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

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "bkey_buf.h"
   5 #include "btree_cache.h"
   6 #include "btree_io.h"
   7 #include "btree_iter.h"
   8 #include "btree_locking.h"
   9 #include "debug.h"
  10 #include "error.h"
  11
  12 #include <linux/prefetch.h>
  13 #include <linux/sched/mm.h>
  14 #include <trace/events/bcachefs.h>
  15
  16 struct lock_class_key bch2_btree_node_lock_key;
  17
  18 const char * const bch2_btree_node_flags[] = {
  19 #define x(f)    #f,
  20         BTREE_FLAGS()
  21 #undef x
  22         NULL
  23 };
  24
  25 void bch2_recalc_btree_reserve(struct bch_fs *c)
  26 {
  27         unsigned i, reserve = 16;
  28
  29         if (!c->btree_roots[0].b)
  30                 reserve += 8;
  31
  32         for (i = 0; i < BTREE_ID_NR; i++)
  33                 if (c->btree_roots[i].b)
  34                         reserve += min_t(unsigned, 1,
  35                                          c->btree_roots[i].b->c.level) * 8;
  36
  37         c->btree_cache.reserve = reserve;
  38 }
  39
  40 static inline unsigned btree_cache_can_free(struct btree_cache *bc)
  41 {
  42         return max_t(int, 0, bc->used - bc->reserve);
  43 }
  44
  45 static void btree_node_to_freedlist(struct btree_cache *bc, struct btree *b)
  46 {
  47         if (b->c.lock.readers)
  48                 list_move(&b->list, &bc->freed_pcpu);
  49         else
  50                 list_move(&b->list, &bc->freed_nonpcpu);
  51 }
  52
  53 static void btree_node_data_free(struct bch_fs *c, struct btree *b)
  54 {
  55         struct btree_cache *bc = &c->btree_cache;
  56
  57         EBUG_ON(btree_node_write_in_flight(b));
  58
  59         kvpfree(b->data, btree_bytes(c));
  60         b->data = NULL;
  61 #ifdef __KERNEL__
  62         vfree(b->aux_data);
  63 #else
  64         munmap(b->aux_data, btree_aux_data_bytes(b));
  65 #endif
  66         b->aux_data = NULL;
  67
  68         bc->used--;
  69
  70         btree_node_to_freedlist(bc, b);
  71 }
  72
  73 static int bch2_btree_cache_cmp_fn(struct rhashtable_compare_arg *arg,
  74                                    const void *obj)
  75 {
  76         const struct btree *b = obj;
  77         const u64 *v = arg->key;
  78
  79         return b->hash_val == *v ? 0 : 1;
  80 }
  81
  82 static const struct rhashtable_params bch_btree_cache_params = {
  83         .head_offset    = offsetof(struct btree, hash),
  84         .key_offset     = offsetof(struct btree, hash_val),
  85         .key_len        = sizeof(u64),
  86         .obj_cmpfn      = bch2_btree_cache_cmp_fn,
  87 };
  88
  89 static int btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
  90 {
  91         BUG_ON(b->data || b->aux_data);
  92
  93         b->data = kvpmalloc(btree_bytes(c), gfp);
  94         if (!b->data)
  95                 return -ENOMEM;
  96 #ifdef __KERNEL__
  97         b->aux_data = vmalloc_exec(btree_aux_data_bytes(b), gfp);
  98 #else
  99         b->aux_data = mmap(NULL, btree_aux_data_bytes(b),
 100                            PROT_READ|PROT_WRITE|PROT_EXEC,
 101                            MAP_PRIVATE|MAP_ANONYMOUS, 0, 0);
 102         if (b->aux_data == MAP_FAILED)
 103                 b->aux_data = NULL;
 104 #endif
 105         if (!b->aux_data) {
 106                 kvpfree(b->data, btree_bytes(c));
 107                 b->data = NULL;
 108                 return -ENOMEM;
 109         }
 110
 111         return 0;
 112 }
 113
 114 static struct btree *__btree_node_mem_alloc(struct bch_fs *c)
 115 {
 116         struct btree *b = kzalloc(sizeof(struct btree), GFP_KERNEL);
 117         if (!b)
 118                 return NULL;
 119
 120         bkey_btree_ptr_init(&b->key);
 121         __six_lock_init(&b->c.lock, "b->c.lock", &bch2_btree_node_lock_key);
 122         INIT_LIST_HEAD(&b->list);
 123         INIT_LIST_HEAD(&b->write_blocked);
 124         b->byte_order = ilog2(btree_bytes(c));
 125         return b;
 126 }
 127
 128 struct btree *__bch2_btree_node_mem_alloc(struct bch_fs *c)
 129 {
 130         struct btree_cache *bc = &c->btree_cache;
 131         struct btree *b = __btree_node_mem_alloc(c);
 132         if (!b)
 133                 return NULL;
 134
 135         if (btree_node_data_alloc(c, b, GFP_KERNEL)) {
 136                 kfree(b);
 137                 return NULL;
 138         }
 139
 140         bc->used++;
 141         list_add(&b->list, &bc->freeable);
 142         return b;
 143 }
 144
 145 /* Btree in memory cache - hash table */
 146
 147 void bch2_btree_node_hash_remove(struct btree_cache *bc, struct btree *b)
 148 {
 149         int ret = rhashtable_remove_fast(&bc->table, &b->hash, bch_btree_cache_params);
 150         BUG_ON(ret);
 151
 152         /* Cause future lookups for this node to fail: */
 153         b->hash_val = 0;
 154
 155         six_lock_wakeup_all(&b->c.lock);
 156 }
 157
 158 int __bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b)
 159 {
 160         BUG_ON(b->hash_val);
 161         b->hash_val = btree_ptr_hash_val(&b->key);
 162
 163         return rhashtable_lookup_insert_fast(&bc->table, &b->hash,
 164                                              bch_btree_cache_params);
 165 }
 166
 167 int bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b,
 168                                 unsigned level, enum btree_id id)
 169 {
 170         int ret;
 171
 172         b->c.level      = level;
 173         b->c.btree_id   = id;
 174
 175         mutex_lock(&bc->lock);
 176         ret = __bch2_btree_node_hash_insert(bc, b);
 177         if (!ret)
 178                 list_add(&b->list, &bc->live);
 179         mutex_unlock(&bc->lock);
 180
 181         return ret;
 182 }
 183
 184 __flatten
 185 static inline struct btree *btree_cache_find(struct btree_cache *bc,
 186                                      const struct bkey_i *k)
 187 {
 188         u64 v = btree_ptr_hash_val(k);
 189
 190         return rhashtable_lookup_fast(&bc->table, &v, bch_btree_cache_params);
 191 }
 192
 193 /*
 194  * this version is for btree nodes that have already been freed (we're not
 195  * reaping a real btree node)
 196  */
 197 static int __btree_node_reclaim(struct bch_fs *c, struct btree *b, bool flush)
 198 {
 199         struct btree_cache *bc = &c->btree_cache;
 200         int ret = 0;
 201
 202         lockdep_assert_held(&bc->lock);
 203 wait_on_io:
 204         if (b->flags & ((1U << BTREE_NODE_dirty)|
 205                         (1U << BTREE_NODE_read_in_flight)|
 206                         (1U << BTREE_NODE_write_in_flight))) {
 207                 if (!flush)
 208                         return -ENOMEM;
 209
 210                 /* XXX: waiting on IO with btree cache lock held */
 211                 bch2_btree_node_wait_on_read(b);
 212                 bch2_btree_node_wait_on_write(b);
 213         }
 214
 215         if (!six_trylock_intent(&b->c.lock))
 216                 return -ENOMEM;
 217
 218         if (!six_trylock_write(&b->c.lock))
 219                 goto out_unlock_intent;
 220
 221         /* recheck under lock */
 222         if (b->flags & ((1U << BTREE_NODE_read_in_flight)|
 223                         (1U << BTREE_NODE_write_in_flight))) {
 224                 if (!flush)
 225                         goto out_unlock;
 226                 six_unlock_write(&b->c.lock);
 227                 six_unlock_intent(&b->c.lock);
 228                 goto wait_on_io;
 229         }
 230
 231         if (btree_node_noevict(b) ||
 232             btree_node_write_blocked(b) ||
 233             btree_node_will_make_reachable(b))
 234                 goto out_unlock;
 235
 236         if (btree_node_dirty(b)) {
 237                 if (!flush)
 238                         goto out_unlock;
 239                 /*
 240                  * Using the underscore version because we don't want to compact
 241                  * bsets after the write, since this node is about to be evicted
 242                  * - unless btree verify mode is enabled, since it runs out of
 243                  * the post write cleanup:
 244                  */
 245                 if (bch2_verify_btree_ondisk)
 246                         bch2_btree_node_write(c, b, SIX_LOCK_intent, 0);
 247                 else
 248                         __bch2_btree_node_write(c, b, 0);
 249
 250                 six_unlock_write(&b->c.lock);
 251                 six_unlock_intent(&b->c.lock);
 252                 goto wait_on_io;
 253         }
 254 out:
 255         if (b->hash_val && !ret)
 256                 trace_btree_node_reap(c, b);
 257         return ret;
 258 out_unlock:
 259         six_unlock_write(&b->c.lock);
 260 out_unlock_intent:
 261         six_unlock_intent(&b->c.lock);
 262         ret = -ENOMEM;
 263         goto out;
 264 }
 265
 266 static int btree_node_reclaim(struct bch_fs *c, struct btree *b)
 267 {
 268         return __btree_node_reclaim(c, b, false);
 269 }
 270
 271 static int btree_node_write_and_reclaim(struct bch_fs *c, struct btree *b)
 272 {
 273         return __btree_node_reclaim(c, b, true);
 274 }
 275
 276 static unsigned long bch2_btree_cache_scan(struct shrinker *shrink,
 277                                            struct shrink_control *sc)
 278 {
 279         struct bch_fs *c = container_of(shrink, struct bch_fs,
 280                                         btree_cache.shrink);
 281         struct btree_cache *bc = &c->btree_cache;
 282         struct btree *b, *t;
 283         unsigned long nr = sc->nr_to_scan;
 284         unsigned long can_free;
 285         unsigned long touched = 0;
 286         unsigned long freed = 0;
 287         unsigned i, flags;
 288         unsigned long ret = SHRINK_STOP;
 289
 290         if (bch2_btree_shrinker_disabled)
 291                 return SHRINK_STOP;
 292
 293         /* Return -1 if we can't do anything right now */
 294         if (sc->gfp_mask & __GFP_FS)
 295                 mutex_lock(&bc->lock);
 296         else if (!mutex_trylock(&bc->lock))
 297                 goto out_norestore;
 298
 299         flags = memalloc_nofs_save();
 300
 301         /*
 302          * It's _really_ critical that we don't free too many btree nodes - we
 303          * have to always leave ourselves a reserve. The reserve is how we
 304          * guarantee that allocating memory for a new btree node can always
 305          * succeed, so that inserting keys into the btree can always succeed and
 306          * IO can always make forward progress:
 307          */
 308         nr /= btree_pages(c);
 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 = (unsigned long) freed * btree_pages(c);
 379         memalloc_nofs_restore(flags);
 380 out_norestore:
 381         trace_btree_cache_scan(sc->nr_to_scan,
 382                                sc->nr_to_scan / btree_pages(c),
 383                                btree_cache_can_free(bc),
 384                                ret);
 385         return ret;
 386 }
 387
 388 static unsigned long bch2_btree_cache_count(struct shrinker *shrink,
 389                                             struct shrink_control *sc)
 390 {
 391         struct bch_fs *c = container_of(shrink, struct bch_fs,
 392                                         btree_cache.shrink);
 393         struct btree_cache *bc = &c->btree_cache;
 394
 395         if (bch2_btree_shrinker_disabled)
 396                 return 0;
 397
 398         return btree_cache_can_free(bc) * btree_pages(c);
 399 }
 400
 401 void bch2_fs_btree_cache_exit(struct bch_fs *c)
 402 {
 403         struct btree_cache *bc = &c->btree_cache;
 404         struct btree *b;
 405         unsigned i, flags;
 406
 407         if (bc->shrink.list.next)
 408                 unregister_shrinker(&bc->shrink);
 409
 410         /* vfree() can allocate memory: */
 411         flags = memalloc_nofs_save();
 412         mutex_lock(&bc->lock);
 413
 414         if (c->verify_data)
 415                 list_move(&c->verify_data->list, &bc->live);
 416
 417         kvpfree(c->verify_ondisk, btree_bytes(c));
 418
 419         for (i = 0; i < BTREE_ID_NR; i++)
 420                 if (c->btree_roots[i].b)
 421                         list_add(&c->btree_roots[i].b->list, &bc->live);
 422
 423         list_splice(&bc->freeable, &bc->live);
 424
 425         while (!list_empty(&bc->live)) {
 426                 b = list_first_entry(&bc->live, struct btree, list);
 427
 428                 BUG_ON(btree_node_read_in_flight(b) ||
 429                        btree_node_write_in_flight(b));
 430
 431                 if (btree_node_dirty(b))
 432                         bch2_btree_complete_write(c, b, btree_current_write(b));
 433                 clear_btree_node_dirty_acct(c, b);
 434
 435                 btree_node_data_free(c, b);
 436         }
 437
 438         BUG_ON(atomic_read(&c->btree_cache.dirty));
 439
 440         list_splice(&bc->freed_pcpu, &bc->freed_nonpcpu);
 441
 442         while (!list_empty(&bc->freed_nonpcpu)) {
 443                 b = list_first_entry(&bc->freed_nonpcpu, struct btree, list);
 444                 list_del(&b->list);
 445                 six_lock_pcpu_free(&b->c.lock);
 446                 kfree(b);
 447         }
 448
 449         mutex_unlock(&bc->lock);
 450         memalloc_nofs_restore(flags);
 451
 452         if (bc->table_init_done)
 453                 rhashtable_destroy(&bc->table);
 454 }
 455
 456 int bch2_fs_btree_cache_init(struct bch_fs *c)
 457 {
 458         struct btree_cache *bc = &c->btree_cache;
 459         unsigned i;
 460         int ret = 0;
 461
 462         pr_verbose_init(c->opts, "");
 463
 464         ret = rhashtable_init(&bc->table, &bch_btree_cache_params);
 465         if (ret)
 466                 goto out;
 467
 468         bc->table_init_done = true;
 469
 470         bch2_recalc_btree_reserve(c);
 471
 472         for (i = 0; i < bc->reserve; i++)
 473                 if (!__bch2_btree_node_mem_alloc(c)) {
 474                         ret = -ENOMEM;
 475                         goto out;
 476                 }
 477
 478         list_splice_init(&bc->live, &bc->freeable);
 479
 480         mutex_init(&c->verify_lock);
 481
 482         bc->shrink.count_objects        = bch2_btree_cache_count;
 483         bc->shrink.scan_objects         = bch2_btree_cache_scan;
 484         bc->shrink.seeks                = 4;
 485         bc->shrink.batch                = btree_pages(c) * 2;
 486         ret = register_shrinker(&bc->shrink);
 487 out:
 488         pr_verbose_init(c->opts, "ret %i", ret);
 489         return ret;
 490 }
 491
 492 void bch2_fs_btree_cache_init_early(struct btree_cache *bc)
 493 {
 494         mutex_init(&bc->lock);
 495         INIT_LIST_HEAD(&bc->live);
 496         INIT_LIST_HEAD(&bc->freeable);
 497         INIT_LIST_HEAD(&bc->freed_pcpu);
 498         INIT_LIST_HEAD(&bc->freed_nonpcpu);
 499 }
 500
 501 /*
 502  * We can only have one thread cannibalizing other cached btree nodes at a time,
 503  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 504  * cannibalize_bucket() will take. This means every time we unlock the root of
 505  * the btree, we need to release this lock if we have it held.
 506  */
 507 void bch2_btree_cache_cannibalize_unlock(struct bch_fs *c)
 508 {
 509         struct btree_cache *bc = &c->btree_cache;
 510
 511         if (bc->alloc_lock == current) {
 512                 trace_btree_node_cannibalize_unlock(c);
 513                 bc->alloc_lock = NULL;
 514                 closure_wake_up(&bc->alloc_wait);
 515         }
 516 }
 517
 518 int bch2_btree_cache_cannibalize_lock(struct bch_fs *c, struct closure *cl)
 519 {
 520         struct btree_cache *bc = &c->btree_cache;
 521         struct task_struct *old;
 522
 523         old = cmpxchg(&bc->alloc_lock, NULL, current);
 524         if (old == NULL || old == current)
 525                 goto success;
 526
 527         if (!cl) {
 528                 trace_btree_node_cannibalize_lock_fail(c);
 529                 return -ENOMEM;
 530         }
 531
 532         closure_wait(&bc->alloc_wait, cl);
 533
 534         /* Try again, after adding ourselves to waitlist */
 535         old = cmpxchg(&bc->alloc_lock, NULL, current);
 536         if (old == NULL || old == current) {
 537                 /* We raced */
 538                 closure_wake_up(&bc->alloc_wait);
 539                 goto success;
 540         }
 541
 542         trace_btree_node_cannibalize_lock_fail(c);
 543         return -EAGAIN;
 544
 545 success:
 546         trace_btree_node_cannibalize_lock(c);
 547         return 0;
 548 }
 549
 550 static struct btree *btree_node_cannibalize(struct bch_fs *c)
 551 {
 552         struct btree_cache *bc = &c->btree_cache;
 553         struct btree *b;
 554
 555         list_for_each_entry_reverse(b, &bc->live, list)
 556                 if (!btree_node_reclaim(c, b))
 557                         return b;
 558
 559         while (1) {
 560                 list_for_each_entry_reverse(b, &bc->live, list)
 561                         if (!btree_node_write_and_reclaim(c, b))
 562                                 return b;
 563
 564                 /*
 565                  * Rare case: all nodes were intent-locked.
 566                  * Just busy-wait.
 567                  */
 568                 WARN_ONCE(1, "btree cache cannibalize failed\n");
 569                 cond_resched();
 570         }
 571 }
 572
 573 struct btree *bch2_btree_node_mem_alloc(struct bch_fs *c, bool pcpu_read_locks)
 574 {
 575         struct btree_cache *bc = &c->btree_cache;
 576         struct list_head *freed = pcpu_read_locks
 577                 ? &bc->freed_pcpu
 578                 : &bc->freed_nonpcpu;
 579         struct btree *b, *b2;
 580         u64 start_time = local_clock();
 581         unsigned flags;
 582
 583         flags = memalloc_nofs_save();
 584         mutex_lock(&bc->lock);
 585
 586         /*
 587          * We never free struct btree itself, just the memory that holds the on
 588          * disk node. Check the freed list before allocating a new one:
 589          */
 590         list_for_each_entry(b, freed, list)
 591                 if (!btree_node_reclaim(c, b)) {
 592                         list_del_init(&b->list);
 593                         goto got_node;
 594                 }
 595
 596         b = __btree_node_mem_alloc(c);
 597         if (!b)
 598                 goto err_locked;
 599
 600         if (pcpu_read_locks)
 601                 six_lock_pcpu_alloc(&b->c.lock);
 602
 603         BUG_ON(!six_trylock_intent(&b->c.lock));
 604         BUG_ON(!six_trylock_write(&b->c.lock));
 605 got_node:
 606
 607         /*
 608          * btree_free() doesn't free memory; it sticks the node on the end of
 609          * the list. Check if there's any freed nodes there:
 610          */
 611         list_for_each_entry(b2, &bc->freeable, list)
 612                 if (!btree_node_reclaim(c, b2)) {
 613                         swap(b->data, b2->data);
 614                         swap(b->aux_data, b2->aux_data);
 615                         btree_node_to_freedlist(bc, b2);
 616                         six_unlock_write(&b2->c.lock);
 617                         six_unlock_intent(&b2->c.lock);
 618                         goto got_mem;
 619                 }
 620
 621         mutex_unlock(&bc->lock);
 622
 623         if (btree_node_data_alloc(c, b, __GFP_NOWARN|GFP_KERNEL))
 624                 goto err;
 625
 626         mutex_lock(&bc->lock);
 627         bc->used++;
 628 got_mem:
 629         mutex_unlock(&bc->lock);
 630
 631         BUG_ON(btree_node_hashed(b));
 632         BUG_ON(btree_node_dirty(b));
 633         BUG_ON(btree_node_write_in_flight(b));
 634 out:
 635         b->flags                = 0;
 636         b->written              = 0;
 637         b->nsets                = 0;
 638         b->sib_u64s[0]          = 0;
 639         b->sib_u64s[1]          = 0;
 640         b->whiteout_u64s        = 0;
 641         bch2_btree_keys_init(b);
 642         set_btree_node_accessed(b);
 643
 644         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],
 645                                start_time);
 646
 647         memalloc_nofs_restore(flags);
 648         return b;
 649 err:
 650         mutex_lock(&bc->lock);
 651 err_locked:
 652         /* Try to cannibalize another cached btree node: */
 653         if (bc->alloc_lock == current) {
 654                 b2 = btree_node_cannibalize(c);
 655                 bch2_btree_node_hash_remove(bc, b2);
 656
 657                 if (b) {
 658                         swap(b->data, b2->data);
 659                         swap(b->aux_data, b2->aux_data);
 660                         btree_node_to_freedlist(bc, b2);
 661                         six_unlock_write(&b2->c.lock);
 662                         six_unlock_intent(&b2->c.lock);
 663                 } else {
 664                         b = b2;
 665                         list_del_init(&b->list);
 666                 }
 667
 668                 mutex_unlock(&bc->lock);
 669
 670                 trace_btree_node_cannibalize(c);
 671                 goto out;
 672         }
 673
 674         mutex_unlock(&bc->lock);
 675         memalloc_nofs_restore(flags);
 676         return ERR_PTR(-ENOMEM);
 677 }
 678
 679 /* Slowpath, don't want it inlined into btree_iter_traverse() */
 680 static noinline struct btree *bch2_btree_node_fill(struct bch_fs *c,
 681                                 struct btree_trans *trans,
 682                                 struct btree_path *path,
 683                                 const struct bkey_i *k,
 684                                 enum btree_id btree_id,
 685                                 unsigned level,
 686                                 enum six_lock_type lock_type,
 687                                 bool sync)
 688 {
 689         struct btree_cache *bc = &c->btree_cache;
 690         struct btree *b;
 691         u32 seq;
 692
 693         BUG_ON(level + 1 >= BTREE_MAX_DEPTH);
 694         /*
 695          * Parent node must be locked, else we could read in a btree node that's
 696          * been freed:
 697          */
 698         if (trans && !bch2_btree_node_relock(trans, path, level + 1)) {
 699                 trace_trans_restart_relock_parent_for_fill(trans->fn,
 700                                         _THIS_IP_, btree_id, &path->pos);
 701                 btree_trans_restart(trans);
 702                 return ERR_PTR(-EINTR);
 703         }
 704
 705         b = bch2_btree_node_mem_alloc(c, level != 0);
 706
 707         if (trans && b == ERR_PTR(-ENOMEM)) {
 708                 trans->memory_allocation_failure = true;
 709                 trace_trans_restart_memory_allocation_failure(trans->fn,
 710                                 _THIS_IP_, btree_id, &path->pos);
 711                 btree_trans_restart(trans);
 712                 return ERR_PTR(-EINTR);
 713         }
 714
 715         if (IS_ERR(b))
 716                 return b;
 717
 718         bkey_copy(&b->key, k);
 719         if (bch2_btree_node_hash_insert(bc, b, level, btree_id)) {
 720                 /* raced with another fill: */
 721
 722                 /* mark as unhashed... */
 723                 b->hash_val = 0;
 724
 725                 mutex_lock(&bc->lock);
 726                 list_add(&b->list, &bc->freeable);
 727                 mutex_unlock(&bc->lock);
 728
 729                 six_unlock_write(&b->c.lock);
 730                 six_unlock_intent(&b->c.lock);
 731                 return NULL;
 732         }
 733
 734         set_btree_node_read_in_flight(b);
 735
 736         six_unlock_write(&b->c.lock);
 737         seq = b->c.lock.state.seq;
 738         six_unlock_intent(&b->c.lock);
 739
 740         /* Unlock before doing IO: */
 741         if (trans && sync)
 742                 bch2_trans_unlock(trans);
 743
 744         bch2_btree_node_read(c, b, sync);
 745
 746         if (!sync)
 747                 return NULL;
 748
 749         if (trans &&
 750             (!bch2_trans_relock(trans) ||
 751              !bch2_btree_path_relock_intent(trans, path))) {
 752                 BUG_ON(!trans->restarted);
 753                 return ERR_PTR(-EINTR);
 754         }
 755
 756         if (!six_relock_type(&b->c.lock, lock_type, seq)) {
 757                 trace_trans_restart_relock_after_fill(trans->fn, _THIS_IP_,
 758                                            btree_id, &path->pos);
 759                 btree_trans_restart(trans);
 760                 return ERR_PTR(-EINTR);
 761         }
 762
 763         return b;
 764 }
 765
 766 static int lock_node_check_fn(struct six_lock *lock, void *p)
 767 {
 768         struct btree *b = container_of(lock, struct btree, c.lock);
 769         const struct bkey_i *k = p;
 770
 771         return b->hash_val == btree_ptr_hash_val(k) ? 0 : -1;
 772 }
 773
 774 static noinline void btree_bad_header(struct bch_fs *c, struct btree *b)
 775 {
 776         struct printbuf buf1 = PRINTBUF;
 777         struct printbuf buf2 = PRINTBUF;
 778         struct printbuf buf3 = PRINTBUF;
 779
 780         if (!test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags))
 781                 return;
 782
 783         bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(&b->key));
 784         bch2_bpos_to_text(&buf2, b->data->min_key);
 785         bch2_bpos_to_text(&buf3, b->data->max_key);
 786
 787         bch2_fs_inconsistent(c, "btree node header doesn't match ptr\n"
 788                              "btree %s level %u\n"
 789                              "ptr: %s\n"
 790                              "header: btree %s level %llu\n"
 791                              "min %s max %s\n",
 792                              bch2_btree_ids[b->c.btree_id], b->c.level,
 793                              buf1.buf,
 794                              bch2_btree_ids[BTREE_NODE_ID(b->data)],
 795                              BTREE_NODE_LEVEL(b->data),
 796                              buf2.buf, buf3.buf);
 797
 798         printbuf_exit(&buf3);
 799         printbuf_exit(&buf2);
 800         printbuf_exit(&buf1);
 801 }
 802
 803 static inline void btree_check_header(struct bch_fs *c, struct btree *b)
 804 {
 805         if (b->c.btree_id != BTREE_NODE_ID(b->data) ||
 806             b->c.level != BTREE_NODE_LEVEL(b->data) ||
 807             bpos_cmp(b->data->max_key, b->key.k.p) ||
 808             (b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 809              bpos_cmp(b->data->min_key,
 810                       bkey_i_to_btree_ptr_v2(&b->key)->v.min_key)))
 811                 btree_bad_header(c, b);
 812 }
 813
 814 /**
 815  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 816  * in from disk if necessary.
 817  *
 818  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
 819  *
 820  * The btree node will have either a read or a write lock held, depending on
 821  * the @write parameter.
 822  */
 823 struct btree *bch2_btree_node_get(struct btree_trans *trans, struct btree_path *path,
 824                                   const struct bkey_i *k, unsigned level,
 825                                   enum six_lock_type lock_type,
 826                                   unsigned long trace_ip)
 827 {
 828         struct bch_fs *c = trans->c;
 829         struct btree_cache *bc = &c->btree_cache;
 830         struct btree *b;
 831         struct bset_tree *t;
 832
 833         EBUG_ON(level >= BTREE_MAX_DEPTH);
 834
 835         b = btree_node_mem_ptr(k);
 836
 837         /*
 838          * Check b->hash_val _before_ calling btree_node_lock() - this might not
 839          * be the node we want anymore, and trying to lock the wrong node could
 840          * cause an unneccessary transaction restart:
 841          */
 842         if (likely(c->opts.btree_node_mem_ptr_optimization &&
 843                    b &&
 844                    b->hash_val == btree_ptr_hash_val(k)))
 845                         goto lock_node;
 846 retry:
 847         b = btree_cache_find(bc, k);
 848         if (unlikely(!b)) {
 849                 /*
 850                  * We must have the parent locked to call bch2_btree_node_fill(),
 851                  * else we could read in a btree node from disk that's been
 852                  * freed:
 853                  */
 854                 b = bch2_btree_node_fill(c, trans, path, k, path->btree_id,
 855                                          level, lock_type, true);
 856
 857                 /* We raced and found the btree node in the cache */
 858                 if (!b)
 859                         goto retry;
 860
 861                 if (IS_ERR(b))
 862                         return b;
 863         } else {
 864 lock_node:
 865                 /*
 866                  * There's a potential deadlock with splits and insertions into
 867                  * interior nodes we have to avoid:
 868                  *
 869                  * The other thread might be holding an intent lock on the node
 870                  * we want, and they want to update its parent node so they're
 871                  * going to upgrade their intent lock on the parent node to a
 872                  * write lock.
 873                  *
 874                  * But if we're holding a read lock on the parent, and we're
 875                  * trying to get the intent lock they're holding, we deadlock.
 876                  *
 877                  * So to avoid this we drop the read locks on parent nodes when
 878                  * we're starting to take intent locks - and handle the race.
 879                  *
 880                  * The race is that they might be about to free the node we
 881                  * want, and dropping our read lock on the parent node lets them
 882                  * update the parent marking the node we want as freed, and then
 883                  * free it:
 884                  *
 885                  * To guard against this, btree nodes are evicted from the cache
 886                  * when they're freed - and b->hash_val is zeroed out, which we
 887                  * check for after we lock the node.
 888                  *
 889                  * Then, bch2_btree_node_relock() on the parent will fail - because
 890                  * the parent was modified, when the pointer to the node we want
 891                  * was removed - and we'll bail out:
 892                  */
 893                 if (btree_node_read_locked(path, level + 1))
 894                         btree_node_unlock(path, level + 1);
 895
 896                 if (!btree_node_lock(trans, path, b, k->k.p, level, lock_type,
 897                                      lock_node_check_fn, (void *) k, trace_ip)) {
 898                         if (!trans->restarted)
 899                                 goto retry;
 900                         return ERR_PTR(-EINTR);
 901                 }
 902
 903                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
 904                              b->c.level != level ||
 905                              race_fault())) {
 906                         six_unlock_type(&b->c.lock, lock_type);
 907                         if (bch2_btree_node_relock(trans, path, level + 1))
 908                                 goto retry;
 909
 910                         trace_trans_restart_btree_node_reused(trans->fn,
 911                                                               trace_ip,
 912                                                               path->btree_id,
 913                                                               &path->pos);
 914                         btree_trans_restart(trans);
 915                         return ERR_PTR(-EINTR);
 916                 }
 917         }
 918
 919         if (unlikely(btree_node_read_in_flight(b))) {
 920                 u32 seq = b->c.lock.state.seq;
 921
 922                 six_unlock_type(&b->c.lock, lock_type);
 923                 bch2_trans_unlock(trans);
 924
 925                 bch2_btree_node_wait_on_read(b);
 926
 927                 /*
 928                  * should_be_locked is not set on this path yet, so we need to
 929                  * relock it specifically:
 930                  */
 931                 if (trans &&
 932                     (!bch2_trans_relock(trans) ||
 933                      !bch2_btree_path_relock_intent(trans, path))) {
 934                         BUG_ON(!trans->restarted);
 935                         return ERR_PTR(-EINTR);
 936                 }
 937
 938                 if (!six_relock_type(&b->c.lock, lock_type, seq))
 939                         goto retry;
 940         }
 941
 942         prefetch(b->aux_data);
 943
 944         for_each_bset(b, t) {
 945                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 946
 947                 prefetch(p + L1_CACHE_BYTES * 0);
 948                 prefetch(p + L1_CACHE_BYTES * 1);
 949                 prefetch(p + L1_CACHE_BYTES * 2);
 950         }
 951
 952         /* avoid atomic set bit if it's not needed: */
 953         if (!btree_node_accessed(b))
 954                 set_btree_node_accessed(b);
 955
 956         if (unlikely(btree_node_read_error(b))) {
 957                 six_unlock_type(&b->c.lock, lock_type);
 958                 return ERR_PTR(-EIO);
 959         }
 960
 961         EBUG_ON(b->c.btree_id != path->btree_id);
 962         EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
 963         btree_check_header(c, b);
 964
 965         return b;
 966 }
 967
 968 struct btree *bch2_btree_node_get_noiter(struct bch_fs *c,
 969                                          const struct bkey_i *k,
 970                                          enum btree_id btree_id,
 971                                          unsigned level,
 972                                          bool nofill)
 973 {
 974         struct btree_cache *bc = &c->btree_cache;
 975         struct btree *b;
 976         struct bset_tree *t;
 977         int ret;
 978
 979         EBUG_ON(level >= BTREE_MAX_DEPTH);
 980
 981         if (c->opts.btree_node_mem_ptr_optimization) {
 982                 b = btree_node_mem_ptr(k);
 983                 if (b)
 984                         goto lock_node;
 985         }
 986 retry:
 987         b = btree_cache_find(bc, k);
 988         if (unlikely(!b)) {
 989                 if (nofill)
 990                         goto out;
 991
 992                 b = bch2_btree_node_fill(c, NULL, NULL, k, btree_id,
 993                                          level, SIX_LOCK_read, true);
 994
 995                 /* We raced and found the btree node in the cache */
 996                 if (!b)
 997                         goto retry;
 998
 999                 if (IS_ERR(b) &&
1000                     !bch2_btree_cache_cannibalize_lock(c, NULL))
1001                         goto retry;
1002
1003                 if (IS_ERR(b))
1004                         goto out;
1005         } else {
1006 lock_node:
1007                 ret = six_lock_read(&b->c.lock, lock_node_check_fn, (void *) k);
1008                 if (ret)
1009                         goto retry;
1010
1011                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
1012                              b->c.btree_id != btree_id ||
1013                              b->c.level != level)) {
1014                         six_unlock_read(&b->c.lock);
1015                         goto retry;
1016                 }
1017         }
1018
1019         /* XXX: waiting on IO with btree locks held: */
1020         __bch2_btree_node_wait_on_read(b);
1021
1022         prefetch(b->aux_data);
1023
1024         for_each_bset(b, t) {
1025                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
1026
1027                 prefetch(p + L1_CACHE_BYTES * 0);
1028                 prefetch(p + L1_CACHE_BYTES * 1);
1029                 prefetch(p + L1_CACHE_BYTES * 2);
1030         }
1031
1032         /* avoid atomic set bit if it's not needed: */
1033         if (!btree_node_accessed(b))
1034                 set_btree_node_accessed(b);
1035
1036         if (unlikely(btree_node_read_error(b))) {
1037                 six_unlock_read(&b->c.lock);
1038                 b = ERR_PTR(-EIO);
1039                 goto out;
1040         }
1041
1042         EBUG_ON(b->c.btree_id != btree_id);
1043         EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
1044         btree_check_header(c, b);
1045 out:
1046         bch2_btree_cache_cannibalize_unlock(c);
1047         return b;
1048 }
1049
1050 int bch2_btree_node_prefetch(struct bch_fs *c,
1051                              struct btree_trans *trans,
1052                              struct btree_path *path,
1053                              const struct bkey_i *k,
1054                              enum btree_id btree_id, unsigned level)
1055 {
1056         struct btree_cache *bc = &c->btree_cache;
1057         struct btree *b;
1058
1059         BUG_ON(trans && !btree_node_locked(path, level + 1));
1060         BUG_ON(level >= BTREE_MAX_DEPTH);
1061
1062         b = btree_cache_find(bc, k);
1063         if (b)
1064                 return 0;
1065
1066         b = bch2_btree_node_fill(c, trans, path, k, btree_id,
1067                                  level, SIX_LOCK_read, false);
1068         return PTR_ERR_OR_ZERO(b);
1069 }
1070
1071 void bch2_btree_node_evict(struct bch_fs *c, const struct bkey_i *k)
1072 {
1073         struct btree_cache *bc = &c->btree_cache;
1074         struct btree *b;
1075
1076         b = btree_cache_find(bc, k);
1077         if (!b)
1078                 return;
1079 wait_on_io:
1080         /* not allowed to wait on io with btree locks held: */
1081
1082         /* XXX we're called from btree_gc which will be holding other btree
1083          * nodes locked
1084          * */
1085         __bch2_btree_node_wait_on_read(b);
1086         __bch2_btree_node_wait_on_write(b);
1087
1088         six_lock_intent(&b->c.lock, NULL, NULL);
1089         six_lock_write(&b->c.lock, NULL, NULL);
1090
1091         if (btree_node_dirty(b)) {
1092                 __bch2_btree_node_write(c, b, 0);
1093                 six_unlock_write(&b->c.lock);
1094                 six_unlock_intent(&b->c.lock);
1095                 goto wait_on_io;
1096         }
1097
1098         BUG_ON(btree_node_dirty(b));
1099
1100         mutex_lock(&bc->lock);
1101         btree_node_data_free(c, b);
1102         bch2_btree_node_hash_remove(bc, b);
1103         mutex_unlock(&bc->lock);
1104
1105         six_unlock_write(&b->c.lock);
1106         six_unlock_intent(&b->c.lock);
1107 }
1108
1109 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
1110                              struct btree *b)
1111 {
1112         const struct bkey_format *f = &b->format;
1113         struct bset_stats stats;
1114
1115         memset(&stats, 0, sizeof(stats));
1116
1117         bch2_btree_keys_stats(b, &stats);
1118
1119         pr_buf(out, "l %u ", b->c.level);
1120         bch2_bpos_to_text(out, b->data->min_key);
1121         pr_buf(out, " - ");
1122         bch2_bpos_to_text(out, b->data->max_key);
1123         pr_buf(out, ":\n"
1124                "    ptrs: ");
1125         bch2_val_to_text(out, c, bkey_i_to_s_c(&b->key));
1126
1127         pr_buf(out, "\n"
1128                "    format: u64s %u fields %u %u %u %u %u\n"
1129                "    unpack fn len: %u\n"
1130                "    bytes used %zu/%zu (%zu%% full)\n"
1131                "    sib u64s: %u, %u (merge threshold %u)\n"
1132                "    nr packed keys %u\n"
1133                "    nr unpacked keys %u\n"
1134                "    floats %zu\n"
1135                "    failed unpacked %zu\n",
1136                f->key_u64s,
1137                f->bits_per_field[0],
1138                f->bits_per_field[1],
1139                f->bits_per_field[2],
1140                f->bits_per_field[3],
1141                f->bits_per_field[4],
1142                b->unpack_fn_len,
1143                b->nr.live_u64s * sizeof(u64),
1144                btree_bytes(c) - sizeof(struct btree_node),
1145                b->nr.live_u64s * 100 / btree_max_u64s(c),
1146                b->sib_u64s[0],
1147                b->sib_u64s[1],
1148                c->btree_foreground_merge_threshold,
1149                b->nr.packed_keys,
1150                b->nr.unpacked_keys,
1151                stats.floats,
1152                stats.failed);
1153 }
1154
1155 void bch2_btree_cache_to_text(struct printbuf *out, struct bch_fs *c)
1156 {
1157         pr_buf(out, "nr nodes:\t\t%u\n", c->btree_cache.used);
1158         pr_buf(out, "nr dirty:\t\t%u\n", atomic_read(&c->btree_cache.dirty));
1159         pr_buf(out, "cannibalize lock:\t%p\n", c->btree_cache.alloc_lock);
1160 }