www.digitalmars.com         C & C++   DMDScript  

digitalmars.D - Suggestion for Condition variables

reply Graham St Jack <grahams acres.com.au> writes:
Here is a suggestion for an extension to D to handle pthread-like 
condition variables in a way that fits neatly with the "synchronized" 
keyword.

I use threads a lot, and quite often find that I need conditions, but 
even though all the functionality is there in the locks library, it is 
galling that I can't use the same mutex as the "synchronized" keyword.

The alternatives seem to be to:
* Eliminate or ignore the synchronized keyword and use a library all the 
   time.
* Make the language "aware" of the library, and provide a way of getting 
an object's mutex.
* Add some more keywords to handle conditions directly.

My suggestion is for the last, because I like synchronized, and because 
I don't like making the language aware of the libraries.


So, the suggestion is to add:

* A "condition" keyword that behaves for the most part just like a bool 
data type, but is actually a condition variable.

* A "waitfor" keyword. It would be used like:
   waitfor (condition_name);
   to wait for the condition to become true, and
   waitfor (condition_name; timeout_value);
   to wait for the condition to come true, or the timeout period to 
expire. Timeout could be ulong nanoseconds, for instance.

Here is a simple example with two conditions in a class in 
fast-and-loose pseudo-code. The two conditions are used to show how 
simple this arrangement can be compared to trying the same thing in 
Java, which has  a limit of one condition per object.

The example Queue class is used to facilitate the transfer of data 
between two threads - a supplier and a worker. The supplier calls 
finalize() when it wants the worker to terminate, blocking until all the 
work has been done.


class Queue {

   bool finalizing;
   condition ready;
   condition done;

   // add an item to the queue - called by a supplier thread
   synchronized void add(int item) {
     ...add item to back of a queue...
     ready = true;
   }

   // remove an item from the queue - called by a worker thread
   synchronized int remove() {
     waitfor (ready);
     if (...queue is empty...) {
       done = true;
       throw new FinalisedException();
     }
     int item = ...remove item from front of queue...
     ready = !(...queue is empty...) || finalizing;
     return item;
   }

   // finalize the queue - normally called by the supplier thread
   synchronized void finalize() {
     finalizing = true;
     ready = true;
     waitfor (done; timeout_value);
     if (!done)
       throw new TimeoutException();
   }
}

The compiler can do all the necessary stuff like:
* re-evaluating the condition after waking up, and waiting again if the 
condition has been cleared by a critical race.
* making it an error to use conditions outside of synchronized methods.

I think this is much tidier than the Java approach, and it is certainly 
a whole lot easier than using pthread conditions - even through a library.
Feb 26 2006
parent reply Sean Kelly <sean f4.ca> writes:
Graham St Jack wrote:
 Here is a suggestion for an extension to D to handle pthread-like 
 condition variables in a way that fits neatly with the "synchronized" 
 keyword.
 
 I use threads a lot, and quite often find that I need conditions, but 
 even though all the functionality is there in the locks library, it is 
 galling that I can't use the same mutex as the "synchronized" keyword.
 
 The alternatives seem to be to:
 * Eliminate or ignore the synchronized keyword and use a library all the 
   time.
 * Make the language "aware" of the library, and provide a way of getting 
 an object's mutex.
 * Add some more keywords to handle conditions directly.
I've been thinking about this as well, and I think it helps that the monitor location is already specified in the ABI. This should allow library use of monitors whether there's any official language support or not.
 My suggestion is for the last, because I like synchronized, and because 
 I don't like making the language aware of the libraries.
 
 
 So, the suggestion is to add:
 
 * A "condition" keyword that behaves for the most part just like a bool 
 data type, but is actually a condition variable.
Personally, I'd be happy with library code for this, so long as it could use the built-in monitors. I'd also like a bit more flexibility with how locks are obtained and such, but perhaps this is better left for later. Sean
Feb 27 2006
parent reply Oskar Linde <oskar.lindeREM OVEgmail.com> writes:
Sean Kelly skrev:
 Graham St Jack wrote:
 Here is a suggestion for an extension to D to handle pthread-like 
 condition variables in a way that fits neatly with the "synchronized" 
 keyword.

 I use threads a lot, and quite often find that I need conditions, but 
 even though all the functionality is there in the locks library, it is 
 galling that I can't use the same mutex as the "synchronized" keyword.

 The alternatives seem to be to:
 * Eliminate or ignore the synchronized keyword and use a library all 
 the   time.
 * Make the language "aware" of the library, and provide a way of 
 getting an object's mutex.
 * Add some more keywords to handle conditions directly.
I've been thinking about this as well, and I think it helps that the monitor location is already specified in the ABI. This should allow library use of monitors whether there's any official language support or not.
 My suggestion is for the last, because I like synchronized, and 
 because I don't like making the language aware of the libraries.


 So, the suggestion is to add:

 * A "condition" keyword that behaves for the most part just like a 
 bool data type, but is actually a condition variable.
Personally, I'd be happy with library code for this, so long as it could use the built-in monitors. I'd also like a bit more flexibility with how locks are obtained and such, but perhaps this is better left for later.
Any comments on Tommie Gannerts implementation? http://www.digitalmars.com/d/archives/digitalmars/D/31340.html /Oskar
Feb 27 2006
next sibling parent Sean Kelly <sean f4.ca> writes:
Oskar Linde wrote:
 
 Any comments on Tommie Gannerts implementation?
 http://www.digitalmars.com/d/archives/digitalmars/D/31340.html
It does pretty much what I want, but i haven't found the time to play with it yet. Sean
Feb 27 2006
prev sibling parent reply Graham St Jack <grahams acres.com.au> writes:
Oskar Linde wrote:
 Sean Kelly skrev:
 
 Graham St Jack wrote:

 Here is a suggestion for an extension to D to handle pthread-like 
 condition variables in a way that fits neatly with the "synchronized" 
 keyword.

 I use threads a lot, and quite often find that I need conditions, but 
 even though all the functionality is there in the locks library, it 
 is galling that I can't use the same mutex as the "synchronized" 
 keyword.

 The alternatives seem to be to:
 * Eliminate or ignore the synchronized keyword and use a library all 
 the   time.
 * Make the language "aware" of the library, and provide a way of 
 getting an object's mutex.
 * Add some more keywords to handle conditions directly.
I've been thinking about this as well, and I think it helps that the monitor location is already specified in the ABI. This should allow library use of monitors whether there's any official language support or not.
 My suggestion is for the last, because I like synchronized, and 
 because I don't like making the language aware of the libraries.


 So, the suggestion is to add:

 * A "condition" keyword that behaves for the most part just like a 
 bool data type, but is actually a condition variable.
Personally, I'd be happy with library code for this, so long as it could use the built-in monitors. I'd also like a bit more flexibility with how locks are obtained and such, but perhaps this is better left for later.
Any comments on Tommie Gannerts implementation? http://www.digitalmars.com/d/archives/digitalmars/D/31340.html /Oskar
The patch neatly implements a fourth option - give a library back-door access to an object's mutex. While I haven't tried it out, it achieves what I need very well. I do have a preference for an addition to the language because the compiler can spot a few kinds of mistakes for you, but this looks a whole lot more achievable. Surely something like this could be slipped into phobos - maybe even into std.thread. How come this didn't go further? Lack of support for conditions is a glaring omission from phobos.
Feb 27 2006
parent Sean Kelly <sean f4.ca> writes:
Graham St Jack wrote:
 
 How come this didn't go further? Lack of support for conditions is a 
 glaring omission from phobos.
Perhaps simply because it wasn't formally submitted? I couldn't say. I definately plan to add conditions to Ares before too terribly long as I find them indispensable. sean
Feb 27 2006