memcg: allocate all page_cgroup at boot
KAMEZAWA Hiroyuki [Sun, 19 Oct 2008 03:28:16 +0000 (20:28 -0700)]
Allocate all page_cgroup at boot and remove page_cgroup poitner from
struct page.  This patch adds an interface as

 struct page_cgroup *lookup_page_cgroup(struct page*)

All FLATMEM/DISCONTIGMEM/SPARSEMEM  and MEMORY_HOTPLUG is supported.

Remove page_cgroup pointer reduces the amount of memory by
 - 4 bytes per PAGE_SIZE.
 - 8 bytes per PAGE_SIZE
if memory controller is disabled. (even if configured.)

On usual 8GB x86-32 server, this saves 8MB of NORMAL_ZONE memory.
On my x86-64 server with 48GB of memory, this saves 96MB of memory.
I think this reduction makes sense.

By pre-allocation, kmalloc/kfree in charge/uncharge are removed.
This means
  - we're not necessary to be afraid of kmalloc faiulre.
    (this can happen because of gfp_mask type.)
  - we can avoid calling kmalloc/kfree.
  - we can avoid allocating tons of small objects which can be fragmented.
  - we can know what amount of memory will be used for this extra-lru handling.

I added printk message as

"allocated %ld bytes of page_cgroup"
        "please try cgroup_disable=memory option if you don't want"

maybe enough informative for users.

Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Reviewed-by: Balbir Singh <balbir@linux.vnet.ibm.com>
Cc: 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/mm_types.h
include/linux/mmzone.h
include/linux/page_cgroup.h [new file with mode: 0644]
mm/Makefile
mm/memcontrol.c
mm/page_alloc.c
mm/page_cgroup.c [new file with mode: 0644]

index ee1b2fc..1fbe14d 100644 (file)
@@ -27,9 +27,6 @@ struct mm_struct;
 
 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
 
-#define page_reset_bad_cgroup(page)    ((page)->page_cgroup = 0)
-
-extern struct page_cgroup *page_get_page_cgroup(struct page *page);
 extern int mem_cgroup_charge(struct page *page, struct mm_struct *mm,
                                gfp_t gfp_mask);
 extern int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
@@ -72,16 +69,8 @@ extern void mem_cgroup_record_reclaim_priority(struct mem_cgroup *mem,
 extern long mem_cgroup_calc_reclaim(struct mem_cgroup *mem, struct zone *zone,
                                        int priority, enum lru_list lru);
 
-#else /* CONFIG_CGROUP_MEM_RES_CTLR */
-static inline void page_reset_bad_cgroup(struct page *page)
-{
-}
-
-static inline struct page_cgroup *page_get_page_cgroup(struct page *page)
-{
-       return NULL;
-}
 
+#else /* CONFIG_CGROUP_MEM_RES_CTLR */
 static inline int mem_cgroup_charge(struct page *page,
                                        struct mm_struct *mm, gfp_t gfp_mask)
 {
index 9d49fa3..fe82547 100644 (file)
@@ -94,9 +94,6 @@ struct page {
        void *virtual;                  /* Kernel virtual address (NULL if
                                           not kmapped, ie. highmem) */
 #endif /* WANT_PAGE_VIRTUAL */
-#ifdef CONFIG_CGROUP_MEM_RES_CTLR
-       unsigned long page_cgroup;
-#endif
 };
 
 /*
index da2d053..35a7b5e 100644 (file)
@@ -601,8 +601,11 @@ typedef struct pglist_data {
        struct zone node_zones[MAX_NR_ZONES];
        struct zonelist node_zonelists[MAX_ZONELISTS];
        int nr_zones;
-#ifdef CONFIG_FLAT_NODE_MEM_MAP
+#ifdef CONFIG_FLAT_NODE_MEM_MAP        /* means !SPARSEMEM */
        struct page *node_mem_map;
+#ifdef CONFIG_CGROUP_MEM_RES_CTLR
+       struct page_cgroup *node_page_cgroup;
+#endif
 #endif
        struct bootmem_data *bdata;
 #ifdef CONFIG_MEMORY_HOTPLUG
@@ -931,6 +934,7 @@ static inline unsigned long early_pfn_to_nid(unsigned long pfn)
 #endif
 
 struct page;
+struct page_cgroup;
 struct mem_section {
        /*
         * This is, logically, a pointer to an array of struct
@@ -948,6 +952,14 @@ struct mem_section {
 
        /* See declaration of similar field in struct zone */
        unsigned long *pageblock_flags;
+#ifdef CONFIG_CGROUP_MEM_RES_CTLR
+       /*
+        * If !SPARSEMEM, pgdat doesn't have page_cgroup pointer. We use
+        * section. (see memcontrol.h/page_cgroup.h about this.)
+        */
+       struct page_cgroup *page_cgroup;
+       unsigned long pad;
+#endif
 };
 
 #ifdef CONFIG_SPARSEMEM_EXTREME
diff --git a/include/linux/page_cgroup.h b/include/linux/page_cgroup.h
new file mode 100644 (file)
index 0000000..0fd39f2
--- /dev/null
@@ -0,0 +1,103 @@
+#ifndef __LINUX_PAGE_CGROUP_H
+#define __LINUX_PAGE_CGROUP_H
+
+#ifdef CONFIG_CGROUP_MEM_RES_CTLR
+#include <linux/bit_spinlock.h>
+/*
+ * Page Cgroup can be considered as an extended mem_map.
+ * A page_cgroup page is associated with every page descriptor. The
+ * page_cgroup helps us identify information about the cgroup
+ * All page cgroups are allocated at boot or memory hotplug event,
+ * then the page cgroup for pfn always exists.
+ */
+struct page_cgroup {
+       unsigned long flags;
+       struct mem_cgroup *mem_cgroup;
+       struct page *page;
+       struct list_head lru;           /* per cgroup LRU list */
+};
+
+void __init pgdat_page_cgroup_init(struct pglist_data *pgdat);
+void __init page_cgroup_init(void);
+struct page_cgroup *lookup_page_cgroup(struct page *page);
+
+enum {
+       /* flags for mem_cgroup */
+       PCG_LOCK,  /* page cgroup is locked */
+       PCG_CACHE, /* charged as cache */
+       PCG_USED, /* this object is in use. */
+       /* flags for LRU placement */
+       PCG_ACTIVE, /* page is active in this cgroup */
+       PCG_FILE, /* page is file system backed */
+       PCG_UNEVICTABLE, /* page is unevictableable */
+};
+
+#define TESTPCGFLAG(uname, lname)                      \
+static inline int PageCgroup##uname(struct page_cgroup *pc)    \
+       { return test_bit(PCG_##lname, &pc->flags); }
+
+#define SETPCGFLAG(uname, lname)                       \
+static inline void SetPageCgroup##uname(struct page_cgroup *pc)\
+       { set_bit(PCG_##lname, &pc->flags);  }
+
+#define CLEARPCGFLAG(uname, lname)                     \
+static inline void ClearPageCgroup##uname(struct page_cgroup *pc)      \
+       { clear_bit(PCG_##lname, &pc->flags);  }
+
+/* Cache flag is set only once (at allocation) */
+TESTPCGFLAG(Cache, CACHE)
+
+TESTPCGFLAG(Used, USED)
+CLEARPCGFLAG(Used, USED)
+
+/* LRU management flags (from global-lru definition) */
+TESTPCGFLAG(File, FILE)
+SETPCGFLAG(File, FILE)
+CLEARPCGFLAG(File, FILE)
+
+TESTPCGFLAG(Active, ACTIVE)
+SETPCGFLAG(Active, ACTIVE)
+CLEARPCGFLAG(Active, ACTIVE)
+
+TESTPCGFLAG(Unevictable, UNEVICTABLE)
+SETPCGFLAG(Unevictable, UNEVICTABLE)
+CLEARPCGFLAG(Unevictable, UNEVICTABLE)
+
+static inline int page_cgroup_nid(struct page_cgroup *pc)
+{
+       return page_to_nid(pc->page);
+}
+
+static inline enum zone_type page_cgroup_zid(struct page_cgroup *pc)
+{
+       return page_zonenum(pc->page);
+}
+
+static inline void lock_page_cgroup(struct page_cgroup *pc)
+{
+       bit_spin_lock(PCG_LOCK, &pc->flags);
+}
+
+static inline int trylock_page_cgroup(struct page_cgroup *pc)
+{
+       return bit_spin_trylock(PCG_LOCK, &pc->flags);
+}
+
+static inline void unlock_page_cgroup(struct page_cgroup *pc)
+{
+       bit_spin_unlock(PCG_LOCK, &pc->flags);
+}
+
+#else /* CONFIG_CGROUP_MEM_RES_CTLR */
+struct page_cgroup;
+
+static inline void pgdat_page_cgroup_init(struct pglist_data *pgdat)
+{
+}
+
+static inline struct page_cgroup *lookup_page_cgroup(struct page *page)
+{
+       return NULL;
+}
+#endif
+#endif
index da4ccf0..c06b45a 100644 (file)
@@ -33,5 +33,4 @@ obj-$(CONFIG_FS_XIP) += filemap_xip.o
 obj-$(CONFIG_MIGRATION) += migrate.o
 obj-$(CONFIG_SMP) += allocpercpu.o
 obj-$(CONFIG_QUICKLIST) += quicklist.o
-obj-$(CONFIG_CGROUP_MEM_RES_CTLR) += memcontrol.o
-
+obj-$(CONFIG_CGROUP_MEM_RES_CTLR) += memcontrol.o page_cgroup.o
index 031682e..d4a92b6 100644 (file)
 #include <linux/seq_file.h>
 #include <linux/vmalloc.h>
 #include <linux/mm_inline.h>
+#include <linux/page_cgroup.h>
 
 #include <asm/uaccess.h>
 
 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
-static struct kmem_cache *page_cgroup_cache __read_mostly;
 #define MEM_CGROUP_RECLAIM_RETRIES     5
 
 /*
@@ -135,79 +135,6 @@ struct mem_cgroup {
 };
 static struct mem_cgroup init_mem_cgroup;
 
-/*
- * We use the lower bit of the page->page_cgroup pointer as a bit spin
- * lock.  We need to ensure that page->page_cgroup is at least two
- * byte aligned (based on comments from Nick Piggin).  But since
- * bit_spin_lock doesn't actually set that lock bit in a non-debug
- * uniprocessor kernel, we should avoid setting it here too.
- */
-#define PAGE_CGROUP_LOCK_BIT   0x0
-#if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_SPINLOCK)
-#define PAGE_CGROUP_LOCK       (1 << PAGE_CGROUP_LOCK_BIT)
-#else
-#define PAGE_CGROUP_LOCK       0x0
-#endif
-
-/*
- * A page_cgroup page is associated with every page descriptor. The
- * page_cgroup helps us identify information about the cgroup
- */
-struct page_cgroup {
-       struct list_head lru;           /* per cgroup LRU list */
-       struct page *page;
-       struct mem_cgroup *mem_cgroup;
-       unsigned long flags;
-};
-
-enum {
-       /* flags for mem_cgroup */
-       PCG_CACHE, /* charged as cache */
-       /* flags for LRU placement */
-       PCG_ACTIVE, /* page is active in this cgroup */
-       PCG_FILE, /* page is file system backed */
-       PCG_UNEVICTABLE, /* page is unevictableable */
-};
-
-#define TESTPCGFLAG(uname, lname)                      \
-static inline int PageCgroup##uname(struct page_cgroup *pc)    \
-       { return test_bit(PCG_##lname, &pc->flags); }
-
-#define SETPCGFLAG(uname, lname)                       \
-static inline void SetPageCgroup##uname(struct page_cgroup *pc)\
-       { set_bit(PCG_##lname, &pc->flags);  }
-
-#define CLEARPCGFLAG(uname, lname)                     \
-static inline void ClearPageCgroup##uname(struct page_cgroup *pc)      \
-       { clear_bit(PCG_##lname, &pc->flags);  }
-
-
-/* Cache flag is set only once (at allocation) */
-TESTPCGFLAG(Cache, CACHE)
-
-/* LRU management flags (from global-lru definition) */
-TESTPCGFLAG(File, FILE)
-SETPCGFLAG(File, FILE)
-CLEARPCGFLAG(File, FILE)
-
-TESTPCGFLAG(Active, ACTIVE)
-SETPCGFLAG(Active, ACTIVE)
-CLEARPCGFLAG(Active, ACTIVE)
-
-TESTPCGFLAG(Unevictable, UNEVICTABLE)
-SETPCGFLAG(Unevictable, UNEVICTABLE)
-CLEARPCGFLAG(Unevictable, UNEVICTABLE)
-
-static int page_cgroup_nid(struct page_cgroup *pc)
-{
-       return page_to_nid(pc->page);
-}
-
-static enum zone_type page_cgroup_zid(struct page_cgroup *pc)
-{
-       return page_zonenum(pc->page);
-}
-
 enum charge_type {
        MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
        MEM_CGROUP_CHARGE_TYPE_MAPPED,
@@ -216,12 +143,18 @@ enum charge_type {
        NR_CHARGE_TYPE,
 };
 
+/* only for here (for easy reading.) */
+#define PCGF_CACHE     (1UL << PCG_CACHE)
+#define PCGF_USED      (1UL << PCG_USED)
+#define PCGF_ACTIVE    (1UL << PCG_ACTIVE)
+#define PCGF_LOCK      (1UL << PCG_LOCK)
+#define PCGF_FILE      (1UL << PCG_FILE)
 static const unsigned long
 pcg_default_flags[NR_CHARGE_TYPE] = {
-       ((1 << PCG_CACHE) | (1 << PCG_FILE)),
-       ((1 << PCG_ACTIVE)),
-       ((1 << PCG_ACTIVE) | (1 << PCG_CACHE)),
-       0,
+       PCGF_CACHE | PCGF_FILE | PCGF_USED | PCGF_LOCK, /* File Cache */
+       PCGF_ACTIVE | PCGF_USED | PCGF_LOCK, /* Anon */
+       PCGF_ACTIVE | PCGF_CACHE | PCGF_USED | PCGF_LOCK, /* Shmem */
+       0, /* FORCE */
 };
 
 /*
@@ -303,37 +236,6 @@ struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
                                struct mem_cgroup, css);
 }
 
-static inline int page_cgroup_locked(struct page *page)
-{
-       return bit_spin_is_locked(PAGE_CGROUP_LOCK_BIT, &page->page_cgroup);
-}
-
-static void page_assign_page_cgroup(struct page *page, struct page_cgroup *pc)
-{
-       VM_BUG_ON(!page_cgroup_locked(page));
-       page->page_cgroup = ((unsigned long)pc | PAGE_CGROUP_LOCK);
-}
-
-struct page_cgroup *page_get_page_cgroup(struct page *page)
-{
-       return (struct page_cgroup *) (page->page_cgroup & ~PAGE_CGROUP_LOCK);
-}
-
-static void lock_page_cgroup(struct page *page)
-{
-       bit_spin_lock(PAGE_CGROUP_LOCK_BIT, &page->page_cgroup);
-}
-
-static int try_lock_page_cgroup(struct page *page)
-{
-       return bit_spin_trylock(PAGE_CGROUP_LOCK_BIT, &page->page_cgroup);
-}
-
-static void unlock_page_cgroup(struct page *page)
-{
-       bit_spin_unlock(PAGE_CGROUP_LOCK_BIT, &page->page_cgroup);
-}
-
 static void __mem_cgroup_remove_list(struct mem_cgroup_per_zone *mz,
                        struct page_cgroup *pc)
 {
@@ -436,17 +338,16 @@ void mem_cgroup_move_lists(struct page *page, enum lru_list lru)
         * safely get to page_cgroup without it, so just try_lock it:
         * mem_cgroup_isolate_pages allows for page left on wrong list.
         */
-       if (!try_lock_page_cgroup(page))
+       pc = lookup_page_cgroup(page);
+       if (!trylock_page_cgroup(pc))
                return;
-
-       pc = page_get_page_cgroup(page);
-       if (pc) {
+       if (pc && PageCgroupUsed(pc)) {
                mz = page_cgroup_zoneinfo(pc);
                spin_lock_irqsave(&mz->lru_lock, flags);
                __mem_cgroup_move_lists(pc, lru);
                spin_unlock_irqrestore(&mz->lru_lock, flags);
        }
-       unlock_page_cgroup(page);
+       unlock_page_cgroup(pc);
 }
 
 /*
@@ -533,6 +434,8 @@ unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
        list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
                if (scan >= nr_to_scan)
                        break;
+               if (unlikely(!PageCgroupUsed(pc)))
+                       continue;
                page = pc->page;
 
                if (unlikely(!PageLRU(page)))
@@ -576,26 +479,27 @@ static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
 {
        struct mem_cgroup *mem;
        struct page_cgroup *pc;
-       unsigned long flags;
        unsigned long nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
        struct mem_cgroup_per_zone *mz;
+       unsigned long flags;
 
-       pc = kmem_cache_alloc(page_cgroup_cache, gfp_mask);
-       if (unlikely(pc == NULL))
-               goto err;
-
+       pc = lookup_page_cgroup(page);
+       /* can happen at boot */
+       if (unlikely(!pc))
+               return 0;
+       prefetchw(pc);
        /*
         * We always charge the cgroup the mm_struct belongs to.
         * The mm_struct's mem_cgroup changes on task migration if the
         * thread group leader migrates. It's possible that mm is not
         * set, if so charge the init_mm (happens for pagecache usage).
         */
+
        if (likely(!memcg)) {
                rcu_read_lock();
                mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
                if (unlikely(!mem)) {
                        rcu_read_unlock();
-                       kmem_cache_free(page_cgroup_cache, pc);
                        return 0;
                }
                /*
@@ -631,36 +535,33 @@ static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
                }
        }
 
+
+       lock_page_cgroup(pc);
+       if (unlikely(PageCgroupUsed(pc))) {
+               unlock_page_cgroup(pc);
+               res_counter_uncharge(&mem->res, PAGE_SIZE);
+               css_put(&mem->css);
+
+               goto done;
+       }
        pc->mem_cgroup = mem;
-       pc->page = page;
        /*
         * If a page is accounted as a page cache, insert to inactive list.
         * If anon, insert to active list.
         */
        pc->flags = pcg_default_flags[ctype];
 
-       lock_page_cgroup(page);
-       if (unlikely(page_get_page_cgroup(page))) {
-               unlock_page_cgroup(page);
-               res_counter_uncharge(&mem->res, PAGE_SIZE);
-               css_put(&mem->css);
-               kmem_cache_free(page_cgroup_cache, pc);
-               goto done;
-       }
-       page_assign_page_cgroup(page, 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(pc);
 
-       unlock_page_cgroup(page);
 done:
        return 0;
 out:
        css_put(&mem->css);
-       kmem_cache_free(page_cgroup_cache, pc);
-err:
        return -ENOMEM;
 }
 
@@ -668,7 +569,8 @@ int mem_cgroup_charge(struct page *page, struct mm_struct *mm, gfp_t gfp_mask)
 {
        if (mem_cgroup_subsys.disabled)
                return 0;
-
+       if (PageCompound(page))
+               return 0;
        /*
         * If already mapped, we don't have to account.
         * If page cache, page->mapping has address_space.
@@ -689,7 +591,8 @@ int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
 {
        if (mem_cgroup_subsys.disabled)
                return 0;
-
+       if (PageCompound(page))
+               return 0;
        /*
         * Corner case handling. This is called from add_to_page_cache()
         * in usual. But some FS (shmem) precharges this page before calling it
@@ -702,15 +605,16 @@ int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
        if (!(gfp_mask & __GFP_WAIT)) {
                struct page_cgroup *pc;
 
-               lock_page_cgroup(page);
-               pc = page_get_page_cgroup(page);
-               if (pc) {
-                       VM_BUG_ON(pc->page != page);
-                       VM_BUG_ON(!pc->mem_cgroup);
-                       unlock_page_cgroup(page);
+
+               pc = lookup_page_cgroup(page);
+               if (!pc)
+                       return 0;
+               lock_page_cgroup(pc);
+               if (PageCgroupUsed(pc)) {
+                       unlock_page_cgroup(pc);
                        return 0;
                }
-               unlock_page_cgroup(page);
+               unlock_page_cgroup(pc);
        }
 
        if (unlikely(!mm))
@@ -741,37 +645,39 @@ __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
        /*
         * Check if our page_cgroup is valid
         */
-       lock_page_cgroup(page);
-       pc = page_get_page_cgroup(page);
-       if (unlikely(!pc))
-               goto unlock;
-
-       VM_BUG_ON(pc->page != page);
+       pc = lookup_page_cgroup(page);
+       if (unlikely(!pc || !PageCgroupUsed(pc)))
+               return;
 
-       if ((ctype == MEM_CGROUP_CHARGE_TYPE_MAPPED)
-           && ((PageCgroupCache(pc) || page_mapped(page))))
-               goto unlock;
+       lock_page_cgroup(pc);
+       if ((ctype == MEM_CGROUP_CHARGE_TYPE_MAPPED && page_mapped(page))
+            || !PageCgroupUsed(pc)) {
+               /* This happens at race in zap_pte_range() and do_swap_page()*/
+               unlock_page_cgroup(pc);
+               return;
+       }
+       ClearPageCgroupUsed(pc);
+       mem = pc->mem_cgroup;
 
        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);
+       unlock_page_cgroup(pc);
 
-       page_assign_page_cgroup(page, NULL);
-       unlock_page_cgroup(page);
-
-       mem = pc->mem_cgroup;
        res_counter_uncharge(&mem->res, PAGE_SIZE);
        css_put(&mem->css);
 
-       kmem_cache_free(page_cgroup_cache, pc);
        return;
-unlock:
-       unlock_page_cgroup(page);
 }
 
 void mem_cgroup_uncharge_page(struct page *page)
 {
+       /* early check. */
+       if (page_mapped(page))
+               return;
+       if (page->mapping && !PageAnon(page))
+               return;
        __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
 }
 
@@ -795,9 +701,9 @@ int mem_cgroup_prepare_migration(struct page *page, struct page *newpage)
        if (mem_cgroup_subsys.disabled)
                return 0;
 
-       lock_page_cgroup(page);
-       pc = page_get_page_cgroup(page);
-       if (pc) {
+       pc = lookup_page_cgroup(page);
+       lock_page_cgroup(pc);
+       if (PageCgroupUsed(pc)) {
                mem = pc->mem_cgroup;
                css_get(&mem->css);
                if (PageCgroupCache(pc)) {
@@ -807,7 +713,7 @@ int mem_cgroup_prepare_migration(struct page *page, struct page *newpage)
                                ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
                }
        }
-       unlock_page_cgroup(page);
+       unlock_page_cgroup(pc);
        if (mem) {
                ret = mem_cgroup_charge_common(newpage, NULL, GFP_KERNEL,
                        ctype, mem);
@@ -832,7 +738,7 @@ void mem_cgroup_end_migration(struct page *newpage)
         */
        if (!newpage->mapping)
                __mem_cgroup_uncharge_common(newpage,
-                                        MEM_CGROUP_CHARGE_TYPE_FORCE);
+                               MEM_CGROUP_CHARGE_TYPE_FORCE);
        else if (PageAnon(newpage))
                mem_cgroup_uncharge_page(newpage);
 }
@@ -918,6 +824,8 @@ static void mem_cgroup_force_empty_list(struct mem_cgroup *mem,
        while (!list_empty(list)) {
                pc = list_entry(list->prev, struct page_cgroup, lru);
                page = pc->page;
+               if (!PageCgroupUsed(pc))
+                       break;
                get_page(page);
                spin_unlock_irqrestore(&mz->lru_lock, flags);
                /*
@@ -932,8 +840,10 @@ static void mem_cgroup_force_empty_list(struct mem_cgroup *mem,
                                count = FORCE_UNCHARGE_BATCH;
                                cond_resched();
                        }
-               } else
-                       cond_resched();
+               } else {
+                       spin_lock_irqsave(&mz->lru_lock, flags);
+                       break;
+               }
                spin_lock_irqsave(&mz->lru_lock, flags);
        }
        spin_unlock_irqrestore(&mz->lru_lock, flags);
@@ -957,6 +867,8 @@ static int mem_cgroup_force_empty(struct mem_cgroup *mem)
        while (mem->res.usage > 0) {
                if (atomic_read(&mem->css.cgroup->count) > 0)
                        goto out;
+               /* This is for making all *used* pages to be on LRU. */
+               lru_add_drain_all();
                for_each_node_state(node, N_POSSIBLE)
                        for (zid = 0; zid < MAX_NR_ZONES; zid++) {
                                struct mem_cgroup_per_zone *mz;
@@ -965,6 +877,7 @@ static int mem_cgroup_force_empty(struct mem_cgroup *mem)
                                for_each_lru(l)
                                        mem_cgroup_force_empty_list(mem, mz, l);
                        }
+               cond_resched();
        }
        ret = 0;
 out:
@@ -1175,8 +1088,8 @@ mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
        int node;
 
        if (unlikely((cont->parent) == NULL)) {
+               page_cgroup_init();
                mem = &init_mem_cgroup;
-               page_cgroup_cache = KMEM_CACHE(page_cgroup, SLAB_PANIC);
        } else {
                mem = mem_cgroup_alloc();
                if (!mem)
index f2fc44e..d0a240f 100644 (file)
@@ -44,7 +44,7 @@
 #include <linux/backing-dev.h>
 #include <linux/fault-inject.h>
 #include <linux/page-isolation.h>
-#include <linux/memcontrol.h>
+#include <linux/page_cgroup.h>
 #include <linux/debugobjects.h>
 
 #include <asm/tlbflush.h>
@@ -223,17 +223,12 @@ static inline int bad_range(struct zone *zone, struct page *page)
 
 static void bad_page(struct page *page)
 {
-       void *pc = page_get_page_cgroup(page);
-
        printk(KERN_EMERG "Bad page state in process '%s'\n" KERN_EMERG
                "page:%p flags:0x%0*lx mapping:%p mapcount:%d count:%d\n",
                current->comm, page, (int)(2*sizeof(unsigned long)),
                (unsigned long)page->flags, page->mapping,
                page_mapcount(page), page_count(page));
-       if (pc) {
-               printk(KERN_EMERG "cgroup:%p\n", pc);
-               page_reset_bad_cgroup(page);
-       }
+
        printk(KERN_EMERG "Trying to fix it up, but a reboot is needed\n"
                KERN_EMERG "Backtrace:\n");
        dump_stack();
@@ -457,7 +452,6 @@ static inline int free_pages_check(struct page *page)
        free_page_mlock(page);
        if (unlikely(page_mapcount(page) |
                (page->mapping != NULL)  |
-               (page_get_page_cgroup(page) != NULL) |
                (page_count(page) != 0)  |
                (page->flags & PAGE_FLAGS_CHECK_AT_FREE)))
                bad_page(page);
@@ -603,7 +597,6 @@ static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
 {
        if (unlikely(page_mapcount(page) |
                (page->mapping != NULL)  |
-               (page_get_page_cgroup(page) != NULL) |
                (page_count(page) != 0)  |
                (page->flags & PAGE_FLAGS_CHECK_AT_PREP)))
                bad_page(page);
@@ -3438,6 +3431,7 @@ static void __paginginit free_area_init_core(struct pglist_data *pgdat,
        pgdat->nr_zones = 0;
        init_waitqueue_head(&pgdat->kswapd_wait);
        pgdat->kswapd_max_order = 0;
+       pgdat_page_cgroup_init(pgdat);
        
        for (j = 0; j < MAX_NR_ZONES; j++) {
                struct zone *zone = pgdat->node_zones + j;
diff --git a/mm/page_cgroup.c b/mm/page_cgroup.c
new file mode 100644 (file)
index 0000000..5d86550
--- /dev/null
@@ -0,0 +1,237 @@
+#include <linux/mm.h>
+#include <linux/mmzone.h>
+#include <linux/bootmem.h>
+#include <linux/bit_spinlock.h>
+#include <linux/page_cgroup.h>
+#include <linux/hash.h>
+#include <linux/memory.h>
+
+static void __meminit
+__init_page_cgroup(struct page_cgroup *pc, unsigned long pfn)
+{
+       pc->flags = 0;
+       pc->mem_cgroup = NULL;
+       pc->page = pfn_to_page(pfn);
+}
+static unsigned long total_usage;
+
+#if !defined(CONFIG_SPARSEMEM)
+
+
+void __init pgdat_page_cgroup_init(struct pglist_data *pgdat)
+{
+       pgdat->node_page_cgroup = NULL;
+}
+
+struct page_cgroup *lookup_page_cgroup(struct page *page)
+{
+       unsigned long pfn = page_to_pfn(page);
+       unsigned long offset;
+       struct page_cgroup *base;
+
+       base = NODE_DATA(page_to_nid(page))->node_page_cgroup;
+       if (unlikely(!base))
+               return NULL;
+
+       offset = pfn - NODE_DATA(page_to_nid(page))->node_start_pfn;
+       return base + offset;
+}
+
+static int __init alloc_node_page_cgroup(int nid)
+{
+       struct page_cgroup *base, *pc;
+       unsigned long table_size;
+       unsigned long start_pfn, nr_pages, index;
+
+       start_pfn = NODE_DATA(nid)->node_start_pfn;
+       nr_pages = NODE_DATA(nid)->node_spanned_pages;
+
+       table_size = sizeof(struct page_cgroup) * nr_pages;
+
+       base = __alloc_bootmem_node_nopanic(NODE_DATA(nid),
+                       table_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
+       if (!base)
+               return -ENOMEM;
+       for (index = 0; index < nr_pages; index++) {
+               pc = base + index;
+               __init_page_cgroup(pc, start_pfn + index);
+       }
+       NODE_DATA(nid)->node_page_cgroup = base;
+       total_usage += table_size;
+       return 0;
+}
+
+void __init page_cgroup_init(void)
+{
+
+       int nid, fail;
+
+       for_each_online_node(nid)  {
+               fail = alloc_node_page_cgroup(nid);
+               if (fail)
+                       goto fail;
+       }
+       printk(KERN_INFO "allocated %ld bytes of page_cgroup\n", total_usage);
+       printk(KERN_INFO "please try cgroup_disable=memory option if you"
+       " don't want\n");
+       return;
+fail:
+       printk(KERN_CRIT "allocation of page_cgroup was failed.\n");
+       printk(KERN_CRIT "please try cgroup_disable=memory boot option\n");
+       panic("Out of memory");
+}
+
+#else /* CONFIG_FLAT_NODE_MEM_MAP */
+
+struct page_cgroup *lookup_page_cgroup(struct page *page)
+{
+       unsigned long pfn = page_to_pfn(page);
+       struct mem_section *section = __pfn_to_section(pfn);
+
+       return section->page_cgroup + pfn;
+}
+
+int __meminit init_section_page_cgroup(unsigned long pfn)
+{
+       struct mem_section *section;
+       struct page_cgroup *base, *pc;
+       unsigned long table_size;
+       int nid, index;
+
+       section = __pfn_to_section(pfn);
+
+       if (section->page_cgroup)
+               return 0;
+
+       nid = page_to_nid(pfn_to_page(pfn));
+
+       table_size = sizeof(struct page_cgroup) * PAGES_PER_SECTION;
+       base = kmalloc_node(table_size, GFP_KERNEL, nid);
+       if (!base)
+               base = vmalloc_node(table_size, nid);
+
+       if (!base) {
+               printk(KERN_ERR "page cgroup allocation failure\n");
+               return -ENOMEM;
+       }
+
+       for (index = 0; index < PAGES_PER_SECTION; index++) {
+               pc = base + index;
+               __init_page_cgroup(pc, pfn + index);
+       }
+
+       section = __pfn_to_section(pfn);
+       section->page_cgroup = base - pfn;
+       total_usage += table_size;
+       return 0;
+}
+#ifdef CONFIG_MEMORY_HOTPLUG
+void __free_page_cgroup(unsigned long pfn)
+{
+       struct mem_section *ms;
+       struct page_cgroup *base;
+
+       ms = __pfn_to_section(pfn);
+       if (!ms || !ms->page_cgroup)
+               return;
+       base = ms->page_cgroup + pfn;
+       ms->page_cgroup = NULL;
+       if (is_vmalloc_addr(base))
+               vfree(base);
+       else
+               kfree(base);
+}
+
+int online_page_cgroup(unsigned long start_pfn,
+                       unsigned long nr_pages,
+                       int nid)
+{
+       unsigned long start, end, pfn;
+       int fail = 0;
+
+       start = start_pfn & (PAGES_PER_SECTION - 1);
+       end = ALIGN(start_pfn + nr_pages, PAGES_PER_SECTION);
+
+       for (pfn = start; !fail && pfn < end; pfn += PAGES_PER_SECTION) {
+               if (!pfn_present(pfn))
+                       continue;
+               fail = init_section_page_cgroup(pfn);
+       }
+       if (!fail)
+               return 0;
+
+       /* rollback */
+       for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
+               __free_page_cgroup(pfn);
+
+       return -ENOMEM;
+}
+
+int offline_page_cgroup(unsigned long start_pfn,
+               unsigned long nr_pages, int nid)
+{
+       unsigned long start, end, pfn;
+
+       start = start_pfn & (PAGES_PER_SECTION - 1);
+       end = ALIGN(start_pfn + nr_pages, PAGES_PER_SECTION);
+
+       for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
+               __free_page_cgroup(pfn);
+       return 0;
+
+}
+
+static int page_cgroup_callback(struct notifier_block *self,
+                              unsigned long action, void *arg)
+{
+       struct memory_notify *mn = arg;
+       int ret = 0;
+       switch (action) {
+       case MEM_GOING_ONLINE:
+               ret = online_page_cgroup(mn->start_pfn,
+                                  mn->nr_pages, mn->status_change_nid);
+               break;
+       case MEM_CANCEL_ONLINE:
+       case MEM_OFFLINE:
+               offline_page_cgroup(mn->start_pfn,
+                               mn->nr_pages, mn->status_change_nid);
+               break;
+       case MEM_GOING_OFFLINE:
+               break;
+       case MEM_ONLINE:
+       case MEM_CANCEL_OFFLINE:
+               break;
+       }
+       ret = notifier_from_errno(ret);
+       return ret;
+}
+
+#endif
+
+void __init page_cgroup_init(void)
+{
+       unsigned long pfn;
+       int fail = 0;
+
+       for (pfn = 0; !fail && pfn < max_pfn; pfn += PAGES_PER_SECTION) {
+               if (!pfn_present(pfn))
+                       continue;
+               fail = init_section_page_cgroup(pfn);
+       }
+       if (fail) {
+               printk(KERN_CRIT "try cgroup_disable=memory boot option\n");
+               panic("Out of memory");
+       } else {
+               hotplug_memory_notifier(page_cgroup_callback, 0);
+       }
+       printk(KERN_INFO "allocated %ld bytes of page_cgroup\n", total_usage);
+       printk(KERN_INFO "please try cgroup_disable=memory option if you don't"
+       " want\n");
+}
+
+void __init pgdat_page_cgroup_init(struct pglist_data *pgdat)
+{
+       return;
+}
+
+#endif