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 with incorrect min_key at btree %s level %u:\n"
104 bch2_btree_ids[b->c.btree_id], b->c.level,
105 buf1.buf, buf2.buf) &&
106 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);
129 "btree node with incorrect max_key at btree %s level %u:\n"
132 bch2_btree_ids[b->c.btree_id], b->c.level,
133 buf1.buf, buf2.buf) &&
134 should_restart_for_topology_repair(c)) {
135 bch_info(c, "Halting mark and sweep to start topology repair pass");
136 ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_topology);
139 set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
143 bch2_bkey_buf_copy(prev, c, cur.k);
146 printbuf_exit(&buf2);
147 printbuf_exit(&buf1);
151 static void btree_ptr_to_v2(struct btree *b, struct bkey_i_btree_ptr_v2 *dst)
153 switch (b->key.k.type) {
154 case KEY_TYPE_btree_ptr: {
155 struct bkey_i_btree_ptr *src = bkey_i_to_btree_ptr(&b->key);
159 dst->v.seq = b->data->keys.seq;
160 dst->v.sectors_written = 0;
162 dst->v.min_key = b->data->min_key;
163 set_bkey_val_bytes(&dst->k, sizeof(dst->v) + bkey_val_bytes(&src->k));
164 memcpy(dst->v.start, src->v.start, bkey_val_bytes(&src->k));
167 case KEY_TYPE_btree_ptr_v2:
168 bkey_copy(&dst->k_i, &b->key);
175 static void bch2_btree_node_update_key_early(struct btree_trans *trans,
176 enum btree_id btree, unsigned level,
177 struct bkey_s_c old, struct bkey_i *new)
179 struct bch_fs *c = trans->c;
184 bch2_bkey_buf_init(&tmp);
185 bch2_bkey_buf_reassemble(&tmp, c, old);
187 b = bch2_btree_node_get_noiter(trans, tmp.k, btree, level, true);
188 if (!IS_ERR_OR_NULL(b)) {
189 mutex_lock(&c->btree_cache.lock);
191 bch2_btree_node_hash_remove(&c->btree_cache, b);
193 bkey_copy(&b->key, new);
194 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
197 mutex_unlock(&c->btree_cache.lock);
198 six_unlock_read(&b->c.lock);
201 bch2_bkey_buf_exit(&tmp, c);
204 static int set_node_min(struct bch_fs *c, struct btree *b, struct bpos new_min)
206 struct bkey_i_btree_ptr_v2 *new;
209 new = kmalloc_array(BKEY_BTREE_PTR_U64s_MAX, sizeof(u64), GFP_KERNEL);
211 return -BCH_ERR_ENOMEM_gc_repair_key;
213 btree_ptr_to_v2(b, new);
214 b->data->min_key = new_min;
215 new->v.min_key = new_min;
216 SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
218 ret = bch2_journal_key_insert_take(c, b->c.btree_id, b->c.level + 1, &new->k_i);
224 bch2_btree_node_drop_keys_outside_node(b);
225 bkey_copy(&b->key, &new->k_i);
229 static int set_node_max(struct bch_fs *c, struct btree *b, struct bpos new_max)
231 struct bkey_i_btree_ptr_v2 *new;
234 ret = bch2_journal_key_delete(c, b->c.btree_id, b->c.level + 1, b->key.k.p);
238 new = kmalloc_array(BKEY_BTREE_PTR_U64s_MAX, sizeof(u64), GFP_KERNEL);
240 return -BCH_ERR_ENOMEM_gc_repair_key;
242 btree_ptr_to_v2(b, new);
243 b->data->max_key = new_max;
245 SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
247 ret = bch2_journal_key_insert_take(c, b->c.btree_id, b->c.level + 1, &new->k_i);
253 bch2_btree_node_drop_keys_outside_node(b);
255 mutex_lock(&c->btree_cache.lock);
256 bch2_btree_node_hash_remove(&c->btree_cache, b);
258 bkey_copy(&b->key, &new->k_i);
259 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
261 mutex_unlock(&c->btree_cache.lock);
265 static int btree_repair_node_boundaries(struct bch_fs *c, struct btree *b,
266 struct btree *prev, struct btree *cur)
268 struct bpos expected_start = !prev
270 : bpos_successor(prev->key.k.p);
271 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
275 prt_printf(&buf1, "start of node: ");
276 bch2_bpos_to_text(&buf1, b->data->min_key);
278 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(&prev->key));
281 bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(&cur->key));
284 bpos_gt(expected_start, cur->data->min_key) &&
285 BTREE_NODE_SEQ(cur->data) > BTREE_NODE_SEQ(prev->data)) {
286 /* cur overwrites prev: */
288 if (mustfix_fsck_err_on(bpos_ge(prev->data->min_key,
289 cur->data->min_key), c,
290 "btree node overwritten by next node at btree %s level %u:\n"
293 bch2_btree_ids[b->c.btree_id], b->c.level,
294 buf1.buf, buf2.buf)) {
295 ret = DROP_PREV_NODE;
299 if (mustfix_fsck_err_on(!bpos_eq(prev->key.k.p,
300 bpos_predecessor(cur->data->min_key)), c,
301 "btree node with incorrect max_key at btree %s level %u:\n"
304 bch2_btree_ids[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 overwritten by prev node at btree %s level %u:\n"
316 bch2_btree_ids[b->c.btree_id], b->c.level,
317 buf1.buf, buf2.buf)) {
318 ret = DROP_THIS_NODE;
322 if (mustfix_fsck_err_on(!bpos_eq(expected_start, cur->data->min_key), c,
323 "btree node with incorrect min_key at btree %s level %u:\n"
326 bch2_btree_ids[b->c.btree_id], b->c.level,
328 ret = set_node_min(c, cur, expected_start);
332 printbuf_exit(&buf2);
333 printbuf_exit(&buf1);
337 static int btree_repair_node_end(struct bch_fs *c, struct btree *b,
340 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
343 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(&child->key));
344 bch2_bpos_to_text(&buf2, b->key.k.p);
346 if (mustfix_fsck_err_on(!bpos_eq(child->key.k.p, b->key.k.p), c,
347 "btree node with incorrect max_key at btree %s level %u:\n"
350 bch2_btree_ids[b->c.btree_id], b->c.level,
351 buf1.buf, buf2.buf)) {
352 ret = set_node_max(c, child, b->key.k.p);
358 printbuf_exit(&buf2);
359 printbuf_exit(&buf1);
363 static int bch2_btree_repair_topology_recurse(struct btree_trans *trans, struct btree *b)
365 struct bch_fs *c = trans->c;
366 struct btree_and_journal_iter iter;
368 struct bkey_buf prev_k, cur_k;
369 struct btree *prev = NULL, *cur = NULL;
370 bool have_child, dropped_children = false;
371 struct printbuf buf = PRINTBUF;
378 have_child = dropped_children = false;
379 bch2_bkey_buf_init(&prev_k);
380 bch2_bkey_buf_init(&cur_k);
381 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
383 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
384 BUG_ON(bpos_lt(k.k->p, b->data->min_key));
385 BUG_ON(bpos_gt(k.k->p, b->data->max_key));
387 bch2_btree_and_journal_iter_advance(&iter);
388 bch2_bkey_buf_reassemble(&cur_k, c, k);
390 cur = bch2_btree_node_get_noiter(trans, cur_k.k,
391 b->c.btree_id, b->c.level - 1,
393 ret = PTR_ERR_OR_ZERO(cur);
395 printbuf_reset(&buf);
396 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(cur_k.k));
398 if (mustfix_fsck_err_on(ret == -EIO, c,
399 "Topology repair: unreadable btree node at btree %s level %u:\n"
401 bch2_btree_ids[b->c.btree_id],
404 bch2_btree_node_evict(trans, cur_k.k);
405 ret = bch2_journal_key_delete(c, b->c.btree_id,
406 b->c.level, cur_k.k->k.p);
414 bch_err_msg(c, ret, "getting btree node");
418 ret = btree_repair_node_boundaries(c, b, prev, cur);
420 if (ret == DROP_THIS_NODE) {
421 six_unlock_read(&cur->c.lock);
422 bch2_btree_node_evict(trans, cur_k.k);
423 ret = bch2_journal_key_delete(c, b->c.btree_id,
424 b->c.level, cur_k.k->k.p);
432 six_unlock_read(&prev->c.lock);
435 if (ret == DROP_PREV_NODE) {
436 bch2_btree_node_evict(trans, prev_k.k);
437 ret = bch2_journal_key_delete(c, b->c.btree_id,
438 b->c.level, prev_k.k->k.p);
442 bch2_btree_and_journal_iter_exit(&iter);
443 bch2_bkey_buf_exit(&prev_k, c);
444 bch2_bkey_buf_exit(&cur_k, c);
451 bch2_bkey_buf_copy(&prev_k, c, cur_k.k);
454 if (!ret && !IS_ERR_OR_NULL(prev)) {
456 ret = btree_repair_node_end(c, b, prev);
459 if (!IS_ERR_OR_NULL(prev))
460 six_unlock_read(&prev->c.lock);
462 if (!IS_ERR_OR_NULL(cur))
463 six_unlock_read(&cur->c.lock);
469 bch2_btree_and_journal_iter_exit(&iter);
470 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
472 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
473 bch2_bkey_buf_reassemble(&cur_k, c, k);
474 bch2_btree_and_journal_iter_advance(&iter);
476 cur = bch2_btree_node_get_noiter(trans, cur_k.k,
477 b->c.btree_id, b->c.level - 1,
479 ret = PTR_ERR_OR_ZERO(cur);
482 bch_err_msg(c, ret, "getting btree node");
486 ret = bch2_btree_repair_topology_recurse(trans, cur);
487 six_unlock_read(&cur->c.lock);
490 if (ret == DROP_THIS_NODE) {
491 bch2_btree_node_evict(trans, cur_k.k);
492 ret = bch2_journal_key_delete(c, b->c.btree_id,
493 b->c.level, cur_k.k->k.p);
494 dropped_children = true;
503 printbuf_reset(&buf);
504 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
506 if (mustfix_fsck_err_on(!have_child, c,
507 "empty interior btree node at btree %s level %u\n"
509 bch2_btree_ids[b->c.btree_id],
510 b->c.level, buf.buf))
511 ret = DROP_THIS_NODE;
514 if (!IS_ERR_OR_NULL(prev))
515 six_unlock_read(&prev->c.lock);
516 if (!IS_ERR_OR_NULL(cur))
517 six_unlock_read(&cur->c.lock);
519 bch2_btree_and_journal_iter_exit(&iter);
520 bch2_bkey_buf_exit(&prev_k, c);
521 bch2_bkey_buf_exit(&cur_k, c);
523 if (!ret && dropped_children)
530 int bch2_check_topology(struct bch_fs *c)
532 struct btree_trans *trans = bch2_trans_get(c);
537 for (i = 0; i < btree_id_nr_alive(c) && !ret; i++) {
538 struct btree_root *r = bch2_btree_id_root(c, i);
544 if (btree_node_fake(b))
547 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
548 ret = bch2_btree_repair_topology_recurse(trans, b);
549 six_unlock_read(&b->c.lock);
551 if (ret == DROP_THIS_NODE) {
552 bch_err(c, "empty btree root - repair unimplemented");
553 ret = -BCH_ERR_fsck_repair_unimplemented;
557 bch2_trans_put(trans);
562 static int bch2_check_fix_ptrs(struct btree_trans *trans, enum btree_id btree_id,
563 unsigned level, bool is_root,
566 struct bch_fs *c = trans->c;
567 struct bkey_ptrs_c ptrs_c = bch2_bkey_ptrs_c(*k);
568 const union bch_extent_entry *entry_c;
569 struct extent_ptr_decoded p = { 0 };
570 bool do_update = false;
571 struct printbuf buf = PRINTBUF;
576 * use check_bucket_ref here
578 bkey_for_each_ptr_decode(k->k, ptrs_c, p, entry_c) {
579 struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
580 struct bucket *g = PTR_GC_BUCKET(ca, &p.ptr);
581 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, &entry_c->ptr);
584 (c->opts.reconstruct_alloc ||
585 fsck_err(c, "bucket %u:%zu data type %s ptr gen %u missing in alloc btree\n"
587 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
588 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
590 (printbuf_reset(&buf),
591 bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))) {
600 if (gen_cmp(p.ptr.gen, g->gen) > 0 &&
601 (c->opts.reconstruct_alloc ||
602 fsck_err(c, "bucket %u:%zu data type %s ptr gen in the future: %u > %u\n"
604 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
605 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
607 (printbuf_reset(&buf),
608 bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))) {
613 g->dirty_sectors = 0;
614 g->cached_sectors = 0;
615 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
621 if (gen_cmp(g->gen, p.ptr.gen) > BUCKET_GC_GEN_MAX &&
622 (c->opts.reconstruct_alloc ||
623 fsck_err(c, "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
625 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr), g->gen,
626 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
628 (printbuf_reset(&buf),
629 bch2_bkey_val_to_text(&buf, c, *k), buf.buf))))
632 if (!p.ptr.cached && gen_cmp(p.ptr.gen, g->gen) < 0 &&
633 (c->opts.reconstruct_alloc ||
634 fsck_err(c, "bucket %u:%zu data type %s stale dirty ptr: %u < %u\n"
636 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
637 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
639 (printbuf_reset(&buf),
640 bch2_bkey_val_to_text(&buf, c, *k), buf.buf))))
643 if (data_type != BCH_DATA_btree && p.ptr.gen != g->gen)
646 if (fsck_err_on(bucket_data_type(g->data_type) &&
647 bucket_data_type(g->data_type) != data_type, c,
648 "bucket %u:%zu different types of data in same bucket: %s, %s\n"
650 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
651 bch2_data_types[g->data_type],
652 bch2_data_types[data_type],
653 (printbuf_reset(&buf),
654 bch2_bkey_val_to_text(&buf, c, *k), buf.buf))) {
655 if (data_type == BCH_DATA_btree) {
656 g->data_type = data_type;
657 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
664 struct gc_stripe *m = genradix_ptr(&c->gc_stripes, p.ec.idx);
666 if (fsck_err_on(!m || !m->alive, c,
667 "pointer to nonexistent stripe %llu\n"
670 (printbuf_reset(&buf),
671 bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
674 if (fsck_err_on(m && m->alive && !bch2_ptr_matches_stripe_m(m, p), c,
675 "pointer does not match stripe %llu\n"
678 (printbuf_reset(&buf),
679 bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
685 struct bkey_ptrs ptrs;
686 union bch_extent_entry *entry;
687 struct bch_extent_ptr *ptr;
691 bch_err(c, "cannot update btree roots yet");
696 new = kmalloc(bkey_bytes(k->k), GFP_KERNEL);
698 bch_err_msg(c, ret, "allocating new key");
699 ret = -BCH_ERR_ENOMEM_gc_repair_key;
703 bkey_reassemble(new, *k);
707 * We don't want to drop btree node pointers - if the
708 * btree node isn't there anymore, the read path will
711 ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
712 bkey_for_each_ptr(ptrs, ptr) {
713 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
714 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
719 bch2_bkey_drop_ptrs(bkey_i_to_s(new), ptr, ({
720 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
721 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
722 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, ptr);
725 (!g->gen_valid || gen_cmp(ptr->gen, g->gen) > 0)) ||
727 gen_cmp(ptr->gen, g->gen) < 0) ||
728 gen_cmp(g->gen, ptr->gen) > BUCKET_GC_GEN_MAX ||
730 g->data_type != data_type);
733 ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
734 bkey_extent_entry_for_each(ptrs, entry) {
735 if (extent_entry_type(entry) == BCH_EXTENT_ENTRY_stripe_ptr) {
736 struct gc_stripe *m = genradix_ptr(&c->gc_stripes,
737 entry->stripe_ptr.idx);
738 union bch_extent_entry *next_ptr;
740 bkey_extent_entry_for_each_from(ptrs, next_ptr, entry)
741 if (extent_entry_type(next_ptr) == BCH_EXTENT_ENTRY_ptr)
746 bch_err(c, "aieee, found stripe ptr with no data ptr");
750 if (!m || !m->alive ||
751 !__bch2_ptr_matches_stripe(&m->ptrs[entry->stripe_ptr.block],
754 bch2_bkey_extent_entry_drop(new, entry);
761 ret = bch2_journal_key_insert_take(c, btree_id, level, new);
768 bch2_btree_node_update_key_early(trans, btree_id, level - 1, *k, new);
771 printbuf_reset(&buf);
772 bch2_bkey_val_to_text(&buf, c, *k);
773 bch_info(c, "updated %s", buf.buf);
775 printbuf_reset(&buf);
776 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(new));
777 bch_info(c, "new key %s", buf.buf);
780 *k = bkey_i_to_s_c(new);
788 /* marking of btree keys/nodes: */
790 static int bch2_gc_mark_key(struct btree_trans *trans, enum btree_id btree_id,
791 unsigned level, bool is_root,
795 struct bch_fs *c = trans->c;
796 struct bkey deleted = KEY(0, 0, 0);
797 struct bkey_s_c old = (struct bkey_s_c) { &deleted, NULL };
800 (initial ? BTREE_TRIGGER_NOATOMIC : 0);
806 BUG_ON(bch2_journal_seq_verify &&
807 k->k->version.lo > atomic64_read(&c->journal.seq));
809 ret = bch2_check_fix_ptrs(trans, btree_id, level, is_root, k);
813 if (fsck_err_on(k->k->version.lo > atomic64_read(&c->key_version), c,
814 "key version number higher than recorded: %llu > %llu",
816 atomic64_read(&c->key_version)))
817 atomic64_set(&c->key_version, k->k->version.lo);
820 ret = commit_do(trans, NULL, NULL, 0,
821 bch2_mark_key(trans, btree_id, level, old, *k, flags));
829 static int btree_gc_mark_node(struct btree_trans *trans, struct btree *b, bool initial)
831 struct bch_fs *c = trans->c;
832 struct btree_node_iter iter;
833 struct bkey unpacked;
835 struct bkey_buf prev, cur;
838 if (!btree_node_type_needs_gc(btree_node_type(b)))
841 bch2_btree_node_iter_init_from_start(&iter, b);
842 bch2_bkey_buf_init(&prev);
843 bch2_bkey_buf_init(&cur);
844 bkey_init(&prev.k->k);
846 while ((k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked)).k) {
847 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, false,
852 bch2_btree_node_iter_advance(&iter, b);
855 bch2_bkey_buf_reassemble(&cur, c, k);
857 ret = bch2_gc_check_topology(c, b, &prev, cur,
858 bch2_btree_node_iter_end(&iter));
864 bch2_bkey_buf_exit(&cur, c);
865 bch2_bkey_buf_exit(&prev, c);
869 static int bch2_gc_btree(struct btree_trans *trans, enum btree_id btree_id,
870 bool initial, bool metadata_only)
872 struct bch_fs *c = trans->c;
873 struct btree_iter iter;
875 unsigned depth = metadata_only ? 1 : 0;
878 gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));
880 __for_each_btree_node(trans, iter, btree_id, POS_MIN,
881 0, depth, BTREE_ITER_PREFETCH, b, ret) {
882 bch2_verify_btree_nr_keys(b);
884 gc_pos_set(c, gc_pos_btree_node(b));
886 ret = btree_gc_mark_node(trans, b, initial);
890 bch2_trans_iter_exit(trans, &iter);
895 mutex_lock(&c->btree_root_lock);
896 b = bch2_btree_id_root(c, btree_id)->b;
897 if (!btree_node_fake(b)) {
898 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
900 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level + 1,
903 gc_pos_set(c, gc_pos_btree_root(b->c.btree_id));
904 mutex_unlock(&c->btree_root_lock);
909 static int bch2_gc_btree_init_recurse(struct btree_trans *trans, struct btree *b,
910 unsigned target_depth)
912 struct bch_fs *c = trans->c;
913 struct btree_and_journal_iter iter;
915 struct bkey_buf cur, prev;
916 struct printbuf buf = PRINTBUF;
919 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
920 bch2_bkey_buf_init(&prev);
921 bch2_bkey_buf_init(&cur);
922 bkey_init(&prev.k->k);
924 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
925 BUG_ON(bpos_lt(k.k->p, b->data->min_key));
926 BUG_ON(bpos_gt(k.k->p, b->data->max_key));
928 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level,
934 bch2_bkey_buf_reassemble(&cur, c, k);
935 k = bkey_i_to_s_c(cur.k);
937 bch2_btree_and_journal_iter_advance(&iter);
939 ret = bch2_gc_check_topology(c, b,
941 !bch2_btree_and_journal_iter_peek(&iter).k);
945 bch2_btree_and_journal_iter_advance(&iter);
949 if (b->c.level > target_depth) {
950 bch2_btree_and_journal_iter_exit(&iter);
951 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
953 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
956 bch2_bkey_buf_reassemble(&cur, c, k);
957 bch2_btree_and_journal_iter_advance(&iter);
959 child = bch2_btree_node_get_noiter(trans, cur.k,
960 b->c.btree_id, b->c.level - 1,
962 ret = PTR_ERR_OR_ZERO(child);
965 bch2_topology_error(c);
971 "Unreadable btree node at btree %s level %u:\n"
973 bch2_btree_ids[b->c.btree_id],
975 (printbuf_reset(&buf),
976 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(cur.k)), buf.buf)) &&
977 should_restart_for_topology_repair(c)) {
978 bch_info(c, "Halting mark and sweep to start topology repair pass");
979 ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_topology);
982 /* Continue marking when opted to not
985 set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
989 bch_err_msg(c, ret, "getting btree node");
993 ret = bch2_gc_btree_init_recurse(trans, child,
995 six_unlock_read(&child->c.lock);
1002 bch2_bkey_buf_exit(&cur, c);
1003 bch2_bkey_buf_exit(&prev, c);
1004 bch2_btree_and_journal_iter_exit(&iter);
1005 printbuf_exit(&buf);
1009 static int bch2_gc_btree_init(struct btree_trans *trans,
1010 enum btree_id btree_id,
1013 struct bch_fs *c = trans->c;
1015 unsigned target_depth = metadata_only ? 1 : 0;
1016 struct printbuf buf = PRINTBUF;
1019 b = bch2_btree_id_root(c, btree_id)->b;
1021 if (btree_node_fake(b))
1024 six_lock_read(&b->c.lock, NULL, NULL);
1025 printbuf_reset(&buf);
1026 bch2_bpos_to_text(&buf, b->data->min_key);
1027 if (mustfix_fsck_err_on(!bpos_eq(b->data->min_key, POS_MIN), c,
1028 "btree root with incorrect min_key: %s", buf.buf)) {
1029 bch_err(c, "repair unimplemented");
1030 ret = -BCH_ERR_fsck_repair_unimplemented;
1034 printbuf_reset(&buf);
1035 bch2_bpos_to_text(&buf, b->data->max_key);
1036 if (mustfix_fsck_err_on(!bpos_eq(b->data->max_key, SPOS_MAX), c,
1037 "btree root with incorrect max_key: %s", buf.buf)) {
1038 bch_err(c, "repair unimplemented");
1039 ret = -BCH_ERR_fsck_repair_unimplemented;
1043 if (b->c.level >= target_depth)
1044 ret = bch2_gc_btree_init_recurse(trans, b, target_depth);
1047 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1049 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level + 1, true,
1053 six_unlock_read(&b->c.lock);
1057 printbuf_exit(&buf);
1061 static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
1063 return (int) btree_id_to_gc_phase(l) -
1064 (int) btree_id_to_gc_phase(r);
1067 static int bch2_gc_btrees(struct bch_fs *c, bool initial, bool metadata_only)
1069 struct btree_trans *trans = bch2_trans_get(c);
1070 enum btree_id ids[BTREE_ID_NR];
1075 trans->is_initial_gc = true;
1077 for (i = 0; i < BTREE_ID_NR; i++)
1079 bubble_sort(ids, BTREE_ID_NR, btree_id_gc_phase_cmp);
1081 for (i = 0; i < BTREE_ID_NR && !ret; i++)
1083 ? bch2_gc_btree_init(trans, ids[i], metadata_only)
1084 : bch2_gc_btree(trans, ids[i], initial, metadata_only);
1086 for (i = BTREE_ID_NR; i < btree_id_nr_alive(c) && !ret; i++) {
1087 if (!bch2_btree_id_root(c, i)->alive)
1091 ? bch2_gc_btree_init(trans, i, metadata_only)
1092 : bch2_gc_btree(trans, i, initial, metadata_only);
1098 bch2_trans_put(trans);
1102 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
1104 enum bch_data_type type,
1107 u64 b = sector_to_bucket(ca, start);
1111 min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1113 bch2_mark_metadata_bucket(c, ca, b, type, sectors,
1114 gc_phase(GC_PHASE_SB), flags);
1117 } while (start < end);
1120 static void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
1123 struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1127 for (i = 0; i < layout->nr_superblocks; i++) {
1128 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1130 if (offset == BCH_SB_SECTOR)
1131 mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
1132 BCH_DATA_sb, flags);
1134 mark_metadata_sectors(c, ca, offset,
1135 offset + (1 << layout->sb_max_size_bits),
1136 BCH_DATA_sb, flags);
1139 for (i = 0; i < ca->journal.nr; i++) {
1140 b = ca->journal.buckets[i];
1141 bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_journal,
1143 gc_phase(GC_PHASE_SB), flags);
1147 static void bch2_mark_superblocks(struct bch_fs *c)
1152 mutex_lock(&c->sb_lock);
1153 gc_pos_set(c, gc_phase(GC_PHASE_SB));
1155 for_each_online_member(ca, c, i)
1156 bch2_mark_dev_superblock(c, ca, BTREE_TRIGGER_GC);
1157 mutex_unlock(&c->sb_lock);
1161 /* Also see bch2_pending_btree_node_free_insert_done() */
1162 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
1164 struct btree_update *as;
1165 struct pending_btree_node_free *d;
1167 mutex_lock(&c->btree_interior_update_lock);
1168 gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
1170 for_each_pending_btree_node_free(c, as, d)
1171 if (d->index_update_done)
1172 bch2_mark_key(c, bkey_i_to_s_c(&d->key), BTREE_TRIGGER_GC);
1174 mutex_unlock(&c->btree_interior_update_lock);
1178 static void bch2_gc_free(struct bch_fs *c)
1183 genradix_free(&c->reflink_gc_table);
1184 genradix_free(&c->gc_stripes);
1186 for_each_member_device(ca, c, i) {
1187 kvpfree(rcu_dereference_protected(ca->buckets_gc, 1),
1188 sizeof(struct bucket_array) +
1189 ca->mi.nbuckets * sizeof(struct bucket));
1190 ca->buckets_gc = NULL;
1192 free_percpu(ca->usage_gc);
1193 ca->usage_gc = NULL;
1196 free_percpu(c->usage_gc);
1200 static int bch2_gc_done(struct bch_fs *c,
1201 bool initial, bool metadata_only)
1203 struct bch_dev *ca = NULL;
1204 struct printbuf buf = PRINTBUF;
1205 bool verify = !metadata_only &&
1206 !c->opts.reconstruct_alloc &&
1207 (!initial || (c->sb.compat & (1ULL << BCH_COMPAT_alloc_info)));
1211 percpu_down_write(&c->mark_lock);
1213 #define copy_field(_f, _msg, ...) \
1214 if (dst->_f != src->_f && \
1216 fsck_err(c, _msg ": got %llu, should be %llu" \
1217 , ##__VA_ARGS__, dst->_f, src->_f))) \
1219 #define copy_dev_field(_f, _msg, ...) \
1220 copy_field(_f, "dev %u has wrong " _msg, dev, ##__VA_ARGS__)
1221 #define copy_fs_field(_f, _msg, ...) \
1222 copy_field(_f, "fs has wrong " _msg, ##__VA_ARGS__)
1224 for (i = 0; i < ARRAY_SIZE(c->usage); i++)
1225 bch2_fs_usage_acc_to_base(c, i);
1227 for_each_member_device(ca, c, dev) {
1228 struct bch_dev_usage *dst = ca->usage_base;
1229 struct bch_dev_usage *src = (void *)
1230 bch2_acc_percpu_u64s((u64 __percpu *) ca->usage_gc,
1233 copy_dev_field(buckets_ec, "buckets_ec");
1235 for (i = 0; i < BCH_DATA_NR; i++) {
1236 copy_dev_field(d[i].buckets, "%s buckets", bch2_data_types[i]);
1237 copy_dev_field(d[i].sectors, "%s sectors", bch2_data_types[i]);
1238 copy_dev_field(d[i].fragmented, "%s fragmented", bch2_data_types[i]);
1243 unsigned nr = fs_usage_u64s(c);
1244 struct bch_fs_usage *dst = c->usage_base;
1245 struct bch_fs_usage *src = (void *)
1246 bch2_acc_percpu_u64s((u64 __percpu *) c->usage_gc, nr);
1248 copy_fs_field(hidden, "hidden");
1249 copy_fs_field(btree, "btree");
1251 if (!metadata_only) {
1252 copy_fs_field(data, "data");
1253 copy_fs_field(cached, "cached");
1254 copy_fs_field(reserved, "reserved");
1255 copy_fs_field(nr_inodes,"nr_inodes");
1257 for (i = 0; i < BCH_REPLICAS_MAX; i++)
1258 copy_fs_field(persistent_reserved[i],
1259 "persistent_reserved[%i]", i);
1262 for (i = 0; i < c->replicas.nr; i++) {
1263 struct bch_replicas_entry *e =
1264 cpu_replicas_entry(&c->replicas, i);
1266 if (metadata_only &&
1267 (e->data_type == BCH_DATA_user ||
1268 e->data_type == BCH_DATA_cached))
1271 printbuf_reset(&buf);
1272 bch2_replicas_entry_to_text(&buf, e);
1274 copy_fs_field(replicas[i], "%s", buf.buf);
1278 #undef copy_fs_field
1279 #undef copy_dev_field
1280 #undef copy_stripe_field
1284 percpu_ref_put(&ca->ref);
1288 percpu_up_write(&c->mark_lock);
1289 printbuf_exit(&buf);
1293 static int bch2_gc_start(struct bch_fs *c)
1295 struct bch_dev *ca = NULL;
1298 BUG_ON(c->usage_gc);
1300 c->usage_gc = __alloc_percpu_gfp(fs_usage_u64s(c) * sizeof(u64),
1301 sizeof(u64), GFP_KERNEL);
1303 bch_err(c, "error allocating c->usage_gc");
1304 return -BCH_ERR_ENOMEM_gc_start;
1307 for_each_member_device(ca, c, i) {
1308 BUG_ON(ca->usage_gc);
1310 ca->usage_gc = alloc_percpu(struct bch_dev_usage);
1311 if (!ca->usage_gc) {
1312 bch_err(c, "error allocating ca->usage_gc");
1313 percpu_ref_put(&ca->ref);
1314 return -BCH_ERR_ENOMEM_gc_start;
1317 this_cpu_write(ca->usage_gc->d[BCH_DATA_free].buckets,
1318 ca->mi.nbuckets - ca->mi.first_bucket);
1324 static int bch2_gc_reset(struct bch_fs *c)
1329 for_each_member_device(ca, c, i) {
1330 free_percpu(ca->usage_gc);
1331 ca->usage_gc = NULL;
1334 free_percpu(c->usage_gc);
1337 return bch2_gc_start(c);
1340 /* returns true if not equal */
1341 static inline bool bch2_alloc_v4_cmp(struct bch_alloc_v4 l,
1342 struct bch_alloc_v4 r)
1344 return l.gen != r.gen ||
1345 l.oldest_gen != r.oldest_gen ||
1346 l.data_type != r.data_type ||
1347 l.dirty_sectors != r.dirty_sectors ||
1348 l.cached_sectors != r.cached_sectors ||
1349 l.stripe_redundancy != r.stripe_redundancy ||
1350 l.stripe != r.stripe;
1353 static int bch2_alloc_write_key(struct btree_trans *trans,
1354 struct btree_iter *iter,
1358 struct bch_fs *c = trans->c;
1359 struct bch_dev *ca = bch_dev_bkey_exists(c, iter->pos.inode);
1360 struct bucket gc, *b;
1361 struct bkey_i_alloc_v4 *a;
1362 struct bch_alloc_v4 old_convert, new;
1363 const struct bch_alloc_v4 *old;
1364 enum bch_data_type type;
1367 if (bkey_ge(iter->pos, POS(ca->dev_idx, ca->mi.nbuckets)))
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 "bucket %llu:%llu gen %u has wrong data_type"
1411 ": got %s, should be %s",
1412 iter->pos.inode, iter->pos.offset,
1414 bch2_data_types[new.data_type],
1415 bch2_data_types[gc.data_type]))
1416 new.data_type = gc.data_type;
1418 #define copy_bucket_field(_f) \
1419 if (c->opts.reconstruct_alloc || \
1420 fsck_err_on(new._f != gc._f, c, \
1421 "bucket %llu:%llu gen %u data type %s has wrong " #_f \
1422 ": got %u, should be %u", \
1423 iter->pos.inode, iter->pos.offset, \
1425 bch2_data_types[gc.data_type], \
1429 copy_bucket_field(gen);
1430 copy_bucket_field(dirty_sectors);
1431 copy_bucket_field(cached_sectors);
1432 copy_bucket_field(stripe_redundancy);
1433 copy_bucket_field(stripe);
1434 #undef copy_bucket_field
1436 if (!bch2_alloc_v4_cmp(*old, new))
1439 a = bch2_alloc_to_v4_mut(trans, k);
1440 ret = PTR_ERR_OR_ZERO(a);
1447 * The trigger normally makes sure this is set, but we're not running
1450 if (a->v.data_type == BCH_DATA_cached && !a->v.io_time[READ])
1451 a->v.io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
1453 ret = bch2_trans_update(trans, iter, &a->k_i, BTREE_TRIGGER_NORUN);
1458 static int bch2_gc_alloc_done(struct bch_fs *c, bool metadata_only)
1460 struct btree_trans *trans = bch2_trans_get(c);
1461 struct btree_iter iter;
1467 for_each_member_device(ca, c, i) {
1468 ret = for_each_btree_key_commit(trans, iter, BTREE_ID_alloc,
1469 POS(ca->dev_idx, ca->mi.first_bucket),
1470 BTREE_ITER_SLOTS|BTREE_ITER_PREFETCH, k,
1471 NULL, NULL, BTREE_INSERT_LAZY_RW,
1472 bch2_alloc_write_key(trans, &iter, k, metadata_only));
1476 percpu_ref_put(&ca->ref);
1481 bch2_trans_put(trans);
1482 return ret < 0 ? ret : 0;
1485 static int bch2_gc_alloc_start(struct bch_fs *c, bool metadata_only)
1488 struct btree_trans *trans = bch2_trans_get(c);
1489 struct btree_iter iter;
1492 struct bch_alloc_v4 a_convert;
1493 const struct bch_alloc_v4 *a;
1497 for_each_member_device(ca, c, i) {
1498 struct bucket_array *buckets = kvpmalloc(sizeof(struct bucket_array) +
1499 ca->mi.nbuckets * sizeof(struct bucket),
1500 GFP_KERNEL|__GFP_ZERO);
1502 percpu_ref_put(&ca->ref);
1503 bch_err(c, "error allocating ca->buckets[gc]");
1504 ret = -BCH_ERR_ENOMEM_gc_alloc_start;
1508 buckets->first_bucket = ca->mi.first_bucket;
1509 buckets->nbuckets = ca->mi.nbuckets;
1510 rcu_assign_pointer(ca->buckets_gc, buckets);
1513 for_each_btree_key(trans, iter, BTREE_ID_alloc, POS_MIN,
1514 BTREE_ITER_PREFETCH, k, ret) {
1515 ca = bch_dev_bkey_exists(c, k.k->p.inode);
1516 g = gc_bucket(ca, k.k->p.offset);
1518 a = bch2_alloc_to_v4(k, &a_convert);
1523 if (metadata_only &&
1524 (a->data_type == BCH_DATA_user ||
1525 a->data_type == BCH_DATA_cached ||
1526 a->data_type == BCH_DATA_parity)) {
1527 g->data_type = a->data_type;
1528 g->dirty_sectors = a->dirty_sectors;
1529 g->cached_sectors = a->cached_sectors;
1530 g->stripe = a->stripe;
1531 g->stripe_redundancy = a->stripe_redundancy;
1534 bch2_trans_iter_exit(trans, &iter);
1536 bch2_trans_put(trans);
1542 static void bch2_gc_alloc_reset(struct bch_fs *c, bool metadata_only)
1547 for_each_member_device(ca, c, i) {
1548 struct bucket_array *buckets = gc_bucket_array(ca);
1551 for_each_bucket(g, buckets) {
1552 if (metadata_only &&
1553 (g->data_type == BCH_DATA_user ||
1554 g->data_type == BCH_DATA_cached ||
1555 g->data_type == BCH_DATA_parity))
1558 g->dirty_sectors = 0;
1559 g->cached_sectors = 0;
1564 static int bch2_gc_write_reflink_key(struct btree_trans *trans,
1565 struct btree_iter *iter,
1569 struct bch_fs *c = trans->c;
1570 const __le64 *refcount = bkey_refcount_c(k);
1571 struct printbuf buf = PRINTBUF;
1572 struct reflink_gc *r;
1578 while ((r = genradix_ptr(&c->reflink_gc_table, *idx)) &&
1579 r->offset < k.k->p.offset)
1583 r->offset != k.k->p.offset ||
1584 r->size != k.k->size) {
1585 bch_err(c, "unexpected inconsistency walking reflink table at gc finish");
1589 if (fsck_err_on(r->refcount != le64_to_cpu(*refcount), c,
1590 "reflink key has wrong refcount:\n"
1593 (bch2_bkey_val_to_text(&buf, c, k), buf.buf),
1595 struct bkey_i *new = bch2_bkey_make_mut(trans, iter, &k, 0);
1597 ret = PTR_ERR_OR_ZERO(new);
1602 new->k.type = KEY_TYPE_deleted;
1604 *bkey_refcount(new) = cpu_to_le64(r->refcount);
1607 printbuf_exit(&buf);
1611 static int bch2_gc_reflink_done(struct bch_fs *c, bool metadata_only)
1613 struct btree_trans *trans;
1614 struct btree_iter iter;
1622 trans = bch2_trans_get(c);
1624 ret = for_each_btree_key_commit(trans, iter,
1625 BTREE_ID_reflink, POS_MIN,
1626 BTREE_ITER_PREFETCH, k,
1627 NULL, NULL, BTREE_INSERT_NOFAIL,
1628 bch2_gc_write_reflink_key(trans, &iter, k, &idx));
1630 c->reflink_gc_nr = 0;
1631 bch2_trans_put(trans);
1635 static int bch2_gc_reflink_start(struct bch_fs *c,
1638 struct btree_trans *trans;
1639 struct btree_iter iter;
1641 struct reflink_gc *r;
1647 trans = bch2_trans_get(c);
1648 c->reflink_gc_nr = 0;
1650 for_each_btree_key(trans, iter, BTREE_ID_reflink, POS_MIN,
1651 BTREE_ITER_PREFETCH, k, ret) {
1652 const __le64 *refcount = bkey_refcount_c(k);
1657 r = genradix_ptr_alloc(&c->reflink_gc_table, c->reflink_gc_nr++,
1660 ret = -BCH_ERR_ENOMEM_gc_reflink_start;
1664 r->offset = k.k->p.offset;
1665 r->size = k.k->size;
1668 bch2_trans_iter_exit(trans, &iter);
1670 bch2_trans_put(trans);
1674 static void bch2_gc_reflink_reset(struct bch_fs *c, bool metadata_only)
1676 struct genradix_iter iter;
1677 struct reflink_gc *r;
1679 genradix_for_each(&c->reflink_gc_table, iter, r)
1683 static int bch2_gc_write_stripes_key(struct btree_trans *trans,
1684 struct btree_iter *iter,
1687 struct bch_fs *c = trans->c;
1688 struct printbuf buf = PRINTBUF;
1689 const struct bch_stripe *s;
1690 struct gc_stripe *m;
1695 if (k.k->type != KEY_TYPE_stripe)
1698 s = bkey_s_c_to_stripe(k).v;
1699 m = genradix_ptr(&c->gc_stripes, k.k->p.offset);
1701 for (i = 0; i < s->nr_blocks; i++) {
1702 u32 old = stripe_blockcount_get(s, i);
1703 u32 new = (m ? m->block_sectors[i] : 0);
1706 prt_printf(&buf, "stripe block %u has wrong sector count: got %u, should be %u\n",
1713 bch2_bkey_val_to_text(&buf, c, k);
1715 if (fsck_err_on(bad, c, "%s", buf.buf)) {
1716 struct bkey_i_stripe *new;
1718 new = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1719 ret = PTR_ERR_OR_ZERO(new);
1723 bkey_reassemble(&new->k_i, k);
1725 for (i = 0; i < new->v.nr_blocks; i++)
1726 stripe_blockcount_set(&new->v, i, m ? m->block_sectors[i] : 0);
1728 ret = bch2_trans_update(trans, iter, &new->k_i, 0);
1731 printbuf_exit(&buf);
1735 static int bch2_gc_stripes_done(struct bch_fs *c, bool metadata_only)
1737 struct btree_trans *trans;
1738 struct btree_iter iter;
1745 trans = bch2_trans_get(c);
1747 ret = for_each_btree_key_commit(trans, iter,
1748 BTREE_ID_stripes, POS_MIN,
1749 BTREE_ITER_PREFETCH, k,
1750 NULL, NULL, BTREE_INSERT_NOFAIL,
1751 bch2_gc_write_stripes_key(trans, &iter, k));
1753 bch2_trans_put(trans);
1757 static void bch2_gc_stripes_reset(struct bch_fs *c, bool metadata_only)
1759 genradix_free(&c->gc_stripes);
1763 * bch2_gc - walk _all_ references to buckets, and recompute them:
1765 * @c: filesystem object
1766 * @initial: are we in recovery?
1767 * @metadata_only: are we just checking metadata references, or everything?
1769 * Returns: 0 on success, or standard errcode on failure
1771 * Order matters here:
1772 * - Concurrent GC relies on the fact that we have a total ordering for
1773 * everything that GC walks - see gc_will_visit_node(),
1774 * gc_will_visit_root()
1776 * - also, references move around in the course of index updates and
1777 * various other crap: everything needs to agree on the ordering
1778 * references are allowed to move around in - e.g., we're allowed to
1779 * start with a reference owned by an open_bucket (the allocator) and
1780 * move it to the btree, but not the reverse.
1782 * This is necessary to ensure that gc doesn't miss references that
1783 * move around - if references move backwards in the ordering GC
1784 * uses, GC could skip past them
1786 int bch2_gc(struct bch_fs *c, bool initial, bool metadata_only)
1791 lockdep_assert_held(&c->state_lock);
1793 down_write(&c->gc_lock);
1795 bch2_btree_interior_updates_flush(c);
1797 ret = bch2_gc_start(c) ?:
1798 bch2_gc_alloc_start(c, metadata_only) ?:
1799 bch2_gc_reflink_start(c, metadata_only);
1803 gc_pos_set(c, gc_phase(GC_PHASE_START));
1805 bch2_mark_superblocks(c);
1807 ret = bch2_gc_btrees(c, initial, metadata_only);
1813 bch2_mark_pending_btree_node_frees(c);
1817 if (test_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags) ||
1818 (!iter && bch2_test_restart_gc)) {
1820 bch_info(c, "Unable to fix bucket gens, looping");
1826 * XXX: make sure gens we fixed got saved
1828 bch_info(c, "Second GC pass needed, restarting:");
1829 clear_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1830 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1832 bch2_gc_stripes_reset(c, metadata_only);
1833 bch2_gc_alloc_reset(c, metadata_only);
1834 bch2_gc_reflink_reset(c, metadata_only);
1835 ret = bch2_gc_reset(c);
1839 /* flush fsck errors, reset counters */
1840 bch2_flush_fsck_errs(c);
1845 bch2_journal_block(&c->journal);
1847 ret = bch2_gc_stripes_done(c, metadata_only) ?:
1848 bch2_gc_reflink_done(c, metadata_only) ?:
1849 bch2_gc_alloc_done(c, metadata_only) ?:
1850 bch2_gc_done(c, initial, metadata_only);
1852 bch2_journal_unblock(&c->journal);
1855 percpu_down_write(&c->mark_lock);
1856 /* Indicates that gc is no longer in progress: */
1857 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1860 percpu_up_write(&c->mark_lock);
1862 up_write(&c->gc_lock);
1865 * At startup, allocations can happen directly instead of via the
1866 * allocator thread - issue wakeup in case they blocked on gc_lock:
1868 closure_wake_up(&c->freelist_wait);
1875 static int gc_btree_gens_key(struct btree_trans *trans,
1876 struct btree_iter *iter,
1879 struct bch_fs *c = trans->c;
1880 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1881 const struct bch_extent_ptr *ptr;
1885 percpu_down_read(&c->mark_lock);
1886 bkey_for_each_ptr(ptrs, ptr) {
1887 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1889 if (ptr_stale(ca, ptr) > 16) {
1890 percpu_up_read(&c->mark_lock);
1895 bkey_for_each_ptr(ptrs, ptr) {
1896 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1897 u8 *gen = &ca->oldest_gen[PTR_BUCKET_NR(ca, ptr)];
1899 if (gen_after(*gen, ptr->gen))
1902 percpu_up_read(&c->mark_lock);
1905 u = bch2_bkey_make_mut(trans, iter, &k, 0);
1906 ret = PTR_ERR_OR_ZERO(u);
1910 bch2_extent_normalize(c, bkey_i_to_s(u));
1914 static int bch2_alloc_write_oldest_gen(struct btree_trans *trans, struct btree_iter *iter,
1917 struct bch_dev *ca = bch_dev_bkey_exists(trans->c, iter->pos.inode);
1918 struct bch_alloc_v4 a_convert;
1919 const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &a_convert);
1920 struct bkey_i_alloc_v4 *a_mut;
1923 if (a->oldest_gen == ca->oldest_gen[iter->pos.offset])
1926 a_mut = bch2_alloc_to_v4_mut(trans, k);
1927 ret = PTR_ERR_OR_ZERO(a_mut);
1931 a_mut->v.oldest_gen = ca->oldest_gen[iter->pos.offset];
1932 a_mut->v.data_type = alloc_data_type(a_mut->v, a_mut->v.data_type);
1934 return bch2_trans_update(trans, iter, &a_mut->k_i, 0);
1937 int bch2_gc_gens(struct bch_fs *c)
1939 struct btree_trans *trans;
1940 struct btree_iter iter;
1943 u64 b, start_time = local_clock();
1948 * Ideally we would be using state_lock and not gc_lock here, but that
1949 * introduces a deadlock in the RO path - we currently take the state
1950 * lock at the start of going RO, thus the gc thread may get stuck:
1952 if (!mutex_trylock(&c->gc_gens_lock))
1955 trace_and_count(c, gc_gens_start, c);
1956 down_read(&c->gc_lock);
1957 trans = bch2_trans_get(c);
1959 for_each_member_device(ca, c, i) {
1960 struct bucket_gens *gens;
1962 BUG_ON(ca->oldest_gen);
1964 ca->oldest_gen = kvmalloc(ca->mi.nbuckets, GFP_KERNEL);
1965 if (!ca->oldest_gen) {
1966 percpu_ref_put(&ca->ref);
1967 ret = -BCH_ERR_ENOMEM_gc_gens;
1971 gens = bucket_gens(ca);
1973 for (b = gens->first_bucket;
1974 b < gens->nbuckets; b++)
1975 ca->oldest_gen[b] = gens->b[b];
1978 for (i = 0; i < BTREE_ID_NR; i++)
1979 if (btree_type_has_ptrs(i)) {
1980 c->gc_gens_btree = i;
1981 c->gc_gens_pos = POS_MIN;
1983 ret = for_each_btree_key_commit(trans, iter, i,
1985 BTREE_ITER_PREFETCH|BTREE_ITER_ALL_SNAPSHOTS,
1988 BTREE_INSERT_NOFAIL,
1989 gc_btree_gens_key(trans, &iter, k));
1990 if (ret && !bch2_err_matches(ret, EROFS))
1996 ret = for_each_btree_key_commit(trans, iter, BTREE_ID_alloc,
1998 BTREE_ITER_PREFETCH,
2001 BTREE_INSERT_NOFAIL,
2002 bch2_alloc_write_oldest_gen(trans, &iter, k));
2003 if (ret && !bch2_err_matches(ret, EROFS))
2008 c->gc_gens_btree = 0;
2009 c->gc_gens_pos = POS_MIN;
2013 bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
2014 trace_and_count(c, gc_gens_end, c);
2016 for_each_member_device(ca, c, i) {
2017 kvfree(ca->oldest_gen);
2018 ca->oldest_gen = NULL;
2021 bch2_trans_put(trans);
2022 up_read(&c->gc_lock);
2023 mutex_unlock(&c->gc_gens_lock);
2027 static int bch2_gc_thread(void *arg)
2029 struct bch_fs *c = arg;
2030 struct io_clock *clock = &c->io_clock[WRITE];
2031 unsigned long last = atomic64_read(&clock->now);
2032 unsigned last_kick = atomic_read(&c->kick_gc);
2039 set_current_state(TASK_INTERRUPTIBLE);
2041 if (kthread_should_stop()) {
2042 __set_current_state(TASK_RUNNING);
2046 if (atomic_read(&c->kick_gc) != last_kick)
2049 if (c->btree_gc_periodic) {
2050 unsigned long next = last + c->capacity / 16;
2052 if (atomic64_read(&clock->now) >= next)
2055 bch2_io_clock_schedule_timeout(clock, next);
2062 __set_current_state(TASK_RUNNING);
2064 last = atomic64_read(&clock->now);
2065 last_kick = atomic_read(&c->kick_gc);
2068 * Full gc is currently incompatible with btree key cache:
2071 ret = bch2_gc(c, false, false);
2073 ret = bch2_gc_gens(c);
2078 debug_check_no_locks_held();
2084 void bch2_gc_thread_stop(struct bch_fs *c)
2086 struct task_struct *p;
2089 c->gc_thread = NULL;
2097 int bch2_gc_thread_start(struct bch_fs *c)
2099 struct task_struct *p;
2104 p = kthread_create(bch2_gc_thread, c, "bch-gc/%s", c->name);
2106 bch_err_fn(c, PTR_ERR(p));