On Wed, Jan 17, 2018 at 03:56:43PM -0800, Linus Torvalds wrote: > On Wed, Jan 17, 2018 at 10:07 AM, Frederic Weisbecker > <frede...@kernel.org> wrote: > > > > I see, so you may want to test (possibly much) higher values of > > MAX_SOFTIRQ_RESTART, > > such as 50 or 100. > > I suspect the "number of softiqs per jiffy" is hardly interesting at all. > > We used to allow up to 2mS or ten iterations per _invocation_, never > mind per timer tick. > > I thought you were going to actally account for time, but I don't > think you ended up doing that.
I did in the first version but then I thought you suggested that count per jiffy. I probably misunderstood :) > > Maybe time isn't necessarily the thing to do, but just pure "count per > jiffy" seems very bad. Indeed, the more I think about it, the more doubts I have too. At least I started to think that this metric alone is not enough. > > What I might suggest using instead: > > - do it by time. This may be too expensive, though. Keeping track of > ns-level timing per invocation can be nasty. Yeah I would like to avoid that if we can. I guess it's ok if it sums up to rdtsc but I fear it's common to have a heavier version. > > - do it by "we got a new softirq event while handling another softirq > event". That was our old count per invocation, except you could do it > per softirq, and just allow *one* (ie keep a bitmask of "I've already > handled this softirq", and if the restart results in it being > triggered *again* you say "ok, I'll just move this to a workqueue" That one is very tempting. > > - .. something else? > > I'd suggest trying the "if we get a new softirq event that we've > already seen while we were already handling softirq events" thing. > That should really take care of the networking case of "90% time spend > in softirq handling during packet storms" thing. If we spend that much > time on softirqs, we *will* get a new softirq while handling an old > one occasionally. Ok I'm going to try that for the v3. Thanks.