memcg: fix mem_cgroup_move_lists locking

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 <[e-mail]>
Cc: David Rientjes <[e-mail]>
Cc: Balbir Singh <[e-mail]>
Acked-by: KAMEZAWA Hiroyuki <[e-mail]>
Cc: Hirokazu Takahashi <[e-mail]>
Cc: YAMAMOTO Takashi <[e-mail]>
Cc: Paul Menage <[e-mail]>
Signed-off-by: Andrew Morton <[e-mail]>
Signed-off-by: Linus Torvalds <[e-mail]>

mm/memcontrol.c

···
 	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);
···
 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);
 }
 
 /*