Re: [PATCH v4 06/14] vfio-user: find and init PCI device

[Stefan Hajnoczi]

On Wed, Dec 15, 2021 at 10:35:30AM -0500, Jagannathan Raman wrote:
> @@ -150,6 +157,38 @@ static void vfu_object_init_ctx(VfuObject *o, Error 
> **errp)
> +    o->pci_dev = PCI_DEVICE(dev);
...
> @@ -190,6 +229,8 @@ static void vfu_object_finalize(Object *obj)
>  
>      o->device = NULL;
>  
> +    o->pci_dev = NULL;

We need to consider how this interacts with device_del hot unplug.
o->pci_dev is a pointer and we don't hold a refcount, so I think
o->pci_dev could disappear at any moment.

A pair of object_ref/unref(OBJECT(o->pci_dev)) calls would not be enough
because device_del will still unrealize the device that's in use by the
vfio-user server.

I suggest adding a check to qdev_unplug() that prevents unplug when the
device is in use by the vfio-user server. That's similar to the code in
that function for preventing unplug during migration.

One way to do that is by adding a new API:

  /*
   * Register an Error that is raised when unplug is attempted on a
   * device. If another blocker has already been registered then that
   * Error may be raised during unplug instead.
   *
   * qdev_del_unplug_blocker() must be called to remove this blocker.
   */
  void qdev_add_unplug_blocker(DeviceState *dev, Error *blocker);

  /*
   * Deregister an Error that was previously registered with
   * qdev_add_unplug_blocker().
   */
  void qdev_del_unplug_blocker(DeviceState *dev, Error *blocker);

The vfio-user server then needs to add an Error *unplug_blocker field to
VfuObject and call qdev_add/del_unplug_blocker() on the PCI device.

>From a user perspective this means that device_del fails with "Device
currently in use by vfio-user server '%s'".

Stefan

signature.asc
Description: PGP signature