Re: [gtkmm] app calls poll() for disconnected connections

[MHL Schulze t-online de (Martin Schulze)]

Am 24.09.2002 21:59 schrieb(en) Eric Buddington:
> stracing my app, I see incessant (to the point of using all CPU time)
> poll() calls that are returning POLLNVAL for file descriptors that
> should no longer be watched. Specifically, my main app forks off
> children and tries to communicate with them via pipes.
>
> In this case, file descriptors 7, 9, and 11 are pipes from short-lived
> children, and 13 and 15 are pipes to long-lived children. Each
> initially is watched by a Connection, created as:
>
> input_connection = Glib::signal_io().connect
>     (SigC::slot(*this, &PluginProxy::on_output),
>      from_child_fd(),
>      Glib::IO_IN | Glib::IO_OUT | Glib::IO_PRI | Glib::IO_ERR |
> Glib::IO_HUP);
>
> All correctly note the "Broken Pipe" error when the child dies, and all
> have disconnect() called on them at that point. After this point, the
> 'on_output' function is no longer called, but the poll() still watches
> the fds and sucks cycles:
>
> poll([{fd=3, events=POLLIN}, {fd=7, events=POLLIN, revents=POLLNVAL},
> {fd=9, events=POLLIN, revents=POLLNVAL}, {fd=11, events=POLLIN,
> revents=POLLNVAL}, {fd=13, events=POLLIN}, {fd=15, events=POLLIN}], 6,
> -1) = 3
>
> Shouldn't GTK stop watching the fd as soon as I call disconnect() on
> the Connection?

I can't find a bug while reading and thinking over the codebase.

SigC::Connection::disconnect() calls
{anonymous}::SourceConnectionNode::notify(false) calls
g_source_destroy(source_) calls
g_source_destroy_internal(source, context, FALSE) calls
g_main_context_remove_poll_unlocked(context, tmp_list->data)

The last function should remove the file descriptors from the poll table.
Obviously it never gets executed - I don't see why.

Could you please file a bug report at bugzilla and provide a simple
testcase? I looks like I had to debug the problem in order to solve it.
You can put my email-adress in the CC field of your bug report.

Regards,

  Martin