Le 05/11/2024 à 04:18, Matt Muggeridge a écrit : > Add a Netlink rtm_flag, RTM_F_RA_ROUTER for the RTM_NEWROUTE message. > This allows an IPv6 Netlink client to indicate the default route came > from an RA. This results in the kernel creating individual default > routes, rather than coalescing multiple default routes into a single > ECMP route. > > Details: > > For IPv6, a Netlink client is unable to create default routes in the > same manner as the kernel. This leads to failures when there are > multiple default routers, as they were being coalesced into a single > ECMP route. When one of the ECMP default routers becomes UNREACHABLE, it > was still being selected as the nexthop. > > Meanwhile, when the kernel processes RAs from multiple default routers, > it sets the fib6_flags: RTF_ADDRCONF | RTF_DEFAULT. The RTF_ADDRCONF > flag is checked by rt6_qualify_for_ecmp(), which returns false when > ADDRCONF is set. As such, the kernel creates separate default routes. > > E.g. compare the routing tables when RAs are processed by the kernel > versus a Netlink client (systemd-networkd, in my case). > > 1) RA Processed by kernel (accept_ra = 2) > $ ip -6 route > 2001:2:0:1000::/64 dev enp0s9 proto kernel metric 256 expires ... > fe80::/64 dev enp0s9 proto kernel metric 256 pref medium > default via fe80::200:10ff:fe10:1060 dev enp0s9 proto ra ... > default via fe80::200:10ff:fe10:1061 dev enp0s9 proto ra ... > > 2) RA Processed by Netlink client (accept_ra = 0) > $ ip -6 route > 2001:2:0:1000::/64 dev enp0s9 proto ra metric 1024 expires ... > fe80::/64 dev enp0s3 proto kernel metric 256 pref medium > fe80::/64 dev enp0s9 proto kernel metric 256 pref medium > default proto ra metric 1024 expires 595sec pref medium > nexthop via fe80::200:10ff:fe10:1060 dev enp0s9 weight 1 > nexthop via fe80::200:10ff:fe10:1061 dev enp0s9 weight 1 > > IPv6 Netlink clients need a mechanism to identify a route as coming from > an RA. i.e. a Netlink client needs a method to set the kernel flags: > > RTF_ADDRCONF | RTF_DEFAULT > > This is needed when there are multiple default routers that each send > an RA. Setting the RTF_ADDRCONF flag ensures their fib entries do not > qualify for ECMP routes, see rt6_qualify_for_ecmp(). > > To achieve this, introduce a user-level flag RTM_F_RA_ROUTER that a > Netlink client can pass to the kernel. > > A Netlink user-level network manager, such as systemd-networkd, may set > the RTM_F_RA_ROUTER flag in the Netlink RTM_NEWROUTE rtmsg. When set, > the kernel sets RTF_RA_ROUTER in the fib6_config fc_flags. This causes a > default route to be created in the same way as if the kernel processed > the RA, via rt6add_dflt_router(). > > This is needed by user-level network managers, like systemd-networkd, > that prefer to do the RA processing themselves. ie. they disable the > kernel's RA processing by setting net.ipv6.conf.<intf>.accept_ra=0. > > Without this flag, when there are mutliple default routers, the kernel > coalesces multiple default routes into an ECMP route. The ECMP route > ignores per-route REACHABILITY information. If one of the default > routers is unresponsive, with a Neighbor Cache entry of INCOMPLETE, then > it can still be selected as the nexthop for outgoing packets. This > results in an inability to communicate with remote hosts, even though > one of the default routers remains REACHABLE. This violates RFC4861 > section 6.3.6, bullet 1. > > Extract from RFC4861 6.3.6 bullet 1: > 1) Routers that are reachable or probably reachable (i.e., in any > state other than INCOMPLETE) SHOULD be preferred over routers > whose reachability is unknown or suspect (i.e., in the > INCOMPLETE state, or for which no Neighbor Cache entry exists). > > This fixes the IPv6 Logo conformance test v6LC_2_2_11, and others that > test with multiple default routers. Also see systemd issue #33470: > https://github.com/systemd/systemd/issues/33470. > > Signed-off-by: Matt Muggeridge <Matt.Muggeridge@xxxxxxx> > Cc: David Ahern <dsahern@xxxxxxxxxx> > Cc: David S. Miller <davem@xxxxxxxxxxxxx> > Cc: linux-api@xxxxxxxxxxxxxxx > Cc: stable@xxxxxxxxxxxxxxx > --- > include/uapi/linux/rtnetlink.h | 9 +++++---- > net/ipv6/route.c | 3 +++ > 2 files changed, 8 insertions(+), 4 deletions(-) > > diff --git a/include/uapi/linux/rtnetlink.h b/include/uapi/linux/rtnetlink.h > index 3b687d20c9ed..9f0259f6e4ed 100644 > --- a/include/uapi/linux/rtnetlink.h > +++ b/include/uapi/linux/rtnetlink.h > @@ -202,7 +202,7 @@ enum { > #define RTM_NR_FAMILIES (RTM_NR_MSGTYPES >> 2) > #define RTM_FAM(cmd) (((cmd) - RTM_BASE) >> 2) > > -/* > +/* > Generic structure for encapsulation of optional route information. > It is reminiscent of sockaddr, but with sa_family replaced > with attribute type. > @@ -242,7 +242,7 @@ struct rtmsg { > > unsigned char rtm_table; /* Routing table id */ > unsigned char rtm_protocol; /* Routing protocol; see below */ > - unsigned char rtm_scope; /* See below */ > + unsigned char rtm_scope; /* See below */ > unsigned char rtm_type; /* See below */ > > unsigned rtm_flags; > @@ -336,6 +336,7 @@ enum rt_scope_t { > #define RTM_F_FIB_MATCH 0x2000 /* return full fib lookup match */ > #define RTM_F_OFFLOAD 0x4000 /* route is offloaded */ > #define RTM_F_TRAP 0x8000 /* route is trapping packets */ > +#define RTM_F_RA_ROUTER 0x10000 /* route is a default route from RA */ Please, don't mix whitespace changes with the changes related to the new flag. https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/Documentation/process/submitting-patches.rst#n166 Regards, Nicolas
