bcachefs: bch2_trans_mark_update()

18 files changed, 702 insertions, 142 deletions

diff --git a/fs/bcachefs/alloc_background.c b/fs/bcachefs/alloc_background.c
index fabe8d364809..f7b6e8c793c5 100644
--- a/fs/bcachefs/alloc_background.c
+++ b/fs/bcachefs/alloc_background.c
@@ -140,8 +140,8 @@ struct bkey_alloc_unpacked bch2_alloc_unpack(const struct bch_alloc *a)
 	return ret;
 }
 
-static void bch2_alloc_pack(struct bkey_i_alloc *dst,
-			    const struct bkey_alloc_unpacked src)
+void bch2_alloc_pack(struct bkey_i_alloc *dst,
+		     const struct bkey_alloc_unpacked src)
 {
 	unsigned idx = 0;
 	void *d = dst->v.data;
@@ -961,7 +961,6 @@ retry:
 
 	invalidating_cached_data = m.cached_sectors != 0;
 
-	//BUG_ON(u.dirty_sectors);
 	u.data_type	= 0;
 	u.dirty_sectors	= 0;
 	u.cached_sectors = 0;
@@ -973,6 +972,7 @@ retry:
 	 * we have to trust the in memory bucket @m, not the version in the
 	 * btree:
 	 */
+	//BUG_ON(u.dirty_sectors);
 	u.gen		= m.gen + 1;
 
 	a = bkey_alloc_init(&alloc_key.k);
diff --git a/fs/bcachefs/alloc_background.h b/fs/bcachefs/alloc_background.h
index 057af2fa9bf4..39c31e391fe4 100644
--- a/fs/bcachefs/alloc_background.h
+++ b/fs/bcachefs/alloc_background.h
@@ -13,6 +13,8 @@ struct bkey_alloc_unpacked {
 };
 
 struct bkey_alloc_unpacked bch2_alloc_unpack(const struct bch_alloc *);
+void bch2_alloc_pack(struct bkey_i_alloc *,
+		     const struct bkey_alloc_unpacked);
 
 #define ALLOC_SCAN_BATCH(ca)		max_t(size_t, 1, (ca)->mi.nbuckets >> 9)
 
diff --git a/fs/bcachefs/bcachefs_format.h b/fs/bcachefs/bcachefs_format.h
index d390ac860d18..be6acec19671 100644
--- a/fs/bcachefs/bcachefs_format.h
+++ b/fs/bcachefs/bcachefs_format.h
@@ -1295,6 +1295,7 @@ enum bch_sb_features {
 
 enum bch_sb_compat {
 	BCH_COMPAT_FEAT_ALLOC_INFO	= 0,
+	BCH_COMPAT_FEAT_ALLOC_METADATA	= 1,
 };
 
 /* options: */
diff --git a/fs/bcachefs/btree_iter.c b/fs/bcachefs/btree_iter.c
index 33cbc2ff5c9a..49ddf05cc9a9 100644
--- a/fs/bcachefs/btree_iter.c
+++ b/fs/bcachefs/btree_iter.c
@@ -1002,7 +1002,7 @@ retry_all:
 			goto retry_all;
 	}
 
-	ret = btree_trans_has_multiple_iters(trans) ? -EINTR : 0;
+	ret = hweight64(trans->iters_live) > 1 ? -EINTR : 0;
 out:
 	bch2_btree_cache_cannibalize_unlock(c);
 	return ret;
@@ -1100,8 +1100,6 @@ int __must_check bch2_btree_iter_traverse(struct btree_iter *iter)
 	if (unlikely(ret))
 		ret = __btree_iter_traverse_all(iter->trans, iter, ret);
 
-	BUG_ON(ret == -EINTR && !btree_trans_has_multiple_iters(iter->trans));
-
 	return ret;
 }
 
diff --git a/fs/bcachefs/btree_types.h b/fs/bcachefs/btree_types.h
index a995efc73fdc..ae273ab7aa1a 100644
--- a/fs/bcachefs/btree_types.h
+++ b/fs/bcachefs/btree_types.h
@@ -6,6 +6,7 @@
 #include <linux/six.h>
 
 #include "bkey_methods.h"
+#include "buckets_types.h"
 #include "journal_types.h"
 
 struct open_bucket;
@@ -260,6 +261,7 @@ struct btree_insert_entry {
 	};
 
 	bool			deferred;
+	bool			triggered;
 };
 
 #define BTREE_ITER_MAX		64
@@ -297,6 +299,8 @@ struct btree_trans {
 
 	struct btree_iter	iters_onstack[2];
 	struct btree_insert_entry updates_onstack[6];
+
+	struct replicas_delta_list fs_usage_deltas;
 };
 
 #define BTREE_FLAG(flag)						\
diff --git a/fs/bcachefs/btree_update.h b/fs/bcachefs/btree_update.h
index 2c4b8c797974..be11efdcbe04 100644
--- a/fs/bcachefs/btree_update.h
+++ b/fs/bcachefs/btree_update.h
@@ -42,8 +42,11 @@ enum {
 	__BTREE_INSERT_USE_ALLOC_RESERVE,
 	__BTREE_INSERT_JOURNAL_REPLAY,
 	__BTREE_INSERT_JOURNAL_RESERVED,
+	__BTREE_INSERT_NOMARK_INSERT,
 	__BTREE_INSERT_NOMARK_OVERWRITES,
 	__BTREE_INSERT_NOMARK,
+	__BTREE_INSERT_MARK_INMEM,
+	__BTREE_INSERT_NO_CLEAR_REPLICAS,
 	__BTREE_INSERT_NOWAIT,
 	__BTREE_INSERT_GC_LOCK_HELD,
 	__BCH_HASH_SET_MUST_CREATE,
@@ -76,12 +79,20 @@ enum {
 
 #define BTREE_INSERT_JOURNAL_RESERVED	(1 << __BTREE_INSERT_JOURNAL_RESERVED)
 
+/* Don't mark new key, just overwrites: */
+#define BTREE_INSERT_NOMARK_INSERT	(1 << __BTREE_INSERT_NOMARK_INSERT)
+
 /* Don't mark overwrites, just new key: */
 #define BTREE_INSERT_NOMARK_OVERWRITES	(1 << __BTREE_INSERT_NOMARK_OVERWRITES)
 
-/* Don't call bch2_mark_key: */
+/* Don't call mark new key at all: */
 #define BTREE_INSERT_NOMARK		(1 << __BTREE_INSERT_NOMARK)
 
+/* Don't mark transactionally: */
+#define BTREE_INSERT_MARK_INMEM		(1 << __BTREE_INSERT_MARK_INMEM)
+
+#define BTREE_INSERT_NO_CLEAR_REPLICAS	(1 << __BTREE_INSERT_NO_CLEAR_REPLICAS)
+
 /* Don't block on allocation failure (for new btree nodes: */
 #define BTREE_INSERT_NOWAIT		(1 << __BTREE_INSERT_NOWAIT)
 #define BTREE_INSERT_GC_LOCK_HELD	(1 << __BTREE_INSERT_GC_LOCK_HELD)
diff --git a/fs/bcachefs/btree_update_leaf.c b/fs/bcachefs/btree_update_leaf.c
index 1a72a584ebef..d052ca541965 100644
--- a/fs/bcachefs/btree_update_leaf.c
+++ b/fs/bcachefs/btree_update_leaf.c
@@ -524,6 +524,22 @@ static inline void do_btree_insert_one(struct btree_trans *trans,
 		btree_insert_key_deferred(trans, insert);
 }
 
+static inline bool update_triggers_transactional(struct btree_trans *trans,
+						 struct btree_insert_entry *i)
+{
+	return likely(!(trans->flags & BTREE_INSERT_MARK_INMEM)) &&
+		(i->iter->btree_id == BTREE_ID_EXTENTS ||
+		 i->iter->btree_id == BTREE_ID_INODES);
+}
+
+static inline bool update_has_triggers(struct btree_trans *trans,
+				       struct btree_insert_entry *i)
+{
+	return likely(!(trans->flags & BTREE_INSERT_NOMARK)) &&
+		!i->deferred &&
+		btree_node_type_needs_gc(i->iter->btree_id);
+}
+
 /*
  * Get journal reservation, take write locks, and attempt to do btree update(s):
  */
@@ -536,29 +552,25 @@ static inline int do_btree_insert_at(struct btree_trans *trans,
 	struct btree_iter *linked;
 	int ret;
 
+	if (likely(!(trans->flags & BTREE_INSERT_NO_CLEAR_REPLICAS))) {
+		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;
+	}
+
 	trans_for_each_update_iter(trans, i)
 		BUG_ON(i->iter->uptodate >= BTREE_ITER_NEED_RELOCK);
 
-	btree_trans_lock_write(c, trans);
-
-	if (likely(!(trans->flags & BTREE_INSERT_NOMARK))) {
-		trans_for_each_update_iter(trans, i) {
-			if (i->deferred ||
-			    !btree_node_type_needs_gc(i->iter->btree_id))
-				continue;
-
-			if (!fs_usage) {
-				percpu_down_read(&c->mark_lock);
-				fs_usage = bch2_fs_usage_scratch_get(c);
-			}
-
-			if (!bch2_bkey_replicas_marked_locked(c,
-					bkey_i_to_s_c(i->k), true)) {
-				ret = BTREE_INSERT_NEED_MARK_REPLICAS;
-				goto out;
-			}
+	trans_for_each_update_iter(trans, i)
+		if (update_has_triggers(trans, i) &&
+		    update_triggers_transactional(trans, i)) {
+			ret = bch2_trans_mark_update(trans, i,
+						&trans->fs_usage_deltas);
+			if (ret)
+				return ret;
 		}
-	}
+
+	btree_trans_lock_write(c, trans);
 
 	if (race_fault()) {
 		ret = -EINTR;
@@ -575,6 +587,23 @@ static inline int do_btree_insert_at(struct btree_trans *trans,
 	if (ret)
 		goto out;
 
+	trans_for_each_update_iter(trans, i) {
+		if (i->deferred ||
+		    !btree_node_type_needs_gc(i->iter->btree_id))
+			continue;
+
+		if (!fs_usage) {
+			percpu_down_read(&c->mark_lock);
+			fs_usage = bch2_fs_usage_scratch_get(c);
+		}
+
+		if (!bch2_bkey_replicas_marked_locked(c,
+			bkey_i_to_s_c(i->k), true)) {
+			ret = BTREE_INSERT_NEED_MARK_REPLICAS;
+			goto out;
+		}
+	}
+
 	/*
 	 * Don't get journal reservation until after we know insert will
 	 * succeed:
@@ -603,20 +632,24 @@ static inline int do_btree_insert_at(struct btree_trans *trans,
 				linked->flags |= BTREE_ITER_NOUNLOCK;
 	}
 
-	if (likely(!(trans->flags & BTREE_INSERT_NOMARK))) {
-		trans_for_each_update_iter(trans, i)
+	trans_for_each_update_iter(trans, i)
+		if (update_has_triggers(trans, i) &&
+		    !update_triggers_transactional(trans, i))
 			bch2_mark_update(trans, i, fs_usage, 0);
-		if (fs_usage)
-			bch2_trans_fs_usage_apply(trans, fs_usage);
-
-		if (unlikely(c->gc_pos.phase)) {
-			trans_for_each_update_iter(trans, i)
-				if (gc_visited(c, gc_pos_btree_node(i->iter->l[0].b)))
-					bch2_mark_update(trans, i, NULL,
-							 BCH_BUCKET_MARK_GC);
-		}
+
+	if (fs_usage) {
+		bch2_replicas_delta_list_apply(c, fs_usage,
+					       &trans->fs_usage_deltas);
+		bch2_trans_fs_usage_apply(trans, fs_usage);
 	}
 
+	if (likely(!(trans->flags & BTREE_INSERT_NOMARK)) &&
+	    unlikely(c->gc_pos.phase))
+		trans_for_each_update_iter(trans, i)
+			if (gc_visited(c, gc_pos_btree_node(i->iter->l[0].b)))
+				bch2_mark_update(trans, i, NULL,
+						 BCH_BUCKET_MARK_GC);
+
 	trans_for_each_update(trans, i)
 		do_btree_insert_one(trans, i);
 out:
@@ -643,6 +676,19 @@ int bch2_trans_commit_error(struct btree_trans *trans,
 {
 	struct bch_fs *c = trans->c;
 	unsigned flags = trans->flags;
+	struct btree_insert_entry *src, *dst;
+
+	src = dst = trans->updates;
+
+	while (src < trans->updates + trans->nr_updates) {
+		if (!src->triggered) {
+			*dst = *src;
+			dst++;
+		}
+		src++;
+	}
+
+	trans->nr_updates = dst - trans->updates;
 
 	/*
 	 * BTREE_INSERT_NOUNLOCK means don't unlock _after_ successful btree
@@ -800,6 +846,7 @@ int bch2_trans_commit(struct btree_trans *trans,
 {
 	struct bch_fs *c = trans->c;
 	struct btree_insert_entry *i;
+	unsigned orig_mem_top = trans->mem_top;
 	int ret = 0;
 
 	if (!trans->nr_updates)
@@ -877,8 +924,16 @@ out_noupdates:
 	return ret;
 err:
 	ret = bch2_trans_commit_error(trans, i, ret);
-	if (!ret)
+
+	/* can't loop if it was passed in and we changed it: */
+	if (unlikely(trans->flags & BTREE_INSERT_NO_CLEAR_REPLICAS) && !ret)
+		ret = -EINTR;
+
+	if (!ret) {
+		/* free memory used by triggers, they'll be reexecuted: */
+		trans->mem_top = orig_mem_top;
 		goto retry;
+	}
 
 	goto out;
 }
@@ -961,6 +1016,7 @@ int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
 	int ret = 0;
 
 	bch2_trans_init(&trans, c);
+	bch2_trans_preload_iters(&trans);
 
 	iter = bch2_trans_get_iter(&trans, id, start, BTREE_ITER_INTENT);
 
@@ -1006,5 +1062,6 @@ int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
 	}
 
 	bch2_trans_exit(&trans);
+	BUG_ON(ret == -EINTR);
 	return ret;
 }
diff --git a/fs/bcachefs/buckets.c b/fs/bcachefs/buckets.c
index 3834b150a0d1..6d452311d924 100644
--- a/fs/bcachefs/buckets.c
+++ b/fs/bcachefs/buckets.c
@@ -646,19 +646,16 @@ static int bch2_mark_alloc(struct bch_fs *c, struct bkey_s_c k,
 	ca = bch_dev_bkey_exists(c, k.k->p.inode);
 	g = __bucket(ca, k.k->p.offset, gc);
 
-	/*
-	 * this should currently only be getting called from the bucket
-	 * invalidate path:
-	 */
-	BUG_ON(u.dirty_sectors);
-	BUG_ON(u.cached_sectors);
-	BUG_ON(!g->mark.owned_by_allocator);
-
 	old = bucket_data_cmpxchg(c, ca, fs_usage, g, m, ({
 		m.gen			= u.gen;
 		m.data_type		= u.data_type;
 		m.dirty_sectors		= u.dirty_sectors;
 		m.cached_sectors	= u.cached_sectors;
+
+		if (!(flags & BCH_BUCKET_MARK_GC)) {
+			m.journal_seq_valid	= 1;
+			m.journal_seq		= journal_seq;
+		}
 	}));
 
 	g->io_time[READ]	= u.read_time;
@@ -666,6 +663,11 @@ static int bch2_mark_alloc(struct bch_fs *c, struct bkey_s_c k,
 	g->oldest_gen		= u.oldest_gen;
 	g->gen_valid		= 1;
 
+	/*
+	 * need to know if we're getting called from the invalidate path or
+	 * not:
+	 */
+
 	if (old.cached_sectors) {
 		update_cached_sectors(c, fs_usage, ca->dev_idx,
 				      -old.cached_sectors);
@@ -751,11 +753,34 @@ static s64 ptr_disk_sectors_delta(struct extent_ptr_decoded p,
 	}
 }
 
-/*
- * Checking against gc's position has to be done here, inside the cmpxchg()
- * loop, to avoid racing with the start of gc clearing all the marks - GC does
- * that with the gc pos seqlock held.
- */
+static void bucket_set_stripe(struct bch_fs *c,
+			      const struct bch_stripe *v,
+			      bool enabled,
+			      struct bch_fs_usage *fs_usage,
+			      u64 journal_seq,
+			      bool gc)
+{
+	unsigned i;
+
+	for (i = 0; i < v->nr_blocks; i++) {
+		const struct bch_extent_ptr *ptr = v->ptrs + i;
+		struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
+		struct bucket *g = PTR_BUCKET(ca, ptr, gc);
+		struct bucket_mark new, old;
+
+		BUG_ON(ptr_stale(ca, ptr));
+
+		old = bucket_data_cmpxchg(c, ca, fs_usage, g, new, ({
+			new.dirty			= true;
+			new.stripe			= enabled;
+			if (journal_seq) {
+				new.journal_seq_valid	= 1;
+				new.journal_seq		= journal_seq;
+			}
+		}));
+	}
+}
+
 static bool bch2_mark_pointer(struct bch_fs *c,
 			      struct extent_ptr_decoded p,
 			      s64 sectors, enum bch_data_type data_type,
@@ -765,8 +790,7 @@ static bool bch2_mark_pointer(struct bch_fs *c,
 {
 	struct bucket_mark old, new;
 	struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
-	size_t b = PTR_BUCKET_NR(ca, &p.ptr);
-	struct bucket *g = __bucket(ca, b, gc);
+	struct bucket *g = PTR_BUCKET(ca, &p.ptr, gc);
 	bool overflow;
 	u64 v;
 
@@ -935,35 +959,6 @@ static int bch2_mark_extent(struct bch_fs *c, struct bkey_s_c k,
 	return 0;
 }
 
-static void bucket_set_stripe(struct bch_fs *c,
-			      const struct bch_stripe *v,
-			      bool enabled,
-			      struct bch_fs_usage *fs_usage,
-			      u64 journal_seq,
-			      bool gc)
-{
-	unsigned i;
-
-	for (i = 0; i < v->nr_blocks; i++) {
-		const struct bch_extent_ptr *ptr = v->ptrs + i;
-		struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
-		size_t b = PTR_BUCKET_NR(ca, ptr);
-		struct bucket *g = __bucket(ca, b, gc);
-		struct bucket_mark new, old;
-
-		BUG_ON(ptr_stale(ca, ptr));
-
-		old = bucket_data_cmpxchg(c, ca, fs_usage, g, new, ({
-			new.dirty			= true;
-			new.stripe			= enabled;
-			if (journal_seq) {
-				new.journal_seq_valid	= 1;
-				new.journal_seq		= journal_seq;
-			}
-		}));
-	}
-}
-
 static int bch2_mark_stripe(struct bch_fs *c, struct bkey_s_c k,
 			    bool inserting,
 			    struct bch_fs_usage *fs_usage,
@@ -995,14 +990,7 @@ static int bch2_mark_stripe(struct bch_fs *c, struct bkey_s_c k,
 		m->nr_blocks	= s.v->nr_blocks;
 		m->nr_redundant	= s.v->nr_redundant;
 
-		memset(&m->r, 0, sizeof(m->r));
-
-		m->r.e.data_type	= BCH_DATA_USER;
-		m->r.e.nr_devs		= s.v->nr_blocks;
-		m->r.e.nr_required	= s.v->nr_blocks - s.v->nr_redundant;
-
-		for (i = 0; i < s.v->nr_blocks; i++)
-			m->r.e.devs[i] = s.v->ptrs[i].dev;
+		bch2_bkey_to_replicas(&m->r.e, k);
 
 	/*
 	 * XXX: account for stripes somehow here
@@ -1169,10 +1157,11 @@ int bch2_mark_update(struct btree_trans *trans,
 	if (!btree_node_type_needs_gc(iter->btree_id))
 		return 0;
 
-	bch2_mark_key_locked(c, bkey_i_to_s_c(insert->k), true,
-		bpos_min(insert->k->k.p, b->key.k.p).offset -
-		bkey_start_offset(&insert->k->k),
-		fs_usage, trans->journal_res.seq, flags);
+	if (!(trans->flags & BTREE_INSERT_NOMARK_INSERT))
+		bch2_mark_key_locked(c, bkey_i_to_s_c(insert->k), true,
+			bpos_min(insert->k->k.p, b->key.k.p).offset -
+			bkey_start_offset(&insert->k->k),
+			fs_usage, trans->journal_res.seq, flags);
 
 	if (unlikely(trans->flags & BTREE_INSERT_NOMARK_OVERWRITES))
 		return 0;
@@ -1251,6 +1240,391 @@ void bch2_trans_fs_usage_apply(struct btree_trans *trans,
 	}
 }
 
+/* trans_mark: */
+
+static inline void update_replicas_list(struct replicas_delta_list *d,
+					struct bch_replicas_entry *r,
+					s64 sectors)
+{
+	d->top->delta = sectors;
+	memcpy(&d->top->r, r, replicas_entry_bytes(r));
+
+	d->top = (void *) d->top + replicas_entry_bytes(r) + 8;
+
+	BUG_ON((void *) d->top > (void *) d->d + sizeof(d->pad));
+}
+
+static inline void update_cached_sectors_list(struct replicas_delta_list *d,
+					      unsigned dev, s64 sectors)
+{
+	struct bch_replicas_padded r;
+
+	bch2_replicas_entry_cached(&r.e, dev);
+
+	update_replicas_list(d, &r.e, sectors);
+}
+
+void bch2_replicas_delta_list_apply(struct bch_fs *c,
+				    struct bch_fs_usage *fs_usage,
+				    struct replicas_delta_list *r)
+{
+	struct replicas_delta *d = r->d;
+
+	acc_u64s((u64 *) fs_usage,
+		 (u64 *) &r->fs_usage, sizeof(*fs_usage) / sizeof(u64));
+
+	while (d != r->top) {
+		BUG_ON((void *) d > (void *) r->top);
+
+		update_replicas(c, fs_usage, &d->r, d->delta);
+
+		d = (void *) d + replicas_entry_bytes(&d->r) + 8;
+	}
+}
+
+static int trans_get_key(struct btree_trans *trans,
+			 enum btree_id btree_id, struct bpos pos,
+			 struct btree_insert_entry **insert,
+			 struct btree_iter **iter,
+			 struct bkey_s_c *k)
+{
+	unsigned i;
+	int ret;
+
+	*insert = NULL;
+
+	for (i = 0; i < trans->nr_updates; i++)
+		if (!trans->updates[i].deferred &&
+		    trans->updates[i].iter->btree_id == btree_id &&
+		    !bkey_cmp(pos, trans->updates[i].iter->pos)) {
+			*insert = &trans->updates[i];
+			*iter	= (*insert)->iter;
+			*k	= bkey_i_to_s_c((*insert)->k);
+			return 0;
+		}
+
+	*iter = __bch2_trans_get_iter(trans, btree_id, pos,
+				   BTREE_ITER_SLOTS|BTREE_ITER_INTENT, 0);
+	if (IS_ERR(*iter))
+		return PTR_ERR(*iter);
+
+	*k = bch2_btree_iter_peek_slot(*iter);
+	ret = bkey_err(*k);
+	if (ret)
+		bch2_trans_iter_put(trans, *iter);
+	return ret;
+}
+
+static int trans_update_key(struct btree_trans *trans,
+			    struct btree_insert_entry **insert,
+			    struct btree_iter *iter,
+			    struct bkey_s_c k,
+			    unsigned extra_u64s)
+{
+	struct bkey_i *new_k;
+
+	if (*insert)
+		return 0;
+
+	new_k = bch2_trans_kmalloc(trans, bkey_bytes(k.k) +
+				   extra_u64s * sizeof(u64));
+	if (IS_ERR(new_k))
+		return PTR_ERR(new_k);
+
+	*insert = bch2_trans_update(trans, ((struct btree_insert_entry) {
+				.iter = iter,
+				.k = new_k,
+				.triggered = true,
+	}));
+
+	bkey_reassemble((*insert)->k, k);
+	return 0;
+}
+
+static int bch2_trans_mark_pointer(struct btree_trans *trans,
+			struct extent_ptr_decoded p,
+			s64 sectors, enum bch_data_type data_type,
+			struct replicas_delta_list *d)
+{
+	struct bch_fs *c = trans->c;
+	struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
+	struct btree_insert_entry *insert;
+	struct btree_iter *iter;
+	struct bkey_s_c k;
+	struct bkey_alloc_unpacked u;
+	struct bkey_i_alloc *a;
+	bool overflow;
+	int ret;
+
+	ret = trans_get_key(trans, BTREE_ID_ALLOC,
+			    POS(p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr)),
+			    &insert, &iter, &k);
+	if (ret)
+		return ret;
+
+	if (k.k->type != KEY_TYPE_alloc) {
+		bch_err_ratelimited(c, "pointer to nonexistent bucket %u:%zu",
+				    p.ptr.dev,
+				    PTR_BUCKET_NR(ca, &p.ptr));
+		ret = -1;
+		goto out;
+	}
+
+	u = bch2_alloc_unpack(bkey_s_c_to_alloc(k).v);
+
+	if (gen_after(u.gen, p.ptr.gen)) {
+		ret = 1;
+		goto out;
+	}
+
+	if (!p.ptr.cached)
+		overflow = checked_add(u.dirty_sectors, sectors);
+	else
+		overflow = checked_add(u.cached_sectors, sectors);
+
+	u.data_type = u.dirty_sectors || u.cached_sectors
+		? data_type : 0;
+
+	bch2_fs_inconsistent_on(overflow, c,
+		"bucket sector count overflow: %u + %lli > U16_MAX",
+		!p.ptr.cached
+		? u.dirty_sectors
+		: u.cached_sectors, sectors);
+
+	ret = trans_update_key(trans, &insert, iter, k, 1);
+	if (ret)
+		goto out;
+
+	a = bkey_alloc_init(insert->k);
+	a->k.p = iter->pos;
+	bch2_alloc_pack(a, u);
+out:
+	bch2_trans_iter_put(trans, iter);
+	return ret;
+}
+
+static int bch2_trans_mark_stripe_ptr(struct btree_trans *trans,
+			struct bch_extent_stripe_ptr p,
+			s64 sectors, enum bch_data_type data_type,
+			struct replicas_delta_list *d)
+{
+	struct bch_replicas_padded r;
+	struct btree_insert_entry *insert;
+	struct btree_iter *iter;
+	struct bkey_s_c k;
+	struct bkey_s_stripe s;
+	unsigned nr_data;
+	s64 parity_sectors;
+	int ret = 0;
+
+	BUG_ON(!sectors);
+
+	ret = trans_get_key(trans, BTREE_ID_EC, POS(0, p.idx),
+			    &insert, &iter, &k);
+	if (ret)
+		return ret;
+
+	if (k.k->type != KEY_TYPE_stripe) {
+		bch_err_ratelimited(trans->c,
+				    "pointer to nonexistent stripe %llu",
+				    (u64) p.idx);
+		ret = -1;
+		goto out;
+	}
+
+	ret = trans_update_key(trans, &insert, iter, k, 1);
+	if (ret)
+		goto out;
+
+	s = bkey_i_to_s_stripe(insert->k);
+
+	nr_data = s.v->nr_blocks - s.v->nr_redundant;
+
+	parity_sectors = DIV_ROUND_UP(abs(sectors) * s.v->nr_redundant, nr_data);
+
+	if (sectors < 0)
+		parity_sectors = -parity_sectors;
+
+	stripe_blockcount_set(s.v, p.block,
+		stripe_blockcount_get(s.v, p.block) +
+		sectors + parity_sectors);
+
+	bch2_bkey_to_replicas(&r.e, s.s_c);
+
+	update_replicas_list(d, &r.e, sectors);
+out:
+	bch2_trans_iter_put(trans, iter);
+	return ret;
+}
+
+static int bch2_trans_mark_extent(struct btree_trans *trans,
+			struct bkey_s_c k,
+			s64 sectors, enum bch_data_type data_type,
+			struct replicas_delta_list *d)
+{
+	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
+	const union bch_extent_entry *entry;
+	struct extent_ptr_decoded p;
+	struct bch_replicas_padded r;
+	s64 dirty_sectors = 0;
+	bool stale;
+	unsigned i;
+	int ret;
+
+	r.e.data_type	= data_type;
+	r.e.nr_devs	= 0;
+	r.e.nr_required	= 1;
+
+	BUG_ON(!sectors);
+
+	bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
+		s64 disk_sectors = data_type == BCH_DATA_BTREE
+			? sectors
+			: ptr_disk_sectors_delta(p, sectors);
+
+		ret = bch2_trans_mark_pointer(trans, p, disk_sectors,
+					      data_type, d);
+		if (ret < 0)
+			return ret;
+
+		stale = ret > 0;
+
+		if (p.ptr.cached) {
+			if (disk_sectors && !stale)
+				update_cached_sectors_list(d, p.ptr.dev,
+							   disk_sectors);
+		} else if (!p.ec_nr) {
+			dirty_sectors	       += disk_sectors;
+			r.e.devs[r.e.nr_devs++]	= p.ptr.dev;
+		} else {
+			for (i = 0; i < p.ec_nr; i++) {
+				ret = bch2_trans_mark_stripe_ptr(trans, p.ec[i],
+						disk_sectors, data_type, d);
+				if (ret)
+					return ret;
+			}
+
+			r.e.nr_required = 0;
+		}
+	}
+
+	if (dirty_sectors)
+		update_replicas_list(d, &r.e, dirty_sectors);
+
+	return 0;
+}
+
+int bch2_trans_mark_key(struct btree_trans *trans,
+			struct bkey_s_c k,
+			bool inserting, s64 sectors,
+			struct replicas_delta_list *d)
+{
+	struct bch_fs *c = trans->c;
+
+	switch (k.k->type) {
+	case KEY_TYPE_btree_ptr:
+		return bch2_trans_mark_extent(trans, k, inserting
+				?  c->opts.btree_node_size
+				: -c->opts.btree_node_size,
+				BCH_DATA_BTREE, d);
+	case KEY_TYPE_extent:
+		return bch2_trans_mark_extent(trans, k,
+				sectors, BCH_DATA_USER, d);
+	case KEY_TYPE_inode:
+		if (inserting)
+			d->fs_usage.nr_inodes++;
+		else
+			d->fs_usage.nr_inodes--;
+		return 0;
+	case KEY_TYPE_reservation: {
+		unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
+
+		sectors *= replicas;
+		replicas = clamp_t(unsigned, replicas, 1,
+				   ARRAY_SIZE(d->fs_usage.persistent_reserved));
+
+		d->fs_usage.reserved				+= sectors;
+		d->fs_usage.persistent_reserved[replicas - 1]	+= sectors;
+		return 0;
+	}
+	default:
+		return 0;
+	}
+}
+
+int bch2_trans_mark_update(struct btree_trans *trans,
+			   struct btree_insert_entry *insert,
+			   struct replicas_delta_list *d)
+{
+	struct btree_iter	*iter = insert->iter;
+	struct btree		*b = iter->l[0].b;
+	struct btree_node_iter	node_iter = iter->l[0].iter;
+	struct bkey_packed	*_k;
+	int ret;
+
+	if (!btree_node_type_needs_gc(iter->btree_id))
+		return 0;
+
+	ret = bch2_trans_mark_key(trans,
+			bkey_i_to_s_c(insert->k), true,
+			bpos_min(insert->k->k.p, b->key.k.p).offset -
+			bkey_start_offset(&insert->k->k), d);
+	if (ret)
+		return ret;
+
+	while ((_k = bch2_btree_node_iter_peek_filter(&node_iter, b,
+						      KEY_TYPE_discard))) {
+		struct bkey		unpacked;
+		struct bkey_s_c		k;
+		s64			sectors = 0;
+
+		k = bkey_disassemble(b, _k, &unpacked);
+
+		if (btree_node_is_extents(b)
+		    ? bkey_cmp(insert->k->k.p, bkey_start_pos(k.k)) <= 0
+		    : bkey_cmp(insert->k->k.p, k.k->p))
+			break;
+
+		if (btree_node_is_extents(b)) {
+			switch (bch2_extent_overlap(&insert->k->k, k.k)) {
+			case BCH_EXTENT_OVERLAP_ALL:
+				sectors = -((s64) k.k->size);
+				break;
+			case BCH_EXTENT_OVERLAP_BACK:
+				sectors = bkey_start_offset(&insert->k->k) -
+					k.k->p.offset;
+				break;
+			case BCH_EXTENT_OVERLAP_FRONT:
+				sectors = bkey_start_offset(k.k) -
+					insert->k->k.p.offset;
+				break;
+			case BCH_EXTENT_OVERLAP_MIDDLE:
+				sectors = k.k->p.offset - insert->k->k.p.offset;
+				BUG_ON(sectors <= 0);
+
+				ret = bch2_trans_mark_key(trans, k, true,
+							  sectors, d);
+				if (ret)
+					return ret;
+
+				sectors = bkey_start_offset(&insert->k->k) -
+					k.k->p.offset;
+				break;
+			}
+
+			BUG_ON(sectors >= 0);
+		}
+
+		ret = bch2_trans_mark_key(trans, k, false, sectors, d);
+		if (ret)
+			return ret;
+
+		bch2_btree_node_iter_advance(&node_iter, b);
+	}
+
+	return 0;
+}
+
 /* Disk reservations: */
 
 static u64 bch2_recalc_sectors_available(struct bch_fs *c)
diff --git a/fs/bcachefs/buckets.h b/fs/bcachefs/buckets.h
index 0aada931322f..a32c25d8f298 100644
--- a/fs/bcachefs/buckets.h
+++ b/fs/bcachefs/buckets.h
@@ -99,7 +99,7 @@ static inline struct bucket_mark ptr_bucket_mark(struct bch_dev *ca,
 	struct bucket_mark m;
 
 	rcu_read_lock();
-	m = READ_ONCE(bucket(ca, PTR_BUCKET_NR(ca, ptr))->mark);
+	m = READ_ONCE(PTR_BUCKET(ca, ptr, 0)->mark);
 	rcu_read_unlock();
 
 	return m;
@@ -265,6 +265,15 @@ int bch2_mark_overwrite(struct btree_trans *, struct btree_iter *,
 			struct bch_fs_usage *, unsigned);
 int bch2_mark_update(struct btree_trans *, struct btree_insert_entry *,
 		     struct bch_fs_usage *, unsigned);
+
+void bch2_replicas_delta_list_apply(struct bch_fs *,
+				    struct bch_fs_usage *,
+				    struct replicas_delta_list *);
+int bch2_trans_mark_key(struct btree_trans *, struct bkey_s_c,
+			bool, s64, struct replicas_delta_list *);
+int bch2_trans_mark_update(struct btree_trans *,
+			   struct btree_insert_entry *,
+			   struct replicas_delta_list *);
 void bch2_trans_fs_usage_apply(struct btree_trans *, struct bch_fs_usage *);
 
 /* disk reservations: */
diff --git a/fs/bcachefs/buckets_types.h b/fs/bcachefs/buckets_types.h
index 2a1fd7a7ec20..974daa7ef2d3 100644
--- a/fs/bcachefs/buckets_types.h
+++ b/fs/bcachefs/buckets_types.h
@@ -94,6 +94,19 @@ struct bch_fs_usage_short {
 	u64			nr_inodes;
 };
 
+struct replicas_delta {
+	s64			delta;
+	struct bch_replicas_entry r;
+};
+
+struct replicas_delta_list {
+	struct bch_fs_usage	fs_usage;
+
+	struct replicas_delta	*top;
+	struct replicas_delta	d[0];
+	u8			pad[256];
+};
+
 /*
  * A reservation for space on disk:
  */
diff --git a/fs/bcachefs/ec.c b/fs/bcachefs/ec.c
index 07c0298cad65..748fc559aae3 100644
--- a/fs/bcachefs/ec.c
+++ b/fs/bcachefs/ec.c
@@ -536,14 +536,17 @@ static int ec_stripe_mem_alloc(struct bch_fs *c,
 			       struct btree_iter *iter)
 {
 	size_t idx = iter->pos.offset;
+	int ret = 0;
 
 	if (!__ec_stripe_mem_alloc(c, idx, GFP_NOWAIT))
-		return 0;
+		return ret;
 
 	bch2_btree_trans_unlock(iter->trans);
+	ret = -EINTR;
 
 	if (!__ec_stripe_mem_alloc(c, idx, GFP_KERNEL))
-		return -EINTR;
+		return ret;
+
 	return -ENOMEM;
 }
 
@@ -689,23 +692,22 @@ retry:
 
 	if (!ret)
 		ret = -ENOSPC;
-	goto out;
+	goto err;
 found_slot:
 	ret = ec_stripe_mem_alloc(c, iter);
-
-	if (ret == -EINTR)
-		goto retry;
 	if (ret)
-		return ret;
+		goto err;
 
 	stripe->k.p = iter->pos;
 
 	bch2_trans_update(&trans, BTREE_INSERT_ENTRY(iter, &stripe->k_i));
 
 	ret = bch2_trans_commit(&trans, NULL, NULL,
-				BTREE_INSERT_NOFAIL|
-				BTREE_INSERT_USE_RESERVE);
-out:
+				BTREE_INSERT_ATOMIC|
+				BTREE_INSERT_NOFAIL);
+err:
+	if (ret == -EINTR)
+		goto retry;
 	bch2_trans_exit(&trans);
 
 	return ret;
@@ -742,6 +744,7 @@ static int ec_stripe_update_ptrs(struct bch_fs *c,
 	int ret = 0, dev, idx;
 
 	bch2_trans_init(&trans, c);
+	bch2_trans_preload_iters(&trans);
 
 	iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS,
 				   bkey_start_pos(pos),
diff --git a/fs/bcachefs/extents.c b/fs/bcachefs/extents.c
index d81e590c9124..f8f29251a5e7 100644
--- a/fs/bcachefs/extents.c
+++ b/fs/bcachefs/extents.c
@@ -872,15 +872,54 @@ static void extent_bset_insert(struct bch_fs *c, struct btree_iter *iter,
 	bch2_btree_iter_verify(iter, l->b);
 }
 
+static unsigned bch2_bkey_nr_alloc_ptrs(struct bkey_s_c k)
+{
+	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
+	const union bch_extent_entry *entry;
+	unsigned ret = 0;
+
+	bkey_extent_entry_for_each(ptrs, entry) {
+		switch (__extent_entry_type(entry)) {
+		case BCH_EXTENT_ENTRY_ptr:
+		case BCH_EXTENT_ENTRY_stripe_ptr:
+			ret++;
+		}
+	}
+
+	return ret;
+}
+
 static inline struct bpos
-bch2_extent_atomic_end(struct bkey_i *k, struct btree_iter *iter)
+bch2_extent_atomic_end(struct bkey_i *insert, struct btree_iter *iter)
 {
 	struct btree *b = iter->l[0].b;
+	struct btree_node_iter	node_iter = iter->l[0].iter;
+	struct bkey_packed	*_k;
+	unsigned		nr_alloc_ptrs =
+		bch2_bkey_nr_alloc_ptrs(bkey_i_to_s_c(insert));
 
 	BUG_ON(iter->uptodate > BTREE_ITER_NEED_PEEK);
-	BUG_ON(bkey_cmp(bkey_start_pos(&k->k), b->data->min_key) < 0);
+	BUG_ON(bkey_cmp(bkey_start_pos(&insert->k), b->data->min_key) < 0);
+
+	while ((_k = bch2_btree_node_iter_peek_filter(&node_iter, b,
+						      KEY_TYPE_discard))) {
+		struct bkey	unpacked;
+		struct bkey_s_c	k = bkey_disassemble(b, _k, &unpacked);
+
+		if (bkey_cmp(insert->k.p, bkey_start_pos(k.k)) <= 0)
+			break;
+
+		nr_alloc_ptrs += bch2_bkey_nr_alloc_ptrs(k);
+
+		if (nr_alloc_ptrs > 20) {
+			BUG_ON(bkey_cmp(k.k->p, bkey_start_pos(&insert->k)) <= 0);
+			return bpos_min(insert->k.p, k.k->p);
+		}
+
+		bch2_btree_node_iter_advance(&node_iter, b);
+	}
 
-	return bpos_min(k->k.p, b->key.k.p);
+	return bpos_min(insert->k.p, b->key.k.p);
 }
 
 void bch2_extent_trim_atomic(struct bkey_i *k, struct btree_iter *iter)
diff --git a/fs/bcachefs/migrate.c b/fs/bcachefs/migrate.c
index 88761d34dc65..822b3fce09b4 100644
--- a/fs/bcachefs/migrate.c
+++ b/fs/bcachefs/migrate.c
@@ -42,6 +42,7 @@ static int bch2_dev_usrdata_drop(struct bch_fs *c, unsigned dev_idx, int flags)
 	int ret = 0;
 
 	bch2_trans_init(&trans, c);
+	bch2_trans_preload_iters(&trans);
 
 	iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS,
 				   POS_MIN, BTREE_ITER_PREFETCH);
@@ -95,6 +96,8 @@ static int bch2_dev_usrdata_drop(struct bch_fs *c, unsigned dev_idx, int flags)
 			break;
 	}
 
+	BUG_ON(ret == -EINTR);
+
 	bch2_trans_exit(&trans);
 
 	bch2_replicas_gc_end(c, ret);
diff --git a/fs/bcachefs/move.c b/fs/bcachefs/move.c
index 7c69a4a51fba..946e616228c9 100644
--- a/fs/bcachefs/move.c
+++ b/fs/bcachefs/move.c
@@ -62,6 +62,7 @@ static int bch2_migrate_index_update(struct bch_write_op *op)
 	int ret = 0;
 
 	bch2_trans_init(&trans, c);
+	bch2_trans_preload_iters(&trans);
 
 	iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS,
 				   bkey_start_pos(&bch2_keylist_front(keys)->k),
@@ -184,6 +185,7 @@ nomatch:
 	}
 out:
 	bch2_trans_exit(&trans);
+	BUG_ON(ret == -EINTR);
 	return ret;
 }
 
diff --git a/fs/bcachefs/recovery.c b/fs/bcachefs/recovery.c
index 3f60b91629b3..75d64b7f2307 100644
--- a/fs/bcachefs/recovery.c
+++ b/fs/bcachefs/recovery.c
@@ -211,11 +211,6 @@ static int bch2_extent_replay_key(struct bch_fs *c, struct bkey_i *k)
 		bch2_disk_reservation_init(c, 0);
 	struct bkey_i *split;
 	bool split_compressed = false;
-	unsigned flags = BTREE_INSERT_ATOMIC|
-		BTREE_INSERT_NOFAIL|
-		BTREE_INSERT_LAZY_RW|
-		BTREE_INSERT_JOURNAL_REPLAY|
-		BTREE_INSERT_NOMARK;
 	int ret;
 
 	bch2_trans_init(&trans, c);
@@ -251,9 +246,6 @@ retry:
 					BCH_DISK_RESERVATION_NOFAIL);
 			BUG_ON(ret);
 
-			flags &= ~BTREE_INSERT_JOURNAL_REPLAY;
-			flags &= ~BTREE_INSERT_NOMARK;
-			flags |=  BTREE_INSERT_NOMARK_OVERWRITES;
 			split_compressed = true;
 		}
 
@@ -265,24 +257,31 @@ retry:
 		bch2_btree_iter_set_pos(iter, split->k.p);
 	} while (bkey_cmp(iter->pos, k->k.p) < 0);
 
-	ret = bch2_trans_commit(&trans, &disk_res, NULL, flags);
-	if (ret)
-		goto err;
-
 	if (split_compressed) {
-		/*
-		 * This isn't strictly correct - we should only be relying on
-		 * the btree node lock for synchronization with gc when we've
-		 * got a write lock held.
-		 *
-		 * but - there are other correctness issues if btree gc were to
-		 * run before journal replay finishes
-		 */
-		BUG_ON(c->gc_pos.phase);
-
-		bch2_mark_key(c, bkey_i_to_s_c(k), false, -((s64) k->k.size),
-			      NULL, 0, 0);
+		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;
@@ -526,7 +525,7 @@ static int verify_superblock_clean(struct bch_fs *c,
 	struct bch_sb_field_clean *clean = *cleanp;
 	int ret = 0;
 
-	if (!clean || !j)
+	if (!c->sb.clean || !j)
 		return 0;
 
 	if (mustfix_fsck_err_on(j->seq != clean->journal_seq, c,
@@ -652,6 +651,7 @@ int bch2_fs_recovery(struct bch_fs *c)
 	u64 journal_seq;
 	LIST_HEAD(journal_entries);
 	struct journal_keys journal_keys = { NULL };
+	bool wrote = false, write_sb = false;
 	int ret;
 
 	if (c->sb.clean)
@@ -676,8 +676,12 @@ int bch2_fs_recovery(struct bch_fs *c)
 		if (ret)
 			goto err;
 
-		fsck_err_on(c->sb.clean && !journal_empty(&journal_entries), 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_entries)) {
 			bch_err(c, "no journal entries found");
@@ -735,12 +739,15 @@ 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_keys);
 	if (ret)
 		goto err;
+	bch_verbose(c, "alloc read done");
 
 	bch_verbose(c, "starting stripes_read");
+	err = "error reading stripes";
 	ret = bch2_stripes_read(c, &journal_keys);
 	if (ret)
 		goto err;
@@ -748,11 +755,26 @@ int bch2_fs_recovery(struct bch_fs *c)
 
 	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_verbose(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";
+		err = "error in mark and sweep";
 		ret = bch2_gc(c, &journal_keys, true, false);
 		if (ret)
 			goto err;
@@ -779,6 +801,16 @@ int bch2_fs_recovery(struct bch_fs *c)
 		goto err;
 	bch_verbose(c, "journal replay done");
 
+	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");
+
 	if (c->opts.norecovery)
 		goto out;
 
@@ -801,13 +833,23 @@ 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 &&
@@ -820,7 +862,7 @@ out:
 	return ret;
 err:
 fsck_err:
-	pr_err("Error in recovery: %s (%i)", err, ret);
+	bch_err(c, "Error in recovery: %s (%i)", err, ret);
 	goto out;
 }
 
diff --git a/fs/bcachefs/replicas.c b/fs/bcachefs/replicas.c
index c2aa330c9932..eb44119466d8 100644
--- a/fs/bcachefs/replicas.c
+++ b/fs/bcachefs/replicas.c
@@ -101,8 +101,8 @@ static void stripe_to_replicas(struct bkey_s_c k,
 		r->devs[r->nr_devs++] = ptr->dev;
 }
 
-static void bkey_to_replicas(struct bch_replicas_entry *e,
-			     struct bkey_s_c k)
+void bch2_bkey_to_replicas(struct bch_replicas_entry *e,
+			   struct bkey_s_c k)
 {
 	e->nr_devs = 0;
 
@@ -435,7 +435,7 @@ bool bch2_bkey_replicas_marked_locked(struct bch_fs *c,
 			return false;
 	}
 
-	bkey_to_replicas(&search.e, k);
+	bch2_bkey_to_replicas(&search.e, k);
 
 	return bch2_replicas_marked_locked(c, &search.e, check_gc_replicas);
 }
@@ -468,7 +468,7 @@ int bch2_mark_bkey_replicas(struct bch_fs *c, struct bkey_s_c k)
 			return ret;
 	}
 
-	bkey_to_replicas(&search.e, k);
+	bch2_bkey_to_replicas(&search.e, k);
 
 	return bch2_mark_replicas(c, &search.e);
 }
diff --git a/fs/bcachefs/replicas.h b/fs/bcachefs/replicas.h
index ad97e3bc6b93..2ffafad7c631 100644
--- a/fs/bcachefs/replicas.h
+++ b/fs/bcachefs/replicas.h
@@ -27,6 +27,7 @@ int bch2_mark_replicas(struct bch_fs *,
 
 bool bch2_bkey_replicas_marked_locked(struct bch_fs *,
 				      struct bkey_s_c, bool);
+void bch2_bkey_to_replicas(struct bch_replicas_entry *, struct bkey_s_c);
 bool bch2_bkey_replicas_marked(struct bch_fs *,
 			       struct bkey_s_c, bool);
 int bch2_mark_bkey_replicas(struct bch_fs *, struct bkey_s_c);
diff --git a/fs/bcachefs/super-io.c b/fs/bcachefs/super-io.c
index c675280cfe03..61eefd2dd1d2 100644
--- a/fs/bcachefs/super-io.c
+++ b/fs/bcachefs/super-io.c
@@ -945,7 +945,7 @@ int bch2_fs_mark_dirty(struct bch_fs *c)
 
 	mutex_lock(&c->sb_lock);
 	SET_BCH_SB_CLEAN(c->disk_sb.sb, false);
-	c->disk_sb.sb->compat[0] &= ~(1ULL << BCH_COMPAT_FEAT_ALLOC_INFO);
+	c->disk_sb.sb->compat[0] &= ~(1ULL << BCH_COMPAT_FEAT_ALLOC_METADATA);
 	ret = bch2_write_super(c);
 	mutex_unlock(&c->sb_lock);
 
@@ -1061,6 +1061,7 @@ void bch2_fs_mark_clean(struct bch_fs *c)
 	SET_BCH_SB_CLEAN(c->disk_sb.sb, true);
 
 	c->disk_sb.sb->compat[0] |= 1ULL << BCH_COMPAT_FEAT_ALLOC_INFO;
+	c->disk_sb.sb->compat[0] |= 1ULL << BCH_COMPAT_FEAT_ALLOC_METADATA;
 
 	u64s = sizeof(*sb_clean) / sizeof(u64) + c->journal.entry_u64s_reserved;