memcg: move all acccounting to parent at rmdir()
authorKAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Thu, 8 Jan 2009 02:07:53 +0000 (18:07 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Thu, 8 Jan 2009 16:31:04 +0000 (08:31 -0800)
This patch provides a function to move account information of a page
between mem_cgroups and rewrite force_empty to make use of this.

This moving of page_cgroup is done under
 - lru_lock of source/destination mem_cgroup is held.
 - lock_page_cgroup() is held.

Then, a routine which touches pc->mem_cgroup without lock_page_cgroup()
should confirm pc->mem_cgroup is still valid or not.  Typical code can be
following.

(while page is not under lock_page())
mem = pc->mem_cgroup;
mz = page_cgroup_zoneinfo(pc)
spin_lock_irqsave(&mz->lru_lock);
if (pc->mem_cgroup == mem)
...../* some list handling */
spin_unlock_irqrestore(&mz->lru_lock);

Of course, better way is
lock_page_cgroup(pc);
....
unlock_page_cgroup(pc);

But you should confirm the nest of lock and avoid deadlock.

If you treats page_cgroup from mem_cgroup's LRU under mz->lru_lock,
you don't have to worry about what pc->mem_cgroup points to.
moved pages are added to head of lru, not to tail.

Expected users of this routine is:
  - force_empty (rmdir)
  - moving tasks between cgroup (for moving account information.)
  - hierarchy (maybe useful.)

force_empty(rmdir) uses this move_account and move pages to its parent.
This "move" will not cause OOM (I added "oom" parameter to try_charge().)

If the parent is busy (not enough memory), force_empty calls try_to_free_page()
and reduce usage.

Purpose of this behavior is
  - Fix "forget all" behavior of force_empty and avoid leak of accounting.
  - By "moving first, free if necessary", keep pages on memory as much as
    possible.

Adding a switch to change behavior of force_empty to
  - free first, move if necessary
  - free all, if there is mlocked/busy pages, return -EBUSY.
is under consideration. (I'll add if someone requtests.)

This patch also removes memory.force_empty file, a brutal debug-only interface.

Reviewed-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Tested-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Balbir Singh <balbir@in.ibm.com>
Cc: Paul Menage <menage@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Documentation/controllers/memory.txt
mm/memcontrol.c

index 1c07547d3f81f28a19edb9de8752f8865b44478d..58f32c166facdf2e8d96a4c1a81d2ae018f23d3d 100644 (file)
@@ -207,12 +207,6 @@ exceeded.
 The memory.stat file gives accounting information. Now, the number of
 caches, RSS and Active pages/Inactive pages are shown.
 
-The memory.force_empty gives an interface to drop *all* charges by force.
-
-# echo 1 > memory.force_empty
-
-will drop all charges in cgroup. Currently, this is maintained for test.
-
 4. Testing
 
 Balbir posted lmbench, AIM9, LTP and vmmstress results [10] and [11].
@@ -242,8 +236,10 @@ reclaimed.
 
 A cgroup can be removed by rmdir, but as discussed in sections 4.1 and 4.2, a
 cgroup might have some charge associated with it, even though all
-tasks have migrated away from it. Such charges are automatically dropped at
-rmdir() if there are no tasks.
+tasks have migrated away from it.
+Such charges are moved to its parent as much as possible and freed if parent
+is full. Both of RSS and CACHES are moved to parent.
+If both of them are busy, rmdir() returns -EBUSY.
 
 5. TODO
 
index b71195e8198b57dd7c782b0914befe74724d36be..49234d93988acf13b2a699408ed6200be2a5b6aa 100644 (file)
@@ -257,7 +257,7 @@ static void __mem_cgroup_remove_list(struct mem_cgroup_per_zone *mz,
 }
 
 static void __mem_cgroup_add_list(struct mem_cgroup_per_zone *mz,
-                               struct page_cgroup *pc)
+                               struct page_cgroup *pc, bool hot)
 {
        int lru = LRU_BASE;
 
@@ -271,7 +271,10 @@ static void __mem_cgroup_add_list(struct mem_cgroup_per_zone *mz,
        }
 
        MEM_CGROUP_ZSTAT(mz, lru) += 1;
-       list_add(&pc->lru, &mz->lists[lru]);
+       if (hot)
+               list_add(&pc->lru, &mz->lists[lru]);
+       else
+               list_add_tail(&pc->lru, &mz->lists[lru]);
 
        mem_cgroup_charge_statistics(pc->mem_cgroup, pc, true);
 }
@@ -467,21 +470,12 @@ unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
        return nr_taken;
 }
 
-
-/**
- * mem_cgroup_try_charge - get charge of PAGE_SIZE.
- * @mm: an mm_struct which is charged against. (when *memcg is NULL)
- * @gfp_mask: gfp_mask for reclaim.
- * @memcg: a pointer to memory cgroup which is charged against.
- *
- * charge against memory cgroup pointed by *memcg. if *memcg == NULL, estimated
- * memory cgroup from @mm is got and stored in *memcg.
- *
- * Returns 0 if success. -ENOMEM at failure.
+/*
+ * Unlike exported interface, "oom" parameter is added. if oom==true,
+ * oom-killer can be invoked.
  */
-
-int mem_cgroup_try_charge(struct mm_struct *mm,
-                       gfp_t gfp_mask, struct mem_cgroup **memcg)
+static int __mem_cgroup_try_charge(struct mm_struct *mm,
+                       gfp_t gfp_mask, struct mem_cgroup **memcg, bool oom)
 {
        struct mem_cgroup *mem;
        int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
@@ -528,7 +522,8 @@ int mem_cgroup_try_charge(struct mm_struct *mm,
                        continue;
 
                if (!nr_retries--) {
-                       mem_cgroup_out_of_memory(mem, gfp_mask);
+                       if (oom)
+                               mem_cgroup_out_of_memory(mem, gfp_mask);
                        goto nomem;
                }
        }
@@ -538,6 +533,25 @@ nomem:
        return -ENOMEM;
 }
 
+/**
+ * mem_cgroup_try_charge - get charge of PAGE_SIZE.
+ * @mm: an mm_struct which is charged against. (when *memcg is NULL)
+ * @gfp_mask: gfp_mask for reclaim.
+ * @memcg: a pointer to memory cgroup which is charged against.
+ *
+ * charge against memory cgroup pointed by *memcg. if *memcg == NULL, estimated
+ * memory cgroup from @mm is got and stored in *memcg.
+ *
+ * Returns 0 if success. -ENOMEM at failure.
+ * This call can invoke OOM-Killer.
+ */
+
+int mem_cgroup_try_charge(struct mm_struct *mm,
+                         gfp_t mask, struct mem_cgroup **memcg)
+{
+       return __mem_cgroup_try_charge(mm, mask, memcg, true);
+}
+
 /*
  * commit a charge got by mem_cgroup_try_charge() and makes page_cgroup to be
  * USED state. If already USED, uncharge and return.
@@ -571,11 +585,109 @@ static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
        mz = page_cgroup_zoneinfo(pc);
 
        spin_lock_irqsave(&mz->lru_lock, flags);
-       __mem_cgroup_add_list(mz, pc);
+       __mem_cgroup_add_list(mz, pc, true);
        spin_unlock_irqrestore(&mz->lru_lock, flags);
        unlock_page_cgroup(pc);
 }
 
+/**
+ * mem_cgroup_move_account - move account of the page
+ * @pc:        page_cgroup of the page.
+ * @from: mem_cgroup which the page is moved from.
+ * @to:        mem_cgroup which the page is moved to. @from != @to.
+ *
+ * The caller must confirm following.
+ * 1. disable irq.
+ * 2. lru_lock of old mem_cgroup(@from) should be held.
+ *
+ * returns 0 at success,
+ * returns -EBUSY when lock is busy or "pc" is unstable.
+ *
+ * This function does "uncharge" from old cgroup but doesn't do "charge" to
+ * new cgroup. It should be done by a caller.
+ */
+
+static int mem_cgroup_move_account(struct page_cgroup *pc,
+       struct mem_cgroup *from, struct mem_cgroup *to)
+{
+       struct mem_cgroup_per_zone *from_mz, *to_mz;
+       int nid, zid;
+       int ret = -EBUSY;
+
+       VM_BUG_ON(!irqs_disabled());
+       VM_BUG_ON(from == to);
+
+       nid = page_cgroup_nid(pc);
+       zid = page_cgroup_zid(pc);
+       from_mz =  mem_cgroup_zoneinfo(from, nid, zid);
+       to_mz =  mem_cgroup_zoneinfo(to, nid, zid);
+
+
+       if (!trylock_page_cgroup(pc))
+               return ret;
+
+       if (!PageCgroupUsed(pc))
+               goto out;
+
+       if (pc->mem_cgroup != from)
+               goto out;
+
+       if (spin_trylock(&to_mz->lru_lock)) {
+               __mem_cgroup_remove_list(from_mz, pc);
+               css_put(&from->css);
+               res_counter_uncharge(&from->res, PAGE_SIZE);
+               pc->mem_cgroup = to;
+               css_get(&to->css);
+               __mem_cgroup_add_list(to_mz, pc, false);
+               ret = 0;
+               spin_unlock(&to_mz->lru_lock);
+       }
+out:
+       unlock_page_cgroup(pc);
+       return ret;
+}
+
+/*
+ * move charges to its parent.
+ */
+
+static int mem_cgroup_move_parent(struct page_cgroup *pc,
+                                 struct mem_cgroup *child,
+                                 gfp_t gfp_mask)
+{
+       struct cgroup *cg = child->css.cgroup;
+       struct cgroup *pcg = cg->parent;
+       struct mem_cgroup *parent;
+       struct mem_cgroup_per_zone *mz;
+       unsigned long flags;
+       int ret;
+
+       /* Is ROOT ? */
+       if (!pcg)
+               return -EINVAL;
+
+       parent = mem_cgroup_from_cont(pcg);
+
+       ret = __mem_cgroup_try_charge(NULL, gfp_mask, &parent, false);
+       if (ret)
+               return ret;
+
+       mz = mem_cgroup_zoneinfo(child,
+                       page_cgroup_nid(pc), page_cgroup_zid(pc));
+
+       spin_lock_irqsave(&mz->lru_lock, flags);
+       ret = mem_cgroup_move_account(pc, child, parent);
+       spin_unlock_irqrestore(&mz->lru_lock, flags);
+
+       /* drop extra refcnt */
+       css_put(&parent->css);
+       /* uncharge if move fails */
+       if (ret)
+               res_counter_uncharge(&parent->res, PAGE_SIZE);
+
+       return ret;
+}
+
 /*
  * Charge the memory controller for page usage.
  * Return
@@ -597,7 +709,7 @@ static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
        prefetchw(pc);
 
        mem = memcg;
-       ret = mem_cgroup_try_charge(mm, gfp_mask, &mem);
+       ret = __mem_cgroup_try_charge(mm, gfp_mask, &mem, true);
        if (ret)
                return ret;
 
@@ -899,46 +1011,52 @@ static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
  * This routine traverse page_cgroup in given list and drop them all.
  * *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 int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
                            struct mem_cgroup_per_zone *mz,
                            enum lru_list lru)
 {
-       struct page_cgroup *pc;
-       struct page *page;
-       int count = FORCE_UNCHARGE_BATCH;
+       struct page_cgroup *pc, *busy;
        unsigned long flags;
+       unsigned long loop;
        struct list_head *list;
+       int ret = 0;
 
        list = &mz->lists[lru];
 
-       spin_lock_irqsave(&mz->lru_lock, flags);
-       while (!list_empty(list)) {
-               pc = list_entry(list->prev, struct page_cgroup, lru);
-               page = pc->page;
-               if (!PageCgroupUsed(pc))
+       loop = MEM_CGROUP_ZSTAT(mz, lru);
+       /* give some margin against EBUSY etc...*/
+       loop += 256;
+       busy = NULL;
+       while (loop--) {
+               ret = 0;
+               spin_lock_irqsave(&mz->lru_lock, flags);
+               if (list_empty(list)) {
+                       spin_unlock_irqrestore(&mz->lru_lock, flags);
                        break;
-               get_page(page);
+               }
+               pc = list_entry(list->prev, struct page_cgroup, lru);
+               if (busy == pc) {
+                       list_move(&pc->lru, list);
+                       busy = 0;
+                       spin_unlock_irqrestore(&mz->lru_lock, flags);
+                       continue;
+               }
                spin_unlock_irqrestore(&mz->lru_lock, flags);
-               /*
-                * Check if this page is on LRU. !LRU page can be found
-                * if it's under page migration.
-                */
-               if (PageLRU(page)) {
-                       __mem_cgroup_uncharge_common(page,
-                                       MEM_CGROUP_CHARGE_TYPE_FORCE);
-                       put_page(page);
-                       if (--count <= 0) {
-                               count = FORCE_UNCHARGE_BATCH;
-                               cond_resched();
-                       }
-               } else {
-                       spin_lock_irqsave(&mz->lru_lock, flags);
+
+               ret = mem_cgroup_move_parent(pc, mem, GFP_HIGHUSER_MOVABLE);
+               if (ret == -ENOMEM)
                        break;
-               }
-               spin_lock_irqsave(&mz->lru_lock, flags);
+
+               if (ret == -EBUSY || ret == -EINVAL) {
+                       /* found lock contention or "pc" is obsolete. */
+                       busy = pc;
+                       cond_resched();
+               } else
+                       busy = NULL;
        }
-       spin_unlock_irqrestore(&mz->lru_lock, flags);
+       if (!ret && !list_empty(list))
+               return -EBUSY;
+       return ret;
 }
 
 /*
@@ -947,34 +1065,68 @@ static void mem_cgroup_force_empty_list(struct mem_cgroup *mem,
  */
 static int mem_cgroup_force_empty(struct mem_cgroup *mem)
 {
-       int ret = -EBUSY;
-       int node, zid;
+       int ret;
+       int node, zid, shrink;
+       int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
 
        css_get(&mem->css);
-       /*
-        * page reclaim code (kswapd etc..) will move pages between
-        * active_list <-> inactive_list while we don't take a lock.
-        * So, we have to do loop here until all lists are empty.
-        */
+
+       shrink = 0;
+move_account:
        while (mem->res.usage > 0) {
+               ret = -EBUSY;
                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++) {
+               ret = 0;
+               for_each_node_state(node, N_POSSIBLE) {
+                       for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
                                struct mem_cgroup_per_zone *mz;
                                enum lru_list l;
                                mz = mem_cgroup_zoneinfo(mem, node, zid);
-                               for_each_lru(l)
-                                       mem_cgroup_force_empty_list(mem, mz, l);
+                               for_each_lru(l) {
+                                       ret = mem_cgroup_force_empty_list(mem,
+                                                                 mz, l);
+                                       if (ret)
+                                               break;
+                               }
                        }
+                       if (ret)
+                               break;
+               }
+               /* it seems parent cgroup doesn't have enough mem */
+               if (ret == -ENOMEM)
+                       goto try_to_free;
                cond_resched();
        }
        ret = 0;
 out:
        css_put(&mem->css);
        return ret;
+
+try_to_free:
+       /* returns EBUSY if we come here twice. */
+       if (shrink)  {
+               ret = -EBUSY;
+               goto out;
+       }
+       /* try to free all pages in this cgroup */
+       shrink = 1;
+       while (nr_retries && mem->res.usage > 0) {
+               int progress;
+               progress = try_to_free_mem_cgroup_pages(mem,
+                                                 GFP_HIGHUSER_MOVABLE);
+               if (!progress)
+                       nr_retries--;
+
+       }
+       /* try move_account...there may be some *locked* pages. */
+       if (mem->res.usage)
+               goto move_account;
+       ret = 0;
+       goto out;
 }
 
 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
@@ -1023,11 +1175,6 @@ static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
        return 0;
 }
 
-static int mem_force_empty_write(struct cgroup *cont, unsigned int event)
-{
-       return mem_cgroup_force_empty(mem_cgroup_from_cont(cont));
-}
-
 static const struct mem_cgroup_stat_desc {
        const char *msg;
        u64 unit;
@@ -1103,10 +1250,6 @@ static struct cftype mem_cgroup_files[] = {
                .trigger = mem_cgroup_reset,
                .read_u64 = mem_cgroup_read,
        },
-       {
-               .name = "force_empty",
-               .trigger = mem_force_empty_write,
-       },
        {
                .name = "stat",
                .read_map = mem_control_stat_show,