diff mbox series

drm/i915/perf: reintroduce wait on OA configuration completion

Message ID 20200227124356.1625616-1-lionel.g.landwerlin@intel.com (mailing list archive)
State New, archived
Headers show
Series drm/i915/perf: reintroduce wait on OA configuration completion | expand

Commit Message

Lionel Landwerlin Feb. 27, 2020, 12:43 p.m. UTC 
We still need to wait for the initial OA configuration to happen
before we enable OA report writes to the OA buffer.

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Fixes: 15d0ace1f876 ("drm/i915/perf: execute OA configuration from command stream")
---
 drivers/gpu/drm/i915/i915_perf.c       | 49 +++++++++++++++++++++++---
 drivers/gpu/drm/i915/i915_perf_types.h |  8 +++++
 2 files changed, 52 insertions(+), 5 deletions(-)

Comments

Chris Wilson Feb. 27, 2020, 5:04 p.m. UTC | #1 
Quoting Lionel Landwerlin (2020-02-27 12:43:56)
> We still need to wait for the initial OA configuration to happen
> before we enable OA report writes to the OA buffer.

I can confirm this fixes the hang Lionel reported on Skylake [still odd
that we can only get this to be an issue on skl]. However, Lionel
mentioned that we should be more careful and ensure the emit_oa_config()
request is scheduled last. We're currently looking at different ways we
can do that.
-Chris
Chris Wilson Feb. 28, 2020, 10:52 a.m. UTC | #2 
Quoting Chris Wilson (2020-02-27 17:04:42)
> Quoting Lionel Landwerlin (2020-02-27 12:43:56)
> > We still need to wait for the initial OA configuration to happen
> > before we enable OA report writes to the OA buffer.
> 
> I can confirm this fixes the hang Lionel reported on Skylake [still odd
> that we can only get this to be an issue on skl]. However, Lionel
> mentioned that we should be more careful and ensure the emit_oa_config()
> request is scheduled last. We're currently looking at different ways we
> can do that.

Fwiw, this patch works and fixes the problem. I prefer not adding
a single use temporary to i915_perf_stream, but I leave that choice to
Lionel.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
-Chris
Lionel Landwerlin Feb. 28, 2020, 11:27 a.m. UTC | #3 
On 28/02/2020 12:52, Chris Wilson wrote:
> Quoting Chris Wilson (2020-02-27 17:04:42)
>> Quoting Lionel Landwerlin (2020-02-27 12:43:56)
>>> We still need to wait for the initial OA configuration to happen
>>> before we enable OA report writes to the OA buffer.
>> I can confirm this fixes the hang Lionel reported on Skylake [still odd
>> that we can only get this to be an issue on skl]. However, Lionel
>> mentioned that we should be more careful and ensure the emit_oa_config()
>> request is scheduled last. We're currently looking at different ways we
>> can do that.
> Fwiw, this patch works and fixes the problem. I prefer not adding
> a single use temporary to i915_perf_stream, but I leave that choice to
> Lionel.
>
> Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
> -Chris

Thanks Chris!


-Lionel
diff mbox series

Patch

diff --git a/drivers/gpu/drm/i915/i915_perf.c b/drivers/gpu/drm/i915/i915_perf.c
index f4e1dd525fa2..3883c21b13b2 100644
--- a/drivers/gpu/drm/i915/i915_perf.c
+++ b/drivers/gpu/drm/i915/i915_perf.c
@@ -1378,6 +1378,23 @@  free_noa_wait(struct i915_perf_stream *stream)
 	i915_vma_unpin_and_release(&stream->noa_wait, 0);
 }
 
+static int
+wait_and_put_configure_request(struct i915_perf_stream *stream)
+{
+	struct i915_request *rq = stream->configure_request;
+	int ret = 0;
+
+	stream->configure_request = NULL;
+	GEM_BUG_ON(rq == NULL);
+
+	if (i915_request_wait(rq, 0, MAX_SCHEDULE_TIMEOUT) < 0)
+		ret = -ETIME;
+
+	i915_request_put(rq);
+
+	return ret;
+}
+
 static void i915_oa_stream_destroy(struct i915_perf_stream *stream)
 {
 	struct i915_perf *perf = stream->perf;
@@ -1390,6 +1407,7 @@  static void i915_oa_stream_destroy(struct i915_perf_stream *stream)
 	 */
 	perf->exclusive_stream = NULL;
 	perf->ops.disable_metric_set(stream);
+	GEM_BUG_ON(stream->configure_request != NULL);
 
 	free_oa_buffer(stream);
 
@@ -1954,7 +1972,8 @@  get_oa_vma(struct i915_perf_stream *stream, struct i915_oa_config *oa_config)
 
 static int emit_oa_config(struct i915_perf_stream *stream,
 			  struct i915_oa_config *oa_config,
-			  struct intel_context *ce)
+			  struct intel_context *ce,
+			  bool store_on_stream)
 {
 	struct i915_request *rq;
 	struct i915_vma *vma;
@@ -1987,6 +2006,12 @@  static int emit_oa_config(struct i915_perf_stream *stream,
 	err = rq->engine->emit_bb_start(rq,
 					vma->node.start, 0,
 					I915_DISPATCH_SECURE);
+
+	if (err == 0 && store_on_stream) {
+		GEM_BUG_ON(stream->configure_request != NULL);
+		stream->configure_request = i915_request_get(rq);
+	}
+
 err_add_request:
 	i915_request_add(rq);
 err_vma_unpin:
@@ -2020,7 +2045,9 @@  static int hsw_enable_metric_set(struct i915_perf_stream *stream)
 	intel_uncore_rmw(uncore, GEN6_UCGCTL1,
 			 0, GEN6_CSUNIT_CLOCK_GATE_DISABLE);
 
-	return emit_oa_config(stream, stream->oa_config, oa_context(stream));
+	return emit_oa_config(stream, stream->oa_config,
+			      oa_context(stream),
+			      true /* store_on_stream */);
 }
 
 static void hsw_disable_metric_set(struct i915_perf_stream *stream)
@@ -2448,7 +2475,9 @@  static int gen8_enable_metric_set(struct i915_perf_stream *stream)
 	if (ret)
 		return ret;
 
-	return emit_oa_config(stream, oa_config, oa_context(stream));
+	return emit_oa_config(stream, oa_config,
+			      oa_context(stream),
+			      true /* store_on_stream */);
 }
 
 static u32 oag_report_ctx_switches(const struct i915_perf_stream *stream)
@@ -2502,7 +2531,9 @@  static int gen12_enable_metric_set(struct i915_perf_stream *stream)
 			return ret;
 	}
 
-	return emit_oa_config(stream, oa_config, oa_context(stream));
+	return emit_oa_config(stream, oa_config,
+			      oa_context(stream),
+			      true /* store_on_stream */);
 }
 
 static void gen8_disable_metric_set(struct i915_perf_stream *stream)
@@ -2837,6 +2868,12 @@  static int i915_oa_stream_init(struct i915_perf_stream *stream,
 		goto err_enable;
 	}
 
+	ret = wait_and_put_configure_request(stream);
+	if (ret) {
+		DRM_DEBUG("Wait on OA config request timed out\n");
+		goto err_enable;
+	}
+
 	DRM_DEBUG("opening stream oa config uuid=%s\n",
 		  stream->oa_config->uuid);
 
@@ -2851,6 +2888,7 @@  static int i915_oa_stream_init(struct i915_perf_stream *stream,
 err_enable:
 	perf->exclusive_stream = NULL;
 	perf->ops.disable_metric_set(stream);
+	GEM_BUG_ON(stream->configure_request != NULL);
 
 	free_oa_buffer(stream);
 
@@ -3160,7 +3198,8 @@  static long i915_perf_config_locked(struct i915_perf_stream *stream,
 		 * When set globally, we use a low priority kernel context,
 		 * so it will effectively take effect when idle.
 		 */
-		err = emit_oa_config(stream, config, oa_context(stream));
+		err = emit_oa_config(stream, config, oa_context(stream),
+				     false /* store_on_stream */);
 		if (err == 0)
 			config = xchg(&stream->oa_config, config);
 		else
diff --git a/drivers/gpu/drm/i915/i915_perf_types.h b/drivers/gpu/drm/i915/i915_perf_types.h
index d994fa6a1c5f..92fd3cf5afcc 100644
--- a/drivers/gpu/drm/i915/i915_perf_types.h
+++ b/drivers/gpu/drm/i915/i915_perf_types.h
@@ -309,6 +309,14 @@  struct i915_perf_stream {
 	 * reprogrammed.
 	 */
 	struct i915_vma *noa_wait;
+
+	/**
+	 * @configure_request: Request on which to wait for HW to complete its
+	 * initial configuration. This is required for applications caring
+	 * about system wide monitoring. We want all the data they can get
+	 * through the OA buffer to be valid.
+	 */
+	struct i915_request *configure_request;
 };
 
 /**