Thu, 08 Apr 2010

Caution: python-multiprocessing, threads and glib don't mix

If you don't want to read this article, then just steer clear of python-multiprocessing, threads and glib in the same application. Let me explain why.

There's a rather famous bug in Gwibber in Ubuntu Lucid, where a gwibber-service process will start taking 100% of the CPU time of one of your cores if it can. While looking in to why this bug happened I learnt a lot about how multiprocessing and GLib work, and wanted to record some of this so that others may avoid the bear traps.

Python's multiprocessing module is a nice module to allow you to easily run some code in a subprocess, to get around the restriction of the GIL for example. It makes it really easy to run a particular function in a subprocess, which is a step up from what you had to do before it existed. However, when using it you should be aware how the way it works can interact with the rest of your app, because there are some possible nasties lurking there.

GLib is a set of building blocks for apps, most notably used by GTK+. It provides an object system, a mainloop and lots more besides. What we are most interested here is the mainloop, signals, and thread integration that it provides.

Let's start the explanation by looking at how multiprocessing does its thing. When you start a subprocess using multiprocessing.Process, or something that uses it, it causes a fork(2), which starts a new process with a copy of the programs current memory, with some exceptions. This is really nice for multiprocessing, as you can just run any code from that program in the subprocess and pass the result back without too much difficulty.

The problems occur because there isn't an exec(3) to accompany the fork(2). This is what makes multiprocessing so easy to use, but doesn't insert a clean process boundary between the processes. Most notably for this example, it means the child inherits the file descriptors of the parent (critically even those marked FD_CLOEXEC).

The other piece to this puzzle is how the GLib mainloop communicates between threads. It requires some mechanism where one thread can alert another that something of interest happened. To do this when you tell GLib that you will be using threads in your app by calling g_thread_init (gobject.threads_init() in Python) then it will create a pipe for use by glib to alert other threads. It also creates a watcher thread that polls one end of this pipe so that it can act when a thread wishes to pass something on to the mainloop.

The final part of the puzzle is what your app does in a subprocess with mutliprocessing. If you purely do something such as number crunching then you won't have any issues. If however you use some glib functions that will cause the child to communicate with the mainloop then you will see problems.

As the child inherits the file descriptors of the parent it will use the same pipe for communication. Therefore if a function in the child writes to this pipe then it can put the parent in to a confused state. What happens in gwibber is that it uses some gnome-keyring functions and that puts the parent in to a state where the watcher thread created by g_thread_init busy-polls on the pipe, taking up as much CPU time as it can get from one core.

In summary, you will see issues if you use python-multiprocessing from a thread and use some glib functions in the children.

There are some ways to fix this, but no silver bullet:

  • Don't use threads, just use multiprocessing. However, you can't communicate with glib signals between subprocesses, and there's no equivalent built in to multiprocessing.
  • Don't use glib functions from the children.
  • Don't use multiprocessing to run the children, use exec(3) a script that does what you want, but this isn't as flexible or as convenient.

It may be possible to use the support for different GMainContexts for different threads to work around this, but:

  • You can't access this from Python, and
  • I'm not sure that every library you use will correctly implement it, and so you may still get issues.

Note that none of the parties here are doing anything particularly wrong, it's a bad interaction caused by some decisions that are known to cause issues with concurrency. I also think there are issues when using DBus from multiprocessing children, but I haven't thoroughly investigated that. I'm not entirely sure why the multiprocessing child seems to have to be run from a non-main thread in the parent to trigger this, any insight would be welcome. You can find a small script to reproduce the problem here.

Or, to put it another way, global state bad for concurrency.

Posted at: 13:01 | category: /tech | Comments (5)


I would have thought that using multiple children as separate processes altogether and sending all the info using the dbus would work very well. It also gives you the advantage of using other languages that have better processing speeds like C or Vala for the children. So you just spawn a subprocess for the children and listen to the dbus for word back. I hope that makes sense.

Posted by Shane Fagan at Thu Apr 8 14:50:37 2010

Really the fault lies in Python for having a GIL and thus unspeakably terrible threading support.

Posted by Alex at Thu Apr 8 17:23:20 2010

could you give an example of why one would want to run threads in the main process of a python application using multiprocessing in connection with the gobject main loop? (in all cases i used gobject main loop and threading, threading was not required any more as soon as i used multiprocessing, at least in the main (gtk using) process.)

Posted by chrysn at Thu Apr 8 23:19:16 2010

Hi chrysn,

As I said, multiprocessing doesn't offer you things like glib signals between the processes.

Sometimes you want things to be in the same address space.

Note that I'm not advocating threads or multiprocessing or anything, just reporting my findings from debugging them in case someone else
can learn from it.

Thanks,

James

Posted by James Westby at Thu Apr 8 23:29:53 2010

Thank you for tracking down the problem and providing such a clear explanation. I really appreciate the effort that you and others at Canonical have put into addressing this issue for Gwibber.

-- Ryan

Posted by Ryan Paul at Fri Apr 9 08:47:45 2010