rcu: More on deadlock between CPU hotplug and expedited grace periods
authorPaul E. McKenney <paulmck@linux.vnet.ibm.com>
Wed, 22 Oct 2014 17:00:05 +0000 (10:00 -0700)
committerPaul E. McKenney <paulmck@linux.vnet.ibm.com>
Thu, 23 Oct 2014 14:51:17 +0000 (07:51 -0700)
Commit dd56af42bd82 (rcu: Eliminate deadlock between CPU hotplug and
expedited grace periods) was incomplete.  Although it did eliminate
deadlocks involving synchronize_sched_expedited()'s acquisition of
cpu_hotplug.lock via get_online_cpus(), it did nothing about the similar
deadlock involving acquisition of this same lock via put_online_cpus().
This deadlock became apparent with testing involving hibernation.

This commit therefore changes put_online_cpus() acquisition of this lock
to be conditional, and increments a new cpu_hotplug.puts_pending field
in case of acquisition failure.  Then cpu_hotplug_begin() checks for this
new field being non-zero, and applies any changes to cpu_hotplug.refcount.

Reported-by: Jiri Kosina <jkosina@suse.cz>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Tested-by: Jiri Kosina <jkosina@suse.cz>
Tested-by: Borislav Petkov <bp@suse.de>
kernel/cpu.c

index 356450f09c1f89d1ee35820a59a59e3a68cebe63..90a3d017b90ce56edee804892492bd5b62442c62 100644 (file)
@@ -64,6 +64,8 @@ static struct {
         * an ongoing cpu hotplug operation.
         */
        int refcount;
+       /* And allows lockless put_online_cpus(). */
+       atomic_t puts_pending;
 
 #ifdef CONFIG_DEBUG_LOCK_ALLOC
        struct lockdep_map dep_map;
@@ -113,7 +115,11 @@ void put_online_cpus(void)
 {
        if (cpu_hotplug.active_writer == current)
                return;
-       mutex_lock(&cpu_hotplug.lock);
+       if (!mutex_trylock(&cpu_hotplug.lock)) {
+               atomic_inc(&cpu_hotplug.puts_pending);
+               cpuhp_lock_release();
+               return;
+       }
 
        if (WARN_ON(!cpu_hotplug.refcount))
                cpu_hotplug.refcount++; /* try to fix things up */
@@ -155,6 +161,12 @@ void cpu_hotplug_begin(void)
        cpuhp_lock_acquire();
        for (;;) {
                mutex_lock(&cpu_hotplug.lock);
+               if (atomic_read(&cpu_hotplug.puts_pending)) {
+                       int delta;
+
+                       delta = atomic_xchg(&cpu_hotplug.puts_pending, 0);
+                       cpu_hotplug.refcount -= delta;
+               }
                if (likely(!cpu_hotplug.refcount))
                        break;
                __set_current_state(TASK_UNINTERRUPTIBLE);