Hi,

this small patch series fixes another instance of broken TAP output that has landed via 4d11b9c218 (Merge branch 'pt/fsmonitor-linux', 2026-05-31).

As this has happened multiple times by now I decided to have a look at whether we can fix this class of issues a bit more holistically. So this series also contains a change that makes prove bail out when it sees invalid TAP output, which uncovers a small set of preexisting issues in our test suite.

Thanks!

Patrick

---
Patrick Steinhardt (4):
      t7527: fix broken TAP output
      t/test-lib: silence EBUSY errors on Windows during test cleanup
      t/lib-git-p4: silence output when killing p4d and its watchdog
      t: let prove fail when parsing invalid TAP output

 t/lib-git-p4.sh              |  3 ++-
 t/t7527-builtin-fsmonitor.sh |  7 ++++---
 t/test-lib.sh                | 10 ++++++++--
 3 files changed, 14 insertions(+), 6 deletions(-)

--- base-commit: 1666c1265231b0bc5f613fbbf3f0a9896cdef76e change-id: 20260601-pks-t7527-fix-tap-output-105da1d73df0

Before running the tests in t7527 we first verify whether the fsmonitor even works, which seems to depend on the actual filesystem that is in use. The verification executes outside of any prerequisite or test body, so its stdout/stderr is not being redirected.

The consequence of this is that any command that prints to stdout/stderr may break the TAP specification by printing invalid lines. And in fact we already do that, as git-init(1) prints the path to the created Git repository by default.

Fix this issue by moving the logic into a lazy prerequisite.

Signed-off-by: Patrick Steinhardt <ps@pks.im>
---
 t/t7527-builtin-fsmonitor.sh | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

When tests have finished we clean up the trash directory via `rm -rf`. On Windows this can fail with EBUSY in cases where a process still holds some of the files open, for example when we have spawned a daemonized process that wasn't properly terminated. We thus retry several times, but every failure will result in error messages being printed, and that in turn breaks the TAP output format.

One such case where this is causing issues is in t921x, which contains tests related to Scalar. Some tests spawn the fsmonitor daemon, and we never properly terminate it.

The obvious fix would be to ensure that we never leak any processes, but that gets ugly fast. Instead, let's work around the issue by silencing error messages printed by the `rm -rf` calls. We already know to print an error when the retry loop fails, so we don't loose much.

Signed-off-by: Patrick Steinhardt <ps@pks.im>
---
 t/test-lib.sh | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

When stopping the p4d watchdog process via "kill -9", the shell may print a job-control notification like:

  ./test-lib.sh: line 1269: 57960 Killed: 9               while true; do
      if test $nr_tries_left -eq 0; then
          kill -9 $p4d_pid; exit 1;
      fi; sleep 1; nr_tries_left=$(($nr_tries_left - 1));
  done 2> /dev/null 4>&2  (wd: ~)

This message is printed asynchronously by the shell when it reaps the process. While harmless right now, this will cause breakage once we enable strict parsing of the TAP protocol in a subsequent commit.

Fix this by using `wait` so that we can synchronously reap the watchdog process and swallow the diagnostic.

Signed-off-by: Patrick Steinhardt <ps@pks.im>
---
 t/lib-git-p4.sh | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

To make the result of our tests accessible we use the TAP protocol. This protocol is parsed by either prove or by Meson. Unfortunately, these two tools differ when it comes to their strictness when parsing the protocol:

  - Prove by default happily accepts lines not specified by the
    protocol.

  - Meson will also accept such lines, but prints a big and ugly warning
    message.

We have fixed our test suite in the past to not print invalid TAP lines anymore via b1dc2e796e (Merge branch 'ps/meson-tap-parse', 2025-06-17). But as none of our tools perform a strict check it's still possible for broken tests to sneak back in, like for example in 362f69547f (Merge branch 'ps/t1006-tap-fix', 2025-07-16). This doesn't hurt at all when using prove, but it's quite annoying when using Meson due to the generated warnings.

Unfortunately, there doesn't seem to be a portable way to make all tools complain about violations of the TAP format. The TAP 14 specification has added pragmas to the protocol that would allow us to say `pragma +strict`, and the effect of that would be to treat invalid TAP lines as a test failure. But the release of TAP 14 is still rather recent, and Test-Harness for example only gained support for it in version 3.48, which was released in 2023.

In fact though, this pragma was already introduced as an inofficial extension of the TAP protocol with Test-Harness 3.10, released in 2008. So while not all tools understand the pragma, at least prove does for a long time.

Unconditionally enable the pragma when using prove so that we'll detect tests that emit broken TAP output right away. This would have detected the issues fixed in preceding commits:

    $ prove t7527-builtin-fsmonitor.sh
    t7527-builtin-fsmonitor.sh .. All 69 subtests passed
            (less 6 skipped subtests: 63 okay)

    Test Summary Report
    -------------------
    t7527-builtin-fsmonitor.sh (Wstat: 0 Tests: 69 Failed: 0)
      Parse errors: Unknown TAP token: "Initialized empty Git repository in /tmp/git/test_fsmonitor_smoke/.git/"

Signed-off-by: Patrick Steinhardt <ps@pks.im>
---
 t/test-lib.sh | 6 ++++++
 1 file changed, 6 insertions(+)

Patrick Steinhardt <ps@pks.im> writes:

Shoudln't we be waiting on $p4d_pid (not $p4d)...

... like we do here?

On Tue, Jun 02, 2026 at 06:32:55PM +0900, Junio C Hamano wrote:

Oh, good catch. The statement basically doesn't do anything, which isn't much of a problem because we really only care about silencing the error message when the watchdog is being terminated. Will fix.

Patrick

Patrick Steinhardt <ps@pks.im> writes:

Thanks. Another thing I noticed is that they look suspiciously similar.

Hi,

this small patch series fixes another instance of broken TAP output that has landed via 4d11b9c218 (Merge branch 'pt/fsmonitor-linux', 2026-05-31).

As this has happened multiple times by now I decided to have a look at whether we can fix this class of issues a bit more holistically. So this series also contains a change that makes prove bail out when it sees invalid TAP output, which uncovers a small set of preexisting issues in our test suite.

Changes in v3:
  - Fix a test gap for AlmaLinux and Debian in GitLab CI, which uncovers
    an issue flagged by Peff.
  - Fix TAP breakage in t7810.
  - Link to v2: https://patch.msgid.link/20260603-pks-t7527-fix-tap-output-v2-0-cf3af5694e20@pks.im

Changes in v2:
  - Fix waiting for p4d, and deduplicate the logic that does this.
  - Link to v1: https://patch.msgid.link/20260602-pks-t7527-fix-tap-output-v1-0-db3da2a1b137@pks.im

Test runs can be found at [1] and [2]. Note that GitHub-side tests are failing on Windows, but that is a preexisting failure on "master".

Thanks!

Patrick

[1]: https://gitlab.com/gitlab-org/git/-/merge_requests/585 [2]: https://github.com/git/git/pull/2320

---
Patrick Steinhardt (8):
      gitlab-ci: rearrange Linux jobs to match GitHub's order
      gitlab-ci: add missing Linux jobs
      ci: unify Linux images across GitLab and GitHub
      t7527: fix broken TAP output
      t7810: turn MB_REGEX check into a lazy prereq
      t/test-lib: silence EBUSY errors on Windows during test cleanup
      t/lib-git-p4: silence output when killing p4d and its watchdog
      t: let prove fail when parsing invalid TAP output

 .github/workflows/main.yml   |  2 +-
 .gitlab-ci.yml               | 23 +++++++++++++++--------
 ci/lib.sh                    |  2 +-
 t/lib-git-p4.sh              |  4 ++--
 t/t7527-builtin-fsmonitor.sh |  7 ++++---
 t/t7810-grep.sh              |  5 +++--
 t/test-lib.sh                | 10 ++++++++--
 7 files changed, 34 insertions(+), 19 deletions(-)

Range-diff versus v2:

-: ---------- > 1: 5e817b102f gitlab-ci: rearrange Linux jobs to match GitHub's order -: ---------- > 2: 83646cc834 gitlab-ci: add missing Linux jobs -: ---------- > 3: cca1567fbf ci: unify Linux images across GitLab and GitHub 1: 52abbd5280 = 4: 430bc51818 t7527: fix broken TAP output -: ---------- > 5: 78ef22df8d t7810: turn MB_REGEX check into a lazy prereq 2: ea1f1eb466 = 6: 7bbaeff48c t/test-lib: silence EBUSY errors on Windows during test cleanup 3: e97a515470 = 7: abf2be09e6 t/lib-git-p4: silence output when killing p4d and its watchdog 4: 436d7d8cf3 = 8: 04367c34be t: let prove fail when parsing invalid TAP output

--- base-commit: 1666c1265231b0bc5f613fbbf3f0a9896cdef76e change-id: 20260601-pks-t7527-fix-tap-output-105da1d73df0

Rearrange the order of Linux jobs that we have defined in GitLab CI so that it matches the order on GitHub's side. This makes it easier to compare whether the list of jobs actually matches on both sides.

Signed-off-by: Patrick Steinhardt <ps@pks.im>
---
 .gitlab-ci.yml | 15 ++++++++-------
 1 file changed, 8 insertions(+), 7 deletions(-)

The GitLab CI definitions are missing jobs for AlmaLinux and Debian, both of which exist in GitHub Workflows. Plug this gap.

Signed-off-by: Patrick Steinhardt <ps@pks.im>
---
 .gitlab-ci.yml | 6 ++++++
 ci/lib.sh      | 2 +-
 2 files changed, 7 insertions(+), 1 deletion(-)

The image for the "linux-breaking-changes" job has drifted apart across GitHub and GitLab. Adapt it to use "ubuntu:rolling" on both systems.

With this change there's only one difference remaining: GitHub uses "ubuntu:focal" for the "linux32" job while GitLab uses "ubuntu:20.04". These are different names for the same image, so there is no actual difference here. Adjust GitHub to use the "20.04" tag -- this matches all the other jobs which use version numbers, and you don't have to learn Ubuntu's release names by heart.

Signed-off-by: Patrick Steinhardt <ps@pks.im>
---
 .github/workflows/main.yml | 2 +-
 .gitlab-ci.yml             | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

Before running the tests in t7527 we first verify whether the fsmonitor even works, which seems to depend on the actual filesystem that is in use. The verification executes outside of any prerequisite or test body, so its stdout/stderr is not being redirected.

The consequence of this is that any command that prints to stdout/stderr may break the TAP specification by printing invalid lines. And in fact we already do that, as git-init(1) prints the path to the created Git repository by default.

Fix this issue by moving the logic into a lazy prerequisite.

Signed-off-by: Patrick Steinhardt <ps@pks.im>
---
 t/t7527-builtin-fsmonitor.sh | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

In t7810 we verify whether the system has proper multibyte locale support by executing `test-tool regex` with a unicode character. When this check fails though we'll output an error that breaks the TAP format.

Fix this issue by turning the logic into a lazy prerequisite.

Reported-by: Jeff King <peff@peff.net>
Signed-off-by: Patrick Steinhardt <ps@pks.im>
---
 t/t7810-grep.sh | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

When tests have finished we clean up the trash directory via `rm -rf`. On Windows this can fail with EBUSY in cases where a process still holds some of the files open, for example when we have spawned a daemonized process that wasn't properly terminated. We thus retry several times, but every failure will result in error messages being printed, and that in turn breaks the TAP output format.

One such case where this is causing issues is in t921x, which contains tests related to Scalar. Some tests spawn the fsmonitor daemon, and we never properly terminate it.

The obvious fix would be to ensure that we never leak any processes, but that gets ugly fast. Instead, let's work around the issue by silencing error messages printed by the `rm -rf` calls. We already know to print an error when the retry loop fails, so we don't loose much.

Signed-off-by: Patrick Steinhardt <ps@pks.im>
---
 t/test-lib.sh | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

When stopping the p4d watchdog process via "kill -9", the shell may print a job-control notification like:

  ./test-lib.sh: line 1269: 57960 Killed: 9               while true; do
      if test $nr_tries_left -eq 0; then
          kill -9 $p4d_pid; exit 1;
      fi; sleep 1; nr_tries_left=$(($nr_tries_left - 1));
  done 2> /dev/null 4>&2  (wd: ~)

This message is printed asynchronously by the shell when it reaps the process. While harmless right now, this will cause breakage once we enable strict parsing of the TAP protocol in a subsequent commit.

Fix this by using `wait` so that we can synchronously reap the watchdog process and swallow the diagnostic.

While at it, deduplicate the logic we have in `stop_p4d_and_watchdog ()` and `stop_and_cleanup_p4d ()`.

Signed-off-by: Patrick Steinhardt <ps@pks.im>
---
 t/lib-git-p4.sh | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

To make the result of our tests accessible we use the TAP protocol. This protocol is parsed by either prove or by Meson. Unfortunately, these two tools differ when it comes to their strictness when parsing the protocol:

  - Prove by default happily accepts lines not specified by the
    protocol.

  - Meson will also accept such lines, but prints a big and ugly warning
    message.

We have fixed our test suite in the past to not print invalid TAP lines anymore via b1dc2e796e (Merge branch 'ps/meson-tap-parse', 2025-06-17). But as none of our tools perform a strict check it's still possible for broken tests to sneak back in, like for example in 362f69547f (Merge branch 'ps/t1006-tap-fix', 2025-07-16). This doesn't hurt at all when using prove, but it's quite annoying when using Meson due to the generated warnings.

Unfortunately, there doesn't seem to be a portable way to make all tools complain about violations of the TAP format. The TAP 14 specification has added pragmas to the protocol that would allow us to say `pragma +strict`, and the effect of that would be to treat invalid TAP lines as a test failure. But the release of TAP 14 is still rather recent, and Test-Harness for example only gained support for it in version 3.48, which was released in 2023.

In fact though, this pragma was already introduced as an inofficial extension of the TAP protocol with Test-Harness 3.10, released in 2008. So while not all tools understand the pragma, at least prove does for a long time.

Unconditionally enable the pragma when using prove so that we'll detect tests that emit broken TAP output right away. This would have detected the issues fixed in preceding commits:

    $ prove t7527-builtin-fsmonitor.sh
    t7527-builtin-fsmonitor.sh .. All 69 subtests passed
            (less 6 skipped subtests: 63 okay)

    Test Summary Report
    -------------------
    t7527-builtin-fsmonitor.sh (Wstat: 0 Tests: 69 Failed: 0)
      Parse errors: Unknown TAP token: "Initialized empty Git repository in /tmp/git/test_fsmonitor_smoke/.git/"

Signed-off-by: Patrick Steinhardt <ps@pks.im>
---
 t/test-lib.sh | 6 ++++++
 1 file changed, 6 insertions(+)

On Thu, Jun 04, 2026 at 12:07:35PM +0200, Patrick Steinhardt wrote:

Oh good. Since the error was coming from the shell, I was worried that the use of LC_ALL inside the test snippets was somehow causing output to leak to stderr. But it was just (yet another) case where we ran a tool outside of a snippet, which is easy to fix.

This also allays my fears that the tests might have been misbehaving on those other platforms. They are (and were) caught by this prereq and not run at all, which is the right thing.

Thanks for fixing.

-Peff