Re: [PATCH] Bluetooth: hci_core: sync hci rx and cmd sync work

[Hillf Danton]

On Thu, 16 Jan 2025 10:31:03 -0500 Luiz Augusto von Dentz
> On Thu, Jan 16, 2025 at 5:42 AM Hillf Danton <hdanton@xxxxxxxx> wrote:
> >
> > This does not work because conn can be freed before bumping refcnt up,
> > and bumpup alone does not help l2cap_send_cmd() in case conn->hchan has been
> > cleared in l2cap_conn_del().
> 
> Well then perhaps we can use hdev lock just to protect l2cap_data:
> 
> diff --git a/net/bluetooth/l2cap_core.c b/net/bluetooth/l2cap_core.c
> index 27b4c4a2ba1f..00ef61609d49 100644
> --- a/net/bluetooth/l2cap_core.c
> +++ b/net/bluetooth/l2cap_core.c
> @@ -7466,14 +7466,33 @@ static void l2cap_recv_reset(struct l2cap_conn *conn)
>         conn->rx_len = 0;
>  }
> 
> +static struct l2cap_conn *l2cap_conn_hold_unless_zero(struct l2cap_conn *c)
> +{
> +       BT_DBG("conn %p orig refcnt %u", c, kref_read(&c->ref));
> +
> +       if (!kref_get_unless_zero(&c->ref))
> +               return NULL;
> +
> +       return c;
> +}
> +
>  void l2cap_recv_acldata(struct hci_conn *hcon, struct sk_buff *skb, u16 flags)
>  {
> -       struct l2cap_conn *conn = hcon->l2cap_data;
> +       struct l2cap_conn *conn;
>         int len;
> 
> +       /* Lock hdev to access l2cap_data to avoid race with l2cap_conn_del */
> +       hci_dev_lock(hcon->hdev);
> +
> +       conn = hcon->l2cap_data;
> +
>         if (!conn)
>                 conn = l2cap_conn_add(hcon);
> 
> +       conn = l2cap_conn_hold_unless_zero(conn);
> +
> +       hci_dev_unlock(hcon->hdev);
> +
This does not work either because after releasing lock bumpup alone can
not prevent conn->hchan from being cleared.