powerpc: Add user-return-notifier support

[Scott Wood]

On Wed, 2015-09-02 at 08:07 +0530, Ananth N Mavinakayanahalli wrote:
> On Tue, Sep 01, 2015 at 07:03:12PM -0500, Scott Wood wrote:
> > On Tue, 2015-09-01 at 12:11 +0530, Ananth N Mavinakayanahalli wrote:
> > > On Mon, Aug 31, 2015 at 08:35:17PM +1000, Michael Ellerman wrote:
> > > > On Tue, 2015-25-08 at 05:41:10 UTC, Ananth N Mavinakayanahalli wrote:
> > > > > Add user return notifier support for powerpc. Similar to x86, this 
> > > > > feature
> > > > > keys off of the KVM Kconfig.
> > > > 
> > > > Please flesh this out.
> > > > 
> > > > What is it, why do we want it, why is your implementation correct.
> > > 
> > > The only current in-kernel user of the infrastructure is KVM on x86. It
> > > is useful for optimizations when MSR values can continue to be used
> > > between the host and guest. Commit log for 18863bdd60f8 upstream has a
> > > more complete explanation.
> > > 
> > > I do not know the inner details of the KVM implementation on Power,
> > > perhaps Paul/Gautham can comment on if a similar optimization will
> > > benefit Power systems too?
> > 
> > "MSR" is x86-specific terminology and is pretty vague.  What specifically 
> > is 
> > the functionality being optimized, in terms of things that actually exist 
> > on 
> > PPC?
> > 
> > In any case, that commit log doesn't explain what user-return-notifier is 
> > or 
> > how it works, just that it's being used.
> 
> User return notifier is a simple mechanism to fire a one-shot custom 
> handler in
> the context of the thread its registered against, just before it returns to
> userspace. It works by setting a TIF flag to indicate that a handler needs 
> to
> run, and on the userspace return path (do_notify_resume()), this flag is 
> checked
> to call any registered handlers.

Thanks.

> > > We could certainly make this a generic config option.. but I am yet to
> > > see a real usecase outside of the KVM thingy. We do use TIF_UPROBE for
> > > something very similar, though that needs to fire much before the
> > > outstanding signals get handled. The user-return notifier is the last
> > > thing executed before return to userspace -- the two cannot be merged.
> > > 
> > > Ananth
> > > 
> > > PS: I am not sure having an 'ok' against the Documentation/features/ for
> > > powerpc is a valid enough argument for this :-)
> > 
> > So how did you test this?  What platforms did you test it on?  What 
> > hardware 
> > support does it need?
> 
> Simple kernel module below... Its fairly evident from the patch that the
> feature is a software only construct and no specific hardware support is
> needed.

It's evident now that I'm aware of what the purpose is... It wasn't 100% 

clear before, at least from a quick glance, given that the explanation was 

"look at this hardware-specific KVM patch", whether there was an assumption 
of something else going on -- and even things that are implemented in 
software often work differently on book3s versus book3e, ppc32 versus ppc64, 
etc.

Why is this selected by KVM on PPC if KVM on PPC doesn't use it?  What is the 
user you're trying to enable?

Where is the "profiler" that Documentation/features/debug/user-ret-
profiler/arch-support.txt hints at?

-Scott