Re: SMP vs. RAIDframe?

[Michael <macallan%netbsd.org@localhost>]

Hello,

On Sep 29, 2008, at 7:28 AM, Holger Weiss wrote:

> * Michael <macallan%netbsd.org@localhost> [2008-09-28 00:19]:
> > On Sep 27, 2008, at 10:40 PM, Holger Weiss wrote:
> > > However, all filesystems (FFS with neither softdep nor WAPBL) are  
> > > on a
> > > RAIDframe mirror, and for some reason, booting GENERIC.MP leads to a
> > > dirty parity.  During the parity rewrite, the system will freeze  
> > > within a
> > > few minutes; that is, I cannot BREAK into DDB.  If I mark one of the
> > > RAIDframe components as failed using raidctl(8), the system seems  
> > > to run
> > > happily until I generate more or less heavy disk I/O.
> > >
> > > So, I'm not sure whether this is RAIDframe related, and I wasn't  
> > > really
> > > able to gather useful information yet, but I'd be interested if  
> > > running
> > > RAIDframe mirros on SMP kernels works for others [or] if such  
> > > problems
> > > are expected.
> >
> > My U60 ( 2x 450MHz ) works just fine with most things on a RAIDframe
> > stripeset ( kernel and loader had to be outside the RAID in order  
> > to get
> > anything booted ).
>
> So it doesn't seem to be a general "RAIDframe + SMP on sparc64" issue.

Yeah, I've been torturing it pretty badly. There are other things  
broken on sparc64 ( not sure if any of them are SMP-related but I  
guess they are ) - the nfs client likes to deadlock ( Martin filed a  
PR a while ago, I can reproduce it but only on sparc64, my macppc and  
sgimips boxes don't do that ) and occasionally I get panics from soft  
interrupt handlers with heavy network load ( I tried to backup the  
machine using tar czf - / | ttcp -t some_host because nfs has been  
unreliable ).
Also, is the hard lockup by any chance an SIR ( the firmware should  
say something like 'Watchdog reset' ) ?

> > Occasionally I get complaints about dirty parity on startup but that
> > doesn't have any negative side effects ( and it's meaningless on a
> > stripeset anyway )
>
> But you didn't get them without SMP?

I never used RAIDframe on any other machine so my experience is rather  
limited. After a dirty reboot I'd expect out-of-sync parity but here I  
occasionally get it after clean reboots as well ( 'clean' as in  
'shutdown -[p|r] now' so shutdown scripts have a chance to do their  
work ). Also, I don't quite understand how exactly my stripeset can  
have dirty parity at all.

have fun
Michael