1 // SPDX-License-Identifier: GPL-2.0
4 #include "alloc_foreground.h"
5 #include "bkey_methods.h"
6 #include "btree_cache.h"
8 #include "btree_update.h"
9 #include "btree_update_interior.h"
11 #include "btree_iter.h"
12 #include "btree_locking.h"
16 #include "journal_reclaim.h"
21 #include <linux/random.h>
22 #include <trace/events/bcachefs.h>
27 * Verify that child nodes correctly span parent node's range:
29 static void btree_node_interior_verify(struct btree *b)
31 #ifdef CONFIG_BCACHEFS_DEBUG
32 struct bpos next_node = b->data->min_key;
33 struct btree_node_iter iter;
35 struct bkey_s_c_btree_ptr_v2 bp;
40 bch2_btree_node_iter_init_from_start(&iter, b);
43 k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked);
44 if (k.k->type != KEY_TYPE_btree_ptr_v2)
46 bp = bkey_s_c_to_btree_ptr_v2(k);
48 BUG_ON(bkey_cmp(next_node, bp.v->min_key));
50 bch2_btree_node_iter_advance(&iter, b);
52 if (bch2_btree_node_iter_end(&iter)) {
53 BUG_ON(bkey_cmp(k.k->p, b->key.k.p));
57 next_node = bkey_successor(k.k->p);
62 /* Calculate ideal packed bkey format for new btree nodes: */
64 void __bch2_btree_calc_format(struct bkey_format_state *s, struct btree *b)
66 struct bkey_packed *k;
70 bch2_bkey_format_add_pos(s, b->data->min_key);
73 bset_tree_for_each_key(b, t, k)
74 if (!bkey_whiteout(k)) {
75 uk = bkey_unpack_key(b, k);
76 bch2_bkey_format_add_key(s, &uk);
80 static struct bkey_format bch2_btree_calc_format(struct btree *b)
82 struct bkey_format_state s;
84 bch2_bkey_format_init(&s);
85 __bch2_btree_calc_format(&s, b);
87 return bch2_bkey_format_done(&s);
90 static size_t btree_node_u64s_with_format(struct btree *b,
91 struct bkey_format *new_f)
93 struct bkey_format *old_f = &b->format;
95 /* stupid integer promotion rules */
97 (((int) new_f->key_u64s - old_f->key_u64s) *
98 (int) b->nr.packed_keys) +
99 (((int) new_f->key_u64s - BKEY_U64s) *
100 (int) b->nr.unpacked_keys);
102 BUG_ON(delta + b->nr.live_u64s < 0);
104 return b->nr.live_u64s + delta;
108 * btree_node_format_fits - check if we could rewrite node with a new format
110 * This assumes all keys can pack with the new format -- it just checks if
111 * the re-packed keys would fit inside the node itself.
113 bool bch2_btree_node_format_fits(struct bch_fs *c, struct btree *b,
114 struct bkey_format *new_f)
116 size_t u64s = btree_node_u64s_with_format(b, new_f);
118 return __vstruct_bytes(struct btree_node, u64s) < btree_bytes(c);
121 /* Btree node freeing/allocation: */
123 static void __btree_node_free(struct bch_fs *c, struct btree *b)
125 trace_btree_node_free(c, b);
127 BUG_ON(btree_node_dirty(b));
128 BUG_ON(btree_node_need_write(b));
129 BUG_ON(b == btree_node_root(c, b));
131 BUG_ON(!list_empty(&b->write_blocked));
132 BUG_ON(b->will_make_reachable);
134 clear_btree_node_noevict(b);
136 bch2_btree_node_hash_remove(&c->btree_cache, b);
138 six_lock_wakeup_all(&b->c.lock);
140 mutex_lock(&c->btree_cache.lock);
141 list_move(&b->list, &c->btree_cache.freeable);
142 mutex_unlock(&c->btree_cache.lock);
145 void bch2_btree_node_free_never_inserted(struct bch_fs *c, struct btree *b)
147 struct open_buckets ob = b->ob;
151 clear_btree_node_dirty(b);
153 btree_node_lock_type(c, b, SIX_LOCK_write);
154 __btree_node_free(c, b);
155 six_unlock_write(&b->c.lock);
157 bch2_open_buckets_put(c, &ob);
160 void bch2_btree_node_free_inmem(struct bch_fs *c, struct btree *b,
161 struct btree_iter *iter)
163 struct btree_iter *linked;
165 trans_for_each_iter(iter->trans, linked)
166 BUG_ON(linked->l[b->c.level].b == b);
168 six_lock_write(&b->c.lock, NULL, NULL);
169 __btree_node_free(c, b);
170 six_unlock_write(&b->c.lock);
171 six_unlock_intent(&b->c.lock);
174 static struct btree *__bch2_btree_node_alloc(struct bch_fs *c,
175 struct disk_reservation *res,
179 struct write_point *wp;
182 struct open_buckets ob = { .nr = 0 };
183 struct bch_devs_list devs_have = (struct bch_devs_list) { 0 };
185 enum alloc_reserve alloc_reserve;
187 if (flags & BTREE_INSERT_USE_ALLOC_RESERVE) {
189 alloc_reserve = RESERVE_ALLOC;
190 } else if (flags & BTREE_INSERT_USE_RESERVE) {
191 nr_reserve = BTREE_NODE_RESERVE / 2;
192 alloc_reserve = RESERVE_BTREE;
194 nr_reserve = BTREE_NODE_RESERVE;
195 alloc_reserve = RESERVE_NONE;
198 mutex_lock(&c->btree_reserve_cache_lock);
199 if (c->btree_reserve_cache_nr > nr_reserve) {
200 struct btree_alloc *a =
201 &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
204 bkey_copy(&tmp.k, &a->k);
205 mutex_unlock(&c->btree_reserve_cache_lock);
208 mutex_unlock(&c->btree_reserve_cache_lock);
211 wp = bch2_alloc_sectors_start(c, c->opts.foreground_target, 0,
212 writepoint_ptr(&c->btree_write_point),
215 c->opts.metadata_replicas_required,
216 alloc_reserve, 0, cl);
220 if (wp->sectors_free < c->opts.btree_node_size) {
221 struct open_bucket *ob;
224 open_bucket_for_each(c, &wp->ptrs, ob, i)
225 if (ob->sectors_free < c->opts.btree_node_size)
226 ob->sectors_free = 0;
228 bch2_alloc_sectors_done(c, wp);
232 if (c->sb.features & (1ULL << BCH_FEATURE_btree_ptr_v2))
233 bkey_btree_ptr_v2_init(&tmp.k);
235 bkey_btree_ptr_init(&tmp.k);
237 bch2_alloc_sectors_append_ptrs(c, wp, &tmp.k, c->opts.btree_node_size);
239 bch2_open_bucket_get(c, wp, &ob);
240 bch2_alloc_sectors_done(c, wp);
242 b = bch2_btree_node_mem_alloc(c);
244 /* we hold cannibalize_lock: */
248 bkey_copy(&b->key, &tmp.k);
254 static struct btree *bch2_btree_node_alloc(struct btree_update *as, unsigned level)
256 struct bch_fs *c = as->c;
260 BUG_ON(level >= BTREE_MAX_DEPTH);
261 BUG_ON(!as->nr_prealloc_nodes);
263 b = as->prealloc_nodes[--as->nr_prealloc_nodes];
265 set_btree_node_accessed(b);
266 set_btree_node_dirty(b);
267 set_btree_node_need_write(b);
269 bch2_bset_init_first(b, &b->data->keys);
271 b->c.btree_id = as->btree_id;
273 memset(&b->nr, 0, sizeof(b->nr));
274 b->data->magic = cpu_to_le64(bset_magic(c));
276 SET_BTREE_NODE_ID(b->data, as->btree_id);
277 SET_BTREE_NODE_LEVEL(b->data, level);
278 b->data->ptr = bch2_bkey_ptrs_c(bkey_i_to_s_c(&b->key)).start->ptr;
280 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
281 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(&b->key);
284 bp->v.seq = b->data->keys.seq;
285 bp->v.sectors_written = 0;
286 bp->v.sectors = cpu_to_le16(c->opts.btree_node_size);
289 if (c->sb.features & (1ULL << BCH_FEATURE_new_extent_overwrite))
290 SET_BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data, true);
292 if (btree_node_is_extents(b) &&
293 !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data))
294 set_btree_node_old_extent_overwrite(b);
296 bch2_btree_build_aux_trees(b);
298 ret = bch2_btree_node_hash_insert(&c->btree_cache, b, level, as->btree_id);
301 trace_btree_node_alloc(c, b);
305 static void btree_set_min(struct btree *b, struct bpos pos)
307 if (b->key.k.type == KEY_TYPE_btree_ptr_v2)
308 bkey_i_to_btree_ptr_v2(&b->key)->v.min_key = pos;
309 b->data->min_key = pos;
312 static void btree_set_max(struct btree *b, struct bpos pos)
315 b->data->max_key = pos;
318 struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *as,
320 struct bkey_format format)
324 n = bch2_btree_node_alloc(as, b->c.level);
326 SET_BTREE_NODE_SEQ(n->data, BTREE_NODE_SEQ(b->data) + 1);
328 btree_set_min(n, b->data->min_key);
329 btree_set_max(n, b->data->max_key);
331 n->data->format = format;
332 btree_node_set_format(n, format);
334 bch2_btree_sort_into(as->c, n, b);
336 btree_node_reset_sib_u64s(n);
338 n->key.k.p = b->key.k.p;
342 static struct btree *bch2_btree_node_alloc_replacement(struct btree_update *as,
345 struct bkey_format new_f = bch2_btree_calc_format(b);
348 * The keys might expand with the new format - if they wouldn't fit in
349 * the btree node anymore, use the old format for now:
351 if (!bch2_btree_node_format_fits(as->c, b, &new_f))
354 return __bch2_btree_node_alloc_replacement(as, b, new_f);
357 static struct btree *__btree_root_alloc(struct btree_update *as, unsigned level)
359 struct btree *b = bch2_btree_node_alloc(as, level);
361 btree_set_min(b, POS_MIN);
362 btree_set_max(b, POS_MAX);
363 b->data->format = bch2_btree_calc_format(b);
365 btree_node_set_format(b, b->data->format);
366 bch2_btree_build_aux_trees(b);
368 bch2_btree_update_add_new_node(as, b);
369 six_unlock_write(&b->c.lock);
374 static void bch2_btree_reserve_put(struct btree_update *as)
376 struct bch_fs *c = as->c;
378 mutex_lock(&c->btree_reserve_cache_lock);
380 while (as->nr_prealloc_nodes) {
381 struct btree *b = as->prealloc_nodes[--as->nr_prealloc_nodes];
383 six_unlock_write(&b->c.lock);
385 if (c->btree_reserve_cache_nr <
386 ARRAY_SIZE(c->btree_reserve_cache)) {
387 struct btree_alloc *a =
388 &c->btree_reserve_cache[c->btree_reserve_cache_nr++];
392 bkey_copy(&a->k, &b->key);
394 bch2_open_buckets_put(c, &b->ob);
397 btree_node_lock_type(c, b, SIX_LOCK_write);
398 __btree_node_free(c, b);
399 six_unlock_write(&b->c.lock);
401 six_unlock_intent(&b->c.lock);
404 mutex_unlock(&c->btree_reserve_cache_lock);
407 static int bch2_btree_reserve_get(struct btree_update *as, unsigned nr_nodes,
408 unsigned flags, struct closure *cl)
410 struct bch_fs *c = as->c;
414 BUG_ON(nr_nodes > BTREE_RESERVE_MAX);
417 * Protects reaping from the btree node cache and using the btree node
418 * open bucket reserve:
420 ret = bch2_btree_cache_cannibalize_lock(c, cl);
424 while (as->nr_prealloc_nodes < nr_nodes) {
425 b = __bch2_btree_node_alloc(c, &as->disk_res,
426 flags & BTREE_INSERT_NOWAIT
433 ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(&b->key));
437 as->prealloc_nodes[as->nr_prealloc_nodes++] = b;
440 bch2_btree_cache_cannibalize_unlock(c);
443 bch2_btree_cache_cannibalize_unlock(c);
444 trace_btree_reserve_get_fail(c, nr_nodes, cl);
448 /* Asynchronous interior node update machinery */
450 static void bch2_btree_update_free(struct btree_update *as)
452 struct bch_fs *c = as->c;
454 bch2_journal_preres_put(&c->journal, &as->journal_preres);
456 bch2_journal_pin_drop(&c->journal, &as->journal);
457 bch2_journal_pin_flush(&c->journal, &as->journal);
458 bch2_disk_reservation_put(c, &as->disk_res);
459 bch2_btree_reserve_put(as);
461 mutex_lock(&c->btree_interior_update_lock);
462 list_del(&as->unwritten_list);
464 mutex_unlock(&c->btree_interior_update_lock);
466 closure_debug_destroy(&as->cl);
467 mempool_free(as, &c->btree_interior_update_pool);
469 closure_wake_up(&c->btree_interior_update_wait);
472 static void btree_update_will_delete_key(struct btree_update *as,
475 BUG_ON(bch2_keylist_u64s(&as->old_keys) + k->k.u64s >
476 ARRAY_SIZE(as->_old_keys));
477 bch2_keylist_add(&as->old_keys, k);
480 static void btree_update_will_add_key(struct btree_update *as,
483 BUG_ON(bch2_keylist_u64s(&as->new_keys) + k->k.u64s >
484 ARRAY_SIZE(as->_new_keys));
485 bch2_keylist_add(&as->new_keys, k);
489 * The transactional part of an interior btree node update, where we journal the
490 * update we did to the interior node and update alloc info:
492 static int btree_update_nodes_written_trans(struct btree_trans *trans,
493 struct btree_update *as)
498 trans->extra_journal_entries = (void *) &as->journal_entries[0];
499 trans->extra_journal_entry_u64s = as->journal_u64s;
500 trans->journal_pin = &as->journal;
502 for_each_keylist_key(&as->new_keys, k) {
503 ret = bch2_trans_mark_key(trans, bkey_i_to_s_c(k),
504 0, 0, BTREE_TRIGGER_INSERT);
509 for_each_keylist_key(&as->old_keys, k) {
510 ret = bch2_trans_mark_key(trans, bkey_i_to_s_c(k),
511 0, 0, BTREE_TRIGGER_OVERWRITE);
519 static void btree_update_nodes_written(struct btree_update *as)
521 struct bch_fs *c = as->c;
522 struct btree *b = as->b;
528 * We did an update to a parent node where the pointers we added pointed
529 * to child nodes that weren't written yet: now, the child nodes have
530 * been written so we can write out the update to the interior node.
534 * We can't call into journal reclaim here: we'd block on the journal
535 * reclaim lock, but we may need to release the open buckets we have
536 * pinned in order for other btree updates to make forward progress, and
537 * journal reclaim does btree updates when flushing bkey_cached entries,
538 * which may require allocations as well.
540 ret = bch2_trans_do(c, &as->disk_res, &journal_seq,
542 BTREE_INSERT_USE_RESERVE|
543 BTREE_INSERT_USE_ALLOC_RESERVE|
544 BTREE_INSERT_NOCHECK_RW|
545 BTREE_INSERT_JOURNAL_RECLAIM|
546 BTREE_INSERT_JOURNAL_RESERVED,
547 btree_update_nodes_written_trans(&trans, as));
548 BUG_ON(ret && !bch2_journal_error(&c->journal));
552 * @b is the node we did the final insert into:
554 * On failure to get a journal reservation, we still have to
555 * unblock the write and allow most of the write path to happen
556 * so that shutdown works, but the i->journal_seq mechanism
557 * won't work to prevent the btree write from being visible (we
558 * didn't get a journal sequence number) - instead
559 * __bch2_btree_node_write() doesn't do the actual write if
560 * we're in journal error state:
563 btree_node_lock_type(c, b, SIX_LOCK_intent);
564 btree_node_lock_type(c, b, SIX_LOCK_write);
565 mutex_lock(&c->btree_interior_update_lock);
567 list_del(&as->write_blocked_list);
569 if (!ret && as->b == b) {
570 struct bset *i = btree_bset_last(b);
573 BUG_ON(!btree_node_dirty(b));
575 i->journal_seq = cpu_to_le64(
577 le64_to_cpu(i->journal_seq)));
579 bch2_btree_add_journal_pin(c, b, journal_seq);
582 mutex_unlock(&c->btree_interior_update_lock);
583 six_unlock_write(&b->c.lock);
585 btree_node_write_if_need(c, b, SIX_LOCK_intent);
586 six_unlock_intent(&b->c.lock);
589 bch2_journal_pin_drop(&c->journal, &as->journal);
591 bch2_journal_preres_put(&c->journal, &as->journal_preres);
593 mutex_lock(&c->btree_interior_update_lock);
594 for (i = 0; i < as->nr_new_nodes; i++) {
595 b = as->new_nodes[i];
597 BUG_ON(b->will_make_reachable != (unsigned long) as);
598 b->will_make_reachable = 0;
600 mutex_unlock(&c->btree_interior_update_lock);
602 for (i = 0; i < as->nr_new_nodes; i++) {
603 b = as->new_nodes[i];
605 btree_node_lock_type(c, b, SIX_LOCK_read);
606 btree_node_write_if_need(c, b, SIX_LOCK_read);
607 six_unlock_read(&b->c.lock);
610 for (i = 0; i < as->nr_open_buckets; i++)
611 bch2_open_bucket_put(c, c->open_buckets + as->open_buckets[i]);
613 bch2_btree_update_free(as);
616 static void btree_interior_update_work(struct work_struct *work)
619 container_of(work, struct bch_fs, btree_interior_update_work);
620 struct btree_update *as;
623 mutex_lock(&c->btree_interior_update_lock);
624 as = list_first_entry_or_null(&c->btree_interior_updates_unwritten,
625 struct btree_update, unwritten_list);
626 if (as && !as->nodes_written)
628 mutex_unlock(&c->btree_interior_update_lock);
633 btree_update_nodes_written(as);
637 static void btree_update_set_nodes_written(struct closure *cl)
639 struct btree_update *as = container_of(cl, struct btree_update, cl);
640 struct bch_fs *c = as->c;
642 mutex_lock(&c->btree_interior_update_lock);
643 as->nodes_written = true;
644 mutex_unlock(&c->btree_interior_update_lock);
646 queue_work(c->btree_interior_update_worker, &c->btree_interior_update_work);
650 * We're updating @b with pointers to nodes that haven't finished writing yet:
651 * block @b from being written until @as completes
653 static void btree_update_updated_node(struct btree_update *as, struct btree *b)
655 struct bch_fs *c = as->c;
657 mutex_lock(&c->btree_interior_update_lock);
658 list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
660 BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
661 BUG_ON(!btree_node_dirty(b));
663 as->mode = BTREE_INTERIOR_UPDATING_NODE;
665 list_add(&as->write_blocked_list, &b->write_blocked);
667 mutex_unlock(&c->btree_interior_update_lock);
670 static void btree_update_reparent(struct btree_update *as,
671 struct btree_update *child)
673 struct bch_fs *c = as->c;
675 lockdep_assert_held(&c->btree_interior_update_lock);
678 child->mode = BTREE_INTERIOR_UPDATING_AS;
681 * When we write a new btree root, we have to drop our journal pin
682 * _before_ the new nodes are technically reachable; see
683 * btree_update_nodes_written().
685 * This goes for journal pins that are recursively blocked on us - so,
686 * just transfer the journal pin to the new interior update so
687 * btree_update_nodes_written() can drop it.
689 bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal, NULL);
690 bch2_journal_pin_drop(&c->journal, &child->journal);
693 static void btree_update_updated_root(struct btree_update *as, struct btree *b)
695 struct bkey_i *insert = &b->key;
696 struct bch_fs *c = as->c;
698 BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
700 BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
701 ARRAY_SIZE(as->journal_entries));
704 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
705 BCH_JSET_ENTRY_btree_root,
706 b->c.btree_id, b->c.level,
707 insert, insert->k.u64s);
709 mutex_lock(&c->btree_interior_update_lock);
710 list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
712 as->mode = BTREE_INTERIOR_UPDATING_ROOT;
713 mutex_unlock(&c->btree_interior_update_lock);
717 * bch2_btree_update_add_new_node:
719 * This causes @as to wait on @b to be written, before it gets to
720 * bch2_btree_update_nodes_written
722 * Additionally, it sets b->will_make_reachable to prevent any additional writes
723 * to @b from happening besides the first until @b is reachable on disk
725 * And it adds @b to the list of @as's new nodes, so that we can update sector
726 * counts in bch2_btree_update_nodes_written:
728 void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b)
730 struct bch_fs *c = as->c;
732 closure_get(&as->cl);
734 mutex_lock(&c->btree_interior_update_lock);
735 BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes));
736 BUG_ON(b->will_make_reachable);
738 as->new_nodes[as->nr_new_nodes++] = b;
739 b->will_make_reachable = 1UL|(unsigned long) as;
741 mutex_unlock(&c->btree_interior_update_lock);
743 btree_update_will_add_key(as, &b->key);
747 * returns true if @b was a new node
749 static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b)
751 struct btree_update *as;
755 mutex_lock(&c->btree_interior_update_lock);
757 * When b->will_make_reachable != 0, it owns a ref on as->cl that's
758 * dropped when it gets written by bch2_btree_complete_write - the
759 * xchg() is for synchronization with bch2_btree_complete_write:
761 v = xchg(&b->will_make_reachable, 0);
762 as = (struct btree_update *) (v & ~1UL);
765 mutex_unlock(&c->btree_interior_update_lock);
769 for (i = 0; i < as->nr_new_nodes; i++)
770 if (as->new_nodes[i] == b)
775 array_remove_item(as->new_nodes, as->nr_new_nodes, i);
776 mutex_unlock(&c->btree_interior_update_lock);
779 closure_put(&as->cl);
782 void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b)
785 as->open_buckets[as->nr_open_buckets++] =
790 * @b is being split/rewritten: it may have pointers to not-yet-written btree
791 * nodes and thus outstanding btree_updates - redirect @b's
792 * btree_updates to point to this btree_update:
794 void bch2_btree_interior_update_will_free_node(struct btree_update *as,
797 struct bch_fs *c = as->c;
798 struct btree_update *p, *n;
799 struct btree_write *w;
801 set_btree_node_dying(b);
803 if (btree_node_fake(b))
806 mutex_lock(&c->btree_interior_update_lock);
809 * Does this node have any btree_update operations preventing
810 * it from being written?
812 * If so, redirect them to point to this btree_update: we can
813 * write out our new nodes, but we won't make them visible until those
814 * operations complete
816 list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) {
817 list_del_init(&p->write_blocked_list);
818 btree_update_reparent(as, p);
821 * for flush_held_btree_writes() waiting on updates to flush or
822 * nodes to be writeable:
824 closure_wake_up(&c->btree_interior_update_wait);
827 clear_btree_node_dirty(b);
828 clear_btree_node_need_write(b);
831 * Does this node have unwritten data that has a pin on the journal?
833 * If so, transfer that pin to the btree_update operation -
834 * note that if we're freeing multiple nodes, we only need to keep the
835 * oldest pin of any of the nodes we're freeing. We'll release the pin
836 * when the new nodes are persistent and reachable on disk:
838 w = btree_current_write(b);
839 bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
840 bch2_journal_pin_drop(&c->journal, &w->journal);
842 w = btree_prev_write(b);
843 bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
844 bch2_journal_pin_drop(&c->journal, &w->journal);
846 mutex_unlock(&c->btree_interior_update_lock);
849 * Is this a node that isn't reachable on disk yet?
851 * Nodes that aren't reachable yet have writes blocked until they're
852 * reachable - now that we've cancelled any pending writes and moved
853 * things waiting on that write to wait on this update, we can drop this
854 * node from the list of nodes that the other update is making
855 * reachable, prior to freeing it:
857 btree_update_drop_new_node(c, b);
859 btree_update_will_delete_key(as, &b->key);
862 void bch2_btree_update_done(struct btree_update *as)
864 BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE);
866 bch2_btree_reserve_put(as);
868 continue_at(&as->cl, btree_update_set_nodes_written, system_freezable_wq);
871 struct btree_update *
872 bch2_btree_update_start(struct btree_trans *trans, enum btree_id id,
873 unsigned nr_nodes, unsigned flags,
876 struct bch_fs *c = trans->c;
877 struct btree_update *as;
878 int disk_res_flags = (flags & BTREE_INSERT_NOFAIL)
879 ? BCH_DISK_RESERVATION_NOFAIL : 0;
880 int journal_flags = (flags & BTREE_INSERT_JOURNAL_RESERVED)
881 ? JOURNAL_RES_GET_RECLAIM : 0;
885 * This check isn't necessary for correctness - it's just to potentially
886 * prevent us from doing a lot of work that'll end up being wasted:
888 ret = bch2_journal_error(&c->journal);
892 as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOIO);
893 memset(as, 0, sizeof(*as));
894 closure_init(&as->cl, NULL);
896 as->mode = BTREE_INTERIOR_NO_UPDATE;
898 INIT_LIST_HEAD(&as->list);
899 INIT_LIST_HEAD(&as->unwritten_list);
900 INIT_LIST_HEAD(&as->write_blocked_list);
901 bch2_keylist_init(&as->old_keys, as->_old_keys);
902 bch2_keylist_init(&as->new_keys, as->_new_keys);
903 bch2_keylist_init(&as->parent_keys, as->inline_keys);
905 ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
906 BTREE_UPDATE_JOURNAL_RES,
907 journal_flags|JOURNAL_RES_GET_NONBLOCK);
908 if (ret == -EAGAIN) {
909 if (flags & BTREE_INSERT_NOUNLOCK)
910 return ERR_PTR(-EINTR);
912 bch2_trans_unlock(trans);
914 ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
915 BTREE_UPDATE_JOURNAL_RES,
920 if (!bch2_trans_relock(trans)) {
926 ret = bch2_disk_reservation_get(c, &as->disk_res,
927 nr_nodes * c->opts.btree_node_size,
928 c->opts.metadata_replicas,
933 ret = bch2_btree_reserve_get(as, nr_nodes, flags, cl);
937 mutex_lock(&c->btree_interior_update_lock);
938 list_add_tail(&as->list, &c->btree_interior_update_list);
939 mutex_unlock(&c->btree_interior_update_lock);
943 bch2_btree_update_free(as);
947 /* Btree root updates: */
949 static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
951 /* Root nodes cannot be reaped */
952 mutex_lock(&c->btree_cache.lock);
953 list_del_init(&b->list);
954 mutex_unlock(&c->btree_cache.lock);
956 mutex_lock(&c->btree_root_lock);
957 BUG_ON(btree_node_root(c, b) &&
958 (b->c.level < btree_node_root(c, b)->c.level ||
959 !btree_node_dying(btree_node_root(c, b))));
961 btree_node_root(c, b) = b;
962 mutex_unlock(&c->btree_root_lock);
964 bch2_recalc_btree_reserve(c);
968 * bch_btree_set_root - update the root in memory and on disk
970 * To ensure forward progress, the current task must not be holding any
971 * btree node write locks. However, you must hold an intent lock on the
974 * Note: This allocates a journal entry but doesn't add any keys to
975 * it. All the btree roots are part of every journal write, so there
976 * is nothing new to be done. This just guarantees that there is a
979 static void bch2_btree_set_root(struct btree_update *as, struct btree *b,
980 struct btree_iter *iter)
982 struct bch_fs *c = as->c;
985 trace_btree_set_root(c, b);
986 BUG_ON(!b->written &&
987 !test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags));
989 old = btree_node_root(c, b);
992 * Ensure no one is using the old root while we switch to the
995 bch2_btree_node_lock_write(old, iter);
997 bch2_btree_set_root_inmem(c, b);
999 btree_update_updated_root(as, b);
1002 * Unlock old root after new root is visible:
1004 * The new root isn't persistent, but that's ok: we still have
1005 * an intent lock on the new root, and any updates that would
1006 * depend on the new root would have to update the new root.
1008 bch2_btree_node_unlock_write(old, iter);
1011 /* Interior node updates: */
1013 static void bch2_insert_fixup_btree_ptr(struct btree_update *as, struct btree *b,
1014 struct btree_iter *iter,
1015 struct bkey_i *insert,
1016 struct btree_node_iter *node_iter)
1018 struct bkey_packed *k;
1020 BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1021 ARRAY_SIZE(as->journal_entries));
1024 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
1025 BCH_JSET_ENTRY_btree_keys,
1026 b->c.btree_id, b->c.level,
1027 insert, insert->k.u64s);
1029 while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1030 bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1031 bch2_btree_node_iter_advance(node_iter, b);
1033 bch2_btree_bset_insert_key(iter, b, node_iter, insert);
1034 set_btree_node_dirty(b);
1035 set_btree_node_need_write(b);
1039 * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1042 static struct btree *__btree_split_node(struct btree_update *as,
1044 struct btree_iter *iter)
1046 size_t nr_packed = 0, nr_unpacked = 0;
1048 struct bset *set1, *set2;
1049 struct bkey_packed *k, *prev = NULL;
1051 n2 = bch2_btree_node_alloc(as, n1->c.level);
1052 bch2_btree_update_add_new_node(as, n2);
1054 n2->data->max_key = n1->data->max_key;
1055 n2->data->format = n1->format;
1056 SET_BTREE_NODE_SEQ(n2->data, BTREE_NODE_SEQ(n1->data));
1057 n2->key.k.p = n1->key.k.p;
1059 btree_node_set_format(n2, n2->data->format);
1061 set1 = btree_bset_first(n1);
1062 set2 = btree_bset_first(n2);
1065 * Has to be a linear search because we don't have an auxiliary
1070 struct bkey_packed *n = bkey_next_skip_noops(k, vstruct_last(set1));
1072 if (n == vstruct_last(set1))
1074 if (k->_data - set1->_data >= (le16_to_cpu(set1->u64s) * 3) / 5)
1088 btree_set_max(n1, bkey_unpack_pos(n1, prev));
1089 btree_set_min(n2, bkey_successor(n1->key.k.p));
1091 set2->u64s = cpu_to_le16((u64 *) vstruct_end(set1) - (u64 *) k);
1092 set1->u64s = cpu_to_le16(le16_to_cpu(set1->u64s) - le16_to_cpu(set2->u64s));
1094 set_btree_bset_end(n1, n1->set);
1095 set_btree_bset_end(n2, n2->set);
1097 n2->nr.live_u64s = le16_to_cpu(set2->u64s);
1098 n2->nr.bset_u64s[0] = le16_to_cpu(set2->u64s);
1099 n2->nr.packed_keys = n1->nr.packed_keys - nr_packed;
1100 n2->nr.unpacked_keys = n1->nr.unpacked_keys - nr_unpacked;
1102 n1->nr.live_u64s = le16_to_cpu(set1->u64s);
1103 n1->nr.bset_u64s[0] = le16_to_cpu(set1->u64s);
1104 n1->nr.packed_keys = nr_packed;
1105 n1->nr.unpacked_keys = nr_unpacked;
1107 BUG_ON(!set1->u64s);
1108 BUG_ON(!set2->u64s);
1110 memcpy_u64s(set2->start,
1112 le16_to_cpu(set2->u64s));
1114 btree_node_reset_sib_u64s(n1);
1115 btree_node_reset_sib_u64s(n2);
1117 bch2_verify_btree_nr_keys(n1);
1118 bch2_verify_btree_nr_keys(n2);
1121 btree_node_interior_verify(n1);
1122 btree_node_interior_verify(n2);
1129 * For updates to interior nodes, we've got to do the insert before we split
1130 * because the stuff we're inserting has to be inserted atomically. Post split,
1131 * the keys might have to go in different nodes and the split would no longer be
1134 * Worse, if the insert is from btree node coalescing, if we do the insert after
1135 * we do the split (and pick the pivot) - the pivot we pick might be between
1136 * nodes that were coalesced, and thus in the middle of a child node post
1139 static void btree_split_insert_keys(struct btree_update *as, struct btree *b,
1140 struct btree_iter *iter,
1141 struct keylist *keys)
1143 struct btree_node_iter node_iter;
1144 struct bkey_i *k = bch2_keylist_front(keys);
1145 struct bkey_packed *src, *dst, *n;
1148 BUG_ON(btree_node_type(b) != BKEY_TYPE_BTREE);
1150 bch2_btree_node_iter_init(&node_iter, b, &k->k.p);
1152 while (!bch2_keylist_empty(keys)) {
1153 k = bch2_keylist_front(keys);
1155 bch2_insert_fixup_btree_ptr(as, b, iter, k, &node_iter);
1156 bch2_keylist_pop_front(keys);
1160 * We can't tolerate whiteouts here - with whiteouts there can be
1161 * duplicate keys, and it would be rather bad if we picked a duplicate
1164 i = btree_bset_first(b);
1165 src = dst = i->start;
1166 while (src != vstruct_last(i)) {
1167 n = bkey_next_skip_noops(src, vstruct_last(i));
1168 if (!bkey_deleted(src)) {
1169 memmove_u64s_down(dst, src, src->u64s);
1170 dst = bkey_next(dst);
1175 i->u64s = cpu_to_le16((u64 *) dst - i->_data);
1176 set_btree_bset_end(b, b->set);
1178 BUG_ON(b->nsets != 1 ||
1179 b->nr.live_u64s != le16_to_cpu(btree_bset_first(b)->u64s));
1181 btree_node_interior_verify(b);
1184 static void btree_split(struct btree_update *as, struct btree *b,
1185 struct btree_iter *iter, struct keylist *keys,
1188 struct bch_fs *c = as->c;
1189 struct btree *parent = btree_node_parent(iter, b);
1190 struct btree *n1, *n2 = NULL, *n3 = NULL;
1191 u64 start_time = local_clock();
1193 BUG_ON(!parent && (b != btree_node_root(c, b)));
1194 BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->c.level));
1196 bch2_btree_interior_update_will_free_node(as, b);
1198 n1 = bch2_btree_node_alloc_replacement(as, b);
1199 bch2_btree_update_add_new_node(as, n1);
1202 btree_split_insert_keys(as, n1, iter, keys);
1204 if (bset_u64s(&n1->set[0]) > BTREE_SPLIT_THRESHOLD(c)) {
1205 trace_btree_split(c, b);
1207 n2 = __btree_split_node(as, n1, iter);
1209 bch2_btree_build_aux_trees(n2);
1210 bch2_btree_build_aux_trees(n1);
1211 six_unlock_write(&n2->c.lock);
1212 six_unlock_write(&n1->c.lock);
1214 bch2_btree_node_write(c, n2, SIX_LOCK_intent);
1217 * Note that on recursive parent_keys == keys, so we
1218 * can't start adding new keys to parent_keys before emptying it
1219 * out (which we did with btree_split_insert_keys() above)
1221 bch2_keylist_add(&as->parent_keys, &n1->key);
1222 bch2_keylist_add(&as->parent_keys, &n2->key);
1225 /* Depth increases, make a new root */
1226 n3 = __btree_root_alloc(as, b->c.level + 1);
1228 n3->sib_u64s[0] = U16_MAX;
1229 n3->sib_u64s[1] = U16_MAX;
1231 btree_split_insert_keys(as, n3, iter, &as->parent_keys);
1233 bch2_btree_node_write(c, n3, SIX_LOCK_intent);
1236 trace_btree_compact(c, b);
1238 bch2_btree_build_aux_trees(n1);
1239 six_unlock_write(&n1->c.lock);
1242 bch2_keylist_add(&as->parent_keys, &n1->key);
1245 bch2_btree_node_write(c, n1, SIX_LOCK_intent);
1247 /* New nodes all written, now make them visible: */
1250 /* Split a non root node */
1251 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1253 bch2_btree_set_root(as, n3, iter);
1255 /* Root filled up but didn't need to be split */
1256 bch2_btree_set_root(as, n1, iter);
1259 bch2_btree_update_get_open_buckets(as, n1);
1261 bch2_btree_update_get_open_buckets(as, n2);
1263 bch2_btree_update_get_open_buckets(as, n3);
1265 /* Successful split, update the iterator to point to the new nodes: */
1267 six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1268 bch2_btree_iter_node_drop(iter, b);
1270 bch2_btree_iter_node_replace(iter, n3);
1272 bch2_btree_iter_node_replace(iter, n2);
1273 bch2_btree_iter_node_replace(iter, n1);
1276 * The old node must be freed (in memory) _before_ unlocking the new
1277 * nodes - else another thread could re-acquire a read lock on the old
1278 * node after another thread has locked and updated the new node, thus
1279 * seeing stale data:
1281 bch2_btree_node_free_inmem(c, b, iter);
1284 six_unlock_intent(&n3->c.lock);
1286 six_unlock_intent(&n2->c.lock);
1287 six_unlock_intent(&n1->c.lock);
1289 bch2_btree_trans_verify_locks(iter->trans);
1291 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_split],
1296 bch2_btree_insert_keys_interior(struct btree_update *as, struct btree *b,
1297 struct btree_iter *iter, struct keylist *keys)
1299 struct btree_iter *linked;
1300 struct btree_node_iter node_iter;
1301 struct bkey_i *insert = bch2_keylist_front(keys);
1302 struct bkey_packed *k;
1304 /* Don't screw up @iter's position: */
1305 node_iter = iter->l[b->c.level].iter;
1308 * btree_split(), btree_gc_coalesce() will insert keys before
1309 * the iterator's current position - they know the keys go in
1310 * the node the iterator points to:
1312 while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1313 (bkey_cmp_packed(b, k, &insert->k) >= 0))
1316 for_each_keylist_key(keys, insert)
1317 bch2_insert_fixup_btree_ptr(as, b, iter, insert, &node_iter);
1319 btree_update_updated_node(as, b);
1321 trans_for_each_iter_with_node(iter->trans, b, linked)
1322 bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b);
1324 bch2_btree_trans_verify_iters(iter->trans, b);
1328 * bch_btree_insert_node - insert bkeys into a given btree node
1330 * @iter: btree iterator
1331 * @keys: list of keys to insert
1332 * @hook: insert callback
1333 * @persistent: if not null, @persistent will wait on journal write
1335 * Inserts as many keys as it can into a given btree node, splitting it if full.
1336 * If a split occurred, this function will return early. This can only happen
1337 * for leaf nodes -- inserts into interior nodes have to be atomic.
1339 void bch2_btree_insert_node(struct btree_update *as, struct btree *b,
1340 struct btree_iter *iter, struct keylist *keys,
1343 struct bch_fs *c = as->c;
1344 int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1345 int old_live_u64s = b->nr.live_u64s;
1346 int live_u64s_added, u64s_added;
1348 BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->c.level));
1349 BUG_ON(!b->c.level);
1350 BUG_ON(!as || as->b);
1351 bch2_verify_keylist_sorted(keys);
1353 if (as->must_rewrite)
1356 bch2_btree_node_lock_for_insert(c, b, iter);
1358 if (!bch2_btree_node_insert_fits(c, b, bch2_keylist_u64s(keys))) {
1359 bch2_btree_node_unlock_write(b, iter);
1363 bch2_btree_insert_keys_interior(as, b, iter, keys);
1365 live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1366 u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1368 if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1369 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1370 if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1371 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1373 if (u64s_added > live_u64s_added &&
1374 bch2_maybe_compact_whiteouts(c, b))
1375 bch2_btree_iter_reinit_node(iter, b);
1377 bch2_btree_node_unlock_write(b, iter);
1379 btree_node_interior_verify(b);
1382 * when called from the btree_split path the new nodes aren't added to
1383 * the btree iterator yet, so the merge path's unlock/wait/relock dance
1386 bch2_foreground_maybe_merge(c, iter, b->c.level,
1387 flags|BTREE_INSERT_NOUNLOCK);
1390 btree_split(as, b, iter, keys, flags);
1393 int bch2_btree_split_leaf(struct bch_fs *c, struct btree_iter *iter,
1396 struct btree_trans *trans = iter->trans;
1397 struct btree *b = iter_l(iter)->b;
1398 struct btree_update *as;
1401 struct btree_insert_entry *i;
1404 * We already have a disk reservation and open buckets pinned; this
1405 * allocation must not block:
1407 trans_for_each_update(trans, i)
1408 if (btree_node_type_needs_gc(i->iter->btree_id))
1409 flags |= BTREE_INSERT_USE_RESERVE;
1411 closure_init_stack(&cl);
1413 /* Hack, because gc and splitting nodes doesn't mix yet: */
1414 if (!(flags & BTREE_INSERT_GC_LOCK_HELD) &&
1415 !down_read_trylock(&c->gc_lock)) {
1416 if (flags & BTREE_INSERT_NOUNLOCK) {
1417 trace_transaction_restart_ip(trans->ip, _THIS_IP_);
1421 bch2_trans_unlock(trans);
1422 down_read(&c->gc_lock);
1424 if (!bch2_trans_relock(trans))
1429 * XXX: figure out how far we might need to split,
1430 * instead of locking/reserving all the way to the root:
1432 if (!bch2_btree_iter_upgrade(iter, U8_MAX)) {
1433 trace_trans_restart_iter_upgrade(trans->ip);
1438 as = bch2_btree_update_start(trans, iter->btree_id,
1439 btree_update_reserve_required(c, b), flags,
1440 !(flags & BTREE_INSERT_NOUNLOCK) ? &cl : NULL);
1443 if (ret == -EAGAIN) {
1444 BUG_ON(flags & BTREE_INSERT_NOUNLOCK);
1445 bch2_trans_unlock(trans);
1448 trace_transaction_restart_ip(trans->ip, _THIS_IP_);
1453 btree_split(as, b, iter, NULL, flags);
1454 bch2_btree_update_done(as);
1457 * We haven't successfully inserted yet, so don't downgrade all the way
1458 * back to read locks;
1460 __bch2_btree_iter_downgrade(iter, 1);
1462 if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1463 up_read(&c->gc_lock);
1468 void __bch2_foreground_maybe_merge(struct bch_fs *c,
1469 struct btree_iter *iter,
1472 enum btree_node_sibling sib)
1474 struct btree_trans *trans = iter->trans;
1475 struct btree_update *as;
1476 struct bkey_format_state new_s;
1477 struct bkey_format new_f;
1478 struct bkey_i delete;
1479 struct btree *b, *m, *n, *prev, *next, *parent;
1484 BUG_ON(!btree_node_locked(iter, level));
1486 closure_init_stack(&cl);
1488 BUG_ON(!btree_node_locked(iter, level));
1490 b = iter->l[level].b;
1492 parent = btree_node_parent(iter, b);
1496 if (b->sib_u64s[sib] > BTREE_FOREGROUND_MERGE_THRESHOLD(c))
1499 /* XXX: can't be holding read locks */
1500 m = bch2_btree_node_get_sibling(c, iter, b, sib);
1506 /* NULL means no sibling: */
1508 b->sib_u64s[sib] = U16_MAX;
1512 if (sib == btree_prev_sib) {
1520 bch2_bkey_format_init(&new_s);
1521 __bch2_btree_calc_format(&new_s, b);
1522 __bch2_btree_calc_format(&new_s, m);
1523 new_f = bch2_bkey_format_done(&new_s);
1525 sib_u64s = btree_node_u64s_with_format(b, &new_f) +
1526 btree_node_u64s_with_format(m, &new_f);
1528 if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1529 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1531 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1534 sib_u64s = min(sib_u64s, btree_max_u64s(c));
1535 b->sib_u64s[sib] = sib_u64s;
1537 if (b->sib_u64s[sib] > BTREE_FOREGROUND_MERGE_THRESHOLD(c)) {
1538 six_unlock_intent(&m->c.lock);
1542 /* We're changing btree topology, doesn't mix with gc: */
1543 if (!(flags & BTREE_INSERT_GC_LOCK_HELD) &&
1544 !down_read_trylock(&c->gc_lock))
1545 goto err_cycle_gc_lock;
1547 if (!bch2_btree_iter_upgrade(iter, U8_MAX)) {
1552 as = bch2_btree_update_start(trans, iter->btree_id,
1553 btree_update_reserve_required(c, parent) + 1,
1555 BTREE_INSERT_NOFAIL|
1556 BTREE_INSERT_USE_RESERVE,
1557 !(flags & BTREE_INSERT_NOUNLOCK) ? &cl : NULL);
1563 trace_btree_merge(c, b);
1565 bch2_btree_interior_update_will_free_node(as, b);
1566 bch2_btree_interior_update_will_free_node(as, m);
1568 n = bch2_btree_node_alloc(as, b->c.level);
1569 bch2_btree_update_add_new_node(as, n);
1571 btree_set_min(n, prev->data->min_key);
1572 btree_set_max(n, next->data->max_key);
1573 n->data->format = new_f;
1575 btree_node_set_format(n, new_f);
1577 bch2_btree_sort_into(c, n, prev);
1578 bch2_btree_sort_into(c, n, next);
1580 bch2_btree_build_aux_trees(n);
1581 six_unlock_write(&n->c.lock);
1583 bkey_init(&delete.k);
1584 delete.k.p = prev->key.k.p;
1585 bch2_keylist_add(&as->parent_keys, &delete);
1586 bch2_keylist_add(&as->parent_keys, &n->key);
1588 bch2_btree_node_write(c, n, SIX_LOCK_intent);
1590 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1592 bch2_btree_update_get_open_buckets(as, n);
1594 six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1595 bch2_btree_iter_node_drop(iter, b);
1596 bch2_btree_iter_node_drop(iter, m);
1598 bch2_btree_iter_node_replace(iter, n);
1600 bch2_btree_trans_verify_iters(trans, n);
1602 bch2_btree_node_free_inmem(c, b, iter);
1603 bch2_btree_node_free_inmem(c, m, iter);
1605 six_unlock_intent(&n->c.lock);
1607 bch2_btree_update_done(as);
1609 if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1610 up_read(&c->gc_lock);
1612 bch2_btree_trans_verify_locks(trans);
1615 * Don't downgrade locks here: we're called after successful insert,
1616 * and the caller will downgrade locks after a successful insert
1617 * anyways (in case e.g. a split was required first)
1619 * And we're also called when inserting into interior nodes in the
1620 * split path, and downgrading to read locks in there is potentially
1627 six_unlock_intent(&m->c.lock);
1629 if (flags & BTREE_INSERT_NOUNLOCK)
1632 bch2_trans_unlock(trans);
1634 down_read(&c->gc_lock);
1635 up_read(&c->gc_lock);
1640 six_unlock_intent(&m->c.lock);
1641 if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1642 up_read(&c->gc_lock);
1644 BUG_ON(ret == -EAGAIN && (flags & BTREE_INSERT_NOUNLOCK));
1646 if ((ret == -EAGAIN || ret == -EINTR) &&
1647 !(flags & BTREE_INSERT_NOUNLOCK)) {
1648 bch2_trans_unlock(trans);
1650 ret = bch2_btree_iter_traverse(iter);
1660 static int __btree_node_rewrite(struct bch_fs *c, struct btree_iter *iter,
1661 struct btree *b, unsigned flags,
1664 struct btree *n, *parent = btree_node_parent(iter, b);
1665 struct btree_update *as;
1667 as = bch2_btree_update_start(iter->trans, iter->btree_id,
1669 ? btree_update_reserve_required(c, parent)
1673 trace_btree_gc_rewrite_node_fail(c, b);
1677 bch2_btree_interior_update_will_free_node(as, b);
1679 n = bch2_btree_node_alloc_replacement(as, b);
1680 bch2_btree_update_add_new_node(as, n);
1682 bch2_btree_build_aux_trees(n);
1683 six_unlock_write(&n->c.lock);
1685 trace_btree_gc_rewrite_node(c, b);
1687 bch2_btree_node_write(c, n, SIX_LOCK_intent);
1690 bch2_keylist_add(&as->parent_keys, &n->key);
1691 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1693 bch2_btree_set_root(as, n, iter);
1696 bch2_btree_update_get_open_buckets(as, n);
1698 six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1699 bch2_btree_iter_node_drop(iter, b);
1700 bch2_btree_iter_node_replace(iter, n);
1701 bch2_btree_node_free_inmem(c, b, iter);
1702 six_unlock_intent(&n->c.lock);
1704 bch2_btree_update_done(as);
1709 * bch_btree_node_rewrite - Rewrite/move a btree node
1711 * Returns 0 on success, -EINTR or -EAGAIN on failure (i.e.
1712 * btree_check_reserve() has to wait)
1714 int bch2_btree_node_rewrite(struct bch_fs *c, struct btree_iter *iter,
1715 __le64 seq, unsigned flags)
1717 struct btree_trans *trans = iter->trans;
1722 flags |= BTREE_INSERT_NOFAIL;
1724 closure_init_stack(&cl);
1726 bch2_btree_iter_upgrade(iter, U8_MAX);
1728 if (!(flags & BTREE_INSERT_GC_LOCK_HELD)) {
1729 if (!down_read_trylock(&c->gc_lock)) {
1730 bch2_trans_unlock(trans);
1731 down_read(&c->gc_lock);
1736 ret = bch2_btree_iter_traverse(iter);
1740 b = bch2_btree_iter_peek_node(iter);
1741 if (!b || b->data->keys.seq != seq)
1744 ret = __btree_node_rewrite(c, iter, b, flags, &cl);
1745 if (ret != -EAGAIN &&
1749 bch2_trans_unlock(trans);
1753 bch2_btree_iter_downgrade(iter);
1755 if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1756 up_read(&c->gc_lock);
1762 static void __bch2_btree_node_update_key(struct bch_fs *c,
1763 struct btree_update *as,
1764 struct btree_iter *iter,
1765 struct btree *b, struct btree *new_hash,
1766 struct bkey_i *new_key)
1768 struct btree *parent;
1771 btree_update_will_delete_key(as, &b->key);
1772 btree_update_will_add_key(as, new_key);
1774 parent = btree_node_parent(iter, b);
1777 bkey_copy(&new_hash->key, new_key);
1778 ret = bch2_btree_node_hash_insert(&c->btree_cache,
1779 new_hash, b->c.level, b->c.btree_id);
1783 bch2_keylist_add(&as->parent_keys, new_key);
1784 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, 0);
1787 mutex_lock(&c->btree_cache.lock);
1788 bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
1790 bch2_btree_node_hash_remove(&c->btree_cache, b);
1792 bkey_copy(&b->key, new_key);
1793 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
1795 mutex_unlock(&c->btree_cache.lock);
1797 bkey_copy(&b->key, new_key);
1800 BUG_ON(btree_node_root(c, b) != b);
1802 bch2_btree_node_lock_write(b, iter);
1803 bkey_copy(&b->key, new_key);
1805 if (btree_ptr_hash_val(&b->key) != b->hash_val) {
1806 mutex_lock(&c->btree_cache.lock);
1807 bch2_btree_node_hash_remove(&c->btree_cache, b);
1809 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
1811 mutex_unlock(&c->btree_cache.lock);
1814 btree_update_updated_root(as, b);
1815 bch2_btree_node_unlock_write(b, iter);
1818 bch2_btree_update_done(as);
1821 int bch2_btree_node_update_key(struct bch_fs *c, struct btree_iter *iter,
1823 struct bkey_i *new_key)
1825 struct btree *parent = btree_node_parent(iter, b);
1826 struct btree_update *as = NULL;
1827 struct btree *new_hash = NULL;
1831 closure_init_stack(&cl);
1833 if (!bch2_btree_iter_upgrade(iter, U8_MAX))
1836 if (!down_read_trylock(&c->gc_lock)) {
1837 bch2_trans_unlock(iter->trans);
1838 down_read(&c->gc_lock);
1840 if (!bch2_trans_relock(iter->trans)) {
1847 * check btree_ptr_hash_val() after @b is locked by
1848 * btree_iter_traverse():
1850 if (btree_ptr_hash_val(new_key) != b->hash_val) {
1851 /* bch2_btree_reserve_get will unlock */
1852 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1854 bch2_trans_unlock(iter->trans);
1855 up_read(&c->gc_lock);
1857 down_read(&c->gc_lock);
1859 if (!bch2_trans_relock(iter->trans)) {
1865 new_hash = bch2_btree_node_mem_alloc(c);
1868 as = bch2_btree_update_start(iter->trans, iter->btree_id,
1869 parent ? btree_update_reserve_required(c, parent) : 0,
1870 BTREE_INSERT_NOFAIL|
1871 BTREE_INSERT_USE_RESERVE|
1872 BTREE_INSERT_USE_ALLOC_RESERVE,
1883 bch2_trans_unlock(iter->trans);
1884 up_read(&c->gc_lock);
1886 down_read(&c->gc_lock);
1888 if (!bch2_trans_relock(iter->trans))
1892 ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(new_key));
1894 goto err_free_update;
1896 __bch2_btree_node_update_key(c, as, iter, b, new_hash, new_key);
1898 bch2_btree_iter_downgrade(iter);
1901 mutex_lock(&c->btree_cache.lock);
1902 list_move(&new_hash->list, &c->btree_cache.freeable);
1903 mutex_unlock(&c->btree_cache.lock);
1905 six_unlock_write(&new_hash->c.lock);
1906 six_unlock_intent(&new_hash->c.lock);
1908 up_read(&c->gc_lock);
1912 bch2_btree_update_free(as);
1919 * Only for filesystem bringup, when first reading the btree roots or allocating
1920 * btree roots when initializing a new filesystem:
1922 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
1924 BUG_ON(btree_node_root(c, b));
1926 bch2_btree_set_root_inmem(c, b);
1929 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
1935 closure_init_stack(&cl);
1938 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1942 b = bch2_btree_node_mem_alloc(c);
1943 bch2_btree_cache_cannibalize_unlock(c);
1945 set_btree_node_fake(b);
1949 bkey_btree_ptr_init(&b->key);
1950 b->key.k.p = POS_MAX;
1951 *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
1953 bch2_bset_init_first(b, &b->data->keys);
1954 bch2_btree_build_aux_trees(b);
1957 btree_set_min(b, POS_MIN);
1958 btree_set_max(b, POS_MAX);
1959 b->data->format = bch2_btree_calc_format(b);
1960 btree_node_set_format(b, b->data->format);
1962 ret = bch2_btree_node_hash_insert(&c->btree_cache, b,
1963 b->c.level, b->c.btree_id);
1966 bch2_btree_set_root_inmem(c, b);
1968 six_unlock_write(&b->c.lock);
1969 six_unlock_intent(&b->c.lock);
1972 ssize_t bch2_btree_updates_print(struct bch_fs *c, char *buf)
1974 struct printbuf out = _PBUF(buf, PAGE_SIZE);
1975 struct btree_update *as;
1977 mutex_lock(&c->btree_interior_update_lock);
1978 list_for_each_entry(as, &c->btree_interior_update_list, list)
1979 pr_buf(&out, "%p m %u w %u r %u j %llu\n",
1983 atomic_read(&as->cl.remaining) & CLOSURE_REMAINING_MASK,
1985 mutex_unlock(&c->btree_interior_update_lock);
1987 return out.pos - buf;
1990 size_t bch2_btree_interior_updates_nr_pending(struct bch_fs *c)
1993 struct list_head *i;
1995 mutex_lock(&c->btree_interior_update_lock);
1996 list_for_each(i, &c->btree_interior_update_list)
1998 mutex_unlock(&c->btree_interior_update_lock);
2003 void bch2_journal_entries_to_btree_roots(struct bch_fs *c, struct jset *jset)
2005 struct btree_root *r;
2006 struct jset_entry *entry;
2008 mutex_lock(&c->btree_root_lock);
2010 vstruct_for_each(jset, entry)
2011 if (entry->type == BCH_JSET_ENTRY_btree_root) {
2012 r = &c->btree_roots[entry->btree_id];
2013 r->level = entry->level;
2015 bkey_copy(&r->key, &entry->start[0]);
2018 mutex_unlock(&c->btree_root_lock);
2022 bch2_btree_roots_to_journal_entries(struct bch_fs *c,
2023 struct jset_entry *start,
2024 struct jset_entry *end)
2026 struct jset_entry *entry;
2027 unsigned long have = 0;
2030 for (entry = start; entry < end; entry = vstruct_next(entry))
2031 if (entry->type == BCH_JSET_ENTRY_btree_root)
2032 __set_bit(entry->btree_id, &have);
2034 mutex_lock(&c->btree_root_lock);
2036 for (i = 0; i < BTREE_ID_NR; i++)
2037 if (c->btree_roots[i].alive && !test_bit(i, &have)) {
2038 journal_entry_set(end,
2039 BCH_JSET_ENTRY_btree_root,
2040 i, c->btree_roots[i].level,
2041 &c->btree_roots[i].key,
2042 c->btree_roots[i].key.u64s);
2043 end = vstruct_next(end);
2046 mutex_unlock(&c->btree_root_lock);
2051 void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
2053 if (c->btree_interior_update_worker)
2054 destroy_workqueue(c->btree_interior_update_worker);
2055 mempool_exit(&c->btree_interior_update_pool);
2058 int bch2_fs_btree_interior_update_init(struct bch_fs *c)
2060 mutex_init(&c->btree_reserve_cache_lock);
2061 INIT_LIST_HEAD(&c->btree_interior_update_list);
2062 INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
2063 mutex_init(&c->btree_interior_update_lock);
2064 INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
2066 c->btree_interior_update_worker =
2067 alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
2068 if (!c->btree_interior_update_worker)
2071 return mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
2072 sizeof(struct btree_update));