RE: [PATCH 0/2] clocksource/Hyper-V: Add Hyper-V specific sched clock function
From: Vitaly Kuznetsov
Date: Wed Aug 21 2019 - 03:15:31 EST
Michael Kelley <mikelley@xxxxxxxxxxxxx> writes:
> From: Vitaly Kuznetsov <vkuznets@xxxxxxxxxx> Sent: Tuesday, August 13, 2019 1:34 AM
>>
>> Michael Kelley <mikelley@xxxxxxxxxxxxx> writes:
>>
>> > From: Tianyu Lan <lantianyu1986@xxxxxxxxx> Sent: Tuesday, July 30, 2019 6:41 AM
>> >>
>> >> On Mon, Jul 29, 2019 at 8:13 PM Vitaly Kuznetsov <vkuznets@xxxxxxxxxx> wrote:
>> >> >
>> >> > Peter Zijlstra <peterz@xxxxxxxxxxxxx> writes:
>> >> >
>> >> > > On Mon, Jul 29, 2019 at 12:59:26PM +0200, Vitaly Kuznetsov wrote:
>> >> > >> lantianyu1986@xxxxxxxxx writes:
>> >> > >>
>> >> > >> > From: Tianyu Lan <Tianyu.Lan@xxxxxxxxxxxxx>
>> >> > >> >
>> >> > >> > Hyper-V guests use the default native_sched_clock() in pv_ops.time.sched_clock
>> >> > >> > on x86. But native_sched_clock() directly uses the raw TSC value, which
>> >> > >> > can be discontinuous in a Hyper-V VM. Add the generic hv_setup_sched_clock()
>> >> > >> > to set the sched clock function appropriately. On x86, this sets
>> >> > >> > pv_ops.time.sched_clock to read the Hyper-V reference TSC value that is
>> >> > >> > scaled and adjusted to be continuous.
>> >> > >>
>> >> > >> Hypervisor can, in theory, disable TSC page and then we're forced to use
>> >> > >> MSR-based clocksource but using it as sched_clock() can be very slow,
>> >> > >> I'm afraid.
>> >> > >>
>> >> > >> On the other hand, what we have now is probably worse: TSC can,
>> >> > >> actually, jump backwards (e.g. on migration) and we're breaking the
>> >> > >> requirements for sched_clock().
>> >> > >
>> >> > > That (obviously) also breaks the requirements for using TSC as
>> >> > > clocksource.
>> >> > >
>> >> > > IOW, it breaks the entire purpose of having TSC in the first place.
>> >> >
>> >> > Currently, we mark raw TSC as unstable when running on Hyper-V (see
>> >> > 88c9281a9fba6), 'TSC page' (which is TSC * scale + offset) is being used
>> >> > instead. The problem is that 'TSC page' can be disabled by the
>> >> > hypervisor and in that case the only remaining clocksource is MSR-based
>> >> > (slow).
>> >> >
>> >>
>> >> Yes, that will be slow if Hyper-V doesn't expose hv tsc page and
>> >> kernel uses MSR based
>> >> clocksource. Each MSR read will trigger one VM-EXIT. This also happens on other
>> >> hypervisors (e,g, KVM doesn't expose KVM clock). Hypervisor should
>> >> take this into
>> >> account and determine which clocksource should be exposed or not.
>> >>
>> >
>> > We've confirmed with the Hyper-V team that the TSC page is always available
>> > on Hyper-V 2016 and later, and on Hyper-V 2012 R2 when the physical
>> > hardware presents an InvariantTSC.
>>
>> Currently we check that TSC page is valid on every read and it seems
>> this is redundant, right? It is either available on boot or not. I can
>> only imagine migrating a VM to a non-InvariantTSC host when Hyper-V will
>> likely disable the page (and we can get reenlightenment notification
>> then).
>
> I think Hyper-V can have brief intervals when the TSC page is not valid, so
> the code checks for the "sequence" value being zero. Otherwise, yes, it
> should always be there or not be there. Is there some other validity
> check on every read that you are thinking of?
>
No, it's this one. In case these 'invalidity periods' are real there's
nothing to improve in the current code.
>>
>> > But the Linux Kconfig's are set up so
>> > the TSC page is not used for 32-bit guests -- all clock reads are synthetic MSR
>> > reads. For 32-bit, this set of changes will add more overhead because the
>> > sched clock reads will now be MSR reads.
>> >
>> > I would be inclined to fix the problem, even with the perf hit on 32-bit Linux.
>> > I donât have any data on 32-bit Linux being used in a Hyper-V guest, but it's not
>> > supported in Azure so usage is pretty small. The alternative would be to continue
>> > to use the raw TSC value on 32-bit, even with the risk of a discontinuity in case of
>> > live migration or similar scenarios.
>>
>> The issue needs fixing, I agree, however using MSR based clocksource as
>> sched clock may give us too big of a performance hit (not sure who cares
>> about 32 bit guest performance nowadays but still). What stops us from
>> enabling TSC page for 32 bit guests if it is available?
>
> I talked to KY Srinivasan for any history about TSC page on 32-bit. He said
> there was no technical reason not to implement it, but our focus was always
> 64-bit Linux, so the 32-bit was much less important. Also, on 32-bit Linux,
> the required 64x64 multiply and shift is more complex and takes more
> more cycles (compare 32-bit implementation of mul_u64_u64_shr vs.
> the 64-bit implementation), so the win over a MSR read is less. I
> don't know of any actual measurements being made to compare vs.
> MSR read.
VMExit is 1000 CPU cycles or so, I would guess that TSC page
calculations are better. Let me try to build 32bit kernel and do some
quick measurements.
--
Vitaly