[PATCH v2 0/6] Per-context and per-client engine busyness

From: Tvrtko Ursulin <tursulin@ursulin.net>
To: Intel-gfx@lists.freedesktop.org
Subject: [PATCH v2 0/6] Per-context and per-client engine busyness
Date: Fri, 19 Jan 2018 13:45:22 +0000	[thread overview]
Message-ID: <20180119134528.4720-1-tvrtko.ursulin@linux.intel.com> (raw)

From: Tvrtko Ursulin <tvrtko.ursulin@intel.com>

I have sent this as part of a larger series back in October '17.

First part of it is implementing a customer requirement to be able to query
engine utilization on their own contexts. This is done in patch 2, which falls
under the standard open source userspace requirements etc.

The rest of the series implements per-client engine utilization stats, something
which I think is quite interesting and missing from our stack at the moment. And
in fact, in the latest other OS update I have noticed they actually can expose
this statistics.

I have prototyped a simple top-like utility which can show you the overall
engine stats (data coming from recently merged PMU), and then using the sysfs
API from this series a sorted breakdown of currently running clients. Output of
that looks like this:

  rcs0:  88.89% busy (  0.27 qd),   0.00% wait,   0.00% sema
  bcs0:  27.50% busy (  0.00 qd),   0.00% wait,   0.00% sema
  vcs0:  86.75% busy (  0.00 qd),   0.00% wait,   0.00% sema
 vecs0:   0.00% busy (  0.00 qd),   0.00% wait,   0.00% sema

         gem_wsim[ 21806]:  rcs0:  43.18%  bcs0:  27.52%  vcs0:  86.80%  vecs0:   0.00%
             Xorg[ 21451]:  rcs0:  33.60%  bcs0:   0.00%  vcs0:   0.00%  vecs0:   0.00%
  chromium-browse[ 21670]:  rcs0:  12.16%  bcs0:   0.00%  vcs0:   0.00%  vecs0:   0.00%
             Xorg[ 21451]:  rcs0:   0.00%  bcs0:   0.00%  vcs0:   0.00%  vecs0:   0.00%
            xfwm4[ 21506]:  rcs0:   0.00%  bcs0:   0.00%  vcs0:   0.00%  vecs0:   0.00%

Patch series needs some more work, for instance the summation of busyness per-
client is not the most efficient, standard debugging and so, but seems to work
reasonably well so far.

What is also possible to do on top of this series, is to extend our current PMU
interface to allow per-task profiling via perf tool as well. I have started
implementing that at one point and could go back to it.

Tvrtko Ursulin (6):
  drm/i915: Track per-context engine busyness
  drm/i915: Allow clients to query own per-engine busyness
  drm/i915: Expose list of clients in sysfs
  drm/i915: Update client name on context create
  drm/i915: Expose per-engine client busyness
  drm/i915: Add sysfs toggle to enable per-client engine stats

 drivers/gpu/drm/i915/i915_drv.h         |  24 +++++
 drivers/gpu/drm/i915/i915_gem.c         | 178 ++++++++++++++++++++++++++++++--
 drivers/gpu/drm/i915/i915_gem_context.c |  56 +++++++++-
 drivers/gpu/drm/i915/i915_gem_context.h |   6 ++
 drivers/gpu/drm/i915/i915_sysfs.c       |  80 ++++++++++++++
 drivers/gpu/drm/i915/intel_engine_cs.c  |  32 ++++++
 drivers/gpu/drm/i915/intel_lrc.c        |  14 ++-
 drivers/gpu/drm/i915/intel_ringbuffer.h |  50 +++++++--
 include/uapi/drm/i915_drm.h             |  11 +-
 9 files changed, 427 insertions(+), 24 deletions(-)

-- 
2.14.1
_______________________________________________
Intel-gfx mailing list
Intel-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/intel-gfx