[Intel-gfx] [PATCH 1/2] drm/i915: time out of load detect busy-waits

[Daniel Vetter]

On Sat, Apr 28, 2012 at 10:00:29PM +0100, Chris Wilson wrote:
> On Wed, 25 Apr 2012 21:26:50 +0200, Daniel Vetter <daniel at ffwll.ch> wrote:
> > On Wed, Apr 25, 2012 at 06:14:37PM +0100, Chris Wilson wrote:
> > > On Fri, 20 Apr 2012 21:03:35 +0200, Daniel Vetter <daniel.vetter at ffwll.ch> wrote:
> > > > If we try to do that and the scanlines just wouldn't advance, we
> > > > busy-hang the machine holding the modeset mutex. Not great for
> > > > debugging.
> > > > 
> > > > References: https://bugzilla.kernel.org/show_bug.cgi?id=43020
> > > > Signed-off-by: Daniel Vetter <daniel.vetter at ffwll.ch>
> > > 
> > > Reviewer hangs head in shame:
> > > 
> > > > +		if (wait_for(I915_READ(pipe_dsl_reg) >= vactive, 1000))
> > > > +			DRM_ERROR("timed out waiting for vactive in "
> > > > +				  "load_detect, scanline: %u\n",
> > > > +				  I915_READ(pipe_dsl_reg));
> > > > +		if (wait_for((dsl = I915_READ(pipe_dsl_reg)) <= vsample, 1000))
> > > > +			DRM_ERROR("timed out waiting for vsample in "
> > > > +				  "load_detect, scanline: %u\n",
> > > > +				  I915_READ(pipe_dsl_reg));
> > > 
> > > wait_for() catches us out everytime we convert and existing while(),
> > > because the predicate is when it stops. Perhaps if we had a wait_until,
> > > but anyway the fix here is:
> > > 
> > > if (wait_for(I915_READ(pipe_dsl_reg) < vactive, 1000))
> > >   ...
> > > if (wait_for((dsl = I915_READ(pipe_dsl_reg)) > vsample, 1000))
> > >   ...
> > dinq rectified, it never happened. Thanks for catching this.
> 
> wait_for() has even more subtleties in store for us, the unwary coder.
> By default, it uses a 1ms sleep between polling the register, chosen to
> be kind whilst waiting for panel bits to power up which do take a fair
> amount of time. Here, that extra delay causes us to sample the vsync
> rather than the border. The quirk of the [vh]sync is that the monitor bit
> of ST00 is always true. And since we always seem to pick that row to read
> we always think there is a CRT present.
> 
> The choice is either to use the busy-polling variant, wait_for_atomic,
> or restructure the entire block to use a single timeout with direct
> reads. And whilst you are modifying the code, convert the polling reads
> to I915_READ_NOTRACE().

Thanks a lot for digging into this, I've only managed to do the bisect
before I've haeded off into the w/e. I've dropped this patch for now, too
much fail in it. The underlying issue of not properly doing load-detect on
a active but disabled crtc is fixed, so this isn't required any more.

I'll look a this again later, hopefully clue strikes me by then.
-Daniel
-- 
Daniel Vetter
Mail: daniel at ffwll.ch
Mobile: +41 (0)79 365 57 48