On Wed, 25 Jan 2017 16:25:20 +0100 Frederic Weisbecker <fweis...@gmail.com> wrote:
> On Mon, Jan 23, 2017 at 10:44:56AM +0100, Martin Schwidefsky wrote: > > On Sun, 22 Jan 2017 19:20:13 +0100 > > Frederic Weisbecker <fweis...@gmail.com> wrote: > > > > > The s390 clock has a higher granularity than nanoseconds. 1 nanosec > > > equals 4.096 in s390 cputime_t. Therefore we leak a remainder while > > > flushing the cputime through cputime_to_nsecs(). > > > > > > For more precision, make sure we keep that remainder on cputime > > > accumulators for later accounting. > > > > > > Reported-by: Martin Schwidefsky <schwidef...@de.ibm.com> > > > Cc: Benjamin Herrenschmidt <b...@kernel.crashing.org> > > > Cc: Paul Mackerras <pau...@samba.org> > > > Cc: Michael Ellerman <m...@ellerman.id.au> > > > Cc: Heiko Carstens <heiko.carst...@de.ibm.com> > > > Cc: Martin Schwidefsky <schwidef...@de.ibm.com> > > > Cc: Tony Luck <tony.l...@intel.com> > > > Cc: Fenghua Yu <fenghua...@intel.com> > > > Cc: Peter Zijlstra <pet...@infradead.org> > > > Cc: Rik van Riel <r...@redhat.com> > > > Cc: Thomas Gleixner <t...@linutronix.de> > > > Cc: Ingo Molnar <mi...@kernel.org> > > > Cc: Stanislaw Gruszka <sgrus...@redhat.com> > > > Cc: Wanpeng Li <wanpeng...@hotmail.com> > > > Signed-off-by: Frederic Weisbecker <fweis...@gmail.com> > > > > NAK. Good intention but the patch is just broken. with 36 of the 37 > > patches applied all looks good but the last one completely breaks the > > accounting for s390. This is from an idle system: > > > > top - 10:39:33 up 0 min, 1 user, load average: 0,00, 0,00, 0,00 > > Tasks: 106 total, 1 running, 105 sleeping, 0 stopped, 0 zombie > > %Cpu0 : 8,9 us, 21,6 sy, 0,0 ni, 0,0 id, 0,0 wa, 10,8 hi, 4,3 si, > > 54,4 st > > %Cpu1 : 0,0 us, 23,5 sy, 0,0 ni, 0,0 id, 0,0 wa, 19,0 hi, 13,1 si, > > 44,3 st > > %Cpu2 : 0,0 us, 30,3 sy, 0,0 ni, 0,0 id, 0,0 wa, 14,7 hi, 14,8 si, > > 40,2 st > > KiB Mem : 1009304 total, 818808 free, 57284 used, 133212 buff/cache > > KiB Swap: 1048556 total, 1048556 free, 0 used. 917356 avail Mem > > Oh ok. I must have done something wrong. > > > > > There is another issue that affects precision, there is no s390 specific > > version of cputime_to_nsecs. The generic version uses cputime_to_usecs > > and mulitplies by 1000 to get nano-seconds. That already looses precision. > > That's right. And that's the point of this patch. I'm not sure we can have a > more precise version of cputime_to_nsecs() if 1 nsec == 4.096 cputime_t I am thinking about a version that does the calculation (cputime * 1000) >> 12 in several steps, like this nsecs = (((((cputime * 5) >> 3) * 5) >> 3) * 5) >> 3 To multiply with 5 is a shift by 2 and an add. In total that would gives us 6 shifts and 3 adds for the conversion. -- blue skies, Martin. "Reality continues to ruin my life." - Calvin.
