Deadlock between device_detach() and usbd_do_request_flags()

[Hans Petter Selasky]

On 09/04/16 23:20, Andriy Voskoboinyk wrote:
> There is a rare, but reproducible deadlock for wlan(4) drivers:
>
> Thread 1:
>  * uhub_explore_handle_re_enumerate() (obtains enum_sx lock)
>  * usbd_set_config_index()
>  * usb_unconfigure()
>  * usb_detach_device()
>  * usb_detach_device_sub()
>  * <device_detach()>
>    typically <device_stop()> is executed here (prevents
>     another possible deadlock?)
>  * ieee80211_ifdetach()
>  * ieee80211_vap_destroy()
>  * <ic->ic_vap_delete>
>  * ieee80211_vap_detach()
>    here it calls ieee80211_stop() and waits for <any> -> INIT state
>    transition
>
> Thread 2 (started from thread 1):
>  * ieee80211_newstate_cb()
>  * vap->iv_newstate()
>    here: if the driver will try to call usbd_do_request_flags()
>    (typically via <drv_write_reg> / <drv_read_reg>) it will hang
>    (because enum_sx lock is already held by thread 1).
>
>
> Another way: execute some periodical task that will try to access
> some registers (urtwn_temp_calib(), rum_ratectl_task(),
> run_ratectl_cb()) while thread 1 is running - deadlock is
> here too, since <device_detach> will wait for them indefinitely
> (via ieee80211_draintask())
>
> Right now the most obvious (and, probably, wrong) way is to just
> detect & release all locks (usbd_enum_unlock()) for
> ieee80211_ifdetach() / ieee80211_draintask() and re-acquire them
> later (not tested yet).
>

Hi,

I think the right solution is to let usbd_do_request_flags() use its own 
SX lock for synchronization, instead of re-using the enumeration SX 
lock. What do you think about that?

--HPS