[PATCH v3 20/24] sched/fair: Introduce sched_smt_siblings_idle()

[Ricardo Neri]

X86 needs to know the idle state of the SMT siblings of a CPU to improve
the accuracy of IPCC classification. X86 implements support for IPC classes
in the thermal HFI driver.

Rename is_core_idle() as sched_smt_siblings_idle() and make it available
outside the scheduler code.

Cc: Ben Segall <bsegall@xxxxxxxxxx>
Cc: Daniel Bristot de Oliveira <bristot@xxxxxxxxxx>
Cc: Dietmar Eggemann <dietmar.eggemann@xxxxxxx>
Cc: Len Brown <len.brown@xxxxxxxxx>
Cc: Mel Gorman <mgorman@xxxxxxx>
Cc: Rafael J. Wysocki <rafael.j.wysocki@xxxxxxxxx>
Cc: Srinivas Pandruvada <srinivas.pandruvada@xxxxxxxxxxxxxxx>
Cc: Steven Rostedt <rostedt@xxxxxxxxxxx>
Cc: Tim C. Chen <tim.c.chen@xxxxxxxxx>
Cc: Valentin Schneider <vschneid@xxxxxxxxxx>
Cc: x86@xxxxxxxxxx
Cc: linux-kernel@xxxxxxxxxxxxxxx
Signed-off-by: Ricardo Neri <ricardo.neri-calderon@xxxxxxxxxxxxxxx>
---
is_core_idle() is no longer an inline function after this patch. To rule
out performance degradation, I compared the execution time of the inline
and non-inline versions on a 4-socket Cascade Lake system using the NUMA
stressor of stress-ng:

        $ stress-ng --numa 1500 -t 10m

is_core_idle() was called ~200,000 times. I measured the value of the TSC
counter before and after calling is_core_idle() and computed the delta
value.

I arbitrarily removed outliers (defined as any delta larger than 5000
counts). This required removing ~40 samples.

The table below summarizes the difference in execution time. All quantities
are expressed in TSC counts, except the standard deviation, expressed as a
percentage of the average.

                              Average  Median  Std(%) Mode
        TSCdelta inline        668.76     626   67.24   42
        TSCdelta non-inline    677.64     624   67.67   46

All metrics are similar for the inline and non-inline cases.
---
Changes since v2:
 * Brought back this previously dropped patch.
 * Profiled inline vs non-inline is_core_idle(). I found not major penalty.
 * Merged is_core_idle() and sched_smt_siblings_idle() into a single
   function. (Dietmar)

Changes since v1:
 * Dropped this patch.
---
 include/linux/sched.h |  2 ++
 kernel/sched/fair.c   | 21 +++++++++++++++------
 2 files changed, 17 insertions(+), 6 deletions(-)

diff --git a/include/linux/sched.h b/include/linux/sched.h
index 45f28a601b3d..7ef9fd84e7ad 100644
--- a/include/linux/sched.h
+++ b/include/linux/sched.h
@@ -2449,4 +2449,6 @@ static inline void sched_core_fork(struct task_struct *p) { }
 
 extern void sched_set_stop_task(int cpu, struct task_struct *stop);
 
+extern bool sched_smt_siblings_idle(int cpu);
+
 #endif
diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index d3c22dc145f7..a66d86c5cb5c 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -1064,7 +1064,14 @@ update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
  * Scheduling class queueing methods:
  */
 
-static inline bool is_core_idle(int cpu)
+/**
+ * sched_smt_siblings_idle - Check whether SMT siblings of a CPU are idle
+ * @cpu:	The CPU to check
+ *
+ * Returns true if all the SMT siblings of @cpu are idle or @cpu does not have
+ * SMT siblings. The idle state of @cpu is not considered.
+ */
+bool sched_smt_siblings_idle(int cpu)
 {
 #ifdef CONFIG_SCHED_SMT
 	int sibling;
@@ -1767,7 +1774,7 @@ static inline int numa_idle_core(int idle_core, int cpu)
 	 * Prefer cores instead of packing HT siblings
 	 * and triggering future load balancing.
 	 */
-	if (is_core_idle(cpu))
+	if (sched_smt_siblings_idle(cpu))
 		idle_core = cpu;
 
 	return idle_core;
@@ -9518,7 +9525,8 @@ sched_asym(struct lb_env *env, struct sd_lb_stats *sds,  struct sg_lb_stats *sgs
 	 * If the destination CPU has SMT siblings, env->idle != CPU_NOT_IDLE
 	 * is not sufficient. We need to make sure the whole core is idle.
 	 */
-	if (sds->local->flags & SD_SHARE_CPUCAPACITY && !is_core_idle(env->dst_cpu))
+	if (sds->local->flags & SD_SHARE_CPUCAPACITY &&
+	    !sched_smt_siblings_idle(env->dst_cpu))
 		return false;
 
 	/* Only do SMT checks if either local or candidate have SMT siblings. */
@@ -10687,7 +10695,8 @@ static struct rq *find_busiest_queue(struct lb_env *env,
 		    sched_asym_prefer(i, env->dst_cpu) &&
 		    nr_running == 1) {
 			if (env->sd->flags & SD_SHARE_CPUCAPACITY ||
-			    (!(env->sd->flags & SD_SHARE_CPUCAPACITY) && is_core_idle(i)))
+			    (!(env->sd->flags & SD_SHARE_CPUCAPACITY) &&
+			     sched_smt_siblings_idle(i)))
 				continue;
 		}
 
@@ -10816,7 +10825,7 @@ asym_active_balance(struct lb_env *env)
 		 * busy sibling.
 		 */
 		return sched_asym_prefer(env->dst_cpu, env->src_cpu) ||
-		       !is_core_idle(env->src_cpu);
+		       !sched_smt_siblings_idle(env->src_cpu);
 	}
 
 	return false;
@@ -11563,7 +11572,7 @@ static void nohz_balancer_kick(struct rq *rq)
 				 */
 				if (sd->flags & SD_SHARE_CPUCAPACITY ||
 				    (!(sd->flags & SD_SHARE_CPUCAPACITY) &&
-				     is_core_idle(i))) {
+				     sched_smt_siblings_idle(i))) {
 					flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
 					goto unlock;
 				}
-- 
2.25.1