[PATCH v4 0/2] drm/xe/shrinker: Runtime PM and walk termination fixes

[Shuicheng Lin <shuicheng.lin@intel.com>]

The xe shrinker walks the SYSTEM and TT LRUs without a runtime PM
reference.  Shrinking a bo outside system memory invalidates its GPU
mappings, so with the device runtime suspended the page table zap trips
an assert and the TLB invalidation returns -ENODEV.  Patch 2 takes a
reference before walking a memory type other than XE_PL_SYSTEM.

Patch 1 is a separate problem found while reviewing that one:
xe_bo_shrink() returns -EBUSY for a bo that is simply not a shrink
candidate, and the walk treats it as fatal.  It is first in the series
because its Fixes commit is a year older, so it applies on its own to
stable trees that must not take patch 2.

Verified with igt@xe_madvise@dontneed-before-exec while the GPU is
runtime suspended:

  unmodified                            15/15 runs hit the WARN
  with patch 2                           0/15
  CI-like, device active, 10 runs        10 pass, 0 skip, no splats

ftrace confirms the new path executes, runtime_usage returns to 0, and
there are no lockdep reports.

For patch 1, tracing shows the walk continuing past the -EBUSY and
reclaiming two further bos that previously ended it.  Note that
xe_shrinker_walk() re-enters the walk for each escalation level, so the
bos are usually still reached on a later pass and the pages reported to
mm were unchanged in that test - the effect is repeated restarts rather
than lost reclaim.

Shuicheng Lin (2):
  drm/xe/shrinker: Do not abort the walk on unshrinkable bos
  drm/xe/shrinker: Take a runtime PM ref before shrinking non-system
    memory

 drivers/gpu/drm/xe/xe_shrinker.c | 42 +++++++++++++++++++++++++-------
 1 file changed, 33 insertions(+), 9 deletions(-)

-- 
2.43.0