Skip to contentSkip to conversation

forcedeth: fix locking bug with netconsole

PATCH In review Ingo Molnar 2 participants 3 messages Mar 21, 2008

Subsystems (2)
Networking Network drivers
lore
Share & tools
lore

Ingo Molnar

author
#0 lore raw

PATCH 1/1 forcedeth: fix locking bug with netconsole

Subject: forcedeth: fix locking bug with netconsole
From: Ingo Molnar <[email protected]>

while using netconsole on forcedeth, lockdep noticed the following locking bug:

=================================
[ INFO: inconsistent lock state ]
2.6.24-rc6 #6

Signed-off-by Ingo Molnar <[email protected]>

---------------------------------
inconsistent {softirq-on-W} -> {in-softirq-W} usage. udevd/719 [HC0[0]:SC1[1]:HE1:SE0] takes:
(_xmit_ETHER){-+..}, at: [<c043062e>] dev_watchdog+0x1c/0xb9
{softirq-on-W} state was registered at:

  [<c0147f67>] mark_held_locks+0x4e/0x66
  [<c014810e>] trace_hardirqs_on+0xfe/0x136
  [<c048ae63>] _spin_unlock_irq+0x22/0x42
  [<c02ec617>] nv_start_xmit_optimized+0x347/0x37a
  [<c042c80d>] netpoll_send_skb+0xa4/0x147
  [<c042d4a6>] netpoll_send_udp+0x238/0x242
  [<c02f44f6>] write_msg+0x6d/0x9b
  [<c012c129>] __call_console_drivers+0x4e/0x5a
  [<c012c18c>] _call_console_drivers+0x57/0x5b
  [<c012c2dd>] release_console_sem+0x11c/0x1b9
  [<c012caeb>] register_console+0x1eb/0x1f3
  [<c06ae673>] init_netconsole+0x119/0x15f
  [<c069149b>] kernel_init+0x147/0x294
  [<c01058cb>] kernel_thread_helper+0x7/0x10
  [<ffffffff>] 0xffffffff
irq event stamp: 950
hardirqs last  enabled at (950): [<c048ae63>] _spin_unlock_irq+0x22/0x42
hardirqs last disabled at (949): [<c048aaf7>] _spin_lock_irq+0xc/0x38
softirqs last  enabled at (0): [<c012a29c>] copy_process+0x375/0x126d
softirqs last disabled at (947): [<c0106d43>] do_softirq+0x61/0xc6

other info that might help us debug this:
no locks held by udevd/719.

stack backtrace:
Pid: 719, comm: udevd Not tainted 2.6.24-rc6 #6
 [<c0105c46>] show_trace_log_lvl+0x12/0x25
 [<c01063ec>] show_trace+0xd/0x10
 [<c010670c>] dump_stack+0x57/0x5f
 [<c0147505>] print_usage_bug+0x10a/0x117
 [<c0147c38>] mark_lock+0x121/0x402
 [<c01488b6>] __lock_acquire+0x3d1/0xb64
 [<c0149405>] lock_acquire+0x4e/0x6a
 [<c048a99b>] _spin_lock+0x23/0x32
 [<c043062e>] dev_watchdog+0x1c/0xb9
 [<c0133e4a>] run_timer_softirq+0x133/0x193
 [<c0130907>] __do_softirq+0x78/0xed
 [<c0106d43>] do_softirq+0x61/0xc6
 =======================

eth1: link down

the fix is to disable/restore irqs instead of disable/enable.

Signed-off-by Ingo Molnar <[email protected]>
---
 drivers/net/forcedeth.c |   18 ++++++++++--------
 net/core/netpoll.c      |    3 +++
 2 files changed, 13 insertions(+), 8 deletions(-)
drivers/net/forcedeth.c +10 -8
Index: linux/drivers/net/forcedeth.c
===================================================================
@@ -1854,6 +1854,7 @@ static int nv_start_xmit(struct sk_buff 
 	struct ring_desc* start_tx;
 	struct ring_desc* prev_tx;
 	struct nv_skb_map* prev_tx_ctx;
+	unsigned long flags;
 
 	/* add fragments to entries count */
 	for (i = 0; i < fragments; i++) {

@@ -1863,10 +1864,10 @@ static int nv_start_xmit(struct sk_buff 
 
 	empty_slots = nv_get_empty_tx_slots(np);
 	if (unlikely(empty_slots <= entries)) {
-		spin_lock_irq(&np->lock);
+		spin_lock_irqsave(&np->lock, flags);
 		netif_stop_queue(dev);
 		np->tx_stop = 1;
-		spin_unlock_irq(&np->lock);
+		spin_unlock_irqrestore(&np->lock, flags);
 		return NETDEV_TX_BUSY;
 	}
 

@@ -1929,13 +1930,13 @@ static int nv_start_xmit(struct sk_buff 
 		tx_flags_extra = skb->ip_summed == CHECKSUM_PARTIAL ?
 			 NV_TX2_CHECKSUM_L3 | NV_TX2_CHECKSUM_L4 : 0;
 
-	spin_lock_irq(&np->lock);
+	spin_lock_irqsave(&np->lock, flags);
 
 	/* set tx flags */
 	start_tx->flaglen |= cpu_to_le32(tx_flags | tx_flags_extra);
 	np->put_tx.orig = put_tx;
 
-	spin_unlock_irq(&np->lock);
+	spin_unlock_irqrestore(&np->lock, flags);
 
 	dprintk(KERN_DEBUG "%s: nv_start_xmit: entries %d queued for transmission. tx_flags_extra: %x\n",
 		dev->name, entries, tx_flags_extra);

@@ -1971,6 +1972,7 @@ static int nv_start_xmit_optimized(struc
 	struct ring_desc_ex* prev_tx;
 	struct nv_skb_map* prev_tx_ctx;
 	struct nv_skb_map* start_tx_ctx;
+	unsigned long flags;
 
 	/* add fragments to entries count */
 	for (i = 0; i < fragments; i++) {

@@ -1980,10 +1982,10 @@ static int nv_start_xmit_optimized(struc
 
 	empty_slots = nv_get_empty_tx_slots(np);
 	if (unlikely(empty_slots <= entries)) {
-		spin_lock_irq(&np->lock);
+		spin_lock_irqsave(&np->lock, flags);
 		netif_stop_queue(dev);
 		np->tx_stop = 1;
-		spin_unlock_irq(&np->lock);
+		spin_unlock_irqrestore(&np->lock, flags);
 		return NETDEV_TX_BUSY;
 	}
 

@@ -2059,7 +2061,7 @@ static int nv_start_xmit_optimized(struc
 			start_tx->txvlan = 0;
 	}
 
-	spin_lock_irq(&np->lock);
+	spin_lock_irqsave(&np->lock, flags);
 
 	if (np->tx_limit) {
 		/* Limit the number of outstanding tx. Setup all fragments, but

@@ -2085,7 +2087,7 @@ static int nv_start_xmit_optimized(struc
 	start_tx->flaglen |= cpu_to_le32(tx_flags | tx_flags_extra);
 	np->put_tx.ex = put_tx;
 
-	spin_unlock_irq(&np->lock);
+	spin_unlock_irqrestore(&np->lock, flags);
 
 	dprintk(KERN_DEBUG "%s: nv_start_xmit_optimized: entries %d queued for transmission. tx_flags_extra: %x\n",
 		dev->name, entries, tx_flags_extra);
net/core/netpoll.c +3 -0
Index: linux/net/core/netpoll.c
===================================================================
@@ -284,6 +284,7 @@ static void netpoll_send_skb(struct netp
 		/* try until next clock tick */
 		for (tries = jiffies_to_usecs(1)/USEC_PER_POLL;
 		     tries > 0; --tries) {
+			WARN_ON_ONCE(!irqs_disabled());
 			if (netif_tx_trylock(dev)) {
 				if (!netif_queue_stopped(dev) &&
 				    !netif_subqueue_stopped(dev, skb))

@@ -297,8 +298,10 @@ static void netpoll_send_skb(struct netp
 
 			/* tickle device maybe there is some cleanup */
 			netpoll_poll(np);
+			WARN_ON_ONCE(!irqs_disabled());
 
 			udelay(USEC_PER_POLL);
+			WARN_ON_ONCE(!irqs_disabled());
 		}
 		local_irq_restore(flags);
 	}
#1 lore raw
Quoted 5 lines · Ingo Molnar

Subject: forcedeth: fix locking bug with netconsole
From: Ingo Molnar <[email protected]>

while using netconsole on forcedeth, lockdep noticed the following locking bug:

[...]

Quoted 1 line

the fix is to disable/restore irqs instead of disable/enable.

So we make _the_ TX fast path slower, to accomodate a path that is rarely executed?

That seems like the wrong fix to me.

Jeff

author
#2 lore raw
Quoted 11 lines · Jeff Garzik
Quoted 5 lines · Ingo Molnar

Subject: forcedeth: fix locking bug with netconsole
From: Ingo Molnar <[email protected]>

while using netconsole on forcedeth, lockdep noticed the following locking bug:

[...]

Quoted 1 line

the fix is to disable/restore irqs instead of disable/enable.

So we make _the_ TX fast path slower, to accomodate a path that is rarely executed?

no strong opinion but a single PCI mmio access far outweighs any flags save/restore, so i'd not fight about something like that. I've never seen flags save/restore costs show up in NMI-driven profiles that had networking load.

Ingo