Re: [Qemu-devel] [PATCH for-2.8 2/3] migration: spapr_drc: defined VMStateDescription struct

[Michael Roth]

Quoting David Gibson (2016-11-18 00:04:33)
> On Thu, Nov 17, 2016 at 07:40:26PM -0600, Michael Roth wrote:
> > From: Jianjun Duan <address@hidden>
> > 
> > To manage hotplug/unplug of dynamic resources such as PCI cards,
> > memory, and CPU on sPAPR guests, a firmware abstraction known as
> > a Dynamic Resource Connector (DRC) is used to assign a particular
> > dynamic resource to the guest, and provide an interface for the
> > guest to manage configuration/removal of the resource associated
> > with it.
> > 
> > To migrate the hotplugged resources in migration, the
> > associated DRC state need be migrated. To migrate the DRC state,
> > we defined the VMStateDescription struct for spapr_drc to enable
> > the transmission of spapr_drc state in migration.
> > 
> > Not all the elements in the DRC state are migrated. Only those
> > ones modifiable or needed by guest actions or device add/remove
> > operation are migrated. From the perspective of device
> > hotplugging, if we hotplug a device on the source, we need to
> > "coldplug" it on the target. The states across two hosts for the
> > same device are not the same. Ideally we want the states be same
> > after migration so that the device would function as hotplugged
> > on the target. For example we can unplug it. The minimum DRC
> > state we need to transfer should cover all the pieces changed by
> > hotplugging. Out of the elements of the DRC state, isolation_state,
> > allocation_sate, and configured are involved in the DR state
> > transition diagram from PAPR+ 2.7, 13.4. configured and signalled
> > are needed in attaching and detaching devices. indicator_state
> > provides users with hardware state information. These 6 elements
> > are migrated.
> > 
> > detach_cb in the DRC state is a function pointer that cannot be
> > migrated. We set it right after DRC state is migrated so that
> > a migrated hot-unplug event could finish its work.
> > 
> > The instance_id is used to identify objects in migration. We set
> > instance_id of DRC using the unique index so that it is the same
> > across migration.
> > 
> > Signed-off-by: Jianjun Duan <address@hidden>
> > * add migration for awaiting_allocation state
> > Signed-off-by: Michael Roth <address@hidden>
> > ---
> >  hw/ppc/spapr_drc.c         | 70 
> > ++++++++++++++++++++++++++++++++++++++++++++++
> >  hw/ppc/spapr_pci.c         | 22 +++++++++++++++
> >  include/hw/ppc/spapr_drc.h |  9 ++++++
> >  3 files changed, 101 insertions(+)
> > 
> > diff --git a/hw/ppc/spapr_drc.c b/hw/ppc/spapr_drc.c
> > index a0c44ee..1ec6551 100644
> > --- a/hw/ppc/spapr_drc.c
> > +++ b/hw/ppc/spapr_drc.c
> > @@ -632,6 +632,72 @@ static void spapr_dr_connector_instance_init(Object 
> > *obj)
> >                          NULL, NULL, NULL, NULL);
> >  }
> >  
> > +static bool spapr_drc_needed(void *opaque)
> > +{
> > +    sPAPRDRConnector *drc = (sPAPRDRConnector *)opaque;
> > +    sPAPRDRConnectorClass *drck = SPAPR_DR_CONNECTOR_GET_CLASS(drc);
> > +    bool rc = false;
> > +    sPAPRDREntitySense value;
> > +
> > +    drck->entity_sense(drc, &value);
> > +    /* If no dev is plugged in there is no need to migrate the DRC state */
> > +    if (value != SPAPR_DR_ENTITY_SENSE_PRESENT) {
> > +        return false;
> > +    }
> > +    /*
> > +     * If there is dev plugged in, we need to migrate the DRC state when
> > +     * it is different from cold-plugged state
> > +     */
> > +    switch(drc->type) {
> > +    /* for PCI type */
> > +    case SPAPR_DR_CONNECTOR_TYPE_PCI:
> > +        rc = !((drc->isolation_state == 
> > SPAPR_DR_ISOLATION_STATE_UNISOLATED) &&
> > +               (drc->allocation_state == SPAPR_DR_ALLOCATION_STATE_USABLE) 
> > &&
> > +               drc->configured && drc->signalled && 
> > !drc->awaiting_release);
> > +        break;
> > +    /* for LMB type */
> > +    case SPAPR_DR_CONNECTOR_TYPE_LMB:
> > +        rc = !((drc->isolation_state == SPAPR_DR_ISOLATION_STATE_ISOLATED) 
> > &&
> > +               (drc->allocation_state == 
> > SPAPR_DR_ALLOCATION_STATE_UNUSABLE) &&
> > +               drc->configured && drc->signalled && 
> > !drc->awaiting_release);
> > +        break;
> > +    default:
> > +        ;
> > +    }
> > +
> > +    return rc;
> > +}
> > +
> > +/* detach_cb needs be set since it is not migrated */
> > +static void postmigrate_set_detach_cb(sPAPRDRConnector *drc,
> > +                                      spapr_drc_detach_cb *detach_cb)
> > +{
> > +    drc->detach_cb = detach_cb;
> > +}
> > +
> > +/* return the unique drc index as instance_id for qom interfaces*/
> > +static int get_instance_id(DeviceState *dev)
> > +{
> > +    return (int)get_index(SPAPR_DR_CONNECTOR(OBJECT(dev)));
> > +}
> > +
> > +static const VMStateDescription vmstate_spapr_drc = {
> > +    .name = "spapr_drc",
> > +    .version_id = 1,
> > +    .minimum_version_id = 1,
> > +    .needed = spapr_drc_needed,
> > +    .fields  = (VMStateField []) {
> > +        VMSTATE_UINT32(isolation_state, sPAPRDRConnector),
> > +        VMSTATE_UINT32(allocation_state, sPAPRDRConnector),
> > +        VMSTATE_UINT32(indicator_state, sPAPRDRConnector),
> > +        VMSTATE_BOOL(configured, sPAPRDRConnector),
> > +        VMSTATE_BOOL(awaiting_release, sPAPRDRConnector),
> > +        VMSTATE_BOOL(awaiting_allocation, sPAPRDRConnector),
> > +        VMSTATE_BOOL(signalled, sPAPRDRConnector),
> > +        VMSTATE_END_OF_LIST()
> > +    }
> > +};
> > +
> >  static void spapr_dr_connector_class_init(ObjectClass *k, void *data)
> >  {
> >      DeviceClass *dk = DEVICE_CLASS(k);
> > @@ -640,6 +706,8 @@ static void spapr_dr_connector_class_init(ObjectClass 
> > *k, void *data)
> >      dk->reset = reset;
> >      dk->realize = realize;
> >      dk->unrealize = unrealize;
> > +    dk->vmsd = &vmstate_spapr_drc;
> > +    dk->dev_get_instance_id = get_instance_id;
> >      drck->set_isolation_state = set_isolation_state;
> >      drck->set_indicator_state = set_indicator_state;
> >      drck->set_allocation_state = set_allocation_state;
> > @@ -653,6 +721,8 @@ static void spapr_dr_connector_class_init(ObjectClass 
> > *k, void *data)
> >      drck->detach = detach;
> >      drck->release_pending = release_pending;
> >      drck->set_signalled = set_signalled;
> > +    drck->postmigrate_set_detach_cb = postmigrate_set_detach_cb;
> > +
> >      /*
> >       * Reason: it crashes FIXME find and document the real reason
> >       */
> > diff --git a/hw/ppc/spapr_pci.c b/hw/ppc/spapr_pci.c
> > index f9661b7..661f7d8 100644
> > --- a/hw/ppc/spapr_pci.c
> > +++ b/hw/ppc/spapr_pci.c
> > @@ -1638,11 +1638,33 @@ static void spapr_pci_pre_save(void *opaque)
> >      }
> >  }
> >  
> > +/*
> > + * detach_cb in the DRC state is a function pointer that cannot be
> > + * migrated. We set it right after migration so that a migrated
> > + * hot-unplug event could finish its work.
> > + */
> > +static void spapr_pci_set_detach_cb(PCIBus *bus, PCIDevice *pdev,
> > +                                 void *opaque)
> > +{
> > +    sPAPRPHBState *sphb = opaque;
> > +    sPAPRDRConnector *drc = spapr_phb_get_pci_drc(sphb, pdev);
> > +    sPAPRDRConnectorClass *drck = SPAPR_DR_CONNECTOR_GET_CLASS(drc);
> > +    drck->postmigrate_set_detach_cb(drc, spapr_phb_remove_pci_device_cb);
> > +}
> > +
> >  static int spapr_pci_post_load(void *opaque, int version_id)
> >  {
> >      sPAPRPHBState *sphb = opaque;
> >      gpointer key, value;
> >      int i;
> > +    PCIBus *bus = PCI_HOST_BRIDGE(sphb)->bus;
> > +    unsigned int bus_no = 0;
> > +
> > +    /* Set detach_cb for the drc unconditionally after migration */
> > +    if (bus) {
> > +        pci_for_each_device(bus, pci_bus_num(bus), spapr_pci_set_detach_cb,
> > +                            &bus_no);
> > +    }
> 
> This reconstruction of the detach_cbs is a bit clunky.  As noted
> elsewhere, I'm not going to delay the patches for that, of course.

Agreed, and I thought about pulling it in as part of this series but
was worried some of the changes in the DRC logic might be a bit risky
for hard-freeze. There are a couple other edge-cases in the DRC logic
that we don't fully handle correctly either, so I think it would be
good to introduce some unit tests for the DRC code and do the
refactoring on top of that. Will work on that for 2.9.

> 
> IIRC, this callback is always either NULL, or the detach callback that
> belongs to the particular DRC type.
> 
> So for the sake of migration, I think it would make sense to change
> the representation of the DRC to instead of including a callback
> directly, simply have a boolean flag for whether the callback is
> active at the moment.  When we need to call the callback, we check if
> it is active, and if it is look up the right function for the DRC type
> (we could use subclasses of DRC to do this).

Sub-classes would be pretty elegant... I considered doing that early on
for some of the more core logic, but there seemed to be a lot of overlap
(though the logic seems to be getting more and more type-specific as
time goes on). For the attach/detach callbacks the distinctions are much
clearer though...

Another approach would be to register all callbacks as part of the
DRC creation. Having control over the callbacks at the DRC creation
callsites might have some benefits WRT doing the unit tests, since
the DRCs remain fairly self-contained. But probably not too hard to
make it work either way.

> 
> >  
> >      for (i = 0; i < sphb->msi_devs_num; ++i) {
> >          key = g_memdup(&sphb->msi_devs[i].key,
> > diff --git a/include/hw/ppc/spapr_drc.h b/include/hw/ppc/spapr_drc.h
> > index fa531d5..17589c8 100644
> > --- a/include/hw/ppc/spapr_drc.h
> > +++ b/include/hw/ppc/spapr_drc.h
> > @@ -192,6 +192,15 @@ typedef struct sPAPRDRConnectorClass {
> >                     void *detach_cb_opaque, Error **errp);
> >      bool (*release_pending)(sPAPRDRConnector *drc);
> >      void (*set_signalled)(sPAPRDRConnector *drc);
> > +
> > +    /*
> > +     * QEMU interface for setting detach_cb after migration.
> > +     * detach_cb in the DRC state is a function pointer that cannot be
> > +     * migrated. We set it right after migration so that a migrated
> > +     * hot-unplug event could finish its work.
> > +     */
> > +    void (*postmigrate_set_detach_cb)(sPAPRDRConnector *drc,
> > +                                      spapr_drc_detach_cb *detach_cb);
> >  } sPAPRDRConnectorClass;
> >  
> >  sPAPRDRConnector *spapr_dr_connector_new(Object *owner,
> 
> -- 
> David Gibson                    | I'll have my music baroque, and my code
> david AT gibson.dropbear.id.au  | minimalist, thank you.  NOT _the_ _other_
>                                 | _way_ _around_!
> http://www.ozlabs.org/~dgibson