| Message ID | 20170608070351.1434-4-sjitindarsingh@gmail.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Thu, Jun 08, 2017 at 05:03:49PM +1000, Suraj Jitindar Singh wrote: > Implement support to migrate a guest which is using the large > decrementer. > > We need to save the decrementer width to the migration stream. > On incoming migration we then need to check that the hypervisor is > capable of letting the guest use the large decrementer and that the > decrementer width is the same on the receiving side. Since there is no > way to tell the guest when the width of the decrementer changes we have > to terminate if the decrementer width is not what the guest expects. > If we can use the large decrementer then we have to tell the hypervisor > to enable it. > > Signed-off-by: Suraj Jitindar Singh <sjitindarsingh@gmail.com> > --- > hw/ppc/spapr.c | 63 ++++++++++++++++++++++++++++++++++++++++++++++++++ > include/hw/ppc/spapr.h | 1 + > 2 files changed, 64 insertions(+) > > diff --git a/hw/ppc/spapr.c b/hw/ppc/spapr.c > index 5d10366..6ba869a 100644 > --- a/hw/ppc/spapr.c > +++ b/hw/ppc/spapr.c > @@ -1420,6 +1420,45 @@ static bool spapr_vga_init(PCIBus *pci_bus, Error **errp) > } > } > > +static int spapr_import_large_decr_bits(sPAPRMachineState *spapr) > +{ > + /* > + * If the guest uses the large decrementer then this hypervisor must also > + * support it and have the exact same width. We must also enable the large > + * decrementer because we have no way to tell the guest to stop using it. > + */ > + if (spapr->large_decr_bits) { > + uint32_t dec_bits = 32; > + PowerPCCPU *cpu = POWERPC_CPU(first_cpu); > + CPUState *cs = CPU(cpu); > + PowerPCCPUClass *pcc = POWERPC_CPU_GET_CLASS(cs); > + > + if (kvm_enabled()) { > + if (!kvmppc_has_cap_large_decr()) { > + error_report("Host doesn't support large decrementer and guest requires it"); > + return -EINVAL; > + } > + dec_bits = kvmppc_get_dec_bits(); > + } else { > + dec_bits = pcc->large_decr_bits; > + } > + > + if (spapr->large_decr_bits != dec_bits) { > + error_report("Host large decrementer size [%u] doesn't match what guest expects [%u]", > + dec_bits, spapr->large_decr_bits); > + return -EINVAL; > + } Could you just use a VMSTATE_EQUAL() rather than explicit post_load logic? > + > + if (kvm_enabled()) { > + CPU_FOREACH(cs) { > + kvmppc_configure_large_decrementer(cs, true); > + } > + } > + } > + > + return 0; > +} > + > static int spapr_post_load(void *opaque, int version_id) > { > sPAPRMachineState *spapr = (sPAPRMachineState *)opaque; > @@ -1439,8 +1478,13 @@ static int spapr_post_load(void *opaque, int version_id) > * value into the RTC device */ > if (version_id < 3) { > err = spapr_rtc_import_offset(&spapr->rtc, spapr->rtc_offset); > + if (err) { > + return err; > + } > } > > + err = spapr_import_large_decr_bits(spapr); > + > return err; > } > > @@ -1529,6 +1573,24 @@ static const VMStateDescription vmstate_spapr_patb_entry = { > }, > }; > > +static bool spapr_large_decr_entry_needed(void *opaque) > +{ > + sPAPRMachineState *spapr = opaque; > + > + return !!spapr->large_decr_bits; Hrm. We have no existing releases - upstream or down - that support POWER9, which is the first thing with large decr. Rather than fancy conditional logic, could we just always transfer the decr size for POWER9? Do we need to support qemu on POWER9 with a pre-large-decr host kernel? Or is POWER9 support new enough that we can just say that you require a host kernel with large decr support to run POWER9 guests. That could simplify several things. > +} > + > +static const VMStateDescription vmstate_spapr_large_decr_entry = { > + .name = "spapr_large_decr_entry", > + .version_id = 1, > + .minimum_version_id = 1, > + .needed = spapr_large_decr_entry_needed, > + .fields = (VMStateField[]) { > + VMSTATE_UINT32(large_decr_bits, sPAPRMachineState), > + VMSTATE_END_OF_LIST() > + }, > +}; > + > static const VMStateDescription vmstate_spapr = { > .name = "spapr", > .version_id = 3, > @@ -1547,6 +1609,7 @@ static const VMStateDescription vmstate_spapr = { > .subsections = (const VMStateDescription*[]) { > &vmstate_spapr_ov5_cas, > &vmstate_spapr_patb_entry, > + &vmstate_spapr_large_decr_entry, > NULL > } > }; > diff --git a/include/hw/ppc/spapr.h b/include/hw/ppc/spapr.h > index 98fb78b..4ba9b89 100644 > --- a/include/hw/ppc/spapr.h > +++ b/include/hw/ppc/spapr.h > @@ -91,6 +91,7 @@ struct sPAPRMachineState { > sPAPROptionVector *ov5_cas; /* negotiated (via CAS) option vectors */ > bool cas_reboot; > bool cas_legacy_guest_workaround; > + uint32_t large_decr_bits; /* Large decrementer width (0 -> not in use) */ Having this here as well as in the CPU class seems a bit icky. > Notifier epow_notifier; > QTAILQ_HEAD(, sPAPREventLogEntry) pending_events;
diff --git a/hw/ppc/spapr.c b/hw/ppc/spapr.c index 5d10366..6ba869a 100644 --- a/hw/ppc/spapr.c +++ b/hw/ppc/spapr.c @@ -1420,6 +1420,45 @@ static bool spapr_vga_init(PCIBus *pci_bus, Error **errp) } } +static int spapr_import_large_decr_bits(sPAPRMachineState *spapr) +{ + /* + * If the guest uses the large decrementer then this hypervisor must also + * support it and have the exact same width. We must also enable the large + * decrementer because we have no way to tell the guest to stop using it. + */ + if (spapr->large_decr_bits) { + uint32_t dec_bits = 32; + PowerPCCPU *cpu = POWERPC_CPU(first_cpu); + CPUState *cs = CPU(cpu); + PowerPCCPUClass *pcc = POWERPC_CPU_GET_CLASS(cs); + + if (kvm_enabled()) { + if (!kvmppc_has_cap_large_decr()) { + error_report("Host doesn't support large decrementer and guest requires it"); + return -EINVAL; + } + dec_bits = kvmppc_get_dec_bits(); + } else { + dec_bits = pcc->large_decr_bits; + } + + if (spapr->large_decr_bits != dec_bits) { + error_report("Host large decrementer size [%u] doesn't match what guest expects [%u]", + dec_bits, spapr->large_decr_bits); + return -EINVAL; + } + + if (kvm_enabled()) { + CPU_FOREACH(cs) { + kvmppc_configure_large_decrementer(cs, true); + } + } + } + + return 0; +} + static int spapr_post_load(void *opaque, int version_id) { sPAPRMachineState *spapr = (sPAPRMachineState *)opaque; @@ -1439,8 +1478,13 @@ static int spapr_post_load(void *opaque, int version_id) * value into the RTC device */ if (version_id < 3) { err = spapr_rtc_import_offset(&spapr->rtc, spapr->rtc_offset); + if (err) { + return err; + } } + err = spapr_import_large_decr_bits(spapr); + return err; } @@ -1529,6 +1573,24 @@ static const VMStateDescription vmstate_spapr_patb_entry = { }, }; +static bool spapr_large_decr_entry_needed(void *opaque) +{ + sPAPRMachineState *spapr = opaque; + + return !!spapr->large_decr_bits; +} + +static const VMStateDescription vmstate_spapr_large_decr_entry = { + .name = "spapr_large_decr_entry", + .version_id = 1, + .minimum_version_id = 1, + .needed = spapr_large_decr_entry_needed, + .fields = (VMStateField[]) { + VMSTATE_UINT32(large_decr_bits, sPAPRMachineState), + VMSTATE_END_OF_LIST() + }, +}; + static const VMStateDescription vmstate_spapr = { .name = "spapr", .version_id = 3, @@ -1547,6 +1609,7 @@ static const VMStateDescription vmstate_spapr = { .subsections = (const VMStateDescription*[]) { &vmstate_spapr_ov5_cas, &vmstate_spapr_patb_entry, + &vmstate_spapr_large_decr_entry, NULL } }; diff --git a/include/hw/ppc/spapr.h b/include/hw/ppc/spapr.h index 98fb78b..4ba9b89 100644 --- a/include/hw/ppc/spapr.h +++ b/include/hw/ppc/spapr.h @@ -91,6 +91,7 @@ struct sPAPRMachineState { sPAPROptionVector *ov5_cas; /* negotiated (via CAS) option vectors */ bool cas_reboot; bool cas_legacy_guest_workaround; + uint32_t large_decr_bits; /* Large decrementer width (0 -> not in use) */ Notifier epow_notifier; QTAILQ_HEAD(, sPAPREventLogEntry) pending_events;
Implement support to migrate a guest which is using the large decrementer. We need to save the decrementer width to the migration stream. On incoming migration we then need to check that the hypervisor is capable of letting the guest use the large decrementer and that the decrementer width is the same on the receiving side. Since there is no way to tell the guest when the width of the decrementer changes we have to terminate if the decrementer width is not what the guest expects. If we can use the large decrementer then we have to tell the hypervisor to enable it. Signed-off-by: Suraj Jitindar Singh <sjitindarsingh@gmail.com> --- hw/ppc/spapr.c | 63 ++++++++++++++++++++++++++++++++++++++++++++++++++ include/hw/ppc/spapr.h | 1 + 2 files changed, 64 insertions(+)