From: Florian Westphal <f...@strlen.de> Date: Sun, 30 Jul 2017 03:57:17 +0200
> During a hallway discussion with Eric Dumazet at Netdev 1.2 in > Tokyo some maybe-not-so-useful-anymore TCP stack features came up, > among these header prediction and prequeueing. > > In brief, TCP prequeue assumes a single-process-blocking-read design, > which is not that common anymore. The most frequently used high-performance > networking program that is an excellent fit for these features is netperf. > > The idea behind prequeueing is to move part of tcp processing, including > retransmit queue cleaning, to process context. > > With (e)poll designs, prequeue is always skipped, so for such programs > this is dead-code removal. > > Header prediction is also less useful nowadays. > For packet trains, GRO will do packet aggregation so we do not get the > per-packet benefit that this had before GRO anymore. > > Because of SACK, header prediction also will be ineffective once > a connection suffers even light packet losses. > > code removal aside, after this change processing always occurs in BH > context, this allows to experiment e.g. with doing bulk freeing of > skb heads when incoming ACKs clean packets from the retransmit queue. > > There are no changes since the RFC, except in last patch (i missed > another no-longer-used mib counter). I also edited a few commit messages. Series applied, thanks a lot for doing this.
