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