[lvc-project] [PATCH RFC net] net: sockmap: avoid race between sock_map_destroy() and sk_psock_put()

[Dmitry Antipov]

On 9/11/24 7:45 PM, Cong Wang wrote:

> I guess you totally misunderstand my point. As a significant sockmap
> contributor, I am certainly aware of sockmap users. My point is that I
> needed to narrow down the problem to CONFIG_RDS when I was debugging it.

I've narrowed down the problem to possible race condition between two
functions. "Narrowing down" the problem to a 17.5Kloc-sized subsystem
is not too helpful.

> So, please let me know if you can still reproduce this after disabling
> CONFIG_RDS, because I could not reproduce it any more. If you can,
> please kindly share the stack trace without rds_* functions.

Yes, this issue requires CONFIG_RDS and CONFIG_RDS_TCP to reproduce. But
syzbot reproducer I'm working with doesn't create RDS sockets explicitly
(with 'socket(AF_RDS, ..., ...)' or so). When two options above are enabled,
the default network namespace has special kernel-space socket which is
created in 'rds_tcp_listen_init()' and (if my understanding of the namespaces
is correct) may be inherited with 'unshare(CLONE_NEWNET)'. So just enabling
these two options makes the kernel vulnerable.

So I'm still gently asking you to check whether there is a race condition
I've talked about. Hopefully this shouldn't be too hard for a significant
sockmap contributor.

Dmitry