On Sun, 17 Jan 2021 10:02:18 +0200 Ido Schimmel wrote: > From: Ido Schimmel <ido...@nvidia.com> > > The RED qdisc currently supports two qevents: "early_drop" and "mark". The > filters added to the block bound to the "early_drop" qevent are executed on > packets for which the RED algorithm decides that they should be > early-dropped. The "mark" filters are similarly executed on ECT packets > that are marked as ECN-CE (Congestion Encountered). > > A previous patchset has offloaded "early_drop" filters on Spectrum-2 and > later, provided that the classifier used is "matchall", that the action > used is either "trap" or "mirred", and a handful or further limitations.
For early_drop trap or mirred makes obvious sense, no explanation needed. But for marked as a user I'd like to see a _copy_ of the packet, while the original continues on its marry way to the destination. I'd venture to say that e.g. for a DCTCP deployment mark+trap is unusable, at least for tracing, because it distorts the operation by effectively dropping instead of marking. Am I reading this right? If that is the case and you really want to keep the mark+trap functionality - I feel like at least better documentation is needed. The current two liner should also be rewritten, quoting from patch 1: > * - ``ecn_mark`` > - ``drop`` > - Traps ECN-capable packets that were marked with CE (Congestion > Encountered) code point by RED algorithm instead of being dropped That needs to say that the trap is for datagrams trapped by a qevent. Otherwise "Traps ... instead of being dropped" is too much of a thought-shortcut, marked packets are not dropped. (I'd also think that trap is better documented next to early_drop, let's look at it from the reader's perspective)
