Review request: 7003789 PTRACE_GETREGS problems with SA on Linux.

Yasumasa Suenaga suenaga.yasumasa at oss.ntt.co.jp
Sun Dec 5 16:14:46 PST 2010 

Hi Kevin,

+1

I agree to this modification.
(Should I do anything? vote?)

(2010/12/03 19:48), Kevin Walls wrote:
>> If PTRACE_GETREGS is undefined and PT_GETREGS is defined, PTRACE_GETREGS_REQ will be defined
>> by undefined value.
> 
> Yes, that was worrying me as well, as it's what I see on my fedora 13 x86_x64 system.
> 
> And yes to your second message - we need to leave in the sparcv9 parts.
> 
> I'm convinced PTRACE_GETREGS64 only exists on 64-bit sparcv9, and PowerPC were we don't build. If it became defined on a platform that is not 64-bit we will only use it if we are doing a 64-bit build, so that looks good.  Further searching finds this thread...
> 
> So here it is as a webrev. Builds and tests OK on x86 and x64.
> 
> http://cr.openjdk.java.net/~kevinw/7003789/webrev.01/
> 
> Regards
> Kevin
> 
> 
> 
> On 03/12/2010 06:47, Yasumasa Suenaga wrote:
>> Sorry...
>>
>>> BTW, this modification is for "agent/src/os/linux/ps_proc.c".
>>> I guess that this source affects for Linux ONLY.
>>> If my guess is correct, we can delete the statement about SPARC from this source code.
>>>
>>> -------
>>> #if defined(sparc) || defined(sparcv9)
>>> #define ptrace_getregs(request, pid, addr, data) ptrace(request, pid, addr, data)
>>> #else
>>> #define ptrace_getregs(request, pid, addr, data) ptrace(request, pid, data, addr)
>>> #endif
>>> -------
>>>
>>> and
>>>
>>> -------
>>> #ifdef PTRACE_GETREGS64
>>> #define PTRACE_GETREGS_REQ PTRACE_GETREGS64
>>> -------
>> This sentence concerns CPU... not OS.
>> I don't have Linux on SPARC. So, I can't confirm that Linux (glibc) ptrace() systemcall on SPARC
>> can handle PTRACE_GETREGS64.
>> If ptrace() interface is different from x86 and SPARC (Linux on SPARC defines PTRACE_GETREGS64),
>> my guess is incorrect...
>>
>> (2010/12/03 15:21), Yasumasa Suenaga wrote:
>>> Hi David, Kevin,
>>>
>>>>     #if defined(PTRACE_GETREGS) || defined(PT_GETREGS)
>>>>       #define PTRACE_GETREGS_REQ PTRACE_GETREGS
>>>>     #endif
>>> If PTRACE_GETREGS is undefined and PT_GETREGS is defined, PTRACE_GETREGS_REQ will be defined
>>> by undefined value. (PTRACE_GETREGS: 1? decided by preprocessor... does that case exist?)
>>>
>>> In Linux manpage PTRACE(2) in Fedora14, PTRACE_GETREGS is described, PTRACE_GETREGS64 and
>>> PT_GETREGS is NOT described.
>>>> My system(Fedora14 x86_64) defines PTRACE_GETREGS and PT_GETREGS, PTRACE_GETREGS64 does NOT define.
>>>
>>> So, how about this?
>>>
>>> -------
>>> #if defined(_LP64)&&   defined(PTRACE_GETREGS64) /* for SPARCV9(and others?) */
>>> #define PTRACE_GETREGS_REQ PTRACE_GETREGS64
>>> #elif defined(PTRACE_GETREGS) /* for Linux */
>>> #define PTRACE_GETREGS_REQ PTRACE_GETREGS
>>> #elif defined(PT_GETREGS) /* last resort */
>>> #define PTRACE_GETREGS_REQ PT_GETREGS
>>> #endif /* _LP64 */
>>> -------
>>>
>>> BTW, this modification is for "agent/src/os/linux/ps_proc.c".
>>> I guess that this source affects for Linux ONLY.
>>> If my guess is correct, we can delete the statement about SPARC from this source code.
>>>
>>> -------
>>> #if defined(sparc) || defined(sparcv9)
>>> #define ptrace_getregs(request, pid, addr, data) ptrace(request, pid, addr, data)
>>> #else
>>> #define ptrace_getregs(request, pid, addr, data) ptrace(request, pid, data, addr)
>>> #endif
>>> -------
>>>
>>> and
>>>
>>> -------
>>> #ifdef PTRACE_GETREGS64
>>> #define PTRACE_GETREGS_REQ PTRACE_GETREGS64
>>> -------
>>>
>>>
>>> Thanks.
>>> Yasumasa
>>>
>>>
>>> (2010/12/03 12:23), David Holmes wrote:
>>>> Kevin Walls said the following on 12/03/10 06:03:
>>>>> Thanks both for pointing out that I didn't notice I was smashing the
>>>>> 32-bit case. 8-)
>>>>>
>>>>> I can now build and test 32 and 64-bit OK with...
>>>>>
>>>>> #ifdef _LP64
>>>>>      #ifdef PTRACE_GETREGS64
>>>>>        #define PTRACE_GETREGS_REQ PTRACE_GETREGS64
>>>>>      #elif defined(PTRACE_GETREGS)
>>>>>        #define PTRACE_GETREGS_REQ PTRACE_GETREGS
>>>>>      #elif defined(PT_GETREGS)
>>>>>        #define PTRACE_GETREGS_REQ PT_GETREGS
>>>>>      #endif
>>>>> #else
>>>>>      #if defined(PTRACE_GETREGS) || defined(PT_GETREGS)
>>>>>        #define PTRACE_GETREGS_REQ PTRACE_GETREGS
>>>>>      #endif
>>>>> #endif /* _LP64 */
>>>>>
>>>>> My 64-bit system is Fedora and really needs to use PT_REGS as that's the
>>>>> only definition. My 32-bit system is a different RedHat and has
>>>>> PTRACE_GETREGS defined so the 32-bit build does work. Am tempted to
>>>>> change the 32-bit side of the ifdef to also use PT_GETREGS as a last
>>>>> resort...
>>>> I'd be tempted to do that too. This seems so inconsistent I wouldn't
>>>> trust what you might find on any particular linux.
>>>>
>>>> If we knew that we'd never find PTRACE_GETREGS64 on a 32-bit system we
>>>> could make it somewhat simpler.
>>>>
>>>> David
>>>>
>>>>> Thanks
>>>>> Kevin
>>>>>
>>>>>
>>>>> On 02/12/2010 02:23, Yasumasa Suenaga wrote:
>>>>>> Thank you for the reply.
>>>>>>
>>>>>> I've read your patch on a webrev.
>>>>>> Your patch will work on AMD64 architecture.
>>>>>> However, this patch will not work on x86 (32bit) architecture.
>>>>>> Preprocessor in GCC for i386 does not define "_LP64" .
>>>>>>
>>>>>> -----------------
>>>>>> [root at RHEL4-4 test]# cat /etc/redhat-release
>>>>>> Red Hat Enterprise Linux ES release 4 (Nahant Update 4)
>>>>>> [root at RHEL4-4 test]# uname -a
>>>>>> Linux RHEL4-4 2.6.9-42.EL #1 Wed Jul 12 23:16:43 EDT 2006 i686 i686 i386 GNU/Linux
>>>>>> [root at RHEL4-4 test]# rpm -q glibc
>>>>>> glibc-2.3.4-2.25
>>>>>> [root at RHEL4-4 test]# rpm -q gcc
>>>>>> gcc-3.4.6-3
>>>>>> [root at RHEL4-4 test]# touch test.h
>>>>>> [root at RHEL4-4 test]# cpp -dM test.h | grep _LP64
>>>>>> [root at RHEL4-4 test]#
>>>>>> -----------------
>>>>>>
>>>>>>
>>>>>> So, you need to modify the patch as follows:
>>>>>>
>>>>>> -----------------
>>>>>> [root at Fedora13 OpenJDK7]# diff -u b118/openjdk/hotspot/agent/src/os/linux/ps_proc.c trunk/
>>>>>> openjdk/hotspot/agent/src/os/linux/ps_proc.c
>>>>>> --- a/agent/src/os/linux/ps_proc.c   2010-11-12 05:43:12.000000000 +0900
>>>>>> +++ b/agent/src/os/linux/ps_proc.c   2010-12-02 10:45:09.117050388 +0900
>>>>>> @@ -124,6 +124,8 @@
>>>>>>     #ifdef _LP64
>>>>>>     #ifdef PTRACE_GETREGS64
>>>>>>     #define PTRACE_GETREGS_REQ PTRACE_GETREGS64
>>>>>> +#elif defined(PTRACE_GETREGS)
>>>>>> +#define PTRACE_GETREGS_REQ PTRACE_GETREGS
>>>>>>     #endif
>>>>>>     #else
>>>>>>     #if defined(PTRACE_GETREGS) || defined(PT_GETREGS)
>>>>>> -----------------
>>>>>>
>>>>>>
>>>>>> Thanks.
>>>>>>
>>>>>>
>>>>>> (2010/12/02 1:41), Kevin Walls wrote:
>>>>>>> Sorry, there was actually a typo in that diff, although it was good
>>>>>>> enough to be a solution on my system.
>>>>>>>
>>>>>>> A webrev also:
>>>>>>>
>>>>>>> http://cr.openjdk.java.net/~kevinw/7003789/webrev.00/
>>>>>>>
>>>>>>> Thanks
>>>>>>> Kevin
>>>>>>>
>>>>>>>
>>>>>>> On 01/12/2010 16:01, Kevin Walls wrote:
>>>>>>>> Hi --
>>>>>>>>
>>>>>>>> I've just been hitting that as well.... I think it needs to be dealt
>>>>>>>> with as a separate bug as there may be a few which are similar - and
>>>>>>>> 6359295 is marked fixed in 5.0 long ago...
>>>>>>>>
>>>>>>>> I just found your message after I'd done some investigating... The patch
>>>>>>>> I was working with is to split one of the existing "if defined"
>>>>>>>> statements, as currently the bug is: if we have PT_GETREGS defined, we
>>>>>>>> use PTRACE_GETREGS to define PTRACE_GETREGS_REQ... We soon realise that
>>>>>>>> may NOT be defined and give the "unsupported" message. So I can log and
>>>>>>>> bug and get this done if it sounds good:
>>>>>>>>
>>>>>>>> [kevin at oldbox make]$ hg diff  ../src/os/linux/ps_proc.c
>>>>>>>> diff --git a/agent/src/os/linux/ps_proc.c b/agent/src/os/linux/ps_proc.c
>>>>>>>> --- a/agent/src/os/linux/ps_proc.c
>>>>>>>> +++ b/agent/src/os/linux/ps_proc.c
>>>>>>>> @@ -124,9 +124,9 @@
>>>>>>>>      #ifdef _LP64
>>>>>>>>      #ifdef PTRACE_GETREGS64
>>>>>>>>      #define PTRACE_GETREGS_REQ PTRACE_GETREGS64
>>>>>>>> -#endif
>>>>>>>> -#else
>>>>>>>> -#if defined(PTRACE_GETREGS) || defined(PT_GETREGS)
>>>>>>>> +#elif defined(PTRACE_GETREGS)
>>>>>>>> +#define PTRACE_GETREGS_REQ PTRACE_GETREGS
>>>>>>>> +#elif defined (PT_GETREGS)
>>>>>>>>      #define PTRACE_GETREGS_REQ PTRACE_GETREGS
>>>>>>>>      #endif
>>>>>>>>      #endif /* _LP64 */
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> Regards
>>>>>>>> Kevin
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> On 15/11/2010 06:46, Yasumasa Suenaga wrote:
>>>>>>>>> Hi.
>>>>>>>>>
>>>>>>>>> I and co-worker use jstack for various trouble shooting.
>>>>>>>>>
>>>>>>>>> We mainly use Java on Linux with AMD64 architecture.
>>>>>>>>> However, jstack -F option doesn't work our platform.
>>>>>>>>>
>>>>>>>>> I ran jstack -F with LIBSAPROC_DEBUG=1 (environment variable),
>>>>>>>>> I got following messages:
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> /***************/
>>>>>>>>> Thread 31933: (state = BLOCKED)
>>>>>>>>> libsaproc DEBUG: ptrace(PTRACE_GETREGS, ...) not supported
>>>>>>>>> Error occurred during stack walking:
>>>>>>>>> sun.jvm.hotspot.debugger.DebuggerException: sun.jvm.hotspot.debugger.DebuggerException: get_thread_regs failed for a lwp
>>>>>>>>> /***************/
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> In order to fix this problem, I made a patch for preprocessor
>>>>>>>>> macro in "agent/src/os/linux/ps_proc.c" .
>>>>>>>>> The patch that attached this mail works well on Fedora 13 x86_64.
>>>>>>>>>
>>>>>>>>> Please merge this patch if you don't fix this problem yet.
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> Best regards.
>>>>>>>>>
>>>>>>>>>
>>>
>>
> 


-- 
日本電信電話株式会社 研究企画部門 OSS センタ
応用技術ユニット　Webグループ 

末永　恭正（すえなが　やすまさ）

TEL: 03-5860-5105 (直通 5069)
E-mail: suenaga.yasumasa at oss.ntt.co.jp

More information about the serviceability-dev mailing list