rcu: Place guard on rcu_all_qs() and rcu_note_context_switch() actions

The rcu_all_qs() and rcu_note_context_switch() do a series of checks,
taking various actions to supply RCU with quiescent states, depending
on the outcomes of the various checks.  This is a bit much for scheduling
fastpaths, so this commit creates a separate ->rcu_urgent_qs field in
the rcu_dynticks structure that acts as a global guard for these checks.
Thus, in the common case, rcu_all_qs() and rcu_note_context_switch()
check the ->rcu_urgent_qs field, find it false, and simply return.

Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>

5 files changed, 44 insertions, 18 deletions

diff --git a/Documentation/RCU/Design/Data-Structures/Data-Structures.html b/Documentation/RCU/Design/Data-Structures/Data-Structures.html
index e4bf20a68fa3..4dec89097559 100644
--- a/Documentation/RCU/Design/Data-Structures/Data-Structures.html
+++ b/Documentation/RCU/Design/Data-Structures/Data-Structures.html
@@ -1106,6 +1106,7 @@ Its fields are as follows:
   3   atomic_t dynticks;
   4   bool rcu_need_heavy_qs;
   5   unsigned long rcu_qs_ctr;
+  6   bool rcu_urgent_qs;
 </pre>
 
 <p>The <tt>-&gt;dynticks_nesting</tt> field counts the
@@ -1131,12 +1132,20 @@ it is willing to call for heavy-weight dyntick-counter operations.
 This flag is checked by RCU's context-switch and <tt>cond_resched()</tt>
 code, which provide a momentary idle sojourn in response.
 
-</p><p>Finally the <tt>-&gt;rcu_qs_ctr</tt> field is used to record
+</p><p>The <tt>-&gt;rcu_qs_ctr</tt> field is used to record
 quiescent states from <tt>cond_resched()</tt>.
 Because <tt>cond_resched()</tt> can execute quite frequently, this
 must be quite lightweight, as in a non-atomic increment of this
 per-CPU field.
 
+</p><p>Finally, the <tt>-&gt;rcu_urgent_qs</tt> field is used to record
+the fact that the RCU core code would really like to see a quiescent
+state from the corresponding CPU, with the various other fields indicating
+just how badly RCU wants this quiescent state.
+This flag is checked by RCU's context-switch and <tt>cond_resched()</tt>
+code, which, if nothing else, non-atomically increment <tt>-&gt;rcu_qs_ctr</tt>
+in response.
+
 <table>
 <tr><th>&nbsp;</th></tr>
 <tr><th align="left">Quick Quiz:</th></tr>
diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
index c2cbc78a0625..530ab6cf7a0b 100644
--- a/kernel/rcu/tree.c
+++ b/kernel/rcu/tree.c
@@ -464,8 +464,14 @@ void rcu_note_context_switch(void)
 	trace_rcu_utilization(TPS("Start context switch"));
 	rcu_sched_qs();
 	rcu_preempt_note_context_switch();
+	/* Load rcu_urgent_qs before other flags. */
+	if (!smp_load_acquire(this_cpu_ptr(&rcu_dynticks.rcu_urgent_qs)))
+		goto out;
+	this_cpu_write(rcu_dynticks.rcu_urgent_qs, false);
 	if (unlikely(raw_cpu_read(rcu_dynticks.rcu_need_heavy_qs)))
 		rcu_momentary_dyntick_idle();
+	this_cpu_inc(rcu_dynticks.rcu_qs_ctr);
+out:
 	trace_rcu_utilization(TPS("End context switch"));
 	barrier(); /* Avoid RCU read-side critical sections leaking up. */
 }
@@ -488,29 +494,26 @@ void rcu_all_qs(void)
 {
 	unsigned long flags;
 
+	if (!raw_cpu_read(rcu_dynticks.rcu_urgent_qs))
+		return;
+	preempt_disable();
+	/* Load rcu_urgent_qs before other flags. */
+	if (!smp_load_acquire(this_cpu_ptr(&rcu_dynticks.rcu_urgent_qs))) {
+		preempt_enable();
+		return;
+	}
+	this_cpu_write(rcu_dynticks.rcu_urgent_qs, false);
 	barrier(); /* Avoid RCU read-side critical sections leaking down. */
 	if (unlikely(raw_cpu_read(rcu_dynticks.rcu_need_heavy_qs))) {
 		local_irq_save(flags);
 		rcu_momentary_dyntick_idle();
 		local_irq_restore(flags);
 	}
-	if (unlikely(raw_cpu_read(rcu_sched_data.cpu_no_qs.b.exp))) {
-		/*
-		 * Yes, we just checked a per-CPU variable with preemption
-		 * enabled, so we might be migrated to some other CPU at
-		 * this point.  That is OK because in that case, the
-		 * migration will supply the needed quiescent state.
-		 * We might end up needlessly disabling preemption and
-		 * invoking rcu_sched_qs() on the destination CPU, but
-		 * the probability and cost are both quite low, so this
-		 * should not be a problem in practice.
-		 */
-		preempt_disable();
+	if (unlikely(raw_cpu_read(rcu_sched_data.cpu_no_qs.b.exp)))
 		rcu_sched_qs();
-		preempt_enable();
-	}
 	this_cpu_inc(rcu_dynticks.rcu_qs_ctr);
 	barrier(); /* Avoid RCU read-side critical sections leaking up. */
+	preempt_enable();
 }
 EXPORT_SYMBOL_GPL(rcu_all_qs);
 
@@ -1246,6 +1249,7 @@ static int rcu_implicit_dynticks_qs(struct rcu_data *rdp,
 {
 	unsigned long jtsq;
 	bool *rnhqp;
+	bool *ruqp;
 	unsigned long rjtsc;
 	struct rcu_node *rnp;
 
@@ -1281,11 +1285,15 @@ static int rcu_implicit_dynticks_qs(struct rcu_data *rdp,
 	 * might not be the case for nohz_full CPUs looping in the kernel.
 	 */
 	rnp = rdp->mynode;
+	ruqp = per_cpu_ptr(&rcu_dynticks.rcu_urgent_qs, rdp->cpu);
 	if (time_after(jiffies, rdp->rsp->gp_start + jtsq) &&
 	    READ_ONCE(rdp->rcu_qs_ctr_snap) != per_cpu(rcu_dynticks.rcu_qs_ctr, rdp->cpu) &&
 	    READ_ONCE(rdp->gpnum) == rnp->gpnum && !rdp->gpwrap) {
 		trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("rqc"));
 		return 1;
+	} else {
+		/* Load rcu_qs_ctr before store to rcu_urgent_qs. */
+		smp_store_release(ruqp, true);
 	}
 
 	/* Check for the CPU being offline. */
@@ -1321,6 +1329,8 @@ static int rcu_implicit_dynticks_qs(struct rcu_data *rdp,
 	    (time_after(jiffies, rdp->rsp->gp_start + jtsq) ||
 	     time_after(jiffies, rdp->rsp->jiffies_resched))) {
 		WRITE_ONCE(*rnhqp, true);
+		/* Store rcu_need_heavy_qs before rcu_urgent_qs. */
+		smp_store_release(ruqp, true);
 		rdp->rsp->jiffies_resched += 5; /* Re-enable beating. */
 	}
 
diff --git a/kernel/rcu/tree.h b/kernel/rcu/tree.h
index b212cd0f22c7..d2f276fc2edc 100644
--- a/kernel/rcu/tree.h
+++ b/kernel/rcu/tree.h
@@ -113,8 +113,9 @@ struct rcu_dynticks {
 				    /* Process level is worth LLONG_MAX/2. */
 	int dynticks_nmi_nesting;   /* Track NMI nesting level. */
 	atomic_t dynticks;	    /* Even value for idle, else odd. */
-	bool rcu_need_heavy_qs;      /* GP old, need heavy quiescent state. */
+	bool rcu_need_heavy_qs;     /* GP old, need heavy quiescent state. */
 	unsigned long rcu_qs_ctr;   /* Light universal quiescent state ctr. */
+	bool rcu_urgent_qs;	    /* GP old need light quiescent state. */
 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
 	long long dynticks_idle_nesting;
 				    /* irq/process nesting level from idle. */
diff --git a/kernel/rcu/tree_exp.h b/kernel/rcu/tree_exp.h
index a7b639ccd46e..a1f52bbe9db6 100644
--- a/kernel/rcu/tree_exp.h
+++ b/kernel/rcu/tree_exp.h
@@ -331,6 +331,8 @@ static void sync_sched_exp_handler(void *data)
 		return;
 	}
 	__this_cpu_write(rcu_sched_data.cpu_no_qs.b.exp, true);
+	/* Store .exp before .rcu_urgent_qs. */
+	smp_store_release(this_cpu_ptr(&rcu_dynticks.rcu_urgent_qs), true);
 	resched_cpu(smp_processor_id());
 }
 
diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h
index 0a62a8f1caac..621296a6694b 100644
--- a/kernel/rcu/tree_plugin.h
+++ b/kernel/rcu/tree_plugin.h
@@ -1860,7 +1860,9 @@ static void __call_rcu_nocb_enqueue(struct rcu_data *rdp,
 			trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
 					    TPS("WakeEmpty"));
 		} else {
-			rdp->nocb_defer_wakeup = RCU_NOGP_WAKE;
+			WRITE_ONCE(rdp->nocb_defer_wakeup, RCU_NOGP_WAKE);
+			/* Store ->nocb_defer_wakeup before ->rcu_urgent_qs. */
+			smp_store_release(this_cpu_ptr(&rcu_dynticks.rcu_urgent_qs), true);
 			trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
 					    TPS("WakeEmptyIsDeferred"));
 		}
@@ -1872,7 +1874,9 @@ static void __call_rcu_nocb_enqueue(struct rcu_data *rdp,
 			trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
 					    TPS("WakeOvf"));
 		} else {
-			rdp->nocb_defer_wakeup = RCU_NOGP_WAKE_FORCE;
+			WRITE_ONCE(rdp->nocb_defer_wakeup, RCU_NOGP_WAKE_FORCE);
+			/* Store ->nocb_defer_wakeup before ->rcu_urgent_qs. */
+			smp_store_release(this_cpu_ptr(&rcu_dynticks.rcu_urgent_qs), true);
 			trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
 					    TPS("WakeOvfIsDeferred"));
 		}