2 * Copyright (C) 2010 Kent Overstreet <kent.overstreet@gmail.com>
3 * Copyright (C) 2014 Datera Inc.
7 #include "alloc_background.h"
8 #include "alloc_foreground.h"
9 #include "bkey_methods.h"
10 #include "btree_locking.h"
11 #include "btree_update_interior.h"
21 #include "journal_io.h"
27 #include <linux/slab.h>
28 #include <linux/bitops.h>
29 #include <linux/freezer.h>
30 #include <linux/kthread.h>
31 #include <linux/preempt.h>
32 #include <linux/rcupdate.h>
33 #include <linux/sched/task.h>
34 #include <trace/events/bcachefs.h>
36 static inline void __gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
38 write_seqcount_begin(&c->gc_pos_lock);
40 write_seqcount_end(&c->gc_pos_lock);
43 static inline void gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
45 BUG_ON(gc_pos_cmp(new_pos, c->gc_pos) <= 0);
46 __gc_pos_set(c, new_pos);
49 /* range_checks - for validating min/max pos of each btree node: */
59 static void btree_node_range_checks_init(struct range_checks *r, unsigned depth)
63 for (i = 0; i < BTREE_MAX_DEPTH; i++)
64 r->l[i].min = r->l[i].max = POS_MIN;
68 static void btree_node_range_checks(struct bch_fs *c, struct btree *b,
69 struct range_checks *r)
71 struct range_level *l = &r->l[b->level];
73 struct bpos expected_min = bkey_cmp(l->min, l->max)
74 ? btree_type_successor(b->btree_id, l->max)
77 bch2_fs_inconsistent_on(bkey_cmp(b->data->min_key, expected_min), c,
78 "btree node has incorrect min key: %llu:%llu != %llu:%llu",
79 b->data->min_key.inode,
80 b->data->min_key.offset,
84 l->max = b->data->max_key;
86 if (b->level > r->depth) {
87 l = &r->l[b->level - 1];
89 bch2_fs_inconsistent_on(bkey_cmp(b->data->min_key, l->min), c,
90 "btree node min doesn't match min of child nodes: %llu:%llu != %llu:%llu",
91 b->data->min_key.inode,
92 b->data->min_key.offset,
96 bch2_fs_inconsistent_on(bkey_cmp(b->data->max_key, l->max), c,
97 "btree node max doesn't match max of child nodes: %llu:%llu != %llu:%llu",
98 b->data->max_key.inode,
99 b->data->max_key.offset,
103 if (bkey_cmp(b->data->max_key, POS_MAX))
105 btree_type_successor(b->btree_id,
110 /* marking of btree keys/nodes: */
112 static int bch2_gc_mark_key(struct bch_fs *c, struct bkey_s_c k,
113 u8 *max_stale, bool initial)
115 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
116 const struct bch_extent_ptr *ptr;
119 (initial ? BCH_BUCKET_MARK_NOATOMIC : 0);
123 BUG_ON(journal_seq_verify(c) &&
124 k.k->version.lo > journal_cur_seq(&c->journal));
126 if (k.k->version.lo > atomic64_read(&c->key_version))
127 atomic64_set(&c->key_version, k.k->version.lo);
129 if (test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) ||
130 fsck_err_on(!bch2_bkey_replicas_marked(c, k, false), c,
131 "superblock not marked as containing replicas (type %u)",
133 ret = bch2_mark_bkey_replicas(c, k);
138 bkey_for_each_ptr(ptrs, ptr) {
139 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
140 struct bucket *g = PTR_BUCKET(ca, ptr, true);
141 struct bucket *g2 = PTR_BUCKET(ca, ptr, false);
143 if (mustfix_fsck_err_on(!g->gen_valid, c,
144 "found ptr with missing gen in alloc btree,\n"
146 k.k->type, ptr->gen)) {
147 g2->_mark.gen = g->_mark.gen = ptr->gen;
148 g2->_mark.dirty = g->_mark.dirty = true;
149 g2->gen_valid = g->gen_valid = true;
152 if (mustfix_fsck_err_on(gen_cmp(ptr->gen, g->mark.gen) > 0, c,
153 "%u ptr gen in the future: %u > %u",
154 k.k->type, ptr->gen, g->mark.gen)) {
155 g2->_mark.gen = g->_mark.gen = ptr->gen;
156 g2->_mark.dirty = g->_mark.dirty = true;
157 g2->gen_valid = g->gen_valid = true;
158 set_bit(BCH_FS_FIXED_GENS, &c->flags);
163 bkey_for_each_ptr(ptrs, ptr) {
164 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
165 struct bucket *g = PTR_BUCKET(ca, ptr, true);
167 if (gen_after(g->oldest_gen, ptr->gen))
168 g->oldest_gen = ptr->gen;
170 *max_stale = max(*max_stale, ptr_stale(ca, ptr));
173 bch2_mark_key(c, k, true, k.k->size, NULL, 0, flags);
178 static int btree_gc_mark_node(struct bch_fs *c, struct btree *b,
179 u8 *max_stale, bool initial)
181 struct btree_node_iter iter;
182 struct bkey unpacked;
188 if (!btree_node_type_needs_gc(btree_node_type(b)))
191 for_each_btree_node_key_unpack(b, k, &iter,
193 bch2_bkey_debugcheck(c, b, k);
195 ret = bch2_gc_mark_key(c, k, max_stale, initial);
203 static int bch2_gc_btree(struct bch_fs *c, enum btree_id btree_id,
204 bool initial, bool metadata_only)
206 struct btree_trans trans;
207 struct btree_iter *iter;
209 struct range_checks r;
210 unsigned depth = btree_node_type_needs_gc(btree_id) ? 0 : 1;
214 bch2_trans_init(&trans, c);
216 gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));
219 * if expensive_debug_checks is on, run range_checks on all leaf nodes:
221 * and on startup, we have to read every btree node (XXX: only if it was
222 * an unclean shutdown)
226 else if (initial || expensive_debug_checks(c))
229 btree_node_range_checks_init(&r, depth);
231 __for_each_btree_node(&trans, iter, btree_id, POS_MIN,
232 0, depth, BTREE_ITER_PREFETCH, b) {
233 btree_node_range_checks(c, b, &r);
235 bch2_verify_btree_nr_keys(b);
237 gc_pos_set(c, gc_pos_btree_node(b));
239 ret = btree_gc_mark_node(c, b, &max_stale, initial);
245 bch2_btree_node_rewrite(c, iter,
247 BTREE_INSERT_USE_RESERVE|
249 BTREE_INSERT_GC_LOCK_HELD);
250 else if (!btree_gc_rewrite_disabled(c) &&
251 (btree_gc_always_rewrite(c) || max_stale > 16))
252 bch2_btree_node_rewrite(c, iter,
255 BTREE_INSERT_GC_LOCK_HELD);
258 bch2_trans_cond_resched(&trans);
260 ret = bch2_trans_exit(&trans) ?: ret;
264 mutex_lock(&c->btree_root_lock);
265 b = c->btree_roots[btree_id].b;
266 if (!btree_node_fake(b))
267 ret = bch2_gc_mark_key(c, bkey_i_to_s_c(&b->key),
268 &max_stale, initial);
269 gc_pos_set(c, gc_pos_btree_root(b->btree_id));
270 mutex_unlock(&c->btree_root_lock);
275 static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
277 return (int) btree_id_to_gc_phase(l) -
278 (int) btree_id_to_gc_phase(r);
281 static int bch2_gc_btrees(struct bch_fs *c, struct list_head *journal,
282 bool initial, bool metadata_only)
284 enum btree_id ids[BTREE_ID_NR];
288 for (i = 0; i < BTREE_ID_NR; i++)
290 bubble_sort(ids, BTREE_ID_NR, btree_id_gc_phase_cmp);
292 for (i = 0; i < BTREE_ID_NR; i++) {
293 enum btree_id id = ids[i];
294 enum btree_node_type type = __btree_node_type(0, id);
296 int ret = bch2_gc_btree(c, id, initial, metadata_only);
300 if (journal && !metadata_only &&
301 btree_node_type_needs_gc(type)) {
302 struct bkey_i *k, *n;
303 struct jset_entry *j;
304 struct journal_replay *r;
307 list_for_each_entry(r, journal, list)
308 for_each_jset_key(k, n, j, &r->j) {
309 if (type == __btree_node_type(j->level, j->btree_id)) {
310 ret = bch2_gc_mark_key(c,
312 &max_stale, initial);
323 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
325 enum bch_data_type type,
328 u64 b = sector_to_bucket(ca, start);
332 min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
334 bch2_mark_metadata_bucket(c, ca, b, type, sectors,
335 gc_phase(GC_PHASE_SB), flags);
338 } while (start < end);
341 void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
344 struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
349 * This conditional is kind of gross, but we may be called from the
350 * device add path, before the new device has actually been added to the
351 * running filesystem:
354 lockdep_assert_held(&c->sb_lock);
355 percpu_down_read_preempt_disable(&c->mark_lock);
360 for (i = 0; i < layout->nr_superblocks; i++) {
361 u64 offset = le64_to_cpu(layout->sb_offset[i]);
363 if (offset == BCH_SB_SECTOR)
364 mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
367 mark_metadata_sectors(c, ca, offset,
368 offset + (1 << layout->sb_max_size_bits),
372 for (i = 0; i < ca->journal.nr; i++) {
373 b = ca->journal.buckets[i];
374 bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_JOURNAL,
376 gc_phase(GC_PHASE_SB), flags);
380 percpu_up_read_preempt_enable(&c->mark_lock);
386 static void bch2_mark_superblocks(struct bch_fs *c)
391 mutex_lock(&c->sb_lock);
392 gc_pos_set(c, gc_phase(GC_PHASE_SB));
394 for_each_online_member(ca, c, i)
395 bch2_mark_dev_superblock(c, ca, BCH_BUCKET_MARK_GC);
396 mutex_unlock(&c->sb_lock);
399 /* Also see bch2_pending_btree_node_free_insert_done() */
400 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
402 struct btree_update *as;
403 struct pending_btree_node_free *d;
405 mutex_lock(&c->btree_interior_update_lock);
406 gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
408 for_each_pending_btree_node_free(c, as, d)
409 if (d->index_update_done)
410 bch2_mark_key(c, bkey_i_to_s_c(&d->key),
414 mutex_unlock(&c->btree_interior_update_lock);
417 static void bch2_mark_allocator_buckets(struct bch_fs *c)
420 struct open_bucket *ob;
424 percpu_down_read_preempt_disable(&c->mark_lock);
426 spin_lock(&c->freelist_lock);
427 gc_pos_set(c, gc_pos_alloc(c, NULL));
429 for_each_member_device(ca, c, ci) {
430 fifo_for_each_entry(i, &ca->free_inc, iter)
431 bch2_mark_alloc_bucket(c, ca, i, true,
432 gc_pos_alloc(c, NULL),
437 for (j = 0; j < RESERVE_NR; j++)
438 fifo_for_each_entry(i, &ca->free[j], iter)
439 bch2_mark_alloc_bucket(c, ca, i, true,
440 gc_pos_alloc(c, NULL),
444 spin_unlock(&c->freelist_lock);
446 for (ob = c->open_buckets;
447 ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
449 spin_lock(&ob->lock);
451 gc_pos_set(c, gc_pos_alloc(c, ob));
452 ca = bch_dev_bkey_exists(c, ob->ptr.dev);
453 bch2_mark_alloc_bucket(c, ca, PTR_BUCKET_NR(ca, &ob->ptr), true,
457 spin_unlock(&ob->lock);
460 percpu_up_read_preempt_enable(&c->mark_lock);
463 static void bch2_gc_free(struct bch_fs *c)
468 genradix_free(&c->stripes[1]);
470 for_each_member_device(ca, c, i) {
471 kvpfree(rcu_dereference_protected(ca->buckets[1], 1),
472 sizeof(struct bucket_array) +
473 ca->mi.nbuckets * sizeof(struct bucket));
474 ca->buckets[1] = NULL;
476 free_percpu(ca->usage[1]);
480 free_percpu(c->usage[1]);
484 static int bch2_gc_done(struct bch_fs *c,
485 bool initial, bool metadata_only)
488 bool verify = !metadata_only &&
490 (c->sb.compat & (1ULL << BCH_COMPAT_FEAT_ALLOC_INFO)));
494 #define copy_field(_f, _msg, ...) \
495 if (dst->_f != src->_f) { \
497 fsck_err(c, _msg ": got %llu, should be %llu" \
498 , ##__VA_ARGS__, dst->_f, src->_f); \
501 #define copy_stripe_field(_f, _msg, ...) \
502 if (dst->_f != src->_f) { \
504 fsck_err(c, "stripe %zu has wrong "_msg \
505 ": got %u, should be %u", \
506 dst_iter.pos, ##__VA_ARGS__, \
511 #define copy_bucket_field(_f) \
512 if (dst->b[b].mark._f != src->b[b].mark._f) { \
514 fsck_err(c, "dev %u bucket %zu has wrong " #_f \
515 ": got %u, should be %u", i, b, \
516 dst->b[b].mark._f, src->b[b].mark._f); \
517 dst->b[b]._mark._f = src->b[b].mark._f; \
518 dst->b[b]._mark.dirty = true; \
520 #define copy_dev_field(_f, _msg, ...) \
521 copy_field(_f, "dev %u has wrong " _msg, i, ##__VA_ARGS__)
522 #define copy_fs_field(_f, _msg, ...) \
523 copy_field(_f, "fs has wrong " _msg, ##__VA_ARGS__)
525 if (!metadata_only) {
526 struct genradix_iter dst_iter = genradix_iter_init(&c->stripes[0], 0);
527 struct genradix_iter src_iter = genradix_iter_init(&c->stripes[1], 0);
528 struct stripe *dst, *src;
531 c->ec_stripes_heap.used = 0;
533 while ((dst = genradix_iter_peek(&dst_iter, &c->stripes[0])) &&
534 (src = genradix_iter_peek(&src_iter, &c->stripes[1]))) {
535 BUG_ON(src_iter.pos != dst_iter.pos);
537 copy_stripe_field(alive, "alive");
538 copy_stripe_field(sectors, "sectors");
539 copy_stripe_field(algorithm, "algorithm");
540 copy_stripe_field(nr_blocks, "nr_blocks");
541 copy_stripe_field(nr_redundant, "nr_redundant");
542 copy_stripe_field(blocks_nonempty,
545 for (i = 0; i < ARRAY_SIZE(dst->block_sectors); i++)
546 copy_stripe_field(block_sectors[i],
547 "block_sectors[%u]", i);
550 bch2_stripes_heap_insert(c, dst, dst_iter.pos);
552 genradix_iter_advance(&dst_iter, &c->stripes[0]);
553 genradix_iter_advance(&src_iter, &c->stripes[1]);
557 for_each_member_device(ca, c, i) {
558 struct bucket_array *dst = __bucket_array(ca, 0);
559 struct bucket_array *src = __bucket_array(ca, 1);
562 for (b = 0; b < src->nbuckets; b++) {
563 copy_bucket_field(gen);
564 copy_bucket_field(data_type);
565 copy_bucket_field(owned_by_allocator);
566 copy_bucket_field(stripe);
567 copy_bucket_field(dirty_sectors);
568 copy_bucket_field(cached_sectors);
570 if (dst->b[b].oldest_gen != src->b[b].oldest_gen) {
571 dst->b[b].oldest_gen = src->b[b].oldest_gen;
572 dst->b[b]._mark.dirty = true;
577 bch2_dev_usage_from_buckets(c);
580 unsigned nr = fs_usage_u64s(c);
581 struct bch_fs_usage *dst = (void *)
582 bch2_acc_percpu_u64s((void *) c->usage[0], nr);
583 struct bch_fs_usage *src = (void *)
584 bch2_acc_percpu_u64s((void *) c->usage[1], nr);
586 copy_fs_field(hidden, "hidden");
587 copy_fs_field(btree, "btree");
589 if (!metadata_only) {
590 copy_fs_field(data, "data");
591 copy_fs_field(cached, "cached");
592 copy_fs_field(reserved, "reserved");
593 copy_fs_field(nr_inodes,"nr_inodes");
595 for (i = 0; i < BCH_REPLICAS_MAX; i++)
596 copy_fs_field(persistent_reserved[i],
597 "persistent_reserved[%i]", i);
600 for (i = 0; i < c->replicas.nr; i++) {
601 struct bch_replicas_entry *e =
602 cpu_replicas_entry(&c->replicas, i);
606 (e->data_type == BCH_DATA_USER ||
607 e->data_type == BCH_DATA_CACHED))
610 bch2_replicas_entry_to_text(&PBUF(buf), e);
612 copy_fs_field(replicas[i], "%s", buf);
617 #undef copy_dev_field
618 #undef copy_bucket_field
619 #undef copy_stripe_field
625 static int bch2_gc_start(struct bch_fs *c,
632 * indicate to stripe code that we need to allocate for the gc stripes
635 gc_pos_set(c, gc_phase(GC_PHASE_START));
639 c->usage[1] = __alloc_percpu_gfp(fs_usage_u64s(c) * sizeof(u64),
640 sizeof(u64), GFP_KERNEL);
644 for_each_member_device(ca, c, i) {
645 BUG_ON(ca->buckets[1]);
646 BUG_ON(ca->usage[1]);
648 ca->buckets[1] = kvpmalloc(sizeof(struct bucket_array) +
649 ca->mi.nbuckets * sizeof(struct bucket),
650 GFP_KERNEL|__GFP_ZERO);
651 if (!ca->buckets[1]) {
652 percpu_ref_put(&ca->ref);
656 ca->usage[1] = alloc_percpu(struct bch_dev_usage);
658 percpu_ref_put(&ca->ref);
663 for_each_member_device(ca, c, i) {
664 struct bucket_array *dst = __bucket_array(ca, 1);
665 struct bucket_array *src = __bucket_array(ca, 0);
668 dst->first_bucket = src->first_bucket;
669 dst->nbuckets = src->nbuckets;
671 for (b = 0; b < src->nbuckets; b++) {
672 struct bucket *d = &dst->b[b];
673 struct bucket *s = &src->b[b];
675 d->_mark.gen = dst->b[b].oldest_gen = s->mark.gen;
676 d->gen_valid = s->gen_valid;
679 (s->mark.data_type == BCH_DATA_USER ||
680 s->mark.data_type == BCH_DATA_CACHED)) {
682 d->_mark.owned_by_allocator = 0;
687 return bch2_ec_mem_alloc(c, true);
691 * bch2_gc - walk _all_ references to buckets, and recompute them:
693 * Order matters here:
694 * - Concurrent GC relies on the fact that we have a total ordering for
695 * everything that GC walks - see gc_will_visit_node(),
696 * gc_will_visit_root()
698 * - also, references move around in the course of index updates and
699 * various other crap: everything needs to agree on the ordering
700 * references are allowed to move around in - e.g., we're allowed to
701 * start with a reference owned by an open_bucket (the allocator) and
702 * move it to the btree, but not the reverse.
704 * This is necessary to ensure that gc doesn't miss references that
705 * move around - if references move backwards in the ordering GC
706 * uses, GC could skip past them
708 int bch2_gc(struct bch_fs *c, struct list_head *journal,
709 bool initial, bool metadata_only)
712 u64 start_time = local_clock();
713 unsigned i, iter = 0;
718 down_write(&c->gc_lock);
720 percpu_down_write(&c->mark_lock);
721 ret = bch2_gc_start(c, metadata_only);
722 percpu_up_write(&c->mark_lock);
727 bch2_mark_superblocks(c);
729 ret = bch2_gc_btrees(c, journal, initial, metadata_only);
733 bch2_mark_pending_btree_node_frees(c);
734 bch2_mark_allocator_buckets(c);
739 (test_bit(BCH_FS_FIXED_GENS, &c->flags) ||
740 (!iter && test_restart_gc(c)))) {
742 * XXX: make sure gens we fixed got saved
745 bch_info(c, "Fixed gens, restarting mark and sweep:");
746 clear_bit(BCH_FS_FIXED_GENS, &c->flags);
747 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
749 percpu_down_write(&c->mark_lock);
751 percpu_up_write(&c->mark_lock);
756 bch_info(c, "Unable to fix bucket gens, looping");
760 percpu_down_write(&c->mark_lock);
763 ret = bch2_gc_done(c, initial, metadata_only);
765 /* Indicates that gc is no longer in progress: */
766 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
769 percpu_up_write(&c->mark_lock);
771 up_write(&c->gc_lock);
774 bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
777 * Wake up allocator in case it was waiting for buckets
778 * because of not being able to inc gens
780 for_each_member_device(ca, c, i)
781 bch2_wake_allocator(ca);
784 * At startup, allocations can happen directly instead of via the
785 * allocator thread - issue wakeup in case they blocked on gc_lock:
787 closure_wake_up(&c->freelist_wait);
791 /* Btree coalescing */
793 static void recalc_packed_keys(struct btree *b)
795 struct bset *i = btree_bset_first(b);
796 struct bkey_packed *k;
798 memset(&b->nr, 0, sizeof(b->nr));
800 BUG_ON(b->nsets != 1);
802 vstruct_for_each(i, k)
803 btree_keys_account_key_add(&b->nr, 0, k);
806 static void bch2_coalesce_nodes(struct bch_fs *c, struct btree_iter *iter,
807 struct btree *old_nodes[GC_MERGE_NODES])
809 struct btree *parent = btree_node_parent(iter, old_nodes[0]);
810 unsigned i, nr_old_nodes, nr_new_nodes, u64s = 0;
811 unsigned blocks = btree_blocks(c) * 2 / 3;
812 struct btree *new_nodes[GC_MERGE_NODES];
813 struct btree_update *as;
814 struct keylist keylist;
815 struct bkey_format_state format_state;
816 struct bkey_format new_format;
818 memset(new_nodes, 0, sizeof(new_nodes));
819 bch2_keylist_init(&keylist, NULL);
821 /* Count keys that are not deleted */
822 for (i = 0; i < GC_MERGE_NODES && old_nodes[i]; i++)
823 u64s += old_nodes[i]->nr.live_u64s;
825 nr_old_nodes = nr_new_nodes = i;
827 /* Check if all keys in @old_nodes could fit in one fewer node */
828 if (nr_old_nodes <= 1 ||
829 __vstruct_blocks(struct btree_node, c->block_bits,
830 DIV_ROUND_UP(u64s, nr_old_nodes - 1)) > blocks)
833 /* Find a format that all keys in @old_nodes can pack into */
834 bch2_bkey_format_init(&format_state);
836 for (i = 0; i < nr_old_nodes; i++)
837 __bch2_btree_calc_format(&format_state, old_nodes[i]);
839 new_format = bch2_bkey_format_done(&format_state);
841 /* Check if repacking would make any nodes too big to fit */
842 for (i = 0; i < nr_old_nodes; i++)
843 if (!bch2_btree_node_format_fits(c, old_nodes[i], &new_format)) {
844 trace_btree_gc_coalesce_fail(c,
845 BTREE_GC_COALESCE_FAIL_FORMAT_FITS);
849 if (bch2_keylist_realloc(&keylist, NULL, 0,
850 (BKEY_U64s + BKEY_EXTENT_U64s_MAX) * nr_old_nodes)) {
851 trace_btree_gc_coalesce_fail(c,
852 BTREE_GC_COALESCE_FAIL_KEYLIST_REALLOC);
856 as = bch2_btree_update_start(c, iter->btree_id,
857 btree_update_reserve_required(c, parent) + nr_old_nodes,
859 BTREE_INSERT_USE_RESERVE,
862 trace_btree_gc_coalesce_fail(c,
863 BTREE_GC_COALESCE_FAIL_RESERVE_GET);
864 bch2_keylist_free(&keylist, NULL);
868 trace_btree_gc_coalesce(c, old_nodes[0]);
870 for (i = 0; i < nr_old_nodes; i++)
871 bch2_btree_interior_update_will_free_node(as, old_nodes[i]);
873 /* Repack everything with @new_format and sort down to one bset */
874 for (i = 0; i < nr_old_nodes; i++)
876 __bch2_btree_node_alloc_replacement(as, old_nodes[i],
880 * Conceptually we concatenate the nodes together and slice them
881 * up at different boundaries.
883 for (i = nr_new_nodes - 1; i > 0; --i) {
884 struct btree *n1 = new_nodes[i];
885 struct btree *n2 = new_nodes[i - 1];
887 struct bset *s1 = btree_bset_first(n1);
888 struct bset *s2 = btree_bset_first(n2);
889 struct bkey_packed *k, *last = NULL;
891 /* Calculate how many keys from @n2 we could fit inside @n1 */
895 k < vstruct_last(s2) &&
896 vstruct_blocks_plus(n1->data, c->block_bits,
897 u64s + k->u64s) <= blocks;
903 if (u64s == le16_to_cpu(s2->u64s)) {
904 /* n2 fits entirely in n1 */
905 n1->key.k.p = n1->data->max_key = n2->data->max_key;
907 memcpy_u64s(vstruct_last(s1),
909 le16_to_cpu(s2->u64s));
910 le16_add_cpu(&s1->u64s, le16_to_cpu(s2->u64s));
912 set_btree_bset_end(n1, n1->set);
914 six_unlock_write(&n2->lock);
915 bch2_btree_node_free_never_inserted(c, n2);
916 six_unlock_intent(&n2->lock);
918 memmove(new_nodes + i - 1,
920 sizeof(new_nodes[0]) * (nr_new_nodes - i));
921 new_nodes[--nr_new_nodes] = NULL;
923 /* move part of n2 into n1 */
924 n1->key.k.p = n1->data->max_key =
925 bkey_unpack_pos(n1, last);
928 btree_type_successor(iter->btree_id,
931 memcpy_u64s(vstruct_last(s1),
933 le16_add_cpu(&s1->u64s, u64s);
936 vstruct_idx(s2, u64s),
937 (le16_to_cpu(s2->u64s) - u64s) * sizeof(u64));
938 s2->u64s = cpu_to_le16(le16_to_cpu(s2->u64s) - u64s);
940 set_btree_bset_end(n1, n1->set);
941 set_btree_bset_end(n2, n2->set);
945 for (i = 0; i < nr_new_nodes; i++) {
946 struct btree *n = new_nodes[i];
948 recalc_packed_keys(n);
949 btree_node_reset_sib_u64s(n);
951 bch2_btree_build_aux_trees(n);
952 six_unlock_write(&n->lock);
954 bch2_btree_node_write(c, n, SIX_LOCK_intent);
958 * The keys for the old nodes get deleted. We don't want to insert keys
959 * that compare equal to the keys for the new nodes we'll also be
960 * inserting - we can't because keys on a keylist must be strictly
961 * greater than the previous keys, and we also don't need to since the
962 * key for the new node will serve the same purpose (overwriting the key
965 for (i = 0; i < nr_old_nodes; i++) {
966 struct bkey_i delete;
969 for (j = 0; j < nr_new_nodes; j++)
970 if (!bkey_cmp(old_nodes[i]->key.k.p,
971 new_nodes[j]->key.k.p))
974 bkey_init(&delete.k);
975 delete.k.p = old_nodes[i]->key.k.p;
976 bch2_keylist_add_in_order(&keylist, &delete);
982 * Keys for the new nodes get inserted: bch2_btree_insert_keys() only
983 * does the lookup once and thus expects the keys to be in sorted order
984 * so we have to make sure the new keys are correctly ordered with
985 * respect to the deleted keys added in the previous loop
987 for (i = 0; i < nr_new_nodes; i++)
988 bch2_keylist_add_in_order(&keylist, &new_nodes[i]->key);
990 /* Insert the newly coalesced nodes */
991 bch2_btree_insert_node(as, parent, iter, &keylist, 0);
993 BUG_ON(!bch2_keylist_empty(&keylist));
995 BUG_ON(iter->l[old_nodes[0]->level].b != old_nodes[0]);
997 bch2_btree_iter_node_replace(iter, new_nodes[0]);
999 for (i = 0; i < nr_new_nodes; i++)
1000 bch2_open_buckets_put(c, &new_nodes[i]->ob);
1002 /* Free the old nodes and update our sliding window */
1003 for (i = 0; i < nr_old_nodes; i++) {
1004 bch2_btree_node_free_inmem(c, old_nodes[i], iter);
1007 * the index update might have triggered a split, in which case
1008 * the nodes we coalesced - the new nodes we just created -
1009 * might not be sibling nodes anymore - don't add them to the
1010 * sliding window (except the first):
1013 old_nodes[i] = new_nodes[i];
1015 old_nodes[i] = NULL;
1017 six_unlock_intent(&new_nodes[i]->lock);
1021 bch2_btree_update_done(as);
1022 bch2_keylist_free(&keylist, NULL);
1025 static int bch2_coalesce_btree(struct bch_fs *c, enum btree_id btree_id)
1027 struct btree_trans trans;
1028 struct btree_iter *iter;
1030 bool kthread = (current->flags & PF_KTHREAD) != 0;
1033 /* Sliding window of adjacent btree nodes */
1034 struct btree *merge[GC_MERGE_NODES];
1035 u32 lock_seq[GC_MERGE_NODES];
1037 bch2_trans_init(&trans, c);
1040 * XXX: We don't have a good way of positively matching on sibling nodes
1041 * that have the same parent - this code works by handling the cases
1042 * where they might not have the same parent, and is thus fragile. Ugh.
1044 * Perhaps redo this to use multiple linked iterators?
1046 memset(merge, 0, sizeof(merge));
1048 __for_each_btree_node(&trans, iter, btree_id, POS_MIN,
1050 BTREE_ITER_PREFETCH, b) {
1051 memmove(merge + 1, merge,
1052 sizeof(merge) - sizeof(merge[0]));
1053 memmove(lock_seq + 1, lock_seq,
1054 sizeof(lock_seq) - sizeof(lock_seq[0]));
1058 for (i = 1; i < GC_MERGE_NODES; i++) {
1060 !six_relock_intent(&merge[i]->lock, lock_seq[i]))
1063 if (merge[i]->level != merge[0]->level) {
1064 six_unlock_intent(&merge[i]->lock);
1068 memset(merge + i, 0, (GC_MERGE_NODES - i) * sizeof(merge[0]));
1070 bch2_coalesce_nodes(c, iter, merge);
1072 for (i = 1; i < GC_MERGE_NODES && merge[i]; i++) {
1073 lock_seq[i] = merge[i]->lock.state.seq;
1074 six_unlock_intent(&merge[i]->lock);
1077 lock_seq[0] = merge[0]->lock.state.seq;
1079 if (kthread && kthread_should_stop()) {
1080 bch2_trans_exit(&trans);
1084 bch2_trans_cond_resched(&trans);
1087 * If the parent node wasn't relocked, it might have been split
1088 * and the nodes in our sliding window might not have the same
1089 * parent anymore - blow away the sliding window:
1091 if (btree_iter_node(iter, iter->level + 1) &&
1092 !btree_node_intent_locked(iter, iter->level + 1))
1093 memset(merge + 1, 0,
1094 (GC_MERGE_NODES - 1) * sizeof(merge[0]));
1096 return bch2_trans_exit(&trans);
1100 * bch_coalesce - coalesce adjacent nodes with low occupancy
1102 void bch2_coalesce(struct bch_fs *c)
1106 down_read(&c->gc_lock);
1107 trace_gc_coalesce_start(c);
1109 for (id = 0; id < BTREE_ID_NR; id++) {
1110 int ret = c->btree_roots[id].b
1111 ? bch2_coalesce_btree(c, id)
1115 if (ret != -ESHUTDOWN)
1116 bch_err(c, "btree coalescing failed: %d", ret);
1121 trace_gc_coalesce_end(c);
1122 up_read(&c->gc_lock);
1125 static int bch2_gc_thread(void *arg)
1127 struct bch_fs *c = arg;
1128 struct io_clock *clock = &c->io_clock[WRITE];
1129 unsigned long last = atomic_long_read(&clock->now);
1130 unsigned last_kick = atomic_read(&c->kick_gc);
1137 set_current_state(TASK_INTERRUPTIBLE);
1139 if (kthread_should_stop()) {
1140 __set_current_state(TASK_RUNNING);
1144 if (atomic_read(&c->kick_gc) != last_kick)
1147 if (c->btree_gc_periodic) {
1148 unsigned long next = last + c->capacity / 16;
1150 if (atomic_long_read(&clock->now) >= next)
1153 bch2_io_clock_schedule_timeout(clock, next);
1160 __set_current_state(TASK_RUNNING);
1162 last = atomic_long_read(&clock->now);
1163 last_kick = atomic_read(&c->kick_gc);
1165 ret = bch2_gc(c, NULL, false, false);
1167 bch_err(c, "btree gc failed: %i", ret);
1169 debug_check_no_locks_held();
1175 void bch2_gc_thread_stop(struct bch_fs *c)
1177 struct task_struct *p;
1180 c->gc_thread = NULL;
1188 int bch2_gc_thread_start(struct bch_fs *c)
1190 struct task_struct *p;
1192 BUG_ON(c->gc_thread);
1194 p = kthread_create(bch2_gc_thread, c, "bch_gc");