Skip to contentSkip to conversation

Collect command stream based OA reports using i915 perf

PATCH In review [email protected] 4 participants 24 messages Mar 16, 2017 – Mar 17, 2017

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

cover Collect command stream based OA reports using i915 perf

From: Sourab Gupta <[email protected]>

This series adds framework for collection of OA reports associated with the render command stream, which are collected around batchbuffer boundaries.

Refloating the series rebased on Robert's latest patch set for 'Enabling OA unit for Gen 8 and 9 in i915 perf', which can be found here:
https://patchwork.freedesktop.org/series/20084/

Since Robert's patches are being reviewed and this patch series extends his framework to collect command stream based OA metrics, it would be good to keep this work in perspective. Looking to receive feedback (and possibly r-b's :)) on the series.

Since the OA reports collected associated with the render command stream, this also gives us the ability to collect other metadata such as ctx_id, pid, etc. with the samples, and thus we can establish the association of samples collected with the corresponding process/workload.

These patches can be found for viewing at
https://github.com/sourabgu/linux/tree/oa-6march2017

Sourab Gupta (8):
drm/i915: Add ctx getparam ioctl parameter to retrieve ctx unique id drm/i915: Expose OA sample source to userspace
drm/i915: Framework for capturing command stream based OA reports drm/i915: flush periodic samples, in case of no pending CS sample
requests
drm/i915: Inform userspace about command stream OA buf overflow drm/i915: Populate ctx ID for periodic OA reports drm/i915: Add support for having pid output with OA report drm/i915: Add support for emitting execbuffer tags through OA counter
reports

 drivers/gpu/drm/i915/i915_drv.h            |  125 ++-
 drivers/gpu/drm/i915/i915_gem_context.c    |    3 +
 drivers/gpu/drm/i915/i915_gem_execbuffer.c |    6 +
 drivers/gpu/drm/i915/i915_perf.c           | 1149 ++++++++++++++++++++++++----
 include/uapi/drm/i915_drm.h                |   49 ++
 5 files changed, 1184 insertions(+), 148 deletions(-)
Signature · 6 lines
1.9.1

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#1 lore raw

PATCH 1/8 drm/i915: Add ctx getparam ioctl parameter to retrieve ctx unique id

From: Sourab Gupta <[email protected]>

This patch adds a new ctx getparam ioctl parameter, which can be used to retrieve ctx unique id by userspace.

This can be used by userspace to map the OA reports received in the i915 perf samples with their associated ctx's (The OA reports have the hw ctx ID information embedded for Gen8+).
Otherwise the userspace has no way of maintaining this association, since it has the knowledge of only per-drm file specific ctx handles.

---
 drivers/gpu/drm/i915/i915_gem_context.c | 3 +++
 include/uapi/drm/i915_drm.h             | 1 +
 2 files changed, 4 insertions(+)
drivers/gpu/drm/i915/i915_gem_context.c +3 -0
index baceca1..b6d2125 100644
@@ -1062,6 +1062,9 @@ int i915_gem_context_getparam_ioctl(struct drm_device *dev, void *data,
 	case I915_CONTEXT_PARAM_BANNABLE:
 		args->value = i915_gem_context_is_bannable(ctx);
 		break;
+	case I915_CONTEXT_PARAM_HW_ID:
+		args->value = ctx->hw_id;
+		break;
 	default:
 		ret = -EINVAL;
 		break;
include/uapi/drm/i915_drm.h +1 -0
index 03b8338..835e711 100644
@@ -1291,6 +1291,7 @@ struct drm_i915_gem_context_param {
 #define I915_CONTEXT_PARAM_GTT_SIZE	0x3
 #define I915_CONTEXT_PARAM_NO_ERROR_CAPTURE	0x4
 #define I915_CONTEXT_PARAM_BANNABLE	0x5
+#define I915_CONTEXT_PARAM_HW_ID	0x6
 	__u64 value;
 };
 
Signature · 6 lines
1.9.1

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Reviewed-by on v1 1/8
#2 lore raw

Re: [PATCH 1/8] drm/i915: Add ctx getparam ioctl parameter to retrieve ctx unique id

Quoted 5 lines · On Thu, Mar 16, 2017 at 11:44:08AM +0530, [email protected] wrote:

From: Sourab Gupta <[email protected]>

This patch adds a new ctx getparam ioctl parameter, which can be used to retrieve ctx unique id by userspace.

Quoted 7 lines

This can be used by userspace to map the OA reports received in the i915 perf samples with their associated ctx's (The OA reports have the hw ctx ID information embedded for Gen8+).
Otherwise the userspace has no way of maintaining this association, since it has the knowledge of only per-drm file specific ctx handles.

-Chris

Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#3 lore raw

PATCH 2/8 drm/i915: Expose OA sample source to userspace

In reply to [email protected]

From: Sourab Gupta <[email protected]>

This patch exposes a new sample source field to userspace. This field can be populated to specify the origin of the OA report.
Currently, the OA samples are being generated only periodically, and hence there's only source flag enum definition right now, but there are other means of generating OA samples, such as via MI_RPC commands. The OA_SOURCE sample type is introducing a mechanism (for userspace) to distinguish various OA reports generated via different sources.

---
 drivers/gpu/drm/i915/i915_perf.c | 18 ++++++++++++++++++
 include/uapi/drm/i915_drm.h      | 14 ++++++++++++++
 2 files changed, 32 insertions(+)
drivers/gpu/drm/i915/i915_perf.c +18 -0
index 4b1db73..540c5b2 100644
@@ -324,6 +324,7 @@
 };
 
 #define SAMPLE_OA_REPORT      (1<<0)
+#define SAMPLE_OA_SOURCE_INFO	(1<<1)
 
 /**
  * struct perf_open_properties - for validated properties given to open a stream
@@ -659,6 +660,15 @@ static int append_oa_sample(struct i915_perf_stream *stream,
 		return -EFAULT;
 	buf += sizeof(header);
 
+	if (sample_flags & SAMPLE_OA_SOURCE_INFO) {
+		enum drm_i915_perf_oa_event_source source =
+				I915_PERF_OA_EVENT_SOURCE_PERIODIC;
+
+		if (copy_to_user(buf, &source, 4))
+			return -EFAULT;
+		buf += 4;
+	}
+
 	if (sample_flags & SAMPLE_OA_REPORT) {
 		if (copy_to_user(buf, report, report_size))
 			return -EFAULT;
@@ -2030,6 +2040,11 @@ static int i915_oa_stream_init(struct i915_perf_stream *stream,
 	stream->sample_flags |= SAMPLE_OA_REPORT;
 	stream->sample_size += format_size;
 
+	if (props->sample_flags & SAMPLE_OA_SOURCE_INFO) {
+		stream->sample_flags |= SAMPLE_OA_SOURCE_INFO;
+		stream->sample_size += 4;
+	}
+
 	dev_priv->perf.oa.oa_buffer.format_size = format_size;
 	if (WARN_ON(dev_priv->perf.oa.oa_buffer.format_size == 0))
 		return -EINVAL;
@@ -2814,6 +2829,9 @@ 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_SAMPLE_OA_SOURCE:
+			props->sample_flags |= SAMPLE_OA_SOURCE_INFO;
+			break;
 		default:
 			MISSING_CASE(id);
 			DRM_DEBUG("Unknown i915 perf property ID\n");
include/uapi/drm/i915_drm.h +14 -0
index 835e711..c597e36 100644
@@ -1312,6 +1312,12 @@ enum drm_i915_oa_format {
 	I915_OA_FORMAT_MAX	    /* non-ABI */
 };
 
+enum drm_i915_perf_oa_event_source {
+	I915_PERF_OA_EVENT_SOURCE_PERIODIC,
+
+	I915_PERF_OA_EVENT_SOURCE_MAX	/* non-ABI */
+};
+
 enum drm_i915_perf_property_id {
 	/**
 	 * Open the stream for a specific context handle (as used with
@@ -1346,6 +1352,13 @@ enum drm_i915_perf_property_id {
 	 */
 	DRM_I915_PERF_PROP_OA_EXPONENT,
 
+	/**
+	 * The value of this property set to 1 requests inclusion of sample
+	 * source field to be given to userspace. The sample source field
+	 * specifies the origin of OA report.
+	 */
+	DRM_I915_PERF_PROP_SAMPLE_OA_SOURCE,
+
 	DRM_I915_PERF_PROP_MAX /* non-ABI */
 };
 
@@ -1411,6 +1424,7 @@ enum drm_i915_perf_record_type {
 	 * struct {
 	 *     struct drm_i915_perf_record_header header;
 	 *
+	 *     { u32 source_info; } && DRM_I915_PERF_PROP_SAMPLE_OA_SOURCE
 	 *     { u32 oa_report[]; } && DRM_I915_PERF_PROP_SAMPLE_OA
 	 * };
 	 */
Signature · 6 lines
1.9.1

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#4 lore raw

Re: [PATCH 2/8] drm/i915: Expose OA sample source to userspace

Quoted 9 lines · On Thu, Mar 16, 2017 at 6:14 AM, <[email protected]> wrote:

From: Sourab Gupta <[email protected]>

This patch exposes a new sample source field to userspace. This field can be populated to specify the origin of the OA report.
Currently, the OA samples are being generated only periodically, and hence there's only source flag enum definition right now, but there are other means of generating OA samples, such as via MI_RPC commands. The OA_SOURCE sample type is introducing a mechanism (for userspace) to distinguish various OA reports generated via different sources.

Maybe we could clarify that it's not intended as a replacement for the reason field that's part of Gen8+ OA reports. For automatically triggered reports written to the OABUFFER then the reason field will distinguish e.g. periodic vs ctx-switch vs GO transition reasons for the OA unit writing a report. The reason field is overloaded as the RPT_ID field for MI_RPC reports so we need our own way of tracking the difference.

Quoted 21 lines
---
 drivers/gpu/drm/i915/i915_perf.c | 18 ++++++++++++++++++
 include/uapi/drm/i915_drm.h      | 14 ++++++++++++++
 2 files changed, 32 insertions(+)
drivers/gpu/drm/i915/i915_perf.c +1 -0
index 4b1db73..540c5b2 100644
@@ -324,6 +324,7 @@
 };
 
 #define SAMPLE_OA_REPORT      (1<<0)
+#define SAMPLE_OA_SOURCE_INFO  (1<<1)
 
 /**
  * struct perf_open_properties - for validated properties given to open a stream

drivers/gpu/drm/i915/i915_perf.c static int append_oa_sample(struct i915_perf_stream *stream, · line 662

Quoted 3 lines
drivers/gpu/drm/i915/i915_perf.c +0 -0
@@ -659,6 +660,15 @@ static int append_oa_sample(struct i915_perf_stream *stream,
                return -EFAULT;
        buf += sizeof(header);
 

I think I'd add a note mentioning that the ordering of _OA_SOURCE_INFO before the _OA_REPORT forms part of the uapi - just as an extra guard against someone reorganising code not realising that. Maybe over paranoid though.

drivers/gpu/drm/i915/i915_perf.c static int append_oa_sample(struct i915_perf_stream *stream, · line 670

Quoted 8 lines
drivers/gpu/drm/i915/i915_perf.c +8 -0
+       if (sample_flags & SAMPLE_OA_SOURCE_INFO) {
+               enum drm_i915_perf_oa_event_source source =
+                               I915_PERF_OA_EVENT_SOURCE_PERIODIC;
+
+               if (copy_to_user(buf, &source, 4))
+                       return -EFAULT;
+               buf += 4;
+       }

I think we should keep all sample sections 8 byte aligned to keep any embedded u64 values we might want within records naturally aligned. Currently we know OA reports all have power-of-two sizes of 64, 128 or 256 bytes. The _OA_REPORT_LOST and _OA_BUFFER_LOST records only have an 8 byte header.

'PERIODIC' may not be an accurate differentiator. As mentioned above with the comment about then Gen8+ 'reason' field there are a number of different automatic triggers for writing OA reports to the OABUFFER. Maybe 'OABUFFER' would be better.

I'm not sure the final '_INFO' suffix is necessary here. To me I think I'd expect richer, more-detailed data with that suffix and I'd expect the data to about the source while this is just about identifying the source. (not too important at this stage - just mentioning how it comes across when I read it currently).

drivers/gpu/drm/i915/i915_perf.c static int i915_oa_stream_init(struct i915_perf_stream *stream, · line 2046

Quoted 12 lines
drivers/gpu/drm/i915/i915_perf.c +5 -0
+
        if (sample_flags & SAMPLE_OA_REPORT) {
                if (copy_to_user(buf, report, report_size))
                        return -EFAULT;
@@ -2030,6 +2040,11 @@ static int i915_oa_stream_init(struct i915_perf_stream *stream,
        stream->sample_flags |= SAMPLE_OA_REPORT;
        stream->sample_size += format_size;
 
+       if (props->sample_flags & SAMPLE_OA_SOURCE_INFO) {
+               stream->sample_flags |= SAMPLE_OA_SOURCE_INFO;
+               stream->sample_size += 4;
+       }

8 - as commented above.

Quoted 46 lines
drivers/gpu/drm/i915/i915_perf.c +4 -0
+
        dev_priv->perf.oa.oa_buffer.format_size = format_size;
        if (WARN_ON(dev_priv->perf.oa.oa_buffer.format_size == 0))
                return -EINVAL;
@@ -2814,6 +2829,9 @@ 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_SAMPLE_OA_SOURCE:
+                       props->sample_flags |= SAMPLE_OA_SOURCE_INFO;
+                       break;
                default:
                        MISSING_CASE(id);
                        DRM_DEBUG("Unknown i915 perf property ID\n");
include/uapi/drm/i915_drm.h +13 -0
index 835e711..c597e36 100644
@@ -1312,6 +1312,12 @@ enum drm_i915_oa_format {
        I915_OA_FORMAT_MAX          /* non-ABI */
 };
 
+enum drm_i915_perf_oa_event_source {
+       I915_PERF_OA_EVENT_SOURCE_PERIODIC,
+
+       I915_PERF_OA_EVENT_SOURCE_MAX   /* non-ABI */
+};
+
 enum drm_i915_perf_property_id {
        /**
         * Open the stream for a specific context handle (as used with
@@ -1346,6 +1352,13 @@ enum drm_i915_perf_property_id {
         */
        DRM_I915_PERF_PROP_OA_EXPONENT,
 
+       /**
+        * The value of this property set to 1 requests inclusion of sample
+        * source field to be given to userspace. The sample source field
+        * specifies the origin of OA report.
+        */
+       DRM_I915_PERF_PROP_SAMPLE_OA_SOURCE,
+
        DRM_I915_PERF_PROP_MAX /* non-ABI */
 };
 

include/uapi/drm/i915_drm.h enum drm_i915_perf_record_type { · line 1427

Quoted 4 lines
include/uapi/drm/i915_drm.h +1 -0
@@ -1411,6 +1424,7 @@ enum drm_i915_perf_record_type {
         * struct {
         *     struct drm_i915_perf_record_header header;
         *
+        *     { u32 source_info; } && DRM_I915_PERF_PROP_SAMPLE_OA_SOURCE

u64

Considering how 8 bytes is obviously overkill for the info here I'm wondering if we should consider maintaining the ability to pack more data into the same section in the future.

Maybe u64 source_info:4, reserved:60 bitfields could keep it easier to combine with more data later? (this is a convention taken from from uapi/linux/perf_event.h

Quoted 6 lines
include/uapi/drm/i915_drm.h +0 -1
         *     { u32 oa_report[]; } && DRM_I915_PERF_PROP_SAMPLE_OA
         * };
         */
--

1.9.1

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#5 lore raw

PATCH 3/8 drm/i915: Framework for capturing command stream based OA reports

In reply to [email protected]

From: Sourab Gupta <[email protected]>

This patch introduces a framework to capture OA counter reports associated with Render command stream. We can then associate the reports captured through this mechanism with their corresponding context id's. This can be further extended to associate any other metadata information with the corresponding samples (since the association with Render command stream gives us the ability to capture these information while inserting the corresponding capture commands into the command stream).

The OA reports generated in this way are associated with a corresponding workload, and thus can be used the delimit the workload (i.e. sample the counters at the workload boundaries), within an ongoing stream of periodic counter snapshots.

There may be usecases wherein we need more than periodic OA capture mode which is supported currently. This mode is primarily used for two usecases:
- Ability to capture system wide metrics, alongwith the ability to map the reports back to individual contexts (particularly for HSW).
- Ability to inject tags for work, into the reports. This provides visibility into the multiple stages of work within single context.

The userspace will be able to distinguish between the periodic and CS based OA reports by the virtue of source_info sample field.

The command MI_REPORT_PERF_COUNT can be used to capture snapshots of OA counters, and is inserted at BB boundaries.
The data thus captured will be stored in a separate buffer, which will be different from the buffer used otherwise for periodic OA capture mode. The metadata information pertaining to snapshot is maintained in a list, which also has offsets into the gem buffer object per captured snapshot. In order to track whether the gpu has completed processing the node, a field pertaining to corresponding gem request is added, which is tracked for completion of the command.

Both periodic and RCS based reports are associated with a single stream (corresponding to render engine), and it is expected to have the samples in the sequential order according to their timestamps. Now, since these reports are collected in separate buffers, these are merge sorted at the time of forwarding to userspace during the read call.

v2: Aligining with the non-perf interface (custom drm ioctl based). Also, few related patches are squashed together for better readability

---
 drivers/gpu/drm/i915/i915_drv.h            |  88 ++-
 drivers/gpu/drm/i915/i915_gem_execbuffer.c |   4 +
 drivers/gpu/drm/i915/i915_perf.c           | 888 ++++++++++++++++++++++++-----
 include/uapi/drm/i915_drm.h                |  15 +
 4 files changed, 861 insertions(+), 134 deletions(-)
drivers/gpu/drm/i915/i915_drv.h +83 -5
index ef104ff5..0b70052 100644
@@ -1972,6 +1972,23 @@ struct i915_perf_stream_ops {
 	 * The stream will always be disabled before this is called.
 	 */
 	void (*destroy)(struct i915_perf_stream *stream);
+
+	/*
+	 * @command_stream_hook: Emit the commands in the command streamer
+	 * for a particular gpu engine.
+	 *
+	 * The commands are inserted to capture the perf sample data at
+	 * specific points during workload execution, such as before and after
+	 * the batch buffer.
+	 */
+	void (*command_stream_hook)(struct i915_perf_stream *stream,
+				struct drm_i915_gem_request *request);
+};
+
+enum i915_perf_stream_state {
+	I915_PERF_STREAM_DISABLED,
+	I915_PERF_STREAM_ENABLE_IN_PROGRESS,
+	I915_PERF_STREAM_ENABLED,
 };
 
 /**
@@ -1989,6 +2006,10 @@ struct i915_perf_stream {
 	struct list_head link;
 
 	/**
+	 * @engine: GPU engine associated with this particular stream
+	 */
+	enum intel_engine_id engine;
+	/**
 	 * @sample_flags: Flags representing the `DRM_I915_PERF_PROP_SAMPLE_*`
 	 * properties given when opening a stream, representing the contents
 	 * of a single sample as read() by userspace.
@@ -2009,11 +2030,25 @@ struct i915_perf_stream {
 	struct i915_gem_context *ctx;
 
 	/**
-	 * @enabled: Whether the stream is currently enabled, considering
-	 * whether the stream was opened in a disabled state and based
-	 * on `I915_PERF_IOCTL_ENABLE` and `I915_PERF_IOCTL_DISABLE` calls.
+	 * @state: Current stream state, which can be either disabled, enabled,
+	 * or enable_in_progress, while considering whether the stream was
+	 * opened in a disabled state and based on `I915_PERF_IOCTL_ENABLE` and
+	 * `I915_PERF_IOCTL_DISABLE` calls.
 	 */
-	bool enabled;
+	enum i915_perf_stream_state state;
+
+	/**
+	 * @cs_mode: Whether command stream based perf sample collection is
+	 * enabled for this stream
+	 */
+	bool cs_mode;
+
+	/**
+	 * @last_request: Contains an RCU guarded pointer to the last request
+	 * associated with the stream, when command stream mode is enabled for
+	 * the stream.
+	 */
+	struct i915_gem_active last_request;
 
 	/**
 	 * @ops: The callbacks providing the implementation of this specific
@@ -2074,7 +2109,8 @@ struct i915_oa_ops {
 	int (*read)(struct i915_perf_stream *stream,
 		    char __user *buf,
 		    size_t count,
-		    size_t *offset);
+		    size_t *offset,
+		    u32 ts);
 
 	/**
 	 * @oa_buffer_check: Check for OA buffer data + update tail
@@ -2093,6 +2129,36 @@ struct i915_oa_ops {
 	bool (*oa_buffer_check)(struct drm_i915_private *dev_priv);
 };
 
+/*
+ * i915_perf_cs_sample - Sample element to hold info about a single perf
+ * sample data associated with a particular GPU command stream.
+ */
+struct i915_perf_cs_sample {
+	/**
+	 * @link: Links the sample into ``&drm_i915_private->perf.cs_samples``
+	 */
+	struct list_head link;
+
+	/**
+	 * @request: Gem request associated with the sample. The commands to
+	 * capture the perf metrics are inserted into the command streamer in
+	 * context of this request.
+	 */
+	struct drm_i915_gem_request *request;
+
+	/**
+	 * @offset: Offset into ``&drm_i915_private->perf.command_stream_buf``
+	 * where the perf metrics will be collected, when the commands inserted
+	 * into the command stream are executed by GPU.
+	 */
+	u32 offset;
+
+	/**
+	 * @ctx_id: Context ID associated with this perf sample
+	 */
+	u32 ctx_id;
+};
+
 struct intel_cdclk_state {
 	unsigned int cdclk, vco, ref;
 };
@@ -2421,6 +2487,8 @@ struct drm_i915_private {
 		struct ctl_table_header *sysctl_header;
 
 		struct mutex lock;
+
+		struct mutex streams_lock;
 		struct list_head streams;
 
 		spinlock_t hook_lock;
@@ -2532,6 +2600,15 @@ struct drm_i915_private {
 			const struct i915_oa_format *oa_formats;
 			int n_builtin_sets;
 		} oa;
+
+		/* Command stream based perf data buffer */
+		struct {
+			struct i915_vma *vma;
+			u8 *vaddr;
+		} command_stream_buf;
+
+		struct list_head cs_samples;
+		spinlock_t sample_lock;
 	} perf;
 
 	/* Abstract the submission mechanism (legacy ringbuffer or execlists) away */
@@ -3593,6 +3670,7 @@ void i915_oa_init_reg_state(struct intel_engine_cs *engine,
 void i915_oa_update_reg_state(struct intel_engine_cs *engine,
 			      struct i915_gem_context *ctx,
 			      uint32_t *reg_state);
+void i915_perf_command_stream_hook(struct drm_i915_gem_request *req);
 
 /* i915_gem_evict.c */
 int __must_check i915_gem_evict_something(struct i915_address_space *vm,
drivers/gpu/drm/i915/i915_gem_execbuffer.c +4 -0
index aa75ea2..7af32c97 100644
@@ -1441,12 +1441,16 @@ static void eb_export_fence(struct drm_i915_gem_object *obj,
 	if (exec_len == 0)
 		exec_len = params->batch->size - params->args_batch_start_offset;
 
+	i915_perf_command_stream_hook(params->request);
+
 	ret = params->engine->emit_bb_start(params->request,
 					    exec_start, exec_len,
 					    params->dispatch_flags);
 	if (ret)
 		return ret;
 
+	i915_perf_command_stream_hook(params->request);
+
 	i915_gem_execbuffer_move_to_active(vmas, params->request);
 
 	return 0;
drivers/gpu/drm/i915/i915_perf.c +759 -129

1,244 lines not on this page · raw message

include/uapi/drm/i915_drm.h +15 -0
index c597e36..4cbefc8 100644
@@ -1314,6 +1314,7 @@ enum drm_i915_oa_format {
 
 enum drm_i915_perf_oa_event_source {
 	I915_PERF_OA_EVENT_SOURCE_PERIODIC,
+	I915_PERF_OA_EVENT_SOURCE_RCS,
 
 	I915_PERF_OA_EVENT_SOURCE_MAX	/* non-ABI */
 };
@@ -1359,6 +1360,19 @@ enum drm_i915_perf_property_id {
 	 */
 	DRM_I915_PERF_PROP_SAMPLE_OA_SOURCE,
 
+	/**
+	 * The value of this property specifies the GPU engine for which
+	 * the samples need to be collected. Specifying this property also
+	 * implies the command stream based sample collection.
+	 */
+	DRM_I915_PERF_PROP_ENGINE,
+
+	/**
+	 * The value of this property set to 1 requests inclusion of context ID
+	 * in the perf sample data.
+	 */
+	DRM_I915_PERF_PROP_SAMPLE_CTX_ID,
+
 	DRM_I915_PERF_PROP_MAX /* non-ABI */
 };
 
@@ -1425,6 +1439,7 @@ enum drm_i915_perf_record_type {
 	 *     struct drm_i915_perf_record_header header;
 	 *
 	 *     { u32 source_info; } && DRM_I915_PERF_PROP_SAMPLE_OA_SOURCE
+	 *     { u32 ctx_id; } && DRM_I915_PERF_PROP_SAMPLE_CTX_ID
 	 *     { u32 oa_report[]; } && DRM_I915_PERF_PROP_SAMPLE_OA
 	 * };
 	 */
Signature · 6 lines
1.9.1

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#6 lore raw

Re: [PATCH 3/8] drm/i915: Framework for capturing command stream based OA reports

Quoted 13 lines · On Thu, Mar 16, 2017 at 11:44:10AM +0530, [email protected] wrote:
drivers/gpu/drm/i915/i915_drv.h +1 -0
@@ -3593,6 +3670,7 @@ void i915_oa_init_reg_state(struct intel_engine_cs *engine,
 void i915_oa_update_reg_state(struct intel_engine_cs *engine,
 			      struct i915_gem_context *ctx,
 			      uint32_t *reg_state);
+void i915_perf_command_stream_hook(struct drm_i915_gem_request *req);
 
 /* i915_gem_evict.c */
 int __must_check i915_gem_evict_something(struct i915_address_space *vm,
drivers/gpu/drm/i915/i915_gem_execbuffer.c +0 -0
index aa75ea2..7af32c97 100644

drivers/gpu/drm/i915/i915_gem_execbuffer.c static void eb_export_fence(struct drm_i915_gem_object *obj, · line 1444

Quoted 4 lines
drivers/gpu/drm/i915/i915_gem_execbuffer.c +1 -0
@@ -1441,12 +1441,16 @@ static void eb_export_fence(struct drm_i915_gem_object *obj,
 	if (exec_len == 0)
 		exec_len = params->batch->size - params->args_batch_start_offset;
 
+	i915_perf_command_stream_hook(params->request);

Could you have named it anything more cyptic and inconsistent?

Quite clearly this can fail, so justify the non handling of errors.

DRM_ERROR is not error handling, it is an indication that this patch isn't ready.

drivers/gpu/drm/i915/i915_gem_execbuffer.c static void eb_export_fence(struct drm_i915_gem_object *obj, · line 1454

Quoted 10 lines
drivers/gpu/drm/i915/i915_gem_execbuffer.c +10 -0
+void i915_perf_command_stream_hook(struct drm_i915_gem_request *request)
+{
+	struct intel_engine_cs *engine = request->engine;
+	struct drm_i915_private *dev_priv = engine->i915;
+	struct i915_perf_stream *stream;
+
+	if (!dev_priv->perf.initialized)
+		return;
+
+	mutex_lock(&dev_priv->perf.streams_lock);

Justify a new global lock very, very carefully on execbuf.

Quoted 24 lines
drivers/gpu/drm/i915/i915_gem_execbuffer.c +23 -0
+	list_for_each_entry(stream, &dev_priv->perf.streams, link) {
+		if ((stream->state == I915_PERF_STREAM_ENABLED) &&
+					stream->cs_mode)
+			stream->ops->command_stream_hook(stream, request);
+	}
+	mutex_unlock(&dev_priv->perf.streams_lock);
+}
 
+static void i915_perf_command_stream_hook_oa(struct i915_perf_stream *stream,
+					struct drm_i915_gem_request *request)
+{
+	struct drm_i915_private *dev_priv = request->i915;
+	struct i915_gem_context *ctx = request->ctx;
+	struct i915_perf_cs_sample *sample;
+	u32 addr = 0;
+	u32 cmd, *cs;
+
+	sample = kzalloc(sizeof(*sample), GFP_KERNEL);
+	if (sample == NULL) {
+		DRM_ERROR("Perf sample alloc failed\n");
+		return;
+	}
+
+	cs = intel_ring_begin(request, 4);

drivers/gpu/drm/i915/i915_gem_execbuffer.c static void eb_export_fence(struct drm_i915_gem_object *obj,

Quoted 10 lines
drivers/gpu/drm/i915/i915_gem_execbuffer.c +9 -0
+	if (IS_ERR(cs)) {
+		kfree(sample);
+		return;
+	}
+
+	sample->ctx_id = ctx->hw_id;
+	i915_gem_request_assign(&sample->request, request);
 
+
+	i915_gem_active_set(&stream->last_request, request);

Hint, you don't need you own request trap.
-Chris

Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#7 lore raw

Re: [PATCH 3/8] drm/i915: Framework for capturing command stream based OA reports

Quoted 17 lines · Chris Wilson
Quoted 16 lines · On Thu, Mar 16, 2017 at 11:44:10AM +0530, [email protected] wrote:
drivers/gpu/drm/i915/i915_drv.h +1 -0
@@ -3593,6 +3670,7 @@ void i915_oa_init_reg_state(struct intel_engine_cs *engine,
 void i915_oa_update_reg_state(struct intel_engine_cs *engine,
 			      struct i915_gem_context *ctx,
 			      uint32_t *reg_state);
+void i915_perf_command_stream_hook(struct drm_i915_gem_request *req);
 
 /* i915_gem_evict.c */
 int __must_check i915_gem_evict_something(struct i915_address_space *vm,
drivers/gpu/drm/i915/i915_gem_execbuffer.c +0 -0
index aa75ea2..7af32c97 100644
@@ -1441,12 +1441,16 @@ static void eb_export_fence(struct drm_i915_gem_object *obj,
 	if (exec_len == 0)
 		exec_len = params->batch->size - params->args_batch_start_offset;
 
Quoted 3 lines

drivers/gpu/drm/i915/i915_gem_execbuffer.c static void eb_export_fence(struct drm_i915_gem_object *obj, · line 1444

Quoted 1 line
drivers/gpu/drm/i915/i915_gem_execbuffer.c +1 -0
+	i915_perf_command_stream_hook(params->request);

Could you have named it anything more cyptic and inconsistent?

Sorry. Would 'i915_perf_capture_metrics' work? Can you please suggest an appropriate name otherwise.

Quoted 5 lines

Quite clearly this can fail, so justify the non handling of errors.

DRM_ERROR is not error handling, it is an indication that this patch isn't ready.

Well, the intent was to have minimal effect to execbuf normal routine, even if we fail. But, I guess I'm mistaken.
I'll rectify this to handle the case, if perf callback fails.

Quoted 13 lines

drivers/gpu/drm/i915/i915_gem_execbuffer.c static void eb_export_fence(struct drm_i915_gem_object *obj, · line 1454

Quoted 11 lines
drivers/gpu/drm/i915/i915_gem_execbuffer.c +10 -0
+void i915_perf_command_stream_hook(struct drm_i915_gem_request *request)
+{
+	struct intel_engine_cs *engine = request->engine;
+	struct drm_i915_private *dev_priv = engine->i915;
+	struct i915_perf_stream *stream;
+
+	if (!dev_priv->perf.initialized)
+		return;
+
+	mutex_lock(&dev_priv->perf.streams_lock);

Justify a new global lock very, very carefully on execbuf.

The lock introduced here is to protect the the perf.streams list against addition/deletion while we're processing the list during execbuf call. The other places where the mutex is taken is when a new stream is being created (using perf_open ioctl) or a stream is being destroyed (perf_release ioctl), which just protect the list_add and list_del into the perf.streams list.
As such, there should not be much impact on execbuf path. Does this seem reasonable?

Quoted 33 lines
Quoted 32 lines
drivers/gpu/drm/i915/i915_gem_execbuffer.c +30 -0
+	list_for_each_entry(stream, &dev_priv->perf.streams, link) {
+		if ((stream->state == I915_PERF_STREAM_ENABLED) &&
+					stream->cs_mode)
+			stream->ops->command_stream_hook(stream, request);
+	}
+	mutex_unlock(&dev_priv->perf.streams_lock);
+}
 
+static void i915_perf_command_stream_hook_oa(struct i915_perf_stream *stream,
+					struct drm_i915_gem_request *request)
+{
+	struct drm_i915_private *dev_priv = request->i915;
+	struct i915_gem_context *ctx = request->ctx;
+	struct i915_perf_cs_sample *sample;
+	u32 addr = 0;
+	u32 cmd, *cs;
+
+	sample = kzalloc(sizeof(*sample), GFP_KERNEL);
+	if (sample == NULL) {
+		DRM_ERROR("Perf sample alloc failed\n");
+		return;
+	}
+
+	cs = intel_ring_begin(request, 4);
+	if (IS_ERR(cs)) {
+		kfree(sample);
+		return;
+	}
+
+	sample->ctx_id = ctx->hw_id;
+	i915_gem_request_assign(&sample->request, request);
Quoted 4 lines

drivers/gpu/drm/i915/i915_gem_execbuffer.c

Quoted 2 lines
drivers/gpu/drm/i915/i915_gem_execbuffer.c +2 -0
+
+	i915_gem_active_set(&stream->last_request, request);

Hint, you don't need you own request trap.

Sorry, I didn't understand you fully here. I'm storing the reference to the last active request associated with the stream inside the last_request 'gem_active' field. Do I need to do something differently here?

Quoted 2 lines

-Chris

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#8 lore raw

Re: [PATCH 3/8] drm/i915: Framework for capturing command stream based OA reports

Quoted 18 lines · sourab gupta
Quoted 17 lines · Chris Wilson
Quoted 16 lines · On Thu, Mar 16, 2017 at 11:44:10AM +0530, [email protected] wrote:
drivers/gpu/drm/i915/i915_drv.h +1 -0
@@ -3593,6 +3670,7 @@ void i915_oa_init_reg_state(struct intel_engine_cs *engine,
 void i915_oa_update_reg_state(struct intel_engine_cs *engine,
 			      struct i915_gem_context *ctx,
 			      uint32_t *reg_state);
+void i915_perf_command_stream_hook(struct drm_i915_gem_request *req);
 
 /* i915_gem_evict.c */
 int __must_check i915_gem_evict_something(struct i915_address_space *vm,
drivers/gpu/drm/i915/i915_gem_execbuffer.c +0 -0
index aa75ea2..7af32c97 100644
@@ -1441,12 +1441,16 @@ static void eb_export_fence(struct drm_i915_gem_object *obj,
 	if (exec_len == 0)
 		exec_len = params->batch->size - params->args_batch_start_offset;
 
Quoted 6 lines
Quoted 3 lines

drivers/gpu/drm/i915/i915_gem_execbuffer.c static void eb_export_fence(struct drm_i915_gem_object *obj, · line 1444

Quoted 1 line
drivers/gpu/drm/i915/i915_gem_execbuffer.c +1 -0
+	i915_perf_command_stream_hook(params->request);

Could you have named it anything more cyptic and inconsistent?

Sorry. Would 'i915_perf_capture_metrics' work? Can you please suggest an appropriate name otherwise.

The verb we use for writting into the command stream is emit. So i915_perf_emit_samples() (emit_record record is clumsy as it is not clear whether it is the verb or noun).

Quoted 9 lines
Quoted 5 lines

Quite clearly this can fail, so justify the non handling of errors.

DRM_ERROR is not error handling, it is an indication that this patch isn't ready.

Well, the intent was to have minimal effect to execbuf normal routine, even if we fail. But, I guess I'm mistaken.
I'll rectify this to handle the case, if perf callback fails.

That all depends on whether or not you can handle the unbalanced metrics. If simply missing a report is fine, then just kill the DRM_ERROR.

The bigger question is whether the following emit can to fail -- once the batch is in the request, no more failures are tolerable. You have to preallocate reserved space.

Don't you need a flush before the emit following the batch?

Quoted 13 lines
Quoted 4 lines
Quoted 4 lines
drivers/gpu/drm/i915/i915_gem_execbuffer.c +4 -0
+void i915_perf_command_stream_hook(struct drm_i915_gem_request *request)
+{
+	struct intel_engine_cs *engine = request->engine;
+	struct drm_i915_private *dev_priv = engine->i915;
Quoted 8 lines

drivers/gpu/drm/i915/i915_gem_execbuffer.c static void eb_export_fence(struct drm_i915_gem_object *obj, · line 1454

Quoted 6 lines
drivers/gpu/drm/i915/i915_gem_execbuffer.c +6 -0
+	struct i915_perf_stream *stream;
+
+	if (!dev_priv->perf.initialized)
+		return;
+
+	mutex_lock(&dev_priv->perf.streams_lock);

Justify a new global lock very, very carefully on execbuf.

Quoted 8 lines

The lock introduced here is to protect the the perf.streams list against addition/deletion while we're processing the list during execbuf call. The other places where the mutex is taken is when a new stream is being created (using perf_open ioctl) or a stream is being destroyed (perf_release ioctl), which just protect the list_add and list_del into the perf.streams list.
As such, there should not be much impact on execbuf path. Does this seem reasonable?

It doesn't sound like it needs a mutex, which will simplify the other users as well (if switched to, for example, an RCU protected list).

Quoted 32 lines
Quoted 32 lines
Quoted 31 lines
drivers/gpu/drm/i915/i915_gem_execbuffer.c +30 -0
+	list_for_each_entry(stream, &dev_priv->perf.streams, link) {
+		if ((stream->state == I915_PERF_STREAM_ENABLED) &&
+					stream->cs_mode)
+			stream->ops->command_stream_hook(stream, request);
+	}
+	mutex_unlock(&dev_priv->perf.streams_lock);
+}
 
+static void i915_perf_command_stream_hook_oa(struct i915_perf_stream *stream,
+					struct drm_i915_gem_request *request)
+{
+	struct drm_i915_private *dev_priv = request->i915;
+	struct i915_gem_context *ctx = request->ctx;
+	struct i915_perf_cs_sample *sample;
+	u32 addr = 0;
+	u32 cmd, *cs;
+
+	sample = kzalloc(sizeof(*sample), GFP_KERNEL);
+	if (sample == NULL) {
+		DRM_ERROR("Perf sample alloc failed\n");
+		return;
+	}
+
+	cs = intel_ring_begin(request, 4);
+	if (IS_ERR(cs)) {
+		kfree(sample);
+		return;
+	}
+
+	sample->ctx_id = ctx->hw_id;
+	i915_gem_request_assign(&sample->request, request);
Quoted 8 lines
Quoted 4 lines

drivers/gpu/drm/i915/i915_gem_execbuffer.c

Quoted 2 lines
drivers/gpu/drm/i915/i915_gem_execbuffer.c +2 -0
+
+	i915_gem_active_set(&stream->last_request, request);

Hint, you don't need you own request trap.

Sorry, I didn't understand you fully here. I'm storing the reference to the last active request associated with the stream inside the last_request 'gem_active' field. Do I need to do something differently here?

It's the duplication.
-Chris

Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#9 lore raw

Re: [PATCH 3/8] drm/i915: Framework for capturing command stream based OA reports

Quoted 23 lines · Chris Wilson
Quoted 22 lines · sourab gupta
Quoted 20 lines · Chris Wilson
Quoted 13 lines · On Thu, Mar 16, 2017 at 11:44:10AM +0530, [email protected] wrote:
drivers/gpu/drm/i915/i915_drv.h +1 -0
@@ -3593,6 +3670,7 @@ void i915_oa_init_reg_state(struct intel_engine_cs *engine,
 void i915_oa_update_reg_state(struct intel_engine_cs *engine,
 			      struct i915_gem_context *ctx,
 			      uint32_t *reg_state);
+void i915_perf_command_stream_hook(struct drm_i915_gem_request *req);
 
 /* i915_gem_evict.c */
 int __must_check i915_gem_evict_something(struct i915_address_space *vm,
drivers/gpu/drm/i915/i915_gem_execbuffer.c +0 -0
index aa75ea2..7af32c97 100644

drivers/gpu/drm/i915/i915_gem_execbuffer.c static void eb_export_fence(struct drm_i915_gem_object *obj, · line 1444

Quoted 4 lines
drivers/gpu/drm/i915/i915_gem_execbuffer.c +1 -0
@@ -1441,12 +1441,16 @@ static void eb_export_fence(struct drm_i915_gem_object *obj,
 	if (exec_len == 0)
 		exec_len = params->batch->size - params->args_batch_start_offset;
 
+	i915_perf_command_stream_hook(params->request);

Could you have named it anything more cyptic and inconsistent?

Quoted 7 lines
Quoted 2 lines

Sorry. Would 'i915_perf_capture_metrics' work? Can you please suggest an appropriate name otherwise.

The verb we use for writting into the command stream is emit. So i915_perf_emit_samples() (emit_record record is clumsy as it is not clear whether it is the verb or noun).

Thanks for suggesting. I'll use 'i915_perf_emit_samples' here.

Quoted 14 lines
Quoted 9 lines
Quoted 5 lines

Quite clearly this can fail, so justify the non handling of errors.

DRM_ERROR is not error handling, it is an indication that this patch isn't ready.

Well, the intent was to have minimal effect to execbuf normal routine, even if we fail. But, I guess I'm mistaken.
I'll rectify this to handle the case, if perf callback fails.

That all depends on whether or not you can handle the unbalanced metrics. If simply missing a report is fine, then just kill the DRM_ERROR.

Quoted 6 lines

The bigger question is whether the following emit can to fail -- once the batch is in the request, no more failures are tolerable. You have to preallocate reserved space.

Don't you need a flush before the emit following the batch?

Ok. So, that would mean that we have to first of all reserve the space required by two 'perf_emit_samples' calls, so that we can't fail for the lack of space in the emit following the batch.
Probably, I could pass an additional boolean parameter 'reserve_space' set to true in the first call, which would reserve the space for both emit_samples() calls (through intel_ring_begin)?

Would it still need the flush before the emit following the batch? Ideally, the metrics should be collected as close to batch as possible. So, if there are cache flush/tlb invalidate commands, it would introduce some lag between the batch and following emit_samples command. Sorry, I'm not able to gauge the need for flush here. I can understand it's needed before the batch is programmed for flushing the cache/TLB entries for the new workload to be submitted. But for the Sample_emit commands, which generally only capture OA/timestamp/mmio metrics, is this still required?

Quoted 22 lines
Quoted 13 lines
Quoted 4 lines
Quoted 4 lines
drivers/gpu/drm/i915/i915_gem_execbuffer.c +4 -0
+void i915_perf_command_stream_hook(struct drm_i915_gem_request *request)
+{
+	struct intel_engine_cs *engine = request->engine;
+	struct drm_i915_private *dev_priv = engine->i915;
Quoted 8 lines

drivers/gpu/drm/i915/i915_gem_execbuffer.c static void eb_export_fence(struct drm_i915_gem_object *obj, · line 1454

Quoted 6 lines
drivers/gpu/drm/i915/i915_gem_execbuffer.c +6 -0
+	struct i915_perf_stream *stream;
+
+	if (!dev_priv->perf.initialized)
+		return;
+
+	mutex_lock(&dev_priv->perf.streams_lock);

Justify a new global lock very, very carefully on execbuf.

Quoted 8 lines

The lock introduced here is to protect the the perf.streams list against addition/deletion while we're processing the list during execbuf call. The other places where the mutex is taken is when a new stream is being created (using perf_open ioctl) or a stream is being destroyed (perf_release ioctl), which just protect the list_add and list_del into the perf.streams list.
As such, there should not be much impact on execbuf path. Does this seem reasonable?

Quoted 2 lines

It doesn't sound like it needs a mutex, which will simplify the other users as well (if switched to, for example, an RCU protected list).

Ok. Sounds reasonable, I'll switch to using an RCU protected list here.

Quoted 37 lines
Quoted 37 lines
Quoted 37 lines
Quoted 25 lines
drivers/gpu/drm/i915/i915_gem_execbuffer.c +23 -0
+	list_for_each_entry(stream, &dev_priv->perf.streams, link) {
+		if ((stream->state == I915_PERF_STREAM_ENABLED) &&
+					stream->cs_mode)
+			stream->ops->command_stream_hook(stream, request);
+	}
+	mutex_unlock(&dev_priv->perf.streams_lock);
+}
 
+static void i915_perf_command_stream_hook_oa(struct i915_perf_stream *stream,
+					struct drm_i915_gem_request *request)
+{
+	struct drm_i915_private *dev_priv = request->i915;
+	struct i915_gem_context *ctx = request->ctx;
+	struct i915_perf_cs_sample *sample;
+	u32 addr = 0;
+	u32 cmd, *cs;
+
+	sample = kzalloc(sizeof(*sample), GFP_KERNEL);
+	if (sample == NULL) {
+		DRM_ERROR("Perf sample alloc failed\n");
+		return;
+	}
+
+	cs = intel_ring_begin(request, 4);

drivers/gpu/drm/i915/i915_gem_execbuffer.c static void eb_export_fence(struct drm_i915_gem_object *obj,

Quoted 10 lines
drivers/gpu/drm/i915/i915_gem_execbuffer.c +9 -0
+	if (IS_ERR(cs)) {
+		kfree(sample);
+		return;
+	}
+
+	sample->ctx_id = ctx->hw_id;
+	i915_gem_request_assign(&sample->request, request);
 
+
+	i915_gem_active_set(&stream->last_request, request);

Hint, you don't need you own request trap.

Quoted 6 lines
Quoted 4 lines

Sorry, I didn't understand you fully here. I'm storing the reference to the last active request associated with the stream inside the last_request 'gem_active' field. Do I need to do something differently here?

It's the duplication.

Are you suggesting that request_assign() and active_set() is duplication?
Actually, I'm using the last_request active tracker for the purpose of waiting on last request to complete, whenever the need.
But still I need to get reference for every request for which the commands for collection of metrics are emitted. This is because I need to check for their completion before collecting the associated metrics. This is done inside 'append_command_stream_samples' function, which also takes care of releasing the reference taken here.
Am I missing something w.r.t. the active_set() functionality?

Quoted 2 lines

-Chris

Thanks,
Sourab

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#10 lore raw

Re: [PATCH 3/8] drm/i915: Framework for capturing command stream based OA reports

Quoted 54 lines · sourab gupta
Quoted 23 lines · Chris Wilson
Quoted 22 lines · sourab gupta
Quoted 20 lines · Chris Wilson
Quoted 13 lines · On Thu, Mar 16, 2017 at 11:44:10AM +0530, [email protected] wrote:
drivers/gpu/drm/i915/i915_drv.h +1 -0
@@ -3593,6 +3670,7 @@ void i915_oa_init_reg_state(struct intel_engine_cs *engine,
 void i915_oa_update_reg_state(struct intel_engine_cs *engine,
 			      struct i915_gem_context *ctx,
 			      uint32_t *reg_state);
+void i915_perf_command_stream_hook(struct drm_i915_gem_request *req);
 
 /* i915_gem_evict.c */
 int __must_check i915_gem_evict_something(struct i915_address_space *vm,
drivers/gpu/drm/i915/i915_gem_execbuffer.c +0 -0
index aa75ea2..7af32c97 100644

drivers/gpu/drm/i915/i915_gem_execbuffer.c static void eb_export_fence(struct drm_i915_gem_object *obj, · line 1444

Quoted 4 lines
drivers/gpu/drm/i915/i915_gem_execbuffer.c +1 -0
@@ -1441,12 +1441,16 @@ static void eb_export_fence(struct drm_i915_gem_object *obj,
 	if (exec_len == 0)
 		exec_len = params->batch->size - params->args_batch_start_offset;
 
+	i915_perf_command_stream_hook(params->request);

Could you have named it anything more cyptic and inconsistent?

Quoted 7 lines
Quoted 2 lines

Sorry. Would 'i915_perf_capture_metrics' work? Can you please suggest an appropriate name otherwise.

The verb we use for writting into the command stream is emit. So i915_perf_emit_samples() (emit_record record is clumsy as it is not clear whether it is the verb or noun).

Thanks for suggesting. I'll use 'i915_perf_emit_samples' here.

Quoted 14 lines
Quoted 9 lines
Quoted 5 lines

Quite clearly this can fail, so justify the non handling of errors.

DRM_ERROR is not error handling, it is an indication that this patch isn't ready.

Well, the intent was to have minimal effect to execbuf normal routine, even if we fail. But, I guess I'm mistaken.
I'll rectify this to handle the case, if perf callback fails.

That all depends on whether or not you can handle the unbalanced metrics. If simply missing a report is fine, then just kill the DRM_ERROR.

Quoted 6 lines

The bigger question is whether the following emit can to fail -- once the batch is in the request, no more failures are tolerable. You have to preallocate reserved space.

Don't you need a flush before the emit following the batch?

Quoted 6 lines

Ok. So, that would mean that we have to first of all reserve the space required by two 'perf_emit_samples' calls, so that we can't fail for the lack of space in the emit following the batch.
Probably, I could pass an additional boolean parameter 'reserve_space' set to true in the first call, which would reserve the space for both emit_samples() calls (through intel_ring_begin)?

Hmm, yes, you can just tweak the request->reserved_space in the first call to preallocate some space (and make it remains available) for the second.

perf_emit_samples(rq, bool preallocate) {
	/* then in each sample callback */
	cs_len = foo;
	if (preallocate)
		rq->reserved_space += cs_len;
	else
		rq->reserved_space -= cs_len;
	cs = intel_ring_begin(rq, cs_len);
}
Quoted 9 lines

Would it still need the flush before the emit following the batch? Ideally, the metrics should be collected as close to batch as possible. So, if there are cache flush/tlb invalidate commands, it would introduce some lag between the batch and following emit_samples command. Sorry, I'm not able to gauge the need for flush here. I can understand it's needed before the batch is programmed for flushing the cache/TLB entries for the new workload to be submitted. But for the Sample_emit commands, which generally only capture OA/timestamp/mmio metrics, is this still required?

Depends on the desired accuracy. If you want your metrics sampled after the user pipeline is completed, you need a stall at least, a flush if your metrics include e.g. fragment data. If you want samples taken in whilst the user's batch is still executing, then no.

Quoted 22 lines
Quoted 22 lines
Quoted 21 lines
Quoted 4 lines
Quoted 4 lines
drivers/gpu/drm/i915/i915_gem_execbuffer.c +4 -0
+void i915_perf_command_stream_hook(struct drm_i915_gem_request *request)
+{
+	struct intel_engine_cs *engine = request->engine;
+	struct drm_i915_private *dev_priv = engine->i915;
Quoted 8 lines

drivers/gpu/drm/i915/i915_gem_execbuffer.c static void eb_export_fence(struct drm_i915_gem_object *obj, · line 1454

Quoted 6 lines
drivers/gpu/drm/i915/i915_gem_execbuffer.c +6 -0
+	struct i915_perf_stream *stream;
+
+	if (!dev_priv->perf.initialized)
+		return;
+
+	mutex_lock(&dev_priv->perf.streams_lock);

Justify a new global lock very, very carefully on execbuf.

The lock introduced here is to protect the the perf.streams list against addition/deletion while we're processing the list during execbuf call. The other places where the mutex is taken is when a new stream is being created (using perf_open ioctl) or a stream is being destroyed (perf_release ioctl), which just protect the list_add and list_del into the perf.streams list.
As such, there should not be much impact on execbuf path. Does this seem reasonable?

Quoted 4 lines
Quoted 2 lines

It doesn't sound like it needs a mutex, which will simplify the other users as well (if switched to, for example, an RCU protected list).

Ok. Sounds reasonable, I'll switch to using an RCU protected list here.

(I may be overthinking this, but it still seems overkill and made the timer callback uglier than expected.)

Quoted 43 lines
Quoted 36 lines
Quoted 36 lines
Quoted 36 lines
Quoted 24 lines
drivers/gpu/drm/i915/i915_gem_execbuffer.c +23 -0
+	list_for_each_entry(stream, &dev_priv->perf.streams, link) {
+		if ((stream->state == I915_PERF_STREAM_ENABLED) &&
+					stream->cs_mode)
+			stream->ops->command_stream_hook(stream, request);
+	}
+	mutex_unlock(&dev_priv->perf.streams_lock);
+}
 
+static void i915_perf_command_stream_hook_oa(struct i915_perf_stream *stream,
+					struct drm_i915_gem_request *request)
+{
+	struct drm_i915_private *dev_priv = request->i915;
+	struct i915_gem_context *ctx = request->ctx;
+	struct i915_perf_cs_sample *sample;
+	u32 addr = 0;
+	u32 cmd, *cs;
+
+	sample = kzalloc(sizeof(*sample), GFP_KERNEL);
+	if (sample == NULL) {
+		DRM_ERROR("Perf sample alloc failed\n");
+		return;
+	}
+
+	cs = intel_ring_begin(request, 4);

drivers/gpu/drm/i915/i915_gem_execbuffer.c static void eb_export_fence(struct drm_i915_gem_object *obj,

Quoted 10 lines
drivers/gpu/drm/i915/i915_gem_execbuffer.c +9 -0
+	if (IS_ERR(cs)) {
+		kfree(sample);
+		return;
+	}
+
+	sample->ctx_id = ctx->hw_id;
+	i915_gem_request_assign(&sample->request, request);
 
+
+	i915_gem_active_set(&stream->last_request, request);

Hint, you don't need you own request trap.

Quoted 6 lines
Quoted 4 lines

Sorry, I didn't understand you fully here. I'm storing the reference to the last active request associated with the stream inside the last_request 'gem_active' field. Do I need to do something differently here?

It's the duplication.

Quoted 7 lines

Are you suggesting that request_assign() and active_set() is duplication?
Actually, I'm using the last_request active tracker for the purpose of waiting on last request to complete, whenever the need.
But still I need to get reference for every request for which the commands for collection of metrics are emitted. This is because I need to check for their completion before collecting the associated metrics.

Missed the sample / stream difference. request_assign means update the request field, had you used

	sample->request = i915_gem_request_get(request)

it would have been clearer.

Note that the requests are not ordered for the stream, so you cannot track the last_request so easily.

Quoted 3 lines

This is done inside 'append_command_stream_samples' function, which also takes care of releasing the reference taken here.
Am I missing something w.r.t. the active_set() functionality?

I was mostly thrown by the idea that you were reassigning requests, which history has shown to be a bad idea (hence i915_gem_active). However, stream->last_request should be a reservation_object. -Chris

Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#11 lore raw

Re: [PATCH 3/8] drm/i915: Framework for capturing command stream based OA reports

Quoted 66 lines · Chris Wilson
Quoted 54 lines · sourab gupta
Quoted 23 lines · Chris Wilson
Quoted 22 lines · sourab gupta
Quoted 20 lines · Chris Wilson
Quoted 13 lines · On Thu, Mar 16, 2017 at 11:44:10AM +0530, [email protected] wrote:
drivers/gpu/drm/i915/i915_drv.h +1 -0
@@ -3593,6 +3670,7 @@ void i915_oa_init_reg_state(struct intel_engine_cs *engine,
 void i915_oa_update_reg_state(struct intel_engine_cs *engine,
 			      struct i915_gem_context *ctx,
 			      uint32_t *reg_state);
+void i915_perf_command_stream_hook(struct drm_i915_gem_request *req);
 
 /* i915_gem_evict.c */
 int __must_check i915_gem_evict_something(struct i915_address_space *vm,
drivers/gpu/drm/i915/i915_gem_execbuffer.c +0 -0
index aa75ea2..7af32c97 100644

drivers/gpu/drm/i915/i915_gem_execbuffer.c static void eb_export_fence(struct drm_i915_gem_object *obj, · line 1444

Quoted 4 lines
drivers/gpu/drm/i915/i915_gem_execbuffer.c +1 -0
@@ -1441,12 +1441,16 @@ static void eb_export_fence(struct drm_i915_gem_object *obj,
 	if (exec_len == 0)
 		exec_len = params->batch->size - params->args_batch_start_offset;
 
+	i915_perf_command_stream_hook(params->request);

Could you have named it anything more cyptic and inconsistent?

Quoted 7 lines
Quoted 2 lines

Sorry. Would 'i915_perf_capture_metrics' work? Can you please suggest an appropriate name otherwise.

The verb we use for writting into the command stream is emit. So i915_perf_emit_samples() (emit_record record is clumsy as it is not clear whether it is the verb or noun).

Thanks for suggesting. I'll use 'i915_perf_emit_samples' here.

Quoted 14 lines
Quoted 9 lines
Quoted 5 lines

Quite clearly this can fail, so justify the non handling of errors.

DRM_ERROR is not error handling, it is an indication that this patch isn't ready.

Well, the intent was to have minimal effect to execbuf normal routine, even if we fail. But, I guess I'm mistaken.
I'll rectify this to handle the case, if perf callback fails.

That all depends on whether or not you can handle the unbalanced metrics. If simply missing a report is fine, then just kill the DRM_ERROR.

Quoted 6 lines

The bigger question is whether the following emit can to fail -- once the batch is in the request, no more failures are tolerable. You have to preallocate reserved space.

Don't you need a flush before the emit following the batch?

Quoted 6 lines

Ok. So, that would mean that we have to first of all reserve the space required by two 'perf_emit_samples' calls, so that we can't fail for the lack of space in the emit following the batch.
Probably, I could pass an additional boolean parameter 'reserve_space' set to true in the first call, which would reserve the space for both emit_samples() calls (through intel_ring_begin)?

Hmm, yes, you can just tweak the request->reserved_space in the first call to preallocate some space (and make it remains available) for the second.

Quoted 10 lines
perf_emit_samples(rq, bool preallocate) {
	/* then in each sample callback */
	cs_len = foo;
	if (preallocate)
		rq->reserved_space += cs_len;
	else
		rq->reserved_space -= cs_len;
	cs = intel_ring_begin(rq, cs_len);
}

Ok. I would do that. Thanks for suggesting.

Quoted 10 lines
Quoted 9 lines

Would it still need the flush before the emit following the batch? Ideally, the metrics should be collected as close to batch as possible. So, if there are cache flush/tlb invalidate commands, it would introduce some lag between the batch and following emit_samples command. Sorry, I'm not able to gauge the need for flush here. I can understand it's needed before the batch is programmed for flushing the cache/TLB entries for the new workload to be submitted. But for the Sample_emit commands, which generally only capture OA/timestamp/mmio metrics, is this still required?

Quoted 5 lines

Depends on the desired accuracy. If you want your metrics sampled after the user pipeline is completed, you need a stall at least, a flush if your metrics include e.g. fragment data. If you want samples taken in whilst the user's batch is still executing, then no.

I guess, in that case, I would atleast need a stall to ensure user pipeline is completed before metrics are collected.
Any additional inputs here, Robert?

Quoted 22 lines
Quoted 22 lines
Quoted 22 lines
Quoted 21 lines
Quoted 4 lines
Quoted 4 lines
drivers/gpu/drm/i915/i915_gem_execbuffer.c +4 -0
+void i915_perf_command_stream_hook(struct drm_i915_gem_request *request)
+{
+	struct intel_engine_cs *engine = request->engine;
+	struct drm_i915_private *dev_priv = engine->i915;
Quoted 8 lines

drivers/gpu/drm/i915/i915_gem_execbuffer.c static void eb_export_fence(struct drm_i915_gem_object *obj, · line 1454

Quoted 6 lines
drivers/gpu/drm/i915/i915_gem_execbuffer.c +6 -0
+	struct i915_perf_stream *stream;
+
+	if (!dev_priv->perf.initialized)
+		return;
+
+	mutex_lock(&dev_priv->perf.streams_lock);

Justify a new global lock very, very carefully on execbuf.

The lock introduced here is to protect the the perf.streams list against addition/deletion while we're processing the list during execbuf call. The other places where the mutex is taken is when a new stream is being created (using perf_open ioctl) or a stream is being destroyed (perf_release ioctl), which just protect the list_add and list_del into the perf.streams list.
As such, there should not be much impact on execbuf path. Does this seem reasonable?

Quoted 8 lines
Quoted 4 lines
Quoted 2 lines

It doesn't sound like it needs a mutex, which will simplify the other users as well (if switched to, for example, an RCU protected list).

Ok. Sounds reasonable, I'll switch to using an RCU protected list here.

(I may be overthinking this, but it still seems overkill and made the timer callback uglier than expected.)

Are you suggesting that using RCU here is overkill, and mutex would do?

Quoted 59 lines
Quoted 43 lines
Quoted 36 lines
Quoted 36 lines
Quoted 36 lines
Quoted 24 lines
drivers/gpu/drm/i915/i915_gem_execbuffer.c +23 -0
+	list_for_each_entry(stream, &dev_priv->perf.streams, link) {
+		if ((stream->state == I915_PERF_STREAM_ENABLED) &&
+					stream->cs_mode)
+			stream->ops->command_stream_hook(stream, request);
+	}
+	mutex_unlock(&dev_priv->perf.streams_lock);
+}
 
+static void i915_perf_command_stream_hook_oa(struct i915_perf_stream *stream,
+					struct drm_i915_gem_request *request)
+{
+	struct drm_i915_private *dev_priv = request->i915;
+	struct i915_gem_context *ctx = request->ctx;
+	struct i915_perf_cs_sample *sample;
+	u32 addr = 0;
+	u32 cmd, *cs;
+
+	sample = kzalloc(sizeof(*sample), GFP_KERNEL);
+	if (sample == NULL) {
+		DRM_ERROR("Perf sample alloc failed\n");
+		return;
+	}
+
+	cs = intel_ring_begin(request, 4);

drivers/gpu/drm/i915/i915_gem_execbuffer.c static void eb_export_fence(struct drm_i915_gem_object *obj,

Quoted 10 lines
drivers/gpu/drm/i915/i915_gem_execbuffer.c +9 -0
+	if (IS_ERR(cs)) {
+		kfree(sample);
+		return;
+	}
+
+	sample->ctx_id = ctx->hw_id;
+	i915_gem_request_assign(&sample->request, request);
 
+
+	i915_gem_active_set(&stream->last_request, request);

Hint, you don't need you own request trap.

Quoted 6 lines
Quoted 4 lines

Sorry, I didn't understand you fully here. I'm storing the reference to the last active request associated with the stream inside the last_request 'gem_active' field. Do I need to do something differently here?

It's the duplication.

Quoted 7 lines

Are you suggesting that request_assign() and active_set() is duplication?
Actually, I'm using the last_request active tracker for the purpose of waiting on last request to complete, whenever the need.
But still I need to get reference for every request for which the commands for collection of metrics are emitted. This is because I need to check for their completion before collecting the associated metrics.

Missed the sample / stream difference. request_assign means update the request field, had you used

	sample->request = i915_gem_request_get(request)

it would have been clearer.

Note that the requests are not ordered for the stream, so you cannot track the last_request so easily.

Quoted 9 lines
Quoted 3 lines

This is done inside 'append_command_stream_samples' function, which also takes care of releasing the reference taken here.
Am I missing something w.r.t. the active_set() functionality?

I was mostly thrown by the idea that you were reassigning requests, which history has shown to be a bad idea (hence i915_gem_active). However, stream->last_request should be a reservation_object. -Chris

If I understand correctly, I need to have last_request of type reservation object to hold all fences corresponding to the requests pertaining to the stream. So, instead of i915_gem_active_set, I need to do something like this:

	reservation_object_add_excl_fence(obj->resv, &rq->fence); /* Or any

other api to be used here ? */
And, when we want to wait for requests submitted on the stream to complete, we have to wait for fences associated with the 'last_request' reservation_object, for e.g. as done in 'i915_gem_object_wait_reservation' function.
Is this all to it, or am I missing some piece of action with this reservation_object?

Thanks,
Sourab

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#12 lore raw

Re: [PATCH 3/8] drm/i915: Framework for capturing command stream based OA reports

Quoted 31 lines · sourab gupta
Quoted 30 lines · Chris Wilson
Quoted 29 lines · sourab gupta
Quoted 27 lines · Chris Wilson
Quoted 15 lines · sourab gupta
Quoted 13 lines · Chris Wilson

Review

Quoted 10 lines · On Thu, Mar 16, 2017 at 11:44:10AM +0530, [email protected] wrote:
(quoted diff) +10 -0
+void i915_perf_command_stream_hook(struct drm_i915_gem_request *request)
+{
+	struct intel_engine_cs *engine = request->engine;
+	struct drm_i915_private *dev_priv = engine->i915;
+	struct i915_perf_stream *stream;
+
+	if (!dev_priv->perf.initialized)
+		return;
+
+	mutex_lock(&dev_priv->perf.streams_lock);

Justify a new global lock very, very carefully on execbuf.

Quoted 8 lines

The lock introduced here is to protect the the perf.streams list against addition/deletion while we're processing the list during execbuf call. The other places where the mutex is taken is when a new stream is being created (using perf_open ioctl) or a stream is being destroyed (perf_release ioctl), which just protect the list_add and list_del into the perf.streams list.
As such, there should not be much impact on execbuf path. Does this seem reasonable?

It doesn't sound like it needs a mutex, which will simplify the other users as well (if switched to, for example, an RCU protected list).

Quoted 6 lines
Quoted 5 lines
Quoted 1 line

Ok. Sounds reasonable, I'll switch to using an RCU protected list here.

(I may be overthinking this, but it still seems overkill and made the timer callback uglier than expected.)

Are you suggesting that using RCU here is overkill, and mutex would do?

No, I still think it can be done without a mutex, just slightly more ugly due to the potential sleep here (perhaps srcu?). It would just be a shame when we get parallel submission of execbuf sorted for it all to be serialised by walking a simple list.

Quoted 10 lines
Quoted 4 lines

I was mostly thrown by the idea that you were reassigning requests, which history has shown to be a bad idea (hence i915_gem_active). However, stream->last_request should be a reservation_object.

If I understand correctly, I need to have last_request of type reservation object to hold all fences corresponding to the requests pertaining to the stream. So, instead of i915_gem_active_set, I need to do something like this:

	reservation_object_add_excl_fence(obj->resv, &rq->fence); /* Or any

other api to be used here ? */

reservation_object_lock();
if (reservation_object_reserve_shared() == 0)

	reservation_object_add_shared_fence();

reservation_object_unlock();

Quoted 6 lines

And, when we want to wait for requests submitted on the stream to complete, we have to wait for fences associated with the 'last_request' reservation_object, for e.g. as done in 'i915_gem_object_wait_reservation' function.
Is this all to it, or am I missing some piece of action with this reservation_object?

reservation_object_test_signaled_rcu(.test_all = true) reservation_object_wait_timeout_rcu(.wait_all = true) -Chris

Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#13 lore raw

Re: [PATCH 3/8] drm/i915: Framework for capturing command stream based OA reports

In reply to [email protected]
Quoted 31 lines · On Thu, Mar 16, 2017 at 11:44:10AM +0530, [email protected] wrote:
drivers/gpu/drm/i915/i915_perf.c +31 -0
+static void i915_perf_command_stream_hook_oa(struct i915_perf_stream *stream,
+					struct drm_i915_gem_request *request)
+{
+	struct drm_i915_private *dev_priv = request->i915;
+	struct i915_gem_context *ctx = request->ctx;
+	struct i915_perf_cs_sample *sample;
+	u32 addr = 0;
+	u32 cmd, *cs;
+
+	sample = kzalloc(sizeof(*sample), GFP_KERNEL);
+	if (sample == NULL) {
+		DRM_ERROR("Perf sample alloc failed\n");
+		return;
+	}
+
+	cs = intel_ring_begin(request, 4);
+	if (IS_ERR(cs)) {
+		kfree(sample);
+		return;
+	}
+
+	sample->ctx_id = ctx->hw_id;
+	i915_gem_request_assign(&sample->request, request);
+
+	insert_perf_sample(dev_priv, sample);
+
+	addr = dev_priv->perf.command_stream_buf.vma->node.start +
+		sample->offset;
+
+	if (WARN_ON(addr & 0x3f)) {
+		DRM_ERROR("OA buffer address not aligned to 64 byte");

drivers/gpu/drm/i915/i915_perf.c struct perf_open_properties { · line 549

Quoted 10 lines
drivers/gpu/drm/i915/i915_perf.c +10 -0
+		return;
+	}
+
+	cmd = MI_REPORT_PERF_COUNT | (1<<0);
+	if (INTEL_GEN(dev_priv) >= 8)
+		cmd |= (2<<0);
+
+	*cs++ = cmd;
+	*cs++ = addr | MI_REPORT_PERF_COUNT_GGTT;
+	*cs++ = request->global_seqno;

request->global_seqno is always zero at this point, and is subject to change after assignment.

[ctx:engine or request->fence.context]:request->fence.seqno is permanent. -Chris

Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#14 lore raw

Re: [PATCH 3/8] drm/i915: Framework for capturing command stream based OA reports

Quoted 46 lines · Chris Wilson
Quoted 31 lines · On Thu, Mar 16, 2017 at 11:44:10AM +0530, [email protected] wrote:
drivers/gpu/drm/i915/i915_perf.c +31 -0
+static void i915_perf_command_stream_hook_oa(struct i915_perf_stream *stream,
+					struct drm_i915_gem_request *request)
+{
+	struct drm_i915_private *dev_priv = request->i915;
+	struct i915_gem_context *ctx = request->ctx;
+	struct i915_perf_cs_sample *sample;
+	u32 addr = 0;
+	u32 cmd, *cs;
+
+	sample = kzalloc(sizeof(*sample), GFP_KERNEL);
+	if (sample == NULL) {
+		DRM_ERROR("Perf sample alloc failed\n");
+		return;
+	}
+
+	cs = intel_ring_begin(request, 4);
+	if (IS_ERR(cs)) {
+		kfree(sample);
+		return;
+	}
+
+	sample->ctx_id = ctx->hw_id;
+	i915_gem_request_assign(&sample->request, request);
+
+	insert_perf_sample(dev_priv, sample);
+
+	addr = dev_priv->perf.command_stream_buf.vma->node.start +
+		sample->offset;
+
+	if (WARN_ON(addr & 0x3f)) {
+		DRM_ERROR("OA buffer address not aligned to 64 byte");

drivers/gpu/drm/i915/i915_perf.c struct perf_open_properties { · line 549

Quoted 10 lines
drivers/gpu/drm/i915/i915_perf.c +10 -0
+		return;
+	}
+
+	cmd = MI_REPORT_PERF_COUNT | (1<<0);
+	if (INTEL_GEN(dev_priv) >= 8)
+		cmd |= (2<<0);
+
+	*cs++ = cmd;
+	*cs++ = addr | MI_REPORT_PERF_COUNT_GGTT;
+	*cs++ = request->global_seqno;

request->global_seqno is always zero at this point, and is subject to change after assignment.

Quoted 3 lines

[ctx:engine or request->fence.context]:request->fence.seqno is permanent. -Chris

Thanks for pointing out. I'll use request->fence.seqno instead.

Regards,
Sourab

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#15 lore raw

PATCH 4/8 drm/i915: flush periodic samples, in case of no pending CS sample requests

In reply to [email protected]

From: Sourab Gupta <[email protected]>

When there are no pending CS OA samples, flush the periodic OA samples collected so far.

We can safely forward the periodic OA samples in the case we have no pending CS samples, but we can't do so in the case we have pending CS samples, since we don't know what the ordering between pending CS samples and periodic samples will eventually be. If we have no pending CS sample, it won't be possible for future pending CS sample to have timestamps earlier than current periodic timestamp.

---
 drivers/gpu/drm/i915/i915_drv.h  |  10 ++-
 drivers/gpu/drm/i915/i915_perf.c | 175 +++++++++++++++++++++++++++++----------
 2 files changed, 139 insertions(+), 46 deletions(-)
drivers/gpu/drm/i915/i915_drv.h +7 -3
index 0b70052..d0f43e9 100644
@@ -2110,10 +2110,11 @@ struct i915_oa_ops {
 		    char __user *buf,
 		    size_t count,
 		    size_t *offset,
-		    u32 ts);
+		    u32 ts,
+		    u32 max_reports);
 
 	/**
-	 * @oa_buffer_check: Check for OA buffer data + update tail
+	 * @oa_buffer_num_reports: Return number of OA reports + update tail
 	 *
 	 * This is either called via fops or the poll check hrtimer (atomic
 	 * ctx) without any locks taken.
@@ -2126,7 +2127,8 @@ struct i915_oa_ops {
 	 * here, which will be handled gracefully - likely resulting in an
 	 * %EAGAIN error for userspace.
 	 */
-	bool (*oa_buffer_check)(struct drm_i915_private *dev_priv);
+	u32 (*oa_buffer_num_reports)(struct drm_i915_private *dev_priv,
+					u32 *last_ts);
 };
 
 /*
@@ -2589,6 +2591,8 @@ struct drm_i915_private {
 			u32 gen7_latched_oastatus1;
 			u32 ctx_oactxctrl_off;
 			u32 ctx_flexeu0_off;
+			u32 n_pending_periodic_samples;
+			u32 pending_periodic_ts;
 
 			/* The RPT_ID/reason field for Gen8+ includes a bit
 			 * to determine if the CTX ID in the report is valid
drivers/gpu/drm/i915/i915_perf.c +132 -43

422 lines not on this page · raw message

Signature · 6 lines
1.9.1

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#16 lore raw

PATCH 5/8 drm/i915: Inform userspace about command stream OA buf overflow

In reply to [email protected]

From: Sourab Gupta <[email protected]>

Considering how we don't currently give userspace control over the OA buffer size and always configure a large 16MB buffer, then a buffer overflow does anyway likely indicate that something has gone quite badly wrong.

Here we set a status flag to detect overflow and inform userspace of the report_lost condition accordingly. This is in line with the behavior of the periodic OA buffer.

---
 drivers/gpu/drm/i915/i915_drv.h  |  2 ++
 drivers/gpu/drm/i915/i915_perf.c | 15 +++++++++++++++
 2 files changed, 17 insertions(+)
drivers/gpu/drm/i915/i915_drv.h +2 -0
index d0f43e9..0f2a552 100644
@@ -2609,6 +2609,8 @@ struct drm_i915_private {
 		struct {
 			struct i915_vma *vma;
 			u8 *vaddr;
+#define I915_PERF_CMD_STREAM_BUF_STATUS_OVERFLOW (1<<0)
+			u32 status;
 		} command_stream_buf;
 
 		struct list_head cs_samples;
drivers/gpu/drm/i915/i915_perf.c +15 -0
index d1d9853..2841d0a 100644
@@ -478,6 +478,8 @@ static void insert_perf_sample(struct drm_i915_private *dev_priv,
 		else {
 			u32 target_size = sample_size - first->offset;
 
+			dev_priv->perf.command_stream_buf.status |=
+				I915_PERF_CMD_STREAM_BUF_STATUS_OVERFLOW;
 			release_perf_samples(dev_priv, target_size);
 			sample->offset = 0;
 		}
@@ -491,6 +493,8 @@ static void insert_perf_sample(struct drm_i915_private *dev_priv,
 				(first->offset - last->offset -
 				sample_size);
 
+			dev_priv->perf.command_stream_buf.status |=
+				I915_PERF_CMD_STREAM_BUF_STATUS_OVERFLOW;
 			release_perf_samples(dev_priv, target_size);
 			sample->offset = last->offset + sample_size;
 		}
@@ -1577,6 +1581,17 @@ static int oa_rcs_append_reports(struct i915_perf_stream *stream,
 	struct i915_perf_cs_sample *entry, *next;
 	LIST_HEAD(free_list);
 	int ret = 0;
+	u32 status = dev_priv->perf.command_stream_buf.status;
+
+	if (unlikely(status & I915_PERF_CMD_STREAM_BUF_STATUS_OVERFLOW)) {
+		ret = append_oa_status(stream, buf, count, offset,
+				       DRM_I915_PERF_RECORD_OA_BUFFER_LOST);
+		if (ret)
+			return ret;
+
+		dev_priv->perf.command_stream_buf.status &=
+				~I915_PERF_CMD_STREAM_BUF_STATUS_OVERFLOW;
+	}
 
 	spin_lock(&dev_priv->perf.sample_lock);
 	if (list_empty(&dev_priv->perf.cs_samples)) {
Signature · 6 lines
1.9.1

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#17 lore raw

PATCH 6/8 drm/i915: Populate ctx ID for periodic OA reports

In reply to [email protected]

From: Sourab Gupta <[email protected]>

This adds support for populating the ctx id for the periodic OA reports when requested through the corresponding property.

For Gen8, the OA reports itself have the ctx ID and it is the one programmed into HW while submitting workloads. Thus it's retrieved from reports itself. For Gen7, the OA reports don't have any such field, and we can populate this field with the last seen ctx ID while sending CS reports.

---
 drivers/gpu/drm/i915/i915_drv.h  |  7 ++++++
 drivers/gpu/drm/i915/i915_perf.c | 53 +++++++++++++++++++++++++++++++++++-----
 2 files changed, 54 insertions(+), 6 deletions(-)
drivers/gpu/drm/i915/i915_drv.h +7 -0
index 0f2a552..7a6dcb3 100644
@@ -2129,6 +2129,12 @@ struct i915_oa_ops {
 	 */
 	u32 (*oa_buffer_num_reports)(struct drm_i915_private *dev_priv,
 					u32 *last_ts);
+
+	/**
+	 * @get_ctx_id: Retrieve the ctx_id associated with the (periodic) OA
+	 * report.
+	 */
+	u32 (*get_ctx_id)(struct i915_perf_stream *stream, const u8 *report);
 };
 
 /*
@@ -2613,6 +2619,7 @@ struct drm_i915_private {
 			u32 status;
 		} command_stream_buf;
 
+		u32 last_cmd_stream_ctx_id;
 		struct list_head cs_samples;
 		spinlock_t sample_lock;
 	} perf;
drivers/gpu/drm/i915/i915_perf.c +47 -6

91 lines not on this page · raw message

Signature · 6 lines
1.9.1

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#18 lore raw

PATCH 7/8 drm/i915: Add support for having pid output with OA report

In reply to [email protected]

From: Sourab Gupta <[email protected]>

This patch introduces flags and adds support for having pid output with the OA reports generated through the RCS commands.

When the stream is opened with pid sample type, the pid information is also captured through the command stream samples and forwarded along with the OA reports.

---
 drivers/gpu/drm/i915/i915_drv.h  |  7 ++++++
 drivers/gpu/drm/i915/i915_perf.c | 50 ++++++++++++++++++++++++++++++++++++++--
 include/uapi/drm/i915_drm.h      |  7 ++++++
 3 files changed, 62 insertions(+), 2 deletions(-)
drivers/gpu/drm/i915/i915_drv.h +7 -0
index 7a6dcb3..fa1d3fc 100644
@@ -2165,6 +2165,12 @@ struct i915_perf_cs_sample {
 	 * @ctx_id: Context ID associated with this perf sample
 	 */
 	u32 ctx_id;
+
+	/**
+	 * @pid: PID of the process in context of which the workload was
+	 * submitted, pertaining to this perf sample
+	 */
+	u32 pid;
 };
 
 struct intel_cdclk_state {
@@ -2620,6 +2626,7 @@ struct drm_i915_private {
 		} command_stream_buf;
 
 		u32 last_cmd_stream_ctx_id;
+		u32 last_pid;
 		struct list_head cs_samples;
 		spinlock_t sample_lock;
 	} perf;
drivers/gpu/drm/i915/i915_perf.c +48 -2

116 lines not on this page · raw message

include/uapi/drm/i915_drm.h +7 -0
index 4cbefc8..25cf612 100644
@@ -1373,6 +1373,12 @@ enum drm_i915_perf_property_id {
 	 */
 	DRM_I915_PERF_PROP_SAMPLE_CTX_ID,
 
+	/**
+	 * The value of this property set to 1 requests inclusion of pid in the
+	 * perf sample data.
+	 */
+	DRM_I915_PERF_PROP_SAMPLE_PID,
+
 	DRM_I915_PERF_PROP_MAX /* non-ABI */
 };
 
@@ -1440,6 +1446,7 @@ enum drm_i915_perf_record_type {
 	 *
 	 *     { u32 source_info; } && DRM_I915_PERF_PROP_SAMPLE_OA_SOURCE
 	 *     { u32 ctx_id; } && DRM_I915_PERF_PROP_SAMPLE_CTX_ID
+	 *     { u32 pid; } && DRM_I915_PERF_PROP_SAMPLE_PID
 	 *     { u32 oa_report[]; } && DRM_I915_PERF_PROP_SAMPLE_OA
 	 * };
 	 */
Signature · 6 lines
1.9.1

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#19 lore raw

PATCH 8/8 drm/i915: Add support for emitting execbuffer tags through OA counter reports

In reply to [email protected]

From: Sourab Gupta <[email protected]>

This patch enables userspace to specify tags (per workload), provided via execbuffer ioctl, which could be added to OA reports, to help associate reports with the corresponding workloads.

There may be multiple stages within a single context, from a userspace perspective. An ability is needed to individually associate the OA reports with their corresponding workloads(execbuffers), which may not be possible solely with ctx_id or pid information. This patch enables such a mechanism.

In this patch, upper 32 bits of rsvd1 field, which were previously unused are now being used to pass in the tag.

---
 drivers/gpu/drm/i915/i915_drv.h            | 17 +++++++++++--
 drivers/gpu/drm/i915/i915_gem_execbuffer.c |  6 +++--
 drivers/gpu/drm/i915/i915_perf.c           | 40 +++++++++++++++++++++++++-----
 include/uapi/drm/i915_drm.h                | 12 +++++++++
 4 files changed, 65 insertions(+), 10 deletions(-)
drivers/gpu/drm/i915/i915_drv.h +15 -2
index fa1d3fc..414afa5 100644
@@ -1982,7 +1982,8 @@ struct i915_perf_stream_ops {
 	 * the batch buffer.
 	 */
 	void (*command_stream_hook)(struct i915_perf_stream *stream,
-				struct drm_i915_gem_request *request);
+				struct drm_i915_gem_request *request,
+				u32 tag);
 };
 
 enum i915_perf_stream_state {
@@ -2171,6 +2172,17 @@ struct i915_perf_cs_sample {
 	 * submitted, pertaining to this perf sample
 	 */
 	u32 pid;
+
+	/**
+	 * @tag: Tag associated with workload, for which the perf sample is
+	 * being collected.
+	 *
+	 * Userspace can specify tags (provided via execbuffer ioctl), which
+	 * can be associated with the perf samples, and be used to functionally
+	 * distinguish different workload stages, and associate samples with
+	 * these different stages.
+	 */
+	u32 tag;
 };
 
 struct intel_cdclk_state {
@@ -2627,6 +2639,7 @@ struct drm_i915_private {
 
 		u32 last_cmd_stream_ctx_id;
 		u32 last_pid;
+		u32 last_tag;
 		struct list_head cs_samples;
 		spinlock_t sample_lock;
 	} perf;
@@ -3690,7 +3703,7 @@ void i915_oa_init_reg_state(struct intel_engine_cs *engine,
 void i915_oa_update_reg_state(struct intel_engine_cs *engine,
 			      struct i915_gem_context *ctx,
 			      uint32_t *reg_state);
-void i915_perf_command_stream_hook(struct drm_i915_gem_request *req);
+void i915_perf_command_stream_hook(struct drm_i915_gem_request *req, u32 tag);
 
 /* i915_gem_evict.c */
 int __must_check i915_gem_evict_something(struct i915_address_space *vm,
drivers/gpu/drm/i915/i915_gem_execbuffer.c +4 -2
index 7af32c97..b42d47e 100644
@@ -59,6 +59,7 @@ struct i915_execbuffer_params {
 	struct intel_engine_cs          *engine;
 	struct i915_gem_context         *ctx;
 	struct drm_i915_gem_request     *request;
+	uint32_t			tag;
 };
 
 struct eb_vmas {
@@ -1441,7 +1442,7 @@ static void eb_export_fence(struct drm_i915_gem_object *obj,
 	if (exec_len == 0)
 		exec_len = params->batch->size - params->args_batch_start_offset;
 
-	i915_perf_command_stream_hook(params->request);
+	i915_perf_command_stream_hook(params->request, params->tag);
 
 	ret = params->engine->emit_bb_start(params->request,
 					    exec_start, exec_len,
@@ -1449,7 +1450,7 @@ static void eb_export_fence(struct drm_i915_gem_object *obj,
 	if (ret)
 		return ret;
 
-	i915_perf_command_stream_hook(params->request);
+	i915_perf_command_stream_hook(params->request, params->tag);
 
 	i915_gem_execbuffer_move_to_active(vmas, params->request);
 
@@ -1791,6 +1792,7 @@ static void eb_export_fence(struct drm_i915_gem_object *obj,
 	params->engine                    = engine;
 	params->dispatch_flags          = dispatch_flags;
 	params->ctx                     = ctx;
+	params->tag			= i915_execbuffer2_get_tag(*args);
 
 	trace_i915_gem_request_queue(params->request, dispatch_flags);
 
drivers/gpu/drm/i915/i915_perf.c +34 -6
index 6e8af2d..759865e 100644
@@ -290,6 +290,7 @@ struct oa_sample_data {
 	u32 source;
 	u32 ctx_id;
 	u32 pid;
+	u32 tag;
 	const u8 *report;
 };
 
@@ -344,6 +345,7 @@ struct oa_sample_data {
 #define SAMPLE_OA_SOURCE_INFO	(1<<1)
 #define SAMPLE_CTX_ID		(1<<2)
 #define SAMPLE_PID		(1<<3)
+#define SAMPLE_TAG		(1<<4)
 
 /**
  * struct perf_open_properties - for validated properties given to open a stream
@@ -387,7 +389,8 @@ struct perf_open_properties {
  * The function provides a hook through which the commands to capture perf
  * metrics, are inserted into the command stream of a GPU engine.
  */
-void i915_perf_command_stream_hook(struct drm_i915_gem_request *request)
+void i915_perf_command_stream_hook(struct drm_i915_gem_request *request,
+					u32 tag)
 {
 	struct intel_engine_cs *engine = request->engine;
 	struct drm_i915_private *dev_priv = engine->i915;
@@ -400,7 +403,7 @@ void i915_perf_command_stream_hook(struct drm_i915_gem_request *request)
 	list_for_each_entry(stream, &dev_priv->perf.streams, link) {
 		if ((stream->state == I915_PERF_STREAM_ENABLED) &&
 					stream->cs_mode)
-			stream->ops->command_stream_hook(stream, request);
+			stream->ops->command_stream_hook(stream, request, tag);
 	}
 	mutex_unlock(&dev_priv->perf.streams_lock);
 }
@@ -510,10 +513,11 @@ static void insert_perf_sample(struct drm_i915_private *dev_priv,
  * metrics into the render command stream
  * @stream: An i915-perf stream opened for OA metrics
  * @request: request in whose context the metrics are being collected.
- *
+ * @tag: userspace provided tag to be associated with the perf sample
  */
 static void i915_perf_command_stream_hook_oa(struct i915_perf_stream *stream,
-					struct drm_i915_gem_request *request)
+					struct drm_i915_gem_request *request,
+					u32 tag)
 {
 	struct drm_i915_private *dev_priv = request->i915;
 	struct i915_gem_context *ctx = request->ctx;
@@ -535,6 +539,7 @@ static void i915_perf_command_stream_hook_oa(struct i915_perf_stream *stream,
 
 	sample->ctx_id = ctx->hw_id;
 	sample->pid = current->pid;
+	sample->tag = tag;
 	i915_gem_request_assign(&sample->request, request);
 
 	insert_perf_sample(dev_priv, sample);
@@ -986,6 +991,12 @@ static int append_oa_sample(struct i915_perf_stream *stream,
 		buf += 4;
 	}
 
+	if (sample_flags & SAMPLE_TAG) {
+		if (copy_to_user(buf, &data->tag, 4))
+			return -EFAULT;
+		buf += 4;
+	}
+
 	if (sample_flags & SAMPLE_OA_REPORT) {
 		if (copy_to_user(buf, data->report, report_size))
 			return -EFAULT;
@@ -1025,6 +1036,9 @@ static int append_oa_buffer_sample(struct i915_perf_stream *stream,
 	if (sample_flags & SAMPLE_PID)
 		data.pid = dev_priv->perf.last_pid;
 
+	if (sample_flags & SAMPLE_TAG)
+		data.tag = dev_priv->perf.last_tag;
+
 	if (sample_flags & SAMPLE_OA_REPORT)
 		data.report = report;
 
@@ -1606,6 +1620,11 @@ static int append_oa_rcs_sample(struct i915_perf_stream *stream,
 		dev_priv->perf.last_pid = node->pid;
 	}
 
+	if (sample_flags & SAMPLE_TAG) {
+		data.tag = node->tag;
+		dev_priv->perf.last_tag = node->tag;
+	}
+
 	if (sample_flags & SAMPLE_OA_REPORT)
 		data.report = report;
 
@@ -2682,7 +2701,8 @@ static int i915_oa_stream_init(struct i915_perf_stream *stream,
 	struct drm_i915_private *dev_priv = stream->dev_priv;
 	bool require_oa_unit = props->sample_flags & (SAMPLE_OA_REPORT |
 						      SAMPLE_OA_SOURCE_INFO);
-	bool require_cs_mode = props->sample_flags & SAMPLE_PID;
+	bool require_cs_mode = props->sample_flags & (SAMPLE_PID |
+						      SAMPLE_TAG);
 	bool cs_sample_data = props->sample_flags & SAMPLE_OA_REPORT;
 	int ret;
 
@@ -2819,7 +2839,7 @@ static int i915_oa_stream_init(struct i915_perf_stream *stream,
 	}
 
 	if (require_cs_mode && !props->cs_mode) {
-		DRM_ERROR("PID sampling requires a ring to be specified");
+		DRM_ERROR("PID/TAG sampling requires a ring to be specified");
 		ret = -EINVAL;
 		goto err_enable;
 	}
@@ -2851,6 +2871,11 @@ static int i915_oa_stream_init(struct i915_perf_stream *stream,
 			stream->sample_size += 4;
 		}
 
+		if (props->sample_flags & SAMPLE_TAG) {
+			stream->sample_flags |= SAMPLE_TAG;
+			stream->sample_size += 4;
+		}
+
 		ret = alloc_command_stream_buf(dev_priv);
 		if (ret)
 			goto err_enable;
@@ -3648,6 +3673,9 @@ static int read_properties_unlocked(struct drm_i915_private *dev_priv,
 		case DRM_I915_PERF_PROP_SAMPLE_PID:
 			props->sample_flags |= SAMPLE_PID;
 			break;
+		case DRM_I915_PERF_PROP_SAMPLE_TAG:
+			props->sample_flags |= SAMPLE_TAG;
+			break;
 		default:
 			MISSING_CASE(id);
 			DRM_DEBUG("Unknown i915 perf property ID\n");
include/uapi/drm/i915_drm.h +12 -0
index 25cf612..ab5fdc0 100644
@@ -899,6 +899,11 @@ struct drm_i915_gem_execbuffer2 {
 #define i915_execbuffer2_get_context_id(eb2) \
 	((eb2).rsvd1 & I915_EXEC_CONTEXT_ID_MASK)
 
+/* upper 32 bits of rsvd1 field contain tag */
+#define I915_EXEC_TAG_MASK		(0xffffffff00000000UL)
+#define i915_execbuffer2_get_tag(eb2) \
+	((eb2).rsvd1 & I915_EXEC_TAG_MASK)
+
 struct drm_i915_gem_pin {
 	/** Handle of the buffer to be pinned. */
 	__u32 handle;
@@ -1379,6 +1384,12 @@ enum drm_i915_perf_property_id {
 	 */
 	DRM_I915_PERF_PROP_SAMPLE_PID,
 
+	/**
+	 * The value of this property set to 1 requests inclusion of tag in the
+	 * perf sample data.
+	 */
+	DRM_I915_PERF_PROP_SAMPLE_TAG,
+
 	DRM_I915_PERF_PROP_MAX /* non-ABI */
 };
 
@@ -1447,6 +1458,7 @@ enum drm_i915_perf_record_type {
 	 *     { u32 source_info; } && DRM_I915_PERF_PROP_SAMPLE_OA_SOURCE
 	 *     { u32 ctx_id; } && DRM_I915_PERF_PROP_SAMPLE_CTX_ID
 	 *     { u32 pid; } && DRM_I915_PERF_PROP_SAMPLE_PID
+	 *     { u32 tag; } && DRM_I915_PERF_PROP_SAMPLE_TAG
 	 *     { u32 oa_report[]; } && DRM_I915_PERF_PROP_SAMPLE_OA
 	 * };
 	 */
Signature · 6 lines
1.9.1

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#20 lore raw

Re: [PATCH 8/8] drm/i915: Add support for emitting execbuffer tags through OA counter reports

Quoted 6 lines · On Thu, Mar 16, 2017 at 11:44:15AM +0530, [email protected] wrote:

From: Sourab Gupta <[email protected]>

This patch enables userspace to specify tags (per workload), provided via execbuffer ioctl, which could be added to OA reports, to help associate reports with the corresponding workloads.

Quoted 7 lines

There may be multiple stages within a single context, from a userspace perspective. An ability is needed to individually associate the OA reports with their corresponding workloads(execbuffers), which may not be possible solely with ctx_id or pid information. This patch enables such a mechanism.

In this patch, upper 32 bits of rsvd1 field, which were previously unused are now being used to pass in the tag.

Clearly this patch hasn't been tested.
-Chris

Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#21 lore raw

✓ Fi.CI.BAT: success for Collect command stream based OA reports using i915 perf

In reply to [email protected]
== Series Details ==

== Series Details ==

Series: Collect command stream based OA reports using i915 perf
URL : https://patchwork.freedesktop.org/series/21351/
State : success

== Summary ==

Series 21351v1 Collect command stream based OA reports using i915 perf
https://patchwork.freedesktop.org/api/1.0/series/21351/revisions/1/mbox/

Test gem_exec_flush:

        Subgroup basic-batch-kernel-default-uc:
                fail       -> PASS       (fi-snb-2600) fdo#100007

fdo#100007 https://bugs.freedesktop.org/show_bug.cgi?id=100007

fi-bdw-5557u     total:278  pass:267  dwarn:0   dfail:0   fail:0   skip:11  time: 455s
fi-bsw-n3050     total:278  pass:239  dwarn:0   dfail:0   fail:0   skip:39  time: 579s
fi-bxt-j4205     total:278  pass:259  dwarn:0   dfail:0   fail:0   skip:19  time: 540s
fi-bxt-t5700     total:278  pass:258  dwarn:0   dfail:0   fail:0   skip:20  time: 534s
fi-byt-j1900     total:278  pass:251  dwarn:0   dfail:0   fail:0   skip:27  time: 499s
fi-byt-n2820     total:278  pass:247  dwarn:0   dfail:0   fail:0   skip:31  time: 496s
fi-hsw-4770      total:278  pass:262  dwarn:0   dfail:0   fail:0   skip:16  time: 439s
fi-hsw-4770r     total:278  pass:262  dwarn:0   dfail:0   fail:0   skip:16  time: 434s
fi-ilk-650       total:278  pass:228  dwarn:0   dfail:0   fail:0   skip:50  time: 436s
fi-ivb-3520m     total:278  pass:260  dwarn:0   dfail:0   fail:0   skip:18  time: 515s
fi-ivb-3770      total:278  pass:260  dwarn:0   dfail:0   fail:0   skip:18  time: 495s
fi-kbl-7500u     total:278  pass:260  dwarn:0   dfail:0   fail:0   skip:18  time: 482s
fi-skl-6260u     total:278  pass:268  dwarn:0   dfail:0   fail:0   skip:10  time: 482s
fi-skl-6700hq    total:278  pass:261  dwarn:0   dfail:0   fail:0   skip:17  time: 604s
fi-skl-6700k     total:278  pass:256  dwarn:4   dfail:0   fail:0   skip:18  time: 491s
fi-skl-6770hq    total:278  pass:268  dwarn:0   dfail:0   fail:0   skip:10  time: 515s
fi-snb-2520m     total:278  pass:250  dwarn:0   dfail:0   fail:0   skip:28  time: 546s
fi-snb-2600      total:278  pass:249  dwarn:0   dfail:0   fail:0   skip:29  time: 411s

d7cb114e1327b50c609ee6e67122cc0293ea515f drm-tip: 2017y-03m-15d-21h-47m-56s UTC integration manifest 1408d1e drm/i915: Expose OA sample source to userspace
dabfab4 drm/i915: Add ctx getparam ioctl parameter to retrieve ctx unique id

== Logs ==

For more details see: https://intel-gfx-ci.01.org/CI/Patchwork_4194/
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#22 lore raw
In reply to [email protected]
Quoted 22 lines · On Thu, Mar 16, 2017 at 6:14 AM, <[email protected]> wrote:

From: Sourab Gupta <[email protected]>

This series adds framework for collection of OA reports associated with the render command stream, which are collected around batchbuffer boundaries.

Refloating the series rebased on Robert's latest patch set for 'Enabling OA unit for Gen 8 and 9 in i915 perf', which can be found here:
https://patchwork.freedesktop.org/series/20084/

Since Robert's patches are being reviewed and this patch series extends his framework to collect command stream based OA metrics, it would be good to keep this work in perspective. Looking to receive feedback (and possibly r-b's :)) on the series.

Since the OA reports collected associated with the render command stream, this also gives us the ability to collect other metadata such as ctx_id, pid, etc. with the samples, and thus we can establish the association of samples collected with the corresponding process/workload.

These patches can be found for viewing at
https://github.com/sourabgu/linux/tree/oa-6march2017

Quoted 11 lines

Sourab Gupta (8):
drm/i915: Add ctx getparam ioctl parameter to retrieve ctx unique id drm/i915: Expose OA sample source to userspace
drm/i915: Framework for capturing command stream based OA reports drm/i915: flush periodic samples, in case of no pending CS sample
requests
drm/i915: Inform userspace about command stream OA buf overflow drm/i915: Populate ctx ID for periodic OA reports drm/i915: Add support for having pid output with OA report drm/i915: Add support for emitting execbuffer tags through OA counter
reports

Thanks for the updated series Sourab.

I think it could really help to have a pointer to some userspace that can be used to exercise these new features. Maybe you could look at adding support to the gputop-csv command line tool which is probably the simplest, usable userspace for i915 perf we have currently. A pointer to some work-in-progress IGT tests could be good too.

Br,
- Robert

Quoted 11 lines
 drivers/gpu/drm/i915/i915_drv.h            |  125 ++-
 drivers/gpu/drm/i915/i915_gem_context.c    |    3 +
 drivers/gpu/drm/i915/i915_gem_execbuffer.c |    6 +
 drivers/gpu/drm/i915/i915_perf.c           | 1149 ++++++++++++++++++++++++----
 include/uapi/drm/i915_drm.h                |   49 ++
 5 files changed, 1184 insertions(+), 148 deletions(-)

--
1.9.1

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#23 lore raw
Quoted 37 lines · Robert Bragg
Quoted 22 lines · On Thu, Mar 16, 2017 at 6:14 AM, <[email protected]> wrote:

From: Sourab Gupta <[email protected]>

This series adds framework for collection of OA reports associated with the render command stream, which are collected around batchbuffer boundaries.

Refloating the series rebased on Robert's latest patch set for 'Enabling OA unit for Gen 8 and 9 in i915 perf', which can be found here:
https://patchwork.freedesktop.org/series/20084/

Since Robert's patches are being reviewed and this patch series extends his framework to collect command stream based OA metrics, it would be good to keep this work in perspective. Looking to receive feedback (and possibly r-b's :)) on the series.

Since the OA reports collected associated with the render command stream, this also gives us the ability to collect other metadata such as ctx_id, pid, etc. with the samples, and thus we can establish the association of samples collected with the corresponding process/workload.

These patches can be found for viewing at
https://github.com/sourabgu/linux/tree/oa-6march2017

Quoted 11 lines

Sourab Gupta (8):
drm/i915: Add ctx getparam ioctl parameter to retrieve ctx unique id drm/i915: Expose OA sample source to userspace
drm/i915: Framework for capturing command stream based OA reports drm/i915: flush periodic samples, in case of no pending CS sample
requests
drm/i915: Inform userspace about command stream OA buf overflow drm/i915: Populate ctx ID for periodic OA reports drm/i915: Add support for having pid output with OA report drm/i915: Add support for emitting execbuffer tags through OA counter
reports

Thanks for the updated series Sourab.

Quoted 9 lines

I think it could really help to have a pointer to some userspace that can be used to exercise these new features. Maybe you could look at adding support to the gputop-csv command line tool which is probably the simplest, usable userspace for i915 perf we have currently. A pointer to some work-in-progress IGT tests could be good too.

Br,
- Robert

Hi Robert,

I'll aim to publish this during next revision.

Thanks,
Sourab

Quoted 11 lines
Quoted 11 lines
 drivers/gpu/drm/i915/i915_drv.h            |  125 ++-
 drivers/gpu/drm/i915/i915_gem_context.c    |    3 +
 drivers/gpu/drm/i915/i915_gem_execbuffer.c |    6 +
 drivers/gpu/drm/i915/i915_perf.c           | 1149 ++++++++++++++++++++++++----
 include/uapi/drm/i915_drm.h                |   49 ++
 5 files changed, 1184 insertions(+), 148 deletions(-)

--
1.9.1

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx