[PATCH] Problem with ruby 1.8.6-p36 (and p39) on Tiger

Hi,

> The problem can be easily reproduced under Tiger with a one-liner:
> % ruby -ve 'require "timeout" ; Timeout.timeout(5) { puts `df -h` }'
> ruby 1.8.6 (2007-06-07 patchlevel 36) [i686-darwin8.10.1]
> /usr/local/lib/ruby/1.8/timeout.rb:54: execution expired (Timeout::Error)
>         from /usr/local/lib/ruby/1.8/timeout.rb:56:in `timeout'
>         from -e:1
I investigated more and the function that blocks is
rb_thread_cancel_timer (eval.c). This function is called in
proc_exec_v (process.h) only in 1.8.6-p36 and -p39, not in the 1.8
branch nor in 1.8.6-p0. And it is empty if pthread is not enabled.

So I investigated a little more to see what was wrong in this
rb_thread_cancel_timer function, and why everything works fine under
systems other than Tiger (Linux, Leopard...). The reason seems to be
due to a bug in Tiger. Here is a little step-by-step of what happens
when calling in Ruby system() or ``:
- the process is forked, the parent does not do much but the child do
what follows ;
- the rb_thread_cancel_timer is called to kill the timer thread.
However, as this process is a child created using fork, it has no
timer thread (fork kills all threads in the child process except the
current one), and the time_thread references a thread in the parent.
So on most systems, the pthread_cancel and pthread_join done by
rb_thread_cancel_timer both fail. But under Tiger pthread_cancel
SUCCEEDS and pthread_join HANGS.

However, after seeing in the ChangeLog why this call to
rb_thread_cancel_timer is done (cf ruby-dev:30581), it's clear that
this cannot be simply removed. However it should not be called just in
a child created by fork in C. The simplest way may be to just put
time_thread_alive_p to false in the children of forks created to start
a process (for example in pipe_open in io.c or in rb_f_system in
process.c). However I do not know well enough Ruby to be 100 % sure of
this.

I attached a small patch using pthread_atfork to do this and it solves
at least on my small tests the problem with system() or `` in
Timeout.timeout, and the problem mentioned in ruby-dev:30581 is still
corrected. But I think it needs some reviews to be sure it does not
add any problem.

Cheers,
Vincent Isambart