On Thursday 07 January 2010 21:16, Wolfgang Grandegger wrote:
> Oliver Hartkopp wrote:
> > Matthias Fuchs wrote:
> >
> >> But I ran into some issues. Perhaps some of you can help:
> >> Q1: Under high TX load the driver runs out of free tx_contexts.
> >> (see "couldn't find free context"). I expected the socket can core
> >> to take care of this so that the lower driver cannot be overrun
> >> by tx jobs. Either this expectation is wrong or there's a bug in
> >> the driver :-)
> >
> > Hi Matthias,
> >
> > i wonder, why you do not stop the netdev tx queue immediately at the
> > beginning of
> >
> >> +static netdev_tx_t esd_usb2_start_xmit(struct sk_buff *skb,
> >> + struct net_device *netdev)
> >
> > but at this (late) point:
> >
> >> + if (err) {
> >> + can_free_echo_skb(netdev, context->echo_index);
> >> +
> >> + atomic_dec(&priv->active_tx_urbs);
> >> + usb_unanchor_urb(urb);
> >> +
> >> + if (err == -ENODEV)
> >> + netif_device_detach(netdev);
> >> + else
> >> + dev_warn(ND2D(netdev), "failed tx_urb %d\n", err);
> >> +
> >> + goto releasebuf;
> >> + } else {
> >> + netdev->trans_start = jiffies;
> >> +
> >> + /* Slow down tx path */
> >> + if (atomic_read(&priv->active_tx_urbs) >= MAX_TX_URBS)
> >> + netif_stop_queue(netdev);
> >> + }
>
> The driver queues more than one TX packet for the sake of throughput.
> Sebastian mentioned that there is a big improvement with 10 TX URBs.
>
> > The netdev queues in Linux will give you skbs until you say 'stop' :-)
> > This is not SocketCAN specific.
> >
> > I assume the handling of netif_[wake|stop]_queue() should be reworked.
> >
> > E.g. invoke netif_stop_queue() at the beginning of esd_usb2_start_xmit() and
> > when everything is really fine, you can invoke netif_wake_queue() at the and
> > of esd_usb2_start_xmit() again.
> >
> > Btw. did you check the above code with
> >
> > if (atomic_read(&priv->active_tx_urbs) >= MAX_TX_URBS - 1)
> >
> > to test for an off-by-one problem in the tx_urbs?
>
> So far and in the EMS USB driver, the stop/wakeup is handled based on
> the "priv->active_tx_urbs" count. If there are no more tx_urbs, netdev
> will be stopped and restarted later in esd_usb2_write_bulk_callback().
> At a first glance, I think there's the problem. The context needs to be
> stored until esd_usb2_tx_done_msg() is called. That function must also
> wakeup the netdev queue.
Yes your are right. I already fixed that and will post a new patch in a couple
of minutes.
Matthias
_______________________________________________
Socketcan-core mailing list
[email protected]
https://lists.berlios.de/mailman/listinfo/socketcan-core
