bcachefs: Fix deadlock

This fixes two deadlocks:

1.pcpu_alloc_mutex involved one as pointed by syzbot[1]
2.recursion deadlock.

The root cause is that we hold the bc lock during alloc_percpu, fix it
by following the pattern used by __btree_node_mem_alloc().

[1] https://lore.kernel.org/all/[email protected]/T/

Reported-by: [email protected]
Tested-by: [email protected]
Signed-off-by: Alan Huang <[email protected]>
Signed-off-by: Kent Overstreet <[email protected]>

Author: alanskind

fs/bcachefs/btree_cache.c

@@ -203,7 +203,7 @@ struct btree *__bch2_btree_node_mem_alloc(struct bch_fs *c)
 		return NULL;
 	}
 
-	bch2_btree_lock_init(&b->c, 0);
+	bch2_btree_lock_init(&b->c, 0, GFP_KERNEL);
 
 	__bch2_btree_node_to_freelist(bc, b);
 	return b;
@@ -795,17 +795,18 @@ struct btree *bch2_btree_node_mem_alloc(struct btree_trans *trans, bool pcpu_rea
 		}
 
 	b = __btree_node_mem_alloc(c, GFP_NOWAIT|__GFP_NOWARN);
-	if (!b) {
+	if (b) {
+		bch2_btree_lock_init(&b->c, pcpu_read_locks ? SIX_LOCK_INIT_PCPU : 0, GFP_NOWAIT);
+	} else {
 		mutex_unlock(&bc->lock);
 		bch2_trans_unlock(trans);
 		b = __btree_node_mem_alloc(c, GFP_KERNEL);
 		if (!b)
 			goto err;
+		bch2_btree_lock_init(&b->c, pcpu_read_locks ? SIX_LOCK_INIT_PCPU : 0, GFP_KERNEL);
 		mutex_lock(&bc->lock);
 	}
 
-	bch2_btree_lock_init(&b->c, pcpu_read_locks ? SIX_LOCK_INIT_PCPU : 0);
-
 	BUG_ON(!six_trylock_intent(&b->c.lock));
 	BUG_ON(!six_trylock_write(&b->c.lock));
 

fs/bcachefs/btree_key_cache.c

@@ -156,7 +156,7 @@ bkey_cached_alloc(struct btree_trans *trans, struct btree_path *path, unsigned k
 	}
 
 	if (ck) {
-		bch2_btree_lock_init(&ck->c, pcpu_readers ? SIX_LOCK_INIT_PCPU : 0);
+		bch2_btree_lock_init(&ck->c, pcpu_readers ? SIX_LOCK_INIT_PCPU : 0, GFP_KERNEL);
 		ck->c.cached = true;
 		goto lock;
 	}

fs/bcachefs/btree_locking.c

@@ -7,9 +7,10 @@
 static struct lock_class_key bch2_btree_node_lock_key;
 
 void bch2_btree_lock_init(struct btree_bkey_cached_common *b,
-			  enum six_lock_init_flags flags)
+			  enum six_lock_init_flags flags,
+			  gfp_t gfp)
 {
-	__six_lock_init(&b->lock, "b->c.lock", &bch2_btree_node_lock_key, flags);
+	__six_lock_init(&b->lock, "b->c.lock", &bch2_btree_node_lock_key, flags, gfp);
 	lockdep_set_notrack_class(&b->lock);
 }
 

fs/bcachefs/btree_locking.h

@@ -13,7 +13,7 @@
 #include "btree_iter.h"
 #include "six.h"
 
-void bch2_btree_lock_init(struct btree_bkey_cached_common *, enum six_lock_init_flags);
+void bch2_btree_lock_init(struct btree_bkey_cached_common *, enum six_lock_init_flags, gfp_t gfp);
 
 void bch2_trans_unlock_noassert(struct btree_trans *);
 void bch2_trans_unlock_write(struct btree_trans *);

fs/bcachefs/six.c

@@ -850,7 +850,8 @@ void six_lock_exit(struct six_lock *lock)
 EXPORT_SYMBOL_GPL(six_lock_exit);
 
 void __six_lock_init(struct six_lock *lock, const char *name,
-		     struct lock_class_key *key, enum six_lock_init_flags flags)
+		     struct lock_class_key *key, enum six_lock_init_flags flags,
+		     gfp_t gfp)
 {
 	atomic_set(&lock->state, 0);
 	raw_spin_lock_init(&lock->wait_lock);
@@ -873,7 +874,7 @@ void __six_lock_init(struct six_lock *lock, const char *name,
 		 * failure if they wish by checking lock->readers, but generally
 		 * will not want to treat it as an error.
 		 */
-		lock->readers = alloc_percpu(unsigned);
+		lock->readers = alloc_percpu_gfp(unsigned, gfp);
 	}
 #endif
 }

fs/bcachefs/six.h

@@ -164,18 +164,19 @@ enum six_lock_init_flags {
 };
 
 void __six_lock_init(struct six_lock *lock, const char *name,
-		     struct lock_class_key *key, enum six_lock_init_flags flags);
+		     struct lock_class_key *key, enum six_lock_init_flags flags,
+		     gfp_t gfp);
 
 /**
  * six_lock_init - initialize a six lock
  * @lock:	lock to initialize
  * @flags:	optional flags, i.e. SIX_LOCK_INIT_PCPU
  */
-#define six_lock_init(lock, flags)					\
+#define six_lock_init(lock, flags, gfp)					\
 do {									\
 	static struct lock_class_key __key;				\
 									\
-	__six_lock_init((lock), #lock, &__key, flags);			\
+	__six_lock_init((lock), #lock, &__key, flags, gfp);			\
 } while (0)
 
 /**