Hi,

I came up with this patch to centralize netns refcounting on kernel sockets,
because `sk_net_refcnt = !kern' is not enough anymore.

The idea is simply to remove the responsibility of a module outside of net/
to have to deal with sockets internals (cf. sk_net_refcnt_upgrade()).

It adds an anonymous enum (just for named values) SOCK_NETNS_REFCNT_* that
can be passed to __sock_create() and sk_alloc() through the @kern arg.
(this was much easier and shorter than e.g. adding another arg)

A sock_create_netns() wrapper is added, for callers who need such refcounting
(e.g. current callers of sk_net_refcnt_upgrade()).

And then, the core change is quite simple in sk_alloc() -- sk_net_refcnt is
set only if it's a user socket, or
(@kern == SOCK_NETNS_REFCNT_KERN_ANY && @net != inet_net).

I have the patches that modifies current users of sk_net_refcnt_upgrade() to
create their sockets with sock_create_netns(), if anyone wants to test or
this gets merged.

I could confirm this works only on cifs, though, by using Kuniyuki's reproducer
in [0], which is quite reliable.  Unfortunately, I don't know enough about the
other modules and/or how to trigger this same bug on those, but I'll be happy
to test it if I can get instructions.


Cheers,

Enzo


[0] - https://lore.kernel.org/linux-cifs/20241031175709.20111-1-kuniyu@amazon.com/

Enzo Matsumiya (1):
  net: socket: hint netns refcounting through @kern arg

 include/linux/net.h | 15 +++++++++++++++
 net/core/sock.c     | 10 ++++++----
 net/socket.c        | 27 +++++++++++++++++++++++++--
 3 files changed, 46 insertions(+), 6 deletions(-)

-- 
2.48.1

From: Enzo Matsumiya <ematsumiya@suse.de>
Date: Fri, 16 May 2025 19:09:18 -0300
> Hi,
> 
> I came up with this patch to centralize netns refcounting on kernel sockets,
> because `sk_net_refcnt = !kern' is not enough anymore.
> 
> The idea is simply to remove the responsibility of a module outside of net/
> to have to deal with sockets internals (cf. sk_net_refcnt_upgrade()).
> 
> It adds an anonymous enum (just for named values) SOCK_NETNS_REFCNT_* that
> can be passed to __sock_create() and sk_alloc() through the @kern arg.
> (this was much easier and shorter than e.g. adding another arg)
> 
> A sock_create_netns() wrapper is added, for callers who need such refcounting
> (e.g. current callers of sk_net_refcnt_upgrade()).
> 
> And then, the core change is quite simple in sk_alloc() -- sk_net_refcnt is
> set only if it's a user socket, or
> (@kern == SOCK_NETNS_REFCNT_KERN_ANY && @net != inet_net).
> 
> I have the patches that modifies current users of sk_net_refcnt_upgrade() to
> create their sockets with sock_create_netns(), if anyone wants to test or
> this gets merged.
> 
> I could confirm this works only on cifs, though, by using Kuniyuki's reproducer
> in [0], which is quite reliable.  Unfortunately, I don't know enough about the
> other modules and/or how to trigger this same bug on those, but I'll be happy
> to test it if I can get instructions.
> 

I was waiting for b013b817f32f to land on net-next to respin
this series [0] as propised in [1].

[0]: https://lore.kernel.org/netdev/20241213092152.14057-1-kuniyu@amazon.com/#t
[1]: https://lore.kernel.org/lkml/20250409190031.38942-1-kuniyu@amazon.com/

So, let me respin it.

FWIW, I posted the most identical patch before the series.
https://lore.kernel.org/netdev/20240227011041.97375-4-kuniyu@amazon.com/