3 #include "alloc_foreground.h"
4 #include "bkey_methods.h"
5 #include "btree_cache.h"
7 #include "btree_update.h"
8 #include "btree_update_interior.h"
10 #include "btree_iter.h"
11 #include "btree_locking.h"
15 #include "journal_reclaim.h"
20 #include <linux/random.h>
21 #include <trace/events/bcachefs.h>
23 static void btree_node_will_make_reachable(struct btree_update *,
25 static void btree_update_drop_new_node(struct bch_fs *, struct btree *);
26 static void bch2_btree_set_root_ondisk(struct bch_fs *, struct btree *, int);
30 static void btree_node_interior_verify(struct btree *b)
32 struct btree_node_iter iter;
33 struct bkey_packed *k;
37 bch2_btree_node_iter_init(&iter, b, &b->key.k.p);
39 BUG_ON(!(k = bch2_btree_node_iter_peek(&iter, b)) ||
40 bkey_cmp_left_packed(b, k, &b->key.k.p));
42 BUG_ON((bch2_btree_node_iter_advance(&iter, b),
43 !bch2_btree_node_iter_end(&iter)));
48 k = bch2_btree_node_iter_peek(&iter, b);
52 msg = "isn't what it should be";
53 if (bkey_cmp_left_packed(b, k, &b->key.k.p))
56 bch2_btree_node_iter_advance(&iter, b);
58 msg = "isn't last key";
59 if (!bch2_btree_node_iter_end(&iter))
63 bch2_dump_btree_node(b);
64 printk(KERN_ERR "last key %llu:%llu %s\n", b->key.k.p.inode,
65 b->key.k.p.offset, msg);
70 /* Calculate ideal packed bkey format for new btree nodes: */
72 void __bch2_btree_calc_format(struct bkey_format_state *s, struct btree *b)
74 struct bkey_packed *k;
78 bch2_bkey_format_add_pos(s, b->data->min_key);
81 for (k = btree_bkey_first(b, t);
82 k != btree_bkey_last(b, t);
84 if (!bkey_whiteout(k)) {
85 uk = bkey_unpack_key(b, k);
86 bch2_bkey_format_add_key(s, &uk);
90 static struct bkey_format bch2_btree_calc_format(struct btree *b)
92 struct bkey_format_state s;
94 bch2_bkey_format_init(&s);
95 __bch2_btree_calc_format(&s, b);
97 return bch2_bkey_format_done(&s);
100 static size_t btree_node_u64s_with_format(struct btree *b,
101 struct bkey_format *new_f)
103 struct bkey_format *old_f = &b->format;
105 /* stupid integer promotion rules */
107 (((int) new_f->key_u64s - old_f->key_u64s) *
108 (int) b->nr.packed_keys) +
109 (((int) new_f->key_u64s - BKEY_U64s) *
110 (int) b->nr.unpacked_keys);
112 BUG_ON(delta + b->nr.live_u64s < 0);
114 return b->nr.live_u64s + delta;
118 * btree_node_format_fits - check if we could rewrite node with a new format
120 * This assumes all keys can pack with the new format -- it just checks if
121 * the re-packed keys would fit inside the node itself.
123 bool bch2_btree_node_format_fits(struct bch_fs *c, struct btree *b,
124 struct bkey_format *new_f)
126 size_t u64s = btree_node_u64s_with_format(b, new_f);
128 return __vstruct_bytes(struct btree_node, u64s) < btree_bytes(c);
131 /* Btree node freeing/allocation: */
133 static bool btree_key_matches(struct bch_fs *c,
134 struct bkey_s_c_extent l,
135 struct bkey_s_c_extent r)
137 const struct bch_extent_ptr *ptr1, *ptr2;
139 extent_for_each_ptr(l, ptr1)
140 extent_for_each_ptr(r, ptr2)
141 if (ptr1->dev == ptr2->dev &&
142 ptr1->gen == ptr2->gen &&
143 ptr1->offset == ptr2->offset)
150 * We're doing the index update that makes @b unreachable, update stuff to
153 * Must be called _before_ btree_update_updated_root() or
154 * btree_update_updated_node:
156 static void bch2_btree_node_free_index(struct btree_update *as, struct btree *b,
158 struct bch_fs_usage *stats)
160 struct bch_fs *c = as->c;
161 struct pending_btree_node_free *d;
165 * btree_update lock is only needed here to avoid racing with
168 mutex_lock(&c->btree_interior_update_lock);
170 for (d = as->pending; d < as->pending + as->nr_pending; d++)
171 if (!bkey_cmp(k.k->p, d->key.k.p) &&
172 btree_key_matches(c, bkey_s_c_to_extent(k),
173 bkey_i_to_s_c_extent(&d->key)))
177 BUG_ON(d->index_update_done);
178 d->index_update_done = true;
181 * Btree nodes are accounted as freed in bch_alloc_stats when they're
182 * freed from the index:
184 replicas = bch2_extent_nr_dirty_ptrs(k);
186 stats->replicas[replicas - 1].data[BCH_DATA_BTREE] -=
187 c->opts.btree_node_size * replicas;
190 * We're dropping @k from the btree, but it's still live until the
191 * index update is persistent so we need to keep a reference around for
192 * mark and sweep to find - that's primarily what the
193 * btree_node_pending_free list is for.
195 * So here (when we set index_update_done = true), we're moving an
196 * existing reference to a different part of the larger "gc keyspace" -
197 * and the new position comes after the old position, since GC marks
198 * the pending free list after it walks the btree.
200 * If we move the reference while mark and sweep is _between_ the old
201 * and the new position, mark and sweep will see the reference twice
202 * and it'll get double accounted - so check for that here and subtract
203 * to cancel out one of mark and sweep's markings if necessary:
207 * bch2_mark_key() compares the current gc pos to the pos we're
208 * moving this reference from, hence one comparison here:
210 if (gc_pos_cmp(c->gc_pos, gc_phase(GC_PHASE_PENDING_DELETE)) < 0) {
211 struct bch_fs_usage tmp = { 0 };
213 bch2_mark_key(c, BKEY_TYPE_BTREE,
214 bkey_i_to_s_c(&d->key),
216 ? gc_pos_btree_node(b)
217 : gc_pos_btree_root(as->btree_id),
220 * Don't apply tmp - pending deletes aren't tracked in
225 mutex_unlock(&c->btree_interior_update_lock);
228 static void __btree_node_free(struct bch_fs *c, struct btree *b)
230 trace_btree_node_free(c, b);
232 BUG_ON(btree_node_dirty(b));
233 BUG_ON(btree_node_need_write(b));
234 BUG_ON(b == btree_node_root(c, b));
236 BUG_ON(!list_empty(&b->write_blocked));
237 BUG_ON(b->will_make_reachable);
239 clear_btree_node_noevict(b);
241 bch2_btree_node_hash_remove(&c->btree_cache, b);
243 mutex_lock(&c->btree_cache.lock);
244 list_move(&b->list, &c->btree_cache.freeable);
245 mutex_unlock(&c->btree_cache.lock);
248 void bch2_btree_node_free_never_inserted(struct bch_fs *c, struct btree *b)
250 struct open_buckets ob = b->ob;
252 btree_update_drop_new_node(c, b);
256 clear_btree_node_dirty(b);
258 btree_node_lock_type(c, b, SIX_LOCK_write);
259 __btree_node_free(c, b);
260 six_unlock_write(&b->lock);
262 bch2_open_buckets_put(c, &ob);
265 void bch2_btree_node_free_inmem(struct bch_fs *c, struct btree *b,
266 struct btree_iter *iter)
269 * Is this a node that isn't reachable on disk yet?
271 * Nodes that aren't reachable yet have writes blocked until they're
272 * reachable - now that we've cancelled any pending writes and moved
273 * things waiting on that write to wait on this update, we can drop this
274 * node from the list of nodes that the other update is making
275 * reachable, prior to freeing it:
277 btree_update_drop_new_node(c, b);
279 __bch2_btree_node_lock_write(b, iter);
280 __btree_node_free(c, b);
281 six_unlock_write(&b->lock);
283 bch2_btree_iter_node_drop(iter, b);
286 static void bch2_btree_node_free_ondisk(struct bch_fs *c,
287 struct pending_btree_node_free *pending)
289 struct bch_fs_usage stats = { 0 };
291 BUG_ON(!pending->index_update_done);
293 bch2_mark_key(c, BKEY_TYPE_BTREE,
294 bkey_i_to_s_c(&pending->key),
296 gc_phase(GC_PHASE_PENDING_DELETE),
299 * Don't apply stats - pending deletes aren't tracked in
304 static struct btree *__bch2_btree_node_alloc(struct bch_fs *c,
305 struct disk_reservation *res,
309 struct write_point *wp;
312 struct bkey_i_extent *e;
313 struct open_buckets ob = { .nr = 0 };
314 struct bch_devs_list devs_have = (struct bch_devs_list) { 0 };
316 enum alloc_reserve alloc_reserve;
318 if (flags & BTREE_INSERT_USE_ALLOC_RESERVE) {
320 alloc_reserve = RESERVE_ALLOC;
321 } else if (flags & BTREE_INSERT_USE_RESERVE) {
322 nr_reserve = BTREE_NODE_RESERVE / 2;
323 alloc_reserve = RESERVE_BTREE;
325 nr_reserve = BTREE_NODE_RESERVE;
326 alloc_reserve = RESERVE_NONE;
329 mutex_lock(&c->btree_reserve_cache_lock);
330 if (c->btree_reserve_cache_nr > nr_reserve) {
331 struct btree_alloc *a =
332 &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
335 bkey_copy(&tmp.k, &a->k);
336 mutex_unlock(&c->btree_reserve_cache_lock);
339 mutex_unlock(&c->btree_reserve_cache_lock);
342 wp = bch2_alloc_sectors_start(c, c->opts.foreground_target,
343 writepoint_ptr(&c->btree_write_point),
346 c->opts.metadata_replicas_required,
347 alloc_reserve, 0, cl);
351 if (wp->sectors_free < c->opts.btree_node_size) {
352 struct open_bucket *ob;
355 open_bucket_for_each(c, &wp->ptrs, ob, i)
356 if (ob->sectors_free < c->opts.btree_node_size)
357 ob->sectors_free = 0;
359 bch2_alloc_sectors_done(c, wp);
363 e = bkey_extent_init(&tmp.k);
364 bch2_alloc_sectors_append_ptrs(c, wp, e, c->opts.btree_node_size);
366 bch2_open_bucket_get(c, wp, &ob);
367 bch2_alloc_sectors_done(c, wp);
369 b = bch2_btree_node_mem_alloc(c);
371 /* we hold cannibalize_lock: */
375 bkey_copy(&b->key, &tmp.k);
381 static struct btree *bch2_btree_node_alloc(struct btree_update *as, unsigned level)
383 struct bch_fs *c = as->c;
386 BUG_ON(level >= BTREE_MAX_DEPTH);
387 BUG_ON(!as->reserve->nr);
389 b = as->reserve->b[--as->reserve->nr];
391 BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, as->btree_id));
393 set_btree_node_accessed(b);
394 set_btree_node_dirty(b);
396 bch2_bset_init_first(b, &b->data->keys);
397 memset(&b->nr, 0, sizeof(b->nr));
398 b->data->magic = cpu_to_le64(bset_magic(c));
400 SET_BTREE_NODE_ID(b->data, as->btree_id);
401 SET_BTREE_NODE_LEVEL(b->data, level);
402 b->data->ptr = bkey_i_to_extent(&b->key)->v.start->ptr;
404 bch2_btree_build_aux_trees(b);
406 btree_node_will_make_reachable(as, b);
408 trace_btree_node_alloc(c, b);
412 struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *as,
414 struct bkey_format format)
418 n = bch2_btree_node_alloc(as, b->level);
420 n->data->min_key = b->data->min_key;
421 n->data->max_key = b->data->max_key;
422 n->data->format = format;
423 SET_BTREE_NODE_SEQ(n->data, BTREE_NODE_SEQ(b->data) + 1);
425 btree_node_set_format(n, format);
427 bch2_btree_sort_into(as->c, n, b);
429 btree_node_reset_sib_u64s(n);
431 n->key.k.p = b->key.k.p;
435 static struct btree *bch2_btree_node_alloc_replacement(struct btree_update *as,
438 struct bkey_format new_f = bch2_btree_calc_format(b);
441 * The keys might expand with the new format - if they wouldn't fit in
442 * the btree node anymore, use the old format for now:
444 if (!bch2_btree_node_format_fits(as->c, b, &new_f))
447 return __bch2_btree_node_alloc_replacement(as, b, new_f);
450 static struct btree *__btree_root_alloc(struct btree_update *as, unsigned level)
452 struct btree *b = bch2_btree_node_alloc(as, level);
454 b->data->min_key = POS_MIN;
455 b->data->max_key = POS_MAX;
456 b->data->format = bch2_btree_calc_format(b);
457 b->key.k.p = POS_MAX;
459 btree_node_set_format(b, b->data->format);
460 bch2_btree_build_aux_trees(b);
462 six_unlock_write(&b->lock);
467 static void bch2_btree_reserve_put(struct bch_fs *c, struct btree_reserve *reserve)
469 bch2_disk_reservation_put(c, &reserve->disk_res);
471 mutex_lock(&c->btree_reserve_cache_lock);
473 while (reserve->nr) {
474 struct btree *b = reserve->b[--reserve->nr];
476 six_unlock_write(&b->lock);
478 if (c->btree_reserve_cache_nr <
479 ARRAY_SIZE(c->btree_reserve_cache)) {
480 struct btree_alloc *a =
481 &c->btree_reserve_cache[c->btree_reserve_cache_nr++];
485 bkey_copy(&a->k, &b->key);
487 bch2_open_buckets_put(c, &b->ob);
490 btree_node_lock_type(c, b, SIX_LOCK_write);
491 __btree_node_free(c, b);
492 six_unlock_write(&b->lock);
494 six_unlock_intent(&b->lock);
497 mutex_unlock(&c->btree_reserve_cache_lock);
499 mempool_free(reserve, &c->btree_reserve_pool);
502 static struct btree_reserve *bch2_btree_reserve_get(struct bch_fs *c,
507 struct btree_reserve *reserve;
509 struct disk_reservation disk_res = { 0, 0 };
510 unsigned sectors = nr_nodes * c->opts.btree_node_size;
511 int ret, disk_res_flags = BCH_DISK_RESERVATION_GC_LOCK_HELD;
513 if (flags & BTREE_INSERT_NOFAIL)
514 disk_res_flags |= BCH_DISK_RESERVATION_NOFAIL;
517 * This check isn't necessary for correctness - it's just to potentially
518 * prevent us from doing a lot of work that'll end up being wasted:
520 ret = bch2_journal_error(&c->journal);
524 if (bch2_disk_reservation_get(c, &disk_res, sectors,
525 c->opts.metadata_replicas,
527 return ERR_PTR(-ENOSPC);
529 BUG_ON(nr_nodes > BTREE_RESERVE_MAX);
532 * Protects reaping from the btree node cache and using the btree node
533 * open bucket reserve:
535 ret = bch2_btree_cache_cannibalize_lock(c, cl);
537 bch2_disk_reservation_put(c, &disk_res);
541 reserve = mempool_alloc(&c->btree_reserve_pool, GFP_NOIO);
543 reserve->disk_res = disk_res;
546 while (reserve->nr < nr_nodes) {
547 b = __bch2_btree_node_alloc(c, &disk_res,
548 flags & BTREE_INSERT_NOWAIT
555 ret = bch2_mark_bkey_replicas(c, BKEY_TYPE_BTREE,
556 bkey_i_to_s_c(&b->key));
560 reserve->b[reserve->nr++] = b;
563 bch2_btree_cache_cannibalize_unlock(c);
566 bch2_btree_reserve_put(c, reserve);
567 bch2_btree_cache_cannibalize_unlock(c);
568 trace_btree_reserve_get_fail(c, nr_nodes, cl);
572 /* Asynchronous interior node update machinery */
574 static void bch2_btree_update_free(struct btree_update *as)
576 struct bch_fs *c = as->c;
578 bch2_journal_pin_flush(&c->journal, &as->journal);
580 BUG_ON(as->nr_new_nodes);
581 BUG_ON(as->nr_pending);
584 bch2_btree_reserve_put(c, as->reserve);
586 mutex_lock(&c->btree_interior_update_lock);
589 closure_debug_destroy(&as->cl);
590 mempool_free(as, &c->btree_interior_update_pool);
591 percpu_ref_put(&c->writes);
593 closure_wake_up(&c->btree_interior_update_wait);
594 mutex_unlock(&c->btree_interior_update_lock);
597 static void btree_update_nodes_reachable(struct closure *cl)
599 struct btree_update *as = container_of(cl, struct btree_update, cl);
600 struct bch_fs *c = as->c;
602 bch2_journal_pin_drop(&c->journal, &as->journal);
604 mutex_lock(&c->btree_interior_update_lock);
606 while (as->nr_new_nodes) {
607 struct btree *b = as->new_nodes[--as->nr_new_nodes];
609 BUG_ON(b->will_make_reachable != (unsigned long) as);
610 b->will_make_reachable = 0;
611 mutex_unlock(&c->btree_interior_update_lock);
614 * b->will_make_reachable prevented it from being written, so
615 * write it now if it needs to be written:
617 btree_node_lock_type(c, b, SIX_LOCK_read);
618 bch2_btree_node_write_cond(c, b, btree_node_need_write(b));
619 six_unlock_read(&b->lock);
620 mutex_lock(&c->btree_interior_update_lock);
623 while (as->nr_pending)
624 bch2_btree_node_free_ondisk(c, &as->pending[--as->nr_pending]);
626 mutex_unlock(&c->btree_interior_update_lock);
628 closure_wake_up(&as->wait);
630 bch2_btree_update_free(as);
633 static void btree_update_wait_on_journal(struct closure *cl)
635 struct btree_update *as = container_of(cl, struct btree_update, cl);
636 struct bch_fs *c = as->c;
639 ret = bch2_journal_open_seq_async(&c->journal, as->journal_seq, cl);
643 continue_at(cl, btree_update_wait_on_journal, system_wq);
647 bch2_journal_flush_seq_async(&c->journal, as->journal_seq, cl);
649 continue_at(cl, btree_update_nodes_reachable, system_wq);
652 static void btree_update_nodes_written(struct closure *cl)
654 struct btree_update *as = container_of(cl, struct btree_update, cl);
655 struct bch_fs *c = as->c;
659 * We did an update to a parent node where the pointers we added pointed
660 * to child nodes that weren't written yet: now, the child nodes have
661 * been written so we can write out the update to the interior node.
664 mutex_lock(&c->btree_interior_update_lock);
665 as->nodes_written = true;
668 case BTREE_INTERIOR_NO_UPDATE:
670 case BTREE_INTERIOR_UPDATING_NODE:
671 /* The usual case: */
672 b = READ_ONCE(as->b);
674 if (!six_trylock_read(&b->lock)) {
675 mutex_unlock(&c->btree_interior_update_lock);
676 btree_node_lock_type(c, b, SIX_LOCK_read);
677 six_unlock_read(&b->lock);
681 BUG_ON(!btree_node_dirty(b));
682 closure_wait(&btree_current_write(b)->wait, cl);
684 list_del(&as->write_blocked_list);
685 mutex_unlock(&c->btree_interior_update_lock);
688 * b->write_blocked prevented it from being written, so
689 * write it now if it needs to be written:
691 bch2_btree_node_write_cond(c, b, true);
692 six_unlock_read(&b->lock);
695 case BTREE_INTERIOR_UPDATING_AS:
697 * The btree node we originally updated has been freed and is
698 * being rewritten - so we need to write anything here, we just
699 * need to signal to that btree_update that it's ok to make the
700 * new replacement node visible:
702 closure_put(&as->parent_as->cl);
705 * and then we have to wait on that btree_update to finish:
707 closure_wait(&as->parent_as->wait, cl);
708 mutex_unlock(&c->btree_interior_update_lock);
711 case BTREE_INTERIOR_UPDATING_ROOT:
712 /* b is the new btree root: */
713 b = READ_ONCE(as->b);
715 if (!six_trylock_read(&b->lock)) {
716 mutex_unlock(&c->btree_interior_update_lock);
717 btree_node_lock_type(c, b, SIX_LOCK_read);
718 six_unlock_read(&b->lock);
722 BUG_ON(c->btree_roots[b->btree_id].as != as);
723 c->btree_roots[b->btree_id].as = NULL;
725 bch2_btree_set_root_ondisk(c, b, WRITE);
728 * We don't have to wait anything anything here (before
729 * btree_update_nodes_reachable frees the old nodes
730 * ondisk) - we've ensured that the very next journal write will
731 * have the pointer to the new root, and before the allocator
732 * can reuse the old nodes it'll have to do a journal commit:
734 six_unlock_read(&b->lock);
735 mutex_unlock(&c->btree_interior_update_lock);
738 * Bit of funny circularity going on here we have to break:
740 * We have to drop our journal pin before writing the journal
741 * entry that points to the new btree root: else, we could
742 * deadlock if the journal currently happens to be full.
744 * This mean we're dropping the journal pin _before_ the new
745 * nodes are technically reachable - but this is safe, because
746 * after the bch2_btree_set_root_ondisk() call above they will
747 * be reachable as of the very next journal write:
749 bch2_journal_pin_drop(&c->journal, &as->journal);
751 as->journal_seq = bch2_journal_last_unwritten_seq(&c->journal);
753 btree_update_wait_on_journal(cl);
757 continue_at(cl, btree_update_nodes_reachable, system_wq);
761 * We're updating @b with pointers to nodes that haven't finished writing yet:
762 * block @b from being written until @as completes
764 static void btree_update_updated_node(struct btree_update *as, struct btree *b)
766 struct bch_fs *c = as->c;
768 mutex_lock(&c->btree_interior_update_lock);
770 BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
771 BUG_ON(!btree_node_dirty(b));
773 as->mode = BTREE_INTERIOR_UPDATING_NODE;
775 list_add(&as->write_blocked_list, &b->write_blocked);
777 mutex_unlock(&c->btree_interior_update_lock);
780 * In general, when you're staging things in a journal that will later
781 * be written elsewhere, and you also want to guarantee ordering: that
782 * is, if you have updates a, b, c, after a crash you should never see c
783 * and not a or b - there's a problem:
785 * If the final destination of the update(s) (i.e. btree node) can be
786 * written/flushed _before_ the relevant journal entry - oops, that
787 * breaks ordering, since the various leaf nodes can be written in any
790 * Normally we use bset->journal_seq to deal with this - if during
791 * recovery we find a btree node write that's newer than the newest
792 * journal entry, we just ignore it - we don't need it, anything we're
793 * supposed to have (that we reported as completed via fsync()) will
794 * still be in the journal, and as far as the state of the journal is
795 * concerned that btree node write never happened.
797 * That breaks when we're rewriting/splitting/merging nodes, since we're
798 * mixing btree node writes that haven't happened yet with previously
799 * written data that has been reported as completed to the journal.
801 * Thus, before making the new nodes reachable, we have to wait the
802 * newest journal sequence number we have data for to be written (if it
805 bch2_journal_wait_on_seq(&c->journal, as->journal_seq, &as->cl);
808 static void interior_update_flush(struct journal *j,
809 struct journal_entry_pin *pin, u64 seq)
811 struct btree_update *as =
812 container_of(pin, struct btree_update, journal);
814 bch2_journal_flush_seq_async(j, as->journal_seq, NULL);
817 static void btree_update_reparent(struct btree_update *as,
818 struct btree_update *child)
820 struct bch_fs *c = as->c;
823 child->mode = BTREE_INTERIOR_UPDATING_AS;
824 child->parent_as = as;
825 closure_get(&as->cl);
828 * When we write a new btree root, we have to drop our journal pin
829 * _before_ the new nodes are technically reachable; see
830 * btree_update_nodes_written().
832 * This goes for journal pins that are recursively blocked on us - so,
833 * just transfer the journal pin to the new interior update so
834 * btree_update_nodes_written() can drop it.
836 bch2_journal_pin_add_if_older(&c->journal, &child->journal,
837 &as->journal, interior_update_flush);
838 bch2_journal_pin_drop(&c->journal, &child->journal);
840 as->journal_seq = max(as->journal_seq, child->journal_seq);
843 static void btree_update_updated_root(struct btree_update *as)
845 struct bch_fs *c = as->c;
846 struct btree_root *r = &c->btree_roots[as->btree_id];
848 mutex_lock(&c->btree_interior_update_lock);
850 BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
853 * Old root might not be persistent yet - if so, redirect its
854 * btree_update operation to point to us:
857 btree_update_reparent(as, r->as);
859 as->mode = BTREE_INTERIOR_UPDATING_ROOT;
863 mutex_unlock(&c->btree_interior_update_lock);
866 * When we're rewriting nodes and updating interior nodes, there's an
867 * issue with updates that haven't been written in the journal getting
868 * mixed together with older data - see btree_update_updated_node()
869 * for the explanation.
871 * However, this doesn't affect us when we're writing a new btree root -
872 * because to make that new root reachable we have to write out a new
873 * journal entry, which must necessarily be newer than as->journal_seq.
877 static void btree_node_will_make_reachable(struct btree_update *as,
880 struct bch_fs *c = as->c;
882 mutex_lock(&c->btree_interior_update_lock);
883 BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes));
884 BUG_ON(b->will_make_reachable);
886 as->new_nodes[as->nr_new_nodes++] = b;
887 b->will_make_reachable = 1UL|(unsigned long) as;
889 closure_get(&as->cl);
890 mutex_unlock(&c->btree_interior_update_lock);
893 static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b)
895 struct btree_update *as;
899 mutex_lock(&c->btree_interior_update_lock);
900 v = xchg(&b->will_make_reachable, 0);
901 as = (struct btree_update *) (v & ~1UL);
904 mutex_unlock(&c->btree_interior_update_lock);
908 for (i = 0; i < as->nr_new_nodes; i++)
909 if (as->new_nodes[i] == b)
914 array_remove_item(as->new_nodes, as->nr_new_nodes, i);
915 mutex_unlock(&c->btree_interior_update_lock);
918 closure_put(&as->cl);
921 static void btree_interior_update_add_node_reference(struct btree_update *as,
924 struct bch_fs *c = as->c;
925 struct pending_btree_node_free *d;
927 mutex_lock(&c->btree_interior_update_lock);
929 /* Add this node to the list of nodes being freed: */
930 BUG_ON(as->nr_pending >= ARRAY_SIZE(as->pending));
932 d = &as->pending[as->nr_pending++];
933 d->index_update_done = false;
934 d->seq = b->data->keys.seq;
935 d->btree_id = b->btree_id;
937 bkey_copy(&d->key, &b->key);
939 mutex_unlock(&c->btree_interior_update_lock);
943 * @b is being split/rewritten: it may have pointers to not-yet-written btree
944 * nodes and thus outstanding btree_updates - redirect @b's
945 * btree_updates to point to this btree_update:
947 void bch2_btree_interior_update_will_free_node(struct btree_update *as,
950 struct bch_fs *c = as->c;
951 struct closure *cl, *cl_n;
952 struct btree_update *p, *n;
953 struct btree_write *w;
956 set_btree_node_dying(b);
958 if (btree_node_fake(b))
961 btree_interior_update_add_node_reference(as, b);
964 * Does this node have data that hasn't been written in the journal?
966 * If so, we have to wait for the corresponding journal entry to be
967 * written before making the new nodes reachable - we can't just carry
968 * over the bset->journal_seq tracking, since we'll be mixing those keys
969 * in with keys that aren't in the journal anymore:
972 as->journal_seq = max(as->journal_seq,
973 le64_to_cpu(bset(b, t)->journal_seq));
975 mutex_lock(&c->btree_interior_update_lock);
978 * Does this node have any btree_update operations preventing
979 * it from being written?
981 * If so, redirect them to point to this btree_update: we can
982 * write out our new nodes, but we won't make them visible until those
983 * operations complete
985 list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) {
986 list_del(&p->write_blocked_list);
987 btree_update_reparent(as, p);
990 clear_btree_node_dirty(b);
991 clear_btree_node_need_write(b);
992 w = btree_current_write(b);
995 * Does this node have any btree_update operations waiting on this node
998 * If so, wake them up when this btree_update operation is reachable:
1000 llist_for_each_entry_safe(cl, cl_n, llist_del_all(&w->wait.list), list)
1001 llist_add(&cl->list, &as->wait.list);
1004 * Does this node have unwritten data that has a pin on the journal?
1006 * If so, transfer that pin to the btree_update operation -
1007 * note that if we're freeing multiple nodes, we only need to keep the
1008 * oldest pin of any of the nodes we're freeing. We'll release the pin
1009 * when the new nodes are persistent and reachable on disk:
1011 bch2_journal_pin_add_if_older(&c->journal, &w->journal,
1012 &as->journal, interior_update_flush);
1013 bch2_journal_pin_drop(&c->journal, &w->journal);
1015 w = btree_prev_write(b);
1016 bch2_journal_pin_add_if_older(&c->journal, &w->journal,
1017 &as->journal, interior_update_flush);
1018 bch2_journal_pin_drop(&c->journal, &w->journal);
1020 mutex_unlock(&c->btree_interior_update_lock);
1023 void bch2_btree_update_done(struct btree_update *as)
1025 BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE);
1027 bch2_btree_reserve_put(as->c, as->reserve);
1030 continue_at(&as->cl, btree_update_nodes_written, system_freezable_wq);
1033 struct btree_update *
1034 bch2_btree_update_start(struct bch_fs *c, enum btree_id id,
1035 unsigned nr_nodes, unsigned flags,
1038 struct btree_reserve *reserve;
1039 struct btree_update *as;
1041 if (unlikely(!percpu_ref_tryget(&c->writes)))
1042 return ERR_PTR(-EROFS);
1044 reserve = bch2_btree_reserve_get(c, nr_nodes, flags, cl);
1045 if (IS_ERR(reserve)) {
1046 percpu_ref_put(&c->writes);
1047 return ERR_CAST(reserve);
1050 as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOIO);
1051 memset(as, 0, sizeof(*as));
1052 closure_init(&as->cl, NULL);
1054 as->mode = BTREE_INTERIOR_NO_UPDATE;
1056 as->reserve = reserve;
1057 INIT_LIST_HEAD(&as->write_blocked_list);
1059 bch2_keylist_init(&as->parent_keys, as->inline_keys);
1061 mutex_lock(&c->btree_interior_update_lock);
1062 list_add_tail(&as->list, &c->btree_interior_update_list);
1063 mutex_unlock(&c->btree_interior_update_lock);
1068 /* Btree root updates: */
1070 static void __bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
1072 /* Root nodes cannot be reaped */
1073 mutex_lock(&c->btree_cache.lock);
1074 list_del_init(&b->list);
1075 mutex_unlock(&c->btree_cache.lock);
1077 mutex_lock(&c->btree_root_lock);
1078 BUG_ON(btree_node_root(c, b) &&
1079 (b->level < btree_node_root(c, b)->level ||
1080 !btree_node_dying(btree_node_root(c, b))));
1082 btree_node_root(c, b) = b;
1083 mutex_unlock(&c->btree_root_lock);
1085 bch2_recalc_btree_reserve(c);
1088 static void bch2_btree_set_root_inmem(struct btree_update *as, struct btree *b)
1090 struct bch_fs *c = as->c;
1091 struct btree *old = btree_node_root(c, b);
1092 struct bch_fs_usage stats = { 0 };
1094 __bch2_btree_set_root_inmem(c, b);
1096 bch2_mark_key(c, BKEY_TYPE_BTREE,
1097 bkey_i_to_s_c(&b->key),
1099 gc_pos_btree_root(b->btree_id),
1102 if (old && !btree_node_fake(old))
1103 bch2_btree_node_free_index(as, NULL,
1104 bkey_i_to_s_c(&old->key),
1106 bch2_fs_usage_apply(c, &stats, &as->reserve->disk_res,
1107 gc_pos_btree_root(b->btree_id));
1110 static void bch2_btree_set_root_ondisk(struct bch_fs *c, struct btree *b, int rw)
1112 struct btree_root *r = &c->btree_roots[b->btree_id];
1114 mutex_lock(&c->btree_root_lock);
1117 bkey_copy(&r->key, &b->key);
1118 r->level = b->level;
1121 c->btree_roots_dirty = true;
1123 mutex_unlock(&c->btree_root_lock);
1127 * bch_btree_set_root - update the root in memory and on disk
1129 * To ensure forward progress, the current task must not be holding any
1130 * btree node write locks. However, you must hold an intent lock on the
1133 * Note: This allocates a journal entry but doesn't add any keys to
1134 * it. All the btree roots are part of every journal write, so there
1135 * is nothing new to be done. This just guarantees that there is a
1138 static void bch2_btree_set_root(struct btree_update *as, struct btree *b,
1139 struct btree_iter *iter)
1141 struct bch_fs *c = as->c;
1144 trace_btree_set_root(c, b);
1145 BUG_ON(!b->written &&
1146 !test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags));
1148 old = btree_node_root(c, b);
1151 * Ensure no one is using the old root while we switch to the
1154 bch2_btree_node_lock_write(old, iter);
1156 bch2_btree_set_root_inmem(as, b);
1158 btree_update_updated_root(as);
1161 * Unlock old root after new root is visible:
1163 * The new root isn't persistent, but that's ok: we still have
1164 * an intent lock on the new root, and any updates that would
1165 * depend on the new root would have to update the new root.
1167 bch2_btree_node_unlock_write(old, iter);
1170 /* Interior node updates: */
1172 static void bch2_insert_fixup_btree_ptr(struct btree_update *as, struct btree *b,
1173 struct btree_iter *iter,
1174 struct bkey_i *insert,
1175 struct btree_node_iter *node_iter)
1177 struct bch_fs *c = as->c;
1178 struct bch_fs_usage stats = { 0 };
1179 struct bkey_packed *k;
1182 BUG_ON(insert->k.u64s > bch_btree_keys_u64s_remaining(c, b));
1184 if (bkey_extent_is_data(&insert->k))
1185 bch2_mark_key(c, BKEY_TYPE_BTREE,
1186 bkey_i_to_s_c(insert),
1188 gc_pos_btree_node(b), &stats, 0, 0);
1190 while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1191 bkey_iter_pos_cmp(b, &insert->k.p, k) > 0)
1192 bch2_btree_node_iter_advance(node_iter, b);
1195 * If we're overwriting, look up pending delete and mark so that gc
1196 * marks it on the pending delete list:
1198 if (k && !bkey_cmp_packed(b, k, &insert->k))
1199 bch2_btree_node_free_index(as, b,
1200 bkey_disassemble(b, k, &tmp),
1203 bch2_fs_usage_apply(c, &stats, &as->reserve->disk_res,
1204 gc_pos_btree_node(b));
1206 bch2_btree_bset_insert_key(iter, b, node_iter, insert);
1207 set_btree_node_dirty(b);
1208 set_btree_node_need_write(b);
1212 * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1215 static struct btree *__btree_split_node(struct btree_update *as,
1217 struct btree_iter *iter)
1219 size_t nr_packed = 0, nr_unpacked = 0;
1221 struct bset *set1, *set2;
1222 struct bkey_packed *k, *prev = NULL;
1224 n2 = bch2_btree_node_alloc(as, n1->level);
1226 n2->data->max_key = n1->data->max_key;
1227 n2->data->format = n1->format;
1228 SET_BTREE_NODE_SEQ(n2->data, BTREE_NODE_SEQ(n1->data));
1229 n2->key.k.p = n1->key.k.p;
1231 btree_node_set_format(n2, n2->data->format);
1233 set1 = btree_bset_first(n1);
1234 set2 = btree_bset_first(n2);
1237 * Has to be a linear search because we don't have an auxiliary
1242 if (bkey_next(k) == vstruct_last(set1))
1244 if (k->_data - set1->_data >= (le16_to_cpu(set1->u64s) * 3) / 5)
1258 n1->key.k.p = bkey_unpack_pos(n1, prev);
1259 n1->data->max_key = n1->key.k.p;
1261 btree_type_successor(n1->btree_id, n1->key.k.p);
1263 set2->u64s = cpu_to_le16((u64 *) vstruct_end(set1) - (u64 *) k);
1264 set1->u64s = cpu_to_le16(le16_to_cpu(set1->u64s) - le16_to_cpu(set2->u64s));
1266 set_btree_bset_end(n1, n1->set);
1267 set_btree_bset_end(n2, n2->set);
1269 n2->nr.live_u64s = le16_to_cpu(set2->u64s);
1270 n2->nr.bset_u64s[0] = le16_to_cpu(set2->u64s);
1271 n2->nr.packed_keys = n1->nr.packed_keys - nr_packed;
1272 n2->nr.unpacked_keys = n1->nr.unpacked_keys - nr_unpacked;
1274 n1->nr.live_u64s = le16_to_cpu(set1->u64s);
1275 n1->nr.bset_u64s[0] = le16_to_cpu(set1->u64s);
1276 n1->nr.packed_keys = nr_packed;
1277 n1->nr.unpacked_keys = nr_unpacked;
1279 BUG_ON(!set1->u64s);
1280 BUG_ON(!set2->u64s);
1282 memcpy_u64s(set2->start,
1284 le16_to_cpu(set2->u64s));
1286 btree_node_reset_sib_u64s(n1);
1287 btree_node_reset_sib_u64s(n2);
1289 bch2_verify_btree_nr_keys(n1);
1290 bch2_verify_btree_nr_keys(n2);
1293 btree_node_interior_verify(n1);
1294 btree_node_interior_verify(n2);
1301 * For updates to interior nodes, we've got to do the insert before we split
1302 * because the stuff we're inserting has to be inserted atomically. Post split,
1303 * the keys might have to go in different nodes and the split would no longer be
1306 * Worse, if the insert is from btree node coalescing, if we do the insert after
1307 * we do the split (and pick the pivot) - the pivot we pick might be between
1308 * nodes that were coalesced, and thus in the middle of a child node post
1311 static void btree_split_insert_keys(struct btree_update *as, struct btree *b,
1312 struct btree_iter *iter,
1313 struct keylist *keys)
1315 struct btree_node_iter node_iter;
1316 struct bkey_i *k = bch2_keylist_front(keys);
1317 struct bkey_packed *p;
1320 BUG_ON(btree_node_type(b) != BKEY_TYPE_BTREE);
1322 bch2_btree_node_iter_init(&node_iter, b, &k->k.p);
1324 while (!bch2_keylist_empty(keys)) {
1325 k = bch2_keylist_front(keys);
1327 BUG_ON(bch_keylist_u64s(keys) >
1328 bch_btree_keys_u64s_remaining(as->c, b));
1329 BUG_ON(bkey_cmp(k->k.p, b->data->min_key) < 0);
1330 BUG_ON(bkey_cmp(k->k.p, b->data->max_key) > 0);
1332 bch2_insert_fixup_btree_ptr(as, b, iter, k, &node_iter);
1333 bch2_keylist_pop_front(keys);
1337 * We can't tolerate whiteouts here - with whiteouts there can be
1338 * duplicate keys, and it would be rather bad if we picked a duplicate
1341 i = btree_bset_first(b);
1343 while (p != vstruct_last(i))
1344 if (bkey_deleted(p)) {
1345 le16_add_cpu(&i->u64s, -p->u64s);
1346 set_btree_bset_end(b, b->set);
1347 memmove_u64s_down(p, bkey_next(p),
1348 (u64 *) vstruct_last(i) -
1353 BUG_ON(b->nsets != 1 ||
1354 b->nr.live_u64s != le16_to_cpu(btree_bset_first(b)->u64s));
1356 btree_node_interior_verify(b);
1359 static void btree_split(struct btree_update *as, struct btree *b,
1360 struct btree_iter *iter, struct keylist *keys,
1363 struct bch_fs *c = as->c;
1364 struct btree *parent = btree_node_parent(iter, b);
1365 struct btree *n1, *n2 = NULL, *n3 = NULL;
1366 u64 start_time = local_clock();
1368 BUG_ON(!parent && (b != btree_node_root(c, b)));
1369 BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->level));
1371 bch2_btree_interior_update_will_free_node(as, b);
1373 n1 = bch2_btree_node_alloc_replacement(as, b);
1376 btree_split_insert_keys(as, n1, iter, keys);
1378 if (vstruct_blocks(n1->data, c->block_bits) > BTREE_SPLIT_THRESHOLD(c)) {
1379 trace_btree_split(c, b);
1381 n2 = __btree_split_node(as, n1, iter);
1383 bch2_btree_build_aux_trees(n2);
1384 bch2_btree_build_aux_trees(n1);
1385 six_unlock_write(&n2->lock);
1386 six_unlock_write(&n1->lock);
1388 bch2_btree_node_write(c, n2, SIX_LOCK_intent);
1391 * Note that on recursive parent_keys == keys, so we
1392 * can't start adding new keys to parent_keys before emptying it
1393 * out (which we did with btree_split_insert_keys() above)
1395 bch2_keylist_add(&as->parent_keys, &n1->key);
1396 bch2_keylist_add(&as->parent_keys, &n2->key);
1399 /* Depth increases, make a new root */
1400 n3 = __btree_root_alloc(as, b->level + 1);
1402 n3->sib_u64s[0] = U16_MAX;
1403 n3->sib_u64s[1] = U16_MAX;
1405 btree_split_insert_keys(as, n3, iter, &as->parent_keys);
1407 bch2_btree_node_write(c, n3, SIX_LOCK_intent);
1410 trace_btree_compact(c, b);
1412 bch2_btree_build_aux_trees(n1);
1413 six_unlock_write(&n1->lock);
1415 bch2_keylist_add(&as->parent_keys, &n1->key);
1418 bch2_btree_node_write(c, n1, SIX_LOCK_intent);
1420 /* New nodes all written, now make them visible: */
1423 /* Split a non root node */
1424 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1426 bch2_btree_set_root(as, n3, iter);
1428 /* Root filled up but didn't need to be split */
1429 bch2_btree_set_root(as, n1, iter);
1432 bch2_open_buckets_put(c, &n1->ob);
1434 bch2_open_buckets_put(c, &n2->ob);
1436 bch2_open_buckets_put(c, &n3->ob);
1439 * Note - at this point other linked iterators could still have @b read
1440 * locked; we're depending on the bch2_btree_iter_node_replace() calls
1441 * below removing all references to @b so we don't return with other
1442 * iterators pointing to a node they have locked that's been freed.
1444 * We have to free the node first because the bch2_iter_node_replace()
1445 * calls will drop _our_ iterator's reference - and intent lock - to @b.
1447 bch2_btree_node_free_inmem(c, b, iter);
1449 /* Successful split, update the iterator to point to the new nodes: */
1452 bch2_btree_iter_node_replace(iter, n3);
1454 bch2_btree_iter_node_replace(iter, n2);
1455 bch2_btree_iter_node_replace(iter, n1);
1457 bch2_time_stats_update(&c->times[BCH_TIME_btree_split], start_time);
1461 bch2_btree_insert_keys_interior(struct btree_update *as, struct btree *b,
1462 struct btree_iter *iter, struct keylist *keys)
1464 struct btree_iter *linked;
1465 struct btree_node_iter node_iter;
1466 struct bkey_i *insert = bch2_keylist_front(keys);
1467 struct bkey_packed *k;
1469 /* Don't screw up @iter's position: */
1470 node_iter = iter->l[b->level].iter;
1473 * btree_split(), btree_gc_coalesce() will insert keys before
1474 * the iterator's current position - they know the keys go in
1475 * the node the iterator points to:
1477 while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1478 (bkey_cmp_packed(b, k, &insert->k) >= 0))
1481 while (!bch2_keylist_empty(keys)) {
1482 insert = bch2_keylist_front(keys);
1484 bch2_insert_fixup_btree_ptr(as, b, iter, insert, &node_iter);
1485 bch2_keylist_pop_front(keys);
1488 btree_update_updated_node(as, b);
1490 for_each_btree_iter_with_node(iter, b, linked)
1491 bch2_btree_node_iter_peek(&linked->l[b->level].iter, b);
1493 bch2_btree_iter_verify(iter, b);
1497 * bch_btree_insert_node - insert bkeys into a given btree node
1499 * @iter: btree iterator
1500 * @keys: list of keys to insert
1501 * @hook: insert callback
1502 * @persistent: if not null, @persistent will wait on journal write
1504 * Inserts as many keys as it can into a given btree node, splitting it if full.
1505 * If a split occurred, this function will return early. This can only happen
1506 * for leaf nodes -- inserts into interior nodes have to be atomic.
1508 void bch2_btree_insert_node(struct btree_update *as, struct btree *b,
1509 struct btree_iter *iter, struct keylist *keys,
1512 struct bch_fs *c = as->c;
1513 int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1514 int old_live_u64s = b->nr.live_u64s;
1515 int live_u64s_added, u64s_added;
1517 BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->level));
1519 BUG_ON(!as || as->b);
1520 bch2_verify_keylist_sorted(keys);
1522 if (as->must_rewrite)
1525 bch2_btree_node_lock_for_insert(c, b, iter);
1527 if (!bch2_btree_node_insert_fits(c, b, bch_keylist_u64s(keys))) {
1528 bch2_btree_node_unlock_write(b, iter);
1532 bch2_btree_insert_keys_interior(as, b, iter, keys);
1534 live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1535 u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1537 if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1538 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1539 if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1540 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1542 if (u64s_added > live_u64s_added &&
1543 bch2_maybe_compact_whiteouts(c, b))
1544 bch2_btree_iter_reinit_node(iter, b);
1546 bch2_btree_node_unlock_write(b, iter);
1548 btree_node_interior_verify(b);
1551 * when called from the btree_split path the new nodes aren't added to
1552 * the btree iterator yet, so the merge path's unlock/wait/relock dance
1555 bch2_foreground_maybe_merge(c, iter, b->level,
1556 flags|BTREE_INSERT_NOUNLOCK);
1559 btree_split(as, b, iter, keys, flags);
1562 int bch2_btree_split_leaf(struct bch_fs *c, struct btree_iter *iter,
1565 struct btree *b = iter->l[0].b;
1566 struct btree_update *as;
1569 struct btree_iter *linked;
1572 * We already have a disk reservation and open buckets pinned; this
1573 * allocation must not block:
1575 for_each_btree_iter(iter, linked)
1576 if (linked->btree_id == BTREE_ID_EXTENTS)
1577 flags |= BTREE_INSERT_USE_RESERVE;
1579 closure_init_stack(&cl);
1581 /* Hack, because gc and splitting nodes doesn't mix yet: */
1582 if (!down_read_trylock(&c->gc_lock)) {
1583 if (flags & BTREE_INSERT_NOUNLOCK)
1586 bch2_btree_iter_unlock(iter);
1587 down_read(&c->gc_lock);
1589 if (btree_iter_linked(iter))
1594 * XXX: figure out how far we might need to split,
1595 * instead of locking/reserving all the way to the root:
1597 if (!bch2_btree_iter_upgrade(iter, U8_MAX,
1598 !(flags & BTREE_INSERT_NOUNLOCK))) {
1603 as = bch2_btree_update_start(c, iter->btree_id,
1604 btree_update_reserve_required(c, b), flags,
1605 !(flags & BTREE_INSERT_NOUNLOCK) ? &cl : NULL);
1608 if (ret == -EAGAIN) {
1609 BUG_ON(flags & BTREE_INSERT_NOUNLOCK);
1610 bch2_btree_iter_unlock(iter);
1616 btree_split(as, b, iter, NULL, flags);
1617 bch2_btree_update_done(as);
1620 * We haven't successfully inserted yet, so don't downgrade all the way
1621 * back to read locks;
1623 __bch2_btree_iter_downgrade(iter, 1);
1625 up_read(&c->gc_lock);
1630 void __bch2_foreground_maybe_merge(struct bch_fs *c,
1631 struct btree_iter *iter,
1634 enum btree_node_sibling sib)
1636 struct btree_update *as;
1637 struct bkey_format_state new_s;
1638 struct bkey_format new_f;
1639 struct bkey_i delete;
1640 struct btree *b, *m, *n, *prev, *next, *parent;
1645 closure_init_stack(&cl);
1647 BUG_ON(!btree_node_locked(iter, level));
1649 b = iter->l[level].b;
1651 parent = btree_node_parent(iter, b);
1655 if (b->sib_u64s[sib] > BTREE_FOREGROUND_MERGE_THRESHOLD(c))
1658 /* XXX: can't be holding read locks */
1659 m = bch2_btree_node_get_sibling(c, iter, b,
1660 !(flags & BTREE_INSERT_NOUNLOCK), sib);
1666 /* NULL means no sibling: */
1668 b->sib_u64s[sib] = U16_MAX;
1672 if (sib == btree_prev_sib) {
1680 bch2_bkey_format_init(&new_s);
1681 __bch2_btree_calc_format(&new_s, b);
1682 __bch2_btree_calc_format(&new_s, m);
1683 new_f = bch2_bkey_format_done(&new_s);
1685 sib_u64s = btree_node_u64s_with_format(b, &new_f) +
1686 btree_node_u64s_with_format(m, &new_f);
1688 if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1689 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1691 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1694 sib_u64s = min(sib_u64s, btree_max_u64s(c));
1695 b->sib_u64s[sib] = sib_u64s;
1697 if (b->sib_u64s[sib] > BTREE_FOREGROUND_MERGE_THRESHOLD(c)) {
1698 six_unlock_intent(&m->lock);
1702 /* We're changing btree topology, doesn't mix with gc: */
1703 if (!down_read_trylock(&c->gc_lock))
1704 goto err_cycle_gc_lock;
1706 if (!bch2_btree_iter_upgrade(iter, U8_MAX,
1707 !(flags & BTREE_INSERT_NOUNLOCK))) {
1712 as = bch2_btree_update_start(c, iter->btree_id,
1713 btree_update_reserve_required(c, parent) + 1,
1714 BTREE_INSERT_NOFAIL|
1715 BTREE_INSERT_USE_RESERVE,
1716 !(flags & BTREE_INSERT_NOUNLOCK) ? &cl : NULL);
1722 trace_btree_merge(c, b);
1724 bch2_btree_interior_update_will_free_node(as, b);
1725 bch2_btree_interior_update_will_free_node(as, m);
1727 n = bch2_btree_node_alloc(as, b->level);
1729 n->data->min_key = prev->data->min_key;
1730 n->data->max_key = next->data->max_key;
1731 n->data->format = new_f;
1732 n->key.k.p = next->key.k.p;
1734 btree_node_set_format(n, new_f);
1736 bch2_btree_sort_into(c, n, prev);
1737 bch2_btree_sort_into(c, n, next);
1739 bch2_btree_build_aux_trees(n);
1740 six_unlock_write(&n->lock);
1742 bkey_init(&delete.k);
1743 delete.k.p = prev->key.k.p;
1744 bch2_keylist_add(&as->parent_keys, &delete);
1745 bch2_keylist_add(&as->parent_keys, &n->key);
1747 bch2_btree_node_write(c, n, SIX_LOCK_intent);
1749 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1751 bch2_open_buckets_put(c, &n->ob);
1752 bch2_btree_node_free_inmem(c, b, iter);
1753 bch2_btree_node_free_inmem(c, m, iter);
1754 bch2_btree_iter_node_replace(iter, n);
1756 bch2_btree_iter_verify(iter, n);
1758 bch2_btree_update_done(as);
1760 six_unlock_intent(&m->lock);
1761 up_read(&c->gc_lock);
1764 * Don't downgrade locks here: we're called after successful insert,
1765 * and the caller will downgrade locks after a successful insert
1766 * anyways (in case e.g. a split was required first)
1768 * And we're also called when inserting into interior nodes in the
1769 * split path, and downgrading to read locks in there is potentially
1776 six_unlock_intent(&m->lock);
1778 if (flags & BTREE_INSERT_NOUNLOCK)
1781 bch2_btree_iter_unlock(iter);
1783 down_read(&c->gc_lock);
1784 up_read(&c->gc_lock);
1789 six_unlock_intent(&m->lock);
1790 up_read(&c->gc_lock);
1792 BUG_ON(ret == -EAGAIN && (flags & BTREE_INSERT_NOUNLOCK));
1794 if ((ret == -EAGAIN || ret == -EINTR) &&
1795 !(flags & BTREE_INSERT_NOUNLOCK)) {
1796 bch2_btree_iter_unlock(iter);
1798 ret = bch2_btree_iter_traverse(iter);
1808 static int __btree_node_rewrite(struct bch_fs *c, struct btree_iter *iter,
1809 struct btree *b, unsigned flags,
1812 struct btree *n, *parent = btree_node_parent(iter, b);
1813 struct btree_update *as;
1815 as = bch2_btree_update_start(c, iter->btree_id,
1817 ? btree_update_reserve_required(c, parent)
1821 trace_btree_gc_rewrite_node_fail(c, b);
1825 bch2_btree_interior_update_will_free_node(as, b);
1827 n = bch2_btree_node_alloc_replacement(as, b);
1829 bch2_btree_build_aux_trees(n);
1830 six_unlock_write(&n->lock);
1832 trace_btree_gc_rewrite_node(c, b);
1834 bch2_btree_node_write(c, n, SIX_LOCK_intent);
1837 bch2_keylist_add(&as->parent_keys, &n->key);
1838 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1840 bch2_btree_set_root(as, n, iter);
1843 bch2_open_buckets_put(c, &n->ob);
1845 bch2_btree_node_free_inmem(c, b, iter);
1847 bch2_btree_iter_node_replace(iter, n);
1849 bch2_btree_update_done(as);
1854 * bch_btree_node_rewrite - Rewrite/move a btree node
1856 * Returns 0 on success, -EINTR or -EAGAIN on failure (i.e.
1857 * btree_check_reserve() has to wait)
1859 int bch2_btree_node_rewrite(struct bch_fs *c, struct btree_iter *iter,
1860 __le64 seq, unsigned flags)
1866 flags |= BTREE_INSERT_NOFAIL;
1868 closure_init_stack(&cl);
1870 bch2_btree_iter_upgrade(iter, U8_MAX, true);
1872 if (!(flags & BTREE_INSERT_GC_LOCK_HELD)) {
1873 if (!down_read_trylock(&c->gc_lock)) {
1874 bch2_btree_iter_unlock(iter);
1875 down_read(&c->gc_lock);
1880 ret = bch2_btree_iter_traverse(iter);
1884 b = bch2_btree_iter_peek_node(iter);
1885 if (!b || b->data->keys.seq != seq)
1888 ret = __btree_node_rewrite(c, iter, b, flags, &cl);
1889 if (ret != -EAGAIN &&
1893 bch2_btree_iter_unlock(iter);
1897 bch2_btree_iter_downgrade(iter);
1899 if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1900 up_read(&c->gc_lock);
1906 static void __bch2_btree_node_update_key(struct bch_fs *c,
1907 struct btree_update *as,
1908 struct btree_iter *iter,
1909 struct btree *b, struct btree *new_hash,
1910 struct bkey_i_extent *new_key)
1912 struct btree *parent;
1916 * Two corner cases that need to be thought about here:
1918 * @b may not be reachable yet - there might be another interior update
1919 * operation waiting on @b to be written, and we're gonna deliver the
1920 * write completion to that interior update operation _before_
1921 * persisting the new_key update
1923 * That ends up working without us having to do anything special here:
1924 * the reason is, we do kick off (and do the in memory updates) for the
1925 * update for @new_key before we return, creating a new interior_update
1928 * The new interior update operation here will in effect override the
1929 * previous one. The previous one was going to terminate - make @b
1930 * reachable - in one of two ways:
1931 * - updating the btree root pointer
1933 * no, this doesn't work. argh.
1936 if (b->will_make_reachable)
1937 as->must_rewrite = true;
1939 btree_interior_update_add_node_reference(as, b);
1941 parent = btree_node_parent(iter, b);
1944 bkey_copy(&new_hash->key, &new_key->k_i);
1945 ret = bch2_btree_node_hash_insert(&c->btree_cache,
1946 new_hash, b->level, b->btree_id);
1950 bch2_keylist_add(&as->parent_keys, &new_key->k_i);
1951 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, 0);
1954 mutex_lock(&c->btree_cache.lock);
1955 bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
1957 bch2_btree_node_hash_remove(&c->btree_cache, b);
1959 bkey_copy(&b->key, &new_key->k_i);
1960 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
1962 mutex_unlock(&c->btree_cache.lock);
1964 bkey_copy(&b->key, &new_key->k_i);
1967 struct bch_fs_usage stats = { 0 };
1969 BUG_ON(btree_node_root(c, b) != b);
1971 bch2_btree_node_lock_write(b, iter);
1973 bch2_mark_key(c, BKEY_TYPE_BTREE,
1974 bkey_i_to_s_c(&new_key->k_i),
1976 gc_pos_btree_root(b->btree_id),
1978 bch2_btree_node_free_index(as, NULL,
1979 bkey_i_to_s_c(&b->key),
1981 bch2_fs_usage_apply(c, &stats, &as->reserve->disk_res,
1982 gc_pos_btree_root(b->btree_id));
1984 if (PTR_HASH(&new_key->k_i) != PTR_HASH(&b->key)) {
1985 mutex_lock(&c->btree_cache.lock);
1986 bch2_btree_node_hash_remove(&c->btree_cache, b);
1988 bkey_copy(&b->key, &new_key->k_i);
1989 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
1991 mutex_unlock(&c->btree_cache.lock);
1993 bkey_copy(&b->key, &new_key->k_i);
1996 btree_update_updated_root(as);
1997 bch2_btree_node_unlock_write(b, iter);
2000 bch2_btree_update_done(as);
2003 int bch2_btree_node_update_key(struct bch_fs *c, struct btree_iter *iter,
2004 struct btree *b, struct bkey_i_extent *new_key)
2006 struct btree *parent = btree_node_parent(iter, b);
2007 struct btree_update *as = NULL;
2008 struct btree *new_hash = NULL;
2012 closure_init_stack(&cl);
2014 if (!bch2_btree_iter_upgrade(iter, U8_MAX, true))
2017 if (!down_read_trylock(&c->gc_lock)) {
2018 bch2_btree_iter_unlock(iter);
2019 down_read(&c->gc_lock);
2021 if (!bch2_btree_iter_relock(iter)) {
2027 /* check PTR_HASH() after @b is locked by btree_iter_traverse(): */
2028 if (PTR_HASH(&new_key->k_i) != PTR_HASH(&b->key)) {
2029 /* bch2_btree_reserve_get will unlock */
2030 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2034 bch2_btree_iter_unlock(iter);
2035 up_read(&c->gc_lock);
2037 down_read(&c->gc_lock);
2039 if (!bch2_btree_iter_relock(iter))
2043 new_hash = bch2_btree_node_mem_alloc(c);
2046 as = bch2_btree_update_start(c, iter->btree_id,
2047 parent ? btree_update_reserve_required(c, parent) : 0,
2048 BTREE_INSERT_NOFAIL|
2049 BTREE_INSERT_USE_RESERVE|
2050 BTREE_INSERT_USE_ALLOC_RESERVE,
2061 bch2_btree_iter_unlock(iter);
2062 up_read(&c->gc_lock);
2064 down_read(&c->gc_lock);
2066 if (!bch2_btree_iter_relock(iter))
2070 ret = bch2_mark_bkey_replicas(c, BKEY_TYPE_BTREE,
2071 extent_i_to_s_c(new_key).s_c);
2073 goto err_free_update;
2075 __bch2_btree_node_update_key(c, as, iter, b, new_hash, new_key);
2077 bch2_btree_iter_downgrade(iter);
2080 mutex_lock(&c->btree_cache.lock);
2081 list_move(&new_hash->list, &c->btree_cache.freeable);
2082 mutex_unlock(&c->btree_cache.lock);
2084 six_unlock_write(&new_hash->lock);
2085 six_unlock_intent(&new_hash->lock);
2087 up_read(&c->gc_lock);
2091 bch2_btree_update_free(as);
2098 * Only for filesystem bringup, when first reading the btree roots or allocating
2099 * btree roots when initializing a new filesystem:
2101 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
2103 BUG_ON(btree_node_root(c, b));
2105 __bch2_btree_set_root_inmem(c, b);
2106 bch2_btree_set_root_ondisk(c, b, READ);
2109 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
2115 closure_init_stack(&cl);
2118 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2122 b = bch2_btree_node_mem_alloc(c);
2123 bch2_btree_cache_cannibalize_unlock(c);
2125 set_btree_node_fake(b);
2129 bkey_extent_init(&b->key);
2130 b->key.k.p = POS_MAX;
2131 bkey_i_to_extent(&b->key)->v._data[0] = U64_MAX - id;
2133 bch2_bset_init_first(b, &b->data->keys);
2134 bch2_btree_build_aux_trees(b);
2136 b->data->min_key = POS_MIN;
2137 b->data->max_key = POS_MAX;
2138 b->data->format = bch2_btree_calc_format(b);
2139 btree_node_set_format(b, b->data->format);
2141 ret = bch2_btree_node_hash_insert(&c->btree_cache, b, b->level, b->btree_id);
2144 __bch2_btree_set_root_inmem(c, b);
2146 six_unlock_write(&b->lock);
2147 six_unlock_intent(&b->lock);
2150 ssize_t bch2_btree_updates_print(struct bch_fs *c, char *buf)
2152 char *out = buf, *end = buf + PAGE_SIZE;
2153 struct btree_update *as;
2155 mutex_lock(&c->btree_interior_update_lock);
2156 list_for_each_entry(as, &c->btree_interior_update_list, list)
2157 out += scnprintf(out, end - out, "%p m %u w %u r %u j %llu\n",
2161 atomic_read(&as->cl.remaining) & CLOSURE_REMAINING_MASK,
2163 mutex_unlock(&c->btree_interior_update_lock);
2168 size_t bch2_btree_interior_updates_nr_pending(struct bch_fs *c)
2171 struct list_head *i;
2173 mutex_lock(&c->btree_interior_update_lock);
2174 list_for_each(i, &c->btree_interior_update_list)
2176 mutex_unlock(&c->btree_interior_update_lock);