PREEMPT_RT with e1000

PREEMPT_RT with e1000

[Daniel Walker]

It looks like Gigabit Ethernet is still having some problems. This is
with the e1000 driver. If I remove all the qdisc_restart changes it
starts to work the warning below goes away, but it has smp_processor_id
warnings.

Daniel


BUG: softirq-net-tx/:5, possible softlockup detected on CPU#0!

Call Trace: <IRQ> <ffffffff8015b92a>{softlockup_detected+56} <ffffffff8015b9d7>{softlockup_tick+164}
       <ffffffff80119a4b>{smp_apic_timer_interrupt+55} <ffffffff8010e6a4>{apic_timer_interrupt+132}
        <EOI> <ffffffff80516f9b>{__down_mutex+694} <ffffffff80482e4d>{qdisc_restart+369}
       <ffffffff80476198>{net_tx_action+200} <ffffffff8013a802>{ksoftirqd+231}
       <ffffffff8013a71b>{ksoftirqd+0} <ffffffff8014a769>{kthread+218}
       <ffffffff8010e9f6>{child_rip+8} <ffffffff8014a68f>{kthread+0}
       <ffffffff8010e9ee>{child_rip+0}
---------------------------
| preempt count: 00010000 ]
| 0-level deep critical section nesting:
----------------------------------------

Re: PREEMPT_RT with e1000

[Ingo Molnar]

* Daniel Walker <dwalker@mvista.com> wrote:

> It looks like Gigabit Ethernet is still having some problems. This is 
> with the e1000 driver. If I remove all the qdisc_restart changes it 
> starts to work the warning below goes away, but it has 
> smp_processor_id warnings.

hmmm. The thing to watch out for there is dev->xmit_lock_owner - it's a 
"spin lock does not preempt" assumption that is open-coded. What the 
qdisc_restart changes do is to remove this assumption. Perhaps it would 
be enough to just keep the trylock logic, and to mute the 
xmit_lock_owner logic.

	Ingo

Re: PREEMPT_RT with e1000

[Ingo Molnar]

* Daniel Walker <dwalker@mvista.com> wrote:

> It looks like Gigabit Ethernet is still having some problems. This is 
> with the e1000 driver. If I remove all the qdisc_restart changes it 
> starts to work the warning below goes away, but it has 
> smp_processor_id warnings.

btw., what does "problems" mean, precisely - does it not work at all, or 
does it produce the lockup under certain loads?

	Ingo

Re: PREEMPT_RT with e1000

[Daniel Walker]

On Thu, 2005-09-01 at 09:10 +0200, Ingo Molnar wrote:
> * Daniel Walker <dwalker@mvista.com> wrote:
> 
> > It looks like Gigabit Ethernet is still having some problems. This is 
> > with the e1000 driver. If I remove all the qdisc_restart changes it 
> > starts to work the warning below goes away, but it has 
> > smp_processor_id warnings.
> 
> btw., what does "problems" mean, precisely - does it not work at all, or 
> does it produce the lockup under certain loads?

It hangs when I un'tar a big archive over NFS, but otherwise it works
slowly . The "possible softlockup" must mean it's doing some major
looping in/around qdisc_restart() . 

Daniel 

Re: PREEMPT_RT with e1000

[Daniel Walker]

On Thu, 2005-09-01 at 07:43 -0700, Daniel Walker wrote:
> On Thu, 2005-09-01 at 09:10 +0200, Ingo Molnar wrote:
> > * Daniel Walker <dwalker@mvista.com> wrote:
> > 
> > > It looks like Gigabit Ethernet is still having some problems. This is 
> > > with the e1000 driver. If I remove all the qdisc_restart changes it 
> > > starts to work the warning below goes away, but it has 
> > > smp_processor_id warnings.
> > 
> > btw., what does "problems" mean, precisely - does it not work at all, or 
> > does it produce the lockup under certain loads?
> 
> It hangs when I un'tar a big archive over NFS, but otherwise it works
> slowly . The "possible softlockup" must mean it's doing some major
> looping in/around qdisc_restart() . 

Another note, the possible softlockup warning is disconnected from the
actual lockup . I get several possible softlockup warnings then nothing,
then a hang with no warning. 

Daniel