[PATCH] Improvements in edid detection timings (final)

* [PATCH] Improvements in edid detection timings (final)
@ 2011-10-17 13:12 Eugeni Dodonov
  2011-10-17 13:12 ` [PATCH 1/2] Give up on edid retries when i2c tells us that bus is not there Eugeni Dodonov
  2011-10-17 13:12 ` [PATCH 2/2] Check if the bus is valid prior to discovering edid Eugeni Dodonov
  0 siblings, 2 replies; 8+ messages in thread
From: Eugeni Dodonov @ 2011-10-17 13:12 UTC (permalink / raw)
  To: intel-gfx; +Cc: dri-devel, Eugeni Dodonov

Those are two identical fixes for improving EDID detection timings, which
also fix extremely slow xrandr queries in case of phantom outputs
(https://bugs.freedesktop.org/show_bug.cgi?id=41059)

The first fix is a small change to drm_edid, which prevents it from talking to
non-existent adapters by detecting them faster.

The second fix replicates the first one within the i915 driver. It does the
same thing, but without touching core DRM files.

Those are some of the testing results from our QA team:

Regressions and functional testing:
         Machine+ports                  result
Ironlake(mobile) eDP/VGA                pass
Ironlake(desktop) DP/VGA                pass
Ironlake(mobile) LVDS/VGA/DP            pass
G45(desktop) VGA/HDMI                   pass
Pineview(mobile) LVDS/VGA               pass

xrandr performance:
                           Without patch   with patch
E6510(Ironlake mobile)     0.119           0.111
PK1(Ironlake desktop)      0.101           0.080
T410b(Ironlake mobile)     0.406           0.114
G45b( G45 desktop)         0.121           0.091
Pnv1(Pineview mobile)      0.043           0.040

Those are the results for machines affected by phantom outputs issue, based on
fd.o #41059 feedback:

xrandr performance:
                           Without patch   with patch
System 1                   0.840           0.290
System 2                   0.690           0.140
System 3                   0.315           0.280
System 4                   0.175           0.140
System 6 (original issue)  4s              0.184

We have observed no regressions in any cases, and performance improvements
of 20-30% for edid detection timing. Combining it with the results obtained
at https://bugs.freedesktop.org/show_bug.cgi?id=41059, besides those
improvements it also improves xrandr timing by up to 20x in the worst case
of phantom outputs.

I believe that the better way to fix this is via the drm_get_edid() fix, as
it is a one-line fix and could benefit all other chipsets. And we won't have
to reinvent the wheel with intel_drm_get_edid, which only duplicates the
existent functionality with no additional benefits.

Could we have any feedback or reviewed-by or from non-intel drm maintainers?

Thanks!

Eugeni Dodonov (2):
  Give up on edid retries when i2c tells us that bus is not there
  Check if the bus is valid prior to discovering edid.

 drivers/gpu/drm/drm_edid.c         |    5 ++++
 drivers/gpu/drm/i915/intel_crt.c   |   46 ++++++++++++++++++------------------
 drivers/gpu/drm/i915/intel_dp.c    |    4 +-
 drivers/gpu/drm/i915/intel_drv.h   |    3 +-
 drivers/gpu/drm/i915/intel_hdmi.c  |    4 +-
 drivers/gpu/drm/i915/intel_i2c.c   |   42 ++++++++++++++++++++++++++++++++
 drivers/gpu/drm/i915/intel_lvds.c  |    2 +-
 drivers/gpu/drm/i915/intel_modes.c |   29 +----------------------
 drivers/gpu/drm/i915/intel_sdvo.c  |    4 +-
 9 files changed, 80 insertions(+), 59 deletions(-)

-- 
1.7.7

^ permalink raw reply	[flat|nested] 8+ messages in thread