On 11/11/15 13:23, Eduardo Habkost wrote:
> On Wed, Nov 11, 2015 at 12:57:44AM +0800, Haozhong Zhang wrote:
> > On 11/09/15 14:01, Eduardo Habkost wrote:
> > > On Mon, Nov 09, 2015 at 08:33:55AM +0800, haozhong.zhang@xxxxxxxxx wrote:
> > > > On 11/06/15 13:12, Eduardo Habkost wrote:
> > > > > On Fri, Nov 06, 2015 at 10:32:24AM +0800, haozhong.zhang@xxxxxxxxx wrote:
> > > > > > On 11/05/15 14:05, Eduardo Habkost wrote:
> > > > > > > On Thu, Nov 05, 2015 at 09:30:51AM +0800, Haozhong Zhang wrote:
> > > > > > > > On 11/04/15 19:42, Eduardo Habkost wrote:
> > > > > [...]
> > > > > > > > > > + env->tsc_khz_saved = r;
> > > > > > > > > > + }
> > > > > > > > >
> > > > > > > > > Why do you need a separate tsc_khz_saved field, and don't simply use
> > > > > > > > > tsc_khz? It would have the additional feature of letting QMP clients
> > > > > > > > > query the current TSC rate by asking for the tsc-freq property on CPU
> > > > > > > > > objects.
> > > > > > > > >
> > > > > > > >
> > > > > > > > It's to avoid overriding env->tsc_khz on the destination in the
> > > > > > > > migration. I can change this line to
> > > > > > > > env->tsc_khz = env->tsc_khz_saved = r;
> > > > > > >
> > > > > > > You are already avoiding overriding env->tsc_khz, because you use
> > > > > > > KVM_GET_TSC_KHZ only if tsc_khz is not set yet. I still don't see why
> > > > > > > you need a tsc_khz_saved field that requires duplicating the SET_TSC_KHZ
> > > > > > > code, if you could just do this:
> > > > > > >
> > > > > > > if (!env->tsc_khz) {
> > > > > > > env->tsc_khz = kvm_vcpu_ioctl(cs, KVM_GET_TSC_KHZ);
> > > > > > > }
> > > > > > >
> > > > > >
> > > > > > Consider an example that we migrate a VM from machine A to machine B
> > > > > > and then to machine C, and QEMU on machine B is launched with the cpu
> > > > > > option 'tsc-freq' (i.e. env->tsc_khz on B is non-zero at the
> > > > > > beginning):
> > > > > > 1) In the migration from B to C, the user-specified TSC frequency by
> > > > > > 'tsc-freq' on B is expected to be migrated to C. That is, the
> > > > > > value of env->tsc_khz on B is migrated.
> > > > > > 2) If TSC frequency is migrated through env->tsc_khz, then
> > > > > > env->tsc_khz on B will be overrode in the migration from A to B
> > > > > > before kvm_arch_setup_tsc_khz(). If the guest TSC frequency is
> > > > > > different than the user-specified TSC frequency on B, the
> > > > > > expectation in 1) will not be satisfied anymore.
> > > > >
> > > > > Setting tsc-freq on B when tsc-freq was not used on A is invalid usage.
> > > > > This is not different from changing the CPU model and adding or removing
> > > > > CPU flags when migrating, which is also incorrect. The command-line
> > > > > parameters defining the VM must be the same when you migrate.
> > > > >
> > > >
> > > > Good to know it's an invalid usage. Then the question is what QEMU is
> > > > expected to do for this invalid usage?
> > > >
> > > > 1) Abort the migration? But I find that the current QEMU does not
> > > > abort the migration between different CPU models (e.g. Nehalem and
> > > > Haswell).
> > > >
> > > > 2) Or do not abort the migration and ignore tsc-freq option? If so,
> > > > tsc_khz_saved will be not needed.
> > >
> > > My first choice is to abort migration. If we decide to abort today and
> > > find it to cause problems, we can easily fix it. If we decide to
> > > continue without aborting, it is difficult to change that behavior
> > > without breaking existing setups.
> > >
> >
> > Two additional questions:
> >
> > 1) Existing QEMU allows 'tsc-freq' on the destination in the
> > migration. If we decided to abort when both 'tsc-freq' and
> > migrated TSC were present on the destination, it would break some
> > existing usages. Considering backward compatibility, would above
> > choice 2) be better?
>
> We shouldn't abort simply because the section is present and tsc-freq is
> set (because we will always send the section in the newer
> machine-types). We should abort only when we know that the command-line
> contradicts what we see in the migration stream.
>
Got it, only abort when there are contradicts.
> >
> > 2) If we do decide to abort, could I use abort()? Or are there other
> > clean approaches to abort?
>
> You don't need to abort QEMU. You just need to tell the migration code
> that migration can't continue. The exact way to do it depends on where
> you are hooking the sanity check code. If you use a
> VMStateDescription.post_load hook, you can use error_report() and return
> a negative errno value. CCing Quintela in case he has suggestions.
>
Yes, I really should put the sanity check in cpu_post_load().
Thanks,
Haozhong
> --
> Eduardo
> --
> To unsubscribe from this list: send the line "unsubscribe kvm" in
> the body of a message to majordomo@xxxxxxxxxxxxxxx
> More majordomo info at http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
