[NET] ifb: set separate lockdep classes for queue locks
authorJarek Poplawski <jarkao2@gmail.com>
Fri, 21 Mar 2008 00:05:13 +0000 (17:05 -0700)
committerDavid S. Miller <davem@davemloft.net>
Fri, 21 Mar 2008 00:05:13 +0000 (17:05 -0700)
[   10.536424] =======================================================
[   10.536424] [ INFO: possible circular locking dependency detected ]
[   10.536424] 2.6.25-rc3-devel #3
[   10.536424] -------------------------------------------------------
[   10.536424] swapper/0 is trying to acquire lock:
[   10.536424]  (&dev->queue_lock){-+..}, at: [<c0299b4a>]
dev_queue_xmit+0x175/0x2f3
[   10.536424]
[   10.536424] but task is already holding lock:
[   10.536424]  (&p->tcfc_lock){-+..}, at: [<f8a67154>] tcf_mirred+0x20/0x178
[act_mirred]
[   10.536424]
[   10.536424] which lock already depends on the new lock.

lockdep warns of locking order while using ifb with sch_ingress and
act_mirred: ingress_lock, tcfc_lock, queue_lock (usually queue_lock
is at the beginning). This patch is only to tell lockdep that ifb is
a different device (e.g. from eth) and has its own pair of queue
locks. (This warning is a false-positive in common scenario of using
ifb; yet there are possible situations, when this order could be
dangerous; lockdep should warn in such a case.) (With suggestions by
David S. Miller)

Reported-and-tested-by: Denys Fedoryshchenko <denys@visp.net.lb>
Signed-off-by: Jarek Poplawski <jarkao2@gmail.com>
Acked-by: Jamal Hadi Salim <hadi@cyberus.ca>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ifb.c

index 15949d3df17e082726033593ae0246ce57e3191f..af233b591534e6128a2b1c6b702ee312a12a49b4 100644 (file)
@@ -35,6 +35,7 @@
 #include <linux/moduleparam.h>
 #include <net/pkt_sched.h>
 #include <net/net_namespace.h>
+#include <linux/lockdep.h>
 
 #define TX_TIMEOUT  (2*HZ)
 
@@ -227,6 +228,16 @@ static struct rtnl_link_ops ifb_link_ops __read_mostly = {
 module_param(numifbs, int, 0);
 MODULE_PARM_DESC(numifbs, "Number of ifb devices");
 
+/*
+ * dev_ifb->queue_lock is usually taken after dev->ingress_lock,
+ * reversely to e.g. qdisc_lock_tree(). It should be safe until
+ * ifb doesn't take dev->queue_lock with dev_ifb->ingress_lock.
+ * But lockdep should know that ifb has different locks from dev.
+ */
+static struct lock_class_key ifb_queue_lock_key;
+static struct lock_class_key ifb_ingress_lock_key;
+
+
 static int __init ifb_init_one(int index)
 {
        struct net_device *dev_ifb;
@@ -246,6 +257,10 @@ static int __init ifb_init_one(int index)
        err = register_netdevice(dev_ifb);
        if (err < 0)
                goto err;
+
+       lockdep_set_class(&dev_ifb->queue_lock, &ifb_queue_lock_key);
+       lockdep_set_class(&dev_ifb->ingress_lock, &ifb_ingress_lock_key);
+
        return 0;
 
 err: