rcutorture: Don't count CPU-stalled time against priority boosting
authorPaul E. McKenney <paulmck@kernel.org>
Wed, 14 Apr 2021 20:00:10 +0000 (13:00 -0700)
committerPaul E. McKenney <paulmck@kernel.org>
Mon, 10 May 2021 23:05:07 +0000 (16:05 -0700)
It will frequently be the case that rcu_torture_boost() will get a
->start_gp_poll() cookie that needs almost all of the current grace period
plus an additional grace period to elapse before ->poll_gp_state() will
return true.  It is quite possible that the current grace period will have
(say) two seconds of stall by a CPU failing to pass through a quiescent
state, followed by 300 milliseconds of delay due to a preempted reader.
The next grace period might suffer only one second of stall by a CPU,
followed by another 300 milliseconds of delay due to a preempted reader.
This is an example of RCU priority boosting doing its job, but the full
elapsed time of 3.6 seconds exceeds the 3.5-second limit.  In addition,
there is no CPU stall in force at the 3.5-second mark, so this would
nevertheless currently be counted as an RCU priority boosting failure.

This commit therefore avoids this sort of false positive by resetting
the gp_state_time timestamp any time that the current grace period is
being blocked by a CPU.  This results in extremely frequent calls to
the ->check_boost_failed() function, so this commit provides a lockless
fastpath that is selected by supplying a NULL CPU-number pointer.

Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
kernel/rcu/rcutorture.c
kernel/rcu/tree_stall.h

index 5ae4dcc..8b347b9 100644 (file)
@@ -918,17 +918,18 @@ static void rcu_torture_enable_rt_throttle(void)
        old_rt_runtime = -1;
 }
 
-static bool rcu_torture_boost_failed(unsigned long gp_state, unsigned long start, unsigned long end)
+static bool rcu_torture_boost_failed(unsigned long gp_state, unsigned long *start)
 {
        int cpu;
        static int dbg_done;
+       unsigned long end = jiffies;
        bool gp_done;
        unsigned long j;
        static unsigned long last_persist;
        unsigned long lp;
        unsigned long mininterval = test_boost_duration * HZ - HZ / 2;
 
-       if (end - start > mininterval) {
+       if (end - *start > mininterval) {
                // Recheck after checking time to avoid false positives.
                smp_mb(); // Time check before grace-period check.
                if (cur_ops->poll_gp_state(gp_state))
@@ -945,7 +946,7 @@ static bool rcu_torture_boost_failed(unsigned long gp_state, unsigned long start
                n_rcu_torture_boost_failure++;
                if (!xchg(&dbg_done, 1) && cur_ops->gp_kthread_dbg) {
                        pr_info("Boost inversion thread ->rt_priority %u gp_state %lu jiffies %lu\n",
-                               current->rt_priority, gp_state, end - start);
+                               current->rt_priority, gp_state, end - *start);
                        cur_ops->gp_kthread_dbg();
                        // Recheck after print to flag grace period ending during splat.
                        gp_done = cur_ops->poll_gp_state(gp_state);
@@ -955,6 +956,8 @@ static bool rcu_torture_boost_failed(unsigned long gp_state, unsigned long start
                }
 
                return true; // failed
+       } else if (cur_ops->check_boost_failed && !cur_ops->check_boost_failed(gp_state, NULL)) {
+               *start = jiffies;
        }
 
        return false; // passed
@@ -995,7 +998,7 @@ static int rcu_torture_boost(void *arg)
                while (time_before(jiffies, endtime)) {
                        // Has current GP gone too long?
                        if (gp_initiated && !failed && !cur_ops->poll_gp_state(gp_state))
-                               failed = rcu_torture_boost_failed(gp_state, gp_state_time, jiffies);
+                               failed = rcu_torture_boost_failed(gp_state, &gp_state_time);
                        // If we don't have a grace period in flight, start one.
                        if (!gp_initiated || cur_ops->poll_gp_state(gp_state)) {
                                gp_state = cur_ops->start_gp_poll();
@@ -1016,7 +1019,7 @@ static int rcu_torture_boost(void *arg)
 
                // In case the grace period extended beyond the end of the loop.
                if (gp_initiated && !failed && !cur_ops->poll_gp_state(gp_state))
-                       rcu_torture_boost_failed(gp_state, gp_state_time, jiffies);
+                       rcu_torture_boost_failed(gp_state, &gp_state_time);
 
                /*
                 * Set the start time of the next test interval.
index af92d9f..8bde1b5 100644 (file)
@@ -723,6 +723,10 @@ static void check_cpu_stall(struct rcu_data *rdp)
  * count this as an RCU priority boosting failure.  A return of true says
  * RCU priority boosting is to blame, and false says otherwise.  If false
  * is returned, the first of the CPUs to blame is stored through cpup.
+ *
+ * If cpup is NULL, then a lockless quick check is carried out, suitable
+ * for high-rate usage.  On the other hand, if cpup is non-NULL, each
+ * rcu_node structure's ->lock is acquired, ruling out high-rate usage.
  */
 bool rcu_check_boost_fail(unsigned long gp_state, int *cpup)
 {
@@ -731,6 +735,12 @@ bool rcu_check_boost_fail(unsigned long gp_state, int *cpup)
        struct rcu_node *rnp;
 
        rcu_for_each_leaf_node(rnp) {
+               if (!cpup) {
+                       if (READ_ONCE(rnp->qsmask))
+                               return false;
+                       else
+                               continue;
+               }
                raw_spin_lock_irqsave_rcu_node(rnp, flags);
                if (!rnp->qsmask) {
                        // No CPUs without quiescent states for this rnp.