[Intel-gfx] [PATCH v4] vfio: fix potential deadlock on vfio group lock

[Matthew Rosato]

On 1/18/23 4:03 AM, Tian, Kevin wrote:
>> From: Alex Williamson
>> Sent: Wednesday, January 18, 2023 5:23 AM
>>
>> On Fri, 13 Jan 2023 19:03:51 -0500
>> Matthew Rosato <mjrosato at linux.ibm.com> wrote:
>>
>>>  void vfio_device_group_close(struct vfio_device *device)
>>>  {
>>> +	void (*put_kvm)(struct kvm *kvm);
>>> +	struct kvm *kvm;
>>> +
>>>  	mutex_lock(&device->group->group_lock);
>>> +	kvm = device->kvm;
>>> +	put_kvm = device->put_kvm;
>>>  	vfio_device_close(device, device->group->iommufd);
>>> +	if (kvm == device->kvm)
>>> +		kvm = NULL;
>>
>> Hmm, so we're using whether the device->kvm pointer gets cleared in
>> last_close to detect whether we should put the kvm reference.  That's a
>> bit obscure.  Our get and put is also asymmetric.
>>
>> Did we decide that we couldn't do this via a schedule_work() from the
>> last_close function, ie. implementing our own version of an async put?
>> It seems like that potentially has a cleaner implementation, symmetric
>> call points, handling all the storing and clearing of kvm related
>> pointers within the get/put wrappers, passing only a vfio_device to the
>> put wrapper, using the "vfio_device_" prefix for both.  Potentially
>> we'd just want an unconditional flush outside of lock here for
>> deterministic release.
>>
>> What's the downside?  Thanks,
>>
> 
> btw I guess this can be also fixed by Yi's work here:
> 
> https://lore.kernel.org/kvm/20230117134942.101112-6-yi.l.liu@intel.com/
> 
> with set_kvm(NULL) moved to the release callback of kvm_vfio device,
> such circular lock dependency can be avoided too.

Oh, interesting...  It seems to me that this would eliminate the reported call chain altogether:

kvm_put_kvm
 -> kvm_destroy_vm
  -> kvm_destroy_devices
   -> kvm_vfio_destroy (starting here -- this would no longer be executed)
    -> kvm_vfio_file_set_kvm
     -> vfio_file_set_kvm
      -> group->group_lock/group_rwsem

because kvm_destroy_devices now can't end up calling kvm_vfio_destroy and friends, it won't try and acquire the group lock a 2nd time making a kvm_put_kvm while the group lock is held OK to do.  The vfio_file_set_kvm call will now always come from a separate thread of execution, kvm_vfio_group_add, kvm_vfio_group_del or the release thread:

kvm_device_release (where the group->group_lock would not be held since vfio does not trigger closing of the kvm fd)
 -> kvm_vfio_destroy (or, kvm_vfio_release)
  -> kvm_vfio_file_set_kvm
   -> vfio_file_set_kvm
    -> group->group_lock/group_rwsem