bcachefs: Fsck for need_discard & freespace btrees

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

fs/bcachefs/alloc_background.c

@@ -580,6 +580,332 @@ int bch2_trans_mark_alloc(struct btree_trans *trans,
 	return 0;
 }
 
+static int bch2_check_alloc_key(struct btree_trans *trans,
+				struct btree_iter *alloc_iter)
+{
+	struct bch_fs *c = trans->c;
+	struct btree_iter discard_iter, freespace_iter;
+	struct bch_alloc_v4 a;
+	unsigned discard_key_type, freespace_key_type;
+	struct bkey_s_c alloc_k, k;
+	struct printbuf buf = PRINTBUF;
+	struct printbuf buf2 = PRINTBUF;
+	int ret;
+
+	alloc_k = bch2_btree_iter_peek(alloc_iter);
+	if (!alloc_k.k)
+		return 0;
+
+	ret = bkey_err(alloc_k);
+	if (ret)
+		return ret;
+
+	bch2_alloc_to_v4(alloc_k, &a);
+	discard_key_type = bucket_state(a) == BUCKET_need_discard
+		? KEY_TYPE_set : 0;
+	freespace_key_type = bucket_state(a) == BUCKET_free
+		? KEY_TYPE_set : 0;
+
+	bch2_trans_iter_init(trans, &discard_iter, BTREE_ID_need_discard,
+			     alloc_k.k->p, 0);
+	bch2_trans_iter_init(trans, &freespace_iter, BTREE_ID_freespace,
+			     alloc_freespace_pos(alloc_k.k->p, a), 0);
+
+	k = bch2_btree_iter_peek_slot(&discard_iter);
+	ret = bkey_err(k);
+	if (ret)
+		goto err;
+
+	if (fsck_err_on(k.k->type != discard_key_type, c,
+			"incorrect key in need_discard btree (got %s should be %s)\n"
+			"  %s",
+			bch2_bkey_types[k.k->type],
+			bch2_bkey_types[discard_key_type],
+			(bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
+		struct bkey_i *update =
+			bch2_trans_kmalloc(trans, sizeof(*update));
+
+		ret = PTR_ERR_OR_ZERO(update);
+		if (ret)
+			goto err;
+
+		bkey_init(&update->k);
+		update->k.type	= discard_key_type;
+		update->k.p	= discard_iter.pos;
+
+		ret = bch2_trans_update(trans, &discard_iter, update, 0);
+		if (ret)
+			goto err;
+	}
+
+	k = bch2_btree_iter_peek_slot(&freespace_iter);
+	ret = bkey_err(k);
+	if (ret)
+		goto err;
+
+	if (fsck_err_on(k.k->type != freespace_key_type, c,
+			"incorrect key in freespace btree (got %s should be %s)\n"
+			"  %s",
+			bch2_bkey_types[k.k->type],
+			bch2_bkey_types[freespace_key_type],
+			(printbuf_reset(&buf),
+			 bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
+		struct bkey_i *update =
+			bch2_trans_kmalloc(trans, sizeof(*update));
+
+		ret = PTR_ERR_OR_ZERO(update);
+		if (ret)
+			goto err;
+
+		bkey_init(&update->k);
+		update->k.type	= freespace_key_type;
+		update->k.p	= freespace_iter.pos;
+		bch2_key_resize(&update->k, 1);
+
+		ret = bch2_trans_update(trans, &freespace_iter, update, 0);
+		if (ret)
+			goto err;
+	}
+err:
+fsck_err:
+	bch2_trans_iter_exit(trans, &freespace_iter);
+	bch2_trans_iter_exit(trans, &discard_iter);
+	printbuf_exit(&buf2);
+	printbuf_exit(&buf);
+	return ret;
+}
+
+static inline bool bch2_dev_bucket_exists(struct bch_fs *c, struct bpos pos)
+{
+	struct bch_dev *ca;
+
+	if (pos.inode >= c->sb.nr_devices || !c->devs[pos.inode])
+		return false;
+
+	ca = bch_dev_bkey_exists(c, pos.inode);
+	return pos.offset >= ca->mi.first_bucket &&
+		pos.offset < ca->mi.nbuckets;
+}
+
+static int bch2_check_freespace_key(struct btree_trans *trans,
+				    struct btree_iter *freespace_iter,
+				    bool initial)
+{
+	struct bch_fs *c = trans->c;
+	struct btree_iter alloc_iter;
+	struct bkey_s_c k, freespace_k;
+	struct bch_alloc_v4 a;
+	u64 genbits;
+	struct bpos pos;
+	struct bkey_i *update;
+	struct printbuf buf = PRINTBUF;
+	int ret;
+
+	freespace_k = bch2_btree_iter_peek(freespace_iter);
+	if (!freespace_k.k)
+		return 1;
+
+	ret = bkey_err(freespace_k);
+	if (ret)
+		return ret;
+
+	pos = freespace_iter->pos;
+	pos.offset &= ~(~0ULL << 56);
+	genbits = freespace_iter->pos.offset & (~0ULL << 56);
+
+	bch2_trans_iter_init(trans, &alloc_iter, BTREE_ID_alloc, pos, 0);
+
+	if (fsck_err_on(!bch2_dev_bucket_exists(c, pos), c,
+			"%llu:%llu set in freespace btree but device or bucket does not exist",
+			pos.inode, pos.offset))
+		goto delete;
+
+	k = bch2_btree_iter_peek_slot(&alloc_iter);
+	ret = bkey_err(k);
+	if (ret)
+		goto err;
+
+	bch2_alloc_to_v4(k, &a);
+
+	if (fsck_err_on(bucket_state(a) != BUCKET_free ||
+			genbits != alloc_freespace_genbits(a), c,
+			"%s\n  incorrectly set in freespace index (free %u, genbits %llu should be %llu)",
+			(bch2_bkey_val_to_text(&buf, c, k), buf.buf),
+			bucket_state(a) == BUCKET_free,
+			genbits >> 56, alloc_freespace_genbits(a) >> 56))
+		goto delete;
+out:
+err:
+fsck_err:
+	bch2_trans_iter_exit(trans, &alloc_iter);
+	printbuf_exit(&buf);
+	return ret;
+delete:
+	update = bch2_trans_kmalloc(trans, sizeof(*update));
+	ret = PTR_ERR_OR_ZERO(update);
+	if (ret)
+		goto err;
+
+	bkey_init(&update->k);
+	update->k.p = freespace_iter->pos;
+	bch2_key_resize(&update->k, 1);
+
+	ret   = bch2_trans_update(trans, freespace_iter, update, 0) ?:
+		bch2_trans_commit(trans, NULL, NULL, 0);
+	goto out;
+}
+
+int bch2_check_alloc_info(struct bch_fs *c, bool initial)
+{
+	struct btree_trans trans;
+	struct btree_iter iter;
+	struct bkey_s_c k;
+	int ret = 0, last_dev = -1;
+
+	bch2_trans_init(&trans, c, 0, 0);
+
+	for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
+			   BTREE_ITER_PREFETCH, k, ret) {
+		if (k.k->p.inode != last_dev) {
+			struct bch_dev *ca = bch_dev_bkey_exists(c, k.k->p.inode);
+
+			if (!ca->mi.freespace_initialized) {
+				bch2_btree_iter_set_pos(&iter, POS(k.k->p.inode + 1, 0));
+				continue;
+			}
+
+			last_dev = k.k->p.inode;
+		}
+
+		ret = __bch2_trans_do(&trans, NULL, NULL, 0,
+			bch2_check_alloc_key(&trans, &iter));
+		if (ret)
+			break;
+	}
+	bch2_trans_iter_exit(&trans, &iter);
+
+	if (ret)
+		goto err;
+
+	bch2_trans_iter_init(&trans, &iter, BTREE_ID_freespace, POS_MIN,
+			     BTREE_ITER_PREFETCH);
+	while (1) {
+		ret = __bch2_trans_do(&trans, NULL, NULL, 0,
+			bch2_check_freespace_key(&trans, &iter, initial));
+		if (ret)
+			break;
+
+		bch2_btree_iter_set_pos(&iter, bpos_nosnap_successor(iter.pos));
+	}
+	bch2_trans_iter_exit(&trans, &iter);
+err:
+	bch2_trans_exit(&trans);
+	return ret < 0 ? ret : 0;
+}
+
+static int bch2_check_alloc_to_lru_ref(struct btree_trans *trans,
+				       struct btree_iter *alloc_iter)
+{
+	struct bch_fs *c = trans->c;
+	struct btree_iter lru_iter;
+	struct bch_alloc_v4 a;
+	struct bkey_s_c alloc_k, k;
+	struct printbuf buf = PRINTBUF;
+	struct printbuf buf2 = PRINTBUF;
+	int ret;
+
+	alloc_k = bch2_btree_iter_peek(alloc_iter);
+	if (!alloc_k.k)
+		return 0;
+
+	ret = bkey_err(alloc_k);
+	if (ret)
+		return ret;
+
+	bch2_alloc_to_v4(alloc_k, &a);
+
+	if (bucket_state(a) != BUCKET_cached)
+		return 0;
+
+	bch2_trans_iter_init(trans, &lru_iter, BTREE_ID_lru,
+			     POS(alloc_k.k->p.inode, a.io_time[READ]), 0);
+
+	k = bch2_btree_iter_peek_slot(&lru_iter);
+	ret = bkey_err(k);
+	if (ret)
+		goto err;
+
+	if (fsck_err_on(!a.io_time[READ], c,
+			"cached bucket with read_time 0\n"
+			"  %s",
+		(printbuf_reset(&buf),
+		 bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)) ||
+	    fsck_err_on(k.k->type != KEY_TYPE_lru ||
+			le64_to_cpu(bkey_s_c_to_lru(k).v->idx) != alloc_k.k->p.offset, c,
+			"incorrect/missing lru entry\n"
+			"  %s\n"
+			"  %s",
+			(printbuf_reset(&buf),
+			 bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf),
+			(bch2_bkey_val_to_text(&buf2, c, k), buf2.buf))) {
+		u64 read_time = a.io_time[READ];
+
+		if (!a.io_time[READ])
+			a.io_time[READ] = atomic64_read(&c->io_clock[READ].now);
+
+		ret = bch2_lru_change(trans,
+				      alloc_k.k->p.inode,
+				      alloc_k.k->p.offset,
+				      0, &a.io_time[READ]);
+		if (ret)
+			goto err;
+
+		if (a.io_time[READ] != read_time) {
+			struct bkey_i_alloc_v4 *a_mut =
+				bch2_alloc_to_v4_mut(trans, alloc_k);
+			ret = PTR_ERR_OR_ZERO(a_mut);
+			if (ret)
+				goto err;
+
+			a_mut->v.io_time[READ] = a.io_time[READ];
+			ret = bch2_trans_update(trans, alloc_iter,
+						&a_mut->k_i, BTREE_TRIGGER_NORUN);
+			if (ret)
+				goto err;
+		}
+	}
+err:
+fsck_err:
+	bch2_trans_iter_exit(trans, &lru_iter);
+	printbuf_exit(&buf2);
+	printbuf_exit(&buf);
+	return ret;
+}
+
+int bch2_check_alloc_to_lru_refs(struct bch_fs *c)
+{
+	struct btree_trans trans;
+	struct btree_iter iter;
+	struct bkey_s_c k;
+	int ret = 0;
+
+	bch2_trans_init(&trans, c, 0, 0);
+
+	for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
+			   BTREE_ITER_PREFETCH, k, ret) {
+		ret = __bch2_trans_do(&trans, NULL, NULL,
+				      BTREE_INSERT_NOFAIL|
+				      BTREE_INSERT_LAZY_RW,
+			bch2_check_alloc_to_lru_ref(&trans, &iter));
+		if (ret)
+			break;
+	}
+	bch2_trans_iter_exit(&trans, &iter);
+
+	bch2_trans_exit(&trans);
+	return ret < 0 ? ret : 0;
+}
+
 static int bch2_clear_need_discard(struct btree_trans *trans, struct bpos pos,
 				   struct bch_dev *ca, bool *discard_done)
 {

fs/bcachefs/alloc_background.h

@@ -113,6 +113,8 @@ int bch2_alloc_read(struct bch_fs *, bool, bool);
 
 int bch2_trans_mark_alloc(struct btree_trans *, struct bkey_s_c,
 			  struct bkey_i *, unsigned);
+int bch2_check_alloc_info(struct bch_fs *, bool);
+int bch2_check_alloc_to_lru_refs(struct bch_fs *);
 void bch2_do_discards(struct bch_fs *);
 
 static inline bool should_invalidate_buckets(struct bch_dev *ca)

fs/bcachefs/lru.c

 // SPDX-License-Identifier: GPL-2.0
 
 #include "bcachefs.h"
+#include "alloc_background.h"
 #include "btree_iter.h"
 #include "btree_update.h"
 #include "error.h"
 #include "lru.h"
+#include "recovery.h"
 
 const char *bch2_lru_invalid(const struct bch_fs *c, struct bkey_s_c k)
 {
@@ -117,3 +119,85 @@ int bch2_lru_change(struct btree_trans *trans, u64 id, u64 idx,
 	return  lru_delete(trans, id, idx, old_time) ?:
 		lru_set(trans, id, idx, new_time);
 }
+
+static int bch2_check_lru_key(struct btree_trans *trans,
+			      struct btree_iter *lru_iter, bool initial)
+{
+	struct bch_fs *c = trans->c;
+	struct btree_iter iter;
+	struct bkey_s_c lru_k, k;
+	struct bch_alloc_v4 a;
+	struct printbuf buf1 = PRINTBUF;
+	struct printbuf buf2 = PRINTBUF;
+	u64 idx;
+	int ret;
+
+	lru_k = bch2_btree_iter_peek(lru_iter);
+	if (!lru_k.k)
+		return 0;
+
+	ret = bkey_err(lru_k);
+	if (ret)
+		return ret;
+
+	idx = le64_to_cpu(bkey_s_c_to_lru(lru_k).v->idx);
+
+	bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
+			     POS(lru_k.k->p.inode, idx), 0);
+	k = bch2_btree_iter_peek_slot(&iter);
+	ret = bkey_err(k);
+	if (ret)
+		goto err;
+
+	bch2_alloc_to_v4(k, &a);
+
+	if (fsck_err_on(bucket_state(a) != BUCKET_cached ||
+			a.io_time[READ] != lru_k.k->p.offset, c,
+			"incorrect lru entry %s\n"
+			"  for %s",
+			(bch2_bkey_val_to_text(&buf1, c, lru_k), buf1.buf),
+			(bch2_bkey_val_to_text(&buf2, c, k), buf2.buf))) {
+		struct bkey_i *update =
+			bch2_trans_kmalloc(trans, sizeof(*update));
+
+		ret = PTR_ERR_OR_ZERO(update);
+		if (ret)
+			goto err;
+
+		bkey_init(&update->k);
+		update->k.p = lru_iter->pos;
+
+		ret = bch2_trans_update(trans, lru_iter, update, 0);
+		if (ret)
+			goto err;
+	}
+err:
+fsck_err:
+	bch2_trans_iter_exit(trans, &iter);
+	printbuf_exit(&buf2);
+	printbuf_exit(&buf1);
+	return ret;
+}
+
+int bch2_check_lrus(struct bch_fs *c, bool initial)
+{
+	struct btree_trans trans;
+	struct btree_iter iter;
+	struct bkey_s_c k;
+	int ret = 0;
+
+	bch2_trans_init(&trans, c, 0, 0);
+
+	for_each_btree_key(&trans, iter, BTREE_ID_lru, POS_MIN,
+			   BTREE_ITER_PREFETCH, k, ret) {
+		ret = __bch2_trans_do(&trans, NULL, NULL, 0,
+			bch2_check_lru_key(&trans, &iter, initial));
+		if (ret)
+			break;
+	}
+	bch2_trans_iter_exit(&trans, &iter);
+
+	bch2_trans_exit(&trans);
+	return ret;
+
+}

fs/bcachefs/lru.h

@@ -12,4 +12,6 @@ void bch2_lru_to_text(struct printbuf *, struct bch_fs *, struct bkey_s_c);
 
 int bch2_lru_change(struct btree_trans *, u64, u64, u64, u64 *);
 
+int bch2_check_lrus(struct bch_fs *, bool);
+
 #endif /* _BCACHEFS_LRU_H */

fs/bcachefs/recovery.c

@@ -16,6 +16,7 @@
 #include "journal_io.h"
 #include "journal_reclaim.h"
 #include "journal_seq_blacklist.h"
+#include "lru.h"
 #include "move.h"
 #include "quota.h"
 #include "recovery.h"
@@ -1166,13 +1167,26 @@ int bch2_fs_recovery(struct bch_fs *c)
 		bool metadata_only = c->opts.norecovery;
 
 		bch_info(c, "checking allocations");
-		err = "error in mark and sweep";
+		err = "error checking allocations";
 		ret = bch2_gc(c, true, metadata_only);
 		if (ret)
 			goto err;
 		bch_verbose(c, "done checking allocations");
 	}
 
+	if (c->opts.fsck) {
+		bch_info(c, "checking need_discard and freespace btrees");
+		err = "error checking need_discard and freespace btrees";
+		ret = bch2_check_alloc_info(c, true);
+		if (ret)
+			goto err;
+
+		ret = bch2_check_lrus(c, true);
+		if (ret)
+			goto err;
+		bch_verbose(c, "done checking need_discard and freespace btrees");
+	}
+
 	bch2_stripes_heap_start(c);
 
 	clear_bit(BCH_FS_REBUILD_REPLICAS, &c->flags);
@@ -1202,6 +1216,19 @@ int bch2_fs_recovery(struct bch_fs *c)
 	if (ret)
 		goto err;
 
+	if (c->opts.fsck) {
+		bch_info(c, "checking alloc to lru refs");
+		err = "error checking alloc to lru refs";
+		ret = bch2_check_alloc_to_lru_refs(c);
+		if (ret)
+			goto err;
+
+		ret = bch2_check_lrus(c, true);
+		if (ret)
+			goto err;
+		bch_verbose(c, "done checking alloc to lru refs");
+	}
+
 	if (c->sb.version < bcachefs_metadata_version_snapshot_2) {
 		bch2_fs_lazy_rw(c);