diff options
| author | Maíra Canal <mcanal@igalia.com> | 2026-11-18 18:47:30 -0300 |
|---|---|---|
| committer | Maíra Canal <mcanal@igalia.com> | 2026-11-19 10:40:54 -0300 |
| commit | c0e1ed41ad86cf5230d64dc5f2c6d2e3aeb59937 (patch) | |
| tree | 0c599c47c837477a8ae51389abe7b444dcab24f3 /scripts | |
| parent | 15da41dff1f66a389d238db60a1c19fa83582d07 (diff) | |
drm/vc4: Start shader code at varying offsets in its BO
VC4 is affected by the HW erratum GFXH-16: the QPUs can get stuck in a
loop where instruction cache loads are evicted before they are used.
The QPUs of a slice share one instruction cache. Shader BOs are page
aligned and the code always starts at offset 0, so instructions at the
same position in different shaders end up mapping to the same cache sets.
When the four QPUs of a slice run four different shaders, their fills
evict each other, no instruction executes and the hangcheck resets the
GPU.
Place each shader's code at one of four offsets, 256 bytes apart, inside
its BO, and add the offset when relocating the shader record. Shader code
is position independent, as the validator only accepts relative branches.
The offsets are tracked per client, so shaders a given client creates in
sequence (which tend to run together) start in different sets.
Notice that this approach makes the livelock unlikely, but not impossible.
The next commit will address the remaining cases.
On a Raspberry Pi 3 running the issue's reproducer, GPU resets went from
dozens in an hour to none.
Cc: stable@vger.kernel.org
Fixes: 463873d57014 ("drm/vc4: Add an API for creating GPU shaders in GEM BOs.")
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/16652
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Link: https://patch.msgid.link/20261118-vc4-icache-workaround-v1-3-5b8adc2c84fe@igalia.com
Signed-off-by: Maíra Canal <mcanal@igalia.com>
Diffstat (limited to 'scripts')
0 files changed, 0 insertions, 0 deletions
