Re: [PATCH 11/18] seqcount: Introduce raw_write_seqcount_barrier()

[Peter Zijlstra]

On Thu, Jun 11, 2015 at 02:45:57PM -0700, Paul E. McKenney wrote:
> Color me slow and stupid.  Maybe due to reviewing a patch too early in
> the morning, who knows?
> 
> There is nothing above that prevents the compiler and the CPU from
> reordering the assignments to X and Y with the increment of s->sequence++.
> One fix would be as follows:
> 
> 	static inline void raw_write_seqcount_barrier(seqcount_t *s)
> 	{
> 		smp_wmb();
> 		s->sequence++;
> 		smp_wmb();
> 		s->sequence++;
> 		smp_wmb();
> 	}
> 
> Of course, this assumes that the accesses surrounding the call to
> raw_write_seqcount_barrier() are writes.  If they can be a reads,
> the two added smp_wmb() calls need to be full barriers.

I have updated the Changelog to hopefully explain things better.

I did leave off the READ/WRITE ONCE stuff, because I could not come up
with a scenario where it makes a difference -- I appreciate paranoia,
but I also think we should not overdo the thing.

---
Subject: seqcount: Introduce raw_write_seqcount_barrier()
From: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
Date: Thu Jun 11 12:35:48 CEST 2015

Introduce raw_write_seqcount_barrier(), a new construct that can be
used to provide write barrier semantics in seqcount read loops instead
of the usual consistency guarantee.

raw_write_seqcount_barier() is equivalent to:

	raw_write_seqcount_begin();
	raw_write_seqcount_end();

But avoids issueing two back-to-back smp_wmb() instructions.

This construct works because the read side will 'stall' when observing
odd values. This means that -- referring to the example in the comment
below -- even though there is no (matching) read barrier between the
loads of X and Y, we cannot observe !x && !y, because:

 - if we observe Y == false we must observe the first sequence
   increment, which makes us loop, until

 - we observe !(seq & 1) -- the second sequence increment -- at which
   time we must also observe T == true.

Cc: Al Viro <viro@xxxxxxxxxxxxxxxxxx>
Cc: Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx>
Cc: Paul McKenney <paulmck@xxxxxxxxxxxxxxxxxx>
Suggested-by: Oleg Nesterov <oleg@xxxxxxxxxx>
Signed-off-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>
---
 include/linux/seqlock.h |   42 ++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 42 insertions(+)

--- a/include/linux/seqlock.h
+++ b/include/linux/seqlock.h
@@ -233,6 +233,47 @@ static inline void raw_write_seqcount_en
 	s->sequence++;
 }
 
+/**
+ * raw_write_seqcount_barrier - do a seq write barrier
+ * @s: pointer to seqcount_t
+ *
+ * This can be used to provide an ordering guarantee instead of the
+ * usual consistency guarantee. It is one wmb cheaper, because we can
+ * collapse the two back-to-back wmb()s.
+ *
+ *      seqcount_t seq;
+ *      bool X = true, Y = false;
+ *
+ *      void read(void)
+ *      {
+ *              bool x, y;
+ *
+ *              do {
+ *                      int s = read_seqcount_begin(&seq);
+ *
+ *                      x = X; y = Y;
+ *
+ *              } while (read_seqcount_retry(&seq, s));
+ *
+ *              BUG_ON(!x && !y);
+ *      }
+ *
+ *      void write(void)
+ *      {
+ *              Y = true;
+ *
+ *              raw_write_seqcount_barrier(seq);
+ *
+ *              X = false;
+ *      }
+ */
+static inline void raw_write_seqcount_barrier(seqcount_t *s)
+{
+	s->sequence++;
+	smp_wmb();
+	s->sequence++;
+}
+
 /*
  * raw_write_seqcount_latch - redirect readers to even/odd copy
  * @s: pointer to seqcount_t
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/