Re-enable t9128, t9141, and t9167 which were disabled due to these
random segfaults. The bug has also been reported against the subversion
perl bindings in Debian¹. Hopefully this will reach upstream subversion
if a fix is made to subversion.
¹ https://bugs.debian.org/888791
Move contrib/hooks/multimail from git-core to git and use python3 rather
than python2.
We still use python2 as the default PYTHON_PATH because not all of the
python scripts in git support python3. None of the other scripts are
included in git-core though.
Primarily, python2 is used by git-p4 and contrib/svn-fe/svnrdump_sim.py
(which is used by t9020-remote-svn.sh). Converting git-p4 to python3 is
not a trivial matter of fixing a few print statements. A simple
conversion using python3's 2to3 tool results in a large number of test
failures.
Add a python3-devel BuildRequires for %{__python3} and add python2-devel
to the tests section since t9020-remote-svn.sh uses python2. (We
already BR python2-devel in git-p4, but having it in the tests section
ensures we don't remove it if/when git-p4 supports python3.)
This release fixes an issue which only affects users on case-insensitive
file systems and repositories which contain paths that differ only in
case. Such circumstances result in a segmentation fault in various git
commands.
This test was passing as recently as last week. Something seems to have
changed or broken in the x86 arch of f28 since then. Disable the test
until the issue is determined and resolved.
With 'prove' as the test harness the tests can be run in parallel on
EPEL as well as Fedora targets.
Move GIT_TEST_OPTS to config.mak along with the new test options and
enable shell tracing (-x). The output from failures when tracing is
enabled should allow us to more easily diagnose test failures.
Explicitly use /bin/bash as the shell for the test suite; it allows
using "-x" reliably across the whole test suite. This is made possible
by changes included in 2.16.0 thanks to Jeff King¹, particularly:
3f824e91c8 t/Makefile: introduce TEST_SHELL_PATH
f5ba2de6bc test-lib: make "-x" work with "--verbose-log"
90c8a1db9d test-lib: silence "-x" cleanup under bash
¹ https://github.com/gitster/git/tree/jk/test-suite-tracing
The paths don't change often, but if they do this is one less place
which will need to be updated. It also makes it easier for anyone
rebuilding the packages to change the defaults.
Use a single macro rather than testing for %{fedora} and %{rhel}
versions repeatedly. The "dist" logic is best kept at the start of the
spec file (as much as possible).
Cleanup a few comments related to "dist" checks while we're at it.
The acl and perl(HTTP::Date) requirements are pulled in by other
packages on Fedora so they were not noticed when the test requirements
were updated in 6dc6285 ("Improve test suite coverage", 2017-11-10).
The highlight package is available in EPEL for EL-6 (on all supported
architectures). Extend the conditional to only exclude it on ppc64 for
EL-7.
The gnome-keyring credential helper was removed in Fedora 26, but
remained in builds for older Fedora and EL-7 releases. No official
release has shipped the gnome-keyring subpackage since 2.11.1-3.
Obsolete it to ensure a clean upgrade path for any users who have it
installed. The Obsoletes can be removed after rawhide is branched for
Fedora 29.
Starting with rpm-build-4.13.0 (in F-22), brp-python-bytecompile
excludes files under /usr/share/doc by default. Restrict the %files
%exclude to EL <= 7 to avoid spurious 'File not found' warnings from
rpm-build.
Ensure these tests are not skipped if some dependency or other condition
is not met. We prefer to see the failure(s) in order to fix the
underlying issue(s).
Disable t9115-git-svn-dcommit-funky-renames on ppc-arches because it
frequently fails. The port it uses (9115) is already in use. It is unclear if
this is due to an issue in the test suite or a conflict with some other
process on the build host.
Add mod_dav_svn BuildRequires for t9115-git-svn-dcommit-funky-renames
(GIT_SVN_TEST_HTTPD).
When building release candidates the path to the perl requires filter
script is wrong. Correct it.
It's tempting to use the output of `pwd` to avoid future changes to the
build dir breaking it, but the call occurs prior to unpacking and thus
`pwd` is not suitable.
The tests appear to work reliably now, after a number of test builds. If
not, this is easily reverted. (The newly added test output may also
help determine the root cause.)
All tests which call 'git svn branch' fail intermittently with SIGSEGV
in the subversion bindings. The failures are not remotely consistent.
Until we can enable shell tracing (-x) in the test suite by default to
attempt to further debug the failure, disable these tests.
An option to enable parallel test runs was added in 6dc6285 ("Improve
test suite coverage", 2017-11-10). After further improvements it seems
to be reliable for use in koji builds.
It is still disabled for EPEL builds because make lacks support for the
-O (--output-sync) option. Without this option the output is much less
readable.
In the future we may be able to move to using 'prove' as the default
test harness which might allow us to run in parallel without requiring
the -O option.
Setting SVNSERVE_PORT enables several tests which require a local
svnserve daemon to be run (in t9113 & t9126). The tests share setup of
the local svnserve via `start_svnserve()`. The function uses svnserve's
`--listen-once` option, which causes svnserve to accept one connection
on the port, serve it, and exit. When running the tests in parallel
this fails if one test tries to start svnserve while the other is still
running.
Use the test number as the svnserve port (similar to httpd tests) to
avoid port conflicts. Set GIT_TEST_SVNSERVE to enable these tests.
The change should make it into 2.16 or 2.17. Until then, setting
GIT_TEST_SVNSERVE will disable the svnserve tests. There are only a few
tests affected and this allows more testing with parallel test runs, so
it's a reasonable trade. Once an upstream release with these changes
arrives, we'll begin running these tests without any further changes.
In c3202fd ("Use prebuilt documentation on EL-5, where asciidoc is too
old", 2013-01-04) the make targets to build and install the
documentation were split into separate make calls to handle prebuilt
docs on EL-5 where asciidoc was too old to build the documentation.
Support for EL-5 was removed in 903d8f3 ("Remove EL-5 and old Fedora
conditionals", 2017-07-22), making the separate make calls unneeded.
Using %autosetup frees us from having to manage %patch entries. It does
require that all patches use the same strip level (-p1 in our case).
There is also a mildly annoying bug in EL-6¹ which chokes on the patch
invocation if a blank line doesn't follow %autosetup. A comment should
prevent any problems.
The %autosetup -S option is deliberately avoided, to prevent needless
bootstrapping problems.
¹ https://bugzilla.redhat.com/1310704
When building in koji or copr and a test fails, the verbose output from
the tests can be very useful in determining the cause. Print the output
from failing tests, borrowing heavily from the upstream code¹ used when
running tests in Travis CI².
This could be done inline in %check rather than in a separate script.
The downside is that rpm would include each command invocation of the
loop in the output, which adds a lot of useless text to the already
copious build output.
¹ https://git.kernel.org/pub/scm/git/git.git/plain/ci/print-test-failures.sh
² https://travis-ci.org/git/git
We used to remove the contrib/credential tree from the build dir to
avoid binaries and other cruft from ending up in the doc dir. Doing so
prevented debuginfo from being generated for the credential helpers we
install. The following warnings were printed during the debuginfo
extraction:
cpio: git-2.15.1/contrib/credential/gnome-keyring: Cannot stat: No such file or directory
cpio: git-2.15.1/contrib/credential/gnome-keyring/git-credential-gnome-keyring.c: Cannot stat: No such file or directory
cpio: git-2.15.1/contrib/credential/libsecret: Cannot stat: No such file or directory
cpio: git-2.15.1/contrib/credential/libsecret/git-credential-libsecret.c: Cannot stat: No such file or directory
Keep contrib/credential in the build dir; remove it from the buildroot
instead.
Building release candidates helps test the upstream code before it's
final. Make it easier to create such builds by defining an rcrev macro,
which is commented out by default.
The git-gui build tries to use tclsh to generate an index file. It
falls back to a static list, which is why this wasn't noticed before.
It's not a real problem, but git may soon also try to test that wish is
installed before building and installing git-gui and gitk, so this helps
prepare for that future.
The code to filter the spurious YAML::Any requires was removed in
903d8f3 (Remove EL-5 and old Fedora conditionals, 2017-07-22). The
comment was simply missed.
In older versions of git, the bash-completion script began with a
needless #!bash, which we removed before installation.
This and other unneeded shebangs were removed upstream in 1.9.0,
11d62145b9 (remove #!interpreter line from shell libraries, 2013-11-25).
3041eeb (Run git test suite, 2016-12-16) enabled the test suite, but
many of the tests were skipped due to missing requirements. Improve the
test suite coverage by ensuring as many requirements are present as
possible.
Also add an option to run the test suite with multiple jobs using the
%{?_smp_mflags} macro. This cut the test suite run time roughly in
half in local testing. The -O option to 'make' helps ensure the output
of each job is collected together rather than interspersed with output
from other jobs. Sadly, the tests fail when run in parallel in koji, so
this is disabled by default. It is included to speed up local builds.
To enable in rpmbuild and mock, use either "--with parallel_tests" or
"--define 'parallel_tests 1'" arguments. This is only supported on
Fedora, as it requires make >= 4.0, which neither EL6 nor EL7 satisfy.
Disable 2 failing tests which run with a limited stack on aarch64, arm*,
and ppc* architectures. Some debugging with the affected hardware is
needed.
Similarly, disable "git svn branch tests" globally for both
t9128-git-svn-cmd-branch and t9167-git-svn-cmd-branch-subproject because
they fail randomly. The bug seems likely to be in the subversion perl
bindings, but requires more testing. The failure output from a manual
run (./t9128-git-svn-cmd-branch.sh -v -x -d):
++ git svn branch a
Copying file:///home/tmz/rpmbuild/BUILD/git-2.14.2/t/trash%20directory.t9128-git-svn-cmd-branch/svnrepo/trunk at r2 to file:///home/tmz/rpmbuild/BUILD/git-2.14.2/t/trash%20directory.t9128-git-svn-cmd-branch/svnrepo/branches/a...
Found possible branch point: file:///home/tmz/rpmbuild/BUILD/git-2.14.2/t/trash%20directory.t9128-git-svn-cmd-branch/svnrepo/trunk => file:///home/tmz/rpmbuild/BUILD/git-2.14.2/t/trash%20directory.t9128-git-svn-cmd-branch/svnrepo/branches/a, 2
Found branch parent: (refs/remotes/origin/a) bb43a88ce94511096d7d774f4d5feae281603eb0
Following parent with do_switch
Successfully followed parent
r3 = 2e27ee45dadc098239114bc13a6ae6cf6122d16f (refs/remotes/origin/a)
error: git-svn died of signal 11
+ test_eval_ret_=139
+ want_trace
+ test t = t
+ test t = t
+ set +x
error: last command exited with $?=139
not ok 3 - git svn branch tests
The failure in t9167-git-svn-cmd-branch-subproject is similar.
A change to the git filter-branch command in git-2.15 added a perl
dependency, bd2c79fbfe (filter-branch: stash away ref map in a branch,
2017-09-21)¹.
This dependency is only present when using the option --state-branch,
which means many users of filter-branch will never notice the perl
dependency.
However, without moving the tool from git-core to git, any users running
filter-branch with the --state-branch option might find that the command
fails badly, potentially leaving their repository in a broken state.
This is not an acceptable trade-off, in my opinion.
I'm not sure if it's worthwhile to suggest a patch upstream to better
handle users running with the --state-branch option when perl is not
installed, allowing us to safely ship git filter-branch in git-core for
use on systems without perl.
¹ https://github.com/git/git/commit/bd2c79fbfe
Commands which required perl were moved from git-core to git in 6fdc5e5
(new subpackage git-core, which is perl-less and part of original git
rpm, 2015-06-03). This was with git-2.4.2. Since then, a number of
commands have been partially or entirely rewritten upstream and no
longer require perl. Move them back to git-core.
- git-am was converted to a builtin in 2.6.0, 783d7e865e (builtin-am:
remove redirection to git-am.sh, 2015-08-04)
- git-submodule's dependence on perl was removed in 2.10.0, 74703a1e4d
(submodule: rewrite `module_list` shell function in C, 2015-09-02)
- git-relink was removed in 2.12.0, ed21e30fef (relink: retire the
command, 2017-01-25)
Additionally, two sample hooks (pre-rebase and prepare-commit-msg) were
intended to be excluded from git-core due to their perl dependencies,
but were included due to the way the exclusion via %files lists are
created. List them explicitly in the git %files section and exclude
them from git-core.
Using the same variable make it clearer that we're referring to the same
thing as upstream when the Makefile uses gitexecdir.
Remove a stale conditional in git-cvs for handling the case where git's
binaries are stored in %{_bindir}. This support was dropped after
EL-5's support ended. Most similar conditionals were removed in 903d8f3
(Remove EL-5 and old Fedora conditionals, 2017-07-22).
Also remove the conditional setting of 'gitexecdir = %{_bindir}' which
was only used on EL-5 to prevent moving the git binaries during the
update to git-1.6.
In ef7ac1a (git-p4: adjust python2 shebang manually, 2017-08-03) we
dropped PYTHON_PATH due to failures in t9020-remote-svn.
Digging deeper, the issue appears to be caused by the test calling a
script which has a hard-coded #!/usr/bin/python shebang. On newer
fedora releases, /usr/bin/python is python3.
Replacing the shebang in contrib/svn-fe/svnrdump_sim.py (the script the
test calls) with #!%{__python2} allows the test to succeed while setting
PYTHON_PATH in config.mak, as desired.
Use '%{summary}.' as the %description for all subpackages where the
summary and description matched. This is one less place to change
if/when we update a summary/description in the future.
While looking at rpmlint complaints, gitk had a spelling warning because
the summary/description used the British English spelling of
"visualiser." Rather than simply change that to the American English
spelling, change the summary/description to match the gitk
documentation. Do the same for git-gui.
Using %summary fixes the %description for perl-Git-SVN, which previously
used the %description from perl-Git mistakenly.
Also improve/clarify the summary of git-gnome-keyring, git-send-email,
and git-svn.
The git-all subpackage added an obsoletes for older git packages in
0c4a11c (git 1.5.4.3 and include krh's change to rename git-core to git,
2008-02-23).
The git-arch subpackage was obsoleted many years ago in 3f997b4
(Obsolete git-arch as needed, 2011-08-05).
We don't need to carry either of these obsoletes any longer.
We cannot be sure that users installing the git package have not mounted
the git bin and lib dirs on different file systems. The cost is a small
amount of extra disk space (1.25M in the current build).
From the release announcement¹
A malicious third-party can give a crafted "ssh://..." URL to an
unsuspecting victim, and an attempt to visit the URL can result in
any program that exists on the victim's machine being executed.
Such a URL could be placed in the .gitmodules file of a malicious
project, and an unsuspecting victim could be tricked into running
"git clone --recurse-submodules" to trigger the vulnerability.
Credits to find and fix the issue go to Brian Neel at GitLab, Joern
Schneeweisz of Recurity Labs and Jeff King at GitHub.
¹ https://public-inbox.org/git/xmqqh8xf482j.fsf@gitster.mtv.corp.google.com/
