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 const char * const bch2_btree_node_flags[] = {
  18 #define x(f)    #f,
  19         BTREE_FLAGS()
  20 #undef x
  21         NULL
  22 };
  23
  24 void bch2_recalc_btree_reserve(struct bch_fs *c)
  25 {
  26         unsigned i, reserve = 16;
  27
  28         if (!c->btree_roots[0].b)
  29                 reserve += 8;
  30
  31         for (i = 0; i < BTREE_ID_NR; i++)
  32                 if (c->btree_roots[i].b)
  33                         reserve += min_t(unsigned, 1,
  34                                          c->btree_roots[i].b->c.level) * 8;
  35
  36         c->btree_cache.reserve = reserve;
  37 }
  38
  39 static inline unsigned btree_cache_can_free(struct btree_cache *bc)
  40 {
  41         return max_t(int, 0, bc->used - bc->reserve);
  42 }
  43
  44 static void btree_node_to_freedlist(struct btree_cache *bc, struct btree *b)
  45 {
  46         if (b->c.lock.readers)
  47                 list_move(&b->list, &bc->freed_pcpu);
  48         else
  49                 list_move(&b->list, &bc->freed_nonpcpu);
  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         EBUG_ON(btree_node_write_in_flight(b));
  57
  58         kvpfree(b->data, btree_bytes(c));
  59         b->data = NULL;
  60 #ifdef __KERNEL__
  61         vfree(b->aux_data);
  62 #else
  63         munmap(b->aux_data, btree_aux_data_bytes(b));
  64 #endif
  65         b->aux_data = NULL;
  66
  67         bc->used--;
  68
  69         btree_node_to_freedlist(bc, b);
  70 }
  71
  72 static int bch2_btree_cache_cmp_fn(struct rhashtable_compare_arg *arg,
  73                                    const void *obj)
  74 {
  75         const struct btree *b = obj;
  76         const u64 *v = arg->key;
  77
  78         return b->hash_val == *v ? 0 : 1;
  79 }
  80
  81 static const struct rhashtable_params bch_btree_cache_params = {
  82         .head_offset    = offsetof(struct btree, hash),
  83         .key_offset     = offsetof(struct btree, hash_val),
  84         .key_len        = sizeof(u64),
  85         .obj_cmpfn      = bch2_btree_cache_cmp_fn,
  86 };
  87
  88 static int btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
  89 {
  90         BUG_ON(b->data || b->aux_data);
  91
  92         b->data = kvpmalloc(btree_bytes(c), gfp);
  93         if (!b->data)
  94                 return -ENOMEM;
  95 #ifdef __KERNEL__
  96         b->aux_data = vmalloc_exec(btree_aux_data_bytes(b), gfp);
  97 #else
  98         b->aux_data = mmap(NULL, btree_aux_data_bytes(b),
  99                            PROT_READ|PROT_WRITE|PROT_EXEC,
 100                            MAP_PRIVATE|MAP_ANONYMOUS, 0, 0);
 101         if (b->aux_data == MAP_FAILED)
 102                 b->aux_data = NULL;
 103 #endif
 104         if (!b->aux_data) {
 105                 kvpfree(b->data, btree_bytes(c));
 106                 b->data = NULL;
 107                 return -ENOMEM;
 108         }
 109
 110         return 0;
 111 }
 112
 113 static struct btree *__btree_node_mem_alloc(struct bch_fs *c)
 114 {
 115         struct btree *b = kzalloc(sizeof(struct btree), GFP_KERNEL);
 116         if (!b)
 117                 return NULL;
 118
 119         bkey_btree_ptr_init(&b->key);
 120         __six_lock_init(&b->c.lock, "b->c.lock", &bch2_btree_node_lock_key);
 121         INIT_LIST_HEAD(&b->list);
 122         INIT_LIST_HEAD(&b->write_blocked);
 123         b->byte_order = ilog2(btree_bytes(c));
 124         return b;
 125 }
 126
 127 struct btree *__bch2_btree_node_mem_alloc(struct bch_fs *c)
 128 {
 129         struct btree_cache *bc = &c->btree_cache;
 130         struct btree *b = __btree_node_mem_alloc(c);
 131         if (!b)
 132                 return NULL;
 133
 134         if (btree_node_data_alloc(c, b, GFP_KERNEL)) {
 135                 kfree(b);
 136                 return NULL;
 137         }
 138
 139         bc->used++;
 140         list_add(&b->list, &bc->freeable);
 141         return b;
 142 }
 143
 144 /* Btree in memory cache - hash table */
 145
 146 void bch2_btree_node_hash_remove(struct btree_cache *bc, struct btree *b)
 147 {
 148         int ret = rhashtable_remove_fast(&bc->table, &b->hash, bch_btree_cache_params);
 149         BUG_ON(ret);
 150
 151         /* Cause future lookups for this node to fail: */
 152         b->hash_val = 0;
 153
 154         six_lock_wakeup_all(&b->c.lock);
 155 }
 156
 157 int __bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b)
 158 {
 159         BUG_ON(b->hash_val);
 160         b->hash_val = btree_ptr_hash_val(&b->key);
 161
 162         return rhashtable_lookup_insert_fast(&bc->table, &b->hash,
 163                                              bch_btree_cache_params);
 164 }
 165
 166 int bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b,
 167                                 unsigned level, enum btree_id id)
 168 {
 169         int ret;
 170
 171         b->c.level      = level;
 172         b->c.btree_id   = id;
 173
 174         mutex_lock(&bc->lock);
 175         ret = __bch2_btree_node_hash_insert(bc, b);
 176         if (!ret)
 177                 list_add(&b->list, &bc->live);
 178         mutex_unlock(&bc->lock);
 179
 180         return ret;
 181 }
 182
 183 __flatten
 184 static inline struct btree *btree_cache_find(struct btree_cache *bc,
 185                                      const struct bkey_i *k)
 186 {
 187         u64 v = btree_ptr_hash_val(k);
 188
 189         return rhashtable_lookup_fast(&bc->table, &v, bch_btree_cache_params);
 190 }
 191
 192 /*
 193  * this version is for btree nodes that have already been freed (we're not
 194  * reaping a real btree node)
 195  */
 196 static int __btree_node_reclaim(struct bch_fs *c, struct btree *b, bool flush)
 197 {
 198         struct btree_cache *bc = &c->btree_cache;
 199         int ret = 0;
 200
 201         lockdep_assert_held(&bc->lock);
 202 wait_on_io:
 203         if (b->flags & ((1U << BTREE_NODE_dirty)|
 204                         (1U << BTREE_NODE_read_in_flight)|
 205                         (1U << BTREE_NODE_write_in_flight))) {
 206                 if (!flush)
 207                         return -ENOMEM;
 208
 209                 /* XXX: waiting on IO with btree cache lock held */
 210                 bch2_btree_node_wait_on_read(b);
 211                 bch2_btree_node_wait_on_write(b);
 212         }
 213
 214         if (!six_trylock_intent(&b->c.lock))
 215                 return -ENOMEM;
 216
 217         if (!six_trylock_write(&b->c.lock))
 218                 goto out_unlock_intent;
 219
 220         /* recheck under lock */
 221         if (b->flags & ((1U << BTREE_NODE_read_in_flight)|
 222                         (1U << BTREE_NODE_write_in_flight))) {
 223                 if (!flush)
 224                         goto out_unlock;
 225                 six_unlock_write(&b->c.lock);
 226                 six_unlock_intent(&b->c.lock);
 227                 goto wait_on_io;
 228         }
 229
 230         if (btree_node_noevict(b) ||
 231             btree_node_write_blocked(b) ||
 232             btree_node_will_make_reachable(b))
 233                 goto out_unlock;
 234
 235         if (btree_node_dirty(b)) {
 236                 if (!flush)
 237                         goto out_unlock;
 238                 /*
 239                  * Using the underscore version because we don't want to compact
 240                  * bsets after the write, since this node is about to be evicted
 241                  * - unless btree verify mode is enabled, since it runs out of
 242                  * the post write cleanup:
 243                  */
 244                 if (bch2_verify_btree_ondisk)
 245                         bch2_btree_node_write(c, b, SIX_LOCK_intent, 0);
 246                 else
 247                         __bch2_btree_node_write(c, b, 0);
 248
 249                 six_unlock_write(&b->c.lock);
 250                 six_unlock_intent(&b->c.lock);
 251                 goto wait_on_io;
 252         }
 253 out:
 254         if (b->hash_val && !ret)
 255                 trace_and_count(c, btree_cache_reap, c, b);
 256         return ret;
 257 out_unlock:
 258         six_unlock_write(&b->c.lock);
 259 out_unlock_intent:
 260         six_unlock_intent(&b->c.lock);
 261         ret = -ENOMEM;
 262         goto out;
 263 }
 264
 265 static int btree_node_reclaim(struct bch_fs *c, struct btree *b)
 266 {
 267         return __btree_node_reclaim(c, b, false);
 268 }
 269
 270 static int btree_node_write_and_reclaim(struct bch_fs *c, struct btree *b)
 271 {
 272         return __btree_node_reclaim(c, b, true);
 273 }
 274
 275 static unsigned long bch2_btree_cache_scan(struct shrinker *shrink,
 276                                            struct shrink_control *sc)
 277 {
 278         struct bch_fs *c = container_of(shrink, struct bch_fs,
 279                                         btree_cache.shrink);
 280         struct btree_cache *bc = &c->btree_cache;
 281         struct btree *b, *t;
 282         unsigned long nr = sc->nr_to_scan;
 283         unsigned long can_free = 0;
 284         unsigned long touched = 0;
 285         unsigned long freed = 0;
 286         unsigned i, flags;
 287         unsigned long ret = SHRINK_STOP;
 288
 289         if (bch2_btree_shrinker_disabled)
 290                 return SHRINK_STOP;
 291
 292         /* Return -1 if we can't do anything right now */
 293         if (sc->gfp_mask & __GFP_FS)
 294                 mutex_lock(&bc->lock);
 295         else if (!mutex_trylock(&bc->lock))
 296                 goto out_norestore;
 297
 298         flags = memalloc_nofs_save();
 299
 300         /*
 301          * It's _really_ critical that we don't free too many btree nodes - we
 302          * have to always leave ourselves a reserve. The reserve is how we
 303          * guarantee that allocating memory for a new btree node can always
 304          * succeed, so that inserting keys into the btree can always succeed and
 305          * IO can always make forward progress:
 306          */
 307         can_free = btree_cache_can_free(bc);
 308         nr = min_t(unsigned long, nr, can_free);
 309
 310         i = 0;
 311         list_for_each_entry_safe(b, t, &bc->freeable, list) {
 312                 /*
 313                  * Leave a few nodes on the freeable list, so that a btree split
 314                  * won't have to hit the system allocator:
 315                  */
 316                 if (++i <= 3)
 317                         continue;
 318
 319                 touched++;
 320
 321                 if (touched >= nr)
 322                         break;
 323
 324                 if (!btree_node_reclaim(c, b)) {
 325                         btree_node_data_free(c, b);
 326                         six_unlock_write(&b->c.lock);
 327                         six_unlock_intent(&b->c.lock);
 328                         freed++;
 329                 }
 330         }
 331 restart:
 332         list_for_each_entry_safe(b, t, &bc->live, list) {
 333                 /* tweak this */
 334                 if (btree_node_accessed(b)) {
 335                         clear_btree_node_accessed(b);
 336                         goto touched;
 337                 }
 338
 339                 if (!btree_node_reclaim(c, b)) {
 340                         /* can't call bch2_btree_node_hash_remove under lock  */
 341                         freed++;
 342                         if (&t->list != &bc->live)
 343                                 list_move_tail(&bc->live, &t->list);
 344
 345                         btree_node_data_free(c, b);
 346                         mutex_unlock(&bc->lock);
 347
 348                         bch2_btree_node_hash_remove(bc, b);
 349                         six_unlock_write(&b->c.lock);
 350                         six_unlock_intent(&b->c.lock);
 351
 352                         if (freed >= nr)
 353                                 goto out;
 354
 355                         if (sc->gfp_mask & __GFP_FS)
 356                                 mutex_lock(&bc->lock);
 357                         else if (!mutex_trylock(&bc->lock))
 358                                 goto out;
 359                         goto restart;
 360                 } else {
 361                         continue;
 362                 }
 363 touched:
 364                 touched++;
 365
 366                 if (touched >= nr) {
 367                         /* Save position */
 368                         if (&t->list != &bc->live)
 369                                 list_move_tail(&bc->live, &t->list);
 370                         break;
 371                 }
 372         }
 373
 374         mutex_unlock(&bc->lock);
 375 out:
 376         ret = freed;
 377         memalloc_nofs_restore(flags);
 378 out_norestore:
 379         trace_and_count(c, btree_cache_scan, sc->nr_to_scan, can_free, ret);
 380         return ret;
 381 }
 382
 383 static unsigned long bch2_btree_cache_count(struct shrinker *shrink,
 384                                             struct shrink_control *sc)
 385 {
 386         struct bch_fs *c = container_of(shrink, struct bch_fs,
 387                                         btree_cache.shrink);
 388         struct btree_cache *bc = &c->btree_cache;
 389
 390         if (bch2_btree_shrinker_disabled)
 391                 return 0;
 392
 393         return btree_cache_can_free(bc);
 394 }
 395
 396 static void bch2_btree_cache_shrinker_to_text(struct printbuf *out, struct shrinker *shrink)
 397 {
 398         struct bch_fs *c = container_of(shrink, struct bch_fs,
 399                                         btree_cache.shrink);
 400
 401         bch2_btree_cache_to_text(out, c);
 402 }
 403
 404 void bch2_fs_btree_cache_exit(struct bch_fs *c)
 405 {
 406         struct btree_cache *bc = &c->btree_cache;
 407         struct btree *b;
 408         unsigned i, flags;
 409
 410         if (bc->shrink.list.next)
 411                 unregister_shrinker(&bc->shrink);
 412
 413         /* vfree() can allocate memory: */
 414         flags = memalloc_nofs_save();
 415         mutex_lock(&bc->lock);
 416
 417         if (c->verify_data)
 418                 list_move(&c->verify_data->list, &bc->live);
 419
 420         kvpfree(c->verify_ondisk, btree_bytes(c));
 421
 422         for (i = 0; i < BTREE_ID_NR; i++)
 423                 if (c->btree_roots[i].b)
 424                         list_add(&c->btree_roots[i].b->list, &bc->live);
 425
 426         list_splice(&bc->freeable, &bc->live);
 427
 428         while (!list_empty(&bc->live)) {
 429                 b = list_first_entry(&bc->live, struct btree, list);
 430
 431                 BUG_ON(btree_node_read_in_flight(b) ||
 432                        btree_node_write_in_flight(b));
 433
 434                 if (btree_node_dirty(b))
 435                         bch2_btree_complete_write(c, b, btree_current_write(b));
 436                 clear_btree_node_dirty_acct(c, b);
 437
 438                 btree_node_data_free(c, b);
 439         }
 440
 441         BUG_ON(atomic_read(&c->btree_cache.dirty));
 442
 443         list_splice(&bc->freed_pcpu, &bc->freed_nonpcpu);
 444
 445         while (!list_empty(&bc->freed_nonpcpu)) {
 446                 b = list_first_entry(&bc->freed_nonpcpu, struct btree, list);
 447                 list_del(&b->list);
 448                 six_lock_pcpu_free(&b->c.lock);
 449                 kfree(b);
 450         }
 451
 452         mutex_unlock(&bc->lock);
 453         memalloc_nofs_restore(flags);
 454
 455         if (bc->table_init_done)
 456                 rhashtable_destroy(&bc->table);
 457 }
 458
 459 int bch2_fs_btree_cache_init(struct bch_fs *c)
 460 {
 461         struct btree_cache *bc = &c->btree_cache;
 462         unsigned i;
 463         int ret = 0;
 464
 465         pr_verbose_init(c->opts, "");
 466
 467         ret = rhashtable_init(&bc->table, &bch_btree_cache_params);
 468         if (ret)
 469                 goto out;
 470
 471         bc->table_init_done = true;
 472
 473         bch2_recalc_btree_reserve(c);
 474
 475         for (i = 0; i < bc->reserve; i++)
 476                 if (!__bch2_btree_node_mem_alloc(c)) {
 477                         ret = -ENOMEM;
 478                         goto out;
 479                 }
 480
 481         list_splice_init(&bc->live, &bc->freeable);
 482
 483         mutex_init(&c->verify_lock);
 484
 485         bc->shrink.count_objects        = bch2_btree_cache_count;
 486         bc->shrink.scan_objects         = bch2_btree_cache_scan;
 487         bc->shrink.to_text              = bch2_btree_cache_shrinker_to_text;
 488         bc->shrink.seeks                = 4;
 489         ret = register_shrinker(&bc->shrink);
 490 out:
 491         pr_verbose_init(c->opts, "ret %i", ret);
 492         return ret;
 493 }
 494
 495 void bch2_fs_btree_cache_init_early(struct btree_cache *bc)
 496 {
 497         mutex_init(&bc->lock);
 498         INIT_LIST_HEAD(&bc->live);
 499         INIT_LIST_HEAD(&bc->freeable);
 500         INIT_LIST_HEAD(&bc->freed_pcpu);
 501         INIT_LIST_HEAD(&bc->freed_nonpcpu);
 502 }
 503
 504 /*
 505  * We can only have one thread cannibalizing other cached btree nodes at a time,
 506  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 507  * cannibalize_bucket() will take. This means every time we unlock the root of
 508  * the btree, we need to release this lock if we have it held.
 509  */
 510 void bch2_btree_cache_cannibalize_unlock(struct bch_fs *c)
 511 {
 512         struct btree_cache *bc = &c->btree_cache;
 513
 514         if (bc->alloc_lock == current) {
 515                 trace_and_count(c, btree_cache_cannibalize_unlock, c);
 516                 bc->alloc_lock = NULL;
 517                 closure_wake_up(&bc->alloc_wait);
 518         }
 519 }
 520
 521 int bch2_btree_cache_cannibalize_lock(struct bch_fs *c, struct closure *cl)
 522 {
 523         struct btree_cache *bc = &c->btree_cache;
 524         struct task_struct *old;
 525
 526         old = cmpxchg(&bc->alloc_lock, NULL, current);
 527         if (old == NULL || old == current)
 528                 goto success;
 529
 530         if (!cl) {
 531                 trace_and_count(c, btree_cache_cannibalize_lock_fail, c);
 532                 return -ENOMEM;
 533         }
 534
 535         closure_wait(&bc->alloc_wait, cl);
 536
 537         /* Try again, after adding ourselves to waitlist */
 538         old = cmpxchg(&bc->alloc_lock, NULL, current);
 539         if (old == NULL || old == current) {
 540                 /* We raced */
 541                 closure_wake_up(&bc->alloc_wait);
 542                 goto success;
 543         }
 544
 545         trace_and_count(c, btree_cache_cannibalize_lock_fail, c);
 546         return -EAGAIN;
 547
 548 success:
 549         trace_and_count(c, btree_cache_cannibalize_lock, c);
 550         return 0;
 551 }
 552
 553 static struct btree *btree_node_cannibalize(struct bch_fs *c)
 554 {
 555         struct btree_cache *bc = &c->btree_cache;
 556         struct btree *b;
 557
 558         list_for_each_entry_reverse(b, &bc->live, list)
 559                 if (!btree_node_reclaim(c, b))
 560                         return b;
 561
 562         while (1) {
 563                 list_for_each_entry_reverse(b, &bc->live, list)
 564                         if (!btree_node_write_and_reclaim(c, b))
 565                                 return b;
 566
 567                 /*
 568                  * Rare case: all nodes were intent-locked.
 569                  * Just busy-wait.
 570                  */
 571                 WARN_ONCE(1, "btree cache cannibalize failed\n");
 572                 cond_resched();
 573         }
 574 }
 575
 576 struct btree *bch2_btree_node_mem_alloc(struct bch_fs *c, bool pcpu_read_locks)
 577 {
 578         struct btree_cache *bc = &c->btree_cache;
 579         struct list_head *freed = pcpu_read_locks
 580                 ? &bc->freed_pcpu
 581                 : &bc->freed_nonpcpu;
 582         struct btree *b, *b2;
 583         u64 start_time = local_clock();
 584         unsigned flags;
 585
 586         flags = memalloc_nofs_save();
 587         mutex_lock(&bc->lock);
 588
 589         /*
 590          * We never free struct btree itself, just the memory that holds the on
 591          * disk node. Check the freed list before allocating a new one:
 592          */
 593         list_for_each_entry(b, freed, list)
 594                 if (!btree_node_reclaim(c, b)) {
 595                         list_del_init(&b->list);
 596                         goto got_node;
 597                 }
 598
 599         b = __btree_node_mem_alloc(c);
 600         if (!b)
 601                 goto err_locked;
 602
 603         if (pcpu_read_locks)
 604                 six_lock_pcpu_alloc(&b->c.lock);
 605
 606         BUG_ON(!six_trylock_intent(&b->c.lock));
 607         BUG_ON(!six_trylock_write(&b->c.lock));
 608 got_node:
 609
 610         /*
 611          * btree_free() doesn't free memory; it sticks the node on the end of
 612          * the list. Check if there's any freed nodes there:
 613          */
 614         list_for_each_entry(b2, &bc->freeable, list)
 615                 if (!btree_node_reclaim(c, b2)) {
 616                         swap(b->data, b2->data);
 617                         swap(b->aux_data, b2->aux_data);
 618                         btree_node_to_freedlist(bc, b2);
 619                         six_unlock_write(&b2->c.lock);
 620                         six_unlock_intent(&b2->c.lock);
 621                         goto got_mem;
 622                 }
 623
 624         mutex_unlock(&bc->lock);
 625
 626         if (btree_node_data_alloc(c, b, __GFP_NOWARN|GFP_KERNEL))
 627                 goto err;
 628
 629         mutex_lock(&bc->lock);
 630         bc->used++;
 631 got_mem:
 632         mutex_unlock(&bc->lock);
 633
 634         BUG_ON(btree_node_hashed(b));
 635         BUG_ON(btree_node_dirty(b));
 636         BUG_ON(btree_node_write_in_flight(b));
 637 out:
 638         b->flags                = 0;
 639         b->written              = 0;
 640         b->nsets                = 0;
 641         b->sib_u64s[0]          = 0;
 642         b->sib_u64s[1]          = 0;
 643         b->whiteout_u64s        = 0;
 644         bch2_btree_keys_init(b);
 645         set_btree_node_accessed(b);
 646
 647         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],
 648                                start_time);
 649
 650         memalloc_nofs_restore(flags);
 651         return b;
 652 err:
 653         mutex_lock(&bc->lock);
 654 err_locked:
 655         /* Try to cannibalize another cached btree node: */
 656         if (bc->alloc_lock == current) {
 657                 b2 = btree_node_cannibalize(c);
 658                 bch2_btree_node_hash_remove(bc, b2);
 659
 660                 if (b) {
 661                         swap(b->data, b2->data);
 662                         swap(b->aux_data, b2->aux_data);
 663                         btree_node_to_freedlist(bc, b2);
 664                         six_unlock_write(&b2->c.lock);
 665                         six_unlock_intent(&b2->c.lock);
 666                 } else {
 667                         b = b2;
 668                         list_del_init(&b->list);
 669                 }
 670
 671                 mutex_unlock(&bc->lock);
 672
 673                 trace_and_count(c, btree_cache_cannibalize, c);
 674                 goto out;
 675         }
 676
 677         mutex_unlock(&bc->lock);
 678         memalloc_nofs_restore(flags);
 679         return ERR_PTR(-ENOMEM);
 680 }
 681
 682 /* Slowpath, don't want it inlined into btree_iter_traverse() */
 683 static noinline struct btree *bch2_btree_node_fill(struct bch_fs *c,
 684                                 struct btree_trans *trans,
 685                                 struct btree_path *path,
 686                                 const struct bkey_i *k,
 687                                 enum btree_id btree_id,
 688                                 unsigned level,
 689                                 enum six_lock_type lock_type,
 690                                 bool sync)
 691 {
 692         struct btree_cache *bc = &c->btree_cache;
 693         struct btree *b;
 694         u32 seq;
 695
 696         BUG_ON(level + 1 >= BTREE_MAX_DEPTH);
 697         /*
 698          * Parent node must be locked, else we could read in a btree node that's
 699          * been freed:
 700          */
 701         if (trans && !bch2_btree_node_relock(trans, path, level + 1)) {
 702                 trace_and_count(c, trans_restart_relock_parent_for_fill, trans, _THIS_IP_, path);
 703                 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_fill_relock));
 704         }
 705
 706         b = bch2_btree_node_mem_alloc(c, level != 0);
 707
 708         if (trans && b == ERR_PTR(-ENOMEM)) {
 709                 trans->memory_allocation_failure = true;
 710                 trace_and_count(c, trans_restart_memory_allocation_failure, trans, _THIS_IP_, path);
 711                 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_fill_mem_alloc_fail));
 712         }
 713
 714         if (IS_ERR(b))
 715                 return b;
 716
 717         bkey_copy(&b->key, k);
 718         if (bch2_btree_node_hash_insert(bc, b, level, btree_id)) {
 719                 /* raced with another fill: */
 720
 721                 /* mark as unhashed... */
 722                 b->hash_val = 0;
 723
 724                 mutex_lock(&bc->lock);
 725                 list_add(&b->list, &bc->freeable);
 726                 mutex_unlock(&bc->lock);
 727
 728                 six_unlock_write(&b->c.lock);
 729                 six_unlock_intent(&b->c.lock);
 730                 return NULL;
 731         }
 732
 733         set_btree_node_read_in_flight(b);
 734
 735         six_unlock_write(&b->c.lock);
 736         seq = b->c.lock.state.seq;
 737         six_unlock_intent(&b->c.lock);
 738
 739         /* Unlock before doing IO: */
 740         if (trans && sync)
 741                 bch2_trans_unlock(trans);
 742
 743         bch2_btree_node_read(c, b, sync);
 744
 745         if (!sync)
 746                 return NULL;
 747
 748         if (trans) {
 749                 int ret = bch2_trans_relock(trans) ?:
 750                         bch2_btree_path_relock_intent(trans, path);
 751                 if (ret) {
 752                         BUG_ON(!trans->restarted);
 753                         return ERR_PTR(ret);
 754                 }
 755         }
 756
 757         if (!six_relock_type(&b->c.lock, lock_type, seq)) {
 758                 if (trans)
 759                         trace_and_count(c, trans_restart_relock_after_fill, trans, _THIS_IP_, path);
 760                 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_relock_after_fill));
 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         if (b->hash_val != btree_ptr_hash_val(k))
 772                 return BCH_ERR_lock_fail_node_reused;
 773         return 0;
 774 }
 775
 776 static noinline void btree_bad_header(struct bch_fs *c, struct btree *b)
 777 {
 778         struct printbuf buf = PRINTBUF;
 779
 780         if (!test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags))
 781                 return;
 782
 783         prt_printf(&buf,
 784                "btree node header doesn't match ptr\n"
 785                "btree %s level %u\n"
 786                "ptr: ",
 787                bch2_btree_ids[b->c.btree_id], b->c.level);
 788         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
 789
 790         prt_printf(&buf, "\nheader: btree %s level %llu\n"
 791                "min ",
 792                bch2_btree_ids[BTREE_NODE_ID(b->data)],
 793                BTREE_NODE_LEVEL(b->data));
 794         bch2_bpos_to_text(&buf, b->data->min_key);
 795
 796         prt_printf(&buf, "\nmax ");
 797         bch2_bpos_to_text(&buf, b->data->max_key);
 798
 799         bch2_fs_inconsistent(c, "%s", buf.buf);
 800         printbuf_exit(&buf);
 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         int ret;
 833
 834         EBUG_ON(level >= BTREE_MAX_DEPTH);
 835
 836         b = btree_node_mem_ptr(k);
 837
 838         /*
 839          * Check b->hash_val _before_ calling btree_node_lock() - this might not
 840          * be the node we want anymore, and trying to lock the wrong node could
 841          * cause an unneccessary transaction restart:
 842          */
 843         if (likely(c->opts.btree_node_mem_ptr_optimization &&
 844                    b &&
 845                    b->hash_val == btree_ptr_hash_val(k)))
 846                         goto lock_node;
 847 retry:
 848         b = btree_cache_find(bc, k);
 849         if (unlikely(!b)) {
 850                 /*
 851                  * We must have the parent locked to call bch2_btree_node_fill(),
 852                  * else we could read in a btree node from disk that's been
 853                  * freed:
 854                  */
 855                 b = bch2_btree_node_fill(c, trans, path, k, path->btree_id,
 856                                          level, lock_type, true);
 857
 858                 /* We raced and found the btree node in the cache */
 859                 if (!b)
 860                         goto retry;
 861
 862                 if (IS_ERR(b))
 863                         return b;
 864         } else {
 865 lock_node:
 866                 /*
 867                  * There's a potential deadlock with splits and insertions into
 868                  * interior nodes we have to avoid:
 869                  *
 870                  * The other thread might be holding an intent lock on the node
 871                  * we want, and they want to update its parent node so they're
 872                  * going to upgrade their intent lock on the parent node to a
 873                  * write lock.
 874                  *
 875                  * But if we're holding a read lock on the parent, and we're
 876                  * trying to get the intent lock they're holding, we deadlock.
 877                  *
 878                  * So to avoid this we drop the read locks on parent nodes when
 879                  * we're starting to take intent locks - and handle the race.
 880                  *
 881                  * The race is that they might be about to free the node we
 882                  * want, and dropping our read lock on the parent node lets them
 883                  * update the parent marking the node we want as freed, and then
 884                  * free it:
 885                  *
 886                  * To guard against this, btree nodes are evicted from the cache
 887                  * when they're freed - and b->hash_val is zeroed out, which we
 888                  * check for after we lock the node.
 889                  *
 890                  * Then, bch2_btree_node_relock() on the parent will fail - because
 891                  * the parent was modified, when the pointer to the node we want
 892                  * was removed - and we'll bail out:
 893                  */
 894                 if (btree_node_read_locked(path, level + 1))
 895                         btree_node_unlock(trans, path, level + 1);
 896
 897                 ret = btree_node_lock(trans, path, &b->c, k->k.p, level, lock_type,
 898                                       lock_node_check_fn, (void *) k, trace_ip);
 899                 if (unlikely(ret)) {
 900                         if (bch2_err_matches(ret, BCH_ERR_lock_fail_node_reused))
 901                                 goto retry;
 902                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 903                                 return ERR_PTR(ret);
 904                         BUG();
 905                 }
 906
 907                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
 908                              b->c.level != level ||
 909                              race_fault())) {
 910                         six_unlock_type(&b->c.lock, lock_type);
 911                         if (bch2_btree_node_relock(trans, path, level + 1))
 912                                 goto retry;
 913
 914                         trace_and_count(c, trans_restart_btree_node_reused, trans, trace_ip, path);
 915                         return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_lock_node_reused));
 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                         int ret = bch2_trans_relock(trans) ?:
 933                                 bch2_btree_path_relock_intent(trans, path);
 934                         if (ret) {
 935                                 BUG_ON(!trans->restarted);
 936                                 return ERR_PTR(ret);
 937                         }
 938                 }
 939
 940                 if (!six_relock_type(&b->c.lock, lock_type, seq))
 941                         goto retry;
 942         }
 943
 944         prefetch(b->aux_data);
 945
 946         for_each_bset(b, t) {
 947                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 948
 949                 prefetch(p + L1_CACHE_BYTES * 0);
 950                 prefetch(p + L1_CACHE_BYTES * 1);
 951                 prefetch(p + L1_CACHE_BYTES * 2);
 952         }
 953
 954         /* avoid atomic set bit if it's not needed: */
 955         if (!btree_node_accessed(b))
 956                 set_btree_node_accessed(b);
 957
 958         if (unlikely(btree_node_read_error(b))) {
 959                 six_unlock_type(&b->c.lock, lock_type);
 960                 return ERR_PTR(-EIO);
 961         }
 962
 963         EBUG_ON(b->c.btree_id != path->btree_id);
 964         EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
 965         btree_check_header(c, b);
 966
 967         return b;
 968 }
 969
 970 struct btree *bch2_btree_node_get_noiter(struct btree_trans *trans,
 971                                          const struct bkey_i *k,
 972                                          enum btree_id btree_id,
 973                                          unsigned level,
 974                                          bool nofill)
 975 {
 976         struct bch_fs *c = trans->c;
 977         struct btree_cache *bc = &c->btree_cache;
 978         struct btree *b;
 979         struct bset_tree *t;
 980         int ret;
 981
 982         EBUG_ON(level >= BTREE_MAX_DEPTH);
 983
 984         if (c->opts.btree_node_mem_ptr_optimization) {
 985                 b = btree_node_mem_ptr(k);
 986                 if (b)
 987                         goto lock_node;
 988         }
 989 retry:
 990         b = btree_cache_find(bc, k);
 991         if (unlikely(!b)) {
 992                 if (nofill)
 993                         goto out;
 994
 995                 b = bch2_btree_node_fill(c, NULL, NULL, k, btree_id,
 996                                          level, SIX_LOCK_read, true);
 997
 998                 /* We raced and found the btree node in the cache */
 999                 if (!b)
1000                         goto retry;
1001
1002                 if (IS_ERR(b) &&
1003                     !bch2_btree_cache_cannibalize_lock(c, NULL))
1004                         goto retry;
1005
1006                 if (IS_ERR(b))
1007                         goto out;
1008         } else {
1009 lock_node:
1010                 ret = btree_node_lock_nopath(trans, &b->c, SIX_LOCK_read);
1011                 if (unlikely(ret)) {
1012                         if (bch2_err_matches(ret, BCH_ERR_lock_fail_node_reused))
1013                                 goto retry;
1014                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1015                                 return ERR_PTR(ret);
1016                         BUG();
1017                 }
1018
1019                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
1020                              b->c.btree_id != btree_id ||
1021                              b->c.level != level)) {
1022                         six_unlock_read(&b->c.lock);
1023                         goto retry;
1024                 }
1025         }
1026
1027         /* XXX: waiting on IO with btree locks held: */
1028         __bch2_btree_node_wait_on_read(b);
1029
1030         prefetch(b->aux_data);
1031
1032         for_each_bset(b, t) {
1033                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
1034
1035                 prefetch(p + L1_CACHE_BYTES * 0);
1036                 prefetch(p + L1_CACHE_BYTES * 1);
1037                 prefetch(p + L1_CACHE_BYTES * 2);
1038         }
1039
1040         /* avoid atomic set bit if it's not needed: */
1041         if (!btree_node_accessed(b))
1042                 set_btree_node_accessed(b);
1043
1044         if (unlikely(btree_node_read_error(b))) {
1045                 six_unlock_read(&b->c.lock);
1046                 b = ERR_PTR(-EIO);
1047                 goto out;
1048         }
1049
1050         EBUG_ON(b->c.btree_id != btree_id);
1051         EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
1052         btree_check_header(c, b);
1053 out:
1054         bch2_btree_cache_cannibalize_unlock(c);
1055         return b;
1056 }
1057
1058 int bch2_btree_node_prefetch(struct bch_fs *c,
1059                              struct btree_trans *trans,
1060                              struct btree_path *path,
1061                              const struct bkey_i *k,
1062                              enum btree_id btree_id, unsigned level)
1063 {
1064         struct btree_cache *bc = &c->btree_cache;
1065         struct btree *b;
1066
1067         BUG_ON(trans && !btree_node_locked(path, level + 1));
1068         BUG_ON(level >= BTREE_MAX_DEPTH);
1069
1070         b = btree_cache_find(bc, k);
1071         if (b)
1072                 return 0;
1073
1074         b = bch2_btree_node_fill(c, trans, path, k, btree_id,
1075                                  level, SIX_LOCK_read, false);
1076         return PTR_ERR_OR_ZERO(b);
1077 }
1078
1079 void bch2_btree_node_evict(struct btree_trans *trans, const struct bkey_i *k)
1080 {
1081         struct bch_fs *c = trans->c;
1082         struct btree_cache *bc = &c->btree_cache;
1083         struct btree *b;
1084
1085         b = btree_cache_find(bc, k);
1086         if (!b)
1087                 return;
1088 wait_on_io:
1089         /* not allowed to wait on io with btree locks held: */
1090
1091         /* XXX we're called from btree_gc which will be holding other btree
1092          * nodes locked
1093          * */
1094         __bch2_btree_node_wait_on_read(b);
1095         __bch2_btree_node_wait_on_write(b);
1096
1097         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
1098         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
1099
1100         if (btree_node_dirty(b)) {
1101                 __bch2_btree_node_write(c, b, 0);
1102                 six_unlock_write(&b->c.lock);
1103                 six_unlock_intent(&b->c.lock);
1104                 goto wait_on_io;
1105         }
1106
1107         BUG_ON(btree_node_dirty(b));
1108
1109         mutex_lock(&bc->lock);
1110         btree_node_data_free(c, b);
1111         bch2_btree_node_hash_remove(bc, b);
1112         mutex_unlock(&bc->lock);
1113
1114         six_unlock_write(&b->c.lock);
1115         six_unlock_intent(&b->c.lock);
1116 }
1117
1118 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
1119                              struct btree *b)
1120 {
1121         const struct bkey_format *f = &b->format;
1122         struct bset_stats stats;
1123
1124         memset(&stats, 0, sizeof(stats));
1125
1126         bch2_btree_keys_stats(b, &stats);
1127
1128         prt_printf(out, "l %u ", b->c.level);
1129         bch2_bpos_to_text(out, b->data->min_key);
1130         prt_printf(out, " - ");
1131         bch2_bpos_to_text(out, b->data->max_key);
1132         prt_printf(out, ":\n"
1133                "    ptrs: ");
1134         bch2_val_to_text(out, c, bkey_i_to_s_c(&b->key));
1135
1136         prt_printf(out, "\n"
1137                "    format: u64s %u fields %u %u %u %u %u\n"
1138                "    unpack fn len: %u\n"
1139                "    bytes used %zu/%zu (%zu%% full)\n"
1140                "    sib u64s: %u, %u (merge threshold %u)\n"
1141                "    nr packed keys %u\n"
1142                "    nr unpacked keys %u\n"
1143                "    floats %zu\n"
1144                "    failed unpacked %zu\n",
1145                f->key_u64s,
1146                f->bits_per_field[0],
1147                f->bits_per_field[1],
1148                f->bits_per_field[2],
1149                f->bits_per_field[3],
1150                f->bits_per_field[4],
1151                b->unpack_fn_len,
1152                b->nr.live_u64s * sizeof(u64),
1153                btree_bytes(c) - sizeof(struct btree_node),
1154                b->nr.live_u64s * 100 / btree_max_u64s(c),
1155                b->sib_u64s[0],
1156                b->sib_u64s[1],
1157                c->btree_foreground_merge_threshold,
1158                b->nr.packed_keys,
1159                b->nr.unpacked_keys,
1160                stats.floats,
1161                stats.failed);
1162 }
1163
1164 void bch2_btree_cache_to_text(struct printbuf *out, struct bch_fs *c)
1165 {
1166         prt_printf(out, "nr nodes:\t\t%u\n", c->btree_cache.used);
1167         prt_printf(out, "nr dirty:\t\t%u\n", atomic_read(&c->btree_cache.dirty));
1168         prt_printf(out, "cannibalize lock:\t%p\n", c->btree_cache.alloc_lock);
1169 }