Re: [bug report] RDMA/iwpm: reentrant iwpm hello message

[Leon Romanovsky <leon@kernel.org>]

On Tue, Dec 24, 2024 at 06:51:27PM +0800, Lin Ma wrote:
> Hello Leon,
> 
> > 
> > I'm not fully understand the lockdep here. We use down_read(), which is
> > reentry safe.
> > 
> 
> Really? To my knowledge, though down_read() itself will not trigger locking
> errors. But below scenario will lead to deadlock, and that's why this
> WARNING is raised.
> 
>    CPU0                CPU1
>    ----                ----
> down_read()[1]
>                   down_write()[2]
> down_read()[3]
> 
> If CPU1 thread not exists, the CPU0 will run smoothly (However, it will keep
> looping and the PoC cannot be killed by any signal, causing Denial-of-Service).
> 
> When CPU1 calls down_write(), it will wait for [1] to be released.
> However, when [3] is called, it will then wait for [2] to be released,
> leading to a deadlock situation.
> 
> Please let me know if I understand this correctly or incorrectly?

The thing is that down_write() is called when we unregistering module
which sent netlink messages. It shouldn't happen.

Thanks

> 
> Thanks,
> Lin