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"
29 #include <linux/slab.h>
30 #include <linux/bitops.h>
31 #include <linux/freezer.h>
32 #include <linux/kthread.h>
33 #include <linux/preempt.h>
34 #include <linux/rcupdate.h>
35 #include <linux/sched/task.h>
36 #include <trace/events/bcachefs.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);
54 * Missing: if an interior btree node is empty, we need to do something -
55 * perhaps just kill it
57 static int bch2_gc_check_topology(struct bch_fs *c,
59 struct bkey_buf *prev,
63 struct bpos node_start = b->data->min_key;
64 struct bpos node_end = b->data->max_key;
65 struct bpos expected_start = bkey_deleted(&prev->k->k)
67 : bpos_successor(prev->k->k.p);
68 char buf1[200], buf2[200];
69 bool update_min = false;
70 bool update_max = false;
73 if (cur.k->k.type == KEY_TYPE_btree_ptr_v2) {
74 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(cur.k);
76 if (bkey_deleted(&prev->k->k)) {
77 struct printbuf out = PBUF(buf1);
78 pr_buf(&out, "start of node: ");
79 bch2_bpos_to_text(&out, node_start);
81 bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(prev->k));
84 if (fsck_err_on(bpos_cmp(expected_start, bp->v.min_key), c,
85 "btree node with incorrect min_key at btree %s level %u:\n"
88 bch2_btree_ids[b->c.btree_id], b->c.level,
90 (bch2_bkey_val_to_text(&PBUF(buf2), c, bkey_i_to_s_c(cur.k)), buf2)))
94 if (fsck_err_on(is_last &&
95 bpos_cmp(cur.k->k.p, node_end), c,
96 "btree node with incorrect max_key at btree %s level %u:\n"
99 bch2_btree_ids[b->c.btree_id], b->c.level,
100 (bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(cur.k)), buf1),
101 (bch2_bpos_to_text(&PBUF(buf2), node_end), buf2)))
104 bch2_bkey_buf_copy(prev, c, cur.k);
106 if (update_min || update_max) {
108 struct bkey_i_btree_ptr_v2 *bp = NULL;
112 ret = bch2_journal_key_delete(c, b->c.btree_id,
113 b->c.level, cur.k->k.p);
118 new = kmalloc(bkey_bytes(&cur.k->k), GFP_KERNEL);
120 bch_err(c, "%s: error allocating new key", __func__);
124 bkey_copy(new, cur.k);
126 if (new->k.type == KEY_TYPE_btree_ptr_v2)
127 bp = bkey_i_to_btree_ptr_v2(new);
130 bp->v.min_key = expected_start;
134 SET_BTREE_PTR_RANGE_UPDATED(&bp->v, true);
136 ret = bch2_journal_key_insert(c, b->c.btree_id, b->c.level, new);
142 n = bch2_btree_node_get_noiter(c, cur.k, b->c.btree_id,
143 b->c.level - 1, true);
145 mutex_lock(&c->btree_cache.lock);
146 bch2_btree_node_hash_remove(&c->btree_cache, n);
148 bkey_copy(&n->key, new);
150 n->data->min_key = expected_start;
152 n->data->max_key = node_end;
154 ret = __bch2_btree_node_hash_insert(&c->btree_cache, n);
156 mutex_unlock(&c->btree_cache.lock);
157 six_unlock_read(&n->c.lock);
164 static int bch2_check_fix_ptrs(struct bch_fs *c, enum btree_id btree_id,
165 unsigned level, bool is_root,
168 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(*k);
169 const union bch_extent_entry *entry;
170 struct extent_ptr_decoded p = { 0 };
171 bool do_update = false;
174 bkey_for_each_ptr_decode(k->k, ptrs, p, entry) {
175 struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
176 struct bucket *g = PTR_BUCKET(ca, &p.ptr, true);
177 struct bucket *g2 = PTR_BUCKET(ca, &p.ptr, false);
179 if (fsck_err_on(!g->gen_valid, c,
180 "bucket %u:%zu data type %s ptr gen %u missing in alloc btree",
181 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
182 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
185 g2->_mark.gen = g->_mark.gen = p.ptr.gen;
186 g2->gen_valid = g->gen_valid = true;
187 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
193 if (fsck_err_on(gen_cmp(p.ptr.gen, g->mark.gen) > 0, c,
194 "bucket %u:%zu data type %s ptr gen in the future: %u > %u",
195 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
196 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
197 p.ptr.gen, g->mark.gen)) {
199 g2->_mark.gen = g->_mark.gen = p.ptr.gen;
200 g2->gen_valid = g->gen_valid = true;
201 g2->_mark.data_type = 0;
202 g2->_mark.dirty_sectors = 0;
203 g2->_mark.cached_sectors = 0;
204 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
205 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
211 if (fsck_err_on(!p.ptr.cached &&
212 gen_cmp(p.ptr.gen, g->mark.gen) < 0, c,
213 "bucket %u:%zu data type %s stale dirty ptr: %u < %u",
214 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
215 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
216 p.ptr.gen, g->mark.gen))
220 struct stripe *m = genradix_ptr(&c->stripes[true], p.ec.idx);
222 if (fsck_err_on(!m || !m->alive, c,
223 "pointer to nonexistent stripe %llu",
227 if (fsck_err_on(!bch2_ptr_matches_stripe_m(m, p), c,
228 "pointer does not match stripe %llu",
235 struct bkey_ptrs ptrs;
236 union bch_extent_entry *entry;
237 struct bch_extent_ptr *ptr;
241 bch_err(c, "cannot update btree roots yet");
245 new = kmalloc(bkey_bytes(k->k), GFP_KERNEL);
247 bch_err(c, "%s: error allocating new key", __func__);
251 bkey_reassemble(new, *k);
253 bch2_bkey_drop_ptrs(bkey_i_to_s(new), ptr, ({
254 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
255 struct bucket *g = PTR_BUCKET(ca, ptr, true);
258 (!g->gen_valid || gen_cmp(ptr->gen, g->mark.gen) > 0)) ||
260 gen_cmp(ptr->gen, g->mark.gen) < 0);
263 ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
264 bkey_extent_entry_for_each(ptrs, entry) {
265 if (extent_entry_type(entry) == BCH_EXTENT_ENTRY_stripe_ptr) {
266 struct stripe *m = genradix_ptr(&c->stripes[true],
267 entry->stripe_ptr.idx);
268 union bch_extent_entry *next_ptr;
270 bkey_extent_entry_for_each_from(ptrs, next_ptr, entry)
271 if (extent_entry_type(next_ptr) == BCH_EXTENT_ENTRY_ptr)
276 bch_err(c, "aieee, found stripe ptr with no data ptr");
280 if (!m || !m->alive ||
281 !__bch2_ptr_matches_stripe(&m->ptrs[entry->stripe_ptr.block],
284 bch2_bkey_extent_entry_drop(new, entry);
290 ret = bch2_journal_key_insert(c, btree_id, level, new);
294 *k = bkey_i_to_s_c(new);
300 /* marking of btree keys/nodes: */
302 static int bch2_gc_mark_key(struct bch_fs *c, enum btree_id btree_id,
303 unsigned level, bool is_root,
305 u8 *max_stale, bool initial)
307 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
308 const struct bch_extent_ptr *ptr;
311 (initial ? BTREE_TRIGGER_NOATOMIC : 0);
315 BUG_ON(bch2_journal_seq_verify &&
316 k.k->version.lo > journal_cur_seq(&c->journal));
318 if (fsck_err_on(k.k->version.lo > atomic64_read(&c->key_version), c,
319 "key version number higher than recorded: %llu > %llu",
321 atomic64_read(&c->key_version)))
322 atomic64_set(&c->key_version, k.k->version.lo);
324 if (test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) ||
325 fsck_err_on(!bch2_bkey_replicas_marked(c, k), c,
326 "superblock not marked as containing replicas (type %u)",
328 ret = bch2_mark_bkey_replicas(c, k);
330 bch_err(c, "error marking bkey replicas: %i", ret);
335 ret = bch2_check_fix_ptrs(c, btree_id, level, is_root, &k);
338 bkey_for_each_ptr(ptrs, ptr) {
339 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
340 struct bucket *g = PTR_BUCKET(ca, ptr, true);
342 if (gen_after(g->oldest_gen, ptr->gen))
343 g->oldest_gen = ptr->gen;
345 *max_stale = max(*max_stale, ptr_stale(ca, ptr));
348 bch2_mark_key(c, k, 0, k.k->size, NULL, 0, flags);
352 bch_err(c, "%s: ret %i", __func__, ret);
356 static int btree_gc_mark_node(struct bch_fs *c, struct btree *b, u8 *max_stale,
359 struct btree_node_iter iter;
360 struct bkey unpacked;
362 struct bkey_buf prev, cur;
367 if (!btree_node_type_needs_gc(btree_node_type(b)))
370 bch2_btree_node_iter_init_from_start(&iter, b);
371 bch2_bkey_buf_init(&prev);
372 bch2_bkey_buf_init(&cur);
373 bkey_init(&prev.k->k);
375 while ((k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked)).k) {
376 ret = bch2_gc_mark_key(c, b->c.btree_id, b->c.level, false,
377 k, max_stale, initial);
381 bch2_btree_node_iter_advance(&iter, b);
384 bch2_bkey_buf_reassemble(&cur, c, k);
386 ret = bch2_gc_check_topology(c, b, &prev, cur,
387 bch2_btree_node_iter_end(&iter));
393 bch2_bkey_buf_exit(&cur, c);
394 bch2_bkey_buf_exit(&prev, c);
398 static int bch2_gc_btree(struct bch_fs *c, enum btree_id btree_id,
401 struct btree_trans trans;
402 struct btree_iter *iter;
404 unsigned depth = bch2_expensive_debug_checks ? 0
405 : !btree_node_type_needs_gc(btree_id) ? 1
410 bch2_trans_init(&trans, c, 0, 0);
412 gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));
414 __for_each_btree_node(&trans, iter, btree_id, POS_MIN,
415 0, depth, BTREE_ITER_PREFETCH, b) {
416 bch2_verify_btree_nr_keys(b);
418 gc_pos_set(c, gc_pos_btree_node(b));
420 ret = btree_gc_mark_node(c, b, &max_stale, initial);
426 bch2_btree_node_rewrite(c, iter,
429 BTREE_INSERT_GC_LOCK_HELD);
430 else if (!bch2_btree_gc_rewrite_disabled &&
431 (bch2_btree_gc_always_rewrite || max_stale > 16))
432 bch2_btree_node_rewrite(c, iter,
435 BTREE_INSERT_GC_LOCK_HELD);
438 bch2_trans_cond_resched(&trans);
440 bch2_trans_iter_put(&trans, iter);
442 ret = bch2_trans_exit(&trans) ?: ret;
446 mutex_lock(&c->btree_root_lock);
447 b = c->btree_roots[btree_id].b;
448 if (!btree_node_fake(b))
449 ret = bch2_gc_mark_key(c, b->c.btree_id, b->c.level, true,
450 bkey_i_to_s_c(&b->key),
451 &max_stale, initial);
452 gc_pos_set(c, gc_pos_btree_root(b->c.btree_id));
453 mutex_unlock(&c->btree_root_lock);
458 static int bch2_gc_btree_init_recurse(struct bch_fs *c, struct btree *b,
459 unsigned target_depth)
461 struct btree_and_journal_iter iter;
463 struct bkey_buf cur, prev;
467 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
468 bch2_bkey_buf_init(&prev);
469 bch2_bkey_buf_init(&cur);
470 bkey_init(&prev.k->k);
472 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
473 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
474 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
476 ret = bch2_gc_mark_key(c, b->c.btree_id, b->c.level, false,
477 k, &max_stale, true);
479 bch_err(c, "%s: error %i from bch2_gc_mark_key", __func__, ret);
484 bch2_bkey_buf_reassemble(&cur, c, k);
485 k = bkey_i_to_s_c(cur.k);
487 bch2_btree_and_journal_iter_advance(&iter);
489 ret = bch2_gc_check_topology(c, b,
491 !bch2_btree_and_journal_iter_peek(&iter).k);
495 bch2_btree_and_journal_iter_advance(&iter);
499 if (b->c.level > target_depth) {
500 bch2_btree_and_journal_iter_exit(&iter);
501 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
503 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
506 bch2_bkey_buf_reassemble(&cur, c, k);
507 bch2_btree_and_journal_iter_advance(&iter);
509 child = bch2_btree_node_get_noiter(c, cur.k,
510 b->c.btree_id, b->c.level - 1,
512 ret = PTR_ERR_OR_ZERO(child);
514 if (fsck_err_on(ret == -EIO, c,
515 "unreadable btree node")) {
516 ret = bch2_journal_key_delete(c, b->c.btree_id,
517 b->c.level, cur.k->k.p);
521 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
526 bch_err(c, "%s: error %i getting btree node",
531 ret = bch2_gc_btree_init_recurse(c, child,
533 six_unlock_read(&child->c.lock);
540 bch2_bkey_buf_exit(&cur, c);
541 bch2_bkey_buf_exit(&prev, c);
542 bch2_btree_and_journal_iter_exit(&iter);
546 static int bch2_gc_btree_init(struct bch_fs *c,
547 enum btree_id btree_id)
550 unsigned target_depth = bch2_expensive_debug_checks ? 0
551 : !btree_node_type_needs_gc(btree_id) ? 1
557 b = c->btree_roots[btree_id].b;
559 if (btree_node_fake(b))
562 six_lock_read(&b->c.lock, NULL, NULL);
563 if (fsck_err_on(bpos_cmp(b->data->min_key, POS_MIN), c,
564 "btree root with incorrect min_key: %s",
565 (bch2_bpos_to_text(&PBUF(buf), b->data->min_key), buf))) {
569 if (fsck_err_on(bpos_cmp(b->data->max_key, POS_MAX), c,
570 "btree root with incorrect max_key: %s",
571 (bch2_bpos_to_text(&PBUF(buf), b->data->max_key), buf))) {
575 if (b->c.level >= target_depth)
576 ret = bch2_gc_btree_init_recurse(c, b, target_depth);
579 ret = bch2_gc_mark_key(c, b->c.btree_id, b->c.level, true,
580 bkey_i_to_s_c(&b->key),
583 six_unlock_read(&b->c.lock);
586 bch_err(c, "%s: ret %i", __func__, ret);
590 static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
592 return (int) btree_id_to_gc_phase(l) -
593 (int) btree_id_to_gc_phase(r);
596 static int bch2_gc_btrees(struct bch_fs *c, bool initial)
598 enum btree_id ids[BTREE_ID_NR];
601 for (i = 0; i < BTREE_ID_NR; i++)
603 bubble_sort(ids, BTREE_ID_NR, btree_id_gc_phase_cmp);
605 for (i = 0; i < BTREE_ID_NR; i++) {
606 enum btree_id id = ids[i];
608 ? bch2_gc_btree_init(c, id)
609 : bch2_gc_btree(c, id, initial);
611 bch_err(c, "%s: ret %i", __func__, ret);
619 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
621 enum bch_data_type type,
624 u64 b = sector_to_bucket(ca, start);
628 min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
630 bch2_mark_metadata_bucket(c, ca, b, type, sectors,
631 gc_phase(GC_PHASE_SB), flags);
634 } while (start < end);
637 void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
640 struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
645 * This conditional is kind of gross, but we may be called from the
646 * device add path, before the new device has actually been added to the
647 * running filesystem:
650 lockdep_assert_held(&c->sb_lock);
651 percpu_down_read(&c->mark_lock);
654 for (i = 0; i < layout->nr_superblocks; i++) {
655 u64 offset = le64_to_cpu(layout->sb_offset[i]);
657 if (offset == BCH_SB_SECTOR)
658 mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
661 mark_metadata_sectors(c, ca, offset,
662 offset + (1 << layout->sb_max_size_bits),
666 for (i = 0; i < ca->journal.nr; i++) {
667 b = ca->journal.buckets[i];
668 bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_journal,
670 gc_phase(GC_PHASE_SB), flags);
674 percpu_up_read(&c->mark_lock);
677 static void bch2_mark_superblocks(struct bch_fs *c)
682 mutex_lock(&c->sb_lock);
683 gc_pos_set(c, gc_phase(GC_PHASE_SB));
685 for_each_online_member(ca, c, i)
686 bch2_mark_dev_superblock(c, ca, BTREE_TRIGGER_GC);
687 mutex_unlock(&c->sb_lock);
691 /* Also see bch2_pending_btree_node_free_insert_done() */
692 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
694 struct btree_update *as;
695 struct pending_btree_node_free *d;
697 mutex_lock(&c->btree_interior_update_lock);
698 gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
700 for_each_pending_btree_node_free(c, as, d)
701 if (d->index_update_done)
702 bch2_mark_key(c, bkey_i_to_s_c(&d->key),
706 mutex_unlock(&c->btree_interior_update_lock);
710 static void bch2_mark_allocator_buckets(struct bch_fs *c)
713 struct open_bucket *ob;
717 percpu_down_read(&c->mark_lock);
719 spin_lock(&c->freelist_lock);
720 gc_pos_set(c, gc_pos_alloc(c, NULL));
722 for_each_member_device(ca, c, ci) {
723 fifo_for_each_entry(i, &ca->free_inc, iter)
724 bch2_mark_alloc_bucket(c, ca, i, true,
725 gc_pos_alloc(c, NULL),
730 for (j = 0; j < RESERVE_NR; j++)
731 fifo_for_each_entry(i, &ca->free[j], iter)
732 bch2_mark_alloc_bucket(c, ca, i, true,
733 gc_pos_alloc(c, NULL),
737 spin_unlock(&c->freelist_lock);
739 for (ob = c->open_buckets;
740 ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
742 spin_lock(&ob->lock);
744 gc_pos_set(c, gc_pos_alloc(c, ob));
745 ca = bch_dev_bkey_exists(c, ob->ptr.dev);
746 bch2_mark_alloc_bucket(c, ca, PTR_BUCKET_NR(ca, &ob->ptr), true,
750 spin_unlock(&ob->lock);
753 percpu_up_read(&c->mark_lock);
756 static void bch2_gc_free(struct bch_fs *c)
761 genradix_free(&c->stripes[1]);
763 for_each_member_device(ca, c, i) {
764 kvpfree(rcu_dereference_protected(ca->buckets[1], 1),
765 sizeof(struct bucket_array) +
766 ca->mi.nbuckets * sizeof(struct bucket));
767 ca->buckets[1] = NULL;
769 free_percpu(ca->usage_gc);
773 free_percpu(c->usage_gc);
777 static int bch2_gc_done(struct bch_fs *c,
781 bool verify = (!initial ||
782 (c->sb.compat & (1ULL << BCH_COMPAT_alloc_info)));
786 #define copy_field(_f, _msg, ...) \
787 if (dst->_f != src->_f) { \
789 fsck_err(c, _msg ": got %llu, should be %llu" \
790 , ##__VA_ARGS__, dst->_f, src->_f); \
792 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags); \
794 #define copy_stripe_field(_f, _msg, ...) \
795 if (dst->_f != src->_f) { \
797 fsck_err(c, "stripe %zu has wrong "_msg \
798 ": got %u, should be %u", \
799 iter.pos, ##__VA_ARGS__, \
802 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags); \
804 #define copy_bucket_field(_f) \
805 if (dst->b[b].mark._f != src->b[b].mark._f) { \
807 fsck_err(c, "bucket %u:%zu gen %u data type %s has wrong " #_f \
808 ": got %u, should be %u", i, b, \
809 dst->b[b].mark.gen, \
810 bch2_data_types[dst->b[b].mark.data_type],\
811 dst->b[b].mark._f, src->b[b].mark._f); \
812 dst->b[b]._mark._f = src->b[b].mark._f; \
813 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags); \
815 #define copy_dev_field(_f, _msg, ...) \
816 copy_field(_f, "dev %u has wrong " _msg, i, ##__VA_ARGS__)
817 #define copy_fs_field(_f, _msg, ...) \
818 copy_field(_f, "fs has wrong " _msg, ##__VA_ARGS__)
821 struct genradix_iter iter = genradix_iter_init(&c->stripes[1], 0);
822 struct stripe *dst, *src;
824 while ((src = genradix_iter_peek(&iter, &c->stripes[1]))) {
825 dst = genradix_ptr_alloc(&c->stripes[0], iter.pos, GFP_KERNEL);
827 if (dst->alive != src->alive ||
828 dst->sectors != src->sectors ||
829 dst->algorithm != src->algorithm ||
830 dst->nr_blocks != src->nr_blocks ||
831 dst->nr_redundant != src->nr_redundant) {
832 bch_err(c, "unexpected stripe inconsistency at bch2_gc_done, confused");
837 for (i = 0; i < ARRAY_SIZE(dst->block_sectors); i++)
838 copy_stripe_field(block_sectors[i],
839 "block_sectors[%u]", i);
841 dst->blocks_nonempty = 0;
842 for (i = 0; i < dst->nr_blocks; i++)
843 dst->blocks_nonempty += dst->block_sectors[i] != 0;
845 genradix_iter_advance(&iter, &c->stripes[1]);
849 for (i = 0; i < ARRAY_SIZE(c->usage); i++)
850 bch2_fs_usage_acc_to_base(c, i);
852 for_each_member_device(ca, c, dev) {
853 struct bucket_array *dst = __bucket_array(ca, 0);
854 struct bucket_array *src = __bucket_array(ca, 1);
857 for (b = 0; b < src->nbuckets; b++) {
858 copy_bucket_field(gen);
859 copy_bucket_field(data_type);
860 copy_bucket_field(owned_by_allocator);
861 copy_bucket_field(stripe);
862 copy_bucket_field(dirty_sectors);
863 copy_bucket_field(cached_sectors);
865 dst->b[b].oldest_gen = src->b[b].oldest_gen;
869 struct bch_dev_usage *dst = ca->usage_base;
870 struct bch_dev_usage *src = (void *)
871 bch2_acc_percpu_u64s((void *) ca->usage_gc,
874 copy_dev_field(buckets_ec, "buckets_ec");
875 copy_dev_field(buckets_unavailable, "buckets_unavailable");
877 for (i = 0; i < BCH_DATA_NR; i++) {
878 copy_dev_field(d[i].buckets, "%s buckets", bch2_data_types[i]);
879 copy_dev_field(d[i].sectors, "%s sectors", bch2_data_types[i]);
880 copy_dev_field(d[i].fragmented, "%s fragmented", bch2_data_types[i]);
886 unsigned nr = fs_usage_u64s(c);
887 struct bch_fs_usage *dst = c->usage_base;
888 struct bch_fs_usage *src = (void *)
889 bch2_acc_percpu_u64s((void *) c->usage_gc, nr);
891 copy_fs_field(hidden, "hidden");
892 copy_fs_field(btree, "btree");
893 copy_fs_field(data, "data");
894 copy_fs_field(cached, "cached");
895 copy_fs_field(reserved, "reserved");
896 copy_fs_field(nr_inodes,"nr_inodes");
898 for (i = 0; i < BCH_REPLICAS_MAX; i++)
899 copy_fs_field(persistent_reserved[i],
900 "persistent_reserved[%i]", i);
902 for (i = 0; i < c->replicas.nr; i++) {
903 struct bch_replicas_entry *e =
904 cpu_replicas_entry(&c->replicas, i);
907 bch2_replicas_entry_to_text(&PBUF(buf), e);
909 copy_fs_field(replicas[i], "%s", buf);
914 #undef copy_dev_field
915 #undef copy_bucket_field
916 #undef copy_stripe_field
920 bch_err(c, "%s: ret %i", __func__, ret);
924 static int bch2_gc_start(struct bch_fs *c)
932 c->usage_gc = __alloc_percpu_gfp(fs_usage_u64s(c) * sizeof(u64),
933 sizeof(u64), GFP_KERNEL);
935 bch_err(c, "error allocating c->usage_gc");
939 for_each_member_device(ca, c, i) {
940 BUG_ON(ca->buckets[1]);
941 BUG_ON(ca->usage_gc);
943 ca->buckets[1] = kvpmalloc(sizeof(struct bucket_array) +
944 ca->mi.nbuckets * sizeof(struct bucket),
945 GFP_KERNEL|__GFP_ZERO);
946 if (!ca->buckets[1]) {
947 percpu_ref_put(&ca->ref);
948 bch_err(c, "error allocating ca->buckets[gc]");
952 ca->usage_gc = alloc_percpu(struct bch_dev_usage);
954 bch_err(c, "error allocating ca->usage_gc");
955 percpu_ref_put(&ca->ref);
960 ret = bch2_ec_mem_alloc(c, true);
962 bch_err(c, "error allocating ec gc mem");
966 percpu_down_write(&c->mark_lock);
969 * indicate to stripe code that we need to allocate for the gc stripes
972 gc_pos_set(c, gc_phase(GC_PHASE_START));
974 for_each_member_device(ca, c, i) {
975 struct bucket_array *dst = __bucket_array(ca, 1);
976 struct bucket_array *src = __bucket_array(ca, 0);
979 dst->first_bucket = src->first_bucket;
980 dst->nbuckets = src->nbuckets;
982 for (b = 0; b < src->nbuckets; b++) {
983 struct bucket *d = &dst->b[b];
984 struct bucket *s = &src->b[b];
986 d->_mark.gen = dst->b[b].oldest_gen = s->mark.gen;
987 d->gen_valid = s->gen_valid;
991 percpu_up_write(&c->mark_lock);
997 * bch2_gc - walk _all_ references to buckets, and recompute them:
999 * Order matters here:
1000 * - Concurrent GC relies on the fact that we have a total ordering for
1001 * everything that GC walks - see gc_will_visit_node(),
1002 * gc_will_visit_root()
1004 * - also, references move around in the course of index updates and
1005 * various other crap: everything needs to agree on the ordering
1006 * references are allowed to move around in - e.g., we're allowed to
1007 * start with a reference owned by an open_bucket (the allocator) and
1008 * move it to the btree, but not the reverse.
1010 * This is necessary to ensure that gc doesn't miss references that
1011 * move around - if references move backwards in the ordering GC
1012 * uses, GC could skip past them
1014 int bch2_gc(struct bch_fs *c, bool initial)
1017 u64 start_time = local_clock();
1018 unsigned i, iter = 0;
1021 lockdep_assert_held(&c->state_lock);
1024 down_write(&c->gc_lock);
1026 /* flush interior btree updates: */
1027 closure_wait_event(&c->btree_interior_update_wait,
1028 !bch2_btree_interior_updates_nr_pending(c));
1030 ret = bch2_gc_start(c);
1034 bch2_mark_superblocks(c);
1036 ret = bch2_gc_btrees(c, initial);
1041 bch2_mark_pending_btree_node_frees(c);
1043 bch2_mark_allocator_buckets(c);
1047 if (test_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags) ||
1048 (!iter && bch2_test_restart_gc)) {
1050 * XXX: make sure gens we fixed got saved
1053 bch_info(c, "Second GC pass needed, restarting:");
1054 clear_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1055 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1057 percpu_down_write(&c->mark_lock);
1059 percpu_up_write(&c->mark_lock);
1060 /* flush fsck errors, reset counters */
1061 bch2_flush_fsck_errs(c);
1066 bch_info(c, "Unable to fix bucket gens, looping");
1071 bch2_journal_block(&c->journal);
1073 percpu_down_write(&c->mark_lock);
1074 ret = bch2_gc_done(c, initial);
1076 bch2_journal_unblock(&c->journal);
1078 percpu_down_write(&c->mark_lock);
1081 /* Indicates that gc is no longer in progress: */
1082 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1085 percpu_up_write(&c->mark_lock);
1087 up_write(&c->gc_lock);
1090 bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
1093 * Wake up allocator in case it was waiting for buckets
1094 * because of not being able to inc gens
1096 for_each_member_device(ca, c, i)
1097 bch2_wake_allocator(ca);
1100 * At startup, allocations can happen directly instead of via the
1101 * allocator thread - issue wakeup in case they blocked on gc_lock:
1103 closure_wake_up(&c->freelist_wait);
1107 static bool gc_btree_gens_key(struct bch_fs *c, struct bkey_s_c k)
1109 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1110 const struct bch_extent_ptr *ptr;
1112 percpu_down_read(&c->mark_lock);
1113 bkey_for_each_ptr(ptrs, ptr) {
1114 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1115 struct bucket *g = PTR_BUCKET(ca, ptr, false);
1117 if (gen_after(g->mark.gen, ptr->gen) > 16) {
1118 percpu_up_read(&c->mark_lock);
1123 bkey_for_each_ptr(ptrs, ptr) {
1124 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1125 struct bucket *g = PTR_BUCKET(ca, ptr, false);
1127 if (gen_after(g->gc_gen, ptr->gen))
1128 g->gc_gen = ptr->gen;
1130 percpu_up_read(&c->mark_lock);
1136 * For recalculating oldest gen, we only need to walk keys in leaf nodes; btree
1137 * node pointers currently never have cached pointers that can become stale:
1139 static int bch2_gc_btree_gens(struct bch_fs *c, enum btree_id btree_id)
1141 struct btree_trans trans;
1142 struct btree_iter *iter;
1147 bch2_bkey_buf_init(&sk);
1148 bch2_trans_init(&trans, c, 0, 0);
1150 iter = bch2_trans_get_iter(&trans, btree_id, POS_MIN,
1151 BTREE_ITER_PREFETCH|
1152 BTREE_ITER_NOT_EXTENTS|
1153 BTREE_ITER_ALL_SNAPSHOTS);
1155 while ((k = bch2_btree_iter_peek(iter)).k &&
1156 !(ret = bkey_err(k))) {
1157 if (gc_btree_gens_key(c, k)) {
1158 bch2_bkey_buf_reassemble(&sk, c, k);
1159 bch2_extent_normalize(c, bkey_i_to_s(sk.k));
1161 bch2_btree_iter_set_pos(iter, bkey_start_pos(&sk.k->k));
1163 bch2_trans_update(&trans, iter, sk.k, 0);
1165 ret = bch2_trans_commit(&trans, NULL, NULL,
1166 BTREE_INSERT_NOFAIL);
1174 bch2_btree_iter_advance(iter);
1176 bch2_trans_iter_put(&trans, iter);
1178 bch2_trans_exit(&trans);
1179 bch2_bkey_buf_exit(&sk, c);
1184 int bch2_gc_gens(struct bch_fs *c)
1187 struct bucket_array *buckets;
1193 * Ideally we would be using state_lock and not gc_lock here, but that
1194 * introduces a deadlock in the RO path - we currently take the state
1195 * lock at the start of going RO, thus the gc thread may get stuck:
1197 down_read(&c->gc_lock);
1199 for_each_member_device(ca, c, i) {
1200 down_read(&ca->bucket_lock);
1201 buckets = bucket_array(ca);
1203 for_each_bucket(g, buckets)
1204 g->gc_gen = g->mark.gen;
1205 up_read(&ca->bucket_lock);
1208 for (i = 0; i < BTREE_ID_NR; i++)
1209 if (btree_node_type_needs_gc(i)) {
1210 ret = bch2_gc_btree_gens(c, i);
1212 bch_err(c, "error recalculating oldest_gen: %i", ret);
1217 for_each_member_device(ca, c, i) {
1218 down_read(&ca->bucket_lock);
1219 buckets = bucket_array(ca);
1221 for_each_bucket(g, buckets)
1222 g->oldest_gen = g->gc_gen;
1223 up_read(&ca->bucket_lock);
1228 up_read(&c->gc_lock);
1232 /* Btree coalescing */
1234 static void recalc_packed_keys(struct btree *b)
1236 struct bset *i = btree_bset_first(b);
1237 struct bkey_packed *k;
1239 memset(&b->nr, 0, sizeof(b->nr));
1241 BUG_ON(b->nsets != 1);
1243 vstruct_for_each(i, k)
1244 btree_keys_account_key_add(&b->nr, 0, k);
1247 static void bch2_coalesce_nodes(struct bch_fs *c, struct btree_iter *iter,
1248 struct btree *old_nodes[GC_MERGE_NODES])
1250 struct btree *parent = btree_node_parent(iter, old_nodes[0]);
1251 unsigned i, nr_old_nodes, nr_new_nodes, u64s = 0;
1252 unsigned blocks = btree_blocks(c) * 2 / 3;
1253 struct btree *new_nodes[GC_MERGE_NODES];
1254 struct btree_update *as;
1255 struct keylist keylist;
1256 struct bkey_format_state format_state;
1257 struct bkey_format new_format;
1259 memset(new_nodes, 0, sizeof(new_nodes));
1260 bch2_keylist_init(&keylist, NULL);
1262 /* Count keys that are not deleted */
1263 for (i = 0; i < GC_MERGE_NODES && old_nodes[i]; i++)
1264 u64s += old_nodes[i]->nr.live_u64s;
1266 nr_old_nodes = nr_new_nodes = i;
1268 /* Check if all keys in @old_nodes could fit in one fewer node */
1269 if (nr_old_nodes <= 1 ||
1270 __vstruct_blocks(struct btree_node, c->block_bits,
1271 DIV_ROUND_UP(u64s, nr_old_nodes - 1)) > blocks)
1274 /* Find a format that all keys in @old_nodes can pack into */
1275 bch2_bkey_format_init(&format_state);
1278 * XXX: this won't correctly take it account the new min/max keys:
1280 for (i = 0; i < nr_old_nodes; i++)
1281 __bch2_btree_calc_format(&format_state, old_nodes[i]);
1283 new_format = bch2_bkey_format_done(&format_state);
1285 /* Check if repacking would make any nodes too big to fit */
1286 for (i = 0; i < nr_old_nodes; i++)
1287 if (!bch2_btree_node_format_fits(c, old_nodes[i], &new_format)) {
1288 trace_btree_gc_coalesce_fail(c,
1289 BTREE_GC_COALESCE_FAIL_FORMAT_FITS);
1293 if (bch2_keylist_realloc(&keylist, NULL, 0,
1294 BKEY_BTREE_PTR_U64s_MAX * nr_old_nodes)) {
1295 trace_btree_gc_coalesce_fail(c,
1296 BTREE_GC_COALESCE_FAIL_KEYLIST_REALLOC);
1300 as = bch2_btree_update_start(iter, old_nodes[0]->c.level,
1301 btree_update_reserve_required(c, parent) + nr_old_nodes,
1302 BTREE_INSERT_NOFAIL|
1303 BTREE_INSERT_USE_RESERVE);
1305 trace_btree_gc_coalesce_fail(c,
1306 BTREE_GC_COALESCE_FAIL_RESERVE_GET);
1307 bch2_keylist_free(&keylist, NULL);
1311 trace_btree_gc_coalesce(c, old_nodes[0]);
1313 for (i = 0; i < nr_old_nodes; i++)
1314 bch2_btree_interior_update_will_free_node(as, old_nodes[i]);
1316 /* Repack everything with @new_format and sort down to one bset */
1317 for (i = 0; i < nr_old_nodes; i++)
1319 __bch2_btree_node_alloc_replacement(as, old_nodes[i],
1323 * Conceptually we concatenate the nodes together and slice them
1324 * up at different boundaries.
1326 for (i = nr_new_nodes - 1; i > 0; --i) {
1327 struct btree *n1 = new_nodes[i];
1328 struct btree *n2 = new_nodes[i - 1];
1330 struct bset *s1 = btree_bset_first(n1);
1331 struct bset *s2 = btree_bset_first(n2);
1332 struct bkey_packed *k, *last = NULL;
1334 /* Calculate how many keys from @n2 we could fit inside @n1 */
1338 k < vstruct_last(s2) &&
1339 vstruct_blocks_plus(n1->data, c->block_bits,
1340 u64s + k->u64s) <= blocks;
1346 if (u64s == le16_to_cpu(s2->u64s)) {
1347 /* n2 fits entirely in n1 */
1348 n1->key.k.p = n1->data->max_key = n2->data->max_key;
1350 memcpy_u64s(vstruct_last(s1),
1352 le16_to_cpu(s2->u64s));
1353 le16_add_cpu(&s1->u64s, le16_to_cpu(s2->u64s));
1355 set_btree_bset_end(n1, n1->set);
1357 six_unlock_write(&n2->c.lock);
1358 bch2_btree_node_free_never_inserted(c, n2);
1359 six_unlock_intent(&n2->c.lock);
1361 memmove(new_nodes + i - 1,
1363 sizeof(new_nodes[0]) * (nr_new_nodes - i));
1364 new_nodes[--nr_new_nodes] = NULL;
1366 /* move part of n2 into n1 */
1367 n1->key.k.p = n1->data->max_key =
1368 bkey_unpack_pos(n1, last);
1370 n2->data->min_key = bpos_successor(n1->data->max_key);
1372 memcpy_u64s(vstruct_last(s1),
1374 le16_add_cpu(&s1->u64s, u64s);
1377 vstruct_idx(s2, u64s),
1378 (le16_to_cpu(s2->u64s) - u64s) * sizeof(u64));
1379 s2->u64s = cpu_to_le16(le16_to_cpu(s2->u64s) - u64s);
1381 set_btree_bset_end(n1, n1->set);
1382 set_btree_bset_end(n2, n2->set);
1386 for (i = 0; i < nr_new_nodes; i++) {
1387 struct btree *n = new_nodes[i];
1389 recalc_packed_keys(n);
1390 btree_node_reset_sib_u64s(n);
1392 bch2_btree_build_aux_trees(n);
1394 bch2_btree_update_add_new_node(as, n);
1395 six_unlock_write(&n->c.lock);
1397 bch2_btree_node_write(c, n, SIX_LOCK_intent);
1401 * The keys for the old nodes get deleted. We don't want to insert keys
1402 * that compare equal to the keys for the new nodes we'll also be
1403 * inserting - we can't because keys on a keylist must be strictly
1404 * greater than the previous keys, and we also don't need to since the
1405 * key for the new node will serve the same purpose (overwriting the key
1406 * for the old node).
1408 for (i = 0; i < nr_old_nodes; i++) {
1409 struct bkey_i delete;
1412 for (j = 0; j < nr_new_nodes; j++)
1413 if (!bpos_cmp(old_nodes[i]->key.k.p,
1414 new_nodes[j]->key.k.p))
1417 bkey_init(&delete.k);
1418 delete.k.p = old_nodes[i]->key.k.p;
1419 bch2_keylist_add_in_order(&keylist, &delete);
1425 * Keys for the new nodes get inserted: bch2_btree_insert_keys() only
1426 * does the lookup once and thus expects the keys to be in sorted order
1427 * so we have to make sure the new keys are correctly ordered with
1428 * respect to the deleted keys added in the previous loop
1430 for (i = 0; i < nr_new_nodes; i++)
1431 bch2_keylist_add_in_order(&keylist, &new_nodes[i]->key);
1433 /* Insert the newly coalesced nodes */
1434 bch2_btree_insert_node(as, parent, iter, &keylist, 0);
1436 BUG_ON(!bch2_keylist_empty(&keylist));
1438 BUG_ON(iter->l[old_nodes[0]->c.level].b != old_nodes[0]);
1440 bch2_btree_iter_node_replace(iter, new_nodes[0]);
1442 for (i = 0; i < nr_new_nodes; i++)
1443 bch2_btree_update_get_open_buckets(as, new_nodes[i]);
1445 /* Free the old nodes and update our sliding window */
1446 for (i = 0; i < nr_old_nodes; i++) {
1447 bch2_btree_node_free_inmem(c, old_nodes[i], iter);
1450 * the index update might have triggered a split, in which case
1451 * the nodes we coalesced - the new nodes we just created -
1452 * might not be sibling nodes anymore - don't add them to the
1453 * sliding window (except the first):
1456 old_nodes[i] = new_nodes[i];
1458 old_nodes[i] = NULL;
1462 for (i = 0; i < nr_new_nodes; i++)
1463 six_unlock_intent(&new_nodes[i]->c.lock);
1465 bch2_btree_update_done(as);
1466 bch2_keylist_free(&keylist, NULL);
1469 static int bch2_coalesce_btree(struct bch_fs *c, enum btree_id btree_id)
1471 struct btree_trans trans;
1472 struct btree_iter *iter;
1474 bool kthread = (current->flags & PF_KTHREAD) != 0;
1478 /* Sliding window of adjacent btree nodes */
1479 struct btree *merge[GC_MERGE_NODES];
1480 u32 lock_seq[GC_MERGE_NODES];
1482 bch2_trans_init(&trans, c, 0, 0);
1485 * XXX: We don't have a good way of positively matching on sibling nodes
1486 * that have the same parent - this code works by handling the cases
1487 * where they might not have the same parent, and is thus fragile. Ugh.
1489 * Perhaps redo this to use multiple linked iterators?
1491 memset(merge, 0, sizeof(merge));
1493 __for_each_btree_node(&trans, iter, btree_id, POS_MIN,
1495 BTREE_ITER_PREFETCH, b) {
1496 memmove(merge + 1, merge,
1497 sizeof(merge) - sizeof(merge[0]));
1498 memmove(lock_seq + 1, lock_seq,
1499 sizeof(lock_seq) - sizeof(lock_seq[0]));
1503 for (i = 1; i < GC_MERGE_NODES; i++) {
1505 !six_relock_intent(&merge[i]->c.lock, lock_seq[i]))
1508 if (merge[i]->c.level != merge[0]->c.level) {
1509 six_unlock_intent(&merge[i]->c.lock);
1513 memset(merge + i, 0, (GC_MERGE_NODES - i) * sizeof(merge[0]));
1515 bch2_coalesce_nodes(c, iter, merge);
1517 for (i = 1; i < GC_MERGE_NODES && merge[i]; i++) {
1518 lock_seq[i] = merge[i]->c.lock.state.seq;
1519 six_unlock_intent(&merge[i]->c.lock);
1522 lock_seq[0] = merge[0]->c.lock.state.seq;
1524 if (kthread && kthread_should_stop()) {
1529 bch2_trans_cond_resched(&trans);
1532 * If the parent node wasn't relocked, it might have been split
1533 * and the nodes in our sliding window might not have the same
1534 * parent anymore - blow away the sliding window:
1536 if (btree_iter_node(iter, iter->level + 1) &&
1537 !btree_node_intent_locked(iter, iter->level + 1))
1538 memset(merge + 1, 0,
1539 (GC_MERGE_NODES - 1) * sizeof(merge[0]));
1541 bch2_trans_iter_put(&trans, iter);
1543 return bch2_trans_exit(&trans) ?: ret;
1547 * bch_coalesce - coalesce adjacent nodes with low occupancy
1549 void bch2_coalesce(struct bch_fs *c)
1553 down_read(&c->gc_lock);
1554 trace_gc_coalesce_start(c);
1556 for (id = 0; id < BTREE_ID_NR; id++) {
1557 int ret = c->btree_roots[id].b
1558 ? bch2_coalesce_btree(c, id)
1562 if (ret != -ESHUTDOWN)
1563 bch_err(c, "btree coalescing failed: %d", ret);
1568 trace_gc_coalesce_end(c);
1569 up_read(&c->gc_lock);
1572 static int bch2_gc_thread(void *arg)
1574 struct bch_fs *c = arg;
1575 struct io_clock *clock = &c->io_clock[WRITE];
1576 unsigned long last = atomic64_read(&clock->now);
1577 unsigned last_kick = atomic_read(&c->kick_gc);
1584 set_current_state(TASK_INTERRUPTIBLE);
1586 if (kthread_should_stop()) {
1587 __set_current_state(TASK_RUNNING);
1591 if (atomic_read(&c->kick_gc) != last_kick)
1594 if (c->btree_gc_periodic) {
1595 unsigned long next = last + c->capacity / 16;
1597 if (atomic64_read(&clock->now) >= next)
1600 bch2_io_clock_schedule_timeout(clock, next);
1607 __set_current_state(TASK_RUNNING);
1609 last = atomic64_read(&clock->now);
1610 last_kick = atomic_read(&c->kick_gc);
1613 * Full gc is currently incompatible with btree key cache:
1616 ret = bch2_gc(c, false, false);
1618 ret = bch2_gc_gens(c);
1621 bch_err(c, "btree gc failed: %i", ret);
1623 debug_check_no_locks_held();
1629 void bch2_gc_thread_stop(struct bch_fs *c)
1631 struct task_struct *p;
1634 c->gc_thread = NULL;
1642 int bch2_gc_thread_start(struct bch_fs *c)
1644 struct task_struct *p;
1649 p = kthread_create(bch2_gc_thread, c, "bch-gc/%s", c->name);
1651 bch_err(c, "error creating gc thread: %li", PTR_ERR(p));