lkml.org 
[lkml]   [2003]   [Dec]   [7]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: Fixes for nforce2 hard lockup, apic, io-apic, udma133 covered
From
Date
On Mon, 2003-12-08 at 03:07, Ross Dickson wrote:
> On Monday 08 December 2003 05:58, you wrote:
> > > There are three parts to this email.
> > > a) apic mods.
> > > Lockups are due to too fast an apic acknowledge of apic timer int.
> > > Apic hard locked up the system - no nmi debug available.
> > > Fixed it by introducing a delay of at least 500ns into
> > > smp_apic_timer_interrupt() just prior to ack_APIC_irq().
> >
> > I find this really odd... It works just fine...
> > As did disabling whats now active ie:
> > 'Halt Disconnect and Stop Grant Disconnect' bit is enabled.
> >
> > So it seems like these are the two most important factors, at least from
> > where i stand. Both enabled me to actually use my machine with IO-APIC.
> > (1, disabling Halt Disconnect and Stop Grand Disconnect bit or 2, Add a
> > delay on the irq ack.)
> > Anyone that has any clues?
>
> I started work on this about 2 weeks ago and have not yet tried the
> "Halt Disconnect patch and Stop Grand Disconnect bit or 2" patch.
> My lockups ceased with just the apic time delay. I agree the delay is
> wasteful but the code branches into other servicing routines which I
> did not want to try to rearrange as yet. Given the infrequent nature
> of the lockup in CPU cycles maybe we can get smarter and read the
> timer register and see if enough time has expired to safely ack it. Is
> the Apic read cycle fast like cache ram (assume as fast as bus
> cycles?) or slow like 8259?

It could be something like cpu-disconnect-while-apic-being-hammered
race... Hell i dunno, i can only see some patterns from my own
experiences.

> After all it counts bus cycles doesn't it. Alternately perhaps there
> is a status bit in the apic somewhere to check against after we ack to
> ensure that it did its job although we would not want to hammer the
> apic with writes it cannot accept? Or maybe it is not too bad for now,
> I note that there is an existing fixed 400ns delay in the IDE command
> routines ide-iops.c ide_execute_command() which we currently tolerate.

A status bit would be preferred since that would fix any races that
might exist. As for ide it would be preferred to have non-static delays
but if it's not possible to solve in another way then, by all means =).

But in this case, this is a specific workaround, and it works without it
if you disable the AMD cooler thing.

> > > b) io-apic mods
> > > So I have fixed it too (tested on both my epox and albatron MOBOs).
> > > Firstly I found 8254 connected directly to pin 0 not pin 2 of io-apic.
> > > I have modified check_timer() in io_apic.c to trial connect pin and
> > > test for it after the existing test for connection to io-apic.
> >
> > Good job, i wonder if it could be more generalized and integrated with
> > the rest of the code (i haven't even checked the rest of the code, but
> > this seemed separated).
> >
> > One thing though, I get a lot more NMI's now than with nmi_watchdog=2...
> > NMI: 85520
> > LOC: 85477
> >
> > I usually had a 3 figure number by now... but.. =)

> I have not tested the nmi against the b) io-apic mods. We may have a
> vector clash? Perhaps the new apic mpparse.c patch lets the existing
> check_timer() routines work properly?
> I have not yet tried it.

I dunno, i have never used a machine with IO-APIC on so i dunno whats
normal. But it grows a lot faster than nmi_watchdog=2.

> > > c) ide driver mods
> >
> > Cool..
> >
> > I applied all patches and it survived my grep test so i think it works.

6h 46 mins tells me that this works.
(peak nonworking uptime is still below 2hours)

--
Ian Kumlien <pomac () vapor ! com> -- http://pomac.netswarm.net
[unhandled content-type:application/pgp-signature]
\
 
 \ /
  Last update: 2005-03-22 13:59    [W:0.080 / U:0.168 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site