Skip to contentSkip to conversation

xe2: Add PAT programming

PATCH Superseded Lucas De Marchi 4 participants 19 messages Sep 29, 2023 – Oct 2, 2023

Subsystems (2)

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

cover xe2: Add PAT programming

Add the PAT table and allow PTEs to be encoded with the additional bit.

This also adds a new debug capability: allow to dump the PAT table via debugfs and to dump it automatically when the driver fails to bind to the device. A wrong PAT table may be related with GuC failing to initialize or other failures that go away with additional debug messages. At least dumping the table allows one to quickly check if the values programmed are the expected ones.

Lucas De Marchi (6):
drm/xe/vm: Prefer xe_assert() over XE_WARN_ON() drm/xe/xe2: Follow XeHPC for TLB invalidation drm/xe/xe2: Add one more bit to encode PAT to ppgtt entries fixup! drm/xe/pat: Prefer the arch/IP names drm/xe/pat: Add debugfs node to dump PAT drm/xe/gt: Dump PAT table when failing to initialize

Matt Roper (1):
drm/xe/xe2: Program PAT tables

 drivers/gpu/drm/xe/xe_bo.h         |   1 +
 drivers/gpu/drm/xe/xe_ggtt.c       |   2 +-
 drivers/gpu/drm/xe/xe_gt.c         |  12 ++
 drivers/gpu/drm/xe/xe_gt_debugfs.c |  12 ++
 drivers/gpu/drm/xe/xe_pat.c        | 190 ++++++++++++++++++++++++++++-
 drivers/gpu/drm/xe/xe_pat.h        |   8 ++
 drivers/gpu/drm/xe/xe_vm.c         |   6 +-
 7 files changed, 227 insertions(+), 4 deletions(-)
Signature · 1 line
2.40.1
#1 lore raw

PATCH 1/7 drm/xe/xe2: Program PAT tables

From: Matt Roper <[email protected]>

The PAT tables become significantly more complicated on Xe2 platforms. They now control L3, L4, and coherency settings, as well as additional characteristics such as compression.

Aside from the main PAT table, there's an additional register that also needs to be programmed with PAT settings for PCI Address Translation Services.

Bspec: 71582

Cc Balasubramani Vivekanandan <[email protected]>
---
 drivers/gpu/drm/xe/xe_pat.c | 92 ++++++++++++++++++++++++++++++++++++-
 1 file changed, 91 insertions(+), 1 deletion(-)
drivers/gpu/drm/xe/xe_pat.c +91 -1
index 869aee64641f..ea2e3fb3a843 100644
@@ -10,10 +10,18 @@
 #include "xe_gt_mcr.h"
 #include "xe_mmio.h"
 
+#define _PAT_ATS				0x47fc
 #define _PAT_INDEX(index)			_PICK_EVEN_2RANGES(index, 8, \
 								   0x4800, 0x4804, \
 								   0x4848, 0x484c)
 
+#define XE2_NO_PROMOTE				REG_BIT(10)
+#define XE2_COMP_EN				REG_BIT(9)
+#define XE2_L3_CLOS				REG_GENMASK(7, 6)
+#define XE2_L3_POLICY				REG_GENMASK(5, 4)
+#define XE2_L4_POLICY				REG_GENMASK(3, 2)
+#define XE2_COH_MODE				REG_GENMASK(1, 0)
+
 #define XELPG_L4_POLICY_MASK			REG_GENMASK(3, 2)
 #define XELPG_PAT_3_UC				REG_FIELD_PREP(XELPG_L4_POLICY_MASK, 3)
 #define XELPG_PAT_1_WT				REG_FIELD_PREP(XELPG_L4_POLICY_MASK, 1)
@@ -67,6 +75,64 @@ static const u32 xelpg_pat_table[] = {
 	[4] = XELPG_PAT_0_WB | XELPG_3_COH_2W,
 };
 
+/*
+ * The Xe2 table is getting large/complicated so it's easier to review if
+ * provided in a form that exactly matches the bspec's formatting.  The meaning
+ * of the fields here are:
+ *   - no_promote:  0=promotable, 1=no promote
+ *   - comp_en:     0=disable, 1=enable
+ *   - l3clos:      L3 class of service (0-3)
+ *   - l3_policy:   0=WB, 1=XD ("WB - Transient Display"), 3=UC
+ *   - l4_policy:   0=WB, 1=WT, 3=UC
+ *   - coh_mode:    0=no snoop, 2=1-way coherent, 3=2-way coherent
+ *
+ * Reserved entries should be programmed with the maximum caching, minimum
+ * coherency (which matches an all-0's encoding), so we can just omit them
+ * in the table.
+ */
+#define XE2_PAT(no_promote, comp_en, l3clos, l3_policy, l4_policy, coh_mode) \
+	(no_promote ? XE2_NO_PROMOTE : 0) | \
+	(comp_en ? XE2_COMP_EN : 0) | \
+	REG_FIELD_PREP(XE2_L3_CLOS, l3clos) | \
+	REG_FIELD_PREP(XE2_L3_POLICY, l3_policy) | \
+	REG_FIELD_PREP(XE2_L4_POLICY, l4_policy) | \
+	REG_FIELD_PREP(XE2_COH_MODE, coh_mode)
+
+static const u32 xe2_pat_table[] = {
+	[ 0] = XE2_PAT( 0, 0, 0, 0, 3, 0 ),
+	[ 1] = XE2_PAT( 0, 0, 0, 0, 3, 2 ),
+	[ 2] = XE2_PAT( 0, 0, 0, 0, 3, 3 ),
+	[ 3] = XE2_PAT( 0, 0, 0, 3, 3, 0 ),
+	[ 4] = XE2_PAT( 0, 0, 0, 3, 0, 2 ),
+	[ 5] = XE2_PAT( 0, 0, 0, 3, 3, 2 ),
+	[ 6] = XE2_PAT( 1, 0, 0, 1, 3, 0 ),
+	[ 7] = XE2_PAT( 0, 0, 0, 3, 0, 3 ),
+	[ 8] = XE2_PAT( 0, 0, 0, 3, 0, 0 ),
+	[ 9] = XE2_PAT( 0, 1, 0, 0, 3, 0 ),
+	[10] = XE2_PAT( 0, 1, 0, 3, 0, 0 ),
+	[11] = XE2_PAT( 1, 1, 0, 1, 3, 0 ),
+	[12] = XE2_PAT( 0, 1, 0, 3, 3, 0 ),
+	[13] = XE2_PAT( 0, 0, 0, 0, 0, 0 ),
+	[14] = XE2_PAT( 0, 1, 0, 0, 0, 0 ),
+	[15] = XE2_PAT( 1, 1, 0, 1, 1, 0 ),
+	/* 16..19 are reserved; leave set to all 0's */
+	[20] = XE2_PAT( 0, 0, 1, 0, 3, 0 ),
+	[21] = XE2_PAT( 0, 1, 1, 0, 3, 0 ),
+	[22] = XE2_PAT( 0, 0, 1, 0, 3, 2 ),
+	[23] = XE2_PAT( 0, 0, 1, 0, 3, 3 ),
+	[24] = XE2_PAT( 0, 0, 2, 0, 3, 0 ),
+	[25] = XE2_PAT( 0, 1, 2, 0, 3, 0 ),
+	[26] = XE2_PAT( 0, 0, 2, 0, 3, 2 ),
+	[27] = XE2_PAT( 0, 0, 2, 0, 3, 3 ),
+	[28] = XE2_PAT( 0, 0, 3, 0, 3, 0 ),
+	[29] = XE2_PAT( 0, 1, 3, 0, 3, 0 ),
+	[30] = XE2_PAT( 0, 0, 3, 0, 3, 2 ),
+	[31] = XE2_PAT( 0, 0, 3, 0, 3, 3 ),
+};
+
+/* Special PAT values programmed outside the main table */
+#define XE2_PAT_ATS	XE2_PAT( 0, 0, 0, 0, 3, 3 )
+
 static void program_pat(struct xe_gt *gt, const u32 table[], int n_entries)
 {
 	for (int i = 0; i < n_entries; i++) {
@@ -102,9 +168,33 @@ static const struct xe_pat_ops xelpg_pat_ops = {
 	.program_media = program_pat_mcr,
 };
 
+static void xe2lpg_program_pat(struct xe_gt *gt, const u32 table[], int n_entries)
+{
+	program_pat_mcr(gt, table, n_entries);
+	xe_gt_mcr_multicast_write(gt, XE_REG_MCR(_PAT_ATS), XE2_PAT_ATS);
+}
+
+static void xe2lpm_program_pat(struct xe_gt *gt, const u32 table[], int n_entries)
+{
+	program_pat(gt, table, n_entries);
+	xe_mmio_write32(gt, XE_REG(_PAT_ATS), XE2_PAT_ATS);
+}
+
+static const struct xe_pat_ops xe2_pat_ops = {
+	.program_graphics = xe2lpg_program_pat,
+	.program_media = xe2lpm_program_pat,
+};
+
 void xe_pat_init_early(struct xe_device *xe)
 {
-	if (xe->info.platform == XE_METEORLAKE) {
+	if (GRAPHICS_VER(xe) == 20) {
+		xe->pat.ops = &xe2_pat_ops;
+		xe->pat.table = xe2_pat_table;
+		xe->pat.n_entries = ARRAY_SIZE(xe2_pat_table);
+		xe->pat.idx[XE_CACHE_NONE] = 3;
+		xe->pat.idx[XE_CACHE_WT] = 15;
+		xe->pat.idx[XE_CACHE_WB] = 2;
+	} else if (xe->info.platform == XE_METEORLAKE) {
 		xe->pat.ops = &xelpg_pat_ops;
 		xe->pat.table = xelpg_pat_table;
 		xe->pat.n_entries = ARRAY_SIZE(xelpg_pat_table);
Signature · 1 line
2.40.1
#2 lore raw

PATCH 2/7 drm/xe/vm: Prefer xe_assert() over XE_WARN_ON()

In reply to Lucas De Marchi

When xelp_pte_encode_addr() was added in commit 99a8844441ba ("drm/xe/migrate: Do not hand-encode pte"), there was no xe pointer for using xe_assert(). This is not the case anymore, so prefer it over XE_WARN_ON().

---
 drivers/gpu/drm/xe/xe_vm.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
drivers/gpu/drm/xe/xe_vm.c +2 -1
index beffbb1039d3..140a70d91d76 100644
@@ -17,6 +17,7 @@
 #include <linux/mm.h>
 #include <linux/swap.h>
 
+#include "xe_assert.h"
 #include "xe_bo.h"
 #include "xe_device.h"
 #include "xe_drm_client.h"
@@ -1293,7 +1294,7 @@ static u64 xelp_pte_encode_addr(struct xe_device *xe, u64 addr,
 	u64 pte;
 
 	/* Avoid passing random bits directly as flags */
-	XE_WARN_ON(flags & ~XE_PTE_PS64);
+	xe_assert(xe, !(flags & ~XE_PTE_PS64));
 
 	pte = addr;
 	pte |= XE_PAGE_PRESENT | XE_PAGE_RW;
Signature · 1 line
2.40.1
Reviewed-by on v1 2/7
#3 lore raw

Re: [Intel-xe] [PATCH 2/7] drm/xe/vm: Prefer xe_assert() over XE_WARN_ON()

Quoted 6 lines · Lucas De Marchi

When xelp_pte_encode_addr() was added in commit 99a8844441ba ("drm/xe/migrate: Do not hand-encode pte"), there was no xe pointer for using xe_assert(). This is not the case anymore, so prefer it over XE_WARN_ON().

Quoted 28 lines
---
 drivers/gpu/drm/xe/xe_vm.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
drivers/gpu/drm/xe/xe_vm.c +2 -1
index beffbb1039d3..140a70d91d76 100644
@@ -17,6 +17,7 @@
 #include <linux/mm.h>
 #include <linux/swap.h>
 
+#include "xe_assert.h"
 #include "xe_bo.h"
 #include "xe_device.h"
 #include "xe_drm_client.h"
@@ -1293,7 +1294,7 @@ static u64 xelp_pte_encode_addr(struct xe_device *xe, u64 addr,
 	u64 pte;
 
 	/* Avoid passing random bits directly as flags */
-	XE_WARN_ON(flags & ~XE_PTE_PS64);
+	xe_assert(xe, !(flags & ~XE_PTE_PS64));
 
 	pte = addr;
 	pte |= XE_PAGE_PRESENT | XE_PAGE_RW;
Signature · 1 line
2.40.1
Signature · 4 lines
Matt Roper
Graphics Software Engineer
Linux GPU Platform Enablement
Intel Corporation
#4 lore raw

PATCH 3/7 drm/xe/xe2: Follow XeHPC for TLB invalidation

In reply to Lucas De Marchi

Register GUC_TLB_INV_CR is gone in xe2. When GuC submission is not yet enabled, make sure to follow the same path as XeHPC.

---
 drivers/gpu/drm/xe/xe_ggtt.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/xe/xe_ggtt.c +1 -1
index 99b54794917e..d7b3d008ce57 100644
@@ -263,7 +263,7 @@ static void ggtt_invalidate_gt_tlb(struct xe_gt *gt)
 	} else if (xe_device_uc_enabled(gt_to_xe(gt))) {
 		struct xe_device *xe = gt_to_xe(gt);
 
-		if (xe->info.platform == XE_PVC) {
+		if (xe->info.platform == XE_PVC || GRAPHICS_VER(xe) >= 20) {
 			xe_mmio_write32(gt, PVC_GUC_TLB_INV_DESC1,
 					PVC_GUC_TLB_INV_DESC1_INVALIDATE);
 			xe_mmio_write32(gt, PVC_GUC_TLB_INV_DESC0,
Signature · 1 line
2.40.1
Reviewed-by on v1 3/7
#5 lore raw

Re: [Intel-xe] [PATCH 3/7] drm/xe/xe2: Follow XeHPC for TLB invalidation

Quoted 4 lines · Lucas De Marchi

Register GUC_TLB_INV_CR is gone in xe2. When GuC submission is not yet enabled, make sure to follow the same path as XeHPC.

Matches the invalidation descriptor structure from bspec 59311.

Although I am uncertain that the general invalidation code here is entirely correct, even for PVC. Since completion only gets delivered to the GuC (which isn't up and running yet at the point we're using this), we should probably be checking that the 'valid' bit has returned to 0 from the previous request before trying to submit another invalidation request. But assuming that's needed, it would be a change for a different patch.

Matt

Quoted 20 lines
---
 drivers/gpu/drm/xe/xe_ggtt.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/xe/xe_ggtt.c +1 -1
index 99b54794917e..d7b3d008ce57 100644
@@ -263,7 +263,7 @@ static void ggtt_invalidate_gt_tlb(struct xe_gt *gt)
 	} else if (xe_device_uc_enabled(gt_to_xe(gt))) {
 		struct xe_device *xe = gt_to_xe(gt);
 
-		if (xe->info.platform == XE_PVC) {
+		if (xe->info.platform == XE_PVC || GRAPHICS_VER(xe) >= 20) {
 			xe_mmio_write32(gt, PVC_GUC_TLB_INV_DESC1,
 					PVC_GUC_TLB_INV_DESC1_INVALIDATE);
 			xe_mmio_write32(gt, PVC_GUC_TLB_INV_DESC0,
Signature · 1 line
2.40.1
Signature · 4 lines
Matt Roper
Graphics Software Engineer
Linux GPU Platform Enablement
Intel Corporation
#6 lore raw

Re: [Intel-xe] [PATCH 3/7] drm/xe/xe2: Follow XeHPC for TLB invalidation

Quoted 6 lines · Matt Roper
Quoted 4 lines · Lucas De Marchi

Register GUC_TLB_INV_CR is gone in xe2. When GuC submission is not yet enabled, make sure to follow the same path as XeHPC.

Quoted 7 lines

Matches the invalidation descriptor structure from bspec 59311.

Although I am uncertain that the general invalidation code here is entirely correct, even for PVC. Since completion only gets delivered to the GuC (which isn't up and running yet at the point we're using this),

I think the point here is not about it being up and running yet, but rather having submission working. These are 2 different things.

Lucas De Marchi

Quoted 34 lines

we should probably be checking that the 'valid' bit has returned to 0 from the previous request before trying to submit another invalidation request. But assuming that's needed, it would be a change for a different patch.

Matt

Quoted 20 lines
---
 drivers/gpu/drm/xe/xe_ggtt.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/xe/xe_ggtt.c +1 -2
index 99b54794917e..d7b3d008ce57 100644
@@ -263,7 +263,7 @@ static void ggtt_invalidate_gt_tlb(struct xe_gt *gt)
 	} else if (xe_device_uc_enabled(gt_to_xe(gt))) {
 		struct xe_device *xe = gt_to_xe(gt);
 
-		if (xe->info.platform == XE_PVC) {
+		if (xe->info.platform == XE_PVC || GRAPHICS_VER(xe) >= 20) {
 			xe_mmio_write32(gt, PVC_GUC_TLB_INV_DESC1,
 					PVC_GUC_TLB_INV_DESC1_INVALIDATE);
 			xe_mmio_write32(gt, PVC_GUC_TLB_INV_DESC0,
--

2.40.1

Signature · 4 lines
Matt Roper
Graphics Software Engineer
Linux GPU Platform Enablement
Intel Corporation
#7 lore raw

PATCH 4/7 drm/xe/xe2: Add one more bit to encode PAT to ppgtt entries

In reply to Lucas De Marchi

Xe2 adds one more bit to cover all the possible 32 entries. Although those entries are not used by internal kernel code paths, it's expected that userspace will make use of it.

Bspec: 59510

---
 drivers/gpu/drm/xe/xe_bo.h | 1 +
 drivers/gpu/drm/xe/xe_vm.c | 3 +++
 2 files changed, 4 insertions(+)
drivers/gpu/drm/xe/xe_bo.h +1 -0
index 5090bdd1e462..7ffc26446d2e 100644
@@ -40,6 +40,7 @@
 #define XE_BO_INTERNAL_64K		BIT(31)
 
 #define XELPG_PPGTT_PTE_PAT3		BIT_ULL(62)
+#define XE2_PPGTT_PTE_PAT4		BIT_ULL(61)
 #define XE_PPGTT_PTE_PAT2		BIT_ULL(7)
 #define XE_PPGTT_PTE_PAT1		BIT_ULL(4)
 #define XE_PPGTT_PTE_PAT0		BIT_ULL(3)
drivers/gpu/drm/xe/xe_vm.c +3 -0
index 140a70d91d76..7e5f4bbea451 100644
@@ -1223,6 +1223,9 @@ static u64 pte_encode_cache(struct xe_device *xe, enum xe_cache_level cache)
 	if (pat_index & BIT(3))
 		pte |= XELPG_PPGTT_PTE_PAT3;
 
+	if (pat_index & (BIT(4)))
+		pte |= XE2_PPGTT_PTE_PAT4;
+
 	return pte;
 }
 
Signature · 1 line
2.40.1
Reviewed-by on v1 4/7
#8 lore raw

Re: [Intel-xe] [PATCH 4/7] drm/xe/xe2: Add one more bit to encode PAT to ppgtt entries

Quoted 40 lines

-----Original Message-----
From: Intel-xe <[email protected]> On Behalf Of Lucas De Marchi
Sent: Thursday, September 28, 2023 10:03 PM
To: [email protected]
Cc: De Marchi, Lucas <[email protected]>
Subject: [Intel-xe] [PATCH 4/7] drm/xe/xe2: Add one more bit to encode PAT to ppgtt entries

Xe2 adds one more bit to cover all the possible 32 entries. Although those entries are not used by internal kernel code paths, it's expected that userspace will make use of it.

Bspec: 59510

---
 drivers/gpu/drm/xe/xe_bo.h | 1 +
 drivers/gpu/drm/xe/xe_vm.c | 3 +++
 2 files changed, 4 insertions(+)
drivers/gpu/drm/xe/xe_bo.h index renamed from drivers/gpu/drm/xe/xe_bo.h +0 -0

5090bdd1e462..7ffc26446d2e 100644

drivers/gpu/drm/xe/xe_bo.h +1 -0
@@ -40,6 +40,7 @@
 #define XE_BO_INTERNAL_64K		BIT(31)
 
 #define XELPG_PPGTT_PTE_PAT3		BIT_ULL(62)
+#define XE2_PPGTT_PTE_PAT4		BIT_ULL(61)
 #define XE_PPGTT_PTE_PAT2		BIT_ULL(7)
 #define XE_PPGTT_PTE_PAT1		BIT_ULL(4)
 #define XE_PPGTT_PTE_PAT0		BIT_ULL(3)
drivers/gpu/drm/xe/xe_vm.c index renamed from drivers/gpu/drm/xe/xe_vm.c +0 -0

140a70d91d76..7e5f4bbea451 100644

drivers/gpu/drm/xe/xe_vm.c static u64 pte_encode_cache(struct xe_device *xe, · line 1231

Quoted 9 lines
drivers/gpu/drm/xe/xe_vm.c +3 -0
@@ -1223,6 +1223,9 @@ static u64 pte_encode_cache(struct xe_device *xe,
 enum xe_cache_level cache)
 	if (pat_index & BIT(3))
 		pte |= XELPG_PPGTT_PTE_PAT3;
 
+	if (pat_index & (BIT(4)))
+		pte |= XE2_PPGTT_PTE_PAT4;
+
 	return pte;
 }
Quoted 2 lines

--
2.40.1

Reviewed-by on v1 4/7
#9 lore raw

Re: [Intel-xe] [PATCH 4/7] drm/xe/xe2: Add one more bit to encode PAT to ppgtt entries

In reply to Lucas De Marchi
Quoted 6 lines · Lucas De Marchi

Xe2 adds one more bit to cover all the possible 32 entries. Although those entries are not used by internal kernel code paths, it's expected that userspace will make use of it.

Bspec: 59510

The same information is also available on Bspec 67095 too.

Quoted 34 lines
---
 drivers/gpu/drm/xe/xe_bo.h | 1 +
 drivers/gpu/drm/xe/xe_vm.c | 3 +++
 2 files changed, 4 insertions(+)
drivers/gpu/drm/xe/xe_bo.h +1 -0
index 5090bdd1e462..7ffc26446d2e 100644
@@ -40,6 +40,7 @@
 #define XE_BO_INTERNAL_64K		BIT(31)
 
 #define XELPG_PPGTT_PTE_PAT3		BIT_ULL(62)
+#define XE2_PPGTT_PTE_PAT4		BIT_ULL(61)
 #define XE_PPGTT_PTE_PAT2		BIT_ULL(7)
 #define XE_PPGTT_PTE_PAT1		BIT_ULL(4)
 #define XE_PPGTT_PTE_PAT0		BIT_ULL(3)
drivers/gpu/drm/xe/xe_vm.c +3 -0
index 140a70d91d76..7e5f4bbea451 100644
@@ -1223,6 +1223,9 @@ static u64 pte_encode_cache(struct xe_device *xe, enum xe_cache_level cache)
 	if (pat_index & BIT(3))
 		pte |= XELPG_PPGTT_PTE_PAT3;
 
+	if (pat_index & (BIT(4)))
+		pte |= XE2_PPGTT_PTE_PAT4;
+
 	return pte;
 }
 
Signature · 1 line
2.40.1
Signature · 4 lines
Matt Roper
Graphics Software Engineer
Linux GPU Platform Enablement
Intel Corporation
#10 lore raw

PATCH 5/7 fixup! drm/xe/pat: Prefer the arch/IP names

In reply to Lucas De Marchi

Fix alignment

---
 drivers/gpu/drm/xe/xe_pat.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/xe/xe_pat.c +1 -1
index ea2e3fb3a843..296763594370 100644
@@ -26,7 +26,7 @@
 #define XELPG_PAT_3_UC				REG_FIELD_PREP(XELPG_L4_POLICY_MASK, 3)
 #define XELPG_PAT_1_WT				REG_FIELD_PREP(XELPG_L4_POLICY_MASK, 1)
 #define XELPG_PAT_0_WB				REG_FIELD_PREP(XELPG_L4_POLICY_MASK, 0)
-#define XELPG_INDEX_COH_MODE_MASK			REG_GENMASK(1, 0)
+#define XELPG_INDEX_COH_MODE_MASK		REG_GENMASK(1, 0)
 #define XELPG_3_COH_2W				REG_FIELD_PREP(XELPG_INDEX_COH_MODE_MASK, 3)
 #define XELPG_2_COH_1W				REG_FIELD_PREP(XELPG_INDEX_COH_MODE_MASK, 2)
 #define XELPG_0_COH_NON				REG_FIELD_PREP(XELPG_INDEX_COH_MODE_MASK, 0)
Signature · 1 line
2.40.1
Reviewed-by on v1 5/7
#11 lore raw

Re: [Intel-xe] [PATCH 5/7] fixup! drm/xe/pat: Prefer the arch/IP names

Quoted 3 lines · Lucas De Marchi

Fix alignment

Quoted 20 lines
---
 drivers/gpu/drm/xe/xe_pat.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/xe/xe_pat.c +1 -1
index ea2e3fb3a843..296763594370 100644
@@ -26,7 +26,7 @@
 #define XELPG_PAT_3_UC				REG_FIELD_PREP(XELPG_L4_POLICY_MASK, 3)
 #define XELPG_PAT_1_WT				REG_FIELD_PREP(XELPG_L4_POLICY_MASK, 1)
 #define XELPG_PAT_0_WB				REG_FIELD_PREP(XELPG_L4_POLICY_MASK, 0)
-#define XELPG_INDEX_COH_MODE_MASK			REG_GENMASK(1, 0)
+#define XELPG_INDEX_COH_MODE_MASK		REG_GENMASK(1, 0)
 #define XELPG_3_COH_2W				REG_FIELD_PREP(XELPG_INDEX_COH_MODE_MASK, 3)
 #define XELPG_2_COH_1W				REG_FIELD_PREP(XELPG_INDEX_COH_MODE_MASK, 2)
 #define XELPG_0_COH_NON				REG_FIELD_PREP(XELPG_INDEX_COH_MODE_MASK, 0)
Signature · 1 line
2.40.1
Signature · 4 lines
Matt Roper
Graphics Software Engineer
Linux GPU Platform Enablement
Intel Corporation
#12 lore raw

PATCH 6/7 drm/xe/pat: Add debugfs node to dump PAT

In reply to Lucas De Marchi

This is useful to debug cache issues, to double check if the PAT indexes match what they were supposed to be set to from spec.

---
 drivers/gpu/drm/xe/xe_gt_debugfs.c | 12 ++++
 drivers/gpu/drm/xe/xe_pat.c        | 96 ++++++++++++++++++++++++++++++
 drivers/gpu/drm/xe/xe_pat.h        |  8 +++
 3 files changed, 116 insertions(+)
drivers/gpu/drm/xe/xe_gt_debugfs.c +12 -0
index ec1ae00f6bfc..cd6d28c7b923 100644
@@ -16,6 +16,7 @@
 #include "xe_gt_topology.h"
 #include "xe_hw_engine.h"
 #include "xe_macros.h"
+#include "xe_pat.h"
 #include "xe_reg_sr.h"
 #include "xe_reg_whitelist.h"
 #include "xe_uc_debugfs.h"
@@ -138,6 +139,16 @@ static int workarounds(struct seq_file *m, void *data)
 	return 0;
 }
 
+static int pat(struct seq_file *m, void *data)
+{
+	struct xe_gt *gt = node_to_gt(m->private);
+	struct drm_printer p = drm_seq_file_printer(m);
+
+	xe_pat_dump(gt, &p);
+
+	return 0;
+}
+
 static const struct drm_info_list debugfs_list[] = {
 	{"hw_engines", hw_engines, 0},
 	{"force_reset", force_reset, 0},
@@ -147,6 +158,7 @@ static const struct drm_info_list debugfs_list[] = {
 	{"ggtt", ggtt, 0},
 	{"register-save-restore", register_save_restore, 0},
 	{"workarounds", workarounds, 0},
+	{"pat", pat, 0},
 };
 
 void xe_gt_debugfs_register(struct xe_gt *gt)
drivers/gpu/drm/xe/xe_pat.c +96 -0

144 lines not on this page · raw message

drivers/gpu/drm/xe/xe_pat.h +8 -0
index 168e80e63809..09c491ab9f15 100644
@@ -6,6 +6,7 @@
 #ifndef _XE_PAT_H_
 #define _XE_PAT_H_
 
+struct drm_printer;
 struct xe_gt;
 struct xe_device;
 
@@ -21,4 +22,11 @@ void xe_pat_init_early(struct xe_device *xe);
  */
 void xe_pat_init(struct xe_gt *gt);
 
+/**
+ * xe_pat_dump - Dump PAT table
+ * @gt: GT structure
+ * @p: Printer to dump info to
+ */
+void xe_pat_dump(struct xe_gt *gt, struct drm_printer *p);
+
 #endif
Signature · 1 line
2.40.1
#13 lore raw

Re: [Intel-xe] [PATCH 6/7] drm/xe/pat: Add debugfs node to dump PAT

Quoted 102 lines · Lucas De Marchi

This is useful to debug cache issues, to double check if the PAT indexes match what they were supposed to be set to from spec.

---
 drivers/gpu/drm/xe/xe_gt_debugfs.c | 12 ++++
 drivers/gpu/drm/xe/xe_pat.c        | 96 ++++++++++++++++++++++++++++++
 drivers/gpu/drm/xe/xe_pat.h        |  8 +++
 3 files changed, 116 insertions(+)
drivers/gpu/drm/xe/xe_gt_debugfs.c +12 -0
index ec1ae00f6bfc..cd6d28c7b923 100644
@@ -16,6 +16,7 @@
 #include "xe_gt_topology.h"
 #include "xe_hw_engine.h"
 #include "xe_macros.h"
+#include "xe_pat.h"
 #include "xe_reg_sr.h"
 #include "xe_reg_whitelist.h"
 #include "xe_uc_debugfs.h"
@@ -138,6 +139,16 @@ static int workarounds(struct seq_file *m, void *data)
 	return 0;
 }
 
+static int pat(struct seq_file *m, void *data)
+{
+	struct xe_gt *gt = node_to_gt(m->private);
+	struct drm_printer p = drm_seq_file_printer(m);
+
+	xe_pat_dump(gt, &p);
+
+	return 0;
+}
+
 static const struct drm_info_list debugfs_list[] = {
 	{"hw_engines", hw_engines, 0},
 	{"force_reset", force_reset, 0},
@@ -147,6 +158,7 @@ static const struct drm_info_list debugfs_list[] = {
 	{"ggtt", ggtt, 0},
 	{"register-save-restore", register_save_restore, 0},
 	{"workarounds", workarounds, 0},
+	{"pat", pat, 0},
 };
 
 void xe_gt_debugfs_register(struct xe_gt *gt)
drivers/gpu/drm/xe/xe_pat.c +26 -0
index 296763594370..20714888a5fa 100644
@@ -6,6 +6,8 @@
 #include "xe_pat.h"
 
 #include "regs/xe_reg_defs.h"
+#include "xe_assert.h"
+#include "xe_device.h"
 #include "xe_gt.h"
 #include "xe_gt_mcr.h"
 #include "xe_mmio.h"
@@ -14,6 +16,7 @@
 #define _PAT_INDEX(index)			_PICK_EVEN_2RANGES(index, 8, \
 								   0x4800, 0x4804, \
 								   0x4848, 0x484c)
+#define _PAT_PTA				0x4820
 
 #define XE2_NO_PROMOTE				REG_BIT(10)
 #define XE2_COMP_EN				REG_BIT(9)
@@ -43,6 +46,7 @@
 struct xe_pat_ops {
 	void (*program_graphics)(struct xe_gt *gt, const u32 table[], int n_entries);
 	void (*program_media)(struct xe_gt *gt, const u32 table[], int n_entries);
+	void (*dump)(struct xe_gt *gt, struct drm_printer *p);
 };
 
 static const u32 xelp_pat_table[] = {
@@ -151,12 +155,38 @@ static void program_pat_mcr(struct xe_gt *gt, const u32 table[], int n_entries)
 	}
 }
 
+void xelp_dump(struct xe_gt *gt, struct drm_printer *p)
+{
+	struct xe_device *xe = gt_to_xe(gt);
+	int i, err;
+
+	xe_device_mem_access_get(xe);
+	err = xe_force_wake_get(gt_to_fw(gt), XE_FW_GT);
+	if (err)
+		goto err_fw;
+
+	drm_printf(p, "PAT table:\n");
+
+	for (i = 0; i < xe->pat.n_entries; i++) {
+		u32 pat = xe_mmio_read32(gt, XE_REG(_PAT_INDEX(i)));
+
+		drm_printf(p, "PAT[%2d] = %#8x\n", i, pat);
+	}
+
+	err = xe_force_wake_put(gt_to_fw(gt), XE_FW_GT);
+err_fw:
+	xe_assert(xe, !err);
+	xe_device_mem_access_put(xe);

drivers/gpu/drm/xe/xe_pat.c static void program_pat_mcr(struct xe_gt *gt, const u32 table[], int n_entries) · line 189

Quoted 10 lines
drivers/gpu/drm/xe/xe_pat.c +4 -0
+}
+
 static const struct xe_pat_ops xelp_pat_ops = {
 	.program_graphics = program_pat,
+	.dump = xelp_dump,
 };
 
 static const struct xe_pat_ops xehp_pat_ops = {
 	.program_graphics = program_pat_mcr,
+	.dump = xelp_dump,

If we use the unicast xelp dumper, some of our reads might happen while another thread is steering us to a group/instance that corresponds to a fused-off mslice, resulting in 0's instead of the correct values. We probably need an xehp-specific variant that does MCR reads.

It might also be nice to have an xehpc-specific dumping function that does field-based printout ( CLOS, MEM_TYPE ).

drivers/gpu/drm/xe/xe_pat.c static const struct xe_pat_ops xehp_pat_ops = { · line 199

Quoted 8 lines
drivers/gpu/drm/xe/xe_pat.c +1 -0
 };
 
 /*
@@ -166,6 +196,7 @@ static const struct xe_pat_ops xehp_pat_ops = {
 static const struct xe_pat_ops xelpg_pat_ops = {
 	.program_graphics = program_pat,
 	.program_media = program_pat_mcr,
+	.dump = xelp_dump,

Xe_LPG / Xe_LPM+ need to choose MCR vs non-MCR based on GT type as you've done with Xe2 below. It would also be nice to have field-based dump output that presents ( L4_POLICY, COH_MODE ) separately.

Matt

Quoted 109 lines
drivers/gpu/drm/xe/xe_pat.c +65 -0
 };
 
 static void xe2lpg_program_pat(struct xe_gt *gt, const u32 table[], int n_entries)
@@ -180,9 +211,64 @@ static void xe2lpm_program_pat(struct xe_gt *gt, const u32 table[], int n_entrie
 	xe_mmio_write32(gt, XE_REG(_PAT_ATS), XE2_PAT_ATS);
 }
 
+void xe2_dump(struct xe_gt *gt, struct drm_printer *p)
+{
+	struct xe_device *xe = gt_to_xe(gt);
+	int i, err;
+	u32 pat;
+
+	xe_device_mem_access_get(xe);
+	err = xe_force_wake_get(gt_to_fw(gt), XE_FW_GT);
+	if (err)
+		goto err_fw;
+
+	drm_printf(p, "PAT table:\n");
+
+	for (i = 0; i < xe->pat.n_entries; i++) {
+		if (xe_gt_is_media_type(gt))
+			pat = xe_mmio_read32(gt, XE_REG(_PAT_INDEX(i)));
+		else
+			pat = xe_gt_mcr_unicast_read_any(gt, XE_REG_MCR(_PAT_INDEX(i)));
+
+		drm_printf(p, "PAT[%2d] = [ %u, %u, %u, %u, %u, %u ]  (%#8x)\n", i,
+			   !!(pat & XE2_NO_PROMOTE),
+			   !!(pat & XE2_COMP_EN),
+			   REG_FIELD_GET(XE2_L3_CLOS, pat),
+			   REG_FIELD_GET(XE2_L3_POLICY, pat),
+			   REG_FIELD_GET(XE2_L4_POLICY, pat),
+			   REG_FIELD_GET(XE2_COH_MODE, pat),
+			   pat);
+	}
+
+	/*
+	 * Also print PTA_MODE, which describes how the hardware accesses
+	 * PPGTT entries.
+	 */
+	if (xe_gt_is_media_type(gt))
+		pat = xe_mmio_read32(gt, XE_REG(_PAT_PTA));
+	else
+		pat = xe_gt_mcr_unicast_read_any(gt, XE_REG_MCR(_PAT_PTA));
+
+	drm_printf(p, "Page Table Access:\n");
+	drm_printf(p, "PTA_MODE= [ %u, %u, %u, %u, %u, %u ]  (%#8x)\n",
+		   !!(pat & XE2_NO_PROMOTE),
+		   !!(pat & XE2_COMP_EN),
+		   REG_FIELD_GET(XE2_L3_CLOS, pat),
+		   REG_FIELD_GET(XE2_L3_POLICY, pat),
+		   REG_FIELD_GET(XE2_L4_POLICY, pat),
+		   REG_FIELD_GET(XE2_COH_MODE, pat),
+		   pat);
+
+	err = xe_force_wake_put(gt_to_fw(gt), XE_FW_GT);
+err_fw:
+	xe_assert(xe, !err);
+	xe_device_mem_access_put(xe);
+}
+
 static const struct xe_pat_ops xe2_pat_ops = {
 	.program_graphics = xe2lpg_program_pat,
 	.program_media = xe2lpm_program_pat,
+	.dump = xe2_dump,
 };
 
 void xe_pat_init_early(struct xe_device *xe)
@@ -252,3 +338,13 @@ void xe_pat_init(struct xe_gt *gt)
 	else
 		xe->pat.ops->program_graphics(gt, xe->pat.table, xe->pat.n_entries);
 }
+
+void xe_pat_dump(struct xe_gt *gt, struct drm_printer *p)
+{
+	struct xe_device *xe = gt_to_xe(gt);
+
+	if (!xe->pat.ops->dump)
+		return;
+
+	xe->pat.ops->dump(gt, p);
+}
drivers/gpu/drm/xe/xe_pat.h +8 -0
index 168e80e63809..09c491ab9f15 100644
@@ -6,6 +6,7 @@
 #ifndef _XE_PAT_H_
 #define _XE_PAT_H_
 
+struct drm_printer;
 struct xe_gt;
 struct xe_device;
 
@@ -21,4 +22,11 @@ void xe_pat_init_early(struct xe_device *xe);
  */
 void xe_pat_init(struct xe_gt *gt);
 
+/**
+ * xe_pat_dump - Dump PAT table
+ * @gt: GT structure
+ * @p: Printer to dump info to
+ */
+void xe_pat_dump(struct xe_gt *gt, struct drm_printer *p);
+
 #endif
Signature · 1 line
2.40.1
Signature · 4 lines
Matt Roper
Graphics Software Engineer
Linux GPU Platform Enablement
Intel Corporation
#14 lore raw

Re: [Intel-xe] [PATCH 6/7] drm/xe/pat: Add debugfs node to dump PAT

Quoted 110 lines · Matt Roper
Quoted 109 lines · Lucas De Marchi

This is useful to debug cache issues, to double check if the PAT indexes match what they were supposed to be set to from spec.

---
 drivers/gpu/drm/xe/xe_gt_debugfs.c | 12 ++++
 drivers/gpu/drm/xe/xe_pat.c        | 96 ++++++++++++++++++++++++++++++
 drivers/gpu/drm/xe/xe_pat.h        |  8 +++
 3 files changed, 116 insertions(+)
drivers/gpu/drm/xe/xe_gt_debugfs.c +12 -0
index ec1ae00f6bfc..cd6d28c7b923 100644
@@ -16,6 +16,7 @@
 #include "xe_gt_topology.h"
 #include "xe_hw_engine.h"
 #include "xe_macros.h"
+#include "xe_pat.h"
 #include "xe_reg_sr.h"
 #include "xe_reg_whitelist.h"
 #include "xe_uc_debugfs.h"
@@ -138,6 +139,16 @@ static int workarounds(struct seq_file *m, void *data)
 	return 0;
 }
 
+static int pat(struct seq_file *m, void *data)
+{
+	struct xe_gt *gt = node_to_gt(m->private);
+	struct drm_printer p = drm_seq_file_printer(m);
+
+	xe_pat_dump(gt, &p);
+
+	return 0;
+}
+
 static const struct drm_info_list debugfs_list[] = {
 	{"hw_engines", hw_engines, 0},
 	{"force_reset", force_reset, 0},
@@ -147,6 +158,7 @@ static const struct drm_info_list debugfs_list[] = {
 	{"ggtt", ggtt, 0},
 	{"register-save-restore", register_save_restore, 0},
 	{"workarounds", workarounds, 0},
+	{"pat", pat, 0},
 };
 
 void xe_gt_debugfs_register(struct xe_gt *gt)
drivers/gpu/drm/xe/xe_pat.c +29 -0
index 296763594370..20714888a5fa 100644
@@ -6,6 +6,8 @@
 #include "xe_pat.h"
 
 #include "regs/xe_reg_defs.h"
+#include "xe_assert.h"
+#include "xe_device.h"
 #include "xe_gt.h"
 #include "xe_gt_mcr.h"
 #include "xe_mmio.h"
@@ -14,6 +16,7 @@
 #define _PAT_INDEX(index)			_PICK_EVEN_2RANGES(index, 8, \
 								   0x4800, 0x4804, \
 								   0x4848, 0x484c)
+#define _PAT_PTA				0x4820
 
 #define XE2_NO_PROMOTE				REG_BIT(10)
 #define XE2_COMP_EN				REG_BIT(9)
@@ -43,6 +46,7 @@
 struct xe_pat_ops {
 	void (*program_graphics)(struct xe_gt *gt, const u32 table[], int n_entries);
 	void (*program_media)(struct xe_gt *gt, const u32 table[], int n_entries);
+	void (*dump)(struct xe_gt *gt, struct drm_printer *p);
 };
 
 static const u32 xelp_pat_table[] = {
@@ -151,12 +155,38 @@ static void program_pat_mcr(struct xe_gt *gt, const u32 table[], int n_entries)
 	}
 }
 
+void xelp_dump(struct xe_gt *gt, struct drm_printer *p)
+{
+	struct xe_device *xe = gt_to_xe(gt);
+	int i, err;
+
+	xe_device_mem_access_get(xe);
+	err = xe_force_wake_get(gt_to_fw(gt), XE_FW_GT);
+	if (err)
+		goto err_fw;
+
+	drm_printf(p, "PAT table:\n");
+
+	for (i = 0; i < xe->pat.n_entries; i++) {
+		u32 pat = xe_mmio_read32(gt, XE_REG(_PAT_INDEX(i)));
+
+		drm_printf(p, "PAT[%2d] = %#8x\n", i, pat);
+	}
+
+	err = xe_force_wake_put(gt_to_fw(gt), XE_FW_GT);
+err_fw:
+	xe_assert(xe, !err);
+	xe_device_mem_access_put(xe);
+}
+
 static const struct xe_pat_ops xelp_pat_ops = {
 	.program_graphics = program_pat,
+	.dump = xelp_dump,
 };
 
Quoted 8 lines

drivers/gpu/drm/xe/xe_pat.c static void program_pat_mcr(struct xe_gt *gt, const u32 table[], int n_entries) · line 189

Quoted 3 lines
drivers/gpu/drm/xe/xe_pat.c +1 -0
 static const struct xe_pat_ops xehp_pat_ops = {
 	.program_graphics = program_pat_mcr,
+	.dump = xelp_dump,

If we use the unicast xelp dumper, some of our reads might happen while another thread is steering us to a group/instance that corresponds to a fused-off mslice, resulting in 0's instead of the correct values. We probably need an xehp-specific variant that does MCR reads.

ugh... I added the additional dump function while coding this, but decided to remove since I thought it was the media gt that did the MCR read. Since we don't have a separate media GT in PVC and DG2 then we wouldn't need the check. I will add that back

Quoted 3 lines

It might also be nice to have an xehpc-specific dumping function that does field-based printout ( CLOS, MEM_TYPE ).

ok

Quoted 12 lines

drivers/gpu/drm/xe/xe_pat.c static const struct xe_pat_ops xehp_pat_ops = { · line 199

Quoted 8 lines
drivers/gpu/drm/xe/xe_pat.c +1 -0
 };
 
 /*
@@ -166,6 +196,7 @@ static const struct xe_pat_ops xehp_pat_ops = {
 static const struct xe_pat_ops xelpg_pat_ops = {
 	.program_graphics = program_pat,
 	.program_media = program_pat_mcr,
+	.dump = xelp_dump,

Xe_LPG / Xe_LPM+ need to choose MCR vs non-MCR based on GT type as you've done with Xe2 below. It would also be nice to have field-based dump output that presents ( L4_POLICY, COH_MODE ) separately.

ok

thanks
Lucas De Marchi

Quoted 119 lines

Matt

Quoted 109 lines
drivers/gpu/drm/xe/xe_pat.c +65 -0
 };
 
 static void xe2lpg_program_pat(struct xe_gt *gt, const u32 table[], int n_entries)
@@ -180,9 +211,64 @@ static void xe2lpm_program_pat(struct xe_gt *gt, const u32 table[], int n_entrie
 	xe_mmio_write32(gt, XE_REG(_PAT_ATS), XE2_PAT_ATS);
 }
 
+void xe2_dump(struct xe_gt *gt, struct drm_printer *p)
+{
+	struct xe_device *xe = gt_to_xe(gt);
+	int i, err;
+	u32 pat;
+
+	xe_device_mem_access_get(xe);
+	err = xe_force_wake_get(gt_to_fw(gt), XE_FW_GT);
+	if (err)
+		goto err_fw;
+
+	drm_printf(p, "PAT table:\n");
+
+	for (i = 0; i < xe->pat.n_entries; i++) {
+		if (xe_gt_is_media_type(gt))
+			pat = xe_mmio_read32(gt, XE_REG(_PAT_INDEX(i)));
+		else
+			pat = xe_gt_mcr_unicast_read_any(gt, XE_REG_MCR(_PAT_INDEX(i)));
+
+		drm_printf(p, "PAT[%2d] = [ %u, %u, %u, %u, %u, %u ]  (%#8x)\n", i,
+			   !!(pat & XE2_NO_PROMOTE),
+			   !!(pat & XE2_COMP_EN),
+			   REG_FIELD_GET(XE2_L3_CLOS, pat),
+			   REG_FIELD_GET(XE2_L3_POLICY, pat),
+			   REG_FIELD_GET(XE2_L4_POLICY, pat),
+			   REG_FIELD_GET(XE2_COH_MODE, pat),
+			   pat);
+	}
+
+	/*
+	 * Also print PTA_MODE, which describes how the hardware accesses
+	 * PPGTT entries.
+	 */
+	if (xe_gt_is_media_type(gt))
+		pat = xe_mmio_read32(gt, XE_REG(_PAT_PTA));
+	else
+		pat = xe_gt_mcr_unicast_read_any(gt, XE_REG_MCR(_PAT_PTA));
+
+	drm_printf(p, "Page Table Access:\n");
+	drm_printf(p, "PTA_MODE= [ %u, %u, %u, %u, %u, %u ]  (%#8x)\n",
+		   !!(pat & XE2_NO_PROMOTE),
+		   !!(pat & XE2_COMP_EN),
+		   REG_FIELD_GET(XE2_L3_CLOS, pat),
+		   REG_FIELD_GET(XE2_L3_POLICY, pat),
+		   REG_FIELD_GET(XE2_L4_POLICY, pat),
+		   REG_FIELD_GET(XE2_COH_MODE, pat),
+		   pat);
+
+	err = xe_force_wake_put(gt_to_fw(gt), XE_FW_GT);
+err_fw:
+	xe_assert(xe, !err);
+	xe_device_mem_access_put(xe);
+}
+
 static const struct xe_pat_ops xe2_pat_ops = {
 	.program_graphics = xe2lpg_program_pat,
 	.program_media = xe2lpm_program_pat,
+	.dump = xe2_dump,
 };
 
 void xe_pat_init_early(struct xe_device *xe)
@@ -252,3 +338,13 @@ void xe_pat_init(struct xe_gt *gt)
 	else
 		xe->pat.ops->program_graphics(gt, xe->pat.table, xe->pat.n_entries);
 }
+
+void xe_pat_dump(struct xe_gt *gt, struct drm_printer *p)
+{
+	struct xe_device *xe = gt_to_xe(gt);
+
+	if (!xe->pat.ops->dump)
+		return;
+
+	xe->pat.ops->dump(gt, p);
+}
drivers/gpu/drm/xe/xe_pat.h +8 -1
index 168e80e63809..09c491ab9f15 100644
@@ -6,6 +6,7 @@
 #ifndef _XE_PAT_H_
 #define _XE_PAT_H_
 
+struct drm_printer;
 struct xe_gt;
 struct xe_device;
 
@@ -21,4 +22,11 @@ void xe_pat_init_early(struct xe_device *xe);
  */
 void xe_pat_init(struct xe_gt *gt);
 
+/**
+ * xe_pat_dump - Dump PAT table
+ * @gt: GT structure
+ * @p: Printer to dump info to
+ */
+void xe_pat_dump(struct xe_gt *gt, struct drm_printer *p);
+
 #endif
--

2.40.1

Signature · 4 lines
Matt Roper
Graphics Software Engineer
Linux GPU Platform Enablement
Intel Corporation
#15 lore raw

PATCH 7/7 drm/xe/gt: Dump PAT table when failing to initialize

In reply to Lucas De Marchi

When failing on early initialization, one cause may be that the PAT configuration is not correct. Dump it for ease of debugging.

---
 drivers/gpu/drm/xe/xe_gt.c | 12 ++++++++++++
 1 file changed, 12 insertions(+)
drivers/gpu/drm/xe/xe_gt.c +12 -0
index 1aa44d4f9ac1..1a750c88cdd6 100644
@@ -307,6 +307,17 @@ int xe_gt_init_early(struct xe_gt *gt)
 	return 0;
 }
 
+static void dump_pat_on_error(struct xe_gt *gt)
+{
+	struct drm_printer p;
+	char prefix[32];
+
+	snprintf(prefix, sizeof(prefix), "[GT%u Error]", gt->info.id);
+	p = drm_debug_printer(prefix);
+
+	xe_pat_dump(gt, &p);
+}
+
 static int gt_fw_domain_init(struct xe_gt *gt)
 {
 	int err, i;
@@ -360,6 +371,7 @@ static int gt_fw_domain_init(struct xe_gt *gt)
 	return 0;
 
 err_force_wake:
+	dump_pat_on_error(gt);
 	xe_force_wake_put(gt_to_fw(gt), XE_FW_GT);
 err_hw_fence_irq:
 	for (i = 0; i < XE_ENGINE_CLASS_MAX; ++i)
Signature · 1 line
2.40.1
Reviewed-by on v1 7/7
#16 lore raw

Re: [Intel-xe] [PATCH 7/7] drm/xe/gt: Dump PAT table when failing to initialize

Quoted 13 lines · Lucas De Marchi

When failing on early initialization, one cause may be that the PAT configuration is not correct. Dump it for ease of debugging.

---
 drivers/gpu/drm/xe/xe_gt.c | 12 ++++++++++++
 1 file changed, 12 insertions(+)
drivers/gpu/drm/xe/xe_gt.c +0 -0
index 1aa44d4f9ac1..1a750c88cdd6 100644

drivers/gpu/drm/xe/xe_gt.c int xe_gt_init_early(struct xe_gt *gt) · line 315

Quoted 9 lines
drivers/gpu/drm/xe/xe_gt.c +6 -0
@@ -307,6 +307,17 @@ int xe_gt_init_early(struct xe_gt *gt)
 	return 0;
 }
 
+static void dump_pat_on_error(struct xe_gt *gt)
+{
+	struct drm_printer p;
+	char prefix[32];
+
+	snprintf(prefix, sizeof(prefix), "[GT%u Error]", gt->info.id);

Should we put 'PAT' somewhere in the string here so it's clear what all the numbers getting dumped are supposed to be? Or else add a drm_printf() to add a "title" before calling the dump command below?

Up to you.

Quoted 19 lines
drivers/gpu/drm/xe/xe_gt.c +6 -0
+	p = drm_debug_printer(prefix);
+
+	xe_pat_dump(gt, &p);
+}
+
 static int gt_fw_domain_init(struct xe_gt *gt)
 {
 	int err, i;
@@ -360,6 +371,7 @@ static int gt_fw_domain_init(struct xe_gt *gt)
 	return 0;
 
 err_force_wake:
+	dump_pat_on_error(gt);
 	xe_force_wake_put(gt_to_fw(gt), XE_FW_GT);
 err_hw_fence_irq:
 	for (i = 0; i < XE_ENGINE_CLASS_MAX; ++i)
Signature · 1 line
2.40.1
Signature · 4 lines
Matt Roper
Graphics Software Engineer
Linux GPU Platform Enablement
Intel Corporation
#17 lore raw

Re: [Intel-xe] [PATCH 7/7] drm/xe/gt: Dump PAT table when failing to initialize

Quoted 17 lines · Matt Roper
Quoted 16 lines · Lucas De Marchi

When failing on early initialization, one cause may be that the PAT configuration is not correct. Dump it for ease of debugging.

---
 drivers/gpu/drm/xe/xe_gt.c | 12 ++++++++++++
 1 file changed, 12 insertions(+)
drivers/gpu/drm/xe/xe_gt.c +0 -0
index 1aa44d4f9ac1..1a750c88cdd6 100644
@@ -307,6 +307,17 @@ int xe_gt_init_early(struct xe_gt *gt)
 	return 0;
 }
 
Quoted 8 lines

drivers/gpu/drm/xe/xe_gt.c int xe_gt_init_early(struct xe_gt *gt) · line 315

Quoted 6 lines
drivers/gpu/drm/xe/xe_gt.c +6 -0
+static void dump_pat_on_error(struct xe_gt *gt)
+{
+	struct drm_printer p;
+	char prefix[32];
+
+	snprintf(prefix, sizeof(prefix), "[GT%u Error]", gt->info.id);

Should we put 'PAT' somewhere in the string here so it's clear what all

the number are being printed with:

drm_printf(p, "PAT table:\n");
...
drm_printf(p, "PAT[%2d] = [ %u, %u, %u, %u, %u, %u ] (%#8x)\n", i,

           !!(pat & XE2_NO_PROMOTE),
           !!(pat & XE2_COMP_EN),
           REG_FIELD_GET(XE2_L3_CLOS, pat),
           REG_FIELD_GET(XE2_L3_POLICY, pat),
           REG_FIELD_GET(XE2_L4_POLICY, pat),
           REG_FIELD_GET(XE2_COH_MODE, pat),
           pat);

So, I don't think we should add another PAT string.

Lucas De Marchi

Quoted 32 lines

the numbers getting dumped are supposed to be? Or else add a drm_printf() to add a "title" before calling the dump command below?

Up to you.

Quoted 19 lines
drivers/gpu/drm/xe/xe_gt.c +6 -1
+	p = drm_debug_printer(prefix);
+
+	xe_pat_dump(gt, &p);
+}
+
 static int gt_fw_domain_init(struct xe_gt *gt)
 {
 	int err, i;
@@ -360,6 +371,7 @@ static int gt_fw_domain_init(struct xe_gt *gt)
 	return 0;
 
 err_force_wake:
+	dump_pat_on_error(gt);
 	xe_force_wake_put(gt_to_fw(gt), XE_FW_GT);
 err_hw_fence_irq:
 	for (i = 0; i < XE_ENGINE_CLASS_MAX; ++i)
--

2.40.1

Signature · 4 lines
Matt Roper
Graphics Software Engineer
Linux GPU Platform Enablement
Intel Corporation
#18 lore raw

[Intel-xe] ✗ CI.Patch_applied: failure for xe2: Add PAT programming

In reply to Lucas De Marchi
== Series Details ==

== Series Details ==

Series: xe2: Add PAT programming
URL : https://patchwork.freedesktop.org/series/124430/
State : failure

== Summary ==

=== Applying kernel patches on branch 'drm-xe-next' with base: === Base commit: 9a795b4f0 drm/xe: timeout needs to be a signed value === git am output follows ===
Applying: drm/xe/xe2: Program PAT tables
Applying: drm/xe/vm: Prefer xe_assert() over XE_WARN_ON()
Applying: drm/xe/xe2: Follow XeHPC for TLB invalidation
Applying: drm/xe/xe2: Add one more bit to encode PAT to ppgtt entries
Applying: fixup! drm/xe/pat: Prefer the arch/IP names
Applying: drm/xe/pat: Add debugfs node to dump PAT
Applying: drm/xe/gt: Dump PAT table when failing to initialize