Re: [PATCH 2/3] drm/amdkfd: Fix the case that vm range is hole at svm_migrate_copy_to_vram

On 2026-08-17 09:59, Xiaogang.Chen wrote:
> From: Xiaogang Chen <[email protected]>
>
> When migration vm range is hole at cpu side(MIGRATE_PFN_MIGRATE set +
> MIGRATE_PFN_VALID unset) driver still allocates device pages. There is no
> dma map of src pages and migration. j is 0 and svm_migrate_copy_memory_gart()
> will return an uninitialized r. That can trigger out_free_vram_pages to drop
> all VRAM just set up.

Good catch. The first fix I would expect here, is that 
svm_migrate_copy_memory_gart should never return an uninitialized r. I'd 
defensively initialize r in svm_migrate_copy_memory_gart as well, even 
if you never plan to call it with npages = 0. With that fixed, the patch is

Reviewed-by: Felix Kuehling <[email protected]>


>
> Initialize r and only call the last svm_migrate_copy_memory_gart if j > 0.
>
> Current code postponed the last page to the final copy. This patch flushes on
> the last page when reach to the end of current drm_buddy_block; avoids another
> svm_migrate_copy_memory_gart.
>
> Signed-off-by: Xiaogang Chen <[email protected]>
> ---
>   drivers/gpu/drm/amd/amdkfd/kfd_migrate.c | 9 ++++++---
>   1 file changed, 6 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c
> index aea82da5f575..263dae49bb3e 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c
> @@ -284,7 +284,7 @@ svm_migrate_copy_to_vram(struct kfd_node *node, struct svm_range *prange,
>   	dma_addr_t *src;
>   	u64 *dst;
>   	u64 i, j;
> -	int r;
> +	int r = 0;
>   
>   	pr_debug("svms 0x%p [0x%lx 0x%lx 0x%llx]\n", prange->svms, prange->start,
>   		 prange->last, ttm_res_offset);
> @@ -310,6 +310,7 @@ svm_migrate_copy_to_vram(struct kfd_node *node, struct svm_range *prange,
>   					      DMA_BIDIRECTIONAL);
>   			r = dma_mapping_error(dev, src[i]);
>   			if (r) {
> +				src[i] = 0;
>   				dev_err(dev, "%s: fail %d dma_map_page\n",
>   					__func__, r);
>   				goto out_free_vram_pages;
> @@ -334,7 +335,8 @@ svm_migrate_copy_to_vram(struct kfd_node *node, struct svm_range *prange,
>   		pr_debug_ratelimited("dma mapping src to 0x%llx, pfn 0x%lx\n",
>   				     src[i] >> PAGE_SHIFT, page_to_pfn(spage));
>   
> -		if (j >= (cursor.size >> PAGE_SHIFT) - 1 && i < npages - 1) {
> +		/* accumulated j + 1 pages reach end of current drm_buddy_block */
> +		if (j + 1 >= (cursor.size >> PAGE_SHIFT)) {
>   			r = svm_migrate_copy_memory_gart(adev, src + i - j,
>   							 dst + i - j, j + 1,
>   							 FROM_RAM_TO_VRAM,
> @@ -348,7 +350,8 @@ svm_migrate_copy_to_vram(struct kfd_node *node, struct svm_range *prange,
>   		}
>   	}
>   
> -	r = svm_migrate_copy_memory_gart(adev, src + i - j, dst + i - j, j,
> +	if (j > 0)
> +		r = svm_migrate_copy_memory_gart(adev, src + i - j, dst + i - j, j,
>   					 FROM_RAM_TO_VRAM, mfence);
>   
>   out_free_vram_pages: