rcu: Remove _rcu_barrier() dependency on __stop_machine()
authorPaul E. McKenney <paul.mckenney@linaro.org>
Fri, 3 Aug 2012 00:43:50 +0000 (17:43 -0700)
committerPaul E. McKenney <paulmck@linux.vnet.ibm.com>
Sun, 23 Sep 2012 14:43:55 +0000 (07:43 -0700)
Currently, _rcu_barrier() relies on preempt_disable() to prevent
any CPU from going offline, which in turn depends on CPU hotplug's
use of __stop_machine().

This patch therefore makes _rcu_barrier() use get_online_cpus() to
block CPU-hotplug operations.  This has the added benefit of removing
the need for _rcu_barrier() to adopt callbacks:  Because CPU-hotplug
operations are excluded, there can be no callbacks to adopt.  This
commit simplifies the code accordingly.

Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Reviewed-by: Josh Triplett <josh@joshtriplett.org>
kernel/rcutree.c
kernel/rcutree.h
kernel/rcutree_trace.c

index f7bcd9e..c45d3f7 100644 (file)
@@ -1392,17 +1392,6 @@ static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
        int i;
        struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
 
-       /*
-        * If there is an rcu_barrier() operation in progress, then
-        * only the task doing that operation is permitted to adopt
-        * callbacks.  To do otherwise breaks rcu_barrier() and friends
-        * by causing them to fail to wait for the callbacks in the
-        * orphanage.
-        */
-       if (rsp->rcu_barrier_in_progress &&
-           rsp->rcu_barrier_in_progress != current)
-               return;
-
        /* Do the accounting first. */
        rdp->qlen_lazy += rsp->qlen_lazy;
        rdp->qlen += rsp->qlen;
@@ -1457,9 +1446,8 @@ static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
  * The CPU has been completely removed, and some other CPU is reporting
  * this fact from process context.  Do the remainder of the cleanup,
  * including orphaning the outgoing CPU's RCU callbacks, and also
- * adopting them, if there is no _rcu_barrier() instance running.
- * There can only be one CPU hotplug operation at a time, so no other
- * CPU can be attempting to update rcu_cpu_kthread_task.
+ * adopting them.  There can only be one CPU hotplug operation at a time,
+ * so no other CPU can be attempting to update rcu_cpu_kthread_task.
  */
 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
 {
@@ -1521,10 +1509,6 @@ static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
 
 #else /* #ifdef CONFIG_HOTPLUG_CPU */
 
-static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
-{
-}
-
 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
 {
 }
@@ -2328,13 +2312,10 @@ static void rcu_barrier_func(void *type)
 static void _rcu_barrier(struct rcu_state *rsp)
 {
        int cpu;
-       unsigned long flags;
        struct rcu_data *rdp;
-       struct rcu_data rd;
        unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done);
        unsigned long snap_done;
 
-       init_rcu_head_on_stack(&rd.barrier_head);
        _rcu_barrier_trace(rsp, "Begin", -1, snap);
 
        /* Take mutex to serialize concurrent rcu_barrier() requests. */
@@ -2374,70 +2355,30 @@ static void _rcu_barrier(struct rcu_state *rsp)
        /*
         * Initialize the count to one rather than to zero in order to
         * avoid a too-soon return to zero in case of a short grace period
-        * (or preemption of this task).  Also flag this task as doing
-        * an rcu_barrier().  This will prevent anyone else from adopting
-        * orphaned callbacks, which could cause otherwise failure if a
-        * CPU went offline and quickly came back online.  To see this,
-        * consider the following sequence of events:
-        *
-        * 1.   We cause CPU 0 to post an rcu_barrier_callback() callback.
-        * 2.   CPU 1 goes offline, orphaning its callbacks.
-        * 3.   CPU 0 adopts CPU 1's orphaned callbacks.
-        * 4.   CPU 1 comes back online.
-        * 5.   We cause CPU 1 to post an rcu_barrier_callback() callback.
-        * 6.   Both rcu_barrier_callback() callbacks are invoked, awakening
-        *      us -- but before CPU 1's orphaned callbacks are invoked!!!
+        * (or preemption of this task).  Exclude CPU-hotplug operations
+        * to ensure that no offline CPU has callbacks queued.
         */
        init_completion(&rsp->barrier_completion);
        atomic_set(&rsp->barrier_cpu_count, 1);
-       raw_spin_lock_irqsave(&rsp->onofflock, flags);
-       rsp->rcu_barrier_in_progress = current;
-       raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
+       get_online_cpus();
 
        /*
-        * Force every CPU with callbacks to register a new callback
-        * that will tell us when all the preceding callbacks have
-        * been invoked.  If an offline CPU has callbacks, wait for
-        * it to either come back online or to finish orphaning those
-        * callbacks.
+        * Force each CPU with callbacks to register a new callback.
+        * When that callback is invoked, we will know that all of the
+        * corresponding CPU's preceding callbacks have been invoked.
         */
-       for_each_possible_cpu(cpu) {
-               preempt_disable();
+       for_each_online_cpu(cpu) {
                rdp = per_cpu_ptr(rsp->rda, cpu);
-               if (cpu_is_offline(cpu)) {
-                       _rcu_barrier_trace(rsp, "Offline", cpu,
-                                          rsp->n_barrier_done);
-                       preempt_enable();
-                       while (cpu_is_offline(cpu) && ACCESS_ONCE(rdp->qlen))
-                               schedule_timeout_interruptible(1);
-               } else if (ACCESS_ONCE(rdp->qlen)) {
+               if (ACCESS_ONCE(rdp->qlen)) {
                        _rcu_barrier_trace(rsp, "OnlineQ", cpu,
                                           rsp->n_barrier_done);
                        smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
-                       preempt_enable();
                } else {
                        _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
                                           rsp->n_barrier_done);
-                       preempt_enable();
                }
        }
-
-       /*
-        * Now that all online CPUs have rcu_barrier_callback() callbacks
-        * posted, we can adopt all of the orphaned callbacks and place
-        * an rcu_barrier_callback() callback after them.  When that is done,
-        * we are guaranteed to have an rcu_barrier_callback() callback
-        * following every callback that could possibly have been
-        * registered before _rcu_barrier() was called.
-        */
-       raw_spin_lock_irqsave(&rsp->onofflock, flags);
-       rcu_adopt_orphan_cbs(rsp);
-       rsp->rcu_barrier_in_progress = NULL;
-       raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
-       atomic_inc(&rsp->barrier_cpu_count);
-       smp_mb__after_atomic_inc(); /* Ensure atomic_inc() before callback. */
-       rd.rsp = rsp;
-       rsp->call(&rd.barrier_head, rcu_barrier_callback);
+       put_online_cpus();
 
        /*
         * Now that we have an rcu_barrier_callback() callback on each
@@ -2458,8 +2399,6 @@ static void _rcu_barrier(struct rcu_state *rsp)
 
        /* Other rcu_barrier() invocations can now safely proceed. */
        mutex_unlock(&rsp->barrier_mutex);
-
-       destroy_rcu_head_on_stack(&rd.barrier_head);
 }
 
 /**
index 4d29169..94dfdf1 100644 (file)
@@ -398,9 +398,6 @@ struct rcu_state {
        struct rcu_head **orphan_donetail;      /* Tail of above. */
        long qlen_lazy;                         /* Number of lazy callbacks. */
        long qlen;                              /* Total number of callbacks. */
-       struct task_struct *rcu_barrier_in_progress;
-                                               /* Task doing rcu_barrier(), */
-                                               /*  or NULL if no barrier. */
        struct mutex barrier_mutex;             /* Guards barrier fields. */
        atomic_t barrier_cpu_count;             /* # CPUs waiting on. */
        struct completion barrier_completion;   /* Wake at barrier end. */
index abffb48..6a2e52a 100644 (file)
@@ -51,8 +51,8 @@ static int show_rcubarrier(struct seq_file *m, void *unused)
        struct rcu_state *rsp;
 
        for_each_rcu_flavor(rsp)
-               seq_printf(m, "%s: %c bcc: %d nbd: %lu\n",
-                          rsp->name, rsp->rcu_barrier_in_progress ? 'B' : '.',
+               seq_printf(m, "%s: bcc: %d nbd: %lu\n",
+                          rsp->name,
                           atomic_read(&rsp->barrier_cpu_count),
                           rsp->n_barrier_done);
        return 0;