PM / Wakeup: Combine atomic counters to avoid reordering issues
authorRafael J. Wysocki <rjw@sisk.pl>
Mon, 31 Jan 2011 10:06:39 +0000 (11:06 +0100)
committerRafael J. Wysocki <rjw@sisk.pl>
Mon, 14 Mar 2011 23:43:13 +0000 (00:43 +0100)
The memory barrier in wakeup_source_deactivate() is supposed to
prevent the callers of pm_wakeup_pending() and pm_get_wakeup_count()
from seeing the new value of events_in_progress (0, in particular)
and the old value of event_count at the same time.  However, if
wakeup_source_deactivate() is executed by CPU0 and, for instance,
pm_wakeup_pending() is executed by CPU1, where both processors can
reorder operations, the memory barrier in wakeup_source_deactivate()
doesn't affect CPU1 which can reorder reads.  In that case CPU1 may
very well decide to fetch event_count before it's modified and
events_in_progress after it's been updated, so pm_wakeup_pending()
may fail to detect a wakeup event.  This issue can be addressed by
using a single atomic variable to store both events_in_progress
and event_count, so that they can be updated together in a single
atomic operation.

Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
drivers/base/power/wakeup.c

index 8ec406d8f548eae6b84ed692236c12ee45139fa4..e5e73b5efc80148a70e92ab6620a5c61b6b75b36 100644 (file)
  */
 bool events_check_enabled;
 
-/* The counter of registered wakeup events. */
-static atomic_t event_count = ATOMIC_INIT(0);
-/* A preserved old value of event_count. */
+/*
+ * Combined counters of registered wakeup events and wakeup events in progress.
+ * They need to be modified together atomically, so it's better to use one
+ * atomic variable to hold them both.
+ */
+static atomic_t combined_event_count = ATOMIC_INIT(0);
+
+#define IN_PROGRESS_BITS       (sizeof(int) * 4)
+#define MAX_IN_PROGRESS                ((1 << IN_PROGRESS_BITS) - 1)
+
+static void split_counters(unsigned int *cnt, unsigned int *inpr)
+{
+       unsigned int comb = atomic_read(&combined_event_count);
+
+       *cnt = (comb >> IN_PROGRESS_BITS);
+       *inpr = comb & MAX_IN_PROGRESS;
+}
+
+/* A preserved old value of the events counter. */
 static unsigned int saved_count;
-/* The counter of wakeup events being processed. */
-static atomic_t events_in_progress = ATOMIC_INIT(0);
 
 static DEFINE_SPINLOCK(events_lock);
 
@@ -307,7 +321,8 @@ static void wakeup_source_activate(struct wakeup_source *ws)
        ws->timer_expires = jiffies;
        ws->last_time = ktime_get();
 
-       atomic_inc(&events_in_progress);
+       /* Increment the counter of events in progress. */
+       atomic_inc(&combined_event_count);
 }
 
 /**
@@ -394,14 +409,10 @@ static void wakeup_source_deactivate(struct wakeup_source *ws)
        del_timer(&ws->timer);
 
        /*
-        * event_count has to be incremented before events_in_progress is
-        * modified, so that the callers of pm_check_wakeup_events() and
-        * pm_save_wakeup_count() don't see the old value of event_count and
-        * events_in_progress equal to zero at the same time.
+        * Increment the counter of registered wakeup events and decrement the
+        * couter of wakeup events in progress simultaneously.
         */
-       atomic_inc(&event_count);
-       smp_mb__before_atomic_dec();
-       atomic_dec(&events_in_progress);
+       atomic_add(MAX_IN_PROGRESS, &combined_event_count);
 }
 
 /**
@@ -556,8 +567,10 @@ bool pm_wakeup_pending(void)
 
        spin_lock_irqsave(&events_lock, flags);
        if (events_check_enabled) {
-               ret = ((unsigned int)atomic_read(&event_count) != saved_count)
-                       || atomic_read(&events_in_progress);
+               unsigned int cnt, inpr;
+
+               split_counters(&cnt, &inpr);
+               ret = (cnt != saved_count || inpr > 0);
                events_check_enabled = !ret;
        }
        spin_unlock_irqrestore(&events_lock, flags);
@@ -579,19 +592,22 @@ bool pm_wakeup_pending(void)
  */
 bool pm_get_wakeup_count(unsigned int *count)
 {
-       bool ret;
+       unsigned int cnt, inpr;
 
        if (capable(CAP_SYS_ADMIN))
                events_check_enabled = false;
 
-       while (atomic_read(&events_in_progress) && !signal_pending(current)) {
+       for (;;) {
+               split_counters(&cnt, &inpr);
+               if (inpr == 0 || signal_pending(current))
+                       break;
                pm_wakeup_update_hit_counts();
                schedule_timeout_interruptible(msecs_to_jiffies(TIMEOUT));
        }
 
-       ret = !atomic_read(&events_in_progress);
-       *count = atomic_read(&event_count);
-       return ret;
+       split_counters(&cnt, &inpr);
+       *count = cnt;
+       return !inpr;
 }
 
 /**
@@ -605,11 +621,12 @@ bool pm_get_wakeup_count(unsigned int *count)
  */
 bool pm_save_wakeup_count(unsigned int count)
 {
+       unsigned int cnt, inpr;
        bool ret = false;
 
        spin_lock_irq(&events_lock);
-       if (count == (unsigned int)atomic_read(&event_count)
-           && !atomic_read(&events_in_progress)) {
+       split_counters(&cnt, &inpr);
+       if (cnt == count && inpr == 0) {
                saved_count = count;
                events_check_enabled = true;
                ret = true;