Re: [PATCH v3 10/16] hw/i386/vmport: Add support for CMD_GETTIME

["Michael S. Tsirkin" <mst@redhat.com>]

On Sat, Mar 14, 2020 at 09:17:30PM +0200, Liran Alon wrote:
> 
> On 14/03/2020 21:14, Michael S. Tsirkin wrote:
> > On Sat, Mar 14, 2020 at 09:04:30PM +0200, Liran Alon wrote:
> > > On 14/03/2020 20:18, Michael S. Tsirkin wrote:
> > > > On Fri, Mar 13, 2020 at 06:26:54PM +0200, Liran Alon wrote:
> > > > > On 13/03/2020 17:47, Michael S. Tsirkin wrote:
> > > > > > On Fri, Mar 13, 2020 at 05:25:20PM +0200, Liran Alon wrote:
> > > > > > > > > @@ -168,6 +169,20 @@ static uint32_t vmport_cmd_ram_size(void *opaque, uint32_t addr)
> > > > > > > > >          return ram_size;
> > > > > > > > >      }
> > > > > > > > > +static uint32_t vmport_cmd_time(void *opaque, uint32_t addr)
> > > > > > > > > +{
> > > > > > > > > +    X86CPU *cpu = X86_CPU(current_cpu);
> > > > > > > > > +    qemu_timeval tv;
> > > > > > > > > +
> > > > > > > > > +    if (qemu_gettimeofday(&tv) < 0) {
> > > > > > > > > +        return UINT32_MAX;
> > > > > > > > > +    }
> > > > > > > > > +
> > > > > > > > > +    cpu->env.regs[R_EBX] = (uint32_t)tv.tv_usec;
> > > > > > > > > +    cpu->env.regs[R_ECX] = port_state->max_time_lag_us;
> > > > > > > > > +    return (uint32_t)tv.tv_sec;
> > > > > > > > > +}
> > > > > > > > > +
> > > > > > > > >      /* vmmouse helpers */
> > > > > > > > >      void vmmouse_get_data(uint32_t *data)
> > > > > > > > >      {
> > > > > > > > That's a very weird thing to return to the guest.
> > > > > > > > For example it's not monotonic across migrations.
> > > > > > > That's the VMware PV interface... I didn't design it. :P
> > > > > > > Regarding how it handles the fact time is not monotonic across migrations,
> > > > > > > see big comment at the start of services/plugins/timeSync/timeSync.c in
> > > > > > > open-vm-tools regarding the time-sync algorithm used by VMware Tools.
> > > > > > > Specifically:
> > > > > > > """
> > > > > > > During normal operation this plugin only steps the time forward and only if
> > > > > > > the error is greater than one second.
> > > > > > Looks like guest assumes this time only moves forward.
> > > > > > So something needs to be done to avoid it moving
> > > > > > backward across migrations.
> > > > > Where do you see this assumption in guest code? I don't think this is true.
> > > > > Guest code seems to handle this by making sure to only step the time
> > > > > forward.
> > > > Exactly. So if host time moved backward e.g. by 100s, then for 100s
> > > > time is not correcting. Which possibly vmware has a way to mitigate
> > > > against e.g. by synchronising host time using their
> > > > management app.
> > > > 
> > > > > Read carefully services/plugins/timeSync/timeSync.c and point me to what I'm
> > > > > missing if you think otherwise (i.e. I missed something).
> > > > I'm just going by what you write in a patch.
> > > > 
> > > So guest doesn't assume that this time only moves forward...
> > > 
> > > Can you clarify then which change do you suggest making to this patch in
> > > this regard? It seems correct to me.
> > > i.e. The CMD_GETTIME implementation seems correct to me and it doesn't need
> > > to do anything special to handle host time moving backwards.
> > > 
> > > -Liran
> > > 
> > I think something needs to be done to make sure host time as reported by
> > this command does not move backwards significantly. Just forwarding
> > gettimeofday won't cut it IMHO. Look at kvm clock for inspiration of
> > things to do.
> > 
> It isn't required by the command interface definition.
> It's explicitly specified in timeSync.c that guest code handles the case
> host time goes backwards.

According to what you wrote it does not crash but also does not do
updates. That will work well only if the time difference isn't large.
Then with time as guest gets interrupted by host, the time difference
shrinks and eventually you are resyncing again.  And I'm guessing there
are hypervisor management interfaces in place to make sure that is so.
Linux is not guaranteed to have such interfaces so you have to come
up with QEMU specific ones.


> We are not inventing the interface, we just implement it according to it's
> definition.
> 
> -Liran

Host time is a vague enough terminology. I suspect this implementation
is too simplistic to cover all cases.

-- 
MST