libffi 3.1~rc1 regression: crashes on i386; python3.4 crashed with SIGSEGV in g_callable_info_free_closure()

[Martin Pitt]

https://launchpad.net/ubuntu/+source/libffi/3.1~rc1-2 has a regression which causes lots of crashes on i386. This particular crash report is a (local) retrace of https://jenkins.qa.ubuntu.com/job/trusty-adt-gtk-3.0/109/ARCH=i386,label=adt/ but this affects all reverse depdencies of libffi, such as

  https://jenkins.qa.ubuntu.com/job/trusty-adt-python-cffi/37/ARCH=i386,label=adt
  https://jenkins.qa.ubuntu.com/job/trusty-adt-gtk-3.0/109/ARCH=i386,label=adt/
  https://jenkins.qa.ubuntu.com/job/trusty-adt-ruby-ffi/26/ARCH=i386,label=adt/

and gvfs, umockdev, ubiquity, etc.

I reproduced this in a local trusty i386 schroot and confirm that downgrading libffi6 to the previous version 3.0.13-12 re-fixes things.

So far a small reproducer is:

  sudo apt-get install python3-gi gir1.2-gtk-3.0 xvfb
  xvfb-run python3 -c 'from gi.repository import GLib, Gtk; GLib.timeout_add_seconds(1, Gtk.main_quit, None); Gtk.main()'

Or you run python-cffi's test suite:

  python3 -m pytest -v -k test_callback_as_function_argument testing

ProblemType: Crash
DistroRelease: Ubuntu 14.04
Package: python3.4-minimal 3.4.0-1
Uname: Linux 3.13.0-19-generic i686
Architecture: i386
Date: Fri Mar 28 07:05:41 2014
ExecutablePath: /usr/bin/python3.4
ExecutableTimestamp: 1395533063
PackageArchitecture: amd64
ProcCmdline: python3
ProcCwd: /tmp/adt-run.QKVQFy/dsc0-build/gtk+3.0-3.10.7
ProcEnviron:
 TERM=linux
 PATH=(custom, no user)
 XDG_RUNTIME_DIR=<set>
 LANG=en_US.UTF-8
 SHELL=/bin/sh
Signal: 11
SourcePackage: python3.4
StacktraceTop:
 g_callable_info_free_closure (callable_info=0x9e8a5b0, closure=0xb6b43008) at girepository/girffi.c:426
 _pygi_invoke_closure_free (data=0x9efdd50) at ../../gi/pygi-closure.c:638
 _pygi_destroy_notify_callback_closure (cif=0x9efddbc, result=0xbfaec770, args=0xbfaec710, data=0x0) at ../../gi/pygi-closure.c:703
 ffi_closure_SYSV_inner (closure=0xb6b43030, respp=0xbfaec77c, args=0xbfaec790) at ../src/x86/ffi.c:503
 ffi_closure_SYSV () at ../src/x86/sysv.S:199
UserGroups: adm audio cdrom dialout dip floppy kvm netdev plugdev sudo video

[Martin Pitt (pitti)]

As a data point, when I rebuild gobject-introspection and pygobject against the new libffi, it works again. This suggests an ABI break, not in terms of added/dropped symbols but something more subtle such as changed macro values, enum values changing, etc.

[Martin Pitt (pitti)]

I tried to do the same with python-cffi, but it fails to build due to test suite failures (due to that segfault). Building it with DEB_BUILD_FLAGS=nocheck and installing the debs doesn't fix the crash for python-cffi's autopkgtest; not surprisingly, as it's the same test suite that gets run during package build.

[Martin Pitt (pitti)]

I prepared a reversion upload which is now sitting in -unapproved for the release team's consideration. If there's a quick fix, we can of course reject this and fix it properly instead.

[Launchpad Janitor (janitor)]

This bug was fixed in the package libffi - 3.1~rc1+r3.0.13-12

---------------
libffi (3.1~rc1+r3.0.13-12) trusty; urgency=medium

  * Revert 3.1~rc1-2 upload, it causes an unannounced ABI break on at least
    i386, causing lots of segfaults. (LP: #1298824)
 -- Martin Pitt <email address hidden> Fri, 28 Mar 2014 09:55:09 +0100