git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/btree_cache.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "bkey_buf.h"
   5 #include "btree_cache.h"
   6 #include "btree_io.h"
   7 #include "btree_iter.h"
   8 #include "btree_locking.h"
   9 #include "debug.h"
  10 #include "errcode.h"
  11 #include "error.h"
  12
  13 #include <linux/prefetch.h>
  14 #include <linux/sched/mm.h>
  15 #include <trace/events/bcachefs.h>
  16
  17 #define BTREE_CACHE_NOT_FREED_INCREMENT(counter) \
  18 do {                                             \
  19         if (shrinker_counter)                    \
  20                 bc->not_freed_##counter++;       \
  21 } while (0)
  22
  23 const char * const bch2_btree_node_flags[] = {
  24 #define x(f)    #f,
  25         BTREE_FLAGS()
  26 #undef x
  27         NULL
  28 };
  29
  30 void bch2_recalc_btree_reserve(struct bch_fs *c)
  31 {
  32         unsigned i, reserve = 16;
  33
  34         if (!c->btree_roots[0].b)
  35                 reserve += 8;
  36
  37         for (i = 0; i < BTREE_ID_NR; i++)
  38                 if (c->btree_roots[i].b)
  39                         reserve += min_t(unsigned, 1,
  40                                          c->btree_roots[i].b->c.level) * 8;
  41
  42         c->btree_cache.reserve = reserve;
  43 }
  44
  45 static inline unsigned btree_cache_can_free(struct btree_cache *bc)
  46 {
  47         return max_t(int, 0, bc->used - bc->reserve);
  48 }
  49
  50 static void btree_node_to_freedlist(struct btree_cache *bc, struct btree *b)
  51 {
  52         if (b->c.lock.readers)
  53                 list_move(&b->list, &bc->freed_pcpu);
  54         else
  55                 list_move(&b->list, &bc->freed_nonpcpu);
  56 }
  57
  58 static void btree_node_data_free(struct bch_fs *c, struct btree *b)
  59 {
  60         struct btree_cache *bc = &c->btree_cache;
  61
  62         EBUG_ON(btree_node_write_in_flight(b));
  63
  64         kvpfree(b->data, btree_bytes(c));
  65         b->data = NULL;
  66 #ifdef __KERNEL__
  67         vfree(b->aux_data);
  68 #else
  69         munmap(b->aux_data, btree_aux_data_bytes(b));
  70 #endif
  71         b->aux_data = NULL;
  72
  73         bc->used--;
  74
  75         btree_node_to_freedlist(bc, b);
  76 }
  77
  78 static int bch2_btree_cache_cmp_fn(struct rhashtable_compare_arg *arg,
  79                                    const void *obj)
  80 {
  81         const struct btree *b = obj;
  82         const u64 *v = arg->key;
  83
  84         return b->hash_val == *v ? 0 : 1;
  85 }
  86
  87 static const struct rhashtable_params bch_btree_cache_params = {
  88         .head_offset    = offsetof(struct btree, hash),
  89         .key_offset     = offsetof(struct btree, hash_val),
  90         .key_len        = sizeof(u64),
  91         .obj_cmpfn      = bch2_btree_cache_cmp_fn,
  92 };
  93
  94 static int btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
  95 {
  96         BUG_ON(b->data || b->aux_data);
  97
  98         b->data = kvpmalloc(btree_bytes(c), gfp);
  99         if (!b->data)
 100                 return -ENOMEM;
 101 #ifdef __KERNEL__
 102         b->aux_data = vmalloc_exec(btree_aux_data_bytes(b), gfp);
 103 #else
 104         b->aux_data = mmap(NULL, btree_aux_data_bytes(b),
 105                            PROT_READ|PROT_WRITE|PROT_EXEC,
 106                            MAP_PRIVATE|MAP_ANONYMOUS, 0, 0);
 107         if (b->aux_data == MAP_FAILED)
 108                 b->aux_data = NULL;
 109 #endif
 110         if (!b->aux_data) {
 111                 kvpfree(b->data, btree_bytes(c));
 112                 b->data = NULL;
 113                 return -ENOMEM;
 114         }
 115
 116         return 0;
 117 }
 118
 119 static struct btree *__btree_node_mem_alloc(struct bch_fs *c, gfp_t gfp)
 120 {
 121         struct btree *b;
 122
 123         b = kzalloc(sizeof(struct btree), gfp);
 124         if (!b)
 125                 return NULL;
 126
 127         bkey_btree_ptr_init(&b->key);
 128         __six_lock_init(&b->c.lock, "b->c.lock", &bch2_btree_node_lock_key);
 129 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 130         lockdep_set_no_check_recursion(&b->c.lock.dep_map);
 131 #endif
 132         INIT_LIST_HEAD(&b->list);
 133         INIT_LIST_HEAD(&b->write_blocked);
 134         b->byte_order = ilog2(btree_bytes(c));
 135         return b;
 136 }
 137
 138 struct btree *__bch2_btree_node_mem_alloc(struct bch_fs *c)
 139 {
 140         struct btree_cache *bc = &c->btree_cache;
 141         struct btree *b;
 142
 143         b = __btree_node_mem_alloc(c, GFP_KERNEL);
 144         if (!b)
 145                 return NULL;
 146
 147         if (btree_node_data_alloc(c, b, GFP_KERNEL)) {
 148                 kfree(b);
 149                 return NULL;
 150         }
 151
 152         bc->used++;
 153         list_add(&b->list, &bc->freeable);
 154         return b;
 155 }
 156
 157 /* Btree in memory cache - hash table */
 158
 159 void bch2_btree_node_hash_remove(struct btree_cache *bc, struct btree *b)
 160 {
 161         int ret = rhashtable_remove_fast(&bc->table, &b->hash, bch_btree_cache_params);
 162
 163         BUG_ON(ret);
 164
 165         /* Cause future lookups for this node to fail: */
 166         b->hash_val = 0;
 167 }
 168
 169 int __bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b)
 170 {
 171         BUG_ON(b->hash_val);
 172         b->hash_val = btree_ptr_hash_val(&b->key);
 173
 174         return rhashtable_lookup_insert_fast(&bc->table, &b->hash,
 175                                              bch_btree_cache_params);
 176 }
 177
 178 int bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b,
 179                                 unsigned level, enum btree_id id)
 180 {
 181         int ret;
 182
 183         b->c.level      = level;
 184         b->c.btree_id   = id;
 185
 186         mutex_lock(&bc->lock);
 187         ret = __bch2_btree_node_hash_insert(bc, b);
 188         if (!ret)
 189                 list_add_tail(&b->list, &bc->live);
 190         mutex_unlock(&bc->lock);
 191
 192         return ret;
 193 }
 194
 195 __flatten
 196 static inline struct btree *btree_cache_find(struct btree_cache *bc,
 197                                      const struct bkey_i *k)
 198 {
 199         u64 v = btree_ptr_hash_val(k);
 200
 201         return rhashtable_lookup_fast(&bc->table, &v, bch_btree_cache_params);
 202 }
 203
 204 /*
 205  * this version is for btree nodes that have already been freed (we're not
 206  * reaping a real btree node)
 207  */
 208 static int __btree_node_reclaim(struct bch_fs *c, struct btree *b, bool flush, bool shrinker_counter)
 209 {
 210         struct btree_cache *bc = &c->btree_cache;
 211         int ret = 0;
 212
 213         lockdep_assert_held(&bc->lock);
 214 wait_on_io:
 215         if (b->flags & ((1U << BTREE_NODE_dirty)|
 216                         (1U << BTREE_NODE_read_in_flight)|
 217                         (1U << BTREE_NODE_write_in_flight))) {
 218                 if (!flush) {
 219                         if (btree_node_dirty(b))
 220                                 BTREE_CACHE_NOT_FREED_INCREMENT(dirty);
 221                         else if (btree_node_read_in_flight(b))
 222                                 BTREE_CACHE_NOT_FREED_INCREMENT(read_in_flight);
 223                         else if (btree_node_write_in_flight(b))
 224                                 BTREE_CACHE_NOT_FREED_INCREMENT(write_in_flight);
 225                         return -ENOMEM;
 226                 }
 227
 228                 /* XXX: waiting on IO with btree cache lock held */
 229                 bch2_btree_node_wait_on_read(b);
 230                 bch2_btree_node_wait_on_write(b);
 231         }
 232
 233         if (!six_trylock_intent(&b->c.lock)) {
 234                 BTREE_CACHE_NOT_FREED_INCREMENT(lock_intent);
 235                 return -ENOMEM;
 236         }
 237
 238         if (!six_trylock_write(&b->c.lock)) {
 239                 BTREE_CACHE_NOT_FREED_INCREMENT(lock_write);
 240                 goto out_unlock_intent;
 241         }
 242
 243         /* recheck under lock */
 244         if (b->flags & ((1U << BTREE_NODE_read_in_flight)|
 245                         (1U << BTREE_NODE_write_in_flight))) {
 246                 if (!flush) {
 247                         if (btree_node_read_in_flight(b))
 248                                 BTREE_CACHE_NOT_FREED_INCREMENT(read_in_flight);
 249                         else if (btree_node_write_in_flight(b))
 250                                 BTREE_CACHE_NOT_FREED_INCREMENT(write_in_flight);
 251                         goto out_unlock;
 252                 }
 253                 six_unlock_write(&b->c.lock);
 254                 six_unlock_intent(&b->c.lock);
 255                 goto wait_on_io;
 256         }
 257
 258         if (btree_node_noevict(b)) {
 259                 BTREE_CACHE_NOT_FREED_INCREMENT(noevict);
 260                 goto out_unlock;
 261         }
 262         if (btree_node_write_blocked(b)) {
 263                 BTREE_CACHE_NOT_FREED_INCREMENT(write_blocked);
 264                 goto out_unlock;
 265         }
 266         if (btree_node_will_make_reachable(b)) {
 267                 BTREE_CACHE_NOT_FREED_INCREMENT(will_make_reachable);
 268                 goto out_unlock;
 269         }
 270
 271         if (btree_node_dirty(b)) {
 272                 if (!flush) {
 273                         BTREE_CACHE_NOT_FREED_INCREMENT(dirty);
 274                         goto out_unlock;
 275                 }
 276                 /*
 277                  * Using the underscore version because we don't want to compact
 278                  * bsets after the write, since this node is about to be evicted
 279                  * - unless btree verify mode is enabled, since it runs out of
 280                  * the post write cleanup:
 281                  */
 282                 if (bch2_verify_btree_ondisk)
 283                         bch2_btree_node_write(c, b, SIX_LOCK_intent, 0);
 284                 else
 285                         __bch2_btree_node_write(c, b, 0);
 286
 287                 six_unlock_write(&b->c.lock);
 288                 six_unlock_intent(&b->c.lock);
 289                 goto wait_on_io;
 290         }
 291 out:
 292         if (b->hash_val && !ret)
 293                 trace_and_count(c, btree_cache_reap, c, b);
 294         return ret;
 295 out_unlock:
 296         six_unlock_write(&b->c.lock);
 297 out_unlock_intent:
 298         six_unlock_intent(&b->c.lock);
 299         ret = -ENOMEM;
 300         goto out;
 301 }
 302
 303 static int btree_node_reclaim(struct bch_fs *c, struct btree *b, bool shrinker_counter)
 304 {
 305         return __btree_node_reclaim(c, b, false, shrinker_counter);
 306 }
 307
 308 static int btree_node_write_and_reclaim(struct bch_fs *c, struct btree *b)
 309 {
 310         return __btree_node_reclaim(c, b, true, false);
 311 }
 312
 313 static unsigned long bch2_btree_cache_scan(struct shrinker *shrink,
 314                                            struct shrink_control *sc)
 315 {
 316         struct bch_fs *c = container_of(shrink, struct bch_fs,
 317                                         btree_cache.shrink);
 318         struct btree_cache *bc = &c->btree_cache;
 319         struct btree *b, *t;
 320         unsigned long nr = sc->nr_to_scan;
 321         unsigned long can_free = 0;
 322         unsigned long freed = 0;
 323         unsigned long touched = 0;
 324         unsigned i, flags;
 325         unsigned long ret = SHRINK_STOP;
 326         bool trigger_writes = atomic_read(&bc->dirty) + nr >=
 327                 bc->used * 3 / 4;
 328
 329         if (bch2_btree_shrinker_disabled)
 330                 return SHRINK_STOP;
 331
 332         mutex_lock(&bc->lock);
 333         flags = memalloc_nofs_save();
 334
 335         /*
 336          * It's _really_ critical that we don't free too many btree nodes - we
 337          * have to always leave ourselves a reserve. The reserve is how we
 338          * guarantee that allocating memory for a new btree node can always
 339          * succeed, so that inserting keys into the btree can always succeed and
 340          * IO can always make forward progress:
 341          */
 342         can_free = btree_cache_can_free(bc);
 343         nr = min_t(unsigned long, nr, can_free);
 344
 345         i = 0;
 346         list_for_each_entry_safe(b, t, &bc->freeable, list) {
 347                 /*
 348                  * Leave a few nodes on the freeable list, so that a btree split
 349                  * won't have to hit the system allocator:
 350                  */
 351                 if (++i <= 3)
 352                         continue;
 353
 354                 touched++;
 355
 356                 if (touched >= nr)
 357                         goto out;
 358
 359                 if (!btree_node_reclaim(c, b, true)) {
 360                         btree_node_data_free(c, b);
 361                         six_unlock_write(&b->c.lock);
 362                         six_unlock_intent(&b->c.lock);
 363                         freed++;
 364                         bc->freed++;
 365                 }
 366         }
 367 restart:
 368         list_for_each_entry_safe(b, t, &bc->live, list) {
 369                 touched++;
 370
 371                 if (btree_node_accessed(b)) {
 372                         clear_btree_node_accessed(b);
 373                         bc->not_freed_access_bit++;
 374                 } else if (!btree_node_reclaim(c, b, true)) {
 375                         freed++;
 376                         btree_node_data_free(c, b);
 377                         bc->freed++;
 378
 379                         bch2_btree_node_hash_remove(bc, b);
 380                         six_unlock_write(&b->c.lock);
 381                         six_unlock_intent(&b->c.lock);
 382
 383                         if (freed == nr)
 384                                 goto out_rotate;
 385                 } else if (trigger_writes &&
 386                            btree_node_dirty(b) &&
 387                            !btree_node_will_make_reachable(b) &&
 388                            !btree_node_write_blocked(b) &&
 389                            six_trylock_read(&b->c.lock)) {
 390                         list_move(&bc->live, &b->list);
 391                         mutex_unlock(&bc->lock);
 392                         __bch2_btree_node_write(c, b, 0);
 393                         six_unlock_read(&b->c.lock);
 394                         if (touched >= nr)
 395                                 goto out_nounlock;
 396                         mutex_lock(&bc->lock);
 397                         goto restart;
 398                 }
 399
 400                 if (touched >= nr)
 401                         break;
 402         }
 403 out_rotate:
 404         if (&t->list != &bc->live)
 405                 list_move_tail(&bc->live, &t->list);
 406 out:
 407         mutex_unlock(&bc->lock);
 408 out_nounlock:
 409         ret = freed;
 410         memalloc_nofs_restore(flags);
 411         trace_and_count(c, btree_cache_scan, sc->nr_to_scan, can_free, ret);
 412         return ret;
 413 }
 414
 415 static unsigned long bch2_btree_cache_count(struct shrinker *shrink,
 416                                             struct shrink_control *sc)
 417 {
 418         struct bch_fs *c = container_of(shrink, struct bch_fs,
 419                                         btree_cache.shrink);
 420         struct btree_cache *bc = &c->btree_cache;
 421
 422         if (bch2_btree_shrinker_disabled)
 423                 return 0;
 424
 425         return btree_cache_can_free(bc);
 426 }
 427
 428 static void bch2_btree_cache_shrinker_to_text(struct printbuf *out, struct shrinker *shrink)
 429 {
 430         struct bch_fs *c = container_of(shrink, struct bch_fs,
 431                                         btree_cache.shrink);
 432
 433         bch2_btree_cache_to_text(out, &c->btree_cache);
 434 }
 435
 436 void bch2_fs_btree_cache_exit(struct bch_fs *c)
 437 {
 438         struct btree_cache *bc = &c->btree_cache;
 439         struct btree *b;
 440         unsigned i, flags;
 441
 442         if (bc->shrink.list.next)
 443                 unregister_shrinker(&bc->shrink);
 444
 445         /* vfree() can allocate memory: */
 446         flags = memalloc_nofs_save();
 447         mutex_lock(&bc->lock);
 448
 449         if (c->verify_data)
 450                 list_move(&c->verify_data->list, &bc->live);
 451
 452         kvpfree(c->verify_ondisk, btree_bytes(c));
 453
 454         for (i = 0; i < BTREE_ID_NR; i++)
 455                 if (c->btree_roots[i].b)
 456                         list_add(&c->btree_roots[i].b->list, &bc->live);
 457
 458         list_splice(&bc->freeable, &bc->live);
 459
 460         while (!list_empty(&bc->live)) {
 461                 b = list_first_entry(&bc->live, struct btree, list);
 462
 463                 BUG_ON(btree_node_read_in_flight(b) ||
 464                        btree_node_write_in_flight(b));
 465
 466                 if (btree_node_dirty(b))
 467                         bch2_btree_complete_write(c, b, btree_current_write(b));
 468                 clear_btree_node_dirty_acct(c, b);
 469
 470                 btree_node_data_free(c, b);
 471         }
 472
 473         BUG_ON(atomic_read(&c->btree_cache.dirty));
 474
 475         list_splice(&bc->freed_pcpu, &bc->freed_nonpcpu);
 476
 477         while (!list_empty(&bc->freed_nonpcpu)) {
 478                 b = list_first_entry(&bc->freed_nonpcpu, struct btree, list);
 479                 list_del(&b->list);
 480                 six_lock_pcpu_free(&b->c.lock);
 481                 kfree(b);
 482         }
 483
 484         mutex_unlock(&bc->lock);
 485         memalloc_nofs_restore(flags);
 486
 487         if (bc->table_init_done)
 488                 rhashtable_destroy(&bc->table);
 489 }
 490
 491 int bch2_fs_btree_cache_init(struct bch_fs *c)
 492 {
 493         struct btree_cache *bc = &c->btree_cache;
 494         unsigned i;
 495         int ret = 0;
 496
 497         pr_verbose_init(c->opts, "");
 498
 499         ret = rhashtable_init(&bc->table, &bch_btree_cache_params);
 500         if (ret)
 501                 goto out;
 502
 503         bc->table_init_done = true;
 504
 505         bch2_recalc_btree_reserve(c);
 506
 507         for (i = 0; i < bc->reserve; i++)
 508                 if (!__bch2_btree_node_mem_alloc(c)) {
 509                         ret = -ENOMEM;
 510                         goto out;
 511                 }
 512
 513         list_splice_init(&bc->live, &bc->freeable);
 514
 515         mutex_init(&c->verify_lock);
 516
 517         bc->shrink.count_objects        = bch2_btree_cache_count;
 518         bc->shrink.scan_objects         = bch2_btree_cache_scan;
 519         bc->shrink.to_text              = bch2_btree_cache_shrinker_to_text;
 520         bc->shrink.seeks                = 4;
 521         ret = register_shrinker(&bc->shrink, "%s/btree_cache", c->name);
 522 out:
 523         pr_verbose_init(c->opts, "ret %i", ret);
 524         return ret;
 525 }
 526
 527 void bch2_fs_btree_cache_init_early(struct btree_cache *bc)
 528 {
 529         mutex_init(&bc->lock);
 530         INIT_LIST_HEAD(&bc->live);
 531         INIT_LIST_HEAD(&bc->freeable);
 532         INIT_LIST_HEAD(&bc->freed_pcpu);
 533         INIT_LIST_HEAD(&bc->freed_nonpcpu);
 534 }
 535
 536 /*
 537  * We can only have one thread cannibalizing other cached btree nodes at a time,
 538  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 539  * cannibalize_bucket() will take. This means every time we unlock the root of
 540  * the btree, we need to release this lock if we have it held.
 541  */
 542 void bch2_btree_cache_cannibalize_unlock(struct bch_fs *c)
 543 {
 544         struct btree_cache *bc = &c->btree_cache;
 545
 546         if (bc->alloc_lock == current) {
 547                 trace_and_count(c, btree_cache_cannibalize_unlock, c);
 548                 bc->alloc_lock = NULL;
 549                 closure_wake_up(&bc->alloc_wait);
 550         }
 551 }
 552
 553 int bch2_btree_cache_cannibalize_lock(struct bch_fs *c, struct closure *cl)
 554 {
 555         struct btree_cache *bc = &c->btree_cache;
 556         struct task_struct *old;
 557
 558         old = cmpxchg(&bc->alloc_lock, NULL, current);
 559         if (old == NULL || old == current)
 560                 goto success;
 561
 562         if (!cl) {
 563                 trace_and_count(c, btree_cache_cannibalize_lock_fail, c);
 564                 return -ENOMEM;
 565         }
 566
 567         closure_wait(&bc->alloc_wait, cl);
 568
 569         /* Try again, after adding ourselves to waitlist */
 570         old = cmpxchg(&bc->alloc_lock, NULL, current);
 571         if (old == NULL || old == current) {
 572                 /* We raced */
 573                 closure_wake_up(&bc->alloc_wait);
 574                 goto success;
 575         }
 576
 577         trace_and_count(c, btree_cache_cannibalize_lock_fail, c);
 578         return -EAGAIN;
 579
 580 success:
 581         trace_and_count(c, btree_cache_cannibalize_lock, c);
 582         return 0;
 583 }
 584
 585 static struct btree *btree_node_cannibalize(struct bch_fs *c)
 586 {
 587         struct btree_cache *bc = &c->btree_cache;
 588         struct btree *b;
 589
 590         list_for_each_entry_reverse(b, &bc->live, list)
 591                 if (!btree_node_reclaim(c, b, false))
 592                         return b;
 593
 594         while (1) {
 595                 list_for_each_entry_reverse(b, &bc->live, list)
 596                         if (!btree_node_write_and_reclaim(c, b))
 597                                 return b;
 598
 599                 /*
 600                  * Rare case: all nodes were intent-locked.
 601                  * Just busy-wait.
 602                  */
 603                 WARN_ONCE(1, "btree cache cannibalize failed\n");
 604                 cond_resched();
 605         }
 606 }
 607
 608 struct btree *bch2_btree_node_mem_alloc(struct bch_fs *c, bool pcpu_read_locks)
 609 {
 610         struct btree_cache *bc = &c->btree_cache;
 611         struct list_head *freed = pcpu_read_locks
 612                 ? &bc->freed_pcpu
 613                 : &bc->freed_nonpcpu;
 614         struct btree *b, *b2;
 615         u64 start_time = local_clock();
 616         unsigned flags;
 617
 618         flags = memalloc_nofs_save();
 619         mutex_lock(&bc->lock);
 620
 621         /*
 622          * We never free struct btree itself, just the memory that holds the on
 623          * disk node. Check the freed list before allocating a new one:
 624          */
 625         list_for_each_entry(b, freed, list)
 626                 if (!btree_node_reclaim(c, b, false)) {
 627                         list_del_init(&b->list);
 628                         goto got_node;
 629                 }
 630
 631         b = __btree_node_mem_alloc(c, __GFP_NOWARN);
 632         if (!b) {
 633                 mutex_unlock(&bc->lock);
 634                 b = __btree_node_mem_alloc(c, GFP_KERNEL);
 635                 if (!b)
 636                         goto err;
 637                 mutex_lock(&bc->lock);
 638         }
 639
 640         if (pcpu_read_locks)
 641                 six_lock_pcpu_alloc(&b->c.lock);
 642
 643         BUG_ON(!six_trylock_intent(&b->c.lock));
 644         BUG_ON(!six_trylock_write(&b->c.lock));
 645 got_node:
 646
 647         /*
 648          * btree_free() doesn't free memory; it sticks the node on the end of
 649          * the list. Check if there's any freed nodes there:
 650          */
 651         list_for_each_entry(b2, &bc->freeable, list)
 652                 if (!btree_node_reclaim(c, b2, false)) {
 653                         swap(b->data, b2->data);
 654                         swap(b->aux_data, b2->aux_data);
 655                         btree_node_to_freedlist(bc, b2);
 656                         six_unlock_write(&b2->c.lock);
 657                         six_unlock_intent(&b2->c.lock);
 658                         goto got_mem;
 659                 }
 660
 661         mutex_unlock(&bc->lock);
 662
 663         if (btree_node_data_alloc(c, b, __GFP_NOWARN|GFP_KERNEL))
 664                 goto err;
 665
 666         mutex_lock(&bc->lock);
 667         bc->used++;
 668 got_mem:
 669         mutex_unlock(&bc->lock);
 670
 671         BUG_ON(btree_node_hashed(b));
 672         BUG_ON(btree_node_dirty(b));
 673         BUG_ON(btree_node_write_in_flight(b));
 674 out:
 675         b->flags                = 0;
 676         b->written              = 0;
 677         b->nsets                = 0;
 678         b->sib_u64s[0]          = 0;
 679         b->sib_u64s[1]          = 0;
 680         b->whiteout_u64s        = 0;
 681         bch2_btree_keys_init(b);
 682         set_btree_node_accessed(b);
 683
 684         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],
 685                                start_time);
 686
 687         memalloc_nofs_restore(flags);
 688         return b;
 689 err:
 690         mutex_lock(&bc->lock);
 691
 692         /* Try to cannibalize another cached btree node: */
 693         if (bc->alloc_lock == current) {
 694                 b2 = btree_node_cannibalize(c);
 695                 bch2_btree_node_hash_remove(bc, b2);
 696
 697                 if (b) {
 698                         swap(b->data, b2->data);
 699                         swap(b->aux_data, b2->aux_data);
 700                         btree_node_to_freedlist(bc, b2);
 701                         six_unlock_write(&b2->c.lock);
 702                         six_unlock_intent(&b2->c.lock);
 703                 } else {
 704                         b = b2;
 705                         list_del_init(&b->list);
 706                 }
 707
 708                 mutex_unlock(&bc->lock);
 709
 710                 trace_and_count(c, btree_cache_cannibalize, c);
 711                 goto out;
 712         }
 713
 714         mutex_unlock(&bc->lock);
 715         memalloc_nofs_restore(flags);
 716         return ERR_PTR(-ENOMEM);
 717 }
 718
 719 /* Slowpath, don't want it inlined into btree_iter_traverse() */
 720 static noinline struct btree *bch2_btree_node_fill(struct bch_fs *c,
 721                                 struct btree_trans *trans,
 722                                 struct btree_path *path,
 723                                 const struct bkey_i *k,
 724                                 enum btree_id btree_id,
 725                                 unsigned level,
 726                                 enum six_lock_type lock_type,
 727                                 bool sync)
 728 {
 729         struct btree_cache *bc = &c->btree_cache;
 730         struct btree *b;
 731         u32 seq;
 732
 733         BUG_ON(level + 1 >= BTREE_MAX_DEPTH);
 734         /*
 735          * Parent node must be locked, else we could read in a btree node that's
 736          * been freed:
 737          */
 738         if (trans && !bch2_btree_node_relock(trans, path, level + 1)) {
 739                 trace_and_count(c, trans_restart_relock_parent_for_fill, trans, _THIS_IP_, path);
 740                 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_fill_relock));
 741         }
 742
 743         b = bch2_btree_node_mem_alloc(c, level != 0);
 744
 745         if (trans && b == ERR_PTR(-ENOMEM)) {
 746                 trans->memory_allocation_failure = true;
 747                 trace_and_count(c, trans_restart_memory_allocation_failure, trans, _THIS_IP_, path);
 748                 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_fill_mem_alloc_fail));
 749         }
 750
 751         if (IS_ERR(b))
 752                 return b;
 753
 754         bkey_copy(&b->key, k);
 755         if (bch2_btree_node_hash_insert(bc, b, level, btree_id)) {
 756                 /* raced with another fill: */
 757
 758                 /* mark as unhashed... */
 759                 b->hash_val = 0;
 760
 761                 mutex_lock(&bc->lock);
 762                 list_add(&b->list, &bc->freeable);
 763                 mutex_unlock(&bc->lock);
 764
 765                 six_unlock_write(&b->c.lock);
 766                 six_unlock_intent(&b->c.lock);
 767                 return NULL;
 768         }
 769
 770         set_btree_node_read_in_flight(b);
 771
 772         six_unlock_write(&b->c.lock);
 773         seq = b->c.lock.state.seq;
 774         six_unlock_intent(&b->c.lock);
 775
 776         /* Unlock before doing IO: */
 777         if (trans && sync)
 778                 bch2_trans_unlock(trans);
 779
 780         bch2_btree_node_read(c, b, sync);
 781
 782         if (!sync)
 783                 return NULL;
 784
 785         if (trans) {
 786                 int ret = bch2_trans_relock(trans) ?:
 787                         bch2_btree_path_relock_intent(trans, path);
 788                 if (ret) {
 789                         BUG_ON(!trans->restarted);
 790                         return ERR_PTR(ret);
 791                 }
 792         }
 793
 794         if (!six_relock_type(&b->c.lock, lock_type, seq)) {
 795                 if (trans)
 796                         trace_and_count(c, trans_restart_relock_after_fill, trans, _THIS_IP_, path);
 797                 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_relock_after_fill));
 798         }
 799
 800         return b;
 801 }
 802
 803 static noinline void btree_bad_header(struct bch_fs *c, struct btree *b)
 804 {
 805         struct printbuf buf = PRINTBUF;
 806
 807         if (!test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags))
 808                 return;
 809
 810         prt_printf(&buf,
 811                "btree node header doesn't match ptr\n"
 812                "btree %s level %u\n"
 813                "ptr: ",
 814                bch2_btree_ids[b->c.btree_id], b->c.level);
 815         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
 816
 817         prt_printf(&buf, "\nheader: btree %s level %llu\n"
 818                "min ",
 819                bch2_btree_ids[BTREE_NODE_ID(b->data)],
 820                BTREE_NODE_LEVEL(b->data));
 821         bch2_bpos_to_text(&buf, b->data->min_key);
 822
 823         prt_printf(&buf, "\nmax ");
 824         bch2_bpos_to_text(&buf, b->data->max_key);
 825
 826         bch2_fs_inconsistent(c, "%s", buf.buf);
 827         printbuf_exit(&buf);
 828 }
 829
 830 static inline void btree_check_header(struct bch_fs *c, struct btree *b)
 831 {
 832         if (b->c.btree_id != BTREE_NODE_ID(b->data) ||
 833             b->c.level != BTREE_NODE_LEVEL(b->data) ||
 834             bpos_cmp(b->data->max_key, b->key.k.p) ||
 835             (b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 836              bpos_cmp(b->data->min_key,
 837                       bkey_i_to_btree_ptr_v2(&b->key)->v.min_key)))
 838                 btree_bad_header(c, b);
 839 }
 840
 841 /**
 842  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 843  * in from disk if necessary.
 844  *
 845  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
 846  *
 847  * The btree node will have either a read or a write lock held, depending on
 848  * the @write parameter.
 849  */
 850 struct btree *bch2_btree_node_get(struct btree_trans *trans, struct btree_path *path,
 851                                   const struct bkey_i *k, unsigned level,
 852                                   enum six_lock_type lock_type,
 853                                   unsigned long trace_ip)
 854 {
 855         struct bch_fs *c = trans->c;
 856         struct btree_cache *bc = &c->btree_cache;
 857         struct btree *b;
 858         struct bset_tree *t;
 859         int ret;
 860
 861         EBUG_ON(level >= BTREE_MAX_DEPTH);
 862
 863         b = btree_node_mem_ptr(k);
 864
 865         /*
 866          * Check b->hash_val _before_ calling btree_node_lock() - this might not
 867          * be the node we want anymore, and trying to lock the wrong node could
 868          * cause an unneccessary transaction restart:
 869          */
 870         if (likely(c->opts.btree_node_mem_ptr_optimization &&
 871                    b &&
 872                    b->hash_val == btree_ptr_hash_val(k)))
 873                 goto lock_node;
 874 retry:
 875         b = btree_cache_find(bc, k);
 876         if (unlikely(!b)) {
 877                 /*
 878                  * We must have the parent locked to call bch2_btree_node_fill(),
 879                  * else we could read in a btree node from disk that's been
 880                  * freed:
 881                  */
 882                 b = bch2_btree_node_fill(c, trans, path, k, path->btree_id,
 883                                          level, lock_type, true);
 884
 885                 /* We raced and found the btree node in the cache */
 886                 if (!b)
 887                         goto retry;
 888
 889                 if (IS_ERR(b))
 890                         return b;
 891         } else {
 892 lock_node:
 893                 /*
 894                  * There's a potential deadlock with splits and insertions into
 895                  * interior nodes we have to avoid:
 896                  *
 897                  * The other thread might be holding an intent lock on the node
 898                  * we want, and they want to update its parent node so they're
 899                  * going to upgrade their intent lock on the parent node to a
 900                  * write lock.
 901                  *
 902                  * But if we're holding a read lock on the parent, and we're
 903                  * trying to get the intent lock they're holding, we deadlock.
 904                  *
 905                  * So to avoid this we drop the read locks on parent nodes when
 906                  * we're starting to take intent locks - and handle the race.
 907                  *
 908                  * The race is that they might be about to free the node we
 909                  * want, and dropping our read lock on the parent node lets them
 910                  * update the parent marking the node we want as freed, and then
 911                  * free it:
 912                  *
 913                  * To guard against this, btree nodes are evicted from the cache
 914                  * when they're freed - and b->hash_val is zeroed out, which we
 915                  * check for after we lock the node.
 916                  *
 917                  * Then, bch2_btree_node_relock() on the parent will fail - because
 918                  * the parent was modified, when the pointer to the node we want
 919                  * was removed - and we'll bail out:
 920                  */
 921                 if (btree_node_read_locked(path, level + 1))
 922                         btree_node_unlock(trans, path, level + 1);
 923
 924                 ret = btree_node_lock(trans, path, &b->c, level, lock_type, trace_ip);
 925                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 926                         return ERR_PTR(ret);
 927
 928                 BUG_ON(ret);
 929
 930                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
 931                              b->c.level != level ||
 932                              race_fault())) {
 933                         six_unlock_type(&b->c.lock, lock_type);
 934                         if (bch2_btree_node_relock(trans, path, level + 1))
 935                                 goto retry;
 936
 937                         trace_and_count(c, trans_restart_btree_node_reused, trans, trace_ip, path);
 938                         return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_lock_node_reused));
 939                 }
 940         }
 941
 942         if (unlikely(btree_node_read_in_flight(b))) {
 943                 u32 seq = b->c.lock.state.seq;
 944
 945                 six_unlock_type(&b->c.lock, lock_type);
 946                 bch2_trans_unlock(trans);
 947
 948                 bch2_btree_node_wait_on_read(b);
 949
 950                 /*
 951                  * should_be_locked is not set on this path yet, so we need to
 952                  * relock it specifically:
 953                  */
 954                 if (trans) {
 955                         int ret = bch2_trans_relock(trans) ?:
 956                                 bch2_btree_path_relock_intent(trans, path);
 957                         if (ret) {
 958                                 BUG_ON(!trans->restarted);
 959                                 return ERR_PTR(ret);
 960                         }
 961                 }
 962
 963                 if (!six_relock_type(&b->c.lock, lock_type, seq))
 964                         goto retry;
 965         }
 966
 967         prefetch(b->aux_data);
 968
 969         for_each_bset(b, t) {
 970                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 971
 972                 prefetch(p + L1_CACHE_BYTES * 0);
 973                 prefetch(p + L1_CACHE_BYTES * 1);
 974                 prefetch(p + L1_CACHE_BYTES * 2);
 975         }
 976
 977         /* avoid atomic set bit if it's not needed: */
 978         if (!btree_node_accessed(b))
 979                 set_btree_node_accessed(b);
 980
 981         if (unlikely(btree_node_read_error(b))) {
 982                 six_unlock_type(&b->c.lock, lock_type);
 983                 return ERR_PTR(-EIO);
 984         }
 985
 986         EBUG_ON(b->c.btree_id != path->btree_id);
 987         EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
 988         btree_check_header(c, b);
 989
 990         return b;
 991 }
 992
 993 struct btree *bch2_btree_node_get_noiter(struct btree_trans *trans,
 994                                          const struct bkey_i *k,
 995                                          enum btree_id btree_id,
 996                                          unsigned level,
 997                                          bool nofill)
 998 {
 999         struct bch_fs *c = trans->c;
1000         struct btree_cache *bc = &c->btree_cache;
1001         struct btree *b;
1002         struct bset_tree *t;
1003         int ret;
1004
1005         EBUG_ON(level >= BTREE_MAX_DEPTH);
1006
1007         if (c->opts.btree_node_mem_ptr_optimization) {
1008                 b = btree_node_mem_ptr(k);
1009                 if (b)
1010                         goto lock_node;
1011         }
1012 retry:
1013         b = btree_cache_find(bc, k);
1014         if (unlikely(!b)) {
1015                 if (nofill)
1016                         goto out;
1017
1018                 b = bch2_btree_node_fill(c, NULL, NULL, k, btree_id,
1019                                          level, SIX_LOCK_read, true);
1020
1021                 /* We raced and found the btree node in the cache */
1022                 if (!b)
1023                         goto retry;
1024
1025                 if (IS_ERR(b) &&
1026                     !bch2_btree_cache_cannibalize_lock(c, NULL))
1027                         goto retry;
1028
1029                 if (IS_ERR(b))
1030                         goto out;
1031         } else {
1032 lock_node:
1033                 ret = btree_node_lock_nopath(trans, &b->c, SIX_LOCK_read);
1034                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1035                         return ERR_PTR(ret);
1036
1037                 BUG_ON(ret);
1038
1039                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
1040                              b->c.btree_id != btree_id ||
1041                              b->c.level != level)) {
1042                         six_unlock_read(&b->c.lock);
1043                         goto retry;
1044                 }
1045         }
1046
1047         /* XXX: waiting on IO with btree locks held: */
1048         __bch2_btree_node_wait_on_read(b);
1049
1050         prefetch(b->aux_data);
1051
1052         for_each_bset(b, t) {
1053                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
1054
1055                 prefetch(p + L1_CACHE_BYTES * 0);
1056                 prefetch(p + L1_CACHE_BYTES * 1);
1057                 prefetch(p + L1_CACHE_BYTES * 2);
1058         }
1059
1060         /* avoid atomic set bit if it's not needed: */
1061         if (!btree_node_accessed(b))
1062                 set_btree_node_accessed(b);
1063
1064         if (unlikely(btree_node_read_error(b))) {
1065                 six_unlock_read(&b->c.lock);
1066                 b = ERR_PTR(-EIO);
1067                 goto out;
1068         }
1069
1070         EBUG_ON(b->c.btree_id != btree_id);
1071         EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
1072         btree_check_header(c, b);
1073 out:
1074         bch2_btree_cache_cannibalize_unlock(c);
1075         return b;
1076 }
1077
1078 int bch2_btree_node_prefetch(struct bch_fs *c,
1079                              struct btree_trans *trans,
1080                              struct btree_path *path,
1081                              const struct bkey_i *k,
1082                              enum btree_id btree_id, unsigned level)
1083 {
1084         struct btree_cache *bc = &c->btree_cache;
1085         struct btree *b;
1086
1087         BUG_ON(trans && !btree_node_locked(path, level + 1));
1088         BUG_ON(level >= BTREE_MAX_DEPTH);
1089
1090         b = btree_cache_find(bc, k);
1091         if (b)
1092                 return 0;
1093
1094         b = bch2_btree_node_fill(c, trans, path, k, btree_id,
1095                                  level, SIX_LOCK_read, false);
1096         return PTR_ERR_OR_ZERO(b);
1097 }
1098
1099 void bch2_btree_node_evict(struct btree_trans *trans, const struct bkey_i *k)
1100 {
1101         struct bch_fs *c = trans->c;
1102         struct btree_cache *bc = &c->btree_cache;
1103         struct btree *b;
1104
1105         b = btree_cache_find(bc, k);
1106         if (!b)
1107                 return;
1108 wait_on_io:
1109         /* not allowed to wait on io with btree locks held: */
1110
1111         /* XXX we're called from btree_gc which will be holding other btree
1112          * nodes locked
1113          */
1114         __bch2_btree_node_wait_on_read(b);
1115         __bch2_btree_node_wait_on_write(b);
1116
1117         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
1118         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
1119
1120         if (btree_node_dirty(b)) {
1121                 __bch2_btree_node_write(c, b, 0);
1122                 six_unlock_write(&b->c.lock);
1123                 six_unlock_intent(&b->c.lock);
1124                 goto wait_on_io;
1125         }
1126
1127         BUG_ON(btree_node_dirty(b));
1128
1129         mutex_lock(&bc->lock);
1130         btree_node_data_free(c, b);
1131         bch2_btree_node_hash_remove(bc, b);
1132         mutex_unlock(&bc->lock);
1133
1134         six_unlock_write(&b->c.lock);
1135         six_unlock_intent(&b->c.lock);
1136 }
1137
1138 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
1139                              struct btree *b)
1140 {
1141         const struct bkey_format *f = &b->format;
1142         struct bset_stats stats;
1143
1144         memset(&stats, 0, sizeof(stats));
1145
1146         bch2_btree_keys_stats(b, &stats);
1147
1148         prt_printf(out, "l %u ", b->c.level);
1149         bch2_bpos_to_text(out, b->data->min_key);
1150         prt_printf(out, " - ");
1151         bch2_bpos_to_text(out, b->data->max_key);
1152         prt_printf(out, ":\n"
1153                "    ptrs: ");
1154         bch2_val_to_text(out, c, bkey_i_to_s_c(&b->key));
1155
1156         prt_printf(out, "\n"
1157                "    format: u64s %u fields %u %u %u %u %u\n"
1158                "    unpack fn len: %u\n"
1159                "    bytes used %zu/%zu (%zu%% full)\n"
1160                "    sib u64s: %u, %u (merge threshold %u)\n"
1161                "    nr packed keys %u\n"
1162                "    nr unpacked keys %u\n"
1163                "    floats %zu\n"
1164                "    failed unpacked %zu\n",
1165                f->key_u64s,
1166                f->bits_per_field[0],
1167                f->bits_per_field[1],
1168                f->bits_per_field[2],
1169                f->bits_per_field[3],
1170                f->bits_per_field[4],
1171                b->unpack_fn_len,
1172                b->nr.live_u64s * sizeof(u64),
1173                btree_bytes(c) - sizeof(struct btree_node),
1174                b->nr.live_u64s * 100 / btree_max_u64s(c),
1175                b->sib_u64s[0],
1176                b->sib_u64s[1],
1177                c->btree_foreground_merge_threshold,
1178                b->nr.packed_keys,
1179                b->nr.unpacked_keys,
1180                stats.floats,
1181                stats.failed);
1182 }
1183
1184 void bch2_btree_cache_to_text(struct printbuf *out, struct btree_cache *bc)
1185 {
1186         prt_printf(out, "nr nodes:\t\t%u\n", bc->used);
1187         prt_printf(out, "nr dirty:\t\t%u\n", atomic_read(&bc->dirty));
1188         prt_printf(out, "cannibalize lock:\t%p\n", bc->alloc_lock);
1189
1190         prt_printf(out, "freed:\t\t\t\t%u\n", bc->freed);
1191         prt_printf(out, "not freed, dirty:\t\t%u\n", bc->not_freed_dirty);
1192         prt_printf(out, "not freed, write in flight:\t%u\n", bc->not_freed_write_in_flight);
1193         prt_printf(out, "not freed, read in flight:\t%u\n", bc->not_freed_read_in_flight);
1194         prt_printf(out, "not freed, lock intent failed:\t%u\n", bc->not_freed_lock_intent);
1195         prt_printf(out, "not freed, lock write failed:\t%u\n", bc->not_freed_lock_write);
1196         prt_printf(out, "not freed, access bit:\t\t%u\n", bc->not_freed_access_bit);
1197         prt_printf(out, "not freed, no evict failed:\t%u\n", bc->not_freed_noevict);
1198         prt_printf(out, "not freed, write blocked:\t%u\n", bc->not_freed_write_blocked);
1199         prt_printf(out, "not freed, will make reachable:\t%u\n", bc->not_freed_will_make_reachable);
1200
1201 }