[PATCH v4 0/5] drm/xe: Allow migrate vm gpu submissions from reclaim context

["Thomas Hellström" <thomas.hellstrom@linux.intel.com>]

On lunar lake, we need to be able to copy CCS compression metadata from
reclaim context when using a shrinker.
ince we do dma_fence allocations under the
struct xe_migrate::job_mutex, we're facing a lockdep violation.
This patchset deals with moving the job + fence allocation out of
the job_mutex by splitting up the fence allocation- and
initialization.

v2:
- Add a patch from Matthew Brost to fix up the use of
  composite fence seqno.
- A couple of bugfixes as detailed in patch 3.
v3:
- Rebase on drm-tip.
v4:
- Add an assert (Matthew Brost)

Matthew Brost (1):
  drm/xe: Decouple job seqno and lrc seqno

Thomas Hellström (4):
  drm/xe: Split lrc seqno fence creation up
  drm/xe: Don't initialize fences at xe_sched_job_create()
  drm/xe: Remove xe_lrc_create_seqno_fence()
  drm/xe: Move job creation out of the struct xe_migrate::job_mutex

 drivers/gpu/drm/xe/xe_exec_queue.c       |   5 -
 drivers/gpu/drm/xe/xe_exec_queue_types.h |  10 --
 drivers/gpu/drm/xe/xe_guc_submit.c       |   5 +-
 drivers/gpu/drm/xe/xe_hw_fence.c         |  59 ++++++--
 drivers/gpu/drm/xe/xe_hw_fence.h         |   7 +-
 drivers/gpu/drm/xe/xe_lrc.c              |  39 +++++-
 drivers/gpu/drm/xe/xe_lrc.h              |   4 +-
 drivers/gpu/drm/xe/xe_migrate.c          |  17 ++-
 drivers/gpu/drm/xe/xe_ring_ops.c         |  22 +--
 drivers/gpu/drm/xe/xe_sched_job.c        | 167 +++++++++++++----------
 drivers/gpu/drm/xe/xe_sched_job.h        |   5 +
 drivers/gpu/drm/xe/xe_sched_job_types.h  |  20 ++-
 drivers/gpu/drm/xe/xe_trace.h            |   9 +-
 13 files changed, 237 insertions(+), 132 deletions(-)

-- 
2.44.0