[PATCH] Fix Incorrect VMIDs passed to HWS

[Felix Kuehling]

Am 2022-03-17 um 15:37 schrieb Tushar Patel:
> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c |  2 +-
>   drivers/gpu/drm/amd/amdkfd/kfd_device.c | 21 ++++++++++++---------
>   2 files changed, 13 insertions(+), 10 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
> index 4c20c23d6ba0..bda1b5132ee8 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
> @@ -680,7 +680,7 @@ MODULE_PARM_DESC(sched_policy,
>    * Maximum number of processes that HWS can schedule concurrently. The maximum is the
>    * number of VMIDs assigned to the HWS, which is also the default.
>    */
> -int hws_max_conc_proc = 8;
> +int hws_max_conc_proc = -1;
>   module_param(hws_max_conc_proc, int, 0444);
>   MODULE_PARM_DESC(hws_max_conc_proc,
>   	"Max # processes HWS can execute concurrently when sched_policy=0 (0 = no concurrency, #VMIDs for KFD = Maximum(default))");
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
> index 339e12c94cff..39073f72fe5f 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
> @@ -483,15 +483,18 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd,
>   	}
>   
>   	/* Verify module parameters regarding mapped process number*/
> -	if ((hws_max_conc_proc < 0)
> -			|| (hws_max_conc_proc > kfd->vm_info.vmid_num_kfd)) {
> -		dev_err(kfd_device,
> -			"hws_max_conc_proc %d must be between 0 and %d, use %d instead\n",
> -			hws_max_conc_proc, kfd->vm_info.vmid_num_kfd,
> -			kfd->vm_info.vmid_num_kfd);
> -		kfd->max_proc_per_quantum = kfd->vm_info.vmid_num_kfd;
> -	} else
> -		kfd->max_proc_per_quantum = hws_max_conc_proc;
> +	kfd->max_proc_per_quantum = kfd->vm_info.vmid_num_kfd;
> +	if (hws_max_conc_proc != -1) {
> +		if ((hws_max_conc_proc > 0)
> +				&& (hws_max_conc_proc < kfd->vm_info.vmid_num_kfd)) {

I think this should be <= kfd->vm_info.vmid_num_kfd.


> +			kfd->max_proc_per_quantum = hws_max_conc_proc;
> +		} else {
> +					dev_err(kfd_device,
> +				"hws_max_conc_proc %d must be between 0 and %d, use %d instead\n",
> +				hws_max_conc_proc, kfd->vm_info.vmid_num_kfd,
> +				kfd->vm_info.vmid_num_kfd);

I think this error message is the wrong approach. hws_max_conc_proc is a 
global setting that affects all GPUs. Different GPUs may have different 
numbers of VMIDs. So we can't treat (hws_max_conc_proc > 
kfd->vm_info.vmid_num_kfd) as an error. It may be an error on one GPU 
but perfectly fine on another.

I think you can simplify this if-else like this and get rid of the dev_err:

     kfd->max_proc_per_quantum = min(hws_max_conc_proc, 
kfd->vm_info.vmid_num_kfd);

Regards,
   Felix


> +		}
> +	}
>   
>   	/* calculate max size of mqds needed for queues */
>   	size = max_num_of_queues_per_device *