bcachefs: Don't block on reclaim_lock from journal_res_get

When we're doing btree updates from journal flush, this becomes a
locking inversion

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

3 files changed, 20 insertions, 9 deletions

diff --git a/fs/bcachefs/journal.c b/fs/bcachefs/journal.c
index 5caa01881d00..ba6adf11ef42 100644
--- a/fs/bcachefs/journal.c
+++ b/fs/bcachefs/journal.c
@@ -390,7 +390,10 @@ retry:
 				goto retry;
 			}
 
-			bch2_journal_reclaim_work(&j->reclaim_work.work);
+			if (mutex_trylock(&j->reclaim_lock)) {
+				bch2_journal_reclaim(j);
+				mutex_unlock(&j->reclaim_lock);
+			}
 		}
 
 		ret = -EAGAIN;
diff --git a/fs/bcachefs/journal_reclaim.c b/fs/bcachefs/journal_reclaim.c
index 0884fc823cdf..a3c53b78ad10 100644
--- a/fs/bcachefs/journal_reclaim.c
+++ b/fs/bcachefs/journal_reclaim.c
@@ -433,7 +433,7 @@ static void journal_flush_pins(struct journal *j, u64 seq_to_flush,
 }
 
 /**
- * bch2_journal_reclaim_work - free up journal buckets
+ * bch2_journal_reclaim - free up journal buckets
  *
  * Background journal reclaim writes out btree nodes. It should be run
  * early enough so that we never completely run out of journal buckets.
@@ -450,18 +450,17 @@ static void journal_flush_pins(struct journal *j, u64 seq_to_flush,
  * 512 journal entries or 25% of all journal buckets, then
  * journal_next_bucket() should not stall.
  */
-void bch2_journal_reclaim_work(struct work_struct *work)
+void bch2_journal_reclaim(struct journal *j)
 {
-	struct bch_fs *c = container_of(to_delayed_work(work),
-				struct bch_fs, journal.reclaim_work);
-	struct journal *j = &c->journal;
+	struct bch_fs *c = container_of(j, struct bch_fs, journal);
 	struct bch_dev *ca;
 	unsigned iter, bucket_to_flush, min_nr = 0;
 	u64 seq_to_flush = 0;
 
+	lockdep_assert_held(&j->reclaim_lock);
+
 	bch2_journal_do_discards(j);
 
-	mutex_lock(&j->reclaim_lock);
 	spin_lock(&j->lock);
 
 	for_each_rw_member(ca, c, iter) {
@@ -493,13 +492,21 @@ void bch2_journal_reclaim_work(struct work_struct *work)
 
 	journal_flush_pins(j, seq_to_flush, min_nr);
 
-	mutex_unlock(&j->reclaim_lock);
-
 	if (!test_bit(BCH_FS_RO, &c->flags))
 		queue_delayed_work(c->journal_reclaim_wq, &j->reclaim_work,
 				   msecs_to_jiffies(j->reclaim_delay_ms));
 }
 
+void bch2_journal_reclaim_work(struct work_struct *work)
+{
+	struct journal *j = container_of(to_delayed_work(work),
+				struct journal, reclaim_work);
+
+	mutex_lock(&j->reclaim_lock);
+	bch2_journal_reclaim(j);
+	mutex_unlock(&j->reclaim_lock);
+}
+
 static int journal_flush_done(struct journal *j, u64 seq_to_flush)
 {
 	int ret;
diff --git a/fs/bcachefs/journal_reclaim.h b/fs/bcachefs/journal_reclaim.h
index 71545ad3bd58..9bf982a17797 100644
--- a/fs/bcachefs/journal_reclaim.h
+++ b/fs/bcachefs/journal_reclaim.h
@@ -42,6 +42,7 @@ void bch2_journal_pin_add_if_older(struct journal *,
 void bch2_journal_pin_flush(struct journal *, struct journal_entry_pin *);
 
 void bch2_journal_do_discards(struct journal *);
+void bch2_journal_reclaim(struct journal *);
 void bch2_journal_reclaim_work(struct work_struct *);
 
 void bch2_journal_flush_pins(struct journal *, u64);