Reduce advertised ipv6 mtu by 16 to fit ethernet header

This is a temporary hack to workaround the inability of current
kernel's ebpf bpf_skb_change_mode() function to prefix a 14-byte
ethernet header on to a packet without going over the upstream
(source, rawip) interface's mtu *before* we bpf_redirect() to
the downstream (destination, ethernet) interface.

Test: build, atest, atest TetheringTests
Bug: 149816401
Signed-off-by: Maciej Żenczykowski <maze@google.com>
Change-Id: I646148ebfd978a2489c0cd065e4b671b01150add
This commit is contained in:
Maciej Żenczykowski
2020-02-19 01:24:39 -08:00
parent 907976e0a2
commit da0fb1bca8

View File

@@ -541,7 +541,12 @@ public class IpServer extends StateMachine {
if (v6only != null) {
params = new RaParams();
params.mtu = v6only.getMtu();
// We advertise an mtu lower by 16, which is the closest multiple of 8 >= 14,
// the ethernet header size. This makes kernel ebpf tethering offload happy.
// This hack should be reverted once we have the kernel fixed up.
// Note: this will automatically clamp to at least 1280 (ipv6 minimum mtu)
// see RouterAdvertisementDaemon.java putMtu()
params.mtu = v6only.getMtu() - 16;
params.hasDefaultRoute = v6only.hasIpv6DefaultRoute();
if (params.hasDefaultRoute) params.hopLimit = getHopLimit(v6only.getInterfaceName());