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"
18 #include "journal_reclaim.h"
25 #include <linux/random.h>
27 static int bch2_btree_insert_node(struct btree_update *, struct btree_trans *,
28 struct btree_path *, struct btree *,
29 struct keylist *, unsigned);
30 static void bch2_btree_update_add_new_node(struct btree_update *, struct btree *);
32 static struct btree_path *get_unlocked_mut_path(struct btree_trans *trans,
33 enum btree_id btree_id,
37 struct btree_path *path;
39 path = bch2_path_get(trans, btree_id, pos, level + 1, level,
40 BTREE_ITER_NOPRESERVE|
41 BTREE_ITER_INTENT, _RET_IP_);
42 path = bch2_btree_path_make_mut(trans, path, true, _RET_IP_);
43 bch2_btree_path_downgrade(trans, path);
44 __bch2_btree_path_unlock(trans, path);
51 * Verify that child nodes correctly span parent node's range:
53 static void btree_node_interior_verify(struct bch_fs *c, struct btree *b)
55 #ifdef CONFIG_BCACHEFS_DEBUG
56 struct bpos next_node = b->data->min_key;
57 struct btree_node_iter iter;
59 struct bkey_s_c_btree_ptr_v2 bp;
61 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
65 if (!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))
68 bch2_btree_node_iter_init_from_start(&iter, b);
71 k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked);
72 if (k.k->type != KEY_TYPE_btree_ptr_v2)
74 bp = bkey_s_c_to_btree_ptr_v2(k);
76 if (!bpos_eq(next_node, bp.v->min_key)) {
77 bch2_dump_btree_node(c, b);
78 bch2_bpos_to_text(&buf1, next_node);
79 bch2_bpos_to_text(&buf2, bp.v->min_key);
80 panic("expected next min_key %s got %s\n", buf1.buf, buf2.buf);
83 bch2_btree_node_iter_advance(&iter, b);
85 if (bch2_btree_node_iter_end(&iter)) {
86 if (!bpos_eq(k.k->p, b->key.k.p)) {
87 bch2_dump_btree_node(c, b);
88 bch2_bpos_to_text(&buf1, b->key.k.p);
89 bch2_bpos_to_text(&buf2, k.k->p);
90 panic("expected end %s got %s\n", buf1.buf, buf2.buf);
95 next_node = bpos_successor(k.k->p);
100 /* Calculate ideal packed bkey format for new btree nodes: */
102 void __bch2_btree_calc_format(struct bkey_format_state *s, struct btree *b)
104 struct bkey_packed *k;
109 bset_tree_for_each_key(b, t, k)
110 if (!bkey_deleted(k)) {
111 uk = bkey_unpack_key(b, k);
112 bch2_bkey_format_add_key(s, &uk);
116 static struct bkey_format bch2_btree_calc_format(struct btree *b)
118 struct bkey_format_state s;
120 bch2_bkey_format_init(&s);
121 bch2_bkey_format_add_pos(&s, b->data->min_key);
122 bch2_bkey_format_add_pos(&s, b->data->max_key);
123 __bch2_btree_calc_format(&s, b);
125 return bch2_bkey_format_done(&s);
128 static size_t btree_node_u64s_with_format(struct btree *b,
129 struct bkey_format *new_f)
131 struct bkey_format *old_f = &b->format;
133 /* stupid integer promotion rules */
135 (((int) new_f->key_u64s - old_f->key_u64s) *
136 (int) b->nr.packed_keys) +
137 (((int) new_f->key_u64s - BKEY_U64s) *
138 (int) b->nr.unpacked_keys);
140 BUG_ON(delta + b->nr.live_u64s < 0);
142 return b->nr.live_u64s + delta;
146 * btree_node_format_fits - check if we could rewrite node with a new format
148 * This assumes all keys can pack with the new format -- it just checks if
149 * the re-packed keys would fit inside the node itself.
151 bool bch2_btree_node_format_fits(struct bch_fs *c, struct btree *b,
152 struct bkey_format *new_f)
154 size_t u64s = btree_node_u64s_with_format(b, new_f);
156 return __vstruct_bytes(struct btree_node, u64s) < btree_bytes(c);
159 /* Btree node freeing/allocation: */
161 static void __btree_node_free(struct bch_fs *c, struct btree *b)
163 trace_and_count(c, btree_node_free, c, b);
165 BUG_ON(btree_node_write_blocked(b));
166 BUG_ON(btree_node_dirty(b));
167 BUG_ON(btree_node_need_write(b));
168 BUG_ON(b == btree_node_root(c, b));
170 BUG_ON(!list_empty(&b->write_blocked));
171 BUG_ON(b->will_make_reachable);
173 clear_btree_node_noevict(b);
175 mutex_lock(&c->btree_cache.lock);
176 list_move(&b->list, &c->btree_cache.freeable);
177 mutex_unlock(&c->btree_cache.lock);
180 static void bch2_btree_node_free_inmem(struct btree_trans *trans,
181 struct btree_path *path,
184 struct bch_fs *c = trans->c;
185 unsigned level = b->c.level;
187 bch2_btree_node_lock_write_nofail(trans, path, &b->c);
188 bch2_btree_node_hash_remove(&c->btree_cache, b);
189 __btree_node_free(c, b);
190 six_unlock_write(&b->c.lock);
191 mark_btree_node_locked_noreset(path, level, SIX_LOCK_intent);
193 trans_for_each_path(trans, path)
194 if (path->l[level].b == b) {
195 btree_node_unlock(trans, path, level);
196 path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
200 static void bch2_btree_node_free_never_used(struct btree_update *as,
201 struct btree_trans *trans,
204 struct bch_fs *c = as->c;
205 struct prealloc_nodes *p = &as->prealloc_nodes[b->c.lock.readers != NULL];
206 struct btree_path *path;
207 unsigned level = b->c.level;
209 BUG_ON(!list_empty(&b->write_blocked));
210 BUG_ON(b->will_make_reachable != (1UL|(unsigned long) as));
212 b->will_make_reachable = 0;
213 closure_put(&as->cl);
215 clear_btree_node_will_make_reachable(b);
216 clear_btree_node_accessed(b);
217 clear_btree_node_dirty_acct(c, b);
218 clear_btree_node_need_write(b);
220 mutex_lock(&c->btree_cache.lock);
221 list_del_init(&b->list);
222 bch2_btree_node_hash_remove(&c->btree_cache, b);
223 mutex_unlock(&c->btree_cache.lock);
225 BUG_ON(p->nr >= ARRAY_SIZE(p->b));
228 six_unlock_intent(&b->c.lock);
230 trans_for_each_path(trans, path)
231 if (path->l[level].b == b) {
232 btree_node_unlock(trans, path, level);
233 path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
237 static struct btree *__bch2_btree_node_alloc(struct btree_trans *trans,
238 struct disk_reservation *res,
243 struct bch_fs *c = trans->c;
244 struct write_point *wp;
246 BKEY_PADDED_ONSTACK(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
247 struct open_buckets ob = { .nr = 0 };
248 struct bch_devs_list devs_have = (struct bch_devs_list) { 0 };
250 enum bch_watermark alloc_reserve;
253 if (flags & BTREE_INSERT_USE_RESERVE) {
255 alloc_reserve = BCH_WATERMARK_btree_copygc;
257 nr_reserve = BTREE_NODE_RESERVE;
258 alloc_reserve = BCH_WATERMARK_btree;
261 mutex_lock(&c->btree_reserve_cache_lock);
262 if (c->btree_reserve_cache_nr > nr_reserve) {
263 struct btree_alloc *a =
264 &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
267 bkey_copy(&tmp.k, &a->k);
268 mutex_unlock(&c->btree_reserve_cache_lock);
271 mutex_unlock(&c->btree_reserve_cache_lock);
274 ret = bch2_alloc_sectors_start_trans(trans,
275 c->opts.metadata_target ?:
276 c->opts.foreground_target,
278 writepoint_ptr(&c->btree_write_point),
281 c->opts.metadata_replicas_required,
282 alloc_reserve, 0, cl, &wp);
286 if (wp->sectors_free < btree_sectors(c)) {
287 struct open_bucket *ob;
290 open_bucket_for_each(c, &wp->ptrs, ob, i)
291 if (ob->sectors_free < btree_sectors(c))
292 ob->sectors_free = 0;
294 bch2_alloc_sectors_done(c, wp);
298 bkey_btree_ptr_v2_init(&tmp.k);
299 bch2_alloc_sectors_append_ptrs(c, wp, &tmp.k, btree_sectors(c), false);
301 bch2_open_bucket_get(c, wp, &ob);
302 bch2_alloc_sectors_done(c, wp);
304 b = bch2_btree_node_mem_alloc(trans, interior_node);
305 six_unlock_write(&b->c.lock);
306 six_unlock_intent(&b->c.lock);
308 /* we hold cannibalize_lock: */
312 bkey_copy(&b->key, &tmp.k);
318 static struct btree *bch2_btree_node_alloc(struct btree_update *as,
319 struct btree_trans *trans,
322 struct bch_fs *c = as->c;
324 struct prealloc_nodes *p = &as->prealloc_nodes[!!level];
327 BUG_ON(level >= BTREE_MAX_DEPTH);
332 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
333 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
335 set_btree_node_accessed(b);
336 set_btree_node_dirty_acct(c, b);
337 set_btree_node_need_write(b);
339 bch2_bset_init_first(b, &b->data->keys);
341 b->c.btree_id = as->btree_id;
342 b->version_ondisk = c->sb.version;
344 memset(&b->nr, 0, sizeof(b->nr));
345 b->data->magic = cpu_to_le64(bset_magic(c));
346 memset(&b->data->_ptr, 0, sizeof(b->data->_ptr));
348 SET_BTREE_NODE_ID(b->data, as->btree_id);
349 SET_BTREE_NODE_LEVEL(b->data, level);
351 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
352 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(&b->key);
355 bp->v.seq = b->data->keys.seq;
356 bp->v.sectors_written = 0;
359 SET_BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data, true);
361 bch2_btree_build_aux_trees(b);
363 ret = bch2_btree_node_hash_insert(&c->btree_cache, b, level, as->btree_id);
366 trace_and_count(c, btree_node_alloc, c, b);
367 bch2_increment_clock(c, btree_sectors(c), WRITE);
371 static void btree_set_min(struct btree *b, struct bpos pos)
373 if (b->key.k.type == KEY_TYPE_btree_ptr_v2)
374 bkey_i_to_btree_ptr_v2(&b->key)->v.min_key = pos;
375 b->data->min_key = pos;
378 static void btree_set_max(struct btree *b, struct bpos pos)
381 b->data->max_key = pos;
384 static struct btree *bch2_btree_node_alloc_replacement(struct btree_update *as,
385 struct btree_trans *trans,
388 struct btree *n = bch2_btree_node_alloc(as, trans, b->c.level);
389 struct bkey_format format = bch2_btree_calc_format(b);
392 * The keys might expand with the new format - if they wouldn't fit in
393 * the btree node anymore, use the old format for now:
395 if (!bch2_btree_node_format_fits(as->c, b, &format))
398 SET_BTREE_NODE_SEQ(n->data, BTREE_NODE_SEQ(b->data) + 1);
400 btree_set_min(n, b->data->min_key);
401 btree_set_max(n, b->data->max_key);
403 n->data->format = format;
404 btree_node_set_format(n, format);
406 bch2_btree_sort_into(as->c, n, b);
408 btree_node_reset_sib_u64s(n);
412 static struct btree *__btree_root_alloc(struct btree_update *as,
413 struct btree_trans *trans, unsigned level)
415 struct btree *b = bch2_btree_node_alloc(as, trans, level);
417 btree_set_min(b, POS_MIN);
418 btree_set_max(b, SPOS_MAX);
419 b->data->format = bch2_btree_calc_format(b);
421 btree_node_set_format(b, b->data->format);
422 bch2_btree_build_aux_trees(b);
427 static void bch2_btree_reserve_put(struct btree_update *as, struct btree_trans *trans)
429 struct bch_fs *c = as->c;
430 struct prealloc_nodes *p;
432 for (p = as->prealloc_nodes;
433 p < as->prealloc_nodes + ARRAY_SIZE(as->prealloc_nodes);
436 struct btree *b = p->b[--p->nr];
438 mutex_lock(&c->btree_reserve_cache_lock);
440 if (c->btree_reserve_cache_nr <
441 ARRAY_SIZE(c->btree_reserve_cache)) {
442 struct btree_alloc *a =
443 &c->btree_reserve_cache[c->btree_reserve_cache_nr++];
447 bkey_copy(&a->k, &b->key);
449 bch2_open_buckets_put(c, &b->ob);
452 mutex_unlock(&c->btree_reserve_cache_lock);
454 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
455 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
456 __btree_node_free(c, b);
457 six_unlock_write(&b->c.lock);
458 six_unlock_intent(&b->c.lock);
463 static int bch2_btree_reserve_get(struct btree_trans *trans,
464 struct btree_update *as,
465 unsigned nr_nodes[2],
469 struct bch_fs *c = as->c;
474 BUG_ON(nr_nodes[0] + nr_nodes[1] > BTREE_RESERVE_MAX);
477 * Protects reaping from the btree node cache and using the btree node
478 * open bucket reserve:
480 * BTREE_INSERT_NOWAIT only applies to btree node allocation, not
481 * blocking on this lock:
483 ret = bch2_btree_cache_cannibalize_lock(c, cl);
487 for (interior = 0; interior < 2; interior++) {
488 struct prealloc_nodes *p = as->prealloc_nodes + interior;
490 while (p->nr < nr_nodes[interior]) {
491 b = __bch2_btree_node_alloc(trans, &as->disk_res,
492 flags & BTREE_INSERT_NOWAIT ? NULL : cl,
503 bch2_btree_cache_cannibalize_unlock(c);
507 /* Asynchronous interior node update machinery */
509 static void bch2_btree_update_free(struct btree_update *as, struct btree_trans *trans)
511 struct bch_fs *c = as->c;
513 if (as->took_gc_lock)
514 up_read(&c->gc_lock);
515 as->took_gc_lock = false;
517 bch2_journal_preres_put(&c->journal, &as->journal_preres);
519 bch2_journal_pin_drop(&c->journal, &as->journal);
520 bch2_journal_pin_flush(&c->journal, &as->journal);
521 bch2_disk_reservation_put(c, &as->disk_res);
522 bch2_btree_reserve_put(as, trans);
524 bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_total],
527 mutex_lock(&c->btree_interior_update_lock);
528 list_del(&as->unwritten_list);
531 closure_debug_destroy(&as->cl);
532 mempool_free(as, &c->btree_interior_update_pool);
535 * Have to do the wakeup with btree_interior_update_lock still held,
536 * since being on btree_interior_update_list is our ref on @c:
538 closure_wake_up(&c->btree_interior_update_wait);
540 mutex_unlock(&c->btree_interior_update_lock);
543 static void btree_update_add_key(struct btree_update *as,
544 struct keylist *keys, struct btree *b)
546 struct bkey_i *k = &b->key;
548 BUG_ON(bch2_keylist_u64s(keys) + k->k.u64s >
549 ARRAY_SIZE(as->_old_keys));
551 bkey_copy(keys->top, k);
552 bkey_i_to_btree_ptr_v2(keys->top)->v.mem_ptr = b->c.level + 1;
554 bch2_keylist_push(keys);
558 * The transactional part of an interior btree node update, where we journal the
559 * update we did to the interior node and update alloc info:
561 static int btree_update_nodes_written_trans(struct btree_trans *trans,
562 struct btree_update *as)
567 ret = darray_make_room(&trans->extra_journal_entries, as->journal_u64s);
571 memcpy(&darray_top(trans->extra_journal_entries),
573 as->journal_u64s * sizeof(u64));
574 trans->extra_journal_entries.nr += as->journal_u64s;
576 trans->journal_pin = &as->journal;
578 for_each_keylist_key(&as->old_keys, k) {
579 unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr;
581 ret = bch2_trans_mark_old(trans, as->btree_id, level, bkey_i_to_s_c(k), 0);
586 for_each_keylist_key(&as->new_keys, k) {
587 unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr;
589 ret = bch2_trans_mark_new(trans, as->btree_id, level, k, 0);
597 static void btree_update_nodes_written(struct btree_update *as)
599 struct bch_fs *c = as->c;
601 struct btree_trans trans;
606 bch2_trans_init(&trans, c, 0, 512);
608 * If we're already in an error state, it might be because a btree node
609 * was never written, and we might be trying to free that same btree
610 * node here, but it won't have been marked as allocated and we'll see
611 * spurious disk usage inconsistencies in the transactional part below
612 * if we don't skip it:
614 ret = bch2_journal_error(&c->journal);
619 * Wait for any in flight writes to finish before we free the old nodes
622 for (i = 0; i < as->nr_old_nodes; i++) {
625 b = as->old_nodes[i];
627 btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read);
628 seq = b->data ? b->data->keys.seq : 0;
629 six_unlock_read(&b->c.lock);
631 if (seq == as->old_nodes_seq[i])
632 wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight_inner,
633 TASK_UNINTERRUPTIBLE);
637 * We did an update to a parent node where the pointers we added pointed
638 * to child nodes that weren't written yet: now, the child nodes have
639 * been written so we can write out the update to the interior node.
643 * We can't call into journal reclaim here: we'd block on the journal
644 * reclaim lock, but we may need to release the open buckets we have
645 * pinned in order for other btree updates to make forward progress, and
646 * journal reclaim does btree updates when flushing bkey_cached entries,
647 * which may require allocations as well.
649 ret = commit_do(&trans, &as->disk_res, &journal_seq,
651 BTREE_INSERT_NOCHECK_RW|
652 BTREE_INSERT_USE_RESERVE|
653 BTREE_INSERT_JOURNAL_RECLAIM|
654 JOURNAL_WATERMARK_reserved,
655 btree_update_nodes_written_trans(&trans, as));
656 bch2_trans_unlock(&trans);
658 bch2_fs_fatal_err_on(ret && !bch2_journal_error(&c->journal), c,
659 "%s(): error %s", __func__, bch2_err_str(ret));
662 struct btree_path *path;
665 path = get_unlocked_mut_path(&trans, as->btree_id, b->c.level, b->key.k.p);
667 * @b is the node we did the final insert into:
669 * On failure to get a journal reservation, we still have to
670 * unblock the write and allow most of the write path to happen
671 * so that shutdown works, but the i->journal_seq mechanism
672 * won't work to prevent the btree write from being visible (we
673 * didn't get a journal sequence number) - instead
674 * __bch2_btree_node_write() doesn't do the actual write if
675 * we're in journal error state:
679 * Ensure transaction is unlocked before using
680 * btree_node_lock_nopath() (the use of which is always suspect,
681 * we need to work on removing this in the future)
683 * It should be, but get_unlocked_mut_path() -> bch2_path_get()
684 * calls bch2_path_upgrade(), before we call path_make_mut(), so
685 * we may rarely end up with a locked path besides the one we
688 bch2_trans_unlock(&trans);
689 btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_intent);
690 mark_btree_node_locked(&trans, path, b->c.level, SIX_LOCK_intent);
691 path->l[b->c.level].lock_seq = six_lock_seq(&b->c.lock);
692 path->l[b->c.level].b = b;
694 bch2_btree_node_lock_write_nofail(&trans, path, &b->c);
696 mutex_lock(&c->btree_interior_update_lock);
698 list_del(&as->write_blocked_list);
699 if (list_empty(&b->write_blocked))
700 clear_btree_node_write_blocked(b);
703 * Node might have been freed, recheck under
704 * btree_interior_update_lock:
707 struct bset *i = btree_bset_last(b);
710 BUG_ON(!btree_node_dirty(b));
713 i->journal_seq = cpu_to_le64(
715 le64_to_cpu(i->journal_seq)));
717 bch2_btree_add_journal_pin(c, b, journal_seq);
720 * If we didn't get a journal sequence number we
721 * can't write this btree node, because recovery
722 * won't know to ignore this write:
724 set_btree_node_never_write(b);
728 mutex_unlock(&c->btree_interior_update_lock);
730 mark_btree_node_locked_noreset(path, b->c.level, SIX_LOCK_intent);
731 six_unlock_write(&b->c.lock);
733 btree_node_write_if_need(c, b, SIX_LOCK_intent);
734 btree_node_unlock(&trans, path, b->c.level);
735 bch2_path_put(&trans, path, true);
738 bch2_journal_pin_drop(&c->journal, &as->journal);
740 bch2_journal_preres_put(&c->journal, &as->journal_preres);
742 mutex_lock(&c->btree_interior_update_lock);
743 for (i = 0; i < as->nr_new_nodes; i++) {
744 b = as->new_nodes[i];
746 BUG_ON(b->will_make_reachable != (unsigned long) as);
747 b->will_make_reachable = 0;
748 clear_btree_node_will_make_reachable(b);
750 mutex_unlock(&c->btree_interior_update_lock);
752 for (i = 0; i < as->nr_new_nodes; i++) {
753 b = as->new_nodes[i];
755 btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read);
756 btree_node_write_if_need(c, b, SIX_LOCK_read);
757 six_unlock_read(&b->c.lock);
760 for (i = 0; i < as->nr_open_buckets; i++)
761 bch2_open_bucket_put(c, c->open_buckets + as->open_buckets[i]);
763 bch2_btree_update_free(as, &trans);
764 bch2_trans_exit(&trans);
767 static void btree_interior_update_work(struct work_struct *work)
770 container_of(work, struct bch_fs, btree_interior_update_work);
771 struct btree_update *as;
774 mutex_lock(&c->btree_interior_update_lock);
775 as = list_first_entry_or_null(&c->btree_interior_updates_unwritten,
776 struct btree_update, unwritten_list);
777 if (as && !as->nodes_written)
779 mutex_unlock(&c->btree_interior_update_lock);
784 btree_update_nodes_written(as);
788 static void btree_update_set_nodes_written(struct closure *cl)
790 struct btree_update *as = container_of(cl, struct btree_update, cl);
791 struct bch_fs *c = as->c;
793 mutex_lock(&c->btree_interior_update_lock);
794 as->nodes_written = true;
795 mutex_unlock(&c->btree_interior_update_lock);
797 queue_work(c->btree_interior_update_worker, &c->btree_interior_update_work);
801 * We're updating @b with pointers to nodes that haven't finished writing yet:
802 * block @b from being written until @as completes
804 static void btree_update_updated_node(struct btree_update *as, struct btree *b)
806 struct bch_fs *c = as->c;
808 mutex_lock(&c->btree_interior_update_lock);
809 list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
811 BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
812 BUG_ON(!btree_node_dirty(b));
815 as->mode = BTREE_INTERIOR_UPDATING_NODE;
818 set_btree_node_write_blocked(b);
819 list_add(&as->write_blocked_list, &b->write_blocked);
821 mutex_unlock(&c->btree_interior_update_lock);
824 static void btree_update_reparent(struct btree_update *as,
825 struct btree_update *child)
827 struct bch_fs *c = as->c;
829 lockdep_assert_held(&c->btree_interior_update_lock);
832 child->mode = BTREE_INTERIOR_UPDATING_AS;
834 bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal, NULL);
837 static void btree_update_updated_root(struct btree_update *as, struct btree *b)
839 struct bkey_i *insert = &b->key;
840 struct bch_fs *c = as->c;
842 BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
844 BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
845 ARRAY_SIZE(as->journal_entries));
848 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
849 BCH_JSET_ENTRY_btree_root,
850 b->c.btree_id, b->c.level,
851 insert, insert->k.u64s);
853 mutex_lock(&c->btree_interior_update_lock);
854 list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
856 as->mode = BTREE_INTERIOR_UPDATING_ROOT;
857 mutex_unlock(&c->btree_interior_update_lock);
861 * bch2_btree_update_add_new_node:
863 * This causes @as to wait on @b to be written, before it gets to
864 * bch2_btree_update_nodes_written
866 * Additionally, it sets b->will_make_reachable to prevent any additional writes
867 * to @b from happening besides the first until @b is reachable on disk
869 * And it adds @b to the list of @as's new nodes, so that we can update sector
870 * counts in bch2_btree_update_nodes_written:
872 static void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b)
874 struct bch_fs *c = as->c;
876 closure_get(&as->cl);
878 mutex_lock(&c->btree_interior_update_lock);
879 BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes));
880 BUG_ON(b->will_make_reachable);
882 as->new_nodes[as->nr_new_nodes++] = b;
883 b->will_make_reachable = 1UL|(unsigned long) as;
884 set_btree_node_will_make_reachable(b);
886 mutex_unlock(&c->btree_interior_update_lock);
888 btree_update_add_key(as, &as->new_keys, b);
890 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
891 unsigned bytes = vstruct_end(&b->data->keys) - (void *) b->data;
892 unsigned sectors = round_up(bytes, block_bytes(c)) >> 9;
894 bkey_i_to_btree_ptr_v2(&b->key)->v.sectors_written =
895 cpu_to_le16(sectors);
900 * returns true if @b was a new node
902 static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b)
904 struct btree_update *as;
908 mutex_lock(&c->btree_interior_update_lock);
910 * When b->will_make_reachable != 0, it owns a ref on as->cl that's
911 * dropped when it gets written by bch2_btree_complete_write - the
912 * xchg() is for synchronization with bch2_btree_complete_write:
914 v = xchg(&b->will_make_reachable, 0);
915 clear_btree_node_will_make_reachable(b);
916 as = (struct btree_update *) (v & ~1UL);
919 mutex_unlock(&c->btree_interior_update_lock);
923 for (i = 0; i < as->nr_new_nodes; i++)
924 if (as->new_nodes[i] == b)
929 array_remove_item(as->new_nodes, as->nr_new_nodes, i);
930 mutex_unlock(&c->btree_interior_update_lock);
933 closure_put(&as->cl);
936 static void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b)
939 as->open_buckets[as->nr_open_buckets++] =
944 * @b is being split/rewritten: it may have pointers to not-yet-written btree
945 * nodes and thus outstanding btree_updates - redirect @b's
946 * btree_updates to point to this btree_update:
948 static void bch2_btree_interior_update_will_free_node(struct btree_update *as,
951 struct bch_fs *c = as->c;
952 struct btree_update *p, *n;
953 struct btree_write *w;
955 set_btree_node_dying(b);
957 if (btree_node_fake(b))
960 mutex_lock(&c->btree_interior_update_lock);
963 * Does this node have any btree_update operations preventing
964 * it from being written?
966 * If so, redirect them to point to this btree_update: we can
967 * write out our new nodes, but we won't make them visible until those
968 * operations complete
970 list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) {
971 list_del_init(&p->write_blocked_list);
972 btree_update_reparent(as, p);
975 * for flush_held_btree_writes() waiting on updates to flush or
976 * nodes to be writeable:
978 closure_wake_up(&c->btree_interior_update_wait);
981 clear_btree_node_dirty_acct(c, b);
982 clear_btree_node_need_write(b);
983 clear_btree_node_write_blocked(b);
986 * Does this node have unwritten data that has a pin on the journal?
988 * If so, transfer that pin to the btree_update operation -
989 * note that if we're freeing multiple nodes, we only need to keep the
990 * oldest pin of any of the nodes we're freeing. We'll release the pin
991 * when the new nodes are persistent and reachable on disk:
993 w = btree_current_write(b);
994 bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
995 bch2_journal_pin_drop(&c->journal, &w->journal);
997 w = btree_prev_write(b);
998 bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
999 bch2_journal_pin_drop(&c->journal, &w->journal);
1001 mutex_unlock(&c->btree_interior_update_lock);
1004 * Is this a node that isn't reachable on disk yet?
1006 * Nodes that aren't reachable yet have writes blocked until they're
1007 * reachable - now that we've cancelled any pending writes and moved
1008 * things waiting on that write to wait on this update, we can drop this
1009 * node from the list of nodes that the other update is making
1010 * reachable, prior to freeing it:
1012 btree_update_drop_new_node(c, b);
1014 btree_update_add_key(as, &as->old_keys, b);
1016 as->old_nodes[as->nr_old_nodes] = b;
1017 as->old_nodes_seq[as->nr_old_nodes] = b->data->keys.seq;
1021 static void bch2_btree_update_done(struct btree_update *as, struct btree_trans *trans)
1023 struct bch_fs *c = as->c;
1024 u64 start_time = as->start_time;
1026 BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE);
1028 if (as->took_gc_lock)
1029 up_read(&as->c->gc_lock);
1030 as->took_gc_lock = false;
1032 bch2_btree_reserve_put(as, trans);
1034 continue_at(&as->cl, btree_update_set_nodes_written,
1035 as->c->btree_interior_update_worker);
1037 bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_foreground],
1041 static struct btree_update *
1042 bch2_btree_update_start(struct btree_trans *trans, struct btree_path *path,
1043 unsigned level, bool split, unsigned flags)
1045 struct bch_fs *c = trans->c;
1046 struct btree_update *as;
1047 u64 start_time = local_clock();
1048 int disk_res_flags = (flags & BTREE_INSERT_NOFAIL)
1049 ? BCH_DISK_RESERVATION_NOFAIL : 0;
1050 unsigned nr_nodes[2] = { 0, 0 };
1051 unsigned update_level = level;
1052 int journal_flags = flags & JOURNAL_WATERMARK_MASK;
1054 u32 restart_count = trans->restart_count;
1056 BUG_ON(!path->should_be_locked);
1058 if (flags & BTREE_INSERT_JOURNAL_RECLAIM)
1059 journal_flags |= JOURNAL_RES_GET_NONBLOCK;
1062 nr_nodes[!!update_level] += 1 + split;
1065 ret = bch2_btree_path_upgrade(trans, path, update_level + 1);
1067 return ERR_PTR(ret);
1069 if (!btree_path_node(path, update_level)) {
1070 /* Allocating new root? */
1071 nr_nodes[1] += split;
1072 update_level = BTREE_MAX_DEPTH;
1076 if (bch2_btree_node_insert_fits(c, path->l[update_level].b,
1077 BKEY_BTREE_PTR_U64s_MAX * (1 + split)))
1080 split = path->l[update_level].b->nr.live_u64s > BTREE_SPLIT_THRESHOLD(c);
1083 if (flags & BTREE_INSERT_GC_LOCK_HELD)
1084 lockdep_assert_held(&c->gc_lock);
1085 else if (!down_read_trylock(&c->gc_lock)) {
1086 ret = drop_locks_do(trans, (down_read(&c->gc_lock), 0));
1088 up_read(&c->gc_lock);
1089 return ERR_PTR(ret);
1093 as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOFS);
1094 memset(as, 0, sizeof(*as));
1095 closure_init(&as->cl, NULL);
1097 as->start_time = start_time;
1098 as->mode = BTREE_INTERIOR_NO_UPDATE;
1099 as->took_gc_lock = !(flags & BTREE_INSERT_GC_LOCK_HELD);
1100 as->btree_id = path->btree_id;
1101 as->update_level = update_level;
1102 INIT_LIST_HEAD(&as->list);
1103 INIT_LIST_HEAD(&as->unwritten_list);
1104 INIT_LIST_HEAD(&as->write_blocked_list);
1105 bch2_keylist_init(&as->old_keys, as->_old_keys);
1106 bch2_keylist_init(&as->new_keys, as->_new_keys);
1107 bch2_keylist_init(&as->parent_keys, as->inline_keys);
1109 mutex_lock(&c->btree_interior_update_lock);
1110 list_add_tail(&as->list, &c->btree_interior_update_list);
1111 mutex_unlock(&c->btree_interior_update_lock);
1114 * We don't want to allocate if we're in an error state, that can cause
1115 * deadlock on emergency shutdown due to open buckets getting stuck in
1116 * the btree_reserve_cache after allocator shutdown has cleared it out.
1117 * This check needs to come after adding us to the btree_interior_update
1118 * list but before calling bch2_btree_reserve_get, to synchronize with
1119 * __bch2_fs_read_only().
1121 ret = bch2_journal_error(&c->journal);
1125 ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
1126 BTREE_UPDATE_JOURNAL_RES,
1127 journal_flags|JOURNAL_RES_GET_NONBLOCK);
1129 if (flags & BTREE_INSERT_JOURNAL_RECLAIM) {
1130 ret = -BCH_ERR_journal_reclaim_would_deadlock;
1134 ret = drop_locks_do(trans,
1135 bch2_journal_preres_get(&c->journal, &as->journal_preres,
1136 BTREE_UPDATE_JOURNAL_RES,
1138 if (ret == -BCH_ERR_journal_preres_get_blocked) {
1139 trace_and_count(c, trans_restart_journal_preres_get, trans, _RET_IP_, journal_flags);
1140 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_journal_preres_get);
1146 ret = bch2_disk_reservation_get(c, &as->disk_res,
1147 (nr_nodes[0] + nr_nodes[1]) * btree_sectors(c),
1148 c->opts.metadata_replicas,
1153 ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, NULL);
1154 if (bch2_err_matches(ret, ENOSPC) ||
1155 bch2_err_matches(ret, ENOMEM)) {
1158 closure_init_stack(&cl);
1161 ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, &cl);
1163 bch2_trans_unlock(trans);
1165 } while (bch2_err_matches(ret, BCH_ERR_operation_blocked));
1169 trace_and_count(c, btree_reserve_get_fail, trans->fn,
1170 _RET_IP_, nr_nodes[0] + nr_nodes[1], ret);
1174 ret = bch2_trans_relock(trans);
1178 bch2_trans_verify_not_restarted(trans, restart_count);
1181 bch2_btree_update_free(as, trans);
1182 return ERR_PTR(ret);
1185 /* Btree root updates: */
1187 static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
1189 /* Root nodes cannot be reaped */
1190 mutex_lock(&c->btree_cache.lock);
1191 list_del_init(&b->list);
1192 mutex_unlock(&c->btree_cache.lock);
1194 mutex_lock(&c->btree_root_lock);
1195 BUG_ON(btree_node_root(c, b) &&
1196 (b->c.level < btree_node_root(c, b)->c.level ||
1197 !btree_node_dying(btree_node_root(c, b))));
1199 btree_node_root(c, b) = b;
1200 mutex_unlock(&c->btree_root_lock);
1202 bch2_recalc_btree_reserve(c);
1206 * bch_btree_set_root - update the root in memory and on disk
1208 * To ensure forward progress, the current task must not be holding any
1209 * btree node write locks. However, you must hold an intent lock on the
1212 * Note: This allocates a journal entry but doesn't add any keys to
1213 * it. All the btree roots are part of every journal write, so there
1214 * is nothing new to be done. This just guarantees that there is a
1217 static void bch2_btree_set_root(struct btree_update *as,
1218 struct btree_trans *trans,
1219 struct btree_path *path,
1222 struct bch_fs *c = as->c;
1225 trace_and_count(c, btree_node_set_root, c, b);
1227 old = btree_node_root(c, b);
1230 * Ensure no one is using the old root while we switch to the
1233 bch2_btree_node_lock_write_nofail(trans, path, &old->c);
1235 bch2_btree_set_root_inmem(c, b);
1237 btree_update_updated_root(as, b);
1240 * Unlock old root after new root is visible:
1242 * The new root isn't persistent, but that's ok: we still have
1243 * an intent lock on the new root, and any updates that would
1244 * depend on the new root would have to update the new root.
1246 bch2_btree_node_unlock_write(trans, path, old);
1249 /* Interior node updates: */
1251 static void bch2_insert_fixup_btree_ptr(struct btree_update *as,
1252 struct btree_trans *trans,
1253 struct btree_path *path,
1255 struct btree_node_iter *node_iter,
1256 struct bkey_i *insert)
1258 struct bch_fs *c = as->c;
1259 struct bkey_packed *k;
1260 struct printbuf buf = PRINTBUF;
1261 unsigned long old, new, v;
1263 BUG_ON(insert->k.type == KEY_TYPE_btree_ptr_v2 &&
1264 !btree_ptr_sectors_written(insert));
1266 if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)))
1267 bch2_journal_key_overwritten(c, b->c.btree_id, b->c.level, insert->k.p);
1269 if (bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1270 btree_node_type(b), WRITE, &buf) ?:
1271 bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert), &buf)) {
1272 printbuf_reset(&buf);
1273 prt_printf(&buf, "inserting invalid bkey\n ");
1274 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
1275 prt_printf(&buf, "\n ");
1276 bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1277 btree_node_type(b), WRITE, &buf);
1278 bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert), &buf);
1280 bch2_fs_inconsistent(c, "%s", buf.buf);
1284 BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1285 ARRAY_SIZE(as->journal_entries));
1288 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
1289 BCH_JSET_ENTRY_btree_keys,
1290 b->c.btree_id, b->c.level,
1291 insert, insert->k.u64s);
1293 while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1294 bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1295 bch2_btree_node_iter_advance(node_iter, b);
1297 bch2_btree_bset_insert_key(trans, path, b, node_iter, insert);
1298 set_btree_node_dirty_acct(c, b);
1300 v = READ_ONCE(b->flags);
1304 new &= ~BTREE_WRITE_TYPE_MASK;
1305 new |= BTREE_WRITE_interior;
1306 new |= 1 << BTREE_NODE_need_write;
1307 } while ((v = cmpxchg(&b->flags, old, new)) != old);
1309 printbuf_exit(&buf);
1313 __bch2_btree_insert_keys_interior(struct btree_update *as,
1314 struct btree_trans *trans,
1315 struct btree_path *path,
1317 struct btree_node_iter node_iter,
1318 struct keylist *keys)
1320 struct bkey_i *insert = bch2_keylist_front(keys);
1321 struct bkey_packed *k;
1323 BUG_ON(btree_node_type(b) != BKEY_TYPE_btree);
1325 while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1326 (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0))
1329 while (!bch2_keylist_empty(keys)) {
1330 struct bkey_i *k = bch2_keylist_front(keys);
1332 if (bpos_gt(k->k.p, b->key.k.p))
1335 bch2_insert_fixup_btree_ptr(as, trans, path, b, &node_iter, k);
1336 bch2_keylist_pop_front(keys);
1341 * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1344 static void __btree_split_node(struct btree_update *as,
1345 struct btree_trans *trans,
1349 struct bkey_packed *k;
1350 struct bpos n1_pos = POS_MIN;
1351 struct btree_node_iter iter;
1352 struct bset *bsets[2];
1353 struct bkey_format_state format[2];
1354 struct bkey_packed *out[2];
1356 unsigned u64s, n1_u64s = (b->nr.live_u64s * 3) / 5;
1359 for (i = 0; i < 2; i++) {
1360 BUG_ON(n[i]->nsets != 1);
1362 bsets[i] = btree_bset_first(n[i]);
1363 out[i] = bsets[i]->start;
1365 SET_BTREE_NODE_SEQ(n[i]->data, BTREE_NODE_SEQ(b->data) + 1);
1366 bch2_bkey_format_init(&format[i]);
1370 for_each_btree_node_key(b, k, &iter) {
1371 if (bkey_deleted(k))
1374 i = u64s >= n1_u64s;
1376 uk = bkey_unpack_key(b, k);
1379 bch2_bkey_format_add_key(&format[i], &uk);
1382 btree_set_min(n[0], b->data->min_key);
1383 btree_set_max(n[0], n1_pos);
1384 btree_set_min(n[1], bpos_successor(n1_pos));
1385 btree_set_max(n[1], b->data->max_key);
1387 for (i = 0; i < 2; i++) {
1388 bch2_bkey_format_add_pos(&format[i], n[i]->data->min_key);
1389 bch2_bkey_format_add_pos(&format[i], n[i]->data->max_key);
1391 n[i]->data->format = bch2_bkey_format_done(&format[i]);
1392 btree_node_set_format(n[i], n[i]->data->format);
1396 for_each_btree_node_key(b, k, &iter) {
1397 if (bkey_deleted(k))
1400 i = u64s >= n1_u64s;
1403 if (bch2_bkey_transform(&n[i]->format, out[i], bkey_packed(k)
1404 ? &b->format: &bch2_bkey_format_current, k))
1405 out[i]->format = KEY_FORMAT_LOCAL_BTREE;
1407 bch2_bkey_unpack(b, (void *) out[i], k);
1409 out[i]->needs_whiteout = false;
1411 btree_keys_account_key_add(&n[i]->nr, 0, out[i]);
1412 out[i] = bkey_p_next(out[i]);
1415 for (i = 0; i < 2; i++) {
1416 bsets[i]->u64s = cpu_to_le16((u64 *) out[i] - bsets[i]->_data);
1418 BUG_ON(!bsets[i]->u64s);
1420 set_btree_bset_end(n[i], n[i]->set);
1422 btree_node_reset_sib_u64s(n[i]);
1424 bch2_verify_btree_nr_keys(n[i]);
1427 btree_node_interior_verify(as->c, n[i]);
1432 * For updates to interior nodes, we've got to do the insert before we split
1433 * because the stuff we're inserting has to be inserted atomically. Post split,
1434 * the keys might have to go in different nodes and the split would no longer be
1437 * Worse, if the insert is from btree node coalescing, if we do the insert after
1438 * we do the split (and pick the pivot) - the pivot we pick might be between
1439 * nodes that were coalesced, and thus in the middle of a child node post
1442 static void btree_split_insert_keys(struct btree_update *as,
1443 struct btree_trans *trans,
1444 struct btree_path *path,
1446 struct keylist *keys)
1448 if (!bch2_keylist_empty(keys) &&
1449 bpos_le(bch2_keylist_front(keys)->k.p, b->data->max_key)) {
1450 struct btree_node_iter node_iter;
1452 bch2_btree_node_iter_init(&node_iter, b, &bch2_keylist_front(keys)->k.p);
1454 __bch2_btree_insert_keys_interior(as, trans, path, b, node_iter, keys);
1456 btree_node_interior_verify(as->c, b);
1460 static int btree_split(struct btree_update *as, struct btree_trans *trans,
1461 struct btree_path *path, struct btree *b,
1462 struct keylist *keys, unsigned flags)
1464 struct bch_fs *c = as->c;
1465 struct btree *parent = btree_node_parent(path, b);
1466 struct btree *n1, *n2 = NULL, *n3 = NULL;
1467 struct btree_path *path1 = NULL, *path2 = NULL;
1468 u64 start_time = local_clock();
1471 BUG_ON(!parent && (b != btree_node_root(c, b)));
1472 BUG_ON(parent && !btree_node_intent_locked(path, b->c.level + 1));
1474 bch2_btree_interior_update_will_free_node(as, b);
1476 if (b->nr.live_u64s > BTREE_SPLIT_THRESHOLD(c)) {
1479 trace_and_count(c, btree_node_split, c, b);
1481 n[0] = n1 = bch2_btree_node_alloc(as, trans, b->c.level);
1482 n[1] = n2 = bch2_btree_node_alloc(as, trans, b->c.level);
1484 __btree_split_node(as, trans, b, n);
1487 btree_split_insert_keys(as, trans, path, n1, keys);
1488 btree_split_insert_keys(as, trans, path, n2, keys);
1489 BUG_ON(!bch2_keylist_empty(keys));
1492 bch2_btree_build_aux_trees(n2);
1493 bch2_btree_build_aux_trees(n1);
1495 bch2_btree_update_add_new_node(as, n1);
1496 bch2_btree_update_add_new_node(as, n2);
1497 six_unlock_write(&n2->c.lock);
1498 six_unlock_write(&n1->c.lock);
1500 path1 = get_unlocked_mut_path(trans, path->btree_id, n1->c.level, n1->key.k.p);
1501 six_lock_increment(&n1->c.lock, SIX_LOCK_intent);
1502 mark_btree_node_locked(trans, path1, n1->c.level, SIX_LOCK_intent);
1503 bch2_btree_path_level_init(trans, path1, n1);
1505 path2 = get_unlocked_mut_path(trans, path->btree_id, n2->c.level, n2->key.k.p);
1506 six_lock_increment(&n2->c.lock, SIX_LOCK_intent);
1507 mark_btree_node_locked(trans, path2, n2->c.level, SIX_LOCK_intent);
1508 bch2_btree_path_level_init(trans, path2, n2);
1511 * Note that on recursive parent_keys == keys, so we
1512 * can't start adding new keys to parent_keys before emptying it
1513 * out (which we did with btree_split_insert_keys() above)
1515 bch2_keylist_add(&as->parent_keys, &n1->key);
1516 bch2_keylist_add(&as->parent_keys, &n2->key);
1519 /* Depth increases, make a new root */
1520 n3 = __btree_root_alloc(as, trans, b->c.level + 1);
1522 bch2_btree_update_add_new_node(as, n3);
1523 six_unlock_write(&n3->c.lock);
1525 path2->locks_want++;
1526 BUG_ON(btree_node_locked(path2, n3->c.level));
1527 six_lock_increment(&n3->c.lock, SIX_LOCK_intent);
1528 mark_btree_node_locked(trans, path2, n3->c.level, SIX_LOCK_intent);
1529 bch2_btree_path_level_init(trans, path2, n3);
1531 n3->sib_u64s[0] = U16_MAX;
1532 n3->sib_u64s[1] = U16_MAX;
1534 btree_split_insert_keys(as, trans, path, n3, &as->parent_keys);
1537 trace_and_count(c, btree_node_compact, c, b);
1539 n1 = bch2_btree_node_alloc_replacement(as, trans, b);
1542 btree_split_insert_keys(as, trans, path, n1, keys);
1543 BUG_ON(!bch2_keylist_empty(keys));
1546 bch2_btree_build_aux_trees(n1);
1547 bch2_btree_update_add_new_node(as, n1);
1548 six_unlock_write(&n1->c.lock);
1550 path1 = get_unlocked_mut_path(trans, path->btree_id, n1->c.level, n1->key.k.p);
1551 six_lock_increment(&n1->c.lock, SIX_LOCK_intent);
1552 mark_btree_node_locked(trans, path1, n1->c.level, SIX_LOCK_intent);
1553 bch2_btree_path_level_init(trans, path1, n1);
1556 bch2_keylist_add(&as->parent_keys, &n1->key);
1559 /* New nodes all written, now make them visible: */
1562 /* Split a non root node */
1563 ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1567 bch2_btree_set_root(as, trans, path, n3);
1569 /* Root filled up but didn't need to be split */
1570 bch2_btree_set_root(as, trans, path, n1);
1574 bch2_btree_update_get_open_buckets(as, n3);
1575 bch2_btree_node_write(c, n3, SIX_LOCK_intent, 0);
1578 bch2_btree_update_get_open_buckets(as, n2);
1579 bch2_btree_node_write(c, n2, SIX_LOCK_intent, 0);
1581 bch2_btree_update_get_open_buckets(as, n1);
1582 bch2_btree_node_write(c, n1, SIX_LOCK_intent, 0);
1585 * The old node must be freed (in memory) _before_ unlocking the new
1586 * nodes - else another thread could re-acquire a read lock on the old
1587 * node after another thread has locked and updated the new node, thus
1588 * seeing stale data:
1590 bch2_btree_node_free_inmem(trans, path, b);
1593 bch2_trans_node_add(trans, n3);
1595 bch2_trans_node_add(trans, n2);
1596 bch2_trans_node_add(trans, n1);
1599 six_unlock_intent(&n3->c.lock);
1601 six_unlock_intent(&n2->c.lock);
1602 six_unlock_intent(&n1->c.lock);
1605 __bch2_btree_path_unlock(trans, path2);
1606 bch2_path_put(trans, path2, true);
1609 __bch2_btree_path_unlock(trans, path1);
1610 bch2_path_put(trans, path1, true);
1613 bch2_trans_verify_locks(trans);
1615 bch2_time_stats_update(&c->times[n2
1616 ? BCH_TIME_btree_node_split
1617 : BCH_TIME_btree_node_compact],
1622 bch2_btree_node_free_never_used(as, trans, n3);
1624 bch2_btree_node_free_never_used(as, trans, n2);
1625 bch2_btree_node_free_never_used(as, trans, n1);
1630 bch2_btree_insert_keys_interior(struct btree_update *as,
1631 struct btree_trans *trans,
1632 struct btree_path *path,
1634 struct keylist *keys)
1636 struct btree_path *linked;
1638 __bch2_btree_insert_keys_interior(as, trans, path, b,
1639 path->l[b->c.level].iter, keys);
1641 btree_update_updated_node(as, b);
1643 trans_for_each_path_with_node(trans, b, linked)
1644 bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b);
1646 bch2_trans_verify_paths(trans);
1650 * bch_btree_insert_node - insert bkeys into a given btree node
1652 * @iter: btree iterator
1653 * @keys: list of keys to insert
1654 * @hook: insert callback
1655 * @persistent: if not null, @persistent will wait on journal write
1657 * Inserts as many keys as it can into a given btree node, splitting it if full.
1658 * If a split occurred, this function will return early. This can only happen
1659 * for leaf nodes -- inserts into interior nodes have to be atomic.
1661 static int bch2_btree_insert_node(struct btree_update *as, struct btree_trans *trans,
1662 struct btree_path *path, struct btree *b,
1663 struct keylist *keys, unsigned flags)
1665 struct bch_fs *c = as->c;
1666 int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1667 int old_live_u64s = b->nr.live_u64s;
1668 int live_u64s_added, u64s_added;
1671 lockdep_assert_held(&c->gc_lock);
1672 BUG_ON(!btree_node_intent_locked(path, b->c.level));
1673 BUG_ON(!b->c.level);
1674 BUG_ON(!as || as->b);
1675 bch2_verify_keylist_sorted(keys);
1677 ret = bch2_btree_node_lock_write(trans, path, &b->c);
1681 bch2_btree_node_prep_for_write(trans, path, b);
1683 if (!bch2_btree_node_insert_fits(c, b, bch2_keylist_u64s(keys))) {
1684 bch2_btree_node_unlock_write(trans, path, b);
1688 btree_node_interior_verify(c, b);
1690 bch2_btree_insert_keys_interior(as, trans, path, b, keys);
1692 live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1693 u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1695 if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1696 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1697 if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1698 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1700 if (u64s_added > live_u64s_added &&
1701 bch2_maybe_compact_whiteouts(c, b))
1702 bch2_trans_node_reinit_iter(trans, b);
1704 bch2_btree_node_unlock_write(trans, path, b);
1706 btree_node_interior_verify(c, b);
1710 * We could attempt to avoid the transaction restart, by calling
1711 * bch2_btree_path_upgrade() and allocating more nodes:
1713 if (b->c.level >= as->update_level) {
1714 trace_and_count(c, trans_restart_split_race, trans, _THIS_IP_, b);
1715 return btree_trans_restart(trans, BCH_ERR_transaction_restart_split_race);
1718 return btree_split(as, trans, path, b, keys, flags);
1721 int bch2_btree_split_leaf(struct btree_trans *trans,
1722 struct btree_path *path,
1725 struct btree *b = path_l(path)->b;
1726 struct btree_update *as;
1730 as = bch2_btree_update_start(trans, path, path->level,
1735 ret = btree_split(as, trans, path, b, NULL, flags);
1737 bch2_btree_update_free(as, trans);
1741 bch2_btree_update_done(as, trans);
1743 for (l = path->level + 1; btree_node_intent_locked(path, l) && !ret; l++)
1744 ret = bch2_foreground_maybe_merge(trans, path, l, flags);
1749 int __bch2_foreground_maybe_merge(struct btree_trans *trans,
1750 struct btree_path *path,
1753 enum btree_node_sibling sib)
1755 struct bch_fs *c = trans->c;
1756 struct btree_path *sib_path = NULL, *new_path = NULL;
1757 struct btree_update *as;
1758 struct bkey_format_state new_s;
1759 struct bkey_format new_f;
1760 struct bkey_i delete;
1761 struct btree *b, *m, *n, *prev, *next, *parent;
1762 struct bpos sib_pos;
1764 u64 start_time = local_clock();
1767 BUG_ON(!path->should_be_locked);
1768 BUG_ON(!btree_node_locked(path, level));
1770 b = path->l[level].b;
1772 if ((sib == btree_prev_sib && bpos_eq(b->data->min_key, POS_MIN)) ||
1773 (sib == btree_next_sib && bpos_eq(b->data->max_key, SPOS_MAX))) {
1774 b->sib_u64s[sib] = U16_MAX;
1778 sib_pos = sib == btree_prev_sib
1779 ? bpos_predecessor(b->data->min_key)
1780 : bpos_successor(b->data->max_key);
1782 sib_path = bch2_path_get(trans, path->btree_id, sib_pos,
1783 U8_MAX, level, BTREE_ITER_INTENT, _THIS_IP_);
1784 ret = bch2_btree_path_traverse(trans, sib_path, false);
1788 btree_path_set_should_be_locked(sib_path);
1790 m = sib_path->l[level].b;
1792 if (btree_node_parent(path, b) !=
1793 btree_node_parent(sib_path, m)) {
1794 b->sib_u64s[sib] = U16_MAX;
1798 if (sib == btree_prev_sib) {
1806 if (!bpos_eq(bpos_successor(prev->data->max_key), next->data->min_key)) {
1807 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
1809 bch2_bpos_to_text(&buf1, prev->data->max_key);
1810 bch2_bpos_to_text(&buf2, next->data->min_key);
1812 "%s(): btree topology error:\n"
1813 " prev ends at %s\n"
1814 " next starts at %s",
1815 __func__, buf1.buf, buf2.buf);
1816 printbuf_exit(&buf1);
1817 printbuf_exit(&buf2);
1818 bch2_topology_error(c);
1823 bch2_bkey_format_init(&new_s);
1824 bch2_bkey_format_add_pos(&new_s, prev->data->min_key);
1825 __bch2_btree_calc_format(&new_s, prev);
1826 __bch2_btree_calc_format(&new_s, next);
1827 bch2_bkey_format_add_pos(&new_s, next->data->max_key);
1828 new_f = bch2_bkey_format_done(&new_s);
1830 sib_u64s = btree_node_u64s_with_format(b, &new_f) +
1831 btree_node_u64s_with_format(m, &new_f);
1833 if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1834 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1836 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1839 sib_u64s = min(sib_u64s, btree_max_u64s(c));
1840 sib_u64s = min(sib_u64s, (size_t) U16_MAX - 1);
1841 b->sib_u64s[sib] = sib_u64s;
1843 if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold)
1846 parent = btree_node_parent(path, b);
1847 as = bch2_btree_update_start(trans, path, level, false,
1848 BTREE_INSERT_NOFAIL|
1849 BTREE_INSERT_USE_RESERVE|
1851 ret = PTR_ERR_OR_ZERO(as);
1855 trace_and_count(c, btree_node_merge, c, b);
1857 bch2_btree_interior_update_will_free_node(as, b);
1858 bch2_btree_interior_update_will_free_node(as, m);
1860 n = bch2_btree_node_alloc(as, trans, b->c.level);
1862 SET_BTREE_NODE_SEQ(n->data,
1863 max(BTREE_NODE_SEQ(b->data),
1864 BTREE_NODE_SEQ(m->data)) + 1);
1866 btree_set_min(n, prev->data->min_key);
1867 btree_set_max(n, next->data->max_key);
1869 n->data->format = new_f;
1870 btree_node_set_format(n, new_f);
1872 bch2_btree_sort_into(c, n, prev);
1873 bch2_btree_sort_into(c, n, next);
1875 bch2_btree_build_aux_trees(n);
1876 bch2_btree_update_add_new_node(as, n);
1877 six_unlock_write(&n->c.lock);
1879 new_path = get_unlocked_mut_path(trans, path->btree_id, n->c.level, n->key.k.p);
1880 six_lock_increment(&n->c.lock, SIX_LOCK_intent);
1881 mark_btree_node_locked(trans, new_path, n->c.level, SIX_LOCK_intent);
1882 bch2_btree_path_level_init(trans, new_path, n);
1884 bkey_init(&delete.k);
1885 delete.k.p = prev->key.k.p;
1886 bch2_keylist_add(&as->parent_keys, &delete);
1887 bch2_keylist_add(&as->parent_keys, &n->key);
1889 bch2_trans_verify_paths(trans);
1891 ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1893 goto err_free_update;
1895 bch2_trans_verify_paths(trans);
1897 bch2_btree_update_get_open_buckets(as, n);
1898 bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1900 bch2_btree_node_free_inmem(trans, path, b);
1901 bch2_btree_node_free_inmem(trans, sib_path, m);
1903 bch2_trans_node_add(trans, n);
1905 bch2_trans_verify_paths(trans);
1907 six_unlock_intent(&n->c.lock);
1909 bch2_btree_update_done(as, trans);
1911 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_merge], start_time);
1915 bch2_path_put(trans, new_path, true);
1916 bch2_path_put(trans, sib_path, true);
1917 bch2_trans_verify_locks(trans);
1920 bch2_btree_node_free_never_used(as, trans, n);
1921 bch2_btree_update_free(as, trans);
1926 * bch_btree_node_rewrite - Rewrite/move a btree node
1928 int bch2_btree_node_rewrite(struct btree_trans *trans,
1929 struct btree_iter *iter,
1933 struct bch_fs *c = trans->c;
1934 struct btree_path *new_path = NULL;
1935 struct btree *n, *parent;
1936 struct btree_update *as;
1939 flags |= BTREE_INSERT_NOFAIL;
1941 parent = btree_node_parent(iter->path, b);
1942 as = bch2_btree_update_start(trans, iter->path, b->c.level,
1944 ret = PTR_ERR_OR_ZERO(as);
1948 bch2_btree_interior_update_will_free_node(as, b);
1950 n = bch2_btree_node_alloc_replacement(as, trans, b);
1952 bch2_btree_build_aux_trees(n);
1953 bch2_btree_update_add_new_node(as, n);
1954 six_unlock_write(&n->c.lock);
1956 new_path = get_unlocked_mut_path(trans, iter->btree_id, n->c.level, n->key.k.p);
1957 six_lock_increment(&n->c.lock, SIX_LOCK_intent);
1958 mark_btree_node_locked(trans, new_path, n->c.level, SIX_LOCK_intent);
1959 bch2_btree_path_level_init(trans, new_path, n);
1961 trace_and_count(c, btree_node_rewrite, c, b);
1964 bch2_keylist_add(&as->parent_keys, &n->key);
1965 ret = bch2_btree_insert_node(as, trans, iter->path, parent,
1966 &as->parent_keys, flags);
1970 bch2_btree_set_root(as, trans, iter->path, n);
1973 bch2_btree_update_get_open_buckets(as, n);
1974 bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1976 bch2_btree_node_free_inmem(trans, iter->path, b);
1978 bch2_trans_node_add(trans, n);
1979 six_unlock_intent(&n->c.lock);
1981 bch2_btree_update_done(as, trans);
1984 bch2_path_put(trans, new_path, true);
1985 bch2_btree_path_downgrade(trans, iter->path);
1988 bch2_btree_node_free_never_used(as, trans, n);
1989 bch2_btree_update_free(as, trans);
1993 struct async_btree_rewrite {
1995 struct work_struct work;
1996 struct list_head list;
1997 enum btree_id btree_id;
2003 static int async_btree_node_rewrite_trans(struct btree_trans *trans,
2004 struct async_btree_rewrite *a)
2006 struct bch_fs *c = trans->c;
2007 struct btree_iter iter;
2011 bch2_trans_node_iter_init(trans, &iter, a->btree_id, a->pos,
2012 BTREE_MAX_DEPTH, a->level, 0);
2013 b = bch2_btree_iter_peek_node(&iter);
2014 ret = PTR_ERR_OR_ZERO(b);
2018 if (!b || b->data->keys.seq != a->seq) {
2019 struct printbuf buf = PRINTBUF;
2022 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
2024 prt_str(&buf, "(null");
2025 bch_info(c, "%s: node to rewrite not found:, searching for seq %llu, got\n%s",
2026 __func__, a->seq, buf.buf);
2027 printbuf_exit(&buf);
2031 ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
2033 bch2_trans_iter_exit(trans, &iter);
2038 void async_btree_node_rewrite_work(struct work_struct *work)
2040 struct async_btree_rewrite *a =
2041 container_of(work, struct async_btree_rewrite, work);
2042 struct bch_fs *c = a->c;
2045 ret = bch2_trans_do(c, NULL, NULL, 0,
2046 async_btree_node_rewrite_trans(&trans, a));
2048 bch_err(c, "%s: error %s", __func__, bch2_err_str(ret));
2049 bch2_write_ref_put(c, BCH_WRITE_REF_node_rewrite);
2053 void bch2_btree_node_rewrite_async(struct bch_fs *c, struct btree *b)
2055 struct async_btree_rewrite *a;
2058 a = kmalloc(sizeof(*a), GFP_NOFS);
2060 bch_err(c, "%s: error allocating memory", __func__);
2065 a->btree_id = b->c.btree_id;
2066 a->level = b->c.level;
2067 a->pos = b->key.k.p;
2068 a->seq = b->data->keys.seq;
2069 INIT_WORK(&a->work, async_btree_node_rewrite_work);
2071 if (unlikely(!test_bit(BCH_FS_MAY_GO_RW, &c->flags))) {
2072 mutex_lock(&c->pending_node_rewrites_lock);
2073 list_add(&a->list, &c->pending_node_rewrites);
2074 mutex_unlock(&c->pending_node_rewrites_lock);
2078 if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_node_rewrite)) {
2079 if (test_bit(BCH_FS_STARTED, &c->flags)) {
2080 bch_err(c, "%s: error getting c->writes ref", __func__);
2085 ret = bch2_fs_read_write_early(c);
2087 bch_err(c, "%s: error going read-write: %s",
2088 __func__, bch2_err_str(ret));
2093 bch2_write_ref_get(c, BCH_WRITE_REF_node_rewrite);
2096 queue_work(c->btree_interior_update_worker, &a->work);
2099 void bch2_do_pending_node_rewrites(struct bch_fs *c)
2101 struct async_btree_rewrite *a, *n;
2103 mutex_lock(&c->pending_node_rewrites_lock);
2104 list_for_each_entry_safe(a, n, &c->pending_node_rewrites, list) {
2107 bch2_write_ref_get(c, BCH_WRITE_REF_node_rewrite);
2108 queue_work(c->btree_interior_update_worker, &a->work);
2110 mutex_unlock(&c->pending_node_rewrites_lock);
2113 void bch2_free_pending_node_rewrites(struct bch_fs *c)
2115 struct async_btree_rewrite *a, *n;
2117 mutex_lock(&c->pending_node_rewrites_lock);
2118 list_for_each_entry_safe(a, n, &c->pending_node_rewrites, list) {
2123 mutex_unlock(&c->pending_node_rewrites_lock);
2126 static int __bch2_btree_node_update_key(struct btree_trans *trans,
2127 struct btree_iter *iter,
2128 struct btree *b, struct btree *new_hash,
2129 struct bkey_i *new_key,
2132 struct bch_fs *c = trans->c;
2133 struct btree_iter iter2 = { NULL };
2134 struct btree *parent;
2137 if (!skip_triggers) {
2138 ret = bch2_trans_mark_old(trans, b->c.btree_id, b->c.level + 1,
2139 bkey_i_to_s_c(&b->key), 0);
2143 ret = bch2_trans_mark_new(trans, b->c.btree_id, b->c.level + 1,
2150 bkey_copy(&new_hash->key, new_key);
2151 ret = bch2_btree_node_hash_insert(&c->btree_cache,
2152 new_hash, b->c.level, b->c.btree_id);
2156 parent = btree_node_parent(iter->path, b);
2158 bch2_trans_copy_iter(&iter2, iter);
2160 iter2.path = bch2_btree_path_make_mut(trans, iter2.path,
2161 iter2.flags & BTREE_ITER_INTENT,
2164 BUG_ON(iter2.path->level != b->c.level);
2165 BUG_ON(!bpos_eq(iter2.path->pos, new_key->k.p));
2167 btree_path_set_level_up(trans, iter2.path);
2169 trans->paths_sorted = false;
2171 ret = bch2_btree_iter_traverse(&iter2) ?:
2172 bch2_trans_update(trans, &iter2, new_key, BTREE_TRIGGER_NORUN);
2176 BUG_ON(btree_node_root(c, b) != b);
2178 ret = darray_make_room(&trans->extra_journal_entries,
2179 jset_u64s(new_key->k.u64s));
2183 journal_entry_set((void *) &darray_top(trans->extra_journal_entries),
2184 BCH_JSET_ENTRY_btree_root,
2185 b->c.btree_id, b->c.level,
2186 new_key, new_key->k.u64s);
2187 trans->extra_journal_entries.nr += jset_u64s(new_key->k.u64s);
2190 ret = bch2_trans_commit(trans, NULL, NULL,
2191 BTREE_INSERT_NOFAIL|
2192 BTREE_INSERT_NOCHECK_RW|
2193 BTREE_INSERT_USE_RESERVE|
2194 BTREE_INSERT_JOURNAL_RECLAIM|
2195 JOURNAL_WATERMARK_reserved);
2199 bch2_btree_node_lock_write_nofail(trans, iter->path, &b->c);
2202 mutex_lock(&c->btree_cache.lock);
2203 bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
2204 bch2_btree_node_hash_remove(&c->btree_cache, b);
2206 bkey_copy(&b->key, new_key);
2207 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
2209 mutex_unlock(&c->btree_cache.lock);
2211 bkey_copy(&b->key, new_key);
2214 bch2_btree_node_unlock_write(trans, iter->path, b);
2216 bch2_trans_iter_exit(trans, &iter2);
2220 mutex_lock(&c->btree_cache.lock);
2221 bch2_btree_node_hash_remove(&c->btree_cache, b);
2222 mutex_unlock(&c->btree_cache.lock);
2227 int bch2_btree_node_update_key(struct btree_trans *trans, struct btree_iter *iter,
2228 struct btree *b, struct bkey_i *new_key,
2231 struct bch_fs *c = trans->c;
2232 struct btree *new_hash = NULL;
2233 struct btree_path *path = iter->path;
2237 ret = bch2_btree_path_upgrade(trans, path, b->c.level + 1);
2241 closure_init_stack(&cl);
2244 * check btree_ptr_hash_val() after @b is locked by
2245 * btree_iter_traverse():
2247 if (btree_ptr_hash_val(new_key) != b->hash_val) {
2248 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2250 ret = drop_locks_do(trans, (closure_sync(&cl), 0));
2255 new_hash = bch2_btree_node_mem_alloc(trans, false);
2259 ret = __bch2_btree_node_update_key(trans, iter, b, new_hash,
2260 new_key, skip_triggers);
2264 mutex_lock(&c->btree_cache.lock);
2265 list_move(&new_hash->list, &c->btree_cache.freeable);
2266 mutex_unlock(&c->btree_cache.lock);
2268 six_unlock_write(&new_hash->c.lock);
2269 six_unlock_intent(&new_hash->c.lock);
2272 bch2_btree_cache_cannibalize_unlock(c);
2276 int bch2_btree_node_update_key_get_iter(struct btree_trans *trans,
2277 struct btree *b, struct bkey_i *new_key,
2280 struct btree_iter iter;
2283 bch2_trans_node_iter_init(trans, &iter, b->c.btree_id, b->key.k.p,
2284 BTREE_MAX_DEPTH, b->c.level,
2286 ret = bch2_btree_iter_traverse(&iter);
2290 /* has node been freed? */
2291 if (iter.path->l[b->c.level].b != b) {
2292 /* node has been freed: */
2293 BUG_ON(!btree_node_dying(b));
2297 BUG_ON(!btree_node_hashed(b));
2299 ret = bch2_btree_node_update_key(trans, &iter, b, new_key, skip_triggers);
2301 bch2_trans_iter_exit(trans, &iter);
2308 * Only for filesystem bringup, when first reading the btree roots or allocating
2309 * btree roots when initializing a new filesystem:
2311 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
2313 BUG_ON(btree_node_root(c, b));
2315 bch2_btree_set_root_inmem(c, b);
2318 static int __bch2_btree_root_alloc(struct btree_trans *trans, enum btree_id id)
2320 struct bch_fs *c = trans->c;
2325 closure_init_stack(&cl);
2328 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2332 b = bch2_btree_node_mem_alloc(trans, false);
2333 bch2_btree_cache_cannibalize_unlock(c);
2335 set_btree_node_fake(b);
2336 set_btree_node_need_rewrite(b);
2340 bkey_btree_ptr_init(&b->key);
2341 b->key.k.p = SPOS_MAX;
2342 *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
2344 bch2_bset_init_first(b, &b->data->keys);
2345 bch2_btree_build_aux_trees(b);
2348 btree_set_min(b, POS_MIN);
2349 btree_set_max(b, SPOS_MAX);
2350 b->data->format = bch2_btree_calc_format(b);
2351 btree_node_set_format(b, b->data->format);
2353 ret = bch2_btree_node_hash_insert(&c->btree_cache, b,
2354 b->c.level, b->c.btree_id);
2357 bch2_btree_set_root_inmem(c, b);
2359 six_unlock_write(&b->c.lock);
2360 six_unlock_intent(&b->c.lock);
2364 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
2366 bch2_trans_run(c, __bch2_btree_root_alloc(&trans, id));
2369 void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c)
2371 struct btree_update *as;
2373 mutex_lock(&c->btree_interior_update_lock);
2374 list_for_each_entry(as, &c->btree_interior_update_list, list)
2375 prt_printf(out, "%p m %u w %u r %u j %llu\n",
2379 atomic_read(&as->cl.remaining) & CLOSURE_REMAINING_MASK,
2381 mutex_unlock(&c->btree_interior_update_lock);
2384 static bool bch2_btree_interior_updates_pending(struct bch_fs *c)
2388 mutex_lock(&c->btree_interior_update_lock);
2389 ret = !list_empty(&c->btree_interior_update_list);
2390 mutex_unlock(&c->btree_interior_update_lock);
2395 bool bch2_btree_interior_updates_flush(struct bch_fs *c)
2397 bool ret = bch2_btree_interior_updates_pending(c);
2400 closure_wait_event(&c->btree_interior_update_wait,
2401 !bch2_btree_interior_updates_pending(c));
2405 void bch2_journal_entry_to_btree_root(struct bch_fs *c, struct jset_entry *entry)
2407 struct btree_root *r = &c->btree_roots[entry->btree_id];
2409 mutex_lock(&c->btree_root_lock);
2411 r->level = entry->level;
2413 bkey_copy(&r->key, &entry->start[0]);
2415 mutex_unlock(&c->btree_root_lock);
2419 bch2_btree_roots_to_journal_entries(struct bch_fs *c,
2420 struct jset_entry *start,
2421 struct jset_entry *end)
2423 struct jset_entry *entry;
2424 unsigned long have = 0;
2427 for (entry = start; entry < end; entry = vstruct_next(entry))
2428 if (entry->type == BCH_JSET_ENTRY_btree_root)
2429 __set_bit(entry->btree_id, &have);
2431 mutex_lock(&c->btree_root_lock);
2433 for (i = 0; i < BTREE_ID_NR; i++)
2434 if (c->btree_roots[i].alive && !test_bit(i, &have)) {
2435 journal_entry_set(end,
2436 BCH_JSET_ENTRY_btree_root,
2437 i, c->btree_roots[i].level,
2438 &c->btree_roots[i].key,
2439 c->btree_roots[i].key.k.u64s);
2440 end = vstruct_next(end);
2443 mutex_unlock(&c->btree_root_lock);
2448 void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
2450 if (c->btree_interior_update_worker)
2451 destroy_workqueue(c->btree_interior_update_worker);
2452 mempool_exit(&c->btree_interior_update_pool);
2455 int bch2_fs_btree_interior_update_init(struct bch_fs *c)
2457 mutex_init(&c->btree_reserve_cache_lock);
2458 INIT_LIST_HEAD(&c->btree_interior_update_list);
2459 INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
2460 mutex_init(&c->btree_interior_update_lock);
2461 INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
2463 INIT_LIST_HEAD(&c->pending_node_rewrites);
2464 mutex_init(&c->pending_node_rewrites_lock);
2466 c->btree_interior_update_worker =
2467 alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
2468 if (!c->btree_interior_update_worker)
2469 return -BCH_ERR_ENOMEM_btree_interior_update_worker_init;
2471 if (mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
2472 sizeof(struct btree_update)))
2473 return -BCH_ERR_ENOMEM_btree_interior_update_pool_init;