#!/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 behaviour'
. ./setup.sh

# Some cases below share one repository and one config, so the fixture
# carries everything they need at once, readmes holding markup that must not
# reach the page as markup and a subdirectory to sort ahead of the files.
# The blob limit is the lowest there is, 1 KB, so a small file shows that a
# blob under it is still served.
test_expect_success 'set up security fixtures' '
	mkrepo repos/sec 1 &&
	(
		cd repos/sec &&
		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
	) &&
	cgitrc_repo seccgitrc sec max-blob-size=1 enable-tree-group-dirs=1
'

# 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. The request check and the log page refuse it
# with the same words, so the page cannot tell which of them did.
test_expect_success 'log id=--output is refused without writing a file' '
	cgitrc_copy nocacherc &&
	cgit_rc nocacherc "url=foo/log&id=--output=$PWD/log-output" >tmp &&
	test_grep "^Status: 400" tmp &&
	test_grep "<div class=.error.>Invalid revision</div>" tmp &&
	test_path_is_missing log-output
'

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

test_expect_success 'a small blob is still served' '
	cgit_ok seccgitrc "url=sec/plain/afile" >tmp &&
	strip_headers <tmp >body &&
	echo top >expect &&
	test_cmp expect body
'

# 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=$PWD/repos/sec/.git" &&
		echo "repo.readme=master:README.md" &&
		echo "repo.url=txt" &&
		echo "repo.path=$PWD/repos/sec/.git" &&
		echo "repo.readme=master:readme.txt"
	} >secreadmerc &&
	cgit_ok secreadmerc "url=md/about/" >tmp &&
	test_grep "pre class=.plaintext.># Title$" tmp &&
	test_grep "^&lt;script&gt;alert(1)&lt;/script&gt;$" tmp &&
	test_grep ! "<script>alert(1)</script>" tmp
'

test_expect_success 'non-markdown readme without a filter is escaped as plain text' '
	cgit_ok secreadmerc "url=txt/about/" >tmp &&
	test_grep "pre class=.plaintext.>&lt;script&gt;alert(2)&lt;/script&gt;$" tmp &&
	test_grep ! "<script>alert(2)</script>" tmp
'

# root-readme is the site's own file rather than repository content, so it
# goes out as written, on an about page of its own.
test_expect_success 'root-readme without a filter is its own page, as written' '
	echo "<b>site</b>" >root.html &&
	{
		echo "virtual-root=/" &&
		echo "cache-size=0" &&
		echo "root-readme=$PWD/root.html"
	} >secrootrc &&
	cgit_rc secrootrc "url=" >tmp &&
	grep "<li><a title=.About this site. href=./?p=about.>about</a>" tmp &&
	! grep "<b>site</b>" tmp &&
	cgit_rc secrootrc "p=about" >tmp &&
	grep "^<b>site</b>$" 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, and nor does any other
# page.
test_expect_success 'diff options submit through a button' '
	sha=$(git -C repos/foo rev-parse HEAD) &&
	cgit_ok_url "foo/commit&id=$sha" >tmp &&
	test_grep "type=.submit. value=.reload." tmp &&
	test_grep ! " on[a-z]*=" tmp
'

test_expect_success 'no page carries an inline handler' '
	for page in repolist: summary:foo/ log:foo/log/ tree:foo/tree/ diff:foo/diff/ \
		refs:foo/refs/
	do
		cgit_ok_url "${page#*:}" >tmp &&
		test_grep "data-page=.${page%%:*}." tmp &&
		test_grep ! " on[a-z]*=" tmp || return 1
	done
'

# Grouping directories ahead of files is behaviour this fork adds, so nothing
# upstream covers it.
test_expect_success 'tree groups directories before files' '
	cgit_rc seccgitrc "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,
# or a hash or question mark, which the attribute escaping leaves alone and
# a browser would read as the end of the path. So the path is percent-encoded
# on its way into them.
test_expect_success 'ssdiff percent-encodes the file path in its line links' '
	mkrepo repos/xss 1 &&
	quoted=$(printf "x\047y.txt") &&
	(
		cd repos/xss &&
		printf "a\nb\n" >"$quoted" &&
		printf "a\nb\n" >"p#q?.txt" &&
		git add -A &&
		git commit -m add &&
		printf "a\nc\n" >"$quoted" &&
		printf "a\nc\n" >"p#q?.txt" &&
		git commit -am change
	) &&
	cgitrc_repo xssrc xss &&
	old=$(git -C repos/xss rev-parse HEAD~1) &&
	new=$(git -C repos/xss rev-parse HEAD) &&
	cgit_ok xssrc "url=xss/diff/&id=$new&ss=1" >tmp &&
	for path in x%27y.txt p%23q%3F.txt
	do
		test_grep "<td class=.lineno.><a href=./xss/tree/$path?id=$old#n2.>2</a></td>" tmp &&
		test_grep "<td class=.lineno.><a href=./xss/tree/$path?id=$new#n2.>2</a></td>" tmp || return 1
	done
'

# 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"
	) &&
	cgitrc_repo nomsgrc nomsg &&
	cgit_rc nomsgrc "url=nomsg/log/&showmsg=1" >tmp &&
	grep "no commit message" tmp &&
	cgit_rc nomsgrc "url=nomsg/" >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
	) &&
	cgitrc_repo oddrefrc oddref branch-sort=age
'

test_expect_success 'the refs, branch and summary pages sort such branches' '
	for page in refs/ refs/heads/ ""
	do
		cgit_rc oddrefrc "url=oddref/$page" >tmp &&
		grep "points-at-tree" tmp &&
		grep "</html>" tmp || return 1
	done
'

test_expect_success 'name-sorted branches are unaffected' '
	sed "s/^branch-sort=age$/branch-sort=name/" oddrefrc >oddrefnamerc &&
	cgit_rc oddrefnamerc "url=oddref/refs/heads/" >tmp &&
	grep "points-at-tree" tmp &&
	grep "</html>" tmp
'

# module-link is a template and a trusted scan 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 &&
	sub=$(git -C repos/modlink rev-parse HEAD) &&
	(
		cd repos/modlink &&
		git update-index --add --cacheinfo 160000,$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 "trust-scan-config=1" &&
		echo "scan-path=$PWD/scan"
	} >modlinkrc
'

test_expect_success 'a surplus conversion does not crash the tree view' '
	cgit_rc modlinkrc "url=modlink/tree/" >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' '
	cgitrc_repo modlinkokrc modlink "module-link=/mod/%s/commit/?id=%s" &&
	cgit_rc modlinkokrc "url=modlink/tree/" >tmp &&
	grep "href=./mod/submod/commit/?id=$sub." tmp
'

test_expect_success 'a per-path module-link takes only the sha1' '
	cgitrc_repo modlinkpathrc modlink "repo.module-link.submod=https://example.com/s/?id=%s" &&
	cgit_rc modlinkpathrc "url=modlink/tree/" >tmp &&
	grep "href=.https://example.com/s/?id=$sub." tmp
'

test_expect_success 'a doubled percent in a module-link renders as one' '
	cgitrc_repo modlinkpctrc modlink "module-link=/m/%s/%%/%s" &&
	cgit_rc modlinkpctrc "url=modlink/tree/" >tmp &&
	grep "href=./m/submod/%/$sub." tmp
'

# An unknown page under a repository used to reach the page header with no
# head resolved and crash in the branch switcher.
test_expect_success 'an unknown page under a repository answers 404 with its header' '
	cgit_url "foo/nonsense/" >tmp &&
	grep "^Status: 404 Not Found" tmp &&
	grep "Invalid request" tmp &&
	grep "<select name=.h. " tmp
'

test_expect_success 'a clone endpoint with http clone off answers the same way' '
	{
		echo "enable-http-clone=0" &&
		cat cgitrc
	} >noclonerc &&
	cgit_rc noclonerc "url=foo/info/refs" >tmp &&
	grep "^Status: 404 Not Found" tmp &&
	grep "<select name=.h. " 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
	) &&
	cgitrc_repo modlinkgmrc modlink enable-gitmodules-links=1
'

test_expect_success 'a javascript url never reaches an href' '
	cgit_rc modlinkgmrc "url=modlink/tree/" >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
'

# The classes behind cgit's published advisories, each pinned against the
# current code: a newline in a file name splitting the headers, a posted
# length overflowing its buffer, a path climbing out on the dumb transport
# and the about page, a commit with an empty author, a percent sign with no
# digits behind it, script in a file name shown by the diff, and shell
# syntax in a file name handed to a filter.
test_expect_success 'set up the advisory fixtures' '
	mkrepo repos/cve 1 &&
	(
		cd repos/cve &&
		printf "x\n" >"$(printf "crlf\r\nX-Injected: 1.txt")" &&
		printf "x\n" >"<script>alert(1)x.txt" &&
		printf "x\n" >"\$(touch pwned).c" &&
		git add -A &&
		git commit -m names &&
		tree=$(git rev-parse HEAD^{tree}) &&
		printf "tree %s\nparent %s\n" "$tree" "$(git rev-parse HEAD)" >raw &&
		printf "author  <> 1735689600 +0000\ncommitter  <> 1735689600 +0000\n\nempty author\n" >>raw &&
		cid=$(git hash-object --literally -t commit -w raw) &&
		git update-ref refs/heads/noname "$cid"
	) &&
	mkdir -p readme &&
	printf "readme text\n" >readme/README &&
	write_script auth.sh <<-\EOF &&
	case "$1" in
	authenticate-post)
		wc -c >received
		printf "Status: 302 Found\nLocation: /\n\n"
		;;
	esac
	exit 1
	EOF
	cgitrc_repo cverc cve "auth-filter=exec:$PWD/auth.sh" "repo.readme=$PWD/readme/README" "repo.source-filter=exec:$FILTER_DIRECTORY/dump.sh"
'

test_expect_success 'a newline in a file name cannot split the headers' '
	cgit_rc cverc "url=cve/plain/crlf%0d%0aX-Injected:%201.txt" >tmp &&
	grep "^Status: 200" tmp &&
	! grep "^X-Injected" tmp &&
	grep "^Content-Disposition: inline; filename=.crlf\\\\r\\\\nX-Injected: 1.txt.$" tmp
'

# The body has to outgrow the 4096 byte buffer, or an unclamped read would
# still stop at the end of the body and nothing would overflow. A length of
# -1 reaches cgit as the largest number strtoul can return.
test_expect_success 'a posted length past any buffer is clamped' '
	test_seq 2000 >body &&
	for length in 99999999999999 -1
	do
		rm -f received &&
		CGIT_CONFIG="$PWD/cverc" REQUEST_METHOD=POST CONTENT_LENGTH=$length \
			QUERY_STRING="p=login" cgit <body >tmp &&
		test_grep "^Status: 302" tmp &&
		test $(cat received) = 4096 || return 1
	done &&
	CGIT_CONFIG="$PWD/cverc" REQUEST_METHOD=POST CONTENT_LENGTH=10 \
		QUERY_STRING="p=login" cgit <body >tmp &&
	test $(cat received) = 10
'

test_expect_success 'the dumb transport refuses a path that climbs out' '
	cgit_rc cverc "url=cve/objects/../config" >tmp &&
	grep "^Status: 404" tmp &&
	! grep "repositoryformatversion" tmp &&
	cgit_rc cverc "url=cve/objects/info/../../config" >tmp &&
	grep "^Status: 404" tmp &&
	! grep "repositoryformatversion" tmp
'

test_expect_success 'the about page serves nothing above the readme directory' '
	cgit_rc cverc "url=cve/about/../cverc" >tmp &&
	grep "^Status: 404" tmp &&
	! grep "virtual-root" tmp &&
	cgit_rc cverc "url=cve/about/" >tmp &&
	grep "readme text" tmp
'

test_expect_success 'a commit with an empty author renders everywhere' '
	cgit_rc cverc "url=cve/log/&h=noname" >tmp &&
	grep ">empty author</a>" tmp &&
	cgit_rc cverc "url=cve/commit/&h=noname" >tmp &&
	grep "<div class=.commit-subject.>empty author<" tmp &&
	cgit_rc cverc "url=cve/atom/&h=noname" >tmp &&
	grep "<name>unknown</name>" tmp &&
	! grep "<email></email>" tmp
'

test_expect_success 'a percent sign without digits behind it is read as itself' '
	cgit_rc cverc "url=cve/log/&q=%" >tmp &&
	grep "^Status: 200" tmp &&
	cgit_rc cverc "url=cve/log/&q=%zz%2" >tmp &&
	grep "^Status: 200" tmp &&
	grep "value=.%zz%2." tmp
'

test_expect_success 'script in a file name is escaped on the diff page' '
	cgit_rc cverc "url=cve/diff/" >tmp &&
	! grep "<script>alert" tmp &&
	grep "&lt;script&gt;alert(1)x.txt" tmp
'

test_expect_success 'shell syntax in a file name reaches a filter as a plain argument' '
	cgit_ok cverc "url=cve/tree/%24(touch%20pwned).c" >tmp &&
	test_grep -F "\$(touch pwned).c X" tmp &&
	test_path_is_missing pwned &&
	test_path_is_missing repos/cve/pwned
'

# An untrusted response hands over repository bytes as they are. An image
# keeps its own type, SVG among them, so script inside one is held back by
# the policy alone. The cache listing prints keys visitors chose.
test_expect_success 'repository bytes go out under a policy that loads nothing' '
	mkrepo repos/svg 1 &&
	(
		cd repos/svg &&
		echo "<svg xmlns=\"http://www.w3.org/2000/svg\"><script>alert(1)</script></svg>" >x.svg &&
		git add x.svg &&
		git commit -m svg
	) &&
	cgitrc_repo svgrc svg mimetype.svg=image/svg+xml &&
	cgit_ok svgrc "url=svg/plain/x.svg" >tmp &&
	test_grep "^Content-Type: image/svg+xml;" tmp &&
	test_grep "<script>alert(1)</script>" tmp &&
	for page in plain/x.svg blob/x.svg
	do
		cgit_ok svgrc "url=svg/$page" >tmp &&
		test_grep "^X-Content-Type-Options: nosniff$" tmp &&
		test_grep "^Content-Security-Policy: default-src .none.$" tmp || return 1
	done &&
	cgitrc_copy lscacherc &&
	cgit_ok lscacherc "p=ls_cache" >tmp &&
	test_grep "^Content-Type: text/plain;" tmp &&
	test_grep "^X-Content-Type-Options: nosniff$" tmp &&
	test_grep "^Content-Security-Policy: default-src .none.$" tmp
'

test_done