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 bool should_restart_for_topology_repair(struct bch_fs *c)
46 return c->opts.fix_errors != FSCK_FIX_no &&
47 !(c->recovery_passes_complete & BIT_ULL(BCH_RECOVERY_PASS_check_topology));
50 static inline void __gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
53 write_seqcount_begin(&c->gc_pos_lock);
55 write_seqcount_end(&c->gc_pos_lock);
59 static inline void gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
61 BUG_ON(gc_pos_cmp(new_pos, c->gc_pos) <= 0);
62 __gc_pos_set(c, new_pos);
66 * Missing: if an interior btree node is empty, we need to do something -
67 * perhaps just kill it
69 static int bch2_gc_check_topology(struct bch_fs *c,
71 struct bkey_buf *prev,
75 struct bpos node_start = b->data->min_key;
76 struct bpos node_end = b->data->max_key;
77 struct bpos expected_start = bkey_deleted(&prev->k->k)
79 : bpos_successor(prev->k->k.p);
80 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
83 if (cur.k->k.type == KEY_TYPE_btree_ptr_v2) {
84 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(cur.k);
86 if (!bpos_eq(expected_start, bp->v.min_key)) {
87 bch2_topology_error(c);
89 if (bkey_deleted(&prev->k->k)) {
90 prt_printf(&buf1, "start of node: ");
91 bch2_bpos_to_text(&buf1, node_start);
93 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(prev->k));
95 bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(cur.k));
101 btree_node_topology_bad_min_key,
102 "btree node with incorrect min_key at btree %s level %u:\n"
105 bch2_btree_id_str(b->c.btree_id), b->c.level,
106 buf1.buf, buf2.buf) && should_restart_for_topology_repair(c)) {
107 bch_info(c, "Halting mark and sweep to start topology repair pass");
108 ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_topology);
111 set_bit(BCH_FS_initial_gc_unfixed, &c->flags);
116 if (is_last && !bpos_eq(cur.k->k.p, node_end)) {
117 bch2_topology_error(c);
119 printbuf_reset(&buf1);
120 printbuf_reset(&buf2);
122 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(cur.k));
123 bch2_bpos_to_text(&buf2, node_end);
125 if (__fsck_err(c, FSCK_CAN_FIX|FSCK_CAN_IGNORE|FSCK_NO_RATELIMIT,
126 btree_node_topology_bad_max_key,
127 "btree node with incorrect max_key at btree %s level %u:\n"
130 bch2_btree_id_str(b->c.btree_id), b->c.level,
131 buf1.buf, buf2.buf) &&
132 should_restart_for_topology_repair(c)) {
133 bch_info(c, "Halting mark and sweep to start topology repair pass");
134 ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_topology);
137 set_bit(BCH_FS_initial_gc_unfixed, &c->flags);
141 bch2_bkey_buf_copy(prev, c, cur.k);
144 printbuf_exit(&buf2);
145 printbuf_exit(&buf1);
149 static void btree_ptr_to_v2(struct btree *b, struct bkey_i_btree_ptr_v2 *dst)
151 switch (b->key.k.type) {
152 case KEY_TYPE_btree_ptr: {
153 struct bkey_i_btree_ptr *src = bkey_i_to_btree_ptr(&b->key);
157 dst->v.seq = b->data->keys.seq;
158 dst->v.sectors_written = 0;
160 dst->v.min_key = b->data->min_key;
161 set_bkey_val_bytes(&dst->k, sizeof(dst->v) + bkey_val_bytes(&src->k));
162 memcpy(dst->v.start, src->v.start, bkey_val_bytes(&src->k));
165 case KEY_TYPE_btree_ptr_v2:
166 bkey_copy(&dst->k_i, &b->key);
173 static void bch2_btree_node_update_key_early(struct btree_trans *trans,
174 enum btree_id btree, unsigned level,
175 struct bkey_s_c old, struct bkey_i *new)
177 struct bch_fs *c = trans->c;
182 bch2_bkey_buf_init(&tmp);
183 bch2_bkey_buf_reassemble(&tmp, c, old);
185 b = bch2_btree_node_get_noiter(trans, tmp.k, btree, level, true);
186 if (!IS_ERR_OR_NULL(b)) {
187 mutex_lock(&c->btree_cache.lock);
189 bch2_btree_node_hash_remove(&c->btree_cache, b);
191 bkey_copy(&b->key, new);
192 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
195 mutex_unlock(&c->btree_cache.lock);
196 six_unlock_read(&b->c.lock);
199 bch2_bkey_buf_exit(&tmp, c);
202 static int set_node_min(struct bch_fs *c, struct btree *b, struct bpos new_min)
204 struct bkey_i_btree_ptr_v2 *new;
207 new = kmalloc_array(BKEY_BTREE_PTR_U64s_MAX, sizeof(u64), GFP_KERNEL);
209 return -BCH_ERR_ENOMEM_gc_repair_key;
211 btree_ptr_to_v2(b, new);
212 b->data->min_key = new_min;
213 new->v.min_key = new_min;
214 SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
216 ret = bch2_journal_key_insert_take(c, b->c.btree_id, b->c.level + 1, &new->k_i);
222 bch2_btree_node_drop_keys_outside_node(b);
223 bkey_copy(&b->key, &new->k_i);
227 static int set_node_max(struct bch_fs *c, struct btree *b, struct bpos new_max)
229 struct bkey_i_btree_ptr_v2 *new;
232 ret = bch2_journal_key_delete(c, b->c.btree_id, b->c.level + 1, b->key.k.p);
236 new = kmalloc_array(BKEY_BTREE_PTR_U64s_MAX, sizeof(u64), GFP_KERNEL);
238 return -BCH_ERR_ENOMEM_gc_repair_key;
240 btree_ptr_to_v2(b, new);
241 b->data->max_key = new_max;
243 SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
245 ret = bch2_journal_key_insert_take(c, b->c.btree_id, b->c.level + 1, &new->k_i);
251 bch2_btree_node_drop_keys_outside_node(b);
253 mutex_lock(&c->btree_cache.lock);
254 bch2_btree_node_hash_remove(&c->btree_cache, b);
256 bkey_copy(&b->key, &new->k_i);
257 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
259 mutex_unlock(&c->btree_cache.lock);
263 static int btree_repair_node_boundaries(struct bch_fs *c, struct btree *b,
264 struct btree *prev, struct btree *cur)
266 struct bpos expected_start = !prev
268 : bpos_successor(prev->key.k.p);
269 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
273 prt_printf(&buf1, "start of node: ");
274 bch2_bpos_to_text(&buf1, b->data->min_key);
276 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(&prev->key));
279 bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(&cur->key));
282 bpos_gt(expected_start, cur->data->min_key) &&
283 BTREE_NODE_SEQ(cur->data) > BTREE_NODE_SEQ(prev->data)) {
284 /* cur overwrites prev: */
286 if (mustfix_fsck_err_on(bpos_ge(prev->data->min_key,
287 cur->data->min_key), c,
288 btree_node_topology_overwritten_by_next_node,
289 "btree node overwritten by next node at btree %s level %u:\n"
292 bch2_btree_id_str(b->c.btree_id), b->c.level,
293 buf1.buf, buf2.buf)) {
294 ret = DROP_PREV_NODE;
298 if (mustfix_fsck_err_on(!bpos_eq(prev->key.k.p,
299 bpos_predecessor(cur->data->min_key)), c,
300 btree_node_topology_bad_max_key,
301 "btree node with incorrect max_key at btree %s level %u:\n"
304 bch2_btree_id_str(b->c.btree_id), b->c.level,
306 ret = set_node_max(c, prev,
307 bpos_predecessor(cur->data->min_key));
309 /* prev overwrites cur: */
311 if (mustfix_fsck_err_on(bpos_ge(expected_start,
312 cur->data->max_key), c,
313 btree_node_topology_overwritten_by_prev_node,
314 "btree node overwritten by prev node at btree %s level %u:\n"
317 bch2_btree_id_str(b->c.btree_id), b->c.level,
318 buf1.buf, buf2.buf)) {
319 ret = DROP_THIS_NODE;
323 if (mustfix_fsck_err_on(!bpos_eq(expected_start, cur->data->min_key), c,
324 btree_node_topology_bad_min_key,
325 "btree node with incorrect min_key at btree %s level %u:\n"
328 bch2_btree_id_str(b->c.btree_id), b->c.level,
330 ret = set_node_min(c, cur, expected_start);
334 printbuf_exit(&buf2);
335 printbuf_exit(&buf1);
339 static int btree_repair_node_end(struct bch_fs *c, struct btree *b,
342 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
345 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(&child->key));
346 bch2_bpos_to_text(&buf2, b->key.k.p);
348 if (mustfix_fsck_err_on(!bpos_eq(child->key.k.p, b->key.k.p), c,
349 btree_node_topology_bad_max_key,
350 "btree node with incorrect max_key at btree %s level %u:\n"
353 bch2_btree_id_str(b->c.btree_id), b->c.level,
354 buf1.buf, buf2.buf)) {
355 ret = set_node_max(c, child, b->key.k.p);
361 printbuf_exit(&buf2);
362 printbuf_exit(&buf1);
366 static int bch2_btree_repair_topology_recurse(struct btree_trans *trans, struct btree *b)
368 struct bch_fs *c = trans->c;
369 struct btree_and_journal_iter iter;
371 struct bkey_buf prev_k, cur_k;
372 struct btree *prev = NULL, *cur = NULL;
373 bool have_child, dropped_children = false;
374 struct printbuf buf = PRINTBUF;
381 have_child = dropped_children = false;
382 bch2_bkey_buf_init(&prev_k);
383 bch2_bkey_buf_init(&cur_k);
384 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
386 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
387 BUG_ON(bpos_lt(k.k->p, b->data->min_key));
388 BUG_ON(bpos_gt(k.k->p, b->data->max_key));
390 bch2_btree_and_journal_iter_advance(&iter);
391 bch2_bkey_buf_reassemble(&cur_k, c, k);
393 cur = bch2_btree_node_get_noiter(trans, cur_k.k,
394 b->c.btree_id, b->c.level - 1,
396 ret = PTR_ERR_OR_ZERO(cur);
398 printbuf_reset(&buf);
399 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(cur_k.k));
401 if (mustfix_fsck_err_on(ret == -EIO, c,
402 btree_node_unreadable,
403 "Topology repair: unreadable btree node at btree %s level %u:\n"
405 bch2_btree_id_str(b->c.btree_id),
408 bch2_btree_node_evict(trans, cur_k.k);
409 ret = bch2_journal_key_delete(c, b->c.btree_id,
410 b->c.level, cur_k.k->k.p);
417 bch_err_msg(c, ret, "getting btree node");
421 ret = btree_repair_node_boundaries(c, b, prev, cur);
423 if (ret == DROP_THIS_NODE) {
424 six_unlock_read(&cur->c.lock);
425 bch2_btree_node_evict(trans, cur_k.k);
426 ret = bch2_journal_key_delete(c, b->c.btree_id,
427 b->c.level, cur_k.k->k.p);
435 six_unlock_read(&prev->c.lock);
438 if (ret == DROP_PREV_NODE) {
439 bch2_btree_node_evict(trans, prev_k.k);
440 ret = bch2_journal_key_delete(c, b->c.btree_id,
441 b->c.level, prev_k.k->k.p);
445 bch2_btree_and_journal_iter_exit(&iter);
446 bch2_bkey_buf_exit(&prev_k, c);
447 bch2_bkey_buf_exit(&cur_k, c);
454 bch2_bkey_buf_copy(&prev_k, c, cur_k.k);
457 if (!ret && !IS_ERR_OR_NULL(prev)) {
459 ret = btree_repair_node_end(c, b, prev);
462 if (!IS_ERR_OR_NULL(prev))
463 six_unlock_read(&prev->c.lock);
465 if (!IS_ERR_OR_NULL(cur))
466 six_unlock_read(&cur->c.lock);
472 bch2_btree_and_journal_iter_exit(&iter);
473 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
475 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
476 bch2_bkey_buf_reassemble(&cur_k, c, k);
477 bch2_btree_and_journal_iter_advance(&iter);
479 cur = bch2_btree_node_get_noiter(trans, cur_k.k,
480 b->c.btree_id, b->c.level - 1,
482 ret = PTR_ERR_OR_ZERO(cur);
484 bch_err_msg(c, ret, "getting btree node");
488 ret = bch2_btree_repair_topology_recurse(trans, cur);
489 six_unlock_read(&cur->c.lock);
492 if (ret == DROP_THIS_NODE) {
493 bch2_btree_node_evict(trans, cur_k.k);
494 ret = bch2_journal_key_delete(c, b->c.btree_id,
495 b->c.level, cur_k.k->k.p);
496 dropped_children = true;
505 printbuf_reset(&buf);
506 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
508 if (mustfix_fsck_err_on(!have_child, c,
509 btree_node_topology_interior_node_empty,
510 "empty interior btree node at btree %s level %u\n"
512 bch2_btree_id_str(b->c.btree_id),
513 b->c.level, buf.buf))
514 ret = DROP_THIS_NODE;
517 if (!IS_ERR_OR_NULL(prev))
518 six_unlock_read(&prev->c.lock);
519 if (!IS_ERR_OR_NULL(cur))
520 six_unlock_read(&cur->c.lock);
522 bch2_btree_and_journal_iter_exit(&iter);
523 bch2_bkey_buf_exit(&prev_k, c);
524 bch2_bkey_buf_exit(&cur_k, c);
526 if (!ret && dropped_children)
533 int bch2_check_topology(struct bch_fs *c)
535 struct btree_trans *trans = bch2_trans_get(c);
540 for (i = 0; i < btree_id_nr_alive(c) && !ret; i++) {
541 struct btree_root *r = bch2_btree_id_root(c, i);
547 if (btree_node_fake(b))
550 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
551 ret = bch2_btree_repair_topology_recurse(trans, b);
552 six_unlock_read(&b->c.lock);
554 if (ret == DROP_THIS_NODE) {
555 bch_err(c, "empty btree root - repair unimplemented");
556 ret = -BCH_ERR_fsck_repair_unimplemented;
560 bch2_trans_put(trans);
565 static int bch2_check_fix_ptrs(struct btree_trans *trans, enum btree_id btree_id,
566 unsigned level, bool is_root,
569 struct bch_fs *c = trans->c;
570 struct bkey_ptrs_c ptrs_c = bch2_bkey_ptrs_c(*k);
571 const union bch_extent_entry *entry_c;
572 struct extent_ptr_decoded p = { 0 };
573 bool do_update = false;
574 struct printbuf buf = PRINTBUF;
579 * use check_bucket_ref here
581 bkey_for_each_ptr_decode(k->k, ptrs_c, p, entry_c) {
582 struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
583 struct bucket *g = PTR_GC_BUCKET(ca, &p.ptr);
584 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, &entry_c->ptr);
587 (c->opts.reconstruct_alloc ||
588 fsck_err(c, ptr_to_missing_alloc_key,
589 "bucket %u:%zu data type %s ptr gen %u missing in alloc btree\n"
591 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
592 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
594 (printbuf_reset(&buf),
595 bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))) {
604 if (gen_cmp(p.ptr.gen, g->gen) > 0 &&
605 (c->opts.reconstruct_alloc ||
606 fsck_err(c, ptr_gen_newer_than_bucket_gen,
607 "bucket %u:%zu data type %s ptr gen in the future: %u > %u\n"
609 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
610 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
612 (printbuf_reset(&buf),
613 bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))) {
618 g->dirty_sectors = 0;
619 g->cached_sectors = 0;
620 set_bit(BCH_FS_need_another_gc, &c->flags);
626 if (gen_cmp(g->gen, p.ptr.gen) > BUCKET_GC_GEN_MAX &&
627 (c->opts.reconstruct_alloc ||
628 fsck_err(c, ptr_gen_newer_than_bucket_gen,
629 "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
631 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr), g->gen,
632 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
634 (printbuf_reset(&buf),
635 bch2_bkey_val_to_text(&buf, c, *k), buf.buf))))
638 if (!p.ptr.cached && gen_cmp(p.ptr.gen, g->gen) < 0 &&
639 (c->opts.reconstruct_alloc ||
640 fsck_err(c, stale_dirty_ptr,
641 "bucket %u:%zu data type %s stale dirty ptr: %u < %u\n"
643 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
644 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
646 (printbuf_reset(&buf),
647 bch2_bkey_val_to_text(&buf, c, *k), buf.buf))))
650 if (data_type != BCH_DATA_btree && p.ptr.gen != g->gen)
653 if (fsck_err_on(bucket_data_type(g->data_type) &&
654 bucket_data_type(g->data_type) != data_type, c,
655 ptr_bucket_data_type_mismatch,
656 "bucket %u:%zu different types of data in same bucket: %s, %s\n"
658 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
659 bch2_data_types[g->data_type],
660 bch2_data_types[data_type],
661 (printbuf_reset(&buf),
662 bch2_bkey_val_to_text(&buf, c, *k), buf.buf))) {
663 if (data_type == BCH_DATA_btree) {
664 g->data_type = data_type;
665 set_bit(BCH_FS_need_another_gc, &c->flags);
672 struct gc_stripe *m = genradix_ptr(&c->gc_stripes, p.ec.idx);
674 if (fsck_err_on(!m || !m->alive, c,
675 ptr_to_missing_stripe,
676 "pointer to nonexistent stripe %llu\n"
679 (printbuf_reset(&buf),
680 bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
683 if (fsck_err_on(m && m->alive && !bch2_ptr_matches_stripe_m(m, p), c,
684 ptr_to_incorrect_stripe,
685 "pointer does not match stripe %llu\n"
688 (printbuf_reset(&buf),
689 bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
695 struct bkey_ptrs ptrs;
696 union bch_extent_entry *entry;
697 struct bch_extent_ptr *ptr;
701 bch_err(c, "cannot update btree roots yet");
706 new = kmalloc(bkey_bytes(k->k), GFP_KERNEL);
708 ret = -BCH_ERR_ENOMEM_gc_repair_key;
709 bch_err_msg(c, ret, "allocating new key");
713 bkey_reassemble(new, *k);
717 * We don't want to drop btree node pointers - if the
718 * btree node isn't there anymore, the read path will
721 ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
722 bkey_for_each_ptr(ptrs, ptr) {
723 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
724 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
729 bch2_bkey_drop_ptrs(bkey_i_to_s(new), ptr, ({
730 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
731 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
732 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, ptr);
735 (!g->gen_valid || gen_cmp(ptr->gen, g->gen) > 0)) ||
737 gen_cmp(ptr->gen, g->gen) < 0) ||
738 gen_cmp(g->gen, ptr->gen) > BUCKET_GC_GEN_MAX ||
740 g->data_type != data_type);
743 ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
744 bkey_extent_entry_for_each(ptrs, entry) {
745 if (extent_entry_type(entry) == BCH_EXTENT_ENTRY_stripe_ptr) {
746 struct gc_stripe *m = genradix_ptr(&c->gc_stripes,
747 entry->stripe_ptr.idx);
748 union bch_extent_entry *next_ptr;
750 bkey_extent_entry_for_each_from(ptrs, next_ptr, entry)
751 if (extent_entry_type(next_ptr) == BCH_EXTENT_ENTRY_ptr)
756 bch_err(c, "aieee, found stripe ptr with no data ptr");
760 if (!m || !m->alive ||
761 !__bch2_ptr_matches_stripe(&m->ptrs[entry->stripe_ptr.block],
764 bch2_bkey_extent_entry_drop(new, entry);
771 ret = bch2_journal_key_insert_take(c, btree_id, level, new);
778 bch2_btree_node_update_key_early(trans, btree_id, level - 1, *k, new);
781 printbuf_reset(&buf);
782 bch2_bkey_val_to_text(&buf, c, *k);
783 bch_info(c, "updated %s", buf.buf);
785 printbuf_reset(&buf);
786 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(new));
787 bch_info(c, "new key %s", buf.buf);
790 *k = bkey_i_to_s_c(new);
798 /* marking of btree keys/nodes: */
800 static int bch2_gc_mark_key(struct btree_trans *trans, enum btree_id btree_id,
801 unsigned level, bool is_root,
805 struct bch_fs *c = trans->c;
806 struct bkey deleted = KEY(0, 0, 0);
807 struct bkey_s_c old = (struct bkey_s_c) { &deleted, NULL };
810 (initial ? BTREE_TRIGGER_NOATOMIC : 0);
816 BUG_ON(bch2_journal_seq_verify &&
817 k->k->version.lo > atomic64_read(&c->journal.seq));
819 ret = bch2_check_fix_ptrs(trans, btree_id, level, is_root, k);
823 if (fsck_err_on(k->k->version.lo > atomic64_read(&c->key_version), c,
824 bkey_version_in_future,
825 "key version number higher than recorded: %llu > %llu",
827 atomic64_read(&c->key_version)))
828 atomic64_set(&c->key_version, k->k->version.lo);
831 ret = commit_do(trans, NULL, NULL, 0,
832 bch2_mark_key(trans, btree_id, level, old, *k, flags));
839 static int btree_gc_mark_node(struct btree_trans *trans, struct btree *b, bool initial)
841 struct bch_fs *c = trans->c;
842 struct btree_node_iter iter;
843 struct bkey unpacked;
845 struct bkey_buf prev, cur;
848 if (!btree_node_type_needs_gc(btree_node_type(b)))
851 bch2_btree_node_iter_init_from_start(&iter, b);
852 bch2_bkey_buf_init(&prev);
853 bch2_bkey_buf_init(&cur);
854 bkey_init(&prev.k->k);
856 while ((k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked)).k) {
857 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, false,
862 bch2_btree_node_iter_advance(&iter, b);
865 bch2_bkey_buf_reassemble(&cur, c, k);
867 ret = bch2_gc_check_topology(c, b, &prev, cur,
868 bch2_btree_node_iter_end(&iter));
874 bch2_bkey_buf_exit(&cur, c);
875 bch2_bkey_buf_exit(&prev, c);
879 static int bch2_gc_btree(struct btree_trans *trans, enum btree_id btree_id,
880 bool initial, bool metadata_only)
882 struct bch_fs *c = trans->c;
883 struct btree_iter iter;
885 unsigned depth = metadata_only ? 1 : 0;
888 gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));
890 __for_each_btree_node(trans, iter, btree_id, POS_MIN,
891 0, depth, BTREE_ITER_PREFETCH, b, ret) {
892 bch2_verify_btree_nr_keys(b);
894 gc_pos_set(c, gc_pos_btree_node(b));
896 ret = btree_gc_mark_node(trans, b, initial);
900 bch2_trans_iter_exit(trans, &iter);
905 mutex_lock(&c->btree_root_lock);
906 b = bch2_btree_id_root(c, btree_id)->b;
907 if (!btree_node_fake(b)) {
908 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
910 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level + 1,
913 gc_pos_set(c, gc_pos_btree_root(b->c.btree_id));
914 mutex_unlock(&c->btree_root_lock);
919 static int bch2_gc_btree_init_recurse(struct btree_trans *trans, struct btree *b,
920 unsigned target_depth)
922 struct bch_fs *c = trans->c;
923 struct btree_and_journal_iter iter;
925 struct bkey_buf cur, prev;
926 struct printbuf buf = PRINTBUF;
929 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
930 bch2_bkey_buf_init(&prev);
931 bch2_bkey_buf_init(&cur);
932 bkey_init(&prev.k->k);
934 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
935 BUG_ON(bpos_lt(k.k->p, b->data->min_key));
936 BUG_ON(bpos_gt(k.k->p, b->data->max_key));
938 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level,
944 bch2_bkey_buf_reassemble(&cur, c, k);
945 k = bkey_i_to_s_c(cur.k);
947 bch2_btree_and_journal_iter_advance(&iter);
949 ret = bch2_gc_check_topology(c, b,
951 !bch2_btree_and_journal_iter_peek(&iter).k);
955 bch2_btree_and_journal_iter_advance(&iter);
959 if (b->c.level > target_depth) {
960 bch2_btree_and_journal_iter_exit(&iter);
961 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
963 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
966 bch2_bkey_buf_reassemble(&cur, c, k);
967 bch2_btree_and_journal_iter_advance(&iter);
969 child = bch2_btree_node_get_noiter(trans, cur.k,
970 b->c.btree_id, b->c.level - 1,
972 ret = PTR_ERR_OR_ZERO(child);
975 bch2_topology_error(c);
981 btree_node_read_error,
982 "Unreadable btree node at btree %s level %u:\n"
984 bch2_btree_id_str(b->c.btree_id),
986 (printbuf_reset(&buf),
987 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(cur.k)), buf.buf)) &&
988 should_restart_for_topology_repair(c)) {
989 bch_info(c, "Halting mark and sweep to start topology repair pass");
990 ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_topology);
993 /* Continue marking when opted to not
996 set_bit(BCH_FS_initial_gc_unfixed, &c->flags);
1000 bch_err_msg(c, ret, "getting btree node");
1004 ret = bch2_gc_btree_init_recurse(trans, child,
1006 six_unlock_read(&child->c.lock);
1013 bch2_bkey_buf_exit(&cur, c);
1014 bch2_bkey_buf_exit(&prev, c);
1015 bch2_btree_and_journal_iter_exit(&iter);
1016 printbuf_exit(&buf);
1020 static int bch2_gc_btree_init(struct btree_trans *trans,
1021 enum btree_id btree_id,
1024 struct bch_fs *c = trans->c;
1026 unsigned target_depth = metadata_only ? 1 : 0;
1027 struct printbuf buf = PRINTBUF;
1030 b = bch2_btree_id_root(c, btree_id)->b;
1032 if (btree_node_fake(b))
1035 six_lock_read(&b->c.lock, NULL, NULL);
1036 printbuf_reset(&buf);
1037 bch2_bpos_to_text(&buf, b->data->min_key);
1038 if (mustfix_fsck_err_on(!bpos_eq(b->data->min_key, POS_MIN), c,
1039 btree_root_bad_min_key,
1040 "btree root with incorrect min_key: %s", buf.buf)) {
1041 bch_err(c, "repair unimplemented");
1042 ret = -BCH_ERR_fsck_repair_unimplemented;
1046 printbuf_reset(&buf);
1047 bch2_bpos_to_text(&buf, b->data->max_key);
1048 if (mustfix_fsck_err_on(!bpos_eq(b->data->max_key, SPOS_MAX), c,
1049 btree_root_bad_max_key,
1050 "btree root with incorrect max_key: %s", buf.buf)) {
1051 bch_err(c, "repair unimplemented");
1052 ret = -BCH_ERR_fsck_repair_unimplemented;
1056 if (b->c.level >= target_depth)
1057 ret = bch2_gc_btree_init_recurse(trans, b, target_depth);
1060 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1062 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level + 1, true,
1066 six_unlock_read(&b->c.lock);
1069 printbuf_exit(&buf);
1073 static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
1075 return (int) btree_id_to_gc_phase(l) -
1076 (int) btree_id_to_gc_phase(r);
1079 static int bch2_gc_btrees(struct bch_fs *c, bool initial, bool metadata_only)
1081 struct btree_trans *trans = bch2_trans_get(c);
1082 enum btree_id ids[BTREE_ID_NR];
1086 for (i = 0; i < BTREE_ID_NR; i++)
1088 bubble_sort(ids, BTREE_ID_NR, btree_id_gc_phase_cmp);
1090 for (i = 0; i < BTREE_ID_NR && !ret; i++)
1092 ? bch2_gc_btree_init(trans, ids[i], metadata_only)
1093 : bch2_gc_btree(trans, ids[i], initial, metadata_only);
1095 for (i = BTREE_ID_NR; i < btree_id_nr_alive(c) && !ret; i++) {
1096 if (!bch2_btree_id_root(c, i)->alive)
1100 ? bch2_gc_btree_init(trans, i, metadata_only)
1101 : bch2_gc_btree(trans, i, initial, metadata_only);
1104 bch2_trans_put(trans);
1109 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
1111 enum bch_data_type type,
1114 u64 b = sector_to_bucket(ca, start);
1118 min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1120 bch2_mark_metadata_bucket(c, ca, b, type, sectors,
1121 gc_phase(GC_PHASE_SB), flags);
1124 } while (start < end);
1127 static void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
1130 struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1134 for (i = 0; i < layout->nr_superblocks; i++) {
1135 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1137 if (offset == BCH_SB_SECTOR)
1138 mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
1139 BCH_DATA_sb, flags);
1141 mark_metadata_sectors(c, ca, offset,
1142 offset + (1 << layout->sb_max_size_bits),
1143 BCH_DATA_sb, flags);
1146 for (i = 0; i < ca->journal.nr; i++) {
1147 b = ca->journal.buckets[i];
1148 bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_journal,
1150 gc_phase(GC_PHASE_SB), flags);
1154 static void bch2_mark_superblocks(struct bch_fs *c)
1156 mutex_lock(&c->sb_lock);
1157 gc_pos_set(c, gc_phase(GC_PHASE_SB));
1159 for_each_online_member(c, ca)
1160 bch2_mark_dev_superblock(c, ca, BTREE_TRIGGER_GC);
1161 mutex_unlock(&c->sb_lock);
1165 /* Also see bch2_pending_btree_node_free_insert_done() */
1166 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
1168 struct btree_update *as;
1169 struct pending_btree_node_free *d;
1171 mutex_lock(&c->btree_interior_update_lock);
1172 gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
1174 for_each_pending_btree_node_free(c, as, d)
1175 if (d->index_update_done)
1176 bch2_mark_key(c, bkey_i_to_s_c(&d->key), BTREE_TRIGGER_GC);
1178 mutex_unlock(&c->btree_interior_update_lock);
1182 static void bch2_gc_free(struct bch_fs *c)
1184 genradix_free(&c->reflink_gc_table);
1185 genradix_free(&c->gc_stripes);
1187 for_each_member_device(c, ca) {
1188 kvpfree(rcu_dereference_protected(ca->buckets_gc, 1),
1189 sizeof(struct bucket_array) +
1190 ca->mi.nbuckets * sizeof(struct bucket));
1191 ca->buckets_gc = NULL;
1193 free_percpu(ca->usage_gc);
1194 ca->usage_gc = NULL;
1197 free_percpu(c->usage_gc);
1201 static int bch2_gc_done(struct bch_fs *c,
1202 bool initial, bool metadata_only)
1204 struct bch_dev *ca = NULL;
1205 struct printbuf buf = PRINTBUF;
1206 bool verify = !metadata_only &&
1207 !c->opts.reconstruct_alloc &&
1208 (!initial || (c->sb.compat & (1ULL << BCH_COMPAT_alloc_info)));
1212 percpu_down_write(&c->mark_lock);
1214 #define copy_field(_err, _f, _msg, ...) \
1215 if (dst->_f != src->_f && \
1217 fsck_err(c, _err, _msg ": got %llu, should be %llu" \
1218 , ##__VA_ARGS__, dst->_f, src->_f))) \
1220 #define copy_dev_field(_err, _f, _msg, ...) \
1221 copy_field(_err, _f, "dev %u has wrong " _msg, ca->dev_idx, ##__VA_ARGS__)
1222 #define copy_fs_field(_err, _f, _msg, ...) \
1223 copy_field(_err, _f, "fs has wrong " _msg, ##__VA_ARGS__)
1225 for (i = 0; i < ARRAY_SIZE(c->usage); i++)
1226 bch2_fs_usage_acc_to_base(c, i);
1228 __for_each_member_device(c, ca) {
1229 struct bch_dev_usage *dst = ca->usage_base;
1230 struct bch_dev_usage *src = (void *)
1231 bch2_acc_percpu_u64s((u64 __percpu *) ca->usage_gc,
1234 for (i = 0; i < BCH_DATA_NR; i++) {
1235 copy_dev_field(dev_usage_buckets_wrong,
1236 d[i].buckets, "%s buckets", bch2_data_types[i]);
1237 copy_dev_field(dev_usage_sectors_wrong,
1238 d[i].sectors, "%s sectors", bch2_data_types[i]);
1239 copy_dev_field(dev_usage_fragmented_wrong,
1240 d[i].fragmented, "%s fragmented", bch2_data_types[i]);
1245 unsigned nr = fs_usage_u64s(c);
1246 struct bch_fs_usage *dst = c->usage_base;
1247 struct bch_fs_usage *src = (void *)
1248 bch2_acc_percpu_u64s((u64 __percpu *) c->usage_gc, nr);
1250 copy_fs_field(fs_usage_hidden_wrong,
1252 copy_fs_field(fs_usage_btree_wrong,
1255 if (!metadata_only) {
1256 copy_fs_field(fs_usage_data_wrong,
1258 copy_fs_field(fs_usage_cached_wrong,
1260 copy_fs_field(fs_usage_reserved_wrong,
1261 reserved, "reserved");
1262 copy_fs_field(fs_usage_nr_inodes_wrong,
1263 nr_inodes,"nr_inodes");
1265 for (i = 0; i < BCH_REPLICAS_MAX; i++)
1266 copy_fs_field(fs_usage_persistent_reserved_wrong,
1267 persistent_reserved[i],
1268 "persistent_reserved[%i]", i);
1271 for (i = 0; i < c->replicas.nr; i++) {
1272 struct bch_replicas_entry_v1 *e =
1273 cpu_replicas_entry(&c->replicas, i);
1275 if (metadata_only &&
1276 (e->data_type == BCH_DATA_user ||
1277 e->data_type == BCH_DATA_cached))
1280 printbuf_reset(&buf);
1281 bch2_replicas_entry_to_text(&buf, e);
1283 copy_fs_field(fs_usage_replicas_wrong,
1284 replicas[i], "%s", buf.buf);
1288 #undef copy_fs_field
1289 #undef copy_dev_field
1290 #undef copy_stripe_field
1294 percpu_ref_put(&ca->ref);
1297 percpu_up_write(&c->mark_lock);
1298 printbuf_exit(&buf);
1302 static int bch2_gc_start(struct bch_fs *c)
1304 BUG_ON(c->usage_gc);
1306 c->usage_gc = __alloc_percpu_gfp(fs_usage_u64s(c) * sizeof(u64),
1307 sizeof(u64), GFP_KERNEL);
1309 bch_err(c, "error allocating c->usage_gc");
1310 return -BCH_ERR_ENOMEM_gc_start;
1313 for_each_member_device(c, ca) {
1314 BUG_ON(ca->usage_gc);
1316 ca->usage_gc = alloc_percpu(struct bch_dev_usage);
1317 if (!ca->usage_gc) {
1318 bch_err(c, "error allocating ca->usage_gc");
1319 percpu_ref_put(&ca->ref);
1320 return -BCH_ERR_ENOMEM_gc_start;
1323 this_cpu_write(ca->usage_gc->d[BCH_DATA_free].buckets,
1324 ca->mi.nbuckets - ca->mi.first_bucket);
1330 static int bch2_gc_reset(struct bch_fs *c)
1332 for_each_member_device(c, ca) {
1333 free_percpu(ca->usage_gc);
1334 ca->usage_gc = NULL;
1337 free_percpu(c->usage_gc);
1340 return bch2_gc_start(c);
1343 /* returns true if not equal */
1344 static inline bool bch2_alloc_v4_cmp(struct bch_alloc_v4 l,
1345 struct bch_alloc_v4 r)
1347 return l.gen != r.gen ||
1348 l.oldest_gen != r.oldest_gen ||
1349 l.data_type != r.data_type ||
1350 l.dirty_sectors != r.dirty_sectors ||
1351 l.cached_sectors != r.cached_sectors ||
1352 l.stripe_redundancy != r.stripe_redundancy ||
1353 l.stripe != r.stripe;
1356 static int bch2_alloc_write_key(struct btree_trans *trans,
1357 struct btree_iter *iter,
1361 struct bch_fs *c = trans->c;
1362 struct bch_dev *ca = bch_dev_bkey_exists(c, iter->pos.inode);
1363 struct bucket gc, *b;
1364 struct bkey_i_alloc_v4 *a;
1365 struct bch_alloc_v4 old_convert, new;
1366 const struct bch_alloc_v4 *old;
1367 enum bch_data_type type;
1370 old = bch2_alloc_to_v4(k, &old_convert);
1373 percpu_down_read(&c->mark_lock);
1374 b = gc_bucket(ca, iter->pos.offset);
1377 * b->data_type doesn't yet include need_discard & need_gc_gen states -
1380 type = __alloc_data_type(b->dirty_sectors,
1385 if (b->data_type != type) {
1386 struct bch_dev_usage *u;
1389 u = this_cpu_ptr(ca->usage_gc);
1390 u->d[b->data_type].buckets--;
1391 b->data_type = type;
1392 u->d[b->data_type].buckets++;
1397 percpu_up_read(&c->mark_lock);
1399 if (metadata_only &&
1400 gc.data_type != BCH_DATA_sb &&
1401 gc.data_type != BCH_DATA_journal &&
1402 gc.data_type != BCH_DATA_btree)
1405 if (gen_after(old->gen, gc.gen))
1408 if (c->opts.reconstruct_alloc ||
1409 fsck_err_on(new.data_type != gc.data_type, c,
1410 alloc_key_data_type_wrong,
1411 "bucket %llu:%llu gen %u has wrong data_type"
1412 ": got %s, should be %s",
1413 iter->pos.inode, iter->pos.offset,
1415 bch2_data_types[new.data_type],
1416 bch2_data_types[gc.data_type]))
1417 new.data_type = gc.data_type;
1419 #define copy_bucket_field(_errtype, _f) \
1420 if (c->opts.reconstruct_alloc || \
1421 fsck_err_on(new._f != gc._f, c, _errtype, \
1422 "bucket %llu:%llu gen %u data type %s has wrong " #_f \
1423 ": got %u, should be %u", \
1424 iter->pos.inode, iter->pos.offset, \
1426 bch2_data_types[gc.data_type], \
1430 copy_bucket_field(alloc_key_gen_wrong,
1432 copy_bucket_field(alloc_key_dirty_sectors_wrong,
1434 copy_bucket_field(alloc_key_cached_sectors_wrong,
1436 copy_bucket_field(alloc_key_stripe_wrong,
1438 copy_bucket_field(alloc_key_stripe_redundancy_wrong,
1440 #undef copy_bucket_field
1442 if (!bch2_alloc_v4_cmp(*old, new))
1445 a = bch2_alloc_to_v4_mut(trans, k);
1446 ret = PTR_ERR_OR_ZERO(a);
1453 * The trigger normally makes sure this is set, but we're not running
1456 if (a->v.data_type == BCH_DATA_cached && !a->v.io_time[READ])
1457 a->v.io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
1459 ret = bch2_trans_update(trans, iter, &a->k_i, BTREE_TRIGGER_NORUN);
1464 static int bch2_gc_alloc_done(struct bch_fs *c, bool metadata_only)
1468 for_each_member_device(c, ca) {
1469 ret = bch2_trans_run(c,
1470 for_each_btree_key_upto_commit(trans, iter, BTREE_ID_alloc,
1471 POS(ca->dev_idx, ca->mi.first_bucket),
1472 POS(ca->dev_idx, ca->mi.nbuckets - 1),
1473 BTREE_ITER_SLOTS|BTREE_ITER_PREFETCH, k,
1474 NULL, NULL, BCH_TRANS_COMMIT_lazy_rw,
1475 bch2_alloc_write_key(trans, &iter, k, metadata_only)));
1477 percpu_ref_put(&ca->ref);
1486 static int bch2_gc_alloc_start(struct bch_fs *c, bool metadata_only)
1488 for_each_member_device(c, ca) {
1489 struct bucket_array *buckets = kvpmalloc(sizeof(struct bucket_array) +
1490 ca->mi.nbuckets * sizeof(struct bucket),
1491 GFP_KERNEL|__GFP_ZERO);
1493 percpu_ref_put(&ca->ref);
1494 bch_err(c, "error allocating ca->buckets[gc]");
1495 return -BCH_ERR_ENOMEM_gc_alloc_start;
1498 buckets->first_bucket = ca->mi.first_bucket;
1499 buckets->nbuckets = ca->mi.nbuckets;
1500 rcu_assign_pointer(ca->buckets_gc, buckets);
1503 int ret = bch2_trans_run(c,
1504 for_each_btree_key(trans, iter, BTREE_ID_alloc, POS_MIN,
1505 BTREE_ITER_PREFETCH, k, ({
1506 struct bch_dev *ca = bch_dev_bkey_exists(c, k.k->p.inode);
1507 struct bucket *g = gc_bucket(ca, k.k->p.offset);
1509 struct bch_alloc_v4 a_convert;
1510 const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &a_convert);
1515 if (metadata_only &&
1516 (a->data_type == BCH_DATA_user ||
1517 a->data_type == BCH_DATA_cached ||
1518 a->data_type == BCH_DATA_parity)) {
1519 g->data_type = a->data_type;
1520 g->dirty_sectors = a->dirty_sectors;
1521 g->cached_sectors = a->cached_sectors;
1522 g->stripe = a->stripe;
1523 g->stripe_redundancy = a->stripe_redundancy;
1532 static void bch2_gc_alloc_reset(struct bch_fs *c, bool metadata_only)
1534 for_each_member_device(c, ca) {
1535 struct bucket_array *buckets = gc_bucket_array(ca);
1538 for_each_bucket(g, buckets) {
1539 if (metadata_only &&
1540 (g->data_type == BCH_DATA_user ||
1541 g->data_type == BCH_DATA_cached ||
1542 g->data_type == BCH_DATA_parity))
1545 g->dirty_sectors = 0;
1546 g->cached_sectors = 0;
1551 static int bch2_gc_write_reflink_key(struct btree_trans *trans,
1552 struct btree_iter *iter,
1556 struct bch_fs *c = trans->c;
1557 const __le64 *refcount = bkey_refcount_c(k);
1558 struct printbuf buf = PRINTBUF;
1559 struct reflink_gc *r;
1565 while ((r = genradix_ptr(&c->reflink_gc_table, *idx)) &&
1566 r->offset < k.k->p.offset)
1570 r->offset != k.k->p.offset ||
1571 r->size != k.k->size) {
1572 bch_err(c, "unexpected inconsistency walking reflink table at gc finish");
1576 if (fsck_err_on(r->refcount != le64_to_cpu(*refcount), c,
1577 reflink_v_refcount_wrong,
1578 "reflink key has wrong refcount:\n"
1581 (bch2_bkey_val_to_text(&buf, c, k), buf.buf),
1583 struct bkey_i *new = bch2_bkey_make_mut(trans, iter, &k, 0);
1585 ret = PTR_ERR_OR_ZERO(new);
1590 new->k.type = KEY_TYPE_deleted;
1592 *bkey_refcount(new) = cpu_to_le64(r->refcount);
1595 printbuf_exit(&buf);
1599 static int bch2_gc_reflink_done(struct bch_fs *c, bool metadata_only)
1606 int ret = bch2_trans_run(c,
1607 for_each_btree_key_commit(trans, iter,
1608 BTREE_ID_reflink, POS_MIN,
1609 BTREE_ITER_PREFETCH, k,
1610 NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
1611 bch2_gc_write_reflink_key(trans, &iter, k, &idx)));
1612 c->reflink_gc_nr = 0;
1616 static int bch2_gc_reflink_start(struct bch_fs *c,
1623 c->reflink_gc_nr = 0;
1625 int ret = bch2_trans_run(c,
1626 for_each_btree_key(trans, iter, BTREE_ID_reflink, POS_MIN,
1627 BTREE_ITER_PREFETCH, k, ({
1628 const __le64 *refcount = bkey_refcount_c(k);
1633 struct reflink_gc *r = genradix_ptr_alloc(&c->reflink_gc_table,
1634 c->reflink_gc_nr++, GFP_KERNEL);
1636 ret = -BCH_ERR_ENOMEM_gc_reflink_start;
1640 r->offset = k.k->p.offset;
1641 r->size = k.k->size;
1650 static void bch2_gc_reflink_reset(struct bch_fs *c, bool metadata_only)
1652 struct genradix_iter iter;
1653 struct reflink_gc *r;
1655 genradix_for_each(&c->reflink_gc_table, iter, r)
1659 static int bch2_gc_write_stripes_key(struct btree_trans *trans,
1660 struct btree_iter *iter,
1663 struct bch_fs *c = trans->c;
1664 struct printbuf buf = PRINTBUF;
1665 const struct bch_stripe *s;
1666 struct gc_stripe *m;
1671 if (k.k->type != KEY_TYPE_stripe)
1674 s = bkey_s_c_to_stripe(k).v;
1675 m = genradix_ptr(&c->gc_stripes, k.k->p.offset);
1677 for (i = 0; i < s->nr_blocks; i++) {
1678 u32 old = stripe_blockcount_get(s, i);
1679 u32 new = (m ? m->block_sectors[i] : 0);
1682 prt_printf(&buf, "stripe block %u has wrong sector count: got %u, should be %u\n",
1689 bch2_bkey_val_to_text(&buf, c, k);
1691 if (fsck_err_on(bad, c, stripe_sector_count_wrong,
1693 struct bkey_i_stripe *new;
1695 new = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1696 ret = PTR_ERR_OR_ZERO(new);
1700 bkey_reassemble(&new->k_i, k);
1702 for (i = 0; i < new->v.nr_blocks; i++)
1703 stripe_blockcount_set(&new->v, i, m ? m->block_sectors[i] : 0);
1705 ret = bch2_trans_update(trans, iter, &new->k_i, 0);
1708 printbuf_exit(&buf);
1712 static int bch2_gc_stripes_done(struct bch_fs *c, bool metadata_only)
1717 return bch2_trans_run(c,
1718 for_each_btree_key_commit(trans, iter,
1719 BTREE_ID_stripes, POS_MIN,
1720 BTREE_ITER_PREFETCH, k,
1721 NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
1722 bch2_gc_write_stripes_key(trans, &iter, k)));
1725 static void bch2_gc_stripes_reset(struct bch_fs *c, bool metadata_only)
1727 genradix_free(&c->gc_stripes);
1731 * bch2_gc - walk _all_ references to buckets, and recompute them:
1733 * @c: filesystem object
1734 * @initial: are we in recovery?
1735 * @metadata_only: are we just checking metadata references, or everything?
1737 * Returns: 0 on success, or standard errcode on failure
1739 * Order matters here:
1740 * - Concurrent GC relies on the fact that we have a total ordering for
1741 * everything that GC walks - see gc_will_visit_node(),
1742 * gc_will_visit_root()
1744 * - also, references move around in the course of index updates and
1745 * various other crap: everything needs to agree on the ordering
1746 * references are allowed to move around in - e.g., we're allowed to
1747 * start with a reference owned by an open_bucket (the allocator) and
1748 * move it to the btree, but not the reverse.
1750 * This is necessary to ensure that gc doesn't miss references that
1751 * move around - if references move backwards in the ordering GC
1752 * uses, GC could skip past them
1754 int bch2_gc(struct bch_fs *c, bool initial, bool metadata_only)
1759 lockdep_assert_held(&c->state_lock);
1761 down_write(&c->gc_lock);
1763 bch2_btree_interior_updates_flush(c);
1765 ret = bch2_gc_start(c) ?:
1766 bch2_gc_alloc_start(c, metadata_only) ?:
1767 bch2_gc_reflink_start(c, metadata_only);
1771 gc_pos_set(c, gc_phase(GC_PHASE_START));
1773 bch2_mark_superblocks(c);
1775 ret = bch2_gc_btrees(c, initial, metadata_only);
1781 bch2_mark_pending_btree_node_frees(c);
1785 if (test_bit(BCH_FS_need_another_gc, &c->flags) ||
1786 (!iter && bch2_test_restart_gc)) {
1788 bch_info(c, "Unable to fix bucket gens, looping");
1794 * XXX: make sure gens we fixed got saved
1796 bch_info(c, "Second GC pass needed, restarting:");
1797 clear_bit(BCH_FS_need_another_gc, &c->flags);
1798 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1800 bch2_gc_stripes_reset(c, metadata_only);
1801 bch2_gc_alloc_reset(c, metadata_only);
1802 bch2_gc_reflink_reset(c, metadata_only);
1803 ret = bch2_gc_reset(c);
1807 /* flush fsck errors, reset counters */
1808 bch2_flush_fsck_errs(c);
1813 bch2_journal_block(&c->journal);
1815 ret = bch2_gc_stripes_done(c, metadata_only) ?:
1816 bch2_gc_reflink_done(c, metadata_only) ?:
1817 bch2_gc_alloc_done(c, metadata_only) ?:
1818 bch2_gc_done(c, initial, metadata_only);
1820 bch2_journal_unblock(&c->journal);
1823 percpu_down_write(&c->mark_lock);
1824 /* Indicates that gc is no longer in progress: */
1825 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1828 percpu_up_write(&c->mark_lock);
1830 up_write(&c->gc_lock);
1833 * At startup, allocations can happen directly instead of via the
1834 * allocator thread - issue wakeup in case they blocked on gc_lock:
1836 closure_wake_up(&c->freelist_wait);
1841 static int gc_btree_gens_key(struct btree_trans *trans,
1842 struct btree_iter *iter,
1845 struct bch_fs *c = trans->c;
1846 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1850 percpu_down_read(&c->mark_lock);
1851 bkey_for_each_ptr(ptrs, ptr) {
1852 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1854 if (ptr_stale(ca, ptr) > 16) {
1855 percpu_up_read(&c->mark_lock);
1860 bkey_for_each_ptr(ptrs, ptr) {
1861 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1862 u8 *gen = &ca->oldest_gen[PTR_BUCKET_NR(ca, ptr)];
1864 if (gen_after(*gen, ptr->gen))
1867 percpu_up_read(&c->mark_lock);
1870 u = bch2_bkey_make_mut(trans, iter, &k, 0);
1871 ret = PTR_ERR_OR_ZERO(u);
1875 bch2_extent_normalize(c, bkey_i_to_s(u));
1879 static int bch2_alloc_write_oldest_gen(struct btree_trans *trans, struct btree_iter *iter,
1882 struct bch_dev *ca = bch_dev_bkey_exists(trans->c, iter->pos.inode);
1883 struct bch_alloc_v4 a_convert;
1884 const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &a_convert);
1885 struct bkey_i_alloc_v4 *a_mut;
1888 if (a->oldest_gen == ca->oldest_gen[iter->pos.offset])
1891 a_mut = bch2_alloc_to_v4_mut(trans, k);
1892 ret = PTR_ERR_OR_ZERO(a_mut);
1896 a_mut->v.oldest_gen = ca->oldest_gen[iter->pos.offset];
1897 a_mut->v.data_type = alloc_data_type(a_mut->v, a_mut->v.data_type);
1899 return bch2_trans_update(trans, iter, &a_mut->k_i, 0);
1902 int bch2_gc_gens(struct bch_fs *c)
1904 u64 b, start_time = local_clock();
1908 * Ideally we would be using state_lock and not gc_lock here, but that
1909 * introduces a deadlock in the RO path - we currently take the state
1910 * lock at the start of going RO, thus the gc thread may get stuck:
1912 if (!mutex_trylock(&c->gc_gens_lock))
1915 trace_and_count(c, gc_gens_start, c);
1916 down_read(&c->gc_lock);
1918 for_each_member_device(c, ca) {
1919 struct bucket_gens *gens = bucket_gens(ca);
1921 BUG_ON(ca->oldest_gen);
1923 ca->oldest_gen = kvmalloc(gens->nbuckets, GFP_KERNEL);
1924 if (!ca->oldest_gen) {
1925 percpu_ref_put(&ca->ref);
1926 ret = -BCH_ERR_ENOMEM_gc_gens;
1930 for (b = gens->first_bucket;
1931 b < gens->nbuckets; b++)
1932 ca->oldest_gen[b] = gens->b[b];
1935 for (unsigned i = 0; i < BTREE_ID_NR; i++)
1936 if (btree_type_has_ptrs(i)) {
1937 c->gc_gens_btree = i;
1938 c->gc_gens_pos = POS_MIN;
1940 ret = bch2_trans_run(c,
1941 for_each_btree_key_commit(trans, iter, i,
1943 BTREE_ITER_PREFETCH|BTREE_ITER_ALL_SNAPSHOTS,
1946 BCH_TRANS_COMMIT_no_enospc,
1947 gc_btree_gens_key(trans, &iter, k)));
1952 ret = bch2_trans_run(c,
1953 for_each_btree_key_commit(trans, iter, BTREE_ID_alloc,
1955 BTREE_ITER_PREFETCH,
1958 BCH_TRANS_COMMIT_no_enospc,
1959 bch2_alloc_write_oldest_gen(trans, &iter, k)));
1963 c->gc_gens_btree = 0;
1964 c->gc_gens_pos = POS_MIN;
1968 bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
1969 trace_and_count(c, gc_gens_end, c);
1971 for_each_member_device(c, ca) {
1972 kvfree(ca->oldest_gen);
1973 ca->oldest_gen = NULL;
1976 up_read(&c->gc_lock);
1977 mutex_unlock(&c->gc_gens_lock);
1978 if (!bch2_err_matches(ret, EROFS))
1983 static int bch2_gc_thread(void *arg)
1985 struct bch_fs *c = arg;
1986 struct io_clock *clock = &c->io_clock[WRITE];
1987 unsigned long last = atomic64_read(&clock->now);
1988 unsigned last_kick = atomic_read(&c->kick_gc);
1994 set_current_state(TASK_INTERRUPTIBLE);
1996 if (kthread_should_stop()) {
1997 __set_current_state(TASK_RUNNING);
2001 if (atomic_read(&c->kick_gc) != last_kick)
2004 if (c->btree_gc_periodic) {
2005 unsigned long next = last + c->capacity / 16;
2007 if (atomic64_read(&clock->now) >= next)
2010 bch2_io_clock_schedule_timeout(clock, next);
2017 __set_current_state(TASK_RUNNING);
2019 last = atomic64_read(&clock->now);
2020 last_kick = atomic_read(&c->kick_gc);
2023 * Full gc is currently incompatible with btree key cache:
2026 ret = bch2_gc(c, false, false);
2030 debug_check_no_locks_held();
2036 void bch2_gc_thread_stop(struct bch_fs *c)
2038 struct task_struct *p;
2041 c->gc_thread = NULL;
2049 int bch2_gc_thread_start(struct bch_fs *c)
2051 struct task_struct *p;
2056 p = kthread_create(bch2_gc_thread, c, "bch-gc/%s", c->name);
2058 bch_err_fn(c, PTR_ERR(p));