memcg: simplify corner case handling of LRU.
authorKAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Fri, 13 Jan 2012 01:18:57 +0000 (17:18 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Fri, 13 Jan 2012 04:13:07 +0000 (20:13 -0800)
This patch simplifies LRU handling of racy case (memcg+SwapCache).  At
charging, SwapCache tend to be on LRU already.  So, before overwriting
pc->mem_cgroup, the page must be removed from LRU and added to LRU
later.

This patch does
        spin_lock(zone->lru_lock);
        if (PageLRU(page))
                remove from LRU
        overwrite pc->mem_cgroup
        if (PageLRU(page))
                add to new LRU.
        spin_unlock(zone->lru_lock);

And guarantee all pages are not on LRU at modifying pc->mem_cgroup.
This patch also unfies lru handling of replace_page_cache() and
swapin.

Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Miklos Szeredi <mszeredi@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.cz>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Ying Han <yinghan@google.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/memcontrol.c

index 71bac4d720d755b5f9f56a339ab7b623f41339a8..d58bb5fa4403861601b867747984362b4a80c028 100644 (file)
@@ -1136,86 +1136,6 @@ struct lruvec *mem_cgroup_lru_move_lists(struct zone *zone,
        return mem_cgroup_lru_add_list(zone, page, to);
 }
 
-/*
- * At handling SwapCache and other FUSE stuff, pc->mem_cgroup may be changed
- * while it's linked to lru because the page may be reused after it's fully
- * uncharged. To handle that, unlink page_cgroup from LRU when charge it again.
- * It's done under lock_page and expected that zone->lru_lock isnever held.
- */
-static void mem_cgroup_lru_del_before_commit(struct page *page)
-{
-       enum lru_list lru;
-       unsigned long flags;
-       struct zone *zone = page_zone(page);
-       struct page_cgroup *pc = lookup_page_cgroup(page);
-
-       /*
-        * Doing this check without taking ->lru_lock seems wrong but this
-        * is safe. Because if page_cgroup's USED bit is unset, the page
-        * will not be added to any memcg's LRU. If page_cgroup's USED bit is
-        * set, the commit after this will fail, anyway.
-        * This all charge/uncharge is done under some mutual execustion.
-        * So, we don't need to taking care of changes in USED bit.
-        */
-       if (likely(!PageLRU(page)))
-               return;
-
-       spin_lock_irqsave(&zone->lru_lock, flags);
-       lru = page_lru(page);
-       /*
-        * The uncharged page could still be registered to the LRU of
-        * the stale pc->mem_cgroup.
-        *
-        * As pc->mem_cgroup is about to get overwritten, the old LRU
-        * accounting needs to be taken care of.  Let root_mem_cgroup
-        * babysit the page until the new memcg is responsible for it.
-        *
-        * The PCG_USED bit is guarded by lock_page() as the page is
-        * swapcache/pagecache.
-        */
-       if (PageLRU(page) && PageCgroupAcctLRU(pc) && !PageCgroupUsed(pc)) {
-               del_page_from_lru_list(zone, page, lru);
-               add_page_to_lru_list(zone, page, lru);
-       }
-       spin_unlock_irqrestore(&zone->lru_lock, flags);
-}
-
-static void mem_cgroup_lru_add_after_commit(struct page *page)
-{
-       enum lru_list lru;
-       unsigned long flags;
-       struct zone *zone = page_zone(page);
-       struct page_cgroup *pc = lookup_page_cgroup(page);
-       /*
-        * putback:                             charge:
-        * SetPageLRU                           SetPageCgroupUsed
-        * smp_mb                               smp_mb
-        * PageCgroupUsed && add to memcg LRU   PageLRU && add to memcg LRU
-        *
-        * Ensure that one of the two sides adds the page to the memcg
-        * LRU during a race.
-        */
-       smp_mb();
-       /* taking care of that the page is added to LRU while we commit it */
-       if (likely(!PageLRU(page)))
-               return;
-       spin_lock_irqsave(&zone->lru_lock, flags);
-       lru = page_lru(page);
-       /*
-        * If the page is not on the LRU, someone will soon put it
-        * there.  If it is, and also already accounted for on the
-        * memcg-side, it must be on the right lruvec as setting
-        * pc->mem_cgroup and PageCgroupUsed is properly ordered.
-        * Otherwise, root_mem_cgroup has been babysitting the page
-        * during the charge.  Move it to the new memcg now.
-        */
-       if (PageLRU(page) && !PageCgroupAcctLRU(pc)) {
-               del_page_from_lru_list(zone, page, lru);
-               add_page_to_lru_list(zone, page, lru);
-       }
-       spin_unlock_irqrestore(&zone->lru_lock, flags);
-}
-
 /*
  * Checks whether given mem is same or in the root_mem_cgroup's
  * hierarchy subtree
@@ -2775,14 +2695,27 @@ __mem_cgroup_commit_charge_lrucare(struct page *page, struct mem_cgroup *memcg,
                                        enum charge_type ctype)
 {
        struct page_cgroup *pc = lookup_page_cgroup(page);
+       struct zone *zone = page_zone(page);
+       unsigned long flags;
+       bool removed = false;
+
        /*
         * In some case, SwapCache, FUSE(splice_buf->radixtree), the page
         * is already on LRU. It means the page may on some other page_cgroup's
         * LRU. Take care of it.
         */
-       mem_cgroup_lru_del_before_commit(page);
+       spin_lock_irqsave(&zone->lru_lock, flags);
+       if (PageLRU(page)) {
+               del_page_from_lru_list(zone, page, page_lru(page));
+               ClearPageLRU(page);
+               removed = true;
+       }
        __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype);
-       mem_cgroup_lru_add_after_commit(page);
+       if (removed) {
+               add_page_to_lru_list(zone, page, page_lru(page));
+               SetPageLRU(page);
+       }
+       spin_unlock_irqrestore(&zone->lru_lock, flags);
        return;
 }
 
@@ -3383,9 +3316,7 @@ void mem_cgroup_replace_page_cache(struct page *oldpage,
 {
        struct mem_cgroup *memcg;
        struct page_cgroup *pc;
-       struct zone *zone;
        enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
-       unsigned long flags;
 
        if (mem_cgroup_disabled())
                return;
@@ -3401,20 +3332,12 @@ void mem_cgroup_replace_page_cache(struct page *oldpage,
        if (PageSwapBacked(oldpage))
                type = MEM_CGROUP_CHARGE_TYPE_SHMEM;
 
-       zone = page_zone(newpage);
-       pc = lookup_page_cgroup(newpage);
        /*
         * Even if newpage->mapping was NULL before starting replacement,
         * the newpage may be on LRU(or pagevec for LRU) already. We lock
         * LRU while we overwrite pc->mem_cgroup.
         */
-       spin_lock_irqsave(&zone->lru_lock, flags);
-       if (PageLRU(newpage))
-               del_page_from_lru_list(zone, newpage, page_lru(newpage));
-       __mem_cgroup_commit_charge(memcg, newpage, 1, pc, type);
-       if (PageLRU(newpage))
-               add_page_to_lru_list(zone, newpage, page_lru(newpage));
-       spin_unlock_irqrestore(&zone->lru_lock, flags);
+       __mem_cgroup_commit_charge_lrucare(newpage, memcg, type);
 }
 
 #ifdef CONFIG_DEBUG_VM