Re: [PATCH net-next] ipv6: check mc_forwarding before calling ip6mr_get_route()
From: Wenbo Sun
Date: Sun Nov 15 2026 - 20:46:03 EST
Thank you for the detailed review comments. Please note that a 2-patch v2
series was already posted prior to this message [1]; below is a detailed
root-cause analysis, our modification plan, and the UML reproduction and
verification results for the upcoming v3 series:
1. Root-Cause Analysis:
(a) [Medium 1] Returning -EMSGSIZE when mc_forwarding != 0 (skb->dev == NULL):
In inet6_rtm_getroute(), `dev = dev_get_by_index_rcu(net, iif)` is only
used for ip6_route_input_lookup(), and `skb->dev` is never assigned on
the newly allocated reply skb (`skb = alloc_skb(NLMSG_GOODSIZE, ...)`).
In addition, `rt = dst_rt6_info(skb_dst(skb))` is the per-CPU route
initialized by ip6_rt_copy_init() from `f6i->fib6_dst` (the `ff00::/8`
prefix installed by addrconf_add_mroute()) rather than the queried
(S,G). Consequently, on a cache miss in ip6mr_get_route(), `skb->dev` is
NULL regardless of `mc_forwarding` and ip6mr_get_route() returns
-ENODEV, which v1 mapped to -EMSGSIZE at `nla_put_failure`.
In the v2 series [1], patch 2/2 ("ipv6: preserve ip6mr_get_route() error
in rt6_fill_node()") already implements the error-preservation approach
mentioned in the review: instead of overwriting negative errors from
ip6mr_get_route() with -EMSGSIZE, rt6_fill_node() preserves and returns
the negative error (such as -ENODEV). Note that simply assigning
`skb->dev = dev` in inet6_rtm_getroute() without reworking
ip6mr_get_route() would not be safe on its own: on an unresolved cache
miss, ip6mr_get_route() allocates a 40-byte `skb2` (`sizeof(struct
ipv6hdr)`) without a netlink header and queues it onto `nl_queries`,
where ip6mr_destroy_unres() / ip6mr_cache_resolve() later pull
`sizeof(struct ipv6hdr)` and write `struct nlmsghdr` into an empty skb
(while inet6_rtm_getroute() also lacks `err == 0` handling and would
send an unfinalized reply skb without nlmsg_end()).
In v3, we also refine the commit message of patch 1/2 so it does not
imply `skb->dev == NULL` is caused by `mc_forwarding == 0`.
(b) [Medium 2] Excluding interface-local / link-local multicast and checking
`dest` instead of `rt6_dst->addr`:
This is a valid gap still present in v2. IPv4 rt_fill_info() guards
ipmr_get_route() with `!ipv4_is_local_multicast(dst)` alongside
`IPV4_DEVCONF_ALL_RO(net, MC_FORWARDING)`. Likewise, IPv6 ip6_mc_input()
never forwards groups matching `(IPV6_ADDR_LOOPBACK |
IPV6_ADDR_LINKLOCAL)` (i.e. interface-local / node-local `ff01::/16` and
link-local `ff02::/16`), so no MFC state can ever exist for them.
Furthermore, in normal route-get queries (`!fibmatch`), `rt6_dst->addr`
is the matched FIB6 prefix `ff00::` (from `ff00::/8`), whose scope is 0,
while the queried group address is passed in `dest` (`&fl6.daddr`). In
`fibmatch` queries (`dest == NULL`), `&rt6_dst->addr` holds the route
prefix. Therefore, checking `mc_dst = dest ? : &rt6_dst->addr` against
`(IPV6_ADDR_LOOPBACK | IPV6_ADDR_LINKLOCAL)` is required so that
interface-local and link-local multicast queries bypass
ip6mr_get_route() and emit RTA_IIF even when `mc_forwarding != 0`.
(c) [Low] Fixes tag / net vs net-next:
Because this series aligns long-standing non-crash route-get reporting
semantics with IPv4 (and full `mc_forwarding != 0` MFC lookup in
inet6_rtm_getroute() + ip6mr_get_route() remains a separate topic), we
target net-next without Fixes tags. If maintainers prefer targeting net
with Fixes tags, we will be happy to adjust.
2. Concrete Modification Plan (v3 Series):
Update patch 1/2 in `rt6_fill_node()` (`net/ipv6/route.c`) to inspect
`mc_dst = dest ? : &rt6_dst->addr` and exclude `IPV6_ADDR_LOOPBACK |
IPV6_ADDR_LINKLOCAL` before calling `ip6mr_get_route()` (and drop Hangbin's
Reviewed-by tag on patch 1/2 since its code changes, while keeping it on
unchanged patch 2/2):
```c
if (iif) {
#ifdef CONFIG_IPV6_MROUTE
const struct in6_addr *mc_dst = dest ? : &rt6_dst->addr;
if (ipv6_addr_is_multicast(mc_dst) &&
!(ipv6_addr_type(mc_dst) &
(IPV6_ADDR_LOOPBACK | IPV6_ADDR_LINKLOCAL)) &&
atomic_read(&net->ipv6.devconf_all->mc_forwarding)) {
int ret = ip6mr_get_route(net, skb, rtm, portid);
if (ret == 0)
return 0;
if (ret < 0) {
err = ret;
goto nla_put_failure;
}
} else
#endif
if (nla_put_u32(skb, RTA_IIF, iif))
goto nla_put_failure;
}
```
3. Reproduction & Verification Steps and Results (UML):
Setup in UML (`CONFIG_IPV6_MROUTE=y`):
ip link add d0 type veth peer name d1
ip link set d0 up && ip link set d1 up
ip -6 addr add 2001:db8::2/64 dev d0
ip -6 addr add ff0e::1 dev d0 autojoin
To test `mc_forwarding=1`, a helper (`test_mc_err`) opens an `AF_INET6`
`SOCK_RAW` (`IPPROTO_ICMPV6`) socket, enables `setsockopt(..., MRT6_INIT,
&one, sizeof(one))`, verifies `/proc/sys/net/ipv6/conf/all/mc_forwarding` is
`1`, and runs `ip -6 route get ... iif d0`:
- Before v3 (with v2 applied, `mc_forwarding=1` via `MRT6_INIT`):
* `ip -6 route get ff0e::1 from 2001:db8::3 iif d0`
-> `RTNETLINK answers: No such device` (`-ENODEV`, preserved by 2/2)
* `ip -6 route get ff02::1 from 2001:db8::3 iif d0`
-> `RTNETLINK answers: No such device` (`ip_exit=512`, FAIL: entered
`ip6mr_get_route()` for link-local multicast)
* `ip -6 route get ff01::1 from 2001:db8::3 iif d0`
-> `RTNETLINK answers: No such device` (`ip_exit=512`, FAIL: entered
`ip6mr_get_route()` for interface-local multicast)
- With v3 applied:
* Test 1 (`mc_forwarding=0`, routable `ff0e::1` with `iif d0`):
`ff0e::1 dev d0 proto kernel metric 256 iif d0 pref medium` (PASS)
* Test 2 (`mc_forwarding=1`, routable `ff0e::1` without VIF/cache):
`RTNETLINK answers: No such device` (`-ENODEV`, PASS)
* Test 3 (`mc_forwarding=1`, link-local `ff02::1` with `iif d0`):
`multicast ff02::1 dev d0 proto kernel metric 256 iif d0 pref medium`
(`ip_exit=0`, PASS)
* Test 4 (`mc_forwarding=1`, interface-local `ff01::1` with `iif d0`):
`multicast ff01::1 dev d0 proto kernel metric 256 iif d0 pref medium`
(`ip_exit=0`, PASS)
* Tests 5-7 (IPv6 unicast with `iif`, IPv6 multicast without `iif`, and
IPv4 `224.0.0.1` / `239.1.1.1` with `iif` regressions): all PASS
(`FAILURES=0, TEST_EXIT_CODE=0`).
[1] https://lore.kernel.org/netdev/20261112094630.197244-1-sunwenbo@xxxxxxxxxxxxxxxxxxxx/
Thanks,
Wenbo
pw-bot: cr