[PATCH] kexec: force x86_64 arches to boot kdump kernels on boot cpu

[Neil Horman]

On Fri, Dec 07, 2007 at 09:39:44AM -0500, Vivek Goyal wrote:
> On Thu, Dec 06, 2007 at 07:10:23PM -0500, Neil Horman wrote:
> > On Thu, Dec 06, 2007 at 05:11:43PM -0500, Vivek Goyal wrote:
> > > On Thu, Dec 06, 2007 at 04:39:51PM -0500, Neil Horman wrote:
> > > > On Fri, Nov 30, 2007 at 09:51:31AM -0500, Neil Horman wrote:
> > > > > On Fri, Nov 30, 2007 at 09:42:50AM -0500, Vivek Goyal wrote:
> > > > <snip>
> > > > > 
> > > > > Thats what I'm doing at the moment.  I'm working on a RHEL5 patch at the moment
> > > > > (since thats whats on the production system thats failing), and will forward
> > > > > port it once its working
> > > > > 
> > > > > And not to split hairs, but techically thats not our _only_ choice.  We could
> > > > > force kdump boots on cpu0 as well ;)
> > > > > 
> > > > > Thanks
> > > > > Neil
> > > > > 
> > > > > > Thanks
> > > > > > Vivek
> > > > > 
> > > > 
> > > > 
> > > > Sorry to have been quiet on this issue for a few days. Interesting news to
> > > > report, though.  So I was working on a patch to do early apic enabling on
> > > > x86_64, and had something working for the old 2.6.18 kernel that we were
> > > > origionally testing on.  Unfortunately while it worked on 2.6.18 it failed
> > > > miserably on 2.6.24-rc3-mm2, causing check_timer to consistently report that the
> > > > timer interrupt wasn't getting received (even though we could successfully run
> > > > calibrate_delay).  Vivek and I were digging into this, when I ran accross the
> > > > description of the hypertransport configuration register in the opteron
> > > > specification.  It contains a bit that, suprise, configures the ht bus to either
> > > > unicast interrupts delivered accross the ht bus to a single cpu, or to broadcast
> > > > it to all cpus.  Since it seemed more likely that the 8259 in the nvidia
> > > > southbridge was transporting legacy mode interrupts over the ht bus than
> > > > directly to cpu0 via an actual wire, I wrote the attached patch to add a quirk
> > > > for nvidia chipsets, which scanned for hypertransport controllers, and ensured
> > > > that that broadcast bit was set.  Test results indicate that this solves the
> > > > problem, and kdump kernels boot just fine on the affected system.
> > > > 
> > > 
> > > Hi Neil,
> > > 
> > > Should we disable this broadcasting feature once we are through? Otherwise
> > > in normal systems it might mean extra traffic on hypertransport. There
> > > is no need for every interrupt to be broadcasted in normal systems?
> > > 
> > > Thanks
> > > Vivek
> > 
> > No, I don't think thats necessecary.  Once the apics are enabled, interrupts
> > shouldn't travel accross the hypertransport bus anyway, opting instead to use
> > the dedicated apic bus (at least thats my understanding).
> 
> I think all interrupt message travel on hypertransport. Even after APICS
> have been enabled.
> 
> Look at the following document.
> 
> http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/24674.pdf
> 
> Have a look at figure 1, figure 2 and section 3.4.2.2 and 3.4.2.3
> 
> That's a different thing that once IOAPIC has formed the vectored message,
> Hypertransport might not touch the destination field.
>  
Ok, that might be the case then.

> Having said that, I am wondering what will happen if a system continues
> to operate the timer through IOAPIC in ExtInt mode. Will hypertransport
> keep on broadcasting that interrupt to every cpu? And every cpu will 
> process that interrupt.
> 
I don't think so.  IIRC once the other cpus are started they all disable the
timer interrupt, except for one cpu, opting instead to get the timer tick via
ipi, So while they all might see the interrupt packet on the ht bus, only one
cpu will process it.

> Hence, I feel it is safe to restore the broadcast bit back to BIOS value once
> we are through calibrate_delay().
> 
I disagree.  Looking at what Yinghai said, the default setting for the broadcast
bit isn't actually to unicast the interrupt, its just to set the broadcast mask
to 0xF, or to 0xFF.  Its use is actually to allow cpus with an extended 8 bit
apic id see interrupts.  So its not so much to direct interrupts to cpu0, but
rather to the first 16 cpus rather than to all 255 available cpus.  From what
I've seen in my testing, systems that 'work' already have this bit set by bios,
and my quirk patch above does nothing to them.  Disabling this bit after
calibrate_dealy is going to introduce more uncertainty in systems that have been
proven to work.  We should leave well enough alone, and just enable the bit if
its off, and we see that we are using extended apic ids via bit 18 of the same
register, as Yinghai pointed out.  By enabling the quirk that way, all we are
really doing is bringing into alignment two bits that should arguably be
set/cleared in unison anyway.

Regards
Neil

> Thanks
> Vivek

-- 
/***************************************************
 *Neil Horman
 *Software Engineer
 *Red Hat, Inc.
 *nhorman at redhat.com
 *gpg keyid: 1024D / 0x92A74FA1
 *http://pgp.mit.edu
 ***************************************************/