[jboss-jira] [JBoss JIRA] (JGRP-1877) System.nanoTime() can be negative

[Bela Ban (JIRA)]

    [ https://issues.jboss.org/browse/JGRP-1877?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13000823#comment-13000823 ] 

Bela Ban commented on JGRP-1877:
--------------------------------

[comment by david]
{quote}
now here's one more problem for you, now that you've accepted nanoTime()'s benefits and quirks  on some JVMs and some OSes, it uses the CPU's TSC - unfortunately, each CPU has its own, which might be slightly off from each other - so if your thread migrates between CPUs at just the wrong time, and you're on such a JVM on such an OS on such a CPU, you could still get a negative time for your elapsed time
{quote}

In other words, the thread measuring t0 and t1 has to be the same thread ! Not a problem, IIRC, as t0 and t1 are always measured by the same thread.

> System.nanoTime() can be negative
> ---------------------------------
>
>                 Key: JGRP-1877
>                 URL: https://issues.jboss.org/browse/JGRP-1877
>             Project: JGroups
>          Issue Type: Bug
>            Reporter: Bela Ban
>            Assignee: Bela Ban
>            Priority: Critical
>             Fix For: 3.5.1, 3.6
>
>
> According to the javadoc, {{System.nanoTime()}} should only be used to measure _elapsed time_, but not compute a _target time in the future_, as {{nanoTime()}} might return a a time in the future. 
> Code like the one below might fail:
> {code:title=Responses.waitFor()|borderStyle=solid}
> public boolean waitFor(long timeout) {
>         long wait_time;
>         final long target_time=System.nanoTime() + TimeUnit.NANOSECONDS.convert(timeout, TimeUnit.MILLISECONDS); // ns
>         lock.lock();
>         try {
>             while(!done && (wait_time=target_time - System.nanoTime()) > 0) {
>                 try {
>                     cond.await(wait_time,TimeUnit.NANOSECONDS);
>                 }
>                 catch(InterruptedException e) {
>                 }
>             }
>             return done;
>         }
>         finally {
>             lock.unlock();
>         }
>     }
> {code}
> When computing {{target_time}}, {{System.nanoTime()}} could return a negative value (numeric overflow) or a value in the future. In the first case, {{target_time}} could be negative, so the method would not block at all. In the latter case, {{target_time}} could be huge, so the method would block for a long time.
> Investigate all occurrences where we use {{nanoTime()}} to compute a time in the future, and see what impact a future value value could have. Possibly replace with {{System.currentTimeMillis()}} or the _time service_.



--
This message was sent by Atlassian JIRA
(v6.3.1#6329)