The "multiprocessing" module uses a bare fork() to create child processes under Linux, so the children get a copy of the entire state of the parent process.  But under Windows, child processes are freshly spun-up Python interpreters with none of the data structures or open connections of the parent process available.  This means that code that tests fine under Linux, because it is depending on residual parent state in a way that the programmer has not noticed, can fail spectacularly under Windows.

Therefore, the "multiprocessing" module should offer an option under Linux that ignores the advantage of being able to do a bare fork() and instead spins up a new interpreter instance just like Windows does.  Some developers will just use this for testing under Linux, so their test results are valid for Windows too; and some developers might even use this in production, preferring to give up a bit of efficiency under Linux in return for an application that will show the same behavior on both platforms.  Either way, an option that lets the developer subvert the simple "sys.platform != 'win32'" check in "forking.py" would go a long way towards helping us write platform-agnostic Python programs.

This is on my wish list; but I have not had time to do it. Patch welcome.

Jesse, it's great to learn it's on your wish list too!

Should I design the patch so that (a) there is some global in the module that needs tweaking to choose the child creation technique, or (b) that an argument to the Process() constructor forces a full interpreter exec to make all platforms match, or (c) that a process object once created has an attribute (like ".daemon") that you set before starting it off?  Or (d) should there be a subclass of Process that, if specifically used, has the fork/exec behavior instead of just doing the fork?

My vote would probably be for (b), but you have a much better feel for the library and its style than I do.

I pretty much agree with (b) an argument - your gut instinct is correct - there's a long standing thread in python-dev which pretty much solidified my thinking about whether or not we need this (we do). 

Any patch has to be backwards compatible by the way, it can not alter the current default behavior, also, it has to target python3 as 2.7 is nearing final, and this is a behavioral change.

+1 for this issue; I've also wished for this feature in the past.

+1

I have suspected that this may be necessary, not just merely useful, for some time, and issue6721 seems to verify that.  In addition to adding the keyword arg to Process, it should also be added to Pool and Manager.

Is anyone working on a patch? If not I will work on a patch asap.

No one is currently working on a patch AFAIK

Here is a patch which adds the following functions:

  forking_disable()
  forking_enable()
  forking_is_enabled()
  set_semaphore_prefix()
  get_semaphore_prefix()

To create child processes using fork+exec on Unix, call
forking_disable() at the beginning of the program.

I have tested the patch on Linux (by adding forking_disable() to
test_multiprocessing), and it seems to work.  However, the patch does
not modify test_multiprocessing, and I am not sure of the best way to
do so.  (See below.)

There are some issues with named semaphores.  When forking is
disabled, the name of the semaphore must be left unlinked so that
child processes can use sem_open() on the name.  The patch therefore
delays unlinking the name (only when forking is disabled) until the
original SemLock object is garbage collected or the process which
created it exits.

But if a process is killed without exiting cleanly then the name may
be left unlinked.  This happens, for instance, if I run
test_multiprocessing and then keep hitting ^C until all the processes
exit.  On Linux this leaves files with names like

  /dev/shm/sem.mp-fa012c80-4019-2

which represent leaked semaphores.  These won't be destroyed until the
computer reboots or the semaphores are manually removed (by using
sem_unlink() or by unlinking the entry from the file system).

If some form of this patch is accepted, then the problem of leaked
semaphores needs to be addressed, otherwise the buildbots are likely
run out of named semaphores.  But I am not sure how best to do this in
a platform agnostic way.  (Maybe a forked process could collect names
of all semaphores created, via a pipe.  Then it could try to
sem_unlink() all those names when all write-ends of the pipe are
closed.)

Small fix to patch.

Thanks for the patch sbt.
I think this is indeed useful, but I'm tempted to go further and say we should make this the default - and only - behavior. This will probably break existing code that accidentaly relied the fact that the implementation uses a bare fork(), but i'd say it's worth it:
- it's cleaner
- it will make it possible to remove all the ad-hoc handlers called after fork()
- it will remove the only place in the whole stdlib where fork() isn't followed by exec(): people who get bitten by issue #6721 will thus only be people calling explicitely fork(), in which case they're the sole responsibles for their misery ;-)

Another - although less common - advantage over the current implementation is that now one can run out of memory pretty easily if the operating system doesn't do overcommitting if you work with a large dataset. If fork() is followed by an exec, no problem.

Thoughts?

> Thanks for the patch sbt.
> I think this is indeed useful, but I'm tempted to go further and say
> we should make this the default - and only - behavior. This will
> probably break existing code that accidentaly relied the fact that the
> implementation uses a bare fork(),

There is probably lots of such code:
- code that passes non-pickleable function object / function args to
execute in the child process
- code that executes code with side effects at module top level
- code that relies (willingly or not) on other stuff such as fds being
inherited

> I think this is indeed useful, but I'm tempted to go further and say we 
> should make this the default - and only - behavior. This will probably 
> break existing code that accidentaly relied the fact that the 
> implementation uses a bare fork(), but i'd say it's worth it:

I'm not convinced about making it the default behaviour, and certainly not the only one.

I have a working patch which ensures that leaked semaphores get cleaned up on exit.  However, I think to add proper tests for the patch, test_multiprocessing needs to be refactored.  Maybe we could end up with

multiprocessing_common.py
test_multiprocessing_processes_fork.py
test_multiprocessing_processes_nofork.py
test_multiprocessing_manager_fork.py
test_multiprocessing_manager_nofork.py
test_multiprocessing_threads.py
test_multiprocessing_others.py

The actual unittests would be in multiprocessing_common.py and test_multiprocessing_others.py.  The other files would run the unittests in multiprocessing_common.py using different configurations.

Thoughts?

> There is probably lots of such code:

> I'm not convinced about making it the default behaviour, and 
> certainly not the only one.

Then I'm not convinced that this patch is useful.
Having three different implentations and code paths doesn't sound like a good idea to me.
fork() vs fork() + exec() is an implementation detail, and exposing such tweakables to the user will only make confusion worse.

On Wednesday, December 21, 2011 at 10:04 AM, Charles-François Natali wrote:

While I would tend to agree with you in theory - I don't think we should make it the default - at least not without a LOT of lead time. There's a surprising amount of code relying on the current behavior that I think the best course is to enable this option, and change the docs to steer users in this direction.

For users jumping from 2.x into 3.x, I think the less surprises they have the better, and changing the default behavior of the stdlib module in this was would qualify as surprising.

See also consolidated Issue13558 for additional justification for processes option on OS X.

Attached is an updated version of the mp_fork_exec.patch.  This one is able to reliably clean up any unlinked semaphores if the program exits abnormally.

mp_split_tests.patch splits up the test_multiprocessing.py:

test_multiprocessing_misc.py
  miscellaneous tests which need not be run with multiple configurations

mp_common.py
  testcases which should be run with multiple configurations

test_multiprocessing_fork.py
test_multiprocessing_nofork.py
test_multiprocessing_manager_fork.py
test_multiprocessing_manager_nofork.py
test_multiprocessing_threads.py
  run the testcases in mp_common.py with various configurations

I don't know what the others think, but I'm still -1 on this patch.
Not that I don't like the principle - it's actually the contrary: in a
perfect world, I think this should be made the default -and only -
behavior on POSIX. But since it may break existing code, we'd have to
keep both implementations for POSIX systems, with - at least to me -
little benefit.
Having three different implementations, with different codepaths, will
increase the cognitive burden, make the code less readable, and
debugging harder:
- user: I'm getting this error with multiprocessing
- dev: On Windows or on Unix?
- user: On Unix
- dev: Do you use the fork()+exec() version or the bare fork() version?
- user: what's fork() and exec()?

A use case for not using fork() is when your parent process opens some system resources of some sort (for example a listening TCP socket). The child will then inherit those resources, which can have all kinds of unforeseen and troublesome consequences (for example that listening TCP socket will be left open in the child when it is closed in the parent, and so trying to bind() to the same port again will fail).

Generally, I think having an option for zero-sharing spawning of processes would help code quality.

By the way, instead of doing fork() + exec() in pure Python, you probably want to use _posixsubprocess.fork_exec().

+1

I still have to use parallel python (pp) in our application stack because the fork() approach causes a lot of strange issues in our application. It might be the punishment for embedding a Java runtime env into a Python process, too. :)

> A use case for not using fork() is when your parent process opens some 
> system resources of some sort (for example a listening TCP socket). The 
> child will then inherit those resources, which can have all kinds of 
> unforeseen and troublesome consequences (for example that listening TCP 
> socket will be left open in the child when it is closed in the parent, 
> and so trying to bind() to the same port again will fail).
>
> Generally, I think having an option for zero-sharing spawning of 
> processes would help code quality.

The patch as it stands still depends on fd inheritance, so you would need to use FD_CLOEXEC on your listening socket.  But yes, it should be possible to use the closefds feature of _posixsubprocess.


BTW, I also have working code (which passes the unittests) that starts a helper process at the beginning of the program and which will fork processes on behalf of the other processes.  This also solves the issue of unintended inheritance of resources (and the mixing of fork with threads) but is as fast as doing normal forks.

For updated code see http://hg.python.org/sandbox/sbt#spawn

This uses _posixsubprocess and closefds=True.

http://hg.python.org/sandbox/sbt#spawn now contains support for starting processes via a separate server process.  This depends on fd passing support.

This also solves the problem of mixing threads and processes, but is much faster than using fork+exec.  It seems to be just as fast as using plain fork.

I have tested it successfully on Linux and a MacOSX buildbot.  (OpenSolaris does not seem to support fd passing.)

At the begining of your program you write

    multiprocessing.set_start_method('forkserver')

to use the fork server.

Alternatively you can use

    multiprocessing.set_start_method('spawn')

to use _posixsubprocess.fork_exec() with closefds=True on Unix or 

    multiprocessing.set_start_method('fork')

to use the standard fork method.

This branch also stops child processes on Windows from automatically inheriting inheritable handles.

The test suite can be run with the different start methods by doing

    python -m test.test_multiprocessing_fork
    python -m test.test_multiprocessing_spawn
    python -m test.test_multiprocessing_forkserver

Richard, apart from performance, what's the advantage of this approach over the fork+exec version?
Because it seems more complicated, and although I didn't have a look a this last patch, I guess that most of the fork+exec version could be factorized with the Windows version, no?
Since it's only intented to be used as a "debugging"/special-purpose replacement - it would probably be better if it could be made as simple as possible. Also, as you've noted, FD passing isn't supported by all Unices out there (and we've had some reliability issues on OS-X, too).

> Richard, apart from performance, what's the advantage of this approach over the 
> fork+exec version?

It is really just performance.  For context running the unittests in a 1 cpu linux VM gives me

fork:
real    0m53.868s
user    0m1.496s
sys     0m9.757s

fork+exec:
real    1m30.951s
user    0m24.598s
sys     0m25.614s

forkserver:
real    0m54.087s
user    0m1.572s             # excludes descendant processes
sys     0m2.336s             # excludes descendant processes

So running the unit tests using fork+exec takes about 4 times as much cpu time.

Starting then immediately joining a trivial process in a loop gives

fork:        0.025 seconds/process
fork+exec:   0.245 seconds/process
forkserver:  0.016 seconds/process

So latency is about 10 times higher with fork+exec.

> Because it seems more complicated, and although I didn't have a look a this last 
> patch, I guess that most of the fork+exec version could be factorized with the
> Windows version, no?

The different fork methods are now implemented in separate files.  The line counts are

  117 popen_spawn_win32.py
   80 popen_fork.py
  184 popen_spawn_posix.py
  191 popen_forkserver.py

I don't think any more sharing between the win32 and posix cases is possible.  (Note that popen_spawn_posix.py implements a cleanup helper process which is also used by the "forkserver" method.)

> Since it's only intented to be used as a "debugging"/special-purpose replacement - it 
> would probably be better if it could be made as simple as possible.

Actually, avoiding the whole fork+threads mess is a big motivation.  multiprocessing uses threads in a few places (like implementing Queue), and tries to do so as safely as possible.  But unless you turn off garbage collection you cannot really control what code might be running in a background thread when the main thread forks.

> Also, as you've noted, FD passing isn't supported by all Unices out there
> (and we've had some reliability issues on OS-X, too).

OSX does not seem to allow passing multiple ancilliary messages at once -- but you can send multiple fds in a single ancilliary message.  Also, when you send fds on OSX you have to wait for a response from the other end before doing anything else.  Not doing that was the cause of the previous fd passing failures in test_multiprocessing.

Numbers when running on Linux on a laptop with 2 cores + hyperthreading.

RUNNING UNITTESTS:
  fork:
    real    0m50.687s
    user    0m9.213s
    sys     0m4.012s

  fork+exec:
    real    1m9.062s
    user    0m48.579s
    sys     0m6.648s

  forkserver:
    real    0m50.702s
    user    0m4.140s        # excluding descendants
    sys     0m0.708s        # excluding descendants

LATENCY:
  fork:       0.0071 secs/proc
  fork+exec:  0.0622 secs/proc
  forkserver: 0.0035 secs/proc

Still 4 times the cpu time and 10 times the latency.  But the latency is far lower than in the VM.

I think the forkserver approach is a good idea. It is what a lot of users will choose.

forkserver won't work everywhere though so the fork+exec option is still desirable to have available.  Threads can be started by non-python code (extension modules, or the larger C/C++ program that is embedding the Python interpreter within it).  In that context, by the time the multiprocessing module is can be too late to start a fork server and there is no easy way for Python code to determine if that is the case.

The safest default would be fork+exec though we need to implement the fork+exec code as a C extension module or have it use subprocess (as I noted in the mb_fork_exec.patch review).

> The safest default would be fork+exec though we need to implement the 
> fork+exec code as a C extension module or have it use subprocess (as I 
> noted in the mb_fork_exec.patch review).

That was an old version of the patch.

In the branch

    http://hg.python.org/sandbox/sbt#spawn

_posixsubprocess is used instead of fork+exec, and all unnecessary fds are closed.  See

    http://hg.python.org/sandbox/sbt/file/8f08d83264a0/Lib/multiprocessing/popen_spawn_posix.py

ah, i missed that update.  cool!  +1