RFR: 8011882: Replace spin loops as back off when suspending

[David Holmes]

On 11/04/2013 10:30 PM, Rickard Bäckman wrote:
> David,
>
> On Apr 11, 2013, at 2:15 PM, David Holmes wrote:
>
>> Rickard,
>> On 11/04/2013 9:51 PM, Rickard Bäckman wrote:
>>> Hi all,
>>>
>>> can I please have reviews for this change.
>>>
>>> In the current implementation do_suspend uses spin loops to wait until a thread has been suspended. I would like to change that to use semaphores to reduce CPU usage and also make it easier to have a deterministic timeout. Since we are posting to the semaphore in the signal handler we can't use pthread_semaphore since it isn't async safe, which is why all implementations uses low-level os-specific semaphore constructs.
>>
>> POSIX Semaphore sem_post() is async-signal safe:
>
> Yes, and they are used in the implementation on Linux and BSD. Solaris uses the sema_{post,wait,…} (since that was we were already using for other things) and on Mac we use semaphore_{signal, wait, …}

So what did you mean about pthread_semaphore (what is that anyway?) ??

I really, really, really don't like seeing three versions of this class 
:( Can't BSD and Linux at least share a POSIX version? (And I wonder if 
we can actually mix-n-match UI threads on Solaris with POSIX semaphores 
on Solaris?)

David

> /R
>
>>
>> http://pubs.opengroup.org/onlinepubs/9699919799/functions/V2_chap02.html#tag_15_04
>>
>> David
>> -----
>>
>>> Bug: http://bugs.sun.com/view_bug.do?bug_id=8011882
>>> Webrev: http://cr.openjdk.java.net/~rbackman/8011882/
>>>
>>> Thanks
>>> /R
>>>
>