fetch.writeCommitGraph fails on first fetch

While dogfooding, Johannes found a bug in the fetch.writeCommitGraph config behavior. While his example initially happened during a clone with --recurse-submodules, we found that it is actually a problem with the first fetch after a new clone and no existing commit-graph file:

$ git clone <url> test
$ cd test
$ git -c fetch.writeCommitGraph=true fetch origin
Computing commit graph generation numbers: 100% (12/12), done.
BUG: commit-graph.c:886: missing parent <hash1> for commit <hash2>
Aborted (core dumped)

In the repo I had cloned, there were really 60 commits to scan, but only 12 were in the list to write when calling compute_generation_numbers(). A commit in the list expects to see a parent, but that parent is not in the list.

The details above are the start of the commit message for [PATCH 1/1]. I tried to include as much detail as I could for how I investigated the problem and why I think this is the right solution.

I would like help creating a test that demonstrates this problem. It does not appear to trigger on the simplest examples. The simple example I used for my testing was https://github.com/derrickstolee/numbers

Thanks, -Stolee

/cc [email protected], [email protected]

Derrick Stolee (1):
commit-graph: fix writing first commit-graph during fetch

 commit-graph.c | 11 +++++++----
 commit-reach.c |  1 -
 object.h       |  3 ++-
 3 files changed, 9 insertions(+), 6 deletions(-)

base-commit: d966095db01190a2196e31195ea6fa0c722aa732
Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-415%2Fderrickstolee%2Ffetch-first-write-fail-v1
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-415/derrickstolee/fetch-first-write-fail-v1
Pull-Request: https://github.com/gitgitgadget/git/pull/415

gitgitgadget

From: Derrick Stolee <[email protected]>

While dogfooding, Johannes found a bug in the fetch.writeCommitGraph config behavior. While his example initially happened during a clone with --recurse-submodules, we found that it is actually a problem with the first fetch after a new clone and no existing commit-graph file:

	$ git clone <url> test
	$ cd test
	$ git -c fetch.writeCommitGraph=true fetch origin
	Computing commit graph generation numbers: 100% (12/12), done.
	BUG: commit-graph.c:886: missing parent <hash1> for commit <hash2>
	Aborted (core dumped)

In the repo I had cloned, there were really 60 commits to scan, but only 12 were in the list to write when calling compute_generation_numbers(). A commit in the list expects to see a parent, but that parent is not in the list.

As an initial test, I converted the code in builtin/fetch.c that calls write_commit_graph_reachable() to instead launch a "git commit-graph write --reachable --split" process. That code worked, but is not how we want the feature to work long-term.

That test did demonstrate that the issue must be something to do with internal state of the 'git fetch' process.

The write_commit_graph() method in commit-graph.c ensures the commits we plan to write are "closed under reachability" using close_reachable(). This method walks from the input commits, and uses the UNINTERESTING flag to mark which commits have already been visited. This allows the walk to take O(N) time, where N is the number of commits, instead of O(P) time, where P is the number of paths. (The number of paths can be exponential in the number of commits.)

However, the UNINTERESTING flag is used in lots of places in the codebase. This flag usually means some barrier to stop a commit walk, such as in revision-walking to compare histories. It is not often cleared after the walk completes because the starting points of those walks do not have the UNINTERESTING flag, and clear_commit_marks() would stop immediately.

This is happening during a 'git fetch' call with a remote. The fetch negotiation is comparing the remote refs with the local refs and marking some commits as UNINTERESTING.

You may ask: did this feature ever work at all? Yes, it did, as long as you had a commit-graph covering all of your local refs. My testing was unfortunately limited to this scenario. The UNINTERESTING commits are always part of the "old" commit-graph, and when we add new commits to a top layer of the commit-graph chain those are not needed. If we happen to merge layers of the chain, then the commits are added as a list, not using a commit walk. Further, the test added for this config option in t5510-fetch.sh uses local filesystem clones, which somehow avoids this logic.

I tested running clear_commit_marks_many() to clear the UNINTERESTING flag inside close_reachable(), but the tips did not have the flag, so that did nothing.

Instead, I finally arrived on the conclusion that I should use a flag that is not used in any other part of the code. In commit-reach.c, a number of flags were defined for commit walk algorithms. The REACHABLE flag seemed like it made the most sense, and it seems it was not actually used in the file.

Add the REACHABLE flag to commit-graph.c and use it instead of UNINTERESTING in close_reachable(). This fixes the bug in manual testing.

I have failed to produce a test using the file:// protocol that demonstrates this bug.

---
 commit-graph.c | 11 +++++++----
 commit-reach.c |  1 -
 object.h       |  3 ++-
 3 files changed, 9 insertions(+), 6 deletions(-)

index fc4a43b8d6..0aea7b2ae5 100644
@@ -41,6 +41,9 @@
 #define GRAPH_MIN_SIZE (GRAPH_HEADER_SIZE + 4 * GRAPH_CHUNKLOOKUP_WIDTH \
 			+ GRAPH_FANOUT_SIZE + the_hash_algo->rawsz)
 
+/* Remember to update object flag allocation in object.h */
+#define REACHABLE       (1u<<15)
+
 char *get_commit_graph_filename(const char *obj_dir)
 {
 	char *filename = xstrfmt("%s/info/commit-graph", obj_dir);

@@ -1030,11 +1033,11 @@ static void add_missing_parents(struct write_commit_graph_context *ctx, struct c
 {
 	struct commit_list *parent;
 	for (parent = commit->parents; parent; parent = parent->next) {
-		if (!(parent->item->object.flags & UNINTERESTING)) {
+		if (!(parent->item->object.flags & REACHABLE)) {
 			ALLOC_GROW(ctx->oids.list, ctx->oids.nr + 1, ctx->oids.alloc);
 			oidcpy(&ctx->oids.list[ctx->oids.nr], &(parent->item->object.oid));
 			ctx->oids.nr++;
-			parent->item->object.flags |= UNINTERESTING;
+			parent->item->object.flags |= REACHABLE;
 		}
 	}
 }

@@ -1052,7 +1055,7 @@ static void close_reachable(struct write_commit_graph_context *ctx)
 		display_progress(ctx->progress, i + 1);
 		commit = lookup_commit(ctx->r, &ctx->oids.list[i]);
 		if (commit)
-			commit->object.flags |= UNINTERESTING;
+			commit->object.flags |= REACHABLE;
 	}
 	stop_progress(&ctx->progress);
 

@@ -1089,7 +1092,7 @@ static void close_reachable(struct write_commit_graph_context *ctx)
 		commit = lookup_commit(ctx->r, &ctx->oids.list[i]);
 
 		if (commit)
-			commit->object.flags &= ~UNINTERESTING;
+			commit->object.flags &= ~REACHABLE;
 	}
 	stop_progress(&ctx->progress);
 }

index 3ea174788a..4ca7e706a1 100644
@@ -10,7 +10,6 @@
 #include "commit-reach.h"
 
 /* Remember to update object flag allocation in object.h */
-#define REACHABLE       (1u<<15)
 #define PARENT1		(1u<<16)
 #define PARENT2		(1u<<17)
 #define STALE		(1u<<18)

index 0120892bbd..25f5ab3d54 100644
@@ -68,7 +68,8 @@ struct object_array {
  * bisect.c:                                        16
  * bundle.c:                                        16
  * http-push.c:                                     16-----19
- * commit-reach.c:                                15-------19
+ * commit-graph.c:                                15
+ * commit-reach.c:                                  16-----19
  * sha1-name.c:                                              20
  * list-objects-filter.c:                                      21
  * builtin/fsck.c:           0--3

gitgitgadget

Oof. Nicely explained, and your fix makes sense.

The global-ness of revision flags always makes me nervous about doing more things in-process (this isn't the first such bug we've had).

I have a dream of converting most uses of flags into using a commit-slab. That provides cheap access to an auxiliary structure, so each traversal, etc, could keep its own flag structure. I'm not sure if it would have a performance impact, though. Even though it's O(1), it is an indirect lookup, which could have some memory-access impact (though my guess is it would be lost in the noise).

One of the sticking points is that all object types, not just commits, use flags. But we only assign slab ids to commits. I noticed recently that "struct object" has quite a few spare bits in it these days, because the switch to a 32-byte oid means 64-bit machines now have an extra 4 bytes of padding. I wonder if we could use that to store an index field.

Anyway, that's getting far off the topic; clearly we need a fix in the meantime, and what you have here looks good to me.

Another option would be clear_object_flags(), which just walks all of the in-memory structs. Your REACHABLE solution is cheaper, though.

Yeah, being able to remove it from commit-reach.c surprised me for a moment. To further add to the confusion, builtin/fsck.c has its own REACHABLE flag (with a different bit and a totally different purpose). I don't think there's any practical problem there, though.

Hmm, from the description, it sounds like it should be easy. I might poke at it a bit.

-Peff

Hmph. I can reproduce it here, but it seems to depend on the repository. If I do this:

index ecabbe1616..8d473a456f 100755
@@ -583,6 +583,14 @@ test_expect_success 'fetch.writeCommitGraph' '
 	)
 '
 
+test_expect_success 'fetch.writeCommitGraph with a bigger repo' '
+	git clone "$TEST_DIRECTORY/.." repo &&
+	(
+		cd repo &&
+		git -c fetch.writeCommitGraph fetch origin
+	)
+'
+
 # configured prune tests
 
 set_config_tristate () {

it reliably triggers the bug. But if I make a synthetic repo, even it has a lot of commits (thousands or more), it doesn't trigger. I thought maybe it had to do with having commits that were not at tips (since the tip ones presumably _are_ fed into the graph generation process). But that doesn't seem to help.

Puzzling...

-Peff

Submodules?

  $ cd ~/src/git/
  $ git quotelog 86cfd61e6b
  86cfd61e6b (sha1dc: optionally use sha1collisiondetection as a submodule, 2017-07-01)
  $ git init --bare good.git
  Initialized empty Git repository in /home/szeder/src/git/good.git/
  $ git push -q good.git 86cfd61e6b^:refs/heads/master
  $ git clone good.git good-clone
  Cloning into 'good-clone'...
  done.
  $ git -c fetch.writeCommitGraph -C good-clone fetch origin
  Computing commit graph generation numbers: 100% (46958/46958), done.
  $ git init --bare bad.git
  Initialized empty Git repository in /home/szeder/src/git/bad.git/
  $ git push -q bad.git 86cfd61e6b:refs/heads/master
  $ git clone bad.git bad-clone
  Cloning into 'bad-clone'...
  done.
  $ git -c fetch.writeCommitGraph -C bad-clone fetch origin
  Computing commit graph generation numbers: 100% (1/1), done.
  BUG: commit-graph.c:886: missing parent 9936c1b52a39fa14fca04f937df3e75f7498ac66 for commit 86cfd61e6bc12745751c43b4f69886b290cd85cb
  Aborted

In the cover letter Derrick mentioned that he used
https://github.com/derrickstolee/numbers for testing, and that repo has a submodule as well.

I completely forgot that I put a submodule in that repo. It makes sense that the Git repo also has one. Thanks for the test! I'll get it into the test suite tomorrow.

-Stolee

Ah, right. Good catch Gábor. The test below fails for me without your patch, and succeeds with it (the extra empty commit is necessary so that we have a parent).

I admit I am puzzled, though, _why_ the presence of the submodule matters. That is, from your explanation, I thought the issue was simply that fetch walked (and marked) some commits, and the flags overlapped with what the commit-graph code expected.

I could guess that the presence of the submodule triggers some analysis for --recurse-submodules. But then we don't actually recurse (maybe because they're not activated? In which case maybe we shouldn't be doing that extra walk to look for submodules if there aren't any activated ones in our local repo).

I'd also wonder if it would be possible to trigger in another way (say, due to want/have negotiation, though I guess most of the walking there is done on the server side). But it may not be worth digging too far into it.

-Peff

---

index ecabbe1616..1b092fec0b 100755
@@ -583,6 +583,21 @@ test_expect_success 'fetch.writeCommitGraph' '
 	)
 '
 
+test_expect_success 'fetch.writeCommitGraph with a submodule' '
+	git init has-submodule &&
+	(
+		cd has-submodule &&
+		git commit --allow-empty -m parent &&
+		git submodule add ../three &&
+		git commit -m "add submodule"
+	) &&
+	git clone has-submodule submod-clone &&
+	(
+		cd submod-clone &&
+		git -c fetch.writeCommitGraph fetch origin
+	)
+'
+
 # configured prune tests
 
 set_config_tristate () {

Indeed, that seems to be it. If I do this:

git init repo
cd repo
cat >.gitmodules <<\EOF
[submodule "foo"]
path = foo
url = https://example.com
EOF
time git fetch /path/to/git.git

then we end up traversing the whole git.git history a second time, even though we should know off the bat that there are no active submodules that we would recurse to.

Doing this makes the problem go away:

index 0f199c5137..0db2f18b93 100644
@@ -1193,7 +1193,7 @@ static void calculate_changed_submodule_paths(struct repository *r,
 	struct string_list_item *name;
 
 	/* No need to check if there are no submodules configured */
-	if (!submodule_from_path(r, NULL, NULL))
+	if (!is_submodule_active(r, NULL))
 		return;
 
 	argv_array_push(&argv, "--"); /* argv[0] program name */

but causes some tests to fail (I think that in some cases we're supposed to auto-initialize, and we'd probably need to cover that case, too).

All of this is outside of your fix, of course, but:

1. I'm satisfied now that I understand why the test triggers the problem.

2. You may want have a real activated submodule in your test. Right now we'll trigger the submodule-recursion check even without that, but in the future we might do something like the hunk above. In which case your test wouldn't be checking anything interesting anymore.

-Peff

UPDATE for V2: We now know the full repro, and a test is added. Thanks Szeder and Peff for your insights!

While dogfooding, Johannes found a bug in the fetch.writeCommitGraph config behavior. While his example initially happened during a clone with --recurse-submodules, (UPDATE) and the submodule is important, but --recurse-submodules is not:

$ git clone <url> test
$ cd test
$ git -c fetch.writeCommitGraph=true fetch origin
Computing commit graph generation numbers: 100% (12/12), done.
BUG: commit-graph.c:886: missing parent <hash1> for commit <hash2>
Aborted (core dumped)

In the repo I had cloned, there were really 60 commits to scan, but only 12 were in the list to write when calling compute_generation_numbers(). A commit in the list expects to see a parent, but that parent is not in the list. The simple example I used for my testing was
https://github.com/derrickstolee/numbers. Thie repo HAS A SUBMODULE, I just forgot. Sorry for derailing the investigation somewhat.

The details above are the start of the commit message for [PATCH 1/2], including a test that fails when fetching after cloning a repo with a submodule.

In [PATCH 2/2], I actually have the fix. I tried to include as much detail as I could for how I investigated the problem and why I think this is the right solution. I added details that have come from the on-list discussion, including what the submodule code is doing and why REACHABLE is no longer used in commit-reach.c.

Thanks, -Stolee

Derrick Stolee (2):
t5510-fetch.sh: demonstrate fetch.writeCommitGraph bug
commit-graph: fix writing first commit-graph during fetch

 commit-graph.c   | 11 +++++++----
 commit-reach.c   |  1 -
 object.h         |  3 ++-
 t/t5510-fetch.sh | 17 +++++++++++++++++
 4 files changed, 26 insertions(+), 6 deletions(-)

base-commit: d966095db01190a2196e31195ea6fa0c722aa732
Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-415%2Fderrickstolee%2Ffetch-first-write-fail-v2
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-415/derrickstolee/fetch-first-write-fail-v2
Pull-Request: https://github.com/gitgitgadget/git/pull/415

Range-diff vs v1:

 -:  ---------- > 1:  6ac0a05746 t5510-fetch.sh: demonstrate fetch.writeCommitGraph bug
 1:  a1e5280d4b ! 2:  ca59b118f1 commit-graph: fix writing first commit-graph during fetch

@@ -2,10 +2,13 @@

commit-graph: fix writing first commit-graph during fetch

- While dogfooding, Johannes found a bug in the fetch.writeCommitGraph
- config behavior. While his example initially happened during a clone
- with --recurse-submodules, we found that it is actually a problem with
- the first fetch after a new clone and no existing commit-graph file:
+ The previous commit includes a failing test for an issue around
+ fetch.writeCommitGraph and fetching in a repo with a submodule. Here, we
+ fix that bug and set the test to "test_expect_success".
+
+ The prolem arises with this set of commands when the remote repo at
+ <url> has a submodule. Note that --recurse-submodules is not needed to
+ demonstrate the bug.

$ git clone <url> test
$ cd test
@@ -14,12 +17,7 @@
BUG: commit-graph.c:886: missing parent <hash1> for commit <hash2> Aborted (core dumped)

- In the repo I had cloned, there were really 60 commits to scan, but
- only 12 were in the list to write when calling
- compute_generation_numbers(). A commit in the list expects to see a
- parent, but that parent is not in the list.
-
- As an initial test, I converted the code in builtin/fetch.c that calls
+ As an initial fix, I converted the code in builtin/fetch.c that calls
write_commit_graph_reachable() to instead launch a "git commit-graph write --reachable --split" process. That code worked, but is not how we want the feature to work long-term.
@@ -60,20 +58,28 @@
flag inside close_reachable(), but the tips did not have the flag, so that did nothing.

+ It turns out that the calculate_changed_submodule_paths() method is at
+ fault. Thanks, Peff, for pointing out this detail! More specifically,
+ for each submodule, the collect_changed_submodules() runs a revision
+ walk to essentially do file-history on the list of submodules. That
+ revision walk marks commits UNININTERESTING if they are simiplified away
+ by not changing the submodule.
+
Instead, I finally arrived on the conclusion that I should use a flag that is not used in any other part of the code. In commit-reach.c, a number of flags were defined for commit walk algorithms. The REACHABLE flag seemed like it made the most sense, and it seems it was not
- actually used in the file.
+ actually used in the file. The REACHABLE flag was used in early versions
+ of commit-reach.c, but was removed by 4fbcca4 (commit-reach: make
+ can_all_from_reach... linear, 2018-07-20).

Add the REACHABLE flag to commit-graph.c and use it instead of UNINTERESTING in close_reachable(). This fixes the bug in manual testing.

- I have failed to produce a test using the file:// protocol that
- demonstrates this bug.
-

+ Helped-by: Jeff King <[email protected]>
+ Helped-by: Szeder Gábor <[email protected]>

diff --git a/commit-graph.c b/commit-graph.c
@@ -147,3 +153,16 @@
* sha1-name.c: 20
* list-objects-filter.c: 21
* builtin/fsck.c: 0--3
+
+ diff --git a/t/t5510-fetch.sh b/t/t5510-fetch.sh
+ --- a/t/t5510-fetch.sh
+ +++ b/t/t5510-fetch.sh
+@@
+ )
+ '
+
+-test_expect_failure 'fetch.writeCommitGraph with submodules' ' ++test_expect_success 'fetch.writeCommitGraph with submodules' '
+ pwd="$(pwd)" &&
+ git clone dups super &&
+ (

gitgitgadget

From: Derrick Stolee <[email protected]>

While dogfooding, Johannes found a bug in the fetch.writeCommitGraph config behavior. His example initially happened during a clone with --recurse-submodules, we found that this happens with the first fetch after cloning a repository that contains a submodule:

	$ git clone <url> test
	$ cd test
	$ git -c fetch.writeCommitGraph=true fetch origin
	Computing commit graph generation numbers: 100% (12/12), done.
	BUG: commit-graph.c:886: missing parent <hash1> for commit <hash2>
	Aborted (core dumped)

In the repo I had cloned, there were really 60 commits to scan, but only 12 were in the list to write when calling compute_generation_numbers(). A commit in the list expects to see a parent, but that parent is not in the list.

A follow-up will fix the bug, but first we create a test that demonstrates the problem.

I used "test_expect_failure" for the entire test instead of "test_must_fail" only on the command that I expect to fail. This is because the BUG() returns an exit code so test_must_fail complains.

---
 t/t5510-fetch.sh | 17 +++++++++++++++++
 1 file changed, 17 insertions(+)

index ecabbe1616..e8ae3af0b6 100755
@@ -583,6 +583,23 @@ test_expect_success 'fetch.writeCommitGraph' '
 	)
 '
 
+test_expect_failure 'fetch.writeCommitGraph with submodules' '
+	pwd="$(pwd)" &&
+	git clone dups super &&
+	(
+		cd super &&
+		git submodule add "file://$pwd/three" &&
+		git commit -m "add submodule"
+	) &&
+	git clone "super" writeError &&
+	(
+		cd writeError &&
+		test_path_is_missing .git/objects/info/commit-graphs/commit-graph-chain &&
+		git -c fetch.writeCommitGraph=true fetch origin &&
+		test_path_is_file .git/objects/info/commit-graphs/commit-graph-chain
+	)
+'
+
 # configured prune tests
 
 set_config_tristate () {

gitgitgadget

I don't think this paragraph is necessary; using 'test_expect_failure' is the way to demonstrate a known breakage.

'test_must_fail' should only be used when the failure is the desired behavior of a git command. (I used the word "desired" here, because you just used the word "expect" above in the sense that "I expect it to fail, because I know it's buggy, and I want to demonstrate that bug")

I guess that I prefer pointing out which line of the test fails, and making that part of the test (that must otherwise pass). However, you are right that test_expect_failure does a good job of communicating that the test script shows an existing bug. Those are not always cleaned up immediately, but at least we can find them easily.

True, that would be better. Thanks!

Good point.

Thanks!
-Stolee

From: Derrick Stolee <[email protected]>

The previous commit includes a failing test for an issue around fetch.writeCommitGraph and fetching in a repo with a submodule. Here, we fix that bug and set the test to "test_expect_success".

The prolem arises with this set of commands when the remote repo at <url> has a submodule. Note that --recurse-submodules is not needed to demonstrate the bug.

	$ git clone <url> test
	$ cd test
	$ git -c fetch.writeCommitGraph=true fetch origin
	Computing commit graph generation numbers: 100% (12/12), done.
	BUG: commit-graph.c:886: missing parent <hash1> for commit <hash2>
	Aborted (core dumped)

As an initial fix, I converted the code in builtin/fetch.c that calls write_commit_graph_reachable() to instead launch a "git commit-graph write --reachable --split" process. That code worked, but is not how we want the feature to work long-term.

That test did demonstrate that the issue must be something to do with internal state of the 'git fetch' process.

The write_commit_graph() method in commit-graph.c ensures the commits we plan to write are "closed under reachability" using close_reachable(). This method walks from the input commits, and uses the UNINTERESTING flag to mark which commits have already been visited. This allows the walk to take O(N) time, where N is the number of commits, instead of O(P) time, where P is the number of paths. (The number of paths can be exponential in the number of commits.)

However, the UNINTERESTING flag is used in lots of places in the codebase. This flag usually means some barrier to stop a commit walk, such as in revision-walking to compare histories. It is not often cleared after the walk completes because the starting points of those walks do not have the UNINTERESTING flag, and clear_commit_marks() would stop immediately.

This is happening during a 'git fetch' call with a remote. The fetch negotiation is comparing the remote refs with the local refs and marking some commits as UNINTERESTING.

You may ask: did this feature ever work at all? Yes, it did, as long as you had a commit-graph covering all of your local refs. My testing was unfortunately limited to this scenario. The UNINTERESTING commits are always part of the "old" commit-graph, and when we add new commits to a top layer of the commit-graph chain those are not needed. If we happen to merge layers of the chain, then the commits are added as a list, not using a commit walk. Further, the test added for this config option in t5510-fetch.sh uses local filesystem clones, which somehow avoids this logic.

I tested running clear_commit_marks_many() to clear the UNINTERESTING flag inside close_reachable(), but the tips did not have the flag, so that did nothing.

It turns out that the calculate_changed_submodule_paths() method is at fault. Thanks, Peff, for pointing out this detail! More specifically, for each submodule, the collect_changed_submodules() runs a revision walk to essentially do file-history on the list of submodules. That revision walk marks commits UNININTERESTING if they are simiplified away by not changing the submodule.

Instead, I finally arrived on the conclusion that I should use a flag that is not used in any other part of the code. In commit-reach.c, a number of flags were defined for commit walk algorithms. The REACHABLE flag seemed like it made the most sense, and it seems it was not actually used in the file. The REACHABLE flag was used in early versions of commit-reach.c, but was removed by 4fbcca4 (commit-reach: make can_all_from_reach... linear, 2018-07-20).

Add the REACHABLE flag to commit-graph.c and use it instead of UNINTERESTING in close_reachable(). This fixes the bug in manual testing.

---
 commit-graph.c   | 11 +++++++----
 commit-reach.c   |  1 -
 object.h         |  3 ++-
 t/t5510-fetch.sh |  2 +-
 4 files changed, 10 insertions(+), 7 deletions(-)

index fc4a43b8d6..0aea7b2ae5 100644
@@ -41,6 +41,9 @@
 #define GRAPH_MIN_SIZE (GRAPH_HEADER_SIZE + 4 * GRAPH_CHUNKLOOKUP_WIDTH \
 			+ GRAPH_FANOUT_SIZE + the_hash_algo->rawsz)
 
+/* Remember to update object flag allocation in object.h */
+#define REACHABLE       (1u<<15)
+
 char *get_commit_graph_filename(const char *obj_dir)
 {
 	char *filename = xstrfmt("%s/info/commit-graph", obj_dir);

@@ -1030,11 +1033,11 @@ static void add_missing_parents(struct write_commit_graph_context *ctx, struct c
 {
 	struct commit_list *parent;
 	for (parent = commit->parents; parent; parent = parent->next) {
-		if (!(parent->item->object.flags & UNINTERESTING)) {
+		if (!(parent->item->object.flags & REACHABLE)) {
 			ALLOC_GROW(ctx->oids.list, ctx->oids.nr + 1, ctx->oids.alloc);
 			oidcpy(&ctx->oids.list[ctx->oids.nr], &(parent->item->object.oid));
 			ctx->oids.nr++;
-			parent->item->object.flags |= UNINTERESTING;
+			parent->item->object.flags |= REACHABLE;
 		}
 	}
 }

@@ -1052,7 +1055,7 @@ static void close_reachable(struct write_commit_graph_context *ctx)
 		display_progress(ctx->progress, i + 1);
 		commit = lookup_commit(ctx->r, &ctx->oids.list[i]);
 		if (commit)
-			commit->object.flags |= UNINTERESTING;
+			commit->object.flags |= REACHABLE;
 	}
 	stop_progress(&ctx->progress);
 

@@ -1089,7 +1092,7 @@ static void close_reachable(struct write_commit_graph_context *ctx)
 		commit = lookup_commit(ctx->r, &ctx->oids.list[i]);
 
 		if (commit)
-			commit->object.flags &= ~UNINTERESTING;
+			commit->object.flags &= ~REACHABLE;
 	}
 	stop_progress(&ctx->progress);
 }

index 3ea174788a..4ca7e706a1 100644
@@ -10,7 +10,6 @@
 #include "commit-reach.h"
 
 /* Remember to update object flag allocation in object.h */
-#define REACHABLE       (1u<<15)
 #define PARENT1		(1u<<16)
 #define PARENT2		(1u<<17)
 #define STALE		(1u<<18)

index 0120892bbd..25f5ab3d54 100644
@@ -68,7 +68,8 @@ struct object_array {
  * bisect.c:                                        16
  * bundle.c:                                        16
  * http-push.c:                                     16-----19
- * commit-reach.c:                                15-------19
+ * commit-graph.c:                                15
+ * commit-reach.c:                                  16-----19
  * sha1-name.c:                                              20
  * list-objects-filter.c:                                      21
  * builtin/fsck.c:           0--3

index e8ae3af0b6..7bfbcc2036 100755
@@ -583,7 +583,7 @@ test_expect_success 'fetch.writeCommitGraph' '
 	)
 '
 
-test_expect_failure 'fetch.writeCommitGraph with submodules' '
+test_expect_success 'fetch.writeCommitGraph with submodules' '
 	pwd="$(pwd)" &&
 	git clone dups super &&
 	(

gitgitgadget

s/prolem/problem/

Does this last sentence still holds, given that a submodule plays a crucial role in triggering this bug? I think it either doesn't, or I still don't completely understand the situation.

s/simiplified/simplified/

I'm inclined to agree that using a flag that is not used anywhere else is the safest thing to do, and at -rcX time safest is good. I'm not sure whether it's the right thing to do in the long term, though.

Furthermore, calling this flag REACHABLE is misleading, because the code actually means SEEN.
Consider the following sequence of commands:

  # Create a pack with two commits
  $ git commit --allow-empty -m one &&
  $ git commit --allow-empty -m two &&
  $ git repack -ad &&
  # Make one of those commits unreachable
  $ git reset --hard HEAD^ &&
  # Not even from reflogs!
  $ git reflog expire --expire-unreachable=now --all
  # Now write a commit-graph from that pack file
  $ git commit-graph write
  Computing commit graph generation numbers: 100% (2/2), done.

It added two commits to the commit-graph, although one of them is clearly not reachable anymore, so marking it as REACHABLE while enumerating all commits feels wrong.

This does not apply anymore. I forgot to delete it.

Since you are using "git commit-graph write", the command is scanning all pack-files for commits to include. Even in this case, the close_reachable() method needs to walk to see if any commits are missing. (It could be that the root commit is loose for some strange reason.)

In this case, we are marking REACHABLE the commits that can be reached from our "starting" commits. In your example we start with every commit.

If you had used git commit-graph write --stdin-packs and provided a small pack name over stdin, the concept would be similar and even more
pronounced: the pack (perhaps downloaded via 'fetch') is not likely to contain every commit, so we need to walk all reachable commits from those included.

I'll have a v3 today with the requested fixes.

Thanks,
-Stolee

That's exactly my point. fsck already uses the REACHABLE flag to mark objects that are reachable not only from all refs (including the HEADs of all worktrees), but their reflogs and the index as well.

However, in close_unreachable() we start with a bunch of commits and then go into a loop adding their parents to an array, while marking each parent to prevent them from being added multiple times in a mergy history. We have a good couple of similar traversals in our code base, and in revision.c, builtin/describe.c, walker.c, negotiator/default.c (and at this point I stopped looking) we mark those parents as SEEN.

On a somewhat related note: 'git commit-graph write --reachable' doesn't include HEAD; should it?

UPDATE for V2: We now know the full repro, and a test is added. Thanks Szeder and Peff for your insights!

UPDATE in V3: Cleaned up the commit messages and some test details.

While dogfooding, Johannes found a bug in the fetch.writeCommitGraph config behavior. While his example initially happened during a clone with --recurse-submodules, (UPDATE) and the submodule is important, but --recurse-submodules is not:

$ git clone <url> test
$ cd test
$ git -c fetch.writeCommitGraph=true fetch origin
Computing commit graph generation numbers: 100% (12/12), done.
BUG: commit-graph.c:886: missing parent <hash1> for commit <hash2>
Aborted (core dumped)

In the repo I had cloned, there were really 60 commits to scan, but only 12 were in the list to write when calling compute_generation_numbers(). A commit in the list expects to see a parent, but that parent is not in the list. The simple example I used for my testing was
https://github.com/derrickstolee/numbers. Thie repo HAS A SUBMODULE, I just forgot. Sorry for derailing the investigation somewhat.

The details above are the start of the commit message for [PATCH 1/2], including a test that fails when fetching after cloning a repo with a submodule.

In [PATCH 2/2], I actually have the fix. I tried to include as much detail as I could for how I investigated the problem and why I think this is the right solution. I added details that have come from the on-list discussion, including what the submodule code is doing and why REACHABLE is no longer used in commit-reach.c.

Thanks, -Stolee

Derrick Stolee (2):
t5510-fetch.sh: demonstrate fetch.writeCommitGraph bug
commit-graph: fix writing first commit-graph during fetch

 commit-graph.c   | 11 +++++++----
 commit-reach.c   |  1 -
 object.h         |  3 ++-
 t/t5510-fetch.sh | 16 ++++++++++++++++
 4 files changed, 25 insertions(+), 6 deletions(-)

base-commit: d966095db01190a2196e31195ea6fa0c722aa732
Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-415%2Fderrickstolee%2Ffetch-first-write-fail-v3
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-415/derrickstolee/fetch-first-write-fail-v3
Pull-Request: https://github.com/gitgitgadget/git/pull/415

Range-diff vs v2:

1: 6ac0a05746 ! 1: ce53b5a7bf t5510-fetch.sh: demonstrate fetch.writeCommitGraph bug

     @@ -22,10 +22,6 @@
          A follow-up will fix the bug, but first we create a test that
          demonstrates the problem.
      
     -    I used "test_expect_failure" for the entire test instead of
     -    "test_must_fail" only on the command that I expect to fail. This is
     -    because the BUG() returns an exit code so test_must_fail complains.
     -
          Helped-by: Jeff King <[email protected]>
          Helped-by: Johannes Schindelin <[email protected]>
          Helped-by: Szeder Gábor <[email protected]>
     @@ -39,16 +35,15 @@
       '
       
      +test_expect_failure 'fetch.writeCommitGraph with submodules' '
     -+	pwd="$(pwd)" &&
      +	git clone dups super &&
      +	(
      +		cd super &&
     -+		git submodule add "file://$pwd/three" &&
     ++		git submodule add "file://$TRASH_DIRECTORY/three" &&
      +		git commit -m "add submodule"
      +	) &&
     -+	git clone "super" writeError &&
     ++	git clone "super" super-clone &&
      +	(
     -+		cd writeError &&
     ++		cd super-clone &&
      +		test_path_is_missing .git/objects/info/commit-graphs/commit-graph-chain &&
      +		git -c fetch.writeCommitGraph=true fetch origin &&
      +		test_path_is_file .git/objects/info/commit-graphs/commit-graph-chain

2: ca59b118f1 ! 2: edacfff490 commit-graph: fix writing first commit-graph during fetch
@@ -6,7 +6,7 @@
fetch.writeCommitGraph and fetching in a repo with a submodule. Here, we fix that bug and set the test to "test_expect_success".

- The prolem arises with this set of commands when the remote repo at
+ The problem arises with this set of commands when the remote repo at
<url> has a submodule. Note that --recurse-submodules is not needed to demonstrate the bug.

@@ -44,16 +44,6 @@
negotiation is comparing the remote refs with the local refs and marking some commits as UNINTERESTING.

- You may ask: did this feature ever work at all? Yes, it did, as long as
- you had a commit-graph covering all of your local refs. My testing was
- unfortunately limited to this scenario. The UNINTERESTING commits are
- always part of the "old" commit-graph, and when we add new commits to a
- top layer of the commit-graph chain those are not needed. If we happen
- to merge layers of the chain, then the commits are added as a list, not
- using a commit walk. Further, the test added for this config option in
- t5510-fetch.sh uses local filesystem clones, which somehow avoids this
- logic.
-
I tested running clear_commit_marks_many() to clear the UNINTERESTING flag inside close_reachable(), but the tips did not have the flag, so that did nothing.
@@ -62,7 +52,7 @@
fault. Thanks, Peff, for pointing out this detail! More specifically, for each submodule, the collect_changed_submodules() runs a revision walk to essentially do file-history on the list of submodules. That
- revision walk marks commits UNININTERESTING if they are simiplified away
+ revision walk marks commits UNININTERESTING if they are simplified away
by not changing the submodule.

Instead, I finally arrived on the conclusion that I should use a flag
@@ -163,6 +153,6 @@

-test_expect_failure 'fetch.writeCommitGraph with submodules' ' +test_expect_success 'fetch.writeCommitGraph with submodules' '
- pwd="$(pwd)" &&
git clone dups super &&
(
+ cd super &&

gitgitgadget

From: Derrick Stolee <[email protected]>

While dogfooding, Johannes found a bug in the fetch.writeCommitGraph config behavior. His example initially happened during a clone with --recurse-submodules, we found that this happens with the first fetch after cloning a repository that contains a submodule:

	$ git clone <url> test
	$ cd test
	$ git -c fetch.writeCommitGraph=true fetch origin
	Computing commit graph generation numbers: 100% (12/12), done.
	BUG: commit-graph.c:886: missing parent <hash1> for commit <hash2>
	Aborted (core dumped)

In the repo I had cloned, there were really 60 commits to scan, but only 12 were in the list to write when calling compute_generation_numbers(). A commit in the list expects to see a parent, but that parent is not in the list.

A follow-up will fix the bug, but first we create a test that demonstrates the problem.

---
 t/t5510-fetch.sh | 16 ++++++++++++++++
 1 file changed, 16 insertions(+)

index ecabbe1616..cc76e1ce50 100755
@@ -583,6 +583,22 @@ test_expect_success 'fetch.writeCommitGraph' '
 	)
 '
 
+test_expect_failure 'fetch.writeCommitGraph with submodules' '
+	git clone dups super &&
+	(
+		cd super &&
+		git submodule add "file://$TRASH_DIRECTORY/three" &&
+		git commit -m "add submodule"
+	) &&
+	git clone "super" super-clone &&
+	(
+		cd super-clone &&
+		test_path_is_missing .git/objects/info/commit-graphs/commit-graph-chain &&
+		git -c fetch.writeCommitGraph=true fetch origin &&
+		test_path_is_file .git/objects/info/commit-graphs/commit-graph-chain
+	)
+'
+
 # configured prune tests
 
 set_config_tristate () {

gitgitgadget

From: Derrick Stolee <[email protected]>

The previous commit includes a failing test for an issue around fetch.writeCommitGraph and fetching in a repo with a submodule. Here, we fix that bug and set the test to "test_expect_success".

The problem arises with this set of commands when the remote repo at <url> has a submodule. Note that --recurse-submodules is not needed to demonstrate the bug.

	$ git clone <url> test
	$ cd test
	$ git -c fetch.writeCommitGraph=true fetch origin
	Computing commit graph generation numbers: 100% (12/12), done.
	BUG: commit-graph.c:886: missing parent <hash1> for commit <hash2>
	Aborted (core dumped)

As an initial fix, I converted the code in builtin/fetch.c that calls write_commit_graph_reachable() to instead launch a "git commit-graph write --reachable --split" process. That code worked, but is not how we want the feature to work long-term.

That test did demonstrate that the issue must be something to do with internal state of the 'git fetch' process.

The write_commit_graph() method in commit-graph.c ensures the commits we plan to write are "closed under reachability" using close_reachable(). This method walks from the input commits, and uses the UNINTERESTING flag to mark which commits have already been visited. This allows the walk to take O(N) time, where N is the number of commits, instead of O(P) time, where P is the number of paths. (The number of paths can be exponential in the number of commits.)

However, the UNINTERESTING flag is used in lots of places in the codebase. This flag usually means some barrier to stop a commit walk, such as in revision-walking to compare histories. It is not often cleared after the walk completes because the starting points of those walks do not have the UNINTERESTING flag, and clear_commit_marks() would stop immediately.

This is happening during a 'git fetch' call with a remote. The fetch negotiation is comparing the remote refs with the local refs and marking some commits as UNINTERESTING.

I tested running clear_commit_marks_many() to clear the UNINTERESTING flag inside close_reachable(), but the tips did not have the flag, so that did nothing.

It turns out that the calculate_changed_submodule_paths() method is at fault. Thanks, Peff, for pointing out this detail! More specifically, for each submodule, the collect_changed_submodules() runs a revision walk to essentially do file-history on the list of submodules. That revision walk marks commits UNININTERESTING if they are simplified away by not changing the submodule.

Instead, I finally arrived on the conclusion that I should use a flag that is not used in any other part of the code. In commit-reach.c, a number of flags were defined for commit walk algorithms. The REACHABLE flag seemed like it made the most sense, and it seems it was not actually used in the file. The REACHABLE flag was used in early versions of commit-reach.c, but was removed by 4fbcca4 (commit-reach: make can_all_from_reach... linear, 2018-07-20).

Add the REACHABLE flag to commit-graph.c and use it instead of UNINTERESTING in close_reachable(). This fixes the bug in manual testing.

---
 commit-graph.c   | 11 +++++++----
 commit-reach.c   |  1 -
 object.h         |  3 ++-
 t/t5510-fetch.sh |  2 +-
 4 files changed, 10 insertions(+), 7 deletions(-)

index fc4a43b8d6..0aea7b2ae5 100644
@@ -41,6 +41,9 @@
 #define GRAPH_MIN_SIZE (GRAPH_HEADER_SIZE + 4 * GRAPH_CHUNKLOOKUP_WIDTH \
 			+ GRAPH_FANOUT_SIZE + the_hash_algo->rawsz)
 
+/* Remember to update object flag allocation in object.h */
+#define REACHABLE       (1u<<15)
+
 char *get_commit_graph_filename(const char *obj_dir)
 {
 	char *filename = xstrfmt("%s/info/commit-graph", obj_dir);

@@ -1030,11 +1033,11 @@ static void add_missing_parents(struct write_commit_graph_context *ctx, struct c
 {
 	struct commit_list *parent;
 	for (parent = commit->parents; parent; parent = parent->next) {
-		if (!(parent->item->object.flags & UNINTERESTING)) {
+		if (!(parent->item->object.flags & REACHABLE)) {
 			ALLOC_GROW(ctx->oids.list, ctx->oids.nr + 1, ctx->oids.alloc);
 			oidcpy(&ctx->oids.list[ctx->oids.nr], &(parent->item->object.oid));
 			ctx->oids.nr++;
-			parent->item->object.flags |= UNINTERESTING;
+			parent->item->object.flags |= REACHABLE;
 		}
 	}
 }

@@ -1052,7 +1055,7 @@ static void close_reachable(struct write_commit_graph_context *ctx)
 		display_progress(ctx->progress, i + 1);
 		commit = lookup_commit(ctx->r, &ctx->oids.list[i]);
 		if (commit)
-			commit->object.flags |= UNINTERESTING;
+			commit->object.flags |= REACHABLE;
 	}
 	stop_progress(&ctx->progress);
 

@@ -1089,7 +1092,7 @@ static void close_reachable(struct write_commit_graph_context *ctx)
 		commit = lookup_commit(ctx->r, &ctx->oids.list[i]);
 
 		if (commit)
-			commit->object.flags &= ~UNINTERESTING;
+			commit->object.flags &= ~REACHABLE;
 	}
 	stop_progress(&ctx->progress);
 }

index 3ea174788a..4ca7e706a1 100644
@@ -10,7 +10,6 @@
 #include "commit-reach.h"
 
 /* Remember to update object flag allocation in object.h */
-#define REACHABLE       (1u<<15)
 #define PARENT1		(1u<<16)
 #define PARENT2		(1u<<17)
 #define STALE		(1u<<18)

index 0120892bbd..25f5ab3d54 100644
@@ -68,7 +68,8 @@ struct object_array {
  * bisect.c:                                        16
  * bundle.c:                                        16
  * http-push.c:                                     16-----19
- * commit-reach.c:                                15-------19
+ * commit-graph.c:                                15
+ * commit-reach.c:                                  16-----19
  * sha1-name.c:                                              20
  * list-objects-filter.c:                                      21
  * builtin/fsck.c:           0--3

index cc76e1ce50..e16b7b8db2 100755
@@ -583,7 +583,7 @@ test_expect_success 'fetch.writeCommitGraph' '
 	)
 '
 
-test_expect_failure 'fetch.writeCommitGraph with submodules' '
+test_expect_success 'fetch.writeCommitGraph with submodules' '
 	git clone dups super &&
 	(
 		cd super &&

gitgitgadget

UPDATE for V2: We now know the full repro, and a test is added. Thanks Szeder and Peff for your insights!

UPDATE in V3: Cleaned up the commit messages and some test details.

UPDATE in V4: There is an unfortunate interaction with GIT_TEST_COMMIT_GRAPH that requires an update to the test. Sorry for not noticing this earlier. Thanks, Johannes for pointing it out!

While dogfooding, Johannes found a bug in the fetch.writeCommitGraph config behavior. While his example initially happened during a clone with --recurse-submodules, (UPDATE) and the submodule is important, but --recurse-submodules is not:

$ git clone <url> test
$ cd test
$ git -c fetch.writeCommitGraph=true fetch origin
Computing commit graph generation numbers: 100% (12/12), done.
BUG: commit-graph.c:886: missing parent <hash1> for commit <hash2>
Aborted (core dumped)

In the repo I had cloned, there were really 60 commits to scan, but only 12 were in the list to write when calling compute_generation_numbers(). A commit in the list expects to see a parent, but that parent is not in the list. The simple example I used for my testing was
https://github.com/derrickstolee/numbers. Thie repo HAS A SUBMODULE, I just forgot. Sorry for derailing the investigation somewhat.

The details above are the start of the commit message for [PATCH 1/2], including a test that fails when fetching after cloning a repo with a submodule.

In [PATCH 2/2], I actually have the fix. I tried to include as much detail as I could for how I investigated the problem and why I think this is the right solution. I added details that have come from the on-list discussion, including what the submodule code is doing and why REACHABLE is no longer used in commit-reach.c.

Thanks, -Stolee

Derrick Stolee (2):
t5510-fetch.sh: demonstrate fetch.writeCommitGraph bug
commit-graph: fix writing first commit-graph during fetch

 commit-graph.c   | 11 +++++++----
 commit-reach.c   |  1 -
 object.h         |  3 ++-
 t/t5510-fetch.sh | 16 ++++++++++++++++
 4 files changed, 25 insertions(+), 6 deletions(-)

base-commit: d966095db01190a2196e31195ea6fa0c722aa732
Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-415%2Fderrickstolee%2Ffetch-first-write-fail-v4
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-415/derrickstolee/fetch-first-write-fail-v4
Pull-Request: https://github.com/gitgitgadget/git/pull/415

Range-diff vs v3:

1: ce53b5a7bf ! 1: 8150108471 t5510-fetch.sh: demonstrate fetch.writeCommitGraph bug
@@ -20,7 +20,10 @@
parent, but that parent is not in the list.

A follow-up will fix the bug, but first we create a test that
- demonstrates the problem.
+ demonstrates the problem. This test must be careful about an existing
+ commit-graph file, since GIT_TEST_COMMIT_GRAPH=1 will cause the repo we
+ are cloning to already have one. This then prevents the incremtnal
+ commit-graph write during the first 'git fetch'.

@@ -44,7 +47,7 @@
+ git clone "super" super-clone &&
+ (
+ cd super-clone &&
-+ test_path_is_missing .git/objects/info/commit-graphs/commit-graph-chain &&
++ rm -rf .git/objects/info &&
+ git -c fetch.writeCommitGraph=true fetch origin &&
+ test_path_is_file .git/objects/info/commit-graphs/commit-graph-chain
+ )
2: edacfff490 = 2: 6d01e90591 commit-graph: fix writing first commit-graph during fetch

gitgitgadget

From: Derrick Stolee <[email protected]>

While dogfooding, Johannes found a bug in the fetch.writeCommitGraph config behavior. His example initially happened during a clone with --recurse-submodules, we found that this happens with the first fetch after cloning a repository that contains a submodule:

	$ git clone <url> test
	$ cd test
	$ git -c fetch.writeCommitGraph=true fetch origin
	Computing commit graph generation numbers: 100% (12/12), done.
	BUG: commit-graph.c:886: missing parent <hash1> for commit <hash2>
	Aborted (core dumped)

In the repo I had cloned, there were really 60 commits to scan, but only 12 were in the list to write when calling compute_generation_numbers(). A commit in the list expects to see a parent, but that parent is not in the list.

A follow-up will fix the bug, but first we create a test that demonstrates the problem. This test must be careful about an existing commit-graph file, since GIT_TEST_COMMIT_GRAPH=1 will cause the repo we are cloning to already have one. This then prevents the incremtnal commit-graph write during the first 'git fetch'.

---
 t/t5510-fetch.sh | 16 ++++++++++++++++
 1 file changed, 16 insertions(+)

index ecabbe1616..195781ce8d 100755
@@ -583,6 +583,22 @@ test_expect_success 'fetch.writeCommitGraph' '
 	)
 '
 
+test_expect_failure 'fetch.writeCommitGraph with submodules' '
+	git clone dups super &&
+	(
+		cd super &&
+		git submodule add "file://$TRASH_DIRECTORY/three" &&
+		git commit -m "add submodule"
+	) &&
+	git clone "super" super-clone &&
+	(
+		cd super-clone &&
+		rm -rf .git/objects/info &&
+		git -c fetch.writeCommitGraph=true fetch origin &&
+		test_path_is_file .git/objects/info/commit-graphs/commit-graph-chain
+	)
+'
+
 # configured prune tests
 
 set_config_tristate () {

gitgitgadget

From: Derrick Stolee <[email protected]>

The previous commit includes a failing test for an issue around fetch.writeCommitGraph and fetching in a repo with a submodule. Here, we fix that bug and set the test to "test_expect_success".

The problem arises with this set of commands when the remote repo at <url> has a submodule. Note that --recurse-submodules is not needed to demonstrate the bug.

	$ git clone <url> test
	$ cd test
	$ git -c fetch.writeCommitGraph=true fetch origin
	Computing commit graph generation numbers: 100% (12/12), done.
	BUG: commit-graph.c:886: missing parent <hash1> for commit <hash2>
	Aborted (core dumped)

As an initial fix, I converted the code in builtin/fetch.c that calls write_commit_graph_reachable() to instead launch a "git commit-graph write --reachable --split" process. That code worked, but is not how we want the feature to work long-term.

That test did demonstrate that the issue must be something to do with internal state of the 'git fetch' process.

The write_commit_graph() method in commit-graph.c ensures the commits we plan to write are "closed under reachability" using close_reachable(). This method walks from the input commits, and uses the UNINTERESTING flag to mark which commits have already been visited. This allows the walk to take O(N) time, where N is the number of commits, instead of O(P) time, where P is the number of paths. (The number of paths can be exponential in the number of commits.)

However, the UNINTERESTING flag is used in lots of places in the codebase. This flag usually means some barrier to stop a commit walk, such as in revision-walking to compare histories. It is not often cleared after the walk completes because the starting points of those walks do not have the UNINTERESTING flag, and clear_commit_marks() would stop immediately.

This is happening during a 'git fetch' call with a remote. The fetch negotiation is comparing the remote refs with the local refs and marking some commits as UNINTERESTING.

I tested running clear_commit_marks_many() to clear the UNINTERESTING flag inside close_reachable(), but the tips did not have the flag, so that did nothing.

It turns out that the calculate_changed_submodule_paths() method is at fault. Thanks, Peff, for pointing out this detail! More specifically, for each submodule, the collect_changed_submodules() runs a revision walk to essentially do file-history on the list of submodules. That revision walk marks commits UNININTERESTING if they are simplified away by not changing the submodule.

Instead, I finally arrived on the conclusion that I should use a flag that is not used in any other part of the code. In commit-reach.c, a number of flags were defined for commit walk algorithms. The REACHABLE flag seemed like it made the most sense, and it seems it was not actually used in the file. The REACHABLE flag was used in early versions of commit-reach.c, but was removed by 4fbcca4 (commit-reach: make can_all_from_reach... linear, 2018-07-20).

Add the REACHABLE flag to commit-graph.c and use it instead of UNINTERESTING in close_reachable(). This fixes the bug in manual testing.

---
 commit-graph.c   | 11 +++++++----
 commit-reach.c   |  1 -
 object.h         |  3 ++-
 t/t5510-fetch.sh |  2 +-
 4 files changed, 10 insertions(+), 7 deletions(-)

index fc4a43b8d6..0aea7b2ae5 100644
@@ -41,6 +41,9 @@
 #define GRAPH_MIN_SIZE (GRAPH_HEADER_SIZE + 4 * GRAPH_CHUNKLOOKUP_WIDTH \
 			+ GRAPH_FANOUT_SIZE + the_hash_algo->rawsz)
 
+/* Remember to update object flag allocation in object.h */
+#define REACHABLE       (1u<<15)
+
 char *get_commit_graph_filename(const char *obj_dir)
 {
 	char *filename = xstrfmt("%s/info/commit-graph", obj_dir);

@@ -1030,11 +1033,11 @@ static void add_missing_parents(struct write_commit_graph_context *ctx, struct c
 {
 	struct commit_list *parent;
 	for (parent = commit->parents; parent; parent = parent->next) {
-		if (!(parent->item->object.flags & UNINTERESTING)) {
+		if (!(parent->item->object.flags & REACHABLE)) {
 			ALLOC_GROW(ctx->oids.list, ctx->oids.nr + 1, ctx->oids.alloc);
 			oidcpy(&ctx->oids.list[ctx->oids.nr], &(parent->item->object.oid));
 			ctx->oids.nr++;
-			parent->item->object.flags |= UNINTERESTING;
+			parent->item->object.flags |= REACHABLE;
 		}
 	}
 }

@@ -1052,7 +1055,7 @@ static void close_reachable(struct write_commit_graph_context *ctx)
 		display_progress(ctx->progress, i + 1);
 		commit = lookup_commit(ctx->r, &ctx->oids.list[i]);
 		if (commit)
-			commit->object.flags |= UNINTERESTING;
+			commit->object.flags |= REACHABLE;
 	}
 	stop_progress(&ctx->progress);
 

@@ -1089,7 +1092,7 @@ static void close_reachable(struct write_commit_graph_context *ctx)
 		commit = lookup_commit(ctx->r, &ctx->oids.list[i]);
 
 		if (commit)
-			commit->object.flags &= ~UNINTERESTING;
+			commit->object.flags &= ~REACHABLE;
 	}
 	stop_progress(&ctx->progress);
 }

index 3ea174788a..4ca7e706a1 100644
@@ -10,7 +10,6 @@
 #include "commit-reach.h"
 
 /* Remember to update object flag allocation in object.h */
-#define REACHABLE       (1u<<15)
 #define PARENT1		(1u<<16)
 #define PARENT2		(1u<<17)
 #define STALE		(1u<<18)

index 0120892bbd..25f5ab3d54 100644
@@ -68,7 +68,8 @@ struct object_array {
  * bisect.c:                                        16
  * bundle.c:                                        16
  * http-push.c:                                     16-----19
- * commit-reach.c:                                15-------19
+ * commit-graph.c:                                15
+ * commit-reach.c:                                  16-----19
  * sha1-name.c:                                              20
  * list-objects-filter.c:                                      21
  * builtin/fsck.c:           0--3

index 195781ce8d..4b60282689 100755
@@ -583,7 +583,7 @@ test_expect_success 'fetch.writeCommitGraph' '
 	)
 '
 
-test_expect_failure 'fetch.writeCommitGraph with submodules' '
+test_expect_success 'fetch.writeCommitGraph with submodules' '
 	git clone dups super &&
 	(
 		cd super &&

gitgitgadget