RFR(S): 8247533: SA stack walking sometimes fails with sun.jvm.hotspot.debugger.DebuggerException: get_thread_regs failed for a lwp

Yasumasa Suenaga yasuenag at gmail.com
Thu Jun 25 00:17:59 UTC 2020 

On 2020/06/25 3:22, Chris Plummer wrote:
> On 6/24/20 12:01 AM, Yasumasa Suenaga wrote:
>> On 2020/06/24 15:32, Chris Plummer wrote:
>>> Hi Yasumasa ,
>>>
>>> I think LinuxAMD64CFrame is used for pstack and what I've been looking at has been jstack, and in particular AMD64CurrentFrameGuess, which does use "last java frame". I'm not sure why LinuxAMD64CFrame does not look at "last java frame". Maybe it should.
>>
>> I thought both pattern (jstack, mixed stack) for this change.
>> As you know, mixed jstack (jstack --mixed) attempt to find top of native stack via LinuxAMD64CFrame, register values are needed for it (so it depends on ptrace() call). So I guess mixed mode jstack (jhsdb jstack --mixed) would not show any stacks (cannot find "last java frame").
> Hi Yasumasa,
> 
> I should have been more clear on what I meant by jstack and pstack. For jstack I meant using StackTrace.java, which is what you get by default with "jhsdb jstack" and also the clhsdb jstack command. For pstack I meant PStack.java, which is what you get with "jhsdb jstack --mixed" or the clhsdb pstack command.
> 
> So this CR impacts both types of stack traces in that they will get null registers when the the lower level API fails to get the register set. For StackTrace.java it will then defer to "last java frame" if available. For PStack.java it will not, and will always result in no stack trace. The code of interest is here:
> 
>         AMD64ThreadContext context = (AMD64ThreadContext) thread.getContext();
>         Address pc  = context.getRegisterAsAddress(AMD64ThreadContext.RIP);
>         if (pc == null) return null;
>         return LinuxAMD64CFrame.getTopFrame(dbg, pc, context);
> 
> So the question is should "last java frame" be used if pc == null. If so, then getTopFrame() would also need to be modified to use "last java frame" when fetching RBP.

I don't think so because CFrame is defined as "Models a "C" programming language frame on the stack" in the javadoc, so it should have *valid* register values IMHO.
In addition, RIP is needed for Linux AMD64 at least because it would use DWARF since JDK-8234624.


Thanks,

Yasumasa


> thanks,
> 
> Chris
>>
>>
>> Thanks,
>>
>> Yasumasa
>>
>>
>>> thanks,
>>>
>>> Chris
>>>
>>> On 6/23/20 11:04 PM, Yasumasa Suenaga wrote:
>>>> Hi Chris,
>>>>
>>>> Thanks you for explanation.
>>>> Your change looks good (but "last java frame" would not be found in Linux AMD64 because RSP is NULL - cf. LinuxAMD64CFrame.java)
>>>>
>>>>
>>>> Thanks,
>>>>
>>>> Yasumasa
>>>>
>>>>
>>>> On 2020/06/24 12:09, Chris Plummer wrote:
>>>>> On 6/23/20 6:05 PM, Yasumasa Suenaga wrote:
>>>>>> Hi Chris,
>>>>>>
>>>>>> Skillful troubleshooters who use jhsdb will aware this warnings, and they will take other appropriate methods.
>>>>>>
>>>>>> However, I'm not sure it is worth to continue to perform even if SA cannot get register values.
>>>>>>
>>>>>> For example, Linux AMD64 depends on RIP and RSP values to find top frame.
>>>>>> According to your change, The caller of getThreadIntegerRegisterSet() has responsible for dealing with the set of null registers. However X86ThreadContext::data (it includes raw register values) would still be zero when it happens.
>>>>> This is  what I intended to have happen. Just end up with a register set of all nulls. Then when stack walking code gets a null, it will revert to "last java frame" if available, otherwise no stack dump is done.
>>>>>>
>>>>>> So I think register holder (e.g. X86ThreadContext) should have tri-state (have registers, fail to get registers, not yet attempt to get registers).
>>>>>> OTOH it might be over-engineering. What do you think?
>>>>> Before implementing this I looked at the what would be the easier approach to get the desired effect of stack walking code simply failing over to using "last java frame", and decided the null set of registers was easiest. Other approaches involved more changes and impacted more files.
>>>>>
>>>>> thanks,
>>>>>
>>>>> Chris
>>>>>>
>>>>>>
>>>>>> Thanks,
>>>>>>
>>>>>> Yasumasa
>>>>>>
>>>>>>
>>>>>> On 2020/06/24 3:16, Chris Plummer wrote:
>>>>>>> On 6/20/20 12:53 AM, Yasumasa Suenaga wrote:
>>>>>>>> Hi Chris,
>>>>>>>>
>>>>>>>> On 2020/06/20 15:20, Chris Plummer wrote:
>>>>>>>>> Hi Yasumasa,
>>>>>>>>>
>>>>>>>>> ptrace is not used for core files, so the EFAULT for a bad core file is not a possibility. However, get_lwp_regs() does redirect to core_get_lwp_regs() for core files. It can fail, but the only reason it ever does is if the LWP can't be found in the core (which is never suppose to happen). I would think if this happened due to the core being truncated, SA would be blowing up all over the place with exceptions, probably before we ever get to this code, but in any cast what we do here wouldn't really make a difference.
>>>>>>>>
>>>>>>>> You are right, sorry.
>>>>>>>>
>>>>>>>>
>>>>>>>>> I'm not sure why you prefer an exception for errors other than ESRCH. Why should they be treated differently? getThreadIntegerRegisterSet0() is used for finding the current frame for stack tracing. With my changes any failure will result in deferring to "last java frame" if set, and otherwise just not produce a stack trace (and the WARNING will be present in the output). This seems preferable to completely abandoning any further thread stack tracking.
>>>>>>>>
>>>>>>>> I'm not sure we can trust call stack when ptrace() returns any errors other than ESRCH even if "last java frame" is available. For example, don't ptrace() return EFAULT or EIO when something wrong? (e.g. stack corruption) If so, it may lead to a wrong analysis for troubleshooter.
>>>>>>>> I think it should be abort dumping call stack for its thread at least.
>>>>>>> Hi Yasumasa,
>>>>>>>
>>>>>>> In general stack walking makes a best effort and can be wrong, even when not getting errors like this. For any actively executing thread SA needs to determine where the stack starts, with register contents being the starting point (SP, FP, and PC). These registers could contain anything, and SA makes a best effort to determine a current frame from them. However, the verification steps it takes are not 100% guaranteed, and can lead to an incorrect assumption of the current frame, which in turn can result in an exception later on when walking the stack. See JDK-8247641.
>>>>>>>
>>>>>>> Keep in mind that the WARNING message will always be there. This should be enough to put the troubleshooter on alert that the stack trace may not be accurate. I think it's better to make an attempt at a stack trace then to just abandon it and not attempt to do something that may be useful.
>>>>>>>
>>>>>>> thanks,
>>>>>>>
>>>>>>> Chris
>>>>>>>>
>>>>>>>>
>>>>>>>> Thanks,
>>>>>>>>
>>>>>>>> Yasumasa
>>>>>>>>
>>>>>>>>
>>>>>>>>> thanks,
>>>>>>>>>
>>>>>>>>> Chris
>>>>>>>>>
>>>>>>>>> On 6/19/20 6:33 PM, Yasumasa Suenaga wrote:
>>>>>>>>>> Hi Chris,
>>>>>>>>>>
>>>>>>>>>> I checked Linux kernel code at a glance, ESRCH seems to be set to errno by default.
>>>>>>>>>> So I guess it is similar to "generic" error code.
>>>>>>>>>>
>>>>>>>>>> https://github.com/torvalds/linux/blob/master/kernel/ptrace.c
>>>>>>>>>>
>>>>>>>>>> According to manpage of ptrace(2), it might return errno other than ESRCH.
>>>>>>>>>> For example, if we analyze broken core (e.g. the core was dumped with disk full), we might get EFAULT.
>>>>>>>>>> Thus I prefer to handle ESRCH only in your patch, and also I think SA should throw DebuggerException if other error is occurred.
>>>>>>>>>>
>>>>>>>>>> https://www.man7.org/linux/man-pages/man2/ptrace.2.html
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> Thanks,
>>>>>>>>>>
>>>>>>>>>> Yasumasa
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> On 2020/06/20 5:51, Chris Plummer wrote:
>>>>>>>>>>> Hello,
>>>>>>>>>>>
>>>>>>>>>>> I've  updated with webrev based on the new finding that a JavaThread cannot be on the ThreadList after its OS thread has been destroyed since the JavaThread removes itself from the ThreadList, and therefore must be running on its OS thread. The logic of the fix is unchanged from the first webrev, but I updated the comments to better reflect what is going on. I also updated the CR:
>>>>>>>>>>>
>>>>>>>>>>> https://bugs.openjdk.java.net/browse/JDK-8247533
>>>>>>>>>>> http://cr.openjdk.java.net/~cjplummer/8247533/webrev.01/index.html
>>>>>>>>>>>
>>>>>>>>>>> thanks,
>>>>>>>>>>>
>>>>>>>>>>> Chris
>>>>>>>>>>>
>>>>>>>>>>> On 6/19/20 12:24 AM, David Holmes wrote:
>>>>>>>>>>>> Hi Chris,
>>>>>>>>>>>>
>>>>>>>>>>>> On 19/06/2020 8:55 am, Chris Plummer wrote:
>>>>>>>>>>>>> On 6/18/20 1:43 AM, David Holmes wrote:
>>>>>>>>>>>>>> On 18/06/2020 4:49 pm, Chris Plummer wrote:
>>>>>>>>>>>>>>> On 6/17/20 10:29 PM, David Holmes wrote:
>>>>>>>>>>>>>>>> On 18/06/2020 3:13 pm, Chris Plummer wrote:
>>>>>>>>>>>>>>>>> On 6/17/20 10:09 PM, David Holmes wrote:
>>>>>>>>>>>>>>>>>> On 18/06/2020 2:33 pm, Chris Plummer wrote:
>>>>>>>>>>>>>>>>>>> On 6/17/20 7:43 PM, David Holmes wrote:
>>>>>>>>>>>>>>>>>>>> Hi Chris,
>>>>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>>>>> On 18/06/2020 6:34 am, Chris Plummer wrote:
>>>>>>>>>>>>>>>>>>>>> Hello,
>>>>>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>>>>>> Please help review the following:
>>>>>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>>>>>> https://bugs.openjdk.java.net/browse/JDK-8247533
>>>>>>>>>>>>>>>>>>>>> http://cr.openjdk.java.net/~cjplummer/8247533/webrev.00/index.html
>>>>>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>>>>>> The CR contains all the needed details. Here's a summary of changes in each file:
>>>>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>>>>> The problem sounds to me like a variation of the more general problem of not ensuring a thread is kept alive whilst acting upon it. I don't know how the SA finds these references to the threads it is going to stackwalk, but is it possible to fix this via appropriate uses of ThreadsListHandle/Iterator?
>>>>>>>>>>>>>>>>>>> It fetches ThreadsSMRSupport::_java_thread_list.
>>>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>>>> Keep in mind that once SA attaches, nothing in the VM changes. For example, SA can't create a wrapper to a JavaThread, only to have the JavaThread be freed later on. It's just not possible.
>>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>>> Then how does it obtain a reference to a JavaThread for which the native OS thread id is invalid? Any thread found in _java_thread_list is either live or still to be started. In the latter case the JavaThread->osThread does not have its thread_id set yet.
>>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>> My assumption was that the JavaThread is in the process of being destroyed, and it has freed its OS thread but is itself still in the thread list. I did notice that the OS thread id being used looked to be in the range of thread id #'s you would expect for the running app, so that to me indicated it was once valid, but is no more.
>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>> Keep in mind that although hotspot may have synchronization code that prevents you from pulling a JavaThread off the thread list when it is in the process of being destroyed (I'm guessing it does), SA has no such protections.
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>> But you stated that once the SA has attached, the target VM can't change. If the SA gets its set of thread from one attach then tries to make queries about those threads in a separate attach, then obviously it could be providing garbage thread information. So you would need to re-validate the JavaThread in the target VM before trying to do anything with it.
>>>>>>>>>>>>>>> That's not what is going on here. It's attaching and doing a stack trace, which involves getting the thread list and iterating through all threads without detaching. 
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> Okay so I restate my original comment - all the JavaThreads must be alive or not yet started, so how are you encountering an invalid thread id? Any thread you find via the ThreadsList can't have destroyed its osThread. In any case the logic should be checking thread->osThread() for NULL, and then osThread()->get_state() to ensure it is >= INITIALIZED before using the thread_id().
>>>>>>>>>>>>> Hi David,
>>>>>>>>>>>>>
>>>>>>>>>>>>> I chatted with Dan about this, and he said since the JavaThread is responsible for removing itself from the ThreadList, it is impossible to have a JavaThread still on the ThreadList, but without and underlying OS Thread. So I'm a bit perplexed as to how I can find a JavaThread on the ThreadList, but that results in ESRCH when trying to access the thread with ptrace. My only conclusion is that this failure is somehow spurious, and maybe the issue it just that the thread is in some temporary state that prevents its access. If so, I still think the approach I'm taking is the correct one, but the comments should be updated.
>>>>>>>>>>>>
>>>>>>>>>>>> ESRCH can have other meanings but I don't know enough about the broader context to know whether they are applicable in this case.
>>>>>>>>>>>>
>>>>>>>>>>>>     ESRCH  The  specified  process  does not exist, or is not currently being traced by the caller, or is not stopped
>>>>>>>>>>>>               (for requests that require a stopped tracee).
>>>>>>>>>>>>
>>>>>>>>>>>> I won't comment further on the fix/workaround as I don't know the code. I'll leave that to other folk.
>>>>>>>>>>>>
>>>>>>>>>>>> Cheers,
>>>>>>>>>>>> David
>>>>>>>>>>>> -----
>>>>>>>>>>>>
>>>>>>>>>>>>> I had one other finding. When this issue first turned up, it prevented the thread from getting a stack trace due to the exception being thrown. What I hadn't realize is that after fixing it to not throw an exception, which resulted in the stack walking code getting all nulls for register values, I actually started to see a stack trace printed:
>>>>>>>>>>>>>
>>>>>>>>>>>>> "JLine terminal non blocking reader thread" #26 daemon prio=5 tid=0x00007f12f0cd6420 nid=0x1f99 runnable [0x00007f125f0f4000]
>>>>>>>>>>>>>     java.lang.Thread.State: RUNNABLE
>>>>>>>>>>>>>     JavaThread state: _thread_in_native
>>>>>>>>>>>>> WARNING: getThreadIntegerRegisterSet0: get_lwp_regs failed for lwp (8089)
>>>>>>>>>>>>> CurrentFrameGuess: choosing last Java frame: sp = 0x00007f125f0f4770, fp = 0x00007f125f0f47c0
>>>>>>>>>>>>>   - java.io.FileInputStream.read0() @bci=0 (Interpreted frame)
>>>>>>>>>>>>>   - java.io.FileInputStream.read() @bci=1, line=223 (Interpreted frame)
>>>>>>>>>>>>>   - jdk.internal.org.jline.utils.NonBlockingInputStreamImpl.run() @bci=108, line=216 (Interpreted frame)
>>>>>>>>>>>>>   - jdk.internal.org.jline.utils.NonBlockingInputStreamImpl$$Lambda$536+0x0000000800daeca0.run() @bci=4 (Interpreted frame)
>>>>>>>>>>>>>   - java.lang.Thread.run() @bci=11, line=832 (Interpreted frame)
>>>>>>>>>>>>>
>>>>>>>>>>>>> The "CurrentFrameGuess" output is some debug tracing I had enabled, and it indicates that the stack walking code is using the "last java frame" setting, which it will do if current registers values don't indicate a valid frame (as would be the case if sp was null). I had previously assumed that without an underling valid LWP, there would be no stack trace. Given that there is one, there must be a valid LWP. Otherwise I don't see how the stack could have been walked. That's another indication that the ptrace failure is spurious in nature.
>>>>>>>>>>>>>
>>>>>>>>>>>>> thanks,
>>>>>>>>>>>>>
>>>>>>>>>>>>> Chris
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> Cheers,
>>>>>>>>>>>>>> David
>>>>>>>>>>>>>> -----
>>>>>>>>>>>>>>
>>>>>>>>>>>>>>> Also, even if you are using something like clhsdb to issue commands on addresses, if the address is no longer valid for the command you are executing, then you would get the appropriate error when there is an attempt to create a wrapper for it. I don't know of any command that operates directly on a JavaThread, but I think there are for InstanceKlass. So if you remembered the address of an InstanceKlass, and then reattached and tried a command that takes an InstanceKlass address, you would get an exception when SA tries to create the wrapper for the InsanceKlass if it were no longer a valid address for one.
>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>> Chris
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>> David
>>>>>>>>>>>>>>>> -----
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>> Chris
>>>>>>>>>>>>>>>>>> David
>>>>>>>>>>>>>>>>>> -----
>>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>>>> Chris
>>>>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>>>>> Cheers,
>>>>>>>>>>>>>>>>>>>> David
>>>>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>>>>>> src/jdk.hotspot.agent/linux/native/libsaproc/LinuxDebuggerLocal.cpp
>>>>>>>>>>>>>>>>>>>>> src/jdk.hotspot.agent/macosx/native/libsaproc/MacosxDebuggerLocal.m
>>>>>>>>>>>>>>>>>>>>> src/jdk.hotspot.agent/windows/native/libsaproc/sawindbg.cpp
>>>>>>>>>>>>>>>>>>>>> -Instead of throwing an exception when the OS ThreadID is invalid, print a warning.
>>>>>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>>>>>> src/jdk.hotspot.agent/linux/native/libsaproc/ps_proc.c
>>>>>>>>>>>>>>>>>>>>> -Improve a print_debug message
>>>>>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>>>>>> src/jdk.hotspot.agent/share/classes/sun/jvm/hotspot/debugger/bsd/BsdThread.java
>>>>>>>>>>>>>>>>>>>>> src/jdk.hotspot.agent/share/classes/sun/jvm/hotspot/debugger/linux/LinuxThread.java
>>>>>>>>>>>>>>>>>>>>> src/jdk.hotspot.agent/share/classes/sun/jvm/hotspot/debugger/windbg/amd64/WindbgAMD64Thread.java
>>>>>>>>>>>>>>>>>>>>> -Deal with the array of registers read in being null due to the OS ThreadID not being valid.
>>>>>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>>>>>> src/jdk.hotspot.agent/share/classes/sun/jvm/hotspot/debugger/bsd/BsdDebuggerLocal.java
>>>>>>>>>>>>>>>>>>>>> src/jdk.hotspot.agent/share/classes/sun/jvm/hotspot/debugger/linux/LinuxDebuggerLocal.java
>>>>>>>>>>>>>>>>>>>>> -Fix issue with "sun.jvm.hotspot.debugger.DebuggerException" appearing twice when printing the exception.
>>>>>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>>>>>> thanks,
>>>>>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>>>>>> Chris
>>>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>
>>>>>>>
>>>>>
>>>>>
>>>
>>>
> 
>

More information about the serviceability-dev mailing list