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
  11 #include <linux/prefetch.h>
  12 #include <linux/sched/mm.h>
  13 #include <trace/events/bcachefs.h>
  14
  15 const char * const bch2_btree_ids[] = {
  16 #define x(kwd, val, name) name,
  17         BCH_BTREE_IDS()
  18 #undef x
  19         NULL
  20 };
  21
  22 void bch2_recalc_btree_reserve(struct bch_fs *c)
  23 {
  24         unsigned i, reserve = 16;
  25
  26         if (!c->btree_roots[0].b)
  27                 reserve += 8;
  28
  29         for (i = 0; i < BTREE_ID_NR; i++)
  30                 if (c->btree_roots[i].b)
  31                         reserve += min_t(unsigned, 1,
  32                                          c->btree_roots[i].b->c.level) * 8;
  33
  34         c->btree_cache.reserve = reserve;
  35 }
  36
  37 static inline unsigned btree_cache_can_free(struct btree_cache *bc)
  38 {
  39         return max_t(int, 0, bc->used - bc->reserve);
  40 }
  41
  42 static void __btree_node_data_free(struct bch_fs *c, struct btree *b)
  43 {
  44         EBUG_ON(btree_node_write_in_flight(b));
  45
  46         kvpfree(b->data, btree_bytes(c));
  47         b->data = NULL;
  48         vfree(b->aux_data);
  49         b->aux_data = NULL;
  50 }
  51
  52 static void btree_node_data_free(struct bch_fs *c, struct btree *b)
  53 {
  54         struct btree_cache *bc = &c->btree_cache;
  55
  56         __btree_node_data_free(c, b);
  57         bc->used--;
  58         list_move(&b->list, &bc->freed);
  59 }
  60
  61 static int bch2_btree_cache_cmp_fn(struct rhashtable_compare_arg *arg,
  62                                    const void *obj)
  63 {
  64         const struct btree *b = obj;
  65         const u64 *v = arg->key;
  66
  67         return b->hash_val == *v ? 0 : 1;
  68 }
  69
  70 static const struct rhashtable_params bch_btree_cache_params = {
  71         .head_offset    = offsetof(struct btree, hash),
  72         .key_offset     = offsetof(struct btree, hash_val),
  73         .key_len        = sizeof(u64),
  74         .obj_cmpfn      = bch2_btree_cache_cmp_fn,
  75 };
  76
  77 static int btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
  78 {
  79         BUG_ON(b->data || b->aux_data);
  80
  81         b->data = kvpmalloc(btree_bytes(c), gfp);
  82         if (!b->data)
  83                 return -ENOMEM;
  84
  85         b->aux_data = vmalloc_exec(btree_aux_data_bytes(b), gfp);
  86         if (!b->aux_data) {
  87                 kvpfree(b->data, btree_bytes(c));
  88                 b->data = NULL;
  89                 return -ENOMEM;
  90         }
  91
  92         return 0;
  93 }
  94
  95 static struct btree *__btree_node_mem_alloc(struct bch_fs *c)
  96 {
  97         struct btree *b = kzalloc(sizeof(struct btree), GFP_KERNEL);
  98         if (!b)
  99                 return NULL;
 100
 101         bkey_btree_ptr_init(&b->key);
 102         six_lock_init(&b->c.lock);
 103         INIT_LIST_HEAD(&b->list);
 104         INIT_LIST_HEAD(&b->write_blocked);
 105         b->byte_order = ilog2(btree_bytes(c));
 106         return b;
 107 }
 108
 109 static struct btree *btree_node_mem_alloc(struct bch_fs *c)
 110 {
 111         struct btree_cache *bc = &c->btree_cache;
 112         struct btree *b = __btree_node_mem_alloc(c);
 113         if (!b)
 114                 return NULL;
 115
 116         if (btree_node_data_alloc(c, b, GFP_KERNEL)) {
 117                 kfree(b);
 118                 return NULL;
 119         }
 120
 121         bc->used++;
 122         list_add(&b->list, &bc->freeable);
 123         return b;
 124 }
 125
 126 /* Btree in memory cache - hash table */
 127
 128 void bch2_btree_node_hash_remove(struct btree_cache *bc, struct btree *b)
 129 {
 130         rhashtable_remove_fast(&bc->table, &b->hash, bch_btree_cache_params);
 131
 132         /* Cause future lookups for this node to fail: */
 133         b->hash_val = 0;
 134
 135         six_lock_wakeup_all(&b->c.lock);
 136 }
 137
 138 int __bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b)
 139 {
 140         BUG_ON(b->hash_val);
 141         b->hash_val = btree_ptr_hash_val(&b->key);
 142
 143         return rhashtable_lookup_insert_fast(&bc->table, &b->hash,
 144                                              bch_btree_cache_params);
 145 }
 146
 147 int bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b,
 148                                 unsigned level, enum btree_id id)
 149 {
 150         int ret;
 151
 152         b->c.level      = level;
 153         b->c.btree_id   = id;
 154
 155         mutex_lock(&bc->lock);
 156         ret = __bch2_btree_node_hash_insert(bc, b);
 157         if (!ret)
 158                 list_add(&b->list, &bc->live);
 159         mutex_unlock(&bc->lock);
 160
 161         return ret;
 162 }
 163
 164 __flatten
 165 static inline struct btree *btree_cache_find(struct btree_cache *bc,
 166                                      const struct bkey_i *k)
 167 {
 168         u64 v = btree_ptr_hash_val(k);
 169
 170         return rhashtable_lookup_fast(&bc->table, &v, bch_btree_cache_params);
 171 }
 172
 173 /*
 174  * this version is for btree nodes that have already been freed (we're not
 175  * reaping a real btree node)
 176  */
 177 static int __btree_node_reclaim(struct bch_fs *c, struct btree *b, bool flush)
 178 {
 179         struct btree_cache *bc = &c->btree_cache;
 180         int ret = 0;
 181
 182         lockdep_assert_held(&bc->lock);
 183
 184         if (!six_trylock_intent(&b->c.lock))
 185                 return -ENOMEM;
 186
 187         if (!six_trylock_write(&b->c.lock))
 188                 goto out_unlock_intent;
 189
 190         if (btree_node_noevict(b))
 191                 goto out_unlock;
 192
 193         if (!btree_node_may_write(b))
 194                 goto out_unlock;
 195
 196         if (btree_node_dirty(b) &&
 197             test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags))
 198                 goto out_unlock;
 199
 200         if (btree_node_dirty(b) ||
 201             btree_node_write_in_flight(b) ||
 202             btree_node_read_in_flight(b)) {
 203                 if (!flush)
 204                         goto out_unlock;
 205
 206                 wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
 207                                TASK_UNINTERRUPTIBLE);
 208
 209                 /*
 210                  * Using the underscore version because we don't want to compact
 211                  * bsets after the write, since this node is about to be evicted
 212                  * - unless btree verify mode is enabled, since it runs out of
 213                  * the post write cleanup:
 214                  */
 215                 if (bch2_verify_btree_ondisk)
 216                         bch2_btree_node_write(c, b, SIX_LOCK_intent);
 217                 else
 218                         __bch2_btree_node_write(c, b, SIX_LOCK_read);
 219
 220                 /* wait for any in flight btree write */
 221                 btree_node_wait_on_io(b);
 222         }
 223 out:
 224         if (b->hash_val && !ret)
 225                 trace_btree_node_reap(c, b);
 226         return ret;
 227 out_unlock:
 228         six_unlock_write(&b->c.lock);
 229 out_unlock_intent:
 230         six_unlock_intent(&b->c.lock);
 231         ret = -ENOMEM;
 232         goto out;
 233 }
 234
 235 static int btree_node_reclaim(struct bch_fs *c, struct btree *b)
 236 {
 237         return __btree_node_reclaim(c, b, false);
 238 }
 239
 240 static int btree_node_write_and_reclaim(struct bch_fs *c, struct btree *b)
 241 {
 242         return __btree_node_reclaim(c, b, true);
 243 }
 244
 245 static unsigned long bch2_btree_cache_scan(struct shrinker *shrink,
 246                                            struct shrink_control *sc)
 247 {
 248         struct bch_fs *c = container_of(shrink, struct bch_fs,
 249                                         btree_cache.shrink);
 250         struct btree_cache *bc = &c->btree_cache;
 251         struct btree *b, *t;
 252         unsigned long nr = sc->nr_to_scan;
 253         unsigned long can_free;
 254         unsigned long touched = 0;
 255         unsigned long freed = 0;
 256         unsigned i, flags;
 257
 258         if (bch2_btree_shrinker_disabled)
 259                 return SHRINK_STOP;
 260
 261         /* Return -1 if we can't do anything right now */
 262         if (sc->gfp_mask & __GFP_FS)
 263                 mutex_lock(&bc->lock);
 264         else if (!mutex_trylock(&bc->lock))
 265                 return -1;
 266
 267         flags = memalloc_nofs_save();
 268
 269         /*
 270          * It's _really_ critical that we don't free too many btree nodes - we
 271          * have to always leave ourselves a reserve. The reserve is how we
 272          * guarantee that allocating memory for a new btree node can always
 273          * succeed, so that inserting keys into the btree can always succeed and
 274          * IO can always make forward progress:
 275          */
 276         nr /= btree_pages(c);
 277         can_free = btree_cache_can_free(bc);
 278         nr = min_t(unsigned long, nr, can_free);
 279
 280         i = 0;
 281         list_for_each_entry_safe(b, t, &bc->freeable, list) {
 282                 touched++;
 283
 284                 if (freed >= nr)
 285                         break;
 286
 287                 if (++i > 3 &&
 288                     !btree_node_reclaim(c, b)) {
 289                         btree_node_data_free(c, b);
 290                         six_unlock_write(&b->c.lock);
 291                         six_unlock_intent(&b->c.lock);
 292                         freed++;
 293                 }
 294         }
 295 restart:
 296         list_for_each_entry_safe(b, t, &bc->live, list) {
 297                 touched++;
 298
 299                 if (freed >= nr) {
 300                         /* Save position */
 301                         if (&t->list != &bc->live)
 302                                 list_move_tail(&bc->live, &t->list);
 303                         break;
 304                 }
 305
 306                 if (!btree_node_accessed(b) &&
 307                     !btree_node_reclaim(c, b)) {
 308                         /* can't call bch2_btree_node_hash_remove under lock  */
 309                         freed++;
 310                         if (&t->list != &bc->live)
 311                                 list_move_tail(&bc->live, &t->list);
 312
 313                         btree_node_data_free(c, b);
 314                         mutex_unlock(&bc->lock);
 315
 316                         bch2_btree_node_hash_remove(bc, b);
 317                         six_unlock_write(&b->c.lock);
 318                         six_unlock_intent(&b->c.lock);
 319
 320                         if (freed >= nr)
 321                                 goto out;
 322
 323                         if (sc->gfp_mask & __GFP_FS)
 324                                 mutex_lock(&bc->lock);
 325                         else if (!mutex_trylock(&bc->lock))
 326                                 goto out;
 327                         goto restart;
 328                 } else
 329                         clear_btree_node_accessed(b);
 330         }
 331
 332         mutex_unlock(&bc->lock);
 333 out:
 334         memalloc_nofs_restore(flags);
 335         return (unsigned long) freed * btree_pages(c);
 336 }
 337
 338 static unsigned long bch2_btree_cache_count(struct shrinker *shrink,
 339                                             struct shrink_control *sc)
 340 {
 341         struct bch_fs *c = container_of(shrink, struct bch_fs,
 342                                         btree_cache.shrink);
 343         struct btree_cache *bc = &c->btree_cache;
 344
 345         if (bch2_btree_shrinker_disabled)
 346                 return 0;
 347
 348         return btree_cache_can_free(bc) * btree_pages(c);
 349 }
 350
 351 void bch2_fs_btree_cache_exit(struct bch_fs *c)
 352 {
 353         struct btree_cache *bc = &c->btree_cache;
 354         struct btree *b;
 355         unsigned i, flags;
 356
 357         if (bc->shrink.list.next)
 358                 unregister_shrinker(&bc->shrink);
 359
 360         /* vfree() can allocate memory: */
 361         flags = memalloc_nofs_save();
 362         mutex_lock(&bc->lock);
 363
 364 #ifdef CONFIG_BCACHEFS_DEBUG
 365         if (c->verify_data)
 366                 list_move(&c->verify_data->list, &bc->live);
 367
 368         kvpfree(c->verify_ondisk, btree_bytes(c));
 369 #endif
 370
 371         for (i = 0; i < BTREE_ID_NR; i++)
 372                 if (c->btree_roots[i].b)
 373                         list_add(&c->btree_roots[i].b->list, &bc->live);
 374
 375         list_splice(&bc->freeable, &bc->live);
 376
 377         while (!list_empty(&bc->live)) {
 378                 b = list_first_entry(&bc->live, struct btree, list);
 379
 380                 BUG_ON(btree_node_read_in_flight(b) ||
 381                        btree_node_write_in_flight(b));
 382
 383                 if (btree_node_dirty(b))
 384                         bch2_btree_complete_write(c, b, btree_current_write(b));
 385                 clear_btree_node_dirty(c, b);
 386
 387                 btree_node_data_free(c, b);
 388         }
 389
 390         BUG_ON(atomic_read(&c->btree_cache.dirty));
 391
 392         while (!list_empty(&bc->freed)) {
 393                 b = list_first_entry(&bc->freed, struct btree, list);
 394                 list_del(&b->list);
 395                 kfree(b);
 396         }
 397
 398         mutex_unlock(&bc->lock);
 399         memalloc_nofs_restore(flags);
 400
 401         if (bc->table_init_done)
 402                 rhashtable_destroy(&bc->table);
 403 }
 404
 405 int bch2_fs_btree_cache_init(struct bch_fs *c)
 406 {
 407         struct btree_cache *bc = &c->btree_cache;
 408         unsigned i;
 409         int ret = 0;
 410
 411         pr_verbose_init(c->opts, "");
 412
 413         ret = rhashtable_init(&bc->table, &bch_btree_cache_params);
 414         if (ret)
 415                 goto out;
 416
 417         bc->table_init_done = true;
 418
 419         bch2_recalc_btree_reserve(c);
 420
 421         for (i = 0; i < bc->reserve; i++)
 422                 if (!btree_node_mem_alloc(c)) {
 423                         ret = -ENOMEM;
 424                         goto out;
 425                 }
 426
 427         list_splice_init(&bc->live, &bc->freeable);
 428
 429 #ifdef CONFIG_BCACHEFS_DEBUG
 430         mutex_init(&c->verify_lock);
 431
 432         c->verify_ondisk = kvpmalloc(btree_bytes(c), GFP_KERNEL);
 433         if (!c->verify_ondisk) {
 434                 ret = -ENOMEM;
 435                 goto out;
 436         }
 437
 438         c->verify_data = btree_node_mem_alloc(c);
 439         if (!c->verify_data) {
 440                 ret = -ENOMEM;
 441                 goto out;
 442         }
 443
 444         list_del_init(&c->verify_data->list);
 445 #endif
 446
 447         bc->shrink.count_objects        = bch2_btree_cache_count;
 448         bc->shrink.scan_objects         = bch2_btree_cache_scan;
 449         bc->shrink.seeks                = 4;
 450         bc->shrink.batch                = btree_pages(c) * 2;
 451         ret = register_shrinker(&bc->shrink);
 452 out:
 453         pr_verbose_init(c->opts, "ret %i", ret);
 454         return ret;
 455 }
 456
 457 void bch2_fs_btree_cache_init_early(struct btree_cache *bc)
 458 {
 459         mutex_init(&bc->lock);
 460         INIT_LIST_HEAD(&bc->live);
 461         INIT_LIST_HEAD(&bc->freeable);
 462         INIT_LIST_HEAD(&bc->freed);
 463 }
 464
 465 /*
 466  * We can only have one thread cannibalizing other cached btree nodes at a time,
 467  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 468  * cannibalize_bucket() will take. This means every time we unlock the root of
 469  * the btree, we need to release this lock if we have it held.
 470  */
 471 void bch2_btree_cache_cannibalize_unlock(struct bch_fs *c)
 472 {
 473         struct btree_cache *bc = &c->btree_cache;
 474
 475         if (bc->alloc_lock == current) {
 476                 trace_btree_node_cannibalize_unlock(c);
 477                 bc->alloc_lock = NULL;
 478                 closure_wake_up(&bc->alloc_wait);
 479         }
 480 }
 481
 482 int bch2_btree_cache_cannibalize_lock(struct bch_fs *c, struct closure *cl)
 483 {
 484         struct btree_cache *bc = &c->btree_cache;
 485         struct task_struct *old;
 486
 487         old = cmpxchg(&bc->alloc_lock, NULL, current);
 488         if (old == NULL || old == current)
 489                 goto success;
 490
 491         if (!cl) {
 492                 trace_btree_node_cannibalize_lock_fail(c);
 493                 return -ENOMEM;
 494         }
 495
 496         closure_wait(&bc->alloc_wait, cl);
 497
 498         /* Try again, after adding ourselves to waitlist */
 499         old = cmpxchg(&bc->alloc_lock, NULL, current);
 500         if (old == NULL || old == current) {
 501                 /* We raced */
 502                 closure_wake_up(&bc->alloc_wait);
 503                 goto success;
 504         }
 505
 506         trace_btree_node_cannibalize_lock_fail(c);
 507         return -EAGAIN;
 508
 509 success:
 510         trace_btree_node_cannibalize_lock(c);
 511         return 0;
 512 }
 513
 514 static struct btree *btree_node_cannibalize(struct bch_fs *c)
 515 {
 516         struct btree_cache *bc = &c->btree_cache;
 517         struct btree *b;
 518
 519         list_for_each_entry_reverse(b, &bc->live, list)
 520                 if (!btree_node_reclaim(c, b))
 521                         return b;
 522
 523         while (1) {
 524                 list_for_each_entry_reverse(b, &bc->live, list)
 525                         if (!btree_node_write_and_reclaim(c, b))
 526                                 return b;
 527
 528                 /*
 529                  * Rare case: all nodes were intent-locked.
 530                  * Just busy-wait.
 531                  */
 532                 WARN_ONCE(1, "btree cache cannibalize failed\n");
 533                 cond_resched();
 534         }
 535 }
 536
 537 struct btree *bch2_btree_node_mem_alloc(struct bch_fs *c)
 538 {
 539         struct btree_cache *bc = &c->btree_cache;
 540         struct btree *b;
 541         u64 start_time = local_clock();
 542         unsigned flags;
 543
 544         flags = memalloc_nofs_save();
 545         mutex_lock(&bc->lock);
 546
 547         /*
 548          * btree_free() doesn't free memory; it sticks the node on the end of
 549          * the list. Check if there's any freed nodes there:
 550          */
 551         list_for_each_entry(b, &bc->freeable, list)
 552                 if (!btree_node_reclaim(c, b))
 553                         goto got_node;
 554
 555         /*
 556          * We never free struct btree itself, just the memory that holds the on
 557          * disk node. Check the freed list before allocating a new one:
 558          */
 559         list_for_each_entry(b, &bc->freed, list)
 560                 if (!btree_node_reclaim(c, b))
 561                         goto got_node;
 562
 563         b = NULL;
 564 got_node:
 565         if (b)
 566                 list_del_init(&b->list);
 567         mutex_unlock(&bc->lock);
 568
 569         if (!b) {
 570                 b = __btree_node_mem_alloc(c);
 571                 if (!b)
 572                         goto err;
 573
 574                 BUG_ON(!six_trylock_intent(&b->c.lock));
 575                 BUG_ON(!six_trylock_write(&b->c.lock));
 576         }
 577
 578         if (!b->data) {
 579                 if (btree_node_data_alloc(c, b, __GFP_NOWARN|GFP_KERNEL))
 580                         goto err;
 581
 582                 mutex_lock(&bc->lock);
 583                 bc->used++;
 584                 mutex_unlock(&bc->lock);
 585         }
 586
 587         BUG_ON(btree_node_hashed(b));
 588         BUG_ON(btree_node_write_in_flight(b));
 589 out:
 590         b->flags                = 0;
 591         b->written              = 0;
 592         b->nsets                = 0;
 593         b->sib_u64s[0]          = 0;
 594         b->sib_u64s[1]          = 0;
 595         b->whiteout_u64s        = 0;
 596         bch2_btree_keys_init(b);
 597
 598         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],
 599                                start_time);
 600
 601         memalloc_nofs_restore(flags);
 602         return b;
 603 err:
 604         mutex_lock(&bc->lock);
 605
 606         if (b) {
 607                 list_add(&b->list, &bc->freed);
 608                 six_unlock_write(&b->c.lock);
 609                 six_unlock_intent(&b->c.lock);
 610         }
 611
 612         /* Try to cannibalize another cached btree node: */
 613         if (bc->alloc_lock == current) {
 614                 b = btree_node_cannibalize(c);
 615                 list_del_init(&b->list);
 616                 mutex_unlock(&bc->lock);
 617
 618                 bch2_btree_node_hash_remove(bc, b);
 619
 620                 trace_btree_node_cannibalize(c);
 621                 goto out;
 622         }
 623
 624         mutex_unlock(&bc->lock);
 625         memalloc_nofs_restore(flags);
 626         return ERR_PTR(-ENOMEM);
 627 }
 628
 629 /* Slowpath, don't want it inlined into btree_iter_traverse() */
 630 static noinline struct btree *bch2_btree_node_fill(struct bch_fs *c,
 631                                 struct btree_iter *iter,
 632                                 const struct bkey_i *k,
 633                                 enum btree_id btree_id,
 634                                 unsigned level,
 635                                 enum six_lock_type lock_type,
 636                                 bool sync)
 637 {
 638         struct btree_cache *bc = &c->btree_cache;
 639         struct btree *b;
 640
 641         BUG_ON(level + 1 >= BTREE_MAX_DEPTH);
 642         /*
 643          * Parent node must be locked, else we could read in a btree node that's
 644          * been freed:
 645          */
 646         if (iter && !bch2_btree_node_relock(iter, level + 1))
 647                 return ERR_PTR(-EINTR);
 648
 649         b = bch2_btree_node_mem_alloc(c);
 650         if (IS_ERR(b))
 651                 return b;
 652
 653         bkey_copy(&b->key, k);
 654         if (bch2_btree_node_hash_insert(bc, b, level, btree_id)) {
 655                 /* raced with another fill: */
 656
 657                 /* mark as unhashed... */
 658                 b->hash_val = 0;
 659
 660                 mutex_lock(&bc->lock);
 661                 list_add(&b->list, &bc->freeable);
 662                 mutex_unlock(&bc->lock);
 663
 664                 six_unlock_write(&b->c.lock);
 665                 six_unlock_intent(&b->c.lock);
 666                 return NULL;
 667         }
 668
 669         /*
 670          * Unlock before doing IO:
 671          *
 672          * XXX: ideally should be dropping all btree node locks here
 673          */
 674         if (iter && btree_node_read_locked(iter, level + 1))
 675                 btree_node_unlock(iter, level + 1);
 676
 677         bch2_btree_node_read(c, b, sync);
 678
 679         six_unlock_write(&b->c.lock);
 680
 681         if (!sync) {
 682                 six_unlock_intent(&b->c.lock);
 683                 return NULL;
 684         }
 685
 686         if (lock_type == SIX_LOCK_read)
 687                 six_lock_downgrade(&b->c.lock);
 688
 689         return b;
 690 }
 691
 692 static int lock_node_check_fn(struct six_lock *lock, void *p)
 693 {
 694         struct btree *b = container_of(lock, struct btree, c.lock);
 695         const struct bkey_i *k = p;
 696
 697         return b->hash_val == btree_ptr_hash_val(k) ? 0 : -1;
 698 }
 699
 700 /**
 701  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 702  * in from disk if necessary.
 703  *
 704  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
 705  *
 706  * The btree node will have either a read or a write lock held, depending on
 707  * the @write parameter.
 708  */
 709 struct btree *bch2_btree_node_get(struct bch_fs *c, struct btree_iter *iter,
 710                                   const struct bkey_i *k, unsigned level,
 711                                   enum six_lock_type lock_type,
 712                                   unsigned long trace_ip)
 713 {
 714         struct btree_cache *bc = &c->btree_cache;
 715         struct btree *b;
 716         struct bset_tree *t;
 717
 718         EBUG_ON(level >= BTREE_MAX_DEPTH);
 719
 720         b = btree_node_mem_ptr(k);
 721         if (b)
 722                 goto lock_node;
 723 retry:
 724         b = btree_cache_find(bc, k);
 725         if (unlikely(!b)) {
 726                 /*
 727                  * We must have the parent locked to call bch2_btree_node_fill(),
 728                  * else we could read in a btree node from disk that's been
 729                  * freed:
 730                  */
 731                 b = bch2_btree_node_fill(c, iter, k, iter->btree_id,
 732                                          level, lock_type, true);
 733
 734                 /* We raced and found the btree node in the cache */
 735                 if (!b)
 736                         goto retry;
 737
 738                 if (IS_ERR(b))
 739                         return b;
 740         } else {
 741 lock_node:
 742                 /*
 743                  * There's a potential deadlock with splits and insertions into
 744                  * interior nodes we have to avoid:
 745                  *
 746                  * The other thread might be holding an intent lock on the node
 747                  * we want, and they want to update its parent node so they're
 748                  * going to upgrade their intent lock on the parent node to a
 749                  * write lock.
 750                  *
 751                  * But if we're holding a read lock on the parent, and we're
 752                  * trying to get the intent lock they're holding, we deadlock.
 753                  *
 754                  * So to avoid this we drop the read locks on parent nodes when
 755                  * we're starting to take intent locks - and handle the race.
 756                  *
 757                  * The race is that they might be about to free the node we
 758                  * want, and dropping our read lock on the parent node lets them
 759                  * update the parent marking the node we want as freed, and then
 760                  * free it:
 761                  *
 762                  * To guard against this, btree nodes are evicted from the cache
 763                  * when they're freed - and b->hash_val is zeroed out, which we
 764                  * check for after we lock the node.
 765                  *
 766                  * Then, bch2_btree_node_relock() on the parent will fail - because
 767                  * the parent was modified, when the pointer to the node we want
 768                  * was removed - and we'll bail out:
 769                  */
 770                 if (btree_node_read_locked(iter, level + 1))
 771                         btree_node_unlock(iter, level + 1);
 772
 773                 if (!btree_node_lock(b, k->k.p, level, iter, lock_type,
 774                                      lock_node_check_fn, (void *) k, trace_ip)) {
 775                         if (b->hash_val != btree_ptr_hash_val(k))
 776                                 goto retry;
 777                         return ERR_PTR(-EINTR);
 778                 }
 779
 780                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
 781                              b->c.level != level ||
 782                              race_fault())) {
 783                         six_unlock_type(&b->c.lock, lock_type);
 784                         if (bch2_btree_node_relock(iter, level + 1))
 785                                 goto retry;
 786
 787                         trace_trans_restart_btree_node_reused(iter->trans->ip);
 788                         return ERR_PTR(-EINTR);
 789                 }
 790         }
 791
 792         /* XXX: waiting on IO with btree locks held: */
 793         wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
 794                        TASK_UNINTERRUPTIBLE);
 795
 796         prefetch(b->aux_data);
 797
 798         for_each_bset(b, t) {
 799                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 800
 801                 prefetch(p + L1_CACHE_BYTES * 0);
 802                 prefetch(p + L1_CACHE_BYTES * 1);
 803                 prefetch(p + L1_CACHE_BYTES * 2);
 804         }
 805
 806         /* avoid atomic set bit if it's not needed: */
 807         if (!btree_node_accessed(b))
 808                 set_btree_node_accessed(b);
 809
 810         if (unlikely(btree_node_read_error(b))) {
 811                 six_unlock_type(&b->c.lock, lock_type);
 812                 return ERR_PTR(-EIO);
 813         }
 814
 815         EBUG_ON(b->c.btree_id != iter->btree_id ||
 816                 BTREE_NODE_LEVEL(b->data) != level ||
 817                 bkey_cmp(b->data->max_key, k->k.p));
 818
 819         return b;
 820 }
 821
 822 struct btree *bch2_btree_node_get_noiter(struct bch_fs *c,
 823                                          const struct bkey_i *k,
 824                                          enum btree_id btree_id,
 825                                          unsigned level)
 826 {
 827         struct btree_cache *bc = &c->btree_cache;
 828         struct btree *b;
 829         struct bset_tree *t;
 830         int ret;
 831
 832         EBUG_ON(level >= BTREE_MAX_DEPTH);
 833
 834         b = btree_node_mem_ptr(k);
 835         if (b)
 836                 goto lock_node;
 837 retry:
 838         b = btree_cache_find(bc, k);
 839         if (unlikely(!b)) {
 840                 b = bch2_btree_node_fill(c, NULL, k, btree_id,
 841                                          level, SIX_LOCK_read, true);
 842
 843                 /* We raced and found the btree node in the cache */
 844                 if (!b)
 845                         goto retry;
 846
 847                 if (IS_ERR(b))
 848                         return b;
 849         } else {
 850 lock_node:
 851                 ret = six_lock_read(&b->c.lock, lock_node_check_fn, (void *) k);
 852                 if (ret)
 853                         goto retry;
 854
 855                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
 856                              b->c.btree_id != btree_id ||
 857                              b->c.level != level)) {
 858                         six_unlock_read(&b->c.lock);
 859                         goto retry;
 860                 }
 861         }
 862
 863         /* XXX: waiting on IO with btree locks held: */
 864         wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
 865                        TASK_UNINTERRUPTIBLE);
 866
 867         prefetch(b->aux_data);
 868
 869         for_each_bset(b, t) {
 870                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 871
 872                 prefetch(p + L1_CACHE_BYTES * 0);
 873                 prefetch(p + L1_CACHE_BYTES * 1);
 874                 prefetch(p + L1_CACHE_BYTES * 2);
 875         }
 876
 877         /* avoid atomic set bit if it's not needed: */
 878         if (!btree_node_accessed(b))
 879                 set_btree_node_accessed(b);
 880
 881         if (unlikely(btree_node_read_error(b))) {
 882                 six_unlock_read(&b->c.lock);
 883                 return ERR_PTR(-EIO);
 884         }
 885
 886         EBUG_ON(b->c.btree_id != btree_id ||
 887                 BTREE_NODE_LEVEL(b->data) != level ||
 888                 bkey_cmp(b->data->max_key, k->k.p));
 889
 890         return b;
 891 }
 892
 893 struct btree *bch2_btree_node_get_sibling(struct bch_fs *c,
 894                                           struct btree_iter *iter,
 895                                           struct btree *b,
 896                                           enum btree_node_sibling sib)
 897 {
 898         struct btree_trans *trans = iter->trans;
 899         struct btree *parent;
 900         struct btree_node_iter node_iter;
 901         struct bkey_packed *k;
 902         struct bkey_buf tmp;
 903         struct btree *ret = NULL;
 904         unsigned level = b->c.level;
 905
 906         bch2_bkey_buf_init(&tmp);
 907
 908         parent = btree_iter_node(iter, level + 1);
 909         if (!parent)
 910                 return NULL;
 911
 912         /*
 913          * There's a corner case where a btree_iter might have a node locked
 914          * that is just outside its current pos - when
 915          * bch2_btree_iter_set_pos_same_leaf() gets to the end of the node.
 916          *
 917          * But the lock ordering checks in __bch2_btree_node_lock() go off of
 918          * iter->pos, not the node's key: so if the iterator is marked as
 919          * needing to be traversed, we risk deadlock if we don't bail out here:
 920          */
 921         if (iter->uptodate >= BTREE_ITER_NEED_TRAVERSE)
 922                 return ERR_PTR(-EINTR);
 923
 924         if (!bch2_btree_node_relock(iter, level + 1)) {
 925                 ret = ERR_PTR(-EINTR);
 926                 goto out;
 927         }
 928
 929         node_iter = iter->l[parent->c.level].iter;
 930
 931         k = bch2_btree_node_iter_peek_all(&node_iter, parent);
 932         BUG_ON(bkey_cmp_left_packed(parent, k, &b->key.k.p));
 933
 934         k = sib == btree_prev_sib
 935                 ? bch2_btree_node_iter_prev(&node_iter, parent)
 936                 : (bch2_btree_node_iter_advance(&node_iter, parent),
 937                    bch2_btree_node_iter_peek(&node_iter, parent));
 938         if (!k)
 939                 goto out;
 940
 941         bch2_bkey_buf_unpack(&tmp, c, parent, k);
 942
 943         ret = bch2_btree_node_get(c, iter, tmp.k, level,
 944                                   SIX_LOCK_intent, _THIS_IP_);
 945
 946         if (PTR_ERR_OR_ZERO(ret) == -EINTR && !trans->nounlock) {
 947                 struct btree_iter *linked;
 948
 949                 if (!bch2_btree_node_relock(iter, level + 1))
 950                         goto out;
 951
 952                 /*
 953                  * We might have got -EINTR because trylock failed, and we're
 954                  * holding other locks that would cause us to deadlock:
 955                  */
 956                 trans_for_each_iter(trans, linked)
 957                         if (btree_iter_lock_cmp(iter, linked) < 0)
 958                                 __bch2_btree_iter_unlock(linked);
 959
 960                 if (sib == btree_prev_sib)
 961                         btree_node_unlock(iter, level);
 962
 963                 ret = bch2_btree_node_get(c, iter, tmp.k, level,
 964                                           SIX_LOCK_intent, _THIS_IP_);
 965
 966                 /*
 967                  * before btree_iter_relock() calls btree_iter_verify_locks():
 968                  */
 969                 if (btree_lock_want(iter, level + 1) == BTREE_NODE_UNLOCKED)
 970                         btree_node_unlock(iter, level + 1);
 971
 972                 if (!bch2_btree_node_relock(iter, level)) {
 973                         btree_iter_set_dirty(iter, BTREE_ITER_NEED_RELOCK);
 974
 975                         if (!IS_ERR(ret)) {
 976                                 six_unlock_intent(&ret->c.lock);
 977                                 ret = ERR_PTR(-EINTR);
 978                         }
 979                 }
 980
 981                 bch2_trans_relock(trans);
 982         }
 983 out:
 984         if (btree_lock_want(iter, level + 1) == BTREE_NODE_UNLOCKED)
 985                 btree_node_unlock(iter, level + 1);
 986
 987         if (PTR_ERR_OR_ZERO(ret) == -EINTR)
 988                 bch2_btree_iter_upgrade(iter, level + 2);
 989
 990         BUG_ON(!IS_ERR(ret) && !btree_node_locked(iter, level));
 991
 992         if (!IS_ERR_OR_NULL(ret)) {
 993                 struct btree *n1 = ret, *n2 = b;
 994
 995                 if (sib != btree_prev_sib)
 996                         swap(n1, n2);
 997
 998                 BUG_ON(bkey_cmp(bkey_successor(n1->key.k.p),
 999                                 n2->data->min_key));
1000         }
1001
1002         bch2_btree_trans_verify_locks(trans);
1003
1004         bch2_bkey_buf_exit(&tmp, c);
1005
1006         return ret;
1007 }
1008
1009 void bch2_btree_node_prefetch(struct bch_fs *c, struct btree_iter *iter,
1010                               const struct bkey_i *k,
1011                               enum btree_id btree_id, unsigned level)
1012 {
1013         struct btree_cache *bc = &c->btree_cache;
1014         struct btree *b;
1015
1016         BUG_ON(iter && !btree_node_locked(iter, level + 1));
1017         BUG_ON(level >= BTREE_MAX_DEPTH);
1018
1019         b = btree_cache_find(bc, k);
1020         if (b)
1021                 return;
1022
1023         bch2_btree_node_fill(c, iter, k, btree_id, level, SIX_LOCK_read, false);
1024 }
1025
1026 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
1027                              struct btree *b)
1028 {
1029         const struct bkey_format *f = &b->format;
1030         struct bset_stats stats;
1031
1032         memset(&stats, 0, sizeof(stats));
1033
1034         bch2_btree_keys_stats(b, &stats);
1035
1036         pr_buf(out,
1037                "l %u %llu:%llu - %llu:%llu:\n"
1038                "    ptrs: ",
1039                b->c.level,
1040                b->data->min_key.inode,
1041                b->data->min_key.offset,
1042                b->data->max_key.inode,
1043                b->data->max_key.offset);
1044         bch2_val_to_text(out, c, bkey_i_to_s_c(&b->key));
1045         pr_buf(out, "\n"
1046                "    format: u64s %u fields %u %u %u %u %u\n"
1047                "    unpack fn len: %u\n"
1048                "    bytes used %zu/%zu (%zu%% full)\n"
1049                "    sib u64s: %u, %u (merge threshold %zu)\n"
1050                "    nr packed keys %u\n"
1051                "    nr unpacked keys %u\n"
1052                "    floats %zu\n"
1053                "    failed unpacked %zu\n",
1054                f->key_u64s,
1055                f->bits_per_field[0],
1056                f->bits_per_field[1],
1057                f->bits_per_field[2],
1058                f->bits_per_field[3],
1059                f->bits_per_field[4],
1060                b->unpack_fn_len,
1061                b->nr.live_u64s * sizeof(u64),
1062                btree_bytes(c) - sizeof(struct btree_node),
1063                b->nr.live_u64s * 100 / btree_max_u64s(c),
1064                b->sib_u64s[0],
1065                b->sib_u64s[1],
1066                BTREE_FOREGROUND_MERGE_THRESHOLD(c),
1067                b->nr.packed_keys,
1068                b->nr.unpacked_keys,
1069                stats.floats,
1070                stats.failed);
1071 }
1072
1073 void bch2_btree_cache_to_text(struct printbuf *out, struct bch_fs *c)
1074 {
1075         pr_buf(out, "nr nodes:\t\t%u\n", c->btree_cache.used);
1076         pr_buf(out, "nr dirty:\t\t%u\n", atomic_read(&c->btree_cache.dirty));
1077         pr_buf(out, "cannibalize lock:\t%p\n", c->btree_cache.alloc_lock);
1078 }