lib: proportion: fix underflow in prop_norm_percpu()
authorPeter Zijlstra <a.p.zijlstra@chello.nl>
Sat, 22 Dec 2007 22:03:29 +0000 (14:03 -0800)
committerLinus Torvalds <torvalds@woody.linux-foundation.org>
Sun, 23 Dec 2007 20:54:37 +0000 (12:54 -0800)
Zhe Jiang noticed that its possible to underflow pl->events in
prop_norm_percpu() when the value returned by percpu_counter_read() is less
than the error on that read and the period delay > 1.  In that case half might
not trigger the batch increment and the value will be identical on the next
iteration, causing the same half to be subtracted again and again.

Fix this by rewriting the division as a single subtraction instead of a
subtraction loop and using percpu_counter_sum() when the value returned by
percpu_counter_read() is smaller than the error.

The latter is still needed if we want pl->events to shrink properly in the
error region.

[akpm@linux-foundation.org: cleanups]
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Reviewed-by: Jiang Zhe <zhe.jiang@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
lib/proportions.c

index 332d8c58184d4964406dfe71179b877fd9f927f6..9508d9a7af3ef9ba482925a9d1cdc34ef2fddd0e 100644 (file)
@@ -190,6 +190,8 @@ prop_adjust_shift(int *pl_shift, unsigned long *pl_period, int new_shift)
  * PERCPU
  */
 
+#define PROP_BATCH (8*(1+ilog2(nr_cpu_ids)))
+
 int prop_local_init_percpu(struct prop_local_percpu *pl)
 {
        spin_lock_init(&pl->lock);
@@ -230,31 +232,24 @@ void prop_norm_percpu(struct prop_global *pg, struct prop_local_percpu *pl)
 
        spin_lock_irqsave(&pl->lock, flags);
        prop_adjust_shift(&pl->shift, &pl->period, pg->shift);
+
        /*
         * For each missed period, we half the local counter.
         * basically:
         *   pl->events >> (global_period - pl->period);
-        *
-        * but since the distributed nature of percpu counters make division
-        * rather hard, use a regular subtraction loop. This is safe, because
-        * the events will only every be incremented, hence the subtraction
-        * can never result in a negative number.
         */
-       while (pl->period != global_period) {
-               unsigned long val = percpu_counter_read(&pl->events);
-               unsigned long half = (val + 1) >> 1;
-
-               /*
-                * Half of zero won't be much less, break out.
-                * This limits the loop to shift iterations, even
-                * if we missed a million.
-                */
-               if (!val)
-                       break;
-
-               percpu_counter_add(&pl->events, -half);
-               pl->period += period;
-       }
+       period = (global_period - pl->period) >> (pg->shift - 1);
+       if (period < BITS_PER_LONG) {
+               s64 val = percpu_counter_read(&pl->events);
+
+               if (val < (nr_cpu_ids * PROP_BATCH))
+                       val = percpu_counter_sum(&pl->events);
+
+               __percpu_counter_add(&pl->events, -val + (val >> period),
+                                       PROP_BATCH);
+       } else
+               percpu_counter_set(&pl->events, 0);
+
        pl->period = global_period;
        spin_unlock_irqrestore(&pl->lock, flags);
 }
@@ -267,7 +262,7 @@ void __prop_inc_percpu(struct prop_descriptor *pd, struct prop_local_percpu *pl)
        struct prop_global *pg = prop_get_global(pd);
 
        prop_norm_percpu(pg, pl);
-       percpu_counter_add(&pl->events, 1);
+       __percpu_counter_add(&pl->events, 1, PROP_BATCH);
        percpu_counter_add(&pg->events, 1);
        prop_put_global(pd, pg);
 }