[PATCH v2 0/4] lib/xe: query engine class capabilities from debugfs info

[Xin Wang <x.wang@intel.com>]

The xe kernel driver recently started exposing engine class capability
bitmasks in the debugfs "info" node:

  multi_lrc_engine_classes vcs vecs
  gt0 multi_queue_engine_classes bcs ccs

These entries were introduced by the following KMD series:
  https://patchwork.freedesktop.org/series/163802/
  - drm/xe: improve readability of debugfs engine info output
  - drm/xe: expose multi-lrc engine classes in debugfs info

This series wires up IGT to consume these debugfs entries:

Patch 1 adds two bitmask fields to struct xe_device (multi_lrc_mask and
multi_queue_engine_class_mask) and populates them in xe_device_get() by
reading the debugfs "info" file line-by-line.  UINT16_MAX is used as a
sentinel meaning "not available" so that older kernels that do not expose
these entries continue to work via the existing hardcoded fallbacks.

Patch 2 implements xe_engine_class_supports_multi_lrc(fd, engine_class).
When the kernel-reported bitmask is available it is used directly; otherwise
the function falls back to returning true only for VIDEO_DECODE and
VIDEO_ENHANCE, matching the previous hardcoded behaviour.

Patch 3 refactors xe_engine_class_supports_multi_queue() in the same way,
adding an fd parameter so it can consult the kernel-reported bitmask.  All
callers (direct calls in xe_exec_threads.c and macro users in
xe_exec_multi_queue.c) are updated accordingly.  Note that
xe_for_each_multi_queue_engine_class() gains a mandatory fd parameter as
part of this change.

Patch 4 guards the balancer and multi-LRC codepaths in xe_exec_balancer,
xe_exec_reset, xe_exec_threads and xe_drm_fdinfo with
xe_engine_class_supports_multi_lrc() so that tests are skipped rather than
failing on engine classes whose hardware does not support multi-LRC
submission.  It also ensures the parallel-queue rejection check in
xe_exec_multi_queue __test_sanity() is only exercised for engine classes
that support multi-LRC.

v2:
  - Merged previous xe_exec_multi_queue expected-error fix into patch 4 two
    patches share the same logical motivation.
  - xe_exec_multi_queue: use `if (n > 1 && xe_engine_class_supports_multi_lrc())`
    instead of a ternary to select the error code (Niranjana)

Xin Wang (4):
  lib/xe: cache engine class masks from debugfs info
  lib/xe: add xe_engine_class_supports_multi_lrc()
  lib/xe: use debugfs info to implement
    xe_engine_class_supports_multi_queue()
  tests/intel: skip or adjust tests for non-multi-LRC engine classes

 lib/xe/xe_query.c                 | 156 +++++++++++++++++++++++++++---
 lib/xe/xe_query.h                 |  21 +++-
 tests/intel/xe_drm_fdinfo.c       |   2 +-
 tests/intel/xe_exec_balancer.c    |   6 +-
 tests/intel/xe_exec_multi_queue.c |   6 +-
 tests/intel/xe_exec_reset.c       |   2 +-
 tests/intel/xe_exec_threads.c     |   5 +-
 7 files changed, 168 insertions(+), 30 deletions(-)

-- 
2.43.0