After investigating some waitpid problems, I was able to repoduce the bug in a simple test case. Make sure to save your
time or re-run your ntp client before executing the script :

-- test.rb (run as root) --
pid = fork { exec 'sleep 100000' }

t = Thread.new do
  puts "Waiting pid #{pid}"
  Process.waitpid(pid)
  puts "Pid gone"
end

sleep 2

system "ps -p#{pid}"
system 'date -s 0000'
system "kill #{pid}"
sleep 1
system "ps -p#{pid}"

puts 'Should have printed "Pid gone"'
t.join
---------------

The critical line, is "date -s 0000". If you remove it, everything is fine. If you don't, you'll see the second
"ps -p#{pid}" show a <defunct> process and "Pid gone" will not be printed.

From a short code review, it seems like the thread delay is stored as an absolute time (eval.c:10904). When setting
the time backwards, the thread won't be executed until it goes back to the time it was last executed. In our case, the
C waitpid WNOHANG won't be called and we'll see the defunct process.

I don't know for other platforms, but Linux has a monotonic timer that doesn't suffer the "ntp" problem, which
could be used to fix that kind of problem.

# ruby -v
ruby 1.8.6 (2007-06-07 patchlevel 36) [i486-linux]

Date: 2007-09-08 09:59
Sender:  

I have investigated the problem a little bit more.

It should be possible to replace gettimeofday in
eval.c:10010 (static double timeofday()) by
clock_gettime(CLOCK_MONOTONIC, ...) under Linux. I say
"should" because it is only used by the scheduler to
update
thread->delay but I am not shure that delay is never used
otherwise than relatively.

Date: 2007-08-31 11:12
Sender: Shyouhei Urabe

I don't know it it's by design. Assigning this to matz.