[Intel-gfx] [PATCH 1/6] intel: Defer setting madv on the bo cache

[Chris Wilson]

On Fri, Jun 05, 2015 at 09:34:23AM -0700, Ben Widawsky wrote:
> Sorry. I know I asked some questions last time around, but I am having trouble
> finding the response.

You thought the existing code used a one second timeout and so wanted a
separated patch to convert from age <= 1 to age < 2.

Other than that you should have pointed out that dontneed should be
moved into madvise_internal() so that it is kept in sync with the kernel
value to avoid EFAULT embarrassment.

> Just to be clear, you're reducing IOCTL traffic by deferring an IOCTL until you
> cleanup the BO cache.

Right and thereby reducing ioctls.

> I think I said it last time, overall the idea seems fine to me.

The change is marginal except when there is a lot of mutex contention
(e.g. other clients clflushing) and then the effect is to push the
contention to until it is unavoidable.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre