Stack unwinder for SA

[Yasumasa Suenaga]

Hi David, Chris,

I want to fix `jhsdb jstack --mixed` in JDK-8245234 with third-party library at first.
Fixing hs_err log is an another issue IMHO, but I just want to say we may divert it for hs_err log.
We need further discussion for it of course.


Thanks,

Yasumasa


On 2021/01/25 12:32, Chris Plummer wrote:
> On 1/24/21 7:14 PM, David Holmes wrote:
>> On 25/01/2021 12:41 pm, Yasumasa Suenaga wrote:
>>> Hi all,
>>>
>>> Chris and I discussed about fixing JDK-8245234 to see all mixed stacks via `jhsdb jstack --mixed`.
>>> In the discussion, we agree to use stack unwinder, but we haven't yet decided which library to use.
>>> Can you suggest the stack unwinder for SA?
>>>
>>> I fixed similar issue in JDK-8234624 to understand DWARF in SA for Linux, however it does not support DWARF completely because it is too complex to implement by scratch.
>>> AMD64 SysV ABI [1] says DWARF is used for stack unwinding. So we need to support DWARF completely if we want to unwind.
>>>
>>> There are some stack unwinder. I think libunwind [2] is suitable for this purpose, however I'm not sure it is the best solution.
>>> And also I saw similar issue on Windows because we cannot depend on base pointer (like SA) in Windows x64 [3]. So I want to use the library which is support both Linux and Windows.
>>>
>>> We also think introducing stack unwinder helps improving native call stacks in hs_err log. It will help troubleshooters.
>>
>> Immediate reaction to this part is that I would not want third-party libraries getting involved in hs_err file generation. Given we often do this from a signal handling context, it would be far too risky IMO.
>>
> This is probably much less of an issues for hs_err stack dumps than for SA, since the crashed thread is almost always in JDK code, so we can still get the stack trace for the hs_err file. The issue with SA and JDK-8234624 is when you try to dump all threads (with --mixed). Any that are waiting in a clib function only show the frame they are in. In debug.cpp there are a couple of thread dumping APIs that probably also suffer from the problem described in JDK-8245234, so that's possibly where some hotspot improvements could be made.
> 
> Chris
>> Cheers,
>> David
>> -----
>>
>>>
>>> I want to hear your opinions!
>>>
>>>
>>> Thanks,
>>>
>>> Yasumasa
>>>
>>>
>>> [1] https://software.intel.com/sites/default/files/article/402129/mpx-linux64-abi.pdf
>>> [2] https://github.com/libunwind/libunwind
>>> [3] https://docs.microsoft.com/en-us/cpp/build/exception-handling-x64#unwind-procedure
>