On Thu, 13 Aug 2020 at 18:39, Marco Elver <elver@xxxxxxxxxx> wrote:
> Previous to the change to distinguish read-write accesses, when
> CONFIG_KCSAN_ASSUME_PLAIN_WRITES_ATOMIC=y is set, KCSAN would consider
> the non-atomic bitops as atomic. We want to partially revert to this
> behaviour, but with one important distinction: report racing
> modifications, since lost bits due to non-atomicity are certainly
> possible.
>
> Given the operations here only modify a single bit, assuming
> non-atomicity of the writer is sufficient may be reasonable for certain
> usage (and follows the permissible nature of the "assume plain writes
> atomic" rule). In other words:
>
> 1. We want non-atomic read-modify-write races to be reported;
> this is accomplished by kcsan_check_read(), where any
> concurrent write (atomic or not) will generate a report.
>
> 2. We do not want to report races with marked readers, but -do-
> want to report races with unmarked readers; this is
> accomplished by the instrument_write() ("assume atomic
> write" with Kconfig option set).
>
> With the above rules, when KCSAN_ASSUME_PLAIN_WRITES_ATOMIC is selected,
> it is hoped that KCSAN's reporting behaviour is better aligned with
> current expected permissible usage for non-atomic bitops.
>
> Note that, a side-effect of not telling KCSAN that the accesses are
> read-writes, is that this information is not displayed in the access
> summary in the report. It is, however, visible in inline-expanded stack
> traces. For now, it does not make sense to introduce yet another special
> case to KCSAN's runtime, only to cater to the case here.
>
> Signed-off-by: Marco Elver <elver@xxxxxxxxxx>
> Cc: Dmitry Vyukov <dvyukov@xxxxxxxxxx>
> Cc: Paul E. McKenney <paulmck@xxxxxxxxxx>
> Cc: Will Deacon <will@xxxxxxxxxx>
> ---
> As discussed, partially reverting behaviour for non-atomic bitops when
> KCSAN_ASSUME_PLAIN_WRITES_ATOMIC is selected.
>
> I'd like to avoid more special cases in KCSAN's runtime to cater to
> cases like this, not only because it adds more complexity, but it
> invites more special cases to be added. If there are other such
> primitives, we likely have to do it on a case-by-case basis as well, and
> justify carefully for each such case. But currently, as far as I can
> tell, the bitops are truly special, simply because we do know each op
> just touches a single bit.
> ---
> .../bitops/instrumented-non-atomic.h | 30 +++++++++++++++++--
> 1 file changed, 27 insertions(+), 3 deletions(-)
Paul, if it looks good to you, feel free to pick it up.
Thanks,
-- Marco
