linux-kernel - Re: [PATCH] KVM: x86/xen: improve accuracy of Xen timers

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: Windows password security audit tool. GUI, reports in PDF.

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

Message-ID: <908ad29b56fbe5e5f4d99e477e96fbd9d03ba4c6.camel@infradead.org>
Date:   Tue, 31 Oct 2023 16:34:46 +0000
From:   David Woodhouse <dwmw2@...radead.org>
To:     paul@....org, Paul Durrant <xadimgnik@...il.com>,
        kvm@...r.kernel.org, linux-kernel@...r.kernel.org
Cc:     Sean Christopherson <seanjc@...gle.com>,
        Paolo Bonzini <pbonzini@...hat.com>,
        Thomas Gleixner <tglx@...utronix.de>,
        Ingo Molnar <mingo@...hat.com>, Borislav Petkov <bp@...en8.de>,
        Dave Hansen <dave.hansen@...ux.intel.com>, x86@...nel.org,
        "H. Peter Anvin" <hpa@...or.com>
Subject: Re: [PATCH] KVM: x86/xen: improve accuracy of Xen timers

On Tue, 2023-10-31 at 12:22 +0000, Paul Durrant wrote:
> 
> > 
> > As I said, this patch stands even *after* we fix kvmclock, because
> > it handles the timer delta calculation from an single TSC read.
> > 
> > But overengineering a timer reset on KVM_REQ_CLOCK_UPDATE would not.
> 
> I'm not sure what you intend to do to kvmlock, so not sure whether we'll 
> still need the __pvclock_read_cycles(&vcpu->arch.hv_clock, guest_tsc)
> but this patch (with the extra check on validity of hv_clock) does fix 
> the drift so...
> 
> Reviewed-by: Paul Durrant <paul@....org>

Ta. And no, I'm not quite sure what I'm going to do with kvmclock for
the general case yet. The more I look at it, the more I realise how
broken it is.

Last week I thought it was just about the way KVM 'jumps' the kvmclock
and yanks it back to the host's CLOCK_MONOTONIC_RAW, but I thought KVM
at *least* managed to do it right between those times. But no, this
patch is addressing the fact that even *without* those clock jumps, KVM
doesn't manage to calculate the guest clock the same way that it tells
the guest to... and thus doesn't get the same results :)

I think it involves get_kvmclock_ns() using the frequency given in the
KVM-wide (not per-vCPU) KVM_SET_TSC_KHZ ioctl, and scaling via that to
get the guest clock. That should match, without having to have a
specific vCPU's hv_clock to play with. And maybe we can also have a
get_kvmclock_ns_at() which takes a host TSC value, and the timer code
from this patch can use that instead of using __pvclock_read_cycles()
directly.

That's probably the easy part. Fixing the 'resync' to
CLOCK_MONOTONIC_RAW, while keeping things working for the now-
considered-pathological !CONSTANT_TSC case, will be slightly more fun.
As well as suspend etc. in the CONSTANT_TSC case, of course.

And replacing that stupid KVM_CLOCK_REALTIME with something that uses
CLOCK_TAI. Or maybe just making it export the tai_offset at the same
moment?

Download attachment "smime.p7s" of type "application/pkcs7-signature" (5965 bytes)