[5/8] seqlock: Require WRITE_ONCE surrounding raw_seqcount_barrier
diff mbox series

Message ID 20191016083959.186860-6-elver@google.com
State New
Headers show
  • Add Kernel Concurrency Sanitizer (KCSAN)
Related show

Commit Message

Marco Elver Oct. 16, 2019, 8:39 a.m. UTC
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@google.com>
 include/linux/seqlock.h | 7 +++++++
 1 file changed, 7 insertions(+)

diff mbox series

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;