Re: [PATCH v4 0/3] qspinlock: Introducing a 4-byte queue spinlock

[Linus Torvalds <torvalds@linux-foundation.org>]

On Thu, Feb 20, 2014 at 9:37 AM, Waiman Long <waiman.long@hp.com> wrote:
>
> Yes, that is true for paravirt, but I was talking about virtualization in
> general.

For non-paravirtualized environments, we should basically optimize for
the actual hardware, and assume that the full virtualization is good
enough that we don't care.

Obviously, sometimes we might take "this is particularly expensive to
virtualize" into account, and if it doesn't hurt hardware, try to help
the virtual environment along. So you might have cases where certain
hardware (typically interrupt controllers etc) should be accessed
certain particular ways, because those not only work on real hardware,
but behave well in virtualized environments too.

But for spinlocks, just make the raw hardware work well when not
*actively* paravirtualized. If some hypervisor deals badly with it,
it's a hypervisor or management problem.

            Linus