On Tue, Jan 9, 2018 at 10:58 AM, Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote: > On Tue, Jan 9, 2018 at 9:57 AM, Eric Dumazet <edumazet@xxxxxxxxxx> wrote: >> >> Your patch considers TASKLET_SOFTIRQ being a candidate for 'immediate >> handling', but TCP Small queues heavily use TASKLET, >> so as far as I am concerned a revert would have the same effect. > > Does it actually? > > TCP ends up dropping packets outside of the window etc, so flooding a > machine with TCP packets and causing some further processing up the > stack sounds very different from the basic packet flooding thing that > happens with NET_RX_SOFTIRQ. > > Also, honestly, the kinds of people who really worry about flooding > tend to have packet filtering in the receive path etc. > > So I really think "you can use up 90% of CPU time with a UDP packet > flood from the same network" is very very very different - and > honestly not at all as important - as "you want to be able to use a > USB DVB receiver and watch/record TV". > > Because that whole "UDP packet flood from the same network" really is > something you _fundamentally_ have other mitigations for. > > I bet that whole commit was introduced because of a benchmark test, > rather than real life. No? > > In contrast, now people are complaining about real loads not working. > > Linus I said that a revert was fine, maybe I was not clear. Clearly we can not touch anything scheduler related without breaking someone workload/assumptions on how system behaved at some point. Your patch wont solve other workloads that might have been impacted by my patch, so in one year (or next week), we will have to cope with another device driver not using tasklet but still relying on immediate softirq processing. Apparently, we have to live with softirq model forever, or switch to RT kernels. Note that we have no mitigation for something that involve flood of valid packets that no firewall can drop (without dropping legitimate packets). The 'benchmark' here is not really the trigger, only a tool validating an idea/patch.
