OS support for thread stopping

I've recently had done some thinking about how can an OS make the shutdown of a multithreaded application or its part easier, to stop the threads quickly but relatively nicely on request. I think a very helpful support can be pretty easy to implement:

• Have system call to request the shutdown of an OS-level thread.
• This call would make any system calls that could block non-blocking (except for the thread synchronization calls, like mutex or condition variable operations), would interrupt any current such blocking calls, and return an error. This includes the timed sleeps, and even the thread synchronization calls with timeouts would return an error immediately instead of sleeping. Optionally, it might also just fail any file operations altogether, possibly except for close(). Note that the file descriptors as such wouldn't be revoked altogether, they would just kind of stop working in one thread, but the other threads would still work.
• A problem with this might be in such things as logging: when your thread is stopping, it might still write the useful log files. Revoking these log files would be  Well, one solution for this would be to buffer the logs in memory and write from a separate thread. Another might be to keep some kind of exception list.
• The threads might also beget threads, and keeping track of all of them is complicated. So there is a need for grouping them, like the process groups in the OS, or Triceps Triead fragments, and then stopping the whole group.

Yet another separate approach might be to say "but if you want the ability to stop something, why don't you run it as a separate process and kill if when needed?". This would make a lot of sense for such things as PAM authentication modules.

There are problems with the separate processes too. But after some thinking, they seem to be fairly easy to solve.

When you think about processes, usually a separate binary comes to mind but it doesn't have to. You can just do a fork() and run from the same binary, and you even get the whole address space copy-on-write out of the box. Though I guess it might be interesting to NOT copy-on-write the whole address space. Instead just allocate some communication area for passing the arguments and results, and get rd of the rest, maybe even re-running the run-time loader from scratch in the new context (that would be kind of like the Java class loader spaces).

Communication with a separate process is kind of painful. But if you map an anonymous memory segment, that segment can be inherited by the child process. Indeed, should be inherited to act as the communication region (truly shared, not COW) even if the rest of address space is dropped. And then you can happily create the mutexes and condition variables in it for synchronization. The arguments and results would have to be serialized to pass through this region but it shouldn't be such a big deal. It would be also possible to keep some large data dictionaries in this communication space.

Of course the callbacks wouldn't work across the process boundary. But there is a solution for this too: create a local thread pool (or simply a thread) for callbacks that would get the requests from the remote process though a portion of the same communications region, execute the callback, and return the results back there.

I think this looks feasible even without any new technology needed, just needs a helper library to implement the interface. This isn't even a new concept, it's a classic microkernel architecture without the common memory space. Yeah, copying the data in the serialized form adds an overhead but it's not that much overhead compared to say JVM, and people use Java and .NET all over the place with no regrets.