[BUG REPORT] t5322: demonstrate a pack-objects bug

[BUG REPORT] t5322: demonstrate a pack-objects bug

[Derrick Stolee]

I was preparing a new "sparse" algorithm for calculating the
interesting objects to send on push. The important steps happen
during 'git pack-objects', so I was creating test cases to see
how the behavior changes in narrow cases. Specifically, when
copying a directory across sibling directories (see test case),
the new logic would accidentally send that object as an extra.

However, I found a bug in the existing logic. The included test
demonstrates this during the final 'git index-pack' call. It
fails with the message

	'fatal: pack has 1 unresolved delta'

It is probable that this is not a minimal test case, but happens
to be the test I had created before discovering the problem.

I compiled v2.17.0 and v2.12.0 as checks to see if I could find
a "good" commit with which to start a bisect, but both failed.
This is an old bug!

Signed-off-by: Derrick Stolee <dstolee@microsoft.com>
---
 t/t5322-pack-objects-sparse.sh | 94 ++++++++++++++++++++++++++++++++++
 1 file changed, 94 insertions(+)
 create mode 100755 t/t5322-pack-objects-sparse.sh

diff --git a/t/t5322-pack-objects-sparse.sh b/t/t5322-pack-objects-sparse.sh
new file mode 100755
index 0000000000..36faa70fe9
--- /dev/null
+++ b/t/t5322-pack-objects-sparse.sh
@@ -0,0 +1,94 @@
+#!/bin/sh
+
+test_description='pack-objects object selection using sparse algorithm'
+. ./test-lib.sh
+
+test_expect_success 'setup repo' '
+	test_commit initial &&
+	for i in $(test_seq 1 3)
+	do
+		mkdir f$i &&
+		for j in $(test_seq 1 3)
+		do
+			mkdir f$i/f$j &&
+			echo $j >f$i/f$j/data.txt
+		done
+	done &&
+	git add . &&
+	git commit -m "Initialized trees" &&
+	for i in $(test_seq 1 3)
+	do
+		git checkout -b topic$i master &&
+		echo change-$i >f$i/f$i/data.txt &&
+		git commit -a -m "Changed f$i/f$i/data.txt"
+	done &&
+	cat >packinput.txt <<-EOF &&
+	topic1
+	^topic2
+	^topic3
+	EOF
+	git rev-parse			\
+		topic1			\
+		topic1^{tree}		\
+		topic1:f1		\
+		topic1:f1/f1		\
+		topic1:f1/f1/data.txt | sort >actual_objects.txt
+'
+
+test_expect_success 'non-sparse pack-objects' '
+	git pack-objects --stdout --thin --revs <packinput.txt >nonsparse.pack &&
+	git index-pack -o nonsparse.idx nonsparse.pack &&
+	git show-index <nonsparse.idx | awk "{print \$2}" >nonsparse_objects.txt &&
+	test_cmp actual_objects.txt nonsparse_objects.txt
+'
+
+# Demonstrate that both algorithms send "extra" objects because
+# they are not in the frontier.
+
+test_expect_success 'duplicate a folder from f3 and commit to topic1' '
+	git checkout topic1 &&
+	echo change-3 >f3/f3/data.txt &&
+	git commit -a -m "Changed f3/f3/data.txt" &&
+	git rev-parse			\
+		topic1~1		\
+		topic1~1^{tree}		\
+		topic1^{tree}		\
+		topic1			\
+		topic1:f1		\
+		topic1:f1/f1		\
+		topic1:f1/f1/data.txt	\
+		topic1:f3		\
+		topic1:f3/f3		\
+		topic1:f3/f3/data.txt | sort >actual_objects.txt
+'
+
+test_expect_success 'non-sparse pack-objects' '
+	git pack-objects --stdout --thin --revs <packinput.txt >nonsparse.pack &&
+	git index-pack -o nonsparse.idx nonsparse.pack &&
+	git show-index <nonsparse.idx | awk "{print \$2}" >nonsparse_objects.txt &&
+	test_cmp actual_objects.txt nonsparse_objects.txt
+'
+
+test_expect_success 'duplicate a folder from f3 and commit to topic1' '
+	mkdir f3/f4 &&
+	cp -r f1/f1/* f3/f4 &&
+	git add f3/f4 &&
+	git commit -m "Copied f1/f1 to f3/f4" &&
+	cat >packinput.txt <<-EOF &&
+	topic1
+	^topic1~1
+	EOF
+	git rev-parse		\
+		topic1		\
+		topic1^{tree}	\
+		topic1:f3 | sort >actual_objects.txt
+'
+
+test_expect_success 'non-sparse pack-objects' '
+	git pack-objects --stdout --thin --revs <packinput.txt >nonsparse.pack &&
+	git index-pack -o nonsparse.idx nonsparse.pack &&
+	git show-index <nonsparse.idx | awk "{print \$2}" >nonsparse_objects.txt &&
+	test_cmp actual_objects.txt nonsparse_objects.txt
+'
+
+test_done
-- 
2.20.0.rc1

Re: [BUG REPORT] t5322: demonstrate a pack-objects bug

[Derrick Stolee]

On 11/28/2018 2:45 PM, Derrick Stolee wrote:
> I was preparing a new "sparse" algorithm for calculating the
> interesting objects to send on push. The important steps happen
> during 'git pack-objects', so I was creating test cases to see
> how the behavior changes in narrow cases. Specifically, when
> copying a directory across sibling directories (see test case),
> the new logic would accidentally send that object as an extra.
>
> However, I found a bug in the existing logic. The included test
> demonstrates this during the final 'git index-pack' call. It
> fails with the message
>
> 	'fatal: pack has 1 unresolved delta'

I realize now that I've gone through this effort that the problem is me 
(of course it is).

> +	git pack-objects --stdout --thin --revs <packinput.txt >nonsparse.pack &&

Since I'm packing thin packs, the index operation is complaining about 
deltas. So, I need to use a different mechanism to list the objects in 
the pack (say, 'git verify-pack -v').

Sorry for the noise!

Thanks,

-Stolee