Re: [PATCH] mce: fix missing stack-dumping in mce_panic()

[Miaohe Lin]

On 2022/12/2 22:44, Luck, Tony wrote:
>> When machine check exception occurs, there is no stack-dumping now in
>> mce_panic(). It's because bust_spinlocks(1) is called prematurely so
>> oops_in_progress will be >= 2 when trying to call dump_stack() in
>> panic(). Thus dump_stack() won't be called as this is considered as
>> nested stack-dumping.
> 

Many thanks for your quick reply. :)

> I had an earlier patch series to just dump from "interesting" machine checks
> (I think the interesting ones are when the kernel hit poison in code that hadn't
> been tagged in the extable as recoverable)
> 
> https://lore.kernel.org/all/20220922195136.54575-1-tony.luck@xxxxxxxxx/

[1]

> 
> Discussion on that fizzled out.
> 
> Thanks for tracking down why panic() didn't provide a stack dump.
> 
> I'm still of the opinion that stack dumps from machine checks aren't
> generally useful. But I'd rather have extra stack dumps than no stack
> dumps at all.

As you mentioned in [1]:
"""
In order to ease the hunt for additional code flows where machine check
errors can be recovered it is useful to know, for example, why the
kernel was copying a page.
"""

So I think it's better to have at least one stack dumps. Also what the commit
6e6f0a1f0fa6 ("panic: don't print redundant backtraces on oops") and commit
026ee1f66aaa ("panic: fix stack dump print on direct call to panic()") want
to do is avoiding nested stack-dumping to have the original oops data being
scrolled away on a 80x50 screen but to have *at least one backtraces*. So
this patch acts more like a BUGFIX to ensure having at least one backtraces
in mce_panic(). What's your thought, Luck?

Thanks,
Miaohe Lin