Skip to contentSkip to conversation

Add DG2 OA support

PATCH Superseded Umesh Nerlige Ramappa 4 participants 27 messages Oct 10, 2022 – Oct 12, 2022

Subsystems (3)

v3 of 6 Superseded Latest is v6 →

lore
Share & tools
lore
#0 lore raw

v3 cover Add DG2 OA support

Add OA format support for DG2 and various fixes for DG2.

This series has 2 uapi changes listed below:

1) drm/i915/perf: Add OAG and OAR formats for DG2

DG2 has new OA formats defined that can be selected by the user. The UMD changes that are consumed by GPUvis are:
https://patchwork.freedesktop.org/patch/504456/?series=107633&rev=5

2) drm/i915/perf: Apply Wa_18013179988

DG2 has a bug where the OA timestamp does not tick at the CS timestamp frequency. Instead it ticks at a multiple that is determined from the CTC_SHIFT value in RPM_CONFIG. Since the timestamp is used by UMD to make sense of all the counters in the report, expose the OA timestamp frequency to the user. The interface is generic and applies to all platforms. On platforms where the bug is not present, this returns the CS timestamp frequency. UMD specific changes consumed by GPUvis are:
https://patchwork.freedesktop.org/patch/504464/?series=107633&rev=5

v2:
- Add review comments
- Update uapi changes in cover letter
- Drop patches for non-production platforms drm/i915/perf: Use helpers to process reports w.r.t. OA buffer size drm/i915/perf: Add Wa_16010703925:dg2

- Drop 64-bit OA format changes for now drm/i915/perf: Parse 64bit report header formats correctly drm/i915/perf: Add Wa_1608133521:dg2

v3:
- Add review comments to patches 02, 04, 05, 14
- Drop Acks

Test-with: [email protected]

Lionel Landwerlin (1):
drm/i915/perf: complete programming whitelisting for XEHPSDV

Umesh Nerlige Ramappa (14):
drm/i915/perf: Fix OA filtering logic for GuC mode drm/i915/perf: Add 32-bit OAG and OAR formats for DG2 drm/i915/perf: Fix noa wait predication for DG2 drm/i915/perf: Determine gen12 oa ctx offset at runtime drm/i915/perf: Enable bytes per clock reporting in OA drm/i915/perf: Simply use stream->ctx
drm/i915/perf: Move gt-specific data from i915->perf to gt->perf drm/i915/perf: Replace gt->perf.lock with stream->lock for file ops drm/i915/perf: Use gt-specific ggtt for OA and noa-wait buffers drm/i915/perf: Store a pointer to oa_format in oa_buffer drm/i915/perf: Add Wa_1508761755:dg2
drm/i915/perf: Apply Wa_18013179988
drm/i915/perf: Save/restore EU flex counters across reset drm/i915/perf: Enable OA for DG2

Vinay Belgaumkar (1):
drm/i915/guc: Support OA when Wa_16011777198 is enabled

 drivers/gpu/drm/i915/gt/intel_engine_regs.h   |   1 +
 drivers/gpu/drm/i915/gt/intel_gpu_commands.h  |   4 +
 drivers/gpu/drm/i915/gt/intel_gt_regs.h       |   1 +
 drivers/gpu/drm/i915/gt/intel_gt_types.h      |   3 +
 drivers/gpu/drm/i915/gt/intel_lrc.h           |   2 +
 drivers/gpu/drm/i915/gt/intel_sseu.c          |   4 +-
 .../drm/i915/gt/uc/abi/guc_actions_slpc_abi.h |   9 +
 drivers/gpu/drm/i915/gt/uc/intel_guc_ads.c    |   8 +
 drivers/gpu/drm/i915/gt/uc/intel_guc_slpc.c   |  66 ++
 drivers/gpu/drm/i915/gt/uc/intel_guc_slpc.h   |   2 +
 drivers/gpu/drm/i915/i915_drv.h               |   5 +
 drivers/gpu/drm/i915/i915_getparam.c          |   3 +
 drivers/gpu/drm/i915/i915_pci.c               |   2 +
 drivers/gpu/drm/i915/i915_perf.c              | 583 ++++++++++++++----
 drivers/gpu/drm/i915/i915_perf.h              |   2 +
 drivers/gpu/drm/i915/i915_perf_oa_regs.h      |   6 +-
 drivers/gpu/drm/i915/i915_perf_types.h        |  47 +-
 drivers/gpu/drm/i915/intel_device_info.h      |   2 +
 drivers/gpu/drm/i915/selftests/i915_perf.c    |  16 +-
 include/uapi/drm/i915_drm.h                   |  10 +
 20 files changed, 629 insertions(+), 147 deletions(-)
Signature · 1 line
2.25.1
#1 lore raw

PATCH v3 1/16 drm/i915/perf: Fix OA filtering logic for GuC mode

With GuC mode of submission, GuC is in control of defining the context id field that is part of the OA reports. To filter reports, UMD and KMD must know what sw context id was chosen by GuC. There is not interface between KMD and GuC to determine this, so read the upper-dword of EXECLIST_STATUS to filter/squash OA reports for the specific context.

v2: Explain guc id stealing w.r.t OA use case

---
 drivers/gpu/drm/i915/gt/intel_lrc.h |   2 +
 drivers/gpu/drm/i915/i915_perf.c    | 144 ++++++++++++++++++++++++----
 2 files changed, 127 insertions(+), 19 deletions(-)
drivers/gpu/drm/i915/gt/intel_lrc.h +2 -0
index a390f0813c8b..7111bae759f3 100644
@@ -110,6 +110,8 @@ enum {
 #define XEHP_SW_CTX_ID_WIDTH			16
 #define XEHP_SW_COUNTER_SHIFT			58
 #define XEHP_SW_COUNTER_WIDTH			6
+#define GEN12_GUC_SW_CTX_ID_SHIFT		39
+#define GEN12_GUC_SW_CTX_ID_WIDTH		16
 
 static inline void lrc_runtime_start(struct intel_context *ce)
 {
drivers/gpu/drm/i915/i915_perf.c +125 -19

168 lines not on this page · raw message

Signature · 1 line
2.25.1
#2 lore raw

PATCH v3 2/16 drm/i915/perf: Add 32-bit OAG and OAR formats for DG2

In reply to Umesh Nerlige Ramappa

Add new OA formats for DG2.

v2:
- Update commit title (Ashutosh)
- Coding style fixes (Lionel)
- 64 bit OA formats need UMD changes in GPUvis, drop for now and send in a separate series with UMD changes

v3:
- Update commit message to drop 64 bit related description

---
 drivers/gpu/drm/i915/i915_perf.c | 7 +++++++
 include/uapi/drm/i915_drm.h      | 4 ++++
 2 files changed, 11 insertions(+)
drivers/gpu/drm/i915/i915_perf.c +7 -0
index 315662329be3..41e9f620ee31 100644
@@ -320,6 +320,8 @@ static const struct i915_oa_format oa_formats[I915_OA_FORMAT_MAX] = {
 	[I915_OA_FORMAT_A12]		    = { 0, 64 },
 	[I915_OA_FORMAT_A12_B8_C8]	    = { 2, 128 },
 	[I915_OA_FORMAT_A32u40_A4u32_B8_C8] = { 5, 256 },
+	[I915_OAR_FORMAT_A32u40_A4u32_B8_C8]    = { 5, 256 },
+	[I915_OA_FORMAT_A24u40_A14u32_B8_C8]    = { 5, 256 },
 };
 
 #define SAMPLE_OA_REPORT      (1<<0)
@@ -4517,6 +4519,11 @@ static void oa_init_supported_formats(struct i915_perf *perf)
 		oa_format_add(perf, I915_OA_FORMAT_C4_B8);
 		break;
 
+	case INTEL_DG2:
+		oa_format_add(perf, I915_OAR_FORMAT_A32u40_A4u32_B8_C8);
+		oa_format_add(perf, I915_OA_FORMAT_A24u40_A14u32_B8_C8);
+		break;
+
 	default:
 		MISSING_CASE(platform);
 	}
include/uapi/drm/i915_drm.h +4 -0
index 520ad2691a99..8b59590e06d4 100644
@@ -2650,6 +2650,10 @@ enum drm_i915_oa_format {
 	I915_OA_FORMAT_A12_B8_C8,
 	I915_OA_FORMAT_A32u40_A4u32_B8_C8,
 
+	/* DG2 */
+	I915_OAR_FORMAT_A32u40_A4u32_B8_C8,
+	I915_OA_FORMAT_A24u40_A14u32_B8_C8,
+
 	I915_OA_FORMAT_MAX	    /* non-ABI */
 };
 
Signature · 1 line
2.25.1
#3 lore raw

PATCH v3 3/16 drm/i915/perf: Fix noa wait predication for DG2

In reply to Umesh Nerlige Ramappa

Predication for batch buffer commands changed in XEHPSDV. MI_BATCH_BUFFER_START predicates based on MI_SET_PREDICATE_RESULT register. The MI_SET_PREDICATE_RESULT register can only be modified with MI_SET_PREDICATE command. When configured, the MI_SET_PREDICATE command sets MI_SET_PREDICATE_RESULT based on bit 0 of MI_PREDICATE_RESULT_2. Use this to configure predication in noa_wait.

---
 drivers/gpu/drm/i915/gt/intel_engine_regs.h |  1 +
 drivers/gpu/drm/i915/i915_perf.c            | 24 +++++++++++++++++----
 2 files changed, 21 insertions(+), 4 deletions(-)
drivers/gpu/drm/i915/gt/intel_engine_regs.h +1 -0
index fe1a0d5fd4b1..ee3efd06ee54 100644
@@ -201,6 +201,7 @@
 #define RING_CONTEXT_STATUS_PTR(base)		_MMIO((base) + 0x3a0)
 #define RING_CTX_TIMESTAMP(base)		_MMIO((base) + 0x3a8) /* gen8+ */
 #define RING_PREDICATE_RESULT(base)		_MMIO((base) + 0x3b8)
+#define MI_PREDICATE_RESULT_2_ENGINE(base)	_MMIO((base) + 0x3bc)
 #define RING_FORCE_TO_NONPRIV(base, i)		_MMIO(((base) + 0x4D0) + (i) * 4)
 #define   RING_FORCE_TO_NONPRIV_DENY		REG_BIT(30)
 #define   RING_FORCE_TO_NONPRIV_ADDRESS_MASK	REG_GENMASK(25, 2)
drivers/gpu/drm/i915/i915_perf.c +20 -4
index 41e9f620ee31..cd57b5836386 100644
@@ -286,6 +286,7 @@ static u32 i915_perf_stream_paranoid = true;
 #define OAREPORT_REASON_CTX_SWITCH     (1<<3)
 #define OAREPORT_REASON_CLK_RATIO      (1<<5)
 
+#define HAS_MI_SET_PREDICATE(i915) (GRAPHICS_VER_FULL(i915) >= IP_VER(12, 50))
 
 /* For sysctl proc_dointvec_minmax of i915_oa_max_sample_rate
  *
@@ -1762,6 +1763,9 @@ static int alloc_noa_wait(struct i915_perf_stream *stream)
 		DELTA_TARGET,
 		N_CS_GPR
 	};
+	i915_reg_t mi_predicate_result = HAS_MI_SET_PREDICATE(i915) ?
+					  MI_PREDICATE_RESULT_2_ENGINE(base) :
+					  MI_PREDICATE_RESULT_1(RENDER_RING_BASE);
 
 	bo = i915_gem_object_create_internal(i915, 4096);
 	if (IS_ERR(bo)) {
@@ -1799,7 +1803,7 @@ static int alloc_noa_wait(struct i915_perf_stream *stream)
 			stream, cs, true /* save */, CS_GPR(i),
 			INTEL_GT_SCRATCH_FIELD_PERF_CS_GPR + 8 * i, 2);
 	cs = save_restore_register(
-		stream, cs, true /* save */, MI_PREDICATE_RESULT_1(RENDER_RING_BASE),
+		stream, cs, true /* save */, mi_predicate_result,
 		INTEL_GT_SCRATCH_FIELD_PERF_PREDICATE_RESULT_1, 1);
 
 	/* First timestamp snapshot location. */
@@ -1853,7 +1857,10 @@ static int alloc_noa_wait(struct i915_perf_stream *stream)
 	 */
 	*cs++ = MI_LOAD_REGISTER_REG | (3 - 2);
 	*cs++ = i915_mmio_reg_offset(CS_GPR(JUMP_PREDICATE));
-	*cs++ = i915_mmio_reg_offset(MI_PREDICATE_RESULT_1(RENDER_RING_BASE));
+	*cs++ = i915_mmio_reg_offset(mi_predicate_result);
+
+	if (HAS_MI_SET_PREDICATE(i915))
+		*cs++ = MI_SET_PREDICATE | 1;
 
 	/* Restart from the beginning if we had timestamps roll over. */
 	*cs++ = (GRAPHICS_VER(i915) < 8 ?
@@ -1863,6 +1870,9 @@ static int alloc_noa_wait(struct i915_perf_stream *stream)
 	*cs++ = i915_ggtt_offset(vma) + (ts0 - batch) * 4;
 	*cs++ = 0;
 
+	if (HAS_MI_SET_PREDICATE(i915))
+		*cs++ = MI_SET_PREDICATE;
+
 	/*
 	 * Now add the diff between to previous timestamps and add it to :
 	 *      (((1 * << 64) - 1) - delay_ns)
@@ -1890,7 +1900,10 @@ static int alloc_noa_wait(struct i915_perf_stream *stream)
 	 */
 	*cs++ = MI_LOAD_REGISTER_REG | (3 - 2);
 	*cs++ = i915_mmio_reg_offset(CS_GPR(JUMP_PREDICATE));
-	*cs++ = i915_mmio_reg_offset(MI_PREDICATE_RESULT_1(RENDER_RING_BASE));
+	*cs++ = i915_mmio_reg_offset(mi_predicate_result);
+
+	if (HAS_MI_SET_PREDICATE(i915))
+		*cs++ = MI_SET_PREDICATE | 1;
 
 	/* Predicate the jump.  */
 	*cs++ = (GRAPHICS_VER(i915) < 8 ?
@@ -1900,13 +1913,16 @@ static int alloc_noa_wait(struct i915_perf_stream *stream)
 	*cs++ = i915_ggtt_offset(vma) + (jump - batch) * 4;
 	*cs++ = 0;
 
+	if (HAS_MI_SET_PREDICATE(i915))
+		*cs++ = MI_SET_PREDICATE;
+
 	/* Restore registers. */
 	for (i = 0; i < N_CS_GPR; i++)
 		cs = save_restore_register(
 			stream, cs, false /* restore */, CS_GPR(i),
 			INTEL_GT_SCRATCH_FIELD_PERF_CS_GPR + 8 * i, 2);
 	cs = save_restore_register(
-		stream, cs, false /* restore */, MI_PREDICATE_RESULT_1(RENDER_RING_BASE),
+		stream, cs, false /* restore */, mi_predicate_result,
 		INTEL_GT_SCRATCH_FIELD_PERF_PREDICATE_RESULT_1, 1);
 
 	/* And return to the ring. */
Signature · 1 line
2.25.1
#4 lore raw

PATCH v3 4/16 drm/i915/perf: Determine gen12 oa ctx offset at runtime

In reply to Umesh Nerlige Ramappa

Some SKUs of same gen12 platform may have different oactxctrl offsets. For gen12, determine oactxctrl offsets at runtime.

v2: (Lionel)
- Move MI definitions to intel_gpu_commands.h
- Ensure __find_reg_in_lri does read past context image size

v3: (Ashutosh)
- Drop unnecessary use of double underscores
- fix find_reg_in_lri
- Return error if oa context offset is U32_MAX
- Error out if oa_ctx_ctrl_offset does not find offset

---
 drivers/gpu/drm/i915/gt/intel_gpu_commands.h |   4 +
 drivers/gpu/drm/i915/i915_perf.c             | 154 +++++++++++++++----
 drivers/gpu/drm/i915/i915_perf_oa_regs.h     |   2 +-
 3 files changed, 129 insertions(+), 31 deletions(-)
drivers/gpu/drm/i915/gt/intel_gpu_commands.h +4 -0
index d4e9702d3c8e..f50ea92910d9 100644
@@ -187,6 +187,10 @@
 #define   MI_BATCH_RESOURCE_STREAMER REG_BIT(10)
 #define   MI_BATCH_PREDICATE         REG_BIT(15) /* HSW+ on RCS only*/
 
+#define MI_OPCODE(x)		(((x) >> 23) & 0x3f)
+#define IS_MI_LRI_CMD(x)	(MI_OPCODE(x) == MI_OPCODE(MI_INSTR(0x22, 0)))
+#define MI_LRI_LEN(x)		(((x) & 0xff) + 1)
+
 /*
  * 3D instructions used by the kernel
  */
drivers/gpu/drm/i915/i915_perf.c +124 -30

224 lines not on this page · raw message

drivers/gpu/drm/i915/i915_perf_oa_regs.h +1 -1
index f31c9f13a9fc..0ef3562ff4aa 100644
@@ -97,7 +97,7 @@
 #define  GEN12_OAR_OACONTROL_COUNTER_FORMAT_SHIFT 1
 #define  GEN12_OAR_OACONTROL_COUNTER_ENABLE       (1 << 0)
 
-#define GEN12_OACTXCONTROL _MMIO(0x2360)
+#define GEN12_OACTXCONTROL(base) _MMIO((base) + 0x360)
 #define GEN12_OAR_OASTATUS _MMIO(0x2968)
 
 /* Gen12 OAG unit */
Signature · 1 line
2.25.1
Reviewed-by on v3 4/16
#5 lore raw

Re: [Intel-gfx] [PATCH v3 04/16] drm/i915/perf: Determine gen12 oa ctx offset at runtime

On Mon, 10 Oct 2022 11:14:22 -0700, Umesh Nerlige Ramappa wrote:

Hi Umesh,

Quoted 5 lines
drivers/gpu/drm/i915/i915_perf.c +0 -0
index cd57b5836386..b292aa39633e 100644

drivers/gpu/drm/i915/i915_perf.c static int gen12_get_render_context_id(struct i915_perf_stream *stream) · line 1366

Quoted 9 lines
drivers/gpu/drm/i915/i915_perf.c +6 -0
@@ -1358,6 +1358,68 @@ static int gen12_get_render_context_id(struct i915_perf_stream *stream)
 	return 0;
 }
 
+#define valid_oactxctrl_offset(x) ((x) && (x) != U32_MAX)
+static bool oa_find_reg_in_lri(u32 *state, u32 reg, u32 *offset, u32 end)
+{
+	u32 idx = *offset;
+	u32 len = min(MI_LRI_LEN(state[idx]) + idx, end);
+	bool found = false;

My recommendation would be to put something like this here:

	if (MI_LRI_LEN(state[idx]) & 0x1)
		drm_warn("MI_LRI instruction with odd length\n");

Because we expect the MI_LRI length to even and I am not sure what is in the context image. But maybe not needed?

drivers/gpu/drm/i915/i915_perf.c static int gen12_configure_oar_context(struct i915_perf_stream *stream, · line 2518

Quoted 10 lines
drivers/gpu/drm/i915/i915_perf.c +2 -4
@@ -2436,15 +2514,18 @@ static int gen12_configure_oar_context(struct i915_perf_stream *stream,
 		},
 	};
 
-	/* Modify the context image of pinned context with regs_context*/
-	err = intel_context_lock_pinned(ce);
-	if (err)
-		return err;
+	/* Modify the context image of pinned context with regs_context */
+	if (valid_oactxctrl_offset(offset)) {

As I said before, this check is not needed, if we didn't have valid a offset we should return error from oa_get_render_ctx_id. For if we have this check and offset is not valid, can we skip this code and will things still work? Or do we need to return an error from gen12_configure_oar_context?

Quoted 4 lines
drivers/gpu/drm/i915/i915_perf.c +3 -0
+		err = intel_context_lock_pinned(ce);
+		if (err)
+			return err;
 

drivers/gpu/drm/i915/i915_perf.c static int gen12_configure_oar_context(struct i915_perf_stream *stream, · line 2528

Quoted 10 lines
drivers/gpu/drm/i915/i915_perf.c +6 -4
-	err = gen8_modify_context(ce, regs_context, ARRAY_SIZE(regs_context));
-	intel_context_unlock_pinned(ce);
-	if (err)
-		return err;
+		err = gen8_modify_context(ce, regs_context,
+					  ARRAY_SIZE(regs_context));
+		intel_context_unlock_pinned(ce);
+		if (err)
+			return err;
+	}

With the if statement removed or handled otherwise, this is:

#6 lore raw

Re: [Intel-gfx] [PATCH v3 04/16] drm/i915/perf: Determine gen12 oa ctx offset at runtime

Quoted 21 lines · Ashutosh

On Mon, 10 Oct 2022 11:14:22 -0700, Umesh Nerlige Ramappa wrote:

Hi Umesh,

Quoted 5 lines
drivers/gpu/drm/i915/i915_perf.c +0 -0
index cd57b5836386..b292aa39633e 100644

drivers/gpu/drm/i915/i915_perf.c static int gen12_get_render_context_id(struct i915_perf_stream *stream) · line 1366

Quoted 9 lines
drivers/gpu/drm/i915/i915_perf.c +6 -0
@@ -1358,6 +1358,68 @@ static int gen12_get_render_context_id(struct i915_perf_stream *stream)
 	return 0;
 }
 
+#define valid_oactxctrl_offset(x) ((x) && (x) != U32_MAX)
+static bool oa_find_reg_in_lri(u32 *state, u32 reg, u32 *offset, u32 end)
+{
+	u32 idx = *offset;
+	u32 len = min(MI_LRI_LEN(state[idx]) + idx, end);
+	bool found = false;

My recommendation would be to put something like this here:

Quoted 5 lines
	if (MI_LRI_LEN(state[idx]) & 0x1)
		drm_warn("MI_LRI instruction with odd length\n");

Because we expect the MI_LRI length to even and I am not sure what is in the context image. But maybe not needed?

will add.

Quoted 5 lines
Quoted 5 lines
drivers/gpu/drm/i915/i915_perf.c +0 -0
@@ -2436,15 +2514,18 @@ static int gen12_configure_oar_context(struct i915_perf_stream *stream,
 		},
 	};
 
Quoted 12 lines

drivers/gpu/drm/i915/i915_perf.c static int gen12_configure_oar_context(struct i915_perf_stream *stream, · line 2518

Quoted 6 lines
drivers/gpu/drm/i915/i915_perf.c +2 -4
-	/* Modify the context image of pinned context with regs_context*/
-	err = intel_context_lock_pinned(ce);
-	if (err)
-		return err;
+	/* Modify the context image of pinned context with regs_context */
+	if (valid_oactxctrl_offset(offset)) {

As I said before, this check is not needed, if we didn't have valid a offset we should return error from oa_get_render_ctx_id. For if we have this check and offset is not valid, can we skip this code and will things still work? Or do we need to return an error from gen12_configure_oar_context?

I missed that from the previous comments. Will add.

Quoted 19 lines
Quoted 5 lines
drivers/gpu/drm/i915/i915_perf.c +3 -0
+		err = intel_context_lock_pinned(ce);
+		if (err)
+			return err;
 

drivers/gpu/drm/i915/i915_perf.c static int gen12_configure_oar_context(struct i915_perf_stream *stream, · line 2528

Quoted 10 lines
drivers/gpu/drm/i915/i915_perf.c +6 -4
-	err = gen8_modify_context(ce, regs_context, ARRAY_SIZE(regs_context));
-	intel_context_unlock_pinned(ce);
-	if (err)
-		return err;
+		err = gen8_modify_context(ce, regs_context,
+					  ARRAY_SIZE(regs_context));
+		intel_context_unlock_pinned(ce);
+		if (err)
+			return err;
+	}

With the if statement removed or handled otherwise, this is:

maintainer
#7 lore raw

Re: [Intel-gfx] [PATCH v3 04/16] drm/i915/perf: Determine gen12 oa ctx offset at runtime

In reply to Umesh Nerlige Ramappa
Quoted 40 lines · Umesh Nerlige Ramappa

Some SKUs of same gen12 platform may have different oactxctrl offsets. For gen12, determine oactxctrl offsets at runtime.

v2: (Lionel)
- Move MI definitions to intel_gpu_commands.h
- Ensure __find_reg_in_lri does read past context image size

v3: (Ashutosh)
- Drop unnecessary use of double underscores
- fix find_reg_in_lri
- Return error if oa context offset is U32_MAX
- Error out if oa_ctx_ctrl_offset does not find offset

---
 drivers/gpu/drm/i915/gt/intel_gpu_commands.h |   4 +
 drivers/gpu/drm/i915/i915_perf.c             | 154 +++++++++++++++----
 drivers/gpu/drm/i915/i915_perf_oa_regs.h     |   2 +-
 3 files changed, 129 insertions(+), 31 deletions(-)
drivers/gpu/drm/i915/gt/intel_gpu_commands.h +4 -0
index d4e9702d3c8e..f50ea92910d9 100644
@@ -187,6 +187,10 @@
 #define   MI_BATCH_RESOURCE_STREAMER REG_BIT(10)
 #define   MI_BATCH_PREDICATE         REG_BIT(15) /* HSW+ on RCS only*/
 
+#define MI_OPCODE(x)		(((x) >> 23) & 0x3f)
+#define IS_MI_LRI_CMD(x)	(MI_OPCODE(x) == MI_OPCODE(MI_INSTR(0x22, 0)))
+#define MI_LRI_LEN(x)		(((x) & 0xff) + 1)
+
 /*
  * 3D instructions used by the kernel
  */
drivers/gpu/drm/i915/i915_perf.c +0 -0
index cd57b5836386..b292aa39633e 100644

drivers/gpu/drm/i915/i915_perf.c static int gen12_get_render_context_id(struct i915_perf_stream *stream) · line 1365

Quoted 8 lines
drivers/gpu/drm/i915/i915_perf.c +5 -0
@@ -1358,6 +1358,68 @@ static int gen12_get_render_context_id(struct i915_perf_stream *stream)
 	return 0;
 }
 
+#define valid_oactxctrl_offset(x) ((x) && (x) != U32_MAX)
+static bool oa_find_reg_in_lri(u32 *state, u32 reg, u32 *offset, u32 end)
+{
+	u32 idx = *offset;
+	u32 len = min(MI_LRI_LEN(state[idx]) + idx, end);

I don't understand any of this stuff, but why are the offset, index and length u32 instead of just, say, int?

drivers/gpu/drm/i915/i915_perf.c static int gen12_get_render_context_id(struct i915_perf_stream *stream) · line 1369

Quoted 4 lines
drivers/gpu/drm/i915/i915_perf.c +4 -0
+	bool found = false;
+
+	idx++;
+	for (; idx < len; idx += 2) {

I find the initialization of idx and len confusing, and thereby the entire loop too.

BR,
Jani.

Quoted 235 lines
drivers/gpu/drm/i915/i915_perf.c +115 -30
+		if (state[idx] == reg) {
+			found = true;
+			break;
+		}
+	}
+
+	*offset = idx;
+	return found;
+}
+
+static u32 oa_context_image_offset(struct intel_context *ce, u32 reg)
+{
+	u32 offset, len = (ce->engine->context_size - PAGE_SIZE) / 4;
+	u32 *state = ce->lrc_reg_state;
+
+	for (offset = 0; offset < len; ) {
+		if (IS_MI_LRI_CMD(state[offset])) {
+			if (oa_find_reg_in_lri(state, reg, &offset, len))
+				break;
+		} else {
+			offset++;
+		}
+	}
+
+	return offset < len ? offset : U32_MAX;
+}
+
+static int set_oa_ctx_ctrl_offset(struct intel_context *ce)
+{
+	i915_reg_t reg = GEN12_OACTXCONTROL(ce->engine->mmio_base);
+	struct i915_perf *perf = &ce->engine->i915->perf;
+	u32 offset = perf->ctx_oactxctrl_offset;
+
+	/* Do this only once. Failure is stored as offset of U32_MAX */
+	if (offset)
+		goto exit;
+
+	offset = oa_context_image_offset(ce, i915_mmio_reg_offset(reg));
+	perf->ctx_oactxctrl_offset = offset;
+
+	drm_dbg(&ce->engine->i915->drm,
+		"%s oa ctx control at 0x%08x dword offset\n",
+		ce->engine->name, offset);
+
+exit:
+	return valid_oactxctrl_offset(offset) ? 0 : -ENODEV;
+}
+
+static bool engine_supports_mi_query(struct intel_engine_cs *engine)
+{
+	return engine->class == RENDER_CLASS;
+}
+
 /**
  * oa_get_render_ctx_id - determine and hold ctx hw id
  * @stream: An i915-perf stream opened for OA metrics
@@ -1377,6 +1439,21 @@ static int oa_get_render_ctx_id(struct i915_perf_stream *stream)
 	if (IS_ERR(ce))
 		return PTR_ERR(ce);
 
+	if (engine_supports_mi_query(stream->engine)) {
+		/*
+		 * We are enabling perf query here. If we don't find the context
+		 * offset here, just return an error.
+		 */
+		ret = set_oa_ctx_ctrl_offset(ce);
+		if (ret) {
+			intel_context_unpin(ce);
+			drm_err(&stream->perf->i915->drm,
+				"Enabling perf query failed for %s\n",
+				stream->engine->name);
+			return ret;
+		}
+	}
+
 	switch (GRAPHICS_VER(ce->engine->i915)) {
 	case 7: {
 		/*
@@ -2408,10 +2485,11 @@ static int gen12_configure_oar_context(struct i915_perf_stream *stream,
 	int err;
 	struct intel_context *ce = stream->pinned_ctx;
 	u32 format = stream->oa_buffer.format;
+	u32 offset = stream->perf->ctx_oactxctrl_offset;
 	struct flex regs_context[] = {
 		{
 			GEN8_OACTXCONTROL,
-			stream->perf->ctx_oactxctrl_offset + 1,
+			offset + 1,
 			active ? GEN8_OA_COUNTER_RESUME : 0,
 		},
 	};
@@ -2436,15 +2514,18 @@ static int gen12_configure_oar_context(struct i915_perf_stream *stream,
 		},
 	};
 
-	/* Modify the context image of pinned context with regs_context*/
-	err = intel_context_lock_pinned(ce);
-	if (err)
-		return err;
+	/* Modify the context image of pinned context with regs_context */
+	if (valid_oactxctrl_offset(offset)) {
+		err = intel_context_lock_pinned(ce);
+		if (err)
+			return err;
 
-	err = gen8_modify_context(ce, regs_context, ARRAY_SIZE(regs_context));
-	intel_context_unlock_pinned(ce);
-	if (err)
-		return err;
+		err = gen8_modify_context(ce, regs_context,
+					  ARRAY_SIZE(regs_context));
+		intel_context_unlock_pinned(ce);
+		if (err)
+			return err;
+	}
 
 	/* Apply regs_lri using LRI with pinned context */
 	return gen8_modify_self(ce, regs_lri, ARRAY_SIZE(regs_lri), active);
@@ -2566,6 +2647,7 @@ lrc_configure_all_contexts(struct i915_perf_stream *stream,
 			   const struct i915_oa_config *oa_config,
 			   struct i915_active *active)
 {
+	u32 ctx_oactxctrl = stream->perf->ctx_oactxctrl_offset;
 	/* The MMIO offsets for Flex EU registers aren't contiguous */
 	const u32 ctx_flexeu0 = stream->perf->ctx_flexeu0_offset;
 #define ctx_flexeuN(N) (ctx_flexeu0 + 2 * (N) + 1)
@@ -2576,7 +2658,7 @@ lrc_configure_all_contexts(struct i915_perf_stream *stream,
 		},
 		{
 			GEN8_OACTXCONTROL,
-			stream->perf->ctx_oactxctrl_offset + 1,
+			ctx_oactxctrl + 1,
 		},
 		{ EU_PERF_CNTL0, ctx_flexeuN(0) },
 		{ EU_PERF_CNTL1, ctx_flexeuN(1) },
@@ -4545,6 +4627,37 @@ static void oa_init_supported_formats(struct i915_perf *perf)
 	}
 }
 
+static void i915_perf_init_info(struct drm_i915_private *i915)
+{
+	struct i915_perf *perf = &i915->perf;
+
+	switch (GRAPHICS_VER(i915)) {
+	case 8:
+		perf->ctx_oactxctrl_offset = 0x120;
+		perf->ctx_flexeu0_offset = 0x2ce;
+		perf->gen8_valid_ctx_bit = BIT(25);
+		break;
+	case 9:
+		perf->ctx_oactxctrl_offset = 0x128;
+		perf->ctx_flexeu0_offset = 0x3de;
+		perf->gen8_valid_ctx_bit = BIT(16);
+		break;
+	case 11:
+		perf->ctx_oactxctrl_offset = 0x124;
+		perf->ctx_flexeu0_offset = 0x78e;
+		perf->gen8_valid_ctx_bit = BIT(16);
+		break;
+	case 12:
+		/*
+		 * Calculate offset at runtime in oa_pin_context for gen12 and
+		 * cache the value in perf->ctx_oactxctrl_offset.
+		 */
+		break;
+	default:
+		MISSING_CASE(GRAPHICS_VER(i915));
+	}
+}
+
 /**
  * i915_perf_init - initialize i915-perf state on module bind
  * @i915: i915 device instance
@@ -4583,6 +4696,7 @@ void i915_perf_init(struct drm_i915_private *i915)
 		 * execlist mode by default.
 		 */
 		perf->ops.read = gen8_oa_read;
+		i915_perf_init_info(i915);
 
 		if (IS_GRAPHICS_VER(i915, 8, 9)) {
 			perf->ops.is_valid_b_counter_reg =
@@ -4602,18 +4716,6 @@ void i915_perf_init(struct drm_i915_private *i915)
 			perf->ops.enable_metric_set = gen8_enable_metric_set;
 			perf->ops.disable_metric_set = gen8_disable_metric_set;
 			perf->ops.oa_hw_tail_read = gen8_oa_hw_tail_read;
-
-			if (GRAPHICS_VER(i915) == 8) {
-				perf->ctx_oactxctrl_offset = 0x120;
-				perf->ctx_flexeu0_offset = 0x2ce;
-
-				perf->gen8_valid_ctx_bit = BIT(25);
-			} else {
-				perf->ctx_oactxctrl_offset = 0x128;
-				perf->ctx_flexeu0_offset = 0x3de;
-
-				perf->gen8_valid_ctx_bit = BIT(16);
-			}
 		} else if (GRAPHICS_VER(i915) == 11) {
 			perf->ops.is_valid_b_counter_reg =
 				gen7_is_valid_b_counter_addr;
@@ -4627,11 +4729,6 @@ void i915_perf_init(struct drm_i915_private *i915)
 			perf->ops.enable_metric_set = gen8_enable_metric_set;
 			perf->ops.disable_metric_set = gen11_disable_metric_set;
 			perf->ops.oa_hw_tail_read = gen8_oa_hw_tail_read;
-
-			perf->ctx_oactxctrl_offset = 0x124;
-			perf->ctx_flexeu0_offset = 0x78e;
-
-			perf->gen8_valid_ctx_bit = BIT(16);
 		} else if (GRAPHICS_VER(i915) == 12) {
 			perf->ops.is_valid_b_counter_reg =
 				gen12_is_valid_b_counter_addr;
@@ -4645,9 +4742,6 @@ void i915_perf_init(struct drm_i915_private *i915)
 			perf->ops.enable_metric_set = gen12_enable_metric_set;
 			perf->ops.disable_metric_set = gen12_disable_metric_set;
 			perf->ops.oa_hw_tail_read = gen12_oa_hw_tail_read;
-
-			perf->ctx_flexeu0_offset = 0;
-			perf->ctx_oactxctrl_offset = 0x144;
 		}
 	}
 
drivers/gpu/drm/i915/i915_perf_oa_regs.h +1 -1
index f31c9f13a9fc..0ef3562ff4aa 100644
@@ -97,7 +97,7 @@
 #define  GEN12_OAR_OACONTROL_COUNTER_FORMAT_SHIFT 1
 #define  GEN12_OAR_OACONTROL_COUNTER_ENABLE       (1 << 0)
 
-#define GEN12_OACTXCONTROL _MMIO(0x2360)
+#define GEN12_OACTXCONTROL(base) _MMIO((base) + 0x360)
 #define GEN12_OAR_OASTATUS _MMIO(0x2968)
 
 /* Gen12 OAG unit */
Signature · 1 line
Jani Nikula, Intel Open Source Graphics Center
#8 lore raw

Re: [Intel-gfx] [PATCH v3 04/16] drm/i915/perf: Determine gen12 oa ctx offset at runtime

Quoted 44 lines · Jani Nikula
Quoted 43 lines · Umesh Nerlige Ramappa

Some SKUs of same gen12 platform may have different oactxctrl offsets. For gen12, determine oactxctrl offsets at runtime.

v2: (Lionel)
- Move MI definitions to intel_gpu_commands.h
- Ensure __find_reg_in_lri does read past context image size

v3: (Ashutosh)
- Drop unnecessary use of double underscores
- fix find_reg_in_lri
- Return error if oa context offset is U32_MAX
- Error out if oa_ctx_ctrl_offset does not find offset

---
 drivers/gpu/drm/i915/gt/intel_gpu_commands.h |   4 +
 drivers/gpu/drm/i915/i915_perf.c             | 154 +++++++++++++++----
 drivers/gpu/drm/i915/i915_perf_oa_regs.h     |   2 +-
 3 files changed, 129 insertions(+), 31 deletions(-)
drivers/gpu/drm/i915/gt/intel_gpu_commands.h +4 -0
index d4e9702d3c8e..f50ea92910d9 100644
@@ -187,6 +187,10 @@
 #define   MI_BATCH_RESOURCE_STREAMER REG_BIT(10)
 #define   MI_BATCH_PREDICATE         REG_BIT(15) /* HSW+ on RCS only*/
 
+#define MI_OPCODE(x)		(((x) >> 23) & 0x3f)
+#define IS_MI_LRI_CMD(x)	(MI_OPCODE(x) == MI_OPCODE(MI_INSTR(0x22, 0)))
+#define MI_LRI_LEN(x)		(((x) & 0xff) + 1)
+
 /*
  * 3D instructions used by the kernel
  */
drivers/gpu/drm/i915/i915_perf.c +0 -0
index cd57b5836386..b292aa39633e 100644
@@ -1358,6 +1358,68 @@ static int gen12_get_render_context_id(struct i915_perf_stream *stream)
 	return 0;
 }
 
Quoted 8 lines

drivers/gpu/drm/i915/i915_perf.c static int gen12_get_render_context_id(struct i915_perf_stream *stream) · line 1365

Quoted 5 lines
drivers/gpu/drm/i915/i915_perf.c +5 -0
+#define valid_oactxctrl_offset(x) ((x) && (x) != U32_MAX)
+static bool oa_find_reg_in_lri(u32 *state, u32 reg, u32 *offset, u32 end)
+{
+	u32 idx = *offset;
+	u32 len = min(MI_LRI_LEN(state[idx]) + idx, end);

I don't understand any of this stuff, but why are the offset, index and length u32 instead of just, say, int?

I can use int, but the ce->engine->context_size is u32 and we are parsing the context image here, so I have just used the same type for offset, index, length.

Any guideline/recommendation to choose int over u32?

Quoted 8 lines

drivers/gpu/drm/i915/i915_perf.c static int gen12_get_render_context_id(struct i915_perf_stream *stream) · line 1369

Quoted 5 lines
drivers/gpu/drm/i915/i915_perf.c +4 -0
+	bool found = false;
+
+	idx++;
+	for (; idx < len; idx += 2) {

I find the initialization of idx and len confusing, and thereby the entire loop too.

Considering that the context image is a collection of MI_LRI commands with each command having this format:

dword 0: MI_LRI
dword 1: reg offset
dword 2: reg value
dword 3: reg offset
dword 4: reg value
...

oa_context_image_offset() and oa_find_reg_in_lri() are parsing this context image to look for a specific reg_offset. Once the offset is known, the OA code programs the reg_value for the context.

Thanks,
Umesh

Quoted 243 lines

BR,
Jani.

Quoted 235 lines
drivers/gpu/drm/i915/i915_perf.c +115 -30
+		if (state[idx] == reg) {
+			found = true;
+			break;
+		}
+	}
+
+	*offset = idx;
+	return found;
+}
+
+static u32 oa_context_image_offset(struct intel_context *ce, u32 reg)
+{
+	u32 offset, len = (ce->engine->context_size - PAGE_SIZE) / 4;
+	u32 *state = ce->lrc_reg_state;
+
+	for (offset = 0; offset < len; ) {
+		if (IS_MI_LRI_CMD(state[offset])) {
+			if (oa_find_reg_in_lri(state, reg, &offset, len))
+				break;
+		} else {
+			offset++;
+		}
+	}
+
+	return offset < len ? offset : U32_MAX;
+}
+
+static int set_oa_ctx_ctrl_offset(struct intel_context *ce)
+{
+	i915_reg_t reg = GEN12_OACTXCONTROL(ce->engine->mmio_base);
+	struct i915_perf *perf = &ce->engine->i915->perf;
+	u32 offset = perf->ctx_oactxctrl_offset;
+
+	/* Do this only once. Failure is stored as offset of U32_MAX */
+	if (offset)
+		goto exit;
+
+	offset = oa_context_image_offset(ce, i915_mmio_reg_offset(reg));
+	perf->ctx_oactxctrl_offset = offset;
+
+	drm_dbg(&ce->engine->i915->drm,
+		"%s oa ctx control at 0x%08x dword offset\n",
+		ce->engine->name, offset);
+
+exit:
+	return valid_oactxctrl_offset(offset) ? 0 : -ENODEV;
+}
+
+static bool engine_supports_mi_query(struct intel_engine_cs *engine)
+{
+	return engine->class == RENDER_CLASS;
+}
+
 /**
  * oa_get_render_ctx_id - determine and hold ctx hw id
  * @stream: An i915-perf stream opened for OA metrics
@@ -1377,6 +1439,21 @@ static int oa_get_render_ctx_id(struct i915_perf_stream *stream)
 	if (IS_ERR(ce))
 		return PTR_ERR(ce);
 
+	if (engine_supports_mi_query(stream->engine)) {
+		/*
+		 * We are enabling perf query here. If we don't find the context
+		 * offset here, just return an error.
+		 */
+		ret = set_oa_ctx_ctrl_offset(ce);
+		if (ret) {
+			intel_context_unpin(ce);
+			drm_err(&stream->perf->i915->drm,
+				"Enabling perf query failed for %s\n",
+				stream->engine->name);
+			return ret;
+		}
+	}
+
 	switch (GRAPHICS_VER(ce->engine->i915)) {
 	case 7: {
 		/*
@@ -2408,10 +2485,11 @@ static int gen12_configure_oar_context(struct i915_perf_stream *stream,
 	int err;
 	struct intel_context *ce = stream->pinned_ctx;
 	u32 format = stream->oa_buffer.format;
+	u32 offset = stream->perf->ctx_oactxctrl_offset;
 	struct flex regs_context[] = {
 		{
 			GEN8_OACTXCONTROL,
-			stream->perf->ctx_oactxctrl_offset + 1,
+			offset + 1,
 			active ? GEN8_OA_COUNTER_RESUME : 0,
 		},
 	};
@@ -2436,15 +2514,18 @@ static int gen12_configure_oar_context(struct i915_perf_stream *stream,
 		},
 	};
 
-	/* Modify the context image of pinned context with regs_context*/
-	err = intel_context_lock_pinned(ce);
-	if (err)
-		return err;
+	/* Modify the context image of pinned context with regs_context */
+	if (valid_oactxctrl_offset(offset)) {
+		err = intel_context_lock_pinned(ce);
+		if (err)
+			return err;
 
-	err = gen8_modify_context(ce, regs_context, ARRAY_SIZE(regs_context));
-	intel_context_unlock_pinned(ce);
-	if (err)
-		return err;
+		err = gen8_modify_context(ce, regs_context,
+					  ARRAY_SIZE(regs_context));
+		intel_context_unlock_pinned(ce);
+		if (err)
+			return err;
+	}
 
 	/* Apply regs_lri using LRI with pinned context */
 	return gen8_modify_self(ce, regs_lri, ARRAY_SIZE(regs_lri), active);
@@ -2566,6 +2647,7 @@ lrc_configure_all_contexts(struct i915_perf_stream *stream,
 			   const struct i915_oa_config *oa_config,
 			   struct i915_active *active)
 {
+	u32 ctx_oactxctrl = stream->perf->ctx_oactxctrl_offset;
 	/* The MMIO offsets for Flex EU registers aren't contiguous */
 	const u32 ctx_flexeu0 = stream->perf->ctx_flexeu0_offset;
 #define ctx_flexeuN(N) (ctx_flexeu0 + 2 * (N) + 1)
@@ -2576,7 +2658,7 @@ lrc_configure_all_contexts(struct i915_perf_stream *stream,
 		},
 		{
 			GEN8_OACTXCONTROL,
-			stream->perf->ctx_oactxctrl_offset + 1,
+			ctx_oactxctrl + 1,
 		},
 		{ EU_PERF_CNTL0, ctx_flexeuN(0) },
 		{ EU_PERF_CNTL1, ctx_flexeuN(1) },
@@ -4545,6 +4627,37 @@ static void oa_init_supported_formats(struct i915_perf *perf)
 	}
 }
 
+static void i915_perf_init_info(struct drm_i915_private *i915)
+{
+	struct i915_perf *perf = &i915->perf;
+
+	switch (GRAPHICS_VER(i915)) {
+	case 8:
+		perf->ctx_oactxctrl_offset = 0x120;
+		perf->ctx_flexeu0_offset = 0x2ce;
+		perf->gen8_valid_ctx_bit = BIT(25);
+		break;
+	case 9:
+		perf->ctx_oactxctrl_offset = 0x128;
+		perf->ctx_flexeu0_offset = 0x3de;
+		perf->gen8_valid_ctx_bit = BIT(16);
+		break;
+	case 11:
+		perf->ctx_oactxctrl_offset = 0x124;
+		perf->ctx_flexeu0_offset = 0x78e;
+		perf->gen8_valid_ctx_bit = BIT(16);
+		break;
+	case 12:
+		/*
+		 * Calculate offset at runtime in oa_pin_context for gen12 and
+		 * cache the value in perf->ctx_oactxctrl_offset.
+		 */
+		break;
+	default:
+		MISSING_CASE(GRAPHICS_VER(i915));
+	}
+}
+
 /**
  * i915_perf_init - initialize i915-perf state on module bind
  * @i915: i915 device instance
@@ -4583,6 +4696,7 @@ void i915_perf_init(struct drm_i915_private *i915)
 		 * execlist mode by default.
 		 */
 		perf->ops.read = gen8_oa_read;
+		i915_perf_init_info(i915);
 
 		if (IS_GRAPHICS_VER(i915, 8, 9)) {
 			perf->ops.is_valid_b_counter_reg =
@@ -4602,18 +4716,6 @@ void i915_perf_init(struct drm_i915_private *i915)
 			perf->ops.enable_metric_set = gen8_enable_metric_set;
 			perf->ops.disable_metric_set = gen8_disable_metric_set;
 			perf->ops.oa_hw_tail_read = gen8_oa_hw_tail_read;
-
-			if (GRAPHICS_VER(i915) == 8) {
-				perf->ctx_oactxctrl_offset = 0x120;
-				perf->ctx_flexeu0_offset = 0x2ce;
-
-				perf->gen8_valid_ctx_bit = BIT(25);
-			} else {
-				perf->ctx_oactxctrl_offset = 0x128;
-				perf->ctx_flexeu0_offset = 0x3de;
-
-				perf->gen8_valid_ctx_bit = BIT(16);
-			}
 		} else if (GRAPHICS_VER(i915) == 11) {
 			perf->ops.is_valid_b_counter_reg =
 				gen7_is_valid_b_counter_addr;
@@ -4627,11 +4729,6 @@ void i915_perf_init(struct drm_i915_private *i915)
 			perf->ops.enable_metric_set = gen8_enable_metric_set;
 			perf->ops.disable_metric_set = gen11_disable_metric_set;
 			perf->ops.oa_hw_tail_read = gen8_oa_hw_tail_read;
-
-			perf->ctx_oactxctrl_offset = 0x124;
-			perf->ctx_flexeu0_offset = 0x78e;
-
-			perf->gen8_valid_ctx_bit = BIT(16);
 		} else if (GRAPHICS_VER(i915) == 12) {
 			perf->ops.is_valid_b_counter_reg =
 				gen12_is_valid_b_counter_addr;
@@ -4645,9 +4742,6 @@ void i915_perf_init(struct drm_i915_private *i915)
 			perf->ops.enable_metric_set = gen12_enable_metric_set;
 			perf->ops.disable_metric_set = gen12_disable_metric_set;
 			perf->ops.oa_hw_tail_read = gen12_oa_hw_tail_read;
-
-			perf->ctx_flexeu0_offset = 0;
-			perf->ctx_oactxctrl_offset = 0x144;
 		}
 	}
 
drivers/gpu/drm/i915/i915_perf_oa_regs.h +1 -1
index f31c9f13a9fc..0ef3562ff4aa 100644
@@ -97,7 +97,7 @@
 #define  GEN12_OAR_OACONTROL_COUNTER_FORMAT_SHIFT 1
 #define  GEN12_OAR_OACONTROL_COUNTER_ENABLE       (1 << 0)
 
-#define GEN12_OACTXCONTROL _MMIO(0x2360)
+#define GEN12_OACTXCONTROL(base) _MMIO((base) + 0x360)
 #define GEN12_OAR_OASTATUS _MMIO(0x2968)
 
 /* Gen12 OAG unit */
Signature · 1 line
Jani Nikula, Intel Open Source Graphics Center
maintainer
#9 lore raw

Re: [Intel-gfx] [PATCH v3 04/16] drm/i915/perf: Determine gen12 oa ctx offset at runtime

Quoted 54 lines · Umesh Nerlige Ramappa
Quoted 44 lines · Jani Nikula
Quoted 43 lines · Umesh Nerlige Ramappa

Some SKUs of same gen12 platform may have different oactxctrl offsets. For gen12, determine oactxctrl offsets at runtime.

v2: (Lionel)
- Move MI definitions to intel_gpu_commands.h
- Ensure __find_reg_in_lri does read past context image size

v3: (Ashutosh)
- Drop unnecessary use of double underscores
- fix find_reg_in_lri
- Return error if oa context offset is U32_MAX
- Error out if oa_ctx_ctrl_offset does not find offset

---
 drivers/gpu/drm/i915/gt/intel_gpu_commands.h |   4 +
 drivers/gpu/drm/i915/i915_perf.c             | 154 +++++++++++++++----
 drivers/gpu/drm/i915/i915_perf_oa_regs.h     |   2 +-
 3 files changed, 129 insertions(+), 31 deletions(-)
drivers/gpu/drm/i915/gt/intel_gpu_commands.h +4 -0
index d4e9702d3c8e..f50ea92910d9 100644
@@ -187,6 +187,10 @@
 #define   MI_BATCH_RESOURCE_STREAMER REG_BIT(10)
 #define   MI_BATCH_PREDICATE         REG_BIT(15) /* HSW+ on RCS only*/
 
+#define MI_OPCODE(x)		(((x) >> 23) & 0x3f)
+#define IS_MI_LRI_CMD(x)	(MI_OPCODE(x) == MI_OPCODE(MI_INSTR(0x22, 0)))
+#define MI_LRI_LEN(x)		(((x) & 0xff) + 1)
+
 /*
  * 3D instructions used by the kernel
  */
drivers/gpu/drm/i915/i915_perf.c +0 -0
index cd57b5836386..b292aa39633e 100644
@@ -1358,6 +1358,68 @@ static int gen12_get_render_context_id(struct i915_perf_stream *stream)
 	return 0;
 }
 
Quoted 8 lines

drivers/gpu/drm/i915/i915_perf.c static int gen12_get_render_context_id(struct i915_perf_stream *stream) · line 1365

Quoted 5 lines
drivers/gpu/drm/i915/i915_perf.c +5 -0
+#define valid_oactxctrl_offset(x) ((x) && (x) != U32_MAX)
+static bool oa_find_reg_in_lri(u32 *state, u32 reg, u32 *offset, u32 end)
+{
+	u32 idx = *offset;
+	u32 len = min(MI_LRI_LEN(state[idx]) + idx, end);

I don't understand any of this stuff, but why are the offset, index and length u32 instead of just, say, int?

Quoted 5 lines

I can use int, but the ce->engine->context_size is u32 and we are parsing the context image here, so I have just used the same type for offset, index, length.

Any guideline/recommendation to choose int over u32?

If it's an index in general, looping, regular computation, just plain old int will often do. Use u32 and friends when you actually need a specific size to map to hardware or registers etc.

I guess it's not so clear cut here, and mixing types not necessarily a bright idea here either.

BR,
Jani.

Quoted 270 lines
Quoted 9 lines

drivers/gpu/drm/i915/i915_perf.c static int gen12_get_render_context_id(struct i915_perf_stream *stream) · line 1369

Quoted 6 lines
drivers/gpu/drm/i915/i915_perf.c +4 -0
+	bool found = false;
+
+	idx++;
+	for (; idx < len; idx += 2) {

I find the initialization of idx and len confusing, and thereby the entire loop too.

Considering that the context image is a collection of MI_LRI commands with each command having this format:

dword 0: MI_LRI
dword 1: reg offset
dword 2: reg value
dword 3: reg offset
dword 4: reg value
...

oa_context_image_offset() and oa_find_reg_in_lri() are parsing this context image to look for a specific reg_offset. Once the offset is known, the OA code programs the reg_value for the context.

Thanks,
Umesh

Quoted 243 lines

BR,
Jani.

Quoted 235 lines
drivers/gpu/drm/i915/i915_perf.c +115 -30
+		if (state[idx] == reg) {
+			found = true;
+			break;
+		}
+	}
+
+	*offset = idx;
+	return found;
+}
+
+static u32 oa_context_image_offset(struct intel_context *ce, u32 reg)
+{
+	u32 offset, len = (ce->engine->context_size - PAGE_SIZE) / 4;
+	u32 *state = ce->lrc_reg_state;
+
+	for (offset = 0; offset < len; ) {
+		if (IS_MI_LRI_CMD(state[offset])) {
+			if (oa_find_reg_in_lri(state, reg, &offset, len))
+				break;
+		} else {
+			offset++;
+		}
+	}
+
+	return offset < len ? offset : U32_MAX;
+}
+
+static int set_oa_ctx_ctrl_offset(struct intel_context *ce)
+{
+	i915_reg_t reg = GEN12_OACTXCONTROL(ce->engine->mmio_base);
+	struct i915_perf *perf = &ce->engine->i915->perf;
+	u32 offset = perf->ctx_oactxctrl_offset;
+
+	/* Do this only once. Failure is stored as offset of U32_MAX */
+	if (offset)
+		goto exit;
+
+	offset = oa_context_image_offset(ce, i915_mmio_reg_offset(reg));
+	perf->ctx_oactxctrl_offset = offset;
+
+	drm_dbg(&ce->engine->i915->drm,
+		"%s oa ctx control at 0x%08x dword offset\n",
+		ce->engine->name, offset);
+
+exit:
+	return valid_oactxctrl_offset(offset) ? 0 : -ENODEV;
+}
+
+static bool engine_supports_mi_query(struct intel_engine_cs *engine)
+{
+	return engine->class == RENDER_CLASS;
+}
+
 /**
  * oa_get_render_ctx_id - determine and hold ctx hw id
  * @stream: An i915-perf stream opened for OA metrics
@@ -1377,6 +1439,21 @@ static int oa_get_render_ctx_id(struct i915_perf_stream *stream)
 	if (IS_ERR(ce))
 		return PTR_ERR(ce);
 
+	if (engine_supports_mi_query(stream->engine)) {
+		/*
+		 * We are enabling perf query here. If we don't find the context
+		 * offset here, just return an error.
+		 */
+		ret = set_oa_ctx_ctrl_offset(ce);
+		if (ret) {
+			intel_context_unpin(ce);
+			drm_err(&stream->perf->i915->drm,
+				"Enabling perf query failed for %s\n",
+				stream->engine->name);
+			return ret;
+		}
+	}
+
 	switch (GRAPHICS_VER(ce->engine->i915)) {
 	case 7: {
 		/*
@@ -2408,10 +2485,11 @@ static int gen12_configure_oar_context(struct i915_perf_stream *stream,
 	int err;
 	struct intel_context *ce = stream->pinned_ctx;
 	u32 format = stream->oa_buffer.format;
+	u32 offset = stream->perf->ctx_oactxctrl_offset;
 	struct flex regs_context[] = {
 		{
 			GEN8_OACTXCONTROL,
-			stream->perf->ctx_oactxctrl_offset + 1,
+			offset + 1,
 			active ? GEN8_OA_COUNTER_RESUME : 0,
 		},
 	};
@@ -2436,15 +2514,18 @@ static int gen12_configure_oar_context(struct i915_perf_stream *stream,
 		},
 	};
 
-	/* Modify the context image of pinned context with regs_context*/
-	err = intel_context_lock_pinned(ce);
-	if (err)
-		return err;
+	/* Modify the context image of pinned context with regs_context */
+	if (valid_oactxctrl_offset(offset)) {
+		err = intel_context_lock_pinned(ce);
+		if (err)
+			return err;
 
-	err = gen8_modify_context(ce, regs_context, ARRAY_SIZE(regs_context));
-	intel_context_unlock_pinned(ce);
-	if (err)
-		return err;
+		err = gen8_modify_context(ce, regs_context,
+					  ARRAY_SIZE(regs_context));
+		intel_context_unlock_pinned(ce);
+		if (err)
+			return err;
+	}
 
 	/* Apply regs_lri using LRI with pinned context */
 	return gen8_modify_self(ce, regs_lri, ARRAY_SIZE(regs_lri), active);
@@ -2566,6 +2647,7 @@ lrc_configure_all_contexts(struct i915_perf_stream *stream,
 			   const struct i915_oa_config *oa_config,
 			   struct i915_active *active)
 {
+	u32 ctx_oactxctrl = stream->perf->ctx_oactxctrl_offset;
 	/* The MMIO offsets for Flex EU registers aren't contiguous */
 	const u32 ctx_flexeu0 = stream->perf->ctx_flexeu0_offset;
 #define ctx_flexeuN(N) (ctx_flexeu0 + 2 * (N) + 1)
@@ -2576,7 +2658,7 @@ lrc_configure_all_contexts(struct i915_perf_stream *stream,
 		},
 		{
 			GEN8_OACTXCONTROL,
-			stream->perf->ctx_oactxctrl_offset + 1,
+			ctx_oactxctrl + 1,
 		},
 		{ EU_PERF_CNTL0, ctx_flexeuN(0) },
 		{ EU_PERF_CNTL1, ctx_flexeuN(1) },
@@ -4545,6 +4627,37 @@ static void oa_init_supported_formats(struct i915_perf *perf)
 	}
 }
 
+static void i915_perf_init_info(struct drm_i915_private *i915)
+{
+	struct i915_perf *perf = &i915->perf;
+
+	switch (GRAPHICS_VER(i915)) {
+	case 8:
+		perf->ctx_oactxctrl_offset = 0x120;
+		perf->ctx_flexeu0_offset = 0x2ce;
+		perf->gen8_valid_ctx_bit = BIT(25);
+		break;
+	case 9:
+		perf->ctx_oactxctrl_offset = 0x128;
+		perf->ctx_flexeu0_offset = 0x3de;
+		perf->gen8_valid_ctx_bit = BIT(16);
+		break;
+	case 11:
+		perf->ctx_oactxctrl_offset = 0x124;
+		perf->ctx_flexeu0_offset = 0x78e;
+		perf->gen8_valid_ctx_bit = BIT(16);
+		break;
+	case 12:
+		/*
+		 * Calculate offset at runtime in oa_pin_context for gen12 and
+		 * cache the value in perf->ctx_oactxctrl_offset.
+		 */
+		break;
+	default:
+		MISSING_CASE(GRAPHICS_VER(i915));
+	}
+}
+
 /**
  * i915_perf_init - initialize i915-perf state on module bind
  * @i915: i915 device instance
@@ -4583,6 +4696,7 @@ void i915_perf_init(struct drm_i915_private *i915)
 		 * execlist mode by default.
 		 */
 		perf->ops.read = gen8_oa_read;
+		i915_perf_init_info(i915);
 
 		if (IS_GRAPHICS_VER(i915, 8, 9)) {
 			perf->ops.is_valid_b_counter_reg =
@@ -4602,18 +4716,6 @@ void i915_perf_init(struct drm_i915_private *i915)
 			perf->ops.enable_metric_set = gen8_enable_metric_set;
 			perf->ops.disable_metric_set = gen8_disable_metric_set;
 			perf->ops.oa_hw_tail_read = gen8_oa_hw_tail_read;
-
-			if (GRAPHICS_VER(i915) == 8) {
-				perf->ctx_oactxctrl_offset = 0x120;
-				perf->ctx_flexeu0_offset = 0x2ce;
-
-				perf->gen8_valid_ctx_bit = BIT(25);
-			} else {
-				perf->ctx_oactxctrl_offset = 0x128;
-				perf->ctx_flexeu0_offset = 0x3de;
-
-				perf->gen8_valid_ctx_bit = BIT(16);
-			}
 		} else if (GRAPHICS_VER(i915) == 11) {
 			perf->ops.is_valid_b_counter_reg =
 				gen7_is_valid_b_counter_addr;
@@ -4627,11 +4729,6 @@ void i915_perf_init(struct drm_i915_private *i915)
 			perf->ops.enable_metric_set = gen8_enable_metric_set;
 			perf->ops.disable_metric_set = gen11_disable_metric_set;
 			perf->ops.oa_hw_tail_read = gen8_oa_hw_tail_read;
-
-			perf->ctx_oactxctrl_offset = 0x124;
-			perf->ctx_flexeu0_offset = 0x78e;
-
-			perf->gen8_valid_ctx_bit = BIT(16);
 		} else if (GRAPHICS_VER(i915) == 12) {
 			perf->ops.is_valid_b_counter_reg =
 				gen12_is_valid_b_counter_addr;
@@ -4645,9 +4742,6 @@ void i915_perf_init(struct drm_i915_private *i915)
 			perf->ops.enable_metric_set = gen12_enable_metric_set;
 			perf->ops.disable_metric_set = gen12_disable_metric_set;
 			perf->ops.oa_hw_tail_read = gen12_oa_hw_tail_read;
-
-			perf->ctx_flexeu0_offset = 0;
-			perf->ctx_oactxctrl_offset = 0x144;
 		}
 	}
 
drivers/gpu/drm/i915/i915_perf_oa_regs.h +1 -1
index f31c9f13a9fc..0ef3562ff4aa 100644
@@ -97,7 +97,7 @@
 #define  GEN12_OAR_OACONTROL_COUNTER_FORMAT_SHIFT 1
 #define  GEN12_OAR_OACONTROL_COUNTER_ENABLE       (1 << 0)
 
-#define GEN12_OACTXCONTROL _MMIO(0x2360)
+#define GEN12_OACTXCONTROL(base) _MMIO((base) + 0x360)
 #define GEN12_OAR_OASTATUS _MMIO(0x2968)
 
 /* Gen12 OAG unit */
Signature · 1 line
Jani Nikula, Intel Open Source Graphics Center
Signature · 1 line
Jani Nikula, Intel Open Source Graphics Center
#10 lore raw

PATCH v3 5/16 drm/i915/perf: Enable bytes per clock reporting in OA

In reply to Umesh Nerlige Ramappa

XEHPSDV and DG2 provide a way to configure bytes per clock vs commands per clock reporting. Enable bytes per clock setting on enabling OA.

Bspec: 51762
Bspec: 52201

v2:
- Fix commit msg (Ashutosh)
- Fix checkpatch issues

v3:
- s/commands/bytes/ in code comment and commmit msg

---
 drivers/gpu/drm/i915/i915_drv.h          |  3 +++
 drivers/gpu/drm/i915/i915_pci.c          |  1 +
 drivers/gpu/drm/i915/i915_perf.c         | 20 ++++++++++++++++++++
 drivers/gpu/drm/i915/i915_perf_oa_regs.h |  4 ++++
 drivers/gpu/drm/i915/intel_device_info.h |  1 +
 5 files changed, 29 insertions(+)
drivers/gpu/drm/i915/i915_drv.h +3 -0
index 9f9372931fd2..ccd54ff54002 100644
@@ -902,6 +902,9 @@ IS_SUBPLATFORM(const struct drm_i915_private *i915,
 #define HAS_RUNTIME_PM(dev_priv) (INTEL_INFO(dev_priv)->has_runtime_pm)
 #define HAS_64BIT_RELOC(dev_priv) (INTEL_INFO(dev_priv)->has_64bit_reloc)
 
+#define HAS_OA_BPC_REPORTING(dev_priv) \
+	(INTEL_INFO(dev_priv)->has_oa_bpc_reporting)
+
 /*
  * Set this flag, when platform requires 64K GTT page sizes or larger for
  * device local memory access.
drivers/gpu/drm/i915/i915_pci.c +1 -0
index 77e7df21f539..6b25e4cb6221 100644
@@ -1022,6 +1022,7 @@ static const struct intel_device_info adl_p_info = {
 	.has_logical_ring_contexts = 1, \
 	.has_logical_ring_elsq = 1, \
 	.has_mslice_steering = 1, \
+	.has_oa_bpc_reporting = 1, \
 	.has_rc6 = 1, \
 	.has_reset_engine = 1, \
 	.has_rps = 1, \
drivers/gpu/drm/i915/i915_perf.c +20 -0
index b292aa39633e..a22f18554ab9 100644
@@ -2746,10 +2746,12 @@ static int
 gen12_enable_metric_set(struct i915_perf_stream *stream,
 			struct i915_active *active)
 {
+	struct drm_i915_private *i915 = stream->perf->i915;
 	struct intel_uncore *uncore = stream->uncore;
 	struct i915_oa_config *oa_config = stream->oa_config;
 	bool periodic = stream->periodic;
 	u32 period_exponent = stream->period_exponent;
+	u32 sqcnt1;
 	int ret;
 
 	intel_uncore_write(uncore, GEN12_OAG_OA_DEBUG,
@@ -2768,6 +2770,16 @@ gen12_enable_metric_set(struct i915_perf_stream *stream,
 			    (period_exponent << GEN12_OAG_OAGLBCTXCTRL_TIMER_PERIOD_SHIFT))
 			    : 0);
 
+	/*
+	 * Initialize Super Queue Internal Cnt Register
+	 * Set PMON Enable in order to collect valid metrics.
+	 * Enable byets per clock reporting in OA for XEHPSDV onward.
+	 */
+	sqcnt1 = GEN12_SQCNT1_PMON_ENABLE |
+		 (HAS_OA_BPC_REPORTING(i915) ? GEN12_SQCNT1_OABPC : 0);
+
+	intel_uncore_rmw(uncore, GEN12_SQCNT1, 0, sqcnt1);
+
 	/*
 	 * Update all contexts prior writing the mux configurations as we need
 	 * to make sure all slices/subslices are ON before writing to NOA
@@ -2817,6 +2829,8 @@ static void gen11_disable_metric_set(struct i915_perf_stream *stream)
 static void gen12_disable_metric_set(struct i915_perf_stream *stream)
 {
 	struct intel_uncore *uncore = stream->uncore;
+	struct drm_i915_private *i915 = stream->perf->i915;
+	u32 sqcnt1;
 
 	/* Reset all contexts' slices/subslices configurations. */
 	gen12_configure_all_contexts(stream, NULL, NULL);
@@ -2827,6 +2841,12 @@ static void gen12_disable_metric_set(struct i915_perf_stream *stream)
 
 	/* Make sure we disable noa to save power. */
 	intel_uncore_rmw(uncore, RPM_CONFIG1, GEN10_GT_NOA_ENABLE, 0);
+
+	sqcnt1 = GEN12_SQCNT1_PMON_ENABLE |
+		 (HAS_OA_BPC_REPORTING(i915) ? GEN12_SQCNT1_OABPC : 0);
+
+	/* Reset PMON Enable to save power. */
+	intel_uncore_rmw(uncore, GEN12_SQCNT1, sqcnt1, 0);
 }
 
 static void gen7_oa_enable(struct i915_perf_stream *stream)
drivers/gpu/drm/i915/i915_perf_oa_regs.h +4 -0
index 0ef3562ff4aa..381d94101610 100644
@@ -134,4 +134,8 @@
 #define GDT_CHICKEN_BITS    _MMIO(0x9840)
 #define   GT_NOA_ENABLE	    0x00000080
 
+#define GEN12_SQCNT1				_MMIO(0x8718)
+#define   GEN12_SQCNT1_PMON_ENABLE		REG_BIT(30)
+#define   GEN12_SQCNT1_OABPC			REG_BIT(29)
+
 #endif /* __INTEL_PERF_OA_REGS__ */
drivers/gpu/drm/i915/intel_device_info.h +1 -0
index 09b18910d3ab..1f7a842cd408 100644
@@ -164,6 +164,7 @@ enum intel_ppgtt_type {
 	func(has_logical_ring_elsq); \
 	func(has_media_ratio_mode); \
 	func(has_mslice_steering); \
+	func(has_oa_bpc_reporting); \
 	func(has_one_eu_per_fuse_bit); \
 	func(has_pxp); \
 	func(has_rc6); \
Signature · 1 line
2.25.1
#11 lore raw

PATCH v3 6/16 drm/i915/perf: Simply use stream->ctx

In reply to Umesh Nerlige Ramappa

Earlier code used exclusive_stream to check for user passed context. Simplify this by accessing stream->ctx.

---
 drivers/gpu/drm/i915/i915_perf.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/gpu/drm/i915/i915_perf.c +2 -2
index a22f18554ab9..b60846f5d33c 100644
@@ -777,7 +777,7 @@ static int gen8_append_oa_reports(struct i915_perf_stream *stream,
 		 * switches since it's not-uncommon for periodic samples to
 		 * identify a switch before any 'context switch' report.
 		 */
-		if (!stream->perf->exclusive_stream->ctx ||
+		if (!stream->ctx ||
 		    stream->specific_ctx_id == ctx_id ||
 		    stream->oa_buffer.last_ctx_id == stream->specific_ctx_id ||
 		    reason & OAREPORT_REASON_CTX_SWITCH) {
@@ -786,7 +786,7 @@ static int gen8_append_oa_reports(struct i915_perf_stream *stream,
 			 * While filtering for a single context we avoid
 			 * leaking the IDs of other contexts.
 			 */
-			if (stream->perf->exclusive_stream->ctx &&
+			if (stream->ctx &&
 			    stream->specific_ctx_id != ctx_id) {
 				report32[2] = INVALID_CTX_ID;
 			}
Signature · 1 line
2.25.1
#12 lore raw

PATCH v3 7/16 drm/i915/perf: Move gt-specific data from i915->perf to gt->perf

In reply to Umesh Nerlige Ramappa

Make perf part of gt as the OAG buffer is specific to a gt. The refactor eventually simplifies programming the right OA buffer and the right HW registers when supporting multiple gts.

---
 drivers/gpu/drm/i915/gt/intel_gt_types.h   |  3 +
 drivers/gpu/drm/i915/gt/intel_sseu.c       |  4 +-
 drivers/gpu/drm/i915/i915_perf.c           | 75 +++++++++++++---------
 drivers/gpu/drm/i915/i915_perf_types.h     | 39 +++++------
 drivers/gpu/drm/i915/selftests/i915_perf.c | 16 +++--
 5 files changed, 80 insertions(+), 57 deletions(-)
drivers/gpu/drm/i915/gt/intel_gt_types.h +3 -0
index f19c2de77ff6..9f653a347cad 100644
@@ -20,6 +20,7 @@
 #include "intel_gsc.h"
 
 #include "i915_vma.h"
+#include "i915_perf_types.h"
 #include "intel_engine_types.h"
 #include "intel_gt_buffer_pool_types.h"
 #include "intel_hwconfig.h"
@@ -286,6 +287,8 @@ struct intel_gt {
 	/* sysfs defaults per gt */
 	struct gt_defaults defaults;
 	struct kobject *sysfs_defaults;
+
+	struct i915_perf_gt perf;
 };
 
 struct intel_gt_definition {
drivers/gpu/drm/i915/gt/intel_sseu.c +2 -2
index 66f21c735d54..6c6198a257ac 100644
@@ -677,8 +677,8 @@ u32 intel_sseu_make_rpcs(struct intel_gt *gt,
 	 * If i915/perf is active, we want a stable powergating configuration
 	 * on the system. Use the configuration pinned by i915/perf.
 	 */
-	if (i915->perf.exclusive_stream)
-		req_sseu = &i915->perf.sseu;
+	if (gt->perf.exclusive_stream)
+		req_sseu = &gt->perf.sseu;
 
 	slices = hweight8(req_sseu->slice_mask);
 	subslices = hweight8(req_sseu->subslice_mask);
drivers/gpu/drm/i915/i915_perf.c +44 -31

241 lines not on this page · raw message

drivers/gpu/drm/i915/i915_perf_types.h +20 -19
index 05cb9a335a97..e888bfab478f 100644
@@ -380,6 +380,26 @@ struct i915_oa_ops {
 	u32 (*oa_hw_tail_read)(struct i915_perf_stream *stream);
 };
 
+struct i915_perf_gt {
+	/*
+	 * Lock associated with anything below within this structure.
+	 */
+	struct mutex lock;
+
+	/**
+	 * @sseu: sseu configuration selected to run while perf is active,
+	 * applies to all contexts.
+	 */
+	struct intel_sseu sseu;
+
+	/*
+	 * @exclusive_stream: The stream currently using the OA unit. This is
+	 * sometimes accessed outside a syscall associated to its file
+	 * descriptor.
+	 */
+	struct i915_perf_stream *exclusive_stream;
+};
+
 struct i915_perf {
 	struct drm_i915_private *i915;
 
@@ -397,25 +417,6 @@ struct i915_perf {
 	 */
 	struct idr metrics_idr;
 
-	/*
-	 * Lock associated with anything below within this structure
-	 * except exclusive_stream.
-	 */
-	struct mutex lock;
-
-	/*
-	 * The stream currently using the OA unit. If accessed
-	 * outside a syscall associated to its file
-	 * descriptor.
-	 */
-	struct i915_perf_stream *exclusive_stream;
-
-	/**
-	 * @sseu: sseu configuration selected to run while perf is active,
-	 * applies to all contexts.
-	 */
-	struct intel_sseu sseu;
-
 	/**
 	 * For rate limiting any notifications of spurious
 	 * invalid OA reports
drivers/gpu/drm/i915/selftests/i915_perf.c +11 -5
index 429c6d73b159..24dde5531423 100644
@@ -102,6 +102,12 @@ test_stream(struct i915_perf *perf)
 		I915_OA_FORMAT_A32u40_A4u32_B8_C8 : I915_OA_FORMAT_C4_B8,
 	};
 	struct i915_perf_stream *stream;
+	struct intel_gt *gt;
+
+	if (!props.engine)
+		return NULL;
+
+	gt = props.engine->gt;
 
 	if (!oa_config)
 		return NULL;
@@ -116,12 +122,12 @@ test_stream(struct i915_perf *perf)
 
 	stream->perf = perf;
 
-	mutex_lock(&perf->lock);
+	mutex_lock(&gt->perf.lock);
 	if (i915_oa_stream_init(stream, &param, &props)) {
 		kfree(stream);
 		stream =  NULL;
 	}
-	mutex_unlock(&perf->lock);
+	mutex_unlock(&gt->perf.lock);
 
 	i915_oa_config_put(oa_config);
 
@@ -130,11 +136,11 @@ test_stream(struct i915_perf *perf)
 
 static void stream_destroy(struct i915_perf_stream *stream)
 {
-	struct i915_perf *perf = stream->perf;
+	struct intel_gt *gt = stream->engine->gt;
 
-	mutex_lock(&perf->lock);
+	mutex_lock(&gt->perf.lock);
 	i915_perf_destroy_locked(stream);
-	mutex_unlock(&perf->lock);
+	mutex_unlock(&gt->perf.lock);
 }
 
 static int live_sanitycheck(void *arg)
Signature · 1 line
2.25.1
#13 lore raw

PATCH v3 8/16 drm/i915/perf: Replace gt->perf.lock with stream->lock for file ops

In reply to Umesh Nerlige Ramappa

With multi-gt, user can access multiple OA buffers concurrently. Use stream->lock instead of gt->perf.lock to serialize file operations.

---
 drivers/gpu/drm/i915/i915_perf.c       | 31 ++++++++++++--------------
 drivers/gpu/drm/i915/i915_perf_types.h |  5 +++++
 2 files changed, 19 insertions(+), 17 deletions(-)
drivers/gpu/drm/i915/i915_perf.c +14 -17
index 446c9c97c3a4..3961e9c9e97b 100644
@@ -3229,6 +3229,7 @@ static int i915_oa_stream_init(struct i915_perf_stream *stream,
 	stream->poll_check_timer.function = oa_poll_check_timer_cb;
 	init_waitqueue_head(&stream->poll_wq);
 	spin_lock_init(&stream->oa_buffer.ptr_lock);
+	mutex_init(&stream->lock);
 
 	return 0;
 
@@ -3292,7 +3293,6 @@ static ssize_t i915_perf_read(struct file *file,
 			      loff_t *ppos)
 {
 	struct i915_perf_stream *stream = file->private_data;
-	struct intel_gt *gt = stream->engine->gt;
 	size_t offset = 0;
 	int ret;
 
@@ -3316,14 +3316,14 @@ static ssize_t i915_perf_read(struct file *file,
 			if (ret)
 				return ret;
 
-			mutex_lock(&gt->perf.lock);
+			mutex_lock(&stream->lock);
 			ret = stream->ops->read(stream, buf, count, &offset);
-			mutex_unlock(&gt->perf.lock);
+			mutex_unlock(&stream->lock);
 		} while (!offset && !ret);
 	} else {
-		mutex_lock(&gt->perf.lock);
+		mutex_lock(&stream->lock);
 		ret = stream->ops->read(stream, buf, count, &offset);
-		mutex_unlock(&gt->perf.lock);
+		mutex_unlock(&stream->lock);
 	}
 
 	/* We allow the poll checking to sometimes report false positive EPOLLIN
@@ -3370,9 +3370,6 @@ static enum hrtimer_restart oa_poll_check_timer_cb(struct hrtimer *hrtimer)
  * &i915_perf_stream_ops->poll_wait to call poll_wait() with a wait queue that
  * will be woken for new stream data.
  *
- * Note: The &gt->perf.lock mutex has been taken to serialize
- * with any non-file-operation driver hooks.
- *
  * Returns: any poll events that are ready without sleeping
  */
 static __poll_t i915_perf_poll_locked(struct i915_perf_stream *stream,
@@ -3411,12 +3408,11 @@ static __poll_t i915_perf_poll_locked(struct i915_perf_stream *stream,
 static __poll_t i915_perf_poll(struct file *file, poll_table *wait)
 {
 	struct i915_perf_stream *stream = file->private_data;
-	struct intel_gt *gt = stream->engine->gt;
 	__poll_t ret;
 
-	mutex_lock(&gt->perf.lock);
+	mutex_lock(&stream->lock);
 	ret = i915_perf_poll_locked(stream, file, wait);
-	mutex_unlock(&gt->perf.lock);
+	mutex_unlock(&stream->lock);
 
 	return ret;
 }
@@ -3515,9 +3511,6 @@ static long i915_perf_config_locked(struct i915_perf_stream *stream,
  * @cmd: the ioctl request
  * @arg: the ioctl data
  *
- * Note: The &gt->perf.lock mutex has been taken to serialize
- * with any non-file-operation driver hooks.
- *
  * Returns: zero on success or a negative error code. Returns -EINVAL for
  * an unknown ioctl request.
  */
@@ -3555,12 +3548,11 @@ static long i915_perf_ioctl(struct file *file,
 			    unsigned long arg)
 {
 	struct i915_perf_stream *stream = file->private_data;
-	struct intel_gt *gt = stream->engine->gt;
 	long ret;
 
-	mutex_lock(&gt->perf.lock);
+	mutex_lock(&stream->lock);
 	ret = i915_perf_ioctl_locked(stream, cmd, arg);
-	mutex_unlock(&gt->perf.lock);
+	mutex_unlock(&stream->lock);
 
 	return ret;
 }
@@ -3606,6 +3598,11 @@ static int i915_perf_release(struct inode *inode, struct file *file)
 	struct i915_perf *perf = stream->perf;
 	struct intel_gt *gt = stream->engine->gt;
 
+	/*
+	 * Within this call, we know that the fd is being closed and we have no
+	 * other user of stream->lock. Use the perf lock to destroy the stream
+	 * here.
+	 */
 	mutex_lock(&gt->perf.lock);
 	i915_perf_destroy_locked(stream);
 	mutex_unlock(&gt->perf.lock);
drivers/gpu/drm/i915/i915_perf_types.h +5 -0
index e888bfab478f..dc9bfd8086cf 100644
@@ -146,6 +146,11 @@ struct i915_perf_stream {
 	 */
 	struct intel_engine_cs *engine;
 
+	/*
+	 * Lock associated with operations on stream
+	 */
+	struct mutex lock;
+
 	/**
 	 * @sample_flags: Flags representing the `DRM_I915_PERF_PROP_SAMPLE_*`
 	 * properties given when opening a stream, representing the contents
Signature · 1 line
2.25.1
#14 lore raw

PATCH v3 9/16 drm/i915/perf: Use gt-specific ggtt for OA and noa-wait buffers

In reply to Umesh Nerlige Ramappa

User passes uabi engine class and instance to the perf OA interface. Use gt corresponding to the engine to pin the buffers to the right ggtt.

---
 drivers/gpu/drm/i915/i915_perf.c | 21 +++++++++++++++++++--
 1 file changed, 19 insertions(+), 2 deletions(-)
drivers/gpu/drm/i915/i915_perf.c +19 -2
index 3961e9c9e97b..2f9e18ee0aab 100644
@@ -1750,6 +1750,7 @@ static void gen12_init_oa_buffer(struct i915_perf_stream *stream)
 static int alloc_oa_buffer(struct i915_perf_stream *stream)
 {
 	struct drm_i915_private *i915 = stream->perf->i915;
+	struct intel_gt *gt = stream->engine->gt;
 	struct drm_i915_gem_object *bo;
 	struct i915_vma *vma;
 	int ret;
@@ -1769,11 +1770,22 @@ static int alloc_oa_buffer(struct i915_perf_stream *stream)
 	i915_gem_object_set_cache_coherency(bo, I915_CACHE_LLC);
 
 	/* PreHSW required 512K alignment, HSW requires 16M */
-	vma = i915_gem_object_ggtt_pin(bo, NULL, 0, SZ_16M, 0);
+	vma = i915_vma_instance(bo, &gt->ggtt->vm, NULL);
 	if (IS_ERR(vma)) {
 		ret = PTR_ERR(vma);
 		goto err_unref;
 	}
+
+	/*
+	 * PreHSW required 512K alignment.
+	 * HSW and onwards, align to requested size of OA buffer.
+	 */
+	ret = i915_vma_pin(vma, 0, SZ_16M, PIN_GLOBAL | PIN_HIGH);
+	if (ret) {
+		drm_err(&gt->i915->drm, "Failed to pin OA buffer %d\n", ret);
+		goto err_unref;
+	}
+
 	stream->oa_buffer.vma = vma;
 
 	stream->oa_buffer.vaddr =
@@ -1823,6 +1835,7 @@ static u32 *save_restore_register(struct i915_perf_stream *stream, u32 *cs,
 static int alloc_noa_wait(struct i915_perf_stream *stream)
 {
 	struct drm_i915_private *i915 = stream->perf->i915;
+	struct intel_gt *gt = stream->engine->gt;
 	struct drm_i915_gem_object *bo;
 	struct i915_vma *vma;
 	const u64 delay_ticks = 0xffffffffffffffff -
@@ -1863,12 +1876,16 @@ static int alloc_noa_wait(struct i915_perf_stream *stream)
 	 * multiple OA config BOs will have a jump to this address and it
 	 * needs to be fixed during the lifetime of the i915/perf stream.
 	 */
-	vma = i915_gem_object_ggtt_pin_ww(bo, &ww, NULL, 0, 0, PIN_HIGH);
+	vma = i915_vma_instance(bo, &gt->ggtt->vm, NULL);
 	if (IS_ERR(vma)) {
 		ret = PTR_ERR(vma);
 		goto out_ww;
 	}
 
+	ret = i915_vma_pin_ww(vma, &ww, 0, 0, PIN_GLOBAL | PIN_HIGH);
+	if (ret)
+		goto out_ww;
+
 	batch = cs = i915_gem_object_pin_map(bo, I915_MAP_WB);
 	if (IS_ERR(batch)) {
 		ret = PTR_ERR(batch);
Signature · 1 line
2.25.1
#15 lore raw

PATCH v3 10/16 drm/i915/perf: Store a pointer to oa_format in oa_buffer

In reply to Umesh Nerlige Ramappa

DG2 introduces OA reports with 64 bit report header fields. Perf OA would need more information about the OA format in order to process such reports. Store all OA format info in oa_buffer instead of just the size and format-id.

v2: Drop format_size variable (Ashutosh)

---
 drivers/gpu/drm/i915/i915_perf.c       | 30 +++++++++++---------------
 drivers/gpu/drm/i915/i915_perf_types.h |  3 +--
 2 files changed, 13 insertions(+), 20 deletions(-)
drivers/gpu/drm/i915/i915_perf.c +12 -18
index 2f9e18ee0aab..ad74fa5847f7 100644
@@ -465,7 +465,7 @@ static u32 gen7_oa_hw_tail_read(struct i915_perf_stream *stream)
 static bool oa_buffer_check_unlocked(struct i915_perf_stream *stream)
 {
 	u32 gtt_offset = i915_ggtt_offset(stream->oa_buffer.vma);
-	int report_size = stream->oa_buffer.format_size;
+	int report_size = stream->oa_buffer.format->size;
 	unsigned long flags;
 	bool pollin;
 	u32 hw_tail;
@@ -602,7 +602,7 @@ static int append_oa_sample(struct i915_perf_stream *stream,
 			    size_t *offset,
 			    const u8 *report)
 {
-	int report_size = stream->oa_buffer.format_size;
+	int report_size = stream->oa_buffer.format->size;
 	struct drm_i915_perf_record_header header;
 
 	header.type = DRM_I915_PERF_RECORD_SAMPLE;
@@ -652,7 +652,7 @@ static int gen8_append_oa_reports(struct i915_perf_stream *stream,
 				  size_t *offset)
 {
 	struct intel_uncore *uncore = stream->uncore;
-	int report_size = stream->oa_buffer.format_size;
+	int report_size = stream->oa_buffer.format->size;
 	u8 *oa_buf_base = stream->oa_buffer.vaddr;
 	u32 gtt_offset = i915_ggtt_offset(stream->oa_buffer.vma);
 	u32 mask = (OA_BUFFER_SIZE - 1);
@@ -946,7 +946,7 @@ static int gen7_append_oa_reports(struct i915_perf_stream *stream,
 				  size_t *offset)
 {
 	struct intel_uncore *uncore = stream->uncore;
-	int report_size = stream->oa_buffer.format_size;
+	int report_size = stream->oa_buffer.format->size;
 	u8 *oa_buf_base = stream->oa_buffer.vaddr;
 	u32 gtt_offset = i915_ggtt_offset(stream->oa_buffer.vma);
 	u32 mask = (OA_BUFFER_SIZE - 1);
@@ -2502,7 +2502,7 @@ static int gen12_configure_oar_context(struct i915_perf_stream *stream,
 {
 	int err;
 	struct intel_context *ce = stream->pinned_ctx;
-	u32 format = stream->oa_buffer.format;
+	u32 format = stream->oa_buffer.format->format;
 	u32 offset = stream->perf->ctx_oactxctrl_offset;
 	struct flex regs_context[] = {
 		{
@@ -2875,7 +2875,7 @@ static void gen7_oa_enable(struct i915_perf_stream *stream)
 	u32 ctx_id = stream->specific_ctx_id;
 	bool periodic = stream->periodic;
 	u32 period_exponent = stream->period_exponent;
-	u32 report_format = stream->oa_buffer.format;
+	u32 report_format = stream->oa_buffer.format->format;
 
 	/*
 	 * Reset buf pointers so we don't forward reports from before now.
@@ -2901,7 +2901,7 @@ static void gen7_oa_enable(struct i915_perf_stream *stream)
 static void gen8_oa_enable(struct i915_perf_stream *stream)
 {
 	struct intel_uncore *uncore = stream->uncore;
-	u32 report_format = stream->oa_buffer.format;
+	u32 report_format = stream->oa_buffer.format->format;
 
 	/*
 	 * Reset buf pointers so we don't forward reports from before now.
@@ -2927,7 +2927,7 @@ static void gen8_oa_enable(struct i915_perf_stream *stream)
 static void gen12_oa_enable(struct i915_perf_stream *stream)
 {
 	struct intel_uncore *uncore = stream->uncore;
-	u32 report_format = stream->oa_buffer.format;
+	u32 report_format = stream->oa_buffer.format->format;
 
 	/*
 	 * If we don't want OA reports from the OA buffer, then we don't even
@@ -3108,7 +3108,6 @@ static int i915_oa_stream_init(struct i915_perf_stream *stream,
 	struct drm_i915_private *i915 = stream->perf->i915;
 	struct i915_perf *perf = stream->perf;
 	struct intel_gt *gt;
-	int format_size;
 	int ret;
 
 	if (!props->engine) {
@@ -3164,20 +3163,15 @@ static int i915_oa_stream_init(struct i915_perf_stream *stream,
 
 	stream->sample_size = sizeof(struct drm_i915_perf_record_header);
 
-	format_size = perf->oa_formats[props->oa_format].size;
+	stream->oa_buffer.format = &perf->oa_formats[props->oa_format];
+	if (drm_WARN_ON(&i915->drm, stream->oa_buffer.format->size == 0))
+		return -EINVAL;
 
 	stream->sample_flags = props->sample_flags;
-	stream->sample_size += format_size;
-
-	stream->oa_buffer.format_size = format_size;
-	if (drm_WARN_ON(&i915->drm, stream->oa_buffer.format_size == 0))
-		return -EINVAL;
+	stream->sample_size += stream->oa_buffer.format->size;
 
 	stream->hold_preemption = props->hold_preemption;
 
-	stream->oa_buffer.format =
-		perf->oa_formats[props->oa_format].format;
-
 	stream->periodic = props->oa_periodic;
 	if (stream->periodic)
 		stream->period_exponent = props->oa_period_exponent;
drivers/gpu/drm/i915/i915_perf_types.h +1 -2
index dc9bfd8086cf..e0c96b44eda8 100644
@@ -250,11 +250,10 @@ struct i915_perf_stream {
 	 * @oa_buffer: State of the OA buffer.
 	 */
 	struct {
+		const struct i915_oa_format *format;
 		struct i915_vma *vma;
 		u8 *vaddr;
 		u32 last_ctx_id;
-		int format;
-		int format_size;
 		int size_exponent;
 
 		/**
Signature · 1 line
2.25.1
#16 lore raw

PATCH v3 11/16 drm/i915/perf: Add Wa_1508761755:dg2

In reply to Umesh Nerlige Ramappa

Disable Clock gating in EU when gathering the events so that EU events are not lost.

v2: Fix checkpatch issues

---
 drivers/gpu/drm/i915/gt/intel_gt_regs.h |  1 +
 drivers/gpu/drm/i915/i915_perf.c        | 23 +++++++++++++++++++++++
 2 files changed, 24 insertions(+)
drivers/gpu/drm/i915/gt/intel_gt_regs.h +1 -0
index 2275ee47da95..8bb1694035a6 100644
@@ -1131,6 +1131,7 @@
 #define   GEN12_DISABLE_EARLY_READ		REG_BIT(14)
 #define   GEN12_ENABLE_LARGE_GRF_MODE		REG_BIT(12)
 #define   GEN12_PUSH_CONST_DEREF_HOLD_DIS	REG_BIT(8)
+#define   GEN12_DISABLE_DOP_GATING              REG_BIT(0)
 
 #define RT_CTRL					_MMIO(0xe530)
 #define   DIS_NULL_QUERY			REG_BIT(10)
drivers/gpu/drm/i915/i915_perf.c +23 -0
index ad74fa5847f7..ccc0bd9f2969 100644
@@ -2773,6 +2773,18 @@ gen12_enable_metric_set(struct i915_perf_stream *stream,
 	u32 sqcnt1;
 	int ret;
 
+	/*
+	 * Wa_1508761755:xehpsdv, dg2
+	 * EU NOA signals behave incorrectly if EU clock gating is enabled.
+	 * Disable thread stall DOP gating and EU DOP gating.
+	 */
+	if (IS_XEHPSDV(i915) || IS_DG2(i915)) {
+		intel_uncore_write(uncore, GEN8_ROW_CHICKEN,
+				   _MASKED_BIT_ENABLE(STALL_DOP_GATING_DISABLE));
+		intel_uncore_write(uncore, GEN7_ROW_CHICKEN2,
+				   _MASKED_BIT_ENABLE(GEN12_DISABLE_DOP_GATING));
+	}
+
 	intel_uncore_write(uncore, GEN12_OAG_OA_DEBUG,
 			   /* Disable clk ratio reports, like previous Gens. */
 			   _MASKED_BIT_ENABLE(GEN12_OAG_OA_DEBUG_DISABLE_CLK_RATIO_REPORTS |
@@ -2851,6 +2863,17 @@ static void gen12_disable_metric_set(struct i915_perf_stream *stream)
 	struct drm_i915_private *i915 = stream->perf->i915;
 	u32 sqcnt1;
 
+	/*
+	 * Wa_1508761755:xehpsdv, dg2
+	 * Enable thread stall DOP gating and EU DOP gating.
+	 */
+	if (IS_XEHPSDV(i915) || IS_DG2(i915)) {
+		intel_uncore_write(uncore, GEN8_ROW_CHICKEN,
+				   _MASKED_BIT_DISABLE(STALL_DOP_GATING_DISABLE));
+		intel_uncore_write(uncore, GEN7_ROW_CHICKEN2,
+				   _MASKED_BIT_DISABLE(GEN12_DISABLE_DOP_GATING));
+	}
+
 	/* Reset all contexts' slices/subslices configurations. */
 	gen12_configure_all_contexts(stream, NULL, NULL);
 
Signature · 1 line
2.25.1
#17 lore raw

PATCH v3 12/16 drm/i915/perf: Apply Wa_18013179988

In reply to Umesh Nerlige Ramappa

OA reports in the OA buffer contain an OA timestamp field that helps user calculate delta between 2 OA reports. The calculation relies on the CS timestamp frequency to convert the timestamp value to nanoseconds. The CS timestamp frequency is a function of the CTC_SHIFT value in RPM_CONFIG0.

In DG2, OA unit assumes that the CTC_SHIFT is 3, instead of using the actual value from RPM_CONFIG0. At the user level, this results in an error in calculating delta between 2 OA reports since the OA timestamp is not shifted in the same manner as CS timestamp. Also the periodicity of the reports is different from what the user configured because of mismatch in the CS and OA frequencies.

The issue also affects MI_REPORT_PERF_COUNT command.

To resolve this, return actual OA timestamp frequency to the user in i915_getparam_ioctl, so that user can calculate the right OA exponent as well as interpret the reports correctly.

v2:
- Use REG_FIELD_GET (Ashutosh)
- Update commit msg

---
 drivers/gpu/drm/i915/i915_getparam.c |  3 +++
 drivers/gpu/drm/i915/i915_perf.c     | 30 ++++++++++++++++++++++++++--
 drivers/gpu/drm/i915/i915_perf.h     |  2 ++
 include/uapi/drm/i915_drm.h          |  6 ++++++
 4 files changed, 39 insertions(+), 2 deletions(-)
drivers/gpu/drm/i915/i915_getparam.c +3 -0
index 342c8ca6414e..3047e80e1163 100644
@@ -175,6 +175,9 @@ int i915_getparam_ioctl(struct drm_device *dev, void *data,
 	case I915_PARAM_PERF_REVISION:
 		value = i915_perf_ioctl_version();
 		break;
+	case I915_PARAM_OA_TIMESTAMP_FREQUENCY:
+		value = i915_perf_oa_timestamp_frequency(i915);
+		break;
 	default:
 		DRM_DEBUG("Unknown parameter %d\n", param->param);
 		return -EINVAL;
drivers/gpu/drm/i915/i915_perf.c +28 -2
index ccc0bd9f2969..ece4b34461ad 100644
@@ -3106,6 +3106,30 @@ get_sseu_config(struct intel_sseu *out_sseu,
 	return i915_gem_user_to_context_sseu(engine->gt, drm_sseu, out_sseu);
 }
 
+/*
+ * OA timestamp frequency = CS timestamp frequency in most platforms. On some
+ * platforms OA unit ignores the CTC_SHIFT and the 2 timestamps differ. In such
+ * cases, return the adjusted CS timestamp frequency to the user.
+ */
+u32 i915_perf_oa_timestamp_frequency(struct drm_i915_private *i915)
+{
+	/* Wa_18013179988:dg2 */
+	if (IS_DG2(i915)) {
+		intel_wakeref_t wakeref;
+		u32 reg, shift;
+
+		with_intel_runtime_pm(to_gt(i915)->uncore->rpm, wakeref)
+			reg = intel_uncore_read(to_gt(i915)->uncore, RPM_CONFIG0);
+
+		shift = REG_FIELD_GET(GEN10_RPM_CONFIG0_CTC_SHIFT_PARAMETER_MASK,
+				      reg);
+
+		return to_gt(i915)->clock_frequency << (3 - shift);
+	}
+
+	return to_gt(i915)->clock_frequency;
+}
+
 /**
  * i915_oa_stream_init - validate combined props for OA stream and init
  * @stream: An i915 perf stream
@@ -3827,8 +3851,10 @@ i915_perf_open_ioctl_locked(struct i915_perf *perf,
 
 static u64 oa_exponent_to_ns(struct i915_perf *perf, int exponent)
 {
-	return intel_gt_clock_interval_to_ns(to_gt(perf->i915),
-					     2ULL << exponent);
+	u64 nom = (2ULL << exponent) * NSEC_PER_SEC;
+	u32 den = i915_perf_oa_timestamp_frequency(perf->i915);
+
+	return div_u64(nom + den - 1, den);
 }
 
 static __always_inline bool
drivers/gpu/drm/i915/i915_perf.h +2 -0
index 1d1329e5af3a..f96e09a4af04 100644
@@ -57,4 +57,6 @@ static inline void i915_oa_config_put(struct i915_oa_config *oa_config)
 	kref_put(&oa_config->ref, i915_oa_config_release);
 }
 
+u32 i915_perf_oa_timestamp_frequency(struct drm_i915_private *i915);
+
 #endif /* __I915_PERF_H__ */
include/uapi/drm/i915_drm.h +6 -0
index 8b59590e06d4..3a714980b9f5 100644
@@ -749,6 +749,12 @@ typedef struct drm_i915_irq_wait {
 /* Query if the kernel supports the I915_USERPTR_PROBE flag. */
 #define I915_PARAM_HAS_USERPTR_PROBE 56
 
+/*
+ * Frequency of the timestamps in OA reports. This used to be the same as the CS
+ * timestamp frequency, but differs on some platforms.
+ */
+#define I915_PARAM_OA_TIMESTAMP_FREQUENCY 57
+
 /* Must be kept compact -- no holes and well documented */
 
 /**
Signature · 1 line
2.25.1
#18 lore raw

PATCH v3 13/16 drm/i915/perf: Save/restore EU flex counters across reset

In reply to Umesh Nerlige Ramappa

If a drm client is killed, then hw contexts used by the client are reset immediately. This reset clears the EU flex counter configuration. If an OA use case is running in parallel, it would start seeing zeroed eu counter values following the reset even if the drm client is restarted. Save/restore the EU flex counter config so that the EU counters can be monitored continuously across resets.

---
 drivers/gpu/drm/i915/gt/uc/intel_guc_ads.c | 8 ++++++++
 1 file changed, 8 insertions(+)
drivers/gpu/drm/i915/gt/uc/intel_guc_ads.c +8 -0
index 74cbe8eaf531..3e152219fcb2 100644
@@ -375,6 +375,14 @@ static int guc_mmio_regset_init(struct temp_regset *regset,
 	for (i = 0; i < GEN9_LNCFCMOCS_REG_COUNT; i++)
 		ret |= GUC_MMIO_REG_ADD(gt, regset, GEN9_LNCFCMOCS(i), false);
 
+	ret |= GUC_MMIO_REG_ADD(gt, regset, EU_PERF_CNTL0, false);
+	ret |= GUC_MMIO_REG_ADD(gt, regset, EU_PERF_CNTL1, false);
+	ret |= GUC_MMIO_REG_ADD(gt, regset, EU_PERF_CNTL2, false);
+	ret |= GUC_MMIO_REG_ADD(gt, regset, EU_PERF_CNTL3, false);
+	ret |= GUC_MMIO_REG_ADD(gt, regset, EU_PERF_CNTL4, false);
+	ret |= GUC_MMIO_REG_ADD(gt, regset, EU_PERF_CNTL5, false);
+	ret |= GUC_MMIO_REG_ADD(gt, regset, EU_PERF_CNTL6, false);
+
 	return ret ? -1 : 0;
 }
 
Signature · 1 line
2.25.1
#19 lore raw

PATCH v3 14/16 drm/i915/guc: Support OA when Wa_16011777198 is enabled

In reply to Umesh Nerlige Ramappa

From: Vinay Belgaumkar <[email protected]>

On DG2, a w/a resets RCS/CCS before it goes into RC6. This breaks OA since OA does not expect engine resets during its use. Fix it by disabling RC6.

v2: (Ashutosh)
- Bring back slpc_unset_param helper
- Update commit msg
- Use with_intel_runtime_pm helper for set/unset

v3: (Ashutosh)
- Just use intel_uc_uses_guc_rc

---
 .../drm/i915/gt/uc/abi/guc_actions_slpc_abi.h |  9 +++
 drivers/gpu/drm/i915/gt/uc/intel_guc_slpc.c   | 66 +++++++++++++++++++
 drivers/gpu/drm/i915/gt/uc/intel_guc_slpc.h   |  2 +
 drivers/gpu/drm/i915/i915_perf.c              | 27 ++++++++
 4 files changed, 104 insertions(+)
drivers/gpu/drm/i915/gt/uc/abi/guc_actions_slpc_abi.h +9 -0
index 4c840a2639dc..811add10c30d 100644
@@ -128,6 +128,15 @@ enum slpc_media_ratio_mode {
 	SLPC_MEDIA_RATIO_MODE_FIXED_ONE_TO_TWO = 2,
 };
 
+enum slpc_gucrc_mode {
+	SLPC_GUCRC_MODE_HW = 0,
+	SLPC_GUCRC_MODE_GUCRC_NO_RC6 = 1,
+	SLPC_GUCRC_MODE_GUCRC_STATIC_TIMEOUT = 2,
+	SLPC_GUCRC_MODE_GUCRC_DYNAMIC_HYSTERESIS = 3,
+
+	SLPC_GUCRC_MODE_MAX,
+};
+
 enum slpc_event_id {
 	SLPC_EVENT_RESET = 0,
 	SLPC_EVENT_SHUTDOWN = 1,
drivers/gpu/drm/i915/gt/uc/intel_guc_slpc.c +66 -0

84 lines not on this page · raw message

drivers/gpu/drm/i915/gt/uc/intel_guc_slpc.h +2 -0
index 82a98f78f96c..ccf483730d9d 100644
@@ -42,5 +42,7 @@ int intel_guc_slpc_set_media_ratio_mode(struct intel_guc_slpc *slpc, u32 val);
 void intel_guc_pm_intrmsk_enable(struct intel_gt *gt);
 void intel_guc_slpc_boost(struct intel_guc_slpc *slpc);
 void intel_guc_slpc_dec_waiters(struct intel_guc_slpc *slpc);
+int intel_guc_slpc_unset_gucrc_mode(struct intel_guc_slpc *slpc);
+int intel_guc_slpc_override_gucrc_mode(struct intel_guc_slpc *slpc, u32 mode);
 
 #endif
drivers/gpu/drm/i915/i915_perf.c +27 -0
index ece4b34461ad..ae7efe8a831e 100644
@@ -208,6 +208,7 @@
 #include "gt/intel_lrc.h"
 #include "gt/intel_lrc_reg.h"
 #include "gt/intel_ring.h"
+#include "gt/uc/intel_guc_slpc.h"
 
 #include "i915_drv.h"
 #include "i915_file_private.h"
@@ -1577,6 +1578,15 @@ static void i915_oa_stream_destroy(struct i915_perf_stream *stream)
 
 	free_oa_buffer(stream);
 
+	/*
+	 * Wa_16011777198:dg2: Unset the override of GUCRC mode to enable rc6.
+	 */
+	if (intel_uc_uses_guc_rc(&gt->uc) &&
+	    (IS_DG2_GRAPHICS_STEP(gt->i915, G10, STEP_A0, STEP_C0) ||
+	     IS_DG2_GRAPHICS_STEP(gt->i915, G11, STEP_A0, STEP_B0)))
+		drm_WARN_ON(&gt->i915->drm,
+			    intel_guc_slpc_unset_gucrc_mode(&gt->uc.guc.slpc));
+
 	intel_uncore_forcewake_put(stream->uncore, FORCEWAKE_ALL);
 	intel_engine_pm_put(stream->engine);
 
@@ -3262,6 +3272,23 @@ static int i915_oa_stream_init(struct i915_perf_stream *stream,
 	intel_engine_pm_get(stream->engine);
 	intel_uncore_forcewake_get(stream->uncore, FORCEWAKE_ALL);
 
+	/*
+	 * Wa_16011777198:dg2: GuC resets render as part of the Wa. This causes
+	 * OA to lose the configuration state. Prevent this by overriding GUCRC
+	 * mode.
+	 */
+	if (intel_uc_uses_guc_rc(&gt->uc) &&
+	    (IS_DG2_GRAPHICS_STEP(gt->i915, G10, STEP_A0, STEP_C0) ||
+	     IS_DG2_GRAPHICS_STEP(gt->i915, G11, STEP_A0, STEP_B0))) {
+		ret = intel_guc_slpc_override_gucrc_mode(&gt->uc.guc.slpc,
+							 SLPC_GUCRC_MODE_GUCRC_NO_RC6);
+		if (ret) {
+			drm_dbg(&stream->perf->i915->drm,
+				"Unable to override gucrc mode\n");
+			goto err_config;
+		}
+	}
+
 	ret = alloc_oa_buffer(stream);
 	if (ret)
 		goto err_oa_buf_alloc;
Signature · 1 line
2.25.1
#20 lore raw

PATCH v3 15/16 drm/i915/perf: complete programming whitelisting for XEHPSDV

In reply to Umesh Nerlige Ramappa

From: Lionel Landwerlin <[email protected]>

We have an additional register to select which slices contribute to OAG/OAG counter increments.

---
 drivers/gpu/drm/i915/i915_drv.h          |  2 ++
 drivers/gpu/drm/i915/i915_pci.c          |  1 +
 drivers/gpu/drm/i915/i915_perf.c         | 13 +++++++++++++
 drivers/gpu/drm/i915/intel_device_info.h |  1 +
 4 files changed, 17 insertions(+)
drivers/gpu/drm/i915/i915_drv.h +2 -0
index ccd54ff54002..992ca7b0aea0 100644
@@ -904,6 +904,8 @@ IS_SUBPLATFORM(const struct drm_i915_private *i915,
 
 #define HAS_OA_BPC_REPORTING(dev_priv) \
 	(INTEL_INFO(dev_priv)->has_oa_bpc_reporting)
+#define HAS_OA_SLICE_CONTRIB_LIMITS(dev_priv) \
+	(INTEL_INFO(dev_priv)->has_oa_slice_contrib_limits)
 
 /*
  * Set this flag, when platform requires 64K GTT page sizes or larger for
drivers/gpu/drm/i915/i915_pci.c +1 -0
index 6b25e4cb6221..f5de5ae0b800 100644
@@ -1023,6 +1023,7 @@ static const struct intel_device_info adl_p_info = {
 	.has_logical_ring_elsq = 1, \
 	.has_mslice_steering = 1, \
 	.has_oa_bpc_reporting = 1, \
+	.has_oa_slice_contrib_limits = 1, \
 	.has_rc6 = 1, \
 	.has_reset_engine = 1, \
 	.has_rps = 1, \
drivers/gpu/drm/i915/i915_perf.c +13 -0
index ae7efe8a831e..c9ab7eaa15ea 100644
@@ -4258,6 +4258,11 @@ static const struct i915_range gen12_oa_b_counters[] = {
 	{}
 };
 
+static const struct i915_range xehp_oa_b_counters[] = {
+	{ .start = 0xdc48, .end = 0xdc48 },	/* OAA_ENABLE_REG */
+	{ .start = 0xdd00, .end = 0xdd48 },	/* OAG_LCE0_0 - OAA_LENABLE_REG */
+};
+
 static const struct i915_range gen7_oa_mux_regs[] = {
 	{ .start = 0x91b8, .end = 0x91cc },	/* OA_PERFCNT[1-2], OA_PERFMATRIX */
 	{ .start = 0x9800, .end = 0x9888 },	/* MICRO_BP0_0 - NOA_WRITE */
@@ -4332,6 +4337,12 @@ static bool gen12_is_valid_b_counter_addr(struct i915_perf *perf, u32 addr)
 	return reg_in_range_table(addr, gen12_oa_b_counters);
 }
 
+static bool xehp_is_valid_b_counter_addr(struct i915_perf *perf, u32 addr)
+{
+	return reg_in_range_table(addr, xehp_oa_b_counters) ||
+		reg_in_range_table(addr, gen12_oa_b_counters);
+}
+
 static bool gen12_is_valid_mux_addr(struct i915_perf *perf, u32 addr)
 {
 	return reg_in_range_table(addr, gen12_oa_mux_regs);
@@ -4844,6 +4855,8 @@ void i915_perf_init(struct drm_i915_private *i915)
 			perf->ops.oa_hw_tail_read = gen8_oa_hw_tail_read;
 		} else if (GRAPHICS_VER(i915) == 12) {
 			perf->ops.is_valid_b_counter_reg =
+				HAS_OA_SLICE_CONTRIB_LIMITS(i915) ?
+				xehp_is_valid_b_counter_addr :
 				gen12_is_valid_b_counter_addr;
 			perf->ops.is_valid_mux_reg =
 				gen12_is_valid_mux_addr;
drivers/gpu/drm/i915/intel_device_info.h +1 -0
index 1f7a842cd408..9975f736bc41 100644
@@ -165,6 +165,7 @@ enum intel_ppgtt_type {
 	func(has_media_ratio_mode); \
 	func(has_mslice_steering); \
 	func(has_oa_bpc_reporting); \
+	func(has_oa_slice_contrib_limits); \
 	func(has_one_eu_per_fuse_bit); \
 	func(has_pxp); \
 	func(has_rc6); \
Signature · 1 line
2.25.1
Reviewed-by on v3 15/16
#21 lore raw

Re: [Intel-gfx] [PATCH v3 15/16] drm/i915/perf: complete programming whitelisting for XEHPSDV

Quoted 20 lines · Umesh Nerlige Ramappa

From: Lionel Landwerlin <[email protected]>

We have an additional register to select which slices contribute to OAG/OAG counter increments.

---
 drivers/gpu/drm/i915/i915_drv.h          |  2 ++
 drivers/gpu/drm/i915/i915_pci.c          |  1 +
 drivers/gpu/drm/i915/i915_perf.c         | 13 +++++++++++++
 drivers/gpu/drm/i915/intel_device_info.h |  1 +
 4 files changed, 17 insertions(+)
drivers/gpu/drm/i915/i915_drv.h +0 -0
index ccd54ff54002..992ca7b0aea0 100644

drivers/gpu/drm/i915/i915_drv.h IS_SUBPLATFORM(const struct drm_i915_private *i915, · line 908

Quoted 5 lines
drivers/gpu/drm/i915/i915_drv.h +2 -0
@@ -904,6 +904,8 @@ IS_SUBPLATFORM(const struct drm_i915_private *i915,
 
 #define HAS_OA_BPC_REPORTING(dev_priv) \
 	(INTEL_INFO(dev_priv)->has_oa_bpc_reporting)
+#define HAS_OA_SLICE_CONTRIB_LIMITS(dev_priv) \
+	(INTEL_INFO(dev_priv)->has_oa_slice_contrib_limits)

If this is just for XEHPSDV maybe we could just have a platform check rather than introduce a flag? Otherwise this is:

#22 lore raw

Re: [Intel-gfx] [PATCH v3 15/16] drm/i915/perf: complete programming whitelisting for XEHPSDV

Quoted 22 lines · Ashutosh
Quoted 21 lines · Umesh Nerlige Ramappa

From: Lionel Landwerlin <[email protected]>

We have an additional register to select which slices contribute to OAG/OAG counter increments.

---
 drivers/gpu/drm/i915/i915_drv.h          |  2 ++
 drivers/gpu/drm/i915/i915_pci.c          |  1 +
 drivers/gpu/drm/i915/i915_perf.c         | 13 +++++++++++++
 drivers/gpu/drm/i915/intel_device_info.h |  1 +
 4 files changed, 17 insertions(+)
drivers/gpu/drm/i915/i915_drv.h +0 -0
index ccd54ff54002..992ca7b0aea0 100644
@@ -904,6 +904,8 @@ IS_SUBPLATFORM(const struct drm_i915_private *i915,
 
Quoted 7 lines

drivers/gpu/drm/i915/i915_drv.h IS_SUBPLATFORM(const struct drm_i915_private *i915, · line 908

Quoted 4 lines
drivers/gpu/drm/i915/i915_drv.h +2 -0
 #define HAS_OA_BPC_REPORTING(dev_priv) \
 	(INTEL_INFO(dev_priv)->has_oa_bpc_reporting)
+#define HAS_OA_SLICE_CONTRIB_LIMITS(dev_priv) \
+	(INTEL_INFO(dev_priv)->has_oa_slice_contrib_limits)

If this is just for XEHPSDV maybe we could just have a platform check rather than introduce a flag? Otherwise this is:

I think when it was added, it was only xehpsdv, but now applies to dg2 and some other platforms.

Thanks,
Umesh

Quoted 2 lines
#23 lore raw

PATCH v3 16/16 drm/i915/perf: Enable OA for DG2

In reply to Umesh Nerlige Ramappa

OA was disabled for DG2 as support was missing. Enable it back now.

---
 drivers/gpu/drm/i915/i915_perf.c | 6 ------
 1 file changed, 6 deletions(-)
drivers/gpu/drm/i915/i915_perf.c +0 -6
index c9ab7eaa15ea..cb3c67abb82e 100644
@@ -4795,12 +4795,6 @@ void i915_perf_init(struct drm_i915_private *i915)
 {
 	struct i915_perf *perf = &i915->perf;
 
-	/* XXX const struct i915_perf_ops! */
-
-	/* i915_perf is not enabled for DG2 yet */
-	if (IS_DG2(i915))
-		return;
-
 	perf->oa_formats = oa_formats;
 	if (IS_HASWELL(i915)) {
 		perf->ops.is_valid_b_counter_reg = gen7_is_valid_b_counter_addr;
Signature · 1 line
2.25.1
Reviewed-by on v3 16/16
#24 lore raw

Re: [Intel-gfx] [PATCH v3 16/16] drm/i915/perf: Enable OA for DG2

Quoted 2 lines · Umesh Nerlige Ramappa

OA was disabled for DG2 as support was missing. Enable it back now.

Quoted 25 lines
---
 drivers/gpu/drm/i915/i915_perf.c | 6 ------
 1 file changed, 6 deletions(-)
drivers/gpu/drm/i915/i915_perf.c +0 -7
index c9ab7eaa15ea..cb3c67abb82e 100644
@@ -4795,12 +4795,6 @@ void i915_perf_init(struct drm_i915_private *i915)
 {
 	struct i915_perf *perf = &i915->perf;
 
-	/* XXX const struct i915_perf_ops! */
-
-	/* i915_perf is not enabled for DG2 yet */
-	if (IS_DG2(i915))
-		return;
-
 	perf->oa_formats = oa_formats;
 	if (IS_HASWELL(i915)) {
 		perf->ops.is_valid_b_counter_reg = gen7_is_valid_b_counter_addr;
--

2.25.1

#25 lore raw

[Intel-gfx] ✗ Fi.CI.CHECKPATCH: warning for Add DG2 OA support (rev5)

In reply to Umesh Nerlige Ramappa
== Series Details ==

== Series Details ==

Series: Add DG2 OA support (rev5)
URL : https://patchwork.freedesktop.org/series/107584/
State : warning

== Summary ==

Error: dim checkpatch failed
e7d2ddc2181d drm/i915/perf: Fix OA filtering logic for GuC mode 1d70790b1303 drm/i915/perf: Add 32-bit OAG and OAR formats for DG2 b97b644bb1d0 drm/i915/perf: Fix noa wait predication for DG2 da8d80f58908 drm/i915/perf: Determine gen12 oa ctx offset at runtime -:44: CHECK:MACRO_ARG_REUSE: Macro argument reuse 'x' - possible side-effects? #44: FILE: drivers/gpu/drm/i915/i915_perf.c:1359:
+#define valid_oactxctrl_offset(x) ((x) && (x) != U32_MAX)

total: 0 errors, 0 warnings, 1 checks, 242 lines checked 0f7174c003eb drm/i915/perf: Enable bytes per clock reporting in OA b286c128b54f drm/i915/perf: Simply use stream->ctx
2efd0183464c drm/i915/perf: Move gt-specific data from i915->perf to gt->perf 5245cfe41005 drm/i915/perf: Replace gt->perf.lock with stream->lock for file ops 1fb05a2f731b drm/i915/perf: Use gt-specific ggtt for OA and noa-wait buffers 6b6fc115fa11 drm/i915/perf: Store a pointer to oa_format in oa_buffer a32dfbc72a72 drm/i915/perf: Add Wa_1508761755:dg2
475cba3e1d18 drm/i915/perf: Apply Wa_18013179988
59eaa4a25577 drm/i915/perf: Save/restore EU flex counters across reset 180bdc18d55d drm/i915/guc: Support OA when Wa_16011777198 is enabled 5c2fb7812b19 drm/i915/perf: complete programming whitelisting for XEHPSDV 20611dd13047 drm/i915/perf: Enable OA for DG2