memcg: better migration handling
authorKAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Fri, 25 Jul 2008 08:47:10 +0000 (01:47 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Fri, 25 Jul 2008 17:53:37 +0000 (10:53 -0700)
This patch changes page migration under memory controller to use a
different algorithm.  (thanks to Christoph for new idea.)

Before:
 - page_cgroup is migrated from an old page to a new page.
After:
 - a new page is accounted , no reuse of page_cgroup.

Pros:

 - We can avoid compliated lock depndencies and races in migration.

Cons:

 - new param to mem_cgroup_charge_common().

 - mem_cgroup_getref() is added for handling ref_cnt ping-pong.

This version simplifies complicated lock dependency in page migraiton
under memory resource controller.

  new refcnt sequence is following.

a mapped page:
  prepage_migration() ..... +1 to NEW page
  try_to_unmap()      ..... all refs to OLD page is gone.
  move_pages()        ..... +1 to NEW page if page cache.
  remap...            ..... all refs from *map* is added to NEW one.
  end_migration()     ..... -1 to New page.

  page's mapcount + (page_is_cache) refs are added to NEW one.

Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Balbir Singh <balbir@in.ibm.com>
Cc: Pavel Emelyanov <xemul@openvz.org>
Cc: Li Zefan <lizf@cn.fujitsu.com>
Cc: YAMAMOTO Takashi <yamamoto@valinux.co.jp>
Cc: Hugh Dickins <hugh@veritas.com>
Cc: Christoph Lameter <cl@linux-foundation.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
include/linux/memcontrol.h
mm/memcontrol.c
mm/migrate.c

index e6608776bc96138c4fac9eea0b2a4126997bc7b8..84ead2aa6f18d46d682d84169bcd105a0e6d03c9 100644 (file)
@@ -50,9 +50,10 @@ extern struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p);
 #define mm_match_cgroup(mm, cgroup)    \
        ((cgroup) == mem_cgroup_from_task((mm)->owner))
 
-extern int mem_cgroup_prepare_migration(struct page *page);
+extern int
+mem_cgroup_prepare_migration(struct page *page, struct page *newpage);
 extern void mem_cgroup_end_migration(struct page *page);
-extern void mem_cgroup_page_migration(struct page *page, struct page *newpage);
+extern int mem_cgroup_getref(struct page *page);
 
 /*
  * For memory reclaim.
@@ -112,7 +113,8 @@ static inline int task_in_mem_cgroup(struct task_struct *task,
        return 1;
 }
 
-static inline int mem_cgroup_prepare_migration(struct page *page)
+static inline int
+mem_cgroup_prepare_migration(struct page *page, struct page *newpage)
 {
        return 0;
 }
@@ -121,8 +123,7 @@ static inline void mem_cgroup_end_migration(struct page *page)
 {
 }
 
-static inline void
-mem_cgroup_page_migration(struct page *page, struct page *newpage)
+static inline void mem_cgroup_getref(struct page *page)
 {
 }
 
index 90ccc1326356a58e8589bed396e058ab5d201390..da5912b845511f8910957ac388f164baa07eade6 100644 (file)
@@ -524,7 +524,8 @@ unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
  * < 0 if the cgroup is over its limit
  */
 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
-                               gfp_t gfp_mask, enum charge_type ctype)
+                               gfp_t gfp_mask, enum charge_type ctype,
+                               struct mem_cgroup *memcg)
 {
        struct mem_cgroup *mem;
        struct page_cgroup *pc;
@@ -569,16 +570,21 @@ retry:
         * thread group leader migrates. It's possible that mm is not
         * set, if so charge the init_mm (happens for pagecache usage).
         */
-       if (!mm)
-               mm = &init_mm;
+       if (!memcg) {
+               if (!mm)
+                       mm = &init_mm;
 
-       rcu_read_lock();
-       mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
-       /*
-        * For every charge from the cgroup, increment reference count
-        */
-       css_get(&mem->css);
-       rcu_read_unlock();
+               rcu_read_lock();
+               mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
+               /*
+                * For every charge from the cgroup, increment reference count
+                */
+               css_get(&mem->css);
+               rcu_read_unlock();
+       } else {
+               mem = memcg;
+               css_get(&memcg->css);
+       }
 
        while (res_counter_charge(&mem->res, PAGE_SIZE)) {
                if (!(gfp_mask & __GFP_WAIT))
@@ -648,7 +654,7 @@ err:
 int mem_cgroup_charge(struct page *page, struct mm_struct *mm, gfp_t gfp_mask)
 {
        return mem_cgroup_charge_common(page, mm, gfp_mask,
-                               MEM_CGROUP_CHARGE_TYPE_MAPPED);
+                               MEM_CGROUP_CHARGE_TYPE_MAPPED, NULL);
 }
 
 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
@@ -657,7 +663,22 @@ int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
        if (!mm)
                mm = &init_mm;
        return mem_cgroup_charge_common(page, mm, gfp_mask,
-                               MEM_CGROUP_CHARGE_TYPE_CACHE);
+                               MEM_CGROUP_CHARGE_TYPE_CACHE, NULL);
+}
+
+int mem_cgroup_getref(struct page *page)
+{
+       struct page_cgroup *pc;
+
+       if (mem_cgroup_subsys.disabled)
+               return 0;
+
+       lock_page_cgroup(page);
+       pc = page_get_page_cgroup(page);
+       VM_BUG_ON(!pc);
+       pc->ref_cnt++;
+       unlock_page_cgroup(page);
+       return 0;
 }
 
 /*
@@ -707,65 +728,39 @@ unlock:
 }
 
 /*
- * Returns non-zero if a page (under migration) has valid page_cgroup member.
- * Refcnt of page_cgroup is incremented.
+ * Before starting migration, account against new page.
  */
-int mem_cgroup_prepare_migration(struct page *page)
+int mem_cgroup_prepare_migration(struct page *page, struct page *newpage)
 {
        struct page_cgroup *pc;
+       struct mem_cgroup *mem = NULL;
+       enum charge_type ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
+       int ret = 0;
 
        if (mem_cgroup_subsys.disabled)
                return 0;
 
        lock_page_cgroup(page);
        pc = page_get_page_cgroup(page);
-       if (pc)
-               pc->ref_cnt++;
+       if (pc) {
+               mem = pc->mem_cgroup;
+               css_get(&mem->css);
+               if (pc->flags & PAGE_CGROUP_FLAG_CACHE)
+                       ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
+       }
        unlock_page_cgroup(page);
-       return pc != NULL;
-}
-
-void mem_cgroup_end_migration(struct page *page)
-{
-       mem_cgroup_uncharge_page(page);
+       if (mem) {
+               ret = mem_cgroup_charge_common(newpage, NULL, GFP_KERNEL,
+                       ctype, mem);
+               css_put(&mem->css);
+       }
+       return ret;
 }
 
-/*
- * We know both *page* and *newpage* are now not-on-LRU and PG_locked.
- * And no race with uncharge() routines because page_cgroup for *page*
- * has extra one reference by mem_cgroup_prepare_migration.
- */
-void mem_cgroup_page_migration(struct page *page, struct page *newpage)
+/* remove redundant charge */
+void mem_cgroup_end_migration(struct page *newpage)
 {
-       struct page_cgroup *pc;
-       struct mem_cgroup_per_zone *mz;
-       unsigned long flags;
-
-       lock_page_cgroup(page);
-       pc = page_get_page_cgroup(page);
-       if (!pc) {
-               unlock_page_cgroup(page);
-               return;
-       }
-
-       mz = page_cgroup_zoneinfo(pc);
-       spin_lock_irqsave(&mz->lru_lock, flags);
-       __mem_cgroup_remove_list(mz, pc);
-       spin_unlock_irqrestore(&mz->lru_lock, flags);
-
-       page_assign_page_cgroup(page, NULL);
-       unlock_page_cgroup(page);
-
-       pc->page = newpage;
-       lock_page_cgroup(newpage);
-       page_assign_page_cgroup(newpage, pc);
-
-       mz = page_cgroup_zoneinfo(pc);
-       spin_lock_irqsave(&mz->lru_lock, flags);
-       __mem_cgroup_add_list(mz, pc);
-       spin_unlock_irqrestore(&mz->lru_lock, flags);
-
-       unlock_page_cgroup(newpage);
+       mem_cgroup_uncharge_page(newpage);
 }
 
 /*
@@ -795,12 +790,19 @@ static void mem_cgroup_force_empty_list(struct mem_cgroup *mem,
                page = pc->page;
                get_page(page);
                spin_unlock_irqrestore(&mz->lru_lock, flags);
-               mem_cgroup_uncharge_page(page);
-               put_page(page);
-               if (--count <= 0) {
-                       count = FORCE_UNCHARGE_BATCH;
+               /*
+                * Check if this page is on LRU. !LRU page can be found
+                * if it's under page migration.
+                */
+               if (PageLRU(page)) {
+                       mem_cgroup_uncharge_page(page);
+                       put_page(page);
+                       if (--count <= 0) {
+                               count = FORCE_UNCHARGE_BATCH;
+                               cond_resched();
+                       }
+               } else
                        cond_resched();
-               }
                spin_lock_irqsave(&mz->lru_lock, flags);
        }
        spin_unlock_irqrestore(&mz->lru_lock, flags);
index 376cceba82f98e7110520acc885f6bce0687482b..f6d7f8efd1a84b5fb1729013890ecaa3d6caa609 100644 (file)
@@ -358,6 +358,10 @@ static int migrate_page_move_mapping(struct address_space *mapping,
        __inc_zone_page_state(newpage, NR_FILE_PAGES);
 
        write_unlock_irq(&mapping->tree_lock);
+       if (!PageSwapCache(newpage)) {
+               mem_cgroup_uncharge_page(page);
+               mem_cgroup_getref(newpage);
+       }
 
        return 0;
 }
@@ -611,7 +615,6 @@ static int move_to_new_page(struct page *newpage, struct page *page)
                rc = fallback_migrate_page(mapping, newpage, page);
 
        if (!rc) {
-               mem_cgroup_page_migration(page, newpage);
                remove_migration_ptes(page, newpage);
        } else
                newpage->mapping = NULL;
@@ -641,6 +644,14 @@ static int unmap_and_move(new_page_t get_new_page, unsigned long private,
                /* page was freed from under us. So we are done. */
                goto move_newpage;
 
+       charge = mem_cgroup_prepare_migration(page, newpage);
+       if (charge == -ENOMEM) {
+               rc = -ENOMEM;
+               goto move_newpage;
+       }
+       /* prepare cgroup just returns 0 or -ENOMEM */
+       BUG_ON(charge);
+
        rc = -EAGAIN;
        if (TestSetPageLocked(page)) {
                if (!force)
@@ -692,19 +703,14 @@ static int unmap_and_move(new_page_t get_new_page, unsigned long private,
                goto rcu_unlock;
        }
 
-       charge = mem_cgroup_prepare_migration(page);
        /* Establish migration ptes or remove ptes */
        try_to_unmap(page, 1);
 
        if (!page_mapped(page))
                rc = move_to_new_page(newpage, page);
 
-       if (rc) {
+       if (rc)
                remove_migration_ptes(page, page);
-               if (charge)
-                       mem_cgroup_end_migration(page);
-       } else if (charge)
-               mem_cgroup_end_migration(newpage);
 rcu_unlock:
        if (rcu_locked)
                rcu_read_unlock();
@@ -725,6 +731,8 @@ unlock:
        }
 
 move_newpage:
+       if (!charge)
+               mem_cgroup_end_migration(newpage);
        /*
         * Move the new page to the LRU. If migration was not successful
         * then this will free the page.