Skip to contentSkip to conversation

FAILED: patch "[PATCH] drm/xe/guc_ads: allocate UM queues in a separate BO" failed to apply to 6.18-stable tree

PATCH New [email protected] 3 participants 19 messages Aug 20 – Aug 27

Subsystems (3)
lore
Share & tools
lore

[email protected]

author maintainer
#0 lore raw

The patch below does not apply to the 6.18-stable tree.
If someone wants it applied there, or to any other stable or longterm tree, then please email the backport, including the original git commit id to <[email protected]>.

To reproduce the conflict and resubmit, you may use the following commands:

git fetch https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/ linux-6.18.y
git checkout FETCH_HEAD
git cherry-pick -x 99b01815957bd7d848420cb697f79ed11f7f215c
# <resolve conflicts, build, test, etc.>
git commit -s
git send-email --to '<[email protected]>' --in-reply-to '2026082036-afterglow-tapered-84a0@gregkh' --subject-prefix 'PATCH 6.18.y' 'HEAD^..'

Possible dependencies:

thanks,

greg k-h

------------------ original commit in Linus's tree ------------------

From 99b01815957bd7d848420cb697f79ed11f7f215c Mon Sep 17 00:00:00 2001
From: Jia Yao <[email protected]>
Date: Tue, 4 Aug 2026 16:50:55 +0000
Subject: [PATCH] drm/xe/guc_ads: allocate UM queues in a separate BO
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Move the UM queues into a dedicated BO (ads->um_queue_bo) and avoid CPU memset operations on it, which eliminates the CPU as a potential cacheline-polluting agent and helps maintain consistency between GAM writes and GuC reads.

We also need to ensure the base_dpa for the queue is contiguous on hw where this is used instead of a GGTT address. Another good reason to split this out to a separate BO.

Fixes 9c57bc08652a ("drm/xe/lnl: Drop force_probe requirement")
Cc Gwan-gyeong Mun <[email protected]>
Cc Matthew Auld <[email protected]>
Cc <[email protected]> # v6.12+
(cherry picked from commit 6af05de0dc19bbf3aaeec2020fe48b37c834b811)
drivers/gpu/drm/xe/xe_bo.h +14 -0
index 7ae1d9ac0574..57039cf42ea7 100644
@@ -6,6 +6,7 @@
 #ifndef _XE_BO_H_
 #define _XE_BO_H_
 
+#include <drm/drm_prime.h>
 #include <drm/ttm/ttm_tt.h>
 
 #include "xe_bo_types.h"
@@ -548,6 +549,19 @@ void xe_bo_dev_fini(struct xe_bo_dev *bo_device);
 
 struct sg_table *xe_bo_sg(struct xe_bo *bo);
 
+/**
+ * xe_bo_sg_is_contiguous() - Check if a BO's DMA address space is contiguous.
+ * @bo: the BO to check (must have a valid sg table, i.e. !xe_bo_is_vram())
+ * @len: required contiguous length in bytes
+ *
+ * Returns true if the first @len bytes of the BO are mapped to a contiguous
+ * DMA address range.
+ */
+static inline bool xe_bo_sg_is_contiguous(struct xe_bo *bo, size_t len)
+{
+	return drm_prime_get_contiguous_size(xe_bo_sg(bo)) >= len;
+}
+
 /*
  * xe_sg_segment_size() - Provides upper limit for sg segment size.
  * @dev: device pointer
drivers/gpu/drm/xe/xe_guc.c +6 -0
index 4023700ff2a9..21cf3edef922 100644
@@ -714,6 +714,12 @@ 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;
 }
 
drivers/gpu/drm/xe/xe_guc_ads.c +58 -32

137 lines not on this page · raw message

drivers/gpu/drm/xe/xe_guc_ads_types.h +5 -0
index 48a8e092023f..845c1fbd93a4 100644
@@ -16,6 +16,11 @@ struct xe_bo;
 struct xe_guc_ads {
 	/** @bo: Xe BO for GuC ads blob */
 	struct xe_bo *bo;
+	/**
+	 * @um_queue_bo: Dedicated BO for the HW fault ring (UM queues).
+	 * NULL if the platform does not support USM.
+	 */
+	struct xe_bo *um_queue_bo;
 	/** @golden_lrc_size: golden LRC size */
 	size_t golden_lrc_size;
 	/** @regset_size: size of register set passed to GuC for save/restore */
#1 lore raw

PATCH 1/1 drm/xe/guc_ads: allocate UM queues in a separate BO

Move the UM queues into a dedicated BO (ads->um_queue_bo) and avoid CPU memset operations on it, which eliminates the CPU as a potential cacheline-polluting agent and helps maintain consistency between GAM writes and GuC reads.

We also need to ensure the base_dpa for the queue is contiguous on hw where this is used instead of a GGTT address. Another good reason to split this out to a separate BO.

Fixes 9c57bc08652a ("drm/xe/lnl: Drop force_probe requirement")
Cc Gwan-gyeong Mun <[email protected]>
Cc Matthew Auld <[email protected]>
Cc <[email protected]> # v6.12+
(cherry picked from commit 6af05de0dc19bbf3aaeec2020fe48b37c834b811)
(cherry picked from commit 99b01815957bd7d848420cb697f79ed11f7f215c)
---
 drivers/gpu/drm/xe/xe_bo.h            | 14 +++++
 drivers/gpu/drm/xe/xe_guc.c           |  6 ++
 drivers/gpu/drm/xe/xe_guc_ads.c       | 90 +++++++++++++++++----------
 drivers/gpu/drm/xe/xe_guc_ads_types.h |  5 ++
 4 files changed, 83 insertions(+), 32 deletions(-)
drivers/gpu/drm/xe/xe_bo.h +14 -0
index fef60c8a4e5b..71169b57fbda 100644
@@ -6,6 +6,7 @@
 #ifndef _XE_BO_H_
 #define _XE_BO_H_
 
+#include <drm/drm_prime.h>
 #include <drm/ttm/ttm_tt.h>
 
 #include "xe_bo_types.h"
@@ -455,6 +456,19 @@ void xe_bo_dev_fini(struct xe_bo_dev *bo_device);
 
 struct sg_table *xe_bo_sg(struct xe_bo *bo);
 
+/**
+ * xe_bo_sg_is_contiguous() - Check if a BO's DMA address space is contiguous.
+ * @bo: the BO to check (must have a valid sg table, i.e. !xe_bo_is_vram())
+ * @len: required contiguous length in bytes
+ *
+ * Returns true if the first @len bytes of the BO are mapped to a contiguous
+ * DMA address range.
+ */
+static inline bool xe_bo_sg_is_contiguous(struct xe_bo *bo, size_t len)
+{
+	return drm_prime_get_contiguous_size(xe_bo_sg(bo)) >= len;
+}
+
 /*
  * xe_sg_segment_size() - Provides upper limit for sg segment size.
  * @dev: device pointer
drivers/gpu/drm/xe/xe_guc.c +6 -0
index ae0c88da422b..cc5741f751b9 100644
@@ -700,6 +700,12 @@ 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;
 }
 
drivers/gpu/drm/xe/xe_guc_ads.c +58 -32

135 lines not on this page · raw message

drivers/gpu/drm/xe/xe_guc_ads_types.h +5 -0
index 70c132458ac3..acb5366f76ef 100644
@@ -16,6 +16,11 @@ struct xe_bo;
 struct xe_guc_ads {
 	/** @bo: XE BO for GuC ads blob */
 	struct xe_bo *bo;
+	/**
+	 * @um_queue_bo: Dedicated BO for the HW fault ring (UM queues).
+	 * NULL if the platform does not support USM.
+	 */
+	struct xe_bo *um_queue_bo;
 	/** @golden_lrc_size: golden LRC size */
 	size_t golden_lrc_size;
 	/** @regset_size: size of register set passed to GuC for save/restore */
Signature · 1 line
2.43.0
#2 lore raw

PATCH 1/3 drm/xe/guc_ads: allocate UM queues in a separate BO

In reply to [email protected]

Move the UM queues into a dedicated BO (ads->um_queue_bo) and avoid CPU memset operations on it, which eliminates the CPU as a potential cacheline-polluting agent and helps maintain consistency between GAM writes and GuC reads.

We also need to ensure the base_dpa for the queue is contiguous on hw where this is used instead of a GGTT address. Another good reason to split this out to a separate BO.

Fixes 9c57bc08652a ("drm/xe/lnl: Drop force_probe requirement")
Cc Gwan-gyeong Mun <[email protected]>
Cc Matthew Auld <[email protected]>
Cc <[email protected]> # v6.12+
(cherry picked from commit 6af05de0dc19bbf3aaeec2020fe48b37c834b811)
(cherry picked from commit 99b01815957bd7d848420cb697f79ed11f7f215c)
---
 drivers/gpu/drm/xe/xe_bo.h            | 14 +++++
 drivers/gpu/drm/xe/xe_guc.c           |  6 ++
 drivers/gpu/drm/xe/xe_guc_ads.c       | 90 +++++++++++++++++----------
 drivers/gpu/drm/xe/xe_guc_ads_types.h |  5 ++
 4 files changed, 83 insertions(+), 32 deletions(-)
drivers/gpu/drm/xe/xe_bo.h +14 -0
index fef60c8a4e5b..71169b57fbda 100644
@@ -6,6 +6,7 @@
 #ifndef _XE_BO_H_
 #define _XE_BO_H_
 
+#include <drm/drm_prime.h>
 #include <drm/ttm/ttm_tt.h>
 
 #include "xe_bo_types.h"
@@ -455,6 +456,19 @@ void xe_bo_dev_fini(struct xe_bo_dev *bo_device);
 
 struct sg_table *xe_bo_sg(struct xe_bo *bo);
 
+/**
+ * xe_bo_sg_is_contiguous() - Check if a BO's DMA address space is contiguous.
+ * @bo: the BO to check (must have a valid sg table, i.e. !xe_bo_is_vram())
+ * @len: required contiguous length in bytes
+ *
+ * Returns true if the first @len bytes of the BO are mapped to a contiguous
+ * DMA address range.
+ */
+static inline bool xe_bo_sg_is_contiguous(struct xe_bo *bo, size_t len)
+{
+	return drm_prime_get_contiguous_size(xe_bo_sg(bo)) >= len;
+}
+
 /*
  * xe_sg_segment_size() - Provides upper limit for sg segment size.
  * @dev: device pointer
drivers/gpu/drm/xe/xe_guc.c +6 -0
index ae0c88da422b..cc5741f751b9 100644
@@ -700,6 +700,12 @@ 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;
 }
 
drivers/gpu/drm/xe/xe_guc_ads.c +58 -32

135 lines not on this page · raw message

drivers/gpu/drm/xe/xe_guc_ads_types.h +5 -0
index 70c132458ac3..acb5366f76ef 100644
@@ -16,6 +16,11 @@ struct xe_bo;
 struct xe_guc_ads {
 	/** @bo: XE BO for GuC ads blob */
 	struct xe_bo *bo;
+	/**
+	 * @um_queue_bo: Dedicated BO for the HW fault ring (UM queues).
+	 * NULL if the platform does not support USM.
+	 */
+	struct xe_bo *um_queue_bo;
 	/** @golden_lrc_size: golden LRC size */
 	size_t golden_lrc_size;
 	/** @regset_size: size of register set passed to GuC for save/restore */
Signature · 1 line
2.43.0
#3 lore raw

PATCH 2/3 drm/xe/guc_ads: allocate UM queues in VRAM on dGFX

On iGPU, the UM queue BO is allocated in system memory. On dGFX, the BO was previously created in system memory and later reallocated in xe_guc_realloc_post_hwconfig(). Allocate the UM queue BO directly in VRAM on dGFX, where it is ultimately required.

Fixes 9c57bc08652a ("drm/xe/lnl: Drop force_probe requirement")
Cc Gwan-gyeong Mun <[email protected]>
Cc Matthew Auld <[email protected]>
Cc <[email protected]> # v6.12+
(cherry picked from commit ace076ef0a854ab5940bacc539bf66afd61d118c)
(cherry picked from commit f342810a141f8a7e8b3786a6e4b6c0695a078a74)
---
 drivers/gpu/drm/xe/xe_guc.c     | 6 ------
 drivers/gpu/drm/xe/xe_guc_ads.c | 8 +++++---
 2 files changed, 5 insertions(+), 9 deletions(-)
drivers/gpu/drm/xe/xe_guc.c +0 -6
index cc5741f751b9..ae0c88da422b 100644
@@ -700,12 +700,6 @@ 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;
 }
 
drivers/gpu/drm/xe/xe_guc_ads.c +5 -3
index 806c2618a7cd..4c3e21b5e60e 100644
@@ -392,15 +392,17 @@ int xe_guc_ads_init(struct xe_guc_ads *ads)
 		/*
 		 * Allocate a separate BO for the HW fault ring (UM queues).
 		 *
-		 * Round the size up to the next power of two so that
+		 * Round the size up to the next power of two so that on iGPU
 		 * (system memory, no IOMMU) the TTM pool issues a single
 		 * alloc_pages(order=N) call, maximising the chance of getting
 		 * a physically contiguous block.  GuC requires contiguous DPA.
 		 */
-		size_t um_size = roundup_pow_of_two(GUC_UM_QUEUE_SIZE *
+		size_t um_size = IS_DGFX(xe) ?
+				 GUC_UM_QUEUE_SIZE * GUC_UM_HW_QUEUE_MAX :
+				 roundup_pow_of_two(GUC_UM_QUEUE_SIZE *
 						    GUC_UM_HW_QUEUE_MAX);
 
-		u32 um_flags = XE_BO_FLAG_SYSTEM |
+		u32 um_flags = XE_BO_FLAG_VRAM_IF_DGFX(tile) |
 			       XE_BO_FLAG_GGTT |
 			       XE_BO_FLAG_GGTT_INVALIDATE |
 			       XE_BO_FLAG_PINNED_NORESTORE;
Signature · 1 line
2.43.0
#4 lore raw

PATCH 3/3 drm/xe/guc_ads: use uncached mapping for UM queue BO

In reply to Jia Yao

On Pre-Xe3p platform, the GAM write the UM queue through DPA using UC. if GuC reads the queue via GGTT (WB), stale data may be observed when the cacheline has been polluted by another agent.

To match the GAM's UC writes, configure the GuC mapping as UC as well.

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

Reviewed by: Matthew Auld <[email protected]>

(cherry picked from commit 9daa302a82590eeee7bdc68023ddad302df4b88c)
(cherry picked from commit 8d5134ae4177fa4f5a9bc8e71e6656cfc2852882)
---
 drivers/gpu/drm/xe/xe_guc_ads.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
drivers/gpu/drm/xe/xe_guc_ads.c +2 -1
index 4c3e21b5e60e..ecd07d76ec4a 100644
@@ -405,7 +405,8 @@ int xe_guc_ads_init(struct xe_guc_ads *ads)
 		u32 um_flags = XE_BO_FLAG_VRAM_IF_DGFX(tile) |
 			       XE_BO_FLAG_GGTT |
 			       XE_BO_FLAG_GGTT_INVALIDATE |
-			       XE_BO_FLAG_PINNED_NORESTORE;
+			       XE_BO_FLAG_PINNED_NORESTORE |
+			       XE_BO_FLAG_NEEDS_UC;
 
 		bo = xe_managed_bo_create_pin_map(xe, tile, um_size, um_flags);
 		if (IS_ERR(bo))
Signature · 1 line
2.43.0
author maintainer
#5 lore raw

Re: [PATCH 1/3] drm/xe/guc_ads: allocate UM queues in a separate BO

In reply to Jia Yao
Quoted 20 lines · Jia Yao

Move the UM queues into a dedicated BO (ads->um_queue_bo) and avoid CPU memset operations on it, which eliminates the CPU as a potential cacheline-polluting agent and helps maintain consistency between GAM writes and GuC reads.

We also need to ensure the base_dpa for the queue is contiguous on hw where this is used instead of a GGTT address. Another good reason to split this out to a separate BO.

Fixes 9c57bc08652a ("drm/xe/lnl: Drop force_probe requirement")
Cc Gwan-gyeong Mun <[email protected]>
Cc Matthew Auld <[email protected]>
Cc <[email protected]> # v6.12+
(cherry picked from commit 6af05de0dc19bbf3aaeec2020fe48b37c834b811)
(cherry picked from commit 99b01815957bd7d848420cb697f79ed11f7f215c)
Quoted 6 lines
---
 drivers/gpu/drm/xe/xe_bo.h            | 14 +++++
 drivers/gpu/drm/xe/xe_guc.c           |  6 ++
 drivers/gpu/drm/xe/xe_guc_ads.c       | 90 +++++++++++++++++----------
 drivers/gpu/drm/xe/xe_guc_ads_types.h |  5 ++
 4 files changed, 83 insertions(+), 32 deletions(-)

What stable tree(s) are these patches for?

And what's with the crazy 2 cherry picked lines? You all know better than to attempt to mess this up even more...

#6 lore raw

RE: [PATCH 1/3] drm/xe/guc_ads: allocate UM queues in a separate BO

This fix is intended for all stable trees containing commit 9c57bc08652a ("drm/xe/lnl: Drop force_probe requirement").

Regarding the duplicate cherry-pick annotations, 99b01815957b was itself cherry-picked from 6af05de0dc19 and the extra annotation was carried over inadvertently. One of the cherry-picked lines can be dropped.

Quoted 45 lines

-----Original Message-----
From: Greg KH <[email protected]>
Sent: Monday, August 24, 2026 10:03 AM
To: Yao, Jia <[email protected]>
Cc: [email protected]; Mun, Gwan-gyeong <gwan- [email protected]>; Auld, Matthew <[email protected]>; Thomas Hellström <[email protected]>
Subject: Re: [PATCH 1/3] drm/xe/guc_ads: allocate UM queues in a separate BO

Quoted 16 lines · Jia Yao

Move the UM queues into a dedicated BO (ads->um_queue_bo) and avoid CPU memset operations on it, which eliminates the CPU as a potential cacheline-polluting agent and helps maintain consistency between GAM writes and GuC reads.

We also need to ensure the base_dpa for the queue is contiguous on hw where this is used instead of a GGTT address. Another good reason to split this out to a separate BO.

Fixes 9c57bc08652a ("drm/xe/lnl: Drop force_probe requirement")
Cc Gwan-gyeong Mun <[email protected]>
Cc Matthew Auld <[email protected]>
Cc <[email protected]> # v6.12+
Quoted 3 lines

Link:
https://patch.msgid.link/[email protected]
(cherry picked from commit

6af05de0dc19bbf3aaeec2020fe48b37c834b811)

Quoted 2 lines
(cherry picked from commit

99b01815957bd7d848420cb697f79ed11f7f215c)

Quoted 6 lines
---
 drivers/gpu/drm/xe/xe_bo.h            | 14 +++++
 drivers/gpu/drm/xe/xe_guc.c           |  6 ++
 drivers/gpu/drm/xe/xe_guc_ads.c       | 90 +++++++++++++++++----------
 drivers/gpu/drm/xe/xe_guc_ads_types.h |  5 ++
 4 files changed, 83 insertions(+), 32 deletions(-)

What stable tree(s) are these patches for?

And what's with the crazy 2 cherry picked lines? You all know better than to attempt to mess this up even more...

#7 lore raw

RE: [PATCH 1/3] drm/xe/guc_ads: allocate UM queues in a separate BO

Resent the patches after removing the redundant cherry-pick information and adding the correct subject.

Quoted 63 lines

-----Original Message-----
From: Yao, Jia
Sent: Monday, August 24, 2026 10:13 AM
To: 'Greg KH' <[email protected]>
Cc: [email protected]; Mun, Gwan-gyeong <Gwan- [email protected]>; Auld, Matthew <[email protected]>; Thomas Hellström <[email protected]>
Subject: RE: [PATCH 1/3] drm/xe/guc_ads: allocate UM queues in a separate BO

This fix is intended for all stable trees containing commit 9c57bc08652a ("drm/xe/lnl: Drop force_probe requirement").

Regarding the duplicate cherry-pick annotations, 99b01815957b was itself cherry-picked from 6af05de0dc19 and the extra annotation was carried over inadvertently. One of the cherry-picked lines can be dropped.

Quoted 6 lines

-----Original Message-----
From: Greg KH <[email protected]>
Sent: Monday, August 24, 2026 10:03 AM
To: Yao, Jia <[email protected]>
Cc: [email protected]; Mun, Gwan-gyeong <gwan- [email protected]>; Auld, Matthew <[email protected]>;

Thomas

Quoted 39 lines

Hellström <[email protected]>
Subject: Re: [PATCH 1/3] drm/xe/guc_ads: allocate UM queues in a separate BO

Quoted 16 lines · Jia Yao

Move the UM queues into a dedicated BO (ads->um_queue_bo) and avoid CPU memset operations on it, which eliminates the CPU as a potential cacheline-polluting agent and helps maintain consistency between GAM writes and GuC reads.

We also need to ensure the base_dpa for the queue is contiguous on hw where this is used instead of a GGTT address. Another good reason to split this out to a separate BO.

Fixes 9c57bc08652a ("drm/xe/lnl: Drop force_probe requirement")
Cc Gwan-gyeong Mun <[email protected]>
Cc Matthew Auld <[email protected]>
Cc <[email protected]> # v6.12+
Quoted 3 lines

Link:
https://patch.msgid.link/[email protected]
(cherry picked from commit

6af05de0dc19bbf3aaeec2020fe48b37c834b811)

Quoted 2 lines
(cherry picked from commit

99b01815957bd7d848420cb697f79ed11f7f215c)

Quoted 6 lines
---
 drivers/gpu/drm/xe/xe_bo.h            | 14 +++++
 drivers/gpu/drm/xe/xe_guc.c           |  6 ++
 drivers/gpu/drm/xe/xe_guc_ads.c       | 90 +++++++++++++++++----------
 drivers/gpu/drm/xe/xe_guc_ads_types.h |  5 ++
 4 files changed, 83 insertions(+), 32 deletions(-)

What stable tree(s) are these patches for?

And what's with the crazy 2 cherry picked lines? You all know better than to attempt to mess this up even more...

#8 lore raw

PATCH 1/3 drm/xe/guc_ads: allocate UM queues in a separate BO

In reply to [email protected]

Move the UM queues into a dedicated BO (ads->um_queue_bo) and avoid CPU memset operations on it, which eliminates the CPU as a potential cacheline-polluting agent and helps maintain consistency between GAM writes and GuC reads.

We also need to ensure the base_dpa for the queue is contiguous on hw where this is used instead of a GGTT address. Another good reason to split this out to a separate BO.

Fixes 9c57bc08652a ("drm/xe/lnl: Drop force_probe requirement")
Cc Gwan-gyeong Mun <[email protected]>
Cc Matthew Auld <[email protected]>
Cc <[email protected]> # v6.12+
(cherry picked from commit 6af05de0dc19bbf3aaeec2020fe48b37c834b811)
---
 drivers/gpu/drm/xe/xe_bo.h            | 14 +++++
 drivers/gpu/drm/xe/xe_guc.c           |  6 ++
 drivers/gpu/drm/xe/xe_guc_ads.c       | 90 +++++++++++++++++----------
 drivers/gpu/drm/xe/xe_guc_ads_types.h |  5 ++
 4 files changed, 83 insertions(+), 32 deletions(-)
drivers/gpu/drm/xe/xe_bo.h +14 -0
index fef60c8a4e5b..71169b57fbda 100644
@@ -6,6 +6,7 @@
 #ifndef _XE_BO_H_
 #define _XE_BO_H_
 
+#include <drm/drm_prime.h>
 #include <drm/ttm/ttm_tt.h>
 
 #include "xe_bo_types.h"
@@ -455,6 +456,19 @@ void xe_bo_dev_fini(struct xe_bo_dev *bo_device);
 
 struct sg_table *xe_bo_sg(struct xe_bo *bo);
 
+/**
+ * xe_bo_sg_is_contiguous() - Check if a BO's DMA address space is contiguous.
+ * @bo: the BO to check (must have a valid sg table, i.e. !xe_bo_is_vram())
+ * @len: required contiguous length in bytes
+ *
+ * Returns true if the first @len bytes of the BO are mapped to a contiguous
+ * DMA address range.
+ */
+static inline bool xe_bo_sg_is_contiguous(struct xe_bo *bo, size_t len)
+{
+	return drm_prime_get_contiguous_size(xe_bo_sg(bo)) >= len;
+}
+
 /*
  * xe_sg_segment_size() - Provides upper limit for sg segment size.
  * @dev: device pointer
drivers/gpu/drm/xe/xe_guc.c +6 -0
index ae0c88da422b..cc5741f751b9 100644
@@ -700,6 +700,12 @@ 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;
 }
 
drivers/gpu/drm/xe/xe_guc_ads.c +58 -32

135 lines not on this page · raw message

drivers/gpu/drm/xe/xe_guc_ads_types.h +5 -0
index 70c132458ac3..acb5366f76ef 100644
@@ -16,6 +16,11 @@ struct xe_bo;
 struct xe_guc_ads {
 	/** @bo: XE BO for GuC ads blob */
 	struct xe_bo *bo;
+	/**
+	 * @um_queue_bo: Dedicated BO for the HW fault ring (UM queues).
+	 * NULL if the platform does not support USM.
+	 */
+	struct xe_bo *um_queue_bo;
 	/** @golden_lrc_size: golden LRC size */
 	size_t golden_lrc_size;
 	/** @regset_size: size of register set passed to GuC for save/restore */
Signature · 1 line
2.43.0
#9 lore raw

PATCH 2/3 drm/xe/guc_ads: allocate UM queues in VRAM on dGFX

On iGPU, the UM queue BO is allocated in system memory. On dGFX, the BO was previously created in system memory and later reallocated in xe_guc_realloc_post_hwconfig(). Allocate the UM queue BO directly in VRAM on dGFX, where it is ultimately required.

Fixes 9c57bc08652a ("drm/xe/lnl: Drop force_probe requirement")
Cc Gwan-gyeong Mun <[email protected]>
Cc Matthew Auld <[email protected]>
Cc <[email protected]> # v6.12+
(cherry picked from commit ace076ef0a854ab5940bacc539bf66afd61d118c)
---
 drivers/gpu/drm/xe/xe_guc.c     | 6 ------
 drivers/gpu/drm/xe/xe_guc_ads.c | 8 +++++---
 2 files changed, 5 insertions(+), 9 deletions(-)
drivers/gpu/drm/xe/xe_guc.c +0 -6
index cc5741f751b9..ae0c88da422b 100644
@@ -700,12 +700,6 @@ 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;
 }
 
drivers/gpu/drm/xe/xe_guc_ads.c +5 -3
index 806c2618a7cd..4c3e21b5e60e 100644
@@ -392,15 +392,17 @@ int xe_guc_ads_init(struct xe_guc_ads *ads)
 		/*
 		 * Allocate a separate BO for the HW fault ring (UM queues).
 		 *
-		 * Round the size up to the next power of two so that
+		 * Round the size up to the next power of two so that on iGPU
 		 * (system memory, no IOMMU) the TTM pool issues a single
 		 * alloc_pages(order=N) call, maximising the chance of getting
 		 * a physically contiguous block.  GuC requires contiguous DPA.
 		 */
-		size_t um_size = roundup_pow_of_two(GUC_UM_QUEUE_SIZE *
+		size_t um_size = IS_DGFX(xe) ?
+				 GUC_UM_QUEUE_SIZE * GUC_UM_HW_QUEUE_MAX :
+				 roundup_pow_of_two(GUC_UM_QUEUE_SIZE *
 						    GUC_UM_HW_QUEUE_MAX);
 
-		u32 um_flags = XE_BO_FLAG_SYSTEM |
+		u32 um_flags = XE_BO_FLAG_VRAM_IF_DGFX(tile) |
 			       XE_BO_FLAG_GGTT |
 			       XE_BO_FLAG_GGTT_INVALIDATE |
 			       XE_BO_FLAG_PINNED_NORESTORE;
Signature · 1 line
2.43.0
author maintainer
#10 lore raw

Re: [PATCH 6.18.y 2/3] drm/xe/guc_ads: allocate UM queues in VRAM on dGFX

Quoted 5 lines · Jia Yao

On iGPU, the UM queue BO is allocated in system memory. On dGFX, the BO was previously created in system memory and later reallocated in xe_guc_realloc_post_hwconfig(). Allocate the UM queue BO directly in VRAM on dGFX, where it is ultimately required.

Quoted 9 lines
Fixes 9c57bc08652a ("drm/xe/lnl: Drop force_probe requirement")
Cc Gwan-gyeong Mun <[email protected]>
Cc Matthew Auld <[email protected]>
Cc <[email protected]> # v6.12+
(cherry picked from commit ace076ef0a854ab5940bacc539bf66afd61d118c)

Same here, this commit id is not in Linus's, or any stable kernel trees. Where is it coming from?

thanks,

greg k-h

#11 lore raw

PATCH 3/3 drm/xe/guc_ads: use uncached mapping for UM queue BO

In reply to Jia Yao

On Pre-Xe3p platform, the GAM write the UM queue through DPA using UC. if GuC reads the queue via GGTT (WB), stale data may be observed when the cacheline has been polluted by another agent.

To match the GAM's UC writes, configure the GuC mapping as UC as well.

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

Reviewed by: Matthew Auld <[email protected]>

(cherry picked from commit 9daa302a82590eeee7bdc68023ddad302df4b88c)
---
 drivers/gpu/drm/xe/xe_guc_ads.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
drivers/gpu/drm/xe/xe_guc_ads.c +2 -1
index 4c3e21b5e60e..ecd07d76ec4a 100644
@@ -405,7 +405,8 @@ int xe_guc_ads_init(struct xe_guc_ads *ads)
 		u32 um_flags = XE_BO_FLAG_VRAM_IF_DGFX(tile) |
 			       XE_BO_FLAG_GGTT |
 			       XE_BO_FLAG_GGTT_INVALIDATE |
-			       XE_BO_FLAG_PINNED_NORESTORE;
+			       XE_BO_FLAG_PINNED_NORESTORE |
+			       XE_BO_FLAG_NEEDS_UC;
 
 		bo = xe_managed_bo_create_pin_map(xe, tile, um_size, um_flags);
 		if (IS_ERR(bo))
Signature · 1 line
2.43.0
author maintainer
#12 lore raw

Re: [PATCH 6.18.y 3/3] drm/xe/guc_ads: use uncached mapping for UM queue BO

Quoted 11 lines · Jia Yao

On Pre-Xe3p platform, the GAM write the UM queue through DPA using UC. if GuC reads the queue via GGTT (WB), stale data may be observed when the cacheline has been polluted by another agent.

To match the GAM's UC writes, configure the GuC mapping as UC as well.

Fixes 9c57bc08652a ("drm/xe/lnl: Drop force_probe requirement")
Cc Gwan-gyeong Mun <[email protected]>
Cc Matthew Auld <[email protected]>
Cc <[email protected]> # v6.12+
Quoted 4 lines

Reviewed by: Matthew Auld <[email protected]>

(cherry picked from commit 9daa302a82590eeee7bdc68023ddad302df4b88c)

Same here, invalid git commit id.

You all really screw yourself with the bizarre "cherry-pick backwards in time" model :(

Please reconsider it as it seems to cause nothing but trouble.

greg k-h

#13 lore raw

RE: [PATCH 6.18.y 3/3] drm/xe/guc_ads: use uncached mapping for UM queue BO

Sorry for that, will fix it and send another one.

Quoted 37 lines

-----Original Message-----
From: Greg KH <[email protected]>
Sent: Tuesday, August 25, 2026 1:29 AM
To: Yao, Jia <[email protected]>
Cc: [email protected]; Mun, Gwan-gyeong <gwan- [email protected]>; Auld, Matthew <[email protected]>; Thomas Hellström <[email protected]>
Subject: Re: [PATCH 6.18.y 3/3] drm/xe/guc_ads: use uncached mapping for UM queue BO

Quoted 11 lines · Jia Yao

On Pre-Xe3p platform, the GAM write the UM queue through DPA using UC. if GuC reads the queue via GGTT (WB), stale data may be observed when the cacheline has been polluted by another agent.

To match the GAM's UC writes, configure the GuC mapping as UC as well.

Fixes 9c57bc08652a ("drm/xe/lnl: Drop force_probe requirement")
Cc Gwan-gyeong Mun <[email protected]>
Cc Matthew Auld <[email protected]>
Cc <[email protected]> # v6.12+
Quoted 5 lines

<[email protected]>

Link:
https://patch.msgid.link/[email protected]
(cherry picked from commit

9daa302a82590eeee7bdc68023ddad302df4b88c)

Same here, invalid git commit id.

You all really screw yourself with the bizarre "cherry-pick backwards in time" model :(

Please reconsider it as it seems to cause nothing but trouble.

greg k-h

author maintainer
#14 lore raw

Re: [PATCH 6.18.y 1/3] drm/xe/guc_ads: allocate UM queues in a separate BO

In reply to Jia Yao
Quoted 9 lines · Jia Yao

Move the UM queues into a dedicated BO (ads->um_queue_bo) and avoid CPU memset operations on it, which eliminates the CPU as a potential cacheline-polluting agent and helps maintain consistency between GAM writes and GuC reads.

We also need to ensure the base_dpa for the queue is contiguous on hw where this is used instead of a GGTT address. Another good reason to split this out to a separate BO.

Quoted 9 lines
Fixes 9c57bc08652a ("drm/xe/lnl: Drop force_probe requirement")
Cc Gwan-gyeong Mun <[email protected]>
Cc Matthew Auld <[email protected]>
Cc <[email protected]> # v6.12+
(cherry picked from commit 6af05de0dc19bbf3aaeec2020fe48b37c834b811)

This is not a valid git commit id anywhere that I can find :(

confused,

greg k-h

#15 lore raw

PATCH 1/3 drm/xe/guc_ads: allocate UM queues in a separate BO

In reply to [email protected]

Move the UM queues into a dedicated BO (ads->um_queue_bo) and avoid CPU memset operations on it, which eliminates the CPU as a potential cacheline-polluting agent and helps maintain consistency between GAM writes and GuC reads.

We also need to ensure the base_dpa for the queue is contiguous on hw where this is used instead of a GGTT address. Another good reason to split this out to a separate BO.

Fixes 9c57bc08652a ("drm/xe/lnl: Drop force_probe requirement")
Cc Gwan-gyeong Mun <[email protected]>
Cc Matthew Auld <[email protected]>
Cc <[email protected]> # v6.12+
(cherry picked from commit 99b01815957bd7d848420cb697f79ed11f7f215c)
---
 drivers/gpu/drm/xe/xe_bo.h            | 14 +++++
 drivers/gpu/drm/xe/xe_guc.c           |  6 ++
 drivers/gpu/drm/xe/xe_guc_ads.c       | 89 +++++++++++++++++----------
 drivers/gpu/drm/xe/xe_guc_ads_types.h |  5 ++
 4 files changed, 82 insertions(+), 32 deletions(-)
drivers/gpu/drm/xe/xe_bo.h +14 -0
index fef60c8a4e5b..71169b57fbda 100644
@@ -6,6 +6,7 @@
 #ifndef _XE_BO_H_
 #define _XE_BO_H_
 
+#include <drm/drm_prime.h>
 #include <drm/ttm/ttm_tt.h>
 
 #include "xe_bo_types.h"
@@ -455,6 +456,19 @@ void xe_bo_dev_fini(struct xe_bo_dev *bo_device);
 
 struct sg_table *xe_bo_sg(struct xe_bo *bo);
 
+/**
+ * xe_bo_sg_is_contiguous() - Check if a BO's DMA address space is contiguous.
+ * @bo: the BO to check (must have a valid sg table, i.e. !xe_bo_is_vram())
+ * @len: required contiguous length in bytes
+ *
+ * Returns true if the first @len bytes of the BO are mapped to a contiguous
+ * DMA address range.
+ */
+static inline bool xe_bo_sg_is_contiguous(struct xe_bo *bo, size_t len)
+{
+	return drm_prime_get_contiguous_size(xe_bo_sg(bo)) >= len;
+}
+
 /*
  * xe_sg_segment_size() - Provides upper limit for sg segment size.
  * @dev: device pointer
drivers/gpu/drm/xe/xe_guc.c +6 -0
index ae0c88da422b..cc5741f751b9 100644
@@ -700,6 +700,12 @@ 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;
 }
 
drivers/gpu/drm/xe/xe_guc_ads.c +57 -32

134 lines not on this page · raw message

drivers/gpu/drm/xe/xe_guc_ads_types.h +5 -0
index 70c132458ac3..acb5366f76ef 100644
@@ -16,6 +16,11 @@ struct xe_bo;
 struct xe_guc_ads {
 	/** @bo: XE BO for GuC ads blob */
 	struct xe_bo *bo;
+	/**
+	 * @um_queue_bo: Dedicated BO for the HW fault ring (UM queues).
+	 * NULL if the platform does not support USM.
+	 */
+	struct xe_bo *um_queue_bo;
 	/** @golden_lrc_size: golden LRC size */
 	size_t golden_lrc_size;
 	/** @regset_size: size of register set passed to GuC for save/restore */
Signature · 1 line
2.43.0
#16 lore raw

PATCH 2/3 drm/xe/guc_ads: allocate UM queues in VRAM on dGFX

On iGPU, the UM queue BO is allocated in system memory. On dGFX, the BO was previously created in system memory and later reallocated in xe_guc_realloc_post_hwconfig(). Allocate the UM queue BO directly in VRAM on dGFX, where it is ultimately required.

Fixes 9c57bc08652a ("drm/xe/lnl: Drop force_probe requirement")
Cc Gwan-gyeong Mun <[email protected]>
Cc Matthew Auld <[email protected]>
Cc <[email protected]> # v6.12+
(cherry picked from commit f342810a141f8a7e8b3786a6e4b6c0695a078a74)
---
 drivers/gpu/drm/xe/xe_guc.c     | 6 ------
 drivers/gpu/drm/xe/xe_guc_ads.c | 8 +++++---
 2 files changed, 5 insertions(+), 9 deletions(-)
drivers/gpu/drm/xe/xe_guc.c +0 -6
index cc5741f751b9..ae0c88da422b 100644
@@ -700,12 +700,6 @@ 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;
 }
 
drivers/gpu/drm/xe/xe_guc_ads.c +5 -3
index 86e6651c0036..4f48356ae622 100644
@@ -392,15 +392,17 @@ int xe_guc_ads_init(struct xe_guc_ads *ads)
 		/*
 		 * Allocate a separate BO for the HW fault ring (UM queues).
 		 *
-		 * Round the size up to the next power of two so that
+		 * Round the size up to the next power of two so that on iGPU
 		 * (system memory, no IOMMU) the TTM pool issues a single
 		 * alloc_pages(order=N) call, maximising the chance of getting
 		 * a physically contiguous block.  GuC requires contiguous DPA.
 		 */
-		size_t um_size = roundup_pow_of_two(GUC_UM_QUEUE_SIZE *
+		size_t um_size = IS_DGFX(xe) ?
+				 GUC_UM_QUEUE_SIZE * GUC_UM_HW_QUEUE_MAX :
+				 roundup_pow_of_two(GUC_UM_QUEUE_SIZE *
 						    GUC_UM_HW_QUEUE_MAX);
 
-		u32 um_flags = XE_BO_FLAG_SYSTEM |
+		u32 um_flags = XE_BO_FLAG_VRAM_IF_DGFX(tile) |
 			       XE_BO_FLAG_GGTT |
 			       XE_BO_FLAG_GGTT_INVALIDATE |
 			       XE_BO_FLAG_PINNED_NORESTORE;
Signature · 1 line
2.43.0
#17 lore raw

PATCH 3/3 drm/xe/guc_ads: use uncached mapping for UM queue BO

In reply to Jia Yao

On Pre-Xe3p platform, the GAM write the UM queue through DPA using UC. if GuC reads the queue via GGTT (WB), stale data may be observed when the cacheline has been polluted by another agent.

To match the GAM's UC writes, configure the GuC mapping as UC as well.

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

Reviewed by: Matthew Auld <[email protected]>

(cherry picked from commit 8d5134ae4177fa4f5a9bc8e71e6656cfc2852882)
---
 drivers/gpu/drm/xe/xe_guc_ads.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
drivers/gpu/drm/xe/xe_guc_ads.c +2 -1
index 4f48356ae622..63017ae2c381 100644
@@ -405,7 +405,8 @@ int xe_guc_ads_init(struct xe_guc_ads *ads)
 		u32 um_flags = XE_BO_FLAG_VRAM_IF_DGFX(tile) |
 			       XE_BO_FLAG_GGTT |
 			       XE_BO_FLAG_GGTT_INVALIDATE |
-			       XE_BO_FLAG_PINNED_NORESTORE;
+			       XE_BO_FLAG_PINNED_NORESTORE |
+			       XE_BO_FLAG_NEEDS_UC;
 
 		bo = xe_managed_bo_create_pin_map(xe, tile, um_size, um_flags);
 		if (IS_ERR(bo))
Signature · 1 line
2.43.0
maintainer Applied
#18 lore raw

Re: [PATCH 6.18.y 1/3] drm/xe/guc_ads: allocate UM queues in a separate BO

In reply to Jia Yao
Quoted 4 lines

Move the UM queues into a dedicated BO (ads->um_queue_bo) and avoid CPU memset operations on it, which eliminates the CPU as a potential cacheline-polluting agent and helps maintain consistency between GAM writes and GuC reads.

Queued for 6.18, thanks.

Signature · 2 lines
Thanks,
Sasha