Re: [RFC][PATCH 01/22] x86 user stack frame reads: switch to explicit __get_user()

[Andy Lutomirski]

On Sun, Mar 29, 2020 at 2:26 AM Ingo Molnar <mingo@xxxxxxxxxx> wrote:
>
>
> * Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote:
>
> > > but the __get_user() API doesn't carry the 'unsafe' tag yet.
> > >
> > > Should we add an __unsafe_get_user() alias to it perhaps, and use it
> > > in all code that adds it, like the chunk above? Or rename it to
> > > __unsafe_get_user() outright? No change to the logic, but it would be
> > > more obvious what code has inherited old __get_user() uses and which
> > > code uses __unsafe_get_user() intentionally.
> > >
> > > Even after your series there's 700 uses of __get_user(), so it would
> > > make sense to make a distinction in name at least and tag all unsafe
> > > APIs with an 'unsafe_' prefix.
> >
> > "unsafe" != "lacks access_ok", it's "done under user_access_begin".
>
> Well, I thought the principle was that we'd mark generic APIs that had
> *either* a missing access_ok() check or a missing
> user_access_begin()/end() wrapping marked unsafe_*(), right?
>
> __get_user() has __uaccess_begin()/end() on the inside, but doesn't have
> the access_ok() check, so those calls are 'unsafe' with regard to not
> being safe to untrusted (ptr,size) ranges.
>
> I agree that all of these topics need equal attention:
>
>  - leaking of cleared SMAP state (CLAC), which results in a silent
>    failure.
>
>  - running user accesses without STAC, which results in a crash.
>
>  - not doing an access_ok() check on untrusted (pointer,size) ranges,
>    which results in a silent failure as well.

My incliniation is to just get rid of the __get_user()-style APIs.
There shouldn't be any __get_user() calls that can't be directly
replaced by get_user(), and a single integer comparison is not that
expensive.  On SMAP systems, the speedup of __get_user vs get_user is
negligible.

(It's possible that some arch code somewhere uses __get_user as a way
to say "access user or kernel memory -- I know what I'm doing".  This
is crap if it exists.  It better not happen in generic code because of
sane architectures like s390x.)