1 // SPDX-License-Identifier: GPL-2.0
5 #include "btree_cache.h"
7 #include "btree_iter.h"
8 #include "btree_locking.h"
14 #include <linux/prefetch.h>
15 #include <linux/sched/mm.h>
16 #include <linux/seq_buf.h>
18 #define BTREE_CACHE_NOT_FREED_INCREMENT(counter) \
20 if (shrinker_counter) \
21 bc->not_freed_##counter++; \
24 const char * const bch2_btree_node_flags[] = {
31 void bch2_recalc_btree_reserve(struct bch_fs *c)
33 unsigned i, reserve = 16;
35 if (!c->btree_roots_known[0].b)
38 for (i = 0; i < btree_id_nr_alive(c); i++) {
39 struct btree_root *r = bch2_btree_id_root(c, i);
42 reserve += min_t(unsigned, 1, r->b->c.level) * 8;
45 c->btree_cache.reserve = reserve;
48 static inline unsigned btree_cache_can_free(struct btree_cache *bc)
50 return max_t(int, 0, bc->used - bc->reserve);
53 static void btree_node_to_freedlist(struct btree_cache *bc, struct btree *b)
55 if (b->c.lock.readers)
56 list_move(&b->list, &bc->freed_pcpu);
58 list_move(&b->list, &bc->freed_nonpcpu);
61 static void btree_node_data_free(struct bch_fs *c, struct btree *b)
63 struct btree_cache *bc = &c->btree_cache;
65 EBUG_ON(btree_node_write_in_flight(b));
67 clear_btree_node_just_written(b);
69 kvpfree(b->data, btree_bytes(c));
74 munmap(b->aux_data, btree_aux_data_bytes(b));
80 btree_node_to_freedlist(bc, b);
83 static int bch2_btree_cache_cmp_fn(struct rhashtable_compare_arg *arg,
86 const struct btree *b = obj;
87 const u64 *v = arg->key;
89 return b->hash_val == *v ? 0 : 1;
92 static const struct rhashtable_params bch_btree_cache_params = {
93 .head_offset = offsetof(struct btree, hash),
94 .key_offset = offsetof(struct btree, hash_val),
95 .key_len = sizeof(u64),
96 .obj_cmpfn = bch2_btree_cache_cmp_fn,
99 static int btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
101 BUG_ON(b->data || b->aux_data);
103 b->data = kvpmalloc(btree_bytes(c), gfp);
105 return -BCH_ERR_ENOMEM_btree_node_mem_alloc;
107 b->aux_data = kvmalloc(btree_aux_data_bytes(b), gfp);
109 b->aux_data = mmap(NULL, btree_aux_data_bytes(b),
110 PROT_READ|PROT_WRITE|PROT_EXEC,
111 MAP_PRIVATE|MAP_ANONYMOUS, 0, 0);
112 if (b->aux_data == MAP_FAILED)
116 kvpfree(b->data, btree_bytes(c));
118 return -BCH_ERR_ENOMEM_btree_node_mem_alloc;
124 static struct btree *__btree_node_mem_alloc(struct bch_fs *c, gfp_t gfp)
128 b = kzalloc(sizeof(struct btree), gfp);
132 bkey_btree_ptr_init(&b->key);
133 INIT_LIST_HEAD(&b->list);
134 INIT_LIST_HEAD(&b->write_blocked);
135 b->byte_order = ilog2(btree_bytes(c));
139 struct btree *__bch2_btree_node_mem_alloc(struct bch_fs *c)
141 struct btree_cache *bc = &c->btree_cache;
144 b = __btree_node_mem_alloc(c, GFP_KERNEL);
148 if (btree_node_data_alloc(c, b, GFP_KERNEL)) {
153 bch2_btree_lock_init(&b->c, 0);
156 list_add(&b->list, &bc->freeable);
160 /* Btree in memory cache - hash table */
162 void bch2_btree_node_hash_remove(struct btree_cache *bc, struct btree *b)
164 int ret = rhashtable_remove_fast(&bc->table, &b->hash, bch_btree_cache_params);
168 /* Cause future lookups for this node to fail: */
172 int __bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b)
175 b->hash_val = btree_ptr_hash_val(&b->key);
177 return rhashtable_lookup_insert_fast(&bc->table, &b->hash,
178 bch_btree_cache_params);
181 int bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b,
182 unsigned level, enum btree_id id)
189 mutex_lock(&bc->lock);
190 ret = __bch2_btree_node_hash_insert(bc, b);
192 list_add_tail(&b->list, &bc->live);
193 mutex_unlock(&bc->lock);
199 static inline struct btree *btree_cache_find(struct btree_cache *bc,
200 const struct bkey_i *k)
202 u64 v = btree_ptr_hash_val(k);
204 return rhashtable_lookup_fast(&bc->table, &v, bch_btree_cache_params);
208 * this version is for btree nodes that have already been freed (we're not
209 * reaping a real btree node)
211 static int __btree_node_reclaim(struct bch_fs *c, struct btree *b, bool flush, bool shrinker_counter)
213 struct btree_cache *bc = &c->btree_cache;
216 lockdep_assert_held(&bc->lock);
218 if (b->flags & ((1U << BTREE_NODE_dirty)|
219 (1U << BTREE_NODE_read_in_flight)|
220 (1U << BTREE_NODE_write_in_flight))) {
222 if (btree_node_dirty(b))
223 BTREE_CACHE_NOT_FREED_INCREMENT(dirty);
224 else if (btree_node_read_in_flight(b))
225 BTREE_CACHE_NOT_FREED_INCREMENT(read_in_flight);
226 else if (btree_node_write_in_flight(b))
227 BTREE_CACHE_NOT_FREED_INCREMENT(write_in_flight);
228 return -BCH_ERR_ENOMEM_btree_node_reclaim;
231 /* XXX: waiting on IO with btree cache lock held */
232 bch2_btree_node_wait_on_read(b);
233 bch2_btree_node_wait_on_write(b);
236 if (!six_trylock_intent(&b->c.lock)) {
237 BTREE_CACHE_NOT_FREED_INCREMENT(lock_intent);
238 return -BCH_ERR_ENOMEM_btree_node_reclaim;
241 if (!six_trylock_write(&b->c.lock)) {
242 BTREE_CACHE_NOT_FREED_INCREMENT(lock_write);
243 goto out_unlock_intent;
246 /* recheck under lock */
247 if (b->flags & ((1U << BTREE_NODE_read_in_flight)|
248 (1U << BTREE_NODE_write_in_flight))) {
250 if (btree_node_read_in_flight(b))
251 BTREE_CACHE_NOT_FREED_INCREMENT(read_in_flight);
252 else if (btree_node_write_in_flight(b))
253 BTREE_CACHE_NOT_FREED_INCREMENT(write_in_flight);
256 six_unlock_write(&b->c.lock);
257 six_unlock_intent(&b->c.lock);
261 if (btree_node_noevict(b)) {
262 BTREE_CACHE_NOT_FREED_INCREMENT(noevict);
265 if (btree_node_write_blocked(b)) {
266 BTREE_CACHE_NOT_FREED_INCREMENT(write_blocked);
269 if (btree_node_will_make_reachable(b)) {
270 BTREE_CACHE_NOT_FREED_INCREMENT(will_make_reachable);
274 if (btree_node_dirty(b)) {
276 BTREE_CACHE_NOT_FREED_INCREMENT(dirty);
280 * Using the underscore version because we don't want to compact
281 * bsets after the write, since this node is about to be evicted
282 * - unless btree verify mode is enabled, since it runs out of
283 * the post write cleanup:
285 if (bch2_verify_btree_ondisk)
286 bch2_btree_node_write(c, b, SIX_LOCK_intent,
287 BTREE_WRITE_cache_reclaim);
289 __bch2_btree_node_write(c, b,
290 BTREE_WRITE_cache_reclaim);
292 six_unlock_write(&b->c.lock);
293 six_unlock_intent(&b->c.lock);
297 if (b->hash_val && !ret)
298 trace_and_count(c, btree_cache_reap, c, b);
301 six_unlock_write(&b->c.lock);
303 six_unlock_intent(&b->c.lock);
304 ret = -BCH_ERR_ENOMEM_btree_node_reclaim;
308 static int btree_node_reclaim(struct bch_fs *c, struct btree *b, bool shrinker_counter)
310 return __btree_node_reclaim(c, b, false, shrinker_counter);
313 static int btree_node_write_and_reclaim(struct bch_fs *c, struct btree *b)
315 return __btree_node_reclaim(c, b, true, false);
318 static unsigned long bch2_btree_cache_scan(struct shrinker *shrink,
319 struct shrink_control *sc)
321 struct bch_fs *c = container_of(shrink, struct bch_fs,
323 struct btree_cache *bc = &c->btree_cache;
325 unsigned long nr = sc->nr_to_scan;
326 unsigned long can_free = 0;
327 unsigned long freed = 0;
328 unsigned long touched = 0;
330 unsigned long ret = SHRINK_STOP;
331 bool trigger_writes = atomic_read(&bc->dirty) + nr >=
334 if (bch2_btree_shrinker_disabled)
337 mutex_lock(&bc->lock);
338 flags = memalloc_nofs_save();
341 * It's _really_ critical that we don't free too many btree nodes - we
342 * have to always leave ourselves a reserve. The reserve is how we
343 * guarantee that allocating memory for a new btree node can always
344 * succeed, so that inserting keys into the btree can always succeed and
345 * IO can always make forward progress:
347 can_free = btree_cache_can_free(bc);
348 nr = min_t(unsigned long, nr, can_free);
351 list_for_each_entry_safe(b, t, &bc->freeable, list) {
353 * Leave a few nodes on the freeable list, so that a btree split
354 * won't have to hit the system allocator:
364 if (!btree_node_reclaim(c, b, true)) {
365 btree_node_data_free(c, b);
366 six_unlock_write(&b->c.lock);
367 six_unlock_intent(&b->c.lock);
373 list_for_each_entry_safe(b, t, &bc->live, list) {
376 if (btree_node_accessed(b)) {
377 clear_btree_node_accessed(b);
378 bc->not_freed_access_bit++;
379 } else if (!btree_node_reclaim(c, b, true)) {
381 btree_node_data_free(c, b);
384 bch2_btree_node_hash_remove(bc, b);
385 six_unlock_write(&b->c.lock);
386 six_unlock_intent(&b->c.lock);
390 } else if (trigger_writes &&
391 btree_node_dirty(b) &&
392 !btree_node_will_make_reachable(b) &&
393 !btree_node_write_blocked(b) &&
394 six_trylock_read(&b->c.lock)) {
395 list_move(&bc->live, &b->list);
396 mutex_unlock(&bc->lock);
397 __bch2_btree_node_write(c, b, BTREE_WRITE_cache_reclaim);
398 six_unlock_read(&b->c.lock);
401 mutex_lock(&bc->lock);
409 if (&t->list != &bc->live)
410 list_move_tail(&bc->live, &t->list);
412 mutex_unlock(&bc->lock);
415 memalloc_nofs_restore(flags);
416 trace_and_count(c, btree_cache_scan, sc->nr_to_scan, can_free, ret);
420 static unsigned long bch2_btree_cache_count(struct shrinker *shrink,
421 struct shrink_control *sc)
423 struct bch_fs *c = container_of(shrink, struct bch_fs,
425 struct btree_cache *bc = &c->btree_cache;
427 if (bch2_btree_shrinker_disabled)
430 return btree_cache_can_free(bc);
433 static void bch2_btree_cache_shrinker_to_text(struct seq_buf *s, struct shrinker *shrink)
435 struct bch_fs *c = container_of(shrink, struct bch_fs,
438 size_t buflen = seq_buf_get_buf(s, &cbuf);
439 struct printbuf out = PRINTBUF_EXTERN(cbuf, buflen);
441 bch2_btree_cache_to_text(&out, &c->btree_cache);
442 seq_buf_commit(s, out.pos);
445 void bch2_fs_btree_cache_exit(struct bch_fs *c)
447 struct btree_cache *bc = &c->btree_cache;
451 unregister_shrinker(&bc->shrink);
453 /* vfree() can allocate memory: */
454 flags = memalloc_nofs_save();
455 mutex_lock(&bc->lock);
458 list_move(&c->verify_data->list, &bc->live);
460 kvpfree(c->verify_ondisk, btree_bytes(c));
462 for (i = 0; i < btree_id_nr_alive(c); i++) {
463 struct btree_root *r = bch2_btree_id_root(c, i);
466 list_add(&r->b->list, &bc->live);
469 list_splice(&bc->freeable, &bc->live);
471 while (!list_empty(&bc->live)) {
472 b = list_first_entry(&bc->live, struct btree, list);
474 BUG_ON(btree_node_read_in_flight(b) ||
475 btree_node_write_in_flight(b));
477 if (btree_node_dirty(b))
478 bch2_btree_complete_write(c, b, btree_current_write(b));
479 clear_btree_node_dirty_acct(c, b);
481 btree_node_data_free(c, b);
484 BUG_ON(atomic_read(&c->btree_cache.dirty));
486 list_splice(&bc->freed_pcpu, &bc->freed_nonpcpu);
488 while (!list_empty(&bc->freed_nonpcpu)) {
489 b = list_first_entry(&bc->freed_nonpcpu, struct btree, list);
491 six_lock_exit(&b->c.lock);
495 mutex_unlock(&bc->lock);
496 memalloc_nofs_restore(flags);
498 if (bc->table_init_done)
499 rhashtable_destroy(&bc->table);
502 int bch2_fs_btree_cache_init(struct bch_fs *c)
504 struct btree_cache *bc = &c->btree_cache;
508 ret = rhashtable_init(&bc->table, &bch_btree_cache_params);
512 bc->table_init_done = true;
514 bch2_recalc_btree_reserve(c);
516 for (i = 0; i < bc->reserve; i++)
517 if (!__bch2_btree_node_mem_alloc(c))
520 list_splice_init(&bc->live, &bc->freeable);
522 mutex_init(&c->verify_lock);
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);
534 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
537 void bch2_fs_btree_cache_init_early(struct btree_cache *bc)
539 mutex_init(&bc->lock);
540 INIT_LIST_HEAD(&bc->live);
541 INIT_LIST_HEAD(&bc->freeable);
542 INIT_LIST_HEAD(&bc->freed_pcpu);
543 INIT_LIST_HEAD(&bc->freed_nonpcpu);
547 * We can only have one thread cannibalizing other cached btree nodes at a time,
548 * or we'll deadlock. We use an open coded mutex to ensure that, which a
549 * cannibalize_bucket() will take. This means every time we unlock the root of
550 * the btree, we need to release this lock if we have it held.
552 void bch2_btree_cache_cannibalize_unlock(struct bch_fs *c)
554 struct btree_cache *bc = &c->btree_cache;
556 if (bc->alloc_lock == current) {
557 trace_and_count(c, btree_cache_cannibalize_unlock, c);
558 bc->alloc_lock = NULL;
559 closure_wake_up(&bc->alloc_wait);
563 int bch2_btree_cache_cannibalize_lock(struct bch_fs *c, struct closure *cl)
565 struct btree_cache *bc = &c->btree_cache;
566 struct task_struct *old;
568 old = cmpxchg(&bc->alloc_lock, NULL, current);
569 if (old == NULL || old == current)
573 trace_and_count(c, btree_cache_cannibalize_lock_fail, c);
574 return -BCH_ERR_ENOMEM_btree_cache_cannibalize_lock;
577 closure_wait(&bc->alloc_wait, cl);
579 /* Try again, after adding ourselves to waitlist */
580 old = cmpxchg(&bc->alloc_lock, NULL, current);
581 if (old == NULL || old == current) {
583 closure_wake_up(&bc->alloc_wait);
587 trace_and_count(c, btree_cache_cannibalize_lock_fail, c);
588 return -BCH_ERR_btree_cache_cannibalize_lock_blocked;
591 trace_and_count(c, btree_cache_cannibalize_lock, c);
595 static struct btree *btree_node_cannibalize(struct bch_fs *c)
597 struct btree_cache *bc = &c->btree_cache;
600 list_for_each_entry_reverse(b, &bc->live, list)
601 if (!btree_node_reclaim(c, b, false))
605 list_for_each_entry_reverse(b, &bc->live, list)
606 if (!btree_node_write_and_reclaim(c, b))
610 * Rare case: all nodes were intent-locked.
613 WARN_ONCE(1, "btree cache cannibalize failed\n");
618 struct btree *bch2_btree_node_mem_alloc(struct btree_trans *trans, bool pcpu_read_locks)
620 struct bch_fs *c = trans->c;
621 struct btree_cache *bc = &c->btree_cache;
622 struct list_head *freed = pcpu_read_locks
624 : &bc->freed_nonpcpu;
625 struct btree *b, *b2;
626 u64 start_time = local_clock();
629 flags = memalloc_nofs_save();
630 mutex_lock(&bc->lock);
633 * We never free struct btree itself, just the memory that holds the on
634 * disk node. Check the freed list before allocating a new one:
636 list_for_each_entry(b, freed, list)
637 if (!btree_node_reclaim(c, b, false)) {
638 list_del_init(&b->list);
642 b = __btree_node_mem_alloc(c, GFP_NOWAIT|__GFP_NOWARN);
644 mutex_unlock(&bc->lock);
645 bch2_trans_unlock(trans);
646 b = __btree_node_mem_alloc(c, GFP_KERNEL);
649 mutex_lock(&bc->lock);
652 bch2_btree_lock_init(&b->c, pcpu_read_locks ? SIX_LOCK_INIT_PCPU : 0);
654 BUG_ON(!six_trylock_intent(&b->c.lock));
655 BUG_ON(!six_trylock_write(&b->c.lock));
659 * btree_free() doesn't free memory; it sticks the node on the end of
660 * the list. Check if there's any freed nodes there:
662 list_for_each_entry(b2, &bc->freeable, list)
663 if (!btree_node_reclaim(c, b2, false)) {
664 swap(b->data, b2->data);
665 swap(b->aux_data, b2->aux_data);
666 btree_node_to_freedlist(bc, b2);
667 six_unlock_write(&b2->c.lock);
668 six_unlock_intent(&b2->c.lock);
672 mutex_unlock(&bc->lock);
674 if (btree_node_data_alloc(c, b, GFP_NOWAIT|__GFP_NOWARN)) {
675 bch2_trans_unlock(trans);
676 if (btree_node_data_alloc(c, b, GFP_KERNEL|__GFP_NOWARN))
680 mutex_lock(&bc->lock);
683 mutex_unlock(&bc->lock);
685 BUG_ON(btree_node_hashed(b));
686 BUG_ON(btree_node_dirty(b));
687 BUG_ON(btree_node_write_in_flight(b));
694 b->whiteout_u64s = 0;
695 bch2_btree_keys_init(b);
696 set_btree_node_accessed(b);
698 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],
701 memalloc_nofs_restore(flags);
704 mutex_lock(&bc->lock);
706 /* Try to cannibalize another cached btree node: */
707 if (bc->alloc_lock == current) {
708 b2 = btree_node_cannibalize(c);
709 clear_btree_node_just_written(b2);
710 bch2_btree_node_hash_remove(bc, b2);
713 swap(b->data, b2->data);
714 swap(b->aux_data, b2->aux_data);
715 btree_node_to_freedlist(bc, b2);
716 six_unlock_write(&b2->c.lock);
717 six_unlock_intent(&b2->c.lock);
720 list_del_init(&b->list);
723 mutex_unlock(&bc->lock);
725 trace_and_count(c, btree_cache_cannibalize, c);
729 mutex_unlock(&bc->lock);
730 memalloc_nofs_restore(flags);
731 return ERR_PTR(-BCH_ERR_ENOMEM_btree_node_mem_alloc);
734 /* Slowpath, don't want it inlined into btree_iter_traverse() */
735 static noinline struct btree *bch2_btree_node_fill(struct btree_trans *trans,
736 struct btree_path *path,
737 const struct bkey_i *k,
738 enum btree_id btree_id,
740 enum six_lock_type lock_type,
743 struct bch_fs *c = trans->c;
744 struct btree_cache *bc = &c->btree_cache;
748 BUG_ON(level + 1 >= BTREE_MAX_DEPTH);
750 * Parent node must be locked, else we could read in a btree node that's
753 if (path && !bch2_btree_node_relock(trans, path, level + 1)) {
754 trace_and_count(c, trans_restart_relock_parent_for_fill, trans, _THIS_IP_, path);
755 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_fill_relock));
758 b = bch2_btree_node_mem_alloc(trans, level != 0);
760 if (bch2_err_matches(PTR_ERR_OR_ZERO(b), ENOMEM)) {
761 trans->memory_allocation_failure = true;
762 trace_and_count(c, trans_restart_memory_allocation_failure, trans, _THIS_IP_, path);
763 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_fill_mem_alloc_fail));
770 * Btree nodes read in from disk should not have the accessed bit set
771 * initially, so that linear scans don't thrash the cache:
773 clear_btree_node_accessed(b);
775 bkey_copy(&b->key, k);
776 if (bch2_btree_node_hash_insert(bc, b, level, btree_id)) {
777 /* raced with another fill: */
779 /* mark as unhashed... */
782 mutex_lock(&bc->lock);
783 list_add(&b->list, &bc->freeable);
784 mutex_unlock(&bc->lock);
786 six_unlock_write(&b->c.lock);
787 six_unlock_intent(&b->c.lock);
791 set_btree_node_read_in_flight(b);
793 six_unlock_write(&b->c.lock);
794 seq = six_lock_seq(&b->c.lock);
795 six_unlock_intent(&b->c.lock);
797 /* Unlock before doing IO: */
799 bch2_trans_unlock_noassert(trans);
801 bch2_btree_node_read(c, b, sync);
807 int ret = bch2_trans_relock(trans) ?:
808 bch2_btree_path_relock_intent(trans, path);
810 BUG_ON(!trans->restarted);
815 if (!six_relock_type(&b->c.lock, lock_type, seq)) {
817 trace_and_count(c, trans_restart_relock_after_fill, trans, _THIS_IP_, path);
818 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_relock_after_fill));
824 static noinline void btree_bad_header(struct bch_fs *c, struct btree *b)
826 struct printbuf buf = PRINTBUF;
828 if (c->curr_recovery_pass <= BCH_RECOVERY_PASS_check_allocations)
832 "btree node header doesn't match ptr\n"
833 "btree %s level %u\n"
835 bch2_btree_ids[b->c.btree_id], b->c.level);
836 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
838 prt_printf(&buf, "\nheader: btree %s level %llu\n"
840 bch2_btree_ids[BTREE_NODE_ID(b->data)],
841 BTREE_NODE_LEVEL(b->data));
842 bch2_bpos_to_text(&buf, b->data->min_key);
844 prt_printf(&buf, "\nmax ");
845 bch2_bpos_to_text(&buf, b->data->max_key);
847 bch2_fs_inconsistent(c, "%s", buf.buf);
851 static inline void btree_check_header(struct bch_fs *c, struct btree *b)
853 if (b->c.btree_id != BTREE_NODE_ID(b->data) ||
854 b->c.level != BTREE_NODE_LEVEL(b->data) ||
855 !bpos_eq(b->data->max_key, b->key.k.p) ||
856 (b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
857 !bpos_eq(b->data->min_key,
858 bkey_i_to_btree_ptr_v2(&b->key)->v.min_key)))
859 btree_bad_header(c, b);
862 static struct btree *__bch2_btree_node_get(struct btree_trans *trans, struct btree_path *path,
863 const struct bkey_i *k, unsigned level,
864 enum six_lock_type lock_type,
865 unsigned long trace_ip)
867 struct bch_fs *c = trans->c;
868 struct btree_cache *bc = &c->btree_cache;
871 bool need_relock = false;
874 EBUG_ON(level >= BTREE_MAX_DEPTH);
876 b = btree_cache_find(bc, k);
879 * We must have the parent locked to call bch2_btree_node_fill(),
880 * else we could read in a btree node from disk that's been
883 b = bch2_btree_node_fill(trans, path, k, path->btree_id,
884 level, lock_type, true);
887 /* We raced and found the btree node in the cache */
894 if (btree_node_read_locked(path, level + 1))
895 btree_node_unlock(trans, path, level + 1);
897 ret = btree_node_lock(trans, path, &b->c, level, lock_type, trace_ip);
898 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
903 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
904 b->c.level != level ||
906 six_unlock_type(&b->c.lock, lock_type);
907 if (bch2_btree_node_relock(trans, path, level + 1))
910 trace_and_count(c, trans_restart_btree_node_reused, trans, trace_ip, path);
911 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_lock_node_reused));
914 /* avoid atomic set bit if it's not needed: */
915 if (!btree_node_accessed(b))
916 set_btree_node_accessed(b);
919 if (unlikely(btree_node_read_in_flight(b))) {
920 u32 seq = six_lock_seq(&b->c.lock);
922 six_unlock_type(&b->c.lock, lock_type);
923 bch2_trans_unlock(trans);
926 bch2_btree_node_wait_on_read(b);
929 * should_be_locked is not set on this path yet, so we need to
930 * relock it specifically:
932 if (!six_relock_type(&b->c.lock, lock_type, seq))
936 if (unlikely(need_relock)) {
937 ret = bch2_trans_relock(trans) ?:
938 bch2_btree_path_relock_intent(trans, path);
940 six_unlock_type(&b->c.lock, lock_type);
945 prefetch(b->aux_data);
947 for_each_bset(b, t) {
948 void *p = (u64 *) b->aux_data + t->aux_data_offset;
950 prefetch(p + L1_CACHE_BYTES * 0);
951 prefetch(p + L1_CACHE_BYTES * 1);
952 prefetch(p + L1_CACHE_BYTES * 2);
955 if (unlikely(btree_node_read_error(b))) {
956 six_unlock_type(&b->c.lock, lock_type);
957 return ERR_PTR(-EIO);
960 EBUG_ON(b->c.btree_id != path->btree_id);
961 EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
962 btree_check_header(c, b);
968 * bch2_btree_node_get - find a btree node in the cache and lock it, reading it
969 * in from disk if necessary.
971 * @trans: btree transaction object
972 * @path: btree_path being traversed
973 * @k: pointer to btree node (generally KEY_TYPE_btree_ptr_v2)
974 * @level: level of btree node being looked up (0 == leaf node)
975 * @lock_type: SIX_LOCK_read or SIX_LOCK_intent
976 * @trace_ip: ip of caller of btree iterator code (i.e. caller of bch2_btree_iter_peek())
978 * The btree node will have either a read or a write lock held, depending on
979 * the @write parameter.
981 * Returns: btree node or ERR_PTR()
983 struct btree *bch2_btree_node_get(struct btree_trans *trans, struct btree_path *path,
984 const struct bkey_i *k, unsigned level,
985 enum six_lock_type lock_type,
986 unsigned long trace_ip)
988 struct bch_fs *c = trans->c;
993 EBUG_ON(level >= BTREE_MAX_DEPTH);
995 b = btree_node_mem_ptr(k);
998 * Check b->hash_val _before_ calling btree_node_lock() - this might not
999 * be the node we want anymore, and trying to lock the wrong node could
1000 * cause an unneccessary transaction restart:
1002 if (unlikely(!c->opts.btree_node_mem_ptr_optimization ||
1004 b->hash_val != btree_ptr_hash_val(k)))
1005 return __bch2_btree_node_get(trans, path, k, level, lock_type, trace_ip);
1007 if (btree_node_read_locked(path, level + 1))
1008 btree_node_unlock(trans, path, level + 1);
1010 ret = btree_node_lock(trans, path, &b->c, level, lock_type, trace_ip);
1011 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1012 return ERR_PTR(ret);
1016 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
1017 b->c.level != level ||
1019 six_unlock_type(&b->c.lock, lock_type);
1020 if (bch2_btree_node_relock(trans, path, level + 1))
1021 return __bch2_btree_node_get(trans, path, k, level, lock_type, trace_ip);
1023 trace_and_count(c, trans_restart_btree_node_reused, trans, trace_ip, path);
1024 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_lock_node_reused));
1027 if (unlikely(btree_node_read_in_flight(b))) {
1028 six_unlock_type(&b->c.lock, lock_type);
1029 return __bch2_btree_node_get(trans, path, k, level, lock_type, trace_ip);
1032 prefetch(b->aux_data);
1034 for_each_bset(b, t) {
1035 void *p = (u64 *) b->aux_data + t->aux_data_offset;
1037 prefetch(p + L1_CACHE_BYTES * 0);
1038 prefetch(p + L1_CACHE_BYTES * 1);
1039 prefetch(p + L1_CACHE_BYTES * 2);
1042 /* avoid atomic set bit if it's not needed: */
1043 if (!btree_node_accessed(b))
1044 set_btree_node_accessed(b);
1046 if (unlikely(btree_node_read_error(b))) {
1047 six_unlock_type(&b->c.lock, lock_type);
1048 return ERR_PTR(-EIO);
1051 EBUG_ON(b->c.btree_id != path->btree_id);
1052 EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
1053 btree_check_header(c, b);
1058 struct btree *bch2_btree_node_get_noiter(struct btree_trans *trans,
1059 const struct bkey_i *k,
1060 enum btree_id btree_id,
1064 struct bch_fs *c = trans->c;
1065 struct btree_cache *bc = &c->btree_cache;
1067 struct bset_tree *t;
1070 EBUG_ON(level >= BTREE_MAX_DEPTH);
1072 if (c->opts.btree_node_mem_ptr_optimization) {
1073 b = btree_node_mem_ptr(k);
1078 b = btree_cache_find(bc, k);
1083 b = bch2_btree_node_fill(trans, NULL, k, btree_id,
1084 level, SIX_LOCK_read, true);
1086 /* We raced and found the btree node in the cache */
1091 !bch2_btree_cache_cannibalize_lock(c, NULL))
1098 ret = btree_node_lock_nopath(trans, &b->c, SIX_LOCK_read, _THIS_IP_);
1099 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1100 return ERR_PTR(ret);
1104 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
1105 b->c.btree_id != btree_id ||
1106 b->c.level != level)) {
1107 six_unlock_read(&b->c.lock);
1112 /* XXX: waiting on IO with btree locks held: */
1113 __bch2_btree_node_wait_on_read(b);
1115 prefetch(b->aux_data);
1117 for_each_bset(b, t) {
1118 void *p = (u64 *) b->aux_data + t->aux_data_offset;
1120 prefetch(p + L1_CACHE_BYTES * 0);
1121 prefetch(p + L1_CACHE_BYTES * 1);
1122 prefetch(p + L1_CACHE_BYTES * 2);
1125 /* avoid atomic set bit if it's not needed: */
1126 if (!btree_node_accessed(b))
1127 set_btree_node_accessed(b);
1129 if (unlikely(btree_node_read_error(b))) {
1130 six_unlock_read(&b->c.lock);
1135 EBUG_ON(b->c.btree_id != btree_id);
1136 EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
1137 btree_check_header(c, b);
1139 bch2_btree_cache_cannibalize_unlock(c);
1143 int bch2_btree_node_prefetch(struct btree_trans *trans,
1144 struct btree_path *path,
1145 const struct bkey_i *k,
1146 enum btree_id btree_id, unsigned level)
1148 struct bch_fs *c = trans->c;
1149 struct btree_cache *bc = &c->btree_cache;
1152 BUG_ON(trans && !btree_node_locked(path, level + 1));
1153 BUG_ON(level >= BTREE_MAX_DEPTH);
1155 b = btree_cache_find(bc, k);
1159 b = bch2_btree_node_fill(trans, path, k, btree_id,
1160 level, SIX_LOCK_read, false);
1161 return PTR_ERR_OR_ZERO(b);
1164 void bch2_btree_node_evict(struct btree_trans *trans, const struct bkey_i *k)
1166 struct bch_fs *c = trans->c;
1167 struct btree_cache *bc = &c->btree_cache;
1170 b = btree_cache_find(bc, k);
1174 /* not allowed to wait on io with btree locks held: */
1176 /* XXX we're called from btree_gc which will be holding other btree
1179 __bch2_btree_node_wait_on_read(b);
1180 __bch2_btree_node_wait_on_write(b);
1182 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
1183 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
1185 if (btree_node_dirty(b)) {
1186 __bch2_btree_node_write(c, b, BTREE_WRITE_cache_reclaim);
1187 six_unlock_write(&b->c.lock);
1188 six_unlock_intent(&b->c.lock);
1192 BUG_ON(btree_node_dirty(b));
1194 mutex_lock(&bc->lock);
1195 btree_node_data_free(c, b);
1196 bch2_btree_node_hash_remove(bc, b);
1197 mutex_unlock(&bc->lock);
1199 six_unlock_write(&b->c.lock);
1200 six_unlock_intent(&b->c.lock);
1203 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
1204 const struct btree *b)
1206 struct bset_stats stats;
1208 memset(&stats, 0, sizeof(stats));
1210 bch2_btree_keys_stats(b, &stats);
1212 prt_printf(out, "l %u ", b->c.level);
1213 bch2_bpos_to_text(out, b->data->min_key);
1214 prt_printf(out, " - ");
1215 bch2_bpos_to_text(out, b->data->max_key);
1216 prt_printf(out, ":\n"
1218 bch2_val_to_text(out, c, bkey_i_to_s_c(&b->key));
1223 bch2_bkey_format_to_text(out, &b->format);
1226 " unpack fn len: %u\n"
1227 " bytes used %zu/%zu (%zu%% full)\n"
1228 " sib u64s: %u, %u (merge threshold %u)\n"
1229 " nr packed keys %u\n"
1230 " nr unpacked keys %u\n"
1232 " failed unpacked %zu\n",
1234 b->nr.live_u64s * sizeof(u64),
1235 btree_bytes(c) - sizeof(struct btree_node),
1236 b->nr.live_u64s * 100 / btree_max_u64s(c),
1239 c->btree_foreground_merge_threshold,
1241 b->nr.unpacked_keys,
1246 void bch2_btree_cache_to_text(struct printbuf *out, const struct btree_cache *bc)
1248 prt_printf(out, "nr nodes:\t\t%u\n", bc->used);
1249 prt_printf(out, "nr dirty:\t\t%u\n", atomic_read(&bc->dirty));
1250 prt_printf(out, "cannibalize lock:\t%p\n", bc->alloc_lock);
1252 prt_printf(out, "freed:\t\t\t\t%u\n", bc->freed);
1253 prt_printf(out, "not freed, dirty:\t\t%u\n", bc->not_freed_dirty);
1254 prt_printf(out, "not freed, write in flight:\t%u\n", bc->not_freed_write_in_flight);
1255 prt_printf(out, "not freed, read in flight:\t%u\n", bc->not_freed_read_in_flight);
1256 prt_printf(out, "not freed, lock intent failed:\t%u\n", bc->not_freed_lock_intent);
1257 prt_printf(out, "not freed, lock write failed:\t%u\n", bc->not_freed_lock_write);
1258 prt_printf(out, "not freed, access bit:\t\t%u\n", bc->not_freed_access_bit);
1259 prt_printf(out, "not freed, no evict failed:\t%u\n", bc->not_freed_noevict);
1260 prt_printf(out, "not freed, write blocked:\t%u\n", bc->not_freed_write_blocked);
1261 prt_printf(out, "not freed, will make reachable:\t%u\n", bc->not_freed_will_make_reachable);