Received-SPF: pass (google.com: domain of linux-kernel+bounces-145366-linux.lists.archive=gmail.com@vger.kernel.org designates 139.178.88.99 as permitted sender) client-ip=139.178.88.99;
From: Valentin Schneider <vschneid@redhat.com>
To: Eric Dumazet <edumazet@google.com>
Cc: dccp@vger.kernel.org, netdev@vger.kernel.org,
 linux-kernel@vger.kernel.org, linux-rt-users@vger.kernel.org, "David S.
 Miller" <davem@davemloft.net>, Jakub Kicinski <kuba@kernel.org>, Paolo
 Abeni <pabeni@redhat.com>, mleitner@redhat.com, David Ahern
 <dsahern@kernel.org>, Juri Lelli <juri.lelli@redhat.com>, Tomas Glozar
 <tglozar@redhat.com>, Sebastian Andrzej Siewior <bigeasy@linutronix.de>,
 Thomas Gleixner <tglx@linutronix.de>
Subject: Re: [PATCH v5 0/2] tcp/dcpp: Un-pin tw_timer
In-Reply-To: <CANn89iJYX8e_3Or9a5Q55NuQ8ZAHfYL+p_SpM0yz91sdj4HqtQ@mail.gmail.com>
References: <20240415113436.3261042-1-vschneid@redhat.com>
 <CANn89iJYX8e_3Or9a5Q55NuQ8ZAHfYL+p_SpM0yz91sdj4HqtQ@mail.gmail.com>
Date: Mon, 15 Apr 2024 16:33:44 +0200
Message-ID: <xhsmhmspu8zlj.mognet@vschneid-thinkpadt14sgen2i.remote.csb>
Precedence: bulk
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: quoted-printable

On 15/04/24 14:35, Eric Dumazet wrote:
> On Mon, Apr 15, 2024 at 1:34=E2=80=AFPM Valentin Schneider <vschneid@redh=
at.com> wrote:
>>
>> Hi,
>>
>> This is v5 of the series where the tw_timer is un-pinned to get rid of
>> interferences in isolated CPUs setups.
>>
>> The first patch is a new one stemming from Jakub's bug reported. It's th=
ere
>> mainly to make the reviewing a bit easier, but as it changes behaviour i=
t should
>> be squashed with the second one.
>>
>> Revisions
>> =3D=3D=3D=3D=3D=3D=3D=3D=3D
>>
>> v4 -> v5
>> ++++++++
>>
>> o Rebased against latest Linus' tree
>> o Converted tw_timer into a delayed work following Jakub's bug report on=
 v4
>>   http://lore.kernel.org/r/20240411100536.224fa1e7@kernel.org
>
> What was the issue again ?
>
> Please explain precisely why it was fundamentally tied to the use of
> timers (and this was not possible to fix the issue without
> adding work queues and more dependencies to TCP stack)

In v4 I added the use of the ehash lock to serialize arming the timewait
timer vs destroying it (inet_twsk_schedule() vs inet_twsk_deschedule_put()).

Unfortunately, holding a lock both in a timer callback and in the context
in which it is destroyed is invalid. AIUI the issue is as follows:

  CPUx                        CPUy
  spin_lock(foo);
                              <timer fires>
                              call_timer_fn()
                                spin_lock(foo) // blocks
  timer_shutdown_sync()
    __timer_delete_sync()
      __try_to_del_timer_sync() // looped as long as timer is running
                       <deadlock>

In our case, we had in v4:

  inet_twsk_deschedule_put()
    spin_lock(ehash_lock);
                                          tw_timer_handler()
                                            inet_twsk_kill()
                                              spin_lock(ehash_lock);
                                              __inet_twsk_kill();
    timer_shutdown_sync(&tw->tw_timer);

The fix here is to move the timer deletion to a non-timer
context. Workqueues fit the bill, and as the tw_timer_handler() would just =
queue
a work item, I converted it to a delayed_work.