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 "btree_cache.h"
   5 #include "btree_io.h"
   6 #include "btree_iter.h"
   7 #include "btree_locking.h"
   8 #include "debug.h"
   9
  10 #include <linux/prefetch.h>
  11 #include <linux/sched/mm.h>
  12 #include <trace/events/bcachefs.h>
  13
  14 const char * const bch2_btree_ids[] = {
  15 #define x(kwd, val, name) name,
  16         BCH_BTREE_IDS()
  17 #undef x
  18         NULL
  19 };
  20
  21 void bch2_recalc_btree_reserve(struct bch_fs *c)
  22 {
  23         unsigned i, reserve = 16;
  24
  25         if (!c->btree_roots[0].b)
  26                 reserve += 8;
  27
  28         for (i = 0; i < BTREE_ID_NR; i++)
  29                 if (c->btree_roots[i].b)
  30                         reserve += min_t(unsigned, 1,
  31                                          c->btree_roots[i].b->level) * 8;
  32
  33         c->btree_cache.reserve = reserve;
  34 }
  35
  36 static inline unsigned btree_cache_can_free(struct btree_cache *bc)
  37 {
  38         return max_t(int, 0, bc->used - bc->reserve);
  39 }
  40
  41 static void __btree_node_data_free(struct bch_fs *c, struct btree *b)
  42 {
  43         EBUG_ON(btree_node_write_in_flight(b));
  44
  45         kvpfree(b->data, btree_bytes(c));
  46         b->data = NULL;
  47         bch2_btree_keys_free(b);
  48 }
  49
  50 static void btree_node_data_free(struct bch_fs *c, struct btree *b)
  51 {
  52         struct btree_cache *bc = &c->btree_cache;
  53
  54         __btree_node_data_free(c, b);
  55         bc->used--;
  56         list_move(&b->list, &bc->freed);
  57 }
  58
  59 static int bch2_btree_cache_cmp_fn(struct rhashtable_compare_arg *arg,
  60                                    const void *obj)
  61 {
  62         const struct btree *b = obj;
  63         const u64 *v = arg->key;
  64
  65         return PTR_HASH(&b->key) == *v ? 0 : 1;
  66 }
  67
  68 static const struct rhashtable_params bch_btree_cache_params = {
  69         .head_offset    = offsetof(struct btree, hash),
  70         .key_offset     = offsetof(struct btree, key.v),
  71         .key_len        = sizeof(struct bch_extent_ptr),
  72         .obj_cmpfn      = bch2_btree_cache_cmp_fn,
  73 };
  74
  75 static void btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
  76 {
  77         struct btree_cache *bc = &c->btree_cache;
  78
  79         b->data = kvpmalloc(btree_bytes(c), gfp);
  80         if (!b->data)
  81                 goto err;
  82
  83         if (bch2_btree_keys_alloc(b, btree_page_order(c), gfp))
  84                 goto err;
  85
  86         memset(&b->data->csum, 0, sizeof b->data->csum);
  87         b->data->flags = 0;
  88
  89         bc->used++;
  90         list_move(&b->list, &bc->freeable);
  91         return;
  92 err:
  93         kvpfree(b->data, btree_bytes(c));
  94         b->data = NULL;
  95         list_move(&b->list, &bc->freed);
  96 }
  97
  98 static struct btree *btree_node_mem_alloc(struct bch_fs *c, gfp_t gfp)
  99 {
 100         struct btree *b = kzalloc(sizeof(struct btree), gfp);
 101         if (!b)
 102                 return NULL;
 103
 104         bkey_btree_ptr_init(&b->key);
 105         six_lock_init(&b->lock);
 106         INIT_LIST_HEAD(&b->list);
 107         INIT_LIST_HEAD(&b->write_blocked);
 108
 109         btree_node_data_alloc(c, b, gfp);
 110         return b->data ? b : NULL;
 111 }
 112
 113 /* Btree in memory cache - hash table */
 114
 115 void bch2_btree_node_hash_remove(struct btree_cache *bc, struct btree *b)
 116 {
 117         rhashtable_remove_fast(&bc->table, &b->hash, bch_btree_cache_params);
 118
 119         /* Cause future lookups for this node to fail: */
 120         PTR_HASH(&b->key) = 0;
 121 }
 122
 123 int __bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b)
 124 {
 125         return rhashtable_lookup_insert_fast(&bc->table, &b->hash,
 126                                              bch_btree_cache_params);
 127 }
 128
 129 int bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b,
 130                                 unsigned level, enum btree_id id)
 131 {
 132         int ret;
 133
 134         b->level        = level;
 135         b->btree_id     = id;
 136
 137         mutex_lock(&bc->lock);
 138         ret = __bch2_btree_node_hash_insert(bc, b);
 139         if (!ret)
 140                 list_add(&b->list, &bc->live);
 141         mutex_unlock(&bc->lock);
 142
 143         return ret;
 144 }
 145
 146 __flatten
 147 static inline struct btree *btree_cache_find(struct btree_cache *bc,
 148                                      const struct bkey_i *k)
 149 {
 150         return rhashtable_lookup_fast(&bc->table, &PTR_HASH(k),
 151                                       bch_btree_cache_params);
 152 }
 153
 154 /*
 155  * this version is for btree nodes that have already been freed (we're not
 156  * reaping a real btree node)
 157  */
 158 static int __btree_node_reclaim(struct bch_fs *c, struct btree *b, bool flush)
 159 {
 160         struct btree_cache *bc = &c->btree_cache;
 161         int ret = 0;
 162
 163         lockdep_assert_held(&bc->lock);
 164
 165         if (!six_trylock_intent(&b->lock))
 166                 return -ENOMEM;
 167
 168         if (!six_trylock_write(&b->lock))
 169                 goto out_unlock_intent;
 170
 171         if (btree_node_noevict(b))
 172                 goto out_unlock;
 173
 174         if (!btree_node_may_write(b))
 175                 goto out_unlock;
 176
 177         if (btree_node_dirty(b) &&
 178             test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags))
 179                 goto out_unlock;
 180
 181         if (btree_node_dirty(b) ||
 182             btree_node_write_in_flight(b) ||
 183             btree_node_read_in_flight(b)) {
 184                 if (!flush)
 185                         goto out_unlock;
 186
 187                 wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
 188                                TASK_UNINTERRUPTIBLE);
 189
 190                 /*
 191                  * Using the underscore version because we don't want to compact
 192                  * bsets after the write, since this node is about to be evicted
 193                  * - unless btree verify mode is enabled, since it runs out of
 194                  * the post write cleanup:
 195                  */
 196                 if (verify_btree_ondisk(c))
 197                         bch2_btree_node_write(c, b, SIX_LOCK_intent);
 198                 else
 199                         __bch2_btree_node_write(c, b, SIX_LOCK_read);
 200
 201                 /* wait for any in flight btree write */
 202                 btree_node_wait_on_io(b);
 203         }
 204 out:
 205         if (PTR_HASH(&b->key) && !ret)
 206                 trace_btree_node_reap(c, b);
 207         return ret;
 208 out_unlock:
 209         six_unlock_write(&b->lock);
 210 out_unlock_intent:
 211         six_unlock_intent(&b->lock);
 212         ret = -ENOMEM;
 213         goto out;
 214 }
 215
 216 static int btree_node_reclaim(struct bch_fs *c, struct btree *b)
 217 {
 218         return __btree_node_reclaim(c, b, false);
 219 }
 220
 221 static int btree_node_write_and_reclaim(struct bch_fs *c, struct btree *b)
 222 {
 223         return __btree_node_reclaim(c, b, true);
 224 }
 225
 226 static unsigned long bch2_btree_cache_scan(struct shrinker *shrink,
 227                                            struct shrink_control *sc)
 228 {
 229         struct bch_fs *c = container_of(shrink, struct bch_fs,
 230                                         btree_cache.shrink);
 231         struct btree_cache *bc = &c->btree_cache;
 232         struct btree *b, *t;
 233         unsigned long nr = sc->nr_to_scan;
 234         unsigned long can_free;
 235         unsigned long touched = 0;
 236         unsigned long freed = 0;
 237         unsigned i;
 238
 239         if (btree_shrinker_disabled(c))
 240                 return SHRINK_STOP;
 241
 242         /* Return -1 if we can't do anything right now */
 243         if (sc->gfp_mask & __GFP_IO)
 244                 mutex_lock(&bc->lock);
 245         else if (!mutex_trylock(&bc->lock))
 246                 return -1;
 247
 248         /*
 249          * It's _really_ critical that we don't free too many btree nodes - we
 250          * have to always leave ourselves a reserve. The reserve is how we
 251          * guarantee that allocating memory for a new btree node can always
 252          * succeed, so that inserting keys into the btree can always succeed and
 253          * IO can always make forward progress:
 254          */
 255         nr /= btree_pages(c);
 256         can_free = btree_cache_can_free(bc);
 257         nr = min_t(unsigned long, nr, can_free);
 258
 259         i = 0;
 260         list_for_each_entry_safe(b, t, &bc->freeable, list) {
 261                 touched++;
 262
 263                 if (freed >= nr)
 264                         break;
 265
 266                 if (++i > 3 &&
 267                     !btree_node_reclaim(c, b)) {
 268                         btree_node_data_free(c, b);
 269                         six_unlock_write(&b->lock);
 270                         six_unlock_intent(&b->lock);
 271                         freed++;
 272                 }
 273         }
 274 restart:
 275         list_for_each_entry_safe(b, t, &bc->live, list) {
 276                 touched++;
 277
 278                 if (freed >= nr) {
 279                         /* Save position */
 280                         if (&t->list != &bc->live)
 281                                 list_move_tail(&bc->live, &t->list);
 282                         break;
 283                 }
 284
 285                 if (!btree_node_accessed(b) &&
 286                     !btree_node_reclaim(c, b)) {
 287                         /* can't call bch2_btree_node_hash_remove under lock  */
 288                         freed++;
 289                         if (&t->list != &bc->live)
 290                                 list_move_tail(&bc->live, &t->list);
 291
 292                         btree_node_data_free(c, b);
 293                         mutex_unlock(&bc->lock);
 294
 295                         bch2_btree_node_hash_remove(bc, b);
 296                         six_unlock_write(&b->lock);
 297                         six_unlock_intent(&b->lock);
 298
 299                         if (freed >= nr)
 300                                 goto out;
 301
 302                         if (sc->gfp_mask & __GFP_IO)
 303                                 mutex_lock(&bc->lock);
 304                         else if (!mutex_trylock(&bc->lock))
 305                                 goto out;
 306                         goto restart;
 307                 } else
 308                         clear_btree_node_accessed(b);
 309         }
 310
 311         mutex_unlock(&bc->lock);
 312 out:
 313         return (unsigned long) freed * btree_pages(c);
 314 }
 315
 316 static unsigned long bch2_btree_cache_count(struct shrinker *shrink,
 317                                             struct shrink_control *sc)
 318 {
 319         struct bch_fs *c = container_of(shrink, struct bch_fs,
 320                                         btree_cache.shrink);
 321         struct btree_cache *bc = &c->btree_cache;
 322
 323         if (btree_shrinker_disabled(c))
 324                 return 0;
 325
 326         return btree_cache_can_free(bc) * btree_pages(c);
 327 }
 328
 329 void bch2_fs_btree_cache_exit(struct bch_fs *c)
 330 {
 331         struct btree_cache *bc = &c->btree_cache;
 332         struct btree *b;
 333         unsigned i;
 334
 335         if (bc->shrink.list.next)
 336                 unregister_shrinker(&bc->shrink);
 337
 338         mutex_lock(&bc->lock);
 339
 340 #ifdef CONFIG_BCACHEFS_DEBUG
 341         if (c->verify_data)
 342                 list_move(&c->verify_data->list, &bc->live);
 343
 344         kvpfree(c->verify_ondisk, btree_bytes(c));
 345 #endif
 346
 347         for (i = 0; i < BTREE_ID_NR; i++)
 348                 if (c->btree_roots[i].b)
 349                         list_add(&c->btree_roots[i].b->list, &bc->live);
 350
 351         list_splice(&bc->freeable, &bc->live);
 352
 353         while (!list_empty(&bc->live)) {
 354                 b = list_first_entry(&bc->live, struct btree, list);
 355
 356                 BUG_ON(btree_node_read_in_flight(b) ||
 357                        btree_node_write_in_flight(b));
 358
 359                 if (btree_node_dirty(b))
 360                         bch2_btree_complete_write(c, b, btree_current_write(b));
 361                 clear_btree_node_dirty(b);
 362
 363                 btree_node_data_free(c, b);
 364         }
 365
 366         while (!list_empty(&bc->freed)) {
 367                 b = list_first_entry(&bc->freed, struct btree, list);
 368                 list_del(&b->list);
 369                 kfree(b);
 370         }
 371
 372         mutex_unlock(&bc->lock);
 373
 374         if (bc->table_init_done)
 375                 rhashtable_destroy(&bc->table);
 376 }
 377
 378 int bch2_fs_btree_cache_init(struct bch_fs *c)
 379 {
 380         struct btree_cache *bc = &c->btree_cache;
 381         unsigned i;
 382         int ret = 0;
 383
 384         pr_verbose_init(c->opts, "");
 385
 386         ret = rhashtable_init(&bc->table, &bch_btree_cache_params);
 387         if (ret)
 388                 goto out;
 389
 390         bc->table_init_done = true;
 391
 392         bch2_recalc_btree_reserve(c);
 393
 394         for (i = 0; i < bc->reserve; i++)
 395                 if (!btree_node_mem_alloc(c, GFP_KERNEL)) {
 396                         ret = -ENOMEM;
 397                         goto out;
 398                 }
 399
 400         list_splice_init(&bc->live, &bc->freeable);
 401
 402 #ifdef CONFIG_BCACHEFS_DEBUG
 403         mutex_init(&c->verify_lock);
 404
 405         c->verify_ondisk = kvpmalloc(btree_bytes(c), GFP_KERNEL);
 406         if (!c->verify_ondisk) {
 407                 ret = -ENOMEM;
 408                 goto out;
 409         }
 410
 411         c->verify_data = btree_node_mem_alloc(c, GFP_KERNEL);
 412         if (!c->verify_data) {
 413                 ret = -ENOMEM;
 414                 goto out;
 415         }
 416
 417         list_del_init(&c->verify_data->list);
 418 #endif
 419
 420         bc->shrink.count_objects        = bch2_btree_cache_count;
 421         bc->shrink.scan_objects         = bch2_btree_cache_scan;
 422         bc->shrink.seeks                = 4;
 423         bc->shrink.batch                = btree_pages(c) * 2;
 424         register_shrinker(&bc->shrink);
 425 out:
 426         pr_verbose_init(c->opts, "ret %i", ret);
 427         return ret;
 428 }
 429
 430 void bch2_fs_btree_cache_init_early(struct btree_cache *bc)
 431 {
 432         mutex_init(&bc->lock);
 433         INIT_LIST_HEAD(&bc->live);
 434         INIT_LIST_HEAD(&bc->freeable);
 435         INIT_LIST_HEAD(&bc->freed);
 436 }
 437
 438 /*
 439  * We can only have one thread cannibalizing other cached btree nodes at a time,
 440  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 441  * cannibalize_bucket() will take. This means every time we unlock the root of
 442  * the btree, we need to release this lock if we have it held.
 443  */
 444 void bch2_btree_cache_cannibalize_unlock(struct bch_fs *c)
 445 {
 446         struct btree_cache *bc = &c->btree_cache;
 447
 448         if (bc->alloc_lock == current) {
 449                 trace_btree_node_cannibalize_unlock(c);
 450                 bc->alloc_lock = NULL;
 451                 closure_wake_up(&bc->alloc_wait);
 452         }
 453 }
 454
 455 int bch2_btree_cache_cannibalize_lock(struct bch_fs *c, struct closure *cl)
 456 {
 457         struct btree_cache *bc = &c->btree_cache;
 458         struct task_struct *old;
 459
 460         old = cmpxchg(&bc->alloc_lock, NULL, current);
 461         if (old == NULL || old == current)
 462                 goto success;
 463
 464         if (!cl) {
 465                 trace_btree_node_cannibalize_lock_fail(c);
 466                 return -ENOMEM;
 467         }
 468
 469         closure_wait(&bc->alloc_wait, cl);
 470
 471         /* Try again, after adding ourselves to waitlist */
 472         old = cmpxchg(&bc->alloc_lock, NULL, current);
 473         if (old == NULL || old == current) {
 474                 /* We raced */
 475                 closure_wake_up(&bc->alloc_wait);
 476                 goto success;
 477         }
 478
 479         trace_btree_node_cannibalize_lock_fail(c);
 480         return -EAGAIN;
 481
 482 success:
 483         trace_btree_node_cannibalize_lock(c);
 484         return 0;
 485 }
 486
 487 static struct btree *btree_node_cannibalize(struct bch_fs *c)
 488 {
 489         struct btree_cache *bc = &c->btree_cache;
 490         struct btree *b;
 491
 492         list_for_each_entry_reverse(b, &bc->live, list)
 493                 if (!btree_node_reclaim(c, b))
 494                         return b;
 495
 496         while (1) {
 497                 list_for_each_entry_reverse(b, &bc->live, list)
 498                         if (!btree_node_write_and_reclaim(c, b))
 499                                 return b;
 500
 501                 /*
 502                  * Rare case: all nodes were intent-locked.
 503                  * Just busy-wait.
 504                  */
 505                 WARN_ONCE(1, "btree cache cannibalize failed\n");
 506                 cond_resched();
 507         }
 508 }
 509
 510 struct btree *bch2_btree_node_mem_alloc(struct bch_fs *c)
 511 {
 512         struct btree_cache *bc = &c->btree_cache;
 513         struct btree *b;
 514         u64 start_time = local_clock();
 515         unsigned flags;
 516
 517         flags = memalloc_nofs_save();
 518         mutex_lock(&bc->lock);
 519
 520         /*
 521          * btree_free() doesn't free memory; it sticks the node on the end of
 522          * the list. Check if there's any freed nodes there:
 523          */
 524         list_for_each_entry(b, &bc->freeable, list)
 525                 if (!btree_node_reclaim(c, b))
 526                         goto out_unlock;
 527
 528         /*
 529          * We never free struct btree itself, just the memory that holds the on
 530          * disk node. Check the freed list before allocating a new one:
 531          */
 532         list_for_each_entry(b, &bc->freed, list)
 533                 if (!btree_node_reclaim(c, b)) {
 534                         btree_node_data_alloc(c, b, __GFP_NOWARN|GFP_NOIO);
 535                         if (b->data)
 536                                 goto out_unlock;
 537
 538                         six_unlock_write(&b->lock);
 539                         six_unlock_intent(&b->lock);
 540                         goto err;
 541                 }
 542
 543         b = btree_node_mem_alloc(c, __GFP_NOWARN|GFP_NOIO);
 544         if (!b)
 545                 goto err;
 546
 547         BUG_ON(!six_trylock_intent(&b->lock));
 548         BUG_ON(!six_trylock_write(&b->lock));
 549 out_unlock:
 550         BUG_ON(btree_node_hashed(b));
 551         BUG_ON(btree_node_write_in_flight(b));
 552
 553         list_del_init(&b->list);
 554         mutex_unlock(&bc->lock);
 555         memalloc_nofs_restore(flags);
 556 out:
 557         b->flags                = 0;
 558         b->written              = 0;
 559         b->nsets                = 0;
 560         b->sib_u64s[0]          = 0;
 561         b->sib_u64s[1]          = 0;
 562         b->whiteout_u64s        = 0;
 563         b->uncompacted_whiteout_u64s = 0;
 564         bch2_btree_keys_init(b, &c->expensive_debug_checks);
 565
 566         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],
 567                                start_time);
 568
 569         return b;
 570 err:
 571         /* Try to cannibalize another cached btree node: */
 572         if (bc->alloc_lock == current) {
 573                 b = btree_node_cannibalize(c);
 574                 list_del_init(&b->list);
 575                 mutex_unlock(&bc->lock);
 576
 577                 bch2_btree_node_hash_remove(bc, b);
 578
 579                 trace_btree_node_cannibalize(c);
 580                 goto out;
 581         }
 582
 583         mutex_unlock(&bc->lock);
 584         return ERR_PTR(-ENOMEM);
 585 }
 586
 587 /* Slowpath, don't want it inlined into btree_iter_traverse() */
 588 static noinline struct btree *bch2_btree_node_fill(struct bch_fs *c,
 589                                 struct btree_iter *iter,
 590                                 const struct bkey_i *k,
 591                                 unsigned level,
 592                                 enum six_lock_type lock_type,
 593                                 bool sync)
 594 {
 595         struct btree_cache *bc = &c->btree_cache;
 596         struct btree *b;
 597
 598         /*
 599          * Parent node must be locked, else we could read in a btree node that's
 600          * been freed:
 601          */
 602         BUG_ON(!btree_node_locked(iter, level + 1));
 603         BUG_ON(level >= BTREE_MAX_DEPTH);
 604
 605         b = bch2_btree_node_mem_alloc(c);
 606         if (IS_ERR(b))
 607                 return b;
 608
 609         bkey_copy(&b->key, k);
 610         if (bch2_btree_node_hash_insert(bc, b, level, iter->btree_id)) {
 611                 /* raced with another fill: */
 612
 613                 /* mark as unhashed... */
 614                 PTR_HASH(&b->key) = 0;
 615
 616                 mutex_lock(&bc->lock);
 617                 list_add(&b->list, &bc->freeable);
 618                 mutex_unlock(&bc->lock);
 619
 620                 six_unlock_write(&b->lock);
 621                 six_unlock_intent(&b->lock);
 622                 return NULL;
 623         }
 624
 625         /*
 626          * If the btree node wasn't cached, we can't drop our lock on
 627          * the parent until after it's added to the cache - because
 628          * otherwise we could race with a btree_split() freeing the node
 629          * we're trying to lock.
 630          *
 631          * But the deadlock described below doesn't exist in this case,
 632          * so it's safe to not drop the parent lock until here:
 633          */
 634         if (btree_node_read_locked(iter, level + 1))
 635                 btree_node_unlock(iter, level + 1);
 636
 637         bch2_btree_node_read(c, b, sync);
 638
 639         six_unlock_write(&b->lock);
 640
 641         if (!sync) {
 642                 six_unlock_intent(&b->lock);
 643                 return NULL;
 644         }
 645
 646         if (lock_type == SIX_LOCK_read)
 647                 six_lock_downgrade(&b->lock);
 648
 649         return b;
 650 }
 651
 652 /**
 653  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 654  * in from disk if necessary.
 655  *
 656  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
 657  *
 658  * The btree node will have either a read or a write lock held, depending on
 659  * the @write parameter.
 660  */
 661 struct btree *bch2_btree_node_get(struct bch_fs *c, struct btree_iter *iter,
 662                                   const struct bkey_i *k, unsigned level,
 663                                   enum six_lock_type lock_type)
 664 {
 665         struct btree_cache *bc = &c->btree_cache;
 666         struct btree *b;
 667         struct bset_tree *t;
 668
 669         /*
 670          * XXX: locking optimization
 671          *
 672          * we can make the locking looser here - caller can drop lock on parent
 673          * node before locking child node (and potentially blocking): we just
 674          * have to have bch2_btree_node_fill() call relock on the parent and
 675          * return -EINTR if that fails
 676          */
 677         EBUG_ON(!btree_node_locked(iter, level + 1));
 678         EBUG_ON(level >= BTREE_MAX_DEPTH);
 679 retry:
 680         b = btree_cache_find(bc, k);
 681         if (unlikely(!b)) {
 682                 /*
 683                  * We must have the parent locked to call bch2_btree_node_fill(),
 684                  * else we could read in a btree node from disk that's been
 685                  * freed:
 686                  */
 687                 b = bch2_btree_node_fill(c, iter, k, level, lock_type, true);
 688
 689                 /* We raced and found the btree node in the cache */
 690                 if (!b)
 691                         goto retry;
 692
 693                 if (IS_ERR(b))
 694                         return b;
 695         } else {
 696                 /*
 697                  * There's a potential deadlock with splits and insertions into
 698                  * interior nodes we have to avoid:
 699                  *
 700                  * The other thread might be holding an intent lock on the node
 701                  * we want, and they want to update its parent node so they're
 702                  * going to upgrade their intent lock on the parent node to a
 703                  * write lock.
 704                  *
 705                  * But if we're holding a read lock on the parent, and we're
 706                  * trying to get the intent lock they're holding, we deadlock.
 707                  *
 708                  * So to avoid this we drop the read locks on parent nodes when
 709                  * we're starting to take intent locks - and handle the race.
 710                  *
 711                  * The race is that they might be about to free the node we
 712                  * want, and dropping our read lock on the parent node lets them
 713                  * update the parent marking the node we want as freed, and then
 714                  * free it:
 715                  *
 716                  * To guard against this, btree nodes are evicted from the cache
 717                  * when they're freed - and PTR_HASH() is zeroed out, which we
 718                  * check for after we lock the node.
 719                  *
 720                  * Then, bch2_btree_node_relock() on the parent will fail - because
 721                  * the parent was modified, when the pointer to the node we want
 722                  * was removed - and we'll bail out:
 723                  */
 724                 if (btree_node_read_locked(iter, level + 1))
 725                         btree_node_unlock(iter, level + 1);
 726
 727                 if (!btree_node_lock(b, k->k.p, level, iter, lock_type))
 728                         return ERR_PTR(-EINTR);
 729
 730                 if (unlikely(PTR_HASH(&b->key) != PTR_HASH(k) ||
 731                              b->level != level ||
 732                              race_fault())) {
 733                         six_unlock_type(&b->lock, lock_type);
 734                         if (bch2_btree_node_relock(iter, level + 1))
 735                                 goto retry;
 736
 737                         trace_trans_restart_btree_node_reused(iter->trans->ip);
 738                         return ERR_PTR(-EINTR);
 739                 }
 740         }
 741
 742         wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
 743                        TASK_UNINTERRUPTIBLE);
 744
 745         prefetch(b->aux_data);
 746
 747         for_each_bset(b, t) {
 748                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 749
 750                 prefetch(p + L1_CACHE_BYTES * 0);
 751                 prefetch(p + L1_CACHE_BYTES * 1);
 752                 prefetch(p + L1_CACHE_BYTES * 2);
 753         }
 754
 755         /* avoid atomic set bit if it's not needed: */
 756         if (btree_node_accessed(b))
 757                 set_btree_node_accessed(b);
 758
 759         if (unlikely(btree_node_read_error(b))) {
 760                 six_unlock_type(&b->lock, lock_type);
 761                 return ERR_PTR(-EIO);
 762         }
 763
 764         EBUG_ON(b->btree_id != iter->btree_id ||
 765                 BTREE_NODE_LEVEL(b->data) != level ||
 766                 bkey_cmp(b->data->max_key, k->k.p));
 767
 768         return b;
 769 }
 770
 771 struct btree *bch2_btree_node_get_sibling(struct bch_fs *c,
 772                                           struct btree_iter *iter,
 773                                           struct btree *b,
 774                                           enum btree_node_sibling sib)
 775 {
 776         struct btree_trans *trans = iter->trans;
 777         struct btree *parent;
 778         struct btree_node_iter node_iter;
 779         struct bkey_packed *k;
 780         BKEY_PADDED(k) tmp;
 781         struct btree *ret = NULL;
 782         unsigned level = b->level;
 783
 784         parent = btree_iter_node(iter, level + 1);
 785         if (!parent)
 786                 return NULL;
 787
 788         if (!bch2_btree_node_relock(iter, level + 1)) {
 789                 ret = ERR_PTR(-EINTR);
 790                 goto out;
 791         }
 792
 793         node_iter = iter->l[parent->level].iter;
 794
 795         k = bch2_btree_node_iter_peek_all(&node_iter, parent);
 796         BUG_ON(bkey_cmp_left_packed(parent, k, &b->key.k.p));
 797
 798         k = sib == btree_prev_sib
 799                 ? bch2_btree_node_iter_prev(&node_iter, parent)
 800                 : (bch2_btree_node_iter_advance(&node_iter, parent),
 801                    bch2_btree_node_iter_peek(&node_iter, parent));
 802         if (!k)
 803                 goto out;
 804
 805         bch2_bkey_unpack(parent, &tmp.k, k);
 806
 807         ret = bch2_btree_node_get(c, iter, &tmp.k, level,
 808                                   SIX_LOCK_intent);
 809
 810         if (PTR_ERR_OR_ZERO(ret) == -EINTR && !trans->nounlock) {
 811                 struct btree_iter *linked;
 812
 813                 if (!bch2_btree_node_relock(iter, level + 1))
 814                         goto out;
 815
 816                 /*
 817                  * We might have got -EINTR because trylock failed, and we're
 818                  * holding other locks that would cause us to deadlock:
 819                  */
 820                 trans_for_each_iter(trans, linked)
 821                         if (btree_iter_cmp(iter, linked) < 0)
 822                                 __bch2_btree_iter_unlock(linked);
 823
 824                 if (sib == btree_prev_sib)
 825                         btree_node_unlock(iter, level);
 826
 827                 ret = bch2_btree_node_get(c, iter, &tmp.k, level,
 828                                           SIX_LOCK_intent);
 829
 830                 /*
 831                  * before btree_iter_relock() calls btree_iter_verify_locks():
 832                  */
 833                 if (btree_lock_want(iter, level + 1) == BTREE_NODE_UNLOCKED)
 834                         btree_node_unlock(iter, level + 1);
 835
 836                 if (!bch2_btree_node_relock(iter, level)) {
 837                         btree_iter_set_dirty(iter, BTREE_ITER_NEED_RELOCK);
 838
 839                         if (!IS_ERR(ret)) {
 840                                 six_unlock_intent(&ret->lock);
 841                                 ret = ERR_PTR(-EINTR);
 842                         }
 843                 }
 844
 845                 bch2_trans_relock(trans);
 846         }
 847 out:
 848         if (btree_lock_want(iter, level + 1) == BTREE_NODE_UNLOCKED)
 849                 btree_node_unlock(iter, level + 1);
 850
 851         if (PTR_ERR_OR_ZERO(ret) == -EINTR)
 852                 bch2_btree_iter_upgrade(iter, level + 2);
 853
 854         BUG_ON(!IS_ERR(ret) && !btree_node_locked(iter, level));
 855
 856         if (!IS_ERR_OR_NULL(ret)) {
 857                 struct btree *n1 = ret, *n2 = b;
 858
 859                 if (sib != btree_prev_sib)
 860                         swap(n1, n2);
 861
 862                 BUG_ON(bkey_cmp(btree_type_successor(n1->btree_id,
 863                                                      n1->key.k.p),
 864                                 n2->data->min_key));
 865         }
 866
 867         bch2_btree_trans_verify_locks(trans);
 868
 869         return ret;
 870 }
 871
 872 void bch2_btree_node_prefetch(struct bch_fs *c, struct btree_iter *iter,
 873                               const struct bkey_i *k, unsigned level)
 874 {
 875         struct btree_cache *bc = &c->btree_cache;
 876         struct btree *b;
 877
 878         BUG_ON(!btree_node_locked(iter, level + 1));
 879         BUG_ON(level >= BTREE_MAX_DEPTH);
 880
 881         b = btree_cache_find(bc, k);
 882         if (b)
 883                 return;
 884
 885         bch2_btree_node_fill(c, iter, k, level, SIX_LOCK_read, false);
 886 }
 887
 888 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
 889                              struct btree *b)
 890 {
 891         const struct bkey_format *f = &b->format;
 892         struct bset_stats stats;
 893
 894         memset(&stats, 0, sizeof(stats));
 895
 896         bch2_btree_keys_stats(b, &stats);
 897
 898         pr_buf(out,
 899                "l %u %llu:%llu - %llu:%llu:\n"
 900                "    ptrs: ",
 901                b->level,
 902                b->data->min_key.inode,
 903                b->data->min_key.offset,
 904                b->data->max_key.inode,
 905                b->data->max_key.offset);
 906         bch2_val_to_text(out, c, bkey_i_to_s_c(&b->key));
 907         pr_buf(out, "\n"
 908                "    format: u64s %u fields %u %u %u %u %u\n"
 909                "    unpack fn len: %u\n"
 910                "    bytes used %zu/%zu (%zu%% full)\n"
 911                "    sib u64s: %u, %u (merge threshold %zu)\n"
 912                "    nr packed keys %u\n"
 913                "    nr unpacked keys %u\n"
 914                "    floats %zu\n"
 915                "    failed unpacked %zu\n"
 916                "    failed prev %zu\n"
 917                "    failed overflow %zu\n",
 918                f->key_u64s,
 919                f->bits_per_field[0],
 920                f->bits_per_field[1],
 921                f->bits_per_field[2],
 922                f->bits_per_field[3],
 923                f->bits_per_field[4],
 924                b->unpack_fn_len,
 925                b->nr.live_u64s * sizeof(u64),
 926                btree_bytes(c) - sizeof(struct btree_node),
 927                b->nr.live_u64s * 100 / btree_max_u64s(c),
 928                b->sib_u64s[0],
 929                b->sib_u64s[1],
 930                BTREE_FOREGROUND_MERGE_THRESHOLD(c),
 931                b->nr.packed_keys,
 932                b->nr.unpacked_keys,
 933                stats.floats,
 934                stats.failed_unpacked,
 935                stats.failed_prev,
 936                stats.failed_overflow);
 937 }