On Mon, Nov 30, 2015 at 10:11:12AM +0000, Chris Wilson wrote:
> On Thu, Nov 26, 2015 at 12:34:34PM +0100, Daniel Vetter wrote:
> > So there's 3 competing proposals for what wait_ioctl should do wrt
> > -EIO:
> >
> > - return -EIO when the gpu is wedged. Not terribly useful for
> > userspace since it might race with a hang and then there's no
> > guarantee that a subsequent execbuf won't end up in an -EIO.
> > Terminally wedge really can only be reliably signalled at execbuf
> > time, and userspace needs to cope with that (or decide not to
> > bother).
> >
> > - EIO for any obj that suffered from a reset. This means big internal
> > reorginazation in the kernel since currently we track reset stats
> > per-ctx and not on the obj. That's also what arb robustness wants.
> > We could do this, but this feels like new ABI territory with the
> > usual userspace requirements and high hurdles.
> >
> > - No -EIO at all. Consistent with set_domain_ioctl and simplest to
> > implement. Which is what this patch does.
>
> Since no one else is weighing into the ABI discussion, I'm happy with
> losing EIO here. I thought it could be useful, but as no one is using or
> seems likely to start using it, begone.
>
> > We can always opt to change this later on if there's a real need.
> >
> > To make the test really exercise this do a full wedged gpu hang, to
> > make sure -EIO doesn't leak out at all.
> >
> > Cc: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx>
> > Signed-off-by: Daniel Vetter <daniel.vetter@xxxxxxxxx>
> > ---
> > tests/gem_eio.c | 6 +++++-
> > 1 file changed, 5 insertions(+), 1 deletion(-)
> >
> > diff --git a/tests/gem_eio.c b/tests/gem_eio.c
> > index a24c8f1c53b5..8345d1a7a429 100644
> > --- a/tests/gem_eio.c
> > +++ b/tests/gem_eio.c
> > @@ -161,10 +161,14 @@ static void test_wait(int fd)
> > {
> > igt_hang_ring_t hang;
> >
> > + igt_require(i915_reset_control(false));
>
> However, this is not required to test the ABI change above as the wait
> itself will still hang, whether or not it wedges the GPU.
Yes it's not strictly required, but without it the testcase is fairly
boring. If we move the check_wedge out of wait_request then a normail gpu
reset would always return 0 (after retrying a few times perhaps), so I
figured testing the wedged case is the only one that's worth it.
Maybe we should dupe the subtests all and have wedged and non-wedged cases
for all of them? That would imo make more sense.
-Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
