[Intel-gfx] [PATCH 07/51] drm/i915: Early alloc request in execbuff

[Daniel Vetter]

On Fri, Feb 27, 2015 at 12:27:15PM +0000, John Harrison wrote:
> On 25/02/2015 22:22, Daniel Vetter wrote:
> >On Fri, Feb 13, 2015 at 11:48:16AM +0000, John.C.Harrison at Intel.com wrote:
> >>From: John Harrison <John.C.Harrison at Intel.com>
> >>
> >>Start of explicit request management in the execbuffer code path. This patch
> >>adds a call to allocate a request structure before all the actual hardware work
> >>is done. Thus guaranteeing that all that work is tagged by a known request. At
> >>present, nothing further is done with the request, the rest comes later in the
> >>series.
> >>
> >>The only noticable change is that failure to get a request (e.g. due to lack of
> >>memory) will be caught earlier in the sequence. It now occurs right at the start
> >>before any un-undoable work has been done.
> >>
> >>For: VIZ-5115
> >>Signed-off-by: John Harrison <John.C.Harrison at Intel.com>
> >>---
> >>  drivers/gpu/drm/i915/i915_gem_execbuffer.c |   13 ++++++++++---
> >>  1 file changed, 10 insertions(+), 3 deletions(-)
> >>
> >>diff --git a/drivers/gpu/drm/i915/i915_gem_execbuffer.c b/drivers/gpu/drm/i915/i915_gem_execbuffer.c
> >>index ca85803..61471e9 100644
> >>--- a/drivers/gpu/drm/i915/i915_gem_execbuffer.c
> >>+++ b/drivers/gpu/drm/i915/i915_gem_execbuffer.c
> >>@@ -1356,7 +1356,7 @@ i915_gem_do_execbuffer(struct drm_device *dev, void *data,
> >>  	const u32 ctx_id = i915_execbuffer2_get_context_id(*args);
> >>  	u32 dispatch_flags;
> >>  	int ret;
> >>-	bool need_relocs;
> >>+	bool need_relocs, batch_pinned = false;
> >>  	if (!i915_gem_check_execbuffer(args))
> >>  		return -EINVAL;
> >>@@ -1525,10 +1525,16 @@ i915_gem_do_execbuffer(struct drm_device *dev, void *data,
> >>  		if (ret)
> >>  			goto err;
> >>+		batch_pinned = true;
> >>  		params->batch_obj_vm_offset = i915_gem_obj_ggtt_offset(batch_obj);
> >>  	} else
> >>  		params->batch_obj_vm_offset = i915_gem_obj_offset(batch_obj, vm);
> >>+	/* Allocate a request for this batch buffer nice and early. */
> >>+	ret = dev_priv->gt.alloc_request(ring, ctx);
> >>+	if (ret)
> >>+		goto err;
> >>+
> >>  	/*
> >>  	 * Save assorted stuff away to pass through to *_submission().
> >>  	 * NB: This data should be 'persistent' and not local as it will
> >>@@ -1544,15 +1550,16 @@ i915_gem_do_execbuffer(struct drm_device *dev, void *data,
> >>  	ret = dev_priv->gt.do_execbuf(params, args, &eb->vmas);
> >>+err:
> >>  	/*
> >>  	 * FIXME: We crucially rely upon the active tracking for the (ppgtt)
> >>  	 * batch vma for correctness. For less ugly and less fragility this
> >>  	 * needs to be adjusted to also track the ggtt batch vma properly as
> >>  	 * active.
> >>  	 */
> >>-	if (dispatch_flags & I915_DISPATCH_SECURE)
> >>+	if (batch_pinned)
> >>  		i915_gem_object_ggtt_unpin(batch_obj);
> >>-err:
> >>+
> >>  	/* the request owns the ref now */
> >>  	i915_gem_context_unreference(ctx);
> >>  	eb_destroy(eb);
> >This hunk here looks wrong, or maybe the context changed sufficiently
> >already (but I can't find that in previous patches). Why do we need to
> >change the pinning for the ggtt batch pin hack when allocating the request
> >earlier?
> >-Daniel
> 
> It doesn't change the behaviour. It is just coping with the extra error
> path. If the alloc request call fails, we need to jump past the do_execbuf()
> call but not past the batch unpin. Hence the 'err:' tag is moved upwards.
> That means it is now possible to get to the batch unpin test from an error
> that occurred before the pin call actually happened. Hence it is no longer
> safe to just test the dispatch flag. Instead an explicit 'did this get
> pinned yet' boolean is required.

Ah I see what's going on. The usual pattern is to just duplicate goto
targets, i.e. here:

err_batch_unpin:
 	/*
 	 * FIXME: We crucially rely upon the active tracking for the (ppgtt)
 	 * batch vma for correctness. For less ugly and less fragility this
 	 * needs to be adjusted to also track the ggtt batch vma properly as
 	 * active.
 	 */
	if (dispatch_flags & I915_DISPATCH_SECURE)
 		i915_gem_object_ggtt_unpin(batch_obj);
err:

 	/* the request owns the ref now */
 	i915_gem_context_unreference(ctx);
 	eb_destroy(eb);

Then you can let gcc deal with how to best structure the control flow for
you. It means a bit larger diff since you need to change a pile of goto
err; to goto err_batch_unpin, but imo that also makes it easier for
reviewers to double-check there's no error case with funky requirements
now broken.

Cheers, Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch