net: qdisc busylock needs lockdep annotations
Eric Dumazet [Wed, 5 Sep 2012 01:02:56 +0000 (01:02 +0000)]
It seems we need to provide ability for stacked devices
to use specific lock_class_key for sch->busylock

We could instead default l2tpeth tx_queue_len to 0 (no qdisc), but
a user might use a qdisc anyway.

(So same fixes are probably needed on non LLTX stacked drivers)

Noticed while stressing L2TPV3 setup :

======================================================
 [ INFO: possible circular locking dependency detected ]
 3.6.0-rc3+ #788 Not tainted
 -------------------------------------------------------
 netperf/4660 is trying to acquire lock:
  (l2tpsock){+.-...}, at: [<ffffffffa0208db2>] l2tp_xmit_skb+0x172/0xa50 [l2tp_core]

 but task is already holding lock:
  (&(&sch->busylock)->rlock){+.-...}, at: [<ffffffff81596595>] dev_queue_xmit+0xd75/0xe00

 which lock already depends on the new lock.

 the existing dependency chain (in reverse order) is:

 -> #1 (&(&sch->busylock)->rlock){+.-...}:
        [<ffffffff810a5df0>] lock_acquire+0x90/0x200
        [<ffffffff817499fc>] _raw_spin_lock_irqsave+0x4c/0x60
        [<ffffffff81074872>] __wake_up+0x32/0x70
        [<ffffffff8136d39e>] tty_wakeup+0x3e/0x80
        [<ffffffff81378fb3>] pty_write+0x73/0x80
        [<ffffffff8136cb4c>] tty_put_char+0x3c/0x40
        [<ffffffff813722b2>] process_echoes+0x142/0x330
        [<ffffffff813742ab>] n_tty_receive_buf+0x8fb/0x1230
        [<ffffffff813777b2>] flush_to_ldisc+0x142/0x1c0
        [<ffffffff81062818>] process_one_work+0x198/0x760
        [<ffffffff81063236>] worker_thread+0x186/0x4b0
        [<ffffffff810694d3>] kthread+0x93/0xa0
        [<ffffffff81753e24>] kernel_thread_helper+0x4/0x10

 -> #0 (l2tpsock){+.-...}:
        [<ffffffff810a5288>] __lock_acquire+0x1628/0x1b10
        [<ffffffff810a5df0>] lock_acquire+0x90/0x200
        [<ffffffff817498c1>] _raw_spin_lock+0x41/0x50
        [<ffffffffa0208db2>] l2tp_xmit_skb+0x172/0xa50 [l2tp_core]
        [<ffffffffa021a802>] l2tp_eth_dev_xmit+0x32/0x60 [l2tp_eth]
        [<ffffffff815952b2>] dev_hard_start_xmit+0x502/0xa70
        [<ffffffff815b63ce>] sch_direct_xmit+0xfe/0x290
        [<ffffffff81595a05>] dev_queue_xmit+0x1e5/0xe00
        [<ffffffff815d9d60>] ip_finish_output+0x3d0/0x890
        [<ffffffff815db019>] ip_output+0x59/0xf0
        [<ffffffff815da36d>] ip_local_out+0x2d/0xa0
        [<ffffffff815da5a3>] ip_queue_xmit+0x1c3/0x680
        [<ffffffff815f4192>] tcp_transmit_skb+0x402/0xa60
        [<ffffffff815f4a94>] tcp_write_xmit+0x1f4/0xa30
        [<ffffffff815f5300>] tcp_push_one+0x30/0x40
        [<ffffffff815e6672>] tcp_sendmsg+0xe82/0x1040
        [<ffffffff81614495>] inet_sendmsg+0x125/0x230
        [<ffffffff81576cdc>] sock_sendmsg+0xdc/0xf0
        [<ffffffff81579ece>] sys_sendto+0xfe/0x130
        [<ffffffff81752c92>] system_call_fastpath+0x16/0x1b
  Possible unsafe locking scenario:

        CPU0                    CPU1
        ----                    ----
   lock(&(&sch->busylock)->rlock);
                                lock(l2tpsock);
                                lock(&(&sch->busylock)->rlock);
   lock(l2tpsock);

  *** DEADLOCK ***

 5 locks held by netperf/4660:
  #0:  (sk_lock-AF_INET){+.+.+.}, at: [<ffffffff815e581c>] tcp_sendmsg+0x2c/0x1040
  #1:  (rcu_read_lock){.+.+..}, at: [<ffffffff815da3e0>] ip_queue_xmit+0x0/0x680
  #2:  (rcu_read_lock_bh){.+....}, at: [<ffffffff815d9ac5>] ip_finish_output+0x135/0x890
  #3:  (rcu_read_lock_bh){.+....}, at: [<ffffffff81595820>] dev_queue_xmit+0x0/0xe00
  #4:  (&(&sch->busylock)->rlock){+.-...}, at: [<ffffffff81596595>] dev_queue_xmit+0xd75/0xe00

 stack backtrace:
 Pid: 4660, comm: netperf Not tainted 3.6.0-rc3+ #788
 Call Trace:
  [<ffffffff8173dbf8>] print_circular_bug+0x1fb/0x20c
  [<ffffffff810a5288>] __lock_acquire+0x1628/0x1b10
  [<ffffffff810a334b>] ? check_usage+0x9b/0x4d0
  [<ffffffff810a3f44>] ? __lock_acquire+0x2e4/0x1b10
  [<ffffffff810a5df0>] lock_acquire+0x90/0x200
  [<ffffffffa0208db2>] ? l2tp_xmit_skb+0x172/0xa50 [l2tp_core]
  [<ffffffff817498c1>] _raw_spin_lock+0x41/0x50
  [<ffffffffa0208db2>] ? l2tp_xmit_skb+0x172/0xa50 [l2tp_core]
  [<ffffffffa0208db2>] l2tp_xmit_skb+0x172/0xa50 [l2tp_core]
  [<ffffffffa021a802>] l2tp_eth_dev_xmit+0x32/0x60 [l2tp_eth]
  [<ffffffff815952b2>] dev_hard_start_xmit+0x502/0xa70
  [<ffffffff81594e0e>] ? dev_hard_start_xmit+0x5e/0xa70
  [<ffffffff81595961>] ? dev_queue_xmit+0x141/0xe00
  [<ffffffff815b63ce>] sch_direct_xmit+0xfe/0x290
  [<ffffffff81595a05>] dev_queue_xmit+0x1e5/0xe00
  [<ffffffff81595820>] ? dev_hard_start_xmit+0xa70/0xa70
  [<ffffffff815d9d60>] ip_finish_output+0x3d0/0x890
  [<ffffffff815d9ac5>] ? ip_finish_output+0x135/0x890
  [<ffffffff815db019>] ip_output+0x59/0xf0
  [<ffffffff815da36d>] ip_local_out+0x2d/0xa0
  [<ffffffff815da5a3>] ip_queue_xmit+0x1c3/0x680
  [<ffffffff815da3e0>] ? ip_local_out+0xa0/0xa0
  [<ffffffff815f4192>] tcp_transmit_skb+0x402/0xa60
  [<ffffffff815fa25e>] ? tcp_md5_do_lookup+0x18e/0x1a0
  [<ffffffff815f4a94>] tcp_write_xmit+0x1f4/0xa30
  [<ffffffff815f5300>] tcp_push_one+0x30/0x40
  [<ffffffff815e6672>] tcp_sendmsg+0xe82/0x1040
  [<ffffffff81614495>] inet_sendmsg+0x125/0x230
  [<ffffffff81614370>] ? inet_create+0x6b0/0x6b0
  [<ffffffff8157e6e2>] ? sock_update_classid+0xc2/0x3b0
  [<ffffffff8157e750>] ? sock_update_classid+0x130/0x3b0
  [<ffffffff81576cdc>] sock_sendmsg+0xdc/0xf0
  [<ffffffff81162579>] ? fget_light+0x3f9/0x4f0
  [<ffffffff81579ece>] sys_sendto+0xfe/0x130
  [<ffffffff810a69ad>] ? trace_hardirqs_on+0xd/0x10
  [<ffffffff8174a0b0>] ? _raw_spin_unlock_irq+0x30/0x50
  [<ffffffff810757e3>] ? finish_task_switch+0x83/0xf0
  [<ffffffff810757a6>] ? finish_task_switch+0x46/0xf0
  [<ffffffff81752cb7>] ? sysret_check+0x1b/0x56
  [<ffffffff81752c92>] system_call_fastpath+0x16/0x1b

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

include/linux/netdevice.h
net/l2tp/l2tp_eth.c
net/sched/sch_generic.c

index ccac82e..ae3153c 100644 (file)
@@ -1322,6 +1322,8 @@ struct net_device {
        /* phy device may attach itself for hardware timestamping */
        struct phy_device *phydev;
 
+       struct lock_class_key *qdisc_tx_busylock;
+
        /* group the device belongs to */
        int group;
 
index f9ee74d..ba89997 100644 (file)
@@ -67,6 +67,7 @@ static inline struct l2tp_eth_net *l2tp_eth_pernet(struct net *net)
        return net_generic(net, l2tp_eth_net_id);
 }
 
+static struct lock_class_key l2tp_eth_tx_busylock;
 static int l2tp_eth_dev_init(struct net_device *dev)
 {
        struct l2tp_eth *priv = netdev_priv(dev);
@@ -74,7 +75,7 @@ static int l2tp_eth_dev_init(struct net_device *dev)
        priv->dev = dev;
        eth_hw_addr_random(dev);
        memset(&dev->broadcast[0], 0xff, 6);
-
+       dev->qdisc_tx_busylock = &l2tp_eth_tx_busylock;
        return 0;
 }
 
index 6c4d5fe..aefc150 100644 (file)
@@ -527,6 +527,8 @@ struct Qdisc_ops pfifo_fast_ops __read_mostly = {
 };
 EXPORT_SYMBOL(pfifo_fast_ops);
 
+static struct lock_class_key qdisc_tx_busylock;
+
 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
                          struct Qdisc_ops *ops)
 {
@@ -534,6 +536,7 @@ struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
        struct Qdisc *sch;
        unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
        int err = -ENOBUFS;
+       struct net_device *dev = dev_queue->dev;
 
        p = kzalloc_node(size, GFP_KERNEL,
                         netdev_queue_numa_node_read(dev_queue));
@@ -553,12 +556,16 @@ struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
        }
        INIT_LIST_HEAD(&sch->list);
        skb_queue_head_init(&sch->q);
+
        spin_lock_init(&sch->busylock);
+       lockdep_set_class(&sch->busylock,
+                         dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
+
        sch->ops = ops;
        sch->enqueue = ops->enqueue;
        sch->dequeue = ops->dequeue;
        sch->dev_queue = dev_queue;
-       dev_hold(qdisc_dev(sch));
+       dev_hold(dev);
        atomic_set(&sch->refcnt, 1);
 
        return sch;