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

[PATCH 4/4] repack: retain cruft packs in MIDXs containing kept packs

From
Taylor Blau <ttaylorr@openai.com>
Date
Sep 30, 2026, 01:28 UTC
Message-ID
<e942c256334e4de31ec0a1cb2d5f8c7465d8696f.1790731662.git.me@ttaylorr.com>
In-Reply-To
<cover.1790731662.git.me@ttaylorr.com>

When performing a geometric repack with 'repack.midxMustContainCruft' set to "false", Git uses '--stdin-packs=follow' to copy (once-cruft) objects needed for reachability closure out of cruft packs. .keep packs do not need to participate in that walk, though they *are* included in the resulting MIDX.

A .keep pack can contain a commit that reaches an object whose only copy is in a cruft pack. When there is no previous MIDX and the repack writes a new pack, neither `midx_has_unknown_packs()` nor the `!names.nr` fallback require that cruft pack to be included. If the kept commit (or a descendant of it) is selected for bitmap coverage, the bitmap writer fails because the MIDX does not contain all of its reachable objects.

Include cruft packs whenever the MIDX contains kept packs. This also retains cruft when the kept packs happen to have full closure, or when '--pack-kept-objects' lets the repack walk them. It avoids having to establish their closure before deciding which packs the MIDX needs.

Add a test that packs the tip commit and its tree into a kept pack, leaving its parent in the cruft pack. The new commit's blob remains loose, making the geometric repack write a new pack and bypass the no-new-packs fallback. Verify that the repack succeeds and that we are able to successfully write a bitmap.

Signed-off-by: Taylor Blau <ttaylorr@openai.com>
---
 repack-midx.c           |  5 +++++
 t/t7704-repack-cruft.sh | 20 ++++++++++++++++++++
 2 files changed, 25 insertions(+)
diff --git a/repack-midx.c b/repack-midx.c
index 64c7f8d0f42..622c3c9d236 100644
--- a/repack-midx.c
+++ b/repack-midx.c
@@ -197,6 +197,7 @@ static void midx_included_packs(struct string_list *include,
 	}
 
 	if (opts->midx_must_contain_cruft ||
+	    existing->kept_packs.nr ||
 	    midx_has_unknown_packs(include, geometry, existing)) {
 		/*
 		 * If there are one or more unknown pack(s) present (see
@@ -209,6 +210,10 @@ static void midx_included_packs(struct string_list *include,
 		 * reachability closure if the MIDX is bitmapped and one
 		 * or more of the bitmap's selected commits reaches a
 		 * once-cruft object that was later made reachable.
+		 *
+		 * Kept packs may also depend on cruft objects, since
+		 * they are included above without necessarily being
+		 * traversed by the repack.
 		 */
 		for_each_string_list_item(item, &existing->cruft_packs) {
 			/*
diff --git a/t/t7704-repack-cruft.sh b/t/t7704-repack-cruft.sh
index f7f83e70ffe..02db2a06d9e 100755
--- a/t/t7704-repack-cruft.sh
+++ b/t/t7704-repack-cruft.sh
@@ -798,6 +798,26 @@ test_expect_success 'incremental repack includes cruft for MIDX bitmaps' '
 	)
 '
 
+test_expect_success 'geometric repack includes cruft for kept packs' '
+	setup_cruft_exclude_tests kept-cruft &&
+	(
+		cd kept-cruft &&
+
+		# Keep HEAD and its tree outside the geometric repack. Its
+		# parent is reachable again, but still in the cruft pack.
+		git rev-parse HEAD HEAD^{tree} >objects &&
+		pack=$(git pack-objects $packdir/pack <objects) &&
+		touch $packdir/pack-$pack.keep &&
+		git prune-packed &&
+
+		# The new blob is still loose, so this writes a pack instead
+		# of taking the no-new-packs fallback.
+		GIT_TEST_MULTI_PACK_INDEX=0 \
+		git repack -d --geometric=2 --write-midx --write-bitmap-index &&
+		git rev-list --test-bitmap HEAD
+	)
+'
+
 test_expect_success 'repack --write-midx includes cruft when instructed' '
 	setup_cruft_exclude_tests exclude-cruft-when-instructed &&
 	(
-- 
2.56.0.4.gbee41d2fc68
Previous: Taylor BlauNext: Jeff King
Message 17 of 41 in “repack: various corner cases for cruft-less MIDXs”
  1. 0/4 repack: various corner cases for cruft-less MIDXsTaylor Blau, Sep 30, 2026
  2. 1/4 pack-objects: introduce `stdin_packs_context` structTaylor Blau, Sep 30, 2026
  3. Junio C HamanoSep 30, 2026
  4. Taylor BlauOct 1, 2026
  5. 2/4 pack-objects: ensure tree/tag closure with '--stdin-packs=follow'Taylor Blau, Sep 30, 2026
  6. Junio C HamanoSep 30, 2026
  7. Derrick StoleeSep 30, 2026
  8. Taylor BlauOct 1, 2026
  9. Elijah NewrenOct 1, 2026
  10. Taylor BlauOct 2, 2026
  11. Jeff KingOct 2, 2026
  12. Jeff KingSep 30, 2026
  13. Taylor BlauOct 1, 2026
  14. 3/4 repack: retain cruft packs in MIDXs after incremental repacksTaylor Blau, Sep 30, 2026
  15. Jeff KingSep 30, 2026
  16. Taylor BlauOct 1, 2026
  17. 4/4 repack: retain cruft packs in MIDXs containing kept packsTaylor Blau, Sep 30, 2026
  18. Jeff KingSep 30, 2026
  19. Taylor BlauOct 1, 2026
  20. Jeff KingSep 30, 2026
  21. Taylor BlauOct 1, 2026
  22. 0/8 repack: various corner cases for cruft-less MIDXsTaylor Blau, Oct 1, 2026
  23. 1/8 pack-objects: introduce `stdin_packs_context` structTaylor Blau, Oct 1, 2026
  24. 2/8 pack-objects: ensure tree/tag closure with '--stdin-packs=follow'Taylor Blau, Oct 1, 2026
  25. Jeff KingOct 2, 2026
  26. Taylor BlauOct 3, 2026
  27. Jeff KingOct 3, 2026
  28. 3/8 repack: retain cruft packs in MIDXs after incremental repacksTaylor Blau, Oct 1, 2026
  29. 4/8 repack: use a sorted list for explicitly kept packsTaylor Blau, Oct 1, 2026
  30. Jeff KingOct 2, 2026
  31. 5/8 repack: follow kept packs when omitting cruft from the MIDXTaylor Blau, Oct 1, 2026
  32. Jeff KingOct 2, 2026
  33. Taylor BlauOct 3, 2026
  34. 6/8 repack: track the preferred pack explicitly in MIDX write stepsTaylor Blau, Oct 1, 2026
  35. Jeff KingOct 2, 2026
  36. Taylor BlauOct 3, 2026
  37. Jeff KingOct 3, 2026
  38. 7/8 repack: defer allocating the append plan's write stepTaylor Blau, Oct 1, 2026
  39. 8/8 repack: include required packs in incremental MIDX writesTaylor Blau, Oct 1, 2026
  40. Jeff KingOct 2, 2026
  41. Taylor BlauOct 3, 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.