[musl] glob(GLOB_BRACE)

[musl] glob(GLOB_BRACE)

[Andrew Cagney]

FYI, It's on the BSDs and glibc.
Found by libreswan's test framework on alpine.

Re: [musl] glob(GLOB_BRACE)

[Rich Felker]

On Thu, Feb 01, 2024 at 08:48:01AM -0500, Andrew Cagney wrote:
> FYI, It's on the BSDs and glibc.
> Found by libreswan's test framework on alpine.

By the usual criteria, this one probably leans towards a no, at least
contingent on how constraining it is to support it.

If it happened at the fnmatch layer, it would very much be a no,
because that effectively imposes regex-equivalent (or at least very
close) matching algorithm complexity and is thereby impossible to do
efficiently in constant-space, vs the standard fnmatch language where
that is possible (and we do it).

For glob and GLOB_BRACE, it looks like it's "just" a matter of
iterating over the different expansions and applying glob to each, but
it seems to be underspecified how it interacts with escaping, special
chars, slashes, etc.

Do you have good reasons in favor of inclusion? My impression is that
everything that wants/needs it is shipping its own version of GNU glob
or whatever that has it, or else very little is using it; otherwise it
would have come up before. But if adding it allowed a lot of things to
drop GNU glob and just use the libc glob, that might be compelling.

Rich

Re: [musl] glob(GLOB_BRACE)

[Andrew Cagney]

On Thu, 1 Feb 2024 at 09:27, Rich Felker <dalias@libc.org> wrote:
>
> On Thu, Feb 01, 2024 at 08:48:01AM -0500, Andrew Cagney wrote:
> > FYI, It's on the BSDs and glibc.
> > Found by libreswan's test framework on alpine.

> For glob and GLOB_BRACE, it looks like it's "just" a matter of
> iterating over the different expansions and applying glob to each, but
> it seems to be underspecified how it interacts with escaping, special
> chars, slashes, etc.
>
> Do you have good reasons in favor of inclusion? My impression is that
> everything that wants/needs it is shipping its own version of GNU glob
> or whatever that has it, or else very little is using it; otherwise it
> would have come up before. But if adding it allowed a lot of things to
> drop GNU glob and just use the libc glob, that might be compelling.

for libreswan, after some digging:

2007 new code added with glob(GLOB_BRACE) call
2014  #ifdef GLOB_BRACE added; it turns out due to musl
2024  while merging some glob() calls; problem re-discovered

So libreswan took the path of least resistance.  I suspect it wasn't alone.

Including it would let libreswan's config files be consistent across
platforms - no need for special documentation.

However, I also wonder if libreswan needs GLOB_BRACE.  It isn't needed
to expand the recommended <<include /etc/ipsec.d/*.conf>>.  I'll take
that up with libreswan.