On 03/24/2017 at 01:46 AM, Michael Holzheu wrote:
> Am Thu, 23 Mar 2017 17:23:53 +0800
> schrieb Xunlei Pang <xp...@redhat.com>:
>
>> On 03/23/2017 at 04:48 AM, Michael Holzheu wrote:
>>> Am Wed, 22 Mar 2017 12:30:04 +0800
>>> schrieb Dave Young <dyo...@redhat.com>:
>>>
>>>> On 03/21/17 at 10:18pm, Eric W. Biederman wrote:
>>>>> Dave Young <dyo...@redhat.com> writes:
>>>>>
>>> [snip]
>>>
>>>>>> I think makedumpfile is using it, but I also vote to remove the
>>>>>> CRASHTIME. It is better not to do this while crashing and a makedumpfile
>>>>>> userspace patch is needed to drop the use of it.
>>>>>>
>>>>>>> As we are looking at reliability concerns removing CRASHTIME should make
>>>>>>> everything in vmcoreinfo a boot time constant. Which should simplify
>>>>>>> everything considerably.
>>>>>> It is a nice improvement..
>>>>> We also need to take a close look at what s390 is doing with vmcoreinfo.
>>>>> As apparently it is reading it in a different kind of crashdump process.
>>>> Yes, need careful review from s390 and maybe ppc64 especially about
>>>> patch 2/3, better to have comments from IBM about s390 dump tool and ppc
>>>> fadump. Added more cc.
>>> On s390 we have at least an issue with patch 1/3. For stand-alone dump
>>> and also because we create the ELF header for kdump in the new
>>> kernel we save the pointer to the vmcoreinfo note in the old kernel on a
>>> defined memory address in our absolute zero lowcore.
>>>
>>> This is done in arch/s390/kernel/setup.c:
>>>
>>> static void __init setup_vmcoreinfo(void)
>>> {
>>> mem_assign_absolute(S390_lowcore.vmcore_info,
>>> paddr_vmcoreinfo_note());
>>> }
>>>
>>> Since with patch 1/3 paddr_vmcoreinfo_note() returns NULL at this point in
>>> time we have a problem here.
>>>
>>> To solve this - I think - we could move the initialization to
>>> arch/s390/kernel/machine_kexec.c:
>>>
>>> void arch_crash_save_vmcoreinfo(void)
>>> {
>>> VMCOREINFO_SYMBOL(lowcore_ptr);
>>> VMCOREINFO_SYMBOL(high_memory);
>>> VMCOREINFO_LENGTH(lowcore_ptr, NR_CPUS);
>>> mem_assign_absolute(S390_lowcore.vmcore_info,
>>> paddr_vmcoreinfo_note());
>>> }
>>>
>>> Probably related to this is my observation that patch 3/3 leads to
>>> an empty VMCOREINFO note for kdump on s390. The note is there ...
>>>
>>> # readelf -n /var/crash/127.0.0.1-2017-03-22-21:14:39/vmcore | grep VMCORE
>>> VMCOREINFO 0x0000068e Unknown note type: (0x00000000)
>>>
>>> But it contains only zeros.
>> Yes, this is a good catch, I will do more tests.
> Hello Xunlei,
>
> After spending some time on this, I now understood the problem:
>
> In patch 3/3 you copy vmcoreinfo into the control page before
> machine_kexec_prepare() is called. For s390 we give back all the
> crashkernel memory to the hypervisor before the new crashkernel
> is loaded:
>
> /*
> * Give back memory to hypervisor before new kdump is loaded
> */
> static int machine_kexec_prepare_kdump(void)
> {
> #ifdef CONFIG_CRASH_DUMP
> if (MACHINE_IS_VM)
> diag10_range(PFN_DOWN(crashk_res.start),
> PFN_DOWN(crashk_res.end - crashk_res.start + 1));
> return 0;
> #else
> return -EINVAL;
> #endif
> }
>
> So after machine_kexec_prepare_kdump() the contents of your control page
> is gone and therefore the vmcorinfo ELF note contains only zeros.
>
> If you call kimage_crash_copy_vmcoreinfo() after
> machine_kexec_prepare_kdump() the problem should be solved for s390.
Will update, thanks for finding the root cause.
Regards,
Xunlei
