[Webkit-unassigned] [Bug 145347] New: [GTK][Stable] Crashes when SoupSession is destroyed in exit handler

[bugzilla-daemon at webkit.org]

https://bugs.webkit.org/show_bug.cgi?id=145347

            Bug ID: 145347
           Summary: [GTK][Stable] Crashes when SoupSession is destroyed in
                    exit handler
    Classification: Unclassified
           Product: WebKit
           Version: 528+ (Nightly build)
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: Normal
          Priority: P2
         Component: WebKit Gtk
          Assignee: webkit-unassigned at lists.webkit.org
          Reporter: mcatanzaro at igalia.com

We have a couple downstream reports of crashes that occur when g_object_unref destroys a SoupSession in an exit handler. E.g.

Truncated backtrace:
Thread no. 1 (10 frames)
 #0 free at /lib64/libc.so.6
 #1 _gnutls_free_dh_info at /lib64/libgnutls.so.28
 #2 _gnutls_free_auth_info at /lib64/libgnutls.so.28
 #3 gnutls_deinit at /lib64/libgnutls.so.28
 #4 g_tls_connection_gnutls_finalize at /usr/lib64/gio/modules/libgiognutls.so
 #6 soup_io_stream_finalize at /lib64/libsoup-2.4.so.1
 #8 soup_socket_finalize at /lib64/libsoup-2.4.so.1
 #10 soup_connection_disconnect at /lib64/libsoup-2.4.so.1
 #11 soup_session_abort at /lib64/libsoup-2.4.so.1
 #12 soup_session_dispose at /lib64/libsoup-2.4.so.1

(Truncated, but soup_session_dispose is called by g_object_unref is called by __run_exit_handlers.)

I spent some time checking libsoup, WebKit, Epiphany, and even GLib for all uses of atexit and on_exit, but didn't find anything suspicious. So my suspicion falls on SoupNetworkSession::defaultSession(). I guess that will be destroyed in an exit handler because it is static? Maybe we should leak it using NeverDestroyed<SoupNetworkSession> instead? This is kind of wild speculation, but it seems reasonable....

We have another backtrace where __run_exit_handlers calls g_object_unref calls soup_session_dispose (as in the backtrace above), then soup_session_dispose calls g_source_destroy calls g_source_destroy_internal calls g_mutex_lock, then boom. That surely is happening because the mutex is already locked, so it's some thread-safety issue for sure.

Full backtraces downstream (in the See Also field).

-- 
You are receiving this mail because:
You are the assignee for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.webkit.org/pipermail/webkit-unassigned/attachments/20150523/00291f3e/attachment-0001.html>