The shell that slept 100 milliseconds at a time

Twenty invocations of /bin/true, timed:

# huck, 8 July
$ time huck -c 'i=0; while [ $i -lt 20 ]; do /bin/true; i=$((i+1)); done'
2.02 s
 
# bash 5.2, same loop
0.10 s

Two seconds to run twenty commands that do nothing. Roughly 100 milliseconds each, almost exactly — and almost exactly is the tell. Real work is never that punctual. That's a sleep.

It was. The code that waits for a foreground command to finish was polling: check whether the child has exited, sleep 100 ms, check again. For a long-running command nobody notices. For /bin/true — or for git rev-parse in a prompt, which is the one that would have made huck feel broken — you pay the full tick every single time. The fix is to stop polling and block on the child until the kernel says it's done.

# huck, 20 July
$ time huck -c 'i=0; while [ $i -lt 20 ]; do /bin/true; i=$((i+1)); done'
0.10 s

Why it hid for so long

Nobody noticed because it never failed. Every test passed; a handful in the interactive suite occasionally hit their timeout and got mentally filed under "flaky PTY tests", which is a phrase that should always be treated as a confession rather than a diagnosis. When those five timeouts were finally looked at together, they weren't five problems. They were two — and the big one was this: not a bug in any feature, but a tax on every feature at once.

The second one was a genuine hang, in the same neighbourhood:

# huck, 8 July
$ huck -c 'exec 3< /etc/hostname; exec 4<&3-; read line <&4; echo "got=$line"'
huck: line 1: bad fd: 3-
huck: line 1: 4: Bad file descriptor
got=
 
# huck, 20 July
got=devbox

4<&3- is the move form of a redirect: duplicate fd 3 onto fd 4 and close the original. huck didn't implement the trailing -, so it rejected the redirect, dropped it, and carried on — which left a later read waiting on the terminal for input that was never coming. A test that hangs forever because a redirect was silently skipped is a good argument for a rule huck now follows: an unsupported redirect is an error that stops the command, not a warning you step over.

The thread this belongs to

Both fixes came out of a week spent on file-descriptor plumbing — how redirects, pipelines, background jobs and captured output all decide where fd 0, 1 and 2 point. That week fixed a lot of individual bugs, and left behind a suspicion that the pattern of the bugs mattered more than any of them: each fix landed in a slightly different code path doing slightly the same thing. Which is a story for August.