Feature #5788: Thread#at_exit

• Status: Rejected
• Assignee: -
• Target version: -

[Glass_saga (Masaki Matsushita)]

=begin
I propose a new method Thread#at_exit.
It is to register a block which will be called when the thread ends.

p Thread.main #=> #<Thread:0x007f828fa97fe0 run>
p t = Thread.new{ # do something } #=> #<Thread:0x007f828fc9e500 run>
t.at_exit{ p Thread.current }
t.join #=> #<Thread:0x007f828fc9e500 run>

Thread.main.at_exit is equal to Kernel.at_exit.

I think this method is useful for waiting plural running threads end.

Without Thread#at_exit:
t1 = Thread.new{ sleep 1 }
t2 = Thread.new{ sleep 2 }

[t1, t2].each do |t|
Thread.new do
t.join
puts "Thread ended!"
end
end

[t1, t2].each(&:join)

To handle threads' end immediately, the same number of threads as ones to wait is needed.

For example, Ruby's standard library thwait.rb does so.

With Thread#at_exit:
t1 = Thread.new{ sleep 1 }
t2 = Thread.new{ sleep 2 }

[t1, t2].each do |t|

It runs when interpreter ends.

t.at_exit{ puts "Thread ended!" }
end

New threads are not necessary.

I made a patch.
Patched ruby passes test-all.
=end

[nobu (Nobuyoshi Nakada)]

• Status changed from Open to Feedback

=begin
Just reviewed briefly.

• (({th->at_exit})) needs to be marked,
• (({th->at_exit})) should be hidden, and
• a hook registered to the main thread seems to be executed in a child
  process forked in a sub-thread too.

Second, what will happen if the thread has terminated already? Also,
it might be useful if the thread or its value is passed to the hook.
=end

[sorah (Sorah Fukumori)]

i don't think the name of this method is good… are there any more suitable name for this method?

[kosaki (Motohiro KOSAKI)]

I don't think ruby impl w/o GVL can implement this feature.

[Glass_saga (Masaki Matsushita)]

• File patch2.diff patch2.diff added

=begin
Nakada-san, thank you for your reviewing.
I improved the patch at some points.

• th->at_exit is to be marked.
• If the thread is dead, Thread#at_exit returns nil.
• The thread is passed to the hooks(including hooks on main thread).
• Registered hooks are executed in reverse order of registration(same as END or Kernel.at_exit).
  =end

[nobu (Nobuyoshi Nakada)]

=begin
Calling (({rb_gc_mark()})) only once at the hook definition makes no sense.
You have to mark it in (({rb_thread_mark()})).
=end

[normalperson (Eric Wong)]

Motohiro KOSAKI [e-mail] wrote:

I don't think ruby impl w/o GVL can implement this feature.

pthread_key_create() can be used to register a destructor function.
I don't know about other thread APIs, though.

[kosaki (Motohiro KOSAKI)]

2011/12/27 Eric Wong [e-mail]:

Motohiro KOSAKI [e-mail] wrote:

I don't think ruby impl w/o GVL can implement this feature.

pthread_key_create() can be used to register a destructor function.
I don't know about other thread APIs, though.

No.
The proposal API can replace with another thread's exiting hook. It's
racy be definition.
pthread_key_create() can't help.

Moreover, the proposal syntax may accelerate careless racy code.
That's bad. We sould
only encourage good custom.

[Glass_saga (Masaki Matsushita)]

• File patch3.diff patch3.diff added

=begin
I modified the patch.

• use mutex on define_at_exit to avoid race.
• mark th->at_exit and th->at_exit_lock in rb_thread_mark().
  =end

[mame (Yusuke Endoh)]

• Target version set to 2.6

[naruse (Yui NARUSE)]

• Target version deleted (2.6)

[Glass_saga (Masaki Matsushita)]

• Status changed from Feedback to Closed

[Glass_saga (Masaki Matsushita)]

• Status changed from Closed to Rejected