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"
17 #include "journal_reclaim.h"
23 #include <linux/random.h>
24 #include <trace/events/bcachefs.h>
26 static int bch2_btree_insert_node(struct btree_update *, struct btree_trans *,
27 struct btree_path *, struct btree *,
28 struct keylist *, unsigned);
29 static void bch2_btree_update_add_new_node(struct btree_update *, struct btree *);
31 static struct btree_path *get_unlocked_mut_path(struct btree_trans *trans,
32 enum btree_id btree_id,
36 struct btree_path *path;
38 path = bch2_path_get(trans, btree_id, pos, level + 1, level,
39 BTREE_ITER_NOPRESERVE|
40 BTREE_ITER_INTENT, _RET_IP_);
41 path = bch2_btree_path_make_mut(trans, path, true, _RET_IP_);
42 bch2_btree_path_downgrade(trans, path);
43 __bch2_btree_path_unlock(trans, path);
50 * Verify that child nodes correctly span parent node's range:
52 static void btree_node_interior_verify(struct bch_fs *c, struct btree *b)
54 #ifdef CONFIG_BCACHEFS_DEBUG
55 struct bpos next_node = b->data->min_key;
56 struct btree_node_iter iter;
58 struct bkey_s_c_btree_ptr_v2 bp;
60 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
64 if (!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))
67 bch2_btree_node_iter_init_from_start(&iter, b);
70 k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked);
71 if (k.k->type != KEY_TYPE_btree_ptr_v2)
73 bp = bkey_s_c_to_btree_ptr_v2(k);
75 if (bpos_cmp(next_node, bp.v->min_key)) {
76 bch2_dump_btree_node(c, b);
77 bch2_bpos_to_text(&buf1, next_node);
78 bch2_bpos_to_text(&buf2, bp.v->min_key);
79 panic("expected next min_key %s got %s\n", buf1.buf, buf2.buf);
82 bch2_btree_node_iter_advance(&iter, b);
84 if (bch2_btree_node_iter_end(&iter)) {
85 if (bpos_cmp(k.k->p, b->key.k.p)) {
86 bch2_dump_btree_node(c, b);
87 bch2_bpos_to_text(&buf1, b->key.k.p);
88 bch2_bpos_to_text(&buf2, k.k->p);
89 panic("expected end %s got %s\n", buf1.buf, buf2.buf);
94 next_node = bpos_successor(k.k->p);
99 /* Calculate ideal packed bkey format for new btree nodes: */
101 void __bch2_btree_calc_format(struct bkey_format_state *s, struct btree *b)
103 struct bkey_packed *k;
108 bset_tree_for_each_key(b, t, k)
109 if (!bkey_deleted(k)) {
110 uk = bkey_unpack_key(b, k);
111 bch2_bkey_format_add_key(s, &uk);
115 static struct bkey_format bch2_btree_calc_format(struct btree *b)
117 struct bkey_format_state s;
119 bch2_bkey_format_init(&s);
120 bch2_bkey_format_add_pos(&s, b->data->min_key);
121 bch2_bkey_format_add_pos(&s, b->data->max_key);
122 __bch2_btree_calc_format(&s, b);
124 return bch2_bkey_format_done(&s);
127 static size_t btree_node_u64s_with_format(struct btree *b,
128 struct bkey_format *new_f)
130 struct bkey_format *old_f = &b->format;
132 /* stupid integer promotion rules */
134 (((int) new_f->key_u64s - old_f->key_u64s) *
135 (int) b->nr.packed_keys) +
136 (((int) new_f->key_u64s - BKEY_U64s) *
137 (int) b->nr.unpacked_keys);
139 BUG_ON(delta + b->nr.live_u64s < 0);
141 return b->nr.live_u64s + delta;
145 * btree_node_format_fits - check if we could rewrite node with a new format
147 * This assumes all keys can pack with the new format -- it just checks if
148 * the re-packed keys would fit inside the node itself.
150 bool bch2_btree_node_format_fits(struct bch_fs *c, struct btree *b,
151 struct bkey_format *new_f)
153 size_t u64s = btree_node_u64s_with_format(b, new_f);
155 return __vstruct_bytes(struct btree_node, u64s) < btree_bytes(c);
158 /* Btree node freeing/allocation: */
160 static void __btree_node_free(struct bch_fs *c, struct btree *b)
162 trace_and_count(c, btree_node_free, c, b);
164 BUG_ON(btree_node_dirty(b));
165 BUG_ON(btree_node_need_write(b));
166 BUG_ON(b == btree_node_root(c, b));
168 BUG_ON(!list_empty(&b->write_blocked));
169 BUG_ON(b->will_make_reachable);
171 clear_btree_node_noevict(b);
173 mutex_lock(&c->btree_cache.lock);
174 list_move(&b->list, &c->btree_cache.freeable);
175 mutex_unlock(&c->btree_cache.lock);
178 static void bch2_btree_node_free_inmem(struct btree_trans *trans,
179 struct btree_path *path,
182 struct bch_fs *c = trans->c;
183 unsigned level = b->c.level;
185 bch2_btree_node_lock_write_nofail(trans, path, &b->c);
186 bch2_btree_node_hash_remove(&c->btree_cache, b);
187 __btree_node_free(c, b);
188 six_unlock_write(&b->c.lock);
189 mark_btree_node_locked_noreset(path, level, SIX_LOCK_intent);
191 trans_for_each_path(trans, path)
192 if (path->l[level].b == b) {
193 btree_node_unlock(trans, path, level);
194 path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
198 static void bch2_btree_node_free_never_used(struct btree_update *as,
199 struct btree_trans *trans,
202 struct bch_fs *c = as->c;
203 struct prealloc_nodes *p = &as->prealloc_nodes[b->c.lock.readers != NULL];
204 struct btree_path *path;
205 unsigned level = b->c.level;
207 BUG_ON(!list_empty(&b->write_blocked));
208 BUG_ON(b->will_make_reachable != (1UL|(unsigned long) as));
210 b->will_make_reachable = 0;
211 closure_put(&as->cl);
213 clear_btree_node_will_make_reachable(b);
214 clear_btree_node_accessed(b);
215 clear_btree_node_dirty_acct(c, b);
216 clear_btree_node_need_write(b);
218 mutex_lock(&c->btree_cache.lock);
219 list_del_init(&b->list);
220 bch2_btree_node_hash_remove(&c->btree_cache, b);
221 mutex_unlock(&c->btree_cache.lock);
223 BUG_ON(p->nr >= ARRAY_SIZE(p->b));
226 six_unlock_intent(&b->c.lock);
228 trans_for_each_path(trans, path)
229 if (path->l[level].b == b) {
230 btree_node_unlock(trans, path, level);
231 path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
235 static struct btree *__bch2_btree_node_alloc(struct btree_trans *trans,
236 struct disk_reservation *res,
241 struct bch_fs *c = trans->c;
242 struct write_point *wp;
244 __BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
245 struct open_buckets ob = { .nr = 0 };
246 struct bch_devs_list devs_have = (struct bch_devs_list) { 0 };
248 enum alloc_reserve alloc_reserve;
250 if (flags & BTREE_INSERT_USE_RESERVE) {
252 alloc_reserve = RESERVE_btree_movinggc;
254 nr_reserve = BTREE_NODE_RESERVE;
255 alloc_reserve = RESERVE_btree;
258 mutex_lock(&c->btree_reserve_cache_lock);
259 if (c->btree_reserve_cache_nr > nr_reserve) {
260 struct btree_alloc *a =
261 &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
264 bkey_copy(&tmp.k, &a->k);
265 mutex_unlock(&c->btree_reserve_cache_lock);
268 mutex_unlock(&c->btree_reserve_cache_lock);
271 wp = bch2_alloc_sectors_start_trans(trans,
272 c->opts.metadata_target ?:
273 c->opts.foreground_target,
275 writepoint_ptr(&c->btree_write_point),
278 c->opts.metadata_replicas_required,
279 alloc_reserve, 0, cl);
283 if (wp->sectors_free < btree_sectors(c)) {
284 struct open_bucket *ob;
287 open_bucket_for_each(c, &wp->ptrs, ob, i)
288 if (ob->sectors_free < btree_sectors(c))
289 ob->sectors_free = 0;
291 bch2_alloc_sectors_done(c, wp);
295 bkey_btree_ptr_v2_init(&tmp.k);
296 bch2_alloc_sectors_append_ptrs(c, wp, &tmp.k, btree_sectors(c), false);
298 bch2_open_bucket_get(c, wp, &ob);
299 bch2_alloc_sectors_done(c, wp);
301 b = bch2_btree_node_mem_alloc(c, interior_node);
302 six_unlock_write(&b->c.lock);
303 six_unlock_intent(&b->c.lock);
305 /* we hold cannibalize_lock: */
309 bkey_copy(&b->key, &tmp.k);
315 static struct btree *bch2_btree_node_alloc(struct btree_update *as,
316 struct btree_trans *trans,
319 struct bch_fs *c = as->c;
321 struct prealloc_nodes *p = &as->prealloc_nodes[!!level];
324 BUG_ON(level >= BTREE_MAX_DEPTH);
329 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
330 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
332 set_btree_node_accessed(b);
333 set_btree_node_dirty_acct(c, b);
334 set_btree_node_need_write(b);
336 bch2_bset_init_first(b, &b->data->keys);
338 b->c.btree_id = as->btree_id;
339 b->version_ondisk = c->sb.version;
341 memset(&b->nr, 0, sizeof(b->nr));
342 b->data->magic = cpu_to_le64(bset_magic(c));
343 memset(&b->data->_ptr, 0, sizeof(b->data->_ptr));
345 SET_BTREE_NODE_ID(b->data, as->btree_id);
346 SET_BTREE_NODE_LEVEL(b->data, level);
348 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
349 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(&b->key);
352 bp->v.seq = b->data->keys.seq;
353 bp->v.sectors_written = 0;
356 SET_BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data, true);
358 bch2_btree_build_aux_trees(b);
360 ret = bch2_btree_node_hash_insert(&c->btree_cache, b, level, as->btree_id);
363 trace_and_count(c, btree_node_alloc, c, b);
367 static void btree_set_min(struct btree *b, struct bpos pos)
369 if (b->key.k.type == KEY_TYPE_btree_ptr_v2)
370 bkey_i_to_btree_ptr_v2(&b->key)->v.min_key = pos;
371 b->data->min_key = pos;
374 static void btree_set_max(struct btree *b, struct bpos pos)
377 b->data->max_key = pos;
380 struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *as,
381 struct btree_trans *trans,
383 struct bkey_format format)
387 n = bch2_btree_node_alloc(as, trans, b->c.level);
389 SET_BTREE_NODE_SEQ(n->data, BTREE_NODE_SEQ(b->data) + 1);
391 btree_set_min(n, b->data->min_key);
392 btree_set_max(n, b->data->max_key);
394 n->data->format = format;
395 btree_node_set_format(n, format);
397 bch2_btree_sort_into(as->c, n, b);
399 btree_node_reset_sib_u64s(n);
401 n->key.k.p = b->key.k.p;
405 static struct btree *bch2_btree_node_alloc_replacement(struct btree_update *as,
406 struct btree_trans *trans,
409 struct bkey_format new_f = bch2_btree_calc_format(b);
412 * The keys might expand with the new format - if they wouldn't fit in
413 * the btree node anymore, use the old format for now:
415 if (!bch2_btree_node_format_fits(as->c, b, &new_f))
418 return __bch2_btree_node_alloc_replacement(as, trans, b, new_f);
421 static struct btree *__btree_root_alloc(struct btree_update *as,
422 struct btree_trans *trans, unsigned level)
424 struct btree *b = bch2_btree_node_alloc(as, trans, level);
426 btree_set_min(b, POS_MIN);
427 btree_set_max(b, SPOS_MAX);
428 b->data->format = bch2_btree_calc_format(b);
430 btree_node_set_format(b, b->data->format);
431 bch2_btree_build_aux_trees(b);
436 static void bch2_btree_reserve_put(struct btree_update *as, struct btree_trans *trans)
438 struct bch_fs *c = as->c;
439 struct prealloc_nodes *p;
441 for (p = as->prealloc_nodes;
442 p < as->prealloc_nodes + ARRAY_SIZE(as->prealloc_nodes);
445 struct btree *b = p->b[--p->nr];
447 mutex_lock(&c->btree_reserve_cache_lock);
449 if (c->btree_reserve_cache_nr <
450 ARRAY_SIZE(c->btree_reserve_cache)) {
451 struct btree_alloc *a =
452 &c->btree_reserve_cache[c->btree_reserve_cache_nr++];
456 bkey_copy(&a->k, &b->key);
458 bch2_open_buckets_put(c, &b->ob);
461 mutex_unlock(&c->btree_reserve_cache_lock);
463 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
464 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
465 __btree_node_free(c, b);
466 six_unlock_write(&b->c.lock);
467 six_unlock_intent(&b->c.lock);
472 static int bch2_btree_reserve_get(struct btree_trans *trans,
473 struct btree_update *as,
474 unsigned nr_nodes[2],
478 struct bch_fs *c = as->c;
483 BUG_ON(nr_nodes[0] + nr_nodes[1] > BTREE_RESERVE_MAX);
486 * Protects reaping from the btree node cache and using the btree node
487 * open bucket reserve:
489 * BTREE_INSERT_NOWAIT only applies to btree node allocation, not
490 * blocking on this lock:
492 ret = bch2_btree_cache_cannibalize_lock(c, cl);
496 for (interior = 0; interior < 2; interior++) {
497 struct prealloc_nodes *p = as->prealloc_nodes + interior;
499 while (p->nr < nr_nodes[interior]) {
500 b = __bch2_btree_node_alloc(trans, &as->disk_res,
501 flags & BTREE_INSERT_NOWAIT ? NULL : cl,
512 bch2_btree_cache_cannibalize_unlock(c);
516 /* Asynchronous interior node update machinery */
518 static void bch2_btree_update_free(struct btree_update *as, struct btree_trans *trans)
520 struct bch_fs *c = as->c;
522 if (as->took_gc_lock)
523 up_read(&c->gc_lock);
524 as->took_gc_lock = false;
526 bch2_journal_preres_put(&c->journal, &as->journal_preres);
528 bch2_journal_pin_drop(&c->journal, &as->journal);
529 bch2_journal_pin_flush(&c->journal, &as->journal);
530 bch2_disk_reservation_put(c, &as->disk_res);
531 bch2_btree_reserve_put(as, trans);
533 bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_total],
536 mutex_lock(&c->btree_interior_update_lock);
537 list_del(&as->unwritten_list);
540 closure_debug_destroy(&as->cl);
541 mempool_free(as, &c->btree_interior_update_pool);
544 * Have to do the wakeup with btree_interior_update_lock still held,
545 * since being on btree_interior_update_list is our ref on @c:
547 closure_wake_up(&c->btree_interior_update_wait);
549 mutex_unlock(&c->btree_interior_update_lock);
552 static void btree_update_add_key(struct btree_update *as,
553 struct keylist *keys, struct btree *b)
555 struct bkey_i *k = &b->key;
557 BUG_ON(bch2_keylist_u64s(keys) + k->k.u64s >
558 ARRAY_SIZE(as->_old_keys));
560 bkey_copy(keys->top, k);
561 bkey_i_to_btree_ptr_v2(keys->top)->v.mem_ptr = b->c.level + 1;
563 bch2_keylist_push(keys);
567 * The transactional part of an interior btree node update, where we journal the
568 * update we did to the interior node and update alloc info:
570 static int btree_update_nodes_written_trans(struct btree_trans *trans,
571 struct btree_update *as)
576 ret = darray_make_room(&trans->extra_journal_entries, as->journal_u64s);
580 memcpy(&darray_top(trans->extra_journal_entries),
582 as->journal_u64s * sizeof(u64));
583 trans->extra_journal_entries.nr += as->journal_u64s;
585 trans->journal_pin = &as->journal;
587 for_each_keylist_key(&as->old_keys, k) {
588 unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr;
590 ret = bch2_trans_mark_old(trans, as->btree_id, level, bkey_i_to_s_c(k), 0);
595 for_each_keylist_key(&as->new_keys, k) {
596 unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr;
598 ret = bch2_trans_mark_new(trans, as->btree_id, level, k, 0);
606 static void btree_update_nodes_written(struct btree_update *as)
608 struct bch_fs *c = as->c;
610 struct btree_trans trans;
615 bch2_trans_init(&trans, c, 0, 512);
617 * If we're already in an error state, it might be because a btree node
618 * was never written, and we might be trying to free that same btree
619 * node here, but it won't have been marked as allocated and we'll see
620 * spurious disk usage inconsistencies in the transactional part below
621 * if we don't skip it:
623 ret = bch2_journal_error(&c->journal);
628 * Wait for any in flight writes to finish before we free the old nodes
631 for (i = 0; i < as->nr_old_nodes; i++) {
634 b = as->old_nodes[i];
636 btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read);
637 seq = b->data ? b->data->keys.seq : 0;
638 six_unlock_read(&b->c.lock);
640 if (seq == as->old_nodes_seq[i])
641 wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight_inner,
642 TASK_UNINTERRUPTIBLE);
646 * We did an update to a parent node where the pointers we added pointed
647 * to child nodes that weren't written yet: now, the child nodes have
648 * been written so we can write out the update to the interior node.
652 * We can't call into journal reclaim here: we'd block on the journal
653 * reclaim lock, but we may need to release the open buckets we have
654 * pinned in order for other btree updates to make forward progress, and
655 * journal reclaim does btree updates when flushing bkey_cached entries,
656 * which may require allocations as well.
658 ret = commit_do(&trans, &as->disk_res, &journal_seq,
660 BTREE_INSERT_NOCHECK_RW|
661 BTREE_INSERT_USE_RESERVE|
662 BTREE_INSERT_JOURNAL_RECLAIM|
663 JOURNAL_WATERMARK_reserved,
664 btree_update_nodes_written_trans(&trans, as));
665 bch2_trans_unlock(&trans);
667 bch2_fs_fatal_err_on(ret && !bch2_journal_error(&c->journal), c,
668 "%s(): error %s", __func__, bch2_err_str(ret));
671 struct btree_path *path;
674 path = get_unlocked_mut_path(&trans, as->btree_id, b->c.level, b->key.k.p);
676 * @b is the node we did the final insert into:
678 * On failure to get a journal reservation, we still have to
679 * unblock the write and allow most of the write path to happen
680 * so that shutdown works, but the i->journal_seq mechanism
681 * won't work to prevent the btree write from being visible (we
682 * didn't get a journal sequence number) - instead
683 * __bch2_btree_node_write() doesn't do the actual write if
684 * we're in journal error state:
688 * Ensure transaction is unlocked before using
689 * btree_node_lock_nopath() (the use of which is always suspect,
690 * we need to work on removing this in the future)
692 * It should be, but get_unlocked_mut_path() -> bch2_path_get()
693 * calls bch2_path_upgrade(), before we call path_make_mut(), so
694 * we may rarely end up with a locked path besides the one we
697 bch2_trans_unlock(&trans);
698 btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_intent);
699 mark_btree_node_locked(&trans, path, b->c.level, SIX_LOCK_intent);
700 bch2_btree_path_level_init(&trans, path, b);
702 bch2_btree_node_lock_write_nofail(&trans, path, &b->c);
704 mutex_lock(&c->btree_interior_update_lock);
706 list_del(&as->write_blocked_list);
707 if (list_empty(&b->write_blocked))
708 clear_btree_node_write_blocked(b);
711 * Node might have been freed, recheck under
712 * btree_interior_update_lock:
715 struct bset *i = btree_bset_last(b);
718 BUG_ON(!btree_node_dirty(b));
721 i->journal_seq = cpu_to_le64(
723 le64_to_cpu(i->journal_seq)));
725 bch2_btree_add_journal_pin(c, b, journal_seq);
728 * If we didn't get a journal sequence number we
729 * can't write this btree node, because recovery
730 * won't know to ignore this write:
732 set_btree_node_never_write(b);
736 mutex_unlock(&c->btree_interior_update_lock);
738 mark_btree_node_locked_noreset(path, b->c.level, SIX_LOCK_intent);
739 six_unlock_write(&b->c.lock);
741 btree_node_write_if_need(c, b, SIX_LOCK_intent);
742 btree_node_unlock(&trans, path, b->c.level);
743 bch2_path_put(&trans, path, true);
746 bch2_journal_pin_drop(&c->journal, &as->journal);
748 bch2_journal_preres_put(&c->journal, &as->journal_preres);
750 mutex_lock(&c->btree_interior_update_lock);
751 for (i = 0; i < as->nr_new_nodes; i++) {
752 b = as->new_nodes[i];
754 BUG_ON(b->will_make_reachable != (unsigned long) as);
755 b->will_make_reachable = 0;
756 clear_btree_node_will_make_reachable(b);
758 mutex_unlock(&c->btree_interior_update_lock);
760 for (i = 0; i < as->nr_new_nodes; i++) {
761 b = as->new_nodes[i];
763 btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read);
764 btree_node_write_if_need(c, b, SIX_LOCK_read);
765 six_unlock_read(&b->c.lock);
768 for (i = 0; i < as->nr_open_buckets; i++)
769 bch2_open_bucket_put(c, c->open_buckets + as->open_buckets[i]);
771 bch2_btree_update_free(as, &trans);
772 bch2_trans_exit(&trans);
775 static void btree_interior_update_work(struct work_struct *work)
778 container_of(work, struct bch_fs, btree_interior_update_work);
779 struct btree_update *as;
782 mutex_lock(&c->btree_interior_update_lock);
783 as = list_first_entry_or_null(&c->btree_interior_updates_unwritten,
784 struct btree_update, unwritten_list);
785 if (as && !as->nodes_written)
787 mutex_unlock(&c->btree_interior_update_lock);
792 btree_update_nodes_written(as);
796 static void btree_update_set_nodes_written(struct closure *cl)
798 struct btree_update *as = container_of(cl, struct btree_update, cl);
799 struct bch_fs *c = as->c;
801 mutex_lock(&c->btree_interior_update_lock);
802 as->nodes_written = true;
803 mutex_unlock(&c->btree_interior_update_lock);
805 queue_work(c->btree_interior_update_worker, &c->btree_interior_update_work);
809 * We're updating @b with pointers to nodes that haven't finished writing yet:
810 * block @b from being written until @as completes
812 static void btree_update_updated_node(struct btree_update *as, struct btree *b)
814 struct bch_fs *c = as->c;
816 mutex_lock(&c->btree_interior_update_lock);
817 list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
819 BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
820 BUG_ON(!btree_node_dirty(b));
822 as->mode = BTREE_INTERIOR_UPDATING_NODE;
825 set_btree_node_write_blocked(b);
826 list_add(&as->write_blocked_list, &b->write_blocked);
828 mutex_unlock(&c->btree_interior_update_lock);
831 static void btree_update_reparent(struct btree_update *as,
832 struct btree_update *child)
834 struct bch_fs *c = as->c;
836 lockdep_assert_held(&c->btree_interior_update_lock);
839 child->mode = BTREE_INTERIOR_UPDATING_AS;
841 bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal, NULL);
844 static void btree_update_updated_root(struct btree_update *as, struct btree *b)
846 struct bkey_i *insert = &b->key;
847 struct bch_fs *c = as->c;
849 BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
851 BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
852 ARRAY_SIZE(as->journal_entries));
855 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
856 BCH_JSET_ENTRY_btree_root,
857 b->c.btree_id, b->c.level,
858 insert, insert->k.u64s);
860 mutex_lock(&c->btree_interior_update_lock);
861 list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
863 as->mode = BTREE_INTERIOR_UPDATING_ROOT;
864 mutex_unlock(&c->btree_interior_update_lock);
868 * bch2_btree_update_add_new_node:
870 * This causes @as to wait on @b to be written, before it gets to
871 * bch2_btree_update_nodes_written
873 * Additionally, it sets b->will_make_reachable to prevent any additional writes
874 * to @b from happening besides the first until @b is reachable on disk
876 * And it adds @b to the list of @as's new nodes, so that we can update sector
877 * counts in bch2_btree_update_nodes_written:
879 static void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b)
881 struct bch_fs *c = as->c;
883 closure_get(&as->cl);
885 mutex_lock(&c->btree_interior_update_lock);
886 BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes));
887 BUG_ON(b->will_make_reachable);
889 as->new_nodes[as->nr_new_nodes++] = b;
890 b->will_make_reachable = 1UL|(unsigned long) as;
891 set_btree_node_will_make_reachable(b);
893 mutex_unlock(&c->btree_interior_update_lock);
895 btree_update_add_key(as, &as->new_keys, b);
897 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
898 unsigned bytes = vstruct_end(&b->data->keys) - (void *) b->data;
899 unsigned sectors = round_up(bytes, block_bytes(c)) >> 9;
901 bkey_i_to_btree_ptr_v2(&b->key)->v.sectors_written =
902 cpu_to_le16(sectors);
907 * returns true if @b was a new node
909 static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b)
911 struct btree_update *as;
915 mutex_lock(&c->btree_interior_update_lock);
917 * When b->will_make_reachable != 0, it owns a ref on as->cl that's
918 * dropped when it gets written by bch2_btree_complete_write - the
919 * xchg() is for synchronization with bch2_btree_complete_write:
921 v = xchg(&b->will_make_reachable, 0);
922 clear_btree_node_will_make_reachable(b);
923 as = (struct btree_update *) (v & ~1UL);
926 mutex_unlock(&c->btree_interior_update_lock);
930 for (i = 0; i < as->nr_new_nodes; i++)
931 if (as->new_nodes[i] == b)
936 array_remove_item(as->new_nodes, as->nr_new_nodes, i);
937 mutex_unlock(&c->btree_interior_update_lock);
940 closure_put(&as->cl);
943 static void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b)
946 as->open_buckets[as->nr_open_buckets++] =
951 * @b is being split/rewritten: it may have pointers to not-yet-written btree
952 * nodes and thus outstanding btree_updates - redirect @b's
953 * btree_updates to point to this btree_update:
955 static void bch2_btree_interior_update_will_free_node(struct btree_update *as,
958 struct bch_fs *c = as->c;
959 struct btree_update *p, *n;
960 struct btree_write *w;
962 set_btree_node_dying(b);
964 if (btree_node_fake(b))
967 mutex_lock(&c->btree_interior_update_lock);
970 * Does this node have any btree_update operations preventing
971 * it from being written?
973 * If so, redirect them to point to this btree_update: we can
974 * write out our new nodes, but we won't make them visible until those
975 * operations complete
977 list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) {
978 list_del_init(&p->write_blocked_list);
979 btree_update_reparent(as, p);
982 * for flush_held_btree_writes() waiting on updates to flush or
983 * nodes to be writeable:
985 closure_wake_up(&c->btree_interior_update_wait);
988 clear_btree_node_dirty_acct(c, b);
989 clear_btree_node_need_write(b);
992 * Does this node have unwritten data that has a pin on the journal?
994 * If so, transfer that pin to the btree_update operation -
995 * note that if we're freeing multiple nodes, we only need to keep the
996 * oldest pin of any of the nodes we're freeing. We'll release the pin
997 * when the new nodes are persistent and reachable on disk:
999 w = btree_current_write(b);
1000 bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
1001 bch2_journal_pin_drop(&c->journal, &w->journal);
1003 w = btree_prev_write(b);
1004 bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
1005 bch2_journal_pin_drop(&c->journal, &w->journal);
1007 mutex_unlock(&c->btree_interior_update_lock);
1010 * Is this a node that isn't reachable on disk yet?
1012 * Nodes that aren't reachable yet have writes blocked until they're
1013 * reachable - now that we've cancelled any pending writes and moved
1014 * things waiting on that write to wait on this update, we can drop this
1015 * node from the list of nodes that the other update is making
1016 * reachable, prior to freeing it:
1018 btree_update_drop_new_node(c, b);
1020 btree_update_add_key(as, &as->old_keys, b);
1022 as->old_nodes[as->nr_old_nodes] = b;
1023 as->old_nodes_seq[as->nr_old_nodes] = b->data->keys.seq;
1027 static void bch2_btree_update_done(struct btree_update *as, struct btree_trans *trans)
1029 struct bch_fs *c = as->c;
1030 u64 start_time = as->start_time;
1032 BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE);
1034 if (as->took_gc_lock)
1035 up_read(&as->c->gc_lock);
1036 as->took_gc_lock = false;
1038 bch2_btree_reserve_put(as, trans);
1040 continue_at(&as->cl, btree_update_set_nodes_written,
1041 as->c->btree_interior_update_worker);
1043 bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_foreground],
1047 static struct btree_update *
1048 bch2_btree_update_start(struct btree_trans *trans, struct btree_path *path,
1049 unsigned level, bool split, unsigned flags)
1051 struct bch_fs *c = trans->c;
1052 struct btree_update *as;
1053 u64 start_time = local_clock();
1054 int disk_res_flags = (flags & BTREE_INSERT_NOFAIL)
1055 ? BCH_DISK_RESERVATION_NOFAIL : 0;
1056 unsigned nr_nodes[2] = { 0, 0 };
1057 unsigned update_level = level;
1058 int journal_flags = flags & JOURNAL_WATERMARK_MASK;
1060 u32 restart_count = trans->restart_count;
1062 BUG_ON(!path->should_be_locked);
1064 if (flags & BTREE_INSERT_JOURNAL_RECLAIM)
1065 journal_flags |= JOURNAL_RES_GET_NONBLOCK;
1068 nr_nodes[!!update_level] += 1 + split;
1071 ret = bch2_btree_path_upgrade(trans, path, update_level + 1);
1073 return ERR_PTR(ret);
1075 if (!btree_path_node(path, update_level)) {
1076 /* Allocating new root? */
1077 nr_nodes[1] += split;
1078 update_level = BTREE_MAX_DEPTH;
1082 if (bch2_btree_node_insert_fits(c, path->l[update_level].b,
1083 BKEY_BTREE_PTR_U64s_MAX * (1 + split)))
1089 if (flags & BTREE_INSERT_GC_LOCK_HELD)
1090 lockdep_assert_held(&c->gc_lock);
1091 else if (!down_read_trylock(&c->gc_lock)) {
1092 bch2_trans_unlock(trans);
1093 down_read(&c->gc_lock);
1094 ret = bch2_trans_relock(trans);
1096 up_read(&c->gc_lock);
1097 return ERR_PTR(ret);
1101 as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOIO);
1102 memset(as, 0, sizeof(*as));
1103 closure_init(&as->cl, NULL);
1105 as->start_time = start_time;
1106 as->mode = BTREE_INTERIOR_NO_UPDATE;
1107 as->took_gc_lock = !(flags & BTREE_INSERT_GC_LOCK_HELD);
1108 as->btree_id = path->btree_id;
1109 as->update_level = update_level;
1110 INIT_LIST_HEAD(&as->list);
1111 INIT_LIST_HEAD(&as->unwritten_list);
1112 INIT_LIST_HEAD(&as->write_blocked_list);
1113 bch2_keylist_init(&as->old_keys, as->_old_keys);
1114 bch2_keylist_init(&as->new_keys, as->_new_keys);
1115 bch2_keylist_init(&as->parent_keys, as->inline_keys);
1117 mutex_lock(&c->btree_interior_update_lock);
1118 list_add_tail(&as->list, &c->btree_interior_update_list);
1119 mutex_unlock(&c->btree_interior_update_lock);
1122 * We don't want to allocate if we're in an error state, that can cause
1123 * deadlock on emergency shutdown due to open buckets getting stuck in
1124 * the btree_reserve_cache after allocator shutdown has cleared it out.
1125 * This check needs to come after adding us to the btree_interior_update
1126 * list but before calling bch2_btree_reserve_get, to synchronize with
1127 * __bch2_fs_read_only().
1129 ret = bch2_journal_error(&c->journal);
1133 ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
1134 BTREE_UPDATE_JOURNAL_RES,
1135 journal_flags|JOURNAL_RES_GET_NONBLOCK);
1137 bch2_trans_unlock(trans);
1139 if (flags & BTREE_INSERT_JOURNAL_RECLAIM) {
1140 ret = -BCH_ERR_journal_reclaim_would_deadlock;
1144 ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
1145 BTREE_UPDATE_JOURNAL_RES,
1148 trace_and_count(c, trans_restart_journal_preres_get, trans, _RET_IP_, journal_flags);
1149 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_journal_preres_get);
1153 ret = bch2_trans_relock(trans);
1158 ret = bch2_disk_reservation_get(c, &as->disk_res,
1159 (nr_nodes[0] + nr_nodes[1]) * btree_sectors(c),
1160 c->opts.metadata_replicas,
1165 ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, NULL);
1166 if (bch2_err_matches(ret, ENOSPC) ||
1167 bch2_err_matches(ret, ENOMEM)) {
1170 closure_init_stack(&cl);
1173 ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, &cl);
1175 bch2_trans_unlock(trans);
1177 } while (ret == -EAGAIN);
1181 trace_and_count(c, btree_reserve_get_fail, trans->fn, _RET_IP_, nr_nodes[0] + nr_nodes[1]);
1185 ret = bch2_trans_relock(trans);
1189 bch2_trans_verify_not_restarted(trans, restart_count);
1192 bch2_btree_update_free(as, trans);
1193 return ERR_PTR(ret);
1196 /* Btree root updates: */
1198 static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
1200 /* Root nodes cannot be reaped */
1201 mutex_lock(&c->btree_cache.lock);
1202 list_del_init(&b->list);
1203 mutex_unlock(&c->btree_cache.lock);
1205 mutex_lock(&c->btree_root_lock);
1206 BUG_ON(btree_node_root(c, b) &&
1207 (b->c.level < btree_node_root(c, b)->c.level ||
1208 !btree_node_dying(btree_node_root(c, b))));
1210 btree_node_root(c, b) = b;
1211 mutex_unlock(&c->btree_root_lock);
1213 bch2_recalc_btree_reserve(c);
1217 * bch_btree_set_root - update the root in memory and on disk
1219 * To ensure forward progress, the current task must not be holding any
1220 * btree node write locks. However, you must hold an intent lock on the
1223 * Note: This allocates a journal entry but doesn't add any keys to
1224 * it. All the btree roots are part of every journal write, so there
1225 * is nothing new to be done. This just guarantees that there is a
1228 static void bch2_btree_set_root(struct btree_update *as,
1229 struct btree_trans *trans,
1230 struct btree_path *path,
1233 struct bch_fs *c = as->c;
1236 trace_and_count(c, btree_node_set_root, c, b);
1238 old = btree_node_root(c, b);
1241 * Ensure no one is using the old root while we switch to the
1244 bch2_btree_node_lock_write_nofail(trans, path, &old->c);
1246 bch2_btree_set_root_inmem(c, b);
1248 btree_update_updated_root(as, b);
1251 * Unlock old root after new root is visible:
1253 * The new root isn't persistent, but that's ok: we still have
1254 * an intent lock on the new root, and any updates that would
1255 * depend on the new root would have to update the new root.
1257 bch2_btree_node_unlock_write(trans, path, old);
1260 /* Interior node updates: */
1262 static void bch2_insert_fixup_btree_ptr(struct btree_update *as,
1263 struct btree_trans *trans,
1264 struct btree_path *path,
1266 struct btree_node_iter *node_iter,
1267 struct bkey_i *insert)
1269 struct bch_fs *c = as->c;
1270 struct bkey_packed *k;
1271 struct printbuf buf = PRINTBUF;
1273 BUG_ON(insert->k.type == KEY_TYPE_btree_ptr_v2 &&
1274 !btree_ptr_sectors_written(insert));
1276 if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)))
1277 bch2_journal_key_overwritten(c, b->c.btree_id, b->c.level, insert->k.p);
1279 if (bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1280 btree_node_type(b), WRITE, &buf) ?:
1281 bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert), &buf)) {
1282 printbuf_reset(&buf);
1283 prt_printf(&buf, "inserting invalid bkey\n ");
1284 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
1285 prt_printf(&buf, "\n ");
1286 bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1287 btree_node_type(b), WRITE, &buf);
1288 bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert), &buf);
1290 bch2_fs_inconsistent(c, "%s", buf.buf);
1294 BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1295 ARRAY_SIZE(as->journal_entries));
1298 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
1299 BCH_JSET_ENTRY_btree_keys,
1300 b->c.btree_id, b->c.level,
1301 insert, insert->k.u64s);
1303 while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1304 bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1305 bch2_btree_node_iter_advance(node_iter, b);
1307 bch2_btree_bset_insert_key(trans, path, b, node_iter, insert);
1308 set_btree_node_dirty_acct(c, b);
1309 set_btree_node_need_write(b);
1311 printbuf_exit(&buf);
1315 __bch2_btree_insert_keys_interior(struct btree_update *as,
1316 struct btree_trans *trans,
1317 struct btree_path *path,
1319 struct btree_node_iter node_iter,
1320 struct keylist *keys)
1322 struct bkey_i *insert = bch2_keylist_front(keys);
1323 struct bkey_packed *k;
1325 BUG_ON(btree_node_type(b) != BKEY_TYPE_btree);
1327 while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1328 (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0))
1331 while (!bch2_keylist_empty(keys)) {
1332 bch2_insert_fixup_btree_ptr(as, trans, path, b,
1333 &node_iter, bch2_keylist_front(keys));
1334 bch2_keylist_pop_front(keys);
1339 * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1342 static struct btree *__btree_split_node(struct btree_update *as,
1343 struct btree_trans *trans,
1346 struct bkey_format_state s;
1347 size_t nr_packed = 0, nr_unpacked = 0;
1349 struct bset *set1, *set2;
1350 struct bkey_packed *k, *set2_start, *set2_end, *out, *prev = NULL;
1353 n2 = bch2_btree_node_alloc(as, trans, n1->c.level);
1355 n2->data->max_key = n1->data->max_key;
1356 n2->data->format = n1->format;
1357 SET_BTREE_NODE_SEQ(n2->data, BTREE_NODE_SEQ(n1->data));
1358 n2->key.k.p = n1->key.k.p;
1360 set1 = btree_bset_first(n1);
1361 set2 = btree_bset_first(n2);
1364 * Has to be a linear search because we don't have an auxiliary
1369 struct bkey_packed *n = bkey_next(k);
1371 if (n == vstruct_last(set1))
1373 if (k->_data - set1->_data >= (le16_to_cpu(set1->u64s) * 3) / 5)
1387 set2_end = vstruct_last(set1);
1389 set1->u64s = cpu_to_le16((u64 *) set2_start - set1->_data);
1390 set_btree_bset_end(n1, n1->set);
1392 n1->nr.live_u64s = le16_to_cpu(set1->u64s);
1393 n1->nr.bset_u64s[0] = le16_to_cpu(set1->u64s);
1394 n1->nr.packed_keys = nr_packed;
1395 n1->nr.unpacked_keys = nr_unpacked;
1397 n1_pos = bkey_unpack_pos(n1, prev);
1398 if (as->c->sb.version < bcachefs_metadata_version_snapshot)
1399 n1_pos.snapshot = U32_MAX;
1401 btree_set_max(n1, n1_pos);
1402 btree_set_min(n2, bpos_successor(n1->key.k.p));
1404 bch2_bkey_format_init(&s);
1405 bch2_bkey_format_add_pos(&s, n2->data->min_key);
1406 bch2_bkey_format_add_pos(&s, n2->data->max_key);
1408 for (k = set2_start; k != set2_end; k = bkey_next(k)) {
1409 struct bkey uk = bkey_unpack_key(n1, k);
1410 bch2_bkey_format_add_key(&s, &uk);
1413 n2->data->format = bch2_bkey_format_done(&s);
1414 btree_node_set_format(n2, n2->data->format);
1417 memset(&n2->nr, 0, sizeof(n2->nr));
1419 for (k = set2_start; k != set2_end; k = bkey_next(k)) {
1420 BUG_ON(!bch2_bkey_transform(&n2->format, out, bkey_packed(k)
1421 ? &n1->format : &bch2_bkey_format_current, k));
1422 out->format = KEY_FORMAT_LOCAL_BTREE;
1423 btree_keys_account_key_add(&n2->nr, 0, out);
1424 out = bkey_next(out);
1427 set2->u64s = cpu_to_le16((u64 *) out - set2->_data);
1428 set_btree_bset_end(n2, n2->set);
1430 BUG_ON(!set1->u64s);
1431 BUG_ON(!set2->u64s);
1433 btree_node_reset_sib_u64s(n1);
1434 btree_node_reset_sib_u64s(n2);
1436 bch2_verify_btree_nr_keys(n1);
1437 bch2_verify_btree_nr_keys(n2);
1440 btree_node_interior_verify(as->c, n1);
1441 btree_node_interior_verify(as->c, n2);
1448 * For updates to interior nodes, we've got to do the insert before we split
1449 * because the stuff we're inserting has to be inserted atomically. Post split,
1450 * the keys might have to go in different nodes and the split would no longer be
1453 * Worse, if the insert is from btree node coalescing, if we do the insert after
1454 * we do the split (and pick the pivot) - the pivot we pick might be between
1455 * nodes that were coalesced, and thus in the middle of a child node post
1458 static void btree_split_insert_keys(struct btree_update *as,
1459 struct btree_trans *trans,
1460 struct btree_path *path,
1462 struct keylist *keys)
1464 struct btree_node_iter node_iter;
1465 struct bkey_i *k = bch2_keylist_front(keys);
1466 struct bkey_packed *src, *dst, *n;
1469 bch2_btree_node_iter_init(&node_iter, b, &k->k.p);
1471 __bch2_btree_insert_keys_interior(as, trans, path, b, node_iter, keys);
1474 * We can't tolerate whiteouts here - with whiteouts there can be
1475 * duplicate keys, and it would be rather bad if we picked a duplicate
1478 i = btree_bset_first(b);
1479 src = dst = i->start;
1480 while (src != vstruct_last(i)) {
1482 if (!bkey_deleted(src)) {
1483 memmove_u64s_down(dst, src, src->u64s);
1484 dst = bkey_next(dst);
1489 /* Also clear out the unwritten whiteouts area: */
1490 b->whiteout_u64s = 0;
1492 i->u64s = cpu_to_le16((u64 *) dst - i->_data);
1493 set_btree_bset_end(b, b->set);
1495 BUG_ON(b->nsets != 1 ||
1496 b->nr.live_u64s != le16_to_cpu(btree_bset_first(b)->u64s));
1498 btree_node_interior_verify(as->c, b);
1501 static int btree_split(struct btree_update *as, struct btree_trans *trans,
1502 struct btree_path *path, struct btree *b,
1503 struct keylist *keys, unsigned flags)
1505 struct bch_fs *c = as->c;
1506 struct btree *parent = btree_node_parent(path, b);
1507 struct btree *n1, *n2 = NULL, *n3 = NULL;
1508 struct btree_path *path1 = NULL, *path2 = NULL;
1509 u64 start_time = local_clock();
1512 BUG_ON(!parent && (b != btree_node_root(c, b)));
1513 BUG_ON(parent && !btree_node_intent_locked(path, b->c.level + 1));
1515 bch2_btree_interior_update_will_free_node(as, b);
1517 n1 = bch2_btree_node_alloc_replacement(as, trans, b);
1520 btree_split_insert_keys(as, trans, path, n1, keys);
1522 if (bset_u64s(&n1->set[0]) > BTREE_SPLIT_THRESHOLD(c)) {
1523 trace_and_count(c, btree_node_split, c, b);
1525 n2 = __btree_split_node(as, trans, n1);
1527 bch2_btree_build_aux_trees(n2);
1528 bch2_btree_build_aux_trees(n1);
1530 bch2_btree_update_add_new_node(as, n1);
1531 bch2_btree_update_add_new_node(as, n2);
1532 six_unlock_write(&n2->c.lock);
1533 six_unlock_write(&n1->c.lock);
1535 path1 = get_unlocked_mut_path(trans, path->btree_id, n1->c.level, n1->key.k.p);
1536 six_lock_increment(&n1->c.lock, SIX_LOCK_intent);
1537 mark_btree_node_locked(trans, path1, n1->c.level, SIX_LOCK_intent);
1538 bch2_btree_path_level_init(trans, path1, n1);
1540 path2 = get_unlocked_mut_path(trans, path->btree_id, n2->c.level, n2->key.k.p);
1541 six_lock_increment(&n2->c.lock, SIX_LOCK_intent);
1542 mark_btree_node_locked(trans, path2, n2->c.level, SIX_LOCK_intent);
1543 bch2_btree_path_level_init(trans, path2, n2);
1546 * Note that on recursive parent_keys == keys, so we
1547 * can't start adding new keys to parent_keys before emptying it
1548 * out (which we did with btree_split_insert_keys() above)
1550 bch2_keylist_add(&as->parent_keys, &n1->key);
1551 bch2_keylist_add(&as->parent_keys, &n2->key);
1554 /* Depth increases, make a new root */
1555 n3 = __btree_root_alloc(as, trans, b->c.level + 1);
1557 bch2_btree_update_add_new_node(as, n3);
1558 six_unlock_write(&n3->c.lock);
1560 path2->locks_want++;
1561 BUG_ON(btree_node_locked(path2, n3->c.level));
1562 six_lock_increment(&n3->c.lock, SIX_LOCK_intent);
1563 mark_btree_node_locked(trans, path2, n3->c.level, SIX_LOCK_intent);
1564 bch2_btree_path_level_init(trans, path2, n3);
1566 n3->sib_u64s[0] = U16_MAX;
1567 n3->sib_u64s[1] = U16_MAX;
1569 btree_split_insert_keys(as, trans, path, n3, &as->parent_keys);
1572 trace_and_count(c, btree_node_compact, c, b);
1574 bch2_btree_build_aux_trees(n1);
1575 bch2_btree_update_add_new_node(as, n1);
1576 six_unlock_write(&n1->c.lock);
1578 path1 = get_unlocked_mut_path(trans, path->btree_id, n1->c.level, n1->key.k.p);
1579 six_lock_increment(&n1->c.lock, SIX_LOCK_intent);
1580 mark_btree_node_locked(trans, path1, n1->c.level, SIX_LOCK_intent);
1581 bch2_btree_path_level_init(trans, path1, n1);
1584 bch2_keylist_add(&as->parent_keys, &n1->key);
1587 /* New nodes all written, now make them visible: */
1590 /* Split a non root node */
1591 ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1595 bch2_btree_set_root(as, trans, path, n3);
1597 /* Root filled up but didn't need to be split */
1598 bch2_btree_set_root(as, trans, path, n1);
1602 bch2_btree_update_get_open_buckets(as, n3);
1603 bch2_btree_node_write(c, n3, SIX_LOCK_intent, 0);
1606 bch2_btree_update_get_open_buckets(as, n2);
1607 bch2_btree_node_write(c, n2, SIX_LOCK_intent, 0);
1609 bch2_btree_update_get_open_buckets(as, n1);
1610 bch2_btree_node_write(c, n1, SIX_LOCK_intent, 0);
1613 * The old node must be freed (in memory) _before_ unlocking the new
1614 * nodes - else another thread could re-acquire a read lock on the old
1615 * node after another thread has locked and updated the new node, thus
1616 * seeing stale data:
1618 bch2_btree_node_free_inmem(trans, path, b);
1621 bch2_trans_node_add(trans, n3);
1623 bch2_trans_node_add(trans, n2);
1624 bch2_trans_node_add(trans, n1);
1627 six_unlock_intent(&n3->c.lock);
1629 six_unlock_intent(&n2->c.lock);
1630 six_unlock_intent(&n1->c.lock);
1633 __bch2_btree_path_unlock(trans, path2);
1634 bch2_path_put(trans, path2, true);
1637 __bch2_btree_path_unlock(trans, path1);
1638 bch2_path_put(trans, path1, true);
1641 bch2_trans_verify_locks(trans);
1643 bch2_time_stats_update(&c->times[n2
1644 ? BCH_TIME_btree_node_split
1645 : BCH_TIME_btree_node_compact],
1650 bch2_btree_node_free_never_used(as, trans, n3);
1652 bch2_btree_node_free_never_used(as, trans, n2);
1653 bch2_btree_node_free_never_used(as, trans, n1);
1658 bch2_btree_insert_keys_interior(struct btree_update *as,
1659 struct btree_trans *trans,
1660 struct btree_path *path,
1662 struct keylist *keys)
1664 struct btree_path *linked;
1666 __bch2_btree_insert_keys_interior(as, trans, path, b,
1667 path->l[b->c.level].iter, keys);
1669 btree_update_updated_node(as, b);
1671 trans_for_each_path_with_node(trans, b, linked)
1672 bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b);
1674 bch2_trans_verify_paths(trans);
1678 * bch_btree_insert_node - insert bkeys into a given btree node
1680 * @iter: btree iterator
1681 * @keys: list of keys to insert
1682 * @hook: insert callback
1683 * @persistent: if not null, @persistent will wait on journal write
1685 * Inserts as many keys as it can into a given btree node, splitting it if full.
1686 * If a split occurred, this function will return early. This can only happen
1687 * for leaf nodes -- inserts into interior nodes have to be atomic.
1689 static int bch2_btree_insert_node(struct btree_update *as, struct btree_trans *trans,
1690 struct btree_path *path, struct btree *b,
1691 struct keylist *keys, unsigned flags)
1693 struct bch_fs *c = as->c;
1694 int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1695 int old_live_u64s = b->nr.live_u64s;
1696 int live_u64s_added, u64s_added;
1699 lockdep_assert_held(&c->gc_lock);
1700 BUG_ON(!btree_node_intent_locked(path, b->c.level));
1701 BUG_ON(!b->c.level);
1702 BUG_ON(!as || as->b);
1703 bch2_verify_keylist_sorted(keys);
1705 if (!(local_clock() & 63))
1706 return btree_trans_restart(trans, BCH_ERR_transaction_restart_split_race);
1708 ret = bch2_btree_node_lock_write(trans, path, &b->c);
1712 bch2_btree_node_prep_for_write(trans, path, b);
1714 if (!bch2_btree_node_insert_fits(c, b, bch2_keylist_u64s(keys))) {
1715 bch2_btree_node_unlock_write(trans, path, b);
1719 btree_node_interior_verify(c, b);
1721 bch2_btree_insert_keys_interior(as, trans, path, b, keys);
1723 live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1724 u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1726 if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1727 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1728 if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1729 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1731 if (u64s_added > live_u64s_added &&
1732 bch2_maybe_compact_whiteouts(c, b))
1733 bch2_trans_node_reinit_iter(trans, b);
1735 bch2_btree_node_unlock_write(trans, path, b);
1737 btree_node_interior_verify(c, b);
1741 * We could attempt to avoid the transaction restart, by calling
1742 * bch2_btree_path_upgrade() and allocating more nodes:
1744 if (b->c.level >= as->update_level)
1745 return btree_trans_restart(trans, BCH_ERR_transaction_restart_split_race);
1747 return btree_split(as, trans, path, b, keys, flags);
1750 int bch2_btree_split_leaf(struct btree_trans *trans,
1751 struct btree_path *path,
1754 struct btree *b = path_l(path)->b;
1755 struct btree_update *as;
1759 as = bch2_btree_update_start(trans, path, path->level,
1764 ret = btree_split(as, trans, path, b, NULL, flags);
1766 bch2_btree_update_free(as, trans);
1770 bch2_btree_update_done(as, trans);
1772 for (l = path->level + 1; btree_node_intent_locked(path, l) && !ret; l++)
1773 ret = bch2_foreground_maybe_merge(trans, path, l, flags);
1778 int __bch2_foreground_maybe_merge(struct btree_trans *trans,
1779 struct btree_path *path,
1782 enum btree_node_sibling sib)
1784 struct bch_fs *c = trans->c;
1785 struct btree_path *sib_path = NULL, *new_path = NULL;
1786 struct btree_update *as;
1787 struct bkey_format_state new_s;
1788 struct bkey_format new_f;
1789 struct bkey_i delete;
1790 struct btree *b, *m, *n, *prev, *next, *parent;
1791 struct bpos sib_pos;
1793 u64 start_time = local_clock();
1796 BUG_ON(!path->should_be_locked);
1797 BUG_ON(!btree_node_locked(path, level));
1799 b = path->l[level].b;
1801 if ((sib == btree_prev_sib && !bpos_cmp(b->data->min_key, POS_MIN)) ||
1802 (sib == btree_next_sib && !bpos_cmp(b->data->max_key, SPOS_MAX))) {
1803 b->sib_u64s[sib] = U16_MAX;
1807 sib_pos = sib == btree_prev_sib
1808 ? bpos_predecessor(b->data->min_key)
1809 : bpos_successor(b->data->max_key);
1811 sib_path = bch2_path_get(trans, path->btree_id, sib_pos,
1812 U8_MAX, level, BTREE_ITER_INTENT, _THIS_IP_);
1813 ret = bch2_btree_path_traverse(trans, sib_path, false);
1817 btree_path_set_should_be_locked(sib_path);
1819 m = sib_path->l[level].b;
1821 if (btree_node_parent(path, b) !=
1822 btree_node_parent(sib_path, m)) {
1823 b->sib_u64s[sib] = U16_MAX;
1827 if (sib == btree_prev_sib) {
1835 if (bkey_cmp(bpos_successor(prev->data->max_key), next->data->min_key)) {
1836 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
1838 bch2_bpos_to_text(&buf1, prev->data->max_key);
1839 bch2_bpos_to_text(&buf2, next->data->min_key);
1841 "%s(): btree topology error:\n"
1842 " prev ends at %s\n"
1843 " next starts at %s",
1844 __func__, buf1.buf, buf2.buf);
1845 printbuf_exit(&buf1);
1846 printbuf_exit(&buf2);
1847 bch2_topology_error(c);
1852 bch2_bkey_format_init(&new_s);
1853 bch2_bkey_format_add_pos(&new_s, prev->data->min_key);
1854 __bch2_btree_calc_format(&new_s, prev);
1855 __bch2_btree_calc_format(&new_s, next);
1856 bch2_bkey_format_add_pos(&new_s, next->data->max_key);
1857 new_f = bch2_bkey_format_done(&new_s);
1859 sib_u64s = btree_node_u64s_with_format(b, &new_f) +
1860 btree_node_u64s_with_format(m, &new_f);
1862 if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1863 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1865 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1868 sib_u64s = min(sib_u64s, btree_max_u64s(c));
1869 sib_u64s = min(sib_u64s, (size_t) U16_MAX - 1);
1870 b->sib_u64s[sib] = sib_u64s;
1872 if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold)
1875 parent = btree_node_parent(path, b);
1876 as = bch2_btree_update_start(trans, path, level, false,
1877 BTREE_INSERT_NOFAIL|
1878 BTREE_INSERT_USE_RESERVE|
1880 ret = PTR_ERR_OR_ZERO(as);
1884 trace_and_count(c, btree_node_merge, c, b);
1886 bch2_btree_interior_update_will_free_node(as, b);
1887 bch2_btree_interior_update_will_free_node(as, m);
1889 n = bch2_btree_node_alloc(as, trans, b->c.level);
1891 SET_BTREE_NODE_SEQ(n->data,
1892 max(BTREE_NODE_SEQ(b->data),
1893 BTREE_NODE_SEQ(m->data)) + 1);
1895 btree_set_min(n, prev->data->min_key);
1896 btree_set_max(n, next->data->max_key);
1898 n->data->format = new_f;
1899 btree_node_set_format(n, new_f);
1901 bch2_btree_sort_into(c, n, prev);
1902 bch2_btree_sort_into(c, n, next);
1904 bch2_btree_build_aux_trees(n);
1905 bch2_btree_update_add_new_node(as, n);
1906 six_unlock_write(&n->c.lock);
1908 new_path = get_unlocked_mut_path(trans, path->btree_id, n->c.level, n->key.k.p);
1909 six_lock_increment(&n->c.lock, SIX_LOCK_intent);
1910 mark_btree_node_locked(trans, new_path, n->c.level, SIX_LOCK_intent);
1911 bch2_btree_path_level_init(trans, new_path, n);
1913 bkey_init(&delete.k);
1914 delete.k.p = prev->key.k.p;
1915 bch2_keylist_add(&as->parent_keys, &delete);
1916 bch2_keylist_add(&as->parent_keys, &n->key);
1918 bch2_trans_verify_paths(trans);
1920 ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1922 goto err_free_update;
1924 bch2_trans_verify_paths(trans);
1926 bch2_btree_update_get_open_buckets(as, n);
1927 bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1929 bch2_btree_node_free_inmem(trans, path, b);
1930 bch2_btree_node_free_inmem(trans, sib_path, m);
1932 bch2_trans_node_add(trans, n);
1934 bch2_trans_verify_paths(trans);
1936 six_unlock_intent(&n->c.lock);
1938 bch2_btree_update_done(as, trans);
1940 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_merge], start_time);
1944 bch2_path_put(trans, new_path, true);
1945 bch2_path_put(trans, sib_path, true);
1946 bch2_trans_verify_locks(trans);
1949 bch2_btree_node_free_never_used(as, trans, n);
1950 bch2_btree_update_free(as, trans);
1955 * bch_btree_node_rewrite - Rewrite/move a btree node
1957 int bch2_btree_node_rewrite(struct btree_trans *trans,
1958 struct btree_iter *iter,
1962 struct bch_fs *c = trans->c;
1963 struct btree_path *new_path = NULL;
1964 struct btree *n, *parent;
1965 struct btree_update *as;
1968 flags |= BTREE_INSERT_NOFAIL;
1970 parent = btree_node_parent(iter->path, b);
1971 as = bch2_btree_update_start(trans, iter->path, b->c.level,
1973 ret = PTR_ERR_OR_ZERO(as);
1977 bch2_btree_interior_update_will_free_node(as, b);
1979 n = bch2_btree_node_alloc_replacement(as, trans, b);
1981 bch2_btree_build_aux_trees(n);
1982 bch2_btree_update_add_new_node(as, n);
1983 six_unlock_write(&n->c.lock);
1985 new_path = get_unlocked_mut_path(trans, iter->btree_id, n->c.level, n->key.k.p);
1986 six_lock_increment(&n->c.lock, SIX_LOCK_intent);
1987 mark_btree_node_locked(trans, new_path, n->c.level, SIX_LOCK_intent);
1988 bch2_btree_path_level_init(trans, new_path, n);
1990 trace_and_count(c, btree_node_rewrite, c, b);
1993 bch2_keylist_add(&as->parent_keys, &n->key);
1994 ret = bch2_btree_insert_node(as, trans, iter->path, parent,
1995 &as->parent_keys, flags);
1999 bch2_btree_set_root(as, trans, iter->path, n);
2002 bch2_btree_update_get_open_buckets(as, n);
2003 bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
2005 bch2_btree_node_free_inmem(trans, iter->path, b);
2007 bch2_trans_node_add(trans, n);
2008 six_unlock_intent(&n->c.lock);
2010 bch2_btree_update_done(as, trans);
2013 bch2_path_put(trans, new_path, true);
2014 bch2_btree_path_downgrade(trans, iter->path);
2017 bch2_btree_node_free_never_used(as, trans, n);
2018 bch2_btree_update_free(as, trans);
2022 struct async_btree_rewrite {
2024 struct work_struct work;
2025 enum btree_id btree_id;
2031 static int async_btree_node_rewrite_trans(struct btree_trans *trans,
2032 struct async_btree_rewrite *a)
2034 struct btree_iter iter;
2038 bch2_trans_node_iter_init(trans, &iter, a->btree_id, a->pos,
2039 BTREE_MAX_DEPTH, a->level, 0);
2040 b = bch2_btree_iter_peek_node(&iter);
2041 ret = PTR_ERR_OR_ZERO(b);
2045 if (!b || b->data->keys.seq != a->seq)
2048 ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
2050 bch2_trans_iter_exit(trans, &iter);
2055 void async_btree_node_rewrite_work(struct work_struct *work)
2057 struct async_btree_rewrite *a =
2058 container_of(work, struct async_btree_rewrite, work);
2059 struct bch_fs *c = a->c;
2061 bch2_trans_do(c, NULL, NULL, 0,
2062 async_btree_node_rewrite_trans(&trans, a));
2063 percpu_ref_put(&c->writes);
2067 void bch2_btree_node_rewrite_async(struct bch_fs *c, struct btree *b)
2069 struct async_btree_rewrite *a;
2071 if (!percpu_ref_tryget_live(&c->writes))
2074 a = kmalloc(sizeof(*a), GFP_NOFS);
2076 percpu_ref_put(&c->writes);
2081 a->btree_id = b->c.btree_id;
2082 a->level = b->c.level;
2083 a->pos = b->key.k.p;
2084 a->seq = b->data->keys.seq;
2086 INIT_WORK(&a->work, async_btree_node_rewrite_work);
2087 queue_work(c->btree_interior_update_worker, &a->work);
2090 static int __bch2_btree_node_update_key(struct btree_trans *trans,
2091 struct btree_iter *iter,
2092 struct btree *b, struct btree *new_hash,
2093 struct bkey_i *new_key,
2096 struct bch_fs *c = trans->c;
2097 struct btree_iter iter2 = { NULL };
2098 struct btree *parent;
2101 if (!skip_triggers) {
2102 ret = bch2_trans_mark_old(trans, b->c.btree_id, b->c.level + 1,
2103 bkey_i_to_s_c(&b->key), 0);
2107 ret = bch2_trans_mark_new(trans, b->c.btree_id, b->c.level + 1,
2114 bkey_copy(&new_hash->key, new_key);
2115 ret = bch2_btree_node_hash_insert(&c->btree_cache,
2116 new_hash, b->c.level, b->c.btree_id);
2120 parent = btree_node_parent(iter->path, b);
2122 bch2_trans_copy_iter(&iter2, iter);
2124 iter2.path = bch2_btree_path_make_mut(trans, iter2.path,
2125 iter2.flags & BTREE_ITER_INTENT,
2128 BUG_ON(iter2.path->level != b->c.level);
2129 BUG_ON(bpos_cmp(iter2.path->pos, new_key->k.p));
2131 btree_path_set_level_up(trans, iter2.path);
2133 bch2_btree_path_check_sort(trans, iter2.path, 0);
2135 ret = bch2_btree_iter_traverse(&iter2) ?:
2136 bch2_trans_update(trans, &iter2, new_key, BTREE_TRIGGER_NORUN);
2140 BUG_ON(btree_node_root(c, b) != b);
2142 ret = darray_make_room(&trans->extra_journal_entries,
2143 jset_u64s(new_key->k.u64s));
2147 journal_entry_set((void *) &darray_top(trans->extra_journal_entries),
2148 BCH_JSET_ENTRY_btree_root,
2149 b->c.btree_id, b->c.level,
2150 new_key, new_key->k.u64s);
2151 trans->extra_journal_entries.nr += jset_u64s(new_key->k.u64s);
2154 ret = bch2_trans_commit(trans, NULL, NULL,
2155 BTREE_INSERT_NOFAIL|
2156 BTREE_INSERT_NOCHECK_RW|
2157 BTREE_INSERT_USE_RESERVE|
2158 BTREE_INSERT_JOURNAL_RECLAIM|
2159 JOURNAL_WATERMARK_reserved);
2163 bch2_btree_node_lock_write_nofail(trans, iter->path, &b->c);
2166 mutex_lock(&c->btree_cache.lock);
2167 bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
2168 bch2_btree_node_hash_remove(&c->btree_cache, b);
2170 bkey_copy(&b->key, new_key);
2171 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
2173 mutex_unlock(&c->btree_cache.lock);
2175 bkey_copy(&b->key, new_key);
2178 bch2_btree_node_unlock_write(trans, iter->path, b);
2180 bch2_trans_iter_exit(trans, &iter2);
2184 mutex_lock(&c->btree_cache.lock);
2185 bch2_btree_node_hash_remove(&c->btree_cache, b);
2186 mutex_unlock(&c->btree_cache.lock);
2191 int bch2_btree_node_update_key(struct btree_trans *trans, struct btree_iter *iter,
2192 struct btree *b, struct bkey_i *new_key,
2195 struct bch_fs *c = trans->c;
2196 struct btree *new_hash = NULL;
2197 struct btree_path *path = iter->path;
2201 ret = bch2_btree_path_upgrade(trans, path, b->c.level + 1);
2205 closure_init_stack(&cl);
2208 * check btree_ptr_hash_val() after @b is locked by
2209 * btree_iter_traverse():
2211 if (btree_ptr_hash_val(new_key) != b->hash_val) {
2212 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2214 bch2_trans_unlock(trans);
2216 ret = bch2_trans_relock(trans);
2221 new_hash = bch2_btree_node_mem_alloc(c, false);
2225 ret = __bch2_btree_node_update_key(trans, iter, b, new_hash,
2226 new_key, skip_triggers);
2230 mutex_lock(&c->btree_cache.lock);
2231 list_move(&new_hash->list, &c->btree_cache.freeable);
2232 mutex_unlock(&c->btree_cache.lock);
2234 six_unlock_write(&new_hash->c.lock);
2235 six_unlock_intent(&new_hash->c.lock);
2238 bch2_btree_cache_cannibalize_unlock(c);
2242 int bch2_btree_node_update_key_get_iter(struct btree_trans *trans,
2243 struct btree *b, struct bkey_i *new_key,
2246 struct btree_iter iter;
2249 bch2_trans_node_iter_init(trans, &iter, b->c.btree_id, b->key.k.p,
2250 BTREE_MAX_DEPTH, b->c.level,
2252 ret = bch2_btree_iter_traverse(&iter);
2256 /* has node been freed? */
2257 if (iter.path->l[b->c.level].b != b) {
2258 /* node has been freed: */
2259 BUG_ON(!btree_node_dying(b));
2263 BUG_ON(!btree_node_hashed(b));
2265 ret = bch2_btree_node_update_key(trans, &iter, b, new_key, skip_triggers);
2267 bch2_trans_iter_exit(trans, &iter);
2274 * Only for filesystem bringup, when first reading the btree roots or allocating
2275 * btree roots when initializing a new filesystem:
2277 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
2279 BUG_ON(btree_node_root(c, b));
2281 bch2_btree_set_root_inmem(c, b);
2284 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
2290 closure_init_stack(&cl);
2293 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2297 b = bch2_btree_node_mem_alloc(c, false);
2298 bch2_btree_cache_cannibalize_unlock(c);
2300 set_btree_node_fake(b);
2301 set_btree_node_need_rewrite(b);
2305 bkey_btree_ptr_init(&b->key);
2306 b->key.k.p = SPOS_MAX;
2307 *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
2309 bch2_bset_init_first(b, &b->data->keys);
2310 bch2_btree_build_aux_trees(b);
2313 btree_set_min(b, POS_MIN);
2314 btree_set_max(b, SPOS_MAX);
2315 b->data->format = bch2_btree_calc_format(b);
2316 btree_node_set_format(b, b->data->format);
2318 ret = bch2_btree_node_hash_insert(&c->btree_cache, b,
2319 b->c.level, b->c.btree_id);
2322 bch2_btree_set_root_inmem(c, b);
2324 six_unlock_write(&b->c.lock);
2325 six_unlock_intent(&b->c.lock);
2328 void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c)
2330 struct btree_update *as;
2332 mutex_lock(&c->btree_interior_update_lock);
2333 list_for_each_entry(as, &c->btree_interior_update_list, list)
2334 prt_printf(out, "%p m %u w %u r %u j %llu\n",
2338 atomic_read(&as->cl.remaining) & CLOSURE_REMAINING_MASK,
2340 mutex_unlock(&c->btree_interior_update_lock);
2343 static bool bch2_btree_interior_updates_pending(struct bch_fs *c)
2347 mutex_lock(&c->btree_interior_update_lock);
2348 ret = !list_empty(&c->btree_interior_update_list);
2349 mutex_unlock(&c->btree_interior_update_lock);
2354 bool bch2_btree_interior_updates_flush(struct bch_fs *c)
2356 bool ret = bch2_btree_interior_updates_pending(c);
2359 closure_wait_event(&c->btree_interior_update_wait,
2360 !bch2_btree_interior_updates_pending(c));
2364 void bch2_journal_entries_to_btree_roots(struct bch_fs *c, struct jset *jset)
2366 struct btree_root *r;
2367 struct jset_entry *entry;
2369 mutex_lock(&c->btree_root_lock);
2371 vstruct_for_each(jset, entry)
2372 if (entry->type == BCH_JSET_ENTRY_btree_root) {
2373 r = &c->btree_roots[entry->btree_id];
2374 r->level = entry->level;
2376 bkey_copy(&r->key, &entry->start[0]);
2379 mutex_unlock(&c->btree_root_lock);
2383 bch2_btree_roots_to_journal_entries(struct bch_fs *c,
2384 struct jset_entry *start,
2385 struct jset_entry *end)
2387 struct jset_entry *entry;
2388 unsigned long have = 0;
2391 for (entry = start; entry < end; entry = vstruct_next(entry))
2392 if (entry->type == BCH_JSET_ENTRY_btree_root)
2393 __set_bit(entry->btree_id, &have);
2395 mutex_lock(&c->btree_root_lock);
2397 for (i = 0; i < BTREE_ID_NR; i++)
2398 if (c->btree_roots[i].alive && !test_bit(i, &have)) {
2399 journal_entry_set(end,
2400 BCH_JSET_ENTRY_btree_root,
2401 i, c->btree_roots[i].level,
2402 &c->btree_roots[i].key,
2403 c->btree_roots[i].key.u64s);
2404 end = vstruct_next(end);
2407 mutex_unlock(&c->btree_root_lock);
2412 void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
2414 if (c->btree_interior_update_worker)
2415 destroy_workqueue(c->btree_interior_update_worker);
2416 mempool_exit(&c->btree_interior_update_pool);
2419 int bch2_fs_btree_interior_update_init(struct bch_fs *c)
2421 mutex_init(&c->btree_reserve_cache_lock);
2422 INIT_LIST_HEAD(&c->btree_interior_update_list);
2423 INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
2424 mutex_init(&c->btree_interior_update_lock);
2425 INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
2427 c->btree_interior_update_worker =
2428 alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
2429 if (!c->btree_interior_update_worker)
2432 return mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
2433 sizeof(struct btree_update));