Re: [PATCH v5 00/18] Rework READ_ONCE() to improve codegen

[Will Deacon <will@kernel.org>]

On Wed, May 13, 2020 at 03:15:55PM +0200, Marco Elver wrote:
> On Wed, 13 May 2020 at 14:40, Will Deacon <will@kernel.org> wrote:
> >
> > On Wed, May 13, 2020 at 02:32:43PM +0200, Peter Zijlstra wrote:
> > > On Wed, May 13, 2020 at 01:48:41PM +0200, Marco Elver wrote:
> > >
> > > > Disabling most instrumentation for arch/x86 is reasonable. Also fine
> > > > with the __READ_ONCE/__WRITE_ONCE changes (your improved
> > > > compiler-friendlier version).
> > > >
> > > > We likely can't have both: still instrument __READ_ONCE/__WRITE_ONCE
> > > > (as Will suggested) *and* avoid double-instrumentation in arch_atomic.
> > > > If most use-cases of __READ_ONCE/__WRITE_ONCE are likely to use
> > > > data_race() or KCSAN_SANITIZE := n anyway, I'd say it's reasonable for
> > > > now.
> >
> > I agree that Peter's patch is the right thing to do for now. I was hoping we
> > could instrument __{READ,WRITE}_ONCE(), but that we before I realised that
> > __no_sanitize_or_inline doesn't seem to do anything.
> >
> > > Right, if/when people want sanitize crud enabled for x86 I need
> > > something that:
> > >
> > >  - can mark a function 'no_sanitize' and all code that gets inlined into
> > >    that function must automagically also not get sanitized. ie. make
> > >    inline work like macros (again).
> > >
> > > And optionally:
> > >
> > >  - can mark a function explicitly 'sanitize', and only when an explicit
> > >    sanitize and no_sanitize mix in inlining give the current
> > >    incompatible attribute splat.
> > >
> > > That way we can have the noinstr function attribute imply no_sanitize
> > > and frob the DEFINE_IDTENTRY*() macros to use (a new) sanitize_or_inline
> > > helper instead of __always_inline for __##func().
> >
> > Sounds like a good plan to me, assuming the compiler folks are onboard.
> > In the meantime, can we kill __no_sanitize_or_inline and put it back to
> > the old __no_kasan_or_inline, which I think simplifies compiler.h and
> > doesn't mislead people into using the function annotation to avoid KCSAN?
> >
> > READ_ONCE_NOCHECK should also probably be READ_ONCE_NOKASAN, but I
> > appreciate that's a noisier change.
> 
> So far so good, except: both __no_sanitize_or_inline and
> __no_kcsan_or_inline *do* avoid KCSAN instrumenting plain accesses, it
> just doesn't avoid explicit kcsan_check calls, like those in
> READ/WRITE_ONCE if KCSAN is enabled for the compilation unit. That's
> just because macros won't be redefined just for __no_sanitize
> functions. Similarly, READ_ONCE_NOCHECK does work as expected, and its
> access is unchecked.
> 
> This will have the expected result:
> __no_sanitize_or_inline void foo(void) { x++; } // no data races reported
> 
> This will not work as expected:
> __no_sanitize_or_inline void foo(void) { READ_ONCE(x); }  // data
> races are reported

But the problem is that *this* does not work as expected:

unsigned long __no_sanitize_or_inline foo(unsigned long *ptr)
{
	return READ_ONCE_NOCHECK(*ptr);
}

which I think means that the function annotation is practically useless.

Will