Re: [syzbot] [rdma?] possible deadlock in siw_create_listen (2)

[Jason Gunthorpe <jgg@ziepe.ca>]

On Fri, Oct 04, 2024 at 04:10:31PM +0000, Bernard Metzler wrote:

> Could one please help me to understand this situation?
> cma.c:5354
> 
>         mutex_lock(&lock);
>         list_add_tail(&cma_dev->list, &dev_list);
>         list_for_each_entry(id_priv, &listen_any_list, listen_any_item) {
>                 ret = cma_listen_on_dev(id_priv, cma_dev, &to_destroy);
>                 if (ret)
>                         goto free_listen;
>         }               
>         mutex_unlock(&lock);
> 
> siw_cm.c:1776
> 	sock_set_reuseaddr(s->sk);
>
> ...which calls lock_sock(sk) on a feshly created socket.

I think this is a smc bug, and lockdep is getting confused about what
to report due to all the different locks.

smc_setsockopt() eventually in ip_setsockopt() does:

	mutex_lock(&smc->clcsock_release_lock);

	if (needs_rtnl)
		rtnl_lock();
	sockopt_lock_sock(sk);
	mutex_unlock(&smc->clcsock_release_lock);


smc_sendmsg() does

	lock_sock(sk);
	mutex_lock(&smc->clcsock_release_lock);

Which is classic deadlock locking.

That the CMA gets involved here seems like wrong reporting because
syzkaller put those lock chains into it.

I guess this is a dup of 

https://lore.kernel.org/netdev/00000000000093078f0622583e6e@google.com/T/

Or at least that should be fixed before looking at this

Jason