On 8/31/26 13:27, Laurent Bercot wrote:
>> #!/usr/bin/env execlineb
>> background { sleep 3 }
>> # wait on init; clearly not our child
>> wait { 1 }
>
> The short answer is that waiting on a pid that isn't a child of yours
> is a programming error and the wait command has undefined behaviour in
> this case. This is a cop-out, because execline is not C, and UB should
> not happen, but it's genuinely a case I haven't considered. ^^"
>
> The longer answer is that POSIX makes it incredibly difficult
> to be fancy with waiting on processes. As long as you are waiting on
> *one* process, or on *any* process, and that you want a response *now*
> or are willing to *wait for as long as it takes* for a process to exit,
> you can do it without trouble; but anything more complex than that,
> such as waiting for *a set of* processes, or having a *timeout*, becomes
> convoluted, sometimes excessively so.
>
> Add to that the fact that execline's wait command has to also support
> the case where it's called without a block, and follow the specification
> ( https://pubs.opengroup.org/onlinepubs/9799919799/utilities/wait.html )
> in this case, and you end up with way more code that should be necessary
> for something that sounds so simple:
>
> [ If you have not visited skarnet.org's cgit in the last week, hit
> https://git.skarnet.org/cgit once before clicking the following URL,
> else
> you will get caught by my LLM blocker and won't be able to access the
> site again ]
>
> https://git.skarnet.org/cgit/execline/tree/src/execline/wait.c?h=v2.9.9.2
>
> And as you have noticed, even with that much code, the behaviour is
> arguably incorrect. wait should report an error, not "succeed" silently.
>
> I did not code fully functional "waitpids" primitives in skalibs
> because
> I did not think I'd need them: in C code, when you correctly track your
> children, you can get by with the POSIX primitives and do not need
> perfect
> reporting in case of user error. But execline's wait command is an
> exception, and that strictly speaking, the way it is currently
> implemented
> is insufficient. I'll see what I can do for the next release.
>
> --
> Laurent
On Linux, one can use pidfds to do much better. Those allow waiting
for any process and are pollable. If one has CAP_SYS_ADMIN, one can
even bind-mount them to a file, providing a race-free alternative to
PID files.
--
Sincerely,
Demi Marie Obenour (she/her/hers)
Received on Mon Aug 31 2026 - 21:37:15 CEST