Hacker News Reader

Oh, yes. I've been looking forward to this since I saw https://www.youtube.com/watch?v=KXuZi9aeGTw (User-level threads....... with threads. - Paul Turner - Google) in 2013.

The way it is implemented in on top of the futex machinery is very elegant.

Unfortunately it still need to go through the kernel, so it is probably going to be orders of magnitude slower than pure userspace based task switching, especially after spectre.

And yet it is an order of magnitude faster than using the kernel thread scheduler, and you get to keep working thread-locals, normal process ID behavior, etc. It's a compromise.

The best solution is to standardize tooling around backtraces and switch in userspace. Kernel switchto() has its place, but only between processes. It’s much, much slower than a userspace function call.

Great! I remember reading about this but I thought it'd never make it into the kernel. What blocked its inclusion last time?

Woah, that's really interesting. I was wondering if something like this had ever been explored.