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

   1
   2 #include "bcachefs.h"
   3 #include "btree_cache.h"
   4 #include "btree_io.h"
   5 #include "btree_iter.h"
   6 #include "btree_locking.h"
   7 #include "debug.h"
   8
   9 #include <linux/prefetch.h>
  10 #include <trace/events/bcachefs.h>
  11
  12 const char * const bch2_btree_ids[] = {
  13 #define x(kwd, val, name) name,
  14         BCH_BTREE_IDS()
  15 #undef x
  16         NULL
  17 };
  18
  19 void bch2_recalc_btree_reserve(struct bch_fs *c)
  20 {
  21         unsigned i, reserve = 16;
  22
  23         if (!c->btree_roots[0].b)
  24                 reserve += 8;
  25
  26         for (i = 0; i < BTREE_ID_NR; i++)
  27                 if (c->btree_roots[i].b)
  28                         reserve += min_t(unsigned, 1,
  29                                          c->btree_roots[i].b->level) * 8;
  30
  31         c->btree_cache.reserve = reserve;
  32 }
  33
  34 static inline unsigned btree_cache_can_free(struct btree_cache *bc)
  35 {
  36         return max_t(int, 0, bc->used - bc->reserve);
  37 }
  38
  39 static void __btree_node_data_free(struct bch_fs *c, struct btree *b)
  40 {
  41         EBUG_ON(btree_node_write_in_flight(b));
  42
  43         kvpfree(b->data, btree_bytes(c));
  44         b->data = NULL;
  45         bch2_btree_keys_free(b);
  46 }
  47
  48 static void btree_node_data_free(struct bch_fs *c, struct btree *b)
  49 {
  50         struct btree_cache *bc = &c->btree_cache;
  51
  52         __btree_node_data_free(c, b);
  53         bc->used--;
  54         list_move(&b->list, &bc->freed);
  55 }
  56
  57 static int bch2_btree_cache_cmp_fn(struct rhashtable_compare_arg *arg,
  58                                    const void *obj)
  59 {
  60         const struct btree *b = obj;
  61         const u64 *v = arg->key;
  62
  63         return PTR_HASH(&b->key) == *v ? 0 : 1;
  64 }
  65
  66 static const struct rhashtable_params bch_btree_cache_params = {
  67         .head_offset    = offsetof(struct btree, hash),
  68         .key_offset     = offsetof(struct btree, key.v),
  69         .key_len        = sizeof(struct bch_extent_ptr),
  70         .obj_cmpfn      = bch2_btree_cache_cmp_fn,
  71 };
  72
  73 static void btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
  74 {
  75         struct btree_cache *bc = &c->btree_cache;
  76
  77         b->data = kvpmalloc(btree_bytes(c), gfp);
  78         if (!b->data)
  79                 goto err;
  80
  81         if (bch2_btree_keys_alloc(b, btree_page_order(c), gfp))
  82                 goto err;
  83
  84         bc->used++;
  85         list_move(&b->list, &bc->freeable);
  86         return;
  87 err:
  88         kvpfree(b->data, btree_bytes(c));
  89         b->data = NULL;
  90         list_move(&b->list, &bc->freed);
  91 }
  92
  93 static struct btree *btree_node_mem_alloc(struct bch_fs *c, gfp_t gfp)
  94 {
  95         struct btree *b = kzalloc(sizeof(struct btree), gfp);
  96         if (!b)
  97                 return NULL;
  98
  99         bkey_btree_ptr_init(&b->key);
 100         six_lock_init(&b->lock);
 101         INIT_LIST_HEAD(&b->list);
 102         INIT_LIST_HEAD(&b->write_blocked);
 103
 104         btree_node_data_alloc(c, b, gfp);
 105         return b->data ? b : NULL;
 106 }
 107
 108 /* Btree in memory cache - hash table */
 109
 110 void bch2_btree_node_hash_remove(struct btree_cache *bc, struct btree *b)
 111 {
 112         rhashtable_remove_fast(&bc->table, &b->hash, bch_btree_cache_params);
 113
 114         /* Cause future lookups for this node to fail: */
 115         PTR_HASH(&b->key) = 0;
 116 }
 117
 118 int __bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b)
 119 {
 120         return rhashtable_lookup_insert_fast(&bc->table, &b->hash,
 121                                              bch_btree_cache_params);
 122 }
 123
 124 int bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b,
 125                                 unsigned level, enum btree_id id)
 126 {
 127         int ret;
 128
 129         b->level        = level;
 130         b->btree_id     = id;
 131
 132         mutex_lock(&bc->lock);
 133         ret = __bch2_btree_node_hash_insert(bc, b);
 134         if (!ret)
 135                 list_add(&b->list, &bc->live);
 136         mutex_unlock(&bc->lock);
 137
 138         return ret;
 139 }
 140
 141 __flatten
 142 static inline struct btree *btree_cache_find(struct btree_cache *bc,
 143                                      const struct bkey_i *k)
 144 {
 145         return rhashtable_lookup_fast(&bc->table, &PTR_HASH(k),
 146                                       bch_btree_cache_params);
 147 }
 148
 149 /*
 150  * this version is for btree nodes that have already been freed (we're not
 151  * reaping a real btree node)
 152  */
 153 static int __btree_node_reclaim(struct bch_fs *c, struct btree *b, bool flush)
 154 {
 155         struct btree_cache *bc = &c->btree_cache;
 156         int ret = 0;
 157
 158         lockdep_assert_held(&bc->lock);
 159
 160         if (!six_trylock_intent(&b->lock))
 161                 return -ENOMEM;
 162
 163         if (!six_trylock_write(&b->lock))
 164                 goto out_unlock_intent;
 165
 166         if (btree_node_noevict(b))
 167                 goto out_unlock;
 168
 169         if (!btree_node_may_write(b))
 170                 goto out_unlock;
 171
 172         if (btree_node_dirty(b) &&
 173             test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags))
 174                 goto out_unlock;
 175
 176         if (btree_node_dirty(b) ||
 177             btree_node_write_in_flight(b) ||
 178             btree_node_read_in_flight(b)) {
 179                 if (!flush)
 180                         goto out_unlock;
 181
 182                 wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
 183                                TASK_UNINTERRUPTIBLE);
 184
 185                 /*
 186                  * Using the underscore version because we don't want to compact
 187                  * bsets after the write, since this node is about to be evicted
 188                  * - unless btree verify mode is enabled, since it runs out of
 189                  * the post write cleanup:
 190                  */
 191                 if (verify_btree_ondisk(c))
 192                         bch2_btree_node_write(c, b, SIX_LOCK_intent);
 193                 else
 194                         __bch2_btree_node_write(c, b, SIX_LOCK_read);
 195
 196                 /* wait for any in flight btree write */
 197                 btree_node_wait_on_io(b);
 198         }
 199 out:
 200         if (PTR_HASH(&b->key) && !ret)
 201                 trace_btree_node_reap(c, b);
 202         return ret;
 203 out_unlock:
 204         six_unlock_write(&b->lock);
 205 out_unlock_intent:
 206         six_unlock_intent(&b->lock);
 207         ret = -ENOMEM;
 208         goto out;
 209 }
 210
 211 static int btree_node_reclaim(struct bch_fs *c, struct btree *b)
 212 {
 213         return __btree_node_reclaim(c, b, false);
 214 }
 215
 216 static int btree_node_write_and_reclaim(struct bch_fs *c, struct btree *b)
 217 {
 218         return __btree_node_reclaim(c, b, true);
 219 }
 220
 221 static unsigned long bch2_btree_cache_scan(struct shrinker *shrink,
 222                                            struct shrink_control *sc)
 223 {
 224         struct bch_fs *c = container_of(shrink, struct bch_fs,
 225                                         btree_cache.shrink);
 226         struct btree_cache *bc = &c->btree_cache;
 227         struct btree *b, *t;
 228         unsigned long nr = sc->nr_to_scan;
 229         unsigned long can_free;
 230         unsigned long touched = 0;
 231         unsigned long freed = 0;
 232         unsigned i;
 233
 234         if (btree_shrinker_disabled(c))
 235                 return SHRINK_STOP;
 236
 237         /* Return -1 if we can't do anything right now */
 238         if (sc->gfp_mask & __GFP_IO)
 239                 mutex_lock(&bc->lock);
 240         else if (!mutex_trylock(&bc->lock))
 241                 return -1;
 242
 243         /*
 244          * It's _really_ critical that we don't free too many btree nodes - we
 245          * have to always leave ourselves a reserve. The reserve is how we
 246          * guarantee that allocating memory for a new btree node can always
 247          * succeed, so that inserting keys into the btree can always succeed and
 248          * IO can always make forward progress:
 249          */
 250         nr /= btree_pages(c);
 251         can_free = btree_cache_can_free(bc);
 252         nr = min_t(unsigned long, nr, can_free);
 253
 254         i = 0;
 255         list_for_each_entry_safe(b, t, &bc->freeable, list) {
 256                 touched++;
 257
 258                 if (freed >= nr)
 259                         break;
 260
 261                 if (++i > 3 &&
 262                     !btree_node_reclaim(c, b)) {
 263                         btree_node_data_free(c, b);
 264                         six_unlock_write(&b->lock);
 265                         six_unlock_intent(&b->lock);
 266                         freed++;
 267                 }
 268         }
 269 restart:
 270         list_for_each_entry_safe(b, t, &bc->live, list) {
 271                 touched++;
 272
 273                 if (freed >= nr) {
 274                         /* Save position */
 275                         if (&t->list != &bc->live)
 276                                 list_move_tail(&bc->live, &t->list);
 277                         break;
 278                 }
 279
 280                 if (!btree_node_accessed(b) &&
 281                     !btree_node_reclaim(c, b)) {
 282                         /* can't call bch2_btree_node_hash_remove under lock  */
 283                         freed++;
 284                         if (&t->list != &bc->live)
 285                                 list_move_tail(&bc->live, &t->list);
 286
 287                         btree_node_data_free(c, b);
 288                         mutex_unlock(&bc->lock);
 289
 290                         bch2_btree_node_hash_remove(bc, b);
 291                         six_unlock_write(&b->lock);
 292                         six_unlock_intent(&b->lock);
 293
 294                         if (freed >= nr)
 295                                 goto out;
 296
 297                         if (sc->gfp_mask & __GFP_IO)
 298                                 mutex_lock(&bc->lock);
 299                         else if (!mutex_trylock(&bc->lock))
 300                                 goto out;
 301                         goto restart;
 302                 } else
 303                         clear_btree_node_accessed(b);
 304         }
 305
 306         mutex_unlock(&bc->lock);
 307 out:
 308         return (unsigned long) freed * btree_pages(c);
 309 }
 310
 311 static unsigned long bch2_btree_cache_count(struct shrinker *shrink,
 312                                             struct shrink_control *sc)
 313 {
 314         struct bch_fs *c = container_of(shrink, struct bch_fs,
 315                                         btree_cache.shrink);
 316         struct btree_cache *bc = &c->btree_cache;
 317
 318         if (btree_shrinker_disabled(c))
 319                 return 0;
 320
 321         return btree_cache_can_free(bc) * btree_pages(c);
 322 }
 323
 324 void bch2_fs_btree_cache_exit(struct bch_fs *c)
 325 {
 326         struct btree_cache *bc = &c->btree_cache;
 327         struct btree *b;
 328         unsigned i;
 329
 330         if (bc->shrink.list.next)
 331                 unregister_shrinker(&bc->shrink);
 332
 333         mutex_lock(&bc->lock);
 334
 335 #ifdef CONFIG_BCACHEFS_DEBUG
 336         if (c->verify_data)
 337                 list_move(&c->verify_data->list, &bc->live);
 338
 339         kvpfree(c->verify_ondisk, btree_bytes(c));
 340 #endif
 341
 342         for (i = 0; i < BTREE_ID_NR; i++)
 343                 if (c->btree_roots[i].b)
 344                         list_add(&c->btree_roots[i].b->list, &bc->live);
 345
 346         list_splice(&bc->freeable, &bc->live);
 347
 348         while (!list_empty(&bc->live)) {
 349                 b = list_first_entry(&bc->live, struct btree, list);
 350
 351                 BUG_ON(btree_node_read_in_flight(b) ||
 352                        btree_node_write_in_flight(b));
 353
 354                 if (btree_node_dirty(b))
 355                         bch2_btree_complete_write(c, b, btree_current_write(b));
 356                 clear_btree_node_dirty(b);
 357
 358                 btree_node_data_free(c, b);
 359         }
 360
 361         while (!list_empty(&bc->freed)) {
 362                 b = list_first_entry(&bc->freed, struct btree, list);
 363                 list_del(&b->list);
 364                 kfree(b);
 365         }
 366
 367         mutex_unlock(&bc->lock);
 368
 369         if (bc->table_init_done)
 370                 rhashtable_destroy(&bc->table);
 371 }
 372
 373 int bch2_fs_btree_cache_init(struct bch_fs *c)
 374 {
 375         struct btree_cache *bc = &c->btree_cache;
 376         unsigned i;
 377         int ret = 0;
 378
 379         pr_verbose_init(c->opts, "");
 380
 381         ret = rhashtable_init(&bc->table, &bch_btree_cache_params);
 382         if (ret)
 383                 goto out;
 384
 385         bc->table_init_done = true;
 386
 387         bch2_recalc_btree_reserve(c);
 388
 389         for (i = 0; i < bc->reserve; i++)
 390                 if (!btree_node_mem_alloc(c, GFP_KERNEL)) {
 391                         ret = -ENOMEM;
 392                         goto out;
 393                 }
 394
 395         list_splice_init(&bc->live, &bc->freeable);
 396
 397 #ifdef CONFIG_BCACHEFS_DEBUG
 398         mutex_init(&c->verify_lock);
 399
 400         c->verify_ondisk = kvpmalloc(btree_bytes(c), GFP_KERNEL);
 401         if (!c->verify_ondisk) {
 402                 ret = -ENOMEM;
 403                 goto out;
 404         }
 405
 406         c->verify_data = btree_node_mem_alloc(c, GFP_KERNEL);
 407         if (!c->verify_data) {
 408                 ret = -ENOMEM;
 409                 goto out;
 410         }
 411
 412         list_del_init(&c->verify_data->list);
 413 #endif
 414
 415         bc->shrink.count_objects        = bch2_btree_cache_count;
 416         bc->shrink.scan_objects         = bch2_btree_cache_scan;
 417         bc->shrink.seeks                = 4;
 418         bc->shrink.batch                = btree_pages(c) * 2;
 419         register_shrinker(&bc->shrink);
 420 out:
 421         pr_verbose_init(c->opts, "ret %i", ret);
 422         return ret;
 423 }
 424
 425 void bch2_fs_btree_cache_init_early(struct btree_cache *bc)
 426 {
 427         mutex_init(&bc->lock);
 428         INIT_LIST_HEAD(&bc->live);
 429         INIT_LIST_HEAD(&bc->freeable);
 430         INIT_LIST_HEAD(&bc->freed);
 431 }
 432
 433 /*
 434  * We can only have one thread cannibalizing other cached btree nodes at a time,
 435  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 436  * cannibalize_bucket() will take. This means every time we unlock the root of
 437  * the btree, we need to release this lock if we have it held.
 438  */
 439 void bch2_btree_cache_cannibalize_unlock(struct bch_fs *c)
 440 {
 441         struct btree_cache *bc = &c->btree_cache;
 442
 443         if (bc->alloc_lock == current) {
 444                 trace_btree_node_cannibalize_unlock(c);
 445                 bc->alloc_lock = NULL;
 446                 closure_wake_up(&bc->alloc_wait);
 447         }
 448 }
 449
 450 int bch2_btree_cache_cannibalize_lock(struct bch_fs *c, struct closure *cl)
 451 {
 452         struct btree_cache *bc = &c->btree_cache;
 453         struct task_struct *old;
 454
 455         old = cmpxchg(&bc->alloc_lock, NULL, current);
 456         if (old == NULL || old == current)
 457                 goto success;
 458
 459         if (!cl) {
 460                 trace_btree_node_cannibalize_lock_fail(c);
 461                 return -ENOMEM;
 462         }
 463
 464         closure_wait(&bc->alloc_wait, cl);
 465
 466         /* Try again, after adding ourselves to waitlist */
 467         old = cmpxchg(&bc->alloc_lock, NULL, current);
 468         if (old == NULL || old == current) {
 469                 /* We raced */
 470                 closure_wake_up(&bc->alloc_wait);
 471                 goto success;
 472         }
 473
 474         trace_btree_node_cannibalize_lock_fail(c);
 475         return -EAGAIN;
 476
 477 success:
 478         trace_btree_node_cannibalize_lock(c);
 479         return 0;
 480 }
 481
 482 static struct btree *btree_node_cannibalize(struct bch_fs *c)
 483 {
 484         struct btree_cache *bc = &c->btree_cache;
 485         struct btree *b;
 486
 487         list_for_each_entry_reverse(b, &bc->live, list)
 488                 if (!btree_node_reclaim(c, b))
 489                         return b;
 490
 491         while (1) {
 492                 list_for_each_entry_reverse(b, &bc->live, list)
 493                         if (!btree_node_write_and_reclaim(c, b))
 494                                 return b;
 495
 496                 /*
 497                  * Rare case: all nodes were intent-locked.
 498                  * Just busy-wait.
 499                  */
 500                 WARN_ONCE(1, "btree cache cannibalize failed\n");
 501                 cond_resched();
 502         }
 503 }
 504
 505 struct btree *bch2_btree_node_mem_alloc(struct bch_fs *c)
 506 {
 507         struct btree_cache *bc = &c->btree_cache;
 508         struct btree *b;
 509         u64 start_time = local_clock();
 510
 511         mutex_lock(&bc->lock);
 512
 513         /*
 514          * btree_free() doesn't free memory; it sticks the node on the end of
 515          * the list. Check if there's any freed nodes there:
 516          */
 517         list_for_each_entry(b, &bc->freeable, list)
 518                 if (!btree_node_reclaim(c, b))
 519                         goto out_unlock;
 520
 521         /*
 522          * We never free struct btree itself, just the memory that holds the on
 523          * disk node. Check the freed list before allocating a new one:
 524          */
 525         list_for_each_entry(b, &bc->freed, list)
 526                 if (!btree_node_reclaim(c, b)) {
 527                         btree_node_data_alloc(c, b, __GFP_NOWARN|GFP_NOIO);
 528                         if (b->data)
 529                                 goto out_unlock;
 530
 531                         six_unlock_write(&b->lock);
 532                         six_unlock_intent(&b->lock);
 533                         goto err;
 534                 }
 535
 536         b = btree_node_mem_alloc(c, __GFP_NOWARN|GFP_NOIO);
 537         if (!b)
 538                 goto err;
 539
 540         BUG_ON(!six_trylock_intent(&b->lock));
 541         BUG_ON(!six_trylock_write(&b->lock));
 542 out_unlock:
 543         BUG_ON(btree_node_hashed(b));
 544         BUG_ON(btree_node_write_in_flight(b));
 545
 546         list_del_init(&b->list);
 547         mutex_unlock(&bc->lock);
 548 out:
 549         b->flags                = 0;
 550         b->written              = 0;
 551         b->nsets                = 0;
 552         b->sib_u64s[0]          = 0;
 553         b->sib_u64s[1]          = 0;
 554         b->whiteout_u64s        = 0;
 555         b->uncompacted_whiteout_u64s = 0;
 556         bch2_btree_keys_init(b, &c->expensive_debug_checks);
 557
 558         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],
 559                                start_time);
 560
 561         return b;
 562 err:
 563         /* Try to cannibalize another cached btree node: */
 564         if (bc->alloc_lock == current) {
 565                 b = btree_node_cannibalize(c);
 566                 list_del_init(&b->list);
 567                 mutex_unlock(&bc->lock);
 568
 569                 bch2_btree_node_hash_remove(bc, b);
 570
 571                 trace_btree_node_cannibalize(c);
 572                 goto out;
 573         }
 574
 575         mutex_unlock(&bc->lock);
 576         return ERR_PTR(-ENOMEM);
 577 }
 578
 579 /* Slowpath, don't want it inlined into btree_iter_traverse() */
 580 static noinline struct btree *bch2_btree_node_fill(struct bch_fs *c,
 581                                 struct btree_iter *iter,
 582                                 const struct bkey_i *k,
 583                                 unsigned level,
 584                                 enum six_lock_type lock_type,
 585                                 bool sync)
 586 {
 587         struct btree_cache *bc = &c->btree_cache;
 588         struct btree *b;
 589
 590         /*
 591          * Parent node must be locked, else we could read in a btree node that's
 592          * been freed:
 593          */
 594         BUG_ON(!btree_node_locked(iter, level + 1));
 595         BUG_ON(level >= BTREE_MAX_DEPTH);
 596
 597         b = bch2_btree_node_mem_alloc(c);
 598         if (IS_ERR(b))
 599                 return b;
 600
 601         bkey_copy(&b->key, k);
 602         if (bch2_btree_node_hash_insert(bc, b, level, iter->btree_id)) {
 603                 /* raced with another fill: */
 604
 605                 /* mark as unhashed... */
 606                 PTR_HASH(&b->key) = 0;
 607
 608                 mutex_lock(&bc->lock);
 609                 list_add(&b->list, &bc->freeable);
 610                 mutex_unlock(&bc->lock);
 611
 612                 six_unlock_write(&b->lock);
 613                 six_unlock_intent(&b->lock);
 614                 return NULL;
 615         }
 616
 617         /*
 618          * If the btree node wasn't cached, we can't drop our lock on
 619          * the parent until after it's added to the cache - because
 620          * otherwise we could race with a btree_split() freeing the node
 621          * we're trying to lock.
 622          *
 623          * But the deadlock described below doesn't exist in this case,
 624          * so it's safe to not drop the parent lock until here:
 625          */
 626         if (btree_node_read_locked(iter, level + 1))
 627                 btree_node_unlock(iter, level + 1);
 628
 629         bch2_btree_node_read(c, b, sync);
 630
 631         six_unlock_write(&b->lock);
 632
 633         if (!sync) {
 634                 six_unlock_intent(&b->lock);
 635                 return NULL;
 636         }
 637
 638         if (lock_type == SIX_LOCK_read)
 639                 six_lock_downgrade(&b->lock);
 640
 641         return b;
 642 }
 643
 644 /**
 645  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 646  * in from disk if necessary.
 647  *
 648  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
 649  *
 650  * The btree node will have either a read or a write lock held, depending on
 651  * the @write parameter.
 652  */
 653 struct btree *bch2_btree_node_get(struct bch_fs *c, struct btree_iter *iter,
 654                                   const struct bkey_i *k, unsigned level,
 655                                   enum six_lock_type lock_type,
 656                                   bool may_drop_locks)
 657 {
 658         struct btree_cache *bc = &c->btree_cache;
 659         struct btree *b;
 660         struct bset_tree *t;
 661
 662         /*
 663          * XXX: locking optimization
 664          *
 665          * we can make the locking looser here - caller can drop lock on parent
 666          * node before locking child node (and potentially blocking): we just
 667          * have to have bch2_btree_node_fill() call relock on the parent and
 668          * return -EINTR if that fails
 669          */
 670         EBUG_ON(!btree_node_locked(iter, level + 1));
 671         EBUG_ON(level >= BTREE_MAX_DEPTH);
 672 retry:
 673         rcu_read_lock();
 674         b = btree_cache_find(bc, k);
 675         rcu_read_unlock();
 676
 677         if (unlikely(!b)) {
 678                 /*
 679                  * We must have the parent locked to call bch2_btree_node_fill(),
 680                  * else we could read in a btree node from disk that's been
 681                  * freed:
 682                  */
 683                 b = bch2_btree_node_fill(c, iter, k, level, lock_type, true);
 684
 685                 /* We raced and found the btree node in the cache */
 686                 if (!b)
 687                         goto retry;
 688
 689                 if (IS_ERR(b))
 690                         return b;
 691         } else {
 692                 /*
 693                  * There's a potential deadlock with splits and insertions into
 694                  * interior nodes we have to avoid:
 695                  *
 696                  * The other thread might be holding an intent lock on the node
 697                  * we want, and they want to update its parent node so they're
 698                  * going to upgrade their intent lock on the parent node to a
 699                  * write lock.
 700                  *
 701                  * But if we're holding a read lock on the parent, and we're
 702                  * trying to get the intent lock they're holding, we deadlock.
 703                  *
 704                  * So to avoid this we drop the read locks on parent nodes when
 705                  * we're starting to take intent locks - and handle the race.
 706                  *
 707                  * The race is that they might be about to free the node we
 708                  * want, and dropping our read lock on the parent node lets them
 709                  * update the parent marking the node we want as freed, and then
 710                  * free it:
 711                  *
 712                  * To guard against this, btree nodes are evicted from the cache
 713                  * when they're freed - and PTR_HASH() is zeroed out, which we
 714                  * check for after we lock the node.
 715                  *
 716                  * Then, bch2_btree_node_relock() on the parent will fail - because
 717                  * the parent was modified, when the pointer to the node we want
 718                  * was removed - and we'll bail out:
 719                  */
 720                 if (btree_node_read_locked(iter, level + 1))
 721                         btree_node_unlock(iter, level + 1);
 722
 723                 if (!btree_node_lock(b, k->k.p, level, iter,
 724                                      lock_type, may_drop_locks))
 725                         return ERR_PTR(-EINTR);
 726
 727                 if (unlikely(PTR_HASH(&b->key) != PTR_HASH(k) ||
 728                              b->level != level ||
 729                              race_fault())) {
 730                         six_unlock_type(&b->lock, lock_type);
 731                         if (bch2_btree_node_relock(iter, level + 1))
 732                                 goto retry;
 733
 734                         trans_restart();
 735                         trace_trans_restart_btree_node_reused(c,
 736                                                 iter->trans->ip);
 737                         return ERR_PTR(-EINTR);
 738                 }
 739         }
 740
 741         wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
 742                        TASK_UNINTERRUPTIBLE);
 743
 744         prefetch(b->aux_data);
 745
 746         for_each_bset(b, t) {
 747                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 748
 749                 prefetch(p + L1_CACHE_BYTES * 0);
 750                 prefetch(p + L1_CACHE_BYTES * 1);
 751                 prefetch(p + L1_CACHE_BYTES * 2);
 752         }
 753
 754         /* avoid atomic set bit if it's not needed: */
 755         if (btree_node_accessed(b))
 756                 set_btree_node_accessed(b);
 757
 758         if (unlikely(btree_node_read_error(b))) {
 759                 six_unlock_type(&b->lock, lock_type);
 760                 return ERR_PTR(-EIO);
 761         }
 762
 763         EBUG_ON(b->btree_id != iter->btree_id ||
 764                 BTREE_NODE_LEVEL(b->data) != level ||
 765                 bkey_cmp(b->data->max_key, k->k.p));
 766
 767         return b;
 768 }
 769
 770 struct btree *bch2_btree_node_get_sibling(struct bch_fs *c,
 771                                           struct btree_iter *iter,
 772                                           struct btree *b,
 773                                           bool may_drop_locks,
 774                                           enum btree_node_sibling sib)
 775 {
 776         struct btree *parent;
 777         struct btree_node_iter node_iter;
 778         struct bkey_packed *k;
 779         BKEY_PADDED(k) tmp;
 780         struct btree *ret = NULL;
 781         unsigned level = b->level;
 782
 783         parent = btree_iter_node(iter, level + 1);
 784         if (!parent)
 785                 return NULL;
 786
 787         if (!bch2_btree_node_relock(iter, level + 1))
 788                 goto out_upgrade;
 789
 790         node_iter = iter->l[parent->level].iter;
 791
 792         k = bch2_btree_node_iter_peek_all(&node_iter, parent);
 793         BUG_ON(bkey_cmp_left_packed(parent, k, &b->key.k.p));
 794
 795         k = sib == btree_prev_sib
 796                 ? bch2_btree_node_iter_prev(&node_iter, parent)
 797                 : (bch2_btree_node_iter_advance(&node_iter, parent),
 798                    bch2_btree_node_iter_peek(&node_iter, parent));
 799         if (!k)
 800                 goto out;
 801
 802         bch2_bkey_unpack(parent, &tmp.k, k);
 803
 804         ret = bch2_btree_node_get(c, iter, &tmp.k, level,
 805                                   SIX_LOCK_intent, may_drop_locks);
 806
 807         if (PTR_ERR_OR_ZERO(ret) == -EINTR && may_drop_locks) {
 808                 struct btree_iter *linked;
 809
 810                 if (!bch2_btree_node_relock(iter, level + 1))
 811                         goto out_upgrade;
 812
 813                 /*
 814                  * We might have got -EINTR because trylock failed, and we're
 815                  * holding other locks that would cause us to deadlock:
 816                  */
 817                 trans_for_each_iter(iter->trans, linked)
 818                         if (btree_iter_cmp(iter, linked) < 0)
 819                                 __bch2_btree_iter_unlock(linked);
 820
 821                 if (sib == btree_prev_sib)
 822                         btree_node_unlock(iter, level);
 823
 824                 ret = bch2_btree_node_get(c, iter, &tmp.k, level,
 825                                           SIX_LOCK_intent, may_drop_locks);
 826
 827                 /*
 828                  * before btree_iter_relock() calls btree_iter_verify_locks():
 829                  */
 830                 if (btree_lock_want(iter, level + 1) == BTREE_NODE_UNLOCKED)
 831                         btree_node_unlock(iter, level + 1);
 832
 833                 if (!bch2_btree_node_relock(iter, level)) {
 834                         btree_iter_set_dirty(iter, BTREE_ITER_NEED_RELOCK);
 835
 836                         if (!IS_ERR(ret)) {
 837                                 six_unlock_intent(&ret->lock);
 838                                 ret = ERR_PTR(-EINTR);
 839                         }
 840                 }
 841
 842                 bch2_btree_trans_relock(iter->trans);
 843         }
 844 out:
 845         if (btree_lock_want(iter, level + 1) == BTREE_NODE_UNLOCKED)
 846                 btree_node_unlock(iter, level + 1);
 847
 848         bch2_btree_trans_verify_locks(iter->trans);
 849
 850         BUG_ON((!may_drop_locks || !IS_ERR(ret)) &&
 851                (iter->uptodate >= BTREE_ITER_NEED_RELOCK ||
 852                 !btree_node_locked(iter, level)));
 853
 854         if (!IS_ERR_OR_NULL(ret)) {
 855                 struct btree *n1 = ret, *n2 = b;
 856
 857                 if (sib != btree_prev_sib)
 858                         swap(n1, n2);
 859
 860                 BUG_ON(bkey_cmp(btree_type_successor(n1->btree_id,
 861                                                      n1->key.k.p),
 862                                 n2->data->min_key));
 863         }
 864
 865         return ret;
 866 out_upgrade:
 867         if (may_drop_locks)
 868                 bch2_btree_iter_upgrade(iter, level + 2, true);
 869         ret = ERR_PTR(-EINTR);
 870         goto out;
 871 }
 872
 873 void bch2_btree_node_prefetch(struct bch_fs *c, struct btree_iter *iter,
 874                               const struct bkey_i *k, unsigned level)
 875 {
 876         struct btree_cache *bc = &c->btree_cache;
 877         struct btree *b;
 878
 879         BUG_ON(!btree_node_locked(iter, level + 1));
 880         BUG_ON(level >= BTREE_MAX_DEPTH);
 881
 882         rcu_read_lock();
 883         b = btree_cache_find(bc, k);
 884         rcu_read_unlock();
 885
 886         if (b)
 887                 return;
 888
 889         bch2_btree_node_fill(c, iter, k, level, SIX_LOCK_read, false);
 890 }
 891
 892 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
 893                              struct btree *b)
 894 {
 895         const struct bkey_format *f = &b->format;
 896         struct bset_stats stats;
 897
 898         memset(&stats, 0, sizeof(stats));
 899
 900         bch2_btree_keys_stats(b, &stats);
 901
 902         pr_buf(out,
 903                "l %u %llu:%llu - %llu:%llu:\n"
 904                "    ptrs: ",
 905                b->level,
 906                b->data->min_key.inode,
 907                b->data->min_key.offset,
 908                b->data->max_key.inode,
 909                b->data->max_key.offset);
 910         bch2_val_to_text(out, c, bkey_i_to_s_c(&b->key));
 911         pr_buf(out, "\n"
 912                "    format: u64s %u fields %u %u %u %u %u\n"
 913                "    unpack fn len: %u\n"
 914                "    bytes used %zu/%zu (%zu%% full)\n"
 915                "    sib u64s: %u, %u (merge threshold %zu)\n"
 916                "    nr packed keys %u\n"
 917                "    nr unpacked keys %u\n"
 918                "    floats %zu\n"
 919                "    failed unpacked %zu\n"
 920                "    failed prev %zu\n"
 921                "    failed overflow %zu\n",
 922                f->key_u64s,
 923                f->bits_per_field[0],
 924                f->bits_per_field[1],
 925                f->bits_per_field[2],
 926                f->bits_per_field[3],
 927                f->bits_per_field[4],
 928                b->unpack_fn_len,
 929                b->nr.live_u64s * sizeof(u64),
 930                btree_bytes(c) - sizeof(struct btree_node),
 931                b->nr.live_u64s * 100 / btree_max_u64s(c),
 932                b->sib_u64s[0],
 933                b->sib_u64s[1],
 934                BTREE_FOREGROUND_MERGE_THRESHOLD(c),
 935                b->nr.packed_keys,
 936                b->nr.unpacked_keys,
 937                stats.floats,
 938                stats.failed_unpacked,
 939                stats.failed_prev,
 940                stats.failed_overflow);
 941 }