Re: [PATCH v5] watchdog: add driver for StreamLabs USB watchdog device

[Alan Stern <stern@rowland.harvard.edu>]

On Sun, Jul 31, 2022 at 01:20:55AM -0700, Guenter Roeck wrote:
> On Sun, Jul 31, 2022 at 03:34:16AM +0100, Alexey Klimov wrote:
> > On Tue, Jul 26, 2022 at 8:48 AM Oliver Neukum <oneukum@suse.com> wrote:
> > >
> > > On 26.07.22 02:21, Alexey Klimov wrote:
> > > > On Mon, Jul 25, 2022 at 9:51 AM Greg KH <gregkh@linuxfoundation.org> wrote:
> > > >>
> > > >> On Mon, Jul 25, 2022 at 04:06:05AM +0100, Alexey Klimov wrote:
> > > >
> > > > [..]
> > > >
> > > >> Anyway, driver looks good to me, nice work!
> > > >>
> > > >> Reviewed-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
> > > >
> > > > Thanks, Greg. If you don't mind I'll use your tag in next version
> > > > after making changes suggested by Guenter since there will be no
> > > > significant functional changes. If code will change a lot, then the
> > > > process (Documentation/process/submitting-patches.rst) will require me
> > > > to drop the tag.
> > >
> > > Hi,
> > >
> > > while thinking about this a question arose. How does this
> > > device react to a USB reset? A watchdog that can be disabled
> > > by a simple reset does not like very reliable to me.
> > > Do you need to implement pre/post_reset() ?
> > 
> > You're right. Upon reset the watchdog is disabled even if it was active before.
> > Adding empty ->pre_reset() and ->post_reset() helps to avoid that, but
> > looking at Documentation and other drivers it seems that I need to do:
> > in pre_reset():
> > mutex_lock() to block any other I/O to the usb device;
> > __usb_streamlabs_wdt_cmd(STOP) to stop the watchdog;
> > and do not unlock the mutex;
> > 
> > in post_reset():
> > if (watchdog_active())
> >         __usb_streamlabs_wdt_cmd(START);
> > mutex_unlock() to allow other's I/O to the usb deivce.
> > 
> > Seems right?
> > 
> Not necessarily. Is other code doing something similar ?
> Using a mutex like this creates the risk for hung tasks.

Are mutexes intended to be used in situations where one function 
acquires the lock, then returns, and then a different function releases 
the lock?  I'm not sure about this.

Perhaps a good old semaphore would be more appropriate.  But it's clear 
that I/O to the device does need to be mutually exclusive with resets, 
one way or another.

Alan Stern