mempool_resize() shrinks a pool by dropping pool->lock around every free_fn() call and, once the pool holds no more than new_min_nr elements, stores new_min_nr into pool->min_nr. new_min_nr was only compared against pool->min_nr before the lock was first dropped, so that store assumes nothing changed meanwhile. The kernel-doc allows mempool_alloc()/mempool_free() to run while a resize is in progress and the grow path also copes with concurrent resizes ("Raced, other resize will do our work"), but the shrink path does not: while its lock is dropped another resize may shrink the pool below new_min_nr and a third one may grow it again to a size between the two, replacing pool->elements with an array of exactly that size. Storing the stale new_min_nr afterwards leaves pool->min_nr larger than the elements array. Interleaving on a pool with min_nr = curr_nr = 10: CPU A: mempool_resize(pool, 8) takes the lock, 8 <= 10 so it shrinks: remove_element(), drops the lock, calls free_fn() and is preempted there CPU B: mempool_resize(pool, 3) completes: curr_nr = 3, min_nr = 3, elements still has 10 slots CPU B: mempool_resize(pool, 5) completes: elements = kmalloc_objs(5), min_nr = 5, curr_nr = 5 CPU A: retakes the lock, 8 < curr_nr (5) is false, stores min_nr = 8 while elements has 5 slots Every later mempool_free() sees curr_nr < min_nr and add_element() stores the returned element at elements[5], [6] and [7], past the end of the 5-slot kmalloc allocation: a heap out-of-bounds write, i.e. slab corruption, a KASAN report, or a crash in whichever unrelated object sits next to the array. The invariant that breaks is that pool->min_nr never exceeds the size of pool->elements; every other path that changes min_nr or elements keeps it. An in-tree caller can hit this: fs/smb/client resizes the global cifs_req_poolp from cifs_demultiplex_thread() (grow, when a TCP session is created) and from clean_demultiplex_info() (shrink, when one goes away). The two are coordinated only through the atomic tcpSesAllocCount, so mounts and unmounts racing on one client run the resizes above concurrently; pool->free is kmem_cache_free() there and the window is a preemption inside it. The other callers (block/blk-zoned.c, drivers/s390/scsi/zfcp_erp.c) serialise their resizes. The shrink path has stored new_min_nr unconditionally since mempool_resize() was introduced, while the grow path has always re-checked pool->min_nr after reacquiring the lock. Re-check new_min_nr against pool->min_nr under the lock before storing it. If a concurrent resize has shrunk the pool below new_min_nr in the meantime, keep that resize's min_nr (and whatever array it installed), as the grow path already does; otherwise the store is unchanged. The elements the loop has already freed stay freed, which is right either way because the pool's limit is now below new_min_nr. Found by formal verification of mm/mempool.c against the kernel binary. Reproduced with a userspace build of the affected functions in which the free_fn() callback drives the interleaving above: before the fix min_nr ends at 8 with a 5-slot array and the next mempool_free() writes past it (AddressSanitizer heap-buffer-overflow); after the fix min_nr stays at 5 and the element is handed to free_fn(). Not reproduced on a running kernel, where the window is a single preemption inside free_fn(). Assisted-by: LLM Signed-off-by: Nickolai Zeldovich --- mm/mempool.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/mm/mempool.c b/mm/mempool.c index cb74e718b2c6..95533e2975ce 100644 --- a/mm/mempool.c +++ b/mm/mempool.c @@ -378,7 +378,9 @@ int mempool_resize(struct mempool *pool, int new_min_nr) pool->free(element, pool->pool_data); spin_lock_irqsave(&pool->lock, flags); } - pool->min_nr = new_min_nr; + /* Raced, the array may now be smaller than new_min_nr */ + if (new_min_nr <= pool->min_nr) + pool->min_nr = new_min_nr; goto out_unlock; } spin_unlock_irqrestore(&pool->lock, flags); base-commit: 90201351a862182bdabfc1ad1c97b1e1b6ed5931 -- 2.53.0