Now CC'ing the previous commenters to the virtio-rtc spec draft, since
this discussion is mostly about the spec, and the Virtio mailing lists
still seem to be in a migration hiatus...
On 13.03.24 19:18, David Woodhouse wrote:
> On 13 March 2024 17:50:48 GMT, Peter Hilber <peter.hilber@xxxxxxxxxxxxxxx> wrote:
>> On 13.03.24 13:45, David Woodhouse wrote:
>>> Surely the whole point of this effort is to provide guests with precise
>>> and *unambiguous* knowledge of what the time is?
>>
>> I would say, a fundamental point of this effort is to enable such
>> implementations, and to detect if a device is promising to support this.
>>
>> Where we might differ is as to whether the Virtio clock *for every
>> implementation* has to be *continuously* accurate w.r.t. a time standard,
>> or whether *for some implementations* it could be enough that all guests in
>> the local system have the same, precise local notion of time, which might
>> be off from the actual time standard.
>
> That makes sense, but remember I don't just want {X, Y, Z} but *also* the error bounds of ±deltaY and ±deltaZ too.
>
> So your example just boils down to "I'm calling it UTC, and it's really precise, but we make no promises about its *accuracy*". And that's fine.
>
>> Also, cf. ptp_kvm, which AFAIU doesn't address leap seconds at all...
>
> KVM is not an exemplar of good time practices.
> Not in *any* respect :)
>
>> With your described use case the UTC_SMEARED clock should of course not be
>> used. The UTC_SMEARED clock would get a distinct name through udev, like
>> /dev/ptp_virtio_utc_smeared, so the incompatibility could at least be
>> detected.
>
> As long as it's clear to all concerned that this is fundamentally not usable as an accurate time source, and is only for the local-sync case you described, sure.
>
>>> Using UTC is bad enough, because for a UTC timestamp in the middle of a
>>> leap second the guest can't know know *which* occurrence of that leap
>>> second it is, so it might be wrong by a second. To resolve that
>>> ambiguity needs a leap indicator and/or tai_offset field.
>>
>> I agree that virtio-rtc should communicate this. The question is, what
>> exactly, and for which clock read request?
>
> Are we now conflating software architecture (and Linux in particular) with "hardware" design?
>
> To a certain extent, as long as the virtio-rtc device is designed to expose time precisely and unambiguously, it's less important if the Linux kernel *today* can use that. Although of course we should strive for that. Let's be...well, *unambiguous*, I suppose... that we've changed topics to discuss that though.
>
As Virtio is extensible (unlike hardware), my approach is to mostly specify
only what also has a PoC user and a use case.
>> As for PTP clocks:
>>
>> - It doesn't fit into the ioctl PTP_SYS_OFFSET_PRECISE2.
>>
>> - The clock_adjtime(2) tai_offset and return value could be set (if
>> upstream will accept this). Would this help? As discussed, user space
>> would need to interpret this (and currently no dynamic POSIX clock sets
>> this).
>
> Hm, maybe?
>
>
>>>> I think I can add a SHOULD requirement which vaguely refers to vCPU 0, or
>>>> boot vCPU. But the Virtio device is not necessarily hosted by a hypervisor,
>>>> so the device might not even know which vCPUs there are. E.g. there is even
>>>> interest to make virtio-rtc work as part of the virtio-net device (which
>>>> might be implemented in hardware).
>>>
>>> Sure, but those implementations aren't going to offer the TSC pairing
>>> at all, are they?
>>>
>>
>> They could offer an Intel ART pairing (some physical PTP NICs are already
>> doing this, look for the convert_art_to_tsc() users).
>
> Right, but isn't that software's problem? The time pairing is defined against the ART in that case.
My point was that such a device would then not necessarily have an idea
what vCPU 0 is. But let's just say that this will be phrased as a SHOULD
best-effort requirement anyway.
Thanks for the comments,
Peter
