memory controller: soft limit reclaim on contention
authorBalbir Singh <balbir@linux.vnet.ibm.com>
Wed, 23 Sep 2009 22:56:39 +0000 (15:56 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Thu, 24 Sep 2009 14:20:59 +0000 (07:20 -0700)
Implement reclaim from groups over their soft limit

Permit reclaim from memory cgroups on contention (via the direct reclaim
path).

memory cgroup soft limit reclaim finds the group that exceeds its soft
limit by the largest number of pages and reclaims pages from it and then
reinserts the cgroup into its correct place in the rbtree.

Add additional checks to mem_cgroup_hierarchical_reclaim() to detect long
loops in case all swap is turned off.  The code has been refactored and
the loop check (loop < 2) has been enhanced for soft limits.  For soft
limits, we try to do more targetted reclaim.  Instead of bailing out after
two loops, the routine now reclaims memory proportional to the size by
which the soft limit is exceeded.  The proportion has been empirically
determined.

[akpm@linux-foundation.org: build fix]
[kamezawa.hiroyu@jp.fujitsu.com: fix softlimit css refcnt handling]
[nishimura@mxp.nes.nec.co.jp: refcount of the "victim" should be decremented before exiting the loop]
Signed-off-by: Balbir Singh <balbir@linux.vnet.ibm.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Li Zefan <lizf@cn.fujitsu.com>
Acked-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Signed-off-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
include/linux/memcontrol.h
include/linux/swap.h
mm/memcontrol.c
mm/vmscan.c

index e46a0734ab6e68c9f4df95ce5f9771923af392ea..bf9213b2db8f2da2a443d8788d2d70eb0a003af8 100644 (file)
@@ -118,6 +118,9 @@ static inline bool mem_cgroup_disabled(void)
 
 extern bool mem_cgroup_oom_called(struct task_struct *task);
 void mem_cgroup_update_mapped_file_stat(struct page *page, int val);
+unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
+                                               gfp_t gfp_mask, int nid,
+                                               int zid);
 #else /* CONFIG_CGROUP_MEM_RES_CTLR */
 struct mem_cgroup;
 
@@ -276,6 +279,13 @@ static inline void mem_cgroup_update_mapped_file_stat(struct page *page,
 {
 }
 
+static inline
+unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
+                                           gfp_t gfp_mask, int nid, int zid)
+{
+       return 0;
+}
+
 #endif /* CONFIG_CGROUP_MEM_CONT */
 
 #endif /* _LINUX_MEMCONTROL_H */
index 6c990e658f4ed33dc2244b1b8ef5d1649cc22a31..4c78fea989b9ca36c805fbec953f77ba458bc9bf 100644 (file)
@@ -217,6 +217,11 @@ extern unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
 extern unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem,
                                                  gfp_t gfp_mask, bool noswap,
                                                  unsigned int swappiness);
+extern unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
+                                               gfp_t gfp_mask, bool noswap,
+                                               unsigned int swappiness,
+                                               struct zone *zone,
+                                               int nid);
 extern int __isolate_lru_page(struct page *page, int mode, int file);
 extern unsigned long shrink_all_memory(unsigned long nr_pages);
 extern int vm_swappiness;
index 90f0b13e1c3c4f41d4e0d164bebc7f5bff37ebea..011aba6cad70fa58a887bf6220a888a8d4985598 100644 (file)
@@ -139,6 +139,8 @@ struct mem_cgroup_per_zone {
        unsigned long long      usage_in_excess;/* Set to the value by which */
                                                /* the soft limit is exceeded*/
        bool                    on_tree;
+       struct mem_cgroup       *mem;           /* Back pointer, we cannot */
+                                               /* use container_of        */
 };
 /* Macro for accessing counter */
 #define MEM_CGROUP_ZSTAT(mz, idx)      ((mz)->count[(idx)])
@@ -228,6 +230,13 @@ struct mem_cgroup {
        struct mem_cgroup_stat stat;
 };
 
+/*
+ * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
+ * limit reclaim to prevent infinite loops, if they ever occur.
+ */
+#define        MEM_CGROUP_MAX_RECLAIM_LOOPS            (100)
+#define        MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
+
 enum charge_type {
        MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
        MEM_CGROUP_CHARGE_TYPE_MAPPED,
@@ -259,6 +268,8 @@ enum charge_type {
 #define MEM_CGROUP_RECLAIM_NOSWAP      (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
 #define MEM_CGROUP_RECLAIM_SHRINK_BIT  0x1
 #define MEM_CGROUP_RECLAIM_SHRINK      (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
+#define MEM_CGROUP_RECLAIM_SOFT_BIT    0x2
+#define MEM_CGROUP_RECLAIM_SOFT                (1 << MEM_CGROUP_RECLAIM_SOFT_BIT)
 
 static void mem_cgroup_get(struct mem_cgroup *mem);
 static void mem_cgroup_put(struct mem_cgroup *mem);
@@ -299,7 +310,7 @@ soft_limit_tree_from_page(struct page *page)
 }
 
 static void
-mem_cgroup_insert_exceeded(struct mem_cgroup *mem,
+__mem_cgroup_insert_exceeded(struct mem_cgroup *mem,
                                struct mem_cgroup_per_zone *mz,
                                struct mem_cgroup_tree_per_zone *mctz)
 {
@@ -311,7 +322,6 @@ mem_cgroup_insert_exceeded(struct mem_cgroup *mem,
                return;
 
        mz->usage_in_excess = res_counter_soft_limit_excess(&mem->res);
-       spin_lock(&mctz->lock);
        while (*p) {
                parent = *p;
                mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
@@ -328,6 +338,26 @@ mem_cgroup_insert_exceeded(struct mem_cgroup *mem,
        rb_link_node(&mz->tree_node, parent, p);
        rb_insert_color(&mz->tree_node, &mctz->rb_root);
        mz->on_tree = true;
+}
+
+static void
+__mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
+                               struct mem_cgroup_per_zone *mz,
+                               struct mem_cgroup_tree_per_zone *mctz)
+{
+       if (!mz->on_tree)
+               return;
+       rb_erase(&mz->tree_node, &mctz->rb_root);
+       mz->on_tree = false;
+}
+
+static void
+mem_cgroup_insert_exceeded(struct mem_cgroup *mem,
+                               struct mem_cgroup_per_zone *mz,
+                               struct mem_cgroup_tree_per_zone *mctz)
+{
+       spin_lock(&mctz->lock);
+       __mem_cgroup_insert_exceeded(mem, mz, mctz);
        spin_unlock(&mctz->lock);
 }
 
@@ -337,8 +367,7 @@ mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
                                struct mem_cgroup_tree_per_zone *mctz)
 {
        spin_lock(&mctz->lock);
-       rb_erase(&mz->tree_node, &mctz->rb_root);
-       mz->on_tree = false;
+       __mem_cgroup_remove_exceeded(mem, mz, mctz);
        spin_unlock(&mctz->lock);
 }
 
@@ -408,6 +437,47 @@ static void mem_cgroup_remove_from_trees(struct mem_cgroup *mem)
        }
 }
 
+static inline unsigned long mem_cgroup_get_excess(struct mem_cgroup *mem)
+{
+       return res_counter_soft_limit_excess(&mem->res) >> PAGE_SHIFT;
+}
+
+static struct mem_cgroup_per_zone *
+__mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
+{
+       struct rb_node *rightmost = NULL;
+       struct mem_cgroup_per_zone *mz = NULL;
+
+retry:
+       rightmost = rb_last(&mctz->rb_root);
+       if (!rightmost)
+               goto done;              /* Nothing to reclaim from */
+
+       mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
+       /*
+        * Remove the node now but someone else can add it back,
+        * we will to add it back at the end of reclaim to its correct
+        * position in the tree.
+        */
+       __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
+       if (!res_counter_soft_limit_excess(&mz->mem->res) ||
+               !css_tryget(&mz->mem->css))
+               goto retry;
+done:
+       return mz;
+}
+
+static struct mem_cgroup_per_zone *
+mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
+{
+       struct mem_cgroup_per_zone *mz;
+
+       spin_lock(&mctz->lock);
+       mz = __mem_cgroup_largest_soft_limit_node(mctz);
+       spin_unlock(&mctz->lock);
+       return mz;
+}
+
 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
                                         struct page_cgroup *pc,
                                         bool charge)
@@ -1037,6 +1107,7 @@ mem_cgroup_select_victim(struct mem_cgroup *root_mem)
  * If shrink==true, for avoiding to free too much, this returns immedieately.
  */
 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
+                                               struct zone *zone,
                                                gfp_t gfp_mask,
                                                unsigned long reclaim_options)
 {
@@ -1045,23 +1116,53 @@ static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
        int loop = 0;
        bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
        bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
+       bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
+       unsigned long excess = mem_cgroup_get_excess(root_mem);
 
        /* If memsw_is_minimum==1, swap-out is of-no-use. */
        if (root_mem->memsw_is_minimum)
                noswap = true;
 
-       while (loop < 2) {
+       while (1) {
                victim = mem_cgroup_select_victim(root_mem);
-               if (victim == root_mem)
+               if (victim == root_mem) {
                        loop++;
+                       if (loop >= 2) {
+                               /*
+                                * If we have not been able to reclaim
+                                * anything, it might because there are
+                                * no reclaimable pages under this hierarchy
+                                */
+                               if (!check_soft || !total) {
+                                       css_put(&victim->css);
+                                       break;
+                               }
+                               /*
+                                * We want to do more targetted reclaim.
+                                * excess >> 2 is not to excessive so as to
+                                * reclaim too much, nor too less that we keep
+                                * coming back to reclaim from this cgroup
+                                */
+                               if (total >= (excess >> 2) ||
+                                       (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) {
+                                       css_put(&victim->css);
+                                       break;
+                               }
+                       }
+               }
                if (!mem_cgroup_local_usage(&victim->stat)) {
                        /* this cgroup's local usage == 0 */
                        css_put(&victim->css);
                        continue;
                }
                /* we use swappiness of local cgroup */
-               ret = try_to_free_mem_cgroup_pages(victim, gfp_mask, noswap,
-                                                  get_swappiness(victim));
+               if (check_soft)
+                       ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
+                               noswap, get_swappiness(victim), zone,
+                               zone->zone_pgdat->node_id);
+               else
+                       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
@@ -1071,7 +1172,10 @@ static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
                if (shrink)
                        return ret;
                total += ret;
-               if (mem_cgroup_check_under_limit(root_mem))
+               if (check_soft) {
+                       if (res_counter_check_under_soft_limit(&root_mem->res))
+                               return total;
+               } else if (mem_cgroup_check_under_limit(root_mem))
                        return 1 + total;
        }
        return total;
@@ -1206,8 +1310,8 @@ static int __mem_cgroup_try_charge(struct mm_struct *mm,
                if (!(gfp_mask & __GFP_WAIT))
                        goto nomem;
 
-               ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, gfp_mask,
-                                                       flags);
+               ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
+                                               gfp_mask, flags);
                if (ret)
                        continue;
 
@@ -2018,8 +2122,9 @@ static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
                if (!ret)
                        break;
 
-               progress = mem_cgroup_hierarchical_reclaim(memcg, GFP_KERNEL,
-                                                  MEM_CGROUP_RECLAIM_SHRINK);
+               progress = mem_cgroup_hierarchical_reclaim(memcg, NULL,
+                                               GFP_KERNEL,
+                                               MEM_CGROUP_RECLAIM_SHRINK);
                curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
                /* Usage is reduced ? */
                if (curusage >= oldusage)
@@ -2071,7 +2176,7 @@ static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
                if (!ret)
                        break;
 
-               mem_cgroup_hierarchical_reclaim(memcg, GFP_KERNEL,
+               mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
                                                MEM_CGROUP_RECLAIM_NOSWAP |
                                                MEM_CGROUP_RECLAIM_SHRINK);
                curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
@@ -2084,6 +2189,97 @@ static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
        return ret;
 }
 
+unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
+                                               gfp_t gfp_mask, int nid,
+                                               int zid)
+{
+       unsigned long nr_reclaimed = 0;
+       struct mem_cgroup_per_zone *mz, *next_mz = NULL;
+       unsigned long reclaimed;
+       int loop = 0;
+       struct mem_cgroup_tree_per_zone *mctz;
+
+       if (order > 0)
+               return 0;
+
+       mctz = soft_limit_tree_node_zone(nid, zid);
+       /*
+        * This loop can run a while, specially if mem_cgroup's continuously
+        * keep exceeding their soft limit and putting the system under
+        * pressure
+        */
+       do {
+               if (next_mz)
+                       mz = next_mz;
+               else
+                       mz = mem_cgroup_largest_soft_limit_node(mctz);
+               if (!mz)
+                       break;
+
+               reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
+                                               gfp_mask,
+                                               MEM_CGROUP_RECLAIM_SOFT);
+               nr_reclaimed += reclaimed;
+               spin_lock(&mctz->lock);
+
+               /*
+                * If we failed to reclaim anything from this memory cgroup
+                * it is time to move on to the next cgroup
+                */
+               next_mz = NULL;
+               if (!reclaimed) {
+                       do {
+                               /*
+                                * Loop until we find yet another one.
+                                *
+                                * By the time we get the soft_limit lock
+                                * again, someone might have aded the
+                                * group back on the RB tree. Iterate to
+                                * make sure we get a different mem.
+                                * mem_cgroup_largest_soft_limit_node returns
+                                * NULL if no other cgroup is present on
+                                * the tree
+                                */
+                               next_mz =
+                               __mem_cgroup_largest_soft_limit_node(mctz);
+                               if (next_mz == mz) {
+                                       css_put(&next_mz->mem->css);
+                                       next_mz = NULL;
+                               } else /* next_mz == NULL or other memcg */
+                                       break;
+                       } while (1);
+               }
+               mz->usage_in_excess =
+                       res_counter_soft_limit_excess(&mz->mem->res);
+               __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
+               /*
+                * One school of thought says that we should not add
+                * back the node to the tree if reclaim returns 0.
+                * But our reclaim could return 0, simply because due
+                * to priority we are exposing a smaller subset of
+                * memory to reclaim from. Consider this as a longer
+                * term TODO.
+                */
+               if (mz->usage_in_excess)
+                       __mem_cgroup_insert_exceeded(mz->mem, mz, mctz);
+               spin_unlock(&mctz->lock);
+               css_put(&mz->mem->css);
+               loop++;
+               /*
+                * Could not reclaim anything and there are no more
+                * mem cgroups to try or we seem to be looping without
+                * reclaiming anything.
+                */
+               if (!nr_reclaimed &&
+                       (next_mz == NULL ||
+                       loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
+                       break;
+       } while (!nr_reclaimed);
+       if (next_mz)
+               css_put(&next_mz->mem->css);
+       return nr_reclaimed;
+}
+
 /*
  * This routine traverse page_cgroup in given list and drop them all.
  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
@@ -2686,6 +2882,8 @@ static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
                for_each_lru(l)
                        INIT_LIST_HEAD(&mz->lists[l]);
                mz->usage_in_excess = 0;
+               mz->on_tree = false;
+               mz->mem = mem;
        }
        return 0;
 }
index 613e89f471d92c2710be84e9e0fdcf17b0976d0b..2423782214abf5da398d17c99d6605bfa265e329 100644 (file)
@@ -1836,11 +1836,45 @@ unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
 
 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
 
+unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
+                                               gfp_t gfp_mask, bool noswap,
+                                               unsigned int swappiness,
+                                               struct zone *zone, int nid)
+{
+       struct scan_control sc = {
+               .may_writepage = !laptop_mode,
+               .may_unmap = 1,
+               .may_swap = !noswap,
+               .swap_cluster_max = SWAP_CLUSTER_MAX,
+               .swappiness = swappiness,
+               .order = 0,
+               .mem_cgroup = mem,
+               .isolate_pages = mem_cgroup_isolate_pages,
+       };
+       nodemask_t nm  = nodemask_of_node(nid);
+
+       sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
+                       (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
+       sc.nodemask = &nm;
+       sc.nr_reclaimed = 0;
+       sc.nr_scanned = 0;
+       /*
+        * NOTE: Although we can get the priority field, using it
+        * here is not a good idea, since it limits the pages we can scan.
+        * if we don't reclaim here, the shrink_zone from balance_pgdat
+        * will pick up pages from other mem cgroup's as well. We hack
+        * the priority and make it zero.
+        */
+       shrink_zone(0, zone, &sc);
+       return sc.nr_reclaimed;
+}
+
 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
                                           gfp_t gfp_mask,
                                           bool noswap,
                                           unsigned int swappiness)
 {
+       struct zonelist *zonelist;
        struct scan_control sc = {
                .may_writepage = !laptop_mode,
                .may_unmap = 1,
@@ -1852,7 +1886,6 @@ unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
                .isolate_pages = mem_cgroup_isolate_pages,
                .nodemask = NULL, /* we don't care the placement */
        };
-       struct zonelist *zonelist;
 
        sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
                        (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
@@ -1974,6 +2007,7 @@ loop_again:
                for (i = 0; i <= end_zone; i++) {
                        struct zone *zone = pgdat->node_zones + i;
                        int nr_slab;
+                       int nid, zid;
 
                        if (!populated_zone(zone))
                                continue;
@@ -1988,6 +2022,15 @@ loop_again:
                        temp_priority[i] = priority;
                        sc.nr_scanned = 0;
                        note_zone_scanning_priority(zone, priority);
+
+                       nid = pgdat->node_id;
+                       zid = zone_idx(zone);
+                       /*
+                        * Call soft limit reclaim before calling shrink_zone.
+                        * For now we ignore the return value
+                        */
+                       mem_cgroup_soft_limit_reclaim(zone, order, sc.gfp_mask,
+                                                       nid, zid);
                        /*
                         * We put equal pressure on every zone, unless one
                         * zone has way too many pages free already.