On Fri, Jul 30, 2021 at 11:49 AM Heiko Carstens <hca@xxxxxxxxxxxxx> wrote: > On Tue, Jul 27, 2021 at 04:48:53PM +0200, Arnd Bergmann wrote: > > Our CI reports this with linux-next and running strace selftest in > compat mode: Thanks a lot for the report! I managed track it down based on your output, it turns out that I end up copying data from the stack according to how much the user asked for, and in this case that was much more than the 8 byte nodemask_t, copying all of the kernel stack all the way into the guard page with CONFIG_VMAP_STACK, where it crashed. Without CONFIG_VMAP_STACK, or with user space that asks for less data, it would just be an information leak, so others probably haven't noticed the problem. The change below should fix that, I'll double-check the other callers as well before sending a proper fixup patch to Andrew. Arnd diff --git a/mm/mempolicy.c b/mm/mempolicy.c index 4fabf2dddbc0..0d1f3be32723 100644 --- a/mm/mempolicy.c +++ b/mm/mempolicy.c @@ -1438,6 +1438,7 @@ static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode, if (clear_user((char __user *)mask + nbytes, copy - nbytes)) return -EFAULT; copy = nbytes; + maxnode = nr_node_ids; } if (compat)