Re: [PATCH net] genetlink: take netlink table lock when (un)registering

[stranche@codeaurora.org]

On 2020-06-27 12:55, Cong Wang wrote:
> On Fri, Jun 26, 2020 at 5:32 PM Sean Tranchetti 
> <stranche@codeaurora.org> wrote:
>> 
>> A potential deadlock can occur during registering or unregistering a 
>> new
>> generic netlink family between the main nl_table_lock and the cb_lock 
>> where
>> each thread wants the lock held by the other, as demonstrated below.
>> 
>> 1) Thread 1 is performing a netlink_bind() operation on a socket. As 
>> part
>>    of this call, it will call netlink_lock_table(), incrementing the
>>    nl_table_users count to 1.
>> 2) Thread 2 is registering (or unregistering) a genl_family via the
>>    genl_(un)register_family() API. The cb_lock semaphore will be taken 
>> for
>>    writing.
>> 3) Thread 1 will call genl_bind() as part of the bind operation to 
>> handle
>>    subscribing to GENL multicast groups at the request of the user. It 
>> will
>>    attempt to take the cb_lock semaphore for reading, but it will fail 
>> and
>>    be scheduled away, waiting for Thread 2 to finish the write.
>> 4) Thread 2 will call netlink_table_grab() during the (un)registration
>>    call. However, as Thread 1 has incremented nl_table_users, it will 
>> not
>>    be able to proceed, and both threads will be stuck waiting for the
>>    other.
>> 
>> To avoid this scenario, the locks should be acquired in the same order 
>> by
>> both threads. Since both the register and unregister functions need to 
>> take
>> the nl_table_lock in their processing, it makes sense to explicitly 
>> acquire
>> them before they lock the genl_mutex and the cb_lock. In 
>> unregistering, no
>> other change is needed aside from this locking change.
> 
> Like the kernel test robot reported, you can not call genl_lock_all 
> while
> holding netlink_table_grab() which is effectively a write lock.
> 
> To me, it seems genl_bind() can be just removed as there is no one
> in-tree uses family->mcast_bind(). Can you test the attached patch?
> It seems sufficient to fix this deadlock.
> 
> Thanks.

Thanks Cong. Yes, removing the genl_bind()/genl_unbind() functions 
eliminates the
potential for this deadlock. Adding Johannes here to comment on removing 
these,
as the family->mcast_bind() capability added by commit c380d9a7afff
("genetlink: pass multicast bind/unbind to families") would be lost.