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         EBUG_ON(btree_node_write_in_flight(b));
  39
  40         kvpfree(b->data, btree_bytes(c));
  41         b->data = NULL;
  42         vfree(b->aux_data);
  43         b->aux_data = NULL;
  44 }
  45
  46 static void btree_node_data_free(struct bch_fs *c, struct btree *b)
  47 {
  48         struct btree_cache *bc = &c->btree_cache;
  49
  50         __btree_node_data_free(c, b);
  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
  79         b->aux_data = vmalloc_exec(btree_aux_data_bytes(b), gfp);
  80         if (!b->aux_data) {
  81                 kvpfree(b->data, btree_bytes(c));
  82                 b->data = NULL;
  83                 return -ENOMEM;
  84         }
  85
  86         return 0;
  87 }
  88
  89 static struct btree *__btree_node_mem_alloc(struct bch_fs *c)
  90 {
  91         struct btree *b = kzalloc(sizeof(struct btree), GFP_KERNEL);
  92         if (!b)
  93                 return NULL;
  94
  95         bkey_btree_ptr_init(&b->key);
  96         six_lock_init(&b->c.lock);
  97         INIT_LIST_HEAD(&b->list);
  98         INIT_LIST_HEAD(&b->write_blocked);
  99         b->byte_order = ilog2(btree_bytes(c));
 100         return b;
 101 }
 102
 103 static struct btree *btree_node_mem_alloc(struct bch_fs *c)
 104 {
 105         struct btree_cache *bc = &c->btree_cache;
 106         struct btree *b = __btree_node_mem_alloc(c);
 107         if (!b)
 108                 return NULL;
 109
 110         if (btree_node_data_alloc(c, b, GFP_KERNEL)) {
 111                 kfree(b);
 112                 return NULL;
 113         }
 114
 115         bc->used++;
 116         list_add(&b->list, &bc->freeable);
 117         return b;
 118 }
 119
 120 /* Btree in memory cache - hash table */
 121
 122 void bch2_btree_node_hash_remove(struct btree_cache *bc, struct btree *b)
 123 {
 124         rhashtable_remove_fast(&bc->table, &b->hash, bch_btree_cache_params);
 125
 126         /* Cause future lookups for this node to fail: */
 127         b->hash_val = 0;
 128
 129         six_lock_wakeup_all(&b->c.lock);
 130 }
 131
 132 int __bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b)
 133 {
 134         BUG_ON(b->hash_val);
 135         b->hash_val = btree_ptr_hash_val(&b->key);
 136
 137         return rhashtable_lookup_insert_fast(&bc->table, &b->hash,
 138                                              bch_btree_cache_params);
 139 }
 140
 141 int bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b,
 142                                 unsigned level, enum btree_id id)
 143 {
 144         int ret;
 145
 146         b->c.level      = level;
 147         b->c.btree_id   = id;
 148
 149         mutex_lock(&bc->lock);
 150         ret = __bch2_btree_node_hash_insert(bc, b);
 151         if (!ret)
 152                 list_add(&b->list, &bc->live);
 153         mutex_unlock(&bc->lock);
 154
 155         return ret;
 156 }
 157
 158 __flatten
 159 static inline struct btree *btree_cache_find(struct btree_cache *bc,
 160                                      const struct bkey_i *k)
 161 {
 162         u64 v = btree_ptr_hash_val(k);
 163
 164         return rhashtable_lookup_fast(&bc->table, &v, bch_btree_cache_params);
 165 }
 166
 167 /*
 168  * this version is for btree nodes that have already been freed (we're not
 169  * reaping a real btree node)
 170  */
 171 static int __btree_node_reclaim(struct bch_fs *c, struct btree *b, bool flush)
 172 {
 173         struct btree_cache *bc = &c->btree_cache;
 174         int ret = 0;
 175
 176         lockdep_assert_held(&bc->lock);
 177
 178         if (!six_trylock_intent(&b->c.lock))
 179                 return -ENOMEM;
 180
 181         if (!six_trylock_write(&b->c.lock))
 182                 goto out_unlock_intent;
 183
 184         if (btree_node_noevict(b))
 185                 goto out_unlock;
 186
 187         if (!btree_node_may_write(b))
 188                 goto out_unlock;
 189
 190         if (btree_node_dirty(b) &&
 191             test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags))
 192                 goto out_unlock;
 193
 194         if (btree_node_dirty(b) ||
 195             btree_node_write_in_flight(b) ||
 196             btree_node_read_in_flight(b)) {
 197                 if (!flush)
 198                         goto out_unlock;
 199
 200                 wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
 201                                TASK_UNINTERRUPTIBLE);
 202
 203                 /*
 204                  * Using the underscore version because we don't want to compact
 205                  * bsets after the write, since this node is about to be evicted
 206                  * - unless btree verify mode is enabled, since it runs out of
 207                  * the post write cleanup:
 208                  */
 209                 if (bch2_verify_btree_ondisk)
 210                         bch2_btree_node_write(c, b, SIX_LOCK_intent);
 211                 else
 212                         __bch2_btree_node_write(c, b, SIX_LOCK_read);
 213
 214                 /* wait for any in flight btree write */
 215                 btree_node_wait_on_io(b);
 216         }
 217 out:
 218         if (b->hash_val && !ret)
 219                 trace_btree_node_reap(c, b);
 220         return ret;
 221 out_unlock:
 222         six_unlock_write(&b->c.lock);
 223 out_unlock_intent:
 224         six_unlock_intent(&b->c.lock);
 225         ret = -ENOMEM;
 226         goto out;
 227 }
 228
 229 static int btree_node_reclaim(struct bch_fs *c, struct btree *b)
 230 {
 231         return __btree_node_reclaim(c, b, false);
 232 }
 233
 234 static int btree_node_write_and_reclaim(struct bch_fs *c, struct btree *b)
 235 {
 236         return __btree_node_reclaim(c, b, true);
 237 }
 238
 239 static unsigned long bch2_btree_cache_scan(struct shrinker *shrink,
 240                                            struct shrink_control *sc)
 241 {
 242         struct bch_fs *c = container_of(shrink, struct bch_fs,
 243                                         btree_cache.shrink);
 244         struct btree_cache *bc = &c->btree_cache;
 245         struct btree *b, *t;
 246         unsigned long nr = sc->nr_to_scan;
 247         unsigned long can_free;
 248         unsigned long touched = 0;
 249         unsigned long freed = 0;
 250         unsigned i, flags;
 251
 252         if (bch2_btree_shrinker_disabled)
 253                 return SHRINK_STOP;
 254
 255         /* Return -1 if we can't do anything right now */
 256         if (sc->gfp_mask & __GFP_FS)
 257                 mutex_lock(&bc->lock);
 258         else if (!mutex_trylock(&bc->lock))
 259                 return -1;
 260
 261         flags = memalloc_nofs_save();
 262
 263         /*
 264          * It's _really_ critical that we don't free too many btree nodes - we
 265          * have to always leave ourselves a reserve. The reserve is how we
 266          * guarantee that allocating memory for a new btree node can always
 267          * succeed, so that inserting keys into the btree can always succeed and
 268          * IO can always make forward progress:
 269          */
 270         nr /= btree_pages(c);
 271         can_free = btree_cache_can_free(bc);
 272         nr = min_t(unsigned long, nr, can_free);
 273
 274         i = 0;
 275         list_for_each_entry_safe(b, t, &bc->freeable, list) {
 276                 touched++;
 277
 278                 if (freed >= nr)
 279                         break;
 280
 281                 if (++i > 3 &&
 282                     !btree_node_reclaim(c, b)) {
 283                         btree_node_data_free(c, b);
 284                         six_unlock_write(&b->c.lock);
 285                         six_unlock_intent(&b->c.lock);
 286                         freed++;
 287                 }
 288         }
 289 restart:
 290         list_for_each_entry_safe(b, t, &bc->live, list) {
 291                 touched++;
 292
 293                 if (freed >= nr) {
 294                         /* Save position */
 295                         if (&t->list != &bc->live)
 296                                 list_move_tail(&bc->live, &t->list);
 297                         break;
 298                 }
 299
 300                 if (!btree_node_accessed(b) &&
 301                     !btree_node_reclaim(c, b)) {
 302                         /* can't call bch2_btree_node_hash_remove under lock  */
 303                         freed++;
 304                         if (&t->list != &bc->live)
 305                                 list_move_tail(&bc->live, &t->list);
 306
 307                         btree_node_data_free(c, b);
 308                         mutex_unlock(&bc->lock);
 309
 310                         bch2_btree_node_hash_remove(bc, b);
 311                         six_unlock_write(&b->c.lock);
 312                         six_unlock_intent(&b->c.lock);
 313
 314                         if (freed >= nr)
 315                                 goto out;
 316
 317                         if (sc->gfp_mask & __GFP_FS)
 318                                 mutex_lock(&bc->lock);
 319                         else if (!mutex_trylock(&bc->lock))
 320                                 goto out;
 321                         goto restart;
 322                 } else
 323                         clear_btree_node_accessed(b);
 324         }
 325
 326         mutex_unlock(&bc->lock);
 327 out:
 328         memalloc_nofs_restore(flags);
 329         return (unsigned long) freed * btree_pages(c);
 330 }
 331
 332 static unsigned long bch2_btree_cache_count(struct shrinker *shrink,
 333                                             struct shrink_control *sc)
 334 {
 335         struct bch_fs *c = container_of(shrink, struct bch_fs,
 336                                         btree_cache.shrink);
 337         struct btree_cache *bc = &c->btree_cache;
 338
 339         if (bch2_btree_shrinker_disabled)
 340                 return 0;
 341
 342         return btree_cache_can_free(bc) * btree_pages(c);
 343 }
 344
 345 void bch2_fs_btree_cache_exit(struct bch_fs *c)
 346 {
 347         struct btree_cache *bc = &c->btree_cache;
 348         struct btree *b;
 349         unsigned i, flags;
 350
 351         if (bc->shrink.list.next)
 352                 unregister_shrinker(&bc->shrink);
 353
 354         /* vfree() can allocate memory: */
 355         flags = memalloc_nofs_save();
 356         mutex_lock(&bc->lock);
 357
 358 #ifdef CONFIG_BCACHEFS_DEBUG
 359         if (c->verify_data)
 360                 list_move(&c->verify_data->list, &bc->live);
 361
 362         kvpfree(c->verify_ondisk, btree_bytes(c));
 363 #endif
 364
 365         for (i = 0; i < BTREE_ID_NR; i++)
 366                 if (c->btree_roots[i].b)
 367                         list_add(&c->btree_roots[i].b->list, &bc->live);
 368
 369         list_splice(&bc->freeable, &bc->live);
 370
 371         while (!list_empty(&bc->live)) {
 372                 b = list_first_entry(&bc->live, struct btree, list);
 373
 374                 BUG_ON(btree_node_read_in_flight(b) ||
 375                        btree_node_write_in_flight(b));
 376
 377                 if (btree_node_dirty(b))
 378                         bch2_btree_complete_write(c, b, btree_current_write(b));
 379                 clear_btree_node_dirty(c, b);
 380
 381                 btree_node_data_free(c, b);
 382         }
 383
 384         BUG_ON(atomic_read(&c->btree_cache.dirty));
 385
 386         while (!list_empty(&bc->freed)) {
 387                 b = list_first_entry(&bc->freed, struct btree, list);
 388                 list_del(&b->list);
 389                 kfree(b);
 390         }
 391
 392         mutex_unlock(&bc->lock);
 393         memalloc_nofs_restore(flags);
 394
 395         if (bc->table_init_done)
 396                 rhashtable_destroy(&bc->table);
 397 }
 398
 399 int bch2_fs_btree_cache_init(struct bch_fs *c)
 400 {
 401         struct btree_cache *bc = &c->btree_cache;
 402         unsigned i;
 403         int ret = 0;
 404
 405         pr_verbose_init(c->opts, "");
 406
 407         ret = rhashtable_init(&bc->table, &bch_btree_cache_params);
 408         if (ret)
 409                 goto out;
 410
 411         bc->table_init_done = true;
 412
 413         bch2_recalc_btree_reserve(c);
 414
 415         for (i = 0; i < bc->reserve; i++)
 416                 if (!btree_node_mem_alloc(c)) {
 417                         ret = -ENOMEM;
 418                         goto out;
 419                 }
 420
 421         list_splice_init(&bc->live, &bc->freeable);
 422
 423 #ifdef CONFIG_BCACHEFS_DEBUG
 424         mutex_init(&c->verify_lock);
 425
 426         c->verify_ondisk = kvpmalloc(btree_bytes(c), GFP_KERNEL);
 427         if (!c->verify_ondisk) {
 428                 ret = -ENOMEM;
 429                 goto out;
 430         }
 431
 432         c->verify_data = btree_node_mem_alloc(c);
 433         if (!c->verify_data) {
 434                 ret = -ENOMEM;
 435                 goto out;
 436         }
 437
 438         list_del_init(&c->verify_data->list);
 439 #endif
 440
 441         bc->shrink.count_objects        = bch2_btree_cache_count;
 442         bc->shrink.scan_objects         = bch2_btree_cache_scan;
 443         bc->shrink.seeks                = 4;
 444         bc->shrink.batch                = btree_pages(c) * 2;
 445         ret = register_shrinker(&bc->shrink);
 446 out:
 447         pr_verbose_init(c->opts, "ret %i", ret);
 448         return ret;
 449 }
 450
 451 void bch2_fs_btree_cache_init_early(struct btree_cache *bc)
 452 {
 453         mutex_init(&bc->lock);
 454         INIT_LIST_HEAD(&bc->live);
 455         INIT_LIST_HEAD(&bc->freeable);
 456         INIT_LIST_HEAD(&bc->freed);
 457 }
 458
 459 /*
 460  * We can only have one thread cannibalizing other cached btree nodes at a time,
 461  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 462  * cannibalize_bucket() will take. This means every time we unlock the root of
 463  * the btree, we need to release this lock if we have it held.
 464  */
 465 void bch2_btree_cache_cannibalize_unlock(struct bch_fs *c)
 466 {
 467         struct btree_cache *bc = &c->btree_cache;
 468
 469         if (bc->alloc_lock == current) {
 470                 trace_btree_node_cannibalize_unlock(c);
 471                 bc->alloc_lock = NULL;
 472                 closure_wake_up(&bc->alloc_wait);
 473         }
 474 }
 475
 476 int bch2_btree_cache_cannibalize_lock(struct bch_fs *c, struct closure *cl)
 477 {
 478         struct btree_cache *bc = &c->btree_cache;
 479         struct task_struct *old;
 480
 481         old = cmpxchg(&bc->alloc_lock, NULL, current);
 482         if (old == NULL || old == current)
 483                 goto success;
 484
 485         if (!cl) {
 486                 trace_btree_node_cannibalize_lock_fail(c);
 487                 return -ENOMEM;
 488         }
 489
 490         closure_wait(&bc->alloc_wait, cl);
 491
 492         /* Try again, after adding ourselves to waitlist */
 493         old = cmpxchg(&bc->alloc_lock, NULL, current);
 494         if (old == NULL || old == current) {
 495                 /* We raced */
 496                 closure_wake_up(&bc->alloc_wait);
 497                 goto success;
 498         }
 499
 500         trace_btree_node_cannibalize_lock_fail(c);
 501         return -EAGAIN;
 502
 503 success:
 504         trace_btree_node_cannibalize_lock(c);
 505         return 0;
 506 }
 507
 508 static struct btree *btree_node_cannibalize(struct bch_fs *c)
 509 {
 510         struct btree_cache *bc = &c->btree_cache;
 511         struct btree *b;
 512
 513         list_for_each_entry_reverse(b, &bc->live, list)
 514                 if (!btree_node_reclaim(c, b))
 515                         return b;
 516
 517         while (1) {
 518                 list_for_each_entry_reverse(b, &bc->live, list)
 519                         if (!btree_node_write_and_reclaim(c, b))
 520                                 return b;
 521
 522                 /*
 523                  * Rare case: all nodes were intent-locked.
 524                  * Just busy-wait.
 525                  */
 526                 WARN_ONCE(1, "btree cache cannibalize failed\n");
 527                 cond_resched();
 528         }
 529 }
 530
 531 struct btree *bch2_btree_node_mem_alloc(struct bch_fs *c)
 532 {
 533         struct btree_cache *bc = &c->btree_cache;
 534         struct btree *b;
 535         u64 start_time = local_clock();
 536         unsigned flags;
 537
 538         flags = memalloc_nofs_save();
 539         mutex_lock(&bc->lock);
 540
 541         /*
 542          * btree_free() doesn't free memory; it sticks the node on the end of
 543          * the list. Check if there's any freed nodes there:
 544          */
 545         list_for_each_entry(b, &bc->freeable, list)
 546                 if (!btree_node_reclaim(c, b))
 547                         goto got_node;
 548
 549         /*
 550          * We never free struct btree itself, just the memory that holds the on
 551          * disk node. Check the freed list before allocating a new one:
 552          */
 553         list_for_each_entry(b, &bc->freed, list)
 554                 if (!btree_node_reclaim(c, b))
 555                         goto got_node;
 556
 557         b = NULL;
 558 got_node:
 559         if (b)
 560                 list_del_init(&b->list);
 561         mutex_unlock(&bc->lock);
 562
 563         if (!b) {
 564                 b = __btree_node_mem_alloc(c);
 565                 if (!b)
 566                         goto err;
 567
 568                 BUG_ON(!six_trylock_intent(&b->c.lock));
 569                 BUG_ON(!six_trylock_write(&b->c.lock));
 570         }
 571
 572         if (!b->data) {
 573                 if (btree_node_data_alloc(c, b, __GFP_NOWARN|GFP_KERNEL))
 574                         goto err;
 575
 576                 mutex_lock(&bc->lock);
 577                 bc->used++;
 578                 mutex_unlock(&bc->lock);
 579         }
 580
 581         BUG_ON(btree_node_hashed(b));
 582         BUG_ON(btree_node_write_in_flight(b));
 583 out:
 584         b->flags                = 0;
 585         b->written              = 0;
 586         b->nsets                = 0;
 587         b->sib_u64s[0]          = 0;
 588         b->sib_u64s[1]          = 0;
 589         b->whiteout_u64s        = 0;
 590         bch2_btree_keys_init(b);
 591
 592         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],
 593                                start_time);
 594
 595         memalloc_nofs_restore(flags);
 596         return b;
 597 err:
 598         mutex_lock(&bc->lock);
 599
 600         if (b) {
 601                 list_add(&b->list, &bc->freed);
 602                 six_unlock_write(&b->c.lock);
 603                 six_unlock_intent(&b->c.lock);
 604         }
 605
 606         /* Try to cannibalize another cached btree node: */
 607         if (bc->alloc_lock == current) {
 608                 b = btree_node_cannibalize(c);
 609                 list_del_init(&b->list);
 610                 mutex_unlock(&bc->lock);
 611
 612                 bch2_btree_node_hash_remove(bc, b);
 613
 614                 trace_btree_node_cannibalize(c);
 615                 goto out;
 616         }
 617
 618         mutex_unlock(&bc->lock);
 619         memalloc_nofs_restore(flags);
 620         return ERR_PTR(-ENOMEM);
 621 }
 622
 623 /* Slowpath, don't want it inlined into btree_iter_traverse() */
 624 static noinline struct btree *bch2_btree_node_fill(struct bch_fs *c,
 625                                 struct btree_iter *iter,
 626                                 const struct bkey_i *k,
 627                                 enum btree_id btree_id,
 628                                 unsigned level,
 629                                 enum six_lock_type lock_type,
 630                                 bool sync)
 631 {
 632         struct btree_cache *bc = &c->btree_cache;
 633         struct btree *b;
 634
 635         BUG_ON(level + 1 >= BTREE_MAX_DEPTH);
 636         /*
 637          * Parent node must be locked, else we could read in a btree node that's
 638          * been freed:
 639          */
 640         if (iter && !bch2_btree_node_relock(iter, level + 1))
 641                 return ERR_PTR(-EINTR);
 642
 643         b = bch2_btree_node_mem_alloc(c);
 644         if (IS_ERR(b))
 645                 return b;
 646
 647         bkey_copy(&b->key, k);
 648         if (bch2_btree_node_hash_insert(bc, b, level, btree_id)) {
 649                 /* raced with another fill: */
 650
 651                 /* mark as unhashed... */
 652                 b->hash_val = 0;
 653
 654                 mutex_lock(&bc->lock);
 655                 list_add(&b->list, &bc->freeable);
 656                 mutex_unlock(&bc->lock);
 657
 658                 six_unlock_write(&b->c.lock);
 659                 six_unlock_intent(&b->c.lock);
 660                 return NULL;
 661         }
 662
 663         /*
 664          * Unlock before doing IO:
 665          *
 666          * XXX: ideally should be dropping all btree node locks here
 667          */
 668         if (iter && btree_node_read_locked(iter, level + 1))
 669                 btree_node_unlock(iter, level + 1);
 670
 671         bch2_btree_node_read(c, b, sync);
 672
 673         six_unlock_write(&b->c.lock);
 674
 675         if (!sync) {
 676                 six_unlock_intent(&b->c.lock);
 677                 return NULL;
 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 /**
 695  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 696  * in from disk if necessary.
 697  *
 698  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
 699  *
 700  * The btree node will have either a read or a write lock held, depending on
 701  * the @write parameter.
 702  */
 703 struct btree *bch2_btree_node_get(struct bch_fs *c, struct btree_iter *iter,
 704                                   const struct bkey_i *k, unsigned level,
 705                                   enum six_lock_type lock_type,
 706                                   unsigned long trace_ip)
 707 {
 708         struct btree_cache *bc = &c->btree_cache;
 709         struct btree *b;
 710         struct bset_tree *t;
 711
 712         EBUG_ON(level >= BTREE_MAX_DEPTH);
 713
 714         b = btree_node_mem_ptr(k);
 715         if (b)
 716                 goto lock_node;
 717 retry:
 718         b = btree_cache_find(bc, k);
 719         if (unlikely(!b)) {
 720                 /*
 721                  * We must have the parent locked to call bch2_btree_node_fill(),
 722                  * else we could read in a btree node from disk that's been
 723                  * freed:
 724                  */
 725                 b = bch2_btree_node_fill(c, iter, k, iter->btree_id,
 726                                          level, lock_type, true);
 727
 728                 /* We raced and found the btree node in the cache */
 729                 if (!b)
 730                         goto retry;
 731
 732                 if (IS_ERR(b))
 733                         return b;
 734         } else {
 735 lock_node:
 736                 /*
 737                  * There's a potential deadlock with splits and insertions into
 738                  * interior nodes we have to avoid:
 739                  *
 740                  * The other thread might be holding an intent lock on the node
 741                  * we want, and they want to update its parent node so they're
 742                  * going to upgrade their intent lock on the parent node to a
 743                  * write lock.
 744                  *
 745                  * But if we're holding a read lock on the parent, and we're
 746                  * trying to get the intent lock they're holding, we deadlock.
 747                  *
 748                  * So to avoid this we drop the read locks on parent nodes when
 749                  * we're starting to take intent locks - and handle the race.
 750                  *
 751                  * The race is that they might be about to free the node we
 752                  * want, and dropping our read lock on the parent node lets them
 753                  * update the parent marking the node we want as freed, and then
 754                  * free it:
 755                  *
 756                  * To guard against this, btree nodes are evicted from the cache
 757                  * when they're freed - and b->hash_val is zeroed out, which we
 758                  * check for after we lock the node.
 759                  *
 760                  * Then, bch2_btree_node_relock() on the parent will fail - because
 761                  * the parent was modified, when the pointer to the node we want
 762                  * was removed - and we'll bail out:
 763                  */
 764                 if (btree_node_read_locked(iter, level + 1))
 765                         btree_node_unlock(iter, level + 1);
 766
 767                 if (!btree_node_lock(b, k->k.p, level, iter, lock_type,
 768                                      lock_node_check_fn, (void *) k, trace_ip)) {
 769                         if (b->hash_val != btree_ptr_hash_val(k))
 770                                 goto retry;
 771                         return ERR_PTR(-EINTR);
 772                 }
 773
 774                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
 775                              b->c.level != level ||
 776                              race_fault())) {
 777                         six_unlock_type(&b->c.lock, lock_type);
 778                         if (bch2_btree_node_relock(iter, level + 1))
 779                                 goto retry;
 780
 781                         trace_trans_restart_btree_node_reused(iter->trans->ip);
 782                         return ERR_PTR(-EINTR);
 783                 }
 784         }
 785
 786         /* XXX: waiting on IO with btree locks held: */
 787         wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
 788                        TASK_UNINTERRUPTIBLE);
 789
 790         prefetch(b->aux_data);
 791
 792         for_each_bset(b, t) {
 793                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 794
 795                 prefetch(p + L1_CACHE_BYTES * 0);
 796                 prefetch(p + L1_CACHE_BYTES * 1);
 797                 prefetch(p + L1_CACHE_BYTES * 2);
 798         }
 799
 800         /* avoid atomic set bit if it's not needed: */
 801         if (!btree_node_accessed(b))
 802                 set_btree_node_accessed(b);
 803
 804         if (unlikely(btree_node_read_error(b))) {
 805                 six_unlock_type(&b->c.lock, lock_type);
 806                 return ERR_PTR(-EIO);
 807         }
 808
 809         EBUG_ON(b->c.btree_id != iter->btree_id);
 810         EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
 811         EBUG_ON(bkey_cmp(b->data->max_key, k->k.p));
 812         EBUG_ON(b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 813                 bkey_cmp(b->data->min_key,
 814                          bkey_i_to_btree_ptr_v2(&b->key)->v.min_key));
 815
 816         return b;
 817 }
 818
 819 struct btree *bch2_btree_node_get_noiter(struct bch_fs *c,
 820                                          const struct bkey_i *k,
 821                                          enum btree_id btree_id,
 822                                          unsigned level,
 823                                          bool nofill)
 824 {
 825         struct btree_cache *bc = &c->btree_cache;
 826         struct btree *b;
 827         struct bset_tree *t;
 828         int ret;
 829
 830         EBUG_ON(level >= BTREE_MAX_DEPTH);
 831
 832         b = btree_node_mem_ptr(k);
 833         if (b)
 834                 goto lock_node;
 835 retry:
 836         b = btree_cache_find(bc, k);
 837         if (unlikely(!b)) {
 838                 if (nofill)
 839                         goto out;
 840
 841                 b = bch2_btree_node_fill(c, NULL, k, btree_id,
 842                                          level, SIX_LOCK_read, true);
 843
 844                 /* We raced and found the btree node in the cache */
 845                 if (!b)
 846                         goto retry;
 847
 848                 if (IS_ERR(b) &&
 849                     !bch2_btree_cache_cannibalize_lock(c, NULL))
 850                         goto retry;
 851
 852                 if (IS_ERR(b))
 853                         goto out;
 854         } else {
 855 lock_node:
 856                 ret = six_lock_read(&b->c.lock, lock_node_check_fn, (void *) k);
 857                 if (ret)
 858                         goto retry;
 859
 860                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
 861                              b->c.btree_id != btree_id ||
 862                              b->c.level != level)) {
 863                         six_unlock_read(&b->c.lock);
 864                         goto retry;
 865                 }
 866         }
 867
 868         /* XXX: waiting on IO with btree locks held: */
 869         wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
 870                        TASK_UNINTERRUPTIBLE);
 871
 872         prefetch(b->aux_data);
 873
 874         for_each_bset(b, t) {
 875                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 876
 877                 prefetch(p + L1_CACHE_BYTES * 0);
 878                 prefetch(p + L1_CACHE_BYTES * 1);
 879                 prefetch(p + L1_CACHE_BYTES * 2);
 880         }
 881
 882         /* avoid atomic set bit if it's not needed: */
 883         if (!btree_node_accessed(b))
 884                 set_btree_node_accessed(b);
 885
 886         if (unlikely(btree_node_read_error(b))) {
 887                 six_unlock_read(&b->c.lock);
 888                 b = ERR_PTR(-EIO);
 889                 goto out;
 890         }
 891
 892         EBUG_ON(b->c.btree_id != btree_id);
 893         EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
 894         EBUG_ON(bkey_cmp(b->data->max_key, k->k.p));
 895         EBUG_ON(b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 896                 bkey_cmp(b->data->min_key,
 897                          bkey_i_to_btree_ptr_v2(&b->key)->v.min_key));
 898 out:
 899         bch2_btree_cache_cannibalize_unlock(c);
 900         return b;
 901 }
 902
 903 struct btree *bch2_btree_node_get_sibling(struct bch_fs *c,
 904                                           struct btree_iter *iter,
 905                                           struct btree *b,
 906                                           enum btree_node_sibling sib)
 907 {
 908         struct btree_trans *trans = iter->trans;
 909         struct btree *parent;
 910         struct btree_node_iter node_iter;
 911         struct bkey_packed *k;
 912         struct bkey_buf tmp;
 913         struct btree *ret = NULL;
 914         unsigned level = b->c.level;
 915
 916         bch2_bkey_buf_init(&tmp);
 917
 918         parent = btree_iter_node(iter, level + 1);
 919         if (!parent)
 920                 return NULL;
 921
 922         /*
 923          * There's a corner case where a btree_iter might have a node locked
 924          * that is just outside its current pos - when
 925          * bch2_btree_iter_set_pos_same_leaf() gets to the end of the node.
 926          *
 927          * But the lock ordering checks in __bch2_btree_node_lock() go off of
 928          * iter->pos, not the node's key: so if the iterator is marked as
 929          * needing to be traversed, we risk deadlock if we don't bail out here:
 930          */
 931         if (iter->uptodate >= BTREE_ITER_NEED_TRAVERSE)
 932                 return ERR_PTR(-EINTR);
 933
 934         if (!bch2_btree_node_relock(iter, level + 1)) {
 935                 ret = ERR_PTR(-EINTR);
 936                 goto out;
 937         }
 938
 939         node_iter = iter->l[parent->c.level].iter;
 940
 941         k = bch2_btree_node_iter_peek_all(&node_iter, parent);
 942         BUG_ON(bkey_cmp_left_packed(parent, k, &b->key.k.p));
 943
 944         k = sib == btree_prev_sib
 945                 ? bch2_btree_node_iter_prev(&node_iter, parent)
 946                 : (bch2_btree_node_iter_advance(&node_iter, parent),
 947                    bch2_btree_node_iter_peek(&node_iter, parent));
 948         if (!k)
 949                 goto out;
 950
 951         bch2_bkey_buf_unpack(&tmp, c, parent, k);
 952
 953         ret = bch2_btree_node_get(c, iter, tmp.k, level,
 954                                   SIX_LOCK_intent, _THIS_IP_);
 955
 956         if (PTR_ERR_OR_ZERO(ret) == -EINTR && !trans->nounlock) {
 957                 struct btree_iter *linked;
 958
 959                 if (!bch2_btree_node_relock(iter, level + 1))
 960                         goto out;
 961
 962                 /*
 963                  * We might have got -EINTR because trylock failed, and we're
 964                  * holding other locks that would cause us to deadlock:
 965                  */
 966                 trans_for_each_iter(trans, linked)
 967                         if (btree_iter_lock_cmp(iter, linked) < 0)
 968                                 __bch2_btree_iter_unlock(linked);
 969
 970                 if (sib == btree_prev_sib)
 971                         btree_node_unlock(iter, level);
 972
 973                 ret = bch2_btree_node_get(c, iter, tmp.k, level,
 974                                           SIX_LOCK_intent, _THIS_IP_);
 975
 976                 /*
 977                  * before btree_iter_relock() calls btree_iter_verify_locks():
 978                  */
 979                 if (btree_lock_want(iter, level + 1) == BTREE_NODE_UNLOCKED)
 980                         btree_node_unlock(iter, level + 1);
 981
 982                 if (!bch2_btree_node_relock(iter, level)) {
 983                         btree_iter_set_dirty(iter, BTREE_ITER_NEED_RELOCK);
 984
 985                         if (!IS_ERR(ret)) {
 986                                 six_unlock_intent(&ret->c.lock);
 987                                 ret = ERR_PTR(-EINTR);
 988                         }
 989                 }
 990
 991                 bch2_trans_relock(trans);
 992         }
 993 out:
 994         if (btree_lock_want(iter, level + 1) == BTREE_NODE_UNLOCKED)
 995                 btree_node_unlock(iter, level + 1);
 996
 997         if (PTR_ERR_OR_ZERO(ret) == -EINTR)
 998                 bch2_btree_iter_upgrade(iter, level + 2);
 999
1000         BUG_ON(!IS_ERR(ret) && !btree_node_locked(iter, level));
1001
1002         if (!IS_ERR_OR_NULL(ret)) {
1003                 struct btree *n1 = ret, *n2 = b;
1004
1005                 if (sib != btree_prev_sib)
1006                         swap(n1, n2);
1007
1008                 if (bkey_cmp(bkey_successor(n1->key.k.p),
1009                              n2->data->min_key)) {
1010                         char buf1[200], buf2[200];
1011
1012                         bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(&n1->key));
1013                         bch2_bkey_val_to_text(&PBUF(buf2), c, bkey_i_to_s_c(&n2->key));
1014
1015                         bch2_fs_inconsistent(c, "btree topology error at btree %s level %u:\n"
1016                                              "prev: %s\n"
1017                                              "next: %s\n",
1018                                              bch2_btree_ids[iter->btree_id], level,
1019                                              buf1, buf2);
1020
1021                         six_unlock_intent(&ret->c.lock);
1022                         ret = NULL;
1023                 }
1024         }
1025
1026         bch2_btree_trans_verify_locks(trans);
1027
1028         bch2_bkey_buf_exit(&tmp, c);
1029
1030         return ret;
1031 }
1032
1033 void bch2_btree_node_prefetch(struct bch_fs *c, struct btree_iter *iter,
1034                               const struct bkey_i *k,
1035                               enum btree_id btree_id, unsigned level)
1036 {
1037         struct btree_cache *bc = &c->btree_cache;
1038         struct btree *b;
1039
1040         BUG_ON(iter && !btree_node_locked(iter, level + 1));
1041         BUG_ON(level >= BTREE_MAX_DEPTH);
1042
1043         b = btree_cache_find(bc, k);
1044         if (b)
1045                 return;
1046
1047         bch2_btree_node_fill(c, iter, k, btree_id, level, SIX_LOCK_read, false);
1048 }
1049
1050 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
1051                              struct btree *b)
1052 {
1053         const struct bkey_format *f = &b->format;
1054         struct bset_stats stats;
1055
1056         memset(&stats, 0, sizeof(stats));
1057
1058         bch2_btree_keys_stats(b, &stats);
1059
1060         pr_buf(out, "l %u ", b->c.level);
1061         bch2_bpos_to_text(out, b->data->min_key);
1062         pr_buf(out, " - ");
1063         bch2_bpos_to_text(out, b->data->max_key);
1064         pr_buf(out, ":\n"
1065                "    ptrs: ");
1066         bch2_val_to_text(out, c, bkey_i_to_s_c(&b->key));
1067
1068         pr_buf(out, "\n"
1069                "    format: u64s %u fields %u %u %u %u %u\n"
1070                "    unpack fn len: %u\n"
1071                "    bytes used %zu/%zu (%zu%% full)\n"
1072                "    sib u64s: %u, %u (merge threshold %zu)\n"
1073                "    nr packed keys %u\n"
1074                "    nr unpacked keys %u\n"
1075                "    floats %zu\n"
1076                "    failed unpacked %zu\n",
1077                f->key_u64s,
1078                f->bits_per_field[0],
1079                f->bits_per_field[1],
1080                f->bits_per_field[2],
1081                f->bits_per_field[3],
1082                f->bits_per_field[4],
1083                b->unpack_fn_len,
1084                b->nr.live_u64s * sizeof(u64),
1085                btree_bytes(c) - sizeof(struct btree_node),
1086                b->nr.live_u64s * 100 / btree_max_u64s(c),
1087                b->sib_u64s[0],
1088                b->sib_u64s[1],
1089                BTREE_FOREGROUND_MERGE_THRESHOLD(c),
1090                b->nr.packed_keys,
1091                b->nr.unpacked_keys,
1092                stats.floats,
1093                stats.failed);
1094 }
1095
1096 void bch2_btree_cache_to_text(struct printbuf *out, struct bch_fs *c)
1097 {
1098         pr_buf(out, "nr nodes:\t\t%u\n", c->btree_cache.used);
1099         pr_buf(out, "nr dirty:\t\t%u\n", atomic_read(&c->btree_cache.dirty));
1100         pr_buf(out, "cannibalize lock:\t%p\n", c->btree_cache.alloc_lock);
1101 }