On Tue, Jun 11, 2019 at 11:56:06PM +0000, Stephen Bates wrote:
> The array_index_nospec() check in io_sq_offload_start() is performed
> before any checks on p->sq_thread_cpu are done. This means cpu is
> clamped and therefore no error occurs when out-of-range values are
> passed in from userspace. This is in violation of the specification
> for io_ring_setup() and causes the io_ring_setup unit test in liburing
> to regress.
>
> Add a new bounds check on sq_thread_cpu at the start of
> io_sq_offload_start() so we can exit the function early when bad
> values are passed in.
>
> Fixes: 975554b03edd ("io_uring: fix SQPOLL cpu validation")
> Signed-off-by: Stephen Bates <sbates@xxxxxxxxxxxx>
Aargh. My original patch [1] handled that correctly, and this case was
explicitly called out in the commit message, which was retained even
when the patch was "simplified". That's rather disappointing. :/
Thanks,
Mark.
[1] https://lore.kernel.org/lkml/20190430123451.44227-1-mark.rutland@xxxxxxx/
> ---
> fs/io_uring.c | 3 +++
> 1 file changed, 3 insertions(+)
>
> diff --git a/fs/io_uring.c b/fs/io_uring.c
> index 30a5687..e458470 100644
> --- a/fs/io_uring.c
> +++ b/fs/io_uring.c
> @@ -2316,6 +2316,9 @@ static int io_sq_offload_start(struct io_ring_ctx *ctx,
> {
> int ret;
>
> + if (p->sq_thread_cpu >= nr_cpu_ids)
> + return -EINVAL;
> +
> init_waitqueue_head(&ctx->sqo_wait);
> mmgrab(current->mm);
> ctx->sqo_mm = current->mm;
> --
> 2.7.4
>
