Re: [PATCH 1/2] VMX: fix VMCS race on context-switch paths

["Jan Beulich" <JBeulich@suse.com>]

>>> On 16.02.17 at 09:29, <sergey.dyasli@citrix.com> wrote:
> On Wed, 2017-02-15 at 08:15 -0700, Jan Beulich wrote:
>> > > > On 15.02.17 at 15:55, <sergey.dyasli@citrix.com> wrote:
>> > Is it worth giving your patch another try with removing ctxt_switch_same()
>> > since we figured out that vmx_do_resume() will reload vmcs either way?
>> 
>> Yes, but that's the cosmetic part, whereras ...
>> 
>> > And I will also update vmx_vmcs_reload() from your last email.
>> 
>> ... this looks to be the actual bug fix. If you agree with my
>> reasoning of removing the loop altogether, you may want to go
>> with that version instead of adding the conditional.
> 
> After extensive night testing, it can be safe to assume that below
> patch fixes the PML issue. I agree about removing the spinning since
> vmx_vmcs_enter/exit are synchronized with the scheduler by schedule_lock.
> But it costs nothing to check so I added a debug message to the loop.
> Needless to say, it was never printed.

Thanks, that's good to know. I'll remove the loop in v2.

> My patch for vmx_vmcs_exit() is obviously a half measure because it
> doesn't protect against VMCS clearing by an external IPI when current
> is idle. I'm not sure such situation is possible but there is nothing
> that prevents it.
> 
> This clearly makes your approach superior and I think you need to
> submit v2 for proper review.

Will do.

Jan


_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
https://lists.xen.org/xen-devel