Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Message-ID: <b5a7efc9-7cfa-3314-fe36-b8da4a25265d@linux.ibm.com>
Date:   Wed, 18 Jan 2023 09:55:39 -0500
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:102.0) Gecko/20100101
 Thunderbird/102.6.0
Subject: Re: [PATCH v4] vfio: fix potential deadlock on vfio group lock
Content-Language: en-US
To:     "Tian, Kevin" <kevin.tian@intel.com>,
        Alex Williamson <alex.williamson@redhat.com>
Cc:     "akrowiak@linux.ibm.com" <akrowiak@linux.ibm.com>,
        "jjherne@linux.ibm.com" <jjherne@linux.ibm.com>,
        "farman@linux.ibm.com" <farman@linux.ibm.com>,
        "imbrenda@linux.ibm.com" <imbrenda@linux.ibm.com>,
        "frankja@linux.ibm.com" <frankja@linux.ibm.com>,
        "pmorel@linux.ibm.com" <pmorel@linux.ibm.com>,
        "david@redhat.com" <david@redhat.com>,
        "Christopherson, , Sean" <seanjc@google.com>,
        "intel-gfx@lists.freedesktop.org" <intel-gfx@lists.freedesktop.org>,
        "cohuck@redhat.com" <cohuck@redhat.com>,
        "linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
        "zhenyuw@linux.intel.com" <zhenyuw@linux.intel.com>,
        "pasic@linux.ibm.com" <pasic@linux.ibm.com>,
        "jgg@nvidia.com" <jgg@nvidia.com>,
        "kvm@vger.kernel.org" <kvm@vger.kernel.org>,
        "pbonzini@redhat.com" <pbonzini@redhat.com>,
        "linux-s390@vger.kernel.org" <linux-s390@vger.kernel.org>,
        "borntraeger@linux.ibm.com" <borntraeger@linux.ibm.com>,
        "Liu, Yi L" <yi.l.liu@intel.com>,
        "intel-gvt-dev@lists.freedesktop.org" 
        <intel-gvt-dev@lists.freedesktop.org>,
        "Wang, Zhi A" <zhi.a.wang@intel.com>
References: <20230114000351.115444-1-mjrosato@linux.ibm.com>
 <20230117142252.70cc85c7.alex.williamson@redhat.com>
 <BN9PR11MB52763D861C254248FD33F65C8CC79@BN9PR11MB5276.namprd11.prod.outlook.com>
From:   Matthew Rosato <mjrosato@linux.ibm.com>
In-Reply-To: <BN9PR11MB52763D861C254248FD33F65C8CC79@BN9PR11MB5276.namprd11.prod.outlook.com>
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 7bit
MIME-Version: 1.0
Precedence: bulk

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@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