Re: [PATCH v3] drm/i915: Workaround to avoid lite restore with HEAD==TAIL

[Chris Wilson <chris@chris-wilson.co.uk>]

On Wed, Apr 15, 2015 at 05:17:13PM +0100, Michel Thierry wrote:
> WaIdleLiteRestore is an execlists-only workaround, and requires the driver
> to ensure that any context always has HEAD!=TAIL when attempting lite
> restore.
> 
> Add two extra MI_NOOP instructions at the end of each request, but keep
> the requests tail pointing before the MI_NOOPs. We may not need to
> executed them, and this is why request->tail must be sampled before adding
> these extra instructions.
> 
> If we submit a context to the ELSP which has previously been submitted,
> move the tail pointer past the MI_NOOPs. This ensures HEAD!=TAIL.
> 
> v2: Move overallocation to gen8_emit_request, and added note about
> sampling request->tail in commit message (Chris).
> 
> v3: Remove redundant request->tail assignment in __i915_add_request, in
> lrc mode this is already set in execlists_context_queue.
> Do not add wa implementation details inside gem (Chris).
> 
> Cc: Chris Wilson <chris@chris-wilson.co.uk>
> Signed-off-by: Thomas Daniel <thomas.daniel@intel.com>
> Signed-off-by: Michel Thierry <michel.thierry@intel.com>
> ---
>  drivers/gpu/drm/i915/i915_gem.c  |  3 ++-
>  drivers/gpu/drm/i915/intel_lrc.c | 35 ++++++++++++++++++++++++++++++++++-
>  2 files changed, 36 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
> index 3d5a5a8..980e17c 100644
> --- a/drivers/gpu/drm/i915/i915_gem.c
> +++ b/drivers/gpu/drm/i915/i915_gem.c
> @@ -2400,10 +2400,11 @@ int __i915_add_request(struct intel_engine_cs *ring,
>  		ret = ring->add_request(ring);
>  		if (ret)
>  			return ret;
> +
> +		request->tail = intel_ring_get_tail(ringbuf);
>  	}
>  
>  	request->head = request_start;
> -	request->tail = intel_ring_get_tail(ringbuf);
>  
>  	/* Whilst this request exists, batch_obj will be on the
>  	 * active_list, and so will hold the active reference. Only when this
> diff --git a/drivers/gpu/drm/i915/intel_lrc.c b/drivers/gpu/drm/i915/intel_lrc.c
> index f4a5ef9..0296350 100644
> --- a/drivers/gpu/drm/i915/intel_lrc.c
> +++ b/drivers/gpu/drm/i915/intel_lrc.c
> @@ -427,6 +427,26 @@ static void execlists_context_unqueue(struct intel_engine_cs *ring)
>  		}
>  	}
>  
> +	if (IS_GEN8(ring->dev) || IS_GEN9(ring->dev)) {
> +		/*
> +		 * WaIdleLiteRestore: make sure we never cause a lite
> +		 * restore with HEAD==TAIL
> +		 */
> +		if (req0 && req0->elsp_submitted == 1) {
> +			/*
> +			 * Consume the buffer NOOPs to ensure HEAD != TAIL when
> +			 * submitting. elsp_submitted can only be >1 after
> +			 * reset, in which case we don't need the workaround as
> +			 * a lite restore will not occur.

I actually think you can remove the == 1 and hence remove comment since
the wa is safe to apply in that case as well.

/* Apply the wa NOOPS to prevent ring:HEAD == rq:TAIL as we
 * resubmit the request. See gen8_emit_request() for where we
 * prepare the padding after the end of the request.
 */

> +			 */
> +			struct intel_ringbuffer *ringbuf;
> +
> +			ringbuf = req0->ctx->engine[ring->id].ringbuf;
> +			req0->tail += 8;
> +			req0->tail &= ringbuf->size - 1;
> +		}
> +	}
> +
>  	WARN_ON(req1 && req1->elsp_submitted);
>  
>  	execlists_submit_contexts(ring, req0->ctx, req0->tail,
> @@ -1289,7 +1309,12 @@ static int gen8_emit_request(struct intel_ringbuffer *ringbuf,
>  	u32 cmd;
>  	int ret;
>  
> -	ret = intel_logical_ring_begin(ringbuf, request->ctx, 6);
> +	/*
> +	 * Reserve space for 2 NOOPs at the end of each request to be
> +	 * used as a workaround for not being allowed to do lite
> +	 * restore with HEAD==TAIL (WaIdleLiteRestore).
> +	 */
> +	ret = intel_logical_ring_begin(ringbuf, request->ctx, 8);
>  	if (ret)
>  		return ret;
>  
> @@ -1307,6 +1332,14 @@ static int gen8_emit_request(struct intel_ringbuffer *ringbuf,
>  	intel_logical_ring_emit(ringbuf, MI_NOOP);
>  	intel_logical_ring_advance_and_submit(ringbuf, request->ctx, request);
>  
> +	/*
> +	 * Here we add two extra NOOPs as padding to avoid
> +	 * lite restore of a context with HEAD==TAIL.
> +	 */
> +	intel_logical_ring_emit(ringbuf, MI_NOOP);
> +	intel_logical_ring_emit(ringbuf, MI_NOOP);
> +	intel_logical_ring_advance(ringbuf);
> +

Ok, looks better.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre
_______________________________________________
Intel-gfx mailing list
Intel-gfx@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/intel-gfx