Re: [PATCH] linux/nospec.h: allow index argument to have const-qualified type

[Linus Torvalds]

On Thu, Feb 15, 2018 at 1:56 PM, Dan Williams <dan.j.williams@xxxxxxxxx> wrote:
>
> So I don't mind removing it, but I don't think it is garbage. It's
> there purely as a notification to the odd kernel developer that wants
> to pass "insane" index values,

But the thing is, the "index" value isn't even kernel-supplied.

Here's a test:  run a 32-bit kernel, and then do an ioctl() or
something with a negative fd.

What I think will happen is:

 - the negative fd will be seen as a big 'unsigned int' here:

        fcheck_files(struct files_struct *files, unsigned int fd)

which then does

                fd = array_index_nospec(fd, fdt->max_fds);

and that existing *STUPID* and *WRONG* WARN_ON() will trigger.

Sure, you can't trigger it on 64-bit kernels because there the
"unsigned int" will be small compared to LONG_MAX, but..

It is simply is *wrong* to check the "index".  It really fundamentally
is complete garbage.

Because the whole - and ONLY - *point* of this is that you have an
untrusted index. So checking it and giving a warning when it's out of
range is pure garbage.

Really. That warning must go away. Stop arguing for it, it's stupid and wrong.

Checking _size_ is one thing, but honestly, that's questionable too.

                  Linus