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

[PATCH 5/5] cache-tree: fix verify_cache() to catch non-adjacent D/F conflicts

From
Elijah Newren via GitGitGadget <gitgitgadget@gmail.com>
Date
Apr 21, 2026, 00:26 UTC
Message-ID
<a87bbaa84fd5dcb2a585f82c4a5dfa1572b54588.1776731171.git.gitgitgadget@gmail.com>
In-Reply-To
<pull.2096.git.1776731171.gitgitgadget@gmail.com>
From: Elijah Newren <newren@gmail.com>

verify_cache() checks that the index does not contain both "path" and "path/file" before writing a tree. It does this by comparing only adjacent entries, relying on the assumption that "path/file" would immediately follow "path" in sorted order. Unfortunately, this assumption does not always hold. For example:

    docs                     <-- submodule entry
    docs-internal/README.md  <-- intervening entry
    docs/requirements.txt    <-- D/F conflict, NOT adjacent to "docs"

When this happens, verify_cache() silently misses the D/F conflict and write-tree produces a corrupt tree object containing duplicate entries (one for the submodule "docs" and one for the tree "docs").

I could not find any caller in current git that both allows the index to get into this state and then tries to write it out without doing other checks beyond the verify_cache() call in cache_tree_update(), but verify_cache() is documented as a safety net for preventing corrupt trees and should actually provide that guarantee. A downstream consumer that relied solely on cache_tree_update()'s internal checking via verify_cache() to prevent duplicate tree entries was bitten by the gap.

Add a test that constructs a corrupt index directly (bypassing the D/F checks in add_index_entry) and verifies that write-tree now rejects it.

Signed-off-by: Elijah Newren <newren@gmail.com>
---
 cache-tree.c                   | 46 ++++++++++++++++++++++++--
 t/meson.build                  |  1 +
 t/t0093-direct-index-write.pl  | 38 ++++++++++++++++++++++
 t/t0093-verify-cache-df-gap.sh | 59 ++++++++++++++++++++++++++++++++++
 4 files changed, 141 insertions(+), 3 deletions(-)
 create mode 100644 t/t0093-direct-index-write.pl
 create mode 100755 t/t0093-verify-cache-df-gap.sh
diff --git a/cache-tree.c b/cache-tree.c
index 7881b42aa2..f11844fe72 100644
--- a/cache-tree.c
+++ b/cache-tree.c
@@ -192,22 +192,62 @@ static int verify_cache(struct index_state *istate, int flags)
 	for (i = 0; i + 1 < istate->cache_nr; i++) {
 		/* path/file always comes after path because of the way
 		 * the cache is sorted.  Also path can appear only once,
-		 * which means conflicting one would immediately follow.
+		 * so path/file is likely the immediately following path
+		 * but might be separated if there is e.g. a
+		 * path-internal/... file.
 		 */
 		const struct cache_entry *this_ce = istate->cache[i];
 		const struct cache_entry *next_ce = istate->cache[i + 1];
 		const char *this_name = this_ce->name;
 		const char *next_name = next_ce->name;
 		int this_len = ce_namelen(this_ce);
+		const char *conflict_name = NULL;
+
 		if (this_len < ce_namelen(next_ce) &&
-		    next_name[this_len] == '/' &&
+		    next_name[this_len] <= '/' &&
 		    strncmp(this_name, next_name, this_len) == 0) {
+			if (next_name[this_len] == '/') {
+				conflict_name = next_name;
+			} else if (next_name[this_len] < '/') {
+				/*
+				 * The immediately next entry shares our
+				 * prefix but sorts before "path/" (e.g.,
+				 * "path-internal" between "path" and
+				 * "path/file", since '-' (0x2D) < '/'
+				 * (0x2F)).  Binary search to find where
+				 * "path/" would be and check for a D/F
+				 * conflict there.
+				 */
+				struct cache_entry *other;
+				struct strbuf probe = STRBUF_INIT;
+				int pos;
+
+				strbuf_add(&probe, this_name, this_len);
+				strbuf_addch(&probe, '/');
+				pos = index_name_pos_sparse(istate,
+							    probe.buf,
+							    probe.len);
+				strbuf_release(&probe);
+
+				if (pos < 0)
+					pos = -pos - 1;
+				if (pos >= (int)istate->cache_nr)
+					continue;
+				other = istate->cache[pos];
+				if (ce_namelen(other) > this_len &&
+				    other->name[this_len] == '/' &&
+				    !strncmp(this_name, other->name, this_len))
+					conflict_name = other->name;
+			}
+		}
+
+		if (conflict_name) {
 			if (10 < ++funny) {
 				fprintf(stderr, "...\n");
 				break;
 			}
 			fprintf(stderr, "You have both %s and %s\n",
-				this_name, next_name);
+				this_name, conflict_name);
 		}
 	}
 	if (funny)
diff --git a/t/meson.build b/t/meson.build
index 7528e5cda5..362177999b 100644
--- a/t/meson.build
+++ b/t/meson.build
@@ -124,6 +124,7 @@ integration_tests = [
   't0090-cache-tree.sh',
   't0091-bugreport.sh',
   't0092-diagnose.sh',
+  't0093-verify-cache-df-gap.sh',
   't0095-bloom.sh',
   't0100-previous.sh',
   't0101-at-syntax.sh',
diff --git a/t/t0093-direct-index-write.pl b/t/t0093-direct-index-write.pl
new file mode 100644
index 0000000000..2881a3ebb2
--- /dev/null
+++ b/t/t0093-direct-index-write.pl
@@ -0,0 +1,38 @@
+#!/usr/bin/perl
+#
+# Build a v2 index file from entries listed on stdin.
+# Each line: "octalmode hex-oid name"
+# Output: binary index written to stdout.
+#
+# This bypasses all D/F safety checks in add_index_entry(), simulating
+# what happens when code uses ADD_CACHE_JUST_APPEND to bulk-load entries.
+use strict;
+use warnings;
+use Digest::SHA qw(sha1 sha256);
+
+my $hash_algo = $ENV{'GIT_DEFAULT_HASH'} || 'sha1';
+my $hash_func = $hash_algo eq 'sha256' ? \&sha256 : \&sha1;
+
+my @entries;
+while (my $line = <STDIN>) {
+	chomp $line;
+	my ($mode, $oid_hex, $name) = split(/ /, $line, 3);
+	push @entries, [$mode, $oid_hex, $name];
+}
+
+my $body = "DIRC" . pack("NN", 2, scalar @entries);
+
+for my $ent (@entries) {
+	my ($mode, $oid_hex, $name) = @{$ent};
+	# 10 x 32-bit stat fields (zeroed), with mode in position 7
+	my $stat = pack("N10", 0, 0, 0, 0, 0, 0, oct($mode), 0, 0, 0);
+	my $oid = pack("H*", $oid_hex);
+	my $flags = pack("n", length($name) & 0xFFF);
+	my $entry = $stat . $oid . $flags . $name . "\0";
+	# Pad to 8-byte boundary
+	while (length($entry) % 8) { $entry .= "\0"; }
+	$body .= $entry;
+}
+
+binmode STDOUT;
+print $body . $hash_func->($body);
diff --git a/t/t0093-verify-cache-df-gap.sh b/t/t0093-verify-cache-df-gap.sh
new file mode 100755
index 0000000000..0b6829d805
--- /dev/null
+++ b/t/t0093-verify-cache-df-gap.sh
@@ -0,0 +1,59 @@
+#!/bin/sh
+
+test_description='verify_cache() must catch non-adjacent D/F conflicts
+
+Ensure that verify_cache() can complain about bad entries like:
+
+  docs               <-- submodule
+  docs-internal/...  <-- sorts here because "-" < "/"
+  docs/...           <-- D/F conflict with "docs" above, not adjacent
+
+In order to test verify_cache, we directly construct a corrupt index
+(bypassing the D/F safety checks in add_index_entry) and verify that
+write-tree rejects it.
+'
+
+. ./test-lib.sh
+
+if ! test_have_prereq PERL
+then
+	skip_all='skipping verify_cache D/F tests; Perl not available'
+	test_done
+fi
+
+# Build a v2 index from entries on stdin, bypassing D/F checks.
+# Each line: "octalmode hex-oid name" (entries must be pre-sorted).
+build_corrupt_index () {
+	perl "$TEST_DIRECTORY/t0093-direct-index-write.pl" >"$1"
+}
+
+test_expect_success 'setup objects' '
+	test_commit base &&
+	BLOB=$(git rev-parse HEAD:base.t) &&
+	SUB_COMMIT=$(git rev-parse HEAD)
+'
+
+test_expect_success 'adjacent D/F conflict is caught by verify_cache' '
+	cat >index-entries <<-EOF &&
+	0160000 $SUB_COMMIT docs
+	0100644 $BLOB docs/requirements.txt
+	EOF
+	build_corrupt_index .git/index <index-entries &&
+
+	test_must_fail git write-tree 2>err &&
+	test_grep "You have both docs and docs/requirements.txt" err
+'
+
+test_expect_success 'non-adjacent D/F conflict is caught by verify_cache' '
+	cat >index-entries <<-EOF &&
+	0160000 $SUB_COMMIT docs
+	0100644 $BLOB docs-internal/README.md
+	0100644 $BLOB docs/requirements.txt
+	EOF
+	build_corrupt_index .git/index <index-entries &&
+
+	test_must_fail git write-tree 2>err &&
+	test_grep "You have both docs and docs/requirements.txt" err
+'
+
+test_done
-- 
gitgitgadget
Previous: Junio C HamanoNext: Junio C Hamano
Message 10 of 23 in “Duplicate entry hardening”
  1. 0/5 Duplicate entry hardeningElijah Newren via GitGitGadget, Apr 21, 2026
  2. 1/5 merge-ort: propagate callback errors from traverse_trees_wrapper()Elijah Newren via GitGitGadget, Apr 21, 2026
  3. Junio C HamanoJun 1, 2026
  4. Elijah NewrenJun 14, 2026
  5. 2/5 merge-ort: drop unnecessary show_all_errors from collect_merge_info()Elijah Newren via GitGitGadget, Apr 21, 2026
  6. Junio C HamanoJun 1, 2026
  7. 3/5 merge-ort: free diff pairs queue in clear_or_reinit_internal_opts()Elijah Newren via GitGitGadget, Apr 21, 2026
  8. 4/5 merge-ort: abort merge when trees have duplicate entriesElijah Newren via GitGitGadget, Apr 21, 2026
  9. Junio C HamanoJun 1, 2026
  10. 5/5 cache-tree: fix verify_cache() to catch non-adjacent D/F conflictsElijah Newren via GitGitGadget, Apr 21, 2026
  11. Junio C HamanoJun 1, 2026
  12. Elijah NewrenJun 14, 2026
  13. Junio C HamanoJun 1, 2026
  14. Patrick SteinhardtJun 1, 2026
  15. Automated reviews by AI (was Re: [PATCH 0/5] Duplicate entry hardening)Christian Couder, Jun 12, 2026
  16. Junio C HamanoJun 12, 2026
  17. 0/5 Duplicate entry hardeningElijah Newren via GitGitGadget, Jun 14, 2026
  18. 1/5 merge-ort: propagate callback errors from traverse_trees_wrapper()Elijah Newren via GitGitGadget, Jun 14, 2026
  19. 2/5 merge-ort: drop unnecessary show_all_errors from collect_merge_info()Elijah Newren via GitGitGadget, Jun 14, 2026
  20. 3/5 merge-ort: free diff pairs queue in clear_or_reinit_internal_opts()Elijah Newren via GitGitGadget, Jun 14, 2026
  21. 4/5 merge-ort: abort merge when trees have duplicate entriesElijah Newren via GitGitGadget, Jun 14, 2026
  22. 5/5 cache-tree: fix verify_cache() to catch non-adjacent D/F conflictsElijah Newren via GitGitGadget, Jun 14, 2026
  23. Junio C HamanoJun 14, 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.