[PATCH v2] Avoid memory barrier in read_seqcount() through load acquire

[Christoph Lameter (Ampere)]

On Fri, 23 Aug 2024, Thomas Gleixner wrote:

> This all can be done without the extra copies of the counter
> accessors. Uncompiled patch below.

Great. Thanks. Tried it too initially but could not make it work right.

One thing that we also want is the use of the smp_cond_load_acquire to
have the cpu power down while waiting for a cacheline change.

The code has several places where loops occur when the last bit is set in
the seqcount.

We could use smp_cond_load_acquire in load_sequence() but what do we do
about the loops at the higher level? Also this does not sync with the lock
checking logic.


diff --git a/include/linux/seqlock.h b/include/linux/seqlock.h
index 68b3af8bd6c6..4442a97ffe9a 100644
--- a/include/linux/seqlock.h
+++ b/include/linux/seqlock.h
@@ -135,7 +135,7 @@ static inline void seqcount_lockdep_reader_access(const seqcount_t *s)
 static __always_inline unsigned __seqprop_load_sequence(const seqcount_t *s, bool acquire)
 {
 	if (acquire && IS_ENABLED(CONFIG_ARCH_HAS_ACQUIRE_RELEASE))
-		return smp_load_acquire(&s->sequence);
+		return smp_cond_load_acquire(&s->sequence, (s->sequence & 1) == 0);
 	else
 		return READ_ONCE(s->sequence);
 }