* Re: [PATCH net 1/2] ipv4: accept nexthop dump attributes in route deletes
2026-11-18 8:45 ` [PATCH net 1/2] ipv4: accept nexthop dump attributes in route deletes Yuen Su
@ 2026-11-18 12:21 ` Ido Schimmel
2026-11-19 15:47 ` David Ahern
2026-11-22 9:50 ` netdev-bot+sashiko
2 siblings, 0 replies; 6+ messages in thread
From: Ido Schimmel @ 2026-11-18 12:21 UTC (permalink / raw)
To: Yuen Su
Cc: netdev, David Ahern, Stephen Hemminger, David S . Miller,
Eric Dumazet, Jakub Kicinski, Paolo Abeni, Simon Horman,
Shuah Khan, linux-kselftest, linux-kernel, linux-api
On Wed, Nov 18, 2026 at 04:45:09PM +0800, Yuen Su wrote:
> ip route flush reuses route dump messages as RTM_DELROUTE requests.
That's an iproute2 design choice. Where does it say that the kernel has
to accept netlink messages that it dumped when they are delivered as
requests (potentially with a different type, as in this case:
RTM_NEWROUTE -> RTM_DELROUTE)?
I assume it was implemented this way because it was easy / convenient,
but you can tell that it doesn't make sense by how fragile it is.
Another example (that wasn't accepted):
https://lore.kernel.org/netdev/20260923-lwt-encap-noflag-v1-0-8de7ab6c86e9@gmail.com/
> With nexthop compatibility mode enabled, the dump includes RTA_NH_ID
> alongside expanded nexthop attributes. rtm_to_fib_config() rejects the
> combination with -EINVAL, leaving the routes installed.
OK, but why do you need this compatibility mode in the first place and
you're not simply disabling it like most production systems that are
using nexthop objects?
Note that iproute2 can include the necessary information in route dumps
even when the compatibility mode is disabled (with -d):
# sysctl -wq net.ipv4.nexthop_compat_mode=0
# ip route show
192.0.2.0/24 nhid 1
# ip -d route show
unicast 192.0.2.0/24 nhid 1 proto boot scope global
nh_info id 1 dev dummy1 scope host proto unspec
# ip nexthop
id 1 dev dummy1 scope host
Is this patch motivated by an actual use case or an LLM report?
>
> Skip the mutual exclusion check for RTM_DELROUTE. fib_nh_match()
> already uses the ID to match the nexthop when an ID is supplied, so
> the expanded fields do not constrain that match. Keep the validation
> for RTM_NEWROUTE and leave the other delete selectors intact.
This leaves save / restore broken.
>
> The failure was reproduced while investigating the iproute2 flush
> report below. An unmodified kernel and iproute2 reproduced it without
> fault injection in a fresh network namespace with compatibility mode
> enabled.
>
> ip link add d0 type dummy
> ip link set d0 up
> ip -4 nexthop add id 101 via 192.0.2.1 dev d0 onlink
> ip -4 route add 198.51.100.0/24 table 1111 nhid 101 metric 10
> ip -4 route flush table 1111 metric 10
>
> Verified the flush and the existing new-route validation in QEMU/KVM.
> The IPv4 compatibility selftest reproduced the failure on the base
> kernel and passed with the fix. The IPv4 group and IPv6 compatibility
> selftests also passed with the fix.
>
> Codex assisted with isolating the IPv4 delete change, writing the
> regression test and running the QEMU comparisons.
>
> Fixes: 493ced1ac47c ("ipv4: Allow routes to use nexthop objects")
Not a regression, never worked. If anything it should be targeted at
net-next, but I don't think that such patches should be accepted. The
fix belongs in iproute2.
> Link: https://lists.openwall.net/netdev/2026/10/26/186
> Signed-off-by: Yuen Su <zypulsecore@gmail.com>
^ permalink raw reply [flat|nested] 6+ messages in thread* Re: [PATCH net 1/2] ipv4: accept nexthop dump attributes in route deletes
2026-11-18 8:45 ` [PATCH net 1/2] ipv4: accept nexthop dump attributes in route deletes Yuen Su
2026-11-18 12:21 ` Ido Schimmel
@ 2026-11-19 15:47 ` David Ahern
2026-11-22 9:50 ` netdev-bot+sashiko
2 siblings, 0 replies; 6+ messages in thread
From: David Ahern @ 2026-11-19 15:47 UTC (permalink / raw)
To: Yuen Su, netdev
Cc: Ido Schimmel, Stephen Hemminger, David S . Miller, Eric Dumazet,
Jakub Kicinski, Paolo Abeni, Simon Horman, Shuah Khan,
linux-kselftest, linux-kernel, linux-api
On 11/18/26 2:45 AM, Yuen Su wrote:
> diff --git a/net/ipv4/fib_frontend.c b/net/ipv4/fib_frontend.c
> index 8c123b1..6b012c8 100644
> --- a/net/ipv4/fib_frontend.c
> +++ b/net/ipv4/fib_frontend.c
> @@ -868,7 +868,10 @@ static int rtm_to_fib_config(struct net *net, struct sk_buff *skb,
> goto errout;
> }
>
> - if (cfg->fc_nh_id) {
> + /* Dumps can include both a nexthop ID and compatibility attributes.
> + * On deletion, fib_nh_match() uses the ID to match the nexthop.
> + */
> + if (cfg->fc_nh_id && nlh->nlmsg_type != RTM_DELROUTE) {
> if (cfg->fc_oif || cfg->fc_gw_family ||
> cfg->fc_encap || cfg->fc_mp) {
> NL_SET_ERR_MSG(extack,
as the error message at this point shows, the 2 specifications must be
mutually exclusive. iproute2 will need to strip the nh_id or the nexthop
spec.
^ permalink raw reply [flat|nested] 6+ messages in thread* Re: [PATCH net 1/2] ipv4: accept nexthop dump attributes in route deletes
2026-11-18 8:45 ` [PATCH net 1/2] ipv4: accept nexthop dump attributes in route deletes Yuen Su
2026-11-18 12:21 ` Ido Schimmel
2026-11-19 15:47 ` David Ahern
@ 2026-11-22 9:50 ` netdev-bot+sashiko
2 siblings, 0 replies; 6+ messages in thread
From: netdev-bot+sashiko @ 2026-11-22 9:50 UTC (permalink / raw)
To: zypulsecore
Cc: netdev, dsahern, idosch, stephen, davem, edumazet, kuba, pabeni,
horms, shuah, linux-kselftest, linux-kernel, linux-api
Thank you for your contribution! Sashiko AI review found 2 potential
issue(s) to consider.
Critical: 0 · High: 1 · Medium: 0 · 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).
- [High] The patch fixes `ip route flush` of nexthop-object routes for
IPv4 only.
- [Low] The commit message says the failure occurs "with nexthop
compatibility mode enabled" and gives a reproducer "in a fresh network…
--
ipv4: accept nexthop dump attributes in route deletes
This commit skips the check in rtm_to_fib_config() that makes RTA_NH_ID
and the expanded nexthop attributes mutually exclusive, but only for
RTM_DELROUTE requests. This lets ip route flush delete routes that use
nexthop objects when the dump carries both. The check stays in place for
RTM_NEWROUTE.
Link: https://lists.openwall.net/netdev/2026/10/26/186
> ip route flush reuses route dump messages as RTM_DELROUTE requests.
> With nexthop compatibility mode enabled, the dump includes RTA_NH_ID
> alongside expanded nexthop attributes. rtm_to_fib_config() rejects the
> combination with -EINVAL, leaving the routes installed.
[ ... ]
> report below. An unmodified kernel and iproute2 reproduced it without
> fault injection in a fresh network namespace with compatibility mode
> enabled.
[Severity: Low]
This isn't a bug, but "with nexthop compatibility mode enabled" reads like
an opt-in setting. Isn't compat mode already on by default in every network
namespace?
net/ipv4/af_inet.c:inet_init_net() {
...
net->ipv4.sysctl_nexthop_compat_mode = 1;
...
}
fib_dump_info() leaves out the expanded attributes only when compat mode
has been set to 0, so default configurations hit this bug.
Could the commit message say that default configurations are affected? That
would help with stable triage.
> diff --git a/net/ipv4/fib_frontend.c b/net/ipv4/fib_frontend.c
> index 8c123b111b3d8..6b012c8c6ea30 100644
> --- a/net/ipv4/fib_frontend.c
> +++ b/net/ipv4/fib_frontend.c
> @@ -868,7 +868,10 @@ static int rtm_to_fib_config(struct net *net, struct sk_buff *skb,
> goto errout;
> }
>
> - if (cfg->fc_nh_id) {
> + /* Dumps can include both a nexthop ID and compatibility attributes.
> + * On deletion, fib_nh_match() uses the ID to match the nexthop.
> + */
> + if (cfg->fc_nh_id && nlh->nlmsg_type != RTM_DELROUTE) {
> if (cfg->fc_oif || cfg->fc_gw_family ||
> cfg->fc_encap || cfg->fc_mp) {
> NL_SET_ERR_MSG(extack,
[Severity: High]
Does IPv6 need the same change? rtm_to_fib6_config() in net/ipv6/route.c
still runs this check for every message type, including RTM_DELROUTE:
net/ipv6/route.c:rtm_to_fib6_config() {
...
if (tb[RTA_NH_ID]) {
if (tb[RTA_GATEWAY] || tb[RTA_OIF] ||
tb[RTA_MULTIPATH] || tb[RTA_ENCAP]) {
NL_SET_ERR_MSG(extack,
"Nexthop specification and nexthop id are mutually exclusive");
goto errout;
}
cfg->fc_nh_id = nla_get_u32(tb[RTA_NH_ID]);
}
...
}
err is already -EINVAL at this point.
In compat mode, rt6_fill_node() adds RTA_NH_ID to the dump and then calls
rt6_fill_node_nexthop(). For a group, that adds RTA_MULTIPATH. For a single
nexthop, fib_nexthop_info() adds RTA_OIF, plus RTA_GATEWAY when there is a
gateway. IPv6 nexthops always have a device, so RTA_OIF is always there.
When iproute2 sends that message back as RTM_DELROUTE, won't ip -6 route
flush on nhid routes fail with -EINVAL the same way, and leave the routes
installed?
ip6_route_del() already matches on the nexthop ID:
if (rt->nh && cfg->fc_nh_id &&
rt->nh->id != cfg->fc_nh_id)
continue;
It also skips routes without a nexthop object when fc_nh_id is set. That
suggests the same RTM_DELROUTE exemption would be safe in
rtm_to_fib6_config().
The commit message also says:
The IPv4 group and IPv6 compatibility selftests also passed with the fix.
This could be read as saying IPv6 is not affected. However, those selftests
never run ip -6 route flush, and this patch leaves the IPv6 path unchanged.
The later patch in the series ("selftests: net: test flushing IPv4 nexthop
routes in compat mode") also covers only IPv4 and does not touch
net/ipv6/route.c.
Could the series be extended to cover the IPv6 parser too?
--
Sashiko AI review · https://netdev-ai.bots.linux.dev/sashiko/#/patchset/20261118084510.332821-1-zypulsecore%40gmail.com
^ permalink raw reply [flat|nested] 6+ messages in thread