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         bch2_btree_keys_free(b);
  48 }
  49
  50 static void btree_node_data_free(struct bch_fs *c, struct btree *b)
  51 {
  52         struct btree_cache *bc = &c->btree_cache;
  53
  54         __btree_node_data_free(c, b);
  55         bc->used--;
  56         list_move(&b->list, &bc->freed);
  57 }
  58
  59 static int bch2_btree_cache_cmp_fn(struct rhashtable_compare_arg *arg,
  60                                    const void *obj)
  61 {
  62         const struct btree *b = obj;
  63         const u64 *v = arg->key;
  64
  65         return b->hash_val == *v ? 0 : 1;
  66 }
  67
  68 static const struct rhashtable_params bch_btree_cache_params = {
  69         .head_offset    = offsetof(struct btree, hash),
  70         .key_offset     = offsetof(struct btree, hash_val),
  71         .key_len        = sizeof(u64),
  72         .obj_cmpfn      = bch2_btree_cache_cmp_fn,
  73 };
  74
  75 static int __btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
  76 {
  77         BUG_ON(b->data || b->aux_data);
  78
  79         b->data = kvpmalloc(btree_bytes(c), gfp);
  80         if (!b->data)
  81                 return -ENOMEM;
  82
  83         if (bch2_btree_keys_alloc(b, btree_page_order(c), gfp)) {
  84                 kvpfree(b->data, btree_bytes(c));
  85                 b->data = NULL;
  86                 return -ENOMEM;
  87         }
  88
  89         return 0;
  90 }
  91
  92 static void btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
  93 {
  94         struct btree_cache *bc = &c->btree_cache;
  95
  96         if (!__btree_node_data_alloc(c, b, gfp)) {
  97                 bc->used++;
  98                 list_move(&b->list, &bc->freeable);
  99         } else {
 100                 list_move(&b->list, &bc->freed);
 101         }
 102 }
 103
 104 static struct btree *btree_node_mem_alloc(struct bch_fs *c, gfp_t gfp)
 105 {
 106         struct btree *b = kzalloc(sizeof(struct btree), gfp);
 107         if (!b)
 108                 return NULL;
 109
 110         bkey_btree_ptr_init(&b->key);
 111         six_lock_init(&b->c.lock);
 112         INIT_LIST_HEAD(&b->list);
 113         INIT_LIST_HEAD(&b->write_blocked);
 114
 115         btree_node_data_alloc(c, b, gfp);
 116         return b->data ? b : NULL;
 117 }
 118
 119 /* Btree in memory cache - hash table */
 120
 121 void bch2_btree_node_hash_remove(struct btree_cache *bc, struct btree *b)
 122 {
 123         rhashtable_remove_fast(&bc->table, &b->hash, bch_btree_cache_params);
 124
 125         /* Cause future lookups for this node to fail: */
 126         b->hash_val = 0;
 127 }
 128
 129 int __bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b)
 130 {
 131         BUG_ON(b->hash_val);
 132         b->hash_val = btree_ptr_hash_val(&b->key);
 133
 134         return rhashtable_lookup_insert_fast(&bc->table, &b->hash,
 135                                              bch_btree_cache_params);
 136 }
 137
 138 int bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b,
 139                                 unsigned level, enum btree_id id)
 140 {
 141         int ret;
 142
 143         b->c.level      = level;
 144         b->c.btree_id   = id;
 145
 146         mutex_lock(&bc->lock);
 147         ret = __bch2_btree_node_hash_insert(bc, b);
 148         if (!ret)
 149                 list_add(&b->list, &bc->live);
 150         mutex_unlock(&bc->lock);
 151
 152         return ret;
 153 }
 154
 155 __flatten
 156 static inline struct btree *btree_cache_find(struct btree_cache *bc,
 157                                      const struct bkey_i *k)
 158 {
 159         u64 v = btree_ptr_hash_val(k);
 160
 161         return rhashtable_lookup_fast(&bc->table, &v, bch_btree_cache_params);
 162 }
 163
 164 /*
 165  * this version is for btree nodes that have already been freed (we're not
 166  * reaping a real btree node)
 167  */
 168 static int __btree_node_reclaim(struct bch_fs *c, struct btree *b, bool flush)
 169 {
 170         struct btree_cache *bc = &c->btree_cache;
 171         int ret = 0;
 172
 173         lockdep_assert_held(&bc->lock);
 174
 175         if (!six_trylock_intent(&b->c.lock))
 176                 return -ENOMEM;
 177
 178         if (!six_trylock_write(&b->c.lock))
 179                 goto out_unlock_intent;
 180
 181         if (btree_node_noevict(b))
 182                 goto out_unlock;
 183
 184         if (!btree_node_may_write(b))
 185                 goto out_unlock;
 186
 187         if (btree_node_dirty(b) &&
 188             test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags))
 189                 goto out_unlock;
 190
 191         if (btree_node_dirty(b) ||
 192             btree_node_write_in_flight(b) ||
 193             btree_node_read_in_flight(b)) {
 194                 if (!flush)
 195                         goto out_unlock;
 196
 197                 wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
 198                                TASK_UNINTERRUPTIBLE);
 199
 200                 /*
 201                  * Using the underscore version because we don't want to compact
 202                  * bsets after the write, since this node is about to be evicted
 203                  * - unless btree verify mode is enabled, since it runs out of
 204                  * the post write cleanup:
 205                  */
 206                 if (verify_btree_ondisk(c))
 207                         bch2_btree_node_write(c, b, SIX_LOCK_intent);
 208                 else
 209                         __bch2_btree_node_write(c, b, SIX_LOCK_read);
 210
 211                 /* wait for any in flight btree write */
 212                 btree_node_wait_on_io(b);
 213         }
 214 out:
 215         if (b->hash_val && !ret)
 216                 trace_btree_node_reap(c, b);
 217         return ret;
 218 out_unlock:
 219         six_unlock_write(&b->c.lock);
 220 out_unlock_intent:
 221         six_unlock_intent(&b->c.lock);
 222         ret = -ENOMEM;
 223         goto out;
 224 }
 225
 226 static int btree_node_reclaim(struct bch_fs *c, struct btree *b)
 227 {
 228         return __btree_node_reclaim(c, b, false);
 229 }
 230
 231 static int btree_node_write_and_reclaim(struct bch_fs *c, struct btree *b)
 232 {
 233         return __btree_node_reclaim(c, b, true);
 234 }
 235
 236 static unsigned long bch2_btree_cache_scan(struct shrinker *shrink,
 237                                            struct shrink_control *sc)
 238 {
 239         struct bch_fs *c = container_of(shrink, struct bch_fs,
 240                                         btree_cache.shrink);
 241         struct btree_cache *bc = &c->btree_cache;
 242         struct btree *b, *t;
 243         unsigned long nr = sc->nr_to_scan;
 244         unsigned long can_free;
 245         unsigned long touched = 0;
 246         unsigned long freed = 0;
 247         unsigned i;
 248
 249         if (btree_shrinker_disabled(c))
 250                 return SHRINK_STOP;
 251
 252         /* Return -1 if we can't do anything right now */
 253         if (sc->gfp_mask & __GFP_FS)
 254                 mutex_lock(&bc->lock);
 255         else if (!mutex_trylock(&bc->lock))
 256                 return -1;
 257
 258         /*
 259          * It's _really_ critical that we don't free too many btree nodes - we
 260          * have to always leave ourselves a reserve. The reserve is how we
 261          * guarantee that allocating memory for a new btree node can always
 262          * succeed, so that inserting keys into the btree can always succeed and
 263          * IO can always make forward progress:
 264          */
 265         nr /= btree_pages(c);
 266         can_free = btree_cache_can_free(bc);
 267         nr = min_t(unsigned long, nr, can_free);
 268
 269         i = 0;
 270         list_for_each_entry_safe(b, t, &bc->freeable, list) {
 271                 touched++;
 272
 273                 if (freed >= nr)
 274                         break;
 275
 276                 if (++i > 3 &&
 277                     !btree_node_reclaim(c, b)) {
 278                         btree_node_data_free(c, b);
 279                         six_unlock_write(&b->c.lock);
 280                         six_unlock_intent(&b->c.lock);
 281                         freed++;
 282                 }
 283         }
 284 restart:
 285         list_for_each_entry_safe(b, t, &bc->live, list) {
 286                 touched++;
 287
 288                 if (freed >= nr) {
 289                         /* Save position */
 290                         if (&t->list != &bc->live)
 291                                 list_move_tail(&bc->live, &t->list);
 292                         break;
 293                 }
 294
 295                 if (!btree_node_accessed(b) &&
 296                     !btree_node_reclaim(c, b)) {
 297                         /* can't call bch2_btree_node_hash_remove under lock  */
 298                         freed++;
 299                         if (&t->list != &bc->live)
 300                                 list_move_tail(&bc->live, &t->list);
 301
 302                         btree_node_data_free(c, b);
 303                         mutex_unlock(&bc->lock);
 304
 305                         bch2_btree_node_hash_remove(bc, b);
 306                         six_unlock_write(&b->c.lock);
 307                         six_unlock_intent(&b->c.lock);
 308
 309                         if (freed >= nr)
 310                                 goto out;
 311
 312                         if (sc->gfp_mask & __GFP_FS)
 313                                 mutex_lock(&bc->lock);
 314                         else if (!mutex_trylock(&bc->lock))
 315                                 goto out;
 316                         goto restart;
 317                 } else
 318                         clear_btree_node_accessed(b);
 319         }
 320
 321         mutex_unlock(&bc->lock);
 322 out:
 323         return (unsigned long) freed * btree_pages(c);
 324 }
 325
 326 static unsigned long bch2_btree_cache_count(struct shrinker *shrink,
 327                                             struct shrink_control *sc)
 328 {
 329         struct bch_fs *c = container_of(shrink, struct bch_fs,
 330                                         btree_cache.shrink);
 331         struct btree_cache *bc = &c->btree_cache;
 332
 333         if (btree_shrinker_disabled(c))
 334                 return 0;
 335
 336         return btree_cache_can_free(bc) * btree_pages(c);
 337 }
 338
 339 void bch2_fs_btree_cache_exit(struct bch_fs *c)
 340 {
 341         struct btree_cache *bc = &c->btree_cache;
 342         struct btree *b;
 343         unsigned i;
 344
 345         if (bc->shrink.list.next)
 346                 unregister_shrinker(&bc->shrink);
 347
 348         mutex_lock(&bc->lock);
 349
 350 #ifdef CONFIG_BCACHEFS_DEBUG
 351         if (c->verify_data)
 352                 list_move(&c->verify_data->list, &bc->live);
 353
 354         kvpfree(c->verify_ondisk, btree_bytes(c));
 355 #endif
 356
 357         for (i = 0; i < BTREE_ID_NR; i++)
 358                 if (c->btree_roots[i].b)
 359                         list_add(&c->btree_roots[i].b->list, &bc->live);
 360
 361         list_splice(&bc->freeable, &bc->live);
 362
 363         while (!list_empty(&bc->live)) {
 364                 b = list_first_entry(&bc->live, struct btree, list);
 365
 366                 BUG_ON(btree_node_read_in_flight(b) ||
 367                        btree_node_write_in_flight(b));
 368
 369                 if (btree_node_dirty(b))
 370                         bch2_btree_complete_write(c, b, btree_current_write(b));
 371                 clear_btree_node_dirty(b);
 372
 373                 btree_node_data_free(c, b);
 374         }
 375
 376         while (!list_empty(&bc->freed)) {
 377                 b = list_first_entry(&bc->freed, struct btree, list);
 378                 list_del(&b->list);
 379                 kfree(b);
 380         }
 381
 382         mutex_unlock(&bc->lock);
 383
 384         if (bc->table_init_done)
 385                 rhashtable_destroy(&bc->table);
 386 }
 387
 388 int bch2_fs_btree_cache_init(struct bch_fs *c)
 389 {
 390         struct btree_cache *bc = &c->btree_cache;
 391         unsigned i;
 392         int ret = 0;
 393
 394         pr_verbose_init(c->opts, "");
 395
 396         ret = rhashtable_init(&bc->table, &bch_btree_cache_params);
 397         if (ret)
 398                 goto out;
 399
 400         bc->table_init_done = true;
 401
 402         bch2_recalc_btree_reserve(c);
 403
 404         for (i = 0; i < bc->reserve; i++)
 405                 if (!btree_node_mem_alloc(c, GFP_KERNEL)) {
 406                         ret = -ENOMEM;
 407                         goto out;
 408                 }
 409
 410         list_splice_init(&bc->live, &bc->freeable);
 411
 412 #ifdef CONFIG_BCACHEFS_DEBUG
 413         mutex_init(&c->verify_lock);
 414
 415         c->verify_ondisk = kvpmalloc(btree_bytes(c), GFP_KERNEL);
 416         if (!c->verify_ondisk) {
 417                 ret = -ENOMEM;
 418                 goto out;
 419         }
 420
 421         c->verify_data = btree_node_mem_alloc(c, GFP_KERNEL);
 422         if (!c->verify_data) {
 423                 ret = -ENOMEM;
 424                 goto out;
 425         }
 426
 427         list_del_init(&c->verify_data->list);
 428 #endif
 429
 430         bc->shrink.count_objects        = bch2_btree_cache_count;
 431         bc->shrink.scan_objects         = bch2_btree_cache_scan;
 432         bc->shrink.seeks                = 4;
 433         bc->shrink.batch                = btree_pages(c) * 2;
 434         register_shrinker(&bc->shrink);
 435 out:
 436         pr_verbose_init(c->opts, "ret %i", ret);
 437         return ret;
 438 }
 439
 440 void bch2_fs_btree_cache_init_early(struct btree_cache *bc)
 441 {
 442         mutex_init(&bc->lock);
 443         INIT_LIST_HEAD(&bc->live);
 444         INIT_LIST_HEAD(&bc->freeable);
 445         INIT_LIST_HEAD(&bc->freed);
 446 }
 447
 448 /*
 449  * We can only have one thread cannibalizing other cached btree nodes at a time,
 450  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 451  * cannibalize_bucket() will take. This means every time we unlock the root of
 452  * the btree, we need to release this lock if we have it held.
 453  */
 454 void bch2_btree_cache_cannibalize_unlock(struct bch_fs *c)
 455 {
 456         struct btree_cache *bc = &c->btree_cache;
 457
 458         if (bc->alloc_lock == current) {
 459                 trace_btree_node_cannibalize_unlock(c);
 460                 bc->alloc_lock = NULL;
 461                 closure_wake_up(&bc->alloc_wait);
 462         }
 463 }
 464
 465 int bch2_btree_cache_cannibalize_lock(struct bch_fs *c, struct closure *cl)
 466 {
 467         struct btree_cache *bc = &c->btree_cache;
 468         struct task_struct *old;
 469
 470         old = cmpxchg(&bc->alloc_lock, NULL, current);
 471         if (old == NULL || old == current)
 472                 goto success;
 473
 474         if (!cl) {
 475                 trace_btree_node_cannibalize_lock_fail(c);
 476                 return -ENOMEM;
 477         }
 478
 479         closure_wait(&bc->alloc_wait, cl);
 480
 481         /* Try again, after adding ourselves to waitlist */
 482         old = cmpxchg(&bc->alloc_lock, NULL, current);
 483         if (old == NULL || old == current) {
 484                 /* We raced */
 485                 closure_wake_up(&bc->alloc_wait);
 486                 goto success;
 487         }
 488
 489         trace_btree_node_cannibalize_lock_fail(c);
 490         return -EAGAIN;
 491
 492 success:
 493         trace_btree_node_cannibalize_lock(c);
 494         return 0;
 495 }
 496
 497 static struct btree *btree_node_cannibalize(struct bch_fs *c)
 498 {
 499         struct btree_cache *bc = &c->btree_cache;
 500         struct btree *b;
 501
 502         list_for_each_entry_reverse(b, &bc->live, list)
 503                 if (!btree_node_reclaim(c, b))
 504                         return b;
 505
 506         while (1) {
 507                 list_for_each_entry_reverse(b, &bc->live, list)
 508                         if (!btree_node_write_and_reclaim(c, b))
 509                                 return b;
 510
 511                 /*
 512                  * Rare case: all nodes were intent-locked.
 513                  * Just busy-wait.
 514                  */
 515                 WARN_ONCE(1, "btree cache cannibalize failed\n");
 516                 cond_resched();
 517         }
 518 }
 519
 520 struct btree *bch2_btree_node_mem_alloc(struct bch_fs *c)
 521 {
 522         struct btree_cache *bc = &c->btree_cache;
 523         struct btree *b;
 524         u64 start_time = local_clock();
 525         unsigned flags;
 526
 527         flags = memalloc_nofs_save();
 528         mutex_lock(&bc->lock);
 529
 530         /*
 531          * btree_free() doesn't free memory; it sticks the node on the end of
 532          * the list. Check if there's any freed nodes there:
 533          */
 534         list_for_each_entry(b, &bc->freeable, list)
 535                 if (!btree_node_reclaim(c, b))
 536                         goto got_node;
 537
 538         /*
 539          * We never free struct btree itself, just the memory that holds the on
 540          * disk node. Check the freed list before allocating a new one:
 541          */
 542         list_for_each_entry(b, &bc->freed, list)
 543                 if (!btree_node_reclaim(c, b))
 544                         goto got_node;
 545
 546         b = NULL;
 547 got_node:
 548         if (b)
 549                 list_del_init(&b->list);
 550         mutex_unlock(&bc->lock);
 551
 552         if (!b) {
 553                 b = kzalloc(sizeof(struct btree), GFP_KERNEL);
 554                 if (!b)
 555                         goto err;
 556
 557                 bkey_btree_ptr_init(&b->key);
 558                 six_lock_init(&b->c.lock);
 559                 INIT_LIST_HEAD(&b->list);
 560                 INIT_LIST_HEAD(&b->write_blocked);
 561
 562                 BUG_ON(!six_trylock_intent(&b->c.lock));
 563                 BUG_ON(!six_trylock_write(&b->c.lock));
 564         }
 565
 566         if (!b->data) {
 567                 if (__btree_node_data_alloc(c, b, __GFP_NOWARN|GFP_KERNEL))
 568                         goto err;
 569
 570                 mutex_lock(&bc->lock);
 571                 bc->used++;
 572                 mutex_unlock(&bc->lock);
 573         }
 574
 575         BUG_ON(btree_node_hashed(b));
 576         BUG_ON(btree_node_write_in_flight(b));
 577 out:
 578         b->flags                = 0;
 579         b->written              = 0;
 580         b->nsets                = 0;
 581         b->sib_u64s[0]          = 0;
 582         b->sib_u64s[1]          = 0;
 583         b->whiteout_u64s        = 0;
 584         bch2_btree_keys_init(b, &c->expensive_debug_checks);
 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         /*
 658          * Unlock before doing IO:
 659          *
 660          * XXX: ideally should be dropping all btree node locks here
 661          */
 662         if (iter && btree_node_read_locked(iter, level + 1))
 663                 btree_node_unlock(iter, level + 1);
 664
 665         bch2_btree_node_read(c, b, sync);
 666
 667         six_unlock_write(&b->c.lock);
 668
 669         if (!sync) {
 670                 six_unlock_intent(&b->c.lock);
 671                 return NULL;
 672         }
 673
 674         if (lock_type == SIX_LOCK_read)
 675                 six_lock_downgrade(&b->c.lock);
 676
 677         return b;
 678 }
 679
 680 static int lock_node_check_fn(struct six_lock *lock, void *p)
 681 {
 682         struct btree *b = container_of(lock, struct btree, c.lock);
 683         const struct bkey_i *k = p;
 684
 685         return b->hash_val == btree_ptr_hash_val(k) ? 0 : -1;
 686 }
 687
 688 /**
 689  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 690  * in from disk if necessary.
 691  *
 692  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
 693  *
 694  * The btree node will have either a read or a write lock held, depending on
 695  * the @write parameter.
 696  */
 697 struct btree *bch2_btree_node_get(struct bch_fs *c, struct btree_iter *iter,
 698                                   const struct bkey_i *k, unsigned level,
 699                                   enum six_lock_type lock_type)
 700 {
 701         struct btree_cache *bc = &c->btree_cache;
 702         struct btree *b;
 703         struct bset_tree *t;
 704
 705         EBUG_ON(level >= BTREE_MAX_DEPTH);
 706
 707         b = btree_node_mem_ptr(k);
 708         if (b)
 709                 goto lock_node;
 710 retry:
 711         b = btree_cache_find(bc, k);
 712         if (unlikely(!b)) {
 713                 /*
 714                  * We must have the parent locked to call bch2_btree_node_fill(),
 715                  * else we could read in a btree node from disk that's been
 716                  * freed:
 717                  */
 718                 b = bch2_btree_node_fill(c, iter, k, iter->btree_id,
 719                                          level, lock_type, true);
 720
 721                 /* We raced and found the btree node in the cache */
 722                 if (!b)
 723                         goto retry;
 724
 725                 if (IS_ERR(b))
 726                         return b;
 727         } else {
 728 lock_node:
 729                 /*
 730                  * There's a potential deadlock with splits and insertions into
 731                  * interior nodes we have to avoid:
 732                  *
 733                  * The other thread might be holding an intent lock on the node
 734                  * we want, and they want to update its parent node so they're
 735                  * going to upgrade their intent lock on the parent node to a
 736                  * write lock.
 737                  *
 738                  * But if we're holding a read lock on the parent, and we're
 739                  * trying to get the intent lock they're holding, we deadlock.
 740                  *
 741                  * So to avoid this we drop the read locks on parent nodes when
 742                  * we're starting to take intent locks - and handle the race.
 743                  *
 744                  * The race is that they might be about to free the node we
 745                  * want, and dropping our read lock on the parent node lets them
 746                  * update the parent marking the node we want as freed, and then
 747                  * free it:
 748                  *
 749                  * To guard against this, btree nodes are evicted from the cache
 750                  * when they're freed - and b->hash_val is zeroed out, which we
 751                  * check for after we lock the node.
 752                  *
 753                  * Then, bch2_btree_node_relock() on the parent will fail - because
 754                  * the parent was modified, when the pointer to the node we want
 755                  * was removed - and we'll bail out:
 756                  */
 757                 if (btree_node_read_locked(iter, level + 1))
 758                         btree_node_unlock(iter, level + 1);
 759
 760                 if (!btree_node_lock(b, k->k.p, level, iter, lock_type,
 761                                      lock_node_check_fn, (void *) k)) {
 762                         if (b->hash_val != btree_ptr_hash_val(k))
 763                                 goto retry;
 764                         return ERR_PTR(-EINTR);
 765                 }
 766
 767                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
 768                              b->c.level != level ||
 769                              race_fault())) {
 770                         six_unlock_type(&b->c.lock, lock_type);
 771                         if (bch2_btree_node_relock(iter, level + 1))
 772                                 goto retry;
 773
 774                         trace_trans_restart_btree_node_reused(iter->trans->ip);
 775                         return ERR_PTR(-EINTR);
 776                 }
 777         }
 778
 779         /* XXX: waiting on IO with btree locks held: */
 780         wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
 781                        TASK_UNINTERRUPTIBLE);
 782
 783         prefetch(b->aux_data);
 784
 785         for_each_bset(b, t) {
 786                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 787
 788                 prefetch(p + L1_CACHE_BYTES * 0);
 789                 prefetch(p + L1_CACHE_BYTES * 1);
 790                 prefetch(p + L1_CACHE_BYTES * 2);
 791         }
 792
 793         /* avoid atomic set bit if it's not needed: */
 794         if (!btree_node_accessed(b))
 795                 set_btree_node_accessed(b);
 796
 797         if (unlikely(btree_node_read_error(b))) {
 798                 six_unlock_type(&b->c.lock, lock_type);
 799                 return ERR_PTR(-EIO);
 800         }
 801
 802         EBUG_ON(b->c.btree_id != iter->btree_id ||
 803                 BTREE_NODE_LEVEL(b->data) != level ||
 804                 bkey_cmp(b->data->max_key, k->k.p));
 805
 806         return b;
 807 }
 808
 809 struct btree *bch2_btree_node_get_noiter(struct bch_fs *c,
 810                                          const struct bkey_i *k,
 811                                          enum btree_id btree_id,
 812                                          unsigned level)
 813 {
 814         struct btree_cache *bc = &c->btree_cache;
 815         struct btree *b;
 816         struct bset_tree *t;
 817         int ret;
 818
 819         EBUG_ON(level >= BTREE_MAX_DEPTH);
 820
 821         b = btree_node_mem_ptr(k);
 822         if (b)
 823                 goto lock_node;
 824 retry:
 825         b = btree_cache_find(bc, k);
 826         if (unlikely(!b)) {
 827                 b = bch2_btree_node_fill(c, NULL, k, btree_id,
 828                                          level, SIX_LOCK_read, true);
 829
 830                 /* We raced and found the btree node in the cache */
 831                 if (!b)
 832                         goto retry;
 833
 834                 if (IS_ERR(b))
 835                         return b;
 836         } else {
 837 lock_node:
 838                 ret = six_lock_read(&b->c.lock, lock_node_check_fn, (void *) k);
 839                 if (ret)
 840                         goto retry;
 841
 842                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
 843                              b->c.btree_id != btree_id ||
 844                              b->c.level != level)) {
 845                         six_unlock_read(&b->c.lock);
 846                         goto retry;
 847                 }
 848         }
 849
 850         /* XXX: waiting on IO with btree locks held: */
 851         wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
 852                        TASK_UNINTERRUPTIBLE);
 853
 854         prefetch(b->aux_data);
 855
 856         for_each_bset(b, t) {
 857                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 858
 859                 prefetch(p + L1_CACHE_BYTES * 0);
 860                 prefetch(p + L1_CACHE_BYTES * 1);
 861                 prefetch(p + L1_CACHE_BYTES * 2);
 862         }
 863
 864         /* avoid atomic set bit if it's not needed: */
 865         if (!btree_node_accessed(b))
 866                 set_btree_node_accessed(b);
 867
 868         if (unlikely(btree_node_read_error(b))) {
 869                 six_unlock_read(&b->c.lock);
 870                 return ERR_PTR(-EIO);
 871         }
 872
 873         EBUG_ON(b->c.btree_id != btree_id ||
 874                 BTREE_NODE_LEVEL(b->data) != level ||
 875                 bkey_cmp(b->data->max_key, k->k.p));
 876
 877         return b;
 878 }
 879
 880 struct btree *bch2_btree_node_get_sibling(struct bch_fs *c,
 881                                           struct btree_iter *iter,
 882                                           struct btree *b,
 883                                           enum btree_node_sibling sib)
 884 {
 885         struct btree_trans *trans = iter->trans;
 886         struct btree *parent;
 887         struct btree_node_iter node_iter;
 888         struct bkey_packed *k;
 889         BKEY_PADDED(k) tmp;
 890         struct btree *ret = NULL;
 891         unsigned level = b->c.level;
 892
 893         parent = btree_iter_node(iter, level + 1);
 894         if (!parent)
 895                 return NULL;
 896
 897         /*
 898          * There's a corner case where a btree_iter might have a node locked
 899          * that is just outside its current pos - when
 900          * bch2_btree_iter_set_pos_same_leaf() gets to the end of the node.
 901          *
 902          * But the lock ordering checks in __bch2_btree_node_lock() go off of
 903          * iter->pos, not the node's key: so if the iterator is marked as
 904          * needing to be traversed, we risk deadlock if we don't bail out here:
 905          */
 906         if (iter->uptodate >= BTREE_ITER_NEED_TRAVERSE)
 907                 return ERR_PTR(-EINTR);
 908
 909         if (!bch2_btree_node_relock(iter, level + 1)) {
 910                 ret = ERR_PTR(-EINTR);
 911                 goto out;
 912         }
 913
 914         node_iter = iter->l[parent->c.level].iter;
 915
 916         k = bch2_btree_node_iter_peek_all(&node_iter, parent);
 917         BUG_ON(bkey_cmp_left_packed(parent, k, &b->key.k.p));
 918
 919         k = sib == btree_prev_sib
 920                 ? bch2_btree_node_iter_prev(&node_iter, parent)
 921                 : (bch2_btree_node_iter_advance(&node_iter, parent),
 922                    bch2_btree_node_iter_peek(&node_iter, parent));
 923         if (!k)
 924                 goto out;
 925
 926         bch2_bkey_unpack(parent, &tmp.k, k);
 927
 928         ret = bch2_btree_node_get(c, iter, &tmp.k, level,
 929                                   SIX_LOCK_intent);
 930
 931         if (PTR_ERR_OR_ZERO(ret) == -EINTR && !trans->nounlock) {
 932                 struct btree_iter *linked;
 933
 934                 if (!bch2_btree_node_relock(iter, level + 1))
 935                         goto out;
 936
 937                 /*
 938                  * We might have got -EINTR because trylock failed, and we're
 939                  * holding other locks that would cause us to deadlock:
 940                  */
 941                 trans_for_each_iter(trans, linked)
 942                         if (btree_iter_cmp(iter, linked) < 0)
 943                                 __bch2_btree_iter_unlock(linked);
 944
 945                 if (sib == btree_prev_sib)
 946                         btree_node_unlock(iter, level);
 947
 948                 ret = bch2_btree_node_get(c, iter, &tmp.k, level,
 949                                           SIX_LOCK_intent);
 950
 951                 /*
 952                  * before btree_iter_relock() calls btree_iter_verify_locks():
 953                  */
 954                 if (btree_lock_want(iter, level + 1) == BTREE_NODE_UNLOCKED)
 955                         btree_node_unlock(iter, level + 1);
 956
 957                 if (!bch2_btree_node_relock(iter, level)) {
 958                         btree_iter_set_dirty(iter, BTREE_ITER_NEED_RELOCK);
 959
 960                         if (!IS_ERR(ret)) {
 961                                 six_unlock_intent(&ret->c.lock);
 962                                 ret = ERR_PTR(-EINTR);
 963                         }
 964                 }
 965
 966                 bch2_trans_relock(trans);
 967         }
 968 out:
 969         if (btree_lock_want(iter, level + 1) == BTREE_NODE_UNLOCKED)
 970                 btree_node_unlock(iter, level + 1);
 971
 972         if (PTR_ERR_OR_ZERO(ret) == -EINTR)
 973                 bch2_btree_iter_upgrade(iter, level + 2);
 974
 975         BUG_ON(!IS_ERR(ret) && !btree_node_locked(iter, level));
 976
 977         if (!IS_ERR_OR_NULL(ret)) {
 978                 struct btree *n1 = ret, *n2 = b;
 979
 980                 if (sib != btree_prev_sib)
 981                         swap(n1, n2);
 982
 983                 BUG_ON(bkey_cmp(bkey_successor(n1->key.k.p),
 984                                 n2->data->min_key));
 985         }
 986
 987         bch2_btree_trans_verify_locks(trans);
 988
 989         return ret;
 990 }
 991
 992 void bch2_btree_node_prefetch(struct bch_fs *c, struct btree_iter *iter,
 993                               const struct bkey_i *k, unsigned level)
 994 {
 995         struct btree_cache *bc = &c->btree_cache;
 996         struct btree *b;
 997
 998         BUG_ON(!btree_node_locked(iter, level + 1));
 999         BUG_ON(level >= BTREE_MAX_DEPTH);
1000
1001         b = btree_cache_find(bc, k);
1002         if (b)
1003                 return;
1004
1005         bch2_btree_node_fill(c, iter, k, iter->btree_id,
1006                              level, SIX_LOCK_read, false);
1007 }
1008
1009 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
1010                              struct btree *b)
1011 {
1012         const struct bkey_format *f = &b->format;
1013         struct bset_stats stats;
1014
1015         memset(&stats, 0, sizeof(stats));
1016
1017         bch2_btree_keys_stats(b, &stats);
1018
1019         pr_buf(out,
1020                "l %u %llu:%llu - %llu:%llu:\n"
1021                "    ptrs: ",
1022                b->c.level,
1023                b->data->min_key.inode,
1024                b->data->min_key.offset,
1025                b->data->max_key.inode,
1026                b->data->max_key.offset);
1027         bch2_val_to_text(out, c, bkey_i_to_s_c(&b->key));
1028         pr_buf(out, "\n"
1029                "    format: u64s %u fields %u %u %u %u %u\n"
1030                "    unpack fn len: %u\n"
1031                "    bytes used %zu/%zu (%zu%% full)\n"
1032                "    sib u64s: %u, %u (merge threshold %zu)\n"
1033                "    nr packed keys %u\n"
1034                "    nr unpacked keys %u\n"
1035                "    floats %zu\n"
1036                "    failed unpacked %zu\n",
1037                f->key_u64s,
1038                f->bits_per_field[0],
1039                f->bits_per_field[1],
1040                f->bits_per_field[2],
1041                f->bits_per_field[3],
1042                f->bits_per_field[4],
1043                b->unpack_fn_len,
1044                b->nr.live_u64s * sizeof(u64),
1045                btree_bytes(c) - sizeof(struct btree_node),
1046                b->nr.live_u64s * 100 / btree_max_u64s(c),
1047                b->sib_u64s[0],
1048                b->sib_u64s[1],
1049                BTREE_FOREGROUND_MERGE_THRESHOLD(c),
1050                b->nr.packed_keys,
1051                b->nr.unpacked_keys,
1052                stats.floats,
1053                stats.failed);
1054 }