memcgroup: fix check for thread being a group leader in memcgroup
[sfrench/cifs-2.6.git] / mm / memcontrol.c
index 83ba13ad31e16d69d32087bf1a02285b72e3844b..9b648bd63451dd81026649f244a6f8ae0d4b0699 100644 (file)
@@ -137,6 +137,7 @@ struct mem_cgroup {
         */
        struct mem_cgroup_stat stat;
 };
+static struct mem_cgroup init_mem_cgroup;
 
 /*
  * We use the lower bit of the page->page_cgroup pointer as a bit spin
@@ -160,37 +161,27 @@ struct page_cgroup {
        struct list_head lru;           /* per cgroup LRU list */
        struct page *page;
        struct mem_cgroup *mem_cgroup;
-       atomic_t ref_cnt;               /* Helpful when pages move b/w  */
-                                       /* mapped and cached states     */
-       int      flags;
+       int ref_cnt;                    /* cached, mapped, migrating */
+       int flags;
 };
 #define PAGE_CGROUP_FLAG_CACHE (0x1)   /* charged as cache */
 #define PAGE_CGROUP_FLAG_ACTIVE (0x2)  /* page is active in this cgroup */
 
-static inline int page_cgroup_nid(struct page_cgroup *pc)
+static 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)
+static enum zone_type page_cgroup_zid(struct page_cgroup *pc)
 {
        return page_zonenum(pc->page);
 }
 
-enum {
-       MEM_CGROUP_TYPE_UNSPEC = 0,
-       MEM_CGROUP_TYPE_MAPPED,
-       MEM_CGROUP_TYPE_CACHED,
-       MEM_CGROUP_TYPE_ALL,
-       MEM_CGROUP_TYPE_MAX,
-};
-
 enum charge_type {
        MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
        MEM_CGROUP_CHARGE_TYPE_MAPPED,
 };
 
-
 /*
  * Always modified under lru lock. Then, not necessary to preempt_disable()
  */
@@ -199,23 +190,21 @@ static void mem_cgroup_charge_statistics(struct mem_cgroup *mem, int flags,
 {
        int val = (charge)? 1 : -1;
        struct mem_cgroup_stat *stat = &mem->stat;
-       VM_BUG_ON(!irqs_disabled());
 
+       VM_BUG_ON(!irqs_disabled());
        if (flags & PAGE_CGROUP_FLAG_CACHE)
-               __mem_cgroup_stat_add_safe(stat,
-                                       MEM_CGROUP_STAT_CACHE, val);
+               __mem_cgroup_stat_add_safe(stat, MEM_CGROUP_STAT_CACHE, val);
        else
                __mem_cgroup_stat_add_safe(stat, MEM_CGROUP_STAT_RSS, val);
 }
 
-static inline struct mem_cgroup_per_zone *
+static struct mem_cgroup_per_zone *
 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
 {
-       BUG_ON(!mem->info.nodeinfo[nid]);
        return &mem->info.nodeinfo[nid]->zoneinfo[zid];
 }
 
-static inline struct mem_cgroup_per_zone *
+static struct mem_cgroup_per_zone *
 page_cgroup_zoneinfo(struct page_cgroup *pc)
 {
        struct mem_cgroup *mem = pc->mem_cgroup;
@@ -240,18 +229,14 @@ static unsigned long mem_cgroup_get_all_zonestat(struct mem_cgroup *mem,
        return total;
 }
 
-static struct mem_cgroup init_mem_cgroup;
-
-static inline
-struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
+static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
 {
        return container_of(cgroup_subsys_state(cont,
                                mem_cgroup_subsys_id), struct mem_cgroup,
                                css);
 }
 
-static inline
-struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
+static struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
 {
        return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
                                struct mem_cgroup, css);
@@ -273,8 +258,7 @@ void mm_free_cgroup(struct mm_struct *mm)
 
 static inline int page_cgroup_locked(struct page *page)
 {
-       return bit_spin_is_locked(PAGE_CGROUP_LOCK_BIT,
-                                       &page->page_cgroup);
+       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)
@@ -285,41 +269,22 @@ static void page_assign_page_cgroup(struct page *page, struct page_cgroup *pc)
 
 struct page_cgroup *page_get_page_cgroup(struct page *page)
 {
-       return (struct page_cgroup *)
-               (page->page_cgroup & ~PAGE_CGROUP_LOCK);
+       return (struct page_cgroup *) (page->page_cgroup & ~PAGE_CGROUP_LOCK);
 }
 
-static void __always_inline lock_page_cgroup(struct page *page)
+static void lock_page_cgroup(struct page *page)
 {
        bit_spin_lock(PAGE_CGROUP_LOCK_BIT, &page->page_cgroup);
-       VM_BUG_ON(!page_cgroup_locked(page));
 }
 
-static void __always_inline unlock_page_cgroup(struct page *page)
+static int try_lock_page_cgroup(struct page *page)
 {
-       bit_spin_unlock(PAGE_CGROUP_LOCK_BIT, &page->page_cgroup);
+       return bit_spin_trylock(PAGE_CGROUP_LOCK_BIT, &page->page_cgroup);
 }
 
-/*
- * Clear page->page_cgroup member under lock_page_cgroup().
- * If given "pc" value is different from one page->page_cgroup,
- * page->cgroup is not cleared.
- * Returns a value of page->page_cgroup at lock taken.
- * A can can detect failure of clearing by following
- *  clear_page_cgroup(page, pc) == pc
- */
-
-static struct page_cgroup *clear_page_cgroup(struct page *page,
-                                               struct page_cgroup *pc)
+static void unlock_page_cgroup(struct page *page)
 {
-       struct page_cgroup *ret;
-       /* lock and clear */
-       lock_page_cgroup(page);
-       ret = page_get_page_cgroup(page);
-       if (likely(ret == pc))
-               page_assign_page_cgroup(page, NULL);
-       unlock_page_cgroup(page);
-       return ret;
+       bit_spin_unlock(PAGE_CGROUP_LOCK_BIT, &page->page_cgroup);
 }
 
 static void __mem_cgroup_remove_list(struct page_cgroup *pc)
@@ -391,14 +356,24 @@ void mem_cgroup_move_lists(struct page *page, bool active)
        struct mem_cgroup_per_zone *mz;
        unsigned long flags;
 
-       pc = page_get_page_cgroup(page);
-       if (!pc)
+       /*
+        * We cannot lock_page_cgroup while holding zone's lru_lock,
+        * because other holders of lock_page_cgroup can be interrupted
+        * with an attempt to rotate_reclaimable_page.  But we cannot
+        * 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))
                return;
 
-       mz = page_cgroup_zoneinfo(pc);
-       spin_lock_irqsave(&mz->lru_lock, flags);
-       __mem_cgroup_move_lists(pc, active);
-       spin_unlock_irqrestore(&mz->lru_lock, flags);
+       pc = page_get_page_cgroup(page);
+       if (pc) {
+               mz = page_cgroup_zoneinfo(pc);
+               spin_lock_irqsave(&mz->lru_lock, flags);
+               __mem_cgroup_move_lists(pc, active);
+               spin_unlock_irqrestore(&mz->lru_lock, flags);
+       }
+       unlock_page_cgroup(page);
 }
 
 /*
@@ -417,6 +392,7 @@ int mem_cgroup_calc_mapped_ratio(struct mem_cgroup *mem)
        rss = (long)mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_RSS);
        return (int)((rss * 100L) / total);
 }
+
 /*
  * This function is called from vmscan.c. In page reclaiming loop. balance
  * between active and inactive list is calculated. For memory controller
@@ -480,7 +456,6 @@ long mem_cgroup_calc_reclaim_inactive(struct mem_cgroup *mem,
        struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(mem, nid, zid);
 
        nr_inactive = MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_INACTIVE);
-
        return (nr_inactive >> priority);
 }
 
@@ -573,15 +548,12 @@ retry:
         * the page has already been accounted.
         */
        if (pc) {
-               if (unlikely(!atomic_inc_not_zero(&pc->ref_cnt))) {
-                       /* this page is under being uncharged ? */
-                       unlock_page_cgroup(page);
-                       cpu_relax();
-                       goto retry;
-               } else {
-                       unlock_page_cgroup(page);
-                       goto done;
-               }
+               VM_BUG_ON(pc->page != page);
+               VM_BUG_ON(pc->ref_cnt <= 0);
+
+               pc->ref_cnt++;
+               unlock_page_cgroup(page);
+               goto done;
        }
        unlock_page_cgroup(page);
 
@@ -601,16 +573,11 @@ retry:
        rcu_read_lock();
        mem = rcu_dereference(mm->mem_cgroup);
        /*
-        * For every charge from the cgroup, increment reference
-        * count
+        * For every charge from the cgroup, increment reference count
         */
        css_get(&mem->css);
        rcu_read_unlock();
 
-       /*
-        * If we created the page_cgroup, we should free it on exceeding
-        * the cgroup limit.
-        */
        while (res_counter_charge(&mem->res, PAGE_SIZE)) {
                if (!(gfp_mask & __GFP_WAIT))
                        goto out;
@@ -619,12 +586,12 @@ retry:
                        continue;
 
                /*
-                * try_to_free_mem_cgroup_pages() might not give us a full
-                * picture of reclaim. Some pages are reclaimed and might be
-                * moved to swap cache or just unmapped from the cgroup.
-                * Check the limit again to see if the reclaim reduced the
-                * current usage of the cgroup before giving up
-                */
+                * try_to_free_mem_cgroup_pages() might not give us a full
+                * picture of reclaim. Some pages are reclaimed and might be
+                * moved to swap cache or just unmapped from the cgroup.
+                * Check the limit again to see if the reclaim reduced the
+                * current usage of the cgroup before giving up
+                */
                if (res_counter_check_under_limit(&mem->res))
                        continue;
 
@@ -635,7 +602,7 @@ retry:
                congestion_wait(WRITE, HZ/10);
        }
 
-       atomic_set(&pc->ref_cnt, 1);
+       pc->ref_cnt = 1;
        pc->mem_cgroup = mem;
        pc->page = page;
        pc->flags = PAGE_CGROUP_FLAG_ACTIVE;
@@ -656,14 +623,13 @@ retry:
                goto retry;
        }
        page_assign_page_cgroup(page, pc);
-       unlock_page_cgroup(page);
 
        mz = page_cgroup_zoneinfo(pc);
        spin_lock_irqsave(&mz->lru_lock, flags);
-       /* Update statistics vector */
        __mem_cgroup_add_list(pc);
        spin_unlock_irqrestore(&mz->lru_lock, flags);
 
+       unlock_page_cgroup(page);
 done:
        return 0;
 out:
@@ -673,70 +639,61 @@ err:
        return -ENOMEM;
 }
 
-int mem_cgroup_charge(struct page *page, struct mm_struct *mm,
-                       gfp_t gfp_mask)
+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);
 }
 
-/*
- * See if the cached pages should be charged at all?
- */
 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
                                gfp_t gfp_mask)
 {
-       int ret = 0;
        if (!mm)
                mm = &init_mm;
-
-       ret = mem_cgroup_charge_common(page, mm, gfp_mask,
+       return mem_cgroup_charge_common(page, mm, gfp_mask,
                                MEM_CGROUP_CHARGE_TYPE_CACHE);
-       return ret;
 }
 
 /*
  * Uncharging is always a welcome operation, we never complain, simply
- * uncharge. This routine should be called with lock_page_cgroup held
+ * uncharge.
  */
-void mem_cgroup_uncharge(struct page_cgroup *pc)
+void mem_cgroup_uncharge_page(struct page *page)
 {
+       struct page_cgroup *pc;
        struct mem_cgroup *mem;
        struct mem_cgroup_per_zone *mz;
-       struct page *page;
        unsigned long flags;
 
        /*
         * Check if our page_cgroup is valid
         */
+       lock_page_cgroup(page);
+       pc = page_get_page_cgroup(page);
        if (!pc)
-               return;
+               goto unlock;
 
-       if (atomic_dec_and_test(&pc->ref_cnt)) {
-               page = pc->page;
+       VM_BUG_ON(pc->page != page);
+       VM_BUG_ON(pc->ref_cnt <= 0);
+
+       if (--(pc->ref_cnt) == 0) {
                mz = page_cgroup_zoneinfo(pc);
-               /*
-                * get page->cgroup and clear it under lock.
-                * force_empty can drop page->cgroup without checking refcnt.
-                */
+               spin_lock_irqsave(&mz->lru_lock, flags);
+               __mem_cgroup_remove_list(pc);
+               spin_unlock_irqrestore(&mz->lru_lock, flags);
+
+               page_assign_page_cgroup(page, NULL);
                unlock_page_cgroup(page);
-               if (clear_page_cgroup(page, pc) == pc) {
-                       mem = pc->mem_cgroup;
-                       css_put(&mem->css);
-                       res_counter_uncharge(&mem->res, PAGE_SIZE);
-                       spin_lock_irqsave(&mz->lru_lock, flags);
-                       __mem_cgroup_remove_list(pc);
-                       spin_unlock_irqrestore(&mz->lru_lock, flags);
-                       kfree(pc);
-               }
-               lock_page_cgroup(page);
+
+               mem = pc->mem_cgroup;
+               res_counter_uncharge(&mem->res, PAGE_SIZE);
+               css_put(&mem->css);
+
+               kfree(pc);
+               return;
        }
-}
 
-void mem_cgroup_uncharge_page(struct page *page)
-{
-       lock_page_cgroup(page);
-       mem_cgroup_uncharge(page_get_page_cgroup(page));
+unlock:
        unlock_page_cgroup(page);
 }
 
@@ -744,63 +701,59 @@ void mem_cgroup_uncharge_page(struct page *page)
  * Returns non-zero if a page (under migration) has valid page_cgroup member.
  * Refcnt of page_cgroup is incremented.
  */
-
 int mem_cgroup_prepare_migration(struct page *page)
 {
        struct page_cgroup *pc;
-       int ret = 0;
+
        lock_page_cgroup(page);
        pc = page_get_page_cgroup(page);
-       if (pc && atomic_inc_not_zero(&pc->ref_cnt))
-               ret = 1;
+       if (pc)
+               pc->ref_cnt++;
        unlock_page_cgroup(page);
-       return ret;
+       return pc != NULL;
 }
 
 void mem_cgroup_end_migration(struct page *page)
 {
-       struct page_cgroup *pc;
-
-       lock_page_cgroup(page);
-       pc = page_get_page_cgroup(page);
-       mem_cgroup_uncharge(pc);
-       unlock_page_cgroup(page);
+       mem_cgroup_uncharge_page(page);
 }
+
 /*
- * We know both *page* and *newpage* are now not-on-LRU and Pg_locked.
+ * 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)
 {
        struct page_cgroup *pc;
-       struct mem_cgroup *mem;
-       unsigned long flags;
        struct mem_cgroup_per_zone *mz;
-retry:
+       unsigned long flags;
+
+       lock_page_cgroup(page);
        pc = page_get_page_cgroup(page);
-       if (!pc)
+       if (!pc) {
+               unlock_page_cgroup(page);
                return;
-       mem = pc->mem_cgroup;
+       }
+
        mz = page_cgroup_zoneinfo(pc);
-       if (clear_page_cgroup(page, pc) != pc)
-               goto retry;
        spin_lock_irqsave(&mz->lru_lock, flags);
-
        __mem_cgroup_remove_list(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);
-       unlock_page_cgroup(newpage);
 
        mz = page_cgroup_zoneinfo(pc);
        spin_lock_irqsave(&mz->lru_lock, flags);
        __mem_cgroup_add_list(pc);
        spin_unlock_irqrestore(&mz->lru_lock, flags);
-       return;
+
+       unlock_page_cgroup(newpage);
 }
 
 /*
@@ -809,14 +762,13 @@ retry:
  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
  */
 #define FORCE_UNCHARGE_BATCH   (128)
-static void
-mem_cgroup_force_empty_list(struct mem_cgroup *mem,
+static void mem_cgroup_force_empty_list(struct mem_cgroup *mem,
                            struct mem_cgroup_per_zone *mz,
                            int active)
 {
        struct page_cgroup *pc;
        struct page *page;
-       int count;
+       int count = FORCE_UNCHARGE_BATCH;
        unsigned long flags;
        struct list_head *list;
 
@@ -825,46 +777,36 @@ mem_cgroup_force_empty_list(struct mem_cgroup *mem,
        else
                list = &mz->inactive_list;
 
-       if (list_empty(list))
-               return;
-retry:
-       count = FORCE_UNCHARGE_BATCH;
        spin_lock_irqsave(&mz->lru_lock, flags);
-
-       while (--count && !list_empty(list)) {
+       while (!list_empty(list)) {
                pc = list_entry(list->prev, struct page_cgroup, lru);
                page = pc->page;
-               /* Avoid race with charge */
-               atomic_set(&pc->ref_cnt, 0);
-               if (clear_page_cgroup(page, pc) == pc) {
-                       css_put(&mem->css);
-                       res_counter_uncharge(&mem->res, PAGE_SIZE);
-                       __mem_cgroup_remove_list(pc);
-                       kfree(pc);
-               } else  /* being uncharged ? ...do relax */
-                       break;
+               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;
+                       cond_resched();
+               }
+               spin_lock_irqsave(&mz->lru_lock, flags);
        }
        spin_unlock_irqrestore(&mz->lru_lock, flags);
-       if (!list_empty(list)) {
-               cond_resched();
-               goto retry;
-       }
-       return;
 }
 
 /*
  * make mem_cgroup's charge to be 0 if there is no task.
  * This enables deleting this mem_cgroup.
  */
-
-int mem_cgroup_force_empty(struct mem_cgroup *mem)
+static int mem_cgroup_force_empty(struct mem_cgroup *mem)
 {
        int ret = -EBUSY;
        int node, zid;
+
        css_get(&mem->css);
        /*
         * page reclaim code (kswapd etc..) will move pages between
-`       * active_list <-> inactive_list while we don't take a lock.
+        * active_list <-> inactive_list while we don't take a lock.
         * So, we have to do loop here until all lists are empty.
         */
        while (mem->res.usage > 0) {
@@ -886,9 +828,7 @@ out:
        return ret;
 }
 
-
-
-int mem_cgroup_write_strategy(char *buf, unsigned long long *tmp)
+static int mem_cgroup_write_strategy(char *buf, unsigned long long *tmp)
 {
        *tmp = memparse(buf, &buf);
        if (*buf != '\0')
@@ -925,8 +865,7 @@ static ssize_t mem_force_empty_write(struct cgroup *cont,
                                size_t nbytes, loff_t *ppos)
 {
        struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
-       int ret;
-       ret = mem_cgroup_force_empty(mem);
+       int ret = mem_cgroup_force_empty(mem);
        if (!ret)
                ret = nbytes;
        return ret;
@@ -935,7 +874,6 @@ static ssize_t mem_force_empty_write(struct cgroup *cont,
 /*
  * Note: This should be removed if cgroup supports write-only file.
  */
-
 static ssize_t mem_force_empty_read(struct cgroup *cont,
                                struct cftype *cft,
                                struct file *file, char __user *userbuf,
@@ -944,7 +882,6 @@ static ssize_t mem_force_empty_read(struct cgroup *cont,
        return -EINVAL;
 }
 
-
 static const struct mem_cgroup_stat_desc {
        const char *msg;
        u64 unit;
@@ -997,8 +934,6 @@ static int mem_control_stat_open(struct inode *unused, struct file *file)
        return single_open(file, mem_control_stat_show, cont);
 }
 
-
-
 static struct cftype mem_cgroup_files[] = {
        {
                .name = "usage_in_bytes",
@@ -1064,9 +999,6 @@ static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
        kfree(mem->info.nodeinfo[node]);
 }
 
-
-static struct mem_cgroup init_mem_cgroup;
-
 static struct cgroup_subsys_state *
 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
 {
@@ -1147,7 +1079,7 @@ static void mem_cgroup_move_task(struct cgroup_subsys *ss,
         * Only thread group leaders are allowed to migrate, the mm_struct is
         * in effect owned by the leader
         */
-       if (p->tgid != p->pid)
+       if (!thread_group_leader(p))
                goto out;
 
        css_get(&mem->css);
@@ -1156,7 +1088,6 @@ static void mem_cgroup_move_task(struct cgroup_subsys *ss,
 
 out:
        mmput(mm);
-       return;
 }
 
 struct cgroup_subsys mem_cgroup_subsys = {