[PATCH] dmatest: abort transfers immediately when asked for

[will.deacon@arm.com (Will Deacon)]

On Tue, May 21, 2013 at 06:24:15PM +0100, Andy Shevchenko wrote:
> On Tue, May 21, 2013 at 6:11 PM, Will Deacon <will.deacon@arm.com> wrote:
> > Hi Andy,
> >
> > On Tue, May 21, 2013 at 01:33:17PM +0100, Andy Shevchenko wrote:
> >> When thread is going to be stopped we have to unconditionally terminate all
> >> ongoing transfers. Otherwise it would be possible that callback function will
> >> be called on the next interrupt and will try to access to already freed
> >> structures.
> >>
> >> The patch introduces specific error message for this, though it doesn't
> >> increase the counter of the failed tests.
> >>
> >> Signed-off-by: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
> >> Reported-by: Will Deacon <will.deacon@arm.com>
> >
> > Thanks for persevering with this! Although this patch definitely fixes the
> > panic I was seeing, I now observe buffer verification failures in subsequent
> > test runs after an aborted run:
> 
> I think the description to the commit adfa543e "dmatest: don't use
> set_freezable_with_signal()" may shed light on this.
> 
> The background (if  I got it correctly) is in race with done flag. So,
> we got a callback call from the DMA engine, but we don't know which
> transfer triggers it.
> I might be wrong. This is just an assumption.

I've not managed to work out exactly what's going on, but it's certainly
something like that. In fact, I just managed to trigger a case where all but
one of the transfers is aborted, whilst the remaining one fails. Looking at
the code, I can't see how that situation comes about, since the threads are
protected with the info mutex and kthread_stop is synchronous.

> Have you ever see such behaviour on pre v3.10-rc1 kernels? (I mean
> with old dmatest module)

No, dmatest from 3.9 is completely reliable in my experience.

Will