mbox series

[0/4] drm/xe/xe_drm_client: Add per drm client reset stats

Message ID 20250214203757.27895-1-jonathan.cavitt@intel.com (mailing list archive)
Headers show
Series drm/xe/xe_drm_client: Add per drm client reset stats | expand

Message

Jonathan Cavitt Feb. 14, 2025, 8:37 p.m. UTC
Add additional information to drm client so it can report the last 50
exec queues to have been banned on it, as well as the last pagefault
seen when said exec queues were banned. Since we cannot reasonably
associate a pagefault to a specific exec queue, we currently report the
last seen pagefault on the associated hw engine instead.

The last pagefault seen per exec queue is saved to the hw engine, and the
pagefault is updated during the pagefault handling process in
xe_gt_pagefault. The last seen pagefault is reset when the engine is
reset because any future exec queue bans likely were not caused by said
pagefault after the reset.

Also add a tracker that counts the number of times the drm client has
experienced an engine reset.

Signed-off-by: Jonathan Cavitt <jonathan.cavitt@intel.com>

Jonathan Cavitt (4):
  drm/xe/xe_exec_queue: Add ID param to exec queue struct
  drm/xe/xe_gt_pagefault: Migrate pagefault struct to header
  FIXME: drm/xe/xe_drm_client: Add per drm client pagefault info
  drm/xe/xe_drm_client: Add per drm client reset stats

 drivers/gpu/drm/xe/xe_drm_client.c       | 130 +++++++++++++++++++++++
 drivers/gpu/drm/xe/xe_drm_client.h       |  38 +++++++
 drivers/gpu/drm/xe/xe_exec_queue.c       |   8 ++
 drivers/gpu/drm/xe/xe_exec_queue_types.h |   2 +
 drivers/gpu/drm/xe/xe_gt_pagefault.c     |  46 ++++----
 drivers/gpu/drm/xe/xe_gt_pagefault.h     |  51 +++++++++
 drivers/gpu/drm/xe/xe_guc_submit.c       |  19 ++++
 drivers/gpu/drm/xe/xe_hw_engine.c        |   4 +
 drivers/gpu/drm/xe/xe_hw_engine_types.h  |   8 ++
 9 files changed, 279 insertions(+), 27 deletions(-)

Comments

André Almeida Feb. 18, 2025, 7:21 p.m. UTC | #1
Hi Jonathan,

Em 14/02/2025 17:37, Jonathan Cavitt escreveu:
> Add additional information to drm client so it can report the last 50
> exec queues to have been banned on it, as well as the last pagefault
> seen when said exec queues were banned. Since we cannot reasonably
> associate a pagefault to a specific exec queue, we currently report the
> last seen pagefault on the associated hw engine instead.
> 
> The last pagefault seen per exec queue is saved to the hw engine, and the
> pagefault is updated during the pagefault handling process in
> xe_gt_pagefault. The last seen pagefault is reset when the engine is
> reset because any future exec queue bans likely were not caused by said
> pagefault after the reset.
> 
> Also add a tracker that counts the number of times the drm client has
> experienced an engine reset.

What's the use case for this? How will userspace consume this information?

> 
> Signed-off-by: Jonathan Cavitt <jonathan.cavitt@intel.com>
> 
> Jonathan Cavitt (4):
>    drm/xe/xe_exec_queue: Add ID param to exec queue struct
>    drm/xe/xe_gt_pagefault: Migrate pagefault struct to header
>    FIXME: drm/xe/xe_drm_client: Add per drm client pagefault info
>    drm/xe/xe_drm_client: Add per drm client reset stats
> 
>   drivers/gpu/drm/xe/xe_drm_client.c       | 130 +++++++++++++++++++++++
>   drivers/gpu/drm/xe/xe_drm_client.h       |  38 +++++++
>   drivers/gpu/drm/xe/xe_exec_queue.c       |   8 ++
>   drivers/gpu/drm/xe/xe_exec_queue_types.h |   2 +
>   drivers/gpu/drm/xe/xe_gt_pagefault.c     |  46 ++++----
>   drivers/gpu/drm/xe/xe_gt_pagefault.h     |  51 +++++++++
>   drivers/gpu/drm/xe/xe_guc_submit.c       |  19 ++++
>   drivers/gpu/drm/xe/xe_hw_engine.c        |   4 +
>   drivers/gpu/drm/xe/xe_hw_engine_types.h  |   8 ++
>   9 files changed, 279 insertions(+), 27 deletions(-)
>