pyplot: Fix exception in _backend_selection during import [backport to 1.4.x]

[wmanley]

if the new style introspection GObject python bindings are in use. With pygobject >= 3.13.4 the following:

from gi.repository import GObject
from matplotlib import pyplot

causes an exception to be raised:

AttributeError: When using gi.repository you must not import static modules
like "gobject". Please change all occurrences of "import gobject" to "from
gi.repository import GObject". See:
https://bugzilla.gnome.org/show_bug.cgi?id=709183

It is not valid to use both non-introspection based and introspection based PyGObject in the same process. Backend probing will import gobject (i.e. the non-introspection bindings) if it sees that the 'gtk' module is loaded.
Unfortunately it wouldn't check if this was the pygi or old-style gtk module. This commit adds this check avoiding the exception.

This check was added to PyGObject in d704033

[wmanley]

Fixes #2901

[tacaswell]

Please use () line continuation, not \.

[wmanley]

Done, thanks.

[tacaswell]

cc @fariza as a gtk expert.

I seem to recall that we have some of the new-style imports someplace in the code. I may be out-of-scope for this PR, but I think if gi.repository is in the path, it should fall back to which ever backend uses the new-style import.

[wmanley]

Well there's Gtk3Agg and GTK3Cairo, so I guess the equivalent would be:

    elif ('gtk' in sys.modules
            and 'gi.repository.GLib' in sys.modules
            and backend not in ('GTK3Agg', 'GTK3Cairo')):
        from gi.repository import GLib
        if GLib.MainLoop().is_running():
            if is_agg_backend:
                rcParams['backend'] = 'GTK3Agg'
            else:
                rcParams['backend'] = 'GTK3Cairo'

but I confess I don't understand the original code. My reading of gobject.MainLoop().is_running() is that it creates a new gobject.MainLoop and then checks if it's running, which it can't be because we just created it. i.e. by my reading this should always return False.

I wrote a little script to test this:

from gi.repository import GLib
m = GLib.MainLoop()
print "Initially running:", m.is_running()

def probe():
    print "New loop:", GLib.MainLoop().is_running()
    print "Old loop:", m.is_running()
    m.quit()

GLib.timeout_add(1, probe)
m.run()

and the Gtk2 style bindings equivalent:

import gobject
m = gobject.MainLoop()
print "Initially running:", m.is_running()

def probe():
    print "New loop:", gobject.MainLoop().is_running()
    print "Old loop:", m.is_running()
    m.quit()

gobject.timeout_add(1, probe)
m.run()

They both give the same result:

Initially running: False
New loop: False
Old loop: True

I agree that this these larger issues should be out-of-scope for this pull request. For one thing I imagine you wouldn't want to backport a patch that may change the selection of the backend to 1.4.0.

[tacaswell]

I don't really understand this code either, I was hoping you did!

[fariza]

Sorry to come late to the party.
I haven't deal with mainloop in a long time. Even less in GTK3.
As soon as I get some time I will try to check this.

[wmanley]

Would it be ok to have this patch merged? It might not implement the full-solution that we'd like, but it does fix the crash and is unlikely to have any undesirable side-effects.

[pelson]

Would it be ok to have this patch merged? It might not implement the full-solution that we'd like, but it does fix the crash and is unlikely to have any undesirable side-effects.

The error message is pretty clear, so it seems reasonable to me. Any objection if I cherry-pick this to v1.4.x @tacaswell?

[tacaswell]

No problem with cherry-picking it back from me.

[fariza]

Sorry.

I didn't have time to check it. I'm in the middle of some Angular mess
On 25 Sep 2014 14:15, "Thomas A Caswell" notifications@github.com wrote:

No problem with cherry-picking it back from me.

—
Reply to this email directly or view it on GitHub
#3541 (comment)
.

[pelson]

Merged with commit 236355c into v1.4.x. f9feefe...236355c