As reported[1], in the current state of master (that is, *without* that[2] patch, yet unmerged), it is possible to trigger Oops/out-of-bounds errors/unbalanced runtime PM by simply compiling DRM_ACCEL_ROCKET built-in (=y) instead of as a module (=m).
This fixes points 1 and 2 reported here[1] by fixing the unwinding in two functions to properly undo everything done in the same function prior to the error.
Note that this doesn't mean the Rocket device is usable if one core is missing. In fact, it seems it doesn't as I'm hit with many rocket fdac0000.npu: NPU job timed out followed by one rocket fdad0000.npu: NPU job timed out (and that, five times) whenever core0 (fdab0000.npu) fails to probe and I'm running the example from https://docs.mesa3d.org/teflon.html#do-some-inference-with-mobilenetv1 so something else probably needs some additional love.
[1] https://lore.kernel.org/linux-rockchip/0b20d760-ad4f-41c0-b733-39db10d6cc41@... [2] https://lore.kernel.org/linux-rockchip/20251205064739.20270-1-rmxpzlb@gmail....
Signed-off-by: Quentin Schulz quentin.schulz@cherry.de --- Quentin Schulz (2): accel/rocket: fix unwinding in error path in rocket_core_init accel/rocket: fix unwinding in error path in rocket_probe
drivers/accel/rocket/rocket_core.c | 7 +++++-- drivers/accel/rocket/rocket_drv.c | 15 ++++++++++++++- 2 files changed, 19 insertions(+), 3 deletions(-) --- base-commit: a619746d25c8adafe294777cc98c47a09759b3ed change-id: 20251212-rocket-error-path-f9784c46a503
Best regards,