Re: [PATCH RFC] cpumask: Randomly distribute the tasks within affinity mask

[Peter Zijlstra <peterz@infradead.org>]

On Wed, Oct 11, 2023 at 12:53:29PM +0200, Peter Zijlstra wrote:
> On Wed, Oct 11, 2023 at 12:49:25PM +0530, Ankit Jain wrote:
> > commit 46a87b3851f0 ("sched/core: Distribute tasks within affinity masks")
> > and commit 14e292f8d453 ("sched,rt: Use cpumask_any*_distribute()")
> > introduced the logic to distribute the tasks at initial wakeup on cpus
> > where load balancing works poorly or disabled at all (isolated cpus).
> > 
> > There are cases in which the distribution of tasks
> > that are spawned on isolcpus does not happen properly.
> > In production deployment, initial wakeup of tasks spawn from
> > housekeeping cpus to isolcpus[nohz_full cpu] happens on first cpu
> > within isolcpus range instead of distributed across isolcpus.
> > 
> > Usage of distribute_cpu_mask_prev from one processes group,
> > will clobber previous value of another or other groups and vice-versa.
> > 
> > When housekeeping cpus spawn multiple child tasks to wakeup on
> > isolcpus[nohz_full cpu], using cpusets.cpus/sched_setaffinity(),
> > distribution is currently performed based on per-cpu
> > distribute_cpu_mask_prev counter.
> > At the same time, on housekeeping cpus there are percpu
> > bounded timers interrupt/rcu threads and other system/user tasks
> > would be running with affinity as housekeeping cpus. In a real-life
> > environment, housekeeping cpus are much fewer and are too much loaded.
> > So, distribute_cpu_mask_prev value from these tasks impacts
> > the offset value for the tasks spawning to wakeup on isolcpus and
> > thus most of the tasks end up waking up on first cpu within the
> > isolcpus set.
> > 
> > Steps to reproduce:
> > Kernel cmdline parameters:
> > isolcpus=2-5 skew_tick=1 nohz=on nohz_full=2-5
> > rcu_nocbs=2-5 rcu_nocb_poll idle=poll irqaffinity=0-1
> > 
> > * pid=$(echo $$)
> > * taskset -pc 0 $pid
> > * cat loop-normal.c
> > int main(void)
> > {
> >         while (1)
> >                 ;
> >         return 0;
> > }
> > * gcc -o loop-normal loop-normal.c
> > * for i in {1..50}; do ./loop-normal & done
> > * pids=$(ps -a | grep loop-normal | cut -d' ' -f5)
> > * for i in $pids; do taskset -pc 2-5 $i ; done
> > 
> > Expected output:
> > * All 50 “loop-normal” tasks should wake up on cpu2-5
> > equally distributed.
> > * ps -eLo cpuid,pid,tid,ppid,cls,psr,cls,cmd | grep "^    [2345]"
> > 
> > Actual output:
> > * All 50 “loop-normal” tasks got woken up on cpu2 only
> 
> Your expectation is wrong. Things work as advertised.

That is, isolcpus results in single CPU balance domains and as such we
must not distribute -- there is no load balancing.

Ideally we'd reject setting cpumasks with multi bits set on domains like
that, but alas, that would break historical behaviour :/

Now, looking at the code, I don't think the current code actually
behaves correct in this case :-(, somewhere along the line we should
truncate cpu_valid_mask to a single bit. Let me see where the sane place
is to do that.