Re: [PATCH v2] rcu: fix a race in hlist_nulls_for_each_entry_rcu macro

[Eric Dumazet <eric.dumazet@gmail.com>]

On Wed, 2013-05-22 at 06:00 -0700, Paul E. McKenney wrote:

> Right, rcu_read_lock() is part of the protection, but rcu_dereference()
> is the other part.
> 
> All that aside, I can't claim that I understand what problem the various
> patches would solve.  ;-)

Problem is that rcu_dereference(expr) might be optimized by the compiler
to cache the dereferenced data in certain circumstances.

Following patch shows the difference if you look at the generated code:

This patch fixes the problem, and its not really obvious why !

diff --git a/net/ipv4/udp.c b/net/ipv4/udp.c
index 0bf5d39..6aa8088 100644
--- a/net/ipv4/udp.c
+++ b/net/ipv4/udp.c
@@ -416,11 +416,12 @@ static struct sock *udp4_lib_lookup2(struct net *net,
 	struct hlist_nulls_node *node;
 	int score, badness, matches = 0, reuseport = 0;
 	u32 hash = 0;
+	struct hlist_nulls_head *head = &hslot2->head;
 
 begin:
 	result = NULL;
 	badness = 0;
-	udp_portaddr_for_each_entry_rcu(sk, node, &hslot2->head) {
+	udp_portaddr_for_each_entry_rcu(sk, node, head) {
 		score = compute_score2(sk, net, saddr, sport,
 				      daddr, hnum, dif);
 		if (score > badness) {