[Intel-gfx] [PATCH v4 2/4] drm/i915/psr: Account for sink CRC raciness on some panels

[Jim Bride]

On Fri, Aug 04, 2017 at 06:38:02PM +0000, Pandiyan, Dhinakaran wrote:
> 
> 
> 
> On Thu, 2017-08-03 at 11:07 -0700, Rodrigo Vivi wrote:
> > On Tue, Jul 18, 2017 at 2:34 PM, Jim Bride <jim.bride at linux.intel.com> wrote:
> > > According to the eDP spec, when the count field in TEST_SINK_MISC
> > > increments then the six bytes of sink CRC information in the DPCD
> > > should be valid.  Unfortunately, this doesn't seem to be the case
> > > on some panels, and as a result we get some incorrect and inconsistent
> > > values from the sink CRC DPCD locations at times.  This problem exhibits
> > > itself more on faster processors (relative failure rates HSW < SKL < KBL.)
> > > In order to try and account for this, we try a lot harder to read the sink
> > > CRC until we get consistent values twice in a row before returning what we
> > > read and delay for a time before trying to read.  We still see some
> > > occasional failures, but reading the sink CRC is much more reliable,
> > > particularly on SKL and KBL, with these changes than without.
> 
> I'm curious if we get the correct crc if we waited a full second.

On SKL, times less than a second work fine generally.  On KBL, the
sink CRC is *way* less reliable, and I've seen runs where I set the
retry counts ridiculously high (> 30) and still not received valid
values.


Jim


> > 
> > Is DK now ok with this description?
> > I believe he requested more info here.
> > 
> > >
> > > v2: * Reduce number of retries when reading the sink CRC (Jani)
> > >     * Refactor to minimize changes to the code (Jani)
> > >     * Rebase
> > > v3: * Rebase
> > > v4: * Switch from do-while to for loop when reading CRC values (Jani)
> > >     * Rebase
> > > Cc: Rodrigo Vivi <rodrigo.vivi at intel.com>
> > > Cc: Paulo Zanoni <paulo.r.zanoni at intel.com>
> > > Cc: Jani Nikula <jani.nikula at intel.com>
> > > Signed-off-by: Jim Bride <jim.bride at linux.intel.com>
> > > ---
> > >  drivers/gpu/drm/i915/intel_dp.c | 33 ++++++++++++++++++++++++++++++---
> > >  1 file changed, 30 insertions(+), 3 deletions(-)
> > >
> > > diff --git a/drivers/gpu/drm/i915/intel_dp.c b/drivers/gpu/drm/i915/intel_dp.c
> > > index 2d42d09..c90ca1c 100644
> > > --- a/drivers/gpu/drm/i915/intel_dp.c
> > > +++ b/drivers/gpu/drm/i915/intel_dp.c
> > > @@ -3906,6 +3906,11 @@ int intel_dp_sink_crc(struct intel_dp *intel_dp, u8 *crc)
> > >         u8 buf;
> > >         int count, ret;
> > >         int attempts = 6;
> > > +       u8 old_crc[6];
> > > +
> > > +       if (crc == NULL) {
> > > +               return -ENOMEM;
> > > +       }
> > 
> > wouldn't we drop this check per DK and Jani request?
> > I believe we don't need it, but even if there are cases that we need
> > we could remove the braces..
> > 
> 
> Yeah, crc is allocated on the stack. If that is null, we'll have bigger
> problems to deal with. And I think it's reasonable to assume the caller
> is sending a valid array to fill data.
> 
> > >
> > >         ret = intel_dp_sink_crc_start(intel_dp);
> > >         if (ret)
> > > @@ -3929,11 +3934,33 @@ int intel_dp_sink_crc(struct intel_dp *intel_dp, u8 *crc)
> > >                 goto stop;
> > >         }
> > >
> > > -       if (drm_dp_dpcd_read(&intel_dp->aux, DP_TEST_CRC_R_CR, crc, 6) < 0) {
> > > -               ret = -EIO;
> > > -               goto stop;
> > > +       /*
> > > +        * Sometimes it takes a while for the "real" CRC values to land in
> > > +        * the DPCD, so try several times until we get two reads in a row
> > > +        * that are the same.  If we're an eDP panel, delay between reads
> > > +        * for a while since the values take a bit longer to propagate.
> > > +        */
> > > +       for (attempts = 0; attempts < 6; attempts++) {
> > > +               intel_wait_for_vblank(dev_priv, intel_crtc->pipe);
> > 
> > DK, we need vblank wait because the crc calculation also may take one vblank.
> > usually 2 actually... one to make sure you have the full screen
> > updated and one for the calculation.
> > In the past when we didn't used the count we were waiting 2 vblanks...
> > 
> 
> 
> Thanks for the clarification. My reasoning was, after the first two
> vblank_waits for the sink to calculate crc, the ones in the retry path
> were unnecessary. We just need some delay before reading the dpcd again
> without having to enable vblank interrupts. Anyway, the number of
> retries is low enough that it shouldn't matter.
> 
> On the other hand, since the only consumers of dp sink crc are tests,
> why can't the kernel just dump what it reads to debugfs and let the test
> deal with erroneous results?
> 
> > > +
> > > +               if (drm_dp_dpcd_read(&intel_dp->aux, DP_TEST_CRC_R_CR,
> > > +                                    crc, 6) < 0) {
> > > +                       ret = -EIO;
> > > +                       break;
> > > +               }
> > > +
> > > +               if (attempts && memcmp(old_crc, crc, 6) == 0)
> > > +                       break;
> > > +               memcpy(old_crc, crc, 6);
> > 
> > little bikeshed: too many hardcoded "6" around... a sizeof would be better...
> > but whatever...
> > 
> > > +
> > > +               if (is_edp(intel_dp))
> > > +                       usleep_range(20000, 25000);
> > >         }
> > >
> > > +       if (attempts == 6) {
> > > +               DRM_DEBUG_KMS("Failed to get CRC after 6 attempts.\n");
> > > +               ret = -ETIMEDOUT;
> > > +       }
> > >  stop:
> > >         intel_dp_sink_crc_stop(intel_dp);
> > >         return ret;
> > > --
> > > 2.7.4
> > >
> > > _______________________________________________
> > > Intel-gfx mailing list
> > > Intel-gfx at lists.freedesktop.org
> > > https://lists.freedesktop.org/mailman/listinfo/intel-gfx
> > 
> > 
> >