[Intel-gfx] [PATCH i-g-t] lib/igt_gt.c : allow changes to stop_rings mode bits

[Daniel Vetter]

On Mon, Jul 13, 2015 at 04:07:14PM +0000, Gore, Tim wrote:
> 
> 
> Tim Gore 
> Intel Corporation (UK) Ltd. - Co. Reg. #1134945 - Pipers Way, Swindon SN3 1RJ
> 
> 
> > -----Original Message-----
> > From: Daniel Vetter [mailto:daniel.vetter at ffwll.ch] On Behalf Of Daniel
> > Vetter
> > Sent: Monday, July 13, 2015 3:59 PM
> > To: Gore, Tim
> > Cc: Daniel Vetter; intel-gfx at lists.freedesktop.org; Wood, Thomas
> > Subject: Re: [Intel-gfx] [PATCH i-g-t] lib/igt_gt.c : allow changes to stop_rings
> > mode bits
> > 
> > On Mon, Jul 13, 2015 at 09:43:11AM +0000, Gore, Tim wrote:
> > >
> > >
> > > Tim Gore
> > > Intel Corporation (UK) Ltd. - Co. Reg. #1134945 - Pipers Way, Swindon
> > > SN3 1RJ
> > >
> > > > -----Original Message-----
> > > > From: Daniel Vetter [mailto:daniel.vetter at ffwll.ch] On Behalf Of
> > > > Daniel Vetter
> > > > Sent: Monday, July 13, 2015 10:30 AM
> > > > To: Gore, Tim
> > > > Cc: intel-gfx at lists.freedesktop.org; Wood, Thomas
> > > > Subject: Re: [Intel-gfx] [PATCH i-g-t] lib/igt_gt.c : allow changes
> > > > to stop_rings mode bits
> > > >
> > > > On Fri, Jul 10, 2015 at 02:06:06PM +0100, tim.gore at intel.com wrote:
> > > > > From: Tim Gore <tim.gore at intel.com>
> > > > >
> > > > > In function igt_set_stop_rings, the test
> > > > >   igt_assert_f(flags == 0 || current == 0, ..
> > > > >
> > > > > will fail if we are trying to force a hang but the
> > > > > STOP_RINGS_ALLOW_BAN or STOP_RINGS_ALLOW_ERROR bit is set.
> > > > > With the introduction of per ring resets in the driver (in
> > > > > android) these bits do not get cleared to zero when an individual
> > > > > ring is reset. This causes subsequent attempt to cause a ring hang
> > > > > via this function to fail, leading to several igt tests failing
> > > > > (ie gem_reset_stats subtest ban-xxx etc).
> > > >
> > > > Fix tdr to reset these instead?
> > > > -Daniel
> > > >
> > > I could change tdr, but why. When the TDR handles a ring hang and
> > > resets the ring, why would it modify the flag that defines if the
> > > driver should ban a frequently hanging context? If we get rid of the
> > > stop_rings interface, as Chris Wilson suggested, we would still need
> > > to keep the STOP_RING_ALLOW_BAN/ALLOW_ERRORS bits in debugfs, but
> > you
> > > would not expect to have to re-write these bits each time there is a ring
> > reset.
> > 
> > The fix current hang recover code to no reset this, add some grace period,
> > then push this patch to igt. We don't have full-blown abi guarantees for
> > debugfs/igt stuff, but I want at least a few months (really last released
> > kernel&igt) of backwards/forward compatibility. And inconsistent behaviour
> > isn't great imo.
> > -Daniel
> 
> Sorry Daniel, I didn't really follow that.
> I didn't want a gpu reset to clear the ALLOW_BAN bit, since this will defeat the
> point of this bit, except perhaps in test situations where you can keep setting it
> each time you deliberately cause a hang. It seems like the ALLOW_BAN bit has
> uses in real world situations, although I don't know it anyone uses it.

ALLOW_BAN is only for testing. It's meant to re-enable auto-banning
because we disable that by default for automated tests. This is all meant
to be used together with the stop_rings stuff only.

> Would you be more comfortable with
> Igt_assert_f ( ( flags ==0 ) || 
>      (( current & STOP_RING_ALL) ==0)  && ((current ^ flags) & ~ STOP_RING_ALL == 0 ) )
> 
> So the either the new flags must be 0 (currently allowed) or the existing flags must
> indicate that  all hangs are cleared (0 except possibly the mode bits) AND the mode
> bits you are writing are the same as the current values. ??

Iirc gem_reset_stats uses stop_rings only to supress gpu reset warnings in
dmesg (since they're expected) and not for the actual stop_rings logic. It
sets stop_rings after submitting the hanging batch, but before that one is
detected as hung. We could just add another bit to stop_rings for that
case.

What I still don't understand is why tdr can't just keep on properly
resetting stop_rings. It'll break tons of existing tests, and I don't
understand the upside. stop_rings has become quasi-abi (that's why the
ALLOW bits have such funky semantics, it's for backwards compat), if you
need to change it you need to extend it, not break it.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch