1 // SPDX-License-Identifier: GPL-2.0 2 3 #include "misc.h" 4 #include "ctree.h" 5 #include "space-info.h" 6 #include "sysfs.h" 7 #include "volumes.h" 8 #include "free-space-cache.h" 9 #include "ordered-data.h" 10 #include "transaction.h" 11 #include "block-group.h" 12 #include "zoned.h" 13 #include "fs.h" 14 15 /* 16 * HOW DOES SPACE RESERVATION WORK 17 * 18 * If you want to know about delalloc specifically, there is a separate comment 19 * for that with the delalloc code. This comment is about how the whole system 20 * works generally. 21 * 22 * BASIC CONCEPTS 23 * 24 * 1) space_info. This is the ultimate arbiter of how much space we can use. 25 * There's a description of the bytes_ fields with the struct declaration, 26 * refer to that for specifics on each field. Suffice it to say that for 27 * reservations we care about total_bytes - SUM(space_info->bytes_) when 28 * determining if there is space to make an allocation. There is a space_info 29 * for METADATA, SYSTEM, and DATA areas. 30 * 31 * 2) block_rsv's. These are basically buckets for every different type of 32 * metadata reservation we have. You can see the comment in the block_rsv 33 * code on the rules for each type, but generally block_rsv->reserved is how 34 * much space is accounted for in space_info->bytes_may_use. 35 * 36 * 3) btrfs_calc*_size. These are the worst case calculations we used based 37 * on the number of items we will want to modify. We have one for changing 38 * items, and one for inserting new items. Generally we use these helpers to 39 * determine the size of the block reserves, and then use the actual bytes 40 * values to adjust the space_info counters. 41 * 42 * MAKING RESERVATIONS, THE NORMAL CASE 43 * 44 * We call into either btrfs_reserve_data_bytes() or 45 * btrfs_reserve_metadata_bytes(), depending on which we're looking for, with 46 * num_bytes we want to reserve. 47 * 48 * ->reserve 49 * space_info->bytes_may_reserve += num_bytes 50 * 51 * ->extent allocation 52 * Call btrfs_add_reserved_bytes() which does 53 * space_info->bytes_may_reserve -= num_bytes 54 * space_info->bytes_reserved += extent_bytes 55 * 56 * ->insert reference 57 * Call btrfs_update_block_group() which does 58 * space_info->bytes_reserved -= extent_bytes 59 * space_info->bytes_used += extent_bytes 60 * 61 * MAKING RESERVATIONS, FLUSHING NORMALLY (non-priority) 62 * 63 * Assume we are unable to simply make the reservation because we do not have 64 * enough space 65 * 66 * -> __reserve_bytes 67 * create a reserve_ticket with ->bytes set to our reservation, add it to 68 * the tail of space_info->tickets, kick async flush thread 69 * 70 * ->handle_reserve_ticket 71 * wait on ticket->wait for ->bytes to be reduced to 0, or ->error to be set 72 * on the ticket. 73 * 74 * -> btrfs_async_reclaim_metadata_space/btrfs_async_reclaim_data_space 75 * Flushes various things attempting to free up space. 76 * 77 * -> btrfs_try_granting_tickets() 78 * This is called by anything that either subtracts space from 79 * space_info->bytes_may_use, ->bytes_pinned, etc, or adds to the 80 * space_info->total_bytes. This loops through the ->priority_tickets and 81 * then the ->tickets list checking to see if the reservation can be 82 * completed. If it can the space is added to space_info->bytes_may_use and 83 * the ticket is woken up. 84 * 85 * -> ticket wakeup 86 * Check if ->bytes == 0, if it does we got our reservation and we can carry 87 * on, if not return the appropriate error (ENOSPC, but can be EINTR if we 88 * were interrupted.) 89 * 90 * MAKING RESERVATIONS, FLUSHING HIGH PRIORITY 91 * 92 * Same as the above, except we add ourselves to the 93 * space_info->priority_tickets, and we do not use ticket->wait, we simply 94 * call flush_space() ourselves for the states that are safe for us to call 95 * without deadlocking and hope for the best. 96 * 97 * THE FLUSHING STATES 98 * 99 * Generally speaking we will have two cases for each state, a "nice" state 100 * and a "ALL THE THINGS" state. In btrfs we delay a lot of work in order to 101 * reduce the locking over head on the various trees, and even to keep from 102 * doing any work at all in the case of delayed refs. Each of these delayed 103 * things however hold reservations, and so letting them run allows us to 104 * reclaim space so we can make new reservations. 105 * 106 * FLUSH_DELAYED_ITEMS 107 * Every inode has a delayed item to update the inode. Take a simple write 108 * for example, we would update the inode item at write time to update the 109 * mtime, and then again at finish_ordered_io() time in order to update the 110 * isize or bytes. We keep these delayed items to coalesce these operations 111 * into a single operation done on demand. These are an easy way to reclaim 112 * metadata space. 113 * 114 * FLUSH_DELALLOC 115 * Look at the delalloc comment to get an idea of how much space is reserved 116 * for delayed allocation. We can reclaim some of this space simply by 117 * running delalloc, but usually we need to wait for ordered extents to 118 * reclaim the bulk of this space. 119 * 120 * FLUSH_DELAYED_REFS 121 * We have a block reserve for the outstanding delayed refs space, and every 122 * delayed ref operation holds a reservation. Running these is a quick way 123 * to reclaim space, but we want to hold this until the end because COW can 124 * churn a lot and we can avoid making some extent tree modifications if we 125 * are able to delay for as long as possible. 126 * 127 * ALLOC_CHUNK 128 * We will skip this the first time through space reservation, because of 129 * overcommit and we don't want to have a lot of useless metadata space when 130 * our worst case reservations will likely never come true. 131 * 132 * RUN_DELAYED_IPUTS 133 * If we're freeing inodes we're likely freeing checksums, file extent 134 * items, and extent tree items. Loads of space could be freed up by these 135 * operations, however they won't be usable until the transaction commits. 136 * 137 * COMMIT_TRANS 138 * This will commit the transaction. Historically we had a lot of logic 139 * surrounding whether or not we'd commit the transaction, but this waits born 140 * out of a pre-tickets era where we could end up committing the transaction 141 * thousands of times in a row without making progress. Now thanks to our 142 * ticketing system we know if we're not making progress and can error 143 * everybody out after a few commits rather than burning the disk hoping for 144 * a different answer. 145 * 146 * OVERCOMMIT 147 * 148 * Because we hold so many reservations for metadata we will allow you to 149 * reserve more space than is currently free in the currently allocate 150 * metadata space. This only happens with metadata, data does not allow 151 * overcommitting. 152 * 153 * You can see the current logic for when we allow overcommit in 154 * btrfs_can_overcommit(), but it only applies to unallocated space. If there 155 * is no unallocated space to be had, all reservations are kept within the 156 * free space in the allocated metadata chunks. 157 * 158 * Because of overcommitting, you generally want to use the 159 * btrfs_can_overcommit() logic for metadata allocations, as it does the right 160 * thing with or without extra unallocated space. 161 */ 162 163 u64 __pure btrfs_space_info_used(struct btrfs_space_info *s_info, 164 bool may_use_included) 165 { 166 ASSERT(s_info); 167 return s_info->bytes_used + s_info->bytes_reserved + 168 s_info->bytes_pinned + s_info->bytes_readonly + 169 s_info->bytes_zone_unusable + 170 (may_use_included ? s_info->bytes_may_use : 0); 171 } 172 173 /* 174 * after adding space to the filesystem, we need to clear the full flags 175 * on all the space infos. 176 */ 177 void btrfs_clear_space_info_full(struct btrfs_fs_info *info) 178 { 179 struct list_head *head = &info->space_info; 180 struct btrfs_space_info *found; 181 182 list_for_each_entry(found, head, list) 183 found->full = 0; 184 } 185 186 /* 187 * Block groups with more than this value (percents) of unusable space will be 188 * scheduled for background reclaim. 189 */ 190 #define BTRFS_DEFAULT_ZONED_RECLAIM_THRESH (75) 191 192 /* 193 * Calculate chunk size depending on volume type (regular or zoned). 194 */ 195 static u64 calc_chunk_size(const struct btrfs_fs_info *fs_info, u64 flags) 196 { 197 if (btrfs_is_zoned(fs_info)) 198 return fs_info->zone_size; 199 200 ASSERT(flags & BTRFS_BLOCK_GROUP_TYPE_MASK); 201 202 if (flags & BTRFS_BLOCK_GROUP_DATA) 203 return BTRFS_MAX_DATA_CHUNK_SIZE; 204 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM) 205 return SZ_32M; 206 207 /* Handle BTRFS_BLOCK_GROUP_METADATA */ 208 if (fs_info->fs_devices->total_rw_bytes > 50ULL * SZ_1G) 209 return SZ_1G; 210 211 return SZ_256M; 212 } 213 214 /* 215 * Update default chunk size. 216 */ 217 void btrfs_update_space_info_chunk_size(struct btrfs_space_info *space_info, 218 u64 chunk_size) 219 { 220 WRITE_ONCE(space_info->chunk_size, chunk_size); 221 } 222 223 static int create_space_info(struct btrfs_fs_info *info, u64 flags) 224 { 225 226 struct btrfs_space_info *space_info; 227 int i; 228 int ret; 229 230 space_info = kzalloc(sizeof(*space_info), GFP_NOFS); 231 if (!space_info) 232 return -ENOMEM; 233 234 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) 235 INIT_LIST_HEAD(&space_info->block_groups[i]); 236 init_rwsem(&space_info->groups_sem); 237 spin_lock_init(&space_info->lock); 238 space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK; 239 space_info->force_alloc = CHUNK_ALLOC_NO_FORCE; 240 INIT_LIST_HEAD(&space_info->ro_bgs); 241 INIT_LIST_HEAD(&space_info->tickets); 242 INIT_LIST_HEAD(&space_info->priority_tickets); 243 space_info->clamp = 1; 244 btrfs_update_space_info_chunk_size(space_info, calc_chunk_size(info, flags)); 245 246 if (btrfs_is_zoned(info)) 247 space_info->bg_reclaim_threshold = BTRFS_DEFAULT_ZONED_RECLAIM_THRESH; 248 249 ret = btrfs_sysfs_add_space_info_type(info, space_info); 250 if (ret) 251 return ret; 252 253 list_add(&space_info->list, &info->space_info); 254 if (flags & BTRFS_BLOCK_GROUP_DATA) 255 info->data_sinfo = space_info; 256 257 return ret; 258 } 259 260 int btrfs_init_space_info(struct btrfs_fs_info *fs_info) 261 { 262 struct btrfs_super_block *disk_super; 263 u64 features; 264 u64 flags; 265 int mixed = 0; 266 int ret; 267 268 disk_super = fs_info->super_copy; 269 if (!btrfs_super_root(disk_super)) 270 return -EINVAL; 271 272 features = btrfs_super_incompat_flags(disk_super); 273 if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS) 274 mixed = 1; 275 276 flags = BTRFS_BLOCK_GROUP_SYSTEM; 277 ret = create_space_info(fs_info, flags); 278 if (ret) 279 goto out; 280 281 if (mixed) { 282 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA; 283 ret = create_space_info(fs_info, flags); 284 } else { 285 flags = BTRFS_BLOCK_GROUP_METADATA; 286 ret = create_space_info(fs_info, flags); 287 if (ret) 288 goto out; 289 290 flags = BTRFS_BLOCK_GROUP_DATA; 291 ret = create_space_info(fs_info, flags); 292 } 293 out: 294 return ret; 295 } 296 297 void btrfs_add_bg_to_space_info(struct btrfs_fs_info *info, 298 struct btrfs_block_group *block_group) 299 { 300 struct btrfs_space_info *found; 301 int factor, index; 302 303 factor = btrfs_bg_type_to_factor(block_group->flags); 304 305 found = btrfs_find_space_info(info, block_group->flags); 306 ASSERT(found); 307 spin_lock(&found->lock); 308 found->total_bytes += block_group->length; 309 if (test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &block_group->runtime_flags)) 310 found->active_total_bytes += block_group->length; 311 found->disk_total += block_group->length * factor; 312 found->bytes_used += block_group->used; 313 found->disk_used += block_group->used * factor; 314 found->bytes_readonly += block_group->bytes_super; 315 found->bytes_zone_unusable += block_group->zone_unusable; 316 if (block_group->length > 0) 317 found->full = 0; 318 btrfs_try_granting_tickets(info, found); 319 spin_unlock(&found->lock); 320 321 block_group->space_info = found; 322 323 index = btrfs_bg_flags_to_raid_index(block_group->flags); 324 down_write(&found->groups_sem); 325 list_add_tail(&block_group->list, &found->block_groups[index]); 326 up_write(&found->groups_sem); 327 } 328 329 struct btrfs_space_info *btrfs_find_space_info(struct btrfs_fs_info *info, 330 u64 flags) 331 { 332 struct list_head *head = &info->space_info; 333 struct btrfs_space_info *found; 334 335 flags &= BTRFS_BLOCK_GROUP_TYPE_MASK; 336 337 list_for_each_entry(found, head, list) { 338 if (found->flags & flags) 339 return found; 340 } 341 return NULL; 342 } 343 344 static u64 calc_available_free_space(struct btrfs_fs_info *fs_info, 345 struct btrfs_space_info *space_info, 346 enum btrfs_reserve_flush_enum flush) 347 { 348 u64 profile; 349 u64 avail; 350 int factor; 351 352 if (space_info->flags & BTRFS_BLOCK_GROUP_SYSTEM) 353 profile = btrfs_system_alloc_profile(fs_info); 354 else 355 profile = btrfs_metadata_alloc_profile(fs_info); 356 357 avail = atomic64_read(&fs_info->free_chunk_space); 358 359 /* 360 * If we have dup, raid1 or raid10 then only half of the free 361 * space is actually usable. For raid56, the space info used 362 * doesn't include the parity drive, so we don't have to 363 * change the math 364 */ 365 factor = btrfs_bg_type_to_factor(profile); 366 avail = div_u64(avail, factor); 367 368 /* 369 * If we aren't flushing all things, let us overcommit up to 370 * 1/2th of the space. If we can flush, don't let us overcommit 371 * too much, let it overcommit up to 1/8 of the space. 372 */ 373 if (flush == BTRFS_RESERVE_FLUSH_ALL) 374 avail >>= 3; 375 else 376 avail >>= 1; 377 return avail; 378 } 379 380 static inline u64 writable_total_bytes(struct btrfs_fs_info *fs_info, 381 struct btrfs_space_info *space_info) 382 { 383 /* 384 * On regular filesystem, all total_bytes are always writable. On zoned 385 * filesystem, there may be a limitation imposed by max_active_zones. 386 * For metadata allocation, we cannot finish an existing active block 387 * group to avoid a deadlock. Thus, we need to consider only the active 388 * groups to be writable for metadata space. 389 */ 390 if (!btrfs_is_zoned(fs_info) || (space_info->flags & BTRFS_BLOCK_GROUP_DATA)) 391 return space_info->total_bytes; 392 393 return space_info->active_total_bytes; 394 } 395 396 int btrfs_can_overcommit(struct btrfs_fs_info *fs_info, 397 struct btrfs_space_info *space_info, u64 bytes, 398 enum btrfs_reserve_flush_enum flush) 399 { 400 u64 avail; 401 u64 used; 402 403 /* Don't overcommit when in mixed mode */ 404 if (space_info->flags & BTRFS_BLOCK_GROUP_DATA) 405 return 0; 406 407 used = btrfs_space_info_used(space_info, true); 408 if (btrfs_is_zoned(fs_info) && (space_info->flags & BTRFS_BLOCK_GROUP_METADATA)) 409 avail = 0; 410 else 411 avail = calc_available_free_space(fs_info, space_info, flush); 412 413 if (used + bytes < writable_total_bytes(fs_info, space_info) + avail) 414 return 1; 415 return 0; 416 } 417 418 static void remove_ticket(struct btrfs_space_info *space_info, 419 struct reserve_ticket *ticket) 420 { 421 if (!list_empty(&ticket->list)) { 422 list_del_init(&ticket->list); 423 ASSERT(space_info->reclaim_size >= ticket->bytes); 424 space_info->reclaim_size -= ticket->bytes; 425 } 426 } 427 428 /* 429 * This is for space we already have accounted in space_info->bytes_may_use, so 430 * basically when we're returning space from block_rsv's. 431 */ 432 void btrfs_try_granting_tickets(struct btrfs_fs_info *fs_info, 433 struct btrfs_space_info *space_info) 434 { 435 struct list_head *head; 436 enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH; 437 438 lockdep_assert_held(&space_info->lock); 439 440 head = &space_info->priority_tickets; 441 again: 442 while (!list_empty(head)) { 443 struct reserve_ticket *ticket; 444 u64 used = btrfs_space_info_used(space_info, true); 445 446 ticket = list_first_entry(head, struct reserve_ticket, list); 447 448 /* Check and see if our ticket can be satisfied now. */ 449 if ((used + ticket->bytes <= writable_total_bytes(fs_info, space_info)) || 450 btrfs_can_overcommit(fs_info, space_info, ticket->bytes, 451 flush)) { 452 btrfs_space_info_update_bytes_may_use(fs_info, 453 space_info, 454 ticket->bytes); 455 remove_ticket(space_info, ticket); 456 ticket->bytes = 0; 457 space_info->tickets_id++; 458 wake_up(&ticket->wait); 459 } else { 460 break; 461 } 462 } 463 464 if (head == &space_info->priority_tickets) { 465 head = &space_info->tickets; 466 flush = BTRFS_RESERVE_FLUSH_ALL; 467 goto again; 468 } 469 } 470 471 #define DUMP_BLOCK_RSV(fs_info, rsv_name) \ 472 do { \ 473 struct btrfs_block_rsv *__rsv = &(fs_info)->rsv_name; \ 474 spin_lock(&__rsv->lock); \ 475 btrfs_info(fs_info, #rsv_name ": size %llu reserved %llu", \ 476 __rsv->size, __rsv->reserved); \ 477 spin_unlock(&__rsv->lock); \ 478 } while (0) 479 480 static const char *space_info_flag_to_str(const struct btrfs_space_info *space_info) 481 { 482 switch (space_info->flags) { 483 case BTRFS_BLOCK_GROUP_SYSTEM: 484 return "SYSTEM"; 485 case BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA: 486 return "DATA+METADATA"; 487 case BTRFS_BLOCK_GROUP_DATA: 488 return "DATA"; 489 case BTRFS_BLOCK_GROUP_METADATA: 490 return "METADATA"; 491 default: 492 return "UNKNOWN"; 493 } 494 } 495 496 static void dump_global_block_rsv(struct btrfs_fs_info *fs_info) 497 { 498 DUMP_BLOCK_RSV(fs_info, global_block_rsv); 499 DUMP_BLOCK_RSV(fs_info, trans_block_rsv); 500 DUMP_BLOCK_RSV(fs_info, chunk_block_rsv); 501 DUMP_BLOCK_RSV(fs_info, delayed_block_rsv); 502 DUMP_BLOCK_RSV(fs_info, delayed_refs_rsv); 503 } 504 505 static void __btrfs_dump_space_info(struct btrfs_fs_info *fs_info, 506 struct btrfs_space_info *info) 507 { 508 const char *flag_str = space_info_flag_to_str(info); 509 lockdep_assert_held(&info->lock); 510 511 /* The free space could be negative in case of overcommit */ 512 btrfs_info(fs_info, "space_info %s has %lld free, is %sfull", 513 flag_str, 514 (s64)(info->total_bytes - btrfs_space_info_used(info, true)), 515 info->full ? "" : "not "); 516 btrfs_info(fs_info, 517 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu zone_unusable=%llu", 518 info->total_bytes, info->bytes_used, info->bytes_pinned, 519 info->bytes_reserved, info->bytes_may_use, 520 info->bytes_readonly, info->bytes_zone_unusable); 521 } 522 523 void btrfs_dump_space_info(struct btrfs_fs_info *fs_info, 524 struct btrfs_space_info *info, u64 bytes, 525 int dump_block_groups) 526 { 527 struct btrfs_block_group *cache; 528 int index = 0; 529 530 spin_lock(&info->lock); 531 __btrfs_dump_space_info(fs_info, info); 532 dump_global_block_rsv(fs_info); 533 spin_unlock(&info->lock); 534 535 if (!dump_block_groups) 536 return; 537 538 down_read(&info->groups_sem); 539 again: 540 list_for_each_entry(cache, &info->block_groups[index], list) { 541 spin_lock(&cache->lock); 542 btrfs_info(fs_info, 543 "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %llu zone_unusable %s", 544 cache->start, cache->length, cache->used, cache->pinned, 545 cache->reserved, cache->zone_unusable, 546 cache->ro ? "[readonly]" : ""); 547 spin_unlock(&cache->lock); 548 btrfs_dump_free_space(cache, bytes); 549 } 550 if (++index < BTRFS_NR_RAID_TYPES) 551 goto again; 552 up_read(&info->groups_sem); 553 } 554 555 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info, 556 u64 to_reclaim) 557 { 558 u64 bytes; 559 u64 nr; 560 561 bytes = btrfs_calc_insert_metadata_size(fs_info, 1); 562 nr = div64_u64(to_reclaim, bytes); 563 if (!nr) 564 nr = 1; 565 return nr; 566 } 567 568 #define EXTENT_SIZE_PER_ITEM SZ_256K 569 570 /* 571 * shrink metadata reservation for delalloc 572 */ 573 static void shrink_delalloc(struct btrfs_fs_info *fs_info, 574 struct btrfs_space_info *space_info, 575 u64 to_reclaim, bool wait_ordered, 576 bool for_preempt) 577 { 578 struct btrfs_trans_handle *trans; 579 u64 delalloc_bytes; 580 u64 ordered_bytes; 581 u64 items; 582 long time_left; 583 int loops; 584 585 delalloc_bytes = percpu_counter_sum_positive(&fs_info->delalloc_bytes); 586 ordered_bytes = percpu_counter_sum_positive(&fs_info->ordered_bytes); 587 if (delalloc_bytes == 0 && ordered_bytes == 0) 588 return; 589 590 /* Calc the number of the pages we need flush for space reservation */ 591 if (to_reclaim == U64_MAX) { 592 items = U64_MAX; 593 } else { 594 /* 595 * to_reclaim is set to however much metadata we need to 596 * reclaim, but reclaiming that much data doesn't really track 597 * exactly. What we really want to do is reclaim full inode's 598 * worth of reservations, however that's not available to us 599 * here. We will take a fraction of the delalloc bytes for our 600 * flushing loops and hope for the best. Delalloc will expand 601 * the amount we write to cover an entire dirty extent, which 602 * will reclaim the metadata reservation for that range. If 603 * it's not enough subsequent flush stages will be more 604 * aggressive. 605 */ 606 to_reclaim = max(to_reclaim, delalloc_bytes >> 3); 607 items = calc_reclaim_items_nr(fs_info, to_reclaim) * 2; 608 } 609 610 trans = current->journal_info; 611 612 /* 613 * If we are doing more ordered than delalloc we need to just wait on 614 * ordered extents, otherwise we'll waste time trying to flush delalloc 615 * that likely won't give us the space back we need. 616 */ 617 if (ordered_bytes > delalloc_bytes && !for_preempt) 618 wait_ordered = true; 619 620 loops = 0; 621 while ((delalloc_bytes || ordered_bytes) && loops < 3) { 622 u64 temp = min(delalloc_bytes, to_reclaim) >> PAGE_SHIFT; 623 long nr_pages = min_t(u64, temp, LONG_MAX); 624 int async_pages; 625 626 btrfs_start_delalloc_roots(fs_info, nr_pages, true); 627 628 /* 629 * We need to make sure any outstanding async pages are now 630 * processed before we continue. This is because things like 631 * sync_inode() try to be smart and skip writing if the inode is 632 * marked clean. We don't use filemap_fwrite for flushing 633 * because we want to control how many pages we write out at a 634 * time, thus this is the only safe way to make sure we've 635 * waited for outstanding compressed workers to have started 636 * their jobs and thus have ordered extents set up properly. 637 * 638 * This exists because we do not want to wait for each 639 * individual inode to finish its async work, we simply want to 640 * start the IO on everybody, and then come back here and wait 641 * for all of the async work to catch up. Once we're done with 642 * that we know we'll have ordered extents for everything and we 643 * can decide if we wait for that or not. 644 * 645 * If we choose to replace this in the future, make absolutely 646 * sure that the proper waiting is being done in the async case, 647 * as there have been bugs in that area before. 648 */ 649 async_pages = atomic_read(&fs_info->async_delalloc_pages); 650 if (!async_pages) 651 goto skip_async; 652 653 /* 654 * We don't want to wait forever, if we wrote less pages in this 655 * loop than we have outstanding, only wait for that number of 656 * pages, otherwise we can wait for all async pages to finish 657 * before continuing. 658 */ 659 if (async_pages > nr_pages) 660 async_pages -= nr_pages; 661 else 662 async_pages = 0; 663 wait_event(fs_info->async_submit_wait, 664 atomic_read(&fs_info->async_delalloc_pages) <= 665 async_pages); 666 skip_async: 667 loops++; 668 if (wait_ordered && !trans) { 669 btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1); 670 } else { 671 time_left = schedule_timeout_killable(1); 672 if (time_left) 673 break; 674 } 675 676 /* 677 * If we are for preemption we just want a one-shot of delalloc 678 * flushing so we can stop flushing if we decide we don't need 679 * to anymore. 680 */ 681 if (for_preempt) 682 break; 683 684 spin_lock(&space_info->lock); 685 if (list_empty(&space_info->tickets) && 686 list_empty(&space_info->priority_tickets)) { 687 spin_unlock(&space_info->lock); 688 break; 689 } 690 spin_unlock(&space_info->lock); 691 692 delalloc_bytes = percpu_counter_sum_positive( 693 &fs_info->delalloc_bytes); 694 ordered_bytes = percpu_counter_sum_positive( 695 &fs_info->ordered_bytes); 696 } 697 } 698 699 /* 700 * Try to flush some data based on policy set by @state. This is only advisory 701 * and may fail for various reasons. The caller is supposed to examine the 702 * state of @space_info to detect the outcome. 703 */ 704 static void flush_space(struct btrfs_fs_info *fs_info, 705 struct btrfs_space_info *space_info, u64 num_bytes, 706 enum btrfs_flush_state state, bool for_preempt) 707 { 708 struct btrfs_root *root = fs_info->tree_root; 709 struct btrfs_trans_handle *trans; 710 int nr; 711 int ret = 0; 712 713 switch (state) { 714 case FLUSH_DELAYED_ITEMS_NR: 715 case FLUSH_DELAYED_ITEMS: 716 if (state == FLUSH_DELAYED_ITEMS_NR) 717 nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2; 718 else 719 nr = -1; 720 721 trans = btrfs_join_transaction(root); 722 if (IS_ERR(trans)) { 723 ret = PTR_ERR(trans); 724 break; 725 } 726 ret = btrfs_run_delayed_items_nr(trans, nr); 727 btrfs_end_transaction(trans); 728 break; 729 case FLUSH_DELALLOC: 730 case FLUSH_DELALLOC_WAIT: 731 case FLUSH_DELALLOC_FULL: 732 if (state == FLUSH_DELALLOC_FULL) 733 num_bytes = U64_MAX; 734 shrink_delalloc(fs_info, space_info, num_bytes, 735 state != FLUSH_DELALLOC, for_preempt); 736 break; 737 case FLUSH_DELAYED_REFS_NR: 738 case FLUSH_DELAYED_REFS: 739 trans = btrfs_join_transaction(root); 740 if (IS_ERR(trans)) { 741 ret = PTR_ERR(trans); 742 break; 743 } 744 if (state == FLUSH_DELAYED_REFS_NR) 745 nr = calc_reclaim_items_nr(fs_info, num_bytes); 746 else 747 nr = 0; 748 btrfs_run_delayed_refs(trans, nr); 749 btrfs_end_transaction(trans); 750 break; 751 case ALLOC_CHUNK: 752 case ALLOC_CHUNK_FORCE: 753 /* 754 * For metadata space on zoned filesystem, reaching here means we 755 * don't have enough space left in active_total_bytes. Try to 756 * activate a block group first, because we may have inactive 757 * block group already allocated. 758 */ 759 ret = btrfs_zoned_activate_one_bg(fs_info, space_info, false); 760 if (ret < 0) 761 break; 762 else if (ret == 1) 763 break; 764 765 trans = btrfs_join_transaction(root); 766 if (IS_ERR(trans)) { 767 ret = PTR_ERR(trans); 768 break; 769 } 770 ret = btrfs_chunk_alloc(trans, 771 btrfs_get_alloc_profile(fs_info, space_info->flags), 772 (state == ALLOC_CHUNK) ? CHUNK_ALLOC_NO_FORCE : 773 CHUNK_ALLOC_FORCE); 774 btrfs_end_transaction(trans); 775 776 /* 777 * For metadata space on zoned filesystem, allocating a new chunk 778 * is not enough. We still need to activate the block * group. 779 * Active the newly allocated block group by (maybe) finishing 780 * a block group. 781 */ 782 if (ret == 1) { 783 ret = btrfs_zoned_activate_one_bg(fs_info, space_info, true); 784 /* 785 * Revert to the original ret regardless we could finish 786 * one block group or not. 787 */ 788 if (ret >= 0) 789 ret = 1; 790 } 791 792 if (ret > 0 || ret == -ENOSPC) 793 ret = 0; 794 break; 795 case RUN_DELAYED_IPUTS: 796 /* 797 * If we have pending delayed iputs then we could free up a 798 * bunch of pinned space, so make sure we run the iputs before 799 * we do our pinned bytes check below. 800 */ 801 btrfs_run_delayed_iputs(fs_info); 802 btrfs_wait_on_delayed_iputs(fs_info); 803 break; 804 case COMMIT_TRANS: 805 ASSERT(current->journal_info == NULL); 806 trans = btrfs_join_transaction(root); 807 if (IS_ERR(trans)) { 808 ret = PTR_ERR(trans); 809 break; 810 } 811 ret = btrfs_commit_transaction(trans); 812 break; 813 default: 814 ret = -ENOSPC; 815 break; 816 } 817 818 trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state, 819 ret, for_preempt); 820 return; 821 } 822 823 static inline u64 824 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info, 825 struct btrfs_space_info *space_info) 826 { 827 u64 used; 828 u64 avail; 829 u64 total; 830 u64 to_reclaim = space_info->reclaim_size; 831 832 lockdep_assert_held(&space_info->lock); 833 834 avail = calc_available_free_space(fs_info, space_info, 835 BTRFS_RESERVE_FLUSH_ALL); 836 used = btrfs_space_info_used(space_info, true); 837 838 /* 839 * We may be flushing because suddenly we have less space than we had 840 * before, and now we're well over-committed based on our current free 841 * space. If that's the case add in our overage so we make sure to put 842 * appropriate pressure on the flushing state machine. 843 */ 844 total = writable_total_bytes(fs_info, space_info); 845 if (total + avail < used) 846 to_reclaim += used - (total + avail); 847 848 return to_reclaim; 849 } 850 851 static bool need_preemptive_reclaim(struct btrfs_fs_info *fs_info, 852 struct btrfs_space_info *space_info) 853 { 854 u64 global_rsv_size = fs_info->global_block_rsv.reserved; 855 u64 ordered, delalloc; 856 u64 total = writable_total_bytes(fs_info, space_info); 857 u64 thresh; 858 u64 used; 859 860 thresh = div_factor_fine(total, 90); 861 862 lockdep_assert_held(&space_info->lock); 863 864 /* If we're just plain full then async reclaim just slows us down. */ 865 if ((space_info->bytes_used + space_info->bytes_reserved + 866 global_rsv_size) >= thresh) 867 return false; 868 869 used = space_info->bytes_may_use + space_info->bytes_pinned; 870 871 /* The total flushable belongs to the global rsv, don't flush. */ 872 if (global_rsv_size >= used) 873 return false; 874 875 /* 876 * 128MiB is 1/4 of the maximum global rsv size. If we have less than 877 * that devoted to other reservations then there's no sense in flushing, 878 * we don't have a lot of things that need flushing. 879 */ 880 if (used - global_rsv_size <= SZ_128M) 881 return false; 882 883 /* 884 * We have tickets queued, bail so we don't compete with the async 885 * flushers. 886 */ 887 if (space_info->reclaim_size) 888 return false; 889 890 /* 891 * If we have over half of the free space occupied by reservations or 892 * pinned then we want to start flushing. 893 * 894 * We do not do the traditional thing here, which is to say 895 * 896 * if (used >= ((total_bytes + avail) / 2)) 897 * return 1; 898 * 899 * because this doesn't quite work how we want. If we had more than 50% 900 * of the space_info used by bytes_used and we had 0 available we'd just 901 * constantly run the background flusher. Instead we want it to kick in 902 * if our reclaimable space exceeds our clamped free space. 903 * 904 * Our clamping range is 2^1 -> 2^8. Practically speaking that means 905 * the following: 906 * 907 * Amount of RAM Minimum threshold Maximum threshold 908 * 909 * 256GiB 1GiB 128GiB 910 * 128GiB 512MiB 64GiB 911 * 64GiB 256MiB 32GiB 912 * 32GiB 128MiB 16GiB 913 * 16GiB 64MiB 8GiB 914 * 915 * These are the range our thresholds will fall in, corresponding to how 916 * much delalloc we need for the background flusher to kick in. 917 */ 918 919 thresh = calc_available_free_space(fs_info, space_info, 920 BTRFS_RESERVE_FLUSH_ALL); 921 used = space_info->bytes_used + space_info->bytes_reserved + 922 space_info->bytes_readonly + global_rsv_size; 923 if (used < total) 924 thresh += total - used; 925 thresh >>= space_info->clamp; 926 927 used = space_info->bytes_pinned; 928 929 /* 930 * If we have more ordered bytes than delalloc bytes then we're either 931 * doing a lot of DIO, or we simply don't have a lot of delalloc waiting 932 * around. Preemptive flushing is only useful in that it can free up 933 * space before tickets need to wait for things to finish. In the case 934 * of ordered extents, preemptively waiting on ordered extents gets us 935 * nothing, if our reservations are tied up in ordered extents we'll 936 * simply have to slow down writers by forcing them to wait on ordered 937 * extents. 938 * 939 * In the case that ordered is larger than delalloc, only include the 940 * block reserves that we would actually be able to directly reclaim 941 * from. In this case if we're heavy on metadata operations this will 942 * clearly be heavy enough to warrant preemptive flushing. In the case 943 * of heavy DIO or ordered reservations, preemptive flushing will just 944 * waste time and cause us to slow down. 945 * 946 * We want to make sure we truly are maxed out on ordered however, so 947 * cut ordered in half, and if it's still higher than delalloc then we 948 * can keep flushing. This is to avoid the case where we start 949 * flushing, and now delalloc == ordered and we stop preemptively 950 * flushing when we could still have several gigs of delalloc to flush. 951 */ 952 ordered = percpu_counter_read_positive(&fs_info->ordered_bytes) >> 1; 953 delalloc = percpu_counter_read_positive(&fs_info->delalloc_bytes); 954 if (ordered >= delalloc) 955 used += fs_info->delayed_refs_rsv.reserved + 956 fs_info->delayed_block_rsv.reserved; 957 else 958 used += space_info->bytes_may_use - global_rsv_size; 959 960 return (used >= thresh && !btrfs_fs_closing(fs_info) && 961 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state)); 962 } 963 964 static bool steal_from_global_rsv(struct btrfs_fs_info *fs_info, 965 struct btrfs_space_info *space_info, 966 struct reserve_ticket *ticket) 967 { 968 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv; 969 u64 min_bytes; 970 971 if (!ticket->steal) 972 return false; 973 974 if (global_rsv->space_info != space_info) 975 return false; 976 977 spin_lock(&global_rsv->lock); 978 min_bytes = div_factor(global_rsv->size, 1); 979 if (global_rsv->reserved < min_bytes + ticket->bytes) { 980 spin_unlock(&global_rsv->lock); 981 return false; 982 } 983 global_rsv->reserved -= ticket->bytes; 984 remove_ticket(space_info, ticket); 985 ticket->bytes = 0; 986 wake_up(&ticket->wait); 987 space_info->tickets_id++; 988 if (global_rsv->reserved < global_rsv->size) 989 global_rsv->full = 0; 990 spin_unlock(&global_rsv->lock); 991 992 return true; 993 } 994 995 /* 996 * maybe_fail_all_tickets - we've exhausted our flushing, start failing tickets 997 * @fs_info - fs_info for this fs 998 * @space_info - the space info we were flushing 999 * 1000 * We call this when we've exhausted our flushing ability and haven't made 1001 * progress in satisfying tickets. The reservation code handles tickets in 1002 * order, so if there is a large ticket first and then smaller ones we could 1003 * very well satisfy the smaller tickets. This will attempt to wake up any 1004 * tickets in the list to catch this case. 1005 * 1006 * This function returns true if it was able to make progress by clearing out 1007 * other tickets, or if it stumbles across a ticket that was smaller than the 1008 * first ticket. 1009 */ 1010 static bool maybe_fail_all_tickets(struct btrfs_fs_info *fs_info, 1011 struct btrfs_space_info *space_info) 1012 { 1013 struct reserve_ticket *ticket; 1014 u64 tickets_id = space_info->tickets_id; 1015 const bool aborted = BTRFS_FS_ERROR(fs_info); 1016 1017 trace_btrfs_fail_all_tickets(fs_info, space_info); 1018 1019 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) { 1020 btrfs_info(fs_info, "cannot satisfy tickets, dumping space info"); 1021 __btrfs_dump_space_info(fs_info, space_info); 1022 } 1023 1024 while (!list_empty(&space_info->tickets) && 1025 tickets_id == space_info->tickets_id) { 1026 ticket = list_first_entry(&space_info->tickets, 1027 struct reserve_ticket, list); 1028 1029 if (!aborted && steal_from_global_rsv(fs_info, space_info, ticket)) 1030 return true; 1031 1032 if (!aborted && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) 1033 btrfs_info(fs_info, "failing ticket with %llu bytes", 1034 ticket->bytes); 1035 1036 remove_ticket(space_info, ticket); 1037 if (aborted) 1038 ticket->error = -EIO; 1039 else 1040 ticket->error = -ENOSPC; 1041 wake_up(&ticket->wait); 1042 1043 /* 1044 * We're just throwing tickets away, so more flushing may not 1045 * trip over btrfs_try_granting_tickets, so we need to call it 1046 * here to see if we can make progress with the next ticket in 1047 * the list. 1048 */ 1049 if (!aborted) 1050 btrfs_try_granting_tickets(fs_info, space_info); 1051 } 1052 return (tickets_id != space_info->tickets_id); 1053 } 1054 1055 /* 1056 * This is for normal flushers, we can wait all goddamned day if we want to. We 1057 * will loop and continuously try to flush as long as we are making progress. 1058 * We count progress as clearing off tickets each time we have to loop. 1059 */ 1060 static void btrfs_async_reclaim_metadata_space(struct work_struct *work) 1061 { 1062 struct btrfs_fs_info *fs_info; 1063 struct btrfs_space_info *space_info; 1064 u64 to_reclaim; 1065 enum btrfs_flush_state flush_state; 1066 int commit_cycles = 0; 1067 u64 last_tickets_id; 1068 1069 fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work); 1070 space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA); 1071 1072 spin_lock(&space_info->lock); 1073 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info); 1074 if (!to_reclaim) { 1075 space_info->flush = 0; 1076 spin_unlock(&space_info->lock); 1077 return; 1078 } 1079 last_tickets_id = space_info->tickets_id; 1080 spin_unlock(&space_info->lock); 1081 1082 flush_state = FLUSH_DELAYED_ITEMS_NR; 1083 do { 1084 flush_space(fs_info, space_info, to_reclaim, flush_state, false); 1085 spin_lock(&space_info->lock); 1086 if (list_empty(&space_info->tickets)) { 1087 space_info->flush = 0; 1088 spin_unlock(&space_info->lock); 1089 return; 1090 } 1091 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, 1092 space_info); 1093 if (last_tickets_id == space_info->tickets_id) { 1094 flush_state++; 1095 } else { 1096 last_tickets_id = space_info->tickets_id; 1097 flush_state = FLUSH_DELAYED_ITEMS_NR; 1098 if (commit_cycles) 1099 commit_cycles--; 1100 } 1101 1102 /* 1103 * We do not want to empty the system of delalloc unless we're 1104 * under heavy pressure, so allow one trip through the flushing 1105 * logic before we start doing a FLUSH_DELALLOC_FULL. 1106 */ 1107 if (flush_state == FLUSH_DELALLOC_FULL && !commit_cycles) 1108 flush_state++; 1109 1110 /* 1111 * We don't want to force a chunk allocation until we've tried 1112 * pretty hard to reclaim space. Think of the case where we 1113 * freed up a bunch of space and so have a lot of pinned space 1114 * to reclaim. We would rather use that than possibly create a 1115 * underutilized metadata chunk. So if this is our first run 1116 * through the flushing state machine skip ALLOC_CHUNK_FORCE and 1117 * commit the transaction. If nothing has changed the next go 1118 * around then we can force a chunk allocation. 1119 */ 1120 if (flush_state == ALLOC_CHUNK_FORCE && !commit_cycles) 1121 flush_state++; 1122 1123 if (flush_state > COMMIT_TRANS) { 1124 commit_cycles++; 1125 if (commit_cycles > 2) { 1126 if (maybe_fail_all_tickets(fs_info, space_info)) { 1127 flush_state = FLUSH_DELAYED_ITEMS_NR; 1128 commit_cycles--; 1129 } else { 1130 space_info->flush = 0; 1131 } 1132 } else { 1133 flush_state = FLUSH_DELAYED_ITEMS_NR; 1134 } 1135 } 1136 spin_unlock(&space_info->lock); 1137 } while (flush_state <= COMMIT_TRANS); 1138 } 1139 1140 /* 1141 * This handles pre-flushing of metadata space before we get to the point that 1142 * we need to start blocking threads on tickets. The logic here is different 1143 * from the other flush paths because it doesn't rely on tickets to tell us how 1144 * much we need to flush, instead it attempts to keep us below the 80% full 1145 * watermark of space by flushing whichever reservation pool is currently the 1146 * largest. 1147 */ 1148 static void btrfs_preempt_reclaim_metadata_space(struct work_struct *work) 1149 { 1150 struct btrfs_fs_info *fs_info; 1151 struct btrfs_space_info *space_info; 1152 struct btrfs_block_rsv *delayed_block_rsv; 1153 struct btrfs_block_rsv *delayed_refs_rsv; 1154 struct btrfs_block_rsv *global_rsv; 1155 struct btrfs_block_rsv *trans_rsv; 1156 int loops = 0; 1157 1158 fs_info = container_of(work, struct btrfs_fs_info, 1159 preempt_reclaim_work); 1160 space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA); 1161 delayed_block_rsv = &fs_info->delayed_block_rsv; 1162 delayed_refs_rsv = &fs_info->delayed_refs_rsv; 1163 global_rsv = &fs_info->global_block_rsv; 1164 trans_rsv = &fs_info->trans_block_rsv; 1165 1166 spin_lock(&space_info->lock); 1167 while (need_preemptive_reclaim(fs_info, space_info)) { 1168 enum btrfs_flush_state flush; 1169 u64 delalloc_size = 0; 1170 u64 to_reclaim, block_rsv_size; 1171 u64 global_rsv_size = global_rsv->reserved; 1172 1173 loops++; 1174 1175 /* 1176 * We don't have a precise counter for the metadata being 1177 * reserved for delalloc, so we'll approximate it by subtracting 1178 * out the block rsv's space from the bytes_may_use. If that 1179 * amount is higher than the individual reserves, then we can 1180 * assume it's tied up in delalloc reservations. 1181 */ 1182 block_rsv_size = global_rsv_size + 1183 delayed_block_rsv->reserved + 1184 delayed_refs_rsv->reserved + 1185 trans_rsv->reserved; 1186 if (block_rsv_size < space_info->bytes_may_use) 1187 delalloc_size = space_info->bytes_may_use - block_rsv_size; 1188 1189 /* 1190 * We don't want to include the global_rsv in our calculation, 1191 * because that's space we can't touch. Subtract it from the 1192 * block_rsv_size for the next checks. 1193 */ 1194 block_rsv_size -= global_rsv_size; 1195 1196 /* 1197 * We really want to avoid flushing delalloc too much, as it 1198 * could result in poor allocation patterns, so only flush it if 1199 * it's larger than the rest of the pools combined. 1200 */ 1201 if (delalloc_size > block_rsv_size) { 1202 to_reclaim = delalloc_size; 1203 flush = FLUSH_DELALLOC; 1204 } else if (space_info->bytes_pinned > 1205 (delayed_block_rsv->reserved + 1206 delayed_refs_rsv->reserved)) { 1207 to_reclaim = space_info->bytes_pinned; 1208 flush = COMMIT_TRANS; 1209 } else if (delayed_block_rsv->reserved > 1210 delayed_refs_rsv->reserved) { 1211 to_reclaim = delayed_block_rsv->reserved; 1212 flush = FLUSH_DELAYED_ITEMS_NR; 1213 } else { 1214 to_reclaim = delayed_refs_rsv->reserved; 1215 flush = FLUSH_DELAYED_REFS_NR; 1216 } 1217 1218 spin_unlock(&space_info->lock); 1219 1220 /* 1221 * We don't want to reclaim everything, just a portion, so scale 1222 * down the to_reclaim by 1/4. If it takes us down to 0, 1223 * reclaim 1 items worth. 1224 */ 1225 to_reclaim >>= 2; 1226 if (!to_reclaim) 1227 to_reclaim = btrfs_calc_insert_metadata_size(fs_info, 1); 1228 flush_space(fs_info, space_info, to_reclaim, flush, true); 1229 cond_resched(); 1230 spin_lock(&space_info->lock); 1231 } 1232 1233 /* We only went through once, back off our clamping. */ 1234 if (loops == 1 && !space_info->reclaim_size) 1235 space_info->clamp = max(1, space_info->clamp - 1); 1236 trace_btrfs_done_preemptive_reclaim(fs_info, space_info); 1237 spin_unlock(&space_info->lock); 1238 } 1239 1240 /* 1241 * FLUSH_DELALLOC_WAIT: 1242 * Space is freed from flushing delalloc in one of two ways. 1243 * 1244 * 1) compression is on and we allocate less space than we reserved 1245 * 2) we are overwriting existing space 1246 * 1247 * For #1 that extra space is reclaimed as soon as the delalloc pages are 1248 * COWed, by way of btrfs_add_reserved_bytes() which adds the actual extent 1249 * length to ->bytes_reserved, and subtracts the reserved space from 1250 * ->bytes_may_use. 1251 * 1252 * For #2 this is trickier. Once the ordered extent runs we will drop the 1253 * extent in the range we are overwriting, which creates a delayed ref for 1254 * that freed extent. This however is not reclaimed until the transaction 1255 * commits, thus the next stages. 1256 * 1257 * RUN_DELAYED_IPUTS 1258 * If we are freeing inodes, we want to make sure all delayed iputs have 1259 * completed, because they could have been on an inode with i_nlink == 0, and 1260 * thus have been truncated and freed up space. But again this space is not 1261 * immediately re-usable, it comes in the form of a delayed ref, which must be 1262 * run and then the transaction must be committed. 1263 * 1264 * COMMIT_TRANS 1265 * This is where we reclaim all of the pinned space generated by running the 1266 * iputs 1267 * 1268 * ALLOC_CHUNK_FORCE 1269 * For data we start with alloc chunk force, however we could have been full 1270 * before, and then the transaction commit could have freed new block groups, 1271 * so if we now have space to allocate do the force chunk allocation. 1272 */ 1273 static const enum btrfs_flush_state data_flush_states[] = { 1274 FLUSH_DELALLOC_FULL, 1275 RUN_DELAYED_IPUTS, 1276 COMMIT_TRANS, 1277 ALLOC_CHUNK_FORCE, 1278 }; 1279 1280 static void btrfs_async_reclaim_data_space(struct work_struct *work) 1281 { 1282 struct btrfs_fs_info *fs_info; 1283 struct btrfs_space_info *space_info; 1284 u64 last_tickets_id; 1285 enum btrfs_flush_state flush_state = 0; 1286 1287 fs_info = container_of(work, struct btrfs_fs_info, async_data_reclaim_work); 1288 space_info = fs_info->data_sinfo; 1289 1290 spin_lock(&space_info->lock); 1291 if (list_empty(&space_info->tickets)) { 1292 space_info->flush = 0; 1293 spin_unlock(&space_info->lock); 1294 return; 1295 } 1296 last_tickets_id = space_info->tickets_id; 1297 spin_unlock(&space_info->lock); 1298 1299 while (!space_info->full) { 1300 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE, false); 1301 spin_lock(&space_info->lock); 1302 if (list_empty(&space_info->tickets)) { 1303 space_info->flush = 0; 1304 spin_unlock(&space_info->lock); 1305 return; 1306 } 1307 1308 /* Something happened, fail everything and bail. */ 1309 if (BTRFS_FS_ERROR(fs_info)) 1310 goto aborted_fs; 1311 last_tickets_id = space_info->tickets_id; 1312 spin_unlock(&space_info->lock); 1313 } 1314 1315 while (flush_state < ARRAY_SIZE(data_flush_states)) { 1316 flush_space(fs_info, space_info, U64_MAX, 1317 data_flush_states[flush_state], false); 1318 spin_lock(&space_info->lock); 1319 if (list_empty(&space_info->tickets)) { 1320 space_info->flush = 0; 1321 spin_unlock(&space_info->lock); 1322 return; 1323 } 1324 1325 if (last_tickets_id == space_info->tickets_id) { 1326 flush_state++; 1327 } else { 1328 last_tickets_id = space_info->tickets_id; 1329 flush_state = 0; 1330 } 1331 1332 if (flush_state >= ARRAY_SIZE(data_flush_states)) { 1333 if (space_info->full) { 1334 if (maybe_fail_all_tickets(fs_info, space_info)) 1335 flush_state = 0; 1336 else 1337 space_info->flush = 0; 1338 } else { 1339 flush_state = 0; 1340 } 1341 1342 /* Something happened, fail everything and bail. */ 1343 if (BTRFS_FS_ERROR(fs_info)) 1344 goto aborted_fs; 1345 1346 } 1347 spin_unlock(&space_info->lock); 1348 } 1349 return; 1350 1351 aborted_fs: 1352 maybe_fail_all_tickets(fs_info, space_info); 1353 space_info->flush = 0; 1354 spin_unlock(&space_info->lock); 1355 } 1356 1357 void btrfs_init_async_reclaim_work(struct btrfs_fs_info *fs_info) 1358 { 1359 INIT_WORK(&fs_info->async_reclaim_work, btrfs_async_reclaim_metadata_space); 1360 INIT_WORK(&fs_info->async_data_reclaim_work, btrfs_async_reclaim_data_space); 1361 INIT_WORK(&fs_info->preempt_reclaim_work, 1362 btrfs_preempt_reclaim_metadata_space); 1363 } 1364 1365 static const enum btrfs_flush_state priority_flush_states[] = { 1366 FLUSH_DELAYED_ITEMS_NR, 1367 FLUSH_DELAYED_ITEMS, 1368 ALLOC_CHUNK, 1369 }; 1370 1371 static const enum btrfs_flush_state evict_flush_states[] = { 1372 FLUSH_DELAYED_ITEMS_NR, 1373 FLUSH_DELAYED_ITEMS, 1374 FLUSH_DELAYED_REFS_NR, 1375 FLUSH_DELAYED_REFS, 1376 FLUSH_DELALLOC, 1377 FLUSH_DELALLOC_WAIT, 1378 FLUSH_DELALLOC_FULL, 1379 ALLOC_CHUNK, 1380 COMMIT_TRANS, 1381 }; 1382 1383 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info, 1384 struct btrfs_space_info *space_info, 1385 struct reserve_ticket *ticket, 1386 const enum btrfs_flush_state *states, 1387 int states_nr) 1388 { 1389 u64 to_reclaim; 1390 int flush_state = 0; 1391 1392 spin_lock(&space_info->lock); 1393 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info); 1394 /* 1395 * This is the priority reclaim path, so to_reclaim could be >0 still 1396 * because we may have only satisfied the priority tickets and still 1397 * left non priority tickets on the list. We would then have 1398 * to_reclaim but ->bytes == 0. 1399 */ 1400 if (ticket->bytes == 0) { 1401 spin_unlock(&space_info->lock); 1402 return; 1403 } 1404 1405 while (flush_state < states_nr) { 1406 spin_unlock(&space_info->lock); 1407 flush_space(fs_info, space_info, to_reclaim, states[flush_state], 1408 false); 1409 flush_state++; 1410 spin_lock(&space_info->lock); 1411 if (ticket->bytes == 0) { 1412 spin_unlock(&space_info->lock); 1413 return; 1414 } 1415 } 1416 1417 /* Attempt to steal from the global rsv if we can. */ 1418 if (!steal_from_global_rsv(fs_info, space_info, ticket)) { 1419 ticket->error = -ENOSPC; 1420 remove_ticket(space_info, ticket); 1421 } 1422 1423 /* 1424 * We must run try_granting_tickets here because we could be a large 1425 * ticket in front of a smaller ticket that can now be satisfied with 1426 * the available space. 1427 */ 1428 btrfs_try_granting_tickets(fs_info, space_info); 1429 spin_unlock(&space_info->lock); 1430 } 1431 1432 static void priority_reclaim_data_space(struct btrfs_fs_info *fs_info, 1433 struct btrfs_space_info *space_info, 1434 struct reserve_ticket *ticket) 1435 { 1436 spin_lock(&space_info->lock); 1437 1438 /* We could have been granted before we got here. */ 1439 if (ticket->bytes == 0) { 1440 spin_unlock(&space_info->lock); 1441 return; 1442 } 1443 1444 while (!space_info->full) { 1445 spin_unlock(&space_info->lock); 1446 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE, false); 1447 spin_lock(&space_info->lock); 1448 if (ticket->bytes == 0) { 1449 spin_unlock(&space_info->lock); 1450 return; 1451 } 1452 } 1453 1454 ticket->error = -ENOSPC; 1455 remove_ticket(space_info, ticket); 1456 btrfs_try_granting_tickets(fs_info, space_info); 1457 spin_unlock(&space_info->lock); 1458 } 1459 1460 static void wait_reserve_ticket(struct btrfs_fs_info *fs_info, 1461 struct btrfs_space_info *space_info, 1462 struct reserve_ticket *ticket) 1463 1464 { 1465 DEFINE_WAIT(wait); 1466 int ret = 0; 1467 1468 spin_lock(&space_info->lock); 1469 while (ticket->bytes > 0 && ticket->error == 0) { 1470 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE); 1471 if (ret) { 1472 /* 1473 * Delete us from the list. After we unlock the space 1474 * info, we don't want the async reclaim job to reserve 1475 * space for this ticket. If that would happen, then the 1476 * ticket's task would not known that space was reserved 1477 * despite getting an error, resulting in a space leak 1478 * (bytes_may_use counter of our space_info). 1479 */ 1480 remove_ticket(space_info, ticket); 1481 ticket->error = -EINTR; 1482 break; 1483 } 1484 spin_unlock(&space_info->lock); 1485 1486 schedule(); 1487 1488 finish_wait(&ticket->wait, &wait); 1489 spin_lock(&space_info->lock); 1490 } 1491 spin_unlock(&space_info->lock); 1492 } 1493 1494 /** 1495 * Do the appropriate flushing and waiting for a ticket 1496 * 1497 * @fs_info: the filesystem 1498 * @space_info: space info for the reservation 1499 * @ticket: ticket for the reservation 1500 * @start_ns: timestamp when the reservation started 1501 * @orig_bytes: amount of bytes originally reserved 1502 * @flush: how much we can flush 1503 * 1504 * This does the work of figuring out how to flush for the ticket, waiting for 1505 * the reservation, and returning the appropriate error if there is one. 1506 */ 1507 static int handle_reserve_ticket(struct btrfs_fs_info *fs_info, 1508 struct btrfs_space_info *space_info, 1509 struct reserve_ticket *ticket, 1510 u64 start_ns, u64 orig_bytes, 1511 enum btrfs_reserve_flush_enum flush) 1512 { 1513 int ret; 1514 1515 switch (flush) { 1516 case BTRFS_RESERVE_FLUSH_DATA: 1517 case BTRFS_RESERVE_FLUSH_ALL: 1518 case BTRFS_RESERVE_FLUSH_ALL_STEAL: 1519 wait_reserve_ticket(fs_info, space_info, ticket); 1520 break; 1521 case BTRFS_RESERVE_FLUSH_LIMIT: 1522 priority_reclaim_metadata_space(fs_info, space_info, ticket, 1523 priority_flush_states, 1524 ARRAY_SIZE(priority_flush_states)); 1525 break; 1526 case BTRFS_RESERVE_FLUSH_EVICT: 1527 priority_reclaim_metadata_space(fs_info, space_info, ticket, 1528 evict_flush_states, 1529 ARRAY_SIZE(evict_flush_states)); 1530 break; 1531 case BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE: 1532 priority_reclaim_data_space(fs_info, space_info, ticket); 1533 break; 1534 default: 1535 ASSERT(0); 1536 break; 1537 } 1538 1539 ret = ticket->error; 1540 ASSERT(list_empty(&ticket->list)); 1541 /* 1542 * Check that we can't have an error set if the reservation succeeded, 1543 * as that would confuse tasks and lead them to error out without 1544 * releasing reserved space (if an error happens the expectation is that 1545 * space wasn't reserved at all). 1546 */ 1547 ASSERT(!(ticket->bytes == 0 && ticket->error)); 1548 trace_btrfs_reserve_ticket(fs_info, space_info->flags, orig_bytes, 1549 start_ns, flush, ticket->error); 1550 return ret; 1551 } 1552 1553 /* 1554 * This returns true if this flush state will go through the ordinary flushing 1555 * code. 1556 */ 1557 static inline bool is_normal_flushing(enum btrfs_reserve_flush_enum flush) 1558 { 1559 return (flush == BTRFS_RESERVE_FLUSH_ALL) || 1560 (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL); 1561 } 1562 1563 static inline void maybe_clamp_preempt(struct btrfs_fs_info *fs_info, 1564 struct btrfs_space_info *space_info) 1565 { 1566 u64 ordered = percpu_counter_sum_positive(&fs_info->ordered_bytes); 1567 u64 delalloc = percpu_counter_sum_positive(&fs_info->delalloc_bytes); 1568 1569 /* 1570 * If we're heavy on ordered operations then clamping won't help us. We 1571 * need to clamp specifically to keep up with dirty'ing buffered 1572 * writers, because there's not a 1:1 correlation of writing delalloc 1573 * and freeing space, like there is with flushing delayed refs or 1574 * delayed nodes. If we're already more ordered than delalloc then 1575 * we're keeping up, otherwise we aren't and should probably clamp. 1576 */ 1577 if (ordered < delalloc) 1578 space_info->clamp = min(space_info->clamp + 1, 8); 1579 } 1580 1581 static inline bool can_steal(enum btrfs_reserve_flush_enum flush) 1582 { 1583 return (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL || 1584 flush == BTRFS_RESERVE_FLUSH_EVICT); 1585 } 1586 1587 /* 1588 * NO_FLUSH and FLUSH_EMERGENCY don't want to create a ticket, they just want to 1589 * fail as quickly as possible. 1590 */ 1591 static inline bool can_ticket(enum btrfs_reserve_flush_enum flush) 1592 { 1593 return (flush != BTRFS_RESERVE_NO_FLUSH && 1594 flush != BTRFS_RESERVE_FLUSH_EMERGENCY); 1595 } 1596 1597 /** 1598 * Try to reserve bytes from the block_rsv's space 1599 * 1600 * @fs_info: the filesystem 1601 * @space_info: space info we want to allocate from 1602 * @orig_bytes: number of bytes we want 1603 * @flush: whether or not we can flush to make our reservation 1604 * 1605 * This will reserve orig_bytes number of bytes from the space info associated 1606 * with the block_rsv. If there is not enough space it will make an attempt to 1607 * flush out space to make room. It will do this by flushing delalloc if 1608 * possible or committing the transaction. If flush is 0 then no attempts to 1609 * regain reservations will be made and this will fail if there is not enough 1610 * space already. 1611 */ 1612 static int __reserve_bytes(struct btrfs_fs_info *fs_info, 1613 struct btrfs_space_info *space_info, u64 orig_bytes, 1614 enum btrfs_reserve_flush_enum flush) 1615 { 1616 struct work_struct *async_work; 1617 struct reserve_ticket ticket; 1618 u64 start_ns = 0; 1619 u64 used; 1620 int ret = 0; 1621 bool pending_tickets; 1622 1623 ASSERT(orig_bytes); 1624 ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL); 1625 1626 if (flush == BTRFS_RESERVE_FLUSH_DATA) 1627 async_work = &fs_info->async_data_reclaim_work; 1628 else 1629 async_work = &fs_info->async_reclaim_work; 1630 1631 spin_lock(&space_info->lock); 1632 ret = -ENOSPC; 1633 used = btrfs_space_info_used(space_info, true); 1634 1635 /* 1636 * We don't want NO_FLUSH allocations to jump everybody, they can 1637 * generally handle ENOSPC in a different way, so treat them the same as 1638 * normal flushers when it comes to skipping pending tickets. 1639 */ 1640 if (is_normal_flushing(flush) || (flush == BTRFS_RESERVE_NO_FLUSH)) 1641 pending_tickets = !list_empty(&space_info->tickets) || 1642 !list_empty(&space_info->priority_tickets); 1643 else 1644 pending_tickets = !list_empty(&space_info->priority_tickets); 1645 1646 /* 1647 * Carry on if we have enough space (short-circuit) OR call 1648 * can_overcommit() to ensure we can overcommit to continue. 1649 */ 1650 if (!pending_tickets && 1651 ((used + orig_bytes <= writable_total_bytes(fs_info, space_info)) || 1652 btrfs_can_overcommit(fs_info, space_info, orig_bytes, flush))) { 1653 btrfs_space_info_update_bytes_may_use(fs_info, space_info, 1654 orig_bytes); 1655 ret = 0; 1656 } 1657 1658 /* 1659 * Things are dire, we need to make a reservation so we don't abort. We 1660 * will let this reservation go through as long as we have actual space 1661 * left to allocate for the block. 1662 */ 1663 if (ret && unlikely(flush == BTRFS_RESERVE_FLUSH_EMERGENCY)) { 1664 used = btrfs_space_info_used(space_info, false); 1665 if (used + orig_bytes <= 1666 writable_total_bytes(fs_info, space_info)) { 1667 btrfs_space_info_update_bytes_may_use(fs_info, space_info, 1668 orig_bytes); 1669 ret = 0; 1670 } 1671 } 1672 1673 /* 1674 * If we couldn't make a reservation then setup our reservation ticket 1675 * and kick the async worker if it's not already running. 1676 * 1677 * If we are a priority flusher then we just need to add our ticket to 1678 * the list and we will do our own flushing further down. 1679 */ 1680 if (ret && can_ticket(flush)) { 1681 ticket.bytes = orig_bytes; 1682 ticket.error = 0; 1683 space_info->reclaim_size += ticket.bytes; 1684 init_waitqueue_head(&ticket.wait); 1685 ticket.steal = can_steal(flush); 1686 if (trace_btrfs_reserve_ticket_enabled()) 1687 start_ns = ktime_get_ns(); 1688 1689 if (flush == BTRFS_RESERVE_FLUSH_ALL || 1690 flush == BTRFS_RESERVE_FLUSH_ALL_STEAL || 1691 flush == BTRFS_RESERVE_FLUSH_DATA) { 1692 list_add_tail(&ticket.list, &space_info->tickets); 1693 if (!space_info->flush) { 1694 /* 1695 * We were forced to add a reserve ticket, so 1696 * our preemptive flushing is unable to keep 1697 * up. Clamp down on the threshold for the 1698 * preemptive flushing in order to keep up with 1699 * the workload. 1700 */ 1701 maybe_clamp_preempt(fs_info, space_info); 1702 1703 space_info->flush = 1; 1704 trace_btrfs_trigger_flush(fs_info, 1705 space_info->flags, 1706 orig_bytes, flush, 1707 "enospc"); 1708 queue_work(system_unbound_wq, async_work); 1709 } 1710 } else { 1711 list_add_tail(&ticket.list, 1712 &space_info->priority_tickets); 1713 } 1714 } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) { 1715 /* 1716 * We will do the space reservation dance during log replay, 1717 * which means we won't have fs_info->fs_root set, so don't do 1718 * the async reclaim as we will panic. 1719 */ 1720 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) && 1721 !work_busy(&fs_info->preempt_reclaim_work) && 1722 need_preemptive_reclaim(fs_info, space_info)) { 1723 trace_btrfs_trigger_flush(fs_info, space_info->flags, 1724 orig_bytes, flush, "preempt"); 1725 queue_work(system_unbound_wq, 1726 &fs_info->preempt_reclaim_work); 1727 } 1728 } 1729 spin_unlock(&space_info->lock); 1730 if (!ret || !can_ticket(flush)) 1731 return ret; 1732 1733 return handle_reserve_ticket(fs_info, space_info, &ticket, start_ns, 1734 orig_bytes, flush); 1735 } 1736 1737 /** 1738 * Trye to reserve metadata bytes from the block_rsv's space 1739 * 1740 * @fs_info: the filesystem 1741 * @block_rsv: block_rsv we're allocating for 1742 * @orig_bytes: number of bytes we want 1743 * @flush: whether or not we can flush to make our reservation 1744 * 1745 * This will reserve orig_bytes number of bytes from the space info associated 1746 * with the block_rsv. If there is not enough space it will make an attempt to 1747 * flush out space to make room. It will do this by flushing delalloc if 1748 * possible or committing the transaction. If flush is 0 then no attempts to 1749 * regain reservations will be made and this will fail if there is not enough 1750 * space already. 1751 */ 1752 int btrfs_reserve_metadata_bytes(struct btrfs_fs_info *fs_info, 1753 struct btrfs_block_rsv *block_rsv, 1754 u64 orig_bytes, 1755 enum btrfs_reserve_flush_enum flush) 1756 { 1757 int ret; 1758 1759 ret = __reserve_bytes(fs_info, block_rsv->space_info, orig_bytes, flush); 1760 if (ret == -ENOSPC) { 1761 trace_btrfs_space_reservation(fs_info, "space_info:enospc", 1762 block_rsv->space_info->flags, 1763 orig_bytes, 1); 1764 1765 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) 1766 btrfs_dump_space_info(fs_info, block_rsv->space_info, 1767 orig_bytes, 0); 1768 } 1769 return ret; 1770 } 1771 1772 /** 1773 * Try to reserve data bytes for an allocation 1774 * 1775 * @fs_info: the filesystem 1776 * @bytes: number of bytes we need 1777 * @flush: how we are allowed to flush 1778 * 1779 * This will reserve bytes from the data space info. If there is not enough 1780 * space then we will attempt to flush space as specified by flush. 1781 */ 1782 int btrfs_reserve_data_bytes(struct btrfs_fs_info *fs_info, u64 bytes, 1783 enum btrfs_reserve_flush_enum flush) 1784 { 1785 struct btrfs_space_info *data_sinfo = fs_info->data_sinfo; 1786 int ret; 1787 1788 ASSERT(flush == BTRFS_RESERVE_FLUSH_DATA || 1789 flush == BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE || 1790 flush == BTRFS_RESERVE_NO_FLUSH); 1791 ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_DATA); 1792 1793 ret = __reserve_bytes(fs_info, data_sinfo, bytes, flush); 1794 if (ret == -ENOSPC) { 1795 trace_btrfs_space_reservation(fs_info, "space_info:enospc", 1796 data_sinfo->flags, bytes, 1); 1797 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) 1798 btrfs_dump_space_info(fs_info, data_sinfo, bytes, 0); 1799 } 1800 return ret; 1801 } 1802 1803 /* Dump all the space infos when we abort a transaction due to ENOSPC. */ 1804 __cold void btrfs_dump_space_info_for_trans_abort(struct btrfs_fs_info *fs_info) 1805 { 1806 struct btrfs_space_info *space_info; 1807 1808 btrfs_info(fs_info, "dumping space info:"); 1809 list_for_each_entry(space_info, &fs_info->space_info, list) { 1810 spin_lock(&space_info->lock); 1811 __btrfs_dump_space_info(fs_info, space_info); 1812 spin_unlock(&space_info->lock); 1813 } 1814 dump_global_block_rsv(fs_info); 1815 } 1816