Re: [Proxool-developer] concurrency bug in Prototyper class?

[Bill H. <bi...@lo...>]

If you want to patch it and try it out in your app I'm certainly open to
including it Proxool. Anybody else got anything to say on the subject?

2009/6/19 Matt Salafia <msa...@gm...>

> In this specific case, I believe making 'connectionCount' a volatile
> variable satisfies both goals: correctness and performance.  It even
> establishes the 'out-of-thin-air' guarantee you don't have when reading a
> non-synchronized, non-volatile 64-bit value.
>
> I think sacrificing correctness for performance is OK in places where stale
> (although correct) values won't hurt you.  I don't believe it works with the
> variable holding keys to the connection pool.
>
> My application is massively multithreaded.  There are a large number of
> concurrent threads acquiring and releasing connections. The fact that these
> threads will always see the 'connectionCount' increment done in
> Prototyper.buildConnection( ) but might not see any decrements done in
> Prototyper.connectionRemoved( ) seems serious to me.
>
>
>
> On Fri, Jun 19, 2009 at 3:59 PM, Bill Horsman <bi...@lo...>wrote:
>
>> I see your point. We've jumped through a few hoops in Proxool to get the
>> balance right between concurrency and excessive blocking. As you'd expect,
>> our first priority is to ensure the code doesn't break in concurrent
>> situations, but it's closely followed by our second priority: that it must
>> be very fast and not do unnecessary locking. I'm not saying you're right or
>> wrong - just that it would be hasty to grab a lock to read the
>> connectionCount unless we explore the implications.
>>
>> 2009/6/19 Matt Salafia <msa...@gm...>
>>
>>> Possibly, but I don't know - it's indeterminate.  Without the proper
>>> memory visibility semantics (e.g. synchronization, immutability, volatile
>>> et. al.) I think all bets are off.
>>>
>>>
>>>
>>> On Fri, Jun 19, 2009 at 1:32 PM, Bill Horsman <bi...@lo...>wrote:
>>>
>>>> Is that a problem that corrects itself? In other words, is it a
>>>> temporary problem where a connection is refused when it didn't have to be
>>>> but that the connection count does end up being correct? - Bill
>>>>
>>>> 2009/6/19 Matt Salafia <msa...@gm...>
>>>>
>>>>>  It seems Class Prototyper should either acquire the lock on private
>>>>> member 'lock' whenever it reads or modifies 'connectionCount' or make
>>>>> 'connectionCount' a volatile variable.
>>>>>
>>>>> Prototyper.buildConnection( ) acquires the lock to read and modify
>>>>> 'connectionCount' but other methods do not: Prototyper.sweep( ),
>>>>> Prototyper.connectionRemoved( ) and Prototyper.getConnectionCount().
>>>>>
>>>>> I started poking around after my application started seeing
>>>>> ProxoolExceptions indicating the maximum connection count had been reached
>>>>> and discovered what appears to be a memory visibility issue.  If calls to
>>>>> Prototyper.connectionRemoved( ) and Prototyper.buildConnection( ) happen in
>>>>> different threads there could be unwarranted ProxoolExceptions for reaching
>>>>> the maximum connection count - a 'connectionPool' decrement might not be
>>>>> visible to the threads calling Prototyper.buildConnection( ).
>>>>>
>>>>>
>>>>>
>>>>>
>>>>> ------------------------------------------------------------------------------
>>>>> Are you an open source citizen? Join us for the Open Source Bridge
>>>>> conference!
>>>>> Portland, OR, June 17-19. Two days of sessions, one day of
>>>>> unconference: $250.
>>>>> Need another reason to go? 24-hour hacker lounge. Register today!
>>>>>
>>>>> http://ad.doubleclick.net/clk;215844324;13503038;v?http://opensourcebridge.org
>>>>> _______________________________________________
>>>>> Proxool-developer mailing list
>>>>> Pro...@li...
>>>>> https://lists.sourceforge.net/lists/listinfo/proxool-developer
>>>>>
>>>>>
>>>>
>>>>
>>>> ------------------------------------------------------------------------------
>>>> Are you an open source citizen? Join us for the Open Source Bridge
>>>> conference!
>>>> Portland, OR, June 17-19. Two days of sessions, one day of unconference:
>>>> $250.
>>>> Need another reason to go? 24-hour hacker lounge. Register today!
>>>>
>>>> http://ad.doubleclick.net/clk;215844324;13503038;v?http://opensourcebridge.org
>>>> _______________________________________________
>>>> Proxool-developer mailing list
>>>> Pro...@li...
>>>> https://lists.sourceforge.net/lists/listinfo/proxool-developer
>>>>
>>>>
>>>
>>>
>>> ------------------------------------------------------------------------------
>>> Are you an open source citizen? Join us for the Open Source Bridge
>>> conference!
>>> Portland, OR, June 17-19. Two days of sessions, one day of unconference:
>>> $250.
>>> Need another reason to go? 24-hour hacker lounge. Register today!
>>>
>>> http://ad.doubleclick.net/clk;215844324;13503038;v?http://opensourcebridge.org
>>> _______________________________________________
>>> Proxool-developer mailing list
>>> Pro...@li...
>>> https://lists.sourceforge.net/lists/listinfo/proxool-developer
>>>
>>>
>>
>>
>> ------------------------------------------------------------------------------
>> Are you an open source citizen? Join us for the Open Source Bridge
>> conference!
>> Portland, OR, June 17-19. Two days of sessions, one day of unconference:
>> $250.
>> Need another reason to go? 24-hour hacker lounge. Register today!
>>
>> http://ad.doubleclick.net/clk;215844324;13503038;v?http://opensourcebridge.org
>> _______________________________________________
>> Proxool-developer mailing list
>> Pro...@li...
>> https://lists.sourceforge.net/lists/listinfo/proxool-developer
>>
>>
>
>
> ------------------------------------------------------------------------------
> Are you an open source citizen? Join us for the Open Source Bridge
> conference!
> Portland, OR, June 17-19. Two days of sessions, one day of unconference:
> $250.
> Need another reason to go? 24-hour hacker lounge. Register today!
>
> http://ad.doubleclick.net/clk;215844324;13503038;v?http://opensourcebridge.org
> _______________________________________________
> Proxool-developer mailing list
> Pro...@li...
> https://lists.sourceforge.net/lists/listinfo/proxool-developer
>
>