linux-kernel - Re: [RFC PATCH] x86, kvm: use kvmclock to compute TSC deadline value

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite for Android: free password hash cracker in your pocket

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Date:	Wed, 6 Jul 2016 15:03:01 +0200
From:	Paolo Bonzini <pbonzini@...hat.com>
To:	Andy Lutomirski <luto@...capital.net>
Cc:	Eduardo Habkost <ehabkost@...hat.com>,
	Radim Krcmar <rkrcmar@...hat.com>,
	Peter Hornyack <peterhornyack@...gle.com>,
	David Matlack <dmatlack@...gle.com>,
	"linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
	kvm list <kvm@...r.kernel.org>
Subject: Re: [RFC PATCH] x86, kvm: use kvmclock to compute TSC deadline value



On 06/07/2016 15:01, Andy Lutomirski wrote:
> On Jul 5, 2016 1:36 PM, "Paolo Bonzini" <pbonzini@...hat.com> wrote:
>>
>> Bad things happen if a guest using the TSC deadline timer is migrated.
>> The guest doesn't re-calibrate the TSC after migration, and the
>> TSC frequency can and will change unless your processor supports TSC
>> scaling (on Intel this is only Skylake) or your data center is perfectly
>> homogeneous.
>>
>> The solution in this patch is to skip tsc_khz, and instead derive the
>> frequency from kvmclock's (mult, shift) pair.  Because kvmclock
>> parameters convert from tsc to nanoseconds, this needs a division
>> but that's the least of our problems when the TSC_DEADLINE_MSR write
>> costs 2000 clock cycles.  Luckily tsc_khz is really used by very little
>> outside the tsc clocksource (which kvmclock replaces) and the TSC
>> deadline timer.
> 
> I'm wondering if it would be more straightforward to blacklist the
> APIC timer entirely (on KVM guests that have the TSC deadline
> capability) and to instead offer a paravirt clockevent that still uses
> the deadline timer under the hood.  After all, these systems should be
> mostly quirk-free, so I imagine that the whole implementation would
> only be a few lines of code.

Very good idea!

Paolo