memcontrol: move oom task exclusion to tasklist scan
[linux-2.6.git] / mm / memcontrol.c
index 9e9ff91..2fadd48 100644 (file)
@@ -29,6 +29,8 @@
 #include <linux/spinlock.h>
 #include <linux/fs.h>
 
+#include <asm/uaccess.h>
+
 struct cgroup_subsys mem_cgroup_subsys;
 static const int MEM_CGROUP_RECLAIM_RETRIES = 5;
 
@@ -60,6 +62,7 @@ struct mem_cgroup {
         * spin_lock to protect the per cgroup LRU
         */
        spinlock_t lru_lock;
+       unsigned long control_type;     /* control RSS or RSS+Pagecache */
 };
 
 /*
@@ -82,6 +85,15 @@ struct page_cgroup {
                                        /* mapped and cached states     */
 };
 
+enum {
+       MEM_CGROUP_TYPE_UNSPEC = 0,
+       MEM_CGROUP_TYPE_MAPPED,
+       MEM_CGROUP_TYPE_CACHED,
+       MEM_CGROUP_TYPE_ALL,
+       MEM_CGROUP_TYPE_MAX,
+};
+
+static struct mem_cgroup init_mem_cgroup;
 
 static inline
 struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
@@ -139,18 +151,18 @@ struct page_cgroup *page_get_page_cgroup(struct page *page)
                (page->page_cgroup & ~PAGE_CGROUP_LOCK);
 }
 
-void __always_inline lock_page_cgroup(struct page *page)
+static void __always_inline lock_page_cgroup(struct page *page)
 {
        bit_spin_lock(PAGE_CGROUP_LOCK_BIT, &page->page_cgroup);
        VM_BUG_ON(!page_cgroup_locked(page));
 }
 
-void __always_inline unlock_page_cgroup(struct page *page)
+static void __always_inline unlock_page_cgroup(struct page *page)
 {
        bit_spin_unlock(PAGE_CGROUP_LOCK_BIT, &page->page_cgroup);
 }
 
-void __mem_cgroup_move_lists(struct page_cgroup *pc, bool active)
+static void __mem_cgroup_move_lists(struct page_cgroup *pc, bool active)
 {
        if (active)
                list_move(&pc->lru, &pc->mem_cgroup->active_list);
@@ -158,6 +170,16 @@ void __mem_cgroup_move_lists(struct page_cgroup *pc, bool active)
                list_move(&pc->lru, &pc->mem_cgroup->inactive_list);
 }
 
+int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
+{
+       int ret;
+
+       task_lock(task);
+       ret = task->mm && mm_cgroup(task->mm) == mem;
+       task_unlock(task);
+       return ret;
+}
+
 /*
  * This routine assumes that the appropriate zone's lru lock is already held
  */
@@ -244,7 +266,8 @@ unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
  * 0 if the charge was successful
  * < 0 if the cgroup is over its limit
  */
-int mem_cgroup_charge(struct page *page, struct mm_struct *mm)
+int mem_cgroup_charge(struct page *page, struct mm_struct *mm,
+                               gfp_t gfp_mask)
 {
        struct mem_cgroup *mem;
        struct page_cgroup *pc, *race_pc;
@@ -276,7 +299,7 @@ retry:
 
        unlock_page_cgroup(page);
 
-       pc = kzalloc(sizeof(struct page_cgroup), GFP_KERNEL);
+       pc = kzalloc(sizeof(struct page_cgroup), gfp_mask);
        if (pc == NULL)
                goto err;
 
@@ -302,8 +325,15 @@ retry:
         * If we created the page_cgroup, we should free it on exceeding
         * the cgroup limit.
         */
-       while (res_counter_charge(&mem->res, 1)) {
-               if (try_to_free_mem_cgroup_pages(mem))
+       while (res_counter_charge(&mem->res, PAGE_SIZE)) {
+               bool is_atomic = gfp_mask & GFP_ATOMIC;
+               /*
+                * We cannot reclaim under GFP_ATOMIC, fail the charge
+                */
+               if (is_atomic)
+                       goto noreclaim;
+
+               if (try_to_free_mem_cgroup_pages(mem, gfp_mask))
                        continue;
 
                /*
@@ -327,8 +357,10 @@ retry:
                        congestion_wait(WRITE, HZ/10);
                        continue;
                }
-
+noreclaim:
                css_put(&mem->css);
+               if (!is_atomic)
+                       mem_cgroup_out_of_memory(mem, GFP_KERNEL);
                goto free_pc;
        }
 
@@ -341,7 +373,7 @@ retry:
                kfree(pc);
                pc = race_pc;
                atomic_inc(&pc->ref_cnt);
-               res_counter_uncharge(&mem->res, 1);
+               res_counter_uncharge(&mem->res, PAGE_SIZE);
                css_put(&mem->css);
                goto done;
        }
@@ -365,6 +397,23 @@ err:
 }
 
 /*
+ * 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)
+{
+       struct mem_cgroup *mem;
+       if (!mm)
+               mm = &init_mm;
+
+       mem = rcu_dereference(mm->mem_cgroup);
+       if (mem->control_type == MEM_CGROUP_TYPE_ALL)
+               return mem_cgroup_charge(page, mm, gfp_mask);
+       else
+               return 0;
+}
+
+/*
  * Uncharging is always a welcome operation, we never complain, simply
  * uncharge.
  */
@@ -374,6 +423,10 @@ void mem_cgroup_uncharge(struct page_cgroup *pc)
        struct page *page;
        unsigned long flags;
 
+       /*
+        * This can handle cases when a page is not charged at all and we
+        * are switching between handling the control_type.
+        */
        if (!pc)
                return;
 
@@ -384,7 +437,7 @@ void mem_cgroup_uncharge(struct page_cgroup *pc)
                css_put(&mem->css);
                page_assign_page_cgroup(page, NULL);
                unlock_page_cgroup(page);
-               res_counter_uncharge(&mem->res, 1);
+               res_counter_uncharge(&mem->res, PAGE_SIZE);
 
                spin_lock_irqsave(&mem->lru_lock, flags);
                list_del_init(&pc->lru);
@@ -393,12 +446,26 @@ void mem_cgroup_uncharge(struct page_cgroup *pc)
        }
 }
 
-static ssize_t mem_cgroup_read(struct cgroup *cont, struct cftype *cft,
-                       struct file *file, char __user *userbuf, size_t nbytes,
-                       loff_t *ppos)
+int mem_cgroup_write_strategy(char *buf, unsigned long long *tmp)
+{
+       *tmp = memparse(buf, &buf);
+       if (*buf != '\0')
+               return -EINVAL;
+
+       /*
+        * Round up the value to the closest page size
+        */
+       *tmp = ((*tmp + PAGE_SIZE - 1) >> PAGE_SHIFT) << PAGE_SHIFT;
+       return 0;
+}
+
+static ssize_t mem_cgroup_read(struct cgroup *cont,
+                       struct cftype *cft, struct file *file,
+                       char __user *userbuf, size_t nbytes, loff_t *ppos)
 {
        return res_counter_read(&mem_cgroup_from_cont(cont)->res,
-                               cft->private, userbuf, nbytes, ppos);
+                               cft->private, userbuf, nbytes, ppos,
+                               NULL);
 }
 
 static ssize_t mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
@@ -406,17 +473,72 @@ static ssize_t mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
                                size_t nbytes, loff_t *ppos)
 {
        return res_counter_write(&mem_cgroup_from_cont(cont)->res,
-                               cft->private, userbuf, nbytes, ppos);
+                               cft->private, userbuf, nbytes, ppos,
+                               mem_cgroup_write_strategy);
+}
+
+static ssize_t mem_control_type_write(struct cgroup *cont,
+                       struct cftype *cft, struct file *file,
+                       const char __user *userbuf,
+                       size_t nbytes, loff_t *pos)
+{
+       int ret;
+       char *buf, *end;
+       unsigned long tmp;
+       struct mem_cgroup *mem;
+
+       mem = mem_cgroup_from_cont(cont);
+       buf = kmalloc(nbytes + 1, GFP_KERNEL);
+       ret = -ENOMEM;
+       if (buf == NULL)
+               goto out;
+
+       buf[nbytes] = 0;
+       ret = -EFAULT;
+       if (copy_from_user(buf, userbuf, nbytes))
+               goto out_free;
+
+       ret = -EINVAL;
+       tmp = simple_strtoul(buf, &end, 10);
+       if (*end != '\0')
+               goto out_free;
+
+       if (tmp <= MEM_CGROUP_TYPE_UNSPEC || tmp >= MEM_CGROUP_TYPE_MAX)
+               goto out_free;
+
+       mem->control_type = tmp;
+       ret = nbytes;
+out_free:
+       kfree(buf);
+out:
+       return ret;
+}
+
+static ssize_t mem_control_type_read(struct cgroup *cont,
+                               struct cftype *cft,
+                               struct file *file, char __user *userbuf,
+                               size_t nbytes, loff_t *ppos)
+{
+       unsigned long val;
+       char buf[64], *s;
+       struct mem_cgroup *mem;
+
+       mem = mem_cgroup_from_cont(cont);
+       s = buf;
+       val = mem->control_type;
+       s += sprintf(s, "%lu\n", val);
+       return simple_read_from_buffer((void __user *)userbuf, nbytes,
+                       ppos, buf, s - buf);
 }
 
 static struct cftype mem_cgroup_files[] = {
        {
-               .name = "usage",
+               .name = "usage_in_bytes",
                .private = RES_USAGE,
                .read = mem_cgroup_read,
        },
        {
-               .name = "limit",
+               .name = "limit_in_bytes",
                .private = RES_LIMIT,
                .write = mem_cgroup_write,
                .read = mem_cgroup_read,
@@ -426,6 +548,11 @@ static struct cftype mem_cgroup_files[] = {
                .private = RES_FAILCNT,
                .read = mem_cgroup_read,
        },
+       {
+               .name = "control_type",
+               .write = mem_control_type_write,
+               .read = mem_control_type_read,
+       },
 };
 
 static struct mem_cgroup init_mem_cgroup;
@@ -448,6 +575,7 @@ mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
        INIT_LIST_HEAD(&mem->active_list);
        INIT_LIST_HEAD(&mem->inactive_list);
        spin_lock_init(&mem->lru_lock);
+       mem->control_type = MEM_CGROUP_TYPE_ALL;
        return &mem->css;
 }