[PATCH] drm/etnaviv: only reject timeouts with tv_nsec >= 2 seconds

[Arnd Bergmann]

On Tue, Jan 21, 2020 at 5:10 PM Lucas Stach <l.stach at pengutronix.de> wrote:
>
> Hi Guido,
>
> On Di, 2020-01-21 at 13:55 +0100, Guido Günther wrote:
> > Hi,
> > On Tue, Jan 21, 2020 at 12:45:25PM +0100, Arnd Bergmann wrote:
> > > As Guido Günther reported, get_abs_timeout() in the etnaviv user space
> > > sometimes passes timeouts with nanosecond values larger than 1000000000,
> > > which gets rejected after my first patch.
> > >
> > > To avoid breaking this, while also not allowing completely arbitrary
> > > values, set the limit to 1999999999 and use set_normalized_timespec64()
> > > to get the correct format before comparing it.
> >
> > I'm seeing values up to 5 seconds so I need
> >
> >      if (args->timeout.tv_nsec > (5 * NSEC_PER_SEC))
> >
> > to unbreak rendering. Which seems to match what mesa's get_abs_timeout()
> > does and how it's invoked.
>
> I have not tested this myself yet, only looked at the code. From the
> code I quoted earlier, I don't see how we end up with 5 * NSEC_PER_SEC
> in the tv_nsec member, even if the timeout passed to get_abs_timeout()
> is 5 seconds.

I can think of two different ways you'd end up with around five seconds here:

a) you have a completely arbitrary 32-bit number through truncation,
    which is up to 4.2 seconds
b) you have the same kind of 32-bit number, but add up to another 999999999
    nanoseconds, so you get up to 5.2 seconds in the 64-bit field.

It could of course be something completely different. If this works correctly
today, we may need to allow any 64-bit input for the nanoseconds and do
an expensive 64-bit div/mod in the kernel for normalization rather than the
cheaper set_normalized_timespec64() from my patch.

        Arnd