Date:   Mon, 8 Jun 2020 21:56:23 +0200
From:   Marco Elver <elver@...gle.com>
To:     Martin Liška <mliska@...e.cz>
Cc:     Borislav Petkov <bp@...en8.de>, Will Deacon <will@...nel.org>,
        Peter Zijlstra <peterz@...radead.org>,
        LKML <linux-kernel@...r.kernel.org>,
        Thomas Gleixner <tglx@...utronix.de>,
        "Paul E. McKenney" <paulmck@...nel.org>,
        Ingo Molnar <mingo@...nel.org>,
        Dmitry Vyukov <dvyukov@...gle.com>,
        Nick Desaulniers <ndesaulniers@...gle.com>
Subject: Re: [PATCH v5 00/18] Rework READ_ONCE() to improve codegen

On Mon, 8 Jun 2020 at 19:32, Martin Liška <mliska@...e.cz> wrote:
>
> On 6/3/20 9:23 PM, Marco Elver wrote:
> >
> >
> > On Wed, 03 Jun 2020, Borislav Petkov wrote:
> >
> >> On Thu, May 14, 2020 at 12:05:38PM +0100, Will Deacon wrote:
> >>> Talking off-list, Clang >= 7 is pretty reasonable wrt inlining decisions
> >>> and the behaviour for __always_inline is:
> >>>
> >>>    * An __always_inline function inlined into a __no_sanitize function is
> >>>      not instrumented
> >>>    * An __always_inline function inlined into an instrumented function is
> >>>      instrumented
> >>>    * You can't mark a function as both __always_inline __no_sanitize, because
> >>>      __no_sanitize functions are never inlined
> >>>
> >>> GCC, on the other hand, may still inline __no_sanitize functions and then
> >>> subsequently instrument them.
> >>
> >> Yeah, about that: I've been looking for a way to trigger this so that
> >> I can show preprocessed source to gcc people. So do you guys have a
> >> .config or somesuch I can try?
> >
> > For example take this:
> >
> >       int x;
> >
> >       static inline __attribute__((no_sanitize_thread)) void do_not_sanitize(void) {
> >         x++;
> >       }
> >
> >       void sanitize_this(void) {
> >         do_not_sanitize();
> >       }
> >
> > Then
> >
> >       gcc-10 -O3 -fsanitize=thread -o example.o -c example.c
> >       objdump -D example.o
>
> Hello.
>
> Thank you for the example. It seems to me that Clang does not inline a no_sanitize_* function
> into one which is instrumented. Is it a documented behavior ([1] doesn't mention that)?
> If so, we can do the same in GCC.

It is not explicitly mentioned in [1]. But the contract of
"no_sanitize" is "that a particular instrumentation or set of
instrumentations should not be applied". That contract is broken if a
function is instrumented, however that may happen. It sadly does
happen with GCC when a function is inlined. Presumably because the
sanitizer passes for TSAN/ASAN/MSAN run after the optimizer -- this
definitely can't change. Also because it currently gives us the
property that __always_inline functions are instrumented according to
the function they are inlined into (a property we want).

The easy fix to no_sanitize seems to be to do what Clang does, and
never inline no_sanitize functions (with or without "inline"
attribute).  always_inline functions should remain unchanged
(specifying no_sanitize on an always_inline function is an error).

Note this applies to all sanitizers (TSAN/ASAN/MSAN) and their
no_sanitize attribute that GCC has.

The list of requirements were also summarized in more detail here:
https://lore.kernel.org/lkml/CANpmjNMTsY_8241bS7=XAfqvZHFLrVEkv_uM4aDUWE_kh3Rvbw@mail.gmail.com/

Hope that makes sense. (I also need to send a v2 for param
tsan-distinguish-volatile, but haven't gotten around to it yet --
hopefully soon. And then we also need a param
tsan-instrument-func-entry-exit, which LLVM has for TSAN. One step at
a time though.)

Thanks,
-- Marco


> Thanks,
> Martin
>
> [1] https://clang.llvm.org/docs/AttributeReference.html#no-sanitize
>
> >
> > will show that do_not_sanitize() was inlined into sanitize_this() and is
> > instrumented. (With Clang this doesn't happen.)
> >
> > Hope this is enough.
> >
> > Thanks,
> > -- Marco
> >
>

Powered by blists - more mailing lists