[PATCH v2 4/4] ath10k: fix issues on non-preemptible systems

Kalle Valo kvalo at qca.qualcomm.com
Tue Aug 27 03:06:13 EDT 2013


Michal Kazior <michal.kazior at tieto.com> writes:

> Workers may not call into a sleepable function
> (e.g. mutex_lock). Since ath10k workers can run
> for a very long time it is necessary to explicitly
> allow process rescheduling in case there's no
> preemption.
>
> This fixes some issues with system freezes, hangs,
> watchdogs being triggered, userspace being
> unresponsive on slow host machines under heavy
> load.

I consider this more as a workaround as a real fix. Would NAPI be a
proper fix for issues like this?

NAPI support was removed from mac80211 six months ago, but I guess we
could try to get it back if we have a good reason:

http://marc.info/?l=linux-wireless&m=136204135907491

> --- a/drivers/net/wireless/ath/ath10k/htt_rx.c
> +++ b/drivers/net/wireless/ath/ath10k/htt_rx.c
> @@ -1229,6 +1229,10 @@ static void ath10k_htt_rx_work(struct work_struct *work)
>  			break;
>  
>  		ath10k_htt_rx_process_skb(htt->ar, skb);
> +
> +#ifndef CONFIG_PREEMPT
> +		cond_resched();
> +#endif

Why the #ifndef? Why should we not call cond_resched() when PREEMPT is
enabled? Does something negative happen then?

-- 
Kalle Valo



More information about the ath10k mailing list