Bug #15367: IO.select is not resumed when io-object gets closed

• Status: Open
• Assignee: -
• Target version: -
• Backport: 2.4: UNKNOWN, 2.5: UNKNOWN
• Tags: core

[printercu (Max Melentiev)]

Here is sample code:

rp, wp = IO.pipe
t2 = Thread.new { IO.select([rp]) }
# This also does not work:
# t2 = Thread.new { IO.select([rp], nil, [rp]) }
sleep 0.01
rp.close
t2
# => #<Thread:0x00000000089b6ce0@(pry):51 sleep>

It happens only on linux, tested with 2.5.1, 2.6.0-preview2. On macOS it gives error, as expected:

#<Thread:0x00007fab3aebce58@(pry):5 run> terminated with exception (report_on_exception is true):
Traceback (most recent call last):
	1: from (pry):5:in `block in <main>'
(pry):5:in `select': Bad file descriptor (Errno::EBADF)
> t2
=> #<Thread:0x00007fab3aebce58@(pry):5 dead>

[normalperson (Eric Wong)]

[e-mail] wrote:

https://bugs.ruby-lang.org/issues/15367
It happens only on linux, tested with 2.5.1, 2.6.0-preview2. On macOS it
gives error, as expected:

Right, it's platform-specific. I brought it up a few months ago
and can implement it; but it'll add some overhead
(current IO#close during IO#write/IO#write already does).

@akr (Akira Tanaka) doesn't want it, @matz (Yukihiro Matsumoto) hasn't responded, yet.

https://bugs.ruby-lang.org/issues/14760

[shevegen (Robert A. Heiler)]

If I may inquire, how significant would the overhead be?

While I think Akira's comment is perfectly fine on its own, I
feel that if other people notice a different behaviour between
different operating systems with regards to IO.select then this
may surprise them. (I myself use almost exclusively Linux.)

(On a not-so-relevant comment, my first project in ruby was an
IRC bot and that was also when I used IO.select for the first
time.)

[MSP-Greg (Greg L)]

On Windows (MinGW), the thread is also sleeping...

[normalperson (Eric Wong)]

[e-mail] wrote:

If I may inquire, how significant would the overhead be?

I'd have to implement it to know for sure...

Thread::Light [Bug #13618] has a process-wide FD map anyways
(similar to the kernel fdtable) to deal with multiple
threads waiting on different operations on the same FD,
so we could take advantage of that.

IO.select is a heavy operation, already

Right now, the IO#close notification isn't so bad if few
threads are operating on FDs, but it's O(n) where `n' is
the number of threads calling rb_io_blocking_region in parallel,
regardless of FD.

Back to the Thread::Light FD map, the `n' would be reduced
to the number of threads for a certain FD, because there's
a per-FD linked-list (and getting to that linked-list is
just an array lookup, so O(1).

[printercu (Max Melentiev)]

I'm using IO.select with ssl socket as it's suggested in docs https://ruby-doc.org/core-2.5.3/IO.html#method-c-select :

    def read_from_socket
      socket.read_nonblock(read_buffer_size)
    rescue IO::WaitReadable
      IO.select([socket.to_io])
      retry
    rescue IO::WaitWritable
      IO.select(nil, [socket.to_io])
      retry
    end

Other code is like this (simplified):

def run
  loop do
    data = read_from_socket
    process(data)
  end
rescue Errno::EBADF, IOError
  raise unless @stopped
end

def stop
  @stopped = true
  socket.close
end

Signal.trap('TERM') { Thread.new { stop } } # thread is just workaround for trap contex
run
# exit

I can't use just Thread#kill to not stop thread while it runs #process.
It works fine on macOS because socket.close makes read_from_socket fail with Errno::EBADF which is rescued later.
However this does not work on linux and #run hangs forever.

Is there a right way for this task to not face multi-thread limitations of IO.select?

For now I use workaround:

    SELECT_TIMEOUT = 10

    def read_from_socket
      socket.read_nonblock(read_buffer_size)
    rescue IO::WaitReadable
      nil until IO.select([socket.to_io], nil, nil, SELECT_TIMEOUT)
      retry
    rescue IO::WaitWritable
      nil until IO.select(nil, [socket.to_io], nil, SELECT_TIMEOUT)
      retry
    end

[normalperson (Eric Wong)]

[e-mail] wrote:

I can't use just Thread#kill to not stop thread while it runs #process.
It works fine on macOS because socket.close makes read_from_socket fail with Errno::EBADF which is rescued later.
However this does not work on linux and #run hangs forever.

Is there a right way for this task to not face multi-thread
limitations of IO.select?

Several ways (there may be more, but I'm tired)

a) You can use Thread#kill with Thread.handle_interrupt around
#process, I think...

b) IO.select allows waiting on multiple FDs, so you could
wait on one process-wide pipe which every IO.select
call checks on:

int_pipe = IO.pipe
trap(:TERM) { int_pipe[1].write('.') }

IO.select([@socket, int_pipe[0]])

IO.select([int_pipe[0]], [@socket])

(Btw, you don't need #to_io when calling IO.select,
it already calls #to_io).

c) if you're dealing with Internet traffic, it's unreliable
and there's malicious clients trying to DoS you.
So you're going to need a SELECT_TIMEOUT anyways, I think...