git/list[1] front-page[2] threads[3] people[4] search[5] about
 

Re: [PATCH 01/12] commit-graph: simplify linking commit graph chains

From
Junio C Hamano <gitster@pobox.com>
Date
Nov 16, 2026, 20:53 UTC
Message-ID
<xmqqqzgkjy57.fsf@gitster.g>
In-Reply-To
<20261116-pks-odb-pluggable-commit-graphs-v1-1-a83fc1ff0808@pks.im>
Patrick Steinhardt <ps@pks.im> writes:
Show 37 quoted lines
> When we load a commit graph chain we verify for every layer that the
> hash of the base commit graph matches what the current tip has in its
> base graphs chunk. To do so we pass both the chain itself as well as the
> array of object IDs that we have read from the commit-graph chain file
> to `add_graph_to_chain()`.
>
> But the array of object IDs is redundant information: we have already
> loaded and linked the base layers, and as part of that we have stored
> their respective hashes in the structure itself. Consequently, we can
> already derive the hashes by walking the chain of base layers.
>
> Refactor the function to only take the new tip and base layers as input
> and derive all the other information ad-hoc and make the function
> public. This prepares for a subsequent commit, where we are about to
> move the logic that creates commit graph chains into the object database
> backend.
>
> Signed-off-by: Patrick Steinhardt <ps@pks.im>
> ---
>  commit-graph.c | 52 ++++++++++++++++++++++++----------------------------
>  commit-graph.h |  8 ++++++++
>  2 files changed, 32 insertions(+), 28 deletions(-)
>
> diff --git a/commit-graph.c b/commit-graph.c
> index da65cd3deb9..94c761907a0 100644
> --- a/commit-graph.c
> +++ b/commit-graph.c
> @@ -575,50 +575,48 @@ static void validate_mixed_bloom_settings(struct commit_graph *g)
>  	}
>  }
>  
> -static int add_graph_to_chain(struct commit_graph *g,
> -			      struct commit_graph *chain,
> -			      struct object_id *oids,
> -			      int n)
> +int commit_graph_link_base(struct commit_graph *g,
> +			   struct commit_graph *base)
Nice simplification, I guess.
Show 7 quoted lines
>  {
> -	struct commit_graph *cur_g = chain;
> +	struct commit_graph *layer;
> +	size_t n = 0;
> +
> +	for (layer = base; layer; layer = layer->base_graph)
> +		n++;

We chase the pointers to count the number of layers. We no longer count in signed int, which is probably a big plus in the sanitation department.

Show 5 quoted lines
>  	if (n && !g->chunk_base_graphs) {
>  		warning(_("commit-graph has no base graphs chunk"));
> -		return 0;
> +		return -1;
>  	}

It might be obvious when reading (and writing) this patch, but it is a welcome change to standardize on "0 means happy" convention. Should the fact that we now use more common error reporting convention be mentioned in the log message?

Show 16 quoted lines
>  	if (g->chunk_base_graphs_size / g->hash_algo->rawsz < n) {
>  		warning(_("commit-graph base graphs chunk is too small"));
> -		return 0;
> +		return -1;
>  	}
>  
> -	while (n) {
> +	for (layer = base; layer; layer = layer->base_graph) {
>  		n--;
> -
> -		if (!cur_g ||
> -		    !oideq(&oids[n], &cur_g->oid) ||
> -		    !hasheq(oids[n].hash, g->chunk_base_graphs + st_mult(g->hash_algo->rawsz, n),
> +		if (!hasheq(layer->oid.hash,
> +			    g->chunk_base_graphs + st_mult(g->hash_algo->rawsz, n),
>  			    g->hash_algo)) {

We no longer check cur_g->oid is what we expect, because we lost the array that holds the value to compare. Would this make cur_g->oid a write-only member that is only computed and written but never used other than reporting purposes? In commit-graph.c before this step, I think this is the only place that reads ->oid for purposes other than passing it to diagnostics like die/warn.

We seem to still be making sure in load_commit_graph_chain_fd_st() that the line we read into local oid there is a hash, so perhaps it is possible to validate g->oid against this oid after we load g?

Show 9 quoted lines
> @@ -682,7 +679,7 @@ struct commit_graph *load_commit_graph_chain_fd_st(struct object_database *odb,
>  			free(graph_name);
>  
>  			if (g) {
> -				if (add_graph_to_chain(g, graph_chain, oids, i)) {
> +				if (!commit_graph_link_base(g, graph_chain)) {
>  					graph_chain = g;
>  					valid = 1;
>  				} else {
Show 10 quoted lines
> +
> +/*
> + * Link the commit graph layer `g` on top of the chain `base`. `base` may be
> + * NULL if `g` is the root of a new chain. Returns 0 on success, a negative
> + * error code otherwise.
> + */
> +int commit_graph_link_base(struct commit_graph *g,
> +			   struct commit_graph *base);
>  struct commit_graph *read_commit_graph_one(struct repository *repo,
>  					   const char *dir);
Previous: Patrick SteinhardtNext: Patrick Steinhardt
Message 3 of 14 in “Introduce pluggable commit graphs”
  1. 00/12 Introduce pluggable commit graphsPatrick Steinhardt, Nov 16, 2026
  2. 01/12 commit-graph: simplify linking commit graph chainsPatrick Steinhardt, Nov 16, 2026
  3. Junio C HamanoNov 16, 2026
  4. 02/12 commit-graph: simplify reading a single commit graphPatrick Steinhardt, Nov 16, 2026
  5. 03/12 commit-graph: expose function to validate a commit graph chainPatrick Steinhardt, Nov 16, 2026
  6. 04/12 commit-graph: decide mergeable and chainable layers by depthPatrick Steinhardt, Nov 16, 2026
  7. 05/12 commit-graph: introduce common exit for `write_commit_graph_file()`Patrick Steinhardt, Nov 16, 2026
  8. 06/12 commit-graph: derive layer infos from the passed-in chainPatrick Steinhardt, Nov 16, 2026
  9. 07/12 commit-graph: split out check whether writing commit graphs is allowedPatrick Steinhardt, Nov 16, 2026
  10. 08/12 commit-graph: split out logic to write a new commit graph layerPatrick Steinhardt, Nov 16, 2026
  11. 09/12 commit-graph: mark incomplete commit graph chainsPatrick Steinhardt, Nov 16, 2026
  12. 10/12 odb/source: make commit graphs pluggablePatrick Steinhardt, Nov 16, 2026
  13. 11/12 odb/source-files: move logic to derive commit graph pathsPatrick Steinhardt, Nov 16, 2026
  14. 12/12 odb/source-files: internalize logic to derive commit graph pathsPatrick Steinhardt, Nov 16, 2026

Read the whole thread, see it on lore, or plain text.

$ cat FOOTERMessages come from the public archive at lore.kernel.org/git, fetched every hour. The front page is chosen and written each morning by an AI editor and can be wrong; the threads themselves are the record. About and API. For agents: an MCP server at https://gitlist.dev/mcp, and any thread, story or person page as Markdown by adding .md to its URL (or sending Accept: text/markdown). Details in /llms.txt.