Re: [PATCH -rcu] asm-generic, kcsan: Add KCSAN instrumentation for bitops

[Marco Elver <elver@google.com>]

On Wed, 15 Jan 2020 at 20:55, Arnd Bergmann <arnd@arndb.de> wrote:
>
> On Wed, Jan 15, 2020 at 8:51 PM Marco Elver <elver@google.com> wrote:
> >
> > On Wed, 15 Jan 2020 at 20:27, Arnd Bergmann <arnd@arndb.de> wrote:
> > >
> > > On Wed, Jan 15, 2020 at 5:58 PM Marco Elver <elver@google.com> wrote:
> > > >   * set_bit - Atomically set a bit in memory
> > > > @@ -26,6 +27,7 @@
> > > >  static inline void set_bit(long nr, volatile unsigned long *addr)
> > > >  {
> > > >         kasan_check_write(addr + BIT_WORD(nr), sizeof(long));
> > > > +       kcsan_check_atomic_write(addr + BIT_WORD(nr), sizeof(long));
> > > >         arch_set_bit(nr, addr);
> > > >  }
> > >
> > > It looks like you add a kcsan_check_atomic_write or kcsan_check_write directly
> > > next to almost any instance of kasan_check_write().
> > >
> > > Are there any cases where we actually just need one of the two but not the
> > > other? If not, maybe it's better to rename the macro and have it do both things
> > > as needed?
> >
> > Do you mean adding an inline helper at the top of each bitops header
> > here, similar to what we did for atomic-instrumented?  Happy to do
> > that if it improves readability.
>
> I was thinking of treewide wrappers, given that there are only a couple of files
> calling kasan_check_write():
>
> $ git grep -wl kasan_check_write
> arch/arm64/include/asm/barrier.h
> arch/arm64/include/asm/uaccess.h
> arch/x86/include/asm/uaccess_64.h
> include/asm-generic/atomic-instrumented.h
> include/asm-generic/bitops/instrumented-atomic.h
> include/asm-generic/bitops/instrumented-lock.h
> include/asm-generic/bitops/instrumented-non-atomic.h
> include/linux/kasan-checks.h
> include/linux/uaccess.h
> lib/iov_iter.c
> lib/strncpy_from_user.c
> lib/usercopy.c
> scripts/atomic/gen-atomic-instrumented.sh
>
> Are there any that really just want kasan_check_write() but not one
> of the kcsan checks?

If I understood correctly, this suggestion would amount to introducing
a new header, e.g. 'ksan-checks.h', that provides unified generic
checks. For completeness, we will also need to consider reads. Since
KCSAN provides 4 check variants ({read,write} x {plain,atomic}), we
will need 4 generic check variants.

I certainly do not feel comfortable blindly introducing kcsan_checks
in all places where we have kasan_checks, but it may be worthwhile
adding this infrastructure and starting with atomic-instrumented and
bitops-instrumented wrappers. The other locations you list above would
need to be evaluated on a case-by-case basis to check if we want to
report data races for those accesses.

As a minor data point, {READ,WRITE}_ONCE in compiler.h currently only
has kcsan_checks and not kasan_checks.

My personal preference would be to keep the various checks explicit,
clearly opting into either KCSAN and/or KASAN. Since I do not think
it's obvious if we want both for the existing and potentially new
locations (in future), the potential for error by blindly using a
generic 'ksan_check' appears worse than potentially adding a dozen
lines or so.

Let me know if you'd like to proceed with 'ksan-checks.h'.

Thanks,
-- Marco