OpenJDK8 Thread.sleep will deadlock while turn down system date time.

[张泽鹏]

I apply the patch and reinstall from ports, it works well for me now.
Thanks Kim! thanks everyone!

发件人: Jung-uk Kim <jkim at freebsd.org> <jkim at freebsd.org>
回复: Jung-uk Kim <jkim at freebsd.org> <jkim at freebsd.org>
日期: 2017年4月13日 at 上午1:17:41
至: Daniel Eischen <deischen at freebsd.org> <deischen at freebsd.org>, Eric van
Gyzen <eric at vangyzen.net> <eric at vangyzen.net>
抄送: redraiment at gmail.com <redraiment at gmail.com>
<redraiment at gmail.com>, Konstantin
Belousov <kostikbel at gmail.com> <kostikbel at gmail.com>,
freebsd-java at freebsd.org <freebsd-java at freebsd.org>
<freebsd-java at freebsd.org>
主题:  Re: OpenJDK8 Thread.sleep will deadlock while turn down system date
time.

On 04/11/2017 19:55, Daniel Eischen wrote:
> On Tue, 11 Apr 2017, Eric van Gyzen wrote:
>
>> On 4/8/17 2:03 AM, Konstantin Belousov wrote:
> [ ... ]
>>>
>>> If JVM sets timeouts using absolute times than it might be fixed in
>>> latest
>>> HEAD and stable/11.
>>
>> The JVM uses pthread_cond_timedwait() to implement Thread.sleep(), so
>> it always uses absolute times. Furthermore, it uses the default
>> clock, CLOCK_REALTIME. My recent change (r315280) does not fix this
>> behavior; in fact, I believe it will make the problem worse, since
>> moving the clock forward will wake the thread prematurely.
>>
>> I think the JVM should be fixed to use CLOCK_MONOTONIC. Would someone
>> like to do the research to determine the correct behavior of
>> Thread.sleep()? Specifically, should the duration of the sleep be
>> affected by adjustments to the real-time clock? I expect that it
>> should /not/ be affected, especially since the API takes a
>> relative/interval time. Ideally, we would find the answer in the
>> formal language specification; failing that, I would be satisfied with
>> empirical data from testing on Windows, Linux, MacOS, and Solaris.
>> I'll be happy to write a patch once we know what it should do.
>>
>> Please keep me CC'd, since I'm not on freebsd-java at . (Thanks for the
>> CC, Kostik.)
>
> It seems CLOCK_REALTIME behavior has bugged Linux in the past, too. I
> think using CLOCK_MONOTONIC, perhaps via pthread_condattr_setclock() in
> our JVM, would be the better approach.

http://bugs.java.com/view_bug.do?bug_id=6900441
http://hg.openjdk.java.net/jdk8u/jdk8u/hotspot/rev/2e6938dd68f2

Basically, it was fixed in Linux source but it wasn't merged to ours
somehow.

Please try the attached patch.

> You'd probably also want to check that whatever java.util.concurrent
> relies on, if different, also behaves similarly (monotonically).

Runtime behaviour depends on HotSpot implementation.

Jung-uk Kim