Feature #19560
openIO#close_on_fork= and IO#close_on_fork?
Description
Context¶
Forking setups are extremely common in the Ruby ecosystem, as they remain the primary way to get parallelism with MRI.
Generally speaking it works very well, however there are two main issues library authors and application owners need to be careful of:
- Restarting threads
- Closing inherited connections and other file descriptors.
I believe we could make the second one much easier.
O_CLOFORK¶
A couple years ago, a new flag was added to the POSIX spec: O_CLOFORK
. Similar to O_CLOEXEC
, this file descriptor flag make it so the file descriptor is automatically closed upon forking.
Unfortunately its support is relatively limited for now. It's supported on macOS and some relatively exotic unixes, but not in Linux nor most BSDs.
The feature was discussed on Linux mailing list, but it seem to have encountered some strong opposition, so it's unclear if we can hope for it to be added.
That said, I don't think it would be too hard for Ruby to shim this feature by closing all IOs with close_on_fork?
right after fork.
Ruby shim¶
This can be implemented as a Ruby shim starting in Ruby 3.1 using the Process._fork
callback
class IO
def close_on_fork=(enabled)
if enabled
::CloseIOOnFork::IOS[self] = true
end
@close_on_fork = enabled
end
def close_on_fork?
@close_on_fork
end
end
module CloseIOOnFork
IOS = ObjectSpace::WeakMap.new
def _fork
pid = super
if pid == 0 # child
::CloseIOOnFork::IOS.each_key do |io|
io.close if io.close_on_fork?
end
end
pid
end
end
Process.singleton_class.prepend(CloseIOOnFork)
rd, rw = IO.pipe
rw.close_on_fork = true
pid = fork do
p rw.closed? # => true
end
Process.wait(pid)
Usage¶
With such feature, many network client would mostly just need to set this flag on their sockets, and just properly handle unexpectedly closed connections, which most already do.
Updated by mame (Yusuke Endoh) over 1 year ago
Discussed at the dev meeting.
One of the purposes of O_CLOFORK
seems to be to avoid thread race condition: if fork
is called immediately after another thread opens a file, the file descriptor would unintentionally leak to a child process. To prevent this, it is essential that O_CLOFORK can be atomically specified at open, not after open. So the API IO#close_on_fork=
is inappropriate for the purpose.
Considering this,
- @matz (Yukihiro Matsumoto) approved to define the constants provided by the OS: O_CLOFORK as
File::Constants::CLOFORK
, and FD_CLOFORK asFcntl::FD_CLOFORK
. - He did not approve an emulation layer for the case where the OS does not provide them. If it is absolutely necessary, more careful consideration and better API proposal will be required.
Incidentally, O_CLOEXEC is set by default for open IOs. Therefore, a similar race condition issue should not occur. In other words, IO#close_on_exec=
is not essential itself. It is just an auxiliary API for opt-out.
It would be difficult to set O_CLOFORK by default because there are already many Ruby programs that pass IOs to a child process implicitly at fork
.
Updated by byroot (Jean Boussier) over 1 year ago
One of the purposes of O_CLOFORK seems to be to avoid thread race condition
Yes, but that's not the reason I want it for Ruby. Agreed that it would be nice to solve both problems though.
So the API IO#close_on_fork= is inappropriate for the purpose.
Based on the discussion log, it seems that the intent I described here was missed. My goal here isn't to protect from race condition caused by threads forking, but to get an automatic close of some IOs in normal conditions (no race condition).
It would be difficult to set O_CLOFORK by default
Yes, it shouldn't be set by default.
approved to define the constants provided by the OS: O_CLOFORK as File::Constants::CLOFORK, and FD_CLOFORK as Fcntl::FD_CLOFORK.
Thanks, I'll implement that at least, and I'll see about how to create IOs with that option enabled to allow for an emulation layer.
Updated by Dan0042 (Daniel DeLorme) over 1 year ago
- @matz (Yukihiro Matsumoto) approved to define the constants provided by the OS: O_CLOFORK as
File::Constants::CLOFORK
, and FD_CLOFORK asFcntl::FD_CLOFORK
.
If it's possible to open a file with CLOFORK, it should also be possible to answer the question "has this file been opened with CLOFORK?" Would that be done via IO#close_on_fork?
or maybe some flags & CLOFORK
?
Updated by akr (Akira Tanaka) over 1 year ago
I found an issue about O_CLOFORK after the meeting.
It closes FD in a child process at fork.
But it does not close IO object.
So, an open IO object can reference a closed FD.
If the IO object is inactive (no method called, not collected by GC), it is not a problem.
But if the FD is reused and the IO object is (accidentally) active, the FD is wrongly operated.