X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libbcachefs%2Frecovery.c;h=ec206139ecd1ed709c9893bf249b827ac85b35ce;hb=3070f45d968575efd42ecf44f94427203b2ab62e;hp=069aafba48acc8e32b554a6bd87f3f2f0f663a5e;hpb=a06dee6da2babd19a392ff30a399b89b85018a36;p=bcachefs-tools-debian

diff --git a/libbcachefs/recovery.c b/libbcachefs/recovery.c
index 069aafb..ec20613 100644
--- a/libbcachefs/recovery.c
+++ b/libbcachefs/recovery.c
@@ -5,6 +5,7 @@
 #include "bkey_buf.h"
 #include "alloc_background.h"
 #include "btree_gc.h"
+#include "btree_journal_iter.h"
 #include "btree_update.h"
 #include "btree_update_interior.h"
 #include "btree_io.h"
@@ -19,10 +20,14 @@
 #include "journal_reclaim.h"
 #include "journal_seq_blacklist.h"
 #include "lru.h"
+#include "logged_ops.h"
 #include "move.h"
 #include "quota.h"
+#include "rebalance.h"
 #include "recovery.h"
 #include "replicas.h"
+#include "sb-clean.h"
+#include "snapshot.h"
 #include "subvolume.h"
 #include "super-io.h"
 
@@ -31,13 +36,27 @@
 
 #define QSTR(n) { { { .len = strlen(n) } }, .name = n }
 
+static bool btree_id_is_alloc(enum btree_id id)
+{
+	switch (id) {
+	case BTREE_ID_alloc:
+	case BTREE_ID_backpointers:
+	case BTREE_ID_need_discard:
+	case BTREE_ID_freespace:
+	case BTREE_ID_bucket_gens:
+		return true;
+	default:
+		return false;
+	}
+}
+
 /* for -o reconstruct_alloc: */
 static void drop_alloc_keys(struct journal_keys *keys)
 {
 	size_t src, dst;
 
 	for (src = 0, dst = 0; src < keys->nr; src++)
-		if (keys->d[src].btree_id != BTREE_ID_alloc)
+		if (!btree_id_is_alloc(keys->d[src].btree_id))
 			keys->d[dst++] = keys->d[src];
 
 	keys->nr = dst;
@@ -57,488 +76,6 @@ static void zero_out_btree_mem_ptr(struct journal_keys *keys)
 			bkey_i_to_btree_ptr_v2(i->k)->v.mem_ptr = 0;
 }
 
-/* iterate over keys read from the journal: */
-
-static int __journal_key_cmp(enum btree_id	l_btree_id,
-			     unsigned		l_level,
-			     struct bpos	l_pos,
-			     const struct journal_key *r)
-{
-	return (cmp_int(l_btree_id,	r->btree_id) ?:
-		cmp_int(l_level,	r->level) ?:
-		bpos_cmp(l_pos,	r->k->k.p));
-}
-
-static int journal_key_cmp(const struct journal_key *l, const struct journal_key *r)
-{
-	return __journal_key_cmp(l->btree_id, l->level, l->k->k.p, r);
-}
-
-static inline size_t idx_to_pos(struct journal_keys *keys, size_t idx)
-{
-	size_t gap_size = keys->size - keys->nr;
-
-	if (idx >= keys->gap)
-		idx += gap_size;
-	return idx;
-}
-
-static inline struct journal_key *idx_to_key(struct journal_keys *keys, size_t idx)
-{
-	return keys->d + idx_to_pos(keys, idx);
-}
-
-static size_t __bch2_journal_key_search(struct journal_keys *keys,
-					enum btree_id id, unsigned level,
-					struct bpos pos)
-{
-	size_t l = 0, r = keys->nr, m;
-
-	while (l < r) {
-		m = l + ((r - l) >> 1);
-		if (__journal_key_cmp(id, level, pos, idx_to_key(keys, m)) > 0)
-			l = m + 1;
-		else
-			r = m;
-	}
-
-	BUG_ON(l < keys->nr &&
-	       __journal_key_cmp(id, level, pos, idx_to_key(keys, l)) > 0);
-
-	BUG_ON(l &&
-	       __journal_key_cmp(id, level, pos, idx_to_key(keys, l - 1)) <= 0);
-
-	return l;
-}
-
-static size_t bch2_journal_key_search(struct journal_keys *keys,
-				      enum btree_id id, unsigned level,
-				      struct bpos pos)
-{
-	return idx_to_pos(keys, __bch2_journal_key_search(keys, id, level, pos));
-}
-
-struct bkey_i *bch2_journal_keys_peek_upto(struct bch_fs *c, enum btree_id btree_id,
-					   unsigned level, struct bpos pos,
-					   struct bpos end_pos, size_t *idx)
-{
-	struct journal_keys *keys = &c->journal_keys;
-	unsigned iters = 0;
-	struct journal_key *k;
-search:
-	if (!*idx)
-		*idx = __bch2_journal_key_search(keys, btree_id, level, pos);
-
-	while ((k = *idx < keys->nr ? idx_to_key(keys, *idx) : NULL)) {
-		if (__journal_key_cmp(btree_id, level, end_pos, k) < 0)
-			return NULL;
-
-		if (__journal_key_cmp(btree_id, level, pos, k) <= 0 &&
-		    !k->overwritten)
-			return k->k;
-
-		(*idx)++;
-		iters++;
-		if (iters == 10) {
-			*idx = 0;
-			goto search;
-		}
-	}
-
-	return NULL;
-}
-
-struct bkey_i *bch2_journal_keys_peek_slot(struct bch_fs *c, enum btree_id btree_id,
-					   unsigned level, struct bpos pos)
-{
-	size_t idx = 0;
-
-	return bch2_journal_keys_peek_upto(c, btree_id, level, pos, pos, &idx);
-}
-
-static void journal_iters_fix(struct bch_fs *c)
-{
-	struct journal_keys *keys = &c->journal_keys;
-	/* The key we just inserted is immediately before the gap: */
-	size_t gap_end = keys->gap + (keys->size - keys->nr);
-	struct btree_and_journal_iter *iter;
-
-	/*
-	 * If an iterator points one after the key we just inserted, decrement
-	 * the iterator so it points at the key we just inserted - if the
-	 * decrement was unnecessary, bch2_btree_and_journal_iter_peek() will
-	 * handle that:
-	 */
-	list_for_each_entry(iter, &c->journal_iters, journal.list)
-		if (iter->journal.idx == gap_end)
-			iter->journal.idx = keys->gap - 1;
-}
-
-static void journal_iters_move_gap(struct bch_fs *c, size_t old_gap, size_t new_gap)
-{
-	struct journal_keys *keys = &c->journal_keys;
-	struct journal_iter *iter;
-	size_t gap_size = keys->size - keys->nr;
-
-	list_for_each_entry(iter, &c->journal_iters, list) {
-		if (iter->idx > old_gap)
-			iter->idx -= gap_size;
-		if (iter->idx >= new_gap)
-			iter->idx += gap_size;
-	}
-}
-
-int bch2_journal_key_insert_take(struct bch_fs *c, enum btree_id id,
-				 unsigned level, struct bkey_i *k)
-{
-	struct journal_key n = {
-		.btree_id	= id,
-		.level		= level,
-		.k		= k,
-		.allocated	= true,
-		/*
-		 * Ensure these keys are done last by journal replay, to unblock
-		 * journal reclaim:
-		 */
-		.journal_seq	= U32_MAX,
-	};
-	struct journal_keys *keys = &c->journal_keys;
-	size_t idx = bch2_journal_key_search(keys, id, level, k->k.p);
-
-	BUG_ON(test_bit(BCH_FS_RW, &c->flags));
-
-	if (idx < keys->size &&
-	    journal_key_cmp(&n, &keys->d[idx]) == 0) {
-		if (keys->d[idx].allocated)
-			kfree(keys->d[idx].k);
-		keys->d[idx] = n;
-		return 0;
-	}
-
-	if (idx > keys->gap)
-		idx -= keys->size - keys->nr;
-
-	if (keys->nr == keys->size) {
-		struct journal_keys new_keys = {
-			.nr			= keys->nr,
-			.size			= max_t(size_t, keys->size, 8) * 2,
-		};
-
-		new_keys.d = kvmalloc_array(new_keys.size, sizeof(new_keys.d[0]), GFP_KERNEL);
-		if (!new_keys.d) {
-			bch_err(c, "%s: error allocating new key array (size %zu)",
-				__func__, new_keys.size);
-			return -ENOMEM;
-		}
-
-		/* Since @keys was full, there was no gap: */
-		memcpy(new_keys.d, keys->d, sizeof(keys->d[0]) * keys->nr);
-		kvfree(keys->d);
-		*keys = new_keys;
-
-		/* And now the gap is at the end: */
-		keys->gap = keys->nr;
-	}
-
-	journal_iters_move_gap(c, keys->gap, idx);
-
-	move_gap(keys->d, keys->nr, keys->size, keys->gap, idx);
-	keys->gap = idx;
-
-	keys->nr++;
-	keys->d[keys->gap++] = n;
-
-	journal_iters_fix(c);
-
-	return 0;
-}
-
-/*
- * Can only be used from the recovery thread while we're still RO - can't be
- * used once we've got RW, as journal_keys is at that point used by multiple
- * threads:
- */
-int bch2_journal_key_insert(struct bch_fs *c, enum btree_id id,
-			    unsigned level, struct bkey_i *k)
-{
-	struct bkey_i *n;
-	int ret;
-
-	n = kmalloc(bkey_bytes(&k->k), GFP_KERNEL);
-	if (!n)
-		return -ENOMEM;
-
-	bkey_copy(n, k);
-	ret = bch2_journal_key_insert_take(c, id, level, n);
-	if (ret)
-		kfree(n);
-	return ret;
-}
-
-int bch2_journal_key_delete(struct bch_fs *c, enum btree_id id,
-			    unsigned level, struct bpos pos)
-{
-	struct bkey_i whiteout;
-
-	bkey_init(&whiteout.k);
-	whiteout.k.p = pos;
-
-	return bch2_journal_key_insert(c, id, level, &whiteout);
-}
-
-void bch2_journal_key_overwritten(struct bch_fs *c, enum btree_id btree,
-				  unsigned level, struct bpos pos)
-{
-	struct journal_keys *keys = &c->journal_keys;
-	size_t idx = bch2_journal_key_search(keys, btree, level, pos);
-
-	if (idx < keys->size &&
-	    keys->d[idx].btree_id	== btree &&
-	    keys->d[idx].level		== level &&
-	    bpos_eq(keys->d[idx].k->k.p, pos))
-		keys->d[idx].overwritten = true;
-}
-
-static void bch2_journal_iter_advance(struct journal_iter *iter)
-{
-	if (iter->idx < iter->keys->size) {
-		iter->idx++;
-		if (iter->idx == iter->keys->gap)
-			iter->idx += iter->keys->size - iter->keys->nr;
-	}
-}
-
-struct bkey_s_c bch2_journal_iter_peek(struct journal_iter *iter)
-{
-	struct journal_key *k = iter->keys->d + iter->idx;
-
-	while (k < iter->keys->d + iter->keys->size &&
-	       k->btree_id	== iter->btree_id &&
-	       k->level		== iter->level) {
-		if (!k->overwritten)
-			return bkey_i_to_s_c(k->k);
-
-		bch2_journal_iter_advance(iter);
-		k = iter->keys->d + iter->idx;
-	}
-
-	return bkey_s_c_null;
-}
-
-static void bch2_journal_iter_exit(struct journal_iter *iter)
-{
-	list_del(&iter->list);
-}
-
-static void bch2_journal_iter_init(struct bch_fs *c,
-				   struct journal_iter *iter,
-				   enum btree_id id, unsigned level,
-				   struct bpos pos)
-{
-	iter->btree_id	= id;
-	iter->level	= level;
-	iter->keys	= &c->journal_keys;
-	iter->idx	= bch2_journal_key_search(&c->journal_keys, id, level, pos);
-}
-
-static struct bkey_s_c bch2_journal_iter_peek_btree(struct btree_and_journal_iter *iter)
-{
-	return bch2_btree_node_iter_peek_unpack(&iter->node_iter,
-						iter->b, &iter->unpacked);
-}
-
-static void bch2_journal_iter_advance_btree(struct btree_and_journal_iter *iter)
-{
-	bch2_btree_node_iter_advance(&iter->node_iter, iter->b);
-}
-
-void bch2_btree_and_journal_iter_advance(struct btree_and_journal_iter *iter)
-{
-	if (bpos_eq(iter->pos, SPOS_MAX))
-		iter->at_end = true;
-	else
-		iter->pos = bpos_successor(iter->pos);
-}
-
-struct bkey_s_c bch2_btree_and_journal_iter_peek(struct btree_and_journal_iter *iter)
-{
-	struct bkey_s_c btree_k, journal_k, ret;
-again:
-	if (iter->at_end)
-		return bkey_s_c_null;
-
-	while ((btree_k = bch2_journal_iter_peek_btree(iter)).k &&
-	       bpos_lt(btree_k.k->p, iter->pos))
-		bch2_journal_iter_advance_btree(iter);
-
-	while ((journal_k = bch2_journal_iter_peek(&iter->journal)).k &&
-	       bpos_lt(journal_k.k->p, iter->pos))
-		bch2_journal_iter_advance(&iter->journal);
-
-	ret = journal_k.k &&
-		(!btree_k.k || bpos_le(journal_k.k->p, btree_k.k->p))
-		? journal_k
-		: btree_k;
-
-	if (ret.k && iter->b && bpos_gt(ret.k->p, iter->b->data->max_key))
-		ret = bkey_s_c_null;
-
-	if (ret.k) {
-		iter->pos = ret.k->p;
-		if (bkey_deleted(ret.k)) {
-			bch2_btree_and_journal_iter_advance(iter);
-			goto again;
-		}
-	} else {
-		iter->pos = SPOS_MAX;
-		iter->at_end = true;
-	}
-
-	return ret;
-}
-
-void bch2_btree_and_journal_iter_exit(struct btree_and_journal_iter *iter)
-{
-	bch2_journal_iter_exit(&iter->journal);
-}
-
-void __bch2_btree_and_journal_iter_init_node_iter(struct btree_and_journal_iter *iter,
-						  struct bch_fs *c,
-						  struct btree *b,
-						  struct btree_node_iter node_iter,
-						  struct bpos pos)
-{
-	memset(iter, 0, sizeof(*iter));
-
-	iter->b = b;
-	iter->node_iter = node_iter;
-	bch2_journal_iter_init(c, &iter->journal, b->c.btree_id, b->c.level, pos);
-	INIT_LIST_HEAD(&iter->journal.list);
-	iter->pos = b->data->min_key;
-	iter->at_end = false;
-}
-
-/*
- * this version is used by btree_gc before filesystem has gone RW and
- * multithreaded, so uses the journal_iters list:
- */
-void bch2_btree_and_journal_iter_init_node_iter(struct btree_and_journal_iter *iter,
-						struct bch_fs *c,
-						struct btree *b)
-{
-	struct btree_node_iter node_iter;
-
-	bch2_btree_node_iter_init_from_start(&node_iter, b);
-	__bch2_btree_and_journal_iter_init_node_iter(iter, c, b, node_iter, b->data->min_key);
-	list_add(&iter->journal.list, &c->journal_iters);
-}
-
-/* sort and dedup all keys in the journal: */
-
-void bch2_journal_entries_free(struct bch_fs *c)
-{
-	struct journal_replay **i;
-	struct genradix_iter iter;
-
-	genradix_for_each(&c->journal_entries, iter, i)
-		if (*i)
-			kvpfree(*i, offsetof(struct journal_replay, j) +
-				vstruct_bytes(&(*i)->j));
-	genradix_free(&c->journal_entries);
-}
-
-/*
- * When keys compare equal, oldest compares first:
- */
-static int journal_sort_key_cmp(const void *_l, const void *_r)
-{
-	const struct journal_key *l = _l;
-	const struct journal_key *r = _r;
-
-	return  journal_key_cmp(l, r) ?:
-		cmp_int(l->journal_seq, r->journal_seq) ?:
-		cmp_int(l->journal_offset, r->journal_offset);
-}
-
-void bch2_journal_keys_free(struct journal_keys *keys)
-{
-	struct journal_key *i;
-
-	move_gap(keys->d, keys->nr, keys->size, keys->gap, keys->nr);
-	keys->gap = keys->nr;
-
-	for (i = keys->d; i < keys->d + keys->nr; i++)
-		if (i->allocated)
-			kfree(i->k);
-
-	kvfree(keys->d);
-	keys->d = NULL;
-	keys->nr = keys->gap = keys->size = 0;
-}
-
-static int journal_keys_sort(struct bch_fs *c)
-{
-	struct genradix_iter iter;
-	struct journal_replay *i, **_i;
-	struct jset_entry *entry;
-	struct bkey_i *k, *_n;
-	struct journal_keys *keys = &c->journal_keys;
-	struct journal_key *src, *dst;
-	size_t nr_keys = 0;
-
-	genradix_for_each(&c->journal_entries, iter, _i) {
-		i = *_i;
-
-		if (!i || i->ignore)
-			continue;
-
-		for_each_jset_key(k, _n, entry, &i->j)
-			nr_keys++;
-	}
-
-	if (!nr_keys)
-		return 0;
-
-	keys->size = roundup_pow_of_two(nr_keys);
-
-	keys->d = kvmalloc_array(keys->size, sizeof(keys->d[0]), GFP_KERNEL);
-	if (!keys->d)
-		return -ENOMEM;
-
-	genradix_for_each(&c->journal_entries, iter, _i) {
-		i = *_i;
-
-		if (!i || i->ignore)
-			continue;
-
-		for_each_jset_key(k, _n, entry, &i->j)
-			keys->d[keys->nr++] = (struct journal_key) {
-				.btree_id	= entry->btree_id,
-				.level		= entry->level,
-				.k		= k,
-				.journal_seq	= le64_to_cpu(i->j.seq),
-				.journal_offset	= k->_data - i->j._data,
-			};
-	}
-
-	sort(keys->d, keys->nr, sizeof(keys->d[0]), journal_sort_key_cmp, NULL);
-
-	src = dst = keys->d;
-	while (src < keys->d + keys->nr) {
-		while (src + 1 < keys->d + keys->nr &&
-		       src[0].btree_id	== src[1].btree_id &&
-		       src[0].level	== src[1].level &&
-		       bpos_eq(src[0].k->k.p, src[1].k->k.p))
-			src++;
-
-		*dst++ = *src++;
-	}
-
-	keys->nr = dst - keys->d;
-	keys->gap = keys->nr;
-	return 0;
-}
-
 /* journal replay: */
 
 static void replay_now_at(struct journal *j, u64 seq)
@@ -558,10 +95,26 @@ static int bch2_journal_replay_key(struct btree_trans *trans,
 	unsigned iter_flags =
 		BTREE_ITER_INTENT|
 		BTREE_ITER_NOT_EXTENTS;
+	unsigned update_flags = BTREE_TRIGGER_NORUN;
 	int ret;
 
+	if (k->overwritten)
+		return 0;
+
+	trans->journal_res.seq = k->journal_seq;
+
+	/*
+	 * BTREE_UPDATE_KEY_CACHE_RECLAIM disables key cache lookup/update to
+	 * keep the key cache coherent with the underlying btree. Nothing
+	 * besides the allocator is doing updates yet so we don't need key cache
+	 * coherency for non-alloc btrees, and key cache fills for snapshots
+	 * btrees use BTREE_ITER_FILTER_SNAPSHOTS, which isn't available until
+	 * the snapshots recovery pass runs.
+	 */
 	if (!k->level && k->btree_id == BTREE_ID_alloc)
 		iter_flags |= BTREE_ITER_CACHED;
+	else
+		update_flags |= BTREE_UPDATE_KEY_CACHE_RECLAIM;
 
 	bch2_trans_node_iter_init(trans, &iter, k->btree_id, k->k->k.p,
 				  BTREE_MAX_DEPTH, k->level,
@@ -574,7 +127,7 @@ static int bch2_journal_replay_key(struct btree_trans *trans,
 	if (k->overwritten)
 		goto out;
 
-	ret = bch2_trans_update(trans, &iter, k->k, BTREE_TRIGGER_NORUN);
+	ret = bch2_trans_update(trans, &iter, k->k, update_flags);
 out:
 	bch2_trans_iter_exit(trans, &iter);
 	return ret;
@@ -591,57 +144,98 @@ static int journal_sort_seq_cmp(const void *_l, const void *_r)
 static int bch2_journal_replay(struct bch_fs *c)
 {
 	struct journal_keys *keys = &c->journal_keys;
-	struct journal_key **keys_sorted, *k;
+	DARRAY(struct journal_key *) keys_sorted = { 0 };
 	struct journal *j = &c->journal;
-	size_t i;
-	int ret;
+	u64 start_seq	= c->journal_replay_seq_start;
+	u64 end_seq	= c->journal_replay_seq_start;
+	struct btree_trans *trans = bch2_trans_get(c);
+	int ret = 0;
 
-	move_gap(keys->d, keys->nr, keys->size, keys->gap, keys->nr);
-	keys->gap = keys->nr;
+	if (keys->nr) {
+		ret = bch2_journal_log_msg(c, "Starting journal replay (%zu keys in entries %llu-%llu)",
+					   keys->nr, start_seq, end_seq);
+		if (ret)
+			goto err;
+	}
+
+	BUG_ON(!atomic_read(&keys->ref));
 
-	keys_sorted = kvmalloc_array(sizeof(*keys_sorted), keys->nr, GFP_KERNEL);
-	if (!keys_sorted)
-		return -ENOMEM;
+	/*
+	 * First, attempt to replay keys in sorted order. This is more
+	 * efficient - better locality of btree access -  but some might fail if
+	 * that would cause a journal deadlock.
+	 */
+	for (size_t i = 0; i < keys->nr; i++) {
+		cond_resched();
 
-	for (i = 0; i < keys->nr; i++)
-		keys_sorted[i] = &keys->d[i];
+		struct journal_key *k = keys->d + i;
 
-	sort(keys_sorted, keys->nr,
-	     sizeof(keys_sorted[0]),
-	     journal_sort_seq_cmp, NULL);
+		/* Skip fastpath if we're low on space in the journal */
+		ret = c->journal.watermark ? -1 :
+			commit_do(trans, NULL, NULL,
+				  BCH_TRANS_COMMIT_no_enospc|
+				  BCH_TRANS_COMMIT_journal_reclaim|
+				  (!k->allocated ? BCH_TRANS_COMMIT_no_journal_res : 0),
+			     bch2_journal_replay_key(trans, k));
+		BUG_ON(!ret && !k->overwritten);
+		if (ret) {
+			ret = darray_push(&keys_sorted, k);
+			if (ret)
+				goto err;
+		}
+	}
 
-	for (i = 0; i < keys->nr; i++) {
-		k = keys_sorted[i];
+	/*
+	 * Now, replay any remaining keys in the order in which they appear in
+	 * the journal, unpinning those journal entries as we go:
+	 */
+	sort(keys_sorted.data, keys_sorted.nr,
+	     sizeof(keys_sorted.data[0]),
+	     journal_sort_seq_cmp, NULL);
 
+	darray_for_each(keys_sorted, kp) {
 		cond_resched();
 
+		struct journal_key *k = *kp;
+
 		replay_now_at(j, k->journal_seq);
 
-		ret = bch2_trans_do(c, NULL, NULL,
-				    BTREE_INSERT_LAZY_RW|
-				    BTREE_INSERT_NOFAIL|
-				    (!k->allocated
-				     ? BTREE_INSERT_JOURNAL_REPLAY|JOURNAL_WATERMARK_reserved
-				     : 0),
-			     bch2_journal_replay_key(&trans, k));
-		if (ret) {
-			bch_err(c, "journal replay: error %d while replaying key at btree %s level %u",
-				ret, bch2_btree_ids[k->btree_id], k->level);
+		ret = commit_do(trans, NULL, NULL,
+				BCH_TRANS_COMMIT_no_enospc|
+				(!k->allocated
+				 ? BCH_TRANS_COMMIT_no_journal_res|BCH_WATERMARK_reclaim
+				 : 0),
+			     bch2_journal_replay_key(trans, k));
+		bch_err_msg(c, ret, "while replaying key at btree %s level %u:",
+			    bch2_btree_id_str(k->btree_id), k->level);
+		if (ret)
 			goto err;
-		}
+
+		BUG_ON(!k->overwritten);
 	}
 
+	/*
+	 * We need to put our btree_trans before calling flush_all_pins(), since
+	 * that will use a btree_trans internally
+	 */
+	bch2_trans_put(trans);
+	trans = NULL;
+
+	if (!c->opts.keep_journal)
+		bch2_journal_keys_put_initial(c);
+
 	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);
-	ret = bch2_journal_error(j);
 
-	if (keys->nr && !ret)
-		bch2_journal_log_msg(&c->journal, "journal replay finished");
+	if (keys->nr)
+		bch2_journal_log_msg(c, "journal replay finished");
 err:
-	kvfree(keys_sorted);
+	if (trans)
+		bch2_trans_put(trans);
+	darray_exit(&keys_sorted);
+	bch_err_fn(c, ret);
 	return ret;
 }
 
@@ -656,17 +250,17 @@ static int journal_replay_entry_early(struct bch_fs *c,
 	case BCH_JSET_ENTRY_btree_root: {
 		struct btree_root *r;
 
-		if (entry->btree_id >= BTREE_ID_NR) {
-			bch_err(c, "filesystem has unknown btree type %u",
-				entry->btree_id);
-			return -EINVAL;
+		while (entry->btree_id >= c->btree_roots_extra.nr + BTREE_ID_NR) {
+			ret = darray_push(&c->btree_roots_extra, (struct btree_root) { NULL });
+			if (ret)
+				return ret;
 		}
 
-		r = &c->btree_roots[entry->btree_id];
+		r = bch2_btree_id_root(c, entry->btree_id);
 
 		if (entry->u64s) {
 			r->level = entry->level;
-			bkey_copy(&r->key, &entry->start[0]);
+			bkey_copy(&r->key, (struct bkey_i *) entry->start);
 			r->error = 0;
 		} else {
 			r->error = -EIO;
@@ -709,8 +303,6 @@ static int journal_replay_entry_early(struct bch_fs *c,
 		struct bch_dev *ca = bch_dev_bkey_exists(c, le32_to_cpu(u->dev));
 		unsigned i, nr_types = jset_entry_dev_usage_nr_types(u);
 
-		ca->usage_base->buckets_ec		= le64_to_cpu(u->buckets_ec);
-
 		for (i = 0; i < min_t(unsigned, nr_types, BCH_DATA_NR); i++) {
 			ca->usage_base->d[i].buckets	= le64_to_cpu(u->d[i].buckets);
 			ca->usage_base->d[i].sectors	= le64_to_cpu(u->d[i].sectors);
@@ -751,191 +343,46 @@ static int journal_replay_entry_early(struct bch_fs *c,
 static int journal_replay_early(struct bch_fs *c,
 				struct bch_sb_field_clean *clean)
 {
-	struct jset_entry *entry;
-	int ret;
-
 	if (clean) {
-		for (entry = clean->start;
+		for (struct jset_entry *entry = clean->start;
 		     entry != vstruct_end(&clean->field);
 		     entry = vstruct_next(entry)) {
-			ret = journal_replay_entry_early(c, entry);
+			int ret = journal_replay_entry_early(c, entry);
 			if (ret)
 				return ret;
 		}
 	} else {
 		struct genradix_iter iter;
-		struct journal_replay *i, **_i;
-
-		genradix_for_each(&c->journal_entries, iter, _i) {
-			i = *_i;
-
-			if (!i || i->ignore)
-				continue;
-
-			vstruct_for_each(&i->j, entry) {
-				ret = journal_replay_entry_early(c, entry);
-				if (ret)
-					return ret;
-			}
-		}
-	}
-
-	bch2_fs_usage_initialize(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;
-	struct printbuf buf1 = PRINTBUF;
-	struct printbuf buf2 = PRINTBUF;
-	int ret = 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;
-	}
-
-	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;
-
-		printbuf_reset(&buf1);
-		printbuf_reset(&buf2);
-
-		if (k1)
-			bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(k1));
-		else
-			prt_printf(&buf1, "(none)");
-
-		if (k2)
-			bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(k2));
-		else
-			prt_printf(&buf2, "(none)");
-
-		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 %u doesn't match journal after clean shutdown\n"
-			"sb:      l=%u %s\n"
-			"journal: l=%u %s\n", i,
-			l1, buf1.buf,
-			l2, buf2.buf);
-	}
-fsck_err:
-	printbuf_exit(&buf2);
-	printbuf_exit(&buf1);
-	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);
+		struct journal_replay *i, **_i;
 
-	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;
-	}
+		genradix_for_each(&c->journal_entries, iter, _i) {
+			i = *_i;
 
-	clean = kmemdup(sb_clean, vstruct_bytes(&sb_clean->field),
-			GFP_KERNEL);
-	if (!clean) {
-		mutex_unlock(&c->sb_lock);
-		return ERR_PTR(-ENOMEM);
-	}
+			if (!i || i->ignore)
+				continue;
 
-	ret = bch2_sb_clean_validate_late(c, clean, READ);
-	if (ret) {
-		mutex_unlock(&c->sb_lock);
-		return ERR_PTR(ret);
+			vstruct_for_each(&i->j, entry) {
+				int ret = journal_replay_entry_early(c, entry);
+				if (ret)
+					return ret;
+			}
+		}
 	}
 
-	mutex_unlock(&c->sb_lock);
+	bch2_fs_usage_initialize(c);
 
-	return clean;
-fsck_err:
-	mutex_unlock(&c->sb_lock);
-	return ERR_PTR(ret);
+	return 0;
 }
 
-static bool btree_id_is_alloc(enum btree_id id)
-{
-	switch (id) {
-	case BTREE_ID_alloc:
-	case BTREE_ID_backpointers:
-	case BTREE_ID_need_discard:
-	case BTREE_ID_freespace:
-	case BTREE_ID_bucket_gens:
-		return true;
-	default:
-		return false;
-	}
-}
+/* sb clean section: */
 
 static int read_btree_roots(struct bch_fs *c)
 {
 	unsigned i;
 	int ret = 0;
 
-	for (i = 0; i < BTREE_ID_NR; i++) {
-		struct btree_root *r = &c->btree_roots[i];
+	for (i = 0; i < btree_id_nr_alive(c); i++) {
+		struct btree_root *r = bch2_btree_id_root(c, i);
 
 		if (!r->alive)
 			continue;
@@ -947,85 +394,90 @@ static int read_btree_roots(struct bch_fs *c)
 		}
 
 		if (r->error) {
-			__fsck_err(c, btree_id_is_alloc(i)
+			__fsck_err(c,
+				   btree_id_is_alloc(i)
 				   ? FSCK_CAN_IGNORE : 0,
+				   btree_root_bkey_invalid,
 				   "invalid btree root %s",
-				   bch2_btree_ids[i]);
+				   bch2_btree_id_str(i));
 			if (i == BTREE_ID_alloc)
 				c->sb.compat &= ~(1ULL << BCH_COMPAT_alloc_info);
 		}
 
 		ret = bch2_btree_root_read(c, i, &r->key, r->level);
 		if (ret) {
-			__fsck_err(c,
-				   btree_id_is_alloc(i)
-				   ? FSCK_CAN_IGNORE : 0,
-				   "error reading btree root %s",
-				   bch2_btree_ids[i]);
-			if (i == BTREE_ID_alloc)
+			fsck_err(c,
+				 btree_root_read_error,
+				 "error reading btree root %s",
+				 bch2_btree_id_str(i));
+			if (btree_id_is_alloc(i))
 				c->sb.compat &= ~(1ULL << BCH_COMPAT_alloc_info);
+			ret = 0;
 		}
 	}
 
-	for (i = 0; i < BTREE_ID_NR; i++)
-		if (!c->btree_roots[i].b)
+	for (i = 0; i < BTREE_ID_NR; i++) {
+		struct btree_root *r = bch2_btree_id_root(c, i);
+
+		if (!r->b) {
+			r->alive = false;
+			r->level = 0;
 			bch2_btree_root_alloc(c, i);
+		}
+	}
 fsck_err:
 	return ret;
 }
 
-static int bch2_fs_initialize_subvolumes(struct bch_fs *c)
+static int bch2_initialize_subvolumes(struct bch_fs *c)
 {
-	struct bkey_i_snapshot	root_snapshot;
-	struct bkey_i_subvolume root_volume;
+	struct bkey_i_snapshot_tree	root_tree;
+	struct bkey_i_snapshot		root_snapshot;
+	struct bkey_i_subvolume		root_volume;
 	int ret;
 
+	bkey_snapshot_tree_init(&root_tree.k_i);
+	root_tree.k.p.offset		= 1;
+	root_tree.v.master_subvol	= cpu_to_le32(1);
+	root_tree.v.root_snapshot	= cpu_to_le32(U32_MAX);
+
 	bkey_snapshot_init(&root_snapshot.k_i);
 	root_snapshot.k.p.offset = U32_MAX;
 	root_snapshot.v.flags	= 0;
 	root_snapshot.v.parent	= 0;
-	root_snapshot.v.subvol	= BCACHEFS_ROOT_SUBVOL;
-	root_snapshot.v.pad	= 0;
+	root_snapshot.v.subvol	= cpu_to_le32(BCACHEFS_ROOT_SUBVOL);
+	root_snapshot.v.tree	= cpu_to_le32(1);
 	SET_BCH_SNAPSHOT_SUBVOL(&root_snapshot.v, true);
 
-	ret = bch2_btree_insert(c, BTREE_ID_snapshots,
-				&root_snapshot.k_i,
-				NULL, NULL, 0);
-	if (ret)
-		return ret;
-
 	bkey_subvolume_init(&root_volume.k_i);
 	root_volume.k.p.offset = BCACHEFS_ROOT_SUBVOL;
 	root_volume.v.flags	= 0;
 	root_volume.v.snapshot	= cpu_to_le32(U32_MAX);
 	root_volume.v.inode	= cpu_to_le64(BCACHEFS_ROOT_INO);
 
-	ret = bch2_btree_insert(c, BTREE_ID_subvolumes,
-				&root_volume.k_i,
-				NULL, NULL, 0);
-	if (ret)
-		return ret;
-
-	return 0;
+	ret =   bch2_btree_insert(c, BTREE_ID_snapshot_trees,	&root_tree.k_i, NULL, 0) ?:
+		bch2_btree_insert(c, BTREE_ID_snapshots,	&root_snapshot.k_i, NULL, 0) ?:
+		bch2_btree_insert(c, BTREE_ID_subvolumes,	&root_volume.k_i, NULL, 0);
+	bch_err_fn(c, ret);
+	return ret;
 }
 
-static int bch2_fs_upgrade_for_subvolumes(struct btree_trans *trans)
+static int __bch2_fs_upgrade_for_subvolumes(struct btree_trans *trans)
 {
 	struct btree_iter iter;
 	struct bkey_s_c k;
 	struct bch_inode_unpacked inode;
 	int ret;
 
-	bch2_trans_iter_init(trans, &iter, BTREE_ID_inodes,
-			     SPOS(0, BCACHEFS_ROOT_INO, U32_MAX), 0);
-	k = bch2_btree_iter_peek_slot(&iter);
+	k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_inodes,
+			       SPOS(0, BCACHEFS_ROOT_INO, U32_MAX), 0);
 	ret = bkey_err(k);
 	if (ret)
-		goto err;
+		return ret;
 
 	if (!bkey_is_inode(k.k)) {
 		bch_err(trans->c, "root inode not found");
-		ret = -ENOENT;
+		ret = -BCH_ERR_ENOENT_inode;
 		goto err;
 	}
 
@@ -1040,26 +492,237 @@ err:
 	return ret;
 }
 
+/* set bi_subvol on root inode */
+noinline_for_stack
+static int bch2_fs_upgrade_for_subvolumes(struct bch_fs *c)
+{
+	int ret = bch2_trans_do(c, NULL, NULL, BCH_TRANS_COMMIT_lazy_rw,
+				__bch2_fs_upgrade_for_subvolumes(trans));
+	bch_err_fn(c, ret);
+	return ret;
+}
+
+const char * const bch2_recovery_passes[] = {
+#define x(_fn, _when)	#_fn,
+	BCH_RECOVERY_PASSES()
+#undef x
+	NULL
+};
+
+static int bch2_check_allocations(struct bch_fs *c)
+{
+	return bch2_gc(c, true, c->opts.norecovery);
+}
+
+static int bch2_set_may_go_rw(struct bch_fs *c)
+{
+	struct journal_keys *keys = &c->journal_keys;
+
+	/*
+	 * After we go RW, the journal keys buffer can't be modified (except for
+	 * setting journal_key->overwritten: it will be accessed by multiple
+	 * threads
+	 */
+	move_gap(keys->d, keys->nr, keys->size, keys->gap, keys->nr);
+	keys->gap = keys->nr;
+
+	set_bit(BCH_FS_may_go_rw, &c->flags);
+
+	if (keys->nr || c->opts.fsck || !c->sb.clean)
+		return bch2_fs_read_write_early(c);
+	return 0;
+}
+
+struct recovery_pass_fn {
+	int		(*fn)(struct bch_fs *);
+	unsigned	when;
+};
+
+static struct recovery_pass_fn recovery_pass_fns[] = {
+#define x(_fn, _when)	{ .fn = bch2_##_fn, .when = _when },
+	BCH_RECOVERY_PASSES()
+#undef x
+};
+
+static void check_version_upgrade(struct bch_fs *c)
+{
+	unsigned latest_compatible = bch2_latest_compatible_version(c->sb.version);
+	unsigned latest_version	= bcachefs_metadata_version_current;
+	unsigned old_version = c->sb.version_upgrade_complete ?: c->sb.version;
+	unsigned new_version = 0;
+	u64 recovery_passes;
+
+	if (old_version < bcachefs_metadata_required_upgrade_below) {
+		if (c->opts.version_upgrade == BCH_VERSION_UPGRADE_incompatible ||
+		    latest_compatible < bcachefs_metadata_required_upgrade_below)
+			new_version = latest_version;
+		else
+			new_version = latest_compatible;
+	} else {
+		switch (c->opts.version_upgrade) {
+		case BCH_VERSION_UPGRADE_compatible:
+			new_version = latest_compatible;
+			break;
+		case BCH_VERSION_UPGRADE_incompatible:
+			new_version = latest_version;
+			break;
+		case BCH_VERSION_UPGRADE_none:
+			new_version = old_version;
+			break;
+		}
+	}
+
+	if (new_version > old_version) {
+		struct printbuf buf = PRINTBUF;
+
+		if (old_version < bcachefs_metadata_required_upgrade_below)
+			prt_str(&buf, "Version upgrade required:\n");
+
+		if (old_version != c->sb.version) {
+			prt_str(&buf, "Version upgrade from ");
+			bch2_version_to_text(&buf, c->sb.version_upgrade_complete);
+			prt_str(&buf, " to ");
+			bch2_version_to_text(&buf, c->sb.version);
+			prt_str(&buf, " incomplete\n");
+		}
+
+		prt_printf(&buf, "Doing %s version upgrade from ",
+			   BCH_VERSION_MAJOR(old_version) != BCH_VERSION_MAJOR(new_version)
+			   ? "incompatible" : "compatible");
+		bch2_version_to_text(&buf, old_version);
+		prt_str(&buf, " to ");
+		bch2_version_to_text(&buf, new_version);
+		prt_newline(&buf);
+
+		recovery_passes = bch2_upgrade_recovery_passes(c, old_version, new_version);
+		if (recovery_passes) {
+			if ((recovery_passes & RECOVERY_PASS_ALL_FSCK) == RECOVERY_PASS_ALL_FSCK)
+				prt_str(&buf, "fsck required");
+			else {
+				prt_str(&buf, "running recovery passes: ");
+				prt_bitflags(&buf, bch2_recovery_passes, recovery_passes);
+			}
+
+			c->recovery_passes_explicit |= recovery_passes;
+			c->opts.fix_errors = FSCK_FIX_yes;
+		}
+
+		bch_info(c, "%s", buf.buf);
+
+		mutex_lock(&c->sb_lock);
+		bch2_sb_upgrade(c, new_version);
+		mutex_unlock(&c->sb_lock);
+
+		printbuf_exit(&buf);
+	}
+}
+
+u64 bch2_fsck_recovery_passes(void)
+{
+	u64 ret = 0;
+
+	for (unsigned i = 0; i < ARRAY_SIZE(recovery_pass_fns); i++)
+		if (recovery_pass_fns[i].when & PASS_FSCK)
+			ret |= BIT_ULL(i);
+	return ret;
+}
+
+static bool should_run_recovery_pass(struct bch_fs *c, enum bch_recovery_pass pass)
+{
+	struct recovery_pass_fn *p = recovery_pass_fns + pass;
+
+	if (c->opts.norecovery && pass > BCH_RECOVERY_PASS_snapshots_read)
+		return false;
+	if (c->recovery_passes_explicit & BIT_ULL(pass))
+		return true;
+	if ((p->when & PASS_FSCK) && c->opts.fsck)
+		return true;
+	if ((p->when & PASS_UNCLEAN) && !c->sb.clean)
+		return true;
+	if (p->when & PASS_ALWAYS)
+		return true;
+	return false;
+}
+
+static int bch2_run_recovery_pass(struct bch_fs *c, enum bch_recovery_pass pass)
+{
+	struct recovery_pass_fn *p = recovery_pass_fns + pass;
+	int ret;
+
+	if (!(p->when & PASS_SILENT))
+		bch2_print(c, KERN_INFO bch2_log_msg(c, "%s..."),
+			   bch2_recovery_passes[pass]);
+	ret = p->fn(c);
+	if (ret)
+		return ret;
+	if (!(p->when & PASS_SILENT))
+		bch2_print(c, KERN_CONT " done\n");
+
+	return 0;
+}
+
+static int bch2_run_recovery_passes(struct bch_fs *c)
+{
+	int ret = 0;
+
+	while (c->curr_recovery_pass < ARRAY_SIZE(recovery_pass_fns)) {
+		if (should_run_recovery_pass(c, c->curr_recovery_pass)) {
+			ret = bch2_run_recovery_pass(c, c->curr_recovery_pass);
+			if (bch2_err_matches(ret, BCH_ERR_restart_recovery))
+				continue;
+			if (ret)
+				break;
+
+			c->recovery_passes_complete |= BIT_ULL(c->curr_recovery_pass);
+		}
+		c->curr_recovery_pass++;
+		c->recovery_pass_done = max(c->recovery_pass_done, c->curr_recovery_pass);
+	}
+
+	return ret;
+}
+
+int bch2_run_online_recovery_passes(struct bch_fs *c)
+{
+	int ret = 0;
+
+	for (unsigned i = 0; i < ARRAY_SIZE(recovery_pass_fns); i++) {
+		struct recovery_pass_fn *p = recovery_pass_fns + i;
+
+		if (!(p->when & PASS_ONLINE))
+			continue;
+
+		ret = bch2_run_recovery_pass(c, i);
+		if (bch2_err_matches(ret, BCH_ERR_restart_recovery)) {
+			i = c->curr_recovery_pass;
+			continue;
+		}
+		if (ret)
+			break;
+	}
+
+	return ret;
+}
+
 int bch2_fs_recovery(struct bch_fs *c)
 {
-	const char *err = "cannot allocate memory";
 	struct bch_sb_field_clean *clean = NULL;
 	struct jset *last_journal_entry = NULL;
-	u64 blacklist_seq, journal_seq;
+	u64 last_seq = 0, blacklist_seq, journal_seq;
 	bool write_sb = false;
 	int ret = 0;
 
-	if (c->sb.clean)
-		clean = read_superblock_clean(c);
-	ret = PTR_ERR_OR_ZERO(clean);
-	if (ret)
-		goto err;
+	if (c->sb.clean) {
+		clean = bch2_read_superblock_clean(c);
+		ret = PTR_ERR_OR_ZERO(clean);
+		if (ret)
+			goto err;
 
-	if (c->sb.clean)
 		bch_info(c, "recovering from clean shutdown, journal seq %llu",
 			 le64_to_cpu(clean->journal_seq));
-	else
+	} else {
 		bch_info(c, "recovering from unclean shutdown");
+	}
 
 	if (!(c->sb.features & (1ULL << BCH_FEATURE_new_extent_overwrite))) {
 		bch_err(c, "feature new_extent_overwrite not set, filesystem no longer supported");
@@ -1074,29 +737,8 @@ int bch2_fs_recovery(struct bch_fs *c)
 		goto err;
 	}
 
-	if (!(c->sb.compat & (1ULL << BCH_COMPAT_bformat_overflow_done))) {
-		bch_err(c, "filesystem may have incompatible bkey formats; run fsck from the compat branch to fix");
-		ret = -EINVAL;
-		goto err;
-	}
-
-	if (!(c->sb.features & (1ULL << BCH_FEATURE_alloc_v2))) {
-		bch_info(c, "alloc_v2 feature bit not set, fsck required");
-		c->opts.fsck = true;
-		c->opts.fix_errors = FSCK_OPT_YES;
-	}
-
-	if (!c->opts.nochanges) {
-		if (c->sb.version < bcachefs_metadata_version_backpointers) {
-			bch_info(c, "version prior to backpointers, upgrade and fsck required");
-			c->opts.version_upgrade	= true;
-			c->opts.fsck		= true;
-			c->opts.fix_errors	= FSCK_OPT_YES;
-		} else if (c->sb.version < bcachefs_metadata_version_inode_v3) {
-			bch_info(c, "version prior to inode_v3, upgrade required");
-			c->opts.version_upgrade	= true;
-		}
-	}
+	if (!(c->opts.nochanges && c->opts.norecovery))
+		check_version_upgrade(c);
 
 	if (c->opts.fsck && c->opts.norecovery) {
 		bch_err(c, "cannot select both norecovery and fsck");
@@ -1115,7 +757,7 @@ int bch2_fs_recovery(struct bch_fs *c)
 		struct journal_replay **i;
 
 		bch_verbose(c, "starting journal read");
-		ret = bch2_journal_read(c, &blacklist_seq, &journal_seq);
+		ret = bch2_journal_read(c, &last_seq, &blacklist_seq, &journal_seq);
 		if (ret)
 			goto err;
 
@@ -1135,6 +777,7 @@ int bch2_fs_recovery(struct bch_fs *c)
 		if (mustfix_fsck_err_on(c->sb.clean &&
 					last_journal_entry &&
 					!journal_entry_empty(last_journal_entry), c,
+				clean_but_journal_not_empty,
 				"filesystem marked clean but journal not empty")) {
 			c->sb.compat &= ~(1ULL << BCH_COMPAT_alloc_info);
 			SET_BCH_SB_CLEAN(c->disk_sb.sb, false);
@@ -1142,16 +785,33 @@ int bch2_fs_recovery(struct bch_fs *c)
 		}
 
 		if (!last_journal_entry) {
-			fsck_err_on(!c->sb.clean, c, "no journal entries found");
-			goto use_clean;
+			fsck_err_on(!c->sb.clean, c,
+				    dirty_but_no_journal_entries,
+				    "no journal entries found");
+			if (clean)
+				goto use_clean;
+
+			genradix_for_each_reverse(&c->journal_entries, iter, i)
+				if (*i) {
+					last_journal_entry = &(*i)->j;
+					(*i)->ignore = false;
+					/*
+					 * This was probably a NO_FLUSH entry,
+					 * so last_seq was garbage - but we know
+					 * we're only using a single journal
+					 * entry, set it here:
+					 */
+					(*i)->j.last_seq = (*i)->j.seq;
+					break;
+				}
 		}
 
-		ret = journal_keys_sort(c);
+		ret = bch2_journal_keys_sort(c);
 		if (ret)
 			goto err;
 
 		if (c->sb.clean && last_journal_entry) {
-			ret = verify_superblock_clean(c, &clean,
+			ret = bch2_verify_superblock_clean(c, &clean,
 						      last_journal_entry);
 			if (ret)
 				goto err;
@@ -1167,6 +827,9 @@ use_clean:
 		blacklist_seq = journal_seq = le64_to_cpu(clean->journal_seq) + 1;
 	}
 
+	c->journal_replay_seq_start	= last_seq;
+	c->journal_replay_seq_end	= blacklist_seq - 1;
+
 	if (c->opts.reconstruct_alloc) {
 		c->sb.compat &= ~(1ULL << BCH_COMPAT_alloc_info);
 		drop_alloc_keys(&c->journal_keys);
@@ -1189,7 +852,9 @@ use_clean:
 		journal_seq += 8;
 
 	if (blacklist_seq != journal_seq) {
-		ret = bch2_journal_seq_blacklist_add(c,
+		ret =   bch2_journal_log_msg(c, "blacklisting entries %llu-%llu",
+					     blacklist_seq, journal_seq) ?:
+			bch2_journal_seq_blacklist_add(c,
 					blacklist_seq, journal_seq);
 		if (ret) {
 			bch_err(c, "error creating new journal seq blacklist entry");
@@ -1197,10 +862,15 @@ use_clean:
 		}
 	}
 
-	ret = bch2_fs_journal_start(&c->journal, journal_seq);
+	ret =   bch2_journal_log_msg(c, "starting journal at entry %llu, replaying %llu-%llu",
+				     journal_seq, last_seq, blacklist_seq - 1) ?:
+		bch2_fs_journal_start(&c->journal, journal_seq);
 	if (ret)
 		goto err;
 
+	if (c->opts.reconstruct_alloc)
+		bch2_journal_log_msg(c, "dropping alloc info");
+
 	/*
 	 * Skip past versions that might have possibly been used (as nonces),
 	 * but hadn't had their pointers written:
@@ -1212,181 +882,38 @@ use_clean:
 	if (ret)
 		goto err;
 
-	bch_verbose(c, "starting alloc read");
-	err = "error reading allocation information";
-
-	down_read(&c->gc_lock);
-	ret = c->sb.version < bcachefs_metadata_version_bucket_gens
-		? bch2_alloc_read(c)
-		: bch2_bucket_gens_read(c);
-	up_read(&c->gc_lock);
-
-	if (ret)
-		goto err;
-	bch_verbose(c, "alloc read done");
+	if (c->opts.fsck &&
+	    (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) ||
+	     BCH_SB_HAS_TOPOLOGY_ERRORS(c->disk_sb.sb)))
+		c->recovery_passes_explicit |= BIT_ULL(BCH_RECOVERY_PASS_check_topology);
 
-	bch_verbose(c, "starting stripes_read");
-	err = "error reading stripes";
-	ret = bch2_stripes_read(c);
+	ret = bch2_run_recovery_passes(c);
 	if (ret)
 		goto err;
-	bch_verbose(c, "stripes_read done");
-
-	bch2_stripes_heap_start(c);
-
-	if (c->opts.fsck) {
-		bool metadata_only = c->opts.norecovery;
-
-		bch_info(c, "checking allocations");
-		err = "error checking allocations";
-		ret = bch2_gc(c, true, metadata_only);
-		if (ret)
-			goto err;
-		bch_verbose(c, "done checking allocations");
-
-		set_bit(BCH_FS_INITIAL_GC_DONE, &c->flags);
-
-		bch_info(c, "checking need_discard and freespace btrees");
-		err = "error checking need_discard and freespace btrees";
-		ret = bch2_check_alloc_info(c);
-		if (ret)
-			goto err;
-		bch_verbose(c, "done checking need_discard and freespace btrees");
-
-		if (c->sb.version < bcachefs_metadata_version_snapshot_2) {
-			err = "error creating root snapshot node";
-			ret = bch2_fs_initialize_subvolumes(c);
-			if (ret)
-				goto err;
-		}
-
-		bch_verbose(c, "reading snapshots table");
-		err = "error reading snapshots table";
-		ret = bch2_fs_snapshots_start(c);
-		if (ret)
-			goto err;
-		bch_verbose(c, "reading snapshots done");
 
-		set_bit(BCH_FS_MAY_GO_RW, &c->flags);
+	/* If we fixed errors, verify that fs is actually clean now: */
+	if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) &&
+	    test_bit(BCH_FS_errors_fixed, &c->flags) &&
+	    !test_bit(BCH_FS_errors_not_fixed, &c->flags) &&
+	    !test_bit(BCH_FS_error, &c->flags)) {
+		bch2_flush_fsck_errs(c);
 
-		bch_info(c, "starting journal replay, %zu keys", c->journal_keys.nr);
-		err = "journal replay failed";
-		ret = bch2_journal_replay(c);
-		if (ret)
-			goto err;
-		if (c->opts.verbose || !c->sb.clean)
-			bch_info(c, "journal replay done");
-
-		bch_info(c, "checking lrus");
-		err = "error checking lrus";
-		ret = bch2_check_lrus(c);
-		if (ret)
-			goto err;
-		bch_verbose(c, "done checking lrus");
-		set_bit(BCH_FS_CHECK_LRUS_DONE, &c->flags);
-
-		bch_info(c, "checking backpointers to alloc keys");
-		err = "error checking backpointers to alloc keys";
-		ret = bch2_check_btree_backpointers(c);
-		if (ret)
-			goto err;
-		bch_verbose(c, "done checking backpointers to alloc keys");
+		bch_info(c, "Fixed errors, running fsck a second time to verify fs is clean");
+		clear_bit(BCH_FS_errors_fixed, &c->flags);
 
-		bch_info(c, "checking backpointers to extents");
-		err = "error checking backpointers to extents";
-		ret = bch2_check_backpointers_to_extents(c);
-		if (ret)
-			goto err;
-		bch_verbose(c, "done checking backpointers to extents");
-
-		bch_info(c, "checking extents to backpointers");
-		err = "error checking extents to backpointers";
-		ret = bch2_check_extents_to_backpointers(c);
-		if (ret)
-			goto err;
-		bch_verbose(c, "done checking extents to backpointers");
-		set_bit(BCH_FS_CHECK_BACKPOINTERS_DONE, &c->flags);
+		c->curr_recovery_pass = BCH_RECOVERY_PASS_check_alloc_info;
 
-		bch_info(c, "checking alloc to lru refs");
-		err = "error checking alloc to lru refs";
-		ret = bch2_check_alloc_to_lru_refs(c);
+		ret = bch2_run_recovery_passes(c);
 		if (ret)
 			goto err;
-		bch_verbose(c, "done checking alloc to lru refs");
-		set_bit(BCH_FS_CHECK_ALLOC_TO_LRU_REFS_DONE, &c->flags);
-	} else {
-		set_bit(BCH_FS_INITIAL_GC_DONE, &c->flags);
-		set_bit(BCH_FS_CHECK_LRUS_DONE, &c->flags);
-		set_bit(BCH_FS_CHECK_BACKPOINTERS_DONE, &c->flags);
-		set_bit(BCH_FS_CHECK_ALLOC_TO_LRU_REFS_DONE, &c->flags);
-		set_bit(BCH_FS_FSCK_DONE, &c->flags);
-
-		if (c->opts.norecovery)
-			goto out;
 
-		if (c->sb.version < bcachefs_metadata_version_snapshot_2) {
-			err = "error creating root snapshot node";
-			ret = bch2_fs_initialize_subvolumes(c);
-			if (ret)
-				goto err;
+		if (test_bit(BCH_FS_errors_fixed, &c->flags) ||
+		    test_bit(BCH_FS_errors_not_fixed, &c->flags)) {
+			bch_err(c, "Second fsck run was not clean");
+			set_bit(BCH_FS_errors_not_fixed, &c->flags);
 		}
 
-		bch_verbose(c, "reading snapshots table");
-		err = "error reading snapshots table";
-		ret = bch2_fs_snapshots_start(c);
-		if (ret)
-			goto err;
-		bch_verbose(c, "reading snapshots done");
-
-		set_bit(BCH_FS_MAY_GO_RW, &c->flags);
-
-		bch_verbose(c, "starting journal replay, %zu keys", c->journal_keys.nr);
-		err = "journal replay failed";
-		ret = bch2_journal_replay(c);
-		if (ret)
-			goto err;
-		if (c->opts.verbose || !c->sb.clean)
-			bch_info(c, "journal replay done");
-	}
-
-	err = "error initializing freespace";
-	ret = bch2_fs_freespace_init(c);
-	if (ret)
-		goto err;
-
-	if (c->sb.version < bcachefs_metadata_version_bucket_gens &&
-	    c->opts.version_upgrade) {
-		bch_info(c, "initializing bucket_gens");
-		err = "error initializing bucket gens";
-		ret = bch2_bucket_gens_init(c);
-		if (ret)
-			goto err;
-		bch_verbose(c, "bucket_gens init done");
-	}
-
-	if (c->sb.version < bcachefs_metadata_version_snapshot_2) {
-		/* set bi_subvol on root inode */
-		err = "error upgrade root inode for subvolumes";
-		ret = bch2_trans_do(c, NULL, NULL, BTREE_INSERT_LAZY_RW,
-				    bch2_fs_upgrade_for_subvolumes(&trans));
-		if (ret)
-			goto err;
-	}
-
-	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");
-	} else if (!c->sb.clean) {
-		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");
+		set_bit(BCH_FS_errors_fixed, &c->flags);
 	}
 
 	if (enabled_qtypes(c)) {
@@ -1398,20 +925,19 @@ use_clean:
 	}
 
 	mutex_lock(&c->sb_lock);
-	if (c->opts.version_upgrade) {
-		c->disk_sb.sb->version = cpu_to_le16(bcachefs_metadata_version_current);
-		c->disk_sb.sb->features[0] |= cpu_to_le64(BCH_SB_FEATURES_ALL);
+	if (BCH_SB_VERSION_UPGRADE_COMPLETE(c->disk_sb.sb) != c->sb.version) {
+		SET_BCH_SB_VERSION_UPGRADE_COMPLETE(c->disk_sb.sb, c->sb.version);
 		write_sb = true;
 	}
 
-	if (!test_bit(BCH_FS_ERROR, &c->flags)) {
+	if (!test_bit(BCH_FS_error, &c->flags)) {
 		c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_alloc_info);
 		write_sb = true;
 	}
 
 	if (c->opts.fsck &&
-	    !test_bit(BCH_FS_ERROR, &c->flags) &&
-	    !test_bit(BCH_FS_ERRORS_NOT_FIXED, &c->flags)) {
+	    !test_bit(BCH_FS_error, &c->flags) &&
+	    !test_bit(BCH_FS_errors_not_fixed, &c->flags)) {
 		SET_BCH_SB_HAS_ERRORS(c->disk_sb.sb, 0);
 		SET_BCH_SB_HAS_TOPOLOGY_ERRORS(c->disk_sb.sb, 0);
 		write_sb = true;
@@ -1422,18 +948,18 @@ use_clean:
 	mutex_unlock(&c->sb_lock);
 
 	if (!(c->sb.compat & (1ULL << BCH_COMPAT_extents_above_btree_updates_done)) ||
-	    !(c->sb.compat & (1ULL << BCH_COMPAT_bformat_overflow_done)) ||
-	    le16_to_cpu(c->sb.version_min) < bcachefs_metadata_version_btree_ptr_sectors_written) {
+	    c->sb.version_min < bcachefs_metadata_version_btree_ptr_sectors_written) {
 		struct bch_move_stats stats;
 
 		bch2_move_stats_init(&stats, "recovery");
 
-		bch_info(c, "scanning for old btree nodes");
-		ret = bch2_fs_read_write(c);
-		if (ret)
-			goto err;
+		struct printbuf buf = PRINTBUF;
+		bch2_version_to_text(&buf, c->sb.version_min);
+		bch_info(c, "scanning for old btree nodes: min_version %s", buf.buf);
+		printbuf_exit(&buf);
 
-		ret = bch2_scan_old_btree_nodes(c, &stats);
+		ret =   bch2_fs_read_write_early(c) ?:
+			bch2_scan_old_btree_nodes(c, &stats);
 		if (ret)
 			goto err;
 		bch_info(c, "scanning for old btree nodes done");
@@ -1445,25 +971,22 @@ use_clean:
 
 	ret = 0;
 out:
-	set_bit(BCH_FS_FSCK_DONE, &c->flags);
+	set_bit(BCH_FS_fsck_done, &c->flags);
 	bch2_flush_fsck_errs(c);
 
 	if (!c->opts.keep_journal &&
-	    test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)) {
-		bch2_journal_keys_free(&c->journal_keys);
-		bch2_journal_entries_free(c);
-	}
+	    test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))
+		bch2_journal_keys_put_initial(c);
 	kfree(clean);
 
-	if (!ret && test_bit(BCH_FS_HAVE_DELETED_SNAPSHOTS, &c->flags)) {
+	if (!ret &&
+	    test_bit(BCH_FS_need_delete_dead_snapshots, &c->flags) &&
+	    !c->opts.nochanges) {
 		bch2_fs_read_write_early(c);
 		bch2_delete_dead_snapshots_async(c);
 	}
 
-	if (ret)
-		bch_err(c, "Error in recovery: %s (%s)", err, bch2_err_str(ret));
-	else
-		bch_verbose(c, "ret %s", bch2_err_str(ret));
+	bch_err_fn(c, ret);
 	return ret;
 err:
 fsck_err:
@@ -1476,9 +999,6 @@ int bch2_fs_initialize(struct bch_fs *c)
 	struct bch_inode_unpacked root_inode, lostfound_inode;
 	struct bkey_inode_buf packed_inode;
 	struct qstr lostfound = QSTR("lost+found");
-	const char *err = "cannot allocate memory";
-	struct bch_dev *ca;
-	unsigned i;
 	int ret;
 
 	bch_notice(c, "initializing new filesystem");
@@ -1487,37 +1007,28 @@ int bch2_fs_initialize(struct bch_fs *c)
 	c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_extents_above_btree_updates_done);
 	c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_bformat_overflow_done);
 
-	if (c->sb.version < bcachefs_metadata_version_inode_v3)
-		c->opts.version_upgrade	= true;
+	bch2_sb_maybe_downgrade(c);
 
-	if (c->opts.version_upgrade) {
-		c->disk_sb.sb->version = cpu_to_le16(bcachefs_metadata_version_current);
-		c->disk_sb.sb->features[0] |= cpu_to_le64(BCH_SB_FEATURES_ALL);
+	if (c->opts.version_upgrade != BCH_VERSION_UPGRADE_none) {
+		bch2_sb_upgrade(c, bcachefs_metadata_version_current);
+		SET_BCH_SB_VERSION_UPGRADE_COMPLETE(c->disk_sb.sb, bcachefs_metadata_version_current);
 		bch2_write_super(c);
 	}
 	mutex_unlock(&c->sb_lock);
 
-	set_bit(BCH_FS_INITIAL_GC_DONE, &c->flags);
-	set_bit(BCH_FS_CHECK_LRUS_DONE, &c->flags);
-	set_bit(BCH_FS_CHECK_BACKPOINTERS_DONE, &c->flags);
-	set_bit(BCH_FS_CHECK_ALLOC_TO_LRU_REFS_DONE, &c->flags);
-	set_bit(BCH_FS_MAY_GO_RW, &c->flags);
-	set_bit(BCH_FS_FSCK_DONE, &c->flags);
+	c->curr_recovery_pass = ARRAY_SIZE(recovery_pass_fns);
+	set_bit(BCH_FS_may_go_rw, &c->flags);
+	set_bit(BCH_FS_fsck_done, &c->flags);
 
-	for (i = 0; i < BTREE_ID_NR; i++)
+	for (unsigned i = 0; i < BTREE_ID_NR; i++)
 		bch2_btree_root_alloc(c, i);
 
-	for_each_online_member(ca, c, i)
+	for_each_member_device(c, ca)
 		bch2_dev_usage_init(ca);
 
-	err = "unable to allocate journal buckets";
-	for_each_online_member(ca, c, i) {
-		ret = bch2_dev_journal_alloc(ca);
-		if (ret) {
-			percpu_ref_put(&ca->io_ref);
-			goto err;
-		}
-	}
+	ret = bch2_fs_journal_alloc(c);
+	if (ret)
+		goto err;
 
 	/*
 	 * journal_res_get() will crash if called before this has
@@ -1526,7 +1037,6 @@ int bch2_fs_initialize(struct bch_fs *c)
 	bch2_fs_journal_start(&c->journal, 1);
 	bch2_journal_set_replay_done(&c->journal);
 
-	err = "error going read-write";
 	ret = bch2_fs_read_write_early(c);
 	if (ret)
 		goto err;
@@ -1536,31 +1046,24 @@ int bch2_fs_initialize(struct bch_fs *c)
 	 * btree updates
 	 */
 	bch_verbose(c, "marking superblocks");
-	err = "error marking superblock and journal";
-	for_each_member_device(ca, c, i) {
-		ret = bch2_trans_mark_dev_sb(c, ca);
-		if (ret) {
-			percpu_ref_put(&ca->ref);
-			goto err;
-		}
+	ret = bch2_trans_mark_dev_sbs(c);
+	bch_err_msg(c, ret, "marking superblocks");
+	if (ret)
+		goto err;
 
+	for_each_online_member(c, ca)
 		ca->new_fs_bucket_idx = 0;
-	}
 
-	bch_verbose(c, "initializing freespace");
-	err = "error initializing freespace";
 	ret = bch2_fs_freespace_init(c);
 	if (ret)
 		goto err;
 
-	err = "error creating root snapshot node";
-	ret = bch2_fs_initialize_subvolumes(c);
+	ret = bch2_initialize_subvolumes(c);
 	if (ret)
 		goto err;
 
 	bch_verbose(c, "reading snapshots table");
-	err = "error reading snapshots table";
-	ret = bch2_fs_snapshots_start(c);
+	ret = bch2_snapshots_read(c);
 	if (ret)
 		goto err;
 	bch_verbose(c, "reading snapshots done");
@@ -1571,27 +1074,25 @@ int bch2_fs_initialize(struct bch_fs *c)
 	bch2_inode_pack(&packed_inode, &root_inode);
 	packed_inode.inode.k.p.snapshot = U32_MAX;
 
-	err = "error creating root directory";
-	ret = bch2_btree_insert(c, BTREE_ID_inodes,
-				&packed_inode.inode.k_i,
-				NULL, NULL, 0);
+	ret = bch2_btree_insert(c, BTREE_ID_inodes, &packed_inode.inode.k_i, NULL, 0);
+	bch_err_msg(c, ret, "creating root directory");
 	if (ret)
 		goto err;
 
 	bch2_inode_init_early(c, &lostfound_inode);
 
-	err = "error creating lost+found";
 	ret = bch2_trans_do(c, NULL, NULL, 0,
-		bch2_create_trans(&trans,
+		bch2_create_trans(trans,
 				  BCACHEFS_ROOT_SUBVOL_INUM,
 				  &root_inode, &lostfound_inode,
 				  &lostfound,
 				  0, 0, S_IFDIR|0700, 0,
 				  NULL, NULL, (subvol_inum) { 0 }, 0));
-	if (ret) {
-		bch_err(c, "error creating lost+found");
+	bch_err_msg(c, ret, "creating lost+found");
+	if (ret)
 		goto err;
-	}
+
+	c->recovery_pass_done = ARRAY_SIZE(recovery_pass_fns) - 1;
 
 	if (enabled_qtypes(c)) {
 		ret = bch2_fs_quota_read(c);
@@ -1599,8 +1100,8 @@ int bch2_fs_initialize(struct bch_fs *c)
 			goto err;
 	}
 
-	err = "error writing first journal entry";
 	ret = bch2_journal_flush(&c->journal);
+	bch_err_msg(c, ret, "writing first journal entry");
 	if (ret)
 		goto err;
 
@@ -1613,6 +1114,6 @@ int bch2_fs_initialize(struct bch_fs *c)
 
 	return 0;
 err:
-	pr_err("Error initializing new filesystem: %s (%i)", err, ret);
+	bch_err_fn(c, ret);
 	return ret;
 }