Replace the browser markdown renderer with a filter

The readme is now escaped plain text unless `about-filter` points at
the new `about-render.lua`, which renders markdown, man pages and plain
text server-side. `enable-markdown` goes away with the renderer.

1 file changed, 592 insertions, 0 deletions

diff --git a/extensions/about-render.lua b/extensions/about-render.lua
new file mode 100644
index 0000000..073b531
--- /dev/null
+++ b/extensions/about-render.lua
@@ -0,0 +1,592 @@
+-- Server-side about-page rendering for cgit, used with the about-filter setting
+-- in cgitrc and the lua: prefix so it runs in cgit's embedded interpreter with
+-- no per-request process.
+--
+--     about-filter=lua:/usr/lib/cgit/filters/about-render.lua
+--
+-- One filter renders the three baseline about formats, chosen by the readme's
+-- file extension.
+--
+--     markdown   .md .markdown .mkd .mdown
+--     man page   .1 through .9 and .man
+--     plain text everything else, and the fallback for anything that fails
+--
+-- Add a format by giving it a render function and a row in the handler table
+-- near the foot of this file. Nothing else needs to change.
+--
+-- SUPPORTED LUA
+--
+-- Lua 5.1 through 5.4 and LuaJIT.
+--
+-- REQUIREMENTS
+--
+-- lpeg, the parsing module, for the Lua cgit is linked against. Markdown and
+-- man pages are parsed with lpeg grammars, so without lpeg both fall back to
+-- escaped plain text. It ships with cgit's syntax highlighter too, so a cgit
+-- that colours source already has it.
+--
+--     # Debian and Ubuntu
+--     sudo apt install lua-lpeg
+--     # Fedora
+--     sudo dnf install lua-lpeg
+--     # Alpine
+--     sudo apk add lua5.1-lpeg
+--     # or with LuaRocks, matched to your Lua version
+--     sudo luarocks --lua-version 5.1 install lpeg
+--
+-- Plain text needs only Lua itself.
+--
+-- SECURITY
+--
+-- The about page renders untrusted repository content, so the safety rule is
+-- that every run of text reaches the page through cgit's own html_txt, every
+-- attribute value through html_attr, and every link or image target through
+-- the safe_url scheme allowlist below before html_attr. Only http, https,
+-- mailto and relative targets survive, so a hostile readme cannot inject markup
+-- or a javascript: url. Because all output is routed through those sinks by
+-- construction, a bug in the parser can only mis-render, never inject.
+--
+-- LIMITATIONS
+--
+-- Markdown is a deliberate subset, not CommonMark. Headings, thematic breaks,
+-- fenced and inline code, blockquotes, pipe tables, single-level lists, links,
+-- images and emphasis. No reference links, no raw HTML passthrough, no nested
+-- lists, no setext headings. Emphasis does not span a hard line break inside a
+-- paragraph. Man rendering covers the common macros (section headings, filled
+-- and no-fill paragraphs, bold and italic, font escapes) and drops the rest.
+-- cgit sends filter output through a C string sink that stops at the first NUL
+-- byte, so content past a NUL is truncated.
+--
+-- OUTPUT
+--
+-- Markdown is wrapped in <div class='markdown'>, man pages in
+-- <div class='markdown manpage'>, plain text in <pre class='plaintext'>, all
+-- styled by assets/cgit.css.
+
+local ok_lpeg, lpeg = pcall(require, "lpeg")
+
+
+-- ===== configuration ===================================================
+
+-- Readmes larger than this are served as escaped plain text rather than parsed.
+local max_bytes = 512 * 1024
+
+-- Ceiling on blockquote and emphasis nesting, so a hostile readme of stacked
+-- markers cannot drive the recursive parser into a stack overflow.
+local max_depth = 24
+
+local filename, chunks = "", {}
+
+
+-- ===== helpers =========================================================
+
+local function trim(s)
+	return (s:gsub("^%s+", ""):gsub("%s+$", ""))
+end
+
+-- Split on newlines after normalising CRLF and CR, returning the lines without
+-- their terminators. A trailing newline yields a final empty line, which every
+-- caller treats as blank.
+local function split_lines(s)
+	s = s:gsub("\r\n?", "\n")
+	local lines, start = {}, 1
+	while true do
+		local nl = s:find("\n", start, true)
+		if not nl then
+			lines[#lines + 1] = s:sub(start)
+			return lines
+		end
+		lines[#lines + 1] = s:sub(start, nl - 1)
+		start = nl + 1
+	end
+end
+
+-- Split a table row into trimmed cells, dropping one optional leading and
+-- trailing pipe.
+local function split_cells(row)
+	row = trim(row):gsub("^|", ""):gsub("|$", "")
+	local cells = {}
+	for c in (row .. "|"):gmatch("(.-)|") do
+		cells[#cells + 1] = trim(c)
+	end
+	return cells
+end
+
+-- Return the url if its scheme is safe, else nil. Control and whitespace bytes
+-- are stripped anywhere first because a browser ignores them when resolving the
+-- scheme, so "java\nscript:" must still be caught as javascript. The class is
+-- %c%s rather than a literal 0x00 range so it is safe on Lua 5.1, where an
+-- embedded zero byte ends a pattern.
+local function safe_url(url)
+	url = url:gsub("[%c%s]", "")
+	-- A leading "//" or "/\" is scheme-relative, which a browser resolves to
+	-- another origin, so it is not the local target the parser meant to allow.
+	if url:sub(1, 2) == "//" or url:sub(1, 2) == "/\\" then
+		return nil
+	end
+	local scheme = url:match("^(%a[%w%+%.%-]*):")
+	if scheme then
+		scheme = scheme:lower()
+		if scheme ~= "http" and scheme ~= "https" and scheme ~= "mailto" then
+			return nil
+		end
+	end
+	return url
+end
+
+
+-- ===== emit ============================================================
+-- html, html_txt and html_attr are injected by cgit's lua filter host. Tag
+-- scaffolding is a literal argument to html; every value that came from the
+-- repository goes through html_txt, html_attr or safe_url.
+
+local emit_inline
+
+emit_inline = function(list)
+	for _, nd in ipairs(list) do
+		local t = nd.t
+		if t == "text" then
+			html_txt(nd.v)
+		elseif t == "code" then
+			html("<code>"); html_txt(nd.v); html("</code>")
+		elseif t == "strong" then
+			html("<strong>"); emit_inline(nd.kids); html("</strong>")
+		elseif t == "em" then
+			html("<em>"); emit_inline(nd.kids); html("</em>")
+		elseif t == "link" then
+			local u = safe_url(nd.url)
+			if u then
+				html("<a href='"); html_attr(u); html("'>")
+				emit_inline(nd.kids)
+				html("</a>")
+			else
+				emit_inline(nd.kids)
+			end
+		elseif t == "image" then
+			local u = safe_url(nd.url)
+			if u then
+				html("<img src='"); html_attr(u)
+				html("' alt='"); html_attr(nd.alt); html("'/>")
+			else
+				html_txt("![" .. nd.alt .. "]")
+			end
+		end
+	end
+end
+
+local emit_blocks
+
+emit_blocks = function(blocks)
+	for _, b in ipairs(blocks) do
+		local t = b.t
+		if t == "heading" then
+			html("<h" .. b.level .. ">")
+			emit_inline(b.kids)
+			html("</h" .. b.level .. ">")
+		elseif t == "hr" then
+			html("<hr/>")
+		elseif t == "code" then
+			html("<pre><code")
+			if b.lang and b.lang ~= "" then
+				html(" data-lang='"); html_attr(b.lang); html("'")
+			end
+			html(">"); html_txt(b.text); html("</code></pre>")
+		elseif t == "blockquote" then
+			html("<blockquote>"); emit_blocks(b.blocks); html("</blockquote>")
+		elseif t == "table" then
+			html("<table><thead><tr>")
+			for _, c in ipairs(b.head) do
+				html("<th>"); emit_inline(c); html("</th>")
+			end
+			html("</tr></thead><tbody>")
+			for _, row in ipairs(b.rows) do
+				html("<tr>")
+				for _, c in ipairs(row) do
+					html("<td>"); emit_inline(c); html("</td>")
+				end
+				html("</tr>")
+			end
+			html("</tbody></table>")
+		elseif t == "list" then
+			local tag = b.ordered and "ol" or "ul"
+			html("<" .. tag .. ">")
+			for _, item in ipairs(b.items) do
+				html("<li>"); emit_inline(item); html("</li>")
+			end
+			html("</" .. tag .. ">")
+		elseif t == "para" then
+			html("<p>")
+			for k, line in ipairs(b.lines) do
+				if k > 1 then html("<br/>") end
+				emit_inline(line)
+			end
+			html("</p>")
+		end
+	end
+end
+
+
+-- ===== plain text ======================================================
+
+local function render_plaintext(text)
+	html("<pre class='plaintext'>")
+	html_txt(text)
+	html("</pre>")
+end
+
+
+-- ===== markdown and man ================================================
+-- Both are parsed with lpeg into the block/inline node trees emit_blocks and
+-- emit_inline above consume. Parsing is pure and builds a tree; emit is the
+-- only step that writes output, so a parse failure falls back to plain text
+-- without leaving half a page behind.
+
+local render_markdown, render_man
+
+if ok_lpeg then
+	local P, R, S, C, Ct = lpeg.P, lpeg.R, lpeg.S, lpeg.C, lpeg.Ct
+
+	local sp = S(" \t")
+	local ws = S(" \t\r\n\f\v")
+	local eol = P(-1)
+
+	-- ----- markdown inline -----
+	local marker = S("`![*_")
+	local urlchar = 1 - S(")") - ws
+	-- Bound the link and image inner scans. Without a cap a readme of unclosed
+	-- brackets ("[[[[...") makes each position scan to end of line for a "]"
+	-- that never comes, which is quadratic. Real link text and urls sit far
+	-- under this, and anything longer simply renders as plain text.
+	local CAP = 512
+
+	local parse_inline
+	local inline_depth = 0
+
+	local function node_text(s) return { t = "text", v = s } end
+	local function node_code(s) return { t = "code", v = s } end
+	local function node_strong(s) return { t = "strong", kids = parse_inline(s) } end
+	local function node_em(s) return { t = "em", kids = parse_inline(s) } end
+	local function node_link(text, url) return { t = "link", kids = parse_inline(text), url = url } end
+	local function node_image(alt, url) return { t = "image", alt = alt, url = url } end
+
+	local inline_grammar = Ct((
+		  (P("`") * C((1 - P("`")) ^ 1) * P("`")) / node_code
+		+ (P("![") * C((1 - P("]")) ^ (-CAP)) * P("]") * P("(") * sp ^ 0
+			* C(urlchar * urlchar ^ (-CAP)) * (1 - P(")")) ^ (-CAP) * P(")")) / node_image
+		+ (P("[") * C((1 - P("]")) ^ (-CAP)) * P("]") * P("(") * sp ^ 0
+			* C(urlchar * urlchar ^ (-CAP)) * (1 - P(")")) ^ (-CAP) * P(")")) / node_link
+		+ (P("**") * C((1 - P("**")) ^ 1) * P("**")) / node_strong
+		+ (P("__") * C((1 - P("__")) ^ 1) * P("__")) / node_strong
+		+ (P("*") * C((1 - ws) * (1 - P("*")) ^ 0) * P("*")) / node_em
+		+ (P("_") * C((1 - ws) * (1 - P("_")) ^ 0) * P("_")) / node_em
+		+ C((1 - marker) ^ 1) / node_text
+		+ C(P(1)) / node_text
+	) ^ 0)
+
+	parse_inline = function(s)
+		if inline_depth >= max_depth then
+			return { node_text(s) }
+		end
+		inline_depth = inline_depth + 1
+		local nodes = inline_grammar:match(s) or { node_text(s) }
+		inline_depth = inline_depth - 1
+		return nodes
+	end
+
+	-- ----- markdown blocks -----
+	-- Line matchers. Each is anchored at the start of a single line and returns
+	-- its captures, or nil when the line is not of that kind.
+	local langchar = R("az", "AZ", "09") + S("_.+#-")
+	local m_heading = C(P("#") * P("#") ^ -5) * sp ^ 1 * C(P(1) ^ 0)
+	local function rule(c) return sp ^ 0 * P(c) * (sp ^ 0 * P(c)) ^ 2 * sp ^ 0 * eol end
+	local m_hr = rule("-") + rule("*") + rule("_")
+	local m_fence = sp ^ 0 * (C(P("`") ^ 3) + C(P("~") ^ 3)) * sp ^ 0 * C(langchar ^ 0)
+	local m_close_bt = sp ^ 0 * P("`") ^ 3 * sp ^ 0 * eol
+	local m_close_ti = sp ^ 0 * P("~") ^ 3 * sp ^ 0 * eol
+	local m_bq = sp ^ 0 * P(">")
+	local bullet = S("-*+")
+	local number = R("09") ^ 1 * S(".)")
+	local m_item = sp ^ 0 * C(bullet + number) * sp ^ 1 * C(P(1) ^ 0)
+	local dcell = sp ^ 0 * P(":") ^ -1 * P("-") ^ 1 * P(":") ^ -1 * sp ^ 0
+	-- A delimiter row is dash-cells joined by pipes. Require at least one pipe,
+	-- a leading one or one between cells, so a bare rule of dashes stays a
+	-- thematic break and an ordinary paragraph line is never taken for a table.
+	local pipe = P("|")
+	local m_tdelim = sp ^ 0 * (
+		  pipe * dcell * (pipe * dcell) ^ 0 * pipe ^ -1
+		+ dcell * (pipe * dcell) ^ 1 * pipe ^ -1
+	) * sp ^ 0 * eol
+	local m_blockstart = sp ^ 0 * ((P("#") * P("#") ^ -5 * sp) + P(">")
+		+ P("`") ^ 3 + P("~") ^ 3 + ((bullet + number) * sp))
+
+	local function ordered(mark) return mark:match("%d") ~= nil end
+
+	local parse_blocks
+
+	parse_blocks = function(lines, depth)
+		local blocks, i, n = {}, 1, #lines
+		while i <= n do
+			local line = lines[i]
+			local fence, lang = m_fence:match(line)
+			local hashes, htext = m_heading:match(line)
+			local mark = m_item:match(line)
+			if line:match("^%s*$") then
+				i = i + 1
+			elseif fence then
+				local close = (fence:sub(1, 1) == "`") and m_close_bt or m_close_ti
+				local code = {}
+				i = i + 1
+				while i <= n and not close:match(lines[i]) do
+					code[#code + 1] = lines[i]; i = i + 1
+				end
+				i = i + 1
+				local body = table.concat(code, "\n")
+				if #code > 0 then body = body .. "\n" end
+				blocks[#blocks + 1] = { t = "code", lang = lang, text = body }
+			elseif hashes then
+				blocks[#blocks + 1] = {
+					t = "heading",
+					level = #hashes,
+					kids = parse_inline((htext:gsub("%s*#*%s*$", ""))),
+				}
+				i = i + 1
+			elseif m_hr:match(line) then
+				blocks[#blocks + 1] = { t = "hr" }
+				i = i + 1
+			elseif m_bq:match(line) then
+				local inner = {}
+				while i <= n and m_bq:match(lines[i]) do
+					inner[#inner + 1] = (lines[i]:gsub("^%s*>%s?", ""))
+					i = i + 1
+				end
+				if depth < max_depth then
+					blocks[#blocks + 1] = { t = "blockquote", blocks = parse_blocks(inner, depth + 1) }
+				else
+					local para = {}
+					for _, l in ipairs(inner) do para[#para + 1] = parse_inline(l) end
+					blocks[#blocks + 1] = { t = "para", lines = para }
+				end
+			elseif i + 1 <= n and line:find("|", 1, true) and m_tdelim:match(lines[i + 1]) then
+				local head = {}
+				for _, c in ipairs(split_cells(line)) do head[#head + 1] = parse_inline(c) end
+				local rows = {}
+				i = i + 2
+				while i <= n and lines[i]:find("|", 1, true) and not lines[i]:match("^%s*$") do
+					local row = {}
+					for _, c in ipairs(split_cells(lines[i])) do row[#row + 1] = parse_inline(c) end
+					rows[#rows + 1] = row
+					i = i + 1
+				end
+				blocks[#blocks + 1] = { t = "table", head = head, rows = rows }
+			elseif mark then
+				local is_ol = ordered(mark)
+				local items = {}
+				while i <= n do
+					local mk, ct = m_item:match(lines[i])
+					if not mk or ordered(mk) ~= is_ol then break end
+					items[#items + 1] = parse_inline(ct)
+					i = i + 1
+				end
+				blocks[#blocks + 1] = { t = "list", ordered = is_ol, items = items }
+			else
+				local para = { parse_inline(line) }
+				i = i + 1
+				while i <= n and not lines[i]:match("^%s*$") and not m_blockstart:match(lines[i]) do
+					para[#para + 1] = parse_inline(lines[i]); i = i + 1
+				end
+				blocks[#blocks + 1] = { t = "para", lines = para }
+			end
+		end
+		return blocks
+	end
+
+	render_markdown = function(text)
+		if #text > max_bytes then
+			return render_plaintext(text)
+		end
+		local ok, blocks = pcall(parse_blocks, split_lines(text), 0)
+		if not ok then
+			return render_plaintext(text)
+		end
+		html("<div class='markdown'>")
+		-- Emit is wrapped so a parser bug can at worst truncate the page, never
+		-- error out and let the outer fallback duplicate what was already sent.
+		pcall(emit_blocks, blocks)
+		html("</div>")
+	end
+
+	-- ----- man pages -----
+	-- Macro lines are matched with lpeg, and the roff inline font and character
+	-- escapes are an lpeg grammar producing a flat token list that a fold turns
+	-- into the same inline nodes markdown emits. Bold and italic are the current
+	-- font, which carries across a run, so the inline pass tokenises and folds
+	-- rather than nesting the way markdown does.
+	local m_macro = S(".'") * sp ^ 0 * C((1 - sp) ^ 1) * sp ^ 0 * C(P(1) ^ 0)
+
+	local one_font = { B = "B", I = "I" }
+	local two_font = { CB = "B", BI = "B", CI = "I" }
+	local man_chars = {
+		aq = "'", cq = "'", oq = "'", dq = '"', lq = '"', rq = '"',
+		hy = "-", en = "-", em = "-",
+	}
+	local esc = P("\\")
+	local function font_tok(f) return { k = "font", f = f } end
+	local function text_tok(v) return { k = "text", v = v } end
+	local skip_tok = { k = "skip" }
+
+	local man_inline_grammar = Ct((
+		  (esc * P("f") * P("(") * C(P(1) * P(1))) / function(nm) return font_tok(two_font[nm] or "R") end
+		+ (esc * P("f") * P("[") * C((1 - P("]")) ^ 0) * P("]")) / function(nm) return font_tok(one_font[nm] or "R") end
+		+ (esc * P("f") * C(P(1))) / function(f) return font_tok(one_font[f] or "R") end
+		+ (esc * P("-")) / function() return text_tok("-") end
+		+ (esc * P("e")) / function() return text_tok("\\") end
+		+ (esc * P(" ")) / function() return text_tok(" ") end
+		+ (esc * S("&|^")) / function() return skip_tok end
+		+ (esc * P("(") * C(P(1) * P(1))) / function(nm) return text_tok(man_chars[nm] or "") end
+		+ (esc * P("*") * (P("(") * P(1) * P(1) + P(1))) / function() return skip_tok end
+		+ (esc * C(P(1))) / text_tok
+		+ esc / function() return skip_tok end
+		+ C((1 - esc) ^ 1) / text_tok
+	) ^ 0)
+
+	local function man_inline(s)
+		local toks = man_inline_grammar:match(s) or {}
+		local nodes, buf, font = {}, {}, "R"
+		local function flush()
+			if #buf == 0 then return end
+			local txt = table.concat(buf); buf = {}
+			if txt == "" then return end
+			if font == "B" then
+				nodes[#nodes + 1] = { t = "strong", kids = { { t = "text", v = txt } } }
+			elseif font == "I" then
+				nodes[#nodes + 1] = { t = "em", kids = { { t = "text", v = txt } } }
+			else
+				nodes[#nodes + 1] = { t = "text", v = txt }
+			end
+		end
+		for _, tk in ipairs(toks) do
+			if tk.k == "text" then
+				buf[#buf + 1] = tk.v
+			elseif tk.k == "font" then
+				flush(); font = tk.f
+			end
+		end
+		flush()
+		return nodes
+	end
+
+	local function parse_man(text)
+		local lines = split_lines(text)
+		local blocks = {}
+		local para, pre, nofill = nil, nil, false
+		local function flush_para()
+			if para and #para > 0 then
+				blocks[#blocks + 1] = { t = "para", lines = para }
+			end
+			para = nil
+		end
+		local function flush_pre()
+			if pre then
+				local body = table.concat(pre, "\n")
+				if #pre > 0 then body = body .. "\n" end
+				blocks[#blocks + 1] = { t = "code", text = body }
+				pre = nil
+			end
+		end
+		local function add_line(nodes)
+			para = para or {}
+			para[#para + 1] = nodes
+		end
+		for _, line in ipairs(lines) do
+			local macro, rest = m_macro:match(line)
+			if rest then rest = (rest:gsub('^"(.*)"$', "%1")) end
+			if nofill then
+				if macro == "fi" then
+					nofill = false; flush_pre()
+				else
+					pre[#pre + 1] = line
+				end
+			elseif line:match("^%s*$") then
+				flush_para()
+			elseif macro and macro:sub(1, 1) == "\\" then
+				-- roff comment, dropped
+			elseif macro == "SH" or macro == "SS" then
+				flush_para()
+				blocks[#blocks + 1] = {
+					t = "heading",
+					level = (macro == "SH") and 2 or 3,
+					kids = man_inline(rest),
+				}
+			elseif macro == "nf" then
+				flush_para(); nofill = true; pre = {}
+			elseif macro == "fi" then
+				flush_pre()
+			elseif macro == "B" or macro == "BR" or macro == "RB"
+			    or macro == "BI" or macro == "IB" then
+				add_line({ { t = "strong", kids = man_inline(rest) } })
+			elseif macro == "I" or macro == "IR" or macro == "RI" then
+				add_line({ { t = "em", kids = man_inline(rest) } })
+			elseif macro == "TH" or macro == "PP" or macro == "LP"
+			    or macro == "P" or macro == "TP" or macro == "IP"
+			    or macro == "HP" or macro == "RS" or macro == "RE"
+			    or macro == "sp" or macro == "br" then
+				flush_para()
+			elseif macro then
+				if rest ~= "" then add_line(man_inline(rest)) end
+			else
+				add_line(man_inline(line))
+			end
+		end
+		flush_para(); flush_pre()
+		return blocks
+	end
+
+	render_man = function(text)
+		if #text > max_bytes then
+			return render_plaintext(text)
+		end
+		local ok, blocks = pcall(parse_man, text)
+		if not ok then
+			return render_plaintext(text)
+		end
+		html("<div class='markdown manpage'>")
+		pcall(emit_blocks, blocks)
+		html("</div>")
+	end
+else
+	-- No lpeg, so markdown and man readmes read as escaped source, not nothing.
+	render_markdown = render_plaintext
+	render_man = render_plaintext
+end
+
+
+-- ===== dispatch ========================================================
+
+local handlers = {}
+for _, ext in ipairs({ "md", "markdown", "mkd", "mdown" }) do
+	handlers[ext] = render_markdown
+end
+for _, ext in ipairs({ "1", "2", "3", "4", "5", "6", "7", "8", "9", "man" }) do
+	handlers[ext] = render_man
+end
+
+function filter_open(name)
+	filename = name or ""
+	chunks = {}
+end
+
+function filter_write(str)
+	chunks[#chunks + 1] = str
+end
+
+function filter_close()
+	local text = table.concat(chunks)
+	chunks = {}
+	local ext = (filename:match("%.([^.]+)$") or ""):lower()
+	local render = handlers[ext] or render_plaintext
+	local ok = pcall(render, text)
+	if not ok then
+		render_plaintext(text)
+	end
+	return 0
+end