[Intel-gfx] [PATCH] drm/i915/perf: don't forget noa wait after oa config

[Chris Wilson]

Quoting Lionel Landwerlin (2019-11-13 15:46:39)
> I'm observing incoherence metric values, changing from run to run.
> 
> It appears the patches introducing noa wait & reconfiguration from
> command stream switched places in the series multiple times during the
> review. This lead to the dependency of one onto the order to go
> missing...

I don't think I dropped it; if I did my apologies. I do feel the
egg-on-face for writing a selftest to verify that noa_wait does what you
said it did, but completely missing that it went unused :)
 
> Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin at intel.com>
> Fixes: 15d0ace1f876 ("drm/i915/perf: execute OA configuration from command stream")
> ---
>  drivers/gpu/drm/i915/i915_perf.c | 9 +++++++--
>  1 file changed, 7 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/gpu/drm/i915/i915_perf.c b/drivers/gpu/drm/i915/i915_perf.c
> index 507236bd41ae..31e47ee23357 100644
> --- a/drivers/gpu/drm/i915/i915_perf.c
> +++ b/drivers/gpu/drm/i915/i915_perf.c
> @@ -1870,7 +1870,7 @@ alloc_oa_config_buffer(struct i915_perf_stream *stream,
>         config_length += num_lri_dwords(oa_config->mux_regs_len);
>         config_length += num_lri_dwords(oa_config->b_counter_regs_len);
>         config_length += num_lri_dwords(oa_config->flex_regs_len);
> -       config_length++; /* MI_BATCH_BUFFER_END */
> +       config_length += 3; /* MI_BATCH_BUFFER_START */
>         config_length = ALIGN(sizeof(u32) * config_length, I915_GTT_PAGE_SIZE);
>  
>         obj = i915_gem_object_create_shmem(stream->perf->i915, config_length);
> @@ -1895,7 +1895,12 @@ alloc_oa_config_buffer(struct i915_perf_stream *stream,
>                              oa_config->flex_regs,
>                              oa_config->flex_regs_len);
>  
> -       *cs++ = MI_BATCH_BUFFER_END;
> +       /* Jump into the active wait. */
> +       *cs++ = (INTEL_GEN(stream->perf->i915) < 8 ?
> +                MI_BATCH_BUFFER_START :
> +                MI_BATCH_BUFFER_START_GEN8);
> +       *cs++ = i915_ggtt_offset(stream->noa_wait);
> +       *cs++ = 0;

Yikes, stream->noa_wait is unused.

Hmm, the noa_wait doesn't have any arbitration points internally, so we
probably do need to make it non-preemptable as well?

With a rq->flags |= I915_REQUEST_NOPREEMPT in emit_oa_config,
Reviewed-by: Chris Wilson <chris at chris-wilson.co.uk>

We need to wrap emit_oa_config() in a similar selftest and verify
that a read of the oa regs are correct and that the TIMESTAMP indicates
the appropriate delay before the read. If you feel bored.
-Chris