Re: [Bug] Kdump does not work when panic triggered due to MCE

[Vivek Goyal <vgoyal@redhat.com>]

On Mon, May 09, 2011 at 10:23:57PM +0530, K.Prasad wrote:
> On Mon, May 09, 2011 at 08:39:02AM -0400, Vivek Goyal wrote:
> > On Fri, May 06, 2011 at 10:24:12PM +0530, K.Prasad wrote:
> > > Hi All,
> > > 	I wanted to test the behaviour of kdump when panic is triggered
> > > due to MCE on x86 and found that kdump is not captured.
> > > 
> > > While the kdump service is configured and running and non-MCE panics
> > > (such as those triggered through to /proc/sysrq-trigger) successfully
> > > capture a kdump, any fatal MCE error injected through the mce-inject
> > > tool causes a reboot of the machine.
> > > 
> > > The code has been traced (using early_serial_putc()) to enter the kexec
> > > path i.e. panic()->crash_kexec()->machine_kexec()->relocate_kernel()
> > > but is untraceable further.
> > > 
> > > Kdump works fine when the same the similar test is carried out inside a
> > > KVM guest.
> > > 
> > > Has anybody tested this before? Or have found kdump working when fatal
> > > MCEs have actually occurred?
> > 
> > Prasad,
> > 
> > I have never tried taking dump in MCE situation. Does kdump work on this
> > machine with normal panic()?
> > 
> 
> Hi Vivek,
> 	kdump worked fine on this machine for non-MCE triggered panic
> calls (the /proc/sysrq-trigger initiated crashes got the kdump fine).
> 
> > Use --debug and --serial option in kexec-tools to print some debug message
> > and look for "I am in purgatory". This will tell you whether you hanged
> > in first kernel or second kernel.
> > 
> 
> There were no boot logs from the second kernel while the "Rebooting in X
> seconds..." message had appeared before the system rebooted, suggesting
> that the second kernel did not boot at all.
> 
> > Then put "outb()" messages in the kernel to trace what happened. 
> > 
> 
> The outb logs showed that the system entered machine_kexec function (traceable
> upto relocate_kernel) but then rebooted from inside the panic() function.

Ok, that means that we returned from crash_kexec() function instead of
transitioning into second kernel. This is strange. machine_kexec() is not
supposed to return until and unless it finds that there is no crash
kernel loaded. As per your mail you can trace it to relocate_kernel()
being entered. So only thing I can suggest is debug relocate_kernel()
code now to see why it is returning.

Thanks
Vivek