Skip to contentSkip to conversation

sky2: initialize napi before registering device

PATCH Applied Stanislaw Gruszka 3 participants 4 messages Jan 25, 2014 – Jan 28, 2014

Subsystems (3)
Networking Network drivers Marvell gigabit Ethernet drivers (skge/sky2)
lore
Share & tools
lore

Stanislaw Gruszka

author
#0 lore raw

PATCH 1/1 sky2: initialize napi before registering device

There is race condition when call netif_napi_add() after register_netdevice(), as ->open() can be called without napi initialized and trigger BUG_ON() on napi_enable(), like on below messages:

[    9.699863] sky2: driver version 1.30
[    9.699960] sky2 0000:02:00.0: Yukon-2 EC Ultra chip revision 2
[    9.700020] sky2 0000:02:00.0: irq 45 for MSI/MSI-X
[    9.700498] ------------[ cut here ]------------
[    9.703391] kernel BUG at include/linux/netdevice.h:501!
[    9.703391] invalid opcode: 0000 [#1] PREEMPT SMP
<snip>
[    9.830018] Call Trace:
[    9.830018]  [<fa996169>] sky2_open+0x309/0x360 [sky2]
[    9.830018]  [<c1007210>] ? via_no_dac+0x40/0x40
[    9.830018]  [<c1007210>] ? via_no_dac+0x40/0x40
[    9.830018]  [<c135ed4b>] __dev_open+0x9b/0x120
[    9.830018]  [<c1431cbe>] ? _raw_spin_unlock_bh+0x1e/0x20
[    9.830018]  [<c135efd9>] __dev_change_flags+0x89/0x150
[    9.830018]  [<c135f148>] dev_change_flags+0x18/0x50
[    9.830018]  [<c13bb8e0>] devinet_ioctl+0x5d0/0x6e0
[    9.830018]  [<c13bcced>] inet_ioctl+0x6d/0xa0

To fix the problem patch changes the order of initialization.

Bug report:
https://bugzilla.kernel.org/show_bug.cgi?id=67151

Reported-and-tested-by [email protected]
Signed-off-by Stanislaw Gruszka <[email protected]>
---
 drivers/net/ethernet/marvell/sky2.c |    4 ++--
 1 files changed, 2 insertions(+), 2 deletions(-)
drivers/net/ethernet/marvell/sky2.c +2 -2
index 43aa7acd..2564eda 100644
@@ -5020,6 +5020,8 @@ static int sky2_probe(struct pci_dev *pdev, const struct pci_device_id *ent)
 		}
  	}
 
+	netif_napi_add(dev, &hw->napi, sky2_poll, NAPI_WEIGHT);
+
 	err = register_netdev(dev);
 	if (err) {
 		dev_err(&pdev->dev, "cannot register net device\n");

@@ -5028,8 +5030,6 @@ static int sky2_probe(struct pci_dev *pdev, const struct pci_device_id *ent)
 
 	netif_carrier_off(dev);
 
-	netif_napi_add(dev, &hw->napi, sky2_poll, NAPI_WEIGHT);
-
 	sky2_show_addr(dev);
 
 	if (hw->ports > 1) {
Signature · 1 line
1.7.4.4
#1 lore raw
Quoted 22 lines · On Sat, 25 Jan 2014 11:34:54 +0100 Stanislaw Gruszka <[email protected]> wrote:

There is race condition when call netif_napi_add() after register_netdevice(), as ->open() can be called without napi initialized and trigger BUG_ON() on napi_enable(), like on below messages:

[    9.699863] sky2: driver version 1.30
[    9.699960] sky2 0000:02:00.0: Yukon-2 EC Ultra chip revision 2
[    9.700020] sky2 0000:02:00.0: irq 45 for MSI/MSI-X
[    9.700498] ------------[ cut here ]------------
[    9.703391] kernel BUG at include/linux/netdevice.h:501!
[    9.703391] invalid opcode: 0000 [#1] PREEMPT SMP
<snip>
[    9.830018] Call Trace:
[    9.830018]  [<fa996169>] sky2_open+0x309/0x360 [sky2]
[    9.830018]  [<c1007210>] ? via_no_dac+0x40/0x40
[    9.830018]  [<c1007210>] ? via_no_dac+0x40/0x40
[    9.830018]  [<c135ed4b>] __dev_open+0x9b/0x120
[    9.830018]  [<c1431cbe>] ? _raw_spin_unlock_bh+0x1e/0x20
[    9.830018]  [<c135efd9>] __dev_change_flags+0x89/0x150
[    9.830018]  [<c135f148>] dev_change_flags+0x18/0x50
[    9.830018]  [<c13bb8e0>] devinet_ioctl+0x5d0/0x6e0
[    9.830018]  [<c13bcced>] inet_ioctl+0x6d/0xa0
Quoted 7 lines

To fix the problem patch changes the order of initialization.

Bug report:
https://bugzilla.kernel.org/show_bug.cgi?id=67151

Reported-and-tested-by [email protected]
Signed-off-by Stanislaw Gruszka <[email protected]>

That looks good, problem was introduced years ago.
I wonder if netif_napi_del() should be in unwind if registration of either devices fails?

author
#2 lore raw
Quoted 30 lines · Stephen Hemminger
Quoted 27 lines · On Sat, 25 Jan 2014 11:34:54 +0100 Stanislaw Gruszka <[email protected]> wrote:

There is race condition when call netif_napi_add() after register_netdevice(), as ->open() can be called without napi initialized and trigger BUG_ON() on napi_enable(), like on below messages:

[    9.699863] sky2: driver version 1.30
[    9.699960] sky2 0000:02:00.0: Yukon-2 EC Ultra chip revision 2
[    9.700020] sky2 0000:02:00.0: irq 45 for MSI/MSI-X
[    9.700498] ------------[ cut here ]------------
[    9.703391] kernel BUG at include/linux/netdevice.h:501!
[    9.703391] invalid opcode: 0000 [#1] PREEMPT SMP
<snip>
[    9.830018] Call Trace:
[    9.830018]  [<fa996169>] sky2_open+0x309/0x360 [sky2]
[    9.830018]  [<c1007210>] ? via_no_dac+0x40/0x40
[    9.830018]  [<c1007210>] ? via_no_dac+0x40/0x40
[    9.830018]  [<c135ed4b>] __dev_open+0x9b/0x120
[    9.830018]  [<c1431cbe>] ? _raw_spin_unlock_bh+0x1e/0x20
[    9.830018]  [<c135efd9>] __dev_change_flags+0x89/0x150
[    9.830018]  [<c135f148>] dev_change_flags+0x18/0x50
[    9.830018]  [<c13bb8e0>] devinet_ioctl+0x5d0/0x6e0
[    9.830018]  [<c13bcced>] inet_ioctl+0x6d/0xa0

To fix the problem patch changes the order of initialization.

Bug report:
https://bugzilla.kernel.org/show_bug.cgi?id=67151

Quoted 6 lines
Quoted 2 lines
Reported-and-tested-by [email protected]
Signed-off-by Stanislaw Gruszka <[email protected]>

That looks good, problem was introduced years ago.
I wonder if netif_napi_del() should be in unwind if registration of either devices fails?

napif_napi_del() is called internally by free_netdev(), so we are fine regarding error unwind and driver remove.

Thanks
Stanislaw

maintainer Applied
#3 lore raw
In reply to Stanislaw Gruszka

From: Stanislaw Gruszka <[email protected]>
Date: Sat, 25 Jan 2014 11:34:54 +0100

Quoted 3 lines

There is race condition when call netif_napi_add() after register_netdevice(), as ->open() can be called without napi initialized and trigger BUG_ON() on napi_enable(), like on below messages:

...

Quoted 7 lines

To fix the problem patch changes the order of initialization.

Bug report:
https://bugzilla.kernel.org/show_bug.cgi?id=67151

Reported-and-tested-by [email protected]
Signed-off-by Stanislaw Gruszka <[email protected]>

Applied, thanks a lot.