Hi Iago,

On 19/11/26 05:17, Iago Toral wrote:
Looks good to me:
Reviewed-by: Iago Toral Quiroga <[email protected]>

I was not aware that hang checks were happening at second boundaries,
regardless of the time at which the hangcheck was queued, it doesn't
sound like a good idea, so I think it is good to change that.

On the other hand, now we rearm the hang check every 100ms instead
which is a much smaller window. I understand this reduces visible
impact of the workaround when the issue happens, but I am wondering
whether this alone has any visible performance impact for the 99% of
the time we are not hitting the problem, I presume you haven't seen
any?


I benchmarked glmark2-wayland --off-screen (800x600, under labwc) and
glmark2-es2-drm (1920x1080 fullscreen), which I believe are
representative applications for RPi 3B+, and I observed no performance
impact. I even tested 50 ms and I didn't observe a performance impact as
well.

If you have other applications you would like to see the benchmark
results, just let me know and I can test it.

Best regards,
- Maíra

Iago

El mié, 18-11-2026 a las 18:47 -0300, Maíra Canal escribió:
Mesa issue #16652 [1] reports GPU resets every few minutes on the
Raspberry Pi 3. They happen when text labels are drawn over their
shadows:
two different threaded fragment shaders, plus untextured single-
threaded
ones, in the same tiles.

During a hang, the performance counters show four QPUs fetching
instructions with 0 cache hits, ~6.5M misses per 100 ms, and 0 valid
instructions executed. This matches HW erratum GFXH-16: the QPUs of a
slice share one instruction cache and can livelock, evicting each
other's
fills. Since shader code always starts at offset 0 of a page-aligned
BO,
every shader competes for the same cache sets.

This series works around it:

   - Patch 3 starts each shader's code at one of four offsets in its
BO,
     spreading shaders across cache sets.

   - Patch 4 clears the slice instruction and uniforms caches when the
     hangcheck finds no progress, and resets the GPU only if the stall
     persists. It is meant to recover from livelocks the offsets don't
     prevent.

Patches 1 and 2 fix an off-by-one in the shader BO handle check and a
NULL pointer dereference in validate_gl_shader_rec(), found while
working
on patch 3.

Tested on a Raspberry Pi 3B+ with the issue's reproducer, 4 x 15 min
runs:

   without the series:  29 GPU resets
   with patches 1-3:     0 GPU resets

[1] https://gitlab.freedesktop.org/mesa/mesa/-/issues/16652

Best regards,
- Maíra

---
Maíra Canal (4):
       drm/vc4: Fix off-by-one in the shader record BO handle check
       drm/vc4: Check for a validated shader before reading its thread
mode
       drm/vc4: Start shader code at varying offsets in its BO
       drm/vc4: Clear the slice caches before resetting a stalled GPU

  drivers/gpu/drm/vc4/vc4_bo.c               | 43
+++++++++++++++++++++++-------
  drivers/gpu/drm/vc4/vc4_drv.h              | 15 ++++++++++-
  drivers/gpu/drm/vc4/vc4_gem.c              | 20 ++++++++++++--
  drivers/gpu/drm/vc4/vc4_validate.c         | 17 +++++++-----
  drivers/gpu/drm/vc4/vc4_validate_shaders.c |  7 ++---
  5 files changed, 79 insertions(+), 23 deletions(-)
---
base-commit: 80ac6d197218c3399d66674167a4d7c9051a25e2
change-id: 20261112-vc4-icache-workaround-d11b444bc7e8




Reply via email to