On 2021/07/18 14:16, Luiz Augusto von Dentz wrote:
Anyway after all these discussion Im even
more convinced that the real problem lies in hci_dev_get/hold, after
all references are usually used to prevent the objects to be freed but
in this case it doesn't and no locking will gonna fix that.
If hci_dev_hold() calls atomic_long_add_unless(&file->f_count, 1, 0) under RCU,
vhci_release(file) would not be called until all sockets using that hdev drops
the reference, and hci_sock_dev_event(hdev, HCI_DEV_UNREG) no longer needs to
traverse sockets on hci_sk_list.head list. This requires adding "struct file *" to
"struct hci_dev". My patch keeps changes be confined to only hci_sock_dev_event().