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 = shrink->private_data;
322 struct btree_cache *bc = &c->btree_cache;
324 unsigned long nr = sc->nr_to_scan;
325 unsigned long can_free = 0;
326 unsigned long freed = 0;
327 unsigned long touched = 0;
329 unsigned long ret = SHRINK_STOP;
330 bool trigger_writes = atomic_read(&bc->dirty) + nr >=
333 if (bch2_btree_shrinker_disabled)
336 mutex_lock(&bc->lock);
337 flags = memalloc_nofs_save();
340 * It's _really_ critical that we don't free too many btree nodes - we
341 * have to always leave ourselves a reserve. The reserve is how we
342 * guarantee that allocating memory for a new btree node can always
343 * succeed, so that inserting keys into the btree can always succeed and
344 * IO can always make forward progress:
346 can_free = btree_cache_can_free(bc);
347 nr = min_t(unsigned long, nr, can_free);
350 list_for_each_entry_safe(b, t, &bc->freeable, list) {
352 * Leave a few nodes on the freeable list, so that a btree split
353 * won't have to hit the system allocator:
363 if (!btree_node_reclaim(c, b, true)) {
364 btree_node_data_free(c, b);
365 six_unlock_write(&b->c.lock);
366 six_unlock_intent(&b->c.lock);
372 list_for_each_entry_safe(b, t, &bc->live, list) {
375 if (btree_node_accessed(b)) {
376 clear_btree_node_accessed(b);
377 bc->not_freed_access_bit++;
378 } else if (!btree_node_reclaim(c, b, true)) {
380 btree_node_data_free(c, b);
383 bch2_btree_node_hash_remove(bc, b);
384 six_unlock_write(&b->c.lock);
385 six_unlock_intent(&b->c.lock);
389 } else if (trigger_writes &&
390 btree_node_dirty(b) &&
391 !btree_node_will_make_reachable(b) &&
392 !btree_node_write_blocked(b) &&
393 six_trylock_read(&b->c.lock)) {
394 list_move(&bc->live, &b->list);
395 mutex_unlock(&bc->lock);
396 __bch2_btree_node_write(c, b, BTREE_WRITE_cache_reclaim);
397 six_unlock_read(&b->c.lock);
400 mutex_lock(&bc->lock);
408 if (&t->list != &bc->live)
409 list_move_tail(&bc->live, &t->list);
411 mutex_unlock(&bc->lock);
414 memalloc_nofs_restore(flags);
415 trace_and_count(c, btree_cache_scan, sc->nr_to_scan, can_free, ret);
419 static unsigned long bch2_btree_cache_count(struct shrinker *shrink,
420 struct shrink_control *sc)
422 struct bch_fs *c = shrink->private_data;
423 struct btree_cache *bc = &c->btree_cache;
425 if (bch2_btree_shrinker_disabled)
428 return btree_cache_can_free(bc);
431 static void bch2_btree_cache_shrinker_to_text(struct seq_buf *s, struct shrinker *shrink)
433 struct bch_fs *c = shrink->private_data;
435 size_t buflen = seq_buf_get_buf(s, &cbuf);
436 struct printbuf out = PRINTBUF_EXTERN(cbuf, buflen);
438 bch2_btree_cache_to_text(&out, &c->btree_cache);
439 seq_buf_commit(s, out.pos);
442 void bch2_fs_btree_cache_exit(struct bch_fs *c)
444 struct btree_cache *bc = &c->btree_cache;
448 shrinker_free(bc->shrink);
450 /* vfree() can allocate memory: */
451 flags = memalloc_nofs_save();
452 mutex_lock(&bc->lock);
455 list_move(&c->verify_data->list, &bc->live);
457 kvpfree(c->verify_ondisk, btree_bytes(c));
459 for (i = 0; i < btree_id_nr_alive(c); i++) {
460 struct btree_root *r = bch2_btree_id_root(c, i);
463 list_add(&r->b->list, &bc->live);
466 list_splice(&bc->freeable, &bc->live);
468 while (!list_empty(&bc->live)) {
469 b = list_first_entry(&bc->live, struct btree, list);
471 BUG_ON(btree_node_read_in_flight(b) ||
472 btree_node_write_in_flight(b));
474 if (btree_node_dirty(b))
475 bch2_btree_complete_write(c, b, btree_current_write(b));
476 clear_btree_node_dirty_acct(c, b);
478 btree_node_data_free(c, b);
481 BUG_ON(atomic_read(&c->btree_cache.dirty));
483 list_splice(&bc->freed_pcpu, &bc->freed_nonpcpu);
485 while (!list_empty(&bc->freed_nonpcpu)) {
486 b = list_first_entry(&bc->freed_nonpcpu, struct btree, list);
488 six_lock_exit(&b->c.lock);
492 mutex_unlock(&bc->lock);
493 memalloc_nofs_restore(flags);
495 if (bc->table_init_done)
496 rhashtable_destroy(&bc->table);
499 int bch2_fs_btree_cache_init(struct bch_fs *c)
501 struct btree_cache *bc = &c->btree_cache;
502 struct shrinker *shrink;
506 ret = rhashtable_init(&bc->table, &bch_btree_cache_params);
510 bc->table_init_done = true;
512 bch2_recalc_btree_reserve(c);
514 for (i = 0; i < bc->reserve; i++)
515 if (!__bch2_btree_node_mem_alloc(c))
518 list_splice_init(&bc->live, &bc->freeable);
520 mutex_init(&c->verify_lock);
522 shrink = shrinker_alloc(0, "%s-btree_cache", c->name);
526 shrink->count_objects = bch2_btree_cache_count;
527 shrink->scan_objects = bch2_btree_cache_scan;
528 shrink->to_text = bch2_btree_cache_shrinker_to_text;
530 shrink->private_data = c;
531 shrinker_register(shrink);
535 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
538 void bch2_fs_btree_cache_init_early(struct btree_cache *bc)
540 mutex_init(&bc->lock);
541 INIT_LIST_HEAD(&bc->live);
542 INIT_LIST_HEAD(&bc->freeable);
543 INIT_LIST_HEAD(&bc->freed_pcpu);
544 INIT_LIST_HEAD(&bc->freed_nonpcpu);
548 * We can only have one thread cannibalizing other cached btree nodes at a time,
549 * or we'll deadlock. We use an open coded mutex to ensure that, which a
550 * cannibalize_bucket() will take. This means every time we unlock the root of
551 * the btree, we need to release this lock if we have it held.
553 void bch2_btree_cache_cannibalize_unlock(struct bch_fs *c)
555 struct btree_cache *bc = &c->btree_cache;
557 if (bc->alloc_lock == current) {
558 trace_and_count(c, btree_cache_cannibalize_unlock, c);
559 bc->alloc_lock = NULL;
560 closure_wake_up(&bc->alloc_wait);
564 int bch2_btree_cache_cannibalize_lock(struct bch_fs *c, struct closure *cl)
566 struct btree_cache *bc = &c->btree_cache;
567 struct task_struct *old;
569 old = cmpxchg(&bc->alloc_lock, NULL, current);
570 if (old == NULL || old == current)
574 trace_and_count(c, btree_cache_cannibalize_lock_fail, c);
575 return -BCH_ERR_ENOMEM_btree_cache_cannibalize_lock;
578 closure_wait(&bc->alloc_wait, cl);
580 /* Try again, after adding ourselves to waitlist */
581 old = cmpxchg(&bc->alloc_lock, NULL, current);
582 if (old == NULL || old == current) {
584 closure_wake_up(&bc->alloc_wait);
588 trace_and_count(c, btree_cache_cannibalize_lock_fail, c);
589 return -BCH_ERR_btree_cache_cannibalize_lock_blocked;
592 trace_and_count(c, btree_cache_cannibalize_lock, c);
596 static struct btree *btree_node_cannibalize(struct bch_fs *c)
598 struct btree_cache *bc = &c->btree_cache;
601 list_for_each_entry_reverse(b, &bc->live, list)
602 if (!btree_node_reclaim(c, b, false))
606 list_for_each_entry_reverse(b, &bc->live, list)
607 if (!btree_node_write_and_reclaim(c, b))
611 * Rare case: all nodes were intent-locked.
614 WARN_ONCE(1, "btree cache cannibalize failed\n");
619 struct btree *bch2_btree_node_mem_alloc(struct btree_trans *trans, bool pcpu_read_locks)
621 struct bch_fs *c = trans->c;
622 struct btree_cache *bc = &c->btree_cache;
623 struct list_head *freed = pcpu_read_locks
625 : &bc->freed_nonpcpu;
626 struct btree *b, *b2;
627 u64 start_time = local_clock();
630 flags = memalloc_nofs_save();
631 mutex_lock(&bc->lock);
634 * We never free struct btree itself, just the memory that holds the on
635 * disk node. Check the freed list before allocating a new one:
637 list_for_each_entry(b, freed, list)
638 if (!btree_node_reclaim(c, b, false)) {
639 list_del_init(&b->list);
643 b = __btree_node_mem_alloc(c, GFP_NOWAIT|__GFP_NOWARN);
645 mutex_unlock(&bc->lock);
646 bch2_trans_unlock(trans);
647 b = __btree_node_mem_alloc(c, GFP_KERNEL);
650 mutex_lock(&bc->lock);
653 bch2_btree_lock_init(&b->c, pcpu_read_locks ? SIX_LOCK_INIT_PCPU : 0);
655 BUG_ON(!six_trylock_intent(&b->c.lock));
656 BUG_ON(!six_trylock_write(&b->c.lock));
660 * btree_free() doesn't free memory; it sticks the node on the end of
661 * the list. Check if there's any freed nodes there:
663 list_for_each_entry(b2, &bc->freeable, list)
664 if (!btree_node_reclaim(c, b2, false)) {
665 swap(b->data, b2->data);
666 swap(b->aux_data, b2->aux_data);
667 btree_node_to_freedlist(bc, b2);
668 six_unlock_write(&b2->c.lock);
669 six_unlock_intent(&b2->c.lock);
673 mutex_unlock(&bc->lock);
675 if (btree_node_data_alloc(c, b, GFP_NOWAIT|__GFP_NOWARN)) {
676 bch2_trans_unlock(trans);
677 if (btree_node_data_alloc(c, b, GFP_KERNEL|__GFP_NOWARN))
681 mutex_lock(&bc->lock);
684 mutex_unlock(&bc->lock);
686 BUG_ON(btree_node_hashed(b));
687 BUG_ON(btree_node_dirty(b));
688 BUG_ON(btree_node_write_in_flight(b));
695 b->whiteout_u64s = 0;
696 bch2_btree_keys_init(b);
697 set_btree_node_accessed(b);
699 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],
702 memalloc_nofs_restore(flags);
705 mutex_lock(&bc->lock);
707 /* Try to cannibalize another cached btree node: */
708 if (bc->alloc_lock == current) {
709 b2 = btree_node_cannibalize(c);
710 clear_btree_node_just_written(b2);
711 bch2_btree_node_hash_remove(bc, b2);
714 swap(b->data, b2->data);
715 swap(b->aux_data, b2->aux_data);
716 btree_node_to_freedlist(bc, b2);
717 six_unlock_write(&b2->c.lock);
718 six_unlock_intent(&b2->c.lock);
721 list_del_init(&b->list);
724 mutex_unlock(&bc->lock);
726 trace_and_count(c, btree_cache_cannibalize, c);
730 mutex_unlock(&bc->lock);
731 memalloc_nofs_restore(flags);
732 return ERR_PTR(-BCH_ERR_ENOMEM_btree_node_mem_alloc);
735 /* Slowpath, don't want it inlined into btree_iter_traverse() */
736 static noinline struct btree *bch2_btree_node_fill(struct btree_trans *trans,
737 struct btree_path *path,
738 const struct bkey_i *k,
739 enum btree_id btree_id,
741 enum six_lock_type lock_type,
744 struct bch_fs *c = trans->c;
745 struct btree_cache *bc = &c->btree_cache;
749 BUG_ON(level + 1 >= BTREE_MAX_DEPTH);
751 * Parent node must be locked, else we could read in a btree node that's
754 if (path && !bch2_btree_node_relock(trans, path, level + 1)) {
755 trace_and_count(c, trans_restart_relock_parent_for_fill, trans, _THIS_IP_, path);
756 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_fill_relock));
759 b = bch2_btree_node_mem_alloc(trans, level != 0);
761 if (bch2_err_matches(PTR_ERR_OR_ZERO(b), ENOMEM)) {
762 trans->memory_allocation_failure = true;
763 trace_and_count(c, trans_restart_memory_allocation_failure, trans, _THIS_IP_, path);
764 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_fill_mem_alloc_fail));
771 * Btree nodes read in from disk should not have the accessed bit set
772 * initially, so that linear scans don't thrash the cache:
774 clear_btree_node_accessed(b);
776 bkey_copy(&b->key, k);
777 if (bch2_btree_node_hash_insert(bc, b, level, btree_id)) {
778 /* raced with another fill: */
780 /* mark as unhashed... */
783 mutex_lock(&bc->lock);
784 list_add(&b->list, &bc->freeable);
785 mutex_unlock(&bc->lock);
787 six_unlock_write(&b->c.lock);
788 six_unlock_intent(&b->c.lock);
792 set_btree_node_read_in_flight(b);
794 six_unlock_write(&b->c.lock);
795 seq = six_lock_seq(&b->c.lock);
796 six_unlock_intent(&b->c.lock);
798 /* Unlock before doing IO: */
800 bch2_trans_unlock_noassert(trans);
802 bch2_btree_node_read(c, b, sync);
808 int ret = bch2_trans_relock(trans) ?:
809 bch2_btree_path_relock_intent(trans, path);
811 BUG_ON(!trans->restarted);
816 if (!six_relock_type(&b->c.lock, lock_type, seq)) {
818 trace_and_count(c, trans_restart_relock_after_fill, trans, _THIS_IP_, path);
819 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_relock_after_fill));
825 static noinline void btree_bad_header(struct bch_fs *c, struct btree *b)
827 struct printbuf buf = PRINTBUF;
829 if (c->curr_recovery_pass <= BCH_RECOVERY_PASS_check_allocations)
833 "btree node header doesn't match ptr\n"
834 "btree %s level %u\n"
836 bch2_btree_id_str(b->c.btree_id), b->c.level);
837 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
839 prt_printf(&buf, "\nheader: btree %s level %llu\n"
841 bch2_btree_id_str(BTREE_NODE_ID(b->data)),
842 BTREE_NODE_LEVEL(b->data));
843 bch2_bpos_to_text(&buf, b->data->min_key);
845 prt_printf(&buf, "\nmax ");
846 bch2_bpos_to_text(&buf, b->data->max_key);
848 bch2_fs_inconsistent(c, "%s", buf.buf);
852 static inline void btree_check_header(struct bch_fs *c, struct btree *b)
854 if (b->c.btree_id != BTREE_NODE_ID(b->data) ||
855 b->c.level != BTREE_NODE_LEVEL(b->data) ||
856 !bpos_eq(b->data->max_key, b->key.k.p) ||
857 (b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
858 !bpos_eq(b->data->min_key,
859 bkey_i_to_btree_ptr_v2(&b->key)->v.min_key)))
860 btree_bad_header(c, b);
863 static struct btree *__bch2_btree_node_get(struct btree_trans *trans, struct btree_path *path,
864 const struct bkey_i *k, unsigned level,
865 enum six_lock_type lock_type,
866 unsigned long trace_ip)
868 struct bch_fs *c = trans->c;
869 struct btree_cache *bc = &c->btree_cache;
872 bool need_relock = false;
875 EBUG_ON(level >= BTREE_MAX_DEPTH);
877 b = btree_cache_find(bc, k);
880 * We must have the parent locked to call bch2_btree_node_fill(),
881 * else we could read in a btree node from disk that's been
884 b = bch2_btree_node_fill(trans, path, k, path->btree_id,
885 level, lock_type, true);
888 /* We raced and found the btree node in the cache */
895 if (btree_node_read_locked(path, level + 1))
896 btree_node_unlock(trans, path, level + 1);
898 ret = btree_node_lock(trans, path, &b->c, level, lock_type, trace_ip);
899 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
904 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
905 b->c.level != level ||
907 six_unlock_type(&b->c.lock, lock_type);
908 if (bch2_btree_node_relock(trans, path, level + 1))
911 trace_and_count(c, trans_restart_btree_node_reused, trans, trace_ip, path);
912 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_lock_node_reused));
915 /* avoid atomic set bit if it's not needed: */
916 if (!btree_node_accessed(b))
917 set_btree_node_accessed(b);
920 if (unlikely(btree_node_read_in_flight(b))) {
921 u32 seq = six_lock_seq(&b->c.lock);
923 six_unlock_type(&b->c.lock, lock_type);
924 bch2_trans_unlock(trans);
927 bch2_btree_node_wait_on_read(b);
930 * should_be_locked is not set on this path yet, so we need to
931 * relock it specifically:
933 if (!six_relock_type(&b->c.lock, lock_type, seq))
937 if (unlikely(need_relock)) {
938 ret = bch2_trans_relock(trans) ?:
939 bch2_btree_path_relock_intent(trans, path);
941 six_unlock_type(&b->c.lock, lock_type);
946 prefetch(b->aux_data);
948 for_each_bset(b, t) {
949 void *p = (u64 *) b->aux_data + t->aux_data_offset;
951 prefetch(p + L1_CACHE_BYTES * 0);
952 prefetch(p + L1_CACHE_BYTES * 1);
953 prefetch(p + L1_CACHE_BYTES * 2);
956 if (unlikely(btree_node_read_error(b))) {
957 six_unlock_type(&b->c.lock, lock_type);
958 return ERR_PTR(-EIO);
961 EBUG_ON(b->c.btree_id != path->btree_id);
962 EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
963 btree_check_header(c, b);
969 * bch2_btree_node_get - find a btree node in the cache and lock it, reading it
970 * in from disk if necessary.
972 * @trans: btree transaction object
973 * @path: btree_path being traversed
974 * @k: pointer to btree node (generally KEY_TYPE_btree_ptr_v2)
975 * @level: level of btree node being looked up (0 == leaf node)
976 * @lock_type: SIX_LOCK_read or SIX_LOCK_intent
977 * @trace_ip: ip of caller of btree iterator code (i.e. caller of bch2_btree_iter_peek())
979 * The btree node will have either a read or a write lock held, depending on
980 * the @write parameter.
982 * Returns: btree node or ERR_PTR()
984 struct btree *bch2_btree_node_get(struct btree_trans *trans, struct btree_path *path,
985 const struct bkey_i *k, unsigned level,
986 enum six_lock_type lock_type,
987 unsigned long trace_ip)
989 struct bch_fs *c = trans->c;
994 EBUG_ON(level >= BTREE_MAX_DEPTH);
996 b = btree_node_mem_ptr(k);
999 * Check b->hash_val _before_ calling btree_node_lock() - this might not
1000 * be the node we want anymore, and trying to lock the wrong node could
1001 * cause an unneccessary transaction restart:
1003 if (unlikely(!c->opts.btree_node_mem_ptr_optimization ||
1005 b->hash_val != btree_ptr_hash_val(k)))
1006 return __bch2_btree_node_get(trans, path, k, level, lock_type, trace_ip);
1008 if (btree_node_read_locked(path, level + 1))
1009 btree_node_unlock(trans, path, level + 1);
1011 ret = btree_node_lock(trans, path, &b->c, level, lock_type, trace_ip);
1012 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1013 return ERR_PTR(ret);
1017 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
1018 b->c.level != level ||
1020 six_unlock_type(&b->c.lock, lock_type);
1021 if (bch2_btree_node_relock(trans, path, level + 1))
1022 return __bch2_btree_node_get(trans, path, k, level, lock_type, trace_ip);
1024 trace_and_count(c, trans_restart_btree_node_reused, trans, trace_ip, path);
1025 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_lock_node_reused));
1028 if (unlikely(btree_node_read_in_flight(b))) {
1029 six_unlock_type(&b->c.lock, lock_type);
1030 return __bch2_btree_node_get(trans, path, k, level, lock_type, trace_ip);
1033 prefetch(b->aux_data);
1035 for_each_bset(b, t) {
1036 void *p = (u64 *) b->aux_data + t->aux_data_offset;
1038 prefetch(p + L1_CACHE_BYTES * 0);
1039 prefetch(p + L1_CACHE_BYTES * 1);
1040 prefetch(p + L1_CACHE_BYTES * 2);
1043 /* avoid atomic set bit if it's not needed: */
1044 if (!btree_node_accessed(b))
1045 set_btree_node_accessed(b);
1047 if (unlikely(btree_node_read_error(b))) {
1048 six_unlock_type(&b->c.lock, lock_type);
1049 return ERR_PTR(-EIO);
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);
1059 struct btree *bch2_btree_node_get_noiter(struct btree_trans *trans,
1060 const struct bkey_i *k,
1061 enum btree_id btree_id,
1065 struct bch_fs *c = trans->c;
1066 struct btree_cache *bc = &c->btree_cache;
1068 struct bset_tree *t;
1071 EBUG_ON(level >= BTREE_MAX_DEPTH);
1073 if (c->opts.btree_node_mem_ptr_optimization) {
1074 b = btree_node_mem_ptr(k);
1079 b = btree_cache_find(bc, k);
1084 b = bch2_btree_node_fill(trans, NULL, k, btree_id,
1085 level, SIX_LOCK_read, true);
1087 /* We raced and found the btree node in the cache */
1092 !bch2_btree_cache_cannibalize_lock(c, NULL))
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);
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);
1113 /* XXX: waiting on IO with btree locks held: */
1114 __bch2_btree_node_wait_on_read(b);
1116 prefetch(b->aux_data);
1118 for_each_bset(b, t) {
1119 void *p = (u64 *) b->aux_data + t->aux_data_offset;
1121 prefetch(p + L1_CACHE_BYTES * 0);
1122 prefetch(p + L1_CACHE_BYTES * 1);
1123 prefetch(p + L1_CACHE_BYTES * 2);
1126 /* avoid atomic set bit if it's not needed: */
1127 if (!btree_node_accessed(b))
1128 set_btree_node_accessed(b);
1130 if (unlikely(btree_node_read_error(b))) {
1131 six_unlock_read(&b->c.lock);
1136 EBUG_ON(b->c.btree_id != btree_id);
1137 EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
1138 btree_check_header(c, b);
1140 bch2_btree_cache_cannibalize_unlock(c);
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)
1149 struct bch_fs *c = trans->c;
1150 struct btree_cache *bc = &c->btree_cache;
1153 BUG_ON(trans && !btree_node_locked(path, level + 1));
1154 BUG_ON(level >= BTREE_MAX_DEPTH);
1156 b = btree_cache_find(bc, k);
1160 b = bch2_btree_node_fill(trans, path, k, btree_id,
1161 level, SIX_LOCK_read, false);
1162 return PTR_ERR_OR_ZERO(b);
1165 void bch2_btree_node_evict(struct btree_trans *trans, const struct bkey_i *k)
1167 struct bch_fs *c = trans->c;
1168 struct btree_cache *bc = &c->btree_cache;
1171 b = btree_cache_find(bc, k);
1175 /* not allowed to wait on io with btree locks held: */
1177 /* XXX we're called from btree_gc which will be holding other btree
1180 __bch2_btree_node_wait_on_read(b);
1181 __bch2_btree_node_wait_on_write(b);
1183 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
1184 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
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);
1193 BUG_ON(btree_node_dirty(b));
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);
1200 six_unlock_write(&b->c.lock);
1201 six_unlock_intent(&b->c.lock);
1204 const char *bch2_btree_id_str(enum btree_id btree)
1206 return btree < BTREE_ID_NR ? __bch2_btree_ids[btree] : "(unknown)";
1209 void bch2_btree_pos_to_text(struct printbuf *out, struct bch_fs *c, const struct btree *b)
1211 prt_printf(out, "%s level %u/%u\n ",
1212 bch2_btree_id_str(b->c.btree_id),
1214 bch2_btree_id_root(c, b->c.btree_id)->level);
1215 bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(&b->key));
1218 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c, const struct btree *b)
1220 struct bset_stats stats;
1222 memset(&stats, 0, sizeof(stats));
1224 bch2_btree_keys_stats(b, &stats);
1226 prt_printf(out, "l %u ", b->c.level);
1227 bch2_bpos_to_text(out, b->data->min_key);
1228 prt_printf(out, " - ");
1229 bch2_bpos_to_text(out, b->data->max_key);
1230 prt_printf(out, ":\n"
1232 bch2_val_to_text(out, c, bkey_i_to_s_c(&b->key));
1237 bch2_bkey_format_to_text(out, &b->format);
1240 " unpack fn len: %u\n"
1241 " bytes used %zu/%zu (%zu%% full)\n"
1242 " sib u64s: %u, %u (merge threshold %u)\n"
1243 " nr packed keys %u\n"
1244 " nr unpacked keys %u\n"
1246 " failed unpacked %zu\n",
1248 b->nr.live_u64s * sizeof(u64),
1249 btree_bytes(c) - sizeof(struct btree_node),
1250 b->nr.live_u64s * 100 / btree_max_u64s(c),
1253 c->btree_foreground_merge_threshold,
1255 b->nr.unpacked_keys,
1260 void bch2_btree_cache_to_text(struct printbuf *out, const struct btree_cache *bc)
1262 prt_printf(out, "nr nodes:\t\t%u\n", bc->used);
1263 prt_printf(out, "nr dirty:\t\t%u\n", atomic_read(&bc->dirty));
1264 prt_printf(out, "cannibalize lock:\t%p\n", bc->alloc_lock);
1266 prt_printf(out, "freed:\t\t\t\t%u\n", bc->freed);
1267 prt_printf(out, "not freed, dirty:\t\t%u\n", bc->not_freed_dirty);
1268 prt_printf(out, "not freed, write in flight:\t%u\n", bc->not_freed_write_in_flight);
1269 prt_printf(out, "not freed, read in flight:\t%u\n", bc->not_freed_read_in_flight);
1270 prt_printf(out, "not freed, lock intent failed:\t%u\n", bc->not_freed_lock_intent);
1271 prt_printf(out, "not freed, lock write failed:\t%u\n", bc->not_freed_lock_write);
1272 prt_printf(out, "not freed, access bit:\t\t%u\n", bc->not_freed_access_bit);
1273 prt_printf(out, "not freed, no evict failed:\t%u\n", bc->not_freed_noevict);
1274 prt_printf(out, "not freed, write blocked:\t%u\n", bc->not_freed_write_blocked);
1275 prt_printf(out, "not freed, will make reachable:\t%u\n", bc->not_freed_will_make_reachable);