Re: [PATCH v2 net] net: sch_generic: aviod concurrent reset and enqueue op for lockless qdisc

[Pai, Vishwanath]

On 10/28/20, 10:51 AM, "Vishwanath Pai" <pai.vishwain.kernel@xxxxxxxxx> wrote:

On Thu, Sep 17, 2020 at 4:26 PM Cong Wang <xiyou.wangcong@xxxxxxxxx> wrote:
>
> On Fri, Sep 11, 2020 at 1:13 AM Yunsheng Lin <linyunsheng@xxxxxxxxxx> wrote:
> >
> > On 2020/9/11 4:07, Cong Wang wrote:
> > > On Tue, Sep 8, 2020 at 4:06 AM Yunsheng Lin <linyunsheng@xxxxxxxxxx> wrote:
> > >>
> > >> Currently there is concurrent reset and enqueue operation for the
> > >> same lockless qdisc when there is no lock to synchronize the
> > >> q->enqueue() in __dev_xmit_skb() with the qdisc reset operation in
> > >> qdisc_deactivate() called by dev_deactivate_queue(), which may cause
> > >> out-of-bounds access for priv->ring[] in hns3 driver if user has
> > >> requested a smaller queue num when __dev_xmit_skb() still enqueue a
> > >> skb with a larger queue_mapping after the corresponding qdisc is
> > >> reset, and call hns3_nic_net_xmit() with that skb later.
> > >>
> > >> Reused the existing synchronize_net() in dev_deactivate_many() to
> > >> make sure skb with larger queue_mapping enqueued to old qdisc(which
> > >> is saved in dev_queue->qdisc_sleeping) will always be reset when
> > >> dev_reset_queue() is called.
> > >>
> > >> Fixes: 6b3ba9146fe6 ("net: sched: allow qdiscs to handle locking")
> > >> Signed-off-by: Yunsheng Lin <linyunsheng@xxxxxxxxxx>
> > >> ---
> > >> ChangeLog V2:
> > >>         Reuse existing synchronize_net().
> > >> ---
> > >>  net/sched/sch_generic.c | 48 +++++++++++++++++++++++++++++++++---------------
> > >>  1 file changed, 33 insertions(+), 15 deletions(-)
> > >>
> > >> diff --git a/net/sched/sch_generic.c b/net/sched/sch_generic.c
> > >> index 265a61d..54c4172 100644
> > >> --- a/net/sched/sch_generic.c
> > >> +++ b/net/sched/sch_generic.c
> > >> @@ -1131,24 +1131,10 @@ EXPORT_SYMBOL(dev_activate);
> > >>
> > >>  static void qdisc_deactivate(struct Qdisc *qdisc)
> > >>  {
> > >> -       bool nolock = qdisc->flags & TCQ_F_NOLOCK;
> > >> -
> > >>         if (qdisc->flags & TCQ_F_BUILTIN)
> > >>                 return;
> > >> -       if (test_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state))
> > >> -               return;
> > >> -
> > >> -       if (nolock)
> > >> -               spin_lock_bh(&qdisc->seqlock);
> > >> -       spin_lock_bh(qdisc_lock(qdisc));
> > >>
> > >>         set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
> > >> -
> > >> -       qdisc_reset(qdisc);
> > >> -
> > >> -       spin_unlock_bh(qdisc_lock(qdisc));
> > >> -       if (nolock)
> > >> -               spin_unlock_bh(&qdisc->seqlock);
> > >>  }
> > >>
> > >>  static void dev_deactivate_queue(struct net_device *dev,
> > >> @@ -1165,6 +1151,30 @@ static void dev_deactivate_queue(struct net_device *dev,
> > >>         }
> > >>  }
> > >>
> > >> +static void dev_reset_queue(struct net_device *dev,
> > >> +                           struct netdev_queue *dev_queue,
> > >> +                           void *_unused)
> > >> +{
> > >> +       struct Qdisc *qdisc;
> > >> +       bool nolock;
> > >> +
> > >> +       qdisc = dev_queue->qdisc_sleeping;
> > >> +       if (!qdisc)
> > >> +               return;
> > >> +
> > >> +       nolock = qdisc->flags & TCQ_F_NOLOCK;
> > >> +
> > >> +       if (nolock)
> > >> +               spin_lock_bh(&qdisc->seqlock);
> > >> +       spin_lock_bh(qdisc_lock(qdisc));
> > >
> > >
> > > I think you do not need this lock for lockless one.
> >
> > It seems so.
> > Maybe another patch to remove qdisc_lock(qdisc) for lockless
> > qdisc?
>
> Yeah, but not sure if we still want this lockless qdisc any more,
> it brings more troubles than gains.
>
> >
> >
> > >
> > >> +
> > >> +       qdisc_reset(qdisc);
> > >> +
> > >> +       spin_unlock_bh(qdisc_lock(qdisc));
> > >> +       if (nolock)
> > >> +               spin_unlock_bh(&qdisc->seqlock);
> > >> +}
> > >> +
> > >>  static bool some_qdisc_is_busy(struct net_device *dev)
> > >>  {
> > >>         unsigned int i;
> > >> @@ -1213,12 +1223,20 @@ void dev_deactivate_many(struct list_head *head)
> > >>                 dev_watchdog_down(dev);
> > >>         }
> > >>
> > >> -       /* Wait for outstanding qdisc-less dev_queue_xmit calls.
> > >> +       /* Wait for outstanding qdisc-less dev_queue_xmit calls or
> > >> +        * outstanding qdisc enqueuing calls.
> > >>          * This is avoided if all devices are in dismantle phase :
> > >>          * Caller will call synchronize_net() for us
> > >>          */
> > >>         synchronize_net();
> > >>
> > >> +       list_for_each_entry(dev, head, close_list) {
> > >> +               netdev_for_each_tx_queue(dev, dev_reset_queue, NULL);
> > >> +
> > >> +               if (dev_ingress_queue(dev))
> > >> +                       dev_reset_queue(dev, dev_ingress_queue(dev), NULL);
> > >> +       }
> > >> +
> > >>         /* Wait for outstanding qdisc_run calls. */
> > >>         list_for_each_entry(dev, head, close_list) {
> > >>                 while (some_qdisc_is_busy(dev)) {
> > >
> > > Do you want to reset before waiting for TX action?
> > >
> > > I think it is safer to do it after, at least prior to commit 759ae57f1b
> > > we did after.
> >
> > The reference to the txq->qdisc is always protected by RCU, so the synchronize_net()
> > should be enought to ensure there is no skb enqueued to the old qdisc that is saved
> > in the dev_queue->qdisc_sleeping, because __dev_queue_xmit can only see the new qdisc
> > after synchronize_net(), which is noop_qdisc, and noop_qdisc will make sure any skb
> > enqueued to it will be dropped and freed, right?
>
> Hmm? In net_tx_action(), we do not hold RCU read lock, and we do not
> reference qdisc via txq->qdisc but via sd->output_queue.
>
>
> >
> > If we do any additional reset that is not related to qdisc in dev_reset_queue(), we
> > can move it after some_qdisc_is_busy() checking.
>
> I am not suggesting to do an additional reset, I am suggesting to move
> your reset after the busy waiting.
>
> Thanks.

Hi,

We noticed some problems when testing the latest 5.4 LTS kernel and traced it
back to this commit using git bisect. When running our tests the machine stops
responding to all traffic and the only way to recover is a reboot. I do not see
a stack trace on the console.

This can be reproduced using the packetdrill test below, it should be run a
few times or in a loop. You should hit this issue within a few tries but
sometimes might take up to 15-20 tries.

*** TEST BEGIN ***

0 `echo 4 > /proc/sys/net/ipv4/tcp_min_tso_segs`

0.400 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3
0.400 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0

// set maxseg to 1000 to work with both ipv4 and ipv6
0.500 setsockopt(3, SOL_TCP, TCP_MAXSEG, [1000], 4) = 0
0.500 bind(3, ..., ...) = 0
0.500 listen(3, 1) = 0

// Establish connection
0.600 < S 0:0(0) win 32792 <mss 1000,sackOK,nop,nop,nop,wscale 5>
0.600 > S. 0:0(0) ack 1 <...>

0.800 < . 1:1(0) ack 1 win 320
0.800 accept(3, ..., ...) = 4

// Send 4 data segments.
+0 write(4, ..., 4000) = 4000
+0 > P. 1:4001(4000) ack 1

// Receive a SACK
+.1 < . 1:1(0) ack 1 win 320 <sack 1001:2001,nop,nop>

+.3 %{ print "TCP CA state: ",tcpi_ca_state  }%

*** TEST END ***

I can reproduce the issue easily on v5.4.68, and after reverting this commit it
does not happen anymore.

Thanks,
Vishwanath