Re: sched_mc_power_savings broken with CGROUPS+CPUSETS

[Peter Zijlstra <peterz@infradead.org>]

On Fri, 2008-08-29 at 13:29 -0700, Max Krasnyansky wrote:
> Peter Zijlstra wrote:
> > On Fri, 2008-08-29 at 18:45 +0530, Vaidyanathan Srinivasan wrote:
> >> Hi,
> >>
> >> sched_mc_power_savings seems to be broken with CGROUPS+CPUSETS.
> >> When CONFIG_CPUSETS=y the attached BUG_ON() is being hit.
> >>
> >> I added a BUG_ON to check if SD_POWERSAVINGS_BALANCE is set at
> >> SD_LV_CPU whenever sched_mc_power_savings is set.
> >>
> >> This BUG is hit when config CONFIG_CPUSETS (depends on CONFIG_CGROUPS)
> >> is just compiled in while this is never hit when they are compiled
> >> out.  The fact that SD_POWERSAVINGS_BALANCE being cleared even when
> >> sched_mc_power_savings = 1 completely breaks the
> >> sched_mc_power_savings heuristics.
> >>
> >> To recreate the problem,
> >> Have sched_mc power savings enabled CONFIG_SCHED_MC=y
> >> Add this BUG_ON()
> >>
> >> echo 1 > /sys/devices/system/cpu/sched_mc_power_savings
> >>
> >> Try these these on a multi core x86 box.
> >>
> >> sched_mc_power_savings seems to be broken from 2.6.26-rc1, but
> >> I do not have a confirmation that the root cause is same in all
> >> successive versions. sched_mc_power_savings works perfect in
> >> 2.6.25.
> >>
> >> Please help me root cause the issue.  Please point me to changes that
> >> may potential cause this bug.
> > 
> > I'm still greatly mistified by all that power savings code.
> > 
> > Its hard to read and utterly hard to comprehend - I've been about to rip
> > the whole stuff out on several occasions. But so far tried to carefully
> > thread around it maintaining its operation even though not fully
> > understood.
> > 
> > Someone with clue - preferably the authors of the code in question -
> > should enlighten us with a patch that adds some comments as to the
> > intent of said lines of code.
> 
> I do not fully understand how balancing is affected by the MC stuff but I can
> explain how the mc power saving settings are applied to the domains and the
> overall mechanism for that.
> Here a quote from one of my emails to Paul
> 	
> > Max wrote:
> > ...
> > Those things (mc_power and topology updates) have to update domain flags based
> > on the mc/smt power and current topology settings.
> > This is done in the
> >   __rebuild_sched_domains()
> >        ...
> >        SD_INIT(sd, ALLNODES);
> >        ...
> >        SD_INIT(sd, MC);
> >        ...
> > 
> > SD_INIT(sd,X) uses one of SD initializers defined in the include/linux/topology.h
> > For example SD_CPU_INIT() includes BALANCE_FOR_PKG_POWER which expands to
> > 
> > #define BALANCE_FOR_PKG_POWER   \
> >         ((sched_mc_power_savings || sched_smt_power_savings) ?  \
> >          SD_POWERSAVINGS_BALANCE : 0)
> > 
> > Yes it's kind convoluted :). Anyway, the point is that we need to rebuild the
> > domains when those settings change. We could probably write a simpler version
> > that just iterates existing domains and updates the flags. Maybe some other dat :)

I don't think iterating the domains and setting the flag is sufficient.
Look at this crap (found in arch/x86/kernel/smpboot.c):

cpumask_t cpu_coregroup_map(int cpu)
{
        struct cpuinfo_x86 *c = &cpu_data(cpu);
        /*
         * For perf, we return last level cache shared map.
         * And for power savings, we return cpu_core_map
         */
        if (sched_mc_power_savings || sched_smt_power_savings)
                return per_cpu(cpu_core_map, cpu);
        else
                return c->llc_shared_map;
}

which means we'll actually end up building different domain/group
configurations depending on power savings settings.

> As I explained in the previous reply I missed the fact the logic that avoids
> redundant rebuilds in partition_sched_domains() will prevent
> arch_reinit_sched_domains() from doing the actual rebuild and hence will not
> apply the SD_POWERSAVINGS_BALANCE until something changes in cpuset setup.
> 
> btw I can certainly attest to the fact that powersaving code is very hard to
> read and comprehend :)

Yeah - I was primarity hinting at the sched_group and find_*_group()
fudge, esp find_busiest_group() is an utter nightmare.

I'm still struggeling to understand _why_ we need those group things to
begin with, why aren't the child domains good enough?