Hacker News Reader

Comment by bandrami | Hacker News Reader

It's more like the concept of Component Object Model failed. For example OLE (the base technology behind DCOM) doesn't fly beside the legacy Office usage.

True, but the idea of the self-registering COM object eventually seemed to limp across the finish line.

Anyways. "Two key kdbus developers" (ahem) have said that the entire kernel signal mechanism should be deprecated because it's "too brittle and complex". When I read that, a big light bulb turned on in my head about what the actual problem is here.

View on HN

Signals are too brittle and complex. That's why you have to read three whole manpages to figure out what happens if a process gets the same signal twice in rapid succession.

They are also un-Unixlike: they are used to communicate three or four different kinds of information, and they do most of them badly.

Exactly. Signal mechanism makes sense to notify synchronous errors that arise from the thread's own execution, like SIGSEGV or SIGILL or SIGFPE.

Most of the rest of the traditional UNIX signals are events that should be communicated asynchronously via file descriptors that a process can poll at its leisure, which would be more UNIX-y.

Well at least Linux has signalfd(2) now.

Except they're not wrong. We have a potentially infinite set of signals and data we'd like to communicate to programs and a very limited set of signals to do it with. I mean apache uses sigwinch as an exit code!

We have a potentially infinite set of signals and data we'd like to communicate to programs

Ah, I think I've identified your problem.

Signals are for alerting programs to a specific (small) set of changes of external states. There's nothing "potentially infinite" there (frankly having 2 different SIGUSR's is generous).

For transmitting arbitrary messages, Linux provides message queues, which are very nearly sockets (and the ways they aren't sockets make using them for IPC easier). What's the limitation there?

It lacks multicast, sure, but then receivers can just make their own queues and have the broadcaster send to them if you want that.

You answered it yourself. It lacks efficient multicast.

What dbus really needs is an efficient kernel-level capability-aware multicast IPC mechanism, and a total rewrite of the user-space daemon to not be rubbish.

All the user-space daemon needs to do is provide a registration point for other processes to find and connect to each other via (existing) unicast or (new) multicast mechanisms.

My initial thoughts were that kdbus /was/ that IPC mechanism, but it sounds like that might not actually be true.

Reply on news.ycombinator.com