rcu: Check cond_resched_rcu_qs() state less often to reduce GP overhead

Commit 4a81e832 ("rcu: Reduce overhead of cond_resched() checks
for RCU") moved quiescent-state generation out of cond_resched()
and commit bde6c3aa ("rcu: Provide cond_resched_rcu_qs() to force
quiescent states in long loops") introduced cond_resched_rcu_qs(), and
commit 5cd37193 ("rcu: Make cond_resched_rcu_qs() apply to normal RCU
flavors") introduced the per-CPU rcu_qs_ctr variable, which is frequently
polled by the RCU core state machine.

This frequent polling can increase grace-period rate, which in turn
increases grace-period overhead, which is visible in some benchmarks
(for example, the "open1" benchmark in Anton Blanchard's "will it scale"
suite).  This commit therefore reduces the rate at which rcu_qs_ctr
is polled by moving that polling into the force-quiescent-state (FQS)
machinery, and by further polling it only after the grace period has
been in effect for at least jiffies_till_sched_qs jiffies.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Reviewed-by: Josh Triplett <josh@joshtriplett.org>

include/trace/events/rcu.h

@@ -385,11 +385,11 @@ TRACE_EVENT(rcu_quiescent_state_report,
 
 /*
  * Tracepoint for quiescent states detected by force_quiescent_state().
- * These trace events include the type of RCU, the grace-period number
- * that was blocked by the CPU, the CPU itself, and the type of quiescent
- * state, which can be "dti" for dyntick-idle mode, "ofl" for CPU offline,
- * or "kick" when kicking a CPU that has been in dyntick-idle mode for
- * too long.
+ * These trace events include the type of RCU, the grace-period number that
+ * was blocked by the CPU, the CPU itself, and the type of quiescent state,
+ * which can be "dti" for dyntick-idle mode, "ofl" for CPU offline, "kick"
+ * when kicking a CPU that has been in dyntick-idle mode for too long, or
+ * "rqc" if the CPU got a quiescent state via its rcu_qs_ctr.
  */
 TRACE_EVENT(rcu_fqs,
 

kernel/rcu/tree.c

@@ -1232,7 +1232,10 @@ static int dyntick_save_progress_counter(struct rcu_data *rdp,
 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp,
 				    bool *isidle, unsigned long *maxj)
 {
+	unsigned long jtsq;
 	int *rcrmp;
+	unsigned long rjtsc;
+	struct rcu_node *rnp;
 
 	/*
 	 * If the CPU passed through or entered a dynticks idle phase with
@@ -1248,6 +1251,31 @@ static int rcu_implicit_dynticks_qs(struct rcu_data *rdp,
 		return 1;
 	}
 
+	/* Compute and saturate jiffies_till_sched_qs. */
+	jtsq = jiffies_till_sched_qs;
+	rjtsc = rcu_jiffies_till_stall_check();
+	if (jtsq > rjtsc / 2) {
+		WRITE_ONCE(jiffies_till_sched_qs, rjtsc);
+		jtsq = rjtsc / 2;
+	} else if (jtsq < 1) {
+		WRITE_ONCE(jiffies_till_sched_qs, 1);
+		jtsq = 1;
+	}
+
+	/*
+	 * Has this CPU encountered a cond_resched_rcu_qs() since the
+	 * beginning of the grace period?  For this to be the case,
+	 * the CPU has to have noticed the current grace period.  This
+	 * might not be the case for nohz_full CPUs looping in the kernel.
+	 */
+	rnp = rdp->mynode;
+	if (time_after(jiffies, rdp->rsp->gp_start + jtsq) &&
+	    READ_ONCE(rdp->rcu_qs_ctr_snap) != per_cpu(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;
+	}
+
 	/*
 	 * Check for the CPU being offline, but only if the grace period
 	 * is old enough.  We don't need to worry about the CPU changing
@@ -1290,9 +1318,8 @@ static int rcu_implicit_dynticks_qs(struct rcu_data *rdp,
 	 * warning delay.
 	 */
 	rcrmp = &per_cpu(rcu_sched_qs_mask, rdp->cpu);
-	if (ULONG_CMP_GE(jiffies,
-			 rdp->rsp->gp_start + jiffies_till_sched_qs) ||
-	    ULONG_CMP_GE(jiffies, rdp->rsp->jiffies_resched)) {
+	if (time_after(jiffies, rdp->rsp->gp_start + jtsq) ||
+	    time_after(jiffies, rdp->rsp->jiffies_resched)) {
 		if (!(READ_ONCE(*rcrmp) & rdp->rsp->flavor_mask)) {
 			WRITE_ONCE(rdp->cond_resched_completed,
 				   READ_ONCE(rdp->mynode->completed));
@@ -2550,10 +2577,8 @@ rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
 
 	rnp = rdp->mynode;
 	raw_spin_lock_irqsave_rcu_node(rnp, flags);
-	if ((rdp->cpu_no_qs.b.norm &&
-	     rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr)) ||
-	    rdp->gpnum != rnp->gpnum || rnp->completed == rnp->gpnum ||
-	    rdp->gpwrap) {
+	if (rdp->cpu_no_qs.b.norm || rdp->gpnum != rnp->gpnum ||
+	    rnp->completed == rnp->gpnum || rdp->gpwrap) {
 
 		/*
 		 * The grace period in which this quiescent state was
@@ -2608,8 +2633,7 @@ rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
 	 * Was there a quiescent state since the beginning of the grace
 	 * period? If no, then exit and wait for the next call.
 	 */
-	if (rdp->cpu_no_qs.b.norm &&
-	    rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr))
+	if (rdp->cpu_no_qs.b.norm)
 		return;
 
 	/*
@@ -3563,9 +3587,7 @@ static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
 	    rdp->core_needs_qs && rdp->cpu_no_qs.b.norm &&
 	    rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr)) {
 		rdp->n_rp_core_needs_qs++;
-	} else if (rdp->core_needs_qs &&
-		   (!rdp->cpu_no_qs.b.norm ||
-		    rdp->rcu_qs_ctr_snap != __this_cpu_read(rcu_qs_ctr))) {
+	} else if (rdp->core_needs_qs && !rdp->cpu_no_qs.b.norm) {
 		rdp->n_rp_report_qs++;
 		return 1;
 	}