# Re: \[PATCH net v2 2/2\] net/sched: sch\_teql: keep skb-\>dev consistent on the arp-queue path

- From: Jamal Hadi Salim \<jhs@mojatatu.com\>
- Date: 2026-11-03 11:19:19 
- List: netdev
- Page: https://ratatoskr.run/netdev/2026/11/17851908
- Thread: https://ratatoskr.run/netdev/2026/11/17839849/t.md

```text
On Mon, Nov 2, 2026 at 11:30 AM [ref] wrote:
>
> Thank you for your contribution! Sashiko AI review found 3 potential
> issue(s) to consider.
>
> Critical: 0 · High: 0 (1) · Medium: 2 · Low: 1
>
> If you believe none of the issues are legitimate, please reply to each
> issue of Medium severity or higher and explain why the AI is wrong.
> If any of them are legitimate and you plan to respin, please let patchwork
> know by including "pw-bot: cr" as a separate line at the end of your reply
> (one such reply per series is enough).
>
> - [Medium] The new gate in neigh_update_process_arp_queue()
>   (net/core/neighbour.c:1343-1344), `if (!n1->dev->header_ops) skb->dev…

This is by design. Covers every headerless neigh_direct_output user.
Nothing to fix.

> - [Medium] teql_master_xmit() now sets skb->dev = slave before
>   teql_resolve() (sch_teql.c:371).

No (that renames the master again. Nothing to fix.

> - [Low] The commit message's reproducer cannot show the use-after-free
>   the patch says it fixes.

Sigh. I will send a v3 to fix the messaging. Dammit, dont whine about
something else after you rat bastard.
pw-bot:cr

> Pre-existing issues:
> - [High] This problem wasn't introduced by this patch, but
>   teql_qdisc_init() (net/sched/sch_teql.c:194-253) never checks that the…

Separate thread. see:
 https://lore.kernel.org/netdev/SER-NSBD.v2.r0.20261101171325-3@mojatatu.com/

cheers,
jamal
> --
>
> net/sched: sch_teql: keep skb->dev consistent on the arp-queue path
>
> This patch moves skb->dev = slave ahead of teql_resolve(). An skb parked
> on the slave's arp_queue then names the slave instead of the teql master.
> It also restores skb->dev = n1->dev in neigh_update_process_arp_queue()
> when the top-level neighbour's device has no header_ops, so the skb goes
> back through teql_master_xmit().
>
> Link: https://lore.kernel.org/netdev/20260824115928.4099988-1-victor@mojatatu.com/
> Link: https://lore.kernel.org/netdev/20261026160747.283702-1-horms@kernel.org/
> Link: https://lore.kernel.org/netdev/20261031165925.780930-1-horms@kernel.org/
>
> > Conditions to recreate the bug:
> > CONFIG_NET_SCH_TEQL=y, CONFIG_VETH=y, run as root. Create a veth pair,
> > make one end a teql slave, give the master an address and a route, flush
> > the slave's neighbour entry and send one packet:
>
> [ ... ]
>
> > On the unfixed tree the skb is parked on v0's arp_queue while skb->dev
> > still names teql0, so the reinjection can reach a master that teql_exit()
> > has freed.
>
> [Severity: Low]
> Can this recipe reach the use-after-free it describes?
>
> With CONFIG_NET_SCH_TEQL=y, teql_exit() is __exit and never runs, so no
> master is ever freed:
>
> net/sched/sch_teql.c:teql_exit() {
>     ...
>                 unregister_qdisc(&master->qops);
>                 unregister_netdev(master->dev);
>                 free_netdev(master->dev);
>     ...
> }
>
> teql_master_setup() installs no rtnl_link_ops, so "ip link del teql0"
> can't free the master either.
>
> Building sch_teql as a module doesn't change this. teql_master_setup()
> sets ops->owner = THIS_MODULE, and qdisc_lookup_ops() takes a module
> reference through bpf_try_module_get() for each slave qdisc. rmmod
> therefore fails while the slave qdisc on v0 exists.
>
> To reach a freed master you seem to need sch_teql as a module, then
> "tc qdisc del dev v0 root", then "rmmod sch_teql". All of this has to
> happen while the skb is still parked on v0's arp_queue. None of these
> steps are in the recipe. Could the recipe be updated to match?
>
> Also, the message only names the reinjection path. Commit dc4b95b8fee9
> showed this KASAN-confirmed path for a stale skb->dev:
>
> neigh_invalidate()->arp_error_report()->ipv4_link_failure()->__icmp_send()
>
> Here __icmp_send() reads skb_in->dev. This patch covers that path too,
> because the parked skb now names the slave. Would it be worth saying so
> in the message?
>
> > diff --git a/net/core/neighbour.c b/net/core/neighbour.c
> > index d6e9c7c59a892..541cf699abae4 100644
> > --- a/net/core/neighbour.c
> > +++ b/net/core/neighbour.c
> > @@ -1339,6 +1339,9 @@ static void neigh_update_process_arp_queue(struct neighbour *neigh)
> >                       if (n2)
> >                               n1 = n2;
> >               }
> > +             /* headerless outputs (teql, shaper) transmit on skb->dev */
> > +             if (!n1->dev->header_ops)
> > +                     skb->dev = n1->dev;
> >               READ_ONCE(n1->output)(n1, skb);
>
> [Severity: Medium]
> Is this check narrow enough to match only teql?
>
> n1 comes from dst_neigh_lookup_skb() on the skb's retained dst. It does
> not come from neigh, the neighbour the skb was parked on. Nothing checks
> that neigh->dev is a teql slave of n1->dev, so any headerless route
> device matches.
>
> Two producers park the skb on another device Y's neighbour while keeping
> the original route's dst:
>
> - nft "fwd ... neigh", via nft_fwd_neigh_eval(). nft_fwd_validate()
>   allows it at NF_NETDEV_EGRESS.
> - bpf_redirect_neigh() with an explicit nexthop.
>   __bpf_redirect_neigh_v4/v6 only replace the dst when nh == NULL.
>
> Suppose such a forward runs at the egress hook of a headerless device X
> and the neighbour on Y is unresolved:
>
> __neigh_event_send()
>   skb_dst_force(skb)                    /* dst->dev is still X */
>   queue on Y's arp_queue
>
> neigh_update_process_arp_queue()
>   n1 = dst_neigh_lookup_skb(dst, skb)   /* X's neighbour */
>   skb->dev = n1->dev                    /* X */
>   neigh_direct_output()->dev_queue_xmit(X)
>
> nf_skip_egress was reset after sch_handle_egress(), so nf_hook_egress()
> runs X's egress chain a second time.
>
> The forward then runs twice. nft_fwd_neigh_eval() decrements TTL again,
> so a TTL=2 packet that should leave Y with TTL 1 is dropped here on the
> second pass:
>
>                 if (iph->ttl <= 1) {
>                         verdict = NF_DROP;
>                         goto out;
>                 }
>
> Quota, limit and counter state are charged twice. The second pass may
> also decide differently, for example with numgen, an expired set element
> or a ruleset change. The packet then leaves through X, the route the
> forward was meant to bypass.
>
> Devices that match as X include:
>
> - point-to-point ipgre with a unicast remote, and collect_md gre
>   (ipgre_tunnel_init() leaves header_ops NULL)
> - ppp, eql and teql
>
> Tunnels using ip_tunnel_header_ops (wireguard, tun, ipip, sit, xfrmi and
> so on) don't match. Before this patch these skbs left Y without a link
> header. That was broken, but they never re-entered X.
>
> The commit message acknowledges this and leaves the producer fix to a
> separate change. This net/core change is going to stable under a
> sch_teql subject. Could the producer fix be included, or the gate
> narrowed, for example by requiring neigh->dev to be a teql slave of
> n1->dev?
>
> > diff --git a/net/sched/sch_teql.c b/net/sched/sch_teql.c
> > index ed8faa848f184..4eac1f9c8ddf6 100644
> > --- a/net/sched/sch_teql.c
> > +++ b/net/sched/sch_teql.c
> > @@ -368,13 +368,13 @@ static netdev_tx_t teql_master_xmit(struct sk_buff *skb, struct net_device *dev)
> >                       continue;
> >               }
> >
> > +             skb->dev = slave;
> >               switch (teql_resolve(skb, skb_res, slave, slave_txq, &hlen)) {
>
> [Severity: Medium]
> What happens to an skb parked here if the neighbour then fails?
>
> In case 1 of teql_resolve(), teql_master_xmit() returns before the
> skb->dev = dev restore at the bottom of the loop. The skb therefore sits
> on the slave's arp_queue with skb->dev = slave.
>
> Only neigh_update_process_arp_queue() was changed to put the master back.
> The other consumer of the queue passes the skb on unchanged:
>
> net/core/neighbour.c:neigh_invalidate() {
>     ...
>         while (neigh->nud_state == NUD_FAILED &&
>                (skb = __skb_dequeue(&neigh->arp_queue)) != NULL) {
>                 write_unlock(&neigh->lock);
>                 neigh->ops->error_report(neigh, skb);
>     ...
> }
>
> For IPv6 the call chain is:
>
> ndisc_error_report()->dst_link_failure()->ip6_link_failure()->
> icmpv6_send()->icmp6_send()
>
> icmp6_send() takes its context from skb->dev:
>
>         net = dev_net_rcu(skb->dev);
>         ...
>                 iif = l3mdev_master_ifindex(skb->dev);
>
> It also uses ipv6_chk_addr(net, &hdr->daddr, skb->dev, 0), and
> icmp6_iif(skb) for link-local sources. That iif becomes flowi6_oif for
> the route lookup of the error.
>
> These values used to come from the teql master. Now they come from the
> slave, which changes the result in these setups:
>
> - teql0 in a VRF with the slave outside it: the ICMPv6 Destination
>   Unreachable is routed in the main table instead of the VRF table.
> - A link-local source on teql0: the oif becomes the slave.
> - A slave in another netns: the error is built in that netns.
>
> Can the local sender's error be dropped or misrouted in these setups?
>
> Would restoring skb->dev on this path as well avoid this? One option is
> skb_dst(skb)->dev, which dst_dev_put() keeps valid.
>
> [Severity: High]
> This isn't a bug introduced by this patch, but should teql_qdisc_init()
> check that the slave and the master are in the same network namespace?
>
> The teql Qdisc_ops are on the global qdisc_base list, so
> qdisc_lookup_ops() finds "teql0" from any netns. rtnetlink_rcv_msg()
> only checks netlink_net_capable(skb, CAP_NET_ADMIN) against the caller's
> own netns. teql_qdisc_init() then checks only hard_header_len,
> m->dev == dev and the root parent:
>
> net/sched/sch_teql.c:teql_qdisc_init() {
>     ...
>         if (dev->hard_header_len > m->dev->hard_header_len)
>                 return -EINVAL;
>
>         if (m->dev == dev)
>                 return -ELOOP;
>     ...
>                 m->dev->mtu = dev->mtu;
>                 m->dev->flags = (m->dev->flags&~FMASK)|(dev->flags&FMASK);
>     ...
> }
>
> As a result, root inside an unprivileged user+net namespace can run
> "tc qdisc add dev <own veth> root teql0". That enslaves its own device
> to the init_net master. If it is the first slave, it also overwrites the
> master's mtu and flags.
>
> teql_master_xmit() then round-robins init_net traffic routed through
> teql0 onto that device. __teql_resolve() creates the neighbour there, so
> the namespace can read or drop that traffic and controls neighbour
> resolution for parked skbs. This needs sch_teql loaded or built in by
> the host. There is no MODULE_ALIAS for autoloading.
>
> With this patch, parked skbs also carry skb->dev = slave. ICMPv6 errors
> from neighbour failure that quote the init_net packet are therefore
> built in that namespace, through dev_net_rcu(skb->dev) in icmp6_send().
>
> [ ... ]
>
> --
> Sashiko AI review · https://netdev-ai.bots.linux.dev/sashiko/#/patchset/QDISC-6HS9.v2.r0.20261101105945%40mojatatu.com
```
