forcedeth: fix lockdep warning on ethtool -s
Tobias Diedrich [Fri, 4 Jul 2008 06:54:56 +0000 (23:54 -0700)]
After enabling CONFIG_LOCKDEP and CONFIG_PROVE_LOCKING I get the
following warning when ethtool -s is first called on one of the
forcedeth ports:

=================================
[ INFO: inconsistent lock state ]
2.6.26-rc4 #28
---------------------------------
inconsistent {in-hardirq-W} -> {hardirq-on-W} usage.
ethtool/1985 [HC0[0]:SC0[1]:HE1:SE0] takes:
 (&np->lock){++..}, at: [<ffffffffa000c5fd>] nv_set_settings+0xc8/0x3de [forcedeth]
{in-hardirq-W} state was registered at:
  [<ffffffffffffffff>] 0xffffffffffffffff
irq event stamp: 3606
hardirqs last  enabled at (3605): [<ffffffff8068106f>] _spin_unlock_irqrestore+0x3f/0x68
hardirqs last disabled at (3604): [<ffffffff80680d38>] _spin_lock_irqsave+0x13/0x46
softirqs last  enabled at (3534): [<ffffffff80246ba5>] __do_softirq+0xbc/0xc5
softirqs last disabled at (3606): [<ffffffff80680b33>] _spin_lock_bh+0x11/0x41

other info that might help us debug this:
2 locks held by ethtool/1985:
 #0:  (rtnl_mutex){--..}, at: [<ffffffff80596072>] rtnl_lock+0x12/0x14
 #1:  (_xmit_ETHER){-+..}, at: [<ffffffffa000c5e8>] nv_set_settings+0xb3/0x3de [forcedeth]
stack backtrace:
Pid: 1985, comm: ethtool Not tainted 2.6.26-rc4 #28
Call Trace:
 [<ffffffff8025f190>] print_usage_bug+0x162/0x173
 [<ffffffff8025fa8b>] mark_lock+0x231/0x41f
 [<ffffffff802607cf>] __lock_acquire+0x4e7/0xcac
 [<ffffffff8025fe64>] ? trace_hardirqs_on+0xf1/0x115
 [<ffffffff80272c3a>] ? disable_irq_nosync+0x6f/0x7b
 [<ffffffff80261375>] lock_acquire+0x55/0x6e
 [<ffffffffa000c5fd>] ? :forcedeth:nv_set_settings+0xc8/0x3de
 [<ffffffff80680b15>] _spin_lock+0x2f/0x3c
 [<ffffffffa000c5fd>] :forcedeth:nv_set_settings+0xc8/0x3de
 [<ffffffff8058f8bb>] dev_ethtool+0x186/0xea3
 [<ffffffff8067f446>] ? mutex_lock_nested+0x243/0x275
 [<ffffffff8025df2b>] ? debug_mutex_free_waiter+0x46/0x4a
 [<ffffffff8067f469>] ? mutex_lock_nested+0x266/0x275
 [<ffffffff8058e1ce>] dev_ioctl+0x4eb/0x600
 [<ffffffff8068106f>] ? _spin_unlock_irqrestore+0x3f/0x68
 [<ffffffff80580f91>] sock_ioctl+0x1f5/0x202
 [<ffffffff802a322e>] vfs_ioctl+0x2a/0x77
 [<ffffffff802a34d6>] do_vfs_ioctl+0x25b/0x270
 [<ffffffff806807b6>] ? trace_hardirqs_on_thunk+0x35/0x3a
 [<ffffffff802a352d>] sys_ioctl+0x42/0x65
 [<ffffffff8021fffb>] system_call_after_swapgs+0x7b/0x80

This is caused by the following snippet in nv_set_settings:

netif_carrier_off(dev);
if (netif_running(dev)) {
nv_disable_irq(dev);
netif_tx_lock_bh(dev);
spin_lock(&np->lock);
/* stop engines */
nv_stop_rxtx(dev);
spin_unlock(&np->lock);
netif_tx_unlock_bh(dev);
}

Because of nv_disable_irq this is probably not really a problem
though (I guess) and replacing the spin_lock with spin_lock_irqsave
could keep interrupts disabled for a longer period of time because
of delays in nv_stop_rx and nv_stop_tx.

Signed-off-by: Tobias Diedrich <ranma+kernel@tdiedrich.de>
Cc: Ayaz Abdulla <aabdulla@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Jeff Garzik <jgarzik@redhat.com>

drivers/net/forcedeth.c

index 2cb2447..20d4fe9 100644 (file)
@@ -4194,12 +4194,23 @@ static int nv_set_settings(struct net_device *dev, struct ethtool_cmd *ecmd)
 
        netif_carrier_off(dev);
        if (netif_running(dev)) {
+               unsigned long flags;
+
                nv_disable_irq(dev);
                netif_tx_lock_bh(dev);
-               spin_lock(&np->lock);
+               /* with plain spinlock lockdep complains */
+               spin_lock_irqsave(&np->lock, flags);
                /* stop engines */
+               /* FIXME:
+                * this can take some time, and interrupts are disabled
+                * due to spin_lock_irqsave, but let's hope no daemon
+                * is going to change the settings very often...
+                * Worst case:
+                * NV_RXSTOP_DELAY1MAX + NV_TXSTOP_DELAY1MAX
+                * + some minor delays, which is up to a second approximately
+                */
                nv_stop_rxtx(dev);
-               spin_unlock(&np->lock);
+               spin_unlock_irqrestore(&np->lock, flags);
                netif_tx_unlock_bh(dev);
        }