Re: [PATCH bpf-next v3 1/3] bpf, sockmap: avoid using sk_socket after free when sending

["Jiayuan Chen" <jiayuan.chen@linux.dev>]

March 20, 2025 at 08:06, "Cong Wang" <xiyou.wangcong@gmail.com> wrote:

> 
> On Wed, Mar 19, 2025 at 11:36:13PM +0000, Jiayuan Chen wrote:
> 
> > 
> > 2025/3/20 07:02, "Cong Wang" <xiyou.wangcong@gmail.com> wrote:
> > 
> >  
> > 
> >  
> > 
> >  On Mon, Mar 17, 2025 at 05:22:54PM +0800, Jiayuan Chen wrote:
> > 
> >  
> > 
> >  > 
> > 
> >  > The sk->sk_socket is not locked or referenced, and during the call to
> > 
> >  > 
> > 
> >  
> > 
> >  Hm? We should have a reference in socket map, whether directly or
> > 
> >  
> > 
> >  indirectly, right? When we add a socket to a socket map, we do call
> > 
> >  
> > 
> >  sock_map_psock_get_checked() to obtain a reference.
> > 
> >  
> > 
> >  
> > 
> >  Yes, but we remove psock from sockmap when sock_map_close() was called
> > 
> >  '''
> > 
> >  sock_map_close
> > 
> >  lock_sock(sk);
> > 
> >  rcu_read_lock();
> > 
> >  psock = sk_psock(sk);
> > 
> >  // here we remove psock and the reference of psock become 0
> > 
> >  sock_map_remove_links(sk, psock)
> > 
> 
> sk_psock_drop() also calls cancel_delayed_work_sync(&psock->work),
> 
> althrough in yet another work. Is this also a contribution to this bug?
>

Maybe it's related. Calling cancel_delayed_work_sync() in sk_psock_drop()
is too late for our scenario.

To be more precise, the core goal of this patch is to prevent sock_map_close()
from executing until the backlog work completes. This is because sock_map_close()
resides in the close(fd) path, once it finishes, subsequent steps will release
the sk_socket. Therefore, performing cancellation in sk_psock_drop() is too late.

Upon reviewing historical commits, I found that the backlog work originally held
lock_sk, which naturally synchronized with lock_sk in sock_map_close. However,
when the backlog work later removed lock_sk, an alternative synchronization
mechanism(just hold psock reference like this patch) became necessary.
> > 
> > psock = sk_psock_get(sk);
> > 
> >  if (unlikely(!psock))
> > 
> >  goto no_psock; <=== jmp to no_psock
> > 
> >  rcu_read_unlock();
> > 
> >  release_sock(sk);
> > 
> >  cancel_delayed_work_sync(&psock->work); <== no chance to run cancel
> > 
> >  '''
> > 
> 
> I have to say sock_map_close() becomes harder and harder to understand
> 
> now. And I am feeling we may have more bugs since we have two flying
> 
> work's here: psock->rwork and psock->work.
> 
> Thanks.

Yes, this patch prevent sock_map_close() from executing
until the backlog work completes. This likely makes the
cancel_delayed_work in sk_psock_destroy redundant.

The code has undergone too many iterations. While sk_psock_destroy certainly
contains redundant operations, we should retain it for now. There may be
hidden dependencies we haven't fully untangled.

Thanks.