RE: [PATCH] x86/hyperv: Set X86_FEATURE_TSC_RELIABLE unconditionally

[Michael Kelley]

From: Stanislav Kinsburskii <skinsburskii@xxxxxxxxxxxxxxxxxxx> Sent: Wednesday, December 11, 2024 10:01 AM
> 
> On Tue, Nov 26, 2024 at 06:11:10AM +0000, Michael Kelley wrote:
> > From: Stanislav Kinsburskii <skinsburskii@xxxxxxxxxxxxxxxxxxx> Sent: Monday, November 25, 2024 2:25 PM
> > >
> > > On Fri, Nov 22, 2024 at 06:33:12PM +0000, Michael Kelley wrote:
> > > > From: Stanislav Kinsburskii <skinsburskii@xxxxxxxxxxxxxxxxxxx> Sent: Tuesday, November 19, 2024 4:51 PM
> > > > >
> > > > > On Tue, Nov 12, 2024 at 07:48:06PM +0000, Michael Kelley wrote:
> > > > > > From: Stanislav Kinsburskii <skinsburskii@xxxxxxxxxxxxxxxxxxx> Sent: Tuesday,
> > > > > November 12, 2024 10:18 AM
> > > > > > >
> > > > > > > Enable X86_FEATURE_TSC_RELIABLE by default as X86_FEATURE_TSC_RELIABLE is
> > > > > > > independent from invariant TSC and should have never been gated by the
> > > > > > > HV_ACCESS_TSC_INVARIANT privilege.
> > > > > >
> > > > > > I think originally X86_FEATURE_TSC_RELIABLE was gated by the Hyper-V
> > > > > > TSC Invariant feature because otherwise VM live migration may cause
> > > > > > the TSC value reported by the RDTSC/RDTSCP instruction in the guest
> > > > > > to abruptly change frequency and value. In such cases, the TSC isn't
> > > > > > useable by the kernel or user space.
> > > > > >
> > > > > > Enabling the Hyper-V TSC Invariant feature fixes that by using the
> > > > > > hardware scaling available in more recent processors to automatically
> > > > > > fixup the TSC value returned by RDTSC/RDTSCP in the guest.
> > > > > >
> > > > > > Is there a practical problem that is fixed by always enabling
> > > > > > X86_FEATURE_TSC_RELIABLE?
> > > > > >
> > > > >
> > > > > The particular problem is that HV_ACCESS_TSC_INVARIANT is not set for the
> > > > > nested root, which in turn leads to keeping tsc clocksource watchdog
> > > > > thread and TSC sycn check timer around.
> > > >
> > > > I have trouble keeping all the different TSC "features" conceptually
> > > > separate. :-( The TSC frequency not changing (and the value not
> > > > abruptly jumping?) should already be represented by
> > > > X86_FEATURE_TSC_CONSTANT.  In the kernel, X86_FEATURE_TSC_RELIABLE
> > > > effectively only controls whether the TSC clocksource watchdog is
> > > > enabled, and in spite of the live migration foibles, I don't see a need
> > > > for that watchdog in a Hyper-V VM. So maybe it's OK to always set
> > > > X86_FEATURE_TSC_RELIABLE in a Hyper-V VM, as you have
> > > > proposed.
> > > >
> > > > The "tsc_reliable" flag is also exposed to user space as part of the
> > > > /proc/cpuinfo "flags" output, so theoretically some user space
> > > > program could change behavior based on that flag. But that seems
> > > > a bit far-fetched. I know there are user space programs that check
> > > > the CPUID INVARIANT_TSC flag to know whether they can use
> > > > the raw RDTSC instruction output to do start/stop timing. The
> > > > Hyper-V TSC Invariant feature makes that work correctly, even
> > > > across live migrations.
> > > >
> > >
> > > It sounds to me that if X86_FEATURE_TSC_CONSTANT is available
> > > on Hyper-V, then we can set X86_FEATURE_TSC_RELIABLE.
> > > Is it what you are saying?
> > >
> >
> > No. Sorry I wasn't clear. X86_FEATURE_TSC_CONSTANT will
> > be set only when the Hyper-V TSC Invariant feature is enabled, so
> > tying X86_FEATURE_TSC_RELIABLE to that is what happens now.
> >
> > What I'm suggesting is to take your patch "as is". In other words,
> > always enable X86_FEATURE_TSC_RELIABLE. From what I can tell,
> > TSC_RELIABLE is only used to disable the TSC watchdog. Since I
> > can't see a use for the TSC watchdog in a VM, always setting
> > TSC_RELIABLE probably makes sense. TSC_RELIABLE doesn't
> > say anything about whether the TSC frequency might change, such
> > as across a VM live migration. TSC_CONSTANT is what tells you that
> > the frequency won't change.
> >
> > My caveat is that I don't know the history of TSC_RELIABLE. I
> > don't see any documentation on the details of what it is supposed
> > to convey, especially in a VM. Maybe someone on the "To:" list
> > who knows for sure can confirm what I'm thinking.
> >
> > Michael
> 
> We had a long ionternal discussion with hypervisor folks and it looks
> like we will propose a more robust solution to go forward.
> The hypervisor will provide an additional CPUID bit, which guarantees
> TSC reliability (including across live migration).

Interesting. Is there a text description of what this new CPUID
bit means? I would be curious, and in particular how it is
different from the various existing CPUID bits describing TSC
behavior.

Michael