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.h"
23 #include <linux/slab.h>
24 #include <linux/bitops.h>
25 #include <linux/freezer.h>
26 #include <linux/kthread.h>
27 #include <linux/rcupdate.h>
28 #include <trace/events/bcachefs.h>
38 static void btree_node_range_checks_init(struct range_checks *r, unsigned depth)
42 for (i = 0; i < BTREE_MAX_DEPTH; i++)
43 r->l[i].min = r->l[i].max = POS_MIN;
47 static void btree_node_range_checks(struct bch_fs *c, struct btree *b,
48 struct range_checks *r)
50 struct range_level *l = &r->l[b->level];
52 struct bpos expected_min = bkey_cmp(l->min, l->max)
53 ? btree_type_successor(b->btree_id, l->max)
56 bch2_fs_inconsistent_on(bkey_cmp(b->data->min_key, expected_min), c,
57 "btree node has incorrect min key: %llu:%llu != %llu:%llu",
58 b->data->min_key.inode,
59 b->data->min_key.offset,
63 l->max = b->data->max_key;
65 if (b->level > r->depth) {
66 l = &r->l[b->level - 1];
68 bch2_fs_inconsistent_on(bkey_cmp(b->data->min_key, l->min), c,
69 "btree node min doesn't match min of child nodes: %llu:%llu != %llu:%llu",
70 b->data->min_key.inode,
71 b->data->min_key.offset,
75 bch2_fs_inconsistent_on(bkey_cmp(b->data->max_key, l->max), c,
76 "btree node max doesn't match max of child nodes: %llu:%llu != %llu:%llu",
77 b->data->max_key.inode,
78 b->data->max_key.offset,
82 if (bkey_cmp(b->data->max_key, POS_MAX))
84 btree_type_successor(b->btree_id,
89 u8 bch2_btree_key_recalc_oldest_gen(struct bch_fs *c, struct bkey_s_c k)
91 const struct bch_extent_ptr *ptr;
94 if (bkey_extent_is_data(k.k)) {
95 struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
97 extent_for_each_ptr(e, ptr) {
98 struct bch_dev *ca = c->devs[ptr->dev];
99 size_t b = PTR_BUCKET_NR(ca, ptr);
101 if (gen_after(ca->oldest_gens[b], ptr->gen))
102 ca->oldest_gens[b] = ptr->gen;
104 max_stale = max(max_stale, ptr_stale(ca, ptr));
112 * For runtime mark and sweep:
114 static u8 bch2_btree_mark_key(struct bch_fs *c, enum bkey_type type,
118 case BKEY_TYPE_BTREE:
119 bch2_gc_mark_key(c, k, c->sb.btree_node_size, true);
121 case BKEY_TYPE_EXTENTS:
122 bch2_gc_mark_key(c, k, k.k->size, false);
123 return bch2_btree_key_recalc_oldest_gen(c, k);
129 int bch2_btree_mark_key_initial(struct bch_fs *c, enum bkey_type type,
136 case BCH_EXTENT_CACHED: {
137 struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
138 const struct bch_extent_ptr *ptr;
140 extent_for_each_ptr(e, ptr) {
141 struct bch_dev *ca = c->devs[ptr->dev];
142 struct bucket *g = PTR_BUCKET(ca, ptr);
143 struct bucket_mark new;
145 if (fsck_err_on(gen_cmp(ptr->gen, g->mark.gen) > 0, c,
146 "%s ptr gen in the future: %u > %u",
147 type == BKEY_TYPE_BTREE
149 ptr->gen, g->mark.gen)) {
150 bucket_cmpxchg(g, new, new.gen = ptr->gen);
151 set_bit(BCH_FS_FIXED_GENS, &c->flags);
152 ca->need_prio_write = true;
160 atomic64_set(&c->key_version,
161 max_t(u64, k.k->version.lo,
162 atomic64_read(&c->key_version)));
164 bch2_btree_mark_key(c, type, k);
169 static bool btree_gc_mark_node(struct bch_fs *c, struct btree *b)
171 if (btree_node_has_ptrs(b)) {
172 struct btree_node_iter iter;
173 struct bkey unpacked;
177 for_each_btree_node_key_unpack(b, k, &iter,
178 btree_node_is_extents(b),
180 bch2_bkey_debugcheck(c, b, k);
181 stale = max(stale, bch2_btree_mark_key(c,
182 btree_node_type(b), k));
185 if (btree_gc_rewrite_disabled(c))
192 if (btree_gc_always_rewrite(c))
198 static inline void __gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
200 write_seqcount_begin(&c->gc_pos_lock);
202 write_seqcount_end(&c->gc_pos_lock);
205 static inline void gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
207 BUG_ON(gc_pos_cmp(new_pos, c->gc_pos) <= 0);
208 __gc_pos_set(c, new_pos);
211 static int bch2_gc_btree(struct bch_fs *c, enum btree_id btree_id)
213 struct btree_iter iter;
216 struct range_checks r;
217 unsigned depth = btree_id == BTREE_ID_EXTENTS ? 0 : 1;
221 * if expensive_debug_checks is on, run range_checks on all leaf nodes:
223 if (expensive_debug_checks(c))
226 btree_node_range_checks_init(&r, depth);
228 for_each_btree_node(&iter, c, btree_id, POS_MIN, depth, b) {
229 btree_node_range_checks(c, b, &r);
231 bch2_verify_btree_nr_keys(b);
233 should_rewrite = btree_gc_mark_node(c, b);
235 gc_pos_set(c, gc_pos_btree_node(b));
238 bch2_btree_node_rewrite(&iter, b, NULL);
240 bch2_btree_iter_cond_resched(&iter);
242 ret = bch2_btree_iter_unlock(&iter);
246 mutex_lock(&c->btree_root_lock);
248 b = c->btree_roots[btree_id].b;
249 bch2_btree_mark_key(c, BKEY_TYPE_BTREE, bkey_i_to_s_c(&b->key));
250 gc_pos_set(c, gc_pos_btree_root(b->btree_id));
252 mutex_unlock(&c->btree_root_lock);
256 static void bch2_mark_allocator_buckets(struct bch_fs *c)
259 struct open_bucket *ob;
263 for_each_member_device(ca, c, ci) {
264 spin_lock(&ca->freelist_lock);
266 fifo_for_each_entry(i, &ca->free_inc, iter)
267 bch2_mark_alloc_bucket(ca, &ca->buckets[i], true);
269 for (j = 0; j < RESERVE_NR; j++)
270 fifo_for_each_entry(i, &ca->free[j], iter)
271 bch2_mark_alloc_bucket(ca, &ca->buckets[i], true);
273 spin_unlock(&ca->freelist_lock);
276 for (ob = c->open_buckets;
277 ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
279 const struct bch_extent_ptr *ptr;
281 mutex_lock(&ob->lock);
282 open_bucket_for_each_ptr(ob, ptr) {
283 ca = c->devs[ptr->dev];
284 bch2_mark_alloc_bucket(ca, PTR_BUCKET(ca, ptr), true);
286 mutex_unlock(&ob->lock);
290 static void mark_metadata_sectors(struct bch_dev *ca, u64 start, u64 end,
291 enum bucket_data_type type)
293 u64 b = start >> ca->bucket_bits;
296 bch2_mark_metadata_bucket(ca, ca->buckets + b, type, true);
298 } while (b < end >> ca->bucket_bits);
301 static void bch2_dev_mark_superblocks(struct bch_dev *ca)
303 struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
306 for (i = 0; i < layout->nr_superblocks; i++) {
307 if (layout->sb_offset[i] == BCH_SB_SECTOR)
308 mark_metadata_sectors(ca, 0, BCH_SB_SECTOR,
311 mark_metadata_sectors(ca,
312 layout->sb_offset[i],
313 layout->sb_offset[i] +
314 (1 << layout->sb_max_size_bits),
320 * Mark non btree metadata - prios, journal
322 void bch2_mark_dev_metadata(struct bch_fs *c, struct bch_dev *ca)
327 lockdep_assert_held(&c->sb_lock);
329 bch2_dev_mark_superblocks(ca);
331 spin_lock(&c->journal.lock);
333 for (i = 0; i < ca->journal.nr; i++) {
334 b = ca->journal.buckets[i];
335 bch2_mark_metadata_bucket(ca, ca->buckets + b,
336 BUCKET_JOURNAL, true);
339 spin_unlock(&c->journal.lock);
341 spin_lock(&ca->prio_buckets_lock);
343 for (i = 0; i < prio_buckets(ca) * 2; i++) {
344 b = ca->prio_buckets[i];
346 bch2_mark_metadata_bucket(ca, ca->buckets + b,
350 spin_unlock(&ca->prio_buckets_lock);
353 static void bch2_mark_metadata(struct bch_fs *c)
358 mutex_lock(&c->sb_lock);
359 gc_pos_set(c, gc_phase(GC_PHASE_SB_METADATA));
361 for_each_online_member(ca, c, i)
362 bch2_mark_dev_metadata(c, ca);
363 mutex_unlock(&c->sb_lock);
366 /* Also see bch2_pending_btree_node_free_insert_done() */
367 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
369 struct bch_fs_usage stats = { 0 };
370 struct btree_interior_update *as;
371 struct pending_btree_node_free *d;
373 mutex_lock(&c->btree_interior_update_lock);
374 gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
376 for_each_pending_btree_node_free(c, as, d)
377 if (d->index_update_done)
378 __bch2_gc_mark_key(c, bkey_i_to_s_c(&d->key),
379 c->sb.btree_node_size, true,
382 * Don't apply stats - pending deletes aren't tracked in
386 mutex_unlock(&c->btree_interior_update_lock);
389 void bch2_gc_start(struct bch_fs *c)
393 struct bucket_mark new;
397 lg_global_lock(&c->usage_lock);
400 * Indicates to buckets code that gc is now in progress - done under
401 * usage_lock to avoid racing with bch2_mark_key():
403 __gc_pos_set(c, GC_POS_MIN);
405 /* Save a copy of the existing bucket stats while we recompute them: */
406 for_each_member_device(ca, c, i) {
407 ca->usage_cached = __bch2_dev_usage_read(ca);
408 for_each_possible_cpu(cpu) {
409 struct bch_dev_usage *p =
410 per_cpu_ptr(ca->usage_percpu, cpu);
411 memset(p, 0, sizeof(*p));
415 c->usage_cached = __bch2_fs_usage_read(c);
416 for_each_possible_cpu(cpu) {
417 struct bch_fs_usage *p =
418 per_cpu_ptr(c->usage_percpu, cpu);
420 memset(p->s, 0, sizeof(p->s));
421 p->persistent_reserved = 0;
424 lg_global_unlock(&c->usage_lock);
426 /* Clear bucket marks: */
427 for_each_member_device(ca, c, i)
428 for_each_bucket(g, ca) {
429 bucket_cmpxchg(g, new, ({
430 new.owned_by_allocator = 0;
432 new.cached_sectors = 0;
433 new.dirty_sectors = 0;
435 ca->oldest_gens[g - ca->buckets] = new.gen;
440 * bch_gc - recompute bucket marks and oldest_gen, rewrite btree nodes
442 void bch2_gc(struct bch_fs *c)
445 u64 start_time = local_clock();
449 * Walk _all_ references to buckets, and recompute them:
451 * Order matters here:
452 * - Concurrent GC relies on the fact that we have a total ordering for
453 * everything that GC walks - see gc_will_visit_node(),
454 * gc_will_visit_root()
456 * - also, references move around in the course of index updates and
457 * various other crap: everything needs to agree on the ordering
458 * references are allowed to move around in - e.g., we're allowed to
459 * start with a reference owned by an open_bucket (the allocator) and
460 * move it to the btree, but not the reverse.
462 * This is necessary to ensure that gc doesn't miss references that
463 * move around - if references move backwards in the ordering GC
464 * uses, GC could skip past them
467 if (test_bit(BCH_FS_GC_FAILURE, &c->flags))
473 * Do this before taking gc_lock - bch2_disk_reservation_get() blocks on
474 * gc_lock if sectors_available goes to 0:
476 bch2_recalc_sectors_available(c);
478 down_write(&c->gc_lock);
482 /* Walk allocator's references: */
483 bch2_mark_allocator_buckets(c);
486 while (c->gc_pos.phase < (int) BTREE_ID_NR) {
487 int ret = c->btree_roots[c->gc_pos.phase].b
488 ? bch2_gc_btree(c, (int) c->gc_pos.phase)
492 bch_err(c, "btree gc failed: %d", ret);
493 set_bit(BCH_FS_GC_FAILURE, &c->flags);
494 up_write(&c->gc_lock);
498 gc_pos_set(c, gc_phase(c->gc_pos.phase + 1));
501 bch2_mark_metadata(c);
502 bch2_mark_pending_btree_node_frees(c);
504 for_each_member_device(ca, c, i)
505 atomic_long_set(&ca->saturated_count, 0);
507 /* Indicates that gc is no longer in progress: */
508 gc_pos_set(c, gc_phase(GC_PHASE_DONE));
510 up_write(&c->gc_lock);
512 bch2_time_stats_update(&c->btree_gc_time, start_time);
515 * Wake up allocator in case it was waiting for buckets
516 * because of not being able to inc gens
518 for_each_member_device(ca, c, i)
519 bch2_wake_allocator(ca);
522 /* Btree coalescing */
524 static void recalc_packed_keys(struct btree *b)
526 struct bkey_packed *k;
528 memset(&b->nr, 0, sizeof(b->nr));
530 BUG_ON(b->nsets != 1);
532 for (k = btree_bkey_first(b, b->set);
533 k != btree_bkey_last(b, b->set);
535 btree_keys_account_key_add(&b->nr, 0, k);
538 static void bch2_coalesce_nodes(struct btree *old_nodes[GC_MERGE_NODES],
539 struct btree_iter *iter)
541 struct btree *parent = iter->nodes[old_nodes[0]->level + 1];
542 struct bch_fs *c = iter->c;
543 unsigned i, nr_old_nodes, nr_new_nodes, u64s = 0;
544 unsigned blocks = btree_blocks(c) * 2 / 3;
545 struct btree *new_nodes[GC_MERGE_NODES];
546 struct btree_interior_update *as;
547 struct btree_reserve *res;
548 struct keylist keylist;
549 struct bkey_format_state format_state;
550 struct bkey_format new_format;
552 memset(new_nodes, 0, sizeof(new_nodes));
553 bch2_keylist_init(&keylist, NULL, 0);
555 /* Count keys that are not deleted */
556 for (i = 0; i < GC_MERGE_NODES && old_nodes[i]; i++)
557 u64s += old_nodes[i]->nr.live_u64s;
559 nr_old_nodes = nr_new_nodes = i;
561 /* Check if all keys in @old_nodes could fit in one fewer node */
562 if (nr_old_nodes <= 1 ||
563 __vstruct_blocks(struct btree_node, c->block_bits,
564 DIV_ROUND_UP(u64s, nr_old_nodes - 1)) > blocks)
567 res = bch2_btree_reserve_get(c, parent, nr_old_nodes,
569 BTREE_INSERT_USE_RESERVE,
572 trace_btree_gc_coalesce_fail(c,
573 BTREE_GC_COALESCE_FAIL_RESERVE_GET);
577 if (bch2_keylist_realloc(&keylist, NULL, 0,
578 (BKEY_U64s + BKEY_EXTENT_U64s_MAX) * nr_old_nodes)) {
579 trace_btree_gc_coalesce_fail(c,
580 BTREE_GC_COALESCE_FAIL_KEYLIST_REALLOC);
584 /* Find a format that all keys in @old_nodes can pack into */
585 bch2_bkey_format_init(&format_state);
587 for (i = 0; i < nr_old_nodes; i++)
588 __bch2_btree_calc_format(&format_state, old_nodes[i]);
590 new_format = bch2_bkey_format_done(&format_state);
592 /* Check if repacking would make any nodes too big to fit */
593 for (i = 0; i < nr_old_nodes; i++)
594 if (!bch2_btree_node_format_fits(c, old_nodes[i], &new_format)) {
595 trace_btree_gc_coalesce_fail(c,
596 BTREE_GC_COALESCE_FAIL_FORMAT_FITS);
600 trace_btree_gc_coalesce(c, parent, nr_old_nodes);
602 as = bch2_btree_interior_update_alloc(c);
604 for (i = 0; i < nr_old_nodes; i++)
605 bch2_btree_interior_update_will_free_node(c, as, old_nodes[i]);
607 /* Repack everything with @new_format and sort down to one bset */
608 for (i = 0; i < nr_old_nodes; i++) {
610 __bch2_btree_node_alloc_replacement(c, old_nodes[i],
612 list_add(&new_nodes[i]->reachable, &as->reachable_list);
616 * Conceptually we concatenate the nodes together and slice them
617 * up at different boundaries.
619 for (i = nr_new_nodes - 1; i > 0; --i) {
620 struct btree *n1 = new_nodes[i];
621 struct btree *n2 = new_nodes[i - 1];
623 struct bset *s1 = btree_bset_first(n1);
624 struct bset *s2 = btree_bset_first(n2);
625 struct bkey_packed *k, *last = NULL;
627 /* Calculate how many keys from @n2 we could fit inside @n1 */
631 k < vstruct_last(s2) &&
632 vstruct_blocks_plus(n1->data, c->block_bits,
633 u64s + k->u64s) <= blocks;
639 if (u64s == le16_to_cpu(s2->u64s)) {
640 /* n2 fits entirely in n1 */
641 n1->key.k.p = n1->data->max_key = n2->data->max_key;
643 memcpy_u64s(vstruct_last(s1),
645 le16_to_cpu(s2->u64s));
646 le16_add_cpu(&s1->u64s, le16_to_cpu(s2->u64s));
648 set_btree_bset_end(n1, n1->set);
650 list_del_init(&n2->reachable);
651 six_unlock_write(&n2->lock);
652 bch2_btree_node_free_never_inserted(c, n2);
653 six_unlock_intent(&n2->lock);
655 memmove(new_nodes + i - 1,
657 sizeof(new_nodes[0]) * (nr_new_nodes - i));
658 new_nodes[--nr_new_nodes] = NULL;
660 /* move part of n2 into n1 */
661 n1->key.k.p = n1->data->max_key =
662 bkey_unpack_pos(n1, last);
665 btree_type_successor(iter->btree_id,
668 memcpy_u64s(vstruct_last(s1),
670 le16_add_cpu(&s1->u64s, u64s);
673 vstruct_idx(s2, u64s),
674 (le16_to_cpu(s2->u64s) - u64s) * sizeof(u64));
675 s2->u64s = cpu_to_le16(le16_to_cpu(s2->u64s) - u64s);
677 set_btree_bset_end(n1, n1->set);
678 set_btree_bset_end(n2, n2->set);
682 for (i = 0; i < nr_new_nodes; i++) {
683 struct btree *n = new_nodes[i];
685 recalc_packed_keys(n);
686 btree_node_reset_sib_u64s(n);
688 bch2_btree_build_aux_trees(n);
689 six_unlock_write(&n->lock);
691 bch2_btree_node_write(c, n, &as->cl, SIX_LOCK_intent);
695 * The keys for the old nodes get deleted. We don't want to insert keys
696 * that compare equal to the keys for the new nodes we'll also be
697 * inserting - we can't because keys on a keylist must be strictly
698 * greater than the previous keys, and we also don't need to since the
699 * key for the new node will serve the same purpose (overwriting the key
702 for (i = 0; i < nr_old_nodes; i++) {
703 struct bkey_i delete;
706 for (j = 0; j < nr_new_nodes; j++)
707 if (!bkey_cmp(old_nodes[i]->key.k.p,
708 new_nodes[j]->key.k.p))
711 bkey_init(&delete.k);
712 delete.k.p = old_nodes[i]->key.k.p;
713 bch2_keylist_add_in_order(&keylist, &delete);
719 * Keys for the new nodes get inserted: bch2_btree_insert_keys() only
720 * does the lookup once and thus expects the keys to be in sorted order
721 * so we have to make sure the new keys are correctly ordered with
722 * respect to the deleted keys added in the previous loop
724 for (i = 0; i < nr_new_nodes; i++)
725 bch2_keylist_add_in_order(&keylist, &new_nodes[i]->key);
727 /* Insert the newly coalesced nodes */
728 bch2_btree_insert_node(parent, iter, &keylist, res, as);
730 BUG_ON(!bch2_keylist_empty(&keylist));
732 BUG_ON(iter->nodes[old_nodes[0]->level] != old_nodes[0]);
734 BUG_ON(!bch2_btree_iter_node_replace(iter, new_nodes[0]));
736 for (i = 0; i < nr_new_nodes; i++)
737 bch2_btree_open_bucket_put(c, new_nodes[i]);
739 /* Free the old nodes and update our sliding window */
740 for (i = 0; i < nr_old_nodes; i++) {
741 bch2_btree_node_free_inmem(iter, old_nodes[i]);
742 six_unlock_intent(&old_nodes[i]->lock);
745 * the index update might have triggered a split, in which case
746 * the nodes we coalesced - the new nodes we just created -
747 * might not be sibling nodes anymore - don't add them to the
748 * sliding window (except the first):
751 old_nodes[i] = new_nodes[i];
755 six_unlock_intent(&new_nodes[i]->lock);
759 bch2_keylist_free(&keylist, NULL);
760 bch2_btree_reserve_put(c, res);
763 static int bch2_coalesce_btree(struct bch_fs *c, enum btree_id btree_id)
765 struct btree_iter iter;
769 /* Sliding window of adjacent btree nodes */
770 struct btree *merge[GC_MERGE_NODES];
771 u32 lock_seq[GC_MERGE_NODES];
774 * XXX: We don't have a good way of positively matching on sibling nodes
775 * that have the same parent - this code works by handling the cases
776 * where they might not have the same parent, and is thus fragile. Ugh.
778 * Perhaps redo this to use multiple linked iterators?
780 memset(merge, 0, sizeof(merge));
782 __for_each_btree_node(&iter, c, btree_id, POS_MIN, 0, b, U8_MAX) {
783 memmove(merge + 1, merge,
784 sizeof(merge) - sizeof(merge[0]));
785 memmove(lock_seq + 1, lock_seq,
786 sizeof(lock_seq) - sizeof(lock_seq[0]));
790 for (i = 1; i < GC_MERGE_NODES; i++) {
792 !six_relock_intent(&merge[i]->lock, lock_seq[i]))
795 if (merge[i]->level != merge[0]->level) {
796 six_unlock_intent(&merge[i]->lock);
800 memset(merge + i, 0, (GC_MERGE_NODES - i) * sizeof(merge[0]));
802 bch2_coalesce_nodes(merge, &iter);
804 for (i = 1; i < GC_MERGE_NODES && merge[i]; i++) {
805 lock_seq[i] = merge[i]->lock.state.seq;
806 six_unlock_intent(&merge[i]->lock);
809 lock_seq[0] = merge[0]->lock.state.seq;
811 if (test_bit(BCH_FS_GC_STOPPING, &c->flags)) {
812 bch2_btree_iter_unlock(&iter);
816 bch2_btree_iter_cond_resched(&iter);
819 * If the parent node wasn't relocked, it might have been split
820 * and the nodes in our sliding window might not have the same
821 * parent anymore - blow away the sliding window:
823 if (iter.nodes[iter.level + 1] &&
824 !btree_node_intent_locked(&iter, iter.level + 1))
826 (GC_MERGE_NODES - 1) * sizeof(merge[0]));
828 return bch2_btree_iter_unlock(&iter);
832 * bch_coalesce - coalesce adjacent nodes with low occupancy
834 void bch2_coalesce(struct bch_fs *c)
839 if (test_bit(BCH_FS_GC_FAILURE, &c->flags))
842 down_read(&c->gc_lock);
843 trace_gc_coalesce_start(c);
844 start_time = local_clock();
846 for (id = 0; id < BTREE_ID_NR; id++) {
847 int ret = c->btree_roots[id].b
848 ? bch2_coalesce_btree(c, id)
852 if (ret != -ESHUTDOWN)
853 bch_err(c, "btree coalescing failed: %d", ret);
854 set_bit(BCH_FS_GC_FAILURE, &c->flags);
859 bch2_time_stats_update(&c->btree_coalesce_time, start_time);
860 trace_gc_coalesce_end(c);
861 up_read(&c->gc_lock);
864 static int bch2_gc_thread(void *arg)
866 struct bch_fs *c = arg;
867 struct io_clock *clock = &c->io_clock[WRITE];
868 unsigned long last = atomic_long_read(&clock->now);
869 unsigned last_kick = atomic_read(&c->kick_gc);
874 unsigned long next = last + c->capacity / 16;
876 while (atomic_long_read(&clock->now) < next) {
877 set_current_state(TASK_INTERRUPTIBLE);
879 if (kthread_should_stop()) {
880 __set_current_state(TASK_RUNNING);
884 if (atomic_read(&c->kick_gc) != last_kick) {
885 __set_current_state(TASK_RUNNING);
889 bch2_io_clock_schedule_timeout(clock, next);
893 last = atomic_long_read(&clock->now);
894 last_kick = atomic_read(&c->kick_gc);
897 if (!btree_gc_coalesce_disabled(c))
900 debug_check_no_locks_held();
906 void bch2_gc_thread_stop(struct bch_fs *c)
908 set_bit(BCH_FS_GC_STOPPING, &c->flags);
911 kthread_stop(c->gc_thread);
914 clear_bit(BCH_FS_GC_STOPPING, &c->flags);
917 int bch2_gc_thread_start(struct bch_fs *c)
919 struct task_struct *p;
921 BUG_ON(c->gc_thread);
923 p = kthread_create(bch2_gc_thread, c, "bcache_gc");
928 wake_up_process(c->gc_thread);
932 /* Initial GC computes bucket marks during startup */
934 static int bch2_initial_gc_btree(struct bch_fs *c, enum btree_id id)
936 struct btree_iter iter;
938 struct range_checks r;
941 btree_node_range_checks_init(&r, 0);
943 if (!c->btree_roots[id].b)
946 ret = bch2_btree_mark_key_initial(c, BKEY_TYPE_BTREE,
947 bkey_i_to_s_c(&c->btree_roots[id].b->key));
952 * We have to hit every btree node before starting journal replay, in
953 * order for the journal seq blacklist machinery to work:
955 for_each_btree_node(&iter, c, id, POS_MIN, 0, b) {
956 btree_node_range_checks(c, b, &r);
958 if (btree_node_has_ptrs(b)) {
959 struct btree_node_iter node_iter;
960 struct bkey unpacked;
963 for_each_btree_node_key_unpack(b, k, &node_iter,
964 btree_node_is_extents(b),
966 ret = bch2_btree_mark_key_initial(c,
967 btree_node_type(b), k);
973 bch2_btree_iter_cond_resched(&iter);
976 bch2_btree_iter_unlock(&iter);
980 int bch2_initial_gc(struct bch_fs *c, struct list_head *journal)
988 for (id = 0; id < BTREE_ID_NR; id++) {
989 ret = bch2_initial_gc_btree(c, id);
995 ret = bch2_journal_mark(c, journal);
1000 bch2_mark_metadata(c);
1002 if (test_bit(BCH_FS_FIXED_GENS, &c->flags)) {
1004 bch_info(c, "Unable to fix bucket gens, looping");
1008 bch_info(c, "Fixed gens, restarting initial mark and sweep:");
1009 clear_bit(BCH_FS_FIXED_GENS, &c->flags);
1014 * Skip past versions that might have possibly been used (as nonces),
1015 * but hadn't had their pointers written:
1017 if (c->sb.encryption_type)
1018 atomic64_add(1 << 16, &c->key_version);
1020 gc_pos_set(c, gc_phase(GC_PHASE_DONE));
1021 set_bit(BCH_FS_INITIAL_GC_DONE, &c->flags);