bugfix for memory cgroup controller: avoid !PageLRU page in mem_cgroup_isolate_pages
[linux-2.6.git] / mm / memcontrol.c
index 4d4805e..e8493fb 100644 (file)
 #include <linux/memcontrol.h>
 #include <linux/cgroup.h>
 #include <linux/mm.h>
+#include <linux/page-flags.h>
+#include <linux/backing-dev.h>
+#include <linux/bit_spinlock.h>
+#include <linux/rcupdate.h>
+#include <linux/swap.h>
+#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;
 
 /*
  * The memory controller data structure. The memory controller controls both
@@ -31,7 +41,9 @@ struct cgroup_subsys mem_cgroup_subsys;
  * to help the administrator determine what knobs to tune.
  *
  * TODO: Add a water mark for the memory controller. Reclaim will begin when
- * we hit the water mark.
+ * we hit the water mark. May be even add a low water mark, such that
+ * no reclaim occurs from a cgroup at it's low water mark, this is
+ * a feature that will be implemented much later in the future.
  */
 struct mem_cgroup {
        struct cgroup_subsys_state css;
@@ -46,9 +58,22 @@ struct mem_cgroup {
         */
        struct list_head active_list;
        struct list_head inactive_list;
+       /*
+        * spin_lock to protect the per cgroup LRU
+        */
+       spinlock_t lru_lock;
+       unsigned long control_type;     /* control RSS or RSS+Pagecache */
 };
 
 /*
+ * 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 atleast two
+ * byte aligned (based on comments from Nick Piggin)
+ */
+#define PAGE_CGROUP_LOCK_BIT   0x0
+#define PAGE_CGROUP_LOCK               (1 << PAGE_CGROUP_LOCK_BIT)
+
+/*
  * A page_cgroup page is associated with every page descriptor. The
  * page_cgroup helps us identify information about the cgroup
  */
@@ -56,8 +81,19 @@ 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     */
 };
 
+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)
@@ -88,22 +124,445 @@ void mm_free_cgroup(struct mm_struct *mm)
        css_put(&mm->mem_cgroup->css);
 }
 
+static inline int page_cgroup_locked(struct page *page)
+{
+       return bit_spin_is_locked(PAGE_CGROUP_LOCK_BIT,
+                                       &page->page_cgroup);
+}
+
 void page_assign_page_cgroup(struct page *page, struct page_cgroup *pc)
 {
-       page->page_cgroup = (unsigned long)pc;
+       int locked;
+
+       /*
+        * While resetting the page_cgroup we might not hold the
+        * page_cgroup lock. free_hot_cold_page() is an example
+        * of such a scenario
+        */
+       if (pc)
+               VM_BUG_ON(!page_cgroup_locked(page));
+       locked = (page->page_cgroup & PAGE_CGROUP_LOCK);
+       page->page_cgroup = ((unsigned long)pc | locked);
 }
 
 struct page_cgroup *page_get_page_cgroup(struct page *page)
 {
-       return page->page_cgroup;
+       return (struct page_cgroup *)
+               (page->page_cgroup & ~PAGE_CGROUP_LOCK);
+}
+
+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));
+}
+
+static void __always_inline unlock_page_cgroup(struct page *page)
+{
+       bit_spin_unlock(PAGE_CGROUP_LOCK_BIT, &page->page_cgroup);
+}
+
+/*
+ * Tie new page_cgroup to struct page under lock_page_cgroup()
+ * This can fail if the page has been tied to a page_cgroup.
+ * If success, returns 0.
+ */
+static inline int
+page_cgroup_assign_new_page_cgroup(struct page *page, struct page_cgroup *pc)
+{
+       int ret = 0;
+
+       lock_page_cgroup(page);
+       if (!page_get_page_cgroup(page))
+               page_assign_page_cgroup(page, pc);
+       else /* A page is tied to other pc. */
+               ret = 1;
+       unlock_page_cgroup(page);
+       return ret;
+}
+
+/*
+ * 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 inline struct page_cgroup *
+clear_page_cgroup(struct page *page, struct page_cgroup *pc)
+{
+       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;
+}
+
+
+static void __mem_cgroup_move_lists(struct page_cgroup *pc, bool active)
+{
+       if (active)
+               list_move(&pc->lru, &pc->mem_cgroup->active_list);
+       else
+               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
+ */
+void mem_cgroup_move_lists(struct page_cgroup *pc, bool active)
+{
+       struct mem_cgroup *mem;
+       if (!pc)
+               return;
+
+       mem = pc->mem_cgroup;
+
+       spin_lock(&mem->lru_lock);
+       __mem_cgroup_move_lists(pc, active);
+       spin_unlock(&mem->lru_lock);
+}
+
+unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
+                                       struct list_head *dst,
+                                       unsigned long *scanned, int order,
+                                       int mode, struct zone *z,
+                                       struct mem_cgroup *mem_cont,
+                                       int active)
+{
+       unsigned long nr_taken = 0;
+       struct page *page;
+       unsigned long scan;
+       LIST_HEAD(pc_list);
+       struct list_head *src;
+       struct page_cgroup *pc, *tmp;
+
+       if (active)
+               src = &mem_cont->active_list;
+       else
+               src = &mem_cont->inactive_list;
+
+       spin_lock(&mem_cont->lru_lock);
+       scan = 0;
+       list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
+               if (scan++ > nr_to_scan)
+                       break;
+               page = pc->page;
+               VM_BUG_ON(!pc);
+
+               if (unlikely(!PageLRU(page))) {
+                       scan--;
+                       continue;
+               }
+
+               if (PageActive(page) && !active) {
+                       __mem_cgroup_move_lists(pc, true);
+                       scan--;
+                       continue;
+               }
+               if (!PageActive(page) && active) {
+                       __mem_cgroup_move_lists(pc, false);
+                       scan--;
+                       continue;
+               }
+
+               /*
+                * Reclaim, per zone
+                * TODO: make the active/inactive lists per zone
+                */
+               if (page_zone(page) != z)
+                       continue;
+
+               /*
+                * Check if the meta page went away from under us
+                */
+               if (!list_empty(&pc->lru))
+                       list_move(&pc->lru, &pc_list);
+               else
+                       continue;
+
+               if (__isolate_lru_page(page, mode) == 0) {
+                       list_move(&page->lru, dst);
+                       nr_taken++;
+               }
+       }
+
+       list_splice(&pc_list, src);
+       spin_unlock(&mem_cont->lru_lock);
+
+       *scanned = scan;
+       return nr_taken;
+}
+
+/*
+ * Charge the memory controller for page usage.
+ * Return
+ * 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,
+                               gfp_t gfp_mask)
+{
+       struct mem_cgroup *mem;
+       struct page_cgroup *pc;
+       unsigned long flags;
+       unsigned long nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
+
+       /*
+        * Should page_cgroup's go to their own slab?
+        * One could optimize the performance of the charging routine
+        * by saving a bit in the page_flags and using it as a lock
+        * to see if the cgroup page already has a page_cgroup associated
+        * with it
+        */
+retry:
+       lock_page_cgroup(page);
+       pc = page_get_page_cgroup(page);
+       /*
+        * The page_cgroup exists and 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;
+               }
+       }
+
+       unlock_page_cgroup(page);
+
+       pc = kzalloc(sizeof(struct page_cgroup), gfp_mask);
+       if (pc == NULL)
+               goto err;
+
+       rcu_read_lock();
+       /*
+        * 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 (!mm)
+               mm = &init_mm;
+
+       mem = rcu_dereference(mm->mem_cgroup);
+       /*
+        * 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)) {
+               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;
+
+               /*
+                * 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;
+                       /*
+                        * Since we control both RSS and cache, we end up with a
+                        * very interesting scenario where we end up reclaiming
+                        * memory (essentially RSS), since the memory is pushed
+                        * to swap cache, we eventually end up adding those
+                        * pages back to our list. Hence we give ourselves a
+                        * few chances before we fail
+                        */
+               else if (nr_retries--) {
+                       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;
+       }
+
+       atomic_set(&pc->ref_cnt, 1);
+       pc->mem_cgroup = mem;
+       pc->page = page;
+       if (page_cgroup_assign_new_page_cgroup(page, pc)) {
+               /*
+                * an another charge is added to this page already.
+                * we do take lock_page_cgroup(page) again and read
+                * page->cgroup, increment refcnt.... just retry is OK.
+                */
+               res_counter_uncharge(&mem->res, PAGE_SIZE);
+               css_put(&mem->css);
+               kfree(pc);
+               goto retry;
+       }
+
+       spin_lock_irqsave(&mem->lru_lock, flags);
+       list_add(&pc->lru, &mem->active_list);
+       spin_unlock_irqrestore(&mem->lru_lock, flags);
+
+done:
+       return 0;
+free_pc:
+       kfree(pc);
+err:
+       return -ENOMEM;
+}
+
+/*
+ * 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.
+ */
+void mem_cgroup_uncharge(struct page_cgroup *pc)
+{
+       struct mem_cgroup *mem;
+       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;
+
+       if (atomic_dec_and_test(&pc->ref_cnt)) {
+               page = pc->page;
+               /*
+                * get page->cgroup and clear it under lock.
+                */
+               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(&mem->lru_lock, flags);
+                       list_del_init(&pc->lru);
+                       spin_unlock_irqrestore(&mem->lru_lock, flags);
+                       kfree(pc);
+               } else {
+                       /*
+                        * Note:This will be removed when force-empty patch is
+                        * applied. just show warning here.
+                        */
+                       printk(KERN_ERR "Race in mem_cgroup_uncharge() ?");
+                       dump_stack();
+               }
+       }
+}
+/*
+ * 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;
+       unlock_page_cgroup(page);
+       return ret;
+}
+
+void mem_cgroup_end_migration(struct page *page)
+{
+       struct page_cgroup *pc = page_get_page_cgroup(page);
+       mem_cgroup_uncharge(pc);
+}
+/*
+ * 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;
+retry:
+       pc = page_get_page_cgroup(page);
+       if (!pc)
+               return;
+       if (clear_page_cgroup(page, pc) != pc)
+               goto retry;
+       pc->page = newpage;
+       lock_page_cgroup(newpage);
+       page_assign_page_cgroup(newpage, pc);
+       unlock_page_cgroup(newpage);
+       return;
+}
+
+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)
+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,
@@ -111,17 +570,72 @@ 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,
@@ -131,6 +645,11 @@ 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;
@@ -150,6 +669,10 @@ mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
                return NULL;
 
        res_counter_init(&mem->res);
+       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;
 }
 
@@ -166,11 +689,46 @@ static int mem_cgroup_populate(struct cgroup_subsys *ss,
                                        ARRAY_SIZE(mem_cgroup_files));
 }
 
+static void mem_cgroup_move_task(struct cgroup_subsys *ss,
+                               struct cgroup *cont,
+                               struct cgroup *old_cont,
+                               struct task_struct *p)
+{
+       struct mm_struct *mm;
+       struct mem_cgroup *mem, *old_mem;
+
+       mm = get_task_mm(p);
+       if (mm == NULL)
+               return;
+
+       mem = mem_cgroup_from_cont(cont);
+       old_mem = mem_cgroup_from_cont(old_cont);
+
+       if (mem == old_mem)
+               goto out;
+
+       /*
+        * Only thread group leaders are allowed to migrate, the mm_struct is
+        * in effect owned by the leader
+        */
+       if (p->tgid != p->pid)
+               goto out;
+
+       css_get(&mem->css);
+       rcu_assign_pointer(mm->mem_cgroup, mem);
+       css_put(&old_mem->css);
+
+out:
+       mmput(mm);
+       return;
+}
+
 struct cgroup_subsys mem_cgroup_subsys = {
        .name = "memory",
        .subsys_id = mem_cgroup_subsys_id,
        .create = mem_cgroup_create,
        .destroy = mem_cgroup_destroy,
        .populate = mem_cgroup_populate,
+       .attach = mem_cgroup_move_task,
        .early_init = 1,
 };