Question about kmsg_dump for OOPS

Question about kmsg_dump for OOPS

[Jin Dongming]

Hello, Simon

I am Jin Dongming.

I have a question about kmsg_dump which needs your help.
The question is as following:
    Why not put the kmsg_dump() for OOPS into oops_end() and before the branch
    of crash_kexec()?

The reason for the question is as following:
    Now the kmsg_dump() for OOPS is added in oops_exit(). When OOPS happened,
    kernel will call oops_end(). If the crash_kexec() is executed first in
    oops_end(), the oops_exit() could not be called. And also the kmsg_dump()
    for PANIC could not be executed. So I think that the kmsg_dump() for OOPS
    will lose its real meaning.
    
    The function tree for OOPS is as following:
      oops_end()
        |
        |-- if (crash_kexec is valid)
        |      |
        |      |-- crash_kexec()    ==> reboot (and the following function will
        |                                       not be executed)
        |
        |-- oops_exit
        |      |
        |      |-- kmsg_dump(OOPS)
        |
        |-- if (panic is valid)
        |      |
        |      |-- kmsg_dump(PANIC)
        |

    The function tree for PANIC is as following:
      panic()
        |
        |-- kmsg_dump(PANIC)
        |
        |-- crash_kexec()
        |
        |-- notifier()


When kernel paniced, kmsg_dump() for PANIC is executed before crash_kexec(). So
I think before crash_kexec() is executed, kmsg_dump() for OOPS should be called
too. How do you think?

Best regards,
Jin Dongming

Re: Question about kmsg_dump for OOPS

[Simon Kagstrom]

Hi Jin!

On Thu, 03 Dec 2009 12:04:46 +0900
Jin Dongming <jin.dongming@np.css.fujitsu.com> wrote:

> I have a question about kmsg_dump which needs your help.
> The question is as following:
>     Why not put the kmsg_dump() for OOPS into oops_end() and before the branch
>     of crash_kexec()?
> 
> The reason for the question is as following:
>     Now the kmsg_dump() for OOPS is added in oops_exit(). When OOPS happened,
>     kernel will call oops_end(). If the crash_kexec() is executed first in
>     oops_end(), the oops_exit() could not be called. And also the kmsg_dump()
>     for PANIC could not be executed. So I think that the kmsg_dump() for OOPS
>     will lose its real meaning.

It would be OK to move it for my part, I understand your reasoning.
How this is handled seems to vary a bit between architectures though.
ARM has (arch/arm/kernel/die.c)

  NORET_TYPE void die(const char *str, struct pt_regs *regs, int err)
  {
	[...]
	if (panic_on_oops)
		panic("Fatal exception");

	oops_exit();
	[...]

while x86 does (arch/x86/kernel/dumpstack.c):

  void __kprobes oops_end(unsigned long flags, struct pt_regs *regs, int signr)
  {
	if (regs && kexec_should_crash(current))
		crash_kexec(regs);
	[...]
	oops_exit();
	[...]
	if (in_interrupt())
		panic("Fatal exception in interrupt");
	if (panic_on_oops)
		panic("Fatal exception");


There was some additional discussion on this a while ago in these two
threads:

  http://lkml.org/lkml/2009/11/11/404

  http://lkml.org/lkml/2009/10/23/131

where there additionally was a request to move

  atomic_notifier_call_chain(&panic_notifier_list, 0, buf);

before kmsg_dump() and crash_kexec(). I can't immediately see any
problem with this approach, but I'm no expert on kexec. The discussion
didn't really conclude on this matter though.

// Simon

Re: Question about kmsg_dump for OOPS

[Jin Dongming]

Hi Simon

I am sorry for replying late.

Thank you for your answer and your information.
I will read the discussion and consider whether there is a better method
to resolve my problem.

Best Regards,
Jin Dongming

Simon Kagstrom wrote:
> Hi Jin!
> 
> On Thu, 03 Dec 2009 12:04:46 +0900
> Jin Dongming <jin.dongming@np.css.fujitsu.com> wrote:
> 
>> I have a question about kmsg_dump which needs your help.
>> The question is as following:
>>     Why not put the kmsg_dump() for OOPS into oops_end() and before the branch
>>     of crash_kexec()?
>>
>> The reason for the question is as following:
>>     Now the kmsg_dump() for OOPS is added in oops_exit(). When OOPS happened,
>>     kernel will call oops_end(). If the crash_kexec() is executed first in
>>     oops_end(), the oops_exit() could not be called. And also the kmsg_dump()
>>     for PANIC could not be executed. So I think that the kmsg_dump() for OOPS
>>     will lose its real meaning.
> 
> It would be OK to move it for my part, I understand your reasoning.
> How this is handled seems to vary a bit between architectures though.
> ARM has (arch/arm/kernel/die.c)
> 
>   NORET_TYPE void die(const char *str, struct pt_regs *regs, int err)
>   {
> 	[...]
> 	if (panic_on_oops)
> 		panic("Fatal exception");
> 
> 	oops_exit();
> 	[...]
> 
> while x86 does (arch/x86/kernel/dumpstack.c):
> 
>   void __kprobes oops_end(unsigned long flags, struct pt_regs *regs, int signr)
>   {
> 	if (regs && kexec_should_crash(current))
> 		crash_kexec(regs);
> 	[...]
> 	oops_exit();
> 	[...]
> 	if (in_interrupt())
> 		panic("Fatal exception in interrupt");
> 	if (panic_on_oops)
> 		panic("Fatal exception");
> 
> 
> There was some additional discussion on this a while ago in these two
> threads:
> 
>   http://lkml.org/lkml/2009/11/11/404
> 
>   http://lkml.org/lkml/2009/10/23/131
> 
> where there additionally was a request to move
> 
>   atomic_notifier_call_chain(&panic_notifier_list, 0, buf);
> 
> before kmsg_dump() and crash_kexec(). I can't immediately see any
> problem with this approach, but I'm no expert on kexec. The discussion
> didn't really conclude on this matter though.
> 
> // Simon
> 
>