[PATCH v3 2/3] drm/amdkfd: handle svm partial migration cpages 0

[Felix Kuehling]

Am 2021-10-12 um 5:00 p.m. schrieb Philip Yang:
> migrate_vma_setup may return cpages 0, means 0 page can be migrated,
> treat this as error case to skip the rest of vma migration steps.
>
> Change svm_migrate_vma_to_vram and svm_migrate_vma_to_ram to return the
> number of pages migrated successfully or error code. The caller add up
> all the successful migration pages and update prange->actual_loc only if
> the total migrated pages is not 0.
>
> This also removes the warning message "VRAM BO missing during
> validation" if migration cpages is 0.
>
> Signed-off-by: Philip Yang <Philip.Yang at amd.com>
> ---
>  drivers/gpu/drm/amd/amdkfd/kfd_migrate.c | 90 ++++++++++++++----------
>  1 file changed, 51 insertions(+), 39 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c
> index b05c0579d0b9..d37f20b17586 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c
> @@ -376,7 +376,7 @@ svm_migrate_copy_to_vram(struct amdgpu_device *adev, struct svm_range *prange,
>  	return r;
>  }
>  
> -static int
> +static long
>  svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange,
>  			struct vm_area_struct *vma, uint64_t start,
>  			uint64_t end)
> @@ -413,32 +413,37 @@ svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange,
>  			prange->start, prange->last);
>  		goto out_free;
>  	}
> -	if (migrate.cpages != npages) {
> -		pr_debug("Partial migration. 0x%lx/0x%llx pages can be migrated\n",
> -			 migrate.cpages,
> -			 npages);
> -	}
>  
> -	if (migrate.cpages) {
> -		r = svm_migrate_copy_to_vram(adev, prange, &migrate, &mfence,
> -					     scratch);
> -		migrate_vma_pages(&migrate);
> -		svm_migrate_copy_done(adev, mfence);
> -		migrate_vma_finalize(&migrate);
> +	if (migrate.cpages != npages)
> +		pr_debug("partial migration, 0x%lx/0x%llx pages migrated\n",
> +			 migrate.cpages, npages);
> +	else
> +		pr_debug("0x%lx pages migrated\n", migrate.cpages);
> +
> +	if (!migrate.cpages) {
> +		pr_debug("failed collect migrate sys pages [0x%lx 0x%lx]\n",
> +			 prange->start, prange->last);
> +		goto out_free;
>  	}
>  
> +	r = svm_migrate_copy_to_vram(adev, prange, &migrate, &mfence, scratch);
> +	migrate_vma_pages(&migrate);
> +	svm_migrate_copy_done(adev, mfence);
> +	migrate_vma_finalize(&migrate);
> +
>  	svm_range_dma_unmap(adev->dev, scratch, 0, npages);
>  	svm_range_free_dma_mappings(prange);
>  
>  out_free:
>  	kvfree(buf);
>  out:
> -	if (!r) {
> +	if (!r && migrate.cpages) {
>  		pdd = svm_range_get_pdd_by_adev(prange, adev);
>  		if (pdd)
>  			WRITE_ONCE(pdd->page_in, pdd->page_in + migrate.cpages);
> -	}
>  
> +		return migrate.cpages;
> +	}
>  	return r;
>  }
>  
> @@ -460,7 +465,8 @@ svm_migrate_ram_to_vram(struct svm_range *prange, uint32_t best_loc,
>  	unsigned long addr, start, end;
>  	struct vm_area_struct *vma;
>  	struct amdgpu_device *adev;
> -	int r = 0;
> +	unsigned long cpages = 0;
> +	long r;
>  
>  	if (prange->actual_loc == best_loc) {
>  		pr_debug("svms 0x%p [0x%lx 0x%lx] already on best_loc 0x%x\n",
> @@ -492,17 +498,16 @@ svm_migrate_ram_to_vram(struct svm_range *prange, uint32_t best_loc,
>  
>  		next = min(vma->vm_end, end);
>  		r = svm_migrate_vma_to_vram(adev, prange, vma, addr, next);
> -		if (r) {
> -			pr_debug("failed to migrate\n");
> -			break;
> -		}
> +		if (r > 0)
> +			cpages += r;

I think you still want to break out of the loop here if r < 0,
potentially with a debug message.


>  		addr = next;
>  	}
>  
> -	if (!r)
> +	if (cpages) {
>  		prange->actual_loc = best_loc;
> -
> -	return r;
> +		return 0;
> +	}
> +	return -ENOMEM;
>  }
>  
>  static void svm_migrate_page_free(struct page *page)
> @@ -603,7 +608,7 @@ svm_migrate_copy_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
>  	return r;
>  }
>  
> -static int
> +static long
>  svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
>  		       struct vm_area_struct *vma, uint64_t start, uint64_t end)
>  {
> @@ -640,29 +645,35 @@ svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
>  		goto out_free;
>  	}
>  
> -	pr_debug("cpages %ld\n", migrate.cpages);
> +	if (migrate.cpages != npages)
> +		pr_debug("partial migration, 0x%lx/0x%llx pages migrated\n",
> +			 migrate.cpages, npages);
> +	else
> +		pr_debug("0x%lx pages migrated\n", migrate.cpages);
>  
> -	if (migrate.cpages) {
> -		r = svm_migrate_copy_to_ram(adev, prange, &migrate, &mfence,
> -					    scratch, npages);
> -		migrate_vma_pages(&migrate);
> -		svm_migrate_copy_done(adev, mfence);
> -		migrate_vma_finalize(&migrate);
> -	} else {
> +	if (!migrate.cpages) {
>  		pr_debug("failed collect migrate device pages [0x%lx 0x%lx]\n",
>  			 prange->start, prange->last);
> +		goto out_free;
>  	}
>  
> +	r = svm_migrate_copy_to_ram(adev, prange, &migrate, &mfence,
> +				    scratch, npages);
> +	migrate_vma_pages(&migrate);
> +	svm_migrate_copy_done(adev, mfence);
> +	migrate_vma_finalize(&migrate);
>  	svm_range_dma_unmap(adev->dev, scratch, 0, npages);
>  
>  out_free:
>  	kvfree(buf);
>  out:
> -	if (!r) {
> +	if (!r && migrate.cpages) {
>  		pdd = svm_range_get_pdd_by_adev(prange, adev);
>  		if (pdd)
>  			WRITE_ONCE(pdd->page_out,
>  				   pdd->page_out + migrate.cpages);
> +
> +		return migrate.cpages;
>  	}
>  	return r;
>  }
> @@ -684,7 +695,8 @@ int svm_migrate_vram_to_ram(struct svm_range *prange, struct mm_struct *mm)
>  	unsigned long addr;
>  	unsigned long start;
>  	unsigned long end;
> -	int r = 0;
> +	unsigned long cpages = 0;
> +	long r;
>  
>  	if (!prange->actual_loc) {
>  		pr_debug("[0x%lx 0x%lx] already migrated to ram\n",
> @@ -715,18 +727,18 @@ int svm_migrate_vram_to_ram(struct svm_range *prange, struct mm_struct *mm)
>  
>  		next = min(vma->vm_end, end);
>  		r = svm_migrate_vma_to_ram(adev, prange, vma, addr, next);
> -		if (r) {
> -			pr_debug("failed %d to migrate\n", r);
> -			break;
> -		}
> +		if (r > 0)
> +			cpages += r;

Same as above.

Regards,
  Felix

>  		addr = next;
>  	}
>  
> -	if (!r) {
> +	if (cpages) {
>  		svm_range_vram_node_free(prange);
>  		prange->actual_loc = 0;
> +
> +		return 0;
>  	}
> -	return r;
> +	return -ENOMEM;
>  }
>  
>  /**