Re: better patch for linux/bitops.h

From: H. Peter Anvin
Date: Thu May 05 2016 - 18:23:12 EST

On 05/05/2016 03:18 PM, tytso@xxxxxxx wrote:
> On Thu, May 05, 2016 at 05:34:50PM -0400, Sandy Harris wrote:
>> I completely fail to see why tests or compiler versions should be
>> part of the discussion. The C standard says the behaviour in
>> certain cases is undefined, so a standard-compliant compiler
>> can generate more-or-less any code there.
>> As long as any of portability, reliability or security are among our
>> goals, any code that can give undefined behaviour should be
>> considered problematic.
> Because compilers have been known not necessarily to obey the specs,
> and/or interpret the specs in way that not everyone agrees with. It's
> also the case that we are *already* disabling certain C optimizations
> which are technically allowed by the spec, but which kernel
> programmers consider insane (e.g., strict aliasing).
> And of course, memzero_explicit() which crypto people understand is
> necessary, is something that technically compilers are allowed to
> optimize according to the spec. So trying to write secure kernel code
> which will work on arbitrary compilers may well be impossible.
> And which is also why people have said (mostly in jest), "A
> sufficiently advanced compiler is indistinguishable from an
> adversary." (I assume people will agree that optimizing away a memset
> needed to clear secrets from memory would be considered adversarial,
> at the very least!)
> So this is why I tend to take a much more pragmatic viewpoint on
> things. Sure, it makes sense to pay attention to what the C standard
> writers are trying to do to us; but if we need to suppress certain
> optimizations to write sane kernel code --- I'm ok with that. And
> this is why using a trust-but-verify on a specific set of compilers
> and ranges of compiler versions is a really good idea....

In theory, theory and practice should agree, but in practice, practice
is what counts. I fully agree we should get rid of UD behavior where
doing so is practical, but not at the cost of breaking real-life
compilers, expecially not gcc, and to a lesser but still very real
extent icc and clang.

I would also agree that we should push the gcc developers to add to the
manual C-standard-UD behavior which are well-defined under the
gnu89/gnu99/gnu11 C dialects.