2017-11-09 17:40:02

by Radim Krčmář

[permalink] [raw]
Subject: Re: [PATCHv3 1/1] locking/qspinlock/x86: Avoid test-and-set when PV_DEDICATED is set

2017-11-09 18:12+0100, Peter Zijlstra:
> On Thu, Nov 09, 2017 at 05:45:23PM +0100, Radim Krcmar wrote:
> > 2017-11-09 17:17+0100, Peter Zijlstra:
> > > On Thu, Nov 09, 2017 at 05:05:36PM +0100, Radim Krcmar wrote:
> > > > 2017-11-09 10:53-0500, Pankaj Gupta:
> > > > > 2] PV TLB should also behave as per option PV_DEDICATED for better performance.
> > > >
> > > > Right,
> > >
> > > Shouldn't KVM do flush_tlb_other() in any case? Not sure how
> > > PV_DEDICATED can help with that.
> >
> > It will, the suggestion was based on recent extension of the
> > flush_tlb_others implementaion, https://lkml.org/lkml/2017/11/8/1146.
> >
> > PV_TLB_FLUSH allows a guest to set a flush bit instead of sending flush
> > IPI if the target VCPU is not running. This would be a waste of time
> > with PV_DEDICATED as all VCPUs are expected to always running.
> >
> > With PV_DEDICATED, the guest should keep using native_flush_tlb_others.
>
> Is saving that for_each_cpu() really worth the effort compared to the
> cost of actually doing the IPIs and CR3 write?

It is one line for a few percent (hopefully better for AMD with AVIC).

Still, keeping the decision completely on userspace would be cleaner.

> Also, you should not put cpumask_t on stack, that's 'broken'.

Good catch, thanks.

From 1583610875375035453@xxx Thu Nov 09 17:36:59 +0000 2017
X-GM-THRID: 1583352031107436177
X-Gmail-Labels: Inbox,Category Forums,HistoricalUnread