Messages in this thread | | | Date | Fri, 12 Jan 2018 19:54:29 +0100 | From | Frederic Weisbecker <> | Subject | Re: [RFC PATCH 1/2] softirq: Account time and iteration stats per vector |
| |
On Fri, Jan 12, 2018 at 10:12:32AM -0800, Linus Torvalds wrote: > On Fri, Jan 12, 2018 at 6:34 AM, Frederic Weisbecker > <frederic@kernel.org> wrote: > > > > That's right. But I thought it was bit large for the stack: > > > > struct { > > u64 time; > > u64 count; > > } [NR_SOFTIRQS] > > Note that you definitely don't want "u64" here. > > Both of these values had better be very limited. The "count" is on the > order of 10 - it fits in 4 _bits_ without any overflow. > > And 'time' is on the order of 2ms, so even if it's in nanoseconds, we > already know that we want to limit it to a single ms or so (yes, yes, > right now our limit is 2ms, but I think that's long). So even that > doesn't need 64-bit.
Ok.
> > Finally, I think you can join them. If we do a "time or count" limit, > let's just make the "count" act as some arbitrary fixed time, so that > we limit things that way. > > Say, if we want to limit it to 2ms, consider one count to be 0.2ms. So > instead of keeping track of count at all, just say "make each softirq > call count as at least 200,000ns even if the scheduler clock says it's > less". End result: we'd loop at most ten times. > > So now you only need one value, and you know it can't be bigger than 2 > million, so it can be a 32-bit one. Boom. Done.
Right.
Now I believe that the time was added as a limit because count alone was not reliable enough to diagnose a softirq overrun. But if everyone is fine with keeping the count as a single metric, I would be much happier because that means less overhead, no need to fetch the clock, etc...
> > Also, don't you want these to be percpu, and keep accumulating them > until you decide to either age them away (just clear it in timer > interrupt?) or if the value gets so big that you want o fall back to > the thread instead (and then the thread can clear it every iteration, > so you don't need to track whether the thread is active or not). > > I don't know. I'm traveling today, so I didn't actually have time to > really look at the patches, I'm just reacting to Eric's reaction.
Clearing the accumulation on tick and flush, that sounds like a good plan. Well I'm probably not going to use the tick for that because of nohz (again) but I can check if jiffies changed since we started the accumulation and reset it if so.
I'm going to respin, thanks!
| |