Message-ID: <dEmtK-Tj-bnNJVo0mNwP1vJ1cj9g0hqnoi-0HJdZeTittbRmmzE4wBRIjapBAFQNZDWgE4hcR27UrTSuiGj_-yRFntfX4Tuv4QP6asVecZQ=@willsroot.io>
Date: Wed, 26 Nov 2025 20:30:21 +0000
From: William Liu <will@...lsroot.io>
To: Cong Wang <xiyou.wangcong@...il.com>
Cc: netdev@...r.kernel.org, stephen@...workplumber.org, kuba@...nel.org, Savino Dicanosa <savy@...t3mfailure.io>, Jamal Hadi Salim <jhs@...atatu.com>
Subject: Re: [Patch net v5 3/9] net_sched: Implement the right netem duplication behavior

Jamal should be added to this - he was the one brainstorming fixes for this bug from the very beginning for a whole month [1].

On Wednesday, November 26th, 2025 at 7:53 PM, Cong Wang <xiyou.wangcong@...il.com> wrote:

> 
> 
> In the old behavior, duplicated packets were sent back to the root qdisc,
> which could create dangerous infinite loops in hierarchical setups -
> imagine a scenario where each level of a multi-stage netem hierarchy kept
> feeding duplicates back to the top, potentially causing system instability
> or resource exhaustion.
> 
> The new behavior elegantly solves this by enqueueing duplicates to the same
> qdisc that created them, ensuring that packet duplication occurs exactly
> once per netem stage in a controlled, predictable manner. This change
> enables users to safely construct complex network emulation scenarios using
> netem hierarchies (like the 4x multiplication demonstrated in testing)
> without worrying about runaway packet generation, while still preserving
> the intended duplication effects.
> 
> Another advantage of this approach is that it eliminates the enqueue reentrant
> behaviour which triggered many vulnerabilities. See the last patch in this
> patchset which updates the test cases for such vulnerabilities.
> 
> Now users can confidently chain multiple netem qdiscs together to achieve
> sophisticated network impairment combinations, knowing that each stage will
> apply its effects exactly once to the packet flow, making network testing
> scenarios more reliable and results more deterministic.
> 

Cong, this approach has an issue we previously raised - please refer to [2]. I re-posted the summary of the issues with the various other approaches in [3] just 2 days ago in a thread with you on it. As both Jamal and Stephen have pointed out, this breaks expected user behavior as well, and the enqueuing at root was done for the sake of proper accounting and rate limit semantics. You pointed out that this doesn't violate manpage semantics, but this is still changing long-term user behavior. It doesn't make sense imo to change one longtime user behavior for another.

Jamal suggested a really reasonable fix with tc_skb_ext - can we please take a look at its soundness and attempt that approach? No user behavior would be affected in that case.

> I tested netem packet duplication in two configurations:
> 1. Nest netem-to-netem hierarchy using parent/child attachment
> 2. Single netem using prio qdisc with netem leaf
> 
> Setup commands and results:
> 
> Single netem hierarchy (prio + netem):
> tc qdisc add dev lo root handle 1: prio bands 3 priomap 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
> tc filter add dev lo parent 1:0 protocol ip matchall classid 1:1
> tc qdisc add dev lo parent 1:1 handle 10: netem limit 4 duplicate 100%
> 
> Result: 2x packet multiplication (1→2 packets)
> 2 echo requests + 4 echo replies = 6 total packets
> 
> Expected behavior: Only one netem stage exists in this hierarchy, so
> 1 ping becomes 2 packets (100% duplication). The 2 echo requests generate
> 2 echo replies, which also get duplicated to 4 replies, yielding the
> predictable total of 6 packets (2 requests + 4 replies).
> 
> Nest netem hierarchy (netem + netem):
> tc qdisc add dev lo root handle 1: netem limit 1000 duplicate 100%
> tc qdisc add dev lo parent 1: handle 2: netem limit 1000 duplicate 100%
> 
> Result: 4x packet multiplication (1→2→4 packets)
> 4 echo requests + 16 echo replies = 20 total packets
> 
> Expected behavior: Root netem duplicates 1 ping to 2 packets, child netem
> receives 2 packets and duplicates each to create 4 total packets. Since
> ping operates bidirectionally, 4 echo requests generate 4 echo replies,
> which also get duplicated through the same hierarchy (4→8→16), resulting
> in the predictable total of 20 packets (4 requests + 16 replies).
> 
> The new netem duplication behavior does not break the documented
> semantics of "creates a copy of the packet before queuing." The man page
> description remains true since duplication occurs before the queuing
> process, creating both original and duplicate packets that are then
> enqueued. The documentation does not specify which qdisc should receive
> the duplicates, only that copying happens before queuing. The implementation
> choice to enqueue duplicates to the same qdisc (rather than root) is an
> internal detail that maintains the documented behavior while preventing
> infinite loops in hierarchical configurations.
> 
> Fixes: 0afb51e72855 ("[PKT_SCHED]: netem: reinsert for duplication")
> Reported-by: William Liu will@...lsroot.io
> 
> Reported-by: Savino Dicanosa savy@...t3mfailure.io
> 
> Signed-off-by: Cong Wang xiyou.wangcong@...il.com
> 
> ---
> net/sched/sch_netem.c | 26 +++++++++++++++-----------
> 1 file changed, 15 insertions(+), 11 deletions(-)
> 
> diff --git a/net/sched/sch_netem.c b/net/sched/sch_netem.c
> index fdd79d3ccd8c..191f64bd68ff 100644
> --- a/net/sched/sch_netem.c
> +++ b/net/sched/sch_netem.c
> @@ -165,6 +165,7 @@ struct netem_sched_data {
> */
> struct netem_skb_cb {
> u64 time_to_send;
> + u8 duplicate : 1;
> };
> 
> static inline struct netem_skb_cb *netem_skb_cb(struct sk_buff *skb)
> @@ -460,8 +461,16 @@ static int netem_enqueue(struct sk_buff *skb, struct Qdisc *sch,
> skb->prev = NULL;
> 
> 
> /* Random duplication */
> - if (q->duplicate && q->duplicate >= get_crandom(&q->dup_cor, &q->prng))
> 
> - ++count;
> + if (q->duplicate) {
> 
> + bool dup = true;
> +
> + if (netem_skb_cb(skb)->duplicate) {
> 
> + netem_skb_cb(skb)->duplicate = 0;
> 
> + dup = false;
> + }
> + if (dup && q->duplicate >= get_crandom(&q->dup_cor, &q->prng))
> 
> + ++count;
> + }
> 
> /* Drop packet? */
> if (loss_event(q)) {
> @@ -532,17 +541,12 @@ static int netem_enqueue(struct sk_buff *skb, struct Qdisc sch,
> }
> 
> /
> - * If doing duplication then re-insert at top of the
> - * qdisc tree, since parent queuer expects that only one
> - * skb will be queued.
> + * If doing duplication then re-insert at the same qdisc,
> + * as going back to the root would induce loops.
> */
> if (skb2) {
> - struct Qdisc rootq = qdisc_root_bh(sch);
> - u32 dupsave = q->duplicate; / prevent duplicating a dup... */
> 
> -
> - q->duplicate = 0;
> 
> - rootq->enqueue(skb2, rootq, to_free);
> 
> - q->duplicate = dupsave;
> 
> + netem_skb_cb(skb2)->duplicate = 1;
> 
> + qdisc_enqueue(skb2, sch, to_free);
> skb2 = NULL;
> }
> 
> --
> 2.34.1

[1] https://lore.kernel.org/netdev/8DuRWwfqjoRDLDmBMlIfbrsZg9Gx50DHJc1ilxsEBNe2D6NMoigR_eIRIG0LOjMc3r10nUUZtArXx4oZBIdUfZQrwjcQhdinnMis_0G7VEk=@willsroot.io/T/#u
[2] https://lore.kernel.org/netdev/CAM0EoMmTZon=nFmLsDPKhDEzHruw701iV9=mq92At9oKo0LGpA@mail.gmail.com/T/#u
[3] https://lore.kernel.org/netdev/PKMd5btHYmJcKSiIJdtxQvZBEfuS4RQkBnE4M-TZkjUq_Rdj6Wgm8wDmX-p6rIkSRGDJN8ufn0HcDI6-r2lgibdSk7cn1mHIdbZEohJFKMg=@willsroot.io/

Powered by blists - more mailing lists