Re: [PATCH] drm/i915/bdw: MU_FLUSH_DW a qword instead of dword

From: Ben Widawsky <ben@bwidawsk.net>
To: Chris Wilson <chris@chris-wilson.co.uk>,
	Ben Widawsky <benjamin.widawsky@intel.com>,
	Daniel Vetter <daniel@ffwll.ch>,
	Intel GFX <intel-gfx@lists.freedesktop.org>
Subject: Re: [PATCH] drm/i915/bdw: MU_FLUSH_DW a qword instead of dword
Date: Wed, 5 Mar 2014 14:38:50 -0800	[thread overview]
Message-ID: <20140305223849.GA24362@bwidawsk.net> (raw)
In-Reply-To: <20140305223021.GA3559@nuc-i3427.alporthouse.com>

On Wed, Mar 05, 2014 at 10:30:21PM +0000, Chris Wilson wrote:
> On Wed, Mar 05, 2014 at 11:05:15AM -0800, Ben Widawsky wrote:
> > On Wed, Mar 05, 2014 at 07:33:11PM +0100, Daniel Vetter wrote:
> > > On Wed, Mar 05, 2014 at 09:24:34AM +0000, Chris Wilson wrote:
> > > > On Tue, Mar 04, 2014 at 09:38:56AM -0800, Ben Widawsky wrote:
> > > > > The actual post sync op is "Write Immediate Data QWord." It is therefore
> > > > > arguable that we should have always done a qword write.
> > > > 
> > > > Not really since the spec explicitly says that we can choose either a
> > > > dword or qword write. Note that qword writes also currently require a
> > > > 64 byte alignment.
> > > 
> > > Yeah, that's also my reading of the spec - the lenght field selects
> > > whether the hw does a qword or dword write, and the qword needs to be
> > > specially aligned.
> > > -Daniel
> > 
> > I think both of you only read this sentence, where I said it was
> > "arguable." The rest of the commit message was what actually mattered.
> 
> I'm just arguing that the changelog is misleading. What we are doing is
> papering over an elephant, and more importantly I think it overlooked
> the extra restrictions imposed upon qwords (though it looks like we
> fortuituously are ok). The changelog also implies that all our other
> code is similarly flawed.

It wasn't completely fortuitous, I did check. I was lucky you think my
check was satisfactory though. I agree it makes future code somewhat
risky so maybe some improvement is needed to safeguard. I also have/had
a patch to lengthen MI_STORE_DATA_INDEX. The decoder however does not
complain about that one, and the windows team did neither. So I didn't
want to change it for the sake of change.

I think the reasons for FLUSH_DW are valid, but as it seems unrelated to
the actual root cause of the bug, I'll leave this one to the fates.

> 
> The actual patch of splitting the code up into separate gen8 routines I
> thought was a nice improvement in readibility.
> -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

-- 
Ben Widawsky, Intel Open Source Technology Center