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_key_cache.h"
13 #include "btree_locking.h"
14 #include "btree_update_interior.h"
31 #include <linux/slab.h>
32 #include <linux/bitops.h>
33 #include <linux/freezer.h>
34 #include <linux/kthread.h>
35 #include <linux/preempt.h>
36 #include <linux/rcupdate.h>
37 #include <linux/sched/task.h>
38 #include <trace/events/bcachefs.h>
40 #define DROP_THIS_NODE 10
41 #define DROP_PREV_NODE 11
43 static inline void __gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
46 write_seqcount_begin(&c->gc_pos_lock);
48 write_seqcount_end(&c->gc_pos_lock);
52 static inline void gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
54 BUG_ON(gc_pos_cmp(new_pos, c->gc_pos) <= 0);
55 __gc_pos_set(c, new_pos);
59 * Missing: if an interior btree node is empty, we need to do something -
60 * perhaps just kill it
62 static int bch2_gc_check_topology(struct bch_fs *c,
64 struct bkey_buf *prev,
68 struct bpos node_start = b->data->min_key;
69 struct bpos node_end = b->data->max_key;
70 struct bpos expected_start = bkey_deleted(&prev->k->k)
72 : bpos_successor(prev->k->k.p);
73 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
76 if (cur.k->k.type == KEY_TYPE_btree_ptr_v2) {
77 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(cur.k);
79 if (bpos_cmp(expected_start, bp->v.min_key)) {
80 bch2_topology_error(c);
82 if (bkey_deleted(&prev->k->k)) {
83 prt_printf(&buf1, "start of node: ");
84 bch2_bpos_to_text(&buf1, node_start);
86 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(prev->k));
88 bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(cur.k));
94 "btree node with incorrect min_key at btree %s level %u:\n"
97 bch2_btree_ids[b->c.btree_id], b->c.level,
98 buf1.buf, buf2.buf) &&
99 !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags)) {
100 bch_info(c, "Halting mark and sweep to start topology repair pass");
101 ret = FSCK_ERR_START_TOPOLOGY_REPAIR;
104 set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
109 if (is_last && bpos_cmp(cur.k->k.p, node_end)) {
110 bch2_topology_error(c);
112 printbuf_reset(&buf1);
113 printbuf_reset(&buf2);
115 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(cur.k));
116 bch2_bpos_to_text(&buf2, node_end);
122 "btree node with incorrect max_key at btree %s level %u:\n"
125 bch2_btree_ids[b->c.btree_id], b->c.level,
126 buf1.buf, buf2.buf) &&
127 !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags)) {
128 bch_info(c, "Halting mark and sweep to start topology repair pass");
129 ret = FSCK_ERR_START_TOPOLOGY_REPAIR;
132 set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
136 bch2_bkey_buf_copy(prev, c, cur.k);
139 printbuf_exit(&buf2);
140 printbuf_exit(&buf1);
144 static void btree_ptr_to_v2(struct btree *b, struct bkey_i_btree_ptr_v2 *dst)
146 switch (b->key.k.type) {
147 case KEY_TYPE_btree_ptr: {
148 struct bkey_i_btree_ptr *src = bkey_i_to_btree_ptr(&b->key);
152 dst->v.seq = b->data->keys.seq;
153 dst->v.sectors_written = 0;
155 dst->v.min_key = b->data->min_key;
156 set_bkey_val_bytes(&dst->k, sizeof(dst->v) + bkey_val_bytes(&src->k));
157 memcpy(dst->v.start, src->v.start, bkey_val_bytes(&src->k));
160 case KEY_TYPE_btree_ptr_v2:
161 bkey_copy(&dst->k_i, &b->key);
168 static void bch2_btree_node_update_key_early(struct bch_fs *c,
169 enum btree_id btree, unsigned level,
170 struct bkey_s_c old, struct bkey_i *new)
176 bch2_bkey_buf_init(&tmp);
177 bch2_bkey_buf_reassemble(&tmp, c, old);
179 b = bch2_btree_node_get_noiter(c, tmp.k, btree, level, true);
180 if (!IS_ERR_OR_NULL(b)) {
181 mutex_lock(&c->btree_cache.lock);
183 bch2_btree_node_hash_remove(&c->btree_cache, b);
185 bkey_copy(&b->key, new);
186 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
189 mutex_unlock(&c->btree_cache.lock);
190 six_unlock_read(&b->c.lock);
193 bch2_bkey_buf_exit(&tmp, c);
196 static int set_node_min(struct bch_fs *c, struct btree *b, struct bpos new_min)
198 struct bkey_i_btree_ptr_v2 *new;
201 new = kmalloc(BKEY_BTREE_PTR_U64s_MAX * sizeof(u64), GFP_KERNEL);
205 btree_ptr_to_v2(b, new);
206 b->data->min_key = new_min;
207 new->v.min_key = new_min;
208 SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
210 ret = bch2_journal_key_insert_take(c, b->c.btree_id, b->c.level + 1, &new->k_i);
216 bch2_btree_node_drop_keys_outside_node(b);
217 bkey_copy(&b->key, &new->k_i);
221 static int set_node_max(struct bch_fs *c, struct btree *b, struct bpos new_max)
223 struct bkey_i_btree_ptr_v2 *new;
226 ret = bch2_journal_key_delete(c, b->c.btree_id, b->c.level + 1, b->key.k.p);
230 new = kmalloc(BKEY_BTREE_PTR_U64s_MAX * sizeof(u64), GFP_KERNEL);
234 btree_ptr_to_v2(b, new);
235 b->data->max_key = new_max;
237 SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
239 ret = bch2_journal_key_insert_take(c, b->c.btree_id, b->c.level + 1, &new->k_i);
245 bch2_btree_node_drop_keys_outside_node(b);
247 mutex_lock(&c->btree_cache.lock);
248 bch2_btree_node_hash_remove(&c->btree_cache, b);
250 bkey_copy(&b->key, &new->k_i);
251 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
253 mutex_unlock(&c->btree_cache.lock);
257 static int btree_repair_node_boundaries(struct bch_fs *c, struct btree *b,
258 struct btree *prev, struct btree *cur)
260 struct bpos expected_start = !prev
262 : bpos_successor(prev->key.k.p);
263 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
267 prt_printf(&buf1, "start of node: ");
268 bch2_bpos_to_text(&buf1, b->data->min_key);
270 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(&prev->key));
273 bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(&cur->key));
276 bpos_cmp(expected_start, cur->data->min_key) > 0 &&
277 BTREE_NODE_SEQ(cur->data) > BTREE_NODE_SEQ(prev->data)) {
278 /* cur overwrites prev: */
280 if (mustfix_fsck_err_on(bpos_cmp(prev->data->min_key,
281 cur->data->min_key) >= 0, c,
282 "btree node overwritten by next node at btree %s level %u:\n"
285 bch2_btree_ids[b->c.btree_id], b->c.level,
286 buf1.buf, buf2.buf)) {
287 ret = DROP_PREV_NODE;
291 if (mustfix_fsck_err_on(bpos_cmp(prev->key.k.p,
292 bpos_predecessor(cur->data->min_key)), c,
293 "btree node with incorrect max_key at btree %s level %u:\n"
296 bch2_btree_ids[b->c.btree_id], b->c.level,
298 ret = set_node_max(c, prev,
299 bpos_predecessor(cur->data->min_key));
301 /* prev overwrites cur: */
303 if (mustfix_fsck_err_on(bpos_cmp(expected_start,
304 cur->data->max_key) >= 0, c,
305 "btree node overwritten by prev node at btree %s level %u:\n"
308 bch2_btree_ids[b->c.btree_id], b->c.level,
309 buf1.buf, buf2.buf)) {
310 ret = DROP_THIS_NODE;
314 if (mustfix_fsck_err_on(bpos_cmp(expected_start, cur->data->min_key), c,
315 "btree node with incorrect min_key at btree %s level %u:\n"
318 bch2_btree_ids[b->c.btree_id], b->c.level,
320 ret = set_node_min(c, cur, expected_start);
324 printbuf_exit(&buf2);
325 printbuf_exit(&buf1);
329 static int btree_repair_node_end(struct bch_fs *c, struct btree *b,
332 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
335 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(&child->key));
336 bch2_bpos_to_text(&buf2, b->key.k.p);
338 if (mustfix_fsck_err_on(bpos_cmp(child->key.k.p, b->key.k.p), c,
339 "btree node with incorrect max_key at btree %s level %u:\n"
342 bch2_btree_ids[b->c.btree_id], b->c.level,
343 buf1.buf, buf2.buf)) {
344 ret = set_node_max(c, child, b->key.k.p);
350 printbuf_exit(&buf2);
351 printbuf_exit(&buf1);
355 static int bch2_btree_repair_topology_recurse(struct bch_fs *c, struct btree *b)
357 struct btree_and_journal_iter iter;
359 struct bkey_buf prev_k, cur_k;
360 struct btree *prev = NULL, *cur = NULL;
361 bool have_child, dropped_children = false;
362 struct printbuf buf = PRINTBUF;
369 have_child = dropped_children = false;
370 bch2_bkey_buf_init(&prev_k);
371 bch2_bkey_buf_init(&cur_k);
372 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
374 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
375 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
376 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
378 bch2_btree_and_journal_iter_advance(&iter);
379 bch2_bkey_buf_reassemble(&cur_k, c, k);
381 cur = bch2_btree_node_get_noiter(c, cur_k.k,
382 b->c.btree_id, b->c.level - 1,
384 ret = PTR_ERR_OR_ZERO(cur);
386 printbuf_reset(&buf);
387 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(cur_k.k));
389 if (mustfix_fsck_err_on(ret == -EIO, c,
390 "Topology repair: unreadable btree node at btree %s level %u:\n"
392 bch2_btree_ids[b->c.btree_id],
395 bch2_btree_node_evict(c, cur_k.k);
396 ret = bch2_journal_key_delete(c, b->c.btree_id,
397 b->c.level, cur_k.k->k.p);
404 bch_err(c, "%s: error %i getting btree node",
409 ret = btree_repair_node_boundaries(c, b, prev, cur);
411 if (ret == DROP_THIS_NODE) {
412 six_unlock_read(&cur->c.lock);
413 bch2_btree_node_evict(c, cur_k.k);
414 ret = bch2_journal_key_delete(c, b->c.btree_id,
415 b->c.level, cur_k.k->k.p);
422 six_unlock_read(&prev->c.lock);
425 if (ret == DROP_PREV_NODE) {
426 bch2_btree_node_evict(c, prev_k.k);
427 ret = bch2_journal_key_delete(c, b->c.btree_id,
428 b->c.level, prev_k.k->k.p);
432 bch2_btree_and_journal_iter_exit(&iter);
433 bch2_bkey_buf_exit(&prev_k, c);
434 bch2_bkey_buf_exit(&cur_k, c);
441 bch2_bkey_buf_copy(&prev_k, c, cur_k.k);
444 if (!ret && !IS_ERR_OR_NULL(prev)) {
446 ret = btree_repair_node_end(c, b, prev);
449 if (!IS_ERR_OR_NULL(prev))
450 six_unlock_read(&prev->c.lock);
452 if (!IS_ERR_OR_NULL(cur))
453 six_unlock_read(&cur->c.lock);
459 bch2_btree_and_journal_iter_exit(&iter);
460 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
462 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
463 bch2_bkey_buf_reassemble(&cur_k, c, k);
464 bch2_btree_and_journal_iter_advance(&iter);
466 cur = bch2_btree_node_get_noiter(c, cur_k.k,
467 b->c.btree_id, b->c.level - 1,
469 ret = PTR_ERR_OR_ZERO(cur);
472 bch_err(c, "%s: error %i getting btree node",
477 ret = bch2_btree_repair_topology_recurse(c, cur);
478 six_unlock_read(&cur->c.lock);
481 if (ret == DROP_THIS_NODE) {
482 bch2_btree_node_evict(c, cur_k.k);
483 ret = bch2_journal_key_delete(c, b->c.btree_id,
484 b->c.level, cur_k.k->k.p);
485 dropped_children = true;
494 printbuf_reset(&buf);
495 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
497 if (mustfix_fsck_err_on(!have_child, c,
498 "empty interior btree node at btree %s level %u\n"
500 bch2_btree_ids[b->c.btree_id],
501 b->c.level, buf.buf))
502 ret = DROP_THIS_NODE;
505 if (!IS_ERR_OR_NULL(prev))
506 six_unlock_read(&prev->c.lock);
507 if (!IS_ERR_OR_NULL(cur))
508 six_unlock_read(&cur->c.lock);
510 bch2_btree_and_journal_iter_exit(&iter);
511 bch2_bkey_buf_exit(&prev_k, c);
512 bch2_bkey_buf_exit(&cur_k, c);
514 if (!ret && dropped_children)
521 static int bch2_repair_topology(struct bch_fs *c)
527 for (i = 0; i < BTREE_ID_NR && !ret; i++) {
528 b = c->btree_roots[i].b;
529 if (btree_node_fake(b))
532 six_lock_read(&b->c.lock, NULL, NULL);
533 ret = bch2_btree_repair_topology_recurse(c, b);
534 six_unlock_read(&b->c.lock);
536 if (ret == DROP_THIS_NODE) {
537 bch_err(c, "empty btree root - repair unimplemented");
545 static int bch2_check_fix_ptrs(struct bch_fs *c, enum btree_id btree_id,
546 unsigned level, bool is_root,
549 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(*k);
550 const union bch_extent_entry *entry;
551 struct extent_ptr_decoded p = { 0 };
552 bool do_update = false;
553 struct printbuf buf = PRINTBUF;
558 * use check_bucket_ref here
560 bkey_for_each_ptr_decode(k->k, ptrs, p, entry) {
561 struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
562 struct bucket *g = PTR_GC_BUCKET(ca, &p.ptr);
563 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, &entry->ptr);
565 if (fsck_err_on(!g->gen_valid, c,
566 "bucket %u:%zu data type %s ptr gen %u missing in alloc btree\n"
568 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
569 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
571 (printbuf_reset(&buf),
572 bch2_bkey_val_to_text(&buf, c, *k), buf.buf))) {
581 if (fsck_err_on(gen_cmp(p.ptr.gen, g->gen) > 0, c,
582 "bucket %u:%zu data type %s ptr gen in the future: %u > %u\n"
584 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
585 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
587 (printbuf_reset(&buf),
588 bch2_bkey_val_to_text(&buf, c, *k), buf.buf))) {
593 g->dirty_sectors = 0;
594 g->cached_sectors = 0;
595 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
601 if (fsck_err_on(gen_cmp(g->gen, p.ptr.gen) > BUCKET_GC_GEN_MAX, c,
602 "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
604 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr), g->gen,
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)))
611 if (fsck_err_on(!p.ptr.cached &&
612 gen_cmp(p.ptr.gen, g->gen) < 0, c,
613 "bucket %u:%zu data type %s stale dirty ptr: %u < %u\n"
615 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
616 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
618 (printbuf_reset(&buf),
619 bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
622 if (data_type != BCH_DATA_btree && p.ptr.gen != g->gen)
625 if (fsck_err_on(g->data_type &&
626 g->data_type != data_type, c,
627 "bucket %u:%zu different types of data in same bucket: %s, %s\n"
629 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
630 bch2_data_types[g->data_type],
631 bch2_data_types[data_type],
632 (printbuf_reset(&buf),
633 bch2_bkey_val_to_text(&buf, c, *k), buf.buf))) {
634 if (data_type == BCH_DATA_btree) {
635 g->data_type = data_type;
636 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
643 struct gc_stripe *m = genradix_ptr(&c->gc_stripes, p.ec.idx);
645 if (fsck_err_on(!m || !m->alive, c,
646 "pointer to nonexistent stripe %llu\n"
649 (printbuf_reset(&buf),
650 bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
653 if (fsck_err_on(!bch2_ptr_matches_stripe_m(m, p), c,
654 "pointer does not match stripe %llu\n"
657 (printbuf_reset(&buf),
658 bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
664 struct bkey_ptrs ptrs;
665 union bch_extent_entry *entry;
666 struct bch_extent_ptr *ptr;
670 bch_err(c, "cannot update btree roots yet");
675 new = kmalloc(bkey_bytes(k->k), GFP_KERNEL);
677 bch_err(c, "%s: error allocating new key", __func__);
682 bkey_reassemble(new, *k);
686 * We don't want to drop btree node pointers - if the
687 * btree node isn't there anymore, the read path will
690 ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
691 bkey_for_each_ptr(ptrs, ptr) {
692 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
693 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
698 bch2_bkey_drop_ptrs(bkey_i_to_s(new), ptr, ({
699 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
700 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
701 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, ptr);
704 (!g->gen_valid || gen_cmp(ptr->gen, g->gen) > 0)) ||
706 gen_cmp(ptr->gen, g->gen) < 0) ||
707 gen_cmp(g->gen, ptr->gen) > BUCKET_GC_GEN_MAX ||
709 g->data_type != data_type);
712 ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
713 bkey_extent_entry_for_each(ptrs, entry) {
714 if (extent_entry_type(entry) == BCH_EXTENT_ENTRY_stripe_ptr) {
715 struct gc_stripe *m = genradix_ptr(&c->gc_stripes,
716 entry->stripe_ptr.idx);
717 union bch_extent_entry *next_ptr;
719 bkey_extent_entry_for_each_from(ptrs, next_ptr, entry)
720 if (extent_entry_type(next_ptr) == BCH_EXTENT_ENTRY_ptr)
725 bch_err(c, "aieee, found stripe ptr with no data ptr");
729 if (!m || !m->alive ||
730 !__bch2_ptr_matches_stripe(&m->ptrs[entry->stripe_ptr.block],
733 bch2_bkey_extent_entry_drop(new, entry);
740 ret = bch2_journal_key_insert_take(c, btree_id, level, new);
747 bch2_btree_node_update_key_early(c, btree_id, level - 1, *k, new);
749 printbuf_reset(&buf);
750 bch2_bkey_val_to_text(&buf, c, *k);
751 bch_info(c, "updated %s", buf.buf);
753 printbuf_reset(&buf);
754 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(new));
755 bch_info(c, "new key %s", buf.buf);
757 *k = bkey_i_to_s_c(new);
765 /* marking of btree keys/nodes: */
767 static int bch2_gc_mark_key(struct btree_trans *trans, enum btree_id btree_id,
768 unsigned level, bool is_root,
772 struct bch_fs *c = trans->c;
773 struct bkey deleted = KEY(0, 0, 0);
774 struct bkey_s_c old = (struct bkey_s_c) { &deleted, NULL };
777 (initial ? BTREE_TRIGGER_NOATOMIC : 0);
783 BUG_ON(bch2_journal_seq_verify &&
784 k->k->version.lo > atomic64_read(&c->journal.seq));
786 ret = bch2_check_fix_ptrs(c, btree_id, level, is_root, k);
790 if (fsck_err_on(k->k->version.lo > atomic64_read(&c->key_version), c,
791 "key version number higher than recorded: %llu > %llu",
793 atomic64_read(&c->key_version)))
794 atomic64_set(&c->key_version, k->k->version.lo);
797 ret = __bch2_trans_do(trans, NULL, NULL, 0,
798 bch2_mark_key(trans, old, *k, flags));
802 bch_err(c, "%s: ret %i", __func__, ret);
806 static int btree_gc_mark_node(struct btree_trans *trans, struct btree *b, bool initial)
808 struct bch_fs *c = trans->c;
809 struct btree_node_iter iter;
810 struct bkey unpacked;
812 struct bkey_buf prev, cur;
815 if (!btree_node_type_needs_gc(btree_node_type(b)))
818 bch2_btree_node_iter_init_from_start(&iter, b);
819 bch2_bkey_buf_init(&prev);
820 bch2_bkey_buf_init(&cur);
821 bkey_init(&prev.k->k);
823 while ((k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked)).k) {
824 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, false,
829 bch2_btree_node_iter_advance(&iter, b);
832 bch2_bkey_buf_reassemble(&cur, c, k);
834 ret = bch2_gc_check_topology(c, b, &prev, cur,
835 bch2_btree_node_iter_end(&iter));
841 bch2_bkey_buf_exit(&cur, c);
842 bch2_bkey_buf_exit(&prev, c);
846 static int bch2_gc_btree(struct btree_trans *trans, enum btree_id btree_id,
847 bool initial, bool metadata_only)
849 struct bch_fs *c = trans->c;
850 struct btree_iter iter;
852 unsigned depth = metadata_only ? 1
853 : bch2_expensive_debug_checks ? 0
854 : !btree_node_type_needs_gc(btree_id) ? 1
858 gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));
860 __for_each_btree_node(trans, iter, btree_id, POS_MIN,
861 0, depth, BTREE_ITER_PREFETCH, b, ret) {
862 bch2_verify_btree_nr_keys(b);
864 gc_pos_set(c, gc_pos_btree_node(b));
866 ret = btree_gc_mark_node(trans, b, initial);
870 bch2_trans_iter_exit(trans, &iter);
875 mutex_lock(&c->btree_root_lock);
876 b = c->btree_roots[btree_id].b;
877 if (!btree_node_fake(b)) {
878 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
880 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level,
883 gc_pos_set(c, gc_pos_btree_root(b->c.btree_id));
884 mutex_unlock(&c->btree_root_lock);
889 static int bch2_gc_btree_init_recurse(struct btree_trans *trans, struct btree *b,
890 unsigned target_depth)
892 struct bch_fs *c = trans->c;
893 struct btree_and_journal_iter iter;
895 struct bkey_buf cur, prev;
896 struct printbuf buf = PRINTBUF;
899 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
900 bch2_bkey_buf_init(&prev);
901 bch2_bkey_buf_init(&cur);
902 bkey_init(&prev.k->k);
904 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
905 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
906 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
908 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level,
911 bch_err(c, "%s: error %i from bch2_gc_mark_key", __func__, ret);
916 bch2_bkey_buf_reassemble(&cur, c, k);
917 k = bkey_i_to_s_c(cur.k);
919 bch2_btree_and_journal_iter_advance(&iter);
921 ret = bch2_gc_check_topology(c, b,
923 !bch2_btree_and_journal_iter_peek(&iter).k);
927 bch2_btree_and_journal_iter_advance(&iter);
931 if (b->c.level > target_depth) {
932 bch2_btree_and_journal_iter_exit(&iter);
933 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
935 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
938 bch2_bkey_buf_reassemble(&cur, c, k);
939 bch2_btree_and_journal_iter_advance(&iter);
941 child = bch2_btree_node_get_noiter(c, cur.k,
942 b->c.btree_id, b->c.level - 1,
944 ret = PTR_ERR_OR_ZERO(child);
947 bch2_topology_error(c);
953 "Unreadable btree node at btree %s level %u:\n"
955 bch2_btree_ids[b->c.btree_id],
957 (printbuf_reset(&buf),
958 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(cur.k)), buf.buf)) &&
959 !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags)) {
960 ret = FSCK_ERR_START_TOPOLOGY_REPAIR;
961 bch_info(c, "Halting mark and sweep to start topology repair pass");
964 /* Continue marking when opted to not
967 set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
971 bch_err(c, "%s: error %i getting btree node",
976 ret = bch2_gc_btree_init_recurse(trans, child,
978 six_unlock_read(&child->c.lock);
985 bch2_bkey_buf_exit(&cur, c);
986 bch2_bkey_buf_exit(&prev, c);
987 bch2_btree_and_journal_iter_exit(&iter);
992 static int bch2_gc_btree_init(struct btree_trans *trans,
993 enum btree_id btree_id,
996 struct bch_fs *c = trans->c;
998 unsigned target_depth = metadata_only ? 1
999 : bch2_expensive_debug_checks ? 0
1000 : !btree_node_type_needs_gc(btree_id) ? 1
1002 struct printbuf buf = PRINTBUF;
1005 b = c->btree_roots[btree_id].b;
1007 if (btree_node_fake(b))
1010 six_lock_read(&b->c.lock, NULL, NULL);
1011 printbuf_reset(&buf);
1012 bch2_bpos_to_text(&buf, b->data->min_key);
1013 if (mustfix_fsck_err_on(bpos_cmp(b->data->min_key, POS_MIN), c,
1014 "btree root with incorrect min_key: %s", buf.buf)) {
1015 bch_err(c, "repair unimplemented");
1016 ret = FSCK_ERR_EXIT;
1020 printbuf_reset(&buf);
1021 bch2_bpos_to_text(&buf, b->data->max_key);
1022 if (mustfix_fsck_err_on(bpos_cmp(b->data->max_key, SPOS_MAX), c,
1023 "btree root with incorrect max_key: %s", buf.buf)) {
1024 bch_err(c, "repair unimplemented");
1025 ret = FSCK_ERR_EXIT;
1029 if (b->c.level >= target_depth)
1030 ret = bch2_gc_btree_init_recurse(trans, b, target_depth);
1033 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1035 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, true,
1039 six_unlock_read(&b->c.lock);
1042 bch_err(c, "%s: ret %i", __func__, ret);
1043 printbuf_exit(&buf);
1047 static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
1049 return (int) btree_id_to_gc_phase(l) -
1050 (int) btree_id_to_gc_phase(r);
1053 static int bch2_gc_btrees(struct bch_fs *c, bool initial, bool metadata_only)
1055 struct btree_trans trans;
1056 enum btree_id ids[BTREE_ID_NR];
1060 bch2_trans_init(&trans, c, 0, 0);
1063 trans.is_initial_gc = true;
1065 for (i = 0; i < BTREE_ID_NR; i++)
1067 bubble_sort(ids, BTREE_ID_NR, btree_id_gc_phase_cmp);
1069 for (i = 0; i < BTREE_ID_NR && !ret; i++)
1071 ? bch2_gc_btree_init(&trans, ids[i], metadata_only)
1072 : bch2_gc_btree(&trans, ids[i], initial, metadata_only);
1075 bch_err(c, "%s: ret %i", __func__, ret);
1077 bch2_trans_exit(&trans);
1081 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
1083 enum bch_data_type type,
1086 u64 b = sector_to_bucket(ca, start);
1090 min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1092 bch2_mark_metadata_bucket(c, ca, b, type, sectors,
1093 gc_phase(GC_PHASE_SB), flags);
1096 } while (start < end);
1099 static void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
1102 struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1106 for (i = 0; i < layout->nr_superblocks; i++) {
1107 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1109 if (offset == BCH_SB_SECTOR)
1110 mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
1111 BCH_DATA_sb, flags);
1113 mark_metadata_sectors(c, ca, offset,
1114 offset + (1 << layout->sb_max_size_bits),
1115 BCH_DATA_sb, flags);
1118 for (i = 0; i < ca->journal.nr; i++) {
1119 b = ca->journal.buckets[i];
1120 bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_journal,
1122 gc_phase(GC_PHASE_SB), flags);
1126 static void bch2_mark_superblocks(struct bch_fs *c)
1131 mutex_lock(&c->sb_lock);
1132 gc_pos_set(c, gc_phase(GC_PHASE_SB));
1134 for_each_online_member(ca, c, i)
1135 bch2_mark_dev_superblock(c, ca, BTREE_TRIGGER_GC);
1136 mutex_unlock(&c->sb_lock);
1140 /* Also see bch2_pending_btree_node_free_insert_done() */
1141 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
1143 struct btree_update *as;
1144 struct pending_btree_node_free *d;
1146 mutex_lock(&c->btree_interior_update_lock);
1147 gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
1149 for_each_pending_btree_node_free(c, as, d)
1150 if (d->index_update_done)
1151 bch2_mark_key(c, bkey_i_to_s_c(&d->key), BTREE_TRIGGER_GC);
1153 mutex_unlock(&c->btree_interior_update_lock);
1157 static void bch2_gc_free(struct bch_fs *c)
1162 genradix_free(&c->reflink_gc_table);
1163 genradix_free(&c->gc_stripes);
1165 for_each_member_device(ca, c, i) {
1166 kvpfree(rcu_dereference_protected(ca->buckets_gc, 1),
1167 sizeof(struct bucket_array) +
1168 ca->mi.nbuckets * sizeof(struct bucket));
1169 ca->buckets_gc = NULL;
1171 free_percpu(ca->usage_gc);
1172 ca->usage_gc = NULL;
1175 free_percpu(c->usage_gc);
1179 static int bch2_gc_done(struct bch_fs *c,
1180 bool initial, bool metadata_only)
1182 struct bch_dev *ca = NULL;
1183 struct printbuf buf = PRINTBUF;
1184 bool verify = !metadata_only && (!initial ||
1185 (c->sb.compat & (1ULL << BCH_COMPAT_alloc_info)));
1189 percpu_down_write(&c->mark_lock);
1191 #define copy_field(_f, _msg, ...) \
1192 if (dst->_f != src->_f) { \
1194 fsck_err(c, _msg ": got %llu, should be %llu" \
1195 , ##__VA_ARGS__, dst->_f, src->_f); \
1196 dst->_f = src->_f; \
1198 #define copy_stripe_field(_f, _msg, ...) \
1199 if (dst->_f != src->_f) { \
1201 fsck_err(c, "stripe %zu has wrong "_msg \
1202 ": got %u, should be %u", \
1203 iter.pos, ##__VA_ARGS__, \
1204 dst->_f, src->_f); \
1205 dst->_f = src->_f; \
1207 #define copy_dev_field(_f, _msg, ...) \
1208 copy_field(_f, "dev %u has wrong " _msg, dev, ##__VA_ARGS__)
1209 #define copy_fs_field(_f, _msg, ...) \
1210 copy_field(_f, "fs has wrong " _msg, ##__VA_ARGS__)
1212 for (i = 0; i < ARRAY_SIZE(c->usage); i++)
1213 bch2_fs_usage_acc_to_base(c, i);
1215 for_each_member_device(ca, c, dev) {
1216 struct bch_dev_usage *dst = ca->usage_base;
1217 struct bch_dev_usage *src = (void *)
1218 bch2_acc_percpu_u64s((void *) ca->usage_gc,
1221 copy_dev_field(buckets_ec, "buckets_ec");
1223 for (i = 0; i < BCH_DATA_NR; i++) {
1224 copy_dev_field(d[i].buckets, "%s buckets", bch2_data_types[i]);
1225 copy_dev_field(d[i].sectors, "%s sectors", bch2_data_types[i]);
1226 copy_dev_field(d[i].fragmented, "%s fragmented", bch2_data_types[i]);
1231 unsigned nr = fs_usage_u64s(c);
1232 struct bch_fs_usage *dst = c->usage_base;
1233 struct bch_fs_usage *src = (void *)
1234 bch2_acc_percpu_u64s((void *) c->usage_gc, nr);
1236 copy_fs_field(hidden, "hidden");
1237 copy_fs_field(btree, "btree");
1239 if (!metadata_only) {
1240 copy_fs_field(data, "data");
1241 copy_fs_field(cached, "cached");
1242 copy_fs_field(reserved, "reserved");
1243 copy_fs_field(nr_inodes,"nr_inodes");
1245 for (i = 0; i < BCH_REPLICAS_MAX; i++)
1246 copy_fs_field(persistent_reserved[i],
1247 "persistent_reserved[%i]", i);
1250 for (i = 0; i < c->replicas.nr; i++) {
1251 struct bch_replicas_entry *e =
1252 cpu_replicas_entry(&c->replicas, i);
1254 if (metadata_only &&
1255 (e->data_type == BCH_DATA_user ||
1256 e->data_type == BCH_DATA_cached))
1259 printbuf_reset(&buf);
1260 bch2_replicas_entry_to_text(&buf, e);
1262 copy_fs_field(replicas[i], "%s", buf.buf);
1266 #undef copy_fs_field
1267 #undef copy_dev_field
1268 #undef copy_stripe_field
1272 percpu_ref_put(&ca->ref);
1274 bch_err(c, "%s: ret %i", __func__, ret);
1276 percpu_up_write(&c->mark_lock);
1277 printbuf_exit(&buf);
1281 static int bch2_gc_start(struct bch_fs *c,
1284 struct bch_dev *ca = NULL;
1287 BUG_ON(c->usage_gc);
1289 c->usage_gc = __alloc_percpu_gfp(fs_usage_u64s(c) * sizeof(u64),
1290 sizeof(u64), GFP_KERNEL);
1292 bch_err(c, "error allocating c->usage_gc");
1296 for_each_member_device(ca, c, i) {
1297 BUG_ON(ca->buckets_gc);
1298 BUG_ON(ca->usage_gc);
1300 ca->usage_gc = alloc_percpu(struct bch_dev_usage);
1301 if (!ca->usage_gc) {
1302 bch_err(c, "error allocating ca->usage_gc");
1303 percpu_ref_put(&ca->ref);
1307 this_cpu_write(ca->usage_gc->d[BCH_DATA_free].buckets,
1308 ca->mi.nbuckets - ca->mi.first_bucket);
1314 /* returns true if not equal */
1315 static inline bool bch2_alloc_v4_cmp(struct bch_alloc_v4 l,
1316 struct bch_alloc_v4 r)
1318 return l.gen != r.gen ||
1319 l.oldest_gen != r.oldest_gen ||
1320 l.data_type != r.data_type ||
1321 l.dirty_sectors != r.dirty_sectors ||
1322 l.cached_sectors != r.cached_sectors ||
1323 l.stripe_redundancy != r.stripe_redundancy ||
1324 l.stripe != r.stripe;
1327 static int bch2_alloc_write_key(struct btree_trans *trans,
1328 struct btree_iter *iter,
1331 struct bch_fs *c = trans->c;
1332 struct bch_dev *ca = bch_dev_bkey_exists(c, iter->pos.inode);
1333 struct bucket gc, *b;
1335 struct bkey_i_alloc_v4 *a;
1336 struct bch_alloc_v4 old, new;
1337 enum bch_data_type type;
1340 k = bch2_btree_iter_peek_slot(iter);
1345 bch2_alloc_to_v4(k, &old);
1348 percpu_down_read(&c->mark_lock);
1349 b = gc_bucket(ca, iter->pos.offset);
1352 * b->data_type doesn't yet include need_discard & need_gc_gen states -
1355 type = __alloc_data_type(b->dirty_sectors,
1360 if (b->data_type != type) {
1361 struct bch_dev_usage *u;
1364 u = this_cpu_ptr(ca->usage_gc);
1365 u->d[b->data_type].buckets--;
1366 b->data_type = type;
1367 u->d[b->data_type].buckets++;
1372 percpu_up_read(&c->mark_lock);
1374 if (metadata_only &&
1375 gc.data_type != BCH_DATA_sb &&
1376 gc.data_type != BCH_DATA_journal &&
1377 gc.data_type != BCH_DATA_btree)
1380 if (gen_after(old.gen, gc.gen))
1383 #define copy_bucket_field(_f) \
1384 if (fsck_err_on(new._f != gc._f, c, \
1385 "bucket %llu:%llu gen %u data type %s has wrong " #_f \
1386 ": got %u, should be %u", \
1387 iter->pos.inode, iter->pos.offset, \
1389 bch2_data_types[gc.data_type], \
1393 copy_bucket_field(gen);
1394 copy_bucket_field(data_type);
1395 copy_bucket_field(dirty_sectors);
1396 copy_bucket_field(cached_sectors);
1397 copy_bucket_field(stripe_redundancy);
1398 copy_bucket_field(stripe);
1399 #undef copy_bucket_field
1401 if (!bch2_alloc_v4_cmp(old, new))
1404 a = bch2_alloc_to_v4_mut(trans, k);
1405 ret = PTR_ERR_OR_ZERO(a);
1412 * The trigger normally makes sure this is set, but we're not running
1415 if (a->v.data_type == BCH_DATA_cached && !a->v.io_time[READ])
1416 a->v.io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
1418 ret = bch2_trans_update(trans, iter, &a->k_i, BTREE_TRIGGER_NORUN);
1423 static int bch2_gc_alloc_done(struct bch_fs *c, bool metadata_only)
1425 struct btree_trans trans;
1426 struct btree_iter iter;
1432 bch2_trans_init(&trans, c, 0, 0);
1434 for_each_member_device(ca, c, i) {
1435 for_each_btree_key(&trans, iter, BTREE_ID_alloc,
1436 POS(ca->dev_idx, ca->mi.first_bucket),
1438 BTREE_ITER_PREFETCH, k, ret) {
1439 if (bkey_cmp(iter.pos, POS(ca->dev_idx, ca->mi.nbuckets)) >= 0)
1442 ret = __bch2_trans_do(&trans, NULL, NULL,
1443 BTREE_INSERT_LAZY_RW,
1444 bch2_alloc_write_key(&trans, &iter,
1449 bch2_trans_iter_exit(&trans, &iter);
1452 bch_err(c, "error writing alloc info: %i", ret);
1453 percpu_ref_put(&ca->ref);
1458 bch2_trans_exit(&trans);
1462 static int bch2_gc_alloc_start(struct bch_fs *c, bool metadata_only)
1465 struct btree_trans trans;
1466 struct btree_iter iter;
1469 struct bch_alloc_v4 a;
1473 for_each_member_device(ca, c, i) {
1474 struct bucket_array *buckets = kvpmalloc(sizeof(struct bucket_array) +
1475 ca->mi.nbuckets * sizeof(struct bucket),
1476 GFP_KERNEL|__GFP_ZERO);
1478 percpu_ref_put(&ca->ref);
1479 bch_err(c, "error allocating ca->buckets[gc]");
1483 buckets->first_bucket = ca->mi.first_bucket;
1484 buckets->nbuckets = ca->mi.nbuckets;
1485 rcu_assign_pointer(ca->buckets_gc, buckets);
1488 bch2_trans_init(&trans, c, 0, 0);
1490 for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
1491 BTREE_ITER_PREFETCH, k, ret) {
1492 ca = bch_dev_bkey_exists(c, k.k->p.inode);
1493 g = gc_bucket(ca, k.k->p.offset);
1495 bch2_alloc_to_v4(k, &a);
1500 if (metadata_only &&
1501 (a.data_type == BCH_DATA_user ||
1502 a.data_type == BCH_DATA_cached ||
1503 a.data_type == BCH_DATA_parity)) {
1504 g->data_type = a.data_type;
1505 g->dirty_sectors = a.dirty_sectors;
1506 g->cached_sectors = a.cached_sectors;
1507 g->stripe = a.stripe;
1508 g->stripe_redundancy = a.stripe_redundancy;
1511 bch2_trans_iter_exit(&trans, &iter);
1513 bch2_trans_exit(&trans);
1516 bch_err(c, "error reading alloc info at gc start: %i", ret);
1521 static void bch2_gc_alloc_reset(struct bch_fs *c, bool metadata_only)
1526 for_each_member_device(ca, c, i) {
1527 struct bucket_array *buckets = gc_bucket_array(ca);
1530 for_each_bucket(g, buckets) {
1531 if (metadata_only &&
1532 (g->data_type == BCH_DATA_user ||
1533 g->data_type == BCH_DATA_cached ||
1534 g->data_type == BCH_DATA_parity))
1537 g->dirty_sectors = 0;
1538 g->cached_sectors = 0;
1543 static int bch2_gc_reflink_done(struct bch_fs *c, bool metadata_only)
1545 struct btree_trans trans;
1546 struct btree_iter iter;
1548 struct reflink_gc *r;
1550 struct printbuf buf = PRINTBUF;
1556 bch2_trans_init(&trans, c, 0, 0);
1558 for_each_btree_key(&trans, iter, BTREE_ID_reflink, POS_MIN,
1559 BTREE_ITER_PREFETCH, k, ret) {
1560 const __le64 *refcount = bkey_refcount_c(k);
1565 r = genradix_ptr(&c->reflink_gc_table, idx++);
1567 r->offset != k.k->p.offset ||
1568 r->size != k.k->size) {
1569 bch_err(c, "unexpected inconsistency walking reflink table at gc finish");
1574 if (fsck_err_on(r->refcount != le64_to_cpu(*refcount), c,
1575 "reflink key has wrong refcount:\n"
1578 (printbuf_reset(&buf),
1579 bch2_bkey_val_to_text(&buf, c, k), buf.buf),
1583 new = kmalloc(bkey_bytes(k.k), GFP_KERNEL);
1589 bkey_reassemble(new, k);
1592 new->k.type = KEY_TYPE_deleted;
1594 *bkey_refcount(new) = cpu_to_le64(r->refcount);
1596 ret = __bch2_trans_do(&trans, NULL, NULL, 0,
1597 __bch2_btree_insert(&trans, BTREE_ID_reflink, new));
1605 bch2_trans_iter_exit(&trans, &iter);
1606 c->reflink_gc_nr = 0;
1607 bch2_trans_exit(&trans);
1608 printbuf_exit(&buf);
1612 static int bch2_gc_reflink_start(struct bch_fs *c,
1615 struct btree_trans trans;
1616 struct btree_iter iter;
1618 struct reflink_gc *r;
1624 bch2_trans_init(&trans, c, 0, 0);
1625 c->reflink_gc_nr = 0;
1627 for_each_btree_key(&trans, iter, BTREE_ID_reflink, POS_MIN,
1628 BTREE_ITER_PREFETCH, k, ret) {
1629 const __le64 *refcount = bkey_refcount_c(k);
1634 r = genradix_ptr_alloc(&c->reflink_gc_table, c->reflink_gc_nr++,
1641 r->offset = k.k->p.offset;
1642 r->size = k.k->size;
1645 bch2_trans_iter_exit(&trans, &iter);
1647 bch2_trans_exit(&trans);
1651 static void bch2_gc_reflink_reset(struct bch_fs *c, bool metadata_only)
1653 struct genradix_iter iter;
1654 struct reflink_gc *r;
1656 genradix_for_each(&c->reflink_gc_table, iter, r)
1660 static int bch2_gc_stripes_done(struct bch_fs *c, bool metadata_only)
1662 struct btree_trans trans;
1663 struct btree_iter iter;
1665 struct gc_stripe *m;
1666 const struct bch_stripe *s;
1667 struct printbuf buf = PRINTBUF;
1674 bch2_trans_init(&trans, c, 0, 0);
1676 for_each_btree_key(&trans, iter, BTREE_ID_stripes, POS_MIN,
1677 BTREE_ITER_PREFETCH, k, ret) {
1678 if (k.k->type != KEY_TYPE_stripe)
1681 s = bkey_s_c_to_stripe(k).v;
1682 m = genradix_ptr(&c->gc_stripes, k.k->p.offset);
1684 for (i = 0; i < s->nr_blocks; i++)
1685 if (stripe_blockcount_get(s, i) != (m ? m->block_sectors[i] : 0))
1689 if (fsck_err_on(true, c,
1690 "stripe has wrong block sector count %u:\n"
1693 (printbuf_reset(&buf),
1694 bch2_bkey_val_to_text(&buf, c, k), buf.buf),
1695 m ? m->block_sectors[i] : 0)) {
1696 struct bkey_i_stripe *new;
1698 new = kmalloc(bkey_bytes(k.k), GFP_KERNEL);
1704 bkey_reassemble(&new->k_i, k);
1706 for (i = 0; i < new->v.nr_blocks; i++)
1707 stripe_blockcount_set(&new->v, i, m ? m->block_sectors[i] : 0);
1709 ret = __bch2_trans_do(&trans, NULL, NULL, 0,
1710 __bch2_btree_insert(&trans, BTREE_ID_reflink, &new->k_i));
1715 bch2_trans_iter_exit(&trans, &iter);
1717 bch2_trans_exit(&trans);
1719 printbuf_exit(&buf);
1723 static void bch2_gc_stripes_reset(struct bch_fs *c, bool metadata_only)
1725 genradix_free(&c->gc_stripes);
1729 * bch2_gc - walk _all_ references to buckets, and recompute them:
1731 * Order matters here:
1732 * - Concurrent GC relies on the fact that we have a total ordering for
1733 * everything that GC walks - see gc_will_visit_node(),
1734 * gc_will_visit_root()
1736 * - also, references move around in the course of index updates and
1737 * various other crap: everything needs to agree on the ordering
1738 * references are allowed to move around in - e.g., we're allowed to
1739 * start with a reference owned by an open_bucket (the allocator) and
1740 * move it to the btree, but not the reverse.
1742 * This is necessary to ensure that gc doesn't miss references that
1743 * move around - if references move backwards in the ordering GC
1744 * uses, GC could skip past them
1746 int bch2_gc(struct bch_fs *c, bool initial, bool metadata_only)
1751 lockdep_assert_held(&c->state_lock);
1753 down_write(&c->gc_lock);
1755 bch2_btree_interior_updates_flush(c);
1757 ret = bch2_gc_start(c, metadata_only) ?:
1758 bch2_gc_alloc_start(c, metadata_only) ?:
1759 bch2_gc_reflink_start(c, metadata_only);
1763 gc_pos_set(c, gc_phase(GC_PHASE_START));
1765 bch2_mark_superblocks(c);
1767 if (BCH_SB_HAS_TOPOLOGY_ERRORS(c->disk_sb.sb) &&
1768 !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags) &&
1769 c->opts.fix_errors != FSCK_OPT_NO) {
1770 bch_info(c, "Starting topology repair pass");
1771 ret = bch2_repair_topology(c);
1774 bch_info(c, "Topology repair pass done");
1776 set_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags);
1779 ret = bch2_gc_btrees(c, initial, metadata_only);
1781 if (ret == FSCK_ERR_START_TOPOLOGY_REPAIR &&
1782 !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags) &&
1783 !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags)) {
1784 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1785 SET_BCH_SB_HAS_TOPOLOGY_ERRORS(c->disk_sb.sb, true);
1789 if (ret == FSCK_ERR_START_TOPOLOGY_REPAIR)
1790 ret = FSCK_ERR_EXIT;
1796 bch2_mark_pending_btree_node_frees(c);
1800 if (test_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags) ||
1801 (!iter && bch2_test_restart_gc)) {
1803 bch_info(c, "Unable to fix bucket gens, looping");
1809 * XXX: make sure gens we fixed got saved
1811 bch_info(c, "Second GC pass needed, restarting:");
1812 clear_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1813 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1815 bch2_gc_stripes_reset(c, metadata_only);
1816 bch2_gc_alloc_reset(c, metadata_only);
1817 bch2_gc_reflink_reset(c, metadata_only);
1819 /* flush fsck errors, reset counters */
1820 bch2_flush_fsck_errs(c);
1825 bch2_journal_block(&c->journal);
1827 ret = bch2_gc_stripes_done(c, metadata_only) ?:
1828 bch2_gc_reflink_done(c, metadata_only) ?:
1829 bch2_gc_alloc_done(c, metadata_only) ?:
1830 bch2_gc_done(c, initial, metadata_only);
1832 bch2_journal_unblock(&c->journal);
1835 percpu_down_write(&c->mark_lock);
1836 /* Indicates that gc is no longer in progress: */
1837 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1840 percpu_up_write(&c->mark_lock);
1842 up_write(&c->gc_lock);
1845 * At startup, allocations can happen directly instead of via the
1846 * allocator thread - issue wakeup in case they blocked on gc_lock:
1848 closure_wake_up(&c->freelist_wait);
1852 static bool gc_btree_gens_key(struct bch_fs *c, struct bkey_s_c k)
1854 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1855 const struct bch_extent_ptr *ptr;
1857 percpu_down_read(&c->mark_lock);
1858 bkey_for_each_ptr(ptrs, ptr) {
1859 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1861 if (ptr_stale(ca, ptr) > 16) {
1862 percpu_up_read(&c->mark_lock);
1867 bkey_for_each_ptr(ptrs, ptr) {
1868 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1869 u8 *gen = &ca->oldest_gen[PTR_BUCKET_NR(ca, ptr)];
1871 if (gen_after(*gen, ptr->gen))
1874 percpu_up_read(&c->mark_lock);
1880 * For recalculating oldest gen, we only need to walk keys in leaf nodes; btree
1881 * node pointers currently never have cached pointers that can become stale:
1883 static int bch2_gc_btree_gens(struct btree_trans *trans, enum btree_id btree_id)
1885 struct bch_fs *c = trans->c;
1886 struct btree_iter iter;
1889 int ret = 0, commit_err = 0;
1891 bch2_bkey_buf_init(&sk);
1893 bch2_trans_iter_init(trans, &iter, btree_id, POS_MIN,
1894 BTREE_ITER_PREFETCH|
1895 BTREE_ITER_NOT_EXTENTS|
1896 BTREE_ITER_ALL_SNAPSHOTS);
1898 while ((bch2_trans_begin(trans),
1899 k = bch2_btree_iter_peek(&iter)).k) {
1907 c->gc_gens_pos = iter.pos;
1909 if (gc_btree_gens_key(c, k) && !commit_err) {
1910 bch2_bkey_buf_reassemble(&sk, c, k);
1911 bch2_extent_normalize(c, bkey_i_to_s(sk.k));
1914 bch2_trans_update(trans, &iter, sk.k, 0) ?:
1915 bch2_trans_commit(trans, NULL, NULL,
1916 BTREE_INSERT_NOWAIT|
1917 BTREE_INSERT_NOFAIL);
1918 if (commit_err == -EINTR) {
1924 bch2_btree_iter_advance(&iter);
1926 bch2_trans_iter_exit(trans, &iter);
1928 bch2_bkey_buf_exit(&sk, c);
1933 static int bch2_alloc_write_oldest_gen(struct btree_trans *trans, struct btree_iter *iter)
1935 struct bch_dev *ca = bch_dev_bkey_exists(trans->c, iter->pos.inode);
1937 struct bch_alloc_v4 a;
1938 struct bkey_i_alloc_v4 *a_mut;
1941 k = bch2_btree_iter_peek_slot(iter);
1946 bch2_alloc_to_v4(k, &a);
1948 if (a.oldest_gen == ca->oldest_gen[iter->pos.offset])
1951 a_mut = bch2_alloc_to_v4_mut(trans, k);
1952 ret = PTR_ERR_OR_ZERO(a_mut);
1956 a_mut->v.oldest_gen = ca->oldest_gen[iter->pos.offset];
1957 a_mut->v.data_type = alloc_data_type(a_mut->v, a_mut->v.data_type);
1959 return bch2_trans_update(trans, iter, &a_mut->k_i, 0);
1962 int bch2_gc_gens(struct bch_fs *c)
1964 struct btree_trans trans;
1965 struct btree_iter iter;
1968 u64 b, start_time = local_clock();
1973 * Ideally we would be using state_lock and not gc_lock here, but that
1974 * introduces a deadlock in the RO path - we currently take the state
1975 * lock at the start of going RO, thus the gc thread may get stuck:
1977 if (!mutex_trylock(&c->gc_gens_lock))
1980 trace_gc_gens_start(c);
1981 down_read(&c->gc_lock);
1982 bch2_trans_init(&trans, c, 0, 0);
1984 for_each_member_device(ca, c, i) {
1985 struct bucket_gens *gens;
1987 BUG_ON(ca->oldest_gen);
1989 ca->oldest_gen = kvmalloc(ca->mi.nbuckets, GFP_KERNEL);
1990 if (!ca->oldest_gen) {
1991 percpu_ref_put(&ca->ref);
1996 gens = bucket_gens(ca);
1998 for (b = gens->first_bucket;
1999 b < gens->nbuckets; b++)
2000 ca->oldest_gen[b] = gens->b[b];
2003 for (i = 0; i < BTREE_ID_NR; i++)
2004 if ((1 << i) & BTREE_ID_HAS_PTRS) {
2005 c->gc_gens_btree = i;
2006 c->gc_gens_pos = POS_MIN;
2007 ret = bch2_gc_btree_gens(&trans, i);
2009 bch_err(c, "error recalculating oldest_gen: %i", ret);
2014 for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
2015 BTREE_ITER_PREFETCH, k, ret) {
2016 ret = __bch2_trans_do(&trans, NULL, NULL,
2017 BTREE_INSERT_NOFAIL,
2018 bch2_alloc_write_oldest_gen(&trans, &iter));
2020 bch_err(c, "error writing oldest_gen: %i", ret);
2024 bch2_trans_iter_exit(&trans, &iter);
2026 c->gc_gens_btree = 0;
2027 c->gc_gens_pos = POS_MIN;
2031 bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
2032 trace_gc_gens_end(c);
2034 for_each_member_device(ca, c, i) {
2035 kvfree(ca->oldest_gen);
2036 ca->oldest_gen = NULL;
2039 bch2_trans_exit(&trans);
2040 up_read(&c->gc_lock);
2041 mutex_unlock(&c->gc_gens_lock);
2045 static int bch2_gc_thread(void *arg)
2047 struct bch_fs *c = arg;
2048 struct io_clock *clock = &c->io_clock[WRITE];
2049 unsigned long last = atomic64_read(&clock->now);
2050 unsigned last_kick = atomic_read(&c->kick_gc);
2057 set_current_state(TASK_INTERRUPTIBLE);
2059 if (kthread_should_stop()) {
2060 __set_current_state(TASK_RUNNING);
2064 if (atomic_read(&c->kick_gc) != last_kick)
2067 if (c->btree_gc_periodic) {
2068 unsigned long next = last + c->capacity / 16;
2070 if (atomic64_read(&clock->now) >= next)
2073 bch2_io_clock_schedule_timeout(clock, next);
2080 __set_current_state(TASK_RUNNING);
2082 last = atomic64_read(&clock->now);
2083 last_kick = atomic_read(&c->kick_gc);
2086 * Full gc is currently incompatible with btree key cache:
2089 ret = bch2_gc(c, false, false);
2091 ret = bch2_gc_gens(c);
2094 bch_err(c, "btree gc failed: %i", ret);
2096 debug_check_no_locks_held();
2102 void bch2_gc_thread_stop(struct bch_fs *c)
2104 struct task_struct *p;
2107 c->gc_thread = NULL;
2115 int bch2_gc_thread_start(struct bch_fs *c)
2117 struct task_struct *p;
2122 p = kthread_create(bch2_gc_thread, c, "bch-gc/%s", c->name);
2124 bch_err(c, "error creating gc thread: %li", PTR_ERR(p));