diff mbox series

[11/26] drm/i915/perf: implement active wait for noa configurations

Message ID 20200505083239.2428948-12-tjaalton@ubuntu.com
State New
Headers show
Series drm/i915: Backport support for VK_INTEL_performance_query extension | expand

Commit Message

Timo Aaltonen May 5, 2020, 8:32 a.m. UTC
From: Lionel Landwerlin <lionel.g.landwerlin@intel.com>

BugLink: https://bugs.launchpad.net/bugs/1871957

NOA configuration take some amount of time to apply. That amount of
time depends on the size of the GT. There is no documented time for
this. For example, past experimentations with powergating
configuration changes seem to indicate a 60~70us delay. We go with
500us as default for now which should be over the required amount of
time (according to HW architects).

v2: Don't forget to save/restore registers used for the wait (Chris)

v3: Name used CS_GPR registers (Chris)
    Fix compile issue due to rebase (Lionel)

v4: Fix save/restore helpers (Umesh)

v5: Move noa_wait from drm_i915_private to i915_perf_stream (Lionel)

v6: Add missing struct declarations in i915_perf.h

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: https://patchwork.freedesktop.org/patch/msgid/20191012072308.30312-2-chris@chris-wilson.co.uk
(cherry picked from commit daed3e44396d178cf2098b754bb8ef5ca4e918bc)
Signed-off-by: Timo Aaltonen <timo.aaltonen@canonical.com>
---
 drivers/gpu/drm/i915/gt/intel_gpu_commands.h  |   4 +-
 drivers/gpu/drm/i915/gt/intel_gt_types.h      |   5 +
 drivers/gpu/drm/i915/i915_debugfs.c           |  32 +++
 drivers/gpu/drm/i915/i915_perf.c              | 224 ++++++++++++++++++
 drivers/gpu/drm/i915/i915_perf_types.h        |   8 +
 drivers/gpu/drm/i915/i915_reg.h               |   4 +-
 .../drm/i915/selftests/i915_live_selftests.h  |   1 +
 drivers/gpu/drm/i915/selftests/i915_perf.c    | 216 +++++++++++++++++
 8 files changed, 492 insertions(+), 2 deletions(-)
 create mode 100644 drivers/gpu/drm/i915/selftests/i915_perf.c
diff mbox series

Patch

diff --git a/drivers/gpu/drm/i915/gt/intel_gpu_commands.h b/drivers/gpu/drm/i915/gt/intel_gpu_commands.h
index fed06f057627..f3b7e8a2e751 100644
--- a/drivers/gpu/drm/i915/gt/intel_gpu_commands.h
+++ b/drivers/gpu/drm/i915/gt/intel_gpu_commands.h
@@ -157,7 +157,8 @@ 
 #define MI_BATCH_BUFFER_START	MI_INSTR(0x31, 0)
 #define   MI_BATCH_GTT		    (2<<6) /* aliased with (1<<7) on gen4 */
 #define MI_BATCH_BUFFER_START_GEN8	MI_INSTR(0x31, 1)
-#define   MI_BATCH_RESOURCE_STREAMER (1<<10)
+#define   MI_BATCH_RESOURCE_STREAMER REG_BIT(10)
+#define   MI_BATCH_PREDICATE         REG_BIT(15) /* HSW+ on RCS only*/
 
 /*
  * 3D instructions used by the kernel
@@ -218,6 +219,7 @@ 
 #define   PIPE_CONTROL_CS_STALL				(1<<20)
 #define   PIPE_CONTROL_TLB_INVALIDATE			(1<<18)
 #define   PIPE_CONTROL_MEDIA_STATE_CLEAR		(1<<16)
+#define   PIPE_CONTROL_WRITE_TIMESTAMP			(3<<14)
 #define   PIPE_CONTROL_QW_WRITE				(1<<14)
 #define   PIPE_CONTROL_POST_SYNC_OP_MASK                (3<<14)
 #define   PIPE_CONTROL_DEPTH_STALL			(1<<13)
diff --git a/drivers/gpu/drm/i915/gt/intel_gt_types.h b/drivers/gpu/drm/i915/gt/intel_gt_types.h
index dc295c196d11..f752b6cf9ea1 100644
--- a/drivers/gpu/drm/i915/gt/intel_gt_types.h
+++ b/drivers/gpu/drm/i915/gt/intel_gt_types.h
@@ -97,6 +97,11 @@  enum intel_gt_scratch_field {
 	/* 8 bytes */
 	INTEL_GT_SCRATCH_FIELD_COHERENTL3_WA = 256,
 
+	/* 6 * 8 bytes */
+	INTEL_GT_SCRATCH_FIELD_PERF_CS_GPR = 2048,
+
+	/* 4 bytes */
+	INTEL_GT_SCRATCH_FIELD_PERF_PREDICATE_RESULT_1 = 2096,
 };
 
 #endif /* __INTEL_GT_TYPES_H__ */
diff --git a/drivers/gpu/drm/i915/i915_debugfs.c b/drivers/gpu/drm/i915/i915_debugfs.c
index c4750354581d..1f3c7387e1c1 100644
--- a/drivers/gpu/drm/i915/i915_debugfs.c
+++ b/drivers/gpu/drm/i915/i915_debugfs.c
@@ -3572,6 +3572,37 @@  DEFINE_SIMPLE_ATTRIBUTE(i915_wedged_fops,
 			i915_wedged_get, i915_wedged_set,
 			"%llu\n");
 
+static int
+i915_perf_noa_delay_set(void *data, u64 val)
+{
+	struct drm_i915_private *i915 = data;
+	const u32 clk = RUNTIME_INFO(i915)->cs_timestamp_frequency_khz;
+
+	/*
+	 * This would lead to infinite waits as we're doing timestamp
+	 * difference on the CS with only 32bits.
+	 */
+	if (val > mul_u32_u32(U32_MAX, clk))
+		return -EINVAL;
+
+	atomic64_set(&i915->perf.noa_programming_delay, val);
+	return 0;
+}
+
+static int
+i915_perf_noa_delay_get(void *data, u64 *val)
+{
+	struct drm_i915_private *i915 = data;
+
+	*val = atomic64_read(&i915->perf.noa_programming_delay);
+	return 0;
+}
+
+DEFINE_SIMPLE_ATTRIBUTE(i915_perf_noa_delay_fops,
+			i915_perf_noa_delay_get,
+			i915_perf_noa_delay_set,
+			"%llu\n");
+
 #define DROP_UNBOUND	BIT(0)
 #define DROP_BOUND	BIT(1)
 #define DROP_RETIRE	BIT(2)
@@ -4338,6 +4369,7 @@  static const struct i915_debugfs_files {
 	const char *name;
 	const struct file_operations *fops;
 } i915_debugfs_files[] = {
+	{"i915_perf_noa_delay", &i915_perf_noa_delay_fops},
 	{"i915_wedged", &i915_wedged_fops},
 	{"i915_cache_sharing", &i915_cache_sharing_fops},
 	{"i915_gem_drop_caches", &i915_drop_caches_fops},
diff --git a/drivers/gpu/drm/i915/i915_perf.c b/drivers/gpu/drm/i915/i915_perf.c
index a06cb0df8c9e..a603992a8304 100644
--- a/drivers/gpu/drm/i915/i915_perf.c
+++ b/drivers/gpu/drm/i915/i915_perf.c
@@ -198,6 +198,7 @@ 
 #include "gem/i915_gem_context.h"
 #include "gt/intel_engine_pm.h"
 #include "gt/intel_engine_user.h"
+#include "gt/intel_gt.h"
 #include "gt/intel_lrc_reg.h"
 
 #include "i915_drv.h"
@@ -1369,6 +1370,12 @@  free_oa_configs(struct i915_perf_stream *stream)
 		free_oa_config_bo(oa_bo);
 }
 
+static void
+free_noa_wait(struct i915_perf_stream *stream)
+{
+	i915_vma_unpin_and_release(&stream->noa_wait, 0);
+}
+
 static void i915_oa_stream_destroy(struct i915_perf_stream *stream)
 {
 	struct i915_perf *perf = stream->perf;
@@ -1393,6 +1400,7 @@  static void i915_oa_stream_destroy(struct i915_perf_stream *stream)
 		oa_put_render_ctx_id(stream);
 
 	free_oa_configs(stream);
+	free_noa_wait(stream);
 
 	if (perf->spurious_report_rs.missed) {
 		DRM_NOTE("%d spurious OA report notices suppressed due to ratelimiting\n",
@@ -1565,6 +1573,206 @@  static int alloc_oa_buffer(struct i915_perf_stream *stream)
 	return ret;
 }
 
+static u32 *save_restore_register(struct i915_perf_stream *stream, u32 *cs,
+				  bool save, i915_reg_t reg, u32 offset,
+				  u32 dword_count)
+{
+	u32 cmd;
+	u32 d;
+
+	cmd = save ? MI_STORE_REGISTER_MEM : MI_LOAD_REGISTER_MEM;
+	if (INTEL_GEN(stream->perf->i915) >= 8)
+		cmd++;
+
+	for (d = 0; d < dword_count; d++) {
+		*cs++ = cmd;
+		*cs++ = i915_mmio_reg_offset(reg) + 4 * d;
+		*cs++ = intel_gt_scratch_offset(stream->engine->gt,
+						offset) + 4 * d;
+		*cs++ = 0;
+	}
+
+	return cs;
+}
+
+static int alloc_noa_wait(struct i915_perf_stream *stream)
+{
+	struct drm_i915_private *i915 = stream->perf->i915;
+	struct drm_i915_gem_object *bo;
+	struct i915_vma *vma;
+	const u64 delay_ticks = 0xffffffffffffffff -
+		DIV64_U64_ROUND_UP(
+			atomic64_read(&stream->perf->noa_programming_delay) *
+			RUNTIME_INFO(i915)->cs_timestamp_frequency_khz,
+			1000000ull);
+	const u32 base = stream->engine->mmio_base;
+#define CS_GPR(x) GEN8_RING_CS_GPR(base, x)
+	u32 *batch, *ts0, *cs, *jump;
+	int ret, i;
+	enum {
+		START_TS,
+		NOW_TS,
+		DELTA_TS,
+		JUMP_PREDICATE,
+		DELTA_TARGET,
+		N_CS_GPR
+	};
+
+	bo = i915_gem_object_create_internal(i915, 4096);
+	if (IS_ERR(bo)) {
+		DRM_ERROR("Failed to allocate NOA wait batchbuffer\n");
+		return PTR_ERR(bo);
+	}
+
+	/*
+	 * We pin in GGTT because we jump into this buffer now because
+	 * 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(bo, NULL, 0, 0, PIN_HIGH);
+	if (IS_ERR(vma)) {
+		ret = PTR_ERR(vma);
+		goto err_unref;
+	}
+
+	batch = cs = i915_gem_object_pin_map(bo, I915_MAP_WB);
+	if (IS_ERR(batch)) {
+		ret = PTR_ERR(batch);
+		goto err_unpin;
+	}
+
+	/* Save registers. */
+	for (i = 0; i < N_CS_GPR; i++)
+		cs = save_restore_register(
+			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,
+		INTEL_GT_SCRATCH_FIELD_PERF_PREDICATE_RESULT_1, 1);
+
+	/* First timestamp snapshot location. */
+	ts0 = cs;
+
+	/*
+	 * Initial snapshot of the timestamp register to implement the wait.
+	 * We work with 32b values, so clear out the top 32b bits of the
+	 * register because the ALU works 64bits.
+	 */
+	*cs++ = MI_LOAD_REGISTER_IMM(1);
+	*cs++ = i915_mmio_reg_offset(CS_GPR(START_TS)) + 4;
+	*cs++ = 0;
+	*cs++ = MI_LOAD_REGISTER_REG | (3 - 2);
+	*cs++ = i915_mmio_reg_offset(RING_TIMESTAMP(base));
+	*cs++ = i915_mmio_reg_offset(CS_GPR(START_TS));
+
+	/*
+	 * This is the location we're going to jump back into until the
+	 * required amount of time has passed.
+	 */
+	jump = cs;
+
+	/*
+	 * Take another snapshot of the timestamp register. Take care to clear
+	 * up the top 32bits of CS_GPR(1) as we're using it for other
+	 * operations below.
+	 */
+	*cs++ = MI_LOAD_REGISTER_IMM(1);
+	*cs++ = i915_mmio_reg_offset(CS_GPR(NOW_TS)) + 4;
+	*cs++ = 0;
+	*cs++ = MI_LOAD_REGISTER_REG | (3 - 2);
+	*cs++ = i915_mmio_reg_offset(RING_TIMESTAMP(base));
+	*cs++ = i915_mmio_reg_offset(CS_GPR(NOW_TS));
+
+	/*
+	 * Do a diff between the 2 timestamps and store the result back into
+	 * CS_GPR(1).
+	 */
+	*cs++ = MI_MATH(5);
+	*cs++ = MI_MATH_LOAD(MI_MATH_REG_SRCA, MI_MATH_REG(NOW_TS));
+	*cs++ = MI_MATH_LOAD(MI_MATH_REG_SRCB, MI_MATH_REG(START_TS));
+	*cs++ = MI_MATH_SUB;
+	*cs++ = MI_MATH_STORE(MI_MATH_REG(DELTA_TS), MI_MATH_REG_ACCU);
+	*cs++ = MI_MATH_STORE(MI_MATH_REG(JUMP_PREDICATE), MI_MATH_REG_CF);
+
+	/*
+	 * Transfer the carry flag (set to 1 if ts1 < ts0, meaning the
+	 * timestamp have rolled over the 32bits) into the predicate register
+	 * to be used for the predicated jump.
+	 */
+	*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);
+
+	/* Restart from the beginning if we had timestamps roll over. */
+	*cs++ = (INTEL_GEN(i915) < 8 ?
+		 MI_BATCH_BUFFER_START :
+		 MI_BATCH_BUFFER_START_GEN8) |
+		MI_BATCH_PREDICATE;
+	*cs++ = i915_ggtt_offset(vma) + (ts0 - batch) * 4;
+	*cs++ = 0;
+
+	/*
+	 * Now add the diff between to previous timestamps and add it to :
+	 *      (((1 * << 64) - 1) - delay_ns)
+	 *
+	 * When the Carry Flag contains 1 this means the elapsed time is
+	 * longer than the expected delay, and we can exit the wait loop.
+	 */
+	*cs++ = MI_LOAD_REGISTER_IMM(2);
+	*cs++ = i915_mmio_reg_offset(CS_GPR(DELTA_TARGET));
+	*cs++ = lower_32_bits(delay_ticks);
+	*cs++ = i915_mmio_reg_offset(CS_GPR(DELTA_TARGET)) + 4;
+	*cs++ = upper_32_bits(delay_ticks);
+
+	*cs++ = MI_MATH(4);
+	*cs++ = MI_MATH_LOAD(MI_MATH_REG_SRCA, MI_MATH_REG(DELTA_TS));
+	*cs++ = MI_MATH_LOAD(MI_MATH_REG_SRCB, MI_MATH_REG(DELTA_TARGET));
+	*cs++ = MI_MATH_ADD;
+	*cs++ = MI_MATH_STOREINV(MI_MATH_REG(JUMP_PREDICATE), MI_MATH_REG_CF);
+
+	/*
+	 * Transfer the result into the predicate register to be used for the
+	 * predicated jump.
+	 */
+	*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);
+
+	/* Predicate the jump.  */
+	*cs++ = (INTEL_GEN(i915) < 8 ?
+		 MI_BATCH_BUFFER_START :
+		 MI_BATCH_BUFFER_START_GEN8) |
+		MI_BATCH_PREDICATE;
+	*cs++ = i915_ggtt_offset(vma) + (jump - batch) * 4;
+	*cs++ = 0;
+
+	/* 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,
+		INTEL_GT_SCRATCH_FIELD_PERF_PREDICATE_RESULT_1, 1);
+
+	/* And return to the ring. */
+	*cs++ = MI_BATCH_BUFFER_END;
+
+	GEM_BUG_ON(cs - batch > PAGE_SIZE / sizeof(*batch));
+
+	i915_gem_object_flush_map(bo);
+	i915_gem_object_unpin_map(bo);
+
+	stream->noa_wait = vma;
+	return 0;
+
+err_unpin:
+	__i915_vma_unpin(vma);
+err_unref:
+	i915_gem_object_put(bo);
+	return ret;
+}
+
 static void config_oa_regs(struct intel_uncore *uncore,
 			   const struct i915_oa_reg *regs,
 			   u32 n_regs)
@@ -2235,6 +2443,12 @@  static int i915_oa_stream_init(struct i915_perf_stream *stream,
 		}
 	}
 
+	ret = alloc_noa_wait(stream);
+	if (ret) {
+		DRM_DEBUG("Unable to allocate NOA wait batch buffer\n");
+		goto err_noa_wait_alloc;
+	}
+
 	stream->oa_config = i915_perf_get_oa_config(perf, props->metrics_set);
 	if (!stream->oa_config) {
 		DRM_DEBUG("Invalid OA config id=%i\n", props->metrics_set);
@@ -2302,6 +2516,9 @@  static int i915_oa_stream_init(struct i915_perf_stream *stream,
 	intel_engine_pm_put(stream->engine);
 
 err_config:
+	free_noa_wait(stream);
+
+err_noa_wait_alloc:
 	if (stream->ctx)
 		oa_put_render_ctx_id(stream);
 
@@ -3689,6 +3906,9 @@  void i915_perf_init(struct drm_i915_private *i915)
 		ratelimit_set_flags(&perf->spurious_report_rs,
 				    RATELIMIT_MSG_ON_RELEASE);
 
+		atomic64_set(&perf->noa_programming_delay,
+			     500 * 1000 /* 500us */);
+
 		perf->i915 = i915;
 	}
 }
@@ -3718,3 +3938,7 @@  void i915_perf_fini(struct drm_i915_private *i915)
 	memset(&perf->ops, 0, sizeof(perf->ops));
 	perf->i915 = NULL;
 }
+
+#if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
+#include "selftests/i915_perf.c"
+#endif
diff --git a/drivers/gpu/drm/i915/i915_perf_types.h b/drivers/gpu/drm/i915/i915_perf_types.h
index 337cd7d2ad77..d35a3c1946c3 100644
--- a/drivers/gpu/drm/i915/i915_perf_types.h
+++ b/drivers/gpu/drm/i915/i915_perf_types.h
@@ -266,6 +266,12 @@  struct i915_perf_stream {
 		 */
 		u32 head;
 	} oa_buffer;
+
+	/**
+	 * A batch buffer doing a wait on the GPU for the NOA logic to be
+	 * reprogrammed.
+	 */
+	struct i915_vma *noa_wait;
 };
 
 /**
@@ -385,6 +391,8 @@  struct i915_perf {
 
 	struct i915_oa_ops ops;
 	const struct i915_oa_format *oa_formats;
+
+	atomic64_t noa_programming_delay;
 };
 
 #endif /* _I915_PERF_TYPES_H_ */
diff --git a/drivers/gpu/drm/i915/i915_reg.h b/drivers/gpu/drm/i915/i915_reg.h
index b5eb382514ec..c877cda32808 100644
--- a/drivers/gpu/drm/i915/i915_reg.h
+++ b/drivers/gpu/drm/i915/i915_reg.h
@@ -547,7 +547,9 @@  static inline bool i915_mmio_reg_valid(i915_reg_t reg)
 #define MI_PREDICATE_SRC0_UDW	_MMIO(0x2400 + 4)
 #define MI_PREDICATE_SRC1	_MMIO(0x2408)
 #define MI_PREDICATE_SRC1_UDW	_MMIO(0x2408 + 4)
-
+#define MI_PREDICATE_DATA       _MMIO(0x2410)
+#define MI_PREDICATE_RESULT     _MMIO(0x2418)
+#define MI_PREDICATE_RESULT_1   _MMIO(0x241c)
 #define MI_PREDICATE_RESULT_2	_MMIO(0x2214)
 #define  LOWER_SLICE_ENABLED	(1 << 0)
 #define  LOWER_SLICE_DISABLED	(0 << 0)
diff --git a/drivers/gpu/drm/i915/selftests/i915_live_selftests.h b/drivers/gpu/drm/i915/selftests/i915_live_selftests.h
index 1ccf0f731ac0..846ede427f37 100644
--- a/drivers/gpu/drm/i915/selftests/i915_live_selftests.h
+++ b/drivers/gpu/drm/i915/selftests/i915_live_selftests.h
@@ -33,3 +33,4 @@  selftest(reset, intel_reset_live_selftests)
 selftest(hangcheck, intel_hangcheck_live_selftests)
 selftest(execlists, intel_execlists_live_selftests)
 selftest(guc, intel_guc_live_selftest)
+selftest(perf, i915_perf_live_selftests)
diff --git a/drivers/gpu/drm/i915/selftests/i915_perf.c b/drivers/gpu/drm/i915/selftests/i915_perf.c
new file mode 100644
index 000000000000..9abd525bfa15
--- /dev/null
+++ b/drivers/gpu/drm/i915/selftests/i915_perf.c
@@ -0,0 +1,216 @@ 
+/*
+ * SPDX-License-Identifier: MIT
+ *
+ * Copyright © 2019 Intel Corporation
+ */
+
+#include <linux/kref.h>
+
+#include "gem/i915_gem_pm.h"
+#include "gt/intel_gt.h"
+
+#include "i915_selftest.h"
+
+#include "igt_flush_test.h"
+#include "lib_sw_fence.h"
+
+static struct i915_perf_stream *
+test_stream(struct i915_perf *perf)
+{
+	struct drm_i915_perf_open_param param = {};
+	struct perf_open_properties props = {
+		.engine = intel_engine_lookup_user(perf->i915,
+						   I915_ENGINE_CLASS_RENDER,
+						   0),
+		.sample_flags = SAMPLE_OA_REPORT,
+		.oa_format = I915_OA_FORMAT_C4_B8,
+		.metrics_set = 1,
+	};
+	struct i915_perf_stream *stream;
+
+	stream = kzalloc(sizeof(*stream), GFP_KERNEL);
+	if (!stream)
+		return NULL;
+
+	stream->perf = perf;
+
+	mutex_lock(&perf->lock);
+	if (i915_oa_stream_init(stream, &param, &props)) {
+		kfree(stream);
+		stream =  NULL;
+	}
+	mutex_unlock(&perf->lock);
+
+	return stream;
+}
+
+static void stream_destroy(struct i915_perf_stream *stream)
+{
+	struct i915_perf *perf = stream->perf;
+
+	mutex_lock(&perf->lock);
+	i915_perf_destroy_locked(stream);
+	mutex_unlock(&perf->lock);
+}
+
+static int live_sanitycheck(void *arg)
+{
+	struct drm_i915_private *i915 = arg;
+	struct i915_perf_stream *stream;
+
+	/* Quick check we can create a perf stream */
+
+	stream = test_stream(&i915->perf);
+	if (!stream)
+		return -EINVAL;
+
+	stream_destroy(stream);
+	return 0;
+}
+
+static int write_timestamp(struct i915_request *rq, int slot)
+{
+	u32 *cs;
+	int len;
+
+	cs = intel_ring_begin(rq, 6);
+	if (IS_ERR(cs))
+		return PTR_ERR(cs);
+
+	len = 5;
+	if (INTEL_GEN(rq->i915) >= 8)
+		len++;
+
+	*cs++ = GFX_OP_PIPE_CONTROL(len);
+	*cs++ = PIPE_CONTROL_GLOBAL_GTT_IVB |
+		PIPE_CONTROL_STORE_DATA_INDEX |
+		PIPE_CONTROL_WRITE_TIMESTAMP;
+	*cs++ = slot * sizeof(u32);
+	*cs++ = 0;
+	*cs++ = 0;
+	*cs++ = 0;
+
+	intel_ring_advance(rq, cs);
+
+	return 0;
+}
+
+static ktime_t poll_status(struct i915_request *rq, int slot)
+{
+	while (!intel_read_status_page(rq->engine, slot) &&
+	       !i915_request_completed(rq))
+		cpu_relax();
+
+	return ktime_get();
+}
+
+static int live_noa_delay(void *arg)
+{
+	struct drm_i915_private *i915 = arg;
+	struct i915_perf_stream *stream;
+	struct i915_request *rq;
+	ktime_t t0, t1;
+	u64 expected;
+	u32 delay;
+	int err;
+	int i;
+
+	/* Check that the GPU delays matches expectations */
+
+	stream = test_stream(&i915->perf);
+	if (!stream)
+		return -ENOMEM;
+
+	expected = atomic64_read(&stream->perf->noa_programming_delay);
+
+	if (stream->engine->class != RENDER_CLASS) {
+		err = -ENODEV;
+		goto out;
+	}
+
+	for (i = 0; i < 4; i++)
+		intel_write_status_page(stream->engine, 0x100 + i, 0);
+
+	rq = i915_request_create(stream->engine->kernel_context);
+	if (IS_ERR(rq)) {
+		err = PTR_ERR(rq);
+		goto out;
+	}
+
+	if (rq->engine->emit_init_breadcrumb &&
+	    rq->timeline->has_initial_breadcrumb) {
+		err = rq->engine->emit_init_breadcrumb(rq);
+		if (err) {
+			i915_request_add(rq);
+			goto out;
+		}
+	}
+
+	err = write_timestamp(rq, 0x100);
+	if (err) {
+		i915_request_add(rq);
+		goto out;
+	}
+
+	err = rq->engine->emit_bb_start(rq,
+					i915_ggtt_offset(stream->noa_wait), 0,
+					I915_DISPATCH_SECURE);
+	if (err) {
+		i915_request_add(rq);
+		goto out;
+	}
+
+	err = write_timestamp(rq, 0x102);
+	if (err) {
+		i915_request_add(rq);
+		goto out;
+	}
+
+	i915_request_get(rq);
+	i915_request_add(rq);
+
+	preempt_disable();
+	t0 = poll_status(rq, 0x100);
+	t1 = poll_status(rq, 0x102);
+	preempt_enable();
+
+	pr_info("CPU delay: %lluns, expected %lluns\n",
+		ktime_sub(t1, t0), expected);
+
+	delay = intel_read_status_page(stream->engine, 0x102);
+	delay -= intel_read_status_page(stream->engine, 0x100);
+	delay = div_u64(mul_u32_u32(delay, 1000 * 1000),
+			RUNTIME_INFO(i915)->cs_timestamp_frequency_khz);
+	pr_info("GPU delay: %uns, expected %lluns\n",
+		delay, expected);
+
+	if (4 * delay < 3 * expected || 2 * delay > 3 * expected) {
+		pr_err("GPU delay [%uus] outside of expected threshold! [%lluus, %lluus]\n",
+		       delay / 1000,
+		       div_u64(3 * expected, 4000),
+		       div_u64(3 * expected, 2000));
+		err = -EINVAL;
+	}
+
+	i915_request_put(rq);
+out:
+	stream_destroy(stream);
+	return err;
+}
+
+int i915_perf_live_selftests(struct drm_i915_private *i915)
+{
+	static const struct i915_subtest tests[] = {
+		SUBTEST(live_sanitycheck),
+		SUBTEST(live_noa_delay),
+	};
+	struct i915_perf *perf = &i915->perf;
+
+	if (!perf->metrics_kobj || !perf->ops.enable_metric_set)
+		return 0;
+
+	if (intel_gt_is_wedged(&i915->gt))
+		return 0;
+
+	return i915_subtests(tests, i915);
+}