[PATCH] Bluetooth: hci_sync: invoke destroy callbacks outside lock
From: Zheng Shuhao
Date: Sat Nov 21 2026 - 07:01:35 EST
hci_cmd_sync_clear() cancels queued command-sync entries while holding
cmd_sync_work_lock. Cancellation invokes entry->destroy(), and
add_ext_adv_params_complete() takes hdev->lock. In the opposite direction,
le_conn_complete_evt() holds hdev->lock while hci_update_passive_scan()
queues work and acquires cmd_sync_work_lock. This creates the circular
locking dependency reported by lockdep during controller teardown.
Detach the queued entries to a local list under cmd_sync_work_lock, then
run the cancellation callbacks after releasing the mutex. This keeps
queue updates serialized without invoking callbacks under the mutex.
Testing:
- Lockdep-enabled QEMU baseline: reproducer completed twice and emitted
the circular-locking warning once.
- Patched QEMU kernel: reproducer completed five times with no such
warning.
- Bluetooth ECDH and SMP self-tests passed in both kernels.
- Both kernels built; baseline build logged a timestamp clock-skew
warning. checkpatch.pl --strict reported no errors or warnings.
Fixes: 1c66bee492a5 ("Bluetooth: Fix race condition in hci_cmd_sync_clear")
Reported-by: syzbot+fc55a609cd9deea4ea37@xxxxxxxxxxxxxxxxxxxxxxxxx
Link: https://syzkaller.appspot.com/bug?extid=fc55a609cd9deea4ea37
Signed-off-by: Zheng Shuhao <viennaoffh@xxxxxxxxx>
diff --git a/net/bluetooth/hci_sync.c b/net/bluetooth/hci_sync.c
index ea2014e74..65b8aa4ad 100644
--- a/net/bluetooth/hci_sync.c
+++ b/net/bluetooth/hci_sync.c
@@ -683,6 +683,7 @@ static void _hci_cmd_sync_cancel_entry(struct hci_dev *hdev,
void hci_cmd_sync_clear(struct hci_dev *hdev)
{
+ LIST_HEAD(cancel_list);
struct hci_cmd_sync_work_entry *entry, *tmp;
/* cmd_work is disabled, the pending request can only time out */
@@ -690,10 +691,13 @@ void hci_cmd_sync_clear(struct hci_dev *hdev)
cancel_work_sync(&hdev->cmd_sync_work);
cancel_work_sync(&hdev->reenable_adv_work);
+ /* Run destroy callbacks without holding cmd_sync_work_lock. */
mutex_lock(&hdev->cmd_sync_work_lock);
- list_for_each_entry_safe(entry, tmp, &hdev->cmd_sync_work_list, list)
- _hci_cmd_sync_cancel_entry(hdev, entry, -ECANCELED);
+ list_splice_init(&hdev->cmd_sync_work_list, &cancel_list);
mutex_unlock(&hdev->cmd_sync_work_lock);
+
+ list_for_each_entry_safe(entry, tmp, &cancel_list, list)
+ _hci_cmd_sync_cancel_entry(hdev, entry, -ECANCELED);
}
void hci_cmd_sync_cancel(struct hci_dev *hdev, int err)