hello, my 2 cents:
On Thu, 2020-06-25 at 21:53 +0200, Toke Høiland-Jørgensen wrote: > I think it depends a little on the use case; some callers actually care > about the VLAN tags themselves and handle that specially (e.g., > act_csum). I remember taht something similar was discussed about 1 year ago [1]. > Whereas others (e.g., sch_dsmark) probably will have the same > issue. I'd say that the issue "propagates" to all qdiscs that mangle the ECN-CE bit (i.e., calling INET_ECN_set_ce() [2]), most notably all the RED variants and "codel/fq_codel". > I guess I can trying going through them all and figuring out if > there's a more generic solution. For sch_cake, I think that the qdisc shouldn't look at the IP header when it schedules packets having a VLAN tag. Probably, when tc_skb_protocol() returns ETH_P_8021Q or ETH_P_8021AD, we should look at the VLAN priority (PCP) bits (and that's something that cake doesn't do currently - but I have a small patch in my stash that implements this: please let me know if you are interested in seeing it :) ). Then, to ensure that the IP precedence is respected, even with different VLAN tags, users should explicitly configure TC filters that "map" the DSCP value to a PCP value. This would ensure that configured priority is respected by the scheduler, and would also be flexible enough to allow different "mappings". Sure, my proposal does not cover the problem of mangling the CE bit inside VLAN-tagged packets, i.e. if we should understand if qdiscs should allow it or not. WDYT? thank you in advance! -- davide [1] https://lore.kernel.org/netdev/CAM_iQpUmuHH8S35ERuJ-sFS=17aa-c8uhswf-wf7toanx2e...@mail.gmail.com/#t [2] https://elixir.bootlin.com/linux/latest/C/ident/INET_ECN_set_ce
