[Intel-gfx] [PATCH 0/8] drm/i915/context
bwidawsk at gmail.com
Wed Feb 2 15:55:31 PST 2011
On Wed, Feb 02, 2011 at 11:29:37PM +0000, Chris Wilson wrote:
> I don't think we should be pinning contexts for their entire lifetime.
> The page should only be pinned whilst it is being referenced by the
> hardware, and it looks like you already parts of the infrastructure to
> be able to track this. You will also need to track seqno and retire
I have given thought to not pinning the contexts multiple times, and
believe it is ultimately what we will end up doing. However, the
current implementation is definitely the easier, and more stable way to
go about it. So I'd definitely prefer to keep this as is until the other
parts of the stack are functional.
We need to consider how many clients we will actually have at one time.
I have no idea. If we're talking a common case of 2 or 3 contexts, the
overhead may not be worth it. Retiring of contexts is very simple the
way it is now too because it only happens voluntarily.
> I'm not convinced by the locking strategy, there seems to be quite a few
> holes left, and is something else that is much easier to make it correct
> from the beginning and incrementally adjust.
Yes, I still need to review this, and I totally agree that it must be
right the first time. Is there something in particular that
bothers you about it?
> And then there are lots of little changes to try and make the code
> tidier... But those are probably best left for another day.
Thanks for the feedback.
More information about the Intel-gfx