Re: [PATCH 2/9] drm/i915: Extract node allocation from bind

[Ben Widawsky <ben@bwidawsk.net>]

On Wed, May 07, 2014 at 04:53:08PM +0100, Chris Wilson wrote:
> On Wed, May 07, 2014 at 08:45:38AM -0700, Ben Widawsky wrote:
> > On Wed, May 07, 2014 at 08:02:38AM +0100, Chris Wilson wrote:
> > > On Tue, May 06, 2014 at 10:21:31PM -0700, Ben Widawsky wrote:
> > > > The DRM node allocation code was already a bit of an ugly bit of code
> > > > within a complex function. Removing it serves the purpose of cleaning
> > > > the function up. More importantly, it provides a way to have a
> > > > preallocated (address space) VMA to easily skip this code. Something
> > > > we're very likely to need.
> > > > 
> > > > This is essentially a wrapper for DRM node allocation with an eviction +
> > > > retry. It is something which could be moved to core DRM eventually, if
> > > > other drivers had similar eviction semantics.
> > > > 
> > > > This removes a goto used for something other than error unwinding, a
> > > > generally reviled (I am neutral) practice, and replaces it with a
> > > > function call to itself that should have the same effect. Note that it's
> > > > not a recursive function as all the problem set reduction is done in the
> > > > eviction code.
> > > > 
> > > > Some might say this change is worse than before because we are using
> > > > stack for each subsequent call. Frankly, I'd rather overflow the stack
> > > > and blow it up than loop forever. In either case, this is addressed in
> > > > the next patch.
> > > > 
> > > > I believe, and intend, that other than the stack usage, there is no
> > > > functional change here.
> > > 
> > > Nope, but decoupling evict_something() further does introduce lots more
> > > implied magic.
> > > -Chris
> > > 
> > 
> > Hmm. I really don't see what's actually upsetting. Can you be a bit more
> > explicit about what's so bothersome to you for my edification?
> 
> evict_something() make assumptions about the ranges of the vm which it
> searches. At the moment, they mirror its parent's function.
> -Chris
> 

Ah, thanks. So is plumbing starting eviction offset into evict something an
appealing solution here?

> -- 
> Chris Wilson, Intel Open Source Technology Centre

-- 
Ben Widawsky, Intel Open Source Technology Center