This patch proposes to require marked atomic accesses surrounding raw_write_seqcount_barrier. We reason that otherwise there is no way to guarantee propagation nor atomicity of writes before/after the barrier [1]. For example, consider the compiler tears stores either before or after the barrier; in this case, readers may observe a partial value, and because readers are unaware that writes are going on (writes are not in a seq-writer critical section), will complete the seq-reader critical section while having observed some partial state. [1] https://lwn.net/Articles/793253/ This came up when designing and implementing KCSAN, because KCSAN would flag these accesses as data-races. After careful analysis, our reasoning as above led us to conclude that the best thing to do is to propose an amendment to the raw_seqcount_barrier usage. Signed-off-by: Marco Elver <elver@xxxxxxxxxx> --- include/linux/seqlock.h | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/include/linux/seqlock.h b/include/linux/seqlock.h index 1e425831a7ed..5d50aad53b47 100644 --- a/include/linux/seqlock.h +++ b/include/linux/seqlock.h @@ -265,6 +265,13 @@ static inline void raw_write_seqcount_end(seqcount_t *s) * usual consistency guarantee. It is one wmb cheaper, because we can * collapse the two back-to-back wmb()s. * + * Note that, writes surrounding the barrier should be declared atomic (e.g. + * via WRITE_ONCE): a) to ensure the writes become visible to other threads + * atomically, avoiding compiler optimizations; b) to document which writes are + * meant to propagate to the reader critical section. This is necessary because + * neither writes before and after the barrier are enclosed in a seq-writer + * critical section that would ensure readers are aware of ongoing writes. + * * seqcount_t seq; * bool X = true, Y = false; * -- 2.23.0.866.gb869b98d4c-goog
