Re: [syzbot] WARNING in vhost_dev_cleanup (2)

[Stefano Garzarella <sgarzare@redhat.com>]

On Fri, Feb 18, 2022 at 12:23:10PM -0600, Mike Christie wrote:
>On 2/18/22 11:53 AM, Mike Christie wrote:
>> On 2/17/22 3:48 AM, Stefano Garzarella wrote:
>>>
>>> On Thu, Feb 17, 2022 at 8:50 AM Michael S. Tsirkin <mst@redhat.com> wrote:
>>>>
>>>> On Thu, Feb 17, 2022 at 03:39:48PM +0800, Jason Wang wrote:
>>>>> On Thu, Feb 17, 2022 at 3:36 PM Michael S. Tsirkin <mst@redhat.com> wrote:
>>>>>>
>>>>>> On Thu, Feb 17, 2022 at 03:34:13PM +0800, Jason Wang wrote:
>>>>>>> On Thu, Feb 17, 2022 at 10:01 AM syzbot
>>>>>>> <syzbot+1e3ea63db39f2b4440e0@syzkaller.appspotmail.com> wrote:
>>>>>>>>
>>>>>>>> Hello,
>>>>>>>>
>>>>>>>> syzbot found the following issue on:
>>>>>>>>
>>>>>>>> HEAD commit:    c5d9ae265b10 Merge tag 'for-linus' of git://git.kernel.org..
>>>>>>>> git tree:       upstream
>>>>>>>> console output: https://urldefense.com/v3/__https://syzkaller.appspot.com/x/log.txt?x=132e687c700000__;!!ACWV5N9M2RV99hQ!fLqQTyosTBm7FK50IVmo0ozZhsvUEPFCivEHFDGU3GjlAHDWl07UdOa-t9uf9YisMihn$
>>>>>>>> kernel config:  https://urldefense.com/v3/__https://syzkaller.appspot.com/x/.config?x=a78b064590b9f912__;!!ACWV5N9M2RV99hQ!fLqQTyosTBm7FK50IVmo0ozZhsvUEPFCivEHFDGU3GjlAHDWl07UdOa-t9uf9RjOhplp$
>>>>>>>> dashboard link: https://urldefense.com/v3/__https://syzkaller.appspot.com/bug?extid=1e3ea63db39f2b4440e0__;!!ACWV5N9M2RV99hQ!fLqQTyosTBm7FK50IVmo0ozZhsvUEPFCivEHFDGU3GjlAHDWl07UdOa-t9uf9bBf5tv0$
>>>>>>>> compiler:       gcc (Debian 10.2.1-6) 10.2.1 20210110, GNU ld (GNU Binutils for Debian) 2.35.2
>>>>>>>>
>>>>>>>> Unfortunately, I don't have any reproducer for this issue yet.
>>>>>>>>
>>>>>>>> IMPORTANT: if you fix the issue, please add the following tag to the commit:
>>>>>>>> Reported-by: syzbot+1e3ea63db39f2b4440e0@syzkaller.appspotmail.com
>>>>>>>>
>>>>>>>> WARNING: CPU: 1 PID: 10828 at drivers/vhost/vhost.c:715 vhost_dev_cleanup+0x8b8/0xbc0 drivers/vhost/vhost.c:715
>>>>>>>> Modules linked in:
>>>>>>>> CPU: 0 PID: 10828 Comm: syz-executor.0 Not tainted 5.17.0-rc4-syzkaller-00051-gc5d9ae265b10 #0
>>>>>>>> Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011
>>>>>>>> RIP: 0010:vhost_dev_cleanup+0x8b8/0xbc0 drivers/vhost/vhost.c:715
>>>>>>>
>>>>>>> Probably a hint that we are missing a flush.
>>>>>>>
>>>>>>> Looking at vhost_vsock_stop() that is called by vhost_vsock_dev_release():
>>>>>>>
>>>>>>> static int vhost_vsock_stop(struct vhost_vsock *vsock)
>>>>>>> {
>>>>>>> size_t i;
>>>>>>>         int ret;
>>>>>>>
>>>>>>>         mutex_lock(&vsock->dev.mutex);
>>>>>>>
>>>>>>>         ret = vhost_dev_check_owner(&vsock->dev);
>>>>>>>         if (ret)
>>>>>>>                 goto err;
>>>>>>>
>>>>>>> Where it could fail so the device is not actually stopped.
>>>>>>>
>>>>>>> I wonder if this is something related.
>>>>>>>
>>>>>>> Thanks
>>>>>>
>>>>>>
>>>>>> But then if that is not the owner then no work should be running, right?
>>>>>
>>>>> Could it be a buggy user space that passes the fd to another process
>>>>> and changes the owner just before the mutex_lock() above?
>>>>>
>>>>> Thanks
>>>>
>>>> Maybe, but can you be a bit more explicit? what is the set of
>>>> conditions you see that can lead to this?
>>>
>>> I think the issue could be in the vhost_vsock_stop() as Jason mentioned,
>>> but not related to fd passing, but related to the do_exit() function.
>>>
>>> Looking the stack trace, we are in exit_task_work(), that is called
>>> after exit_mm(), so the vhost_dev_check_owner() can fail because
>>> current->mm should be NULL at that point.
>>>
>>> It seems the fput work is queued by fput_many() in a worker queue, and
>>> in some cases (maybe a lot of files opened?) the work is still queued
>>> when we enter in do_exit().
>> It normally happens if userspace doesn't do a close() when the VM
>
>Just one clarification. I meant to say it "always" happens when userspace
>doesn't do a close.
>
>It doesn't have anything to do with lots of files or something like that.
>We are actually running the vhost device's release function from
>do_exit->task_work_run and so all those __fputs are done from something
>like qemu's context (current == that process).
>
>We are *not* hitting the case:
>
>do_exit->exit_files->put_files_struct->filp_close->fput->fput_many
>
>and then in there hitting the schedule_delayed_work path. For that
>the last __fput would be done from a workqueue thread and so the current
>pointer would point to a completely different thread.
>
>
>
>> is shutdown and instead let's the kernel's reaper code cleanup. The qemu
>> vhost-scsi code doesn't do a close() during shutdown and so this is our
>> normal code path. It also happens when something like qemu is not
>> gracefully shutdown like during a crash.
>>
>> So fire up qemu, start IO, then crash it or kill 9 it while IO is still
>> running and you can hit it.

Thank you very much for this detailed explanation!

>>
>>>
>>> That said, I don't know if we can simply remove that check in
>>> vhost_vsock_stop(), or check if current->mm is NULL, to understand if
>>> the process is exiting.
>>>
>>
>> Should the caller do the vhost_dev_check_owner or tell vhost_vsock_stop
>> when to check?
>>
>> - vhost_vsock_dev_ioctl always wants to check for ownership right?
>>
>> - For vhost_vsock_dev_release ownership doesn't matter because we
>> always want to clean up or it doesn't hurt too much.
>>
>> For the case where we just do open then close and no ioctls then
>> running vhost_vq_set_backend in vhost_vsock_stop is just a minor
>> hit of extra work. If we've done ioctls, but are now in
>> vhost_vsock_dev_release then we know for the graceful and ungraceful
>> case that nothing is going to be accessing this device in the future
>> and it's getting completely freed so we must completely clean it up.

Yep, I think the easiest way is to add a parameter to vhost_vsock_stop() 
to tell when to call vhost_dev_check_owner() or not.  This is because 
dev->mm is protected by dev->mutex, acquired in vhost_vsock_stop().

I will send a patch right away, it would be great if you can review.

Thanks,
Stefano