log: --remerge-diff needs to keep around commit parents

From: Johannes Schindelin <[email protected]>

To show a remerge diff, the merge needs to be recreated. For that to work, the merge base(s) need to be found, which means that the commits' parents have to be traversed until common ancestors are found (if any).

However, one optimization that hails all the way back to cb115748ec0d (Some more memory leak avoidance, 2006-06-17) is to release the commit's list of parents immediately after showing it. This can break the merge base computation.

Note that it matters more clearly when traversing the commits in
reverse: In that instance, if a parent of a merge commit has been shown as part of the git log command, by the time the merge commit's diff needs to be computed, that parent commit's list of parent commits will have been set to NULL and as a result no merge base will be found.

Let's fix this by special-casing the remerge_diff mode, similar to what we did with reflogs in f35650dff6a4 (log: do not free parents when walking reflog, 2017-07-07).

---
log: --remerge-diff needs to keep around commit parents

This fixes a bug that is pretty much as old as the remerge-diff machinery itself. I noticed it while writing my reply to Hannes Sixt's concerns about my range-diff --diff-merges=<mode> patch
(https://lore.kernel.org/git/[email protected]/).

Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-1825%2Fdscho%2Flog-remerge-diff-needs-commit-parents-v1
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-1825/dscho/log-remerge-diff-needs-commit-parents-v1
Pull-Request: https://github.com/gitgitgadget/git/pull/1825

 builtin/log.c           | 2 +-
 t/t4069-remerge-diff.sh | 7 +++++++
 2 files changed, 8 insertions(+), 1 deletion(-)

index c0a8bb95e98..a297c6caf59 100644
@@ -522,7 +522,7 @@ static int cmd_log_walk_no_free(struct rev_info *rev)
 			 * but we didn't actually show the commit.
 			 */
 			rev->max_count++;
-		if (!rev->reflog_info) {
+		if (!rev->reflog_info && !rev->remerge_diff) {
 			/*
 			 * We may show a given commit multiple times when
 			 * walking the reflogs.

index 07323ebafe0..a68c6bfa036 100755
@@ -317,4 +317,11 @@ test_expect_success 'remerge-diff turns off history simplification' '
 	test_cmp expect actual
 '
 
+test_expect_success 'remerge-diff with --reverse' '
+	git log -1 --remerge-diff --oneline ab_resolution^ >expect &&
+	git log -1 --remerge-diff --oneline ab_resolution >>expect &&
+	git log -2 --remerge-diff --oneline ab_resolution --reverse >actual &&
+	test_cmp expect actual
+'
+
 test_done

base-commit: bea9ecd24b0c3bf06cab4a851694fe09e7e51408

gitgitgadget

Wow, nice find, and I appreciate the nice simple testcase demonstrating what had been the problem.

"Johannes Schindelin via GitGitGadget" <[email protected]>

Ouch.

I am curious about "more clearly" in the above, though. Do we also lose parents before a base can be computed during a forward traversal? Unlike the reflog walk, we won't show the same commit twice, so unless we are traversing some of the history of our parents *and* showing these ancestors found there _before_ we need to show the merge in question, I do not think the issue would surface during a forward traversal. So the problematic case is when we are not doing topological display, and traversing down to a commit with skewed timestamp causes it to be shown (and lose its parents due to memory optimization) because it appears much newer than our merge, and there is an ancestry chain leading to it that passes commits other than our merge, or something?

OK.

Ah, we probably should. Especially if the newly discovered failure mode is not due to having to show the same commit multiple times.

During reflog walk, we may show the same commit more than once, so do not discard the parents. There may be a merge commit that is descendent from the current commit, and in order to show it with remerge-diff enabled, we need its ancestry chain, including our parents, to compute the merge base.

or something? It is still not clear to me in what scenario a merge commit, in a forward traversal, is shown _after_ one of its ancestor commit is shown (and due to the memory optimization, loses its parents).

Thanks, both.

Hi Junio,

I consider these examples less clear, but they are still affected:

	git show --remerge-diff v2.45.2^0
	vs
	git show --remerge-diff v2.44.2^0 v2.45.2^0

	git show --remerge-diff v2.45.1~1
	vs
	git log --topo-order --first-parent --remerge-diff v2.44.2 v2.45.1

Concretely, these diffs should be empty, but are not:

	git diff --no-index \
		<(git show --remerge-diff v2.45.2^0 | sed 1d) \
		<(git show --remerge-diff v2.44.2^0 v2.45.2^0 | sed 1,/^commit/d)

	and

	git diff --no-index \
		<(git show --remerge-diff v2.45.1~1 | grep -v ^commit) \
		<(git log --topo-order --first-parent -6 --remerge-diff v2.44.2 v2.45.1 |
		  sed '1,/^commit 1c00f92eb5ee4a48ab615eefa41f2dd6024d43bc/d;/^commit/,$d')

No --reverse required, not even clock skew.

Ciao,
Johannes

From: Johannes Schindelin <[email protected]>

To show a remerge diff, the merge needs to be recreated. For that to work, the merge base(s) need to be found, which means that the commits' parents have to be traversed until common ancestors are found (if any).

However, one optimization that hails all the way back to cb115748ec0d (Some more memory leak avoidance, 2006-06-17) is to release the commit's list of parents immediately after showing it. This can break the merge base computation.

Note that it matters more clearly when traversing the commits in
reverse: In that instance, if a parent of a merge commit has been shown as part of the git log command, by the time the merge commit's diff needs to be computed, that parent commit's list of parent commits will have been set to NULL and as a result no merge base will be found.

Let's fix this by special-casing the remerge_diff mode, similar to what we did with reflogs in f35650dff6a4 (log: do not free parents when walking reflog, 2017-07-07).

---
log: --remerge-diff needs to keep around commit parents

This fixes a bug that is pretty much as old as the remerge-diff machinery itself. I noticed it while writing my reply to Hannes Sixt's concerns about my range-diff --diff-merges=<mode> patch
(https://lore.kernel.org/git/[email protected]/).

Changes since v1:

* Amended the code comment of the affected conditional code block in harmony with the newly-added condition.

Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-1825%2Fdscho%2Flog-remerge-diff-needs-commit-parents-v2
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-1825/dscho/log-remerge-diff-needs-commit-parents-v2
Pull-Request: https://github.com/gitgitgadget/git/pull/1825

Range-diff vs v1:

1: 68238315c14 ! 1: afff27f8222 log: --remerge-diff needs to keep around commit parents
@@ builtin/log.c: static int cmd_log_walk_no_free(struct rev_info *rev)
+ if (!rev->reflog_info && !rev->remerge_diff) {
/*
* We may show a given commit multiple times when
- * walking the reflogs.
+- * walking the reflogs.
++ * walking the reflogs. Therefore we still need it.
++ *
++ * Likewise, we potentially still need the parents
++ * of * already shown commits to determine merge
++ * bases when showing remerge diffs.
+ */
+ free_commit_buffer(the_repository->parsed_objects,
+ commit);

## t/t4069-remerge-diff.sh ##
@@ t/t4069-remerge-diff.sh: test_expect_success 'remerge-diff turns off history simplification' '

 builtin/log.c           | 8 ++++++--
 t/t4069-remerge-diff.sh | 7 +++++++
 2 files changed, 13 insertions(+), 2 deletions(-)

index c0a8bb95e98..2f88a3e607d 100644
@@ -522,10 +522,14 @@ static int cmd_log_walk_no_free(struct rev_info *rev)
 			 * but we didn't actually show the commit.
 			 */
 			rev->max_count++;
-		if (!rev->reflog_info) {
+		if (!rev->reflog_info && !rev->remerge_diff) {
 			/*
 			 * We may show a given commit multiple times when
-			 * walking the reflogs.
+			 * walking the reflogs. Therefore we still need it.
+			 *
+			 * Likewise, we potentially still need the parents
+			 * of * already shown commits to determine merge
+			 * bases when showing remerge diffs.
 			 */
 			free_commit_buffer(the_repository->parsed_objects,
 					   commit);

index 07323ebafe0..a68c6bfa036 100755
@@ -317,4 +317,11 @@ test_expect_success 'remerge-diff turns off history simplification' '
 	test_cmp expect actual
 '
 
+test_expect_success 'remerge-diff with --reverse' '
+	git log -1 --remerge-diff --oneline ab_resolution^ >expect &&
+	git log -1 --remerge-diff --oneline ab_resolution >>expect &&
+	git log -2 --remerge-diff --oneline ab_resolution --reverse >actual &&
+	test_cmp expect actual
+'
+
 test_done

base-commit: bea9ecd24b0c3bf06cab4a851694fe09e7e51408

gitgitgadget

As we often tell contributors, questions in reviewer comments should be also answered in an updated patch, so that future readers of "git log", who cannot ask direct questions to the author of the patch, do not have to ask the same questions. Can we add an explanation how this affects forward traversal in a three-line paragraph?

Thanks.

From: Johannes Schindelin <[email protected]>

To show a remerge diff, the merge needs to be recreated. For that to work, the merge base(s) need to be found, which means that the commits' parents have to be traversed until common ancestors are found (if any).

However, one optimization that hails all the way back to cb115748ec0d (Some more memory leak avoidance, 2006-06-17) is to release the commit's list of parents immediately after showing it _and to set that parent list to NULL_. This can break the merge base computation.

This problem is most obvious when traversing the commits in reverse: In that instance, if a parent of a merge commit has been shown as part of the git log command, by the time the merge commit's diff needs to be computed, that parent commit's list of parent commits will have been set to NULL and as a result no merge base will be found (even if one should be found).

Traversing commits in reverse is far from the only circumstance in which this problem occurs, though. There are many avenues to traversing at least one commit in the revision walk that will later be part of a merge base computation, for example when not even walking any revisions in git show <merge1> <merge2> where <merge1> is part of the commit graph between the parents of <merge2>.

Another way to force a scenario where a commit is traversed before it has to be traversed again as part of a merge base computation is to start with two revisions (where the first one is reachable from the second but not in a first-parent ancestry) and show the commit log with --topo-order and --first-parent.

Let's fix this by special-casing the remerge_diff mode, similar to what we did with reflogs in f35650dff6a4 (log: do not free parents when walking reflog, 2017-07-07).

---
log: --remerge-diff needs to keep around commit parents

This fixes a bug that is pretty much as old as the remerge-diff machinery itself. I noticed it while writing my reply to Hannes Sixt's concerns about my range-diff --diff-merges=<mode> patch
(https://lore.kernel.org/git/[email protected]/).

Changes since v2:

* Amended the commit message to illustrate why reverse walks are not the only instances where this fix is relevant.

Changes since v1:

* Amended the code comment of the affected conditional code block in harmony with the newly-added condition.

Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-1825%2Fdscho%2Flog-remerge-diff-needs-commit-parents-v3
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-1825/dscho/log-remerge-diff-needs-commit-parents-v3
Pull-Request: https://github.com/gitgitgadget/git/pull/1825

Range-diff vs v2:

1: afff27f8222 ! 1: ab75a565b4c log: --remerge-diff needs to keep around commit parents
@@ Commit message
work, the merge base(s) need to be found, which means that the commits' parents have to be traversed until common ancestors are found (if any).

- However, one optimization that hails all the way back to
- cb115748ec0d (Some more memory leak avoidance, 2006-06-17) is to release
- the commit's list of parents immediately after showing it. This can break
- the merge base computation.
-
- Note that it matters more clearly when traversing the commits in
- reverse: In that instance, if a parent of a merge commit has been shown
- as part of the git log command, by the time the merge commit's diff
- needs to be computed, that parent commit's list of parent commits will
- have been set to NULL and as a result no merge base will be found.
+ However, one optimization that hails all the way back to cb115748ec0d
+ (Some more memory leak avoidance, 2006-06-17) is to release the commit's
+ list of parents immediately after showing it _and to set that parent
+ list to NULL_. This can break the merge base computation.
+
+ This problem is most obvious when traversing the commits in reverse: In
+ that instance, if a parent of a merge commit has been shown as part of
+ the git log command, by the time the merge commit's diff needs to be
+ computed, that parent commit's list of parent commits will have been set
+ to NULL and as a result no merge base will be found (even if one
+ should be found).
+
+ Traversing commits in reverse is far from the only circumstance in which
+ this problem occurs, though. There are many avenues to traversing at
+ least one commit in the revision walk that will later be part of a merge
+ base computation, for example when not even walking any revisions in
+ git show <merge1> <merge2> where <merge1> is part of the commit
+ graph between the parents of <merge2>.
+
+ Another way to force a scenario where a commit is traversed before it
+ has to be traversed again as part of a merge base computation is to
+ start with two revisions (where the first one is reachable from the
+ second but not in a first-parent ancestry) and show the commit log with
+ --topo-order and --first-parent.

Let's fix this by special-casing the remerge_diff mode, similar to what we did with reflogs in f35650dff6a4 (log: do not free parents when

 builtin/log.c           | 8 ++++++--
 t/t4069-remerge-diff.sh | 7 +++++++
 2 files changed, 13 insertions(+), 2 deletions(-)

index c0a8bb95e98..2f88a3e607d 100644
@@ -522,10 +522,14 @@ static int cmd_log_walk_no_free(struct rev_info *rev)
 			 * but we didn't actually show the commit.
 			 */
 			rev->max_count++;
-		if (!rev->reflog_info) {
+		if (!rev->reflog_info && !rev->remerge_diff) {
 			/*
 			 * We may show a given commit multiple times when
-			 * walking the reflogs.
+			 * walking the reflogs. Therefore we still need it.
+			 *
+			 * Likewise, we potentially still need the parents
+			 * of * already shown commits to determine merge
+			 * bases when showing remerge diffs.
 			 */
 			free_commit_buffer(the_repository->parsed_objects,
 					   commit);

index 07323ebafe0..a68c6bfa036 100755
@@ -317,4 +317,11 @@ test_expect_success 'remerge-diff turns off history simplification' '
 	test_cmp expect actual
 '
 
+test_expect_success 'remerge-diff with --reverse' '
+	git log -1 --remerge-diff --oneline ab_resolution^ >expect &&
+	git log -1 --remerge-diff --oneline ab_resolution >>expect &&
+	git log -2 --remerge-diff --oneline ab_resolution --reverse >actual &&
+	test_cmp expect actual
+'
+
 test_done

base-commit: bea9ecd24b0c3bf06cab4a851694fe09e7e51408

gitgitgadget

"Johannes Schindelin via GitGitGadget" <[email protected]>

Nicely described.