Re: [PATCH 3/7] commit-graph: start parsing generation v2 (again)

[Derrick Stolee <derrickstolee@github.com>]

On 3/3/2022 6:19 AM, Patrick Steinhardt wrote:
> On Wed, Mar 02, 2022 at 09:57:17AM -0500, Derrick Stolee wrote:
>> On 3/2/2022 8:57 AM, Patrick Steinhardt wrote:
>>> On Tue, Mar 01, 2022 at 10:25:46AM -0500, Derrick Stolee wrote:
>>>> On 3/1/2022 9:53 AM, Patrick Steinhardt wrote:
>>
>>>>> Hum. I have re-verified, and this indeed seems to play out. So I must've
>>>>> accidentally ran all my testing with the state generated without the
>>>>> final patch which fixes the corruption. I do see lots of the following
>>>>> warnings, but overall I can verify and write the commit-graph just fine:
>>>>>
>>>>>     commit-graph generation for commit c80a42de8803e2d77818d0c82f88e748d7f9425f is 1623362063 < 1623362139
>>>>
>>>> But I'm not able to generate these warnings from either version. I
>>>> tried generating different levels of a split commit-graph, but
>>>> could not reproduce it. If you have reproduction steps using current
>>>> 'master' (or any released Git version) and the four patches here,
>>>> then I would love to get a full understanding of your errors.
>>>>
>>>> Thanks,
>>>> -Stolee
>>>
>>> I haven't yet been able to reproduce it with publicly available data,
>>> but with the internal references I'm able to evoke the warnings
>>> reliably. It only works when I have two repositories connected via
>>> alternates, when generating the commit-graph in the linked-to repo
>>> first, and then generating the commit-graph in the linking repo.
>>>
>>> The following recipe allows me to reproduce, but rely on private data:
>>>
>>>     $ git --version
>>>     git version 2.35.1
>>>
>>>     # The pool repository is the one we're linked to from the fork.
>>>     $ cd "$pool"
>>>     $ rm -rf objects/info/commit-graph objects/info/commit-graph
>>>     $ git commit-graph write --split
>>>
>>>     $ cd "$fork"
>>>     $ rm -rf objects/info/commit-graph objects/info/commit-graph
>>>     $ git commit-graph write --split
>>>
>>>     $ git commit-graph verify --no-progress
>>>     $ echo $?
>>>     0
>>>
>>>     # This is 715d08a9e51251ad8290b181b6ac3b9e1f9719d7 with your full v2
>>>     # applied on top.
>>>     $ ~/Development/git/bin-wrappers/git --version
>>>     git version 2.35.1.358.g7ede1bea24
>>>
>>>     $ ~/Development/git/bin-wrappers/git commit-graph verify --no-progress
>>>     commit-graph generation for commit 06a91bac00ed11128becd48d5ae77eacd8f24c97 is 1623273624 < 1623273710
>>>     commit-graph generation for commit 0ae91029f27238e8f8e109c6bb3907f864dda14f is 1622151146 < 1622151220
>>>     commit-graph generation for commit 0d4582a33d8c8e3eb01adbf564f5e1deeb3b56a2 is 1631045222 < 1631045225
>>>     commit-graph generation for commit 0daf8976439d7e0bb9710c5ee63b570580e0dc03 is 1620347739 < 1620347789
>>>     commit-graph generation for commit 0e0ee8ffb3fa22cee7d28e21cbd6df26454932cf is 1623783297 < 1623783380
>>>     commit-graph generation for commit 0f08ab3de6ec115ea8a956a1996cb9759e640e74 is 1621543278 < 1621543339
>>>     commit-graph generation for commit 133ed0319b5a66ae0c2be76e5a887b880452b111 is 1620949864 < 1620949915
>>>     commit-graph generation for commit 1341b3e6c63343ae94a8a473fa057126ddd4669a is 1637344364 < 1637344384
>>>     commit-graph generation for commit 15bdfc501c2c9f23e9353bf6e6a5facd9c32a07a is 1623348103 < 1623348133
>>>     ...
>>>     $ echo $?
>>>     1
>>>
>>> When generating commit-graphs with your patches applied the `verify`
>>> step works alright.
>>>
>>> I've also by accident stumbled over the original error again:
>>>
>>>     fatal: commit-graph requires overflow generation data but has none
>>>
>>> This time it's definitely not caused by generating commit-graphs with an
>>> in-between state of your patch series because the data comes straight
>>> from production with no changes to the commit-graphs performed by
>>> myself. There we're running Git v2.33.1 with a couple of backported
>>> patches (see [1]). While those patches cause us to make more use of the
>>> commit-graph, none modify the way we generate them.
>>>
>>> Of note is that the commit-graph contains references to commits which
>>> don't exist in the ODB anymore.
>>>
>>> Patrick
>>>
>>> [1]: https://gitlab.com/gitlab-org/gitlab-git/-/commits/pks-v2.33.1.gl3
>>
>> Thank you for your diligence here, Patrick. I really appreciate the
>> work you're putting in to verify the situation.
>>
>> Since our repro relies on private information, but is consistent, I
>> wonder if we should take the patch below, which starts to ignore the
>> older generation number v2 data and only writes freshly-computed
>> numbers.
>>
>> Thanks,
>> -Stolee
> 
> Thanks. With your patch below the `fatal:` error is gone, but I'm still
> seeing the same errors with regards to the commit-graph generations.

This is disappointing and unexpected. Thanks for verifying.

> So to summarize my findings:
> 
>     - This bug occurs when writing commit-graphs with v2.35.1, but
>       reading them with your patches.
> 
>     - This bug occurs when I have two repositories connected via an
>       alternates file. I haven't yet been able to reproduce it in a
>       single repository that is not connected to a separate ODB.

This is an interesting distinction. One that I didn't think would
matter, but I'll look into the code to see how that could affect
things.

>     - This bug only occurs when I first generate the commit-graph in the
>       repository I'm borrowing objects from.
> 
>     - This bug only occurs when I write commit-graphs with `--split` in
>       both repositories. "Normal" commit-graphs don't have this issue,
>       and neither can I see it with `--split=replace` or mixed-type
>       commit-graphs.
> 
> Beware, the following explanation is based on my very basic
> understanding of the commit-graph code and thus more likely to be wrong
> than right:
> 
> With the old Git version, we've been mis-parsing the generation because
> `read_generation_data` wasn't ever set. As a result it can happen that
> the second split commit-graph we're generating computes its own
> generation numbers from the wrong starting point because it uses the
> mis-parsed generation numbers from the parent commit-graph.
> 
> With your patches, we start to correctly account for overflows and would
> thus end up with a different value for the generation depending on where
> we parse the commit from: if we parse it from the first commit-graph it
> would be correct because it's contains the "root" of the generation
> numbers. But if we parse a commit from the second commit-graph we may
> have a mismatch because the generation numbers in there may have been
> derived from generation numbers mis-parsed from the first commit-graph.
> And because these would be wrong in case there was an overflow it is
> clear that the new corrected generation number may be wrong, as well.

Hm. My expectation was that the older layers of the split commit-graph
would have read_generation_data disabled (because the new Git version
cannot read the GDAT chunk) and then the validate_mixed_generation_chain()
method would remove read_generation_data from all of the graphs in the
list.

Combining this with your thoughts on cross-alternate split commit-graphs,
this makes me think we should try this:

--- >8 ---

diff --git a/commit-graph.c b/commit-graph.c
index fb2ced0bd6..74c6534f56 100644
--- a/commit-graph.c
+++ b/commit-graph.c
@@ -609,8 +609,6 @@ struct commit_graph *read_commit_graph_one(struct repository *r,
 	if (!g)
 		g = load_commit_graph_chain(r, odb);
 
-	validate_mixed_generation_chain(g);
-
 	return g;
 }
 
@@ -668,7 +666,13 @@ static int prepare_commit_graph(struct repository *r)
 	     !r->objects->commit_graph && odb;
 	     odb = odb->next)
 		prepare_commit_graph_one(r, odb);
-	return !!r->objects->commit_graph;
+
+	if (r->objects->commit_graph) {
+		validate_mixed_generation_chain(r->objects->commit_graph);
+		return 1;
+	}
+
+	return 0;
 }
 
 int generation_numbers_enabled(struct repository *r)


--- >8 ---

Notice that I'm moving the validate_mixed_generation_chain() call
out of read_commit_graph_one() and into prepare_commit_graph(). To
my understanding, this _should_ have an equivalent end state as the
old code, but might be worth trying just as a quick check.

I will continue investigating and try to reproduce with this
additional constraint of working across an alternate.

Thanks,
-Stolee