Re: [qemu-s390x] [PATCH v2] s390x/tod: Properly stop the KVM TOD while the guest is not running

[Thomas Huth]

On 2018-11-28 10:59, David Hildenbrand wrote:
> Just like on other architectures, we should stop the clock while the guest
> is not running. This is already properly done for TCG. Right now, doing an
> offline migration (stop, migrate, cont) can easily trigger stalls in the
> guest.
> 
> Even doing a
>     (hmp) stop
>     ... wait 2 minutes ...
>     (hmp) cont
> will already trigger stalls.
> 
> So whenever the guest stops, backup the KVM TOD. When continuing to run
> the guest, restore the KVM TOD.
> 
> One special case is starting a simple VM: Reading the TOD from KVM to
> stop it right away until the guest is actually started means that the
> time of any simple VM will already differ to the host time. We can
> simply leave the TOD running and the guest won't be able to recognize
> it.
> 
> For migration, we actually want to keep the TOD stopped until really
> starting the guest. To be able to catch most errors, we should however
> try to set the TOD in addition to simply storing it. So we can still
> catch basic migration problems.
> 
> If anything goes wrong while backing up/restoring the TOD, we have to
> ignore it (but print a warning). This is then basically a fallback to
> old behavior (TOD remains running).
> 
> I tested this very basically with an initrd:
>     1. Start a simple VM. Observed that the TOD is kept running. Old
>        behavior.
>     2. Ordinary live migration. Observed that the TOD is temporarily
>        stopped on the destination when setting the new value and
>        correctly started when finally starting the guest.
>     3. Offline live migration. (stop, migrate, cont). Observed that the
>        TOD will be stopped on the source with the "stop" command. On the
>        destination, the TOD is temporarily stopped when setting the new
>        value and correctly started when finally starting the guest via
>        "cont".
>     4. Simple stop/cont correctly stops/starts the TOD. (multiple stops
>        or conts in a row have no effect, so works as expected)
> 
> In the future, we might want to send the guest a special kind of time sync
> interrupt under some conditions, so it can synchronize its tod to the
> host tod. This is interesting for migration scenarios but also when we
> get time sync interrupts ourselves. This however will most probably have
> to be handled in KVM (e.g. when the tods differ too much) and is not
> desired e.g. when debugging the guest. (single stepping should not
> result in permanent time syncs). I consider something like that an add-on
> on top of this basic "don't break the guest" handling.
> 
> Signed-off-by: David Hildenbrand <address@hidden>
> ---
> 
> v1 -> v2:
> - Add time sync idea suggested by Christian to description
> - Drop one unnecessary "return"
> - Register in realize() and not in instance_init()
> 
> 
>  hw/s390x/tod-kvm.c     | 91 +++++++++++++++++++++++++++++++++++++++++-
>  hw/s390x/tod.c         |  5 +++
>  include/hw/s390x/tod.h |  8 +++-
>  3 files changed, 101 insertions(+), 3 deletions(-)
> 
> diff --git a/hw/s390x/tod-kvm.c b/hw/s390x/tod-kvm.c
> index df564ab89c..87717d0be1 100644
> --- a/hw/s390x/tod-kvm.c
> +++ b/hw/s390x/tod-kvm.c
> @@ -10,10 +10,11 @@
>  
>  #include "qemu/osdep.h"
>  #include "qapi/error.h"
> +#include "sysemu/sysemu.h"
>  #include "hw/s390x/tod.h"
>  #include "kvm_s390x.h"
>  
> -static void kvm_s390_tod_get(const S390TODState *td, S390TOD *tod, Error 
> **errp)
> +static void kvm_s390_get_tod_raw(S390TOD *tod, Error **errp)
>  {
>      int r;
>  
> @@ -27,7 +28,17 @@ static void kvm_s390_tod_get(const S390TODState *td, 
> S390TOD *tod, Error **errp)
>      }
>  }
>  
> -static void kvm_s390_tod_set(S390TODState *td, const S390TOD *tod, Error 
> **errp)
> +static void kvm_s390_tod_get(const S390TODState *td, S390TOD *tod, Error 
> **errp)
> +{
> +    if (td->stopped) {
> +        *tod = td->base;
> +        return;
> +    }
> +
> +    kvm_s390_get_tod_raw(tod, errp);
> +}
> +
> +static void kvm_s390_set_tod_raw(const S390TOD *tod, Error **errp)
>  {
>      int r;
>  
> @@ -41,18 +52,94 @@ static void kvm_s390_tod_set(S390TODState *td, const 
> S390TOD *tod, Error **errp)
>      }
>  }
>  
> +static void kvm_s390_tod_set(S390TODState *td, const S390TOD *tod, Error 
> **errp)
> +{
> +    Error *local_err = NULL;
> +
> +    /*
> +     * Somebody (e.g. migration) set the TOD. We'll store it into KVM to
> +     * properly detect errors now but take a look at the runstate to decide
> +     * whether really to keep the tod running. E.g. during migration, this
> +     * is the point where we want to stop the initially running TOD to fire
> +     * it back up when actually starting the migrated guest.
> +     */
> +    kvm_s390_set_tod_raw(tod, &local_err);
> +    if (local_err) {
> +        error_propagate(errp, local_err);
> +        return;
> +    }
> +
> +    if (runstate_is_running()) {
> +        td->stopped = false;
> +    } else {
> +        td->stopped = true;
> +        td->base = *tod;
> +    }
> +}
> +
> +static void kvm_s390_tod_vm_state_change(void *opaque, int running,
> +                                         RunState state)
> +{
> +    S390TODState *td = opaque;
> +    Error *local_err = NULL;
> +
> +    if (running && td->stopped) {
> +        /* Set the old TOD when running the VM - start the TOD clock. */
> +        kvm_s390_set_tod_raw(&td->base, &local_err);
> +        if (local_err) {
> +            warn_report_err(local_err);
> +        }
> +        /* Treat errors like the TOD was running all the time. */
> +        td->stopped = false;
> +    } else if (!running && !td->stopped) {
> +        /* Store the TOD when stopping the VM - stop the TOD clock. */
> +        kvm_s390_get_tod_raw(&td->base, &local_err);
> +        if (local_err) {
> +            /* Keep the TOD running in case we could not back it up. */
> +            warn_report_err(local_err);
> +        } else {
> +            td->stopped = true;
> +        }
> +    }
> +}
> +
> +static void kvm_s390_tod_realize(S390TODState *td, Error **errp)
> +{
> +    /*
> +     * We need to know when the VM gets started/stopped to start/stop the 
> TOD.
> +     * As we can never have more than one TOD instance (and that will never 
> be
> +     * removed), registering here and never unregistering is good enough.
> +     */
> +    qemu_add_vm_change_state_handler(kvm_s390_tod_vm_state_change, td);
> +}
> +
>  static void kvm_s390_tod_class_init(ObjectClass *oc, void *data)
>  {
>      S390TODClass *tdc = S390_TOD_CLASS(oc);
>  
> +    tdc->realize = kvm_s390_tod_realize;
>      tdc->get = kvm_s390_tod_get;
>      tdc->set = kvm_s390_tod_set;
>  }
>  
> +static void kvm_s390_tod_init(Object *obj)
> +{
> +    S390TODState *td = S390_TOD(obj);
> +
> +    /*
> +     * The TOD is initially running (value stored in KVM). Avoid needless
> +     * loading/storing of the TOD when starting a simple VM, so let it
> +     * run although the (never started) VM is stopped. For migration, we
> +     * will properly set the TOD later.
> +     */
> +    td->stopped = false;
> +}
> +
>  static TypeInfo kvm_s390_tod_info = {
>      .name = TYPE_KVM_S390_TOD,
>      .parent = TYPE_S390_TOD,
>      .instance_size = sizeof(S390TODState),
> +    .instance_init = kvm_s390_tod_init,
>      .class_init = kvm_s390_tod_class_init,
>      .class_size = sizeof(S390TODClass),
>  };
> diff --git a/hw/s390x/tod.c b/hw/s390x/tod.c
> index 1c63f411e6..82ea6554ba 100644
> --- a/hw/s390x/tod.c
> +++ b/hw/s390x/tod.c
> @@ -97,9 +97,14 @@ static SaveVMHandlers savevm_tod = {
>  static void s390_tod_realize(DeviceState *dev, Error **errp)
>  {
>      S390TODState *td = S390_TOD(dev);
> +    S390TODClass *tdc = S390_TOD_GET_CLASS(td);
>  
>      /* Legacy migration interface */
>      register_savevm_live(NULL, "todclock", 0, 1, &savevm_tod, td);
> +
> +    if (tdc->realize) {
> +        tdc->realize(td, errp);
> +    }
>  }

I think the more usual way to deal with this, is to use
device_class_set_parent_realize() in the child's class init function,
and then to call ->parent_realize() from the child's realize function.

I guess it doesn't really matter in the end which way we do it here, but
it would be nice to be consistent with the other devices, so I'd prefer
to use device_class_set_parent_realize() for this here, too.

Apart from that, the patch looks fine to me.

 Thomas