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

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "bkey_buf.h"
   5 #include "btree_cache.h"
   6 #include "btree_io.h"
   7 #include "btree_iter.h"
   8 #include "btree_locking.h"
   9 #include "debug.h"
  10 #include "error.h"
  11
  12 #include <linux/prefetch.h>
  13 #include <linux/sched/mm.h>
  14 #include <trace/events/bcachefs.h>
  15
  16 void bch2_recalc_btree_reserve(struct bch_fs *c)
  17 {
  18         unsigned i, reserve = 16;
  19
  20         if (!c->btree_roots[0].b)
  21                 reserve += 8;
  22
  23         for (i = 0; i < BTREE_ID_NR; i++)
  24                 if (c->btree_roots[i].b)
  25                         reserve += min_t(unsigned, 1,
  26                                          c->btree_roots[i].b->c.level) * 8;
  27
  28         c->btree_cache.reserve = reserve;
  29 }
  30
  31 static inline unsigned btree_cache_can_free(struct btree_cache *bc)
  32 {
  33         return max_t(int, 0, bc->used - bc->reserve);
  34 }
  35
  36 static void btree_node_data_free(struct bch_fs *c, struct btree *b)
  37 {
  38         struct btree_cache *bc = &c->btree_cache;
  39
  40         EBUG_ON(btree_node_write_in_flight(b));
  41
  42         kvpfree(b->data, btree_bytes(c));
  43         b->data = NULL;
  44 #ifdef __KERNEL__
  45         vfree(b->aux_data);
  46 #else
  47         munmap(b->aux_data, btree_aux_data_bytes(b));
  48 #endif
  49         b->aux_data = NULL;
  50
  51         bc->used--;
  52         list_move(&b->list, &bc->freed);
  53 }
  54
  55 static int bch2_btree_cache_cmp_fn(struct rhashtable_compare_arg *arg,
  56                                    const void *obj)
  57 {
  58         const struct btree *b = obj;
  59         const u64 *v = arg->key;
  60
  61         return b->hash_val == *v ? 0 : 1;
  62 }
  63
  64 static const struct rhashtable_params bch_btree_cache_params = {
  65         .head_offset    = offsetof(struct btree, hash),
  66         .key_offset     = offsetof(struct btree, hash_val),
  67         .key_len        = sizeof(u64),
  68         .obj_cmpfn      = bch2_btree_cache_cmp_fn,
  69 };
  70
  71 static int btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
  72 {
  73         BUG_ON(b->data || b->aux_data);
  74
  75         b->data = kvpmalloc(btree_bytes(c), gfp);
  76         if (!b->data)
  77                 return -ENOMEM;
  78 #ifdef __KERNEL__
  79         b->aux_data = vmalloc_exec(btree_aux_data_bytes(b), gfp);
  80 #else
  81         b->aux_data = mmap(NULL, btree_aux_data_bytes(b),
  82                            PROT_READ|PROT_WRITE|PROT_EXEC,
  83                            MAP_PRIVATE|MAP_ANONYMOUS, 0, 0);
  84 #endif
  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 struct btree *__bch2_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         if (level)
 155                 six_lock_pcpu_alloc(&b->c.lock);
 156         else
 157                 six_lock_pcpu_free_rcu(&b->c.lock);
 158
 159         mutex_lock(&bc->lock);
 160         ret = __bch2_btree_node_hash_insert(bc, b);
 161         if (!ret)
 162                 list_add(&b->list, &bc->live);
 163         mutex_unlock(&bc->lock);
 164
 165         return ret;
 166 }
 167
 168 __flatten
 169 static inline struct btree *btree_cache_find(struct btree_cache *bc,
 170                                      const struct bkey_i *k)
 171 {
 172         u64 v = btree_ptr_hash_val(k);
 173
 174         return rhashtable_lookup_fast(&bc->table, &v, bch_btree_cache_params);
 175 }
 176
 177 /*
 178  * this version is for btree nodes that have already been freed (we're not
 179  * reaping a real btree node)
 180  */
 181 static int __btree_node_reclaim(struct bch_fs *c, struct btree *b, bool flush)
 182 {
 183         struct btree_cache *bc = &c->btree_cache;
 184         int ret = 0;
 185
 186         lockdep_assert_held(&bc->lock);
 187
 188         if (!six_trylock_intent(&b->c.lock))
 189                 return -ENOMEM;
 190
 191         if (!six_trylock_write(&b->c.lock))
 192                 goto out_unlock_intent;
 193
 194         if (btree_node_noevict(b))
 195                 goto out_unlock;
 196
 197         if (!btree_node_may_write(b))
 198                 goto out_unlock;
 199
 200         if (btree_node_dirty(b) &&
 201             test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags))
 202                 goto out_unlock;
 203
 204         if (btree_node_dirty(b) ||
 205             btree_node_write_in_flight(b) ||
 206             btree_node_read_in_flight(b)) {
 207                 if (!flush)
 208                         goto out_unlock;
 209
 210                 wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
 211                                TASK_UNINTERRUPTIBLE);
 212
 213                 /*
 214                  * Using the underscore version because we don't want to compact
 215                  * bsets after the write, since this node is about to be evicted
 216                  * - unless btree verify mode is enabled, since it runs out of
 217                  * the post write cleanup:
 218                  */
 219                 if (bch2_verify_btree_ondisk)
 220                         bch2_btree_node_write(c, b, SIX_LOCK_intent);
 221                 else
 222                         __bch2_btree_node_write(c, b);
 223
 224                 /* wait for any in flight btree write */
 225                 btree_node_wait_on_io(b);
 226         }
 227 out:
 228         if (b->hash_val && !ret)
 229                 trace_btree_node_reap(c, b);
 230         return ret;
 231 out_unlock:
 232         six_unlock_write(&b->c.lock);
 233 out_unlock_intent:
 234         six_unlock_intent(&b->c.lock);
 235         ret = -ENOMEM;
 236         goto out;
 237 }
 238
 239 static int btree_node_reclaim(struct bch_fs *c, struct btree *b)
 240 {
 241         return __btree_node_reclaim(c, b, false);
 242 }
 243
 244 static int btree_node_write_and_reclaim(struct bch_fs *c, struct btree *b)
 245 {
 246         return __btree_node_reclaim(c, b, true);
 247 }
 248
 249 static unsigned long bch2_btree_cache_scan(struct shrinker *shrink,
 250                                            struct shrink_control *sc)
 251 {
 252         struct bch_fs *c = container_of(shrink, struct bch_fs,
 253                                         btree_cache.shrink);
 254         struct btree_cache *bc = &c->btree_cache;
 255         struct btree *b, *t;
 256         unsigned long nr = sc->nr_to_scan;
 257         unsigned long can_free;
 258         unsigned long touched = 0;
 259         unsigned long freed = 0;
 260         unsigned i, flags;
 261
 262         if (bch2_btree_shrinker_disabled)
 263                 return SHRINK_STOP;
 264
 265         /* Return -1 if we can't do anything right now */
 266         if (sc->gfp_mask & __GFP_FS)
 267                 mutex_lock(&bc->lock);
 268         else if (!mutex_trylock(&bc->lock))
 269                 return -1;
 270
 271         flags = memalloc_nofs_save();
 272
 273         /*
 274          * It's _really_ critical that we don't free too many btree nodes - we
 275          * have to always leave ourselves a reserve. The reserve is how we
 276          * guarantee that allocating memory for a new btree node can always
 277          * succeed, so that inserting keys into the btree can always succeed and
 278          * IO can always make forward progress:
 279          */
 280         nr /= btree_pages(c);
 281         can_free = btree_cache_can_free(bc);
 282         nr = min_t(unsigned long, nr, can_free);
 283
 284         i = 0;
 285         list_for_each_entry_safe(b, t, &bc->freeable, list) {
 286                 touched++;
 287
 288                 if (freed >= nr)
 289                         break;
 290
 291                 if (++i > 3 &&
 292                     !btree_node_reclaim(c, b)) {
 293                         btree_node_data_free(c, b);
 294                         six_unlock_write(&b->c.lock);
 295                         six_unlock_intent(&b->c.lock);
 296                         freed++;
 297                 }
 298         }
 299 restart:
 300         list_for_each_entry_safe(b, t, &bc->live, list) {
 301                 touched++;
 302
 303                 if (freed >= nr) {
 304                         /* Save position */
 305                         if (&t->list != &bc->live)
 306                                 list_move_tail(&bc->live, &t->list);
 307                         break;
 308                 }
 309
 310                 if (!btree_node_accessed(b) &&
 311                     !btree_node_reclaim(c, b)) {
 312                         /* can't call bch2_btree_node_hash_remove under lock  */
 313                         freed++;
 314                         if (&t->list != &bc->live)
 315                                 list_move_tail(&bc->live, &t->list);
 316
 317                         btree_node_data_free(c, b);
 318                         mutex_unlock(&bc->lock);
 319
 320                         bch2_btree_node_hash_remove(bc, b);
 321                         six_unlock_write(&b->c.lock);
 322                         six_unlock_intent(&b->c.lock);
 323
 324                         if (freed >= nr)
 325                                 goto out;
 326
 327                         if (sc->gfp_mask & __GFP_FS)
 328                                 mutex_lock(&bc->lock);
 329                         else if (!mutex_trylock(&bc->lock))
 330                                 goto out;
 331                         goto restart;
 332                 } else
 333                         clear_btree_node_accessed(b);
 334         }
 335
 336         mutex_unlock(&bc->lock);
 337 out:
 338         memalloc_nofs_restore(flags);
 339         return (unsigned long) freed * btree_pages(c);
 340 }
 341
 342 static unsigned long bch2_btree_cache_count(struct shrinker *shrink,
 343                                             struct shrink_control *sc)
 344 {
 345         struct bch_fs *c = container_of(shrink, struct bch_fs,
 346                                         btree_cache.shrink);
 347         struct btree_cache *bc = &c->btree_cache;
 348
 349         if (bch2_btree_shrinker_disabled)
 350                 return 0;
 351
 352         return btree_cache_can_free(bc) * btree_pages(c);
 353 }
 354
 355 void bch2_fs_btree_cache_exit(struct bch_fs *c)
 356 {
 357         struct btree_cache *bc = &c->btree_cache;
 358         struct btree *b;
 359         unsigned i, flags;
 360
 361         if (bc->shrink.list.next)
 362                 unregister_shrinker(&bc->shrink);
 363
 364         /* vfree() can allocate memory: */
 365         flags = memalloc_nofs_save();
 366         mutex_lock(&bc->lock);
 367
 368         if (c->verify_data)
 369                 list_move(&c->verify_data->list, &bc->live);
 370
 371         kvpfree(c->verify_ondisk, btree_bytes(c));
 372
 373         for (i = 0; i < BTREE_ID_NR; i++)
 374                 if (c->btree_roots[i].b)
 375                         list_add(&c->btree_roots[i].b->list, &bc->live);
 376
 377         list_splice(&bc->freeable, &bc->live);
 378
 379         while (!list_empty(&bc->live)) {
 380                 b = list_first_entry(&bc->live, struct btree, list);
 381
 382                 BUG_ON(btree_node_read_in_flight(b) ||
 383                        btree_node_write_in_flight(b));
 384
 385                 if (btree_node_dirty(b))
 386                         bch2_btree_complete_write(c, b, btree_current_write(b));
 387                 clear_btree_node_dirty(c, b);
 388
 389                 btree_node_data_free(c, b);
 390         }
 391
 392         BUG_ON(atomic_read(&c->btree_cache.dirty));
 393
 394         while (!list_empty(&bc->freed)) {
 395                 b = list_first_entry(&bc->freed, struct btree, list);
 396                 list_del(&b->list);
 397                 six_lock_pcpu_free(&b->c.lock);
 398                 kfree(b);
 399         }
 400
 401         mutex_unlock(&bc->lock);
 402         memalloc_nofs_restore(flags);
 403
 404         if (bc->table_init_done)
 405                 rhashtable_destroy(&bc->table);
 406 }
 407
 408 int bch2_fs_btree_cache_init(struct bch_fs *c)
 409 {
 410         struct btree_cache *bc = &c->btree_cache;
 411         unsigned i;
 412         int ret = 0;
 413
 414         pr_verbose_init(c->opts, "");
 415
 416         ret = rhashtable_init(&bc->table, &bch_btree_cache_params);
 417         if (ret)
 418                 goto out;
 419
 420         bc->table_init_done = true;
 421
 422         bch2_recalc_btree_reserve(c);
 423
 424         for (i = 0; i < bc->reserve; i++)
 425                 if (!__bch2_btree_node_mem_alloc(c)) {
 426                         ret = -ENOMEM;
 427                         goto out;
 428                 }
 429
 430         list_splice_init(&bc->live, &bc->freeable);
 431
 432         mutex_init(&c->verify_lock);
 433
 434         bc->shrink.count_objects        = bch2_btree_cache_count;
 435         bc->shrink.scan_objects         = bch2_btree_cache_scan;
 436         bc->shrink.seeks                = 4;
 437         bc->shrink.batch                = btree_pages(c) * 2;
 438         ret = register_shrinker(&bc->shrink);
 439 out:
 440         pr_verbose_init(c->opts, "ret %i", ret);
 441         return ret;
 442 }
 443
 444 void bch2_fs_btree_cache_init_early(struct btree_cache *bc)
 445 {
 446         mutex_init(&bc->lock);
 447         INIT_LIST_HEAD(&bc->live);
 448         INIT_LIST_HEAD(&bc->freeable);
 449         INIT_LIST_HEAD(&bc->freed);
 450 }
 451
 452 /*
 453  * We can only have one thread cannibalizing other cached btree nodes at a time,
 454  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 455  * cannibalize_bucket() will take. This means every time we unlock the root of
 456  * the btree, we need to release this lock if we have it held.
 457  */
 458 void bch2_btree_cache_cannibalize_unlock(struct bch_fs *c)
 459 {
 460         struct btree_cache *bc = &c->btree_cache;
 461
 462         if (bc->alloc_lock == current) {
 463                 trace_btree_node_cannibalize_unlock(c);
 464                 bc->alloc_lock = NULL;
 465                 closure_wake_up(&bc->alloc_wait);
 466         }
 467 }
 468
 469 int bch2_btree_cache_cannibalize_lock(struct bch_fs *c, struct closure *cl)
 470 {
 471         struct btree_cache *bc = &c->btree_cache;
 472         struct task_struct *old;
 473
 474         old = cmpxchg(&bc->alloc_lock, NULL, current);
 475         if (old == NULL || old == current)
 476                 goto success;
 477
 478         if (!cl) {
 479                 trace_btree_node_cannibalize_lock_fail(c);
 480                 return -ENOMEM;
 481         }
 482
 483         closure_wait(&bc->alloc_wait, cl);
 484
 485         /* Try again, after adding ourselves to waitlist */
 486         old = cmpxchg(&bc->alloc_lock, NULL, current);
 487         if (old == NULL || old == current) {
 488                 /* We raced */
 489                 closure_wake_up(&bc->alloc_wait);
 490                 goto success;
 491         }
 492
 493         trace_btree_node_cannibalize_lock_fail(c);
 494         return -EAGAIN;
 495
 496 success:
 497         trace_btree_node_cannibalize_lock(c);
 498         return 0;
 499 }
 500
 501 static struct btree *btree_node_cannibalize(struct bch_fs *c)
 502 {
 503         struct btree_cache *bc = &c->btree_cache;
 504         struct btree *b;
 505
 506         list_for_each_entry_reverse(b, &bc->live, list)
 507                 if (!btree_node_reclaim(c, b))
 508                         return b;
 509
 510         while (1) {
 511                 list_for_each_entry_reverse(b, &bc->live, list)
 512                         if (!btree_node_write_and_reclaim(c, b))
 513                                 return b;
 514
 515                 /*
 516                  * Rare case: all nodes were intent-locked.
 517                  * Just busy-wait.
 518                  */
 519                 WARN_ONCE(1, "btree cache cannibalize failed\n");
 520                 cond_resched();
 521         }
 522 }
 523
 524 struct btree *bch2_btree_node_mem_alloc(struct bch_fs *c)
 525 {
 526         struct btree_cache *bc = &c->btree_cache;
 527         struct btree *b;
 528         u64 start_time = local_clock();
 529         unsigned flags;
 530
 531         flags = memalloc_nofs_save();
 532         mutex_lock(&bc->lock);
 533
 534         /*
 535          * btree_free() doesn't free memory; it sticks the node on the end of
 536          * the list. Check if there's any freed nodes there:
 537          */
 538         list_for_each_entry(b, &bc->freeable, list)
 539                 if (!btree_node_reclaim(c, b))
 540                         goto got_node;
 541
 542         /*
 543          * We never free struct btree itself, just the memory that holds the on
 544          * disk node. Check the freed list before allocating a new one:
 545          */
 546         list_for_each_entry(b, &bc->freed, list)
 547                 if (!btree_node_reclaim(c, b))
 548                         goto got_node;
 549
 550         b = NULL;
 551 got_node:
 552         if (b)
 553                 list_del_init(&b->list);
 554         mutex_unlock(&bc->lock);
 555
 556         if (!b) {
 557                 b = __btree_node_mem_alloc(c);
 558                 if (!b)
 559                         goto err;
 560
 561                 BUG_ON(!six_trylock_intent(&b->c.lock));
 562                 BUG_ON(!six_trylock_write(&b->c.lock));
 563         }
 564
 565         if (!b->data) {
 566                 if (btree_node_data_alloc(c, b, __GFP_NOWARN|GFP_KERNEL))
 567                         goto err;
 568
 569                 mutex_lock(&bc->lock);
 570                 bc->used++;
 571                 mutex_unlock(&bc->lock);
 572         }
 573
 574         BUG_ON(btree_node_hashed(b));
 575         BUG_ON(btree_node_write_in_flight(b));
 576 out:
 577         b->flags                = 0;
 578         b->written              = 0;
 579         b->nsets                = 0;
 580         b->sib_u64s[0]          = 0;
 581         b->sib_u64s[1]          = 0;
 582         b->whiteout_u64s        = 0;
 583         bch2_btree_keys_init(b);
 584         set_btree_node_accessed(b);
 585
 586         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],
 587                                start_time);
 588
 589         memalloc_nofs_restore(flags);
 590         return b;
 591 err:
 592         mutex_lock(&bc->lock);
 593
 594         if (b) {
 595                 list_add(&b->list, &bc->freed);
 596                 six_unlock_write(&b->c.lock);
 597                 six_unlock_intent(&b->c.lock);
 598         }
 599
 600         /* Try to cannibalize another cached btree node: */
 601         if (bc->alloc_lock == current) {
 602                 b = btree_node_cannibalize(c);
 603                 list_del_init(&b->list);
 604                 mutex_unlock(&bc->lock);
 605
 606                 bch2_btree_node_hash_remove(bc, b);
 607
 608                 trace_btree_node_cannibalize(c);
 609                 goto out;
 610         }
 611
 612         mutex_unlock(&bc->lock);
 613         memalloc_nofs_restore(flags);
 614         return ERR_PTR(-ENOMEM);
 615 }
 616
 617 /* Slowpath, don't want it inlined into btree_iter_traverse() */
 618 static noinline struct btree *bch2_btree_node_fill(struct bch_fs *c,
 619                                 struct btree_iter *iter,
 620                                 const struct bkey_i *k,
 621                                 enum btree_id btree_id,
 622                                 unsigned level,
 623                                 enum six_lock_type lock_type,
 624                                 bool sync)
 625 {
 626         struct btree_cache *bc = &c->btree_cache;
 627         struct btree *b;
 628
 629         BUG_ON(level + 1 >= BTREE_MAX_DEPTH);
 630         /*
 631          * Parent node must be locked, else we could read in a btree node that's
 632          * been freed:
 633          */
 634         if (iter && !bch2_btree_node_relock(iter, level + 1))
 635                 return ERR_PTR(-EINTR);
 636
 637         b = bch2_btree_node_mem_alloc(c);
 638         if (IS_ERR(b))
 639                 return b;
 640
 641         bkey_copy(&b->key, k);
 642         if (bch2_btree_node_hash_insert(bc, b, level, btree_id)) {
 643                 /* raced with another fill: */
 644
 645                 /* mark as unhashed... */
 646                 b->hash_val = 0;
 647
 648                 mutex_lock(&bc->lock);
 649                 list_add(&b->list, &bc->freeable);
 650                 mutex_unlock(&bc->lock);
 651
 652                 six_unlock_write(&b->c.lock);
 653                 six_unlock_intent(&b->c.lock);
 654                 return NULL;
 655         }
 656
 657         /* Unlock before doing IO: */
 658         if (iter && sync)
 659                 bch2_trans_unlock(iter->trans);
 660
 661         bch2_btree_node_read(c, b, sync);
 662
 663         six_unlock_write(&b->c.lock);
 664
 665         if (!sync) {
 666                 six_unlock_intent(&b->c.lock);
 667                 return NULL;
 668         }
 669
 670         /*
 671          * XXX: this will probably always fail because btree_iter_relock()
 672          * currently fails for iterators that aren't pointed at a valid btree
 673          * node
 674          */
 675         if (iter && !bch2_trans_relock(iter->trans)) {
 676                 six_unlock_intent(&b->c.lock);
 677                 return ERR_PTR(-EINTR);
 678         }
 679
 680         if (lock_type == SIX_LOCK_read)
 681                 six_lock_downgrade(&b->c.lock);
 682
 683         return b;
 684 }
 685
 686 static int lock_node_check_fn(struct six_lock *lock, void *p)
 687 {
 688         struct btree *b = container_of(lock, struct btree, c.lock);
 689         const struct bkey_i *k = p;
 690
 691         return b->hash_val == btree_ptr_hash_val(k) ? 0 : -1;
 692 }
 693
 694 static noinline void btree_bad_header(struct bch_fs *c, struct btree *b)
 695 {
 696         char buf1[100], buf2[100], buf3[100], buf4[100];
 697
 698         if (!test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags))
 699                 return;
 700
 701         bch2_bpos_to_text(&PBUF(buf1), b->key.k.type == KEY_TYPE_btree_ptr_v2
 702                 ? bkey_i_to_btree_ptr_v2(&b->key)->v.min_key
 703                 : POS_MIN);
 704         bch2_bpos_to_text(&PBUF(buf2), b->data->min_key);
 705
 706         bch2_bpos_to_text(&PBUF(buf3), b->key.k.p);
 707         bch2_bpos_to_text(&PBUF(buf4), b->data->max_key);
 708         bch2_fs_inconsistent(c, "btree node header doesn't match ptr\n"
 709                              "btree: ptr %u header %llu\n"
 710                              "level: ptr %u header %llu\n"
 711                              "min ptr %s node header %s\n"
 712                              "max ptr %s node header %s",
 713                              b->c.btree_id,     BTREE_NODE_ID(b->data),
 714                              b->c.level,        BTREE_NODE_LEVEL(b->data),
 715                              buf1, buf2, buf3, buf4);
 716 }
 717
 718 static inline void btree_check_header(struct bch_fs *c, struct btree *b)
 719 {
 720         if (b->c.btree_id != BTREE_NODE_ID(b->data) ||
 721             b->c.level != BTREE_NODE_LEVEL(b->data) ||
 722             bpos_cmp(b->data->max_key, b->key.k.p) ||
 723             (b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 724              bpos_cmp(b->data->min_key,
 725                       bkey_i_to_btree_ptr_v2(&b->key)->v.min_key)))
 726                 btree_bad_header(c, b);
 727 }
 728
 729 /**
 730  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 731  * in from disk if necessary.
 732  *
 733  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
 734  *
 735  * The btree node will have either a read or a write lock held, depending on
 736  * the @write parameter.
 737  */
 738 struct btree *bch2_btree_node_get(struct bch_fs *c, struct btree_iter *iter,
 739                                   const struct bkey_i *k, unsigned level,
 740                                   enum six_lock_type lock_type,
 741                                   unsigned long trace_ip)
 742 {
 743         struct btree_cache *bc = &c->btree_cache;
 744         struct btree *b;
 745         struct bset_tree *t;
 746
 747         EBUG_ON(level >= BTREE_MAX_DEPTH);
 748
 749         b = btree_node_mem_ptr(k);
 750         if (b)
 751                 goto lock_node;
 752 retry:
 753         b = btree_cache_find(bc, k);
 754         if (unlikely(!b)) {
 755                 /*
 756                  * We must have the parent locked to call bch2_btree_node_fill(),
 757                  * else we could read in a btree node from disk that's been
 758                  * freed:
 759                  */
 760                 b = bch2_btree_node_fill(c, iter, k, iter->btree_id,
 761                                          level, lock_type, true);
 762
 763                 /* We raced and found the btree node in the cache */
 764                 if (!b)
 765                         goto retry;
 766
 767                 if (IS_ERR(b))
 768                         return b;
 769         } else {
 770 lock_node:
 771                 /*
 772                  * There's a potential deadlock with splits and insertions into
 773                  * interior nodes we have to avoid:
 774                  *
 775                  * The other thread might be holding an intent lock on the node
 776                  * we want, and they want to update its parent node so they're
 777                  * going to upgrade their intent lock on the parent node to a
 778                  * write lock.
 779                  *
 780                  * But if we're holding a read lock on the parent, and we're
 781                  * trying to get the intent lock they're holding, we deadlock.
 782                  *
 783                  * So to avoid this we drop the read locks on parent nodes when
 784                  * we're starting to take intent locks - and handle the race.
 785                  *
 786                  * The race is that they might be about to free the node we
 787                  * want, and dropping our read lock on the parent node lets them
 788                  * update the parent marking the node we want as freed, and then
 789                  * free it:
 790                  *
 791                  * To guard against this, btree nodes are evicted from the cache
 792                  * when they're freed - and b->hash_val is zeroed out, which we
 793                  * check for after we lock the node.
 794                  *
 795                  * Then, bch2_btree_node_relock() on the parent will fail - because
 796                  * the parent was modified, when the pointer to the node we want
 797                  * was removed - and we'll bail out:
 798                  */
 799                 if (btree_node_read_locked(iter, level + 1))
 800                         btree_node_unlock(iter, level + 1);
 801
 802                 if (!btree_node_lock(b, k->k.p, level, iter, lock_type,
 803                                      lock_node_check_fn, (void *) k, trace_ip)) {
 804                         if (b->hash_val != btree_ptr_hash_val(k))
 805                                 goto retry;
 806                         return ERR_PTR(-EINTR);
 807                 }
 808
 809                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
 810                              b->c.level != level ||
 811                              race_fault())) {
 812                         six_unlock_type(&b->c.lock, lock_type);
 813                         if (bch2_btree_node_relock(iter, level + 1))
 814                                 goto retry;
 815
 816                         trace_trans_restart_btree_node_reused(iter->trans->ip,
 817                                                               trace_ip,
 818                                                               iter->btree_id,
 819                                                               &iter->real_pos);
 820                         return ERR_PTR(-EINTR);
 821                 }
 822         }
 823
 824         if (unlikely(btree_node_read_in_flight(b))) {
 825                 six_unlock_type(&b->c.lock, lock_type);
 826                 bch2_trans_unlock(iter->trans);
 827
 828                 wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
 829                                TASK_UNINTERRUPTIBLE);
 830
 831                 /*
 832                  * XXX: check if this always fails - btree_iter_relock()
 833                  * currently fails for iterators that aren't pointed at a valid
 834                  * btree node
 835                  */
 836                 if (iter && !bch2_trans_relock(iter->trans))
 837                         return ERR_PTR(-EINTR);
 838                 goto retry;
 839         }
 840
 841         prefetch(b->aux_data);
 842
 843         for_each_bset(b, t) {
 844                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 845
 846                 prefetch(p + L1_CACHE_BYTES * 0);
 847                 prefetch(p + L1_CACHE_BYTES * 1);
 848                 prefetch(p + L1_CACHE_BYTES * 2);
 849         }
 850
 851         /* avoid atomic set bit if it's not needed: */
 852         if (!btree_node_accessed(b))
 853                 set_btree_node_accessed(b);
 854
 855         if (unlikely(btree_node_read_error(b))) {
 856                 six_unlock_type(&b->c.lock, lock_type);
 857                 return ERR_PTR(-EIO);
 858         }
 859
 860         EBUG_ON(b->c.btree_id != iter->btree_id);
 861         EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
 862         btree_check_header(c, b);
 863
 864         return b;
 865 }
 866
 867 struct btree *bch2_btree_node_get_noiter(struct bch_fs *c,
 868                                          const struct bkey_i *k,
 869                                          enum btree_id btree_id,
 870                                          unsigned level,
 871                                          bool nofill)
 872 {
 873         struct btree_cache *bc = &c->btree_cache;
 874         struct btree *b;
 875         struct bset_tree *t;
 876         int ret;
 877
 878         EBUG_ON(level >= BTREE_MAX_DEPTH);
 879
 880         b = btree_node_mem_ptr(k);
 881         if (b)
 882                 goto lock_node;
 883 retry:
 884         b = btree_cache_find(bc, k);
 885         if (unlikely(!b)) {
 886                 if (nofill)
 887                         goto out;
 888
 889                 b = bch2_btree_node_fill(c, NULL, k, btree_id,
 890                                          level, SIX_LOCK_read, true);
 891
 892                 /* We raced and found the btree node in the cache */
 893                 if (!b)
 894                         goto retry;
 895
 896                 if (IS_ERR(b) &&
 897                     !bch2_btree_cache_cannibalize_lock(c, NULL))
 898                         goto retry;
 899
 900                 if (IS_ERR(b))
 901                         goto out;
 902         } else {
 903 lock_node:
 904                 ret = six_lock_read(&b->c.lock, lock_node_check_fn, (void *) k);
 905                 if (ret)
 906                         goto retry;
 907
 908                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
 909                              b->c.btree_id != btree_id ||
 910                              b->c.level != level)) {
 911                         six_unlock_read(&b->c.lock);
 912                         goto retry;
 913                 }
 914         }
 915
 916         /* XXX: waiting on IO with btree locks held: */
 917         wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
 918                        TASK_UNINTERRUPTIBLE);
 919
 920         prefetch(b->aux_data);
 921
 922         for_each_bset(b, t) {
 923                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 924
 925                 prefetch(p + L1_CACHE_BYTES * 0);
 926                 prefetch(p + L1_CACHE_BYTES * 1);
 927                 prefetch(p + L1_CACHE_BYTES * 2);
 928         }
 929
 930         /* avoid atomic set bit if it's not needed: */
 931         if (!btree_node_accessed(b))
 932                 set_btree_node_accessed(b);
 933
 934         if (unlikely(btree_node_read_error(b))) {
 935                 six_unlock_read(&b->c.lock);
 936                 b = ERR_PTR(-EIO);
 937                 goto out;
 938         }
 939
 940         EBUG_ON(b->c.btree_id != btree_id);
 941         EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
 942         btree_check_header(c, b);
 943 out:
 944         bch2_btree_cache_cannibalize_unlock(c);
 945         return b;
 946 }
 947
 948 void bch2_btree_node_prefetch(struct bch_fs *c, struct btree_iter *iter,
 949                               const struct bkey_i *k,
 950                               enum btree_id btree_id, unsigned level)
 951 {
 952         struct btree_cache *bc = &c->btree_cache;
 953         struct btree *b;
 954
 955         BUG_ON(iter && !btree_node_locked(iter, level + 1));
 956         BUG_ON(level >= BTREE_MAX_DEPTH);
 957
 958         b = btree_cache_find(bc, k);
 959         if (b)
 960                 return;
 961
 962         bch2_btree_node_fill(c, iter, k, btree_id, level, SIX_LOCK_read, false);
 963 }
 964
 965 void bch2_btree_node_evict(struct bch_fs *c, const struct bkey_i *k)
 966 {
 967         struct btree_cache *bc = &c->btree_cache;
 968         struct btree *b;
 969
 970         b = btree_cache_find(bc, k);
 971         if (!b)
 972                 return;
 973
 974         six_lock_intent(&b->c.lock, NULL, NULL);
 975         six_lock_write(&b->c.lock, NULL, NULL);
 976
 977         wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
 978                        TASK_UNINTERRUPTIBLE);
 979         __bch2_btree_node_write(c, b);
 980
 981         /* wait for any in flight btree write */
 982         btree_node_wait_on_io(b);
 983
 984         BUG_ON(btree_node_dirty(b));
 985
 986         mutex_lock(&bc->lock);
 987         btree_node_data_free(c, b);
 988         bch2_btree_node_hash_remove(bc, b);
 989         mutex_unlock(&bc->lock);
 990
 991         six_unlock_write(&b->c.lock);
 992         six_unlock_intent(&b->c.lock);
 993 }
 994
 995 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
 996                              struct btree *b)
 997 {
 998         const struct bkey_format *f = &b->format;
 999         struct bset_stats stats;
1000
1001         memset(&stats, 0, sizeof(stats));
1002
1003         bch2_btree_keys_stats(b, &stats);
1004
1005         pr_buf(out, "l %u ", b->c.level);
1006         bch2_bpos_to_text(out, b->data->min_key);
1007         pr_buf(out, " - ");
1008         bch2_bpos_to_text(out, b->data->max_key);
1009         pr_buf(out, ":\n"
1010                "    ptrs: ");
1011         bch2_val_to_text(out, c, bkey_i_to_s_c(&b->key));
1012
1013         pr_buf(out, "\n"
1014                "    format: u64s %u fields %u %u %u %u %u\n"
1015                "    unpack fn len: %u\n"
1016                "    bytes used %zu/%zu (%zu%% full)\n"
1017                "    sib u64s: %u, %u (merge threshold %u)\n"
1018                "    nr packed keys %u\n"
1019                "    nr unpacked keys %u\n"
1020                "    floats %zu\n"
1021                "    failed unpacked %zu\n",
1022                f->key_u64s,
1023                f->bits_per_field[0],
1024                f->bits_per_field[1],
1025                f->bits_per_field[2],
1026                f->bits_per_field[3],
1027                f->bits_per_field[4],
1028                b->unpack_fn_len,
1029                b->nr.live_u64s * sizeof(u64),
1030                btree_bytes(c) - sizeof(struct btree_node),
1031                b->nr.live_u64s * 100 / btree_max_u64s(c),
1032                b->sib_u64s[0],
1033                b->sib_u64s[1],
1034                c->btree_foreground_merge_threshold,
1035                b->nr.packed_keys,
1036                b->nr.unpacked_keys,
1037                stats.floats,
1038                stats.failed);
1039 }
1040
1041 void bch2_btree_cache_to_text(struct printbuf *out, struct bch_fs *c)
1042 {
1043         pr_buf(out, "nr nodes:\t\t%u\n", c->btree_cache.used);
1044         pr_buf(out, "nr dirty:\t\t%u\n", atomic_read(&c->btree_cache.dirty));
1045         pr_buf(out, "cannibalize lock:\t%p\n", c->btree_cache.alloc_lock);
1046 }