Re: deferred cancellation of ipc

[Marcus Brinkmann]

At Tue, 14 Oct 2003 16:00:35 -0700,
Volkmar Uhlig wrote:
> 
> After reading your initial scenario again there is a rather simple
> solution--use zero timeouts.  Instead of storing the cancel flag you
> store the intended send timeout value, which could be zero.  Of course
> you still run into a race on MP if you don't use some synchronization
> primitive, but that is a different issue you have to deal with anyway.  
> This solution assumes that you _only_ want to ensure that the thread
> does not block.

Mmmh, that might even work.  Although it is quite finicky.  However,
it can not work "instead" of a cancel flag, but in addition.
To-be-canceled thread could do:

1. Set desired IPC timeout
2. Check cancel flag
3. Ipc()
4. If timeout, check cancel flag.

The cancelling thread can then make the canceled thread jump into a
routine so that it resets its own timeout value and sets the cancel
flag.

I am not sure about MP synchronization, but I wasn't sure about it
before, either.

The by far easiest approach would be to have an in-kernel flag that
can be set with ExchangeRegisters and that is checked (and cleared) in
the Ipc() system call.  At least it defines semantically what feature
I want to have.

I'll look further into the timeout approach.

Thanks,
Marcus