#!/bin/sh

# Collects the regression tests for the security fixes and for the behaviour
# this fork adds on top of upstream cgit. Each case builds the smallest
# repository and config that reproduce the original problem and then asks
# for the page that used to mishandle it.

test_description='Check security fixes and fork-specific behavior'
. ./setup.sh

# Most of what follows shares one repository and one config, so the fixture
# carries everything they need at once, a blob over the size limit, readmes
# holding markup that must not reach the page as markup, and a subdirectory
# to sort ahead of the files.
test_expect_success 'set up security fixtures' '
	mkrepo repos/sec 1 &&
	(
		cd repos/sec &&
		dd if=/dev/zero bs=1024 count=4 2>/dev/null | tr "\0" "X" >big.txt &&
		printf "# Title\n<script>alert(1)</script>\n" >README.md &&
		printf "<script>alert(2)</script>\n" >readme.txt &&
		printf "top\n" >afile &&
		mkdir zsub &&
		printf "inner\n" >zsub/inner &&
		git add -A &&
		git commit -m fixtures
	) &&
	{
		echo "virtual-root=/" &&
		echo "cache-size=0" &&
		echo "max-blob-size=1" &&
		echo "enable-blame=1" &&
		echo "enable-tree-group-dirs=1" &&
		echo "repo.url=sec" &&
		echo "repo.path=$PWD/repos/sec/.git"
	} >seccgitrc
'

secq() { CGIT_CONFIG="$PWD/seccgitrc" QUERY_STRING="$1" cgit; }

# A revision beginning with a dash reaches git as an option rather than as a
# tip, so a request for id=--output=<path> could create or truncate any file
# the server is able to write.
test_expect_success 'log id=--output does not write a file' '
	rm -f pwned &&
	cgit_query "url=foo/log&id=--output=$PWD/pwned" >tmp 2>&1 &&
	! test -e pwned
'

test_expect_success 'log id=--output is rejected as an invalid revision' '
	grep -i "invalid revision" tmp
'

test_expect_success 'a normal log still renders' '
	cgit_query "url=foo/log" >tmp &&
	grep -i "commit 5" tmp
'

test_expect_success 'a valid id= still renders the log' '
	sha=$(git -C repos/foo rev-parse HEAD) &&
	cgit_query "url=foo/log&id=$sha" >tmp &&
	grep -i "commit 5" tmp
'

# max-blob-size is enforced before the object is read, so every view that
# would otherwise inline a file has to turn the same one away rather than
# inflate it first and think better of it afterwards.
test_expect_success 'tree view refuses an oversized blob' '
	secq "url=sec/tree/big.txt" | grep -iE "exceeds|too large"
'

test_expect_success 'plain view refuses an oversized blob' '
	secq "url=sec/plain/big.txt" | grep -iE "exceeds|too large|413"
'

test_expect_success 'blame view refuses an oversized blob' '
	secq "url=sec/blame/big.txt" | grep -iE "exceeds|too large"
'

test_expect_success 'a small blob is still served' '
	secq "url=sec/plain/afile" | grep -F "top"
'

# A readme is repository content, so with no about filter configured it has
# to reach the page escaped instead of as live markup, whatever its name
# suggests about the format.
test_expect_success 'markdown readme without a filter is escaped as plain text' '
	{
		echo "virtual-root=/" &&
		echo "cache-size=0" &&
		echo "repo.url=md" &&
		echo "repo.path=$PWD/repos/sec/.git" &&
		echo "repo.readme=master:README.md"
	} >secmdrc &&
	CGIT_CONFIG="$PWD/secmdrc" QUERY_STRING="url=md/about/" cgit >tmp &&
	grep "pre class=.plaintext." tmp &&
	grep "&lt;script&gt;" tmp &&
	! grep "<script>alert(1)</script>" tmp
'

test_expect_success 'non-markdown readme without a filter is escaped' '
	{
		echo "virtual-root=/" &&
		echo "cache-size=0" &&
		echo "repo.url=txt" &&
		echo "repo.path=$PWD/repos/sec/.git" &&
		echo "repo.readme=master:readme.txt"
	} >sectxtrc &&
	CGIT_CONFIG="$PWD/sectxtrc" QUERY_STRING="url=txt/about/" cgit >tmp &&
	grep "&lt;script&gt;" tmp &&
	! grep "<script>alert(2)</script>" tmp
'

test_expect_success 'non-markdown readme keeps its line structure' '
	grep "pre class=.plaintext." tmp
'

# A Content-Security-Policy without unsafe-inline stops an inline onchange
# handler from ever running, so the option form submits through a plain
# button and its selects carry no handlers at all.
test_expect_success 'diff options submit through a button' '
	sha=$(git -C repos/foo rev-parse HEAD) &&
	cgit_query "url=foo/commit&id=$sha" >tmp &&
	grep "type=.submit. value=.reload." tmp &&
	! grep "onchange" tmp
'

# Grouping directories ahead of files is behaviour this fork adds, so nothing
# upstream covers it.
test_expect_success 'tree groups directories before files' '
	secq "url=sec/tree/" >tmp &&
	dirline=$(grep -n "tree/zsub" tmp | head -1 | cut -d: -f1) &&
	fileline=$(grep -n "tree/afile" tmp | head -1 | cut -d: -f1) &&
	test -n "$dirline" &&
	test -n "$fileline" &&
	test "$dirline" -lt "$fileline"
'

# A file name is repository content and may hold a quote, which would break
# out of the href attribute on the line number links of a side by side diff,
# so the path is percent-encoded on its way into them.
test_expect_success 'ssdiff percent-encodes a quoted file path' '
	mkrepo repos/xss 1 &&
	name=$(printf "x\047y.txt") &&
	(
		cd repos/xss &&
		printf "a\nb\n" >"$name" &&
		git add -A &&
		git commit -m add &&
		printf "a\nc\n" >"$name" &&
		git commit -am change
	) &&
	{
		echo "virtual-root=/" &&
		echo "cache-size=0" &&
		echo "repo.url=xss" &&
		echo "repo.path=$PWD/repos/xss/.git"
	} >xssrc &&
	sha=$(git -C repos/xss rev-parse HEAD) &&
	CGIT_CONFIG="$PWD/xssrc" QUERY_STRING="url=xss/diff/&id=$sha&ss=1" cgit >tmp &&
	grep "tree/x%27y.txt" tmp &&
	! grep "href=.[^>]*x.y.txt.[^>]*>" tmp
'

# git itself will write a commit with an empty message, so the log and the
# summary both have to have something to print where the subject goes.
test_expect_success 'a message-less commit renders without crashing' '
	mkrepo repos/nomsg 1 &&
	(
		cd repos/nomsg &&
		tree=$(git write-tree) &&
		printf "tree %s\nauthor a <a@b> 1735689600 +0000\ncommitter a <a@b> 1735689600 +0000\n" "$tree" >raw &&
		cid=$(git hash-object -t commit -w raw) &&
		git update-ref refs/heads/master "$cid"
	) &&
	{
		echo "virtual-root=/" &&
		echo "cache-size=0" &&
		echo "repo.url=nomsg" &&
		echo "repo.path=$PWD/repos/nomsg/.git"
	} >nomsgrc &&
	CGIT_CONFIG="$PWD/nomsgrc" QUERY_STRING="url=nomsg/log/&showmsg=1" cgit >tmp &&
	grep "no commit message" tmp &&
	CGIT_CONFIG="$PWD/nomsgrc" QUERY_STRING="url=nomsg/" cgit >tmp &&
	grep "no commit message" tmp
'

# A branch need not point at a commit. struct refinfo keeps taginfo and
# commitinfo in a union and fills only the member matching the object type,
# so sorting branches through the commit member read past the end of the
# smaller taginfo, and read NULL for a tree, which crashed. git update-ref
# refuses to create such a ref, hence the loose files written by hand below,
# and a repository is only files on disk so cgit meets whatever is there.
test_expect_success 'set up a repo whose branches point at odd objects' '
	mkrepo repos/oddref 2 &&
	(
		cd repos/oddref &&
		git tag -a annotated -m note &&
		git rev-parse annotated >.git/refs/heads/points-at-tag &&
		git rev-parse HEAD^{tree} >.git/refs/heads/points-at-tree &&
		git for-each-ref refs/heads/ >refs.out &&
		grep -q "tree.refs/heads/points-at-tree" refs.out &&
		grep -q "tag.refs/heads/points-at-tag" refs.out
	) &&
	{
		echo "virtual-root=/" &&
		echo "cache-size=0" &&
		echo "branch-sort=age" &&
		echo "repo.url=oddref" &&
		echo "repo.path=$PWD/repos/oddref/.git"
	} >oddrefrc
'

test_expect_success 'refs page sorts such branches without crashing' '
	CGIT_CONFIG="$PWD/oddrefrc" QUERY_STRING="url=oddref/refs/" cgit >tmp &&
	grep "points-at-tree" tmp &&
	grep "</html>" tmp
'

test_expect_success 'the branch page sorts them without crashing' '
	CGIT_CONFIG="$PWD/oddrefrc" QUERY_STRING="url=oddref/refs/heads/" cgit >tmp &&
	grep "points-at-tree" tmp &&
	grep "</html>" tmp
'

test_expect_success 'the summary page sorts them without crashing' '
	CGIT_CONFIG="$PWD/oddrefrc" QUERY_STRING="url=oddref/" cgit >tmp &&
	grep "points-at-tree" tmp &&
	grep "</html>" tmp
'

test_expect_success 'name-sorted branches are unaffected' '
	{
		echo "virtual-root=/" &&
		echo "cache-size=0" &&
		echo "branch-sort=name" &&
		echo "repo.url=oddref" &&
		echo "repo.path=$PWD/repos/oddref/.git"
	} >oddrefnamerc &&
	CGIT_CONFIG="$PWD/oddrefnamerc" QUERY_STRING="url=oddref/refs/heads/" cgit >tmp &&
	grep "points-at-tree" tmp &&
	grep "</html>" tmp
'

# module-link is a template and scan-path lets a repository set its own from
# a cgitrc in the tree, so handing that string to printf let the owner of a
# scanned repository crash the process, or read stack memory into the served
# page, merely by adding conversions past the two that are filled. The tests
# after the fixture also pin down the templates that must keep working.
test_expect_success 'set up a submodule fixture with a hostile module-link' '
	mkrepo repos/modlink 1 &&
	(
		cd repos/modlink &&
		sub=$(git rev-parse HEAD) &&
		git update-index --add --cacheinfo 160000,$sub,submod &&
		git commit -m gitlink
	) &&
	mkdir -p scan &&
	cp -R repos/modlink scan/modlink &&
	printf "module-link=/m/%%s/%%s/%%s/%%s/%%s/%%s/%%s/%%s/%%s/%%s\n" >scan/modlink/.git/cgitrc &&
	{
		echo "virtual-root=/" &&
		echo "cache-size=0" &&
		echo "scan-path=$PWD/scan"
	} >modlinkrc
'

test_expect_success 'a surplus conversion does not crash the tree view' '
	CGIT_CONFIG="$PWD/modlinkrc" QUERY_STRING="url=modlink/tree/" cgit >tmp &&
	grep "ls-mod" tmp
'

test_expect_success 'a surplus conversion is shown literally, not filled' '
	grep "href=./m/submod/[0-9a-f]*/%s/%s/" tmp
'

test_expect_success 'a well-formed module-link still takes name and sha1' '
	{
		echo "virtual-root=/" &&
		echo "cache-size=0" &&
		echo "module-link=/mod/%s/commit/?id=%s" &&
		echo "repo.url=modlink" &&
		echo "repo.path=$PWD/repos/modlink/.git"
	} >modlinkokrc &&
	sub=$(git -C repos/modlink rev-parse HEAD~1) &&
	CGIT_CONFIG="$PWD/modlinkokrc" QUERY_STRING="url=modlink/tree/" cgit >tmp &&
	grep "href=./mod/submod/commit/?id=$sub." tmp
'

test_expect_success 'a per-path module-link takes only the sha1' '
	{
		echo "virtual-root=/" &&
		echo "cache-size=0" &&
		echo "repo.url=modlink" &&
		echo "repo.path=$PWD/repos/modlink/.git" &&
		echo "repo.module-link.submod=https://example.com/s/?id=%s"
	} >modlinkpathrc &&
	sub=$(git -C repos/modlink rev-parse HEAD~1) &&
	CGIT_CONFIG="$PWD/modlinkpathrc" QUERY_STRING="url=modlink/tree/" cgit >tmp &&
	grep "href=.https://example.com/s/?id=$sub." tmp
'

test_expect_success 'a doubled percent in a module-link renders as one' '
	{
		echo "virtual-root=/" &&
		echo "cache-size=0" &&
		echo "module-link=/m/%s/%%/%s" &&
		echo "repo.url=modlink" &&
		echo "repo.path=$PWD/repos/modlink/.git"
	} >modlinkpctrc &&
	sub=$(git -C repos/modlink rev-parse HEAD~1) &&
	CGIT_CONFIG="$PWD/modlinkpctrc" QUERY_STRING="url=modlink/tree/" cgit >tmp &&
	grep "href=./m/submod/%/$sub." tmp
'

# The .gitmodules file is commit-controlled content, so a derived link may
# carry any scheme and any byte an author can commit. Only http and https
# may reach an href, and everything lands attribute-escaped.
test_expect_success 'set up a hostile .gitmodules' '
	(
		cd repos/modlink &&
		sub=$(git rev-parse HEAD) &&
		git update-index --add --cacheinfo 160000,$sub,quoted &&
		cat >.gitmodules <<-EOF &&
		[submodule "submod"]
			path = submod
			url = javascript:alert(1)
		[submodule "quoted"]
			path = quoted
			url = https://example.com/x'\''><script>alert(1)</script>
		EOF
		git add .gitmodules &&
		git commit -m hostile
	) &&
	{
		echo "virtual-root=/" &&
		echo "cache-size=0" &&
		echo "enable-gitmodules-links=1" &&
		echo "repo.url=modlink" &&
		echo "repo.path=$PWD/repos/modlink/.git"
	} >modlinkgmrc
'

test_expect_success 'a javascript url never reaches an href' '
	CGIT_CONFIG="$PWD/modlinkgmrc" QUERY_STRING="url=modlink/tree/" cgit >tmp &&
	! grep "href=.javascript:" tmp &&
	grep "class=.ls-mod. title=.javascript:alert(1).>submod</span>" tmp
'

test_expect_success 'a quote in a web url cannot break out of the href' '
	! grep "<script>alert" tmp &&
	grep "href=.https://example.com/x&#x27;&gt;&lt;script&gt;" tmp
'

test_done