Messages in this thread | | | Date | Mon, 22 Jul 2019 22:00:34 +0200 | From | Peter Zijlstra <> | Subject | Re: [PATCH] sched/cputime: make scale_stime() more precise |
| |
On Mon, Jul 22, 2019 at 12:52:41PM +0200, Stanislaw Gruszka wrote: > On Fri, Jul 19, 2019 at 01:03:49PM +0200, Peter Zijlstra wrote: > > > shows the problem even when sum_exec_runtime is not that big: 300000 secs. > > > > > > The new implementation of scale_stime() does the additional div64_u64_rem() > > > in a loop but see the comment, as long it is used by cputime_adjust() this > > > can happen only once. > > > > That only shows something after long long staring :/ There's no words on > > what the output actually means or what would've been expected. > > > > Also, your example is incomplete; the below is a test for scale_stime(); > > from this we can see that the division results in too large a number, > > but, important for our use-case in cputime_adjust(), it is a step > > function (due to loss in precision) and for every plateau we shift > > runtime into the wrong bucket. > > > > Your proposed function works; but is atrocious, esp. on 32bit. That > > said, before we 'fixed' it, it had similar horrible divisions in, see > > commit 55eaa7c1f511 ("sched: Avoid cputime scaling overflow"). > > > > Included below is also an x86_64 implementation in 2 instructions. > > > > I'm still trying see if there's anything saner we can do... > > I was always proponent of removing scaling and export raw values > and sum_exec_runtime. But that has obvious drawback, reintroduce > 'top hiding' issue.
I think (but didn't grep) that we actually export sum_exec_runtime in /proc/ *somewhere*.
> But maybe we can export raw values in separate file i.e. > /proc/[pid]/raw_cpu_times ? So applications that require more precise > cputime values for very long-living processes can use this file.
There are no raw cpu_times, there are system and user samples, and samples are, by their very nature, an approximation. We just happen to track the samples in TICK_NSEC resolution these days, but they're still ticks (except on s390 and maybe other archs, which do time accounting in the syscall path).
But I think you'll find x86 people are quite opposed to doing TSC reads in syscall entry and exit :-)
| |