[v3,1/4] bundle-uri: verify oid before writing refs

Commit Message

Xing Xin May 27, 2024, 3:41 p.m. UTC

From: Xing Xin <[e-mail]>

When using the bundle-uri mechanism with a bundle list containing
multiple interrelated bundles, we encountered a bug where tips from
downloaded bundles were not discovered, thus resulting in rather slow
clones. This was particularly problematic when employing the heuristic
`creationTokens`.

And this is easy to reproduce. Suppose we have a repository with a
single branch `main` pointing to commit `A`, firstly we create a base
bundle with

  git bundle create base.bundle main

Then let's add a new commit `B` on top of `A`, so that an incremental
bundle for `main` can be created with

  git bundle create incr.bundle A..main

Now we can generate a bundle list with the following content:

  [bundle]
      version = 1
      mode = all
      heuristic = creationToken

  [bundle "base"]
      uri = base.bundle
      creationToken = 1

  [bundle "incr"]
      uri = incr.bundle
      creationToken = 2

A fresh clone with the bundle list above would give the expected
`refs/bundles/main` pointing at `B` in new repository, in other words we
already had everything locally from the bundles, but git would still
download everything from server as if we got nothing.

So why the `refs/bundles/main` is not discovered? After some digging I
found that:

1. Bundles in bundle list are downloaded to local files via
   `download_bundle_list` or via `fetch_bundles_by_token` for the
   creationToken heuristic case.
2. Then it tries to unbundle each bundle via `unbundle_from_file`, which
   is called by `unbundle_all_bundles` or called within
   `fetch_bundles_by_token` for the creationToken heuristic case.
3. Here, we first read the bundle header to get all the prerequisites
   for the bundle, this is done in `read_bundle_header`.
4. Then we call `unbundle`, which calls `verify_bundle` to ensure that
   the repository does indeed contain the prerequisites mentioned in the
   bundle.
5. The `verify_bundle` will call `parse_object`, within which the
   `prepare_packed_git` or `reprepare_packed_git` is eventually called,
   which means that the `raw_object_store->packed_git` data gets filled
   in and ``packed_git_initialized` is set. This also means consecutive
   calls to `prepare_packed_git` doesn't re-initiate
   `raw_object_store->packed_git` since `packed_git_initialized` already
   is set.
6. If `unbundle` succeeds, it writes some refs via `refs_update_ref`
   with `REF_SKIP_OID_VERIFICATION` set. So the bundle refs which can
   target arbitrary objects are written to the repository.
7. Finally in `do_fetch_pack_v2`, `mark_complete_and_common_ref` and
   `mark_tips` are called with `OBJECT_INFO_QUICK` set to find local
   tips. Here it won't call `reprepare_packed_git` anymore so it would
   fail to parse oids that only reside in the last bundle.

Back to the example above, when unbunding `incr.bundle`, `base.pack` is
enlisted to `packed_git` bacause of the prerequisites to verify. While
we can not find `B` for negotiation at a latter time because `B` exists
in `incr.pack` which is not enlisted in `packed_git`.

This commit fixes this bug by dropping the `REF_SKIP_OID_VERIFICATION`
flag when writing bundle refs, so we can:

1. Ensure that the bundle refs we are writing are pointing to valid
   objects.
2. Ensure all the tips from bundle refs can be correctly parsed.

And a set of negotiation related tests for bundle-uri are added.

Signed-off-by: Xing Xin <[e-mail]>
---
 bundle-uri.c                |   3 +-
 t/t5558-clone-bundle-uri.sh | 129 ++++++++++++++++++++++++++++++++++++
 2 files changed, 130 insertions(+), 2 deletions(-)

Comments

Patrick Steinhardt May 28, 2024, 11:55 a.m. UTC | #1

On Mon, May 27, 2024 at 03:41:54PM +0000, Xing Xin via GitGitGadget wrote:
> From: Xing Xin <[e-mail]>
[snip]
> 5. The `verify_bundle` will call `parse_object`, within which the
>    `prepare_packed_git` or `reprepare_packed_git` is eventually called,
>    which means that the `raw_object_store->packed_git` data gets filled
>    in and ``packed_git_initialized` is set. This also means consecutive

s/``/`/

[snip]
> This commit fixes this bug by dropping the `REF_SKIP_OID_VERIFICATION`
> flag when writing bundle refs, so we can:
> 
> 1. Ensure that the bundle refs we are writing are pointing to valid
>    objects.
> 2. Ensure all the tips from bundle refs can be correctly parsed.

I think one angle that your explanation doesn't cover is why exactly
dropping the flag fixes the observed issue.

> And a set of negotiation related tests for bundle-uri are added.

s/And/Add/

[snip]
> +#########################################################################
> +# Clone negotiation related tests begin here
> +
> +test_expect_success 'negotiation: bundle with part of wanted commits' '
> +	test_when_finished rm -rf trace*.txt &&
> +	GIT_TRACE_PACKET="$(pwd)/trace-packet.txt" \
> +	git clone --no-local --bundle-uri="clone-from/A.bundle" \
> +		clone-from nego-bundle-part &&
> +	git -C nego-bundle-part for-each-ref --format="%(refname)" >refs &&
> +	grep "refs/bundles/" refs >actual &&
> +	cat >expect <<-\EOF &&
> +	refs/bundles/topic
> +	EOF
> +	test_cmp expect actual &&
> +	# Ensure that refs/bundles/topic are sent as "have".
> +	grep "clone> have $(git -C clone-from rev-parse A)" trace-packet.txt
> +'

As far as I can see there is no test that verifies the case where the
bundle contains refs, but misses the objects to satisfy the refs. Can we
craft such a bundle and exercise this new failure mode?

Patrick

Xing Xin May 30, 2024, 8:32 a.m. UTC | #2

At 2024-05-28 19:55:15, "Patrick Steinhardt" <[e-mail]> wrote:
[snip]
>>    which means that the `raw_object_store->packed_git` data gets filled
>>    in and ``packed_git_initialized` is set. This also means consecutive
>
>s/``/`/

Copy that.

[snip]
>
>> And a set of negotiation related tests for bundle-uri are added.
>
>s/And/Add/

The "And" is correct here, though I changed the commit message in the new
version. :)


[snip]
>As far as I can see there is no test that verifies the case where the
>bundle contains refs, but misses the objects to satisfy the refs. Can we
>craft such a bundle and exercise this new failure mode?

A new test is added in [PATCH v4 1/4], which shows it correctly refuses to
target a bundle reference to a non-existent object.

Xing Xin

Patch

diff --git a/bundle-uri.c b/bundle-uri.c
index 91b3319a5c1..65666a11d9c 100644
--- a/bundle-uri.c
+++ b/bundle-uri.c
@@ -400,8 +400,7 @@  static int unbundle_from_file(struct repository *r, const char *file)
 		refs_update_ref(get_main_ref_store(the_repository),
 				"fetched bundle", bundle_ref.buf, oid,
 				has_old ? &old_oid : NULL,
-				REF_SKIP_OID_VERIFICATION,
-				UPDATE_REFS_MSG_ON_ERR);
+				0, UPDATE_REFS_MSG_ON_ERR);
 	}
 
 	bundle_header_release(&header);
diff --git a/t/t5558-clone-bundle-uri.sh b/t/t5558-clone-bundle-uri.sh
index 1ca5f745e73..a5b04d6f187 100755
--- a/t/t5558-clone-bundle-uri.sh
+++ b/t/t5558-clone-bundle-uri.sh
@@ -20,7 +20,10 @@  test_expect_success 'fail to clone from non-bundle file' '
 test_expect_success 'create bundle' '
 	git init clone-from &&
 	git -C clone-from checkout -b topic &&
+
 	test_commit -C clone-from A &&
+	git -C clone-from bundle create A.bundle topic &&
+
 	test_commit -C clone-from B &&
 	git -C clone-from bundle create B.bundle topic
 '
@@ -259,6 +262,132 @@  test_expect_success 'clone bundle list (file, any mode, all failures)' '
 	! grep "refs/bundles/" refs
 '
 
+#########################################################################
+# Clone negotiation related tests begin here
+
+test_expect_success 'negotiation: bundle with part of wanted commits' '
+	test_when_finished rm -rf trace*.txt &&
+	GIT_TRACE_PACKET="$(pwd)/trace-packet.txt" \
+	git clone --no-local --bundle-uri="clone-from/A.bundle" \
+		clone-from nego-bundle-part &&
+	git -C nego-bundle-part for-each-ref --format="%(refname)" >refs &&
+	grep "refs/bundles/" refs >actual &&
+	cat >expect <<-\EOF &&
+	refs/bundles/topic
+	EOF
+	test_cmp expect actual &&
+	# Ensure that refs/bundles/topic are sent as "have".
+	grep "clone> have $(git -C clone-from rev-parse A)" trace-packet.txt
+'
+
+test_expect_success 'negotiation: bundle with all wanted commits' '
+	test_when_finished rm -rf trace*.txt &&
+	GIT_TRACE_PACKET="$(pwd)/trace-packet.txt" \
+	git clone --no-local --single-branch --branch=topic --no-tags \
+		--bundle-uri="clone-from/B.bundle" \
+		clone-from nego-bundle-all &&
+	git -C nego-bundle-all for-each-ref --format="%(refname)" >refs &&
+	grep "refs/bundles/" refs >actual &&
+	cat >expect <<-\EOF &&
+	refs/bundles/topic
+	EOF
+	test_cmp expect actual &&
+	# We already have all needed commits so no "want" needed.
+	! grep "clone> want " trace-packet.txt
+'
+
+test_expect_success 'negotiation: bundle list (no heuristic)' '
+	test_when_finished rm -f trace*.txt &&
+	cat >bundle-list <<-EOF &&
+	[bundle]
+		version = 1
+		mode = all
+
+	[bundle "bundle-1"]
+		uri = file://$(pwd)/clone-from/bundle-1.bundle
+
+	[bundle "bundle-2"]
+		uri = file://$(pwd)/clone-from/bundle-2.bundle
+	EOF
+
+	GIT_TRACE_PACKET="$(pwd)/trace-packet.txt" \
+	git clone --no-local --bundle-uri="file://$(pwd)/bundle-list" \
+		clone-from nego-bundle-list-no-heuristic &&
+
+	git -C nego-bundle-list-no-heuristic for-each-ref --format="%(refname)" >refs &&
+	grep "refs/bundles/" refs >actual &&
+	cat >expect <<-\EOF &&
+	refs/bundles/base
+	refs/bundles/left
+	EOF
+	test_cmp expect actual &&
+	grep "clone> have $(git -C nego-bundle-list-no-heuristic rev-parse refs/bundles/left)" trace-packet.txt
+'
+
+test_expect_success 'negotiation: bundle list (creationToken)' '
+	test_when_finished rm -f trace*.txt &&
+	cat >bundle-list <<-EOF &&
+	[bundle]
+		version = 1
+		mode = all
+		heuristic = creationToken
+
+	[bundle "bundle-1"]
+		uri = file://$(pwd)/clone-from/bundle-1.bundle
+		creationToken = 1
+
+	[bundle "bundle-2"]
+		uri = file://$(pwd)/clone-from/bundle-2.bundle
+		creationToken = 2
+	EOF
+
+	GIT_TRACE_PACKET="$(pwd)/trace-packet.txt" \
+	git clone --no-local --bundle-uri="file://$(pwd)/bundle-list" \
+		clone-from nego-bundle-list-heuristic &&
+
+	git -C nego-bundle-list-heuristic for-each-ref --format="%(refname)" >refs &&
+	grep "refs/bundles/" refs >actual &&
+	cat >expect <<-\EOF &&
+	refs/bundles/base
+	refs/bundles/left
+	EOF
+	test_cmp expect actual &&
+	grep "clone> have $(git -C nego-bundle-list-heuristic rev-parse refs/bundles/left)" trace-packet.txt
+'
+
+test_expect_success 'negotiation: bundle list with all wanted commits' '
+	test_when_finished rm -f trace*.txt &&
+	cat >bundle-list <<-EOF &&
+	[bundle]
+		version = 1
+		mode = all
+		heuristic = creationToken
+
+	[bundle "bundle-1"]
+		uri = file://$(pwd)/clone-from/bundle-1.bundle
+		creationToken = 1
+
+	[bundle "bundle-2"]
+		uri = file://$(pwd)/clone-from/bundle-2.bundle
+		creationToken = 2
+	EOF
+
+	GIT_TRACE_PACKET="$(pwd)/trace-packet.txt" \
+	git clone --no-local --single-branch --branch=left --no-tags \
+		--bundle-uri="file://$(pwd)/bundle-list" \
+		clone-from nego-bundle-list-all &&
+
+	git -C nego-bundle-list-all for-each-ref --format="%(refname)" >refs &&
+	grep "refs/bundles/" refs >actual &&
+	cat >expect <<-\EOF &&
+	refs/bundles/base
+	refs/bundles/left
+	EOF
+	test_cmp expect actual &&
+	# We already have all needed commits so no "want" needed.
+	! grep "clone> want " trace-packet.txt
+'
+
 #########################################################################
 # HTTP tests begin here