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_journal_iter.h"
13 #include "btree_key_cache.h"
14 #include "btree_locking.h"
15 #include "btree_update_interior.h"
33 #include <linux/slab.h>
34 #include <linux/bitops.h>
35 #include <linux/freezer.h>
36 #include <linux/kthread.h>
37 #include <linux/preempt.h>
38 #include <linux/rcupdate.h>
39 #include <linux/sched/task.h>
41 #define DROP_THIS_NODE 10
42 #define DROP_PREV_NODE 11
44 static struct bkey_s unsafe_bkey_s_c_to_s(struct bkey_s_c k)
46 return (struct bkey_s) {{{
48 (struct bch_val *) k.v
52 static bool should_restart_for_topology_repair(struct bch_fs *c)
54 return c->opts.fix_errors != FSCK_FIX_no &&
55 !(c->recovery_passes_complete & BIT_ULL(BCH_RECOVERY_PASS_check_topology));
58 static inline void __gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
61 write_seqcount_begin(&c->gc_pos_lock);
63 write_seqcount_end(&c->gc_pos_lock);
67 static inline void gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
69 BUG_ON(gc_pos_cmp(new_pos, c->gc_pos) <= 0);
70 __gc_pos_set(c, new_pos);
74 * Missing: if an interior btree node is empty, we need to do something -
75 * perhaps just kill it
77 static int bch2_gc_check_topology(struct bch_fs *c,
79 struct bkey_buf *prev,
83 struct bpos node_start = b->data->min_key;
84 struct bpos node_end = b->data->max_key;
85 struct bpos expected_start = bkey_deleted(&prev->k->k)
87 : bpos_successor(prev->k->k.p);
88 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
91 if (cur.k->k.type == KEY_TYPE_btree_ptr_v2) {
92 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(cur.k);
94 if (!bpos_eq(expected_start, bp->v.min_key)) {
95 bch2_topology_error(c);
97 if (bkey_deleted(&prev->k->k)) {
98 prt_printf(&buf1, "start of node: ");
99 bch2_bpos_to_text(&buf1, node_start);
101 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(prev->k));
103 bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(cur.k));
109 btree_node_topology_bad_min_key,
110 "btree node with incorrect min_key at btree %s level %u:\n"
113 bch2_btree_id_str(b->c.btree_id), b->c.level,
114 buf1.buf, buf2.buf) && should_restart_for_topology_repair(c)) {
115 bch_info(c, "Halting mark and sweep to start topology repair pass");
116 ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_topology);
119 set_bit(BCH_FS_initial_gc_unfixed, &c->flags);
124 if (is_last && !bpos_eq(cur.k->k.p, node_end)) {
125 bch2_topology_error(c);
127 printbuf_reset(&buf1);
128 printbuf_reset(&buf2);
130 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(cur.k));
131 bch2_bpos_to_text(&buf2, node_end);
133 if (__fsck_err(c, FSCK_CAN_FIX|FSCK_CAN_IGNORE|FSCK_NO_RATELIMIT,
134 btree_node_topology_bad_max_key,
135 "btree node with incorrect max_key at btree %s level %u:\n"
138 bch2_btree_id_str(b->c.btree_id), b->c.level,
139 buf1.buf, buf2.buf) &&
140 should_restart_for_topology_repair(c)) {
141 bch_info(c, "Halting mark and sweep to start topology repair pass");
142 ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_topology);
145 set_bit(BCH_FS_initial_gc_unfixed, &c->flags);
149 bch2_bkey_buf_copy(prev, c, cur.k);
152 printbuf_exit(&buf2);
153 printbuf_exit(&buf1);
157 static void btree_ptr_to_v2(struct btree *b, struct bkey_i_btree_ptr_v2 *dst)
159 switch (b->key.k.type) {
160 case KEY_TYPE_btree_ptr: {
161 struct bkey_i_btree_ptr *src = bkey_i_to_btree_ptr(&b->key);
165 dst->v.seq = b->data->keys.seq;
166 dst->v.sectors_written = 0;
168 dst->v.min_key = b->data->min_key;
169 set_bkey_val_bytes(&dst->k, sizeof(dst->v) + bkey_val_bytes(&src->k));
170 memcpy(dst->v.start, src->v.start, bkey_val_bytes(&src->k));
173 case KEY_TYPE_btree_ptr_v2:
174 bkey_copy(&dst->k_i, &b->key);
181 static void bch2_btree_node_update_key_early(struct btree_trans *trans,
182 enum btree_id btree, unsigned level,
183 struct bkey_s_c old, struct bkey_i *new)
185 struct bch_fs *c = trans->c;
190 bch2_bkey_buf_init(&tmp);
191 bch2_bkey_buf_reassemble(&tmp, c, old);
193 b = bch2_btree_node_get_noiter(trans, tmp.k, btree, level, true);
194 if (!IS_ERR_OR_NULL(b)) {
195 mutex_lock(&c->btree_cache.lock);
197 bch2_btree_node_hash_remove(&c->btree_cache, b);
199 bkey_copy(&b->key, new);
200 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
203 mutex_unlock(&c->btree_cache.lock);
204 six_unlock_read(&b->c.lock);
207 bch2_bkey_buf_exit(&tmp, c);
210 static int set_node_min(struct bch_fs *c, struct btree *b, struct bpos new_min)
212 struct bkey_i_btree_ptr_v2 *new;
215 new = kmalloc_array(BKEY_BTREE_PTR_U64s_MAX, sizeof(u64), GFP_KERNEL);
217 return -BCH_ERR_ENOMEM_gc_repair_key;
219 btree_ptr_to_v2(b, new);
220 b->data->min_key = new_min;
221 new->v.min_key = new_min;
222 SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
224 ret = bch2_journal_key_insert_take(c, b->c.btree_id, b->c.level + 1, &new->k_i);
230 bch2_btree_node_drop_keys_outside_node(b);
231 bkey_copy(&b->key, &new->k_i);
235 static int set_node_max(struct bch_fs *c, struct btree *b, struct bpos new_max)
237 struct bkey_i_btree_ptr_v2 *new;
240 ret = bch2_journal_key_delete(c, b->c.btree_id, b->c.level + 1, b->key.k.p);
244 new = kmalloc_array(BKEY_BTREE_PTR_U64s_MAX, sizeof(u64), GFP_KERNEL);
246 return -BCH_ERR_ENOMEM_gc_repair_key;
248 btree_ptr_to_v2(b, new);
249 b->data->max_key = new_max;
251 SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
253 ret = bch2_journal_key_insert_take(c, b->c.btree_id, b->c.level + 1, &new->k_i);
259 bch2_btree_node_drop_keys_outside_node(b);
261 mutex_lock(&c->btree_cache.lock);
262 bch2_btree_node_hash_remove(&c->btree_cache, b);
264 bkey_copy(&b->key, &new->k_i);
265 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
267 mutex_unlock(&c->btree_cache.lock);
271 static int btree_repair_node_boundaries(struct bch_fs *c, struct btree *b,
272 struct btree *prev, struct btree *cur)
274 struct bpos expected_start = !prev
276 : bpos_successor(prev->key.k.p);
277 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
281 prt_printf(&buf1, "start of node: ");
282 bch2_bpos_to_text(&buf1, b->data->min_key);
284 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(&prev->key));
287 bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(&cur->key));
290 bpos_gt(expected_start, cur->data->min_key) &&
291 BTREE_NODE_SEQ(cur->data) > BTREE_NODE_SEQ(prev->data)) {
292 /* cur overwrites prev: */
294 if (mustfix_fsck_err_on(bpos_ge(prev->data->min_key,
295 cur->data->min_key), c,
296 btree_node_topology_overwritten_by_next_node,
297 "btree node overwritten by next node at btree %s level %u:\n"
300 bch2_btree_id_str(b->c.btree_id), b->c.level,
301 buf1.buf, buf2.buf)) {
302 ret = DROP_PREV_NODE;
306 if (mustfix_fsck_err_on(!bpos_eq(prev->key.k.p,
307 bpos_predecessor(cur->data->min_key)), c,
308 btree_node_topology_bad_max_key,
309 "btree node with incorrect max_key at btree %s level %u:\n"
312 bch2_btree_id_str(b->c.btree_id), b->c.level,
314 ret = set_node_max(c, prev,
315 bpos_predecessor(cur->data->min_key));
317 /* prev overwrites cur: */
319 if (mustfix_fsck_err_on(bpos_ge(expected_start,
320 cur->data->max_key), c,
321 btree_node_topology_overwritten_by_prev_node,
322 "btree node overwritten by prev node at btree %s level %u:\n"
325 bch2_btree_id_str(b->c.btree_id), b->c.level,
326 buf1.buf, buf2.buf)) {
327 ret = DROP_THIS_NODE;
331 if (mustfix_fsck_err_on(!bpos_eq(expected_start, cur->data->min_key), c,
332 btree_node_topology_bad_min_key,
333 "btree node with incorrect min_key at btree %s level %u:\n"
336 bch2_btree_id_str(b->c.btree_id), b->c.level,
338 ret = set_node_min(c, cur, expected_start);
342 printbuf_exit(&buf2);
343 printbuf_exit(&buf1);
347 static int btree_repair_node_end(struct bch_fs *c, struct btree *b,
350 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
353 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(&child->key));
354 bch2_bpos_to_text(&buf2, b->key.k.p);
356 if (mustfix_fsck_err_on(!bpos_eq(child->key.k.p, b->key.k.p), c,
357 btree_node_topology_bad_max_key,
358 "btree node with incorrect max_key at btree %s level %u:\n"
361 bch2_btree_id_str(b->c.btree_id), b->c.level,
362 buf1.buf, buf2.buf)) {
363 ret = set_node_max(c, child, b->key.k.p);
369 printbuf_exit(&buf2);
370 printbuf_exit(&buf1);
374 static int bch2_btree_repair_topology_recurse(struct btree_trans *trans, struct btree *b)
376 struct bch_fs *c = trans->c;
377 struct btree_and_journal_iter iter;
379 struct bkey_buf prev_k, cur_k;
380 struct btree *prev = NULL, *cur = NULL;
381 bool have_child, dropped_children = false;
382 struct printbuf buf = PRINTBUF;
389 have_child = dropped_children = false;
390 bch2_bkey_buf_init(&prev_k);
391 bch2_bkey_buf_init(&cur_k);
392 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
394 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
395 BUG_ON(bpos_lt(k.k->p, b->data->min_key));
396 BUG_ON(bpos_gt(k.k->p, b->data->max_key));
398 bch2_btree_and_journal_iter_advance(&iter);
399 bch2_bkey_buf_reassemble(&cur_k, c, k);
401 cur = bch2_btree_node_get_noiter(trans, cur_k.k,
402 b->c.btree_id, b->c.level - 1,
404 ret = PTR_ERR_OR_ZERO(cur);
406 printbuf_reset(&buf);
407 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(cur_k.k));
409 if (mustfix_fsck_err_on(ret == -EIO, c,
410 btree_node_unreadable,
411 "Topology repair: unreadable btree node at btree %s level %u:\n"
413 bch2_btree_id_str(b->c.btree_id),
416 bch2_btree_node_evict(trans, cur_k.k);
417 ret = bch2_journal_key_delete(c, b->c.btree_id,
418 b->c.level, cur_k.k->k.p);
425 bch_err_msg(c, ret, "getting btree node");
429 ret = btree_repair_node_boundaries(c, b, prev, cur);
431 if (ret == DROP_THIS_NODE) {
432 six_unlock_read(&cur->c.lock);
433 bch2_btree_node_evict(trans, cur_k.k);
434 ret = bch2_journal_key_delete(c, b->c.btree_id,
435 b->c.level, cur_k.k->k.p);
443 six_unlock_read(&prev->c.lock);
446 if (ret == DROP_PREV_NODE) {
447 bch2_btree_node_evict(trans, prev_k.k);
448 ret = bch2_journal_key_delete(c, b->c.btree_id,
449 b->c.level, prev_k.k->k.p);
453 bch2_btree_and_journal_iter_exit(&iter);
454 bch2_bkey_buf_exit(&prev_k, c);
455 bch2_bkey_buf_exit(&cur_k, c);
462 bch2_bkey_buf_copy(&prev_k, c, cur_k.k);
465 if (!ret && !IS_ERR_OR_NULL(prev)) {
467 ret = btree_repair_node_end(c, b, prev);
470 if (!IS_ERR_OR_NULL(prev))
471 six_unlock_read(&prev->c.lock);
473 if (!IS_ERR_OR_NULL(cur))
474 six_unlock_read(&cur->c.lock);
480 bch2_btree_and_journal_iter_exit(&iter);
481 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
483 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
484 bch2_bkey_buf_reassemble(&cur_k, c, k);
485 bch2_btree_and_journal_iter_advance(&iter);
487 cur = bch2_btree_node_get_noiter(trans, cur_k.k,
488 b->c.btree_id, b->c.level - 1,
490 ret = PTR_ERR_OR_ZERO(cur);
492 bch_err_msg(c, ret, "getting btree node");
496 ret = bch2_btree_repair_topology_recurse(trans, cur);
497 six_unlock_read(&cur->c.lock);
500 if (ret == DROP_THIS_NODE) {
501 bch2_btree_node_evict(trans, cur_k.k);
502 ret = bch2_journal_key_delete(c, b->c.btree_id,
503 b->c.level, cur_k.k->k.p);
504 dropped_children = true;
513 printbuf_reset(&buf);
514 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
516 if (mustfix_fsck_err_on(!have_child, c,
517 btree_node_topology_interior_node_empty,
518 "empty interior btree node at btree %s level %u\n"
520 bch2_btree_id_str(b->c.btree_id),
521 b->c.level, buf.buf))
522 ret = DROP_THIS_NODE;
525 if (!IS_ERR_OR_NULL(prev))
526 six_unlock_read(&prev->c.lock);
527 if (!IS_ERR_OR_NULL(cur))
528 six_unlock_read(&cur->c.lock);
530 bch2_btree_and_journal_iter_exit(&iter);
531 bch2_bkey_buf_exit(&prev_k, c);
532 bch2_bkey_buf_exit(&cur_k, c);
534 if (!ret && dropped_children)
541 int bch2_check_topology(struct bch_fs *c)
543 struct btree_trans *trans = bch2_trans_get(c);
548 for (i = 0; i < btree_id_nr_alive(c) && !ret; i++) {
549 struct btree_root *r = bch2_btree_id_root(c, i);
555 if (btree_node_fake(b))
558 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
559 ret = bch2_btree_repair_topology_recurse(trans, b);
560 six_unlock_read(&b->c.lock);
562 if (ret == DROP_THIS_NODE) {
563 bch_err(c, "empty btree root - repair unimplemented");
564 ret = -BCH_ERR_fsck_repair_unimplemented;
568 bch2_trans_put(trans);
573 static int bch2_check_fix_ptrs(struct btree_trans *trans, enum btree_id btree_id,
574 unsigned level, bool is_root,
577 struct bch_fs *c = trans->c;
578 struct bkey_ptrs_c ptrs_c = bch2_bkey_ptrs_c(*k);
579 const union bch_extent_entry *entry_c;
580 struct extent_ptr_decoded p = { 0 };
581 bool do_update = false;
582 struct printbuf buf = PRINTBUF;
587 * use check_bucket_ref here
589 bkey_for_each_ptr_decode(k->k, ptrs_c, p, entry_c) {
590 struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
591 struct bucket *g = PTR_GC_BUCKET(ca, &p.ptr);
592 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, &entry_c->ptr);
595 (c->opts.reconstruct_alloc ||
596 fsck_err(c, ptr_to_missing_alloc_key,
597 "bucket %u:%zu data type %s ptr gen %u missing in alloc btree\n"
599 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
600 bch2_data_type_str(ptr_data_type(k->k, &p.ptr)),
602 (printbuf_reset(&buf),
603 bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))) {
612 if (gen_cmp(p.ptr.gen, g->gen) > 0 &&
613 (c->opts.reconstruct_alloc ||
614 fsck_err(c, ptr_gen_newer_than_bucket_gen,
615 "bucket %u:%zu data type %s ptr gen in the future: %u > %u\n"
617 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
618 bch2_data_type_str(ptr_data_type(k->k, &p.ptr)),
620 (printbuf_reset(&buf),
621 bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))) {
626 g->dirty_sectors = 0;
627 g->cached_sectors = 0;
628 set_bit(BCH_FS_need_another_gc, &c->flags);
634 if (gen_cmp(g->gen, p.ptr.gen) > BUCKET_GC_GEN_MAX &&
635 (c->opts.reconstruct_alloc ||
636 fsck_err(c, ptr_gen_newer_than_bucket_gen,
637 "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
639 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr), g->gen,
640 bch2_data_type_str(ptr_data_type(k->k, &p.ptr)),
642 (printbuf_reset(&buf),
643 bch2_bkey_val_to_text(&buf, c, *k), buf.buf))))
646 if (!p.ptr.cached && gen_cmp(p.ptr.gen, g->gen) < 0 &&
647 (c->opts.reconstruct_alloc ||
648 fsck_err(c, stale_dirty_ptr,
649 "bucket %u:%zu data type %s stale dirty ptr: %u < %u\n"
651 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
652 bch2_data_type_str(ptr_data_type(k->k, &p.ptr)),
654 (printbuf_reset(&buf),
655 bch2_bkey_val_to_text(&buf, c, *k), buf.buf))))
658 if (data_type != BCH_DATA_btree && p.ptr.gen != g->gen)
661 if (fsck_err_on(bucket_data_type(g->data_type) &&
662 bucket_data_type(g->data_type) != data_type, c,
663 ptr_bucket_data_type_mismatch,
664 "bucket %u:%zu different types of data in same bucket: %s, %s\n"
666 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
667 bch2_data_type_str(g->data_type),
668 bch2_data_type_str(data_type),
669 (printbuf_reset(&buf),
670 bch2_bkey_val_to_text(&buf, c, *k), buf.buf))) {
671 if (data_type == BCH_DATA_btree) {
672 g->data_type = data_type;
673 set_bit(BCH_FS_need_another_gc, &c->flags);
680 struct gc_stripe *m = genradix_ptr(&c->gc_stripes, p.ec.idx);
682 if (fsck_err_on(!m || !m->alive, c,
683 ptr_to_missing_stripe,
684 "pointer to nonexistent stripe %llu\n"
687 (printbuf_reset(&buf),
688 bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
691 if (fsck_err_on(m && m->alive && !bch2_ptr_matches_stripe_m(m, p), c,
692 ptr_to_incorrect_stripe,
693 "pointer does not match stripe %llu\n"
696 (printbuf_reset(&buf),
697 bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
703 struct bkey_ptrs ptrs;
704 union bch_extent_entry *entry;
705 struct bch_extent_ptr *ptr;
709 bch_err(c, "cannot update btree roots yet");
714 new = kmalloc(bkey_bytes(k->k), GFP_KERNEL);
716 ret = -BCH_ERR_ENOMEM_gc_repair_key;
717 bch_err_msg(c, ret, "allocating new key");
721 bkey_reassemble(new, *k);
725 * We don't want to drop btree node pointers - if the
726 * btree node isn't there anymore, the read path will
729 ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
730 bkey_for_each_ptr(ptrs, ptr) {
731 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
732 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
737 bch2_bkey_drop_ptrs(bkey_i_to_s(new), ptr, ({
738 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
739 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
740 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, ptr);
743 (!g->gen_valid || gen_cmp(ptr->gen, g->gen) > 0)) ||
745 gen_cmp(ptr->gen, g->gen) < 0) ||
746 gen_cmp(g->gen, ptr->gen) > BUCKET_GC_GEN_MAX ||
748 g->data_type != data_type);
751 ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
752 bkey_extent_entry_for_each(ptrs, entry) {
753 if (extent_entry_type(entry) == BCH_EXTENT_ENTRY_stripe_ptr) {
754 struct gc_stripe *m = genradix_ptr(&c->gc_stripes,
755 entry->stripe_ptr.idx);
756 union bch_extent_entry *next_ptr;
758 bkey_extent_entry_for_each_from(ptrs, next_ptr, entry)
759 if (extent_entry_type(next_ptr) == BCH_EXTENT_ENTRY_ptr)
764 bch_err(c, "aieee, found stripe ptr with no data ptr");
768 if (!m || !m->alive ||
769 !__bch2_ptr_matches_stripe(&m->ptrs[entry->stripe_ptr.block],
772 bch2_bkey_extent_entry_drop(new, entry);
779 ret = bch2_journal_key_insert_take(c, btree_id, level, new);
786 bch2_btree_node_update_key_early(trans, btree_id, level - 1, *k, new);
789 printbuf_reset(&buf);
790 bch2_bkey_val_to_text(&buf, c, *k);
791 bch_info(c, "updated %s", buf.buf);
793 printbuf_reset(&buf);
794 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(new));
795 bch_info(c, "new key %s", buf.buf);
798 *k = bkey_i_to_s_c(new);
806 /* marking of btree keys/nodes: */
808 static int bch2_gc_mark_key(struct btree_trans *trans, enum btree_id btree_id,
809 unsigned level, bool is_root,
813 struct bch_fs *c = trans->c;
814 struct bkey deleted = KEY(0, 0, 0);
815 struct bkey_s_c old = (struct bkey_s_c) { &deleted, NULL };
821 BUG_ON(bch2_journal_seq_verify &&
822 k->k->version.lo > atomic64_read(&c->journal.seq));
824 ret = bch2_check_fix_ptrs(trans, btree_id, level, is_root, k);
828 if (fsck_err_on(k->k->version.lo > atomic64_read(&c->key_version), c,
829 bkey_version_in_future,
830 "key version number higher than recorded: %llu > %llu",
832 atomic64_read(&c->key_version)))
833 atomic64_set(&c->key_version, k->k->version.lo);
836 ret = commit_do(trans, NULL, NULL, 0,
837 bch2_key_trigger(trans, btree_id, level, old, unsafe_bkey_s_c_to_s(*k), BTREE_TRIGGER_GC));
844 static int btree_gc_mark_node(struct btree_trans *trans, struct btree *b, bool initial)
846 struct bch_fs *c = trans->c;
847 struct btree_node_iter iter;
848 struct bkey unpacked;
850 struct bkey_buf prev, cur;
853 if (!btree_node_type_needs_gc(btree_node_type(b)))
856 bch2_btree_node_iter_init_from_start(&iter, b);
857 bch2_bkey_buf_init(&prev);
858 bch2_bkey_buf_init(&cur);
859 bkey_init(&prev.k->k);
861 while ((k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked)).k) {
862 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, false,
867 bch2_btree_node_iter_advance(&iter, b);
870 bch2_bkey_buf_reassemble(&cur, c, k);
872 ret = bch2_gc_check_topology(c, b, &prev, cur,
873 bch2_btree_node_iter_end(&iter));
879 bch2_bkey_buf_exit(&cur, c);
880 bch2_bkey_buf_exit(&prev, c);
884 static int bch2_gc_btree(struct btree_trans *trans, enum btree_id btree_id,
885 bool initial, bool metadata_only)
887 struct bch_fs *c = trans->c;
888 struct btree_iter iter;
890 unsigned depth = metadata_only ? 1 : 0;
893 gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));
895 __for_each_btree_node(trans, iter, btree_id, POS_MIN,
896 0, depth, BTREE_ITER_PREFETCH, b, ret) {
897 bch2_verify_btree_nr_keys(b);
899 gc_pos_set(c, gc_pos_btree_node(b));
901 ret = btree_gc_mark_node(trans, b, initial);
905 bch2_trans_iter_exit(trans, &iter);
910 mutex_lock(&c->btree_root_lock);
911 b = bch2_btree_id_root(c, btree_id)->b;
912 if (!btree_node_fake(b)) {
913 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
915 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level + 1,
918 gc_pos_set(c, gc_pos_btree_root(b->c.btree_id));
919 mutex_unlock(&c->btree_root_lock);
924 static int bch2_gc_btree_init_recurse(struct btree_trans *trans, struct btree *b,
925 unsigned target_depth)
927 struct bch_fs *c = trans->c;
928 struct btree_and_journal_iter iter;
930 struct bkey_buf cur, prev;
931 struct printbuf buf = PRINTBUF;
934 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
935 bch2_bkey_buf_init(&prev);
936 bch2_bkey_buf_init(&cur);
937 bkey_init(&prev.k->k);
939 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
940 BUG_ON(bpos_lt(k.k->p, b->data->min_key));
941 BUG_ON(bpos_gt(k.k->p, b->data->max_key));
943 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level,
949 bch2_bkey_buf_reassemble(&cur, c, k);
950 k = bkey_i_to_s_c(cur.k);
952 bch2_btree_and_journal_iter_advance(&iter);
954 ret = bch2_gc_check_topology(c, b,
956 !bch2_btree_and_journal_iter_peek(&iter).k);
960 bch2_btree_and_journal_iter_advance(&iter);
964 if (b->c.level > target_depth) {
965 bch2_btree_and_journal_iter_exit(&iter);
966 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
968 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
971 bch2_bkey_buf_reassemble(&cur, c, k);
972 bch2_btree_and_journal_iter_advance(&iter);
974 child = bch2_btree_node_get_noiter(trans, cur.k,
975 b->c.btree_id, b->c.level - 1,
977 ret = PTR_ERR_OR_ZERO(child);
980 bch2_topology_error(c);
986 btree_node_read_error,
987 "Unreadable btree node at btree %s level %u:\n"
989 bch2_btree_id_str(b->c.btree_id),
991 (printbuf_reset(&buf),
992 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(cur.k)), buf.buf)) &&
993 should_restart_for_topology_repair(c)) {
994 bch_info(c, "Halting mark and sweep to start topology repair pass");
995 ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_topology);
998 /* Continue marking when opted to not
1001 set_bit(BCH_FS_initial_gc_unfixed, &c->flags);
1005 bch_err_msg(c, ret, "getting btree node");
1009 ret = bch2_gc_btree_init_recurse(trans, child,
1011 six_unlock_read(&child->c.lock);
1018 bch2_bkey_buf_exit(&cur, c);
1019 bch2_bkey_buf_exit(&prev, c);
1020 bch2_btree_and_journal_iter_exit(&iter);
1021 printbuf_exit(&buf);
1025 static int bch2_gc_btree_init(struct btree_trans *trans,
1026 enum btree_id btree_id,
1029 struct bch_fs *c = trans->c;
1031 unsigned target_depth = metadata_only ? 1 : 0;
1032 struct printbuf buf = PRINTBUF;
1035 b = bch2_btree_id_root(c, btree_id)->b;
1037 if (btree_node_fake(b))
1040 six_lock_read(&b->c.lock, NULL, NULL);
1041 printbuf_reset(&buf);
1042 bch2_bpos_to_text(&buf, b->data->min_key);
1043 if (mustfix_fsck_err_on(!bpos_eq(b->data->min_key, POS_MIN), c,
1044 btree_root_bad_min_key,
1045 "btree root with incorrect min_key: %s", buf.buf)) {
1046 bch_err(c, "repair unimplemented");
1047 ret = -BCH_ERR_fsck_repair_unimplemented;
1051 printbuf_reset(&buf);
1052 bch2_bpos_to_text(&buf, b->data->max_key);
1053 if (mustfix_fsck_err_on(!bpos_eq(b->data->max_key, SPOS_MAX), c,
1054 btree_root_bad_max_key,
1055 "btree root with incorrect max_key: %s", buf.buf)) {
1056 bch_err(c, "repair unimplemented");
1057 ret = -BCH_ERR_fsck_repair_unimplemented;
1061 if (b->c.level >= target_depth)
1062 ret = bch2_gc_btree_init_recurse(trans, b, target_depth);
1065 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1067 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level + 1, true,
1071 six_unlock_read(&b->c.lock);
1074 printbuf_exit(&buf);
1078 static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
1080 return (int) btree_id_to_gc_phase(l) -
1081 (int) btree_id_to_gc_phase(r);
1084 static int bch2_gc_btrees(struct bch_fs *c, bool initial, bool metadata_only)
1086 struct btree_trans *trans = bch2_trans_get(c);
1087 enum btree_id ids[BTREE_ID_NR];
1091 for (i = 0; i < BTREE_ID_NR; i++)
1093 bubble_sort(ids, BTREE_ID_NR, btree_id_gc_phase_cmp);
1095 for (i = 0; i < BTREE_ID_NR && !ret; i++)
1097 ? bch2_gc_btree_init(trans, ids[i], metadata_only)
1098 : bch2_gc_btree(trans, ids[i], initial, metadata_only);
1100 for (i = BTREE_ID_NR; i < btree_id_nr_alive(c) && !ret; i++) {
1101 if (!bch2_btree_id_root(c, i)->alive)
1105 ? bch2_gc_btree_init(trans, i, metadata_only)
1106 : bch2_gc_btree(trans, i, initial, metadata_only);
1109 bch2_trans_put(trans);
1114 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
1116 enum bch_data_type type,
1119 u64 b = sector_to_bucket(ca, start);
1123 min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1125 bch2_mark_metadata_bucket(c, ca, b, type, sectors,
1126 gc_phase(GC_PHASE_SB), flags);
1129 } while (start < end);
1132 static void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
1135 struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1139 for (i = 0; i < layout->nr_superblocks; i++) {
1140 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1142 if (offset == BCH_SB_SECTOR)
1143 mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
1144 BCH_DATA_sb, flags);
1146 mark_metadata_sectors(c, ca, offset,
1147 offset + (1 << layout->sb_max_size_bits),
1148 BCH_DATA_sb, flags);
1151 for (i = 0; i < ca->journal.nr; i++) {
1152 b = ca->journal.buckets[i];
1153 bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_journal,
1155 gc_phase(GC_PHASE_SB), flags);
1159 static void bch2_mark_superblocks(struct bch_fs *c)
1161 mutex_lock(&c->sb_lock);
1162 gc_pos_set(c, gc_phase(GC_PHASE_SB));
1164 for_each_online_member(c, ca)
1165 bch2_mark_dev_superblock(c, ca, BTREE_TRIGGER_GC);
1166 mutex_unlock(&c->sb_lock);
1170 /* Also see bch2_pending_btree_node_free_insert_done() */
1171 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
1173 struct btree_update *as;
1174 struct pending_btree_node_free *d;
1176 mutex_lock(&c->btree_interior_update_lock);
1177 gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
1179 for_each_pending_btree_node_free(c, as, d)
1180 if (d->index_update_done)
1181 bch2_mark_key(c, bkey_i_to_s_c(&d->key), BTREE_TRIGGER_GC);
1183 mutex_unlock(&c->btree_interior_update_lock);
1187 static void bch2_gc_free(struct bch_fs *c)
1189 genradix_free(&c->reflink_gc_table);
1190 genradix_free(&c->gc_stripes);
1192 for_each_member_device(c, ca) {
1193 kvpfree(rcu_dereference_protected(ca->buckets_gc, 1),
1194 sizeof(struct bucket_array) +
1195 ca->mi.nbuckets * sizeof(struct bucket));
1196 ca->buckets_gc = NULL;
1198 free_percpu(ca->usage_gc);
1199 ca->usage_gc = NULL;
1202 free_percpu(c->usage_gc);
1206 static int bch2_gc_done(struct bch_fs *c,
1207 bool initial, bool metadata_only)
1209 struct bch_dev *ca = NULL;
1210 struct printbuf buf = PRINTBUF;
1211 bool verify = !metadata_only &&
1212 !c->opts.reconstruct_alloc &&
1213 (!initial || (c->sb.compat & (1ULL << BCH_COMPAT_alloc_info)));
1217 percpu_down_write(&c->mark_lock);
1219 #define copy_field(_err, _f, _msg, ...) \
1220 if (dst->_f != src->_f && \
1222 fsck_err(c, _err, _msg ": got %llu, should be %llu" \
1223 , ##__VA_ARGS__, dst->_f, src->_f))) \
1225 #define copy_dev_field(_err, _f, _msg, ...) \
1226 copy_field(_err, _f, "dev %u has wrong " _msg, ca->dev_idx, ##__VA_ARGS__)
1227 #define copy_fs_field(_err, _f, _msg, ...) \
1228 copy_field(_err, _f, "fs has wrong " _msg, ##__VA_ARGS__)
1230 for (i = 0; i < ARRAY_SIZE(c->usage); i++)
1231 bch2_fs_usage_acc_to_base(c, i);
1233 __for_each_member_device(c, ca) {
1234 struct bch_dev_usage *dst = ca->usage_base;
1235 struct bch_dev_usage *src = (void *)
1236 bch2_acc_percpu_u64s((u64 __percpu *) ca->usage_gc,
1239 for (i = 0; i < BCH_DATA_NR; i++) {
1240 copy_dev_field(dev_usage_buckets_wrong,
1241 d[i].buckets, "%s buckets", bch2_data_type_str(i));
1242 copy_dev_field(dev_usage_sectors_wrong,
1243 d[i].sectors, "%s sectors", bch2_data_type_str(i));
1244 copy_dev_field(dev_usage_fragmented_wrong,
1245 d[i].fragmented, "%s fragmented", bch2_data_type_str(i));
1250 unsigned nr = fs_usage_u64s(c);
1251 struct bch_fs_usage *dst = c->usage_base;
1252 struct bch_fs_usage *src = (void *)
1253 bch2_acc_percpu_u64s((u64 __percpu *) c->usage_gc, nr);
1255 copy_fs_field(fs_usage_hidden_wrong,
1256 b.hidden, "hidden");
1257 copy_fs_field(fs_usage_btree_wrong,
1260 if (!metadata_only) {
1261 copy_fs_field(fs_usage_data_wrong,
1263 copy_fs_field(fs_usage_cached_wrong,
1264 b.cached, "cached");
1265 copy_fs_field(fs_usage_reserved_wrong,
1266 b.reserved, "reserved");
1267 copy_fs_field(fs_usage_nr_inodes_wrong,
1268 b.nr_inodes,"nr_inodes");
1270 for (i = 0; i < BCH_REPLICAS_MAX; i++)
1271 copy_fs_field(fs_usage_persistent_reserved_wrong,
1272 persistent_reserved[i],
1273 "persistent_reserved[%i]", i);
1276 for (i = 0; i < c->replicas.nr; i++) {
1277 struct bch_replicas_entry_v1 *e =
1278 cpu_replicas_entry(&c->replicas, i);
1280 if (metadata_only &&
1281 (e->data_type == BCH_DATA_user ||
1282 e->data_type == BCH_DATA_cached))
1285 printbuf_reset(&buf);
1286 bch2_replicas_entry_to_text(&buf, e);
1288 copy_fs_field(fs_usage_replicas_wrong,
1289 replicas[i], "%s", buf.buf);
1293 #undef copy_fs_field
1294 #undef copy_dev_field
1295 #undef copy_stripe_field
1299 percpu_ref_put(&ca->ref);
1302 percpu_up_write(&c->mark_lock);
1303 printbuf_exit(&buf);
1307 static int bch2_gc_start(struct bch_fs *c)
1309 BUG_ON(c->usage_gc);
1311 c->usage_gc = __alloc_percpu_gfp(fs_usage_u64s(c) * sizeof(u64),
1312 sizeof(u64), GFP_KERNEL);
1314 bch_err(c, "error allocating c->usage_gc");
1315 return -BCH_ERR_ENOMEM_gc_start;
1318 for_each_member_device(c, ca) {
1319 BUG_ON(ca->usage_gc);
1321 ca->usage_gc = alloc_percpu(struct bch_dev_usage);
1322 if (!ca->usage_gc) {
1323 bch_err(c, "error allocating ca->usage_gc");
1324 percpu_ref_put(&ca->ref);
1325 return -BCH_ERR_ENOMEM_gc_start;
1328 this_cpu_write(ca->usage_gc->d[BCH_DATA_free].buckets,
1329 ca->mi.nbuckets - ca->mi.first_bucket);
1335 static int bch2_gc_reset(struct bch_fs *c)
1337 for_each_member_device(c, ca) {
1338 free_percpu(ca->usage_gc);
1339 ca->usage_gc = NULL;
1342 free_percpu(c->usage_gc);
1345 return bch2_gc_start(c);
1348 /* returns true if not equal */
1349 static inline bool bch2_alloc_v4_cmp(struct bch_alloc_v4 l,
1350 struct bch_alloc_v4 r)
1352 return l.gen != r.gen ||
1353 l.oldest_gen != r.oldest_gen ||
1354 l.data_type != r.data_type ||
1355 l.dirty_sectors != r.dirty_sectors ||
1356 l.cached_sectors != r.cached_sectors ||
1357 l.stripe_redundancy != r.stripe_redundancy ||
1358 l.stripe != r.stripe;
1361 static int bch2_alloc_write_key(struct btree_trans *trans,
1362 struct btree_iter *iter,
1366 struct bch_fs *c = trans->c;
1367 struct bch_dev *ca = bch_dev_bkey_exists(c, iter->pos.inode);
1368 struct bucket gc, *b;
1369 struct bkey_i_alloc_v4 *a;
1370 struct bch_alloc_v4 old_convert, new;
1371 const struct bch_alloc_v4 *old;
1372 enum bch_data_type type;
1375 old = bch2_alloc_to_v4(k, &old_convert);
1378 percpu_down_read(&c->mark_lock);
1379 b = gc_bucket(ca, iter->pos.offset);
1382 * b->data_type doesn't yet include need_discard & need_gc_gen states -
1385 type = __alloc_data_type(b->dirty_sectors,
1390 if (b->data_type != type) {
1391 struct bch_dev_usage *u;
1394 u = this_cpu_ptr(ca->usage_gc);
1395 u->d[b->data_type].buckets--;
1396 b->data_type = type;
1397 u->d[b->data_type].buckets++;
1402 percpu_up_read(&c->mark_lock);
1404 if (metadata_only &&
1405 gc.data_type != BCH_DATA_sb &&
1406 gc.data_type != BCH_DATA_journal &&
1407 gc.data_type != BCH_DATA_btree)
1410 if (gen_after(old->gen, gc.gen))
1413 if (c->opts.reconstruct_alloc ||
1414 fsck_err_on(new.data_type != gc.data_type, c,
1415 alloc_key_data_type_wrong,
1416 "bucket %llu:%llu gen %u has wrong data_type"
1417 ": got %s, should be %s",
1418 iter->pos.inode, iter->pos.offset,
1420 bch2_data_type_str(new.data_type),
1421 bch2_data_type_str(gc.data_type)))
1422 new.data_type = gc.data_type;
1424 #define copy_bucket_field(_errtype, _f) \
1425 if (c->opts.reconstruct_alloc || \
1426 fsck_err_on(new._f != gc._f, c, _errtype, \
1427 "bucket %llu:%llu gen %u data type %s has wrong " #_f \
1428 ": got %u, should be %u", \
1429 iter->pos.inode, iter->pos.offset, \
1431 bch2_data_type_str(gc.data_type), \
1435 copy_bucket_field(alloc_key_gen_wrong,
1437 copy_bucket_field(alloc_key_dirty_sectors_wrong,
1439 copy_bucket_field(alloc_key_cached_sectors_wrong,
1441 copy_bucket_field(alloc_key_stripe_wrong,
1443 copy_bucket_field(alloc_key_stripe_redundancy_wrong,
1445 #undef copy_bucket_field
1447 if (!bch2_alloc_v4_cmp(*old, new))
1450 a = bch2_alloc_to_v4_mut(trans, k);
1451 ret = PTR_ERR_OR_ZERO(a);
1458 * The trigger normally makes sure this is set, but we're not running
1461 if (a->v.data_type == BCH_DATA_cached && !a->v.io_time[READ])
1462 a->v.io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
1464 ret = bch2_trans_update(trans, iter, &a->k_i, BTREE_TRIGGER_NORUN);
1469 static int bch2_gc_alloc_done(struct bch_fs *c, bool metadata_only)
1473 for_each_member_device(c, ca) {
1474 ret = bch2_trans_run(c,
1475 for_each_btree_key_upto_commit(trans, iter, BTREE_ID_alloc,
1476 POS(ca->dev_idx, ca->mi.first_bucket),
1477 POS(ca->dev_idx, ca->mi.nbuckets - 1),
1478 BTREE_ITER_SLOTS|BTREE_ITER_PREFETCH, k,
1479 NULL, NULL, BCH_TRANS_COMMIT_lazy_rw,
1480 bch2_alloc_write_key(trans, &iter, k, metadata_only)));
1482 percpu_ref_put(&ca->ref);
1491 static int bch2_gc_alloc_start(struct bch_fs *c, bool metadata_only)
1493 for_each_member_device(c, ca) {
1494 struct bucket_array *buckets = kvpmalloc(sizeof(struct bucket_array) +
1495 ca->mi.nbuckets * sizeof(struct bucket),
1496 GFP_KERNEL|__GFP_ZERO);
1498 percpu_ref_put(&ca->ref);
1499 bch_err(c, "error allocating ca->buckets[gc]");
1500 return -BCH_ERR_ENOMEM_gc_alloc_start;
1503 buckets->first_bucket = ca->mi.first_bucket;
1504 buckets->nbuckets = ca->mi.nbuckets;
1505 rcu_assign_pointer(ca->buckets_gc, buckets);
1508 int ret = bch2_trans_run(c,
1509 for_each_btree_key(trans, iter, BTREE_ID_alloc, POS_MIN,
1510 BTREE_ITER_PREFETCH, k, ({
1511 struct bch_dev *ca = bch_dev_bkey_exists(c, k.k->p.inode);
1512 struct bucket *g = gc_bucket(ca, k.k->p.offset);
1514 struct bch_alloc_v4 a_convert;
1515 const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &a_convert);
1520 if (metadata_only &&
1521 (a->data_type == BCH_DATA_user ||
1522 a->data_type == BCH_DATA_cached ||
1523 a->data_type == BCH_DATA_parity)) {
1524 g->data_type = a->data_type;
1525 g->dirty_sectors = a->dirty_sectors;
1526 g->cached_sectors = a->cached_sectors;
1527 g->stripe = a->stripe;
1528 g->stripe_redundancy = a->stripe_redundancy;
1537 static void bch2_gc_alloc_reset(struct bch_fs *c, bool metadata_only)
1539 for_each_member_device(c, ca) {
1540 struct bucket_array *buckets = gc_bucket_array(ca);
1543 for_each_bucket(g, buckets) {
1544 if (metadata_only &&
1545 (g->data_type == BCH_DATA_user ||
1546 g->data_type == BCH_DATA_cached ||
1547 g->data_type == BCH_DATA_parity))
1550 g->dirty_sectors = 0;
1551 g->cached_sectors = 0;
1556 static int bch2_gc_write_reflink_key(struct btree_trans *trans,
1557 struct btree_iter *iter,
1561 struct bch_fs *c = trans->c;
1562 const __le64 *refcount = bkey_refcount_c(k);
1563 struct printbuf buf = PRINTBUF;
1564 struct reflink_gc *r;
1570 while ((r = genradix_ptr(&c->reflink_gc_table, *idx)) &&
1571 r->offset < k.k->p.offset)
1575 r->offset != k.k->p.offset ||
1576 r->size != k.k->size) {
1577 bch_err(c, "unexpected inconsistency walking reflink table at gc finish");
1581 if (fsck_err_on(r->refcount != le64_to_cpu(*refcount), c,
1582 reflink_v_refcount_wrong,
1583 "reflink key has wrong refcount:\n"
1586 (bch2_bkey_val_to_text(&buf, c, k), buf.buf),
1588 struct bkey_i *new = bch2_bkey_make_mut(trans, iter, &k, 0);
1590 ret = PTR_ERR_OR_ZERO(new);
1595 new->k.type = KEY_TYPE_deleted;
1597 *bkey_refcount(bkey_i_to_s(new)) = cpu_to_le64(r->refcount);
1600 printbuf_exit(&buf);
1604 static int bch2_gc_reflink_done(struct bch_fs *c, bool metadata_only)
1611 int ret = bch2_trans_run(c,
1612 for_each_btree_key_commit(trans, iter,
1613 BTREE_ID_reflink, POS_MIN,
1614 BTREE_ITER_PREFETCH, k,
1615 NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
1616 bch2_gc_write_reflink_key(trans, &iter, k, &idx)));
1617 c->reflink_gc_nr = 0;
1621 static int bch2_gc_reflink_start(struct bch_fs *c,
1628 c->reflink_gc_nr = 0;
1630 int ret = bch2_trans_run(c,
1631 for_each_btree_key(trans, iter, BTREE_ID_reflink, POS_MIN,
1632 BTREE_ITER_PREFETCH, k, ({
1633 const __le64 *refcount = bkey_refcount_c(k);
1638 struct reflink_gc *r = genradix_ptr_alloc(&c->reflink_gc_table,
1639 c->reflink_gc_nr++, GFP_KERNEL);
1641 ret = -BCH_ERR_ENOMEM_gc_reflink_start;
1645 r->offset = k.k->p.offset;
1646 r->size = k.k->size;
1655 static void bch2_gc_reflink_reset(struct bch_fs *c, bool metadata_only)
1657 struct genradix_iter iter;
1658 struct reflink_gc *r;
1660 genradix_for_each(&c->reflink_gc_table, iter, r)
1664 static int bch2_gc_write_stripes_key(struct btree_trans *trans,
1665 struct btree_iter *iter,
1668 struct bch_fs *c = trans->c;
1669 struct printbuf buf = PRINTBUF;
1670 const struct bch_stripe *s;
1671 struct gc_stripe *m;
1676 if (k.k->type != KEY_TYPE_stripe)
1679 s = bkey_s_c_to_stripe(k).v;
1680 m = genradix_ptr(&c->gc_stripes, k.k->p.offset);
1682 for (i = 0; i < s->nr_blocks; i++) {
1683 u32 old = stripe_blockcount_get(s, i);
1684 u32 new = (m ? m->block_sectors[i] : 0);
1687 prt_printf(&buf, "stripe block %u has wrong sector count: got %u, should be %u\n",
1694 bch2_bkey_val_to_text(&buf, c, k);
1696 if (fsck_err_on(bad, c, stripe_sector_count_wrong,
1698 struct bkey_i_stripe *new;
1700 new = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1701 ret = PTR_ERR_OR_ZERO(new);
1705 bkey_reassemble(&new->k_i, k);
1707 for (i = 0; i < new->v.nr_blocks; i++)
1708 stripe_blockcount_set(&new->v, i, m ? m->block_sectors[i] : 0);
1710 ret = bch2_trans_update(trans, iter, &new->k_i, 0);
1713 printbuf_exit(&buf);
1717 static int bch2_gc_stripes_done(struct bch_fs *c, bool metadata_only)
1722 return bch2_trans_run(c,
1723 for_each_btree_key_commit(trans, iter,
1724 BTREE_ID_stripes, POS_MIN,
1725 BTREE_ITER_PREFETCH, k,
1726 NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
1727 bch2_gc_write_stripes_key(trans, &iter, k)));
1730 static void bch2_gc_stripes_reset(struct bch_fs *c, bool metadata_only)
1732 genradix_free(&c->gc_stripes);
1736 * bch2_gc - walk _all_ references to buckets, and recompute them:
1738 * @c: filesystem object
1739 * @initial: are we in recovery?
1740 * @metadata_only: are we just checking metadata references, or everything?
1742 * Returns: 0 on success, or standard errcode on failure
1744 * Order matters here:
1745 * - Concurrent GC relies on the fact that we have a total ordering for
1746 * everything that GC walks - see gc_will_visit_node(),
1747 * gc_will_visit_root()
1749 * - also, references move around in the course of index updates and
1750 * various other crap: everything needs to agree on the ordering
1751 * references are allowed to move around in - e.g., we're allowed to
1752 * start with a reference owned by an open_bucket (the allocator) and
1753 * move it to the btree, but not the reverse.
1755 * This is necessary to ensure that gc doesn't miss references that
1756 * move around - if references move backwards in the ordering GC
1757 * uses, GC could skip past them
1759 int bch2_gc(struct bch_fs *c, bool initial, bool metadata_only)
1764 lockdep_assert_held(&c->state_lock);
1766 down_write(&c->gc_lock);
1768 bch2_btree_interior_updates_flush(c);
1770 ret = bch2_gc_start(c) ?:
1771 bch2_gc_alloc_start(c, metadata_only) ?:
1772 bch2_gc_reflink_start(c, metadata_only);
1776 gc_pos_set(c, gc_phase(GC_PHASE_START));
1778 bch2_mark_superblocks(c);
1780 ret = bch2_gc_btrees(c, initial, metadata_only);
1786 bch2_mark_pending_btree_node_frees(c);
1790 if (test_bit(BCH_FS_need_another_gc, &c->flags) ||
1791 (!iter && bch2_test_restart_gc)) {
1793 bch_info(c, "Unable to fix bucket gens, looping");
1799 * XXX: make sure gens we fixed got saved
1801 bch_info(c, "Second GC pass needed, restarting:");
1802 clear_bit(BCH_FS_need_another_gc, &c->flags);
1803 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1805 bch2_gc_stripes_reset(c, metadata_only);
1806 bch2_gc_alloc_reset(c, metadata_only);
1807 bch2_gc_reflink_reset(c, metadata_only);
1808 ret = bch2_gc_reset(c);
1812 /* flush fsck errors, reset counters */
1813 bch2_flush_fsck_errs(c);
1818 bch2_journal_block(&c->journal);
1820 ret = bch2_gc_stripes_done(c, metadata_only) ?:
1821 bch2_gc_reflink_done(c, metadata_only) ?:
1822 bch2_gc_alloc_done(c, metadata_only) ?:
1823 bch2_gc_done(c, initial, metadata_only);
1825 bch2_journal_unblock(&c->journal);
1828 percpu_down_write(&c->mark_lock);
1829 /* Indicates that gc is no longer in progress: */
1830 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1833 percpu_up_write(&c->mark_lock);
1835 up_write(&c->gc_lock);
1838 * At startup, allocations can happen directly instead of via the
1839 * allocator thread - issue wakeup in case they blocked on gc_lock:
1841 closure_wake_up(&c->freelist_wait);
1846 static int gc_btree_gens_key(struct btree_trans *trans,
1847 struct btree_iter *iter,
1850 struct bch_fs *c = trans->c;
1851 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1855 percpu_down_read(&c->mark_lock);
1856 bkey_for_each_ptr(ptrs, ptr) {
1857 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1859 if (ptr_stale(ca, ptr) > 16) {
1860 percpu_up_read(&c->mark_lock);
1865 bkey_for_each_ptr(ptrs, ptr) {
1866 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1867 u8 *gen = &ca->oldest_gen[PTR_BUCKET_NR(ca, ptr)];
1869 if (gen_after(*gen, ptr->gen))
1872 percpu_up_read(&c->mark_lock);
1875 u = bch2_bkey_make_mut(trans, iter, &k, 0);
1876 ret = PTR_ERR_OR_ZERO(u);
1880 bch2_extent_normalize(c, bkey_i_to_s(u));
1884 static int bch2_alloc_write_oldest_gen(struct btree_trans *trans, struct btree_iter *iter,
1887 struct bch_dev *ca = bch_dev_bkey_exists(trans->c, iter->pos.inode);
1888 struct bch_alloc_v4 a_convert;
1889 const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &a_convert);
1890 struct bkey_i_alloc_v4 *a_mut;
1893 if (a->oldest_gen == ca->oldest_gen[iter->pos.offset])
1896 a_mut = bch2_alloc_to_v4_mut(trans, k);
1897 ret = PTR_ERR_OR_ZERO(a_mut);
1901 a_mut->v.oldest_gen = ca->oldest_gen[iter->pos.offset];
1902 a_mut->v.data_type = alloc_data_type(a_mut->v, a_mut->v.data_type);
1904 return bch2_trans_update(trans, iter, &a_mut->k_i, 0);
1907 int bch2_gc_gens(struct bch_fs *c)
1909 u64 b, start_time = local_clock();
1913 * Ideally we would be using state_lock and not gc_lock here, but that
1914 * introduces a deadlock in the RO path - we currently take the state
1915 * lock at the start of going RO, thus the gc thread may get stuck:
1917 if (!mutex_trylock(&c->gc_gens_lock))
1920 trace_and_count(c, gc_gens_start, c);
1921 down_read(&c->gc_lock);
1923 for_each_member_device(c, ca) {
1924 struct bucket_gens *gens = bucket_gens(ca);
1926 BUG_ON(ca->oldest_gen);
1928 ca->oldest_gen = kvmalloc(gens->nbuckets, GFP_KERNEL);
1929 if (!ca->oldest_gen) {
1930 percpu_ref_put(&ca->ref);
1931 ret = -BCH_ERR_ENOMEM_gc_gens;
1935 for (b = gens->first_bucket;
1936 b < gens->nbuckets; b++)
1937 ca->oldest_gen[b] = gens->b[b];
1940 for (unsigned i = 0; i < BTREE_ID_NR; i++)
1941 if (btree_type_has_ptrs(i)) {
1942 c->gc_gens_btree = i;
1943 c->gc_gens_pos = POS_MIN;
1945 ret = bch2_trans_run(c,
1946 for_each_btree_key_commit(trans, iter, i,
1948 BTREE_ITER_PREFETCH|BTREE_ITER_ALL_SNAPSHOTS,
1951 BCH_TRANS_COMMIT_no_enospc,
1952 gc_btree_gens_key(trans, &iter, k)));
1957 ret = bch2_trans_run(c,
1958 for_each_btree_key_commit(trans, iter, BTREE_ID_alloc,
1960 BTREE_ITER_PREFETCH,
1963 BCH_TRANS_COMMIT_no_enospc,
1964 bch2_alloc_write_oldest_gen(trans, &iter, k)));
1968 c->gc_gens_btree = 0;
1969 c->gc_gens_pos = POS_MIN;
1973 bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
1974 trace_and_count(c, gc_gens_end, c);
1976 for_each_member_device(c, ca) {
1977 kvfree(ca->oldest_gen);
1978 ca->oldest_gen = NULL;
1981 up_read(&c->gc_lock);
1982 mutex_unlock(&c->gc_gens_lock);
1983 if (!bch2_err_matches(ret, EROFS))
1988 static int bch2_gc_thread(void *arg)
1990 struct bch_fs *c = arg;
1991 struct io_clock *clock = &c->io_clock[WRITE];
1992 unsigned long last = atomic64_read(&clock->now);
1993 unsigned last_kick = atomic_read(&c->kick_gc);
1999 set_current_state(TASK_INTERRUPTIBLE);
2001 if (kthread_should_stop()) {
2002 __set_current_state(TASK_RUNNING);
2006 if (atomic_read(&c->kick_gc) != last_kick)
2009 if (c->btree_gc_periodic) {
2010 unsigned long next = last + c->capacity / 16;
2012 if (atomic64_read(&clock->now) >= next)
2015 bch2_io_clock_schedule_timeout(clock, next);
2022 __set_current_state(TASK_RUNNING);
2024 last = atomic64_read(&clock->now);
2025 last_kick = atomic_read(&c->kick_gc);
2028 * Full gc is currently incompatible with btree key cache:
2031 ret = bch2_gc(c, false, false);
2035 debug_check_no_locks_held();
2041 void bch2_gc_thread_stop(struct bch_fs *c)
2043 struct task_struct *p;
2046 c->gc_thread = NULL;
2054 int bch2_gc_thread_start(struct bch_fs *c)
2056 struct task_struct *p;
2061 p = kthread_create(bch2_gc_thread, c, "bch-gc/%s", c->name);
2063 bch_err_fn(c, PTR_ERR(p));