[Intel-gfx] [PATCH 03/12] drm/i915: Framework for capturing command stream based OA reports and ctx id info.

[Chris Wilson]

Quoting Sagar Arun Kamble (2017-07-31 08:59:36)
> +/**
> + * i915_perf_stream_emit_sample_capture - Insert the commands to capture perf
> + * metrics into the GPU command stream
> + * @stream: An i915-perf stream opened for GPU metrics
> + * @request: request in whose context the metrics are being collected.
> + * @preallocate: allocate space in ring for related sample.
> + */
> +static void i915_perf_stream_emit_sample_capture(
> +                                       struct i915_perf_stream *stream,
> +                                       struct drm_i915_gem_request *request,
> +                                       bool preallocate)
> +{
> +       struct reservation_object *resv = stream->cs_buffer.vma->resv;
> +       struct i915_perf_cs_sample *sample;
> +       unsigned long flags;
> +       int ret;
> +
> +       sample = kzalloc(sizeof(*sample), GFP_KERNEL);
> +       if (sample == NULL) {
> +               DRM_ERROR("Perf sample alloc failed\n");
> +               return;
> +       }
> +
> +       sample->request = i915_gem_request_get(request);
> +       sample->ctx_id = request->ctx->hw_id;
> +
> +       insert_perf_sample(stream, sample);
> +
> +       if (stream->sample_flags & SAMPLE_OA_REPORT) {
> +               ret = i915_emit_oa_report_capture(request,
> +                                                 preallocate,
> +                                                 sample->offset);
> +               if (ret)
> +                       goto err_unref;
> +       }

This is incorrect as the requests may be reordered. You either need to
declare the linear ordering of requests through the sample buffer, or we
have to delay setting sample->offset until execution, and even then we
need to disable preemption when using SAMPLE_OA_REPORT.
-Chris