arp: fix rcu lockdep splat in arp_process()
authorEric Dumazet <eric.dumazet@gmail.com>
Mon, 22 Aug 2011 19:32:42 +0000 (19:32 +0000)
committerDavid S. Miller <davem@davemloft.net>
Thu, 25 Aug 2011 00:55:00 +0000 (17:55 -0700)
Dave Jones reported a lockdep splat triggered by an arp_process() call
from parp_redo().

Commit faa9dcf793be (arp: RCU changes) is the origin of the bug, since
it assumed arp_process() was called under rcu_read_lock(), which is not
true in this particular path.

Instead of adding rcu_read_lock() in parp_redo(), I chose to add it in
neigh_proxy_process() to take care of IPv6 side too.

 ===================================================
 [ INFO: suspicious rcu_dereference_check() usage. ]
 ---------------------------------------------------
 include/linux/inetdevice.h:209 invoked rcu_dereference_check() without
protection!

 other info that might help us debug this:

 rcu_scheduler_active = 1, debug_locks = 0
 4 locks held by setfiles/2123:
  #0:  (&sb->s_type->i_mutex_key#13){+.+.+.}, at: [<ffffffff8114cbc4>]
walk_component+0x1ef/0x3e8
  #1:  (&isec->lock){+.+.+.}, at: [<ffffffff81204bca>]
inode_doinit_with_dentry+0x3f/0x41f
  #2:  (&tbl->proxy_timer){+.-...}, at: [<ffffffff8106a803>]
run_timer_softirq+0x157/0x372
  #3:  (class){+.-...}, at: [<ffffffff8141f256>] neigh_proxy_process
+0x36/0x103

 stack backtrace:
 Pid: 2123, comm: setfiles Tainted: G        W
3.1.0-0.rc2.git7.2.fc16.x86_64 #1
 Call Trace:
  <IRQ>  [<ffffffff8108ca23>] lockdep_rcu_dereference+0xa7/0xaf
  [<ffffffff8146a0b7>] __in_dev_get_rcu+0x55/0x5d
  [<ffffffff8146a751>] arp_process+0x25/0x4d7
  [<ffffffff8146ac11>] parp_redo+0xe/0x10
  [<ffffffff8141f2ba>] neigh_proxy_process+0x9a/0x103
  [<ffffffff8106a8c4>] run_timer_softirq+0x218/0x372
  [<ffffffff8106a803>] ? run_timer_softirq+0x157/0x372
  [<ffffffff8141f220>] ? neigh_stat_seq_open+0x41/0x41
  [<ffffffff8108f2f0>] ? mark_held_locks+0x6d/0x95
  [<ffffffff81062bb6>] __do_softirq+0x112/0x25a
  [<ffffffff8150d27c>] call_softirq+0x1c/0x30
  [<ffffffff81010bf5>] do_softirq+0x4b/0xa2
  [<ffffffff81062f65>] irq_exit+0x5d/0xcf
  [<ffffffff8150dc11>] smp_apic_timer_interrupt+0x7c/0x8a
  [<ffffffff8150baf3>] apic_timer_interrupt+0x73/0x80
  <EOI>  [<ffffffff8108f439>] ? trace_hardirqs_on_caller+0x121/0x158
  [<ffffffff814fc285>] ? __slab_free+0x30/0x24c
  [<ffffffff814fc283>] ? __slab_free+0x2e/0x24c
  [<ffffffff81204e74>] ? inode_doinit_with_dentry+0x2e9/0x41f
  [<ffffffff81204e74>] ? inode_doinit_with_dentry+0x2e9/0x41f
  [<ffffffff81204e74>] ? inode_doinit_with_dentry+0x2e9/0x41f
  [<ffffffff81130cb0>] kfree+0x108/0x131
  [<ffffffff81204e74>] inode_doinit_with_dentry+0x2e9/0x41f
  [<ffffffff81204fc6>] selinux_d_instantiate+0x1c/0x1e
  [<ffffffff81200f4f>] security_d_instantiate+0x21/0x23
  [<ffffffff81154625>] d_instantiate+0x5c/0x61
  [<ffffffff811563ca>] d_splice_alias+0xbc/0xd2
  [<ffffffff811b17ff>] ext4_lookup+0xba/0xeb
  [<ffffffff8114bf1e>] d_alloc_and_lookup+0x45/0x6b
  [<ffffffff8114cbea>] walk_component+0x215/0x3e8
  [<ffffffff8114cdf8>] lookup_last+0x3b/0x3d
  [<ffffffff8114daf3>] path_lookupat+0x82/0x2af
  [<ffffffff8110fc53>] ? might_fault+0xa5/0xac
  [<ffffffff8110fc0a>] ? might_fault+0x5c/0xac
  [<ffffffff8114c564>] ? getname_flags+0x31/0x1ca
  [<ffffffff8114dd48>] do_path_lookup+0x28/0x97
  [<ffffffff8114df2c>] user_path_at+0x59/0x96
  [<ffffffff811467ad>] ? cp_new_stat+0xf7/0x10d
  [<ffffffff811469a6>] vfs_fstatat+0x44/0x6e
  [<ffffffff811469ee>] vfs_lstat+0x1e/0x20
  [<ffffffff81146b3d>] sys_newlstat+0x1a/0x33
  [<ffffffff8108f439>] ? trace_hardirqs_on_caller+0x121/0x158
  [<ffffffff812535fe>] ? trace_hardirqs_on_thunk+0x3a/0x3f
  [<ffffffff8150af82>] system_call_fastpath+0x16/0x1b

Reported-by: Dave Jones <davej@redhat.com>
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/core/neighbour.c

index 8fab9b0bb2036784e002a4b3bb0464db059bd5ad..1334d7e56f020d5fac869f528590842483b68178 100644 (file)
@@ -1319,11 +1319,15 @@ static void neigh_proxy_process(unsigned long arg)
 
                if (tdif <= 0) {
                        struct net_device *dev = skb->dev;
+
                        __skb_unlink(skb, &tbl->proxy_queue);
-                       if (tbl->proxy_redo && netif_running(dev))
+                       if (tbl->proxy_redo && netif_running(dev)) {
+                               rcu_read_lock();
                                tbl->proxy_redo(skb);
-                       else
+                               rcu_read_unlock();
+                       } else {
                                kfree_skb(skb);
+                       }
 
                        dev_put(dev);
                } else if (!sched_next || tdif < sched_next)