Re: [PATCH 8/9] psi: pressure stall information for CPU, memory, and IO

[Johannes Weiner <hannes@cmpxchg.org>]

On Fri, Sep 07, 2018 at 12:16:34PM +0200, Peter Zijlstra wrote:
> On Tue, Aug 28, 2018 at 01:22:57PM -0400, Johannes Weiner wrote:
> > +enum psi_states {
> > +	PSI_IO_SOME,
> > +	PSI_IO_FULL,
> > +	PSI_MEM_SOME,
> > +	PSI_MEM_FULL,
> > +	PSI_CPU_SOME,
> > +	/* Only per-CPU, to weigh the CPU in the global average: */
> > +	PSI_NONIDLE,
> > +	NR_PSI_STATES,
> > +};
> 
> > +static u32 get_recent_time(struct psi_group *group, int cpu,
> > +			   enum psi_states state)
> > +{
> > +	struct psi_group_cpu *groupc = per_cpu_ptr(group->pcpu, cpu);
> > +	unsigned int seq;
> > +	u32 time, delta;
> > +
> > +	do {
> > +		seq = read_seqcount_begin(&groupc->seq);
> > +
> > +		time = groupc->times[state];
> > +		/*
> > +		 * In addition to already concluded states, we also
> > +		 * incorporate currently active states on the CPU,
> > +		 * since states may last for many sampling periods.
> > +		 *
> > +		 * This way we keep our delta sampling buckets small
> > +		 * (u32) and our reported pressure close to what's
> > +		 * actually happening.
> > +		 */
> > +		if (test_state(groupc->tasks, state))
> > +			time += cpu_clock(cpu) - groupc->state_start;
> > +	} while (read_seqcount_retry(&groupc->seq, seq));
> > +
> > +	delta = time - groupc->times_prev[state];
> > +	groupc->times_prev[state] = time;
> > +
> > +	return delta;
> > +}
> 
> > +static bool update_stats(struct psi_group *group)
> > +{
> > +	u64 deltas[NR_PSI_STATES - 1] = { 0, };
> > +	unsigned long missed_periods = 0;
> > +	unsigned long nonidle_total = 0;
> > +	u64 now, expires, period;
> > +	int cpu;
> > +	int s;
> > +
> > +	mutex_lock(&group->stat_lock);
> > +
> > +	/*
> > +	 * Collect the per-cpu time buckets and average them into a
> > +	 * single time sample that is normalized to wallclock time.
> > +	 *
> > +	 * For averaging, each CPU is weighted by its non-idle time in
> > +	 * the sampling period. This eliminates artifacts from uneven
> > +	 * loading, or even entirely idle CPUs.
> > +	 */
> > +	for_each_possible_cpu(cpu) {
> > +		u32 nonidle;
> > +
> > +		nonidle = get_recent_time(group, cpu, PSI_NONIDLE);
> > +		nonidle = nsecs_to_jiffies(nonidle);
> > +		nonidle_total += nonidle;
> > +
> > +		for (s = 0; s < PSI_NONIDLE; s++) {
> > +			u32 delta;
> > +
> > +			delta = get_recent_time(group, cpu, s);
> > +			deltas[s] += (u64)delta * nonidle;
> > +		}
> > +	}
> 
> This does the whole seqcount thing 6x, which is a bit of a waste.

[...]

> It's a bit cumbersome, but that's because of C.

I was actually debating exactly this with Suren before, but since this
is a super cold path I went with readability. I was also thinking that
restarts could happen quite regularly under heavy scheduler load, and
so keeping the individual retry sections small could be helpful - but
I didn't instrument this in any way.

No strong opinion from me, I can send an updated patch if you prefer.