Re: ipw2100: race between isr_indicate_associated and rx path

[Helmut Schaa <helmut.schaa@googlemail.com>]

Am Freitag, 27. Februar 2009 schrieb Zhu Yi:
> On Wed, 2009-02-25 at 20:39 +0800, Helmut Schaa wrote:
> > Argh! Just found out why dev_activate is called such late after
> > netif_carrier_on:
> > 
> > ipw2100 calls netif_carrier_on followed by netif_wake_queue when the
> > driver
> > moves from associating to associated state. netif_carrier_on will then
> > call linkwatch_fire_event. However the carrier_on event is not treated
> > as
> > urgent and as such the event is delayed (and thus dev_activate too).
> > 
> > An event is considered urgent if the netdev is running, has a carrier
> > _and_ any of the TX qdiscs changed. Since ipw2100 first calls
> > carrier_on,
> > the last condition is not met and thus the event is not considered
> > urgent and gets delayed.
> 
> Calling netif_wake_queue() before netif_carrier_on() is not correct in
> semantics. Even it works, it looks like a workaround hack. I still think
> making a sync version of netif_carrier_on is the way to go.

Hmm, we could use register_netdevice_notifier to get notified when the
carrier event is sent out and thus the tx queues are available.

The flow in ipw2100 would then be:

1) isr_indicate_associated:
- move to ASSOCIATING state
- start buffering incoming frames

2) ipw2100_ex_event_work:
- call carrier_on and wake_queues

3) netdevice_notification:
- move to ASSOCIATED state and replay buffered frames

I hope I'll find some time to try that soon.

Helmut