1 files changed, 538 insertions, 158 deletions

diff --git a/libbcachefs/recovery.c b/libbcachefs/recovery.c
index a5651a9c..70fd9a27 100644
--- a/libbcachefs/recovery.c
+++ b/libbcachefs/recovery.c
@@ -11,92 +11,337 @@
 #include "error.h"
 #include "fsck.h"
 #include "journal_io.h"
+#include "journal_reclaim.h"
 #include "journal_seq_blacklist.h"
 #include "quota.h"
 #include "recovery.h"
 #include "replicas.h"
 #include "super-io.h"
 
+#include <linux/sort.h>
 #include <linux/stat.h>
 
 #define QSTR(n) { { { .len = strlen(n) } }, .name = n }
 
-static struct bkey_i *btree_root_find(struct bch_fs *c,
-				      struct bch_sb_field_clean *clean,
-				      struct jset *j,
-				      enum btree_id id, unsigned *level)
+/* sort and dedup all keys in the journal: */
+
+static void journal_entries_free(struct list_head *list)
 {
-	struct bkey_i *k;
-	struct jset_entry *entry, *start, *end;
 
-	if (clean) {
-		start = clean->start;
-		end = vstruct_end(&clean->field);
-	} else {
-		start = j->start;
-		end = vstruct_last(j);
+	while (!list_empty(list)) {
+		struct journal_replay *i =
+			list_first_entry(list, struct journal_replay, list);
+		list_del(&i->list);
+		kvpfree(i, offsetof(struct journal_replay, j) +
+			vstruct_bytes(&i->j));
 	}
+}
 
-	for (entry = start; entry < end; entry = vstruct_next(entry))
-		if (entry->type == BCH_JSET_ENTRY_btree_root &&
-		    entry->btree_id == id)
-			goto found;
-
-	return NULL;
-found:
-	if (!entry->u64s)
-		return ERR_PTR(-EINVAL);
+static int journal_sort_key_cmp(const void *_l, const void *_r)
+{
+	const struct journal_key *l = _l;
+	const struct journal_key *r = _r;
 
-	k = entry->start;
-	*level = entry->level;
-	return k;
+	return cmp_int(l->btree_id, r->btree_id) ?:
+		bkey_cmp(l->pos, r->pos) ?:
+		cmp_int(l->journal_seq, r->journal_seq) ?:
+		cmp_int(l->journal_offset, r->journal_offset);
 }
 
-static int verify_superblock_clean(struct bch_fs *c,
-				   struct bch_sb_field_clean **cleanp,
-				   struct jset *j)
+static int journal_sort_seq_cmp(const void *_l, const void *_r)
 {
-	unsigned i;
-	struct bch_sb_field_clean *clean = *cleanp;
-	int ret = 0;
+	const struct journal_key *l = _l;
+	const struct journal_key *r = _r;
 
-	if (!clean || !j)
-		return 0;
+	return cmp_int(l->journal_seq, r->journal_seq) ?:
+		cmp_int(l->btree_id, r->btree_id) ?:
+		bkey_cmp(l->pos, r->pos);
+}
 
-	if (mustfix_fsck_err_on(j->seq != clean->journal_seq, c,
-			"superblock journal seq (%llu) doesn't match journal (%llu) after clean shutdown",
-			le64_to_cpu(clean->journal_seq),
-			le64_to_cpu(j->seq))) {
-		kfree(clean);
-		*cleanp = NULL;
-		return 0;
+static void journal_keys_sift(struct journal_keys *keys, struct journal_key *i)
+{
+	while (i + 1 < keys->d + keys->nr &&
+	       journal_sort_key_cmp(i, i + 1) > 0) {
+		swap(i[0], i[1]);
+		i++;
 	}
+}
 
-	mustfix_fsck_err_on(j->read_clock != clean->read_clock, c,
-			"superblock read clock doesn't match journal after clean shutdown");
-	mustfix_fsck_err_on(j->write_clock != clean->write_clock, c,
-			"superblock read clock doesn't match journal after clean shutdown");
+static void journal_keys_free(struct journal_keys *keys)
+{
+	struct journal_key *i;
+
+	for_each_journal_key(*keys, i)
+		if (i->allocated)
+			kfree(i->k);
+	kvfree(keys->d);
+	keys->d = NULL;
+	keys->nr = 0;
+}
 
-	for (i = 0; i < BTREE_ID_NR; i++) {
-		struct bkey_i *k1, *k2;
-		unsigned l1 = 0, l2 = 0;
+static struct journal_keys journal_keys_sort(struct list_head *journal_entries)
+{
+	struct journal_replay *p;
+	struct jset_entry *entry;
+	struct bkey_i *k, *_n;
+	struct journal_keys keys = { NULL }, keys_deduped = { NULL };
+	struct journal_key *i;
+	size_t nr_keys = 0;
+
+	list_for_each_entry(p, journal_entries, list)
+		for_each_jset_key(k, _n, entry, &p->j)
+			nr_keys++;
+
+	keys.journal_seq_base = keys_deduped.journal_seq_base =
+		le64_to_cpu(list_first_entry(journal_entries,
+					     struct journal_replay,
+					     list)->j.seq);
+
+	keys.d = kvmalloc(sizeof(keys.d[0]) * nr_keys, GFP_KERNEL);
+	if (!keys.d)
+		goto err;
 
-		k1 = btree_root_find(c, clean, NULL, i, &l1);
-		k2 = btree_root_find(c, NULL, j, i, &l2);
+	keys_deduped.d = kvmalloc(sizeof(keys.d[0]) * nr_keys * 2, GFP_KERNEL);
+	if (!keys_deduped.d)
+		goto err;
 
-		if (!k1 && !k2)
+	list_for_each_entry(p, journal_entries, list)
+		for_each_jset_key(k, _n, entry, &p->j)
+			keys.d[keys.nr++] = (struct journal_key) {
+				.btree_id	= entry->btree_id,
+				.pos		= bkey_start_pos(&k->k),
+				.k		= k,
+				.journal_seq	= le64_to_cpu(p->j.seq) -
+					keys.journal_seq_base,
+				.journal_offset	= k->_data - p->j._data,
+			};
+
+	sort(keys.d, nr_keys, sizeof(keys.d[0]), journal_sort_key_cmp, NULL);
+
+	i = keys.d;
+	while (i < keys.d + keys.nr) {
+		if (i + 1 < keys.d + keys.nr &&
+		    i[0].btree_id == i[1].btree_id &&
+		    !bkey_cmp(i[0].pos, i[1].pos)) {
+			if (bkey_cmp(i[0].k->k.p, i[1].k->k.p) <= 0) {
+				i++;
+			} else {
+				bch2_cut_front(i[1].k->k.p, i[0].k);
+				i[0].pos = i[1].k->k.p;
+				journal_keys_sift(&keys, i);
+			}
 			continue;
+		}
 
-		mustfix_fsck_err_on(!k1 || !k2 ||
-				    IS_ERR(k1) ||
-				    IS_ERR(k2) ||
-				    k1->k.u64s != k2->k.u64s ||
-				    memcmp(k1, k2, bkey_bytes(k1)) ||
-				    l1 != l2, c,
-			"superblock btree root doesn't match journal after clean shutdown");
+		if (i + 1 < keys.d + keys.nr &&
+		    i[0].btree_id == i[1].btree_id &&
+		    bkey_cmp(i[0].k->k.p, bkey_start_pos(&i[1].k->k)) > 0) {
+			if ((cmp_int(i[0].journal_seq, i[1].journal_seq) ?:
+			     cmp_int(i[0].journal_offset, i[1].journal_offset)) < 0) {
+				if (bkey_cmp(i[0].k->k.p, i[1].k->k.p) <= 0) {
+					bch2_cut_back(bkey_start_pos(&i[1].k->k), &i[0].k->k);
+				} else {
+					struct bkey_i *split =
+						kmalloc(bkey_bytes(i[0].k), GFP_KERNEL);
+
+					if (!split)
+						goto err;
+
+					bkey_copy(split, i[0].k);
+					bch2_cut_back(bkey_start_pos(&i[1].k->k), &split->k);
+					keys_deduped.d[keys_deduped.nr++] = (struct journal_key) {
+						.btree_id	= i[0].btree_id,
+						.allocated	= true,
+						.pos		= bkey_start_pos(&split->k),
+						.k		= split,
+						.journal_seq	= i[0].journal_seq,
+						.journal_offset	= i[0].journal_offset,
+					};
+
+					bch2_cut_front(i[1].k->k.p, i[0].k);
+					i[0].pos = i[1].k->k.p;
+					journal_keys_sift(&keys, i);
+					continue;
+				}
+			} else {
+				if (bkey_cmp(i[0].k->k.p, i[1].k->k.p) >= 0) {
+					i[1] = i[0];
+					i++;
+					continue;
+				} else {
+					bch2_cut_front(i[0].k->k.p, i[1].k);
+					i[1].pos = i[0].k->k.p;
+					journal_keys_sift(&keys, i + 1);
+					continue;
+				}
+			}
+		}
+
+		keys_deduped.d[keys_deduped.nr++] = *i++;
 	}
-fsck_err:
-	return ret;
+
+	kvfree(keys.d);
+	return keys_deduped;
+err:
+	journal_keys_free(&keys_deduped);
+	kvfree(keys.d);
+	return (struct journal_keys) { NULL };
+}
+
+/* journal replay: */
+
+static void replay_now_at(struct journal *j, u64 seq)
+{
+	BUG_ON(seq < j->replay_journal_seq);
+	BUG_ON(seq > j->replay_journal_seq_end);
+
+	while (j->replay_journal_seq < seq)
+		bch2_journal_pin_put(j, j->replay_journal_seq++);
+}
+
+static int bch2_extent_replay_key(struct bch_fs *c, struct bkey_i *k)
+{
+	struct btree_trans trans;
+	struct btree_iter *iter, *split_iter;
+	/*
+	 * We might cause compressed extents to be split, so we need to pass in
+	 * a disk_reservation:
+	 */
+	struct disk_reservation disk_res =
+		bch2_disk_reservation_init(c, 0);
+	struct bkey_i *split;
+	bool split_compressed = false;
+	int ret;
+
+	bch2_trans_init(&trans, c);
+	bch2_trans_preload_iters(&trans);
+retry:
+	bch2_trans_begin(&trans);
+
+	iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS,
+				   bkey_start_pos(&k->k),
+				   BTREE_ITER_INTENT);
+
+	do {
+		ret = bch2_btree_iter_traverse(iter);
+		if (ret)
+			goto err;
+
+		split_iter = bch2_trans_copy_iter(&trans, iter);
+		ret = PTR_ERR_OR_ZERO(split_iter);
+		if (ret)
+			goto err;
+
+		split = bch2_trans_kmalloc(&trans, bkey_bytes(&k->k));
+		ret = PTR_ERR_OR_ZERO(split);
+		if (ret)
+			goto err;
+
+		if (!split_compressed &&
+		    bch2_extent_is_compressed(bkey_i_to_s_c(k)) &&
+		    !bch2_extent_is_atomic(k, split_iter)) {
+			ret = bch2_disk_reservation_add(c, &disk_res,
+					k->k.size *
+					bch2_bkey_nr_dirty_ptrs(bkey_i_to_s_c(k)),
+					BCH_DISK_RESERVATION_NOFAIL);
+			BUG_ON(ret);
+
+			split_compressed = true;
+		}
+
+		bkey_copy(split, k);
+		bch2_cut_front(split_iter->pos, split);
+		bch2_extent_trim_atomic(split, split_iter);
+
+		bch2_trans_update(&trans, BTREE_INSERT_ENTRY(split_iter, split));
+		bch2_btree_iter_set_pos(iter, split->k.p);
+	} while (bkey_cmp(iter->pos, k->k.p) < 0);
+
+	if (split_compressed) {
+		memset(&trans.fs_usage_deltas.fs_usage, 0,
+		       sizeof(trans.fs_usage_deltas.fs_usage));
+		trans.fs_usage_deltas.top = trans.fs_usage_deltas.d;
+
+		ret = bch2_trans_mark_key(&trans, bkey_i_to_s_c(k), false,
+					  -((s64) k->k.size),
+					  &trans.fs_usage_deltas) ?:
+		      bch2_trans_commit(&trans, &disk_res, NULL,
+					BTREE_INSERT_ATOMIC|
+					BTREE_INSERT_NOFAIL|
+					BTREE_INSERT_LAZY_RW|
+					BTREE_INSERT_NOMARK_OVERWRITES|
+					BTREE_INSERT_NO_CLEAR_REPLICAS);
+	} else {
+		ret = bch2_trans_commit(&trans, &disk_res, NULL,
+					BTREE_INSERT_ATOMIC|
+					BTREE_INSERT_NOFAIL|
+					BTREE_INSERT_LAZY_RW|
+					BTREE_INSERT_JOURNAL_REPLAY|
+					BTREE_INSERT_NOMARK);
+	}
+
+	if (ret)
+		goto err;
+err:
+	if (ret == -EINTR)
+		goto retry;
+
+	bch2_disk_reservation_put(c, &disk_res);
+
+	return bch2_trans_exit(&trans) ?: ret;
+}
+
+static int bch2_journal_replay(struct bch_fs *c,
+			       struct journal_keys keys)
+{
+	struct journal *j = &c->journal;
+	struct journal_key *i;
+	int ret;
+
+	sort(keys.d, keys.nr, sizeof(keys.d[0]), journal_sort_seq_cmp, NULL);
+
+	for_each_journal_key(keys, i) {
+		replay_now_at(j, keys.journal_seq_base + i->journal_seq);
+
+		switch (i->btree_id) {
+		case BTREE_ID_ALLOC:
+			ret = bch2_alloc_replay_key(c, i->k);
+			break;
+		case BTREE_ID_EXTENTS:
+			ret = bch2_extent_replay_key(c, i->k);
+			break;
+		default:
+			ret = bch2_btree_insert(c, i->btree_id, i->k,
+						NULL, NULL,
+						BTREE_INSERT_NOFAIL|
+						BTREE_INSERT_LAZY_RW|
+						BTREE_INSERT_JOURNAL_REPLAY|
+						BTREE_INSERT_NOMARK);
+			break;
+		}
+
+		if (ret) {
+			bch_err(c, "journal replay: error %d while replaying key",
+				ret);
+			return ret;
+		}
+
+		cond_resched();
+	}
+
+	replay_now_at(j, j->replay_journal_seq_end);
+	j->replay_journal_seq = 0;
+
+	bch2_journal_set_replay_done(j);
+	bch2_journal_flush_all_pins(j);
+	return bch2_journal_error(j);
+}
+
+static bool journal_empty(struct list_head *journal)
+{
+	return list_empty(journal) ||
+		journal_entry_empty(&list_last_entry(journal,
+					struct journal_replay, list)->j);
 }
 
 static int
@@ -129,40 +374,7 @@ fsck_err:
 	return ret;
 }
 
-static struct bch_sb_field_clean *read_superblock_clean(struct bch_fs *c)
-{
-	struct bch_sb_field_clean *clean, *sb_clean;
-	int ret;
-
-	mutex_lock(&c->sb_lock);
-	sb_clean = bch2_sb_get_clean(c->disk_sb.sb);
-
-	if (fsck_err_on(!sb_clean, c,
-			"superblock marked clean but clean section not present")) {
-		SET_BCH_SB_CLEAN(c->disk_sb.sb, false);
-		c->sb.clean = false;
-		mutex_unlock(&c->sb_lock);
-		return NULL;
-	}
-
-	clean = kmemdup(sb_clean, vstruct_bytes(&sb_clean->field),
-			GFP_KERNEL);
-	if (!clean) {
-		mutex_unlock(&c->sb_lock);
-		return ERR_PTR(-ENOMEM);
-	}
-
-	if (le16_to_cpu(c->disk_sb.sb->version) <
-	    bcachefs_metadata_version_bkey_renumber)
-		bch2_sb_clean_renumber(clean, READ);
-
-	mutex_unlock(&c->sb_lock);
-
-	return clean;
-fsck_err:
-	mutex_unlock(&c->sb_lock);
-	return ERR_PTR(ret);
-}
+/* journal replay early: */
 
 static int journal_replay_entry_early(struct bch_fs *c,
 				      struct jset_entry *entry)
@@ -190,13 +402,11 @@ static int journal_replay_entry_early(struct bch_fs *c,
 		switch (entry->btree_id) {
 		case FS_USAGE_RESERVED:
 			if (entry->level < BCH_REPLICAS_MAX)
-				percpu_u64_set(&c->usage[0]->
-					       persistent_reserved[entry->level],
-					       le64_to_cpu(u->v));
+				c->usage_base->persistent_reserved[entry->level] =
+					le64_to_cpu(u->v);
 			break;
 		case FS_USAGE_INODES:
-			percpu_u64_set(&c->usage[0]->nr_inodes,
-				       le64_to_cpu(u->v));
+			c->usage_base->nr_inodes = le64_to_cpu(u->v);
 			break;
 		case FS_USAGE_KEY_VERSION:
 			atomic64_set(&c->key_version,
@@ -274,6 +484,121 @@ static int journal_replay_early(struct bch_fs *c,
 	return 0;
 }
 
+/* sb clean section: */
+
+static struct bkey_i *btree_root_find(struct bch_fs *c,
+				      struct bch_sb_field_clean *clean,
+				      struct jset *j,
+				      enum btree_id id, unsigned *level)
+{
+	struct bkey_i *k;
+	struct jset_entry *entry, *start, *end;
+
+	if (clean) {
+		start = clean->start;
+		end = vstruct_end(&clean->field);
+	} else {
+		start = j->start;
+		end = vstruct_last(j);
+	}
+
+	for (entry = start; entry < end; entry = vstruct_next(entry))
+		if (entry->type == BCH_JSET_ENTRY_btree_root &&
+		    entry->btree_id == id)
+			goto found;
+
+	return NULL;
+found:
+	if (!entry->u64s)
+		return ERR_PTR(-EINVAL);
+
+	k = entry->start;
+	*level = entry->level;
+	return k;
+}
+
+static int verify_superblock_clean(struct bch_fs *c,
+				   struct bch_sb_field_clean **cleanp,
+				   struct jset *j)
+{
+	unsigned i;
+	struct bch_sb_field_clean *clean = *cleanp;
+	int ret = 0;
+
+	if (!c->sb.clean || !j)
+		return 0;
+
+	if (mustfix_fsck_err_on(j->seq != clean->journal_seq, c,
+			"superblock journal seq (%llu) doesn't match journal (%llu) after clean shutdown",
+			le64_to_cpu(clean->journal_seq),
+			le64_to_cpu(j->seq))) {
+		kfree(clean);
+		*cleanp = NULL;
+		return 0;
+	}
+
+	mustfix_fsck_err_on(j->read_clock != clean->read_clock, c,
+			"superblock read clock doesn't match journal after clean shutdown");
+	mustfix_fsck_err_on(j->write_clock != clean->write_clock, c,
+			"superblock read clock doesn't match journal after clean shutdown");
+
+	for (i = 0; i < BTREE_ID_NR; i++) {
+		struct bkey_i *k1, *k2;
+		unsigned l1 = 0, l2 = 0;
+
+		k1 = btree_root_find(c, clean, NULL, i, &l1);
+		k2 = btree_root_find(c, NULL, j, i, &l2);
+
+		if (!k1 && !k2)
+			continue;
+
+		mustfix_fsck_err_on(!k1 || !k2 ||
+				    IS_ERR(k1) ||
+				    IS_ERR(k2) ||
+				    k1->k.u64s != k2->k.u64s ||
+				    memcmp(k1, k2, bkey_bytes(k1)) ||
+				    l1 != l2, c,
+			"superblock btree root doesn't match journal after clean shutdown");
+	}
+fsck_err:
+	return ret;
+}
+
+static struct bch_sb_field_clean *read_superblock_clean(struct bch_fs *c)
+{
+	struct bch_sb_field_clean *clean, *sb_clean;
+	int ret;
+
+	mutex_lock(&c->sb_lock);
+	sb_clean = bch2_sb_get_clean(c->disk_sb.sb);
+
+	if (fsck_err_on(!sb_clean, c,
+			"superblock marked clean but clean section not present")) {
+		SET_BCH_SB_CLEAN(c->disk_sb.sb, false);
+		c->sb.clean = false;
+		mutex_unlock(&c->sb_lock);
+		return NULL;
+	}
+
+	clean = kmemdup(sb_clean, vstruct_bytes(&sb_clean->field),
+			GFP_KERNEL);
+	if (!clean) {
+		mutex_unlock(&c->sb_lock);
+		return ERR_PTR(-ENOMEM);
+	}
+
+	if (le16_to_cpu(c->disk_sb.sb->version) <
+	    bcachefs_metadata_version_bkey_renumber)
+		bch2_sb_clean_renumber(clean, READ);
+
+	mutex_unlock(&c->sb_lock);
+
+	return clean;
+fsck_err:
+	mutex_unlock(&c->sb_lock);
+	return ERR_PTR(ret);
+}
+
 static int read_btree_roots(struct bch_fs *c)
 {
 	unsigned i;
@@ -319,42 +644,14 @@ fsck_err:
 	return ret;
 }
 
-static bool journal_empty(struct list_head *journal)
-{
-	struct journal_replay *i;
-	struct jset_entry *entry;
-
-	if (list_empty(journal))
-		return true;
-
-	i = list_last_entry(journal, struct journal_replay, list);
-
-	if (i->j.last_seq != i->j.seq)
-		return false;
-
-	list_for_each_entry(i, journal, list) {
-		vstruct_for_each(&i->j, entry) {
-			if (entry->type == BCH_JSET_ENTRY_btree_root ||
-			    entry->type == BCH_JSET_ENTRY_usage ||
-			    entry->type == BCH_JSET_ENTRY_data_usage)
-				continue;
-
-			if (entry->type == BCH_JSET_ENTRY_btree_keys &&
-			    !entry->u64s)
-				continue;
-			return false;
-		}
-	}
-
-	return true;
-}
-
 int bch2_fs_recovery(struct bch_fs *c)
 {
 	const char *err = "cannot allocate memory";
 	struct bch_sb_field_clean *clean = NULL;
 	u64 journal_seq;
-	LIST_HEAD(journal);
+	LIST_HEAD(journal_entries);
+	struct journal_keys journal_keys = { NULL };
+	bool wrote = false, write_sb = false;
 	int ret;
 
 	if (c->sb.clean)
@@ -375,20 +672,31 @@ int bch2_fs_recovery(struct bch_fs *c)
 	if (!c->sb.clean || c->opts.fsck) {
 		struct jset *j;
 
-		ret = bch2_journal_read(c, &journal);
+		ret = bch2_journal_read(c, &journal_entries);
 		if (ret)
 			goto err;
 
-		fsck_err_on(c->sb.clean && !journal_empty(&journal), c,
-			    "filesystem marked clean but journal not empty");
+		if (mustfix_fsck_err_on(c->sb.clean && !journal_empty(&journal_entries), c,
+				"filesystem marked clean but journal not empty")) {
+			c->sb.compat &= ~(1ULL << BCH_COMPAT_FEAT_ALLOC_INFO);
+			SET_BCH_SB_CLEAN(c->disk_sb.sb, false);
+			c->sb.clean = false;
+		}
 
-		if (!c->sb.clean && list_empty(&journal)){
+		if (!c->sb.clean && list_empty(&journal_entries)) {
 			bch_err(c, "no journal entries found");
 			ret = BCH_FSCK_REPAIR_IMPOSSIBLE;
 			goto err;
 		}
 
-		j = &list_last_entry(&journal, struct journal_replay, list)->j;
+		journal_keys = journal_keys_sort(&journal_entries);
+		if (!journal_keys.d) {
+			ret = -ENOMEM;
+			goto err;
+		}
+
+		j = &list_last_entry(&journal_entries,
+				     struct journal_replay, list)->j;
 
 		ret = verify_superblock_clean(c, &clean, j);
 		if (ret)
@@ -399,14 +707,14 @@ int bch2_fs_recovery(struct bch_fs *c)
 		journal_seq = le64_to_cpu(clean->journal_seq) + 1;
 	}
 
-	ret = journal_replay_early(c, clean, &journal);
+	ret = journal_replay_early(c, clean, &journal_entries);
 	if (ret)
 		goto err;
 
 	if (!c->sb.clean) {
 		ret = bch2_journal_seq_blacklist_add(c,
-				journal_seq,
-				journal_seq + 4);
+						     journal_seq,
+						     journal_seq + 4);
 		if (ret) {
 			bch_err(c, "error creating new journal seq blacklist entry");
 			goto err;
@@ -417,11 +725,13 @@ int bch2_fs_recovery(struct bch_fs *c)
 
 	ret = bch2_blacklist_table_initialize(c);
 
-	ret = verify_journal_entries_not_blacklisted_or_missing(c, &journal);
+	ret = verify_journal_entries_not_blacklisted_or_missing(c,
+						&journal_entries);
 	if (ret)
 		goto err;
 
-	ret = bch2_fs_journal_start(&c->journal, journal_seq, &journal);
+	ret = bch2_fs_journal_start(&c->journal, journal_seq,
+				    &journal_entries);
 	if (ret)
 		goto err;
 
@@ -429,25 +739,43 @@ int bch2_fs_recovery(struct bch_fs *c)
 	if (ret)
 		goto err;
 
+	bch_verbose(c, "starting alloc read");
 	err = "error reading allocation information";
-	ret = bch2_alloc_read(c, &journal);
+	ret = bch2_alloc_read(c, &journal_keys);
 	if (ret)
 		goto err;
+	bch_verbose(c, "alloc read done");
 
 	bch_verbose(c, "starting stripes_read");
-	ret = bch2_stripes_read(c, &journal);
+	err = "error reading stripes";
+	ret = bch2_stripes_read(c, &journal_keys);
 	if (ret)
 		goto err;
 	bch_verbose(c, "stripes_read done");
 
 	set_bit(BCH_FS_ALLOC_READ_DONE, &c->flags);
 
+	if ((c->sb.compat & (1ULL << BCH_COMPAT_FEAT_ALLOC_INFO)) &&
+	    !(c->sb.compat & (1ULL << BCH_COMPAT_FEAT_ALLOC_METADATA))) {
+		/*
+		 * interior btree node updates aren't consistent with the
+		 * journal; after an unclean shutdown we have to walk all
+		 * pointers to metadata:
+		 */
+		bch_info(c, "starting metadata mark and sweep");
+		err = "error in mark and sweep";
+		ret = bch2_gc(c, NULL, true, true);
+		if (ret)
+			goto err;
+		bch_verbose(c, "mark and sweep done");
+	}
+
 	if (c->opts.fsck ||
 	    !(c->sb.compat & (1ULL << BCH_COMPAT_FEAT_ALLOC_INFO)) ||
 	    test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags)) {
-		bch_verbose(c, "starting mark and sweep:");
-		err = "error in recovery";
-		ret = bch2_gc(c, &journal, true, false);
+		bch_info(c, "starting mark and sweep");
+		err = "error in mark and sweep";
+		ret = bch2_gc(c, &journal_keys, true, false);
 		if (ret)
 			goto err;
 		bch_verbose(c, "mark and sweep done");
@@ -463,26 +791,63 @@ int bch2_fs_recovery(struct bch_fs *c)
 	if (c->sb.encryption_type && !c->sb.clean)
 		atomic64_add(1 << 16, &c->key_version);
 
-	if (c->opts.noreplay)
+	if (c->opts.norecovery)
 		goto out;
 
-	bch_verbose(c, "starting journal replay:");
+	bch_verbose(c, "starting journal replay");
 	err = "journal replay failed";
-	ret = bch2_journal_replay(c, &journal);
+	ret = bch2_journal_replay(c, journal_keys);
 	if (ret)
 		goto err;
 	bch_verbose(c, "journal replay done");
 
-	if (c->opts.norecovery)
-		goto out;
+	if (!c->opts.nochanges) {
+		/*
+		 * note that even when filesystem was clean there might be work
+		 * to do here, if we ran gc (because of fsck) which recalculated
+		 * oldest_gen:
+		 */
+		bch_verbose(c, "writing allocation info");
+		err = "error writing out alloc info";
+		ret = bch2_stripes_write(c, BTREE_INSERT_LAZY_RW, &wrote) ?:
+			bch2_alloc_write(c, BTREE_INSERT_LAZY_RW, &wrote);
+		if (ret) {
+			bch_err(c, "error writing alloc info");
+			goto err;
+		}
+		bch_verbose(c, "alloc write done");
+	}
 
-	err = "error in fsck";
-	ret = bch2_fsck(c);
-	if (ret)
-		goto err;
+	if (!c->sb.clean) {
+		if (!(c->sb.features & (1 << BCH_FEATURE_ATOMIC_NLINK))) {
+			bch_info(c, "checking inode link counts");
+			err = "error in recovery";
+			ret = bch2_fsck_inode_nlink(c);
+			if (ret)
+				goto err;
+			bch_verbose(c, "check inodes done");
+
+		} else {
+			bch_verbose(c, "checking for deleted inodes");
+			err = "error in recovery";
+			ret = bch2_fsck_walk_inodes_only(c);
+			if (ret)
+				goto err;
+			bch_verbose(c, "check inodes done");
+		}
+	}
+
+	if (c->opts.fsck) {
+		bch_info(c, "starting fsck");
+		err = "error in fsck";
+		ret = bch2_fsck_full(c);
+		if (ret)
+			goto err;
+		bch_verbose(c, "fsck done");
+	}
 
 	if (enabled_qtypes(c)) {
-		bch_verbose(c, "reading quotas:");
+		bch_verbose(c, "reading quotas");
 		ret = bch2_fs_quota_read(c);
 		if (ret)
 			goto err;
@@ -495,26 +860,41 @@ int bch2_fs_recovery(struct bch_fs *c)
 			c->disk_sb.sb->version_min =
 				le16_to_cpu(bcachefs_metadata_version_min);
 		c->disk_sb.sb->version = le16_to_cpu(bcachefs_metadata_version_current);
+		write_sb = true;
+	}
+
+	if (!test_bit(BCH_FS_ERROR, &c->flags)) {
+		c->disk_sb.sb->compat[0] |= 1ULL << BCH_COMPAT_FEAT_ALLOC_INFO;
+		write_sb = true;
 	}
 
 	if (c->opts.fsck &&
 	    !test_bit(BCH_FS_ERROR, &c->flags)) {
 		c->disk_sb.sb->features[0] |= 1ULL << BCH_FEATURE_ATOMIC_NLINK;
 		SET_BCH_SB_HAS_ERRORS(c->disk_sb.sb, 0);
+		write_sb = true;
 	}
+
+	if (write_sb)
+		bch2_write_super(c);
 	mutex_unlock(&c->sb_lock);
 
 	if (c->journal_seq_blacklist_table &&
 	    c->journal_seq_blacklist_table->nr > 128)
 		queue_work(system_long_wq, &c->journal_seq_blacklist_gc_work);
 out:
-	bch2_journal_entries_free(&journal);
-	kfree(clean);
-	return ret;
+	ret = 0;
 err:
 fsck_err:
-	pr_err("Error in recovery: %s (%i)", err, ret);
-	goto out;
+	bch2_flush_fsck_errs(c);
+	journal_keys_free(&journal_keys);
+	journal_entries_free(&journal_entries);
+	kfree(clean);
+	if (ret)
+		bch_err(c, "Error in recovery: %s (%i)", err, ret);
+	else
+		bch_verbose(c, "ret %i", ret);
+	return ret;
 }
 
 int bch2_fs_initialize(struct bch_fs *c)