Matt attempted to add CLOCK_TAI support to the VDSO clock_gettime()
implementation, which extended the clockid switch case and added yet
another slightly different copy of the same code.
Especially the extended switch case is problematic as the compiler tends to
generate a jump table which then requires to use retpolines. If jump tables
are disabled it adds yet another conditional to the existing maze.
This series takes a different approach by consolidating the almost
identical functions into one implementation for high resolution clocks and
one for the coarse grained clock ids by storing the base data for each
clock id in an array which is indexed by the clock id.
This completely eliminates the switch case and allows further
simplifications of the code base, which at the end all together gain a few
cycles performance or at least stay on par with todays code. The resulting
performance depends heavily on the micro architecture and the compiler.
Thanks,
tglx
8<-------------------
arch/x86/Kconfig | 1
arch/x86/entry/vdso/vclock_gettime.c | 199 ++++++++------------------------
arch/x86/entry/vsyscall/vsyscall_gtod.c | 55 ++++----
arch/x86/include/asm/vgtod.h | 46 ++++---
arch/x86/kernel/time.c | 22 +++
include/linux/clocksource.h | 5
kernel/time/Kconfig | 4
kernel/time/clocksource.c | 2
8 files changed, 144 insertions(+), 190 deletions(-)
On 09/14/2018 02:50 PM, Thomas Gleixner wrote:
> Matt attempted to add CLOCK_TAI support to the VDSO clock_gettime()
> implementation, which extended the clockid switch case and added yet
> another slightly different copy of the same code.
>
> Especially the extended switch case is problematic as the compiler tends to
> generate a jump table which then requires to use retpolines.
Does vDSO code really have to use retpolines? It's in userspace, after all.
Thanks,
Florian
On 09/14/2018 03:05 PM, Thomas Gleixner wrote:
> On Fri, 14 Sep 2018, Florian Weimer wrote:
>
>> On 09/14/2018 02:50 PM, Thomas Gleixner wrote:
>>> Matt attempted to add CLOCK_TAI support to the VDSO clock_gettime()
>>> implementation, which extended the clockid switch case and added yet
>>> another slightly different copy of the same code.
>>>
>>> Especially the extended switch case is problematic as the compiler tends to
>>> generate a jump table which then requires to use retpolines.
>>
>> Does vDSO code really have to use retpolines? It's in userspace, after all.
>
> Unless you have IBRS/STIPB enabled, you need user space ratpoutine as well.
I don't think this is a consensus position, and it obviously depends on
the (sub)architecture.
Thanks,
Florian
On Fri, 14 Sep 2018, Florian Weimer wrote:
> On 09/14/2018 02:50 PM, Thomas Gleixner wrote:
> > Matt attempted to add CLOCK_TAI support to the VDSO clock_gettime()
> > implementation, which extended the clockid switch case and added yet
> > another slightly different copy of the same code.
> >
> > Especially the extended switch case is problematic as the compiler tends to
> > generate a jump table which then requires to use retpolines.
>
> Does vDSO code really have to use retpolines? It's in userspace, after all.
Unless you have IBRS/STIPB enabled, you need user space ratpoutine as well.
Thanks,
tglx
On Fri, Sep 14, 2018 at 02:56:46PM +0200, Florian Weimer wrote:
> On 09/14/2018 02:50 PM, Thomas Gleixner wrote:
> > Matt attempted to add CLOCK_TAI support to the VDSO clock_gettime()
> > implementation, which extended the clockid switch case and added yet
> > another slightly different copy of the same code.
> >
> > Especially the extended switch case is problematic as the compiler tends to
> > generate a jump table which then requires to use retpolines.
>
> Does vDSO code really have to use retpolines? It's in userspace, after all.
Userspace is equally susceptible to spectre-v2. Ideally we'd recompile
world with retpoline, but given the amount of inline asm in say things
like openssl and similar projects, validating that there are indeed no
indirect calls/jumps left is nontrivial.
There are currently pending patches to otherwise address user-user
spectre-v2 attacks.
On Fri, 14 Sep 2018, Florian Weimer wrote:
> On 09/14/2018 03:05 PM, Thomas Gleixner wrote:
> > On Fri, 14 Sep 2018, Florian Weimer wrote:
> >
> > > On 09/14/2018 02:50 PM, Thomas Gleixner wrote:
> > > > Matt attempted to add CLOCK_TAI support to the VDSO clock_gettime()
> > > > implementation, which extended the clockid switch case and added yet
> > > > another slightly different copy of the same code.
> > > >
> > > > Especially the extended switch case is problematic as the compiler tends
> > > > to
> > > > generate a jump table which then requires to use retpolines.
> > >
> > > Does vDSO code really have to use retpolines? It's in userspace, after
> > > all.
> >
> > Unless you have IBRS/STIPB enabled, you need user space ratpoutine as well.
>
> I don't think this is a consensus position, and it obviously depends on the
> (sub)architecture.
It does, but we are not building kernels for specific micro architectures
nor do distros AFAIK.
But that aside, even with jump tables the generated code (ratpoutine
disabled) is not any better than with the current approach. It's even worse
and needs the same optimizations at the end and then the main difference is
that you have 3 copies of one function and 2 of the other. Not a win at
all.
Thanks,
tglx
On Fri, Sep 14, 2018 at 2:52 PM Thomas Gleixner <[email protected]> wrote:
>
> Matt attempted to add CLOCK_TAI support to the VDSO clock_gettime()
> implementation, which extended the clockid switch case and added yet
> another slightly different copy of the same code.
>
> Especially the extended switch case is problematic as the compiler tends to
> generate a jump table which then requires to use retpolines. If jump tables
> are disabled it adds yet another conditional to the existing maze.
>
> This series takes a different approach by consolidating the almost
> identical functions into one implementation for high resolution clocks and
> one for the coarse grained clock ids by storing the base data for each
> clock id in an array which is indexed by the clock id.
>
> This completely eliminates the switch case and allows further
> simplifications of the code base, which at the end all together gain a few
> cycles performance or at least stay on par with todays code. The resulting
> performance depends heavily on the micro architecture and the compiler.
No objections from my side, just a few general remarks: Deepa
and I have discussed the vdso in the past for 2038. I have started
a patch that I'll have to redo on top of yours, but that is fine, your
cleanup is only going to help here.
More generally speaking, Deepa said that she would like to see
some generalization on the vdso before adding the time64_t
based variants. Your series probably makes x86 diverge more
from the others, which makes it harder to consolidate them again,
but we might just as well use your new implementation to base the
generic one on, and just move the other ones over to that.
A couple of architectures (s390, ia64, riscv, powerpc, arm64)
implement the vdso as assembler code at the moment, so they
won't be as easy to consolidate (other than outright replacing all
the code).
The other five:
arch/x86/entry/vdso/vclock_gettime.c
arch/sparc/vdso/vclock_gettime.c
arch/nds32/kernel/vdso/gettimeofday.c
arch/mips/vdso/gettimeofday.c
arch/arm/vdso/vgettimeofday.c
are basically all minor variations of the same code base and could be
consolidated to some degree.
Any suggestions here? Should we plan to do that consolitdation based on
your new version, or just add clock_gettime64 in arm32 and x86-32, and then
be done with it? The other ones will obviously still be fast for 32-bit time_t
and will have a working non-vdso sys_clock_getttime64().
I also wonder about clock_getres(): half the architectures seem to implement
it in vdso, but notably arm32 and x86 don't, and I had not expected it to be
performance critical given that the result is easily cached in user space.
Arnd
On Fri, 14 Sep 2018, Arnd Bergmann wrote:
> On Fri, Sep 14, 2018 at 2:52 PM Thomas Gleixner <[email protected]> wrote:
> A couple of architectures (s390, ia64, riscv, powerpc, arm64)
> implement the vdso as assembler code at the moment, so they
> won't be as easy to consolidate (other than outright replacing all
> the code).
>
> The other five:
> arch/x86/entry/vdso/vclock_gettime.c
> arch/sparc/vdso/vclock_gettime.c
> arch/nds32/kernel/vdso/gettimeofday.c
> arch/mips/vdso/gettimeofday.c
> arch/arm/vdso/vgettimeofday.c
>
> are basically all minor variations of the same code base and could be
> consolidated to some degree.
> Any suggestions here? Should we plan to do that consolitdation based on
> your new version, or just add clock_gettime64 in arm32 and x86-32, and then
> be done with it? The other ones will obviously still be fast for 32-bit time_t
> and will have a working non-vdso sys_clock_getttime64().
In principle consolidating all those implementations should be possible to
some extent and probably worthwhile. What's arch specific are the actual
accessors to the hardware clocks.
> I also wonder about clock_getres(): half the architectures seem to implement
> it in vdso, but notably arm32 and x86 don't, and I had not expected it to be
> performance critical given that the result is easily cached in user space.
getres() is not really performance critical, but adding it does not create
a huge problem either.
Thanks,
tglx
On Mon, Sep 17, 2018 at 3:00 PM Thomas Gleixner <[email protected]> wrote:
>
> On Fri, 14 Sep 2018, Arnd Bergmann wrote:
> > On Fri, Sep 14, 2018 at 2:52 PM Thomas Gleixner <[email protected]> wrote:
> > A couple of architectures (s390, ia64, riscv, powerpc, arm64)
> > implement the vdso as assembler code at the moment, so they
> > won't be as easy to consolidate (other than outright replacing all
> > the code).
> >
> > The other five:
> > arch/x86/entry/vdso/vclock_gettime.c
> > arch/sparc/vdso/vclock_gettime.c
> > arch/nds32/kernel/vdso/gettimeofday.c
> > arch/mips/vdso/gettimeofday.c
> > arch/arm/vdso/vgettimeofday.c
> >
> > are basically all minor variations of the same code base and could be
> > consolidated to some degree.
> > Any suggestions here? Should we plan to do that consolitdation based on
> > your new version, or just add clock_gettime64 in arm32 and x86-32, and then
> > be done with it? The other ones will obviously still be fast for 32-bit time_t
> > and will have a working non-vdso sys_clock_getttime64().
>
> In principle consolidating all those implementations should be possible to
> some extent and probably worthwhile. What's arch specific are the actual
> accessors to the hardware clocks.
Ok.
> > I also wonder about clock_getres(): half the architectures seem to implement
> > it in vdso, but notably arm32 and x86 don't, and I had not expected it to be
> > performance critical given that the result is easily cached in user space.
>
> getres() is not really performance critical, but adding it does not create
> a huge problem either.
Right, I'd just not add a getres_time64() to the vdso then.
Arnd
Hi Vitaly, Paolo, Radim, etc.,
On Fri, Sep 14, 2018 at 5:52 AM Thomas Gleixner <[email protected]> wrote:
>
> Matt attempted to add CLOCK_TAI support to the VDSO clock_gettime()
> implementation, which extended the clockid switch case and added yet
> another slightly different copy of the same code.
>
> Especially the extended switch case is problematic as the compiler tends to
> generate a jump table which then requires to use retpolines. If jump tables
> are disabled it adds yet another conditional to the existing maze.
>
> This series takes a different approach by consolidating the almost
> identical functions into one implementation for high resolution clocks and
> one for the coarse grained clock ids by storing the base data for each
> clock id in an array which is indexed by the clock id.
>
I was trying to understand more of the implications of this patch
series, and I was again reminded that there is an entire extra copy of
the vclock reading code in arch/x86/kvm/x86.c. And the purpose of
that code is very, very opaque.
Can one of you explain what the code is even doing? From a couple of
attempts to read through it, it's a whole bunch of
probably-extremely-buggy code that, drumroll please, tries to
atomically read the TSC value and the time. And decide whether the
result is "based on the TSC". And then synthesizes a TSC-to-ns
multiplier and shift, based on *something other than the actual
multiply and shift used*.
IOW, unless I'm totally misunderstanding it, the code digs into the
private arch clocksource data intended for the vDSO, uses a poorly
maintained copy of the vDSO code to read the time (instead of doing
the sane thing and using the kernel interfaces for this), and
propagates a totally made up copy to the guest. And gets it entirely
wrong when doing nested virt, since, unless there's some secret in
this maze, it doesn't acutlaly use the scaling factor from the host
when it tells the guest what to do.
I am really, seriously tempted to send a patch to simply delete all
this code. The correct way to do it is to hook
And I don't see how it's even possible to pass kvmclock correctly to
the L2 guest when L0 is hyperv. KVM could pass *hyperv's* clock, but
L1 isn't notified when the data structure changes, so how the heck is
it supposed to update the kvmclock structure?
--Andy
Andy Lutomirski <[email protected]> writes:
> Hi Vitaly, Paolo, Radim, etc.,
>
> On Fri, Sep 14, 2018 at 5:52 AM Thomas Gleixner <[email protected]> wrote:
>>
>> Matt attempted to add CLOCK_TAI support to the VDSO clock_gettime()
>> implementation, which extended the clockid switch case and added yet
>> another slightly different copy of the same code.
>>
>> Especially the extended switch case is problematic as the compiler tends to
>> generate a jump table which then requires to use retpolines. If jump tables
>> are disabled it adds yet another conditional to the existing maze.
>>
>> This series takes a different approach by consolidating the almost
>> identical functions into one implementation for high resolution clocks and
>> one for the coarse grained clock ids by storing the base data for each
>> clock id in an array which is indexed by the clock id.
>>
>
> I was trying to understand more of the implications of this patch
> series, and I was again reminded that there is an entire extra copy of
> the vclock reading code in arch/x86/kvm/x86.c. And the purpose of
> that code is very, very opaque.
>
> Can one of you explain what the code is even doing? From a couple of
> attempts to read through it, it's a whole bunch of
> probably-extremely-buggy code that, drumroll please, tries to
> atomically read the TSC value and the time. And decide whether the
> result is "based on the TSC". And then synthesizes a TSC-to-ns
> multiplier and shift, based on *something other than the actual
> multiply and shift used*.
>
> IOW, unless I'm totally misunderstanding it, the code digs into the
> private arch clocksource data intended for the vDSO, uses a poorly
> maintained copy of the vDSO code to read the time (instead of doing
> the sane thing and using the kernel interfaces for this), and
> propagates a totally made up copy to the guest. And gets it entirely
> wrong when doing nested virt, since, unless there's some secret in
> this maze, it doesn't acutlaly use the scaling factor from the host
> when it tells the guest what to do.
>
> I am really, seriously tempted to send a patch to simply delete all
> this code. The correct way to do it is to hook
"I have discovered a truly marvelous proof of this, which this margin is
too narrow to contain" :-)
There is a very long history of different (hardware) issues Marcelo was
fighting with and the current code is the survived Frankenstein. E.g. it
is very, very unclear what "catchup", "always catchup" and
masterclock-less mode in general are and if we still need them.
That said I'm all for simplification. I'm not sure if we still need to
care about buggy hardware though.
>
> And I don't see how it's even possible to pass kvmclock correctly to
> the L2 guest when L0 is hyperv. KVM could pass *hyperv's* clock, but
> L1 isn't notified when the data structure changes, so how the heck is
> it supposed to update the kvmclock structure?
Well, this kind of works in the the followin way:
L1's clocksource is 'tsc_page' which is, basically, a compliment to TSC:
two numbers provided by L0: offset and scale and KVM was tought to treat
this clocksource as a good one (see b0c39dc68e3b "x86/kvm: Pass stable
clocksource to guests when running nested on Hyper-V").
The notification you're talking about exists, it is called
Reenligntenment, see 0092e4346f49 "x86/kvm: Support Hyper-V
reenlightenment"). When TSC page changes (and this only happens when L1
is migrated to a different host with a different TSC frequency and TSC
scaling is not supported by the CPU) we receive an interrupt in L1 (at
this moment all TSC accesses are emulated which guarantees the
correctness of the readings), pause all L2 guests, update their kvmclock
structures with new data (we already know the new TSC frequency) and
then tell L0 that we're done and it can stop emulating TSC accesses.
(Nothing like this exists for KVM-on-KVM, by the way, when L1's
clocksource is 'kvmclock' L2s won't get a stable kvmclock clocksource.)
--
Vitaly
> On Oct 3, 2018, at 2:22 AM, Vitaly Kuznetsov <[email protected]> wrote:
>
> Andy Lutomirski <[email protected]> writes:
>
>> Hi Vitaly, Paolo, Radim, etc.,
>>
>>> On Fri, Sep 14, 2018 at 5:52 AM Thomas Gleixner <[email protected]> wrote:
>>>
>>> Matt attempted to add CLOCK_TAI support to the VDSO clock_gettime()
>>> implementation, which extended the clockid switch case and added yet
>>> another slightly different copy of the same code.
>>>
>>> Especially the extended switch case is problematic as the compiler tends to
>>> generate a jump table which then requires to use retpolines. If jump tables
>>> are disabled it adds yet another conditional to the existing maze.
>>>
>>> This series takes a different approach by consolidating the almost
>>> identical functions into one implementation for high resolution clocks and
>>> one for the coarse grained clock ids by storing the base data for each
>>> clock id in an array which is indexed by the clock id.
>>>
>>
>> I was trying to understand more of the implications of this patch
>> series, and I was again reminded that there is an entire extra copy of
>> the vclock reading code in arch/x86/kvm/x86.c. And the purpose of
>> that code is very, very opaque.
>>
>> Can one of you explain what the code is even doing? From a couple of
>> attempts to read through it, it's a whole bunch of
>> probably-extremely-buggy code that, drumroll please, tries to
>> atomically read the TSC value and the time. And decide whether the
>> result is "based on the TSC". And then synthesizes a TSC-to-ns
>> multiplier and shift, based on *something other than the actual
>> multiply and shift used*.
>>
>> IOW, unless I'm totally misunderstanding it, the code digs into the
>> private arch clocksource data intended for the vDSO, uses a poorly
>> maintained copy of the vDSO code to read the time (instead of doing
>> the sane thing and using the kernel interfaces for this), and
>> propagates a totally made up copy to the guest. And gets it entirely
>> wrong when doing nested virt, since, unless there's some secret in
>> this maze, it doesn't acutlaly use the scaling factor from the host
>> when it tells the guest what to do.
>>
>> I am really, seriously tempted to send a patch to simply delete all
>> this code. The correct way to do it is to hook
>
> "I have discovered a truly marvelous proof of this, which this margin is
> too narrow to contain" :-)
>
> There is a very long history of different (hardware) issues Marcelo was
> fighting with and the current code is the survived Frankenstein. E.g. it
> is very, very unclear what "catchup", "always catchup" and
> masterclock-less mode in general are and if we still need them.
>
> That said I'm all for simplification. I'm not sure if we still need to
> care about buggy hardware though.
>
>>
>> And I don't see how it's even possible to pass kvmclock correctly to
>> the L2 guest when L0 is hyperv. KVM could pass *hyperv's* clock, but
>> L1 isn't notified when the data structure changes, so how the heck is
>> it supposed to update the kvmclock structure?
>
> Well, this kind of works in the the followin way:
> L1's clocksource is 'tsc_page' which is, basically, a compliment to TSC:
> two numbers provided by L0: offset and scale and KVM was tought to treat
> this clocksource as a good one (see b0c39dc68e3b "x86/kvm: Pass stable
> clocksource to guests when running nested on Hyper-V").
>
> The notification you're talking about exists, it is called
> Reenligntenment, see 0092e4346f49 "x86/kvm: Support Hyper-V
> reenlightenment"). When TSC page changes (and this only happens when L1
> is migrated to a different host with a different TSC frequency and TSC
> scaling is not supported by the CPU) we receive an interrupt in L1 (at
> this moment all TSC accesses are emulated which guarantees the
> correctness of the readings), pause all L2 guests, update their kvmclock
> structures with new data (we already know the new TSC frequency) and
> then tell L0 that we're done and it can stop emulating TSC accesses.
That’s delightful! Does the emulation magic also work for L1 user mode? If so, couldn’t we drop the HyperV vclock entirely and just fold the adjustment into the core timekeeping data? (Preferably the actual core data, which would require core changes, but it could plausibly be done in arch code, too.)
>
> (Nothing like this exists for KVM-on-KVM, by the way, when L1's
> clocksource is 'kvmclock' L2s won't get a stable kvmclock clocksource.)
>
>
Andy Lutomirski <[email protected]> writes:
>> On Oct 3, 2018, at 2:22 AM, Vitaly Kuznetsov <[email protected]> wrote:
>>
>> Andy Lutomirski <[email protected]> writes:
>>
>>> Hi Vitaly, Paolo, Radim, etc.,
>>>
>> The notification you're talking about exists, it is called
>> Reenligntenment, see 0092e4346f49 "x86/kvm: Support Hyper-V
>> reenlightenment"). When TSC page changes (and this only happens when L1
>> is migrated to a different host with a different TSC frequency and TSC
>> scaling is not supported by the CPU) we receive an interrupt in L1 (at
>> this moment all TSC accesses are emulated which guarantees the
>> correctness of the readings), pause all L2 guests, update their kvmclock
>> structures with new data (we already know the new TSC frequency) and
>> then tell L0 that we're done and it can stop emulating TSC accesses.
>
> That’s delightful! Does the emulation magic also work for L1 user
> mode?
As far as I understand - yes, all rdtsc* calls will trap into L0.
> If so, couldn’t we drop the HyperV vclock entirely and just
> fold the adjustment into the core timekeeping data? (Preferably the
> actual core data, which would require core changes, but it could
> plausibly be done in arch code, too.)
Not all Hyper-V hosts support reenlightenment notifications (and, if I'm
not mistaken, you need to enable nesting for the VM to get the feature -
and most VMs don't have this) so I think we'll have to keep Hyper-V
vclock for the time being.
--
Vitaly
> On Oct 3, 2018, at 5:01 AM, Vitaly Kuznetsov <[email protected]> wrote:
>
> Andy Lutomirski <[email protected]> writes:
>
>>> On Oct 3, 2018, at 2:22 AM, Vitaly Kuznetsov <[email protected]> wrote:
>>>
>>> Andy Lutomirski <[email protected]> writes:
>>>
>>>> Hi Vitaly, Paolo, Radim, etc.,
>>>>
>>> The notification you're talking about exists, it is called
>>> Reenligntenment, see 0092e4346f49 "x86/kvm: Support Hyper-V
>>> reenlightenment"). When TSC page changes (and this only happens when L1
>>> is migrated to a different host with a different TSC frequency and TSC
>>> scaling is not supported by the CPU) we receive an interrupt in L1 (at
>>> this moment all TSC accesses are emulated which guarantees the
>>> correctness of the readings), pause all L2 guests, update their kvmclock
>>> structures with new data (we already know the new TSC frequency) and
>>> then tell L0 that we're done and it can stop emulating TSC accesses.
>>
>> That’s delightful! Does the emulation magic also work for L1 user
>> mode?
>
> As far as I understand - yes, all rdtsc* calls will trap into L0.
>
>> If so, couldn’t we drop the HyperV vclock entirely and just
>> fold the adjustment into the core timekeeping data? (Preferably the
>> actual core data, which would require core changes, but it could
>> plausibly be done in arch code, too.)
>
> Not all Hyper-V hosts support reenlightenment notifications (and, if I'm
> not mistaken, you need to enable nesting for the VM to get the feature -
> and most VMs don't have this) so I think we'll have to keep Hyper-V
> vclock for the time being.
>
>
But this does suggest that the correct way to pass a clock through to an L2 guest where L0 is HV is to make L1 use the “tsc” clock and L2 use kvmclock (or something newer and better). This would require adding support for atomic frequency changes all the way through the timekeeping and arch code.
John, tglx, would that be okay or crazy?
On Wed, 3 Oct 2018, Andy Lutomirski wrote:
> > On Oct 3, 2018, at 5:01 AM, Vitaly Kuznetsov <[email protected]> wrote:
> > Not all Hyper-V hosts support reenlightenment notifications (and, if I'm
> > not mistaken, you need to enable nesting for the VM to get the feature -
> > and most VMs don't have this) so I think we'll have to keep Hyper-V
> > vclock for the time being.
> >
> But this does suggest that the correct way to pass a clock through to an
> L2 guest where L0 is HV is to make L1 use the “tsc” clock and L2 use
> kvmclock (or something newer and better). This would require adding
> support for atomic frequency changes all the way through the timekeeping
> and arch code.
>
> John, tglx, would that be okay or crazy?
Not sure what you mean. I think I lost you somewhere on the way.
Thanks,
tglx
On Wed, Oct 3, 2018 at 8:10 AM Thomas Gleixner <[email protected]> wrote:
>
> On Wed, 3 Oct 2018, Andy Lutomirski wrote:
> > > On Oct 3, 2018, at 5:01 AM, Vitaly Kuznetsov <[email protected]> wrote:
> > > Not all Hyper-V hosts support reenlightenment notifications (and, if I'm
> > > not mistaken, you need to enable nesting for the VM to get the feature -
> > > and most VMs don't have this) so I think we'll have to keep Hyper-V
> > > vclock for the time being.
> > >
> > But this does suggest that the correct way to pass a clock through to an
> > L2 guest where L0 is HV is to make L1 use the “tsc” clock and L2 use
> > kvmclock (or something newer and better). This would require adding
> > support for atomic frequency changes all the way through the timekeeping
> > and arch code.
> >
> > John, tglx, would that be okay or crazy?
>
> Not sure what you mean. I think I lost you somewhere on the way.
>
What I mean is: currently we have a clocksource called
""hyperv_clocksource_tsc_page". Reading it does:
static u64 read_hv_clock_tsc(struct clocksource *arg)
{
u64 current_tick = hv_read_tsc_page(tsc_pg);
if (current_tick == U64_MAX)
rdmsrl(HV_X64_MSR_TIME_REF_COUNT, current_tick);
return current_tick;
}
From Vitaly's email, it sounds like, on most (all?) hyperv systems
with nesting enabled, this clock is better behaved than it appears.
It sounds like the read behavior is that current_tick will never be
U64_MAX -- instead, the clock always works and, more importantly, the
actual scaling factor and offset only change observably on *guest*
request.
So why don't we we improve the actual "tsc" clocksource to understand
this? ISTM the best model would be where the
__clocksource_update_freq_xyz() mechanism gets called so we can use it
like this:
clocksource_begin_update();
clocksource_update_mult_shift();
tell_hv_that_we_reenlightened();
clocksource_end_update();
Where clocksource_begin_update() bumps the seqcount for the vDSO and
takes all the locks, clocksource_update_mult_shift() updates
everything, and clocksource_end_update() makes the updated parameters
usable.
(AFAICT there are currently no clocksources at all in the entire
kernel that update their frequency on the fly using
__clocksource_update_xyz(). Unless I'm missing something, the x86 tsc
cpufreq hooks don't call into the core timekeeping at all, so I'm
assuming that the tsc clocksource is just unusable as a clocksource on
systems that change its frequency.)
Or we could keep the hyperv_clocksource_tsc_page clocksource but make
it use VCLOCK_TSC and a similar update mechanism.
I don't personally want to do this, because the timekeeping code is
subtle and I'm unfamiliar with it. And I don't have *that* many spare
cycles :)
On Tue, Oct 02, 2018 at 10:15:49PM -0700, Andy Lutomirski wrote:
> Hi Vitaly, Paolo, Radim, etc.,
>
> On Fri, Sep 14, 2018 at 5:52 AM Thomas Gleixner <[email protected]> wrote:
> >
> > Matt attempted to add CLOCK_TAI support to the VDSO clock_gettime()
> > implementation, which extended the clockid switch case and added yet
> > another slightly different copy of the same code.
> >
> > Especially the extended switch case is problematic as the compiler tends to
> > generate a jump table which then requires to use retpolines. If jump tables
> > are disabled it adds yet another conditional to the existing maze.
> >
> > This series takes a different approach by consolidating the almost
> > identical functions into one implementation for high resolution clocks and
> > one for the coarse grained clock ids by storing the base data for each
> > clock id in an array which is indexed by the clock id.
> >
>
> I was trying to understand more of the implications of this patch
> series, and I was again reminded that there is an entire extra copy of
> the vclock reading code in arch/x86/kvm/x86.c. And the purpose of
> that code is very, very opaque.
>
> Can one of you explain what the code is even doing? From a couple of
> attempts to read through it, it's a whole bunch of
> probably-extremely-buggy code that,
Yes, probably.
> drumroll please, tries to atomically read the TSC value and the time. And decide whether the
> result is "based on the TSC".
I think "based on the TSC" refers to whether TSC clocksource is being
used.
> And then synthesizes a TSC-to-ns
> multiplier and shift, based on *something other than the actual
> multiply and shift used*.
>
> IOW, unless I'm totally misunderstanding it, the code digs into the
> private arch clocksource data intended for the vDSO, uses a poorly
> maintained copy of the vDSO code to read the time (instead of doing
> the sane thing and using the kernel interfaces for this), and
> propagates a totally made up copy to the guest.
I posted kernel interfaces for this, and it was suggested to
instead write a "in-kernel user of pvclock data".
If you can get kernel interfaces to replace that, go for it. I prefer
kernel interfaces as well.
> And gets it entirely
> wrong when doing nested virt, since, unless there's some secret in
> this maze, it doesn't acutlaly use the scaling factor from the host
> when it tells the guest what to do.
>
> I am really, seriously tempted to send a patch to simply delete all
> this code.
If your patch which deletes the code gets the necessary features right,
sure, go for it.
> The correct way to do it is to hook
Can you expand on the correct way to do it?
> And I don't see how it's even possible to pass kvmclock correctly to
> the L2 guest when L0 is hyperv. KVM could pass *hyperv's* clock, but
> L1 isn't notified when the data structure changes, so how the heck is
> it supposed to update the kvmclock structure?
I don't parse your question.
>
> --Andy
On Wed, Oct 03, 2018 at 11:22:58AM +0200, Vitaly Kuznetsov wrote:
> Andy Lutomirski <[email protected]> writes:
>
> > Hi Vitaly, Paolo, Radim, etc.,
> >
> > On Fri, Sep 14, 2018 at 5:52 AM Thomas Gleixner <[email protected]> wrote:
> >>
> >> Matt attempted to add CLOCK_TAI support to the VDSO clock_gettime()
> >> implementation, which extended the clockid switch case and added yet
> >> another slightly different copy of the same code.
> >>
> >> Especially the extended switch case is problematic as the compiler tends to
> >> generate a jump table which then requires to use retpolines. If jump tables
> >> are disabled it adds yet another conditional to the existing maze.
> >>
> >> This series takes a different approach by consolidating the almost
> >> identical functions into one implementation for high resolution clocks and
> >> one for the coarse grained clock ids by storing the base data for each
> >> clock id in an array which is indexed by the clock id.
> >>
> >
> > I was trying to understand more of the implications of this patch
> > series, and I was again reminded that there is an entire extra copy of
> > the vclock reading code in arch/x86/kvm/x86.c. And the purpose of
> > that code is very, very opaque.
> >
> > Can one of you explain what the code is even doing? From a couple of
> > attempts to read through it, it's a whole bunch of
> > probably-extremely-buggy code that, drumroll please, tries to
> > atomically read the TSC value and the time. And decide whether the
> > result is "based on the TSC". And then synthesizes a TSC-to-ns
> > multiplier and shift, based on *something other than the actual
> > multiply and shift used*.
> >
> > IOW, unless I'm totally misunderstanding it, the code digs into the
> > private arch clocksource data intended for the vDSO, uses a poorly
> > maintained copy of the vDSO code to read the time (instead of doing
> > the sane thing and using the kernel interfaces for this), and
> > propagates a totally made up copy to the guest. And gets it entirely
> > wrong when doing nested virt, since, unless there's some secret in
> > this maze, it doesn't acutlaly use the scaling factor from the host
> > when it tells the guest what to do.
> >
> > I am really, seriously tempted to send a patch to simply delete all
> > this code. The correct way to do it is to hook
>
> "I have discovered a truly marvelous proof of this, which this margin is
> too narrow to contain" :-)
>
> There is a very long history of different (hardware) issues Marcelo was
> fighting with and the current code is the survived Frankenstein.
Right, the code has to handle different TSC modes.
> E.g. it
> is very, very unclear what "catchup", "always catchup" and
> masterclock-less mode in general are and if we still need them.
Catchup means handling exposed (to guest) TSC frequency smaller than
HW TSC frequency.
That is "frankenstein" code, could be removed.
> That said I'm all for simplification. I'm not sure if we still need to
> care about buggy hardware though.
What simplification is that again?
> >
> > And I don't see how it's even possible to pass kvmclock correctly to
> > the L2 guest when L0 is hyperv. KVM could pass *hyperv's* clock, but
> > L1 isn't notified when the data structure changes, so how the heck is
> > it supposed to update the kvmclock structure?
>
> Well, this kind of works in the the followin way:
> L1's clocksource is 'tsc_page' which is, basically, a compliment to TSC:
> two numbers provided by L0: offset and scale and KVM was tought to treat
> this clocksource as a good one (see b0c39dc68e3b "x86/kvm: Pass stable
> clocksource to guests when running nested on Hyper-V").
>
> The notification you're talking about exists, it is called
> Reenligntenment, see 0092e4346f49 "x86/kvm: Support Hyper-V
> reenlightenment"). When TSC page changes (and this only happens when L1
> is migrated to a different host with a different TSC frequency and TSC
> scaling is not supported by the CPU) we receive an interrupt in L1 (at
> this moment all TSC accesses are emulated which guarantees the
> correctness of the readings), pause all L2 guests, update their kvmclock
> structures with new data (we already know the new TSC frequency) and
> then tell L0 that we're done and it can stop emulating TSC accesses.
>
> (Nothing like this exists for KVM-on-KVM, by the way, when L1's
> clocksource is 'kvmclock' L2s won't get a stable kvmclock clocksource.)
>
> --
> Vitaly
On Wed, Oct 03, 2018 at 04:00:29PM -0300, Marcelo Tosatti wrote:
> On Tue, Oct 02, 2018 at 10:15:49PM -0700, Andy Lutomirski wrote:
> > Hi Vitaly, Paolo, Radim, etc.,
> >
> > On Fri, Sep 14, 2018 at 5:52 AM Thomas Gleixner <[email protected]> wrote:
> > >
> > > Matt attempted to add CLOCK_TAI support to the VDSO clock_gettime()
> > > implementation, which extended the clockid switch case and added yet
> > > another slightly different copy of the same code.
> > >
> > > Especially the extended switch case is problematic as the compiler tends to
> > > generate a jump table which then requires to use retpolines. If jump tables
> > > are disabled it adds yet another conditional to the existing maze.
> > >
> > > This series takes a different approach by consolidating the almost
> > > identical functions into one implementation for high resolution clocks and
> > > one for the coarse grained clock ids by storing the base data for each
> > > clock id in an array which is indexed by the clock id.
> > >
> >
> > I was trying to understand more of the implications of this patch
> > series, and I was again reminded that there is an entire extra copy of
> > the vclock reading code in arch/x86/kvm/x86.c. And the purpose of
> > that code is very, very opaque.
> >
> > Can one of you explain what the code is even doing? From a couple of
> > attempts to read through it, it's a whole bunch of
> > probably-extremely-buggy code that,
>
> Yes, probably.
>
> > drumroll please, tries to atomically read the TSC value and the time. And decide whether the
> > result is "based on the TSC".
>
> I think "based on the TSC" refers to whether TSC clocksource is being
> used.
>
> > And then synthesizes a TSC-to-ns
> > multiplier and shift, based on *something other than the actual
> > multiply and shift used*.
> >
> > IOW, unless I'm totally misunderstanding it, the code digs into the
> > private arch clocksource data intended for the vDSO, uses a poorly
> > maintained copy of the vDSO code to read the time (instead of doing
> > the sane thing and using the kernel interfaces for this), and
> > propagates a totally made up copy to the guest.
>
> I posted kernel interfaces for this, and it was suggested to
> instead write a "in-kernel user of pvclock data".
>
> If you can get kernel interfaces to replace that, go for it. I prefer
> kernel interfaces as well.
And cleanup patches, to make that code look nicer, are also very
welcome!
On Wed, Oct 3, 2018 at 12:01 PM Marcelo Tosatti <[email protected]> wrote:
>
> On Tue, Oct 02, 2018 at 10:15:49PM -0700, Andy Lutomirski wrote:
> > Hi Vitaly, Paolo, Radim, etc.,
> >
> > On Fri, Sep 14, 2018 at 5:52 AM Thomas Gleixner <[email protected]> wrote:
> > >
> > > Matt attempted to add CLOCK_TAI support to the VDSO clock_gettime()
> > > implementation, which extended the clockid switch case and added yet
> > > another slightly different copy of the same code.
> > >
> > > Especially the extended switch case is problematic as the compiler tends to
> > > generate a jump table which then requires to use retpolines. If jump tables
> > > are disabled it adds yet another conditional to the existing maze.
> > >
> > > This series takes a different approach by consolidating the almost
> > > identical functions into one implementation for high resolution clocks and
> > > one for the coarse grained clock ids by storing the base data for each
> > > clock id in an array which is indexed by the clock id.
> > >
> >
> > I was trying to understand more of the implications of this patch
> > series, and I was again reminded that there is an entire extra copy of
> > the vclock reading code in arch/x86/kvm/x86.c. And the purpose of
> > that code is very, very opaque.
> >
> > Can one of you explain what the code is even doing? From a couple of
> > attempts to read through it, it's a whole bunch of
> > probably-extremely-buggy code that,
>
> Yes, probably.
>
> > drumroll please, tries to atomically read the TSC value and the time. And decide whether the
> > result is "based on the TSC".
>
> I think "based on the TSC" refers to whether TSC clocksource is being
> used.
>
> > And then synthesizes a TSC-to-ns
> > multiplier and shift, based on *something other than the actual
> > multiply and shift used*.
> >
> > IOW, unless I'm totally misunderstanding it, the code digs into the
> > private arch clocksource data intended for the vDSO, uses a poorly
> > maintained copy of the vDSO code to read the time (instead of doing
> > the sane thing and using the kernel interfaces for this), and
> > propagates a totally made up copy to the guest.
>
> I posted kernel interfaces for this, and it was suggested to
> instead write a "in-kernel user of pvclock data".
>
> If you can get kernel interfaces to replace that, go for it. I prefer
> kernel interfaces as well.
>
> > And gets it entirely
> > wrong when doing nested virt, since, unless there's some secret in
> > this maze, it doesn't acutlaly use the scaling factor from the host
> > when it tells the guest what to do.
> >
> > I am really, seriously tempted to send a patch to simply delete all
> > this code.
>
> If your patch which deletes the code gets the necessary features right,
> sure, go for it.
>
> > The correct way to do it is to hook
>
> Can you expand on the correct way to do it?
>
> > And I don't see how it's even possible to pass kvmclock correctly to
> > the L2 guest when L0 is hyperv. KVM could pass *hyperv's* clock, but
> > L1 isn't notified when the data structure changes, so how the heck is
> > it supposed to update the kvmclock structure?
>
> I don't parse your question.
Let me ask it more intelligently: when the "reenlightenment" IRQ
happens, what tells KVM to do its own update for its guests?
Marcelo Tosatti <[email protected]> writes:
> On Wed, Oct 03, 2018 at 11:22:58AM +0200, Vitaly Kuznetsov wrote:
>>
>> There is a very long history of different (hardware) issues Marcelo was
>> fighting with and the current code is the survived Frankenstein.
>
> Right, the code has to handle different TSC modes.
>
>> E.g. it
>> is very, very unclear what "catchup", "always catchup" and
>> masterclock-less mode in general are and if we still need them.
>
> Catchup means handling exposed (to guest) TSC frequency smaller than
> HW TSC frequency.
>
> That is "frankenstein" code, could be removed.
>
>> That said I'm all for simplification. I'm not sure if we still need to
>> care about buggy hardware though.
>
> What simplification is that again?
>
I was hoping to hear this from you :-) If I am to suggest how we can
move forward I'd propose:
- Check if pure TSC can be used on SkyLake+ systems (where TSC scaling
is supported).
- Check if non-masterclock mode is still needed. E.g. HyperV's TSC page
clocksource is a single page for the whole VM, not a per-cpu thing. Can
we think that all the buggy hardware is already gone?
--
Vitaly
On Thu, Oct 04, 2018 at 09:54:45AM +0200, Vitaly Kuznetsov wrote:
> I was hoping to hear this from you :-) If I am to suggest how we can
> move forward I'd propose:
> - Check if pure TSC can be used on SkyLake+ systems (where TSC scaling
> is supported).
> - Check if non-masterclock mode is still needed. E.g. HyperV's TSC page
> clocksource is a single page for the whole VM, not a per-cpu thing. Can
> we think that all the buggy hardware is already gone?
No, and it is not the hardware you have to worry about (mostly), it is
the frigging PoS firmware people put on it.
Ever since Nehalem TSC is stable (unless you get to >4 socket systems,
after which it still can be, but bets are off). But even relatively
recent systems fail the TSC sync test because firmware messes it up by
writing to either MSR_TSC or MSR_TSC_ADJUST.
But the thing is, if the TSC is not synced, you cannot use it for
timekeeping, full stop. So having a single page is fine, it either
contains a mult/shift that is valid, or it indicates TSC is messed up
and you fall back to something else.
There is no inbetween there.
For sched_clock we can still use the global page, because the rate will
still be the same for each cpu, it's just offset between CPUs and the
code compensates for that.
On 04/10/2018 09:54, Vitaly Kuznetsov wrote:
> - Check if pure TSC can be used on SkyLake+ systems (where TSC scaling
> is supported).
Not if you want to migrate to pre-Skylake systems.
> - Check if non-masterclock mode is still needed. E.g. HyperV's TSC page
> clocksource is a single page for the whole VM, not a per-cpu thing. Can
> we think that all the buggy hardware is already gone?
No. :( We still get reports whenever we break 2007-2008 hardware.
Paolo
> On Oct 4, 2018, at 1:11 AM, Peter Zijlstra <[email protected]> wrote:
>
>> On Thu, Oct 04, 2018 at 09:54:45AM +0200, Vitaly Kuznetsov wrote:
>> I was hoping to hear this from you :-) If I am to suggest how we can
>> move forward I'd propose:
>> - Check if pure TSC can be used on SkyLake+ systems (where TSC scaling
>> is supported).
>> - Check if non-masterclock mode is still needed. E.g. HyperV's TSC page
>> clocksource is a single page for the whole VM, not a per-cpu thing. Can
>> we think that all the buggy hardware is already gone?
>
> No, and it is not the hardware you have to worry about (mostly), it is
> the frigging PoS firmware people put on it.
>
> Ever since Nehalem TSC is stable (unless you get to >4 socket systems,
> after which it still can be, but bets are off). But even relatively
> recent systems fail the TSC sync test because firmware messes it up by
> writing to either MSR_TSC or MSR_TSC_ADJUST.
>
> But the thing is, if the TSC is not synced, you cannot use it for
> timekeeping, full stop. So having a single page is fine, it either
> contains a mult/shift that is valid, or it indicates TSC is messed up
> and you fall back to something else.
>
> There is no inbetween there.
>
> For sched_clock we can still use the global page, because the rate will
> still be the same for each cpu, it's just offset between CPUs and the
> code compensates for that.
But if we’re in a KVM guest, then the clock will jump around on the same *vCPU* when the vCPU migrates.
But I don’t see how kvmclock helps here, since I don’t think it’s used for sched_clock.
> On Oct 4, 2018, at 5:00 AM, Paolo Bonzini <[email protected]> wrote:
>
>> On 04/10/2018 09:54, Vitaly Kuznetsov wrote:
>> - Check if pure TSC can be used on SkyLake+ systems (where TSC scaling
>> is supported).
>
> Not if you want to migrate to pre-Skylake systems.
>
>> - Check if non-masterclock mode is still needed. E.g. HyperV's TSC page
>> clocksource is a single page for the whole VM, not a per-cpu thing. Can
>> we think that all the buggy hardware is already gone?
>
> No. :( We still get reports whenever we break 2007-2008 hardware.
>
>
Does the KVM non-masterclock mode actually help? It’s not clear to me exactly how it’s supposed to work, but it seems like it’s trying to expose per-vCPU adjustments to the guest. Which is dubious at best, since the guest can’t validly use them for anything other than sched_clock, since they aren’t fully corrected by anything KVM can do.
On Wed, Oct 03, 2018 at 03:32:08PM -0700, Andy Lutomirski wrote:
> On Wed, Oct 3, 2018 at 12:01 PM Marcelo Tosatti <[email protected]> wrote:
> >
> > On Tue, Oct 02, 2018 at 10:15:49PM -0700, Andy Lutomirski wrote:
> > > Hi Vitaly, Paolo, Radim, etc.,
> > >
> > > On Fri, Sep 14, 2018 at 5:52 AM Thomas Gleixner <[email protected]> wrote:
> > > >
> > > > Matt attempted to add CLOCK_TAI support to the VDSO clock_gettime()
> > > > implementation, which extended the clockid switch case and added yet
> > > > another slightly different copy of the same code.
> > > >
> > > > Especially the extended switch case is problematic as the compiler tends to
> > > > generate a jump table which then requires to use retpolines. If jump tables
> > > > are disabled it adds yet another conditional to the existing maze.
> > > >
> > > > This series takes a different approach by consolidating the almost
> > > > identical functions into one implementation for high resolution clocks and
> > > > one for the coarse grained clock ids by storing the base data for each
> > > > clock id in an array which is indexed by the clock id.
> > > >
> > >
> > > I was trying to understand more of the implications of this patch
> > > series, and I was again reminded that there is an entire extra copy of
> > > the vclock reading code in arch/x86/kvm/x86.c. And the purpose of
> > > that code is very, very opaque.
> > >
> > > Can one of you explain what the code is even doing? From a couple of
> > > attempts to read through it, it's a whole bunch of
> > > probably-extremely-buggy code that,
> >
> > Yes, probably.
> >
> > > drumroll please, tries to atomically read the TSC value and the time. And decide whether the
> > > result is "based on the TSC".
> >
> > I think "based on the TSC" refers to whether TSC clocksource is being
> > used.
> >
> > > And then synthesizes a TSC-to-ns
> > > multiplier and shift, based on *something other than the actual
> > > multiply and shift used*.
> > >
> > > IOW, unless I'm totally misunderstanding it, the code digs into the
> > > private arch clocksource data intended for the vDSO, uses a poorly
> > > maintained copy of the vDSO code to read the time (instead of doing
> > > the sane thing and using the kernel interfaces for this), and
> > > propagates a totally made up copy to the guest.
> >
> > I posted kernel interfaces for this, and it was suggested to
> > instead write a "in-kernel user of pvclock data".
> >
> > If you can get kernel interfaces to replace that, go for it. I prefer
> > kernel interfaces as well.
> >
> > > And gets it entirely
> > > wrong when doing nested virt, since, unless there's some secret in
> > > this maze, it doesn't acutlaly use the scaling factor from the host
> > > when it tells the guest what to do.
> > >
> > > I am really, seriously tempted to send a patch to simply delete all
> > > this code.
> >
> > If your patch which deletes the code gets the necessary features right,
> > sure, go for it.
> >
> > > The correct way to do it is to hook
> >
> > Can you expand on the correct way to do it?
> >
> > > And I don't see how it's even possible to pass kvmclock correctly to
> > > the L2 guest when L0 is hyperv. KVM could pass *hyperv's* clock, but
> > > L1 isn't notified when the data structure changes, so how the heck is
> > > it supposed to update the kvmclock structure?
> >
> > I don't parse your question.
>
> Let me ask it more intelligently: when the "reenlightenment" IRQ
> happens, what tells KVM to do its own update for its guests?
Update of what, and why it needs to update anything from IRQ?
The update i can think of is from host kernel clocksource,
which there is a notifier for.
On Thu, Oct 4, 2018 at 9:43 AM Marcelo Tosatti <[email protected]> wrote:
>
> On Wed, Oct 03, 2018 at 03:32:08PM -0700, Andy Lutomirski wrote:
> > On Wed, Oct 3, 2018 at 12:01 PM Marcelo Tosatti <[email protected]> wrote:
> > >
> > > On Tue, Oct 02, 2018 at 10:15:49PM -0700, Andy Lutomirski wrote:
> > > > Hi Vitaly, Paolo, Radim, etc.,
> > > >
> > > > On Fri, Sep 14, 2018 at 5:52 AM Thomas Gleixner <[email protected]> wrote:
> > > > >
> > > > > Matt attempted to add CLOCK_TAI support to the VDSO clock_gettime()
> > > > > implementation, which extended the clockid switch case and added yet
> > > > > another slightly different copy of the same code.
> > > > >
> > > > > Especially the extended switch case is problematic as the compiler tends to
> > > > > generate a jump table which then requires to use retpolines. If jump tables
> > > > > are disabled it adds yet another conditional to the existing maze.
> > > > >
> > > > > This series takes a different approach by consolidating the almost
> > > > > identical functions into one implementation for high resolution clocks and
> > > > > one for the coarse grained clock ids by storing the base data for each
> > > > > clock id in an array which is indexed by the clock id.
> > > > >
> > > >
> > > > I was trying to understand more of the implications of this patch
> > > > series, and I was again reminded that there is an entire extra copy of
> > > > the vclock reading code in arch/x86/kvm/x86.c. And the purpose of
> > > > that code is very, very opaque.
> > > >
> > > > Can one of you explain what the code is even doing? From a couple of
> > > > attempts to read through it, it's a whole bunch of
> > > > probably-extremely-buggy code that,
> > >
> > > Yes, probably.
> > >
> > > > drumroll please, tries to atomically read the TSC value and the time. And decide whether the
> > > > result is "based on the TSC".
> > >
> > > I think "based on the TSC" refers to whether TSC clocksource is being
> > > used.
> > >
> > > > And then synthesizes a TSC-to-ns
> > > > multiplier and shift, based on *something other than the actual
> > > > multiply and shift used*.
> > > >
> > > > IOW, unless I'm totally misunderstanding it, the code digs into the
> > > > private arch clocksource data intended for the vDSO, uses a poorly
> > > > maintained copy of the vDSO code to read the time (instead of doing
> > > > the sane thing and using the kernel interfaces for this), and
> > > > propagates a totally made up copy to the guest.
> > >
> > > I posted kernel interfaces for this, and it was suggested to
> > > instead write a "in-kernel user of pvclock data".
> > >
> > > If you can get kernel interfaces to replace that, go for it. I prefer
> > > kernel interfaces as well.
> > >
> > > > And gets it entirely
> > > > wrong when doing nested virt, since, unless there's some secret in
> > > > this maze, it doesn't acutlaly use the scaling factor from the host
> > > > when it tells the guest what to do.
> > > >
> > > > I am really, seriously tempted to send a patch to simply delete all
> > > > this code.
> > >
> > > If your patch which deletes the code gets the necessary features right,
> > > sure, go for it.
> > >
> > > > The correct way to do it is to hook
> > >
> > > Can you expand on the correct way to do it?
> > >
> > > > And I don't see how it's even possible to pass kvmclock correctly to
> > > > the L2 guest when L0 is hyperv. KVM could pass *hyperv's* clock, but
> > > > L1 isn't notified when the data structure changes, so how the heck is
> > > > it supposed to update the kvmclock structure?
> > >
> > > I don't parse your question.
> >
> > Let me ask it more intelligently: when the "reenlightenment" IRQ
> > happens, what tells KVM to do its own update for its guests?
>
> Update of what, and why it needs to update anything from IRQ?
>
> The update i can think of is from host kernel clocksource,
> which there is a notifier for.
>
>
Unless I've missed some serious magic, L2 guests see kvmclock, not hv.
So we have the following sequence of events:
- L0 migrates the whole VM. Starting now, RDTSC is emulated to match
the old host, which applies in L1 and L2.
- An IRQ is queued to L1.
- L1 acknowledges that it noticed the TSC change. RDTSC stops being
emulated for L1 and L2.
- L2 reads the TSC. It has no idea that anything changed, and it
gets the wrong answer.
- At some point, kvm clock updates.
What prevents this? Vitaly, am I missing some subtlety of what
actually happens?
Andy Lutomirski <[email protected]> writes:
> On Thu, Oct 4, 2018 at 9:43 AM Marcelo Tosatti <[email protected]> wrote:
>>
>> On Wed, Oct 03, 2018 at 03:32:08PM -0700, Andy Lutomirski wrote:
>> > On Wed, Oct 3, 2018 at 12:01 PM Marcelo Tosatti <[email protected]> wrote:
>> > >
>> > > On Tue, Oct 02, 2018 at 10:15:49PM -0700, Andy Lutomirski wrote:
>> > > > Hi Vitaly, Paolo, Radim, etc.,
>> > > >
>> > > > On Fri, Sep 14, 2018 at 5:52 AM Thomas Gleixner <[email protected]> wrote:
>> > > > >
>> > > > > Matt attempted to add CLOCK_TAI support to the VDSO clock_gettime()
>> > > > > implementation, which extended the clockid switch case and added yet
>> > > > > another slightly different copy of the same code.
>> > > > >
>> > > > > Especially the extended switch case is problematic as the compiler tends to
>> > > > > generate a jump table which then requires to use retpolines. If jump tables
>> > > > > are disabled it adds yet another conditional to the existing maze.
>> > > > >
>> > > > > This series takes a different approach by consolidating the almost
>> > > > > identical functions into one implementation for high resolution clocks and
>> > > > > one for the coarse grained clock ids by storing the base data for each
>> > > > > clock id in an array which is indexed by the clock id.
>> > > > >
>> > > >
>> > > > I was trying to understand more of the implications of this patch
>> > > > series, and I was again reminded that there is an entire extra copy of
>> > > > the vclock reading code in arch/x86/kvm/x86.c. And the purpose of
>> > > > that code is very, very opaque.
>> > > >
>> > > > Can one of you explain what the code is even doing? From a couple of
>> > > > attempts to read through it, it's a whole bunch of
>> > > > probably-extremely-buggy code that,
>> > >
>> > > Yes, probably.
>> > >
>> > > > drumroll please, tries to atomically read the TSC value and the time. And decide whether the
>> > > > result is "based on the TSC".
>> > >
>> > > I think "based on the TSC" refers to whether TSC clocksource is being
>> > > used.
>> > >
>> > > > And then synthesizes a TSC-to-ns
>> > > > multiplier and shift, based on *something other than the actual
>> > > > multiply and shift used*.
>> > > >
>> > > > IOW, unless I'm totally misunderstanding it, the code digs into the
>> > > > private arch clocksource data intended for the vDSO, uses a poorly
>> > > > maintained copy of the vDSO code to read the time (instead of doing
>> > > > the sane thing and using the kernel interfaces for this), and
>> > > > propagates a totally made up copy to the guest.
>> > >
>> > > I posted kernel interfaces for this, and it was suggested to
>> > > instead write a "in-kernel user of pvclock data".
>> > >
>> > > If you can get kernel interfaces to replace that, go for it. I prefer
>> > > kernel interfaces as well.
>> > >
>> > > > And gets it entirely
>> > > > wrong when doing nested virt, since, unless there's some secret in
>> > > > this maze, it doesn't acutlaly use the scaling factor from the host
>> > > > when it tells the guest what to do.
>> > > >
>> > > > I am really, seriously tempted to send a patch to simply delete all
>> > > > this code.
>> > >
>> > > If your patch which deletes the code gets the necessary features right,
>> > > sure, go for it.
>> > >
>> > > > The correct way to do it is to hook
>> > >
>> > > Can you expand on the correct way to do it?
>> > >
>> > > > And I don't see how it's even possible to pass kvmclock correctly to
>> > > > the L2 guest when L0 is hyperv. KVM could pass *hyperv's* clock, but
>> > > > L1 isn't notified when the data structure changes, so how the heck is
>> > > > it supposed to update the kvmclock structure?
>> > >
>> > > I don't parse your question.
>> >
>> > Let me ask it more intelligently: when the "reenlightenment" IRQ
>> > happens, what tells KVM to do its own update for its guests?
>>
>> Update of what, and why it needs to update anything from IRQ?
>>
>> The update i can think of is from host kernel clocksource,
>> which there is a notifier for.
>>
>>
>
> Unless I've missed some serious magic, L2 guests see kvmclock, not hv.
> So we have the following sequence of events:
>
> - L0 migrates the whole VM. Starting now, RDTSC is emulated to match
> the old host, which applies in L1 and L2.
>
> - An IRQ is queued to L1.
>
> - L1 acknowledges that it noticed the TSC change.
Before the acknowledgement we actually pause all guests so they don't
notice the change ....
> RDTSC stops being emulated for L1 and L2.
.... and right after that we update all kvmclocks for all L2s and
unpause them so all their readings are still correct (see
kvm_hyperv_tsc_notifier()).
>
> - L2 reads the TSC. It has no idea that anything changed, and it
> gets the wrong answer.
I have to admit I forgot what happens if L2 uses raw TSC. I *think* that
we actually adjust TSC offset along with adjusting kvmclocks so the
reading is still correct. I'll have to check this.
All bets are off in case L2 was using TSC for time interval
measurements: frequency, of course, changes.
>
> - At some point, kvm clock updates.
>
> What prevents this? Vitaly, am I missing some subtlety of what
> actually happens?
--
Vitaly
On Thu, Oct 04, 2018 at 07:00:45AM -0700, Andy Lutomirski wrote:
> > On Oct 4, 2018, at 1:11 AM, Peter Zijlstra <[email protected]> wrote:
> >
> >> On Thu, Oct 04, 2018 at 09:54:45AM +0200, Vitaly Kuznetsov wrote:
> >> I was hoping to hear this from you :-) If I am to suggest how we can
> >> move forward I'd propose:
> >> - Check if pure TSC can be used on SkyLake+ systems (where TSC scaling
> >> is supported).
> >> - Check if non-masterclock mode is still needed. E.g. HyperV's TSC page
> >> clocksource is a single page for the whole VM, not a per-cpu thing. Can
> >> we think that all the buggy hardware is already gone?
> >
> > No, and it is not the hardware you have to worry about (mostly), it is
> > the frigging PoS firmware people put on it.
> >
> > Ever since Nehalem TSC is stable (unless you get to >4 socket systems,
> > after which it still can be, but bets are off). But even relatively
> > recent systems fail the TSC sync test because firmware messes it up by
> > writing to either MSR_TSC or MSR_TSC_ADJUST.
> >
> > But the thing is, if the TSC is not synced, you cannot use it for
> > timekeeping, full stop. So having a single page is fine, it either
> > contains a mult/shift that is valid, or it indicates TSC is messed up
> > and you fall back to something else.
> >
> > There is no inbetween there.
> >
> > For sched_clock we can still use the global page, because the rate will
> > still be the same for each cpu, it's just offset between CPUs and the
> > code compensates for that.
>
> But if we’re in a KVM guest, then the clock will jump around on the
> same *vCPU* when the vCPU migrates.
Urgh yes..
> But I don’t see how kvmclock helps here, since I don’t think it’s used
> for sched_clock.
I get hits on kvm_sched_clock, but haven't looked further.
Anyway, Most of the argument still holds, either TSC is synced or it is
not and it _really_ should not be used. Not much middle ground there.
> On Oct 4, 2018, at 12:31 PM, Peter Zijlstra <[email protected]> wrote:
>
> On Thu, Oct 04, 2018 at 07:00:45AM -0700, Andy Lutomirski wrote:
>>> On Oct 4, 2018, at 1:11 AM, Peter Zijlstra <[email protected]> wrote:
>>>
>>>> On Thu, Oct 04, 2018 at 09:54:45AM +0200, Vitaly Kuznetsov wrote:
>>>> I was hoping to hear this from you :-) If I am to suggest how we can
>>>> move forward I'd propose:
>>>> - Check if pure TSC can be used on SkyLake+ systems (where TSC scaling
>>>> is supported).
>>>> - Check if non-masterclock mode is still needed. E.g. HyperV's TSC page
>>>> clocksource is a single page for the whole VM, not a per-cpu thing. Can
>>>> we think that all the buggy hardware is already gone?
>>>
>>> No, and it is not the hardware you have to worry about (mostly), it is
>>> the frigging PoS firmware people put on it.
>>>
>>> Ever since Nehalem TSC is stable (unless you get to >4 socket systems,
>>> after which it still can be, but bets are off). But even relatively
>>> recent systems fail the TSC sync test because firmware messes it up by
>>> writing to either MSR_TSC or MSR_TSC_ADJUST.
>>>
>>> But the thing is, if the TSC is not synced, you cannot use it for
>>> timekeeping, full stop. So having a single page is fine, it either
>>> contains a mult/shift that is valid, or it indicates TSC is messed up
>>> and you fall back to something else.
>>>
>>> There is no inbetween there.
>>>
>>> For sched_clock we can still use the global page, because the rate will
>>> still be the same for each cpu, it's just offset between CPUs and the
>>> code compensates for that.
>>
>> But if we’re in a KVM guest, then the clock will jump around on the
>> same *vCPU* when the vCPU migrates.
>
> Urgh yes..
>
>> But I don’t see how kvmclock helps here, since I don’t think it’s used
>> for sched_clock.
>
> I get hits on kvm_sched_clock, but haven't looked further.
Ok, so KVM is using the per-vCPU pvclock data for sched_clock. Which hopefully does something intelligent.
Regardless of any TSC syncing issues, a paravirt clock should presumably be used for sched_clock to account for time that the vCPU was stopped.
It would be fantastic if this stuff were documented much better, both in terms of the data structures and the Linux code.
On Fri, Sep 14, 2018 at 5:52 AM Thomas Gleixner <[email protected]> wrote:
>
> Matt attempted to add CLOCK_TAI support to the VDSO clock_gettime()
> implementation, which extended the clockid switch case and added yet
> another slightly different copy of the same code.
>
> Especially the extended switch case is problematic as the compiler tends to
> generate a jump table which then requires to use retpolines. If jump tables
> are disabled it adds yet another conditional to the existing maze.
>
> This series takes a different approach by consolidating the almost
> identical functions into one implementation for high resolution clocks and
> one for the coarse grained clock ids by storing the base data for each
> clock id in an array which is indexed by the clock id.
>
> This completely eliminates the switch case and allows further
> simplifications of the code base, which at the end all together gain a few
> cycles performance or at least stay on par with todays code. The resulting
> performance depends heavily on the micro architecture and the compiler.
tglx, please consider this whole series Acked-by: Andy Lutomirski
<[email protected]>
Please feel free to push it top tip:x86/vdso, as long as you pull in
tip/x86/urgent first. Once it lands, I'll email out a couple of
follow-up patches.
--Andy
For better or for worse, I'm trying to understand this code. So far,
I've come up with this patch:
https://git.kernel.org/pub/scm/linux/kernel/git/luto/linux.git/commit/?h=x86/vdso-tglx&id=14fd71e12b1c4492a06f368f75041f263e6862bf
Is it correct, or am I missing some subtlety?
On Thu, Oct 04, 2018 at 09:54:45AM +0200, Vitaly Kuznetsov wrote:
> Marcelo Tosatti <[email protected]> writes:
>
> > On Wed, Oct 03, 2018 at 11:22:58AM +0200, Vitaly Kuznetsov wrote:
> >>
> >> There is a very long history of different (hardware) issues Marcelo was
> >> fighting with and the current code is the survived Frankenstein.
> >
> > Right, the code has to handle different TSC modes.
> >
> >> E.g. it
> >> is very, very unclear what "catchup", "always catchup" and
> >> masterclock-less mode in general are and if we still need them.
> >
> > Catchup means handling exposed (to guest) TSC frequency smaller than
> > HW TSC frequency.
> >
> > That is "frankenstein" code, could be removed.
> >
> >> That said I'm all for simplification. I'm not sure if we still need to
> >> care about buggy hardware though.
> >
> > What simplification is that again?
> >
>
> I was hoping to hear this from you :-) If I am to suggest how we can
> move forward I'd propose:
> - Check if pure TSC can be used on SkyLake+ systems (where TSC scaling
> is supported).
In that case just use TSC clocksource on the guest directly: i am
writing code for that now (its faster than pvclock syscall).
> - Check if non-masterclock mode is still needed. E.g. HyperV's TSC page
> clocksource is a single page for the whole VM, not a per-cpu thing. Can
> we think that all the buggy hardware is already gone?
As Peter mentioned, non sync TSC hardware might exist in the future.
And older hardware must be supported.
On Thu, Oct 04, 2018 at 03:15:32PM -0700, Andy Lutomirski wrote:
> For better or for worse, I'm trying to understand this code. So far,
> I've come up with this patch:
>
> https://git.kernel.org/pub/scm/linux/kernel/git/luto/linux.git/commit/?h=x86/vdso-tglx&id=14fd71e12b1c4492a06f368f75041f263e6862bf
>
> Is it correct, or am I missing some subtlety?
The master clock, when initialized, has a pair
masterclockvalues=(TSC value, time-of-day data).
When updating the guest clock, we only update relative to (TSC value)
that was read on masterclock initialization.
See the following comment on x86.c:
/*
*
* Assuming a stable TSC across physical CPUS, and a stable TSC
* across virtual CPUs, the following condition is possible.
* Each numbered line represents an event visible to both
* CPUs at the next numbered event.
...
When updating the "masterclockvalues" pair, all vcpus are
stopped.
On Sat, Oct 6, 2018 at 1:29 PM Marcelo Tosatti <[email protected]> wrote:
>
> On Thu, Oct 04, 2018 at 03:15:32PM -0700, Andy Lutomirski wrote:
> > For better or for worse, I'm trying to understand this code. So far,
> > I've come up with this patch:
> >
> > https://git.kernel.org/pub/scm/linux/kernel/git/luto/linux.git/commit/?h=x86/vdso-tglx&id=14fd71e12b1c4492a06f368f75041f263e6862bf
> >
> > Is it correct, or am I missing some subtlety?
>
> The master clock, when initialized, has a pair
>
> masterclockvalues=(TSC value, time-of-day data).
>
> When updating the guest clock, we only update relative to (TSC value)
> that was read on masterclock initialization.
I don't see the problem. The masterclock data is updated here:
host_tsc_clocksource = kvm_get_time_and_clockread(
&ka->master_kernel_ns,
&ka->master_cycle_now);
kvm_get_time_and_clockread() gets those values from
do_monotonic_boot(), which, barring bugs, should cause
get_kvmclock_ns() to return exactly the same thing as
ktime_get_boot_ns() + ka->kvmclock_offset, albeit in a rather
roundabout manner.
So what am I missing? Is there actually something wrong with my patch?
>
> See the following comment on x86.c:
I read that comment, and it's not obvious to me how it's related.
On Thu, Oct 04, 2018 at 03:15:32PM -0700, Andy Lutomirski wrote:
> For better or for worse, I'm trying to understand this code. So far,
> I've come up with this patch:
>
> https://git.kernel.org/pub/scm/linux/kernel/git/luto/linux.git/commit/?h=x86/vdso-tglx&id=14fd71e12b1c4492a06f368f75041f263e6862bf
>
> Is it correct, or am I missing some subtlety?
"In the non-fallback case, a bunch of gnarly arithmetic is done: a
hopefully matched pair of (TSC, boot time) is read, then all locks
are dropped, then the TSC frequency is read, a branch new multiplier
and shift is read, and the result is turned into a time.
This seems quite racy to me. Because locks are not held, I don't
see what keeps TSC frequency used in sync with the master clock
data."
If tsc_khz changes, the host TSC clocksource will not be used, which
disables masterclock:
if ((val == CPUFREQ_PRECHANGE && freq->old < freq->new) ||
(val == CPUFREQ_POSTCHANGE && freq->old >
freq->new)) {
*lpj = cpufreq_scale(loops_per_jiffy_ref, ref_freq,
freq->new);
tsc_khz = cpufreq_scale(tsc_khz_ref, ref_freq,
freq->new);
if (!(freq->flags & CPUFREQ_CONST_LOOPS))
mark_tsc_unstable("cpufreq changes");
In which case it ends up in:
- spin_lock(&ka->pvclock_gtod_sync_lock);
- if (!ka->use_master_clock) {
- spin_unlock(&ka->pvclock_gtod_sync_lock);
- return ktime_get_boot_ns() + ka->kvmclock_offset;
- }
masterclock -> non masterclock transition sets
a REQUEST bit on each vCPU, so as to invalidate any previous
clock reads.
static void kvm_gen_update_masterclock(struct kvm *kvm)
{
#ifdef CONFIG_X86_64
int i;
struct kvm_vcpu *vcpu;
struct kvm_arch *ka = &kvm->arch;
spin_lock(&ka->pvclock_gtod_sync_lock);
kvm_make_mclock_inprogress_request(kvm);
/* no guest entries from this point */
pvclock_update_vm_gtod_copy(kvm);
kvm_for_each_vcpu(i, vcpu, kvm)
kvm_make_request(KVM_REQ_CLOCK_UPDATE, vcpu);
/* guest entries allowed */
kvm_for_each_vcpu(i, vcpu, kvm)
kvm_clear_request(KVM_REQ_MCLOCK_INPROGRESS, vcpu);
spin_unlock(&ka->pvclock_gtod_sync_lock);
#endif
/*
* If the host uses TSC clock, then passthrough TSC as stable
* to the guest.
*/
host_tsc_clocksource = kvm_get_time_and_clockread(
&ka->master_kernel_ns,
&ka->master_cycle_now);
ka->use_master_clock = host_tsc_clocksource && vcpus_matched
&& !ka->backwards_tsc_observed
&& !ka->boot_vcpu_runs_old_kvmclock;
On Sat, Oct 06, 2018 at 03:28:05PM -0700, Andy Lutomirski wrote:
> On Sat, Oct 6, 2018 at 1:29 PM Marcelo Tosatti <[email protected]> wrote:
> >
> > On Thu, Oct 04, 2018 at 03:15:32PM -0700, Andy Lutomirski wrote:
> > > For better or for worse, I'm trying to understand this code. So far,
> > > I've come up with this patch:
> > >
> > > https://git.kernel.org/pub/scm/linux/kernel/git/luto/linux.git/commit/?h=x86/vdso-tglx&id=14fd71e12b1c4492a06f368f75041f263e6862bf
> > >
> > > Is it correct, or am I missing some subtlety?
> >
> > The master clock, when initialized, has a pair
> >
> > masterclockvalues=(TSC value, time-of-day data).
> >
> > When updating the guest clock, we only update relative to (TSC value)
> > that was read on masterclock initialization.
>
> I don't see the problem. The masterclock data is updated here:
>
> host_tsc_clocksource = kvm_get_time_and_clockread(
> &ka->master_kernel_ns,
> &ka->master_cycle_now);
>
> kvm_get_time_and_clockread() gets those values from
> do_monotonic_boot(), which, barring bugs, should cause
> get_kvmclock_ns() to return exactly the same thing as
> ktime_get_boot_ns() + ka->kvmclock_offset, albeit in a rather
> roundabout manner.
>
> So what am I missing? Is there actually something wrong with my patch?
For the bug mentioned in the comment not to happen, you must only read
TSC and add it as offset to (TSC value, time-of-day data).
Its more than "a roundabout manner".
Read the comment again.
>
>
> >
> > See the following comment on x86.c:
>
> I read that comment, and it's not obvious to me how it's related.
On Mon, Oct 8, 2018 at 8:27 AM Marcelo Tosatti <[email protected]> wrote:
>
> On Sat, Oct 06, 2018 at 03:28:05PM -0700, Andy Lutomirski wrote:
> > On Sat, Oct 6, 2018 at 1:29 PM Marcelo Tosatti <[email protected]> wrote:
> > >
> > > On Thu, Oct 04, 2018 at 03:15:32PM -0700, Andy Lutomirski wrote:
> > > > For better or for worse, I'm trying to understand this code. So far,
> > > > I've come up with this patch:
> > > >
> > > > https://git.kernel.org/pub/scm/linux/kernel/git/luto/linux.git/commit/?h=x86/vdso-tglx&id=14fd71e12b1c4492a06f368f75041f263e6862bf
> > > >
> > > > Is it correct, or am I missing some subtlety?
> > >
> > > The master clock, when initialized, has a pair
> > >
> > > masterclockvalues=(TSC value, time-of-day data).
> > >
> > > When updating the guest clock, we only update relative to (TSC value)
> > > that was read on masterclock initialization.
> >
> > I don't see the problem. The masterclock data is updated here:
> >
> > host_tsc_clocksource = kvm_get_time_and_clockread(
> > &ka->master_kernel_ns,
> > &ka->master_cycle_now);
> >
> > kvm_get_time_and_clockread() gets those values from
> > do_monotonic_boot(), which, barring bugs, should cause
> > get_kvmclock_ns() to return exactly the same thing as
> > ktime_get_boot_ns() + ka->kvmclock_offset, albeit in a rather
> > roundabout manner.
> >
> > So what am I missing? Is there actually something wrong with my patch?
>
> For the bug mentioned in the comment not to happen, you must only read
> TSC and add it as offset to (TSC value, time-of-day data).
>
> Its more than "a roundabout manner".
>
> Read the comment again.
>
I read the comment three more times and even dug through the git
history. It seems like what you're saying is that, under certain
conditions (which arguably would be bugs in the core Linux timing
code), actually calling ktime_get_boot_ns() could be non-monotonic
with respect to the kvmclock timing. But get_kvmclock_ns() isn't used
for VM timing as such -- it's used for the IOCTL interfaces for
updating the time offset. So can you explain how my patch is
incorrect?
--Andy
On Mon, Oct 08, 2018 at 10:38:22AM -0700, Andy Lutomirski wrote:
> On Mon, Oct 8, 2018 at 8:27 AM Marcelo Tosatti <[email protected]> wrote:
> >
> > On Sat, Oct 06, 2018 at 03:28:05PM -0700, Andy Lutomirski wrote:
> > > On Sat, Oct 6, 2018 at 1:29 PM Marcelo Tosatti <[email protected]> wrote:
> > > >
> > > > On Thu, Oct 04, 2018 at 03:15:32PM -0700, Andy Lutomirski wrote:
> > > > > For better or for worse, I'm trying to understand this code. So far,
> > > > > I've come up with this patch:
> > > > >
> > > > > https://git.kernel.org/pub/scm/linux/kernel/git/luto/linux.git/commit/?h=x86/vdso-tglx&id=14fd71e12b1c4492a06f368f75041f263e6862bf
> > > > >
> > > > > Is it correct, or am I missing some subtlety?
> > > >
> > > > The master clock, when initialized, has a pair
> > > >
> > > > masterclockvalues=(TSC value, time-of-day data).
> > > >
> > > > When updating the guest clock, we only update relative to (TSC value)
> > > > that was read on masterclock initialization.
> > >
> > > I don't see the problem. The masterclock data is updated here:
> > >
> > > host_tsc_clocksource = kvm_get_time_and_clockread(
> > > &ka->master_kernel_ns,
> > > &ka->master_cycle_now);
> > >
> > > kvm_get_time_and_clockread() gets those values from
> > > do_monotonic_boot(), which, barring bugs, should cause
> > > get_kvmclock_ns() to return exactly the same thing as
> > > ktime_get_boot_ns() + ka->kvmclock_offset, albeit in a rather
> > > roundabout manner.
> > >
> > > So what am I missing? Is there actually something wrong with my patch?
> >
> > For the bug mentioned in the comment not to happen, you must only read
> > TSC and add it as offset to (TSC value, time-of-day data).
> >
> > Its more than "a roundabout manner".
> >
> > Read the comment again.
> >
>
> I read the comment three more times and even dug through the git
> history. It seems like what you're saying is that, under certain
> conditions (which arguably would be bugs in the core Linux timing
> code),
I don't see that as a bug. Its just a side effect of reading two
different clocks (one is CLOCK_MONOTONIC and the other is TSC),
and using those two clocks to as a "base + offset".
As the comment explains, if you do that, can't guarantee monotonicity.
> actually calling ktime_get_boot_ns() could be non-monotonic
> with respect to the kvmclock timing. But get_kvmclock_ns() isn't used
> for VM timing as such -- it's used for the IOCTL interfaces for
> updating the time offset. So can you explain how my patch is
> incorrect?
ktime_get_boot_ns() has frequency correction applied, while
reading masterclock + TSC offset does not.
So the clock reads differ.
On Tue, Oct 9, 2018 at 8:28 AM Marcelo Tosatti <[email protected]> wrote:
>
> On Mon, Oct 08, 2018 at 10:38:22AM -0700, Andy Lutomirski wrote:
> > On Mon, Oct 8, 2018 at 8:27 AM Marcelo Tosatti <[email protected]> wrote:
> > I read the comment three more times and even dug through the git
> > history. It seems like what you're saying is that, under certain
> > conditions (which arguably would be bugs in the core Linux timing
> > code),
>
> I don't see that as a bug. Its just a side effect of reading two
> different clocks (one is CLOCK_MONOTONIC and the other is TSC),
> and using those two clocks to as a "base + offset".
>
> As the comment explains, if you do that, can't guarantee monotonicity.
>
> > actually calling ktime_get_boot_ns() could be non-monotonic
> > with respect to the kvmclock timing. But get_kvmclock_ns() isn't used
> > for VM timing as such -- it's used for the IOCTL interfaces for
> > updating the time offset. So can you explain how my patch is
> > incorrect?
>
> ktime_get_boot_ns() has frequency correction applied, while
> reading masterclock + TSC offset does not.
>
> So the clock reads differ.
>
Ah, okay, I finally think I see what's going on. In the kvmclock data
exposed to the guest, tsc_shift and tsc_to_system_mul come from
tgt_tsc_khz, whereas master_kernel_ns and master_cycle_now come from
CLOCK_BOOTTIME. So the kvmclock and kernel clock drift apart at a
rate given by the frequency shift and then suddenly agree again every
time the pvclock data is updated.
Is there a reason to do it this way?
On Tue, Oct 09, 2018 at 01:09:42PM -0700, Andy Lutomirski wrote:
> On Tue, Oct 9, 2018 at 8:28 AM Marcelo Tosatti <[email protected]> wrote:
> >
> > On Mon, Oct 08, 2018 at 10:38:22AM -0700, Andy Lutomirski wrote:
> > > On Mon, Oct 8, 2018 at 8:27 AM Marcelo Tosatti <[email protected]> wrote:
>
> > > I read the comment three more times and even dug through the git
> > > history. It seems like what you're saying is that, under certain
> > > conditions (which arguably would be bugs in the core Linux timing
> > > code),
> >
> > I don't see that as a bug. Its just a side effect of reading two
> > different clocks (one is CLOCK_MONOTONIC and the other is TSC),
> > and using those two clocks to as a "base + offset".
> >
> > As the comment explains, if you do that, can't guarantee monotonicity.
> >
> > > actually calling ktime_get_boot_ns() could be non-monotonic
> > > with respect to the kvmclock timing. But get_kvmclock_ns() isn't used
> > > for VM timing as such -- it's used for the IOCTL interfaces for
> > > updating the time offset. So can you explain how my patch is
> > > incorrect?
> >
> > ktime_get_boot_ns() has frequency correction applied, while
> > reading masterclock + TSC offset does not.
> >
> > So the clock reads differ.
> >
>
> Ah, okay, I finally think I see what's going on. In the kvmclock data
> exposed to the guest, tsc_shift and tsc_to_system_mul come from
> tgt_tsc_khz, whereas master_kernel_ns and master_cycle_now come from
> CLOCK_BOOTTIME. So the kvmclock and kernel clock drift apart at a
> rate given by the frequency shift and then suddenly agree again every
> time the pvclock data is updated.
Yes.
> Is there a reason to do it this way?
Since pvclock updates which update system_timestamp are expensive (must stop all vcpus),
they should be avoided.
So only HW TSC counts, and used as offset against vcpu's tsc_timestamp.
On Thu, Oct 11, 2018 at 3:28 PM Marcelo Tosatti <[email protected]> wrote:
>
> On Tue, Oct 09, 2018 at 01:09:42PM -0700, Andy Lutomirski wrote:
> > On Tue, Oct 9, 2018 at 8:28 AM Marcelo Tosatti <[email protected]> wrote:
> > >
> > > On Mon, Oct 08, 2018 at 10:38:22AM -0700, Andy Lutomirski wrote:
> > > > On Mon, Oct 8, 2018 at 8:27 AM Marcelo Tosatti <[email protected]> wrote:
> >
> > > > I read the comment three more times and even dug through the git
> > > > history. It seems like what you're saying is that, under certain
> > > > conditions (which arguably would be bugs in the core Linux timing
> > > > code),
> > >
> > > I don't see that as a bug. Its just a side effect of reading two
> > > different clocks (one is CLOCK_MONOTONIC and the other is TSC),
> > > and using those two clocks to as a "base + offset".
> > >
> > > As the comment explains, if you do that, can't guarantee monotonicity.
> > >
> > > > actually calling ktime_get_boot_ns() could be non-monotonic
> > > > with respect to the kvmclock timing. But get_kvmclock_ns() isn't used
> > > > for VM timing as such -- it's used for the IOCTL interfaces for
> > > > updating the time offset. So can you explain how my patch is
> > > > incorrect?
> > >
> > > ktime_get_boot_ns() has frequency correction applied, while
> > > reading masterclock + TSC offset does not.
> > >
> > > So the clock reads differ.
> > >
> >
> > Ah, okay, I finally think I see what's going on. In the kvmclock data
> > exposed to the guest, tsc_shift and tsc_to_system_mul come from
> > tgt_tsc_khz, whereas master_kernel_ns and master_cycle_now come from
> > CLOCK_BOOTTIME. So the kvmclock and kernel clock drift apart at a
> > rate given by the frequency shift and then suddenly agree again every
> > time the pvclock data is updated.
>
> Yes.
>
> > Is there a reason to do it this way?
>
> Since pvclock updates which update system_timestamp are expensive (must stop all vcpus),
> they should be avoided.
>
Fair enough.
> So only HW TSC counts
makes sense.
>, and used as offset against vcpu's tsc_timestamp.
>
Why don't you just expose CLOCK_MONTONIC_RAW or CLOCK_MONOTONIC_RAW
plus suspend time, though? Then you would actually be tracking a real
kernel timekeeping mode, and you wouldn't need all this complicated
offsetting work to avoid accidentally going backwards.
On Thu, Oct 11, 2018 at 04:00:29PM -0700, Andy Lutomirski wrote:
> On Thu, Oct 11, 2018 at 3:28 PM Marcelo Tosatti <[email protected]> wrote:
> >
> > On Tue, Oct 09, 2018 at 01:09:42PM -0700, Andy Lutomirski wrote:
> > > On Tue, Oct 9, 2018 at 8:28 AM Marcelo Tosatti <[email protected]> wrote:
> > > >
> > > > On Mon, Oct 08, 2018 at 10:38:22AM -0700, Andy Lutomirski wrote:
> > > > > On Mon, Oct 8, 2018 at 8:27 AM Marcelo Tosatti <[email protected]> wrote:
> > >
> > > > > I read the comment three more times and even dug through the git
> > > > > history. It seems like what you're saying is that, under certain
> > > > > conditions (which arguably would be bugs in the core Linux timing
> > > > > code),
> > > >
> > > > I don't see that as a bug. Its just a side effect of reading two
> > > > different clocks (one is CLOCK_MONOTONIC and the other is TSC),
> > > > and using those two clocks to as a "base + offset".
> > > >
> > > > As the comment explains, if you do that, can't guarantee monotonicity.
> > > >
> > > > > actually calling ktime_get_boot_ns() could be non-monotonic
> > > > > with respect to the kvmclock timing. But get_kvmclock_ns() isn't used
> > > > > for VM timing as such -- it's used for the IOCTL interfaces for
> > > > > updating the time offset. So can you explain how my patch is
> > > > > incorrect?
> > > >
> > > > ktime_get_boot_ns() has frequency correction applied, while
> > > > reading masterclock + TSC offset does not.
> > > >
> > > > So the clock reads differ.
> > > >
> > >
> > > Ah, okay, I finally think I see what's going on. In the kvmclock data
> > > exposed to the guest, tsc_shift and tsc_to_system_mul come from
> > > tgt_tsc_khz, whereas master_kernel_ns and master_cycle_now come from
> > > CLOCK_BOOTTIME. So the kvmclock and kernel clock drift apart at a
> > > rate given by the frequency shift and then suddenly agree again every
> > > time the pvclock data is updated.
> >
> > Yes.
> >
> > > Is there a reason to do it this way?
> >
> > Since pvclock updates which update system_timestamp are expensive (must stop all vcpus),
> > they should be avoided.
> >
>
> Fair enough.
>
> > So only HW TSC counts
>
> makes sense.
>
> >, and used as offset against vcpu's tsc_timestamp.
> >
>
> Why don't you just expose CLOCK_MONTONIC_RAW or CLOCK_MONOTONIC_RAW
> plus suspend time, though? Then you would actually be tracking a real
> kernel timekeeping mode, and you wouldn't need all this complicated
> offsetting work to avoid accidentally going backwards.
Can you outline how that would work ?