Skip to contentSkip to conversation

upload-pack: disable commit graph more gently for shallow traversal

PATCH In review Jeff King 5 participants 25 messages Sep 12, 2019 – Sep 13, 2019

lore
Share & tools
lore

Jeff King

author
#0 lore raw

PATCH 1/1 upload-pack: disable commit graph more gently for shallow traversal

When the client has asked for certain shallow options like "deepen-since", we do a custom rev-list walk that pretends to be shallow. Before doing so, we have to disable the commit-graph, since it is not compatible with the shallow view of the repository. That's handled by 829a321569 (commit-graph: close_commit_graph before shallow walk, 2018-08-20). That commit literally closes and frees our repo->objects->commit_graph struct.

That creates an interesting problem for commits that have _already_ been parsed using the commit graph. Their commit->object.parsed flag is set, their commit->graph_pos is set, but their commit->maybe_tree may still be NULL. When somebody later calls repo_get_commit_tree(), we see that we haven't loaded the tree oid yet and try to get it from the commit graph. But since it has been freed, we segfault!

So the root of the issue is a data dependency between the commit's lazy-load of the tree oid and the fact that the commit graph can go away mid-process. How can we resolve it?

There are a couple of general approaches:

1. The obvious answer is to avoid loading the tree from the graph when we see that it's NULL. But then what do we return for the tree oid? If we return NULL, our caller in do_traverse() will rightly complain that we have no tree. We'd have to fallback to loading the actual commit object and re-parsing it. That requires teaching parse_commit_buffer() to understand re-parsing (i.e., not starting from a clean slate and not leaking any allocated bits like parent list pointers).

2. When we close the commit graph, walk through the set of in-memory objects and clear any graph_pos pointers. But this means we also have to "unparse" any such commits so that we know they still need to open the commit object to fill in their trees. So it's no less complicated than (1), and is more expensive (since we clear objects we might not later need).

3. Stop freeing the commit-graph struct. Continue to let it be used for lazy-loads of tree oids, but let upload-pack specify that it shouldn't be used for further commit parsing.

4. Push the whole shallow rev-list out to its own sub-process, with the commit-graph disabled from the start, giving it a clean memory space to work from.

I've chosen (3) here. Options (1) and (2) would work, but are non-trivial to implement. Option (4) is more expensive, and I'm not sure how complicated it is (shelling out for the actual rev-list part is easy, but we do then parse the resulting commits internally, and I'm not clear which parts need to be handling shallow-ness).

The new test in t5500 triggers this segfault, but see the comments there for how horribly intimate it has to be with how both upload-pack and commit graphs work.

---
 commit-graph.c        | 10 ++++++++++
 commit-graph.h        |  6 ++++++
 t/t5500-fetch-pack.sh | 38 ++++++++++++++++++++++++++++++++++++++
 upload-pack.c         |  2 +-
 4 files changed, 55 insertions(+), 1 deletion(-)
commit-graph.c +10 -0
index 9b02d2c426..bc5dd5913f 100644
@@ -41,6 +41,8 @@
 #define GRAPH_MIN_SIZE (GRAPH_HEADER_SIZE + 4 * GRAPH_CHUNKLOOKUP_WIDTH \
 			+ GRAPH_FANOUT_SIZE + the_hash_algo->rawsz)
 
+static int commit_graph_disabled;
+
 char *get_commit_graph_filename(const char *obj_dir)
 {
 	char *filename = xstrfmt("%s/info/commit-graph", obj_dir);
@@ -472,6 +474,9 @@ static int prepare_commit_graph(struct repository *r)
 		die("dying as requested by the '%s' variable on commit-graph load!",
 		    GIT_TEST_COMMIT_GRAPH_DIE_ON_LOAD);
 
+	if (commit_graph_disabled)
+		return 0;
+
 	if (r->objects->commit_graph_attempted)
 		return !!r->objects->commit_graph;
 	r->objects->commit_graph_attempted = 1;
@@ -2101,3 +2106,8 @@ void free_commit_graph(struct commit_graph *g)
 	free(g->filename);
 	free(g);
 }
+
+void disable_commit_graph(void)
+{
+	commit_graph_disabled = 1;
+}
commit-graph.h +6 -0
index 486e64e591..42d6e7c289 100644
@@ -107,4 +107,10 @@ int verify_commit_graph(struct repository *r, struct commit_graph *g, int flags)
 void close_commit_graph(struct raw_object_store *);
 void free_commit_graph(struct commit_graph *);
 
+/*
+ * Disable further use of the commit graph in this process when parsing a
+ * "struct commit".
+ */
+void disable_commit_graph(void);
+
 #endif
t/t5500-fetch-pack.sh +38 -0
index 8210f63d41..7601664b74 100755
@@ -783,6 +783,44 @@ test_expect_success 'clone shallow since selects no commits' '
 	)
 '
 
+# A few subtle things about the request in this test:
+#
+#  - the server must have commit-graphs present and enabled
+#
+#  - the history is such that our want/have share a common ancestor ("base"
+#    here)
+#
+#  - we send only a single have, which is fewer than a normal client would
+#    send. This ensures that we don't parse "base" up front with
+#    parse_object(), but rather traverse to it as a parent while deciding if we
+#    can stop the "have" negotiation, and call parse_commit(). The former
+#    sees the actual object data and so always loads the three oid, whereas the
+#    latter will try to lod it lazily.
+#
+#  - we must use protocol v2, because it handles the "have" negotiation before
+#    processing the shallow direectives
+#
+test_expect_success 'shallow since with commit graph and already-seen commit' '
+	test_create_repo shallow-since-graph &&
+	(
+	cd shallow-since-graph &&
+	test_commit base &&
+	test_commit master &&
+	git checkout -b other HEAD^ &&
+	test_commit other &&
+	git commit-graph write --reachable &&
+	git config core.commitGraph true &&
+
+	GIT_PROTOCOL=version=2 git upload-pack . <<-EOF >/dev/null
+	0012command=fetch
+	00010013deepen-since 1
+	0032want $(git rev-parse other)
+	0032have $(git rev-parse master)
+	0000
+	EOF
+	)
+'
+
 test_expect_success 'shallow clone exclude tag two' '
 	test_create_repo shallow-exclude &&
 	(
upload-pack.c +1 -1
index 222cd3ad89..a260b6b50d 100644
@@ -722,7 +722,7 @@ static void deepen_by_rev_list(struct packet_writer *writer, int ac,
 {
 	struct commit_list *result;
 
-	close_commit_graph(the_repository->objects);
+	disable_commit_graph();
 	result = get_shallow_commits_by_rev_list(ac, av, SHALLOW, NOT_SHALLOW);
 	send_shallow(writer, result);
 	free_commit_list(result);
Signature · 1 line
2.23.0.663.gbe3d117559
author
#1 lore raw
Quoted 7 lines · Jeff King

When the client has asked for certain shallow options like "deepen-since", we do a custom rev-list walk that pretends to be shallow. Before doing so, we have to disable the commit-graph, since it is not compatible with the shallow view of the repository. That's handled by 829a321569 (commit-graph: close_commit_graph before shallow walk, 2018-08-20). That commit literally closes and frees our repo->objects->commit_graph struct.

A few notes and curiosities on my patch and this general area.

The test suite passes with my patch both with and without GIT_TEST_COMMIT_GRAPH=1. But to my surprise, it also passes if I delete the close_commit_graph() line added by 829a321569!

So it's not clear to me whether this whole thing is truly unnecessary (and Stolee was just being overly cautious because the code is related to shallow-ness, even though it is OK doing a true-parent traversal itself), or if we just don't have good test coverage for the case that requires it.

If it _is_ necessary, I'm a little worried there are other problems lurking. The whole issue is that we've seen and parsed some commits before we get to this shallow deepen-since code-path. So just disabling commit-graphs isn't enough. Even without them, we might have parsed some commits the old-fashioned way and filled in their parent pointers. Is that a problem?

If so, then I think we either have to "unparse" all of the existing in-memory commits, or call out to a rev-list process with a fresh memory space.

One especially interesting bit is this:

t/t5500-fetch-pack.sh test_expect_success 'clone shallow since selects no commits' ' · line 801

Quoted 2 lines
t/t5500-fetch-pack.sh +2 -0
+#  - we must use protocol v2, because it handles the "have" negotiation before
+#    processing the shallow direectives

In the v0 protocol, we handle shallows at the end of receive_needs(). So it happens before we call into get_common_commits(), which may do further traversal. That makes it much more likely for the shallow code to see a "clean" slate. Should v2 be ordering things in the same way? But then would the have negotiation see the shallow parents? If so, is that good or bad?

I'm pretty ignorant of how the shallow bits of upload-pack are supposed to work.

Quoted 6 lines

That creates an interesting problem for commits that have _already_ been parsed using the commit graph. Their commit->object.parsed flag is set, their commit->graph_pos is set, but their commit->maybe_tree may still be NULL. When somebody later calls repo_get_commit_tree(), we see that we haven't loaded the tree oid yet and try to get it from the commit graph. But since it has been freed, we segfault!

I was surprised we ever called repo_get_commit_tree() at all, since we're literally just traversing commits here. It looks like list-objects.c is very happy to queue pending trees for each commit, even if we're just going to throw them away when we get to process_tree()! I wonder if could be checking revs->tree_objects here and saving ourselves some work.

Quoted 3 lines

The new test in t5500 triggers this segfault, but see the comments there for how horribly intimate it has to be with how both upload-pack and commit graphs work.

This bug was triggered by a real world case. I'm not entirely clear what the client-side state was; I had to reverse engineer the whole thing out of the server-side coredump. The test here is my attempt to cut it down to a minimum. I don't like having to manually generate the upload-pack input, but I had trouble finding a fetch command that would trigger it. And anyway, given how weirdly specific the requirements are for generating this case, it seemed sensible to me to keep the input close to the source of the problem.

-Peff

author
#2 lore raw

PATCH 1/1 list-objects: don't queue root trees unless revs->tree_objects is set

Quoted 7 lines · Jeff King
Quoted 6 lines

That creates an interesting problem for commits that have _already_ been parsed using the commit graph. Their commit->object.parsed flag is set, their commit->graph_pos is set, but their commit->maybe_tree may still be NULL. When somebody later calls repo_get_commit_tree(), we see that we haven't loaded the tree oid yet and try to get it from the commit graph. But since it has been freed, we segfault!

Quoted 6 lines

I was surprised we ever called repo_get_commit_tree() at all, since we're literally just traversing commits here. It looks like list-objects.c is very happy to queue pending trees for each commit, even if we're just going to throw them away when we get to process_tree()! I wonder if could be checking revs->tree_objects here and saving ourselves some work.

Indeed, this seems to help quite a bit in the commit-graph case. I think it's worth doing (and is independent of the other patch).

-- >8 --
Subject: list-objects: don't queue root trees unless revs->tree_objects is set

When traverse_commit_list() processes each commit, it queues the commit's root tree in the pending array. Then, after all commits are processed, it calls traverse_trees_and_blobs() to walk over the pending list, calling process_tree() on each. But if revs->tree_objects is not set, process_tree() just exists immediately!

We can save ourselves some work by not even bothering to queue these trees in the first place. There are a few subtle points to make:

- we also detect commits with a NULL tree pointer here. But this isn't an interesting check for broken commits, since the lookup_tree() we'd have done during commit parsing doesn't actually check that we have the tree on disk. So we're not losing any robustness.

- besides queueing, we also set the NOT_USER_GIVEN flag on the tree object. This is used by the traverse_commit_list_filtered() variant. But if we're not exploring trees, then we won't actually care about this flag, which is used only inside process_tree() code-paths.

- queueing trees eventually leads to us queueing blobs, too. But we don't need to check revs->blob_objects here. Even in the current code, we still wouldn't find those blobs, because we'd never open up the tree objects to list their contents.

- the user-visible impact to the caller is minimal. The pending trees are all cleared by the time the function returns anyway, by traverse_trees_and_blobs(). We do call a show_commit() callback, which technically could be looking at revs->pending during the callback. But it seems like a rather unlikely thing to do (if you want the tree of the current commit, then accessing the tree struct member is a lot simpler).

So this should be safe to do. Let's look at the benefits:

[before]
Benchmark #1: git -C linux rev-list HEAD >/dev/null

    Time (mean ± σ):      7.651 s ±  0.021 s    [User: 7.399 s, System: 0.252 s]
    Range (min … max):    7.607 s …  7.683 s    10 runs

[after]
Benchmark #1: git -C linux rev-list HEAD >/dev/null

    Time (mean ± σ):      7.593 s ±  0.023 s    [User: 7.329 s, System: 0.264 s]
    Range (min … max):    7.565 s …  7.634 s    10 runs

Not too impressive, but then we're really just avoiding sticking a pointer into a growable array. But still, I'll take a free 0.75% speedup.

Let's try it after running "git commit-graph write":

[before]
Benchmark #1: git -C linux rev-list HEAD >/dev/null

    Time (mean ± σ):      1.458 s ±  0.011 s    [User: 1.199 s, System: 0.259 s]
    Range (min … max):    1.447 s …  1.481 s    10 runs

[after]
Benchmark #1: git -C linux rev-list HEAD >/dev/null

    Time (mean ± σ):      1.126 s ±  0.023 s    [User: 896.5 ms, System: 229.0 ms]
    Range (min … max):    1.106 s …  1.181 s    10 runs

Now that's more like it. We saved over 22% of the total time. Part of that is because the runtime is shorter overall, but the absolute improvement is also much larger. What's going on?

When we fill in a commit struct using the commit graph, we don't bother to set the tree pointer, and instead lazy-load it when somebody calls get_commit_tree(). So we're not only skipping the pointer write to the pending queue, but we're skipping the lazy-load of the tree entirely.

---
 list-objects.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)
list-objects.c +3 -1
index b5651ddd5b..c837bcaca8 100644
@@ -370,7 +370,9 @@ static void do_traverse(struct traversal_context *ctx)
 		 * an uninteresting boundary commit may not have its tree
 		 * parsed yet, but we are not going to show them anyway
 		 */
-		if (get_commit_tree(commit)) {
+		if (!ctx->revs->tree_objects)
+			; /* do not bother loading tree */
+		else if (get_commit_tree(commit)) {
 			struct tree *tree = get_commit_tree(commit);
 			tree->object.flags |= NOT_USER_GIVEN;
 			add_pending_tree(ctx->revs, tree);
Signature · 1 line
2.23.0.663.gbe3d117559
author
#3 lore raw

Re: [PATCH] list-objects: don't queue root trees unless revs->tree_objects is set

Quoted 7 lines · Jeff King

Let's try it after running "git commit-graph write":

[before]
Benchmark #1: git -C linux rev-list HEAD >/dev/null

    Time (mean ± σ):      1.458 s ±  0.011 s    [User: 1.199 s, System: 0.259 s]
    Range (min … max):    1.447 s …  1.481 s    10 runs
Quoted 8 lines

[after]
Benchmark #1: git -C linux rev-list HEAD >/dev/null

    Time (mean ± σ):      1.126 s ±  0.023 s    [User: 896.5 ms, System: 229.0 ms]
    Range (min … max):    1.106 s …  1.181 s    10 runs

Now that's more like it. We saved over 22% of the total time. Part of that is because the runtime is shorter overall, but the absolute improvement is also much larger. What's going on?

Another thing I noticed is that rev-list line-buffers when we're writing to /dev/null. This is actually the doing of glibc's stdio, as it consider the character device special enough to turn off full buffering (we also do our own manual flush after each commit).

I think it's probably a fairer test to time it that way (quite often you'd be writing to a pipe, which would have the same behavior). But our improvement is even better as a percentage when writing to a file:

[before]
Benchmark #1: git -C linux rev-list HEAD >file
Time (mean ± σ): 1.046 s ± 0.017 s [User: 922.7 ms, System: 104.3 ms]
Range (min … max): 1.031 s … 1.087 s 10 runs

[after]
Benchmark #1: git -C linux rev-list HEAD >file
Time (mean ± σ): 741.4 ms ± 14.1 ms [User: 644.8 ms, System: 75.9 ms]
Range (min … max): 721.2 ms … 766.8 ms 10 runs

That's a 29% improvement instead of 22% (and shows that write() syscalls are wasting close to 30% of our runtime, a well).

I wonder if it would be worth teaching rev-list a --buffer option. Or just kicking it in automatically when we're just printing single oids. Once upon a time the single-record flushing was useful for:

git rev-list HEAD -- <pathspec> | git diff-tree ...

to feed incremental results as soon as we have them (imagine we see one commit which touches the pathspec, then go through 100,000 that don't). But these days "git log" does that at all internally (and typically outputs quite a bit more between each flush, though one could argue that "log --oneline" might want the same behavior).

I dunno. Maybe it's not worth micro-optimizing too hard, but I was surprised how big a difference it made.

-Peff

#4 lore raw

Re: [PATCH] list-objects: don't queue root trees unless revs->tree_objects is set

In reply to Jeff King
Quoted 16 lines · Jeff King
Quoted 7 lines · Jeff King
Quoted 6 lines

That creates an interesting problem for commits that have _already_ been parsed using the commit graph. Their commit->object.parsed flag is set, their commit->graph_pos is set, but their commit->maybe_tree may still be NULL. When somebody later calls repo_get_commit_tree(), we see that we haven't loaded the tree oid yet and try to get it from the commit graph. But since it has been freed, we segfault!

Quoted 6 lines

I was surprised we ever called repo_get_commit_tree() at all, since we're literally just traversing commits here. It looks like list-objects.c is very happy to queue pending trees for each commit, even if we're just going to throw them away when we get to process_tree()! I wonder if could be checking revs->tree_objects here and saving ourselves some work.

Quoted 2 lines

Indeed, this seems to help quite a bit in the commit-graph case. I think it's worth doing (and is independent of the other patch).

Good find!

Quoted 27 lines

-- >8 --
Subject: list-objects: don't queue root trees unless revs->tree_objects is set

When traverse_commit_list() processes each commit, it queues the commit's root tree in the pending array. Then, after all commits are processed, it calls traverse_trees_and_blobs() to walk over the pending list, calling process_tree() on each. But if revs->tree_objects is not set, process_tree() just exists immediately!

We can save ourselves some work by not even bothering to queue these trees in the first place. There are a few subtle points to make:

- we also detect commits with a NULL tree pointer here. But this isn't an interesting check for broken commits, since the lookup_tree() we'd have done during commit parsing doesn't actually check that we have the tree on disk. So we're not losing any robustness.

- besides queueing, we also set the NOT_USER_GIVEN flag on the tree object. This is used by the traverse_commit_list_filtered() variant. But if we're not exploring trees, then we won't actually care about this flag, which is used only inside process_tree() code-paths.

- queueing trees eventually leads to us queueing blobs, too. But we don't need to check revs->blob_objects here. Even in the current code, we still wouldn't find those blobs, because we'd never open up the tree objects to list their contents.

Quoted 7 lines

- the user-visible impact to the caller is minimal. The pending trees are all cleared by the time the function returns anyway, by traverse_trees_and_blobs(). We do call a show_commit() callback, which technically could be looking at revs->pending during the callback. But it seems like a rather unlikely thing to do (if you want the tree of the current commit, then accessing the tree struct member is a lot simpler).

These all look reasonable. We shouldn't need to do any of that any more.

Quoted 23 lines

So this should be safe to do. Let's look at the benefits:

[before]
Benchmark #1: git -C linux rev-list HEAD >/dev/null

    Time (mean ± σ):      7.651 s ±  0.021 s    [User: 7.399 s, System: 0.252 s]
    Range (min … max):    7.607 s …  7.683 s    10 runs

[after]
Benchmark #1: git -C linux rev-list HEAD >/dev/null

    Time (mean ± σ):      7.593 s ±  0.023 s    [User: 7.329 s, System: 0.264 s]
    Range (min … max):    7.565 s …  7.634 s    10 runs

Not too impressive, but then we're really just avoiding sticking a pointer into a growable array. But still, I'll take a free 0.75% speedup.

Let's try it after running "git commit-graph write":

[before]
Benchmark #1: git -C linux rev-list HEAD >/dev/null

    Time (mean ± σ):      1.458 s ±  0.011 s    [User: 1.199 s, System: 0.259 s]
    Range (min … max):    1.447 s …  1.481 s    10 runs
Quoted 8 lines

[after]
Benchmark #1: git -C linux rev-list HEAD >/dev/null

    Time (mean ± σ):      1.126 s ±  0.023 s    [User: 896.5 ms, System: 229.0 ms]
    Range (min … max):    1.106 s …  1.181 s    10 runs

Now that's more like it. We saved over 22% of the total time. Part of that is because the runtime is shorter overall, but the absolute improvement is also much larger. What's going on?

Very cool.

Quoted 15 lines

When we fill in a commit struct using the commit graph, we don't bother to set the tree pointer, and instead lazy-load it when somebody calls get_commit_tree(). So we're not only skipping the pointer write to the pending queue, but we're skipping the lazy-load of the tree entirely.

---
 list-objects.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)
list-objects.c +0 -0
index b5651ddd5b..c837bcaca8 100644

list-objects.c static void do_traverse(struct traversal_context *ctx) · line 378

Quoted 10 lines
list-objects.c +3 -1
@@ -370,7 +370,9 @@ static void do_traverse(struct traversal_context *ctx)
 		 * an uninteresting boundary commit may not have its tree
 		 * parsed yet, but we are not going to show them anyway
 		 */
-		if (get_commit_tree(commit)) {
+		if (!ctx->revs->tree_objects)
+			; /* do not bother loading tree */
+		else if (get_commit_tree(commit)) {
 			struct tree *tree = get_commit_tree(commit);
 			tree->object.flags |= NOT_USER_GIVEN;
 			add_pending_tree(ctx->revs, tree);

And a simple code fix. LGTM.

Thanks!
-Stolee

#5 lore raw

Re: [PATCH] list-objects: don't queue root trees unless revs->tree_objects is set

In reply to Jeff King
Quoted 9 lines · Jeff King
Quoted 6 lines

I was surprised we ever called repo_get_commit_tree() at all, since we're literally just traversing commits here. It looks like list-objects.c is very happy to queue pending trees for each commit, even if we're just going to throw them away when we get to process_tree()! I wonder if could be checking revs->tree_objects here and saving ourselves some work.

Indeed, this seems to help quite a bit in the commit-graph case. I think it's worth doing (and is independent of the other patch).

Yeah, I agree this is very much worth doing and is orthogonal to the other one.

Thanks for spotting it. I wonder if it was broken like this forever since the beginning X-<.

Quoted 93 lines

-- >8 --
Subject: list-objects: don't queue root trees unless revs->tree_objects is set

When traverse_commit_list() processes each commit, it queues the commit's root tree in the pending array. Then, after all commits are processed, it calls traverse_trees_and_blobs() to walk over the pending list, calling process_tree() on each. But if revs->tree_objects is not set, process_tree() just exists immediately!

We can save ourselves some work by not even bothering to queue these trees in the first place. There are a few subtle points to make:

- we also detect commits with a NULL tree pointer here. But this isn't an interesting check for broken commits, since the lookup_tree() we'd have done during commit parsing doesn't actually check that we have the tree on disk. So we're not losing any robustness.

- besides queueing, we also set the NOT_USER_GIVEN flag on the tree object. This is used by the traverse_commit_list_filtered() variant. But if we're not exploring trees, then we won't actually care about this flag, which is used only inside process_tree() code-paths.

- queueing trees eventually leads to us queueing blobs, too. But we don't need to check revs->blob_objects here. Even in the current code, we still wouldn't find those blobs, because we'd never open up the tree objects to list their contents.

- the user-visible impact to the caller is minimal. The pending trees are all cleared by the time the function returns anyway, by traverse_trees_and_blobs(). We do call a show_commit() callback, which technically could be looking at revs->pending during the callback. But it seems like a rather unlikely thing to do (if you want the tree of the current commit, then accessing the tree struct member is a lot simpler).

So this should be safe to do. Let's look at the benefits:

[before]
Benchmark #1: git -C linux rev-list HEAD >/dev/null

    Time (mean ± σ):      7.651 s ±  0.021 s    [User: 7.399 s, System: 0.252 s]
    Range (min … max):    7.607 s …  7.683 s    10 runs

[after]
Benchmark #1: git -C linux rev-list HEAD >/dev/null

    Time (mean ± σ):      7.593 s ±  0.023 s    [User: 7.329 s, System: 0.264 s]
    Range (min … max):    7.565 s …  7.634 s    10 runs

Not too impressive, but then we're really just avoiding sticking a pointer into a growable array. But still, I'll take a free 0.75% speedup.

Let's try it after running "git commit-graph write":

[before]
Benchmark #1: git -C linux rev-list HEAD >/dev/null

    Time (mean ± σ):      1.458 s ±  0.011 s    [User: 1.199 s, System: 0.259 s]
    Range (min … max):    1.447 s …  1.481 s    10 runs

[after]
Benchmark #1: git -C linux rev-list HEAD >/dev/null

    Time (mean ± σ):      1.126 s ±  0.023 s    [User: 896.5 ms, System: 229.0 ms]
    Range (min … max):    1.106 s …  1.181 s    10 runs

Now that's more like it. We saved over 22% of the total time. Part of that is because the runtime is shorter overall, but the absolute improvement is also much larger. What's going on?

When we fill in a commit struct using the commit graph, we don't bother to set the tree pointer, and instead lazy-load it when somebody calls get_commit_tree(). So we're not only skipping the pointer write to the pending queue, but we're skipping the lazy-load of the tree entirely.

---
 list-objects.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)
list-objects.c +3 -1
index b5651ddd5b..c837bcaca8 100644
@@ -370,7 +370,9 @@ static void do_traverse(struct traversal_context *ctx)
 		 * an uninteresting boundary commit may not have its tree
 		 * parsed yet, but we are not going to show them anyway
 		 */
-		if (get_commit_tree(commit)) {
+		if (!ctx->revs->tree_objects)
+			; /* do not bother loading tree */
+		else if (get_commit_tree(commit)) {
 			struct tree *tree = get_commit_tree(commit);
 			tree->object.flags |= NOT_USER_GIVEN;
 			add_pending_tree(ctx->revs, tree);
author
#6 lore raw

Re: [PATCH] list-objects: don't queue root trees unless revs->tree_objects is set

Quoted 9 lines · Junio C Hamano
Quoted 7 lines · Jeff King
Quoted 6 lines

I was surprised we ever called repo_get_commit_tree() at all, since we're literally just traversing commits here. It looks like list-objects.c is very happy to queue pending trees for each commit, even if we're just going to throw them away when we get to process_tree()! I wonder if could be checking revs->tree_objects here and saving ourselves some work.

Quoted 8 lines
Quoted 2 lines

Indeed, this seems to help quite a bit in the commit-graph case. I think it's worth doing (and is independent of the other patch).

Yeah, I agree this is very much worth doing and is orthogonal to the other one.

Thanks for spotting it. I wonder if it was broken like this forever since the beginning X-<.

Not quite since the beginning; it comes from 8d2dfc49b1 (process_{tree,blob}: show objects without buffering, 2009-04-10). I suspect nobody noticed because the cost for the normal case is really just shuffling some pointers around. It's only because it actively works against the commit-graph optimizations that it's so expensive there.

I was surprised (and pleased) by how much such a simple thing helped.

-Peff

#7 lore raw

Re: [PATCH] list-objects: don't queue root trees unless revs->tree_objects is set

Quoted 8 lines · Jeff King
Quoted 2 lines

Thanks for spotting it. I wonder if it was broken like this forever since the beginning X-<.

Not quite since the beginning; it comes from 8d2dfc49b1 (process_{tree,blob}: show objects without buffering, 2009-04-10). I suspect nobody noticed because the cost for the normal case is really just shuffling some pointers around. It's only because it actively works against the commit-graph optimizations that it's so expensive there.

Yeah, that is what I meant by "since the beginning" (of commit-graph, that is).

Quoted 1 line

I was surprised (and pleased) by how much such a simple thing helped.

Yes, it is very pleasing. Thanks.

#8 lore raw
In reply to Jeff King
Quoted 22 lines · Jeff King
Quoted 7 lines · Jeff King

When the client has asked for certain shallow options like "deepen-since", we do a custom rev-list walk that pretends to be shallow. Before doing so, we have to disable the commit-graph, since it is not compatible with the shallow view of the repository. That's handled by 829a321569 (commit-graph: close_commit_graph before shallow walk, 2018-08-20). That commit literally closes and frees our repo->objects->commit_graph struct.

A few notes and curiosities on my patch and this general area.

The test suite passes with my patch both with and without GIT_TEST_COMMIT_GRAPH=1. But to my surprise, it also passes if I delete the close_commit_graph() line added by 829a321569!

So it's not clear to me whether this whole thing is truly unnecessary (and Stolee was just being overly cautious because the code is related to shallow-ness, even though it is OK doing a true-parent traversal itself), or if we just don't have good test coverage for the case that requires it.

Quoted 6 lines

If it _is_ necessary, I'm a little worried there are other problems lurking. The whole issue is that we've seen and parsed some commits before we get to this shallow deepen-since code-path. So just disabling commit-graphs isn't enough. Even without them, we might have parsed some commits the old-fashioned way and filled in their parent pointers. Is that a problem?

I am, too, but I don't think we should hold this patch up which is obviously improving the situation in the meantime while we figure that out.

Quoted 1 line

[snip]

Thanks,
Taylor

author
#9 lore raw
Quoted 17 lines · Taylor Blau
Quoted 10 lines

The test suite passes with my patch both with and without GIT_TEST_COMMIT_GRAPH=1. But to my surprise, it also passes if I delete the close_commit_graph() line added by 829a321569!

So it's not clear to me whether this whole thing is truly unnecessary (and Stolee was just being overly cautious because the code is related to shallow-ness, even though it is OK doing a true-parent traversal itself), or if we just don't have good test coverage for the case that requires it.

Quoted 6 lines

If it _is_ necessary, I'm a little worried there are other problems lurking. The whole issue is that we've seen and parsed some commits before we get to this shallow deepen-since code-path. So just disabling commit-graphs isn't enough. Even without them, we might have parsed some commits the old-fashioned way and filled in their parent pointers. Is that a problem?

Quoted 3 lines

I am, too, but I don't think we should hold this patch up which is obviously improving the situation in the meantime while we figure that out.

Yeah, I'd agree here, unless we determine quickly that we do need the bigger fix, and somebody with a bit more knowledge of this shallow code offers a fix. I believe my patch is a strict improvement, and puts the commit-graph code path on par with the regular one.

-Peff

#10 lore raw
In reply to Jeff King

Hi Peff,

Thanks in advance for digging into all of this. I feel guilty for having found the issue myself, and then gotten a headache for just long enough to have you completely fix the issue by the time that I got back. So, thanks :-).

Quoted 45 lines · Jeff King

When the client has asked for certain shallow options like "deepen-since", we do a custom rev-list walk that pretends to be shallow. Before doing so, we have to disable the commit-graph, since it is not compatible with the shallow view of the repository. That's handled by 829a321569 (commit-graph: close_commit_graph before shallow walk, 2018-08-20). That commit literally closes and frees our repo->objects->commit_graph struct.

That creates an interesting problem for commits that have _already_ been parsed using the commit graph. Their commit->object.parsed flag is set, their commit->graph_pos is set, but their commit->maybe_tree may still be NULL. When somebody later calls repo_get_commit_tree(), we see that we haven't loaded the tree oid yet and try to get it from the commit graph. But since it has been freed, we segfault!

So the root of the issue is a data dependency between the commit's lazy-load of the tree oid and the fact that the commit graph can go away mid-process. How can we resolve it?

There are a couple of general approaches:

1. The obvious answer is to avoid loading the tree from the graph when we see that it's NULL. But then what do we return for the tree oid? If we return NULL, our caller in do_traverse() will rightly complain that we have no tree. We'd have to fallback to loading the actual commit object and re-parsing it. That requires teaching parse_commit_buffer() to understand re-parsing (i.e., not starting from a clean slate and not leaking any allocated bits like parent list pointers).

2. When we close the commit graph, walk through the set of in-memory objects and clear any graph_pos pointers. But this means we also have to "unparse" any such commits so that we know they still need to open the commit object to fill in their trees. So it's no less complicated than (1), and is more expensive (since we clear objects we might not later need).

3. Stop freeing the commit-graph struct. Continue to let it be used for lazy-loads of tree oids, but let upload-pack specify that it shouldn't be used for further commit parsing.

4. Push the whole shallow rev-list out to its own sub-process, with the commit-graph disabled from the start, giving it a clean memory space to work from.

Quoted 5 lines

I've chosen (3) here. Options (1) and (2) would work, but are non-trivial to implement. Option (4) is more expensive, and I'm not sure how complicated it is (shelling out for the actual rev-list part is easy, but we do then parse the resulting commits internally, and I'm not clear which parts need to be handling shallow-ness).

Option (3) make the most sense to me, too. When I started hacking on a similar patch, I implemented roughly what you describe in (1), which is to say that I added an extra:

&& r->objects->commit_graph

To ensure that the 'commit_graph' about to cause a segfault was non-NULL after calling 'close_commit_graph'. But, as you pointed out:
what are we to return after that? NULL, since we expected it to be in the commit-graph, but we don't have a commit-graph to check anymore? I don't think that that makes much sense, and what you implemented here is clearly better.

Quoted 3 lines

The new test in t5500 triggers this segfault, but see the comments there for how horribly intimate it has to be with how both upload-pack and commit graphs work.

Thanks for the comment, too. I agree that protocol-level hacking is somewhat smelly, at least as far as t5500 is concerned, but for this particular case it seems the only sensible option.

I'm still left scratching my (sore) head about how someone triggered this in production, but maybe that's a riddle for another day.

Quoted 60 lines
---
 commit-graph.c        | 10 ++++++++++
 commit-graph.h        |  6 ++++++
 t/t5500-fetch-pack.sh | 38 ++++++++++++++++++++++++++++++++++++++
 upload-pack.c         |  2 +-
 4 files changed, 55 insertions(+), 1 deletion(-)
commit-graph.c +10 -0
index 9b02d2c426..bc5dd5913f 100644
@@ -41,6 +41,8 @@
 #define GRAPH_MIN_SIZE (GRAPH_HEADER_SIZE + 4 * GRAPH_CHUNKLOOKUP_WIDTH \
 			+ GRAPH_FANOUT_SIZE + the_hash_algo->rawsz)
 
+static int commit_graph_disabled;
+
 char *get_commit_graph_filename(const char *obj_dir)
 {
 	char *filename = xstrfmt("%s/info/commit-graph", obj_dir);
@@ -472,6 +474,9 @@ static int prepare_commit_graph(struct repository *r)
 		die("dying as requested by the '%s' variable on commit-graph load!",
 		    GIT_TEST_COMMIT_GRAPH_DIE_ON_LOAD);
 
+	if (commit_graph_disabled)
+		return 0;
+
 	if (r->objects->commit_graph_attempted)
 		return !!r->objects->commit_graph;
 	r->objects->commit_graph_attempted = 1;
@@ -2101,3 +2106,8 @@ void free_commit_graph(struct commit_graph *g)
 	free(g->filename);
 	free(g);
 }
+
+void disable_commit_graph(void)
+{
+	commit_graph_disabled = 1;
+}
commit-graph.h +6 -0
index 486e64e591..42d6e7c289 100644
@@ -107,4 +107,10 @@ int verify_commit_graph(struct repository *r, struct commit_graph *g, int flags)
 void close_commit_graph(struct raw_object_store *);
 void free_commit_graph(struct commit_graph *);
 
+/*
+ * Disable further use of the commit graph in this process when parsing a
+ * "struct commit".
+ */
+void disable_commit_graph(void);
+
 #endif
t/t5500-fetch-pack.sh +0 -0
index 8210f63d41..7601664b74 100755

t/t5500-fetch-pack.sh test_expect_success 'clone shallow since selects no commits' ' · line 788

Quoted 6 lines
t/t5500-fetch-pack.sh +3 -0
@@ -783,6 +783,44 @@ test_expect_success 'clone shallow since selects no commits' '
 	)
 '
 
+# A few subtle things about the request in this test:
+#
+#  - the server must have commit-graphs present and enabled

I think "enabled" may be somewhat redundant, at least since some recent changes to enable this by default. (As an aside, I tried to dig up the patches that Stolee sent to actually enable this and back up my claim, but I couldn't find them on 'master'. I'm not sure if that's my poor use of 'git log', or misremembering the fact that these were enabled by default.)

t/t5500-fetch-pack.sh test_expect_success 'clone shallow since selects no commits' ' · line 798

Quoted 10 lines
t/t5500-fetch-pack.sh +10 -0
+#
+#  - the history is such that our want/have share a common ancestor ("base"
+#    here)
+#
+#  - we send only a single have, which is fewer than a normal client would
+#    send. This ensures that we don't parse "base" up front with
+#    parse_object(), but rather traverse to it as a parent while deciding if we
+#    can stop the "have" negotiation, and call parse_commit(). The former
+#    sees the actual object data and so always loads the three oid, whereas the
+#    latter will try to lod it lazily.

s/lod/load

t/t5500-fetch-pack.sh test_expect_success 'clone shallow since selects no commits' ' · line 805

Quoted 7 lines
t/t5500-fetch-pack.sh +7 -0
+#
+#  - we must use protocol v2, because it handles the "have" negotiation before
+#    processing the shallow direectives
+#
+test_expect_success 'shallow since with commit graph and already-seen commit' '
+	test_create_repo shallow-since-graph &&
+	(

I'm not sure if this same-level indentation is common, or you're missing an extra tab here. Either way.

t/t5500-fetch-pack.sh test_expect_success 'clone shallow since selects no commits' ' · line 812

Quoted 7 lines
t/t5500-fetch-pack.sh +7 -0
+	cd shallow-since-graph &&
+	test_commit base &&
+	test_commit master &&
+	git checkout -b other HEAD^ &&
+	test_commit other &&
+	git commit-graph write --reachable &&
+	git config core.commitGraph true &&

Another nit, but do you have any thoughts about using 'test_config' here instead of a pure 'git config'? I don't think that it really would matter much (since none of the other tests hopefully have anything to do with commit-graph, and doubly so if it is enabled by default, _and_ since you're using your own repository), but anyway.

t/t5500-fetch-pack.sh test_expect_success 'clone shallow since selects no commits' ' · line 823

Quoted 11 lines
t/t5500-fetch-pack.sh +11 -0
+
+	GIT_PROTOCOL=version=2 git upload-pack . <<-EOF >/dev/null
+	0012command=fetch
+	00010013deepen-since 1
+	0032want $(git rev-parse other)
+	0032have $(git rev-parse master)
+	0000
+	EOF
+	)
+'
+

Aside from my nit-picking above, the protocol exchange here looks correct, and your rationale backs it up. Likewise, this causes a segfault for me on the tip of 'master', and applying the later part of your patch fixes it for me. Thanks.

upload-pack.c static void deepen_by_rev_list(struct packet_writer *writer, int ac, · line 725

Quoted 13 lines
t/t5500-fetch-pack.sh +0 -0
 test_expect_success 'shallow clone exclude tag two' '
 	test_create_repo shallow-exclude &&
 	(
upload-pack.c +1 -1
index 222cd3ad89..a260b6b50d 100644
@@ -722,7 +722,7 @@ static void deepen_by_rev_list(struct packet_writer *writer, int ac,
 {
 	struct commit_list *result;
 
-	close_commit_graph(the_repository->objects);
+	disable_commit_graph();

This line made me think to check if we could remove 'close_commit_graph' all together, but there is a remaining callsite in packfile.c, and closing the commit-graph _is_ the right thing to do there, so I think we ought to keep it around.

Quoted 5 lines
upload-pack.c +0 -1
 	result = get_shallow_commits_by_rev_list(ac, av, SHALLOW, NOT_SHALLOW);
 	send_shallow(writer, result);
 	free_commit_list(result);
--

2.23.0.663.gbe3d117559

Thanks,
Taylor

#11 lore raw
Quoted 7 lines · Taylor Blau

t/t5500-fetch-pack.sh test_expect_success 'clone shallow since selects no commits' ' · line 801

Quoted 7 lines
t/t5500-fetch-pack.sh +2 -0
index 8210f63d41..7601664b74 100755
 
+#  - we must use protocol v2, because it handles the "have" negotiation before
+#    processing the shallow direectives

s/ee/e/

Quoted 8 lines

t/t5500-fetch-pack.sh test_expect_success 'clone shallow since selects no commits' ' · line 805

Quoted 4 lines
t/t5500-fetch-pack.sh +4 -0
+#
+test_expect_success 'shallow since with commit graph and already-seen commit' '
+	test_create_repo shallow-since-graph &&
+	(

I'm not sure if this same-level indentation is common, or you're missing an extra tab here. Either way.

Quoted 13 lines

t/t5500-fetch-pack.sh test_expect_success 'clone shallow since selects no commits' ' · line 812

Quoted 7 lines
t/t5500-fetch-pack.sh +7 -0
+	cd shallow-since-graph &&
+	test_commit base &&
+	test_commit master &&
+	git checkout -b other HEAD^ &&
+	test_commit other &&
+	git commit-graph write --reachable &&
+	git config core.commitGraph true &&

Another nit, but do you have any thoughts about using 'test_config' here instead of a pure 'git config'? I don't think that it really would matter much (since none of the other tests hopefully have anything to do with commit-graph, and doubly so if it is enabled by default, _and_ since you're using your own repository), but anyway.

We can't simply replace that 'git config' command with 'test_config', because it is implemented using 'test_when_finished', which doesn't work in a subshell [1]. What we could do is:

test_create_repo shallow-since-graph && test_config -C shallow-since-graph core.commitGraph true && (

     cd shallow-since-graph &&
     ....

Or we could entirely avoid the subshell by passing '-C shallow-since-graph' to every single command... [2]

However, since this repo was specifically created for this test, it doesn't really matter in what state it's left behind, so I don't think it's worth it.

[1] 0968f12a99 (test-lib-functions: detect test_when_finished in

    subshell, 2015-09-05)

[2] Or bite the bullet, and declare that every test case shall start in $TRASH_DIRECTORY.

author
#12 lore raw
Quoted 4 lines · SZEDER Gábor
Quoted 2 lines

t/t5500-fetch-pack.sh test_expect_success 'clone shallow since selects no commits' ' · line 801

Quoted 2 lines
t/t5500-fetch-pack.sh +2 -0
+#  - we must use protocol v2, because it handles the "have" negotiation before
+#    processing the shallow direectives

s/ee/e/

Thanks, fixed.

Quoted 10 lines

We can't simply replace that 'git config' command with 'test_config', because it is implemented using 'test_when_finished', which doesn't work in a subshell [1]. What we could do is:

test_create_repo shallow-since-graph && test_config -C shallow-since-graph core.commitGraph true && (

     cd shallow-since-graph &&
     ....
Quoted 6 lines

Or we could entirely avoid the subshell by passing '-C shallow-since-graph' to every single command... [2]

However, since this repo was specifically created for this test, it doesn't really matter in what state it's left behind, so I don't think it's worth it.

Yep, agreed on all of this. :)

-Peff

#13 lore raw
In reply to Taylor Blau

On 9/11/2019 10:07 PM, Taylor Blau wrote:>>

Quoted 10 lines

t/t5500-fetch-pack.sh test_expect_success 'clone shallow since selects no commits' ' · line 788

Quoted 3 lines
t/t5500-fetch-pack.sh +3 -0
+# A few subtle things about the request in this test:
+#
+#  - the server must have commit-graphs present and enabled

I think "enabled" may be somewhat redundant, at least since some recent changes to enable this by default. (As an aside, I tried to dig up the patches that Stolee sent to actually enable this and back up my claim, but I couldn't find them on 'master'. I'm not sure if that's my poor use of 'git log', or misremembering the fact that these were enabled by default.)

Commit 31b1de6 ("commit-graph: turn on commit-graph by default" 2019-08-13) is part of ds/feature-macros and seems to be in master (at least in gitster/git).

Having core.commitGraph=true does very little if the commit-graph is not written, but gc.writeCommitGraph is also enabled by default in that commit.

Thanks,
-Stolee

author
#14 lore raw
In reply to Taylor Blau
Quoted 10 lines · Taylor Blau
Quoted 3 lines

The new test in t5500 triggers this segfault, but see the comments there for how horribly intimate it has to be with how both upload-pack and commit graphs work.

Thanks for the comment, too. I agree that protocol-level hacking is somewhat smelly, at least as far as t5500 is concerned, but for this particular case it seems the only sensible option.

I'm still left scratching my (sore) head about how someone triggered this in production, but maybe that's a riddle for another day.

I'll admit that part of my motivation was my inability to generate a working case just using Git commands (so my justification is that if it's so hard to do, then the test is inherently fragile, but you can also just accuse me of being lazy).

I think the key is something to do with the shape of history related to the requests, such that we walk over a commit during the have negotiation, but then also need to walk over it during the deepen-since. So maybe I am just missing something obvious, or maybe it just needs to be a deeper history. I do like that the case I showed is the minimal history, at least.

Quoted 10 lines

t/t5500-fetch-pack.sh test_expect_success 'clone shallow since selects no commits' ' · line 788

Quoted 3 lines
t/t5500-fetch-pack.sh +3 -0
+# A few subtle things about the request in this test:
+#
+#  - the server must have commit-graphs present and enabled

I think "enabled" may be somewhat redundant, at least since some recent changes to enable this by default. (As an aside, I tried to dig up the patches that Stolee sent to actually enable this and back up my claim, but I couldn't find them on 'master'. I'm not sure if that's my poor use of 'git log', or misremembering the fact that these were enabled by default.)

Yeah, it is redundant these days. I figured this might go to 'maint', though, and 31b1de6a09 (commit-graph: turn on commit-graph by default, 2019-08-13) is only in master.

Quoted 3 lines
Quoted 1 line

+# latter will try to lod it lazily.

s/lod/load

Thanks, fixed.

Quoted 10 lines

t/t5500-fetch-pack.sh test_expect_success 'clone shallow since selects no commits' ' · line 805

Quoted 7 lines
t/t5500-fetch-pack.sh +7 -0
+#
+#  - we must use protocol v2, because it handles the "have" negotiation before
+#    processing the shallow direectives
+#
+test_expect_success 'shallow since with commit graph and already-seen commit' '
+	test_create_repo shallow-since-graph &&
+	(

I'm not sure if this same-level indentation is common, or you're missing an extra tab here. Either way.

It's not common, but I was matching the surrounding tests for style (and use of a separate repo, and non-use of test_config).

Quoted 8 lines
Quoted 8 lines
upload-pack.c +0 -0
index 222cd3ad89..a260b6b50d 100644
@@ -722,7 +722,7 @@ static void deepen_by_rev_list(struct packet_writer *writer, int ac,
 {
 	struct commit_list *result;
 
Quoted 7 lines

upload-pack.c static void deepen_by_rev_list(struct packet_writer *writer, int ac, · line 725

Quoted 2 lines
upload-pack.c +1 -1
-	close_commit_graph(the_repository->objects);
+	disable_commit_graph();

This line made me think to check if we could remove 'close_commit_graph' all together, but there is a remaining callsite in packfile.c, and closing the commit-graph _is_ the right thing to do there, so I think we ought to keep it around.

Yeah, it's sort of inherently dangerous, as it doesn't scrub any in-memory commit structs that are in this "have a graph_pos but not maybe_tree" state.

The call in close_object_store() is probably fine, though, as the whole point there is getting rid of everything related to the object store. And since 14ba97f81c (alloc: allow arbitrary repositories for alloc functions, 2018-05-15) or thereabouts, that includes the object structs.

There's another call in write_commit_graph_file(), right before we rename a new file into place. This generally happens in a separate "commit-graph write" process, so I think it's OK. But it could eventually happen as part of another process, which would maybe be a problem. I'm actually not convinced the in-memory one needs to be closed here, but maybe it's a Windows thing (closing the file before renaming over it)?

-Peff

#15 lore raw
In reply to Jeff King
Quoted 8 lines · Jeff King

When the client has asked for certain shallow options like "deepen-since", we do a custom rev-list walk that pretends to be shallow. Before doing so, we have to disable the commit-graph, since it is not compatible with the shallow view of the repository. That's handled by 829a321569 (commit-graph: close_commit_graph before shallow walk, 2018-08-20). That commit literally closes and frees our repo->objects->commit_graph struct.

Quoted 6 lines

That creates an interesting problem for commits that have _already_ been parsed using the commit graph. Their commit->object.parsed flag is set, their commit->graph_pos is set, but their commit->maybe_tree may still be NULL. When somebody later calls repo_get_commit_tree(), we see that we haven't loaded the tree oid yet and try to get it from the commit graph. But since it has been freed, we segfault!

OOPS! That is certainly a bad thing. I'm glad you found it, but I am sorry for how you (probably) found it.

Quoted 30 lines

So the root of the issue is a data dependency between the commit's lazy-load of the tree oid and the fact that the commit graph can go away mid-process. How can we resolve it?

There are a couple of general approaches:

1. The obvious answer is to avoid loading the tree from the graph when we see that it's NULL. But then what do we return for the tree oid? If we return NULL, our caller in do_traverse() will rightly complain that we have no tree. We'd have to fallback to loading the actual commit object and re-parsing it. That requires teaching parse_commit_buffer() to understand re-parsing (i.e., not starting from a clean slate and not leaking any allocated bits like parent list pointers).

2. When we close the commit graph, walk through the set of in-memory objects and clear any graph_pos pointers. But this means we also have to "unparse" any such commits so that we know they still need to open the commit object to fill in their trees. So it's no less complicated than (1), and is more expensive (since we clear objects we might not later need).

3. Stop freeing the commit-graph struct. Continue to let it be used for lazy-loads of tree oids, but let upload-pack specify that it shouldn't be used for further commit parsing.

4. Push the whole shallow rev-list out to its own sub-process, with the commit-graph disabled from the start, giving it a clean memory space to work from.

Quoted 5 lines

I've chosen (3) here. Options (1) and (2) would work, but are non-trivial to implement. Option (4) is more expensive, and I'm not sure how complicated it is (shelling out for the actual rev-list part is easy, but we do then parse the resulting commits internally, and I'm not clear which parts need to be handling shallow-ness).

I agree that (3) is the best option. The commit-graph is just a database of commit data, and we are choosing to interpret the parent data differently. The tree data is perfectly good.

Quoted 17 lines

The new test in t5500 triggers this segfault, but see the comments there for how horribly intimate it has to be with how both upload-pack and commit graphs work.

---
 commit-graph.c        | 10 ++++++++++
 commit-graph.h        |  6 ++++++
 t/t5500-fetch-pack.sh | 38 ++++++++++++++++++++++++++++++++++++++
 upload-pack.c         |  2 +-
 4 files changed, 55 insertions(+), 1 deletion(-)
commit-graph.c +0 -0
index 9b02d2c426..bc5dd5913f 100644

commit-graph.c line 44

Quoted 4 lines
commit-graph.c +1 -0
@@ -41,6 +41,8 @@
 #define GRAPH_MIN_SIZE (GRAPH_HEADER_SIZE + 4 * GRAPH_CHUNKLOOKUP_WIDTH \
 			+ GRAPH_FANOUT_SIZE + the_hash_algo->rawsz)
 
+static int commit_graph_disabled;

Should we be putting this inside the repository struct instead?

Quoted 15 lines
commit-graph.c +4 -0
+
 char *get_commit_graph_filename(const char *obj_dir)
 {
 	char *filename = xstrfmt("%s/info/commit-graph", obj_dir);
@@ -472,6 +474,9 @@ static int prepare_commit_graph(struct repository *r)
 		die("dying as requested by the '%s' variable on commit-graph load!",
 		    GIT_TEST_COMMIT_GRAPH_DIE_ON_LOAD);
 
+	if (commit_graph_disabled)
+		return 0;
+
 	if (r->objects->commit_graph_attempted)
 		return !!r->objects->commit_graph;
 	r->objects->commit_graph_attempted = 1;

commit-graph.c void free_commit_graph(struct commit_graph *g) · line 2113

Quoted 8 lines
commit-graph.c +5 -0
@@ -2101,3 +2106,8 @@ void free_commit_graph(struct commit_graph *g)
 	free(g->filename);
 	free(g);
 }
+
+void disable_commit_graph(void)
+{
+	commit_graph_disabled = 1;
+}

Then this would take a struct repository *r.

Quoted 70 lines
commit-graph.h +6 -0
index 486e64e591..42d6e7c289 100644
@@ -107,4 +107,10 @@ int verify_commit_graph(struct repository *r, struct commit_graph *g, int flags)
 void close_commit_graph(struct raw_object_store *);
 void free_commit_graph(struct commit_graph *);
 
+/*
+ * Disable further use of the commit graph in this process when parsing a
+ * "struct commit".
+ */
+void disable_commit_graph(void);
+
 #endif
t/t5500-fetch-pack.sh +38 -0
index 8210f63d41..7601664b74 100755
@@ -783,6 +783,44 @@ test_expect_success 'clone shallow since selects no commits' '
 	)
 '
 
+# A few subtle things about the request in this test:
+#
+#  - the server must have commit-graphs present and enabled
+#
+#  - the history is such that our want/have share a common ancestor ("base"
+#    here)
+#
+#  - we send only a single have, which is fewer than a normal client would
+#    send. This ensures that we don't parse "base" up front with
+#    parse_object(), but rather traverse to it as a parent while deciding if we
+#    can stop the "have" negotiation, and call parse_commit(). The former
+#    sees the actual object data and so always loads the three oid, whereas the
+#    latter will try to lod it lazily.
+#
+#  - we must use protocol v2, because it handles the "have" negotiation before
+#    processing the shallow direectives
+#
+test_expect_success 'shallow since with commit graph and already-seen commit' '
+	test_create_repo shallow-since-graph &&
+	(
+	cd shallow-since-graph &&
+	test_commit base &&
+	test_commit master &&
+	git checkout -b other HEAD^ &&
+	test_commit other &&
+	git commit-graph write --reachable &&
+	git config core.commitGraph true &&
+
+	GIT_PROTOCOL=version=2 git upload-pack . <<-EOF >/dev/null
+	0012command=fetch
+	00010013deepen-since 1
+	0032want $(git rev-parse other)
+	0032have $(git rev-parse master)
+	0000
+	EOF
+	)
+'
+
 test_expect_success 'shallow clone exclude tag two' '
 	test_create_repo shallow-exclude &&
 	(
upload-pack.c +0 -0
index 222cd3ad89..a260b6b50d 100644

upload-pack.c static void deepen_by_rev_list(struct packet_writer *writer, int ac, · line 728

Quoted 8 lines
upload-pack.c +1 -1
@@ -722,7 +722,7 @@ static void deepen_by_rev_list(struct packet_writer *writer, int ac,
 {
 	struct commit_list *result;
 
-	close_commit_graph(the_repository->objects);
+	disable_commit_graph();
 	result = get_shallow_commits_by_rev_list(ac, av, SHALLOW, NOT_SHALLOW);
 	send_shallow(writer, result);
 	free_commit_list(result);

Your patch does not seem to actually cover the "I've already parsed some commits" case, as you are only preventing the commit-graph from being prepared. Instead, we need to have a short-circuit inside parse_commit() to avoid future parsing from the commit-graph file.

(I'm going to the rest of the thread messages now to see if I just repeated someone else's concerns.)

Thanks,
-Stolee

author
#16 lore raw
Quoted 9 lines · Derrick Stolee
Quoted 6 lines

That creates an interesting problem for commits that have _already_ been parsed using the commit graph. Their commit->object.parsed flag is set, their commit->graph_pos is set, but their commit->maybe_tree may still be NULL. When somebody later calls repo_get_commit_tree(), we see that we haven't loaded the tree oid yet and try to get it from the commit graph. But since it has been freed, we segfault!

OOPS! That is certainly a bad thing. I'm glad you found it, but I am sorry for how you (probably) found it.

Heh. I'll admit it was quite a slog of debugging, but _most_ of that was figuring out in which circumstance we'd have actually parsed the object. Finding the problematic end state was pretty easy from a coredump. :)

Quoted 11 lines

commit-graph.c line 44

Quoted 9 lines
commit-graph.c +1 -0
index 9b02d2c426..bc5dd5913f 100644
@@ -41,6 +41,8 @@
 #define GRAPH_MIN_SIZE (GRAPH_HEADER_SIZE + 4 * GRAPH_CHUNKLOOKUP_WIDTH \
 			+ GRAPH_FANOUT_SIZE + the_hash_algo->rawsz)
 
+static int commit_graph_disabled;

Should we be putting this inside the repository struct instead?

Probably. The only caller will just pass the_repository, but it doesn't hurt to scope it down now.

It could potentially go into the commit_graph itself, but it looks like with the incremental work we may have multiple such structs. It could also go into raw_object_store, but I think conceptually it's a repo-level thing.

So I put it straight into "struct repository".

Quoted 4 lines

Your patch does not seem to actually cover the "I've already parsed some commits" case, as you are only preventing the commit-graph from being prepared. Instead, we need to have a short-circuit inside parse_commit() to avoid future parsing from the commit-graph file.

Maybe I was too clever, then. :)

I didn't want to have to sprinkle "are we disabled" in parse_commit(), etc. But any such uses of the commit graph have to do:

if (!prepare_commit_graph(r))

	return;

to lazy-load it. So the logic to prepare becomes (roughly):

if (disabled)

	return 0;
  if (already_loaded)
	return 1;

return actually_load() ? 1 : 0;

and "disabled" takes precedence.

I've added this comment in prepare_commit_graph():

        /*
         * This must come before the "already attempted?" check below, because
         * we want to disable even an already-loaded graph file.
         */
        if (r->commit_graph_disabled)
                return 0;

        if (r->objects->commit_graph_attempted)
                return !!r->objects->commit_graph;
        r->objects->commit_graph_attempted = 1;

Does that make more sense?

Unrelated, but I also notice the top of prepare_commit_graph() has:

        if (git_env_bool(GIT_TEST_COMMIT_GRAPH_DIE_ON_LOAD, 0))
                die("dying as requested by the '%s' variable on commit-graph load!",
                    GIT_TEST_COMMIT_GRAPH_DIE_ON_LOAD);

as the very first thing. Meaning we're calling getenv() as part of every single parse_commit(), rather than just once per process. Seems like an easy efficiency win.

-Peff

#17 lore raw
Quoted 74 lines · Jeff King
Quoted 9 lines · Derrick Stolee
Quoted 6 lines

That creates an interesting problem for commits that have _already_ been parsed using the commit graph. Their commit->object.parsed flag is set, their commit->graph_pos is set, but their commit->maybe_tree may still be NULL. When somebody later calls repo_get_commit_tree(), we see that we haven't loaded the tree oid yet and try to get it from the commit graph. But since it has been freed, we segfault!

OOPS! That is certainly a bad thing. I'm glad you found it, but I am sorry for how you (probably) found it.

Heh. I'll admit it was quite a slog of debugging, but _most_ of that was figuring out in which circumstance we'd have actually parsed the object. Finding the problematic end state was pretty easy from a coredump. :)

Quoted 11 lines

commit-graph.c line 44

Quoted 9 lines
commit-graph.c +1 -0
index 9b02d2c426..bc5dd5913f 100644
@@ -41,6 +41,8 @@
 #define GRAPH_MIN_SIZE (GRAPH_HEADER_SIZE + 4 * GRAPH_CHUNKLOOKUP_WIDTH \
 			+ GRAPH_FANOUT_SIZE + the_hash_algo->rawsz)
 
+static int commit_graph_disabled;

Should we be putting this inside the repository struct instead?

Probably. The only caller will just pass the_repository, but it doesn't hurt to scope it down now.

It could potentially go into the commit_graph itself, but it looks like with the incremental work we may have multiple such structs. It could also go into raw_object_store, but I think conceptually it's a repo-level thing.

So I put it straight into "struct repository".

Quoted 4 lines

Your patch does not seem to actually cover the "I've already parsed some commits" case, as you are only preventing the commit-graph from being prepared. Instead, we need to have a short-circuit inside parse_commit() to avoid future parsing from the commit-graph file.

Maybe I was too clever, then. :)

I didn't want to have to sprinkle "are we disabled" in parse_commit(), etc. But any such uses of the commit graph have to do:

if (!prepare_commit_graph(r))

	return;

to lazy-load it. So the logic to prepare becomes (roughly):

if (disabled)

	return 0;
  if (already_loaded)
	return 1;

return actually_load() ? 1 : 0;

and "disabled" takes precedence.

I've added this comment in prepare_commit_graph():

        /*
         * This must come before the "already attempted?" check below, because
         * we want to disable even an already-loaded graph file.
         */
        if (r->commit_graph_disabled)
                return 0;

        if (r->objects->commit_graph_attempted)
                return !!r->objects->commit_graph;
        r->objects->commit_graph_attempted = 1;
Quoted 1 line

Does that make more sense?

Ah. That does make sense. I now see the connection between parsing and this change.

Quoted 9 lines

Unrelated, but I also notice the top of prepare_commit_graph() has:

        if (git_env_bool(GIT_TEST_COMMIT_GRAPH_DIE_ON_LOAD, 0))
                die("dying as requested by the '%s' variable on commit-graph load!",
                    GIT_TEST_COMMIT_GRAPH_DIE_ON_LOAD);

as the very first thing. Meaning we're calling getenv() as part of every single parse_commit(), rather than just once per process. Seems like an easy efficiency win.

Absolutely. Move this to after the "have we attempted already?" condition.

Thanks,
-Stolee

author
#18 lore raw

PATCH v2 1/1 upload-pack commit graph segfault fix

In reply to Jeff King

Here's a re-roll of my "disable commit graph more gently" fix. Versus v1:

- I've included a preparatory patch that speeds up prepare_commit_graph(). It's not strictly related, but there's a textual dependency. It could be easily spun off to its own series.

- a comment points out that the ordering of the "disable" check is important

- disable_commit_graph() now works on a repository struct

- typo fixes in the test comments

[1/2]: commit-graph: bump DIE_ON_LOAD check to actual load-time [2/2]: upload-pack: disable commit graph more gently for shallow traversal

 commit-graph.c        | 18 +++++++++++++++---
 commit-graph.h        |  6 ++++++
 repository.h          |  3 +++
 t/t5500-fetch-pack.sh | 38 ++++++++++++++++++++++++++++++++++++++
 upload-pack.c         |  2 +-
 5 files changed, 63 insertions(+), 4 deletions(-)

-Peff

author
#19 lore raw

Re: [PATCH v2] upload-pack commit graph segfault fix

Quoted 6 lines · Jeff King

Here's a re-roll of my "disable commit graph more gently" fix. Versus v1:

- I've included a preparatory patch that speeds up prepare_commit_graph(). It's not strictly related, but there's a textual dependency. It could be easily spun off to its own series.

I _didn't_ include here the other speedup from this thread:

https://public-inbox.org/git/[email protected]/

That's worth doing, but is totally independent (conceptually and textually).

-Peff

author
#20 lore raw

PATCH v2 1/2 commit-graph: bump DIE_ON_LOAD check to actual load-time

In reply to Jeff King

Commit 43d3561805 (commit-graph write: don't die if the existing graph is corrupt, 2019-03-25) added an environment variable we use only in the test suite, $GIT_TEST_COMMIT_GRAPH_DIE_ON_LOAD. But it put the check for this variable at the very top of prepare_commit_graph(), which is called every time we want to use the commit graph. Most importantly, it comes _before_ we check the fast-path "did we already try to load?", meaning we end up calling getenv() for every single use of the commit graph, rather than just when we load.

getenv() is allowed to have unexpected side effects, but that shouldn't be a problem here; we're lazy-loading the graph so it's clear that at least _one_ invocation of this function is going to call it.

But it is inefficient. getenv() typically has to do a linear search through the environment space.

We could memoize the call, but it's simpler still to just bump the check down to the actual loading step. That's fine for our sole user in t5318, and produces this minor real-world speedup:

[before]
Benchmark #1: git -C linux rev-list HEAD >/dev/null

    Time (mean ± σ):      1.460 s ±  0.017 s    [User: 1.174 s, System: 0.285 s]
    Range (min … max):    1.440 s …  1.491 s    10 runs

[after]
Benchmark #1: git -C linux rev-list HEAD >/dev/null

    Time (mean ± σ):      1.391 s ±  0.005 s    [User: 1.118 s, System: 0.273 s]
    Range (min … max):    1.385 s …  1.399 s    10 runs

Of course that actual speedup depends on how big your environment is. We can game it like this:

  for i in $(seq 10000); do
    export dummy$i=$i
  done

in which case I get:

[before]
Benchmark #1: git -C linux rev-list HEAD >/dev/null

    Time (mean ± σ):      6.257 s ±  0.061 s    [User: 6.005 s, System: 0.250 s]
    Range (min … max):    6.174 s …  6.337 s    10 runs

[after]
Benchmark #1: git -C linux rev-list HEAD >/dev/null
Time (mean ± σ): 1.403 s ± 0.005 s [User: 1.146 s, System: 0.256 s]
Range (min … max): 1.396 s … 1.412 s 10 runs

So this is really more about avoiding the pathological case than providing a big real-world speedup.

---
 commit-graph.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)
commit-graph.c +4 -4
index 9b02d2c426..baeaf0d1bf 100644
@@ -468,14 +468,14 @@ static int prepare_commit_graph(struct repository *r)
 {
 	struct object_directory *odb;
 
-	if (git_env_bool(GIT_TEST_COMMIT_GRAPH_DIE_ON_LOAD, 0))
-		die("dying as requested by the '%s' variable on commit-graph load!",
-		    GIT_TEST_COMMIT_GRAPH_DIE_ON_LOAD);
-
 	if (r->objects->commit_graph_attempted)
 		return !!r->objects->commit_graph;
 	r->objects->commit_graph_attempted = 1;
 
+	if (git_env_bool(GIT_TEST_COMMIT_GRAPH_DIE_ON_LOAD, 0))
+		die("dying as requested by the '%s' variable on commit-graph load!",
+		    GIT_TEST_COMMIT_GRAPH_DIE_ON_LOAD);
+
 	prepare_repo_settings(r);
 
 	if (!git_env_bool(GIT_TEST_COMMIT_GRAPH, 0) &&
Signature · 1 line
2.23.0.667.gcccf1fbb03
#21 lore raw

Re: [PATCH v2 1/2] commit-graph: bump DIE_ON_LOAD check to actual load-time

Quoted 22 lines · Jeff King

Commit 43d3561805 (commit-graph write: don't die if the existing graph is corrupt, 2019-03-25) added an environment variable we use only in the test suite, $GIT_TEST_COMMIT_GRAPH_DIE_ON_LOAD. But it put the check for this variable at the very top of prepare_commit_graph(), which is called every time we want to use the commit graph. Most importantly, it comes _before_ we check the fast-path "did we already try to load?", meaning we end up calling getenv() for every single use of the commit graph, rather than just when we load.

getenv() is allowed to have unexpected side effects, but that shouldn't be a problem here; we're lazy-loading the graph so it's clear that at least _one_ invocation of this function is going to call it.

But it is inefficient. getenv() typically has to do a linear search through the environment space.

We could memoize the call, but it's simpler still to just bump the check down to the actual loading step. That's fine for our sole user in t5318, and produces this minor real-world speedup:

[before]
Benchmark #1: git -C linux rev-list HEAD >/dev/null

Quoted 7 lines
    Time (mean ± σ):      1.460 s ±  0.017 s    [User: 1.174 s, System: 0.285 s]
    Range (min … max):    1.440 s …  1.491 s    10 runs

[after]
Benchmark #1: git -C linux rev-list HEAD >/dev/null

    Time (mean ± σ):      1.391 s ±  0.005 s    [User: 1.118 s, System: 0.273 s]
    Range (min … max):    1.385 s …  1.399 s    10 runs

This looks like an important improvement on its own.

Quoted 14 lines

Of course that actual speedup depends on how big your environment is. We can game it like this:

  for i in $(seq 10000); do
    export dummy$i=$i
  done

in which case I get:

[before]
Benchmark #1: git -C linux rev-list HEAD >/dev/null

    Time (mean ± σ):      6.257 s ±  0.061 s    [User: 6.005 s, System: 0.250 s]
    Range (min … max):    6.174 s …  6.337 s    10 runs
Quoted 7 lines

[after]
Benchmark #1: git -C linux rev-list HEAD >/dev/null
Time (mean ± σ): 1.403 s ± 0.005 s [User: 1.146 s, System: 0.256 s]
Range (min … max): 1.396 s … 1.412 s 10 runs

So this is really more about avoiding the pathological case than providing a big real-world speedup.

This change is stunning. I'm _sure_ someone is hurting with this.

Quoted 18 lines
---
 commit-graph.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)
commit-graph.c +0 -4
index 9b02d2c426..baeaf0d1bf 100644
@@ -468,14 +468,14 @@ static int prepare_commit_graph(struct repository *r)
 {
 	struct object_directory *odb;
 
-	if (git_env_bool(GIT_TEST_COMMIT_GRAPH_DIE_ON_LOAD, 0))
-		die("dying as requested by the '%s' variable on commit-graph load!",
-		    GIT_TEST_COMMIT_GRAPH_DIE_ON_LOAD);
-
 	if (r->objects->commit_graph_attempted)

commit-graph.c static int prepare_commit_graph(struct repository *r) · line 481

Quoted 10 lines
commit-graph.c +4 -0
 		return !!r->objects->commit_graph;
 	r->objects->commit_graph_attempted = 1;
 
+	if (git_env_bool(GIT_TEST_COMMIT_GRAPH_DIE_ON_LOAD, 0))
+		die("dying as requested by the '%s' variable on commit-graph load!",
+		    GIT_TEST_COMMIT_GRAPH_DIE_ON_LOAD);
+
 	prepare_repo_settings(r);
 
 	if (!git_env_bool(GIT_TEST_COMMIT_GRAPH, 0) &&

LGTM, thanks!

-Stolee

author
#22 lore raw

PATCH v2 2/2 upload-pack: disable commit graph more gently for shallow traversal

In reply to Jeff King

When the client has asked for certain shallow options like "deepen-since", we do a custom rev-list walk that pretends to be shallow. Before doing so, we have to disable the commit-graph, since it is not compatible with the shallow view of the repository. That's handled by 829a321569 (commit-graph: close_commit_graph before shallow walk, 2018-08-20). That commit literally closes and frees our repo->objects->commit_graph struct.

That creates an interesting problem for commits that have _already_ been parsed using the commit graph. Their commit->object.parsed flag is set, their commit->graph_pos is set, but their commit->maybe_tree may still be NULL. When somebody later calls repo_get_commit_tree(), we see that we haven't loaded the tree oid yet and try to get it from the commit graph. But since it has been freed, we segfault!

So the root of the issue is a data dependency between the commit's lazy-load of the tree oid and the fact that the commit graph can go away mid-process. How can we resolve it?

There are a couple of general approaches:

1. The obvious answer is to avoid loading the tree from the graph when we see that it's NULL. But then what do we return for the tree oid? If we return NULL, our caller in do_traverse() will rightly complain that we have no tree. We'd have to fallback to loading the actual commit object and re-parsing it. That requires teaching parse_commit_buffer() to understand re-parsing (i.e., not starting from a clean slate and not leaking any allocated bits like parent list pointers).

2. When we close the commit graph, walk through the set of in-memory objects and clear any graph_pos pointers. But this means we also have to "unparse" any such commits so that we know they still need to open the commit object to fill in their trees. So it's no less complicated than (1), and is more expensive (since we clear objects we might not later need).

3. Stop freeing the commit-graph struct. Continue to let it be used for lazy-loads of tree oids, but let upload-pack specify that it shouldn't be used for further commit parsing.

4. Push the whole shallow rev-list out to its own sub-process, with the commit-graph disabled from the start, giving it a clean memory space to work from.

I've chosen (3) here. Options (1) and (2) would work, but are non-trivial to implement. Option (4) is more expensive, and I'm not sure how complicated it is (shelling out for the actual rev-list part is easy, but we do then parse the resulting commits internally, and I'm not clear which parts need to be handling shallow-ness).

The new test in t5500 triggers this segfault, but see the comments there for how horribly intimate it has to be with how both upload-pack and commit graphs work.

---
 commit-graph.c        | 12 ++++++++++++
 commit-graph.h        |  6 ++++++
 repository.h          |  3 +++
 t/t5500-fetch-pack.sh | 38 ++++++++++++++++++++++++++++++++++++++
 upload-pack.c         |  2 +-
 5 files changed, 60 insertions(+), 1 deletion(-)
commit-graph.c +12 -0
index baeaf0d1bf..bbde647f8b 100644
@@ -468,6 +468,13 @@ static int prepare_commit_graph(struct repository *r)
 {
 	struct object_directory *odb;
 
+	/*
+	 * This must come before the "already attempted?" check below, because
+	 * we want to disable even an already-loaded graph file.
+	 */
+	if (r->commit_graph_disabled)
+		return 0;
+
 	if (r->objects->commit_graph_attempted)
 		return !!r->objects->commit_graph;
 	r->objects->commit_graph_attempted = 1;
@@ -2101,3 +2108,8 @@ void free_commit_graph(struct commit_graph *g)
 	free(g->filename);
 	free(g);
 }
+
+void disable_commit_graph(struct repository *r)
+{
+	r->commit_graph_disabled = 1;
+}
commit-graph.h +6 -0
index 486e64e591..7f5c933fa2 100644
@@ -107,4 +107,10 @@ int verify_commit_graph(struct repository *r, struct commit_graph *g, int flags)
 void close_commit_graph(struct raw_object_store *);
 void free_commit_graph(struct commit_graph *);
 
+/*
+ * Disable further use of the commit graph in this process when parsing a
+ * "struct commit".
+ */
+void disable_commit_graph(struct repository *r);
+
 #endif
repository.h +3 -0
index 4da275e73f..84335292cd 100644
@@ -124,6 +124,9 @@ struct repository {
 	/* A unique-id for tracing purposes. */
 	int trace2_repo_id;
 
+	/* True if commit-graph has been disabled within this process. */
+	int commit_graph_disabled;
+
 	/* Configurations */
 
 	/* Indicate if a repository has a different 'commondir' from 'gitdir' */
t/t5500-fetch-pack.sh +38 -0
index 8210f63d41..244c3e7062 100755
@@ -783,6 +783,44 @@ test_expect_success 'clone shallow since selects no commits' '
 	)
 '
 
+# A few subtle things about the request in this test:
+#
+#  - the server must have commit-graphs present and enabled
+#
+#  - the history is such that our want/have share a common ancestor ("base"
+#    here)
+#
+#  - we send only a single have, which is fewer than a normal client would
+#    send. This ensures that we don't parse "base" up front with
+#    parse_object(), but rather traverse to it as a parent while deciding if we
+#    can stop the "have" negotiation, and call parse_commit(). The former
+#    sees the actual object data and so always loads the three oid, whereas the
+#    latter will try to load it lazily.
+#
+#  - we must use protocol v2, because it handles the "have" negotiation before
+#    processing the shallow directives
+#
+test_expect_success 'shallow since with commit graph and already-seen commit' '
+	test_create_repo shallow-since-graph &&
+	(
+	cd shallow-since-graph &&
+	test_commit base &&
+	test_commit master &&
+	git checkout -b other HEAD^ &&
+	test_commit other &&
+	git commit-graph write --reachable &&
+	git config core.commitGraph true &&
+
+	GIT_PROTOCOL=version=2 git upload-pack . <<-EOF >/dev/null
+	0012command=fetch
+	00010013deepen-since 1
+	0032want $(git rev-parse other)
+	0032have $(git rev-parse master)
+	0000
+	EOF
+	)
+'
+
 test_expect_success 'shallow clone exclude tag two' '
 	test_create_repo shallow-exclude &&
 	(
upload-pack.c +1 -1
index 222cd3ad89..135bb3f6cc 100644
@@ -722,7 +722,7 @@ static void deepen_by_rev_list(struct packet_writer *writer, int ac,
 {
 	struct commit_list *result;
 
-	close_commit_graph(the_repository->objects);
+	disable_commit_graph(the_repository);
 	result = get_shallow_commits_by_rev_list(ac, av, SHALLOW, NOT_SHALLOW);
 	send_shallow(writer, result);
 	free_commit_list(result);
Signature · 1 line
2.23.0.667.gcccf1fbb03
#23 lore raw
Quoted 83 lines · Jeff King

When the client has asked for certain shallow options like "deepen-since", we do a custom rev-list walk that pretends to be shallow. Before doing so, we have to disable the commit-graph, since it is not compatible with the shallow view of the repository. That's handled by 829a321569 (commit-graph: close_commit_graph before shallow walk, 2018-08-20). That commit literally closes and frees our repo->objects->commit_graph struct.

That creates an interesting problem for commits that have _already_ been parsed using the commit graph. Their commit->object.parsed flag is set, their commit->graph_pos is set, but their commit->maybe_tree may still be NULL. When somebody later calls repo_get_commit_tree(), we see that we haven't loaded the tree oid yet and try to get it from the commit graph. But since it has been freed, we segfault!

So the root of the issue is a data dependency between the commit's lazy-load of the tree oid and the fact that the commit graph can go away mid-process. How can we resolve it?

There are a couple of general approaches:

1. The obvious answer is to avoid loading the tree from the graph when we see that it's NULL. But then what do we return for the tree oid? If we return NULL, our caller in do_traverse() will rightly complain that we have no tree. We'd have to fallback to loading the actual commit object and re-parsing it. That requires teaching parse_commit_buffer() to understand re-parsing (i.e., not starting from a clean slate and not leaking any allocated bits like parent list pointers).

2. When we close the commit graph, walk through the set of in-memory objects and clear any graph_pos pointers. But this means we also have to "unparse" any such commits so that we know they still need to open the commit object to fill in their trees. So it's no less complicated than (1), and is more expensive (since we clear objects we might not later need).

3. Stop freeing the commit-graph struct. Continue to let it be used for lazy-loads of tree oids, but let upload-pack specify that it shouldn't be used for further commit parsing.

4. Push the whole shallow rev-list out to its own sub-process, with the commit-graph disabled from the start, giving it a clean memory space to work from.

I've chosen (3) here. Options (1) and (2) would work, but are non-trivial to implement. Option (4) is more expensive, and I'm not sure how complicated it is (shelling out for the actual rev-list part is easy, but we do then parse the resulting commits internally, and I'm not clear which parts need to be handling shallow-ness).

The new test in t5500 triggers this segfault, but see the comments there for how horribly intimate it has to be with how both upload-pack and commit graphs work.

---
 commit-graph.c        | 12 ++++++++++++
 commit-graph.h        |  6 ++++++
 repository.h          |  3 +++
 t/t5500-fetch-pack.sh | 38 ++++++++++++++++++++++++++++++++++++++
 upload-pack.c         |  2 +-
 5 files changed, 60 insertions(+), 1 deletion(-)
commit-graph.c +7 -0
index baeaf0d1bf..bbde647f8b 100644
@@ -468,6 +468,13 @@ static int prepare_commit_graph(struct repository *r)
 {
 	struct object_directory *odb;
 
+	/*
+	 * This must come before the "already attempted?" check below, because
+	 * we want to disable even an already-loaded graph file.
+	 */
+	if (r->commit_graph_disabled)
+		return 0;
+
 	if (r->objects->commit_graph_attempted)
 		return !!r->objects->commit_graph;
 	r->objects->commit_graph_attempted = 1;

commit-graph.c void free_commit_graph(struct commit_graph *g) · line 2115

Quoted 8 lines
commit-graph.c +5 -0
@@ -2101,3 +2108,8 @@ void free_commit_graph(struct commit_graph *g)
 	free(g->filename);
 	free(g);
 }
+
+void disable_commit_graph(struct repository *r)
+{
+	r->commit_graph_disabled = 1;
+}

Thanks for the additional comment and using the repo struct. LGTM.

-Stolee

#24 lore raw

Re: [PATCH v2] upload-pack commit graph segfault fix

In reply to Jeff King
Quoted 2 lines · Jeff King

Here's a re-roll of my "disable commit graph more gently" fix. Versus v1:

Thanks for sending a re-roll. I deployed this out to all of our servers running git at GitHub, and it seems to be working fine.

For what it's worth, I don't have 'core.commitGraph' enabled on any of those servers for now, but I'll turn it back on shortly.

Quoted 14 lines

- I've included a preparatory patch that speeds up prepare_commit_graph(). It's not strictly related, but there's a textual dependency. It could be easily spun off to its own series.

- a comment points out that the ordering of the "disable" check is important

- disable_commit_graph() now works on a repository struct

- typo fixes in the test comments

[1/2]: commit-graph: bump DIE_ON_LOAD check to actual load-time [2/2]: upload-pack: disable commit graph more gently for shallow traversal

Quoted 8 lines
 commit-graph.c        | 18 +++++++++++++++---
 commit-graph.h        |  6 ++++++
 repository.h          |  3 +++
 t/t5500-fetch-pack.sh | 38 ++++++++++++++++++++++++++++++++++++++
 upload-pack.c         |  2 +-
 5 files changed, 63 insertions(+), 4 deletions(-)

-Peff

Thanks,
Taylor