There's an instance variable in Process, threadId, that is nil (unbound) or holds the threadId of the thread it can run on. There is one primitive to bind it to the current thread and one to unbind. It is unsafe to manipulate the threadId inst var directly.
Sent: Tuesday, February 08, 2011 8:57 PM
Subject: Re: [Pharo-project] Status of Alien
On Tue, Feb 8, 2011 at 5:45 PM, Schwab,Wilhelm K <
[hidden email]<mailto:
[hidden email]>> wrote:
Eliot,
No lack of patience implied: just mentioning some things that I have picked up as off in the future and that would make us all stronger.
Re the calls on OS threads, that is GREAT news. What strategy are you using to assign calls to the threads? I ask because there is a tradeoff between using a pool and taking what is ready to go, and coping with thread affinity. Example: OpenSSL is "easy" to call if one uses blocking calls and simply moves them out of the main thread (at least it worked for me on Dolphin). When I did this, I had to write a wrapper DLL to combined calls (connect, read, send, etc.) and checks for error status because otherwise, Dolphin would typically use different threads for the calls and the error test call was meaningless. IIRC, now they always use the same OS thread for any given Smalltalk Process - I think.
There is no pool. Instead, arbitrary threads can run the VM, but only one at a time. So the strategy to control calls is to decide which processes are bound to which threads. A new Smalltalk process can be executed by any thread. There is a well-known thread id for "the GUI thread" and it may be necessary to have more. So to arrange that a call is made on the GUI thread one arranges that the cal is made from a Smalltalk process that is bound to the GUI thread. To arrange that calls are made form some specific thread once again one binds one or more processes to that thread and makes calls from those processes. The VM's Smalltalk scheduler is modified to do a thread switch when a process switch is made to a process bound to other than the current thread (i.e. it is now a classic two-level scheduler). Hence scheduling semantics don't change; it is still a real-time scheduler that always runs the highest-priority runnable Smalltalk process. Its just that under the covers an implicit thread switch might also occur when a process switch occurs. Make sense?
best
Eliot
I fixed a few more things that have been pinning me to my hacked vms, so I will hopefully be able to more easily experiment. I will look around for Alien on Linux and give it a try if I find it. I am also looking forward to running my "real" images through Cog.
Bill
________________________________________
Sent: Tuesday, February 08, 2011 11:19 AM
To: [hidden email]<mailto:[hidden email]>
Cc: Squeak Virtual Machine Development Discussion
Subject: Re: [Pharo-project] Status of Alien
On Tue, Feb 8, 2011 at 4:16 AM, Schwab,Wilhelm K <
[hidden email]<mailto:
[hidden email]><mailto:
[hidden email]<mailto:
[hidden email]>>> wrote:
Torsten,
I have been under the impression that it has never worked on Linux - it would be great to be wrong on that. Does anyone have Alien on Linux waiting for someone to try it?
I would be very surprised if it didn't work. We at least use the data manipulation side of things routinely at Teleplace.
One thing slowing my search for FFI/Alien improvements has been that I was initially dependent on some vm modifications that I made. Some off-line correspondence with Dave Lewis and the benefit of some time might have yielded a solution that will let me return to stock vms.
I need to integrate my new callback support and the ReentrantFFIPlugin into the standard interpreter as soon as possible. The ReentrantFFIPlugin provides compatibility with the old squeak FFI plugin but also allows aliens to be passed as arguments, at least to the extent of passing Alien Callbacks. This means its possible to discard the Alien callout support and use the FFI instead (but just to reiterate to keep the Alien data representation and callback facilities). The project for the next few weekends has to be harmonising the primitive set and functionality of Cog and the standard interpreter now that the StackToRegisterMappingCogit looks complete.
FFI has generally worked for me. I am doing a horrible job of handling double arrays, but I am at least able to use them. My "solution" makes assumptions about byte order and I have declare double* as void*.
There was mention of extending FFI with callbacks. Given that capability, the only thing missing would then to be make calls on OS threads to avoid blocking the image (but blocking the calling Smalltalk Process for the duration of the call).
I have this in prototype and want to release soon. But there is too much else going on right now. Patience please.
best
Eliot
Bill
________________________________________
Sent: Tuesday, February 08, 2011 4:04 AM
To: [hidden email]<mailto:[hidden email]><mailto:[hidden email]<mailto:[hidden email]>>