Re: [PATCH v5 2/3] drm/xe/guc_ads: allocate UM queues in VRAM on dGFX

On 31/07/2026 00:27, Jia Yao wrote:
> On pre-Xe3p dGFX, GuC/GAM accesses the UM queue ring buffer via
> base_dpa, which is interpreted as a raw device-physical address.
> Unlike the GGTT PTE, there is no flag to indicate whether base_dpa
> points to VRAM or system memory; the hardware assumes it is always
> a VRAM address.
> 
> A separate SYSTEM BO for the UM queues would cause xe_bo_main_addr()
> to return an IOMMU IOVA instead of a true VRAM device-physical address.
> GuC then accesses the wrong memory and forwards all-zero fault
> descriptors to the driver (Faulted Address = 0, ASID = 0).
> 
> Use XE_BO_FLAG_VRAM_IF_DGFX() so that on dGFX the UM queue BO is
> placed directly in VRAM and xe_bo_main_addr() returns the correct
> device-physical address.
> 
> Fixes: dd08ebf6c352 ("drm/xe: Introduce a new DRM driver for Intel GPUs")

Fixes: 9c57bc08652a ("drm/xe/lnl: Drop force_probe requirement")
Cc: <[email protected]> # v6.12+

> Cc: Gwan-gyeong Mun <[email protected]>
> Cc: Matthew Auld <[email protected]>
> Cc: [email protected]
> Signed-off-by: Jia Yao <[email protected]>
> ---
>   drivers/gpu/drm/xe/xe_guc.c     |  6 ------
>   drivers/gpu/drm/xe/xe_guc_ads.c | 11 ++++++-----
>   2 files changed, 6 insertions(+), 11 deletions(-)
> 
> diff --git a/drivers/gpu/drm/xe/xe_guc.c b/drivers/gpu/drm/xe/xe_guc.c
> index 4f132bf8539d..4286bd05c686 100644
> --- a/drivers/gpu/drm/xe/xe_guc.c
> +++ b/drivers/gpu/drm/xe/xe_guc.c
> @@ -732,12 +732,6 @@ static int xe_guc_realloc_post_hwconfig(struct xe_guc *guc)
>   	if (ret)
>   		return ret;
>   
> -	if (guc->ads.um_queue_bo) {
> -		ret = xe_managed_bo_reinit_in_vram(xe, tile, &guc->ads.um_queue_bo);
> -		if (ret)
> -			return ret;
> -	}
> -
>   	return 0;
>   }
>   
> diff --git a/drivers/gpu/drm/xe/xe_guc_ads.c b/drivers/gpu/drm/xe/xe_guc_ads.c
> index 3b6e1c45a0e9..57a0614dbf6d 100644
> --- a/drivers/gpu/drm/xe/xe_guc_ads.c
> +++ b/drivers/gpu/drm/xe/xe_guc_ads.c
> @@ -461,11 +461,12 @@ int xe_guc_ads_init(struct xe_guc_ads *ads)
>   		size_t um_size = roundup_pow_of_two(GUC_UM_QUEUE_SIZE *
>   						    GUC_UM_HW_QUEUE_MAX);

If you want we could also make the roundup conditional on igpu, to save 
a small amount of VRAM.

Either way,
Reviewed-by: Matthew Auld <[email protected]>

>   
> -		bo = xe_managed_bo_create_pin_map(xe, tile, um_size,
> -						  XE_BO_FLAG_SYSTEM |
> -						  XE_BO_FLAG_GGTT |
> -						  XE_BO_FLAG_GGTT_INVALIDATE |
> -						  XE_BO_FLAG_PINNED_NORESTORE);
> +		u32 um_flags = XE_BO_FLAG_VRAM_IF_DGFX(tile) |
> +			       XE_BO_FLAG_GGTT |
> +			       XE_BO_FLAG_GGTT_INVALIDATE |
> +			       XE_BO_FLAG_PINNED_NORESTORE;
> +
> +		bo = xe_managed_bo_create_pin_map(xe, tile, um_size, um_flags);
>   		if (IS_ERR(bo))
>   			return PTR_ERR(bo);
>