rcu: Use softirq to address performance regression

Commit a26ac2455ffcf3(rcu: move TREE_RCU from softirq to kthread)
introduced performance regression. In an AIM7 test, this commit degraded
performance by about 40%.

The commit runs rcu callbacks in a kthread instead of softirq. We observed
high rate of context switch which is caused by this. Out test system has
64 CPUs and HZ is 1000, so we saw more than 64k context switch per second
which is caused by RCU's per-CPU kthread.  A trace showed that most of
the time the RCU per-CPU kthread doesn't actually handle any callbacks,
but instead just does a very small amount of work handling grace periods.
This means that RCU's per-CPU kthreads are making the scheduler do quite
a bit of work in order to allow a very small amount of RCU-related
processing to be done.

Alex Shi's analysis determined that this slowdown is due to lock
contention within the scheduler.  Unfortunately, as Peter Zijlstra points
out, the scheduler's real-time semantics require global action, which
means that this contention is inherent in real-time scheduling.  (Yes,
perhaps someone will come up with a workaround -- otherwise, -rt is not
going to do well on large SMP systems -- but this patch will work around
this issue in the meantime.  And "the meantime" might well be forever.)

This patch therefore re-introduces softirq processing to RCU, but only
for core RCU work.  RCU callbacks are still executed in kthread context,
so that only a small amount of RCU work runs in softirq context in the
common case.  This should minimize ksoftirqd execution, allowing us to
skip boosting of ksoftirqd for CONFIG_RCU_BOOST=y kernels.

Signed-off-by: Shaohua Li <shaohua.li@intel.com>
Tested-by: "Alex,Shi" <alex.shi@intel.com>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
diff --git a/Documentation/filesystems/proc.txt b/Documentation/filesystems/proc.txt
index f481780..db3b1ab 100644
--- a/Documentation/filesystems/proc.txt
+++ b/Documentation/filesystems/proc.txt
@@ -843,6 +843,7 @@
  TASKLET:          0          0          0        290
    SCHED:      27035      26983      26971      26746
  HRTIMER:          0          0          0          0
+     RCU:       1678       1769       2178       2250
 
 
 1.3 IDE devices in /proc/ide
diff --git a/include/linux/interrupt.h b/include/linux/interrupt.h
index 6c12989..f6efed0 100644
--- a/include/linux/interrupt.h
+++ b/include/linux/interrupt.h
@@ -414,6 +414,7 @@
 	TASKLET_SOFTIRQ,
 	SCHED_SOFTIRQ,
 	HRTIMER_SOFTIRQ,
+	RCU_SOFTIRQ,    /* Preferable RCU should always be the last softirq */
 
 	NR_SOFTIRQS
 };
diff --git a/include/trace/events/irq.h b/include/trace/events/irq.h
index ae045ca..1c09820 100644
--- a/include/trace/events/irq.h
+++ b/include/trace/events/irq.h
@@ -20,7 +20,8 @@
 			 softirq_name(BLOCK_IOPOLL),	\
 			 softirq_name(TASKLET),		\
 			 softirq_name(SCHED),		\
-			 softirq_name(HRTIMER))
+			 softirq_name(HRTIMER),		\
+			 softirq_name(RCU))
 
 /**
  * irq_handler_entry - called immediately before the irq action handler
diff --git a/kernel/rcutree.c b/kernel/rcutree.c
index 0a8ec5b..ae5c9ea 100644
--- a/kernel/rcutree.c
+++ b/kernel/rcutree.c
@@ -100,6 +100,7 @@
 
 static void rcu_node_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
 static void invoke_rcu_cpu_kthread(void);
+static void __invoke_rcu_cpu_kthread(void);
 
 #define RCU_KTHREAD_PRIO 1	/* RT priority for per-CPU kthreads. */
 
@@ -1442,13 +1443,21 @@
 	}
 
 	/* If there are callbacks ready, invoke them. */
-	rcu_do_batch(rsp, rdp);
+	if (cpu_has_callbacks_ready_to_invoke(rdp))
+		__invoke_rcu_cpu_kthread();
+}
+
+static void rcu_kthread_do_work(void)
+{
+	rcu_do_batch(&rcu_sched_state, &__get_cpu_var(rcu_sched_data));
+	rcu_do_batch(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
+	rcu_preempt_do_callbacks();
 }
 
 /*
  * Do softirq processing for the current CPU.
  */
-static void rcu_process_callbacks(void)
+static void rcu_process_callbacks(struct softirq_action *unused)
 {
 	__rcu_process_callbacks(&rcu_sched_state,
 				&__get_cpu_var(rcu_sched_data));
@@ -1465,7 +1474,7 @@
  * the current CPU with interrupts disabled, the rcu_cpu_kthread_task
  * cannot disappear out from under us.
  */
-static void invoke_rcu_cpu_kthread(void)
+static void __invoke_rcu_cpu_kthread(void)
 {
 	unsigned long flags;
 
@@ -1479,6 +1488,11 @@
 	local_irq_restore(flags);
 }
 
+static void invoke_rcu_cpu_kthread(void)
+{
+	raise_softirq(RCU_SOFTIRQ);
+}
+
 /*
  * Wake up the specified per-rcu_node-structure kthread.
  * Because the per-rcu_node kthreads are immortal, we don't need
@@ -1613,7 +1627,7 @@
 		*workp = 0;
 		local_irq_restore(flags);
 		if (work)
-			rcu_process_callbacks();
+			rcu_kthread_do_work();
 		local_bh_enable();
 		if (*workp != 0)
 			spincnt++;
@@ -2387,6 +2401,7 @@
 	rcu_init_one(&rcu_sched_state, &rcu_sched_data);
 	rcu_init_one(&rcu_bh_state, &rcu_bh_data);
 	__rcu_init_preempt();
+	 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
 
 	/*
 	 * We don't need protection against CPU-hotplug here because
diff --git a/kernel/rcutree.h b/kernel/rcutree.h
index 7b9a08b..0fed6b9 100644
--- a/kernel/rcutree.h
+++ b/kernel/rcutree.h
@@ -439,6 +439,7 @@
 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
 static void rcu_preempt_check_callbacks(int cpu);
 static void rcu_preempt_process_callbacks(void);
+static void rcu_preempt_do_callbacks(void);
 void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu));
 #if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_TREE_PREEMPT_RCU)
 static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp);
diff --git a/kernel/rcutree_plugin.h b/kernel/rcutree_plugin.h
index ea2e2fb..38d09c5 100644
--- a/kernel/rcutree_plugin.h
+++ b/kernel/rcutree_plugin.h
@@ -602,6 +602,11 @@
 				&__get_cpu_var(rcu_preempt_data));
 }
 
+static void rcu_preempt_do_callbacks(void)
+{
+	rcu_do_batch(&rcu_preempt_state, &__get_cpu_var(rcu_preempt_data));
+}
+
 /*
  * Queue a preemptible-RCU callback for invocation after a grace period.
  */
@@ -997,6 +1002,10 @@
 {
 }
 
+static void rcu_preempt_do_callbacks(void)
+{
+}
+
 /*
  * Wait for an rcu-preempt grace period, but make it happen quickly.
  * But because preemptible RCU does not exist, map to rcu-sched.
diff --git a/kernel/softirq.c b/kernel/softirq.c
index 1396017..40cf63d 100644
--- a/kernel/softirq.c
+++ b/kernel/softirq.c
@@ -58,7 +58,7 @@
 
 char *softirq_to_name[NR_SOFTIRQS] = {
 	"HI", "TIMER", "NET_TX", "NET_RX", "BLOCK", "BLOCK_IOPOLL",
-	"TASKLET", "SCHED", "HRTIMER"
+	"TASKLET", "SCHED", "HRTIMER", "RCU"
 };
 
 /*
diff --git a/tools/perf/util/trace-event-parse.c b/tools/perf/util/trace-event-parse.c
index 1e88485..0a7ed5b 100644
--- a/tools/perf/util/trace-event-parse.c
+++ b/tools/perf/util/trace-event-parse.c
@@ -2187,6 +2187,7 @@
 	{ "TASKLET_SOFTIRQ", 6 },
 	{ "SCHED_SOFTIRQ", 7 },
 	{ "HRTIMER_SOFTIRQ", 8 },
+	{ "RCU_SOFTIRQ", 9 },
 
 	{ "HRTIMER_NORESTART", 0 },
 	{ "HRTIMER_RESTART", 1 },