Re: [Dbus-cxx-devel] Race hack fix

[Patrick A. <pa...@ph...>]

On Thu, 2010-03-18 at 10:02 -0600, Rick L. Vinyard, Jr. wrote:
> However, I did noticed that in the newest release of dbus (1.2.22) there
> is a fix for a race condition in the watch handlers.

Sadly, it's the wrong fix. There are two threads, of course, and so the
bug happens in two ways:

1) watch thread calls watch_handle while another thread has I/O path
2) another thread calls dbus_connection_send while watch thread has I/O
path

That patch fixes case #1. Case #2 is still unhandled.

> I keep hoping to "acquire" a test case by stumbling across it, but I
> haven't been bitten by the bug yet.

I can try to write one up soon. The best way to have it happen is to
have D-Bus method A emit D-Bus signal B, and then call method A a
bajillion times. Needs to be run on an SMP machine, obviously.

Still takes about 5-10 minutes, though.

Patrick