Re: [PATCH] iov_iter: separate direction from flavour

From: Linus Torvalds
Date: Sun Jul 04 2021 - 16:42:14 EST


On Sun, Jul 4, 2021 at 1:28 PM Guenter Roeck <linux@xxxxxxxxxxxx> wrote:
>
> Turns out that, at least on m68k/nommu, USER_DS and KERNEL_DS are the same.
>
> #define USER_DS MAKE_MM_SEG(TASK_SIZE)
> #define KERNEL_DS MAKE_MM_SEG(0xFFFFFFFF)

Ahh. So the code is fine, it's just that "uaccess_kernel()" isn't
something that can be reliably even tested for, and it will always
return true on those nommu platforms.

And we don't have a "uaccess_user()" macro that would test if it
matches USER_DS (and that also would always return true on those
configurations), so we can't just change the

WARN_ON_ONCE(uaccess_kernel());

into a

WARN_ON_ONCE(!uaccess_user());

instead.

Very annoying. Basically, every single use of "uaccess_kernel()" is unreliable.

There aren't all that many of them, and most of them are irrelevant
for no-mmu anyway (like the bpf tracing ones, or mm/memory.c). So this
iov_iter.c case is likely the only one that would be an issue.

That warning is something that should go away eventually anyway, but I
_like_ that warning for now, just to get coverage. But apparently it's
just not going to be the case for these situations.

My inclination is to keep it around for a while - to see if it catches
anything else - but remove it for the final 5.14 release because of
these nommu issues.

Of course, I will almost certainly not remember to do that unless
somebody reminds me...

The other alternative would be to just make nommu platforms that have
KERNEL_DS==USER_DS simply do

#define uaccess_kernel() (false)

and avoid it that way, since that's closer to what the modern
non-CONFIG_SET_FS world view is, and is what include/linux/uaccess.h
does for that case..

Linus