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 <linux/seq_buf.h>
  16 #include <trace/events/bcachefs.h>
  17
  18 #define BTREE_CACHE_NOT_FREED_INCREMENT(counter) \
  19 do {                                             \
  20         if (shrinker_counter)                    \
  21                 bc->not_freed_##counter++;       \
  22 } while (0)
  23
  24 const char * const bch2_btree_node_flags[] = {
  25 #define x(f)    #f,
  26         BTREE_FLAGS()
  27 #undef x
  28         NULL
  29 };
  30
  31 void bch2_recalc_btree_reserve(struct bch_fs *c)
  32 {
  33         unsigned i, reserve = 16;
  34
  35         if (!c->btree_roots[0].b)
  36                 reserve += 8;
  37
  38         for (i = 0; i < BTREE_ID_NR; i++)
  39                 if (c->btree_roots[i].b)
  40                         reserve += min_t(unsigned, 1,
  41                                          c->btree_roots[i].b->c.level) * 8;
  42
  43         c->btree_cache.reserve = reserve;
  44 }
  45
  46 static inline unsigned btree_cache_can_free(struct btree_cache *bc)
  47 {
  48         return max_t(int, 0, bc->used - bc->reserve);
  49 }
  50
  51 static void btree_node_to_freedlist(struct btree_cache *bc, struct btree *b)
  52 {
  53         if (b->c.lock.readers)
  54                 list_move(&b->list, &bc->freed_pcpu);
  55         else
  56                 list_move(&b->list, &bc->freed_nonpcpu);
  57 }
  58
  59 static void btree_node_data_free(struct bch_fs *c, struct btree *b)
  60 {
  61         struct btree_cache *bc = &c->btree_cache;
  62
  63         EBUG_ON(btree_node_write_in_flight(b));
  64
  65         kvpfree(b->data, btree_bytes(c));
  66         b->data = NULL;
  67 #ifdef __KERNEL__
  68         vfree(b->aux_data);
  69 #else
  70         munmap(b->aux_data, btree_aux_data_bytes(b));
  71 #endif
  72         b->aux_data = NULL;
  73
  74         bc->used--;
  75
  76         btree_node_to_freedlist(bc, b);
  77 }
  78
  79 static int bch2_btree_cache_cmp_fn(struct rhashtable_compare_arg *arg,
  80                                    const void *obj)
  81 {
  82         const struct btree *b = obj;
  83         const u64 *v = arg->key;
  84
  85         return b->hash_val == *v ? 0 : 1;
  86 }
  87
  88 static const struct rhashtable_params bch_btree_cache_params = {
  89         .head_offset    = offsetof(struct btree, hash),
  90         .key_offset     = offsetof(struct btree, hash_val),
  91         .key_len        = sizeof(u64),
  92         .obj_cmpfn      = bch2_btree_cache_cmp_fn,
  93 };
  94
  95 static int btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
  96 {
  97         BUG_ON(b->data || b->aux_data);
  98
  99         b->data = kvpmalloc(btree_bytes(c), gfp);
 100         if (!b->data)
 101                 return -ENOMEM;
 102 #ifdef __KERNEL__
 103         b->aux_data = vmalloc_exec(btree_aux_data_bytes(b), gfp);
 104 #else
 105         b->aux_data = mmap(NULL, btree_aux_data_bytes(b),
 106                            PROT_READ|PROT_WRITE|PROT_EXEC,
 107                            MAP_PRIVATE|MAP_ANONYMOUS, 0, 0);
 108         if (b->aux_data == MAP_FAILED)
 109                 b->aux_data = NULL;
 110 #endif
 111         if (!b->aux_data) {
 112                 kvpfree(b->data, btree_bytes(c));
 113                 b->data = NULL;
 114                 return -ENOMEM;
 115         }
 116
 117         return 0;
 118 }
 119
 120 static struct btree *__btree_node_mem_alloc(struct bch_fs *c, gfp_t gfp)
 121 {
 122         struct btree *b;
 123
 124         b = kzalloc(sizeof(struct btree), gfp);
 125         if (!b)
 126                 return NULL;
 127
 128         bkey_btree_ptr_init(&b->key);
 129         bch2_btree_lock_init(&b->c);
 130 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 131         lockdep_set_no_check_recursion(&b->c.lock.dep_map);
 132 #endif
 133         INIT_LIST_HEAD(&b->list);
 134         INIT_LIST_HEAD(&b->write_blocked);
 135         b->byte_order = ilog2(btree_bytes(c));
 136         return b;
 137 }
 138
 139 struct btree *__bch2_btree_node_mem_alloc(struct bch_fs *c)
 140 {
 141         struct btree_cache *bc = &c->btree_cache;
 142         struct btree *b;
 143
 144         b = __btree_node_mem_alloc(c, GFP_KERNEL);
 145         if (!b)
 146                 return NULL;
 147
 148         if (btree_node_data_alloc(c, b, GFP_KERNEL)) {
 149                 kfree(b);
 150                 return NULL;
 151         }
 152
 153         bc->used++;
 154         list_add(&b->list, &bc->freeable);
 155         return b;
 156 }
 157
 158 /* Btree in memory cache - hash table */
 159
 160 void bch2_btree_node_hash_remove(struct btree_cache *bc, struct btree *b)
 161 {
 162         int ret = rhashtable_remove_fast(&bc->table, &b->hash, bch_btree_cache_params);
 163
 164         BUG_ON(ret);
 165
 166         /* Cause future lookups for this node to fail: */
 167         b->hash_val = 0;
 168 }
 169
 170 int __bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b)
 171 {
 172         BUG_ON(b->hash_val);
 173         b->hash_val = btree_ptr_hash_val(&b->key);
 174
 175         return rhashtable_lookup_insert_fast(&bc->table, &b->hash,
 176                                              bch_btree_cache_params);
 177 }
 178
 179 int bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b,
 180                                 unsigned level, enum btree_id id)
 181 {
 182         int ret;
 183
 184         b->c.level      = level;
 185         b->c.btree_id   = id;
 186
 187         mutex_lock(&bc->lock);
 188         ret = __bch2_btree_node_hash_insert(bc, b);
 189         if (!ret)
 190                 list_add_tail(&b->list, &bc->live);
 191         mutex_unlock(&bc->lock);
 192
 193         return ret;
 194 }
 195
 196 __flatten
 197 static inline struct btree *btree_cache_find(struct btree_cache *bc,
 198                                      const struct bkey_i *k)
 199 {
 200         u64 v = btree_ptr_hash_val(k);
 201
 202         return rhashtable_lookup_fast(&bc->table, &v, bch_btree_cache_params);
 203 }
 204
 205 /*
 206  * this version is for btree nodes that have already been freed (we're not
 207  * reaping a real btree node)
 208  */
 209 static int __btree_node_reclaim(struct bch_fs *c, struct btree *b, bool flush, bool shrinker_counter)
 210 {
 211         struct btree_cache *bc = &c->btree_cache;
 212         int ret = 0;
 213
 214         lockdep_assert_held(&bc->lock);
 215 wait_on_io:
 216         if (b->flags & ((1U << BTREE_NODE_dirty)|
 217                         (1U << BTREE_NODE_read_in_flight)|
 218                         (1U << BTREE_NODE_write_in_flight))) {
 219                 if (!flush) {
 220                         if (btree_node_dirty(b))
 221                                 BTREE_CACHE_NOT_FREED_INCREMENT(dirty);
 222                         else if (btree_node_read_in_flight(b))
 223                                 BTREE_CACHE_NOT_FREED_INCREMENT(read_in_flight);
 224                         else if (btree_node_write_in_flight(b))
 225                                 BTREE_CACHE_NOT_FREED_INCREMENT(write_in_flight);
 226                         return -ENOMEM;
 227                 }
 228
 229                 /* XXX: waiting on IO with btree cache lock held */
 230                 bch2_btree_node_wait_on_read(b);
 231                 bch2_btree_node_wait_on_write(b);
 232         }
 233
 234         if (!six_trylock_intent(&b->c.lock)) {
 235                 BTREE_CACHE_NOT_FREED_INCREMENT(lock_intent);
 236                 return -ENOMEM;
 237         }
 238
 239         if (!six_trylock_write(&b->c.lock)) {
 240                 BTREE_CACHE_NOT_FREED_INCREMENT(lock_write);
 241                 goto out_unlock_intent;
 242         }
 243
 244         /* recheck under lock */
 245         if (b->flags & ((1U << BTREE_NODE_read_in_flight)|
 246                         (1U << BTREE_NODE_write_in_flight))) {
 247                 if (!flush) {
 248                         if (btree_node_read_in_flight(b))
 249                                 BTREE_CACHE_NOT_FREED_INCREMENT(read_in_flight);
 250                         else if (btree_node_write_in_flight(b))
 251                                 BTREE_CACHE_NOT_FREED_INCREMENT(write_in_flight);
 252                         goto out_unlock;
 253                 }
 254                 six_unlock_write(&b->c.lock);
 255                 six_unlock_intent(&b->c.lock);
 256                 goto wait_on_io;
 257         }
 258
 259         if (btree_node_noevict(b)) {
 260                 BTREE_CACHE_NOT_FREED_INCREMENT(noevict);
 261                 goto out_unlock;
 262         }
 263         if (btree_node_write_blocked(b)) {
 264                 BTREE_CACHE_NOT_FREED_INCREMENT(write_blocked);
 265                 goto out_unlock;
 266         }
 267         if (btree_node_will_make_reachable(b)) {
 268                 BTREE_CACHE_NOT_FREED_INCREMENT(will_make_reachable);
 269                 goto out_unlock;
 270         }
 271
 272         if (btree_node_dirty(b)) {
 273                 if (!flush) {
 274                         BTREE_CACHE_NOT_FREED_INCREMENT(dirty);
 275                         goto out_unlock;
 276                 }
 277                 /*
 278                  * Using the underscore version because we don't want to compact
 279                  * bsets after the write, since this node is about to be evicted
 280                  * - unless btree verify mode is enabled, since it runs out of
 281                  * the post write cleanup:
 282                  */
 283                 if (bch2_verify_btree_ondisk)
 284                         bch2_btree_node_write(c, b, SIX_LOCK_intent,
 285                                               BTREE_WRITE_cache_reclaim);
 286                 else
 287                         __bch2_btree_node_write(c, b,
 288                                                 BTREE_WRITE_cache_reclaim);
 289
 290                 six_unlock_write(&b->c.lock);
 291                 six_unlock_intent(&b->c.lock);
 292                 goto wait_on_io;
 293         }
 294 out:
 295         if (b->hash_val && !ret)
 296                 trace_and_count(c, btree_cache_reap, c, b);
 297         return ret;
 298 out_unlock:
 299         six_unlock_write(&b->c.lock);
 300 out_unlock_intent:
 301         six_unlock_intent(&b->c.lock);
 302         ret = -ENOMEM;
 303         goto out;
 304 }
 305
 306 static int btree_node_reclaim(struct bch_fs *c, struct btree *b, bool shrinker_counter)
 307 {
 308         return __btree_node_reclaim(c, b, false, shrinker_counter);
 309 }
 310
 311 static int btree_node_write_and_reclaim(struct bch_fs *c, struct btree *b)
 312 {
 313         return __btree_node_reclaim(c, b, true, false);
 314 }
 315
 316 static unsigned long bch2_btree_cache_scan(struct shrinker *shrink,
 317                                            struct shrink_control *sc)
 318 {
 319         struct bch_fs *c = container_of(shrink, struct bch_fs,
 320                                         btree_cache.shrink);
 321         struct btree_cache *bc = &c->btree_cache;
 322         struct btree *b, *t;
 323         unsigned long nr = sc->nr_to_scan;
 324         unsigned long can_free = 0;
 325         unsigned long freed = 0;
 326         unsigned long touched = 0;
 327         unsigned i, flags;
 328         unsigned long ret = SHRINK_STOP;
 329         bool trigger_writes = atomic_read(&bc->dirty) + nr >=
 330                 bc->used * 3 / 4;
 331
 332         if (bch2_btree_shrinker_disabled)
 333                 return SHRINK_STOP;
 334
 335         mutex_lock(&bc->lock);
 336         flags = memalloc_nofs_save();
 337
 338         /*
 339          * It's _really_ critical that we don't free too many btree nodes - we
 340          * have to always leave ourselves a reserve. The reserve is how we
 341          * guarantee that allocating memory for a new btree node can always
 342          * succeed, so that inserting keys into the btree can always succeed and
 343          * IO can always make forward progress:
 344          */
 345         can_free = btree_cache_can_free(bc);
 346         nr = min_t(unsigned long, nr, can_free);
 347
 348         i = 0;
 349         list_for_each_entry_safe(b, t, &bc->freeable, list) {
 350                 /*
 351                  * Leave a few nodes on the freeable list, so that a btree split
 352                  * won't have to hit the system allocator:
 353                  */
 354                 if (++i <= 3)
 355                         continue;
 356
 357                 touched++;
 358
 359                 if (touched >= nr)
 360                         goto out;
 361
 362                 if (!btree_node_reclaim(c, b, true)) {
 363                         btree_node_data_free(c, b);
 364                         six_unlock_write(&b->c.lock);
 365                         six_unlock_intent(&b->c.lock);
 366                         freed++;
 367                         bc->freed++;
 368                 }
 369         }
 370 restart:
 371         list_for_each_entry_safe(b, t, &bc->live, list) {
 372                 touched++;
 373
 374                 if (btree_node_accessed(b)) {
 375                         clear_btree_node_accessed(b);
 376                         bc->not_freed_access_bit++;
 377                 } else if (!btree_node_reclaim(c, b, true)) {
 378                         freed++;
 379                         btree_node_data_free(c, b);
 380                         bc->freed++;
 381
 382                         bch2_btree_node_hash_remove(bc, b);
 383                         six_unlock_write(&b->c.lock);
 384                         six_unlock_intent(&b->c.lock);
 385
 386                         if (freed == nr)
 387                                 goto out_rotate;
 388                 } else if (trigger_writes &&
 389                            btree_node_dirty(b) &&
 390                            !btree_node_will_make_reachable(b) &&
 391                            !btree_node_write_blocked(b) &&
 392                            six_trylock_read(&b->c.lock)) {
 393                         list_move(&bc->live, &b->list);
 394                         mutex_unlock(&bc->lock);
 395                         __bch2_btree_node_write(c, b, BTREE_WRITE_cache_reclaim);
 396                         six_unlock_read(&b->c.lock);
 397                         if (touched >= nr)
 398                                 goto out_nounlock;
 399                         mutex_lock(&bc->lock);
 400                         goto restart;
 401                 }
 402
 403                 if (touched >= nr)
 404                         break;
 405         }
 406 out_rotate:
 407         if (&t->list != &bc->live)
 408                 list_move_tail(&bc->live, &t->list);
 409 out:
 410         mutex_unlock(&bc->lock);
 411 out_nounlock:
 412         ret = freed;
 413         memalloc_nofs_restore(flags);
 414         trace_and_count(c, btree_cache_scan, sc->nr_to_scan, can_free, ret);
 415         return ret;
 416 }
 417
 418 static unsigned long bch2_btree_cache_count(struct shrinker *shrink,
 419                                             struct shrink_control *sc)
 420 {
 421         struct bch_fs *c = container_of(shrink, struct bch_fs,
 422                                         btree_cache.shrink);
 423         struct btree_cache *bc = &c->btree_cache;
 424
 425         if (bch2_btree_shrinker_disabled)
 426                 return 0;
 427
 428         return btree_cache_can_free(bc);
 429 }
 430
 431 static void bch2_btree_cache_shrinker_to_text(struct seq_buf *s, struct shrinker *shrink)
 432 {
 433         struct bch_fs *c = container_of(shrink, struct bch_fs,
 434                                         btree_cache.shrink);
 435         char *cbuf;
 436         size_t buflen = seq_buf_get_buf(s, &cbuf);
 437         struct printbuf out = PRINTBUF_EXTERN(cbuf, buflen);
 438
 439         bch2_btree_cache_to_text(&out, &c->btree_cache);
 440         seq_buf_commit(s, out.pos);
 441 }
 442
 443 void bch2_fs_btree_cache_exit(struct bch_fs *c)
 444 {
 445         struct btree_cache *bc = &c->btree_cache;
 446         struct btree *b;
 447         unsigned i, flags;
 448
 449         if (bc->shrink.list.next)
 450                 unregister_shrinker(&bc->shrink);
 451
 452         /* vfree() can allocate memory: */
 453         flags = memalloc_nofs_save();
 454         mutex_lock(&bc->lock);
 455
 456         if (c->verify_data)
 457                 list_move(&c->verify_data->list, &bc->live);
 458
 459         kvpfree(c->verify_ondisk, btree_bytes(c));
 460
 461         for (i = 0; i < BTREE_ID_NR; i++)
 462                 if (c->btree_roots[i].b)
 463                         list_add(&c->btree_roots[i].b->list, &bc->live);
 464
 465         list_splice(&bc->freeable, &bc->live);
 466
 467         while (!list_empty(&bc->live)) {
 468                 b = list_first_entry(&bc->live, struct btree, list);
 469
 470                 BUG_ON(btree_node_read_in_flight(b) ||
 471                        btree_node_write_in_flight(b));
 472
 473                 if (btree_node_dirty(b))
 474                         bch2_btree_complete_write(c, b, btree_current_write(b));
 475                 clear_btree_node_dirty_acct(c, b);
 476
 477                 btree_node_data_free(c, b);
 478         }
 479
 480         BUG_ON(atomic_read(&c->btree_cache.dirty));
 481
 482         list_splice(&bc->freed_pcpu, &bc->freed_nonpcpu);
 483
 484         while (!list_empty(&bc->freed_nonpcpu)) {
 485                 b = list_first_entry(&bc->freed_nonpcpu, struct btree, list);
 486                 list_del(&b->list);
 487                 six_lock_pcpu_free(&b->c.lock);
 488                 kfree(b);
 489         }
 490
 491         mutex_unlock(&bc->lock);
 492         memalloc_nofs_restore(flags);
 493
 494         if (bc->table_init_done)
 495                 rhashtable_destroy(&bc->table);
 496 }
 497
 498 int bch2_fs_btree_cache_init(struct bch_fs *c)
 499 {
 500         struct btree_cache *bc = &c->btree_cache;
 501         unsigned i;
 502         int ret = 0;
 503
 504         pr_verbose_init(c->opts, "");
 505
 506         ret = rhashtable_init(&bc->table, &bch_btree_cache_params);
 507         if (ret)
 508                 goto out;
 509
 510         bc->table_init_done = true;
 511
 512         bch2_recalc_btree_reserve(c);
 513
 514         for (i = 0; i < bc->reserve; i++)
 515                 if (!__bch2_btree_node_mem_alloc(c)) {
 516                         ret = -ENOMEM;
 517                         goto out;
 518                 }
 519
 520         list_splice_init(&bc->live, &bc->freeable);
 521
 522         mutex_init(&c->verify_lock);
 523
 524         bc->shrink.count_objects        = bch2_btree_cache_count;
 525         bc->shrink.scan_objects         = bch2_btree_cache_scan;
 526         bc->shrink.to_text              = bch2_btree_cache_shrinker_to_text;
 527         bc->shrink.seeks                = 4;
 528         ret = register_shrinker(&bc->shrink, "%s/btree_cache", c->name);
 529 out:
 530         pr_verbose_init(c->opts, "ret %i", ret);
 531         return ret;
 532 }
 533
 534 void bch2_fs_btree_cache_init_early(struct btree_cache *bc)
 535 {
 536         mutex_init(&bc->lock);
 537         INIT_LIST_HEAD(&bc->live);
 538         INIT_LIST_HEAD(&bc->freeable);
 539         INIT_LIST_HEAD(&bc->freed_pcpu);
 540         INIT_LIST_HEAD(&bc->freed_nonpcpu);
 541 }
 542
 543 /*
 544  * We can only have one thread cannibalizing other cached btree nodes at a time,
 545  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 546  * cannibalize_bucket() will take. This means every time we unlock the root of
 547  * the btree, we need to release this lock if we have it held.
 548  */
 549 void bch2_btree_cache_cannibalize_unlock(struct bch_fs *c)
 550 {
 551         struct btree_cache *bc = &c->btree_cache;
 552
 553         if (bc->alloc_lock == current) {
 554                 trace_and_count(c, btree_cache_cannibalize_unlock, c);
 555                 bc->alloc_lock = NULL;
 556                 closure_wake_up(&bc->alloc_wait);
 557         }
 558 }
 559
 560 int bch2_btree_cache_cannibalize_lock(struct bch_fs *c, struct closure *cl)
 561 {
 562         struct btree_cache *bc = &c->btree_cache;
 563         struct task_struct *old;
 564
 565         old = cmpxchg(&bc->alloc_lock, NULL, current);
 566         if (old == NULL || old == current)
 567                 goto success;
 568
 569         if (!cl) {
 570                 trace_and_count(c, btree_cache_cannibalize_lock_fail, c);
 571                 return -ENOMEM;
 572         }
 573
 574         closure_wait(&bc->alloc_wait, cl);
 575
 576         /* Try again, after adding ourselves to waitlist */
 577         old = cmpxchg(&bc->alloc_lock, NULL, current);
 578         if (old == NULL || old == current) {
 579                 /* We raced */
 580                 closure_wake_up(&bc->alloc_wait);
 581                 goto success;
 582         }
 583
 584         trace_and_count(c, btree_cache_cannibalize_lock_fail, c);
 585         return -BCH_ERR_btree_cache_cannibalize_lock_blocked;
 586
 587 success:
 588         trace_and_count(c, btree_cache_cannibalize_lock, c);
 589         return 0;
 590 }
 591
 592 static struct btree *btree_node_cannibalize(struct bch_fs *c)
 593 {
 594         struct btree_cache *bc = &c->btree_cache;
 595         struct btree *b;
 596
 597         list_for_each_entry_reverse(b, &bc->live, list)
 598                 if (!btree_node_reclaim(c, b, false))
 599                         return b;
 600
 601         while (1) {
 602                 list_for_each_entry_reverse(b, &bc->live, list)
 603                         if (!btree_node_write_and_reclaim(c, b))
 604                                 return b;
 605
 606                 /*
 607                  * Rare case: all nodes were intent-locked.
 608                  * Just busy-wait.
 609                  */
 610                 WARN_ONCE(1, "btree cache cannibalize failed\n");
 611                 cond_resched();
 612         }
 613 }
 614
 615 struct btree *bch2_btree_node_mem_alloc(struct btree_trans *trans, bool pcpu_read_locks)
 616 {
 617         struct bch_fs *c = trans->c;
 618         struct btree_cache *bc = &c->btree_cache;
 619         struct list_head *freed = pcpu_read_locks
 620                 ? &bc->freed_pcpu
 621                 : &bc->freed_nonpcpu;
 622         struct btree *b, *b2;
 623         u64 start_time = local_clock();
 624         unsigned flags;
 625
 626         flags = memalloc_nofs_save();
 627         mutex_lock(&bc->lock);
 628
 629         /*
 630          * We never free struct btree itself, just the memory that holds the on
 631          * disk node. Check the freed list before allocating a new one:
 632          */
 633         list_for_each_entry(b, freed, list)
 634                 if (!btree_node_reclaim(c, b, false)) {
 635                         list_del_init(&b->list);
 636                         goto got_node;
 637                 }
 638
 639         b = __btree_node_mem_alloc(c, __GFP_NOWARN);
 640         if (!b) {
 641                 mutex_unlock(&bc->lock);
 642                 b = __btree_node_mem_alloc(c, GFP_KERNEL);
 643                 if (!b)
 644                         goto err;
 645                 mutex_lock(&bc->lock);
 646         }
 647
 648         if (pcpu_read_locks)
 649                 six_lock_pcpu_alloc(&b->c.lock);
 650
 651         BUG_ON(!six_trylock_intent(&b->c.lock));
 652         BUG_ON(!six_trylock_write(&b->c.lock));
 653 got_node:
 654
 655         /*
 656          * btree_free() doesn't free memory; it sticks the node on the end of
 657          * the list. Check if there's any freed nodes there:
 658          */
 659         list_for_each_entry(b2, &bc->freeable, list)
 660                 if (!btree_node_reclaim(c, b2, false)) {
 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                         goto got_mem;
 667                 }
 668
 669         mutex_unlock(&bc->lock);
 670
 671         if (btree_node_data_alloc(c, b, __GFP_NOWARN|GFP_KERNEL))
 672                 goto err;
 673
 674         mutex_lock(&bc->lock);
 675         bc->used++;
 676 got_mem:
 677         mutex_unlock(&bc->lock);
 678
 679         BUG_ON(btree_node_hashed(b));
 680         BUG_ON(btree_node_dirty(b));
 681         BUG_ON(btree_node_write_in_flight(b));
 682 out:
 683         b->flags                = 0;
 684         b->written              = 0;
 685         b->nsets                = 0;
 686         b->sib_u64s[0]          = 0;
 687         b->sib_u64s[1]          = 0;
 688         b->whiteout_u64s        = 0;
 689         bch2_btree_keys_init(b);
 690         set_btree_node_accessed(b);
 691
 692         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],
 693                                start_time);
 694
 695         memalloc_nofs_restore(flags);
 696         return b;
 697 err:
 698         mutex_lock(&bc->lock);
 699
 700         /* Try to cannibalize another cached btree node: */
 701         if (bc->alloc_lock == current) {
 702                 b2 = btree_node_cannibalize(c);
 703                 bch2_btree_node_hash_remove(bc, b2);
 704
 705                 if (b) {
 706                         swap(b->data, b2->data);
 707                         swap(b->aux_data, b2->aux_data);
 708                         btree_node_to_freedlist(bc, b2);
 709                         six_unlock_write(&b2->c.lock);
 710                         six_unlock_intent(&b2->c.lock);
 711                 } else {
 712                         b = b2;
 713                         list_del_init(&b->list);
 714                 }
 715
 716                 mutex_unlock(&bc->lock);
 717
 718                 trace_and_count(c, btree_cache_cannibalize, c);
 719                 goto out;
 720         }
 721
 722         mutex_unlock(&bc->lock);
 723         memalloc_nofs_restore(flags);
 724         return ERR_PTR(-ENOMEM);
 725 }
 726
 727 /* Slowpath, don't want it inlined into btree_iter_traverse() */
 728 static noinline struct btree *bch2_btree_node_fill(struct btree_trans *trans,
 729                                 struct btree_path *path,
 730                                 const struct bkey_i *k,
 731                                 enum btree_id btree_id,
 732                                 unsigned level,
 733                                 enum six_lock_type lock_type,
 734                                 bool sync)
 735 {
 736         struct bch_fs *c = trans->c;
 737         struct btree_cache *bc = &c->btree_cache;
 738         struct btree *b;
 739         u32 seq;
 740
 741         BUG_ON(level + 1 >= BTREE_MAX_DEPTH);
 742         /*
 743          * Parent node must be locked, else we could read in a btree node that's
 744          * been freed:
 745          */
 746         if (path && !bch2_btree_node_relock(trans, path, level + 1)) {
 747                 trace_and_count(c, trans_restart_relock_parent_for_fill, trans, _THIS_IP_, path);
 748                 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_fill_relock));
 749         }
 750
 751         b = bch2_btree_node_mem_alloc(trans, level != 0);
 752
 753         if (b == ERR_PTR(-ENOMEM)) {
 754                 trans->memory_allocation_failure = true;
 755                 trace_and_count(c, trans_restart_memory_allocation_failure, trans, _THIS_IP_, path);
 756                 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_fill_mem_alloc_fail));
 757         }
 758
 759         if (IS_ERR(b))
 760                 return b;
 761
 762         /*
 763          * Btree nodes read in from disk should not have the accessed bit set
 764          * initially, so that linear scans don't thrash the cache:
 765          */
 766         clear_btree_node_accessed(b);
 767
 768         bkey_copy(&b->key, k);
 769         if (bch2_btree_node_hash_insert(bc, b, level, btree_id)) {
 770                 /* raced with another fill: */
 771
 772                 /* mark as unhashed... */
 773                 b->hash_val = 0;
 774
 775                 mutex_lock(&bc->lock);
 776                 list_add(&b->list, &bc->freeable);
 777                 mutex_unlock(&bc->lock);
 778
 779                 six_unlock_write(&b->c.lock);
 780                 six_unlock_intent(&b->c.lock);
 781                 return NULL;
 782         }
 783
 784         set_btree_node_read_in_flight(b);
 785
 786         six_unlock_write(&b->c.lock);
 787         seq = b->c.lock.state.seq;
 788         six_unlock_intent(&b->c.lock);
 789
 790         /* Unlock before doing IO: */
 791         if (trans && sync)
 792                 bch2_trans_unlock(trans);
 793
 794         bch2_btree_node_read(c, b, sync);
 795
 796         if (!sync)
 797                 return NULL;
 798
 799         if (path) {
 800                 int ret = bch2_trans_relock(trans) ?:
 801                         bch2_btree_path_relock_intent(trans, path);
 802                 if (ret) {
 803                         BUG_ON(!trans->restarted);
 804                         return ERR_PTR(ret);
 805                 }
 806         }
 807
 808         if (!six_relock_type(&b->c.lock, lock_type, seq)) {
 809                 if (path)
 810                         trace_and_count(c, trans_restart_relock_after_fill, trans, _THIS_IP_, path);
 811                 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_relock_after_fill));
 812         }
 813
 814         return b;
 815 }
 816
 817 static noinline void btree_bad_header(struct bch_fs *c, struct btree *b)
 818 {
 819         struct printbuf buf = PRINTBUF;
 820
 821         if (!test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags))
 822                 return;
 823
 824         prt_printf(&buf,
 825                "btree node header doesn't match ptr\n"
 826                "btree %s level %u\n"
 827                "ptr: ",
 828                bch2_btree_ids[b->c.btree_id], b->c.level);
 829         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
 830
 831         prt_printf(&buf, "\nheader: btree %s level %llu\n"
 832                "min ",
 833                bch2_btree_ids[BTREE_NODE_ID(b->data)],
 834                BTREE_NODE_LEVEL(b->data));
 835         bch2_bpos_to_text(&buf, b->data->min_key);
 836
 837         prt_printf(&buf, "\nmax ");
 838         bch2_bpos_to_text(&buf, b->data->max_key);
 839
 840         bch2_fs_inconsistent(c, "%s", buf.buf);
 841         printbuf_exit(&buf);
 842 }
 843
 844 static inline void btree_check_header(struct bch_fs *c, struct btree *b)
 845 {
 846         if (b->c.btree_id != BTREE_NODE_ID(b->data) ||
 847             b->c.level != BTREE_NODE_LEVEL(b->data) ||
 848             !bpos_eq(b->data->max_key, b->key.k.p) ||
 849             (b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 850              !bpos_eq(b->data->min_key,
 851                       bkey_i_to_btree_ptr_v2(&b->key)->v.min_key)))
 852                 btree_bad_header(c, b);
 853 }
 854
 855 static struct btree *__bch2_btree_node_get(struct btree_trans *trans, struct btree_path *path,
 856                                            const struct bkey_i *k, unsigned level,
 857                                            enum six_lock_type lock_type,
 858                                            unsigned long trace_ip)
 859 {
 860         struct bch_fs *c = trans->c;
 861         struct btree_cache *bc = &c->btree_cache;
 862         struct btree *b;
 863         struct bset_tree *t;
 864         int ret;
 865
 866         EBUG_ON(level >= BTREE_MAX_DEPTH);
 867 retry:
 868         b = btree_cache_find(bc, k);
 869         if (unlikely(!b)) {
 870                 /*
 871                  * We must have the parent locked to call bch2_btree_node_fill(),
 872                  * else we could read in a btree node from disk that's been
 873                  * freed:
 874                  */
 875                 b = bch2_btree_node_fill(trans, path, k, path->btree_id,
 876                                          level, lock_type, true);
 877
 878                 /* We raced and found the btree node in the cache */
 879                 if (!b)
 880                         goto retry;
 881
 882                 if (IS_ERR(b))
 883                         return b;
 884         } else {
 885                 if (btree_node_read_locked(path, level + 1))
 886                         btree_node_unlock(trans, path, level + 1);
 887
 888                 ret = btree_node_lock(trans, path, &b->c, level, lock_type, trace_ip);
 889                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 890                         return ERR_PTR(ret);
 891
 892                 BUG_ON(ret);
 893
 894                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
 895                              b->c.level != level ||
 896                              race_fault())) {
 897                         six_unlock_type(&b->c.lock, lock_type);
 898                         if (bch2_btree_node_relock(trans, path, level + 1))
 899                                 goto retry;
 900
 901                         trace_and_count(c, trans_restart_btree_node_reused, trans, trace_ip, path);
 902                         return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_lock_node_reused));
 903                 }
 904
 905                 /* avoid atomic set bit if it's not needed: */
 906                 if (!btree_node_accessed(b))
 907                         set_btree_node_accessed(b);
 908         }
 909
 910         if (unlikely(btree_node_read_in_flight(b))) {
 911                 u32 seq = b->c.lock.state.seq;
 912
 913                 six_unlock_type(&b->c.lock, lock_type);
 914                 bch2_trans_unlock(trans);
 915
 916                 bch2_btree_node_wait_on_read(b);
 917
 918                 /*
 919                  * should_be_locked is not set on this path yet, so we need to
 920                  * relock it specifically:
 921                  */
 922                 if (trans) {
 923                         int ret = bch2_trans_relock(trans) ?:
 924                                 bch2_btree_path_relock_intent(trans, path);
 925                         if (ret) {
 926                                 BUG_ON(!trans->restarted);
 927                                 return ERR_PTR(ret);
 928                         }
 929                 }
 930
 931                 if (!six_relock_type(&b->c.lock, lock_type, seq))
 932                         goto retry;
 933         }
 934
 935         prefetch(b->aux_data);
 936
 937         for_each_bset(b, t) {
 938                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 939
 940                 prefetch(p + L1_CACHE_BYTES * 0);
 941                 prefetch(p + L1_CACHE_BYTES * 1);
 942                 prefetch(p + L1_CACHE_BYTES * 2);
 943         }
 944
 945         if (unlikely(btree_node_read_error(b))) {
 946                 six_unlock_type(&b->c.lock, lock_type);
 947                 return ERR_PTR(-EIO);
 948         }
 949
 950         EBUG_ON(b->c.btree_id != path->btree_id);
 951         EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
 952         btree_check_header(c, b);
 953
 954         return b;
 955 }
 956
 957 /**
 958  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 959  * in from disk if necessary.
 960  *
 961  * The btree node will have either a read or a write lock held, depending on
 962  * the @write parameter.
 963  */
 964 struct btree *bch2_btree_node_get(struct btree_trans *trans, struct btree_path *path,
 965                                   const struct bkey_i *k, unsigned level,
 966                                   enum six_lock_type lock_type,
 967                                   unsigned long trace_ip)
 968 {
 969         struct bch_fs *c = trans->c;
 970         struct btree *b;
 971         struct bset_tree *t;
 972         int ret;
 973
 974         EBUG_ON(level >= BTREE_MAX_DEPTH);
 975
 976         b = btree_node_mem_ptr(k);
 977
 978         /*
 979          * Check b->hash_val _before_ calling btree_node_lock() - this might not
 980          * be the node we want anymore, and trying to lock the wrong node could
 981          * cause an unneccessary transaction restart:
 982          */
 983         if (unlikely(!c->opts.btree_node_mem_ptr_optimization ||
 984                      !b ||
 985                      b->hash_val != btree_ptr_hash_val(k)))
 986                 return __bch2_btree_node_get(trans, path, k, level, lock_type, trace_ip);
 987
 988         if (btree_node_read_locked(path, level + 1))
 989                 btree_node_unlock(trans, path, level + 1);
 990
 991         ret = btree_node_lock(trans, path, &b->c, level, lock_type, trace_ip);
 992         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 993                 return ERR_PTR(ret);
 994
 995         BUG_ON(ret);
 996
 997         if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
 998                      b->c.level != level ||
 999                      race_fault())) {
1000                 six_unlock_type(&b->c.lock, lock_type);
1001                 if (bch2_btree_node_relock(trans, path, level + 1))
1002                         return __bch2_btree_node_get(trans, path, k, level, lock_type, trace_ip);
1003
1004                 trace_and_count(c, trans_restart_btree_node_reused, trans, trace_ip, path);
1005                 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_lock_node_reused));
1006         }
1007
1008         if (unlikely(btree_node_read_in_flight(b))) {
1009                 u32 seq = b->c.lock.state.seq;
1010
1011                 six_unlock_type(&b->c.lock, lock_type);
1012                 bch2_trans_unlock(trans);
1013
1014                 bch2_btree_node_wait_on_read(b);
1015
1016                 /*
1017                  * should_be_locked is not set on this path yet, so we need to
1018                  * relock it specifically:
1019                  */
1020                 if (trans) {
1021                         int ret = bch2_trans_relock(trans) ?:
1022                                 bch2_btree_path_relock_intent(trans, path);
1023                         if (ret) {
1024                                 BUG_ON(!trans->restarted);
1025                                 return ERR_PTR(ret);
1026                         }
1027                 }
1028
1029                 if (!six_relock_type(&b->c.lock, lock_type, seq))
1030                         return __bch2_btree_node_get(trans, path, k, level, lock_type, trace_ip);
1031         }
1032
1033         prefetch(b->aux_data);
1034
1035         for_each_bset(b, t) {
1036                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
1037
1038                 prefetch(p + L1_CACHE_BYTES * 0);
1039                 prefetch(p + L1_CACHE_BYTES * 1);
1040                 prefetch(p + L1_CACHE_BYTES * 2);
1041         }
1042
1043         /* avoid atomic set bit if it's not needed: */
1044         if (!btree_node_accessed(b))
1045                 set_btree_node_accessed(b);
1046
1047         if (unlikely(btree_node_read_error(b))) {
1048                 six_unlock_type(&b->c.lock, lock_type);
1049                 return ERR_PTR(-EIO);
1050         }
1051
1052         EBUG_ON(b->c.btree_id != path->btree_id);
1053         EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
1054         btree_check_header(c, b);
1055
1056         return b;
1057 }
1058
1059 struct btree *bch2_btree_node_get_noiter(struct btree_trans *trans,
1060                                          const struct bkey_i *k,
1061                                          enum btree_id btree_id,
1062                                          unsigned level,
1063                                          bool nofill)
1064 {
1065         struct bch_fs *c = trans->c;
1066         struct btree_cache *bc = &c->btree_cache;
1067         struct btree *b;
1068         struct bset_tree *t;
1069         int ret;
1070
1071         EBUG_ON(level >= BTREE_MAX_DEPTH);
1072
1073         if (c->opts.btree_node_mem_ptr_optimization) {
1074                 b = btree_node_mem_ptr(k);
1075                 if (b)
1076                         goto lock_node;
1077         }
1078 retry:
1079         b = btree_cache_find(bc, k);
1080         if (unlikely(!b)) {
1081                 if (nofill)
1082                         goto out;
1083
1084                 b = bch2_btree_node_fill(trans, NULL, k, btree_id,
1085                                          level, SIX_LOCK_read, true);
1086
1087                 /* We raced and found the btree node in the cache */
1088                 if (!b)
1089                         goto retry;
1090
1091                 if (IS_ERR(b) &&
1092                     !bch2_btree_cache_cannibalize_lock(c, NULL))
1093                         goto retry;
1094
1095                 if (IS_ERR(b))
1096                         goto out;
1097         } else {
1098 lock_node:
1099                 ret = btree_node_lock_nopath(trans, &b->c, SIX_LOCK_read, _THIS_IP_);
1100                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1101                         return ERR_PTR(ret);
1102
1103                 BUG_ON(ret);
1104
1105                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
1106                              b->c.btree_id != btree_id ||
1107                              b->c.level != level)) {
1108                         six_unlock_read(&b->c.lock);
1109                         goto retry;
1110                 }
1111         }
1112
1113         /* XXX: waiting on IO with btree locks held: */
1114         __bch2_btree_node_wait_on_read(b);
1115
1116         prefetch(b->aux_data);
1117
1118         for_each_bset(b, t) {
1119                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
1120
1121                 prefetch(p + L1_CACHE_BYTES * 0);
1122                 prefetch(p + L1_CACHE_BYTES * 1);
1123                 prefetch(p + L1_CACHE_BYTES * 2);
1124         }
1125
1126         /* avoid atomic set bit if it's not needed: */
1127         if (!btree_node_accessed(b))
1128                 set_btree_node_accessed(b);
1129
1130         if (unlikely(btree_node_read_error(b))) {
1131                 six_unlock_read(&b->c.lock);
1132                 b = ERR_PTR(-EIO);
1133                 goto out;
1134         }
1135
1136         EBUG_ON(b->c.btree_id != btree_id);
1137         EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
1138         btree_check_header(c, b);
1139 out:
1140         bch2_btree_cache_cannibalize_unlock(c);
1141         return b;
1142 }
1143
1144 int bch2_btree_node_prefetch(struct btree_trans *trans,
1145                              struct btree_path *path,
1146                              const struct bkey_i *k,
1147                              enum btree_id btree_id, unsigned level)
1148 {
1149         struct bch_fs *c = trans->c;
1150         struct btree_cache *bc = &c->btree_cache;
1151         struct btree *b;
1152
1153         BUG_ON(trans && !btree_node_locked(path, level + 1));
1154         BUG_ON(level >= BTREE_MAX_DEPTH);
1155
1156         b = btree_cache_find(bc, k);
1157         if (b)
1158                 return 0;
1159
1160         b = bch2_btree_node_fill(trans, path, k, btree_id,
1161                                  level, SIX_LOCK_read, false);
1162         return PTR_ERR_OR_ZERO(b);
1163 }
1164
1165 void bch2_btree_node_evict(struct btree_trans *trans, const struct bkey_i *k)
1166 {
1167         struct bch_fs *c = trans->c;
1168         struct btree_cache *bc = &c->btree_cache;
1169         struct btree *b;
1170
1171         b = btree_cache_find(bc, k);
1172         if (!b)
1173                 return;
1174 wait_on_io:
1175         /* not allowed to wait on io with btree locks held: */
1176
1177         /* XXX we're called from btree_gc which will be holding other btree
1178          * nodes locked
1179          */
1180         __bch2_btree_node_wait_on_read(b);
1181         __bch2_btree_node_wait_on_write(b);
1182
1183         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
1184         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
1185
1186         if (btree_node_dirty(b)) {
1187                 __bch2_btree_node_write(c, b, BTREE_WRITE_cache_reclaim);
1188                 six_unlock_write(&b->c.lock);
1189                 six_unlock_intent(&b->c.lock);
1190                 goto wait_on_io;
1191         }
1192
1193         BUG_ON(btree_node_dirty(b));
1194
1195         mutex_lock(&bc->lock);
1196         btree_node_data_free(c, b);
1197         bch2_btree_node_hash_remove(bc, b);
1198         mutex_unlock(&bc->lock);
1199
1200         six_unlock_write(&b->c.lock);
1201         six_unlock_intent(&b->c.lock);
1202 }
1203
1204 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
1205                              const struct btree *b)
1206 {
1207         const struct bkey_format *f = &b->format;
1208         struct bset_stats stats;
1209
1210         memset(&stats, 0, sizeof(stats));
1211
1212         bch2_btree_keys_stats(b, &stats);
1213
1214         prt_printf(out, "l %u ", b->c.level);
1215         bch2_bpos_to_text(out, b->data->min_key);
1216         prt_printf(out, " - ");
1217         bch2_bpos_to_text(out, b->data->max_key);
1218         prt_printf(out, ":\n"
1219                "    ptrs: ");
1220         bch2_val_to_text(out, c, bkey_i_to_s_c(&b->key));
1221
1222         prt_printf(out, "\n"
1223                "    format: u64s %u fields %u %u %u %u %u\n"
1224                "    unpack fn len: %u\n"
1225                "    bytes used %zu/%zu (%zu%% full)\n"
1226                "    sib u64s: %u, %u (merge threshold %u)\n"
1227                "    nr packed keys %u\n"
1228                "    nr unpacked keys %u\n"
1229                "    floats %zu\n"
1230                "    failed unpacked %zu\n",
1231                f->key_u64s,
1232                f->bits_per_field[0],
1233                f->bits_per_field[1],
1234                f->bits_per_field[2],
1235                f->bits_per_field[3],
1236                f->bits_per_field[4],
1237                b->unpack_fn_len,
1238                b->nr.live_u64s * sizeof(u64),
1239                btree_bytes(c) - sizeof(struct btree_node),
1240                b->nr.live_u64s * 100 / btree_max_u64s(c),
1241                b->sib_u64s[0],
1242                b->sib_u64s[1],
1243                c->btree_foreground_merge_threshold,
1244                b->nr.packed_keys,
1245                b->nr.unpacked_keys,
1246                stats.floats,
1247                stats.failed);
1248 }
1249
1250 void bch2_btree_cache_to_text(struct printbuf *out, const struct btree_cache *bc)
1251 {
1252         prt_printf(out, "nr nodes:\t\t%u\n", bc->used);
1253         prt_printf(out, "nr dirty:\t\t%u\n", atomic_read(&bc->dirty));
1254         prt_printf(out, "cannibalize lock:\t%p\n", bc->alloc_lock);
1255
1256         prt_printf(out, "freed:\t\t\t\t%u\n", bc->freed);
1257         prt_printf(out, "not freed, dirty:\t\t%u\n", bc->not_freed_dirty);
1258         prt_printf(out, "not freed, write in flight:\t%u\n", bc->not_freed_write_in_flight);
1259         prt_printf(out, "not freed, read in flight:\t%u\n", bc->not_freed_read_in_flight);
1260         prt_printf(out, "not freed, lock intent failed:\t%u\n", bc->not_freed_lock_intent);
1261         prt_printf(out, "not freed, lock write failed:\t%u\n", bc->not_freed_lock_write);
1262         prt_printf(out, "not freed, access bit:\t\t%u\n", bc->not_freed_access_bit);
1263         prt_printf(out, "not freed, no evict failed:\t%u\n", bc->not_freed_noevict);
1264         prt_printf(out, "not freed, write blocked:\t%u\n", bc->not_freed_write_blocked);
1265         prt_printf(out, "not freed, will make reachable:\t%u\n", bc->not_freed_will_make_reachable);
1266
1267 }