Re: [BUG] staging: wfx: possible deadlock in wfx_conf_tx() and wfx_add_interface()

[Jia-Ju Bai]

On 2022/2/1 19:33, Hillf Danton wrote:
> On Tue, 1 Feb 2022 15:09:34 +0800 Jia-Ju Bai wrote:
> > Hello,
> >
> > My static analysis tool reports a possible deadlock in the wfx driver in
> > Linux 5.16:
> >
> > wfx_conf_tx()
> >     mutex_lock(&wdev->conf_mutex); --> Line 225 (Lock A)
> >     wfx_update_pm()
> >       wait_for_completion_timeout(&wvif->set_pm_mode_complete, ...); -->
> > Line 3019 (Wait X)
> >
> > wfx_add_interface()
> >     mutex_lock(&wdev->conf_mutex); --> Line 737 (Lock A)
> >     complete(&wvif->set_pm_mode_complete); --> Line 758 (Wake X)
> >
> > When wfx_conf_tx() is executed, "Wait X" is performed by holding "Lock
> > A". If wfx_add_interface() is executed at this time, "Wake X" cannot be
> > performed to wake up "Wait X" in wfx_conf_tx(), because "Lock A" has
> > been already hold by wfx_conf_tx(), causing a possible deadlock.
> > I find that "Wait X" is performed with a timeout, to relieve the
> > possible deadlock; but I think this timeout can cause inefficient execution.
> >
> > I am not quite sure whether this possible problem is real and how to fix
> > it if it is real.
> > Any feedback would be appreciated, thanks :)
> >
> >
> > Best wishes,
> > Jia-Ju Bai
> Hey Jia-Ju
>
> Thank you for reporting it.
>
> Given the init_completion() prior to complete() in wfx_add_interface(),
> no waiter is waken up by the complete(), so it has nothing to do with
> the waiter in the conf path.

Hi Hillf,

Thanks for your reply and detailed explanation :)

> BTW if the unusual wfx init is a real use case then we can add a new helper.
>
> Mind introducing your toy to LKML? How much time have you been put in it?
> Its current status and future works?

Do you mean my static analysis tool that generated the report?
In fact, I spent 3-4 months of my part time on implementing this tool, 
which can detect deadlocks caused by locking cycles and improper 
waiting/waking operations.
This tool still reports some false positives, due to missing some 
special patterns in the kernel code, such as "init_completion() prior to 
complete()" in this false bug.
Thus, I am improving the tool to reduce false positives now.
Any suggestion on deadlock detection or the tool would be appreciated, 
thanks :)


Best wishes,
Jia-Ju Bai