Re: [PATCH 2/2] x86, kvm: use kvmclock to compute TSC deadline value

From: Radim KrÄmÃÅ
Date: Fri Sep 16 2016 - 11:25:02 EST

2016-09-16 17:06+0200, Paolo Bonzini:
> On 16/09/2016 16:59, Radim KrÄmÃÅ wrote:
>> KVM_MSR_DEADLINE would be interface in kvmclock nanosecond values and
>> MSR_IA32_TSCDEADLINE in TSC values. KVM_MSR_DEADLINE would follow
>> similar rules as MSR_IA32_TSCDEADLINE -- the interrupt fires when
>> kvmclock reaches the value, you read what you write, and 0 disarms it.
>> If the TSC deadline timer was enabled, then the guest could write to
>> both MSR_IA32_TSCDEADLINE and KVM_MSR_DEADLINE, but only one could be
>> armed at any time (non-zero write to one will set the other to 0).
>> The dual interface would allow unconditinal addition of the PV feature
>> without regressing users that currently use MSR_IA32_TSCDEADLINE and
>> adapted their stack to handle KVM's TSC shortcomings ...
> So far so good. My question is: what happens if you write to
> KVM_MSR_DEADLINE and read from MSR_IA32_TSCDEADLINE, or vice versa?

(The second paragraph covered it ;])

> The possibilities are:
> a) you read a 0

This one.

> b) you read the value converted to the other unit

Too much hassle. :)

> c) you read another value such as -1

Having common "disarmed" value is nicer and MSR_IA32_TSCDEADLINE has 0.

> (a) and (c) are the simplest of course. (c) may make sense when writing
> to MSR_IA32_TSCDEADLINE and reading from KVM_MSR_DEADLINE, since we can
> decide which values are valid or not; -1 is technically a valid TSC
> deadline.
> I'm not sure about whether to allow (b). In the end KVM is going to
> convert a nsec deadline to a TSC value internally, and vice versa.

It is not necessary to convert nsec deadline to guest-TSC, only to
host-TSC in case the VMX_PREEPTION_TIMER is used.
I would only have the host-TSC internal representation, which is not
exportable to the guest or migratable.

> On
> the other hand, if we do, userspace needs to figure out (on migration)
> whether the guest set up a TSC or a nanosecond deadline.

Yeah, I think the solution described below (writing 0 doesn't disarm the
other one) is not bad.

>>> this lets userspace decide whether to set a nsec-based
>>> deadline or a TSC-based deadline after migration.
>> Hm, isn't switching to TSC-based deadline after migration pointless?
> Yes, but I didn't mean that. I meant preserving which MSR was written
> to arm the timer, and redoing the same on the destination.

Ah, I see. Both MSRs read what deadline written to them (if they are
armed) and at most one can be non-zero.
KVM will add MSR_IA32_TSCDEADLINE to the list of emulated MSRs, so
userspace will save/restore both deadline MSRs and zero writes will not
disarm the other timer, so the correct timer will be armed.

No special logic to try to avoid TSC-related bugs.

>>>>> This still wouldn't handle old hosts of course.
>>>> The question is whether we want to carry around 150 LOC because of old
>>>> hosts. I'd just fix Linux to avoid deadline TSC without invariant TSC.
>>>> :)
>>> Yes, that would automatically blacklist it on KVM. You'd also need to
>>> update the recent optimization to the TSC deadline timer, to also work
>>> on other APIC timer modes or at least in your new PV mode.
>> All modes shouldn't be much harder than just the PV mode.
> The PV mode would still be a bit easier since it's still the TSC
> deadline timer just with a nicer interface that is not based on the TSC.
> Depends on how you code it though, I guess.

Yeah, we'll see. I am planning to carry around the deadline value in
nanoseconds (to avoid needless conversions), so it would have similar
requirements as the APIC timer.