path)
3436 {
3437         struct inode *inode = NULL;
3438
3439         spin_lock(&root->ino_cache_lock);
3440         if (root->ino_cache_inode)
3441                 inode = igrab(root->ino_cache_inode);
3442         spin_unlock(&root->ino_cache_lock);
3443         if (inode)
3444                 return inode;
3445
3446         inode = __lookup_free_space_inode(root, path, 0);
3447         if (IS_ERR(inode))
3448                 return inode;
3449
3450         spin_lock(&root->ino_cache_lock);
3451         if (!btrfs_fs_closing(root->fs_info))
3452                 root->ino_cache_inode = igrab(inode);
3453         spin_unlock(&root->ino_cache_lock);
3454
3455         return inode;
3456 }
3457
3458 int create_free_ino_inode(struct btrfs_root *root,
3459                           struct btrfs_trans_handle *trans,
3460                           struct btrfs_path *path)
3461 {
3462         return __create_free_space_inode(root, trans, path,
3463                                          BTRFS_FREE_INO_OBJECTID, 0);
3464 }
3465
3466 int load_free_ino_cache(struct btrfs_fs_info *fs_info, struct btrfs_root *root)
3467 {
3468         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
3469         struct btrfs_path *path;
3470         struct inode *inode;
3471         int ret = 0;
3472         u64 root_gen = btrfs_root_generation(&root->root_item);
3473
3474         if (!btrfs_test_opt(root->fs_info, INODE_MAP_CACHE))
3475                 return 0;
3476
3477         /*
3478          * If we're unmounting then just return, since this does a search on the
3479          * normal root and not the commit root and we could deadlock.
3480          */
3481         if (btrfs_fs_closing(fs_info))
3482                 return 0;
3483
3484         path = btrfs_alloc_path();
3485         if (!path)
3486                 return 0;
3487
3488         inode = lookup_free_ino_inode(root, path);
3489         if (IS_ERR(inode))
3490                 goto out;
3491
3492         if (root_gen != BTRFS_I(inode)->generation)
3493                 goto out_put;
3494
3495         ret = __load_free_space_cache(root, inode, ctl, path, 0);
3496
3497         if (ret < 0)
3498                 btrfs_err(fs_info,
3499                         "failed to load free ino cache for root %llu",
3500                         root->root_key.objectid);
3501 out_put:
3502         iput(inode);
3503 out:
3504         btrfs_free_path(path);
3505         return ret;
3506 }
3507
3508 int btrfs_write_out_ino_cache(struct btrfs_root *root,
3509                               struct btrfs_trans_handle *trans,
3510                               struct btrfs_path *path,
3511                               struct inode *inode)
3512 {
3513         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
3514         int ret;
3515         struct btrfs_io_ctl io_ctl;
3516         bool release_metadata = true;
3517
3518         if (!btrfs_test_opt(root->fs_info, INODE_MAP_CACHE))
3519                 return 0;
3520
3521         memset(&io_ctl, 0, sizeof(io_ctl));
3522         ret = __btrfs_write_out_cache(root, inode, ctl, NULL, &io_ctl,
3523                                       trans, path, 0);
3524         if (!ret) {
3525                 /*
3526                  * At this point writepages() didn't error out, so our metadata
3527                  * reservation is released when the writeback finishes, at
3528                  * inode.c:btrfs_finish_ordered_io(), regardless of it finishing
3529                  * with or without an error.
3530                  */
3531                 release_metadata = false;
3532                 ret = btrfs_wait_cache_io(root, trans, NULL, &io_ctl, path, 0);
3533         }
3534
3535         if (ret) {
3536                 if (release_metadata)
3537                         btrfs_delalloc_release_metadata(inode, inode->i_size);
3538 #ifdef DEBUG
3539                 btrfs_err(root->fs_info,
3540                         "failed to write free ino cache for root %llu",
3541                         root->root_key.objectid);
3542 #endif
3543         }
3544
3545         return ret;
3546 }
3547
3548 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
3549 /*
3550  * Use this if you need to make a bitmap or extent entry specifically, it
3551  * doesn't do any of the merging that add_free_space does, this acts a lot like
3552  * how the free space cache loading stuff works, so you can get really weird
3553  * configurations.
3554  */
3555 int test_add_free_space_entry(struct btrfs_block_group_cache *cache,
3556                               u64 offset, u64 bytes, bool bitmap)
3557 {
3558         struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
3559         struct btrfs_free_space *info = NULL, *bitmap_info;
3560         void *map = NULL;
3561         u64 bytes_added;
3562         int ret;
3563
3564 again:
3565         if (!info) {
3566                 info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
3567                 if (!info)
3568                         return -ENOMEM;
3569         }
3570
3571         if (!bitmap) {
3572                 spin_lock(&ctl->tree_lock);
3573                 info->offset = offset;
3574                 info->bytes = bytes;
3575                 info->max_extent_size = 0;
3576                 ret = link_free_space(ctl, info);
3577                 spin_unlock(&ctl->tree_lock);
3578                 if (ret)
3579                         kmem_cache_free(btrfs_free_space_cachep, info);
3580                 return ret;
3581         }
3582
3583         if (!map) {
3584                 map = kzalloc(PAGE_SIZE, GFP_NOFS);
3585                 if (!map) {
3586                         kmem_cache_free(btrfs_free_space_cachep, info);
3587                         return -ENOMEM;
3588                 }
3589         }
3590
3591         spin_lock(&ctl->tree_lock);
3592         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
3593                                          1, 0);
3594         if (!bitmap_info) {
3595                 info->bitmap = map;
3596                 map = NULL;
3597                 add_new_bitmap(ctl, info, offset);
3598                 bitmap_info = info;
3599                 info = NULL;
3600         }
3601
3602         bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes);
3603
3604         bytes -= bytes_added;
3605         offset += bytes_added;
3606         spin_unlock(&ctl->tree_lock);
3607
3608         if (bytes)
3609                 goto again;
3610
3611         if (info)
3612                 kmem_cache_free(btrfs_free_space_cachep, info);
3613         if (map)
3614                 kfree(map);
3615         return 0;
3616 }
3617
3618 /*
3619  * Checks to see if the given range is in the free space cache.  This is really
3620  * just used to check the absence of space, so if there is free space in the
3621  * range at all we will return 1.
3622  */
3623 int test_check_exists(struct btrfs_block_group_cache *cache,
3624                       u64 offset, u64 bytes)
3625 {
3626         struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
3627         struct btrfs_free_space *info;
3628         int ret = 0;
3629
3630         spin_lock(&ctl->tree_lock);
3631         info = tree_search_offset(ctl, offset, 0, 0);
3632         if (!info) {
3633                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
3634                                           1, 0);
3635                 if (!info)
3636                         goto out;
3637         }
3638
3639 have_info:
3640         if (info->bitmap) {
3641                 u64 bit_off, bit_bytes;
3642                 struct rb_node *n;
3643                 struct btrfs_free_space *tmp;
3644
3645                 bit_off = offset;
3646                 bit_bytes = ctl->unit;
3647                 ret = search_bitmap(ctl, info, &bit_off, &bit_bytes, false);
3648                 if (!ret) {
3649                         if (bit_off == offset) {
3650                                 ret = 1;
3651                                 goto out;
3652                         } else if (bit_off > offset &&
3653                                    offset + bytes > bit_off) {
3654                                 ret = 1;
3655                                 goto out;
3656                         }
3657                 }
3658
3659                 n = rb_prev(&info->offset_index);
3660                 while (n) {
3661                         tmp = rb_entry(n, struct btrfs_free_space,
3662                                        offset_index);
3663                         if (tmp->offset + tmp->bytes < offset)
3664                                 break;
3665                         if (offset + bytes < tmp->offset) {
3666                                 n = rb_prev(&tmp->offset_index);
3667                                 continue;
3668                         }
3669                         info = tmp;
3670                         goto have_info;
3671                 }
3672
3673                 n = rb_next(&info->offset_index);
3674                 while (n) {
3675                         tmp = rb_entry(n, struct btrfs_free_space,
3676                                        offset_index);
3677                         if (offset + bytes < tmp->offset)
3678                                 break;
3679                         if (tmp->offset + tmp->bytes < offset) {
3680                                 n = rb_next(&tmp->offset_index);
3681                                 continue;
3682                         }
3683                         info = tmp;
3684                         goto have_info;
3685                 }
3686
3687                 ret = 0;
3688                 goto out;
3689         }
3690
3691         if (info->offset == offset) {
3692                 ret = 1;
3693                 goto out;
3694         }
3695
3696         if (offset > info->offset && offset < info->offset + info->bytes)
3697                 ret = 1;
3698 out:
3699         spin_unlock(&ctl->tree_lock);
3700         return ret;
3701 }
3702 #endif /* CONFIG_BTRFS_FS_RUN_SANITY_TESTS */