[PATCH 0/2] Eliminate extraneous ref log entries

[PATCH 0/2] Eliminate extraneous ref log entries

[Kyle J. McKay]

Since Git version v2.29.0, the `git symbolic-ref` command has started
adding extraneous entries to the ref log of the symbolic ref it's
updating.

This change was inadvertently introduced in commit 523fa69c36744ae6
("reflog: cleanse messages in the refs.c layer", 2020-07-10, v2.29.0).

A bug report [1] was made about a failing test in the TopGit test
suite.  Further investigations into the cause led to this patch set.

1/2 - adds new tests to monitor this behavior
2/2 - corrects the problem

The tests added in 1/2 are marked `test_expect_failure` and then
changed to `test_expect_success` in 2/2.

-Kyle

[1]: <https://github.com/mackyle/topgit/issues/17>

Kyle J. McKay (2):
  t/t1417: test symbolic-ref effects on ref logs
  refs.c: avoid creating extra unwanted reflog entries

 refs.c                   | 16 +++----
 t/t1417-reflog-symref.sh | 91 ++++++++++++++++++++++++++++++++++++++++
 2 files changed, 100 insertions(+), 7 deletions(-)
 create mode 100755 t/t1417-reflog-symref.sh

-- 

[PATCH 1/2] t/t1417: test symbolic-ref effects on ref logs

[Kyle J. McKay]

The git command `git symbolic-ref <refname1> <refname2>` updates
<refname1> to be a "symbolic" pointer to <refname2>.  No matter
what future value <refname2> takes on, <refname1> will continue
to refer to that future value since it's "symbolic".

Since commit 523fa69c36744ae6 ("reflog: cleanse messages in the refs.c
layer", 2020-07-10, v2.29.0), the effect of using the aforementioned
"symbolic-ref" command on ref logs has changed in an unexpected way.

Add a new set of tests to exercise and demonstrate this change
in preparation for correcting it (at which point the failing tests
will be flipped from `test_expect_failure` to `test_expect_success`).

The new test file can be used unchanged to examine this behavior
in much older Git versions (likely to as far back as v2.6.0).

Signed-off-by: Kyle J. McKay <mackyle@gmail.com>
---
 t/t1417-reflog-symref.sh | 91 ++++++++++++++++++++++++++++++++++++++++
 1 file changed, 91 insertions(+)
 create mode 100755 t/t1417-reflog-symref.sh

diff --git a/t/t1417-reflog-symref.sh b/t/t1417-reflog-symref.sh
new file mode 100755
index 00000000..6149531f
--- /dev/null
+++ b/t/t1417-reflog-symref.sh
@@ -0,0 +1,91 @@
+#!/bin/sh
+#
+# Copyright (c) 2021 Kyle J. McKay
+#
+
+test_description='Test symbolic-ref effects on reflogs'
+GIT_TEST_DEFAULT_INITIAL_BRANCH_NAME=main
+export GIT_TEST_DEFAULT_INITIAL_BRANCH_NAME
+
+. ./test-lib.sh
+
+test_expect_success setup '
+	test_commit 'initial' &&
+	git checkout -b unu &&
+	test_commit 'one' &&
+	git checkout -b du &&
+	test_commit 'two' &&
+	git checkout -b tri &&
+	test_commit 'three' &&
+	git checkout du &&
+	test_commit 'twofour' &&
+	git checkout -b KVAR du &&
+	test_commit 'four' &&
+	unu="$(git rev-parse --verify unu)" &&
+	du="$(git rev-parse --verify du)" &&
+	tri="$(git rev-parse --verify tri)" &&
+	kvar="$(git rev-parse --verify KVAR)" &&
+	test -n "$unu" &&
+	test -n "$du" &&
+	test -n "$tri" &&
+	test -n "$kvar" &&
+	test "$unu" != "$du" &&
+	test "$unu" != "$tri" &&
+	test "$unu" != "$kvar" &&
+	test "$du" != "$tri" &&
+	test "$du" != "$kvar" &&
+	test "$tri" != "$kvar" &&
+	git symbolic-ref refs/heads/KVAR refs/heads/du &&
+	kvarsym="$(git rev-parse --verify KVAR)" &&
+	test "$kvarsym" = "$du" &&
+	test "$kvarsym" != "$kvar" &&
+	git reflog exists HEAD &&
+	git reflog exists refs/heads/KVAR &&
+	git symbolic-ref HEAD >/dev/null &&
+	git symbolic-ref refs/heads/KVAR &&
+	git checkout unu &&
+	hcnt=$(git reflog show HEAD | wc -l) &&
+	kcnt=$(git reflog show refs/heads/KVAR | wc -l) &&
+	test -n "$hcnt" &&
+	test -n "$kcnt" &&
+	test $hcnt -gt 1 &&
+	test $kcnt -gt 1 &&
+	test $hcnt -ne $kcnt
+'
+
+test_expect_failure 'HEAD reflog symbolic-ref' '
+	hcnt1=$(git reflog show HEAD | wc -l) &&
+	git symbolic-ref HEAD refs/heads/unu &&
+	git symbolic-ref HEAD refs/heads/du &&
+	git symbolic-ref HEAD refs/heads/tri &&
+	hcnt2=$(git reflog show HEAD | wc -l) &&
+	test $hcnt1 = $hcnt2
+'
+
+test_expect_failure 'refs/heads/KVAR reflog symbolic-ref' '
+	kcnt1=$(git reflog show refs/heads/KVAR | wc -l) &&
+	git symbolic-ref refs/heads/KVAR refs/heads/tri &&
+	git symbolic-ref refs/heads/KVAR refs/heads/du &&
+	git symbolic-ref refs/heads/KVAR refs/heads/unu &&
+	kcnt2=$(git reflog show refs/heads/KVAR | wc -l) &&
+	test $kcnt1 = $kcnt2
+'
+
+test_expect_failure 'double symref reflog symbolic-ref' '
+	hcnt1=$(git reflog show HEAD | wc -l) &&
+	kcnt1=$(git reflog show refs/heads/KVAR | wc -l) &&
+	git symbolic-ref HEAD refs/heads/KVAR &&
+	git symbolic-ref refs/heads/KVAR refs/heads/du &&
+	git symbolic-ref refs/heads/KVAR refs/heads/unu &&
+	git symbolic-ref refs/heads/KVAR refs/heads/tri &&
+	git symbolic-ref HEAD refs/heads/du &&
+	git symbolic-ref HEAD refs/heads/tri &&
+	git symbolic-ref HEAD refs/heads/unu &&
+	hcnt2=$(git reflog show HEAD | wc -l) &&
+	kcnt2=$(git reflog show refs/heads/KVAR | wc -l) &&
+	test $hcnt1 = $hcnt2 &&
+	test $kcnt1 = $kcnt2 &&
+	test $hcnt2 != $kcnt2
+'
+
+test_done
-- 

[PATCH 2/2] refs.c: avoid creating extra unwanted reflog entries

[Kyle J. McKay]

Since commit 523fa69c36744ae6 ("reflog: cleanse messages in the refs.c
layer", 2020-07-10, v2.29.0), ref log messages are now being "cleansed"
to make sure they do not end up breaking the ref log files.  A laudable
endeavor.

Unfortunately, that commit had an unintended side effect that causes
the `git symbolic-ref <refname1> <refname2>` command to suddenly start
adding new entries to the ref log for <refname1> whenever it's run.

These new entries have a completely empty message and do not provide
any useful information.  In fact, there was no mention that the change
to "cleanse" ref log messages was intended to add these new ref log
entries at all.

What happened is that when the change to "cleanse" the incoming ref
log message was made, the code started inadvertently transforming
a NULL ref log message pointer into an empty string "".

This created the observed effect that using the `symbolic-ref` command
suddenly started causing ref log entries to be added.

The original code that predated the "cleanse" commit called the
`xstrdup_or_null` function to retain the original NULL pointer and
avoid introducing unwanted extra ref log entries.

After the "cleanse" commit, ref log messages are now funnelled through
a new static function named `normalize_reflog_message`.

Eliminate the unwanted extra blank ref log entries by returning a NULL
pointer when NULL is passed into `normalize_reflog_message` rather
than returning a pointer to an empty string ("").

To reflect this new behavior, rename the function to
`normalize_reflog_message_or_null` in the same spirit as the name
of the `xstrdup_or_null` function that was called pre-"cleanse".

Flip the `test_expect_failure` tests to `test_expect_success`
as they now pass again.

Signed-off-by: Kyle J. McKay <mackyle@gmail.com>
---
 refs.c                   | 16 +++++++++-------
 t/t1417-reflog-symref.sh |  6 +++---
 2 files changed, 12 insertions(+), 10 deletions(-)

diff --git a/refs.c b/refs.c
index 03968ad7..790b1ff0 100644
--- a/refs.c
+++ b/refs.c
@@ -835,11 +835,13 @@ static void copy_reflog_msg(struct strbuf *sb, const char *msg)
 	strbuf_rtrim(sb);
 }
 
-static char *normalize_reflog_message(const char *msg)
+static char *normalize_reflog_message_or_null(const char *msg)
 {
 	struct strbuf sb = STRBUF_INIT;
 
-	if (msg && *msg)
+	if (!msg)
+		return NULL;
+	if (*msg)
 		copy_reflog_msg(&sb, msg);
 	return strbuf_detach(&sb, NULL);
 }
@@ -1067,7 +1069,7 @@ struct ref_update *ref_transaction_add_update(
 		oidcpy(&update->new_oid, new_oid);
 	if (flags & REF_HAVE_OLD)
 		oidcpy(&update->old_oid, old_oid);
-	update->msg = normalize_reflog_message(msg);
+	update->msg = normalize_reflog_message_or_null(msg);
 	return update;
 }
 
@@ -1951,7 +1953,7 @@ int refs_create_symref(struct ref_store *refs,
 	char *msg;
 	int retval;
 
-	msg = normalize_reflog_message(logmsg);
+	msg = normalize_reflog_message_or_null(logmsg);
 	retval = refs->be->create_symref(refs, ref_target, refs_heads_master,
 					 msg);
 	free(msg);
@@ -2339,7 +2341,7 @@ int refs_delete_refs(struct ref_store *refs, const char *logmsg,
 	char *msg;
 	int retval;
 
-	msg = normalize_reflog_message(logmsg);
+	msg = normalize_reflog_message_or_null(logmsg);
 	retval = refs->be->delete_refs(refs, msg, refnames, flags);
 	free(msg);
 	return retval;
@@ -2357,7 +2359,7 @@ int refs_rename_ref(struct ref_store *refs, const char *oldref,
 	char *msg;
 	int retval;
 
-	msg = normalize_reflog_message(logmsg);
+	msg = normalize_reflog_message_or_null(logmsg);
 	retval = refs->be->rename_ref(refs, oldref, newref, msg);
 	free(msg);
 	return retval;
@@ -2374,7 +2376,7 @@ int refs_copy_existing_ref(struct ref_store *refs, const char *oldref,
 	char *msg;
 	int retval;
 
-	msg = normalize_reflog_message(logmsg);
+	msg = normalize_reflog_message_or_null(logmsg);
 	retval = refs->be->copy_ref(refs, oldref, newref, msg);
 	free(msg);
 	return retval;
diff --git a/t/t1417-reflog-symref.sh b/t/t1417-reflog-symref.sh
index 6149531f..3687b058 100755
--- a/t/t1417-reflog-symref.sh
+++ b/t/t1417-reflog-symref.sh
@@ -53,7 +53,7 @@ test_expect_success setup '
 	test $hcnt -ne $kcnt
 '
 
-test_expect_failure 'HEAD reflog symbolic-ref' '
+test_expect_success 'HEAD reflog symbolic-ref' '
 	hcnt1=$(git reflog show HEAD | wc -l) &&
 	git symbolic-ref HEAD refs/heads/unu &&
 	git symbolic-ref HEAD refs/heads/du &&
@@ -62,7 +62,7 @@ test_expect_failure 'HEAD reflog symbolic-ref' '
 	test $hcnt1 = $hcnt2
 '
 
-test_expect_failure 'refs/heads/KVAR reflog symbolic-ref' '
+test_expect_success 'refs/heads/KVAR reflog symbolic-ref' '
 	kcnt1=$(git reflog show refs/heads/KVAR | wc -l) &&
 	git symbolic-ref refs/heads/KVAR refs/heads/tri &&
 	git symbolic-ref refs/heads/KVAR refs/heads/du &&
@@ -71,7 +71,7 @@ test_expect_failure 'refs/heads/KVAR reflog symbolic-ref' '
 	test $kcnt1 = $kcnt2
 '
 
-test_expect_failure 'double symref reflog symbolic-ref' '
+test_expect_success 'double symref reflog symbolic-ref' '
 	hcnt1=$(git reflog show HEAD | wc -l) &&
 	kcnt1=$(git reflog show refs/heads/KVAR | wc -l) &&
 	git symbolic-ref HEAD refs/heads/KVAR &&
-- 

Re: [PATCH 1/2] t/t1417: test symbolic-ref effects on ref logs

[Junio C Hamano]

"Kyle J. McKay" <mackyle@gmail.com> writes:

> The git command `git symbolic-ref <refname1> <refname2>` updates
> <refname1> to be a "symbolic" pointer to <refname2>.  No matter
> what future value <refname2> takes on, <refname1> will continue
> to refer to that future value since it's "symbolic".

Correct.  While you are on the my-topic branch, HEAD (that's the
<refname1> in your description) points at refs/heads/my-topic
(that's the <refname2> in your description).

And when you create a new commit from this state, the logs of these
two refs will gain one entry each.  "git log HEAD@{now} would show
the recent progress of HEAD, "git log my-topic@{now}" would show the
recent progress of my-topic ("my-topic@"now}" can also be spelled as
"@{now}).

The <refname1> (HEAD) will keep referring to <refname2> (my-topic)
until you switch branches, and does not change even if <refname2>
points at a different commit, as it is "symbolic".

> Since commit 523fa69c36744ae6 ("reflog: cleanse messages in the refs.c
> layer", 2020-07-10, v2.29.0), the effect of using the aforementioned
> "symbolic-ref" command on ref logs has changed in an unexpected way.

Please explain "in an unexpected way" here in the log message.  Not
every reader can read your mind and would expect the same as you do.

The said commit came as part of this topic, ...

https://lore.kernel.org/git/pull.669.v2.git.1594401593.gitgitgadget@gmail.com/

... so I've added the true author of it on the Cc: list.

> Add a new set of tests to exercise and demonstrate this change
> in preparation for correcting it (at which point the failing tests
> will be flipped from `test_expect_failure` to `test_expect_success`).

We usually prefer not to do the two-step "expect_failure first and
then in a separate patch flip that to _success", as it makes it hard
to see the "fix" step (because the change in the test would be shown
only 3 lines before and after _failure->_success line, and what the
test is doing cannot be seen in the patch by itself).

Thanks.

Re: [PATCH 1/2] t/t1417: test symbolic-ref effects on ref logs

[Kyle J. McKay]

On Jan 30, 2021, at 11:56, Junio C Hamano wrote:

> The said commit came as part of this topic, ...
>
> https://lore.kernel.org/git/pull.669.v2.git.1594401593.gitgitgadget@gmail.com/
>
> ... so I've added the true author of it on the Cc: list.

Out of curiosity, if Han-Wen Nienhuys is the true author of commit  
523fa69c36744ae6 why is it that you are both the committer and author  
of that commit in the commit's header?

The answer may also inform others trying to determine the best list of  
recipients for patches...

Re: [PATCH 1/2] t/t1417: test symbolic-ref effects on ref logs

[Kyle J. McKay]

On Jan 30, 2021, at 11:56, Junio C Hamano wrote:

>> Add a new set of tests to exercise and demonstrate this change
>> in preparation for correcting it (at which point the failing tests
>> will be flipped from `test_expect_failure` to `test_expect_success`).
>
> We usually prefer not to do the two-step "expect_failure first and
> then in a separate patch flip that to _success", as it makes it hard
> to see the "fix" step (because the change in the test would be shown
> only 3 lines before and after _failure->_success line, and what the
> test is doing cannot be seen in the patch by itself).

I'm having a bit of trouble parsing that into expectations.  A little  
help please.

Given the scenario that a bug is found that is not being caught by a  
test (irrespective of whether or not the outcome of this particular  
series discussion results in it being determined to be a "bug").

Further, if the fix is simple enough that it warrants only a single  
patch, what is the preferred order of patches then?

I would like the patch series to demonstrate:

1) that the test actually catches the bug (in case it comes back in  
the future)
2) the fix isolated (as much as possible) in a patch distinct from the  
test
3) that the test now passes, preferably with minimal changes to be  
sure that it hasn't accidentally been rendered ineffective

All the while, at no point after commits for (1), (2), or (3) is the  
test suite allowed to generate extra failures (that are not marked  
"expect failure").

patch 1/2 accomplishes (1)
patch 2/2 does both (2) and (3)

Are you suggesting that (1) just be omitted?  Or that it be modified  
so that it's an "expect success" patch?  But then (2) would break it  
and introduce a failing test into the test suite.  Should (2) then  
just flip the test from test_expect_success to test_expect_failure and  
then a separate commit flip it back to test_expect_success along with  
the minor change to the test itself to make it start passing again?   
In that case, there's always a risk that changing the test itself  
could accidentally make it no longer detect the problem anymore and  
just always succeed even if the problem comes back.

Without an initial expect_failure step (1), there's no commit in the  
repository that can demonstrate that the test actually catches the  
problem.

I understand that when new code is added, the tests come after, but it  
seems to me that when fixing a pre-existing problem, the test that  
demonstrates the problem should come first and be an expect failure  
since it is considered to be a problem.

What exactly is the order of test/fix changes that you're expecting to  
see here?

Thanks.

Re: [PATCH 1/2] t/t1417: test symbolic-ref effects on ref logs

[Junio C Hamano]

"Kyle J. McKay" <mackyle@gmail.com> writes:

> On Jan 30, 2021, at 11:56, Junio C Hamano wrote:
>
>> The said commit came as part of this topic, ...
>>
>> https://lore.kernel.org/git/pull.669.v2.git.1594401593.gitgitgadget@gmail.com/
>>
>> ... so I've added the true author of it on the Cc: list.
>
> Out of curiosity, if Han-Wen Nienhuys is the true author of commit
> 523fa69c36744ae6 why is it that you are both the committer and author  
> of that commit in the commit's header?

See how the e-mail message was formatted in that thread.  I just ran
"am" on it (which makes me responsible for committing), and the
authorship comes from the "From:" that was in the body.  I suspect
he may have based the patch on some of the "how about doing it like
so" suggestions I made during an earlier discussion and wanted to
give me credit for the input, but I do not remember the context the
patch was originally written in X-<.

Re: [PATCH 1/2] t/t1417: test symbolic-ref effects on ref logs

[Junio C Hamano]

"Kyle J. McKay" <mackyle@gmail.com> writes:

> I'm having a bit of trouble parsing that into expectations.  A little
> help please.
> Are you suggesting that (1) just be omitted?  Or that it be modified
> so that it's an "expect success" patch?

Neither.

The result of applying the current 1/2 and 2/2 on top of, say
'master', would be the shape of the tree you would want to be in.

Our preference is just to have it as a single patch, not as "first
expect failure and then flip it to expect success while modifying
the code".  That approach makes the second step harder to review
than necessary, because the "git show" output and "format-patch"
output from the step would show only very little about the test
that changes behaviour.

Even with a single patch, if somebody wants a demonstration of what
used to be broken without the code modification, it is easy to apply
only the test part of the single patch without using the code change
to see how it breaks, so "I want to demonstrate the breakage" is not
a reason to have it as a separate step.

Thanks.

Re: [PATCH 1/2] t/t1417: test symbolic-ref effects on ref logs

[Han-Wen Nienhuys]

On Sun, Jan 31, 2021 at 12:48 AM Junio C Hamano <gitster@pobox.com> wrote:
> > On Jan 30, 2021, at 11:56, Junio C Hamano wrote:
> >
> >> The said commit came as part of this topic, ...
> >>
> >> https://lore.kernel.org/git/pull.669.v2.git.1594401593.gitgitgadget@gmail.com/
> >>
> >> ... so I've added the true author of it on the Cc: list.
> >
> > Out of curiosity, if Han-Wen Nienhuys is the true author of commit
> > 523fa69c36744ae6 why is it that you are both the committer and author
> > of that commit in the commit's header?
>
> See how the e-mail message was formatted in that thread.  I just ran
> "am" on it (which makes me responsible for committing), and the
> authorship comes from the "From:" that was in the body.  I suspect
> he may have based the patch on some of the "how about doing it like
> so" suggestions I made during an earlier discussion and wanted to
> give me credit for the input, but I do not remember the context the
> patch was originally written in X-<.

The classic reflog format doesn't allow '\n' in messages, but
different parts of the code did try to write '\n'. This patch was
supposed to sanitize the messages in a central location, so alternate
ref backends do not trigger spurious differences in how reflogs are
represented.

Your patch says

> has changed in an unexpected way.

Can you make the expectations and current behavior explicit?

-- 
Han-Wen Nienhuys - Google Munich
I work 80%. Don't expect answers from me on Fridays.
--
Google Germany GmbH, Erika-Mann-Strasse 33, 80636 Munich
Registergericht und -nummer: Hamburg, HRB 86891
Sitz der Gesellschaft: Hamburg
Geschäftsführer: Paul Manicle, Halimah DeLaine Prado