Re: [RFC] net: inet: Potential sleep in atomic context in inet_twsk_hashdance_schedule on PREEMPT_RT

[Eric Dumazet]

On Mon, Aug 18, 2025 at 9:46 AM Yunseong Kim <ysk@xxxxxxxxxxx> wrote:
>
> Hi everyone,
>
> I'm looking at the inet_twsk_hashdance_schedule() function in
> net/ipv4/inet_timewait_sock.c and noticed a pattern that could be
> problematic for PREEMPT_RT kernels.
>
> The code in question is:
>
>  void inet_twsk_hashdance_schedule(struct inet_timewait_sock *tw,
>                                    struct sock *sk,
>                                    struct inet_hashinfo *hashinfo,
>                                    int timeo)
>  {
>      ...
>      local_bh_disable();
>      spin_lock(&bhead->lock);

Note this pattern is quite common, you should look at other instances
like inet_put_port(),
inet_csk_listen_stop(), __inet_hash(), __tcp_close(), tcp_abort(),

>      spin_lock(&bhead2->lock);
>      ...
>  }
>
> The sequence local_bh_disable() followed by spin_lock(), In a PREEMPT_RT
> enabled kernel, spin_lock() is replaced by a mutex that an sleep.
> However, local_bh_disable() creates an atomic context by incrementing
> preempt_count, where sleeping is forbidden.
>
> If the spinlock is contended, this code would attempt to sleep inside an
> atomic context, leading to a "BUG: sleeping function called from invalid
> context" kernel panic.
>
> While this pattern is correct for non-RT kernels (and is essentially what
> spin_lock_bh() expands to), it causes critical issues in an RT environment.
>
> A possible fix would be to replace this sequence with calls to
> spin_lock_bh(). Given that two separate locks are acquired, the most direct
> change would look like this:
>
>  // local_bh_disable();  <- removed
>  spin_lock_bh(&bhead->lock);
>  // The second lock is already protected from BH by the first one
>  spin_lock(&bhead2->lock);
>
> Or, to be more explicit and safe if the logic ever changes:
>
>  spin_lock_bh(&bhead->lock);
>  spin_lock_bh(&bhead2->lock);
>
> However, since spin_lock_bh() on the first lock already disables bottom
> halves, the second lock only needs to be a plain spin_lock().
>
> I would like to ask for your thoughts on this. Is my understanding correct,
> and would a patch to change this locking pattern be welcome?
>
> It's possible the PREEMPT_RT implications were not a primary concern at
> the time.
>
> Thanks for your time and guidance.
>
> Best regards,
> Yunseong Kim