tun: remove possible false sharing in tun_flow_update()
authorEric Dumazet <edumazet@google.com>
Wed, 9 Oct 2019 16:20:02 +0000 (09:20 -0700)
committerJakub Kicinski <jakub.kicinski@netronome.com>
Thu, 10 Oct 2019 04:29:33 +0000 (21:29 -0700)
As mentioned in https://github.com/google/ktsan/wiki/READ_ONCE-and-WRITE_ONCE#it-may-improve-performance
a C compiler can legally transform

if (e->queue_index != queue_index)
e->queue_index = queue_index;

to :

e->queue_index = queue_index;

Note that the code using jiffies has no issue, since jiffies
has volatile attribute.

if (e->updated != jiffies)
    e->updated = jiffies;

Fixes: 83b1bc122cab ("tun: align write-heavy flow entry members to a cache line")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Zhang Yu <zhangyu31@baidu.com>
Cc: Wang Li <wangli39@baidu.com>
Cc: Li RongQing <lirongqing@baidu.com>
Acked-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
drivers/net/tun.c

index 812dc3a..a8d3141 100644 (file)
@@ -526,8 +526,8 @@ static void tun_flow_update(struct tun_struct *tun, u32 rxhash,
        e = tun_flow_find(head, rxhash);
        if (likely(e)) {
                /* TODO: keep queueing to old queue until it's empty? */
-               if (e->queue_index != queue_index)
-                       e->queue_index = queue_index;
+               if (READ_ONCE(e->queue_index) != queue_index)
+                       WRITE_ONCE(e->queue_index, queue_index);
                if (e->updated != jiffies)
                        e->updated = jiffies;
                sock_rps_record_flow_hash(e->rps_rxhash);