drm/ttm: Fix possible stack overflow by recursive shrinker calls.

commit 71336e01 upstream.

While ttm_dma_pool_shrink_scan() tries to take mutex before doing GFP_KERNEL
allocation, ttm_pool_shrink_scan() does not do it. This can result in stack
overflow if kmalloc() in ttm_page_pool_free() triggered recursion due to
memory pressure.

  shrink_slab()
  => ttm_pool_shrink_scan()
     => ttm_page_pool_free()
        => kmalloc(GFP_KERNEL)
           => shrink_slab()
              => ttm_pool_shrink_scan()
                 => ttm_page_pool_free()
                    => kmalloc(GFP_KERNEL)

Change ttm_pool_shrink_scan() to do like ttm_dma_pool_shrink_scan() does.
Signed-off-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Signed-off-by: Dave Airlie <airlied@redhat.com>
[bwh: Backported to 3.2:
 - Adjust context
 - Change return value in the contended case to follow the old shrinker
   API]   
Signed-off-by: Ben Hutchings <ben@decadent.org.uk>

drivers/gpu/drm/ttm/ttm_page_alloc.c

@@ -398,13 +398,18 @@ static int ttm_pool_get_num_unused_pages(void)
 static int ttm_pool_mm_shrink(struct shrinker *shrink,
 			      struct shrink_control *sc)
 {
-	static atomic_t start_pool = ATOMIC_INIT(0);
+	static DEFINE_MUTEX(lock);
+	static unsigned start_pool;
 	unsigned i;
-	unsigned pool_offset = atomic_add_return(1, &start_pool);
+	unsigned pool_offset;
 	struct ttm_page_pool *pool;
 	int shrink_pages = sc->nr_to_scan;
 
-	pool_offset = pool_offset % NUM_POOLS;
+	if (shrink_pages == 0)
+		goto out;
+	if (!mutex_trylock(&lock))
+		return -1;
+	pool_offset = ++start_pool % NUM_POOLS;
 	/* select start pool in round robin fashion */
 	for (i = 0; i < NUM_POOLS; ++i) {
 		unsigned nr_free = shrink_pages;
@@ -413,6 +418,8 @@ static int ttm_pool_mm_shrink(struct shrinker *shrink,
 		pool = &_manager->pools[(i + pool_offset)%NUM_POOLS];
 		shrink_pages = ttm_page_pool_free(pool, nr_free);
 	}
+	mutex_unlock(&lock);
+out:
 	/* return estimated number of unused pages in pool */
 	return ttm_pool_get_num_unused_pages();
 }