commit-graph: ignore duplicates when merging layers

Commit Message

Philippe Blain via GitGitGadget Oct. 8, 2020, 1:56 p.m. UTC

From: Derrick Stolee <[e-mail]>

Thomas reported [1] that a "git fetch" command was failing with an error
saying "unexpected duplicate commit id". The root cause is that they had
fetch.writeCommitGraph enabled which generates commit-graph chains, and
this instance was merging two layers that both contained the same commit
ID.

[1] https://lore.kernel.org/git/[e-mail]/

The initial assumption is that Git would not write a commit ID into a
commit-graph layer if it already exists in a lower commit-graph layer.
Somehow, this specific case did get into that situation, leading to this
error.

While unexpected, this isn't actually invalid (as long as the two layers
agree on the metadata for the commit). When we parse a commit that does
not have a graph_pos in the commit_graph_data_slab, we use binary search
in the commit-graph layers to find the commit and set graph_pos. That
position is never used again in this case. However, when we parse a
commit from the commit-graph file, we load its parents from the
commit-graph and assign graph_pos at that point. If those parents were
already parsed from the commit-graph, then nothing needs to be done.
Otherwise, this graph_pos is a valid position in the commit-graph so we
can parse the parents, when necessary.

Thus, this die() is too aggignoring the duplicates.

This leads to some additional complication that we did no have before:
if we only ignore the duplicates, then we will produce a commit-graph
that has identical commit IDs listed in adjacent positions. This excess
data will never be removed from the commit-graph, which could cascade
into significantly bloated file sizes.

Begrudgingly, the best way to fix this is to copy the commit pointers
into a new list that only contains de-duplicated commit IDs. This adds a
slight memory overhead, but it is small compared to having all of these
commits parsed in memory, so it should be an acceptable cost for
avoiding these failures.

Since the root cause for producing commit-graph layers with these
duplicate commits is currently unknown, it is difficult to create a test
for this scenario. For now, we must rely on testing the example data
graciously provided in [1]. My local test successfully merged layers,
and 'git commit-graph verify' passed.

Reported-by: Thomas Braun <[e-mail]>
Co-authored-by: Jeff King <[e-mail]>
Signed-off-by: Derrick Stolee <[e-mail]>
---
    commit-graph: ignore duplicates when merging layers
    
    This wasn't quite as simple as what Peff had posted, since we really
    don't want to keep duplicate commits around in the new merged layer.
    
    I still don't have a grasp on how this happened in the first place, but
    will keep looking.
    
    Thanks, -Stolee

Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-747%2Fderrickstolee%2Fcommit-graph-dup-commits-v1
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-747/derrickstolee/commit-graph-dup-commits-v1
Pull-Request: https://github.com/gitgitgadget/git/pull/747

 commit-graph.c | 20 ++++++++++++++++++--
 1 file changed, 18 insertions(+), 2 deletions(-)


base-commit: d98273ba77e1ab9ec755576bc86c716a97bf59d7

Comments

Taylor Blau Oct. 8, 2020, 2:15 p.m. UTC | #1

On Thu, Oct 08, 2020 at 01:56:52PM +0000, Derrick Stolee via GitGitGadget wrote:
> From: Derrick Stolee <[e-mail]>
>
> Thomas reported [1] that a "git fetch" command was failing with an error
> saying "unexpected duplicate commit id". The root cause is that they had
> fetch.writeCommitGraph enabled which generates commit-graph chains, and
> this instance was merging two layers that both contained the same commit
> ID.
>
> [1] https://lore.kernel.org/git/[e-mail]/
>
> The initial assumption is that Git would not write a commit ID into a
> commit-graph layer if it already exists in a lower commit-graph layer.
> Somehow, this specific case did get into that situation, leading to this
> error.
>
> While unexpected, this isn't actually invalid (as long as the two layers
> agree on the metadata for the commit). When we parse a commit that does
> not have a graph_pos in the commit_graph_data_slab, we use binary search
> in the commit-graph layers to find the commit and set graph_pos. That
> position is never used again in this case. However, when we parse a
> commit from the commit-graph file, we load its parents from the
> commit-graph and assign graph_pos at that point. If those parents were
> already parsed from the commit-graph, then nothing needs to be done.
> Otherwise, this graph_pos is a valid position in the commit-graph so we
> can parse the parents, when necessary.
>
> Thus, this die() is too aggignoring the duplicates.

s/aggignoring/aggressively ignoring ?

>
> This leads to some additional complication that we did no have before:

s/no/not, but I am more wondering about what "This" is. I think what
you're saying is: "Suppose we didn't die on duplicates, what would
happen? Well, there'd be some additional problems, but here's a way that
we can fix them (storing the de-duplicated OIDs separately)".

> if we only ignore the duplicates, then we will produce a commit-graph
> that has identical commit IDs listed in adjacent positions. This excess
> data will never be removed from the commit-graph, which could cascade
> into significantly bloated file sizes.
>
> Begrudgingly, the best way to fix this is to copy the commit pointers
> into a new list that only contains de-duplicated commit IDs. This adds a
> slight memory overhead, but it is small compared to having all of these
> commits parsed in memory, so it should be an acceptable cost for
> avoiding these failures.
>
> Since the root cause for producing commit-graph layers with these
> duplicate commits is currently unknown, it is difficult to create a test
> for this scenario. For now, we must rely on testing the example data
> graciously provided in [1]. My local test successfully merged layers,
> and 'git commit-graph verify' passed.
>
> Reported-by: Thomas Braun <[e-mail]>
> Co-authored-by: Jeff King <[e-mail]>
> Signed-off-by: Derrick Stolee <[e-mail]>
> ---
>     commit-graph: ignore duplicates when merging layers
>
>     This wasn't quite as simple as what Peff had posted, since we really
>     don't want to keep duplicate commits around in the new merged layer.
>
>     I still don't have a grasp on how this happened in the first place, but
>     will keep looking.

I'm looking as well, but I haven't found any smoking guns yet. I could
imagine that this is a problem that existed before 0bd52e27e3
(commit-graph.h: store an odb in 'struct write_commit_graph_context',
2020-02-03), and simply couldn't be tickled because of how brittle
comparing ODB paths is. I could equally imagine that 0bd52e27e3 did
introduce this problem.
>
>     Thanks, -Stolee
>
> Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-747%2Fderrickstolee%2Fcommit-graph-dup-commits-v1
> Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-747/derrickstolee/commit-graph-dup-commits-v1
> Pull-Request: https://github.com/gitgitgadget/git/pull/747
>
>  commit-graph.c | 20 ++++++++++++++++++--
>  1 file changed, 18 insertions(+), 2 deletions(-)
>
> diff --git a/commit-graph.c b/commit-graph.c
> index cb042bdba8..29bac78dc3 100644
> --- a/commit-graph.c
> +++ b/commit-graph.c
> @@ -2009,6 +2009,7 @@ static int commit_compare(const void *_a, const void *_b)
>  static void sort_and_scan_merged_commits(struct write_commit_graph_context *ctx)
>  {
>  	uint32_t i;
> +	struct packed_commit_list deduped_commits = { NULL, 0, 0 };
>
>  	if (ctx->report_progress)
>  		ctx->progress = start_delayed_progress(
> @@ -2016,6 +2017,8 @@ static void sort_and_scan_merged_commits(struct write_commit_graph_context *ctx)
>  					ctx->commits.nr);
>
>  	QSORT(ctx->commits.list, ctx->commits.nr, commit_compare);
> +	deduped_commits.alloc = ctx->commits.nr;
> +	ALLOC_ARRAY(deduped_commits.list, deduped_commits.alloc);

I'm not sure that this deduped_commits list is actually necessary.

It would be nice for this caller if ctx->commits were a linked list
since it would make deleting duplicates easy, but I think that it would
be a burden for other callers. So, that's a dead end.

But what about marking the duplicate positions by NULL-ing them out, and
then taking another pass over the list to (1) compact it (i.e., push
everything down so that all of the NULLs occur at the end), and then (2)
truncate the length to the number of unique commits.

I could imagine that something like that is a little trickier, but it
seems worth it to avoid doubling the memory cost of this function.

>  	ctx->num_extra_edges = 0;
>  	for (i = 0; i < ctx->commits.nr; i++) {
> @@ -2023,17 +2026,30 @@ static void sort_and_scan_merged_commits(struct write_commit_graph_context *ctx)
>
>  		if (i && oideq(&ctx->commits.list[i - 1]->object.oid,
>  			  &ctx->commits.list[i]->object.oid)) {
> -			die(_("unexpected duplicate commit id %s"),
> -			    oid_to_hex(&ctx->commits.list[i]->object.oid));
> +			/*
> +			 * Silently ignore duplicates. These were likely
> +			 * created due to a commit appearing in multiple
> +			 * layers of the chain, which is unexpected but
> +			 * not invalid. We should make sure there is a
> +			 * unique copy in the new layer.
> +			 */
>  		} else {
>  			unsigned int num_parents;
>
> +			deduped_commits.list[deduped_commits.nr] = ctx->commits.list[i];
> +			deduped_commits.nr++;
> +
>  			num_parents = commit_list_count(ctx->commits.list[i]->parents);
>  			if (num_parents > 2)
>  				ctx->num_extra_edges += num_parents - 1;
>  		}
>  	}
>
> +	free(ctx->commits.list);
> +	ctx->commits.list = deduped_commits.list;
> +	ctx->commits.nr = deduped_commits.nr;
> +	ctx->commits.alloc = deduped_commits.alloc;
> +
>  	stop_progress(&ctx->progress);
>  }
>
>
> base-commit: d98273ba77e1ab9ec755576bc86c716a97bf59d7
> --
> gitgitgadget

Thanks,
Taylor

Derrick Stolee Oct. 8, 2020, 2:29 p.m. UTC | #2

On 10/8/2020 10:15 AM, Taylor Blau wrote:
> On Thu, Oct 08, 2020 at 01:56:52PM +0000, Derrick Stolee via GitGitGadget wrote:
>> Thus, this die() is too aggignoring the duplicates.
> 
> s/aggignoring/aggressively ignoring ?
> 
>>
>> This leads to some additional complication that we did no have before:
> 
> s/no/not, but I am more wondering about what "This" is. I think what
> you're saying is: "Suppose we didn't die on duplicates, what would
> happen? Well, there'd be some additional problems, but here's a way that
> we can fix them (storing the de-duplicated OIDs separately)".

Thanks. The message will be edited to fix these brain farts.

>>     I still don't have a grasp on how this happened in the first place, but
>>     will keep looking.
> 
> I'm looking as well, but I haven't found any smoking guns yet. I could
> imagine that this is a problem that existed before 0bd52e27e3
> (commit-graph.h: store an odb in 'struct write_commit_graph_context',
> 2020-02-03), and simply couldn't be tickled because of how brittle
> comparing ODB paths is. I could equally imagine that 0bd52e27e3 did
> introduce this problem.

Thanks.

>> +	ALLOC_ARRAY(deduped_commits.list, deduped_commits.alloc);
> 
> I'm not sure that this deduped_commits list is actually necessary.
> 
> It would be nice for this caller if ctx->commits were a linked list
> since it would make deleting duplicates easy, but I think that it would
> be a burden for other callers. So, that's a dead end.
> 
> But what about marking the duplicate positions by NULL-ing them out, and
> then taking another pass over the list to (1) compact it (i.e., push
> everything down so that all of the NULLs occur at the end), and then (2)
> truncate the length to the number of unique commits.
> 
> I could imagine that something like that is a little trickier, but it
> seems worth it to avoid doubling the memory cost of this function.

You are correct that we can just re-use the commits.list by "collapsing"
the list on top of duplicate entries. I'll send a new version that does
exactly that.

Thanks,
-Stolee

Patch

diff --git a/commit-graph.c b/commit-graph.c
index cb042bdba8..29bac78dc3 100644
--- a/commit-graph.c
+++ b/commit-graph.c
@@ -2009,6 +2009,7 @@  static int commit_compare(const void *_a, const void *_b)
 static void sort_and_scan_merged_commits(struct write_commit_graph_context *ctx)
 {
 	uint32_t i;
+	struct packed_commit_list deduped_commits = { NULL, 0, 0 };
 
 	if (ctx->report_progress)
 		ctx->progress = start_delayed_progress(
@@ -2016,6 +2017,8 @@  static void sort_and_scan_merged_commits(struct write_commit_graph_context *ctx)
 					ctx->commits.nr);
 
 	QSORT(ctx->commits.list, ctx->commits.nr, commit_compare);
+	deduped_commits.alloc = ctx->commits.nr;
+	ALLOC_ARRAY(deduped_commits.list, deduped_commits.alloc);
 
 	ctx->num_extra_edges = 0;
 	for (i = 0; i < ctx->commits.nr; i++) {
@@ -2023,17 +2026,30 @@  static void sort_and_scan_merged_commits(struct write_commit_graph_context *ctx)
 
 		if (i && oideq(&ctx->commits.list[i - 1]->object.oid,
 			  &ctx->commits.list[i]->object.oid)) {
-			die(_("unexpected duplicate commit id %s"),
-			    oid_to_hex(&ctx->commits.list[i]->object.oid));
+			/*
+			 * Silently ignore duplicates. These were likely
+			 * created due to a commit appearing in multiple
+			 * layers of the chain, which is unexpected but
+			 * not invalid. We should make sure there is a
+			 * unique copy in the new layer.
+			 */
 		} else {
 			unsigned int num_parents;
 
+			deduped_commits.list[deduped_commits.nr] = ctx->commits.list[i];
+			deduped_commits.nr++;
+
 			num_parents = commit_list_count(ctx->commits.list[i]->parents);
 			if (num_parents > 2)
 				ctx->num_extra_edges += num_parents - 1;
 		}
 	}
 
+	free(ctx->commits.list);
+	ctx->commits.list = deduped_commits.list;
+	ctx->commits.nr = deduped_commits.nr;
+	ctx->commits.alloc = deduped_commits.alloc;
+
 	stop_progress(&ctx->progress);
 }