team: avoid race condition in scheduling delayed work
authorJoe Lawrence <Joe.Lawrence@stratus.com>
Fri, 3 Oct 2014 13:58:34 +0000 (09:58 -0400)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Wed, 15 Oct 2014 06:36:42 +0000 (08:36 +0200)
[ Upstream commit 47549650abd13d873fd2e5fc218db19e21031074 ]

When team_notify_peers and team_mcast_rejoin are called, they both reset
their respective .count_pending atomic variable. Then when the actual
worker function is executed, the variable is atomically decremented.
This pattern introduces a potential race condition where the
.count_pending rolls over and the worker function keeps rescheduling
until .count_pending decrements to zero again:

THREAD 1                           THREAD 2

========                           ========
team_notify_peers(teamX)
  atomic_set count_pending = 1
  schedule_delayed_work
                                   team_notify_peers(teamX)
                                   atomic_set count_pending = 1
team_notify_peers_work
  atomic_dec_and_test
    count_pending = 0
  (return)
                                   schedule_delayed_work
                                   team_notify_peers_work
                                   atomic_dec_and_test
                                     count_pending = -1
                                   schedule_delayed_work
                                   (repeat until count_pending = 0)

Instead of assigning a new value to .count_pending, use atomic_add to
tack-on the additional desired worker function invocations.

Signed-off-by: Joe Lawrence <joe.lawrence@stratus.com>
Acked-by: Jiri Pirko <jiri@resnulli.us>
Fixes: fc423ff00df3a19554414ee ("team: add peer notification")
Fixes: 492b200efdd20b8fcfdac87 ("team: add support for sending multicast rejoins")
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
drivers/net/team/team.c

index 26d8c29..979fe43 100644 (file)
@@ -647,7 +647,7 @@ static void team_notify_peers(struct team *team)
 {
        if (!team->notify_peers.count || !netif_running(team->dev))
                return;
-       atomic_set(&team->notify_peers.count_pending, team->notify_peers.count);
+       atomic_add(team->notify_peers.count, &team->notify_peers.count_pending);
        schedule_delayed_work(&team->notify_peers.dw, 0);
 }
 
@@ -687,7 +687,7 @@ static void team_mcast_rejoin(struct team *team)
 {
        if (!team->mcast_rejoin.count || !netif_running(team->dev))
                return;
-       atomic_set(&team->mcast_rejoin.count_pending, team->mcast_rejoin.count);
+       atomic_add(team->mcast_rejoin.count, &team->mcast_rejoin.count_pending);
        schedule_delayed_work(&team->mcast_rejoin.dw, 0);
 }