On Mon, Oct 31, 2022 at 04:00:25PM +0100, Vlastimil Babka wrote:
> +++ b/mm/mempool.c
> @@ -57,8 +57,10 @@ static void __check_element(mempool_t *pool, void *element, size_t size)
> static void check_element(mempool_t *pool, void *element)
> {
> /* Mempools backed by slab allocator */
> - if (pool->free == mempool_free_slab || pool->free == mempool_kfree) {
> + if (pool->free == mempool_kfree) {
> __check_element(pool, element, (size_t)pool->pool_data);
> + } else if (pool->free == mempool_free_slab) {
> + __check_element(pool, element, kmem_cache_size(pool->pool_data));
> } else if (pool->free == mempool_free_pages) {
> /* Mempools backed by page allocator */
> int order = (int)(long)pool->pool_data;
I had a quick look at this to be sure I understood what was going on,
and I found a grotesque bug that has been with us since the introduction
of check_element() in 2015.
+ if (pool->free == mempool_free_pages) {
+ int order = (int)(long)pool->pool_data;
+ void *addr = kmap_atomic((struct page *)element);
+
+ __check_element(pool, addr, 1UL << (PAGE_SHIFT + order));
+ kunmap_atomic(addr);
kmap_atomic() and friends only map a single page. So this is all
nonsense for HIGHMEM kernels, GFP_HIGHMEM allocations and order > 0.
The consequence of doing that will be calling memset(POISON_INUSE)
on random pages that we don't own.
