> Date: Mon, 14 Nov 2016 13:11:58 +0100
> From: Martin Pieuchot <[email protected]>
>
> On 12/11/16(Sat) 15:52, patrick keshishian wrote:
> > Ahh... seems the culprit is softclock_thread added 2016/09/22
> > (kern/kern_timeout.c mpi@).
>
> I'd suggest we simply skip kernel thread when calculating the load.
Hmm, I'd say that counting runnable kernel threads towards the load is
the right thing to do.
> Since we're slowly moving code executed in software interrupt
> context to kernel threads this will keep the original behavior.
A different way of viewing is that this move now enables us to
properly take into account the work done by the network stack.
> Index: uvm/uvm_meter.c
> ===================================================================
> RCS file: /cvs/src/sys/uvm/uvm_meter.c,v
> retrieving revision 1.36
> diff -u -p -r1.36 uvm_meter.c
> --- uvm/uvm_meter.c 14 Mar 2015 03:38:53 -0000 1.36
> +++ uvm/uvm_meter.c 14 Nov 2016 12:08:11 -0000
> @@ -107,6 +107,9 @@ uvm_loadav(struct loadavg *avg)
> memset(nrun_cpu, 0, sizeof(nrun_cpu));
>
> LIST_FOREACH(p, &allproc, p_list) {
> + if (p->p_flag & P_SYSTEM)
> + continue;
> +
> switch (p->p_stat) {
> case SSLEEP:
> if (p->p_priority > PZERO || p->p_slptime > 1)
> @@ -114,8 +117,6 @@ uvm_loadav(struct loadavg *avg)
> /* FALLTHROUGH */
> case SRUN:
> case SONPROC:
> - if (p == p->p_cpu->ci_schedstate.spc_idleproc)
> - continue;
> case SIDL:
> nrun++;
> if (p->p_cpu)
> @@ -136,7 +137,7 @@ uvm_loadav(struct loadavg *avg)
> spc->spc_ldavg = (cexp[0] * spc->spc_ldavg +
> nrun_cpu[CPU_INFO_UNIT(ci)] * FSCALE *
> (FSCALE - cexp[0])) >> FSHIFT;
> - }
> + }
> }
>
> /*
>
>
