exiting a terminal causes 100% cpu usage and breaks restarting that terminal with node >= v18.18

[williamstein]

With Node.js v18.18.0 on cocalc-docker (and also cocalc.com) the following steps result in the nodejs hub process getting stuck at 100% making the project almost unusably slow. The following is 100% reproducible on cocalc-docker. It's I think even worse on cocalc.com.

1. Open a terminal in a CoCalc project and split it in half using the split button.
2. Run the top command in one frame.
3. Hit ctrl+d to exit the other frame.
4. The first time, this works fine. Now restart the project, open the same terminal and do 1-3 again. Every single time, the cpu gets pegged at 100%.
5. Every time the terminal attempts to start again it fails immediately with "signal 1" (and exit code 0).

Switch to v18.17.1 and the problem goes away entirely.

I noticed a few things when working on this.

• This is the upstream bug ptys unexpectedly interrupted and exit early in GitHub Actions, Linux with io_uring enabled microsoft/node-pty#630 !
• Here is what is happening in that 100% cpu loop, according to strace:

io_uring_enter(19, 1, 1, IORING_ENTER_GETEVENTS, NULL, 0) = 1
epoll_pwait(17, [{events=EPOLLHUP, data={u32=38, u64=38}}, {events=EPOLLIN, data={u32=40, u64=40}}], 1024, 0, NULL, 8) = 2
read(40, "read(40, \"read(40, \\\"read(40, \\\\"..., 65536) = 272
writev(36, [{iov_base="\201~\1m", iov_len=4}, {iov_base="[0,\"yw5so45vo29\",\"terminal:.2023"..., iov_len=365}], 2) = 369

• I have no clue if this post by Jared Sumner about something similar is relevant.
• Note that v18.18.0 is the next release after 18.17.1, and 18.18.0 was released on Sep 16, 2023.
• I haven't yet reproduced this bug outside of the complicated environment of "running a cocalc project" (so a daemon, etc.).
• It's not clear whether this is a bug in node-pty (perhaps the C extension code) or in nodejs itself.

WORKAROUND: Use node v18.17.1.

[williamstein]

NOTE: I've now pinned version 18.17.1 of nodejs everywhere, e.g., in the latest cocalc-docker image and on cocalc.com, so this issue is mitigated for now.

[louislam]

Thanks for the workaround, it saved my time. I also added your name in the release note, because I literally couldn't solve it without seeing your info here.

For your information, I was experiencing this on Node.js 20.10.0.

Instead of the top command, my project runs docker compose start/restart/down commands.

Not sure if it is related, but they are also "interactive terminals". I don't know the exact name of this.

And also it doesn't seem to happen on Windows' Node.js 20.10.0.