Re: [Qemu-devel] [RFC PATCH V4 1/4] vfio: introduce a new VFIO subregion for mdev device migration support

[Alex Williamson]

On Tue, 17 Apr 2018 10:36:53 +0000
"Zhang, Yulei" <address@hidden> wrote:

> > -----Original Message-----
> > From: Alex Williamson [mailto:address@hidden
> > Sent: Tuesday, April 17, 2018 4:15 AM
> > To: Kirti Wankhede <address@hidden>
> > Cc: Zhang, Yulei <address@hidden>; address@hidden; Tian,
> > Kevin <address@hidden>; address@hidden;
> > address@hidden; Wang, Zhi A <address@hidden>;
> > address@hidden; address@hidden
> > Subject: Re: [RFC PATCH V4 1/4] vfio: introduce a new VFIO subregion for
> > mdev device migration support
> > 
> > On Mon, 16 Apr 2018 20:14:03 +0530
> > Kirti Wankhede <address@hidden> wrote:
> >   
> > > On 4/10/2018 11:32 AM, Yulei Zhang wrote:  
> > > > New VFIO sub region VFIO_REGION_SUBTYPE_DEVICE_STATE is added
> > > > to fetch and restore the status of mdev device vGPU during the
> > > > live migration.
> > > >
> > > > Signed-off-by: Yulei Zhang <address@hidden>
> > > > ---
> > > >  hw/vfio/pci.c              | 25 ++++++++++++++++++++++++-
> > > >  hw/vfio/pci.h              |  2 ++
> > > >  linux-headers/linux/vfio.h |  9 ++++++---
> > > >  3 files changed, 32 insertions(+), 4 deletions(-)
> > > >
> > > > diff --git a/hw/vfio/pci.c b/hw/vfio/pci.c
> > > > index c977ee3..f98a9dd 100644
> > > > --- a/hw/vfio/pci.c
> > > > +++ b/hw/vfio/pci.c
> > > > @@ -32,6 +32,7 @@
> > > >  #include "pci.h"
> > > >  #include "trace.h"
> > > >  #include "qapi/error.h"
> > > > +#include "migration/blocker.h"
> > > >
> > > >  #define MSIX_CAP_LENGTH 12
> > > >
> > > > @@ -2821,6 +2822,25 @@ static void vfio_realize(PCIDevice *pdev, Error  
> > **errp)  
> > > >          vfio_vga_quirk_setup(vdev);
> > > >      }
> > > >
> > > > +    struct vfio_region_info *device_state;
> > > > +    /* device state region setup */
> > > > +    if (!vfio_get_dev_region_info(&vdev->vbasedev,
> > > > +                VFIO_REGION_TYPE_PCI_VENDOR_TYPE,  
> > 
> > This is not how VENDOR_TYPE is meant to be used.  We have a 32-bit type
> > and 32-bit sub-type.  When bit 31 of type is set (ie. VENDOR_TYPE),
> > then the low 16-bits (VENDOR_MASK) defines a vendor specific set of
> > sub-types.  Using it as above, you're stomping on vendor 0x0000's
> > vendor sub-types.  I would expect non-vendor specific types to leave
> > bit 31 of the type field clear.  We could then define a universal type
> > for device state with a sub-type specifying the API such that we could
> > revise the API by providing an updated sub-type.  Or perhaps there are
> > devices that want separate save vs restore regions, we could do that
> > with the sub-type.  
> 
> Got it. For the subtype, can we use it to specify how to use the region,
> to indicate it can be mmaped or not. 

No, REGION_INFO tells us about mmap'ability and device state regions
can certainly support the sparse mmap capability, we already have full
ability to describe the mmap features of a region without dedicating a
sub-type to it.  A sub-type must have a defined API, but use the
existing mechanisms for mmap.

> > > > +                VFIO_REGION_SUBTYPE_DEVICE_STATE, &device_state)) {
> > > > +        memcpy(&vdev->device_state, device_state,
> > > > +               sizeof(struct vfio_region_info));
> > > > +        g_free(device_state);
> > > > +    } else {
> > > > +        error_setg(&vdev->migration_blocker,
> > > > +                "Migration disabled: cannot support device state 
> > > > region");
> > > > +        migrate_add_blocker(vdev->migration_blocker, &err);  
> > 
> > This appears as if it's going to be rather verbose and generate errors
> > for anything not supporting migration, which is currently everything.
> > Maybe there should be an OnOffAuto vfio-pci device option for the user
> > to specify migration such that if migration=on is specified the device
> > will fail if it's not available.  Otherwise the default would be auto.
> >   
> We only get this error message when we try to migrate the guest with
> vfio-pci device which doesn't contain the device_state region, just like 
> currently we set unmigratable=1.

Ok, I still wonder though if management tools might want the ability to
fail the device if it doesn't support migration via an OnOffAuto
option.  I can imagine a customer that doesn't want to compromise the
migrate'ability of their VM and doesn't want to wait until a migration
is attempted to find out it's blocked.  Thanks,

Alex
 
> > > > +        if (err) {
> > > > +            error_propagate(errp, err);
> > > > +            error_free(vdev->migration_blocker);
> > > > +            goto error;
> > > > +        }
> > > > +    }
> > > > +  
> > >
> > > I think there should be a _PROBE ioctl before trying to setup region for
> > > migration. If vfio device driver or vendor driver for mdev device
> > > supports migration capability, _PROBE ioctl should return success along
> > > with region's information <index, size> which can be used to setup the
> > > region.  
> > 
> > How is _PROBE different from _REGION_INFO which already exists to tell
> > us information about the region?
> >   
> > > >      for (i = 0; i < PCI_ROM_SLOT; i++) {
> > > >          vfio_bar_quirk_setup(vdev, i);
> > > >      }
> > > > @@ -2884,6 +2904,10 @@ out_teardown:
> > > >      vfio_teardown_msi(vdev);
> > > >      vfio_bars_exit(vdev);
> > > >  error:
> > > > +    if (vdev->migration_blocker) {
> > > > +        migrate_del_blocker(vdev->migration_blocker);
> > > > +        error_free(vdev->migration_blocker);
> > > > +    }
> > > >      error_prepend(errp, ERR_PREFIX, vdev->vbasedev.name);
> > > >  }
> > > >
> > > > @@ -3009,7 +3033,6 @@ static Property vfio_pci_dev_properties[] = {
> > > >
> > > >  static const VMStateDescription vfio_pci_vmstate = {
> > > >      .name = "vfio-pci",
> > > > -    .unmigratable = 1,  
> > >
> > > Based on the result of above _PROBE ioctl, 'unmigratable' should be set
> > > if vendor driver doesn't support migration capability.  
> > 
> > .unmigratable isn't really dynamically settable aiui, thus the
> > existence of the migration blocker.
> >   
> > > >  };
> > > >
> > > >  static void vfio_pci_dev_class_init(ObjectClass *klass, void *data)
> > > > diff --git a/hw/vfio/pci.h b/hw/vfio/pci.h
> > > > index 502a575..0ee1724 100644
> > > > --- a/hw/vfio/pci.h
> > > > +++ b/hw/vfio/pci.h
> > > > @@ -116,6 +116,8 @@ typedef struct VFIOPCIDevice {
> > > >      VFIOBAR bars[PCI_NUM_REGIONS - 1]; /* No ROM */
> > > >      VFIOVGA *vga; /* 0xa0000, 0x3b0, 0x3c0 */
> > > >      void *igd_opregion;
> > > > +    struct vfio_region_info device_state;
> > > > +    Error *migration_blocker;
> > > >      PCIHostDeviceAddress host;
> > > >      EventNotifier err_notifier;
> > > >      EventNotifier req_notifier;
> > > > diff --git a/linux-headers/linux/vfio.h b/linux-headers/linux/vfio.h
> > > > index 4312e96..e3380ad 100644
> > > > --- a/linux-headers/linux/vfio.h
> > > > +++ b/linux-headers/linux/vfio.h
> > > > @@ -297,9 +297,12 @@ struct vfio_region_info_cap_type {
> > > >  #define VFIO_REGION_TYPE_PCI_VENDOR_MASK       (0xffff)
> > > >
> > > >  /* 8086 Vendor sub-types */
> > > > -#define VFIO_REGION_SUBTYPE_INTEL_IGD_OPREGION (1)
> > > > -#define VFIO_REGION_SUBTYPE_INTEL_IGD_HOST_CFG (2)
> > > > -#define VFIO_REGION_SUBTYPE_INTEL_IGD_LPC_CFG  (3)
> > > > +#define VFIO_REGION_SUBTYPE_INTEL_IGD_OPREGION         (1)
> > > > +#define VFIO_REGION_SUBTYPE_INTEL_IGD_HOST_CFG         (2)
> > > > +#define VFIO_REGION_SUBTYPE_INTEL_IGD_LPC_CFG          (3)
> > > > +
> > > > +/* Mdev sub-type for device state save and restore */
> > > > +#define VFIO_REGION_SUBTYPE_DEVICE_STATE       (4)  
> > 
> > This continues on with the 0x8086 vendor sub-types, which is entirely
> > unnecessary as each vendor has an entire sub-type address space and we
> > shouldn't be using a vendor type anyway.  Additionally, we need
> > significantly more documentation on the API for this region.  Thanks,
> > 
> > Alex
> >   
> > > >
> > > >  /**
> > > >   * VFIO_DEVICE_GET_IRQ_INFO - _IOWR(VFIO_TYPE, VFIO_BASE + 9,
> > > >  
>