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 "errcode.h"
  11 #include "error.h"
  12
  13 #include <linux/prefetch.h>
  14 #include <linux/sched/mm.h>
  15 #include <trace/events/bcachefs.h>
  16
  17 struct lock_class_key bch2_btree_node_lock_key;
  18
  19 const char * const bch2_btree_node_flags[] = {
  20 #define x(f)    #f,
  21         BTREE_FLAGS()
  22 #undef x
  23         NULL
  24 };
  25
  26 void bch2_recalc_btree_reserve(struct bch_fs *c)
  27 {
  28         unsigned i, reserve = 16;
  29
  30         if (!c->btree_roots[0].b)
  31                 reserve += 8;
  32
  33         for (i = 0; i < BTREE_ID_NR; i++)
  34                 if (c->btree_roots[i].b)
  35                         reserve += min_t(unsigned, 1,
  36                                          c->btree_roots[i].b->c.level) * 8;
  37
  38         c->btree_cache.reserve = reserve;
  39 }
  40
  41 static inline unsigned btree_cache_can_free(struct btree_cache *bc)
  42 {
  43         return max_t(int, 0, bc->used - bc->reserve);
  44 }
  45
  46 static void btree_node_to_freedlist(struct btree_cache *bc, struct btree *b)
  47 {
  48         if (b->c.lock.readers)
  49                 list_move(&b->list, &bc->freed_pcpu);
  50         else
  51                 list_move(&b->list, &bc->freed_nonpcpu);
  52 }
  53
  54 static void btree_node_data_free(struct bch_fs *c, struct btree *b)
  55 {
  56         struct btree_cache *bc = &c->btree_cache;
  57
  58         EBUG_ON(btree_node_write_in_flight(b));
  59
  60         kvpfree(b->data, btree_bytes(c));
  61         b->data = NULL;
  62 #ifdef __KERNEL__
  63         vfree(b->aux_data);
  64 #else
  65         munmap(b->aux_data, btree_aux_data_bytes(b));
  66 #endif
  67         b->aux_data = NULL;
  68
  69         bc->used--;
  70
  71         btree_node_to_freedlist(bc, b);
  72 }
  73
  74 static int bch2_btree_cache_cmp_fn(struct rhashtable_compare_arg *arg,
  75                                    const void *obj)
  76 {
  77         const struct btree *b = obj;
  78         const u64 *v = arg->key;
  79
  80         return b->hash_val == *v ? 0 : 1;
  81 }
  82
  83 static const struct rhashtable_params bch_btree_cache_params = {
  84         .head_offset    = offsetof(struct btree, hash),
  85         .key_offset     = offsetof(struct btree, hash_val),
  86         .key_len        = sizeof(u64),
  87         .obj_cmpfn      = bch2_btree_cache_cmp_fn,
  88 };
  89
  90 static int btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
  91 {
  92         BUG_ON(b->data || b->aux_data);
  93
  94         b->data = kvpmalloc(btree_bytes(c), gfp);
  95         if (!b->data)
  96                 return -ENOMEM;
  97 #ifdef __KERNEL__
  98         b->aux_data = vmalloc_exec(btree_aux_data_bytes(b), gfp);
  99 #else
 100         b->aux_data = mmap(NULL, btree_aux_data_bytes(b),
 101                            PROT_READ|PROT_WRITE|PROT_EXEC,
 102                            MAP_PRIVATE|MAP_ANONYMOUS, 0, 0);
 103         if (b->aux_data == MAP_FAILED)
 104                 b->aux_data = NULL;
 105 #endif
 106         if (!b->aux_data) {
 107                 kvpfree(b->data, btree_bytes(c));
 108                 b->data = NULL;
 109                 return -ENOMEM;
 110         }
 111
 112         return 0;
 113 }
 114
 115 static struct btree *__btree_node_mem_alloc(struct bch_fs *c)
 116 {
 117         struct btree *b = kzalloc(sizeof(struct btree), GFP_KERNEL);
 118         if (!b)
 119                 return NULL;
 120
 121         bkey_btree_ptr_init(&b->key);
 122         __six_lock_init(&b->c.lock, "b->c.lock", &bch2_btree_node_lock_key);
 123         INIT_LIST_HEAD(&b->list);
 124         INIT_LIST_HEAD(&b->write_blocked);
 125         b->byte_order = ilog2(btree_bytes(c));
 126         return b;
 127 }
 128
 129 struct btree *__bch2_btree_node_mem_alloc(struct bch_fs *c)
 130 {
 131         struct btree_cache *bc = &c->btree_cache;
 132         struct btree *b = __btree_node_mem_alloc(c);
 133         if (!b)
 134                 return NULL;
 135
 136         if (btree_node_data_alloc(c, b, GFP_KERNEL)) {
 137                 kfree(b);
 138                 return NULL;
 139         }
 140
 141         bc->used++;
 142         list_add(&b->list, &bc->freeable);
 143         return b;
 144 }
 145
 146 /* Btree in memory cache - hash table */
 147
 148 void bch2_btree_node_hash_remove(struct btree_cache *bc, struct btree *b)
 149 {
 150         int ret = rhashtable_remove_fast(&bc->table, &b->hash, bch_btree_cache_params);
 151         BUG_ON(ret);
 152
 153         /* Cause future lookups for this node to fail: */
 154         b->hash_val = 0;
 155
 156         six_lock_wakeup_all(&b->c.lock);
 157 }
 158
 159 int __bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b)
 160 {
 161         BUG_ON(b->hash_val);
 162         b->hash_val = btree_ptr_hash_val(&b->key);
 163
 164         return rhashtable_lookup_insert_fast(&bc->table, &b->hash,
 165                                              bch_btree_cache_params);
 166 }
 167
 168 int bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b,
 169                                 unsigned level, enum btree_id id)
 170 {
 171         int ret;
 172
 173         b->c.level      = level;
 174         b->c.btree_id   = id;
 175
 176         mutex_lock(&bc->lock);
 177         ret = __bch2_btree_node_hash_insert(bc, b);
 178         if (!ret)
 179                 list_add(&b->list, &bc->live);
 180         mutex_unlock(&bc->lock);
 181
 182         return ret;
 183 }
 184
 185 __flatten
 186 static inline struct btree *btree_cache_find(struct btree_cache *bc,
 187                                      const struct bkey_i *k)
 188 {
 189         u64 v = btree_ptr_hash_val(k);
 190
 191         return rhashtable_lookup_fast(&bc->table, &v, bch_btree_cache_params);
 192 }
 193
 194 /*
 195  * this version is for btree nodes that have already been freed (we're not
 196  * reaping a real btree node)
 197  */
 198 static int __btree_node_reclaim(struct bch_fs *c, struct btree *b, bool flush)
 199 {
 200         struct btree_cache *bc = &c->btree_cache;
 201         int ret = 0;
 202
 203         lockdep_assert_held(&bc->lock);
 204 wait_on_io:
 205         if (b->flags & ((1U << BTREE_NODE_dirty)|
 206                         (1U << BTREE_NODE_read_in_flight)|
 207                         (1U << BTREE_NODE_write_in_flight))) {
 208                 if (!flush)
 209                         return -ENOMEM;
 210
 211                 /* XXX: waiting on IO with btree cache lock held */
 212                 bch2_btree_node_wait_on_read(b);
 213                 bch2_btree_node_wait_on_write(b);
 214         }
 215
 216         if (!six_trylock_intent(&b->c.lock))
 217                 return -ENOMEM;
 218
 219         if (!six_trylock_write(&b->c.lock))
 220                 goto out_unlock_intent;
 221
 222         /* recheck under lock */
 223         if (b->flags & ((1U << BTREE_NODE_read_in_flight)|
 224                         (1U << BTREE_NODE_write_in_flight))) {
 225                 if (!flush)
 226                         goto out_unlock;
 227                 six_unlock_write(&b->c.lock);
 228                 six_unlock_intent(&b->c.lock);
 229                 goto wait_on_io;
 230         }
 231
 232         if (btree_node_noevict(b) ||
 233             btree_node_write_blocked(b) ||
 234             btree_node_will_make_reachable(b))
 235                 goto out_unlock;
 236
 237         if (btree_node_dirty(b)) {
 238                 if (!flush)
 239                         goto out_unlock;
 240                 /*
 241                  * Using the underscore version because we don't want to compact
 242                  * bsets after the write, since this node is about to be evicted
 243                  * - unless btree verify mode is enabled, since it runs out of
 244                  * the post write cleanup:
 245                  */
 246                 if (bch2_verify_btree_ondisk)
 247                         bch2_btree_node_write(c, b, SIX_LOCK_intent, 0);
 248                 else
 249                         __bch2_btree_node_write(c, b, 0);
 250
 251                 six_unlock_write(&b->c.lock);
 252                 six_unlock_intent(&b->c.lock);
 253                 goto wait_on_io;
 254         }
 255 out:
 256         if (b->hash_val && !ret)
 257                 trace_btree_node_reap(c, b);
 258         return ret;
 259 out_unlock:
 260         six_unlock_write(&b->c.lock);
 261 out_unlock_intent:
 262         six_unlock_intent(&b->c.lock);
 263         ret = -ENOMEM;
 264         goto out;
 265 }
 266
 267 static int btree_node_reclaim(struct bch_fs *c, struct btree *b)
 268 {
 269         return __btree_node_reclaim(c, b, false);
 270 }
 271
 272 static int btree_node_write_and_reclaim(struct bch_fs *c, struct btree *b)
 273 {
 274         return __btree_node_reclaim(c, b, true);
 275 }
 276
 277 static unsigned long bch2_btree_cache_scan(struct shrinker *shrink,
 278                                            struct shrink_control *sc)
 279 {
 280         struct bch_fs *c = container_of(shrink, struct bch_fs,
 281                                         btree_cache.shrink);
 282         struct btree_cache *bc = &c->btree_cache;
 283         struct btree *b, *t;
 284         unsigned long nr = sc->nr_to_scan;
 285         unsigned long can_free = 0;
 286         unsigned long touched = 0;
 287         unsigned long freed = 0;
 288         unsigned i, flags;
 289         unsigned long ret = SHRINK_STOP;
 290
 291         if (bch2_btree_shrinker_disabled)
 292                 return SHRINK_STOP;
 293
 294         /* Return -1 if we can't do anything right now */
 295         if (sc->gfp_mask & __GFP_FS)
 296                 mutex_lock(&bc->lock);
 297         else if (!mutex_trylock(&bc->lock))
 298                 goto out_norestore;
 299
 300         flags = memalloc_nofs_save();
 301
 302         /*
 303          * It's _really_ critical that we don't free too many btree nodes - we
 304          * have to always leave ourselves a reserve. The reserve is how we
 305          * guarantee that allocating memory for a new btree node can always
 306          * succeed, so that inserting keys into the btree can always succeed and
 307          * IO can always make forward progress:
 308          */
 309         can_free = btree_cache_can_free(bc);
 310         nr = min_t(unsigned long, nr, can_free);
 311
 312         i = 0;
 313         list_for_each_entry_safe(b, t, &bc->freeable, list) {
 314                 /*
 315                  * Leave a few nodes on the freeable list, so that a btree split
 316                  * won't have to hit the system allocator:
 317                  */
 318                 if (++i <= 3)
 319                         continue;
 320
 321                 touched++;
 322
 323                 if (touched >= nr)
 324                         break;
 325
 326                 if (!btree_node_reclaim(c, b)) {
 327                         btree_node_data_free(c, b);
 328                         six_unlock_write(&b->c.lock);
 329                         six_unlock_intent(&b->c.lock);
 330                         freed++;
 331                 }
 332         }
 333 restart:
 334         list_for_each_entry_safe(b, t, &bc->live, list) {
 335                 /* tweak this */
 336                 if (btree_node_accessed(b)) {
 337                         clear_btree_node_accessed(b);
 338                         goto touched;
 339                 }
 340
 341                 if (!btree_node_reclaim(c, b)) {
 342                         /* can't call bch2_btree_node_hash_remove under lock  */
 343                         freed++;
 344                         if (&t->list != &bc->live)
 345                                 list_move_tail(&bc->live, &t->list);
 346
 347                         btree_node_data_free(c, b);
 348                         mutex_unlock(&bc->lock);
 349
 350                         bch2_btree_node_hash_remove(bc, b);
 351                         six_unlock_write(&b->c.lock);
 352                         six_unlock_intent(&b->c.lock);
 353
 354                         if (freed >= nr)
 355                                 goto out;
 356
 357                         if (sc->gfp_mask & __GFP_FS)
 358                                 mutex_lock(&bc->lock);
 359                         else if (!mutex_trylock(&bc->lock))
 360                                 goto out;
 361                         goto restart;
 362                 } else {
 363                         continue;
 364                 }
 365 touched:
 366                 touched++;
 367
 368                 if (touched >= nr) {
 369                         /* Save position */
 370                         if (&t->list != &bc->live)
 371                                 list_move_tail(&bc->live, &t->list);
 372                         break;
 373                 }
 374         }
 375
 376         mutex_unlock(&bc->lock);
 377 out:
 378         ret = freed;
 379         memalloc_nofs_restore(flags);
 380 out_norestore:
 381         trace_btree_cache_scan(sc->nr_to_scan, can_free, ret);
 382         return ret;
 383 }
 384
 385 static unsigned long bch2_btree_cache_count(struct shrinker *shrink,
 386                                             struct shrink_control *sc)
 387 {
 388         struct bch_fs *c = container_of(shrink, struct bch_fs,
 389                                         btree_cache.shrink);
 390         struct btree_cache *bc = &c->btree_cache;
 391
 392         if (bch2_btree_shrinker_disabled)
 393                 return 0;
 394
 395         return btree_cache_can_free(bc);
 396 }
 397
 398 static void bch2_btree_cache_shrinker_to_text(struct printbuf *out, struct shrinker *shrink)
 399 {
 400         struct bch_fs *c = container_of(shrink, struct bch_fs,
 401                                         btree_cache.shrink);
 402
 403         bch2_btree_cache_to_text(out, c);
 404 }
 405
 406 void bch2_fs_btree_cache_exit(struct bch_fs *c)
 407 {
 408         struct btree_cache *bc = &c->btree_cache;
 409         struct btree *b;
 410         unsigned i, flags;
 411
 412         if (bc->shrink.list.next)
 413                 unregister_shrinker(&bc->shrink);
 414
 415         /* vfree() can allocate memory: */
 416         flags = memalloc_nofs_save();
 417         mutex_lock(&bc->lock);
 418
 419         if (c->verify_data)
 420                 list_move(&c->verify_data->list, &bc->live);
 421
 422         kvpfree(c->verify_ondisk, btree_bytes(c));
 423
 424         for (i = 0; i < BTREE_ID_NR; i++)
 425                 if (c->btree_roots[i].b)
 426                         list_add(&c->btree_roots[i].b->list, &bc->live);
 427
 428         list_splice(&bc->freeable, &bc->live);
 429
 430         while (!list_empty(&bc->live)) {
 431                 b = list_first_entry(&bc->live, struct btree, list);
 432
 433                 BUG_ON(btree_node_read_in_flight(b) ||
 434                        btree_node_write_in_flight(b));
 435
 436                 if (btree_node_dirty(b))
 437                         bch2_btree_complete_write(c, b, btree_current_write(b));
 438                 clear_btree_node_dirty_acct(c, b);
 439
 440                 btree_node_data_free(c, b);
 441         }
 442
 443         BUG_ON(atomic_read(&c->btree_cache.dirty));
 444
 445         list_splice(&bc->freed_pcpu, &bc->freed_nonpcpu);
 446
 447         while (!list_empty(&bc->freed_nonpcpu)) {
 448                 b = list_first_entry(&bc->freed_nonpcpu, struct btree, list);
 449                 list_del(&b->list);
 450                 six_lock_pcpu_free(&b->c.lock);
 451                 kfree(b);
 452         }
 453
 454         mutex_unlock(&bc->lock);
 455         memalloc_nofs_restore(flags);
 456
 457         if (bc->table_init_done)
 458                 rhashtable_destroy(&bc->table);
 459 }
 460
 461 int bch2_fs_btree_cache_init(struct bch_fs *c)
 462 {
 463         struct btree_cache *bc = &c->btree_cache;
 464         unsigned i;
 465         int ret = 0;
 466
 467         pr_verbose_init(c->opts, "");
 468
 469         ret = rhashtable_init(&bc->table, &bch_btree_cache_params);
 470         if (ret)
 471                 goto out;
 472
 473         bc->table_init_done = true;
 474
 475         bch2_recalc_btree_reserve(c);
 476
 477         for (i = 0; i < bc->reserve; i++)
 478                 if (!__bch2_btree_node_mem_alloc(c)) {
 479                         ret = -ENOMEM;
 480                         goto out;
 481                 }
 482
 483         list_splice_init(&bc->live, &bc->freeable);
 484
 485         mutex_init(&c->verify_lock);
 486
 487         bc->shrink.count_objects        = bch2_btree_cache_count;
 488         bc->shrink.scan_objects         = bch2_btree_cache_scan;
 489         bc->shrink.to_text              = bch2_btree_cache_shrinker_to_text;
 490         bc->shrink.seeks                = 4;
 491         ret = register_shrinker(&bc->shrink);
 492 out:
 493         pr_verbose_init(c->opts, "ret %i", ret);
 494         return ret;
 495 }
 496
 497 void bch2_fs_btree_cache_init_early(struct btree_cache *bc)
 498 {
 499         mutex_init(&bc->lock);
 500         INIT_LIST_HEAD(&bc->live);
 501         INIT_LIST_HEAD(&bc->freeable);
 502         INIT_LIST_HEAD(&bc->freed_pcpu);
 503         INIT_LIST_HEAD(&bc->freed_nonpcpu);
 504 }
 505
 506 /*
 507  * We can only have one thread cannibalizing other cached btree nodes at a time,
 508  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 509  * cannibalize_bucket() will take. This means every time we unlock the root of
 510  * the btree, we need to release this lock if we have it held.
 511  */
 512 void bch2_btree_cache_cannibalize_unlock(struct bch_fs *c)
 513 {
 514         struct btree_cache *bc = &c->btree_cache;
 515
 516         if (bc->alloc_lock == current) {
 517                 trace_btree_node_cannibalize_unlock(c);
 518                 bc->alloc_lock = NULL;
 519                 closure_wake_up(&bc->alloc_wait);
 520         }
 521 }
 522
 523 int bch2_btree_cache_cannibalize_lock(struct bch_fs *c, struct closure *cl)
 524 {
 525         struct btree_cache *bc = &c->btree_cache;
 526         struct task_struct *old;
 527
 528         old = cmpxchg(&bc->alloc_lock, NULL, current);
 529         if (old == NULL || old == current)
 530                 goto success;
 531
 532         if (!cl) {
 533                 trace_btree_node_cannibalize_lock_fail(c);
 534                 return -ENOMEM;
 535         }
 536
 537         closure_wait(&bc->alloc_wait, cl);
 538
 539         /* Try again, after adding ourselves to waitlist */
 540         old = cmpxchg(&bc->alloc_lock, NULL, current);
 541         if (old == NULL || old == current) {
 542                 /* We raced */
 543                 closure_wake_up(&bc->alloc_wait);
 544                 goto success;
 545         }
 546
 547         trace_btree_node_cannibalize_lock_fail(c);
 548         return -EAGAIN;
 549
 550 success:
 551         trace_btree_node_cannibalize_lock(c);
 552         return 0;
 553 }
 554
 555 static struct btree *btree_node_cannibalize(struct bch_fs *c)
 556 {
 557         struct btree_cache *bc = &c->btree_cache;
 558         struct btree *b;
 559
 560         list_for_each_entry_reverse(b, &bc->live, list)
 561                 if (!btree_node_reclaim(c, b))
 562                         return b;
 563
 564         while (1) {
 565                 list_for_each_entry_reverse(b, &bc->live, list)
 566                         if (!btree_node_write_and_reclaim(c, b))
 567                                 return b;
 568
 569                 /*
 570                  * Rare case: all nodes were intent-locked.
 571                  * Just busy-wait.
 572                  */
 573                 WARN_ONCE(1, "btree cache cannibalize failed\n");
 574                 cond_resched();
 575         }
 576 }
 577
 578 struct btree *bch2_btree_node_mem_alloc(struct bch_fs *c, bool pcpu_read_locks)
 579 {
 580         struct btree_cache *bc = &c->btree_cache;
 581         struct list_head *freed = pcpu_read_locks
 582                 ? &bc->freed_pcpu
 583                 : &bc->freed_nonpcpu;
 584         struct btree *b, *b2;
 585         u64 start_time = local_clock();
 586         unsigned flags;
 587
 588         flags = memalloc_nofs_save();
 589         mutex_lock(&bc->lock);
 590
 591         /*
 592          * We never free struct btree itself, just the memory that holds the on
 593          * disk node. Check the freed list before allocating a new one:
 594          */
 595         list_for_each_entry(b, freed, list)
 596                 if (!btree_node_reclaim(c, b)) {
 597                         list_del_init(&b->list);
 598                         goto got_node;
 599                 }
 600
 601         b = __btree_node_mem_alloc(c);
 602         if (!b)
 603                 goto err_locked;
 604
 605         if (pcpu_read_locks)
 606                 six_lock_pcpu_alloc(&b->c.lock);
 607
 608         BUG_ON(!six_trylock_intent(&b->c.lock));
 609         BUG_ON(!six_trylock_write(&b->c.lock));
 610 got_node:
 611
 612         /*
 613          * btree_free() doesn't free memory; it sticks the node on the end of
 614          * the list. Check if there's any freed nodes there:
 615          */
 616         list_for_each_entry(b2, &bc->freeable, list)
 617                 if (!btree_node_reclaim(c, b2)) {
 618                         swap(b->data, b2->data);
 619                         swap(b->aux_data, b2->aux_data);
 620                         btree_node_to_freedlist(bc, b2);
 621                         six_unlock_write(&b2->c.lock);
 622                         six_unlock_intent(&b2->c.lock);
 623                         goto got_mem;
 624                 }
 625
 626         mutex_unlock(&bc->lock);
 627
 628         if (btree_node_data_alloc(c, b, __GFP_NOWARN|GFP_KERNEL))
 629                 goto err;
 630
 631         mutex_lock(&bc->lock);
 632         bc->used++;
 633 got_mem:
 634         mutex_unlock(&bc->lock);
 635
 636         BUG_ON(btree_node_hashed(b));
 637         BUG_ON(btree_node_dirty(b));
 638         BUG_ON(btree_node_write_in_flight(b));
 639 out:
 640         b->flags                = 0;
 641         b->written              = 0;
 642         b->nsets                = 0;
 643         b->sib_u64s[0]          = 0;
 644         b->sib_u64s[1]          = 0;
 645         b->whiteout_u64s        = 0;
 646         bch2_btree_keys_init(b);
 647         set_btree_node_accessed(b);
 648
 649         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],
 650                                start_time);
 651
 652         memalloc_nofs_restore(flags);
 653         return b;
 654 err:
 655         mutex_lock(&bc->lock);
 656 err_locked:
 657         /* Try to cannibalize another cached btree node: */
 658         if (bc->alloc_lock == current) {
 659                 b2 = btree_node_cannibalize(c);
 660                 bch2_btree_node_hash_remove(bc, b2);
 661
 662                 if (b) {
 663                         swap(b->data, b2->data);
 664                         swap(b->aux_data, b2->aux_data);
 665                         btree_node_to_freedlist(bc, b2);
 666                         six_unlock_write(&b2->c.lock);
 667                         six_unlock_intent(&b2->c.lock);
 668                 } else {
 669                         b = b2;
 670                         list_del_init(&b->list);
 671                 }
 672
 673                 mutex_unlock(&bc->lock);
 674
 675                 trace_btree_node_cannibalize(c);
 676                 goto out;
 677         }
 678
 679         mutex_unlock(&bc->lock);
 680         memalloc_nofs_restore(flags);
 681         return ERR_PTR(-ENOMEM);
 682 }
 683
 684 /* Slowpath, don't want it inlined into btree_iter_traverse() */
 685 static noinline struct btree *bch2_btree_node_fill(struct bch_fs *c,
 686                                 struct btree_trans *trans,
 687                                 struct btree_path *path,
 688                                 const struct bkey_i *k,
 689                                 enum btree_id btree_id,
 690                                 unsigned level,
 691                                 enum six_lock_type lock_type,
 692                                 bool sync)
 693 {
 694         struct btree_cache *bc = &c->btree_cache;
 695         struct btree *b;
 696         u32 seq;
 697
 698         BUG_ON(level + 1 >= BTREE_MAX_DEPTH);
 699         /*
 700          * Parent node must be locked, else we could read in a btree node that's
 701          * been freed:
 702          */
 703         if (trans && !bch2_btree_node_relock(trans, path, level + 1)) {
 704                 trace_trans_restart_relock_parent_for_fill(trans, _THIS_IP_, path);
 705                 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_fill_relock));
 706         }
 707
 708         b = bch2_btree_node_mem_alloc(c, level != 0);
 709
 710         if (trans && b == ERR_PTR(-ENOMEM)) {
 711                 trans->memory_allocation_failure = true;
 712                 trace_trans_restart_memory_allocation_failure(trans, _THIS_IP_, path);
 713                 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_fill_mem_alloc_fail));
 714         }
 715
 716         if (IS_ERR(b))
 717                 return b;
 718
 719         bkey_copy(&b->key, k);
 720         if (bch2_btree_node_hash_insert(bc, b, level, btree_id)) {
 721                 /* raced with another fill: */
 722
 723                 /* mark as unhashed... */
 724                 b->hash_val = 0;
 725
 726                 mutex_lock(&bc->lock);
 727                 list_add(&b->list, &bc->freeable);
 728                 mutex_unlock(&bc->lock);
 729
 730                 six_unlock_write(&b->c.lock);
 731                 six_unlock_intent(&b->c.lock);
 732                 return NULL;
 733         }
 734
 735         set_btree_node_read_in_flight(b);
 736
 737         six_unlock_write(&b->c.lock);
 738         seq = b->c.lock.state.seq;
 739         six_unlock_intent(&b->c.lock);
 740
 741         /* Unlock before doing IO: */
 742         if (trans && sync)
 743                 bch2_trans_unlock(trans);
 744
 745         bch2_btree_node_read(c, b, sync);
 746
 747         if (!sync)
 748                 return NULL;
 749
 750         if (trans) {
 751                 int ret = bch2_trans_relock(trans) ?:
 752                         bch2_btree_path_relock_intent(trans, path);
 753                 if (ret) {
 754                         BUG_ON(!trans->restarted);
 755                         return ERR_PTR(ret);
 756                 }
 757         }
 758
 759         if (!six_relock_type(&b->c.lock, lock_type, seq)) {
 760                 if (trans)
 761                         trace_trans_restart_relock_after_fill(trans, _THIS_IP_, path);
 762                 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_relock_after_fill));
 763         }
 764
 765         return b;
 766 }
 767
 768 static int lock_node_check_fn(struct six_lock *lock, void *p)
 769 {
 770         struct btree *b = container_of(lock, struct btree, c.lock);
 771         const struct bkey_i *k = p;
 772
 773         if (b->hash_val != btree_ptr_hash_val(k))
 774                 return BCH_ERR_lock_fail_node_reused;
 775         return 0;
 776 }
 777
 778 static noinline void btree_bad_header(struct bch_fs *c, struct btree *b)
 779 {
 780         struct printbuf buf = PRINTBUF;
 781
 782         if (!test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags))
 783                 return;
 784
 785         prt_printf(&buf,
 786                "btree node header doesn't match ptr\n"
 787                "btree %s level %u\n"
 788                "ptr: ",
 789                bch2_btree_ids[b->c.btree_id], b->c.level);
 790         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
 791
 792         prt_printf(&buf, "\nheader: btree %s level %llu\n"
 793                "min ",
 794                bch2_btree_ids[BTREE_NODE_ID(b->data)],
 795                BTREE_NODE_LEVEL(b->data));
 796         bch2_bpos_to_text(&buf, b->data->min_key);
 797
 798         prt_printf(&buf, "\nmax ");
 799         bch2_bpos_to_text(&buf, b->data->max_key);
 800
 801         bch2_fs_inconsistent(c, "%s", buf.buf);
 802         printbuf_exit(&buf);
 803 }
 804
 805 static inline void btree_check_header(struct bch_fs *c, struct btree *b)
 806 {
 807         if (b->c.btree_id != BTREE_NODE_ID(b->data) ||
 808             b->c.level != BTREE_NODE_LEVEL(b->data) ||
 809             bpos_cmp(b->data->max_key, b->key.k.p) ||
 810             (b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 811              bpos_cmp(b->data->min_key,
 812                       bkey_i_to_btree_ptr_v2(&b->key)->v.min_key)))
 813                 btree_bad_header(c, b);
 814 }
 815
 816 /**
 817  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 818  * in from disk if necessary.
 819  *
 820  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
 821  *
 822  * The btree node will have either a read or a write lock held, depending on
 823  * the @write parameter.
 824  */
 825 struct btree *bch2_btree_node_get(struct btree_trans *trans, struct btree_path *path,
 826                                   const struct bkey_i *k, unsigned level,
 827                                   enum six_lock_type lock_type,
 828                                   unsigned long trace_ip)
 829 {
 830         struct bch_fs *c = trans->c;
 831         struct btree_cache *bc = &c->btree_cache;
 832         struct btree *b;
 833         struct bset_tree *t;
 834         int ret;
 835
 836         EBUG_ON(level >= BTREE_MAX_DEPTH);
 837
 838         b = btree_node_mem_ptr(k);
 839
 840         /*
 841          * Check b->hash_val _before_ calling btree_node_lock() - this might not
 842          * be the node we want anymore, and trying to lock the wrong node could
 843          * cause an unneccessary transaction restart:
 844          */
 845         if (likely(c->opts.btree_node_mem_ptr_optimization &&
 846                    b &&
 847                    b->hash_val == btree_ptr_hash_val(k)))
 848                         goto lock_node;
 849 retry:
 850         b = btree_cache_find(bc, k);
 851         if (unlikely(!b)) {
 852                 /*
 853                  * We must have the parent locked to call bch2_btree_node_fill(),
 854                  * else we could read in a btree node from disk that's been
 855                  * freed:
 856                  */
 857                 b = bch2_btree_node_fill(c, trans, path, k, path->btree_id,
 858                                          level, lock_type, true);
 859
 860                 /* We raced and found the btree node in the cache */
 861                 if (!b)
 862                         goto retry;
 863
 864                 if (IS_ERR(b))
 865                         return b;
 866         } else {
 867 lock_node:
 868                 /*
 869                  * There's a potential deadlock with splits and insertions into
 870                  * interior nodes we have to avoid:
 871                  *
 872                  * The other thread might be holding an intent lock on the node
 873                  * we want, and they want to update its parent node so they're
 874                  * going to upgrade their intent lock on the parent node to a
 875                  * write lock.
 876                  *
 877                  * But if we're holding a read lock on the parent, and we're
 878                  * trying to get the intent lock they're holding, we deadlock.
 879                  *
 880                  * So to avoid this we drop the read locks on parent nodes when
 881                  * we're starting to take intent locks - and handle the race.
 882                  *
 883                  * The race is that they might be about to free the node we
 884                  * want, and dropping our read lock on the parent node lets them
 885                  * update the parent marking the node we want as freed, and then
 886                  * free it:
 887                  *
 888                  * To guard against this, btree nodes are evicted from the cache
 889                  * when they're freed - and b->hash_val is zeroed out, which we
 890                  * check for after we lock the node.
 891                  *
 892                  * Then, bch2_btree_node_relock() on the parent will fail - because
 893                  * the parent was modified, when the pointer to the node we want
 894                  * was removed - and we'll bail out:
 895                  */
 896                 if (btree_node_read_locked(path, level + 1))
 897                         btree_node_unlock(trans, path, level + 1);
 898
 899                 ret = btree_node_lock(trans, path, b, k->k.p, level, lock_type,
 900                                       lock_node_check_fn, (void *) k, trace_ip);
 901                 if (unlikely(ret)) {
 902                         if (bch2_err_matches(ret, BCH_ERR_lock_fail_node_reused))
 903                                 goto retry;
 904                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 905                                 return ERR_PTR(ret);
 906                         BUG();
 907                 }
 908
 909                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
 910                              b->c.level != level ||
 911                              race_fault())) {
 912                         six_unlock_type(&b->c.lock, lock_type);
 913                         if (bch2_btree_node_relock(trans, path, level + 1))
 914                                 goto retry;
 915
 916                         trace_trans_restart_btree_node_reused(trans, trace_ip, path);
 917                         return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_lock_node_reused));
 918                 }
 919         }
 920
 921         if (unlikely(btree_node_read_in_flight(b))) {
 922                 u32 seq = b->c.lock.state.seq;
 923
 924                 six_unlock_type(&b->c.lock, lock_type);
 925                 bch2_trans_unlock(trans);
 926
 927                 bch2_btree_node_wait_on_read(b);
 928
 929                 /*
 930                  * should_be_locked is not set on this path yet, so we need to
 931                  * relock it specifically:
 932                  */
 933                 if (trans) {
 934                         int ret = bch2_trans_relock(trans) ?:
 935                                 bch2_btree_path_relock_intent(trans, path);
 936                         if (ret) {
 937                                 BUG_ON(!trans->restarted);
 938                                 return ERR_PTR(ret);
 939                         }
 940                 }
 941
 942                 if (!six_relock_type(&b->c.lock, lock_type, seq))
 943                         goto retry;
 944         }
 945
 946         prefetch(b->aux_data);
 947
 948         for_each_bset(b, t) {
 949                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 950
 951                 prefetch(p + L1_CACHE_BYTES * 0);
 952                 prefetch(p + L1_CACHE_BYTES * 1);
 953                 prefetch(p + L1_CACHE_BYTES * 2);
 954         }
 955
 956         /* avoid atomic set bit if it's not needed: */
 957         if (!btree_node_accessed(b))
 958                 set_btree_node_accessed(b);
 959
 960         if (unlikely(btree_node_read_error(b))) {
 961                 six_unlock_type(&b->c.lock, lock_type);
 962                 return ERR_PTR(-EIO);
 963         }
 964
 965         EBUG_ON(b->c.btree_id != path->btree_id);
 966         EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
 967         btree_check_header(c, b);
 968
 969         return b;
 970 }
 971
 972 struct btree *bch2_btree_node_get_noiter(struct bch_fs *c,
 973                                          const struct bkey_i *k,
 974                                          enum btree_id btree_id,
 975                                          unsigned level,
 976                                          bool nofill)
 977 {
 978         struct btree_cache *bc = &c->btree_cache;
 979         struct btree *b;
 980         struct bset_tree *t;
 981         int ret;
 982
 983         EBUG_ON(level >= BTREE_MAX_DEPTH);
 984
 985         if (c->opts.btree_node_mem_ptr_optimization) {
 986                 b = btree_node_mem_ptr(k);
 987                 if (b)
 988                         goto lock_node;
 989         }
 990 retry:
 991         b = btree_cache_find(bc, k);
 992         if (unlikely(!b)) {
 993                 if (nofill)
 994                         goto out;
 995
 996                 b = bch2_btree_node_fill(c, NULL, NULL, k, btree_id,
 997                                          level, SIX_LOCK_read, true);
 998
 999                 /* We raced and found the btree node in the cache */
1000                 if (!b)
1001                         goto retry;
1002
1003                 if (IS_ERR(b) &&
1004                     !bch2_btree_cache_cannibalize_lock(c, NULL))
1005                         goto retry;
1006
1007                 if (IS_ERR(b))
1008                         goto out;
1009         } else {
1010 lock_node:
1011                 ret = six_lock_read(&b->c.lock, lock_node_check_fn, (void *) k);
1012                 if (ret)
1013                         goto retry;
1014
1015                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
1016                              b->c.btree_id != btree_id ||
1017                              b->c.level != level)) {
1018                         six_unlock_read(&b->c.lock);
1019                         goto retry;
1020                 }
1021         }
1022
1023         /* XXX: waiting on IO with btree locks held: */
1024         __bch2_btree_node_wait_on_read(b);
1025
1026         prefetch(b->aux_data);
1027
1028         for_each_bset(b, t) {
1029                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
1030
1031                 prefetch(p + L1_CACHE_BYTES * 0);
1032                 prefetch(p + L1_CACHE_BYTES * 1);
1033                 prefetch(p + L1_CACHE_BYTES * 2);
1034         }
1035
1036         /* avoid atomic set bit if it's not needed: */
1037         if (!btree_node_accessed(b))
1038                 set_btree_node_accessed(b);
1039
1040         if (unlikely(btree_node_read_error(b))) {
1041                 six_unlock_read(&b->c.lock);
1042                 b = ERR_PTR(-EIO);
1043                 goto out;
1044         }
1045
1046         EBUG_ON(b->c.btree_id != btree_id);
1047         EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
1048         btree_check_header(c, b);
1049 out:
1050         bch2_btree_cache_cannibalize_unlock(c);
1051         return b;
1052 }
1053
1054 int bch2_btree_node_prefetch(struct bch_fs *c,
1055                              struct btree_trans *trans,
1056                              struct btree_path *path,
1057                              const struct bkey_i *k,
1058                              enum btree_id btree_id, unsigned level)
1059 {
1060         struct btree_cache *bc = &c->btree_cache;
1061         struct btree *b;
1062
1063         BUG_ON(trans && !btree_node_locked(path, level + 1));
1064         BUG_ON(level >= BTREE_MAX_DEPTH);
1065
1066         b = btree_cache_find(bc, k);
1067         if (b)
1068                 return 0;
1069
1070         b = bch2_btree_node_fill(c, trans, path, k, btree_id,
1071                                  level, SIX_LOCK_read, false);
1072         return PTR_ERR_OR_ZERO(b);
1073 }
1074
1075 void bch2_btree_node_evict(struct bch_fs *c, const struct bkey_i *k)
1076 {
1077         struct btree_cache *bc = &c->btree_cache;
1078         struct btree *b;
1079
1080         b = btree_cache_find(bc, k);
1081         if (!b)
1082                 return;
1083 wait_on_io:
1084         /* not allowed to wait on io with btree locks held: */
1085
1086         /* XXX we're called from btree_gc which will be holding other btree
1087          * nodes locked
1088          * */
1089         __bch2_btree_node_wait_on_read(b);
1090         __bch2_btree_node_wait_on_write(b);
1091
1092         six_lock_intent(&b->c.lock, NULL, NULL);
1093         six_lock_write(&b->c.lock, NULL, NULL);
1094
1095         if (btree_node_dirty(b)) {
1096                 __bch2_btree_node_write(c, b, 0);
1097                 six_unlock_write(&b->c.lock);
1098                 six_unlock_intent(&b->c.lock);
1099                 goto wait_on_io;
1100         }
1101
1102         BUG_ON(btree_node_dirty(b));
1103
1104         mutex_lock(&bc->lock);
1105         btree_node_data_free(c, b);
1106         bch2_btree_node_hash_remove(bc, b);
1107         mutex_unlock(&bc->lock);
1108
1109         six_unlock_write(&b->c.lock);
1110         six_unlock_intent(&b->c.lock);
1111 }
1112
1113 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
1114                              struct btree *b)
1115 {
1116         const struct bkey_format *f = &b->format;
1117         struct bset_stats stats;
1118
1119         memset(&stats, 0, sizeof(stats));
1120
1121         bch2_btree_keys_stats(b, &stats);
1122
1123         prt_printf(out, "l %u ", b->c.level);
1124         bch2_bpos_to_text(out, b->data->min_key);
1125         prt_printf(out, " - ");
1126         bch2_bpos_to_text(out, b->data->max_key);
1127         prt_printf(out, ":\n"
1128                "    ptrs: ");
1129         bch2_val_to_text(out, c, bkey_i_to_s_c(&b->key));
1130
1131         prt_printf(out, "\n"
1132                "    format: u64s %u fields %u %u %u %u %u\n"
1133                "    unpack fn len: %u\n"
1134                "    bytes used %zu/%zu (%zu%% full)\n"
1135                "    sib u64s: %u, %u (merge threshold %u)\n"
1136                "    nr packed keys %u\n"
1137                "    nr unpacked keys %u\n"
1138                "    floats %zu\n"
1139                "    failed unpacked %zu\n",
1140                f->key_u64s,
1141                f->bits_per_field[0],
1142                f->bits_per_field[1],
1143                f->bits_per_field[2],
1144                f->bits_per_field[3],
1145                f->bits_per_field[4],
1146                b->unpack_fn_len,
1147                b->nr.live_u64s * sizeof(u64),
1148                btree_bytes(c) - sizeof(struct btree_node),
1149                b->nr.live_u64s * 100 / btree_max_u64s(c),
1150                b->sib_u64s[0],
1151                b->sib_u64s[1],
1152                c->btree_foreground_merge_threshold,
1153                b->nr.packed_keys,
1154                b->nr.unpacked_keys,
1155                stats.floats,
1156                stats.failed);
1157 }
1158
1159 void bch2_btree_cache_to_text(struct printbuf *out, struct bch_fs *c)
1160 {
1161         prt_printf(out, "nr nodes:\t\t%u\n", c->btree_cache.used);
1162         prt_printf(out, "nr dirty:\t\t%u\n", atomic_read(&c->btree_cache.dirty));
1163         prt_printf(out, "cannibalize lock:\t%p\n", c->btree_cache.alloc_lock);
1164 }