Re: [PATCH net] net: ipv6: fix NOREF dst use in seg6 and rpl lwtunnels

[Andrea Mayer]

On Thu, 23 Apr 2026 10:00:56 +0200
Sebastian Andrzej Siewior <bigeasy@xxxxxxxxxxxxx> wrote:

Hi Sebastian,

thanks for the review, and to Simon and Justin as well.


> On 2026-04-21 11:47:35 [+0200], Andrea Mayer wrote:
> >
> > [snip]
> 
> So the dst passed to skb_dst_set_noref() has no reference count. The fix
> is to use skb_dst_force() to increment the refcount on it. But this
> requires that we are in the same RCU section. And I guess we are since
> none of the warnings are visible.
 
Yes. lwtunnel_input() holds rcu_read_lock() around ops->input(), which is
where seg6_input_core()/rpl_input() execute. The skb_dst_force() is called
within that RCU section.
 

> Doesn't this make ip6_route_input() on RT fragile in general due to the
> RT6_LOOKUP_F_DST_NOREF usage or here something special about the two
> files that are patched?
> Based on your explanation it all makes sense, I am just not sure if this
> race is limited to those two are if there is more to it.

seg6_input_core() and rpl_input() cache the dst via dst_cache_set_ip6(), which
invokes dst_hold(). The dst_hold() calls rcuref_get(), failing on a zero
refcount and triggering a WARN, but the pointer is still stored in the cache.
After the RCU grace period completes the dst is freed, and a subsequent
dst_cache_get() returns a dangling pointer.
 
The other callers of ip6_route_input() (e.g., ipv6_srh_rcv, ipv6_rpl_srh_rcv,
ip6_rcv_finish_core) consume the NOREF dst without caching it. Even if the
pcpu_rt's refcount is concurrently dropped to zero, the dst memory remains
valid because dst_release() defers the actual free via call_rcu_hurry() and the
caller is still inside the RCU read-side critical section.


> > [snip]
> >
> > Fixes: af4a2209b134 ("ipv6: sr: use dst_cache in seg6_input")
> > Fixes: a7a29f9c361f ("net: ipv6: add rpl sr tunnel")
> 
> If having PREEMPT_RT_NEEDS_BH_LOCK unset is the requirement then the
> right fixes: would be
> Fixes: 3253cb49cbad4 ("softirq: Allow to drop the softirq-BKL lock on PREEMPT_RT")
> 
> as prior this commit the race is not possible, right?

I built and tested kernels at 3253cb49cbad and its parent fd4e876f59b7 (both
CONFIG_PREEMPT_RT=y, without the fix): no issues at fd4e876f59b7.
At 3253cb49cbad, a pcpu_rt cmpxchg contention in rt6_make_pcpu_route() shows
up, which was addressed in 1adaea51c61b. I also tested at 1adaea51c61b, and at
that point the dst_hold() race described in this patch appears.
 
The seg6/rpl code obtains a NOREF dst from ip6_route_input(), does not promote
it via skb_dst_force(), and passes it to dst_cache_set_ip6() which calls
dst_hold(). This pattern has been present since af4a2209b134 and a7a29f9c361f,
and the current Fixes: tags point to the commits where it was introduced.
Does that seem reasonable?


> Does this mean that rpl_input() does a local_bh_disable() while
> obtaining the dst but it never runs outside of bh-disabled section?
> Because if it can run in preemptible context then it would not be to
> PREEMPT_RT at which point the Fixes: tags from above would make sense
> again.
> 

rpl_input() and seg6_input_core() run in softirq context via lwtunnel_input().
They do local_bh_disable() around dst_cache_get() and dst_cache_set_ip6(), but
not around ip6_route_input(). The race window is between ip6_route_input()
returning and dst_cache_set_ip6().
 
> Sebastian

Ciao,
Andrea