[PATCH 0/6] Fix performance when pagefaults and 3d/display share resources

[Matthew Brost <matthew.brost@intel.com>]

xello_raytracer [1] showed a significant performance regression (~40×
slowdown) when the compute UMD enabled page faults. The issue was traced
to the implementation that allows page-fault workloads and dma-fence
workloads (3D/display) to coexist.

This series fixes several issues, bringing page-fault performance in
this benchmark to within 6% of preempt-fence mode (no page faults).

Matt

[1] https://github.com/intel-sandbox/xello_raytracer

Matthew Brost (6):
  drm/xe: Adjust long-running workload timeslices to reasonable values
  drm/xe: Use usleep_range for accurate long-running workload
    timeslicing
  drm/xe: Add debugfs knobs to control long running workload timeslicing
  drm/xe: Skip exec queue schedule toggle if queue is idle during
    suspend
  drm/xe: Wait on in-syncs when swicthing to dma-fence mode
  drm/xe: Add more GT stats around pagefault mode switch flows

 drivers/gpu/drm/xe/xe_debugfs.c         | 74 +++++++++++++++++++++++++
 drivers/gpu/drm/xe/xe_device.c          |  1 +
 drivers/gpu/drm/xe/xe_device_types.h    |  6 ++
 drivers/gpu/drm/xe/xe_exec.c            |  3 +-
 drivers/gpu/drm/xe/xe_exec_queue.h      | 17 ++++++
 drivers/gpu/drm/xe/xe_gt_stats.c        |  6 ++
 drivers/gpu/drm/xe/xe_gt_stats_types.h  |  3 +
 drivers/gpu/drm/xe/xe_guc_submit.c      | 46 +++++++++++++--
 drivers/gpu/drm/xe/xe_hw_engine_group.c | 66 +++++++++++++++++++---
 drivers/gpu/drm/xe/xe_hw_engine_group.h |  4 +-
 drivers/gpu/drm/xe/xe_sync.c            | 14 +++++
 drivers/gpu/drm/xe/xe_sync.h            |  1 +
 drivers/gpu/drm/xe/xe_vm.c              |  5 +-
 drivers/gpu/drm/xe/xe_vm_types.h        |  2 +-
 14 files changed, 232 insertions(+), 16 deletions(-)

-- 
2.34.1