On 09/09/2020 14:23, Steven Price wrote:
> On 08/09/2020 16:18, Neil Armstrong wrote:
>> The T820, G31 & G52 GPUs integratewd by Amlogic in the respective GXM, G12A/SM1 & G12B
>> SoCs needs a quirk in the PWR registers at the GPU reset time.
>>
>> This adds a callback in the device compatible struct of permit this.
>>
>> Signed-off-by: Neil Armstrong <narmstrong@xxxxxxxxxxxx>
>> ---
>> drivers/gpu/drm/panfrost/panfrost_device.h | 3 +++
>> drivers/gpu/drm/panfrost/panfrost_gpu.c | 4 ++++
>> 2 files changed, 7 insertions(+)
>>
>> diff --git a/drivers/gpu/drm/panfrost/panfrost_device.h b/drivers/gpu/drm/panfrost/panfrost_device.h
>> index 2cf1a6a13af8..4c9cd5452ba5 100644
>> --- a/drivers/gpu/drm/panfrost/panfrost_device.h
>> +++ b/drivers/gpu/drm/panfrost/panfrost_device.h
>> @@ -73,6 +73,9 @@ struct panfrost_compatible {
>> /* IOMMU quirks flags */
>> unsigned long pgtbl_quirks;
>> +
>> + /* Vendor implementation quirks at reset time callback */
>> + void (*vendor_reset_quirk)(struct panfrost_device *pfdev);
>> };
>> struct panfrost_device {
>> diff --git a/drivers/gpu/drm/panfrost/panfrost_gpu.c b/drivers/gpu/drm/panfrost/panfrost_gpu.c
>> index e0f190e43813..c129aaf77790 100644
>> --- a/drivers/gpu/drm/panfrost/panfrost_gpu.c
>> +++ b/drivers/gpu/drm/panfrost/panfrost_gpu.c
>> @@ -62,6 +62,10 @@ int panfrost_gpu_soft_reset(struct panfrost_device *pfdev)
>> gpu_write(pfdev, GPU_INT_CLEAR, GPU_IRQ_RESET_COMPLETED);
>> gpu_write(pfdev, GPU_CMD, GPU_CMD_SOFT_RESET);
>> + /* The Amlogic GPU integration needs quirks at this stage */
>> + if (pfdev->comp->vendor_reset_quirk)
>> + pfdev->comp->vendor_reset_quirk(pfdev);
>> +
>> ret = readl_relaxed_poll_timeout(pfdev->iomem + GPU_INT_RAWSTAT,
>> val, val & GPU_IRQ_RESET_COMPLETED, 100, 10000);
>
> Placing the quirk before the reset has completed is dodgy. Can this be ordered after the GPU_IRQ_RESET_COMPLETED signal has been seen? The problem is the reset could (in theory) cause a power transition (e.g. if the GPU is reset while a core is powered) and changing the PWR_OVERRIDEx registers during a transition is undefined. But I don't know the details of how the hardware is broken so it is possible the override is needed for the reset to complete so this would need testing.
>
> I also wonder if this could live in panfrost_gpu_init_quirks() instead? Although that is mostly about quirks common to all Mali GPU implementations rather than a specific implementation. Although now I've looked I've noticed we have a bug as we don't appear to reapply those quirks after a reset - I'll send a patch!
Indeed, it needs to be applied after each reset, so if you send a patch for this pretty sure it could live in panfrost_gpu_init_quirks().
Neil
>
> Steve
_______________________________________________
dri-devel mailing list
dri-devel@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/dri-devel
