In order to reduce the interrupt times in the embedded system,
a receiving workqueue is introduced.
This modification also enhanced the overall throughput as the
benefits of reducing interrupt occurrence.
This work was derived from previous work:
u_ether: move hardware transmit to RX workqueue.
Which should be base on codeaurora's work.
This patch replaced the rx workqueue into rx NAPI.
However, the benchmark on my platform shows the throughput
with workqueue is slightly better than NAPI.
Signed-off-by: Weinn Jheng <clanlab.proj@xxxxxxxxx>
Cc: David S. Miller <davem@xxxxxxxxxxxxx>
Cc: Stephen Hemminger <shemminger@xxxxxxxxxx>
Cc: Felipe Balbi <balbi@xxxxxx>
Cc: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx>
Cc: Manu Gautam <mgautam@xxxxxxxxxxxxxx>
Cc: Andrzej Pietrasiewicz <andrzej.p@xxxxxxxxxxx>
---
Changes for v2:
- Trying to fix the plug/unplug problem which cause KE.
drivers/usb/gadget/u_ether.c | 140 +++++++++++++++++++++++++++----------------
1 file changed, 88 insertions(+), 52 deletions(-)
diff --git a/drivers/usb/gadget/u_ether.c b/drivers/usb/gadget/u_ether.c
index b7d4f82..104fdde 100644
--- a/drivers/usb/gadget/u_ether.c
+++ b/drivers/usb/gadget/u_ether.c
@@ -48,6 +48,8 @@
#define UETH__VERSION "29-May-2008"
+#define GETHER_NAPI_WEIGHT 32
+
struct eth_dev {
/* lock is held while accessing port_usb
*/
@@ -72,6 +74,7 @@ struct eth_dev {
struct sk_buff_head *list);
struct work_struct work;
+ struct napi_struct rx_napi;
unsigned long todo;
#define WORK_RX_MEMORY 0
@@ -262,9 +265,10 @@ enomem:
static void rx_complete(struct usb_ep *ep, struct usb_request *req)
{
- struct sk_buff *skb = req->context, *skb2;
+ struct sk_buff *skb = req->context;
struct eth_dev *dev = ep->driver_data;
int status = req->status;
+ bool rx_queue = 0;
switch (status) {
@@ -288,30 +292,8 @@ static void rx_complete(struct usb_ep *ep, struct usb_request *req)
} else {
skb_queue_tail(&dev->rx_frames, skb);
}
- skb = NULL;
-
- skb2 = skb_dequeue(&dev->rx_frames);
- while (skb2) {
- if (status < 0
- || ETH_HLEN > skb2->len
- || skb2->len > VLAN_ETH_FRAME_LEN) {
- dev->net->stats.rx_errors++;
- dev->net->stats.rx_length_errors++;
- DBG(dev, "rx length %d\n", skb2->len);
- dev_kfree_skb_any(skb2);
- goto next_frame;
- }
- skb2->protocol = eth_type_trans(skb2, dev->net);
- dev->net->stats.rx_packets++;
- dev->net->stats.rx_bytes += skb2->len;
-
- /* no buffer copies needed, unless hardware can't
- * use skb buffers.
- */
- status = netif_rx(skb2);
-next_frame:
- skb2 = skb_dequeue(&dev->rx_frames);
- }
+ if (!status)
+ rx_queue = 1;
break;
/* software-driven interface shutdown */
@@ -334,22 +316,20 @@ quiesce:
/* FALLTHROUGH */
default:
+ rx_queue = 1;
+ dev_kfree_skb_any(skb);
dev->net->stats.rx_errors++;
DBG(dev, "rx status %d\n", status);
break;
}
- if (skb)
- dev_kfree_skb_any(skb);
- if (!netif_running(dev->net)) {
clean:
spin_lock(&dev->req_lock);
list_add(&req->list, &dev->rx_reqs);
spin_unlock(&dev->req_lock);
- req = NULL;
- }
- if (req)
- rx_submit(dev, req, GFP_ATOMIC);
+
+ if (rx_queue && likely(napi_schedule_prep(&dev->rx_napi)))
+ __napi_schedule(&dev->rx_napi);
}
static int prealloc(struct list_head *list, struct usb_ep *ep, unsigned n)
@@ -414,16 +394,24 @@ static void rx_fill(struct eth_dev *dev, gfp_t gfp_flags)
{
struct usb_request *req;
unsigned long flags;
+ int rx_counts = 0;
/* fill unused rxq slots with some skb */
spin_lock_irqsave(&dev->req_lock, flags);
while (!list_empty(&dev->rx_reqs)) {
+
+ if (++rx_counts > qlen(dev->gadget, dev->qmult))
+ break;
+
req = container_of(dev->rx_reqs.next,
struct usb_request, list);
list_del_init(&req->list);
spin_unlock_irqrestore(&dev->req_lock, flags);
if (rx_submit(dev, req, gfp_flags) < 0) {
+ spin_lock_irqsave(&dev->req_lock, flags);
+ list_add(&req->list, &dev->rx_reqs);
+ spin_unlock_irqrestore(&dev->req_lock, flags);
defer_kevent(dev, WORK_RX_MEMORY);
return;
}
@@ -433,6 +421,41 @@ static void rx_fill(struct eth_dev *dev, gfp_t gfp_flags)
spin_unlock_irqrestore(&dev->req_lock, flags);
}
+static int gether_poll(struct napi_struct *napi, int budget)
+{
+ struct eth_dev *dev = container_of(napi, struct eth_dev, rx_napi);
+ struct sk_buff *skb;
+ unsigned int work_done = 0;
+ int status = 0;
+
+ while ((skb = skb_dequeue(&dev->rx_frames))) {
+ if (status < 0
+ || ETH_HLEN > skb->len
+ || skb->len > VLAN_ETH_FRAME_LEN) {
+ dev->net->stats.rx_errors++;
+ dev->net->stats.rx_length_errors++;
+ DBG(dev, "rx length %d\n", skb->len);
+ dev_kfree_skb_any(skb);
+ continue;
+ }
+ skb->protocol = eth_type_trans(skb, dev->net);
+ dev->net->stats.rx_packets++;
+ dev->net->stats.rx_bytes += skb->len;
+
+ status = netif_rx_ni(skb);
+ }
+
+ if (netif_running(dev->net)) {
+ rx_fill(dev, GFP_KERNEL);
+ work_done++;
+ }
+
+ if (work_done < budget)
+ napi_complete(&dev->rx_napi);
+
+ return work_done;
+}
+
static void eth_work(struct work_struct *work)
{
struct eth_dev *dev = container_of(work, struct eth_dev, work);
@@ -550,18 +573,18 @@ static netdev_tx_t eth_start_xmit(struct sk_buff *skb,
* or the hardware can't use skb buffers.
* or there's not enough space for extra headers we need
*/
+ spin_lock_irqsave(&dev->lock, flags);
if (dev->wrap) {
- unsigned long flags;
-
- spin_lock_irqsave(&dev->lock, flags);
if (dev->port_usb)
skb = dev->wrap(dev->port_usb, skb);
- spin_unlock_irqrestore(&dev->lock, flags);
- if (!skb)
+ if (!skb) {
+ spin_unlock_irqrestore(&dev->lock, flags);
goto drop;
-
- length = skb->len;
+ }
}
+
+ spin_unlock_irqrestore(&dev->lock, flags);
+ length = skb->len;
req->buf = skb->data;
req->context = skb;
req->complete = tx_complete;
@@ -625,6 +648,7 @@ static void eth_start(struct eth_dev *dev, gfp_t gfp_flags)
/* and open the tx floodgates */
atomic_set(&dev->tx_qlen, 0);
netif_wake_queue(dev->net);
+ napi_enable(&dev->rx_napi);
}
static int eth_open(struct net_device *net)
@@ -651,6 +675,7 @@ static int eth_stop(struct net_device *net)
unsigned long flags;
VDBG(dev, "%s\n", __func__);
+ napi_disable(&dev->rx_napi);
netif_stop_queue(net);
DBG(dev, "stop stats: rx/tx %ld/%ld, errs %ld/%ld\n",
@@ -662,8 +687,6 @@ static int eth_stop(struct net_device *net)
spin_lock_irqsave(&dev->lock, flags);
if (dev->port_usb) {
struct gether *link = dev->port_usb;
- const struct usb_endpoint_descriptor *in;
- const struct usb_endpoint_descriptor *out;
if (link->close)
link->close(link);
@@ -677,14 +700,18 @@ static int eth_stop(struct net_device *net)
* their own pace; the network stack can handle old packets.
* For the moment we leave this here, since it works.
*/
- in = link->in_ep->desc;
- out = link->out_ep->desc;
usb_ep_disable(link->in_ep);
usb_ep_disable(link->out_ep);
if (netif_carrier_ok(net)) {
+ if (config_ep_by_speed(dev->gadget, &link->func,
+ link->in_ep) ||
+ config_ep_by_speed(dev->gadget, &link->func,
+ link->out_ep)) {
+ link->in_ep->desc = NULL;
+ link->out_ep->desc = NULL;
+ return -EINVAL;
+ }
DBG(dev, "host still using in/out endpoints\n");
- link->in_ep->desc = in;
- link->out_ep->desc = out;
usb_ep_enable(link->in_ep);
usb_ep_enable(link->out_ep);
}
@@ -768,6 +795,7 @@ struct eth_dev *gether_setup_name(struct usb_gadget *g,
return ERR_PTR(-ENOMEM);
dev = netdev_priv(net);
+ netif_napi_add(net, &dev->rx_napi, gether_poll, GETHER_NAPI_WEIGHT);
spin_lock_init(&dev->lock);
spin_lock_init(&dev->req_lock);
INIT_WORK(&dev->work, eth_work);
@@ -795,6 +823,13 @@ struct eth_dev *gether_setup_name(struct usb_gadget *g,
SET_ETHTOOL_OPS(net, &ops);
+ /*
+ * two kinds of host-initiated state changes:
+ * - iff DATA transfer is active, carrier is "on"
+ * - tx queueing enabled if open *and* carrier is "on"
+ */
+ netif_carrier_off(net);
+
dev->gadget = g;
SET_NETDEV_DEV(net, &g->dev);
SET_NETDEV_DEVTYPE(net, &gadget_type);
@@ -807,13 +842,6 @@ struct eth_dev *gether_setup_name(struct usb_gadget *g,
} else {
INFO(dev, "MAC %pM\n", net->dev_addr);
INFO(dev, "HOST MAC %pM\n", dev->host_mac);
-
- /*
- * two kinds of host-initiated state changes:
- * - iff DATA transfer is active, carrier is "on"
- * - tx queueing enabled if open *and* carrier is "on"
- */
- netif_carrier_off(net);
}
return dev;
@@ -830,6 +858,7 @@ struct net_device *gether_setup_name_default(const char *netname)
return ERR_PTR(-ENOMEM);
dev = netdev_priv(net);
+ netif_napi_add(net, &dev->rx_napi, gether_poll, GETHER_NAPI_WEIGHT);
spin_lock_init(&dev->lock);
spin_lock_init(&dev->req_lock);
INIT_WORK(&dev->work, eth_work);
@@ -1113,6 +1142,7 @@ void gether_disconnect(struct gether *link)
{
struct eth_dev *dev = link->ioport;
struct usb_request *req;
+ struct sk_buff *skb;
WARN_ON(!dev);
if (!dev)
@@ -1154,6 +1184,12 @@ void gether_disconnect(struct gether *link)
spin_lock(&dev->req_lock);
}
spin_unlock(&dev->req_lock);
+
+ spin_lock(&dev->rx_frames.lock);
+ while ((skb = __skb_dequeue(&dev->rx_frames)))
+ dev_kfree_skb_any(skb);
+ spin_unlock(&dev->rx_frames.lock);
+
link->out_ep->driver_data = NULL;
link->out_ep->desc = NULL;
--
1.9.rc1
--
To unsubscribe from this list: send the line "unsubscribe linux-usb" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
