2 * Copyright (C) 2010 Kent Overstreet <kent.overstreet@gmail.com>
3 * Copyright (C) 2014 Datera Inc.
8 #include "bkey_methods.h"
9 #include "btree_locking.h"
10 #include "btree_update_interior.h"
23 #include <linux/slab.h>
24 #include <linux/bitops.h>
25 #include <linux/freezer.h>
26 #include <linux/kthread.h>
27 #include <linux/preempt.h>
28 #include <linux/rcupdate.h>
29 #include <trace/events/bcachefs.h>
39 static void btree_node_range_checks_init(struct range_checks *r, unsigned depth)
43 for (i = 0; i < BTREE_MAX_DEPTH; i++)
44 r->l[i].min = r->l[i].max = POS_MIN;
48 static void btree_node_range_checks(struct bch_fs *c, struct btree *b,
49 struct range_checks *r)
51 struct range_level *l = &r->l[b->level];
53 struct bpos expected_min = bkey_cmp(l->min, l->max)
54 ? btree_type_successor(b->btree_id, l->max)
57 bch2_fs_inconsistent_on(bkey_cmp(b->data->min_key, expected_min), c,
58 "btree node has incorrect min key: %llu:%llu != %llu:%llu",
59 b->data->min_key.inode,
60 b->data->min_key.offset,
64 l->max = b->data->max_key;
66 if (b->level > r->depth) {
67 l = &r->l[b->level - 1];
69 bch2_fs_inconsistent_on(bkey_cmp(b->data->min_key, l->min), c,
70 "btree node min doesn't match min of child nodes: %llu:%llu != %llu:%llu",
71 b->data->min_key.inode,
72 b->data->min_key.offset,
76 bch2_fs_inconsistent_on(bkey_cmp(b->data->max_key, l->max), c,
77 "btree node max doesn't match max of child nodes: %llu:%llu != %llu:%llu",
78 b->data->max_key.inode,
79 b->data->max_key.offset,
83 if (bkey_cmp(b->data->max_key, POS_MAX))
85 btree_type_successor(b->btree_id,
90 u8 bch2_btree_key_recalc_oldest_gen(struct bch_fs *c, struct bkey_s_c k)
92 const struct bch_extent_ptr *ptr;
95 if (bkey_extent_is_data(k.k)) {
96 struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
98 extent_for_each_ptr(e, ptr) {
99 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
100 size_t b = PTR_BUCKET_NR(ca, ptr);
102 if (gen_after(ca->oldest_gens[b], ptr->gen))
103 ca->oldest_gens[b] = ptr->gen;
105 max_stale = max(max_stale, ptr_stale(ca, ptr));
113 * For runtime mark and sweep:
115 static u8 bch2_gc_mark_key(struct bch_fs *c, enum bkey_type type,
116 struct bkey_s_c k, unsigned flags)
118 struct gc_pos pos = { 0 };
119 struct bch_fs_usage *stats;
123 stats = this_cpu_ptr(c->usage_percpu);
125 case BKEY_TYPE_BTREE:
126 bch2_mark_key(c, k, c->opts.btree_node_size, true, pos, stats,
128 BCH_BUCKET_MARK_MAY_MAKE_UNAVAILABLE|
129 BCH_BUCKET_MARK_GC_LOCK_HELD);
131 case BKEY_TYPE_EXTENTS:
132 bch2_mark_key(c, k, k.k->size, false, pos, stats,
134 BCH_BUCKET_MARK_MAY_MAKE_UNAVAILABLE|
135 BCH_BUCKET_MARK_GC_LOCK_HELD);
136 ret = bch2_btree_key_recalc_oldest_gen(c, k);
146 int bch2_btree_mark_key_initial(struct bch_fs *c, enum bkey_type type,
149 enum bch_data_type data_type = type == BKEY_TYPE_BTREE
150 ? BCH_DATA_BTREE : BCH_DATA_USER;
155 case BCH_EXTENT_CACHED: {
156 struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
157 const struct bch_extent_ptr *ptr;
159 if (test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) ||
161 fsck_err_on(!bch2_sb_has_replicas(c, e, data_type), c,
162 "superblock not marked as containing replicas (type %u)",
164 ret = bch2_check_mark_super(c, e, data_type);
169 extent_for_each_ptr(e, ptr) {
170 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
171 struct bucket *g = PTR_BUCKET(ca, ptr);
173 if (mustfix_fsck_err_on(!g->mark.gen_valid, c,
174 "found ptr with missing gen in alloc btree,\n"
176 bch2_data_types[data_type],
178 g->_mark.gen = ptr->gen;
179 g->_mark.gen_valid = 1;
180 set_bit(g - ca->buckets, ca->bucket_dirty);
183 if (mustfix_fsck_err_on(gen_cmp(ptr->gen, g->mark.gen) > 0, c,
184 "%s ptr gen in the future: %u > %u",
185 bch2_data_types[data_type],
186 ptr->gen, g->mark.gen)) {
187 g->_mark.gen = ptr->gen;
188 g->_mark.gen_valid = 1;
189 set_bit(g - ca->buckets, ca->bucket_dirty);
190 set_bit(BCH_FS_FIXED_GENS, &c->flags);
199 atomic64_set(&c->key_version,
200 max_t(u64, k.k->version.lo,
201 atomic64_read(&c->key_version)));
203 bch2_gc_mark_key(c, type, k, BCH_BUCKET_MARK_NOATOMIC);
208 static unsigned btree_gc_mark_node(struct bch_fs *c, struct btree *b)
210 enum bkey_type type = btree_node_type(b);
211 struct btree_node_iter iter;
212 struct bkey unpacked;
216 if (btree_node_has_ptrs(b))
217 for_each_btree_node_key_unpack(b, k, &iter,
218 btree_node_is_extents(b),
220 bch2_bkey_debugcheck(c, b, k);
221 stale = max(stale, bch2_gc_mark_key(c, type, k, 0));
227 static inline void __gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
229 write_seqcount_begin(&c->gc_pos_lock);
231 write_seqcount_end(&c->gc_pos_lock);
234 static inline void gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
236 BUG_ON(gc_pos_cmp(new_pos, c->gc_pos) <= 0);
237 __gc_pos_set(c, new_pos);
240 static int bch2_gc_btree(struct bch_fs *c, enum btree_id btree_id)
242 struct btree_iter iter;
244 struct range_checks r;
245 unsigned depth = btree_id == BTREE_ID_EXTENTS ? 0 : 1;
250 * if expensive_debug_checks is on, run range_checks on all leaf nodes:
252 if (expensive_debug_checks(c))
255 btree_node_range_checks_init(&r, depth);
257 __for_each_btree_node(&iter, c, btree_id, POS_MIN,
258 0, depth, BTREE_ITER_PREFETCH, b) {
259 btree_node_range_checks(c, b, &r);
261 bch2_verify_btree_nr_keys(b);
263 max_stale = btree_gc_mark_node(c, b);
265 gc_pos_set(c, gc_pos_btree_node(b));
268 bch2_btree_node_rewrite(c, &iter,
270 BTREE_INSERT_USE_RESERVE|
271 BTREE_INSERT_GC_LOCK_HELD);
272 else if (!btree_gc_rewrite_disabled(c) &&
273 (btree_gc_always_rewrite(c) || max_stale > 16))
274 bch2_btree_node_rewrite(c, &iter,
277 BTREE_INSERT_GC_LOCK_HELD);
279 bch2_btree_iter_cond_resched(&iter);
281 ret = bch2_btree_iter_unlock(&iter);
285 mutex_lock(&c->btree_root_lock);
287 b = c->btree_roots[btree_id].b;
288 bch2_gc_mark_key(c, BKEY_TYPE_BTREE, bkey_i_to_s_c(&b->key), 0);
289 gc_pos_set(c, gc_pos_btree_root(b->btree_id));
291 mutex_unlock(&c->btree_root_lock);
295 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
297 enum bucket_data_type type,
300 u64 b = sector_to_bucket(ca, start);
303 bch2_mark_metadata_bucket(c, ca, ca->buckets + b, type,
304 gc_phase(GC_PHASE_SB), flags);
306 } while (b < sector_to_bucket(ca, end));
309 void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
312 struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
316 lockdep_assert_held(&c->sb_lock);
318 for (i = 0; i < layout->nr_superblocks; i++) {
319 u64 offset = le64_to_cpu(layout->sb_offset[i]);
321 if (offset == BCH_SB_SECTOR)
322 mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
325 mark_metadata_sectors(c, ca, offset,
326 offset + (1 << layout->sb_max_size_bits),
330 spin_lock(&c->journal.lock);
332 for (i = 0; i < ca->journal.nr; i++) {
333 b = ca->journal.buckets[i];
334 bch2_mark_metadata_bucket(c, ca, ca->buckets + b,
336 gc_phase(GC_PHASE_SB), flags);
339 spin_unlock(&c->journal.lock);
342 static void bch2_mark_superblocks(struct bch_fs *c)
347 mutex_lock(&c->sb_lock);
348 gc_pos_set(c, gc_phase(GC_PHASE_SB));
350 for_each_online_member(ca, c, i)
351 bch2_mark_dev_superblock(c, ca,
352 BCH_BUCKET_MARK_MAY_MAKE_UNAVAILABLE|
353 BCH_BUCKET_MARK_GC_LOCK_HELD);
354 mutex_unlock(&c->sb_lock);
357 /* Also see bch2_pending_btree_node_free_insert_done() */
358 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
360 struct gc_pos pos = { 0 };
361 struct bch_fs_usage stats = { 0 };
362 struct btree_update *as;
363 struct pending_btree_node_free *d;
365 mutex_lock(&c->btree_interior_update_lock);
366 gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
368 for_each_pending_btree_node_free(c, as, d)
369 if (d->index_update_done)
370 bch2_mark_key(c, bkey_i_to_s_c(&d->key),
371 c->opts.btree_node_size, true, pos,
373 BCH_BUCKET_MARK_MAY_MAKE_UNAVAILABLE|
374 BCH_BUCKET_MARK_GC_LOCK_HELD);
376 * Don't apply stats - pending deletes aren't tracked in
380 mutex_unlock(&c->btree_interior_update_lock);
383 static void bch2_mark_allocator_buckets(struct bch_fs *c)
386 struct open_bucket *ob;
390 spin_lock(&c->freelist_lock);
391 gc_pos_set(c, gc_pos_alloc(c, NULL));
393 for_each_member_device(ca, c, ci) {
394 fifo_for_each_entry(i, &ca->free_inc, iter)
395 bch2_mark_alloc_bucket(c, ca, &ca->buckets[i], true,
396 gc_pos_alloc(c, NULL),
397 BCH_BUCKET_MARK_MAY_MAKE_UNAVAILABLE|
398 BCH_BUCKET_MARK_GC_LOCK_HELD);
402 for (j = 0; j < RESERVE_NR; j++)
403 fifo_for_each_entry(i, &ca->free[j], iter)
404 bch2_mark_alloc_bucket(c, ca, &ca->buckets[i], true,
405 gc_pos_alloc(c, NULL),
406 BCH_BUCKET_MARK_MAY_MAKE_UNAVAILABLE|
407 BCH_BUCKET_MARK_GC_LOCK_HELD);
410 spin_unlock(&c->freelist_lock);
412 for (ob = c->open_buckets;
413 ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
415 spin_lock(&ob->lock);
417 gc_pos_set(c, gc_pos_alloc(c, ob));
418 ca = bch_dev_bkey_exists(c, ob->ptr.dev);
419 bch2_mark_alloc_bucket(c, ca, PTR_BUCKET(ca, &ob->ptr), true,
421 BCH_BUCKET_MARK_MAY_MAKE_UNAVAILABLE|
422 BCH_BUCKET_MARK_GC_LOCK_HELD);
424 spin_unlock(&ob->lock);
428 static void bch2_gc_start(struct bch_fs *c)
432 struct bucket_mark new;
436 lg_global_lock(&c->usage_lock);
439 * Indicates to buckets code that gc is now in progress - done under
440 * usage_lock to avoid racing with bch2_mark_key():
442 __gc_pos_set(c, GC_POS_MIN);
444 /* Save a copy of the existing bucket stats while we recompute them: */
445 for_each_member_device(ca, c, i) {
446 ca->usage_cached = __bch2_dev_usage_read(ca);
447 for_each_possible_cpu(cpu) {
448 struct bch_dev_usage *p =
449 per_cpu_ptr(ca->usage_percpu, cpu);
450 memset(p, 0, sizeof(*p));
454 c->usage_cached = __bch2_fs_usage_read(c);
455 for_each_possible_cpu(cpu) {
456 struct bch_fs_usage *p =
457 per_cpu_ptr(c->usage_percpu, cpu);
459 memset(p->s, 0, sizeof(p->s));
462 lg_global_unlock(&c->usage_lock);
464 /* Clear bucket marks: */
465 for_each_member_device(ca, c, i)
466 for_each_bucket(g, ca) {
467 bucket_cmpxchg(g, new, ({
468 new.owned_by_allocator = 0;
470 new.cached_sectors = 0;
471 new.dirty_sectors = 0;
473 ca->oldest_gens[g - ca->buckets] = new.gen;
478 * bch_gc - recompute bucket marks and oldest_gen, rewrite btree nodes
480 void bch2_gc(struct bch_fs *c)
483 u64 start_time = local_clock();
487 * Walk _all_ references to buckets, and recompute them:
489 * Order matters here:
490 * - Concurrent GC relies on the fact that we have a total ordering for
491 * everything that GC walks - see gc_will_visit_node(),
492 * gc_will_visit_root()
494 * - also, references move around in the course of index updates and
495 * various other crap: everything needs to agree on the ordering
496 * references are allowed to move around in - e.g., we're allowed to
497 * start with a reference owned by an open_bucket (the allocator) and
498 * move it to the btree, but not the reverse.
500 * This is necessary to ensure that gc doesn't miss references that
501 * move around - if references move backwards in the ordering GC
502 * uses, GC could skip past them
507 * Do this before taking gc_lock - bch2_disk_reservation_get() blocks on
508 * gc_lock if sectors_available goes to 0:
510 bch2_recalc_sectors_available(c);
512 down_write(&c->gc_lock);
513 if (test_bit(BCH_FS_GC_FAILURE, &c->flags))
519 while (c->gc_pos.phase < (int) BTREE_ID_NR) {
520 int ret = c->btree_roots[c->gc_pos.phase].b
521 ? bch2_gc_btree(c, (int) c->gc_pos.phase)
525 bch_err(c, "btree gc failed: %d", ret);
526 set_bit(BCH_FS_GC_FAILURE, &c->flags);
530 gc_pos_set(c, gc_phase(c->gc_pos.phase + 1));
533 bch2_mark_superblocks(c);
534 bch2_mark_pending_btree_node_frees(c);
535 bch2_mark_allocator_buckets(c);
537 for_each_member_device(ca, c, i)
538 atomic_long_set(&ca->saturated_count, 0);
540 /* Indicates that gc is no longer in progress: */
541 gc_pos_set(c, gc_phase(GC_PHASE_DONE));
544 up_write(&c->gc_lock);
546 bch2_time_stats_update(&c->btree_gc_time, start_time);
549 * Wake up allocator in case it was waiting for buckets
550 * because of not being able to inc gens
552 for_each_member_device(ca, c, i)
553 bch2_wake_allocator(ca);
556 * At startup, allocations can happen directly instead of via the
557 * allocator thread - issue wakeup in case they blocked on gc_lock:
559 closure_wake_up(&c->freelist_wait);
562 /* Btree coalescing */
564 static void recalc_packed_keys(struct btree *b)
566 struct bkey_packed *k;
568 memset(&b->nr, 0, sizeof(b->nr));
570 BUG_ON(b->nsets != 1);
572 for (k = btree_bkey_first(b, b->set);
573 k != btree_bkey_last(b, b->set);
575 btree_keys_account_key_add(&b->nr, 0, k);
578 static void bch2_coalesce_nodes(struct bch_fs *c, struct btree_iter *iter,
579 struct btree *old_nodes[GC_MERGE_NODES])
581 struct btree *parent = iter->nodes[old_nodes[0]->level + 1];
582 unsigned i, nr_old_nodes, nr_new_nodes, u64s = 0;
583 unsigned blocks = btree_blocks(c) * 2 / 3;
584 struct btree *new_nodes[GC_MERGE_NODES];
585 struct btree_update *as;
586 struct keylist keylist;
587 struct bkey_format_state format_state;
588 struct bkey_format new_format;
590 memset(new_nodes, 0, sizeof(new_nodes));
591 bch2_keylist_init(&keylist, NULL);
593 /* Count keys that are not deleted */
594 for (i = 0; i < GC_MERGE_NODES && old_nodes[i]; i++)
595 u64s += old_nodes[i]->nr.live_u64s;
597 nr_old_nodes = nr_new_nodes = i;
599 /* Check if all keys in @old_nodes could fit in one fewer node */
600 if (nr_old_nodes <= 1 ||
601 __vstruct_blocks(struct btree_node, c->block_bits,
602 DIV_ROUND_UP(u64s, nr_old_nodes - 1)) > blocks)
605 /* Find a format that all keys in @old_nodes can pack into */
606 bch2_bkey_format_init(&format_state);
608 for (i = 0; i < nr_old_nodes; i++)
609 __bch2_btree_calc_format(&format_state, old_nodes[i]);
611 new_format = bch2_bkey_format_done(&format_state);
613 /* Check if repacking would make any nodes too big to fit */
614 for (i = 0; i < nr_old_nodes; i++)
615 if (!bch2_btree_node_format_fits(c, old_nodes[i], &new_format)) {
616 trace_btree_gc_coalesce_fail(c,
617 BTREE_GC_COALESCE_FAIL_FORMAT_FITS);
621 if (bch2_keylist_realloc(&keylist, NULL, 0,
622 (BKEY_U64s + BKEY_EXTENT_U64s_MAX) * nr_old_nodes)) {
623 trace_btree_gc_coalesce_fail(c,
624 BTREE_GC_COALESCE_FAIL_KEYLIST_REALLOC);
628 as = bch2_btree_update_start(c, iter->btree_id,
629 btree_update_reserve_required(c, parent) + nr_old_nodes,
631 BTREE_INSERT_USE_RESERVE,
634 trace_btree_gc_coalesce_fail(c,
635 BTREE_GC_COALESCE_FAIL_RESERVE_GET);
636 bch2_keylist_free(&keylist, NULL);
640 trace_btree_gc_coalesce(c, old_nodes[0]);
642 for (i = 0; i < nr_old_nodes; i++)
643 bch2_btree_interior_update_will_free_node(as, old_nodes[i]);
645 /* Repack everything with @new_format and sort down to one bset */
646 for (i = 0; i < nr_old_nodes; i++)
648 __bch2_btree_node_alloc_replacement(as, old_nodes[i],
652 * Conceptually we concatenate the nodes together and slice them
653 * up at different boundaries.
655 for (i = nr_new_nodes - 1; i > 0; --i) {
656 struct btree *n1 = new_nodes[i];
657 struct btree *n2 = new_nodes[i - 1];
659 struct bset *s1 = btree_bset_first(n1);
660 struct bset *s2 = btree_bset_first(n2);
661 struct bkey_packed *k, *last = NULL;
663 /* Calculate how many keys from @n2 we could fit inside @n1 */
667 k < vstruct_last(s2) &&
668 vstruct_blocks_plus(n1->data, c->block_bits,
669 u64s + k->u64s) <= blocks;
675 if (u64s == le16_to_cpu(s2->u64s)) {
676 /* n2 fits entirely in n1 */
677 n1->key.k.p = n1->data->max_key = n2->data->max_key;
679 memcpy_u64s(vstruct_last(s1),
681 le16_to_cpu(s2->u64s));
682 le16_add_cpu(&s1->u64s, le16_to_cpu(s2->u64s));
684 set_btree_bset_end(n1, n1->set);
686 six_unlock_write(&n2->lock);
687 bch2_btree_node_free_never_inserted(c, n2);
688 six_unlock_intent(&n2->lock);
690 memmove(new_nodes + i - 1,
692 sizeof(new_nodes[0]) * (nr_new_nodes - i));
693 new_nodes[--nr_new_nodes] = NULL;
695 /* move part of n2 into n1 */
696 n1->key.k.p = n1->data->max_key =
697 bkey_unpack_pos(n1, last);
700 btree_type_successor(iter->btree_id,
703 memcpy_u64s(vstruct_last(s1),
705 le16_add_cpu(&s1->u64s, u64s);
708 vstruct_idx(s2, u64s),
709 (le16_to_cpu(s2->u64s) - u64s) * sizeof(u64));
710 s2->u64s = cpu_to_le16(le16_to_cpu(s2->u64s) - u64s);
712 set_btree_bset_end(n1, n1->set);
713 set_btree_bset_end(n2, n2->set);
717 for (i = 0; i < nr_new_nodes; i++) {
718 struct btree *n = new_nodes[i];
720 recalc_packed_keys(n);
721 btree_node_reset_sib_u64s(n);
723 bch2_btree_build_aux_trees(n);
724 six_unlock_write(&n->lock);
726 bch2_btree_node_write(c, n, &as->cl, SIX_LOCK_intent);
730 * The keys for the old nodes get deleted. We don't want to insert keys
731 * that compare equal to the keys for the new nodes we'll also be
732 * inserting - we can't because keys on a keylist must be strictly
733 * greater than the previous keys, and we also don't need to since the
734 * key for the new node will serve the same purpose (overwriting the key
737 for (i = 0; i < nr_old_nodes; i++) {
738 struct bkey_i delete;
741 for (j = 0; j < nr_new_nodes; j++)
742 if (!bkey_cmp(old_nodes[i]->key.k.p,
743 new_nodes[j]->key.k.p))
746 bkey_init(&delete.k);
747 delete.k.p = old_nodes[i]->key.k.p;
748 bch2_keylist_add_in_order(&keylist, &delete);
754 * Keys for the new nodes get inserted: bch2_btree_insert_keys() only
755 * does the lookup once and thus expects the keys to be in sorted order
756 * so we have to make sure the new keys are correctly ordered with
757 * respect to the deleted keys added in the previous loop
759 for (i = 0; i < nr_new_nodes; i++)
760 bch2_keylist_add_in_order(&keylist, &new_nodes[i]->key);
762 /* Insert the newly coalesced nodes */
763 bch2_btree_insert_node(as, parent, iter, &keylist);
765 BUG_ON(!bch2_keylist_empty(&keylist));
767 BUG_ON(iter->nodes[old_nodes[0]->level] != old_nodes[0]);
769 BUG_ON(!bch2_btree_iter_node_replace(iter, new_nodes[0]));
771 for (i = 0; i < nr_new_nodes; i++)
772 bch2_btree_open_bucket_put(c, new_nodes[i]);
774 /* Free the old nodes and update our sliding window */
775 for (i = 0; i < nr_old_nodes; i++) {
776 bch2_btree_node_free_inmem(c, old_nodes[i], iter);
777 six_unlock_intent(&old_nodes[i]->lock);
780 * the index update might have triggered a split, in which case
781 * the nodes we coalesced - the new nodes we just created -
782 * might not be sibling nodes anymore - don't add them to the
783 * sliding window (except the first):
786 old_nodes[i] = new_nodes[i];
790 six_unlock_intent(&new_nodes[i]->lock);
794 bch2_btree_update_done(as);
795 bch2_keylist_free(&keylist, NULL);
798 static int bch2_coalesce_btree(struct bch_fs *c, enum btree_id btree_id)
800 struct btree_iter iter;
804 /* Sliding window of adjacent btree nodes */
805 struct btree *merge[GC_MERGE_NODES];
806 u32 lock_seq[GC_MERGE_NODES];
809 * XXX: We don't have a good way of positively matching on sibling nodes
810 * that have the same parent - this code works by handling the cases
811 * where they might not have the same parent, and is thus fragile. Ugh.
813 * Perhaps redo this to use multiple linked iterators?
815 memset(merge, 0, sizeof(merge));
817 __for_each_btree_node(&iter, c, btree_id, POS_MIN,
819 BTREE_ITER_PREFETCH, b) {
820 memmove(merge + 1, merge,
821 sizeof(merge) - sizeof(merge[0]));
822 memmove(lock_seq + 1, lock_seq,
823 sizeof(lock_seq) - sizeof(lock_seq[0]));
827 for (i = 1; i < GC_MERGE_NODES; i++) {
829 !six_relock_intent(&merge[i]->lock, lock_seq[i]))
832 if (merge[i]->level != merge[0]->level) {
833 six_unlock_intent(&merge[i]->lock);
837 memset(merge + i, 0, (GC_MERGE_NODES - i) * sizeof(merge[0]));
839 bch2_coalesce_nodes(c, &iter, merge);
841 for (i = 1; i < GC_MERGE_NODES && merge[i]; i++) {
842 lock_seq[i] = merge[i]->lock.state.seq;
843 six_unlock_intent(&merge[i]->lock);
846 lock_seq[0] = merge[0]->lock.state.seq;
848 if (test_bit(BCH_FS_GC_STOPPING, &c->flags)) {
849 bch2_btree_iter_unlock(&iter);
853 bch2_btree_iter_cond_resched(&iter);
856 * If the parent node wasn't relocked, it might have been split
857 * and the nodes in our sliding window might not have the same
858 * parent anymore - blow away the sliding window:
860 if (iter.nodes[iter.level + 1] &&
861 !btree_node_intent_locked(&iter, iter.level + 1))
863 (GC_MERGE_NODES - 1) * sizeof(merge[0]));
865 return bch2_btree_iter_unlock(&iter);
869 * bch_coalesce - coalesce adjacent nodes with low occupancy
871 void bch2_coalesce(struct bch_fs *c)
875 if (test_bit(BCH_FS_GC_FAILURE, &c->flags))
878 down_read(&c->gc_lock);
879 trace_gc_coalesce_start(c);
881 for (id = 0; id < BTREE_ID_NR; id++) {
882 int ret = c->btree_roots[id].b
883 ? bch2_coalesce_btree(c, id)
887 if (ret != -ESHUTDOWN)
888 bch_err(c, "btree coalescing failed: %d", ret);
889 set_bit(BCH_FS_GC_FAILURE, &c->flags);
894 trace_gc_coalesce_end(c);
895 up_read(&c->gc_lock);
898 static int bch2_gc_thread(void *arg)
900 struct bch_fs *c = arg;
901 struct io_clock *clock = &c->io_clock[WRITE];
902 unsigned long last = atomic_long_read(&clock->now);
903 unsigned last_kick = atomic_read(&c->kick_gc);
909 set_current_state(TASK_INTERRUPTIBLE);
911 if (kthread_should_stop()) {
912 __set_current_state(TASK_RUNNING);
916 if (atomic_read(&c->kick_gc) != last_kick)
919 if (c->btree_gc_periodic) {
920 unsigned long next = last + c->capacity / 16;
922 if (atomic_long_read(&clock->now) >= next)
925 bch2_io_clock_schedule_timeout(clock, next);
932 __set_current_state(TASK_RUNNING);
934 last = atomic_long_read(&clock->now);
935 last_kick = atomic_read(&c->kick_gc);
939 debug_check_no_locks_held();
945 void bch2_gc_thread_stop(struct bch_fs *c)
947 set_bit(BCH_FS_GC_STOPPING, &c->flags);
950 kthread_stop(c->gc_thread);
953 clear_bit(BCH_FS_GC_STOPPING, &c->flags);
956 int bch2_gc_thread_start(struct bch_fs *c)
958 struct task_struct *p;
960 BUG_ON(c->gc_thread);
962 p = kthread_create(bch2_gc_thread, c, "bcache_gc");
967 wake_up_process(c->gc_thread);
971 /* Initial GC computes bucket marks during startup */
973 static int bch2_initial_gc_btree(struct bch_fs *c, enum btree_id id)
975 struct btree_iter iter;
977 struct range_checks r;
980 btree_node_range_checks_init(&r, 0);
982 if (!c->btree_roots[id].b)
985 ret = bch2_btree_mark_key_initial(c, BKEY_TYPE_BTREE,
986 bkey_i_to_s_c(&c->btree_roots[id].b->key));
991 * We have to hit every btree node before starting journal replay, in
992 * order for the journal seq blacklist machinery to work:
994 for_each_btree_node(&iter, c, id, POS_MIN, BTREE_ITER_PREFETCH, b) {
995 btree_node_range_checks(c, b, &r);
997 if (btree_node_has_ptrs(b)) {
998 struct btree_node_iter node_iter;
999 struct bkey unpacked;
1002 for_each_btree_node_key_unpack(b, k, &node_iter,
1003 btree_node_is_extents(b),
1005 ret = bch2_btree_mark_key_initial(c,
1006 btree_node_type(b), k);
1012 bch2_btree_iter_cond_resched(&iter);
1015 return bch2_btree_iter_unlock(&iter) ?: ret;
1018 int bch2_initial_gc(struct bch_fs *c, struct list_head *journal)
1024 mutex_lock(&c->sb_lock);
1025 if (!bch2_sb_get_replicas(c->disk_sb)) {
1026 if (BCH_SB_INITIALIZED(c->disk_sb))
1027 bch_info(c, "building replicas info");
1028 set_bit(BCH_FS_REBUILD_REPLICAS, &c->flags);
1030 mutex_unlock(&c->sb_lock);
1034 for (id = 0; id < BTREE_ID_NR; id++) {
1035 ret = bch2_initial_gc_btree(c, id);
1040 ret = bch2_journal_mark(c, journal);
1044 if (test_bit(BCH_FS_FIXED_GENS, &c->flags)) {
1046 bch_info(c, "Unable to fix bucket gens, looping");
1050 bch_info(c, "Fixed gens, restarting initial mark and sweep:");
1051 clear_bit(BCH_FS_FIXED_GENS, &c->flags);
1056 * Skip past versions that might have possibly been used (as nonces),
1057 * but hadn't had their pointers written:
1059 if (c->sb.encryption_type)
1060 atomic64_add(1 << 16, &c->key_version);
1062 bch2_mark_superblocks(c);
1064 gc_pos_set(c, gc_phase(GC_PHASE_DONE));
1065 set_bit(BCH_FS_INITIAL_GC_DONE, &c->flags);