mm: memcontrol: revert use of root_mem_cgroup res_counter
authorJohannes Weiner <hannes@cmpxchg.org>
Fri, 5 Sep 2014 12:43:57 +0000 (08:43 -0400)
committerLinus Torvalds <torvalds@linux-foundation.org>
Fri, 5 Sep 2014 15:19:02 +0000 (08:19 -0700)
Dave Hansen reports a massive scalability regression in an uncontained
page fault benchmark with more than 30 concurrent threads, which he
bisected down to 05b843012335 ("mm: memcontrol: use root_mem_cgroup
res_counter") and pin-pointed on res_counter spinlock contention.

That change relied on the per-cpu charge caches to mostly swallow the
res_counter costs, but it's apparent that the caches don't scale yet.

Revert memcg back to bypassing res_counters on the root level in order
to restore performance for uncontained workloads.

Reported-by: Dave Hansen <dave@sr71.net>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Tested-by: Dave Hansen <dave.hansen@intel.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Acked-by: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/memcontrol.c

index ec4dcf1b9562b6299f215e754768da18a36b156e..085dc6d2f876374c07a518866ecb1d25a2e8c419 100644 (file)
@@ -2534,6 +2534,8 @@ static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
        unsigned long long size;
        int ret = 0;
 
+       if (mem_cgroup_is_root(memcg))
+               goto done;
 retry:
        if (consume_stock(memcg, nr_pages))
                goto done;
@@ -2611,9 +2613,7 @@ nomem:
        if (!(gfp_mask & __GFP_NOFAIL))
                return -ENOMEM;
 bypass:
-       memcg = root_mem_cgroup;
-       ret = -EINTR;
-       goto retry;
+       return -EINTR;
 
 done_restock:
        if (batch > nr_pages)
@@ -2626,6 +2626,9 @@ static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
 {
        unsigned long bytes = nr_pages * PAGE_SIZE;
 
+       if (mem_cgroup_is_root(memcg))
+               return;
+
        res_counter_uncharge(&memcg->res, bytes);
        if (do_swap_account)
                res_counter_uncharge(&memcg->memsw, bytes);
@@ -2640,6 +2643,9 @@ static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
 {
        unsigned long bytes = nr_pages * PAGE_SIZE;
 
+       if (mem_cgroup_is_root(memcg))
+               return;
+
        res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
        if (do_swap_account)
                res_counter_uncharge_until(&memcg->memsw,
@@ -4093,6 +4099,46 @@ out:
        return retval;
 }
 
+static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
+                                              enum mem_cgroup_stat_index idx)
+{
+       struct mem_cgroup *iter;
+       long val = 0;
+
+       /* Per-cpu values can be negative, use a signed accumulator */
+       for_each_mem_cgroup_tree(iter, memcg)
+               val += mem_cgroup_read_stat(iter, idx);
+
+       if (val < 0) /* race ? */
+               val = 0;
+       return val;
+}
+
+static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
+{
+       u64 val;
+
+       if (!mem_cgroup_is_root(memcg)) {
+               if (!swap)
+                       return res_counter_read_u64(&memcg->res, RES_USAGE);
+               else
+                       return res_counter_read_u64(&memcg->memsw, RES_USAGE);
+       }
+
+       /*
+        * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
+        * as well as in MEM_CGROUP_STAT_RSS_HUGE.
+        */
+       val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
+       val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
+
+       if (swap)
+               val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
+
+       return val << PAGE_SHIFT;
+}
+
+
 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
                               struct cftype *cft)
 {
@@ -4102,8 +4148,12 @@ static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
 
        switch (type) {
        case _MEM:
+               if (name == RES_USAGE)
+                       return mem_cgroup_usage(memcg, false);
                return res_counter_read_u64(&memcg->res, name);
        case _MEMSWAP:
+               if (name == RES_USAGE)
+                       return mem_cgroup_usage(memcg, true);
                return res_counter_read_u64(&memcg->memsw, name);
        case _KMEM:
                return res_counter_read_u64(&memcg->kmem, name);
@@ -4572,10 +4622,7 @@ static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
        if (!t)
                goto unlock;
 
-       if (!swap)
-               usage = res_counter_read_u64(&memcg->res, RES_USAGE);
-       else
-               usage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
+       usage = mem_cgroup_usage(memcg, swap);
 
        /*
         * current_threshold points to threshold just below or equal to usage.
@@ -4673,10 +4720,10 @@ static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
 
        if (type == _MEM) {
                thresholds = &memcg->thresholds;
-               usage = res_counter_read_u64(&memcg->res, RES_USAGE);
+               usage = mem_cgroup_usage(memcg, false);
        } else if (type == _MEMSWAP) {
                thresholds = &memcg->memsw_thresholds;
-               usage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
+               usage = mem_cgroup_usage(memcg, true);
        } else
                BUG();
 
@@ -4762,10 +4809,10 @@ static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
 
        if (type == _MEM) {
                thresholds = &memcg->thresholds;
-               usage = res_counter_read_u64(&memcg->res, RES_USAGE);
+               usage = mem_cgroup_usage(memcg, false);
        } else if (type == _MEMSWAP) {
                thresholds = &memcg->memsw_thresholds;
-               usage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
+               usage = mem_cgroup_usage(memcg, true);
        } else
                BUG();
 
@@ -5525,9 +5572,9 @@ mem_cgroup_css_online(struct cgroup_subsys_state *css)
                 * core guarantees its existence.
                 */
        } else {
-               res_counter_init(&memcg->res, &root_mem_cgroup->res);
-               res_counter_init(&memcg->memsw, &root_mem_cgroup->memsw);
-               res_counter_init(&memcg->kmem, &root_mem_cgroup->kmem);
+               res_counter_init(&memcg->res, NULL);
+               res_counter_init(&memcg->memsw, NULL);
+               res_counter_init(&memcg->kmem, NULL);
                /*
                 * Deeper hierachy with use_hierarchy == false doesn't make
                 * much sense so let cgroup subsystem know about this
@@ -5969,8 +6016,9 @@ static void __mem_cgroup_clear_mc(void)
        /* we must fixup refcnts and charges */
        if (mc.moved_swap) {
                /* uncharge swap account from the old cgroup */
-               res_counter_uncharge(&mc.from->memsw,
-                                    PAGE_SIZE * mc.moved_swap);
+               if (!mem_cgroup_is_root(mc.from))
+                       res_counter_uncharge(&mc.from->memsw,
+                                            PAGE_SIZE * mc.moved_swap);
 
                for (i = 0; i < mc.moved_swap; i++)
                        css_put(&mc.from->css);
@@ -5979,8 +6027,9 @@ static void __mem_cgroup_clear_mc(void)
                 * we charged both to->res and to->memsw, so we should
                 * uncharge to->res.
                 */
-               res_counter_uncharge(&mc.to->res,
-                                    PAGE_SIZE * mc.moved_swap);
+               if (!mem_cgroup_is_root(mc.to))
+                       res_counter_uncharge(&mc.to->res,
+                                            PAGE_SIZE * mc.moved_swap);
                /* we've already done css_get(mc.to) */
                mc.moved_swap = 0;
        }
@@ -6345,7 +6394,8 @@ void mem_cgroup_uncharge_swap(swp_entry_t entry)
        rcu_read_lock();
        memcg = mem_cgroup_lookup(id);
        if (memcg) {
-               res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
+               if (!mem_cgroup_is_root(memcg))
+                       res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
                mem_cgroup_swap_statistics(memcg, false);
                css_put(&memcg->css);
        }
@@ -6509,12 +6559,15 @@ static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
 {
        unsigned long flags;
 
-       if (nr_mem)
-               res_counter_uncharge(&memcg->res, nr_mem * PAGE_SIZE);
-       if (nr_memsw)
-               res_counter_uncharge(&memcg->memsw, nr_memsw * PAGE_SIZE);
-
-       memcg_oom_recover(memcg);
+       if (!mem_cgroup_is_root(memcg)) {
+               if (nr_mem)
+                       res_counter_uncharge(&memcg->res,
+                                            nr_mem * PAGE_SIZE);
+               if (nr_memsw)
+                       res_counter_uncharge(&memcg->memsw,
+                                            nr_memsw * PAGE_SIZE);
+               memcg_oom_recover(memcg);
+       }
 
        local_irq_save(flags);
        __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);