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

   1
   2 #include "bcache.h"
   3 #include "btree_cache.h"
   4 #include "btree_io.h"
   5 #include "btree_iter.h"
   6 #include "btree_locking.h"
   7 #include "debug.h"
   8 #include "extents.h"
   9
  10 #include <trace/events/bcache.h>
  11
  12 #define DEF_BTREE_ID(kwd, val, name) name,
  13
  14 const char * const bch_btree_ids[] = {
  15         DEFINE_BCH_BTREE_IDS()
  16         NULL
  17 };
  18
  19 #undef DEF_BTREE_ID
  20
  21 void bch_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 #define mca_can_free(c)                                         \
  37         max_t(int, 0, c->btree_cache_used - c->btree_cache_reserve)
  38
  39 static void __mca_data_free(struct bch_fs *c, struct btree *b)
  40 {
  41         EBUG_ON(btree_node_write_in_flight(b));
  42
  43         free_pages((unsigned long) b->data, btree_page_order(c));
  44         b->data = NULL;
  45         bch_btree_keys_free(b);
  46 }
  47
  48 static void mca_data_free(struct bch_fs *c, struct btree *b)
  49 {
  50         __mca_data_free(c, b);
  51         c->btree_cache_used--;
  52         list_move(&b->list, &c->btree_cache_freed);
  53 }
  54
  55 #define PTR_HASH(_k)    (bkey_i_to_extent_c(_k)->v._data[0])
  56
  57 static const struct rhashtable_params bch_btree_cache_params = {
  58         .head_offset    = offsetof(struct btree, hash),
  59         .key_offset     = offsetof(struct btree, key.v),
  60         .key_len        = sizeof(struct bch_extent_ptr),
  61 };
  62
  63 static void mca_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
  64 {
  65         unsigned order = ilog2(btree_pages(c));
  66
  67         b->data = (void *) __get_free_pages(gfp, order);
  68         if (!b->data)
  69                 goto err;
  70
  71         if (bch_btree_keys_alloc(b, order, gfp))
  72                 goto err;
  73
  74         c->btree_cache_used++;
  75         list_move(&b->list, &c->btree_cache_freeable);
  76         return;
  77 err:
  78         free_pages((unsigned long) b->data, order);
  79         b->data = NULL;
  80         list_move(&b->list, &c->btree_cache_freed);
  81 }
  82
  83 static struct btree *mca_bucket_alloc(struct bch_fs *c, gfp_t gfp)
  84 {
  85         struct btree *b = kzalloc(sizeof(struct btree), gfp);
  86         if (!b)
  87                 return NULL;
  88
  89         six_lock_init(&b->lock);
  90         INIT_LIST_HEAD(&b->list);
  91         INIT_LIST_HEAD(&b->write_blocked);
  92
  93         mca_data_alloc(c, b, gfp);
  94         return b->data ? b : NULL;
  95 }
  96
  97 /* Btree in memory cache - hash table */
  98
  99 void mca_hash_remove(struct bch_fs *c, struct btree *b)
 100 {
 101         BUG_ON(btree_node_dirty(b));
 102
 103         b->nsets = 0;
 104
 105         rhashtable_remove_fast(&c->btree_cache_table, &b->hash,
 106                                bch_btree_cache_params);
 107
 108         /* Cause future lookups for this node to fail: */
 109         bkey_i_to_extent(&b->key)->v._data[0] = 0;
 110 }
 111
 112 int mca_hash_insert(struct bch_fs *c, struct btree *b,
 113                     unsigned level, enum btree_id id)
 114 {
 115         int ret;
 116         b->level        = level;
 117         b->btree_id     = id;
 118
 119         ret = rhashtable_lookup_insert_fast(&c->btree_cache_table, &b->hash,
 120                                             bch_btree_cache_params);
 121         if (ret)
 122                 return ret;
 123
 124         mutex_lock(&c->btree_cache_lock);
 125         list_add(&b->list, &c->btree_cache);
 126         mutex_unlock(&c->btree_cache_lock);
 127
 128         return 0;
 129 }
 130
 131 __flatten
 132 static inline struct btree *mca_find(struct bch_fs *c,
 133                                      const struct bkey_i *k)
 134 {
 135         return rhashtable_lookup_fast(&c->btree_cache_table, &PTR_HASH(k),
 136                                       bch_btree_cache_params);
 137 }
 138
 139 /*
 140  * this version is for btree nodes that have already been freed (we're not
 141  * reaping a real btree node)
 142  */
 143 static int mca_reap_notrace(struct bch_fs *c, struct btree *b, bool flush)
 144 {
 145         lockdep_assert_held(&c->btree_cache_lock);
 146
 147         if (!six_trylock_intent(&b->lock))
 148                 return -ENOMEM;
 149
 150         if (!six_trylock_write(&b->lock))
 151                 goto out_unlock_intent;
 152
 153         if (btree_node_write_error(b) ||
 154             btree_node_noevict(b))
 155                 goto out_unlock;
 156
 157         if (!list_empty(&b->write_blocked))
 158                 goto out_unlock;
 159
 160         if (!flush &&
 161             (btree_node_dirty(b) ||
 162              btree_node_write_in_flight(b)))
 163                 goto out_unlock;
 164
 165         /*
 166          * Using the underscore version because we don't want to compact bsets
 167          * after the write, since this node is about to be evicted - unless
 168          * btree verify mode is enabled, since it runs out of the post write
 169          * cleanup:
 170          */
 171         if (btree_node_dirty(b)) {
 172                 if (verify_btree_ondisk(c))
 173                         bch_btree_node_write(c, b, NULL, SIX_LOCK_intent, -1);
 174                 else
 175                         __bch_btree_node_write(c, b, NULL, SIX_LOCK_read, -1);
 176         }
 177
 178         /* wait for any in flight btree write */
 179         wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight,
 180                        TASK_UNINTERRUPTIBLE);
 181
 182         return 0;
 183 out_unlock:
 184         six_unlock_write(&b->lock);
 185 out_unlock_intent:
 186         six_unlock_intent(&b->lock);
 187         return -ENOMEM;
 188 }
 189
 190 static int mca_reap(struct bch_fs *c, struct btree *b, bool flush)
 191 {
 192         int ret = mca_reap_notrace(c, b, flush);
 193
 194         trace_bcache_mca_reap(c, b, ret);
 195         return ret;
 196 }
 197
 198 static unsigned long bch_mca_scan(struct shrinker *shrink,
 199                                   struct shrink_control *sc)
 200 {
 201         struct bch_fs *c = container_of(shrink, struct bch_fs,
 202                                            btree_cache_shrink);
 203         struct btree *b, *t;
 204         unsigned long nr = sc->nr_to_scan;
 205         unsigned long can_free;
 206         unsigned long touched = 0;
 207         unsigned long freed = 0;
 208         unsigned i;
 209
 210         u64 start_time = local_clock();
 211
 212         if (btree_shrinker_disabled(c))
 213                 return SHRINK_STOP;
 214
 215         if (c->btree_cache_alloc_lock)
 216                 return SHRINK_STOP;
 217
 218         /* Return -1 if we can't do anything right now */
 219         if (sc->gfp_mask & __GFP_IO)
 220                 mutex_lock(&c->btree_cache_lock);
 221         else if (!mutex_trylock(&c->btree_cache_lock))
 222                 return -1;
 223
 224         /*
 225          * It's _really_ critical that we don't free too many btree nodes - we
 226          * have to always leave ourselves a reserve. The reserve is how we
 227          * guarantee that allocating memory for a new btree node can always
 228          * succeed, so that inserting keys into the btree can always succeed and
 229          * IO can always make forward progress:
 230          */
 231         nr /= btree_pages(c);
 232         can_free = mca_can_free(c);
 233         nr = min_t(unsigned long, nr, can_free);
 234
 235         i = 0;
 236         list_for_each_entry_safe(b, t, &c->btree_cache_freeable, list) {
 237                 touched++;
 238
 239                 if (freed >= nr)
 240                         break;
 241
 242                 if (++i > 3 &&
 243                     !mca_reap_notrace(c, b, false)) {
 244                         mca_data_free(c, b);
 245                         six_unlock_write(&b->lock);
 246                         six_unlock_intent(&b->lock);
 247                         freed++;
 248                 }
 249         }
 250 restart:
 251         list_for_each_entry_safe(b, t, &c->btree_cache, list) {
 252                 touched++;
 253
 254                 if (freed >= nr) {
 255                         /* Save position */
 256                         if (&t->list != &c->btree_cache)
 257                                 list_move_tail(&c->btree_cache, &t->list);
 258                         break;
 259                 }
 260
 261                 if (!btree_node_accessed(b) &&
 262                     !mca_reap(c, b, false)) {
 263                         /* can't call mca_hash_remove under btree_cache_lock  */
 264                         freed++;
 265                         if (&t->list != &c->btree_cache)
 266                                 list_move_tail(&c->btree_cache, &t->list);
 267
 268                         mca_data_free(c, b);
 269                         mutex_unlock(&c->btree_cache_lock);
 270
 271                         mca_hash_remove(c, b);
 272                         six_unlock_write(&b->lock);
 273                         six_unlock_intent(&b->lock);
 274
 275                         if (freed >= nr)
 276                                 goto out;
 277
 278                         if (sc->gfp_mask & __GFP_IO)
 279                                 mutex_lock(&c->btree_cache_lock);
 280                         else if (!mutex_trylock(&c->btree_cache_lock))
 281                                 goto out;
 282                         goto restart;
 283                 } else
 284                         clear_btree_node_accessed(b);
 285         }
 286
 287         mutex_unlock(&c->btree_cache_lock);
 288 out:
 289         bch_time_stats_update(&c->mca_scan_time, start_time);
 290
 291         trace_bcache_mca_scan(c,
 292                               touched * btree_pages(c),
 293                               freed * btree_pages(c),
 294                               can_free * btree_pages(c),
 295                               sc->nr_to_scan);
 296
 297         return (unsigned long) freed * btree_pages(c);
 298 }
 299
 300 static unsigned long bch_mca_count(struct shrinker *shrink,
 301                                    struct shrink_control *sc)
 302 {
 303         struct bch_fs *c = container_of(shrink, struct bch_fs,
 304                                            btree_cache_shrink);
 305
 306         if (btree_shrinker_disabled(c))
 307                 return 0;
 308
 309         if (c->btree_cache_alloc_lock)
 310                 return 0;
 311
 312         return mca_can_free(c) * btree_pages(c);
 313 }
 314
 315 void bch_fs_btree_exit(struct bch_fs *c)
 316 {
 317         struct btree *b;
 318         unsigned i;
 319
 320         if (c->btree_cache_shrink.list.next)
 321                 unregister_shrinker(&c->btree_cache_shrink);
 322
 323         mutex_lock(&c->btree_cache_lock);
 324
 325 #ifdef CONFIG_BCACHE_DEBUG
 326         if (c->verify_data)
 327                 list_move(&c->verify_data->list, &c->btree_cache);
 328
 329         free_pages((unsigned long) c->verify_ondisk, ilog2(btree_pages(c)));
 330 #endif
 331
 332         for (i = 0; i < BTREE_ID_NR; i++)
 333                 if (c->btree_roots[i].b)
 334                         list_add(&c->btree_roots[i].b->list, &c->btree_cache);
 335
 336         list_splice(&c->btree_cache_freeable,
 337                     &c->btree_cache);
 338
 339         while (!list_empty(&c->btree_cache)) {
 340                 b = list_first_entry(&c->btree_cache, struct btree, list);
 341
 342                 if (btree_node_dirty(b))
 343                         bch_btree_complete_write(c, b, btree_current_write(b));
 344                 clear_btree_node_dirty(b);
 345
 346                 mca_data_free(c, b);
 347         }
 348
 349         while (!list_empty(&c->btree_cache_freed)) {
 350                 b = list_first_entry(&c->btree_cache_freed,
 351                                      struct btree, list);
 352                 list_del(&b->list);
 353                 kfree(b);
 354         }
 355
 356         mutex_unlock(&c->btree_cache_lock);
 357
 358         if (c->btree_cache_table_init_done)
 359                 rhashtable_destroy(&c->btree_cache_table);
 360 }
 361
 362 int bch_fs_btree_init(struct bch_fs *c)
 363 {
 364         unsigned i;
 365         int ret;
 366
 367         ret = rhashtable_init(&c->btree_cache_table, &bch_btree_cache_params);
 368         if (ret)
 369                 return ret;
 370
 371         c->btree_cache_table_init_done = true;
 372
 373         bch_recalc_btree_reserve(c);
 374
 375         for (i = 0; i < c->btree_cache_reserve; i++)
 376                 if (!mca_bucket_alloc(c, GFP_KERNEL))
 377                         return -ENOMEM;
 378
 379         list_splice_init(&c->btree_cache,
 380                          &c->btree_cache_freeable);
 381
 382 #ifdef CONFIG_BCACHE_DEBUG
 383         mutex_init(&c->verify_lock);
 384
 385         c->verify_ondisk = (void *)
 386                 __get_free_pages(GFP_KERNEL, ilog2(btree_pages(c)));
 387         if (!c->verify_ondisk)
 388                 return -ENOMEM;
 389
 390         c->verify_data = mca_bucket_alloc(c, GFP_KERNEL);
 391         if (!c->verify_data)
 392                 return -ENOMEM;
 393
 394         list_del_init(&c->verify_data->list);
 395 #endif
 396
 397         c->btree_cache_shrink.count_objects = bch_mca_count;
 398         c->btree_cache_shrink.scan_objects = bch_mca_scan;
 399         c->btree_cache_shrink.seeks = 4;
 400         c->btree_cache_shrink.batch = btree_pages(c) * 2;
 401         register_shrinker(&c->btree_cache_shrink);
 402
 403         return 0;
 404 }
 405
 406 /*
 407  * We can only have one thread cannibalizing other cached btree nodes at a time,
 408  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 409  * cannibalize_bucket() will take. This means every time we unlock the root of
 410  * the btree, we need to release this lock if we have it held.
 411  */
 412 void mca_cannibalize_unlock(struct bch_fs *c)
 413 {
 414         if (c->btree_cache_alloc_lock == current) {
 415                 trace_bcache_mca_cannibalize_unlock(c);
 416                 c->btree_cache_alloc_lock = NULL;
 417                 closure_wake_up(&c->mca_wait);
 418         }
 419 }
 420
 421 int mca_cannibalize_lock(struct bch_fs *c, struct closure *cl)
 422 {
 423         struct task_struct *old;
 424
 425         old = cmpxchg(&c->btree_cache_alloc_lock, NULL, current);
 426         if (old == NULL || old == current)
 427                 goto success;
 428
 429         if (!cl) {
 430                 trace_bcache_mca_cannibalize_lock_fail(c);
 431                 return -ENOMEM;
 432         }
 433
 434         closure_wait(&c->mca_wait, cl);
 435
 436         /* Try again, after adding ourselves to waitlist */
 437         old = cmpxchg(&c->btree_cache_alloc_lock, NULL, current);
 438         if (old == NULL || old == current) {
 439                 /* We raced */
 440                 closure_wake_up(&c->mca_wait);
 441                 goto success;
 442         }
 443
 444         trace_bcache_mca_cannibalize_lock_fail(c);
 445         return -EAGAIN;
 446
 447 success:
 448         trace_bcache_mca_cannibalize_lock(c);
 449         return 0;
 450 }
 451
 452 static struct btree *mca_cannibalize(struct bch_fs *c)
 453 {
 454         struct btree *b;
 455
 456         list_for_each_entry_reverse(b, &c->btree_cache, list)
 457                 if (!mca_reap(c, b, false))
 458                         return b;
 459
 460         while (1) {
 461                 list_for_each_entry_reverse(b, &c->btree_cache, list)
 462                         if (!mca_reap(c, b, true))
 463                                 return b;
 464
 465                 /*
 466                  * Rare case: all nodes were intent-locked.
 467                  * Just busy-wait.
 468                  */
 469                 WARN_ONCE(1, "btree cache cannibalize failed\n");
 470                 cond_resched();
 471         }
 472 }
 473
 474 struct btree *mca_alloc(struct bch_fs *c)
 475 {
 476         struct btree *b;
 477         u64 start_time = local_clock();
 478
 479         mutex_lock(&c->btree_cache_lock);
 480
 481         /*
 482          * btree_free() doesn't free memory; it sticks the node on the end of
 483          * the list. Check if there's any freed nodes there:
 484          */
 485         list_for_each_entry(b, &c->btree_cache_freeable, list)
 486                 if (!mca_reap_notrace(c, b, false))
 487                         goto out_unlock;
 488
 489         /*
 490          * We never free struct btree itself, just the memory that holds the on
 491          * disk node. Check the freed list before allocating a new one:
 492          */
 493         list_for_each_entry(b, &c->btree_cache_freed, list)
 494                 if (!mca_reap_notrace(c, b, false)) {
 495                         mca_data_alloc(c, b, __GFP_NOWARN|GFP_NOIO);
 496                         if (b->data)
 497                                 goto out_unlock;
 498
 499                         six_unlock_write(&b->lock);
 500                         six_unlock_intent(&b->lock);
 501                         goto err;
 502                 }
 503
 504         b = mca_bucket_alloc(c, __GFP_NOWARN|GFP_NOIO);
 505         if (!b)
 506                 goto err;
 507
 508         BUG_ON(!six_trylock_intent(&b->lock));
 509         BUG_ON(!six_trylock_write(&b->lock));
 510 out_unlock:
 511         BUG_ON(bkey_extent_is_data(&b->key.k) && PTR_HASH(&b->key));
 512         BUG_ON(btree_node_write_in_flight(b));
 513
 514         list_del_init(&b->list);
 515         mutex_unlock(&c->btree_cache_lock);
 516 out:
 517         b->flags                = 0;
 518         b->written              = 0;
 519         b->nsets                = 0;
 520         b->sib_u64s[0]          = 0;
 521         b->sib_u64s[1]          = 0;
 522         b->whiteout_u64s        = 0;
 523         b->uncompacted_whiteout_u64s = 0;
 524         bch_btree_keys_init(b, &c->expensive_debug_checks);
 525
 526         bch_time_stats_update(&c->mca_alloc_time, start_time);
 527
 528         return b;
 529 err:
 530         /* Try to cannibalize another cached btree node: */
 531         if (c->btree_cache_alloc_lock == current) {
 532                 b = mca_cannibalize(c);
 533                 list_del_init(&b->list);
 534                 mutex_unlock(&c->btree_cache_lock);
 535
 536                 mca_hash_remove(c, b);
 537
 538                 trace_bcache_mca_cannibalize(c);
 539                 goto out;
 540         }
 541
 542         mutex_unlock(&c->btree_cache_lock);
 543         return ERR_PTR(-ENOMEM);
 544 }
 545
 546 /* Slowpath, don't want it inlined into btree_iter_traverse() */
 547 static noinline struct btree *bch_btree_node_fill(struct btree_iter *iter,
 548                                                   const struct bkey_i *k,
 549                                                   unsigned level,
 550                                                   enum six_lock_type lock_type)
 551 {
 552         struct bch_fs *c = iter->c;
 553         struct btree *b;
 554
 555         b = mca_alloc(c);
 556         if (IS_ERR(b))
 557                 return b;
 558
 559         bkey_copy(&b->key, k);
 560         if (mca_hash_insert(c, b, level, iter->btree_id)) {
 561                 /* raced with another fill: */
 562
 563                 /* mark as unhashed... */
 564                 bkey_i_to_extent(&b->key)->v._data[0] = 0;
 565
 566                 mutex_lock(&c->btree_cache_lock);
 567                 list_add(&b->list, &c->btree_cache_freeable);
 568                 mutex_unlock(&c->btree_cache_lock);
 569
 570                 six_unlock_write(&b->lock);
 571                 six_unlock_intent(&b->lock);
 572                 return NULL;
 573         }
 574
 575         /*
 576          * If the btree node wasn't cached, we can't drop our lock on
 577          * the parent until after it's added to the cache - because
 578          * otherwise we could race with a btree_split() freeing the node
 579          * we're trying to lock.
 580          *
 581          * But the deadlock described below doesn't exist in this case,
 582          * so it's safe to not drop the parent lock until here:
 583          */
 584         if (btree_node_read_locked(iter, level + 1))
 585                 btree_node_unlock(iter, level + 1);
 586
 587         bch_btree_node_read(c, b);
 588         six_unlock_write(&b->lock);
 589
 590         if (lock_type == SIX_LOCK_read)
 591                 six_lock_downgrade(&b->lock);
 592
 593         return b;
 594 }
 595
 596 /**
 597  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 598  * in from disk if necessary.
 599  *
 600  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
 601  *
 602  * The btree node will have either a read or a write lock held, depending on
 603  * the @write parameter.
 604  */
 605 struct btree *bch_btree_node_get(struct btree_iter *iter,
 606                                  const struct bkey_i *k, unsigned level,
 607                                  enum six_lock_type lock_type)
 608 {
 609         struct btree *b;
 610         struct bset_tree *t;
 611
 612         BUG_ON(level >= BTREE_MAX_DEPTH);
 613 retry:
 614         rcu_read_lock();
 615         b = mca_find(iter->c, k);
 616         rcu_read_unlock();
 617
 618         if (unlikely(!b)) {
 619                 /*
 620                  * We must have the parent locked to call bch_btree_node_fill(),
 621                  * else we could read in a btree node from disk that's been
 622                  * freed:
 623                  */
 624                 b = bch_btree_node_fill(iter, k, level, lock_type);
 625
 626                 /* We raced and found the btree node in the cache */
 627                 if (!b)
 628                         goto retry;
 629
 630                 if (IS_ERR(b))
 631                         return b;
 632         } else {
 633                 /*
 634                  * There's a potential deadlock with splits and insertions into
 635                  * interior nodes we have to avoid:
 636                  *
 637                  * The other thread might be holding an intent lock on the node
 638                  * we want, and they want to update its parent node so they're
 639                  * going to upgrade their intent lock on the parent node to a
 640                  * write lock.
 641                  *
 642                  * But if we're holding a read lock on the parent, and we're
 643                  * trying to get the intent lock they're holding, we deadlock.
 644                  *
 645                  * So to avoid this we drop the read locks on parent nodes when
 646                  * we're starting to take intent locks - and handle the race.
 647                  *
 648                  * The race is that they might be about to free the node we
 649                  * want, and dropping our read lock on the parent node lets them
 650                  * update the parent marking the node we want as freed, and then
 651                  * free it:
 652                  *
 653                  * To guard against this, btree nodes are evicted from the cache
 654                  * when they're freed - and PTR_HASH() is zeroed out, which we
 655                  * check for after we lock the node.
 656                  *
 657                  * Then, btree_node_relock() on the parent will fail - because
 658                  * the parent was modified, when the pointer to the node we want
 659                  * was removed - and we'll bail out:
 660                  */
 661                 if (btree_node_read_locked(iter, level + 1))
 662                         btree_node_unlock(iter, level + 1);
 663
 664                 if (!btree_node_lock(b, k->k.p, level, iter, lock_type))
 665                         return ERR_PTR(-EINTR);
 666
 667                 if (unlikely(PTR_HASH(&b->key) != PTR_HASH(k) ||
 668                              b->level != level ||
 669                              race_fault())) {
 670                         six_unlock_type(&b->lock, lock_type);
 671                         if (btree_node_relock(iter, level + 1))
 672                                 goto retry;
 673
 674                         return ERR_PTR(-EINTR);
 675                 }
 676         }
 677
 678         prefetch(b->aux_data);
 679
 680         for_each_bset(b, t) {
 681                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 682
 683                 prefetch(p + L1_CACHE_BYTES * 0);
 684                 prefetch(p + L1_CACHE_BYTES * 1);
 685                 prefetch(p + L1_CACHE_BYTES * 2);
 686         }
 687
 688         /* avoid atomic set bit if it's not needed: */
 689         if (btree_node_accessed(b))
 690                 set_btree_node_accessed(b);
 691
 692         if (unlikely(btree_node_read_error(b))) {
 693                 six_unlock_type(&b->lock, lock_type);
 694                 return ERR_PTR(-EIO);
 695         }
 696
 697         EBUG_ON(!b->written);
 698         EBUG_ON(b->btree_id != iter->btree_id ||
 699                 BTREE_NODE_LEVEL(b->data) != level ||
 700                 bkey_cmp(b->data->max_key, k->k.p));
 701
 702         return b;
 703 }
 704
 705 int bch_print_btree_node(struct bch_fs *c, struct btree *b,
 706                          char *buf, size_t len)
 707 {
 708         const struct bkey_format *f = &b->format;
 709         struct bset_stats stats;
 710         char ptrs[100];
 711
 712         memset(&stats, 0, sizeof(stats));
 713
 714         bch_val_to_text(c, BKEY_TYPE_BTREE, ptrs, sizeof(ptrs),
 715                         bkey_i_to_s_c(&b->key));
 716         bch_btree_keys_stats(b, &stats);
 717
 718         return scnprintf(buf, len,
 719                          "l %u %llu:%llu - %llu:%llu:\n"
 720                          "    ptrs: %s\n"
 721                          "    format: u64s %u fields %u %u %u %u %u\n"
 722                          "    unpack fn len: %u\n"
 723                          "    bytes used %zu/%zu (%zu%% full)\n"
 724                          "    sib u64s: %u, %u (merge threshold %zu)\n"
 725                          "    nr packed keys %u\n"
 726                          "    nr unpacked keys %u\n"
 727                          "    floats %zu\n"
 728                          "    failed unpacked %zu\n"
 729                          "    failed prev %zu\n"
 730                          "    failed overflow %zu\n",
 731                          b->level,
 732                          b->data->min_key.inode,
 733                          b->data->min_key.offset,
 734                          b->data->max_key.inode,
 735                          b->data->max_key.offset,
 736                          ptrs,
 737                          f->key_u64s,
 738                          f->bits_per_field[0],
 739                          f->bits_per_field[1],
 740                          f->bits_per_field[2],
 741                          f->bits_per_field[3],
 742                          f->bits_per_field[4],
 743                          b->unpack_fn_len,
 744                          b->nr.live_u64s * sizeof(u64),
 745                          btree_bytes(c) - sizeof(struct btree_node),
 746                          b->nr.live_u64s * 100 / btree_max_u64s(c),
 747                          b->sib_u64s[0],
 748                          b->sib_u64s[1],
 749                          BTREE_FOREGROUND_MERGE_THRESHOLD(c),
 750                          b->nr.packed_keys,
 751                          b->nr.unpacked_keys,
 752                          stats.floats,
 753                          stats.failed_unpacked,
 754                          stats.failed_prev,
 755                          stats.failed_overflow);
 756 }