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