On 02/09/2016 16:51, Roman Kagan wrote:
> On Fri, Sep 02, 2016 at 04:09:42PM +0200, Paolo Bonzini wrote:
>> On 02/09/2016 15:52, Roman Kagan wrote:
>> vs. using a single offset as in the TSC ref page is one nanosecond---and
>> the ref page only has a resolution of 100 ns.
>
> AFAICS it's not a matter of resolution. If one calculation flips from
> value T to T+1 at tsc1, while the other at tsc2, during the window
> between tsc1 and tsc2 we can have monotonicity violation.
Ok, tried "empirically" (throw numbers in a spreadsheet :)) and indeed
the maximum error is not 1 ns but 100 ns (1 unit in the time reference
count MSR).
You can get a flip between T/T+1 because the time reference counter may
be more precise with its rounding due to the separation between
tsc_timestamp and system_time. This separation provides some extra
decimal digits to the offset, which the TSC page lacks. For example:
51256391 tsc_timestamp
3311474323 tsc_to_system_mul
254246 system_time
-1 shift
-195054.1816 offset (computed exactly)
So the flip happens when the nanoseconds are around 81/82:
tsc kvmclock refcount TSC page
51256391 254246 2542 2542
51256483 254281 2542 2542
51256484 254281 2542 2543
51256486 254282 2542 2543
51256746 254382 2543 2544
I'll change patch 4 to store the parameters and use them when accessing
the time reference counter MSR. I'll still keep the procedure that goes
through kvmclock. It's a bit more involved for the scale, but
vcpu->last_guest_tsc only provides a part of the offset computation; the
other half is vcpu->hv_clock.system_time and it's not stored anywhere.
Paolo
