Update: strace(1) shows the following in regard to Open vSwitch in an endless loop:
socket(AF_UNIX, SOCK_DGRAM|SOCK_CLOEXEC, 0) = 40 ioctl(40, SIOCGIFNAME, {ifr_index=6, ifr_name="vlan1105"}) = 0 close(40) = 0
This doesn't make any sense (at least, to me) since the interface vlan1105 isn't related to any bridge configuration at all. Only outbound routes (originating from BGP) point to an address via interface vlan1105. As the name implies this is just a VLAN on a direct physical network port. And BIRD (2.0.4-1):
poll([{fd=4, events=POLLIN}, {fd=7, events=POLLIN}, {fd=8, events=POLLIN}, {fd=9, events=POLLIN}], 4, 3000) = 1 ([{fd=7, revents=POLLIN}]) recvmsg(7, {msg_name={sa_family=AF_NETLINK, nl_pid=0, nl_groups=0x000040}, msg_namelen=12, msg_iov=[{iov_base=[{{len=68, type=0x19 /* NLMSG_??? */, flags=0, seq=806042, pid=2639368684}, "\2\26\0\0\376\f\0\1\0\0\0\0\10\0\17\0\376\0\0\0\10\0\1\0X\335d\0\10\0\6\0"...}, {{len=0, type=0 /* NLMSG_??? */, flags=0, seq=0, pid=0}}], iov_len=8192}], msg_iovlen=1, msg_controllen=0, msg_flags=0}, 0) = 68
There's just a little of polling going on but it seems as every poll results in the socket being opened by ovs-vswitchd, for each route present. And again, and again. Regards, Kees On 06-05-19 20:22, Kees Meijs wrote:
Sorry: shouldn't happen.
Meanwhile we tested BIRD 2.0.4 as well (compiled from source) with the same result. The process ovs-vswitchd completely consumes a CPU thread.
When disabling the exportation of the routes (full feed, so 700k+ routes) the load drops back to nothing.