Re: [PATCH v4 2/2] drm/xe: Add prefetch fault support for Xe3p

[Matthew Brost <matthew.brost@intel.com>]

On Fri, Feb 20, 2026 at 10:05:02AM +0530, Varun Gupta wrote:
> Xe3p hardware prefetches memory ranges and notifies software via an
> additional bit (bit 11) in the page fault descriptor that the fault
> was caused by prefetch.
> 
> Extract the prefetch bit from the fault descriptor and echo it in the
> response (bit 6) only when the page fault handling fails. This allows
> the HW to suppress CAT errors for unsuccessful prefetch faults.
> 
> For prefetch faults that fail, increment stats counter without verbose
> logging to avoid spamming the log. Clear the prefetch bit on successful
> handling so it's not echoed in the response.
> 
> Based on original patches by Brian Welty <brian.welty@intel.com> and
> Priyanka Dandamudi <priyanka.dandamudi@intel.com>.
> 
> Bspec: 59311
> Originally-by: Lucas De Marchi <lucas.demarchi@intel.com>
> Cc: Matthew Brost <matthew.brost@intel.com>
> Cc: Priyanka Dandamudi <priyanka.dandamudi@intel.com>
> Cc: Matt Roper <matthew.d.roper@intel.com>
> Signed-off-by: Lucas De Marchi <lucas.demarchi@intel.com>
> Signed-off-by: Varun Gupta <varun.gupta@intel.com>
> 
> ---
> v4:
>  - Downgrade prefetch error log to xe_gt_dbg to avoid spam (Matt Brost/ Stuart)
> 
> v3:
>  - Drop the rename patch, keep xe_pagefault_print() unchanged (Matt Brost)
>  - Move prefetch check to caller instead of inside print function (Matt Brost)
>  - Remove XE3P_ prefix from prefetch bit defines and add platform comment (Matt Brost)
>  - Show prefetch bit in error messages for debugging (Matt Brost)
>  - Split stats counter into separate patch (Matt Brost)
> 
> v2:
>  - Changed comment wording from "repairs" to "handling" for clarity (Matt Roper)
> ---
>  drivers/gpu/drm/xe/xe_guc_fwif.h        |  5 +++--
>  drivers/gpu/drm/xe/xe_guc_pagefault.c   |  2 ++
>  drivers/gpu/drm/xe/xe_pagefault.c       | 18 +++++++++++++++---
>  drivers/gpu/drm/xe/xe_pagefault_types.h |  6 ++++++
>  4 files changed, 26 insertions(+), 5 deletions(-)
> 
> diff --git a/drivers/gpu/drm/xe/xe_guc_fwif.h b/drivers/gpu/drm/xe/xe_guc_fwif.h
> index a33ea288b907..bb8f71d38611 100644
> --- a/drivers/gpu/drm/xe/xe_guc_fwif.h
> +++ b/drivers/gpu/drm/xe/xe_guc_fwif.h
> @@ -261,7 +261,8 @@ struct xe_guc_pagefault_desc {
>  #define PFD_ACCESS_TYPE		GENMASK(1, 0)
>  #define PFD_FAULT_TYPE		GENMASK(3, 2)
>  #define PFD_VFID		GENMASK(9, 4)
> -#define PFD_RSVD_1		GENMASK(11, 10)
> +#define PFD_RSVD_1		BIT(10)
> +#define PFD_PREFETCH		BIT(11) /* Only valid on Xe3+, reserved on prior platforms */
>  #define PFD_VIRTUAL_ADDR_LO	GENMASK(31, 12)
>  #define PFD_VIRTUAL_ADDR_LO_SHIFT 12
>  
> @@ -281,7 +282,7 @@ struct xe_guc_pagefault_reply {
>  
>  	u32 dw1;
>  #define PFR_VFID		GENMASK(5, 0)
> -#define PFR_RSVD_1		BIT(6)
> +#define PFR_PREFETCH		BIT(6)  /* Only valid on Xe3+, reserved on prior platforms */
>  #define PFR_ENG_INSTANCE	GENMASK(12, 7)
>  #define PFR_ENG_CLASS		GENMASK(15, 13)
>  #define PFR_PDATA		GENMASK(31, 16)
> diff --git a/drivers/gpu/drm/xe/xe_guc_pagefault.c b/drivers/gpu/drm/xe/xe_guc_pagefault.c
> index d48f6ed103bb..0cb68e956af9 100644
> --- a/drivers/gpu/drm/xe/xe_guc_pagefault.c
> +++ b/drivers/gpu/drm/xe/xe_guc_pagefault.c
> @@ -28,6 +28,7 @@ static void guc_ack_fault(struct xe_pagefault *pf, int err)
>  		FIELD_PREP(PFR_ASID, asid),
>  
>  		FIELD_PREP(PFR_VFID, vfid) |
> +		FIELD_PREP(PFR_PREFETCH, pf->consumer.prefetch) |

You can set this up based on the producer data and error.

I'd like to avoid touching consumer fields in this path, see [1].

I think the logic would be...

u32 prefetch = FIELD_GET(PFD_PREFETCH, pf->producer.msg[2]);
....

FIELD_PREP(PFR_PREFETCH, !err ? 0 : prefetch);

This is a better logical divide as the producer is the one who knows if
it should set the PFR_PREFETCH in the acknowledgement - not the
consumer.

[1] https://patchwork.freedesktop.org/patch/704579/?series=161560&rev=1

>  		FIELD_PREP(PFR_ENG_INSTANCE, engine_instance) |
>  		FIELD_PREP(PFR_ENG_CLASS, engine_class) |
>  		FIELD_PREP(PFR_PDATA, pdata),
> @@ -77,6 +78,7 @@ int xe_guc_pagefault_handler(struct xe_guc *guc, u32 *msg, u32 len)
>  		 PFD_VIRTUAL_ADDR_LO_SHIFT);
>  	pf.consumer.asid = FIELD_GET(PFD_ASID, msg[1]);
>  	pf.consumer.access_type = FIELD_GET(PFD_ACCESS_TYPE, msg[2]);
> +	pf.consumer.prefetch = FIELD_GET(PFD_PREFETCH, msg[2]);
>  	if (FIELD_GET(XE2_PFD_TRVA_FAULT, msg[0]))
>  		pf.consumer.fault_type_level = XE_PAGEFAULT_TYPE_LEVEL_NACK;
>  	else
> diff --git a/drivers/gpu/drm/xe/xe_pagefault.c b/drivers/gpu/drm/xe/xe_pagefault.c
> index 72f589fd2b64..b05860b3607d 100644
> --- a/drivers/gpu/drm/xe/xe_pagefault.c
> +++ b/drivers/gpu/drm/xe/xe_pagefault.c
> @@ -261,9 +261,21 @@ static void xe_pagefault_queue_work(struct work_struct *w)
>  
>  		err = xe_pagefault_service(&pf);
>  		if (err) {
> -			xe_pagefault_print(&pf);
> -			xe_gt_info(pf.gt, "Fault response: Unsuccessful %pe\n",
> -				   ERR_PTR(err));
> +			if (!pf.consumer.prefetch) {
> +				xe_pagefault_print(&pf);
> +				xe_gt_info(pf.gt, "Fault response: Unsuccessful %pe\n",
> +					   ERR_PTR(err));
> +			} else {
> +				xe_gt_stats_incr(pf.gt, XE_GT_STATS_ID_INVALID_PREFETCH_PAGEFAULT_COUNT, 1);
> +				xe_gt_dbg(pf.gt, "Fault response: Unsuccessful %pe, prefetch=%d\n",
> +					  ERR_PTR(err), pf.consumer.prefetch);
> +			}
> +		} else {
> +			/*
> +			 * Clear prefetch bit - only needed to suppress CAT errors
> +			 * on unsuccessful handling.
> +			 */
> +			pf.consumer.prefetch = 0;

With above you can drop this logic.

>  		}
>  
>  		pf.producer.ops->ack_fault(&pf, err);
> diff --git a/drivers/gpu/drm/xe/xe_pagefault_types.h b/drivers/gpu/drm/xe/xe_pagefault_types.h
> index 0e378f41ede6..23a6a986d5b2 100644
> --- a/drivers/gpu/drm/xe/xe_pagefault_types.h
> +++ b/drivers/gpu/drm/xe/xe_pagefault_types.h
> @@ -84,6 +84,12 @@ struct xe_pagefault {
>  		u8 engine_class;
>  		/** @consumer.engine_instance: engine instance */
>  		u8 engine_instance;
> +		/**
> +		 * @consumer.prefetch: fault is caused by HW prefetch.
> +		 * Echo in response to suppress CAT errors on
> +		 * unsuccessful handling.
> +		 */
> +		u8 prefetch;

This will grow this structure from 64 bytes in the page fault queue to
128 which is not desirable.

>  		/** consumer.reserved: reserved bits for future expansion */
>  		u64 reserved;

I also we like to leave this u64 reserved as I plan on using it as
single linked list of pagefaults that come in a storm that all match the
same address [2]. This is reason I'd packed fault type and level into a
single u8 here [3]. Maybe pack prefetch into the upper bits of the
access_type field?

Matt

[2] https://gitlab.freedesktop.org/mbrost/xe-kernel-driver-svn-perf-6-15-2025/-/commit/93669c7f4e00ec13d0a18e28d34dfcb41803b7c9#d0e961d15e218e9be8819f1311674d537b25ebdd_91_109 
[3] https://patchwork.freedesktop.org/patch/704578/?series=161560&rev=1

>  	} consumer;
> -- 
> 2.43.0
>