Re: [musl] Feature request: posix_spawnattr_setrlimit_np()

[Adhemerval Zanella Netto <adhemerval.zanella@linaro.org>]

On 15/11/23 21:29, Tavian Barnes wrote:
> On Wed, Nov 15, 2023 at 5:38 PM Rich Felker <dalias@libc.org> wrote:
>>
>> On Wed, Nov 15, 2023 at 05:00:51PM -0500, Tavian Barnes wrote:
>>> I have a program that raises its soft RLIMIT_NOFILE, but wants to
>>> spawn processes with the original value (in case they use select(),
>>> for example).  There seems to be no nice way to do this with
>>> posix_spawn().  I can temporarily lower the rlimit in the parent, but
>>> that interferes with other threads, and can make posix_spawn() fail
>>> with EMFILE.
>>>
>>> Corresponding glibc feature request:
>>> https://sourceware.org/bugzilla/show_bug.cgi?id=31049
>>
>> I don't have any objection to this as long as it's coordinated and
>> there's agreement from other implementors, but there *is* a way to do
>> it already. You posix_spawnp:
>>
>> sh -c 'ulimit -n whatever && exec "$0" "$@"' your_program args...
> 
> True!  Except that ulimit -n is not POSIX.  There are ways around
> that, like making a dedicated shim binary that just does
> 
> setrlimit(RLIMIT_NOFILE, make_rlimit(argv[1]));
> execv(argv[2], argv + 2);
> 
> (or making it a special mode of your_program).  But I think it would
> be better to have a convenient interface for it.  And the
> double-exec() is not free either.
> 
> I'm hoping if some libcs implement this, POSIX will reconsider
> https://www.austingroupbugs.net/view.php?id=603.  Alternatively I
> suppose POSIX could standardize ulimit -n, which is already existing
> practice ~everywhere.

The helper process is what OpenJDK [1] to apply some file actions,
close all file descriptors (similar to close_range), and change the 
working directory.  At least with a recent glibc, all can be accomplish
with posix_spawn.

And I agree that the traditional resource limits seems to be missing 
features from posix_spawn, even when recent Linux interfaces are leaning
to a more fine-grained (and way more complex) interfaces like cgroupv2.

> 
>> This is the general solution to doing all sorts of "child process
>> state setup" things that posix_spawn doesn't have a dedicated
>> attribute for.
>>
>> Note that for a proposal for setting rlimits via an attribute, one
>> complication that needs to be specified is whether the limits take
>> place before or after file actions, since they could change the
>> outcome of file actions. I'm not sure what the answer is, but just
>> YOLO'ing an implementation without thinking about that is a bad idea.
> 
> True!  Actually when I implemented my own posix_spawn()-like
> interface, I had setrlimit() as a file_action.  Then at least the
> order is unambiguous.

I am not sure if the resource limits really maps to a file action,
although it does solve the ordering issue.

For a glibc prototype [2], I added before the file actions because 
caller may define how many file descriptors the new process 
would require (since glibc provides the posix_spawn_file_actions_addclosefrom_np
extension).  However, since this might not be applicable to all libc
implementations I am also not sure which would be best ordering.

[1] https://github.com/openjdk/jdk/blob/master/src/java.base/unix/native/libjava/childproc.c
[2] https://sourceware.org/bugzilla/show_bug.cgi?id=31049