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 "misc.h"
  15 #include "ctree.h"
  16 #include "free-space-cache.h"
  17 #include "transaction.h"
  18 #include "disk-io.h"
  19 #include "extent_io.h"
  20 #include "volumes.h"
  21 #include "space-info.h"
  22 #include "delalloc-space.h"
  23 #include "block-group.h"
  24 #include "discard.h"
  25 #include "subpage.h"
  26 #include "inode-item.h"
  27
  28 #define BITS_PER_BITMAP         (PAGE_SIZE * 8UL)
  29 #define MAX_CACHE_BYTES_PER_GIG SZ_64K
  30 #define FORCE_EXTENT_THRESHOLD  SZ_1M
  31
  32 struct btrfs_trim_range {
  33         u64 start;
  34         u64 bytes;
  35         struct list_head list;
  36 };
  37
  38 static int link_free_space(struct btrfs_free_space_ctl *ctl,
  39                            struct btrfs_free_space *info);
  40 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
  41                               struct btrfs_free_space *info, bool update_stat);
  42 static int search_bitmap(struct btrfs_free_space_ctl *ctl,
  43                          struct btrfs_free_space *bitmap_info, u64 *offset,
  44                          u64 *bytes, bool for_alloc);
  45 static void free_bitmap(struct btrfs_free_space_ctl *ctl,
  46                         struct btrfs_free_space *bitmap_info);
  47 static void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
  48                               struct btrfs_free_space *info, u64 offset,
  49                               u64 bytes, bool update_stats);
  50
  51 static void __btrfs_remove_free_space_cache(struct btrfs_free_space_ctl *ctl)
  52 {
  53         struct btrfs_free_space *info;
  54         struct rb_node *node;
  55
  56         while ((node = rb_last(&ctl->free_space_offset)) != NULL) {
  57                 info = rb_entry(node, struct btrfs_free_space, offset_index);
  58                 if (!info->bitmap) {
  59                         unlink_free_space(ctl, info, true);
  60                         kmem_cache_free(btrfs_free_space_cachep, info);
  61                 } else {
  62                         free_bitmap(ctl, info);
  63                 }
  64
  65                 cond_resched_lock(&ctl->tree_lock);
  66         }
  67 }
  68
  69 static struct inode *__lookup_free_space_inode(struct btrfs_root *root,
  70                                                struct btrfs_path *path,
  71                                                u64 offset)
  72 {
  73         struct btrfs_fs_info *fs_info = root->fs_info;
  74         struct btrfs_key key;
  75         struct btrfs_key location;
  76         struct btrfs_disk_key disk_key;
  77         struct btrfs_free_space_header *header;
  78         struct extent_buffer *leaf;
  79         struct inode *inode = NULL;
  80         unsigned nofs_flag;
  81         int ret;
  82
  83         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
  84         key.offset = offset;
  85         key.type = 0;
  86
  87         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
  88         if (ret < 0)
  89                 return ERR_PTR(ret);
  90         if (ret > 0) {
  91                 btrfs_release_path(path);
  92                 return ERR_PTR(-ENOENT);
  93         }
  94
  95         leaf = path->nodes[0];
  96         header = btrfs_item_ptr(leaf, path->slots[0],
  97                                 struct btrfs_free_space_header);
  98         btrfs_free_space_key(leaf, header, &disk_key);
  99         btrfs_disk_key_to_cpu(&location, &disk_key);
 100         btrfs_release_path(path);
 101
 102         /*
 103          * We are often under a trans handle at this point, so we need to make
 104          * sure NOFS is set to keep us from deadlocking.
 105          */
 106         nofs_flag = memalloc_nofs_save();
 107         inode = btrfs_iget_path(fs_info->sb, location.objectid, root, path);
 108         btrfs_release_path(path);
 109         memalloc_nofs_restore(nofs_flag);
 110         if (IS_ERR(inode))
 111                 return inode;
 112
 113         mapping_set_gfp_mask(inode->i_mapping,
 114                         mapping_gfp_constraint(inode->i_mapping,
 115                         ~(__GFP_FS | __GFP_HIGHMEM)));
 116
 117         return inode;
 118 }
 119
 120 struct inode *lookup_free_space_inode(struct btrfs_block_group *block_group,
 121                 struct btrfs_path *path)
 122 {
 123         struct btrfs_fs_info *fs_info = block_group->fs_info;
 124         struct inode *inode = NULL;
 125         u32 flags = BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
 126
 127         spin_lock(&block_group->lock);
 128         if (block_group->inode)
 129                 inode = igrab(block_group->inode);
 130         spin_unlock(&block_group->lock);
 131         if (inode)
 132                 return inode;
 133
 134         inode = __lookup_free_space_inode(fs_info->tree_root, path,
 135                                           block_group->start);
 136         if (IS_ERR(inode))
 137                 return inode;
 138
 139         spin_lock(&block_group->lock);
 140         if (!((BTRFS_I(inode)->flags & flags) == flags)) {
 141                 btrfs_info(fs_info, "Old style space inode found, converting.");
 142                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM |
 143                         BTRFS_INODE_NODATACOW;
 144                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 145         }
 146
 147         if (!test_and_set_bit(BLOCK_GROUP_FLAG_IREF, &block_group->runtime_flags))
 148                 block_group->inode = igrab(inode);
 149         spin_unlock(&block_group->lock);
 150
 151         return inode;
 152 }
 153
 154 static int __create_free_space_inode(struct btrfs_root *root,
 155                                      struct btrfs_trans_handle *trans,
 156                                      struct btrfs_path *path,
 157                                      u64 ino, u64 offset)
 158 {
 159         struct btrfs_key key;
 160         struct btrfs_disk_key disk_key;
 161         struct btrfs_free_space_header *header;
 162         struct btrfs_inode_item *inode_item;
 163         struct extent_buffer *leaf;
 164         /* We inline CRCs for the free disk space cache */
 165         const u64 flags = BTRFS_INODE_NOCOMPRESS | BTRFS_INODE_PREALLOC |
 166                           BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
 167         int ret;
 168
 169         ret = btrfs_insert_empty_inode(trans, root, path, ino);
 170         if (ret)
 171                 return ret;
 172
 173         leaf = path->nodes[0];
 174         inode_item = btrfs_item_ptr(leaf, path->slots[0],
 175                                     struct btrfs_inode_item);
 176         btrfs_item_key(leaf, &disk_key, path->slots[0]);
 177         memzero_extent_buffer(leaf, (unsigned long)inode_item,
 178                              sizeof(*inode_item));
 179         btrfs_set_inode_generation(leaf, inode_item, trans->transid);
 180         btrfs_set_inode_size(leaf, inode_item, 0);
 181         btrfs_set_inode_nbytes(leaf, inode_item, 0);
 182         btrfs_set_inode_uid(leaf, inode_item, 0);
 183         btrfs_set_inode_gid(leaf, inode_item, 0);
 184         btrfs_set_inode_mode(leaf, inode_item, S_IFREG | 0600);
 185         btrfs_set_inode_flags(leaf, inode_item, flags);
 186         btrfs_set_inode_nlink(leaf, inode_item, 1);
 187         btrfs_set_inode_transid(leaf, inode_item, trans->transid);
 188         btrfs_set_inode_block_group(leaf, inode_item, offset);
 189         btrfs_mark_buffer_dirty(leaf);
 190         btrfs_release_path(path);
 191
 192         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 193         key.offset = offset;
 194         key.type = 0;
 195         ret = btrfs_insert_empty_item(trans, root, path, &key,
 196                                       sizeof(struct btrfs_free_space_header));
 197         if (ret < 0) {
 198                 btrfs_release_path(path);
 199                 return ret;
 200         }
 201
 202         leaf = path->nodes[0];
 203         header = btrfs_item_ptr(leaf, path->slots[0],
 204                                 struct btrfs_free_space_header);
 205         memzero_extent_buffer(leaf, (unsigned long)header, sizeof(*header));
 206         btrfs_set_free_space_key(leaf, header, &disk_key);
 207         btrfs_mark_buffer_dirty(leaf);
 208         btrfs_release_path(path);
 209
 210         return 0;
 211 }
 212
 213 int create_free_space_inode(struct btrfs_trans_handle *trans,
 214                             struct btrfs_block_group *block_group,
 215                             struct btrfs_path *path)
 216 {
 217         int ret;
 218         u64 ino;
 219
 220         ret = btrfs_get_free_objectid(trans->fs_info->tree_root, &ino);
 221         if (ret < 0)
 222                 return ret;
 223
 224         return __create_free_space_inode(trans->fs_info->tree_root, trans, path,
 225                                          ino, block_group->start);
 226 }
 227
 228 /*
 229  * inode is an optional sink: if it is NULL, btrfs_remove_free_space_inode
 230  * handles lookup, otherwise it takes ownership and iputs the inode.
 231  * Don't reuse an inode pointer after passing it into this function.
 232  */
 233 int btrfs_remove_free_space_inode(struct btrfs_trans_handle *trans,
 234                                   struct inode *inode,
 235                                   struct btrfs_block_group *block_group)
 236 {
 237         struct btrfs_path *path;
 238         struct btrfs_key key;
 239         int ret = 0;
 240
 241         path = btrfs_alloc_path();
 242         if (!path)
 243                 return -ENOMEM;
 244
 245         if (!inode)
 246                 inode = lookup_free_space_inode(block_group, path);
 247         if (IS_ERR(inode)) {
 248                 if (PTR_ERR(inode) != -ENOENT)
 249                         ret = PTR_ERR(inode);
 250                 goto out;
 251         }
 252         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
 253         if (ret) {
 254                 btrfs_add_delayed_iput(inode);
 255                 goto out;
 256         }
 257         clear_nlink(inode);
 258         /* One for the block groups ref */
 259         spin_lock(&block_group->lock);
 260         if (test_and_clear_bit(BLOCK_GROUP_FLAG_IREF, &block_group->runtime_flags)) {
 261                 block_group->inode = NULL;
 262                 spin_unlock(&block_group->lock);
 263                 iput(inode);
 264         } else {
 265                 spin_unlock(&block_group->lock);
 266         }
 267         /* One for the lookup ref */
 268         btrfs_add_delayed_iput(inode);
 269
 270         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 271         key.type = 0;
 272         key.offset = block_group->start;
 273         ret = btrfs_search_slot(trans, trans->fs_info->tree_root, &key, path,
 274                                 -1, 1);
 275         if (ret) {
 276                 if (ret > 0)
 277                         ret = 0;
 278                 goto out;
 279         }
 280         ret = btrfs_del_item(trans, trans->fs_info->tree_root, path);
 281 out:
 282         btrfs_free_path(path);
 283         return ret;
 284 }
 285
 286 int btrfs_check_trunc_cache_free_space(struct btrfs_fs_info *fs_info,
 287                                        struct btrfs_block_rsv *rsv)
 288 {
 289         u64 needed_bytes;
 290         int ret;
 291
 292         /* 1 for slack space, 1 for updating the inode */
 293         needed_bytes = btrfs_calc_insert_metadata_size(fs_info, 1) +
 294                 btrfs_calc_metadata_size(fs_info, 1);
 295
 296         spin_lock(&rsv->lock);
 297         if (rsv->reserved < needed_bytes)
 298                 ret = -ENOSPC;
 299         else
 300                 ret = 0;
 301         spin_unlock(&rsv->lock);
 302         return ret;
 303 }
 304
 305 int btrfs_truncate_free_space_cache(struct btrfs_trans_handle *trans,
 306                                     struct btrfs_block_group *block_group,
 307                                     struct inode *vfs_inode)
 308 {
 309         struct btrfs_truncate_control control = {
 310                 .inode = BTRFS_I(vfs_inode),
 311                 .new_size = 0,
 312                 .ino = btrfs_ino(BTRFS_I(vfs_inode)),
 313                 .min_type = BTRFS_EXTENT_DATA_KEY,
 314                 .clear_extent_range = true,
 315         };
 316         struct btrfs_inode *inode = BTRFS_I(vfs_inode);
 317         struct btrfs_root *root = inode->root;
 318         struct extent_state *cached_state = NULL;
 319         int ret = 0;
 320         bool locked = false;
 321
 322         if (block_group) {
 323                 struct btrfs_path *path = btrfs_alloc_path();
 324
 325                 if (!path) {
 326                         ret = -ENOMEM;
 327                         goto fail;
 328                 }
 329                 locked = true;
 330                 mutex_lock(&trans->transaction->cache_write_mutex);
 331                 if (!list_empty(&block_group->io_list)) {
 332                         list_del_init(&block_group->io_list);
 333
 334                         btrfs_wait_cache_io(trans, block_group, path);
 335                         btrfs_put_block_group(block_group);
 336                 }
 337
 338                 /*
 339                  * now that we've truncated the cache away, its no longer
 340                  * setup or written
 341                  */
 342                 spin_lock(&block_group->lock);
 343                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 344                 spin_unlock(&block_group->lock);
 345                 btrfs_free_path(path);
 346         }
 347
 348         btrfs_i_size_write(inode, 0);
 349         truncate_pagecache(vfs_inode, 0);
 350
 351         lock_extent(&inode->io_tree, 0, (u64)-1, &cached_state);
 352         btrfs_drop_extent_map_range(inode, 0, (u64)-1, false);
 353
 354         /*
 355          * We skip the throttling logic for free space cache inodes, so we don't
 356          * need to check for -EAGAIN.
 357          */
 358         ret = btrfs_truncate_inode_items(trans, root, &control);
 359
 360         inode_sub_bytes(&inode->vfs_inode, control.sub_bytes);
 361         btrfs_inode_safe_disk_i_size_write(inode, control.last_size);
 362
 363         unlock_extent(&inode->io_tree, 0, (u64)-1, &cached_state);
 364         if (ret)
 365                 goto fail;
 366
 367         ret = btrfs_update_inode(trans, root, inode);
 368
 369 fail:
 370         if (locked)
 371                 mutex_unlock(&trans->transaction->cache_write_mutex);
 372         if (ret)
 373                 btrfs_abort_transaction(trans, ret);
 374
 375         return ret;
 376 }
 377
 378 static void readahead_cache(struct inode *inode)
 379 {
 380         struct file_ra_state ra;
 381         unsigned long last_index;
 382
 383         file_ra_state_init(&ra, inode->i_mapping);
 384         last_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
 385
 386         page_cache_sync_readahead(inode->i_mapping, &ra, NULL, 0, last_index);
 387 }
 388
 389 static int io_ctl_init(struct btrfs_io_ctl *io_ctl, struct inode *inode,
 390                        int write)
 391 {
 392         int num_pages;
 393
 394         num_pages = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
 395
 396         /* Make sure we can fit our crcs and generation into the first page */
 397         if (write && (num_pages * sizeof(u32) + sizeof(u64)) > PAGE_SIZE)
 398                 return -ENOSPC;
 399
 400         memset(io_ctl, 0, sizeof(struct btrfs_io_ctl));
 401
 402         io_ctl->pages = kcalloc(num_pages, sizeof(struct page *), GFP_NOFS);
 403         if (!io_ctl->pages)
 404                 return -ENOMEM;
 405
 406         io_ctl->num_pages = num_pages;
 407         io_ctl->fs_info = btrfs_sb(inode->i_sb);
 408         io_ctl->inode = inode;
 409
 410         return 0;
 411 }
 412 ALLOW_ERROR_INJECTION(io_ctl_init, ERRNO);
 413
 414 static void io_ctl_free(struct btrfs_io_ctl *io_ctl)
 415 {
 416         kfree(io_ctl->pages);
 417         io_ctl->pages = NULL;
 418 }
 419
 420 static void io_ctl_unmap_page(struct btrfs_io_ctl *io_ctl)
 421 {
 422         if (io_ctl->cur) {
 423                 io_ctl->cur = NULL;
 424                 io_ctl->orig = NULL;
 425         }
 426 }
 427
 428 static void io_ctl_map_page(struct btrfs_io_ctl *io_ctl, int clear)
 429 {
 430         ASSERT(io_ctl->index < io_ctl->num_pages);
 431         io_ctl->page = io_ctl->pages[io_ctl->index++];
 432         io_ctl->cur = page_address(io_ctl->page);
 433         io_ctl->orig = io_ctl->cur;
 434         io_ctl->size = PAGE_SIZE;
 435         if (clear)
 436                 clear_page(io_ctl->cur);
 437 }
 438
 439 static void io_ctl_drop_pages(struct btrfs_io_ctl *io_ctl)
 440 {
 441         int i;
 442
 443         io_ctl_unmap_page(io_ctl);
 444
 445         for (i = 0; i < io_ctl->num_pages; i++) {
 446                 if (io_ctl->pages[i]) {
 447                         btrfs_page_clear_checked(io_ctl->fs_info,
 448                                         io_ctl->pages[i],
 449                                         page_offset(io_ctl->pages[i]),
 450                                         PAGE_SIZE);
 451                         unlock_page(io_ctl->pages[i]);
 452                         put_page(io_ctl->pages[i]);
 453                 }
 454         }
 455 }
 456
 457 static int io_ctl_prepare_pages(struct btrfs_io_ctl *io_ctl, bool uptodate)
 458 {
 459         struct page *page;
 460         struct inode *inode = io_ctl->inode;
 461         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
 462         int i;
 463
 464         for (i = 0; i < io_ctl->num_pages; i++) {
 465                 int ret;
 466
 467                 page = find_or_create_page(inode->i_mapping, i, mask);
 468                 if (!page) {
 469                         io_ctl_drop_pages(io_ctl);
 470                         return -ENOMEM;
 471                 }
 472
 473                 ret = set_page_extent_mapped(page);
 474                 if (ret < 0) {
 475                         unlock_page(page);
 476                         put_page(page);
 477                         io_ctl_drop_pages(io_ctl);
 478                         return ret;
 479                 }
 480
 481                 io_ctl->pages[i] = page;
 482                 if (uptodate && !PageUptodate(page)) {
 483                         btrfs_read_folio(NULL, page_folio(page));
 484                         lock_page(page);
 485                         if (page->mapping != inode->i_mapping) {
 486                                 btrfs_err(BTRFS_I(inode)->root->fs_info,
 487                                           "free space cache page truncated");
 488                                 io_ctl_drop_pages(io_ctl);
 489                                 return -EIO;
 490                         }
 491                         if (!PageUptodate(page)) {
 492                                 btrfs_err(BTRFS_I(inode)->root->fs_info,
 493                                            "error reading free space cache");
 494                                 io_ctl_drop_pages(io_ctl);
 495                                 return -EIO;
 496                         }
 497                 }
 498         }
 499
 500         for (i = 0; i < io_ctl->num_pages; i++)
 501                 clear_page_dirty_for_io(io_ctl->pages[i]);
 502
 503         return 0;
 504 }
 505
 506 static void io_ctl_set_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
 507 {
 508         io_ctl_map_page(io_ctl, 1);
 509
 510         /*
 511          * Skip the csum areas.  If we don't check crcs then we just have a
 512          * 64bit chunk at the front of the first page.
 513          */
 514         io_ctl->cur += (sizeof(u32) * io_ctl->num_pages);
 515         io_ctl->size -= sizeof(u64) + (sizeof(u32) * io_ctl->num_pages);
 516
 517         put_unaligned_le64(generation, io_ctl->cur);
 518         io_ctl->cur += sizeof(u64);
 519 }
 520
 521 static int io_ctl_check_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
 522 {
 523         u64 cache_gen;
 524
 525         /*
 526          * Skip the crc area.  If we don't check crcs then we just have a 64bit
 527          * chunk at the front of the first page.
 528          */
 529         io_ctl->cur += sizeof(u32) * io_ctl->num_pages;
 530         io_ctl->size -= sizeof(u64) + (sizeof(u32) * io_ctl->num_pages);
 531
 532         cache_gen = get_unaligned_le64(io_ctl->cur);
 533         if (cache_gen != generation) {
 534                 btrfs_err_rl(io_ctl->fs_info,
 535                         "space cache generation (%llu) does not match inode (%llu)",
 536                                 cache_gen, generation);
 537                 io_ctl_unmap_page(io_ctl);
 538                 return -EIO;
 539         }
 540         io_ctl->cur += sizeof(u64);
 541         return 0;
 542 }
 543
 544 static void io_ctl_set_crc(struct btrfs_io_ctl *io_ctl, int index)
 545 {
 546         u32 *tmp;
 547         u32 crc = ~(u32)0;
 548         unsigned offset = 0;
 549
 550         if (index == 0)
 551                 offset = sizeof(u32) * io_ctl->num_pages;
 552
 553         crc = btrfs_crc32c(crc, io_ctl->orig + offset, PAGE_SIZE - offset);
 554         btrfs_crc32c_final(crc, (u8 *)&crc);
 555         io_ctl_unmap_page(io_ctl);
 556         tmp = page_address(io_ctl->pages[0]);
 557         tmp += index;
 558         *tmp = crc;
 559 }
 560
 561 static int io_ctl_check_crc(struct btrfs_io_ctl *io_ctl, int index)
 562 {
 563         u32 *tmp, val;
 564         u32 crc = ~(u32)0;
 565         unsigned offset = 0;
 566
 567         if (index == 0)
 568                 offset = sizeof(u32) * io_ctl->num_pages;
 569
 570         tmp = page_address(io_ctl->pages[0]);
 571         tmp += index;
 572         val = *tmp;
 573
 574         io_ctl_map_page(io_ctl, 0);
 575         crc = btrfs_crc32c(crc, io_ctl->orig + offset, PAGE_SIZE - offset);
 576         btrfs_crc32c_final(crc, (u8 *)&crc);
 577         if (val != crc) {
 578                 btrfs_err_rl(io_ctl->fs_info,
 579                         "csum mismatch on free space cache");
 580                 io_ctl_unmap_page(io_ctl);
 581                 return -EIO;
 582         }
 583
 584         return 0;
 585 }
 586
 587 static int io_ctl_add_entry(struct btrfs_io_ctl *io_ctl, u64 offset, u64 bytes,
 588                             void *bitmap)
 589 {
 590         struct btrfs_free_space_entry *entry;
 591
 592         if (!io_ctl->cur)
 593                 return -ENOSPC;
 594
 595         entry = io_ctl->cur;
 596         put_unaligned_le64(offset, &entry->offset);
 597         put_unaligned_le64(bytes, &entry->bytes);
 598         entry->type = (bitmap) ? BTRFS_FREE_SPACE_BITMAP :
 599                 BTRFS_FREE_SPACE_EXTENT;
 600         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
 601         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
 602
 603         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
 604                 return 0;
 605
 606         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 607
 608         /* No more pages to map */
 609         if (io_ctl->index >= io_ctl->num_pages)
 610                 return 0;
 611
 612         /* map the next page */
 613         io_ctl_map_page(io_ctl, 1);
 614         return 0;
 615 }
 616
 617 static int io_ctl_add_bitmap(struct btrfs_io_ctl *io_ctl, void *bitmap)
 618 {
 619         if (!io_ctl->cur)
 620                 return -ENOSPC;
 621
 622         /*
 623          * If we aren't at the start of the current page, unmap this one and
 624          * map the next one if there is any left.
 625          */
 626         if (io_ctl->cur != io_ctl->orig) {
 627                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 628                 if (io_ctl->index >= io_ctl->num_pages)
 629                         return -ENOSPC;
 630                 io_ctl_map_page(io_ctl, 0);
 631         }
 632
 633         copy_page(io_ctl->cur, bitmap);
 634         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 635         if (io_ctl->index < io_ctl->num_pages)
 636                 io_ctl_map_page(io_ctl, 0);
 637         return 0;
 638 }
 639
 640 static void io_ctl_zero_remaining_pages(struct btrfs_io_ctl *io_ctl)
 641 {
 642         /*
 643          * If we're not on the boundary we know we've modified the page and we
 644          * need to crc the page.
 645          */
 646         if (io_ctl->cur != io_ctl->orig)
 647                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 648         else
 649                 io_ctl_unmap_page(io_ctl);
 650
 651         while (io_ctl->index < io_ctl->num_pages) {
 652                 io_ctl_map_page(io_ctl, 1);
 653                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 654         }
 655 }
 656
 657 static int io_ctl_read_entry(struct btrfs_io_ctl *io_ctl,
 658                             struct btrfs_free_space *entry, u8 *type)
 659 {
 660         struct btrfs_free_space_entry *e;
 661         int ret;
 662
 663         if (!io_ctl->cur) {
 664                 ret = io_ctl_check_crc(io_ctl, io_ctl->index);
 665                 if (ret)
 666                         return ret;
 667         }
 668
 669         e = io_ctl->cur;
 670         entry->offset = get_unaligned_le64(&e->offset);
 671         entry->bytes = get_unaligned_le64(&e->bytes);
 672         *type = e->type;
 673         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
 674         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
 675
 676         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
 677                 return 0;
 678
 679         io_ctl_unmap_page(io_ctl);
 680
 681         return 0;
 682 }
 683
 684 static int io_ctl_read_bitmap(struct btrfs_io_ctl *io_ctl,
 685                               struct btrfs_free_space *entry)
 686 {
 687         int ret;
 688
 689         ret = io_ctl_check_crc(io_ctl, io_ctl->index);
 690         if (ret)
 691                 return ret;
 692
 693         copy_page(entry->bitmap, io_ctl->cur);
 694         io_ctl_unmap_page(io_ctl);
 695
 696         return 0;
 697 }
 698
 699 static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
 700 {
 701         struct btrfs_block_group *block_group = ctl->block_group;
 702         u64 max_bytes;
 703         u64 bitmap_bytes;
 704         u64 extent_bytes;
 705         u64 size = block_group->length;
 706         u64 bytes_per_bg = BITS_PER_BITMAP * ctl->unit;
 707         u64 max_bitmaps = div64_u64(size + bytes_per_bg - 1, bytes_per_bg);
 708
 709         max_bitmaps = max_t(u64, max_bitmaps, 1);
 710
 711         if (ctl->total_bitmaps > max_bitmaps)
 712                 btrfs_err(block_group->fs_info,
 713 "invalid free space control: bg start=%llu len=%llu total_bitmaps=%u unit=%u max_bitmaps=%llu bytes_per_bg=%llu",
 714                           block_group->start, block_group->length,
 715                           ctl->total_bitmaps, ctl->unit, max_bitmaps,
 716                           bytes_per_bg);
 717         ASSERT(ctl->total_bitmaps <= max_bitmaps);
 718
 719         /*
 720          * We are trying to keep the total amount of memory used per 1GiB of
 721          * space to be MAX_CACHE_BYTES_PER_GIG.  However, with a reclamation
 722          * mechanism of pulling extents >= FORCE_EXTENT_THRESHOLD out of
 723          * bitmaps, we may end up using more memory than this.
 724          */
 725         if (size < SZ_1G)
 726                 max_bytes = MAX_CACHE_BYTES_PER_GIG;
 727         else
 728                 max_bytes = MAX_CACHE_BYTES_PER_GIG * div_u64(size, SZ_1G);
 729
 730         bitmap_bytes = ctl->total_bitmaps * ctl->unit;
 731
 732         /*
 733          * we want the extent entry threshold to always be at most 1/2 the max
 734          * bytes we can have, or whatever is less than that.
 735          */
 736         extent_bytes = max_bytes - bitmap_bytes;
 737         extent_bytes = min_t(u64, extent_bytes, max_bytes >> 1);
 738
 739         ctl->extents_thresh =
 740                 div_u64(extent_bytes, sizeof(struct btrfs_free_space));
 741 }
 742
 743 static int __load_free_space_cache(struct btrfs_root *root, struct inode *inode,
 744                                    struct btrfs_free_space_ctl *ctl,
 745                                    struct btrfs_path *path, u64 offset)
 746 {
 747         struct btrfs_fs_info *fs_info = root->fs_info;
 748         struct btrfs_free_space_header *header;
 749         struct extent_buffer *leaf;
 750         struct btrfs_io_ctl io_ctl;
 751         struct btrfs_key key;
 752         struct btrfs_free_space *e, *n;
 753         LIST_HEAD(bitmaps);
 754         u64 num_entries;
 755         u64 num_bitmaps;
 756         u64 generation;
 757         u8 type;
 758         int ret = 0;
 759
 760         /* Nothing in the space cache, goodbye */
 761         if (!i_size_read(inode))
 762                 return 0;
 763
 764         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 765         key.offset = offset;
 766         key.type = 0;
 767
 768         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
 769         if (ret < 0)
 770                 return 0;
 771         else if (ret > 0) {
 772                 btrfs_release_path(path);
 773                 return 0;
 774         }
 775
 776         ret = -1;
 777
 778         leaf = path->nodes[0];
 779         header = btrfs_item_ptr(leaf, path->slots[0],
 780                                 struct btrfs_free_space_header);
 781         num_entries = btrfs_free_space_entries(leaf, header);
 782         num_bitmaps = btrfs_free_space_bitmaps(leaf, header);
 783         generation = btrfs_free_space_generation(leaf, header);
 784         btrfs_release_path(path);
 785
 786         if (!BTRFS_I(inode)->generation) {
 787                 btrfs_info(fs_info,
 788                            "the free space cache file (%llu) is invalid, skip it",
 789                            offset);
 790                 return 0;
 791         }
 792
 793         if (BTRFS_I(inode)->generation != generation) {
 794                 btrfs_err(fs_info,
 795                           "free space inode generation (%llu) did not match free space cache generation (%llu)",
 796                           BTRFS_I(inode)->generation, generation);
 797                 return 0;
 798         }
 799
 800         if (!num_entries)
 801                 return 0;
 802
 803         ret = io_ctl_init(&io_ctl, inode, 0);
 804         if (ret)
 805                 return ret;
 806
 807         readahead_cache(inode);
 808
 809         ret = io_ctl_prepare_pages(&io_ctl, true);
 810         if (ret)
 811                 goto out;
 812
 813         ret = io_ctl_check_crc(&io_ctl, 0);
 814         if (ret)
 815                 goto free_cache;
 816
 817         ret = io_ctl_check_generation(&io_ctl, generation);
 818         if (ret)
 819                 goto free_cache;
 820
 821         while (num_entries) {
 822                 e = kmem_cache_zalloc(btrfs_free_space_cachep,
 823                                       GFP_NOFS);
 824                 if (!e) {
 825                         ret = -ENOMEM;
 826                         goto free_cache;
 827                 }
 828
 829                 ret = io_ctl_read_entry(&io_ctl, e, &type);
 830                 if (ret) {
 831                         kmem_cache_free(btrfs_free_space_cachep, e);
 832                         goto free_cache;
 833                 }
 834
 835                 if (!e->bytes) {
 836                         ret = -1;
 837                         kmem_cache_free(btrfs_free_space_cachep, e);
 838                         goto free_cache;
 839                 }
 840
 841                 if (type == BTRFS_FREE_SPACE_EXTENT) {
 842                         spin_lock(&ctl->tree_lock);
 843                         ret = link_free_space(ctl, e);
 844                         spin_unlock(&ctl->tree_lock);
 845                         if (ret) {
 846                                 btrfs_err(fs_info,
 847                                         "Duplicate entries in free space cache, dumping");
 848                                 kmem_cache_free(btrfs_free_space_cachep, e);
 849                                 goto free_cache;
 850                         }
 851                 } else {
 852                         ASSERT(num_bitmaps);
 853                         num_bitmaps--;
 854                         e->bitmap = kmem_cache_zalloc(
 855                                         btrfs_free_space_bitmap_cachep, GFP_NOFS);
 856                         if (!e->bitmap) {
 857                                 ret = -ENOMEM;
 858                                 kmem_cache_free(
 859                                         btrfs_free_space_cachep, e);
 860                                 goto free_cache;
 861                         }
 862                         spin_lock(&ctl->tree_lock);
 863                         ret = link_free_space(ctl, e);
 864                         ctl->total_bitmaps++;
 865                         recalculate_thresholds(ctl);
 866                         spin_unlock(&ctl->tree_lock);
 867                         if (ret) {
 868                                 btrfs_err(fs_info,
 869                                         "Duplicate entries in free space cache, dumping");
 870                                 kmem_cache_free(btrfs_free_space_cachep, e);
 871                                 goto free_cache;
 872                         }
 873                         list_add_tail(&e->list, &bitmaps);
 874                 }
 875
 876                 num_entries--;
 877         }
 878
 879         io_ctl_unmap_page(&io_ctl);
 880
 881         /*
 882          * We add the bitmaps at the end of the entries in order that
 883          * the bitmap entries are added to the cache.
 884          */
 885         list_for_each_entry_safe(e, n, &bitmaps, list) {
 886                 list_del_init(&e->list);
 887                 ret = io_ctl_read_bitmap(&io_ctl, e);
 888                 if (ret)
 889                         goto free_cache;
 890         }
 891
 892         io_ctl_drop_pages(&io_ctl);
 893         ret = 1;
 894 out:
 895         io_ctl_free(&io_ctl);
 896         return ret;
 897 free_cache:
 898         io_ctl_drop_pages(&io_ctl);
 899
 900         spin_lock(&ctl->tree_lock);
 901         __btrfs_remove_free_space_cache(ctl);
 902         spin_unlock(&ctl->tree_lock);
 903         goto out;
 904 }
 905
 906 static int copy_free_space_cache(struct btrfs_block_group *block_group,
 907                                  struct btrfs_free_space_ctl *ctl)
 908 {
 909         struct btrfs_free_space *info;
 910         struct rb_node *n;
 911         int ret = 0;
 912
 913         while (!ret && (n = rb_first(&ctl->free_space_offset)) != NULL) {
 914                 info = rb_entry(n, struct btrfs_free_space, offset_index);
 915                 if (!info->bitmap) {
 916                         unlink_free_space(ctl, info, true);
 917                         ret = btrfs_add_free_space(block_group, info->offset,
 918                                                    info->bytes);
 919                         kmem_cache_free(btrfs_free_space_cachep, info);
 920                 } else {
 921                         u64 offset = info->offset;
 922                         u64 bytes = ctl->unit;
 923
 924                         while (search_bitmap(ctl, info, &offset, &bytes,
 925                                              false) == 0) {
 926                                 ret = btrfs_add_free_space(block_group, offset,
 927                                                            bytes);
 928                                 if (ret)
 929                                         break;
 930                                 bitmap_clear_bits(ctl, info, offset, bytes, true);
 931                                 offset = info->offset;
 932                                 bytes = ctl->unit;
 933                         }
 934                         free_bitmap(ctl, info);
 935                 }
 936                 cond_resched();
 937         }
 938         return ret;
 939 }
 940
 941 static struct lock_class_key btrfs_free_space_inode_key;
 942
 943 int load_free_space_cache(struct btrfs_block_group *block_group)
 944 {
 945         struct btrfs_fs_info *fs_info = block_group->fs_info;
 946         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
 947         struct btrfs_free_space_ctl tmp_ctl = {};
 948         struct inode *inode;
 949         struct btrfs_path *path;
 950         int ret = 0;
 951         bool matched;
 952         u64 used = block_group->used;
 953
 954         /*
 955          * Because we could potentially discard our loaded free space, we want
 956          * to load everything into a temporary structure first, and then if it's
 957          * valid copy it all into the actual free space ctl.
 958          */
 959         btrfs_init_free_space_ctl(block_group, &tmp_ctl);
 960
 961         /*
 962          * If this block group has been marked to be cleared for one reason or
 963          * another then we can't trust the on disk cache, so just return.
 964          */
 965         spin_lock(&block_group->lock);
 966         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
 967                 spin_unlock(&block_group->lock);
 968                 return 0;
 969         }
 970         spin_unlock(&block_group->lock);
 971
 972         path = btrfs_alloc_path();
 973         if (!path)
 974                 return 0;
 975         path->search_commit_root = 1;
 976         path->skip_locking = 1;
 977
 978         /*
 979          * We must pass a path with search_commit_root set to btrfs_iget in
 980          * order to avoid a deadlock when allocating extents for the tree root.
 981          *
 982          * When we are COWing an extent buffer from the tree root, when looking
 983          * for a free extent, at extent-tree.c:find_free_extent(), we can find
 984          * block group without its free space cache loaded. When we find one
 985          * we must load its space cache which requires reading its free space
 986          * cache's inode item from the root tree. If this inode item is located
 987          * in the same leaf that we started COWing before, then we end up in
 988          * deadlock on the extent buffer (trying to read lock it when we
 989          * previously write locked it).
 990          *
 991          * It's safe to read the inode item using the commit root because
 992          * block groups, once loaded, stay in memory forever (until they are
 993          * removed) as well as their space caches once loaded. New block groups
 994          * once created get their ->cached field set to BTRFS_CACHE_FINISHED so
 995          * we will never try to read their inode item while the fs is mounted.
 996          */
 997         inode = lookup_free_space_inode(block_group, path);
 998         if (IS_ERR(inode)) {
 999                 btrfs_free_path(path);
1000                 return 0;
1001         }
1002
1003         /* We may have converted the inode and made the cache invalid. */
1004         spin_lock(&block_group->lock);
1005         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
1006                 spin_unlock(&block_group->lock);
1007                 btrfs_free_path(path);
1008                 goto out;
1009         }
1010         spin_unlock(&block_group->lock);
1011
1012         /*
1013          * Reinitialize the class of struct inode's mapping->invalidate_lock for
1014          * free space inodes to prevent false positives related to locks for normal
1015          * inodes.
1016          */
1017         lockdep_set_class(&(&inode->i_data)->invalidate_lock,
1018                           &btrfs_free_space_inode_key);
1019
1020         ret = __load_free_space_cache(fs_info->tree_root, inode, &tmp_ctl,
1021                                       path, block_group->start);
1022         btrfs_free_path(path);
1023         if (ret <= 0)
1024                 goto out;
1025
1026         matched = (tmp_ctl.free_space == (block_group->length - used -
1027                                           block_group->bytes_super));
1028
1029         if (matched) {
1030                 ret = copy_free_space_cache(block_group, &tmp_ctl);
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(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                 ret = find_first_extent_bit(unpin, start,
1226                                             &extent_start, &extent_end,
1227                                             EXTENT_DIRTY, NULL);
1228                 if (ret)
1229                         return 0;
1230
1231                 /* This pinned extent is out of our range */
1232                 if (extent_start >= block_group->start + block_group->length)
1233                         return 0;
1234
1235                 extent_start = max(extent_start, start);
1236                 extent_end = min(block_group->start + block_group->length,
1237                                  extent_end + 1);
1238                 len = extent_end - extent_start;
1239
1240                 *entries += 1;
1241                 ret = io_ctl_add_entry(io_ctl, extent_start, len, NULL);
1242                 if (ret)
1243                         return -ENOSPC;
1244
1245                 start = extent_end;
1246         }
1247
1248         return 0;
1249 }
1250
1251 static noinline_for_stack int
1252 write_bitmap_entries(struct btrfs_io_ctl *io_ctl, struct list_head *bitmap_list)
1253 {
1254         struct btrfs_free_space *entry, *next;
1255         int ret;
1256
1257         /* Write out the bitmaps */
1258         list_for_each_entry_safe(entry, next, bitmap_list, list) {
1259                 ret = io_ctl_add_bitmap(io_ctl, entry->bitmap);
1260                 if (ret)
1261                         return -ENOSPC;
1262                 list_del_init(&entry->list);
1263         }
1264
1265         return 0;
1266 }
1267
1268 static int flush_dirty_cache(struct inode *inode)
1269 {
1270         int ret;
1271
1272         ret = btrfs_wait_ordered_range(inode, 0, (u64)-1);
1273         if (ret)
1274                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1275                                  EXTENT_DELALLOC, NULL);
1276
1277         return ret;
1278 }
1279
1280 static void noinline_for_stack
1281 cleanup_bitmap_list(struct list_head *bitmap_list)
1282 {
1283         struct btrfs_free_space *entry, *next;
1284
1285         list_for_each_entry_safe(entry, next, bitmap_list, list)
1286                 list_del_init(&entry->list);
1287 }
1288
1289 static void noinline_for_stack
1290 cleanup_write_cache_enospc(struct inode *inode,
1291                            struct btrfs_io_ctl *io_ctl,
1292                            struct extent_state **cached_state)
1293 {
1294         io_ctl_drop_pages(io_ctl);
1295         unlock_extent(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
1296                       cached_state);
1297 }
1298
1299 static int __btrfs_wait_cache_io(struct btrfs_root *root,
1300                                  struct btrfs_trans_handle *trans,
1301                                  struct btrfs_block_group *block_group,
1302                                  struct btrfs_io_ctl *io_ctl,
1303                                  struct btrfs_path *path, u64 offset)
1304 {
1305         int ret;
1306         struct inode *inode = io_ctl->inode;
1307
1308         if (!inode)
1309                 return 0;
1310
1311         /* Flush the dirty pages in the cache file. */
1312         ret = flush_dirty_cache(inode);
1313         if (ret)
1314                 goto out;
1315
1316         /* Update the cache item to tell everyone this cache file is valid. */
1317         ret = update_cache_item(trans, root, inode, path, offset,
1318                                 io_ctl->entries, io_ctl->bitmaps);
1319 out:
1320         if (ret) {
1321                 invalidate_inode_pages2(inode->i_mapping);
1322                 BTRFS_I(inode)->generation = 0;
1323                 if (block_group)
1324                         btrfs_debug(root->fs_info,
1325           "failed to write free space cache for block group %llu error %d",
1326                                   block_group->start, ret);
1327         }
1328         btrfs_update_inode(trans, root, BTRFS_I(inode));
1329
1330         if (block_group) {
1331                 /* the dirty list is protected by the dirty_bgs_lock */
1332                 spin_lock(&trans->transaction->dirty_bgs_lock);
1333
1334                 /* the disk_cache_state is protected by the block group lock */
1335                 spin_lock(&block_group->lock);
1336
1337                 /*
1338                  * only mark this as written if we didn't get put back on
1339                  * the dirty list while waiting for IO.   Otherwise our
1340                  * cache state won't be right, and we won't get written again
1341                  */
1342                 if (!ret && list_empty(&block_group->dirty_list))
1343                         block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1344                 else if (ret)
1345                         block_group->disk_cache_state = BTRFS_DC_ERROR;
1346
1347                 spin_unlock(&block_group->lock);
1348                 spin_unlock(&trans->transaction->dirty_bgs_lock);
1349                 io_ctl->inode = NULL;
1350                 iput(inode);
1351         }
1352
1353         return ret;
1354
1355 }
1356
1357 int btrfs_wait_cache_io(struct btrfs_trans_handle *trans,
1358                         struct btrfs_block_group *block_group,
1359                         struct btrfs_path *path)
1360 {
1361         return __btrfs_wait_cache_io(block_group->fs_info->tree_root, trans,
1362                                      block_group, &block_group->io_ctl,
1363                                      path, block_group->start);
1364 }
1365
1366 /**
1367  * Write out cached info to an inode
1368  *
1369  * @root:        root the inode belongs to
1370  * @inode:       freespace inode we are writing out
1371  * @ctl:         free space cache we are going to write out
1372  * @block_group: block_group for this cache if it belongs to a block_group
1373  * @io_ctl:      holds context for the io
1374  * @trans:       the trans handle
1375  *
1376  * This function writes out a free space cache struct to disk for quick recovery
1377  * on mount.  This will return 0 if it was successful in writing the cache out,
1378  * or an errno if it was not.
1379  */
1380 static int __btrfs_write_out_cache(struct btrfs_root *root, struct inode *inode,
1381                                    struct btrfs_free_space_ctl *ctl,
1382                                    struct btrfs_block_group *block_group,
1383                                    struct btrfs_io_ctl *io_ctl,
1384                                    struct btrfs_trans_handle *trans)
1385 {
1386         struct extent_state *cached_state = NULL;
1387         LIST_HEAD(bitmap_list);
1388         int entries = 0;
1389         int bitmaps = 0;
1390         int ret;
1391         int must_iput = 0;
1392
1393         if (!i_size_read(inode))
1394                 return -EIO;
1395
1396         WARN_ON(io_ctl->pages);
1397         ret = io_ctl_init(io_ctl, inode, 1);
1398         if (ret)
1399                 return ret;
1400
1401         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA)) {
1402                 down_write(&block_group->data_rwsem);
1403                 spin_lock(&block_group->lock);
1404                 if (block_group->delalloc_bytes) {
1405                         block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1406                         spin_unlock(&block_group->lock);
1407                         up_write(&block_group->data_rwsem);
1408                         BTRFS_I(inode)->generation = 0;
1409                         ret = 0;
1410                         must_iput = 1;
1411                         goto out;
1412                 }
1413                 spin_unlock(&block_group->lock);
1414         }
1415
1416         /* Lock all pages first so we can lock the extent safely. */
1417         ret = io_ctl_prepare_pages(io_ctl, false);
1418         if (ret)
1419                 goto out_unlock;
1420
1421         lock_extent(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
1422                     &cached_state);
1423
1424         io_ctl_set_generation(io_ctl, trans->transid);
1425
1426         mutex_lock(&ctl->cache_writeout_mutex);
1427         /* Write out the extent entries in the free space cache */
1428         spin_lock(&ctl->tree_lock);
1429         ret = write_cache_extent_entries(io_ctl, ctl,
1430                                          block_group, &entries, &bitmaps,
1431                                          &bitmap_list);
1432         if (ret)
1433                 goto out_nospc_locked;
1434
1435         /*
1436          * Some spaces that are freed in the current transaction are pinned,
1437          * they will be added into free space cache after the transaction is
1438          * committed, we shouldn't lose them.
1439          *
1440          * If this changes while we are working we'll get added back to
1441          * the dirty list and redo it.  No locking needed
1442          */
1443         ret = write_pinned_extent_entries(trans, block_group, io_ctl, &entries);
1444         if (ret)
1445                 goto out_nospc_locked;
1446
1447         /*
1448          * At last, we write out all the bitmaps and keep cache_writeout_mutex
1449          * locked while doing it because a concurrent trim can be manipulating
1450          * or freeing the bitmap.
1451          */
1452         ret = write_bitmap_entries(io_ctl, &bitmap_list);
1453         spin_unlock(&ctl->tree_lock);
1454         mutex_unlock(&ctl->cache_writeout_mutex);
1455         if (ret)
1456                 goto out_nospc;
1457
1458         /* Zero out the rest of the pages just to make sure */
1459         io_ctl_zero_remaining_pages(io_ctl);
1460
1461         /* Everything is written out, now we dirty the pages in the file. */
1462         ret = btrfs_dirty_pages(BTRFS_I(inode), io_ctl->pages,
1463                                 io_ctl->num_pages, 0, i_size_read(inode),
1464                                 &cached_state, false);
1465         if (ret)
1466                 goto out_nospc;
1467
1468         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1469                 up_write(&block_group->data_rwsem);
1470         /*
1471          * Release the pages and unlock the extent, we will flush
1472          * them out later
1473          */
1474         io_ctl_drop_pages(io_ctl);
1475         io_ctl_free(io_ctl);
1476
1477         unlock_extent(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
1478                       &cached_state);
1479
1480         /*
1481          * at this point the pages are under IO and we're happy,
1482          * The caller is responsible for waiting on them and updating
1483          * the cache and the inode
1484          */
1485         io_ctl->entries = entries;
1486         io_ctl->bitmaps = bitmaps;
1487
1488         ret = btrfs_fdatawrite_range(inode, 0, (u64)-1);
1489         if (ret)
1490                 goto out;
1491
1492         return 0;
1493
1494 out_nospc_locked:
1495         cleanup_bitmap_list(&bitmap_list);
1496         spin_unlock(&ctl->tree_lock);
1497         mutex_unlock(&ctl->cache_writeout_mutex);
1498
1499 out_nospc:
1500         cleanup_write_cache_enospc(inode, io_ctl, &cached_state);
1501
1502 out_unlock:
1503         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1504                 up_write(&block_group->data_rwsem);
1505
1506 out:
1507         io_ctl->inode = NULL;
1508         io_ctl_free(io_ctl);
1509         if (ret) {
1510                 invalidate_inode_pages2(inode->i_mapping);
1511                 BTRFS_I(inode)->generation = 0;
1512         }
1513         btrfs_update_inode(trans, root, BTRFS_I(inode));
1514         if (must_iput)
1515                 iput(inode);
1516         return ret;
1517 }
1518
1519 int btrfs_write_out_cache(struct btrfs_trans_handle *trans,
1520                           struct btrfs_block_group *block_group,
1521                           struct btrfs_path *path)
1522 {
1523         struct btrfs_fs_info *fs_info = trans->fs_info;
1524         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1525         struct inode *inode;
1526         int ret = 0;
1527
1528         spin_lock(&block_group->lock);
1529         if (block_group->disk_cache_state < BTRFS_DC_SETUP) {
1530                 spin_unlock(&block_group->lock);
1531                 return 0;
1532         }
1533         spin_unlock(&block_group->lock);
1534
1535         inode = lookup_free_space_inode(block_group, path);
1536         if (IS_ERR(inode))
1537                 return 0;
1538
1539         ret = __btrfs_write_out_cache(fs_info->tree_root, inode, ctl,
1540                                 block_group, &block_group->io_ctl, trans);
1541         if (ret) {
1542                 btrfs_debug(fs_info,
1543           "failed to write free space cache for block group %llu error %d",
1544                           block_group->start, ret);
1545                 spin_lock(&block_group->lock);
1546                 block_group->disk_cache_state = BTRFS_DC_ERROR;
1547                 spin_unlock(&block_group->lock);
1548
1549                 block_group->io_ctl.inode = NULL;
1550                 iput(inode);
1551         }
1552
1553         /*
1554          * if ret == 0 the caller is expected to call btrfs_wait_cache_io
1555          * to wait for IO and put the inode
1556          */
1557
1558         return ret;
1559 }
1560
1561 static inline unsigned long offset_to_bit(u64 bitmap_start, u32 unit,
1562                                           u64 offset)
1563 {
1564         ASSERT(offset >= bitmap_start);
1565         offset -= bitmap_start;
1566         return (unsigned long)(div_u64(offset, unit));
1567 }
1568
1569 static inline unsigned long bytes_to_bits(u64 bytes, u32 unit)
1570 {
1571         return (unsigned long)(div_u64(bytes, unit));
1572 }
1573
1574 static inline u64 offset_to_bitmap(struct btrfs_free_space_ctl *ctl,
1575                                    u64 offset)
1576 {
1577         u64 bitmap_start;
1578         u64 bytes_per_bitmap;
1579
1580         bytes_per_bitmap = BITS_PER_BITMAP * ctl->unit;
1581         bitmap_start = offset - ctl->start;
1582         bitmap_start = div64_u64(bitmap_start, bytes_per_bitmap);
1583         bitmap_start *= bytes_per_bitmap;
1584         bitmap_start += ctl->start;
1585
1586         return bitmap_start;
1587 }
1588
1589 static int tree_insert_offset(struct rb_root *root, u64 offset,
1590                               struct rb_node *node, int bitmap)
1591 {
1592         struct rb_node **p = &root->rb_node;
1593         struct rb_node *parent = NULL;
1594         struct btrfs_free_space *info;
1595
1596         while (*p) {
1597                 parent = *p;
1598                 info = rb_entry(parent, struct btrfs_free_space, offset_index);
1599
1600                 if (offset < info->offset) {
1601                         p = &(*p)->rb_left;
1602                 } else if (offset > info->offset) {
1603                         p = &(*p)->rb_right;
1604                 } else {
1605                         /*
1606                          * we could have a bitmap entry and an extent entry
1607                          * share the same offset.  If this is the case, we want
1608                          * the extent entry to always be found first if we do a
1609                          * linear search through the tree, since we want to have
1610                          * the quickest allocation time, and allocating from an
1611                          * extent is faster than allocating from a bitmap.  So
1612                          * if we're inserting a bitmap and we find an entry at
1613                          * this offset, we want to go right, or after this entry
1614                          * logically.  If we are inserting an extent and we've
1615                          * found a bitmap, we want to go left, or before
1616                          * logically.
1617                          */
1618                         if (bitmap) {
1619                                 if (info->bitmap) {
1620                                         WARN_ON_ONCE(1);
1621                                         return -EEXIST;
1622                                 }
1623                                 p = &(*p)->rb_right;
1624                         } else {
1625                                 if (!info->bitmap) {
1626                                         WARN_ON_ONCE(1);
1627                                         return -EEXIST;
1628                                 }
1629                                 p = &(*p)->rb_left;
1630                         }
1631                 }
1632         }
1633
1634         rb_link_node(node, parent, p);
1635         rb_insert_color(node, root);
1636
1637         return 0;
1638 }
1639
1640 /*
1641  * This is a little subtle.  We *only* have ->max_extent_size set if we actually
1642  * searched through the bitmap and figured out the largest ->max_extent_size,
1643  * otherwise it's 0.  In the case that it's 0 we don't want to tell the
1644  * allocator the wrong thing, we want to use the actual real max_extent_size
1645  * we've found already if it's larger, or we want to use ->bytes.
1646  *
1647  * This matters because find_free_space() will skip entries who's ->bytes is
1648  * less than the required bytes.  So if we didn't search down this bitmap, we
1649  * may pick some previous entry that has a smaller ->max_extent_size than we
1650  * have.  For example, assume we have two entries, one that has
1651  * ->max_extent_size set to 4K and ->bytes set to 1M.  A second entry hasn't set
1652  * ->max_extent_size yet, has ->bytes set to 8K and it's contiguous.  We will
1653  *  call into find_free_space(), and return with max_extent_size == 4K, because
1654  *  that first bitmap entry had ->max_extent_size set, but the second one did
1655  *  not.  If instead we returned 8K we'd come in searching for 8K, and find the
1656  *  8K contiguous range.
1657  *
1658  *  Consider the other case, we have 2 8K chunks in that second entry and still
1659  *  don't have ->max_extent_size set.  We'll return 16K, and the next time the
1660  *  allocator comes in it'll fully search our second bitmap, and this time it'll
1661  *  get an uptodate value of 8K as the maximum chunk size.  Then we'll get the
1662  *  right allocation the next loop through.
1663  */
1664 static inline u64 get_max_extent_size(const struct btrfs_free_space *entry)
1665 {
1666         if (entry->bitmap && entry->max_extent_size)
1667                 return entry->max_extent_size;
1668         return entry->bytes;
1669 }
1670
1671 /*
1672  * We want the largest entry to be leftmost, so this is inverted from what you'd
1673  * normally expect.
1674  */
1675 static bool entry_less(struct rb_node *node, const struct rb_node *parent)
1676 {
1677         const struct btrfs_free_space *entry, *exist;
1678
1679         entry = rb_entry(node, struct btrfs_free_space, bytes_index);
1680         exist = rb_entry(parent, struct btrfs_free_space, bytes_index);
1681         return get_max_extent_size(exist) < get_max_extent_size(entry);
1682 }
1683
1684 /*
1685  * searches the tree for the given offset.
1686  *
1687  * fuzzy - If this is set, then we are trying to make an allocation, and we just
1688  * want a section that has at least bytes size and comes at or after the given
1689  * offset.
1690  */
1691 static struct btrfs_free_space *
1692 tree_search_offset(struct btrfs_free_space_ctl *ctl,
1693                    u64 offset, int bitmap_only, int fuzzy)
1694 {
1695         struct rb_node *n = ctl->free_space_offset.rb_node;
1696         struct btrfs_free_space *entry = NULL, *prev = NULL;
1697
1698         /* find entry that is closest to the 'offset' */
1699         while (n) {
1700                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1701                 prev = entry;
1702
1703                 if (offset < entry->offset)
1704                         n = n->rb_left;
1705                 else if (offset > entry->offset)
1706                         n = n->rb_right;
1707                 else
1708                         break;
1709
1710                 entry = NULL;
1711         }
1712
1713         if (bitmap_only) {
1714                 if (!entry)
1715                         return NULL;
1716                 if (entry->bitmap)
1717                         return entry;
1718
1719                 /*
1720                  * bitmap entry and extent entry may share same offset,
1721                  * in that case, bitmap entry comes after extent entry.
1722                  */
1723                 n = rb_next(n);
1724                 if (!n)
1725                         return NULL;
1726                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1727                 if (entry->offset != offset)
1728                         return NULL;
1729
1730                 WARN_ON(!entry->bitmap);
1731                 return entry;
1732         } else if (entry) {
1733                 if (entry->bitmap) {
1734                         /*
1735                          * if previous extent entry covers the offset,
1736                          * we should return it instead of the bitmap entry
1737                          */
1738                         n = rb_prev(&entry->offset_index);
1739                         if (n) {
1740                                 prev = rb_entry(n, struct btrfs_free_space,
1741                                                 offset_index);
1742                                 if (!prev->bitmap &&
1743                                     prev->offset + prev->bytes > offset)
1744                                         entry = prev;
1745                         }
1746                 }
1747                 return entry;
1748         }
1749
1750         if (!prev)
1751                 return NULL;
1752
1753         /* find last entry before the 'offset' */
1754         entry = prev;
1755         if (entry->offset > offset) {
1756                 n = rb_prev(&entry->offset_index);
1757                 if (n) {
1758                         entry = rb_entry(n, struct btrfs_free_space,
1759                                         offset_index);
1760                         ASSERT(entry->offset <= offset);
1761                 } else {
1762                         if (fuzzy)
1763                                 return entry;
1764                         else
1765                                 return NULL;
1766                 }
1767         }
1768
1769         if (entry->bitmap) {
1770                 n = rb_prev(&entry->offset_index);
1771                 if (n) {
1772                         prev = rb_entry(n, struct btrfs_free_space,
1773                                         offset_index);
1774                         if (!prev->bitmap &&
1775                             prev->offset + prev->bytes > offset)
1776                                 return prev;
1777                 }
1778                 if (entry->offset + BITS_PER_BITMAP * ctl->unit > offset)
1779                         return entry;
1780         } else if (entry->offset + entry->bytes > offset)
1781                 return entry;
1782
1783         if (!fuzzy)
1784                 return NULL;
1785
1786         while (1) {
1787                 n = rb_next(&entry->offset_index);
1788                 if (!n)
1789                         return NULL;
1790                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1791                 if (entry->bitmap) {
1792                         if (entry->offset + BITS_PER_BITMAP *
1793                             ctl->unit > offset)
1794                                 break;
1795                 } else {
1796                         if (entry->offset + entry->bytes > offset)
1797                                 break;
1798                 }
1799         }
1800         return entry;
1801 }
1802
1803 static inline void unlink_free_space(struct btrfs_free_space_ctl *ctl,
1804                                      struct btrfs_free_space *info,
1805                                      bool update_stat)
1806 {
1807         rb_erase(&info->offset_index, &ctl->free_space_offset);
1808         rb_erase_cached(&info->bytes_index, &ctl->free_space_bytes);
1809         ctl->free_extents--;
1810
1811         if (!info->bitmap && !btrfs_free_space_trimmed(info)) {
1812                 ctl->discardable_extents[BTRFS_STAT_CURR]--;
1813                 ctl->discardable_bytes[BTRFS_STAT_CURR] -= info->bytes;
1814         }
1815
1816         if (update_stat)
1817                 ctl->free_space -= info->bytes;
1818 }
1819
1820 static int link_free_space(struct btrfs_free_space_ctl *ctl,
1821                            struct btrfs_free_space *info)
1822 {
1823         int ret = 0;
1824
1825         ASSERT(info->bytes || info->bitmap);
1826         ret = tree_insert_offset(&ctl->free_space_offset, info->offset,
1827                                  &info->offset_index, (info->bitmap != NULL));
1828         if (ret)
1829                 return ret;
1830
1831         rb_add_cached(&info->bytes_index, &ctl->free_space_bytes, entry_less);
1832
1833         if (!info->bitmap && !btrfs_free_space_trimmed(info)) {
1834                 ctl->discardable_extents[BTRFS_STAT_CURR]++;
1835                 ctl->discardable_bytes[BTRFS_STAT_CURR] += info->bytes;
1836         }
1837
1838         ctl->free_space += info->bytes;
1839         ctl->free_extents++;
1840         return ret;
1841 }
1842
1843 static void relink_bitmap_entry(struct btrfs_free_space_ctl *ctl,
1844                                 struct btrfs_free_space *info)
1845 {
1846         ASSERT(info->bitmap);
1847
1848         /*
1849          * If our entry is empty it's because we're on a cluster and we don't
1850          * want to re-link it into our ctl bytes index.
1851          */
1852         if (RB_EMPTY_NODE(&info->bytes_index))
1853                 return;
1854
1855         rb_erase_cached(&info->bytes_index, &ctl->free_space_bytes);
1856         rb_add_cached(&info->bytes_index, &ctl->free_space_bytes, entry_less);
1857 }
1858
1859 static inline void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1860                                      struct btrfs_free_space *info,
1861                                      u64 offset, u64 bytes, bool update_stat)
1862 {
1863         unsigned long start, count, end;
1864         int extent_delta = -1;
1865
1866         start = offset_to_bit(info->offset, ctl->unit, offset);
1867         count = bytes_to_bits(bytes, ctl->unit);
1868         end = start + count;
1869         ASSERT(end <= BITS_PER_BITMAP);
1870
1871         bitmap_clear(info->bitmap, start, count);
1872
1873         info->bytes -= bytes;
1874         if (info->max_extent_size > ctl->unit)
1875                 info->max_extent_size = 0;
1876
1877         relink_bitmap_entry(ctl, info);
1878
1879         if (start && test_bit(start - 1, info->bitmap))
1880                 extent_delta++;
1881
1882         if (end < BITS_PER_BITMAP && test_bit(end, info->bitmap))
1883                 extent_delta++;
1884
1885         info->bitmap_extents += extent_delta;
1886         if (!btrfs_free_space_trimmed(info)) {
1887                 ctl->discardable_extents[BTRFS_STAT_CURR] += extent_delta;
1888                 ctl->discardable_bytes[BTRFS_STAT_CURR] -= bytes;
1889         }
1890
1891         if (update_stat)
1892                 ctl->free_space -= bytes;
1893 }
1894
1895 static void bitmap_set_bits(struct btrfs_free_space_ctl *ctl,
1896                             struct btrfs_free_space *info, u64 offset,
1897                             u64 bytes)
1898 {
1899         unsigned long start, count, end;
1900         int extent_delta = 1;
1901
1902         start = offset_to_bit(info->offset, ctl->unit, offset);
1903         count = bytes_to_bits(bytes, ctl->unit);
1904         end = start + count;
1905         ASSERT(end <= BITS_PER_BITMAP);
1906
1907         bitmap_set(info->bitmap, start, count);
1908
1909         /*
1910          * We set some bytes, we have no idea what the max extent size is
1911          * anymore.
1912          */
1913         info->max_extent_size = 0;
1914         info->bytes += bytes;
1915         ctl->free_space += bytes;
1916
1917         relink_bitmap_entry(ctl, info);
1918
1919         if (start && test_bit(start - 1, info->bitmap))
1920                 extent_delta--;
1921
1922         if (end < BITS_PER_BITMAP && test_bit(end, info->bitmap))
1923                 extent_delta--;
1924
1925         info->bitmap_extents += extent_delta;
1926         if (!btrfs_free_space_trimmed(info)) {
1927                 ctl->discardable_extents[BTRFS_STAT_CURR] += extent_delta;
1928                 ctl->discardable_bytes[BTRFS_STAT_CURR] += bytes;
1929         }
1930 }
1931
1932 /*
1933  * If we can not find suitable extent, we will use bytes to record
1934  * the size of the max extent.
1935  */
1936 static int search_bitmap(struct btrfs_free_space_ctl *ctl,
1937                          struct btrfs_free_space *bitmap_info, u64 *offset,
1938                          u64 *bytes, bool for_alloc)
1939 {
1940         unsigned long found_bits = 0;
1941         unsigned long max_bits = 0;
1942         unsigned long bits, i;
1943         unsigned long next_zero;
1944         unsigned long extent_bits;
1945
1946         /*
1947          * Skip searching the bitmap if we don't have a contiguous section that
1948          * is large enough for this allocation.
1949          */
1950         if (for_alloc &&
1951             bitmap_info->max_extent_size &&
1952             bitmap_info->max_extent_size < *bytes) {
1953                 *bytes = bitmap_info->max_extent_size;
1954                 return -1;
1955         }
1956
1957         i = offset_to_bit(bitmap_info->offset, ctl->unit,
1958                           max_t(u64, *offset, bitmap_info->offset));
1959         bits = bytes_to_bits(*bytes, ctl->unit);
1960
1961         for_each_set_bit_from(i, bitmap_info->bitmap, BITS_PER_BITMAP) {
1962                 if (for_alloc && bits == 1) {
1963                         found_bits = 1;
1964                         break;
1965                 }
1966                 next_zero = find_next_zero_bit(bitmap_info->bitmap,
1967                                                BITS_PER_BITMAP, i);
1968                 extent_bits = next_zero - i;
1969                 if (extent_bits >= bits) {
1970                         found_bits = extent_bits;
1971                         break;
1972                 } else if (extent_bits > max_bits) {
1973                         max_bits = extent_bits;
1974                 }
1975                 i = next_zero;
1976         }
1977
1978         if (found_bits) {
1979                 *offset = (u64)(i * ctl->unit) + bitmap_info->offset;
1980                 *bytes = (u64)(found_bits) * ctl->unit;
1981                 return 0;
1982         }
1983
1984         *bytes = (u64)(max_bits) * ctl->unit;
1985         bitmap_info->max_extent_size = *bytes;
1986         relink_bitmap_entry(ctl, bitmap_info);
1987         return -1;
1988 }
1989
1990 /* Cache the size of the max extent in bytes */
1991 static struct btrfs_free_space *
1992 find_free_space(struct btrfs_free_space_ctl *ctl, u64 *offset, u64 *bytes,
1993                 unsigned long align, u64 *max_extent_size, bool use_bytes_index)
1994 {
1995         struct btrfs_free_space *entry;
1996         struct rb_node *node;
1997         u64 tmp;
1998         u64 align_off;
1999         int ret;
2000
2001         if (!ctl->free_space_offset.rb_node)
2002                 goto out;
2003 again:
2004         if (use_bytes_index) {
2005                 node = rb_first_cached(&ctl->free_space_bytes);
2006         } else {
2007                 entry = tree_search_offset(ctl, offset_to_bitmap(ctl, *offset),
2008                                            0, 1);
2009                 if (!entry)
2010                         goto out;
2011                 node = &entry->offset_index;
2012         }
2013
2014         for (; node; node = rb_next(node)) {
2015                 if (use_bytes_index)
2016                         entry = rb_entry(node, struct btrfs_free_space,
2017                                          bytes_index);
2018                 else
2019                         entry = rb_entry(node, struct btrfs_free_space,
2020                                          offset_index);
2021
2022                 /*
2023                  * If we are using the bytes index then all subsequent entries
2024                  * in this tree are going to be < bytes, so simply set the max
2025                  * extent size and exit the loop.
2026                  *
2027                  * If we're using the offset index then we need to keep going
2028                  * through the rest of the tree.
2029                  */
2030                 if (entry->bytes < *bytes) {
2031                         *max_extent_size = max(get_max_extent_size(entry),
2032                                                *max_extent_size);
2033                         if (use_bytes_index)
2034                                 break;
2035                         continue;
2036                 }
2037
2038                 /* make sure the space returned is big enough
2039                  * to match our requested alignment
2040                  */
2041                 if (*bytes >= align) {
2042                         tmp = entry->offset - ctl->start + align - 1;
2043                         tmp = div64_u64(tmp, align);
2044                         tmp = tmp * align + ctl->start;
2045                         align_off = tmp - entry->offset;
2046                 } else {
2047                         align_off = 0;
2048                         tmp = entry->offset;
2049                 }
2050
2051                 /*
2052                  * We don't break here if we're using the bytes index because we
2053                  * may have another entry that has the correct alignment that is
2054                  * the right size, so we don't want to miss that possibility.
2055                  * At worst this adds another loop through the logic, but if we
2056                  * broke here we could prematurely ENOSPC.
2057                  */
2058                 if (entry->bytes < *bytes + align_off) {
2059                         *max_extent_size = max(get_max_extent_size(entry),
2060                                                *max_extent_size);
2061                         continue;
2062                 }
2063
2064                 if (entry->bitmap) {
2065                         struct rb_node *old_next = rb_next(node);
2066                         u64 size = *bytes;
2067
2068                         ret = search_bitmap(ctl, entry, &tmp, &size, true);
2069                         if (!ret) {
2070                                 *offset = tmp;
2071                                 *bytes = size;
2072                                 return entry;
2073                         } else {
2074                                 *max_extent_size =
2075                                         max(get_max_extent_size(entry),
2076                                             *max_extent_size);
2077                         }
2078
2079                         /*
2080                          * The bitmap may have gotten re-arranged in the space
2081                          * index here because the max_extent_size may have been
2082                          * updated.  Start from the beginning again if this
2083                          * happened.
2084                          */
2085                         if (use_bytes_index && old_next != rb_next(node))
2086                                 goto again;
2087                         continue;
2088                 }
2089
2090                 *offset = tmp;
2091                 *bytes = entry->bytes - align_off;
2092                 return entry;
2093         }
2094 out:
2095         return NULL;
2096 }
2097
2098 static void add_new_bitmap(struct btrfs_free_space_ctl *ctl,
2099                            struct btrfs_free_space *info, u64 offset)
2100 {
2101         info->offset = offset_to_bitmap(ctl, offset);
2102         info->bytes = 0;
2103         info->bitmap_extents = 0;
2104         INIT_LIST_HEAD(&info->list);
2105         link_free_space(ctl, info);
2106         ctl->total_bitmaps++;
2107         recalculate_thresholds(ctl);
2108 }
2109
2110 static void free_bitmap(struct btrfs_free_space_ctl *ctl,
2111                         struct btrfs_free_space *bitmap_info)
2112 {
2113         /*
2114          * Normally when this is called, the bitmap is completely empty. However,
2115          * if we are blowing up the free space cache for one reason or another
2116          * via __btrfs_remove_free_space_cache(), then it may not be freed and
2117          * we may leave stats on the table.
2118          */
2119         if (bitmap_info->bytes && !btrfs_free_space_trimmed(bitmap_info)) {
2120                 ctl->discardable_extents[BTRFS_STAT_CURR] -=
2121                         bitmap_info->bitmap_extents;
2122                 ctl->discardable_bytes[BTRFS_STAT_CURR] -= bitmap_info->bytes;
2123
2124         }
2125         unlink_free_space(ctl, bitmap_info, true);
2126         kmem_cache_free(btrfs_free_space_bitmap_cachep, bitmap_info->bitmap);
2127         kmem_cache_free(btrfs_free_space_cachep, bitmap_info);
2128         ctl->total_bitmaps--;
2129         recalculate_thresholds(ctl);
2130 }
2131
2132 static noinline int remove_from_bitmap(struct btrfs_free_space_ctl *ctl,
2133                               struct btrfs_free_space *bitmap_info,
2134                               u64 *offset, u64 *bytes)
2135 {
2136         u64 end;
2137         u64 search_start, search_bytes;
2138         int ret;
2139
2140 again:
2141         end = bitmap_info->offset + (u64)(BITS_PER_BITMAP * ctl->unit) - 1;
2142
2143         /*
2144          * We need to search for bits in this bitmap.  We could only cover some
2145          * of the extent in this bitmap thanks to how we add space, so we need
2146          * to search for as much as it as we can and clear that amount, and then
2147          * go searching for the next bit.
2148          */
2149         search_start = *offset;
2150         search_bytes = ctl->unit;
2151         search_bytes = min(search_bytes, end - search_start + 1);
2152         ret = search_bitmap(ctl, bitmap_info, &search_start, &search_bytes,
2153                             false);
2154         if (ret < 0 || search_start != *offset)
2155                 return -EINVAL;
2156
2157         /* We may have found more bits than what we need */
2158         search_bytes = min(search_bytes, *bytes);
2159
2160         /* Cannot clear past the end of the bitmap */
2161         search_bytes = min(search_bytes, end - search_start + 1);
2162
2163         bitmap_clear_bits(ctl, bitmap_info, search_start, search_bytes, true);
2164         *offset += search_bytes;
2165         *bytes -= search_bytes;
2166
2167         if (*bytes) {
2168                 struct rb_node *next = rb_next(&bitmap_info->offset_index);
2169                 if (!bitmap_info->bytes)
2170                         free_bitmap(ctl, bitmap_info);
2171
2172                 /*
2173                  * no entry after this bitmap, but we still have bytes to
2174                  * remove, so something has gone wrong.
2175                  */
2176                 if (!next)
2177                         return -EINVAL;
2178
2179                 bitmap_info = rb_entry(next, struct btrfs_free_space,
2180                                        offset_index);
2181
2182                 /*
2183                  * if the next entry isn't a bitmap we need to return to let the
2184                  * extent stuff do its work.
2185                  */
2186                 if (!bitmap_info->bitmap)
2187                         return -EAGAIN;
2188
2189                 /*
2190                  * Ok the next item is a bitmap, but it may not actually hold
2191                  * the information for the rest of this free space stuff, so
2192                  * look for it, and if we don't find it return so we can try
2193                  * everything over again.
2194                  */
2195                 search_start = *offset;
2196                 search_bytes = ctl->unit;
2197                 ret = search_bitmap(ctl, bitmap_info, &search_start,
2198                                     &search_bytes, false);
2199                 if (ret < 0 || search_start != *offset)
2200                         return -EAGAIN;
2201
2202                 goto again;
2203         } else if (!bitmap_info->bytes)
2204                 free_bitmap(ctl, bitmap_info);
2205
2206         return 0;
2207 }
2208
2209 static u64 add_bytes_to_bitmap(struct btrfs_free_space_ctl *ctl,
2210                                struct btrfs_free_space *info, u64 offset,
2211                                u64 bytes, enum btrfs_trim_state trim_state)
2212 {
2213         u64 bytes_to_set = 0;
2214         u64 end;
2215
2216         /*
2217          * This is a tradeoff to make bitmap trim state minimal.  We mark the
2218          * whole bitmap untrimmed if at any point we add untrimmed regions.
2219          */
2220         if (trim_state == BTRFS_TRIM_STATE_UNTRIMMED) {
2221                 if (btrfs_free_space_trimmed(info)) {
2222                         ctl->discardable_extents[BTRFS_STAT_CURR] +=
2223                                 info->bitmap_extents;
2224                         ctl->discardable_bytes[BTRFS_STAT_CURR] += info->bytes;
2225                 }
2226                 info->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2227         }
2228
2229         end = info->offset + (u64)(BITS_PER_BITMAP * ctl->unit);
2230
2231         bytes_to_set = min(end - offset, bytes);
2232
2233         bitmap_set_bits(ctl, info, offset, bytes_to_set);
2234
2235         return bytes_to_set;
2236
2237 }
2238
2239 static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
2240                       struct btrfs_free_space *info)
2241 {
2242         struct btrfs_block_group *block_group = ctl->block_group;
2243         struct btrfs_fs_info *fs_info = block_group->fs_info;
2244         bool forced = false;
2245
2246 #ifdef CONFIG_BTRFS_DEBUG
2247         if (btrfs_should_fragment_free_space(block_group))
2248                 forced = true;
2249 #endif
2250
2251         /* This is a way to reclaim large regions from the bitmaps. */
2252         if (!forced && info->bytes >= FORCE_EXTENT_THRESHOLD)
2253                 return false;
2254
2255         /*
2256          * If we are below the extents threshold then we can add this as an
2257          * extent, and don't have to deal with the bitmap
2258          */
2259         if (!forced && ctl->free_extents < ctl->extents_thresh) {
2260                 /*
2261                  * If this block group has some small extents we don't want to
2262                  * use up all of our free slots in the cache with them, we want
2263                  * to reserve them to larger extents, however if we have plenty
2264                  * of cache left then go ahead an dadd them, no sense in adding
2265                  * the overhead of a bitmap if we don't have to.
2266                  */
2267                 if (info->bytes <= fs_info->sectorsize * 8) {
2268                         if (ctl->free_extents * 3 <= ctl->extents_thresh)
2269                                 return false;
2270                 } else {
2271                         return false;
2272                 }
2273         }
2274
2275         /*
2276          * The original block groups from mkfs can be really small, like 8
2277          * megabytes, so don't bother with a bitmap for those entries.  However
2278          * some block groups can be smaller than what a bitmap would cover but
2279          * are still large enough that they could overflow the 32k memory limit,
2280          * so allow those block groups to still be allowed to have a bitmap
2281          * entry.
2282          */
2283         if (((BITS_PER_BITMAP * ctl->unit) >> 1) > block_group->length)
2284                 return false;
2285
2286         return true;
2287 }
2288
2289 static const struct btrfs_free_space_op free_space_op = {
2290         .use_bitmap             = use_bitmap,
2291 };
2292
2293 static int insert_into_bitmap(struct btrfs_free_space_ctl *ctl,
2294                               struct btrfs_free_space *info)
2295 {
2296         struct btrfs_free_space *bitmap_info;
2297         struct btrfs_block_group *block_group = NULL;
2298         int added = 0;
2299         u64 bytes, offset, bytes_added;
2300         enum btrfs_trim_state trim_state;
2301         int ret;
2302
2303         bytes = info->bytes;
2304         offset = info->offset;
2305         trim_state = info->trim_state;
2306
2307         if (!ctl->op->use_bitmap(ctl, info))
2308                 return 0;
2309
2310         if (ctl->op == &free_space_op)
2311                 block_group = ctl->block_group;
2312 again:
2313         /*
2314          * Since we link bitmaps right into the cluster we need to see if we
2315          * have a cluster here, and if so and it has our bitmap we need to add
2316          * the free space to that bitmap.
2317          */
2318         if (block_group && !list_empty(&block_group->cluster_list)) {
2319                 struct btrfs_free_cluster *cluster;
2320                 struct rb_node *node;
2321                 struct btrfs_free_space *entry;
2322
2323                 cluster = list_entry(block_group->cluster_list.next,
2324                                      struct btrfs_free_cluster,
2325                                      block_group_list);
2326                 spin_lock(&cluster->lock);
2327                 node = rb_first(&cluster->root);
2328                 if (!node) {
2329                         spin_unlock(&cluster->lock);
2330                         goto no_cluster_bitmap;
2331                 }
2332
2333                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2334                 if (!entry->bitmap) {
2335                         spin_unlock(&cluster->lock);
2336                         goto no_cluster_bitmap;
2337                 }
2338
2339                 if (entry->offset == offset_to_bitmap(ctl, offset)) {
2340                         bytes_added = add_bytes_to_bitmap(ctl, entry, offset,
2341                                                           bytes, trim_state);
2342                         bytes -= bytes_added;
2343                         offset += bytes_added;
2344                 }
2345                 spin_unlock(&cluster->lock);
2346                 if (!bytes) {
2347                         ret = 1;
2348                         goto out;
2349                 }
2350         }
2351
2352 no_cluster_bitmap:
2353         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2354                                          1, 0);
2355         if (!bitmap_info) {
2356                 ASSERT(added == 0);
2357                 goto new_bitmap;
2358         }
2359
2360         bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes,
2361                                           trim_state);
2362         bytes -= bytes_added;
2363         offset += bytes_added;
2364         added = 0;
2365
2366         if (!bytes) {
2367                 ret = 1;
2368                 goto out;
2369         } else
2370                 goto again;
2371
2372 new_bitmap:
2373         if (info && info->bitmap) {
2374                 add_new_bitmap(ctl, info, offset);
2375                 added = 1;
2376                 info = NULL;
2377                 goto again;
2378         } else {
2379                 spin_unlock(&ctl->tree_lock);
2380
2381                 /* no pre-allocated info, allocate a new one */
2382                 if (!info) {
2383                         info = kmem_cache_zalloc(btrfs_free_space_cachep,
2384                                                  GFP_NOFS);
2385                         if (!info) {
2386                                 spin_lock(&ctl->tree_lock);
2387                                 ret = -ENOMEM;
2388                                 goto out;
2389                         }
2390                 }
2391
2392                 /* allocate the bitmap */
2393                 info->bitmap = kmem_cache_zalloc(btrfs_free_space_bitmap_cachep,
2394                                                  GFP_NOFS);
2395                 info->trim_state = BTRFS_TRIM_STATE_TRIMMED;
2396                 spin_lock(&ctl->tree_lock);
2397                 if (!info->bitmap) {
2398                         ret = -ENOMEM;
2399                         goto out;
2400                 }
2401                 goto again;
2402         }
2403
2404 out:
2405         if (info) {
2406                 if (info->bitmap)
2407                         kmem_cache_free(btrfs_free_space_bitmap_cachep,
2408                                         info->bitmap);
2409                 kmem_cache_free(btrfs_free_space_cachep, info);
2410         }
2411
2412         return ret;
2413 }
2414
2415 /*
2416  * Free space merging rules:
2417  *  1) Merge trimmed areas together
2418  *  2) Let untrimmed areas coalesce with trimmed areas
2419  *  3) Always pull neighboring regions from bitmaps
2420  *
2421  * The above rules are for when we merge free space based on btrfs_trim_state.
2422  * Rules 2 and 3 are subtle because they are suboptimal, but are done for the
2423  * same reason: to promote larger extent regions which makes life easier for
2424  * find_free_extent().  Rule 2 enables coalescing based on the common path
2425  * being returning free space from btrfs_finish_extent_commit().  So when free
2426  * space is trimmed, it will prevent aggregating trimmed new region and
2427  * untrimmed regions in the rb_tree.  Rule 3 is purely to obtain larger extents
2428  * and provide find_free_extent() with the largest extents possible hoping for
2429  * the reuse path.
2430  */
2431 static bool try_merge_free_space(struct btrfs_free_space_ctl *ctl,
2432                           struct btrfs_free_space *info, bool update_stat)
2433 {
2434         struct btrfs_free_space *left_info = NULL;
2435         struct btrfs_free_space *right_info;
2436         bool merged = false;
2437         u64 offset = info->offset;
2438         u64 bytes = info->bytes;
2439         const bool is_trimmed = btrfs_free_space_trimmed(info);
2440
2441         /*
2442          * first we want to see if there is free space adjacent to the range we
2443          * are adding, if there is remove that struct and add a new one to
2444          * cover the entire range
2445          */
2446         right_info = tree_search_offset(ctl, offset + bytes, 0, 0);
2447         if (right_info && rb_prev(&right_info->offset_index))
2448                 left_info = rb_entry(rb_prev(&right_info->offset_index),
2449                                      struct btrfs_free_space, offset_index);
2450         else if (!right_info)
2451                 left_info = tree_search_offset(ctl, offset - 1, 0, 0);
2452
2453         /* See try_merge_free_space() comment. */
2454         if (right_info && !right_info->bitmap &&
2455             (!is_trimmed || btrfs_free_space_trimmed(right_info))) {
2456                 unlink_free_space(ctl, right_info, update_stat);
2457                 info->bytes += right_info->bytes;
2458                 kmem_cache_free(btrfs_free_space_cachep, right_info);
2459                 merged = true;
2460         }
2461
2462         /* See try_merge_free_space() comment. */
2463         if (left_info && !left_info->bitmap &&
2464             left_info->offset + left_info->bytes == offset &&
2465             (!is_trimmed || btrfs_free_space_trimmed(left_info))) {
2466                 unlink_free_space(ctl, left_info, update_stat);
2467                 info->offset = left_info->offset;
2468                 info->bytes += left_info->bytes;
2469                 kmem_cache_free(btrfs_free_space_cachep, left_info);
2470                 merged = true;
2471         }
2472
2473         return merged;
2474 }
2475
2476 static bool steal_from_bitmap_to_end(struct btrfs_free_space_ctl *ctl,
2477                                      struct btrfs_free_space *info,
2478                                      bool update_stat)
2479 {
2480         struct btrfs_free_space *bitmap;
2481         unsigned long i;
2482         unsigned long j;
2483         const u64 end = info->offset + info->bytes;
2484         const u64 bitmap_offset = offset_to_bitmap(ctl, end);
2485         u64 bytes;
2486
2487         bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2488         if (!bitmap)
2489                 return false;
2490
2491         i = offset_to_bit(bitmap->offset, ctl->unit, end);
2492         j = find_next_zero_bit(bitmap->bitmap, BITS_PER_BITMAP, i);
2493         if (j == i)
2494                 return false;
2495         bytes = (j - i) * ctl->unit;
2496         info->bytes += bytes;
2497
2498         /* See try_merge_free_space() comment. */
2499         if (!btrfs_free_space_trimmed(bitmap))
2500                 info->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2501
2502         bitmap_clear_bits(ctl, bitmap, end, bytes, update_stat);
2503
2504         if (!bitmap->bytes)
2505                 free_bitmap(ctl, bitmap);
2506
2507         return true;
2508 }
2509
2510 static bool steal_from_bitmap_to_front(struct btrfs_free_space_ctl *ctl,
2511                                        struct btrfs_free_space *info,
2512                                        bool update_stat)
2513 {
2514         struct btrfs_free_space *bitmap;
2515         u64 bitmap_offset;
2516         unsigned long i;
2517         unsigned long j;
2518         unsigned long prev_j;
2519         u64 bytes;
2520
2521         bitmap_offset = offset_to_bitmap(ctl, info->offset);
2522         /* If we're on a boundary, try the previous logical bitmap. */
2523         if (bitmap_offset == info->offset) {
2524                 if (info->offset == 0)
2525                         return false;
2526                 bitmap_offset = offset_to_bitmap(ctl, info->offset - 1);
2527         }
2528
2529         bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2530         if (!bitmap)
2531                 return false;
2532
2533         i = offset_to_bit(bitmap->offset, ctl->unit, info->offset) - 1;
2534         j = 0;
2535         prev_j = (unsigned long)-1;
2536         for_each_clear_bit_from(j, bitmap->bitmap, BITS_PER_BITMAP) {
2537                 if (j > i)
2538                         break;
2539                 prev_j = j;
2540         }
2541         if (prev_j == i)
2542                 return false;
2543
2544         if (prev_j == (unsigned long)-1)
2545                 bytes = (i + 1) * ctl->unit;
2546         else
2547                 bytes = (i - prev_j) * ctl->unit;
2548
2549         info->offset -= bytes;
2550         info->bytes += bytes;
2551
2552         /* See try_merge_free_space() comment. */
2553         if (!btrfs_free_space_trimmed(bitmap))
2554                 info->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2555
2556         bitmap_clear_bits(ctl, bitmap, info->offset, bytes, update_stat);
2557
2558         if (!bitmap->bytes)
2559                 free_bitmap(ctl, bitmap);
2560
2561         return true;
2562 }
2563
2564 /*
2565  * We prefer always to allocate from extent entries, both for clustered and
2566  * non-clustered allocation requests. So when attempting to add a new extent
2567  * entry, try to see if there's adjacent free space in bitmap entries, and if
2568  * there is, migrate that space from the bitmaps to the extent.
2569  * Like this we get better chances of satisfying space allocation requests
2570  * because we attempt to satisfy them based on a single cache entry, and never
2571  * on 2 or more entries - even if the entries represent a contiguous free space
2572  * region (e.g. 1 extent entry + 1 bitmap entry starting where the extent entry
2573  * ends).
2574  */
2575 static void steal_from_bitmap(struct btrfs_free_space_ctl *ctl,
2576                               struct btrfs_free_space *info,
2577                               bool update_stat)
2578 {
2579         /*
2580          * Only work with disconnected entries, as we can change their offset,
2581          * and must be extent entries.
2582          */
2583         ASSERT(!info->bitmap);
2584         ASSERT(RB_EMPTY_NODE(&info->offset_index));
2585
2586         if (ctl->total_bitmaps > 0) {
2587                 bool stole_end;
2588                 bool stole_front = false;
2589
2590                 stole_end = steal_from_bitmap_to_end(ctl, info, update_stat);
2591                 if (ctl->total_bitmaps > 0)
2592                         stole_front = steal_from_bitmap_to_front(ctl, info,
2593                                                                  update_stat);
2594
2595                 if (stole_end || stole_front)
2596                         try_merge_free_space(ctl, info, update_stat);
2597         }
2598 }
2599
2600 int __btrfs_add_free_space(struct btrfs_block_group *block_group,
2601                            u64 offset, u64 bytes,
2602                            enum btrfs_trim_state trim_state)
2603 {
2604         struct btrfs_fs_info *fs_info = block_group->fs_info;
2605         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2606         struct btrfs_free_space *info;
2607         int ret = 0;
2608         u64 filter_bytes = bytes;
2609
2610         ASSERT(!btrfs_is_zoned(fs_info));
2611
2612         info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
2613         if (!info)
2614                 return -ENOMEM;
2615
2616         info->offset = offset;
2617         info->bytes = bytes;
2618         info->trim_state = trim_state;
2619         RB_CLEAR_NODE(&info->offset_index);
2620         RB_CLEAR_NODE(&info->bytes_index);
2621
2622         spin_lock(&ctl->tree_lock);
2623
2624         if (try_merge_free_space(ctl, info, true))
2625                 goto link;
2626
2627         /*
2628          * There was no extent directly to the left or right of this new
2629          * extent then we know we're going to have to allocate a new extent, so
2630          * before we do that see if we need to drop this into a bitmap
2631          */
2632         ret = insert_into_bitmap(ctl, info);
2633         if (ret < 0) {
2634                 goto out;
2635         } else if (ret) {
2636                 ret = 0;
2637                 goto out;
2638         }
2639 link:
2640         /*
2641          * Only steal free space from adjacent bitmaps if we're sure we're not
2642          * going to add the new free space to existing bitmap entries - because
2643          * that would mean unnecessary work that would be reverted. Therefore
2644          * attempt to steal space from bitmaps if we're adding an extent entry.
2645          */
2646         steal_from_bitmap(ctl, info, true);
2647
2648         filter_bytes = max(filter_bytes, info->bytes);
2649
2650         ret = link_free_space(ctl, info);
2651         if (ret)
2652                 kmem_cache_free(btrfs_free_space_cachep, info);
2653 out:
2654         btrfs_discard_update_discardable(block_group);
2655         spin_unlock(&ctl->tree_lock);
2656
2657         if (ret) {
2658                 btrfs_crit(fs_info, "unable to add free space :%d", ret);
2659                 ASSERT(ret != -EEXIST);
2660         }
2661
2662         if (trim_state != BTRFS_TRIM_STATE_TRIMMED) {
2663                 btrfs_discard_check_filter(block_group, filter_bytes);
2664                 btrfs_discard_queue_work(&fs_info->discard_ctl, block_group);
2665         }
2666
2667         return ret;
2668 }
2669
2670 static int __btrfs_add_free_space_zoned(struct btrfs_block_group *block_group,
2671                                         u64 bytenr, u64 size, bool used)
2672 {
2673         struct btrfs_space_info *sinfo = block_group->space_info;
2674         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2675         u64 offset = bytenr - block_group->start;
2676         u64 to_free, to_unusable;
2677         int bg_reclaim_threshold = 0;
2678         bool initial = (size == block_group->length);
2679         u64 reclaimable_unusable;
2680
2681         WARN_ON(!initial && offset + size > block_group->zone_capacity);
2682
2683         if (!initial)
2684                 bg_reclaim_threshold = READ_ONCE(sinfo->bg_reclaim_threshold);
2685
2686         spin_lock(&ctl->tree_lock);
2687         if (!used)
2688                 to_free = size;
2689         else if (initial)
2690                 to_free = block_group->zone_capacity;
2691         else if (offset >= block_group->alloc_offset)
2692                 to_free = size;
2693         else if (offset + size <= block_group->alloc_offset)
2694                 to_free = 0;
2695         else
2696                 to_free = offset + size - block_group->alloc_offset;
2697         to_unusable = size - to_free;
2698
2699         ctl->free_space += to_free;
2700         /*
2701          * If the block group is read-only, we should account freed space into
2702          * bytes_readonly.
2703          */
2704         if (!block_group->ro)
2705                 block_group->zone_unusable += to_unusable;
2706         spin_unlock(&ctl->tree_lock);
2707         if (!used) {
2708                 spin_lock(&block_group->lock);
2709                 block_group->alloc_offset -= size;
2710                 spin_unlock(&block_group->lock);
2711         }
2712
2713         reclaimable_unusable = block_group->zone_unusable -
2714                                (block_group->length - block_group->zone_capacity);
2715         /* All the region is now unusable. Mark it as unused and reclaim */
2716         if (block_group->zone_unusable == block_group->length) {
2717                 btrfs_mark_bg_unused(block_group);
2718         } else if (bg_reclaim_threshold &&
2719                    reclaimable_unusable >=
2720                    div_factor_fine(block_group->zone_capacity,
2721                                    bg_reclaim_threshold)) {
2722                 btrfs_mark_bg_to_reclaim(block_group);
2723         }
2724
2725         return 0;
2726 }
2727
2728 int btrfs_add_free_space(struct btrfs_block_group *block_group,
2729                          u64 bytenr, u64 size)
2730 {
2731         enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2732
2733         if (btrfs_is_zoned(block_group->fs_info))
2734                 return __btrfs_add_free_space_zoned(block_group, bytenr, size,
2735                                                     true);
2736
2737         if (btrfs_test_opt(block_group->fs_info, DISCARD_SYNC))
2738                 trim_state = BTRFS_TRIM_STATE_TRIMMED;
2739
2740         return __btrfs_add_free_space(block_group, bytenr, size, trim_state);
2741 }
2742
2743 int btrfs_add_free_space_unused(struct btrfs_block_group *block_group,
2744                                 u64 bytenr, u64 size)
2745 {
2746         if (btrfs_is_zoned(block_group->fs_info))
2747                 return __btrfs_add_free_space_zoned(block_group, bytenr, size,
2748                                                     false);
2749
2750         return btrfs_add_free_space(block_group, bytenr, size);
2751 }
2752
2753 /*
2754  * This is a subtle distinction because when adding free space back in general,
2755  * we want it to be added as untrimmed for async. But in the case where we add
2756  * it on loading of a block group, we want to consider it trimmed.
2757  */
2758 int btrfs_add_free_space_async_trimmed(struct btrfs_block_group *block_group,
2759                                        u64 bytenr, u64 size)
2760 {
2761         enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2762
2763         if (btrfs_is_zoned(block_group->fs_info))
2764                 return __btrfs_add_free_space_zoned(block_group, bytenr, size,
2765                                                     true);
2766
2767         if (btrfs_test_opt(block_group->fs_info, DISCARD_SYNC) ||
2768             btrfs_test_opt(block_group->fs_info, DISCARD_ASYNC))
2769                 trim_state = BTRFS_TRIM_STATE_TRIMMED;
2770
2771         return __btrfs_add_free_space(block_group, bytenr, size, trim_state);
2772 }
2773
2774 int btrfs_remove_free_space(struct btrfs_block_group *block_group,
2775                             u64 offset, u64 bytes)
2776 {
2777         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2778         struct btrfs_free_space *info;
2779         int ret;
2780         bool re_search = false;
2781
2782         if (btrfs_is_zoned(block_group->fs_info)) {
2783                 /*
2784                  * This can happen with conventional zones when replaying log.
2785                  * Since the allocation info of tree-log nodes are not recorded
2786                  * to the extent-tree, calculate_alloc_pointer() failed to
2787                  * advance the allocation pointer after last allocated tree log
2788                  * node blocks.
2789                  *
2790                  * This function is called from
2791                  * btrfs_pin_extent_for_log_replay() when replaying the log.
2792                  * Advance the pointer not to overwrite the tree-log nodes.
2793                  */
2794                 if (block_group->start + block_group->alloc_offset <
2795                     offset + bytes) {
2796                         block_group->alloc_offset =
2797                                 offset + bytes - block_group->start;
2798                 }
2799                 return 0;
2800         }
2801
2802         spin_lock(&ctl->tree_lock);
2803
2804 again:
2805         ret = 0;
2806         if (!bytes)
2807                 goto out_lock;
2808
2809         info = tree_search_offset(ctl, offset, 0, 0);
2810         if (!info) {
2811                 /*
2812                  * oops didn't find an extent that matched the space we wanted
2813                  * to remove, look for a bitmap instead
2814                  */
2815                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2816                                           1, 0);
2817                 if (!info) {
2818                         /*
2819                          * If we found a partial bit of our free space in a
2820                          * bitmap but then couldn't find the other part this may
2821                          * be a problem, so WARN about it.
2822                          */
2823                         WARN_ON(re_search);
2824                         goto out_lock;
2825                 }
2826         }
2827
2828         re_search = false;
2829         if (!info->bitmap) {
2830                 unlink_free_space(ctl, info, true);
2831                 if (offset == info->offset) {
2832                         u64 to_free = min(bytes, info->bytes);
2833
2834                         info->bytes -= to_free;
2835                         info->offset += to_free;
2836                         if (info->bytes) {
2837                                 ret = link_free_space(ctl, info);
2838                                 WARN_ON(ret);
2839                         } else {
2840                                 kmem_cache_free(btrfs_free_space_cachep, info);
2841                         }
2842
2843                         offset += to_free;
2844                         bytes -= to_free;
2845                         goto again;
2846                 } else {
2847                         u64 old_end = info->bytes + info->offset;
2848
2849                         info->bytes = offset - info->offset;
2850                         ret = link_free_space(ctl, info);
2851                         WARN_ON(ret);
2852                         if (ret)
2853                                 goto out_lock;
2854
2855                         /* Not enough bytes in this entry to satisfy us */
2856                         if (old_end < offset + bytes) {
2857                                 bytes -= old_end - offset;
2858                                 offset = old_end;
2859                                 goto again;
2860                         } else if (old_end == offset + bytes) {
2861                                 /* all done */
2862                                 goto out_lock;
2863                         }
2864                         spin_unlock(&ctl->tree_lock);
2865
2866                         ret = __btrfs_add_free_space(block_group,
2867                                                      offset + bytes,
2868                                                      old_end - (offset + bytes),
2869                                                      info->trim_state);
2870                         WARN_ON(ret);
2871                         goto out;
2872                 }
2873         }
2874
2875         ret = remove_from_bitmap(ctl, info, &offset, &bytes);
2876         if (ret == -EAGAIN) {
2877                 re_search = true;
2878                 goto again;
2879         }
2880 out_lock:
2881         btrfs_discard_update_discardable(block_group);
2882         spin_unlock(&ctl->tree_lock);
2883 out:
2884         return ret;
2885 }
2886
2887 void btrfs_dump_free_space(struct btrfs_block_group *block_group,
2888                            u64 bytes)
2889 {
2890         struct btrfs_fs_info *fs_info = block_group->fs_info;
2891         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2892         struct btrfs_free_space *info;
2893         struct rb_node *n;
2894         int count = 0;
2895
2896         /*
2897          * Zoned btrfs does not use free space tree and cluster. Just print
2898          * out the free space after the allocation offset.
2899          */
2900         if (btrfs_is_zoned(fs_info)) {
2901                 btrfs_info(fs_info, "free space %llu active %d",
2902                            block_group->zone_capacity - block_group->alloc_offset,
2903                            test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE,
2904                                     &block_group->runtime_flags));
2905                 return;
2906         }
2907
2908         spin_lock(&ctl->tree_lock);
2909         for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
2910                 info = rb_entry(n, struct btrfs_free_space, offset_index);
2911                 if (info->bytes >= bytes && !block_group->ro)
2912                         count++;
2913                 btrfs_crit(fs_info, "entry offset %llu, bytes %llu, bitmap %s",
2914                            info->offset, info->bytes,
2915                        (info->bitmap) ? "yes" : "no");
2916         }
2917         spin_unlock(&ctl->tree_lock);
2918         btrfs_info(fs_info, "block group has cluster?: %s",
2919                list_empty(&block_group->cluster_list) ? "no" : "yes");
2920         btrfs_info(fs_info,
2921                    "%d blocks of free space at or bigger than bytes is", count);
2922 }
2923
2924 void btrfs_init_free_space_ctl(struct btrfs_block_group *block_group,
2925                                struct btrfs_free_space_ctl *ctl)
2926 {
2927         struct btrfs_fs_info *fs_info = block_group->fs_info;
2928
2929         spin_lock_init(&ctl->tree_lock);
2930         ctl->unit = fs_info->sectorsize;
2931         ctl->start = block_group->start;
2932         ctl->block_group = block_group;
2933         ctl->op = &free_space_op;
2934         ctl->free_space_bytes = RB_ROOT_CACHED;
2935         INIT_LIST_HEAD(&ctl->trimming_ranges);
2936         mutex_init(&ctl->cache_writeout_mutex);
2937
2938         /*
2939          * we only want to have 32k of ram per block group for keeping
2940          * track of free space, and if we pass 1/2 of that we want to
2941          * start converting things over to using bitmaps
2942          */
2943         ctl->extents_thresh = (SZ_32K / 2) / sizeof(struct btrfs_free_space);
2944 }
2945
2946 /*
2947  * for a given cluster, put all of its extents back into the free
2948  * space cache.  If the block group passed doesn't match the block group
2949  * pointed to by the cluster, someone else raced in and freed the
2950  * cluster already.  In that case, we just return without changing anything
2951  */
2952 static void __btrfs_return_cluster_to_free_space(
2953                              struct btrfs_block_group *block_group,
2954                              struct btrfs_free_cluster *cluster)
2955 {
2956         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2957         struct btrfs_free_space *entry;
2958         struct rb_node *node;
2959
2960         spin_lock(&cluster->lock);
2961         if (cluster->block_group != block_group) {
2962                 spin_unlock(&cluster->lock);
2963                 return;
2964         }
2965
2966         cluster->block_group = NULL;
2967         cluster->window_start = 0;
2968         list_del_init(&cluster->block_group_list);
2969
2970         node = rb_first(&cluster->root);
2971         while (node) {
2972                 bool bitmap;
2973
2974                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2975                 node = rb_next(&entry->offset_index);
2976                 rb_erase(&entry->offset_index, &cluster->root);
2977                 RB_CLEAR_NODE(&entry->offset_index);
2978
2979                 bitmap = (entry->bitmap != NULL);
2980                 if (!bitmap) {
2981                         /* Merging treats extents as if they were new */
2982                         if (!btrfs_free_space_trimmed(entry)) {
2983                                 ctl->discardable_extents[BTRFS_STAT_CURR]--;
2984                                 ctl->discardable_bytes[BTRFS_STAT_CURR] -=
2985                                         entry->bytes;
2986                         }
2987
2988                         try_merge_free_space(ctl, entry, false);
2989                         steal_from_bitmap(ctl, entry, false);
2990
2991                         /* As we insert directly, update these statistics */
2992                         if (!btrfs_free_space_trimmed(entry)) {
2993                                 ctl->discardable_extents[BTRFS_STAT_CURR]++;
2994                                 ctl->discardable_bytes[BTRFS_STAT_CURR] +=
2995                                         entry->bytes;
2996                         }
2997                 }
2998                 tree_insert_offset(&ctl->free_space_offset,
2999                                    entry->offset, &entry->offset_index, bitmap);
3000                 rb_add_cached(&entry->bytes_index, &ctl->free_space_bytes,
3001                               entry_less);
3002         }
3003         cluster->root = RB_ROOT;
3004         spin_unlock(&cluster->lock);
3005         btrfs_put_block_group(block_group);
3006 }
3007
3008 void btrfs_remove_free_space_cache(struct btrfs_block_group *block_group)
3009 {
3010         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3011         struct btrfs_free_cluster *cluster;
3012         struct list_head *head;
3013
3014         spin_lock(&ctl->tree_lock);
3015         while ((head = block_group->cluster_list.next) !=
3016                &block_group->cluster_list) {
3017                 cluster = list_entry(head, struct btrfs_free_cluster,
3018                                      block_group_list);
3019
3020                 WARN_ON(cluster->block_group != block_group);
3021                 __btrfs_return_cluster_to_free_space(block_group, cluster);
3022
3023                 cond_resched_lock(&ctl->tree_lock);
3024         }
3025         __btrfs_remove_free_space_cache(ctl);
3026         btrfs_discard_update_discardable(block_group);
3027         spin_unlock(&ctl->tree_lock);
3028
3029 }
3030
3031 /**
3032  * btrfs_is_free_space_trimmed - see if everything is trimmed
3033  * @block_group: block_group of interest
3034  *
3035  * Walk @block_group's free space rb_tree to determine if everything is trimmed.
3036  */
3037 bool btrfs_is_free_space_trimmed(struct btrfs_block_group *block_group)
3038 {
3039         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3040         struct btrfs_free_space *info;
3041         struct rb_node *node;
3042         bool ret = true;
3043
3044         spin_lock(&ctl->tree_lock);
3045         node = rb_first(&ctl->free_space_offset);
3046
3047         while (node) {
3048                 info = rb_entry(node, struct btrfs_free_space, offset_index);
3049
3050                 if (!btrfs_free_space_trimmed(info)) {
3051                         ret = false;
3052                         break;
3053                 }
3054
3055                 node = rb_next(node);
3056         }
3057
3058         spin_unlock(&ctl->tree_lock);
3059         return ret;
3060 }
3061
3062 u64 btrfs_find_space_for_alloc(struct btrfs_block_group *block_group,
3063                                u64 offset, u64 bytes, u64 empty_size,
3064                                u64 *max_extent_size)
3065 {
3066         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3067         struct btrfs_discard_ctl *discard_ctl =
3068                                         &block_group->fs_info->discard_ctl;
3069         struct btrfs_free_space *entry = NULL;
3070         u64 bytes_search = bytes + empty_size;
3071         u64 ret = 0;
3072         u64 align_gap = 0;
3073         u64 align_gap_len = 0;
3074         enum btrfs_trim_state align_gap_trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3075         bool use_bytes_index = (offset == block_group->start);
3076
3077         ASSERT(!btrfs_is_zoned(block_group->fs_info));
3078
3079         spin_lock(&ctl->tree_lock);
3080         entry = find_free_space(ctl, &offset, &bytes_search,
3081                                 block_group->full_stripe_len, max_extent_size,
3082                                 use_bytes_index);
3083         if (!entry)
3084                 goto out;
3085
3086         ret = offset;
3087         if (entry->bitmap) {
3088                 bitmap_clear_bits(ctl, entry, offset, bytes, true);
3089
3090                 if (!btrfs_free_space_trimmed(entry))
3091                         atomic64_add(bytes, &discard_ctl->discard_bytes_saved);
3092
3093                 if (!entry->bytes)
3094                         free_bitmap(ctl, entry);
3095         } else {
3096                 unlink_free_space(ctl, entry, true);
3097                 align_gap_len = offset - entry->offset;
3098                 align_gap = entry->offset;
3099                 align_gap_trim_state = entry->trim_state;
3100
3101                 if (!btrfs_free_space_trimmed(entry))
3102                         atomic64_add(bytes, &discard_ctl->discard_bytes_saved);
3103
3104                 entry->offset = offset + bytes;
3105                 WARN_ON(entry->bytes < bytes + align_gap_len);
3106
3107                 entry->bytes -= bytes + align_gap_len;
3108                 if (!entry->bytes)
3109                         kmem_cache_free(btrfs_free_space_cachep, entry);
3110                 else
3111                         link_free_space(ctl, entry);
3112         }
3113 out:
3114         btrfs_discard_update_discardable(block_group);
3115         spin_unlock(&ctl->tree_lock);
3116
3117         if (align_gap_len)
3118                 __btrfs_add_free_space(block_group, align_gap, align_gap_len,
3119                                        align_gap_trim_state);
3120         return ret;
3121 }
3122
3123 /*
3124  * given a cluster, put all of its extents back into the free space
3125  * cache.  If a block group is passed, this function will only free
3126  * a cluster that belongs to the passed block group.
3127  *
3128  * Otherwise, it'll get a reference on the block group pointed to by the
3129  * cluster and remove the cluster from it.
3130  */
3131 void btrfs_return_cluster_to_free_space(
3132                                struct btrfs_block_group *block_group,
3133                                struct btrfs_free_cluster *cluster)
3134 {
3135         struct btrfs_free_space_ctl *ctl;
3136
3137         /* first, get a safe pointer to the block group */
3138         spin_lock(&cluster->lock);
3139         if (!block_group) {
3140                 block_group = cluster->block_group;
3141                 if (!block_group) {
3142                         spin_unlock(&cluster->lock);
3143                         return;
3144                 }
3145         } else if (cluster->block_group != block_group) {
3146                 /* someone else has already freed it don't redo their work */
3147                 spin_unlock(&cluster->lock);
3148                 return;
3149         }
3150         btrfs_get_block_group(block_group);
3151         spin_unlock(&cluster->lock);
3152
3153         ctl = block_group->free_space_ctl;
3154
3155         /* now return any extents the cluster had on it */
3156         spin_lock(&ctl->tree_lock);
3157         __btrfs_return_cluster_to_free_space(block_group, cluster);
3158         spin_unlock(&ctl->tree_lock);
3159
3160         btrfs_discard_queue_work(&block_group->fs_info->discard_ctl, block_group);
3161
3162         /* finally drop our ref */
3163         btrfs_put_block_group(block_group);
3164 }
3165
3166 static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group *block_group,
3167                                    struct btrfs_free_cluster *cluster,
3168                                    struct btrfs_free_space *entry,
3169                                    u64 bytes, u64 min_start,
3170                                    u64 *max_extent_size)
3171 {
3172         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3173         int err;
3174         u64 search_start = cluster->window_start;
3175         u64 search_bytes = bytes;
3176         u64 ret = 0;
3177
3178         search_start = min_start;
3179         search_bytes = bytes;
3180
3181         err = search_bitmap(ctl, entry, &search_start, &search_bytes, true);
3182         if (err) {
3183                 *max_extent_size = max(get_max_extent_size(entry),
3184                                        *max_extent_size);
3185                 return 0;
3186         }
3187
3188         ret = search_start;
3189         bitmap_clear_bits(ctl, entry, ret, bytes, false);
3190
3191         return ret;
3192 }
3193
3194 /*
3195  * given a cluster, try to allocate 'bytes' from it, returns 0
3196  * if it couldn't find anything suitably large, or a logical disk offset
3197  * if things worked out
3198  */
3199 u64 btrfs_alloc_from_cluster(struct btrfs_block_group *block_group,
3200                              struct btrfs_free_cluster *cluster, u64 bytes,
3201                              u64 min_start, u64 *max_extent_size)
3202 {
3203         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3204         struct btrfs_discard_ctl *discard_ctl =
3205                                         &block_group->fs_info->discard_ctl;
3206         struct btrfs_free_space *entry = NULL;
3207         struct rb_node *node;
3208         u64 ret = 0;
3209
3210         ASSERT(!btrfs_is_zoned(block_group->fs_info));
3211
3212         spin_lock(&cluster->lock);
3213         if (bytes > cluster->max_size)
3214                 goto out;
3215
3216         if (cluster->block_group != block_group)
3217                 goto out;
3218
3219         node = rb_first(&cluster->root);
3220         if (!node)
3221                 goto out;
3222
3223         entry = rb_entry(node, struct btrfs_free_space, offset_index);
3224         while (1) {
3225                 if (entry->bytes < bytes)
3226                         *max_extent_size = max(get_max_extent_size(entry),
3227                                                *max_extent_size);
3228
3229                 if (entry->bytes < bytes ||
3230                     (!entry->bitmap && entry->offset < min_start)) {
3231                         node = rb_next(&entry->offset_index);
3232                         if (!node)
3233                                 break;
3234                         entry = rb_entry(node, struct btrfs_free_space,
3235                                          offset_index);
3236                         continue;
3237                 }
3238
3239                 if (entry->bitmap) {
3240                         ret = btrfs_alloc_from_bitmap(block_group,
3241                                                       cluster, entry, bytes,
3242                                                       cluster->window_start,
3243                                                       max_extent_size);
3244                         if (ret == 0) {
3245                                 node = rb_next(&entry->offset_index);
3246                                 if (!node)
3247                                         break;
3248                                 entry = rb_entry(node, struct btrfs_free_space,
3249                                                  offset_index);
3250                                 continue;
3251                         }
3252                         cluster->window_start += bytes;
3253                 } else {
3254                         ret = entry->offset;
3255
3256                         entry->offset += bytes;
3257                         entry->bytes -= bytes;
3258                 }
3259
3260                 break;
3261         }
3262 out:
3263         spin_unlock(&cluster->lock);
3264
3265         if (!ret)
3266                 return 0;
3267
3268         spin_lock(&ctl->tree_lock);
3269
3270         if (!btrfs_free_space_trimmed(entry))
3271                 atomic64_add(bytes, &discard_ctl->discard_bytes_saved);
3272
3273         ctl->free_space -= bytes;
3274         if (!entry->bitmap && !btrfs_free_space_trimmed(entry))
3275                 ctl->discardable_bytes[BTRFS_STAT_CURR] -= bytes;
3276
3277         spin_lock(&cluster->lock);
3278         if (entry->bytes == 0) {
3279                 rb_erase(&entry->offset_index, &cluster->root);
3280                 ctl->free_extents--;
3281                 if (entry->bitmap) {
3282                         kmem_cache_free(btrfs_free_space_bitmap_cachep,
3283                                         entry->bitmap);
3284                         ctl->total_bitmaps--;
3285                         recalculate_thresholds(ctl);
3286                 } else if (!btrfs_free_space_trimmed(entry)) {
3287                         ctl->discardable_extents[BTRFS_STAT_CURR]--;
3288                 }
3289                 kmem_cache_free(btrfs_free_space_cachep, entry);
3290         }
3291
3292         spin_unlock(&cluster->lock);
3293         spin_unlock(&ctl->tree_lock);
3294
3295         return ret;
3296 }
3297
3298 static int btrfs_bitmap_cluster(struct btrfs_block_group *block_group,
3299                                 struct btrfs_free_space *entry,
3300                                 struct btrfs_free_cluster *cluster,
3301                                 u64 offset, u64 bytes,
3302                                 u64 cont1_bytes, u64 min_bytes)
3303 {
3304         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3305         unsigned long next_zero;
3306         unsigned long i;
3307         unsigned long want_bits;
3308         unsigned long min_bits;
3309         unsigned long found_bits;
3310         unsigned long max_bits = 0;
3311         unsigned long start = 0;
3312         unsigned long total_found = 0;
3313         int ret;
3314
3315         i = offset_to_bit(entry->offset, ctl->unit,
3316                           max_t(u64, offset, entry->offset));
3317         want_bits = bytes_to_bits(bytes, ctl->unit);
3318         min_bits = bytes_to_bits(min_bytes, ctl->unit);
3319
3320         /*
3321          * Don't bother looking for a cluster in this bitmap if it's heavily
3322          * fragmented.
3323          */
3324         if (entry->max_extent_size &&
3325             entry->max_extent_size < cont1_bytes)
3326                 return -ENOSPC;
3327 again:
3328         found_bits = 0;
3329         for_each_set_bit_from(i, entry->bitmap, BITS_PER_BITMAP) {
3330                 next_zero = find_next_zero_bit(entry->bitmap,
3331                                                BITS_PER_BITMAP, i);
3332                 if (next_zero - i >= min_bits) {
3333                         found_bits = next_zero - i;
3334                         if (found_bits > max_bits)
3335                                 max_bits = found_bits;
3336                         break;
3337                 }
3338                 if (next_zero - i > max_bits)
3339                         max_bits = next_zero - i;
3340                 i = next_zero;
3341         }
3342
3343         if (!found_bits) {
3344                 entry->max_extent_size = (u64)max_bits * ctl->unit;
3345                 return -ENOSPC;
3346         }
3347
3348         if (!total_found) {
3349                 start = i;
3350                 cluster->max_size = 0;
3351         }
3352
3353         total_found += found_bits;
3354
3355         if (cluster->max_size < found_bits * ctl->unit)
3356                 cluster->max_size = found_bits * ctl->unit;
3357
3358         if (total_found < want_bits || cluster->max_size < cont1_bytes) {
3359                 i = next_zero + 1;
3360                 goto again;
3361         }
3362
3363         cluster->window_start = start * ctl->unit + entry->offset;
3364         rb_erase(&entry->offset_index, &ctl->free_space_offset);
3365         rb_erase_cached(&entry->bytes_index, &ctl->free_space_bytes);
3366
3367         /*
3368          * We need to know if we're currently on the normal space index when we
3369          * manipulate the bitmap so that we know we need to remove and re-insert
3370          * it into the space_index tree.  Clear the bytes_index node here so the
3371          * bitmap manipulation helpers know not to mess with the space_index
3372          * until this bitmap entry is added back into the normal cache.
3373          */
3374         RB_CLEAR_NODE(&entry->bytes_index);
3375
3376         ret = tree_insert_offset(&cluster->root, entry->offset,
3377                                  &entry->offset_index, 1);
3378         ASSERT(!ret); /* -EEXIST; Logic error */
3379
3380         trace_btrfs_setup_cluster(block_group, cluster,
3381                                   total_found * ctl->unit, 1);
3382         return 0;
3383 }
3384
3385 /*
3386  * This searches the block group for just extents to fill the cluster with.
3387  * Try to find a cluster with at least bytes total bytes, at least one
3388  * extent of cont1_bytes, and other clusters of at least min_bytes.
3389  */
3390 static noinline int
3391 setup_cluster_no_bitmap(struct btrfs_block_group *block_group,
3392                         struct btrfs_free_cluster *cluster,
3393                         struct list_head *bitmaps, u64 offset, u64 bytes,
3394                         u64 cont1_bytes, u64 min_bytes)
3395 {
3396         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3397         struct btrfs_free_space *first = NULL;
3398         struct btrfs_free_space *entry = NULL;
3399         struct btrfs_free_space *last;
3400         struct rb_node *node;
3401         u64 window_free;
3402         u64 max_extent;
3403         u64 total_size = 0;
3404
3405         entry = tree_search_offset(ctl, offset, 0, 1);
3406         if (!entry)
3407                 return -ENOSPC;
3408
3409         /*
3410          * We don't want bitmaps, so just move along until we find a normal
3411          * extent entry.
3412          */
3413         while (entry->bitmap || entry->bytes < min_bytes) {
3414                 if (entry->bitmap && list_empty(&entry->list))
3415                         list_add_tail(&entry->list, bitmaps);
3416                 node = rb_next(&entry->offset_index);
3417                 if (!node)
3418                         return -ENOSPC;
3419                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3420         }
3421
3422         window_free = entry->bytes;
3423         max_extent = entry->bytes;
3424         first = entry;
3425         last = entry;
3426
3427         for (node = rb_next(&entry->offset_index); node;
3428              node = rb_next(&entry->offset_index)) {
3429                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3430
3431                 if (entry->bitmap) {
3432                         if (list_empty(&entry->list))
3433                                 list_add_tail(&entry->list, bitmaps);
3434                         continue;
3435                 }
3436
3437                 if (entry->bytes < min_bytes)
3438                         continue;
3439
3440                 last = entry;
3441                 window_free += entry->bytes;
3442                 if (entry->bytes > max_extent)
3443                         max_extent = entry->bytes;
3444         }
3445
3446         if (window_free < bytes || max_extent < cont1_bytes)
3447                 return -ENOSPC;
3448
3449         cluster->window_start = first->offset;
3450
3451         node = &first->offset_index;
3452
3453         /*
3454          * now we've found our entries, pull them out of the free space
3455          * cache and put them into the cluster rbtree
3456          */
3457         do {
3458                 int ret;
3459
3460                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3461                 node = rb_next(&entry->offset_index);
3462                 if (entry->bitmap || entry->bytes < min_bytes)
3463                         continue;
3464
3465                 rb_erase(&entry->offset_index, &ctl->free_space_offset);
3466                 rb_erase_cached(&entry->bytes_index, &ctl->free_space_bytes);
3467                 ret = tree_insert_offset(&cluster->root, entry->offset,
3468                                          &entry->offset_index, 0);
3469                 total_size += entry->bytes;
3470                 ASSERT(!ret); /* -EEXIST; Logic error */
3471         } while (node && entry != last);
3472
3473         cluster->max_size = max_extent;
3474         trace_btrfs_setup_cluster(block_group, cluster, total_size, 0);
3475         return 0;
3476 }
3477
3478 /*
3479  * This specifically looks for bitmaps that may work in the cluster, we assume
3480  * that we have already failed to find extents that will work.
3481  */
3482 static noinline int
3483 setup_cluster_bitmap(struct btrfs_block_group *block_group,
3484                      struct btrfs_free_cluster *cluster,
3485                      struct list_head *bitmaps, u64 offset, u64 bytes,
3486                      u64 cont1_bytes, u64 min_bytes)
3487 {
3488         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3489         struct btrfs_free_space *entry = NULL;
3490         int ret = -ENOSPC;
3491         u64 bitmap_offset = offset_to_bitmap(ctl, offset);
3492
3493         if (ctl->total_bitmaps == 0)
3494                 return -ENOSPC;
3495
3496         /*
3497          * The bitmap that covers offset won't be in the list unless offset
3498          * is just its start offset.
3499          */
3500         if (!list_empty(bitmaps))
3501                 entry = list_first_entry(bitmaps, struct btrfs_free_space, list);
3502
3503         if (!entry || entry->offset != bitmap_offset) {
3504                 entry = tree_search_offset(ctl, bitmap_offset, 1, 0);
3505                 if (entry && list_empty(&entry->list))
3506                         list_add(&entry->list, bitmaps);
3507         }
3508
3509         list_for_each_entry(entry, bitmaps, list) {
3510                 if (entry->bytes < bytes)
3511                         continue;
3512                 ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
3513                                            bytes, cont1_bytes, min_bytes);
3514                 if (!ret)
3515                         return 0;
3516         }
3517
3518         /*
3519          * The bitmaps list has all the bitmaps that record free space
3520          * starting after offset, so no more search is required.
3521          */
3522         return -ENOSPC;
3523 }
3524
3525 /*
3526  * here we try to find a cluster of blocks in a block group.  The goal
3527  * is to find at least bytes+empty_size.
3528  * We might not find them all in one contiguous area.
3529  *
3530  * returns zero and sets up cluster if things worked out, otherwise
3531  * it returns -enospc
3532  */
3533 int btrfs_find_space_cluster(struct btrfs_block_group *block_group,
3534                              struct btrfs_free_cluster *cluster,
3535                              u64 offset, u64 bytes, u64 empty_size)
3536 {
3537         struct btrfs_fs_info *fs_info = block_group->fs_info;
3538         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3539         struct btrfs_free_space *entry, *tmp;
3540         LIST_HEAD(bitmaps);
3541         u64 min_bytes;
3542         u64 cont1_bytes;
3543         int ret;
3544
3545         /*
3546          * Choose the minimum extent size we'll require for this
3547          * cluster.  For SSD_SPREAD, don't allow any fragmentation.
3548          * For metadata, allow allocates with smaller extents.  For
3549          * data, keep it dense.
3550          */
3551         if (btrfs_test_opt(fs_info, SSD_SPREAD)) {
3552                 cont1_bytes = bytes + empty_size;
3553                 min_bytes = cont1_bytes;
3554         } else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
3555                 cont1_bytes = bytes;
3556                 min_bytes = fs_info->sectorsize;
3557         } else {
3558                 cont1_bytes = max(bytes, (bytes + empty_size) >> 2);
3559                 min_bytes = fs_info->sectorsize;
3560         }
3561
3562         spin_lock(&ctl->tree_lock);
3563
3564         /*
3565          * If we know we don't have enough space to make a cluster don't even
3566          * bother doing all the work to try and find one.
3567          */
3568         if (ctl->free_space < bytes) {
3569                 spin_unlock(&ctl->tree_lock);
3570                 return -ENOSPC;
3571         }
3572
3573         spin_lock(&cluster->lock);
3574
3575         /* someone already found a cluster, hooray */
3576         if (cluster->block_group) {
3577                 ret = 0;
3578                 goto out;
3579         }
3580
3581         trace_btrfs_find_cluster(block_group, offset, bytes, empty_size,
3582                                  min_bytes);
3583
3584         ret = setup_cluster_no_bitmap(block_group, cluster, &bitmaps, offset,
3585                                       bytes + empty_size,
3586                                       cont1_bytes, min_bytes);
3587         if (ret)
3588                 ret = setup_cluster_bitmap(block_group, cluster, &bitmaps,
3589                                            offset, bytes + empty_size,
3590                                            cont1_bytes, min_bytes);
3591
3592         /* Clear our temporary list */
3593         list_for_each_entry_safe(entry, tmp, &bitmaps, list)
3594                 list_del_init(&entry->list);
3595
3596         if (!ret) {
3597                 btrfs_get_block_group(block_group);
3598                 list_add_tail(&cluster->block_group_list,
3599                               &block_group->cluster_list);
3600                 cluster->block_group = block_group;
3601         } else {
3602                 trace_btrfs_failed_cluster_setup(block_group);
3603         }
3604 out:
3605         spin_unlock(&cluster->lock);
3606         spin_unlock(&ctl->tree_lock);
3607
3608         return ret;
3609 }
3610
3611 /*
3612  * simple code to zero out a cluster
3613  */
3614 void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
3615 {
3616         spin_lock_init(&cluster->lock);
3617         spin_lock_init(&cluster->refill_lock);
3618         cluster->root = RB_ROOT;
3619         cluster->max_size = 0;
3620         cluster->fragmented = false;
3621         INIT_LIST_HEAD(&cluster->block_group_list);
3622         cluster->block_group = NULL;
3623 }
3624
3625 static int do_trimming(struct btrfs_block_group *block_group,
3626                        u64 *total_trimmed, u64 start, u64 bytes,
3627                        u64 reserved_start, u64 reserved_bytes,
3628                        enum btrfs_trim_state reserved_trim_state,
3629                        struct btrfs_trim_range *trim_entry)
3630 {
3631         struct btrfs_space_info *space_info = block_group->space_info;
3632         struct btrfs_fs_info *fs_info = block_group->fs_info;
3633         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3634         int ret;
3635         int update = 0;
3636         const u64 end = start + bytes;
3637         const u64 reserved_end = reserved_start + reserved_bytes;
3638         enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3639         u64 trimmed = 0;
3640
3641         spin_lock(&space_info->lock);
3642         spin_lock(&block_group->lock);
3643         if (!block_group->ro) {
3644                 block_group->reserved += reserved_bytes;
3645                 space_info->bytes_reserved += reserved_bytes;
3646                 update = 1;
3647         }
3648         spin_unlock(&block_group->lock);
3649         spin_unlock(&space_info->lock);
3650
3651         ret = btrfs_discard_extent(fs_info, start, bytes, &trimmed);
3652         if (!ret) {
3653                 *total_trimmed += trimmed;
3654                 trim_state = BTRFS_TRIM_STATE_TRIMMED;
3655         }
3656
3657         mutex_lock(&ctl->cache_writeout_mutex);
3658         if (reserved_start < start)
3659                 __btrfs_add_free_space(block_group, reserved_start,
3660                                        start - reserved_start,
3661                                        reserved_trim_state);
3662         if (start + bytes < reserved_start + reserved_bytes)
3663                 __btrfs_add_free_space(block_group, end, reserved_end - end,
3664                                        reserved_trim_state);
3665         __btrfs_add_free_space(block_group, start, bytes, trim_state);
3666         list_del(&trim_entry->list);
3667         mutex_unlock(&ctl->cache_writeout_mutex);
3668
3669         if (update) {
3670                 spin_lock(&space_info->lock);
3671                 spin_lock(&block_group->lock);
3672                 if (block_group->ro)
3673                         space_info->bytes_readonly += reserved_bytes;
3674                 block_group->reserved -= reserved_bytes;
3675                 space_info->bytes_reserved -= reserved_bytes;
3676                 spin_unlock(&block_group->lock);
3677                 spin_unlock(&space_info->lock);
3678         }
3679
3680         return ret;
3681 }
3682
3683 /*
3684  * If @async is set, then we will trim 1 region and return.
3685  */
3686 static int trim_no_bitmap(struct btrfs_block_group *block_group,
3687                           u64 *total_trimmed, u64 start, u64 end, u64 minlen,
3688                           bool async)
3689 {
3690         struct btrfs_discard_ctl *discard_ctl =
3691                                         &block_group->fs_info->discard_ctl;
3692         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3693         struct btrfs_free_space *entry;
3694         struct rb_node *node;
3695         int ret = 0;
3696         u64 extent_start;
3697         u64 extent_bytes;
3698         enum btrfs_trim_state extent_trim_state;
3699         u64 bytes;
3700         const u64 max_discard_size = READ_ONCE(discard_ctl->max_discard_size);
3701
3702         while (start < end) {
3703                 struct btrfs_trim_range trim_entry;
3704
3705                 mutex_lock(&ctl->cache_writeout_mutex);
3706                 spin_lock(&ctl->tree_lock);
3707
3708                 if (ctl->free_space < minlen)
3709                         goto out_unlock;
3710
3711                 entry = tree_search_offset(ctl, start, 0, 1);
3712                 if (!entry)
3713                         goto out_unlock;
3714
3715                 /* Skip bitmaps and if async, already trimmed entries */
3716                 while (entry->bitmap ||
3717                        (async && btrfs_free_space_trimmed(entry))) {
3718                         node = rb_next(&entry->offset_index);
3719                         if (!node)
3720                                 goto out_unlock;
3721                         entry = rb_entry(node, struct btrfs_free_space,
3722                                          offset_index);
3723                 }
3724
3725                 if (entry->offset >= end)
3726                         goto out_unlock;
3727
3728                 extent_start = entry->offset;
3729                 extent_bytes = entry->bytes;
3730                 extent_trim_state = entry->trim_state;
3731                 if (async) {
3732                         start = entry->offset;
3733                         bytes = entry->bytes;
3734                         if (bytes < minlen) {
3735                                 spin_unlock(&ctl->tree_lock);
3736                                 mutex_unlock(&ctl->cache_writeout_mutex);
3737                                 goto next;
3738                         }
3739                         unlink_free_space(ctl, entry, true);
3740                         /*
3741                          * Let bytes = BTRFS_MAX_DISCARD_SIZE + X.
3742                          * If X < BTRFS_ASYNC_DISCARD_MIN_FILTER, we won't trim
3743                          * X when we come back around.  So trim it now.
3744                          */
3745                         if (max_discard_size &&
3746                             bytes >= (max_discard_size +
3747                                       BTRFS_ASYNC_DISCARD_MIN_FILTER)) {
3748                                 bytes = max_discard_size;
3749                                 extent_bytes = max_discard_size;
3750                                 entry->offset += max_discard_size;
3751                                 entry->bytes -= max_discard_size;
3752                                 link_free_space(ctl, entry);
3753                         } else {
3754                                 kmem_cache_free(btrfs_free_space_cachep, entry);
3755                         }
3756                 } else {
3757                         start = max(start, extent_start);
3758                         bytes = min(extent_start + extent_bytes, end) - start;
3759                         if (bytes < minlen) {
3760                                 spin_unlock(&ctl->tree_lock);
3761                                 mutex_unlock(&ctl->cache_writeout_mutex);
3762                                 goto next;
3763                         }
3764
3765                         unlink_free_space(ctl, entry, true);
3766                         kmem_cache_free(btrfs_free_space_cachep, entry);
3767                 }
3768
3769                 spin_unlock(&ctl->tree_lock);
3770                 trim_entry.start = extent_start;
3771                 trim_entry.bytes = extent_bytes;
3772                 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3773                 mutex_unlock(&ctl->cache_writeout_mutex);
3774
3775                 ret = do_trimming(block_group, total_trimmed, start, bytes,
3776                                   extent_start, extent_bytes, extent_trim_state,
3777                                   &trim_entry);
3778                 if (ret) {
3779                         block_group->discard_cursor = start + bytes;
3780                         break;
3781                 }
3782 next:
3783                 start += bytes;
3784                 block_group->discard_cursor = start;
3785                 if (async && *total_trimmed)
3786                         break;
3787
3788                 if (fatal_signal_pending(current)) {
3789                         ret = -ERESTARTSYS;
3790                         break;
3791                 }
3792
3793                 cond_resched();
3794         }
3795
3796         return ret;
3797
3798 out_unlock:
3799         block_group->discard_cursor = btrfs_block_group_end(block_group);
3800         spin_unlock(&ctl->tree_lock);
3801         mutex_unlock(&ctl->cache_writeout_mutex);
3802
3803         return ret;
3804 }
3805
3806 /*
3807  * If we break out of trimming a bitmap prematurely, we should reset the
3808  * trimming bit.  In a rather contrieved case, it's possible to race here so
3809  * reset the state to BTRFS_TRIM_STATE_UNTRIMMED.
3810  *
3811  * start = start of bitmap
3812  * end = near end of bitmap
3813  *
3814  * Thread 1:                    Thread 2:
3815  * trim_bitmaps(start)
3816  *                              trim_bitmaps(end)
3817  *                              end_trimming_bitmap()
3818  * reset_trimming_bitmap()
3819  */
3820 static void reset_trimming_bitmap(struct btrfs_free_space_ctl *ctl, u64 offset)
3821 {
3822         struct btrfs_free_space *entry;
3823
3824         spin_lock(&ctl->tree_lock);
3825         entry = tree_search_offset(ctl, offset, 1, 0);
3826         if (entry) {
3827                 if (btrfs_free_space_trimmed(entry)) {
3828                         ctl->discardable_extents[BTRFS_STAT_CURR] +=
3829                                 entry->bitmap_extents;
3830                         ctl->discardable_bytes[BTRFS_STAT_CURR] += entry->bytes;
3831                 }
3832                 entry->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3833         }
3834
3835         spin_unlock(&ctl->tree_lock);
3836 }
3837
3838 static void end_trimming_bitmap(struct btrfs_free_space_ctl *ctl,
3839                                 struct btrfs_free_space *entry)
3840 {
3841         if (btrfs_free_space_trimming_bitmap(entry)) {
3842                 entry->trim_state = BTRFS_TRIM_STATE_TRIMMED;
3843                 ctl->discardable_extents[BTRFS_STAT_CURR] -=
3844                         entry->bitmap_extents;
3845                 ctl->discardable_bytes[BTRFS_STAT_CURR] -= entry->bytes;
3846         }
3847 }
3848
3849 /*
3850  * If @async is set, then we will trim 1 region and return.
3851  */
3852 static int trim_bitmaps(struct btrfs_block_group *block_group,
3853                         u64 *total_trimmed, u64 start, u64 end, u64 minlen,
3854                         u64 maxlen, bool async)
3855 {
3856         struct btrfs_discard_ctl *discard_ctl =
3857                                         &block_group->fs_info->discard_ctl;
3858         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3859         struct btrfs_free_space *entry;
3860         int ret = 0;
3861         int ret2;
3862         u64 bytes;
3863         u64 offset = offset_to_bitmap(ctl, start);
3864         const u64 max_discard_size = READ_ONCE(discard_ctl->max_discard_size);
3865
3866         while (offset < end) {
3867                 bool next_bitmap = false;
3868                 struct btrfs_trim_range trim_entry;
3869
3870                 mutex_lock(&ctl->cache_writeout_mutex);
3871                 spin_lock(&ctl->tree_lock);
3872
3873                 if (ctl->free_space < minlen) {
3874                         block_group->discard_cursor =
3875                                 btrfs_block_group_end(block_group);
3876                         spin_unlock(&ctl->tree_lock);
3877                         mutex_unlock(&ctl->cache_writeout_mutex);
3878                         break;
3879                 }
3880
3881                 entry = tree_search_offset(ctl, offset, 1, 0);
3882                 /*
3883                  * Bitmaps are marked trimmed lossily now to prevent constant
3884                  * discarding of the same bitmap (the reason why we are bound
3885                  * by the filters).  So, retrim the block group bitmaps when we
3886                  * are preparing to punt to the unused_bgs list.  This uses
3887                  * @minlen to determine if we are in BTRFS_DISCARD_INDEX_UNUSED
3888                  * which is the only discard index which sets minlen to 0.
3889                  */
3890                 if (!entry || (async && minlen && start == offset &&
3891                                btrfs_free_space_trimmed(entry))) {
3892                         spin_unlock(&ctl->tree_lock);
3893                         mutex_unlock(&ctl->cache_writeout_mutex);
3894                         next_bitmap = true;
3895                         goto next;
3896                 }
3897
3898                 /*
3899                  * Async discard bitmap trimming begins at by setting the start
3900                  * to be key.objectid and the offset_to_bitmap() aligns to the
3901                  * start of the bitmap.  This lets us know we are fully
3902                  * scanning the bitmap rather than only some portion of it.
3903                  */
3904                 if (start == offset)
3905                         entry->trim_state = BTRFS_TRIM_STATE_TRIMMING;
3906
3907                 bytes = minlen;
3908                 ret2 = search_bitmap(ctl, entry, &start, &bytes, false);
3909                 if (ret2 || start >= end) {
3910                         /*
3911                          * We lossily consider a bitmap trimmed if we only skip
3912                          * over regions <= BTRFS_ASYNC_DISCARD_MIN_FILTER.
3913                          */
3914                         if (ret2 && minlen <= BTRFS_ASYNC_DISCARD_MIN_FILTER)
3915                                 end_trimming_bitmap(ctl, entry);
3916                         else
3917                                 entry->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3918                         spin_unlock(&ctl->tree_lock);
3919                         mutex_unlock(&ctl->cache_writeout_mutex);
3920                         next_bitmap = true;
3921                         goto next;
3922                 }
3923
3924                 /*
3925                  * We already trimmed a region, but are using the locking above
3926                  * to reset the trim_state.
3927                  */
3928                 if (async && *total_trimmed) {
3929                         spin_unlock(&ctl->tree_lock);
3930                         mutex_unlock(&ctl->cache_writeout_mutex);
3931                         goto out;
3932                 }
3933
3934                 bytes = min(bytes, end - start);
3935                 if (bytes < minlen || (async && maxlen && bytes > maxlen)) {
3936                         spin_unlock(&ctl->tree_lock);
3937                         mutex_unlock(&ctl->cache_writeout_mutex);
3938                         goto next;
3939                 }
3940
3941                 /*
3942                  * Let bytes = BTRFS_MAX_DISCARD_SIZE + X.
3943                  * If X < @minlen, we won't trim X when we come back around.
3944                  * So trim it now.  We differ here from trimming extents as we
3945                  * don't keep individual state per bit.
3946                  */
3947                 if (async &&
3948                     max_discard_size &&
3949                     bytes > (max_discard_size + minlen))
3950                         bytes = max_discard_size;
3951
3952                 bitmap_clear_bits(ctl, entry, start, bytes, true);
3953                 if (entry->bytes == 0)
3954                         free_bitmap(ctl, entry);
3955
3956                 spin_unlock(&ctl->tree_lock);
3957                 trim_entry.start = start;
3958                 trim_entry.bytes = bytes;
3959                 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3960                 mutex_unlock(&ctl->cache_writeout_mutex);
3961
3962                 ret = do_trimming(block_group, total_trimmed, start, bytes,
3963                                   start, bytes, 0, &trim_entry);
3964                 if (ret) {
3965                         reset_trimming_bitmap(ctl, offset);
3966                         block_group->discard_cursor =
3967                                 btrfs_block_group_end(block_group);
3968                         break;
3969                 }
3970 next:
3971                 if (next_bitmap) {
3972                         offset += BITS_PER_BITMAP * ctl->unit;
3973                         start = offset;
3974                 } else {
3975                         start += bytes;
3976                 }
3977                 block_group->discard_cursor = start;
3978
3979                 if (fatal_signal_pending(current)) {
3980                         if (start != offset)
3981                                 reset_trimming_bitmap(ctl, offset);
3982                         ret = -ERESTARTSYS;
3983                         break;
3984                 }
3985
3986                 cond_resched();
3987         }
3988
3989         if (offset >= end)
3990                 block_group->discard_cursor = end;
3991
3992 out:
3993         return ret;
3994 }
3995
3996 int btrfs_trim_block_group(struct btrfs_block_group *block_group,
3997                            u64 *trimmed, u64 start, u64 end, u64 minlen)
3998 {
3999         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
4000         int ret;
4001         u64 rem = 0;
4002
4003         ASSERT(!btrfs_is_zoned(block_group->fs_info));
4004
4005         *trimmed = 0;
4006
4007         spin_lock(&block_group->lock);
4008         if (test_bit(BLOCK_GROUP_FLAG_REMOVED, &block_group->runtime_flags)) {
4009                 spin_unlock(&block_group->lock);
4010                 return 0;
4011         }
4012         btrfs_freeze_block_group(block_group);
4013         spin_unlock(&block_group->lock);
4014
4015         ret = trim_no_bitmap(block_group, trimmed, start, end, minlen, false);
4016         if (ret)
4017                 goto out;
4018
4019         ret = trim_bitmaps(block_group, trimmed, start, end, minlen, 0, false);
4020         div64_u64_rem(end, BITS_PER_BITMAP * ctl->unit, &rem);
4021         /* If we ended in the middle of a bitmap, reset the trimming flag */
4022         if (rem)
4023                 reset_trimming_bitmap(ctl, offset_to_bitmap(ctl, end));
4024 out:
4025         btrfs_unfreeze_block_group(block_group);
4026         return ret;
4027 }
4028
4029 int btrfs_trim_block_group_extents(struct btrfs_block_group *block_group,
4030                                    u64 *trimmed, u64 start, u64 end, u64 minlen,
4031                                    bool async)
4032 {
4033         int ret;
4034
4035         *trimmed = 0;
4036
4037         spin_lock(&block_group->lock);
4038         if (test_bit(BLOCK_GROUP_FLAG_REMOVED, &block_group->runtime_flags)) {
4039                 spin_unlock(&block_group->lock);
4040                 return 0;
4041         }
4042         btrfs_freeze_block_group(block_group);
4043         spin_unlock(&block_group->lock);
4044
4045         ret = trim_no_bitmap(block_group, trimmed, start, end, minlen, async);
4046         btrfs_unfreeze_block_group(block_group);
4047
4048         return ret;
4049 }
4050
4051 int btrfs_trim_block_group_bitmaps(struct btrfs_block_group *block_group,
4052                                    u64 *trimmed, u64 start, u64 end, u64 minlen,
4053                                    u64 maxlen, bool async)
4054 {
4055         int ret;
4056
4057         *trimmed = 0;
4058
4059         spin_lock(&block_group->lock);
4060         if (test_bit(BLOCK_GROUP_FLAG_REMOVED, &block_group->runtime_flags)) {
4061                 spin_unlock(&block_group->lock);
4062                 return 0;
4063         }
4064         btrfs_freeze_block_group(block_group);
4065         spin_unlock(&block_group->lock);
4066
4067         ret = trim_bitmaps(block_group, trimmed, start, end, minlen, maxlen,
4068                            async);
4069
4070         btrfs_unfreeze_block_group(block_group);
4071
4072         return ret;
4073 }
4074
4075 bool btrfs_free_space_cache_v1_active(struct btrfs_fs_info *fs_info)
4076 {
4077         return btrfs_super_cache_generation(fs_info->super_copy);
4078 }
4079
4080 static int cleanup_free_space_cache_v1(struct btrfs_fs_info *fs_info,
4081                                        struct btrfs_trans_handle *trans)
4082 {
4083         struct btrfs_block_group *block_group;
4084         struct rb_node *node;
4085         int ret = 0;
4086
4087         btrfs_info(fs_info, "cleaning free space cache v1");
4088
4089         node = rb_first_cached(&fs_info->block_group_cache_tree);
4090         while (node) {
4091                 block_group = rb_entry(node, struct btrfs_block_group, cache_node);
4092                 ret = btrfs_remove_free_space_inode(trans, NULL, block_group);
4093                 if (ret)
4094                         goto out;
4095                 node = rb_next(node);
4096         }
4097 out:
4098         return ret;
4099 }
4100
4101 int btrfs_set_free_space_cache_v1_active(struct btrfs_fs_info *fs_info, bool active)
4102 {
4103         struct btrfs_trans_handle *trans;
4104         int ret;
4105
4106         /*
4107          * update_super_roots will appropriately set or unset
4108          * super_copy->cache_generation based on SPACE_CACHE and
4109          * BTRFS_FS_CLEANUP_SPACE_CACHE_V1. For this reason, we need a
4110          * transaction commit whether we are enabling space cache v1 and don't
4111          * have any other work to do, or are disabling it and removing free
4112          * space inodes.
4113          */
4114         trans = btrfs_start_transaction(fs_info->tree_root, 0);
4115         if (IS_ERR(trans))
4116                 return PTR_ERR(trans);
4117
4118         if (!active) {
4119                 set_bit(BTRFS_FS_CLEANUP_SPACE_CACHE_V1, &fs_info->flags);
4120                 ret = cleanup_free_space_cache_v1(fs_info, trans);
4121                 if (ret) {
4122                         btrfs_abort_transaction(trans, ret);
4123                         btrfs_end_transaction(trans);
4124                         goto out;
4125                 }
4126         }
4127
4128         ret = btrfs_commit_transaction(trans);
4129 out:
4130         clear_bit(BTRFS_FS_CLEANUP_SPACE_CACHE_V1, &fs_info->flags);
4131
4132         return ret;
4133 }
4134
4135 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4136 /*
4137  * Use this if you need to make a bitmap or extent entry specifically, it
4138  * doesn't do any of the merging that add_free_space does, this acts a lot like
4139  * how the free space cache loading stuff works, so you can get really weird
4140  * configurations.
4141  */
4142 int test_add_free_space_entry(struct btrfs_block_group *cache,
4143                               u64 offset, u64 bytes, bool bitmap)
4144 {
4145         struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
4146         struct btrfs_free_space *info = NULL, *bitmap_info;
4147         void *map = NULL;
4148         enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_TRIMMED;
4149         u64 bytes_added;
4150         int ret;
4151
4152 again:
4153         if (!info) {
4154                 info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
4155                 if (!info)
4156                         return -ENOMEM;
4157         }
4158
4159         if (!bitmap) {
4160                 spin_lock(&ctl->tree_lock);
4161                 info->offset = offset;
4162                 info->bytes = bytes;
4163                 info->max_extent_size = 0;
4164                 ret = link_free_space(ctl, info);
4165                 spin_unlock(&ctl->tree_lock);
4166                 if (ret)
4167                         kmem_cache_free(btrfs_free_space_cachep, info);
4168                 return ret;
4169         }
4170
4171         if (!map) {
4172                 map = kmem_cache_zalloc(btrfs_free_space_bitmap_cachep, GFP_NOFS);
4173                 if (!map) {
4174                         kmem_cache_free(btrfs_free_space_cachep, info);
4175                         return -ENOMEM;
4176                 }
4177         }
4178
4179         spin_lock(&ctl->tree_lock);
4180         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
4181                                          1, 0);
4182         if (!bitmap_info) {
4183                 info->bitmap = map;
4184                 map = NULL;
4185                 add_new_bitmap(ctl, info, offset);
4186                 bitmap_info = info;
4187                 info = NULL;
4188         }
4189
4190         bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes,
4191                                           trim_state);
4192
4193         bytes -= bytes_added;
4194         offset += bytes_added;
4195         spin_unlock(&ctl->tree_lock);
4196
4197         if (bytes)
4198                 goto again;
4199
4200         if (info)
4201                 kmem_cache_free(btrfs_free_space_cachep, info);
4202         if (map)
4203                 kmem_cache_free(btrfs_free_space_bitmap_cachep, map);
4204         return 0;
4205 }
4206
4207 /*
4208  * Checks to see if the given range is in the free space cache.  This is really
4209  * just used to check the absence of space, so if there is free space in the
4210  * range at all we will return 1.
4211  */
4212 int test_check_exists(struct btrfs_block_group *cache,
4213                       u64 offset, u64 bytes)
4214 {
4215         struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
4216         struct btrfs_free_space *info;
4217         int ret = 0;
4218
4219         spin_lock(&ctl->tree_lock);
4220         info = tree_search_offset(ctl, offset, 0, 0);
4221         if (!info) {
4222                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
4223                                           1, 0);
4224                 if (!info)
4225                         goto out;
4226         }
4227
4228 have_info:
4229         if (info->bitmap) {
4230                 u64 bit_off, bit_bytes;
4231                 struct rb_node *n;
4232                 struct btrfs_free_space *tmp;
4233
4234                 bit_off = offset;
4235                 bit_bytes = ctl->unit;
4236                 ret = search_bitmap(ctl, info, &bit_off, &bit_bytes, false);
4237                 if (!ret) {
4238                         if (bit_off == offset) {
4239                                 ret = 1;
4240                                 goto out;
4241                         } else if (bit_off > offset &&
4242                                    offset + bytes > bit_off) {
4243                                 ret = 1;
4244                                 goto out;
4245                         }
4246                 }
4247
4248                 n = rb_prev(&info->offset_index);
4249                 while (n) {
4250                         tmp = rb_entry(n, struct btrfs_free_space,
4251                                        offset_index);
4252                         if (tmp->offset + tmp->bytes < offset)
4253                                 break;
4254                         if (offset + bytes < tmp->offset) {
4255                                 n = rb_prev(&tmp->offset_index);
4256                                 continue;
4257                         }
4258                         info = tmp;
4259                         goto have_info;
4260                 }
4261
4262                 n = rb_next(&info->offset_index);
4263                 while (n) {
4264                         tmp = rb_entry(n, struct btrfs_free_space,
4265                                        offset_index);
4266                         if (offset + bytes < tmp->offset)
4267                                 break;
4268                         if (tmp->offset + tmp->bytes < offset) {
4269                                 n = rb_next(&tmp->offset_index);
4270                                 continue;
4271                         }
4272                         info = tmp;
4273                         goto have_info;
4274                 }
4275
4276                 ret = 0;
4277                 goto out;
4278         }
4279
4280         if (info->offset == offset) {
4281                 ret = 1;
4282                 goto out;
4283         }
4284
4285         if (offset > info->offset && offset < info->offset + info->bytes)
4286                 ret = 1;
4287 out:
4288         spin_unlock(&ctl->tree_lock);
4289         return ret;
4290 }
4291 #endif /* CONFIG_BTRFS_FS_RUN_SANITY_TESTS */