Re: [4.2, Regression] Queued spinlocks cause major XFS performance regression

[Davidlohr Bueso <dave@stgolabs.net>]

On Mon, 07 Sep 2015, Linus Torvalds wrote:

>On Sun, Sep 6, 2015 at 11:57 PM, Peter Zijlstra <peterz@infradead.org> wrote:
>>
>> Just to continue the argument for arguments sake, the function is named
>> 'virt' (not paravirt) and tests the HYPERVISOR CPUID bit. How is that
>> not appropriately named?
>
>Well, I think right now one issue is that you can't avoid it, even
>when you want pure "raw hardware" spinlocks.
>
>I really think it should at the very least be inside CONFIG_PARAVIRT.

Yeah, I think we all agree here.

>Because it *is* about helping the hypervisor, so really is about
>paravirtualization.

Yes, this is how I interpret it as well.

CONFIG_PARAVIRT seems like a suitable place as while it is known to induce
in overhead for baremetal and distros tend to enable it by default - mainly
for mem pvops (ie: in page fault paths), and having this function doesn't
add any complexity or add make things much different than they already are.

While it is true that CONFIG_HYPERVISOR_GUEST is a pre-req for anything pv,
technically I think it's not as good a fit as CONFIG_PARAVIRT because the
former is really about describing the hypervisor, and that's not what we're
doing here.


Thanks,
Davidlohr