Re: [PATCH v4 2/2] add icswx support

[Benjamin Herrenschmidt <benh@au1.ibm.com>]

On Fri, 2011-03-04 at 16:22 -0600, Tseng-Hui (Frank) Lin wrote:

> > Well, I don't know how you use icswx on P7+, but on Prism it's
> > definitely issued directly by userspace.
> > 
> OK. You've got a point. I wasn't aware of Prism. HFI device driver is
> currently the only icswx user on P7. Could you point me to more
> information about how Prism uses icswx from user space?

Let's ignore that for now, there's a whole infrastructure for it and it
hasn't been published yet, maybe later this year... We can look at
exposing the feature to userspace then.

> > You don't need to. You allocate and initialize the structure, and you
> > compare & swap the pointer. If somebody beat you, you trash your copy. 
> > 
> Is atomic_cmpxchg() the one to do the trick?

No, just cmpxchg()

> The lazy switching checks the shadow variable first before setting ACOP
> register. This saves mtspr() only if the new value is the same as
> current. If there are several coprocessors on the system, the ACOP
> register may have to be changed frequently. In that case, the lazy
> switching will not save time. In extreme case when the ACOP register
> needs to be changed every time, it actually slows down the execution by
> the additional shadow variable checking.

By how much ? Is it even measurable ?

Cheers,
Ben.