Re: [PATCHv2 4/8] drm/amdgpu: Attribute UALink loopback messages to the correct sender

[Felix Kuehling <felix.kuehling@amd.com>]

On 2026-10-27 15:54, Mukul Joshi wrote:
> In UALink loopback mode a single GPU acts as both the exporter and the
> importer, so the sender accel ID of a message depends on its direction.
>
> - Send HELLO as the opposite role of the destination: a message to the
>    exporter is sent as the importer and vice versa. Outside loopback
>    this is still the GPU's own accel ID.
>
> - The MP nHT stamps the IH cookie of every HW-delivered message with
>    the GPU ID of the single physical station, the exporter. The only HW
>    path is importer -> exporter, so attribute these messages to the
>    importer. SW-emulated exporter -> importer messages carry no IH ring
>    and already name the exporter as the sender.
>
> Signed-off-by: Mukul Joshi <mukul.joshi@amd.com>

Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>


> ---
> v1->v2:
> - Use ualink_exp_accel_id/ualink_imp_accel_id() wrappers in
>    amdgpu_ualink_process_irq(). (Lijo)
>
>   drivers/gpu/drm/amd/amdgpu/amdgpu_ualink.c | 32 +++++++++++++++++++++-
>   1 file changed, 31 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ualink.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ualink.c
> index 6c70be7badec..96a6023f51c5 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ualink.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ualink.c
> @@ -1090,6 +1090,23 @@ static inline u32 ualink_imp_accel_id(struct amdgpu_device *adev)
>   	return adev->ualink.info->ppod.accel_id;
>   }
>   
> +/*
> + * amdgpu_ualink_local_sender_accel_id - sender accel_id for an outgoing message
> + *
> + * Normally a GPU sends messages as its own (single) accelerator. In loopback
> + * mode the same GPU plays both roles, so the sender identity depends on the
> + * destination: a message to the exporter is sent as the importer and vice
> + * versa. Outside loopback both roles collapse to the GPU's own accel_id.
> + */
> +static u32 amdgpu_ualink_local_sender_accel_id(struct amdgpu_device *adev,
> +					       u32 remote_acc_id)
> +{
> +	if (remote_acc_id == ualink_exp_accel_id(adev))
> +		return ualink_imp_accel_id(adev);
> +
> +	return ualink_exp_accel_id(adev);
> +}
> +
>   static bool amdgpu_ualink_is_local_accel(struct amdgpu_device *adev,
>   					 u32 accel_id)
>   {
> @@ -2238,7 +2255,10 @@ static int amdgpu_ualink_send_hello_msg(struct amdgpu_device *adev,
>   
>   	dw0 = AMDGPU_UALINK_HELLO_MSG;
>   	dw0 |= (remote_acc_id << AMDGPU_UALINK_HELLO_MSG_RECV_ACCID_SHIFT);
> -	dw0 |= (adev->ualink.info->ppod.accel_id <<
> +	/* In loopback the sender is the opposite role of the destination, so
> +	 * the exporter side tracks the connection under the importer accel_id.
> +	 */
> +	dw0 |= (amdgpu_ualink_local_sender_accel_id(adev, remote_acc_id) <<
>   		AMDGPU_UALINK_HELLO_MSG_SENDER_ACCID_SHIFT);
>   
>   	dev_dbg(adev->dev, "SEND HELLO: HELLO message to remote AccId:%u\n",
> @@ -6548,6 +6568,16 @@ static int amdgpu_ualink_process_irq(struct amdgpu_device *adev,
>   		return handled;
>   	}
>   
> +	/* UALink loopback: the MP nHT stamps every HW-delivered message with
> +	 * the GPU ID of the single physical station, the exporter. The only
> +	 * HW path is importer -> exporter, so the real sender is the importer.
> +	 * SW-emulated exporter -> importer messages carry no IH ring and
> +	 * already name the exporter as sender.
> +	 */
> +	if (adev->ualink.loopback.enabled && entry->ih &&
> +	    src_acc_id == ualink_exp_accel_id(adev))
> +		src_acc_id = ualink_imp_accel_id(adev);
> +
>   	msg_type = dw0 & AMDGPU_UALINK_MESSAGE_HEADER_MASK;
>   	local_acc_id = adev->ualink.info->ppod.accel_id;
>   	dev_dbg(adev->dev, "Got MSG: remote acc_id %u msg_type %u\n",