Version française
Home     About     Download     Resources     Contact us    
Browse thread
[Caml-list] Resource acquisition is initialization
[ Home ] [ Index: by date | by threads ]
[ Search: ]

[ Message by date: previous | next ] [ Message in thread: previous | next ] [ Thread: previous | next ]
Date: -- (:)
From: Brian Hurt <brian.hurt@q...>
Subject: Re: [Caml-list] Resource acquisition is initialization
On Thu, 12 Dec 2002, Alessandro Baretta wrote:

> 
> 
> Brian Hurt wrote:
> > On Wed, 11 Dec 2002, Blair Zajac wrote:
> 
> > The only thing I can think of which cannot tolerate the cleanup slop is 
> > releasing a mutex.  Deleting the mutex itself is a good application for 
> > GC, unlocking the mutex is not.
> 
> It is not the GC which unlocks the mutex but the application 
> code. See Mutex.unlock .

Calling a function to unlock a mutex is precise- so long as you remember 
to always unlock the mutex.  At least, as I'm understanding things.  It's 
the difference between (in Java-esque):

    {
        mutex.lock();
        /* do stuff needing the mutex to be locked */
        mutex.unlock();
    }

and:
    {
        mutex_lock lock_obj = mutex.get_lock();
        /* do stuff needing the mutex to be locked */
    }
    /* lock_obj is now garbage- eventually the GC will clean it up, thus
     * unlocking the mutex.  But this may be a while.  Until that happens,
     * everyone still thinks we're holding the mutex locked.
     */

The problem with the first representation is that you can forget to unlock 
the mutex, especially on error paths.  Consider:
    {
        mutex.lock();
        /* do stuff */
        if (error) { throw AnException(); }
        /* Opps- did I just forget to unlock the mutex? */
        mutex.unlock();
    }

The problem with the second way is that you continue holding the lock for 
some indefinate amount of time until the GC frees the object.  This is 
what I meant by "the time you hold the lock is defined by the lifetime of 
an object".

Note that for most other resources, holding them a little too long isn't a 
problem.  Consider file handles, for example.  OK, you don't want to leak 
filehandles anymore than you want to leak memory, opening them and 
forgetting to ever close them.  But if a file handle stays open a little 
longer than you really need it to be, it's not a catastrophe.  There are 
generally enough spare handles that you can still open new files before 
all the old files get closed.  So the lag isn't critical.

> 
> > And for that, I think representing 
> > holding the lock as the lifetime of an object to be a bad idea.  Were Caml 
> > interested more in doing multithreaded code, I'd recommend something like 
> > Java's synchronize keyword.
> 
> What's the difference with respect to locking a Mutex?
> 

Java explicitly binds the time you hold a mutex to a certain code block- 
literally, a set of curly braces {} defines the length of time when you 
hold a lock.  When you leave that code block, via any means, the lock is 
automatically released.  So you have code like:

    synchronized(foo) {
        /* All the code in here has foo locked */
    }
    /* Code here has foo unlocked */

This is usefull because the programmer doesn't have to remember to unlock 
the mutex- the simple act of leaving the block automatically unlocks the 
mutex.  So:

    synchronized(foo) {
       if (error) throw AnException();
    }

isn't an error.  The throw will unlock the mutex on leaving the code 
block.  This does slow down exceptions a little- now you need to have a 
stack of locks to unlock as you throw the exception.  However this is way 
simpler to implement, and way faster, than having to keep track of all 
objects (and their types) on the stack and call the appropriate 
destructors as you unwind the stack.  And dealing with the corner cases 
implied (what happens when you throw an exception, which calls a 
destructor for an object on the stack, and that destructor throws another 
exception?).

This is the behavior I think was wanted- to bind the mutex lock to a 
specific peice of code.  I'd much rather do it explicitly, with a language 
extension, than a kludge which causes more problems than it solves.

Speaking of which, I've been noodling around with wether it's possible to 
implement synchronization automatically in the common case.  What objects 
need to have synchronization is easy to determine- those objects which can 
be reached from 2 or more threads need synchronization.  The question is, 
what synchronization.  Or how much.  Can you protect a whole slew of 
objects with a single lock?

Obviously, removing as much synchronization as possible is good for
performance.  Also, it's possible to construct examples where if multiple 
objects are protected by a single lock you don't have deadlock, while if 
the objects are each protected by their own lock deadlock is possible.

Brian


-------------------
To unsubscribe, mail caml-list-request@inria.fr Archives: http://caml.inria.fr
Bug reports: http://caml.inria.fr/bin/caml-bugs FAQ: http://caml.inria.fr/FAQ/
Beginner's list: http://groups.yahoo.com/group/ocaml_beginners