* re-enable queues to get higher bandwidth utilization (without needing
* to tweak MTU for larger packets).
*
- * Add support for more "network cable" chips; interop with their Win32
- * drivers may be a good thing. Test the AnchorChip 2720 support..
- * Figure out the initialization protocol used by the Prolific chips,
- * for better robustness ... there's some powerup/reset handshake that's
- * needed when only one end reboots.
+ * - AN2720 ... not widely available, but reportedly works well
*
- * Use interrupt on PL230x to detect peer connect/disconnect, and call
- * netif_carrier_{on,off} (?) appropriately. For Net1080, detect peer
- * connect/disconnect with async control messages.
+ * - Belkin/eTEK ... no known issues
*
- * Find some way to report "peer connected" network hotplug events; it'll
- * likely mean updating the networking layer. (This has been discussed
- * on the netdev list...)
+ * - Both GeneSys and PL-230x use interrupt transfers for driver-to-driver
+ * handshaking; it'd be worth implementing those as "carrier detect".
+ * Prefer generic hooks, not minidriver-specific hacks.
+ *
+ * - Linux devices ... the www.handhelds.org SA-1100 support works nicely,
+ * but the Sharp Zaurus uses an incompatible protocol (extra checksums).
+ * No reason not to merge the Zaurus protocol here too (got patch? :)
+ *
+ * - For Netchip, use keventd to poll via control requests to detect hardware
+ * level "carrier detect".
+ *
+ * - PL-230x ... the initialization protocol doesn't seem to match chip data
+ * sheets, sometimes it's not needed and sometimes it hangs. Prolific has
+ * not responded to repeated support/information requests.
+ *
+ * Interop with more Win32 drivers may be a good thing.
+ *
+ * Seems like reporting "peer connected" (carrier present) events may end
+ * up going through the netlink event system, not hotplug ... that may be
+ * awkward in terms of automatic configuration though.
+ *
+ * There are reports that bridging gives lower-than-usual throughput.
*
* Craft smarter hotplug policy scripts ... ones that know how to arrange
* bridging with "brctl", and can handle static and dynamic ("pump") setups.
* Level of diagnostics is more configurable; they use device
* location (usb_device->devpath) instead of address (2.5).
* For tx_fixup, memflags can't be NOIO.
+ * 07-may-2002 Generalize/cleanup keventd support, handling rx stalls (mostly
+ * for USB 2.0 TTs) and memory shortages (potential) too. (db)
+ * Use "locally assigned" IEEE802 address space. (Brad Hards)
*
*-------------------------------------------------------------------------*/
#include <linux/usb.h>
+/* minidrivers _could_ be individually configured */
#define CONFIG_USB_AN2720
#define CONFIG_USB_BELKIN
#define CONFIG_USB_GENESYS
#define CONFIG_USB_PL2301
-#define DRIVER_VERSION "26-Apr-2002"
+#define DRIVER_VERSION "07-May-2002"
/*-------------------------------------------------------------------------*/
struct sk_buff_head txq;
struct sk_buff_head done;
struct tasklet_struct bh;
- struct tq_struct ctrl_task;
+
+ struct tq_struct kevent;
+ unsigned long flags;
+# define EVENT_TX_HALT 0
+# define EVENT_RX_HALT 1
+# define EVENT_RX_MEMORY 2
};
// device-specific info used by the driver
spin_unlock_irqrestore (&dev->done.lock, flags);
}
+/* some work can't be done in tasklets, so we use keventd
+ *
+ * NOTE: annoying asymmetry: if it's active, schedule_task() fails,
+ * but tasklet_schedule() doesn't. hope the failure is rare.
+ */
+static void defer_kevent (struct usbnet *dev, int work)
+{
+ set_bit (work, &dev->flags);
+ if (!schedule_task (&dev->kevent))
+ err ("%s: kevent %d may have been dropped",
+ dev->net.name, work);
+ else
+ dbg ("%s: kevent %d scheduled", dev->net.name, work);
+}
+
/*-------------------------------------------------------------------------*/
static void rx_complete (struct urb *urb);
if ((skb = alloc_skb (size, flags)) == 0) {
dbg ("no rx skb");
- tasklet_schedule (&dev->bh);
+ defer_kevent (dev, EVENT_RX_MEMORY);
usb_free_urb (urb);
return;
}
spin_lock_irqsave (&dev->rxq.lock, lockflags);
- if (netif_running (&dev->net)) {
- if ((retval = usb_submit_urb (urb, GFP_ATOMIC)) != 0) {
+ if (netif_running (&dev->net)
+ && !test_bit (EVENT_RX_HALT, &dev->flags)) {
+ switch (retval = usb_submit_urb (urb, GFP_ATOMIC)){
+ case -EPIPE:
+ defer_kevent (dev, EVENT_RX_HALT);
+ break;
+ case -ENOMEM:
+ defer_kevent (dev, EVENT_RX_MEMORY);
+ break;
+ default:
dbg ("%s rx submit, %d", dev->net.name, retval);
tasklet_schedule (&dev->bh);
- } else {
+ break;
+ case 0:
__skb_queue_tail (&dev->rxq, skb);
}
} else {
}
break;
+ // stalls need manual reset. this is rare ... except that
+ // when going through USB 2.0 TTs, unplug appears this way.
+ // we avoid the highspeed version of the ETIMEOUT/EILSEQ
+ // storm, recovering as needed.
+ case -EPIPE:
+ defer_kevent (dev, EVENT_RX_HALT);
+ // FALLTHROUGH
+
// software-driven interface shutdown
- case -ECONNRESET: // usb-ohci, usb-uhci
- case -ECONNABORTED: // uhci ... for usb-uhci, INTR
- dbg ("%s shutdown, code %d", dev->net.name, urb_status);
+ case -ECONNRESET: // according to API spec
+ case -ECONNABORTED: // some (now fixed?) UHCI bugs
+ dbg ("%s rx shutdown, code %d", dev->net.name, urb_status);
entry->state = rx_cleanup;
- // do urb frees only in the tasklet
+ // do urb frees only in the tasklet (UHCI has oopsed ...)
entry->urb = urb;
urb = 0;
break;
// FALLTHROUGH
default:
- // on unplug we'll get a burst of ETIMEDOUT/EILSEQ
- // till the khubd gets and handles its interrupt.
+ // on unplug we get ETIMEDOUT (ohci) or EILSEQ (uhci)
+ // until khubd sees its interrupt and disconnects us.
+ // that can easily be hundreds of passes through here.
entry->state = rx_cleanup;
dev->stats.rx_errors++;
dbg ("%s rx: status %d", dev->net.name, urb_status);
defer_bh (dev, skb);
if (urb) {
- if (netif_running (&dev->net)) {
+ if (netif_running (&dev->net)
+ && !test_bit (EVENT_RX_HALT, &dev->flags)) {
rx_submit (dev, urb, GFP_ATOMIC);
return;
}
+ usb_free_urb (urb);
}
#ifdef VERBOSE
dbg ("no read resubmitted");
// during some PM-driven resume scenarios,
// these (async) unlinks complete immediately
retval = usb_unlink_urb (urb);
- if (retval < 0)
+ if (retval != -EINPROGRESS && retval != 0)
dbg ("unlink urb err, %d", retval);
else
count++;
/*-------------------------------------------------------------------------*/
-/* usb_clear_halt cannot be called in interrupt context */
-
+/* work that cannot be done in interrupt context uses keventd.
+ *
+ * NOTE: "uhci" and "usb-uhci" may have trouble with this since they don't
+ * queue control transfers to individual devices, and other threads could
+ * trigger control requests concurrently. hope that's rare.
+ */
static void
-tx_clear_halt (void *data)
+kevent (void *data)
{
struct usbnet *dev = data;
+ int status;
- usb_clear_halt (dev->udev,
- usb_sndbulkpipe (dev->udev, dev->driver_info->out));
- netif_wake_queue (&dev->net);
+ /* usb_clear_halt() needs a thread context */
+ if (test_bit (EVENT_TX_HALT, &dev->flags)) {
+ unlink_urbs (&dev->txq);
+ status = usb_clear_halt (dev->udev,
+ usb_sndbulkpipe (dev->udev, dev->driver_info->out));
+ if (status < 0)
+ err ("%s: can't clear tx halt, status %d",
+ dev->net.name, status);
+ else {
+ clear_bit (EVENT_TX_HALT, &dev->flags);
+ netif_wake_queue (&dev->net);
+ }
+ }
+ if (test_bit (EVENT_RX_HALT, &dev->flags)) {
+ unlink_urbs (&dev->rxq);
+ status = usb_clear_halt (dev->udev,
+ usb_rcvbulkpipe (dev->udev, dev->driver_info->in));
+ if (status < 0)
+ err ("%s: can't clear rx halt, status %d",
+ dev->net.name, status);
+ else {
+ clear_bit (EVENT_RX_HALT, &dev->flags);
+ tasklet_schedule (&dev->bh);
+ }
+ }
+
+ /* tasklet could resubmit itself forever if memory is tight */
+ if (test_bit (EVENT_RX_MEMORY, &dev->flags)) {
+ struct urb *urb = 0;
+
+ if (netif_running (&dev->net))
+ urb = usb_alloc_urb (0, GFP_KERNEL);
+ else
+ clear_bit (EVENT_RX_MEMORY, &dev->flags);
+ if (urb != 0) {
+ clear_bit (EVENT_RX_MEMORY, &dev->flags);
+ rx_submit (dev, urb, GFP_KERNEL);
+ tasklet_schedule (&dev->bh);
+ }
+ }
+
+ if (dev->flags)
+ dbg ("%s: kevent done, flags = 0x%lx",
+ dev->net.name, dev->flags);
}
/*-------------------------------------------------------------------------*/
struct skb_data *entry = (struct skb_data *) skb->cb;
struct usbnet *dev = entry->dev;
- if (urb->status == -EPIPE) {
- if (dev->ctrl_task.sync == 0) {
- dev->ctrl_task.routine = tx_clear_halt;
- dev->ctrl_task.data = dev;
- schedule_task (&dev->ctrl_task);
- } else {
- dbg ("Cannot clear TX stall");
- }
- }
+ if (urb->status == -EPIPE)
+ defer_kevent (dev, EVENT_TX_HALT);
urb->dev = 0;
entry->state = tx_done;
defer_bh (dev, skb);
#endif /* CONFIG_USB_NET1080 */
netif_stop_queue (net);
- if ((retval = usb_submit_urb (urb, GFP_ATOMIC)) != 0) {
+ switch ((retval = usb_submit_urb (urb, GFP_ATOMIC))) {
+ case -EPIPE:
+ defer_kevent (dev, EVENT_TX_HALT);
+ break;
+ default:
netif_start_queue (net);
dbg ("%s tx: submit urb err %d", net->name, retval);
- } else {
+ break;
+ case 0:
net->trans_start = jiffies;
__skb_queue_tail (&dev->txq, skb);
if (dev->txq.qlen < TX_QLEN)
}
// or are we maybe short a few urbs?
- } else if (netif_running (&dev->net)) {
+ } else if (netif_running (&dev->net)
+ && !test_bit (EVENT_RX_HALT, &dev->flags)) {
int temp = dev->rxq.qlen;
if (temp < RX_QLEN) {
list_del (&dev->dev_list);
mutex_unlock (&usbnet_mutex);
+ // assuming we used keventd, it must quiesce too
+ flush_scheduled_tasks ();
+
kfree (dev);
usb_dec_dev_use (udev);
}
skb_queue_head_init (&dev->done);
dev->bh.func = usbnet_bh;
dev->bh.data = (unsigned long) dev;
+ INIT_TQUEUE (&dev->kevent, kevent, dev);
// set up network interface records
net = &dev->net;
get_random_bytes (node_id, sizeof node_id);
node_id [0] &= 0xfe; // clear multicast bit
+ node_id [0] |= 0x02; // set local assignment bit (IEEE802)
if (usb_register (&usbnet_driver) < 0)
return -1;
projects / linux / commitdiff