Hi Lars, Hi Andrew, I test it now in the environment that the problem reproduces with a patch of Mr. Lars. * The patch of msgfromIPC_ll does not apply it. * The patch of crm_trigger_prepare applies it.
The problem does not reappear on the test of several days for the moment. I carry out a test a little more and finally report a result. And I intend to carry out the same test with a patch of Mr. Andrew afterwards. * https://github.com/beekhof/pacemaker/commit/2a6b296 Best Regards, Hideo Yamauchi. --- On Fri, 2012/1/20, Lars Ellenberg <lars.ellenb...@linbit.com> wrote: > On Fri, Jan 20, 2012 at 09:21:58AM +1100, Angus Salkeld wrote: > > On 19/01/12 22:23 +0100, Lars Ellenberg wrote: > > >On Tue, Jan 17, 2012 at 12:13:37AM +0100, Lars Ellenberg wrote: > > >>On Tue, Jan 17, 2012 at 09:52:35AM +1100, Andrew Beekhof wrote: > > >>> > > >>> Ok, done: > > >>> > > >>> https://github.com/beekhof/pacemaker/commit/2a6b296 > > >>> > > >>> If I'm adding voodoo, I at least want the reason well documented so it > > >>> can be removed again if the reason goes away. > > >> > > >>That about sums it up, then ;-) > > > > > >But as having to do this was just "too ugly to be true", > > >I dug a little deeper... > > > > > >The way to do this is obviously to use the glib api ;-) > > >http://developer.gnome.org/glib/2.30/glib-UNIX-specific-utilities-and-integration.html#g-unix-signal-add-full > > > > > >(Since glib 2.30, yay; if you don't have that yet, read on anyways) > > > > > >What it does internally, and what other people have also done for a long > > >time to solve this and similar problems, is: > > > > > >Add to the main context a "wakeup pipe", > > >which is an eventfd if available, > > >or an actual pipe if not. > > >If it is a pipe, set those file descriptors non-blocking. > > >And, of course, add the evenfd (or the read end of the pipe) > > >to the poll loop (with default priority, btw, > > >which is good enough to have the poll terminate). > > > > > >That is done internally when creating the main context. > > >http://git.gnome.org/browse/glib/tree/glib/gmain.c#n548 > > >http://git.gnome.org/browse/glib/tree/glib/gwakeup.c#n138 > > > > > >(the line numbers are correct for glib master as of today, > > >which should correspond to 41fbf42) > > > > > >The g_unix_signal_handler then sets the triggers variables, > > >and calls g_wakeup_signal(that internal wakeup source), > > >which simply posts and event to the eventfd, > > >or does a short (1 byte) write to the write end of the pipe. > > >http://git.gnome.org/browse/glib/tree/glib/gmain.c#n4442 > > >http://git.gnome.org/browse/glib/tree/glib/gwakeup.c#n230 > > > > > >Problem solved, without having to do a full check() everything, > > >prepare() everything, and poll() again cycle every 500ms. > > > > > >"back in those days", when this mechanism was not really there, > > >you could do all that "by hand". > > >And people did. Very common idiom in glib and other mainloop > > >applications, also frequently used to "signal" availability of work > > >or completion of tasks between threads. > > > > > >static int my_wakeup_fds[2] = { -1, -1 }; > > > > > >Then just pipe2(my_wakeup_fds, O_NONBLOCK), add my_wakeup_fds[0] as > > >normal read fd source, and add a write(my_wakeup_fds[1], "", 1); to the > > >signal handlers. > > > > signalfd makes this much easier too "man 2 signalfd" > > See https://bugzilla.gnome.org/show_bug.cgi?id=652072#c32 > following (or the whole bug, if you like). > > Also, pipes are portable. > > Lars > > > _______________________________________________ > Pacemaker mailing list: Pacemaker@oss.clusterlabs.org > http://oss.clusterlabs.org/mailman/listinfo/pacemaker > > Project Home: http://www.clusterlabs.org > Getting started: http://www.clusterlabs.org/doc/Cluster_from_Scratch.pdf > Bugs: http://bugs.clusterlabs.org > _______________________________________________ Pacemaker mailing list: Pacemaker@oss.clusterlabs.org http://oss.clusterlabs.org/mailman/listinfo/pacemaker Project Home: http://www.clusterlabs.org Getting started: http://www.clusterlabs.org/doc/Cluster_from_Scratch.pdf Bugs: http://bugs.clusterlabs.org