RFR: 8032901: WaitForMultipleObjects() return value not handled appropriately (original) (raw)

Aleksej Efimov aleksej.efimov at oracle.com
Wed Jun 4 11:43:05 UTC 2014

Thank you David. Thank you Staffan.

On 06/04/2014 01:46 PM, David Holmes wrote:

Me too.

Thanks, David On 4/06/2014 5:23 PM, Staffan Larsen wrote: Looks ok to me.

/Staffan On 3 jun 2014, at 15:49, Aleksej Efimov <aleksej.efimov at oracle.com> wrote:

Staffan, David,

Returning back to our WaitForMultipleObjects()/WAITABANDONED discussions: Thank you for your comments and remarks. I can't disagree with motivation that it's better to have a fatal error during the incorrect mutex handling then data corruption (the consequence of the previous fix). In case of such error it'll be much more easier to debug/catch it (but as Staffan said - we have tried to check all call paths and don't think that we'll encounter such behavior). I have modified the discussed code according to your suggestions: http://cr.openjdk.java.net/~aefimov/8032901/9/webrev.01/ To abort the process the 'exitTransportWithError' function was utilized. Also I have tried to check that behavior isn't changed by running "svc" regression tests set. There was no related test failures observed. Best Regards, Aleksej On 05/15/2014 01:11 PM, Staffan Larsen wrote: On 15 maj 2014, at 03:48, David Holmes <david.holmes at oracle.com> wrote:

On 14/05/2014 11:18 PM, Aleksej Efimov wrote: David, Vitaly,

I totally agree with Vitaly's explanation (Vitaly - thank you for that) and code in shmemBase.c (the usage of enterMutex() function for sending/receiving bytes through shared memory connection) illustrates on how the connection shutdown event is used as a "cancellation token". Thanks for clarifying that. So if we were to encounter an abandoned mutex the code would presently have acquired the mutex but return an error, thus preventing a subsequent release, and preventing other threads from acquiring (but allowing current thread to recurisevely acquire. So this could both hang and cause data corruption. The new code will still return an error but release the mutex. So no more hangs (other than by conditions caused by data corruption) but more opportunity for data corruption. Obviously it depends on exactly what happens in the critical sections guarded by this mutex, but in general I don't agree with this rationale for making the change: 204 /* If the mutex is abandoned we want to return an error 205 * and also release the mutex so that we don't cause 206 * other threads to be blocked. If a mutex was abandoned 207 * we are in scary state. Data may be corrupted or inconsistent 208 * but it is still better to let other threads run (and possibly 209 * crash) than having them blocked (on the premise that a crash 210 * is always easier to debug than a hang). Considering something has to have gone drastically wrong for the mutex to become abandoned, I'm more inclined to consider this a fatal error and abort. But I'll let the serviceability folk chime in here. I was involved in fixing this and writing the comment, so obviously I thought it a good solution :-) I do agree that it would probably be a good idea to consider this a fatal error and abort. At that point in the code we don’t have any really nice ways of doing that, though. We could just print an error and call abort(). What we are doing now is returning an error from sysIPMutexEnter() and delegating the error handling to the caller. We have tried to check all call paths to verify that they do “the right thing” in the face of an error. It is obviously hard to verify, but it looks like they all terminate the connection with some kind of error message. /Staffan

Thanks, David

Thank you, -Aleksej

On 05/14/2014 01:05 PM, David Holmes wrote: On 14/05/2014 11:06 AM, Vitaly Davidovich wrote: In windows, you acquire a mutex by waiting on it using one of the wait functions, one of them employed in the code in question. If WaitForMultipleObjects succeeds and returns the index of the mutex, current thread has ownership now. Yes I understand the basic mechanics :) It's also common to use multi wait functions where the event is a "cancelation token", e.g. manual reset event; this allows someone to cancel waiting on mutex acquisition and return from the wait function. Presumably that's the case here, but I'll let Aleksej confirm; just wanted to throw this out there in the meantime :). Ah I see - yes cancellable lock acquisition would make sense. Thanks, David Sent from my phone On May 13, 2014 6:46 PM, "David Holmes" <david.holmes at oracle.com_ _<mailto:david.holmes at oracle.com>> wrote: Hi Aleksej, Thanks for the doc references regarding abandonment. Let me rephrase my question. What is this logic trying to achieve by waiting on both a mutex and an event? Do we already own the mutex when this function is called? David On 13/05/2014 11:19 PM, Aleksej Efimov wrote: David, The Windows has a different terminology for mutex objects (much differs from the POSIX one). This one link gave me some understanding of it [1]. Here is the MSDN [1] description of what "abandoned mutex" is: " If a thread terminates without releasing its ownership of a mutex object, the mutex object is considered to be abandoned. A waiting thread can acquire ownership of an abandoned mutex object, but the wait function will returnWAITABANDONEDto indicate that the mutex object is abandoned. An abandoned mutex object indicates that an error has occurred and that any shared resource being protected by the mutex object is in an undefined state. If the thread proceeds as though the mutex object had not been abandoned, it is no longer considered abandoned after the thread releases its ownership. This restores normal behavior if a handle to the mutex object is subsequently specified in a wait function."

What does it mean to wait on mutex and ownership of the mutex object: "Any thread with a handle to a mutex object can use one of thewait functions <http://msdn.microsoft.com/en-_gb/library/windows/desktop/_ms687069%28v=vs.85%29.aspx_ _<http://msdn.microsoft.com/en-gb/library/windows/desktop/ms687069%28v=vs.85%29.aspx>>to request ownership of the mutex object. If the mutex object is owned by another thread, the wait function blocks the requesting thread until the owning thread releases the mutex object using theReleaseMutex _<http://msdn.microsoft.com/en-_gb/library/windows/desktop/_ms685066%28v=vs.85%29.aspx_ _<http://msdn.microsoft.com/en-gb/library/windows/desktop/ms685066%28v=vs.85%29.aspx>>function." How we can release mutex and wait on already owned mutex: " After a thread obtains ownership of a mutex, it can specify the same mutex in repeated calls to the wait-functions _<http://msdn.microsoft.com/en-_gb/library/windows/desktop/_ms687069%28v=vs.85%29.aspx_ _<http://msdn.microsoft.com/en-gb/library/windows/desktop/ms687069%28v=vs.85%29.aspx>>without blocking its execution. This prevents a thread from deadlocking itself while waiting for a mutex that it already owns. To release its ownership under such circumstances, the thread must callReleaseMutex _<http://msdn.microsoft.com/en-_gb/library/windows/desktop/_ms685066%28v=vs.85%29.aspx_ _<http://msdn.microsoft.com/en-gb/library/windows/desktop/ms685066%28v=vs.85%29.aspx>>once for each time that the mutex satisfied the conditions of a wait function." [1] http://msdn.microsoft.com/en-_gb/library/windows/desktop/_ms684266(v=vs.85).aspx <http://msdn.microsoft.com/en-gb/library/windows/desktop/ms684266(v=vs.85).aspx> -Aleksej On 05/13/2014 04:00 PM, David Holmes wrote: I don't understand this one at all. What is an "abandoned mutex"? For that matter why does the code wait on a mutex and an event? Do we already own the mutex? If so what does it mean to wait on it? If not then how can we release it? ??? Thanks, David On 13/05/2014 8:57 PM, Alan Bateman wrote: This is debugger's shared memory transport so cc'ing serviceability-dev as that is there this code is maintained. Is there a test case or any outline of the conditions that cause this? I think that would be useful to understand the issue further. -Alan On 13/05/2014 11:46, Aleksej Efimov wrote: Hi, Can I have a review for 8032901 bug [1] fix [2]. There is a possible case when 'WaitForMultipleObjects' function can return the WAITABANDONED0 [3] error value. In such case it's better to release the mutex and return error value. This will prevent other threads to be blocked on abandoned mutex. Thank you, Aleksej [1] https://bugs.openjdk.java.net/_browse/JDK-8032901 <https://bugs.openjdk.java.net/browse/JDK-8032901> [2] http://cr.openjdk.java.net/~_aefimov/8032901/9/webrev.00/ <http://cr.openjdk.java.net/~aefimov/8032901/9/webrev.00/> [3] http://msdn.microsoft.com/en-_gb/library/windows/desktop/_ms687025(v=vs.85).aspx <http://msdn.microsoft.com/en-gb/library/windows/desktop/ms687025(v=vs.85).aspx>

More information about the core-libs-dev mailing list