From 8235c4747dcc92de2ea991f78cdf9c6b8fa7f522 Mon Sep 17 00:00:00 2001 From: "Ondrej Zajicek (work)" Date: Mon, 15 Jul 2019 16:23:18 +0200 Subject: Netlink: Use route replace for IPv4 Use route replace netlink op instead of delete+add netlink ops for kernel IPv4 route replace. This avoids some packetloss during route replace. Still use the old behavior for IPv6, as some kernel bugs are hidden in IPv6 ECMP handling. --- sysdep/linux/netlink.c | 37 +++++++++++++++++++++++++++---------- 1 file changed, 27 insertions(+), 10 deletions(-) (limited to 'sysdep/linux/netlink.c') diff --git a/sysdep/linux/netlink.c b/sysdep/linux/netlink.c index 6e3fe488..85067bf5 100644 --- a/sysdep/linux/netlink.c +++ b/sysdep/linux/netlink.c @@ -69,6 +69,7 @@ #define RTA_ENCAP 22 #endif +#define krt_ipv4(p) ((p)->af == AF_INET) #define krt_ecmp6(p) ((p)->af == AF_INET6) const int rt_default_ecmp = 16; @@ -1368,27 +1369,43 @@ nl_delete_rte(struct krt_proto *p, rte *e) return err; } +static inline int +nl_replace_rte(struct krt_proto *p, rte *e) +{ + rta *a = e->attrs; + return nl_send_route(p, e, NL_OP_REPLACE, a->dest, &(a->nh)); +} + + void krt_replace_rte(struct krt_proto *p, net *n, rte *new, rte *old) { int err = 0; /* - * We could use NL_OP_REPLACE, but route replace on Linux has some problems: + * We use NL_OP_REPLACE for IPv4, it has an issue with not checking for + * matching rtm_protocol, but that is OK when dedicated priority is used. * - * 1) Does not check for matching rtm_protocol - * 2) Has broken semantics for IPv6 ECMP - * 3) Crashes some kernel version when used for IPv6 ECMP + * We do not use NL_OP_REPLACE for IPv6, as it has broken semantics for ECMP + * and with some kernel versions ECMP replace crashes kernel. Would need more + * testing and checks for kernel versions. * - * So we use NL_OP_DELETE and then NL_OP_ADD. We also do not trust the old - * route value, so we do not try to optimize IPv6 ECMP reconfigurations. + * For IPv6, we use NL_OP_DELETE and then NL_OP_ADD. We also do not trust the + * old route value, so we do not try to optimize IPv6 ECMP reconfigurations. */ - if (old) - nl_delete_rte(p, old); + if (krt_ipv4(p) && old && new) + { + err = nl_replace_rte(p, new); + } + else + { + if (old) + nl_delete_rte(p, old); - if (new) - err = nl_add_rte(p, new); + if (new) + err = nl_add_rte(p, new); + } if (err < 0) n->n.flags |= KRF_SYNC_ERROR; -- cgit v1.2.3