rcu: Move RCU grace-period cleanup into kthread
authorPaul E. McKenney <paulmck@linux.vnet.ibm.com>
Thu, 21 Jun 2012 00:07:14 +0000 (17:07 -0700)
committerPaul E. McKenney <paulmck@linux.vnet.ibm.com>
Sun, 23 Sep 2012 14:41:52 +0000 (07:41 -0700)
As a first step towards allowing grace-period cleanup to be preemptible,
this commit moves the RCU grace-period cleanup into the same kthread
that is now used to initialize grace periods.  This is needed to keep
scheduling latency down to a dull roar.

[ paulmck: Get rid of stray spin_lock_irqsave() calls. ]

Reported-by: Mike Galbraith <mgalbraith@suse.de>
Reported-by: Dimitri Sivanich <sivanich@sgi.com>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Reviewed-by: Josh Triplett <josh@joshtriplett.org>
kernel/rcutree.c

index 781e5f0b7b1754bc6f559cade0ba23fab7b2d465..52c3102dc5f7a987936c2e79b887519404454988 100644 (file)
@@ -1032,6 +1032,7 @@ rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_dat
  */
 static int __noreturn rcu_gp_kthread(void *arg)
 {
+       unsigned long gp_duration;
        struct rcu_data *rdp;
        struct rcu_node *rnp;
        struct rcu_state *rsp = arg;
@@ -1116,6 +1117,65 @@ static int __noreturn rcu_gp_kthread(void *arg)
                rsp->fqs_state = RCU_SIGNAL_INIT;
                raw_spin_unlock_irq(&rnp->lock);
                put_online_cpus();
+
+               /* Handle grace-period end. */
+               rnp = rcu_get_root(rsp);
+               for (;;) {
+                       wait_event_interruptible(rsp->gp_wq,
+                                                !ACCESS_ONCE(rnp->qsmask) &&
+                                                !rcu_preempt_blocked_readers_cgp(rnp));
+                       if (!ACCESS_ONCE(rnp->qsmask) &&
+                           !rcu_preempt_blocked_readers_cgp(rnp))
+                               break;
+                       flush_signals(current);
+               }
+
+               raw_spin_lock_irq(&rnp->lock);
+               gp_duration = jiffies - rsp->gp_start;
+               if (gp_duration > rsp->gp_max)
+                       rsp->gp_max = gp_duration;
+
+               /*
+                * We know the grace period is complete, but to everyone else
+                * it appears to still be ongoing.  But it is also the case
+                * that to everyone else it looks like there is nothing that
+                * they can do to advance the grace period.  It is therefore
+                * safe for us to drop the lock in order to mark the grace
+                * period as completed in all of the rcu_node structures.
+                *
+                * But if this CPU needs another grace period, it will take
+                * care of this while initializing the next grace period.
+                * We use RCU_WAIT_TAIL instead of the usual RCU_DONE_TAIL
+                * because the callbacks have not yet been advanced: Those
+                * callbacks are waiting on the grace period that just now
+                * completed.
+                */
+               if (*rdp->nxttail[RCU_WAIT_TAIL] == NULL) {
+                       raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
+
+                       /*
+                        * Propagate new ->completed value to rcu_node
+                        * structures so that other CPUs don't have to
+                        * wait until the start of the next grace period
+                        * to process their callbacks.
+                        */
+                       rcu_for_each_node_breadth_first(rsp, rnp) {
+                               /* irqs already disabled. */
+                               raw_spin_lock(&rnp->lock);
+                               rnp->completed = rsp->gpnum;
+                               /* irqs remain disabled. */
+                               raw_spin_unlock(&rnp->lock);
+                       }
+                       rnp = rcu_get_root(rsp);
+                       raw_spin_lock(&rnp->lock); /* irqs already disabled. */
+               }
+
+               rsp->completed = rsp->gpnum; /* Declare grace period done. */
+               trace_rcu_grace_period(rsp->name, rsp->completed, "end");
+               rsp->fqs_state = RCU_GP_IDLE;
+               if (cpu_needs_another_gp(rsp, rdp))
+                       rsp->gp_flags = 1;
+               raw_spin_unlock_irq(&rnp->lock);
        }
 }
 
@@ -1162,57 +1222,9 @@ rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
        __releases(rcu_get_root(rsp)->lock)
 {
-       unsigned long gp_duration;
-       struct rcu_node *rnp = rcu_get_root(rsp);
-       struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
-
        WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
-
-       /*
-        * Ensure that all grace-period and pre-grace-period activity
-        * is seen before the assignment to rsp->completed.
-        */
-       smp_mb(); /* See above block comment. */
-       gp_duration = jiffies - rsp->gp_start;
-       if (gp_duration > rsp->gp_max)
-               rsp->gp_max = gp_duration;
-
-       /*
-        * We know the grace period is complete, but to everyone else
-        * it appears to still be ongoing.  But it is also the case
-        * that to everyone else it looks like there is nothing that
-        * they can do to advance the grace period.  It is therefore
-        * safe for us to drop the lock in order to mark the grace
-        * period as completed in all of the rcu_node structures.
-        *
-        * But if this CPU needs another grace period, it will take
-        * care of this while initializing the next grace period.
-        * We use RCU_WAIT_TAIL instead of the usual RCU_DONE_TAIL
-        * because the callbacks have not yet been advanced: Those
-        * callbacks are waiting on the grace period that just now
-        * completed.
-        */
-       if (*rdp->nxttail[RCU_WAIT_TAIL] == NULL) {
-               raw_spin_unlock(&rnp->lock);     /* irqs remain disabled. */
-
-               /*
-                * Propagate new ->completed value to rcu_node structures
-                * so that other CPUs don't have to wait until the start
-                * of the next grace period to process their callbacks.
-                */
-               rcu_for_each_node_breadth_first(rsp, rnp) {
-                       raw_spin_lock(&rnp->lock); /* irqs already disabled. */
-                       rnp->completed = rsp->gpnum;
-                       raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
-               }
-               rnp = rcu_get_root(rsp);
-               raw_spin_lock(&rnp->lock); /* irqs already disabled. */
-       }
-
-       rsp->completed = rsp->gpnum;  /* Declare the grace period complete. */
-       trace_rcu_grace_period(rsp->name, rsp->completed, "end");
-       rsp->fqs_state = RCU_GP_IDLE;
-       rcu_start_gp(rsp, flags);  /* releases root node's rnp->lock. */
+       raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
+       wake_up(&rsp->gp_wq);  /* Memory barrier implied by wake_up() path. */
 }
 
 /*