Description of problem:
This doesn't look correct, the io_submit systemcall calls do_io_submit() directly:

SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
        struct iocb __user * __user *, iocbpp)
{
    return do_io_submit(ctx_id, nr, iocbpp, 0);
}

do_io_submit only checks if nr < 0, but doesnt check if the access_ok
multiplication will overflow.

long do_io_submit(aio_context_t ctx_id, long nr,
          struct iocb __user *__user *iocbpp, bool compat)
{
...
    if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
        return -EFAULT;
...

Then it uses this loop with __get_user_nocheck(), which doesn't look safe.

...
    for (i=0; i<nr; i++) {
        struct iocb __user *user_iocb;
        struct iocb tmp;

        if (unlikely(__get_user(user_iocb, iocbpp + i))) {
            ret = -EFAULT;
            break;
        }
...

Acknowledgements:

Red Hat would like to thank Tavis Ormandy for reporting this issue.