Re: clang memcpy calls

[Peter Zijlstra <peterz@infradead.org>]

On Mon, Mar 28, 2022 at 12:20:39PM +0200, Jakub Jelinek wrote:
> On Mon, Mar 28, 2022 at 10:52:54AM +0100, Mark Rutland wrote:
> > I think we're talking past each other here, so let me be more precise. :)
> > 
> > The key thing is that when the user passes `-fsantize=address`, instrumentation
> > is added by (a part of) the compiler. That instrumentation is added under some
> > assumptions as to how the compiler as a whole will behave.
> > 
> > With that in mind, the question is how is __attribute__((no_sanitize_address))
> > intended to work when considering all the usual expectations around how the
> > compiler can play with memcpy and similar?
> 
> no_sanitize_address or lack thereof is whether the current function
> shouldn't be or should be ASan instrumented, not on whether other functions
> it calls are instrumented or not.  memcpy/memmove/memset are just a tiny bit
> special case because the compiler can add them on their own even if they
> aren't present in the source (there are a few others the compiler can
> pattern match too) and various builtins can be on the other side expanded
> inline instead of called, so one then gets the sanitization status of the
> function in which it is used rather than whether the out of line
> implementation of the function is sanitized.
> 
> If coexistence of instrumented and non-instrumented memcpy etc. was the goal
> (it clearly wasn't), then the sanitizer libraries wouldn't be overriding
> memcpy calls, but instead the compiler would redirect calls to memcpy in
> instrumented functions to say __asan_memcpy which then would be
> instrumented.

This then leaves us holding the pieces because this behaviour is
actively wrong.

A non-instrumented function *MUST*NOT* call an instrumented function,
ever. This very much violates how we use/expect
__attribute__((no_sanitize_address)) to work.

If we use that on a function, we expect/rely on that function (nor any
compiler tranformation thereof) to *NOT* have instrumentation. This is a
hard correctness requirement that cannot be argued with.


So there's two options:

 A) compiler generates implicit mem*() calls with the knowledge that
    mem*() is not instrumentet, and as such will also emit
    instrumentation for it when so required (or calls mem*_asan() like
    functions.

 B) compiler knows mem*() are instrumented, at which point the implicit
    mem*() calls are no longer equivalent under
    __attribute__((no_sanitize_address)) and will no longer perform
    these substitutions.


At some point this becomes a choice between being able to boot or having
KASAN, choice is simple.