net/ipv4/nexthop.c | 11 +++++++++++ 1 file changed, 11 insertions(+)
From: Jiayuan Chen <jiayuan.chen@shopee.com>
fib_check_nexthop() does not validate that the nexthop family matches
the route family. This allows an IPv4 route to reference an IPv6
nexthop object. When the IPv4 route is looked up, __mkroute_output()
accesses nhc->nhc_pcpu_rth_output which is never allocated for IPv6
nexthops (fib6_nh_init does not call fib_nh_common_init), causing a
NULL pointer dereference.
Note that this is not about IPv4 routes with IPv6 gateways (RFC 5549),
which uses an AF_INET nexthop with nhc_gw_family=AF_INET6 and properly
allocates nhc_pcpu_rth_output via fib_nh_common_init(). The bug here
is an AF_INET6 nexthop object being directly referenced by an IPv4
route, which is an invalid combination.
Add the missing family check in fib_check_nexthop(), mirroring what
fib6_check_nexthop() already does for the reverse direction (rejecting
IPv6 routes that reference IPv4 nexthop objects).
Reproducer:
unshare -rn
ip link set lo up
ip nexthop add id 100 via fe80::1 dev lo
ip route add 172.20.20.0/24 nhid 100
ping -c1 172.20.20.1
After fix:
...
$ ip route add 172.20.20.0/24 nhid 100
Error: IPv4 routes can not use an IPv6 nexthop.
Reported-by: syzbot+334190e097a98a1b81bb@syzkaller.appspotmail.com
Closes: https://lore.kernel.org/all/698f8482.a70a0220.2c38d7.00ca.GAE@google.com/T/
Fixes: 4c7e8084fd46 ("ipv4: Plumb support for nexthop object in a fib_info")
Signed-off-by: Jiayuan Chen <jiayuan.chen@shopee.com>
---
net/ipv4/nexthop.c | 11 +++++++++++
1 file changed, 11 insertions(+)
diff --git a/net/ipv4/nexthop.c b/net/ipv4/nexthop.c
index 7b9d70f9b31c..0f236110cd58 100644
--- a/net/ipv4/nexthop.c
+++ b/net/ipv4/nexthop.c
@@ -1634,6 +1634,12 @@ int fib_check_nexthop(struct nexthop *nh, u8 scope,
goto out;
}
+ if (!nhg->has_v4) {
+ NL_SET_ERR_MSG(extack, "IPv4 routes can not use an IPv6 nexthop");
+ err = -EINVAL;
+ goto out;
+ }
+
if (scope == RT_SCOPE_HOST) {
NL_SET_ERR_MSG(extack, "Route with host scope can not have multiple nexthops");
err = -EINVAL;
@@ -1650,6 +1656,11 @@ int fib_check_nexthop(struct nexthop *nh, u8 scope,
err = -EINVAL;
goto out;
}
+ if (nhi->family != AF_INET) {
+ NL_SET_ERR_MSG(extack, "IPv4 routes can not use an IPv6 nexthop");
+ err = -EINVAL;
+ goto out;
+ }
err = nexthop_check_scope(nhi, scope, extack);
}
--
2.43.0On Sat, 28 Feb 2026 11:13:59 +0800 Jiayuan Chen wrote: > From: Jiayuan Chen <jiayuan.chen@shopee.com> > > fib_check_nexthop() does not validate that the nexthop family matches > the route family. This allows an IPv4 route to reference an IPv6 > nexthop object. When the IPv4 route is looked up, __mkroute_output() > accesses nhc->nhc_pcpu_rth_output which is never allocated for IPv6 > nexthops (fib6_nh_init does not call fib_nh_common_init), causing a > NULL pointer dereference. > > Note that this is not about IPv4 routes with IPv6 gateways (RFC 5549), > which uses an AF_INET nexthop with nhc_gw_family=AF_INET6 and properly > allocates nhc_pcpu_rth_output via fib_nh_common_init(). The bug here > is an AF_INET6 nexthop object being directly referenced by an IPv4 > route, which is an invalid combination. > > Add the missing family check in fib_check_nexthop(), mirroring what > fib6_check_nexthop() already does for the reverse direction (rejecting > IPv6 routes that reference IPv4 nexthop objects). AFAICT this breaks a bunch of tests, quickest to repro with is gre_multipath_nh.sh but you should probably run fib_nexthops.sh on your fix as well. > Reproducer: > > unshare -rn > ip link set lo up > ip nexthop add id 100 via fe80::1 dev lo > ip route add 172.20.20.0/24 nhid 100 > ping -c1 172.20.20.1 -- pw-bot: cr
On 2/28/26 8:39 AM, Jakub Kicinski wrote: > On Sat, 28 Feb 2026 11:13:59 +0800 Jiayuan Chen wrote: >> From: Jiayuan Chen <jiayuan.chen@shopee.com> >> >> fib_check_nexthop() does not validate that the nexthop family matches >> the route family. This allows an IPv4 route to reference an IPv6 >> nexthop object. When the IPv4 route is looked up, __mkroute_output() >> accesses nhc->nhc_pcpu_rth_output which is never allocated for IPv6 >> nexthops (fib6_nh_init does not call fib_nh_common_init), causing a >> NULL pointer dereference. >> >> Note that this is not about IPv4 routes with IPv6 gateways (RFC 5549), >> which uses an AF_INET nexthop with nhc_gw_family=AF_INET6 and properly >> allocates nhc_pcpu_rth_output via fib_nh_common_init(). The bug here >> is an AF_INET6 nexthop object being directly referenced by an IPv4 >> route, which is an invalid combination. >> >> Add the missing family check in fib_check_nexthop(), mirroring what >> fib6_check_nexthop() already does for the reverse direction (rejecting >> IPv6 routes that reference IPv4 nexthop objects). > > AFAICT this breaks a bunch of tests, quickest to repro with is > gre_multipath_nh.sh but you should probably run fib_nexthops.sh > on your fix as well. > nothing to fix. The patch is wrong. IPv4 supports IPv6 gateways; that is a known feature. please post the stack trace for the panic
On Sat, Feb 28, 2026 at 5:33 PM David Ahern <dsahern@kernel.org> wrote: > > On 2/28/26 8:39 AM, Jakub Kicinski wrote: > > On Sat, 28 Feb 2026 11:13:59 +0800 Jiayuan Chen wrote: > >> From: Jiayuan Chen <jiayuan.chen@shopee.com> > >> > >> fib_check_nexthop() does not validate that the nexthop family matches > >> the route family. This allows an IPv4 route to reference an IPv6 > >> nexthop object. When the IPv4 route is looked up, __mkroute_output() > >> accesses nhc->nhc_pcpu_rth_output which is never allocated for IPv6 > >> nexthops (fib6_nh_init does not call fib_nh_common_init), causing a > >> NULL pointer dereference. > >> > >> Note that this is not about IPv4 routes with IPv6 gateways (RFC 5549), > >> which uses an AF_INET nexthop with nhc_gw_family=AF_INET6 and properly > >> allocates nhc_pcpu_rth_output via fib_nh_common_init(). The bug here > >> is an AF_INET6 nexthop object being directly referenced by an IPv4 > >> route, which is an invalid combination. > >> > >> Add the missing family check in fib_check_nexthop(), mirroring what > >> fib6_check_nexthop() already does for the reverse direction (rejecting > >> IPv6 routes that reference IPv4 nexthop objects). > > > > AFAICT this breaks a bunch of tests, quickest to repro with is > > gre_multipath_nh.sh but you should probably run fib_nexthops.sh > > on your fix as well. > > > > nothing to fix. The patch is wrong. IPv4 supports IPv6 gateways; that is > a known feature. > > please post the stack trace for the panic https://lore.kernel.org/all/698f8482.a70a0220.2c38d7.00ca.GAE@google.com/T/
March 1, 2026 at 01:04, "Eric Dumazet" <edumazet@google.com mailto:edumazet@google.com?to=%22Eric%20Dumazet%22%20%3Cedumazet%40google.com%3E > wrote:
>
> On Sat, Feb 28, 2026 at 5:33 PM David Ahern <dsahern@kernel.org> wrote:
>
> >
> > On 2/28/26 8:39 AM, Jakub Kicinski wrote:
> > On Sat, 28 Feb 2026 11:13:59 +0800 Jiayuan Chen wrote:
> > From: Jiayuan Chen <jiayuan.chen@shopee.com>
> >
> > fib_check_nexthop() does not validate that the nexthop family matches
> > the route family. This allows an IPv4 route to reference an IPv6
> > nexthop object. When the IPv4 route is looked up, __mkroute_output()
> > accesses nhc->nhc_pcpu_rth_output which is never allocated for IPv6
> > nexthops (fib6_nh_init does not call fib_nh_common_init), causing a
> > NULL pointer dereference.
> >
> > Note that this is not about IPv4 routes with IPv6 gateways (RFC 5549),
> > which uses an AF_INET nexthop with nhc_gw_family=AF_INET6 and properly
> > allocates nhc_pcpu_rth_output via fib_nh_common_init(). The bug here
> > is an AF_INET6 nexthop object being directly referenced by an IPv4
> > route, which is an invalid combination.
> >
> > Add the missing family check in fib_check_nexthop(), mirroring what
> > fib6_check_nexthop() already does for the reverse direction (rejecting
> > IPv6 routes that reference IPv4 nexthop objects).
> >
> > AFAICT this breaks a bunch of tests, quickest to repro with is
> > gre_multipath_nh.sh but you should probably run fib_nexthops.sh
> > on your fix as well.
> >
> > nothing to fix. The patch is wrong. IPv4 supports IPv6 gateways; that is
> > a known feature.
> >
> > please post the stack trace for the panic
> >
> https://lore.kernel.org/all/698f8482.a70a0220.2c38d7.00ca.GAE@google.com/T/
>
My bad, the previous fix was wrong - IPv4 routes referencing IPv6
nexthop objects is totally via this path.
The crash actually only happens with loopback nexthops, e.g.:
ip nexthop add id 100 via fe80::1 dev lo
In fib6_nh_init(), nexthop objects always have fc_dst=:: (no
destination prefix), so fib6_is_reject() returns true for any
nexthop using loopback device. This causes it to skip
fib_nh_common_init(), leaving nhc_pcpu_rth_output, nhc_exceptions
and nhc_rth_input all NULL. When an IPv4 route later references
this nexthop, __mkroute_output() hits raw_cpu_ptr(NULL) and crashes.
The simplest fix is just allocating nhc_pcpu_rth_output in the
reject path of fib6_nh_init(). The release path already handles
it correctly.
diff --git a/net/ipv6/route.c b/net/ipv6/route.c
index c0350d97307e..4e7c44101709 100644
--- a/net/ipv6/route.c
+++ b/net/ipv6/route.c
@@ -3643,6 +3643,12 @@ int fib6_nh_init(struct net *net, struct fib6_nh *fib6_nh,
goto out;
}
}
+ fib6_nh->nh_common.nhc_pcpu_rth_output =
+ alloc_percpu_gfp(struct rtable __rcu *, gfp_flags);
+ if (!fib6_nh->nh_common.nhc_pcpu_rth_output) {
+ err = -ENOMEM;
+ goto out;
+ }
goto pcpu_alloc;
}
./fib_nexthops.sh
Tests passed: 244
Tests failed: 0
Tests skipped: 2
root@bms-ytl-d1-ap
On 2/28/26 6:57 PM, Jiayuan Chen wrote: > The crash actually only happens with loopback nexthops, e.g.: > > ip nexthop add id 100 via fe80::1 dev lo > > In fib6_nh_init(), nexthop objects always have fc_dst=:: (no > destination prefix), so fib6_is_reject() returns true for any > nexthop using loopback device. This causes it to skip > fib_nh_common_init(), leaving nhc_pcpu_rth_output, nhc_exceptions > and nhc_rth_input all NULL. When an IPv4 route later references > this nexthop, __mkroute_output() hits raw_cpu_ptr(NULL) and crashes. > > The simplest fix is just allocating nhc_pcpu_rth_output in the > reject path of fib6_nh_init(). The release path already handles > it correctly. should have read this before that last response. fib_nh_common_init exists to avoid putting ipv4 logic in ipv6 code, so your proposed patch is wrong. Perhaps the better change is to not use fib6_is_reject in fib6_nh_init or relax the loopback check or move the jump label pcpu_alloc to still call fib_nh_common_init. Also, a test should be added to the nexthops test for this case.
On 2/28/26 6:57 PM, Jiayuan Chen wrote:
> diff --git a/net/ipv6/route.c b/net/ipv6/route.c
> index c0350d97307e..4e7c44101709 100644
> --- a/net/ipv6/route.c
> +++ b/net/ipv6/route.c
> @@ -3643,6 +3643,12 @@ int fib6_nh_init(struct net *net, struct fib6_nh *fib6_nh,
> goto out;
> }
> }
> + fib6_nh->nh_common.nhc_pcpu_rth_output =
> + alloc_percpu_gfp(struct rtable __rcu *, gfp_flags);
> + if (!fib6_nh->nh_common.nhc_pcpu_rth_output) {
> + err = -ENOMEM;
> + goto out;
> + }
> goto pcpu_alloc;
> }
>
>
Already done by fib6_nh_init -> fib_nh_common_init():
nhc->nhc_pcpu_rth_output = alloc_percpu_gfp(struct rtable __rcu *,
gfp_flags);
if (!nhc->nhc_pcpu_rth_output)
return -ENOMEM;
© 2016 - 2026 Red Hat, Inc.