memcg: fix VM_BUG_ON from page migration

Page migration gave me free_hot_cold_page's VM_BUG_ON page->page_cgroup.
remove_migration_pte was calling mem_cgroup_charge on the new page whenever it
found a swap pte, before it had determined it to be a migration entry.  That
left a surplus reference count on the page_cgroup, so it was still attached
when the page was later freed.

Move that mem_cgroup_charge down to where we're sure it's a migration entry.
We were already under i_mmap_lock or anon_vma->lock, so its GFP_KERNEL was
already inappropriate: change that to GFP_ATOMIC.

It's essential that remove_migration_pte removes all the migration entries,
other crashes follow if not.  So proceed even when the charge fails: normally
it cannot, but after a mem_cgroup_force_empty it might - comment in the code.

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/migrate.c

@@ -153,11 +153,6 @@ static void remove_migration_pte(struct vm_area_struct *vma,
  		return;
  	}
 
-	if (mem_cgroup_charge(new, mm, GFP_KERNEL)) {
-		pte_unmap(ptep);
-		return;
-	}
-
  	ptl = pte_lockptr(mm, pmd);
  	spin_lock(ptl);
 	pte = *ptep;
@@ -169,6 +164,20 @@ static void remove_migration_pte(struct vm_area_struct *vma,
 	if (!is_migration_entry(entry) || migration_entry_to_page(entry) != old)
 		goto out;
 
+	/*
+	 * Yes, ignore the return value from a GFP_ATOMIC mem_cgroup_charge.
+	 * Failure is not an option here: we're now expected to remove every
+	 * migration pte, and will cause crashes otherwise.  Normally this
+	 * is not an issue: mem_cgroup_prepare_migration bumped up the old
+	 * page_cgroup count for safety, that's now attached to the new page,
+	 * so this charge should just be another incrementation of the count,
+	 * to keep in balance with rmap.c's mem_cgroup_uncharging.  But if
+	 * there's been a force_empty, those reference counts may no longer
+	 * be reliable, and this charge can actually fail: oh well, we don't
+	 * make the situation any worse by proceeding as if it had succeeded.
+	 */
+	mem_cgroup_charge(new, mm, GFP_ATOMIC);
+
 	get_page(new);
 	pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
 	if (is_write_migration_entry(entry))