linux-hardening - Re: [PATCH v6 2/4] Compiler Attributes: Add _

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Date:   Thu, 3 Feb 2022 16:26:12 -0800
From:   Kees Cook <keescook@...omium.org>
To:     Nick Desaulniers <ndesaulniers@...gle.com>
Cc:     Miguel Ojeda <ojeda@...nel.org>,
        Nathan Chancellor <nathan@...nel.org>, llvm@...ts.linux.dev,
        George Burgess IV <gbiv@...gle.com>,
        linux-kernel@...r.kernel.org, linux-hardening@...r.kernel.org
Subject: Re: [PATCH v6 2/4] Compiler Attributes: Add __overloadable for Clang

On Thu, Feb 03, 2022 at 02:11:48PM -0800, Nick Desaulniers wrote:
> On Thu, Feb 3, 2022 at 1:04 PM Kees Cook <keescook@...omium.org> wrote:
> >
> > On Thu, Feb 03, 2022 at 12:26:15PM -0800, Nick Desaulniers wrote:
> > > On Thu, Feb 3, 2022 at 9:33 AM Kees Cook <keescook@...omium.org> wrote:
> > > >
> > > > must be marked as being overloadable (i.e. different prototypes).
> > > > This allows the __pass_object_size versions to take precedence.
> > >
> > > Is this because of the `const` additions to the function signatures?
> >
> > That might be an issue, but the *real* issue is the implicit mutation of
> > the function into an inline with _additional_ arguments. i.e.
> >
> > char *strcpy(char * POS p, const char * POS q)
> >
> > is really
> >
> > char *strcpy(char * const p, const char * const q, size_t __size_of_p, size_t __size_of_q)
> >
> > (i.e. what I was doing with macros, but all internally and still an
> > extern inline)
> 
> What do you mean "is really"? 4/4 doesn't change the number of
> parameters in strcpy explicitly in the definition AFAICT.

It really does change the number of parameters. See the IR difference:

$ cat example.c
#ifdef USE_POS
# define POS __attribute__((pass_object_size(1)))
#else
# define POS
#endif

int func(void * const POS);

struct foo
{
        int a;
        char *b;
};

void usage(struct foo *example)
{
        func(example);
}

$ IR="-O2 -Xclang -disable-llvm-passes -emit-llvm -S"
$ clang           example.c $IR -o normal.ll
$ clang -DUSE_POS example.c $IR -o pos.ll
$ diff -u normal.ll pos.ll
--- normal.ll   2022-02-03 16:23:39.734065036 -0800
+++ pos.ll      2022-02-03 16:23:49.518083451 -0800
@@ -11,14 +11,19 @@
   store %struct.foo* %0, %struct.foo** %2, align 8, !tbaa !3
   %3 = load %struct.foo*, %struct.foo** %2, align 8, !tbaa !3
   %4 = bitcast %struct.foo* %3 to i8*
-  %5 = call i32 @func(i8* noundef %4)
+  %5 = call i64 @llvm.objectsize.i64.p0i8(i8* %4, i1 false, i1 true, i1 false)
+  %6 = call i32 @func(i8* noundef %4, i64 noundef %5)
   ret void
 }
...

This is basically doing internally exactly what I was doing in v4 and
earlier with macros (passing in the caller's view of __bos(arg, 1)).

-- 
Kees Cook