On Fri, Dec 11, 2020 at 3:35 AM Guoqing Jiang
<guoqing.jiang@xxxxxxxxxxxxxxx> wrote:
>
>
>
> On 12/10/20 15:56, Jinpu Wang wrote:
> > On Wed, Dec 9, 2020 at 5:45 PM Jack Wang <jinpu.wang@xxxxxxxxxxxxxxx> wrote:
> >>
> >> If there are many establishments/teardowns, we need to make sure
> >> we do not consume too much system memory. Thus let on going
> >> session closing to finish before accepting new connection.
> >>
> >> Inspired by commit 777dc82395de ("nvmet-rdma: occasionally flush ongoing controller teardown")
> >> Signed-off-by: Jack Wang <jinpu.wang@xxxxxxxxxxxxxxx>
> >> Reviewed-by: Guoqing Jiang <guoqing.jiang@xxxxxxxxxxxxxxx>
> > Please ignore this one, it could lead to deadlock, due to the fact
> > cma_ib_req_handler is holding
> > mutex_lock(&listen_id->handler_mutex) when calling into
> > rtrs_rdma_connect, we call close_work which will call rdma_destroy_id,
> > which
> > could try to hold the same handler_mutex, so deadlock.
> >
>
> I am wondering if nvmet-rdma has the similar issue or not, if so, maybe
> introduce a locked version of rdma_destroy_id.
>
> Thanks,
> Guoqing
No, I was wrong. I rechecked the code, it's not a valid deadlock, in
cma_ib_req_handler, the conn_id is newly created in
https://elixir.bootlin.com/linux/latest/source/drivers/infiniband/core/cma.c#L2185.
Flush_workqueue will only flush close_work for any other cm_id, but
not the newly created one conn_id, it has not associated with anything
yet.
The same applies to nvme-rdma. so it's a false alarm by lockdep.
Regards!
Jack
