On Sat, Apr 4, 2026 at 4:30 AM Nazir Bilal Yavuz <[email protected]> wrote: > On Fri, 3 Apr 2026 at 17:59, Sami Imseih <[email protected]> wrote: > > > On Sat, Apr 4, 2026 at 12:38 AM Nazir Bilal Yavuz <[email protected]> > > > wrote: > > > > I also tried Thomas' > > > > "v2-0001-Improve-tar-portability-logic-from-ebba64c0" [3] but it > > > > didn't fix the problem on OpenBSD [4]. > > > > > > Apparently it wants -F ustar, like this. Funny that it passed on the > > > build farm animals though. Oh, it looks like they changed the default > > > fairly recently. > > > > LGTM with just a correction of my earlier comment. > > Thanks for the patches! I confirm that both v3 and v4 fix the problem > for OpenBSD CI.
Pushed, after testing on an OpenBSD VM and making some corrections: * I'd screwed up the test command line in a way that worked by coincidence ** OpenBSD tar writes to a tape device by default, so use -f /dev/null ** I'd forgotten == 0, so the result was inverted, hiding that screwup * -f /dev/null is a better form for all of them because the default destination is a build option * needed elsif instead of if, or BSD tar finished up getting both --format=ustar and -F ustar * ran perltidy, keeping only the hunks due to this patch CI passes and shows "212 subtests passed" for all five Unixen + Windows/mingw, but only "156 subtests passed" for Windows/MSVC. .cirrus.tasks.yml appears to use the same $TAR for both, namely the system tar, so I think we can say that *this* thing is working, but something else might be wrong with our scripting glue somewhere? The other OSes on our list are AIX and Solaris. From a quick look at their manuals, I don't foresee issues with pax or large UIDs. Hopefully that covers everything!
