memcg: fix mem_cgroup_move_lists locking
Hugh Dickins [Tue, 4 Mar 2008 22:29:13 +0000 (14:29 -0800)]
Ever since the VM_BUG_ON(page_get_page_cgroup(page)) (now Bad page state) went
into page freeing, I've hit it from time to time in testing on some machines,
sometimes only after many days.  Recently found a machine which could usually
produce it within a few hours, which got me there at last.

The culprit is mem_cgroup_move_lists, whose locking is inadequate; and the
arrangement of structures was such that you got page_cgroups from the lru list
neatly put on to SLUB's freelist.  Kamezawa-san identified the same hole
independently.

The main problem was that it was missing the lock_page_cgroup it needs to
safely page_get_page_cgroup; but it's tricky to go beyond that too, and I
couldn't do it with SLAB_DESTROY_BY_RCU as I'd expected.  See the code for
comments on the constraints.

This patch immediately gets replaced by a simpler one from Hirokazu-san; but
is it just foolish pride that tells me to put this one on record, in case we
need to come back to it later?

Signed-off-by: Hugh Dickins <hugh@veritas.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Balbir Singh <balbir@linux.vnet.ibm.com>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Hirokazu Takahashi <taka@valinux.co.jp>
Cc: YAMAMOTO Takashi <yamamoto@valinux.co.jp>
Cc: Paul Menage <menage@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/memcontrol.c

index 66d0e84..dcbe30a 100644 (file)
@@ -277,6 +277,11 @@ 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);
@@ -348,17 +353,49 @@ int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
 void mem_cgroup_move_lists(struct page *page, bool active)
 {
        struct page_cgroup *pc;
+       struct mem_cgroup *mem;
        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);
+       /*
+        * Now page_cgroup is stable, but we cannot acquire mz->lru_lock
+        * while holding it, because mem_cgroup_force_empty_list does the
+        * reverse.  Get a hold on the mem_cgroup before unlocking, so that
+        * the zoneinfo remains stable, then take mz->lru_lock; then check
+        * that page still points to pc and pc (even if freed and reassigned
+        * to that same page meanwhile) still points to the same mem_cgroup.
+        * Then we know mz still points to the right spinlock, so it's safe
+        * to move_lists (page->page_cgroup might be reset while we do so, but
+        * that doesn't matter: pc->page is stable till we drop mz->lru_lock).
+        * We're being a little naughty not to try_lock_page_cgroup again
+        * inside there, but we are safe, aren't we?  Aren't we?  Whistle...
+        */
+       pc = page_get_page_cgroup(page);
+       if (pc) {
+               mem = pc->mem_cgroup;
+               mz = page_cgroup_zoneinfo(pc);
+               css_get(&mem->css);
+
+               unlock_page_cgroup(page);
+
+               spin_lock_irqsave(&mz->lru_lock, flags);
+               if (page_get_page_cgroup(page) == pc && pc->mem_cgroup == mem)
+                       __mem_cgroup_move_lists(pc, active);
+               spin_unlock_irqrestore(&mz->lru_lock, flags);
+
+               css_put(&mem->css);
+       } else
+               unlock_page_cgroup(page);
 }
 
 /*