[Intel-gfx] [RFC PATCH 4/4] drm/i915: reprogram NOA muxes on context switch when using perf

[Chris Wilson]

Quoting Lionel Landwerlin (2017-08-30 19:20:06)
> If some of the contexts submitting workloads to the GPU have been
> configured to shutdown slices/subslices, we might loose the NOA
> configurations written in the NOA muxes. We need to reprogram then at
> context switch.
> 
> Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin at intel.com>
> ---
>  drivers/gpu/drm/i915/i915_drv.h  |  2 ++
>  drivers/gpu/drm/i915/i915_perf.c | 77 ++++++++++++++++++++++++++++++++++++++++
>  drivers/gpu/drm/i915/intel_lrc.c | 64 ++++++++++++++++++++++++++++++---
>  drivers/gpu/drm/i915/intel_lrc.h |  1 +
>  4 files changed, 140 insertions(+), 4 deletions(-)
> 
> diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
> index 0003b46b6840..d4b3e5da9009 100644
> --- a/drivers/gpu/drm/i915/i915_drv.h
> +++ b/drivers/gpu/drm/i915/i915_drv.h
> @@ -3685,6 +3685,8 @@ int i915_perf_remove_config_ioctl(struct drm_device *dev, void *data,
>  void i915_oa_init_reg_state(struct intel_engine_cs *engine,
>                             struct i915_gem_context *ctx,
>                             uint32_t *reg_state);
> +u32 i915_oa_get_perctx_bb_size(struct drm_i915_private *dev_priv);
> +u32 *i915_oa_emit_perctx_bb(struct intel_engine_cs *engine, u32 *batch);
>  
>  /* i915_gem_evict.c */
>  int __must_check i915_gem_evict_something(struct i915_address_space *vm,
> diff --git a/drivers/gpu/drm/i915/i915_perf.c b/drivers/gpu/drm/i915/i915_perf.c
> index 94185d610673..b74ffbb47879 100644
> --- a/drivers/gpu/drm/i915/i915_perf.c
> +++ b/drivers/gpu/drm/i915/i915_perf.c
> @@ -1687,6 +1687,74 @@ static int gen8_emit_oa_config(struct drm_i915_gem_request *req,
>         return 0;
>  }
>  
> +#define MAX_LRI_SIZE (125U)
> +
> +u32 i915_oa_get_perctx_bb_size(struct drm_i915_private *dev_priv)
> +{
> +       struct i915_perf_stream *stream = dev_priv->perf.oa.exclusive_stream;
> +
> +       lockdep_assert_held(&dev_priv->drm.struct_mutex);

Still not happy by this coupling to struct_mutex. :-p

Missed RCS check.

> +
> +       /* Perf not supported. */
> +       if (!dev_priv->perf.initialized)
> +               return 0;
> +
> +       /* OA not currently configured. */
> +       if (!stream)
> +               return 0;
> +
> +       /* Very unlikely but possible that we have no muxes to configure. */
> +       if (!stream->oa_config->mux_regs_len)
> +               return 0;
> +
> +       /* Return the size of MI_LOAD_REGISTER_IMMs. */
> +       return (stream->oa_config->mux_regs_len / MAX_LRI_SIZE) * 4 + 4 +
> +               stream->oa_config->mux_regs_len * 8;
> +}
> +
> +u32 *i915_oa_emit_perctx_bb(struct intel_engine_cs *engine, u32 *batch)
> +{
> +       struct drm_i915_private *dev_priv = engine->i915;
> +       struct i915_perf_stream *stream = dev_priv->perf.oa.exclusive_stream;
> +       u32 n_lri, n_mux_regs;
> +       u32 i;
> +
> +       lockdep_assert_held(&dev_priv->drm.struct_mutex);
> +
> +       /* We only care about RCS. */
> +       if (engine->id != RCS)
> +               return batch;
> +
> +       /* Perf not supported. */
> +       if (!dev_priv->perf.initialized)
> +               return batch;
> +
> +       /* OA not currently configured. */
> +       if (!stream)
> +               return batch;
> +
> +       /* It's very unlikely, but possible that we're dealing with a config
> +        * with no mux to configure.
> +        */
> +       if (!stream->oa_config->mux_regs_len)
> +               return batch;

The above could be condensed into
if (i915_oa_get_perctx_bb_size() == 0)
	return;

> +
> +       n_mux_regs = stream->oa_config->mux_regs_len;
> +       n_lri = (n_mux_regs / MAX_LRI_SIZE) + (n_mux_regs % MAX_LRI_SIZE) != 0;
> +
> +       for (i = 0; i < n_mux_regs; i++) {
> +               if ((i % MAX_LRI_SIZE) == 0) {
> +                       n_lri = min(n_mux_regs - i, MAX_LRI_SIZE);
> +                       *batch++ = MI_LOAD_REGISTER_IMM(n_lri);
> +               }
> +
> +               *batch++ = i915_mmio_reg_offset(stream->oa_config->mux_regs[i].addr);
> +               *batch++ = stream->oa_config->mux_regs[i].value;
> +       }

I would have personally used a double loop. But at least kill that first
n_lri, that was a moment of confusion spent trying to work out what you
were using it for.

> +
> +       return batch;
> +}

>  /**
>   * intel_sanitize_enable_execlists() - sanitize i915.enable_execlists
> @@ -1055,6 +1057,8 @@ static u32 *gen8_init_indirectctx_bb(struct intel_engine_cs *engine, u32 *batch)
>   */
>  static u32 *gen8_init_perctx_bb(struct intel_engine_cs *engine, u32 *batch)
>  {
> +       batch = i915_oa_emit_perctx_bb(engine, batch);
> +
>         /* WaDisableCtxRestoreArbitration:bdw,chv */
>         *batch++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;
>         *batch++ = MI_BATCH_BUFFER_END;
> @@ -1118,21 +1122,27 @@ static u32 *gen9_init_indirectctx_bb(struct intel_engine_cs *engine, u32 *batch)
>  
>  static u32 *gen9_init_perctx_bb(struct intel_engine_cs *engine, u32 *batch)
>  {
> +       batch = i915_oa_emit_perctx_bb(engine, batch);

Wrong wa_bb. This is emitted at the start of every bb, you want
indirectctx_bb which is emitted after a context switch. Or at least I
hope you don't need such a heavy handed approach. If you do, can you
convince me that you are really not in conflict with the application?
(Earlier you said it only needs ctx switch.)

> +int logical_render_ring_reload_wa_bb(struct intel_engine_cs *engine)

intel_lrc_update_wa_bb().

> +{
> +       struct drm_i915_private *dev_priv = engine->i915;
> +       struct i915_ctx_workarounds new_wa_ctx;
> +       struct i915_gem_context *ctx;
> +       int ret;
> +
> +       if (WARN_ON(engine->id != RCS))
> +               return -EINVAL;
> +
> +       memset(&new_wa_ctx, 0, sizeof(new_wa_ctx));
> +       ret = intel_init_workaround_bb(engine, &new_wa_ctx);
> +       if (ret)
> +               return ret;
> +
> +       if (engine->wa_ctx.vma)
> +               lrc_destroy_wa_ctx(engine);

Couldn't we at least try to reuse the existing vma first?