Stream: t-compiler/wg-rls-2.0

Topic: Kill proc-macro child process

Edwin Cheng (Apr 19 2020 at 12:11, on Zulip):

I just found that in some weird situations, if rust-analzyer panic / crash accidentally, the child process spawned for proc-macro did not be killed in Windows. I know it is the default behavior for Windows program. I can't reproduce this in linux, but I doubt it might happens there too.

Edwin Cheng (Apr 19 2020 at 12:12, on Zulip):

I know is a proper solution for this problem in Windows. but it will make the code complicated as it is platform specific, and I don't know it is worth or not.

Christopher Durham (Apr 19 2020 at 17:02, on Zulip):

This feels like the kind of thing that should be handled by a crate in the ecosystem

Christopher Durham (Apr 19 2020 at 17:03, on Zulip):

"spawn a child process with automatic cleanup"

Christopher Durham (Apr 19 2020 at 17:07, on Zulip):

If you assume unwinding, storing the handle somewhere and terminating the child on Drop is enough on any platform

Edwin Cheng (Apr 19 2020 at 17:12, on Zulip):

Even I kill the parent process externally ?

Christopher Durham (Apr 19 2020 at 17:17, on Zulip):

here's duct talking about the adjacent issue of grandchild processes

Christopher Durham (Apr 19 2020 at 17:18, on Zulip):

TL;DR of duct's info page is that job objects seem to solve the problem proposed (correct transitive child process cleanup) but *nix doesn't have a good solution yet

Christopher Durham (Apr 19 2020 at 17:19, on Zulip):

And killing the process externally would not be able to assume unwinding

matklad (Apr 20 2020 at 07:22, on Zulip):

I think if we panic, we should kill the child process (in Drop). If we crash abnormaly (like via sigkill) I don't feel like it is worth it to clean up the child. Although this is possible, cargo implements this here:

matklad (Apr 20 2020 at 07:23, on Zulip):

But, I think that we should add an JChild struct to stdx::process, which kills the child on drop.

simulacrum (Apr 20 2020 at 12:14, on Zulip):

@matklad iirc, the job thing on windows is just for ctrl+c as well

bjorn3 (Apr 20 2020 at 12:17, on Zulip):

From the doc comment:

Job objects have a mode of operation where when all handles to the object are closed it causes all child processes associated with the object to be terminated immediately.

I think this means that any way of exiting the process will kill all child processes.

Last update: Sep 30 2020 at 17:00UTC