-- 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 5.5 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 = "", "", "", {}

-- How often each heading anchor has been handed out. The ids already on the
-- page around the readme start at one, so no heading takes them.
local id_counts = {}


-- 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

-- A pipe escaped with a backslash belongs to its cell, even inside a code
-- span, so as on GitHub the backslash goes and the pipe stays. An even run of
-- backslashes escapes only itself, so the pipe after it still divides.
local function split_cells(row)
	row = trim(row)
	local cells, parts, start = {}, {}, 1
	if row:sub(1, 1) == "|" then
		start = 2
	end
	while true do
		local bar = row:find("|", start, true)
		if not bar then
			break
		end
		local slashes = 0
		while row:sub(bar - slashes - 1, bar - slashes - 1) == "\\" do
			slashes = slashes + 1
		end
		if slashes % 2 == 1 then
			parts[#parts + 1] = row:sub(start, bar - 2) .. "|"
		else
			parts[#parts + 1] = row:sub(start, bar - 1)
			cells[#cells + 1] = trim(table.concat(parts))
			parts = {}
		end
		start = bar + 1
	end
	if start <= #row or #parts > 0 then
		parts[#parts + 1] = row:sub(start)
		cells[#cells + 1] = trim(table.concat(parts))
	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 "//" is scheme-relative, which a browser resolves to
	-- another origin, and it reads a backslash there as a slash, so no
	-- pair of the two is the local target the parser meant to allow.
	local head = url:sub(1, 2):gsub("\\", "/")
	if head == "//" 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 and the character
-- references whose shape the markdown grammar checked.

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 == "reference" then
			html(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 function plain_text(nodes, parts)
	for _, node in ipairs(nodes) do
		if node.kind == "text" or node.kind == "code" then
			parts[#parts + 1] = node.text
		elseif node.kids then
			plain_text(node.kids, parts)
		end
	end
	return parts
end

-- Gives a heading the anchor GitHub would, so a readme's links to its own
-- sections work. That is its text in lower case, with punctuation dropped
-- and each space a hyphen, and a count added to a repeat. Bytes past ASCII
-- stay, keeping the letters of other scripts.
local function emit_id(nodes)
	local base = table.concat(plain_text(nodes, {})):lower()
		:gsub("[^%w_%- \128-\255]", ""):gsub(" ", "-")
	if base == "" then
		return
	end
	local count = id_counts[base] or 0
	id_counts[base] = count + 1
	html(" id='")
	html_attr(count > 0 and base .. "-" .. count or base)
	html("'")
end

local function emit_lines(lines)
	for i, line in ipairs(lines) do
		if i > 1 then
			html("\n")
		end
		emit_inline(line)
	end
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_id(block.kids)
			html(">")
			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_lines(item)
				html("</li>")
			end
			html("</" .. tag .. ">")
		elseif kind == "para" and shows_something(block.lines) then
			html("<p>")
			emit_lines(block.lines)
			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 and nested lists, and inline markup
-- does not span the end of a line. Man rendering covers the man(7) macros and
-- drops the roff requests and definitions that generators put around them.
-- 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, B, C, Ct = lpeg.P, lpeg.R, lpeg.S, lpeg.B, lpeg.C, lpeg.Ct

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

	local marker = S("`![*_<\\&")
	local punctuation = 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 function node_reference(text)
		return { kind = "reference", text = text }
	end
	local function node_autolink(url)
		return { kind = "link", kids = { node_text(url) }, url = url }
	end
	local function node_mail(address)
		return { kind = "link", kids = { node_text(address) }, url = "mailto:" .. address }
	end

	-- CommonMark's autolinks, an absolute url or an address between angle
	-- brackets. Anything else in brackets stays escaped text.
	local alnum = R("az", "AZ", "09")
	local scheme_char = alnum + S("+.-")
	local uri = R("az", "AZ") * scheme_char * scheme_char ^ -30 * P(":")
		* (1 - S("<>") - R("\0 ") - P("\127")) ^ 0
	local address = (alnum + S(".!#$%&'*+/=?^_`{|}~-")) ^ 1
		* P("@") * alnum * (alnum + S(".-")) ^ 0

	-- An underscore inside a word belongs to the word, as in snake_case,
	-- so it neither opens nor closes emphasis.
	local word = alnum + R("\128\255")
	local inner = B(word) * P("_") ^ 1 * #word

	-- A character reference passes through for the browser to decode,
	-- within CommonMark's lengths. Its shape leaves no room for markup, and
	-- a name the browser does not know shows as written.
	local digit, hex = R("09"), R("09", "af", "AF")
	local reference = P("&") * (R("az", "AZ") * alnum ^ -31
		+ P("#") * digit * digit ^ -6
		+ P("#") * S("xX") * hex * hex ^ -5) * P(";")

	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("\\") * C(punctuation)) / node_text
		+ C(reference) / node_reference
		+ (P("<") * C(uri) * P(">")) / node_autolink
		+ (P("<") * C(address) * P(">")) / node_mail
		+ (P("[!") * label * target * P("]") * target) / node_image
		+ (P("!") * label * target) / node_image
		+ (label * target) / node_link
		+ (P("**") * C((1 - P("**")) ^ 1) * P("**")) / node_strong
		+ (-B(word) * P("__") * C((inner + (1 - P("__"))) ^ 1) * P("__") * -word) / node_strong
		+ (P("*") * C((1 - whitespace) * (1 - P("*")) ^ 0) * P("*")) / node_em
		+ (-B(word) * P("_") * C((1 - whitespace) * (inner + (1 - P("_"))) ^ 0) * P("_") * -word) / node_em
		+ C((1 - marker + inner) ^ 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)
	-- A run of equals signs or dashes under a paragraph turns it into a
	-- heading, as on GitHub, rather than reading as more of its text.
	local setext_line = space ^ 0 * (C(P("=") ^ 1) + C(P("-") ^ 1))
		* space ^ 0 * eol
	local code_indent = P("    ") + P("\t")
	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 follows_list(blocks)
		local last = blocks[#blocks]
		return last ~= nil and last.kind == "list"
	end

	-- An indented block runs on through blank lines, and up to four spaces
	-- or one tab come off each of its lines.
	local function continues_code(line)
		return code_indent:match(line) or not line:find("%S")
	end

	local function unindent(line)
		return line:match("^\t(.*)$") or line:match("^ ? ? ? ?(.*)$")
	end

	local function setext_level(line)
		local run = setext_line:match(line)
		if run then
			return run:sub(1, 1) == "=" and 1 or 2
		end
	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) 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
					local item = { content }
					i = i + 1
					-- A wrapped line carries the item on, whatever its
					-- indent, while a rule still ends the list.
					while i <= n and not ends_paragraph(lines, i)
						and not break_line:match(lines[i]) do
						item[#item + 1] = lines[i]
						i = i + 1
					end
					items[#items + 1] = paragraph_lines(item)
				end
				blocks[#blocks + 1] = {
					kind = "list",
					ordered = is_ordered,
					items = items,
				}
			-- Four spaces or a tab start a code block, except straight
			-- after a list, where an indented paragraph more likely
			-- carries an item on.
			elseif code_indent:match(line) and not follows_list(blocks) then
				local code = {}
				while i <= n and continues_code(lines[i]) do
					code[#code + 1] = unindent(lines[i])
					i = i + 1
				end
				while code[#code]:match("^%s*$") do
					code[#code] = nil
				end
				blocks[#blocks + 1] = {
					kind = "code",
					text = table.concat(code, "\n") .. "\n",
				}
			else
				local paragraph, level = { line }, nil
				i = i + 1
				while i <= n do
					level = setext_level(lines[i])
					if level or ends_paragraph(lines, i)
						or break_line:match(lines[i]) then
						break
					end
					paragraph[#paragraph + 1] = lines[i]
					i = i + 1
				end
				if level then
					local words = {}
					for k, text in ipairs(paragraph) do
						words[k] = trim(text)
					end
					blocks[#blocks + 1] = {
						kind = "heading",
						level = level,
						kids = parse_inline(table.concat(words, " ")),
					}
					i = i + 1
				else
					blocks[#blocks + 1] = {
						kind = "para",
						lines = paragraph_lines(paragraph),
					}
				end
			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)

	-- Fonts by name and by position. Anything else, \fP for the previous
	-- font among them, falls back to roman.
	local fonts = {
		B = "B", I = "I", BI = "B", CB = "B", CI = "I",
		["2"] = "I", ["3"] = "B",
	}

	-- The named characters and strings generated pages use most. A name
	-- with no entry renders as nothing rather than as its spelling.
	local man_chars = {
		aq = "'", dq = '"', cq = "’", oq = "‘", lq = "“", rq = "”",
		hy = "-", en = "–", em = "—", bu = "•", co = "©", rg = "®",
		tm = "™", ti = "~", ha = "^", rs = "\\",
	}
	local man_strings = {
		lq = "“", rq = "”", R = "®", Tm = "™", Aq = "'",
		['L"'] = "“", ['R"'] = "”",
	}

	local backslash = P("\\")
	local function text_token(text) return { kind = "text", text = text } end
	local skip_token = { kind = "skip" }
	local function skip() return skip_token end
	local function font_token(name)
		return { kind = "font", font = fonts[name] or "R" }
	end
	local function char_token(name) return text_token(man_chars[name] or "") end
	local function string_token(name)
		return text_token(man_strings[name] or "")
	end

	-- \( takes a two character name and \[ a longer one, bounded so that a
	-- line of unclosed openers stays linear.
	local named = P("(") * C(P(1) * P(1))
		+ P("[") * C((1 - P("]")) ^ -64) * P("]")

	-- Comments, size changes, colours, the half line moves DocBook puts
	-- around a footnote mark and conditional braces draw nothing, and nor do
	-- the escapes that only steer hyphenation or breaks.
	local man_inline_grammar = Ct((
		  (backslash * P("f") * (named + C(P(1)))) / font_token
		+ (backslash * S('"#') * P(1) ^ 0) / skip
		+ (backslash * P("-")) / function() return text_token("-") end
		+ (backslash * P("e")) / function() return text_token("\\") end
		+ (backslash * S(" 0~")) / function() return text_token(" ") end
		+ (backslash * S("&|^%:cud{}")) / skip
		+ (backslash * named) / char_token
		+ (backslash * P("*") * (named + C(P(1)))) / string_token
		+ (backslash * P("s") * S("+-") ^ -1 * (named + R("09") * R("09") ^ -1)) / skip
		+ (backslash * S("mM") * (named + P(1))) / skip
		+ (backslash * C(P(1))) / text_token
		+ backslash / skip
		+ 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

	-- A no fill block keeps its lines as they are, so only the text of
	-- each one is wanted there, without its fonts.
	local function man_plain(text)
		local parts = {}
		for _, token in ipairs(man_inline_grammar:match(text) or {}) do
			if token.kind == "text" then
				parts[#parts + 1] = token.text
			end
		end
		return table.concat(parts)
	end

	-- A macro's arguments divide at blanks, except inside double quotes.
	local function man_args(rest)
		local args, i = {}, 1
		while true do
			i = rest:find("%S", i)
			if not i then
				return args
			end
			local stop
			if rest:sub(i, i) == '"' then
				stop = rest:find('"', i + 1, true) or #rest + 1
				args[#args + 1] = rest:sub(i + 1, stop - 1)
			else
				stop = rest:find("%s", i) or #rest + 1
				args[#args + 1] = rest:sub(i, stop - 1)
			end
			i = stop + 1
		end
	end

	local function in_font(font, nodes)
		if font == "B" then
			return { { kind = "strong", kids = nodes } }
		elseif font == "I" then
			return { { kind = "em", kids = nodes } }
		end
		return nodes
	end

	-- Each alternating macro sets its arguments in its two fonts by turns
	-- with nothing between them, as in .BR ls (1).
	local single_font = { B = "B", I = "I" }
	local alternating = {
		BR = { "B", "R" }, RB = { "R", "B" }, IR = { "I", "R" },
		RI = { "R", "I" }, BI = { "B", "I" }, IB = { "I", "B" },
	}

	-- No fill blocks under their man(7) and groff names, and under the Vb
	-- and Ve that pod2man defines for its verbatim paragraphs.
	local nofill_start = { nf = true, EX = true, Vb = true }
	local nofill_end = { fi = true, EE = true, Ve = true }

	-- Macros that only begin a paragraph, Sp being pod2man's, ones that only
	-- set spacing or an index entry, and the requests that open a definition,
	-- which runs to a line of two dots.
	local breaks = {
		TH = true, PP = true, LP = true, P = true, HP = true, RS = true,
		RE = true, sp = true, Sp = true,
	}
	local silent = { PD = true, DT = true, IX = true }
	local definitions = { de = true, de1 = true, am = true, ig = true }

	local function parse_man(text)
		local lines = split_lines(text)
		local blocks, defined = {}, {}
		local paragraph, nofill_lines, nofill = nil, nil, false
		local skipping, tag_pending = false, 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
		-- A tagged paragraph puts its tag on a line of its own above the
		-- text it describes.
		local function add_line(nodes)
			if tag_pending then
				nodes[#nodes + 1] = { kind = "break" }
				tag_pending = false
			end
			paragraph = paragraph or {}
			paragraph[#paragraph + 1] = nodes
		end
		for _, line in ipairs(lines) do
			local macro, rest = macro_line:match(line)
			if skipping then
				skipping = not line:find("^%.%.%s*$")
			elseif macro and macro:sub(1, 1) == "\\"
				or line:find("^[.']%s*$") then
				-- A comment, the close of a conditional block or an empty
				-- request.
			elseif nofill then
				if nofill_end[macro] then
					nofill = false
					flush_nofill()
				elseif not macro then
					nofill_lines[#nofill_lines + 1] = man_plain(line)
				end
			elseif definitions[macro] then
				-- What a page defines for itself cannot be expanded here,
				-- so the definition is skipped and so are its uses.
				local name = man_args(rest)[1]
				if name and macro ~= "ig" then
					defined[name] = true
				end
				skipping = true
			elseif line:match("^%s*$") then
				flush_paragraph()
			elseif macro == "SH" or macro == "SS" then
				flush_paragraph()
				tag_pending = false
				blocks[#blocks + 1] = {
					kind = "heading",
					level = (macro == "SH") and 2 or 3,
					kids = man_inline(table.concat(man_args(rest), " ")),
				}
			elseif nofill_start[macro] then
				flush_paragraph()
				nofill = true
				nofill_lines = {}
			elseif nofill_end[macro] then
				flush_nofill()
			elseif single_font[macro] then
				local nodes = man_inline(table.concat(man_args(rest), " "))
				add_line(in_font(single_font[macro], nodes))
			elseif alternating[macro] then
				local pair, nodes = alternating[macro], {}
				for k, arg in ipairs(man_args(rest)) do
					local font = pair[(k - 1) % 2 + 1]
					for _, node in ipairs(in_font(font, man_inline(arg))) do
						nodes[#nodes + 1] = node
					end
				end
				add_line(nodes)
			elseif macro == "TP" then
				flush_paragraph()
				tag_pending = true
			elseif macro == "IP" then
				-- A bullet or a number leads the text on its line, and any
				-- other tag takes a line of its own.
				flush_paragraph()
				local tag = man_args(rest)[1]
				if tag and tag ~= "" then
					local nodes, plain = man_inline(tag), man_plain(tag)
					if plain:find("%w") and not plain:find("^%d+[.)]?$") then
						nodes[#nodes + 1] = { kind = "break" }
					else
						nodes[#nodes + 1] = { kind = "text", text = " " }
					end
					add_line(nodes)
				end
			elseif macro == "br" then
				if paragraph then
					local last = paragraph[#paragraph]
					last[#last + 1] = { kind = "break" }
				end
			elseif breaks[macro] then
				flush_paragraph()
			elseif macro
				and (silent[macro] or defined[macro] or macro:find("^%l")) then
				-- A macro the page defined, or a roff request, which in
				-- lower case only steers the layout.
			elseif macro then
				local args = man_args(rest)
				if #args > 0 then
					add_line(man_inline(table.concat(args, " ")))
				end
			else
				-- A line of nothing but escapes that draw nothing, such as
				-- a conditional's brace, adds no line.
				local nodes = man_inline(line)
				if #nodes > 0 then
					add_line(nodes)
				end
			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 = {}
	id_counts = { cgit = 1, header = 1, about = 1 }
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 each one becomes U+FFFD, as a browser would show it, and the
-- text past it survives. A plain find works on Lua 5.1, where a NUL cannot
-- sit in a pattern.
local function replace_nul(text)
	if not text:find("\0", 1, true) then
		return text
	end
	local parts, start = {}, 1
	while true do
		local at = text:find("\0", start, true)
		if not at then
			parts[#parts + 1] = text:sub(start)
			return table.concat(parts, "\239\191\189")
		end
		parts[#parts + 1] = text:sub(start, at - 1)
		start = at + 1
	end
end

function filter_close()
	local text = replace_nul(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