Intel-XE Archive on lore.kernel.org
 help / color / mirror / Atom feed
* [PATCH v6 0/6] Fix serialization on burst of unbinds - v2
@ 2025-10-31 18:09 Matthew Brost
  2025-10-31 18:09 ` [PATCH v6 1/6] drm/xe: Enforce correct user fence signaling order using Matthew Brost
                   ` (9 more replies)
  0 siblings, 10 replies; 11+ messages in thread
From: Matthew Brost @ 2025-10-31 18:09 UTC (permalink / raw)
  To: intel-xe; +Cc: thomas.hellstrom

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 (6):
  drm/xe: Enforce correct user fence signaling order using
  drm/xe: Attach last fence to TLB invalidation job queues
  drm/xe: Decouple bind queue last fence from TLB invalidations
  drm/xe: Skip TLB invalidation waits in page fault binds
  drm/xe: Disallow input fences on zero batch execs and zero binds
  drm/xe: Remove last fence dependency check from binds and execs

 drivers/gpu/drm/xe/xe_exec.c             |   7 +-
 drivers/gpu/drm/xe/xe_exec_queue.c       | 119 ++++++++++++++++++++---
 drivers/gpu/drm/xe/xe_exec_queue.h       |  23 ++++-
 drivers/gpu/drm/xe/xe_exec_queue_types.h |  12 +++
 drivers/gpu/drm/xe/xe_migrate.c          |  14 +++
 drivers/gpu/drm/xe/xe_migrate.h          |   8 ++
 drivers/gpu/drm/xe/xe_oa.c               |  45 ++++++---
 drivers/gpu/drm/xe/xe_oa_types.h         |   8 ++
 drivers/gpu/drm/xe/xe_pt.c               |  80 +++++----------
 drivers/gpu/drm/xe/xe_sched_job.c        |  17 ----
 drivers/gpu/drm/xe/xe_sched_job.h        |   1 -
 drivers/gpu/drm/xe/xe_sync.c             |  91 ++++++++++-------
 drivers/gpu/drm/xe/xe_sync.h             |   3 +
 drivers/gpu/drm/xe/xe_sync_types.h       |   3 +
 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               |  97 ++++++++++--------
 drivers/gpu/drm/xe/xe_vm_types.h         |   6 +-
 18 files changed, 376 insertions(+), 194 deletions(-)

-- 
2.34.1


^ permalink raw reply	[flat|nested] 11+ messages in thread

end of thread, other threads:[~2025-11-01  9:44 UTC | newest]

Thread overview: 11+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2025-10-31 18:09 [PATCH v6 0/6] Fix serialization on burst of unbinds - v2 Matthew Brost
2025-10-31 18:09 ` [PATCH v6 1/6] drm/xe: Enforce correct user fence signaling order using Matthew Brost
2025-10-31 18:09 ` [PATCH v6 2/6] drm/xe: Attach last fence to TLB invalidation job queues Matthew Brost
2025-10-31 18:09 ` [PATCH v6 3/6] drm/xe: Decouple bind queue last fence from TLB invalidations Matthew Brost
2025-10-31 18:09 ` [PATCH v6 4/6] drm/xe: Skip TLB invalidation waits in page fault binds Matthew Brost
2025-10-31 18:09 ` [PATCH v6 5/6] drm/xe: Disallow input fences on zero batch execs and zero binds Matthew Brost
2025-10-31 18:09 ` [PATCH v6 6/6] drm/xe: Remove last fence dependency check from binds and execs Matthew Brost
2025-10-31 19:53 ` ✗ CI.checkpatch: warning for Fix serialization on burst of unbinds - v2 Patchwork
2025-10-31 19:54 ` ✓ CI.KUnit: success " Patchwork
2025-10-31 21:10 ` ✗ Xe.CI.BAT: failure " Patchwork
2025-11-01  9:44 ` ✗ Xe.CI.Full: " Patchwork

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox