Re: [Qemu-devel] [PATCH] virtio-pci: fix missing device properties

[Marc-André Lureau]

Hi

On Wed, Jun 26, 2019 at 3:56 AM Eduardo Habkost <address@hidden> wrote:
>
> On Wed, Jun 26, 2019 at 01:23:33AM +0200, Marc-André Lureau wrote:
> > Since commit a4ee4c8baa37154 ("virtio: Helper for registering virtio
> > device types"), virtio-gpu-pci, virtio-vga, and virtio-crypto-pci lost
> > some properties: "ioeventfd" and "vectors". This may cause various
> > issues, such as failing migration or invalid properties.
> >
> > Since those VirtioPCI devices do not have a base name, their class are
> > initialized with virtio_pci_generic_base_class_init(). However, if the
> > VirtioPCIDeviceTypeInfo provided a class_init which sets dc->props,
> > the properties were overwritten by virtio_pci_generic_class_init().
> >
> > Instead, introduce an intermediary base-type to register the generic
> > properties.
> >
> > Fixes: a4ee4c8baa37154f42b4dc6a13fee79268d15238
> > Cc: address@hidden
> > Signed-off-by: Marc-André Lureau <address@hidden>
> > ---
> >  hw/virtio/virtio-pci.c | 28 ++++++++++++++--------------
> >  1 file changed, 14 insertions(+), 14 deletions(-)
> >
> > diff --git a/hw/virtio/virtio-pci.c b/hw/virtio/virtio-pci.c
> > index e6d5467e54..62c4977332 100644
> > --- a/hw/virtio/virtio-pci.c
> > +++ b/hw/virtio/virtio-pci.c
> > @@ -1913,13 +1913,6 @@ static void 
> > virtio_pci_generic_class_init(ObjectClass *klass, void *data)
> >      dc->props = virtio_pci_generic_properties;
> >  }
> >
> > -/* Used when the generic type and the base type is the same */
> > -static void virtio_pci_generic_base_class_init(ObjectClass *klass, void 
> > *data)
> > -{
> > -    virtio_pci_base_class_init(klass, data);
> > -    virtio_pci_generic_class_init(klass, NULL);
> > -}
> > -
> >  static void virtio_pci_transitional_instance_init(Object *obj)
> >  {
> >      VirtIOPCIProxy *proxy = VIRTIO_PCI(obj);
> > @@ -1938,14 +1931,13 @@ static void 
> > virtio_pci_non_transitional_instance_init(Object *obj)
> >
> >  void virtio_pci_types_register(const VirtioPCIDeviceTypeInfo *t)
> >  {
> > +    char *base_name = NULL;
> >      TypeInfo base_type_info = {
> >          .name          = t->base_name,
> >          .parent        = t->parent ? t->parent : TYPE_VIRTIO_PCI,
> >          .instance_size = t->instance_size,
> >          .instance_init = t->instance_init,
> >          .class_size    = t->class_size,
> > -        .class_init    = virtio_pci_base_class_init,
> > -        .class_data    = (void *)t,
> >          .abstract      = true,
> >      };
> >      TypeInfo generic_type_info = {
> > @@ -1961,13 +1953,20 @@ void virtio_pci_types_register(const 
> > VirtioPCIDeviceTypeInfo *t)
> >
> >      if (!base_type_info.name) {
> >          /* No base type -> register a single generic device type */
> > -        base_type_info.name = t->generic_name;
> > -        base_type_info.class_init = virtio_pci_generic_base_class_init;
> > -        base_type_info.interfaces = generic_type_info.interfaces;
> > -        base_type_info.abstract = false;
> > -        generic_type_info.name = NULL;
> > +        /* use intermediate %s-base-type to add generic device props */
> > +        base_name = g_strdup_printf("%s-base-type", t->generic_name);
> > +        base_type_info.name = base_name;
> > +        base_type_info.class_init = virtio_pci_generic_class_init;
> > +
> > +        generic_type_info.parent = base_name;
> > +        generic_type_info.class_init = virtio_pci_base_class_init;
> > +        generic_type_info.class_data = (void *)t;
>
> Why are you using virtio_pci_generic_class_init for the base
> class, and virtio_pci_base_class_init for the subclass, but doing
> exactly the opposite when t->base_name is set?
>
> Isn't it simpler to just initialize base_type_info.name and leave
> all the rest alone?  Patch below.

That was my initial approach. But then I tested the backport on v4.0.0, you get:

hw/display/virtio-vga.c:200:virtio_vga_class_init: Object
0x56247edbc0e0 is not an instance of type virtio-vga

The problem is that the introduced base class calls the t->class_init,
which expects the final class (virtio-vga, not virtio-vga-base-type).

This seems to be limited to virtio-vga, and gone in upstream since the
registration changed. But I would rather have the same patch
reviewed/applied.


>
> Signed-off-by: Eduardo Habkost <address@hidden>
> ---
> diff --git a/hw/virtio/virtio-pci.c b/hw/virtio/virtio-pci.c
> index e6d5467e54..3ee50a0783 100644
> --- a/hw/virtio/virtio-pci.c
> +++ b/hw/virtio/virtio-pci.c
> @@ -1913,13 +1913,6 @@ static void virtio_pci_generic_class_init(ObjectClass 
> *klass, void *data)
>      dc->props = virtio_pci_generic_properties;
>  }
>
> -/* Used when the generic type and the base type is the same */
> -static void virtio_pci_generic_base_class_init(ObjectClass *klass, void 
> *data)
> -{
> -    virtio_pci_base_class_init(klass, data);
> -    virtio_pci_generic_class_init(klass, NULL);
> -}
> -
>  static void virtio_pci_transitional_instance_init(Object *obj)
>  {
>      VirtIOPCIProxy *proxy = VIRTIO_PCI(obj);
> @@ -1938,8 +1931,11 @@ static void 
> virtio_pci_non_transitional_instance_init(Object *obj)
>
>  void virtio_pci_types_register(const VirtioPCIDeviceTypeInfo *t)
>  {
> +    char *base_name = t->base_name ?
> +                      NULL :
> +                      g_strdup_printf("%s-base-type", t->generic_name);
>      TypeInfo base_type_info = {
> -        .name          = t->base_name,
> +        .name          = t->base_name ?: base_name,
>          .parent        = t->parent ? t->parent : TYPE_VIRTIO_PCI,
>          .instance_size = t->instance_size,
>          .instance_init = t->instance_init,
> @@ -1959,21 +1955,8 @@ void virtio_pci_types_register(const 
> VirtioPCIDeviceTypeInfo *t)
>          },
>      };
>
> -    if (!base_type_info.name) {
> -        /* No base type -> register a single generic device type */
> -        base_type_info.name = t->generic_name;
> -        base_type_info.class_init = virtio_pci_generic_base_class_init;
> -        base_type_info.interfaces = generic_type_info.interfaces;
> -        base_type_info.abstract = false;
> -        generic_type_info.name = NULL;
> -        assert(!t->non_transitional_name);
> -        assert(!t->transitional_name);
> -    }
> -
>      type_register(&base_type_info);
> -    if (generic_type_info.name) {
> -        type_register(&generic_type_info);
> -    }
> +    type_register(&generic_type_info);

t->generic_name can't be NULL anymore? I thought the condition was
done for a good reason.

>
>      if (t->non_transitional_name) {
>          const TypeInfo non_transitional_type_info = {
> @@ -2005,6 +1988,7 @@ void virtio_pci_types_register(const 
> VirtioPCIDeviceTypeInfo *t)
>          };
>          type_register(&transitional_type_info);
>      }
> +    g_free(base_name);
>  }
>
>  /* virtio-pci-bus */
>
>
> --
> Eduardo
>


-- 
Marc-André Lureau