Hi Alex,
> > @@ -84,7 +118,7 @@ void ieee802154_xmit_complete(struct ieee802154_hw *hw, struct sk_buff *skb,
> > hw->phy->sifs_period * NSEC_PER_USEC,
> > HRTIMER_MODE_REL);
> > } else {
> > - ieee802154_wake_queue(hw);
> > + ieee802154_release_queue(local);
> > }
> >
> > dev_consume_skb_any(skb);
> > @@ -98,7 +132,7 @@ void ieee802154_xmit_error(struct ieee802154_hw *hw, struct sk_buff *skb,
> > struct ieee802154_local *local = hw_to_local(hw);
> >
> > local->tx_result = reason;
> > - ieee802154_wake_queue(hw);
> > + ieee802154_release_queue(local);
> > dev_kfree_skb_any(skb);
> > atomic_dec(&hw->phy->ongoing_txs);
>
> I am pretty sure that will end in a scheduling while atomic warning
> with hwsim. If you don't hit it you have the wrong config, you need to
> enable such warnings and have the right preemption model setting.
I was using the "desktop" kernel preemption model (voluntary), I've
switched to CONFIG_PREEMPT ("Preemptible kernel (Low-latency)"),
and enabled CONFIG_DEBUG_ATOMIC_SLEEP. You are right that we should use
a spinlock instead of a mutex here. However I don't think disabling
IRQs is necessary, so I'll switch to spin_(un)lock() calls.
> These calls xmit complete/error should even be allowed to be called
> from a hardware irq context, however I _think_ we don't have a driver
> which currently does that, but the mutex will break stuff here in the
> xmit_do() callback of netdev which hwsim is calling it from.
>
> Please check again...
Will perform a new set of test runs with the new configuration, yes.
Thanks,
Miquèl
