Tree RCU guarantees that every online CPU has a memory barrier between
any given grace period and any of that CPU's RCU read-side sections that
must be ordered against that grace period. Since RCU doesn't always
know where read-side critical sections are, the actual implementation
guarantees order against prior and subsequent non-idle non-offline code,
whether in an RCU read-side critical section or not. As a result, there
does not need to be a memory barrier at the end of synchronize_rcu()
and friends because the ordering internal to the grace period has
ordered every CPU's post-grace-period execution against each CPU's
pre-grace-period execution, again for all non-idle online CPUs.
In contrast, SRCU can have non-idle online CPUs that are completely
uninvolved in a given SRCU grace period, for example, a CPU that
never runs any SRCU read-side critical sections and took no part in
the grace-period processing. It is in theory possible for a given
synchronize_srcu()'s wakeup to be delivered to a CPU that was completely
uninvolved in the prior SRCU grace period, which could mean that the
code following that synchronize_srcu() would end up being unordered with
respect to both the grace period and any pre-existing SRCU read-side
critical sections.
This commit therefore adds an smp_mb() to the end of __synchronize_srcu(),
which prevents this scenario from occurring.
Reported-by: Lance Roy <ldr...@gmail.com>
Signed-off-by: Paul E. McKenney <paul...@linux.vnet.ibm.com>
Acked-by: Lance Roy <ldr...@gmail.com>
---
kernel/rcu/srcutree.c | 9 +++++++++
1 file changed, 9 insertions(+)
diff --git a/kernel/rcu/srcutree.c b/kernel/rcu/srcutree.c
index 94bd6ed43ea3..c1c0ee3cce3b 100644
--- a/kernel/rcu/srcutree.c
+++ b/kernel/rcu/srcutree.c
@@ -897,6 +897,15 @@ static void __synchronize_srcu(struct srcu_struct *sp,
bool do_norm)
__call_srcu(sp, &rcu.head, wakeme_after_rcu, do_norm);
wait_for_completion(&rcu.completion);
destroy_rcu_head_on_stack(&rcu.head);
+
+ /*
+ * Make sure that later code is ordered after the SRCU grace
+ * period. This pairs with the raw_spin_lock_irq_rcu_node()
+ * in srcu_invoke_callbacks(). Unlike Tree RCU, this is needed
+ * because the current CPU might have been totally uninvolved with
+ * (and thus unordered against) that grace period.
+ */
+ smp_mb();
}
/**
--
2.5.2
