RFR (trivial): 8250826: jhsdb does not work with coredump which comes from Substrate VM

[Yasumasa Suenaga]

Hi Chris,

On 2020/07/31 7:29, Chris Plummer wrote:
> Hi Yasumasa,
> 
> If I understand correctly we first call add_map_info() for all the PT_LOAD segments in the core file. We then process all the library segments, calling add_map_info() for them if the target_vaddr has not already been addded. If has already been added, which I assume is the case for any library segment that is already in the core file, then the core file version is replaced the the library version.  I'm a little unclear of the purpose of this replacing of the core PT_LOAD segments with those found in the libraries. If you could explain this that would help me understand your change.

Read only segments in ELF should not be any different from PT_LOAD segments in the core.
And head of ELF header might be included in coredump (See JDK-7133122). Thus we need to replace PT_LOAD segments the library version.

> I'm also unsure why existing_map->fd would ever be something other than the core file. Why would another library map the same target_vaddr.

When mmap() is called to read-only ELF segments / sections, Linux kernel seems to allocate other memory segments which has same top virtual memory address. I've not yet found out from the code of Linux kernel, but I confirmed this behavior on GDB.


Thanks,

Yasumasa


> thanks,
> 
> Chris
> 
> On 7/30/20 1:18 PM, Chris Plummer wrote:
>> Hi Yasumasa,
>>
>> I'm reviewing this RFR, and I'd like to ask that it not be pushed as trivial. Although it is just a one line change, it takes an extensive knowledge to understand the impact. I'll read up on the filed graal issue and try to understand the ELF code a bit better.
>>
>> thanks,
>>
>> Chris
>>
>> On 7/30/20 6:45 AM, Yasumasa Suenaga wrote:
>>> Hi all,
>>>
>>> Please review this trivial change:
>>>
>>>   JBS: https://bugs.openjdk.java.net/browse/JDK-8250826
>>>   webrev: http://cr.openjdk.java.net/~ysuenaga/JDK-8250826/webrev.00/
>>>
>>> I played Truffle NFI on GraalVM, but I cannot get Java stacks from coredump via jhsdb.
>>>
>>> I've reported this issue to GraalVM community [1], and I 've found out the cause of this issue is .svm_heap would be separated to RO and RW areas by mprotect() calls in run time in spite of .svm_heap is RO section in ELF (please see [1] for details).
>>>
>>> It is corner case, but we will see same problem on jhsdb when we attempt to analyze coredump which comes from some applications / libraries which would separate RO sections in ELF like Substrate VM.
>>>
>>> I sent PR to fix libsaproc.so in LabsJDK 11 for this issue [2], then community members suggested me to discuss in serviceability-dev.
>>>
>>>
>>> Thanks,
>>>
>>> Yasumasa
>>>
>>>
>>> [1] https://github.com/oracle/graal/issues/2579
>>> [2] https://github.com/graalvm/labs-openjdk-11/pull/9
>>
>>
> 
>