memcg: fix shrinking memory to return -EBUSY by fixing retry algorithm
authorKAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Thu, 2 Apr 2009 23:57:36 +0000 (16:57 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Fri, 3 Apr 2009 02:04:55 +0000 (19:04 -0700)
As pointed out, shrinking memcg's limit should return -EBUSY after
reasonable retries.  This patch tries to fix the current behavior of
shrink_usage.

Before looking into "shrink should return -EBUSY" problem, we should fix
hierarchical reclaim code.  It compares current usage and current limit,
but it only makes sense when the kernel reclaims memory because hit
limits.  This is also a problem.

What this patch does are.

  1. add new argument "shrink" to hierarchical reclaim. If "shrink==true",
     hierarchical reclaim returns immediately and the caller checks the kernel
     should shrink more or not.
     (At shrinking memory, usage is always smaller than limit. So check for
      usage < limit is useless.)

  2. For adjusting to above change, 2 changes in "shrink"'s retry path.
     2-a. retry_count depends on # of children because the kernel visits
  the children under hierarchy one by one.
     2-b. rather than checking return value of hierarchical_reclaim's progress,
  compares usage-before-shrink and usage-after-shrink.
  If usage-before-shrink <= usage-after-shrink, retry_count is
  decremented.

Reported-by: Li Zefan <lizf@cn.fujitsu.com>
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Paul Menage <menage@google.com>
Cc: Balbir Singh <balbir@in.ibm.com>
Cc: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/memcontrol.c

index 33fc0302e29e8a0e1eac76354ce18a2b80bae5c2..6f6a575e77ad970156b2fceee7a022707ce1805f 100644 (file)
@@ -702,6 +702,23 @@ static unsigned int get_swappiness(struct mem_cgroup *memcg)
        return swappiness;
 }
 
+static int mem_cgroup_count_children_cb(struct mem_cgroup *mem, void *data)
+{
+       int *val = data;
+       (*val)++;
+       return 0;
+}
+/*
+ * This function returns the number of memcg under hierarchy tree. Returns
+ * 1(self count) if no children.
+ */
+static int mem_cgroup_count_children(struct mem_cgroup *mem)
+{
+       int num = 0;
+       mem_cgroup_walk_tree(mem, &num, mem_cgroup_count_children_cb);
+       return num;
+}
+
 /*
  * Visit the first child (need not be the first child as per the ordering
  * of the cgroup list, since we track last_scanned_child) of @mem and use
@@ -750,9 +767,11 @@ mem_cgroup_select_victim(struct mem_cgroup *root_mem)
  *
  * We give up and return to the caller when we visit root_mem twice.
  * (other groups can be removed while we're walking....)
+ *
+ * If shrink==true, for avoiding to free too much, this returns immedieately.
  */
 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
-                                               gfp_t gfp_mask, bool noswap)
+                                  gfp_t gfp_mask, bool noswap, bool shrink)
 {
        struct mem_cgroup *victim;
        int ret, total = 0;
@@ -771,6 +790,13 @@ static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
                ret = try_to_free_mem_cgroup_pages(victim, gfp_mask, noswap,
                                                   get_swappiness(victim));
                css_put(&victim->css);
+               /*
+                * At shrinking usage, we can't check we should stop here or
+                * reclaim more. It's depends on callers. last_scanned_child
+                * will work enough for keeping fairness under tree.
+                */
+               if (shrink)
+                       return ret;
                total += ret;
                if (mem_cgroup_check_under_limit(root_mem))
                        return 1 + total;
@@ -856,7 +882,7 @@ static int __mem_cgroup_try_charge(struct mm_struct *mm,
                        goto nomem;
 
                ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, gfp_mask,
-                                                       noswap);
+                                                       noswap, false);
                if (ret)
                        continue;
 
@@ -1489,7 +1515,8 @@ int mem_cgroup_shrink_usage(struct page *page,
                return 0;
 
        do {
-               progress = mem_cgroup_hierarchical_reclaim(mem, gfp_mask, true);
+               progress = mem_cgroup_hierarchical_reclaim(mem,
+                                       gfp_mask, true, false);
                progress += mem_cgroup_check_under_limit(mem);
        } while (!progress && --retry);
 
@@ -1504,11 +1531,21 @@ static DEFINE_MUTEX(set_limit_mutex);
 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
                                unsigned long long val)
 {
-
-       int retry_count = MEM_CGROUP_RECLAIM_RETRIES;
+       int retry_count;
        int progress;
        u64 memswlimit;
        int ret = 0;
+       int children = mem_cgroup_count_children(memcg);
+       u64 curusage, oldusage;
+
+       /*
+        * For keeping hierarchical_reclaim simple, how long we should retry
+        * is depends on callers. We set our retry-count to be function
+        * of # of children which we should visit in this loop.
+        */
+       retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
+
+       oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
 
        while (retry_count) {
                if (signal_pending(current)) {
@@ -1534,8 +1571,13 @@ static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
                        break;
 
                progress = mem_cgroup_hierarchical_reclaim(memcg, GFP_KERNEL,
-                                                          false);
-               if (!progress)                  retry_count--;
+                                                  false, true);
+               curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
+               /* Usage is reduced ? */
+               if (curusage >= oldusage)
+                       retry_count--;
+               else
+                       oldusage = curusage;
        }
 
        return ret;
@@ -1544,13 +1586,16 @@ static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
 int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
                                unsigned long long val)
 {
-       int retry_count = MEM_CGROUP_RECLAIM_RETRIES;
+       int retry_count;
        u64 memlimit, oldusage, curusage;
-       int ret;
+       int children = mem_cgroup_count_children(memcg);
+       int ret = -EBUSY;
 
        if (!do_swap_account)
                return -EINVAL;
-
+       /* see mem_cgroup_resize_res_limit */
+       retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
+       oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
        while (retry_count) {
                if (signal_pending(current)) {
                        ret = -EINTR;
@@ -1574,11 +1619,13 @@ int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
                if (!ret)
                        break;
 
-               oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
-               mem_cgroup_hierarchical_reclaim(memcg, GFP_KERNEL, true);
+               mem_cgroup_hierarchical_reclaim(memcg, GFP_KERNEL, true, true);
                curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
+               /* Usage is reduced ? */
                if (curusage >= oldusage)
                        retry_count--;
+               else
+                       oldusage = curusage;
        }
        return ret;
 }