sched/idle: introduce smart halt polling

hulk inclusion
category: feature
bugzilla: 176961 https://gitee.com/openeuler/kernel/issues/I4E05T

-------------------------------------------------

In guest, Before entering the real idle, polling for a while. if
the current task is set TIF_NEED_RESCHED during the polling process,
it will immediately break from the polling loop.

The polling time poll_threshold_ns can be adjusted by sysctl to avoid
100% of the CPU usage in host. This value can be adjusted according to
the requirements.

This optimization needs to support _TIF_POLLING_NRFLAG, that can
optimize the overhead of ttwu IPI. Wakeup response delay reduced
from 4us to 1us.
Signed-off-by: NXiangyou Xie <xiexiangyou@huawei.com>
Signed-off-by: NLi Hua <hucool.lihua@huawei.com>
Reviewed-by: NChen Hui <judy.chenhui@huawei.com>
Signed-off-by: NChen Jun <chenjun102@huawei.com>
Signed-off-by: NZheng Zengkai <zhengzengkai@huawei.com>

include/linux/kernel.h

@@ -555,6 +555,8 @@ extern int sysctl_panic_on_rcu_stall;
 extern int sysctl_panic_on_stackoverflow;
 
 extern bool crash_kexec_post_notifiers;
+extern unsigned long poll_threshold_ns;
+
 
 /*
  * panic_cpu is used for synchronizing panic() and crash_kexec() execution. It

kernel/sched/idle.c

@@ -13,6 +13,12 @@
 /* Linker adds these: start and end of __cpuidle functions */
 extern char __cpuidle_text_start[], __cpuidle_text_end[];
 
+/*
+ * Poll_threshold_ns indicates the maximum polling time before
+ * entering real idle.
+ */
+unsigned long poll_threshold_ns;
+
 /**
  * sched_idle_set_state - Record idle state for the current CPU.
  * @idle_state: State to record.
@@ -52,12 +58,31 @@ static int __init cpu_idle_nopoll_setup(char *__unused)
 __setup("hlt", cpu_idle_nopoll_setup);
 #endif
 
+static void smart_idle_poll(void)
+{
+	unsigned long poll_duration = poll_threshold_ns;
+	ktime_t cur, stop;
+
+	if (!poll_duration)
+		return;
+
+	stop = ktime_add_ns(ktime_get(), poll_duration);
+
+	do {
+		cpu_relax();
+		if (tif_need_resched())
+			break;
+		cur = ktime_get();
+	} while (ktime_before(cur, stop));
+}
+
 static noinline int __cpuidle cpu_idle_poll(void)
 {
 	trace_cpu_idle(0, smp_processor_id());
 	stop_critical_timings();
 	rcu_idle_enter();
 	local_irq_enable();
+	smart_idle_poll();
 
 	while (!tif_need_resched() &&
 	       (cpu_idle_force_poll || tick_check_broadcast_expired()))
@@ -261,6 +286,7 @@ static void cpuidle_idle_call(void)
 static void do_idle(void)
 {
 	int cpu = smp_processor_id();
+	unsigned long idle_poll_flag = poll_threshold_ns;
 	/*
 	 * If the arch has a polling bit, we maintain an invariant:
 	 *
@@ -293,9 +319,11 @@ static void do_idle(void)
 		 * broadcast device expired for us, we don't want to go deep
 		 * idle as we know that the IPI is going to arrive right away.
 		 */
-		if (cpu_idle_force_poll || tick_check_broadcast_expired()) {
+		if (cpu_idle_force_poll || tick_check_broadcast_expired() ||
+			  idle_poll_flag) {
 			tick_nohz_idle_restart_tick();
 			cpu_idle_poll();
+			idle_poll_flag = 0;
 		} else {
 			cpuidle_idle_call();
 		}

kernel/sysctl.c

@@ -1849,6 +1849,13 @@ static struct ctl_table kern_table[] = {
 		.proc_handler	= sysctl_sched_uclamp_handler,
 	},
 #endif
+	{
+		.procname	= "halt_poll_threshold",
+		.data		= &poll_threshold_ns,
+		.maxlen		= sizeof(unsigned long),
+		.mode		= 0644,
+		.proc_handler	= proc_doulongvec_minmax,
+	},
 #ifdef CONFIG_SCHED_AUTOGROUP
 	{
 		.procname	= "sched_autogroup_enabled",