[PATCH v6] sched/psi: Skip CPUs with zero non-idle delta in per-CPU aggregation

collect_percpu_times() iterates over every possible CPU to build a
non-idle-weighted average of the PSI state times. When a CPU has no
PSI_NONIDLE delta for the current sampling interval:

    nonidle    = nsecs_to_jiffies(times[PSI_NONIDLE]) = 0
    deltas[s] += times[s] * nonidle                /* += 0 */

so the weighted accumulation contributes nothing.

get_recent_times() already sets the PSI_NONIDLE bit in cpu_changed_states
iff the PSI_NONIDLE delta is non-zero. Use that bit to skip such CPUs
early, as suggested by Johannes, avoiding the nsecs_to_jiffies() call and
the PSI_NONIDLE * u64 mul-adds that follow.

No functional change: on the skipped path the old code adds zero to
deltas[] and zero to nonidle_total, which is exactly the result of not
iterating. The PSI_NONIDLE bit is folded into changed_states before the
skip, so the aggregator's reschedule decision is unaffected.

The cost is worth trimming because collect_percpu_times() is
O(nr_possible_cpus) per group and is not only called from the 2s
averaging work: when a PSI trigger is armed, psi_rtpoll_work() calls it at
the trigger cadence. On a 12-thread host with a "some 50000 500000"
trigger armed, bpftrace measured collect_percpu_times() running ~160
times/s (vs ~0.5/s on the averaging path), and on a partially loaded box a
large fraction of the per-CPU iterations per call had no PSI_NONIDLE delta
and are skipped.

read(/proc/pressure/cpu) median latency, QEMU/KVM A/B on mainline
v7.1-rc2+ (identical config, the patch the only difference), idle guest,
100k iterations, varying -smp:

    -smp    baseline   patched      delta
      2      1834 ns    1799 ns     -1.9%
      4      1957 ns    1897 ns     -3.1%
      8      2204 ns    2103 ns     -4.6%
     12      2369 ns    2272 ns     -4.1%
     16      2596 ns    2363 ns     -9.0%
     24      3083 ns    2677 ns    -13.2%
     32      3740 ns    3050 ns    -18.4%

The saving scales with CPU count, toward the many-CPU systems that run
pressure-monitoring agents. (The high -smp points use KVM oversubscription
on a 12-thread host and carry scheduling noise; the trend is the signal.
The all-busy case is within run-to-run noise.)

Suggested-by: Johannes Weiner <[email protected]>
Signed-off-by: Zhan Xusheng <[email protected]>
---
Changes in v6:
 - Add the benchmark data requested on v4: the -smp scaling A/B table and
   the psi_rtpoll_work() call-frequency measurement. No code change vs v5.
v5: reword changelog (consistency rationale; "No functional change")
    https://lore.kernel.org/all/[email protected]/
v2: https://lore.kernel.org/all/[email protected]/

 kernel/sched/psi.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/kernel/sched/psi.c b/kernel/sched/psi.c
index e2e825dcd088..e2805d32743c 100644
--- a/kernel/sched/psi.c
+++ b/kernel/sched/psi.c
@@ -386,6 +386,9 @@ static void collect_percpu_times(struct psi_group *group,
 				&cpu_changed_states);
 		changed_states |= cpu_changed_states;
 
+		if (!(cpu_changed_states & (1 << PSI_NONIDLE)))
+			continue;
+
 		nonidle = nsecs_to_jiffies(times[PSI_NONIDLE]);
 		nonidle_total += nonidle;
 
-- 
2.43.0