usbnet: ipheth: fix racing condition

Commit Message

Bernd Eckstein May 20, 2019, 3:31 p.m. UTC

Fix a racing condition in ipheth.c that can lead to slow performance.

Bug: In ipheth_tx(), netif_wake_queue() may be called on the callback
ipheth_sndbulk_callback(), _before_ netif_stop_queue() is called.
When this happens, the queue is stopped longer than it needs to be,
thus reducing network performance.

Fix: Move netif_stop_queue() in front of usb_submit_urb(). Now the order
is always correct. In case, usb_submit_urb() fails, the queue is woken up
again as callback will not fire.

Testing: This racing condition is usually not noticeable, as it has to
occur very frequently to slowdown the network. The callback from the USB
is usually triggered slow enough, so the situation does not appear.
However, on a Ubuntu Linux on VMWare Workstation, running on Windows 10,
the we loose the race quite often and the following speedup can be noticed:

Without this patch: Download:  4.10 Mbit/s, Upload:  4.01 Mbit/s
With this patch:    Download: 36.23 Mbit/s, Upload: 17.61 Mbit/s

Signed-off-by: Oliver Zweigle <[e-mail]>
Signed-off-by: Bernd Eckstein <[e-mail]>

---
 drivers/net/usb/ipheth.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

Comments

David Miller May 21, 2019, 8:11 p.m. UTC | #1

From: Bernd Eckstein <[e-mail]>
Date: Mon, 20 May 2019 17:31:09 +0200

> Fix a racing condition in ipheth.c that can lead to slow performance.
> 
> Bug: In ipheth_tx(), netif_wake_queue() may be called on the callback
> ipheth_sndbulk_callback(), _before_ netif_stop_queue() is called.
> When this happens, the queue is stopped longer than it needs to be,
> thus reducing network performance.
> 
> Fix: Move netif_stop_queue() in front of usb_submit_urb(). Now the order
> is always correct. In case, usb_submit_urb() fails, the queue is woken up
> again as callback will not fire.
> 
> Testing: This racing condition is usually not noticeable, as it has to
> occur very frequently to slowdown the network. The callback from the USB
> is usually triggered slow enough, so the situation does not appear.
> However, on a Ubuntu Linux on VMWare Workstation, running on Windows 10,
> the we loose the race quite often and the following speedup can be noticed:
> 
> Without this patch: Download:  4.10 Mbit/s, Upload:  4.01 Mbit/s
> With this patch:    Download: 36.23 Mbit/s, Upload: 17.61 Mbit/s
> 
> Signed-off-by: Oliver Zweigle <[e-mail]>
> Signed-off-by: Bernd Eckstein <[e-mail]>

Applied.

Patch

diff --git a/drivers/net/usb/ipheth.c b/drivers/net/usb/ipheth.c
index c247aed..8c01fbf 100644
--- a/drivers/net/usb/ipheth.c
+++ b/drivers/net/usb/ipheth.c
@@ -383,17 +383,18 @@  static int ipheth_tx(struct sk_buff *skb, struct net_device *net)
 			  dev);
 	dev->tx_urb->transfer_flags |= URB_NO_TRANSFER_DMA_MAP;
 
+	netif_stop_queue(net);
 	retval = usb_submit_urb(dev->tx_urb, GFP_ATOMIC);
 	if (retval) {
 		dev_err(&dev->intf->dev, "%s: usb_submit_urb: %d\n",
 			__func__, retval);
 		dev->net->stats.tx_errors++;
 		dev_kfree_skb_any(skb);
+		netif_wake_queue(net);
 	} else {
 		dev->net->stats.tx_packets++;
 		dev->net->stats.tx_bytes += skb->len;
 		dev_consume_skb_any(skb);
-		netif_stop_queue(net);
 	}
 
 	return NETDEV_TX_OK;