Add two missing fences with comments describing them. These were found by

inspection and after a lengthy discussion with jhb and kib.  They have not
produced test failures.

Don't pointer chase through cpu0's smr.  Use cpu correct smr even when not
in a critical section to reduce the likelihood of false sharing.

1 files changed, 14 insertions, 3 deletions

diff --git a/sys/kern/subr_smr.c b/sys/kern/subr_smr.c
index c6a8f0e89015..c344fdcb7f5f 100644
--- a/sys/kern/subr_smr.c
+++ b/sys/kern/subr_smr.c
@@ -195,7 +195,7 @@ smr_advance(smr_t smr)
 	 * odd and an observed value of 0 in a particular CPU means
 	 * it is not currently in a read section.
 	 */
-	s = smr->c_shared;
+	s = zpcpu_get(smr)->c_shared;
 	goal = atomic_fetchadd_int(&s->s_wr_seq, SMR_SEQ_INCR) + SMR_SEQ_INCR;
 
 	/*
@@ -242,14 +242,19 @@ smr_poll(smr_t smr, smr_seq_t goal, bool wait)
 	 */
 	success = true;
 	critical_enter();
-	s = smr->c_shared;
+	s = zpcpu_get(smr)->c_shared;
 
 	/*
 	 * Acquire barrier loads s_wr_seq after s_rd_seq so that we can not
 	 * observe an updated read sequence that is larger than write.
 	 */
 	s_rd_seq = atomic_load_acq_int(&s->s_rd_seq);
-	s_wr_seq = smr_current(smr);
+
+	/*
+	 * wr_seq must be loaded prior to any c_seq value so that a stale
+	 * c_seq can only reference time after this wr_seq.
+	 */
+	s_wr_seq = atomic_load_acq_int(&s->s_wr_seq);
 
 	/*
 	 * Detect whether the goal is valid and has already been observed.
@@ -336,6 +341,12 @@ smr_poll(smr_t smr, smr_seq_t goal, bool wait)
 out:
 	critical_exit();
 
+	/*
+	 * Serialize with smr_advance()/smr_exit().  The caller is now free
+	 * to modify memory as expected.
+	 */
+	atomic_thread_fence_acq();
+
 	return (success);
 }