July 19, 2011
On 2011-07-19 05:40, Steven Schveighoffer wrote:
> On Mon, 18 Jul 2011 18:39:01 -0400, Jonathan M Davis <jmdavisProg@gmx.com>
> 
> wrote:
> > On 2011-07-18 15:15, teo wrote:
> >> On Mon, 18 Jul 2011 18:14:45 +0000, Jonathan M Davis wrote:
> >> > On 2011-07-18 10:54, Simen Kjaeraas wrote:
> >> >> On Mon, 18 Jul 2011 18:06:46 +0200, Jonathan M Davis <jmdavisProg@gmx.com>
> >> >> 
> >> >> wrote:
> >> >> > On Monday 18 July 2011 15:55:52 teo wrote:
> >> >> >> On Mon, 18 Jul 2011 10:26:27 -0400, Steven Schveighoffer wrote:
> >> >> >> > On Sun, 17 Jul 2011 15:29:02 -0400, teo <teo.ubuntu@yahoo.com>
> >> >> >> > 
> >> >> >> > wrote:
> >> >> >> >> It looks like std.concurrency.spawn does not accept delegates.
> >> 
> >> Is
> >> 
> >> >> >> >> there
> >> >> >> >> any reason for that?
> >> >> >> > 
> >> >> >> > There is no type attached to the hidden 'this' pointer. So spawn
> >> >> >> 
> >> >> >> cannot
> >> >> >> 
> >> >> >> > guarantee it doesn't point to unshared data.
> >> >> >> > 
> >> >> >> > -Steve
> >> >> >> 
> >> >> >> Bad. I tried to pass as an argument a pointer to an instance of a
> >> >> >> class -
> >> >> >> a this pointer. That didn't work.
> >> >> > 
> >> >> > Only stuff that's immutable or implicitly convertible immutable can
> >> >> > be passed
> >> >> > across threads using spawn and send. Otherwise, there's a risk of
> >> 
> >> it
> >> 
> >> >> > ending up
> >> >> > getting altered by both threads (or altered by one when the other
> >> 
> >> one
> >> 
> >> >> > is using
> >> >> > at it). Sometimes, that can be a bit restrictive (particularly when
> >> >> > you _know_
> >> >> > that something isn't going to be altered by the thread sending it
> >> >> > after its
> >> >> > sent but the compiler doesn't), but it avoids all kinds of
> >> 
> >> problems.
> >> 
> >> >> > If you
> >> >> > want to send a class object across, then it needs to be immutable
> >> >> > (which tends
> >> >> > to be a bit of a pain to do for classes, since they need to have an
> >> >> > immutable
> >> >> > constructor, which is often a pain to do).
> >> >> 
> >> >> It could be that assumeUnique should handle this, by transforming the delegate into something spawn() and friends could handle.
> >> >> 
> >> >> Might be worth an enhancement request.
> >> > 
> >> > There have been discussions about how to do it in the past. Whether assumeUnique will work depends on whether casting to immutable(C) will work (where C is the class' type), and I don't know whether that cast will work or not. If it does, then assumeUnique will do it, since all
> >> 
> >> it
> >> 
> >> > does is cast to immutable, but it's the sort of thing that requires language support. There _might_ be a way to solve the problem with
> >> 
> >> some
> >> 
> >> > sort of Unique template that spawn and send knew about, but nothing of the sort has been done yet. But until a number of the issues with
> >> 
> >> const
> >> 
> >> > and immutable in the compiler have been sorted out, that sort of thing would probably be problematic anyway.
> >> > 
> >> > - Jonathan M Davis
> >> 
> >> This is a bit too restrictive in my opinion. Only the shared data
> >> between
> >> two threads should be immutable. But the threads can access all sorts of
> >> mutable data as well. And in this case we are actually talking about the
> >> control function (or start routine) of a thread. As long as it's address
> >> is fixed within the memory of a process its usage for that purpose
> >> should
> >> be fine. Nobody is going to pass that address around. Am I missing
> >> anything here?
> > 
> > When passing data between threads, it must be immutable.
> 
> I have to jump in and correct you, nobody else has.
> 
> You can also pass data marked as shared.
> 
> A solution could be to cast the class as shared, pass it, then cast it back to unshared (ensuring you don't access the class from the originator anymore).
> 
> This is not a compiler-enforced solution, but it gets the job done. But there is risk of concurrency errors if you don't do it right. My recommendation is to isolate the parts that create and pass the shared data.

I thought that spawn and send disallowed shared and that you had to deal with shared separately, but if that's not the case, then that's not the case. Regardless, you can't just pass anything with spawn or send (as the OP seems to be trying to do). They have restrictions to avoid concurrency bugs.

- Jonathan M Davis
July 19, 2011
On Tue, 19 Jul 2011 14:31:19 -0400, Jonathan M Davis <jmdavisProg@gmx.com> wrote:

> On 2011-07-19 05:40, Steven Schveighoffer wrote:
>> On Mon, 18 Jul 2011 18:39:01 -0400, Jonathan M Davis I have to jump in and correct you, nobody else has.
>>
>> You can also pass data marked as shared.
>>
>> A solution could be to cast the class as shared, pass it, then cast it
>> back to unshared (ensuring you don't access the class from the originator
>> anymore).
>>
>> This is not a compiler-enforced solution, but it gets the job done. But
>> there is risk of concurrency errors if you don't do it right. My
>> recommendation is to isolate the parts that create and pass the shared
>> data.
>
> I thought that spawn and send disallowed shared and that you had to deal with
> shared separately, but if that's not the case, then that's not the case.
> Regardless, you can't just pass anything with spawn or send (as the OP seems
> to be trying to do). They have restrictions to avoid concurrency bugs.

send requires the data pass this test:

!hasLocalAliasing!(T)

which maps to:

http://www.digitalmars.com/d/2.0/phobos/std_traits.html#hasUnsharedAliasing

So yes, you can pass shared data.  However, there is still no sanctioned way to "pass and forget" mutable unique thread-local data.  That is, you pass data to another thread, and it becomes local to that other thread instead of to the thread passed from.

You are correct that you can't just pass anything (i.e. thread-local mutable data), but in some cases, you have to force the issue.  Just because the compiler can't prove it's valid doesn't mean it isn't.  But it's definitely opening up a possibility for concurrency bugs.

-Steve
July 19, 2011
On Tue, 19 Jul 2011 18:31:19 +0000, Jonathan M Davis wrote:

> On 2011-07-19 05:40, Steven Schveighoffer wrote:
>> On Mon, 18 Jul 2011 18:39:01 -0400, Jonathan M Davis <jmdavisProg@gmx.com>
>> 
>> wrote:
>> > On 2011-07-18 15:15, teo wrote:
>> >> On Mon, 18 Jul 2011 18:14:45 +0000, Jonathan M Davis wrote:
>> >> > On 2011-07-18 10:54, Simen Kjaeraas wrote:
>> >> >> On Mon, 18 Jul 2011 18:06:46 +0200, Jonathan M Davis <jmdavisProg@gmx.com>
>> >> >> 
>> >> >> wrote:
>> >> >> > On Monday 18 July 2011 15:55:52 teo wrote:
>> >> >> >> On Mon, 18 Jul 2011 10:26:27 -0400, Steven Schveighoffer wrote:
>> >> >> >> > On Sun, 17 Jul 2011 15:29:02 -0400, teo <teo.ubuntu@yahoo.com>
>> >> >> >> > 
>> >> >> >> > wrote:
>> >> >> >> >> It looks like std.concurrency.spawn does not accept delegates.
>> >> 
>> >> Is
>> >> 
>> >> >> >> >> there
>> >> >> >> >> any reason for that?
>> >> >> >> > 
>> >> >> >> > There is no type attached to the hidden 'this' pointer. So spawn
>> >> >> >> 
>> >> >> >> cannot
>> >> >> >> 
>> >> >> >> > guarantee it doesn't point to unshared data.
>> >> >> >> > 
>> >> >> >> > -Steve
>> >> >> >> 
>> >> >> >> Bad. I tried to pass as an argument a pointer to an instance
>> >> >> >> of a class -
>> >> >> >> a this pointer. That didn't work.
>> >> >> > 
>> >> >> > Only stuff that's immutable or implicitly convertible immutable
>> >> >> > can be passed
>> >> >> > across threads using spawn and send. Otherwise, there's a risk
>> >> >> > of
>> >> 
>> >> it
>> >> 
>> >> >> > ending up
>> >> >> > getting altered by both threads (or altered by one when the
>> >> >> > other
>> >> 
>> >> one
>> >> 
>> >> >> > is using
>> >> >> > at it). Sometimes, that can be a bit restrictive (particularly
>> >> >> > when you _know_
>> >> >> > that something isn't going to be altered by the thread sending
>> >> >> > it after its
>> >> >> > sent but the compiler doesn't), but it avoids all kinds of
>> >> 
>> >> problems.
>> >> 
>> >> >> > If you
>> >> >> > want to send a class object across, then it needs to be
>> >> >> > immutable (which tends
>> >> >> > to be a bit of a pain to do for classes, since they need to
>> >> >> > have an immutable
>> >> >> > constructor, which is often a pain to do).
>> >> >> 
>> >> >> It could be that assumeUnique should handle this, by transforming the delegate into something spawn() and friends could handle.
>> >> >> 
>> >> >> Might be worth an enhancement request.
>> >> > 
>> >> > There have been discussions about how to do it in the past. Whether assumeUnique will work depends on whether casting to immutable(C) will work (where C is the class' type), and I don't know whether that cast will work or not. If it does, then assumeUnique will do it, since all
>> >> 
>> >> it
>> >> 
>> >> > does is cast to immutable, but it's the sort of thing that requires language support. There _might_ be a way to solve the problem with
>> >> 
>> >> some
>> >> 
>> >> > sort of Unique template that spawn and send knew about, but nothing of the sort has been done yet. But until a number of the issues with
>> >> 
>> >> const
>> >> 
>> >> > and immutable in the compiler have been sorted out, that sort of thing would probably be problematic anyway.
>> >> > 
>> >> > - Jonathan M Davis
>> >> 
>> >> This is a bit too restrictive in my opinion. Only the shared data
>> >> between
>> >> two threads should be immutable. But the threads can access all
>> >> sorts of mutable data as well. And in this case we are actually
>> >> talking about the control function (or start routine) of a thread.
>> >> As long as it's address is fixed within the memory of a process its
>> >> usage for that purpose should
>> >> be fine. Nobody is going to pass that address around. Am I missing
>> >> anything here?
>> > 
>> > When passing data between threads, it must be immutable.
>> 
>> I have to jump in and correct you, nobody else has.
>> 
>> You can also pass data marked as shared.
>> 
>> A solution could be to cast the class as shared, pass it, then cast it back to unshared (ensuring you don't access the class from the originator anymore).
>> 
>> This is not a compiler-enforced solution, but it gets the job done. But there is risk of concurrency errors if you don't do it right. My recommendation is to isolate the parts that create and pass the shared data.
> 
> I thought that spawn and send disallowed shared and that you had to deal with shared separately, but if that's not the case, then that's not the case. Regardless, you can't just pass anything with spawn or send (as the OP seems to be trying to do). They have restrictions to avoid concurrency bugs.
> 
> - Jonathan M Davis


Well, I tried to pass a this pointer, because I wasn't able to use a delegate as a control function of the thread. All I need is access to the data within a class instance. That isn't static data.

Basically I have an object which encapsulates the access to some resource. Let's call it a generator. One can configure some parameters and later just switch it on. That should be in a worker thread, because some operations are blocking. I also need certain events to be communicated to other threads within the process. Passing messages is fine, because the data is immutable.

Now with spawn I have to use either a regular function or a static method of the class. The only other option that comes to my mind is to copy the implementation of the message box from std.concurrency and use the core.thread.Thread directly instead of spawn. Perhaps that will work.

Next ›   Last »
1 2
Top | Discussion index | About this forum | D home