Jerry/bird
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

Netlink: Use route replace for IPv4

Browse source 
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.
This commit is contained in:
Ondrej Zajicek (work) 2019-07-15 16:23:18 +02:00
parent 1aec7112f7
commit 8235c4747d
1 changed files with 27 additions and 10 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

37
sysdep/linux/netlink.c
View file

@ -69,6 +69,7 @@
#define RTA_ENCAP 22 #define RTA_ENCAP 22
#endif #endif
#define krt_ipv4(p) ((p)->af == AF_INET)
#define krt_ecmp6(p) ((p)->af == AF_INET6) #define krt_ecmp6(p) ((p)->af == AF_INET6)
const int rt_default_ecmp = 16; const int rt_default_ecmp = 16;
@ -1368,27 +1369,43 @@ nl_delete_rte(struct krt_proto *p, rte *e)
return err; 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 void
krt_replace_rte(struct krt_proto *p, net *n, rte *new, rte *old) krt_replace_rte(struct krt_proto *p, net *n, rte *new, rte *old)
{ {
int err = 0; 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 * We do not use NL_OP_REPLACE for IPv6, as it has broken semantics for ECMP
* 2) Has broken semantics for IPv6 ECMP * and with some kernel versions ECMP replace crashes kernel. Would need more
* 3) Crashes some kernel version when used for IPv6 ECMP * testing and checks for kernel versions.
* *
* So we use NL_OP_DELETE and then NL_OP_ADD. We also do not trust the old * For IPv6, we use NL_OP_DELETE and then NL_OP_ADD. We also do not trust the
* route value, so we do not try to optimize IPv6 ECMP reconfigurations. * old route value, so we do not try to optimize IPv6 ECMP reconfigurations.
*/ */
if (old) if (krt_ipv4(p) && old && new)
nl_delete_rte(p, old); {
err = nl_replace_rte(p, new);
}
else
{
if (old)
nl_delete_rte(p, old);
if (new) if (new)
err = nl_add_rte(p, new); err = nl_add_rte(p, new);
}
if (err < 0) if (err < 0)
n->n.flags |= KRF_SYNC_ERROR; n->n.flags |= KRF_SYNC_ERROR;