[Qemu-devel] Re: [PATCH 15/15] kvm: x86: Introduce kvmclock device to save/restore its state

[Glauber Costa <glommer@redhat.com>]

On Mon, 2011-02-07 at 19:12 +0100, Jan Kiszka wrote:
> On 2011-02-07 19:04, Glauber Costa wrote:
> > On Mon, 2011-02-07 at 15:03 +0100, Jan Kiszka wrote:
> >> On 2011-02-07 14:40, Glauber Costa wrote:
> >>> On Mon, 2011-02-07 at 13:36 +0100, Jan Kiszka wrote:
> >>>> On 2011-02-07 13:27, Glauber Costa wrote:
> >>>>> On Mon, 2011-02-07 at 12:19 +0100, Jan Kiszka wrote:
> >>>>>> If kvmclock is used, which implies the kernel supports it, register a
> >>>>>> kvmclock device with the sysbus. Its main purpose is to save and restore
> >>>>>> the kernel state on migration, but this will also allow to visualize it
> >>>>>> one day.
> >>>>>>
> >>>>>> Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
> >>>>>> CC: Glauber Costa <glommer@redhat.com>
> >>>>>> ---
> >>>>>>  Makefile.target |    4 +-
> >>>>>>  hw/kvmclock.c   |  125 +++++++++++++++++++++++++++++++++++++++++++++++++++++++
> >>>>>>  hw/kvmclock.h   |   14 ++++++
> >>>>>>  hw/pc_piix.c    |   31 +++++++++++---
> >>>>>>  4 files changed, 165 insertions(+), 9 deletions(-)
> >>>>>>  create mode 100644 hw/kvmclock.c
> >>>>>>  create mode 100644 hw/kvmclock.h
> >>>>>>
> >>>>>> diff --git a/Makefile.target b/Makefile.target
> >>>>>> index b0ba95f..30232fa 100644
> >>>>>> --- a/Makefile.target
> >>>>>> +++ b/Makefile.target
> >>>>>> @@ -37,7 +37,7 @@ ifndef CONFIG_HAIKU
> >>>>>>  LIBS+=-lm
> >>>>>>  endif
> >>>>>>  
> >>>>>> -kvm.o kvm-all.o vhost.o vhost_net.o: QEMU_CFLAGS+=$(KVM_CFLAGS)
> >>>>>> +kvm.o kvm-all.o vhost.o vhost_net.o kvmclock.o: QEMU_CFLAGS+=$(KVM_CFLAGS)
> >>>>>>  
> >>>>>>  config-target.h: config-target.h-timestamp
> >>>>>>  config-target.h-timestamp: config-target.mak
> >>>>>> @@ -218,7 +218,7 @@ obj-i386-y += cirrus_vga.o apic.o ioapic.o piix_pci.o
> >>>>>>  obj-i386-y += vmmouse.o vmport.o hpet.o applesmc.o
> >>>>>>  obj-i386-y += device-hotplug.o pci-hotplug.o smbios.o wdt_ib700.o
> >>>>>>  obj-i386-y += debugcon.o multiboot.o
> >>>>>> -obj-i386-y += pc_piix.o
> >>>>>> +obj-i386-y += pc_piix.o kvmclock.o
> >>>>>>  obj-i386-$(CONFIG_SPICE) += qxl.o qxl-logger.o qxl-render.o
> >>>>>>  
> >>>>>>  # shared objects
> >>>>>> diff --git a/hw/kvmclock.c b/hw/kvmclock.c
> >>>>>> new file mode 100644
> >>>>>> index 0000000..b6ceddf
> >>>>>> --- /dev/null
> >>>>>> +++ b/hw/kvmclock.c
> >>>>>> @@ -0,0 +1,125 @@
> >>>>>> +/*
> >>>>>> + * QEMU KVM support, paravirtual clock device
> >>>>>> + *
> >>>>>> + * Copyright (C) 2011 Siemens AG
> >>>>>> + *
> >>>>>> + * Authors:
> >>>>>> + *  Jan Kiszka        <jan.kiszka@siemens.com>
> >>>>>> + *
> >>>>>> + * This work is licensed under the terms of the GNU GPL version 2.
> >>>>>> + * See the COPYING file in the top-level directory.
> >>>>>> + *
> >>>>>> + */
> >>>>>> +
> >>>>>> +#include "qemu-common.h"
> >>>>>> +#include "sysemu.h"
> >>>>>> +#include "sysbus.h"
> >>>>>> +#include "kvm.h"
> >>>>>> +#include "kvmclock.h"
> >>>>>> +
> >>>>>> +#if defined(CONFIG_KVM_PARA) && defined(KVM_CAP_ADJUST_CLOCK)
> >>>>>> +
> >>>>>> +#include <linux/kvm.h>
> >>>>>> +#include <linux/kvm_para.h>
> >>>>>> +
> >>>>>> +typedef struct KVMClockState {
> >>>>>> +    SysBusDevice busdev;
> >>>>>> +    uint64_t clock;
> >>>>>> +    bool clock_valid;
> >>>>>> +} KVMClockState;
> >>>>>> +
> >>>>>> +static void kvmclock_pre_save(void *opaque)
> >>>>>> +{
> >>>>>> +    KVMClockState *s = opaque;
> >>>>>> +    struct kvm_clock_data data;
> >>>>>> +    int ret;
> >>>>>> +
> >>>>>> +    if (s->clock_valid) {
> >>>>>> +        return;
> >>>>>> +    }
> >>>>>> +    ret = kvm_vm_ioctl(kvm_state, KVM_GET_CLOCK, &data);
> >>>>>> +    if (ret < 0) {
> >>>>>> +        fprintf(stderr, "KVM_GET_CLOCK failed: %s\n", strerror(ret));
> >>>>>> +        data.clock = 0;
> >>>>>> +    }
> >>>>>> +    s->clock = data.clock;
> >>>>>> +    /*
> >>>>>> +     * If the VM is stopped, declare the clock state valid to avoid re-reading
> >>>>>> +     * it on next vmsave (which would return a different value). Will be reset
> >>>>>> +     * when the VM is continued.
> >>>>>> +     */
> >>>>>> +    s->clock_valid = !vm_running;
> >>>>>> +}
> >>>>>> +
> >>>>>> +static int kvmclock_post_load(void *opaque, int version_id)
> >>>>>> +{
> >>>>>> +    KVMClockState *s = opaque;
> >>>>>> +    struct kvm_clock_data data;
> >>>>>> +
> >>>>>> +    data.clock = s->clock;
> >>>>>> +    data.flags = 0;
> >>>>>> +    return kvm_vm_ioctl(kvm_state, KVM_SET_CLOCK, &data);
> >>>>>> +}
> >>>>>> +
> >>>>>> +static void kvmclock_vm_state_change(void *opaque, int running, int reason)
> >>>>>> +{
> >>>>>> +    KVMClockState *s = opaque;
> >>>>>> +
> >>>>>> +    if (running) {
> >>>>>> +        s->clock_valid = false;
> >>>>>> +    }
> >>>>>> +}
> >>>>>> +
> >>>>>> +static int kvmclock_init(SysBusDevice *dev)
> >>>>>> +{
> >>>>>> +    KVMClockState *s = FROM_SYSBUS(KVMClockState, dev);
> >>>>>> +
> >>>>>> +    qemu_add_vm_change_state_handler(kvmclock_vm_state_change, s);
> >>>>>> +    return 0;
> >>>>>> +}
> >>>>>> +
> >>>>>> +static const VMStateDescription kvmclock_vmsd = {
> >>>>>> +    .name = "kvmclock",
> >>>>>> +    .version_id = 1,
> >>>>>> +    .minimum_version_id = 1,
> >>>>>> +    .minimum_version_id_old = 1,
> >>>>>> +    .pre_save = kvmclock_pre_save,
> >>>>>> +    .post_load = kvmclock_post_load,
> >>>>>> +    .fields = (VMStateField[]) {
> >>>>>> +        VMSTATE_UINT64(clock, KVMClockState),
> >>>>>> +        VMSTATE_END_OF_LIST()
> >>>>>> +    }
> >>>>>> +};
> >>>>>> +
> >>>>>> +static SysBusDeviceInfo kvmclock_info = {
> >>>>>> +    .qdev.name = "kvmclock",
> >>>>>> +    .qdev.size = sizeof(KVMClockState),
> >>>>>> +    .qdev.vmsd = &kvmclock_vmsd,
> >>>>>> +    .qdev.no_user = 1,
> >>>>>> +    .init = kvmclock_init,
> >>>>>> +};
> >>>>>> +
> >>>>>> +/* Note: Must be called after VCPU initialization. */
> >>>>>> +void kvmclock_create(void)
> >>>>>> +{
> >>>>>> +    if (kvm_enabled() &&
> >>>>>> +        first_cpu->cpuid_kvm_features & (1ULL << KVM_FEATURE_CLOCKSOURCE)) {
> >>>>>> +        sysbus_create_simple("kvmclock", -1, NULL);
> >>>>>> +    }
> >>>>>> +}
> >>>>>> +
> >>>>>> +static void kvmclock_register_device(void)
> >>>>>> +{
> >>>>>> +    if (kvm_enabled()) {
> >>>>>> +        sysbus_register_withprop(&kvmclock_info);
> >>>>>> +    }
> >>>>>> +}
> >>>>>> +
> >>>>>> +device_init(kvmclock_register_device);
> >>>>>> +
> >>>>>> +#else /* !(CONFIG_KVM_PARA && KVM_CAP_ADJUST_CLOCK) */
> >>>>>> +
> >>>>>> +void kvmclock_create(void)
> >>>>>> +{
> >>>>>> +}
> >>>>>> +#endif /* !(CONFIG_KVM_PARA && KVM_CAP_ADJUST_CLOCK) */
> >>>>>> diff --git a/hw/kvmclock.h b/hw/kvmclock.h
> >>>>>> new file mode 100644
> >>>>>> index 0000000..7a83cbe
> >>>>>> --- /dev/null
> >>>>>> +++ b/hw/kvmclock.h
> >>>>>> @@ -0,0 +1,14 @@
> >>>>>> +/*
> >>>>>> + * QEMU KVM support, paravirtual clock device
> >>>>>> + *
> >>>>>> + * Copyright (C) 2011 Siemens AG
> >>>>>> + *
> >>>>>> + * Authors:
> >>>>>> + *  Jan Kiszka        <jan.kiszka@siemens.com>
> >>>>>> + *
> >>>>>> + * This work is licensed under the terms of the GNU GPL version 2.
> >>>>>> + * See the COPYING file in the top-level directory.
> >>>>>> + *
> >>>>>> + */
> >>>>>> +
> >>>>>> +void kvmclock_create(void);
> >>>>>> diff --git a/hw/pc_piix.c b/hw/pc_piix.c
> >>>>>> index 7b74473..9bc4659 100644
> >>>>>> --- a/hw/pc_piix.c
> >>>>>> +++ b/hw/pc_piix.c
> >>>>>> @@ -32,6 +32,7 @@
> >>>>>>  #include "boards.h"
> >>>>>>  #include "ide.h"
> >>>>>>  #include "kvm.h"
> >>>>>> +#include "kvmclock.h"
> >>>>>>  #include "sysemu.h"
> >>>>>>  #include "sysbus.h"
> >>>>>>  #include "arch_init.h"
> >>>>>> @@ -66,7 +67,8 @@ static void pc_init1(ram_addr_t ram_size,
> >>>>>>                       const char *kernel_cmdline,
> >>>>>>                       const char *initrd_filename,
> >>>>>>                       const char *cpu_model,
> >>>>>> -                     int pci_enabled)
> >>>>>> +                     int pci_enabled,
> >>>>>> +                     int kvmclock_enabled)
> >>>>>>  
> >>>>> What exactly is your motivation to that ? I think mid/long-term
> >>>>> we should be making machine initialization more common among
> >>>>> architectures, not introducing more arch specific, or even worse, kvm
> >>>>> specific parameters here.
> >>>>>
> >>>>> I'd like to understand what do we gain from that, since opting kvmclock
> >>>>> in our out is done by cpuid anyway - no need for a specific machine.
> >>>>
> >>>> Is that really the case? I thought we were already shipping versions
> >>>> where that CPU feature was enabled by default. If not, I'll happily drop
> >>>> that admittedly clumsy approach above.
> >>>
> >>> Yes, AFAIK, kvmclock is enabled by default, disabled by cpuid-leaf, as
> >>> in
> >>> -cpu kvm64,-kvmclock
> >>>
> >>> So your test for cpuid bit before starting kvmclock should already cover
> >>> it.
> >>>
> >>
> >> No, just the contrary: As kvmclock was always enabled in older versions
> >> and the compat machines also expose it, we cannot rely on the flag to
> >> enable this new (and therefore 0.15-only) vmstate.
> > 
> > I see. You're not enabling kvmclock functionality, but rather, kvmclock
> > device.
> > 
> > It makes sense then, but I still would prefer it encoded somewhere else
> > (but not a hard constraint anymore). Can't qdev encode it, say, in the
> > value field of the device? 
> 
> It would take some machine properties that could be pre-set according to
> compat requirements. I have such machine-specific properties in mind
> anyway in order to implement -machine in a more flexible way than
> proposed so far.
> 
> I'll surely clean up the small ugliness this patch introduces to
> pc_piix.c once there is a better way, I just don't want to let kvmclock
> wait for that perfect infrastructure.

Perfect is the enemy of good, I'm all for it.