[RFC,0/4] drm/i915/perf: add OA interrupt support

Message

Lionel Landwerlin Dec. 19, 2018, 2:37 p.m. UTC

Hi all,

Tests run by the GPA/vTune teams reported a high CPU consumption on
polling data from the i915-perf stream. This is due to a design
decision to check the OA buffer head/tail pointers 200 times per
seconds. The reasonning being not using the interrupt was initially
that you could want to process data as soon as it's available.

To avoid breaking the existing behavior of the i915-perf stream, we'll
introduce 2 new options at opening of the i915-perf stream to allow
the application to choose :
    - how often it wants the OA circular buffer registers to be checked
    - whether to make sure of the interrupt

In the case of GPA, collecting 100k OA reports would initially consume
17~20% CPU. With these 2 new parameters set to 1 second for OA
circular buffer checks and the interrupt enabled, CPU usage drops to
2~3%.

I'm looking for feedback as to whether those 2 new opening parameters
are alright.

Since this might be the first time that we add a new parameter to the
perf stream open ioctl, we would also need a way to detect their
availability. So far in my experiments I've used the following trick :
specify an invalid context id parameter at open and expect ENOENT when
the parameter is available, EINVAL otherwise. I'm open to other ways
of doing this.

Cheers,

Lionel Landwerlin (4):
  drm/i915/perf: rework aging tail workaround
  drm/i915/perf: add new open param to configure polling of OA buffer
  drm/i915/perf: handle interrupts from the OA unit
  drm/i915/perf: add interrupt enabling parameter

 drivers/gpu/drm/i915/i915_drv.h         |  59 +++--
 drivers/gpu/drm/i915/i915_irq.c         |  39 +++-
 drivers/gpu/drm/i915/i915_perf.c        | 273 +++++++++++++++---------
 drivers/gpu/drm/i915/i915_reg.h         |   7 +
 drivers/gpu/drm/i915/intel_ringbuffer.c |   2 +
 include/uapi/drm/i915_drm.h             |  15 ++
 6 files changed, 276 insertions(+), 119 deletions(-)

--
2.20.1