bcachefs: Evict btree nodes we're deleting

There was a bug that led to duplicate btree node pointers being inserted
at the wrong level. The new topology repair code can fix that, except
that the btree cache code gets confused when we read in a btree node
from the pointer that was at the wrong level. This patch evicts nodes
that we're deleting to, which nicely solves the problem.

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

3 files changed, 34 insertions, 0 deletions

diff --git a/fs/bcachefs/btree_cache.c b/fs/bcachefs/btree_cache.c
index a42e0922f5e9..85c19e4e5216 100644
--- a/fs/bcachefs/btree_cache.c
+++ b/fs/bcachefs/btree_cache.c
@@ -948,6 +948,36 @@ void bch2_btree_node_prefetch(struct bch_fs *c, struct btree_iter *iter,
 	bch2_btree_node_fill(c, iter, k, btree_id, level, SIX_LOCK_read, false);
 }
 
+void bch2_btree_node_evict(struct bch_fs *c, const struct bkey_i *k)
+{
+	struct btree_cache *bc = &c->btree_cache;
+	struct btree *b;
+
+	b = btree_cache_find(bc, k);
+	if (!b)
+		return;
+
+	six_lock_intent(&b->c.lock, NULL, NULL);
+	six_lock_write(&b->c.lock, NULL, NULL);
+
+	wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
+		       TASK_UNINTERRUPTIBLE);
+	__bch2_btree_node_write(c, b);
+
+	/* wait for any in flight btree write */
+	btree_node_wait_on_io(b);
+
+	BUG_ON(btree_node_dirty(b));
+
+	mutex_lock(&bc->lock);
+	btree_node_data_free(c, b);
+	bch2_btree_node_hash_remove(bc, b);
+	mutex_unlock(&bc->lock);
+
+	six_unlock_write(&b->c.lock);
+	six_unlock_intent(&b->c.lock);
+}
+
 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
 			     struct btree *b)
 {
diff --git a/fs/bcachefs/btree_cache.h b/fs/bcachefs/btree_cache.h
index c517cc029454..40dd263a7caa 100644
--- a/fs/bcachefs/btree_cache.h
+++ b/fs/bcachefs/btree_cache.h
@@ -30,6 +30,8 @@ struct btree *bch2_btree_node_get_noiter(struct bch_fs *, const struct bkey_i *,
 void bch2_btree_node_prefetch(struct bch_fs *, struct btree_iter *,
 			      const struct bkey_i *, enum btree_id, unsigned);
 
+void bch2_btree_node_evict(struct bch_fs *, const struct bkey_i *);
+
 void bch2_fs_btree_cache_exit(struct bch_fs *);
 int bch2_fs_btree_cache_init(struct bch_fs *);
 void bch2_fs_btree_cache_init_early(struct btree_cache *);
diff --git a/fs/bcachefs/btree_gc.c b/fs/bcachefs/btree_gc.c
index 09a49dc63144..1c2eab41f7ca 100644
--- a/fs/bcachefs/btree_gc.c
+++ b/fs/bcachefs/btree_gc.c
@@ -298,6 +298,7 @@ again:
 				bch2_btree_ids[b->c.btree_id],
 				b->c.level - 1,
 				(bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(tmp.k)), buf))) {
+			bch2_btree_node_evict(c, tmp.k);
 			ret = bch2_journal_key_delete(c, b->c.btree_id,
 						      b->c.level, tmp.k->k.p);
 			if (ret)
@@ -359,6 +360,7 @@ again:
 		cur = NULL;
 
 		if (ret == DROP_THIS_NODE) {
+			bch2_btree_node_evict(c, tmp.k);
 			ret = bch2_journal_key_delete(c, b->c.btree_id,
 						      b->c.level, tmp.k->k.p);
 			dropped_children = true;