[0/4] A bit of struct drm_sched_job cleanup

Message

Tvrtko Ursulin Jan. 20, 2025, 4:52 p.m. UTC

At one point I thought I wanted to add a member to struct drm_sched_job. As I
noticed there is a hole in the struct, I went to re-order some members to get
rid of it (the hole), at which point I was greeted by a subtle bug cause by the
frequent pattern of:

 job = to_drm_sched_job(spsc_queue_peek|pop(...))

Because to_drm_sched_job is a container of, it implies a lot of the scheduler
(and amdgpu) assumes job->queue_node is the first struct element. Dare to change
that, code will compile but weird things will happen at runtime.

So I thought lets clean that up and remove that hardcoded sneaky assumption and
that is what this series does.

1) Adds a helper so amdgpu can touch the scheduler internals a little bit less.
2) Removes the assumption job->queue_node must be the first element.
3) And finally removes the hole from struct drm_sched_job, fixing one instance
   of type confusion in passing too.

Cc: Christian König <[e-mail]>
Cc: Danilo Krummrich <[e-mail]>
Cc: Matthew Brost <[e-mail]>
Cc: Philipp Stanner <[e-mail]>

Tvrtko Ursulin (4):
  drm/sched: Add job popping API
  drm/amdgpu: Use the new low level job popping helper
  drm/sched: Remove to_drm_sched_job internal helper
  drm/sched: Make the type of drm_sched_job->last_dependency consistent

 drivers/gpu/drm/amd/amdgpu/amdgpu_job.c  |  5 +-
 drivers/gpu/drm/scheduler/sched_entity.c | 20 ++++----
 drivers/gpu/drm/scheduler/sched_main.c   | 10 ++--
 include/drm/gpu_scheduler.h              | 59 ++++++++++++++++--------
 4 files changed, 57 insertions(+), 37 deletions(-)