# HG changeset patch
# User Jonas Schäfer <jonas@wielicki.name>
# Date 1641835434 -3600
# Node ID 53e0ae7709171d087c45dcde29413b00302c152f
# Parent  5e68635cdc2c0a3fbe05167c089971f421e7dcb0
util.xml: Do not allow doctypes, comments or processing instructions

Yes. This is as bad as it sounds. CVE pending.

In Prosody itself, this only affects mod_websocket, which uses util.xml
to parse the <open/> frame, thus allowing unauthenticated remote DoS
using Billion Laughs. However, third-party modules using util.xml may
also be affected by this.

This commit installs handlers which disallow the use of doctype
declarations and processing instructions without any escape hatch. It,
by default, also introduces such a handler for comments, however, there
is a way to enable comments nontheless.

This is because util.xml is used to parse human-facing data, where
comments are generally a desirable feature, and also because comments
are generally harmless.

diff -r 5e68635cdc2c -r 53e0ae770917 spec/util_xml_spec.lua
--- a/spec/util_xml_spec.lua	Tue Jan 11 04:15:29 2022 +0100
+++ b/spec/util_xml_spec.lua	Mon Jan 10 18:23:54 2022 +0100
@@ -12,9 +12,41 @@
 	<a:z/> <!-- prefix 'a' is nil here, but should be 'b' -->
 </x>
 ]]
-			local stanza = xml.parse(x);
+			local stanza = xml.parse(x, {allow_comments = true});
 			assert.are.equal(stanza.tags[2].attr.xmlns, "b");
 			assert.are.equal(stanza.tags[2].namespaces["a"], "b");
 		end);
+
+		it("should reject doctypes", function()
+			local x = "<!DOCTYPE foo []><foo/>";
+			local ok = xml.parse(x);
+			assert.falsy(ok);
+		end);
+
+		it("should reject comments by default", function()
+			local x = "<foo><!-- foo --></foo>";
+			local ok = xml.parse(x);
+			assert.falsy(ok);
+		end);
+
+		it("should allow comments if asked nicely", function()
+			local x = "<foo><!-- foo --></foo>";
+			local stanza = xml.parse(x, {allow_comments = true});
+			assert.are.equal(stanza.name, "foo");
+			assert.are.equal(#stanza, 0);
+		end);
+
+		it("should reject processing instructions", function()
+			local x = "<foo><?php die(); ?></foo>";
+			local ok = xml.parse(x);
+			assert.falsy(ok);
+		end);
+
+		it("should allow an xml declaration", function()
+			local x = "<?xml version='1.0'?><foo/>";
+			local stanza = xml.parse(x);
+			assert.truthy(stanza);
+			assert.are.equal(stanza.name, "foo");
+		end);
 	end);
 end);
diff -r 5e68635cdc2c -r 53e0ae770917 util/xml.lua
--- a/util/xml.lua	Tue Jan 11 04:15:29 2022 +0100
+++ b/util/xml.lua	Mon Jan 10 18:23:54 2022 +0100
@@ -3,6 +3,7 @@
 local lxp = require "lxp";
 local t_insert = table.insert;
 local t_remove = table.remove;
+local error = error;
 
 local _ENV = nil;
 -- luacheck: std none
@@ -13,7 +14,7 @@
 	};
 	local ns_separator = "\1";
 	local ns_pattern = "^([^"..ns_separator.."]*)"..ns_separator.."?(.*)$";
-	return function(xml)
+	return function(xml, options)
 		--luacheck: ignore 212/self
 		local handler = {};
 		local stanza = st.stanza("root");
@@ -64,7 +65,27 @@
 		function handler:EndElement()
 			stanza:up();
 		end
-		local parser = lxp.new(handler, ns_separator);
+		local parser;
+		-- SECURITY: These two handlers, especially the Doctype one, are required to prevent exploits such as Billion Laughs.
+		function handler:StartDoctypeDecl()
+			if not parser.stop or not parser:stop() then
+				error("Failed to abort parsing");
+			end
+		end
+		function handler:ProcessingInstruction()
+			if not parser.stop or not parser:stop() then
+				error("Failed to abort parsing");
+			end
+		end
+		if not options or not options.allow_comments then
+			-- NOTE: comments are generally harmless and can be useful when parsing configuration files or other data, even user-provided data
+			function handler:Comment()
+				if not parser.stop or not parser:stop() then
+					error("Failed to abort parsing");
+				end
+			end
+		end
+		parser = lxp.new(handler, ns_separator);
 		local ok, err, line, col = parser:parse(xml);
 		if ok then ok, err, line, col = parser:parse(); end
 		--parser:close();