On Mon, Feb 10, 2014 at 03:42:45PM -0700, Eric Blake wrote: > On 02/10/2014 03:10 PM, Marcelo Tosatti wrote: >> >> Since QEMU commit "kvmclock: clock should count only if vm is running" >> (http://lists.gnu.org/archive/html/qemu-devel/2013-06/msg01225.html), >> guests have their realtime clock stopped during pause. >> >> To correct the situation, invoke guest agent to sync time from >> host time. > > This needs to be conditional on new API (either a flag to existing API > [except that virDomainResumeFlags is not quite existing yet], or new API > altogether). > > We cannot require an API to depend on guest interaction (which an agent > command does) without an explicit flag mentioning that it is okay. We > may even decide to have conditional ACL checks (allowing some users the > ability to restart a domain, but not to interact with the guest agent). > > For that matter, when we proposed adding virDomainResumeFlags as the way > to add in a flag for requesting a sync via this API, Dan suggested it > would be better to add a new API altogether instead of piggybacking on > top of the existing API: > > https://www.redhat.com/archives/libvir-list/2014-February/msg00104.html "So there's interesting, and I'm not sure I entirely agree about adding flags here. It seems to me that if the QEMU agent has a "set-time" capability we'd be better off having an explicit API virDomainSetTime(...). The action of resume + set time cannot be atomic so I don't see any point in overloading the "set time" functionality into the resume API call. Just let apps call the set time method if they so desire." Apps desire the guest time to be synchronized to host time, which is the standard 99% of usecases configuration. Ok, i was thinking of adding a new element GuestSyncTimeOnResume to control this, which would default to on (as users expect their guests time to be correct at all times, without additional commands, and rightly so). It seems unreasonable to me that every API user has to keep track of every location which a guest can possibly resume execution from a paused state: this can be hidden inside libvirt. Daniel, all the logic behind the necessity to set-time after guest resume (*) can be hidden inside libvirt. (*) all instances of guest resume: upon migration continuation, upon ENOSPACE pause continuation, all of them. Don't see the need to force the knowledge and maintenance of this state on libvirt users. Are you OK with the new knob, default on, to control sync time on guest resume, then?

> I think we may need both approaches. I too think that resume with > syncing guest time is so common use case that is deserves to be > exposed as a single operation outside the libvirt. > > On the other hand, we certainly want to expose this as a new > virDomain{Get,Set}Time() API. Why a new API which requires modifications to every libvirt user (read: current libvirt users are broken until patched) ?

Note it requires applications to keep track of every pause/resume instance (see item 1 above).

On 02/11/2014 02:37 PM, Marcelo Tosatti wrote: >> I think we may need both approaches. I too think that resume with >> syncing guest time is so common use case that is deserves to be >> exposed as a single operation outside the libvirt. >> >> On the other hand, we certainly want to expose this as a new >> virDomain{Get,Set}Time() API. > > Why a new API which requires modifications to every libvirt user (read: > current libvirt users are broken until patched) ? ALL existing libvirt users are "broken", in that NO ONE has been able to automatically resync guest time yet (at least, using supported libvirt API). This is a new feature; but as a new feature that involves guest cooperation, it must be explicitly requested by any client that wants to take advantage of it. We cannot change the default and automatically turn this feature on, because it has back-compat implications to existing libvirt clients that are not expecting libvirt to automatically try and munge guest time. > > Note it requires applications to keep track of every pause/resume > instance (see item 1 above). Whether we make it a new API call that all clients must call every time they want, or a single knob that says that for the given VM, libvirt should automatically attempt to resync time after any resume action, isn't as important to me as the fact that it IS something new that gets explicitly requested. We already have <on_poweroff> and friends in the guest XML that control what the guest does on a state transition to poweroff, maybe we could add an <on_resume> element that has instructions on what actions libvirt should take when a guest transitions from paused back to running. But the point is that such an XML element is an explicit request, and only appropriate for guests where the management stack trusts interacting with the guest agent. Maybe even both additions would be appropriate (a knob for automatic libvirt syncing as well as a new API for explicit syncing outside the events where libvirt would normally do it automatically).

On 02/13/2014 02:09 AM, Michal Privoznik wrote: >> Maybe even both additions would be appropriate (a knob for automatic >> libvirt syncing as well as a new API for explicit syncing outside the >> events where libvirt would normally do it automatically). > > I don't think we should add the <on_resume/> element. After all, it > would be the same as: > > virDomainResumeFlags(dom, VIR_SYNC_TIME) > > which as Dan correctly argues, makes sensible error reporting > impossible. If the domain is resumed, libvirt should execute the > <on_resume/> action, which may fail. Or maybe we document that the virDomainResume() api tracks just the resume state, then we add a new libvirt event that says the result of any attempt at a sync operation. Then when you resume, and <on_resume> says to attempt an auto-sync, then you can also listen for the libvirt event that tells you the status of the sync (whether it succeeded (and to what timestamp) or failed).

On 11.02.2014 01:04, Marcelo Tosatti wrote: >On Mon, Feb 10, 2014 at 03:42:45PM -0700, Eric Blake wrote: >>On 02/10/2014 03:10 PM, Marcelo Tosatti wrote: >>> >>>Since QEMU commit "kvmclock: clock should count only if vm is running" >>>(http://lists.gnu.org/archive/html/qemu-devel/2013-06/msg01225.html), >>>guests have their realtime clock stopped during pause. >>> >>>To correct the situation, invoke guest agent to sync time from >>>host time. >> >>This needs to be conditional on new API (either a flag to existing API >>[except that virDomainResumeFlags is not quite existing yet], or new API >>altogether). >> >>We cannot require an API to depend on guest interaction (which an agent >>command does) without an explicit flag mentioning that it is okay. We >>may even decide to have conditional ACL checks (allowing some users the >>ability to restart a domain, but not to interact with the guest agent). >> >>For that matter, when we proposed adding virDomainResumeFlags as the way >>to add in a flag for requesting a sync via this API, Dan suggested it >>would be better to add a new API altogether instead of piggybacking on >>top of the existing API: >> >>https://www.redhat.com/archives/libvir-list/2014-February/msg00104.html > >"So there's interesting, and I'm not sure I entirely agree about >adding flags here. It seems to me that if the QEMU agent has a >"set-time" capability we'd be better off having an explicit API >virDomainSetTime(...). The action of resume + set time cannot >be atomic so I don't see any point in overloading the "set time" >functionality into the resume API call. Just let apps call the >set time method if they so desire." > >Apps desire the guest time to be synchronized to host time, which is the >standard 99% of usecases configuration. > >Ok, i was thinking of adding a new element GuestSyncTimeOnResume to >control this, which would default to on (as users expect their guests >time to be correct at all times, without additional commands, and >rightly so). > >It seems unreasonable to me that every API user has to keep track of >every location which a guest can possibly resume execution from a paused >state: this can be hidden inside libvirt. > >Daniel, all the logic behind the necessity to set-time after guest >resume (*) can be hidden inside libvirt. > >(*) all instances of guest resume: upon migration continuation, upon >ENOSPACE pause continuation, all of them. > >Don't see the need to force the knowledge and maintenance of this state >on libvirt users. > >Are you OK with the new knob, default on, to control sync time on >guest resume, then? > I think we may need both approaches. I too think that resume with syncing guest time is so common use case that is deserves to be exposed as a single operation outside the libvirt. On the other hand, we certainly want to expose this as a new virDomain{Get,Set}Time() API.

On Wed, Feb 12, 2014 at 10:22:11AM +0000, Daniel P. Berrange wrote: > I agree that this should be a completely separate command, not > merely a flag to the Resume API. The reason is that you cannot > do sensible error reporting if you overload this in one API > call. ie consider that resuming the CPUs succeeds, but syncing > time fails. If you return "success" to the caller you are lieing > about the result of time sync. If you return "error" to the caller > you are lieing about the result of resuming the CPUs. > > If there is a separate API to invoke then the caller can clearly > see which operation succeeded vs failed. Well then just require a new guest agent command. No need for a separate command. Going from "resume" to "resume-and-sync" versus Going from "resume" to "resume; send-guest-agent-command" Is not very different is it? In both cases qemu guest agent channel must be operational.

I'll go write a virsh alias and virt-manager patch. (BTW, error reporting would be in the libvirt logs, GuestAgentError with all the details why time-sync failed).

> > I'll go write a virsh alias and virt-manager patch.