Date: Wed, 20 Mar 2024 13:50:15 -0400
From: Jamal Hadi Salim <jhs@...atatu.com>
To: Eric Dumazet <edumazet@...gle.com>
Cc: renmingshuai <renmingshuai@...wei.com>, xiyou.wangcong@...il.com, jiri@...nulli.us, 
	davem@...emloft.net, vladbu@...dia.com, netdev@...r.kernel.org, 
	yanan@...wei.com, liaichun@...wei.com, caowangbao@...wei.com, 
	Eric Dumazet <eric.dumazet@...il.com>, Jakub Kicinski <kuba@...nel.org>, 
	Paolo Abeni <pabeni@...hat.com>, Victor Nogueira <victor@...atatu.com>, 
	Pedro Tammela <pctammela@...atatu.com>, Davide Caratti <dcaratti@...hat.com>
Subject: Re: [PATCH] net/sched: Forbid assigning mirred action to a filter
 attached to the egress

On Wed, Mar 20, 2024 at 1:31 PM Jamal Hadi Salim <jhs@...atatu.com> wrote:
>
> On Wed, Mar 20, 2024 at 12:58 PM Eric Dumazet <edumazet@...gle.com> wrote:
> >
> > On Tue, Mar 19, 2024 at 9:54 PM Jamal Hadi Salim <jhs@...atatu.com> wrote:
> > >
> > > On Tue, Mar 19, 2024 at 5:38 AM Eric Dumazet <edumazet@...gle.com> wrote:
> > > >
> > > > On Mon, Mar 18, 2024 at 11:05 PM Jamal Hadi Salim <jhs@...atatu.com> wrote:
> > > > >
> > > > > On Mon, Mar 18, 2024 at 3:11 PM Eric Dumazet <edumazet@...gle.com> wrote:
> > > > > >
> > > > > > On Mon, Mar 18, 2024 at 6:36 PM Jamal Hadi Salim <jhs@...atatu.com> wrote:
> > > > > > >
> > > > > > > On Mon, Mar 18, 2024 at 11:46 AM Eric Dumazet <edumazet@...gle.com> wrote:
> > > > > > > >
> > > > > > > > On Mon, Mar 18, 2024 at 3:27 PM Jamal Hadi Salim <jhs@...atatu.com> wrote:
> > > > > > > > >
> > > > > > > > > On Sun, Mar 17, 2024 at 12:10 PM Jamal Hadi Salim <jhs@...atatu.com> wrote:
> > > > > > > > > >
> > > > > > > > > > On Thu, Mar 14, 2024 at 1:14 PM Jamal Hadi Salim <jhs@...atatu.com> wrote:
> > > > > > > > > > >
> > > > > > > > > > > On Thu, Mar 14, 2024 at 7:18 AM renmingshuai <renmingshuai@...wei.com> wrote:
> > > > > > > > > > > >
> > > > > > > > > > > > As we all know the mirred action is used to mirroring or redirecting the
> > > > > > > > > > > > packet it receives. Howerver, add mirred action to a filter attached to
> > > > > > > > > > > > a egress qdisc might cause a deadlock. To reproduce the problem, perform
> > > > > > > > > > > > the following steps:
> > > > > > > > > > > > (1)tc qdisc add dev eth0 root handle 1: htb default 30 \n
> > > > > > > > > > > > (2)tc filter add dev eth2 protocol ip prio 2 flower verbose \
> > > > > > > > > > > >      action police rate 100mbit burst 12m conform-exceed jump 1 \
> > > > > > > > > > > >      / pipe mirred egress redirect dev eth2 action drop
> > > > > > > > > > > >
> > > > > > > > > > >
> > > > > > > > > > > I think you meant both to be the same device eth0 or eth2?
> > > > > > > > > > >
> > > > > > > > > > > > The stack is show as below:
> > > > > > > > > > > > [28848.883915]  _raw_spin_lock+0x1e/0x30
> > > > > > > > > > > > [28848.884367]  __dev_queue_xmit+0x160/0x850
> > > > > > > > > > > > [28848.884851]  ? 0xffffffffc031906a
> > > > > > > > > > > > [28848.885279]  tcf_mirred_act+0x3ab/0x596 [act_mirred]
> > > > > > > > > > > > [28848.885863]  tcf_action_exec.part.0+0x88/0x130
> > > > > > > > > > > > [28848.886401]  fl_classify+0x1ca/0x1e0 [cls_flower]
> > > > > > > > > > > > [28848.886970]  ? dequeue_entity+0x145/0x9e0
> > > > > > > > > > > > [28848.887464]  ? newidle_balance+0x23f/0x2f0
> > > > > > > > > > > > [28848.887973]  ? nft_lookup_eval+0x57/0x170 [nf_tables]
> > > > > > > > > > > > [28848.888566]  ? nft_do_chain+0xef/0x430 [nf_tables]
> > > > > > > > > > > > [28848.889137]  ? __flush_work.isra.0+0x35/0x80
> > > > > > > > > > > > [28848.889657]  ? nf_ct_get_tuple+0x1cf/0x210 [nf_conntrack]
> > > > > > > > > > > > [28848.890293]  ? do_select+0x637/0x870
> > > > > > > > > > > > [28848.890735]  tcf_classify+0x52/0xf0
> > > > > > > > > > > > [28848.891177]  htb_classify+0x9d/0x1c0 [sch_htb]
> > > > > > > > > > > > [28848.891722]  htb_enqueue+0x3a/0x1c0 [sch_htb]
> > > > > > > > > > > > [28848.892251]  __dev_queue_xmit+0x2d8/0x850
> > > > > > > > > > > > [28848.892738]  ? nf_hook_slow+0x3c/0xb0
> > > > > > > > > > > > [28848.893198]  ip_finish_output2+0x272/0x580
> > > > > > > > > > > > [28848.893692]  __ip_queue_xmit+0x193/0x420
> > > > > > > > > > > > [28848.894179]  __tcp_transmit_skb+0x8cc/0x970
> > > > > > > > > > > >
> > > > > > > > > > > > In this case, the process has hold the qdisc spin lock in __dev_queue_xmit
> > > > > > > > > > > > before the egress packets are mirred, and it will attempt to obtain the
> > > > > > > > > > > > spin lock again after packets are mirred, which cause a deadlock.
> > > > > > > > > > > >
> > > > > > > > > > > > Fix the issue by forbidding assigning mirred action to a filter attached
> > > > > > > > > > > > to the egress.
> > > > > > > > > > > >
> > > > > > > > > > > > Signed-off-by: Mingshuai Ren <renmingshuai@...wei.com>
> > > > > > > > > > > > ---
> > > > > > > > > > > >  net/sched/act_mirred.c                        |  4 +++
> > > > > > > > > > > >  .../tc-testing/tc-tests/actions/mirred.json   | 32 +++++++++++++++++++
> > > > > > > > > > > >  2 files changed, 36 insertions(+)
> > > > > > > > > > > >
> > > > > > > > > > > > diff --git a/net/sched/act_mirred.c b/net/sched/act_mirred.c
> > > > > > > > > > > > index 5b3814365924..fc96705285fb 100644
> > > > > > > > > > > > --- a/net/sched/act_mirred.c
> > > > > > > > > > > > +++ b/net/sched/act_mirred.c
> > > > > > > > > > > > @@ -120,6 +120,10 @@ static int tcf_mirred_init(struct net *net, struct nlattr *nla,
> > > > > > > > > > > >                 NL_SET_ERR_MSG_MOD(extack, "Mirred requires attributes to be passed");
> > > > > > > > > > > >                 return -EINVAL;
> > > > > > > > > > > >         }
> > > > > > > > > > > > +       if (tp->chain->block->q->parent != TC_H_INGRESS) {
> > > > > > > > > > > > +               NL_SET_ERR_MSG_MOD(extack, "Mirred can only be assigned to the filter attached to ingress");
> > > > > > > > > > > > +               return -EINVAL;
> > > > > > > > > > > > +       }
> > > > > > > > > > >
> > > > > > > > > > > Sorry, this is too restrictive as Jiri said. We'll try to reproduce. I
> > > > > > > > > > > am almost certain this used to work in the old days.
> > > > > > > > > >
> > > > > > > > > > Ok, i looked at old notes - it did work at "some point" pre-tdc.
> > > > > > > > > > Conclusion is things broke around this time frame:
> > > > > > > > > > https://lore.kernel.org/netdev/1431679850-31896-1-git-send-email-fw@strlen.de/
> > > > > > > > > > https://lore.kernel.org/netdev/1465095748.2968.45.camel@edumazet-glaptop3.roam.corp.google.com/
> > > > > > > > > >
> > > > > > > > > > Looking further into it.
> > > > > > > > >
> > > > > > > > > This is what we came up with. Eric, please take a look...
> > > > > > > > >
> > > > > > > > > cheers,
> > > > > > > > > jamal
> > > > > > > > >
> > > > > > > > >
> > > > > > > > > --- a/net/core/dev.c
> > > > > > > > > +++ b/net/core/dev.c
> > > > > > > > > @@ -3789,7 +3789,14 @@ static inline int __dev_xmit_skb(struct sk_buff
> > > > > > > > > *skb, struct Qdisc *q,
> > > > > > > > >         if (unlikely(contended))
> > > > > > > > >                 spin_lock(&q->busylock);
> > > > > > > > >
> > > > > > > > > +       if (dev_recursion_level()) {
> > > > > > > >
> > > > > > > > I am not sure what your intent is, but this seems wrong to me.
> > > > > > > >
> > > > > > >
> > > > > > > There is a deadlock if you reenter the same device which has a qdisc
> > > > > > > attached to it more than once.
> > > > > > > Essentially entering __dev_xmit_skb() we grab the root qdisc lock then
> > > > > > > run some action which requires it to grab the root qdisc lock (again).
> > > > > > > This is easy to show with mirred (although i am wondering if syzkaller
> > > > > > > may have produced this at some point)..
> > > > > > > $TC qdisc add dev $DEV root handle 1: htb default 1
> > > > > > > $TC filter add dev $DEV protocol ip u32 match ip protocol 1 0xff
> > > > > > > action mirred egress mirror dev $DEV
> > > > > > >
> > > > > > > Above example is essentially egress $DEV-> egress $DEV in both cases
> > > > > > > "egress $DEV" grabs the root qdisc lock. You could also create another
> > > > > > > example with egress($DEV1->$DEV2->back to $DEV1).
> > > > > > >
> > > > > > > > Some valid setup use :
> > > > > > > >
> > > > > > > > A bonding device, with HTB qdisc (or other qdisc)
> > > > > > > >   (This also could be a tunnel device with a qdisc)
> > > > > > > >
> > > > > > > > -> one or multiple physical NIC, wth FQ or other qdisc.
> > > > > > > >
> > > > > > > > Packets would be dropped here when we try to reach the physical device.
> > > > > > > >
> > > > > > >
> > > > > > > If you have an example handy please send it. I am trying to imagine
> > > > > > > how those would have worked if they have to reenter the root qdisc of
> > > > > > > the same dev multiple times..
> > > > > >
> > > > > > Any virtual device like a GRE/SIT/IPIP/... tunnel, add a qdisc on it ?
> > > > > >
> > > > > > dev_xmit_recursion_inc() is global (per-cpu), it is not per-device.
> > > > > >
> > > > > > A stack of devices A -> B -> C  would elevate the recursion level to
> > > > > > three just fine.
> > > > > >
> > > > > > After your patch, a stack of devices would no longer work.
> > > > > >
> > > > > > It seems mirred correctly injects packets to the top of the stack for
> > > > > > ingress (via netif_rx() / netif_receive_skb()),
> > > > > > but thinks it is okay to call dev_queue_xmit(), regardless of the context ?
> > > > > >
> > > > > > Perhaps safe-guard mirred, instead of adding more code to fast path.
> > > > >
> > > > > I agree not to penalize everybody for a "bad config" like this
> > > > > (surprising syzkaller hasnt caught this). But i dont see how doing the
> > > > > checking within mirred will catch this (we cant detect the A->B->A
> > > > > case).
> > > > > I think you are suggesting a backlog-like queue for mirred? Not far
> > > > > off from that is how it used to work before
> > > > > (https://lore.kernel.org/netdev/1465095748.2968.45.camel@edumazet-glaptop3.roam.corp.google.com/)
> > > >
> > > >
> > > > spin_trylock() had to go. There is no way we could keep this.
> > > >
> > >
> > > Not asking for it to come back... just pointing out why it worked before.
> > >
> > > > > - i.e we had a trylock for the qdisc lock and if it failed we tagged
> > > > > the rx softirq for a reschedule. That in itself is insufficient, we
> > > > > would need a loop check which is per-skb (which we had before
> > > > > https://lore.kernel.org/netdev/1431679850-31896-1-git-send-email-fw@strlen.de/).
> > > > > There are other gotchas there, potentially packet reordering.
> > > >
> > > > If we want to make sure dev_queue_xmit() is called from the top (no
> > > > spinlock held),
> > > > then we need a queue, serviced from another context.
> > > >
> > > > This extra queueing could happen if
> > > > __this_cpu_read(softnet_data.xmit.recursion) > 0
> > > >
> > >
> > > I dont see a way to detect softnet_data.xmit.recursion > 0 at mirred
> > > level. The first time we enter it will be 0.
> >
> > Then it is fine, no qdisc spinlock is held at this point.
> >
> >  The second time we would
> > > deadlock before we hit mirred.
> >
> > This is not how I see the trace.
> >
> > Mirred would detect that and either drop or queue the packet to a work
> > queue or something.
> >
> > diff --git a/net/sched/act_mirred.c b/net/sched/act_mirred.c
> > index 5b38143659249e66718348e0ec4ed3c7bc21c13d..a2c53e200629a17130f38246ab3cdb8c89c6d30e
> > 100644
> > --- a/net/sched/act_mirred.c
> > +++ b/net/sched/act_mirred.c
> > @@ -237,9 +237,15 @@ tcf_mirred_forward(bool at_ingress, bool
> > want_ingress, struct sk_buff *skb)
> >  {
> >         int err;
> >
> > -       if (!want_ingress)
> > +       if (!want_ingress) {
> > +               if (__this_cpu_read(softnet_data.xmit.recursion) > 0) {
> > +                       // TODO increment a drop counter perhaps ?
> > +                       kfree_skb(skb);
> > +                       return -EINVAL;
> > +               }
> > +
> >                 err = tcf_dev_queue_xmit(skb, dev_queue_xmit);
> > -       else if (!at_ingress)
> > +       } else if (!at_ingress)
> >                 err = netif_rx(skb);
> >         else
> >                 err = netif_receive_skb(skb);
>
>
> I doubt this will work - who increments softnet_data.xmit.recursion?
> We enter __dev_xmit_skb (grab lock) --> qdisc_enq-->classify-->mirred
> (recursion is zero) which redirects entering back into  __dev_xmit_skb
> again and deadlocks trying to grab lock.
>  Maybe something is not clear to me, trying your suggestion...
>

jhs@...aone:~$
[   82.890330] __this_cpu_read(softnet_data.xmit.recursion) 0 in
tcf_mirred_forward
[   82.890906]
[   82.890906] ============================================
[   82.890906] WARNING: possible recursive locking detected
[   82.890906] 6.8.0-05205-g77fadd89fe2d-dirty #213 Tainted: G        W
[   82.890906] --------------------------------------------
[   82.890906] ping/418 is trying to acquire lock:
[   82.890906] ffff888006994110 (&sch->q.lock){+.-.}-{3:3}, at:
__dev_queue_xmit+0x1778/0x3550
[   82.890906]
[   82.890906] but task is already holding lock:
[   82.890906] ffff888006994110 (&sch->q.lock){+.-.}-{3:3}, at:
__dev_queue_xmit+0x1778/0x3550
[   82.890906]
[   82.890906] other info that might help us debug this:
[   82.890906]  Possible unsafe locking scenario:
[   82.890906]
[   82.890906]        CPU0
[   82.890906]        ----
[   82.890906]   lock(&sch->q.lock);
[   82.890906]   lock(&sch->q.lock);
[   82.890906]
[   82.890906]  *** DEADLOCK ***
[   82.890906]
[..... other info removed for brevity....]

Needs more thinking...
A fool proof solution is to add the per-recursion counter to be per
netdevice but that maybe considered blasphemy? ;->

cheers,
jamal

> cheers,
> jamal

Powered by blists - more mailing lists