1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2010 Kent Overstreet <kent.overstreet@gmail.com>
4 * Copyright (C) 2014 Datera Inc.
8 #include "alloc_background.h"
9 #include "alloc_foreground.h"
10 #include "bkey_methods.h"
12 #include "btree_locking.h"
13 #include "btree_update_interior.h"
30 #include <linux/slab.h>
31 #include <linux/bitops.h>
32 #include <linux/freezer.h>
33 #include <linux/kthread.h>
34 #include <linux/preempt.h>
35 #include <linux/rcupdate.h>
36 #include <linux/sched/task.h>
38 static inline void __gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
41 write_seqcount_begin(&c->gc_pos_lock);
43 write_seqcount_end(&c->gc_pos_lock);
47 static inline void gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
49 BUG_ON(gc_pos_cmp(new_pos, c->gc_pos) <= 0);
50 __gc_pos_set(c, new_pos);
53 static int bch2_gc_check_topology(struct bch_fs *c,
55 struct bpos *expected_start,
56 struct bpos expected_end,
61 if (k.k->type == KEY_TYPE_btree_ptr_v2) {
62 struct bkey_s_c_btree_ptr_v2 bp = bkey_s_c_to_btree_ptr_v2(k);
64 if (fsck_err_on(bkey_cmp(*expected_start, bp.v->min_key), c,
65 "btree node with incorrect min_key: got %llu:%llu, should be %llu:%llu",
68 expected_start->inode,
69 expected_start->offset)) {
74 *expected_start = bkey_cmp(k.k->p, POS_MAX)
75 ? bkey_successor(k.k->p)
78 if (fsck_err_on(is_last &&
79 bkey_cmp(k.k->p, expected_end), c,
80 "btree node with incorrect max_key: got %llu:%llu, should be %llu:%llu",
84 expected_end.offset)) {
91 /* marking of btree keys/nodes: */
93 static int bch2_gc_mark_key(struct bch_fs *c, struct bkey_s_c k,
94 u8 *max_stale, bool initial)
96 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
97 const struct bch_extent_ptr *ptr;
100 (initial ? BTREE_TRIGGER_NOATOMIC : 0);
104 BUG_ON(bch2_journal_seq_verify &&
105 k.k->version.lo > journal_cur_seq(&c->journal));
107 /* XXX change to fsck check */
108 if (fsck_err_on(k.k->version.lo > atomic64_read(&c->key_version), c,
109 "key version number higher than recorded: %llu > %llu",
111 atomic64_read(&c->key_version)))
112 atomic64_set(&c->key_version, k.k->version.lo);
114 if (test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) ||
115 fsck_err_on(!bch2_bkey_replicas_marked(c, k), c,
116 "superblock not marked as containing replicas (type %u)",
118 ret = bch2_mark_bkey_replicas(c, k);
123 bkey_for_each_ptr(ptrs, ptr) {
124 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
125 struct bucket *g = PTR_BUCKET(ca, ptr, true);
126 struct bucket *g2 = PTR_BUCKET(ca, ptr, false);
128 if (mustfix_fsck_err_on(!g->gen_valid, c,
129 "bucket %u:%zu data type %s ptr gen %u missing in alloc btree",
130 ptr->dev, PTR_BUCKET_NR(ca, ptr),
131 bch2_data_types[ptr_data_type(k.k, ptr)],
133 g2->_mark.gen = g->_mark.gen = ptr->gen;
134 g2->gen_valid = g->gen_valid = true;
135 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
138 if (mustfix_fsck_err_on(gen_cmp(ptr->gen, g->mark.gen) > 0, c,
139 "bucket %u:%zu data type %s ptr gen in the future: %u > %u",
140 ptr->dev, PTR_BUCKET_NR(ca, ptr),
141 bch2_data_types[ptr_data_type(k.k, ptr)],
142 ptr->gen, g->mark.gen)) {
143 g2->_mark.gen = g->_mark.gen = ptr->gen;
144 g2->gen_valid = g->gen_valid = true;
145 g2->_mark.data_type = 0;
146 g2->_mark.dirty_sectors = 0;
147 g2->_mark.cached_sectors = 0;
148 set_bit(BCH_FS_FIXED_GENS, &c->flags);
149 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
154 bkey_for_each_ptr(ptrs, ptr) {
155 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
156 struct bucket *g = PTR_BUCKET(ca, ptr, true);
158 if (gen_after(g->oldest_gen, ptr->gen))
159 g->oldest_gen = ptr->gen;
161 *max_stale = max(*max_stale, ptr_stale(ca, ptr));
164 bch2_mark_key(c, k, 0, k.k->size, NULL, 0, flags);
169 static int btree_gc_mark_node(struct bch_fs *c, struct btree *b, u8 *max_stale,
172 struct bpos next_node_start = b->data->min_key;
173 struct btree_node_iter iter;
174 struct bkey unpacked;
180 if (!btree_node_type_needs_gc(btree_node_type(b)))
183 bch2_btree_node_iter_init_from_start(&iter, b);
185 while ((k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked)).k) {
186 bch2_bkey_debugcheck(c, b, k);
188 ret = bch2_gc_mark_key(c, k, max_stale, initial);
192 bch2_btree_node_iter_advance(&iter, b);
195 ret = bch2_gc_check_topology(c, k,
198 bch2_btree_node_iter_end(&iter));
207 static int bch2_gc_btree(struct bch_fs *c, enum btree_id btree_id,
208 bool initial, bool metadata_only)
210 struct btree_trans trans;
211 struct btree_iter *iter;
213 unsigned depth = metadata_only ? 1
214 : bch2_expensive_debug_checks ? 0
215 : !btree_node_type_needs_gc(btree_id) ? 1
220 bch2_trans_init(&trans, c, 0, 0);
222 gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));
224 __for_each_btree_node(&trans, iter, btree_id, POS_MIN,
225 0, depth, BTREE_ITER_PREFETCH, b) {
226 bch2_verify_btree_nr_keys(b);
228 gc_pos_set(c, gc_pos_btree_node(b));
230 ret = btree_gc_mark_node(c, b, &max_stale, initial);
236 bch2_btree_node_rewrite(c, iter,
239 BTREE_INSERT_GC_LOCK_HELD);
240 else if (!bch2_btree_gc_rewrite_disabled &&
241 (bch2_btree_gc_always_rewrite || max_stale > 16))
242 bch2_btree_node_rewrite(c, iter,
245 BTREE_INSERT_GC_LOCK_HELD);
248 bch2_trans_cond_resched(&trans);
250 ret = bch2_trans_exit(&trans) ?: ret;
254 mutex_lock(&c->btree_root_lock);
255 b = c->btree_roots[btree_id].b;
256 if (!btree_node_fake(b))
257 ret = bch2_gc_mark_key(c, bkey_i_to_s_c(&b->key),
258 &max_stale, initial);
259 gc_pos_set(c, gc_pos_btree_root(b->c.btree_id));
260 mutex_unlock(&c->btree_root_lock);
265 static int bch2_gc_btree_init_recurse(struct bch_fs *c, struct btree *b,
266 struct journal_keys *journal_keys,
267 unsigned target_depth)
269 struct btree_and_journal_iter iter;
271 struct bpos next_node_start = b->data->min_key;
276 bch2_btree_and_journal_iter_init_node_iter(&iter, journal_keys, b);
277 bch2_bkey_buf_init(&tmp);
279 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
280 bch2_bkey_debugcheck(c, b, k);
282 BUG_ON(bkey_cmp(k.k->p, b->data->min_key) < 0);
283 BUG_ON(bkey_cmp(k.k->p, b->data->max_key) > 0);
285 ret = bch2_gc_mark_key(c, k, &max_stale, true);
292 bch2_bkey_buf_reassemble(&tmp, c, k);
293 k = bkey_i_to_s_c(tmp.k);
295 bch2_btree_and_journal_iter_advance(&iter);
297 ret = bch2_gc_check_topology(c, k,
300 !bch2_btree_and_journal_iter_peek(&iter).k);
304 if (b->c.level > target_depth) {
305 child = bch2_btree_node_get_noiter(c, tmp.k,
306 b->c.btree_id, b->c.level - 1);
307 ret = PTR_ERR_OR_ZERO(child);
311 ret = bch2_gc_btree_init_recurse(c, child,
312 journal_keys, target_depth);
313 six_unlock_read(&child->c.lock);
319 bch2_btree_and_journal_iter_advance(&iter);
323 bch2_bkey_buf_exit(&tmp, c);
327 static int bch2_gc_btree_init(struct bch_fs *c,
328 struct journal_keys *journal_keys,
329 enum btree_id btree_id,
333 unsigned target_depth = metadata_only ? 1
334 : bch2_expensive_debug_checks ? 0
335 : !btree_node_type_needs_gc(btree_id) ? 1
340 b = c->btree_roots[btree_id].b;
342 if (btree_node_fake(b))
345 six_lock_read(&b->c.lock, NULL, NULL);
346 if (fsck_err_on(bkey_cmp(b->data->min_key, POS_MIN), c,
347 "btree root with incorrect min_key: %llu:%llu",
348 b->data->min_key.inode,
349 b->data->min_key.offset)) {
353 if (fsck_err_on(bkey_cmp(b->data->max_key, POS_MAX), c,
354 "btree root with incorrect min_key: %llu:%llu",
355 b->data->max_key.inode,
356 b->data->max_key.offset)) {
360 if (b->c.level >= target_depth)
361 ret = bch2_gc_btree_init_recurse(c, b,
362 journal_keys, target_depth);
365 ret = bch2_gc_mark_key(c, bkey_i_to_s_c(&b->key),
368 six_unlock_read(&b->c.lock);
373 static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
375 return (int) btree_id_to_gc_phase(l) -
376 (int) btree_id_to_gc_phase(r);
379 static int bch2_gc_btrees(struct bch_fs *c, struct journal_keys *journal_keys,
380 bool initial, bool metadata_only)
382 enum btree_id ids[BTREE_ID_NR];
385 for (i = 0; i < BTREE_ID_NR; i++)
387 bubble_sort(ids, BTREE_ID_NR, btree_id_gc_phase_cmp);
389 for (i = 0; i < BTREE_ID_NR; i++) {
390 enum btree_id id = ids[i];
392 ? bch2_gc_btree_init(c, journal_keys,
394 : bch2_gc_btree(c, id, initial, metadata_only);
402 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
404 enum bch_data_type type,
407 u64 b = sector_to_bucket(ca, start);
411 min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
413 bch2_mark_metadata_bucket(c, ca, b, type, sectors,
414 gc_phase(GC_PHASE_SB), flags);
417 } while (start < end);
420 void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
423 struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
428 * This conditional is kind of gross, but we may be called from the
429 * device add path, before the new device has actually been added to the
430 * running filesystem:
433 lockdep_assert_held(&c->sb_lock);
434 percpu_down_read(&c->mark_lock);
437 for (i = 0; i < layout->nr_superblocks; i++) {
438 u64 offset = le64_to_cpu(layout->sb_offset[i]);
440 if (offset == BCH_SB_SECTOR)
441 mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
444 mark_metadata_sectors(c, ca, offset,
445 offset + (1 << layout->sb_max_size_bits),
449 for (i = 0; i < ca->journal.nr; i++) {
450 b = ca->journal.buckets[i];
451 bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_journal,
453 gc_phase(GC_PHASE_SB), flags);
457 percpu_up_read(&c->mark_lock);
460 static void bch2_mark_superblocks(struct bch_fs *c)
465 mutex_lock(&c->sb_lock);
466 gc_pos_set(c, gc_phase(GC_PHASE_SB));
468 for_each_online_member(ca, c, i)
469 bch2_mark_dev_superblock(c, ca, BTREE_TRIGGER_GC);
470 mutex_unlock(&c->sb_lock);
474 /* Also see bch2_pending_btree_node_free_insert_done() */
475 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
477 struct btree_update *as;
478 struct pending_btree_node_free *d;
480 mutex_lock(&c->btree_interior_update_lock);
481 gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
483 for_each_pending_btree_node_free(c, as, d)
484 if (d->index_update_done)
485 bch2_mark_key(c, bkey_i_to_s_c(&d->key),
489 mutex_unlock(&c->btree_interior_update_lock);
493 static void bch2_mark_allocator_buckets(struct bch_fs *c)
496 struct open_bucket *ob;
500 percpu_down_read(&c->mark_lock);
502 spin_lock(&c->freelist_lock);
503 gc_pos_set(c, gc_pos_alloc(c, NULL));
505 for_each_member_device(ca, c, ci) {
506 fifo_for_each_entry(i, &ca->free_inc, iter)
507 bch2_mark_alloc_bucket(c, ca, i, true,
508 gc_pos_alloc(c, NULL),
513 for (j = 0; j < RESERVE_NR; j++)
514 fifo_for_each_entry(i, &ca->free[j], iter)
515 bch2_mark_alloc_bucket(c, ca, i, true,
516 gc_pos_alloc(c, NULL),
520 spin_unlock(&c->freelist_lock);
522 for (ob = c->open_buckets;
523 ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
525 spin_lock(&ob->lock);
527 gc_pos_set(c, gc_pos_alloc(c, ob));
528 ca = bch_dev_bkey_exists(c, ob->ptr.dev);
529 bch2_mark_alloc_bucket(c, ca, PTR_BUCKET_NR(ca, &ob->ptr), true,
533 spin_unlock(&ob->lock);
536 percpu_up_read(&c->mark_lock);
539 static void bch2_gc_free(struct bch_fs *c)
544 genradix_free(&c->stripes[1]);
546 for_each_member_device(ca, c, i) {
547 kvpfree(rcu_dereference_protected(ca->buckets[1], 1),
548 sizeof(struct bucket_array) +
549 ca->mi.nbuckets * sizeof(struct bucket));
550 ca->buckets[1] = NULL;
552 free_percpu(ca->usage[1]);
556 free_percpu(c->usage_gc);
560 static int bch2_gc_done(struct bch_fs *c,
561 bool initial, bool metadata_only)
564 bool verify = !metadata_only &&
566 (c->sb.compat & (1ULL << BCH_COMPAT_FEAT_ALLOC_INFO)));
570 #define copy_field(_f, _msg, ...) \
571 if (dst->_f != src->_f) { \
573 fsck_err(c, _msg ": got %llu, should be %llu" \
574 , ##__VA_ARGS__, dst->_f, src->_f); \
576 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags); \
578 #define copy_stripe_field(_f, _msg, ...) \
579 if (dst->_f != src->_f) { \
581 fsck_err(c, "stripe %zu has wrong "_msg \
582 ": got %u, should be %u", \
583 dst_iter.pos, ##__VA_ARGS__, \
587 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags); \
589 #define copy_bucket_field(_f) \
590 if (dst->b[b].mark._f != src->b[b].mark._f) { \
592 fsck_err(c, "bucket %u:%zu gen %u data type %s has wrong " #_f \
593 ": got %u, should be %u", i, b, \
594 dst->b[b].mark.gen, \
595 bch2_data_types[dst->b[b].mark.data_type],\
596 dst->b[b].mark._f, src->b[b].mark._f); \
597 dst->b[b]._mark._f = src->b[b].mark._f; \
598 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags); \
600 #define copy_dev_field(_f, _msg, ...) \
601 copy_field(_f, "dev %u has wrong " _msg, i, ##__VA_ARGS__)
602 #define copy_fs_field(_f, _msg, ...) \
603 copy_field(_f, "fs has wrong " _msg, ##__VA_ARGS__)
605 if (!metadata_only) {
606 struct genradix_iter dst_iter = genradix_iter_init(&c->stripes[0], 0);
607 struct genradix_iter src_iter = genradix_iter_init(&c->stripes[1], 0);
608 struct stripe *dst, *src;
610 c->ec_stripes_heap.used = 0;
612 while ((dst = genradix_iter_peek(&dst_iter, &c->stripes[0])) &&
613 (src = genradix_iter_peek(&src_iter, &c->stripes[1]))) {
614 BUG_ON(src_iter.pos != dst_iter.pos);
616 copy_stripe_field(alive, "alive");
617 copy_stripe_field(sectors, "sectors");
618 copy_stripe_field(algorithm, "algorithm");
619 copy_stripe_field(nr_blocks, "nr_blocks");
620 copy_stripe_field(nr_redundant, "nr_redundant");
621 copy_stripe_field(blocks_nonempty,
624 for (i = 0; i < ARRAY_SIZE(dst->block_sectors); i++)
625 copy_stripe_field(block_sectors[i],
626 "block_sectors[%u]", i);
629 spin_lock(&c->ec_stripes_heap_lock);
630 bch2_stripes_heap_insert(c, dst, dst_iter.pos);
631 spin_unlock(&c->ec_stripes_heap_lock);
634 genradix_iter_advance(&dst_iter, &c->stripes[0]);
635 genradix_iter_advance(&src_iter, &c->stripes[1]);
639 for_each_member_device(ca, c, i) {
640 struct bucket_array *dst = __bucket_array(ca, 0);
641 struct bucket_array *src = __bucket_array(ca, 1);
644 for (b = 0; b < src->nbuckets; b++) {
645 copy_bucket_field(gen);
646 copy_bucket_field(data_type);
647 copy_bucket_field(owned_by_allocator);
648 copy_bucket_field(stripe);
649 copy_bucket_field(dirty_sectors);
650 copy_bucket_field(cached_sectors);
652 dst->b[b].oldest_gen = src->b[b].oldest_gen;
656 for (i = 0; i < ARRAY_SIZE(c->usage); i++)
657 bch2_fs_usage_acc_to_base(c, i);
659 bch2_dev_usage_from_buckets(c);
662 unsigned nr = fs_usage_u64s(c);
663 struct bch_fs_usage *dst = c->usage_base;
664 struct bch_fs_usage *src = (void *)
665 bch2_acc_percpu_u64s((void *) c->usage_gc, nr);
667 copy_fs_field(hidden, "hidden");
668 copy_fs_field(btree, "btree");
670 if (!metadata_only) {
671 copy_fs_field(data, "data");
672 copy_fs_field(cached, "cached");
673 copy_fs_field(reserved, "reserved");
674 copy_fs_field(nr_inodes,"nr_inodes");
676 for (i = 0; i < BCH_REPLICAS_MAX; i++)
677 copy_fs_field(persistent_reserved[i],
678 "persistent_reserved[%i]", i);
681 for (i = 0; i < c->replicas.nr; i++) {
682 struct bch_replicas_entry *e =
683 cpu_replicas_entry(&c->replicas, i);
687 (e->data_type == BCH_DATA_user ||
688 e->data_type == BCH_DATA_cached))
691 bch2_replicas_entry_to_text(&PBUF(buf), e);
693 copy_fs_field(replicas[i], "%s", buf);
698 #undef copy_dev_field
699 #undef copy_bucket_field
700 #undef copy_stripe_field
706 static int bch2_gc_start(struct bch_fs *c,
715 c->usage_gc = __alloc_percpu_gfp(fs_usage_u64s(c) * sizeof(u64),
716 sizeof(u64), GFP_KERNEL);
718 bch_err(c, "error allocating c->usage_gc");
722 for_each_member_device(ca, c, i) {
723 BUG_ON(ca->buckets[1]);
724 BUG_ON(ca->usage[1]);
726 ca->buckets[1] = kvpmalloc(sizeof(struct bucket_array) +
727 ca->mi.nbuckets * sizeof(struct bucket),
728 GFP_KERNEL|__GFP_ZERO);
729 if (!ca->buckets[1]) {
730 percpu_ref_put(&ca->ref);
731 bch_err(c, "error allocating ca->buckets[gc]");
735 ca->usage[1] = alloc_percpu(struct bch_dev_usage);
737 bch_err(c, "error allocating ca->usage[gc]");
738 percpu_ref_put(&ca->ref);
743 ret = bch2_ec_mem_alloc(c, true);
745 bch_err(c, "error allocating ec gc mem");
749 percpu_down_write(&c->mark_lock);
752 * indicate to stripe code that we need to allocate for the gc stripes
755 gc_pos_set(c, gc_phase(GC_PHASE_START));
757 for_each_member_device(ca, c, i) {
758 struct bucket_array *dst = __bucket_array(ca, 1);
759 struct bucket_array *src = __bucket_array(ca, 0);
762 dst->first_bucket = src->first_bucket;
763 dst->nbuckets = src->nbuckets;
765 for (b = 0; b < src->nbuckets; b++) {
766 struct bucket *d = &dst->b[b];
767 struct bucket *s = &src->b[b];
769 d->_mark.gen = dst->b[b].oldest_gen = s->mark.gen;
770 d->gen_valid = s->gen_valid;
773 (s->mark.data_type == BCH_DATA_user ||
774 s->mark.data_type == BCH_DATA_cached)) {
776 d->_mark.owned_by_allocator = 0;
781 percpu_up_write(&c->mark_lock);
787 * bch2_gc - walk _all_ references to buckets, and recompute them:
789 * Order matters here:
790 * - Concurrent GC relies on the fact that we have a total ordering for
791 * everything that GC walks - see gc_will_visit_node(),
792 * gc_will_visit_root()
794 * - also, references move around in the course of index updates and
795 * various other crap: everything needs to agree on the ordering
796 * references are allowed to move around in - e.g., we're allowed to
797 * start with a reference owned by an open_bucket (the allocator) and
798 * move it to the btree, but not the reverse.
800 * This is necessary to ensure that gc doesn't miss references that
801 * move around - if references move backwards in the ordering GC
802 * uses, GC could skip past them
804 int bch2_gc(struct bch_fs *c, struct journal_keys *journal_keys,
805 bool initial, bool metadata_only)
808 u64 start_time = local_clock();
809 unsigned i, iter = 0;
812 lockdep_assert_held(&c->state_lock);
815 down_write(&c->gc_lock);
817 /* flush interior btree updates: */
818 closure_wait_event(&c->btree_interior_update_wait,
819 !bch2_btree_interior_updates_nr_pending(c));
821 ret = bch2_gc_start(c, metadata_only);
825 bch2_mark_superblocks(c);
827 ret = bch2_gc_btrees(c, journal_keys, initial, metadata_only);
832 bch2_mark_pending_btree_node_frees(c);
834 bch2_mark_allocator_buckets(c);
839 (test_bit(BCH_FS_FIXED_GENS, &c->flags) ||
840 (!iter && bch2_test_restart_gc))) {
842 * XXX: make sure gens we fixed got saved
845 bch_info(c, "Fixed gens, restarting mark and sweep:");
846 clear_bit(BCH_FS_FIXED_GENS, &c->flags);
847 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
849 percpu_down_write(&c->mark_lock);
851 percpu_up_write(&c->mark_lock);
852 /* flush fsck errors, reset counters */
853 bch2_flush_fsck_errs(c);
858 bch_info(c, "Unable to fix bucket gens, looping");
863 bch2_journal_block(&c->journal);
865 percpu_down_write(&c->mark_lock);
866 ret = bch2_gc_done(c, initial, metadata_only);
868 bch2_journal_unblock(&c->journal);
870 percpu_down_write(&c->mark_lock);
873 /* Indicates that gc is no longer in progress: */
874 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
877 percpu_up_write(&c->mark_lock);
879 up_write(&c->gc_lock);
882 bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
885 * Wake up allocator in case it was waiting for buckets
886 * because of not being able to inc gens
888 for_each_member_device(ca, c, i)
889 bch2_wake_allocator(ca);
892 * At startup, allocations can happen directly instead of via the
893 * allocator thread - issue wakeup in case they blocked on gc_lock:
895 closure_wake_up(&c->freelist_wait);
899 static bool gc_btree_gens_key(struct bch_fs *c, struct bkey_s_c k)
901 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
902 const struct bch_extent_ptr *ptr;
904 percpu_down_read(&c->mark_lock);
905 bkey_for_each_ptr(ptrs, ptr) {
906 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
907 struct bucket *g = PTR_BUCKET(ca, ptr, false);
909 if (gen_after(g->mark.gen, ptr->gen) > 16) {
910 percpu_up_read(&c->mark_lock);
915 bkey_for_each_ptr(ptrs, ptr) {
916 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
917 struct bucket *g = PTR_BUCKET(ca, ptr, false);
919 if (gen_after(g->gc_gen, ptr->gen))
920 g->gc_gen = ptr->gen;
922 percpu_up_read(&c->mark_lock);
928 * For recalculating oldest gen, we only need to walk keys in leaf nodes; btree
929 * node pointers currently never have cached pointers that can become stale:
931 static int bch2_gc_btree_gens(struct bch_fs *c, enum btree_id btree_id)
933 struct btree_trans trans;
934 struct btree_iter *iter;
939 bch2_bkey_buf_init(&sk);
940 bch2_trans_init(&trans, c, 0, 0);
942 iter = bch2_trans_get_iter(&trans, btree_id, POS_MIN,
943 BTREE_ITER_PREFETCH);
945 while ((k = bch2_btree_iter_peek(iter)).k &&
946 !(ret = bkey_err(k))) {
947 if (gc_btree_gens_key(c, k)) {
948 bch2_bkey_buf_reassemble(&sk, c, k);
949 bch2_extent_normalize(c, bkey_i_to_s(sk.k));
951 bch2_btree_iter_set_pos(iter, bkey_start_pos(&sk.k->k));
953 bch2_trans_update(&trans, iter, sk.k, 0);
955 ret = bch2_trans_commit(&trans, NULL, NULL,
956 BTREE_INSERT_NOFAIL);
964 bch2_btree_iter_next(iter);
967 bch2_trans_exit(&trans);
968 bch2_bkey_buf_exit(&sk, c);
973 int bch2_gc_gens(struct bch_fs *c)
976 struct bucket_array *buckets;
982 * Ideally we would be using state_lock and not gc_lock here, but that
983 * introduces a deadlock in the RO path - we currently take the state
984 * lock at the start of going RO, thus the gc thread may get stuck:
986 down_read(&c->gc_lock);
988 for_each_member_device(ca, c, i) {
989 down_read(&ca->bucket_lock);
990 buckets = bucket_array(ca);
992 for_each_bucket(g, buckets)
993 g->gc_gen = g->mark.gen;
994 up_read(&ca->bucket_lock);
997 for (i = 0; i < BTREE_ID_NR; i++)
998 if (btree_node_type_needs_gc(i)) {
999 ret = bch2_gc_btree_gens(c, i);
1001 bch_err(c, "error recalculating oldest_gen: %i", ret);
1006 for_each_member_device(ca, c, i) {
1007 down_read(&ca->bucket_lock);
1008 buckets = bucket_array(ca);
1010 for_each_bucket(g, buckets)
1011 g->oldest_gen = g->gc_gen;
1012 up_read(&ca->bucket_lock);
1017 up_read(&c->gc_lock);
1021 /* Btree coalescing */
1023 static void recalc_packed_keys(struct btree *b)
1025 struct bset *i = btree_bset_first(b);
1026 struct bkey_packed *k;
1028 memset(&b->nr, 0, sizeof(b->nr));
1030 BUG_ON(b->nsets != 1);
1032 vstruct_for_each(i, k)
1033 btree_keys_account_key_add(&b->nr, 0, k);
1036 static void bch2_coalesce_nodes(struct bch_fs *c, struct btree_iter *iter,
1037 struct btree *old_nodes[GC_MERGE_NODES])
1039 struct btree *parent = btree_node_parent(iter, old_nodes[0]);
1040 unsigned i, nr_old_nodes, nr_new_nodes, u64s = 0;
1041 unsigned blocks = btree_blocks(c) * 2 / 3;
1042 struct btree *new_nodes[GC_MERGE_NODES];
1043 struct btree_update *as;
1044 struct keylist keylist;
1045 struct bkey_format_state format_state;
1046 struct bkey_format new_format;
1048 memset(new_nodes, 0, sizeof(new_nodes));
1049 bch2_keylist_init(&keylist, NULL);
1051 /* Count keys that are not deleted */
1052 for (i = 0; i < GC_MERGE_NODES && old_nodes[i]; i++)
1053 u64s += old_nodes[i]->nr.live_u64s;
1055 nr_old_nodes = nr_new_nodes = i;
1057 /* Check if all keys in @old_nodes could fit in one fewer node */
1058 if (nr_old_nodes <= 1 ||
1059 __vstruct_blocks(struct btree_node, c->block_bits,
1060 DIV_ROUND_UP(u64s, nr_old_nodes - 1)) > blocks)
1063 /* Find a format that all keys in @old_nodes can pack into */
1064 bch2_bkey_format_init(&format_state);
1066 for (i = 0; i < nr_old_nodes; i++)
1067 __bch2_btree_calc_format(&format_state, old_nodes[i]);
1069 new_format = bch2_bkey_format_done(&format_state);
1071 /* Check if repacking would make any nodes too big to fit */
1072 for (i = 0; i < nr_old_nodes; i++)
1073 if (!bch2_btree_node_format_fits(c, old_nodes[i], &new_format)) {
1074 trace_btree_gc_coalesce_fail(c,
1075 BTREE_GC_COALESCE_FAIL_FORMAT_FITS);
1079 if (bch2_keylist_realloc(&keylist, NULL, 0,
1080 BKEY_BTREE_PTR_U64s_MAX * nr_old_nodes)) {
1081 trace_btree_gc_coalesce_fail(c,
1082 BTREE_GC_COALESCE_FAIL_KEYLIST_REALLOC);
1086 as = bch2_btree_update_start(iter->trans, iter->btree_id,
1087 btree_update_reserve_required(c, parent) + nr_old_nodes,
1088 BTREE_INSERT_NOFAIL|
1089 BTREE_INSERT_USE_RESERVE,
1092 trace_btree_gc_coalesce_fail(c,
1093 BTREE_GC_COALESCE_FAIL_RESERVE_GET);
1094 bch2_keylist_free(&keylist, NULL);
1098 trace_btree_gc_coalesce(c, old_nodes[0]);
1100 for (i = 0; i < nr_old_nodes; i++)
1101 bch2_btree_interior_update_will_free_node(as, old_nodes[i]);
1103 /* Repack everything with @new_format and sort down to one bset */
1104 for (i = 0; i < nr_old_nodes; i++)
1106 __bch2_btree_node_alloc_replacement(as, old_nodes[i],
1110 * Conceptually we concatenate the nodes together and slice them
1111 * up at different boundaries.
1113 for (i = nr_new_nodes - 1; i > 0; --i) {
1114 struct btree *n1 = new_nodes[i];
1115 struct btree *n2 = new_nodes[i - 1];
1117 struct bset *s1 = btree_bset_first(n1);
1118 struct bset *s2 = btree_bset_first(n2);
1119 struct bkey_packed *k, *last = NULL;
1121 /* Calculate how many keys from @n2 we could fit inside @n1 */
1125 k < vstruct_last(s2) &&
1126 vstruct_blocks_plus(n1->data, c->block_bits,
1127 u64s + k->u64s) <= blocks;
1128 k = bkey_next_skip_noops(k, vstruct_last(s2))) {
1133 if (u64s == le16_to_cpu(s2->u64s)) {
1134 /* n2 fits entirely in n1 */
1135 n1->key.k.p = n1->data->max_key = n2->data->max_key;
1137 memcpy_u64s(vstruct_last(s1),
1139 le16_to_cpu(s2->u64s));
1140 le16_add_cpu(&s1->u64s, le16_to_cpu(s2->u64s));
1142 set_btree_bset_end(n1, n1->set);
1144 six_unlock_write(&n2->c.lock);
1145 bch2_btree_node_free_never_inserted(c, n2);
1146 six_unlock_intent(&n2->c.lock);
1148 memmove(new_nodes + i - 1,
1150 sizeof(new_nodes[0]) * (nr_new_nodes - i));
1151 new_nodes[--nr_new_nodes] = NULL;
1153 /* move part of n2 into n1 */
1154 n1->key.k.p = n1->data->max_key =
1155 bkey_unpack_pos(n1, last);
1157 n2->data->min_key = bkey_successor(n1->data->max_key);
1159 memcpy_u64s(vstruct_last(s1),
1161 le16_add_cpu(&s1->u64s, u64s);
1164 vstruct_idx(s2, u64s),
1165 (le16_to_cpu(s2->u64s) - u64s) * sizeof(u64));
1166 s2->u64s = cpu_to_le16(le16_to_cpu(s2->u64s) - u64s);
1168 set_btree_bset_end(n1, n1->set);
1169 set_btree_bset_end(n2, n2->set);
1173 for (i = 0; i < nr_new_nodes; i++) {
1174 struct btree *n = new_nodes[i];
1176 recalc_packed_keys(n);
1177 btree_node_reset_sib_u64s(n);
1179 bch2_btree_build_aux_trees(n);
1181 bch2_btree_update_add_new_node(as, n);
1182 six_unlock_write(&n->c.lock);
1184 bch2_btree_node_write(c, n, SIX_LOCK_intent);
1188 * The keys for the old nodes get deleted. We don't want to insert keys
1189 * that compare equal to the keys for the new nodes we'll also be
1190 * inserting - we can't because keys on a keylist must be strictly
1191 * greater than the previous keys, and we also don't need to since the
1192 * key for the new node will serve the same purpose (overwriting the key
1193 * for the old node).
1195 for (i = 0; i < nr_old_nodes; i++) {
1196 struct bkey_i delete;
1199 for (j = 0; j < nr_new_nodes; j++)
1200 if (!bkey_cmp(old_nodes[i]->key.k.p,
1201 new_nodes[j]->key.k.p))
1204 bkey_init(&delete.k);
1205 delete.k.p = old_nodes[i]->key.k.p;
1206 bch2_keylist_add_in_order(&keylist, &delete);
1212 * Keys for the new nodes get inserted: bch2_btree_insert_keys() only
1213 * does the lookup once and thus expects the keys to be in sorted order
1214 * so we have to make sure the new keys are correctly ordered with
1215 * respect to the deleted keys added in the previous loop
1217 for (i = 0; i < nr_new_nodes; i++)
1218 bch2_keylist_add_in_order(&keylist, &new_nodes[i]->key);
1220 /* Insert the newly coalesced nodes */
1221 bch2_btree_insert_node(as, parent, iter, &keylist, 0);
1223 BUG_ON(!bch2_keylist_empty(&keylist));
1225 BUG_ON(iter->l[old_nodes[0]->c.level].b != old_nodes[0]);
1227 bch2_btree_iter_node_replace(iter, new_nodes[0]);
1229 for (i = 0; i < nr_new_nodes; i++)
1230 bch2_btree_update_get_open_buckets(as, new_nodes[i]);
1232 /* Free the old nodes and update our sliding window */
1233 for (i = 0; i < nr_old_nodes; i++) {
1234 bch2_btree_node_free_inmem(c, old_nodes[i], iter);
1237 * the index update might have triggered a split, in which case
1238 * the nodes we coalesced - the new nodes we just created -
1239 * might not be sibling nodes anymore - don't add them to the
1240 * sliding window (except the first):
1243 old_nodes[i] = new_nodes[i];
1245 old_nodes[i] = NULL;
1249 for (i = 0; i < nr_new_nodes; i++)
1250 six_unlock_intent(&new_nodes[i]->c.lock);
1252 bch2_btree_update_done(as);
1253 bch2_keylist_free(&keylist, NULL);
1256 static int bch2_coalesce_btree(struct bch_fs *c, enum btree_id btree_id)
1258 struct btree_trans trans;
1259 struct btree_iter *iter;
1261 bool kthread = (current->flags & PF_KTHREAD) != 0;
1264 /* Sliding window of adjacent btree nodes */
1265 struct btree *merge[GC_MERGE_NODES];
1266 u32 lock_seq[GC_MERGE_NODES];
1268 bch2_trans_init(&trans, c, 0, 0);
1271 * XXX: We don't have a good way of positively matching on sibling nodes
1272 * that have the same parent - this code works by handling the cases
1273 * where they might not have the same parent, and is thus fragile. Ugh.
1275 * Perhaps redo this to use multiple linked iterators?
1277 memset(merge, 0, sizeof(merge));
1279 __for_each_btree_node(&trans, iter, btree_id, POS_MIN,
1281 BTREE_ITER_PREFETCH, b) {
1282 memmove(merge + 1, merge,
1283 sizeof(merge) - sizeof(merge[0]));
1284 memmove(lock_seq + 1, lock_seq,
1285 sizeof(lock_seq) - sizeof(lock_seq[0]));
1289 for (i = 1; i < GC_MERGE_NODES; i++) {
1291 !six_relock_intent(&merge[i]->c.lock, lock_seq[i]))
1294 if (merge[i]->c.level != merge[0]->c.level) {
1295 six_unlock_intent(&merge[i]->c.lock);
1299 memset(merge + i, 0, (GC_MERGE_NODES - i) * sizeof(merge[0]));
1301 bch2_coalesce_nodes(c, iter, merge);
1303 for (i = 1; i < GC_MERGE_NODES && merge[i]; i++) {
1304 lock_seq[i] = merge[i]->c.lock.state.seq;
1305 six_unlock_intent(&merge[i]->c.lock);
1308 lock_seq[0] = merge[0]->c.lock.state.seq;
1310 if (kthread && kthread_should_stop()) {
1311 bch2_trans_exit(&trans);
1315 bch2_trans_cond_resched(&trans);
1318 * If the parent node wasn't relocked, it might have been split
1319 * and the nodes in our sliding window might not have the same
1320 * parent anymore - blow away the sliding window:
1322 if (btree_iter_node(iter, iter->level + 1) &&
1323 !btree_node_intent_locked(iter, iter->level + 1))
1324 memset(merge + 1, 0,
1325 (GC_MERGE_NODES - 1) * sizeof(merge[0]));
1327 return bch2_trans_exit(&trans);
1331 * bch_coalesce - coalesce adjacent nodes with low occupancy
1333 void bch2_coalesce(struct bch_fs *c)
1337 down_read(&c->gc_lock);
1338 trace_gc_coalesce_start(c);
1340 for (id = 0; id < BTREE_ID_NR; id++) {
1341 int ret = c->btree_roots[id].b
1342 ? bch2_coalesce_btree(c, id)
1346 if (ret != -ESHUTDOWN)
1347 bch_err(c, "btree coalescing failed: %d", ret);
1352 trace_gc_coalesce_end(c);
1353 up_read(&c->gc_lock);
1356 static int bch2_gc_thread(void *arg)
1358 struct bch_fs *c = arg;
1359 struct io_clock *clock = &c->io_clock[WRITE];
1360 unsigned long last = atomic_long_read(&clock->now);
1361 unsigned last_kick = atomic_read(&c->kick_gc);
1368 set_current_state(TASK_INTERRUPTIBLE);
1370 if (kthread_should_stop()) {
1371 __set_current_state(TASK_RUNNING);
1375 if (atomic_read(&c->kick_gc) != last_kick)
1378 if (c->btree_gc_periodic) {
1379 unsigned long next = last + c->capacity / 16;
1381 if (atomic_long_read(&clock->now) >= next)
1384 bch2_io_clock_schedule_timeout(clock, next);
1391 __set_current_state(TASK_RUNNING);
1393 last = atomic_long_read(&clock->now);
1394 last_kick = atomic_read(&c->kick_gc);
1397 * Full gc is currently incompatible with btree key cache:
1400 ret = bch2_gc(c, NULL, false, false);
1402 ret = bch2_gc_gens(c);
1405 bch_err(c, "btree gc failed: %i", ret);
1407 debug_check_no_locks_held();
1413 void bch2_gc_thread_stop(struct bch_fs *c)
1415 struct task_struct *p;
1418 c->gc_thread = NULL;
1426 int bch2_gc_thread_start(struct bch_fs *c)
1428 struct task_struct *p;
1430 BUG_ON(c->gc_thread);
1432 p = kthread_create(bch2_gc_thread, c, "bch-gc/%s", c->name);