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 "error.h"
  11
  12 #include <linux/prefetch.h>
  13 #include <linux/sched/mm.h>
  14 #include <trace/events/bcachefs.h>
  15
  16 struct lock_class_key bch2_btree_node_lock_key;
  17
  18 void bch2_recalc_btree_reserve(struct bch_fs *c)
  19 {
  20         unsigned i, reserve = 16;
  21
  22         if (!c->btree_roots[0].b)
  23                 reserve += 8;
  24
  25         for (i = 0; i < BTREE_ID_NR; i++)
  26                 if (c->btree_roots[i].b)
  27                         reserve += min_t(unsigned, 1,
  28                                          c->btree_roots[i].b->c.level) * 8;
  29
  30         c->btree_cache.reserve = reserve;
  31 }
  32
  33 static inline unsigned btree_cache_can_free(struct btree_cache *bc)
  34 {
  35         return max_t(int, 0, bc->used - bc->reserve);
  36 }
  37
  38 static void btree_node_data_free(struct bch_fs *c, struct btree *b)
  39 {
  40         struct btree_cache *bc = &c->btree_cache;
  41
  42         EBUG_ON(btree_node_write_in_flight(b));
  43
  44         kvpfree(b->data, btree_bytes(c));
  45         b->data = NULL;
  46 #ifdef __KERNEL__
  47         vfree(b->aux_data);
  48 #else
  49         munmap(b->aux_data, btree_aux_data_bytes(b));
  50 #endif
  51         b->aux_data = NULL;
  52
  53         bc->used--;
  54         list_move(&b->list, &bc->freed);
  55 }
  56
  57 static int bch2_btree_cache_cmp_fn(struct rhashtable_compare_arg *arg,
  58                                    const void *obj)
  59 {
  60         const struct btree *b = obj;
  61         const u64 *v = arg->key;
  62
  63         return b->hash_val == *v ? 0 : 1;
  64 }
  65
  66 static const struct rhashtable_params bch_btree_cache_params = {
  67         .head_offset    = offsetof(struct btree, hash),
  68         .key_offset     = offsetof(struct btree, hash_val),
  69         .key_len        = sizeof(u64),
  70         .obj_cmpfn      = bch2_btree_cache_cmp_fn,
  71 };
  72
  73 static int btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
  74 {
  75         BUG_ON(b->data || b->aux_data);
  76
  77         b->data = kvpmalloc(btree_bytes(c), gfp);
  78         if (!b->data)
  79                 return -ENOMEM;
  80 #ifdef __KERNEL__
  81         b->aux_data = vmalloc_exec(btree_aux_data_bytes(b), gfp);
  82 #else
  83         b->aux_data = mmap(NULL, btree_aux_data_bytes(b),
  84                            PROT_READ|PROT_WRITE|PROT_EXEC,
  85                            MAP_PRIVATE|MAP_ANONYMOUS, 0, 0);
  86         if (b->aux_data == MAP_FAILED)
  87                 b->aux_data = NULL;
  88 #endif
  89         if (!b->aux_data) {
  90                 kvpfree(b->data, btree_bytes(c));
  91                 b->data = NULL;
  92                 return -ENOMEM;
  93         }
  94
  95         return 0;
  96 }
  97
  98 static struct btree *__btree_node_mem_alloc(struct bch_fs *c)
  99 {
 100         struct btree *b = kzalloc(sizeof(struct btree), GFP_KERNEL);
 101         if (!b)
 102                 return NULL;
 103
 104         bkey_btree_ptr_init(&b->key);
 105         __six_lock_init(&b->c.lock, "b->c.lock", &bch2_btree_node_lock_key);
 106         INIT_LIST_HEAD(&b->list);
 107         INIT_LIST_HEAD(&b->write_blocked);
 108         b->byte_order = ilog2(btree_bytes(c));
 109         return b;
 110 }
 111
 112 struct btree *__bch2_btree_node_mem_alloc(struct bch_fs *c)
 113 {
 114         struct btree_cache *bc = &c->btree_cache;
 115         struct btree *b = __btree_node_mem_alloc(c);
 116         if (!b)
 117                 return NULL;
 118
 119         if (btree_node_data_alloc(c, b, GFP_KERNEL)) {
 120                 kfree(b);
 121                 return NULL;
 122         }
 123
 124         bc->used++;
 125         list_add(&b->list, &bc->freeable);
 126         return b;
 127 }
 128
 129 /* Btree in memory cache - hash table */
 130
 131 void bch2_btree_node_hash_remove(struct btree_cache *bc, struct btree *b)
 132 {
 133         int ret = rhashtable_remove_fast(&bc->table, &b->hash, bch_btree_cache_params);
 134         BUG_ON(ret);
 135
 136         /* Cause future lookups for this node to fail: */
 137         b->hash_val = 0;
 138
 139         six_lock_wakeup_all(&b->c.lock);
 140 }
 141
 142 int __bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b)
 143 {
 144         BUG_ON(b->hash_val);
 145         b->hash_val = btree_ptr_hash_val(&b->key);
 146
 147         return rhashtable_lookup_insert_fast(&bc->table, &b->hash,
 148                                              bch_btree_cache_params);
 149 }
 150
 151 int bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b,
 152                                 unsigned level, enum btree_id id)
 153 {
 154         int ret;
 155
 156         b->c.level      = level;
 157         b->c.btree_id   = id;
 158
 159         if (level)
 160                 six_lock_pcpu_alloc(&b->c.lock);
 161         else
 162                 six_lock_pcpu_free_rcu(&b->c.lock);
 163
 164         mutex_lock(&bc->lock);
 165         ret = __bch2_btree_node_hash_insert(bc, b);
 166         if (!ret)
 167                 list_add(&b->list, &bc->live);
 168         mutex_unlock(&bc->lock);
 169
 170         return ret;
 171 }
 172
 173 __flatten
 174 static inline struct btree *btree_cache_find(struct btree_cache *bc,
 175                                      const struct bkey_i *k)
 176 {
 177         u64 v = btree_ptr_hash_val(k);
 178
 179         return rhashtable_lookup_fast(&bc->table, &v, bch_btree_cache_params);
 180 }
 181
 182 /*
 183  * this version is for btree nodes that have already been freed (we're not
 184  * reaping a real btree node)
 185  */
 186 static int __btree_node_reclaim(struct bch_fs *c, struct btree *b, bool flush)
 187 {
 188         struct btree_cache *bc = &c->btree_cache;
 189         int ret = 0;
 190
 191         lockdep_assert_held(&bc->lock);
 192 wait_on_io:
 193         if (b->flags & ((1U << BTREE_NODE_dirty)|
 194                         (1U << BTREE_NODE_read_in_flight)|
 195                         (1U << BTREE_NODE_write_in_flight))) {
 196                 if (!flush)
 197                         return -ENOMEM;
 198
 199                 /* XXX: waiting on IO with btree cache lock held */
 200                 bch2_btree_node_wait_on_read(b);
 201                 bch2_btree_node_wait_on_write(b);
 202         }
 203
 204         if (!six_trylock_intent(&b->c.lock))
 205                 return -ENOMEM;
 206
 207         if (!six_trylock_write(&b->c.lock))
 208                 goto out_unlock_intent;
 209
 210         /* recheck under lock */
 211         if (b->flags & ((1U << BTREE_NODE_read_in_flight)|
 212                         (1U << BTREE_NODE_write_in_flight))) {
 213                 if (!flush)
 214                         goto out_unlock;
 215                 six_unlock_write(&b->c.lock);
 216                 six_unlock_intent(&b->c.lock);
 217                 goto wait_on_io;
 218         }
 219
 220         if (btree_node_noevict(b))
 221                 goto out_unlock;
 222
 223         if (!btree_node_may_write(b))
 224                 goto out_unlock;
 225
 226         if (btree_node_dirty(b)) {
 227                 if (!flush ||
 228                     test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags))
 229                         goto out_unlock;
 230                 /*
 231                  * Using the underscore version because we don't want to compact
 232                  * bsets after the write, since this node is about to be evicted
 233                  * - unless btree verify mode is enabled, since it runs out of
 234                  * the post write cleanup:
 235                  */
 236                 if (bch2_verify_btree_ondisk)
 237                         bch2_btree_node_write(c, b, SIX_LOCK_intent);
 238                 else
 239                         __bch2_btree_node_write(c, b, false);
 240
 241                 six_unlock_write(&b->c.lock);
 242                 six_unlock_intent(&b->c.lock);
 243                 goto wait_on_io;
 244         }
 245 out:
 246         if (b->hash_val && !ret)
 247                 trace_btree_node_reap(c, b);
 248         return ret;
 249 out_unlock:
 250         six_unlock_write(&b->c.lock);
 251 out_unlock_intent:
 252         six_unlock_intent(&b->c.lock);
 253         ret = -ENOMEM;
 254         goto out;
 255 }
 256
 257 static int btree_node_reclaim(struct bch_fs *c, struct btree *b)
 258 {
 259         return __btree_node_reclaim(c, b, false);
 260 }
 261
 262 static int btree_node_write_and_reclaim(struct bch_fs *c, struct btree *b)
 263 {
 264         return __btree_node_reclaim(c, b, true);
 265 }
 266
 267 static unsigned long bch2_btree_cache_scan(struct shrinker *shrink,
 268                                            struct shrink_control *sc)
 269 {
 270         struct bch_fs *c = container_of(shrink, struct bch_fs,
 271                                         btree_cache.shrink);
 272         struct btree_cache *bc = &c->btree_cache;
 273         struct btree *b, *t;
 274         unsigned long nr = sc->nr_to_scan;
 275         unsigned long can_free;
 276         unsigned long touched = 0;
 277         unsigned long freed = 0;
 278         unsigned i, flags;
 279         unsigned long ret = SHRINK_STOP;
 280
 281         if (bch2_btree_shrinker_disabled)
 282                 return SHRINK_STOP;
 283
 284         /* Return -1 if we can't do anything right now */
 285         if (sc->gfp_mask & __GFP_FS)
 286                 mutex_lock(&bc->lock);
 287         else if (!mutex_trylock(&bc->lock))
 288                 goto out_norestore;
 289
 290         flags = memalloc_nofs_save();
 291
 292         /*
 293          * It's _really_ critical that we don't free too many btree nodes - we
 294          * have to always leave ourselves a reserve. The reserve is how we
 295          * guarantee that allocating memory for a new btree node can always
 296          * succeed, so that inserting keys into the btree can always succeed and
 297          * IO can always make forward progress:
 298          */
 299         nr /= btree_pages(c);
 300         can_free = btree_cache_can_free(bc);
 301         nr = min_t(unsigned long, nr, can_free);
 302
 303         i = 0;
 304         list_for_each_entry_safe(b, t, &bc->freeable, list) {
 305                 /*
 306                  * Leave a few nodes on the freeable list, so that a btree split
 307                  * won't have to hit the system allocator:
 308                  */
 309                 if (++i <= 3)
 310                         continue;
 311
 312                 touched++;
 313
 314                 if (touched >= nr)
 315                         break;
 316
 317                 if (!btree_node_reclaim(c, b)) {
 318                         btree_node_data_free(c, b);
 319                         six_unlock_write(&b->c.lock);
 320                         six_unlock_intent(&b->c.lock);
 321                         freed++;
 322                 }
 323         }
 324 restart:
 325         list_for_each_entry_safe(b, t, &bc->live, list) {
 326                 touched++;
 327
 328                 if (touched >= nr) {
 329                         /* Save position */
 330                         if (&t->list != &bc->live)
 331                                 list_move_tail(&bc->live, &t->list);
 332                         break;
 333                 }
 334
 335                 if (!btree_node_accessed(b) &&
 336                     !btree_node_reclaim(c, b)) {
 337                         /* can't call bch2_btree_node_hash_remove under lock  */
 338                         freed++;
 339                         if (&t->list != &bc->live)
 340                                 list_move_tail(&bc->live, &t->list);
 341
 342                         btree_node_data_free(c, b);
 343                         mutex_unlock(&bc->lock);
 344
 345                         bch2_btree_node_hash_remove(bc, b);
 346                         six_unlock_write(&b->c.lock);
 347                         six_unlock_intent(&b->c.lock);
 348
 349                         if (freed >= nr)
 350                                 goto out;
 351
 352                         if (sc->gfp_mask & __GFP_FS)
 353                                 mutex_lock(&bc->lock);
 354                         else if (!mutex_trylock(&bc->lock))
 355                                 goto out;
 356                         goto restart;
 357                 } else
 358                         clear_btree_node_accessed(b);
 359         }
 360
 361         mutex_unlock(&bc->lock);
 362 out:
 363         ret = (unsigned long) freed * btree_pages(c);
 364         memalloc_nofs_restore(flags);
 365 out_norestore:
 366         trace_btree_cache_scan(sc->nr_to_scan,
 367                                sc->nr_to_scan / btree_pages(c),
 368                                btree_cache_can_free(bc),
 369                                ret);
 370         return ret;
 371 }
 372
 373 static unsigned long bch2_btree_cache_count(struct shrinker *shrink,
 374                                             struct shrink_control *sc)
 375 {
 376         struct bch_fs *c = container_of(shrink, struct bch_fs,
 377                                         btree_cache.shrink);
 378         struct btree_cache *bc = &c->btree_cache;
 379
 380         if (bch2_btree_shrinker_disabled)
 381                 return 0;
 382
 383         return btree_cache_can_free(bc) * btree_pages(c);
 384 }
 385
 386 void bch2_fs_btree_cache_exit(struct bch_fs *c)
 387 {
 388         struct btree_cache *bc = &c->btree_cache;
 389         struct btree *b;
 390         unsigned i, flags;
 391
 392         if (bc->shrink.list.next)
 393                 unregister_shrinker(&bc->shrink);
 394
 395         /* vfree() can allocate memory: */
 396         flags = memalloc_nofs_save();
 397         mutex_lock(&bc->lock);
 398
 399         if (c->verify_data)
 400                 list_move(&c->verify_data->list, &bc->live);
 401
 402         kvpfree(c->verify_ondisk, btree_bytes(c));
 403
 404         for (i = 0; i < BTREE_ID_NR; i++)
 405                 if (c->btree_roots[i].b)
 406                         list_add(&c->btree_roots[i].b->list, &bc->live);
 407
 408         list_splice(&bc->freeable, &bc->live);
 409
 410         while (!list_empty(&bc->live)) {
 411                 b = list_first_entry(&bc->live, struct btree, list);
 412
 413                 BUG_ON(btree_node_read_in_flight(b) ||
 414                        btree_node_write_in_flight(b));
 415
 416                 if (btree_node_dirty(b))
 417                         bch2_btree_complete_write(c, b, btree_current_write(b));
 418                 clear_btree_node_dirty(c, b);
 419
 420                 btree_node_data_free(c, b);
 421         }
 422
 423         BUG_ON(atomic_read(&c->btree_cache.dirty));
 424
 425         while (!list_empty(&bc->freed)) {
 426                 b = list_first_entry(&bc->freed, struct btree, list);
 427                 list_del(&b->list);
 428                 six_lock_pcpu_free(&b->c.lock);
 429                 kfree(b);
 430         }
 431
 432         mutex_unlock(&bc->lock);
 433         memalloc_nofs_restore(flags);
 434
 435         if (bc->table_init_done)
 436                 rhashtable_destroy(&bc->table);
 437 }
 438
 439 int bch2_fs_btree_cache_init(struct bch_fs *c)
 440 {
 441         struct btree_cache *bc = &c->btree_cache;
 442         unsigned i;
 443         int ret = 0;
 444
 445         pr_verbose_init(c->opts, "");
 446
 447         ret = rhashtable_init(&bc->table, &bch_btree_cache_params);
 448         if (ret)
 449                 goto out;
 450
 451         bc->table_init_done = true;
 452
 453         bch2_recalc_btree_reserve(c);
 454
 455         for (i = 0; i < bc->reserve; i++)
 456                 if (!__bch2_btree_node_mem_alloc(c)) {
 457                         ret = -ENOMEM;
 458                         goto out;
 459                 }
 460
 461         list_splice_init(&bc->live, &bc->freeable);
 462
 463         mutex_init(&c->verify_lock);
 464
 465         bc->shrink.count_objects        = bch2_btree_cache_count;
 466         bc->shrink.scan_objects         = bch2_btree_cache_scan;
 467         bc->shrink.seeks                = 4;
 468         bc->shrink.batch                = btree_pages(c) * 2;
 469         ret = register_shrinker(&bc->shrink);
 470 out:
 471         pr_verbose_init(c->opts, "ret %i", ret);
 472         return ret;
 473 }
 474
 475 void bch2_fs_btree_cache_init_early(struct btree_cache *bc)
 476 {
 477         mutex_init(&bc->lock);
 478         INIT_LIST_HEAD(&bc->live);
 479         INIT_LIST_HEAD(&bc->freeable);
 480         INIT_LIST_HEAD(&bc->freed);
 481 }
 482
 483 /*
 484  * We can only have one thread cannibalizing other cached btree nodes at a time,
 485  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 486  * cannibalize_bucket() will take. This means every time we unlock the root of
 487  * the btree, we need to release this lock if we have it held.
 488  */
 489 void bch2_btree_cache_cannibalize_unlock(struct bch_fs *c)
 490 {
 491         struct btree_cache *bc = &c->btree_cache;
 492
 493         if (bc->alloc_lock == current) {
 494                 trace_btree_node_cannibalize_unlock(c);
 495                 bc->alloc_lock = NULL;
 496                 closure_wake_up(&bc->alloc_wait);
 497         }
 498 }
 499
 500 int bch2_btree_cache_cannibalize_lock(struct bch_fs *c, struct closure *cl)
 501 {
 502         struct btree_cache *bc = &c->btree_cache;
 503         struct task_struct *old;
 504
 505         old = cmpxchg(&bc->alloc_lock, NULL, current);
 506         if (old == NULL || old == current)
 507                 goto success;
 508
 509         if (!cl) {
 510                 trace_btree_node_cannibalize_lock_fail(c);
 511                 return -ENOMEM;
 512         }
 513
 514         closure_wait(&bc->alloc_wait, cl);
 515
 516         /* Try again, after adding ourselves to waitlist */
 517         old = cmpxchg(&bc->alloc_lock, NULL, current);
 518         if (old == NULL || old == current) {
 519                 /* We raced */
 520                 closure_wake_up(&bc->alloc_wait);
 521                 goto success;
 522         }
 523
 524         trace_btree_node_cannibalize_lock_fail(c);
 525         return -EAGAIN;
 526
 527 success:
 528         trace_btree_node_cannibalize_lock(c);
 529         return 0;
 530 }
 531
 532 static struct btree *btree_node_cannibalize(struct bch_fs *c)
 533 {
 534         struct btree_cache *bc = &c->btree_cache;
 535         struct btree *b;
 536
 537         list_for_each_entry_reverse(b, &bc->live, list)
 538                 if (!btree_node_reclaim(c, b))
 539                         return b;
 540
 541         while (1) {
 542                 list_for_each_entry_reverse(b, &bc->live, list)
 543                         if (!btree_node_write_and_reclaim(c, b))
 544                                 return b;
 545
 546                 /*
 547                  * Rare case: all nodes were intent-locked.
 548                  * Just busy-wait.
 549                  */
 550                 WARN_ONCE(1, "btree cache cannibalize failed\n");
 551                 cond_resched();
 552         }
 553 }
 554
 555 struct btree *bch2_btree_node_mem_alloc(struct bch_fs *c)
 556 {
 557         struct btree_cache *bc = &c->btree_cache;
 558         struct btree *b;
 559         u64 start_time = local_clock();
 560         unsigned flags;
 561
 562         flags = memalloc_nofs_save();
 563         mutex_lock(&bc->lock);
 564
 565         /*
 566          * btree_free() doesn't free memory; it sticks the node on the end of
 567          * the list. Check if there's any freed nodes there:
 568          */
 569         list_for_each_entry(b, &bc->freeable, list)
 570                 if (!btree_node_reclaim(c, b))
 571                         goto got_node;
 572
 573         /*
 574          * We never free struct btree itself, just the memory that holds the on
 575          * disk node. Check the freed list before allocating a new one:
 576          */
 577         list_for_each_entry(b, &bc->freed, list)
 578                 if (!btree_node_reclaim(c, b))
 579                         goto got_node;
 580
 581         b = NULL;
 582 got_node:
 583         if (b)
 584                 list_del_init(&b->list);
 585         mutex_unlock(&bc->lock);
 586
 587         if (!b) {
 588                 b = __btree_node_mem_alloc(c);
 589                 if (!b)
 590                         goto err;
 591
 592                 BUG_ON(!six_trylock_intent(&b->c.lock));
 593                 BUG_ON(!six_trylock_write(&b->c.lock));
 594         }
 595
 596         if (!b->data) {
 597                 if (btree_node_data_alloc(c, b, __GFP_NOWARN|GFP_KERNEL))
 598                         goto err;
 599
 600                 mutex_lock(&bc->lock);
 601                 bc->used++;
 602                 mutex_unlock(&bc->lock);
 603         }
 604
 605         BUG_ON(btree_node_hashed(b));
 606         BUG_ON(btree_node_dirty(b));
 607         BUG_ON(btree_node_write_in_flight(b));
 608 out:
 609         b->flags                = 0;
 610         b->written              = 0;
 611         b->nsets                = 0;
 612         b->sib_u64s[0]          = 0;
 613         b->sib_u64s[1]          = 0;
 614         b->whiteout_u64s        = 0;
 615         bch2_btree_keys_init(b);
 616         set_btree_node_accessed(b);
 617
 618         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],
 619                                start_time);
 620
 621         memalloc_nofs_restore(flags);
 622         return b;
 623 err:
 624         mutex_lock(&bc->lock);
 625
 626         if (b) {
 627                 list_add(&b->list, &bc->freed);
 628                 six_unlock_write(&b->c.lock);
 629                 six_unlock_intent(&b->c.lock);
 630         }
 631
 632         /* Try to cannibalize another cached btree node: */
 633         if (bc->alloc_lock == current) {
 634                 b = btree_node_cannibalize(c);
 635                 list_del_init(&b->list);
 636                 mutex_unlock(&bc->lock);
 637
 638                 bch2_btree_node_hash_remove(bc, b);
 639
 640                 trace_btree_node_cannibalize(c);
 641                 goto out;
 642         }
 643
 644         mutex_unlock(&bc->lock);
 645         memalloc_nofs_restore(flags);
 646         return ERR_PTR(-ENOMEM);
 647 }
 648
 649 /* Slowpath, don't want it inlined into btree_iter_traverse() */
 650 static noinline struct btree *bch2_btree_node_fill(struct bch_fs *c,
 651                                 struct btree_trans *trans,
 652                                 struct btree_path *path,
 653                                 const struct bkey_i *k,
 654                                 enum btree_id btree_id,
 655                                 unsigned level,
 656                                 enum six_lock_type lock_type,
 657                                 bool sync)
 658 {
 659         struct btree_cache *bc = &c->btree_cache;
 660         struct btree *b;
 661         u32 seq;
 662
 663         BUG_ON(level + 1 >= BTREE_MAX_DEPTH);
 664         /*
 665          * Parent node must be locked, else we could read in a btree node that's
 666          * been freed:
 667          */
 668         if (trans && !bch2_btree_node_relock(trans, path, level + 1)) {
 669                 trace_trans_restart_relock_parent_for_fill(trans->fn,
 670                                         _THIS_IP_, btree_id, &path->pos);
 671                 btree_trans_restart(trans);
 672                 return ERR_PTR(-EINTR);
 673         }
 674
 675         b = bch2_btree_node_mem_alloc(c);
 676
 677         if (trans && b == ERR_PTR(-ENOMEM)) {
 678                 trans->memory_allocation_failure = true;
 679                 trace_trans_restart_memory_allocation_failure(trans->fn,
 680                                 _THIS_IP_, btree_id, &path->pos);
 681                 btree_trans_restart(trans);
 682                 return ERR_PTR(-EINTR);
 683         }
 684
 685         if (IS_ERR(b))
 686                 return b;
 687
 688         bkey_copy(&b->key, k);
 689         if (bch2_btree_node_hash_insert(bc, b, level, btree_id)) {
 690                 /* raced with another fill: */
 691
 692                 /* mark as unhashed... */
 693                 b->hash_val = 0;
 694
 695                 mutex_lock(&bc->lock);
 696                 list_add(&b->list, &bc->freeable);
 697                 mutex_unlock(&bc->lock);
 698
 699                 six_unlock_write(&b->c.lock);
 700                 six_unlock_intent(&b->c.lock);
 701                 return NULL;
 702         }
 703
 704         set_btree_node_read_in_flight(b);
 705
 706         six_unlock_write(&b->c.lock);
 707         seq = b->c.lock.state.seq;
 708         six_unlock_intent(&b->c.lock);
 709
 710         /* Unlock before doing IO: */
 711         if (trans && sync)
 712                 bch2_trans_unlock(trans);
 713
 714         bch2_btree_node_read(c, b, sync);
 715
 716         if (!sync)
 717                 return NULL;
 718
 719         if (trans &&
 720             (!bch2_trans_relock(trans) ||
 721              !bch2_btree_path_relock_intent(trans, path))) {
 722                 BUG_ON(!trans->restarted);
 723                 return ERR_PTR(-EINTR);
 724         }
 725
 726         if (!six_relock_type(&b->c.lock, lock_type, seq)) {
 727                 trace_trans_restart_relock_after_fill(trans->fn, _THIS_IP_,
 728                                            btree_id, &path->pos);
 729                 btree_trans_restart(trans);
 730                 return ERR_PTR(-EINTR);
 731         }
 732
 733         return b;
 734 }
 735
 736 static int lock_node_check_fn(struct six_lock *lock, void *p)
 737 {
 738         struct btree *b = container_of(lock, struct btree, c.lock);
 739         const struct bkey_i *k = p;
 740
 741         return b->hash_val == btree_ptr_hash_val(k) ? 0 : -1;
 742 }
 743
 744 static noinline void btree_bad_header(struct bch_fs *c, struct btree *b)
 745 {
 746         char buf1[200], buf2[100], buf3[100];
 747
 748         if (!test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags))
 749                 return;
 750
 751         bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(&b->key));
 752         bch2_bpos_to_text(&PBUF(buf2), b->data->min_key);
 753         bch2_bpos_to_text(&PBUF(buf3), b->data->max_key);
 754
 755         bch2_fs_inconsistent(c, "btree node header doesn't match ptr\n"
 756                              "btree %s level %u\n"
 757                              "ptr: %s\n"
 758                              "header: btree %s level %llu\n"
 759                              "min %s max %s\n",
 760                              bch2_btree_ids[b->c.btree_id], b->c.level,
 761                              buf1,
 762                              bch2_btree_ids[BTREE_NODE_ID(b->data)],
 763                              BTREE_NODE_LEVEL(b->data),
 764                              buf2, buf3);
 765 }
 766
 767 static inline void btree_check_header(struct bch_fs *c, struct btree *b)
 768 {
 769         if (b->c.btree_id != BTREE_NODE_ID(b->data) ||
 770             b->c.level != BTREE_NODE_LEVEL(b->data) ||
 771             bpos_cmp(b->data->max_key, b->key.k.p) ||
 772             (b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 773              bpos_cmp(b->data->min_key,
 774                       bkey_i_to_btree_ptr_v2(&b->key)->v.min_key)))
 775                 btree_bad_header(c, b);
 776 }
 777
 778 /**
 779  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 780  * in from disk if necessary.
 781  *
 782  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
 783  *
 784  * The btree node will have either a read or a write lock held, depending on
 785  * the @write parameter.
 786  */
 787 struct btree *bch2_btree_node_get(struct btree_trans *trans, struct btree_path *path,
 788                                   const struct bkey_i *k, unsigned level,
 789                                   enum six_lock_type lock_type,
 790                                   unsigned long trace_ip)
 791 {
 792         struct bch_fs *c = trans->c;
 793         struct btree_cache *bc = &c->btree_cache;
 794         struct btree *b;
 795         struct bset_tree *t;
 796
 797         EBUG_ON(level >= BTREE_MAX_DEPTH);
 798
 799         b = btree_node_mem_ptr(k);
 800
 801         /*
 802          * Check b->hash_val _before_ calling btree_node_lock() - this might not
 803          * be the node we want anymore, and trying to lock the wrong node could
 804          * cause an unneccessary transaction restart:
 805          */
 806         if (likely(c->opts.btree_node_mem_ptr_optimization &&
 807                    b &&
 808                    b->hash_val == btree_ptr_hash_val(k)))
 809                         goto lock_node;
 810 retry:
 811         b = btree_cache_find(bc, k);
 812         if (unlikely(!b)) {
 813                 /*
 814                  * We must have the parent locked to call bch2_btree_node_fill(),
 815                  * else we could read in a btree node from disk that's been
 816                  * freed:
 817                  */
 818                 b = bch2_btree_node_fill(c, trans, path, k, path->btree_id,
 819                                          level, lock_type, true);
 820
 821                 /* We raced and found the btree node in the cache */
 822                 if (!b)
 823                         goto retry;
 824
 825                 if (IS_ERR(b))
 826                         return b;
 827         } else {
 828 lock_node:
 829                 /*
 830                  * There's a potential deadlock with splits and insertions into
 831                  * interior nodes we have to avoid:
 832                  *
 833                  * The other thread might be holding an intent lock on the node
 834                  * we want, and they want to update its parent node so they're
 835                  * going to upgrade their intent lock on the parent node to a
 836                  * write lock.
 837                  *
 838                  * But if we're holding a read lock on the parent, and we're
 839                  * trying to get the intent lock they're holding, we deadlock.
 840                  *
 841                  * So to avoid this we drop the read locks on parent nodes when
 842                  * we're starting to take intent locks - and handle the race.
 843                  *
 844                  * The race is that they might be about to free the node we
 845                  * want, and dropping our read lock on the parent node lets them
 846                  * update the parent marking the node we want as freed, and then
 847                  * free it:
 848                  *
 849                  * To guard against this, btree nodes are evicted from the cache
 850                  * when they're freed - and b->hash_val is zeroed out, which we
 851                  * check for after we lock the node.
 852                  *
 853                  * Then, bch2_btree_node_relock() on the parent will fail - because
 854                  * the parent was modified, when the pointer to the node we want
 855                  * was removed - and we'll bail out:
 856                  */
 857                 if (btree_node_read_locked(path, level + 1))
 858                         btree_node_unlock(path, level + 1);
 859
 860                 if (!btree_node_lock(trans, path, b, k->k.p, level, lock_type,
 861                                      lock_node_check_fn, (void *) k, trace_ip)) {
 862                         if (!trans->restarted)
 863                                 goto retry;
 864                         return ERR_PTR(-EINTR);
 865                 }
 866
 867                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
 868                              b->c.level != level ||
 869                              race_fault())) {
 870                         six_unlock_type(&b->c.lock, lock_type);
 871                         if (bch2_btree_node_relock(trans, path, level + 1))
 872                                 goto retry;
 873
 874                         trace_trans_restart_btree_node_reused(trans->fn,
 875                                                               trace_ip,
 876                                                               path->btree_id,
 877                                                               &path->pos);
 878                         btree_trans_restart(trans);
 879                         return ERR_PTR(-EINTR);
 880                 }
 881         }
 882
 883         if (unlikely(btree_node_read_in_flight(b))) {
 884                 u32 seq = b->c.lock.state.seq;
 885
 886                 six_unlock_type(&b->c.lock, lock_type);
 887                 bch2_trans_unlock(trans);
 888
 889                 bch2_btree_node_wait_on_read(b);
 890
 891                 /*
 892                  * should_be_locked is not set on this path yet, so we need to
 893                  * relock it specifically:
 894                  */
 895                 if (trans &&
 896                     (!bch2_trans_relock(trans) ||
 897                      !bch2_btree_path_relock_intent(trans, path))) {
 898                         BUG_ON(!trans->restarted);
 899                         return ERR_PTR(-EINTR);
 900                 }
 901
 902                 if (!six_relock_type(&b->c.lock, lock_type, seq))
 903                         goto retry;
 904         }
 905
 906         prefetch(b->aux_data);
 907
 908         for_each_bset(b, t) {
 909                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 910
 911                 prefetch(p + L1_CACHE_BYTES * 0);
 912                 prefetch(p + L1_CACHE_BYTES * 1);
 913                 prefetch(p + L1_CACHE_BYTES * 2);
 914         }
 915
 916         /* avoid atomic set bit if it's not needed: */
 917         if (!btree_node_accessed(b))
 918                 set_btree_node_accessed(b);
 919
 920         if (unlikely(btree_node_read_error(b))) {
 921                 six_unlock_type(&b->c.lock, lock_type);
 922                 return ERR_PTR(-EIO);
 923         }
 924
 925         EBUG_ON(b->c.btree_id != path->btree_id);
 926         EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
 927         btree_check_header(c, b);
 928
 929         return b;
 930 }
 931
 932 struct btree *bch2_btree_node_get_noiter(struct bch_fs *c,
 933                                          const struct bkey_i *k,
 934                                          enum btree_id btree_id,
 935                                          unsigned level,
 936                                          bool nofill)
 937 {
 938         struct btree_cache *bc = &c->btree_cache;
 939         struct btree *b;
 940         struct bset_tree *t;
 941         int ret;
 942
 943         EBUG_ON(level >= BTREE_MAX_DEPTH);
 944
 945         if (c->opts.btree_node_mem_ptr_optimization) {
 946                 b = btree_node_mem_ptr(k);
 947                 if (b)
 948                         goto lock_node;
 949         }
 950 retry:
 951         b = btree_cache_find(bc, k);
 952         if (unlikely(!b)) {
 953                 if (nofill)
 954                         goto out;
 955
 956                 b = bch2_btree_node_fill(c, NULL, NULL, k, btree_id,
 957                                          level, SIX_LOCK_read, true);
 958
 959                 /* We raced and found the btree node in the cache */
 960                 if (!b)
 961                         goto retry;
 962
 963                 if (IS_ERR(b) &&
 964                     !bch2_btree_cache_cannibalize_lock(c, NULL))
 965                         goto retry;
 966
 967                 if (IS_ERR(b))
 968                         goto out;
 969         } else {
 970 lock_node:
 971                 ret = six_lock_read(&b->c.lock, lock_node_check_fn, (void *) k);
 972                 if (ret)
 973                         goto retry;
 974
 975                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
 976                              b->c.btree_id != btree_id ||
 977                              b->c.level != level)) {
 978                         six_unlock_read(&b->c.lock);
 979                         goto retry;
 980                 }
 981         }
 982
 983         /* XXX: waiting on IO with btree locks held: */
 984         __bch2_btree_node_wait_on_read(b);
 985
 986         prefetch(b->aux_data);
 987
 988         for_each_bset(b, t) {
 989                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 990
 991                 prefetch(p + L1_CACHE_BYTES * 0);
 992                 prefetch(p + L1_CACHE_BYTES * 1);
 993                 prefetch(p + L1_CACHE_BYTES * 2);
 994         }
 995
 996         /* avoid atomic set bit if it's not needed: */
 997         if (!btree_node_accessed(b))
 998                 set_btree_node_accessed(b);
 999
1000         if (unlikely(btree_node_read_error(b))) {
1001                 six_unlock_read(&b->c.lock);
1002                 b = ERR_PTR(-EIO);
1003                 goto out;
1004         }
1005
1006         EBUG_ON(b->c.btree_id != btree_id);
1007         EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
1008         btree_check_header(c, b);
1009 out:
1010         bch2_btree_cache_cannibalize_unlock(c);
1011         return b;
1012 }
1013
1014 int bch2_btree_node_prefetch(struct bch_fs *c,
1015                              struct btree_trans *trans,
1016                              struct btree_path *path,
1017                              const struct bkey_i *k,
1018                              enum btree_id btree_id, unsigned level)
1019 {
1020         struct btree_cache *bc = &c->btree_cache;
1021         struct btree *b;
1022
1023         BUG_ON(trans && !btree_node_locked(path, level + 1));
1024         BUG_ON(level >= BTREE_MAX_DEPTH);
1025
1026         b = btree_cache_find(bc, k);
1027         if (b)
1028                 return 0;
1029
1030         b = bch2_btree_node_fill(c, trans, path, k, btree_id,
1031                                  level, SIX_LOCK_read, false);
1032         return PTR_ERR_OR_ZERO(b);
1033 }
1034
1035 void bch2_btree_node_evict(struct bch_fs *c, const struct bkey_i *k)
1036 {
1037         struct btree_cache *bc = &c->btree_cache;
1038         struct btree *b;
1039
1040         b = btree_cache_find(bc, k);
1041         if (!b)
1042                 return;
1043 wait_on_io:
1044         /* not allowed to wait on io with btree locks held: */
1045
1046         /* XXX we're called from btree_gc which will be holding other btree
1047          * nodes locked
1048          * */
1049         __bch2_btree_node_wait_on_read(b);
1050         __bch2_btree_node_wait_on_write(b);
1051
1052         six_lock_intent(&b->c.lock, NULL, NULL);
1053         six_lock_write(&b->c.lock, NULL, NULL);
1054
1055         if (btree_node_dirty(b)) {
1056                 __bch2_btree_node_write(c, b, false);
1057                 six_unlock_write(&b->c.lock);
1058                 six_unlock_intent(&b->c.lock);
1059                 goto wait_on_io;
1060         }
1061
1062         BUG_ON(btree_node_dirty(b));
1063
1064         mutex_lock(&bc->lock);
1065         btree_node_data_free(c, b);
1066         bch2_btree_node_hash_remove(bc, b);
1067         mutex_unlock(&bc->lock);
1068
1069         six_unlock_write(&b->c.lock);
1070         six_unlock_intent(&b->c.lock);
1071 }
1072
1073 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
1074                              struct btree *b)
1075 {
1076         const struct bkey_format *f = &b->format;
1077         struct bset_stats stats;
1078
1079         memset(&stats, 0, sizeof(stats));
1080
1081         bch2_btree_keys_stats(b, &stats);
1082
1083         pr_buf(out, "l %u ", b->c.level);
1084         bch2_bpos_to_text(out, b->data->min_key);
1085         pr_buf(out, " - ");
1086         bch2_bpos_to_text(out, b->data->max_key);
1087         pr_buf(out, ":\n"
1088                "    ptrs: ");
1089         bch2_val_to_text(out, c, bkey_i_to_s_c(&b->key));
1090
1091         pr_buf(out, "\n"
1092                "    format: u64s %u fields %u %u %u %u %u\n"
1093                "    unpack fn len: %u\n"
1094                "    bytes used %zu/%zu (%zu%% full)\n"
1095                "    sib u64s: %u, %u (merge threshold %u)\n"
1096                "    nr packed keys %u\n"
1097                "    nr unpacked keys %u\n"
1098                "    floats %zu\n"
1099                "    failed unpacked %zu\n",
1100                f->key_u64s,
1101                f->bits_per_field[0],
1102                f->bits_per_field[1],
1103                f->bits_per_field[2],
1104                f->bits_per_field[3],
1105                f->bits_per_field[4],
1106                b->unpack_fn_len,
1107                b->nr.live_u64s * sizeof(u64),
1108                btree_bytes(c) - sizeof(struct btree_node),
1109                b->nr.live_u64s * 100 / btree_max_u64s(c),
1110                b->sib_u64s[0],
1111                b->sib_u64s[1],
1112                c->btree_foreground_merge_threshold,
1113                b->nr.packed_keys,
1114                b->nr.unpacked_keys,
1115                stats.floats,
1116                stats.failed);
1117 }
1118
1119 void bch2_btree_cache_to_text(struct printbuf *out, struct bch_fs *c)
1120 {
1121         pr_buf(out, "nr nodes:\t\t%u\n", c->btree_cache.used);
1122         pr_buf(out, "nr dirty:\t\t%u\n", atomic_read(&c->btree_cache.dirty));
1123         pr_buf(out, "cannibalize lock:\t%p\n", c->btree_cache.alloc_lock);
1124 }