From: Charles Oliver Nutter <charles.nutter@...>
Date: 2007-12-27T05:27:58+09:00
Subject: Re: Experimental PATCH to improve thread performance

Sylvain Joyeux wrote:
> On Thu, Dec 27, 2007 at 01:57:22AM +0900, Charles Oliver Nutter wrote:
>> Except that requiring users to use synchronization points in order to make 
>> a feature safe basically means everyone will do it the unsafe way. 
> What I meant was in the same line of though than Brent's: only act upon
> queued exceptions if the thread is inside a synchronization point. That
> way, the exception mechanism, without being completely safe, is reduced
> to cases where it can be safe. Ensuring that the child thread is able to
> understand the exception is another matter altogether: it is the
> responsibility of the program designer.

Certainly, the feature could be made safe. And in JRuby's internal 
implementation of Thread#raise, we do exactly that kind of checkpointing 
to make sure Java "finally" blocks and JRuby runtime structures are not 
damaged. But there's nothing equivalent in Ruby code; Thread#raise is 
not safe to use in its current form without additional gymnastics on the 
user's part. When it finds common use in stdlibs like thread.rb--which 
many libraries and applications use without *any* sort of 
checkpointing--the problem is made even worse.

>> a clear checkpoint and agreement from both parent and child thread that 
>> it's ok to raise, you can't *ever* know that a thread is safe to raise in,
> Threads are never created out of the void. Threads are created by the
> program itself and multi-thread messaging is part of the design of the
> whole program. Inter-thread exceptions are a flexible inter-thread error
> messaging and I certainly think that removing them altogether would
> remove a nice idea from the Ruby language.

If made safe and moved toward a more message-passing-like logic, sure. 
But as it's implemented now, it's broken and dangerous.

- Charlie