index : cgit

-- Server-side rendering of a repository's about page, named by the
-- about-filter setting in cgitrc and run inside cgit's embedded Lua
-- interpreter so a readme costs no extra process per request. Markdown, man
-- pages and plain text are the three formats, chosen from the readme's file
-- extension, and anything that fails to parse falls back to escaped plain
-- text. The wrappers it emits carry the classes that assets/cgit.css styles.
-- It runs on Lua 5.1 through 5.4 and LuaJIT.
--
--     about-filter=lua:/usr/local/lib/cgit/filters/about-render.lua

-- Markdown and man pages need lpeg, built for the Lua cgit is linked
-- against. Without it both fall back to escaped plain text.
--
--     # 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
local has_lpeg, lpeg = pcall(require, "lpeg")


-- 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, tree_base, tree_query, chunks = "", "", "", {}


-- Lua patterns backtrack, so "%s+$" costs the square of a run of spaces in
-- the middle of a line. Finding each end directly stays linear.
local function trim(text)
	local first = text:find("%S")
	if first == nil then
		return ""
	end
	local last = #text
	while text:find("^%s", last) do
		last = last - 1
	end
	return text:sub(first, last)
end

-- A heading may close with a run of hashes, which only counts as markup when
-- a space comes before it, so "# C#" keeps its sharp.
local function heading_text(title)
	title = trim(title)
	local last = #title
	while last > 0 and title:sub(last, last) == "#" do
		last = last - 1
	end
	if last == 0 or title:find("^%s", last) then
		return trim(title:sub(1, last))
	end
	return title
end

-- Normalise CRLF and CR to newlines, then split. A trailing newline yields
-- a final empty line, which every caller treats as blank.
local function split_lines(text)
	text = text:gsub("\r\n?", "\n")
	local lines, start = {}, 1
	while true do
		local newline = text:find("\n", start, true)
		if not newline then
			lines[#lines + 1] = text:sub(start)
			return lines
		end
		lines[#lines + 1] = text:sub(start, newline - 1)
		start = newline + 1
	end
end

local function split_cells(row)
	row = trim(row):gsub("^|", ""):gsub("|$", "")
	local cells = {}
	for cell in (row .. "|"):gmatch("(.-)|") do
		cells[#cells + 1] = trim(cell)
	end
	return cells
end

-- Control and whitespace bytes are stripped before the scheme check because
-- a browser ignores them when resolving it, 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

-- A readme percent-encodes what a link path cannot carry, and the path is
-- encoded again on its way out, so it is decoded first. An encoded NUL stays
-- encoded, since cgit's sinks would stop at a real one.
local function unescape_path(path)
	return (path:gsub("%%(%x%x)", function(hex)
		local byte = tonumber(hex, 16)
		if byte > 0 then
			return string.char(byte)
		end
	end))
end

-- The path in the tree that a relative link names. Like GitHub, it resolves
-- against the readme's own directory, or against the root when it starts
-- with a slash, and dot segments fold without climbing above the root.
local function tree_path(path)
	if path:sub(1, 1) ~= "/" then
		path = (filename:match("^(.*/)") or "") .. path
	end
	local segments = {}
	for segment in path:gmatch("[^/]+") do
		if segment == ".." then
			segments[#segments] = nil
		elseif segment ~= "." then
			segments[#segments + 1] = segment
		end
	end
	return table.concat(segments, "/")
end

-- Writes a link target that safe_url passed. A relative one leads into the
-- tree at the ref the readme came from, keeping its fragment but not a
-- query, which the tree page would read as its own. A readme with no tree
-- behind it keeps its links as written.
local function emit_href(url)
	if tree_base == "" or url:find("^#") or url:find("^%a[%w%+%.%-]*:") then
		html_attr(url)
		return
	end
	html_attr(tree_base)
	html_url_path(tree_path(unescape_path(url:match("^[^?#]*"))))
	html_attr(tree_query)
	html_attr(url:match("#.*$") or "")
end


-- The page renders untrusted repository content. Every run of text goes out
-- through cgit's html_txt, every attribute value through html_attr or, for a
-- path in the tree, html_url_path, and every link target through safe_url
-- first. html() only ever gets literal tag scaffolding.

local emit_inline

emit_inline = function(list)
	for _, node in ipairs(list) do
		local kind = node.kind
		if kind == "text" then
			html_txt(node.text)
		elseif kind == "break" then
			html("<br>")
		elseif kind == "code" then
			html("<code>"); html_txt(node.text); html("</code>")
		elseif kind == "strong" then
			html("<strong>"); emit_inline(node.kids); html("</strong>")
		elseif kind == "em" then
			html("<em>"); emit_inline(node.kids); html("</em>")
		elseif kind == "link" then
			local target = safe_url(node.url)
			if target then
				html("<a href='"); emit_href(target); html("'>")
				emit_inline(node.kids)
				html("</a>")
			else
				emit_inline(node.kids)
			end
		end
	end
end

-- Whether a paragraph has anything to show once images are left out, so a
-- row of badges leaves no empty paragraph behind.
local function shows_something(lines)
	for _, line in ipairs(lines) do
		for _, node in ipairs(line) do
			local blank = node.kind == "text" and not node.text:find("%S")
			if node.kind ~= "image" and not blank then
				return true
			end
		end
	end
	return false
end

local emit_blocks

emit_blocks = function(blocks)
	for _, block in ipairs(blocks) do
		local kind = block.kind
		if kind == "heading" then
			html("<h" .. block.level .. ">")
			emit_inline(block.kids)
			html("</h" .. block.level .. ">")
		elseif kind == "hr" then
			html("<hr>")
		elseif kind == "code" then
			html("<pre><code")
			if block.lang and block.lang ~= "" then
				html(" data-lang='")
				html_attr(block.lang)
				html("'")
			end
			html(">")
			html_txt(block.text)
			html("</code></pre>")
		elseif kind == "blockquote" then
			html("<blockquote>")
			emit_blocks(block.blocks)
			html("</blockquote>")
		elseif kind == "table" then
			html("<table><thead><tr>")
			for _, cell in ipairs(block.head) do
				html("<th>")
				emit_inline(cell)
				html("</th>")
			end
			html("</tr></thead><tbody>")
			for _, row in ipairs(block.rows) do
				html("<tr>")
				for _, cell in ipairs(row) do
					html("<td>")
					emit_inline(cell)
					html("</td>")
				end
				html("</tr>")
			end
			html("</tbody></table>")
		elseif kind == "list" then
			local tag = block.ordered and "ol" or "ul"
			html("<" .. tag .. ">")
			for _, item in ipairs(block.items) do
				html("<li>")
				emit_inline(item)
				html("</li>")
			end
			html("</" .. tag .. ">")
		elseif kind == "para" and shows_something(block.lines) then
			html("<p>")
			for i, line in ipairs(block.lines) do
				if i > 1 then
					html("\n")
				end
				emit_inline(line)
			end
			html("</p>")
		end
	end
end


local function render_plaintext(text)
	html("<pre class='plaintext'>")
	html_txt(text)
	html("</pre>")
end


-- Markdown is a deliberate subset rather than CommonMark, leaving out
-- reference links, raw HTML passthrough, nested lists and setext headings,
-- and inline markup does not span the end of a line. Man rendering covers the
-- common macros and drops the rest. Both parse into a node tree before
-- anything is emitted, so a failed parse falls back to plain text without
-- leaving half a page behind.

local render_markdown, render_man

if has_lpeg then
	local P, R, S, C, Ct = lpeg.P, lpeg.R, lpeg.S, lpeg.C, lpeg.Ct

	local space = S(" \t")
	local whitespace = S(" \t\r\n\f\v")
	local eol = P(-1)

	local marker = S("`![*_")
	local url_char = 1 - S(")") - whitespace
	-- Bound the link and image inner scans. Without a cap a readme of
	-- unclosed brackets ("[[[[...") makes every position scan to the
	-- end of the line for a "]" that never comes, which is quadratic.
	local max_scan = 512

	local parse_inline
	local inline_depth = 0

	local function node_text(text)
		return { kind = "text", text = text }
	end
	local function node_code(text)
		return { kind = "code", text = text }
	end
	local function node_strong(text)
		return { kind = "strong", kids = parse_inline(text) }
	end
	local function node_em(text)
		return { kind = "em", kids = parse_inline(text) }
	end
	local function node_link(text, url)
		return { kind = "link", kids = parse_inline(text), url = url }
	end
	local function node_image()
		return { kind = "image" }
	end

	local label = P("[") * C((1 - P("]")) ^ (-max_scan)) * P("]")
	local target = P("(") * space ^ 0 * C(url_char * url_char ^ (-max_scan))
		* (1 - P(")")) ^ (-max_scan) * P(")")

	-- Images are parsed only to be left out, along with the link around a
	-- badge. The about page serves no files for a relative one to load, and
	-- a remote one would have every visitor's browser call a third party.
	local inline_grammar = Ct((
		  (P("`") * C((1 - P("`")) ^ 1) * P("`")) / node_code
		+ (P("[!") * label * target * P("]") * target) / node_image
		+ (P("!") * label * target) / node_image
		+ (label * target) / node_link
		+ (P("**") * C((1 - P("**")) ^ 1) * P("**")) / node_strong
		+ (P("__") * C((1 - P("__")) ^ 1) * P("__")) / node_strong
		+ (P("*") * C((1 - whitespace) * (1 - P("*")) ^ 0) * P("*")) / node_em
		+ (P("_") * C((1 - whitespace) * (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

	local lang_char = R("az", "AZ", "09") + S("_.+#-")
	local heading_line = C(P("#") * P("#") ^ -5) * space ^ 1 * C(P(1) ^ 0)
	local function thematic(mark)
		return space ^ 0 * P(mark) * (space ^ 0 * P(mark)) ^ 2 * space ^ 0 * eol
	end
	local break_line = thematic("-") + thematic("*") + thematic("_")
	local fence_line = space ^ 0 * (C(P("`") ^ 3) + C(P("~") ^ 3)) * space ^ 0 * C(lang_char ^ 0)
	local close_backtick = space ^ 0 * P("`") ^ 3 * space ^ 0 * eol
	local close_tilde = space ^ 0 * P("~") ^ 3 * space ^ 0 * eol
	local quote_line = space ^ 0 * P(">")
	local bullet = S("-*+")
	local number = R("09") ^ 1 * S(".)")
	local item_line = space ^ 0 * C(bullet + number) * space ^ 1 * C(P(1) ^ 0)
	local dash_cell = space ^ 0 * P(":") ^ -1 * P("-") ^ 1 * P(":") ^ -1 * space ^ 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 table_delimiter = space ^ 0 * (
		  pipe * dash_cell * (pipe * dash_cell) ^ 0 * pipe ^ -1
		+ dash_cell * (pipe * dash_cell) ^ 1 * pipe ^ -1
	) * space ^ 0 * eol
	local block_start = space ^ 0 * ((P("#") * P("#") ^ -5 * space) + P(">")
		+ P("`") ^ 3 + P("~") ^ 3 + ((bullet + number) * space))

	local function ordered(mark) return mark:match("%d") ~= nil end

	local function table_starts(lines, i)
		return i < #lines and lines[i]:find("|", 1, true) ~= nil
			and table_delimiter:match(lines[i + 1]) ~= nil
	end

	-- A blank line ends a paragraph, and so does the start of any other
	-- block. That includes a table, which GitHub's own parser lets cut a
	-- paragraph short.
	local function ends_paragraph(lines, i)
		return lines[i]:match("^%s*$") or block_start:match(lines[i])
			or table_starts(lines, i)
	end

	-- A paragraph's lines run together the way GitHub shows them, unless one
	-- ends in two spaces or a backslash, which breaks the line there. The
	-- last line of a paragraph has nothing to break before.
	local function paragraph_lines(raw)
		local lines = {}
		for i, line in ipairs(raw) do
			local hard = false
			if i < #raw then
				if line:sub(-1) == "\\" then
					line, hard = line:sub(1, -2), true
				else
					hard = line:sub(-2) == "  "
				end
			end
			lines[i] = parse_inline(trim(line))
			if hard then
				lines[i][#lines[i] + 1] = { kind = "break" }
			end
		end
		return lines
	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 = fence_line:match(line)
			local hashes, title = heading_line:match(line)
			local mark = item_line:match(line)
			if line:match("^%s*$") then
				i = i + 1
			elseif fence then
				local close = (fence:sub(1, 1) == "`") and close_backtick or close_tilde
				local code = {}
				i = i + 1
				while i <= n and not close:match(lines[i]) do
					code[#code + 1] = lines[i]
					i = i + 1
				end
				i = i + 1
				local body = table.concat(code, "\n")
				if #code > 0 then
					body = body .. "\n"
				end
				blocks[#blocks + 1] = {
					kind = "code",
					lang = lang,
					text = body,
				}
			elseif hashes then
				blocks[#blocks + 1] = {
					kind = "heading",
					level = #hashes,
					kids = parse_inline(heading_text(title)),
				}
				i = i + 1
			elseif break_line:match(line) then
				blocks[#blocks + 1] = { kind = "hr" }
				i = i + 1
			elseif quote_line:match(line) then
				local inner = {}
				while i <= n and quote_line:match(lines[i]) do
					inner[#inner + 1] = (lines[i]:gsub("^%s*>%s?", ""))
					i = i + 1
				end
				if depth < max_depth then
					blocks[#blocks + 1] = {
						kind = "blockquote",
						blocks = parse_blocks(inner, depth + 1),
					}
				else
					blocks[#blocks + 1] = {
						kind = "para",
						lines = paragraph_lines(inner),
					}
				end
			elseif table_starts(lines, i) then
				local head = {}
				for _, cell in ipairs(split_cells(line)) do
					head[#head + 1] = parse_inline(cell)
				end
				local rows = {}
				i = i + 2
				while i <= n and lines[i]:find("|", 1, true) and not lines[i]:match("^%s*$") do
					local row = {}
					for _, cell in ipairs(split_cells(lines[i])) do
						row[#row + 1] = parse_inline(cell)
					end
					rows[#rows + 1] = row
					i = i + 1
				end
				blocks[#blocks + 1] = {
					kind = "table",
					head = head,
					rows = rows,
				}
			elseif mark then
				local is_ordered = ordered(mark)
				local items = {}
				while i <= n do
					local item_mark, content = item_line:match(lines[i])
					if not item_mark or ordered(item_mark) ~= is_ordered then
						break
					end
					items[#items + 1] = parse_inline(content)
					i = i + 1
				end
				blocks[#blocks + 1] = {
					kind = "list",
					ordered = is_ordered,
					items = items,
				}
			else
				local paragraph = { line }
				i = i + 1
				while i <= n and not ends_paragraph(lines, i) do
					paragraph[#paragraph + 1] = lines[i]
					i = i + 1
				end
				blocks[#blocks + 1] = {
					kind = "para",
					lines = paragraph_lines(paragraph),
				}
			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'>")
		-- Emitting is wrapped so a bug there can at worst
		-- truncate the page, never raise and let the outer
		-- fallback duplicate what has already been sent.
		pcall(emit_blocks, blocks)
		html("</div>")
	end

	-- Roff fonts are a current state carried across a run rather than
	-- opened and closed, so the inline pass tokenises to a flat list
	-- and folds it into the nodes markdown emits, instead of nesting.
	local macro_line = S(".'") * space ^ 0 * C((1 - space) ^ 1) * space ^ 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 backslash = P("\\")
	local function font_token(name) return { kind = "font", font = name } end
	local function text_token(text) return { kind = "text", text = text } end
	local skip_token = { kind = "skip" }

	local man_inline_grammar = Ct((
		  (backslash * P("f") * P("(") * C(P(1) * P(1)))
			/ function(name) return font_token(two_font[name] or "R") end
		+ (backslash * P("f") * P("[") * C((1 - P("]")) ^ 0) * P("]"))
			/ function(name) return font_token(one_font[name] or "R") end
		+ (backslash * P("f") * C(P(1)))
			/ function(name) return font_token(one_font[name] or "R") end
		+ (backslash * P("-")) / function() return text_token("-") end
		+ (backslash * P("e")) / function() return text_token("\\") end
		+ (backslash * P(" ")) / function() return text_token(" ") end
		+ (backslash * S("&|^")) / function() return skip_token end
		+ (backslash * P("(") * C(P(1) * P(1)))
			/ function(name) return text_token(man_chars[name] or "") end
		+ (backslash * P("*") * (P("(") * P(1) * P(1) + P(1)))
			/ function() return skip_token end
		+ (backslash * C(P(1))) / text_token
		+ backslash / function() return skip_token end
		+ C((1 - backslash) ^ 1) / text_token
	) ^ 0)

	local function man_inline(text)
		local tokens = man_inline_grammar:match(text) or {}
		local nodes, parts, font = {}, {}, "R"
		local function flush()
			if #parts == 0 then
				return
			end
			local run = table.concat(parts)
			parts = {}
			if run == "" then
				return
			end
			if font == "B" then
				nodes[#nodes + 1] = {
					kind = "strong",
					kids = { { kind = "text", text = run } },
				}
			elseif font == "I" then
				nodes[#nodes + 1] = {
					kind = "em",
					kids = { { kind = "text", text = run } },
				}
			else
				nodes[#nodes + 1] = { kind = "text", text = run }
			end
		end
		for _, token in ipairs(tokens) do
			if token.kind == "text" then
				parts[#parts + 1] = token.text
			elseif token.kind == "font" then
				flush()
				font = token.font
			end
		end
		flush()
		return nodes
	end

	local function parse_man(text)
		local lines = split_lines(text)
		local blocks = {}
		local paragraph, nofill_lines, nofill = nil, nil, false
		local function flush_paragraph()
			if paragraph and #paragraph > 0 then
				blocks[#blocks + 1] = { kind = "para", lines = paragraph }
			end
			paragraph = nil
		end
		local function flush_nofill()
			if nofill_lines then
				local body = table.concat(nofill_lines, "\n")
				if #nofill_lines > 0 then
					body = body .. "\n"
				end
				blocks[#blocks + 1] = { kind = "code", text = body }
				nofill_lines = nil
			end
		end
		local function add_line(nodes)
			paragraph = paragraph or {}
			paragraph[#paragraph + 1] = nodes
		end
		for _, line in ipairs(lines) do
			local macro, rest = macro_line:match(line)
			if rest then
				rest = (rest:gsub('^"(.*)"$', "%1"))
			end
			if nofill then
				if macro == "fi" then
					nofill = false
					flush_nofill()
				else
					nofill_lines[#nofill_lines + 1] = line
				end
			elseif line:match("^%s*$") then
				flush_paragraph()
			elseif macro and macro:sub(1, 1) == "\\" then
				-- A roff comment, dropped.
			elseif macro == "SH" or macro == "SS" then
				flush_paragraph()
				blocks[#blocks + 1] = {
					kind = "heading",
					level = (macro == "SH") and 2 or 3,
					kids = man_inline(rest),
				}
			elseif macro == "nf" then
				flush_paragraph()
				nofill = true
				nofill_lines = {}
			elseif macro == "fi" then
				flush_nofill()
			elseif macro == "B" or macro == "BR" or macro == "RB" or macro == "BI" or macro == "IB" then
				add_line({ { kind = "strong", kids = man_inline(rest) } })
			elseif macro == "I" or macro == "IR" or macro == "RI" then
				add_line({ { kind = "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_paragraph()
			elseif macro then
				if rest ~= "" then
					add_line(man_inline(rest))
				end
			else
				add_line(man_inline(line))
			end
		end
		flush_paragraph()
		flush_nofill()
		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


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

-- A readme tracked in the repository comes with the url of the tree page at
-- its ref and the query that follows a path there, both escaped for a url.
function filter_open(name, base, query)
	filename = name or ""
	tree_base = base or ""
	tree_query = query or ""
	chunks = {}
end

function filter_write(str)
	chunks[#chunks + 1] = str
end

-- cgit takes filter output through a C string sink that stops at the first
-- NUL byte, so a write holding one loses everything from the NUL on.
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