sequencer: honor signoff opt in run_git_commit

Commit Message

David Bimmler April 4, 2024, 9:39 a.m. UTC

From: David Bimmler <[e-mail]>

When rebasing interactively, --signoff would not take effect for commits
which conflict. That is, commits applying cleanly would be signed off,
but commits requiring intervention would miss the sign off trailer.

The reason is that run_git_commit did not check for the signoff replay
opt, and hence even though the option was picked up and passed
correctly, the actual committing dropped the ball.

The patch adds a test specifically for this case, as well as amending a
squash test which codified the broken behaviour.

Signed-off-by: David Bimmler <[e-mail]>
---
    sequencer: honor signoff opt in run_git_commit

Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-1707%2Fbimmlerd%2Fsignoff-conflicting-commits-in-rebase-v1
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-1707/bimmlerd/signoff-conflicting-commits-in-rebase-v1
Pull-Request: https://github.com/gitgitgadget/git/pull/1707

 sequencer.c                     |  2 ++
 t/t3428-rebase-signoff.sh       | 33 +++++++++++++++++++++++++++++++++
 t/t3437/expected-squash-message |  2 ++
 3 files changed, 37 insertions(+)


base-commit: 7774cfed6261ce2900c84e55906da708c711d601

Comments

Junio C Hamano April 4, 2024, 4:22 p.m. UTC | #1

"David Bimmler via GitGitGadget" <[e-mail]> writes:

> From: David Bimmler <[e-mail]>
>
> When rebasing interactively, --signoff would not take effect for commits
> which conflict. That is, commits applying cleanly would be signed off,
> but commits requiring intervention would miss the sign off trailer.

Good finding.

> The reason is that run_git_commit did not check for the signoff replay
> opt, and hence even though the option was picked up and passed
> correctly, the actual committing dropped the ball.

We record and read back the initial "--signoff" given from the
command line (and options.signoff from the "options sheet" when not
REPLAY_INTERACTIVE_REBASE) in opts->signoff, but that piece of
information is used only to decide if we call append_signoff().
When do_commit(), after try_to_commit() punts, calls "git commit",
the invocation lacks "--signoff" on the command line.

    Side note: the "let's avoid spawning 'git commit' and duplicate
    it in-process" codepath in try_to_commit() does not seem to pay
    attention to opts->signoff at all, and it does not do the
    append_signoff() to add the trailer.  Is this something this
    patch needs to fix as well, or am I misreading the control flow
    (e.g., perhaps opts->signoff is set we somehow refrain from
    doing the "let's do it in-process" logic)?

If we forgot to tell "git commit" that we want the result
signed-off, of course it would not add the trailer.  Perhaps

    -correctly, the actual committing dropped the ball" 
    +correctly to run_git_commit(), it failed to pass it to
    +underlying 'git commit'

is less misleading.

> The patch adds a test specifically for this case, as well as amending a
> squash test which codified the broken behaviour.

"The patch adds" -> "Add"

I love this kind of test update that corrects a mistaken expectation
after a careful analysis of how the commands should behave.

Very good job.  Nicely done.

> diff --git a/sequencer.c b/sequencer.c
> index fa838f264f5..16595e26a17 100644
> --- a/sequencer.c
> +++ b/sequencer.c
> @@ -1121,6 +1121,8 @@ static int run_git_commit(const char *defmsg,
>  		strvec_pushf(&cmd.args, "-S%s", opts->gpg_sign);
>  	else
>  		strvec_push(&cmd.args, "--no-gpg-sign");
> +	if (opts->signoff)
> +		strvec_push(&cmd.args, "--signoff");
>  	if (defmsg)
>  		strvec_pushl(&cmd.args, "-F", defmsg, NULL);
>  	else if (!(flags & EDIT_MSG))

This is straight-forward and obviously correct.

> diff --git a/t/t3428-rebase-signoff.sh b/t/t3428-rebase-signoff.sh
> index e1b1e947647..fcecdf41978 100755
> --- a/t/t3428-rebase-signoff.sh
> +++ b/t/t3428-rebase-signoff.sh
> @@ -27,6 +27,13 @@ first
>  Signed-off-by: $(git var GIT_COMMITTER_IDENT | sed -e "s/>.*/>/")
>  EOF
>  
> +# Expected signed off message after resolving the conflict
> +cat >expected-signed-after-conflict <<EOF
> +update file on side
> +
> +Signed-off-by: $(git var GIT_COMMITTER_IDENT | sed -e "s/>.*/>/")
> +EOF
> +

The t3428 script uses ancient style to prepare these expectations
outside the test_expect_success blocks, and following the pattern
is fine (until we clean-up the style of the whole script).  Cleaning
it up is better done as a separate series, and when the dust settles
after this fix lands.  #leftoverbits 

>  # Expected commit message after rebase without --signoff (or with --no-signoff)
>  cat >expected-unsigned <<EOF
>  first
> @@ -82,4 +89,30 @@ test_expect_success 'rebase -m --signoff fails' '
>  	git cat-file commit HEAD | sed -e "1,/^\$/d" >actual &&
>  	test_cmp expected-signed actual
>  '
> +
> +test_expect_success 'rebase -i signs commits even if a conflict occurs' '
> +	git branch -M main &&
> +
> +	git branch side &&
> +	echo "b" >file &&
> +	git add file &&
> +	git commit -m"update file" &&

Have a SP between "-m" and its value, imitating existing tests.

> +	test_tick &&

Why?  Does the rest of the test depend on commits to have different
timestamps?

If you want to use test_tick, the pattern to follow is to call it
_before_ creating a commit, not after.  Up to this point in the
original script nobody calls it (or test_commit that by default
calls it), so "test_tick && git_commit" would be OK.

But if there is no dependency on exact sequence of timestamps,
letting the commit share the same initial timestamp (hardcoded in
test-lib.sh for better reproducibility) is more preferable, as use
of test_tick sends a wrong message that the test results may be
affected by the exact timestamps.

> +	test_must_fail git rebase -i --signoff main &&
> +
> +	echo "merged" >file &&
> +	git add file &&
> +	git rebase --continue &&
> +
> +	git cat-file commit HEAD | sed -e "1,/^\$/d" >actual &&
> +	test_cmp expected-signed-after-conflict actual

Running any git command on the left hand side of a pipe is frowned
upon, as it will hide exit status from it when it fails.

In this case, the primary thing we care about is that we have added
the sign off that did not exist in the original, so I wonder

	git cat-file commit HEAD >actual &&
	test_grep "Signed-off-by: " actual

would be sufficient?

> +'
> +
>  test_done
> diff --git a/t/t3437/expected-squash-message b/t/t3437/expected-squash-message
> index ab2434f90ed..d74af0bcf6b 100644
> --- a/t/t3437/expected-squash-message
> +++ b/t/t3437/expected-squash-message
> @@ -48,4 +48,6 @@ edited 1
>  edited 2
>  
>  edited 3
> +
> +Signed-off-by: Rebase Committer <[e-mail]>
>  squashed

This defines the expected outcome from

    git rebase -ik --signoff A

in t3437-rebase-fixup-options.sh; I am not sure if the location the
sign-off is inserted is sensible, though.

Thanks.

Junio C Hamano April 4, 2024, 4:45 p.m. UTC | #2

Junio C Hamano <[e-mail]> writes:

>> +	git cat-file commit HEAD | sed -e "1,/^\$/d" >actual &&
>> +	test_cmp expected-signed-after-conflict actual
>
> Running any git command on the left hand side of a pipe is frowned
> upon, as it will hide exit status from it when it fails.
>
> In this case, the primary thing we care about is that we have added
> the sign off that did not exist in the original, so I wonder
>
> 	git cat-file commit HEAD >actual &&
> 	test_grep "Signed-off-by: " actual
>
> would be sufficient?

The answer is No.  It is plausible that somebody else in a future
may think that a better fix is to always prepare the final commit
message inside sequencer.c and call append_signoff(), and use it in
both the "in-process commit" codepath in try_to_commit() and in the
code that was fixed in the patch we are discussing.  If such a
future update is done carelessly, we might end up adding duplicate
sign off.  A "at least one instance must be there" test_grep would
not be a good tool to catch such a breakage.

	git show -s --format=%B HEAD >actual &&
	test_cmp expect actual

may be a good replacement.

But having said that, and then after having looked at the existing
tests in the file, I see that it is littered with the same "do not
run git on the upstream of the pipe" violation.  So let's not worry
about this one.  The whole t3428 script needs to be cleaned up after
the dust settles.

Thanks.

Phillip Wood April 4, 2024, 7:21 p.m. UTC | #3

On 04/04/2024 17:22, Junio C Hamano wrote:
> "David Bimmler via GitGitGadget" <[e-mail]> writes:
> 
>> From: David Bimmler <[e-mail]>
>>
>> When rebasing interactively, --signoff would not take effect for commits
>> which conflict. That is, commits applying cleanly would be signed off,
>> but commits requiring intervention would miss the sign off trailer.
> 
> Good finding.

Yes, thanks for posting this patch.

>> The reason is that run_git_commit did not check for the signoff replay
>> opt, and hence even though the option was picked up and passed
>> correctly, the actual committing dropped the ball.
> 
> We record and read back the initial "--signoff" given from the
> command line (and options.signoff from the "options sheet" when not
> REPLAY_INTERACTIVE_REBASE) in opts->signoff, but that piece of
> information is used only to decide if we call append_signoff().
> When do_commit(), after try_to_commit() punts, calls "git commit",
> the invocation lacks "--signoff" on the command line.
> 
>      Side note: the "let's avoid spawning 'git commit' and duplicate
>      it in-process" codepath in try_to_commit() does not seem to pay
>      attention to opts->signoff at all, and it does not do the
>      append_signoff() to add the trailer.  Is this something this
>      patch needs to fix as well, or am I misreading the control flow
>      (e.g., perhaps opts->signoff is set we somehow refrain from
>      doing the "let's do it in-process" logic)?

I think it is more complicated than that. We do not use the "--signoff" 
option of "git commit" because we do not want to append the 
"Signed-off-by:" trailer when processing "fixup" and "squash" commands. 
The trailer is appended to the commit message by the sequencer in 
do_pick_commits(). The problem is that when we commit a conflict 
resolution we end up using the original commit message rather than the 
file containing the commit message that we would have used if we had not 
stopped for conflicts. I've got some patches which need their commit 
messages cleaning up at 
https://github.com/phillipwood/git/commits/wip/fix-rebase-signoff-with-conflicts/ 
which use the correct file when committing conflict resolutions. I'll 
try and clean them up next week.

Best Wishes

Phillip

> If we forgot to tell "git commit" that we want the result
> signed-off, of course it would not add the trailer.  Perhaps
> 
>      -correctly, the actual committing dropped the ball"
>      +correctly to run_git_commit(), it failed to pass it to
>      +underlying 'git commit'
> 
> is less misleading.
> 
>> The patch adds a test specifically for this case, as well as amending a
>> squash test which codified the broken behaviour.
> 
> "The patch adds" -> "Add"
> 
> I love this kind of test update that corrects a mistaken expectation
> after a careful analysis of how the commands should behave.
> 
> Very good job.  Nicely done.
> 
>> diff --git a/sequencer.c b/sequencer.c
>> index fa838f264f5..16595e26a17 100644
>> --- a/sequencer.c
>> +++ b/sequencer.c
>> @@ -1121,6 +1121,8 @@ static int run_git_commit(const char *defmsg,
>>   		strvec_pushf(&cmd.args, "-S%s", opts->gpg_sign);
>>   	else
>>   		strvec_push(&cmd.args, "--no-gpg-sign");
>> +	if (opts->signoff)
>> +		strvec_push(&cmd.args, "--signoff");
>>   	if (defmsg)
>>   		strvec_pushl(&cmd.args, "-F", defmsg, NULL);
>>   	else if (!(flags & EDIT_MSG))
> 
> This is straight-forward and obviously correct.
> 
>> diff --git a/t/t3428-rebase-signoff.sh b/t/t3428-rebase-signoff.sh
>> index e1b1e947647..fcecdf41978 100755
>> --- a/t/t3428-rebase-signoff.sh
>> +++ b/t/t3428-rebase-signoff.sh
>> @@ -27,6 +27,13 @@ first
>>   Signed-off-by: $(git var GIT_COMMITTER_IDENT | sed -e "s/>.*/>/")
>>   EOF
>>   
>> +# Expected signed off message after resolving the conflict
>> +cat >expected-signed-after-conflict <<EOF
>> +update file on side
>> +
>> +Signed-off-by: $(git var GIT_COMMITTER_IDENT | sed -e "s/>.*/>/")
>> +EOF
>> +
> 
> The t3428 script uses ancient style to prepare these expectations
> outside the test_expect_success blocks, and following the pattern
> is fine (until we clean-up the style of the whole script).  Cleaning
> it up is better done as a separate series, and when the dust settles
> after this fix lands.  #leftoverbits
> 
>>   # Expected commit message after rebase without --signoff (or with --no-signoff)
>>   cat >expected-unsigned <<EOF
>>   first
>> @@ -82,4 +89,30 @@ test_expect_success 'rebase -m --signoff fails' '
>>   	git cat-file commit HEAD | sed -e "1,/^\$/d" >actual &&
>>   	test_cmp expected-signed actual
>>   '
>> +
>> +test_expect_success 'rebase -i signs commits even if a conflict occurs' '
>> +	git branch -M main &&
>> +
>> +	git branch side &&
>> +	echo "b" >file &&
>> +	git add file &&
>> +	git commit -m"update file" &&
> 
> Have a SP between "-m" and its value, imitating existing tests.
> 
>> +	test_tick &&
> 
> Why?  Does the rest of the test depend on commits to have different
> timestamps?
> 
> If you want to use test_tick, the pattern to follow is to call it
> _before_ creating a commit, not after.  Up to this point in the
> original script nobody calls it (or test_commit that by default
> calls it), so "test_tick && git_commit" would be OK.
> 
> But if there is no dependency on exact sequence of timestamps,
> letting the commit share the same initial timestamp (hardcoded in
> test-lib.sh for better reproducibility) is more preferable, as use
> of test_tick sends a wrong message that the test results may be
> affected by the exact timestamps.
> 
>> +	test_must_fail git rebase -i --signoff main &&
>> +
>> +	echo "merged" >file &&
>> +	git add file &&
>> +	git rebase --continue &&
>> +
>> +	git cat-file commit HEAD | sed -e "1,/^\$/d" >actual &&
>> +	test_cmp expected-signed-after-conflict actual
> 
> Running any git command on the left hand side of a pipe is frowned
> upon, as it will hide exit status from it when it fails.
> 
> In this case, the primary thing we care about is that we have added
> the sign off that did not exist in the original, so I wonder
> 
> 	git cat-file commit HEAD >actual &&
> 	test_grep "Signed-off-by: " actual
> 
> would be sufficient?
> 
>> +'
>> +
>>   test_done
>> diff --git a/t/t3437/expected-squash-message b/t/t3437/expected-squash-message
>> index ab2434f90ed..d74af0bcf6b 100644
>> --- a/t/t3437/expected-squash-message
>> +++ b/t/t3437/expected-squash-message
>> @@ -48,4 +48,6 @@ edited 1
>>   edited 2
>>   
>>   edited 3
>> +
>> +Signed-off-by: Rebase Committer <[e-mail]>
>>   squashed
> 
> This defines the expected outcome from
> 
>      git rebase -ik --signoff A
> 
> in t3437-rebase-fixup-options.sh; I am not sure if the location the
> sign-off is inserted is sensible, though.
> 
> Thanks.
>

Phillip Wood April 4, 2024, 7:22 p.m. UTC | #4

On 04/04/2024 17:45, Junio C Hamano wrote:
> 	git show -s --format=%B HEAD >actual &&
> 	test_cmp expect actual
> 
> may be a good replacement.

we have test_commit_message to do that these days.

Best Wishes

Phillip

> But having said that, and then after having looked at the existing
> tests in the file, I see that it is littered with the same "do not
> run git on the upstream of the pipe" violation.  So let's not worry
> about this one.  The whole t3428 script needs to be cleaned up after
> the dust settles.
> 
> Thanks.
>

David Bimmler April 5, 2024, 11:50 a.m. UTC | #5

On 04.04.24 21:21, Phillip Wood wrote:

> I think it is more complicated than that. We do not use the "--signoff" 
> option of "git commit" because we do not want to append the 
> "Signed-off-by:" trailer when processing "fixup" and "squash" commands. 
> The trailer is appended to the commit message by the sequencer in 
> do_pick_commits(). The problem is that when we commit a conflict 
> resolution we end up using the original commit message rather than the 
> file containing the commit message that we would have used if we had not 
> stopped for conflicts. I've got some patches which need their commit 
> messages cleaning up at 
> https://github.com/phillipwood/git/commits/wip/fix-rebase-signoff-with-conflicts/ which use the correct file when committing conflict resolutions. I'll try and clean them up next week.

Took a look at that series, I agree that they fix the issue too. I 
wasn't aware of the intended behaviour for signoffs in squash and fixup, 
I'll thus not pursue this patch further.

Thanks for the context and review,
David

Patch

diff --git a/sequencer.c b/sequencer.c
index fa838f264f5..16595e26a17 100644
--- a/sequencer.c
+++ b/sequencer.c
@@ -1121,6 +1121,8 @@  static int run_git_commit(const char *defmsg,
 		strvec_pushf(&cmd.args, "-S%s", opts->gpg_sign);
 	else
 		strvec_push(&cmd.args, "--no-gpg-sign");
+	if (opts->signoff)
+		strvec_push(&cmd.args, "--signoff");
 	if (defmsg)
 		strvec_pushl(&cmd.args, "-F", defmsg, NULL);
 	else if (!(flags & EDIT_MSG))
diff --git a/t/t3428-rebase-signoff.sh b/t/t3428-rebase-signoff.sh
index e1b1e947647..fcecdf41978 100755
--- a/t/t3428-rebase-signoff.sh
+++ b/t/t3428-rebase-signoff.sh
@@ -27,6 +27,13 @@  first
 Signed-off-by: $(git var GIT_COMMITTER_IDENT | sed -e "s/>.*/>/")
 EOF
 
+# Expected signed off message after resolving the conflict
+cat >expected-signed-after-conflict <<EOF
+update file on side
+
+Signed-off-by: $(git var GIT_COMMITTER_IDENT | sed -e "s/>.*/>/")
+EOF
+
 # Expected commit message after rebase without --signoff (or with --no-signoff)
 cat >expected-unsigned <<EOF
 first
@@ -82,4 +89,30 @@  test_expect_success 'rebase -m --signoff fails' '
 	git cat-file commit HEAD | sed -e "1,/^\$/d" >actual &&
 	test_cmp expected-signed actual
 '
+
+test_expect_success 'rebase -i signs commits even if a conflict occurs' '
+	git branch -M main &&
+
+	git branch side &&
+	echo "b" >file &&
+	git add file &&
+	git commit -m"update file" &&
+	test_tick &&
+
+	git checkout side &&
+	echo "side" >file &&
+	git add file &&
+	git commit -m"update file on side" &&
+	test_tick &&
+
+	test_must_fail git rebase -i --signoff main &&
+
+	echo "merged" >file &&
+	git add file &&
+	git rebase --continue &&
+
+	git cat-file commit HEAD | sed -e "1,/^\$/d" >actual &&
+	test_cmp expected-signed-after-conflict actual
+'
+
 test_done
diff --git a/t/t3437/expected-squash-message b/t/t3437/expected-squash-message
index ab2434f90ed..d74af0bcf6b 100644
--- a/t/t3437/expected-squash-message
+++ b/t/t3437/expected-squash-message
@@ -48,4 +48,6 @@  edited 1
 edited 2
 
 edited 3
+
+Signed-off-by: Rebase Committer <[e-mail]>
 squashed