[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20240603133158.GC8496@breakpoint.cc>
Date: Mon, 3 Jun 2024 15:31:58 +0200
From: Florian Westphal <fw@...len.de>
To: Florian Westphal <fw@...len.de>
Cc: Eric Dumazet <edumazet@...gle.com>, netdev@...r.kernel.org,
Paolo Abeni <pabeni@...hat.com>,
"David S. Miller" <davem@...emloft.net>,
Jakub Kicinski <kuba@...nel.org>, mleitner@...hat.com,
juri.lelli@...hat.com, vschneid@...hat.com, tglozar@...hat.com,
dsahern@...nel.org, bigeasy@...utronix.de, tglx@...utronix.de
Subject: Re: [PATCH net-next v6 1/3] net: tcp/dcpp: prepare for tw_timer
un-pinning
Florian Westphal <fw@...len.de> wrote:
> Eric Dumazet <edumazet@...gle.com> wrote:
> > On Mon, Jun 3, 2024 at 11:37 AM Florian Westphal <fw@...len.de> wrote:
> > > + spin_lock(lock);
> > > + if (timer_shutdown(&tw->tw_timer)) {
> > > + /* releases @lock */
> > > + __inet_twsk_kill(tw, lock);
> > > + } else {
> >
> > If we do not have a sync variant here, I think that inet_twsk_purge()
> > could return while ongoing timers are alive.
>
> Yes.
>
> We can't use sync variant, it would deadlock on ehash spinlock.
>
> > tcp_sk_exit_batch() would then possibly hit :
> >
> > WARN_ON_ONCE(!refcount_dec_and_test(&net->ipv4.tcp_death_row.tw_refcount));
> >
> > The alive timer are releasing tw->tw_dr->tw_refcount at the end of
> > inet_twsk_kill()
>
> Theoretically the tw socket can be unlinked from the tw hash already
> (inet_twsk_purge won't encounter it), but timer is still running.
>
> Only solution I see is to schedule() in tcp_sk_exit_batch() until
> tw_refcount has dropped to the expected value, i.e. something like
>
> static void tcp_wait_for_tw_timers(struct net *n)
> {
> while (refcount_read(&n->ipv4.tcp_death_row.tw_refcount) > 1))
> schedule();
> }
>
> Any better idea?
Actually, I think we can solve this in a much simpler way.
Instead of replacing:
void inet_twsk_deschedule_put(struct inet_timewait_sock *tw)
{
if (del_timer_sync(&tw->tw_timer))
inet_twsk_kill(tw);
inet_twsk_put(tw);
}
With:
spinlock_t *lock = inet_ehash_lockp(hashinfo, tw->tw_hash);
spin_lock(lock);
if (timer_shutdown(&tw->tw_timer)) {
(Which gets us into the tcp_sk_exit_batch trouble Eric points out),
we can simply add "empty" ehash lock unlock pair before calling
del_timer_sync():
void inet_twsk_deschedule_put(struct inet_timewait_sock *tw)
{
+ spinlock_t *lock = inet_ehash_lockp(hashinfo, tw->tw_hash);
+ spin_lock(lock)
+ spin_unlock(lock)
if (del_timer_sync(&tw->tw_timer))
inet_twsk_kill(tw);
inet_twsk_put(tw);
}
Rationale:
inet_twsk_deschedule_put() cannot be called before hashdance_schedule
calls refcount_set(&tw->tw_refcnt, 3).
Before this any refcount_inc_not_zero fails so we never get into
deschedule_put.
Hashdance_schedule holds the ehash lock when it sets the tw refcount.
The lock is released only after the timer is up and running.
When inet_twsk_deschedule_put() is called, and hashdance_schedule
is not yet done, the spinlock/unlock pair will guarantee that
the timer is up after the spin_unlock.
I think this is much better than the schedule loop waiting for tw_dr
refcount to drop, it mainly needs a comment to explain what this is
doing.
Thoughts?
Powered by blists - more mailing lists