[RFC] tcp/dcpp: Un-pin tw_timer

  The TCP timewait timer is proving to be problematic for setups where scheduler
CPU isolation is achieved at runtime via cpusets (as opposed to statically via
isolcpus=domains).

What happens there is a CPU goes through tcp_time_wait(), arming the time_wait
timer, then gets isolated. TCP_TIMEWAIT_LEN later, the timer fires, causing
interference for the now-isolated CPU. This is conceptually similar to the issue
described in
  e02b93124855 ("workqueue: Unbind kworkers before sending them to exit()")

Making the timer un-pinned would resolve this, as it would be queued onto
HK_FLAG_TIMER CPUs. It would Unfortunately go against
  ed2e92394589 ("tcp/dccp: fix timewait races in timer handling")
as we'd need to arm the timer after the *hashdance() to not have it fire before
we've finished setting up the timewait_socket.

However, looking into this, I cannot grok what race is fixed by having the timer
*armed* before the hashdance.

[this next segment is brought to you by Cunningham's Law]

Using [1] as an example, inet_twsk_schedule() only arms the timer and increments
the deathrow refcount, which by itself does not affect
__inet_lookup_established(). AFAICT it only comes in handy if:
1) A CPU ends up livelocking in __inet_lookup_established() (cf. [1], though per
   inet_twsk_alloc() I don't see how a timewait socket can hit the
   forever-looping conditions with how the sk_hash and addr/port pairs are copied)
2) the initialization context can be interrupted by NET_RX (it can, cf.
   cfac7f836a71 ("tcp/dccp: block bh before arming time_wait timer"))

In this scenario, we need the timer to fire to go through
  inet_twsk_kill()
    sk_nulls_del_node_init_rcu()
and break out of the loop.

Keep softirqs disabled, but make the timer un-pinned and arm it after the
hashdance. Remote CPUs may start using the timewait socket before the timer is
armed, but their execution of __inet_lookup_established() won't prevent the
arming of the timer.

This partially reverts
  ed2e92394589 ("tcp/dccp: fix timewait races in timer handling")
and
  ec94c2696f0b ("tcp/dccp: avoid one atomic operation for timewait hashdance")

Link: [1] https://lore.kernel.org/lkml/56941035.9040000@fastly.com/
Link: https://lore.kernel.org/all/ZPhpfMjSiHVjQkTk@localhost.localdomain/
Signed-off-by: Valentin Schneider <vschneid@redhat.com>
---
 net/dccp/minisocks.c          | 18 ++++++++++--------
 net/ipv4/inet_timewait_sock.c |  9 ++++-----
 net/ipv4/tcp_minisocks.c      | 18 ++++++++++--------
 3 files changed, 24 insertions(+), 21 deletions(-)
  

On Mon, Oct 16, 2023 at 3:00 PM Valentin Schneider <vschneid@redhat.com> wrote:
>
> The TCP timewait timer is proving to be problematic for setups where scheduler
> CPU isolation is achieved at runtime via cpusets (as opposed to statically via
> isolcpus=domains).
>
> What happens there is a CPU goes through tcp_time_wait(), arming the time_wait
> timer, then gets isolated. TCP_TIMEWAIT_LEN later, the timer fires, causing
> interference for the now-isolated CPU. This is conceptually similar to the issue
> described in
>   e02b93124855 ("workqueue: Unbind kworkers before sending them to exit()")
>
> Making the timer un-pinned would resolve this, as it would be queued onto
> HK_FLAG_TIMER CPUs. It would Unfortunately go against
>   ed2e92394589 ("tcp/dccp: fix timewait races in timer handling")
> as we'd need to arm the timer after the *hashdance() to not have it fire before
> we've finished setting up the timewait_socket.
>
> However, looking into this, I cannot grok what race is fixed by having the timer
> *armed* before the hashdance.

That was because :

1) the timer could expire before we had a chance to set refcnt to
a non zero value. I guess this is fine if we use an extra atomic decrement.

OR

2) another cpu could find the TW and delete it (trying to cancel the
tw_timer) before
   we could arm the timer.  ( inet_twsk_deschedule_put() is using
del_timer_sync() followed by inet_twsk_kill())

Thus the tw timer would be armed for 60 seconds, then we would have to
wait for the timer to really
get rid of the tw structure.

I think you also need to change inet_twsk_deschedule_put() logic ?

>
> [this next segment is brought to you by Cunningham's Law]
>

I guess this is not really relevant to the potential issue.

> Using [1] as an example, inet_twsk_schedule() only arms the timer and increments
> the deathrow refcount, which by itself does not affect
> __inet_lookup_established(). AFAICT it only comes in handy if:
> 1) A CPU ends up livelocking in __inet_lookup_established() (cf. [1], though per
>    inet_twsk_alloc() I don't see how a timewait socket can hit the
>    forever-looping conditions with how the sk_hash and addr/port pairs are copied)
> 2) the initialization context can be interrupted by NET_RX (it can, cf.
>    cfac7f836a71 ("tcp/dccp: block bh before arming time_wait timer"))
>
> In this scenario, we need the timer to fire to go through
>   inet_twsk_kill()
>     sk_nulls_del_node_init_rcu()
> and break out of the loop.
>
> Keep softirqs disabled, but make the timer un-pinned and arm it after the
> hashdance. Remote CPUs may start using the timewait socket before the timer is
> armed, but their execution of __inet_lookup_established() won't prevent the
> arming of the timer.

OK, I guess we can live with the following race :

CPU0

   allocates a tw, insert it in hash table

CPU1:                               finds the TW and removes it (timer
cancel does nothing)

CPU0
   arms a TW timer, lasting

>
> This partially reverts
>   ed2e92394589 ("tcp/dccp: fix timewait races in timer handling")
> and
>   ec94c2696f0b ("tcp/dccp: avoid one atomic operation for timewait hashdance")
>
> Link: [1] https://lore.kernel.org/lkml/56941035.9040000@fastly.com/
> Link: https://lore.kernel.org/all/ZPhpfMjSiHVjQkTk@localhost.localdomain/
> Signed-off-by: Valentin Schneider <vschneid@redhat.com>
> ---
>  net/dccp/minisocks.c          | 18 ++++++++++--------
>  net/ipv4/inet_timewait_sock.c |  9 ++++-----
>  net/ipv4/tcp_minisocks.c      | 18 ++++++++++--------
>  3 files changed, 24 insertions(+), 21 deletions(-)
>
> diff --git a/net/dccp/minisocks.c b/net/dccp/minisocks.c
> index 64d805b27adde..188a29a1aef49 100644
> --- a/net/dccp/minisocks.c
> +++ b/net/dccp/minisocks.c
> @@ -53,16 +53,18 @@ void dccp_time_wait(struct sock *sk, int state, int timeo)
>                 if (state == DCCP_TIME_WAIT)
>                         timeo = DCCP_TIMEWAIT_LEN;
>
> -               /* tw_timer is pinned, so we need to make sure BH are disabled
> -                * in following section, otherwise timer handler could run before
> -                * we complete the initialization.
> -                */
> -               local_bh_disable();
> -               inet_twsk_schedule(tw, timeo);
> -               /* Linkage updates.
> -                * Note that access to tw after this point is illegal.
> +               /* tw_timer is armed after the hashdance and recount update, so
> +                * we need to make sure BH are disabled in following section to
> +                * ensure the timer is armed before we handle any further skb's.
>                  */
> +              local_bh_disable();
> +
> +               // Linkage updates
>                 inet_twsk_hashdance(tw, sk, &dccp_hashinfo);
> +               inet_twsk_schedule(tw, timeo);
> +               // Access to tw after this point is illegal.
> +               inet_twsk_put(tw);
> +
>                 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 dd37a5bf68811..ba59c2c6ef4a2 100644
> --- a/net/ipv4/inet_timewait_sock.c
> +++ b/net/ipv4/inet_timewait_sock.c
> @@ -152,16 +152,15 @@ void inet_twsk_hashdance(struct inet_timewait_sock *tw, struct sock *sk,
>
>         spin_unlock(lock);
>
> -       /* tw_refcnt is set to 3 because we have :
> +       /* tw_refcnt is set to 4 because we have :
>          * - one reference for bhash chain.
>          * - one reference for ehash chain.
>          * - one reference for timer.
> +        * - One reference for ourself (our caller will release it).
>          * We can use atomic_set() because prior spin_lock()/spin_unlock()
>          * committed into memory all tw fields.
> -        * Also note that after this point, we lost our implicit reference
> -        * so we are not allowed to use tw anymore.
>          */
> -       refcount_set(&tw->tw_refcnt, 3);
> +       refcount_set(&tw->tw_refcnt, 4);
>  }
>  EXPORT_SYMBOL_GPL(inet_twsk_hashdance);
>
> @@ -207,7 +206,7 @@ struct inet_timewait_sock *inet_twsk_alloc(const struct sock *sk,
>                 tw->tw_prot         = sk->sk_prot_creator;
>                 atomic64_set(&tw->tw_cookie, atomic64_read(&sk->sk_cookie));
>                 twsk_net_set(tw, sock_net(sk));
> -               timer_setup(&tw->tw_timer, tw_timer_handler, TIMER_PINNED);
> +               timer_setup(&tw->tw_timer, tw_timer_handler, 0);
>                 /*
>                  * Because we use RCU lookups, we should not set tw_refcnt
>                  * to a non null value before everything is setup for this
> diff --git a/net/ipv4/tcp_minisocks.c b/net/ipv4/tcp_minisocks.c
> index b98d476f1594b..269d4aa14a49e 100644
> --- a/net/ipv4/tcp_minisocks.c
> +++ b/net/ipv4/tcp_minisocks.c
> @@ -324,16 +324,18 @@ void tcp_time_wait(struct sock *sk, int state, int timeo)
>                 if (state == TCP_TIME_WAIT)
>                         timeo = TCP_TIMEWAIT_LEN;
>
> -               /* tw_timer is pinned, so we need to make sure BH are disabled
> -                * in following section, otherwise timer handler could run before
> -                * we complete the initialization.
> -                */
> -               local_bh_disable();
> -               inet_twsk_schedule(tw, timeo);
> -               /* Linkage updates.
> -                * Note that access to tw after this point is illegal.
> +               /* tw_timer is armed after the hashdance and recount update, so
> +                * we need to make sure BH are disabled in following section to
> +                * ensure the timer is armed before we handle any further skb's.
>                  */
> +              local_bh_disable();
> +
> +               // Linkage updates.
>                 inet_twsk_hashdance(tw, sk, net->ipv4.tcp_death_row.hashinfo);
> +               inet_twsk_schedule(tw, timeo);
> +               // Access to tw after this point is illegal.
> +               inet_twsk_put(tw);
> +
>                 local_bh_enable();
>         } else {
>                 /* Sorry, if we're out of memory, just CLOSE this
> --
> 2.39.3
>
  

On 16/10/23 17:40, Eric Dumazet wrote:
> On Mon, Oct 16, 2023 at 3:00 PM Valentin Schneider <vschneid@redhat.com> wrote:
>>
>> The TCP timewait timer is proving to be problematic for setups where scheduler
>> CPU isolation is achieved at runtime via cpusets (as opposed to statically via
>> isolcpus=domains).
>>
>> What happens there is a CPU goes through tcp_time_wait(), arming the time_wait
>> timer, then gets isolated. TCP_TIMEWAIT_LEN later, the timer fires, causing
>> interference for the now-isolated CPU. This is conceptually similar to the issue
>> described in
>>   e02b93124855 ("workqueue: Unbind kworkers before sending them to exit()")
>>
>> Making the timer un-pinned would resolve this, as it would be queued onto
>> HK_FLAG_TIMER CPUs. It would Unfortunately go against
>>   ed2e92394589 ("tcp/dccp: fix timewait races in timer handling")
>> as we'd need to arm the timer after the *hashdance() to not have it fire before
>> we've finished setting up the timewait_socket.
>>
>> However, looking into this, I cannot grok what race is fixed by having the timer
>> *armed* before the hashdance.
>
> That was because :
>
> 1) the timer could expire before we had a chance to set refcnt to
> a non zero value. I guess this is fine if we use an extra atomic decrement.
>
> OR
>
> 2) another cpu could find the TW and delete it (trying to cancel the
> tw_timer) before
>    we could arm the timer.  ( inet_twsk_deschedule_put() is using
> del_timer_sync() followed by inet_twsk_kill())
>
> Thus the tw timer would be armed for 60 seconds, then we would have to
> wait for the timer to really
> get rid of the tw structure.
>
> I think you also need to change inet_twsk_deschedule_put() logic ?
>

Gotcha, thank you for pointing it out.

>> Keep softirqs disabled, but make the timer un-pinned and arm it after the
>> hashdance. Remote CPUs may start using the timewait socket before the timer is
>> armed, but their execution of __inet_lookup_established() won't prevent the
>> arming of the timer.
>
> OK, I guess we can live with the following race :
>
> CPU0
>
>    allocates a tw, insert it in hash table
>
> CPU1:                               finds the TW and removes it (timer
> cancel does nothing)
>
> CPU0
>    arms a TW timer, lasting
>

Looks reasonable to me, I'll go write v2.

Thanks for the help!
  

On Wed, Oct 18, 2023 at 4:57 PM Valentin Schneider <vschneid@redhat.com> wrote:

>
> Looks reasonable to me, I'll go write v2.
>
> Thanks for the help!

Sure thing !

BTW, we also use TIMER_PINNED for req->rsk_timer, are you working on it too ?
  

On 18/10/23 17:00, Eric Dumazet wrote:
> On Wed, Oct 18, 2023 at 4:57 PM Valentin Schneider <vschneid@redhat.com> wrote:
>
>>
>> Looks reasonable to me, I'll go write v2.
>>
>> Thanks for the help!
>
> Sure thing !
>
> BTW, we also use TIMER_PINNED for req->rsk_timer, are you working on it too ?

Ah, no, that wasn't on my radar. This hasn't shown up on our systems
yet. From a cursory look it does look like it could lead to similar issues,
I'll add that to my todolist. Thanks!