bcachefs: Fix race between btree updates & journal replay

Add a flag to indicate whether a journal replay key has been
overwritten, and set/test it with appropriate btree locks held.

This fixes a race between the allocator - invalidating buckets, and
doing btree updates - and journal replay, which before this patch could
clobber the allocator thread's update with an older version of the key
from the journal.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>

fs/bcachefs/bcachefs.h

@@ -561,6 +561,7 @@ struct journal_keys {
 		enum btree_id	btree_id:8;
 		unsigned	level:8;
 		bool		allocated;
+		bool		overwritten;
 		struct bkey_i	*k;
 		u32		journal_seq;
 		u32		journal_offset;

fs/bcachefs/btree_update_leaf.c

@@ -15,6 +15,7 @@
 #include "journal.h"
 #include "journal_reclaim.h"
 #include "keylist.h"
+#include "recovery.h"
 #include "subvolume.h"
 #include "replicas.h"
 #include "trace.h"
@@ -625,6 +626,14 @@ static inline int trans_lock_write(struct btree_trans *trans)
 	return btree_trans_restart(trans);
 }
 
+static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans)
+{
+	struct btree_insert_entry *i;
+
+	trans_for_each_update(trans, i)
+		bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p);
+}
+
 /*
  * Get journal reservation, take write locks, and attempt to do btree update(s):
  */
@@ -702,6 +711,9 @@ static inline int do_bch2_trans_commit(struct btree_trans *trans,
 
 	ret = bch2_trans_commit_write_locked(trans, stopped_at, trace_ip);
 
+	if (!ret && unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)))
+		bch2_drop_overwrites_from_journal(trans);
+
 	trans_for_each_update(trans, i)
 		if (!same_leaf_as_prev(trans, i))
 			bch2_btree_node_unlock_write_inlined(trans, i->path,

fs/bcachefs/recovery.c

@@ -185,6 +185,19 @@ int bch2_journal_key_delete(struct bch_fs *c, enum btree_id id,
 	return bch2_journal_key_insert(c, id, level, &whiteout);
 }
 
+void bch2_journal_key_overwritten(struct bch_fs *c, enum btree_id btree,
+				  unsigned level, struct bpos pos)
+{
+	struct journal_keys *keys = &c->journal_keys;
+	size_t idx = journal_key_search(keys, btree, level, pos);
+
+	if (idx < keys->nr &&
+	    keys->d[idx].btree_id	== btree &&
+	    keys->d[idx].level		== level &&
+	    !bpos_cmp(keys->d[idx].k->k.p, pos))
+		keys->d[idx].overwritten = true;
+}
+
 static struct bkey_i *bch2_journal_iter_peek(struct journal_iter *iter)
 {
 	struct journal_key *k = iter->idx - iter->keys->nr
@@ -539,8 +552,16 @@ static int __bch2_journal_replay_key(struct btree_trans *trans,
 	bch2_trans_node_iter_init(trans, &iter, k->btree_id, k->k->k.p,
 				  BTREE_MAX_DEPTH, k->level,
 				  iter_flags);
-	ret   = bch2_btree_iter_traverse(&iter) ?:
-		bch2_trans_update(trans, &iter, k->k, BTREE_TRIGGER_NORUN);
+	ret = bch2_btree_iter_traverse(&iter);
+	if (ret)
+		goto out;
+
+	/* Must be checked with btree locked: */
+	if (k->overwritten)
+		goto out;
+
+	ret = bch2_trans_update(trans, &iter, k->k, BTREE_TRIGGER_NORUN);
+out:
 	bch2_trans_iter_exit(trans, &iter);
 	return ret;
 }

fs/bcachefs/recovery.h

@@ -37,6 +37,8 @@ int bch2_journal_key_insert(struct bch_fs *, enum btree_id,
 			    unsigned, struct bkey_i *);
 int bch2_journal_key_delete(struct bch_fs *, enum btree_id,
 			    unsigned, struct bpos);
+void bch2_journal_key_overwritten(struct bch_fs *, enum btree_id,
+				  unsigned, struct bpos);
 
 void bch2_btree_and_journal_iter_advance(struct btree_and_journal_iter *);
 struct bkey_s_c bch2_btree_and_journal_iter_peek(struct btree_and_journal_iter *);