On 19/10/2025 23:57, Collin Funk wrote:
Pádraig Brady <[email protected]> writes:
That flag is a no-op
since:https://sourceware.org/git/?p=glibc.git;a=commitdiff;h=ccfb2964and only
there to allow old code to compile.
That change was after the clone implementation. It looks like before
then the POSIX_SPAWN_USEVFORK implementation actually used vfork which
would likewise use clone with CLONE_VM.
The default settings may be just fine these days,
though I've not looked into the details.
The default works well, that is what I used in my original message.
Gnulib's version uses vfork when POSIX_SPAWN_USEVFORK is used, which
might help other systems, though.
Oh right. I suppose the attr management isn't too onerous.
I was wondering if POSIX_SPAWN_USEVFORK was always defined,
but it seems so since gnulib's spawn-h module says:
"<spawn.h> is always overridden, because of GNULIB_POSIXCHECK".
I notice that glibc's posix_spawn() reaps the child itself upon exec error
(which I confirmed with `strace -ff`) but means we only get a single error
now from the following:
# Old behavior
$ install src/install.c foo -s --strip-program=./foo
install: cannot run './foo': Permission denied
install: strip process terminated abnormally
# New behavior
$ src/ginstall src/install.c foo -s --strip-program=./foo
ginstall: cannot run './foo': Permission denied
That's fine though, and probably a slight improvement.
I'd add something in NEWS under "Improvements". Perhaps:
'install' now uses posix_spawn() to invoke the strip program more efficiently.
We can augment that message for sort(1) and split(1) if we make the change
there too.
Otherwise the patch looks good.
thanks!
Padraig
