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->c.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         vfree(b->aux_data);
  48         b->aux_data = NULL;
  49 }
  50
  51 static void btree_node_data_free(struct bch_fs *c, struct btree *b)
  52 {
  53         struct btree_cache *bc = &c->btree_cache;
  54
  55         __btree_node_data_free(c, b);
  56         bc->used--;
  57         list_move(&b->list, &bc->freed);
  58 }
  59
  60 static int bch2_btree_cache_cmp_fn(struct rhashtable_compare_arg *arg,
  61                                    const void *obj)
  62 {
  63         const struct btree *b = obj;
  64         const u64 *v = arg->key;
  65
  66         return b->hash_val == *v ? 0 : 1;
  67 }
  68
  69 static const struct rhashtable_params bch_btree_cache_params = {
  70         .head_offset    = offsetof(struct btree, hash),
  71         .key_offset     = offsetof(struct btree, hash_val),
  72         .key_len        = sizeof(u64),
  73         .obj_cmpfn      = bch2_btree_cache_cmp_fn,
  74 };
  75
  76 static int btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
  77 {
  78         BUG_ON(b->data || b->aux_data);
  79
  80         b->data = kvpmalloc(btree_bytes(c), gfp);
  81         if (!b->data)
  82                 return -ENOMEM;
  83
  84         b->aux_data = vmalloc_exec(btree_aux_data_bytes(b), gfp);
  85         if (!b->aux_data) {
  86                 kvpfree(b->data, btree_bytes(c));
  87                 b->data = NULL;
  88                 return -ENOMEM;
  89         }
  90
  91         return 0;
  92 }
  93
  94 static struct btree *__btree_node_mem_alloc(struct bch_fs *c)
  95 {
  96         struct btree *b = kzalloc(sizeof(struct btree), GFP_KERNEL);
  97         if (!b)
  98                 return NULL;
  99
 100         bkey_btree_ptr_init(&b->key);
 101         six_lock_init(&b->c.lock);
 102         INIT_LIST_HEAD(&b->list);
 103         INIT_LIST_HEAD(&b->write_blocked);
 104         b->byte_order = ilog2(btree_bytes(c));
 105         return b;
 106 }
 107
 108 static struct btree *btree_node_mem_alloc(struct bch_fs *c)
 109 {
 110         struct btree_cache *bc = &c->btree_cache;
 111         struct btree *b = __btree_node_mem_alloc(c);
 112         if (!b)
 113                 return NULL;
 114
 115         if (btree_node_data_alloc(c, b, GFP_KERNEL)) {
 116                 kfree(b);
 117                 return NULL;
 118         }
 119
 120         bc->used++;
 121         list_add(&b->list, &bc->freeable);
 122         return b;
 123 }
 124
 125 /* Btree in memory cache - hash table */
 126
 127 void bch2_btree_node_hash_remove(struct btree_cache *bc, struct btree *b)
 128 {
 129         rhashtable_remove_fast(&bc->table, &b->hash, bch_btree_cache_params);
 130
 131         /* Cause future lookups for this node to fail: */
 132         b->hash_val = 0;
 133
 134         six_lock_wakeup_all(&b->c.lock);
 135 }
 136
 137 int __bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b)
 138 {
 139         BUG_ON(b->hash_val);
 140         b->hash_val = btree_ptr_hash_val(&b->key);
 141
 142         return rhashtable_lookup_insert_fast(&bc->table, &b->hash,
 143                                              bch_btree_cache_params);
 144 }
 145
 146 int bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b,
 147                                 unsigned level, enum btree_id id)
 148 {
 149         int ret;
 150
 151         b->c.level      = level;
 152         b->c.btree_id   = id;
 153
 154         mutex_lock(&bc->lock);
 155         ret = __bch2_btree_node_hash_insert(bc, b);
 156         if (!ret)
 157                 list_add(&b->list, &bc->live);
 158         mutex_unlock(&bc->lock);
 159
 160         return ret;
 161 }
 162
 163 __flatten
 164 static inline struct btree *btree_cache_find(struct btree_cache *bc,
 165                                      const struct bkey_i *k)
 166 {
 167         u64 v = btree_ptr_hash_val(k);
 168
 169         return rhashtable_lookup_fast(&bc->table, &v, bch_btree_cache_params);
 170 }
 171
 172 /*
 173  * this version is for btree nodes that have already been freed (we're not
 174  * reaping a real btree node)
 175  */
 176 static int __btree_node_reclaim(struct bch_fs *c, struct btree *b, bool flush)
 177 {
 178         struct btree_cache *bc = &c->btree_cache;
 179         int ret = 0;
 180
 181         lockdep_assert_held(&bc->lock);
 182
 183         if (!six_trylock_intent(&b->c.lock))
 184                 return -ENOMEM;
 185
 186         if (!six_trylock_write(&b->c.lock))
 187                 goto out_unlock_intent;
 188
 189         if (btree_node_noevict(b))
 190                 goto out_unlock;
 191
 192         if (!btree_node_may_write(b))
 193                 goto out_unlock;
 194
 195         if (btree_node_dirty(b) &&
 196             test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags))
 197                 goto out_unlock;
 198
 199         if (btree_node_dirty(b) ||
 200             btree_node_write_in_flight(b) ||
 201             btree_node_read_in_flight(b)) {
 202                 if (!flush)
 203                         goto out_unlock;
 204
 205                 wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
 206                                TASK_UNINTERRUPTIBLE);
 207
 208                 /*
 209                  * Using the underscore version because we don't want to compact
 210                  * bsets after the write, since this node is about to be evicted
 211                  * - unless btree verify mode is enabled, since it runs out of
 212                  * the post write cleanup:
 213                  */
 214                 if (verify_btree_ondisk(c))
 215                         bch2_btree_node_write(c, b, SIX_LOCK_intent);
 216                 else
 217                         __bch2_btree_node_write(c, b, SIX_LOCK_read);
 218
 219                 /* wait for any in flight btree write */
 220                 btree_node_wait_on_io(b);
 221         }
 222 out:
 223         if (b->hash_val && !ret)
 224                 trace_btree_node_reap(c, b);
 225         return ret;
 226 out_unlock:
 227         six_unlock_write(&b->c.lock);
 228 out_unlock_intent:
 229         six_unlock_intent(&b->c.lock);
 230         ret = -ENOMEM;
 231         goto out;
 232 }
 233
 234 static int btree_node_reclaim(struct bch_fs *c, struct btree *b)
 235 {
 236         return __btree_node_reclaim(c, b, false);
 237 }
 238
 239 static int btree_node_write_and_reclaim(struct bch_fs *c, struct btree *b)
 240 {
 241         return __btree_node_reclaim(c, b, true);
 242 }
 243
 244 static unsigned long bch2_btree_cache_scan(struct shrinker *shrink,
 245                                            struct shrink_control *sc)
 246 {
 247         struct bch_fs *c = container_of(shrink, struct bch_fs,
 248                                         btree_cache.shrink);
 249         struct btree_cache *bc = &c->btree_cache;
 250         struct btree *b, *t;
 251         unsigned long nr = sc->nr_to_scan;
 252         unsigned long can_free;
 253         unsigned long touched = 0;
 254         unsigned long freed = 0;
 255         unsigned i;
 256
 257         if (btree_shrinker_disabled(c))
 258                 return SHRINK_STOP;
 259
 260         /* Return -1 if we can't do anything right now */
 261         if (sc->gfp_mask & __GFP_FS)
 262                 mutex_lock(&bc->lock);
 263         else if (!mutex_trylock(&bc->lock))
 264                 return -1;
 265
 266         /*
 267          * It's _really_ critical that we don't free too many btree nodes - we
 268          * have to always leave ourselves a reserve. The reserve is how we
 269          * guarantee that allocating memory for a new btree node can always
 270          * succeed, so that inserting keys into the btree can always succeed and
 271          * IO can always make forward progress:
 272          */
 273         nr /= btree_pages(c);
 274         can_free = btree_cache_can_free(bc);
 275         nr = min_t(unsigned long, nr, can_free);
 276
 277         i = 0;
 278         list_for_each_entry_safe(b, t, &bc->freeable, list) {
 279                 touched++;
 280
 281                 if (freed >= nr)
 282                         break;
 283
 284                 if (++i > 3 &&
 285                     !btree_node_reclaim(c, b)) {
 286                         btree_node_data_free(c, b);
 287                         six_unlock_write(&b->c.lock);
 288                         six_unlock_intent(&b->c.lock);
 289                         freed++;
 290                 }
 291         }
 292 restart:
 293         list_for_each_entry_safe(b, t, &bc->live, list) {
 294                 touched++;
 295
 296                 if (freed >= nr) {
 297                         /* Save position */
 298                         if (&t->list != &bc->live)
 299                                 list_move_tail(&bc->live, &t->list);
 300                         break;
 301                 }
 302
 303                 if (!btree_node_accessed(b) &&
 304                     !btree_node_reclaim(c, b)) {
 305                         /* can't call bch2_btree_node_hash_remove under lock  */
 306                         freed++;
 307                         if (&t->list != &bc->live)
 308                                 list_move_tail(&bc->live, &t->list);
 309
 310                         btree_node_data_free(c, b);
 311                         mutex_unlock(&bc->lock);
 312
 313                         bch2_btree_node_hash_remove(bc, b);
 314                         six_unlock_write(&b->c.lock);
 315                         six_unlock_intent(&b->c.lock);
 316
 317                         if (freed >= nr)
 318                                 goto out;
 319
 320                         if (sc->gfp_mask & __GFP_FS)
 321                                 mutex_lock(&bc->lock);
 322                         else if (!mutex_trylock(&bc->lock))
 323                                 goto out;
 324                         goto restart;
 325                 } else
 326                         clear_btree_node_accessed(b);
 327         }
 328
 329         mutex_unlock(&bc->lock);
 330 out:
 331         return (unsigned long) freed * btree_pages(c);
 332 }
 333
 334 static unsigned long bch2_btree_cache_count(struct shrinker *shrink,
 335                                             struct shrink_control *sc)
 336 {
 337         struct bch_fs *c = container_of(shrink, struct bch_fs,
 338                                         btree_cache.shrink);
 339         struct btree_cache *bc = &c->btree_cache;
 340
 341         if (btree_shrinker_disabled(c))
 342                 return 0;
 343
 344         return btree_cache_can_free(bc) * btree_pages(c);
 345 }
 346
 347 void bch2_fs_btree_cache_exit(struct bch_fs *c)
 348 {
 349         struct btree_cache *bc = &c->btree_cache;
 350         struct btree *b;
 351         unsigned i;
 352
 353         if (bc->shrink.list.next)
 354                 unregister_shrinker(&bc->shrink);
 355
 356         mutex_lock(&bc->lock);
 357
 358 #ifdef CONFIG_BCACHEFS_DEBUG
 359         if (c->verify_data)
 360                 list_move(&c->verify_data->list, &bc->live);
 361
 362         kvpfree(c->verify_ondisk, btree_bytes(c));
 363 #endif
 364
 365         for (i = 0; i < BTREE_ID_NR; i++)
 366                 if (c->btree_roots[i].b)
 367                         list_add(&c->btree_roots[i].b->list, &bc->live);
 368
 369         list_splice(&bc->freeable, &bc->live);
 370
 371         while (!list_empty(&bc->live)) {
 372                 b = list_first_entry(&bc->live, struct btree, list);
 373
 374                 BUG_ON(btree_node_read_in_flight(b) ||
 375                        btree_node_write_in_flight(b));
 376
 377                 if (btree_node_dirty(b))
 378                         bch2_btree_complete_write(c, b, btree_current_write(b));
 379                 clear_btree_node_dirty(b);
 380
 381                 btree_node_data_free(c, b);
 382         }
 383
 384         while (!list_empty(&bc->freed)) {
 385                 b = list_first_entry(&bc->freed, struct btree, list);
 386                 list_del(&b->list);
 387                 kfree(b);
 388         }
 389
 390         mutex_unlock(&bc->lock);
 391
 392         if (bc->table_init_done)
 393                 rhashtable_destroy(&bc->table);
 394 }
 395
 396 int bch2_fs_btree_cache_init(struct bch_fs *c)
 397 {
 398         struct btree_cache *bc = &c->btree_cache;
 399         unsigned i;
 400         int ret = 0;
 401
 402         pr_verbose_init(c->opts, "");
 403
 404         ret = rhashtable_init(&bc->table, &bch_btree_cache_params);
 405         if (ret)
 406                 goto out;
 407
 408         bc->table_init_done = true;
 409
 410         bch2_recalc_btree_reserve(c);
 411
 412         for (i = 0; i < bc->reserve; i++)
 413                 if (!btree_node_mem_alloc(c)) {
 414                         ret = -ENOMEM;
 415                         goto out;
 416                 }
 417
 418         list_splice_init(&bc->live, &bc->freeable);
 419
 420 #ifdef CONFIG_BCACHEFS_DEBUG
 421         mutex_init(&c->verify_lock);
 422
 423         c->verify_ondisk = kvpmalloc(btree_bytes(c), GFP_KERNEL);
 424         if (!c->verify_ondisk) {
 425                 ret = -ENOMEM;
 426                 goto out;
 427         }
 428
 429         c->verify_data = btree_node_mem_alloc(c);
 430         if (!c->verify_data) {
 431                 ret = -ENOMEM;
 432                 goto out;
 433         }
 434
 435         list_del_init(&c->verify_data->list);
 436 #endif
 437
 438         bc->shrink.count_objects        = bch2_btree_cache_count;
 439         bc->shrink.scan_objects         = bch2_btree_cache_scan;
 440         bc->shrink.seeks                = 4;
 441         bc->shrink.batch                = btree_pages(c) * 2;
 442         register_shrinker(&bc->shrink);
 443 out:
 444         pr_verbose_init(c->opts, "ret %i", ret);
 445         return ret;
 446 }
 447
 448 void bch2_fs_btree_cache_init_early(struct btree_cache *bc)
 449 {
 450         mutex_init(&bc->lock);
 451         INIT_LIST_HEAD(&bc->live);
 452         INIT_LIST_HEAD(&bc->freeable);
 453         INIT_LIST_HEAD(&bc->freed);
 454 }
 455
 456 /*
 457  * We can only have one thread cannibalizing other cached btree nodes at a time,
 458  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 459  * cannibalize_bucket() will take. This means every time we unlock the root of
 460  * the btree, we need to release this lock if we have it held.
 461  */
 462 void bch2_btree_cache_cannibalize_unlock(struct bch_fs *c)
 463 {
 464         struct btree_cache *bc = &c->btree_cache;
 465
 466         if (bc->alloc_lock == current) {
 467                 trace_btree_node_cannibalize_unlock(c);
 468                 bc->alloc_lock = NULL;
 469                 closure_wake_up(&bc->alloc_wait);
 470         }
 471 }
 472
 473 int bch2_btree_cache_cannibalize_lock(struct bch_fs *c, struct closure *cl)
 474 {
 475         struct btree_cache *bc = &c->btree_cache;
 476         struct task_struct *old;
 477
 478         old = cmpxchg(&bc->alloc_lock, NULL, current);
 479         if (old == NULL || old == current)
 480                 goto success;
 481
 482         if (!cl) {
 483                 trace_btree_node_cannibalize_lock_fail(c);
 484                 return -ENOMEM;
 485         }
 486
 487         closure_wait(&bc->alloc_wait, cl);
 488
 489         /* Try again, after adding ourselves to waitlist */
 490         old = cmpxchg(&bc->alloc_lock, NULL, current);
 491         if (old == NULL || old == current) {
 492                 /* We raced */
 493                 closure_wake_up(&bc->alloc_wait);
 494                 goto success;
 495         }
 496
 497         trace_btree_node_cannibalize_lock_fail(c);
 498         return -EAGAIN;
 499
 500 success:
 501         trace_btree_node_cannibalize_lock(c);
 502         return 0;
 503 }
 504
 505 static struct btree *btree_node_cannibalize(struct bch_fs *c)
 506 {
 507         struct btree_cache *bc = &c->btree_cache;
 508         struct btree *b;
 509
 510         list_for_each_entry_reverse(b, &bc->live, list)
 511                 if (!btree_node_reclaim(c, b))
 512                         return b;
 513
 514         while (1) {
 515                 list_for_each_entry_reverse(b, &bc->live, list)
 516                         if (!btree_node_write_and_reclaim(c, b))
 517                                 return b;
 518
 519                 /*
 520                  * Rare case: all nodes were intent-locked.
 521                  * Just busy-wait.
 522                  */
 523                 WARN_ONCE(1, "btree cache cannibalize failed\n");
 524                 cond_resched();
 525         }
 526 }
 527
 528 struct btree *bch2_btree_node_mem_alloc(struct bch_fs *c)
 529 {
 530         struct btree_cache *bc = &c->btree_cache;
 531         struct btree *b;
 532         u64 start_time = local_clock();
 533         unsigned flags;
 534
 535         flags = memalloc_nofs_save();
 536         mutex_lock(&bc->lock);
 537
 538         /*
 539          * btree_free() doesn't free memory; it sticks the node on the end of
 540          * the list. Check if there's any freed nodes there:
 541          */
 542         list_for_each_entry(b, &bc->freeable, list)
 543                 if (!btree_node_reclaim(c, b))
 544                         goto got_node;
 545
 546         /*
 547          * We never free struct btree itself, just the memory that holds the on
 548          * disk node. Check the freed list before allocating a new one:
 549          */
 550         list_for_each_entry(b, &bc->freed, list)
 551                 if (!btree_node_reclaim(c, b))
 552                         goto got_node;
 553
 554         b = NULL;
 555 got_node:
 556         if (b)
 557                 list_del_init(&b->list);
 558         mutex_unlock(&bc->lock);
 559
 560         if (!b) {
 561                 b = __btree_node_mem_alloc(c);
 562                 if (!b)
 563                         goto err;
 564
 565                 BUG_ON(!six_trylock_intent(&b->c.lock));
 566                 BUG_ON(!six_trylock_write(&b->c.lock));
 567         }
 568
 569         if (!b->data) {
 570                 if (btree_node_data_alloc(c, b, __GFP_NOWARN|GFP_KERNEL))
 571                         goto err;
 572
 573                 mutex_lock(&bc->lock);
 574                 bc->used++;
 575                 mutex_unlock(&bc->lock);
 576         }
 577
 578         BUG_ON(btree_node_hashed(b));
 579         BUG_ON(btree_node_write_in_flight(b));
 580 out:
 581         b->flags                = 0;
 582         b->written              = 0;
 583         b->nsets                = 0;
 584         b->sib_u64s[0]          = 0;
 585         b->sib_u64s[1]          = 0;
 586         b->whiteout_u64s        = 0;
 587         bch2_btree_keys_init(b, &c->expensive_debug_checks);
 588
 589         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],
 590                                start_time);
 591
 592         memalloc_nofs_restore(flags);
 593         return b;
 594 err:
 595         mutex_lock(&bc->lock);
 596
 597         if (b) {
 598                 list_add(&b->list, &bc->freed);
 599                 six_unlock_write(&b->c.lock);
 600                 six_unlock_intent(&b->c.lock);
 601         }
 602
 603         /* Try to cannibalize another cached btree node: */
 604         if (bc->alloc_lock == current) {
 605                 b = btree_node_cannibalize(c);
 606                 list_del_init(&b->list);
 607                 mutex_unlock(&bc->lock);
 608
 609                 bch2_btree_node_hash_remove(bc, b);
 610
 611                 trace_btree_node_cannibalize(c);
 612                 goto out;
 613         }
 614
 615         mutex_unlock(&bc->lock);
 616         memalloc_nofs_restore(flags);
 617         return ERR_PTR(-ENOMEM);
 618 }
 619
 620 /* Slowpath, don't want it inlined into btree_iter_traverse() */
 621 static noinline struct btree *bch2_btree_node_fill(struct bch_fs *c,
 622                                 struct btree_iter *iter,
 623                                 const struct bkey_i *k,
 624                                 enum btree_id btree_id,
 625                                 unsigned level,
 626                                 enum six_lock_type lock_type,
 627                                 bool sync)
 628 {
 629         struct btree_cache *bc = &c->btree_cache;
 630         struct btree *b;
 631
 632         BUG_ON(level + 1 >= BTREE_MAX_DEPTH);
 633         /*
 634          * Parent node must be locked, else we could read in a btree node that's
 635          * been freed:
 636          */
 637         if (iter && !bch2_btree_node_relock(iter, level + 1))
 638                 return ERR_PTR(-EINTR);
 639
 640         b = bch2_btree_node_mem_alloc(c);
 641         if (IS_ERR(b))
 642                 return b;
 643
 644         bkey_copy(&b->key, k);
 645         if (bch2_btree_node_hash_insert(bc, b, level, btree_id)) {
 646                 /* raced with another fill: */
 647
 648                 /* mark as unhashed... */
 649                 b->hash_val = 0;
 650
 651                 mutex_lock(&bc->lock);
 652                 list_add(&b->list, &bc->freeable);
 653                 mutex_unlock(&bc->lock);
 654
 655                 six_unlock_write(&b->c.lock);
 656                 six_unlock_intent(&b->c.lock);
 657                 return NULL;
 658         }
 659
 660         /*
 661          * Unlock before doing IO:
 662          *
 663          * XXX: ideally should be dropping all btree node locks here
 664          */
 665         if (iter && btree_node_read_locked(iter, level + 1))
 666                 btree_node_unlock(iter, level + 1);
 667
 668         bch2_btree_node_read(c, b, sync);
 669
 670         six_unlock_write(&b->c.lock);
 671
 672         if (!sync) {
 673                 six_unlock_intent(&b->c.lock);
 674                 return NULL;
 675         }
 676
 677         if (lock_type == SIX_LOCK_read)
 678                 six_lock_downgrade(&b->c.lock);
 679
 680         return b;
 681 }
 682
 683 static int lock_node_check_fn(struct six_lock *lock, void *p)
 684 {
 685         struct btree *b = container_of(lock, struct btree, c.lock);
 686         const struct bkey_i *k = p;
 687
 688         return b->hash_val == btree_ptr_hash_val(k) ? 0 : -1;
 689 }
 690
 691 /**
 692  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 693  * in from disk if necessary.
 694  *
 695  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
 696  *
 697  * The btree node will have either a read or a write lock held, depending on
 698  * the @write parameter.
 699  */
 700 struct btree *bch2_btree_node_get(struct bch_fs *c, struct btree_iter *iter,
 701                                   const struct bkey_i *k, unsigned level,
 702                                   enum six_lock_type lock_type)
 703 {
 704         struct btree_cache *bc = &c->btree_cache;
 705         struct btree *b;
 706         struct bset_tree *t;
 707
 708         EBUG_ON(level >= BTREE_MAX_DEPTH);
 709
 710         b = btree_node_mem_ptr(k);
 711         if (b)
 712                 goto lock_node;
 713 retry:
 714         b = btree_cache_find(bc, k);
 715         if (unlikely(!b)) {
 716                 /*
 717                  * We must have the parent locked to call bch2_btree_node_fill(),
 718                  * else we could read in a btree node from disk that's been
 719                  * freed:
 720                  */
 721                 b = bch2_btree_node_fill(c, iter, k, iter->btree_id,
 722                                          level, lock_type, true);
 723
 724                 /* We raced and found the btree node in the cache */
 725                 if (!b)
 726                         goto retry;
 727
 728                 if (IS_ERR(b))
 729                         return b;
 730         } else {
 731 lock_node:
 732                 /*
 733                  * There's a potential deadlock with splits and insertions into
 734                  * interior nodes we have to avoid:
 735                  *
 736                  * The other thread might be holding an intent lock on the node
 737                  * we want, and they want to update its parent node so they're
 738                  * going to upgrade their intent lock on the parent node to a
 739                  * write lock.
 740                  *
 741                  * But if we're holding a read lock on the parent, and we're
 742                  * trying to get the intent lock they're holding, we deadlock.
 743                  *
 744                  * So to avoid this we drop the read locks on parent nodes when
 745                  * we're starting to take intent locks - and handle the race.
 746                  *
 747                  * The race is that they might be about to free the node we
 748                  * want, and dropping our read lock on the parent node lets them
 749                  * update the parent marking the node we want as freed, and then
 750                  * free it:
 751                  *
 752                  * To guard against this, btree nodes are evicted from the cache
 753                  * when they're freed - and b->hash_val is zeroed out, which we
 754                  * check for after we lock the node.
 755                  *
 756                  * Then, bch2_btree_node_relock() on the parent will fail - because
 757                  * the parent was modified, when the pointer to the node we want
 758                  * was removed - and we'll bail out:
 759                  */
 760                 if (btree_node_read_locked(iter, level + 1))
 761                         btree_node_unlock(iter, level + 1);
 762
 763                 if (!btree_node_lock(b, k->k.p, level, iter, lock_type,
 764                                      lock_node_check_fn, (void *) k)) {
 765                         if (b->hash_val != btree_ptr_hash_val(k))
 766                                 goto retry;
 767                         return ERR_PTR(-EINTR);
 768                 }
 769
 770                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
 771                              b->c.level != level ||
 772                              race_fault())) {
 773                         six_unlock_type(&b->c.lock, lock_type);
 774                         if (bch2_btree_node_relock(iter, level + 1))
 775                                 goto retry;
 776
 777                         trace_trans_restart_btree_node_reused(iter->trans->ip);
 778                         return ERR_PTR(-EINTR);
 779                 }
 780         }
 781
 782         /* XXX: waiting on IO with btree locks held: */
 783         wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
 784                        TASK_UNINTERRUPTIBLE);
 785
 786         prefetch(b->aux_data);
 787
 788         for_each_bset(b, t) {
 789                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 790
 791                 prefetch(p + L1_CACHE_BYTES * 0);
 792                 prefetch(p + L1_CACHE_BYTES * 1);
 793                 prefetch(p + L1_CACHE_BYTES * 2);
 794         }
 795
 796         /* avoid atomic set bit if it's not needed: */
 797         if (!btree_node_accessed(b))
 798                 set_btree_node_accessed(b);
 799
 800         if (unlikely(btree_node_read_error(b))) {
 801                 six_unlock_type(&b->c.lock, lock_type);
 802                 return ERR_PTR(-EIO);
 803         }
 804
 805         EBUG_ON(b->c.btree_id != iter->btree_id ||
 806                 BTREE_NODE_LEVEL(b->data) != level ||
 807                 bkey_cmp(b->data->max_key, k->k.p));
 808
 809         return b;
 810 }
 811
 812 struct btree *bch2_btree_node_get_noiter(struct bch_fs *c,
 813                                          const struct bkey_i *k,
 814                                          enum btree_id btree_id,
 815                                          unsigned level)
 816 {
 817         struct btree_cache *bc = &c->btree_cache;
 818         struct btree *b;
 819         struct bset_tree *t;
 820         int ret;
 821
 822         EBUG_ON(level >= BTREE_MAX_DEPTH);
 823
 824         b = btree_node_mem_ptr(k);
 825         if (b)
 826                 goto lock_node;
 827 retry:
 828         b = btree_cache_find(bc, k);
 829         if (unlikely(!b)) {
 830                 b = bch2_btree_node_fill(c, NULL, k, btree_id,
 831                                          level, SIX_LOCK_read, true);
 832
 833                 /* We raced and found the btree node in the cache */
 834                 if (!b)
 835                         goto retry;
 836
 837                 if (IS_ERR(b))
 838                         return b;
 839         } else {
 840 lock_node:
 841                 ret = six_lock_read(&b->c.lock, lock_node_check_fn, (void *) k);
 842                 if (ret)
 843                         goto retry;
 844
 845                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
 846                              b->c.btree_id != btree_id ||
 847                              b->c.level != level)) {
 848                         six_unlock_read(&b->c.lock);
 849                         goto retry;
 850                 }
 851         }
 852
 853         /* XXX: waiting on IO with btree locks held: */
 854         wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
 855                        TASK_UNINTERRUPTIBLE);
 856
 857         prefetch(b->aux_data);
 858
 859         for_each_bset(b, t) {
 860                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 861
 862                 prefetch(p + L1_CACHE_BYTES * 0);
 863                 prefetch(p + L1_CACHE_BYTES * 1);
 864                 prefetch(p + L1_CACHE_BYTES * 2);
 865         }
 866
 867         /* avoid atomic set bit if it's not needed: */
 868         if (!btree_node_accessed(b))
 869                 set_btree_node_accessed(b);
 870
 871         if (unlikely(btree_node_read_error(b))) {
 872                 six_unlock_read(&b->c.lock);
 873                 return ERR_PTR(-EIO);
 874         }
 875
 876         EBUG_ON(b->c.btree_id != btree_id ||
 877                 BTREE_NODE_LEVEL(b->data) != level ||
 878                 bkey_cmp(b->data->max_key, k->k.p));
 879
 880         return b;
 881 }
 882
 883 struct btree *bch2_btree_node_get_sibling(struct bch_fs *c,
 884                                           struct btree_iter *iter,
 885                                           struct btree *b,
 886                                           enum btree_node_sibling sib)
 887 {
 888         struct btree_trans *trans = iter->trans;
 889         struct btree *parent;
 890         struct btree_node_iter node_iter;
 891         struct bkey_packed *k;
 892         BKEY_PADDED(k) tmp;
 893         struct btree *ret = NULL;
 894         unsigned level = b->c.level;
 895
 896         parent = btree_iter_node(iter, level + 1);
 897         if (!parent)
 898                 return NULL;
 899
 900         /*
 901          * There's a corner case where a btree_iter might have a node locked
 902          * that is just outside its current pos - when
 903          * bch2_btree_iter_set_pos_same_leaf() gets to the end of the node.
 904          *
 905          * But the lock ordering checks in __bch2_btree_node_lock() go off of
 906          * iter->pos, not the node's key: so if the iterator is marked as
 907          * needing to be traversed, we risk deadlock if we don't bail out here:
 908          */
 909         if (iter->uptodate >= BTREE_ITER_NEED_TRAVERSE)
 910                 return ERR_PTR(-EINTR);
 911
 912         if (!bch2_btree_node_relock(iter, level + 1)) {
 913                 ret = ERR_PTR(-EINTR);
 914                 goto out;
 915         }
 916
 917         node_iter = iter->l[parent->c.level].iter;
 918
 919         k = bch2_btree_node_iter_peek_all(&node_iter, parent);
 920         BUG_ON(bkey_cmp_left_packed(parent, k, &b->key.k.p));
 921
 922         k = sib == btree_prev_sib
 923                 ? bch2_btree_node_iter_prev(&node_iter, parent)
 924                 : (bch2_btree_node_iter_advance(&node_iter, parent),
 925                    bch2_btree_node_iter_peek(&node_iter, parent));
 926         if (!k)
 927                 goto out;
 928
 929         bch2_bkey_unpack(parent, &tmp.k, k);
 930
 931         ret = bch2_btree_node_get(c, iter, &tmp.k, level,
 932                                   SIX_LOCK_intent);
 933
 934         if (PTR_ERR_OR_ZERO(ret) == -EINTR && !trans->nounlock) {
 935                 struct btree_iter *linked;
 936
 937                 if (!bch2_btree_node_relock(iter, level + 1))
 938                         goto out;
 939
 940                 /*
 941                  * We might have got -EINTR because trylock failed, and we're
 942                  * holding other locks that would cause us to deadlock:
 943                  */
 944                 trans_for_each_iter(trans, linked)
 945                         if (btree_iter_cmp(iter, linked) < 0)
 946                                 __bch2_btree_iter_unlock(linked);
 947
 948                 if (sib == btree_prev_sib)
 949                         btree_node_unlock(iter, level);
 950
 951                 ret = bch2_btree_node_get(c, iter, &tmp.k, level,
 952                                           SIX_LOCK_intent);
 953
 954                 /*
 955                  * before btree_iter_relock() calls btree_iter_verify_locks():
 956                  */
 957                 if (btree_lock_want(iter, level + 1) == BTREE_NODE_UNLOCKED)
 958                         btree_node_unlock(iter, level + 1);
 959
 960                 if (!bch2_btree_node_relock(iter, level)) {
 961                         btree_iter_set_dirty(iter, BTREE_ITER_NEED_RELOCK);
 962
 963                         if (!IS_ERR(ret)) {
 964                                 six_unlock_intent(&ret->c.lock);
 965                                 ret = ERR_PTR(-EINTR);
 966                         }
 967                 }
 968
 969                 bch2_trans_relock(trans);
 970         }
 971 out:
 972         if (btree_lock_want(iter, level + 1) == BTREE_NODE_UNLOCKED)
 973                 btree_node_unlock(iter, level + 1);
 974
 975         if (PTR_ERR_OR_ZERO(ret) == -EINTR)
 976                 bch2_btree_iter_upgrade(iter, level + 2);
 977
 978         BUG_ON(!IS_ERR(ret) && !btree_node_locked(iter, level));
 979
 980         if (!IS_ERR_OR_NULL(ret)) {
 981                 struct btree *n1 = ret, *n2 = b;
 982
 983                 if (sib != btree_prev_sib)
 984                         swap(n1, n2);
 985
 986                 BUG_ON(bkey_cmp(bkey_successor(n1->key.k.p),
 987                                 n2->data->min_key));
 988         }
 989
 990         bch2_btree_trans_verify_locks(trans);
 991
 992         return ret;
 993 }
 994
 995 void bch2_btree_node_prefetch(struct bch_fs *c, struct btree_iter *iter,
 996                               const struct bkey_i *k, unsigned level)
 997 {
 998         struct btree_cache *bc = &c->btree_cache;
 999         struct btree *b;
1000
1001         BUG_ON(!btree_node_locked(iter, level + 1));
1002         BUG_ON(level >= BTREE_MAX_DEPTH);
1003
1004         b = btree_cache_find(bc, k);
1005         if (b)
1006                 return;
1007
1008         bch2_btree_node_fill(c, iter, k, iter->btree_id,
1009                              level, SIX_LOCK_read, false);
1010 }
1011
1012 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
1013                              struct btree *b)
1014 {
1015         const struct bkey_format *f = &b->format;
1016         struct bset_stats stats;
1017
1018         memset(&stats, 0, sizeof(stats));
1019
1020         bch2_btree_keys_stats(b, &stats);
1021
1022         pr_buf(out,
1023                "l %u %llu:%llu - %llu:%llu:\n"
1024                "    ptrs: ",
1025                b->c.level,
1026                b->data->min_key.inode,
1027                b->data->min_key.offset,
1028                b->data->max_key.inode,
1029                b->data->max_key.offset);
1030         bch2_val_to_text(out, c, bkey_i_to_s_c(&b->key));
1031         pr_buf(out, "\n"
1032                "    format: u64s %u fields %u %u %u %u %u\n"
1033                "    unpack fn len: %u\n"
1034                "    bytes used %zu/%zu (%zu%% full)\n"
1035                "    sib u64s: %u, %u (merge threshold %zu)\n"
1036                "    nr packed keys %u\n"
1037                "    nr unpacked keys %u\n"
1038                "    floats %zu\n"
1039                "    failed unpacked %zu\n",
1040                f->key_u64s,
1041                f->bits_per_field[0],
1042                f->bits_per_field[1],
1043                f->bits_per_field[2],
1044                f->bits_per_field[3],
1045                f->bits_per_field[4],
1046                b->unpack_fn_len,
1047                b->nr.live_u64s * sizeof(u64),
1048                btree_bytes(c) - sizeof(struct btree_node),
1049                b->nr.live_u64s * 100 / btree_max_u64s(c),
1050                b->sib_u64s[0],
1051                b->sib_u64s[1],
1052                BTREE_FOREGROUND_MERGE_THRESHOLD(c),
1053                b->nr.packed_keys,
1054                b->nr.unpacked_keys,
1055                stats.floats,
1056                stats.failed);
1057 }