Re: [PATCH 14/18] maccess: allow architectures to provide kernel probing directly

[Linus Torvalds <torvalds@linux-foundation.org>]

On Wed, May 13, 2020 at 9:01 AM Christoph Hellwig <hch@lst.de> wrote:
>
> +               arch_kernel_read(dst, src, type, err_label);            \

I'm wondering if

 (a) we shouldn't expose this as an interface in general

 (b) it wouldn't be named differently..

The reason for (a) is that several users of the
"copy_from_kernel_nofault()" interfaces just seem to want a single
access from kernel mode.

The reason for (b) is that if we do expose this as a normal interface,
it shouldn't be called "arch_kernel_read", and it should have the same
semantics as "get_user_unsafe()".

IOW, maybe we should simply do exactly that: have a
"get_kernel_nofault()" thing that looks exactly like
unsafe_get_user().

On x86, it would basically be identical to unsafe_get_user().

And on architectures that only have the copy function, you'd just have
a fallback something like this:

  #define get_kernel_nofault(dst, src, err_label) do {  \
        typeof (*src) __gkn_result;                     \
        if (probe_kernel_read(&__gkn_result, src) < 0)  \
                goto err_label;                         \
        (dst) = __gkn_result;                           \
  } while (0)

and now the people who want to read a single kernel word can just do

        get_kernel_nofault(n, untrusted_pointer, error);

and they're done.

And some day - when we get reliably "asm goto" wiith outputs - that
"get_kernel_fault()" will literally be a single instruction asm with
the proper exception handler marker, the way "put_user_unsafe()"
already works (and the way "put_kernel_nofault()" would already work
if it does the above).

             Linus