Re: [PATCH][REGRESSION] panic: fix stack dump print on direct call to panic()

[Jason Wessel <jason.wessel@windriver.com>]

On 04/09/2012 01:34 PM, Andi Kleen wrote:
>> The proper way to resolve the problem that original commit tried to
>> solve is to avoid printing a stack dump from panic() when the either
>> of the following conditions is true:
>>
>>    1) TAINT_DIE has been set (this is done by oops_end())
>>       This indicates and oops has already been printed.
>>    2) oops_in_progress>  1
>>       This guards against the rare case where panic() is invoked
>>       a second time, or in between oops_begin() and oops_end()
> Oops. I guess can just revert it for now. Thanks for catching, Jason.
>
> The proper solution is probably some variant of
> http://git.kernel.org/?p=linux/kernel/git/ak/linux-mce-2.6.git;a=shortlog;h=refs/heads/mce/xpanic
>
> Let the caller pass in the proper action instead of all these hacks.

That is an interesting patch series, but I am not sure I agree about the caller propagating a flag to control what you see in panic.  I intentionally set CONFIG_DEBUG_VERBOSE=y for a reason, I want a stack trace if and when panic(), oops, BUG, etc... is ever called.  This might be a personal preference, but I do not wish to be searching on a string, to find out where the kernel execution terminated.  I want to see stack traces all the time, with the exception of the cases you pointed out in the original patch.

I had taken the time to instrument and the edge cases for the oops and panic() nesting.  I would agree that it is not useful to print further traces if the kernel has already processed an oops.  I also checked non-x86 archs to see that the original behavior was intact with the new version of the patch.  My opinion is that checking for the panic recursion is a necessary evil here, we just have to agree about the right approach.  :-)

Cheers,
Jason.