Message-ID: <CALOAHbCKFBVJz1KOyUXrvDu6svTXF3Q-O5boCkMtu-Tsia7aAQ@mail.gmail.com>
Date: Tue, 19 Aug 2025 21:51:12 +0800
From: Yafang Shao <laoar.shao@...il.com>
To: Daniel Borkmann <daniel@...earbox.net>
Cc: davem@...emloft.net, edumazet@...gle.com, kuba@...nel.org, 
	pabeni@...hat.com, horms@...nel.org, netdev@...r.kernel.org, 
	linux-kernel@...r.kernel.org, Thomas Graf <tgraf@...g.ch>, 
	Paul McKenney <paulmck@...nel.org>, Sebastian Andrzej Siewior <bigeasy@...utronix.de>, 
	"open list:BPF [RINGBUF]" <bpf@...r.kernel.org>, Willem de Bruijn <willemb@...gle.com>
Subject: Re: [PATCH] net/cls_cgroup: Fix task_get_classid() during qdisc run

On Tue, Aug 19, 2025 at 8:42 PM Daniel Borkmann <daniel@...earbox.net> wrote:
>
> On 8/19/25 11:37 AM, Yafang Shao wrote:
> > During recent testing with the netem qdisc to inject delays into TCP
> > traffic, we observed that our CLS BPF program failed to function correctly
> > due to incorrect classid retrieval from task_get_classid(). The issue
> > manifests in the following call stack:
> >
> >          bpf_get_cgroup_classid+5
> >          cls_bpf_classify+507
> >          __tcf_classify+90
> >          tcf_classify+217
> >          __dev_queue_xmit+798
> >          bond_dev_queue_xmit+43
> >          __bond_start_xmit+211
> >          bond_start_xmit+70
> >          dev_hard_start_xmit+142
> >          sch_direct_xmit+161
> >          __qdisc_run+102             <<<<< Issue location
> >          __dev_xmit_skb+1015
> >          __dev_queue_xmit+637
> >          neigh_hh_output+159
> >          ip_finish_output2+461
> >          __ip_finish_output+183
> >          ip_finish_output+41
> >          ip_output+120
> >          ip_local_out+94
> >          __ip_queue_xmit+394
> >          ip_queue_xmit+21
> >          __tcp_transmit_skb+2169
> >          tcp_write_xmit+959
> >          __tcp_push_pending_frames+55
> >          tcp_push+264
> >          tcp_sendmsg_locked+661
> >          tcp_sendmsg+45
> >          inet_sendmsg+67
> >          sock_sendmsg+98
> >          sock_write_iter+147
> >          vfs_write+786
> >          ksys_write+181
> >          __x64_sys_write+25
> >          do_syscall_64+56
> >          entry_SYSCALL_64_after_hwframe+100
> >
> > The problem occurs when multiple tasks share a single qdisc. In such cases,
> > __qdisc_run() may transmit skbs created by different tasks. Consequently,
> > task_get_classid() retrieves an incorrect classid since it references the
> > current task's context rather than the skb's originating task.
> >
> > Given that dev_queue_xmit() always executes with bh disabled, we can safely
> > use in_softirq() instead of in_serving_softirq() to properly identify the
> > softirq context and obtain the correct classid.
> >
> > Signed-off-by: Yafang Shao <laoar.shao@...il.com>
> > Cc: Daniel Borkmann <daniel@...earbox.net>
> > Cc: Thomas Graf <tgraf@...g.ch>
> > ---
> >   include/net/cls_cgroup.h | 2 +-
> >   1 file changed, 1 insertion(+), 1 deletion(-)
> >
> > diff --git a/include/net/cls_cgroup.h b/include/net/cls_cgroup.h
> > index 7e78e7d6f015..fc9e0617a73c 100644
> > --- a/include/net/cls_cgroup.h
> > +++ b/include/net/cls_cgroup.h
> > @@ -63,7 +63,7 @@ static inline u32 task_get_classid(const struct sk_buff *skb)
> >        * calls by looking at the number of nested bh disable calls because
> >        * softirqs always disables bh.
> >        */
> > -     if (in_serving_softirq()) {
> > +     if (in_softirq()) {
> >               struct sock *sk = skb_to_full_sk(skb);
> >
> >               /* If there is an sock_cgroup_classid we'll use that. */
>
> Hm, essentially you only want to use the fallback method of retrieving cgroup from
> the socket when the dev_queue_xmit() was triggered via ksoftirqd, rather than from
> a process via syscall all the way into dev_queue_xmit(). It gets more fuzzy presumably
> when skbs are queued somewhere and then some other kthread calls the dev_queue_xmit().

Not only can other kthreads call dev_queue_xmit(), but user tasks can
also invoke it directly.
In the execution path:

  __dev_xmit_skb
  → __qdisc_run
    → qdisc_restart
      → skb = dequeue_skb(q, &validate, packets);

The SKB dequeued at this point might have been originally enqueued by
completely different tasks. This creates a scenario where the current
execution context doesn't match the original context that created and
enqueued the SKB.

>
> Looking at in_softirq(), the comment says "the following macros are deprecated and
> should not be used in new code", see commit 15115830c887 ("preempt: Cleanup the
> macro maze a bit"). Maybe Sebastian or Paul has input on whether in_softirq() is
> still supposed to be used.
>
> Side question, did you investigate where the skb->sk association got orphaned
> somewhere along the way?

The skb->sk is not orphaned in our observed scenario.

The setup to reproduce this issue are as follows,
1. Add network delay to the network interface:
   tc qdisc add dev bond0 root netem delay 1.5ms
2. Create two distinct net_cls cgroups, each running a network-intensive task
3. Initiate parallel TCP streams from both tasks to external servers.
Under this specific condition, the issue reliably occurs.
   The kernel eventually dequeues an SKB that originated from Task-A
while executing in the context of Task-B.

-- 
Regards
Yafang

Powered by blists - more mailing lists