Message ID | 1488800894-22421-4-git-send-email-eric.auger@redhat.com (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
On 6 March 2017 at 12:48, Eric Auger <eric.auger@redhat.com> wrote: > We change the restoration priority of both the GICv3 and ITS. The > GICv3 must be restored before the ITS and the ITS needs to be restored > before PCIe devices since it translates their MSI transactions. > We typically observe the virtio-pci-net device sending MSI transactions > very early (even before the first vcpu run) which looks weird. It > appears that not servicing those transactions cause the virtio-pci-net > to stall. This does seem rather weird and worth closer investigation. A stopped VM ought IMHO to be completely stopped, not still doing things... > Signed-off-by: Eric Auger <eric.auger@redhat.com> > > --- > > v1 -> v2: > - handle case where migrate_add_blocker fails > - add comments along with ITS and GICv3 migration priorities > --- > hw/intc/arm_gicv3_common.c | 1 + > hw/intc/arm_gicv3_its_common.c | 3 ++- > hw/intc/arm_gicv3_its_kvm.c | 23 +++++++++++------------ > include/migration/vmstate.h | 2 ++ > 4 files changed, 16 insertions(+), 13 deletions(-) > > diff --git a/hw/intc/arm_gicv3_common.c b/hw/intc/arm_gicv3_common.c > index c6493d6..4228b7c 100644 > --- a/hw/intc/arm_gicv3_common.c > +++ b/hw/intc/arm_gicv3_common.c > @@ -145,6 +145,7 @@ static const VMStateDescription vmstate_gicv3 = { > .minimum_version_id = 1, > .pre_save = gicv3_pre_save, > .post_load = gicv3_post_load, > + .priority = MIG_PRI_GICV3, > .fields = (VMStateField[]) { > VMSTATE_UINT32(gicd_ctlr, GICv3State), > VMSTATE_UINT32_ARRAY(gicd_statusr, GICv3State, 2), > diff --git a/hw/intc/arm_gicv3_its_common.c b/hw/intc/arm_gicv3_its_common.c > index 75b9f04..854709f 100644 > --- a/hw/intc/arm_gicv3_its_common.c > +++ b/hw/intc/arm_gicv3_its_common.c > @@ -48,7 +48,8 @@ static const VMStateDescription vmstate_its = { > .name = "arm_gicv3_its", > .pre_save = gicv3_its_pre_save, > .post_load = gicv3_its_post_load, > - .unmigratable = true, > + .unmigratable = false, unmigratable = false is the default, so we can just delete the line. > + .priority = MIG_PRI_GICV3_ITS, > .fields = (VMStateField[]) { > VMSTATE_UINT32(ctlr, GICv3ITSState), > VMSTATE_UINT64(cbaser, GICv3ITSState), > diff --git a/hw/intc/arm_gicv3_its_kvm.c b/hw/intc/arm_gicv3_its_kvm.c > index 45e57d6..3bd2873 100644 > --- a/hw/intc/arm_gicv3_its_kvm.c > +++ b/hw/intc/arm_gicv3_its_kvm.c > @@ -76,18 +76,6 @@ static void kvm_arm_its_realize(DeviceState *dev, Error **errp) > GICv3ITSState *s = ARM_GICV3_ITS_COMMON(dev); > Error *local_err = NULL; > > - /* > - * Block migration of a KVM GICv3 ITS device: the API for saving and > - * restoring the state in the kernel is not yet available > - */ > - error_setg(&s->migration_blocker, "vITS migration is not implemented"); > - migrate_add_blocker(s->migration_blocker, &local_err); > - if (local_err) { > - error_propagate(errp, local_err); > - error_free(s->migration_blocker); > - return; > - } > - > s->dev_fd = kvm_create_device(kvm_state, KVM_DEV_TYPE_ARM_VGIC_ITS, false); > if (s->dev_fd < 0) { > error_setg_errno(errp, -s->dev_fd, "error creating in-kernel ITS"); > @@ -104,6 +92,17 @@ static void kvm_arm_its_realize(DeviceState *dev, Error **errp) > > gicv3_its_init_mmio(s, NULL); > > + if (!kvm_device_check_attr(s->dev_fd, KVM_DEV_ARM_VGIC_GRP_ITS_REGS, > + GITS_CTLR)) { > + error_setg(&s->migration_blocker, "vITS migration is not implemented"); I think we should specifically say that it's the host kernel that doesn't support ITS migration (ie not QEMU that's missing support). The GICv3 uses the message "This operating system kernel does not support vGICv3 migration" which is slightly odd phrasing but I guess we should follow that. > + migrate_add_blocker(s->migration_blocker, &local_err); > + if (local_err) { > + error_propagate(errp, local_err); > + error_free(s->migration_blocker); > + return; > + } > + } > + > kvm_msi_use_devid = true; > kvm_gsi_direct_mapping = false; > kvm_msi_via_irqfd_allowed = kvm_irqfds_enabled(); > diff --git a/include/migration/vmstate.h b/include/migration/vmstate.h > index f2dbf84..8dab9c7 100644 > --- a/include/migration/vmstate.h > +++ b/include/migration/vmstate.h > @@ -198,6 +198,8 @@ enum VMStateFlags { > typedef enum { > MIG_PRI_DEFAULT = 0, > MIG_PRI_IOMMU, /* Must happen before PCI devices */ > + MIG_PRI_GICV3_ITS, /* Must happen before PCI devices */ > + MIG_PRI_GICV3, /* Must happen before the ITS */ > MIG_PRI_MAX, > } MigrationPriority; > > -- > 2.5.5 > thanks -- PMM
Hi Peter, On 13/03/2017 19:03, Peter Maydell wrote: > On 6 March 2017 at 12:48, Eric Auger <eric.auger@redhat.com> wrote: >> We change the restoration priority of both the GICv3 and ITS. The >> GICv3 must be restored before the ITS and the ITS needs to be restored >> before PCIe devices since it translates their MSI transactions. > >> We typically observe the virtio-pci-net device sending MSI transactions >> very early (even before the first vcpu run) which looks weird. It >> appears that not servicing those transactions cause the virtio-pci-net >> to stall. > > This does seem rather weird and worth closer investigation. > A stopped VM ought IMHO to be completely stopped, not still > doing things... Yes I will take time to investigate what does happen. > >> Signed-off-by: Eric Auger <eric.auger@redhat.com> >> >> --- >> >> v1 -> v2: >> - handle case where migrate_add_blocker fails >> - add comments along with ITS and GICv3 migration priorities >> --- >> hw/intc/arm_gicv3_common.c | 1 + >> hw/intc/arm_gicv3_its_common.c | 3 ++- >> hw/intc/arm_gicv3_its_kvm.c | 23 +++++++++++------------ >> include/migration/vmstate.h | 2 ++ >> 4 files changed, 16 insertions(+), 13 deletions(-) >> >> diff --git a/hw/intc/arm_gicv3_common.c b/hw/intc/arm_gicv3_common.c >> index c6493d6..4228b7c 100644 >> --- a/hw/intc/arm_gicv3_common.c >> +++ b/hw/intc/arm_gicv3_common.c >> @@ -145,6 +145,7 @@ static const VMStateDescription vmstate_gicv3 = { >> .minimum_version_id = 1, >> .pre_save = gicv3_pre_save, >> .post_load = gicv3_post_load, >> + .priority = MIG_PRI_GICV3, >> .fields = (VMStateField[]) { >> VMSTATE_UINT32(gicd_ctlr, GICv3State), >> VMSTATE_UINT32_ARRAY(gicd_statusr, GICv3State, 2), >> diff --git a/hw/intc/arm_gicv3_its_common.c b/hw/intc/arm_gicv3_its_common.c >> index 75b9f04..854709f 100644 >> --- a/hw/intc/arm_gicv3_its_common.c >> +++ b/hw/intc/arm_gicv3_its_common.c >> @@ -48,7 +48,8 @@ static const VMStateDescription vmstate_its = { >> .name = "arm_gicv3_its", >> .pre_save = gicv3_its_pre_save, >> .post_load = gicv3_its_post_load, >> - .unmigratable = true, >> + .unmigratable = false, > > unmigratable = false is the default, so we can just delete the line. OK thanks > >> + .priority = MIG_PRI_GICV3_ITS, >> .fields = (VMStateField[]) { >> VMSTATE_UINT32(ctlr, GICv3ITSState), >> VMSTATE_UINT64(cbaser, GICv3ITSState), >> diff --git a/hw/intc/arm_gicv3_its_kvm.c b/hw/intc/arm_gicv3_its_kvm.c >> index 45e57d6..3bd2873 100644 >> --- a/hw/intc/arm_gicv3_its_kvm.c >> +++ b/hw/intc/arm_gicv3_its_kvm.c >> @@ -76,18 +76,6 @@ static void kvm_arm_its_realize(DeviceState *dev, Error **errp) >> GICv3ITSState *s = ARM_GICV3_ITS_COMMON(dev); >> Error *local_err = NULL; >> >> - /* >> - * Block migration of a KVM GICv3 ITS device: the API for saving and >> - * restoring the state in the kernel is not yet available >> - */ >> - error_setg(&s->migration_blocker, "vITS migration is not implemented"); >> - migrate_add_blocker(s->migration_blocker, &local_err); >> - if (local_err) { >> - error_propagate(errp, local_err); >> - error_free(s->migration_blocker); >> - return; >> - } >> - >> s->dev_fd = kvm_create_device(kvm_state, KVM_DEV_TYPE_ARM_VGIC_ITS, false); >> if (s->dev_fd < 0) { >> error_setg_errno(errp, -s->dev_fd, "error creating in-kernel ITS"); >> @@ -104,6 +92,17 @@ static void kvm_arm_its_realize(DeviceState *dev, Error **errp) >> >> gicv3_its_init_mmio(s, NULL); >> >> + if (!kvm_device_check_attr(s->dev_fd, KVM_DEV_ARM_VGIC_GRP_ITS_REGS, >> + GITS_CTLR)) { >> + error_setg(&s->migration_blocker, "vITS migration is not implemented"); > > I think we should specifically say that it's the host kernel > that doesn't support ITS migration (ie not QEMU that's missing > support). > > The GICv3 uses the message > "This operating system kernel does not support vGICv3 migration" > which is slightly odd phrasing but I guess we should follow that. Sure I will use that message. Thank you for the review Best Regards Eric > >> + migrate_add_blocker(s->migration_blocker, &local_err); >> + if (local_err) { >> + error_propagate(errp, local_err); >> + error_free(s->migration_blocker); >> + return; >> + } >> + } >> + >> kvm_msi_use_devid = true; >> kvm_gsi_direct_mapping = false; >> kvm_msi_via_irqfd_allowed = kvm_irqfds_enabled(); >> diff --git a/include/migration/vmstate.h b/include/migration/vmstate.h >> index f2dbf84..8dab9c7 100644 >> --- a/include/migration/vmstate.h >> +++ b/include/migration/vmstate.h >> @@ -198,6 +198,8 @@ enum VMStateFlags { >> typedef enum { >> MIG_PRI_DEFAULT = 0, >> MIG_PRI_IOMMU, /* Must happen before PCI devices */ >> + MIG_PRI_GICV3_ITS, /* Must happen before PCI devices */ >> + MIG_PRI_GICV3, /* Must happen before the ITS */ >> MIG_PRI_MAX, >> } MigrationPriority; >> >> -- >> 2.5.5 >> > > thanks > -- PMM >
Hi Peter, On 13/03/2017 19:03, Peter Maydell wrote: > On 6 March 2017 at 12:48, Eric Auger <eric.auger@redhat.com> wrote: >> We change the restoration priority of both the GICv3 and ITS. The >> GICv3 must be restored before the ITS and the ITS needs to be restored >> before PCIe devices since it translates their MSI transactions. > >> We typically observe the virtio-pci-net device sending MSI transactions >> very early (even before the first vcpu run) which looks weird. It >> appears that not servicing those transactions cause the virtio-pci-net >> to stall. > > This does seem rather weird and worth closer investigation. > A stopped VM ought IMHO to be completely stopped, not still > doing things... After further investigations, here is what I observe: Between the "running" vm_change_state_handler and the 1st KVM_RUN I observe virtio_notify_config that calls msix_notify(). This eventually attempts to inject an MSI. This looks valid as the VM is in its running state. Unfortunately the VGIC gets its final init stage, map_resources(), executed on the first KVM_RUN. Until that point no MMIO access can be performed. This looks the bad part to me. So at the moment I trigger a map_resources() at the end of the ITS table flush and that works. If I don't do that, the virtio-net based network is not functional on restore path. Hope this clarifies. Thanks Eric > >> Signed-off-by: Eric Auger <eric.auger@redhat.com> >> >> --- >> >> v1 -> v2: >> - handle case where migrate_add_blocker fails >> - add comments along with ITS and GICv3 migration priorities >> --- >> hw/intc/arm_gicv3_common.c | 1 + >> hw/intc/arm_gicv3_its_common.c | 3 ++- >> hw/intc/arm_gicv3_its_kvm.c | 23 +++++++++++------------ >> include/migration/vmstate.h | 2 ++ >> 4 files changed, 16 insertions(+), 13 deletions(-) >> >> diff --git a/hw/intc/arm_gicv3_common.c b/hw/intc/arm_gicv3_common.c >> index c6493d6..4228b7c 100644 >> --- a/hw/intc/arm_gicv3_common.c >> +++ b/hw/intc/arm_gicv3_common.c >> @@ -145,6 +145,7 @@ static const VMStateDescription vmstate_gicv3 = { >> .minimum_version_id = 1, >> .pre_save = gicv3_pre_save, >> .post_load = gicv3_post_load, >> + .priority = MIG_PRI_GICV3, >> .fields = (VMStateField[]) { >> VMSTATE_UINT32(gicd_ctlr, GICv3State), >> VMSTATE_UINT32_ARRAY(gicd_statusr, GICv3State, 2), >> diff --git a/hw/intc/arm_gicv3_its_common.c b/hw/intc/arm_gicv3_its_common.c >> index 75b9f04..854709f 100644 >> --- a/hw/intc/arm_gicv3_its_common.c >> +++ b/hw/intc/arm_gicv3_its_common.c >> @@ -48,7 +48,8 @@ static const VMStateDescription vmstate_its = { >> .name = "arm_gicv3_its", >> .pre_save = gicv3_its_pre_save, >> .post_load = gicv3_its_post_load, >> - .unmigratable = true, >> + .unmigratable = false, > > unmigratable = false is the default, so we can just delete the line. > >> + .priority = MIG_PRI_GICV3_ITS, >> .fields = (VMStateField[]) { >> VMSTATE_UINT32(ctlr, GICv3ITSState), >> VMSTATE_UINT64(cbaser, GICv3ITSState), >> diff --git a/hw/intc/arm_gicv3_its_kvm.c b/hw/intc/arm_gicv3_its_kvm.c >> index 45e57d6..3bd2873 100644 >> --- a/hw/intc/arm_gicv3_its_kvm.c >> +++ b/hw/intc/arm_gicv3_its_kvm.c >> @@ -76,18 +76,6 @@ static void kvm_arm_its_realize(DeviceState *dev, Error **errp) >> GICv3ITSState *s = ARM_GICV3_ITS_COMMON(dev); >> Error *local_err = NULL; >> >> - /* >> - * Block migration of a KVM GICv3 ITS device: the API for saving and >> - * restoring the state in the kernel is not yet available >> - */ >> - error_setg(&s->migration_blocker, "vITS migration is not implemented"); >> - migrate_add_blocker(s->migration_blocker, &local_err); >> - if (local_err) { >> - error_propagate(errp, local_err); >> - error_free(s->migration_blocker); >> - return; >> - } >> - >> s->dev_fd = kvm_create_device(kvm_state, KVM_DEV_TYPE_ARM_VGIC_ITS, false); >> if (s->dev_fd < 0) { >> error_setg_errno(errp, -s->dev_fd, "error creating in-kernel ITS"); >> @@ -104,6 +92,17 @@ static void kvm_arm_its_realize(DeviceState *dev, Error **errp) >> >> gicv3_its_init_mmio(s, NULL); >> >> + if (!kvm_device_check_attr(s->dev_fd, KVM_DEV_ARM_VGIC_GRP_ITS_REGS, >> + GITS_CTLR)) { >> + error_setg(&s->migration_blocker, "vITS migration is not implemented"); > > I think we should specifically say that it's the host kernel > that doesn't support ITS migration (ie not QEMU that's missing > support). > > The GICv3 uses the message > "This operating system kernel does not support vGICv3 migration" > which is slightly odd phrasing but I guess we should follow that. > >> + migrate_add_blocker(s->migration_blocker, &local_err); >> + if (local_err) { >> + error_propagate(errp, local_err); >> + error_free(s->migration_blocker); >> + return; >> + } >> + } >> + >> kvm_msi_use_devid = true; >> kvm_gsi_direct_mapping = false; >> kvm_msi_via_irqfd_allowed = kvm_irqfds_enabled(); >> diff --git a/include/migration/vmstate.h b/include/migration/vmstate.h >> index f2dbf84..8dab9c7 100644 >> --- a/include/migration/vmstate.h >> +++ b/include/migration/vmstate.h >> @@ -198,6 +198,8 @@ enum VMStateFlags { >> typedef enum { >> MIG_PRI_DEFAULT = 0, >> MIG_PRI_IOMMU, /* Must happen before PCI devices */ >> + MIG_PRI_GICV3_ITS, /* Must happen before PCI devices */ >> + MIG_PRI_GICV3, /* Must happen before the ITS */ >> MIG_PRI_MAX, >> } MigrationPriority; >> >> -- >> 2.5.5 >> > > thanks > -- PMM >
diff --git a/hw/intc/arm_gicv3_common.c b/hw/intc/arm_gicv3_common.c index c6493d6..4228b7c 100644 --- a/hw/intc/arm_gicv3_common.c +++ b/hw/intc/arm_gicv3_common.c @@ -145,6 +145,7 @@ static const VMStateDescription vmstate_gicv3 = { .minimum_version_id = 1, .pre_save = gicv3_pre_save, .post_load = gicv3_post_load, + .priority = MIG_PRI_GICV3, .fields = (VMStateField[]) { VMSTATE_UINT32(gicd_ctlr, GICv3State), VMSTATE_UINT32_ARRAY(gicd_statusr, GICv3State, 2), diff --git a/hw/intc/arm_gicv3_its_common.c b/hw/intc/arm_gicv3_its_common.c index 75b9f04..854709f 100644 --- a/hw/intc/arm_gicv3_its_common.c +++ b/hw/intc/arm_gicv3_its_common.c @@ -48,7 +48,8 @@ static const VMStateDescription vmstate_its = { .name = "arm_gicv3_its", .pre_save = gicv3_its_pre_save, .post_load = gicv3_its_post_load, - .unmigratable = true, + .unmigratable = false, + .priority = MIG_PRI_GICV3_ITS, .fields = (VMStateField[]) { VMSTATE_UINT32(ctlr, GICv3ITSState), VMSTATE_UINT64(cbaser, GICv3ITSState), diff --git a/hw/intc/arm_gicv3_its_kvm.c b/hw/intc/arm_gicv3_its_kvm.c index 45e57d6..3bd2873 100644 --- a/hw/intc/arm_gicv3_its_kvm.c +++ b/hw/intc/arm_gicv3_its_kvm.c @@ -76,18 +76,6 @@ static void kvm_arm_its_realize(DeviceState *dev, Error **errp) GICv3ITSState *s = ARM_GICV3_ITS_COMMON(dev); Error *local_err = NULL; - /* - * Block migration of a KVM GICv3 ITS device: the API for saving and - * restoring the state in the kernel is not yet available - */ - error_setg(&s->migration_blocker, "vITS migration is not implemented"); - migrate_add_blocker(s->migration_blocker, &local_err); - if (local_err) { - error_propagate(errp, local_err); - error_free(s->migration_blocker); - return; - } - s->dev_fd = kvm_create_device(kvm_state, KVM_DEV_TYPE_ARM_VGIC_ITS, false); if (s->dev_fd < 0) { error_setg_errno(errp, -s->dev_fd, "error creating in-kernel ITS"); @@ -104,6 +92,17 @@ static void kvm_arm_its_realize(DeviceState *dev, Error **errp) gicv3_its_init_mmio(s, NULL); + if (!kvm_device_check_attr(s->dev_fd, KVM_DEV_ARM_VGIC_GRP_ITS_REGS, + GITS_CTLR)) { + error_setg(&s->migration_blocker, "vITS migration is not implemented"); + migrate_add_blocker(s->migration_blocker, &local_err); + if (local_err) { + error_propagate(errp, local_err); + error_free(s->migration_blocker); + return; + } + } + kvm_msi_use_devid = true; kvm_gsi_direct_mapping = false; kvm_msi_via_irqfd_allowed = kvm_irqfds_enabled(); diff --git a/include/migration/vmstate.h b/include/migration/vmstate.h index f2dbf84..8dab9c7 100644 --- a/include/migration/vmstate.h +++ b/include/migration/vmstate.h @@ -198,6 +198,8 @@ enum VMStateFlags { typedef enum { MIG_PRI_DEFAULT = 0, MIG_PRI_IOMMU, /* Must happen before PCI devices */ + MIG_PRI_GICV3_ITS, /* Must happen before PCI devices */ + MIG_PRI_GICV3, /* Must happen before the ITS */ MIG_PRI_MAX, } MigrationPriority;
We change the restoration priority of both the GICv3 and ITS. The GICv3 must be restored before the ITS and the ITS needs to be restored before PCIe devices since it translates their MSI transactions. We typically observe the virtio-pci-net device sending MSI transactions very early (even before the first vcpu run) which looks weird. It appears that not servicing those transactions cause the virtio-pci-net to stall. Signed-off-by: Eric Auger <eric.auger@redhat.com> --- v1 -> v2: - handle case where migrate_add_blocker fails - add comments along with ITS and GICv3 migration priorities --- hw/intc/arm_gicv3_common.c | 1 + hw/intc/arm_gicv3_its_common.c | 3 ++- hw/intc/arm_gicv3_its_kvm.c | 23 +++++++++++------------ include/migration/vmstate.h | 2 ++ 4 files changed, 16 insertions(+), 13 deletions(-)