[PATCH v3] drm/amdkfd: Use partial mapping in GPU page faults

[Felix Kuehling]

On 2023-10-20 17:53, Xiaogang.Chen wrote:
> From: Xiaogang Chen <xiaogang.chen at amd.com>
>
> After partial migration to recover GPU page fault this patch does GPU vm
> space mapping for same page range that got migrated intead of mapping all
> pages of svm range in which the page fault happened.
>
> Signed-off-by: Xiaogang Chen<Xiaogang.Chen at amd.com>
> ---
>   drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 29 ++++++++++++++++++++--------
>   1 file changed, 21 insertions(+), 8 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
> index 54af7a2b29f8..3a71d04779b1 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
> @@ -1619,6 +1619,7 @@ static void *kfd_svm_page_owner(struct kfd_process *p, int32_t gpuidx)
>    * 5. Release page table (and SVM BO) reservation
>    */
>   static int svm_range_validate_and_map(struct mm_struct *mm,
> +				      unsigned long map_start, unsigned long map_last,
>   				      struct svm_range *prange, int32_t gpuidx,
>   				      bool intr, bool wait, bool flush_tlb)
>   {
> @@ -1699,6 +1700,8 @@ static int svm_range_validate_and_map(struct mm_struct *mm,
>   	end = (prange->last + 1) << PAGE_SHIFT;
>   	for (addr = start; !r && addr < end; ) {
>   		struct hmm_range *hmm_range;
> +		unsigned long map_start_vma;
> +		unsigned long map_last_vma;
>   		struct vm_area_struct *vma;
>   		uint64_t vram_pages_vma;
>   		unsigned long next = 0;
> @@ -1747,9 +1750,16 @@ static int svm_range_validate_and_map(struct mm_struct *mm,
>   			r = -EAGAIN;
>   		}
>   
> -		if (!r)
> -			r = svm_range_map_to_gpus(prange, offset, npages, readonly,
> -						  ctx->bitmap, wait, flush_tlb);
> +		if (!r) {
> +			map_start_vma = max(map_start, prange->start + offset);
> +			map_last_vma = min(map_last, prange->start + offset + npages - 1);
> +			if (map_start_vma <= map_last_vma) {
> +				offset = map_start_vma - prange->start;
> +				npages = map_last_vma - map_start_vma + 1;
> +				r = svm_range_map_to_gpus(prange, offset, npages, readonly,
> +							  ctx->bitmap, wait, flush_tlb);
> +			}
> +		}
>   
>   		if (!r && next == end)
>   			prange->mapped_to_gpu = true;
> @@ -1855,8 +1865,8 @@ static void svm_range_restore_work(struct work_struct *work)
>   		 */
>   		mutex_lock(&prange->migrate_mutex);
>   
> -		r = svm_range_validate_and_map(mm, prange, MAX_GPU_INSTANCE,
> -					       false, true, false);
> +		r = svm_range_validate_and_map(mm, prange->start, prange->last, prange,
> +					       MAX_GPU_INSTANCE, false, true, false);
>   		if (r)
>   			pr_debug("failed %d to map 0x%lx to gpus\n", r,
>   				 prange->start);
> @@ -3069,6 +3079,8 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,
>   	kfd_smi_event_page_fault_start(node, p->lead_thread->pid, addr,
>   				       write_fault, timestamp);
>   
> +	start = prange->start;
> +	last = prange->last;

This means, page faults that don't migrate will map the whole range. 
Should we move the proper assignment of start and last out of the 
condition below, so it applies equally to page faults that migrate and 
those that don't?

Regards,
   Felix


>   	if (prange->actual_loc != 0 || best_loc != 0) {
>   		migration = true;
>   		/* Align migration range start and size to granularity size */
> @@ -3102,10 +3114,11 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,
>   		}
>   	}
>   
> -	r = svm_range_validate_and_map(mm, prange, gpuidx, false, false, false);
> +	r = svm_range_validate_and_map(mm, start, last, prange, gpuidx, false,
> +				       false, false);
>   	if (r)
>   		pr_debug("failed %d to map svms 0x%p [0x%lx 0x%lx] to gpus\n",
> -			 r, svms, prange->start, prange->last);
> +			 r, svms, start, last);
>   
>   	kfd_smi_event_page_fault_end(node, p->lead_thread->pid, addr,
>   				     migration);
> @@ -3650,7 +3663,7 @@ svm_range_set_attr(struct kfd_process *p, struct mm_struct *mm,
>   
>   		flush_tlb = !migrated && update_mapping && prange->mapped_to_gpu;
>   
> -		r = svm_range_validate_and_map(mm, prange, MAX_GPU_INSTANCE,
> +		r = svm_range_validate_and_map(mm, prange->start, prange->last, prange, MAX_GPU_INSTANCE,
>   					       true, true, flush_tlb);
>   		if (r)
>   			pr_debug("failed %d to map svm range\n", r);