Re: [PATCH 1/4] drm/i915: Clearing buffer objects via blitter engine

[Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com>]

On 07/01/2015 05:30 PM, Chris Wilson wrote:
> On Wed, Jul 01, 2015 at 03:54:55PM +0100, Tvrtko Ursulin wrote:
>>> +static int i915_gem_exec_flush_object(struct drm_i915_gem_object *obj,
>>> +				      struct intel_engine_cs *ring,
>>> +				      struct intel_context *ctx,
>>> +				      struct drm_i915_gem_request **req)
>>> +{
>>> +	int ret;
>>> +
>>> +	ret = i915_gem_object_sync(obj, ring, req);
>>> +	if (ret)
>>> +		return ret;
>>> +
>>> +	if (obj->base.write_domain & I915_GEM_DOMAIN_CPU) {
>>> +		if (i915_gem_clflush_object(obj, false))
>>> +			i915_gem_chipset_flush(obj->base.dev);
>>> +		obj->base.write_domain &= ~I915_GEM_DOMAIN_CPU;
>>> +	}
>>> +	if (obj->base.write_domain & I915_GEM_DOMAIN_GTT) {
>>> +		wmb();
>>> +		obj->base.write_domain &= ~I915_GEM_DOMAIN_GTT;
>>> +	}
>>
>> All this could be replaced with i915_gem_object_set_to_gtt_domain, no?
>
> No. Technically this is i915_gem_execbuffer_move_to_gpu().

Aha.. I see now what was my confusion. It doesn't help that 
i915_gem_execbuffer_move_to_gpu and execlist_move_to_gpu are implemented 
at different places logically.

It would be nice to extract the loop body then call it something like 
i915_gem_execbuffer_move_vma_to_gpu, it would avoid at least three 
instances of the same code.

>>> +
>>> +	return i915.enable_execlists ?
>>> +			logical_ring_invalidate_all_caches(*req) :
>>> +			intel_ring_invalidate_all_caches(*req);
>>
>> And this is done on actual submission for you by the lower levels so
>> you don't need to call it directly.
>
> What submission? We don't build a batch, we are building a raw request
> to do the operation from the ring.

I was confused to where execlist_move_to_gpu is in the stack.

>>> +	lockdep_assert_held(&dev->struct_mutex);
>>
>> It think there was some guidance that lockdep_assert_held is
>> compiled out when lockdep is not in the kernel and that WARN_ON is
>> preferred. In this case that would probably be WARN_ON_ONCE and
>> return error.
>
> Hah, this predates that and I still disagree.

Predates or not is not relevant. :) It is not a clean cut situation I 
agree. Maybe we need our own amalgamation on WARN_ON_ONCE and 
lockdep_assert_held but I think we either check for these things or not, 
or have a really good assurance of test coverage with lockdep enabled 
during QA.

>>> +	ring = &dev_priv->ring[HAS_BLT(dev) ? BCS : RCS];
>>> +	ctx = i915_gem_context_get(file_priv, DEFAULT_CONTEXT_HANDLE);
>>> +	/* Allocate a request for this operation nice and early. */
>>> +	ret = i915_gem_request_alloc(ring, ctx, &req);
>>> +	if (ret)
>>> +		return ret;
>>> +
>>> +	if (ctx->ppgtt)
>>> +		vm = &ctx->ppgtt->base;
>>> +	else
>>> +		vm = &dev_priv->gtt.base;
>>> +
>>> +	if (i915.enable_execlists && !ctx->engine[ring->id].state) {
>>> +		ret = intel_lr_context_deferred_create(ctx, ring);
>>
>> i915_gem_context_get above and this call are very similar to what
>> i915_gem_validate_context does. It seems to me it would be better to
>> call the latter function here.
>
> No, the intel_lrc API is absolute garbage and needs to be taken out the
> back and shot. Until that is done, I wouldn't bother continuing to try
> and use the interface at all.
>
> All that needs to happen here is:
>
> req = i915_gem_request_alloc(ring, ring->default_context);
>
> and for the request/lrc to go off and dtrt.

Well.. I the meantime why duplicate it when i915_gem_validate_context 
does i915_gem_context_get and deferred create if needed. I don't see the 
downside. Snippet from above becomes:

   ring = &dev_priv->ring[HAS_BLT(dev) ? BCS : RCS];
   ctx = i915_gem_validate_context(dev, file, ring,
				DFAULT_CONTEXT_HANDLE);
   ... handle error...
   /* Allocate a request for this operation nice and early. */
   ret = i915_gem_request_alloc(ring, ctx, &req);

Why would this code have to know about deferred create.

>>> +	}
>>> +
>>> +	ringbuf = ctx->engine[ring->id].ringbuf;
>>> +
>>> +	ret = i915_gem_object_pin(obj, vm, PAGE_SIZE, 0);
>>> +	if (ret)
>>> +		return ret;
>>> +
>>> +	if (obj->tiling_mode && INTEL_INFO(dev)->gen <= 3) {
>>> +		ret = i915_gem_object_put_fence(obj);
>>> +		if (ret)
>>> +			goto unpin;
>>> +	}
>>
>> Why is this needed?
>
> Because it's a requirement of the op being used on those gen. Later gen
> can keep the fence.
 >
>> Could it be called unconditionally and still work?
>>
>> At least I would recommend a comment explaining it.

It is ugly to sprinkle platform knowledge to the callers - I think I saw 
a callsites which call i915_gem_object_put_fence unconditionally so why 
would that not work here?

>>> +	if (i915.enable_execlists) {
>>> +		if (dev_priv->info.gen >= 8) {
>>> +			ret = intel_logical_ring_begin(req, 8);
>>> +			if (ret)
>>> +				goto unpin;
>>> +
>>> +			intel_logical_ring_emit(ringbuf, GEN8_COLOR_BLT_CMD |
>>> +							 BLT_WRITE_RGBA |
>>> +							 (7-2));
>>> +			intel_logical_ring_emit(ringbuf, BPP_32 |
>>> +							 ROP_FILL_COPY |
>>> +							 PAGE_SIZE);
>>> +			intel_logical_ring_emit(ringbuf, 0);
>>> +			intel_logical_ring_emit(ringbuf,
>>> +						obj->base.size >> PAGE_SHIFT
>>> +						<< 16 | PAGE_SIZE / 4);
>>> +			intel_logical_ring_emit(ringbuf,
>>> +						i915_gem_obj_offset(obj, vm));
>>> +			intel_logical_ring_emit(ringbuf, 0);
>>> +			intel_logical_ring_emit(ringbuf, 0);
>>> +			intel_logical_ring_emit(ringbuf, MI_NOOP);
>>> +
>>> +			intel_logical_ring_advance(ringbuf);
>>> +		} else {
>>> +			DRM_ERROR("Execlists not supported for gen %d\n",
>>> +				  dev_priv->info.gen);
>>> +			ret = -EINVAL;
>>
>> I would put a WARN_ON_ONCE here, or even just return -EINVAL. If the
>> driver is so messed up in general that execlists are enabled < gen8
>> I think there is no point logging errors about it from here. Would
>> also save you one indentation level.
>
> I would just rewrite this to have a logical interface to the rings. Oh
> wait, I did.

That is out of my jurisdiction, but I think my comment to the above is 
not an unreasonable one since it indicates total driver confusion and 
could/should be handled somewhere else.

Regards,

Tvrtko
_______________________________________________
Intel-gfx mailing list
Intel-gfx@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/intel-gfx