Messages in this thread |  | | | Date | Fri, 8 Jun 2012 22:41:54 +0200 (CEST) | | From | Thomas Gleixner <> | | Subject | Re: NOHZ: local_softirq_pending 08 |
| |
On Fri, 8 Jun 2012, Francois Romieu wrote:
> Thomas Gleixner <tglx@linutronix.de> : > [...] > > I can't find the point where the warning is issued, but I think I > > found the cause of the problem. > > > > static void rtl_slow_event_work(struct rtl8169_private *tp) > > { > > ..... > > napi_schedule(&tp->napi); > > --> __napi_schedule(); > > --> list_add_tail(&napi->poll_list, &sd->poll_list); > > __raise_softirq_irqoff(NET_RX_SOFTIRQ); > > > > This merily sets the softirq bit. > > > > > > So this code is really wrong. It's called from full preemptible > > context of the workqueue. And if the next thing is a context switch to > > idle then the pending softirq check will trigger. > > void __napi_schedule(struct napi_struct *n) > { > unsigned long flags; > > local_irq_save(flags); > ____napi_schedule(&__get_cpu_var(softnet_data), n); > local_irq_restore(flags); > } > > Are you saying that this stuff should be considered "preemptible" ?
Gah, crap. Looked at the wrong ___________underscore level. I _____so __love _that. Though the problem is, that it is neither called in interrupt context nor with bh disabled, so nothing invokes the softirq before it reaches idle.
In hard interrupt context the pending flag is evaluated in irq_exit() and the softirqs are invoked from there. If you call that from thread context, then a bh_disable/enable pair will make sure that the pending softirq is invoked. Did I miss some more ___underscore magic which does that ?
Thanks,
tglx
|  |