For the record: As mentioned into the mailing-list [1], adding "options RADIX_MPATH" doesn't allow any more to use ECMP. The new ip_tryforward (like the previous fastforward) didn't support ECMP: a FreeBSD router sticks to its first route entry and never try the second entry. [1] https://lists.freebsd.org/pipermail/freebsd-net/2017-February/047203.html
Created attachment 206809 [details] ip_tryforward_rout_mpath @Olivier I assume this is still an issue in stable/11,12 and CURRENT, can you confirm? ^Triage: Attach patch from Ermal (eri@) on the mailing list thread, which may (likely?) need a re-base against latest CURRENT
Add ae@ to CC as they mentioned some code would need to be written (which was), in case they can assist with review
(In reply to Kubilay Kocak from comment #1) Yes this is still in issue on stable/11/12/current.
Created attachment 208751 [details] add radix_mpath.h back IPSEC policy or set sendredirect to "1" would disable tryforward(). Then the packet will fall into the slowpath. AKA. normal ip_input process. But ip_input.c missed some .h files, eg. radix_mpath.h, then ECMP is broken. Set sendredirect to "1" is not a good idea for a router, so I add "fastforwarding" back. Now you can set it to "0" to disable fastpath. The attached patch I've tested on FB12-stable r354079 is passed.
Multipath routing support has been rewritten in FreeBSD 13. It's dubbed ROUTE_MPATH and it is included in the GENERIC kernel. In the new implementation, ECMP works for both IPv4/IPv6 fast/slow path.
(In reply to Alexander V. Chernikov from comment #5) Is that to say we can/want to resolve this OBE? One question is what this means for stable/12 users (supported version)
(In reply to Kubilay Kocak from comment #6) Currently, I don't have plans for merging route stack changes to 12. Someone can try to merge/test the changes in the diff to check if they fix the situation in 12.
^Triage: Resolved (partial OBE) via new ROUTE_MPATH (13), without merge to stable 12/11, so it doesn't resolve the issue as reported (for 11.x). If this is inaccurate, please re-open the issue with additional details.