Re: [PATCH 4/6] usbip: fix stub_dev usbip_sockfd_store() races leading to gpf

[Shuah Khan <skhan@linuxfoundation.org>]

On 3/8/21 3:10 AM, Tetsuo Handa wrote:
> On 2021/03/08 16:35, Tetsuo Handa wrote:
>> On 2021/03/08 12:53, Shuah Khan wrote:
>>> Fix the above problems:
>>> - Stop using kthread_get_run() macro to create/start threads.
>>> - Create threads and get task struct reference.
>>> - Add kthread_create() failure handling and bail out.
>>> - Hold usbip_device lock to update local and shared states after
>>>    creating rx and tx threads.
>>> - Update usbip_device status to SDEV_ST_USED.
>>> - Update usbip_device tcp_socket, sockfd, tcp_rx, and tcp_tx
>>> - Start threads after usbip_device (tcp_socket, sockfd, tcp_rx, tcp_tx,
>>>    and status) is complete.
>>
>> No, the whole usbip_sockfd_store() etc. should be serialized using a mutex,
>> for two different threads can open same file and write the same content at
>> the same moment. This results in seeing SDEV_ST_AVAILABLE and creating two
>> threads and overwiting global variables and setting SDEV_ST_USED and starting
>> two threads by each of two thread, which will later fail to call kthread_stop()
>> on one of two thread because global variables are overwritten.
>>
>> kthread_crate() (which involves GFP_KERNEL allocation) can take long time
>> enough to hit
>>
>>    usbip_sockfd_store() must perform
>>
>>        if (sdev->ud.status != SDEV_ST_AVAILABLE) {
> 
> Oops. This is
> 
> 	if (sdev->ud.status == SDEV_ST_AVAILABLE) {
> 
> of course.
> 
>>          /* misc assignments for attach operation */
>>          sdev->ud.status = SDEV_ST_USED;
>>        }
>>
>>    under a lock, or multiple ud->tcp_{tx,rx} are created (which will later
>>    cause a crash like [1]) and refcount on ud->tcp_socket is leaked when
>>    usbip_sockfd_store() is concurrently called.
>>
>> problem. That's why my patch introduced usbip_event_mutex lock.
>>
> 
> And I think that same serialization is required between "rh_port_connect() from attach_store()" and
> "rh_port_disconnect() from vhci_shutdown_connection() via usbip_event_add(&vdev->ud, VDEV_EVENT_DOWN)
>   from vhci_port_disconnect() from detach_store()", for both vhci_rx_pdu() from vhci_rx_loop() and
> vhci_port_disconnect() from detach_store() can queue VDEV_EVENT_DOWN event which can be processed
> without waiting for attach_store() to complete.
> 

Yes. We might need synchronization between events, threads, and shutdown
in usbip_host side and in connection polling and threads in vhci.

I am also looking at the shutdown sequences closely as well since the
local state is referenced without usbip_device lock in these paths.

I am approaching these problems as peeling the onion an expression so
we can limit the changes and take a spot fix approach. We have the
goal to address these crashes and not introduce regressions.

I don't seem to be able to reproduce these problems consistently in my
env. with the reproducer. I couldn't reproduce them in normal case at
all. Hence, the this cautious approach that reduces the chance of
regressions and if we see regressions, they can fixed easily.

https://syzkaller.appspot.com/text?tag=ReproC&x=14801034d00000

If this patch series fixes the problems you are seeing, I would like
get these fixes in and address the other two potential race conditions
in another round of patches. I also want to soak these in the next
for a few weeks.

Please let me know if these patches fix the problems you are seeing in 
your env.

thanks,
-- Shuah