Re: [PATCH v3 1/1] tcp/dcpp: Un-pin tw_timer

[Paolo Abeni]

On Fri, 2024-03-22 at 21:58 +0100, Valentin Schneider wrote:
> On 21/03/24 20:03, Paolo Abeni wrote:
> > Something alike the following (completely untested!!!):
> > 
> > WDYT?
> 
> Thanks for the suggestion! I've been preempted by other things and haven't
> had time to think more about this, so I really appreciate it :)
> 
> > ---
> > diff --git a/include/net/inet_timewait_sock.h b/include/net/inet_timewait_sock.h
> > index f28da08a37b4..d696d10dc8ae 100644
> > --- a/include/net/inet_timewait_sock.h
> > +++ b/include/net/inet_timewait_sock.h
> > @@ -93,8 +93,10 @@ struct inet_timewait_sock *inet_twsk_alloc(const struct sock *sk,
> >  					   struct inet_timewait_death_row *dr,
> >  					   const int state);
> >  
> > -void inet_twsk_hashdance(struct inet_timewait_sock *tw, struct sock *sk,
> > -			 struct inet_hashinfo *hashinfo);
> > +void inet_twsk_hashdance_schedule(struct inet_timewait_sock *tw,
> > +				  struct sock *sk,
> > +				  struct inet_hashinfo *hashinfo,
> > +				  int timeo);
> >  
> >  void __inet_twsk_schedule(struct inet_timewait_sock *tw, int timeo,
> >  			  bool rearm);
> > diff --git a/net/dccp/minisocks.c b/net/dccp/minisocks.c
> > index 64d805b27add..8e108a89d8e4 100644
> > --- a/net/dccp/minisocks.c
> > +++ b/net/dccp/minisocks.c
> > @@ -58,11 +58,10 @@ void dccp_time_wait(struct sock *sk, int state, int timeo)
> >  		 * we complete the initialization.
> >  		 */
> >  		local_bh_disable();
> > -		inet_twsk_schedule(tw, timeo);
> >  		/* Linkage updates.
> >  		 * Note that access to tw after this point is illegal.
> >  		 */
> > -		inet_twsk_hashdance(tw, sk, &dccp_hashinfo);
> > +		inet_twsk_hashdance_schedule(tw, sk, &dccp_hashinfo, timeo);
> >  		local_bh_enable();
> >  	} else {
> >  		/* Sorry, if we're out of memory, just CLOSE this
> > diff --git a/net/ipv4/inet_timewait_sock.c b/net/ipv4/inet_timewait_sock.c
> > index e8de45d34d56..dd314b06c0cd 100644
> > --- a/net/ipv4/inet_timewait_sock.c
> > +++ b/net/ipv4/inet_timewait_sock.c
> > @@ -97,8 +97,10 @@ static void inet_twsk_add_node_rcu(struct inet_timewait_sock *tw,
> >   * Essentially we whip up a timewait bucket, copy the relevant info into it
> >   * from the SK, and mess with hash chains and list linkage.
> >   */
> > -void inet_twsk_hashdance(struct inet_timewait_sock *tw, struct sock *sk,
> > -			   struct inet_hashinfo *hashinfo)
> > +void inet_twsk_hashdance_schedule(struct inet_timewait_sock *tw,
> > +				  struct sock *sk,
> > +				  struct inet_hashinfo *hashinfo,
> > +				  int timeo)
> >  {
> >  	const struct inet_sock *inet = inet_sk(sk);
> >  	const struct inet_connection_sock *icsk = inet_csk(sk);
> > @@ -135,6 +137,8 @@ void inet_twsk_hashdance(struct inet_timewait_sock *tw, struct sock *sk,
> >  	if (__sk_nulls_del_node_init_rcu(sk))
> >  		sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
> >  
> > +	inet_twsk_schedule(tw, timeo);
> > +
> >  	spin_unlock(lock);
> > 
> 
> That arms the timer before the refcounts are set up in the tail end of
> the hashdance, which is what we have upstream ATM.
> 
> Unfortunately this relies on the timer being TIMER_PINNED and having
> softirqs disabled: the timer can only be enqueued on the local CPU, and it
> can't fire until softirqs are enabled, so refcounts can safely be updated
> after it is armed because it can't fire.
> 
> For dynamic CPU isolation we want to make this timer not-pinned, so that it
> can be scheduled on housekeeping CPUs. However that means the
> local_bh_disable() won't prevent the timer from firing, and that means the
> refcounts need to be written to before it is armed.

Ouch, right you are, I underlooked that.


> Using the ehash lock is clever though, and the first thing inet_twsk_kill()
> does is grabbing it... Maybe something like the below? It (should) prevent
> this interleaving race:
> 
>                              tcp_time_wait()
>                                inet_twsk_hashdance()
>   inet_twsk_deschedule_put()
>     del_timer_sync()
>                                inet_twsk_schedule()
> 
> whether it is sane is another thing :-)

[...]

That looks safe to me but, compared to the current code, will need an
additional WMB in tcp_time_wait() and will take the hash lock
unconditionally in inet_twsk_deschedule_put(). The latter should not be
fast-path, I'm unsure if the whole thing be acceptable from performance
perspective??? Eric WDYT?

Thanks,

Paolo