[PATCH v4 0/5] Fix serialization on burst of unbinds - v2

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

Attempting to resolve part of [1]; this solution differs than v1 [2] by
changing last fence semantics detailed in the patches.

Overview of issue in [1]:

When a burst of unbind jobs is issued, a dependency chain can form
between the TLB invalidation of a previous unbind job and the current
one. This leads to undesirable serialization, causing current jobs to
wait unnecessarily for prior TLB invalidations, execute on the GPU when
not needed, and significantly slow down the unbind burst—resulting in up
to a 4× to 8× slowdown depending on platform.
 
Matt

[1] https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/6047
[2] https://patchwork.freedesktop.org/series/156144/

Matthew Brost (5):
  drm/xe: Add last fence attachment to TLB invalidation job queues
  drm/xe: Decouple bind queue last fence from TLB invalidations
  drm/xe: Do not wait on TLB invalidations in page fault binds
  drm/xe: Don't allow in fences on zero batch exec or zero binds
  drm/xe: Remove last fence dependecy check from binds

 drivers/gpu/drm/xe/xe_exec_queue.c       | 108 ++++++++++++++++++++---
 drivers/gpu/drm/xe/xe_exec_queue.h       |  20 ++++-
 drivers/gpu/drm/xe/xe_exec_queue_types.h |   5 ++
 drivers/gpu/drm/xe/xe_migrate.c          |  14 +++
 drivers/gpu/drm/xe/xe_migrate.h          |   8 ++
 drivers/gpu/drm/xe/xe_pt.c               |  80 ++++++-----------
 drivers/gpu/drm/xe/xe_sync.c             |  74 +++++++++-------
 drivers/gpu/drm/xe/xe_tlb_inval_job.c    |  31 +++++--
 drivers/gpu/drm/xe/xe_tlb_inval_job.h    |   5 +-
 drivers/gpu/drm/xe/xe_vm.c               |  93 +++++++++++--------
 drivers/gpu/drm/xe/xe_vm_types.h         |   6 +-
 11 files changed, 290 insertions(+), 154 deletions(-)

-- 
2.34.1