Update bcachefs sources to 55a65a994ed5 bcachefs: bcachefs_metadata_version_persistent_inode_cursors

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

53 files changed, 874 insertions, 677 deletions

diff --git a/.bcachefs_revision b/.bcachefs_revision
index 534b898f..64016eff 100644
--- a/.bcachefs_revision
+++ b/.bcachefs_revision
@@ -1 +1 @@
-4a32728376a81dd9e75851a49159ff33602840f4
+55a65a994ed5fba038fda00f78416faf6f308bb8
diff --git a/bch_bindgen/src/bkey.rs b/bch_bindgen/src/bkey.rs
index 91d515b7..2a012508 100644
--- a/bch_bindgen/src/bkey.rs
+++ b/bch_bindgen/src/bkey.rs
@@ -50,6 +50,7 @@ pub enum BkeyValC<'a> {
     logged_op_truncate(&'a c::bch_logged_op_truncate),
     logged_op_finsert(&'a c::bch_logged_op_finsert),
     accounting(&'a c::bch_accounting),
+    inode_alloc_cursor(&'a c::bch_inode_alloc_cursor),
 }
 
 impl<'a, 'b> BkeySC<'a> {
@@ -106,6 +107,7 @@ impl<'a, 'b> BkeySC<'a> {
                 KEY_TYPE_logged_op_truncate => logged_op_truncate(transmute(self.v)),
                 KEY_TYPE_logged_op_finsert => logged_op_finsert(transmute(self.v)),
                 KEY_TYPE_accounting => accounting(transmute(self.v)),
+                KEY_TYPE_inode_alloc_cursor => inode_alloc_cursor(transmute(self.v)),
                 KEY_TYPE_MAX => unreachable!(),
             }
         }
diff --git a/include/linux/min_heap.h b/include/linux/min_heap.h
index 43a7b9dc..fe17b482 100644
--- a/include/linux/min_heap.h
+++ b/include/linux/min_heap.h
@@ -15,8 +15,8 @@
  */
 #define MIN_HEAP_PREALLOCATED(_type, _name, _nr)	\
 struct _name {	\
-	int nr;	\
-	int size;	\
+	size_t nr;	\
+	size_t size;	\
 	_type *data;	\
 	_type preallocated[_nr];	\
 }
diff --git a/libbcachefs/alloc_background.c b/libbcachefs/alloc_background.c
index b2d57045..7641a3b4 100644
--- a/libbcachefs/alloc_background.c
+++ b/libbcachefs/alloc_background.c
@@ -323,7 +323,8 @@ void bch2_alloc_v4_swab(struct bkey_s k)
 {
 	struct bch_alloc_v4 *a = bkey_s_to_alloc_v4(k).v;
 
-	a->journal_seq		= swab64(a->journal_seq);
+	a->journal_seq_nonempty	= swab64(a->journal_seq_nonempty);
+	a->journal_seq_empty	= swab64(a->journal_seq_empty);
 	a->flags		= swab32(a->flags);
 	a->dirty_sectors	= swab32(a->dirty_sectors);
 	a->cached_sectors	= swab32(a->cached_sectors);
@@ -346,16 +347,17 @@ void bch2_alloc_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c
 	prt_printf(out, "gen %u oldest_gen %u data_type ", a->gen, a->oldest_gen);
 	bch2_prt_data_type(out, a->data_type);
 	prt_newline(out);
-	prt_printf(out, "journal_seq       %llu\n",	a->journal_seq);
-	prt_printf(out, "need_discard      %llu\n",	BCH_ALLOC_V4_NEED_DISCARD(a));
-	prt_printf(out, "need_inc_gen      %llu\n",	BCH_ALLOC_V4_NEED_INC_GEN(a));
-	prt_printf(out, "dirty_sectors     %u\n",	a->dirty_sectors);
-	prt_printf(out, "stripe_sectors    %u\n",	a->stripe_sectors);
-	prt_printf(out, "cached_sectors    %u\n",	a->cached_sectors);
-	prt_printf(out, "stripe            %u\n",	a->stripe);
-	prt_printf(out, "stripe_redundancy %u\n",	a->stripe_redundancy);
-	prt_printf(out, "io_time[READ]     %llu\n",	a->io_time[READ]);
-	prt_printf(out, "io_time[WRITE]    %llu\n",	a->io_time[WRITE]);
+	prt_printf(out, "journal_seq_nonempty %llu\n",	a->journal_seq_nonempty);
+	prt_printf(out, "journal_seq_empty    %llu\n",	a->journal_seq_empty);
+	prt_printf(out, "need_discard         %llu\n",	BCH_ALLOC_V4_NEED_DISCARD(a));
+	prt_printf(out, "need_inc_gen         %llu\n",	BCH_ALLOC_V4_NEED_INC_GEN(a));
+	prt_printf(out, "dirty_sectors        %u\n",	a->dirty_sectors);
+	prt_printf(out, "stripe_sectors       %u\n",	a->stripe_sectors);
+	prt_printf(out, "cached_sectors       %u\n",	a->cached_sectors);
+	prt_printf(out, "stripe               %u\n",	a->stripe);
+	prt_printf(out, "stripe_redundancy    %u\n",	a->stripe_redundancy);
+	prt_printf(out, "io_time[READ]        %llu\n",	a->io_time[READ]);
+	prt_printf(out, "io_time[WRITE]       %llu\n",	a->io_time[WRITE]);
 
 	if (ca)
 		prt_printf(out, "fragmentation     %llu\n",	alloc_lru_idx_fragmentation(*a, ca));
@@ -384,7 +386,7 @@ void __bch2_alloc_to_v4(struct bkey_s_c k, struct bch_alloc_v4 *out)
 		struct bkey_alloc_unpacked u = bch2_alloc_unpack(k);
 
 		*out = (struct bch_alloc_v4) {
-			.journal_seq		= u.journal_seq,
+			.journal_seq_nonempty	= u.journal_seq,
 			.flags			= u.need_discard,
 			.gen			= u.gen,
 			.oldest_gen		= u.oldest_gen,
@@ -856,7 +858,10 @@ int bch2_trigger_alloc(struct btree_trans *trans,
 	if (flags & BTREE_TRIGGER_transactional) {
 		alloc_data_type_set(new_a, new_a->data_type);
 
-		if (bch2_bucket_sectors_total(*new_a) > bch2_bucket_sectors_total(*old_a)) {
+		int is_empty_delta = (int) data_type_is_empty(new_a->data_type) -
+				     (int) data_type_is_empty(old_a->data_type);
+
+		if (is_empty_delta < 0) {
 			new_a->io_time[READ] = bch2_current_io_time(c, READ);
 			new_a->io_time[WRITE]= bch2_current_io_time(c, WRITE);
 			SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, true);
@@ -927,20 +932,31 @@ int bch2_trigger_alloc(struct btree_trans *trans,
 
 	if ((flags & BTREE_TRIGGER_atomic) && (flags & BTREE_TRIGGER_insert)) {
 		u64 transaction_seq = trans->journal_res.seq;
+		BUG_ON(!transaction_seq);
+		BUG_ON(transaction_seq < new_a->journal_seq_nonempty);
+		BUG_ON(transaction_seq < new_a->journal_seq_empty);
 
-		if (log_fsck_err_on(transaction_seq && new_a->journal_seq > transaction_seq,
+		if (log_fsck_err_on(transaction_seq && new_a->journal_seq_nonempty > transaction_seq,
 				    trans, alloc_key_journal_seq_in_future,
 				    "bucket journal seq in future (currently at %llu)\n%s",
 				    journal_cur_seq(&c->journal),
 				    (bch2_bkey_val_to_text(&buf, c, new.s_c), buf.buf)))
-			new_a->journal_seq = transaction_seq;
+			new_a->journal_seq_nonempty = transaction_seq;
 
 		int is_empty_delta = (int) data_type_is_empty(new_a->data_type) -
 				     (int) data_type_is_empty(old_a->data_type);
 
-		/* Record journal sequence number of empty -> nonempty transition: */
-		if (is_empty_delta < 0)
-			new_a->journal_seq = max(new_a->journal_seq, transaction_seq);
+		/*
+		 * Record journal sequence number of empty -> nonempty transition:
+		 * Note that there may be multiple empty -> nonempty
+		 * transitions, data in a bucket may be overwritten while we're
+		 * still writing to it - so be careful to only record the first:
+		 * */
+		if (is_empty_delta < 0 &&
+		    new_a->journal_seq_empty <= c->journal.flushed_seq_ondisk) {
+			new_a->journal_seq_nonempty	= transaction_seq;
+			new_a->journal_seq_empty	= 0;
+		}
 
 		/*
 		 * Bucket becomes empty: mark it as waiting for a journal flush,
@@ -949,18 +965,21 @@ int bch2_trigger_alloc(struct btree_trans *trans,
 		 * intermediate sequence numbers:
 		 */
 		if (is_empty_delta > 0) {
-			if (new_a->journal_seq == transaction_seq ||
-			    bch2_journal_noflush_seq(&c->journal, new_a->journal_seq))
-				new_a->journal_seq = 0;
-			else {
-				new_a->journal_seq = transaction_seq;
+			if (new_a->journal_seq_nonempty == transaction_seq ||
+			    bch2_journal_noflush_seq(&c->journal,
+						     new_a->journal_seq_nonempty,
+						     transaction_seq)) {
+				new_a->journal_seq_nonempty = new_a->journal_seq_empty = 0;
+			} else {
+				new_a->journal_seq_empty = transaction_seq;
 
 				ret = bch2_set_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
-						c->journal.flushed_seq_ondisk,
-						new.k->p.inode, new.k->p.offset,
-						transaction_seq);
+									   c->journal.flushed_seq_ondisk,
+									   new.k->p.inode, new.k->p.offset,
+									   transaction_seq);
 				if (bch2_fs_fatal_err_on(ret, c,
-						"setting bucket_needs_journal_commit: %s", bch2_err_str(ret)))
+						"setting bucket_needs_journal_commit: %s",
+						bch2_err_str(ret)))
 					goto err;
 			}
 		}
@@ -978,7 +997,7 @@ int bch2_trigger_alloc(struct btree_trans *trans,
 
 #define eval_state(_a, expr)		({ const struct bch_alloc_v4 *a = _a; expr; })
 #define statechange(expr)		!eval_state(old_a, expr) && eval_state(new_a, expr)
-#define bucket_flushed(a)		(!a->journal_seq || a->journal_seq <= c->journal.flushed_seq_ondisk)
+#define bucket_flushed(a)		(a->journal_seq_empty <= c->journal.flushed_seq_ondisk)
 
 		if (statechange(a->data_type == BCH_DATA_free) &&
 		    bucket_flushed(new_a))
@@ -1840,16 +1859,6 @@ static int bch2_discard_one_bucket(struct btree_trans *trans,
 		goto out;
 	}
 
-	if (a->v.journal_seq > c->journal.flushed_seq_ondisk) {
-		if (bch2_trans_inconsistent_on(c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_info,
-					       trans, "clearing need_discard but journal_seq %llu > flushed_seq %llu\n%s",
-					       a->v.journal_seq,
-					       c->journal.flushed_seq_ondisk,
-					       (bch2_bkey_val_to_text(&buf, c, k), buf.buf)))
-			ret = -EIO;
-		goto out;
-	}
-
 	if (!fastpath) {
 		if (discard_in_flight_add(ca, iter.pos.offset, true))
 			goto out;
diff --git a/libbcachefs/alloc_background_format.h b/libbcachefs/alloc_background_format.h
index befdaa95..74023836 100644
--- a/libbcachefs/alloc_background_format.h
+++ b/libbcachefs/alloc_background_format.h
@@ -58,7 +58,7 @@ LE32_BITMASK(BCH_ALLOC_V3_NEED_INC_GEN,struct bch_alloc_v3, flags,  1,  2)
 
 struct bch_alloc_v4 {
 	struct bch_val		v;
-	__u64			journal_seq;
+	__u64			journal_seq_nonempty;
 	__u32			flags;
 	__u8			gen;
 	__u8			oldest_gen;
@@ -70,7 +70,7 @@ struct bch_alloc_v4 {
 	__u32			stripe;
 	__u32			nr_external_backpointers;
 	/* end of fields in original version of alloc_v4 */
-	__u64			_fragmentation_lru; /* obsolete */
+	__u64			journal_seq_empty;
 	__u32			stripe_sectors;
 	__u32			pad;
 } __packed __aligned(8);
diff --git a/libbcachefs/alloc_foreground.c b/libbcachefs/alloc_foreground.c
index 095bfe7c..57d5f14c 100644
--- a/libbcachefs/alloc_foreground.c
+++ b/libbcachefs/alloc_foreground.c
@@ -200,14 +200,35 @@ static inline unsigned open_buckets_reserved(enum bch_watermark watermark)
 	}
 }
 
+static inline bool may_alloc_bucket(struct bch_fs *c,
+				    struct bpos bucket,
+				    struct bucket_alloc_state *s)
+{
+	if (bch2_bucket_is_open(c, bucket.inode, bucket.offset)) {
+		s->skipped_open++;
+		return false;
+	}
+
+	if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
+			c->journal.flushed_seq_ondisk, bucket.inode, bucket.offset)) {
+		s->skipped_need_journal_commit++;
+		return false;
+	}
+
+	if (bch2_bucket_nocow_is_locked(&c->nocow_locks, bucket)) {
+		s->skipped_nocow++;
+		return false;
+	}
+
+	return true;
+}
+
 static struct open_bucket *__try_alloc_bucket(struct bch_fs *c, struct bch_dev *ca,
 					      u64 bucket, u8 gen,
 					      enum bch_watermark watermark,
 					      struct bucket_alloc_state *s,
 					      struct closure *cl)
 {
-	struct open_bucket *ob;
-
 	if (unlikely(is_superblock_bucket(c, ca, bucket)))
 		return NULL;
 
@@ -216,22 +237,6 @@ static struct open_bucket *__try_alloc_bucket(struct bch_fs *c, struct bch_dev *
 		return NULL;
 	}
 
-	if (bch2_bucket_is_open(c, ca->dev_idx, bucket)) {
-		s->skipped_open++;
-		return NULL;
-	}
-
-	if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
-			c->journal.flushed_seq_ondisk, ca->dev_idx, bucket)) {
-		s->skipped_need_journal_commit++;
-		return NULL;
-	}
-
-	if (bch2_bucket_nocow_is_locked(&c->nocow_locks, POS(ca->dev_idx, bucket))) {
-		s->skipped_nocow++;
-		return NULL;
-	}
-
 	spin_lock(&c->freelist_lock);
 
 	if (unlikely(c->open_buckets_nr_free <= open_buckets_reserved(watermark))) {
@@ -250,10 +255,9 @@ static struct open_bucket *__try_alloc_bucket(struct bch_fs *c, struct bch_dev *
 		return NULL;
 	}
 
-	ob = bch2_open_bucket_alloc(c);
+	struct open_bucket *ob = bch2_open_bucket_alloc(c);
 
 	spin_lock(&ob->lock);
-
 	ob->valid	= true;
 	ob->sectors_free = ca->mi.bucket_size;
 	ob->dev		= ca->dev_idx;
@@ -279,8 +283,11 @@ static struct open_bucket *try_alloc_bucket(struct btree_trans *trans, struct bc
 {
 	struct bch_fs *c = trans->c;
 	u64 b = freespace_iter->pos.offset & ~(~0ULL << 56);
-	u8 gen;
 
+	if (!may_alloc_bucket(c, POS(ca->dev_idx, b), s))
+		return NULL;
+
+	u8 gen;
 	int ret = bch2_check_discard_freespace_key(trans, freespace_iter, &gen, true);
 	if (ret < 0)
 		return ERR_PTR(ret);
@@ -300,6 +307,7 @@ bch2_bucket_alloc_early(struct btree_trans *trans,
 			struct bucket_alloc_state *s,
 			struct closure *cl)
 {
+	struct bch_fs *c = trans->c;
 	struct btree_iter iter, citer;
 	struct bkey_s_c k, ck;
 	struct open_bucket *ob = NULL;
@@ -359,7 +367,10 @@ again:
 
 		s->buckets_seen++;
 
-		ob = __try_alloc_bucket(trans->c, ca, k.k->p.offset, a->gen, watermark, s, cl);
+		ob = may_alloc_bucket(c, k.k->p, s)
+			? __try_alloc_bucket(c, ca, k.k->p.offset, a->gen,
+					     watermark, s, cl)
+			: NULL;
 next:
 		bch2_set_btree_iter_dontneed(&citer);
 		bch2_trans_iter_exit(trans, &citer);
@@ -626,9 +637,9 @@ struct dev_alloc_list bch2_dev_alloc_list(struct bch_fs *c,
 	unsigned i;
 
 	for_each_set_bit(i, devs->d, BCH_SB_MEMBERS_MAX)
-		ret.devs[ret.nr++] = i;
+		ret.data[ret.nr++] = i;
 
-	bubble_sort(ret.devs, ret.nr, dev_stripe_cmp);
+	bubble_sort(ret.data, ret.nr, dev_stripe_cmp);
 	return ret;
 }
 
@@ -700,18 +711,13 @@ int bch2_bucket_alloc_set_trans(struct btree_trans *trans,
 		      struct closure *cl)
 {
 	struct bch_fs *c = trans->c;
-	struct dev_alloc_list devs_sorted =
-		bch2_dev_alloc_list(c, stripe, devs_may_alloc);
 	int ret = -BCH_ERR_insufficient_devices;
 
 	BUG_ON(*nr_effective >= nr_replicas);
 
-	for (unsigned i = 0; i < devs_sorted.nr; i++) {
-		struct bch_dev_usage usage;
-		struct open_bucket *ob;
-
-		unsigned dev = devs_sorted.devs[i];
-		struct bch_dev *ca = bch2_dev_tryget_noerror(c, dev);
+	struct dev_alloc_list devs_sorted = bch2_dev_alloc_list(c, stripe, devs_may_alloc);
+	darray_for_each(devs_sorted, i) {
+		struct bch_dev *ca = bch2_dev_tryget_noerror(c, *i);
 		if (!ca)
 			continue;
 
@@ -720,8 +726,9 @@ int bch2_bucket_alloc_set_trans(struct btree_trans *trans,
 			continue;
 		}
 
-		ob = bch2_bucket_alloc_trans(trans, ca, watermark, data_type,
-					     cl, flags & BCH_WRITE_ALLOC_NOWAIT, &usage);
+		struct bch_dev_usage usage;
+		struct open_bucket *ob = bch2_bucket_alloc_trans(trans, ca, watermark, data_type,
+						     cl, flags & BCH_WRITE_ALLOC_NOWAIT, &usage);
 		if (!IS_ERR(ob))
 			bch2_dev_stripe_increment_inlined(ca, stripe, &usage);
 		bch2_dev_put(ca);
@@ -765,10 +772,6 @@ static int bucket_alloc_from_stripe(struct btree_trans *trans,
 			 struct closure *cl)
 {
 	struct bch_fs *c = trans->c;
-	struct dev_alloc_list devs_sorted;
-	struct ec_stripe_head *h;
-	struct open_bucket *ob;
-	unsigned i, ec_idx;
 	int ret = 0;
 
 	if (nr_replicas < 2)
@@ -777,34 +780,32 @@ static int bucket_alloc_from_stripe(struct btree_trans *trans,
 	if (ec_open_bucket(c, ptrs))
 		return 0;
 
-	h = bch2_ec_stripe_head_get(trans, target, 0, nr_replicas - 1, watermark, cl);
+	struct ec_stripe_head *h =
+		bch2_ec_stripe_head_get(trans, target, 0, nr_replicas - 1, watermark, cl);
 	if (IS_ERR(h))
 		return PTR_ERR(h);
 	if (!h)
 		return 0;
 
-	devs_sorted = bch2_dev_alloc_list(c, &wp->stripe, devs_may_alloc);
-
-	for (i = 0; i < devs_sorted.nr; i++)
-		for (ec_idx = 0; ec_idx < h->s->nr_data; ec_idx++) {
+	struct dev_alloc_list devs_sorted = bch2_dev_alloc_list(c, &wp->stripe, devs_may_alloc);
+	darray_for_each(devs_sorted, i)
+		for (unsigned ec_idx = 0; ec_idx < h->s->nr_data; ec_idx++) {
 			if (!h->s->blocks[ec_idx])
 				continue;
 
-			ob = c->open_buckets + h->s->blocks[ec_idx];
-			if (ob->dev == devs_sorted.devs[i] &&
-			    !test_and_set_bit(ec_idx, h->s->blocks_allocated))
-				goto got_bucket;
+			struct open_bucket *ob = c->open_buckets + h->s->blocks[ec_idx];
+			if (ob->dev == *i && !test_and_set_bit(ec_idx, h->s->blocks_allocated)) {
+				ob->ec_idx	= ec_idx;
+				ob->ec		= h->s;
+				ec_stripe_new_get(h->s, STRIPE_REF_io);
+
+				ret = add_new_bucket(c, ptrs, devs_may_alloc,
+						     nr_replicas, nr_effective,
+						     have_cache, ob);
+				goto out;
+			}
 		}
-	goto out_put_head;
-got_bucket:
-	ob->ec_idx	= ec_idx;
-	ob->ec		= h->s;
-	ec_stripe_new_get(h->s, STRIPE_REF_io);
-
-	ret = add_new_bucket(c, ptrs, devs_may_alloc,
-			     nr_replicas, nr_effective,
-			     have_cache, ob);
-out_put_head:
+out:
 	bch2_ec_stripe_head_put(c, h);
 	return ret;
 }
diff --git a/libbcachefs/alloc_foreground.h b/libbcachefs/alloc_foreground.h
index 4f87745d..f25481a0 100644
--- a/libbcachefs/alloc_foreground.h
+++ b/libbcachefs/alloc_foreground.h
@@ -20,7 +20,7 @@ void bch2_reset_alloc_cursors(struct bch_fs *);
 
 struct dev_alloc_list {
 	unsigned	nr;
-	u8		devs[BCH_SB_MEMBERS_MAX];
+	u8		data[BCH_SB_MEMBERS_MAX];
 };
 
 struct dev_alloc_list bch2_dev_alloc_list(struct bch_fs *,
diff --git a/libbcachefs/backpointers.c b/libbcachefs/backpointers.c
index 72f37171..72311dc7 100644
--- a/libbcachefs/backpointers.c
+++ b/libbcachefs/backpointers.c
@@ -42,16 +42,17 @@ void bch2_backpointer_to_text(struct printbuf *out, struct bch_fs *c, struct bke
 		u32 bucket_offset;
 		struct bpos bucket = bp_pos_to_bucket_and_offset(ca, bp.k->p, &bucket_offset);
 		rcu_read_unlock();
-		prt_printf(out, "bucket=%llu:%llu:%u", bucket.inode, bucket.offset, bucket_offset);
+		prt_printf(out, "bucket=%llu:%llu:%u ", bucket.inode, bucket.offset, bucket_offset);
 	} else {
 		rcu_read_unlock();
-		prt_printf(out, "sector=%llu:%llu", bp.k->p.inode, bp.k->p.offset >> MAX_EXTENT_COMPRESS_RATIO_SHIFT);
+		prt_printf(out, "sector=%llu:%llu ", bp.k->p.inode, bp.k->p.offset >> MAX_EXTENT_COMPRESS_RATIO_SHIFT);
 	}
 
 	bch2_btree_id_level_to_text(out, bp.v->btree_id, bp.v->level);
-	prt_printf(out, " suboffset=%u len=%u pos=",
+	prt_printf(out, " suboffset=%u len=%u gen=%u pos=",
 		   (u32) bp.k->p.offset & ~(~0U << MAX_EXTENT_COMPRESS_RATIO_SHIFT),
-		   bp.v->bucket_len);
+		   bp.v->bucket_len,
+		   bp.v->bucket_gen);
 	bch2_bpos_to_text(out, bp.v->pos);
 }
 
diff --git a/libbcachefs/bcachefs.h b/libbcachefs/bcachefs.h
index 58d42b3f..d27550ef 100644
--- a/libbcachefs/bcachefs.h
+++ b/libbcachefs/bcachefs.h
@@ -1048,6 +1048,7 @@ struct bch_fs {
 	 * for signaling to the toplevel code which pass we want to run now.
 	 */
 	enum bch_recovery_pass	curr_recovery_pass;
+	enum bch_recovery_pass	next_recovery_pass;
 	/* bitmask of recovery passes that we actually ran */
 	u64			recovery_passes_complete;
 	/* never rewinds version of curr_recovery_pass */
@@ -1063,9 +1064,6 @@ struct bch_fs {
 	struct btree_node	*verify_ondisk;
 	struct mutex		verify_lock;
 
-	u64			*unused_inode_hints;
-	unsigned		inode_shard_bits;
-
 	/*
 	 * A btree node on disk could have too many bsets for an iterator to fit
 	 * on the stack - have to dynamically allocate them
diff --git a/libbcachefs/bcachefs_format.h b/libbcachefs/bcachefs_format.h
index f140c336..09e53bef 100644
--- a/libbcachefs/bcachefs_format.h
+++ b/libbcachefs/bcachefs_format.h
@@ -418,7 +418,8 @@ static inline void bkey_init(struct bkey *k)
 	x(snapshot_tree,	31)			\
 	x(logged_op_truncate,	32)			\
 	x(logged_op_finsert,	33)			\
-	x(accounting,		34)
+	x(accounting,		34)			\
+	x(inode_alloc_cursor,	35)
 
 enum bch_bkey_type {
 #define x(name, nr) KEY_TYPE_##name	= nr,
@@ -682,7 +683,8 @@ struct bch_sb_field_ext {
 	x(backpointer_bucket_gen,	BCH_VERSION(1, 14))		\
 	x(disk_accounting_big_endian,	BCH_VERSION(1, 15))		\
 	x(reflink_p_may_update_opts,	BCH_VERSION(1, 16))		\
-	x(inode_depth,			BCH_VERSION(1, 17))
+	x(inode_depth,			BCH_VERSION(1, 17))		\
+	x(persistent_inode_cursors,	BCH_VERSION(1, 18))
 
 enum bcachefs_metadata_version {
 	bcachefs_metadata_version_min = 9,
@@ -850,6 +852,7 @@ LE64_BITMASK(BCH_SB_ALLOCATOR_STUCK_TIMEOUT,
 LE64_BITMASK(BCH_SB_VERSION_INCOMPAT,	struct bch_sb, flags[5], 32, 48);
 LE64_BITMASK(BCH_SB_VERSION_INCOMPAT_ALLOWED,
 					struct bch_sb, flags[5], 48, 64);
+LE64_BITMASK(BCH_SB_SHARD_INUMS_NBITS,	struct bch_sb, flags[6],  0,  4);
 
 static inline __u64 BCH_SB_COMPRESSION_TYPE(const struct bch_sb *sb)
 {
@@ -1347,7 +1350,8 @@ enum btree_id_flags {
 	  BIT_ULL(KEY_TYPE_set))						\
 	x(logged_ops,		17,	0,					\
 	  BIT_ULL(KEY_TYPE_logged_op_truncate)|					\
-	  BIT_ULL(KEY_TYPE_logged_op_finsert))					\
+	  BIT_ULL(KEY_TYPE_logged_op_finsert)|					\
+	  BIT_ULL(KEY_TYPE_inode_alloc_cursor))					\
 	x(rebalance_work,	18,	BTREE_ID_SNAPSHOT_FIELD,		\
 	  BIT_ULL(KEY_TYPE_set)|BIT_ULL(KEY_TYPE_cookie))			\
 	x(subvolume_children,	19,	0,					\
diff --git a/libbcachefs/bkey_types.h b/libbcachefs/bkey_types.h
index 2af6279b..b4f328f9 100644
--- a/libbcachefs/bkey_types.h
+++ b/libbcachefs/bkey_types.h
@@ -213,16 +213,16 @@ BCH_BKEY_TYPES();
 enum bch_validate_flags {
 	BCH_VALIDATE_write		= BIT(0),
 	BCH_VALIDATE_commit		= BIT(1),
-	BCH_VALIDATE_journal		= BIT(2),
-	BCH_VALIDATE_silent		= BIT(3),
+	BCH_VALIDATE_silent		= BIT(2),
 };
 
 #define BKEY_VALIDATE_CONTEXTS()	\
 	x(unknown)			\
-	x(commit)			\
+	x(superblock)			\
 	x(journal)			\
 	x(btree_root)			\
-	x(btree_node)
+	x(btree_node)			\
+	x(commit)
 
 struct bkey_validate_context {
 	enum {
@@ -230,10 +230,12 @@ struct bkey_validate_context {
 	BKEY_VALIDATE_CONTEXTS()
 #undef x
 	}			from:8;
+	enum bch_validate_flags	flags:8;
 	u8			level;
 	enum btree_id		btree;
 	bool			root:1;
-	enum bch_validate_flags	flags:8;
+	unsigned		journal_offset;
+	u64			journal_seq;
 };
 
 #endif /* _BCACHEFS_BKEY_TYPES_H */
diff --git a/libbcachefs/btree_gc.c b/libbcachefs/btree_gc.c
index e59924cf..24f2f3bd 100644
--- a/libbcachefs/btree_gc.c
+++ b/libbcachefs/btree_gc.c
@@ -29,6 +29,7 @@
 #include "move.h"
 #include "recovery_passes.h"
 #include "reflink.h"
+#include "recovery.h"
 #include "replicas.h"
 #include "super-io.h"
 #include "trace.h"
@@ -359,11 +360,9 @@ again:
 			if (ret)
 				break;
 
-			if (!btree_id_is_alloc(b->c.btree_id)) {
-				ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_scan_for_btree_nodes);
-				if (ret)
-					break;
-			}
+			ret = bch2_btree_lost_data(c, b->c.btree_id);
+			if (ret)
+				break;
 			continue;
 		}
 
@@ -525,7 +524,7 @@ int bch2_check_topology(struct bch_fs *c)
 		bch2_btree_id_to_text(&buf, i);
 
 		if (r->error) {
-			ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_scan_for_btree_nodes);
+			ret = bch2_btree_lost_data(c, i);
 			if (ret)
 				break;
 reconstruct_root:
@@ -741,7 +740,7 @@ static int bch2_gc_btrees(struct bch_fs *c)
 			       (printbuf_reset(&buf),
 				bch2_btree_id_to_text(&buf, btree),
 				buf.buf)))
-			ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_topology);
+			ret = bch2_btree_lost_data(c, btree);
 	}
 fsck_err:
 	printbuf_exit(&buf);
diff --git a/libbcachefs/btree_io.c b/libbcachefs/btree_io.c
index 9df9fc1c..d99f8a78 100644
--- a/libbcachefs/btree_io.c
+++ b/libbcachefs/btree_io.c
@@ -26,7 +26,7 @@
 static void bch2_btree_node_header_to_text(struct printbuf *out, struct btree_node *bn)
 {
 	bch2_btree_id_level_to_text(out, BTREE_NODE_ID(bn), BTREE_NODE_LEVEL(bn));
-	prt_printf(out, " seq %llux\n", bn->keys.seq);
+	prt_printf(out, " seq %llx %llu\n", bn->keys.seq, BTREE_NODE_SEQ(bn));
 	prt_str(out, "min: ");
 	bch2_bpos_to_text(out, bn->min_key);
 	prt_newline(out);
diff --git a/libbcachefs/btree_node_scan.c b/libbcachefs/btree_node_scan.c
index eeafb5e7..a5282ff1 100644
--- a/libbcachefs/btree_node_scan.c
+++ b/libbcachefs/btree_node_scan.c
@@ -12,6 +12,7 @@
 #include "recovery_passes.h"
 
 #include <linux/kthread.h>
+#include <linux/min_heap.h>
 #include <linux/sort.h>
 
 struct find_btree_nodes_worker {
@@ -31,8 +32,6 @@ static void found_btree_node_to_text(struct printbuf *out, struct bch_fs *c, con
 
 	if (n->range_updated)
 		prt_str(out, " range updated");
-	if (n->overwritten)
-		prt_str(out, " overwritten");
 
 	for (unsigned i = 0; i < n->nr_ptrs; i++) {
 		prt_char(out, ' ');
@@ -140,6 +139,24 @@ static int found_btree_node_cmp_pos(const void *_l, const void *_r)
 	       -found_btree_node_cmp_time(l, r);
 }
 
+static inline bool found_btree_node_cmp_pos_less(const void *l, const void *r, void *arg)
+{
+	return found_btree_node_cmp_pos(l, r) < 0;
+}
+
+static inline void found_btree_node_swap(void *_l, void *_r, void *arg)
+{
+	struct found_btree_node *l = _l;
+	struct found_btree_node *r = _r;
+
+	swap(*l, *r);
+}
+
+const struct min_heap_callbacks found_btree_node_heap_cbs = {
+	.less = found_btree_node_cmp_pos_less,
+	.swp = found_btree_node_swap,
+};
+
 static void try_read_btree_node(struct find_btree_nodes *f, struct bch_dev *ca,
 				struct bio *bio, struct btree_node *bn, u64 offset)
 {
@@ -295,55 +312,48 @@ err:
 	return f->ret ?: ret;
 }
 
-static void bubble_up(struct found_btree_node *n, struct found_btree_node *end)
+static bool nodes_overlap(const struct found_btree_node *l,
+			  const struct found_btree_node *r)
 {
-	while (n + 1 < end &&
-	       found_btree_node_cmp_pos(n, n + 1) > 0) {
-		swap(n[0], n[1]);
-		n++;
-	}
+	return (l->btree_id	== r->btree_id &&
+		l->level	== r->level &&
+		bpos_gt(l->max_key, r->min_key));
 }
 
 static int handle_overwrites(struct bch_fs *c,
-			     struct found_btree_node *start,
-			     struct found_btree_node *end)
+			     struct found_btree_node *l,
+			     found_btree_nodes *nodes_heap)
 {
-	struct found_btree_node *n;
-again:
-	for (n = start + 1;
-	     n < end &&
-	     n->btree_id	== start->btree_id &&
-	     n->level		== start->level &&
-	     bpos_lt(n->min_key, start->max_key);
-	     n++)  {
-		int cmp = found_btree_node_cmp_time(start, n);
+	struct found_btree_node *r;
+
+	while ((r = min_heap_peek(nodes_heap)) &&
+	       nodes_overlap(l, r)) {
+		int cmp = found_btree_node_cmp_time(l, r);
 
 		if (cmp > 0) {
-			if (bpos_cmp(start->max_key, n->max_key) >= 0)
-				n->overwritten = true;
+			if (bpos_cmp(l->max_key, r->max_key) >= 0)
+				min_heap_pop(nodes_heap, &found_btree_node_heap_cbs, NULL);
 			else {
-				n->range_updated = true;
-				n->min_key = bpos_successor(start->max_key);
-				n->range_updated = true;
-				bubble_up(n, end);
-				goto again;
+				r->range_updated = true;
+				r->min_key = bpos_successor(l->max_key);
+				r->range_updated = true;
+				min_heap_sift_down(nodes_heap, 0, &found_btree_node_heap_cbs, NULL);
 			}
 		} else if (cmp < 0) {
-			BUG_ON(bpos_cmp(n->min_key, start->min_key) <= 0);
+			BUG_ON(bpos_eq(l->min_key, r->min_key));
 
-			start->max_key = bpos_predecessor(n->min_key);
-			start->range_updated = true;
-		} else if (n->level) {
-			n->overwritten = true;
+			l->max_key = bpos_predecessor(r->min_key);
+			l->range_updated = true;
+		} else if (r->level) {
+			min_heap_pop(nodes_heap, &found_btree_node_heap_cbs, NULL);
 		} else {
-			if (bpos_cmp(start->max_key, n->max_key) >= 0)
-				n->overwritten = true;
+			if (bpos_cmp(l->max_key, r->max_key) >= 0)
+				min_heap_pop(nodes_heap, &found_btree_node_heap_cbs, NULL);
 			else {
-				n->range_updated = true;
-				n->min_key = bpos_successor(start->max_key);
-				n->range_updated = true;
-				bubble_up(n, end);
-				goto again;
+				r->range_updated = true;
+				r->min_key = bpos_successor(l->max_key);
+				r->range_updated = true;
+				min_heap_sift_down(nodes_heap, 0, &found_btree_node_heap_cbs, NULL);
 			}
 		}
 	}
@@ -355,6 +365,7 @@ int bch2_scan_for_btree_nodes(struct bch_fs *c)
 {
 	struct find_btree_nodes *f = &c->found_btree_nodes;
 	struct printbuf buf = PRINTBUF;
+	found_btree_nodes nodes_heap = {};
 	size_t dst;
 	int ret = 0;
 
@@ -409,29 +420,57 @@ int bch2_scan_for_btree_nodes(struct bch_fs *c)
 		bch2_print_string_as_lines(KERN_INFO, buf.buf);
 	}
 
-	dst = 0;
-	darray_for_each(f->nodes, i) {
-		if (i->overwritten)
-			continue;
+	swap(nodes_heap, f->nodes);
+
+	{
+		/* darray must have same layout as a heap */
+		min_heap_char real_heap;
+		BUILD_BUG_ON(sizeof(nodes_heap.nr)	!= sizeof(real_heap.nr));
+		BUILD_BUG_ON(sizeof(nodes_heap.size)	!= sizeof(real_heap.size));
+		BUILD_BUG_ON(offsetof(found_btree_nodes, nr)	!= offsetof(min_heap_char, nr));
+		BUILD_BUG_ON(offsetof(found_btree_nodes, size)	!= offsetof(min_heap_char, size));
+	}
 
-		ret = handle_overwrites(c, i, &darray_top(f->nodes));
+	min_heapify_all(&nodes_heap, &found_btree_node_heap_cbs, NULL);
+
+	if (nodes_heap.nr) {
+		ret = darray_push(&f->nodes, *min_heap_peek(&nodes_heap));
 		if (ret)
 			goto err;
 
-		BUG_ON(i->overwritten);
-		f->nodes.data[dst++] = *i;
+		min_heap_pop(&nodes_heap, &found_btree_node_heap_cbs, NULL);
 	}
-	f->nodes.nr = dst;
 
-	if (c->opts.verbose) {
+	while (true) {
+		ret = handle_overwrites(c, &darray_last(f->nodes), &nodes_heap);
+		if (ret)
+			goto err;
+
+		if (!nodes_heap.nr)
+			break;
+
+		ret = darray_push(&f->nodes, *min_heap_peek(&nodes_heap));
+		if (ret)
+			goto err;
+
+		min_heap_pop(&nodes_heap, &found_btree_node_heap_cbs, NULL);
+	}
+
+	for (struct found_btree_node *n = f->nodes.data; n < &darray_last(f->nodes); n++)
+		BUG_ON(nodes_overlap(n, n + 1));
+
+	if (0 && c->opts.verbose) {
 		printbuf_reset(&buf);
 		prt_printf(&buf, "%s: nodes found after overwrites:\n", __func__);
 		found_btree_nodes_to_text(&buf, c, f->nodes);
 		bch2_print_string_as_lines(KERN_INFO, buf.buf);
+	} else {
+		bch_info(c, "btree node scan found %zu nodes after overwrites", f->nodes.nr);
 	}
 
 	eytzinger0_sort(f->nodes.data, f->nodes.nr, sizeof(f->nodes.data[0]), found_btree_node_cmp_pos, NULL);
 err:
+	darray_exit(&nodes_heap);
 	printbuf_exit(&buf);
 	return ret;
 }
diff --git a/libbcachefs/btree_node_scan_types.h b/libbcachefs/btree_node_scan_types.h
index b6c36c45..2811b685 100644
--- a/libbcachefs/btree_node_scan_types.h
+++ b/libbcachefs/btree_node_scan_types.h
@@ -6,7 +6,6 @@
 
 struct found_btree_node {
 	bool			range_updated:1;
-	bool			overwritten:1;
 	u8			btree_id;
 	u8			level;
 	unsigned		sectors_written;
diff --git a/libbcachefs/btree_trans_commit.c b/libbcachefs/btree_trans_commit.c
index 78d72c26..9011cc3f 100644
--- a/libbcachefs/btree_trans_commit.c
+++ b/libbcachefs/btree_trans_commit.c
@@ -719,38 +719,17 @@ bch2_trans_commit_write_locked(struct btree_trans *trans, unsigned flags,
 			goto fatal_err;
 	}
 
-	trans_for_each_update(trans, i) {
-		enum bch_validate_flags invalid_flags = 0;
-
-		if (!(flags & BCH_TRANS_COMMIT_no_journal_res))
-			invalid_flags |= BCH_VALIDATE_write|BCH_VALIDATE_commit;
-
-		ret = bch2_bkey_validate(c, bkey_i_to_s_c(i->k),
-					 (struct bkey_validate_context) {
-						.from	= BKEY_VALIDATE_commit,
-						.level	= i->level,
-						.btree	= i->btree_id,
-						.flags	= invalid_flags,
-					 });
-		if (unlikely(ret)){
-			bch2_trans_inconsistent(trans, "invalid bkey on insert from %s -> %ps\n",
-						trans->fn, (void *) i->ip_allocated);
-			goto fatal_err;
-		}
-		btree_insert_entry_checks(trans, i);
-	}
+	struct bkey_validate_context validate_context = { .from	= BKEY_VALIDATE_commit };
+
+	if (!(flags & BCH_TRANS_COMMIT_no_journal_res))
+		validate_context.flags = BCH_VALIDATE_write|BCH_VALIDATE_commit;
 
 	for (struct jset_entry *i = trans->journal_entries;
 	     i != (void *) ((u64 *) trans->journal_entries + trans->journal_entries_u64s);
 	     i = vstruct_next(i)) {
-		enum bch_validate_flags invalid_flags = 0;
-
-		if (!(flags & BCH_TRANS_COMMIT_no_journal_res))
-			invalid_flags |= BCH_VALIDATE_write|BCH_VALIDATE_commit;
-
 		ret = bch2_journal_entry_validate(c, NULL, i,
 						  bcachefs_metadata_version_current,
-						  CPU_BIG_ENDIAN, invalid_flags);
+						  CPU_BIG_ENDIAN, validate_context);
 		if (unlikely(ret)) {
 			bch2_trans_inconsistent(trans, "invalid journal entry on insert from %s\n",
 						trans->fn);
@@ -758,6 +737,19 @@ bch2_trans_commit_write_locked(struct btree_trans *trans, unsigned flags,
 		}
 	}
 
+	trans_for_each_update(trans, i) {
+		validate_context.level	= i->level;
+		validate_context.btree	= i->btree_id;
+
+		ret = bch2_bkey_validate(c, bkey_i_to_s_c(i->k), validate_context);
+		if (unlikely(ret)){
+			bch2_trans_inconsistent(trans, "invalid bkey on insert from %s -> %ps\n",
+						trans->fn, (void *) i->ip_allocated);
+			goto fatal_err;
+		}
+		btree_insert_entry_checks(trans, i);
+	}
+
 	if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res))) {
 		struct journal *j = &c->journal;
 		struct jset_entry *entry;
diff --git a/libbcachefs/btree_update.c b/libbcachefs/btree_update.c
index 06fd5aa6..b1da6dba 100644
--- a/libbcachefs/btree_update.c
+++ b/libbcachefs/btree_update.c
@@ -588,7 +588,7 @@ struct jset_entry *__bch2_trans_jset_entry_alloc(struct btree_trans *trans, unsi
 int bch2_bkey_get_empty_slot(struct btree_trans *trans, struct btree_iter *iter,
 			     enum btree_id btree, struct bpos end)
 {
-	bch2_trans_iter_init(trans, iter, btree, POS_MAX, BTREE_ITER_intent);
+	bch2_trans_iter_init(trans, iter, btree, end, BTREE_ITER_intent);
 	struct bkey_s_c k = bch2_btree_iter_peek_prev(iter);
 	int ret = bkey_err(k);
 	if (ret)
diff --git a/libbcachefs/btree_update_interior.c b/libbcachefs/btree_update_interior.c
index f2a1d5d3..7d9dab95 100644
--- a/libbcachefs/btree_update_interior.c
+++ b/libbcachefs/btree_update_interior.c
@@ -1607,8 +1607,6 @@ static int btree_split(struct btree_update *as, struct btree_trans *trans,
 	if (ret)
 		return ret;
 
-	bch2_btree_interior_update_will_free_node(as, b);
-
 	if (b->nr.live_u64s > BTREE_SPLIT_THRESHOLD(c)) {
 		struct btree *n[2];
 
@@ -1707,6 +1705,8 @@ static int btree_split(struct btree_update *as, struct btree_trans *trans,
 	if (ret)
 		goto err;
 
+	bch2_btree_interior_update_will_free_node(as, b);
+
 	if (n3) {
 		bch2_btree_update_get_open_buckets(as, n3);
 		bch2_btree_node_write(c, n3, SIX_LOCK_intent, 0);
@@ -2063,9 +2063,6 @@ int __bch2_foreground_maybe_merge(struct btree_trans *trans,
 
 	trace_and_count(c, btree_node_merge, trans, b);
 
-	bch2_btree_interior_update_will_free_node(as, b);
-	bch2_btree_interior_update_will_free_node(as, m);
-
 	n = bch2_btree_node_alloc(as, trans, b->c.level);
 
 	SET_BTREE_NODE_SEQ(n->data,
@@ -2101,6 +2098,9 @@ int __bch2_foreground_maybe_merge(struct btree_trans *trans,
 	if (ret)
 		goto err_free_update;
 
+	bch2_btree_interior_update_will_free_node(as, b);
+	bch2_btree_interior_update_will_free_node(as, m);
+
 	bch2_trans_verify_paths(trans);
 
 	bch2_btree_update_get_open_buckets(as, n);
@@ -2155,8 +2155,6 @@ int bch2_btree_node_rewrite(struct btree_trans *trans,
 	if (ret)
 		goto out;
 
-	bch2_btree_interior_update_will_free_node(as, b);
-
 	n = bch2_btree_node_alloc_replacement(as, trans, b);
 
 	bch2_btree_build_aux_trees(n);
@@ -2180,6 +2178,8 @@ int bch2_btree_node_rewrite(struct btree_trans *trans,
 	if (ret)
 		goto err;
 
+	bch2_btree_interior_update_will_free_node(as, b);
+
 	bch2_btree_update_get_open_buckets(as, n);
 	bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
 
diff --git a/libbcachefs/darray.h b/libbcachefs/darray.h
index 8f4c3f06..c6151495 100644
--- a/libbcachefs/darray.h
+++ b/libbcachefs/darray.h
@@ -83,7 +83,7 @@ int __bch2_darray_resize_noprof(darray_char *, size_t, size_t, gfp_t);
 	for (typeof(&(_d).data[0]) _i = (_d).data; _i < (_d).data + (_d).nr; _i++)
 
 #define darray_for_each_reverse(_d, _i)					\
-	for (typeof(&(_d).data[0]) _i = (_d).data + (_d).nr - 1; _i >= (_d).data; --_i)
+	for (typeof(&(_d).data[0]) _i = (_d).data + (_d).nr - 1; _i >= (_d).data && (_d).nr; --_i)
 
 #define darray_init(_d)							\
 do {									\
diff --git a/libbcachefs/dirent.c b/libbcachefs/dirent.c
index 41813f9c..600eee93 100644
--- a/libbcachefs/dirent.c
+++ b/libbcachefs/dirent.c
@@ -266,7 +266,7 @@ int bch2_dirent_read_target(struct btree_trans *trans, subvol_inum dir,
 	} else {
 		target->subvol	= le32_to_cpu(d.v->d_child_subvol);
 
-		ret = bch2_subvolume_get(trans, target->subvol, true, BTREE_ITER_cached, &s);
+		ret = bch2_subvolume_get(trans, target->subvol, true, &s);
 
 		target->inum	= le64_to_cpu(s.inode);
 	}
diff --git a/libbcachefs/disk_accounting.c b/libbcachefs/disk_accounting.c
index 77534838..a0e49cbe 100644
--- a/libbcachefs/disk_accounting.c
+++ b/libbcachefs/disk_accounting.c
@@ -324,6 +324,14 @@ static int __bch2_accounting_mem_insert(struct bch_fs *c, struct bkey_s_c_accoun
 
 	eytzinger0_sort(acc->k.data, acc->k.nr, sizeof(acc->k.data[0]),
 			accounting_pos_cmp, NULL);
+
+	if (trace_accounting_mem_insert_enabled()) {
+		struct printbuf buf = PRINTBUF;
+
+		bch2_accounting_to_text(&buf, c, a.s_c);
+		trace_accounting_mem_insert(c, buf.buf);
+		printbuf_exit(&buf);
+	}
 	return 0;
 err:
 	free_percpu(n.v[1]);
@@ -722,11 +730,18 @@ int bch2_accounting_read(struct bch_fs *c)
 	iter.flags &= ~BTREE_ITER_with_journal;
 	int ret = for_each_btree_key_continue(trans, iter,
 				BTREE_ITER_prefetch|BTREE_ITER_all_snapshots, k, ({
+			struct bkey u;
+			struct bkey_s_c k = bch2_btree_path_peek_slot_exact(btree_iter_path(trans, &iter), &u);
+
 			if (k.k->type != KEY_TYPE_accounting)
 				continue;
 
 			struct disk_accounting_pos acc_k;
 			bpos_to_disk_accounting_pos(&acc_k, k.k->p);
+
+			if (acc_k.type >= BCH_DISK_ACCOUNTING_TYPE_NR)
+				break;
+
 			if (!bch2_accounting_is_mem(acc_k)) {
 				struct disk_accounting_pos next = { .type = acc_k.type + 1 };
 				bch2_btree_iter_set_pos(&iter, disk_accounting_pos_to_bpos(&next));
@@ -746,6 +761,7 @@ int bch2_accounting_read(struct bch_fs *c)
 		if (i->k->k.type == KEY_TYPE_accounting) {
 			struct disk_accounting_pos acc_k;
 			bpos_to_disk_accounting_pos(&acc_k, i->k->k.p);
+
 			if (!bch2_accounting_is_mem(acc_k))
 				continue;
 
@@ -782,15 +798,16 @@ int bch2_accounting_read(struct bch_fs *c)
 	keys->gap = keys->nr = dst - keys->data;
 
 	percpu_down_write(&c->mark_lock);
-	unsigned i = 0;
-	while (i < acc->k.nr) {
-		unsigned idx = inorder_to_eytzinger0(i, acc->k.nr);
 
+	darray_for_each_reverse(acc->k, i) {
 		struct disk_accounting_pos acc_k;
-		bpos_to_disk_accounting_pos(&acc_k, acc->k.data[idx].pos);
+		bpos_to_disk_accounting_pos(&acc_k, i->pos);
 
 		u64 v[BCH_ACCOUNTING_MAX_COUNTERS];
-		bch2_accounting_mem_read_counters(acc, idx, v, ARRAY_SIZE(v), false);
+		memset(v, 0, sizeof(v));
+
+		for (unsigned j = 0; j < i->nr_counters; j++)
+			v[j] = percpu_u64_get(i->v[0] + j);
 
 		/*
 		 * If the entry counters are zeroed, it should be treated as
@@ -799,26 +816,25 @@ int bch2_accounting_read(struct bch_fs *c)
 		 * Remove it, so that if it's re-added it gets re-marked in the
 		 * superblock:
 		 */
-		ret = bch2_is_zero(v, sizeof(v[0]) * acc->k.data[idx].nr_counters)
+		ret = bch2_is_zero(v, sizeof(v[0]) * i->nr_counters)
 			? -BCH_ERR_remove_disk_accounting_entry
-			: bch2_disk_accounting_validate_late(trans, acc_k,
-							v, acc->k.data[idx].nr_counters);
+			: bch2_disk_accounting_validate_late(trans, acc_k, v, i->nr_counters);
 
 		if (ret == -BCH_ERR_remove_disk_accounting_entry) {
-			free_percpu(acc->k.data[idx].v[0]);
-			free_percpu(acc->k.data[idx].v[1]);
-			darray_remove_item(&acc->k, &acc->k.data[idx]);
-			eytzinger0_sort(acc->k.data, acc->k.nr, sizeof(acc->k.data[0]),
-					accounting_pos_cmp, NULL);
+			free_percpu(i->v[0]);
+			free_percpu(i->v[1]);
+			darray_remove_item(&acc->k, i);
 			ret = 0;
 			continue;
 		}
 
 		if (ret)
 			goto fsck_err;
-		i++;
 	}
 
+	eytzinger0_sort(acc->k.data, acc->k.nr, sizeof(acc->k.data[0]),
+			accounting_pos_cmp, NULL);
+
 	preempt_disable();
 	struct bch_fs_usage_base *usage = this_cpu_ptr(c->usage);
 
diff --git a/libbcachefs/disk_accounting.h b/libbcachefs/disk_accounting.h
index cb20f723..4a3d9ff8 100644
--- a/libbcachefs/disk_accounting.h
+++ b/libbcachefs/disk_accounting.h
@@ -120,7 +120,8 @@ void bch2_accounting_mem_gc(struct bch_fs *);
 
 static inline bool bch2_accounting_is_mem(struct disk_accounting_pos acc)
 {
-	return acc.type != BCH_DISK_ACCOUNTING_inum;
+	return acc.type < BCH_DISK_ACCOUNTING_TYPE_NR &&
+		acc.type != BCH_DISK_ACCOUNTING_inum;
 }
 
 /*
diff --git a/libbcachefs/errcode.h b/libbcachefs/errcode.h
index 924d2400..775425df 100644
--- a/libbcachefs/errcode.h
+++ b/libbcachefs/errcode.h
@@ -264,6 +264,7 @@
 	x(EIO,				missing_indirect_extent)		\
 	x(EIO,				invalidate_stripe_to_dev)		\
 	x(EIO,				no_encryption_key)			\
+	x(EIO,				insufficient_journal_devices)		\
 	x(BCH_ERR_btree_node_read_err,	btree_node_read_err_fixable)		\
 	x(BCH_ERR_btree_node_read_err,	btree_node_read_err_want_retry)		\
 	x(BCH_ERR_btree_node_read_err,	btree_node_read_err_must_retry)		\
diff --git a/libbcachefs/error.c b/libbcachefs/error.c
index 7af5c594..038da6a6 100644
--- a/libbcachefs/error.c
+++ b/libbcachefs/error.c
@@ -477,8 +477,8 @@ int __bch2_bkey_fsck_err(struct bch_fs *c,
 
 	unsigned fsck_flags = 0;
 	if (!(from.flags & (BCH_VALIDATE_write|BCH_VALIDATE_commit))) {
-		//if (test_bit(err, c->sb.errors_silent))
-		//	return -BCH_ERR_fsck_delete_bkey;
+		if (test_bit(err, c->sb.errors_silent))
+			return -BCH_ERR_fsck_delete_bkey;
 
 		fsck_flags |= FSCK_AUTOFIX|FSCK_CAN_FIX;
 	}
@@ -486,9 +486,14 @@ int __bch2_bkey_fsck_err(struct bch_fs *c,
 		fsck_flags |= fsck_flags_extra[err];
 
 	struct printbuf buf = PRINTBUF;
-
-	prt_printf(&buf, "invalid bkey in %s btree=",
+	prt_printf(&buf, "invalid bkey in %s",
 		   bch2_bkey_validate_contexts[from.from]);
+
+	if (from.from == BKEY_VALIDATE_journal)
+		prt_printf(&buf, " journal seq=%llu offset=%u",
+			   from.journal_seq, from.journal_offset);
+
+	prt_str(&buf, " btree=");
 	bch2_btree_id_to_text(&buf, from.btree);
 	prt_printf(&buf, " level=%u: ", from.level);
 
diff --git a/libbcachefs/extents.c b/libbcachefs/extents.c
index 2fc9ace5..05d5f71a 100644
--- a/libbcachefs/extents.c
+++ b/libbcachefs/extents.c
@@ -1238,6 +1238,12 @@ static int extent_ptr_validate(struct bch_fs *c,
 {
 	int ret = 0;
 
+	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
+	bkey_for_each_ptr(ptrs, ptr2)
+		bkey_fsck_err_on(ptr != ptr2 && ptr->dev == ptr2->dev,
+				 c, ptr_to_duplicate_device,
+				 "multiple pointers to same device (%u)", ptr->dev);
+
 	/* bad pointers are repaired by check_fix_ptrs(): */
 	rcu_read_lock();
 	struct bch_dev *ca = bch2_dev_rcu_noerror(c, ptr->dev);
@@ -1252,13 +1258,6 @@ static int extent_ptr_validate(struct bch_fs *c,
 	unsigned bucket_size	= ca->mi.bucket_size;
 	rcu_read_unlock();
 
-	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
-	bkey_for_each_ptr(ptrs, ptr2)
-		bkey_fsck_err_on(ptr != ptr2 && ptr->dev == ptr2->dev,
-				 c, ptr_to_duplicate_device,
-				 "multiple pointers to same device (%u)", ptr->dev);
-
-
 	bkey_fsck_err_on(bucket >= nbuckets,
 			 c, ptr_after_last_bucket,
 			 "pointer past last bucket (%llu > %llu)", bucket, nbuckets);
diff --git a/libbcachefs/fs-common.c b/libbcachefs/fs-common.c
index d2966b67..7d279f21 100644
--- a/libbcachefs/fs-common.c
+++ b/libbcachefs/fs-common.c
@@ -69,9 +69,7 @@ int bch2_create_trans(struct btree_trans *trans,
 		if (!snapshot_src.inum) {
 			/* Inode wasn't specified, just snapshot: */
 			struct bch_subvolume s;
-
-			ret = bch2_subvolume_get(trans, snapshot_src.subvol, true,
-						 BTREE_ITER_cached, &s);
+			ret = bch2_subvolume_get(trans, snapshot_src.subvol, true, &s);
 			if (ret)
 				goto err;
 
diff --git a/libbcachefs/fs-io-direct.c b/libbcachefs/fs-io-direct.c
index b0367b9d..2089c36b 100644
--- a/libbcachefs/fs-io-direct.c
+++ b/libbcachefs/fs-io-direct.c
@@ -226,7 +226,6 @@ struct dio_write {
 	struct mm_struct		*mm;
 	const struct iovec		*iov;
 	unsigned			loop:1,
-					have_mm_ref:1,
 					extending:1,
 					sync:1,
 					flush:1;
@@ -391,9 +390,6 @@ static __always_inline long bch2_dio_write_done(struct dio_write *dio)
 
 	kfree(dio->iov);
 
-	if (dio->have_mm_ref)
-		mmdrop(dio->mm);
-
 	ret = dio->op.error ?: ((long) dio->written << 9);
 	bio_put(&dio->op.wbio.bio);
 
@@ -533,24 +529,9 @@ static __always_inline long bch2_dio_write_loop(struct dio_write *dio)
 
 		if (unlikely(dio->iter.count) &&
 		    !dio->sync &&
-		    !dio->loop) {
-			/*
-			 * Rest of write will be submitted asynchronously -
-			 * unless copying the iov fails:
-			 */
-			if (likely(!bch2_dio_write_copy_iov(dio))) {
-				/*
-				 * aio guarantees that mm_struct outlives the
-				 * request, but io_uring does not
-				 */
-				if (dio->mm) {
-					mmgrab(dio->mm);
-					dio->have_mm_ref = true;
-				}
-			} else {
-				dio->sync = sync = true;
-			}
-		}
+		    !dio->loop &&
+		    bch2_dio_write_copy_iov(dio))
+			dio->sync = sync = true;
 
 		dio->loop = true;
 		closure_call(&dio->op.cl, bch2_write, NULL, NULL);
@@ -578,25 +559,15 @@ err:
 
 static noinline __cold void bch2_dio_write_continue(struct dio_write *dio)
 {
-	struct mm_struct *mm = dio->have_mm_ref ? dio->mm: NULL;
+	struct mm_struct *mm = dio->mm;
 
 	bio_reset(&dio->op.wbio.bio, NULL, REQ_OP_WRITE);
 
-	if (mm) {
-		if (unlikely(!mmget_not_zero(mm))) {
-			/* process exited */
-			dio->op.error = -ESRCH;
-			bch2_dio_write_done(dio);
-			return;
-		}
-
+	if (mm)
 		kthread_use_mm(mm);
-	}
 	bch2_dio_write_loop(dio);
-	if (mm) {
+	if (mm)
 		kthread_unuse_mm(mm);
-		mmput(mm);
-	}
 }
 
 static void bch2_dio_write_loop_async(struct bch_write_op *op)
@@ -670,7 +641,6 @@ ssize_t bch2_direct_write(struct kiocb *req, struct iov_iter *iter)
 	dio->mm			= current->mm;
 	dio->iov		= NULL;
 	dio->loop		= false;
-	dio->have_mm_ref	= false;
 	dio->extending		= extending;
 	dio->sync		= is_sync_kiocb(req) || extending;
 	dio->flush		= iocb_is_dsync(req) && !c->opts.journal_flush_disabled;
diff --git a/libbcachefs/fs.c b/libbcachefs/fs.c
index c6e7df7c..3f83f131 100644
--- a/libbcachefs/fs.c
+++ b/libbcachefs/fs.c
@@ -499,7 +499,7 @@ struct inode *bch2_vfs_inode_get(struct bch_fs *c, subvol_inum inum)
 	struct bch_inode_unpacked inode_u;
 	struct bch_subvolume subvol;
 	int ret = lockrestart_do(trans,
-		bch2_subvolume_get(trans, inum.subvol, true, 0, &subvol) ?:
+		bch2_subvolume_get(trans, inum.subvol, true, &subvol) ?:
 		bch2_inode_find_by_inum_trans(trans, inum, &inode_u)) ?:
 		PTR_ERR_OR_ZERO(inode = bch2_inode_hash_init_insert(trans, inum, &inode_u, &subvol));
 	bch2_trans_put(trans);
@@ -569,8 +569,7 @@ retry:
 	inum.subvol = inode_u.bi_subvol ?: dir->ei_inum.subvol;
 	inum.inum = inode_u.bi_inum;
 
-	ret   = bch2_subvolume_get(trans, inum.subvol, true,
-				   BTREE_ITER_with_updates, &subvol) ?:
+	ret   = bch2_subvolume_get(trans, inum.subvol, true, &subvol) ?:
 		bch2_trans_commit(trans, NULL, &journal_seq, 0);
 	if (unlikely(ret)) {
 		bch2_quota_acct(c, bch_qid(&inode_u), Q_INO, -1,
@@ -651,7 +650,7 @@ static struct bch_inode_info *bch2_lookup_trans(struct btree_trans *trans,
 
 	struct bch_subvolume subvol;
 	struct bch_inode_unpacked inode_u;
-	ret =   bch2_subvolume_get(trans, inum.subvol, true, 0, &subvol) ?:
+	ret =   bch2_subvolume_get(trans, inum.subvol, true, &subvol) ?:
 		bch2_inode_find_by_inum_nowarn_trans(trans, inum, &inode_u) ?:
 		PTR_ERR_OR_ZERO(inode = bch2_inode_hash_init_insert(trans, inum, &inode_u, &subvol));
 
diff --git a/libbcachefs/fsck.c b/libbcachefs/fsck.c
index d67a3f52..a436c072 100644
--- a/libbcachefs/fsck.c
+++ b/libbcachefs/fsck.c
@@ -109,7 +109,7 @@ static int subvol_lookup(struct btree_trans *trans, u32 subvol,
 			 u32 *snapshot, u64 *inum)
 {
 	struct bch_subvolume s;
-	int ret = bch2_subvolume_get(trans, subvol, false, 0, &s);
+	int ret = bch2_subvolume_get(trans, subvol, false, &s);
 
 	*snapshot = le32_to_cpu(s.snapshot);
 	*inum = le64_to_cpu(s.inode);
@@ -226,8 +226,7 @@ static int lookup_lostfound(struct btree_trans *trans, u32 snapshot,
 	subvol_inum root_inum = { .subvol = le32_to_cpu(st.master_subvol) };
 
 	struct bch_subvolume subvol;
-	ret = bch2_subvolume_get(trans, le32_to_cpu(st.master_subvol),
-				 false, 0, &subvol);
+	ret = bch2_subvolume_get(trans, le32_to_cpu(st.master_subvol), false, &subvol);
 	bch_err_msg(c, ret, "looking up root subvol %u for snapshot %u",
 		    le32_to_cpu(st.master_subvol), snapshot);
 	if (ret)
@@ -942,69 +941,16 @@ static int get_visible_inodes(struct btree_trans *trans,
 	return ret;
 }
 
-static int dirent_has_target(struct btree_trans *trans, struct bkey_s_c_dirent d)
-{
-	if (d.v->d_type == DT_SUBVOL) {
-		u32 snap;
-		u64 inum;
-		int ret = subvol_lookup(trans, le32_to_cpu(d.v->d_child_subvol), &snap, &inum);
-		if (ret && !bch2_err_matches(ret, ENOENT))
-			return ret;
-		return !ret;
-	} else {
-		struct btree_iter iter;
-		struct bkey_s_c k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_inodes,
-				SPOS(0, le64_to_cpu(d.v->d_inum), d.k->p.snapshot), 0);
-		int ret = bkey_err(k);
-		if (ret)
-			return ret;
-
-		ret = bkey_is_inode(k.k);
-		bch2_trans_iter_exit(trans, &iter);
-		return ret;
-	}
-}
-
 /*
  * Prefer to delete the first one, since that will be the one at the wrong
  * offset:
  * return value: 0 -> delete k1, 1 -> delete k2
  */
-static int hash_pick_winner(struct btree_trans *trans,
-			    const struct bch_hash_desc desc,
-			    struct bch_hash_info *hash_info,
-			    struct bkey_s_c k1,
-			    struct bkey_s_c k2)
-{
-	if (bkey_val_bytes(k1.k) == bkey_val_bytes(k2.k) &&
-	    !memcmp(k1.v, k2.v, bkey_val_bytes(k1.k)))
-		return 0;
-
-	switch (desc.btree_id) {
-	case BTREE_ID_dirents: {
-		int ret = dirent_has_target(trans, bkey_s_c_to_dirent(k1));
-		if (ret < 0)
-			return ret;
-		if (!ret)
-			return 0;
-
-		ret = dirent_has_target(trans, bkey_s_c_to_dirent(k2));
-		if (ret < 0)
-			return ret;
-		if (!ret)
-			return 1;
-		return 2;
-	}
-	default:
-		return 0;
-	}
-}
-
-static int fsck_update_backpointers(struct btree_trans *trans,
-				    struct snapshots_seen *s,
-				    const struct bch_hash_desc desc,
-				    struct bch_hash_info *hash_info,
-				    struct bkey_i *new)
+int bch2_fsck_update_backpointers(struct btree_trans *trans,
+				  struct snapshots_seen *s,
+				  const struct bch_hash_desc desc,
+				  struct bch_hash_info *hash_info,
+				  struct bkey_i *new)
 {
 	if (new->k.type != KEY_TYPE_dirent)
 		return 0;
@@ -1032,153 +978,6 @@ err:
 	return ret;
 }
 
-static int fsck_rename_dirent(struct btree_trans *trans,
-			      struct snapshots_seen *s,
-			      const struct bch_hash_desc desc,
-			      struct bch_hash_info *hash_info,
-			      struct bkey_s_c_dirent old)
-{
-	struct qstr old_name = bch2_dirent_get_name(old);
-	struct bkey_i_dirent *new = bch2_trans_kmalloc(trans, bkey_bytes(old.k) + 32);
-	int ret = PTR_ERR_OR_ZERO(new);
-	if (ret)
-		return ret;
-
-	bkey_dirent_init(&new->k_i);
-	dirent_copy_target(new, old);
-	new->k.p = old.k->p;
-
-	for (unsigned i = 0; i < 1000; i++) {
-		unsigned len = sprintf(new->v.d_name, "%.*s.fsck_renamed-%u",
-				       old_name.len, old_name.name, i);
-		unsigned u64s = BKEY_U64s + dirent_val_u64s(len);
-
-		if (u64s > U8_MAX)
-			return -EINVAL;
-
-		new->k.u64s = u64s;
-
-		ret = bch2_hash_set_in_snapshot(trans, bch2_dirent_hash_desc, hash_info,
-						(subvol_inum) { 0, old.k->p.inode },
-						old.k->p.snapshot, &new->k_i,
-						BTREE_UPDATE_internal_snapshot_node);
-		if (!bch2_err_matches(ret, EEXIST))
-			break;
-	}
-
-	if (ret)
-		return ret;
-
-	return fsck_update_backpointers(trans, s, desc, hash_info, &new->k_i);
-}
-
-static int hash_check_key(struct btree_trans *trans,
-			  struct snapshots_seen *s,
-			  const struct bch_hash_desc desc,
-			  struct bch_hash_info *hash_info,
-			  struct btree_iter *k_iter, struct bkey_s_c hash_k)
-{
-	struct bch_fs *c = trans->c;
-	struct btree_iter iter = { NULL };
-	struct printbuf buf = PRINTBUF;
-	struct bkey_s_c k;
-	u64 hash;
-	int ret = 0;
-
-	if (hash_k.k->type != desc.key_type)
-		return 0;
-
-	hash = desc.hash_bkey(hash_info, hash_k);
-
-	if (likely(hash == hash_k.k->p.offset))
-		return 0;
-
-	if (hash_k.k->p.offset < hash)
-		goto bad_hash;
-
-	for_each_btree_key_norestart(trans, iter, desc.btree_id,
-				     SPOS(hash_k.k->p.inode, hash, hash_k.k->p.snapshot),
-				     BTREE_ITER_slots, k, ret) {
-		if (bkey_eq(k.k->p, hash_k.k->p))
-			break;
-
-		if (k.k->type == desc.key_type &&
-		    !desc.cmp_bkey(k, hash_k))
-			goto duplicate_entries;
-
-		if (bkey_deleted(k.k)) {
-			bch2_trans_iter_exit(trans, &iter);
-			goto bad_hash;
-		}
-	}
-out:
-	bch2_trans_iter_exit(trans, &iter);
-	printbuf_exit(&buf);
-	return ret;
-bad_hash:
-	if (fsck_err(trans, hash_table_key_wrong_offset,
-		     "hash table key at wrong offset: btree %s inode %llu offset %llu, hashed to %llu\n  %s",
-		     bch2_btree_id_str(desc.btree_id), hash_k.k->p.inode, hash_k.k->p.offset, hash,
-		     (printbuf_reset(&buf),
-		      bch2_bkey_val_to_text(&buf, c, hash_k), buf.buf))) {
-		struct bkey_i *new = bch2_bkey_make_mut_noupdate(trans, hash_k);
-		if (IS_ERR(new))
-			return PTR_ERR(new);
-
-		k = bch2_hash_set_or_get_in_snapshot(trans, &iter, desc, hash_info,
-				       (subvol_inum) { 0, hash_k.k->p.inode },
-				       hash_k.k->p.snapshot, new,
-				       STR_HASH_must_create|
-				       BTREE_ITER_with_updates|
-				       BTREE_UPDATE_internal_snapshot_node);
-		ret = bkey_err(k);
-		if (ret)
-			goto out;
-		if (k.k)
-			goto duplicate_entries;
-
-		ret =   bch2_hash_delete_at(trans, desc, hash_info, k_iter,
-					    BTREE_UPDATE_internal_snapshot_node) ?:
-			fsck_update_backpointers(trans, s, desc, hash_info, new) ?:
-			bch2_trans_commit(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc) ?:
-			-BCH_ERR_transaction_restart_nested;
-		goto out;
-	}
-fsck_err:
-	goto out;
-duplicate_entries:
-	ret = hash_pick_winner(trans, desc, hash_info, hash_k, k);
-	if (ret < 0)
-		goto out;
-
-	if (!fsck_err(trans, hash_table_key_duplicate,
-		      "duplicate hash table keys%s:\n%s",
-		      ret != 2 ? "" : ", both point to valid inodes",
-		      (printbuf_reset(&buf),
-		       bch2_bkey_val_to_text(&buf, c, hash_k),
-		       prt_newline(&buf),
-		       bch2_bkey_val_to_text(&buf, c, k),
-		       buf.buf)))
-		goto out;
-
-	switch (ret) {
-	case 0:
-		ret = bch2_hash_delete_at(trans, desc, hash_info, k_iter, 0);
-		break;
-	case 1:
-		ret = bch2_hash_delete_at(trans, desc, hash_info, &iter, 0);
-		break;
-	case 2:
-		ret = fsck_rename_dirent(trans, s, desc, hash_info, bkey_s_c_to_dirent(hash_k)) ?:
-			bch2_hash_delete_at(trans, desc, hash_info, k_iter, 0);
-		goto out;
-	}
-
-	ret = bch2_trans_commit(trans, NULL, NULL, 0) ?:
-		-BCH_ERR_transaction_restart_nested;
-	goto out;
-}
-
 static struct bkey_s_c_dirent inode_get_dirent(struct btree_trans *trans,
 					       struct btree_iter *iter,
 					       struct bch_inode_unpacked *inode,
@@ -1421,7 +1220,7 @@ static int check_inode(struct btree_trans *trans,
 	if (u.bi_subvol) {
 		struct bch_subvolume s;
 
-		ret = bch2_subvolume_get(trans, u.bi_subvol, false, 0, &s);
+		ret = bch2_subvolume_get(trans, u.bi_subvol, false, &s);
 		if (ret && !bch2_err_matches(ret, ENOENT))
 			goto err;
 
@@ -2497,7 +2296,7 @@ static int check_dirent(struct btree_trans *trans, struct btree_iter *iter,
 		*hash_info = bch2_hash_info_init(c, &i->inode);
 	dir->first_this_inode = false;
 
-	ret = hash_check_key(trans, s, bch2_dirent_hash_desc, hash_info, iter, k);
+	ret = bch2_str_hash_check_key(trans, s, bch2_dirent_hash_desc, hash_info, iter, k);
 	if (ret < 0)
 		goto err;
 	if (ret) {
@@ -2611,7 +2410,7 @@ static int check_xattr(struct btree_trans *trans, struct btree_iter *iter,
 		*hash_info = bch2_hash_info_init(c, &i->inode);
 	inode->first_this_inode = false;
 
-	ret = hash_check_key(trans, NULL, bch2_xattr_hash_desc, hash_info, iter, k);
+	ret = bch2_str_hash_check_key(trans, NULL, bch2_xattr_hash_desc, hash_info, iter, k);
 	bch_err_fn(c, ret);
 	return ret;
 }
@@ -2864,7 +2663,7 @@ static int check_path_loop(struct btree_trans *trans, struct bkey_s_c inode_k)
 		d = inode_get_dirent(trans, &dirent_iter, &inode, &parent_snapshot);
 		ret = bkey_err(d.s_c);
 		if (ret && !bch2_err_matches(ret, ENOENT))
-			break;
+			goto out;
 
 		if (!ret && (ret = dirent_points_to_inode(c, d, &inode)))
 			bch2_trans_iter_exit(trans, &dirent_iter);
@@ -2899,7 +2698,7 @@ static int check_path_loop(struct btree_trans *trans, struct bkey_s_c inode_k)
 		if (ret) {
 			/* Should have been caught in dirents pass */
 			bch_err_msg(c, ret, "error looking up parent directory");
-			break;
+			goto out;
 		}
 
 		min_bi_depth = parent_inode.bi_depth;
@@ -2930,8 +2729,7 @@ static int check_path_loop(struct btree_trans *trans, struct bkey_s_c inode_k)
 				bch_err_msg(c, ret, "reattaching inode %llu", inode.bi_inum);
 			}
 
-			redo_bi_depth = false;
-			break;
+			goto out;
 		}
 	}
 
diff --git a/libbcachefs/fsck.h b/libbcachefs/fsck.h
index 4481b40a..57494827 100644
--- a/libbcachefs/fsck.h
+++ b/libbcachefs/fsck.h
@@ -2,6 +2,14 @@
 #ifndef _BCACHEFS_FSCK_H
 #define _BCACHEFS_FSCK_H
 
+#include "str_hash.h"
+
+int bch2_fsck_update_backpointers(struct btree_trans *,
+				  struct snapshots_seen *,
+				  const struct bch_hash_desc,
+				  struct bch_hash_info *,
+				  struct bkey_i *);
+
 int bch2_check_inodes(struct bch_fs *);
 int bch2_check_extents(struct bch_fs *);
 int bch2_check_indirect_extents(struct bch_fs *);
diff --git a/libbcachefs/inode.c b/libbcachefs/inode.c
index 8818e418..d1fc44a7 100644
--- a/libbcachefs/inode.c
+++ b/libbcachefs/inode.c
@@ -799,6 +799,28 @@ void bch2_inode_generation_to_text(struct printbuf *out, struct bch_fs *c,
 	prt_printf(out, "generation: %u", le32_to_cpu(gen.v->bi_generation));
 }
 
+int bch2_inode_alloc_cursor_validate(struct bch_fs *c, struct bkey_s_c k,
+				   struct bkey_validate_context from)
+{
+	int ret = 0;
+
+	bkey_fsck_err_on(k.k->p.inode != LOGGED_OPS_INUM_inode_cursors,
+			 c, inode_alloc_cursor_inode_bad,
+			 "k.p.inode bad");
+fsck_err:
+	return ret;
+}
+
+void bch2_inode_alloc_cursor_to_text(struct printbuf *out, struct bch_fs *c,
+				     struct bkey_s_c k)
+{
+	struct bkey_s_c_inode_alloc_cursor i = bkey_s_c_to_inode_alloc_cursor(k);
+
+	prt_printf(out, "idx %llu generation %llu",
+		   le64_to_cpu(i.v->idx),
+		   le64_to_cpu(i.v->gen));
+}
+
 void bch2_inode_init_early(struct bch_fs *c,
 			   struct bch_inode_unpacked *inode_u)
 {
@@ -859,43 +881,81 @@ static inline u32 bkey_generation(struct bkey_s_c k)
 	}
 }
 
-/*
- * This just finds an empty slot:
- */
-int bch2_inode_create(struct btree_trans *trans,
-		      struct btree_iter *iter,
-		      struct bch_inode_unpacked *inode_u,
-		      u32 snapshot, u64 cpu)
+static struct bkey_i_inode_alloc_cursor *
+bch2_inode_alloc_cursor_get(struct btree_trans *trans, u64 cpu, u64 *min, u64 *max)
 {
 	struct bch_fs *c = trans->c;
-	struct bkey_s_c k;
-	u64 min, max, start, pos, *hint;
-	int ret = 0;
-	unsigned bits = (c->opts.inodes_32bit ? 31 : 63);
 
+	u64 cursor_idx = c->opts.shard_inode_numbers ? cpu : 0;
+
+	cursor_idx &= ~(~0ULL << c->opts.shard_inode_numbers_bits);
+
+	struct btree_iter iter;
+	struct bkey_s_c k = bch2_bkey_get_iter(trans, &iter,
+					BTREE_ID_logged_ops,
+					POS(LOGGED_OPS_INUM_inode_cursors, cursor_idx),
+					BTREE_ITER_cached);
+	int ret = bkey_err(k);
+	if (ret)
+		return ERR_PTR(ret);
+
+	struct bkey_i_inode_alloc_cursor *cursor =
+		k.k->type == KEY_TYPE_inode_alloc_cursor
+		? bch2_bkey_make_mut_typed(trans, &iter, &k, 0, inode_alloc_cursor)
+		: bch2_bkey_alloc(trans, &iter, 0, inode_alloc_cursor);
+	ret = PTR_ERR_OR_ZERO(cursor);
+	if (ret)
+		goto err;
+
+	cursor->v.bits = c->opts.shard_inode_numbers_bits;
+
+	unsigned bits = (c->opts.inodes_32bit ? 31 : 63);
 	if (c->opts.shard_inode_numbers) {
-		bits -= c->inode_shard_bits;
+		bits -= cursor->v.bits;
 
-		min = (cpu << bits);
-		max = (cpu << bits) | ~(ULLONG_MAX << bits);
+		*min = (cpu << bits);
+		*max = (cpu << bits) | ~(ULLONG_MAX << bits);
 
-		min = max_t(u64, min, BLOCKDEV_INODE_MAX);
-		hint = c->unused_inode_hints + cpu;
+		*min = max_t(u64, *min, BLOCKDEV_INODE_MAX);
 	} else {
-		min = BLOCKDEV_INODE_MAX;
-		max = ~(ULLONG_MAX << bits);
-		hint = c->unused_inode_hints;
+		*min = BLOCKDEV_INODE_MAX;
+		*max = ~(ULLONG_MAX << bits);
 	}
 
-	start = READ_ONCE(*hint);
+	if (le64_to_cpu(cursor->v.idx)  < *min)
+		cursor->v.idx = cpu_to_le64(*min);
+
+	if (le64_to_cpu(cursor->v.idx) >= *max) {
+		cursor->v.idx = cpu_to_le64(*min);
+		le32_add_cpu(&cursor->v.gen, 1);
+	}
+err:
+	bch2_trans_iter_exit(trans, &iter);
+	return ret ? ERR_PTR(ret) : cursor;
+}
+
+/*
+ * This just finds an empty slot:
+ */
+int bch2_inode_create(struct btree_trans *trans,
+		      struct btree_iter *iter,
+		      struct bch_inode_unpacked *inode_u,
+		      u32 snapshot, u64 cpu)
+{
+	u64 min, max;
+	struct bkey_i_inode_alloc_cursor *cursor =
+		bch2_inode_alloc_cursor_get(trans, cpu, &min, &max);
+	int ret = PTR_ERR_OR_ZERO(cursor);
+	if (ret)
+		return ret;
 
-	if (start >= max || start < min)
-		start = min;
+	u64 start = le64_to_cpu(cursor->v.idx);
+	u64 pos = start;
 
-	pos = start;
 	bch2_trans_iter_init(trans, iter, BTREE_ID_inodes, POS(0, pos),
 			     BTREE_ITER_all_snapshots|
 			     BTREE_ITER_intent);
+	struct bkey_s_c k;
 again:
 	while ((k = bch2_btree_iter_peek(iter)).k &&
 	       !(ret = bkey_err(k)) &&
@@ -925,6 +985,7 @@ again:
 	/* Retry from start */
 	pos = start = min;
 	bch2_btree_iter_set_pos(iter, POS(0, pos));
+	le32_add_cpu(&cursor->v.gen, 1);
 	goto again;
 found_slot:
 	bch2_btree_iter_set_pos(iter, SPOS(0, pos, snapshot));
@@ -935,9 +996,9 @@ found_slot:
 		return ret;
 	}
 
-	*hint			= k.k->p.offset;
 	inode_u->bi_inum	= k.k->p.offset;
-	inode_u->bi_generation	= bkey_generation(k);
+	inode_u->bi_generation	= le64_to_cpu(cursor->v.gen);
+	cursor->v.idx		= cpu_to_le64(k.k->p.offset + 1);
 	return 0;
 }
 
@@ -999,8 +1060,6 @@ int bch2_inode_rm(struct bch_fs *c, subvol_inum inum)
 {
 	struct btree_trans *trans = bch2_trans_get(c);
 	struct btree_iter iter = { NULL };
-	struct bkey_i_inode_generation delete;
-	struct bch_inode_unpacked inode_u;
 	struct bkey_s_c k;
 	u32 snapshot;
 	int ret;
@@ -1040,13 +1099,7 @@ retry:
 		goto err;
 	}
 
-	bch2_inode_unpack(k, &inode_u);
-
-	bkey_inode_generation_init(&delete.k_i);
-	delete.k.p = iter.pos;
-	delete.v.bi_generation = cpu_to_le32(inode_u.bi_generation + 1);
-
-	ret   = bch2_trans_update(trans, &iter, &delete.k_i, 0) ?:
+	ret   = bch2_btree_delete_at(trans, &iter, 0) ?:
 		bch2_trans_commit(trans, NULL, NULL,
 				BCH_TRANS_COMMIT_no_enospc);
 err:
diff --git a/libbcachefs/inode.h b/libbcachefs/inode.h
index 5bca6950..d2e13452 100644
--- a/libbcachefs/inode.h
+++ b/libbcachefs/inode.h
@@ -68,6 +68,16 @@ void bch2_inode_generation_to_text(struct printbuf *, struct bch_fs *, struct bk
 	.min_val_size	= 8,					\
 })
 
+int bch2_inode_alloc_cursor_validate(struct bch_fs *, struct bkey_s_c,
+				     struct bkey_validate_context);
+void bch2_inode_alloc_cursor_to_text(struct printbuf *, struct bch_fs *, struct bkey_s_c);
+
+#define bch2_bkey_ops_inode_alloc_cursor ((struct bkey_ops) {	\
+	.key_validate	= bch2_inode_alloc_cursor_validate,	\
+	.val_to_text	= bch2_inode_alloc_cursor_to_text,	\
+	.min_val_size	= 16,					\
+})
+
 #if 0
 typedef struct {
 	u64			lo;
diff --git a/libbcachefs/inode_format.h b/libbcachefs/inode_format.h
index be1e7476..1b93e189 100644
--- a/libbcachefs/inode_format.h
+++ b/libbcachefs/inode_format.h
@@ -165,4 +165,12 @@ LE64_BITMASK(INODEv3_FIELDS_START,
 				struct bch_inode_v3, bi_flags, 31, 36);
 LE64_BITMASK(INODEv3_MODE,	struct bch_inode_v3, bi_flags, 36, 52);
 
+struct bch_inode_alloc_cursor {
+	struct bch_val		v;
+	__u8			bits;
+	__u8			pad;
+	__le32			gen;
+	__le64			idx;
+};
+
 #endif /* _BCACHEFS_INODE_FORMAT_H */
diff --git a/libbcachefs/journal.c b/libbcachefs/journal.c
index dc665219..2cd20114 100644
--- a/libbcachefs/journal.c
+++ b/libbcachefs/journal.c
@@ -807,10 +807,11 @@ int bch2_journal_flush(struct journal *j)
 }
 
 /*
- * bch2_journal_noflush_seq - tell the journal not to issue any flushes before
+ * bch2_journal_noflush_seq - ask the journal not to issue any flushes in the
+ * range [start, end)
  * @seq
  */
-bool bch2_journal_noflush_seq(struct journal *j, u64 seq)
+bool bch2_journal_noflush_seq(struct journal *j, u64 start, u64 end)
 {
 	struct bch_fs *c = container_of(j, struct bch_fs, journal);
 	u64 unwritten_seq;
@@ -819,15 +820,15 @@ bool bch2_journal_noflush_seq(struct journal *j, u64 seq)
 	if (!(c->sb.features & (1ULL << BCH_FEATURE_journal_no_flush)))
 		return false;
 
-	if (seq <= c->journal.flushed_seq_ondisk)
+	if (c->journal.flushed_seq_ondisk >= start)
 		return false;
 
 	spin_lock(&j->lock);
-	if (seq <= c->journal.flushed_seq_ondisk)
+	if (c->journal.flushed_seq_ondisk >= start)
 		goto out;
 
 	for (unwritten_seq = journal_last_unwritten_seq(j);
-	     unwritten_seq < seq;
+	     unwritten_seq < end;
 	     unwritten_seq++) {
 		struct journal_buf *buf = journal_seq_to_buf(j, unwritten_seq);
 
@@ -1564,6 +1565,9 @@ void __bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
 	printbuf_indent_sub(out, 2);
 
 	for_each_member_device_rcu(c, ca, &c->rw_devs[BCH_DATA_journal]) {
+		if (!ca->mi.durability)
+			continue;
+
 		struct journal_device *ja = &ca->journal;
 
 		if (!test_bit(ca->dev_idx, c->rw_devs[BCH_DATA_journal].d))
@@ -1573,6 +1577,7 @@ void __bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
 			continue;
 
 		prt_printf(out, "dev %u:\n",			ca->dev_idx);
+		prt_printf(out, "durability %u:\n",		ca->mi.durability);
 		printbuf_indent_add(out, 2);
 		prt_printf(out, "nr\t%u\n",			ja->nr);
 		prt_printf(out, "bucket size\t%u\n",		ca->mi.bucket_size);
@@ -1584,6 +1589,8 @@ void __bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
 		printbuf_indent_sub(out, 2);
 	}
 
+	prt_printf(out, "replicas want %u need %u\n", c->opts.metadata_replicas, c->opts.metadata_replicas_required);
+
 	rcu_read_unlock();
 
 	--out->atomic;
diff --git a/libbcachefs/journal.h b/libbcachefs/journal.h
index a6a2e888..cb0df066 100644
--- a/libbcachefs/journal.h
+++ b/libbcachefs/journal.h
@@ -404,7 +404,7 @@ void bch2_journal_flush_async(struct journal *, struct closure *);
 
 int bch2_journal_flush_seq(struct journal *, u64, unsigned);
 int bch2_journal_flush(struct journal *);
-bool bch2_journal_noflush_seq(struct journal *, u64);
+bool bch2_journal_noflush_seq(struct journal *, u64, u64);
 int bch2_journal_meta(struct journal *);
 
 void bch2_journal_halt(struct journal *);
diff --git a/libbcachefs/journal_io.c b/libbcachefs/journal_io.c
index bb69d808..7f2efe85 100644
--- a/libbcachefs/journal_io.c
+++ b/libbcachefs/journal_io.c
@@ -301,7 +301,7 @@ static void journal_entry_err_msg(struct printbuf *out,
 	journal_entry_err_msg(&_buf, version, jset, entry);		\
 	prt_printf(&_buf, msg, ##__VA_ARGS__);				\
 									\
-	switch (flags & BCH_VALIDATE_write) {				\
+	switch (from.flags & BCH_VALIDATE_write) {			\
 	case READ:							\
 		mustfix_fsck_err(c, _err, "%s", _buf.buf);		\
 		break;							\
@@ -390,15 +390,12 @@ static int journal_entry_btree_keys_validate(struct bch_fs *c,
 				struct jset *jset,
 				struct jset_entry *entry,
 				unsigned version, int big_endian,
-				enum bch_validate_flags flags)
+				struct bkey_validate_context from)
 {
 	struct bkey_i *k = entry->start;
-	struct bkey_validate_context from = {
-		.from	= BKEY_VALIDATE_journal,
-		.level	= entry->level,
-		.btree	= entry->btree_id,
-		.flags	= flags|BCH_VALIDATE_journal,
-	};
+
+	from.level	= entry->level;
+	from.btree	= entry->btree_id;
 
 	while (k != vstruct_last(entry)) {
 		int ret = journal_validate_key(c, jset, entry, k, from, version, big_endian);
@@ -435,11 +432,15 @@ static int journal_entry_btree_root_validate(struct bch_fs *c,
 				struct jset *jset,
 				struct jset_entry *entry,
 				unsigned version, int big_endian,
-				enum bch_validate_flags flags)
+				struct bkey_validate_context from)
 {
 	struct bkey_i *k = entry->start;
 	int ret = 0;
 
+	from.root	= true;
+	from.level	= entry->level + 1;
+	from.btree	= entry->btree_id;
+
 	if (journal_entry_err_on(!entry->u64s ||
 				 le16_to_cpu(entry->u64s) != k->k.u64s,
 				 c, version, jset, entry,
@@ -456,13 +457,6 @@ static int journal_entry_btree_root_validate(struct bch_fs *c,
 		return 0;
 	}
 
-	struct bkey_validate_context from = {
-		.from	= BKEY_VALIDATE_journal,
-		.level	= entry->level + 1,
-		.btree	= entry->btree_id,
-		.root	= true,
-		.flags	= flags,
-	};
 	ret = journal_validate_key(c, jset, entry, k, from, version, big_endian);
 	if (ret == FSCK_DELETED_KEY)
 		ret = 0;
@@ -480,7 +474,7 @@ static int journal_entry_prio_ptrs_validate(struct bch_fs *c,
 				struct jset *jset,
 				struct jset_entry *entry,
 				unsigned version, int big_endian,
-				enum bch_validate_flags flags)
+				struct bkey_validate_context from)
 {
 	/* obsolete, don't care: */
 	return 0;
@@ -495,7 +489,7 @@ static int journal_entry_blacklist_validate(struct bch_fs *c,
 				struct jset *jset,
 				struct jset_entry *entry,
 				unsigned version, int big_endian,
-				enum bch_validate_flags flags)
+				struct bkey_validate_context from)
 {
 	int ret = 0;
 
@@ -522,7 +516,7 @@ static int journal_entry_blacklist_v2_validate(struct bch_fs *c,
 				struct jset *jset,
 				struct jset_entry *entry,
 				unsigned version, int big_endian,
-				enum bch_validate_flags flags)
+				struct bkey_validate_context from)
 {
 	struct jset_entry_blacklist_v2 *bl_entry;
 	int ret = 0;
@@ -564,7 +558,7 @@ static int journal_entry_usage_validate(struct bch_fs *c,
 				struct jset *jset,
 				struct jset_entry *entry,
 				unsigned version, int big_endian,
-				enum bch_validate_flags flags)
+				struct bkey_validate_context from)
 {
 	struct jset_entry_usage *u =
 		container_of(entry, struct jset_entry_usage, entry);
@@ -598,7 +592,7 @@ static int journal_entry_data_usage_validate(struct bch_fs *c,
 				struct jset *jset,
 				struct jset_entry *entry,
 				unsigned version, int big_endian,
-				enum bch_validate_flags flags)
+				struct bkey_validate_context from)
 {
 	struct jset_entry_data_usage *u =
 		container_of(entry, struct jset_entry_data_usage, entry);
@@ -642,7 +636,7 @@ static int journal_entry_clock_validate(struct bch_fs *c,
 				struct jset *jset,
 				struct jset_entry *entry,
 				unsigned version, int big_endian,
-				enum bch_validate_flags flags)
+				struct bkey_validate_context from)
 {
 	struct jset_entry_clock *clock =
 		container_of(entry, struct jset_entry_clock, entry);
@@ -682,7 +676,7 @@ static int journal_entry_dev_usage_validate(struct bch_fs *c,
 				struct jset *jset,
 				struct jset_entry *entry,
 				unsigned version, int big_endian,
-				enum bch_validate_flags flags)
+				struct bkey_validate_context from)
 {
 	struct jset_entry_dev_usage *u =
 		container_of(entry, struct jset_entry_dev_usage, entry);
@@ -739,7 +733,7 @@ static int journal_entry_log_validate(struct bch_fs *c,
 				struct jset *jset,
 				struct jset_entry *entry,
 				unsigned version, int big_endian,
-				enum bch_validate_flags flags)
+				struct bkey_validate_context from)
 {
 	return 0;
 }
@@ -756,10 +750,11 @@ static int journal_entry_overwrite_validate(struct bch_fs *c,
 				struct jset *jset,
 				struct jset_entry *entry,
 				unsigned version, int big_endian,
-				enum bch_validate_flags flags)
+				struct bkey_validate_context from)
 {
+	from.flags = 0;
 	return journal_entry_btree_keys_validate(c, jset, entry,
-				version, big_endian, READ);
+				version, big_endian, from);
 }
 
 static void journal_entry_overwrite_to_text(struct printbuf *out, struct bch_fs *c,
@@ -772,10 +767,10 @@ static int journal_entry_write_buffer_keys_validate(struct bch_fs *c,
 				struct jset *jset,
 				struct jset_entry *entry,
 				unsigned version, int big_endian,
-				enum bch_validate_flags flags)
+				struct bkey_validate_context from)
 {
 	return journal_entry_btree_keys_validate(c, jset, entry,
-				version, big_endian, READ);
+				version, big_endian, from);
 }
 
 static void journal_entry_write_buffer_keys_to_text(struct printbuf *out, struct bch_fs *c,
@@ -788,7 +783,7 @@ static int journal_entry_datetime_validate(struct bch_fs *c,
 				struct jset *jset,
 				struct jset_entry *entry,
 				unsigned version, int big_endian,
-				enum bch_validate_flags flags)
+				struct bkey_validate_context from)
 {
 	unsigned bytes = vstruct_bytes(entry);
 	unsigned expected = 16;
@@ -818,7 +813,7 @@ static void journal_entry_datetime_to_text(struct printbuf *out, struct bch_fs *
 struct jset_entry_ops {
 	int (*validate)(struct bch_fs *, struct jset *,
 			struct jset_entry *, unsigned, int,
-			enum bch_validate_flags);
+			struct bkey_validate_context);
 	void (*to_text)(struct printbuf *, struct bch_fs *, struct jset_entry *);
 };
 
@@ -836,11 +831,11 @@ int bch2_journal_entry_validate(struct bch_fs *c,
 				struct jset *jset,
 				struct jset_entry *entry,
 				unsigned version, int big_endian,
-				enum bch_validate_flags flags)
+				struct bkey_validate_context from)
 {
 	return entry->type < BCH_JSET_ENTRY_NR
 		? bch2_jset_entry_ops[entry->type].validate(c, jset, entry,
-				version, big_endian, flags)
+				version, big_endian, from)
 		: 0;
 }
 
@@ -858,10 +853,18 @@ void bch2_journal_entry_to_text(struct printbuf *out, struct bch_fs *c,
 static int jset_validate_entries(struct bch_fs *c, struct jset *jset,
 				 enum bch_validate_flags flags)
 {
+	struct bkey_validate_context from = {
+		.flags		= flags,
+		.from		= BKEY_VALIDATE_journal,
+		.journal_seq	= le64_to_cpu(jset->seq),
+	};
+
 	unsigned version = le32_to_cpu(jset->version);
 	int ret = 0;
 
 	vstruct_for_each(jset, entry) {
+		from.journal_offset = (u64 *) entry - jset->_data;
+
 		if (journal_entry_err_on(vstruct_next(entry) > vstruct_last(jset),
 				c, version, jset, entry,
 				journal_entry_past_jset_end,
@@ -870,8 +873,8 @@ static int jset_validate_entries(struct bch_fs *c, struct jset *jset,
 			break;
 		}
 
-		ret = bch2_journal_entry_validate(c, jset, entry,
-					version, JSET_BIG_ENDIAN(jset), flags);
+		ret = bch2_journal_entry_validate(c, jset, entry, version,
+						  JSET_BIG_ENDIAN(jset), from);
 		if (ret)
 			break;
 	}
@@ -884,13 +887,17 @@ static int jset_validate(struct bch_fs *c,
 			 struct jset *jset, u64 sector,
 			 enum bch_validate_flags flags)
 {
-	unsigned version;
+	struct bkey_validate_context from = {
+		.flags		= flags,
+		.from		= BKEY_VALIDATE_journal,
+		.journal_seq	= le64_to_cpu(jset->seq),
+	};
 	int ret = 0;
 
 	if (le64_to_cpu(jset->magic) != jset_magic(c))
 		return JOURNAL_ENTRY_NONE;
 
-	version = le32_to_cpu(jset->version);
+	unsigned version = le32_to_cpu(jset->version);
 	if (journal_entry_err_on(!bch2_version_compatible(version),
 			c, version, jset, NULL,
 			jset_unsupported_version,
@@ -935,15 +942,16 @@ static int jset_validate_early(struct bch_fs *c,
 			 unsigned bucket_sectors_left,
 			 unsigned sectors_read)
 {
-	size_t bytes = vstruct_bytes(jset);
-	unsigned version;
-	enum bch_validate_flags flags = BCH_VALIDATE_journal;
+	struct bkey_validate_context from = {
+		.from		= BKEY_VALIDATE_journal,
+		.journal_seq	= le64_to_cpu(jset->seq),
+	};
 	int ret = 0;
 
 	if (le64_to_cpu(jset->magic) != jset_magic(c))
 		return JOURNAL_ENTRY_NONE;
 
-	version = le32_to_cpu(jset->version);
+	unsigned version = le32_to_cpu(jset->version);
 	if (journal_entry_err_on(!bch2_version_compatible(version),
 			c, version, jset, NULL,
 			jset_unsupported_version,
@@ -956,6 +964,7 @@ static int jset_validate_early(struct bch_fs *c,
 		return -EINVAL;
 	}
 
+	size_t bytes = vstruct_bytes(jset);
 	if (bytes > (sectors_read << 9) &&
 	    sectors_read < bucket_sectors_left)
 		return JOURNAL_ENTRY_REREAD;
@@ -1240,8 +1249,6 @@ int bch2_journal_read(struct bch_fs *c,
 	 * those entries will be blacklisted:
 	 */
 	genradix_for_each_reverse(&c->journal_entries, radix_iter, _i) {
-		enum bch_validate_flags flags = BCH_VALIDATE_journal;
-
 		i = *_i;
 
 		if (journal_replay_ignore(i))
@@ -1261,6 +1268,10 @@ int bch2_journal_read(struct bch_fs *c,
 			continue;
 		}
 
+		struct bkey_validate_context from = {
+			.from		= BKEY_VALIDATE_journal,
+			.journal_seq	= le64_to_cpu(i->j.seq),
+		};
 		if (journal_entry_err_on(le64_to_cpu(i->j.last_seq) > le64_to_cpu(i->j.seq),
 					 c, le32_to_cpu(i->j.version), &i->j, NULL,
 					 jset_last_seq_newer_than_seq,
@@ -1420,27 +1431,50 @@ fsck_err:
 
 /* journal write: */
 
+static void journal_advance_devs_to_next_bucket(struct journal *j,
+						struct dev_alloc_list *devs,
+						unsigned sectors, u64 seq)
+{
+	struct bch_fs *c = container_of(j, struct bch_fs, journal);
+
+	darray_for_each(*devs, i) {
+		struct bch_dev *ca = rcu_dereference(c->devs[*i]);
+		if (!ca)
+			continue;
+
+		struct journal_device *ja = &ca->journal;
+
+		if (sectors > ja->sectors_free &&
+		    sectors <= ca->mi.bucket_size &&
+		    bch2_journal_dev_buckets_available(j, ja,
+					journal_space_discarded)) {
+			ja->cur_idx = (ja->cur_idx + 1) % ja->nr;
+			ja->sectors_free = ca->mi.bucket_size;
+
+			/*
+			 * ja->bucket_seq[ja->cur_idx] must always have
+			 * something sensible:
+			 */
+			ja->bucket_seq[ja->cur_idx] = le64_to_cpu(seq);
+		}
+	}
+}
+
 static void __journal_write_alloc(struct journal *j,
 				  struct journal_buf *w,
-				  struct dev_alloc_list *devs_sorted,
+				  struct dev_alloc_list *devs,
 				  unsigned sectors,
 				  unsigned *replicas,
 				  unsigned replicas_want)
 {
 	struct bch_fs *c = container_of(j, struct bch_fs, journal);
-	struct journal_device *ja;
-	struct bch_dev *ca;
-	unsigned i;
-
-	if (*replicas >= replicas_want)
-		return;
 
-	for (i = 0; i < devs_sorted->nr; i++) {
-		ca = rcu_dereference(c->devs[devs_sorted->devs[i]]);
+	darray_for_each(*devs, i) {
+		struct bch_dev *ca = rcu_dereference(c->devs[*i]);
 		if (!ca)
 			continue;
 
-		ja = &ca->journal;
+		struct journal_device *ja = &ca->journal;
 
 		/*
 		 * Check that we can use this device, and aren't already using
@@ -1486,65 +1520,53 @@ static int journal_write_alloc(struct journal *j, struct journal_buf *w)
 {
 	struct bch_fs *c = container_of(j, struct bch_fs, journal);
 	struct bch_devs_mask devs;
-	struct journal_device *ja;
-	struct bch_dev *ca;
 	struct dev_alloc_list devs_sorted;
 	unsigned sectors = vstruct_sectors(w->data, c->block_bits);
 	unsigned target = c->opts.metadata_target ?:
 		c->opts.foreground_target;
-	unsigned i, replicas = 0, replicas_want =
+	unsigned replicas = 0, replicas_want =
 		READ_ONCE(c->opts.metadata_replicas);
 	unsigned replicas_need = min_t(unsigned, replicas_want,
 				       READ_ONCE(c->opts.metadata_replicas_required));
+	bool advance_done = false;
 
 	rcu_read_lock();
-retry:
-	devs = target_rw_devs(c, BCH_DATA_journal, target);
 
-	devs_sorted = bch2_dev_alloc_list(c, &j->wp.stripe, &devs);
+	/* We might run more than once if we have to stop and do discards: */
+	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(bkey_i_to_s_c(&w->key));
+	bkey_for_each_ptr(ptrs, p) {
+		struct bch_dev *ca = bch2_dev_rcu_noerror(c, p->dev);
+		if (ca)
+			replicas += ca->mi.durability;
+	}
 
-	__journal_write_alloc(j, w, &devs_sorted,
-			      sectors, &replicas, replicas_want);
+retry_target:
+	devs = target_rw_devs(c, BCH_DATA_journal, target);
+	devs_sorted = bch2_dev_alloc_list(c, &j->wp.stripe, &devs);
+retry_alloc:
+	__journal_write_alloc(j, w, &devs_sorted, sectors, &replicas, replicas_want);
 
-	if (replicas >= replicas_want)
+	if (likely(replicas >= replicas_want))
 		goto done;
 
-	for (i = 0; i < devs_sorted.nr; i++) {
-		ca = rcu_dereference(c->devs[devs_sorted.devs[i]]);
-		if (!ca)
-			continue;
-
-		ja = &ca->journal;
-
-		if (sectors > ja->sectors_free &&
-		    sectors <= ca->mi.bucket_size &&
-		    bch2_journal_dev_buckets_available(j, ja,
-					journal_space_discarded)) {
-			ja->cur_idx = (ja->cur_idx + 1) % ja->nr;
-			ja->sectors_free = ca->mi.bucket_size;
-
-			/*
-			 * ja->bucket_seq[ja->cur_idx] must always have
-			 * something sensible:
-			 */
-			ja->bucket_seq[ja->cur_idx] = le64_to_cpu(w->data->seq);
-		}
+	if (!advance_done) {
+		journal_advance_devs_to_next_bucket(j, &devs_sorted, sectors, w->data->seq);
+		advance_done = true;
+		goto retry_alloc;
 	}
 
-	__journal_write_alloc(j, w, &devs_sorted,
-			      sectors, &replicas, replicas_want);
-
 	if (replicas < replicas_want && target) {
 		/* Retry from all devices: */
 		target = 0;
-		goto retry;
+		advance_done = false;
+		goto retry_target;
 	}
 done:
 	rcu_read_unlock();
 
 	BUG_ON(bkey_val_u64s(&w->key.k) > BCH_REPLICAS_MAX);
 
-	return replicas >= replicas_need ? 0 : -EROFS;
+	return replicas >= replicas_need ? 0 : -BCH_ERR_insufficient_journal_devices;
 }
 
 static void journal_buf_realloc(struct journal *j, struct journal_buf *buf)
@@ -2032,7 +2054,7 @@ CLOSURE_CALLBACK(bch2_journal_write)
 		bch2_journal_do_discards(j);
 	}
 
-	if (ret) {
+	if (ret && !bch2_journal_error(j)) {
 		struct printbuf buf = PRINTBUF;
 		buf.atomic++;
 
@@ -2044,8 +2066,9 @@ CLOSURE_CALLBACK(bch2_journal_write)
 		spin_unlock(&j->lock);
 		bch2_print_string_as_lines(KERN_ERR, buf.buf);
 		printbuf_exit(&buf);
-		goto err;
 	}
+	if (ret)
+		goto err;
 
 	/*
 	 * write is allocated, no longer need to account for it in
diff --git a/libbcachefs/journal_io.h b/libbcachefs/journal_io.h
index 2ca9cde3..12b39fcb 100644
--- a/libbcachefs/journal_io.h
+++ b/libbcachefs/journal_io.h
@@ -63,7 +63,7 @@ static inline struct jset_entry *__jset_entry_type_next(struct jset *jset,
 
 int bch2_journal_entry_validate(struct bch_fs *, struct jset *,
 				struct jset_entry *, unsigned, int,
-				enum bch_validate_flags);
+				struct bkey_validate_context);
 void bch2_journal_entry_to_text(struct printbuf *, struct bch_fs *,
 				struct jset_entry *);
 
diff --git a/libbcachefs/journal_reclaim.c b/libbcachefs/journal_reclaim.c
index b7936ad3..3c824260 100644
--- a/libbcachefs/journal_reclaim.c
+++ b/libbcachefs/journal_reclaim.c
@@ -146,7 +146,8 @@ static struct journal_space __journal_space_available(struct journal *j, unsigne
 
 	rcu_read_lock();
 	for_each_member_device_rcu(c, ca, &c->rw_devs[BCH_DATA_journal]) {
-		if (!ca->journal.nr)
+		if (!ca->journal.nr ||
+		    !ca->mi.durability)
 			continue;
 
 		min_bucket_size = min(min_bucket_size, ca->mi.bucket_size);
diff --git a/libbcachefs/logged_ops.c b/libbcachefs/logged_ops.c
index 1ac51af1..75f27ec2 100644
--- a/libbcachefs/logged_ops.c
+++ b/libbcachefs/logged_ops.c
@@ -65,7 +65,8 @@ int bch2_resume_logged_ops(struct bch_fs *c)
 	int ret = bch2_trans_run(c,
 		for_each_btree_key_max(trans, iter,
 				   BTREE_ID_logged_ops,
-				   POS(LOGGED_OPS_INUM, 0), POS(LOGGED_OPS_INUM, U64_MAX),
+				   POS(LOGGED_OPS_INUM_logged_ops, 0),
+				   POS(LOGGED_OPS_INUM_logged_ops, U64_MAX),
 				   BTREE_ITER_prefetch, k,
 			resume_logged_op(trans, &iter, k)));
 	bch_err_fn(c, ret);
@@ -76,7 +77,7 @@ static int __bch2_logged_op_start(struct btree_trans *trans, struct bkey_i *k)
 {
 	struct btree_iter iter;
 	int ret = bch2_bkey_get_empty_slot(trans, &iter,
-				 BTREE_ID_logged_ops, POS(LOGGED_OPS_INUM, U64_MAX));
+				 BTREE_ID_logged_ops, POS(LOGGED_OPS_INUM_logged_ops, U64_MAX));
 	if (ret)
 		return ret;
 
diff --git a/libbcachefs/logged_ops_format.h b/libbcachefs/logged_ops_format.h
index 0b370a96..cfb67c95 100644
--- a/libbcachefs/logged_ops_format.h
+++ b/libbcachefs/logged_ops_format.h
@@ -2,7 +2,10 @@
 #ifndef _BCACHEFS_LOGGED_OPS_FORMAT_H
 #define _BCACHEFS_LOGGED_OPS_FORMAT_H
 
-#define LOGGED_OPS_INUM		0
+enum logged_ops_inums {
+	LOGGED_OPS_INUM_logged_ops,
+	LOGGED_OPS_INUM_inode_cursors,
+};
 
 struct bch_logged_op_truncate {
 	struct bch_val		v;
diff --git a/libbcachefs/move.c b/libbcachefs/move.c
index a50f27f4..9b066c61 100644
--- a/libbcachefs/move.c
+++ b/libbcachefs/move.c
@@ -825,7 +825,7 @@ int bch2_evacuate_bucket(struct moving_context *ctxt,
 			b = bch2_backpointer_get_node(trans, bp, &iter, &last_flushed);
 			ret = PTR_ERR_OR_ZERO(b);
 			if (ret == -BCH_ERR_backpointer_to_overwritten_btree_node)
-				continue;
+				goto next;
 			if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 				continue;
 			if (ret)
diff --git a/libbcachefs/opts.h b/libbcachefs/opts.h
index ea69099e..2df5c8f7 100644
--- a/libbcachefs/opts.h
+++ b/libbcachefs/opts.h
@@ -222,15 +222,20 @@ enum fsck_err_opts {
 	  BCH_SB_ERASURE_CODE,		false,				\
 	  NULL,		"Enable erasure coding (DO NOT USE YET)")	\
 	x(inodes_32bit,			u8,				\
-	  OPT_FS|OPT_FORMAT|OPT_MOUNT|OPT_RUNTIME,			\
+	  OPT_FS|OPT_FORMAT,						\
 	  OPT_BOOL(),							\
 	  BCH_SB_INODE_32BIT,		true,				\
 	  NULL,		"Constrain inode numbers to 32 bits")		\
 	x(shard_inode_numbers,		u8,				\
-	  OPT_FS|OPT_FORMAT|OPT_MOUNT|OPT_RUNTIME,			\
+	  OPT_FS|OPT_FORMAT,						\
 	  OPT_BOOL(),							\
 	  BCH_SB_SHARD_INUMS,		true,				\
 	  NULL,		"Shard new inode numbers by CPU id")		\
+	x(shard_inode_numbers_bits,	u8,				\
+	  OPT_FS|OPT_FORMAT,						\
+	  OPT_UINT(0, 8),						\
+	  BCH_SB_SHARD_INUMS_NBITS,	0,				\
+	  NULL,		"Shard new inode numbers by CPU id")		\
 	x(inodes_use_key_cache,	u8,					\
 	  OPT_FS|OPT_FORMAT|OPT_MOUNT,					\
 	  OPT_BOOL(),							\
diff --git a/libbcachefs/recovery_passes.c b/libbcachefs/recovery_passes.c
index f6d3a99c..0b3c951c 100644
--- a/libbcachefs/recovery_passes.c
+++ b/libbcachefs/recovery_passes.c
@@ -103,27 +103,31 @@ u64 bch2_recovery_passes_from_stable(u64 v)
 static int __bch2_run_explicit_recovery_pass(struct bch_fs *c,
 					     enum bch_recovery_pass pass)
 {
-	if (c->opts.recovery_passes & BIT_ULL(pass))
-		return 0;
-
 	if (c->curr_recovery_pass == ARRAY_SIZE(recovery_pass_fns))
 		return -BCH_ERR_not_in_recovery;
 
+	if (c->recovery_passes_complete & BIT_ULL(pass))
+		return 0;
+
+	bool print = !(c->opts.recovery_passes & BIT_ULL(pass));
+
 	if (pass < BCH_RECOVERY_PASS_set_may_go_rw &&
 	    c->curr_recovery_pass >= BCH_RECOVERY_PASS_set_may_go_rw) {
-		bch_info(c, "need recovery pass %s (%u), but already rw",
-			 bch2_recovery_passes[pass], pass);
+		if (print)
+			bch_info(c, "need recovery pass %s (%u), but already rw",
+				 bch2_recovery_passes[pass], pass);
 		return -BCH_ERR_cannot_rewind_recovery;
 	}
 
-	bch_info(c, "running explicit recovery pass %s (%u), currently at %s (%u)",
-		 bch2_recovery_passes[pass], pass,
-		 bch2_recovery_passes[c->curr_recovery_pass], c->curr_recovery_pass);
+	if (print)
+		bch_info(c, "running explicit recovery pass %s (%u), currently at %s (%u)",
+			 bch2_recovery_passes[pass], pass,
+			 bch2_recovery_passes[c->curr_recovery_pass], c->curr_recovery_pass);
 
 	c->opts.recovery_passes |= BIT_ULL(pass);
 
-	if (c->curr_recovery_pass >= pass) {
-		c->curr_recovery_pass = pass;
+	if (c->curr_recovery_pass > pass) {
+		c->next_recovery_pass = pass;
 		c->recovery_passes_complete &= (1ULL << pass) >> 1;
 		return -BCH_ERR_restart_recovery;
 	} else {
@@ -264,7 +268,9 @@ int bch2_run_recovery_passes(struct bch_fs *c)
 	 */
 	c->opts.recovery_passes_exclude &= ~BCH_RECOVERY_PASS_set_may_go_rw;
 
-	while (c->curr_recovery_pass < ARRAY_SIZE(recovery_pass_fns)) {
+	while (c->curr_recovery_pass < ARRAY_SIZE(recovery_pass_fns) && !ret) {
+		c->next_recovery_pass = c->curr_recovery_pass + 1;
+
 		spin_lock_irq(&c->recovery_pass_lock);
 		unsigned pass = c->curr_recovery_pass;
 
@@ -285,31 +291,25 @@ int bch2_run_recovery_passes(struct bch_fs *c)
 		ret =   bch2_run_recovery_pass(c, pass) ?:
 			bch2_journal_flush(&c->journal);
 
+		if (!ret && !test_bit(BCH_FS_error, &c->flags))
+			bch2_clear_recovery_pass_required(c, pass);
+
 		spin_lock_irq(&c->recovery_pass_lock);
-		if (c->curr_recovery_pass < pass) {
+		if (c->next_recovery_pass < c->curr_recovery_pass) {
 			/*
 			 * bch2_run_explicit_recovery_pass() was called: we
 			 * can't always catch -BCH_ERR_restart_recovery because
 			 * it may have been called from another thread (btree
 			 * node read completion)
 			 */
-			spin_unlock_irq(&c->recovery_pass_lock);
-			continue;
-		} else if (c->curr_recovery_pass == pass) {
-			c->curr_recovery_pass++;
+			ret = 0;
+			c->recovery_passes_complete &= ~(~0ULL << c->curr_recovery_pass);
 		} else {
-			BUG();
+			c->recovery_passes_complete |= BIT_ULL(pass);
+			c->recovery_pass_done = max(c->recovery_pass_done, pass);
 		}
+		c->curr_recovery_pass = c->next_recovery_pass;
 		spin_unlock_irq(&c->recovery_pass_lock);
-
-		if (ret)
-			break;
-
-		c->recovery_passes_complete |= BIT_ULL(pass);
-		c->recovery_pass_done = max(c->recovery_pass_done, pass);
-
-		if (!test_bit(BCH_FS_error, &c->flags))
-			bch2_clear_recovery_pass_required(c, pass);
 	}
 
 	return ret;
diff --git a/libbcachefs/sb-clean.c b/libbcachefs/sb-clean.c
index 00527528..59c8770e 100644
--- a/libbcachefs/sb-clean.c
+++ b/libbcachefs/sb-clean.c
@@ -23,6 +23,10 @@
 int bch2_sb_clean_validate_late(struct bch_fs *c, struct bch_sb_field_clean *clean,
 				int write)
 {
+	struct bkey_validate_context from = {
+		.flags		= write,
+		.from		= BKEY_VALIDATE_superblock,
+	};
 	struct jset_entry *entry;
 	int ret;
 
@@ -40,7 +44,7 @@ int bch2_sb_clean_validate_late(struct bch_fs *c, struct bch_sb_field_clean *cle
 		ret = bch2_journal_entry_validate(c, NULL, entry,
 						  le16_to_cpu(c->disk_sb.sb->version),
 						  BCH_SB_BIG_ENDIAN(c->disk_sb.sb),
-						  write);
+						  from);
 		if (ret)
 			return ret;
 	}
diff --git a/libbcachefs/sb-errors_format.h b/libbcachefs/sb-errors_format.h
index 4248ff3c..e26317c3 100644
--- a/libbcachefs/sb-errors_format.h
+++ b/libbcachefs/sb-errors_format.h
@@ -58,7 +58,7 @@ enum bch_fsck_flags {
 	x(bset_empty,						 45,	0)		\
 	x(bset_bad_seq,						 46,	0)		\
 	x(bset_blacklisted_journal_seq,				 47,	0)		\
-	x(first_bset_blacklisted_journal_seq,			 48,	0)		\
+	x(first_bset_blacklisted_journal_seq,			 48,	FSCK_AUTOFIX)	\
 	x(btree_node_bad_btree,					 49,	0)		\
 	x(btree_node_bad_level,					 50,	0)		\
 	x(btree_node_bad_min_key,				 51,	0)		\
@@ -168,7 +168,7 @@ enum bch_fsck_flags {
 	x(ptr_to_incorrect_stripe,				151,	0)		\
 	x(ptr_gen_newer_than_bucket_gen,			152,	0)		\
 	x(ptr_too_stale,					153,	0)		\
-	x(stale_dirty_ptr,					154,	0)		\
+	x(stale_dirty_ptr,					154,	FSCK_AUTOFIX)	\
 	x(ptr_bucket_data_type_mismatch,			155,	0)		\
 	x(ptr_cached_and_erasure_coded,				156,	0)		\
 	x(ptr_crc_uncompressed_size_too_small,			157,	0)		\
@@ -211,6 +211,7 @@ enum bch_fsck_flags {
 	x(bkey_in_missing_snapshot,				190,	0)		\
 	x(inode_pos_inode_nonzero,				191,	0)		\
 	x(inode_pos_blockdev_range,				192,	0)		\
+	x(inode_alloc_cursor_inode_bad,				301,	0)		\
 	x(inode_unpack_error,					193,	0)		\
 	x(inode_str_hash_invalid,				194,	0)		\
 	x(inode_v3_fields_start_bad,				195,	0)		\
@@ -311,7 +312,7 @@ enum bch_fsck_flags {
 	x(logged_op_but_clean,					283,	FSCK_AUTOFIX)	\
 	x(compression_opt_not_marked_in_sb,			295,	FSCK_AUTOFIX)	\
 	x(compression_type_not_marked_in_sb,			296,	FSCK_AUTOFIX)	\
-	x(MAX,							301,	0)
+	x(MAX,							302,	0)
 
 enum bch_sb_error_id {
 #define x(t, n, ...) BCH_FSCK_ERR_##t = n,
diff --git a/libbcachefs/snapshot.c b/libbcachefs/snapshot.c
index f368270d..99f04551 100644
--- a/libbcachefs/snapshot.c
+++ b/libbcachefs/snapshot.c
@@ -570,8 +570,7 @@ static int check_snapshot_tree(struct btree_trans *trans,
 		goto err;
 	}
 
-	ret = bch2_subvolume_get(trans, le32_to_cpu(st.v->master_subvol),
-				 false, 0, &subvol);
+	ret = bch2_subvolume_get(trans, le32_to_cpu(st.v->master_subvol), false, &subvol);
 	if (ret && !bch2_err_matches(ret, ENOENT))
 		goto err;
 
@@ -811,7 +810,7 @@ static int check_snapshot(struct btree_trans *trans,
 
 	if (should_have_subvol) {
 		id = le32_to_cpu(s.subvol);
-		ret = bch2_subvolume_get(trans, id, 0, false, &subvol);
+		ret = bch2_subvolume_get(trans, id, false, &subvol);
 		if (bch2_err_matches(ret, ENOENT))
 			bch_err(c, "snapshot points to nonexistent subvolume:\n  %s",
 				(bch2_bkey_val_to_text(&buf, c, k), buf.buf));
diff --git a/libbcachefs/str_hash.c b/libbcachefs/str_hash.c
new file mode 100644
index 00000000..c3276a7e
--- /dev/null
+++ b/libbcachefs/str_hash.c
@@ -0,0 +1,209 @@
+// SPDX-License-Identifier: GPL-2.0
+
+#include "bcachefs.h"
+#include "btree_cache.h"
+#include "btree_update.h"
+#include "dirent.h"
+#include "fsck.h"
+#include "str_hash.h"
+#include "subvolume.h"
+
+static int bch2_dirent_has_target(struct btree_trans *trans, struct bkey_s_c_dirent d)
+{
+	if (d.v->d_type == DT_SUBVOL) {
+		struct bch_subvolume subvol;
+		int ret = bch2_subvolume_get(trans, le32_to_cpu(d.v->d_child_subvol),
+					     false, &subvol);
+		if (ret && !bch2_err_matches(ret, ENOENT))
+			return ret;
+		return !ret;
+	} else {
+		struct btree_iter iter;
+		struct bkey_s_c k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_inodes,
+				SPOS(0, le64_to_cpu(d.v->d_inum), d.k->p.snapshot), 0);
+		int ret = bkey_err(k);
+		if (ret)
+			return ret;
+
+		ret = bkey_is_inode(k.k);
+		bch2_trans_iter_exit(trans, &iter);
+		return ret;
+	}
+}
+
+static int fsck_rename_dirent(struct btree_trans *trans,
+			      struct snapshots_seen *s,
+			      const struct bch_hash_desc desc,
+			      struct bch_hash_info *hash_info,
+			      struct bkey_s_c_dirent old)
+{
+	struct qstr old_name = bch2_dirent_get_name(old);
+	struct bkey_i_dirent *new = bch2_trans_kmalloc(trans, bkey_bytes(old.k) + 32);
+	int ret = PTR_ERR_OR_ZERO(new);
+	if (ret)
+		return ret;
+
+	bkey_dirent_init(&new->k_i);
+	dirent_copy_target(new, old);
+	new->k.p = old.k->p;
+
+	for (unsigned i = 0; i < 1000; i++) {
+		unsigned len = sprintf(new->v.d_name, "%.*s.fsck_renamed-%u",
+				       old_name.len, old_name.name, i);
+		unsigned u64s = BKEY_U64s + dirent_val_u64s(len);
+
+		if (u64s > U8_MAX)
+			return -EINVAL;
+
+		new->k.u64s = u64s;
+
+		ret = bch2_hash_set_in_snapshot(trans, bch2_dirent_hash_desc, hash_info,
+						(subvol_inum) { 0, old.k->p.inode },
+						old.k->p.snapshot, &new->k_i,
+						BTREE_UPDATE_internal_snapshot_node);
+		if (!bch2_err_matches(ret, EEXIST))
+			break;
+	}
+
+	if (ret)
+		return ret;
+
+	return bch2_fsck_update_backpointers(trans, s, desc, hash_info, &new->k_i);
+}
+
+static int hash_pick_winner(struct btree_trans *trans,
+			    const struct bch_hash_desc desc,
+			    struct bch_hash_info *hash_info,
+			    struct bkey_s_c k1,
+			    struct bkey_s_c k2)
+{
+	if (bkey_val_bytes(k1.k) == bkey_val_bytes(k2.k) &&
+	    !memcmp(k1.v, k2.v, bkey_val_bytes(k1.k)))
+		return 0;
+
+	switch (desc.btree_id) {
+	case BTREE_ID_dirents: {
+		int ret = bch2_dirent_has_target(trans, bkey_s_c_to_dirent(k1));
+		if (ret < 0)
+			return ret;
+		if (!ret)
+			return 0;
+
+		ret = bch2_dirent_has_target(trans, bkey_s_c_to_dirent(k2));
+		if (ret < 0)
+			return ret;
+		if (!ret)
+			return 1;
+		return 2;
+	}
+	default:
+		return 0;
+	}
+}
+
+int bch2_str_hash_check_key(struct btree_trans *trans,
+			    struct snapshots_seen *s,
+			    const struct bch_hash_desc desc,
+			    struct bch_hash_info *hash_info,
+			    struct btree_iter *k_iter, struct bkey_s_c hash_k)
+{
+	struct bch_fs *c = trans->c;
+	struct btree_iter iter = { NULL };
+	struct printbuf buf = PRINTBUF;
+	struct bkey_s_c k;
+	u64 hash;
+	int ret = 0;
+
+	if (hash_k.k->type != desc.key_type)
+		return 0;
+
+	hash = desc.hash_bkey(hash_info, hash_k);
+
+	if (likely(hash == hash_k.k->p.offset))
+		return 0;
+
+	if (hash_k.k->p.offset < hash)
+		goto bad_hash;
+
+	for_each_btree_key_norestart(trans, iter, desc.btree_id,
+				     SPOS(hash_k.k->p.inode, hash, hash_k.k->p.snapshot),
+				     BTREE_ITER_slots, k, ret) {
+		if (bkey_eq(k.k->p, hash_k.k->p))
+			break;
+
+		if (k.k->type == desc.key_type &&
+		    !desc.cmp_bkey(k, hash_k))
+			goto duplicate_entries;
+
+		if (bkey_deleted(k.k)) {
+			bch2_trans_iter_exit(trans, &iter);
+			goto bad_hash;
+		}
+	}
+out:
+	bch2_trans_iter_exit(trans, &iter);
+	printbuf_exit(&buf);
+	return ret;
+bad_hash:
+	if (fsck_err(trans, hash_table_key_wrong_offset,
+		     "hash table key at wrong offset: btree %s inode %llu offset %llu, hashed to %llu\n  %s",
+		     bch2_btree_id_str(desc.btree_id), hash_k.k->p.inode, hash_k.k->p.offset, hash,
+		     (printbuf_reset(&buf),
+		      bch2_bkey_val_to_text(&buf, c, hash_k), buf.buf))) {
+		struct bkey_i *new = bch2_bkey_make_mut_noupdate(trans, hash_k);
+		if (IS_ERR(new))
+			return PTR_ERR(new);
+
+		k = bch2_hash_set_or_get_in_snapshot(trans, &iter, desc, hash_info,
+				       (subvol_inum) { 0, hash_k.k->p.inode },
+				       hash_k.k->p.snapshot, new,
+				       STR_HASH_must_create|
+				       BTREE_ITER_with_updates|
+				       BTREE_UPDATE_internal_snapshot_node);
+		ret = bkey_err(k);
+		if (ret)
+			goto out;
+		if (k.k)
+			goto duplicate_entries;
+
+		ret =   bch2_hash_delete_at(trans, desc, hash_info, k_iter,
+					    BTREE_UPDATE_internal_snapshot_node) ?:
+			bch2_fsck_update_backpointers(trans, s, desc, hash_info, new) ?:
+			bch2_trans_commit(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc) ?:
+			-BCH_ERR_transaction_restart_nested;
+		goto out;
+	}
+fsck_err:
+	goto out;
+duplicate_entries:
+	ret = hash_pick_winner(trans, desc, hash_info, hash_k, k);
+	if (ret < 0)
+		goto out;
+
+	if (!fsck_err(trans, hash_table_key_duplicate,
+		      "duplicate hash table keys%s:\n%s",
+		      ret != 2 ? "" : ", both point to valid inodes",
+		      (printbuf_reset(&buf),
+		       bch2_bkey_val_to_text(&buf, c, hash_k),
+		       prt_newline(&buf),
+		       bch2_bkey_val_to_text(&buf, c, k),
+		       buf.buf)))
+		goto out;
+
+	switch (ret) {
+	case 0:
+		ret = bch2_hash_delete_at(trans, desc, hash_info, k_iter, 0);
+		break;
+	case 1:
+		ret = bch2_hash_delete_at(trans, desc, hash_info, &iter, 0);
+		break;
+	case 2:
+		ret = fsck_rename_dirent(trans, s, desc, hash_info, bkey_s_c_to_dirent(hash_k)) ?:
+			bch2_hash_delete_at(trans, desc, hash_info, k_iter, 0);
+		goto out;
+	}
+
+	ret = bch2_trans_commit(trans, NULL, NULL, 0) ?:
+		-BCH_ERR_transaction_restart_nested;
+	goto out;
+}
diff --git a/libbcachefs/str_hash.h b/libbcachefs/str_hash.h
index 00c78505..0c20f3af 100644
--- a/libbcachefs/str_hash.h
+++ b/libbcachefs/str_hash.h
@@ -393,4 +393,11 @@ int bch2_hash_delete(struct btree_trans *trans,
 	return ret;
 }
 
+struct snapshots_seen;
+int bch2_str_hash_check_key(struct btree_trans *,
+			    struct snapshots_seen *,
+			    const struct bch_hash_desc,
+			    struct bch_hash_info *,
+			    struct btree_iter *, struct bkey_s_c);
+
 #endif /* _BCACHEFS_STR_HASH_H */
diff --git a/libbcachefs/subvolume.c b/libbcachefs/subvolume.c
index 5e5ae405..0e756e35 100644
--- a/libbcachefs/subvolume.c
+++ b/libbcachefs/subvolume.c
@@ -286,11 +286,11 @@ int bch2_subvol_has_children(struct btree_trans *trans, u32 subvol)
 static __always_inline int
 bch2_subvolume_get_inlined(struct btree_trans *trans, unsigned subvol,
 			   bool inconsistent_if_not_found,
-			   int iter_flags,
 			   struct bch_subvolume *s)
 {
 	int ret = bch2_bkey_get_val_typed(trans, BTREE_ID_subvolumes, POS(0, subvol),
-					  iter_flags, subvolume, s);
+					  BTREE_ITER_cached|
+					  BTREE_ITER_with_updates, subvolume, s);
 	bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOENT) &&
 				inconsistent_if_not_found,
 				trans->c, "missing subvolume %u", subvol);
@@ -299,16 +299,15 @@ bch2_subvolume_get_inlined(struct btree_trans *trans, unsigned subvol,
 
 int bch2_subvolume_get(struct btree_trans *trans, unsigned subvol,
 		       bool inconsistent_if_not_found,
-		       int iter_flags,
 		       struct bch_subvolume *s)
 {
-	return bch2_subvolume_get_inlined(trans, subvol, inconsistent_if_not_found, iter_flags, s);
+	return bch2_subvolume_get_inlined(trans, subvol, inconsistent_if_not_found, s);
 }
 
 int bch2_subvol_is_ro_trans(struct btree_trans *trans, u32 subvol)
 {
 	struct bch_subvolume s;
-	int ret = bch2_subvolume_get_inlined(trans, subvol, true, 0, &s);
+	int ret = bch2_subvolume_get_inlined(trans, subvol, true, &s);
 	if (ret)
 		return ret;
 
@@ -328,7 +327,7 @@ int bch2_snapshot_get_subvol(struct btree_trans *trans, u32 snapshot,
 	struct bch_snapshot snap;
 
 	return  bch2_snapshot_lookup(trans, snapshot, &snap) ?:
-		bch2_subvolume_get(trans, le32_to_cpu(snap.subvol), true, 0, subvol);
+		bch2_subvolume_get(trans, le32_to_cpu(snap.subvol), true, subvol);
 }
 
 int __bch2_subvolume_get_snapshot(struct btree_trans *trans, u32 subvolid,
@@ -396,8 +395,7 @@ static int bch2_subvolumes_reparent(struct btree_trans *trans, u32 subvolid_to_d
 	struct bch_subvolume s;
 
 	return lockrestart_do(trans,
-			bch2_subvolume_get(trans, subvolid_to_delete, true,
-				   BTREE_ITER_cached, &s)) ?:
+			bch2_subvolume_get(trans, subvolid_to_delete, true, &s)) ?:
 		for_each_btree_key_commit(trans, iter,
 				BTREE_ID_subvolumes, POS_MIN, BTREE_ITER_prefetch, k,
 				NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
diff --git a/libbcachefs/subvolume.h b/libbcachefs/subvolume.h
index d53d292c..910f6196 100644
--- a/libbcachefs/subvolume.h
+++ b/libbcachefs/subvolume.h
@@ -24,7 +24,7 @@ int bch2_subvolume_trigger(struct btree_trans *, enum btree_id, unsigned,
 
 int bch2_subvol_has_children(struct btree_trans *, u32);
 int bch2_subvolume_get(struct btree_trans *, unsigned,
-		       bool, int, struct bch_subvolume *);
+		       bool, struct bch_subvolume *);
 int __bch2_subvolume_get_snapshot(struct btree_trans *, u32,
 				  u32 *, bool);
 int bch2_subvolume_get_snapshot(struct btree_trans *, u32, u32 *);
diff --git a/libbcachefs/super-io.c b/libbcachefs/super-io.c
index 912c3696..dbc09e30 100644
--- a/libbcachefs/super-io.c
+++ b/libbcachefs/super-io.c
@@ -432,7 +432,10 @@ static int bch2_sb_validate(struct bch_sb_handle *disk_sb,
 		bch2_version_to_text(out, BCH_SB_VERSION_INCOMPAT(sb));
 		prt_str(out, " > incompat_allowed ");
 		bch2_version_to_text(out, BCH_SB_VERSION_INCOMPAT_ALLOWED(sb));
-		return -BCH_ERR_invalid_sb_version;
+		if (flags & BCH_VALIDATE_write)
+			return -BCH_ERR_invalid_sb_version;
+		else
+			SET_BCH_SB_VERSION_INCOMPAT_ALLOWED(sb, BCH_SB_VERSION_INCOMPAT(sb));
 	}
 
 	if (!flags) {
@@ -457,6 +460,11 @@ static int bch2_sb_validate(struct bch_sb_handle *disk_sb,
 			SET_BCH_SB_PROMOTE_WHOLE_EXTENTS(sb, true);
 	}
 
+#ifdef __KERNEL__
+	if (!BCH_SB_SHARD_INUMS_NBITS(sb))
+		SET_BCH_SB_SHARD_INUMS_NBITS(sb, ilog2(roundup_pow_of_two(num_online_cpus())));
+#endif
+
 	for (opt_id = 0; opt_id < bch2_opts_nr; opt_id++) {
 		const struct bch_option *opt = bch2_opt_table + opt_id;
 
diff --git a/libbcachefs/super.c b/libbcachefs/super.c
index 14157820..bd6aecea 100644
--- a/libbcachefs/super.c
+++ b/libbcachefs/super.c
@@ -586,7 +586,6 @@ static void __bch2_fs_free(struct bch_fs *c)
 #endif
 	kfree(rcu_dereference_protected(c->disk_groups, 1));
 	kfree(c->journal_seq_blacklist_table);
-	kfree(c->unused_inode_hints);
 
 	if (c->write_ref_wq)
 		destroy_workqueue(c->write_ref_wq);
@@ -872,8 +871,6 @@ static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts)
 		(btree_blocks(c) + 1) * 2 *
 		sizeof(struct sort_iter_set);
 
-	c->inode_shard_bits = ilog2(roundup_pow_of_two(num_possible_cpus()));
-
 	if (!(c->btree_update_wq = alloc_workqueue("bcachefs",
 				WQ_HIGHPRI|WQ_FREEZABLE|WQ_MEM_RECLAIM|WQ_UNBOUND, 512)) ||
 	    !(c->btree_io_complete_wq = alloc_workqueue("bcachefs_btree_io",
@@ -900,9 +897,7 @@ static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts)
 	    !(c->online_reserved = alloc_percpu(u64)) ||
 	    mempool_init_kvmalloc_pool(&c->btree_bounce_pool, 1,
 				       c->opts.btree_node_size) ||
-	    mempool_init_kmalloc_pool(&c->large_bkey_pool, 1, 2048) ||
-	    !(c->unused_inode_hints = kcalloc(1U << c->inode_shard_bits,
-					      sizeof(u64), GFP_KERNEL))) {
+	    mempool_init_kmalloc_pool(&c->large_bkey_pool, 1, 2048)) {
 		ret = -BCH_ERR_ENOMEM_fs_other_alloc;
 		goto err;
 	}
diff --git a/libbcachefs/trace.h b/libbcachefs/trace.h
index 2d5932d2..7baf66be 100644
--- a/libbcachefs/trace.h
+++ b/libbcachefs/trace.h
@@ -199,6 +199,30 @@ DECLARE_EVENT_CLASS(bio,
 		  (unsigned long long)__entry->sector, __entry->nr_sector)
 );
 
+/* disk_accounting.c */
+
+TRACE_EVENT(accounting_mem_insert,
+	TP_PROTO(struct bch_fs *c, const char *acc),
+	TP_ARGS(c, acc),
+
+	TP_STRUCT__entry(
+		__field(dev_t,		dev			)
+		__field(unsigned,	new_nr			)
+		__string(acc,		acc			)
+	),
+
+	TP_fast_assign(
+		__entry->dev		= c->dev;
+		__entry->new_nr		= c->accounting.k.nr;
+		__assign_str(acc);
+	),
+
+	TP_printk("%d,%d entries %u added %s",
+		  MAJOR(__entry->dev), MINOR(__entry->dev),
+		  __entry->new_nr,
+		  __get_str(acc))
+);
+
 /* fs.c: */
 TRACE_EVENT(bch2_sync_fs,
 	TP_PROTO(struct super_block *sb, int wait),