memcg: avoid css_get()
authorKAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Wed, 11 Aug 2010 01:03:02 +0000 (18:03 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Wed, 11 Aug 2010 15:59:19 +0000 (08:59 -0700)
Now, memory cgroup increments css(cgroup subsys state)'s reference count
per a charged page.  And the reference count is kept until the page is
uncharged.  But this has 2 bad effect.

 1. Because css_get/put calls atomic_inc()/dec, heavy call of them
    on large smp will not scale well.
 2. Because css's refcnt cannot be in a state as "ready-to-release",
    cgroup's notify_on_release handler can't work with memcg.
 3. css's refcnt is atomic_t, it means smaller than 32bit. Maybe too small.

This has been a problem since the 1st merge of memcg.

This is a trial to remove css's refcnt per a page. Even if we remove
refcnt, pre_destroy() does enough synchronization as
  - check res->usage == 0.
  - check no pages on LRU.

This patch removes css's refcnt per page.  Even after this patch, at the
1st look, it seems css_get() is still called in try_charge().

But the logic is.

  - If a memcg of mm->owner is cached one, consume_stock() will work.
    At success, return immediately.
  - If consume_stock returns false, css_get() is called and go to
    slow path which may be blocked. At the end of slow path,
    css_put() is called and restart from the start if necessary.

So, in the fast path, we don't call css_get() and can avoid access to
shared counter. This patch can make the most possible case fast.

Here is a result of multi-threaded page fault benchmark.

[Before]
    25.32%  multi-fault-all  [kernel.kallsyms]      [k] clear_page_c
     9.30%  multi-fault-all  [kernel.kallsyms]      [k] _raw_spin_lock_irqsave
     8.02%  multi-fault-all  [kernel.kallsyms]      [k] try_get_mem_cgroup_from_mm <=====(*)
     7.83%  multi-fault-all  [kernel.kallsyms]      [k] down_read_trylock
     5.38%  multi-fault-all  [kernel.kallsyms]      [k] __css_put
     5.29%  multi-fault-all  [kernel.kallsyms]      [k] __alloc_pages_nodemask
     4.92%  multi-fault-all  [kernel.kallsyms]      [k] _raw_spin_lock_irq
     4.24%  multi-fault-all  [kernel.kallsyms]      [k] up_read
     3.53%  multi-fault-all  [kernel.kallsyms]      [k] css_put
     2.11%  multi-fault-all  [kernel.kallsyms]      [k] handle_mm_fault
     1.76%  multi-fault-all  [kernel.kallsyms]      [k] __rmqueue
     1.64%  multi-fault-all  [kernel.kallsyms]      [k] __mem_cgroup_commit_charge

[After]
    28.41%  multi-fault-all  [kernel.kallsyms]      [k] clear_page_c
    10.08%  multi-fault-all  [kernel.kallsyms]      [k] _raw_spin_lock_irq
     9.58%  multi-fault-all  [kernel.kallsyms]      [k] down_read_trylock
     9.38%  multi-fault-all  [kernel.kallsyms]      [k] _raw_spin_lock_irqsave
     5.86%  multi-fault-all  [kernel.kallsyms]      [k] __alloc_pages_nodemask
     5.65%  multi-fault-all  [kernel.kallsyms]      [k] up_read
     2.82%  multi-fault-all  [kernel.kallsyms]      [k] handle_mm_fault
     2.64%  multi-fault-all  [kernel.kallsyms]      [k] mem_cgroup_add_lru_list
     2.48%  multi-fault-all  [kernel.kallsyms]      [k] __mem_cgroup_commit_charge

Then, 8.02% of try_get_mem_cgroup_from_mm() disappears because this patch
removes css_tryget() in it. (But yes, this is an extreme case.)

Signed-off-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Cc: Balbir Singh <balbir@in.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/memcontrol.c

index f52b0a1861c4624b7e99879e4aa538872604c0ac..a1c3c317a4dd92fa34314b72c2bf07e634530e5e 100644 (file)
@@ -1714,28 +1714,66 @@ static int __mem_cgroup_try_charge(struct mm_struct *mm,
         * thread group leader migrates. It's possible that mm is not
         * set, if so charge the init_mm (happens for pagecache usage).
         */
-       if (*memcg) {
+       if (!*memcg && !mm)
+               goto bypass;
+again:
+       if (*memcg) { /* css should be a valid one */
                mem = *memcg;
+               VM_BUG_ON(css_is_removed(&mem->css));
+               if (mem_cgroup_is_root(mem))
+                       goto done;
+               if (consume_stock(mem))
+                       goto done;
                css_get(&mem->css);
        } else {
-               mem = try_get_mem_cgroup_from_mm(mm);
-               if (unlikely(!mem))
-                       return 0;
-               *memcg = mem;
-       }
+               struct task_struct *p;
 
-       VM_BUG_ON(css_is_removed(&mem->css));
-       if (mem_cgroup_is_root(mem))
-               goto done;
+               rcu_read_lock();
+               p = rcu_dereference(mm->owner);
+               VM_BUG_ON(!p);
+               /*
+                * because we don't have task_lock(), "p" can exit while
+                * we're here. In that case, "mem" can point to root
+                * cgroup but never be NULL. (and task_struct itself is freed
+                * by RCU, cgroup itself is RCU safe.) Then, we have small
+                * risk here to get wrong cgroup. But such kind of mis-account
+                * by race always happens because we don't have cgroup_mutex().
+                * It's overkill and we allow that small race, here.
+                */
+               mem = mem_cgroup_from_task(p);
+               VM_BUG_ON(!mem);
+               if (mem_cgroup_is_root(mem)) {
+                       rcu_read_unlock();
+                       goto done;
+               }
+               if (consume_stock(mem)) {
+                       /*
+                        * It seems dagerous to access memcg without css_get().
+                        * But considering how consume_stok works, it's not
+                        * necessary. If consume_stock success, some charges
+                        * from this memcg are cached on this cpu. So, we
+                        * don't need to call css_get()/css_tryget() before
+                        * calling consume_stock().
+                        */
+                       rcu_read_unlock();
+                       goto done;
+               }
+               /* after here, we may be blocked. we need to get refcnt */
+               if (!css_tryget(&mem->css)) {
+                       rcu_read_unlock();
+                       goto again;
+               }
+               rcu_read_unlock();
+       }
 
        do {
                bool oom_check;
 
-               if (consume_stock(mem))
-                       goto done; /* don't need to fill stock */
                /* If killed, bypass charge */
-               if (fatal_signal_pending(current))
+               if (fatal_signal_pending(current)) {
+                       css_put(&mem->css);
                        goto bypass;
+               }
 
                oom_check = false;
                if (oom && !nr_oom_retries) {
@@ -1750,30 +1788,36 @@ static int __mem_cgroup_try_charge(struct mm_struct *mm,
                        break;
                case CHARGE_RETRY: /* not in OOM situation but retry */
                        csize = PAGE_SIZE;
-                       break;
+                       css_put(&mem->css);
+                       mem = NULL;
+                       goto again;
                case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
+                       css_put(&mem->css);
                        goto nomem;
                case CHARGE_NOMEM: /* OOM routine works */
-                       if (!oom)
+                       if (!oom) {
+                               css_put(&mem->css);
                                goto nomem;
+                       }
                        /* If oom, we never return -ENOMEM */
                        nr_oom_retries--;
                        break;
                case CHARGE_OOM_DIE: /* Killed by OOM Killer */
+                       css_put(&mem->css);
                        goto bypass;
                }
        } while (ret != CHARGE_OK);
 
        if (csize > PAGE_SIZE)
                refill_stock(mem, csize - PAGE_SIZE);
+       css_put(&mem->css);
 done:
+       *memcg = mem;
        return 0;
 nomem:
-       css_put(&mem->css);
+       *memcg = NULL;
        return -ENOMEM;
 bypass:
-       if (mem)
-               css_put(&mem->css);
        *memcg = NULL;
        return 0;
 }
@@ -1790,11 +1834,7 @@ static void __mem_cgroup_cancel_charge(struct mem_cgroup *mem,
                res_counter_uncharge(&mem->res, PAGE_SIZE * count);
                if (do_swap_account)
                        res_counter_uncharge(&mem->memsw, PAGE_SIZE * count);
-               VM_BUG_ON(test_bit(CSS_ROOT, &mem->css.flags));
-               WARN_ON_ONCE(count > INT_MAX);
-               __css_put(&mem->css, (int)count);
        }
-       /* we don't need css_put for root */
 }
 
 static void mem_cgroup_cancel_charge(struct mem_cgroup *mem)
@@ -2155,7 +2195,6 @@ int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
                goto charge_cur_mm;
        *ptr = mem;
        ret = __mem_cgroup_try_charge(NULL, mask, ptr, true);
-       /* drop extra refcnt from tryget */
        css_put(&mem->css);
        return ret;
 charge_cur_mm:
@@ -2325,10 +2364,6 @@ __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
                break;
        }
 
-       if (!mem_cgroup_is_root(mem))
-               __do_uncharge(mem, ctype);
-       if (ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
-               mem_cgroup_swap_statistics(mem, true);
        mem_cgroup_charge_statistics(mem, pc, false);
 
        ClearPageCgroupUsed(pc);
@@ -2340,11 +2375,17 @@ __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
         */
 
        unlock_page_cgroup(pc);
-
+       /*
+        * even after unlock, we have mem->res.usage here and this memcg
+        * will never be freed.
+        */
        memcg_check_events(mem, page);
-       /* at swapout, this memcg will be accessed to record to swap */
-       if (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
-               css_put(&mem->css);
+       if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
+               mem_cgroup_swap_statistics(mem, true);
+               mem_cgroup_get(mem);
+       }
+       if (!mem_cgroup_is_root(mem))
+               __do_uncharge(mem, ctype);
 
        return mem;
 
@@ -2431,13 +2472,12 @@ mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
 
        memcg = __mem_cgroup_uncharge_common(page, ctype);
 
-       /* record memcg information */
-       if (do_swap_account && swapout && memcg) {
+       /*
+        * record memcg information,  if swapout && memcg != NULL,
+        * mem_cgroup_get() was called in uncharge().
+        */
+       if (do_swap_account && swapout && memcg)
                swap_cgroup_record(ent, css_id(&memcg->css));
-               mem_cgroup_get(memcg);
-       }
-       if (swapout && memcg)
-               css_put(&memcg->css);
 }
 #endif
 
@@ -2515,7 +2555,6 @@ static int mem_cgroup_move_swap_account(swp_entry_t entry,
                         */
                        if (!mem_cgroup_is_root(to))
                                res_counter_uncharge(&to->res, PAGE_SIZE);
-                       css_put(&to->css);
                }
                return 0;
        }
@@ -4214,9 +4253,6 @@ static int mem_cgroup_do_precharge(unsigned long count)
                        goto one_by_one;
                }
                mc.precharge += count;
-               VM_BUG_ON(test_bit(CSS_ROOT, &mem->css.flags));
-               WARN_ON_ONCE(count > INT_MAX);
-               __css_get(&mem->css, (int)count);
                return ret;
        }
 one_by_one:
@@ -4452,7 +4488,6 @@ static void mem_cgroup_clear_mc(void)
        }
        /* we must fixup refcnts and charges */
        if (mc.moved_swap) {
-               WARN_ON_ONCE(mc.moved_swap > INT_MAX);
                /* uncharge swap account from the old cgroup */
                if (!mem_cgroup_is_root(mc.from))
                        res_counter_uncharge(&mc.from->memsw,
@@ -4466,8 +4501,6 @@ static void mem_cgroup_clear_mc(void)
                         */
                        res_counter_uncharge(&mc.to->res,
                                                PAGE_SIZE * mc.moved_swap);
-                       VM_BUG_ON(test_bit(CSS_ROOT, &mc.to->css.flags));
-                       __css_put(&mc.to->css, mc.moved_swap);
                }
                /* we've already done mem_cgroup_get(mc.to) */