diff mbox series

[v3,6/7] drm/i915/perf: allow holding preemption on filtered ctx

Message ID 20190604131140.12647-7-lionel.g.landwerlin@intel.com (mailing list archive)
State New, archived
Headers show
Series drm/i915: Vulkan performance query support | expand

Commit Message

Lionel Landwerlin June 4, 2019, 1:11 p.m. UTC
We would like to make use of perf in Vulkan. The Vulkan API is much
lower level than OpenGL, with applications directly exposed to the
concept of command buffers (pretty much equivalent to our batch
buffers). In Vulkan, queries are always limited in scope to a command
buffer. In OpenGL, the lack of command buffer concept meant that
queries' duration could span multiple command buffers.

With that restriction gone in Vulkan, we would like to simplify
measuring performance just by measuring the deltas between the counter
snapshots written by 2 MI_RECORD_PERF_COUNT commands, rather than the
more complex scheme we currently have in the GL driver, using 2
MI_RECORD_PERF_COUNT commands and doing some post processing on the
stream of OA reports, coming from the global OA buffer, to remove any
unrelated deltas in between the 2 MI_RECORD_PERF_COUNT.

Disabling preemption only apply to a single context with which want to
query performance counters for and is considered a privileged
operation, by default protected by CAP_SYS_ADMIN. It is possible to
enable it for a normal user by disabling the paranoid stream setting.

v2: Store preemption setting in intel_context (Chris)

v3: Use priorities to avoid preemption rather than the HW mechanism

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
---
 .../gpu/drm/i915/gem/i915_gem_execbuffer.c    |  8 +++++++
 drivers/gpu/drm/i915/gt/intel_lrc.c           |  6 +++++
 drivers/gpu/drm/i915/i915_drv.c               |  2 +-
 drivers/gpu/drm/i915/i915_drv.h               |  8 +++++++
 drivers/gpu/drm/i915/i915_perf.c              | 22 +++++++++++++++++--
 drivers/gpu/drm/i915/i915_priolist_types.h    |  7 ++++++
 drivers/gpu/drm/i915/i915_request.c           |  1 +
 drivers/gpu/drm/i915/i915_request.h           |  1 +
 drivers/gpu/drm/i915/intel_guc_submission.c   |  6 +++++
 include/uapi/drm/i915_drm.h                   | 10 +++++++++
 10 files changed, 68 insertions(+), 3 deletions(-)

Comments

Chris Wilson June 4, 2019, 1:45 p.m. UTC | #1
Quoting Lionel Landwerlin (2019-06-04 14:11:39)
> diff --git a/drivers/gpu/drm/i915/gt/intel_lrc.c b/drivers/gpu/drm/i915/gt/intel_lrc.c
> index ed19f4e53d31..4046f045408b 100644
> --- a/drivers/gpu/drm/i915/gt/intel_lrc.c
> +++ b/drivers/gpu/drm/i915/gt/intel_lrc.c
> @@ -683,6 +683,12 @@ static void port_assign(struct execlist_port *port, struct i915_request *rq)
>         if (port_isset(port))
>                 i915_request_put(port_request(port));
>  
> +       if (rq->has_perf) {
> +               rq->sched.attr.priority =
> +                       (I915_PRIORITY_MASK & rq->sched.attr.priority) |
> +                       I915_USER_PRIORITY(I915_PRIORITY_PERF);
> +       }

This is broken. You can not ignore PI here. If you bump the priority here you
must increase the priority of all of its cross-engine dependencies as
they may still be inflight and later reordered causing deadlocks. (Note
you cannot take the locks required for bumping other engines here.)
-Chris
diff mbox series

Patch

diff --git a/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c b/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c
index 4a785999a9c5..59ec1c0a8879 100644
--- a/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c
+++ b/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c
@@ -2100,6 +2100,14 @@  static int eb_oa_config(struct i915_execbuffer *eb)
 	if (!eb->oa_config)
 		return 0;
 
+	/*
+	 * If the perf stream was opened with hold preemption, flag the
+	 * request properly so that the priority of the request is bumped once
+	 * it reaches the execlist ports.
+	 */
+	if (eb->i915->perf.oa.exclusive_stream->hold_preemption)
+		eb->request->has_perf = true;
+
 	/*
 	 * If the config hasn't changed, skip reconfiguring the HW (this is
 	 * subject to a delay we want to avoid has much as possible).
diff --git a/drivers/gpu/drm/i915/gt/intel_lrc.c b/drivers/gpu/drm/i915/gt/intel_lrc.c
index ed19f4e53d31..4046f045408b 100644
--- a/drivers/gpu/drm/i915/gt/intel_lrc.c
+++ b/drivers/gpu/drm/i915/gt/intel_lrc.c
@@ -683,6 +683,12 @@  static void port_assign(struct execlist_port *port, struct i915_request *rq)
 	if (port_isset(port))
 		i915_request_put(port_request(port));
 
+	if (rq->has_perf) {
+		rq->sched.attr.priority =
+			(I915_PRIORITY_MASK & rq->sched.attr.priority) |
+			I915_USER_PRIORITY(I915_PRIORITY_PERF);
+	}
+
 	port_set(port, port_pack(i915_request_get(rq), port_count(port)));
 }
 
diff --git a/drivers/gpu/drm/i915/i915_drv.c b/drivers/gpu/drm/i915/i915_drv.c
index 75029d1a3802..984b76a09cfa 100644
--- a/drivers/gpu/drm/i915/i915_drv.c
+++ b/drivers/gpu/drm/i915/i915_drv.c
@@ -476,7 +476,7 @@  static int i915_getparam_ioctl(struct drm_device *dev, void *data,
 		value = INTEL_INFO(dev_priv)->has_coherent_ggtt;
 		break;
 	case I915_PARAM_PERF_REVISION:
-		value = 1;
+		value = 2;
 		break;
 	case I915_PARAM_HAS_EXEC_PERF_CONFIG:
 		/* Obviously requires perf support. */
diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
index aefdae856e77..57f915751d76 100644
--- a/drivers/gpu/drm/i915/i915_drv.h
+++ b/drivers/gpu/drm/i915/i915_drv.h
@@ -1396,6 +1396,14 @@  struct i915_perf_stream {
 	 */
 	bool enabled;
 
+	/**
+	 * @hold_preemption: Whether preemption is put on hold for command
+	 * submissions done on the @ctx. This is useful for some drivers that
+	 * cannot easily post process the OA buffer context to subtract delta
+	 * of performance counters not associated with @ctx.
+	 */
+	bool hold_preemption;
+
 	/**
 	 * @ops: The callbacks providing the implementation of this specific
 	 * type of configured stream.
diff --git a/drivers/gpu/drm/i915/i915_perf.c b/drivers/gpu/drm/i915/i915_perf.c
index 82a282f668c0..f8897e3cc862 100644
--- a/drivers/gpu/drm/i915/i915_perf.c
+++ b/drivers/gpu/drm/i915/i915_perf.c
@@ -343,6 +343,8 @@  static const struct i915_oa_format gen8_plus_oa_formats[I915_OA_FORMAT_MAX] = {
  * struct perf_open_properties - for validated properties given to open a stream
  * @sample_flags: `DRM_I915_PERF_PROP_SAMPLE_*` properties are tracked as flags
  * @single_context: Whether a single or all gpu contexts should be monitored
+ * @hold_preemption: Whether the preemption is disabled for the filtered
+ *                   context
  * @ctx_handle: A gem ctx handle for use with @single_context
  * @metrics_set: An ID for an OA unit metric set advertised via sysfs
  * @oa_format: An OA unit HW report format
@@ -357,6 +359,7 @@  struct perf_open_properties {
 	u32 sample_flags;
 
 	u64 single_context:1;
+	u64 hold_preemption:1;
 	u64 ctx_handle;
 
 	/* OA sampling state */
@@ -2170,6 +2173,8 @@  static int i915_oa_stream_init(struct i915_perf_stream *stream,
 	stream->sample_flags |= SAMPLE_OA_REPORT;
 	stream->sample_size += format_size;
 
+	stream->hold_preemption = props->hold_preemption;
+
 	dev_priv->perf.oa.oa_buffer.format_size = format_size;
 	if (WARN_ON(dev_priv->perf.oa.oa_buffer.format_size == 0))
 		return -EINVAL;
@@ -2695,6 +2700,15 @@  i915_perf_open_ioctl_locked(struct drm_i915_private *dev_priv,
 		}
 	}
 
+	if (props->hold_preemption) {
+		if (!props->single_context) {
+			DRM_DEBUG("preemption disable with no context\n");
+			ret = -EINVAL;
+			goto err;
+		}
+		privileged_op = true;
+	}
+
 	/*
 	 * On Haswell the OA unit supports clock gating off for a specific
 	 * context and in this mode there's no visibility of metrics for the
@@ -2709,8 +2723,9 @@  i915_perf_open_ioctl_locked(struct drm_i915_private *dev_priv,
 	 * MI_REPORT_PERF_COUNT commands and so consider it a privileged op to
 	 * enable the OA unit by default.
 	 */
-	if (IS_HASWELL(dev_priv) && specific_ctx)
+	if (IS_HASWELL(dev_priv) && specific_ctx && !props->hold_preemption) {
 		privileged_op = false;
+	}
 
 	/* Similar to perf's kernel.perf_paranoid_cpu sysctl option
 	 * we check a dev.i915.perf_stream_paranoid sysctl option
@@ -2719,7 +2734,7 @@  i915_perf_open_ioctl_locked(struct drm_i915_private *dev_priv,
 	 */
 	if (privileged_op &&
 	    i915_perf_stream_paranoid && !capable(CAP_SYS_ADMIN)) {
-		DRM_DEBUG("Insufficient privileges to open system-wide i915 perf stream\n");
+		DRM_DEBUG("Insufficient privileges to open i915 perf stream\n");
 		ret = -EACCES;
 		goto err_ctx;
 	}
@@ -2911,6 +2926,9 @@  static int read_properties_unlocked(struct drm_i915_private *dev_priv,
 			props->oa_periodic = true;
 			props->oa_period_exponent = value;
 			break;
+		case DRM_I915_PERF_PROP_HOLD_PREEMPTION:
+			props->hold_preemption = value != 0 ? 1 : 0;
+			break;
 		case DRM_I915_PERF_PROP_MAX:
 			MISSING_CASE(id);
 			return -EINVAL;
diff --git a/drivers/gpu/drm/i915/i915_priolist_types.h b/drivers/gpu/drm/i915/i915_priolist_types.h
index 49709de69875..fc1c1fe7ce58 100644
--- a/drivers/gpu/drm/i915/i915_priolist_types.h
+++ b/drivers/gpu/drm/i915/i915_priolist_types.h
@@ -17,6 +17,13 @@  enum {
 	I915_PRIORITY_NORMAL = I915_CONTEXT_DEFAULT_PRIORITY,
 	I915_PRIORITY_MAX = I915_CONTEXT_MAX_USER_PRIORITY + 1,
 
+	/* Requests containing performance queries must not be preempted by
+	 * another context. They get scheduled with their default priority and
+	 * once they reach the execlist ports we bump them to
+	 * I915_PRIORITY_PERF so that they stick to the HW until they finish.
+	 */
+	I915_PRIORITY_PERF = I915_CONTEXT_MAX_USER_PRIORITY + 2,
+
 	I915_PRIORITY_INVALID = INT_MIN
 };
 
diff --git a/drivers/gpu/drm/i915/i915_request.c b/drivers/gpu/drm/i915/i915_request.c
index da1e6984a8cc..48dbe988db23 100644
--- a/drivers/gpu/drm/i915/i915_request.c
+++ b/drivers/gpu/drm/i915/i915_request.c
@@ -725,6 +725,7 @@  __i915_request_create(struct intel_context *ce, gfp_t gfp)
 	rq->batch = NULL;
 	rq->capture_list = NULL;
 	rq->waitboost = false;
+	rq->has_perf = false;
 	rq->execution_mask = ALL_ENGINES;
 
 	INIT_LIST_HEAD(&rq->active_list);
diff --git a/drivers/gpu/drm/i915/i915_request.h b/drivers/gpu/drm/i915/i915_request.h
index c9f7d07991c8..ab09b5bb89f7 100644
--- a/drivers/gpu/drm/i915/i915_request.h
+++ b/drivers/gpu/drm/i915/i915_request.h
@@ -216,6 +216,7 @@  struct i915_request {
 	unsigned long emitted_jiffies;
 
 	bool waitboost;
+	bool has_perf;
 
 	/** engine->request_list entry for this request */
 	struct list_head link;
diff --git a/drivers/gpu/drm/i915/intel_guc_submission.c b/drivers/gpu/drm/i915/intel_guc_submission.c
index a4f98ccef0fe..cba3f50dc1fe 100644
--- a/drivers/gpu/drm/i915/intel_guc_submission.c
+++ b/drivers/gpu/drm/i915/intel_guc_submission.c
@@ -717,6 +717,12 @@  static void port_assign(struct execlist_port *port, struct i915_request *rq)
 {
 	GEM_BUG_ON(port_isset(port));
 
+	if (rq->has_perf) {
+		rq->sched.attr.priority =
+			(I915_PRIORITY_MASK & rq->sched.attr.priority) |
+			I915_USER_PRIORITY(I915_PRIORITY_PERF);
+	}
+
 	port_set(port, i915_request_get(rq));
 }
 
diff --git a/include/uapi/drm/i915_drm.h b/include/uapi/drm/i915_drm.h
index 7f770183ee31..8ffdbc0e4616 100644
--- a/include/uapi/drm/i915_drm.h
+++ b/include/uapi/drm/i915_drm.h
@@ -2001,6 +2001,16 @@  enum drm_i915_perf_property_id {
 	 */
 	DRM_I915_PERF_PROP_OA_EXPONENT,
 
+	/**
+	 * Specifying this property is only valid when specify a context to
+	 * filter with DRM_I915_PERF_PROP_CTX_HANDLE. Specifying this property
+	 * will hold preemption of the particular context we want to gather
+	 * performance data about.
+	 *
+	 * This property is available in perf revision 2.
+	 */
+	DRM_I915_PERF_PROP_HOLD_PREEMPTION,
+
 	DRM_I915_PERF_PROP_MAX /* non-ABI */
 };