Re: [PATCH 1/2] kcsan: Instrument memcpy/memset/memmove with newer Clang

[Marco Elver]

On Wed, 7 Sept 2022 at 20:17, Boqun Feng <boqun.feng@xxxxxxxxx> wrote:
>
> On Wed, Sep 07, 2022 at 07:39:02PM +0200, Marco Elver wrote:
> > With Clang version 16+, -fsanitize=thread will turn
> > memcpy/memset/memmove calls in instrumented functions into
> > __tsan_memcpy/__tsan_memset/__tsan_memmove calls respectively.
> >
> > Add these functions to the core KCSAN runtime, so that we (a) catch data
> > races with mem* functions, and (b) won't run into linker errors with
> > such newer compilers.
> >
> > Cc: stable@xxxxxxxxxxxxxxx # v5.10+
>
> For (b) I think this is Ok, but for (a), what the atomic guarantee of
> our mem* functions? Per-byte atomic or something more complicated (for
> example, providing best effort atomic if a memory location in the range
> is naturally-aligned to a machine word)?

There should be no atomicity guarantee of mem*() functions, anything
else would never be safe, given compilers love to optimize all of them
(replacing the calls with inline versions etc.).

> If it's a per-byte atomicity, then maybe another KCSAN_ACCESS_* flags is
> needed, otherwise memset(0x8, 0, 0x2) is considered as atomic if
> ASSUME_PLAIN_WRITES_ATOMIC=y. Unless I'm missing something.
>
> Anyway, this may be worth another patch and some discussion/doc, because
> it just improve the accuracy of the tool. In other words, this patch and
> the "stable" tag look good to me.

Right, this will treat write accesses done by mem*() functions with a
size less than or equal to word size as atomic if that option is on.
However, I feel the more interesting cases will be
memcpy/memset/memmove with much larger sizes. That being said, note
that even though we pretend smaller than word size writes might be
atomic, for no data race to be detected, both accesses need to be
atomic.

If that behaviour should be changed for mem*() functions in the
default non-strict config is, like you say, something to ponder. In
general, I find the ASSUME_PLAIN_WRITES_ATOMIC=y a pretty bad default,
and I'd rather just change that default. But unfortunately, I think
the kernel isn't ready for that, given opinions on this still diverge.

Thanks,
-- Marco