Hard locks under high interrupt load?

Hard locks under high interrupt load?

[Dan Creswell]

I'm having zero success in getting 2.6.3 stable under interrupt load.  I 
can kill my machine in a variety of fashions all of which appear, to my 
naive eye, related to interrupt load:

(1)   LAN traffic via E1000 card (X is not running)
(2)   Running X for more than a few minutes - starting up a couple of 
applications whilst performing some disk-based activity (such as a 
compile) usually seems to do the trick.

(2) is worth a little more examination.  I have an NVIDIA card (I can 
hear you all groan) *but* I get the same results with the XFree driver 
*or* the proprietary NVIDIA driver.

Disabling IO-APIC usage seems to resolve the problem.

Machine is a dual Xeon, Tyan S2665 (E7505 chipset) with an MPT-Fusion 
SCSI controller.

2.4.26-pre1 and various other 2.4 kernels give me no problems at all.  I 
really want to switch my machines over to 2.6 but I can't whilst this 
problem persists.

Cheers,

Dan.

Re: Hard locks under high interrupt load?

[Zwane Mwaikambo]

On Fri, 27 Feb 2004, Dan Creswell wrote:

> I'm having zero success in getting 2.6.3 stable under interrupt load.  I
> can kill my machine in a variety of fashions all of which appear, to my
> naive eye, related to interrupt load:
>
> (1)   LAN traffic via E1000 card (X is not running)
> (2)   Running X for more than a few minutes - starting up a couple of
> applications whilst performing some disk-based activity (such as a
> compile) usually seems to do the trick.
>
> (2) is worth a little more examination.  I have an NVIDIA card (I can
> hear you all groan) *but* I get the same results with the XFree driver
> *or* the proprietary NVIDIA driver.
>
> Disabling IO-APIC usage seems to resolve the problem.

Does the 'noirqbalance' kernel parameter also serve as a workaround? Are
you using any userspace irq balancers?

> Machine is a dual Xeon, Tyan S2665 (E7505 chipset) with an MPT-Fusion
> SCSI controller.
>
> 2.4.26-pre1 and various other 2.4 kernels give me no problems at all.  I
> really want to switch my machines over to 2.6 but I can't whilst this
> problem persists.

Re: Hard locks under high interrupt load?

[Dan Creswell]

This box is running Fedora Core 1 and so, yes, it's running a userspace 
balance daemon (Redhat's).

I'll try out the noirqbalance option and get back to the list on that one.

Zwane Mwaikambo wrote:

>On Fri, 27 Feb 2004, Dan Creswell wrote:
>
>  
>
>>I'm having zero success in getting 2.6.3 stable under interrupt load.  I
>>can kill my machine in a variety of fashions all of which appear, to my
>>naive eye, related to interrupt load:
>>
>>(1)   LAN traffic via E1000 card (X is not running)
>>(2)   Running X for more than a few minutes - starting up a couple of
>>applications whilst performing some disk-based activity (such as a
>>compile) usually seems to do the trick.
>>
>>(2) is worth a little more examination.  I have an NVIDIA card (I can
>>hear you all groan) *but* I get the same results with the XFree driver
>>*or* the proprietary NVIDIA driver.
>>
>>Disabling IO-APIC usage seems to resolve the problem.
>>    
>>
>
>Does the 'noirqbalance' kernel parameter also serve as a workaround? Are
>you using any userspace irq balancers?
>
>  
>
>>Machine is a dual Xeon, Tyan S2665 (E7505 chipset) with an MPT-Fusion
>>SCSI controller.
>>
>>2.4.26-pre1 and various other 2.4 kernels give me no problems at all.  I
>>really want to switch my machines over to 2.6 but I can't whilst this
>>problem persists.
>>    
>>
>
>  
>

Re: Hard locks under high interrupt load?

[Zwane Mwaikambo]

On Fri, 27 Feb 2004, Dan Creswell wrote:

> This box is running Fedora Core 1 and so, yes, it's running a userspace
> balance daemon (Redhat's).
>
> I'll try out the noirqbalance option and get back to the list on that one.

Could you also give a test run without the userspace balancer daemon and
only the kernel one.

Thanks

Re: Hard locks under high interrupt load?

[Jeff Garzik]

Dan Creswell wrote:
> I'm having zero success in getting 2.6.3 stable under interrupt load.  I 
> can kill my machine in a variety of fashions all of which appear, to my 
> naive eye, related to interrupt load:


There is a lockup problem in 2.6.3's e1000 that's solved in 2.6.3-bk 
snapshot.

	Jeff