[Patch v4 16/22] sched/cache: Disable cache aware scheduling for processes with high thread counts

[Tim Chen <tim.c.chen@linux.intel.com>]

From: Chen Yu <yu.c.chen@intel.com>

A performance regression was observed by Prateek when running hackbench
with many threads per process (high fd count). To avoid this, processes
with a large number of active threads are excluded from cache-aware
scheduling.

With sched_cache enabled, record the number of active threads in each
process during the periodic task_cache_work(). While iterating over
CPUs, if the currently running task belongs to the same process as the
task that launched task_cache_work(), increment the active thread count.

If the number of active threads within the process exceeds the number
of Cores(divided by SMTs number) in the LLC, do not enable cache-aware
scheduling. However, on system with smaller number of CPUs within 1 LLC,
like Power10/Power11 with SMT4 and LLC size of 4, this check effectively
disables cache-aware scheduling for any process. One possible solution
suggested by Peter is to use a LLC-mask instead of a single LLC value
for preference. Once there are a 'few' LLCs as preference, this constraint
becomes a little easier. It could be an enhancement in the future.

For users who wish to perform task aggregation regardless, a debugfs knob
is provided for tuning in a subsequent patch.

Suggested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Suggested-by: Aaron Lu <ziqianlu@bytedance.com>
Signed-off-by: Chen Yu <yu.c.chen@intel.com>
Co-developed-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com>
---

Notes:
    v3->v4:
       Use cpu_smt_num_threads instead of cpumask_weight(cpu_smt_mask(cpu))
       (Peter Zijlstra)

 include/linux/sched.h |  1 +
 kernel/sched/fair.c   | 54 +++++++++++++++++++++++++++++++++++++++----
 2 files changed, 51 insertions(+), 4 deletions(-)

diff --git a/include/linux/sched.h b/include/linux/sched.h
index 526108acc483..dfa4bfd099c6 100644
--- a/include/linux/sched.h
+++ b/include/linux/sched.h
@@ -2392,6 +2392,7 @@ struct sched_cache_stat {
 	struct sched_cache_time __percpu *pcpu_sched;
 	raw_spinlock_t lock;
 	unsigned long epoch;
+	u64 nr_running_avg;
 	int cpu;
 } ____cacheline_aligned_in_smp;
 
diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 9541e94370e7..077ae7875e2e 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -1316,6 +1316,12 @@ static inline bool valid_llc_buf(struct sched_domain *sd,
 	return true;
 }
 
+static bool exceed_llc_nr(struct mm_struct *mm, int cpu)
+{
+	return !fits_capacity((mm->sc_stat.nr_running_avg * cpu_smt_num_threads),
+			per_cpu(sd_llc_size, cpu));
+}
+
 static void account_llc_enqueue(struct rq *rq, struct task_struct *p)
 {
 	struct sched_domain *sd;
@@ -1507,7 +1513,8 @@ void account_mm_sched(struct rq *rq, struct task_struct *p, s64 delta_exec)
 	 */
 	if (time_after(epoch,
 		       READ_ONCE(mm->sc_stat.epoch) + EPOCH_LLC_AFFINITY_TIMEOUT) ||
-	    get_nr_threads(p) <= 1) {
+	    get_nr_threads(p) <= 1 ||
+	    exceed_llc_nr(mm, cpu_of(rq))) {
 		if (mm->sc_stat.cpu != -1)
 			mm->sc_stat.cpu = -1;
 	}
@@ -1592,13 +1599,31 @@ static void get_scan_cpumasks(cpumask_var_t cpus, struct task_struct *p)
 	cpumask_copy(cpus, cpu_online_mask);
 }
 
+static inline void update_avg_scale(u64 *avg, u64 sample)
+{
+	int factor = per_cpu(sd_llc_size, raw_smp_processor_id());
+	s64 diff = sample - *avg;
+	u32 divisor;
+
+	/*
+	 * Scale the divisor based on the number of CPUs contained
+	 * in the LLC. This scaling ensures smaller LLC domains use
+	 * a smaller divisor to achieve more precise sensitivity to
+	 * changes in nr_running, while larger LLC domains are capped
+	 * at a maximum divisor of 8 which is the default smoothing
+	 * factor of EWMA in update_avg().
+	 */
+	divisor = clamp_t(u32, (factor >> 2), 2, 8);
+	*avg += div64_s64(diff, divisor);
+}
+
 static void task_cache_work(struct callback_head *work)
 {
-	struct task_struct *p = current;
+	struct task_struct *p = current, *cur;
+	int cpu, m_a_cpu = -1, nr_running = 0;
+	unsigned long curr_m_a_occ = 0;
 	struct mm_struct *mm = p->mm;
 	unsigned long m_a_occ = 0;
-	unsigned long curr_m_a_occ = 0;
-	int cpu, m_a_cpu = -1;
 	cpumask_var_t cpus;
 
 	WARN_ON_ONCE(work != &p->cache_work);
@@ -1608,6 +1633,13 @@ static void task_cache_work(struct callback_head *work)
 	if (p->flags & PF_EXITING)
 		return;
 
+	if (get_nr_threads(p) <= 1) {
+		if (mm->sc_stat.cpu != -1)
+			mm->sc_stat.cpu = -1;
+
+		return;
+	}
+
 	if (!zalloc_cpumask_var(&cpus, GFP_KERNEL))
 		return;
 
@@ -1631,6 +1663,12 @@ static void task_cache_work(struct callback_head *work)
 					m_occ = occ;
 					m_cpu = i;
 				}
+				scoped_guard (rcu) {
+					cur = rcu_dereference_all(cpu_rq(i)->curr);
+					if (cur && !(cur->flags & (PF_EXITING | PF_KTHREAD)) &&
+					    cur->mm == mm)
+						nr_running++;
+				}
 			}
 
 			/*
@@ -1674,6 +1712,7 @@ static void task_cache_work(struct callback_head *work)
 		mm->sc_stat.cpu = m_a_cpu;
 	}
 
+	update_avg_scale(&mm->sc_stat.nr_running_avg, nr_running);
 	free_cpumask_var(cpus);
 }
 
@@ -10105,6 +10144,13 @@ static enum llc_mig can_migrate_llc_task(int src_cpu, int dst_cpu,
 	if (cpu < 0 || cpus_share_cache(src_cpu, dst_cpu))
 		return mig_unrestricted;
 
+	/* skip cache aware load balance for single/too many threads */
+	if (get_nr_threads(p) <= 1 || exceed_llc_nr(mm, dst_cpu)) {
+		if (mm->sc_stat.cpu != -1)
+			mm->sc_stat.cpu = -1;
+		return mig_unrestricted;
+	}
+
 	if (cpus_share_cache(dst_cpu, cpu))
 		to_pref = true;
 	else if (cpus_share_cache(src_cpu, cpu))
-- 
2.32.0