Ok found the link, the paper's called:
Position paper: Nondeterminism is unavoidable, but data races are
pure evil
http://www.hpl.hp.com/techreports/2012/HPL-2012-218.pdf
Wheather Thread's name field is an issue or not depends on the code that
uses it (Object.toString() and Thread.getName() methods do) and what
that code does with the information.
Outrigger uses it in lifecycle logs, although it doesn't appear to be
called on any threads from ThreadPool. Instead Outrigger calls
getName() on threads it references from fields in OutriggerServerImpl.
These fields have since been changed to final, so if the name is set
during construction and not changed after publication, it's thread safe,
in addition these threads aren't started until after
OutriggerServerImpl's constructor returns.
The concurrency issues I most recently experienced were related to
Outrigger, I couldn't determine the cause of these failures, perhaps I
lack sufficient ability to reason deeply about these issues, my fix for
the problem was akin to carpet bombing; fix every concurrency problem
and that appears to have paid off.
It's very difficult to determine if toString() may be called on
ThreadPool threads. I reasoned it could be and that was sufficient
justification to not use setName().
Cheers,
Peter.
On 27/04/2013 1:20 AM, Patricia Shanahan wrote:
I've looked at the source code. The field "name" that is shared
between threads doing getName or setName on a given Thread is a
reference. Writes and reads of references are always atomic.
The worst that could happen is that the change to the name does not
propagate to all threads that might display information about the
thread. The proposed fix ensures that worst case outcome happens all
the time.
Patricia
On 4/26/2013 5:44 AM, Greg Trasuk wrote:
I'm curious, as I don't see any indication in the Javadocs that
setName() isn't thread safe. Is there another reference that calls that
out? And what would be the failure mode, apart from a mangled string in
a log output?
Personally, if the potential failure mode wasn't onerous, I'd opt for
more descriptive logging. Comprehensibility is everything when you're
troubleshooting.
Cheers,
Greg.
On Fri, 2013-04-26 at 05:48, Peter Firmstone wrote:
Hope you don't mind, I've removed the call to Thread.setName in
com.sun.jini.ThreadPool
As a result threads will be less descriptive, unfortunately setName
isn't thread safe, it's final and cannot be overridden. Thread.getName
is only thread safe if a Thread's name isn't changed after publication.
ThreadPool was the only instance of Thread.setName in River.
Regards,
Peter.
