search

raforg / daemon

turns other processes into daemons
http://libslack.org/daemon/
GNU General Public License v2.0
14 stars 2 forks source link

Sometimes creates zombies. #9

Closed shamefulCake1 closed 1 month ago

shamefulCake1 commented 2 months ago

I have seen this when the target service fails to start from the beginning.

The process tree would look like:

init->daemon(root)->daemon(zombie)

And that is it. Parent daemon never ends up successfully waitpid' ing the child.

raforg commented 1 month ago

Thanks. I've always thought this would be possible but it has never been reported in 25 years. There is a place between two lines of code at the top of the event loop where a signal would be lost if it arrived (and there's a comment there to that effect). But it's a very small time window (between handling signals and starting select).

Are you able to give me specific instructions to reproduce it? What exactly do you mean by "fails to start". Presumably, it must start in order to become a process that can become a zombie. Does it start and exit quickly? It would really help if you could provide a complete command and the resulting process details (like ps or pstree output perhaps).

If it's what I think it is, I can fix it by giving the event loop a timeout to give it a chance to respond to a SIGCHLD. I haven't done that because it's extra cpu cycles that would mostly be a waste. But if I can test it, and the fix works, then I'll know I'm right. Actually, I should use pselect which is standard.

But I really need to be able to reproduce this so I'm not guessing what the reason is. The fact that you are showing two daemon processes (rather than daemon and the client process) makes me wonder if it is something else.

shamefulCake1 commented 1 month ago

1.

What exactly do you mean by "fails to start". Presumably, it must start in order to become a process that can become a zombie. Does it start and exit quickly?

When command is a non-existent path and a few arguments.

2. When I had created this issue I didn't yet realize that the issue might have been due to the sanitizer rather than the program itself.

So, I compiled daemon with leak sanitizer, which somehow (I don't think using a thread), spawns llvm-symbolizer as a child process, and in my case the parent daemon had a zombie daemon child, and the llvm-symbolizer child. Presumably llvm-symbolizer wouldn't exit until daemon ripes all its children, and daemon wouldn't ripe the child daemon until it is the only remaining child, or something like that.

But maybe this is an issue with llvm-symbolizer, and not with daemon, so maybe this issue should be just closed.

raforg commented 1 month ago

OK. Thanks.