fs/btrfs/free-space-cache.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2008 Red Hat.  All rights reserved.
   4  */
   5
   6 #include <linux/pagemap.h>
   7 #include <linux/sched.h>
   8 #include <linux/sched/signal.h>
   9 #include <linux/slab.h>
  10 #include <linux/math64.h>
  11 #include <linux/ratelimit.h>
  12 #include <linux/error-injection.h>
  13 #include <linux/sched/mm.h>
  14 #include "ctree.h"
  15 #include "fs.h"
  16 #include "messages.h"
  17 #include "misc.h"
  18 #include "free-space-cache.h"
  19 #include "transaction.h"
  20 #include "disk-io.h"
  21 #include "extent_io.h"
  22 #include "space-info.h"
  23 #include "block-group.h"
  24 #include "discard.h"
  25 #include "subpage.h"
  26 #include "inode-item.h"
  27 #include "accessors.h"
  28 #include "file-item.h"
  29 #include "file.h"
  30 #include "super.h"
  31
  32 #define BITS_PER_BITMAP         (PAGE_SIZE * 8UL)
  33 #define MAX_CACHE_BYTES_PER_GIG SZ_64K
  34 #define FORCE_EXTENT_THRESHOLD  SZ_1M
  35
  36 static struct kmem_cache *btrfs_free_space_cachep;
  37 static struct kmem_cache *btrfs_free_space_bitmap_cachep;
  38
  39 struct btrfs_trim_range {
  40         u64 start;
  41         u64 bytes;
  42         struct list_head list;
  43 };
  44
  45 static int link_free_space(struct btrfs_free_space_ctl *ctl,
  46                            struct btrfs_free_space *info);
  47 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
  48                               struct btrfs_free_space *info, bool update_stat);
  49 static int search_bitmap(struct btrfs_free_space_ctl *ctl,
  50                          struct btrfs_free_space *bitmap_info, u64 *offset,
  51                          u64 *bytes, bool for_alloc);
  52 static void free_bitmap(struct btrfs_free_space_ctl *ctl,
  53                         struct btrfs_free_space *bitmap_info);
  54 static void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
  55                               struct btrfs_free_space *info, u64 offset,
  56                               u64 bytes, bool update_stats);
  57
  58 static void btrfs_crc32c_final(u32 crc, u8 *result)
  59 {
  60         put_unaligned_le32(~crc, result);
  61 }
  62
  63 static void __btrfs_remove_free_space_cache(struct btrfs_free_space_ctl *ctl)
  64 {
  65         struct btrfs_free_space *info;
  66         struct rb_node *node;
  67
  68         while ((node = rb_last(&ctl->free_space_offset)) != NULL) {
  69                 info = rb_entry(node, struct btrfs_free_space, offset_index);
  70                 if (!info->bitmap) {
  71                         unlink_free_space(ctl, info, true);
  72                         kmem_cache_free(btrfs_free_space_cachep, info);
  73                 } else {
  74                         free_bitmap(ctl, info);
  75                 }
  76
  77                 cond_resched_lock(&ctl->tree_lock);
  78         }
  79 }
  80
  81 static struct inode *__lookup_free_space_inode(struct btrfs_root *root,
  82                                                struct btrfs_path *path,
  83                                                u64 offset)
  84 {
  85         struct btrfs_fs_info *fs_info = root->fs_info;
  86         struct btrfs_key key;
  87         struct btrfs_key location;
  88         struct btrfs_disk_key disk_key;
  89         struct btrfs_free_space_header *header;
  90         struct extent_buffer *leaf;
  91         struct inode *inode = NULL;
  92         unsigned nofs_flag;
  93         int ret;
  94
  95         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
  96         key.offset = offset;
  97         key.type = 0;
  98
  99         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
 100         if (ret < 0)
 101                 return ERR_PTR(ret);
 102         if (ret > 0) {
 103                 btrfs_release_path(path);
 104                 return ERR_PTR(-ENOENT);
 105         }
 106
 107         leaf = path->nodes[0];
 108         header = btrfs_item_ptr(leaf, path->slots[0],
 109                                 struct btrfs_free_space_header);
 110         btrfs_free_space_key(leaf, header, &disk_key);
 111         btrfs_disk_key_to_cpu(&location, &disk_key);
 112         btrfs_release_path(path);
 113
 114         /*
 115          * We are often under a trans handle at this point, so we need to make
 116          * sure NOFS is set to keep us from deadlocking.
 117          */
 118         nofs_flag = memalloc_nofs_save();
 119         inode = btrfs_iget_path(fs_info->sb, location.objectid, root, path);
 120         btrfs_release_path(path);
 121         memalloc_nofs_restore(nofs_flag);
 122         if (IS_ERR(inode))
 123                 return inode;
 124
 125         mapping_set_gfp_mask(inode->i_mapping,
 126                         mapping_gfp_constraint(inode->i_mapping,
 127                         ~(__GFP_FS | __GFP_HIGHMEM)));
 128
 129         return inode;
 130 }
 131
 132 struct inode *lookup_free_space_inode(struct btrfs_block_group *block_group,
 133                 struct btrfs_path *path)
 134 {
 135         struct btrfs_fs_info *fs_info = block_group->fs_info;
 136         struct inode *inode = NULL;
 137         u32 flags = BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
 138
 139         spin_lock(&block_group->lock);
 140         if (block_group->inode)
 141                 inode = igrab(block_group->inode);
 142         spin_unlock(&block_group->lock);
 143         if (inode)
 144                 return inode;
 145
 146         inode = __lookup_free_space_inode(fs_info->tree_root, path,
 147                                           block_group->start);
 148         if (IS_ERR(inode))
 149                 return inode;
 150
 151         spin_lock(&block_group->lock);
 152         if (!((BTRFS_I(inode)->flags & flags) == flags)) {
 153                 btrfs_info(fs_info, "Old style space inode found, converting.");
 154                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM |
 155                         BTRFS_INODE_NODATACOW;
 156                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 157         }
 158
 159         if (!test_and_set_bit(BLOCK_GROUP_FLAG_IREF, &block_group->runtime_flags))
 160                 block_group->inode = igrab(inode);
 161         spin_unlock(&block_group->lock);
 162
 163         return inode;
 164 }
 165
 166 static int __create_free_space_inode(struct btrfs_root *root,
 167                                      struct btrfs_trans_handle *trans,
 168                                      struct btrfs_path *path,
 169                                      u64 ino, u64 offset)
 170 {
 171         struct btrfs_key key;
 172         struct btrfs_disk_key disk_key;
 173         struct btrfs_free_space_header *header;
 174         struct btrfs_inode_item *inode_item;
 175         struct extent_buffer *leaf;
 176         /* We inline CRCs for the free disk space cache */
 177         const u64 flags = BTRFS_INODE_NOCOMPRESS | BTRFS_INODE_PREALLOC |
 178                           BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
 179         int ret;
 180
 181         ret = btrfs_insert_empty_inode(trans, root, path, ino);
 182         if (ret)
 183                 return ret;
 184
 185         leaf = path->nodes[0];
 186         inode_item = btrfs_item_ptr(leaf, path->slots[0],
 187                                     struct btrfs_inode_item);
 188         btrfs_item_key(leaf, &disk_key, path->slots[0]);
 189         memzero_extent_buffer(leaf, (unsigned long)inode_item,
 190                              sizeof(*inode_item));
 191         btrfs_set_inode_generation(leaf, inode_item, trans->transid);
 192         btrfs_set_inode_size(leaf, inode_item, 0);
 193         btrfs_set_inode_nbytes(leaf, inode_item, 0);
 194         btrfs_set_inode_uid(leaf, inode_item, 0);
 195         btrfs_set_inode_gid(leaf, inode_item, 0);
 196         btrfs_set_inode_mode(leaf, inode_item, S_IFREG | 0600);
 197         btrfs_set_inode_flags(leaf, inode_item, flags);
 198         btrfs_set_inode_nlink(leaf, inode_item, 1);
 199         btrfs_set_inode_transid(leaf, inode_item, trans->transid);
 200         btrfs_set_inode_block_group(leaf, inode_item, offset);
 201         btrfs_mark_buffer_dirty(trans, leaf);
 202         btrfs_release_path(path);
 203
 204         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 205         key.offset = offset;
 206         key.type = 0;
 207         ret = btrfs_insert_empty_item(trans, root, path, &key,
 208                                       sizeof(struct btrfs_free_space_header));
 209         if (ret < 0) {
 210                 btrfs_release_path(path);
 211                 return ret;
 212         }
 213
 214         leaf = path->nodes[0];
 215         header = btrfs_item_ptr(leaf, path->slots[0],
 216                                 struct btrfs_free_space_header);
 217         memzero_extent_buffer(leaf, (unsigned long)header, sizeof(*header));
 218         btrfs_set_free_space_key(leaf, header, &disk_key);
 219         btrfs_mark_buffer_dirty(trans, leaf);
 220         btrfs_release_path(path);
 221
 222         return 0;
 223 }
 224
 225 int create_free_space_inode(struct btrfs_trans_handle *trans,
 226                             struct btrfs_block_group *block_group,
 227                             struct btrfs_path *path)
 228 {
 229         int ret;
 230         u64 ino;
 231
 232         ret = btrfs_get_free_objectid(trans->fs_info->tree_root, &ino);
 233         if (ret < 0)
 234                 return ret;
 235
 236         return __create_free_space_inode(trans->fs_info->tree_root, trans, path,
 237                                          ino, block_group->start);
 238 }
 239
 240 /*
 241  * inode is an optional sink: if it is NULL, btrfs_remove_free_space_inode
 242  * handles lookup, otherwise it takes ownership and iputs the inode.
 243  * Don't reuse an inode pointer after passing it into this function.
 244  */
 245 int btrfs_remove_free_space_inode(struct btrfs_trans_handle *trans,
 246                                   struct inode *inode,
 247                                   struct btrfs_block_group *block_group)
 248 {
 249         struct btrfs_path *path;
 250         struct btrfs_key key;
 251         int ret = 0;
 252
 253         path = btrfs_alloc_path();
 254         if (!path)
 255                 return -ENOMEM;
 256
 257         if (!inode)
 258                 inode = lookup_free_space_inode(block_group, path);
 259         if (IS_ERR(inode)) {
 260                 if (PTR_ERR(inode) != -ENOENT)
 261                         ret = PTR_ERR(inode);
 262                 goto out;
 263         }
 264         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
 265         if (ret) {
 266                 btrfs_add_delayed_iput(BTRFS_I(inode));
 267                 goto out;
 268         }
 269         clear_nlink(inode);
 270         /* One for the block groups ref */
 271         spin_lock(&block_group->lock);
 272         if (test_and_clear_bit(BLOCK_GROUP_FLAG_IREF, &block_group->runtime_flags)) {
 273                 block_group->inode = NULL;
 274                 spin_unlock(&block_group->lock);
 275                 iput(inode);
 276         } else {
 277                 spin_unlock(&block_group->lock);
 278         }
 279         /* One for the lookup ref */
 280         btrfs_add_delayed_iput(BTRFS_I(inode));
 281
 282         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 283         key.type = 0;
 284         key.offset = block_group->start;
 285         ret = btrfs_search_slot(trans, trans->fs_info->tree_root, &key, path,
 286                                 -1, 1);
 287         if (ret) {
 288                 if (ret > 0)
 289                         ret = 0;
 290                 goto out;
 291         }
 292         ret = btrfs_del_item(trans, trans->fs_info->tree_root, path);
 293 out:
 294         btrfs_free_path(path);
 295         return ret;
 296 }
 297
 298 int btrfs_truncate_free_space_cache(struct btrfs_trans_handle *trans,
 299                                     struct btrfs_block_group *block_group,
 300                                     struct inode *vfs_inode)
 301 {
 302         struct btrfs_truncate_control control = {
 303                 .inode = BTRFS_I(vfs_inode),
 304                 .new_size = 0,
 305                 .ino = btrfs_ino(BTRFS_I(vfs_inode)),
 306                 .min_type = BTRFS_EXTENT_DATA_KEY,
 307                 .clear_extent_range = true,
 308         };
 309         struct btrfs_inode *inode = BTRFS_I(vfs_inode);
 310         struct btrfs_root *root = inode->root;
 311         struct extent_state *cached_state = NULL;
 312         int ret = 0;
 313         bool locked = false;
 314
 315         if (block_group) {
 316                 struct btrfs_path *path = btrfs_alloc_path();
 317
 318                 if (!path) {
 319                         ret = -ENOMEM;
 320                         goto fail;
 321                 }
 322                 locked = true;
 323                 mutex_lock(&trans->transaction->cache_write_mutex);
 324                 if (!list_empty(&block_group->io_list)) {
 325                         list_del_init(&block_group->io_list);
 326
 327                         btrfs_wait_cache_io(trans, block_group, path);
 328                         btrfs_put_block_group(block_group);
 329                 }
 330
 331                 /*
 332                  * now that we've truncated the cache away, its no longer
 333                  * setup or written
 334                  */
 335                 spin_lock(&block_group->lock);
 336                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 337                 spin_unlock(&block_group->lock);
 338                 btrfs_free_path(path);
 339         }
 340
 341         btrfs_i_size_write(inode, 0);
 342         truncate_pagecache(vfs_inode, 0);
 343
 344         lock_extent(&inode->io_tree, 0, (u64)-1, &cached_state);
 345         btrfs_drop_extent_map_range(inode, 0, (u64)-1, false);
 346
 347         /*
 348          * We skip the throttling logic for free space cache inodes, so we don't
 349          * need to check for -EAGAIN.
 350          */
 351         ret = btrfs_truncate_inode_items(trans, root, &control);
 352
 353         inode_sub_bytes(&inode->vfs_inode, control.sub_bytes);
 354         btrfs_inode_safe_disk_i_size_write(inode, control.last_size);
 355
 356         unlock_extent(&inode->io_tree, 0, (u64)-1, &cached_state);
 357         if (ret)
 358                 goto fail;
 359
 360         ret = btrfs_update_inode(trans, inode);
 361
 362 fail:
 363         if (locked)
 364                 mutex_unlock(&trans->transaction->cache_write_mutex);
 365         if (ret)
 366                 btrfs_abort_transaction(trans, ret);
 367
 368         return ret;
 369 }
 370
 371 static void readahead_cache(struct inode *inode)
 372 {
 373         struct file_ra_state ra;
 374         unsigned long last_index;
 375
 376         file_ra_state_init(&ra, inode->i_mapping);
 377         last_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
 378
 379         page_cache_sync_readahead(inode->i_mapping, &ra, NULL, 0, last_index);
 380 }
 381
 382 static int io_ctl_init(struct btrfs_io_ctl *io_ctl, struct inode *inode,
 383                        int write)
 384 {
 385         int num_pages;
 386
 387         num_pages = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
 388
 389         /* Make sure we can fit our crcs and generation into the first page */
 390         if (write && (num_pages * sizeof(u32) + sizeof(u64)) > PAGE_SIZE)
 391                 return -ENOSPC;
 392
 393         memset(io_ctl, 0, sizeof(struct btrfs_io_ctl));
 394
 395         io_ctl->pages = kcalloc(num_pages, sizeof(struct page *), GFP_NOFS);
 396         if (!io_ctl->pages)
 397                 return -ENOMEM;
 398
 399         io_ctl->num_pages = num_pages;
 400         io_ctl->fs_info = inode_to_fs_info(inode);
 401         io_ctl->inode = inode;
 402
 403         return 0;
 404 }
 405 ALLOW_ERROR_INJECTION(io_ctl_init, ERRNO);
 406
 407 static void io_ctl_free(struct btrfs_io_ctl *io_ctl)
 408 {
 409         kfree(io_ctl->pages);
 410         io_ctl->pages = NULL;
 411 }
 412
 413 static void io_ctl_unmap_page(struct btrfs_io_ctl *io_ctl)
 414 {
 415         if (io_ctl->cur) {
 416                 io_ctl->cur = NULL;
 417                 io_ctl->orig = NULL;
 418         }
 419 }
 420
 421 static void io_ctl_map_page(struct btrfs_io_ctl *io_ctl, int clear)
 422 {
 423         ASSERT(io_ctl->index < io_ctl->num_pages);
 424         io_ctl->page = io_ctl->pages[io_ctl->index++];
 425         io_ctl->cur = page_address(io_ctl->page);
 426         io_ctl->orig = io_ctl->cur;
 427         io_ctl->size = PAGE_SIZE;
 428         if (clear)
 429                 clear_page(io_ctl->cur);
 430 }
 431
 432 static void io_ctl_drop_pages(struct btrfs_io_ctl *io_ctl)
 433 {
 434         int i;
 435
 436         io_ctl_unmap_page(io_ctl);
 437
 438         for (i = 0; i < io_ctl->num_pages; i++) {
 439                 if (io_ctl->pages[i]) {
 440                         btrfs_folio_clear_checked(io_ctl->fs_info,
 441                                         page_folio(io_ctl->pages[i]),
 442                                         page_offset(io_ctl->pages[i]),
 443                                         PAGE_SIZE);
 444                         unlock_page(io_ctl->pages[i]);
 445                         put_page(io_ctl->pages[i]);
 446                 }
 447         }
 448 }
 449
 450 static int io_ctl_prepare_pages(struct btrfs_io_ctl *io_ctl, bool uptodate)
 451 {
 452         struct page *page;
 453         struct inode *inode = io_ctl->inode;
 454         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
 455         int i;
 456
 457         for (i = 0; i < io_ctl->num_pages; i++) {
 458                 int ret;
 459
 460                 page = find_or_create_page(inode->i_mapping, i, mask);
 461                 if (!page) {
 462                         io_ctl_drop_pages(io_ctl);
 463                         return -ENOMEM;
 464                 }
 465
 466                 ret = set_page_extent_mapped(page);
 467                 if (ret < 0) {
 468                         unlock_page(page);
 469                         put_page(page);
 470                         io_ctl_drop_pages(io_ctl);
 471                         return ret;
 472                 }
 473
 474                 io_ctl->pages[i] = page;
 475                 if (uptodate && !PageUptodate(page)) {
 476                         btrfs_read_folio(NULL, page_folio(page));
 477                         lock_page(page);
 478                         if (page->mapping != inode->i_mapping) {
 479                                 btrfs_err(BTRFS_I(inode)->root->fs_info,
 480                                           "free space cache page truncated");
 481                                 io_ctl_drop_pages(io_ctl);
 482                                 return -EIO;
 483                         }
 484                         if (!PageUptodate(page)) {
 485                                 btrfs_err(BTRFS_I(inode)->root->fs_info,
 486                                            "error reading free space cache");
 487                                 io_ctl_drop_pages(io_ctl);
 488                                 return -EIO;
 489                         }
 490                 }
 491         }
 492
 493         for (i = 0; i < io_ctl->num_pages; i++)
 494                 clear_page_dirty_for_io(io_ctl->pages[i]);
 495
 496         return 0;
 497 }
 498
 499 static void io_ctl_set_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
 500 {
 501         io_ctl_map_page(io_ctl, 1);
 502
 503         /*
 504          * Skip the csum areas.  If we don't check crcs then we just have a
 505          * 64bit chunk at the front of the first page.
 506          */
 507         io_ctl->cur += (sizeof(u32) * io_ctl->num_pages);
 508         io_ctl->size -= sizeof(u64) + (sizeof(u32) * io_ctl->num_pages);
 509
 510         put_unaligned_le64(generation, io_ctl->cur);
 511         io_ctl->cur += sizeof(u64);
 512 }
 513
 514 static int io_ctl_check_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
 515 {
 516         u64 cache_gen;
 517
 518         /*
 519          * Skip the crc area.  If we don't check crcs then we just have a 64bit
 520          * chunk at the front of the first page.
 521          */
 522         io_ctl->cur += sizeof(u32) * io_ctl->num_pages;
 523         io_ctl->size -= sizeof(u64) + (sizeof(u32) * io_ctl->num_pages);
 524
 525         cache_gen = get_unaligned_le64(io_ctl->cur);
 526         if (cache_gen != generation) {
 527                 btrfs_err_rl(io_ctl->fs_info,
 528                         "space cache generation (%llu) does not match inode (%llu)",
 529                                 cache_gen, generation);
 530                 io_ctl_unmap_page(io_ctl);
 531                 return -EIO;
 532         }
 533         io_ctl->cur += sizeof(u64);
 534         return 0;
 535 }
 536
 537 static void io_ctl_set_crc(struct btrfs_io_ctl *io_ctl, int index)
 538 {
 539         u32 *tmp;
 540         u32 crc = ~(u32)0;
 541         unsigned offset = 0;
 542
 543         if (index == 0)
 544                 offset = sizeof(u32) * io_ctl->num_pages;
 545
 546         crc = crc32c(crc, io_ctl->orig + offset, PAGE_SIZE - offset);
 547         btrfs_crc32c_final(crc, (u8 *)&crc);
 548         io_ctl_unmap_page(io_ctl);
 549         tmp = page_address(io_ctl->pages[0]);
 550         tmp += index;
 551         *tmp = crc;
 552 }
 553
 554 static int io_ctl_check_crc(struct btrfs_io_ctl *io_ctl, int index)
 555 {
 556         u32 *tmp, val;
 557         u32 crc = ~(u32)0;
 558         unsigned offset = 0;
 559
 560         if (index == 0)
 561                 offset = sizeof(u32) * io_ctl->num_pages;
 562
 563         tmp = page_address(io_ctl->pages[0]);
 564         tmp += index;
 565         val = *tmp;
 566
 567         io_ctl_map_page(io_ctl, 0);
 568         crc = crc32c(crc, io_ctl->orig + offset, PAGE_SIZE - offset);
 569         btrfs_crc32c_final(crc, (u8 *)&crc);
 570         if (val != crc) {
 571                 btrfs_err_rl(io_ctl->fs_info,
 572                         "csum mismatch on free space cache");
 573                 io_ctl_unmap_page(io_ctl);
 574                 return -EIO;
 575         }
 576
 577         return 0;
 578 }
 579
 580 static int io_ctl_add_entry(struct btrfs_io_ctl *io_ctl, u64 offset, u64 bytes,
 581                             void *bitmap)
 582 {
 583         struct btrfs_free_space_entry *entry;
 584
 585         if (!io_ctl->cur)
 586                 return -ENOSPC;
 587
 588         entry = io_ctl->cur;
 589         put_unaligned_le64(offset, &entry->offset);
 590         put_unaligned_le64(bytes, &entry->bytes);
 591         entry->type = (bitmap) ? BTRFS_FREE_SPACE_BITMAP :
 592                 BTRFS_FREE_SPACE_EXTENT;
 593         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
 594         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
 595
 596         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
 597                 return 0;
 598
 599         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 600
 601         /* No more pages to map */
 602         if (io_ctl->index >= io_ctl->num_pages)
 603                 return 0;
 604
 605         /* map the next page */
 606         io_ctl_map_page(io_ctl, 1);
 607         return 0;
 608 }
 609
 610 static int io_ctl_add_bitmap(struct btrfs_io_ctl *io_ctl, void *bitmap)
 611 {
 612         if (!io_ctl->cur)
 613                 return -ENOSPC;
 614
 615         /*
 616          * If we aren't at the start of the current page, unmap this one and
 617          * map the next one if there is any left.
 618          */
 619         if (io_ctl->cur != io_ctl->orig) {
 620                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 621                 if (io_ctl->index >= io_ctl->num_pages)
 622                         return -ENOSPC;
 623                 io_ctl_map_page(io_ctl, 0);
 624         }
 625
 626         copy_page(io_ctl->cur, bitmap);
 627         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 628         if (io_ctl->index < io_ctl->num_pages)
 629                 io_ctl_map_page(io_ctl, 0);
 630         return 0;
 631 }
 632
 633 static void io_ctl_zero_remaining_pages(struct btrfs_io_ctl *io_ctl)
 634 {
 635         /*
 636          * If we're not on the boundary we know we've modified the page and we
 637          * need to crc the page.
 638          */
 639         if (io_ctl->cur != io_ctl->orig)
 640                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 641         else
 642                 io_ctl_unmap_page(io_ctl);
 643
 644         while (io_ctl->index < io_ctl->num_pages) {
 645                 io_ctl_map_page(io_ctl, 1);
 646                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 647         }
 648 }
 649
 650 static int io_ctl_read_entry(struct btrfs_io_ctl *io_ctl,
 651                             struct btrfs_free_space *entry, u8 *type)
 652 {
 653         struct btrfs_free_space_entry *e;
 654         int ret;
 655
 656         if (!io_ctl->cur) {
 657                 ret = io_ctl_check_crc(io_ctl, io_ctl->index);
 658                 if (ret)
 659                         return ret;
 660         }
 661
 662         e = io_ctl->cur;
 663         entry->offset = get_unaligned_le64(&e->offset);
 664         entry->bytes = get_unaligned_le64(&e->bytes);
 665         *type = e->type;
 666         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
 667         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
 668
 669         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
 670                 return 0;
 671
 672         io_ctl_unmap_page(io_ctl);
 673
 674         return 0;
 675 }
 676
 677 static int io_ctl_read_bitmap(struct btrfs_io_ctl *io_ctl,
 678                               struct btrfs_free_space *entry)
 679 {
 680         int ret;
 681
 682         ret = io_ctl_check_crc(io_ctl, io_ctl->index);
 683         if (ret)
 684                 return ret;
 685
 686         copy_page(entry->bitmap, io_ctl->cur);
 687         io_ctl_unmap_page(io_ctl);
 688
 689         return 0;
 690 }
 691
 692 static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
 693 {
 694         struct btrfs_block_group *block_group = ctl->block_group;
 695         u64 max_bytes;
 696         u64 bitmap_bytes;
 697         u64 extent_bytes;
 698         u64 size = block_group->length;
 699         u64 bytes_per_bg = BITS_PER_BITMAP * ctl->unit;
 700         u64 max_bitmaps = div64_u64(size + bytes_per_bg - 1, bytes_per_bg);
 701
 702         max_bitmaps = max_t(u64, max_bitmaps, 1);
 703
 704         if (ctl->total_bitmaps > max_bitmaps)
 705                 btrfs_err(block_group->fs_info,
 706 "invalid free space control: bg start=%llu len=%llu total_bitmaps=%u unit=%u max_bitmaps=%llu bytes_per_bg=%llu",
 707                           block_group->start, block_group->length,
 708                           ctl->total_bitmaps, ctl->unit, max_bitmaps,
 709                           bytes_per_bg);
 710         ASSERT(ctl->total_bitmaps <= max_bitmaps);
 711
 712         /*
 713          * We are trying to keep the total amount of memory used per 1GiB of
 714          * space to be MAX_CACHE_BYTES_PER_GIG.  However, with a reclamation
 715          * mechanism of pulling extents >= FORCE_EXTENT_THRESHOLD out of
 716          * bitmaps, we may end up using more memory than this.
 717          */
 718         if (size < SZ_1G)
 719                 max_bytes = MAX_CACHE_BYTES_PER_GIG;
 720         else
 721                 max_bytes = MAX_CACHE_BYTES_PER_GIG * div_u64(size, SZ_1G);
 722
 723         bitmap_bytes = ctl->total_bitmaps * ctl->unit;
 724
 725         /*
 726          * we want the extent entry threshold to always be at most 1/2 the max
 727          * bytes we can have, or whatever is less than that.
 728          */
 729         extent_bytes = max_bytes - bitmap_bytes;
 730         extent_bytes = min_t(u64, extent_bytes, max_bytes >> 1);
 731
 732         ctl->extents_thresh =
 733                 div_u64(extent_bytes, sizeof(struct btrfs_free_space));
 734 }
 735
 736 static int __load_free_space_cache(struct btrfs_root *root, struct inode *inode,
 737                                    struct btrfs_free_space_ctl *ctl,
 738                                    struct btrfs_path *path, u64 offset)
 739 {
 740         struct btrfs_fs_info *fs_info = root->fs_info;
 741         struct btrfs_free_space_header *header;
 742         struct extent_buffer *leaf;
 743         struct btrfs_io_ctl io_ctl;
 744         struct btrfs_key key;
 745         struct btrfs_free_space *e, *n;
 746         LIST_HEAD(bitmaps);
 747         u64 num_entries;
 748         u64 num_bitmaps;
 749         u64 generation;
 750         u8 type;
 751         int ret = 0;
 752
 753         /* Nothing in the space cache, goodbye */
 754         if (!i_size_read(inode))
 755                 return 0;
 756
 757         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 758         key.offset = offset;
 759         key.type = 0;
 760
 761         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
 762         if (ret < 0)
 763                 return 0;
 764         else if (ret > 0) {
 765                 btrfs_release_path(path);
 766                 return 0;
 767         }
 768
 769         ret = -1;
 770
 771         leaf = path->nodes[0];
 772         header = btrfs_item_ptr(leaf, path->slots[0],
 773                                 struct btrfs_free_space_header);
 774         num_entries = btrfs_free_space_entries(leaf, header);
 775         num_bitmaps = btrfs_free_space_bitmaps(leaf, header);
 776         generation = btrfs_free_space_generation(leaf, header);
 777         btrfs_release_path(path);
 778
 779         if (!BTRFS_I(inode)->generation) {
 780                 btrfs_info(fs_info,
 781                            "the free space cache file (%llu) is invalid, skip it",
 782                            offset);
 783                 return 0;
 784         }
 785
 786         if (BTRFS_I(inode)->generation != generation) {
 787                 btrfs_err(fs_info,
 788                           "free space inode generation (%llu) did not match free space cache generation (%llu)",
 789                           BTRFS_I(inode)->generation, generation);
 790                 return 0;
 791         }
 792
 793         if (!num_entries)
 794                 return 0;
 795
 796         ret = io_ctl_init(&io_ctl, inode, 0);
 797         if (ret)
 798                 return ret;
 799
 800         readahead_cache(inode);
 801
 802         ret = io_ctl_prepare_pages(&io_ctl, true);
 803         if (ret)
 804                 goto out;
 805
 806         ret = io_ctl_check_crc(&io_ctl, 0);
 807         if (ret)
 808                 goto free_cache;
 809
 810         ret = io_ctl_check_generation(&io_ctl, generation);
 811         if (ret)
 812                 goto free_cache;
 813
 814         while (num_entries) {
 815                 e = kmem_cache_zalloc(btrfs_free_space_cachep,
 816                                       GFP_NOFS);
 817                 if (!e) {
 818                         ret = -ENOMEM;
 819                         goto free_cache;
 820                 }
 821
 822                 ret = io_ctl_read_entry(&io_ctl, e, &type);
 823                 if (ret) {
 824                         kmem_cache_free(btrfs_free_space_cachep, e);
 825                         goto free_cache;
 826                 }
 827
 828                 if (!e->bytes) {
 829                         ret = -1;
 830                         kmem_cache_free(btrfs_free_space_cachep, e);
 831                         goto free_cache;
 832                 }
 833
 834                 if (type == BTRFS_FREE_SPACE_EXTENT) {
 835                         spin_lock(&ctl->tree_lock);
 836                         ret = link_free_space(ctl, e);
 837                         spin_unlock(&ctl->tree_lock);
 838                         if (ret) {
 839                                 btrfs_err(fs_info,
 840                                         "Duplicate entries in free space cache, dumping");
 841                                 kmem_cache_free(btrfs_free_space_cachep, e);
 842                                 goto free_cache;
 843                         }
 844                 } else {
 845                         ASSERT(num_bitmaps);
 846                         num_bitmaps--;
 847                         e->bitmap = kmem_cache_zalloc(
 848                                         btrfs_free_space_bitmap_cachep, GFP_NOFS);
 849                         if (!e->bitmap) {
 850                                 ret = -ENOMEM;
 851                                 kmem_cache_free(
 852                                         btrfs_free_space_cachep, e);
 853                                 goto free_cache;
 854                         }
 855                         spin_lock(&ctl->tree_lock);
 856                         ret = link_free_space(ctl, e);
 857                         if (ret) {
 858                                 spin_unlock(&ctl->tree_lock);
 859                                 btrfs_err(fs_info,
 860                                         "Duplicate entries in free space cache, dumping");
 861                                 kmem_cache_free(btrfs_free_space_cachep, e);
 862                                 goto free_cache;
 863                         }
 864                         ctl->total_bitmaps++;
 865                         recalculate_thresholds(ctl);
 866                         spin_unlock(&ctl->tree_lock);
 867                         list_add_tail(&e->list, &bitmaps);
 868                 }
 869
 870                 num_entries--;
 871         }
 872
 873         io_ctl_unmap_page(&io_ctl);
 874
 875         /*
 876          * We add the bitmaps at the end of the entries in order that
 877          * the bitmap entries are added to the cache.
 878          */
 879         list_for_each_entry_safe(e, n, &bitmaps, list) {
 880                 list_del_init(&e->list);
 881                 ret = io_ctl_read_bitmap(&io_ctl, e);
 882                 if (ret)
 883                         goto free_cache;
 884         }
 885
 886         io_ctl_drop_pages(&io_ctl);
 887         ret = 1;
 888 out:
 889         io_ctl_free(&io_ctl);
 890         return ret;
 891 free_cache:
 892         io_ctl_drop_pages(&io_ctl);
 893
 894         spin_lock(&ctl->tree_lock);
 895         __btrfs_remove_free_space_cache(ctl);
 896         spin_unlock(&ctl->tree_lock);
 897         goto out;
 898 }
 899
 900 static int copy_free_space_cache(struct btrfs_block_group *block_group,
 901                                  struct btrfs_free_space_ctl *ctl)
 902 {
 903         struct btrfs_free_space *info;
 904         struct rb_node *n;
 905         int ret = 0;
 906
 907         while (!ret && (n = rb_first(&ctl->free_space_offset)) != NULL) {
 908                 info = rb_entry(n, struct btrfs_free_space, offset_index);
 909                 if (!info->bitmap) {
 910                         const u64 offset = info->offset;
 911                         const u64 bytes = info->bytes;
 912
 913                         unlink_free_space(ctl, info, true);
 914                         spin_unlock(&ctl->tree_lock);
 915                         kmem_cache_free(btrfs_free_space_cachep, info);
 916                         ret = btrfs_add_free_space(block_group, offset, bytes);
 917                         spin_lock(&ctl->tree_lock);
 918                 } else {
 919                         u64 offset = info->offset;
 920                         u64 bytes = ctl->unit;
 921
 922                         ret = search_bitmap(ctl, info, &offset, &bytes, false);
 923                         if (ret == 0) {
 924                                 bitmap_clear_bits(ctl, info, offset, bytes, true);
 925                                 spin_unlock(&ctl->tree_lock);
 926                                 ret = btrfs_add_free_space(block_group, offset,
 927                                                            bytes);
 928                                 spin_lock(&ctl->tree_lock);
 929                         } else {
 930                                 free_bitmap(ctl, info);
 931                                 ret = 0;
 932                         }
 933                 }
 934                 cond_resched_lock(&ctl->tree_lock);
 935         }
 936         return ret;
 937 }
 938
 939 static struct lock_class_key btrfs_free_space_inode_key;
 940
 941 int load_free_space_cache(struct btrfs_block_group *block_group)
 942 {
 943         struct btrfs_fs_info *fs_info = block_group->fs_info;
 944         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
 945         struct btrfs_free_space_ctl tmp_ctl = {};
 946         struct inode *inode;
 947         struct btrfs_path *path;
 948         int ret = 0;
 949         bool matched;
 950         u64 used = block_group->used;
 951
 952         /*
 953          * Because we could potentially discard our loaded free space, we want
 954          * to load everything into a temporary structure first, and then if it's
 955          * valid copy it all into the actual free space ctl.
 956          */
 957         btrfs_init_free_space_ctl(block_group, &tmp_ctl);
 958
 959         /*
 960          * If this block group has been marked to be cleared for one reason or
 961          * another then we can't trust the on disk cache, so just return.
 962          */
 963         spin_lock(&block_group->lock);
 964         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
 965                 spin_unlock(&block_group->lock);
 966                 return 0;
 967         }
 968         spin_unlock(&block_group->lock);
 969
 970         path = btrfs_alloc_path();
 971         if (!path)
 972                 return 0;
 973         path->search_commit_root = 1;
 974         path->skip_locking = 1;
 975
 976         /*
 977          * We must pass a path with search_commit_root set to btrfs_iget in
 978          * order to avoid a deadlock when allocating extents for the tree root.
 979          *
 980          * When we are COWing an extent buffer from the tree root, when looking
 981          * for a free extent, at extent-tree.c:find_free_extent(), we can find
 982          * block group without its free space cache loaded. When we find one
 983          * we must load its space cache which requires reading its free space
 984          * cache's inode item from the root tree. If this inode item is located
 985          * in the same leaf that we started COWing before, then we end up in
 986          * deadlock on the extent buffer (trying to read lock it when we
 987          * previously write locked it).
 988          *
 989          * It's safe to read the inode item using the commit root because
 990          * block groups, once loaded, stay in memory forever (until they are
 991          * removed) as well as their space caches once loaded. New block groups
 992          * once created get their ->cached field set to BTRFS_CACHE_FINISHED so
 993          * we will never try to read their inode item while the fs is mounted.
 994          */
 995         inode = lookup_free_space_inode(block_group, path);
 996         if (IS_ERR(inode)) {
 997                 btrfs_free_path(path);
 998                 return 0;
 999         }
1000
1001         /* We may have converted the inode and made the cache invalid. */
1002         spin_lock(&block_group->lock);
1003         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
1004                 spin_unlock(&block_group->lock);
1005                 btrfs_free_path(path);
1006                 goto out;
1007         }
1008         spin_unlock(&block_group->lock);
1009
1010         /*
1011          * Reinitialize the class of struct inode's mapping->invalidate_lock for
1012          * free space inodes to prevent false positives related to locks for normal
1013          * inodes.
1014          */
1015         lockdep_set_class(&(&inode->i_data)->invalidate_lock,
1016                           &btrfs_free_space_inode_key);
1017
1018         ret = __load_free_space_cache(fs_info->tree_root, inode, &tmp_ctl,
1019                                       path, block_group->start);
1020         btrfs_free_path(path);
1021         if (ret <= 0)
1022                 goto out;
1023
1024         matched = (tmp_ctl.free_space == (block_group->length - used -
1025                                           block_group->bytes_super));
1026
1027         if (matched) {
1028                 spin_lock(&tmp_ctl.tree_lock);
1029                 ret = copy_free_space_cache(block_group, &tmp_ctl);
1030                 spin_unlock(&tmp_ctl.tree_lock);
1031                 /*
1032                  * ret == 1 means we successfully loaded the free space cache,
1033                  * so we need to re-set it here.
1034                  */
1035                 if (ret == 0)
1036                         ret = 1;
1037         } else {
1038                 /*
1039                  * We need to call the _locked variant so we don't try to update
1040                  * the discard counters.
1041                  */
1042                 spin_lock(&tmp_ctl.tree_lock);
1043                 __btrfs_remove_free_space_cache(&tmp_ctl);
1044                 spin_unlock(&tmp_ctl.tree_lock);
1045                 btrfs_warn(fs_info,
1046                            "block group %llu has wrong amount of free space",
1047                            block_group->start);
1048                 ret = -1;
1049         }
1050 out:
1051         if (ret < 0) {
1052                 /* This cache is bogus, make sure it gets cleared */
1053                 spin_lock(&block_group->lock);
1054                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
1055                 spin_unlock(&block_group->lock);
1056                 ret = 0;
1057
1058                 btrfs_warn(fs_info,
1059                            "failed to load free space cache for block group %llu, rebuilding it now",
1060                            block_group->start);
1061         }
1062
1063         spin_lock(&ctl->tree_lock);
1064         btrfs_discard_update_discardable(block_group);
1065         spin_unlock(&ctl->tree_lock);
1066         iput(inode);
1067         return ret;
1068 }
1069
1070 static noinline_for_stack
1071 int write_cache_extent_entries(struct btrfs_io_ctl *io_ctl,
1072                               struct btrfs_free_space_ctl *ctl,
1073                               struct btrfs_block_group *block_group,
1074                               int *entries, int *bitmaps,
1075                               struct list_head *bitmap_list)
1076 {
1077         int ret;
1078         struct btrfs_free_cluster *cluster = NULL;
1079         struct btrfs_free_cluster *cluster_locked = NULL;
1080         struct rb_node *node = rb_first(&ctl->free_space_offset);
1081         struct btrfs_trim_range *trim_entry;
1082
1083         /* Get the cluster for this block_group if it exists */
1084         if (block_group && !list_empty(&block_group->cluster_list)) {
1085                 cluster = list_entry(block_group->cluster_list.next,
1086                                      struct btrfs_free_cluster,
1087                                      block_group_list);
1088         }
1089
1090         if (!node && cluster) {
1091                 cluster_locked = cluster;
1092                 spin_lock(&cluster_locked->lock);
1093                 node = rb_first(&cluster->root);
1094                 cluster = NULL;
1095         }
1096
1097         /* Write out the extent entries */
1098         while (node) {
1099                 struct btrfs_free_space *e;
1100
1101                 e = rb_entry(node, struct btrfs_free_space, offset_index);
1102                 *entries += 1;
1103
1104                 ret = io_ctl_add_entry(io_ctl, e->offset, e->bytes,
1105                                        e->bitmap);
1106                 if (ret)
1107                         goto fail;
1108
1109                 if (e->bitmap) {
1110                         list_add_tail(&e->list, bitmap_list);
1111                         *bitmaps += 1;
1112                 }
1113                 node = rb_next(node);
1114                 if (!node && cluster) {
1115                         node = rb_first(&cluster->root);
1116                         cluster_locked = cluster;
1117                         spin_lock(&cluster_locked->lock);
1118                         cluster = NULL;
1119                 }
1120         }
1121         if (cluster_locked) {
1122                 spin_unlock(&cluster_locked->lock);
1123                 cluster_locked = NULL;
1124         }
1125
1126         /*
1127          * Make sure we don't miss any range that was removed from our rbtree
1128          * because trimming is running. Otherwise after a umount+mount (or crash
1129          * after committing the transaction) we would leak free space and get
1130          * an inconsistent free space cache report from fsck.
1131          */
1132         list_for_each_entry(trim_entry, &ctl->trimming_ranges, list) {
1133                 ret = io_ctl_add_entry(io_ctl, trim_entry->start,
1134                                        trim_entry->bytes, NULL);
1135                 if (ret)
1136                         goto fail;
1137                 *entries += 1;
1138         }
1139
1140         return 0;
1141 fail:
1142         if (cluster_locked)
1143                 spin_unlock(&cluster_locked->lock);
1144         return -ENOSPC;
1145 }
1146
1147 static noinline_for_stack int
1148 update_cache_item(struct btrfs_trans_handle *trans,
1149                   struct btrfs_root *root,
1150                   struct inode *inode,
1151                   struct btrfs_path *path, u64 offset,
1152                   int entries, int bitmaps)
1153 {
1154         struct btrfs_key key;
1155         struct btrfs_free_space_header *header;
1156         struct extent_buffer *leaf;
1157         int ret;
1158
1159         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
1160         key.offset = offset;
1161         key.type = 0;
1162
1163         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
1164         if (ret < 0) {
1165                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1166                                  EXTENT_DELALLOC, NULL);
1167                 goto fail;
1168         }
1169         leaf = path->nodes[0];
1170         if (ret > 0) {
1171                 struct btrfs_key found_key;
1172                 ASSERT(path->slots[0]);
1173                 path->slots[0]--;
1174                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1175                 if (found_key.objectid != BTRFS_FREE_SPACE_OBJECTID ||
1176                     found_key.offset != offset) {
1177                         clear_extent_bit(&BTRFS_I(inode)->io_tree, 0,
1178                                          inode->i_size - 1, EXTENT_DELALLOC,
1179                                          NULL);
1180                         btrfs_release_path(path);
1181                         goto fail;
1182                 }
1183         }
1184
1185         BTRFS_I(inode)->generation = trans->transid;
1186         header = btrfs_item_ptr(leaf, path->slots[0],
1187                                 struct btrfs_free_space_header);
1188         btrfs_set_free_space_entries(leaf, header, entries);
1189         btrfs_set_free_space_bitmaps(leaf, header, bitmaps);
1190         btrfs_set_free_space_generation(leaf, header, trans->transid);
1191         btrfs_mark_buffer_dirty(trans, leaf);
1192         btrfs_release_path(path);
1193
1194         return 0;
1195
1196 fail:
1197         return -1;
1198 }
1199
1200 static noinline_for_stack int write_pinned_extent_entries(
1201                             struct btrfs_trans_handle *trans,
1202                             struct btrfs_block_group *block_group,
1203                             struct btrfs_io_ctl *io_ctl,
1204                             int *entries)
1205 {
1206         u64 start, extent_start, extent_end, len;
1207         struct extent_io_tree *unpin = NULL;
1208         int ret;
1209
1210         if (!block_group)
1211                 return 0;
1212
1213         /*
1214          * We want to add any pinned extents to our free space cache
1215          * so we don't leak the space
1216          *
1217          * We shouldn't have switched the pinned extents yet so this is the
1218          * right one
1219          */
1220         unpin = &trans->transaction->pinned_extents;
1221
1222         start = block_group->start;
1223
1224         while (start < block_group->start + block_group->length) {
1225                 if (!find_first_extent_bit(unpin, start,
1226                                            &extent_start, &extent_end,
1227                                            EXTENT_DIRTY, NULL))
1228                         return 0;
1229
1230                 /* This pinned extent is out of our range */
1231                 if (extent_start >= block_group->start + block_group->length)
1232                         return 0;
1233
1234                 extent_start = max(extent_start, start);
1235                 extent_end = min(block_group->start + block_group->length,
1236                                  extent_end + 1);
1237                 len = extent_end - extent_start;
1238
1239                 *entries += 1;
1240                 ret = io_ctl_add_entry(io_ctl, extent_start, len, NULL);
1241                 if (ret)
1242                         return -ENOSPC;
1243
1244                 start = extent_end;
1245         }
1246
1247         return 0;
1248 }
1249
1250 static noinline_for_stack int
1251 write_bitmap_entries(struct btrfs_io_ctl *io_ctl, struct list_head *bitmap_list)
1252 {
1253         struct btrfs_free_space *entry, *next;
1254         int ret;
1255
1256         /* Write out the bitmaps */
1257         list_for_each_entry_safe(entry, next, bitmap_list, list) {
1258                 ret = io_ctl_add_bitmap(io_ctl, entry->bitmap);
1259                 if (ret)
1260                         return -ENOSPC;
1261                 list_del_init(&entry->list);
1262         }
1263
1264         return 0;
1265 }
1266
1267 static int flush_dirty_cache(struct inode *inode)
1268 {
1269         int ret;
1270
1271         ret = btrfs_wait_ordered_range(inode, 0, (u64)-1);
1272         if (ret)
1273                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1274                                  EXTENT_DELALLOC, NULL);
1275
1276         return ret;
1277 }
1278
1279 static void noinline_for_stack
1280 cleanup_bitmap_list(struct list_head *bitmap_list)
1281 {
1282         struct btrfs_free_space *entry, *next;
1283
1284         list_for_each_entry_safe(entry, next, bitmap_list, list)
1285                 list_del_init(&entry->list);
1286 }
1287
1288 static void noinline_for_stack
1289 cleanup_write_cache_enospc(struct inode *inode,
1290                            struct btrfs_io_ctl *io_ctl,
1291                            struct extent_state **cached_state)
1292 {
1293         io_ctl_drop_pages(io_ctl);
1294         unlock_extent(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
1295                       cached_state);
1296 }
1297
1298 static int __btrfs_wait_cache_io(struct btrfs_root *root,
1299                                  struct btrfs_trans_handle *trans,
1300                                  struct btrfs_block_group *block_group,
1301                                  struct btrfs_io_ctl *io_ctl,
1302                                  struct btrfs_path *path, u64 offset)
1303 {
1304         int ret;
1305         struct inode *inode = io_ctl->inode;
1306
1307         if (!inode)
1308                 return 0;
1309
1310         /* Flush the dirty pages in the cache file. */
1311         ret = flush_dirty_cache(inode);
1312         if (ret)
1313                 goto out;
1314
1315         /* Update the cache item to tell everyone this cache file is valid. */
1316         ret = update_cache_item(trans, root, inode, path, offset,
1317                                 io_ctl->entries, io_ctl->bitmaps);
1318 out:
1319         if (ret) {
1320                 invalidate_inode_pages2(inode->i_mapping);
1321                 BTRFS_I(inode)->generation = 0;
1322                 if (block_group)
1323                         btrfs_debug(root->fs_info,
1324           "failed to write free space cache for block group %llu error %d",
1325                                   block_group->start, ret);
1326         }
1327         btrfs_update_inode(trans, BTRFS_I(inode));
1328
1329         if (block_group) {
1330                 /* the dirty list is protected by the dirty_bgs_lock */
1331                 spin_lock(&trans->transaction->dirty_bgs_lock);
1332
1333                 /* the disk_cache_state is protected by the block group lock */
1334                 spin_lock(&block_group->lock);
1335
1336                 /*
1337                  * only mark this as written if we didn't get put back on
1338                  * the dirty list while waiting for IO.   Otherwise our
1339                  * cache state won't be right, and we won't get written again
1340                  */
1341                 if (!ret && list_empty(&block_group->dirty_list))
1342                         block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1343                 else if (ret)
1344                         block_group->disk_cache_state = BTRFS_DC_ERROR;
1345
1346                 spin_unlock(&block_group->lock);
1347                 spin_unlock(&trans->transaction->dirty_bgs_lock);
1348                 io_ctl->inode = NULL;
1349                 iput(inode);
1350         }
1351
1352         return ret;
1353
1354 }
1355
1356 int btrfs_wait_cache_io(struct btrfs_trans_handle *trans,
1357                         struct btrfs_block_group *block_group,
1358                         struct btrfs_path *path)
1359 {
1360         return __btrfs_wait_cache_io(block_group->fs_info->tree_root, trans,
1361                                      block_group, &block_group->io_ctl,
1362                                      path, block_group->start);
1363 }
1364
1365 /*
1366  * Write out cached info to an inode.
1367  *
1368  * @inode:       freespace inode we are writing out
1369  * @ctl:         free space cache we are going to write out
1370  * @block_group: block_group for this cache if it belongs to a block_group
1371  * @io_ctl:      holds context for the io
1372  * @trans:       the trans handle
1373  *
1374  * This function writes out a free space cache struct to disk for quick recovery
1375  * on mount.  This will return 0 if it was successful in writing the cache out,
1376  * or an errno if it was not.
1377  */
1378 static int __btrfs_write_out_cache(struct inode *inode,
1379                                    struct btrfs_free_space_ctl *ctl,
1380                                    struct btrfs_block_group *block_group,
1381                                    struct btrfs_io_ctl *io_ctl,
1382                                    struct btrfs_trans_handle *trans)
1383 {
1384         struct extent_state *cached_state = NULL;
1385         LIST_HEAD(bitmap_list);
1386         int entries = 0;
1387         int bitmaps = 0;
1388         int ret;
1389         int must_iput = 0;
1390
1391         if (!i_size_read(inode))
1392                 return -EIO;
1393
1394         WARN_ON(io_ctl->pages);
1395         ret = io_ctl_init(io_ctl, inode, 1);
1396         if (ret)
1397                 return ret;
1398
1399         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA)) {
1400                 down_write(&block_group->data_rwsem);
1401                 spin_lock(&block_group->lock);
1402                 if (block_group->delalloc_bytes) {
1403                         block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1404                         spin_unlock(&block_group->lock);
1405                         up_write(&block_group->data_rwsem);
1406                         BTRFS_I(inode)->generation = 0;
1407                         ret = 0;
1408                         must_iput = 1;
1409                         goto out;
1410                 }
1411                 spin_unlock(&block_group->lock);
1412         }
1413
1414         /* Lock all pages first so we can lock the extent safely. */
1415         ret = io_ctl_prepare_pages(io_ctl, false);
1416         if (ret)
1417                 goto out_unlock;
1418
1419         lock_extent(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
1420                     &cached_state);
1421
1422         io_ctl_set_generation(io_ctl, trans->transid);
1423
1424         mutex_lock(&ctl->cache_writeout_mutex);
1425         /* Write out the extent entries in the free space cache */
1426         spin_lock(&ctl->tree_lock);
1427         ret = write_cache_extent_entries(io_ctl, ctl,
1428                                          block_group, &entries, &bitmaps,
1429                                          &bitmap_list);
1430         if (ret)
1431                 goto out_nospc_locked;
1432
1433         /*
1434          * Some spaces that are freed in the current transaction are pinned,
1435          * they will be added into free space cache after the transaction is
1436          * committed, we shouldn't lose them.
1437          *
1438          * If this changes while we are working we'll get added back to
1439          * the dirty list and redo it.  No locking needed
1440          */
1441         ret = write_pinned_extent_entries(trans, block_group, io_ctl, &entries);
1442         if (ret)
1443                 goto out_nospc_locked;
1444
1445         /*
1446          * At last, we write out all the bitmaps and keep cache_writeout_mutex
1447          * locked while doing it because a concurrent trim can be manipulating
1448          * or freeing the bitmap.
1449          */
1450         ret = write_bitmap_entries(io_ctl, &bitmap_list);
1451         spin_unlock(&ctl->tree_lock);
1452         mutex_unlock(&ctl->cache_writeout_mutex);
1453         if (ret)
1454                 goto out_nospc;
1455
1456         /* Zero out the rest of the pages just to make sure */
1457         io_ctl_zero_remaining_pages(io_ctl);
1458
1459         /* Everything is written out, now we dirty the pages in the file. */
1460         ret = btrfs_dirty_pages(BTRFS_I(inode), io_ctl->pages,
1461                                 io_ctl->num_pages, 0, i_size_read(inode),
1462                                 &cached_state, false);
1463         if (ret)
1464                 goto out_nospc;
1465
1466         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1467                 up_write(&block_group->data_rwsem);
1468         /*
1469          * Release the pages and unlock the extent, we will flush
1470          * them out later
1471          */
1472         io_ctl_drop_pages(io_ctl);
1473         io_ctl_free(io_ctl);
1474
1475         unlock_extent(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
1476                       &cached_state);
1477
1478         /*
1479          * at this point the pages are under IO and we're happy,
1480          * The caller is responsible for waiting on them and updating
1481          * the cache and the inode
1482          */
1483         io_ctl->entries = entries;
1484         io_ctl->bitmaps = bitmaps;
1485
1486         ret = btrfs_fdatawrite_range(inode, 0, (u64)-1);
1487         if (ret)
1488                 goto out;
1489
1490         return 0;
1491
1492 out_nospc_locked:
1493         cleanup_bitmap_list(&bitmap_list);
1494         spin_unlock(&ctl->tree_lock);
1495         mutex_unlock(&ctl->cache_writeout_mutex);
1496
1497 out_nospc:
1498         cleanup_write_cache_enospc(inode, io_ctl, &cached_state);
1499
1500 out_unlock:
1501         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1502                 up_write(&block_group->data_rwsem);
1503
1504 out:
1505         io_ctl->inode = NULL;
1506         io_ctl_free(io_ctl);
1507         if (ret) {
1508                 invalidate_inode_pages2(inode->i_mapping);
1509                 BTRFS_I(inode)->generation = 0;
1510         }
1511         btrfs_update_inode(trans, BTRFS_I(inode));
1512         if (must_iput)
1513                 iput(inode);
1514         return ret;
1515 }
1516
1517 int btrfs_write_out_cache(struct btrfs_trans_handle *trans,
1518                           struct btrfs_block_group *block_group,
1519                           struct btrfs_path *path)
1520 {
1521         struct btrfs_fs_info *fs_info = trans->fs_info;
1522         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1523         struct inode *inode;
1524         int ret = 0;
1525
1526         spin_lock(&block_group->lock);
1527         if (block_group->disk_cache_state < BTRFS_DC_SETUP) {
1528                 spin_unlock(&block_group->lock);
1529                 return 0;
1530         }
1531         spin_unlock(&block_group->lock);
1532
1533         inode = lookup_free_space_inode(block_group, path);
1534         if (IS_ERR(inode))
1535                 return 0;
1536
1537         ret = __btrfs_write_out_cache(inode, ctl, block_group,
1538                                       &block_group->io_ctl, trans);
1539         if (ret) {
1540                 btrfs_debug(fs_info,
1541           "failed to write free space cache for block group %llu error %d",
1542                           block_group->start, ret);
1543                 spin_lock(&block_group->lock);
1544                 block_group->disk_cache_state = BTRFS_DC_ERROR;
1545                 spin_unlock(&block_group->lock);
1546
1547                 block_group->io_ctl.inode = NULL;
1548                 iput(inode);
1549         }
1550
1551         /*
1552          * if ret == 0 the caller is expected to call btrfs_wait_cache_io
1553          * to wait for IO and put the inode
1554          */
1555
1556         return ret;
1557 }
1558
1559 static inline unsigned long offset_to_bit(u64 bitmap_start, u32 unit,
1560                                           u64 offset)
1561 {
1562         ASSERT(offset >= bitmap_start);
1563         offset -= bitmap_start;
1564         return (unsigned long)(div_u64(offset, unit));
1565 }
1566
1567 static inline unsigned long bytes_to_bits(u64 bytes, u32 unit)
1568 {
1569         return (unsigned long)(div_u64(bytes, unit));
1570 }
1571
1572 static inline u64 offset_to_bitmap(struct btrfs_free_space_ctl *ctl,
1573                                    u64 offset)
1574 {
1575         u64 bitmap_start;
1576         u64 bytes_per_bitmap;
1577
1578         bytes_per_bitmap = BITS_PER_BITMAP * ctl->unit;
1579         bitmap_start = offset - ctl->start;
1580         bitmap_start = div64_u64(bitmap_start, bytes_per_bitmap);
1581         bitmap_start *= bytes_per_bitmap;
1582         bitmap_start += ctl->start;
1583
1584         return bitmap_start;
1585 }
1586
1587 static int tree_insert_offset(struct btrfs_free_space_ctl *ctl,
1588                               struct btrfs_free_cluster *cluster,
1589                               struct btrfs_free_space *new_entry)
1590 {
1591         struct rb_root *root;
1592         struct rb_node **p;
1593         struct rb_node *parent = NULL;
1594
1595         lockdep_assert_held(&ctl->tree_lock);
1596
1597         if (cluster) {
1598                 lockdep_assert_held(&cluster->lock);
1599                 root = &cluster->root;
1600         } else {
1601                 root = &ctl->free_space_offset;
1602         }
1603
1604         p = &root->rb_node;
1605
1606         while (*p) {
1607                 struct btrfs_free_space *info;
1608
1609                 parent = *p;
1610                 info = rb_entry(parent, struct btrfs_free_space, offset_index);
1611
1612                 if (new_entry->offset < info->offset) {
1613                         p = &(*p)->rb_left;
1614                 } else if (new_entry->offset > info->offset) {
1615                         p = &(*p)->rb_right;
1616                 } else {
1617                         /*
1618                          * we could have a bitmap entry and an extent entry
1619                          * share the same offset.  If this is the case, we want
1620                          * the extent entry to always be found first if we do a
1621                          * linear search through the tree, since we want to have
1622                          * the quickest allocation time, and allocating from an
1623                          * extent is faster than allocating from a bitmap.  So
1624                          * if we're inserting a bitmap and we find an entry at
1625                          * this offset, we want to go right, or after this entry
1626                          * logically.  If we are inserting an extent and we've
1627                          * found a bitmap, we want to go left, or before
1628                          * logically.
1629                          */
1630                         if (new_entry->bitmap) {
1631                                 if (info->bitmap) {
1632                                         WARN_ON_ONCE(1);
1633                                         return -EEXIST;
1634                                 }
1635                                 p = &(*p)->rb_right;
1636                         } else {
1637                                 if (!info->bitmap) {
1638                                         WARN_ON_ONCE(1);
1639                                         return -EEXIST;
1640                                 }
1641                                 p = &(*p)->rb_left;
1642                         }
1643                 }
1644         }
1645
1646         rb_link_node(&new_entry->offset_index, parent, p);
1647         rb_insert_color(&new_entry->offset_index, root);
1648
1649         return 0;
1650 }
1651
1652 /*
1653  * This is a little subtle.  We *only* have ->max_extent_size set if we actually
1654  * searched through the bitmap and figured out the largest ->max_extent_size,
1655  * otherwise it's 0.  In the case that it's 0 we don't want to tell the
1656  * allocator the wrong thing, we want to use the actual real max_extent_size
1657  * we've found already if it's larger, or we want to use ->bytes.
1658  *
1659  * This matters because find_free_space() will skip entries who's ->bytes is
1660  * less than the required bytes.  So if we didn't search down this bitmap, we
1661  * may pick some previous entry that has a smaller ->max_extent_size than we
1662  * have.  For example, assume we have two entries, one that has
1663  * ->max_extent_size set to 4K and ->bytes set to 1M.  A second entry hasn't set
1664  * ->max_extent_size yet, has ->bytes set to 8K and it's contiguous.  We will
1665  *  call into find_free_space(), and return with max_extent_size == 4K, because
1666  *  that first bitmap entry had ->max_extent_size set, but the second one did
1667  *  not.  If instead we returned 8K we'd come in searching for 8K, and find the
1668  *  8K contiguous range.
1669  *
1670  *  Consider the other case, we have 2 8K chunks in that second entry and still
1671  *  don't have ->max_extent_size set.  We'll return 16K, and the next time the
1672  *  allocator comes in it'll fully search our second bitmap, and this time it'll
1673  *  get an uptodate value of 8K as the maximum chunk size.  Then we'll get the
1674  *  right allocation the next loop through.
1675  */
1676 static inline u64 get_max_extent_size(const struct btrfs_free_space *entry)
1677 {
1678         if (entry->bitmap && entry->max_extent_size)
1679                 return entry->max_extent_size;
1680         return entry->bytes;
1681 }
1682
1683 /*
1684  * We want the largest entry to be leftmost, so this is inverted from what you'd
1685  * normally expect.
1686  */
1687 static bool entry_less(struct rb_node *node, const struct rb_node *parent)
1688 {
1689         const struct btrfs_free_space *entry, *exist;
1690
1691         entry = rb_entry(node, struct btrfs_free_space, bytes_index);
1692         exist = rb_entry(parent, struct btrfs_free_space, bytes_index);
1693         return get_max_extent_size(exist) < get_max_extent_size(entry);
1694 }
1695
1696 /*
1697  * searches the tree for the given offset.
1698  *
1699  * fuzzy - If this is set, then we are trying to make an allocation, and we just
1700  * want a section that has at least bytes size and comes at or after the given
1701  * offset.
1702  */
1703 static struct btrfs_free_space *
1704 tree_search_offset(struct btrfs_free_space_ctl *ctl,
1705                    u64 offset, int bitmap_only, int fuzzy)
1706 {
1707         struct rb_node *n = ctl->free_space_offset.rb_node;
1708         struct btrfs_free_space *entry = NULL, *prev = NULL;
1709
1710         lockdep_assert_held(&ctl->tree_lock);
1711
1712         /* find entry that is closest to the 'offset' */
1713         while (n) {
1714                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1715                 prev = entry;
1716
1717                 if (offset < entry->offset)
1718                         n = n->rb_left;
1719                 else if (offset > entry->offset)
1720                         n = n->rb_right;
1721                 else
1722                         break;
1723
1724                 entry = NULL;
1725         }
1726
1727         if (bitmap_only) {
1728                 if (!entry)
1729                         return NULL;
1730                 if (entry->bitmap)
1731                         return entry;
1732
1733                 /*
1734                  * bitmap entry and extent entry may share same offset,
1735                  * in that case, bitmap entry comes after extent entry.
1736                  */
1737                 n = rb_next(n);
1738                 if (!n)
1739                         return NULL;
1740                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1741                 if (entry->offset != offset)
1742                         return NULL;
1743
1744                 WARN_ON(!entry->bitmap);
1745                 return entry;
1746         } else if (entry) {
1747                 if (entry->bitmap) {
1748                         /*
1749                          * if previous extent entry covers the offset,
1750                          * we should return it instead of the bitmap entry
1751                          */
1752                         n = rb_prev(&entry->offset_index);
1753                         if (n) {
1754                                 prev = rb_entry(n, struct btrfs_free_space,
1755                                                 offset_index);
1756                                 if (!prev->bitmap &&
1757                                     prev->offset + prev->bytes > offset)
1758                                         entry = prev;
1759                         }
1760                 }
1761                 return entry;
1762         }
1763
1764         if (!prev)
1765                 return NULL;
1766
1767         /* find last entry before the 'offset' */
1768         entry = prev;
1769         if (entry->offset > offset) {
1770                 n = rb_prev(&entry->offset_index);
1771                 if (n) {
1772                         entry = rb_entry(n, struct btrfs_free_space,
1773                                         offset_index);
1774                         ASSERT(entry->offset <= offset);
1775                 } else {
1776                         if (fuzzy)
1777                                 return entry;
1778                         else
1779                                 return NULL;
1780                 }
1781         }
1782
1783         if (entry->bitmap) {
1784                 n = rb_prev(&entry->offset_index);
1785                 if (n) {
1786                         prev = rb_entry(n, struct btrfs_free_space,
1787                                         offset_index);
1788                         if (!prev->bitmap &&
1789                             prev->offset + prev->bytes > offset)
1790                                 return prev;
1791                 }
1792                 if (entry->offset + BITS_PER_BITMAP * ctl->unit > offset)
1793                         return entry;
1794         } else if (entry->offset + entry->bytes > offset)
1795                 return entry;
1796
1797         if (!fuzzy)
1798                 return NULL;
1799
1800         while (1) {
1801                 n = rb_next(&entry->offset_index);
1802                 if (!n)
1803                         return NULL;
1804                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1805                 if (entry->bitmap) {
1806                         if (entry->offset + BITS_PER_BITMAP *
1807                             ctl->unit > offset)
1808                                 break;
1809                 } else {
1810                         if (entry->offset + entry->bytes > offset)
1811                                 break;
1812                 }
1813         }
1814         return entry;
1815 }
1816
1817 static inline void unlink_free_space(struct btrfs_free_space_ctl *ctl,
1818                                      struct btrfs_free_space *info,
1819                                      bool update_stat)
1820 {
1821         lockdep_assert_held(&ctl->tree_lock);
1822
1823         rb_erase(&info->offset_index, &ctl->free_space_offset);
1824         rb_erase_cached(&info->bytes_index, &ctl->free_space_bytes);
1825         ctl->free_extents--;
1826
1827         if (!info->bitmap && !btrfs_free_space_trimmed(info)) {
1828                 ctl->discardable_extents[BTRFS_STAT_CURR]--;
1829                 ctl->discardable_bytes[BTRFS_STAT_CURR] -= info->bytes;
1830         }
1831
1832         if (update_stat)
1833                 ctl->free_space -= info->bytes;
1834 }
1835
1836 static int link_free_space(struct btrfs_free_space_ctl *ctl,
1837                            struct btrfs_free_space *info)
1838 {
1839         int ret = 0;
1840
1841         lockdep_assert_held(&ctl->tree_lock);
1842
1843         ASSERT(info->bytes || info->bitmap);
1844         ret = tree_insert_offset(ctl, NULL, info);
1845         if (ret)
1846                 return ret;
1847
1848         rb_add_cached(&info->bytes_index, &ctl->free_space_bytes, entry_less);
1849
1850         if (!info->bitmap && !btrfs_free_space_trimmed(info)) {
1851                 ctl->discardable_extents[BTRFS_STAT_CURR]++;
1852                 ctl->discardable_bytes[BTRFS_STAT_CURR] += info->bytes;
1853         }
1854
1855         ctl->free_space += info->bytes;
1856         ctl->free_extents++;
1857         return ret;
1858 }
1859
1860 static void relink_bitmap_entry(struct btrfs_free_space_ctl *ctl,
1861                                 struct btrfs_free_space *info)
1862 {
1863         ASSERT(info->bitmap);
1864
1865         /*
1866          * If our entry is empty it's because we're on a cluster and we don't
1867          * want to re-link it into our ctl bytes index.
1868          */
1869         if (RB_EMPTY_NODE(&info->bytes_index))
1870                 return;
1871
1872         lockdep_assert_held(&ctl->tree_lock);
1873
1874         rb_erase_cached(&info->bytes_index, &ctl->free_space_bytes);
1875         rb_add_cached(&info->bytes_index, &ctl->free_space_bytes, entry_less);
1876 }
1877
1878 static inline void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1879                                      struct btrfs_free_space *info,
1880                                      u64 offset, u64 bytes, bool update_stat)
1881 {
1882         unsigned long start, count, end;
1883         int extent_delta = -1;
1884
1885         start = offset_to_bit(info->offset, ctl->unit, offset);
1886         count = bytes_to_bits(bytes, ctl->unit);
1887         end = start + count;
1888         ASSERT(end <= BITS_PER_BITMAP);
1889
1890         bitmap_clear(info->bitmap, start, count);
1891
1892         info->bytes -= bytes;
1893         if (info->max_extent_size > ctl->unit)
1894                 info->max_extent_size = 0;
1895
1896         relink_bitmap_entry(ctl, info);
1897
1898         if (start && test_bit(start - 1, info->bitmap))
1899                 extent_delta++;
1900
1901         if (end < BITS_PER_BITMAP && test_bit(end, info->bitmap))
1902                 extent_delta++;
1903
1904         info->bitmap_extents += extent_delta;
1905         if (!btrfs_free_space_trimmed(info)) {
1906                 ctl->discardable_extents[BTRFS_STAT_CURR] += extent_delta;
1907                 ctl->discardable_bytes[BTRFS_STAT_CURR] -= bytes;
1908         }
1909
1910         if (update_stat)
1911                 ctl->free_space -= bytes;
1912 }
1913
1914 static void btrfs_bitmap_set_bits(struct btrfs_free_space_ctl *ctl,
1915                                   struct btrfs_free_space *info, u64 offset,
1916                                   u64 bytes)
1917 {
1918         unsigned long start, count, end;
1919         int extent_delta = 1;
1920
1921         start = offset_to_bit(info->offset, ctl->unit, offset);
1922         count = bytes_to_bits(bytes, ctl->unit);
1923         end = start + count;
1924         ASSERT(end <= BITS_PER_BITMAP);
1925
1926         bitmap_set(info->bitmap, start, count);
1927
1928         /*
1929          * We set some bytes, we have no idea what the max extent size is
1930          * anymore.
1931          */
1932         info->max_extent_size = 0;
1933         info->bytes += bytes;
1934         ctl->free_space += bytes;
1935
1936         relink_bitmap_entry(ctl, info);
1937
1938         if (start && test_bit(start - 1, info->bitmap))
1939                 extent_delta--;
1940
1941         if (end < BITS_PER_BITMAP && test_bit(end, info->bitmap))
1942                 extent_delta--;
1943
1944         info->bitmap_extents += extent_delta;
1945         if (!btrfs_free_space_trimmed(info)) {
1946                 ctl->discardable_extents[BTRFS_STAT_CURR] += extent_delta;
1947                 ctl->discardable_bytes[BTRFS_STAT_CURR] += bytes;
1948         }
1949 }
1950
1951 /*
1952  * If we can not find suitable extent, we will use bytes to record
1953  * the size of the max extent.
1954  */
1955 static int search_bitmap(struct btrfs_free_space_ctl *ctl,
1956                          struct btrfs_free_space *bitmap_info, u64 *offset,
1957                          u64 *bytes, bool for_alloc)
1958 {
1959         unsigned long found_bits = 0;
1960         unsigned long max_bits = 0;
1961         unsigned long bits, i;
1962         unsigned long next_zero;
1963         unsigned long extent_bits;
1964
1965         /*
1966          * Skip searching the bitmap if we don't have a contiguous section that
1967          * is large enough for this allocation.
1968          */
1969         if (for_alloc &&
1970             bitmap_info->max_extent_size &&
1971             bitmap_info->max_extent_size < *bytes) {
1972                 *bytes = bitmap_info->max_extent_size;
1973                 return -1;
1974         }
1975
1976         i = offset_to_bit(bitmap_info->offset, ctl->unit,
1977                           max_t(u64, *offset, bitmap_info->offset));
1978         bits = bytes_to_bits(*bytes, ctl->unit);
1979
1980         for_each_set_bit_from(i, bitmap_info->bitmap, BITS_PER_BITMAP) {
1981                 if (for_alloc && bits == 1) {
1982                         found_bits = 1;
1983                         break;
1984                 }
1985                 next_zero = find_next_zero_bit(bitmap_info->bitmap,
1986                                                BITS_PER_BITMAP, i);
1987                 extent_bits = next_zero - i;
1988                 if (extent_bits >= bits) {
1989                         found_bits = extent_bits;
1990                         break;
1991                 } else if (extent_bits > max_bits) {
1992                         max_bits = extent_bits;
1993                 }
1994                 i = next_zero;
1995         }
1996
1997         if (found_bits) {
1998                 *offset = (u64)(i * ctl->unit) + bitmap_info->offset;
1999                 *bytes = (u64)(found_bits) * ctl->unit;
2000                 return 0;
2001         }
2002
2003         *bytes = (u64)(max_bits) * ctl->unit;
2004         bitmap_info->max_extent_size = *bytes;
2005         relink_bitmap_entry(ctl, bitmap_info);
2006         return -1;
2007 }
2008
2009 /* Cache the size of the max extent in bytes */
2010 static struct btrfs_free_space *
2011 find_free_space(struct btrfs_free_space_ctl *ctl, u64 *offset, u64 *bytes,
2012                 unsigned long align, u64 *max_extent_size, bool use_bytes_index)
2013 {
2014         struct btrfs_free_space *entry;
2015         struct rb_node *node;
2016         u64 tmp;
2017         u64 align_off;
2018         int ret;
2019
2020         if (!ctl->free_space_offset.rb_node)
2021                 goto out;
2022 again:
2023         if (use_bytes_index) {
2024                 node = rb_first_cached(&ctl->free_space_bytes);
2025         } else {
2026                 entry = tree_search_offset(ctl, offset_to_bitmap(ctl, *offset),
2027                                            0, 1);
2028                 if (!entry)
2029                         goto out;
2030                 node = &entry->offset_index;
2031         }
2032
2033         for (; node; node = rb_next(node)) {
2034                 if (use_bytes_index)
2035                         entry = rb_entry(node, struct btrfs_free_space,
2036                                          bytes_index);
2037                 else
2038                         entry = rb_entry(node, struct btrfs_free_space,
2039                                          offset_index);
2040
2041                 /*
2042                  * If we are using the bytes index then all subsequent entries
2043                  * in this tree are going to be < bytes, so simply set the max
2044                  * extent size and exit the loop.
2045                  *
2046                  * If we're using the offset index then we need to keep going
2047                  * through the rest of the tree.
2048                  */
2049                 if (entry->bytes < *bytes) {
2050                         *max_extent_size = max(get_max_extent_size(entry),
2051                                                *max_extent_size);
2052                         if (use_bytes_index)
2053                                 break;
2054                         continue;
2055                 }
2056
2057                 /* make sure the space returned is big enough
2058                  * to match our requested alignment
2059                  */
2060                 if (*bytes >= align) {
2061                         tmp = entry->offset - ctl->start + align - 1;
2062                         tmp = div64_u64(tmp, align);
2063                         tmp = tmp * align + ctl->start;
2064                         align_off = tmp - entry->offset;
2065                 } else {
2066                         align_off = 0;
2067                         tmp = entry->offset;
2068                 }
2069
2070                 /*
2071                  * We don't break here if we're using the bytes index because we
2072                  * may have another entry that has the correct alignment that is
2073                  * the right size, so we don't want to miss that possibility.
2074                  * At worst this adds another loop through the logic, but if we
2075                  * broke here we could prematurely ENOSPC.
2076                  */
2077                 if (entry->bytes < *bytes + align_off) {
2078                         *max_extent_size = max(get_max_extent_size(entry),
2079                                                *max_extent_size);
2080                         continue;
2081                 }
2082
2083                 if (entry->bitmap) {
2084                         struct rb_node *old_next = rb_next(node);
2085                         u64 size = *bytes;
2086
2087                         ret = search_bitmap(ctl, entry, &tmp, &size, true);
2088                         if (!ret) {
2089                                 *offset = tmp;
2090                                 *bytes = size;
2091                                 return entry;
2092                         } else {
2093                                 *max_extent_size =
2094                                         max(get_max_extent_size(entry),
2095                                             *max_extent_size);
2096                         }
2097
2098                         /*
2099                          * The bitmap may have gotten re-arranged in the space
2100                          * index here because the max_extent_size may have been
2101                          * updated.  Start from the beginning again if this
2102                          * happened.
2103                          */
2104                         if (use_bytes_index && old_next != rb_next(node))
2105                                 goto again;
2106                         continue;
2107                 }
2108
2109                 *offset = tmp;
2110                 *bytes = entry->bytes - align_off;
2111                 return entry;
2112         }
2113 out:
2114         return NULL;
2115 }
2116
2117 static void add_new_bitmap(struct btrfs_free_space_ctl *ctl,
2118                            struct btrfs_free_space *info, u64 offset)
2119 {
2120         info->offset = offset_to_bitmap(ctl, offset);
2121         info->bytes = 0;
2122         info->bitmap_extents = 0;
2123         INIT_LIST_HEAD(&info->list);
2124         link_free_space(ctl, info);
2125         ctl->total_bitmaps++;
2126         recalculate_thresholds(ctl);
2127 }
2128
2129 static void free_bitmap(struct btrfs_free_space_ctl *ctl,
2130                         struct btrfs_free_space *bitmap_info)
2131 {
2132         /*
2133          * Normally when this is called, the bitmap is completely empty. However,
2134          * if we are blowing up the free space cache for one reason or another
2135          * via __btrfs_remove_free_space_cache(), then it may not be freed and
2136          * we may leave stats on the table.
2137          */
2138         if (bitmap_info->bytes && !btrfs_free_space_trimmed(bitmap_info)) {
2139                 ctl->discardable_extents[BTRFS_STAT_CURR] -=
2140                         bitmap_info->bitmap_extents;
2141                 ctl->discardable_bytes[BTRFS_STAT_CURR] -= bitmap_info->bytes;
2142
2143         }
2144         unlink_free_space(ctl, bitmap_info, true);
2145         kmem_cache_free(btrfs_free_space_bitmap_cachep, bitmap_info->bitmap);
2146         kmem_cache_free(btrfs_free_space_cachep, bitmap_info);
2147         ctl->total_bitmaps--;
2148         recalculate_thresholds(ctl);
2149 }
2150
2151 static noinline int remove_from_bitmap(struct btrfs_free_space_ctl *ctl,
2152                               struct btrfs_free_space *bitmap_info,
2153                               u64 *offset, u64 *bytes)
2154 {
2155         u64 end;
2156         u64 search_start, search_bytes;
2157         int ret;
2158
2159 again:
2160         end = bitmap_info->offset + (u64)(BITS_PER_BITMAP * ctl->unit) - 1;
2161
2162         /*
2163          * We need to search for bits in this bitmap.  We could only cover some
2164          * of the extent in this bitmap thanks to how we add space, so we need
2165          * to search for as much as it as we can and clear that amount, and then
2166          * go searching for the next bit.
2167          */
2168         search_start = *offset;
2169         search_bytes = ctl->unit;
2170         search_bytes = min(search_bytes, end - search_start + 1);
2171         ret = search_bitmap(ctl, bitmap_info, &search_start, &search_bytes,
2172                             false);
2173         if (ret < 0 || search_start != *offset)
2174                 return -EINVAL;
2175
2176         /* We may have found more bits than what we need */
2177         search_bytes = min(search_bytes, *bytes);
2178
2179         /* Cannot clear past the end of the bitmap */
2180         search_bytes = min(search_bytes, end - search_start + 1);
2181
2182         bitmap_clear_bits(ctl, bitmap_info, search_start, search_bytes, true);
2183         *offset += search_bytes;
2184         *bytes -= search_bytes;
2185
2186         if (*bytes) {
2187                 struct rb_node *next = rb_next(&bitmap_info->offset_index);
2188                 if (!bitmap_info->bytes)
2189                         free_bitmap(ctl, bitmap_info);
2190
2191                 /*
2192                  * no entry after this bitmap, but we still have bytes to
2193                  * remove, so something has gone wrong.
2194                  */
2195                 if (!next)
2196                         return -EINVAL;
2197
2198                 bitmap_info = rb_entry(next, struct btrfs_free_space,
2199                                        offset_index);
2200
2201                 /*
2202                  * if the next entry isn't a bitmap we need to return to let the
2203                  * extent stuff do its work.
2204                  */
2205                 if (!bitmap_info->bitmap)
2206                         return -EAGAIN;
2207
2208                 /*
2209                  * Ok the next item is a bitmap, but it may not actually hold
2210                  * the information for the rest of this free space stuff, so
2211                  * look for it, and if we don't find it return so we can try
2212                  * everything over again.
2213                  */
2214                 search_start = *offset;
2215                 search_bytes = ctl->unit;
2216                 ret = search_bitmap(ctl, bitmap_info, &search_start,
2217                                     &search_bytes, false);
2218                 if (ret < 0 || search_start != *offset)
2219                         return -EAGAIN;
2220
2221                 goto again;
2222         } else if (!bitmap_info->bytes)
2223                 free_bitmap(ctl, bitmap_info);
2224
2225         return 0;
2226 }
2227
2228 static u64 add_bytes_to_bitmap(struct btrfs_free_space_ctl *ctl,
2229                                struct btrfs_free_space *info, u64 offset,
2230                                u64 bytes, enum btrfs_trim_state trim_state)
2231 {
2232         u64 bytes_to_set = 0;
2233         u64 end;
2234
2235         /*
2236          * This is a tradeoff to make bitmap trim state minimal.  We mark the
2237          * whole bitmap untrimmed if at any point we add untrimmed regions.
2238          */
2239         if (trim_state == BTRFS_TRIM_STATE_UNTRIMMED) {
2240                 if (btrfs_free_space_trimmed(info)) {
2241                         ctl->discardable_extents[BTRFS_STAT_CURR] +=
2242                                 info->bitmap_extents;
2243                         ctl->discardable_bytes[BTRFS_STAT_CURR] += info->bytes;
2244                 }
2245                 info->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2246         }
2247
2248         end = info->offset + (u64)(BITS_PER_BITMAP * ctl->unit);
2249
2250         bytes_to_set = min(end - offset, bytes);
2251
2252         btrfs_bitmap_set_bits(ctl, info, offset, bytes_to_set);
2253
2254         return bytes_to_set;
2255
2256 }
2257
2258 static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
2259                       struct btrfs_free_space *info)
2260 {
2261         struct btrfs_block_group *block_group = ctl->block_group;
2262         struct btrfs_fs_info *fs_info = block_group->fs_info;
2263         bool forced = false;
2264
2265 #ifdef CONFIG_BTRFS_DEBUG
2266         if (btrfs_should_fragment_free_space(block_group))
2267                 forced = true;
2268 #endif
2269
2270         /* This is a way to reclaim large regions from the bitmaps. */
2271         if (!forced && info->bytes >= FORCE_EXTENT_THRESHOLD)
2272                 return false;
2273
2274         /*
2275          * If we are below the extents threshold then we can add this as an
2276          * extent, and don't have to deal with the bitmap
2277          */
2278         if (!forced && ctl->free_extents < ctl->extents_thresh) {
2279                 /*
2280                  * If this block group has some small extents we don't want to
2281                  * use up all of our free slots in the cache with them, we want
2282                  * to reserve them to larger extents, however if we have plenty
2283                  * of cache left then go ahead an dadd them, no sense in adding
2284                  * the overhead of a bitmap if we don't have to.
2285                  */
2286                 if (info->bytes <= fs_info->sectorsize * 8) {
2287                         if (ctl->free_extents * 3 <= ctl->extents_thresh)
2288                                 return false;
2289                 } else {
2290                         return false;
2291                 }
2292         }
2293
2294         /*
2295          * The original block groups from mkfs can be really small, like 8
2296          * megabytes, so don't bother with a bitmap for those entries.  However
2297          * some block groups can be smaller than what a bitmap would cover but
2298          * are still large enough that they could overflow the 32k memory limit,
2299          * so allow those block groups to still be allowed to have a bitmap
2300          * entry.
2301          */
2302         if (((BITS_PER_BITMAP * ctl->unit) >> 1) > block_group->length)
2303                 return false;
2304
2305         return true;
2306 }
2307
2308 static const struct btrfs_free_space_op free_space_op = {
2309         .use_bitmap             = use_bitmap,
2310 };
2311
2312 static int insert_into_bitmap(struct btrfs_free_space_ctl *ctl,
2313                               struct btrfs_free_space *info)
2314 {
2315         struct btrfs_free_space *bitmap_info;
2316         struct btrfs_block_group *block_group = NULL;
2317         int added = 0;
2318         u64 bytes, offset, bytes_added;
2319         enum btrfs_trim_state trim_state;
2320         int ret;
2321
2322         bytes = info->bytes;
2323         offset = info->offset;
2324         trim_state = info->trim_state;
2325
2326         if (!ctl->op->use_bitmap(ctl, info))
2327                 return 0;
2328
2329         if (ctl->op == &free_space_op)
2330                 block_group = ctl->block_group;
2331 again:
2332         /*
2333          * Since we link bitmaps right into the cluster we need to see if we
2334          * have a cluster here, and if so and it has our bitmap we need to add
2335          * the free space to that bitmap.
2336          */
2337         if (block_group && !list_empty(&block_group->cluster_list)) {
2338                 struct btrfs_free_cluster *cluster;
2339                 struct rb_node *node;
2340                 struct btrfs_free_space *entry;
2341
2342                 cluster = list_entry(block_group->cluster_list.next,
2343                                      struct btrfs_free_cluster,
2344                                      block_group_list);
2345                 spin_lock(&cluster->lock);
2346                 node = rb_first(&cluster->root);
2347                 if (!node) {
2348                         spin_unlock(&cluster->lock);
2349                         goto no_cluster_bitmap;
2350                 }
2351
2352                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2353                 if (!entry->bitmap) {
2354                         spin_unlock(&cluster->lock);
2355                         goto no_cluster_bitmap;
2356                 }
2357
2358                 if (entry->offset == offset_to_bitmap(ctl, offset)) {
2359                         bytes_added = add_bytes_to_bitmap(ctl, entry, offset,
2360                                                           bytes, trim_state);
2361                         bytes -= bytes_added;
2362                         offset += bytes_added;
2363                 }
2364                 spin_unlock(&cluster->lock);
2365                 if (!bytes) {
2366                         ret = 1;
2367                         goto out;
2368                 }
2369         }
2370
2371 no_cluster_bitmap:
2372         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2373                                          1, 0);
2374         if (!bitmap_info) {
2375                 ASSERT(added == 0);
2376                 goto new_bitmap;
2377         }
2378
2379         bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes,
2380                                           trim_state);
2381         bytes -= bytes_added;
2382         offset += bytes_added;
2383         added = 0;
2384
2385         if (!bytes) {
2386                 ret = 1;
2387                 goto out;
2388         } else
2389                 goto again;
2390
2391 new_bitmap:
2392         if (info && info->bitmap) {
2393                 add_new_bitmap(ctl, info, offset);
2394                 added = 1;
2395                 info = NULL;
2396                 goto again;
2397         } else {
2398                 spin_unlock(&ctl->tree_lock);
2399
2400                 /* no pre-allocated info, allocate a new one */
2401                 if (!info) {
2402                         info = kmem_cache_zalloc(btrfs_free_space_cachep,
2403                                                  GFP_NOFS);
2404                         if (!info) {
2405                                 spin_lock(&ctl->tree_lock);
2406                                 ret = -ENOMEM;
2407                                 goto out;
2408                         }
2409                 }
2410
2411                 /* allocate the bitmap */
2412                 info->bitmap = kmem_cache_zalloc(btrfs_free_space_bitmap_cachep,
2413                                                  GFP_NOFS);
2414                 info->trim_state = BTRFS_TRIM_STATE_TRIMMED;
2415                 spin_lock(&ctl->tree_lock);
2416                 if (!info->bitmap) {
2417                         ret = -ENOMEM;
2418                         goto out;
2419                 }
2420                 goto again;
2421         }
2422
2423 out:
2424         if (info) {
2425                 if (info->bitmap)
2426                         kmem_cache_free(btrfs_free_space_bitmap_cachep,
2427                                         info->bitmap);
2428                 kmem_cache_free(btrfs_free_space_cachep, info);
2429         }
2430
2431         return ret;
2432 }
2433
2434 /*
2435  * Free space merging rules:
2436  *  1) Merge trimmed areas together
2437  *  2) Let untrimmed areas coalesce with trimmed areas
2438  *  3) Always pull neighboring regions from bitmaps
2439  *
2440  * The above rules are for when we merge free space based on btrfs_trim_state.
2441  * Rules 2 and 3 are subtle because they are suboptimal, but are done for the
2442  * same reason: to promote larger extent regions which makes life easier for
2443  * find_free_extent().  Rule 2 enables coalescing based on the common path
2444  * being returning free space from btrfs_finish_extent_commit().  So when free
2445  * space is trimmed, it will prevent aggregating trimmed new region and
2446  * untrimmed regions in the rb_tree.  Rule 3 is purely to obtain larger extents
2447  * and provide find_free_extent() with the largest extents possible hoping for
2448  * the reuse path.
2449  */
2450 static bool try_merge_free_space(struct btrfs_free_space_ctl *ctl,
2451                           struct btrfs_free_space *info, bool update_stat)
2452 {
2453         struct btrfs_free_space *left_info = NULL;
2454         struct btrfs_free_space *right_info;
2455         bool merged = false;
2456         u64 offset = info->offset;
2457         u64 bytes = info->bytes;
2458         const bool is_trimmed = btrfs_free_space_trimmed(info);
2459         struct rb_node *right_prev = NULL;
2460
2461         /*
2462          * first we want to see if there is free space adjacent to the range we
2463          * are adding, if there is remove that struct and add a new one to
2464          * cover the entire range
2465          */
2466         right_info = tree_search_offset(ctl, offset + bytes, 0, 0);
2467         if (right_info)
2468                 right_prev = rb_prev(&right_info->offset_index);
2469
2470         if (right_prev)
2471                 left_info = rb_entry(right_prev, struct btrfs_free_space, offset_index);
2472         else if (!right_info)
2473                 left_info = tree_search_offset(ctl, offset - 1, 0, 0);
2474
2475         /* See try_merge_free_space() comment. */
2476         if (right_info && !right_info->bitmap &&
2477             (!is_trimmed || btrfs_free_space_trimmed(right_info))) {
2478                 unlink_free_space(ctl, right_info, update_stat);
2479                 info->bytes += right_info->bytes;
2480                 kmem_cache_free(btrfs_free_space_cachep, right_info);
2481                 merged = true;
2482         }
2483
2484         /* See try_merge_free_space() comment. */
2485         if (left_info && !left_info->bitmap &&
2486             left_info->offset + left_info->bytes == offset &&
2487             (!is_trimmed || btrfs_free_space_trimmed(left_info))) {
2488                 unlink_free_space(ctl, left_info, update_stat);
2489                 info->offset = left_info->offset;
2490                 info->bytes += left_info->bytes;
2491                 kmem_cache_free(btrfs_free_space_cachep, left_info);
2492                 merged = true;
2493         }
2494
2495         return merged;
2496 }
2497
2498 static bool steal_from_bitmap_to_end(struct btrfs_free_space_ctl *ctl,
2499                                      struct btrfs_free_space *info,
2500                                      bool update_stat)
2501 {
2502         struct btrfs_free_space *bitmap;
2503         unsigned long i;
2504         unsigned long j;
2505         const u64 end = info->offset + info->bytes;
2506         const u64 bitmap_offset = offset_to_bitmap(ctl, end);
2507         u64 bytes;
2508
2509         bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2510         if (!bitmap)
2511                 return false;
2512
2513         i = offset_to_bit(bitmap->offset, ctl->unit, end);
2514         j = find_next_zero_bit(bitmap->bitmap, BITS_PER_BITMAP, i);
2515         if (j == i)
2516                 return false;
2517         bytes = (j - i) * ctl->unit;
2518         info->bytes += bytes;
2519
2520         /* See try_merge_free_space() comment. */
2521         if (!btrfs_free_space_trimmed(bitmap))
2522                 info->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2523
2524         bitmap_clear_bits(ctl, bitmap, end, bytes, update_stat);
2525
2526         if (!bitmap->bytes)
2527                 free_bitmap(ctl, bitmap);
2528
2529         return true;
2530 }
2531
2532 static bool steal_from_bitmap_to_front(struct btrfs_free_space_ctl *ctl,
2533                                        struct btrfs_free_space *info,
2534                                        bool update_stat)
2535 {
2536         struct btrfs_free_space *bitmap;
2537         u64 bitmap_offset;
2538         unsigned long i;
2539         unsigned long j;
2540         unsigned long prev_j;
2541         u64 bytes;
2542
2543         bitmap_offset = offset_to_bitmap(ctl, info->offset);
2544         /* If we're on a boundary, try the previous logical bitmap. */
2545         if (bitmap_offset == info->offset) {
2546                 if (info->offset == 0)
2547                         return false;
2548                 bitmap_offset = offset_to_bitmap(ctl, info->offset - 1);
2549         }
2550
2551         bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2552         if (!bitmap)
2553                 return false;
2554
2555         i = offset_to_bit(bitmap->offset, ctl->unit, info->offset) - 1;
2556         j = 0;
2557         prev_j = (unsigned long)-1;
2558         for_each_clear_bit_from(j, bitmap->bitmap, BITS_PER_BITMAP) {
2559                 if (j > i)
2560                         break;
2561                 prev_j = j;
2562         }
2563         if (prev_j == i)
2564                 return false;
2565
2566         if (prev_j == (unsigned long)-1)
2567                 bytes = (i + 1) * ctl->unit;
2568         else
2569                 bytes = (i - prev_j) * ctl->unit;
2570
2571         info->offset -= bytes;
2572         info->bytes += bytes;
2573
2574         /* See try_merge_free_space() comment. */
2575         if (!btrfs_free_space_trimmed(bitmap))
2576                 info->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2577
2578         bitmap_clear_bits(ctl, bitmap, info->offset, bytes, update_stat);
2579
2580         if (!bitmap->bytes)
2581                 free_bitmap(ctl, bitmap);
2582
2583         return true;
2584 }
2585
2586 /*
2587  * We prefer always to allocate from extent entries, both for clustered and
2588  * non-clustered allocation requests. So when attempting to add a new extent
2589  * entry, try to see if there's adjacent free space in bitmap entries, and if
2590  * there is, migrate that space from the bitmaps to the extent.
2591  * Like this we get better chances of satisfying space allocation requests
2592  * because we attempt to satisfy them based on a single cache entry, and never
2593  * on 2 or more entries - even if the entries represent a contiguous free space
2594  * region (e.g. 1 extent entry + 1 bitmap entry starting where the extent entry
2595  * ends).
2596  */
2597 static void steal_from_bitmap(struct btrfs_free_space_ctl *ctl,
2598                               struct btrfs_free_space *info,
2599                               bool update_stat)
2600 {
2601         /*
2602          * Only work with disconnected entries, as we can change their offset,
2603          * and must be extent entries.
2604          */
2605         ASSERT(!info->bitmap);
2606         ASSERT(RB_EMPTY_NODE(&info->offset_index));
2607
2608         if (ctl->total_bitmaps > 0) {
2609                 bool stole_end;
2610                 bool stole_front = false;
2611
2612                 stole_end = steal_from_bitmap_to_end(ctl, info, update_stat);
2613                 if (ctl->total_bitmaps > 0)
2614                         stole_front = steal_from_bitmap_to_front(ctl, info,
2615                                                                  update_stat);
2616
2617                 if (stole_end || stole_front)
2618                         try_merge_free_space(ctl, info, update_stat);
2619         }
2620 }
2621
2622 static int __btrfs_add_free_space(struct btrfs_block_group *block_group,
2623                            u64 offset, u64 bytes,
2624                            enum btrfs_trim_state trim_state)
2625 {
2626         struct btrfs_fs_info *fs_info = block_group->fs_info;
2627         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2628         struct btrfs_free_space *info;
2629         int ret = 0;
2630         u64 filter_bytes = bytes;
2631
2632         ASSERT(!btrfs_is_zoned(fs_info));
2633
2634         info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
2635         if (!info)
2636                 return -ENOMEM;
2637
2638         info->offset = offset;
2639         info->bytes = bytes;
2640         info->trim_state = trim_state;
2641         RB_CLEAR_NODE(&info->offset_index);
2642         RB_CLEAR_NODE(&info->bytes_index);
2643
2644         spin_lock(&ctl->tree_lock);
2645
2646         if (try_merge_free_space(ctl, info, true))
2647                 goto link;
2648
2649         /*
2650          * There was no extent directly to the left or right of this new
2651          * extent then we know we're going to have to allocate a new extent, so
2652          * before we do that see if we need to drop this into a bitmap
2653          */
2654         ret = insert_into_bitmap(ctl, info);
2655         if (ret < 0) {
2656                 goto out;
2657         } else if (ret) {
2658                 ret = 0;
2659                 goto out;
2660         }
2661 link:
2662         /*
2663          * Only steal free space from adjacent bitmaps if we're sure we're not
2664          * going to add the new free space to existing bitmap entries - because
2665          * that would mean unnecessary work that would be reverted. Therefore
2666          * attempt to steal space from bitmaps if we're adding an extent entry.
2667          */
2668         steal_from_bitmap(ctl, info, true);
2669
2670         filter_bytes = max(filter_bytes, info->bytes);
2671
2672         ret = link_free_space(ctl, info);
2673         if (ret)
2674                 kmem_cache_free(btrfs_free_space_cachep, info);
2675 out:
2676         btrfs_discard_update_discardable(block_group);
2677         spin_unlock(&ctl->tree_lock);
2678
2679         if (ret) {
2680                 btrfs_crit(fs_info, "unable to add free space :%d", ret);
2681                 ASSERT(ret != -EEXIST);
2682         }
2683
2684         if (trim_state != BTRFS_TRIM_STATE_TRIMMED) {
2685                 btrfs_discard_check_filter(block_group, filter_bytes);
2686                 btrfs_discard_queue_work(&fs_info->discard_ctl, block_group);
2687         }
2688
2689         return ret;
2690 }
2691
2692 static int __btrfs_add_free_space_zoned(struct btrfs_block_group *block_group,
2693                                         u64 bytenr, u64 size, bool used)
2694 {
2695         struct btrfs_space_info *sinfo = block_group->space_info;
2696         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2697         u64 offset = bytenr - block_group->start;
2698         u64 to_free, to_unusable;
2699         int bg_reclaim_threshold = 0;
2700         bool initial = ((size == block_group->length) && (block_group->alloc_offset == 0));
2701         u64 reclaimable_unusable;
2702
2703         WARN_ON(!initial && offset + size > block_group->zone_capacity);
2704
2705         if (!initial)
2706                 bg_reclaim_threshold = READ_ONCE(sinfo->bg_reclaim_threshold);
2707
2708         spin_lock(&ctl->tree_lock);
2709         if (!used)
2710                 to_free = size;
2711         else if (initial)
2712                 to_free = block_group->zone_capacity;
2713         else if (offset >= block_group->alloc_offset)
2714                 to_free = size;
2715         else if (offset + size <= block_group->alloc_offset)
2716                 to_free = 0;
2717         else
2718                 to_free = offset + size - block_group->alloc_offset;
2719         to_unusable = size - to_free;
2720
2721         ctl->free_space += to_free;
2722         /*
2723          * If the block group is read-only, we should account freed space into
2724          * bytes_readonly.
2725          */
2726         if (!block_group->ro)
2727                 block_group->zone_unusable += to_unusable;
2728         spin_unlock(&ctl->tree_lock);
2729         if (!used) {
2730                 spin_lock(&block_group->lock);
2731                 block_group->alloc_offset -= size;
2732                 spin_unlock(&block_group->lock);
2733         }
2734
2735         reclaimable_unusable = block_group->zone_unusable -
2736                                (block_group->length - block_group->zone_capacity);
2737         /* All the region is now unusable. Mark it as unused and reclaim */
2738         if (block_group->zone_unusable == block_group->length) {
2739                 btrfs_mark_bg_unused(block_group);
2740         } else if (bg_reclaim_threshold &&
2741                    reclaimable_unusable >=
2742                    mult_perc(block_group->zone_capacity, bg_reclaim_threshold)) {
2743                 btrfs_mark_bg_to_reclaim(block_group);
2744         }
2745
2746         return 0;
2747 }
2748
2749 int btrfs_add_free_space(struct btrfs_block_group *block_group,
2750                          u64 bytenr, u64 size)
2751 {
2752         enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2753
2754         if (btrfs_is_zoned(block_group->fs_info))
2755                 return __btrfs_add_free_space_zoned(block_group, bytenr, size,
2756                                                     true);
2757
2758         if (btrfs_test_opt(block_group->fs_info, DISCARD_SYNC))
2759                 trim_state = BTRFS_TRIM_STATE_TRIMMED;
2760
2761         return __btrfs_add_free_space(block_group, bytenr, size, trim_state);
2762 }
2763
2764 int btrfs_add_free_space_unused(struct btrfs_block_group *block_group,
2765                                 u64 bytenr, u64 size)
2766 {
2767         if (btrfs_is_zoned(block_group->fs_info))
2768                 return __btrfs_add_free_space_zoned(block_group, bytenr, size,
2769                                                     false);
2770
2771         return btrfs_add_free_space(block_group, bytenr, size);
2772 }
2773
2774 /*
2775  * This is a subtle distinction because when adding free space back in general,
2776  * we want it to be added as untrimmed for async. But in the case where we add
2777  * it on loading of a block group, we want to consider it trimmed.
2778  */
2779 int btrfs_add_free_space_async_trimmed(struct btrfs_block_group *block_group,
2780                                        u64 bytenr, u64 size)
2781 {
2782         enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2783
2784         if (btrfs_is_zoned(block_group->fs_info))
2785                 return __btrfs_add_free_space_zoned(block_group, bytenr, size,
2786                                                     true);
2787
2788         if (btrfs_test_opt(block_group->fs_info, DISCARD_SYNC) ||
2789             btrfs_test_opt(block_group->fs_info, DISCARD_ASYNC))
2790                 trim_state = BTRFS_TRIM_STATE_TRIMMED;
2791
2792         return __btrfs_add_free_space(block_group, bytenr, size, trim_state);
2793 }
2794
2795 int btrfs_remove_free_space(struct btrfs_block_group *block_group,
2796                             u64 offset, u64 bytes)
2797 {
2798         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2799         struct btrfs_free_space *info;
2800         int ret;
2801         bool re_search = false;
2802
2803         if (btrfs_is_zoned(block_group->fs_info)) {
2804                 /*
2805                  * This can happen with conventional zones when replaying log.
2806                  * Since the allocation info of tree-log nodes are not recorded
2807                  * to the extent-tree, calculate_alloc_pointer() failed to
2808                  * advance the allocation pointer after last allocated tree log
2809                  * node blocks.
2810                  *
2811                  * This function is called from
2812                  * btrfs_pin_extent_for_log_replay() when replaying the log.
2813                  * Advance the pointer not to overwrite the tree-log nodes.
2814                  */
2815                 if (block_group->start + block_group->alloc_offset <
2816                     offset + bytes) {
2817                         block_group->alloc_offset =
2818                                 offset + bytes - block_group->start;
2819                 }
2820                 return 0;
2821         }
2822
2823         spin_lock(&ctl->tree_lock);
2824
2825 again:
2826         ret = 0;
2827         if (!bytes)
2828                 goto out_lock;
2829
2830         info = tree_search_offset(ctl, offset, 0, 0);
2831         if (!info) {
2832                 /*
2833                  * oops didn't find an extent that matched the space we wanted
2834                  * to remove, look for a bitmap instead
2835                  */
2836                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2837                                           1, 0);
2838                 if (!info) {
2839                         /*
2840                          * If we found a partial bit of our free space in a
2841                          * bitmap but then couldn't find the other part this may
2842                          * be a problem, so WARN about it.
2843                          */
2844                         WARN_ON(re_search);
2845                         goto out_lock;
2846                 }
2847         }
2848
2849         re_search = false;
2850         if (!info->bitmap) {
2851                 unlink_free_space(ctl, info, true);
2852                 if (offset == info->offset) {
2853                         u64 to_free = min(bytes, info->bytes);
2854
2855                         info->bytes -= to_free;
2856                         info->offset += to_free;
2857                         if (info->bytes) {
2858                                 ret = link_free_space(ctl, info);
2859                                 WARN_ON(ret);
2860                         } else {
2861                                 kmem_cache_free(btrfs_free_space_cachep, info);
2862                         }
2863
2864                         offset += to_free;
2865                         bytes -= to_free;
2866                         goto again;
2867                 } else {
2868                         u64 old_end = info->bytes + info->offset;
2869
2870                         info->bytes = offset - info->offset;
2871                         ret = link_free_space(ctl, info);
2872                         WARN_ON(ret);
2873                         if (ret)
2874                                 goto out_lock;
2875
2876                         /* Not enough bytes in this entry to satisfy us */
2877                         if (old_end < offset + bytes) {
2878                                 bytes -= old_end - offset;
2879                                 offset = old_end;
2880                                 goto again;
2881                         } else if (old_end == offset + bytes) {
2882                                 /* all done */
2883                                 goto out_lock;
2884                         }
2885                         spin_unlock(&ctl->tree_lock);
2886
2887                         ret = __btrfs_add_free_space(block_group,
2888                                                      offset + bytes,
2889                                                      old_end - (offset + bytes),
2890                                                      info->trim_state);
2891                         WARN_ON(ret);
2892                         goto out;
2893                 }
2894         }
2895
2896         ret = remove_from_bitmap(ctl, info, &offset, &bytes);
2897         if (ret == -EAGAIN) {
2898                 re_search = true;
2899                 goto again;
2900         }
2901 out_lock:
2902         btrfs_discard_update_discardable(block_group);
2903         spin_unlock(&ctl->tree_lock);
2904 out:
2905         return ret;
2906 }
2907
2908 void btrfs_dump_free_space(struct btrfs_block_group *block_group,
2909                            u64 bytes)
2910 {
2911         struct btrfs_fs_info *fs_info = block_group->fs_info;
2912         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2913         struct btrfs_free_space *info;
2914         struct rb_node *n;
2915         int count = 0;
2916
2917         /*
2918          * Zoned btrfs does not use free space tree and cluster. Just print
2919          * out the free space after the allocation offset.
2920          */
2921         if (btrfs_is_zoned(fs_info)) {
2922                 btrfs_info(fs_info, "free space %llu active %d",
2923                            block_group->zone_capacity - block_group->alloc_offset,
2924                            test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE,
2925                                     &block_group->runtime_flags));
2926                 return;
2927         }
2928
2929         spin_lock(&ctl->tree_lock);
2930         for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
2931                 info = rb_entry(n, struct btrfs_free_space, offset_index);
2932                 if (info->bytes >= bytes && !block_group->ro)
2933                         count++;
2934                 btrfs_crit(fs_info, "entry offset %llu, bytes %llu, bitmap %s",
2935                            info->offset, info->bytes,
2936                        (info->bitmap) ? "yes" : "no");
2937         }
2938         spin_unlock(&ctl->tree_lock);
2939         btrfs_info(fs_info, "block group has cluster?: %s",
2940                list_empty(&block_group->cluster_list) ? "no" : "yes");
2941         btrfs_info(fs_info,
2942                    "%d free space entries at or bigger than %llu bytes",
2943                    count, bytes);
2944 }
2945
2946 void btrfs_init_free_space_ctl(struct btrfs_block_group *block_group,
2947                                struct btrfs_free_space_ctl *ctl)
2948 {
2949         struct btrfs_fs_info *fs_info = block_group->fs_info;
2950
2951         spin_lock_init(&ctl->tree_lock);
2952         ctl->unit = fs_info->sectorsize;
2953         ctl->start = block_group->start;
2954         ctl->block_group = block_group;
2955         ctl->op = &free_space_op;
2956         ctl->free_space_bytes = RB_ROOT_CACHED;
2957         INIT_LIST_HEAD(&ctl->trimming_ranges);
2958         mutex_init(&ctl->cache_writeout_mutex);
2959
2960         /*
2961          * we only want to have 32k of ram per block group for keeping
2962          * track of free space, and if we pass 1/2 of that we want to
2963          * start converting things over to using bitmaps
2964          */
2965         ctl->extents_thresh = (SZ_32K / 2) / sizeof(struct btrfs_free_space);
2966 }
2967
2968 /*
2969  * for a given cluster, put all of its extents back into the free
2970  * space cache.  If the block group passed doesn't match the block group
2971  * pointed to by the cluster, someone else raced in and freed the
2972  * cluster already.  In that case, we just return without changing anything
2973  */
2974 static void __btrfs_return_cluster_to_free_space(
2975                              struct btrfs_block_group *block_group,
2976                              struct btrfs_free_cluster *cluster)
2977 {
2978         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2979         struct rb_node *node;
2980
2981         lockdep_assert_held(&ctl->tree_lock);
2982
2983         spin_lock(&cluster->lock);
2984         if (cluster->block_group != block_group) {
2985                 spin_unlock(&cluster->lock);
2986                 return;
2987         }
2988
2989         cluster->block_group = NULL;
2990         cluster->window_start = 0;
2991         list_del_init(&cluster->block_group_list);
2992
2993         node = rb_first(&cluster->root);
2994         while (node) {
2995                 struct btrfs_free_space *entry;
2996
2997                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2998                 node = rb_next(&entry->offset_index);
2999                 rb_erase(&entry->offset_index, &cluster->root);
3000                 RB_CLEAR_NODE(&entry->offset_index);
3001
3002                 if (!entry->bitmap) {
3003                         /* Merging treats extents as if they were new */
3004                         if (!btrfs_free_space_trimmed(entry)) {
3005                                 ctl->discardable_extents[BTRFS_STAT_CURR]--;
3006                                 ctl->discardable_bytes[BTRFS_STAT_CURR] -=
3007                                         entry->bytes;
3008                         }
3009
3010                         try_merge_free_space(ctl, entry, false);
3011                         steal_from_bitmap(ctl, entry, false);
3012
3013                         /* As we insert directly, update these statistics */
3014                         if (!btrfs_free_space_trimmed(entry)) {
3015                                 ctl->discardable_extents[BTRFS_STAT_CURR]++;
3016                                 ctl->discardable_bytes[BTRFS_STAT_CURR] +=
3017                                         entry->bytes;
3018                         }
3019                 }
3020                 tree_insert_offset(ctl, NULL, entry);
3021                 rb_add_cached(&entry->bytes_index, &ctl->free_space_bytes,
3022                               entry_less);
3023         }
3024         cluster->root = RB_ROOT;
3025         spin_unlock(&cluster->lock);
3026         btrfs_put_block_group(block_group);
3027 }
3028
3029 void btrfs_remove_free_space_cache(struct btrfs_block_group *block_group)
3030 {
3031         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3032         struct btrfs_free_cluster *cluster;
3033         struct list_head *head;
3034
3035         spin_lock(&ctl->tree_lock);
3036         while ((head = block_group->cluster_list.next) !=
3037                &block_group->cluster_list) {
3038                 cluster = list_entry(head, struct btrfs_free_cluster,
3039                                      block_group_list);
3040
3041                 WARN_ON(cluster->block_group != block_group);
3042                 __btrfs_return_cluster_to_free_space(block_group, cluster);
3043
3044                 cond_resched_lock(&ctl->tree_lock);
3045         }
3046         __btrfs_remove_free_space_cache(ctl);
3047         btrfs_discard_update_discardable(block_group);
3048         spin_unlock(&ctl->tree_lock);
3049
3050 }
3051
3052 /*
3053  * Walk @block_group's free space rb_tree to determine if everything is trimmed.
3054  */
3055 bool btrfs_is_free_space_trimmed(struct btrfs_block_group *block_group)
3056 {
3057         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3058         struct btrfs_free_space *info;
3059         struct rb_node *node;
3060         bool ret = true;
3061
3062         spin_lock(&ctl->tree_lock);
3063         node = rb_first(&ctl->free_space_offset);
3064
3065         while (node) {
3066                 info = rb_entry(node, struct btrfs_free_space, offset_index);
3067
3068                 if (!btrfs_free_space_trimmed(info)) {
3069                         ret = false;
3070                         break;
3071                 }
3072
3073                 node = rb_next(node);
3074         }
3075
3076         spin_unlock(&ctl->tree_lock);
3077         return ret;
3078 }
3079
3080 u64 btrfs_find_space_for_alloc(struct btrfs_block_group *block_group,
3081                                u64 offset, u64 bytes, u64 empty_size,
3082                                u64 *max_extent_size)
3083 {
3084         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3085         struct btrfs_discard_ctl *discard_ctl =
3086                                         &block_group->fs_info->discard_ctl;
3087         struct btrfs_free_space *entry = NULL;
3088         u64 bytes_search = bytes + empty_size;
3089         u64 ret = 0;
3090         u64 align_gap = 0;
3091         u64 align_gap_len = 0;
3092         enum btrfs_trim_state align_gap_trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3093         bool use_bytes_index = (offset == block_group->start);
3094
3095         ASSERT(!btrfs_is_zoned(block_group->fs_info));
3096
3097         spin_lock(&ctl->tree_lock);
3098         entry = find_free_space(ctl, &offset, &bytes_search,
3099                                 block_group->full_stripe_len, max_extent_size,
3100                                 use_bytes_index);
3101         if (!entry)
3102                 goto out;
3103
3104         ret = offset;
3105         if (entry->bitmap) {
3106                 bitmap_clear_bits(ctl, entry, offset, bytes, true);
3107
3108                 if (!btrfs_free_space_trimmed(entry))
3109                         atomic64_add(bytes, &discard_ctl->discard_bytes_saved);
3110
3111                 if (!entry->bytes)
3112                         free_bitmap(ctl, entry);
3113         } else {
3114                 unlink_free_space(ctl, entry, true);
3115                 align_gap_len = offset - entry->offset;
3116                 align_gap = entry->offset;
3117                 align_gap_trim_state = entry->trim_state;
3118
3119                 if (!btrfs_free_space_trimmed(entry))
3120                         atomic64_add(bytes, &discard_ctl->discard_bytes_saved);
3121
3122                 entry->offset = offset + bytes;
3123                 WARN_ON(entry->bytes < bytes + align_gap_len);
3124
3125                 entry->bytes -= bytes + align_gap_len;
3126                 if (!entry->bytes)
3127                         kmem_cache_free(btrfs_free_space_cachep, entry);
3128                 else
3129                         link_free_space(ctl, entry);
3130         }
3131 out:
3132         btrfs_discard_update_discardable(block_group);
3133         spin_unlock(&ctl->tree_lock);
3134
3135         if (align_gap_len)
3136                 __btrfs_add_free_space(block_group, align_gap, align_gap_len,
3137                                        align_gap_trim_state);
3138         return ret;
3139 }
3140
3141 /*
3142  * given a cluster, put all of its extents back into the free space
3143  * cache.  If a block group is passed, this function will only free
3144  * a cluster that belongs to the passed block group.
3145  *
3146  * Otherwise, it'll get a reference on the block group pointed to by the
3147  * cluster and remove the cluster from it.
3148  */
3149 void btrfs_return_cluster_to_free_space(
3150                                struct btrfs_block_group *block_group,
3151                                struct btrfs_free_cluster *cluster)
3152 {
3153         struct btrfs_free_space_ctl *ctl;
3154
3155         /* first, get a safe pointer to the block group */
3156         spin_lock(&cluster->lock);
3157         if (!block_group) {
3158                 block_group = cluster->block_group;
3159                 if (!block_group) {
3160                         spin_unlock(&cluster->lock);
3161                         return;
3162                 }
3163         } else if (cluster->block_group != block_group) {
3164                 /* someone else has already freed it don't redo their work */
3165                 spin_unlock(&cluster->lock);
3166                 return;
3167         }
3168         btrfs_get_block_group(block_group);
3169         spin_unlock(&cluster->lock);
3170
3171         ctl = block_group->free_space_ctl;
3172
3173         /* now return any extents the cluster had on it */
3174         spin_lock(&ctl->tree_lock);
3175         __btrfs_return_cluster_to_free_space(block_group, cluster);
3176         spin_unlock(&ctl->tree_lock);
3177
3178         btrfs_discard_queue_work(&block_group->fs_info->discard_ctl, block_group);
3179
3180         /* finally drop our ref */
3181         btrfs_put_block_group(block_group);
3182 }
3183
3184 static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group *block_group,
3185                                    struct btrfs_free_cluster *cluster,
3186                                    struct btrfs_free_space *entry,
3187                                    u64 bytes, u64 min_start,
3188                                    u64 *max_extent_size)
3189 {
3190         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3191         int err;
3192         u64 search_start = cluster->window_start;
3193         u64 search_bytes = bytes;
3194         u64 ret = 0;
3195
3196         search_start = min_start;
3197         search_bytes = bytes;
3198
3199         err = search_bitmap(ctl, entry, &search_start, &search_bytes, true);
3200         if (err) {
3201                 *max_extent_size = max(get_max_extent_size(entry),
3202                                        *max_extent_size);
3203                 return 0;
3204         }
3205
3206         ret = search_start;
3207         bitmap_clear_bits(ctl, entry, ret, bytes, false);
3208
3209         return ret;
3210 }
3211
3212 /*
3213  * given a cluster, try to allocate 'bytes' from it, returns 0
3214  * if it couldn't find anything suitably large, or a logical disk offset
3215  * if things worked out
3216  */
3217 u64 btrfs_alloc_from_cluster(struct btrfs_block_group *block_group,
3218                              struct btrfs_free_cluster *cluster, u64 bytes,
3219                              u64 min_start, u64 *max_extent_size)
3220 {
3221         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3222         struct btrfs_discard_ctl *discard_ctl =
3223                                         &block_group->fs_info->discard_ctl;
3224         struct btrfs_free_space *entry = NULL;
3225         struct rb_node *node;
3226         u64 ret = 0;
3227
3228         ASSERT(!btrfs_is_zoned(block_group->fs_info));
3229
3230         spin_lock(&cluster->lock);
3231         if (bytes > cluster->max_size)
3232                 goto out;
3233
3234         if (cluster->block_group != block_group)
3235                 goto out;
3236
3237         node = rb_first(&cluster->root);
3238         if (!node)
3239                 goto out;
3240
3241         entry = rb_entry(node, struct btrfs_free_space, offset_index);
3242         while (1) {
3243                 if (entry->bytes < bytes)
3244                         *max_extent_size = max(get_max_extent_size(entry),
3245                                                *max_extent_size);
3246
3247                 if (entry->bytes < bytes ||
3248                     (!entry->bitmap && entry->offset < min_start)) {
3249                         node = rb_next(&entry->offset_index);
3250                         if (!node)
3251                                 break;
3252                         entry = rb_entry(node, struct btrfs_free_space,
3253                                          offset_index);
3254                         continue;
3255                 }
3256
3257                 if (entry->bitmap) {
3258                         ret = btrfs_alloc_from_bitmap(block_group,
3259                                                       cluster, entry, bytes,
3260                                                       cluster->window_start,
3261                                                       max_extent_size);
3262                         if (ret == 0) {
3263                                 node = rb_next(&entry->offset_index);
3264                                 if (!node)
3265                                         break;
3266                                 entry = rb_entry(node, struct btrfs_free_space,
3267                                                  offset_index);
3268                                 continue;
3269                         }
3270                         cluster->window_start += bytes;
3271                 } else {
3272                         ret = entry->offset;
3273
3274                         entry->offset += bytes;
3275                         entry->bytes -= bytes;
3276                 }
3277
3278                 break;
3279         }
3280 out:
3281         spin_unlock(&cluster->lock);
3282
3283         if (!ret)
3284                 return 0;
3285
3286         spin_lock(&ctl->tree_lock);
3287
3288         if (!btrfs_free_space_trimmed(entry))
3289                 atomic64_add(bytes, &discard_ctl->discard_bytes_saved);
3290
3291         ctl->free_space -= bytes;
3292         if (!entry->bitmap && !btrfs_free_space_trimmed(entry))
3293                 ctl->discardable_bytes[BTRFS_STAT_CURR] -= bytes;
3294
3295         spin_lock(&cluster->lock);
3296         if (entry->bytes == 0) {
3297                 rb_erase(&entry->offset_index, &cluster->root);
3298                 ctl->free_extents--;
3299                 if (entry->bitmap) {
3300                         kmem_cache_free(btrfs_free_space_bitmap_cachep,
3301                                         entry->bitmap);
3302                         ctl->total_bitmaps--;
3303                         recalculate_thresholds(ctl);
3304                 } else if (!btrfs_free_space_trimmed(entry)) {
3305                         ctl->discardable_extents[BTRFS_STAT_CURR]--;
3306                 }
3307                 kmem_cache_free(btrfs_free_space_cachep, entry);
3308         }
3309
3310         spin_unlock(&cluster->lock);
3311         spin_unlock(&ctl->tree_lock);
3312
3313         return ret;
3314 }
3315
3316 static int btrfs_bitmap_cluster(struct btrfs_block_group *block_group,
3317                                 struct btrfs_free_space *entry,
3318                                 struct btrfs_free_cluster *cluster,
3319                                 u64 offset, u64 bytes,
3320                                 u64 cont1_bytes, u64 min_bytes)
3321 {
3322         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3323         unsigned long next_zero;
3324         unsigned long i;
3325         unsigned long want_bits;
3326         unsigned long min_bits;
3327         unsigned long found_bits;
3328         unsigned long max_bits = 0;
3329         unsigned long start = 0;
3330         unsigned long total_found = 0;
3331         int ret;
3332
3333         lockdep_assert_held(&ctl->tree_lock);
3334
3335         i = offset_to_bit(entry->offset, ctl->unit,
3336                           max_t(u64, offset, entry->offset));
3337         want_bits = bytes_to_bits(bytes, ctl->unit);
3338         min_bits = bytes_to_bits(min_bytes, ctl->unit);
3339
3340         /*
3341          * Don't bother looking for a cluster in this bitmap if it's heavily
3342          * fragmented.
3343          */
3344         if (entry->max_extent_size &&
3345             entry->max_extent_size < cont1_bytes)
3346                 return -ENOSPC;
3347 again:
3348         found_bits = 0;
3349         for_each_set_bit_from(i, entry->bitmap, BITS_PER_BITMAP) {
3350                 next_zero = find_next_zero_bit(entry->bitmap,
3351                                                BITS_PER_BITMAP, i);
3352                 if (next_zero - i >= min_bits) {
3353                         found_bits = next_zero - i;
3354                         if (found_bits > max_bits)
3355                                 max_bits = found_bits;
3356                         break;
3357                 }
3358                 if (next_zero - i > max_bits)
3359                         max_bits = next_zero - i;
3360                 i = next_zero;
3361         }
3362
3363         if (!found_bits) {
3364                 entry->max_extent_size = (u64)max_bits * ctl->unit;
3365                 return -ENOSPC;
3366         }
3367
3368         if (!total_found) {
3369                 start = i;
3370                 cluster->max_size = 0;
3371         }
3372
3373         total_found += found_bits;
3374
3375         if (cluster->max_size < found_bits * ctl->unit)
3376                 cluster->max_size = found_bits * ctl->unit;
3377
3378         if (total_found < want_bits || cluster->max_size < cont1_bytes) {
3379                 i = next_zero + 1;
3380                 goto again;
3381         }
3382
3383         cluster->window_start = start * ctl->unit + entry->offset;
3384         rb_erase(&entry->offset_index, &ctl->free_space_offset);
3385         rb_erase_cached(&entry->bytes_index, &ctl->free_space_bytes);
3386
3387         /*
3388          * We need to know if we're currently on the normal space index when we
3389          * manipulate the bitmap so that we know we need to remove and re-insert
3390          * it into the space_index tree.  Clear the bytes_index node here so the
3391          * bitmap manipulation helpers know not to mess with the space_index
3392          * until this bitmap entry is added back into the normal cache.
3393          */
3394         RB_CLEAR_NODE(&entry->bytes_index);
3395
3396         ret = tree_insert_offset(ctl, cluster, entry);
3397         ASSERT(!ret); /* -EEXIST; Logic error */
3398
3399         trace_btrfs_setup_cluster(block_group, cluster,
3400                                   total_found * ctl->unit, 1);
3401         return 0;
3402 }
3403
3404 /*
3405  * This searches the block group for just extents to fill the cluster with.
3406  * Try to find a cluster with at least bytes total bytes, at least one
3407  * extent of cont1_bytes, and other clusters of at least min_bytes.
3408  */
3409 static noinline int
3410 setup_cluster_no_bitmap(struct btrfs_block_group *block_group,
3411                         struct btrfs_free_cluster *cluster,
3412                         struct list_head *bitmaps, u64 offset, u64 bytes,
3413                         u64 cont1_bytes, u64 min_bytes)
3414 {
3415         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3416         struct btrfs_free_space *first = NULL;
3417         struct btrfs_free_space *entry = NULL;
3418         struct btrfs_free_space *last;
3419         struct rb_node *node;
3420         u64 window_free;
3421         u64 max_extent;
3422         u64 total_size = 0;
3423
3424         lockdep_assert_held(&ctl->tree_lock);
3425
3426         entry = tree_search_offset(ctl, offset, 0, 1);
3427         if (!entry)
3428                 return -ENOSPC;
3429
3430         /*
3431          * We don't want bitmaps, so just move along until we find a normal
3432          * extent entry.
3433          */
3434         while (entry->bitmap || entry->bytes < min_bytes) {
3435                 if (entry->bitmap && list_empty(&entry->list))
3436                         list_add_tail(&entry->list, bitmaps);
3437                 node = rb_next(&entry->offset_index);
3438                 if (!node)
3439                         return -ENOSPC;
3440                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3441         }
3442
3443         window_free = entry->bytes;
3444         max_extent = entry->bytes;
3445         first = entry;
3446         last = entry;
3447
3448         for (node = rb_next(&entry->offset_index); node;
3449              node = rb_next(&entry->offset_index)) {
3450                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3451
3452                 if (entry->bitmap) {
3453                         if (list_empty(&entry->list))
3454                                 list_add_tail(&entry->list, bitmaps);
3455                         continue;
3456                 }
3457
3458                 if (entry->bytes < min_bytes)
3459                         continue;
3460
3461                 last = entry;
3462                 window_free += entry->bytes;
3463                 if (entry->bytes > max_extent)
3464                         max_extent = entry->bytes;
3465         }
3466
3467         if (window_free < bytes || max_extent < cont1_bytes)
3468                 return -ENOSPC;
3469
3470         cluster->window_start = first->offset;
3471
3472         node = &first->offset_index;
3473
3474         /*
3475          * now we've found our entries, pull them out of the free space
3476          * cache and put them into the cluster rbtree
3477          */
3478         do {
3479                 int ret;
3480
3481                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3482                 node = rb_next(&entry->offset_index);
3483                 if (entry->bitmap || entry->bytes < min_bytes)
3484                         continue;
3485
3486                 rb_erase(&entry->offset_index, &ctl->free_space_offset);
3487                 rb_erase_cached(&entry->bytes_index, &ctl->free_space_bytes);
3488                 ret = tree_insert_offset(ctl, cluster, entry);
3489                 total_size += entry->bytes;
3490                 ASSERT(!ret); /* -EEXIST; Logic error */
3491         } while (node && entry != last);
3492
3493         cluster->max_size = max_extent;
3494         trace_btrfs_setup_cluster(block_group, cluster, total_size, 0);
3495         return 0;
3496 }
3497
3498 /*
3499  * This specifically looks for bitmaps that may work in the cluster, we assume
3500  * that we have already failed to find extents that will work.
3501  */
3502 static noinline int
3503 setup_cluster_bitmap(struct btrfs_block_group *block_group,
3504                      struct btrfs_free_cluster *cluster,
3505                      struct list_head *bitmaps, u64 offset, u64 bytes,
3506                      u64 cont1_bytes, u64 min_bytes)
3507 {
3508         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3509         struct btrfs_free_space *entry = NULL;
3510         int ret = -ENOSPC;
3511         u64 bitmap_offset = offset_to_bitmap(ctl, offset);
3512
3513         if (ctl->total_bitmaps == 0)
3514                 return -ENOSPC;
3515
3516         /*
3517          * The bitmap that covers offset won't be in the list unless offset
3518          * is just its start offset.
3519          */
3520         if (!list_empty(bitmaps))
3521                 entry = list_first_entry(bitmaps, struct btrfs_free_space, list);
3522
3523         if (!entry || entry->offset != bitmap_offset) {
3524                 entry = tree_search_offset(ctl, bitmap_offset, 1, 0);
3525                 if (entry && list_empty(&entry->list))
3526                         list_add(&entry->list, bitmaps);
3527         }
3528
3529         list_for_each_entry(entry, bitmaps, list) {
3530                 if (entry->bytes < bytes)
3531                         continue;
3532                 ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
3533                                            bytes, cont1_bytes, min_bytes);
3534                 if (!ret)
3535                         return 0;
3536         }
3537
3538         /*
3539          * The bitmaps list has all the bitmaps that record free space
3540          * starting after offset, so no more search is required.
3541          */
3542         return -ENOSPC;
3543 }
3544
3545 /*
3546  * here we try to find a cluster of blocks in a block group.  The goal
3547  * is to find at least bytes+empty_size.
3548  * We might not find them all in one contiguous area.
3549  *
3550  * returns zero and sets up cluster if things worked out, otherwise
3551  * it returns -enospc
3552  */
3553 int btrfs_find_space_cluster(struct btrfs_block_group *block_group,
3554                              struct btrfs_free_cluster *cluster,
3555                              u64 offset, u64 bytes, u64 empty_size)
3556 {
3557         struct btrfs_fs_info *fs_info = block_group->fs_info;
3558         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3559         struct btrfs_free_space *entry, *tmp;
3560         LIST_HEAD(bitmaps);
3561         u64 min_bytes;
3562         u64 cont1_bytes;
3563         int ret;
3564
3565         /*
3566          * Choose the minimum extent size we'll require for this
3567          * cluster.  For SSD_SPREAD, don't allow any fragmentation.
3568          * For metadata, allow allocates with smaller extents.  For
3569          * data, keep it dense.
3570          */
3571         if (btrfs_test_opt(fs_info, SSD_SPREAD)) {
3572                 cont1_bytes = bytes + empty_size;
3573                 min_bytes = cont1_bytes;
3574         } else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
3575                 cont1_bytes = bytes;
3576                 min_bytes = fs_info->sectorsize;
3577         } else {
3578                 cont1_bytes = max(bytes, (bytes + empty_size) >> 2);
3579                 min_bytes = fs_info->sectorsize;
3580         }
3581
3582         spin_lock(&ctl->tree_lock);
3583
3584         /*
3585          * If we know we don't have enough space to make a cluster don't even
3586          * bother doing all the work to try and find one.
3587          */
3588         if (ctl->free_space < bytes) {
3589                 spin_unlock(&ctl->tree_lock);
3590                 return -ENOSPC;
3591         }
3592
3593         spin_lock(&cluster->lock);
3594
3595         /* someone already found a cluster, hooray */
3596         if (cluster->block_group) {
3597                 ret = 0;
3598                 goto out;
3599         }
3600
3601         trace_btrfs_find_cluster(block_group, offset, bytes, empty_size,
3602                                  min_bytes);
3603
3604         ret = setup_cluster_no_bitmap(block_group, cluster, &bitmaps, offset,
3605                                       bytes + empty_size,
3606                                       cont1_bytes, min_bytes);
3607         if (ret)
3608                 ret = setup_cluster_bitmap(block_group, cluster, &bitmaps,
3609                                            offset, bytes + empty_size,
3610                                            cont1_bytes, min_bytes);
3611
3612         /* Clear our temporary list */
3613         list_for_each_entry_safe(entry, tmp, &bitmaps, list)
3614                 list_del_init(&entry->list);
3615
3616         if (!ret) {
3617                 btrfs_get_block_group(block_group);
3618                 list_add_tail(&cluster->block_group_list,
3619                               &block_group->cluster_list);
3620                 cluster->block_group = block_group;
3621         } else {
3622                 trace_btrfs_failed_cluster_setup(block_group);
3623         }
3624 out:
3625         spin_unlock(&cluster->lock);
3626         spin_unlock(&ctl->tree_lock);
3627
3628         return ret;
3629 }
3630
3631 /*
3632  * simple code to zero out a cluster
3633  */
3634 void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
3635 {
3636         spin_lock_init(&cluster->lock);
3637         spin_lock_init(&cluster->refill_lock);
3638         cluster->root = RB_ROOT;
3639         cluster->max_size = 0;
3640         cluster->fragmented = false;
3641         INIT_LIST_HEAD(&cluster->block_group_list);
3642         cluster->block_group = NULL;
3643 }
3644
3645 static int do_trimming(struct btrfs_block_group *block_group,
3646                        u64 *total_trimmed, u64 start, u64 bytes,
3647                        u64 reserved_start, u64 reserved_bytes,
3648                        enum btrfs_trim_state reserved_trim_state,
3649                        struct btrfs_trim_range *trim_entry)
3650 {
3651         struct btrfs_space_info *space_info = block_group->space_info;
3652         struct btrfs_fs_info *fs_info = block_group->fs_info;
3653         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3654         int ret;
3655         int update = 0;
3656         const u64 end = start + bytes;
3657         const u64 reserved_end = reserved_start + reserved_bytes;
3658         enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3659         u64 trimmed = 0;
3660
3661         spin_lock(&space_info->lock);
3662         spin_lock(&block_group->lock);
3663         if (!block_group->ro) {
3664                 block_group->reserved += reserved_bytes;
3665                 space_info->bytes_reserved += reserved_bytes;
3666                 update = 1;
3667         }
3668         spin_unlock(&block_group->lock);
3669         spin_unlock(&space_info->lock);
3670
3671         ret = btrfs_discard_extent(fs_info, start, bytes, &trimmed);
3672         if (!ret) {
3673                 *total_trimmed += trimmed;
3674                 trim_state = BTRFS_TRIM_STATE_TRIMMED;
3675         }
3676
3677         mutex_lock(&ctl->cache_writeout_mutex);
3678         if (reserved_start < start)
3679                 __btrfs_add_free_space(block_group, reserved_start,
3680                                        start - reserved_start,
3681                                        reserved_trim_state);
3682         if (end < reserved_end)
3683                 __btrfs_add_free_space(block_group, end, reserved_end - end,
3684                                        reserved_trim_state);
3685         __btrfs_add_free_space(block_group, start, bytes, trim_state);
3686         list_del(&trim_entry->list);
3687         mutex_unlock(&ctl->cache_writeout_mutex);
3688
3689         if (update) {
3690                 spin_lock(&space_info->lock);
3691                 spin_lock(&block_group->lock);
3692                 if (block_group->ro)
3693                         space_info->bytes_readonly += reserved_bytes;
3694                 block_group->reserved -= reserved_bytes;
3695                 space_info->bytes_reserved -= reserved_bytes;
3696                 spin_unlock(&block_group->lock);
3697                 spin_unlock(&space_info->lock);
3698         }
3699
3700         return ret;
3701 }
3702
3703 /*
3704  * If @async is set, then we will trim 1 region and return.
3705  */
3706 static int trim_no_bitmap(struct btrfs_block_group *block_group,
3707                           u64 *total_trimmed, u64 start, u64 end, u64 minlen,
3708                           bool async)
3709 {
3710         struct btrfs_discard_ctl *discard_ctl =
3711                                         &block_group->fs_info->discard_ctl;
3712         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3713         struct btrfs_free_space *entry;
3714         struct rb_node *node;
3715         int ret = 0;
3716         u64 extent_start;
3717         u64 extent_bytes;
3718         enum btrfs_trim_state extent_trim_state;
3719         u64 bytes;
3720         const u64 max_discard_size = READ_ONCE(discard_ctl->max_discard_size);
3721
3722         while (start < end) {
3723                 struct btrfs_trim_range trim_entry;
3724
3725                 mutex_lock(&ctl->cache_writeout_mutex);
3726                 spin_lock(&ctl->tree_lock);
3727
3728                 if (ctl->free_space < minlen)
3729                         goto out_unlock;
3730
3731                 entry = tree_search_offset(ctl, start, 0, 1);
3732                 if (!entry)
3733                         goto out_unlock;
3734
3735                 /* Skip bitmaps and if async, already trimmed entries */
3736                 while (entry->bitmap ||
3737                        (async && btrfs_free_space_trimmed(entry))) {
3738                         node = rb_next(&entry->offset_index);
3739                         if (!node)
3740                                 goto out_unlock;
3741                         entry = rb_entry(node, struct btrfs_free_space,
3742                                          offset_index);
3743                 }
3744
3745                 if (entry->offset >= end)
3746                         goto out_unlock;
3747
3748                 extent_start = entry->offset;
3749                 extent_bytes = entry->bytes;
3750                 extent_trim_state = entry->trim_state;
3751                 if (async) {
3752                         start = entry->offset;
3753                         bytes = entry->bytes;
3754                         if (bytes < minlen) {
3755                                 spin_unlock(&ctl->tree_lock);
3756                                 mutex_unlock(&ctl->cache_writeout_mutex);
3757                                 goto next;
3758                         }
3759                         unlink_free_space(ctl, entry, true);
3760                         /*
3761                          * Let bytes = BTRFS_MAX_DISCARD_SIZE + X.
3762                          * If X < BTRFS_ASYNC_DISCARD_MIN_FILTER, we won't trim
3763                          * X when we come back around.  So trim it now.
3764                          */
3765                         if (max_discard_size &&
3766                             bytes >= (max_discard_size +
3767                                       BTRFS_ASYNC_DISCARD_MIN_FILTER)) {
3768                                 bytes = max_discard_size;
3769                                 extent_bytes = max_discard_size;
3770                                 entry->offset += max_discard_size;
3771                                 entry->bytes -= max_discard_size;
3772                                 link_free_space(ctl, entry);
3773                         } else {
3774                                 kmem_cache_free(btrfs_free_space_cachep, entry);
3775                         }
3776                 } else {
3777                         start = max(start, extent_start);
3778                         bytes = min(extent_start + extent_bytes, end) - start;
3779                         if (bytes < minlen) {
3780                                 spin_unlock(&ctl->tree_lock);
3781                                 mutex_unlock(&ctl->cache_writeout_mutex);
3782                                 goto next;
3783                         }
3784
3785                         unlink_free_space(ctl, entry, true);
3786                         kmem_cache_free(btrfs_free_space_cachep, entry);
3787                 }
3788
3789                 spin_unlock(&ctl->tree_lock);
3790                 trim_entry.start = extent_start;
3791                 trim_entry.bytes = extent_bytes;
3792                 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3793                 mutex_unlock(&ctl->cache_writeout_mutex);
3794
3795                 ret = do_trimming(block_group, total_trimmed, start, bytes,
3796                                   extent_start, extent_bytes, extent_trim_state,
3797                                   &trim_entry);
3798                 if (ret) {
3799                         block_group->discard_cursor = start + bytes;
3800                         break;
3801                 }
3802 next:
3803                 start += bytes;
3804                 block_group->discard_cursor = start;
3805                 if (async && *total_trimmed)
3806                         break;
3807
3808                 if (fatal_signal_pending(current)) {
3809                         ret = -ERESTARTSYS;
3810                         break;
3811                 }
3812
3813                 cond_resched();
3814         }
3815
3816         return ret;
3817
3818 out_unlock:
3819         block_group->discard_cursor = btrfs_block_group_end(block_group);
3820         spin_unlock(&ctl->tree_lock);
3821         mutex_unlock(&ctl->cache_writeout_mutex);
3822
3823         return ret;
3824 }
3825
3826 /*
3827  * If we break out of trimming a bitmap prematurely, we should reset the
3828  * trimming bit.  In a rather contrieved case, it's possible to race here so
3829  * reset the state to BTRFS_TRIM_STATE_UNTRIMMED.
3830  *
3831  * start = start of bitmap
3832  * end = near end of bitmap
3833  *
3834  * Thread 1:                    Thread 2:
3835  * trim_bitmaps(start)
3836  *                              trim_bitmaps(end)
3837  *                              end_trimming_bitmap()
3838  * reset_trimming_bitmap()
3839  */
3840 static void reset_trimming_bitmap(struct btrfs_free_space_ctl *ctl, u64 offset)
3841 {
3842         struct btrfs_free_space *entry;
3843
3844         spin_lock(&ctl->tree_lock);
3845         entry = tree_search_offset(ctl, offset, 1, 0);
3846         if (entry) {
3847                 if (btrfs_free_space_trimmed(entry)) {
3848                         ctl->discardable_extents[BTRFS_STAT_CURR] +=
3849                                 entry->bitmap_extents;
3850                         ctl->discardable_bytes[BTRFS_STAT_CURR] += entry->bytes;
3851                 }
3852                 entry->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3853         }
3854
3855         spin_unlock(&ctl->tree_lock);
3856 }
3857
3858 static void end_trimming_bitmap(struct btrfs_free_space_ctl *ctl,
3859                                 struct btrfs_free_space *entry)
3860 {
3861         if (btrfs_free_space_trimming_bitmap(entry)) {
3862                 entry->trim_state = BTRFS_TRIM_STATE_TRIMMED;
3863                 ctl->discardable_extents[BTRFS_STAT_CURR] -=
3864                         entry->bitmap_extents;
3865                 ctl->discardable_bytes[BTRFS_STAT_CURR] -= entry->bytes;
3866         }
3867 }
3868
3869 /*
3870  * If @async is set, then we will trim 1 region and return.
3871  */
3872 static int trim_bitmaps(struct btrfs_block_group *block_group,
3873                         u64 *total_trimmed, u64 start, u64 end, u64 minlen,
3874                         u64 maxlen, bool async)
3875 {
3876         struct btrfs_discard_ctl *discard_ctl =
3877                                         &block_group->fs_info->discard_ctl;
3878         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3879         struct btrfs_free_space *entry;
3880         int ret = 0;
3881         int ret2;
3882         u64 bytes;
3883         u64 offset = offset_to_bitmap(ctl, start);
3884         const u64 max_discard_size = READ_ONCE(discard_ctl->max_discard_size);
3885
3886         while (offset < end) {
3887                 bool next_bitmap = false;
3888                 struct btrfs_trim_range trim_entry;
3889
3890                 mutex_lock(&ctl->cache_writeout_mutex);
3891                 spin_lock(&ctl->tree_lock);
3892
3893                 if (ctl->free_space < minlen) {
3894                         block_group->discard_cursor =
3895                                 btrfs_block_group_end(block_group);
3896                         spin_unlock(&ctl->tree_lock);
3897                         mutex_unlock(&ctl->cache_writeout_mutex);
3898                         break;
3899                 }
3900
3901                 entry = tree_search_offset(ctl, offset, 1, 0);
3902                 /*
3903                  * Bitmaps are marked trimmed lossily now to prevent constant
3904                  * discarding of the same bitmap (the reason why we are bound
3905                  * by the filters).  So, retrim the block group bitmaps when we
3906                  * are preparing to punt to the unused_bgs list.  This uses
3907                  * @minlen to determine if we are in BTRFS_DISCARD_INDEX_UNUSED
3908                  * which is the only discard index which sets minlen to 0.
3909                  */
3910                 if (!entry || (async && minlen && start == offset &&
3911                                btrfs_free_space_trimmed(entry))) {
3912                         spin_unlock(&ctl->tree_lock);
3913                         mutex_unlock(&ctl->cache_writeout_mutex);
3914                         next_bitmap = true;
3915                         goto next;
3916                 }
3917
3918                 /*
3919                  * Async discard bitmap trimming begins at by setting the start
3920                  * to be key.objectid and the offset_to_bitmap() aligns to the
3921                  * start of the bitmap.  This lets us know we are fully
3922                  * scanning the bitmap rather than only some portion of it.
3923                  */
3924                 if (start == offset)
3925                         entry->trim_state = BTRFS_TRIM_STATE_TRIMMING;
3926
3927                 bytes = minlen;
3928                 ret2 = search_bitmap(ctl, entry, &start, &bytes, false);
3929                 if (ret2 || start >= end) {
3930                         /*
3931                          * We lossily consider a bitmap trimmed if we only skip
3932                          * over regions <= BTRFS_ASYNC_DISCARD_MIN_FILTER.
3933                          */
3934                         if (ret2 && minlen <= BTRFS_ASYNC_DISCARD_MIN_FILTER)
3935                                 end_trimming_bitmap(ctl, entry);
3936                         else
3937                                 entry->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3938                         spin_unlock(&ctl->tree_lock);
3939                         mutex_unlock(&ctl->cache_writeout_mutex);
3940                         next_bitmap = true;
3941                         goto next;
3942                 }
3943
3944                 /*
3945                  * We already trimmed a region, but are using the locking above
3946                  * to reset the trim_state.
3947                  */
3948                 if (async && *total_trimmed) {
3949                         spin_unlock(&ctl->tree_lock);
3950                         mutex_unlock(&ctl->cache_writeout_mutex);
3951                         goto out;
3952                 }
3953
3954                 bytes = min(bytes, end - start);
3955                 if (bytes < minlen || (async && maxlen && bytes > maxlen)) {
3956                         spin_unlock(&ctl->tree_lock);
3957                         mutex_unlock(&ctl->cache_writeout_mutex);
3958                         goto next;
3959                 }
3960
3961                 /*
3962                  * Let bytes = BTRFS_MAX_DISCARD_SIZE + X.
3963                  * If X < @minlen, we won't trim X when we come back around.
3964                  * So trim it now.  We differ here from trimming extents as we
3965                  * don't keep individual state per bit.
3966                  */
3967                 if (async &&
3968                     max_discard_size &&
3969                     bytes > (max_discard_size + minlen))
3970                         bytes = max_discard_size;
3971
3972                 bitmap_clear_bits(ctl, entry, start, bytes, true);
3973                 if (entry->bytes == 0)
3974                         free_bitmap(ctl, entry);
3975
3976                 spin_unlock(&ctl->tree_lock);
3977                 trim_entry.start = start;
3978                 trim_entry.bytes = bytes;
3979                 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3980                 mutex_unlock(&ctl->cache_writeout_mutex);
3981
3982                 ret = do_trimming(block_group, total_trimmed, start, bytes,
3983                                   start, bytes, 0, &trim_entry);
3984                 if (ret) {
3985                         reset_trimming_bitmap(ctl, offset);
3986                         block_group->discard_cursor =
3987                                 btrfs_block_group_end(block_group);
3988                         break;
3989                 }
3990 next:
3991                 if (next_bitmap) {
3992                         offset += BITS_PER_BITMAP * ctl->unit;
3993                         start = offset;
3994                 } else {
3995                         start += bytes;
3996                 }
3997                 block_group->discard_cursor = start;
3998
3999                 if (fatal_signal_pending(current)) {
4000                         if (start != offset)
4001                                 reset_trimming_bitmap(ctl, offset);
4002                         ret = -ERESTARTSYS;
4003                         break;
4004                 }
4005
4006                 cond_resched();
4007         }
4008
4009         if (offset >= end)
4010                 block_group->discard_cursor = end;
4011
4012 out:
4013         return ret;
4014 }
4015
4016 int btrfs_trim_block_group(struct btrfs_block_group *block_group,
4017                            u64 *trimmed, u64 start, u64 end, u64 minlen)
4018 {
4019         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
4020         int ret;
4021         u64 rem = 0;
4022
4023         ASSERT(!btrfs_is_zoned(block_group->fs_info));
4024
4025         *trimmed = 0;
4026
4027         spin_lock(&block_group->lock);
4028         if (test_bit(BLOCK_GROUP_FLAG_REMOVED, &block_group->runtime_flags)) {
4029                 spin_unlock(&block_group->lock);
4030                 return 0;
4031         }
4032         btrfs_freeze_block_group(block_group);
4033         spin_unlock(&block_group->lock);
4034
4035         ret = trim_no_bitmap(block_group, trimmed, start, end, minlen, false);
4036         if (ret)
4037                 goto out;
4038
4039         ret = trim_bitmaps(block_group, trimmed, start, end, minlen, 0, false);
4040         div64_u64_rem(end, BITS_PER_BITMAP * ctl->unit, &rem);
4041         /* If we ended in the middle of a bitmap, reset the trimming flag */
4042         if (rem)
4043                 reset_trimming_bitmap(ctl, offset_to_bitmap(ctl, end));
4044 out:
4045         btrfs_unfreeze_block_group(block_group);
4046         return ret;
4047 }
4048
4049 int btrfs_trim_block_group_extents(struct btrfs_block_group *block_group,
4050                                    u64 *trimmed, u64 start, u64 end, u64 minlen,
4051                                    bool async)
4052 {
4053         int ret;
4054
4055         *trimmed = 0;
4056
4057         spin_lock(&block_group->lock);
4058         if (test_bit(BLOCK_GROUP_FLAG_REMOVED, &block_group->runtime_flags)) {
4059                 spin_unlock(&block_group->lock);
4060                 return 0;
4061         }
4062         btrfs_freeze_block_group(block_group);
4063         spin_unlock(&block_group->lock);
4064
4065         ret = trim_no_bitmap(block_group, trimmed, start, end, minlen, async);
4066         btrfs_unfreeze_block_group(block_group);
4067
4068         return ret;
4069 }
4070
4071 int btrfs_trim_block_group_bitmaps(struct btrfs_block_group *block_group,
4072                                    u64 *trimmed, u64 start, u64 end, u64 minlen,
4073                                    u64 maxlen, bool async)
4074 {
4075         int ret;
4076
4077         *trimmed = 0;
4078
4079         spin_lock(&block_group->lock);
4080         if (test_bit(BLOCK_GROUP_FLAG_REMOVED, &block_group->runtime_flags)) {
4081                 spin_unlock(&block_group->lock);
4082                 return 0;
4083         }
4084         btrfs_freeze_block_group(block_group);
4085         spin_unlock(&block_group->lock);
4086
4087         ret = trim_bitmaps(block_group, trimmed, start, end, minlen, maxlen,
4088                            async);
4089
4090         btrfs_unfreeze_block_group(block_group);
4091
4092         return ret;
4093 }
4094
4095 bool btrfs_free_space_cache_v1_active(struct btrfs_fs_info *fs_info)
4096 {
4097         return btrfs_super_cache_generation(fs_info->super_copy);
4098 }
4099
4100 static int cleanup_free_space_cache_v1(struct btrfs_fs_info *fs_info,
4101                                        struct btrfs_trans_handle *trans)
4102 {
4103         struct btrfs_block_group *block_group;
4104         struct rb_node *node;
4105         int ret = 0;
4106
4107         btrfs_info(fs_info, "cleaning free space cache v1");
4108
4109         node = rb_first_cached(&fs_info->block_group_cache_tree);
4110         while (node) {
4111                 block_group = rb_entry(node, struct btrfs_block_group, cache_node);
4112                 ret = btrfs_remove_free_space_inode(trans, NULL, block_group);
4113                 if (ret)
4114                         goto out;
4115                 node = rb_next(node);
4116         }
4117 out:
4118         return ret;
4119 }
4120
4121 int btrfs_set_free_space_cache_v1_active(struct btrfs_fs_info *fs_info, bool active)
4122 {
4123         struct btrfs_trans_handle *trans;
4124         int ret;
4125
4126         /*
4127          * update_super_roots will appropriately set or unset
4128          * super_copy->cache_generation based on SPACE_CACHE and
4129          * BTRFS_FS_CLEANUP_SPACE_CACHE_V1. For this reason, we need a
4130          * transaction commit whether we are enabling space cache v1 and don't
4131          * have any other work to do, or are disabling it and removing free
4132          * space inodes.
4133          */
4134         trans = btrfs_start_transaction(fs_info->tree_root, 0);
4135         if (IS_ERR(trans))
4136                 return PTR_ERR(trans);
4137
4138         if (!active) {
4139                 set_bit(BTRFS_FS_CLEANUP_SPACE_CACHE_V1, &fs_info->flags);
4140                 ret = cleanup_free_space_cache_v1(fs_info, trans);
4141                 if (ret) {
4142                         btrfs_abort_transaction(trans, ret);
4143                         btrfs_end_transaction(trans);
4144                         goto out;
4145                 }
4146         }
4147
4148         ret = btrfs_commit_transaction(trans);
4149 out:
4150         clear_bit(BTRFS_FS_CLEANUP_SPACE_CACHE_V1, &fs_info->flags);
4151
4152         return ret;
4153 }
4154
4155 int __init btrfs_free_space_init(void)
4156 {
4157         btrfs_free_space_cachep = KMEM_CACHE(btrfs_free_space, 0);
4158         if (!btrfs_free_space_cachep)
4159                 return -ENOMEM;
4160
4161         btrfs_free_space_bitmap_cachep = kmem_cache_create("btrfs_free_space_bitmap",
4162                                                         PAGE_SIZE, PAGE_SIZE,
4163                                                         0, NULL);
4164         if (!btrfs_free_space_bitmap_cachep) {
4165                 kmem_cache_destroy(btrfs_free_space_cachep);
4166                 return -ENOMEM;
4167         }
4168
4169         return 0;
4170 }
4171
4172 void __cold btrfs_free_space_exit(void)
4173 {
4174         kmem_cache_destroy(btrfs_free_space_cachep);
4175         kmem_cache_destroy(btrfs_free_space_bitmap_cachep);
4176 }
4177
4178 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4179 /*
4180  * Use this if you need to make a bitmap or extent entry specifically, it
4181  * doesn't do any of the merging that add_free_space does, this acts a lot like
4182  * how the free space cache loading stuff works, so you can get really weird
4183  * configurations.
4184  */
4185 int test_add_free_space_entry(struct btrfs_block_group *cache,
4186                               u64 offset, u64 bytes, bool bitmap)
4187 {
4188         struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
4189         struct btrfs_free_space *info = NULL, *bitmap_info;
4190         void *map = NULL;
4191         enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_TRIMMED;
4192         u64 bytes_added;
4193         int ret;
4194
4195 again:
4196         if (!info) {
4197                 info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
4198                 if (!info)
4199                         return -ENOMEM;
4200         }
4201
4202         if (!bitmap) {
4203                 spin_lock(&ctl->tree_lock);
4204                 info->offset = offset;
4205                 info->bytes = bytes;
4206                 info->max_extent_size = 0;
4207                 ret = link_free_space(ctl, info);
4208                 spin_unlock(&ctl->tree_lock);
4209                 if (ret)
4210                         kmem_cache_free(btrfs_free_space_cachep, info);
4211                 return ret;
4212         }
4213
4214         if (!map) {
4215                 map = kmem_cache_zalloc(btrfs_free_space_bitmap_cachep, GFP_NOFS);
4216                 if (!map) {
4217                         kmem_cache_free(btrfs_free_space_cachep, info);
4218                         return -ENOMEM;
4219                 }
4220         }
4221
4222         spin_lock(&ctl->tree_lock);
4223         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
4224                                          1, 0);
4225         if (!bitmap_info) {
4226                 info->bitmap = map;
4227                 map = NULL;
4228                 add_new_bitmap(ctl, info, offset);
4229                 bitmap_info = info;
4230                 info = NULL;
4231         }
4232
4233         bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes,
4234                                           trim_state);
4235
4236         bytes -= bytes_added;
4237         offset += bytes_added;
4238         spin_unlock(&ctl->tree_lock);
4239
4240         if (bytes)
4241                 goto again;
4242
4243         if (info)
4244                 kmem_cache_free(btrfs_free_space_cachep, info);
4245         if (map)
4246                 kmem_cache_free(btrfs_free_space_bitmap_cachep, map);
4247         return 0;
4248 }
4249
4250 /*
4251  * Checks to see if the given range is in the free space cache.  This is really
4252  * just used to check the absence of space, so if there is free space in the
4253  * range at all we will return 1.
4254  */
4255 int test_check_exists(struct btrfs_block_group *cache,
4256                       u64 offset, u64 bytes)
4257 {
4258         struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
4259         struct btrfs_free_space *info;
4260         int ret = 0;
4261
4262         spin_lock(&ctl->tree_lock);
4263         info = tree_search_offset(ctl, offset, 0, 0);
4264         if (!info) {
4265                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
4266                                           1, 0);
4267                 if (!info)
4268                         goto out;
4269         }
4270
4271 have_info:
4272         if (info->bitmap) {
4273                 u64 bit_off, bit_bytes;
4274                 struct rb_node *n;
4275                 struct btrfs_free_space *tmp;
4276
4277                 bit_off = offset;
4278                 bit_bytes = ctl->unit;
4279                 ret = search_bitmap(ctl, info, &bit_off, &bit_bytes, false);
4280                 if (!ret) {
4281                         if (bit_off == offset) {
4282                                 ret = 1;
4283                                 goto out;
4284                         } else if (bit_off > offset &&
4285                                    offset + bytes > bit_off) {
4286                                 ret = 1;
4287                                 goto out;
4288                         }
4289                 }
4290
4291                 n = rb_prev(&info->offset_index);
4292                 while (n) {
4293                         tmp = rb_entry(n, struct btrfs_free_space,
4294                                        offset_index);
4295                         if (tmp->offset + tmp->bytes < offset)
4296                                 break;
4297                         if (offset + bytes < tmp->offset) {
4298                                 n = rb_prev(&tmp->offset_index);
4299                                 continue;
4300                         }
4301                         info = tmp;
4302                         goto have_info;
4303                 }
4304
4305                 n = rb_next(&info->offset_index);
4306                 while (n) {
4307                         tmp = rb_entry(n, struct btrfs_free_space,
4308                                        offset_index);
4309                         if (offset + bytes < tmp->offset)
4310                                 break;
4311                         if (tmp->offset + tmp->bytes < offset) {
4312                                 n = rb_next(&tmp->offset_index);
4313                                 continue;
4314                         }
4315                         info = tmp;
4316                         goto have_info;
4317                 }
4318
4319                 ret = 0;
4320                 goto out;
4321         }
4322
4323         if (info->offset == offset) {
4324                 ret = 1;
4325                 goto out;
4326         }
4327
4328         if (offset > info->offset && offset < info->offset + info->bytes)
4329                 ret = 1;
4330 out:
4331         spin_unlock(&ctl->tree_lock);
4332         return ret;
4333 }
4334 #endif /* CONFIG_BTRFS_FS_RUN_SANITY_TESTS */