memcg: coalesce charging via percpu storage
authorKAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Wed, 16 Dec 2009 00:47:08 +0000 (16:47 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Wed, 16 Dec 2009 15:20:07 +0000 (07:20 -0800)
This is a patch for coalescing access to res_counter at charging by percpu
caching.  At charge, memcg charges 64pages and remember it in percpu
cache.  Because it's cache, drain/flush if necessary.

This version uses public percpu area.
 2 benefits for using public percpu area.
 1. Sum of stocked charge in the system is limited to # of cpus
    not to the number of memcg. This shows better synchonization.
 2. drain code for flush/cpuhotplug is very easy (and quick)

The most important point of this patch is that we never touch res_counter
in fast path. The res_counter is system-wide shared counter which is modified
very frequently. We shouldn't touch it as far as we can for avoiding
false sharing.

On x86-64 8cpu server, I tested overheads of memcg at page fault by
running a program which does map/fault/unmap in a loop. Running
a task per a cpu by taskset and see sum of the number of page faults
in 60secs.

[without memcg config]
  40156968  page-faults              #      0.085 M/sec   ( +-   0.046% )
  27.67 cache-miss/faults

[root cgroup]
  36659599  page-faults              #      0.077 M/sec   ( +-   0.247% )
  31.58 cache miss/faults

[in a child cgroup]
  18444157  page-faults              #      0.039 M/sec   ( +-   0.133% )
  69.96 cache miss/faults

[ + coalescing uncharge patch]
  27133719  page-faults              #      0.057 M/sec   ( +-   0.155% )
  47.16 cache miss/faults

[ + coalescing uncharge patch + this patch ]
  34224709  page-faults              #      0.072 M/sec   ( +-   0.173% )
  34.69 cache miss/faults

Changelog (since Oct/2):
  - updated comments
  - replaced get_cpu_var() with __get_cpu_var() if possible.
  - removed mutex for system-wide drain. adds a counter instead of it.
  - removed CONFIG_HOTPLUG_CPU

Changelog (old):
  - rebased onto the latest mmotm
  - moved charge size check before __GFP_WAIT check for avoiding unnecesary
  - added asynchronous flush routine.
  - fixed bugs pointed out by Nishimura-san.

[akpm@linux-foundation.org: tweak comments]
[nishimura@mxp.nes.nec.co.jp: don't do INIT_WORK() repeatedly against the same work_struct]
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Balbir Singh <balbir@in.ibm.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>
mm/memcontrol.c

index a730c91..6587f65 100644 (file)
@@ -38,6 +38,7 @@
 #include <linux/vmalloc.h>
 #include <linux/mm_inline.h>
 #include <linux/page_cgroup.h>
+#include <linux/cpu.h>
 #include "internal.h"
 
 #include <asm/uaccess.h>
@@ -275,6 +276,7 @@ enum charge_type {
 static void mem_cgroup_get(struct mem_cgroup *mem);
 static void mem_cgroup_put(struct mem_cgroup *mem);
 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem);
+static void drain_all_stock_async(void);
 
 static struct mem_cgroup_per_zone *
 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
@@ -1137,6 +1139,8 @@ static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
                victim = mem_cgroup_select_victim(root_mem);
                if (victim == root_mem) {
                        loop++;
+                       if (loop >= 1)
+                               drain_all_stock_async();
                        if (loop >= 2) {
                                /*
                                 * If we have not been able to reclaim
@@ -1259,6 +1263,133 @@ done:
 }
 
 /*
+ * size of first charge trial. "32" comes from vmscan.c's magic value.
+ * TODO: maybe necessary to use big numbers in big irons.
+ */
+#define CHARGE_SIZE    (32 * PAGE_SIZE)
+struct memcg_stock_pcp {
+       struct mem_cgroup *cached; /* this never be root cgroup */
+       int charge;
+       struct work_struct work;
+};
+static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
+static atomic_t memcg_drain_count;
+
+/*
+ * Try to consume stocked charge on this cpu. If success, PAGE_SIZE is consumed
+ * from local stock and true is returned. If the stock is 0 or charges from a
+ * cgroup which is not current target, returns false. This stock will be
+ * refilled.
+ */
+static bool consume_stock(struct mem_cgroup *mem)
+{
+       struct memcg_stock_pcp *stock;
+       bool ret = true;
+
+       stock = &get_cpu_var(memcg_stock);
+       if (mem == stock->cached && stock->charge)
+               stock->charge -= PAGE_SIZE;
+       else /* need to call res_counter_charge */
+               ret = false;
+       put_cpu_var(memcg_stock);
+       return ret;
+}
+
+/*
+ * Returns stocks cached in percpu to res_counter and reset cached information.
+ */
+static void drain_stock(struct memcg_stock_pcp *stock)
+{
+       struct mem_cgroup *old = stock->cached;
+
+       if (stock->charge) {
+               res_counter_uncharge(&old->res, stock->charge);
+               if (do_swap_account)
+                       res_counter_uncharge(&old->memsw, stock->charge);
+       }
+       stock->cached = NULL;
+       stock->charge = 0;
+}
+
+/*
+ * This must be called under preempt disabled or must be called by
+ * a thread which is pinned to local cpu.
+ */
+static void drain_local_stock(struct work_struct *dummy)
+{
+       struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
+       drain_stock(stock);
+}
+
+/*
+ * Cache charges(val) which is from res_counter, to local per_cpu area.
+ * This will be consumed by consumt_stock() function, later.
+ */
+static void refill_stock(struct mem_cgroup *mem, int val)
+{
+       struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
+
+       if (stock->cached != mem) { /* reset if necessary */
+               drain_stock(stock);
+               stock->cached = mem;
+       }
+       stock->charge += val;
+       put_cpu_var(memcg_stock);
+}
+
+/*
+ * Tries to drain stocked charges in other cpus. This function is asynchronous
+ * and just put a work per cpu for draining localy on each cpu. Caller can
+ * expects some charges will be back to res_counter later but cannot wait for
+ * it.
+ */
+static void drain_all_stock_async(void)
+{
+       int cpu;
+       /* This function is for scheduling "drain" in asynchronous way.
+        * The result of "drain" is not directly handled by callers. Then,
+        * if someone is calling drain, we don't have to call drain more.
+        * Anyway, WORK_STRUCT_PENDING check in queue_work_on() will catch if
+        * there is a race. We just do loose check here.
+        */
+       if (atomic_read(&memcg_drain_count))
+               return;
+       /* Notify other cpus that system-wide "drain" is running */
+       atomic_inc(&memcg_drain_count);
+       get_online_cpus();
+       for_each_online_cpu(cpu) {
+               struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
+               schedule_work_on(cpu, &stock->work);
+       }
+       put_online_cpus();
+       atomic_dec(&memcg_drain_count);
+       /* We don't wait for flush_work */
+}
+
+/* This is a synchronous drain interface. */
+static void drain_all_stock_sync(void)
+{
+       /* called when force_empty is called */
+       atomic_inc(&memcg_drain_count);
+       schedule_on_each_cpu(drain_local_stock);
+       atomic_dec(&memcg_drain_count);
+}
+
+static int __cpuinit memcg_stock_cpu_callback(struct notifier_block *nb,
+                                       unsigned long action,
+                                       void *hcpu)
+{
+       int cpu = (unsigned long)hcpu;
+       struct memcg_stock_pcp *stock;
+
+       if (action != CPU_DEAD)
+               return NOTIFY_OK;
+       stock = &per_cpu(memcg_stock, cpu);
+       drain_stock(stock);
+       return NOTIFY_OK;
+}
+
+/*
  * Unlike exported interface, "oom" parameter is added. if oom==true,
  * oom-killer can be invoked.
  */
@@ -1269,6 +1400,7 @@ static int __mem_cgroup_try_charge(struct mm_struct *mm,
        struct mem_cgroup *mem, *mem_over_limit;
        int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
        struct res_counter *fail_res;
+       int csize = CHARGE_SIZE;
 
        if (unlikely(test_thread_flag(TIF_MEMDIE))) {
                /* Don't account this! */
@@ -1293,23 +1425,25 @@ static int __mem_cgroup_try_charge(struct mm_struct *mm,
                return 0;
 
        VM_BUG_ON(css_is_removed(&mem->css));
+       if (mem_cgroup_is_root(mem))
+               goto done;
 
        while (1) {
                int ret = 0;
                unsigned long flags = 0;
 
-               if (mem_cgroup_is_root(mem))
-                       goto done;
-               ret = res_counter_charge(&mem->res, PAGE_SIZE, &fail_res);
+               if (consume_stock(mem))
+                       goto charged;
+
+               ret = res_counter_charge(&mem->res, csize, &fail_res);
                if (likely(!ret)) {
                        if (!do_swap_account)
                                break;
-                       ret = res_counter_charge(&mem->memsw, PAGE_SIZE,
-                                                       &fail_res);
+                       ret = res_counter_charge(&mem->memsw, csize, &fail_res);
                        if (likely(!ret))
                                break;
                        /* mem+swap counter fails */
-                       res_counter_uncharge(&mem->res, PAGE_SIZE);
+                       res_counter_uncharge(&mem->res, csize);
                        flags |= MEM_CGROUP_RECLAIM_NOSWAP;
                        mem_over_limit = mem_cgroup_from_res_counter(fail_res,
                                                                        memsw);
@@ -1318,6 +1452,11 @@ static int __mem_cgroup_try_charge(struct mm_struct *mm,
                        mem_over_limit = mem_cgroup_from_res_counter(fail_res,
                                                                        res);
 
+               /* reduce request size and retry */
+               if (csize > PAGE_SIZE) {
+                       csize = PAGE_SIZE;
+                       continue;
+               }
                if (!(gfp_mask & __GFP_WAIT))
                        goto nomem;
 
@@ -1347,6 +1486,9 @@ static int __mem_cgroup_try_charge(struct mm_struct *mm,
                        goto nomem;
                }
        }
+       if (csize > PAGE_SIZE)
+               refill_stock(mem, csize - PAGE_SIZE);
+charged:
        /*
         * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
         * if they exceeds softlimit.
@@ -2469,6 +2611,7 @@ move_account:
                        goto out;
                /* This is for making all *used* pages to be on LRU. */
                lru_add_drain_all();
+               drain_all_stock_sync();
                ret = 0;
                for_each_node_state(node, N_HIGH_MEMORY) {
                        for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
@@ -3183,11 +3326,18 @@ mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
 
        /* root ? */
        if (cont->parent == NULL) {
+               int cpu;
                enable_swap_cgroup();
                parent = NULL;
                root_mem_cgroup = mem;
                if (mem_cgroup_soft_limit_tree_init())
                        goto free_out;
+               for_each_possible_cpu(cpu) {
+                       struct memcg_stock_pcp *stock =
+                                               &per_cpu(memcg_stock, cpu);
+                       INIT_WORK(&stock->work, drain_local_stock);
+               }
+               hotcpu_notifier(memcg_stock_cpu_callback, 0);
 
        } else {
                parent = mem_cgroup_from_cont(cont->parent);