On 07/06/2018 08:47 PM, Xiubo Li wrote:
> On 2018/7/7 2:58, Mike Christie wrote:
>> On 07/05/2018 09:57 PM, xiu...@redhat.com wrote:
>>> void uio_event_notify(struct uio_info *info)
>>> {
>>> - struct uio_device *idev = info->uio_dev;
>>> + struct uio_device *idev;
>>> +
>>> + if (!info)
>>> + return;
>>> +
>>> + idev = info->uio_dev;
>>>
>> For this one too, I am not sure if it is needed.
>>
>> uio_interrupt -> uio_event_notify. See other mail.
>>
>> driver XYZ -> uio_event_notify. I think drivers need to handle this and
>> set some bits and/or perform some cleanup to make sure they are not
>> calling uio_event_notify after it has called uio_unregister_device. The
>> problem with the above test is if they do not they could have called
>> uio_unregister_device right after the info test so you could still hit
>> the problem.
>
> When we are tcmu_destroy_device(), if the netlink notify event to
> userspace is not successful then the TCMU will call the uio unregister,
> which will set the idev->info = NULL, without close and deleting the
> device in userspace. But the TCMU could still queue cmds to the ring
> buffer, then the uio_event_notify will be called.
Before tcmu_destroy_device is called LIO will have stopped new IO and
waited for outstanding IO to be completed, so it would never be called
after uio_unregister_device. If it does, it's a bug in LIO.
