rcu: Make rcu_gp_kthread() check for early-boot activity
authorPaul E. McKenney <paulmck@linux.vnet.ibm.com>
Wed, 11 Apr 2018 23:41:20 +0000 (16:41 -0700)
committerPaul E. McKenney <paulmck@linux.vnet.ibm.com>
Tue, 15 May 2018 17:29:24 +0000 (10:29 -0700)
The rcu_gp_kthread() function immediately sleeps waiting to be notified
of the need for a new grace period, which currently works because there
are a number of code sequences that will provide the needed wakeup later.
However, some of these code sequences need to acquire the root rcu_node
structure's ->lock, and contention on that lock has started manifesting.
This commit therefore makes rcu_gp_kthread() check for early-boot activity
when it starts up, omitting the initial sleep in that case.

Reported-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Tested-by: Nicholas Piggin <npiggin@gmail.com>
kernel/rcu/tree.c

index 79fb99951a0cdf219359635fed9cb91d229d1768..497f139056c7f72db2273b9387d8be1ffc2b5052 100644 (file)
@@ -2192,6 +2192,12 @@ static int __noreturn rcu_gp_kthread(void *arg)
        struct rcu_state *rsp = arg;
        struct rcu_node *rnp = rcu_get_root(rsp);
 
+       /* Check for early-boot work. */
+       raw_spin_lock_irq_rcu_node(rnp);
+       if (need_any_future_gp(rnp))
+               WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT);
+       raw_spin_unlock_irq_rcu_node(rnp);
+
        rcu_bind_gp_kthread();
        for (;;) {