Re: [PATCH] drm/i915: paper over missed irq issues with force wake vodoo

From: Keith Packard <keithp@keithp.com>
To: Daniel Vetter <daniel@ffwll.ch>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>,
	intel-gfx <intel-gfx@lists.freedesktop.org>,
	Eugeni Dodonov <eugeni.dodonov@intel.com>,
	stable@kernel.org
Subject: Re: [PATCH] drm/i915: paper over missed irq issues with force wake vodoo
Date: Sat, 14 Jan 2012 22:35:44 -0800	[thread overview]
Message-ID: <861ur14gn3.fsf@sumi.keithp.com> (raw)
In-Reply-To: <20120114121152.GA3583@phenom.ffwll.local>

[-- Attachment #1.1: Type: text/plain, Size: 1137 bytes --]

On Sat, 14 Jan 2012 13:12:07 +0100, Daniel Vetter <daniel@ffwll.ch> wrote:

> I think that race is air-tight with your patch to rework the reset code
> already. But better safe than sorry. And as I've said a good cleanup
> anyway.

Sounds good. I like clearer code, especially when it doesn't cost
performance. I think I'd like my version in -fixes so that we don't
change anything with -next; no reason to have two slightly different
versions out there in case one (or the other) is broken?

> One of the reasons Chris originally shot down Ben's forcewake patches
> which protected everything with a spinlock (i.e. also writes) is the
> overhead. And writes to advance the ring are actually rather common. Iirc
> Chris even wrote a patch to cut down on the overhead by caching the fifo
> count. So I think we actually want this asymmetry in locking for
> performance reasons.

Ah, that's a good reason to use different locking for each path
then. Suitable documentation, and a WARN_ON in the write path to check
for the struct_mutex should suffice to prevent mistakes in the future.

-- 
keith.packard@intel.com

[-- Attachment #1.2: Type: application/pgp-signature, Size: 827 bytes --]

[-- Attachment #2: Type: text/plain, Size: 159 bytes --]

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