[PATCH 2/2] drm/amdkfd: change svm range evict

[Felix Kuehling]

On 2022-06-28 17:43, Eric Huang wrote:
> Two changes:
> 1. reducing unnecessary evict/unmap when range is not mapped to gpu.
> 2. adding always evict when flags is set to always_mapped.
>
> Signed-off-by: Eric Huang <jinhuieric.huang at amd.com>
> ---
>   drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 10 ++++++++--
>   1 file changed, 8 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
> index 4bf2f75f853b..76e817687ef9 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
> @@ -1767,12 +1767,16 @@ svm_range_evict(struct svm_range *prange, struct mm_struct *mm,
>   	struct kfd_process *p;
>   	int r = 0;
>   
> +	if (!prange->mapped_to_gpu)
> +		return 0;

This feels like an unrelated optimization that should be in a separate 
patch.

But I'm not sure this is correct, because it doesn't consider child 
ranges. svm_range_unmap_from_gpus already contains this check, so ranges 
should not be unmapped unnecessarily either way. Is there any other 
benefit to this change that I'm missing?

Regards,
   Felix


> +
>   	p = container_of(svms, struct kfd_process, svms);
>   
>   	pr_debug("invalidate svms 0x%p prange [0x%lx 0x%lx] [0x%lx 0x%lx]\n",
>   		 svms, prange->start, prange->last, start, last);
>   
> -	if (!p->xnack_enabled) {
> +	if (!p->xnack_enabled ||
> +	    (prange->flags & KFD_IOCTL_SVM_FLAG_GPU_ALWAYS_MAPPED)) {
>   		int evicted_ranges;
>   
>   		list_for_each_entry(pchild, &prange->child_list, child_list) {
> @@ -3321,7 +3325,9 @@ svm_range_set_attr(struct kfd_process *p, struct mm_struct *mm,
>   		if (r)
>   			goto out_unlock_range;
>   
> -		if (migrated && !p->xnack_enabled) {
> +		if (migrated && (!p->xnack_enabled ||
> +		    (prange->flags & KFD_IOCTL_SVM_FLAG_GPU_ALWAYS_MAPPED)) &&
> +		    prange->mapped_to_gpu) {
>   			pr_debug("restore_work will update mappings of GPUs\n");
>   			mutex_unlock(&prange->migrate_mutex);
>   			continue;