[PATCH v3 0/2] drm/etnaviv: Add GPU reset counters for robustness

[Christian Gmeiner <christian.gmeiner@gmail.com>]

The OpenGL robustness extensions (GL_KHR_robustness) and Vulkan
(VK_ERROR_DEVICE_LOST) let an application detect a GPU reset and check
whether its own context caused it, so it can throw away the broken
context and build a new one. etnaviv already resets the GPU after a
hang, but userspace has no way to learn about it.

This series adds two counters and a RESET_QUERY ioctl that returns both
in one call:

- a global counter that counts every reset of a GPU core.
- a context counter that only counts the resets the calling context was
  guilty of.

Userspace samples both values and compares them later: if the context
counter moved the context was guilty, if only the global counter moved
the context was an innocent victim. That is all that is needed to
implement glGetGraphicsResetStatus() and Vulkan device loss.

The global counter is kept per GPU core and not per device, so a hang
on one pipe does not look like an innocent reset to contexts that only
use another pipe.

The first patch is preparation: the counters are updated from the
scheduler timeout worker, which can race with the DRM file being
closed, so struct etnaviv_file_private becomes reference counted and
every submit holds a reference.

Link to the Mesa MR implementing the userspace side:
https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/42826

Signed-off-by: Christian Gmeiner <cgmeiner@igalia.com>
---
Changes in v3:
- Update the reset counters and read them back under the GPU lock, so
  the RESET_QUERY ioctl cannot run between the two updates and report
  an innocent reset to the guilty context (Lucas).
- Link to v2: https://lore.kernel.org/r/20260710-etnaviv-reset-notification-v2-0-90cfebcf22a2@igalia.com

Changes in v2:
- Replace the two GET_PARAM values with a dedicated RESET_QUERY ioctl
  that returns both counters in one call. The ioctl has a flags field
  that must be zero for now, so the query can later be extended to name
  a specific context once a DRM file can hold more than one context
  (Lucas).
- Make the reset counters plain u32 instead of atomics (Lucas).
- Use __u32 for the ioctl counter fields to match the internal
  counters, following i915 GET_RESET_STATS.
- Link to v1: https://lore.kernel.org/r/20260709-etnaviv-reset-notification-v1-0-64c617496958@igalia.com

---
Christian Gmeiner (2):
      drm/etnaviv: Reference count struct etnaviv_file_private
      drm/etnaviv: Add GPU reset counters

 drivers/gpu/drm/etnaviv/etnaviv_drv.c        | 46 ++++++++++++++++++++++++++--
 drivers/gpu/drm/etnaviv/etnaviv_drv.h        | 12 ++++++++
 drivers/gpu/drm/etnaviv/etnaviv_gem_submit.c |  5 ++-
 drivers/gpu/drm/etnaviv/etnaviv_gpu.h        |  2 ++
 drivers/gpu/drm/etnaviv/etnaviv_sched.c      |  5 +++
 include/uapi/drm/etnaviv_drm.h               | 19 +++++++++++-
 6 files changed, 85 insertions(+), 4 deletions(-)
---
base-commit: 8cdeaa50eae8dad34885515f62559ee83e7e8dda
change-id: 20260708-etnaviv-reset-notification-b037153a1aab

Best regards,
-- 
Christian Gmeiner <cgmeiner@igalia.com>