mm: page allocator: adjust the per-cpu counter threshold when memory is low
[~shefty/rdma-dev.git] / mm / vmstat.c
index 312d728976f1661c4fa335a1ead46356b3bf091b..bc0f095791b4ff026279c54cf667f50da410ad05 100644 (file)
@@ -83,6 +83,30 @@ EXPORT_SYMBOL(vm_stat);
 
 #ifdef CONFIG_SMP
 
+static int calculate_pressure_threshold(struct zone *zone)
+{
+       int threshold;
+       int watermark_distance;
+
+       /*
+        * As vmstats are not up to date, there is drift between the estimated
+        * and real values. For high thresholds and a high number of CPUs, it
+        * is possible for the min watermark to be breached while the estimated
+        * value looks fine. The pressure threshold is a reduced value such
+        * that even the maximum amount of drift will not accidentally breach
+        * the min watermark
+        */
+       watermark_distance = low_wmark_pages(zone) - min_wmark_pages(zone);
+       threshold = max(1, (int)(watermark_distance / num_online_cpus()));
+
+       /*
+        * Maximum threshold is 125
+        */
+       threshold = min(125, threshold);
+
+       return threshold;
+}
+
 static int calculate_threshold(struct zone *zone)
 {
        int threshold;
@@ -161,6 +185,48 @@ static void refresh_zone_stat_thresholds(void)
        }
 }
 
+void reduce_pgdat_percpu_threshold(pg_data_t *pgdat)
+{
+       struct zone *zone;
+       int cpu;
+       int threshold;
+       int i;
+
+       get_online_cpus();
+       for (i = 0; i < pgdat->nr_zones; i++) {
+               zone = &pgdat->node_zones[i];
+               if (!zone->percpu_drift_mark)
+                       continue;
+
+               threshold = calculate_pressure_threshold(zone);
+               for_each_online_cpu(cpu)
+                       per_cpu_ptr(zone->pageset, cpu)->stat_threshold
+                                                       = threshold;
+       }
+       put_online_cpus();
+}
+
+void restore_pgdat_percpu_threshold(pg_data_t *pgdat)
+{
+       struct zone *zone;
+       int cpu;
+       int threshold;
+       int i;
+
+       get_online_cpus();
+       for (i = 0; i < pgdat->nr_zones; i++) {
+               zone = &pgdat->node_zones[i];
+               if (!zone->percpu_drift_mark)
+                       continue;
+
+               threshold = calculate_threshold(zone);
+               for_each_online_cpu(cpu)
+                       per_cpu_ptr(zone->pageset, cpu)->stat_threshold
+                                                       = threshold;
+       }
+       put_online_cpus();
+}
+
 /*
  * For use when we know that interrupts are disabled.
  */
@@ -911,7 +977,7 @@ static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
                   "\n        scanned  %lu"
                   "\n        spanned  %lu"
                   "\n        present  %lu",
-                  zone_nr_free_pages(zone),
+                  zone_page_state(zone, NR_FREE_PAGES),
                   min_wmark_pages(zone),
                   low_wmark_pages(zone),
                   high_wmark_pages(zone),