Re: [TuxOnIce-devel] [RFC] TuxOnIce

[Rafael J. Wysocki]

On Friday 08 May 2009, Jesse Barnes wrote:
> On Fri, 08 May 2009 09:32:34 +1000
> Nigel Cunningham <nigel@xxxxxxxxxxxx> wrote:
> 
> > Hi.
> > 
> > On Thu, 2009-05-07 at 16:14 -0700, Jesse Barnes wrote:
> > > On Fri, 08 May 2009 06:41:00 +1000
> > > Nigel Cunningham <nigel@xxxxxxxxxxxx> wrote:
> > > 
> > > > Hi.
> > > > 
> > > > On Thu, 2009-05-07 at 21:27 +0200, Rafael J. Wysocki wrote:
> > > > > In fact I agree, but there's a catch.  The way in which TuxOnIce
> > > > > operates LRU pages is based on some assumptions that may or may
> > > > > not be satisfied in future, so if we decide to merge it, then
> > > > > we'll have to make sure these assumptions will be satisfied.
> > > > > That in turn is going to require quite some discussion I guess.
> > > > 
> > > > Agreed. That's why I've got that GEMS patch - it's putting pages
> > > > on the LRU that don't satisfy the former assumptions: they are
> > > > used during hibernating and need to be atomically copied. If
> > > > there are further developments in that area, I would hope we
> > > > could just extend what's been done with GEMS.
> > > 
> > > Another option here would be to suspend all DRM operations earlier.
> > > The suspend hook for i915 already does this, but maybe it needs to
> > > happen sooner?  We'll probably want a generic DRM suspend hook soon
> > > too (as the radeon memory manager lands) to shut down GPU activity
> > > in the suspend and hibernate cases.
> > > 
> > > All that assumes I understand what's going on here though. :)  It
> > > appears you delay saving the GEM (just GEM by the way, for
> > > Graphics/GPU Execution Manager) backing store until late to avoid
> > > having the pages move around out from under you?
> > 
> > Yeah. TuxOnIce saves some pages without doing an atomic copy of them.
> > Up 'til now, the algorithm has been LRU pages - pages used for
> > TuxOnIce's userspace helpers. With GEM, we also need to make sure GEM
> > pages are atomically copied and so also 'subtract' them from the list
> > of pages that aren't atomically copied.
> > 
> > It's no great problem to do this, so I wouldn't ask you to change GEM
> > to suspend DRM operations earlier. It's more important that GEM
> > doesn't allocate extra pages unexpectedly - and I don't think that's
> > likely anyway since we've switched away from X. This is important
> > because TuxOnIce depends (for reliability) on having memory usage
> > being predictable much more than swsusp and uswsusp do. (Larger
> > images, less free RAM to begin with).
> 
> Yeah X is typically the one causing GEM allocations and performing
> execution, but there are other possibilities too.  E.g. Wayland is a
> non-X based display system that may be running instead, or maybe
> there's an EGL or GPGPU program running in the background.
> 
> So I think it's best if we suspend DRM fairly early, otherwise you
> *may* get extra allocations and will probably see all sorts of GPU
> memory mapping activity and execution while you're trying to hibernate
> things.  On the plus side I don't think this is a radical redesign or
> anything, and mostly something we can do in our suspend and hibernate
> callbacks.

I think hibernate callbacks are too late for that.  PM notifiers, OTOH, are
probably too early.

So, I really think we'll need to do something special about it.

Thanks,
Rafael
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/