So, reviewing how far I got today, I realise that I didn't make it as far as "remove threads from the session on exit". We need some way of notifying the thread that created the exiting thread that its child exited, then it needs to clean up. Briefly, then
- Remove CLONE_PARENT from clone() flags. We don't need it now we're not doing stop-for-GC with ptrace(). Add instead an RT signal called, say, SIG_THREAD_EXIT, so that our thread death signals queue up when lots of children exit at once.
- Add a C handler for thread exit, which does normal pseudo-atomic stalling then calls a Lisp handler.
- Write the Lisp handler to do appropriate surgery on SESSION and foreign-call destroy_thread.
Since we're here, we may as well make the new handlers call-behind. Since we're doing that, we may as well add a general Lisp interface for establishing call-behind handlers, and have a look at the existing handlers to see what could be converted to the new style
Many of the existing handlers just drop into the debugger with an appropriate message, but they expect the sigcontext and stuff to be on the stack so they can dig into it for PC and so on. This is not the case for a call-behind handler: Something Must Be Done. I'm quite tempted to make them all call some internal interface and repurpose the existing enable-interrupt to do call-behind. User code won't (shouldn't) notice the difference, after all.
Removing CLONE_PARENT also means that (a) the parent thread has getting on for nothing at all to do whatsoever, and it might even be possible to get rid of it completely, and (b) that the system might get a bit easier to port to BSD. But apparently FreeBSD doesn't have POSIX RT signals, so still not actually possible.