APIC blues with SMP kernel

[Wolfgang Gill]

On Wed, 19 May 2004 22:28:49 -0300, Andre Costa wrote
> Hi,
> 
> I have a P4 at work, and I have been having bad times trying to make 
> it work with hyper-threading enabled.
> 
> What happens are occasional lockups -- system appears to hang 
> (keyboard, mouse and X freeze, although machine still responds to 
> pings; all other services -- ssh, web, postgres -- are also dead).
> 
> Last time it froze, its last words were:
> 
> May 19 14:11:19 dali kernel: APIC error on CPU1: 00(40)
> 
> May 19 14:11:39 dali kernel: hda: dma_timer_expiry: dma status == 0x24
> 
> I know about APIC problems, but the fact is that if I pass 'noapic'
> to the kernel before boot, the sibling CPU is detected but it doesn't
> seem to be activated: it appears on top and gkrellm but it never does
> anything (i.e. it remains with 0% usage).
> 
> I googled around, and this dma_timer_expiry problem seems to be kernel
> 2.6.x related. This leads me to think this is specific to Fedora 
> kernel, since RH backported many 2.6 features into its 2.4 kernel...
> 
> I am compiling a vanilla kernel as we speak, will try it out during the
> weekend. However, given the fact that Fedora kernel is supposed to
> provide superior performance compared to vanilla kernels, I would 
> like to use it instead, if at all possible...
> 
> So, my questions are:
> 
> 1. any success stories out there about SMP + Fedora on P4s?
> 2. any black-magic to use besides 'noapic', 'acpi=ht' and 'apm=off'?
> 3. do vanilla kernels work with Fedora just fine (w/ or w/o SMP)?
> 
> BTW: FC2 is not an option in the near future, so please don't 
> suggest I upgrade ;)
> 
> TIA
> 
> Andre
> 
> -- 
> Andre Oliveira da Costa
<snip>

My first suggestion would be to update your mother board BIOS. There maybe
there are some bugs in the APIC code in the BIOS. Or change the APIC level
from 1.4 to 1.1.

I don't have any APIC problems with my servers, as I run Tyan MPX mother
boards, and have not seen APIC problems at all. I've run RH9, FC1 & RHEL3 WS
on these boards without any problems.

Wolf
--
Open WebMail Project (http://openwebmail.org)