[Intel-gfx] [PATCH] drm/i915: Only mark the ctx as initialised after a SET_CONTEXT operation

[Chris Wilson]

On Fri, May 30, 2014 at 10:44:53AM -0700, Ben Widawsky wrote:
> On Fri, May 30, 2014 at 02:16:30PM +0100, Chris Wilson wrote:
> > Fallout from
> > 
> > commit 46470fc932ac8a0e8317a220b3f4ea4ed903338e
> > Author: Mika Kuoppala <mika.kuoppala at linux.intel.com>
> > Date:   Wed May 21 19:01:06 2014 +0300
> > 
> >     drm/i915: Add null state batch to active list
> > 
> > undid the earlier fix of only marking the ctx as initialised after it is
> > saved by the hardware during a SET_CONTEXT operation.
> > 
> > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> > Cc: Ville Syrjälä <ville.syrjala at linux.intel.com>
> > Cc: Damien Lespiau <damien.lespiau at intel.com>
> > Cc: Mika Kuoppala <mika.kuoppala at intel.com>
> > Cc: Ben Widawsky <ben at bwidawsk.net>
> > ---
> >  drivers/gpu/drm/i915/i915_gem_context.c | 8 +++++---
> >  1 file changed, 5 insertions(+), 3 deletions(-)
> > 
> > diff --git a/drivers/gpu/drm/i915/i915_gem_context.c b/drivers/gpu/drm/i915/i915_gem_context.c
> > index 5a71ef1975b3..34a0b49e6add 100644
> > --- a/drivers/gpu/drm/i915/i915_gem_context.c
> > +++ b/drivers/gpu/drm/i915/i915_gem_context.c
> > @@ -641,6 +641,7 @@ static int do_switch(struct intel_engine_cs *ring,
> >  	struct intel_context *from = ring->last_context;
> >  	struct i915_hw_ppgtt *ppgtt = ctx_to_ppgtt(to);
> >  	u32 hw_flags = 0;
> > +	bool uninitialized = false;
> >  	int ret, i;
> >  
> >  	if (from != NULL && ring == &dev_priv->ring[RCS]) {
> > @@ -739,18 +740,19 @@ static int do_switch(struct intel_engine_cs *ring,
> >  		i915_gem_context_unreference(from);
> >  	}
> >  
> > +	uninitialized = !to->is_initialized && from == NULL;
> > +	to->is_initialized = true;
> 
> Aren't you missing some error paths if you do this?  I think we need to
> set the state after we've successfully emitted the MI_SET_CONTEXT
> command (though really, until we move the ringbuffer tail, it's all
> lies, on that note, I'm scared to look at if reset dtrt).

I also think that until we successfully emit the MI_SET_CONTEXT and
allow the execbuffer to proceed, we can continue to treat the context as
garbage and force-inhibit on the next attempt to run the execbuffer.
 
> > +
> >  done:
> >  	i915_gem_context_reference(to);
> >  	ring->last_context = to;
> >  
> > -	if (ring->id == RCS && !to->is_initialized && from == NULL) {
> > +	if (uninitialized) {
> >  		ret = i915_gem_render_state_init(ring);
> >  		if (ret)
> >  			DRM_ERROR("init render state: %d\n", ret);
> >  	}
> >  
> > -	to->is_initialized = true;
> > -
> >  	return 0;
> >  
> >  unpin_out:
> 
> I just realized the from == NULL check is something which seems fragile
> to me given how much churn we might see with init/reset paths. It'd
> probably be good to eventually switch to some global state in dev_priv
> which we can track across init/reset/fini

It's as fragile as the bug we are papering over. Presumably we do need
to apply fresh paper after a GPU reset, but it is hard to tell without
knowing the bug.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre