Skip to contentSkip to conversation

usb/gadget/function/u_ether.c: Allow jumbo frames

PATCH In review Mike Looijmans 2 participants 4 messages Aug 5, 2015 – Sep 3, 2015 USB

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 usb/gadget/function/u_ether.c: Allow jumbo frames

USB network adapters support Jumbo frames. The only thing blocking that feature is the code in the gadget driver that disposes of packets larger than 1518 bytes, and the limit on the ioctl to set the mtu.

This patch relaxes these limits, and allows up to 15k frames sizes. The 15k value was chosen because 16k does not work on all platforms, and usingclose to 16k will result in allocating 5 or 8 4k pages to store the skb, wasting pages at no measurable performance gain.

On a topic-miami board (Zynq-7000), iperf3 performance reports:
MTU= 1500, PC-to-gadget: 139 Mbps, Gadget-to-PC: 116 Mbps MTU=15000, PC-to-gadget: 239 Mbps, Gadget-to-PC: 361 Mbps

On boards with slower CPUs the performance improvement will be relatively much larger, e.g. an OMAP-L138 increased from 40 to 220 Mbps using a similar patch on an 2.6.37 kernel.

---
 drivers/usb/gadget/function/u_ether.c | 9 +++++++--
 1 file changed, 7 insertions(+), 2 deletions(-)
drivers/usb/gadget/function/u_ether.c +7 -2
index f1fd777..6828ea2 100644
@@ -48,6 +48,11 @@
 
 #define UETH__VERSION	"29-May-2008"
 
+/* Experiments show that both Linux and Windows hosts allow up to 16k
+ * frame sizes. Set the max size to 15k+52 to prevent allocating 32k
+ * blocks and still have efficient handling. */
+#define GETHER_MAX_ETH_FRAME_LEN 15412
+
 struct eth_dev {
 	/* lock is held while accessing port_usb
 	 */
@@ -146,7 +151,7 @@ static int ueth_change_mtu(struct net_device *net, int new_mtu)
 	spin_lock_irqsave(&dev->lock, flags);
 	if (dev->port_usb)
 		status = -EBUSY;
-	else if (new_mtu <= ETH_HLEN || new_mtu > ETH_FRAME_LEN)
+	else if (new_mtu <= ETH_HLEN || new_mtu > GETHER_MAX_ETH_FRAME_LEN)
 		status = -ERANGE;
 	else
 		net->mtu = new_mtu;
@@ -294,7 +299,7 @@ static void rx_complete(struct usb_ep *ep, struct usb_request *req)
 		while (skb2) {
 			if (status < 0
 					|| ETH_HLEN > skb2->len
-					|| skb2->len > VLAN_ETH_FRAME_LEN) {
+					|| skb2->len > GETHER_MAX_ETH_FRAME_LEN) {
 				dev->net->stats.rx_errors++;
 				dev->net->stats.rx_length_errors++;
 				DBG(dev, "rx length %d\n", skb2->len);
Signature · 1 line
1.9.1
#1 lore raw

ping...

More than a week has passed, haven't seen any response though. Any comments?

Quoted 50 lines · Mike Looijmans

USB network adapters support Jumbo frames. The only thing blocking that feature is the code in the gadget driver that disposes of packets larger than 1518 bytes, and the limit on the ioctl to set the mtu.

This patch relaxes these limits, and allows up to 15k frames sizes. The 15k value was chosen because 16k does not work on all platforms, and usingclose to 16k will result in allocating 5 or 8 4k pages to store the skb, wasting pages at no measurable performance gain.

On a topic-miami board (Zynq-7000), iperf3 performance reports:
MTU= 1500, PC-to-gadget: 139 Mbps, Gadget-to-PC: 116 Mbps MTU=15000, PC-to-gadget: 239 Mbps, Gadget-to-PC: 361 Mbps

On boards with slower CPUs the performance improvement will be relatively much larger, e.g. an OMAP-L138 increased from 40 to 220 Mbps using a similar patch on an 2.6.37 kernel.

---
  drivers/usb/gadget/function/u_ether.c | 9 +++++++--
  1 file changed, 7 insertions(+), 2 deletions(-)
drivers/usb/gadget/function/u_ether.c +6 -1
index f1fd777..6828ea2 100644
@@ -48,6 +48,11 @@
 
  #define UETH__VERSION	"29-May-2008"
 
+/* Experiments show that both Linux and Windows hosts allow up to 16k
+ * frame sizes. Set the max size to 15k+52 to prevent allocating 32k
+ * blocks and still have efficient handling. */
+#define GETHER_MAX_ETH_FRAME_LEN 15412
+
  struct eth_dev {
  	/* lock is held while accessing port_usb
  	 */
@@ -146,7 +151,7 @@ static int ueth_change_mtu(struct net_device *net, int new_mtu)
  	spin_lock_irqsave(&dev->lock, flags);
  	if (dev->port_usb)
  		status = -EBUSY;
-	else if (new_mtu <= ETH_HLEN || new_mtu > ETH_FRAME_LEN)
+	else if (new_mtu <= ETH_HLEN || new_mtu > GETHER_MAX_ETH_FRAME_LEN)
  		status = -ERANGE;
  	else
  		net->mtu = new_mtu;

drivers/usb/gadget/function/u_ether.c static void rx_complete(struct usb_ep *ep, struct usb_request *req) · line 305

Quoted 8 lines
drivers/usb/gadget/function/u_ether.c +1 -1
@@ -294,7 +299,7 @@ static void rx_complete(struct usb_ep *ep, struct usb_request *req)
  		while (skb2) {
  			if (status < 0
  					|| ETH_HLEN > skb2->len
-					|| skb2->len > VLAN_ETH_FRAME_LEN) {
+					|| skb2->len > GETHER_MAX_ETH_FRAME_LEN) {
  				dev->net->stats.rx_errors++;
  				dev->net->stats.rx_length_errors++;
  				DBG(dev, "rx length %d\n", skb2->len);

Kind regards,

Mike Looijmans
System Expert

TOPIC Embedded Products Eindhovenseweg 32-C, NL-5683 KH Best Postbus 440, NL-5680 AK Best
Telefoon: +31 (0) 499 33 69 79
Telefax: +31 (0) 499 33 69 70
E-mail: [email protected]
Website: www.topicproducts.com

Please consider the environment before printing this e-mail

#2 lore raw
In reply to Mike Looijmans

I'd like to bring this to your attention again please.

If there is something wrong about this patch, please tell me so.

And also note that this patch does not enable jumbo frames on itself, it just removes the artificial limits in the kernel prohibiting mtus above 1500. The MTU can be set from user space, and the impact on performance is quite impressive.

Quoted 50 lines · Mike Looijmans

USB network adapters support Jumbo frames. The only thing blocking that feature is the code in the gadget driver that disposes of packets larger than 1518 bytes, and the limit on the ioctl to set the mtu.

This patch relaxes these limits, and allows up to 15k frames sizes. The 15k value was chosen because 16k does not work on all platforms, and usingclose to 16k will result in allocating 5 or 8 4k pages to store the skb, wasting pages at no measurable performance gain.

On a topic-miami board (Zynq-7000), iperf3 performance reports:
MTU= 1500, PC-to-gadget: 139 Mbps, Gadget-to-PC: 116 Mbps MTU=15000, PC-to-gadget: 239 Mbps, Gadget-to-PC: 361 Mbps

On boards with slower CPUs the performance improvement will be relatively much larger, e.g. an OMAP-L138 increased from 40 to 220 Mbps using a similar patch on an 2.6.37 kernel.

---
  drivers/usb/gadget/function/u_ether.c | 9 +++++++--
  1 file changed, 7 insertions(+), 2 deletions(-)
drivers/usb/gadget/function/u_ether.c +6 -1
index f1fd777..6828ea2 100644
@@ -48,6 +48,11 @@
 
  #define UETH__VERSION	"29-May-2008"
 
+/* Experiments show that both Linux and Windows hosts allow up to 16k
+ * frame sizes. Set the max size to 15k+52 to prevent allocating 32k
+ * blocks and still have efficient handling. */
+#define GETHER_MAX_ETH_FRAME_LEN 15412
+
  struct eth_dev {
  	/* lock is held while accessing port_usb
  	 */
@@ -146,7 +151,7 @@ static int ueth_change_mtu(struct net_device *net, int new_mtu)
  	spin_lock_irqsave(&dev->lock, flags);
  	if (dev->port_usb)
  		status = -EBUSY;
-	else if (new_mtu <= ETH_HLEN || new_mtu > ETH_FRAME_LEN)
+	else if (new_mtu <= ETH_HLEN || new_mtu > GETHER_MAX_ETH_FRAME_LEN)
  		status = -ERANGE;
  	else
  		net->mtu = new_mtu;

drivers/usb/gadget/function/u_ether.c static void rx_complete(struct usb_ep *ep, struct usb_request *req) · line 305

Quoted 8 lines
drivers/usb/gadget/function/u_ether.c +1 -1
@@ -294,7 +299,7 @@ static void rx_complete(struct usb_ep *ep, struct usb_request *req)
  		while (skb2) {
  			if (status < 0
  					|| ETH_HLEN > skb2->len
-					|| skb2->len > VLAN_ETH_FRAME_LEN) {
+					|| skb2->len > GETHER_MAX_ETH_FRAME_LEN) {
  				dev->net->stats.rx_errors++;
  				dev->net->stats.rx_length_errors++;
  				DBG(dev, "rx length %d\n", skb2->len);

Kind regards,

Mike Looijmans
System Expert

TOPIC Embedded Products Eindhovenseweg 32-C, NL-5683 KH Best Postbus 440, NL-5680 AK Best
Telefoon: +31 (0) 499 33 69 79
Telefax: +31 (0) 499 33 69 70
E-mail: [email protected]
Website: www.topicproducts.com

Please consider the environment before printing this e-mail

maintainer
#3 lore raw
Quoted 3 lines · Mike Looijmans

I'd like to bring this to your attention again please.

If there is something wrong about this patch, please tell me so.

It's the middle of the merge window and we can't do anything with new patches at the moment until 4.3-rc1 comes out. Please be patient.

greg k-h