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 };
249 enum bch_watermark watermark = flags & BCH_WATERMARK_MASK;
250 unsigned nr_reserve = watermark > BCH_WATERMARK_reclaim
255 mutex_lock(&c->btree_reserve_cache_lock);
256 if (c->btree_reserve_cache_nr > nr_reserve) {
257 struct btree_alloc *a =
258 &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
261 bkey_copy(&tmp.k, &a->k);
262 mutex_unlock(&c->btree_reserve_cache_lock);
265 mutex_unlock(&c->btree_reserve_cache_lock);
268 ret = bch2_alloc_sectors_start_trans(trans,
269 c->opts.metadata_target ?:
270 c->opts.foreground_target,
272 writepoint_ptr(&c->btree_write_point),
275 c->opts.metadata_replicas_required,
276 watermark, 0, cl, &wp);
280 if (wp->sectors_free < btree_sectors(c)) {
281 struct open_bucket *ob;
284 open_bucket_for_each(c, &wp->ptrs, ob, i)
285 if (ob->sectors_free < btree_sectors(c))
286 ob->sectors_free = 0;
288 bch2_alloc_sectors_done(c, wp);
292 bkey_btree_ptr_v2_init(&tmp.k);
293 bch2_alloc_sectors_append_ptrs(c, wp, &tmp.k, btree_sectors(c), false);
295 bch2_open_bucket_get(c, wp, &ob);
296 bch2_alloc_sectors_done(c, wp);
298 b = bch2_btree_node_mem_alloc(trans, interior_node);
299 six_unlock_write(&b->c.lock);
300 six_unlock_intent(&b->c.lock);
302 /* we hold cannibalize_lock: */
306 bkey_copy(&b->key, &tmp.k);
312 static struct btree *bch2_btree_node_alloc(struct btree_update *as,
313 struct btree_trans *trans,
316 struct bch_fs *c = as->c;
318 struct prealloc_nodes *p = &as->prealloc_nodes[!!level];
321 BUG_ON(level >= BTREE_MAX_DEPTH);
326 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
327 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
329 set_btree_node_accessed(b);
330 set_btree_node_dirty_acct(c, b);
331 set_btree_node_need_write(b);
333 bch2_bset_init_first(b, &b->data->keys);
335 b->c.btree_id = as->btree_id;
336 b->version_ondisk = c->sb.version;
338 memset(&b->nr, 0, sizeof(b->nr));
339 b->data->magic = cpu_to_le64(bset_magic(c));
340 memset(&b->data->_ptr, 0, sizeof(b->data->_ptr));
342 SET_BTREE_NODE_ID(b->data, as->btree_id);
343 SET_BTREE_NODE_LEVEL(b->data, level);
345 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
346 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(&b->key);
349 bp->v.seq = b->data->keys.seq;
350 bp->v.sectors_written = 0;
353 SET_BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data, true);
355 bch2_btree_build_aux_trees(b);
357 ret = bch2_btree_node_hash_insert(&c->btree_cache, b, level, as->btree_id);
360 trace_and_count(c, btree_node_alloc, c, b);
361 bch2_increment_clock(c, btree_sectors(c), WRITE);
365 static void btree_set_min(struct btree *b, struct bpos pos)
367 if (b->key.k.type == KEY_TYPE_btree_ptr_v2)
368 bkey_i_to_btree_ptr_v2(&b->key)->v.min_key = pos;
369 b->data->min_key = pos;
372 static void btree_set_max(struct btree *b, struct bpos pos)
375 b->data->max_key = pos;
378 static struct btree *bch2_btree_node_alloc_replacement(struct btree_update *as,
379 struct btree_trans *trans,
382 struct btree *n = bch2_btree_node_alloc(as, trans, b->c.level);
383 struct bkey_format format = bch2_btree_calc_format(b);
386 * The keys might expand with the new format - if they wouldn't fit in
387 * the btree node anymore, use the old format for now:
389 if (!bch2_btree_node_format_fits(as->c, b, &format))
392 SET_BTREE_NODE_SEQ(n->data, BTREE_NODE_SEQ(b->data) + 1);
394 btree_set_min(n, b->data->min_key);
395 btree_set_max(n, b->data->max_key);
397 n->data->format = format;
398 btree_node_set_format(n, format);
400 bch2_btree_sort_into(as->c, n, b);
402 btree_node_reset_sib_u64s(n);
406 static struct btree *__btree_root_alloc(struct btree_update *as,
407 struct btree_trans *trans, unsigned level)
409 struct btree *b = bch2_btree_node_alloc(as, trans, level);
411 btree_set_min(b, POS_MIN);
412 btree_set_max(b, SPOS_MAX);
413 b->data->format = bch2_btree_calc_format(b);
415 btree_node_set_format(b, b->data->format);
416 bch2_btree_build_aux_trees(b);
421 static void bch2_btree_reserve_put(struct btree_update *as, struct btree_trans *trans)
423 struct bch_fs *c = as->c;
424 struct prealloc_nodes *p;
426 for (p = as->prealloc_nodes;
427 p < as->prealloc_nodes + ARRAY_SIZE(as->prealloc_nodes);
430 struct btree *b = p->b[--p->nr];
432 mutex_lock(&c->btree_reserve_cache_lock);
434 if (c->btree_reserve_cache_nr <
435 ARRAY_SIZE(c->btree_reserve_cache)) {
436 struct btree_alloc *a =
437 &c->btree_reserve_cache[c->btree_reserve_cache_nr++];
441 bkey_copy(&a->k, &b->key);
443 bch2_open_buckets_put(c, &b->ob);
446 mutex_unlock(&c->btree_reserve_cache_lock);
448 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
449 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
450 __btree_node_free(c, b);
451 six_unlock_write(&b->c.lock);
452 six_unlock_intent(&b->c.lock);
457 static int bch2_btree_reserve_get(struct btree_trans *trans,
458 struct btree_update *as,
459 unsigned nr_nodes[2],
463 struct bch_fs *c = as->c;
468 BUG_ON(nr_nodes[0] + nr_nodes[1] > BTREE_RESERVE_MAX);
471 * Protects reaping from the btree node cache and using the btree node
472 * open bucket reserve:
474 * BTREE_INSERT_NOWAIT only applies to btree node allocation, not
475 * blocking on this lock:
477 ret = bch2_btree_cache_cannibalize_lock(c, cl);
481 for (interior = 0; interior < 2; interior++) {
482 struct prealloc_nodes *p = as->prealloc_nodes + interior;
484 while (p->nr < nr_nodes[interior]) {
485 b = __bch2_btree_node_alloc(trans, &as->disk_res,
486 flags & BTREE_INSERT_NOWAIT ? NULL : cl,
497 bch2_btree_cache_cannibalize_unlock(c);
501 /* Asynchronous interior node update machinery */
503 static void bch2_btree_update_free(struct btree_update *as, struct btree_trans *trans)
505 struct bch_fs *c = as->c;
507 if (as->took_gc_lock)
508 up_read(&c->gc_lock);
509 as->took_gc_lock = false;
511 bch2_journal_preres_put(&c->journal, &as->journal_preres);
513 bch2_journal_pin_drop(&c->journal, &as->journal);
514 bch2_journal_pin_flush(&c->journal, &as->journal);
515 bch2_disk_reservation_put(c, &as->disk_res);
516 bch2_btree_reserve_put(as, trans);
518 bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_total],
521 mutex_lock(&c->btree_interior_update_lock);
522 list_del(&as->unwritten_list);
525 closure_debug_destroy(&as->cl);
526 mempool_free(as, &c->btree_interior_update_pool);
529 * Have to do the wakeup with btree_interior_update_lock still held,
530 * since being on btree_interior_update_list is our ref on @c:
532 closure_wake_up(&c->btree_interior_update_wait);
534 mutex_unlock(&c->btree_interior_update_lock);
537 static void btree_update_add_key(struct btree_update *as,
538 struct keylist *keys, struct btree *b)
540 struct bkey_i *k = &b->key;
542 BUG_ON(bch2_keylist_u64s(keys) + k->k.u64s >
543 ARRAY_SIZE(as->_old_keys));
545 bkey_copy(keys->top, k);
546 bkey_i_to_btree_ptr_v2(keys->top)->v.mem_ptr = b->c.level + 1;
548 bch2_keylist_push(keys);
552 * The transactional part of an interior btree node update, where we journal the
553 * update we did to the interior node and update alloc info:
555 static int btree_update_nodes_written_trans(struct btree_trans *trans,
556 struct btree_update *as)
561 ret = darray_make_room(&trans->extra_journal_entries, as->journal_u64s);
565 memcpy(&darray_top(trans->extra_journal_entries),
567 as->journal_u64s * sizeof(u64));
568 trans->extra_journal_entries.nr += as->journal_u64s;
570 trans->journal_pin = &as->journal;
572 for_each_keylist_key(&as->old_keys, k) {
573 unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr;
575 ret = bch2_trans_mark_old(trans, as->btree_id, level, bkey_i_to_s_c(k), 0);
580 for_each_keylist_key(&as->new_keys, k) {
581 unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr;
583 ret = bch2_trans_mark_new(trans, as->btree_id, level, k, 0);
591 static void btree_update_nodes_written(struct btree_update *as)
593 struct bch_fs *c = as->c;
595 struct btree_trans trans;
600 bch2_trans_init(&trans, c, 0, 512);
602 * If we're already in an error state, it might be because a btree node
603 * was never written, and we might be trying to free that same btree
604 * node here, but it won't have been marked as allocated and we'll see
605 * spurious disk usage inconsistencies in the transactional part below
606 * if we don't skip it:
608 ret = bch2_journal_error(&c->journal);
613 * Wait for any in flight writes to finish before we free the old nodes
616 for (i = 0; i < as->nr_old_nodes; i++) {
619 b = as->old_nodes[i];
621 btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read);
622 seq = b->data ? b->data->keys.seq : 0;
623 six_unlock_read(&b->c.lock);
625 if (seq == as->old_nodes_seq[i])
626 wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight_inner,
627 TASK_UNINTERRUPTIBLE);
631 * We did an update to a parent node where the pointers we added pointed
632 * to child nodes that weren't written yet: now, the child nodes have
633 * been written so we can write out the update to the interior node.
637 * We can't call into journal reclaim here: we'd block on the journal
638 * reclaim lock, but we may need to release the open buckets we have
639 * pinned in order for other btree updates to make forward progress, and
640 * journal reclaim does btree updates when flushing bkey_cached entries,
641 * which may require allocations as well.
643 ret = commit_do(&trans, &as->disk_res, &journal_seq,
644 BCH_WATERMARK_reclaim|
646 BTREE_INSERT_NOCHECK_RW|
647 BTREE_INSERT_JOURNAL_RECLAIM,
648 btree_update_nodes_written_trans(&trans, as));
649 bch2_trans_unlock(&trans);
651 bch2_fs_fatal_err_on(ret && !bch2_journal_error(&c->journal), c,
652 "%s(): error %s", __func__, bch2_err_str(ret));
655 struct btree_path *path;
658 path = get_unlocked_mut_path(&trans, as->btree_id, b->c.level, b->key.k.p);
660 * @b is the node we did the final insert into:
662 * On failure to get a journal reservation, we still have to
663 * unblock the write and allow most of the write path to happen
664 * so that shutdown works, but the i->journal_seq mechanism
665 * won't work to prevent the btree write from being visible (we
666 * didn't get a journal sequence number) - instead
667 * __bch2_btree_node_write() doesn't do the actual write if
668 * we're in journal error state:
672 * Ensure transaction is unlocked before using
673 * btree_node_lock_nopath() (the use of which is always suspect,
674 * we need to work on removing this in the future)
676 * It should be, but get_unlocked_mut_path() -> bch2_path_get()
677 * calls bch2_path_upgrade(), before we call path_make_mut(), so
678 * we may rarely end up with a locked path besides the one we
681 bch2_trans_unlock(&trans);
682 btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_intent);
683 mark_btree_node_locked(&trans, path, b->c.level, SIX_LOCK_intent);
684 path->l[b->c.level].lock_seq = six_lock_seq(&b->c.lock);
685 path->l[b->c.level].b = b;
687 bch2_btree_node_lock_write_nofail(&trans, path, &b->c);
689 mutex_lock(&c->btree_interior_update_lock);
691 list_del(&as->write_blocked_list);
692 if (list_empty(&b->write_blocked))
693 clear_btree_node_write_blocked(b);
696 * Node might have been freed, recheck under
697 * btree_interior_update_lock:
700 struct bset *i = btree_bset_last(b);
703 BUG_ON(!btree_node_dirty(b));
706 i->journal_seq = cpu_to_le64(
708 le64_to_cpu(i->journal_seq)));
710 bch2_btree_add_journal_pin(c, b, journal_seq);
713 * If we didn't get a journal sequence number we
714 * can't write this btree node, because recovery
715 * won't know to ignore this write:
717 set_btree_node_never_write(b);
721 mutex_unlock(&c->btree_interior_update_lock);
723 mark_btree_node_locked_noreset(path, b->c.level, SIX_LOCK_intent);
724 six_unlock_write(&b->c.lock);
726 btree_node_write_if_need(c, b, SIX_LOCK_intent);
727 btree_node_unlock(&trans, path, b->c.level);
728 bch2_path_put(&trans, path, true);
731 bch2_journal_pin_drop(&c->journal, &as->journal);
733 bch2_journal_preres_put(&c->journal, &as->journal_preres);
735 mutex_lock(&c->btree_interior_update_lock);
736 for (i = 0; i < as->nr_new_nodes; i++) {
737 b = as->new_nodes[i];
739 BUG_ON(b->will_make_reachable != (unsigned long) as);
740 b->will_make_reachable = 0;
741 clear_btree_node_will_make_reachable(b);
743 mutex_unlock(&c->btree_interior_update_lock);
745 for (i = 0; i < as->nr_new_nodes; i++) {
746 b = as->new_nodes[i];
748 btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read);
749 btree_node_write_if_need(c, b, SIX_LOCK_read);
750 six_unlock_read(&b->c.lock);
753 for (i = 0; i < as->nr_open_buckets; i++)
754 bch2_open_bucket_put(c, c->open_buckets + as->open_buckets[i]);
756 bch2_btree_update_free(as, &trans);
757 bch2_trans_exit(&trans);
760 static void btree_interior_update_work(struct work_struct *work)
763 container_of(work, struct bch_fs, btree_interior_update_work);
764 struct btree_update *as;
767 mutex_lock(&c->btree_interior_update_lock);
768 as = list_first_entry_or_null(&c->btree_interior_updates_unwritten,
769 struct btree_update, unwritten_list);
770 if (as && !as->nodes_written)
772 mutex_unlock(&c->btree_interior_update_lock);
777 btree_update_nodes_written(as);
781 static void btree_update_set_nodes_written(struct closure *cl)
783 struct btree_update *as = container_of(cl, struct btree_update, cl);
784 struct bch_fs *c = as->c;
786 mutex_lock(&c->btree_interior_update_lock);
787 as->nodes_written = true;
788 mutex_unlock(&c->btree_interior_update_lock);
790 queue_work(c->btree_interior_update_worker, &c->btree_interior_update_work);
794 * We're updating @b with pointers to nodes that haven't finished writing yet:
795 * block @b from being written until @as completes
797 static void btree_update_updated_node(struct btree_update *as, struct btree *b)
799 struct bch_fs *c = as->c;
801 mutex_lock(&c->btree_interior_update_lock);
802 list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
804 BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
805 BUG_ON(!btree_node_dirty(b));
808 as->mode = BTREE_INTERIOR_UPDATING_NODE;
811 set_btree_node_write_blocked(b);
812 list_add(&as->write_blocked_list, &b->write_blocked);
814 mutex_unlock(&c->btree_interior_update_lock);
817 static void btree_update_reparent(struct btree_update *as,
818 struct btree_update *child)
820 struct bch_fs *c = as->c;
822 lockdep_assert_held(&c->btree_interior_update_lock);
825 child->mode = BTREE_INTERIOR_UPDATING_AS;
827 bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal, NULL);
830 static void btree_update_updated_root(struct btree_update *as, struct btree *b)
832 struct bkey_i *insert = &b->key;
833 struct bch_fs *c = as->c;
835 BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
837 BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
838 ARRAY_SIZE(as->journal_entries));
841 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
842 BCH_JSET_ENTRY_btree_root,
843 b->c.btree_id, b->c.level,
844 insert, insert->k.u64s);
846 mutex_lock(&c->btree_interior_update_lock);
847 list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
849 as->mode = BTREE_INTERIOR_UPDATING_ROOT;
850 mutex_unlock(&c->btree_interior_update_lock);
854 * bch2_btree_update_add_new_node:
856 * This causes @as to wait on @b to be written, before it gets to
857 * bch2_btree_update_nodes_written
859 * Additionally, it sets b->will_make_reachable to prevent any additional writes
860 * to @b from happening besides the first until @b is reachable on disk
862 * And it adds @b to the list of @as's new nodes, so that we can update sector
863 * counts in bch2_btree_update_nodes_written:
865 static void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b)
867 struct bch_fs *c = as->c;
869 closure_get(&as->cl);
871 mutex_lock(&c->btree_interior_update_lock);
872 BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes));
873 BUG_ON(b->will_make_reachable);
875 as->new_nodes[as->nr_new_nodes++] = b;
876 b->will_make_reachable = 1UL|(unsigned long) as;
877 set_btree_node_will_make_reachable(b);
879 mutex_unlock(&c->btree_interior_update_lock);
881 btree_update_add_key(as, &as->new_keys, b);
883 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
884 unsigned bytes = vstruct_end(&b->data->keys) - (void *) b->data;
885 unsigned sectors = round_up(bytes, block_bytes(c)) >> 9;
887 bkey_i_to_btree_ptr_v2(&b->key)->v.sectors_written =
888 cpu_to_le16(sectors);
893 * returns true if @b was a new node
895 static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b)
897 struct btree_update *as;
901 mutex_lock(&c->btree_interior_update_lock);
903 * When b->will_make_reachable != 0, it owns a ref on as->cl that's
904 * dropped when it gets written by bch2_btree_complete_write - the
905 * xchg() is for synchronization with bch2_btree_complete_write:
907 v = xchg(&b->will_make_reachable, 0);
908 clear_btree_node_will_make_reachable(b);
909 as = (struct btree_update *) (v & ~1UL);
912 mutex_unlock(&c->btree_interior_update_lock);
916 for (i = 0; i < as->nr_new_nodes; i++)
917 if (as->new_nodes[i] == b)
922 array_remove_item(as->new_nodes, as->nr_new_nodes, i);
923 mutex_unlock(&c->btree_interior_update_lock);
926 closure_put(&as->cl);
929 static void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b)
932 as->open_buckets[as->nr_open_buckets++] =
937 * @b is being split/rewritten: it may have pointers to not-yet-written btree
938 * nodes and thus outstanding btree_updates - redirect @b's
939 * btree_updates to point to this btree_update:
941 static void bch2_btree_interior_update_will_free_node(struct btree_update *as,
944 struct bch_fs *c = as->c;
945 struct btree_update *p, *n;
946 struct btree_write *w;
948 set_btree_node_dying(b);
950 if (btree_node_fake(b))
953 mutex_lock(&c->btree_interior_update_lock);
956 * Does this node have any btree_update operations preventing
957 * it from being written?
959 * If so, redirect them to point to this btree_update: we can
960 * write out our new nodes, but we won't make them visible until those
961 * operations complete
963 list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) {
964 list_del_init(&p->write_blocked_list);
965 btree_update_reparent(as, p);
968 * for flush_held_btree_writes() waiting on updates to flush or
969 * nodes to be writeable:
971 closure_wake_up(&c->btree_interior_update_wait);
974 clear_btree_node_dirty_acct(c, b);
975 clear_btree_node_need_write(b);
976 clear_btree_node_write_blocked(b);
979 * Does this node have unwritten data that has a pin on the journal?
981 * If so, transfer that pin to the btree_update operation -
982 * note that if we're freeing multiple nodes, we only need to keep the
983 * oldest pin of any of the nodes we're freeing. We'll release the pin
984 * when the new nodes are persistent and reachable on disk:
986 w = btree_current_write(b);
987 bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
988 bch2_journal_pin_drop(&c->journal, &w->journal);
990 w = btree_prev_write(b);
991 bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
992 bch2_journal_pin_drop(&c->journal, &w->journal);
994 mutex_unlock(&c->btree_interior_update_lock);
997 * Is this a node that isn't reachable on disk yet?
999 * Nodes that aren't reachable yet have writes blocked until they're
1000 * reachable - now that we've cancelled any pending writes and moved
1001 * things waiting on that write to wait on this update, we can drop this
1002 * node from the list of nodes that the other update is making
1003 * reachable, prior to freeing it:
1005 btree_update_drop_new_node(c, b);
1007 btree_update_add_key(as, &as->old_keys, b);
1009 as->old_nodes[as->nr_old_nodes] = b;
1010 as->old_nodes_seq[as->nr_old_nodes] = b->data->keys.seq;
1014 static void bch2_btree_update_done(struct btree_update *as, struct btree_trans *trans)
1016 struct bch_fs *c = as->c;
1017 u64 start_time = as->start_time;
1019 BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE);
1021 if (as->took_gc_lock)
1022 up_read(&as->c->gc_lock);
1023 as->took_gc_lock = false;
1025 bch2_btree_reserve_put(as, trans);
1027 continue_at(&as->cl, btree_update_set_nodes_written,
1028 as->c->btree_interior_update_worker);
1030 bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_foreground],
1034 static struct btree_update *
1035 bch2_btree_update_start(struct btree_trans *trans, struct btree_path *path,
1036 unsigned level, bool split, unsigned flags)
1038 struct bch_fs *c = trans->c;
1039 struct btree_update *as;
1040 u64 start_time = local_clock();
1041 int disk_res_flags = (flags & BTREE_INSERT_NOFAIL)
1042 ? BCH_DISK_RESERVATION_NOFAIL : 0;
1043 unsigned nr_nodes[2] = { 0, 0 };
1044 unsigned update_level = level;
1045 enum bch_watermark watermark = flags & BCH_WATERMARK_MASK;
1046 unsigned journal_flags = 0;
1048 u32 restart_count = trans->restart_count;
1050 BUG_ON(!path->should_be_locked);
1052 if (watermark == BCH_WATERMARK_copygc)
1053 watermark = BCH_WATERMARK_btree_copygc;
1054 if (watermark < BCH_WATERMARK_btree)
1055 watermark = BCH_WATERMARK_btree;
1057 flags &= ~BCH_WATERMARK_MASK;
1060 if (flags & BTREE_INSERT_JOURNAL_RECLAIM)
1061 journal_flags |= JOURNAL_RES_GET_NONBLOCK;
1062 journal_flags |= watermark;
1065 nr_nodes[!!update_level] += 1 + split;
1068 ret = bch2_btree_path_upgrade(trans, path, update_level + 1);
1070 return ERR_PTR(ret);
1072 if (!btree_path_node(path, update_level)) {
1073 /* Allocating new root? */
1074 nr_nodes[1] += split;
1075 update_level = BTREE_MAX_DEPTH;
1079 if (bch2_btree_node_insert_fits(c, path->l[update_level].b,
1080 BKEY_BTREE_PTR_U64s_MAX * (1 + split)))
1083 split = path->l[update_level].b->nr.live_u64s > BTREE_SPLIT_THRESHOLD(c);
1086 if (flags & BTREE_INSERT_GC_LOCK_HELD)
1087 lockdep_assert_held(&c->gc_lock);
1088 else if (!down_read_trylock(&c->gc_lock)) {
1089 ret = drop_locks_do(trans, (down_read(&c->gc_lock), 0));
1091 up_read(&c->gc_lock);
1092 return ERR_PTR(ret);
1096 as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOFS);
1097 memset(as, 0, sizeof(*as));
1098 closure_init(&as->cl, NULL);
1100 as->start_time = start_time;
1101 as->mode = BTREE_INTERIOR_NO_UPDATE;
1102 as->took_gc_lock = !(flags & BTREE_INSERT_GC_LOCK_HELD);
1103 as->btree_id = path->btree_id;
1104 as->update_level = update_level;
1105 INIT_LIST_HEAD(&as->list);
1106 INIT_LIST_HEAD(&as->unwritten_list);
1107 INIT_LIST_HEAD(&as->write_blocked_list);
1108 bch2_keylist_init(&as->old_keys, as->_old_keys);
1109 bch2_keylist_init(&as->new_keys, as->_new_keys);
1110 bch2_keylist_init(&as->parent_keys, as->inline_keys);
1112 mutex_lock(&c->btree_interior_update_lock);
1113 list_add_tail(&as->list, &c->btree_interior_update_list);
1114 mutex_unlock(&c->btree_interior_update_lock);
1117 * We don't want to allocate if we're in an error state, that can cause
1118 * deadlock on emergency shutdown due to open buckets getting stuck in
1119 * the btree_reserve_cache after allocator shutdown has cleared it out.
1120 * This check needs to come after adding us to the btree_interior_update
1121 * list but before calling bch2_btree_reserve_get, to synchronize with
1122 * __bch2_fs_read_only().
1124 ret = bch2_journal_error(&c->journal);
1128 ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
1129 BTREE_UPDATE_JOURNAL_RES,
1130 journal_flags|JOURNAL_RES_GET_NONBLOCK);
1132 if (flags & BTREE_INSERT_JOURNAL_RECLAIM) {
1133 ret = -BCH_ERR_journal_reclaim_would_deadlock;
1137 ret = drop_locks_do(trans,
1138 bch2_journal_preres_get(&c->journal, &as->journal_preres,
1139 BTREE_UPDATE_JOURNAL_RES,
1141 if (ret == -BCH_ERR_journal_preres_get_blocked) {
1142 trace_and_count(c, trans_restart_journal_preres_get, trans, _RET_IP_, journal_flags);
1143 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_journal_preres_get);
1149 ret = bch2_disk_reservation_get(c, &as->disk_res,
1150 (nr_nodes[0] + nr_nodes[1]) * btree_sectors(c),
1151 c->opts.metadata_replicas,
1156 ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, NULL);
1157 if (bch2_err_matches(ret, ENOSPC) ||
1158 bch2_err_matches(ret, ENOMEM)) {
1162 * XXX: this should probably be a separate BTREE_INSERT_NONBLOCK
1165 if (bch2_err_matches(ret, ENOSPC) &&
1166 (flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
1167 watermark != BCH_WATERMARK_reclaim) {
1168 ret = -BCH_ERR_journal_reclaim_would_deadlock;
1172 closure_init_stack(&cl);
1175 ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, &cl);
1177 bch2_trans_unlock(trans);
1179 } while (bch2_err_matches(ret, BCH_ERR_operation_blocked));
1183 trace_and_count(c, btree_reserve_get_fail, trans->fn,
1184 _RET_IP_, nr_nodes[0] + nr_nodes[1], ret);
1188 ret = bch2_trans_relock(trans);
1192 bch2_trans_verify_not_restarted(trans, restart_count);
1195 bch2_btree_update_free(as, trans);
1196 return ERR_PTR(ret);
1199 /* Btree root updates: */
1201 static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
1203 /* Root nodes cannot be reaped */
1204 mutex_lock(&c->btree_cache.lock);
1205 list_del_init(&b->list);
1206 mutex_unlock(&c->btree_cache.lock);
1208 mutex_lock(&c->btree_root_lock);
1209 BUG_ON(btree_node_root(c, b) &&
1210 (b->c.level < btree_node_root(c, b)->c.level ||
1211 !btree_node_dying(btree_node_root(c, b))));
1213 bch2_btree_id_root(c, b->c.btree_id)->b = b;
1214 mutex_unlock(&c->btree_root_lock);
1216 bch2_recalc_btree_reserve(c);
1220 * bch_btree_set_root - update the root in memory and on disk
1222 * To ensure forward progress, the current task must not be holding any
1223 * btree node write locks. However, you must hold an intent lock on the
1226 * Note: This allocates a journal entry but doesn't add any keys to
1227 * it. All the btree roots are part of every journal write, so there
1228 * is nothing new to be done. This just guarantees that there is a
1231 static void bch2_btree_set_root(struct btree_update *as,
1232 struct btree_trans *trans,
1233 struct btree_path *path,
1236 struct bch_fs *c = as->c;
1239 trace_and_count(c, btree_node_set_root, c, b);
1241 old = btree_node_root(c, b);
1244 * Ensure no one is using the old root while we switch to the
1247 bch2_btree_node_lock_write_nofail(trans, path, &old->c);
1249 bch2_btree_set_root_inmem(c, b);
1251 btree_update_updated_root(as, b);
1254 * Unlock old root after new root is visible:
1256 * The new root isn't persistent, but that's ok: we still have
1257 * an intent lock on the new root, and any updates that would
1258 * depend on the new root would have to update the new root.
1260 bch2_btree_node_unlock_write(trans, path, old);
1263 /* Interior node updates: */
1265 static void bch2_insert_fixup_btree_ptr(struct btree_update *as,
1266 struct btree_trans *trans,
1267 struct btree_path *path,
1269 struct btree_node_iter *node_iter,
1270 struct bkey_i *insert)
1272 struct bch_fs *c = as->c;
1273 struct bkey_packed *k;
1274 struct printbuf buf = PRINTBUF;
1275 unsigned long old, new, v;
1277 BUG_ON(insert->k.type == KEY_TYPE_btree_ptr_v2 &&
1278 !btree_ptr_sectors_written(insert));
1280 if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)))
1281 bch2_journal_key_overwritten(c, b->c.btree_id, b->c.level, insert->k.p);
1283 if (bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1284 btree_node_type(b), WRITE, &buf) ?:
1285 bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert), &buf)) {
1286 printbuf_reset(&buf);
1287 prt_printf(&buf, "inserting invalid bkey\n ");
1288 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
1289 prt_printf(&buf, "\n ");
1290 bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1291 btree_node_type(b), WRITE, &buf);
1292 bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert), &buf);
1294 bch2_fs_inconsistent(c, "%s", buf.buf);
1298 BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1299 ARRAY_SIZE(as->journal_entries));
1302 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
1303 BCH_JSET_ENTRY_btree_keys,
1304 b->c.btree_id, b->c.level,
1305 insert, insert->k.u64s);
1307 while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1308 bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1309 bch2_btree_node_iter_advance(node_iter, b);
1311 bch2_btree_bset_insert_key(trans, path, b, node_iter, insert);
1312 set_btree_node_dirty_acct(c, b);
1314 v = READ_ONCE(b->flags);
1318 new &= ~BTREE_WRITE_TYPE_MASK;
1319 new |= BTREE_WRITE_interior;
1320 new |= 1 << BTREE_NODE_need_write;
1321 } while ((v = cmpxchg(&b->flags, old, new)) != old);
1323 printbuf_exit(&buf);
1327 __bch2_btree_insert_keys_interior(struct btree_update *as,
1328 struct btree_trans *trans,
1329 struct btree_path *path,
1331 struct btree_node_iter node_iter,
1332 struct keylist *keys)
1334 struct bkey_i *insert = bch2_keylist_front(keys);
1335 struct bkey_packed *k;
1337 BUG_ON(btree_node_type(b) != BKEY_TYPE_btree);
1339 while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1340 (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0))
1343 while (!bch2_keylist_empty(keys)) {
1344 struct bkey_i *k = bch2_keylist_front(keys);
1346 if (bpos_gt(k->k.p, b->key.k.p))
1349 bch2_insert_fixup_btree_ptr(as, trans, path, b, &node_iter, k);
1350 bch2_keylist_pop_front(keys);
1355 * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1358 static void __btree_split_node(struct btree_update *as,
1359 struct btree_trans *trans,
1363 struct bkey_packed *k;
1364 struct bpos n1_pos = POS_MIN;
1365 struct btree_node_iter iter;
1366 struct bset *bsets[2];
1367 struct bkey_format_state format[2];
1368 struct bkey_packed *out[2];
1370 unsigned u64s, n1_u64s = (b->nr.live_u64s * 3) / 5;
1373 for (i = 0; i < 2; i++) {
1374 BUG_ON(n[i]->nsets != 1);
1376 bsets[i] = btree_bset_first(n[i]);
1377 out[i] = bsets[i]->start;
1379 SET_BTREE_NODE_SEQ(n[i]->data, BTREE_NODE_SEQ(b->data) + 1);
1380 bch2_bkey_format_init(&format[i]);
1384 for_each_btree_node_key(b, k, &iter) {
1385 if (bkey_deleted(k))
1388 i = u64s >= n1_u64s;
1390 uk = bkey_unpack_key(b, k);
1393 bch2_bkey_format_add_key(&format[i], &uk);
1396 btree_set_min(n[0], b->data->min_key);
1397 btree_set_max(n[0], n1_pos);
1398 btree_set_min(n[1], bpos_successor(n1_pos));
1399 btree_set_max(n[1], b->data->max_key);
1401 for (i = 0; i < 2; i++) {
1402 bch2_bkey_format_add_pos(&format[i], n[i]->data->min_key);
1403 bch2_bkey_format_add_pos(&format[i], n[i]->data->max_key);
1405 n[i]->data->format = bch2_bkey_format_done(&format[i]);
1406 btree_node_set_format(n[i], n[i]->data->format);
1410 for_each_btree_node_key(b, k, &iter) {
1411 if (bkey_deleted(k))
1414 i = u64s >= n1_u64s;
1417 if (bch2_bkey_transform(&n[i]->format, out[i], bkey_packed(k)
1418 ? &b->format: &bch2_bkey_format_current, k))
1419 out[i]->format = KEY_FORMAT_LOCAL_BTREE;
1421 bch2_bkey_unpack(b, (void *) out[i], k);
1423 out[i]->needs_whiteout = false;
1425 btree_keys_account_key_add(&n[i]->nr, 0, out[i]);
1426 out[i] = bkey_p_next(out[i]);
1429 for (i = 0; i < 2; i++) {
1430 bsets[i]->u64s = cpu_to_le16((u64 *) out[i] - bsets[i]->_data);
1432 BUG_ON(!bsets[i]->u64s);
1434 set_btree_bset_end(n[i], n[i]->set);
1436 btree_node_reset_sib_u64s(n[i]);
1438 bch2_verify_btree_nr_keys(n[i]);
1441 btree_node_interior_verify(as->c, n[i]);
1446 * For updates to interior nodes, we've got to do the insert before we split
1447 * because the stuff we're inserting has to be inserted atomically. Post split,
1448 * the keys might have to go in different nodes and the split would no longer be
1451 * Worse, if the insert is from btree node coalescing, if we do the insert after
1452 * we do the split (and pick the pivot) - the pivot we pick might be between
1453 * nodes that were coalesced, and thus in the middle of a child node post
1456 static void btree_split_insert_keys(struct btree_update *as,
1457 struct btree_trans *trans,
1458 struct btree_path *path,
1460 struct keylist *keys)
1462 if (!bch2_keylist_empty(keys) &&
1463 bpos_le(bch2_keylist_front(keys)->k.p, b->data->max_key)) {
1464 struct btree_node_iter node_iter;
1466 bch2_btree_node_iter_init(&node_iter, b, &bch2_keylist_front(keys)->k.p);
1468 __bch2_btree_insert_keys_interior(as, trans, path, b, node_iter, keys);
1470 btree_node_interior_verify(as->c, b);
1474 static int btree_split(struct btree_update *as, struct btree_trans *trans,
1475 struct btree_path *path, struct btree *b,
1476 struct keylist *keys, unsigned flags)
1478 struct bch_fs *c = as->c;
1479 struct btree *parent = btree_node_parent(path, b);
1480 struct btree *n1, *n2 = NULL, *n3 = NULL;
1481 struct btree_path *path1 = NULL, *path2 = NULL;
1482 u64 start_time = local_clock();
1485 BUG_ON(!parent && (b != btree_node_root(c, b)));
1486 BUG_ON(parent && !btree_node_intent_locked(path, b->c.level + 1));
1488 bch2_btree_interior_update_will_free_node(as, b);
1490 if (b->nr.live_u64s > BTREE_SPLIT_THRESHOLD(c)) {
1493 trace_and_count(c, btree_node_split, c, b);
1495 n[0] = n1 = bch2_btree_node_alloc(as, trans, b->c.level);
1496 n[1] = n2 = bch2_btree_node_alloc(as, trans, b->c.level);
1498 __btree_split_node(as, trans, b, n);
1501 btree_split_insert_keys(as, trans, path, n1, keys);
1502 btree_split_insert_keys(as, trans, path, n2, keys);
1503 BUG_ON(!bch2_keylist_empty(keys));
1506 bch2_btree_build_aux_trees(n2);
1507 bch2_btree_build_aux_trees(n1);
1509 bch2_btree_update_add_new_node(as, n1);
1510 bch2_btree_update_add_new_node(as, n2);
1511 six_unlock_write(&n2->c.lock);
1512 six_unlock_write(&n1->c.lock);
1514 path1 = get_unlocked_mut_path(trans, path->btree_id, n1->c.level, n1->key.k.p);
1515 six_lock_increment(&n1->c.lock, SIX_LOCK_intent);
1516 mark_btree_node_locked(trans, path1, n1->c.level, SIX_LOCK_intent);
1517 bch2_btree_path_level_init(trans, path1, n1);
1519 path2 = get_unlocked_mut_path(trans, path->btree_id, n2->c.level, n2->key.k.p);
1520 six_lock_increment(&n2->c.lock, SIX_LOCK_intent);
1521 mark_btree_node_locked(trans, path2, n2->c.level, SIX_LOCK_intent);
1522 bch2_btree_path_level_init(trans, path2, n2);
1525 * Note that on recursive parent_keys == keys, so we
1526 * can't start adding new keys to parent_keys before emptying it
1527 * out (which we did with btree_split_insert_keys() above)
1529 bch2_keylist_add(&as->parent_keys, &n1->key);
1530 bch2_keylist_add(&as->parent_keys, &n2->key);
1533 /* Depth increases, make a new root */
1534 n3 = __btree_root_alloc(as, trans, b->c.level + 1);
1536 bch2_btree_update_add_new_node(as, n3);
1537 six_unlock_write(&n3->c.lock);
1539 path2->locks_want++;
1540 BUG_ON(btree_node_locked(path2, n3->c.level));
1541 six_lock_increment(&n3->c.lock, SIX_LOCK_intent);
1542 mark_btree_node_locked(trans, path2, n3->c.level, SIX_LOCK_intent);
1543 bch2_btree_path_level_init(trans, path2, n3);
1545 n3->sib_u64s[0] = U16_MAX;
1546 n3->sib_u64s[1] = U16_MAX;
1548 btree_split_insert_keys(as, trans, path, n3, &as->parent_keys);
1551 trace_and_count(c, btree_node_compact, c, b);
1553 n1 = bch2_btree_node_alloc_replacement(as, trans, b);
1556 btree_split_insert_keys(as, trans, path, n1, keys);
1557 BUG_ON(!bch2_keylist_empty(keys));
1560 bch2_btree_build_aux_trees(n1);
1561 bch2_btree_update_add_new_node(as, n1);
1562 six_unlock_write(&n1->c.lock);
1564 path1 = get_unlocked_mut_path(trans, path->btree_id, n1->c.level, n1->key.k.p);
1565 six_lock_increment(&n1->c.lock, SIX_LOCK_intent);
1566 mark_btree_node_locked(trans, path1, n1->c.level, SIX_LOCK_intent);
1567 bch2_btree_path_level_init(trans, path1, n1);
1570 bch2_keylist_add(&as->parent_keys, &n1->key);
1573 /* New nodes all written, now make them visible: */
1576 /* Split a non root node */
1577 ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1581 bch2_btree_set_root(as, trans, path, n3);
1583 /* Root filled up but didn't need to be split */
1584 bch2_btree_set_root(as, trans, path, n1);
1588 bch2_btree_update_get_open_buckets(as, n3);
1589 bch2_btree_node_write(c, n3, SIX_LOCK_intent, 0);
1592 bch2_btree_update_get_open_buckets(as, n2);
1593 bch2_btree_node_write(c, n2, SIX_LOCK_intent, 0);
1595 bch2_btree_update_get_open_buckets(as, n1);
1596 bch2_btree_node_write(c, n1, SIX_LOCK_intent, 0);
1599 * The old node must be freed (in memory) _before_ unlocking the new
1600 * nodes - else another thread could re-acquire a read lock on the old
1601 * node after another thread has locked and updated the new node, thus
1602 * seeing stale data:
1604 bch2_btree_node_free_inmem(trans, path, b);
1607 bch2_trans_node_add(trans, n3);
1609 bch2_trans_node_add(trans, n2);
1610 bch2_trans_node_add(trans, n1);
1613 six_unlock_intent(&n3->c.lock);
1615 six_unlock_intent(&n2->c.lock);
1616 six_unlock_intent(&n1->c.lock);
1619 __bch2_btree_path_unlock(trans, path2);
1620 bch2_path_put(trans, path2, true);
1623 __bch2_btree_path_unlock(trans, path1);
1624 bch2_path_put(trans, path1, true);
1627 bch2_trans_verify_locks(trans);
1629 bch2_time_stats_update(&c->times[n2
1630 ? BCH_TIME_btree_node_split
1631 : BCH_TIME_btree_node_compact],
1636 bch2_btree_node_free_never_used(as, trans, n3);
1638 bch2_btree_node_free_never_used(as, trans, n2);
1639 bch2_btree_node_free_never_used(as, trans, n1);
1644 bch2_btree_insert_keys_interior(struct btree_update *as,
1645 struct btree_trans *trans,
1646 struct btree_path *path,
1648 struct keylist *keys)
1650 struct btree_path *linked;
1652 __bch2_btree_insert_keys_interior(as, trans, path, b,
1653 path->l[b->c.level].iter, keys);
1655 btree_update_updated_node(as, b);
1657 trans_for_each_path_with_node(trans, b, linked)
1658 bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b);
1660 bch2_trans_verify_paths(trans);
1664 * bch_btree_insert_node - insert bkeys into a given btree node
1666 * @iter: btree iterator
1667 * @keys: list of keys to insert
1668 * @hook: insert callback
1669 * @persistent: if not null, @persistent will wait on journal write
1671 * Inserts as many keys as it can into a given btree node, splitting it if full.
1672 * If a split occurred, this function will return early. This can only happen
1673 * for leaf nodes -- inserts into interior nodes have to be atomic.
1675 static int bch2_btree_insert_node(struct btree_update *as, struct btree_trans *trans,
1676 struct btree_path *path, struct btree *b,
1677 struct keylist *keys, unsigned flags)
1679 struct bch_fs *c = as->c;
1680 int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1681 int old_live_u64s = b->nr.live_u64s;
1682 int live_u64s_added, u64s_added;
1685 lockdep_assert_held(&c->gc_lock);
1686 BUG_ON(!btree_node_intent_locked(path, b->c.level));
1687 BUG_ON(!b->c.level);
1688 BUG_ON(!as || as->b);
1689 bch2_verify_keylist_sorted(keys);
1691 ret = bch2_btree_node_lock_write(trans, path, &b->c);
1695 bch2_btree_node_prep_for_write(trans, path, b);
1697 if (!bch2_btree_node_insert_fits(c, b, bch2_keylist_u64s(keys))) {
1698 bch2_btree_node_unlock_write(trans, path, b);
1702 btree_node_interior_verify(c, b);
1704 bch2_btree_insert_keys_interior(as, trans, path, b, keys);
1706 live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1707 u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1709 if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1710 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1711 if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1712 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1714 if (u64s_added > live_u64s_added &&
1715 bch2_maybe_compact_whiteouts(c, b))
1716 bch2_trans_node_reinit_iter(trans, b);
1718 bch2_btree_node_unlock_write(trans, path, b);
1720 btree_node_interior_verify(c, b);
1724 * We could attempt to avoid the transaction restart, by calling
1725 * bch2_btree_path_upgrade() and allocating more nodes:
1727 if (b->c.level >= as->update_level) {
1728 trace_and_count(c, trans_restart_split_race, trans, _THIS_IP_, b);
1729 return btree_trans_restart(trans, BCH_ERR_transaction_restart_split_race);
1732 return btree_split(as, trans, path, b, keys, flags);
1735 int bch2_btree_split_leaf(struct btree_trans *trans,
1736 struct btree_path *path,
1739 struct btree *b = path_l(path)->b;
1740 struct btree_update *as;
1744 as = bch2_btree_update_start(trans, path, path->level,
1749 ret = btree_split(as, trans, path, b, NULL, flags);
1751 bch2_btree_update_free(as, trans);
1755 bch2_btree_update_done(as, trans);
1757 for (l = path->level + 1; btree_node_intent_locked(path, l) && !ret; l++)
1758 ret = bch2_foreground_maybe_merge(trans, path, l, flags);
1763 int __bch2_foreground_maybe_merge(struct btree_trans *trans,
1764 struct btree_path *path,
1767 enum btree_node_sibling sib)
1769 struct bch_fs *c = trans->c;
1770 struct btree_path *sib_path = NULL, *new_path = NULL;
1771 struct btree_update *as;
1772 struct bkey_format_state new_s;
1773 struct bkey_format new_f;
1774 struct bkey_i delete;
1775 struct btree *b, *m, *n, *prev, *next, *parent;
1776 struct bpos sib_pos;
1778 u64 start_time = local_clock();
1781 BUG_ON(!path->should_be_locked);
1782 BUG_ON(!btree_node_locked(path, level));
1784 b = path->l[level].b;
1786 if ((sib == btree_prev_sib && bpos_eq(b->data->min_key, POS_MIN)) ||
1787 (sib == btree_next_sib && bpos_eq(b->data->max_key, SPOS_MAX))) {
1788 b->sib_u64s[sib] = U16_MAX;
1792 sib_pos = sib == btree_prev_sib
1793 ? bpos_predecessor(b->data->min_key)
1794 : bpos_successor(b->data->max_key);
1796 sib_path = bch2_path_get(trans, path->btree_id, sib_pos,
1797 U8_MAX, level, BTREE_ITER_INTENT, _THIS_IP_);
1798 ret = bch2_btree_path_traverse(trans, sib_path, false);
1802 btree_path_set_should_be_locked(sib_path);
1804 m = sib_path->l[level].b;
1806 if (btree_node_parent(path, b) !=
1807 btree_node_parent(sib_path, m)) {
1808 b->sib_u64s[sib] = U16_MAX;
1812 if (sib == btree_prev_sib) {
1820 if (!bpos_eq(bpos_successor(prev->data->max_key), next->data->min_key)) {
1821 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
1823 bch2_bpos_to_text(&buf1, prev->data->max_key);
1824 bch2_bpos_to_text(&buf2, next->data->min_key);
1826 "%s(): btree topology error:\n"
1827 " prev ends at %s\n"
1828 " next starts at %s",
1829 __func__, buf1.buf, buf2.buf);
1830 printbuf_exit(&buf1);
1831 printbuf_exit(&buf2);
1832 bch2_topology_error(c);
1837 bch2_bkey_format_init(&new_s);
1838 bch2_bkey_format_add_pos(&new_s, prev->data->min_key);
1839 __bch2_btree_calc_format(&new_s, prev);
1840 __bch2_btree_calc_format(&new_s, next);
1841 bch2_bkey_format_add_pos(&new_s, next->data->max_key);
1842 new_f = bch2_bkey_format_done(&new_s);
1844 sib_u64s = btree_node_u64s_with_format(b, &new_f) +
1845 btree_node_u64s_with_format(m, &new_f);
1847 if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1848 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1850 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1853 sib_u64s = min(sib_u64s, btree_max_u64s(c));
1854 sib_u64s = min(sib_u64s, (size_t) U16_MAX - 1);
1855 b->sib_u64s[sib] = sib_u64s;
1857 if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold)
1860 parent = btree_node_parent(path, b);
1861 as = bch2_btree_update_start(trans, path, level, false,
1862 BTREE_INSERT_NOFAIL|flags);
1863 ret = PTR_ERR_OR_ZERO(as);
1867 trace_and_count(c, btree_node_merge, c, b);
1869 bch2_btree_interior_update_will_free_node(as, b);
1870 bch2_btree_interior_update_will_free_node(as, m);
1872 n = bch2_btree_node_alloc(as, trans, b->c.level);
1874 SET_BTREE_NODE_SEQ(n->data,
1875 max(BTREE_NODE_SEQ(b->data),
1876 BTREE_NODE_SEQ(m->data)) + 1);
1878 btree_set_min(n, prev->data->min_key);
1879 btree_set_max(n, next->data->max_key);
1881 n->data->format = new_f;
1882 btree_node_set_format(n, new_f);
1884 bch2_btree_sort_into(c, n, prev);
1885 bch2_btree_sort_into(c, n, next);
1887 bch2_btree_build_aux_trees(n);
1888 bch2_btree_update_add_new_node(as, n);
1889 six_unlock_write(&n->c.lock);
1891 new_path = get_unlocked_mut_path(trans, path->btree_id, n->c.level, n->key.k.p);
1892 six_lock_increment(&n->c.lock, SIX_LOCK_intent);
1893 mark_btree_node_locked(trans, new_path, n->c.level, SIX_LOCK_intent);
1894 bch2_btree_path_level_init(trans, new_path, n);
1896 bkey_init(&delete.k);
1897 delete.k.p = prev->key.k.p;
1898 bch2_keylist_add(&as->parent_keys, &delete);
1899 bch2_keylist_add(&as->parent_keys, &n->key);
1901 bch2_trans_verify_paths(trans);
1903 ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1905 goto err_free_update;
1907 bch2_trans_verify_paths(trans);
1909 bch2_btree_update_get_open_buckets(as, n);
1910 bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1912 bch2_btree_node_free_inmem(trans, path, b);
1913 bch2_btree_node_free_inmem(trans, sib_path, m);
1915 bch2_trans_node_add(trans, n);
1917 bch2_trans_verify_paths(trans);
1919 six_unlock_intent(&n->c.lock);
1921 bch2_btree_update_done(as, trans);
1923 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_merge], start_time);
1927 bch2_path_put(trans, new_path, true);
1928 bch2_path_put(trans, sib_path, true);
1929 bch2_trans_verify_locks(trans);
1932 bch2_btree_node_free_never_used(as, trans, n);
1933 bch2_btree_update_free(as, trans);
1938 * bch_btree_node_rewrite - Rewrite/move a btree node
1940 int bch2_btree_node_rewrite(struct btree_trans *trans,
1941 struct btree_iter *iter,
1945 struct bch_fs *c = trans->c;
1946 struct btree_path *new_path = NULL;
1947 struct btree *n, *parent;
1948 struct btree_update *as;
1951 flags |= BTREE_INSERT_NOFAIL;
1953 parent = btree_node_parent(iter->path, b);
1954 as = bch2_btree_update_start(trans, iter->path, b->c.level,
1956 ret = PTR_ERR_OR_ZERO(as);
1960 bch2_btree_interior_update_will_free_node(as, b);
1962 n = bch2_btree_node_alloc_replacement(as, trans, b);
1964 bch2_btree_build_aux_trees(n);
1965 bch2_btree_update_add_new_node(as, n);
1966 six_unlock_write(&n->c.lock);
1968 new_path = get_unlocked_mut_path(trans, iter->btree_id, n->c.level, n->key.k.p);
1969 six_lock_increment(&n->c.lock, SIX_LOCK_intent);
1970 mark_btree_node_locked(trans, new_path, n->c.level, SIX_LOCK_intent);
1971 bch2_btree_path_level_init(trans, new_path, n);
1973 trace_and_count(c, btree_node_rewrite, c, b);
1976 bch2_keylist_add(&as->parent_keys, &n->key);
1977 ret = bch2_btree_insert_node(as, trans, iter->path, parent,
1978 &as->parent_keys, flags);
1982 bch2_btree_set_root(as, trans, iter->path, n);
1985 bch2_btree_update_get_open_buckets(as, n);
1986 bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1988 bch2_btree_node_free_inmem(trans, iter->path, b);
1990 bch2_trans_node_add(trans, n);
1991 six_unlock_intent(&n->c.lock);
1993 bch2_btree_update_done(as, trans);
1996 bch2_path_put(trans, new_path, true);
1997 bch2_btree_path_downgrade(trans, iter->path);
2000 bch2_btree_node_free_never_used(as, trans, n);
2001 bch2_btree_update_free(as, trans);
2005 struct async_btree_rewrite {
2007 struct work_struct work;
2008 struct list_head list;
2009 enum btree_id btree_id;
2015 static int async_btree_node_rewrite_trans(struct btree_trans *trans,
2016 struct async_btree_rewrite *a)
2018 struct bch_fs *c = trans->c;
2019 struct btree_iter iter;
2023 bch2_trans_node_iter_init(trans, &iter, a->btree_id, a->pos,
2024 BTREE_MAX_DEPTH, a->level, 0);
2025 b = bch2_btree_iter_peek_node(&iter);
2026 ret = PTR_ERR_OR_ZERO(b);
2030 if (!b || b->data->keys.seq != a->seq) {
2031 struct printbuf buf = PRINTBUF;
2034 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
2036 prt_str(&buf, "(null");
2037 bch_info(c, "%s: node to rewrite not found:, searching for seq %llu, got\n%s",
2038 __func__, a->seq, buf.buf);
2039 printbuf_exit(&buf);
2043 ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
2045 bch2_trans_iter_exit(trans, &iter);
2050 static void async_btree_node_rewrite_work(struct work_struct *work)
2052 struct async_btree_rewrite *a =
2053 container_of(work, struct async_btree_rewrite, work);
2054 struct bch_fs *c = a->c;
2057 ret = bch2_trans_do(c, NULL, NULL, 0,
2058 async_btree_node_rewrite_trans(&trans, a));
2060 bch_err(c, "%s: error %s", __func__, bch2_err_str(ret));
2061 bch2_write_ref_put(c, BCH_WRITE_REF_node_rewrite);
2065 void bch2_btree_node_rewrite_async(struct bch_fs *c, struct btree *b)
2067 struct async_btree_rewrite *a;
2070 a = kmalloc(sizeof(*a), GFP_NOFS);
2072 bch_err(c, "%s: error allocating memory", __func__);
2077 a->btree_id = b->c.btree_id;
2078 a->level = b->c.level;
2079 a->pos = b->key.k.p;
2080 a->seq = b->data->keys.seq;
2081 INIT_WORK(&a->work, async_btree_node_rewrite_work);
2083 if (unlikely(!test_bit(BCH_FS_MAY_GO_RW, &c->flags))) {
2084 mutex_lock(&c->pending_node_rewrites_lock);
2085 list_add(&a->list, &c->pending_node_rewrites);
2086 mutex_unlock(&c->pending_node_rewrites_lock);
2090 if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_node_rewrite)) {
2091 if (test_bit(BCH_FS_STARTED, &c->flags)) {
2092 bch_err(c, "%s: error getting c->writes ref", __func__);
2097 ret = bch2_fs_read_write_early(c);
2099 bch_err(c, "%s: error going read-write: %s",
2100 __func__, bch2_err_str(ret));
2105 bch2_write_ref_get(c, BCH_WRITE_REF_node_rewrite);
2108 queue_work(c->btree_interior_update_worker, &a->work);
2111 void bch2_do_pending_node_rewrites(struct bch_fs *c)
2113 struct async_btree_rewrite *a, *n;
2115 mutex_lock(&c->pending_node_rewrites_lock);
2116 list_for_each_entry_safe(a, n, &c->pending_node_rewrites, list) {
2119 bch2_write_ref_get(c, BCH_WRITE_REF_node_rewrite);
2120 queue_work(c->btree_interior_update_worker, &a->work);
2122 mutex_unlock(&c->pending_node_rewrites_lock);
2125 void bch2_free_pending_node_rewrites(struct bch_fs *c)
2127 struct async_btree_rewrite *a, *n;
2129 mutex_lock(&c->pending_node_rewrites_lock);
2130 list_for_each_entry_safe(a, n, &c->pending_node_rewrites, list) {
2135 mutex_unlock(&c->pending_node_rewrites_lock);
2138 static int __bch2_btree_node_update_key(struct btree_trans *trans,
2139 struct btree_iter *iter,
2140 struct btree *b, struct btree *new_hash,
2141 struct bkey_i *new_key,
2142 unsigned commit_flags,
2145 struct bch_fs *c = trans->c;
2146 struct btree_iter iter2 = { NULL };
2147 struct btree *parent;
2150 if (!skip_triggers) {
2151 ret = bch2_trans_mark_old(trans, b->c.btree_id, b->c.level + 1,
2152 bkey_i_to_s_c(&b->key), 0);
2156 ret = bch2_trans_mark_new(trans, b->c.btree_id, b->c.level + 1,
2163 bkey_copy(&new_hash->key, new_key);
2164 ret = bch2_btree_node_hash_insert(&c->btree_cache,
2165 new_hash, b->c.level, b->c.btree_id);
2169 parent = btree_node_parent(iter->path, b);
2171 bch2_trans_copy_iter(&iter2, iter);
2173 iter2.path = bch2_btree_path_make_mut(trans, iter2.path,
2174 iter2.flags & BTREE_ITER_INTENT,
2177 BUG_ON(iter2.path->level != b->c.level);
2178 BUG_ON(!bpos_eq(iter2.path->pos, new_key->k.p));
2180 btree_path_set_level_up(trans, iter2.path);
2182 trans->paths_sorted = false;
2184 ret = bch2_btree_iter_traverse(&iter2) ?:
2185 bch2_trans_update(trans, &iter2, new_key, BTREE_TRIGGER_NORUN);
2189 BUG_ON(btree_node_root(c, b) != b);
2191 ret = darray_make_room(&trans->extra_journal_entries,
2192 jset_u64s(new_key->k.u64s));
2196 journal_entry_set((void *) &darray_top(trans->extra_journal_entries),
2197 BCH_JSET_ENTRY_btree_root,
2198 b->c.btree_id, b->c.level,
2199 new_key, new_key->k.u64s);
2200 trans->extra_journal_entries.nr += jset_u64s(new_key->k.u64s);
2203 ret = bch2_trans_commit(trans, NULL, NULL, commit_flags);
2207 bch2_btree_node_lock_write_nofail(trans, iter->path, &b->c);
2210 mutex_lock(&c->btree_cache.lock);
2211 bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
2212 bch2_btree_node_hash_remove(&c->btree_cache, b);
2214 bkey_copy(&b->key, new_key);
2215 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
2217 mutex_unlock(&c->btree_cache.lock);
2219 bkey_copy(&b->key, new_key);
2222 bch2_btree_node_unlock_write(trans, iter->path, b);
2224 bch2_trans_iter_exit(trans, &iter2);
2228 mutex_lock(&c->btree_cache.lock);
2229 bch2_btree_node_hash_remove(&c->btree_cache, b);
2230 mutex_unlock(&c->btree_cache.lock);
2235 int bch2_btree_node_update_key(struct btree_trans *trans, struct btree_iter *iter,
2236 struct btree *b, struct bkey_i *new_key,
2237 unsigned commit_flags, bool skip_triggers)
2239 struct bch_fs *c = trans->c;
2240 struct btree *new_hash = NULL;
2241 struct btree_path *path = iter->path;
2245 ret = bch2_btree_path_upgrade(trans, path, b->c.level + 1);
2249 closure_init_stack(&cl);
2252 * check btree_ptr_hash_val() after @b is locked by
2253 * btree_iter_traverse():
2255 if (btree_ptr_hash_val(new_key) != b->hash_val) {
2256 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2258 ret = drop_locks_do(trans, (closure_sync(&cl), 0));
2263 new_hash = bch2_btree_node_mem_alloc(trans, false);
2267 ret = __bch2_btree_node_update_key(trans, iter, b, new_hash, new_key,
2268 commit_flags, skip_triggers);
2272 mutex_lock(&c->btree_cache.lock);
2273 list_move(&new_hash->list, &c->btree_cache.freeable);
2274 mutex_unlock(&c->btree_cache.lock);
2276 six_unlock_write(&new_hash->c.lock);
2277 six_unlock_intent(&new_hash->c.lock);
2280 bch2_btree_cache_cannibalize_unlock(c);
2284 int bch2_btree_node_update_key_get_iter(struct btree_trans *trans,
2285 struct btree *b, struct bkey_i *new_key,
2286 unsigned commit_flags, bool skip_triggers)
2288 struct btree_iter iter;
2291 bch2_trans_node_iter_init(trans, &iter, b->c.btree_id, b->key.k.p,
2292 BTREE_MAX_DEPTH, b->c.level,
2294 ret = bch2_btree_iter_traverse(&iter);
2298 /* has node been freed? */
2299 if (iter.path->l[b->c.level].b != b) {
2300 /* node has been freed: */
2301 BUG_ON(!btree_node_dying(b));
2305 BUG_ON(!btree_node_hashed(b));
2307 ret = bch2_btree_node_update_key(trans, &iter, b, new_key,
2308 commit_flags, skip_triggers);
2310 bch2_trans_iter_exit(trans, &iter);
2317 * Only for filesystem bringup, when first reading the btree roots or allocating
2318 * btree roots when initializing a new filesystem:
2320 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
2322 BUG_ON(btree_node_root(c, b));
2324 bch2_btree_set_root_inmem(c, b);
2327 static int __bch2_btree_root_alloc(struct btree_trans *trans, enum btree_id id)
2329 struct bch_fs *c = trans->c;
2334 closure_init_stack(&cl);
2337 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2341 b = bch2_btree_node_mem_alloc(trans, false);
2342 bch2_btree_cache_cannibalize_unlock(c);
2344 set_btree_node_fake(b);
2345 set_btree_node_need_rewrite(b);
2349 bkey_btree_ptr_init(&b->key);
2350 b->key.k.p = SPOS_MAX;
2351 *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
2353 bch2_bset_init_first(b, &b->data->keys);
2354 bch2_btree_build_aux_trees(b);
2357 btree_set_min(b, POS_MIN);
2358 btree_set_max(b, SPOS_MAX);
2359 b->data->format = bch2_btree_calc_format(b);
2360 btree_node_set_format(b, b->data->format);
2362 ret = bch2_btree_node_hash_insert(&c->btree_cache, b,
2363 b->c.level, b->c.btree_id);
2366 bch2_btree_set_root_inmem(c, b);
2368 six_unlock_write(&b->c.lock);
2369 six_unlock_intent(&b->c.lock);
2373 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
2375 bch2_trans_run(c, __bch2_btree_root_alloc(&trans, id));
2378 void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c)
2380 struct btree_update *as;
2382 mutex_lock(&c->btree_interior_update_lock);
2383 list_for_each_entry(as, &c->btree_interior_update_list, list)
2384 prt_printf(out, "%p m %u w %u r %u j %llu\n",
2388 atomic_read(&as->cl.remaining) & CLOSURE_REMAINING_MASK,
2390 mutex_unlock(&c->btree_interior_update_lock);
2393 static bool bch2_btree_interior_updates_pending(struct bch_fs *c)
2397 mutex_lock(&c->btree_interior_update_lock);
2398 ret = !list_empty(&c->btree_interior_update_list);
2399 mutex_unlock(&c->btree_interior_update_lock);
2404 bool bch2_btree_interior_updates_flush(struct bch_fs *c)
2406 bool ret = bch2_btree_interior_updates_pending(c);
2409 closure_wait_event(&c->btree_interior_update_wait,
2410 !bch2_btree_interior_updates_pending(c));
2414 void bch2_journal_entry_to_btree_root(struct bch_fs *c, struct jset_entry *entry)
2416 struct btree_root *r = bch2_btree_id_root(c, entry->btree_id);
2418 mutex_lock(&c->btree_root_lock);
2420 r->level = entry->level;
2422 bkey_copy(&r->key, &entry->start[0]);
2424 mutex_unlock(&c->btree_root_lock);
2428 bch2_btree_roots_to_journal_entries(struct bch_fs *c,
2429 struct jset_entry *start,
2430 struct jset_entry *end)
2432 struct jset_entry *entry;
2433 unsigned long have = 0;
2436 for (entry = start; entry < end; entry = vstruct_next(entry))
2437 if (entry->type == BCH_JSET_ENTRY_btree_root)
2438 __set_bit(entry->btree_id, &have);
2440 mutex_lock(&c->btree_root_lock);
2442 for (i = 0; i < btree_id_nr_alive(c); i++) {
2443 struct btree_root *r = bch2_btree_id_root(c, i);
2445 if (r->alive && !test_bit(i, &have)) {
2446 journal_entry_set(end, BCH_JSET_ENTRY_btree_root,
2447 i, r->level, &r->key, r->key.k.u64s);
2448 end = vstruct_next(end);
2452 mutex_unlock(&c->btree_root_lock);
2457 void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
2459 if (c->btree_interior_update_worker)
2460 destroy_workqueue(c->btree_interior_update_worker);
2461 mempool_exit(&c->btree_interior_update_pool);
2464 void bch2_fs_btree_interior_update_init_early(struct bch_fs *c)
2466 mutex_init(&c->btree_reserve_cache_lock);
2467 INIT_LIST_HEAD(&c->btree_interior_update_list);
2468 INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
2469 mutex_init(&c->btree_interior_update_lock);
2470 INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
2472 INIT_LIST_HEAD(&c->pending_node_rewrites);
2473 mutex_init(&c->pending_node_rewrites_lock);
2476 int bch2_fs_btree_interior_update_init(struct bch_fs *c)
2478 c->btree_interior_update_worker =
2479 alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
2480 if (!c->btree_interior_update_worker)
2481 return -BCH_ERR_ENOMEM_btree_interior_update_worker_init;
2483 if (mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
2484 sizeof(struct btree_update)))
2485 return -BCH_ERR_ENOMEM_btree_interior_update_pool_init;