Re: [LHF] making t5000 "tar xf" tests more lenient

Am 24.12.2012 21:49, schrieb Junio C Hamano:

I've been running testsuite on a few platforms that are unfamiliar
to me, and was bitten by BSD implementation of tar that do not grok
the extended pax headers.  I've already fixed one in t9502 [*1*]
where we produce a tarball with "git archive" and then try to
validate it with the platform implementation of "tar" so that the
test won't barf when it sees the extended pax header.

t5000 is littered with similar tests that break unnecessarily with
BSD implementations.  I think what it wants to test are:

  - "archive" produces tarball, and "tar" can extract from it.  If
    your "tar" does not understand pax header, you may get it as an
    extra file that shouldn't be there, but that should not cause the
    test to fail---in real life, people without a pax-aware "tar"
    will just ignore and remove the header and can go on.

  - "get-tar-commmit-id" can inspect a tarball produced by "archive"
    to recover the object name of the commit even on a platform
    without a pax-aware "tar".

Perhaps t5000 can be restructured like so:

  - create a tarball with the commit-id and test with
    "get-tar-commit-id" to validate it; also create a tarball out of
    a tree to make sure it does not have commit-id and check with
    "get-tar-commit-id".  Do this on any and all platform, even on
    the ones without a pax-aware "tar".

  - check platform implementation of "tar" early to see if extracting
    a simple output from "git archive" results in an extra pax header
    file.  If so, remember this fact and produce any and all tarballs
    used in the remainder of the test by forcing ^{tree}.

so that people on platforms without pax-aware "tar" do not have to
install GNU tar only to pass this test.

It would be a good exercise during the holiday week for somebody on
BSD (it seems NetBSD is more troublesome than OpenBSD) to come up
with a patch to help users on these platforms.

I got around to building a virtual machine with NetBSD 6.0.1, which
involved a bit of cursing because networking only seems to work when I
turn off ACPI and SMP -- and running tests is a lot more pleasant with
more than one CPU.

Anyway, I don't think the pax headers are to blame here.  The patch below
fixes the tar failures for me, but I'm not sure why.  There must be
something special about some (not all!) directory entries with trailing
slashes after directory names.

Several ZIP tests still fail, however, because NetBSD's unzip doesn't
seem to support (our flavour of) symlinks and streamed files.

I'll take a deeper look into it over the weekend.

René


---
 archive-tar.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/archive-tar.c b/archive-tar.c
index 0ba3f25..fd2d3e8 100644
--- a/archive-tar.c
+++ b/archive-tar.c

@@ -215,6 +215,8 @@ static int write_tar_entry(struct archiver_args *args,
 	if (S_ISDIR(mode) || S_ISGITLINK(mode)) {
 		*header.typeflag = TYPEFLAG_DIR;
 		mode = (mode | 0777) & ~tar_umask;
+		if (pathlen && path[pathlen - 1] == '/')
+			pathlen--;
 	} else if (S_ISLNK(mode)) {
 		*header.typeflag = TYPEFLAG_LNK;
 		mode |= 0777;

-- 
1.7.12