Re: [PATCH] RDMA/nldev: add mutual exclusion in nldev_dellink()

[David Ahern <dsahern@kernel.org>]

On 5/14/26 5:50 AM, Jason Gunthorpe wrote:
> On Thu, May 14, 2026 at 03:31:22PM +0800, Edward Adam Davis wrote:
>> On Wed, 13 May 2026 20:46:55 -0300, Jason Gunthorpe wrote:
>>> On Wed, May 13, 2026 at 02:17:28PM -0400, Leon Romanovsky wrote:
>>>>
>>>> On Thu, 07 May 2026 20:50:10 +0800, Edward Adam Davis wrote:
>>>>> We must serialize calls to nldev_dellink() or risk a crash as syzbot
>>>>> reported:
>>>>>
>>>>> Call Trace:
>>>>>  udp_tunnel_sock_release+0x6d/0x80 net/ipv4/udp_tunnel_core.c:197
>>>>>  rxe_release_udp_tunnel drivers/infiniband/sw/rxe/rxe_net.c:294 [inline]
>>>>>  rxe_sock_put drivers/infiniband/sw/rxe/rxe_net.c:639 [inline]
>>>>>  rxe_net_del+0xfb/0x290 drivers/infiniband/sw/rxe/rxe_net.c:660
>>>>>  rxe_dellink+0x15/0x20 drivers/infiniband/sw/rxe/rxe.c:254
>>>>>
>>>>> [...]
>>>>
>>>> Applied, thanks!
>>>>
>>>> [1/1] RDMA/nldev: add mutual exclusion in nldev_dellink()
>>>>       https://git.kernel.org/rdma/rdma/c/0b28000b64f40d
>>>
>>> This seems like a rxe bug, I would have expected the lock to be inside
>>> rxe to protect its racy implementation of rxe_net_del(), which looks
>>> like it is possibly also triggered by NETDEV_UNREGISTER...
>> No, it was triggered by RDMA_NLDEV_CMD_DELLINK, you can see the "call trace".

Not that Jason's point. Code wise

rxe_dellink -> rxe_net_del

netdev NETDEV_UNREGISTER:
 rxe_notify -> rxe_net_del

both can lead to the same problem

>>>
>>> ie it should not change nldev_dellink().
>> While this could be fixed within RXE, the same issue affects all other
>> RXE-like submodules when they subsequently support the "dellink" interface,
>> therefore, handling this within nldev_dellink() is relatively more appropriate.
> 
> Why would other modules have an issue? The problem is rxe's racey
> refcounting scheme for its lazy socket creation. There is nothing
> wrong with nldev, and now you've created some nasty BKL in the nldev
> code to fix rxe while ignoring its other races.

+1