Restyle the sources and fix the audit's findings

author: Bryce Kwon <bryce@brycekwon.com> 2026-08-10 19:29:44 -1000
committer: Bryce Kwon <bryce@brycekwon.com> 2026-08-10 19:29:44 -1000
commit: 6c62bdc268ce8a8d98ab999846fcabd972507ad6 (patch)
parent: b8c11770bc2d7aa8638d6f1a99074e2ce626f2b2 (diff)
tree: 5e6dc6aa5bd640cb3fe2975586573fdafc10261d
download: cgit-6c62bdc268ce8a8d98ab999846fcabd972507ad6.tar.gz
cgit-6c62bdc268ce8a8d98ab999846fcabd972507ad6.zip
62 files changed, 6506 insertions, 5458 deletions
diff --git a/Makefile b/Makefile
index 010427b..e4ff38a 100644
--- a/Makefile
+++ b/Makefile
@@ -1,13 +1,17 @@
-# cgit top-level Makefile. cgit itself is compiled inside the bundled Git tree
-# by source/cgit.mk, which this file invokes; everything else (docs, install,
-# packaging, maintenance) is driven from here.
+# The entry point for building, documenting and installing cgit. The program
+# itself is compiled inside the bundled Git tree by source/cgit.mk, which this
+# file invokes with the paths and the defaults it needs. Everything that is not
+# compilation, that is the manual page, the install and uninstall rules and the
+# housekeeping targets, is driven from here.
 
 .DEFAULT_GOAL := all
 .SUFFIXES:
 
-# Project layout. Sources live in source/, web assets in assets/, Lua filters
-# in custom/extensions/, the bundled Git tree in vendor/git, and every generated
-# file lands in build/. CGIT_ROOT reaches from the Git tree back to the root.
+# Sources live in source/, web assets in assets/, Lua filters in
+# custom/extensions/, the bundled Git tree in vendor/git, and everything
+# generated lands in build/. CGIT_ROOT leads from inside the Git tree back to
+# the project root, because the cgit build runs with vendor/git as its working
+# directory.
 SRCDIR   = source
 ASSETDIR = assets
 EXTDIR   = custom/extensions
@@ -15,13 +19,14 @@ GITDIR   = vendor/git
 BUILDDIR = build
 CGIT_ROOT = ../..
 
-# Git release fetched by get-git when the submodule is not checked out. Keep in
-# step with the submodule pin in .gitmodules.
+# The Git release that get-git fetches when the submodule is not checked out.
+# Keep it in step with the submodule pin in .gitmodules.
 GIT_VERSION = 2.55.0
 GIT_URL = https://www.kernel.org/pub/software/scm/git/git-$(GIT_VERSION).tar.xz
 
-# Compiled-in cgit defaults, overridable through cgit.conf. All carry the CGIT_
-# prefix except CACHE_ROOT, inherited from upstream under that name.
+# Compiled-in cgit defaults, all overridable through cgit.conf. Every one
+# carries the CGIT_ prefix except CACHE_ROOT, which is inherited from upstream
+# under that name.
 CGIT_VERSION = v2.2.0
 CGIT_SCRIPT_NAME = cgit.cgi
 CGIT_SCRIPT_PATH = /var/www/htdocs/cgit
@@ -29,8 +34,8 @@ CGIT_DATA_PATH = $(CGIT_SCRIPT_PATH)
 CGIT_CONFIG = /etc/cgitrc
 CACHE_ROOT = /var/cache/cgit
 
-# Standard GNU-style install directories, lowercase so packaging can override
-# them; filterdir is the one cgit-specific addition.
+# The standard GNU install directories, kept lowercase so that packaging can
+# override them. filterdir is the one cgit-specific addition.
 prefix = /usr/local
 libdir = $(prefix)/lib
 filterdir = $(libdir)/cgit/filters
@@ -38,13 +43,14 @@ docdir = $(prefix)/share/doc/cgit
 htmldir = $(docdir)
 mandir = $(prefix)/share/man
 
-# Files installed into the data path and the filter dir. One list each drives
-# both install and uninstall, so the two can never drift apart.
+# The files that go into the data path and the filter directory. One list each
+# drives both install and uninstall, so the two can never drift apart.
 ASSETS       = cgit.css cgit.js cgit.png favicon.ico robots.txt
 FILTER_FILES = $(notdir $(wildcard $(EXTDIR)/*.lua))
 
-# The man-page source (*.5.txt) sits at the project root and is rendered into
-# build/. Only one man section exists, so DOC_* derive straight from MAN5_TXT.
+# The manual page source, *.5.txt, sits at the project root and is rendered
+# into build/. Only one man section exists, so both document lists derive
+# straight from MAN5_TXT.
 MAN5_TXT = $(wildcard *.5.txt)
 DOC_MAN5 = $(patsubst %.txt,$(BUILDDIR)/%,$(MAN5_TXT))
 DOC_HTML = $(patsubst %.txt,$(BUILDDIR)/%.html,$(MAN5_TXT))
@@ -56,8 +62,8 @@ ASCIIDOC_HTML = xhtml11
 ASCIIDOC_COMMON = $(ASCIIDOC) $(ASCIIDOC_EXTRA)
 TXT_TO_HTML = $(ASCIIDOC_COMMON) -b $(ASCIIDOC_HTML)
 
-# Reuse Git's platform detection, then let the user override any of the settings
-# above through cgit.conf. Both includes must stay after the defaults.
+# Reuse Git's platform detection, then let the user override any of the
+# settings above through cgit.conf. Both includes must stay after the defaults.
 -include $(GITDIR)/config.mak.uname
 -include cgit.conf
 
@@ -65,13 +71,10 @@ export CGIT_VERSION CGIT_SCRIPT_NAME CGIT_SCRIPT_PATH CGIT_DATA_PATH CGIT_CONFIG
 
 all: cgit
 
-# cgit is compiled inside the bundled Git tree so it can reuse Git's build
-# machinery (see source/cgit.mk). Objects and the binary go to build/.
-#
-# NO_RUST keeps Cargo out of the build. Git 2.55 inverted the flag that used
-# to be WITH_RUST, so Rust is now compiled in unless it is turned off, and
-# leaving this out would quietly make Cargo a build requirement. Git 3.0 drops
-# the choice entirely, at which point this stops working and Rust becomes
+# NO_RUST keeps Cargo out of the build. Git 2.55 inverted the flag that used to
+# be WITH_RUST, so Rust is now compiled in unless it is turned off, and leaving
+# this out would quietly make Cargo a build requirement. Git 3.0 drops the
+# choice entirely, at which point this stops working and Rust becomes
 # mandatory.
 GIT_BUILD_FLAGS = NO_CURL=1 NO_OPENSSL=1 NO_RUST=1
 
@@ -86,8 +89,8 @@ sparse:
 test: cgit-with-git
 	+@$(MAKE) -C tests all
 
-# The test helpers link against a fully built Git tree, so this asks the
-# sub-make for Git's own `all` alongside cgit.
+# The test helpers link against a fully built Git tree, so this asks the sub
+# make for Git's own all target alongside cgit.
 cgit-with-git:
 	+@$(MAKE) cgit EXTRA_GIT_TARGETS=all
 
@@ -138,9 +141,10 @@ uninstall-html:
 	    rm -fv $(DESTDIR)$(htmldir)/$$i; \
 	done
 
-# Compiled output lives under build/; the test suite writes into tests/, so
-# cleaning removes both. The tests descent is guarded because tests/Makefile
-# includes git's config.mak.uname, which needs the submodule present.
+# Compiled output lives under build/ and the test suite writes into tests/, so
+# cleaning removes both. The descent into tests is guarded because
+# tests/Makefile includes git's config.mak.uname, which needs the submodule
+# present.
 clean:
 	$(RM) -r $(BUILDDIR)
 	@if test -e $(GITDIR)/config.mak.uname; then $(MAKE) -C tests clean; fi
@@ -148,7 +152,6 @@ clean:
 cleanall: clean
 	@if test -e $(GITDIR)/Makefile; then $(MAKE) -C $(GITDIR) clean; fi
 
-# Fetch a Git release tarball for builds without the submodule checked out.
 get-git:
 	mkdir -p $(dir $(GITDIR))
 	curl -L $(GIT_URL) | tar -xJf - && rm -rf $(GITDIR) && mv git-$(GIT_VERSION) $(GITDIR)
diff --git a/source/cache.c b/source/cache.c
index c6d0427..d6e450a 100644
--- a/source/cache.c
+++ b/source/cache.c
@@ -1,33 +1,49 @@
-/* cache.c: cache management
- *
- * Copyright (C) 2006-2014 cgit Development Team <cgit@lists.zx2c4.com>
- *
- * Licensed under GNU General Public License v2
- *   (see LICENSE.txt for full license text)
- *
- *
- * The cache is just a directory structure where each file is a cache slot,
- * and each filename is based on the hash of some key (e.g. the cgit url).
- * Each file contains the full key followed by the cached content for that
- * key.
- *
+/*
+ * The cache that lets a repeated request be answered from disk instead of
+ * being rendered again. A slot is one file named after the hash of the
+ * request key, holding that key and then the page it rendered to, and a lock
+ * file beside it is where a replacement page is written before being renamed
+ * over the slot. Only the process holding that lock rebuilds a slot, so a
+ * request arriving while a stale slot is being rebuilt is served the stale
+ * page, and a request with no usable slot at all renders straight to the
+ * client without caching anything.
  */
 
-#include "cgit.h"
 #include "cache.h"
+#include "cgit.h"
 #include "html.h"
+#include "shared.h"
 #ifdef HAVE_LINUX_SENDFILE
 #include <sys/sendfile.h>
 #endif
 
+// One read of a slot file. The stored key has to be recognised out of a
+// single such read, so this also bounds how long a cacheable key can be, see
+// key_fits_slot.
 #define CACHE_BUFSIZE (1024 * 4)
-/* Crude implementation of 32-bit FNV-1 hash algorithm,
- * see http://www.isthe.com/chongo/tech/comp/fnv/ for details
- * about the magic numbers.
- */
+
+// A slot is named by this many hex digits of the key hash, which is also how
+// cache_ls tells slots from the lock files sitting beside them.
+#define SLOT_NAME_LEN 8
+
+// The 32 bit FNV-1 offset basis and prime.
 #define FNV_OFFSET 0x811c9dc5
 #define FNV_PRIME  0x01000193
 
+/*
+ * Cache trouble goes to stderr, which under CGI is the web server's error
+ * log, so that it cannot land in the middle of the page being written to
+ * stdout.
+ */
+__attribute__((format (printf,1,2)))
+static void log_error(const char *format, ...)
+{
+	va_list args;
+	va_start(args, format);
+	vfprintf(stderr, format, args);
+	va_end(args);
+}
+
 struct cache_slot {
 	const char *key;
 	size_t keylen;
@@ -35,56 +51,56 @@ struct cache_slot {
 	cache_fill_fn fn;
 	int cache_fd;
 	int lock_fd;
-	int stdout_fd;
-	const char *cache_name;
-	const char *lock_name;
-	int match;
-	struct stat cache_st;
-	int bufsize;
+	int saved_stdout;
+	const char *path;
+	const char *lock_path;
+	int key_matches;
+	// The slot as it was when it was opened, or the lock file once
+	// fill_slot has written a page into it.
+	struct stat st;
+	// How much of the slot was read into buf, not the size of buf.
+	int buflen;
 	char buf[CACHE_BUFSIZE];
 };
 
-/* Open an existing cache slot and fill the cache buffer with
- * (part of) the content of the cache file. Return 0 on success
- * and errno otherwise.
- */
 static int open_slot(struct cache_slot *slot)
 {
-	char *bufz;
-	ssize_t bufkeylen = -1;
+	char *nul;
+	ssize_t keylen = -1;
 
-	slot->cache_fd = open(slot->cache_name, O_RDONLY);
+	slot->cache_fd = open(slot->path, O_RDONLY);
 	if (slot->cache_fd == -1)
 		return errno;
 
-	if (fstat(slot->cache_fd, &slot->cache_st))
+	if (fstat(slot->cache_fd, &slot->st))
 		return errno;
 
-	slot->bufsize = xread(slot->cache_fd, slot->buf, sizeof(slot->buf));
-	if (slot->bufsize < 0)
+	slot->buflen = xread(slot->cache_fd, slot->buf, sizeof(slot->buf));
+	if (slot->buflen < 0)
 		return errno;
 
-	bufz = memchr(slot->buf, 0, slot->bufsize);
-	if (bufz)
-		bufkeylen = bufz - slot->buf;
+	nul = memchr(slot->buf, 0, slot->buflen);
+	if (nul)
+		keylen = nul - slot->buf;
 
 	if (slot->key)
-		slot->match = bufkeylen >= 0 && (size_t)bufkeylen == slot->keylen &&
-		    !memcmp(slot->key, slot->buf, bufkeylen + 1);
+		slot->key_matches = keylen >= 0 &&
+		    (size_t)keylen == slot->keylen &&
+		    !memcmp(slot->key, slot->buf, keylen + 1);
 
 	return 0;
 }
 
-/* A key longer than the buffer above can never be read back, so a slot keyed
- * on one would never match and every such request would regenerate its page
- * while still writing a slot nothing can use. Those requests skip the cache
- * instead. */
+/*
+ * A key longer than the buffer above can never be read back by open_slot, so
+ * a slot keyed on one would never match and every such request would
+ * regenerate its page while still writing a slot nothing can use.
+ */
 static int key_fits_slot(const char *key)
 {
 	return strlen(key) + 1 <= CACHE_BUFSIZE;
 }
 
-/* Close the active cache slot */
 static int close_slot(struct cache_slot *slot)
 {
 	int err = 0;
@@ -97,7 +113,6 @@ static int close_slot(struct cache_slot *slot)
 	return err;
 }
 
-/* Print the content of the active cache slot (but skip the key). */
 static int print_slot(struct cache_slot *slot)
 {
 	off_t off;
@@ -108,7 +123,7 @@ static int print_slot(struct cache_slot *slot)
 	off = slot->keylen + 1;
 
 #ifdef HAVE_LINUX_SENDFILE
-	size = slot->cache_st.st_size;
+	size = slot->st.st_size;
 
 	do {
 		ssize_t ret;
@@ -116,7 +131,10 @@ static int print_slot(struct cache_slot *slot)
 		if (ret < 0) {
 			if (errno == EAGAIN || errno == EINTR)
 				continue;
-			/* Fall back to read/write on EINVAL or ENOSYS */
+			// EINVAL and ENOSYS mean this kernel or this pair of
+			// descriptors cannot do sendfile at all, so fall back
+			// to the read and write loop rather than fail the
+			// request.
 			if (errno == EINVAL || errno == ENOSYS)
 				break;
 			return errno;
@@ -141,30 +159,41 @@ static int print_slot(struct cache_slot *slot)
 	} while (1);
 }
 
-/* Check if the slot has expired */
+static int serve_slot(struct cache_slot *slot)
+{
+	int err;
+
+	err = print_slot(slot);
+	if (err)
+		log_error("[cgit] error printing cache %s: %s (%d)\n",
+			  slot->path,
+			  strerror(err),
+			  err);
+	return err;
+}
+
 static int is_expired(struct cache_slot *slot)
 {
 	if (slot->ttl < 0)
 		return 0;
-	else
-		return slot->cache_st.st_mtime + slot->ttl * 60 < time(NULL);
+	return slot->st.st_mtime + slot->ttl * SECONDS_PER_MINUTE < time(NULL);
 }
 
-/* Check if the slot has been modified since we opened it.
- * NB: If stat() fails, we pretend the file is modified.
+/*
+ * A stat that fails counts as modified, so that the caller leaves alone a file
+ * it was unable to look at.
  */
 static int is_modified(struct cache_slot *slot)
 {
-	struct stat st;
+	struct stat current;
 
-	if (stat(slot->cache_name, &st))
+	if (stat(slot->path, &current))
 		return 1;
-	return (st.st_ino != slot->cache_st.st_ino ||
-		st.st_mtime != slot->cache_st.st_mtime ||
-		st.st_size != slot->cache_st.st_size);
+	return (current.st_ino != slot->st.st_ino ||
+		current.st_mtime != slot->st.st_mtime ||
+		current.st_size != slot->st.st_size);
 }
 
-/* Close an open lockfile */
 static int close_lock(struct cache_slot *slot)
 {
 	int err = 0;
@@ -177,9 +206,11 @@ static int close_lock(struct cache_slot *slot)
 	return err;
 }
 
-/* Create a lockfile used to store the generated content for a cache
- * slot, and write the slot key + \0 into it.
- * Returns 0 on success and errno otherwise.
+/*
+ * The lock file becomes the slot once it is renamed, so it has to open with
+ * the key the same way a slot does. The lock is taken without blocking,
+ * because failing to get it is how a second process learns that this slot is
+ * already being rebuilt, so it returns an errno instead of waiting.
  */
 static int lock_slot(struct cache_slot *slot)
 {
@@ -190,7 +221,7 @@ static int lock_slot(struct cache_slot *slot)
 		.l_len = 0,
 	};
 
-	slot->lock_fd = open(slot->lock_name, O_RDWR | O_CREAT,
+	slot->lock_fd = open(slot->lock_path, O_RDWR | O_CREAT,
 			     S_IRUSR | S_IWUSR);
 	if (slot->lock_fd == -1)
 		return errno;
@@ -200,6 +231,8 @@ static int lock_slot(struct cache_slot *slot)
 		slot->lock_fd = -1;
 		return saved_errno;
 	}
+	// A run that died before its rename leaves the lock file behind, so
+	// start from empty now that nobody else can be writing it.
 	if (ftruncate(slot->lock_fd, 0) < 0)
 		return errno;
 	if (xwrite(slot->lock_fd, slot->key, slot->keylen + 1) < 0)
@@ -207,24 +240,19 @@ static int lock_slot(struct cache_slot *slot)
 	return 0;
 }
 
-/* Release the current lockfile. If `replace_old_slot` is set the
- * lockfile replaces the old cache slot, otherwise the lockfile is
- * just deleted.
- */
 static int unlock_slot(struct cache_slot *slot, int replace_old_slot)
 {
 	int err;
 
 	if (replace_old_slot)
-		err = rename(slot->lock_name, slot->cache_name);
+		err = rename(slot->lock_path, slot->path);
 	else
-		err = unlink(slot->lock_name);
+		err = unlink(slot->lock_path);
 
-	/* Restore stdout and close the temporary FD. */
-	if (slot->stdout_fd >= 0) {
-		dup2(slot->stdout_fd, STDOUT_FILENO);
-		close(slot->stdout_fd);
-		slot->stdout_fd = -1;
+	if (slot->saved_stdout >= 0) {
+		dup2(slot->saved_stdout, STDOUT_FILENO);
+		close(slot->saved_stdout);
+		slot->saved_stdout = -1;
 	}
 
 	if (err)
@@ -233,48 +261,84 @@ static int unlock_slot(struct cache_slot *slot, int replace_old_slot)
 	return 0;
 }
 
-/* Generate the content for the current cache slot by redirecting
- * stdout to the lock-fd and invoking the callback function
+// Only one slot is ever being filled at a time, so a single pointer is enough
+// for cache_abandon_fill to find its way back to the client.
+static struct cache_slot *slot_being_filled;
+
+void cache_abandon_fill(void)
+{
+	struct cache_slot *slot = slot_being_filled;
+
+	if (!slot)
+		return;
+	slot_being_filled = NULL;
+
+	// Emptied while stdout still points at the lock file, so the half
+	// rendered page goes into the file about to be removed rather than
+	// reaching the client ahead of whatever is written next.
+	html_flush();
+
+	if (slot->saved_stdout >= 0) {
+		dup2(slot->saved_stdout, STDOUT_FILENO);
+		close(slot->saved_stdout);
+		slot->saved_stdout = -1;
+	}
+	unlink(slot->lock_path);
+}
+
+/*
+ * Renders with stdout pointed at the lock file, and on success or failure
+ * alike it is unlock_slot that gives stdout back.
  */
 static int fill_slot(struct cache_slot *slot)
 {
-	/* Preserve stdout */
-	slot->stdout_fd = dup(STDOUT_FILENO);
-	if (slot->stdout_fd == -1)
+	slot->saved_stdout = dup(STDOUT_FILENO);
+	if (slot->saved_stdout == -1)
 		return errno;
 
-	/* Redirect stdout to lockfile */
 	if (dup2(slot->lock_fd, STDOUT_FILENO) == -1)
 		return errno;
 
-	/* Generate cache content */
+	slot_being_filled = slot;
 	slot->fn();
+	slot_being_filled = NULL;
 
-	/* Make sure any buffered data is flushed to the file */
+	// The page is sitting in html.c's buffer and then in stdio's, and all
+	// of it has to reach the lock file before that file is renamed into
+	// place.
 	html_flush();
 	if (fflush(stdout))
 		return errno;
 
-	/* update stat info */
-	if (fstat(slot->lock_fd, &slot->cache_st))
+	// print_slot takes the length of what it copies from here, and what
+	// it copies after a fill is the lock file rather than the old slot.
+	if (fstat(slot->lock_fd, &slot->st))
 		return errno;
 
 	return 0;
 }
 
-unsigned long cache_hash_str(const char *str)
+/*
+ * Giving up is always a valid outcome, because the caller still has the
+ * expired copy open and can serve that.
+ */
+static void refresh_slot(struct cache_slot *slot)
 {
-	unsigned long h = FNV_OFFSET;
-	unsigned char *s = (unsigned char *)str;
-
-	if (!s)
-		return h;
-
-	while (*s) {
-		h *= FNV_PRIME;
-		h ^= *s++;
+	if (lock_slot(slot))
+		return;
+
+	// If another process replaced the slot between open_slot and
+	// lock_slot, the copy already open is served rather than the newer
+	// one, which would mean opening that file and comparing the key in it,
+	// not worth a second descriptor and read on every expiry.
+	if (is_modified(slot) || fill_slot(slot)) {
+		unlock_slot(slot, 0);
+		close_lock(slot);
+	} else {
+		close_slot(slot);
+		unlock_slot(slot, 1);
+		slot->cache_fd = slot->lock_fd;
 	}
-	return h;
 }
 
 static int process_slot(struct cache_slot *slot)
@@ -282,216 +346,180 @@ static int process_slot(struct cache_slot *slot)
 	int err;
 
 	err = open_slot(slot);
-	if (!err && slot->match) {
-		if (is_expired(slot)) {
-			if (!lock_slot(slot)) {
-				/* If the cachefile has been replaced between
-				 * `open_slot` and `lock_slot`, we'll just
-				 * serve the stale content from the original
-				 * cachefile. This way we avoid pruning the
-				 * newly generated slot. The same code-path
-				 * is chosen if fill_slot() fails for some
-				 * reason.
-				 *
-				 * TODO? check if the new slot contains the
-				 * same key as the old one, since we would
-				 * prefer to serve the newest content.
-				 * This will require us to open yet another
-				 * file-descriptor and read and compare the
-				 * key from the new file, so for now we're
-				 * lazy and just ignore the new file.
-				 */
-				if (is_modified(slot) || fill_slot(slot)) {
-					unlock_slot(slot, 0);
-					close_lock(slot);
-				} else {
-					close_slot(slot);
-					unlock_slot(slot, 1);
-					slot->cache_fd = slot->lock_fd;
-				}
-			}
-		}
-		if ((err = print_slot(slot)) != 0) {
-			cache_log("[cgit] error printing cache %s: %s (%d)\n",
-				  slot->cache_name,
-				  strerror(err),
-				  err);
-		}
+	if (!err && slot->key_matches) {
+		if (is_expired(slot))
+			refresh_slot(slot);
+		err = serve_slot(slot);
 		close_slot(slot);
 		return err;
 	}
 
-	/* If the cache slot does not exist (or its key doesn't match the
-	 * current key), lets try to create a new cache slot for this
-	 * request. If this fails (for whatever reason), lets just generate
-	 * the content without caching it and fool the caller to believe
-	 * everything worked out (but print a warning on stdout).
-	 */
-
+	// If any part of creating a slot fails the page is still rendered
+	// straight to the client and the caller is told the request succeeded,
+	// because it did.
 	close_slot(slot);
 	if ((err = lock_slot(slot)) != 0) {
-		cache_log("[cgit] Unable to lock slot %s: %s (%d)\n",
-			  slot->lock_name, strerror(err), err);
+		log_error("[cgit] Unable to lock slot %s: %s (%d)\n",
+			  slot->lock_path, strerror(err), err);
 		slot->fn();
 		return 0;
 	}
 
 	if ((err = fill_slot(slot)) != 0) {
-		cache_log("[cgit] Unable to fill slot %s: %s (%d)\n",
-			  slot->lock_name, strerror(err), err);
+		log_error("[cgit] Unable to fill slot %s: %s (%d)\n",
+			  slot->lock_path, strerror(err), err);
 		unlock_slot(slot, 0);
 		close_lock(slot);
 		slot->fn();
 		return 0;
 	}
-	// We've got a valid cache slot in the lock file, which
-	// is about to replace the old cache slot. But if we
-	// release the lockfile and then try to open the new cache
-	// slot, we might get a race condition with a concurrent
-	// writer for the same cache slot (with a different key).
-	// Lets avoid such a race by just printing the content of
-	// the lock file.
+
+	// Opening the slot by name after the rename could land on a file a
+	// concurrent writer put there for a different key, so what gets
+	// printed is the descriptor still open on the lock file.
 	slot->cache_fd = slot->lock_fd;
 	unlock_slot(slot, 1);
-	if ((err = print_slot(slot)) != 0) {
-		cache_log("[cgit] error printing cache %s: %s (%d)\n",
-			  slot->cache_name,
-			  strerror(err),
-			  err);
-	}
+	err = serve_slot(slot);
 	close_slot(slot);
 	return err;
 }
 
-/* Print cached content to stdout, generate the content if necessary. */
+// The result lives in a static buffer and is only good until the next call.
+static char *format_time(const char *format, time_t when)
+{
+	static char buf[64];
+	struct tm tm;
+
+	if (!when)
+		return NULL;
+	gmtime_r(&when, &tm);
+	strftime(buf, sizeof(buf) - 1, format, &tm);
+	return buf;
+}
+
+/*
+ * The accumulator is an unsigned long rather than a fixed 32 bit type, so on a
+ * 64 bit host this is not the published FNV-1 value. All that decides is which
+ * slot a key lands in, and nothing outside a single build has to agree on the
+ * answer.
+ */
+unsigned long cache_hash_str(const char *str)
+{
+	unsigned long h = FNV_OFFSET;
+	unsigned char *s = (unsigned char *)str;
+
+	if (!s)
+		return h;
+
+	while (*s) {
+		h *= FNV_PRIME;
+		h ^= *s++;
+	}
+	return h;
+}
+
 int cache_process(int size, const char *path, const char *key, int ttl,
 		  cache_fill_fn fn)
 {
 	unsigned long hash;
 	int i;
-	struct strbuf filename = STRBUF_INIT;
-	struct strbuf lockname = STRBUF_INIT;
+	struct strbuf slot_path = STRBUF_INIT;
+	struct strbuf lock_path = STRBUF_INIT;
 	struct cache_slot slot;
 	int result;
 
-	/* If the cache is disabled, just generate the content */
 	if (size <= 0 || ttl == 0) {
 		fn();
 		return 0;
 	}
 
-	/* Verify input, calculate filenames */
 	if (!path) {
-		cache_log("[cgit] Cache path not specified, caching is disabled\n");
+		log_error("[cgit] Cache path not specified, caching is disabled\n");
 		fn();
 		return 0;
 	}
 	if (!key)
 		key = "";
 	if (!key_fits_slot(key)) {
-		cache_log("[cgit] Cache key too long for a slot, caching is "
+		log_error("[cgit] Cache key too long for a slot, caching is "
 			  "disabled for this request\n");
 		fn();
 		return 0;
 	}
 	hash = cache_hash_str(key) % size;
-	strbuf_addstr(&filename, path);
-	strbuf_ensure_end(&filename, '/');
-	for (i = 0; i < 8; i++) {
-		strbuf_addf(&filename, "%x", (unsigned char)(hash & 0xf));
+	strbuf_addstr(&slot_path, path);
+	strbuf_ensure_end(&slot_path, '/');
+	for (i = 0; i < SLOT_NAME_LEN; i++) {
+		strbuf_addf(&slot_path, "%x", (unsigned char)(hash & 0xf));
 		hash >>= 4;
 	}
-	strbuf_addbuf(&lockname, &filename);
-	strbuf_addstr(&lockname, ".lock");
+	strbuf_addbuf(&lock_path, &slot_path);
+	strbuf_addstr(&lock_path, ".lock");
 	slot.fn = fn;
 	slot.ttl = ttl;
-	slot.stdout_fd = -1;
-	slot.cache_name = filename.buf;
-	slot.lock_name = lockname.buf;
+	slot.saved_stdout = -1;
+	slot.path = slot_path.buf;
+	slot.lock_path = lock_path.buf;
 	slot.key = key;
 	slot.keylen = strlen(key);
 	result = process_slot(&slot);
 
-	strbuf_release(&filename);
-	strbuf_release(&lockname);
+	strbuf_release(&slot_path);
+	strbuf_release(&lock_path);
 	return result;
 }
 
-/* Return a strftime formatted date/time
- * NB: the result from this function is to shared memory
- */
-static char *sprintftime(const char *format, time_t time)
-{
-	static char buf[64];
-	struct tm tm;
-
-	if (!time)
-		return NULL;
-	gmtime_r(&time, &tm);
-	strftime(buf, sizeof(buf)-1, format, &tm);
-	return buf;
-}
-
 int cache_ls(const char *path)
 {
 	DIR *dir;
 	struct dirent *ent;
 	int err = 0;
+	// A NULL key leaves open_slot with nothing to compare against, so
+	// every slot it opens is simply read.
 	struct cache_slot slot = { NULL };
-	struct strbuf fullname = STRBUF_INIT;
+	struct strbuf slot_path = STRBUF_INIT;
 	size_t prefixlen;
 	char *nul;
 	int keylen;
 
 	if (!path) {
-		cache_log("[cgit] cache path not specified\n");
+		log_error("[cgit] cache path not specified\n");
 		return -1;
 	}
 	dir = opendir(path);
 	if (!dir) {
 		err = errno;
-		cache_log("[cgit] unable to open path %s: %s (%d)\n",
+		log_error("[cgit] unable to open path %s: %s (%d)\n",
 			  path, strerror(err), err);
 		return err;
 	}
-	strbuf_addstr(&fullname, path);
-	strbuf_ensure_end(&fullname, '/');
-	prefixlen = fullname.len;
+	strbuf_addstr(&slot_path, path);
+	strbuf_ensure_end(&slot_path, '/');
+	prefixlen = slot_path.len;
 	while ((ent = readdir(dir)) != NULL) {
-		if (strlen(ent->d_name) != 8)
+		if (strlen(ent->d_name) != SLOT_NAME_LEN)
 			continue;
-		strbuf_setlen(&fullname, prefixlen);
-		strbuf_addstr(&fullname, ent->d_name);
-		slot.cache_name = fullname.buf;
+		strbuf_setlen(&slot_path, prefixlen);
+		strbuf_addstr(&slot_path, ent->d_name);
+		slot.path = slot_path.buf;
 		if ((err = open_slot(&slot)) != 0) {
-			cache_log("[cgit] unable to open path %s: %s (%d)\n",
-				  fullname.buf, strerror(err), err);
+			log_error("[cgit] unable to open path %s: %s (%d)\n",
+				  slot_path.buf, strerror(err), err);
 			continue;
 		}
-		// The stored key is NUL-terminated within the buffer, but a
-		// truncated or corrupt slot may not be. Bound the print to
-		// what was read so %s cannot run off the end.
-		nul = memchr(slot.buf, 0, slot.bufsize);
-		keylen = nul ? (int)(nul - slot.buf) : slot.bufsize;
+		// A truncated or corrupt slot may hold no NUL, so the print is
+		// bounded by what was read and cannot run off the end.
+		nul = memchr(slot.buf, 0, slot.buflen);
+		keylen = nul ? (int)(nul - slot.buf) : slot.buflen;
 		htmlf("%s %s %10"PRIuMAX" %.*s\n",
-		      fullname.buf,
-		      sprintftime("%Y-%m-%d %H:%M:%S",
-				  slot.cache_st.st_mtime),
-		      (uintmax_t)slot.cache_st.st_size,
+		      slot_path.buf,
+		      format_time("%Y-%m-%d %H:%M:%S",
+				  slot.st.st_mtime),
+		      (uintmax_t)slot.st.st_size,
 		      keylen, slot.buf);
 		close_slot(&slot);
 	}
 	closedir(dir);
-	strbuf_release(&fullname);
+	strbuf_release(&slot_path);
 	return 0;
 }
-
-/* Print a message to stdout */
-void cache_log(const char *format, ...)
-{
-	va_list args;
-	va_start(args, format);
-	vfprintf(stderr, format, args);
-	va_end(args);
-}
-
diff --git a/source/cache.h b/source/cache.h
index f8018be..db77999 100644
--- a/source/cache.h
+++ b/source/cache.h
@@ -1,6 +1,7 @@
 /*
- * Since git has it's own cache.h which we include,
- * lets test on CGIT_CACHE_H to avoid confusion
+ * The front door to cgit's page cache. A caller hands over a key identifying
+ * the request and a callback that renders the page, and gets back either the
+ * copy already on disk or a freshly rendered one.
  */
 
 #ifndef CGIT_CACHE_H
@@ -8,30 +9,30 @@
 
 typedef void (*cache_fill_fn)(void);
 
-
-/* Print cached content to stdout, generate the content if necessary.
- *
- * Parameters
- *   size    max number of cache files
- *   path    directory used to store cache files
- *   key     the key used to lookup cache files
- *   ttl     max cache time in seconds for this key
- *   fn      content generator function for this key
- *
- * Return value
- *   0 indicates success, everything else is an error
+/*
+ * Write the page for key to stdout, taking it from the cache when a fresh
+ * slot for that key is there and rendering it through fn when it is not. size
+ * is how many slots the cache may use and path is the directory holding them.
+ * ttl is how many minutes a slot for this key stays fresh, where a negative
+ * ttl never expires and a ttl of zero skips the cache for this request.
+ * Returns 0 when the page was written, and an errno value when it was not.
  */
 extern int cache_process(int size, const char *path, const char *key, int ttl,
 			 cache_fill_fn fn);
 
-
-/* List info about all cache entries on stdout */
+// Write one line per cache slot to stdout, giving its path, modification
+// time, size and key.
 extern int cache_ls(const char *path);
 
-/* Print a message to stdout */
-__attribute__((format (printf,1,2)))
-extern void cache_log(const char *format, ...);
+/*
+ * Give up on the slot being filled, discarding what has been rendered into it
+ * and putting stdout back on the client. Anything that ends a request part way
+ * through rendering has to call this before it writes what the visitor should
+ * see, because until then stdout is the cache file and the visitor is on
+ * course to receive nothing at all. Does nothing when no slot is being filled.
+ */
+extern void cache_abandon_fill(void);
 
 extern unsigned long cache_hash_str(const char *str);
 
-#endif /* CGIT_CACHE_H */
+#endif // CGIT_CACHE_H
diff --git a/source/cgit.c b/source/cgit.c
index 7cce3d3..5c11a93 100644
--- a/source/cgit.c
+++ b/source/cgit.c
This diff is too large to be rendered inline. View it on its own page.
diff --git a/source/cgit.h b/source/cgit.h
index b05876e..eadb680 100644
--- a/source/cgit.h
+++ b/source/cgit.h
@@ -1,95 +1,71 @@
+/*
+ * The vocabulary shared by every part of cgit. It pulls in the git headers
+ * the renderers are written against, declares the record types they pass
+ * around, and declares the single global ctx holding the parsed request, the
+ * effective configuration and the repository being served. Anything that
+ * belongs to one module is declared in that module's own header instead.
+ */
+
 #ifndef CGIT_H
 #define CGIT_H
 
-#include <stdbool.h>
-
+// git-compat-util.h sets the feature test macros that git's own headers and
+// the system headers are then compiled against, so it has to come first.
 #include <git-compat-util.h>
 
 #include <archive.h>
+#include <blame.h>
 #include <commit.h>
+#include <config.h>
 #include <date.h>
-#include <diffcore.h>
 #include <diff.h>
+#include <diffcore.h>
 #include <environment.h>
 #include <graph.h>
 #include <grep.h>
 #include <hex.h>
 #include <log-tree.h>
 #include <notes.h>
-#include <object.h>
 #include <object-name.h>
+#include <object.h>
 #include <odb.h>
+#include <packfile.h>
 #include <path.h>
 #include <refs.h>
 #include <revision.h>
 #include <setup.h>
+#include <stdbool.h>
 #include <string-list.h>
 #include <strvec.h>
 #include <tag.h>
 #include <tree.h>
+#include <url.h>
 #include <utf8.h>
+#include <version.h>
 #include <wrapper.h>
 #include <xdiff-interface.h>
 #include <xdiff/xdiff.h>
 
-/* Add isgraph(x) to Git's sane ctype support (see git-compat-util.h) */
-#undef isgraph
-#define isgraph(x) (isprint((x)) && !isspace((x)))
-
-
-/*
- * Limits used for relative dates
- */
-#define TM_MIN    60
-#define TM_HOUR  (TM_MIN * 60)
-#define TM_DAY   (TM_HOUR * 24)
-#define TM_WEEK  (TM_DAY * 7)
-#define TM_YEAR  (TM_DAY * 365)
-#define TM_MONTH (TM_YEAR / 12.0)
-
-
-/*
- * Default encoding
- */
+// Every page cgit renders is UTF-8, so commit text in another encoding is
+// converted on the way in.
 #define PAGE_ENCODING "UTF-8"
 
-#define BIT(x)	(1U << (x))
-
-/*
- * Longest search string a request may supply. The filter bar matches its
- * query against every item a page lists, so this bounds the work one request
- * can ask for. It is not a configuration knob, in the same way as the ofs
- * ceiling in querystring_cb.
- */
-#define CGIT_MAX_SEARCH_LEN 512
-
-typedef void (*configfn)(const char *name, const char *value);
-typedef void (*filepair_fn)(struct diff_filepair *pair);
-typedef void (*linediff_fn)(char *line, int len);
+// The month is a double because a twelfth of a year is not a whole number of
+// seconds.
+#define SECONDS_PER_MINUTE 60
+#define SECONDS_PER_HOUR   (SECONDS_PER_MINUTE * 60)
+#define SECONDS_PER_DAY    (SECONDS_PER_HOUR * 24)
+#define SECONDS_PER_WEEK   (SECONDS_PER_DAY * 7)
+#define SECONDS_PER_YEAR   (SECONDS_PER_DAY * 365)
+#define SECONDS_PER_MONTH  (SECONDS_PER_YEAR / 12.0)
 
 typedef enum {
 	DIFF_UNIFIED, DIFF_SSDIFF, DIFF_STATONLY
 } diff_type;
 
-typedef enum {
-	ABOUT, COMMIT, SOURCE, EMAIL, AUTH
-} filter_type;
-
-struct cgit_filter {
-	int (*open)(struct cgit_filter *, va_list ap);
-	int (*close)(struct cgit_filter *);
-	void (*fprintfp)(struct cgit_filter *, FILE *, const char *prefix);
-	void (*cleanup)(struct cgit_filter *);
-	int argument_count;
-};
-
-struct cgit_exec_filter {
-	struct cgit_filter base;
-	char *cmd;
-	char **argv;
-	int old_stdout;
-	int pid;
-};
+// Filters are only ever held by pointer here, so their layout stays private
+// to filter.h.
+struct cgit_filter;
 
 struct cgit_repo {
 	char *url;
@@ -140,11 +116,13 @@ struct commitinfo {
 	struct commit *commit;
 	char *author;
 	char *author_email;
-	unsigned long author_date;
+	// git's own width for a commit timestamp, which is wider than a long
+	// on a 32 bit host and on Windows.
+	timestamp_t author_date;
 	int author_tz;
 	char *committer;
 	char *committer_email;
-	unsigned long committer_date;
+	timestamp_t committer_date;
 	int committer_tz;
 	char *subject;
 	char *msg;
@@ -154,7 +132,7 @@ struct commitinfo {
 struct taginfo {
 	char *tagger;
 	char *tagger_email;
-	unsigned long tagger_date;
+	timestamp_t tagger_date;
 	int tagger_tz;
 	char *msg;
 };
@@ -224,7 +202,7 @@ struct cgit_config {
 	char *script_name;
 	char *section;
 	char *repository_sort;
-	char *virtual_root;	/* Always ends with '/'. */
+	char *virtual_root;	// Always ends with a slash.
 	char *strict_export;
 	struct date_mode date_mode;
 	int cache_size;
@@ -316,7 +294,9 @@ struct cgit_environment {
 	const char *server_port;
 	const char *http_cookie;
 	const char *http_referer;
-	unsigned int content_length;
+	// A byte count, so size_t rather than a fixed width type that would
+	// wrap at four gigabytes on the way in.
+	size_t content_length;
 	int authenticated;
 };
 
@@ -326,97 +306,17 @@ struct cgit_context {
 	struct cgit_config cfg;
 	struct cgit_repo *repo;
 	struct cgit_page page;
-	// Set once the repository is known to hold no commits, so the page
-	// header can drop the tabs that cannot render anything.
+	// Set when the repository holds no commits, so the page header can drop
+	// the tabs that cannot render anything.
 	int empty_repo;
 };
 
-typedef int (*write_archive_fn_t)(const char *, const char *);
-
-struct cgit_snapshot_format {
-	const char *suffix;
-	const char *mimetype;
-	write_archive_fn_t write_func;
-};
-
 extern const char *cgit_version;
 
 extern struct cgit_repolist cgit_repolist;
 extern struct cgit_context ctx;
-extern const struct cgit_snapshot_format cgit_snapshot_formats[];
 
-extern char *cgit_default_repo_desc;
-extern struct cgit_repo *cgit_add_repo(const char *url);
-extern struct cgit_repo *cgit_get_repoinfo(const char *url);
 extern void cgit_repo_config(struct cgit_repo *repo, const char *name,
 			     const char *value);
 
-extern int cgit_die_unless_zero(int result, const char *msg);
-extern int cgit_die_unless_positive(int result, const char *msg);
-extern int cgit_die_unless_non_negative(int result, const char *msg);
-
-extern char *cgit_trim_end(const char *str, char c);
-extern char *cgit_ensure_end(const char *str, char c);
-
-extern void strbuf_ensure_end(struct strbuf *sb, char c);
-
-extern void cgit_add_ref(struct reflist *list, struct refinfo *ref);
-extern void cgit_free_reflist_inner(struct reflist *list);
-extern int cgit_refs_cb(const struct reference *ref, void *cb_data);
-
-extern void cgit_free_commitinfo(struct commitinfo *info);
-extern void cgit_free_taginfo(struct taginfo *info);
-
-void cgit_diff_tree_cb(struct diff_queue_struct *q,
-		       struct diff_options *options, void *data);
-
-extern int cgit_diff_files(const struct object_id *old_oid,
-			   const struct object_id *new_oid,
-			   unsigned long *old_size, unsigned long *new_size,
-			   int *binary, int context, int ignorews,
-			   linediff_fn fn);
-
-extern void cgit_diff_tree(const struct object_id *old_oid,
-			   const struct object_id *new_oid,
-			   filepair_fn fn, const char *prefix, int ignorews);
-
-extern void cgit_diff_commit(struct commit *commit, filepair_fn fn,
-			     const char *prefix);
-
-__attribute__((format (printf,1,2)))
-extern char *cgit_fmt(const char *format,...);
-
-__attribute__((format (printf,1,2)))
-extern char *cgit_fmtalloc(const char *format,...);
-
-extern struct commitinfo *cgit_parse_commit(struct commit *commit);
-extern struct taginfo *cgit_parse_tag(struct tag *tag);
-extern void cgit_parse_url(const char *url);
-
-extern const char *cgit_repobasename(const char *reponame);
-
-extern int cgit_parse_snapshots_mask(const char *str);
-extern void cgit_parse_date_format(const char *format, struct date_mode *mode);
-extern const struct object_id *cgit_snapshot_get_sig(const char *ref,
-						     const struct cgit_snapshot_format *f);
-extern unsigned cgit_snapshot_format_bit(const struct cgit_snapshot_format *f);
-
-extern int cgit_open_filter(struct cgit_filter *filter, ...);
-extern int cgit_close_filter(struct cgit_filter *filter);
-extern void cgit_fprintf_filter(struct cgit_filter *filter, FILE *f, const char *prefix);
-extern void cgit_exec_filter_init(struct cgit_exec_filter *filter, char *cmd, char **argv);
-extern struct cgit_filter *cgit_new_filter(const char *cmd, filter_type filtertype);
-extern void cgit_cleanup_filters(void);
-extern void cgit_init_filters(void);
-
-extern void cgit_prepare_repo_env(struct cgit_repo * repo);
-
-extern int cgit_read_first_line(const char *path, char **buf, size_t *size);
-
-extern char *cgit_strdup_first_line(const char *txt);
-
-extern char *cgit_expand_macros(const char *txt);
-
-extern char *cgit_get_mimetype_for_filename(const char *filename);
-
-#endif /* CGIT_H */
+#endif // CGIT_H
diff --git a/source/cgit.mk b/source/cgit.mk
index e0a3312..1cee1fb 100644
--- a/source/cgit.mk
+++ b/source/cgit.mk
@@ -1,34 +1,30 @@
-# This Makefile runs inside the bundled Git tree (vendor/git) so that it
-# can reuse Git's build variables and platform detection. The top-level
-# Makefile invokes it as:
+# The rules that compile and link cgit itself. They run with the bundled Git
+# tree as the working directory so that Git's build variables and platform
+# detection can be reused, which is why every path back into the project leads
+# through CGIT_ROOT. The top level Makefile invokes it roughly as
 #
 #     make -C vendor/git -f ../../source/cgit.mk ../../build/cgit
 #
-# so every path back into the project reaches through CGIT_ROOT ("../.."):
-# sources come from ../../source and all build output goes to ../../build.
+# so sources are read from ../../source and all output is written to
+# ../../build.
 include Makefile
 
-# Locations relative to vendor/git, where this file is run. SRCDIR and
-# BUILDDIR name the root-relative subdirs (matching the top-level Makefile and
-# used by the version recipe, which cds to the root); CGIT_SRC and CGIT_BUILD
-# are their full paths from here.
+# SRCDIR and BUILDDIR are named relative to the project root, matching the top
+# level Makefile, because the version recipe changes into the root before using
+# them. CGIT_SRC and CGIT_BUILD are the same two directories reached from
+# vendor/git, where everything else here runs.
 CGIT_ROOT  = ../..
 SRCDIR     = source
 BUILDDIR   = build
 CGIT_SRC   = $(CGIT_ROOT)/$(SRCDIR)
 CGIT_BUILD = $(CGIT_ROOT)/$(BUILDDIR)
 
-# Emit zero-initialised globals as plain definitions instead of common
-# symbols, the default everywhere but Apple clang. ld64 otherwise derives
-# a 32 KB alignment from the size of git's 64 KB packet_buffer and warns
-# about reducing it on every macOS link. Use override so a command-line
-# CFLAGS (as tools/release-build.sh passes) still keeps the flag.
-override CFLAGS += -fno-common
-
+# Read again here because a sub make inherits only the variables the top level
+# Makefile exports, which leaves out the build options this file reads.
 -include $(CGIT_ROOT)/cgit.conf
 
-# The CGIT_* variables are inherited from the top-level Makefile.
-
+# CGIT_VERSION and the other CGIT_ values used below come from the top level
+# Makefile, which exports them, rather than being defined in this file.
 $(CGIT_BUILD)/VERSION: force-version
 	@mkdir -p $(CGIT_BUILD)/
 	@cd $(CGIT_ROOT) && '$(SHELL_PATH_SQ)' $(SRCDIR)/gen-version.sh "$(CGIT_VERSION)" $(BUILDDIR)/VERSION
@@ -37,12 +33,12 @@ $(CGIT_BUILD)/VERSION: force-version
 
 # The language the cgit sources are written in. Both GCC and Clang default to
 # this today, so pinning it changes nothing now and stops the meaning of the
-# sources drifting when a compiler moves its default on (GCC 15 defaults to
-# gnu23). The GNU dialect rather than plain c17 because git's headers use GNU
-# extensions, and because dlsym cannot be used through a conforming cast.
-#
-# Only the cgit objects are held to this. Git keeps whatever its own build
-# decides, which on some platforms is a different standard again.
+# sources drifting when a compiler moves its default on, as GCC 15 did by
+# defaulting to gnu23. The GNU dialect rather than plain c17 because git's
+# headers use GNU extensions, and because dlsym cannot be used through a
+# conforming cast. Only the cgit objects are held to this, and Git keeps
+# whatever its own build decides, which on some platforms is a different
+# standard again.
 CGIT_STD ?= gnu17
 
 # CGIT_CFLAGS is tracked separately so that changing it does not force a
@@ -52,8 +48,8 @@ CGIT_CFLAGS += -DCGIT_CONFIG='"$(CGIT_CONFIG)"'
 CGIT_CFLAGS += -DCGIT_SCRIPT_NAME='"$(CGIT_SCRIPT_NAME)"'
 CGIT_CFLAGS += -DCGIT_CACHE_ROOT='"$(CACHE_ROOT)"'
 
-# Reaches only the cgit objects, so a caller can tighten the build (CI passes
-# -Werror) without holding git's own sources to the same standard.
+# Reaches only the cgit objects, so a caller can tighten the build, the way CI
+# passes -Werror, without holding git's own sources to the same standard.
 CGIT_CFLAGS += $(CGIT_EXTRA_CFLAGS)
 
 PKG_CONFIG ?= pkg-config
@@ -88,12 +84,15 @@ endif
 
 endif
 
-# Add -ldl to linker flags on systems that commonly use GNU libc.
+# The filters reach libc's write through dlsym, which lives in a library of its
+# own on the systems that use GNU libc.
 ifneq (,$(filter $(uname_S),Linux GNU GNU/kFreeBSD))
 	CGIT_LIBS += -ldl
 endif
 
-# glibc 2.1+ offers sendfile which the most common C library on Linux
+# The cache sends a stored page straight to stdout with sendfile where it can.
+# Only Linux is assumed to offer it, so everywhere else falls back to reading
+# and writing the file by hand.
 ifeq ($(uname_S),Linux)
 	HAVE_LINUX_SENDFILE = YesPlease
 endif
@@ -105,7 +104,7 @@ endif
 CGIT_OBJ_NAMES += cgit.o
 CGIT_OBJ_NAMES += cache.o
 CGIT_OBJ_NAMES += cmd.o
-CGIT_OBJ_NAMES += configfile.o
+CGIT_OBJ_NAMES += config.o
 CGIT_OBJ_NAMES += filter.o
 CGIT_OBJ_NAMES += html.o
 CGIT_OBJ_NAMES += parsing.o
@@ -140,8 +139,8 @@ $(CGIT_VERSION_OBJS): $(CGIT_BUILD)/VERSION
 $(CGIT_VERSION_OBJS): EXTRA_CPPFLAGS = \
 	-DCGIT_VERSION='"$(CGIT_VERSION)"'
 
-# Git handles dependencies using ":=" so dependencies in CGIT_OBJS are not
-# handled by that and we must handle them ourselves.
+# Git builds its list of dependency files with := before this file adds the
+# cgit objects, so those are missing from it and have to be picked up here.
 cgit_dep_files := $(foreach f,$(CGIT_OBJS),$(dir $f).depend/$(notdir $f).d)
 cgit_dep_files_present := $(wildcard $(cgit_dep_files))
 ifneq ($(cgit_dep_files_present),)
diff --git a/source/cmd.c b/source/cmd.c
index c17d8d9..456a937 100644
--- a/source/cmd.c
+++ b/source/cmd.c
@@ -1,15 +1,14 @@
-/* cmd.c: the cgit command dispatcher
- *
- * Copyright (C) 2006-2017 cgit Development Team <cgit@lists.zx2c4.com>
- *
- * Licensed under GNU General Public License v2
- *   (see LICENSE.txt for full license text)
+/*
+ * One thunk per page, each pulling the parts of the request its renderer
+ * needs out of the global ctx, plus the table that maps a page name onto its
+ * thunk. Keeping the thunks here lets every renderer take ordinary arguments
+ * instead of reaching into the request state itself.
  */
 
+#include "cache.h"
 #include "cgit.h"
 #include "cmd.h"
-#include "cache.h"
-#include "ui-shared.h"
+#include "html.h"
 #include "ui-atom.h"
 #include "ui-blame.h"
 #include "ui-blob.h"
@@ -21,50 +20,56 @@
 #include "ui-plain.h"
 #include "ui-refs.h"
 #include "ui-repolist.h"
+#include "ui-shared.h"
 #include "ui-snapshot.h"
 #include "ui-stats.h"
 #include "ui-summary.h"
 #include "ui-tag.h"
 #include "ui-tree.h"
-#define def_cmd(name, want_repo, want_vpath, is_clone) \
-	{#name, name##_fn, want_repo, want_vpath, is_clone}
 
-static void HEAD_fn(void)
+static void head_fn(void)
 {
 	cgit_clone_head();
 }
 
-static void atom_fn(void)
+static void about_fn(void)
 {
-	cgit_print_atom(ctx.qry.head, ctx.qry.path, ctx.cfg.max_atom_items);
+	char *currenturl, *redirect;
+	size_t path_info_len;
+
+	if (!ctx.repo) {
+		cgit_print_site_readme();
+		return;
+	}
+
+	// The about page resolves relative links against its own URL, so it
+	// only works with a trailing slash.
+	path_info_len = ctx.env.path_info ? strlen(ctx.env.path_info) : 0;
+	if (!ctx.qry.path &&
+	    (!ctx.qry.url || !*ctx.qry.url ||
+	     ctx.qry.url[strlen(ctx.qry.url) - 1] != '/') &&
+	    (!path_info_len || ctx.env.path_info[path_info_len - 1] != '/')) {
+		currenturl = cgit_currenturl();
+		redirect = cgit_fmtalloc("%s/", currenturl);
+		cgit_redirect(redirect, true);
+		free(currenturl);
+		free(redirect);
+	} else if (ctx.repo->readme.nr) {
+		cgit_print_repo_readme(ctx.qry.path);
+	} else if (ctx.repo->homepage) {
+		cgit_redirect(ctx.repo->homepage, false);
+	} else {
+		currenturl = cgit_currenturl();
+		redirect = cgit_fmtalloc("%s../", currenturl);
+		cgit_redirect(redirect, false);
+		free(currenturl);
+		free(redirect);
+	}
 }
 
-static void about_fn(void)
+static void atom_fn(void)
 {
-	if (ctx.repo) {
-		size_t path_info_len = ctx.env.path_info ? strlen(ctx.env.path_info) : 0;
-		if (!ctx.qry.path &&
-		    (!ctx.qry.url || !*ctx.qry.url ||
-		     ctx.qry.url[strlen(ctx.qry.url) - 1] != '/') &&
-		    (!path_info_len || ctx.env.path_info[path_info_len - 1] != '/')) {
-			char *currenturl = cgit_currenturl();
-			char *redirect = cgit_fmtalloc("%s/", currenturl);
-			cgit_redirect(redirect, true);
-			free(currenturl);
-			free(redirect);
-		} else if (ctx.repo->readme.nr)
-			cgit_print_repo_readme(ctx.qry.path);
-		else if (ctx.repo->homepage)
-			cgit_redirect(ctx.repo->homepage, false);
-		else {
-			char *currenturl = cgit_currenturl();
-			char *redirect = cgit_fmtalloc("%s../", currenturl);
-			cgit_redirect(redirect, false);
-			free(currenturl);
-			free(redirect);
-		}
-	} else
-		cgit_print_site_readme();
+	cgit_print_atom(ctx.qry.head, ctx.qry.path, ctx.cfg.max_atom_items);
 }
 
 static void blame_fn(void)
@@ -90,11 +95,6 @@ static void diff_fn(void)
 	cgit_print_diff(ctx.qry.oid, ctx.qry.oid2, ctx.qry.path, 1, 0);
 }
 
-static void rawdiff_fn(void)
-{
-	cgit_print_diff(ctx.qry.oid, ctx.qry.oid2, ctx.qry.path, 1, 1);
-}
-
 static void info_fn(void)
 {
 	cgit_clone_info();
@@ -110,8 +110,8 @@ static void log_fn(void)
 
 static void ls_cache_fn(void)
 {
-	/* The listing exposes the cache path and the URLs other visitors
-	 * requested, so it stays off unless an admin opts in. */
+	// The listing names the cache path and the URLs other visitors asked
+	// for, so it stays off until an administrator opts in.
 	if (!ctx.cfg.enable_cache_list) {
 		cgit_print_error_page(404, "Not found", "Not found");
 		return;
@@ -127,11 +127,6 @@ static void objects_fn(void)
 	cgit_clone_objects();
 }
 
-static void repolist_fn(void)
-{
-	cgit_print_repolist();
-}
-
 static void patch_fn(void)
 {
 	cgit_print_patch(ctx.qry.oid, ctx.qry.oid2, ctx.qry.path);
@@ -142,11 +137,21 @@ static void plain_fn(void)
 	cgit_print_plain();
 }
 
+static void rawdiff_fn(void)
+{
+	cgit_print_diff(ctx.qry.oid, ctx.qry.oid2, ctx.qry.path, 1, 1);
+}
+
 static void refs_fn(void)
 {
 	cgit_print_refs();
 }
 
+static void repolist_fn(void)
+{
+	cgit_print_repolist();
+}
+
 static void snapshot_fn(void)
 {
 	cgit_print_snapshot(ctx.qry.head, ctx.qry.oid, ctx.qry.path,
@@ -176,31 +181,34 @@ static void tree_fn(void)
 	cgit_print_tree(ctx.qry.oid, ctx.qry.path);
 }
 
-struct cgit_cmd *cgit_get_cmd(void)
-{
-	static struct cgit_cmd cmds[] = {
-		def_cmd(HEAD, 1, 0, 1),
-		def_cmd(atom, 1, 0, 0),
-		def_cmd(about, 0, 0, 0),
-		def_cmd(blame, 1, 1, 0),
-		def_cmd(blob, 1, 0, 0),
-		def_cmd(commit, 1, 1, 0),
-		def_cmd(diff, 1, 1, 0),
-		def_cmd(info, 1, 0, 1),
-		def_cmd(log, 1, 1, 0),
-		def_cmd(ls_cache, 0, 0, 0),
-		def_cmd(objects, 1, 0, 1),
-		def_cmd(patch, 1, 1, 0),
-		def_cmd(plain, 1, 0, 0),
-		def_cmd(rawdiff, 1, 1, 0),
-		def_cmd(refs, 1, 0, 0),
-		def_cmd(repolist, 0, 0, 0),
-		def_cmd(snapshot, 1, 0, 0),
-		def_cmd(stats, 1, 1, 0),
-		def_cmd(summary, 1, 0, 0),
-		def_cmd(tag, 1, 0, 0),
-		def_cmd(tree, 1, 1, 0),
-	};
+// The name is what a request spells as p=, so these strings are part of the
+// URL space and cannot be renamed.
+static const struct cgit_cmd commands[] = {
+	{ .name = "HEAD", .fn = head_fn, .want_repo = 1, .is_clone = 1 },
+	{ .name = "about", .fn = about_fn },
+	{ .name = "atom", .fn = atom_fn, .want_repo = 1 },
+	{ .name = "blame", .fn = blame_fn, .want_repo = 1, .want_vpath = 1 },
+	{ .name = "blob", .fn = blob_fn, .want_repo = 1 },
+	{ .name = "commit", .fn = commit_fn, .want_repo = 1, .want_vpath = 1 },
+	{ .name = "diff", .fn = diff_fn, .want_repo = 1, .want_vpath = 1 },
+	{ .name = "info", .fn = info_fn, .want_repo = 1, .is_clone = 1 },
+	{ .name = "log", .fn = log_fn, .want_repo = 1, .want_vpath = 1 },
+	{ .name = "ls_cache", .fn = ls_cache_fn },
+	{ .name = "objects", .fn = objects_fn, .want_repo = 1, .is_clone = 1 },
+	{ .name = "patch", .fn = patch_fn, .want_repo = 1, .want_vpath = 1 },
+	{ .name = "plain", .fn = plain_fn, .want_repo = 1 },
+	{ .name = "rawdiff", .fn = rawdiff_fn, .want_repo = 1, .want_vpath = 1 },
+	{ .name = "refs", .fn = refs_fn, .want_repo = 1 },
+	{ .name = "repolist", .fn = repolist_fn },
+	{ .name = "snapshot", .fn = snapshot_fn, .want_repo = 1 },
+	{ .name = "stats", .fn = stats_fn, .want_repo = 1, .want_vpath = 1 },
+	{ .name = "summary", .fn = summary_fn, .want_repo = 1 },
+	{ .name = "tag", .fn = tag_fn, .want_repo = 1 },
+	{ .name = "tree", .fn = tree_fn, .want_repo = 1, .want_vpath = 1 },
+};
+
+const struct cgit_cmd *cgit_get_cmd(void)
+{
 	size_t i;
 
 	if (ctx.qry.page == NULL) {
@@ -210,8 +218,8 @@ struct cgit_cmd *cgit_get_cmd(void)
 			ctx.qry.page = "repolist";
 	}
 
-	for (i = 0; i < ARRAY_SIZE(cmds); i++)
-		if (!strcmp(ctx.qry.page, cmds[i].name))
-			return &cmds[i];
+	for (i = 0; i < ARRAY_SIZE(commands); i++)
+		if (!strcmp(ctx.qry.page, commands[i].name))
+			return &commands[i];
 	return NULL;
 }
diff --git a/source/cmd.h b/source/cmd.h
index 6249b1d..7299028 100644
--- a/source/cmd.h
+++ b/source/cmd.h
@@ -1,5 +1,12 @@
-#ifndef CMD_H
-#define CMD_H
+/*
+ * The dispatch table that maps the page name in a request to the function
+ * rendering it. Each entry also records what the request needs resolved
+ * first, so the caller can look up the repository or the path within the tree
+ * and reject a bad request once, rather than every page repeating the check.
+ */
+
+#ifndef CGIT_CMD_H
+#define CGIT_CMD_H
 
 typedef void (*cgit_cmd_fn)(void);
 
@@ -11,6 +18,7 @@ struct cgit_cmd {
 		is_clone:1;
 };
 
-extern struct cgit_cmd *cgit_get_cmd(void);
+// The entry naming ctx.qry.page, or NULL when no page goes by that name.
+extern const struct cgit_cmd *cgit_get_cmd(void);
 
-#endif /* CMD_H */
+#endif // CGIT_CMD_H
diff --git a/source/config.c b/source/config.c
new file mode 100644
index 0000000..6a5d136
--- /dev/null
+++ b/source/config.c
@@ -0,0 +1,113 @@
+/*
+ * The reader for cgit's config files, which are the main cgitrc, any file it
+ * pulls in with an include line, the cgitrc that sits beside a scanned
+ * repository, and the cached repolist. A file is a sequence of name=value
+ * lines, and each pair is handed to a callback that decides what it means, so
+ * nothing here knows a single key by name. A value runs to the end of its line
+ * and keeps whatever spacing it has, there is no quoting. A line whose first
+ * non blank character is a hash or a semicolon is a comment, and blank lines
+ * are ignored.
+ */
+
+#include "cgit.h"
+#include "config.h"
+
+#define MAX_INCLUDE_NESTING 8
+
+/*
+ * Folds away a carriage return that sits right before a newline so a file
+ * saved with DOS line endings parses the same as one saved with Unix endings.
+ */
+static int next_char(FILE *f)
+{
+	int c = fgetc(f);
+
+	if (c == '\r') {
+		c = fgetc(f);
+		if (c != '\n') {
+			ungetc(c, f);
+			c = '\r';
+		}
+	}
+	return c;
+}
+
+static void skip_line(FILE *f)
+{
+	int c;
+
+	while ((c = next_char(f)) && c != '\n' && c != EOF)
+		;
+}
+
+static int next_content_char(FILE *f)
+{
+	int c = next_char(f);
+
+	for (;;) {
+		if (c == EOF)
+			return EOF;
+		if (c == '#' || c == ';')
+			skip_line(f);
+		else if (!isspace(c))
+			return c;
+		c = next_char(f);
+	}
+}
+
+static int read_entry(FILE *f, struct strbuf *name, struct strbuf *value)
+{
+	int c;
+
+	strbuf_reset(value);
+
+	for (;;) {
+		c = next_content_char(f);
+		if (c == EOF)
+			return 0;
+
+		strbuf_reset(name);
+		while (c != '=' && c != '\n' && c != EOF) {
+			strbuf_addch(name, c);
+			c = next_char(f);
+		}
+		if (c == EOF)
+			return 0;
+		// Dropping just the line with no equals sign keeps one typo
+		// from hiding the rest of the file.
+		if (c != '=')
+			continue;
+
+		c = next_char(f);
+		while (c != '\n' && c != EOF) {
+			strbuf_addch(value, c);
+			c = next_char(f);
+		}
+
+		return 1;
+	}
+}
+
+int config_file_parse(const char *filename, config_file_value_fn fn)
+{
+	static int nesting;
+	struct strbuf name = STRBUF_INIT;
+	struct strbuf value = STRBUF_INIT;
+	FILE *f;
+
+	// An include line calls back into here, so a file that includes itself,
+	// directly or round a longer loop, would recurse until the stack gave
+	// out.
+	if (nesting > MAX_INCLUDE_NESTING)
+		return -1;
+	if (!(f = fopen(filename, "r")))
+		return -1;
+	nesting++;
+	while (read_entry(f, &name, &value))
+		fn(name.buf, value.buf);
+	nesting--;
+	fclose(f);
+	strbuf_release(&name);
+	strbuf_release(&value);
+	return 0;
+}
diff --git a/source/config.h b/source/config.h
new file mode 100644
index 0000000..ab42e2f
--- /dev/null
+++ b/source/config.h
@@ -0,0 +1,16 @@
+/*
+ * The interface to cgit's config file reader. A caller names a file and passes
+ * a callback that receives every name and value in it, in the order they
+ * appear.
+ */
+
+#ifndef CGIT_CONFIG_H
+#define CGIT_CONFIG_H
+
+#include "cgit.h"
+
+typedef void (*config_file_value_fn)(const char *name, const char *value);
+
+extern int config_file_parse(const char *filename, config_file_value_fn fn);
+
+#endif // CGIT_CONFIG_H
diff --git a/source/configfile.c b/source/configfile.c
deleted file mode 100644
index 9bf2da1..0000000
--- a/source/configfile.c
+++ /dev/null
@@ -1,100 +0,0 @@
-/* configfile.c: parsing of config files
- *
- * Copyright (C) 2006-2014 cgit Development Team <cgit@lists.zx2c4.com>
- *
- * Licensed under GNU General Public License v2
- *   (see LICENSE.txt for full license text)
- */
-
-#include <git-compat-util.h>
-#include "configfile.h"
-
-static int next_char(FILE *f)
-{
-	int c = fgetc(f);
-	if (c == '\r') {
-		c = fgetc(f);
-		if (c != '\n') {
-			ungetc(c, f);
-			c = '\r';
-		}
-	}
-	return c;
-}
-
-static void skip_line(FILE *f)
-{
-	int c;
-
-	while ((c = next_char(f)) && c != '\n' && c != EOF)
-		;
-}
-
-static int read_config_line(FILE *f, struct strbuf *name, struct strbuf *value)
-{
-	int c;
-
-	strbuf_reset(value);
-
-	for (;;) {
-		c = next_char(f);
-
-		// Skip comments and preceding spaces.
-		for (;;) {
-			if (c == EOF)
-				return 0;
-			else if (c == '#' || c == ';')
-				skip_line(f);
-			else if (!isspace(c))
-				break;
-			c = next_char(f);
-		}
-
-		// Read variable name.
-		strbuf_reset(name);
-		while (c != '=') {
-			if (c == EOF)
-				return 0;
-			// A line without '=' is malformed. Skip it and keep
-			// reading so a typo does not drop the rest of the file.
-			if (c == '\n')
-				break;
-			strbuf_addch(name, c);
-			c = next_char(f);
-		}
-		if (c != '=')
-			continue;
-
-		// Read variable value.
-		c = next_char(f);
-		while (c != '\n' && c != EOF) {
-			strbuf_addch(value, c);
-			c = next_char(f);
-		}
-
-		return 1;
-	}
-}
-
-int parse_configfile(const char *filename, configfile_value_fn fn)
-{
-	static int nesting;
-	struct strbuf name = STRBUF_INIT;
-	struct strbuf value = STRBUF_INIT;
-	FILE *f;
-
-	/* cancel deeply nested include-commands */
-	if (nesting > 8)
-		return -1;
-	if (!(f = fopen(filename, "r")))
-		return -1;
-	nesting++;
-	while (read_config_line(f, &name, &value))
-		fn(name.buf, value.buf);
-	nesting--;
-	fclose(f);
-	strbuf_release(&name);
-	strbuf_release(&value);
-	return 0;
-}
-
diff --git a/source/configfile.h b/source/configfile.h
deleted file mode 100644
index af7ca19..0000000
--- a/source/configfile.h
+++ /dev/null
@@ -1,10 +0,0 @@
-#ifndef CONFIGFILE_H
-#define CONFIGFILE_H
-
-#include "cgit.h"
-
-typedef void (*configfile_value_fn)(const char *name, const char *value);
-
-extern int parse_configfile(const char *filename, configfile_value_fn fn);
-
-#endif /* CONFIGFILE_H */
diff --git a/source/filter.c b/source/filter.c
index f195534..3c0c7e4 100644
--- a/source/filter.c
+++ b/source/filter.c
@@ -1,13 +1,16 @@
-/* filter.c: filter framework functions
- *
- * Copyright (C) 2006-2014 cgit Development Team <cgit@lists.zx2c4.com>
- *
- * Licensed under GNU General Public License v2
- *   (see LICENSE.txt for full license text)
+/*
+ * The two backends behind a cgit filter and the plumbing that hands stdout to
+ * whichever one is open. An exec filter forks a program and feeds it the page
+ * through a pipe, while a Lua filter runs a script inside cgit and feeds it
+ * the page through a write that has been interposed in front of libc's. A
+ * cgitrc command picks its backend with an exec or lua prefix, and a command
+ * with no prefix is taken as a program to run.
  */
 
 #include "cgit.h"
+#include "filter.h"
 #include "html.h"
+#include "shared.h"
 #ifndef NO_LUA
 #include <dlfcn.h>
 #include <lua.h>
@@ -15,32 +18,10 @@
 #include <lauxlib.h>
 #endif
 
-static inline void reap_filter(struct cgit_filter *filter)
-{
-	if (filter && filter->cleanup)
-		filter->cleanup(filter);
-}
-
-void cgit_cleanup_filters(void)
-{
-	int i;
-	reap_filter(ctx.cfg.about_filter);
-	reap_filter(ctx.cfg.commit_filter);
-	reap_filter(ctx.cfg.source_filter);
-	reap_filter(ctx.cfg.email_filter);
-	reap_filter(ctx.cfg.auth_filter);
-	for (i = 0; i < cgit_repolist.count; ++i) {
-		reap_filter(cgit_repolist.repos[i].about_filter);
-		reap_filter(cgit_repolist.repos[i].commit_filter);
-		reap_filter(cgit_repolist.repos[i].source_filter);
-		reap_filter(cgit_repolist.repos[i].email_filter);
-	}
-}
-
 static int open_exec_filter(struct cgit_filter *base, va_list ap)
 {
 	struct cgit_exec_filter *filter = (struct cgit_exec_filter *)base;
-	int pipe_fh[2];
+	int pipefd[2];
 	int i;
 
 	for (i = 0; i < filter->base.argument_count; i++)
@@ -48,19 +29,21 @@ static int open_exec_filter(struct cgit_filter *base, va_list ap)
 
 	filter->old_stdout = cgit_die_unless_positive(dup(STDOUT_FILENO),
 		"Unable to duplicate STDOUT");
-	cgit_die_unless_zero(pipe(pipe_fh), "Unable to create pipe to subprocess");
-	filter->pid = cgit_die_unless_non_negative(fork(), "Unable to create subprocess");
+	cgit_die_unless_zero(pipe(pipefd),
+		"Unable to create pipe to subprocess");
+	filter->pid = cgit_die_unless_non_negative(fork(),
+		"Unable to create subprocess");
 	if (filter->pid == 0) {
-		close(pipe_fh[1]);
-		cgit_die_unless_non_negative(dup2(pipe_fh[0], STDIN_FILENO),
+		close(pipefd[1]);
+		cgit_die_unless_non_negative(dup2(pipefd[0], STDIN_FILENO),
 			"Unable to use pipe as STDIN");
 		execvp(filter->cmd, filter->argv);
 		die_errno("Unable to exec subprocess %s", filter->cmd);
 	}
-	close(pipe_fh[0]);
-	cgit_die_unless_non_negative(dup2(pipe_fh[1], STDOUT_FILENO),
+	close(pipefd[0]);
+	cgit_die_unless_non_negative(dup2(pipefd[1], STDOUT_FILENO),
 		"Unable to use pipe as STDOUT");
-	close(pipe_fh[1]);
+	close(pipefd[1]);
 	return 0;
 }
 
@@ -83,10 +66,10 @@ done:
 	for (i = 0; i < filter->base.argument_count; i++)
 		filter->argv[i + 1] = NULL;
 	return WEXITSTATUS(exit_status);
-
 }
 
-static void fprintf_exec_filter(struct cgit_filter *base, FILE *f, const char *prefix)
+static void fprintf_exec_filter(struct cgit_filter *base, FILE *f,
+				const char *prefix)
 {
 	struct cgit_exec_filter *filter = (struct cgit_exec_filter *)base;
 	fprintf(f, "%sexec:%s\n", prefix, filter->cmd);
@@ -107,21 +90,23 @@ static void cleanup_exec_filter(struct cgit_filter *base)
 
 static struct cgit_filter *new_exec_filter(const char *cmd, int argument_count)
 {
-	struct cgit_exec_filter *f;
-	int args_size = 0;
+	struct cgit_exec_filter *filter;
+	int argv_size;
 
-	f = xmalloc(sizeof(*f));
-	/* We leave argv for now and assign it below. */
-	cgit_exec_filter_init(f, cgit_strdup_first_line(cmd), NULL);
-	f->base.argument_count = argument_count;
-	args_size = (2 + argument_count) * sizeof(char *);
-	f->argv = xmalloc(args_size);
-	memset(f->argv, 0, args_size);
-	f->argv[0] = f->cmd;
-	return &f->base;
+	filter = xmalloc(sizeof(*filter));
+	cgit_exec_filter_init(filter, cgit_strdup_first_line(cmd), NULL);
+	filter->base.argument_count = argument_count;
+	// argv is the command, then a slot per argument, then the NULL that
+	// execvp needs to find the end.
+	argv_size = (2 + argument_count) * sizeof(char *);
+	filter->argv = xmalloc(argv_size);
+	memset(filter->argv, 0, argv_size);
+	filter->argv[0] = filter->cmd;
+	return &filter->base;
 }
 
-void cgit_exec_filter_init(struct cgit_exec_filter *filter, char *cmd, char **argv)
+void cgit_exec_filter_init(struct cgit_exec_filter *filter, char *cmd,
+			   char **argv)
 {
 	memset(filter, 0, sizeof(*filter));
 	filter->base.open = open_exec_filter;
@@ -130,7 +115,6 @@ void cgit_exec_filter_init(struct cgit_exec_filter *filter, char *cmd, char **ar
 	filter->base.cleanup = cleanup_exec_filter;
 	filter->cmd = cmd;
 	filter->argv = argv;
-	/* The argument count for open_filter is zero by default, unless called from new_filter, above. */
 	filter->base.argument_count = 0;
 }
 
@@ -141,8 +125,17 @@ void cgit_init_filters(void)
 #endif
 
 #ifndef NO_LUA
+struct lua_filter {
+	struct cgit_filter base;
+	char *script_file;
+	lua_State *lua_state;
+};
+
+typedef ssize_t (*filter_write_fn)(struct cgit_filter *base, const void *buf,
+				   size_t count);
+
 static ssize_t (*libc_write)(int fd, const void *buf, size_t count);
-static ssize_t (*filter_write)(struct cgit_filter *base, const void *buf, size_t count) = NULL;
+static filter_write_fn filter_write = NULL;
 static struct cgit_filter *current_write_filter = NULL;
 
 void cgit_init_filters(void)
@@ -152,6 +145,10 @@ void cgit_init_filters(void)
 		die("Could not locate libc's write function");
 }
 
+/*
+ * Interposing on write is what lets a Lua filter see output produced by code
+ * that has no idea a filter is running.
+ */
 ssize_t write(int fd, const void *buf, size_t count)
 {
 	if (fd != STDOUT_FILENO || !filter_write)
@@ -159,13 +156,15 @@ ssize_t write(int fd, const void *buf, size_t count)
 	return filter_write(current_write_filter, buf, count);
 }
 
-static inline void hook_write(struct cgit_filter *filter, ssize_t (*new_write)(struct cgit_filter *base, const void *buf, size_t count))
+static inline void hook_write(struct cgit_filter *filter,
+			      filter_write_fn write_fn)
 {
-	/* We want to avoid buggy nested patterns. */
+	// Filters cannot nest, because there is one stdout and one hook, so a
+	// second one would strand the first.
 	assert(filter_write == NULL);
 	assert(current_write_filter == NULL);
 	current_write_filter = filter;
-	filter_write = new_write;
+	filter_write = write_fn;
 }
 
 static inline void unhook_write(void)
@@ -176,102 +175,113 @@ static inline void unhook_write(void)
 	current_write_filter = NULL;
 }
 
-struct lua_filter {
-	struct cgit_filter base;
-	char *script_file;
-	lua_State *lua_state;
-};
-
-static void error_lua_filter(struct lua_filter *filter)
+static void die_lua_error(struct lua_filter *filter)
 {
-	die("Lua error in %s: %s", filter->script_file, lua_tostring(filter->lua_state, -1));
+	die("Lua error in %s: %s", filter->script_file,
+	    lua_tostring(filter->lua_state, -1));
 	lua_pop(filter->lua_state, 1);
 }
 
-static ssize_t write_lua_filter(struct cgit_filter *base, const void *buf, size_t count)
+static ssize_t write_lua_filter(struct cgit_filter *base, const void *buf,
+				size_t count)
 {
 	struct lua_filter *filter = (struct lua_filter *)base;
 
 	lua_getglobal(filter->lua_state, "filter_write");
 	lua_pushlstring(filter->lua_state, buf, count);
 	if (lua_pcall(filter->lua_state, 1, 0, 0)) {
-		error_lua_filter(filter);
+		die_lua_error(filter);
 		errno = EIO;
 		return -1;
 	}
 	return count;
 }
 
-static inline int hook_lua_filter(lua_State *lua_state, void (*fn)(const char *txt))
+/*
+ * Output a script asks for belongs on the page and not back in its own filter,
+ * so the hook comes off around the call. The zero returned is Lua's count of
+ * values pushed for the script, not a success code.
+ */
+static inline int emit_unfiltered(lua_State *lua_state,
+				  void (*emit)(const char *text))
 {
-	const char *str;
-	ssize_t (*save_filter_write)(struct cgit_filter *base, const void *buf, size_t count);
-	struct cgit_filter *save_filter;
+	const char *text;
+	filter_write_fn saved_write;
+	struct cgit_filter *saved_filter;
 
-	str = lua_tostring(lua_state, 1);
-	if (!str)
+	text = lua_tostring(lua_state, 1);
+	if (!text)
 		return 0;
 
-	save_filter_write = filter_write;
-	save_filter = current_write_filter;
+	saved_write = filter_write;
+	saved_filter = current_write_filter;
 	unhook_write();
-	fn(str);
-	// fn buffers, so empty it while the hook is still off. Re-hooking first
-	// would send the page's own bytes back into the filter that asked for
-	// them to be written.
+	emit(text);
+	// emit leaves bytes in the html buffer, so empty it while the hook is
+	// still off. Re-hooking first would send the page's own bytes back into
+	// the filter.
 	html_flush();
-	hook_write(save_filter, save_filter_write);
+	hook_write(saved_filter, saved_write);
 
 	return 0;
 }
 
-static int html_lua_filter(lua_State *lua_state)
+static int script_html(lua_State *lua_state)
 {
-	return hook_lua_filter(lua_state, html);
+	return emit_unfiltered(lua_state, html);
 }
 
-static int html_txt_lua_filter(lua_State *lua_state)
+static int script_html_txt(lua_State *lua_state)
 {
-	return hook_lua_filter(lua_state, html_txt);
+	return emit_unfiltered(lua_state, html_txt);
 }
 
-static int html_attr_lua_filter(lua_State *lua_state)
+static int script_html_attr(lua_State *lua_state)
 {
-	return hook_lua_filter(lua_state, html_attr);
+	return emit_unfiltered(lua_state, html_attr);
 }
 
-static int html_url_path_lua_filter(lua_State *lua_state)
+static int script_html_url_path(lua_State *lua_state)
 {
-	return hook_lua_filter(lua_state, html_url_path);
+	return emit_unfiltered(lua_state, html_url_path);
 }
 
-static int html_url_arg_lua_filter(lua_State *lua_state)
+static int script_html_url_arg(lua_State *lua_state)
 {
-	return hook_lua_filter(lua_state, html_url_arg);
+	return emit_unfiltered(lua_state, html_url_arg);
 }
 
-static int html_include_lua_filter(lua_State *lua_state)
+/*
+ * html_include returns whether it found the file, and calling it through a
+ * pointer that claims it returns nothing would be undefined, so the result is
+ * dropped in a wrapper of the right shape.
+ */
+static void include_and_discard_result(const char *filename)
 {
-	return hook_lua_filter(lua_state, (void (*)(const char *))html_include);
+	html_include(filename);
 }
 
-static void cleanup_lua_filter(struct cgit_filter *base)
+static int script_html_include(lua_State *lua_state)
 {
-	struct lua_filter *filter = (struct lua_filter *)base;
-
-	if (!filter->lua_state)
-		return;
-
-	lua_close(filter->lua_state);
-	filter->lua_state = NULL;
-	if (filter->script_file) {
-		free(filter->script_file);
-		filter->script_file = NULL;
-	}
+	return emit_unfiltered(lua_state, include_and_discard_result);
 }
 
+static const struct {
+	const char *name;
+	lua_CFunction fn;
+} script_globals[] = {
+	{ "html", script_html },
+	{ "html_txt", script_html_txt },
+	{ "html_attr", script_html_attr },
+	{ "html_url_path", script_html_url_path },
+	{ "html_url_arg", script_html_url_arg },
+	{ "html_include", script_html_include },
+};
+
 static int init_lua_filter(struct lua_filter *filter)
 {
+	size_t i;
+
 	if (filter->lua_state)
 		return 0;
 
@@ -280,21 +290,13 @@ static int init_lua_filter(struct lua_filter *filter)
 
 	luaL_openlibs(filter->lua_state);
 
-	lua_pushcfunction(filter->lua_state, html_lua_filter);
-	lua_setglobal(filter->lua_state, "html");
-	lua_pushcfunction(filter->lua_state, html_txt_lua_filter);
-	lua_setglobal(filter->lua_state, "html_txt");
-	lua_pushcfunction(filter->lua_state, html_attr_lua_filter);
-	lua_setglobal(filter->lua_state, "html_attr");
-	lua_pushcfunction(filter->lua_state, html_url_path_lua_filter);
-	lua_setglobal(filter->lua_state, "html_url_path");
-	lua_pushcfunction(filter->lua_state, html_url_arg_lua_filter);
-	lua_setglobal(filter->lua_state, "html_url_arg");
-	lua_pushcfunction(filter->lua_state, html_include_lua_filter);
-	lua_setglobal(filter->lua_state, "html_include");
+	for (i = 0; i < ARRAY_SIZE(script_globals); i++) {
+		lua_pushcfunction(filter->lua_state, script_globals[i].fn);
+		lua_setglobal(filter->lua_state, script_globals[i].name);
+	}
 
 	if (luaL_dofile(filter->lua_state, filter->script_file)) {
-		error_lua_filter(filter);
+		die_lua_error(filter);
 		lua_close(filter->lua_state);
 		filter->lua_state = NULL;
 		return 1;
@@ -316,7 +318,7 @@ static int open_lua_filter(struct cgit_filter *base, va_list ap)
 	for (i = 0; i < filter->base.argument_count; ++i)
 		lua_pushstring(filter->lua_state, va_arg(ap, char *));
 	if (lua_pcall(filter->lua_state, filter->base.argument_count, 0, 0)) {
-		error_lua_filter(filter);
+		die_lua_error(filter);
 		return 1;
 	}
 	return 0;
@@ -329,7 +331,7 @@ static int close_lua_filter(struct cgit_filter *base)
 
 	lua_getglobal(filter->lua_state, "filter_close");
 	if (lua_pcall(filter->lua_state, 0, 1, 0)) {
-		error_lua_filter(filter);
+		die_lua_error(filter);
 		ret = -1;
 	} else {
 		ret = lua_tonumber(filter->lua_state, -1);
@@ -340,12 +342,27 @@ static int close_lua_filter(struct cgit_filter *base)
 	return ret;
 }
 
-static void fprintf_lua_filter(struct cgit_filter *base, FILE *f, const char *prefix)
+static void fprintf_lua_filter(struct cgit_filter *base, FILE *f,
+			       const char *prefix)
 {
 	struct lua_filter *filter = (struct lua_filter *)base;
 	fprintf(f, "%slua:%s\n", prefix, filter->script_file);
 }
 
+static void cleanup_lua_filter(struct cgit_filter *base)
+{
+	struct lua_filter *filter = (struct lua_filter *)base;
+
+	if (!filter->lua_state)
+		return;
+
+	lua_close(filter->lua_state);
+	filter->lua_state = NULL;
+	if (filter->script_file) {
+		free(filter->script_file);
+		filter->script_file = NULL;
+	}
+}
 
 static struct cgit_filter *new_lua_filter(const char *cmd, int argument_count)
 {
@@ -362,10 +379,8 @@ static struct cgit_filter *new_lua_filter(const char *cmd, int argument_count)
 
 	return &filter->base;
 }
-
 #endif
 
-
 int cgit_open_filter(struct cgit_filter *filter, ...)
 {
 	int result;
@@ -391,16 +406,37 @@ int cgit_close_filter(struct cgit_filter *filter)
 	return filter->close(filter);
 }
 
-void cgit_fprintf_filter(struct cgit_filter *filter, FILE *f, const char *prefix)
+void cgit_fprintf_filter(struct cgit_filter *filter, FILE *f,
+			 const char *prefix)
 {
 	filter->fprintfp(filter, f, prefix);
 }
 
+static inline void cleanup_filter(struct cgit_filter *filter)
+{
+	if (filter && filter->cleanup)
+		filter->cleanup(filter);
+}
 
+void cgit_cleanup_filters(void)
+{
+	int i;
+	cleanup_filter(ctx.cfg.about_filter);
+	cleanup_filter(ctx.cfg.commit_filter);
+	cleanup_filter(ctx.cfg.source_filter);
+	cleanup_filter(ctx.cfg.email_filter);
+	cleanup_filter(ctx.cfg.auth_filter);
+	for (i = 0; i < cgit_repolist.count; ++i) {
+		cleanup_filter(cgit_repolist.repos[i].about_filter);
+		cleanup_filter(cgit_repolist.repos[i].commit_filter);
+		cleanup_filter(cgit_repolist.repos[i].source_filter);
+		cleanup_filter(cgit_repolist.repos[i].email_filter);
+	}
+}
 
 static const struct {
 	const char *prefix;
-	struct cgit_filter *(*ctor)(const char *cmd, int argument_count);
+	struct cgit_filter *(*create)(const char *cmd, int argument_count);
 } filter_specs[] = {
 	{ "exec", new_exec_filter },
 #ifndef NO_LUA
@@ -419,42 +455,38 @@ struct cgit_filter *cgit_new_filter(const char *cmd, filter_type filtertype)
 		return NULL;
 
 	colon = strchr(cmd, ':');
-	len = colon - cmd;
-	/*
-	 * In case we're running on Windows, don't allow a single letter before
-	 * the colon.
-	 */
+	// Measured only once there is a colon to measure against, because
+	// subtracting from a null pointer is not something C defines.
+	len = colon ? (size_t)(colon - cmd) : 0;
+	// A single letter before the colon is a Windows drive letter, not a
+	// filter prefix.
 	if (len == 1)
 		colon = NULL;
 
 	switch (filtertype) {
-		case AUTH:
-			argument_count = 12;
-			break;
-
-		case EMAIL:
-			argument_count = 2;
-			break;
-
-		case SOURCE:
-		case ABOUT:
-			argument_count = 1;
-			break;
-
-		case COMMIT:
-		default:
-			argument_count = 0;
-			break;
+	case AUTH:
+		argument_count = 12;
+		break;
+	case EMAIL:
+		argument_count = 2;
+		break;
+	case SOURCE:
+	case ABOUT:
+		argument_count = 1;
+		break;
+	case COMMIT:
+	default:
+		argument_count = 0;
+		break;
 	}
 
-	/* If no prefix is given, exec filter is the default. */
 	if (!colon)
 		return new_exec_filter(cmd, argument_count);
 
 	for (i = 0; i < ARRAY_SIZE(filter_specs); i++) {
 		if (len == strlen(filter_specs[i].prefix) &&
 		    !strncmp(filter_specs[i].prefix, cmd, len))
-			return filter_specs[i].ctor(colon + 1, argument_count);
+			return filter_specs[i].create(colon + 1, argument_count);
 	}
 
 	die("Invalid filter type: %.*s", (int) len, cmd);
diff --git a/source/filter.h b/source/filter.h
new file mode 100644
index 0000000..861578c
--- /dev/null
+++ b/source/filter.h
@@ -0,0 +1,54 @@
+/*
+ * Output filters, the hook that lets a site pass page fragments through an
+ * external program or an embedded Lua script on their way to the browser. A
+ * filter takes over stdout between an open and a close call, so at most one
+ * can be active at a time. The exec and Lua backends both hide behind the
+ * struct cgit_filter call table, so callers never learn which is in use.
+ */
+
+#ifndef CGIT_FILTER_H
+#define CGIT_FILTER_H
+
+#include <stdarg.h>
+#include <stdio.h>
+
+typedef enum {
+	ABOUT, COMMIT, SOURCE, EMAIL, AUTH
+} filter_type;
+
+struct cgit_filter {
+	int (*open)(struct cgit_filter *, va_list ap);
+	int (*close)(struct cgit_filter *);
+	void (*fprintfp)(struct cgit_filter *, FILE *, const char *prefix);
+	void (*cleanup)(struct cgit_filter *);
+	int argument_count;
+};
+
+struct cgit_exec_filter {
+	struct cgit_filter base;
+	char *cmd;
+	char **argv;
+	int old_stdout;
+	int pid;
+};
+
+extern struct cgit_filter *cgit_new_filter(const char *cmd,
+					   filter_type filtertype);
+extern void cgit_exec_filter_init(struct cgit_exec_filter *filter, char *cmd,
+				  char **argv);
+
+/*
+ * Redirect stdout into the filter and pass it the variadic arguments its type
+ * expects, then hand stdout back and return the filter's exit status.
+ */
+extern int cgit_open_filter(struct cgit_filter *filter, ...);
+extern int cgit_close_filter(struct cgit_filter *filter);
+
+// Describe the filter the way cgitrc would have configured it.
+extern void cgit_fprintf_filter(struct cgit_filter *filter, FILE *f,
+				const char *prefix);
+
+extern void cgit_init_filters(void);
+extern void cgit_cleanup_filters(void);
+
+#endif // CGIT_FILTER_H
diff --git a/source/gen-version.sh b/source/gen-version.sh
index fb7d459..b28e987 100755
--- a/source/gen-version.sh
+++ b/source/gen-version.sh
@@ -1,27 +1,27 @@
 #!/bin/sh
+# Writes a "CGIT_VERSION = ..." line into the output file, which the build
+# then includes. Run it from the project root, as gen-version.sh <fallback>
+# [output], so that git describe reports cgit rather than the bundled git
+# submodule. The output file defaults to VERSION.
 
-# Writes a "CGIT_VERSION = ..." line to the output file (default: VERSION),
-# for the build to include. Run from the project root so `git describe` acts
-# on cgit rather than the bundled Git submodule.
-#
-# Usage: gen-version.sh <fallback-version> [output-file]
+version=$1
+output=${2:-VERSION}
 
-V=$1
-OUT=${2:-VERSION}
-
-# Prefer `git describe` when run at the top of the cgit working tree,
-# keeping the fallback when it fails, as in a shallow tagless clone.
+# Prefer what git describe reports, but only at the top of the cgit working
+# tree, and keep the fallback when it has nothing to say, as in a shallow or
+# tagless clone.
 if test "$(git rev-parse --git-dir 2>/dev/null)" = '.git'
 then
-	D=$(git describe --abbrev=4 HEAD 2>/dev/null)
-	test -n "$D" && V=$D
+	described=$(git describe --abbrev=4 HEAD 2>/dev/null)
+	test -n "$described" && version=$described
 fi
 
-new="CGIT_VERSION = $V"
-old=$(cat "$OUT" 2>/dev/null)
+new="CGIT_VERSION = $version"
+old=$(cat "$output" 2>/dev/null)
 
-# Exit if the version is already up to date.
+# Leaving the file untouched when nothing changed keeps make from rebuilding
+# everything that depends on it.
 test "$old" = "$new" && exit 0
 
-echo "$new" > "$OUT"
-cat "$OUT"
+echo "$new" > "$output"
+cat "$output"
diff --git a/source/html.c b/source/html.c
index 6e023d1..58ccf3b 100644
--- a/source/html.c
+++ b/source/html.c
@@ -1,20 +1,22 @@
-/* html.c: helper functions for html output
- *
- * Copyright (C) 2006-2014 cgit Development Team <cgit@lists.zx2c4.com>
- *
- * Licensed under GNU General Public License v2
- *   (see LICENSE.txt for full license text)
+/*
+ * The output layer every cgit page is built with, holding the escaping rules
+ * for page text, attribute values, URL paths and query arguments along with
+ * the small formatting helpers the rest of the code prints through. What is
+ * written here is gathered into one buffer and handed to stdout in whole
+ * blocks, because a page is made of a great many small fragments and a write
+ * apiece spent more time in the kernel than rendering the page did. cgit
+ * shares stdout with the filters it runs and with git itself, so that buffer
+ * has to be emptied wherever another writer takes over.
  */
 
 #include "cgit.h"
 #include "html.h"
-#include "url.h"
+
 #define HTML_WRITE_BUFSIZE (64 * 1024)
-static char html_buf[HTML_WRITE_BUFSIZE];
-static size_t html_buflen;
 
-/* Percent-encoding of each character, except: a-zA-Z0-9!$()*,./:;@- */
-static const char* url_escape_table[256] = {
+// The percent encoding for each byte, with NULL marking the bytes a URL may
+// carry as themselves. Those are the letters, the digits, and !$()*,-./:;@[]_~
+static const char *url_escape_table[256] = {
 	"%00", "%01", "%02", "%03", "%04", "%05", "%06", "%07",
 	"%08", "%09", "%0a", "%0b", "%0c", "%0d", "%0e", "%0f",
 	"%10", "%11", "%12", "%13", "%14", "%15", "%16", "%17",
@@ -49,24 +51,42 @@ static const char* url_escape_table[256] = {
 	"%f8", "%f9", "%fa", "%fb", "%fc", "%fd", "%fe", "%ff"
 };
 
+static char out_buf[HTML_WRITE_BUFSIZE];
+static size_t out_len;
+static struct strbuf *capture;
+
+static void write_out(const char *data, size_t size)
+{
+	// A blob, a snapshot or a patch reaches this with a size well past what
+	// one write can move onto a pipe, so a short write is ordinary rather
+	// than an error and has to be resumed instead of reported.
+	if (write_in_full(STDOUT_FILENO, data, size) < 0)
+		die_errno("write error on html output");
+}
+
+/*
+ * Format into one of a rotating set of static buffers, so that a few results
+ * can be alive at once, for example as several arguments to one call. The slot
+ * count has to stay a power of two for the wrap below.
+ */
 char *cgit_fmt(const char *format, ...)
 {
 	static char buf[8][1024];
-	static int bufidx;
+	static int slot;
 	int len;
 	va_list args;
 
-	bufidx++;
-	bufidx &= 7;
+	slot++;
+	slot &= ARRAY_SIZE(buf) - 1;
 
 	va_start(args, format);
-	len = vsnprintf(buf[bufidx], sizeof(buf[bufidx]), format, args);
+	len = vsnprintf(buf[slot], sizeof(buf[slot]), format, args);
 	va_end(args);
-	if (len < 0 || (size_t)len >= sizeof(buf[bufidx])) {
+	if (len < 0 || (size_t)len >= sizeof(buf[slot])) {
 		fprintf(stderr, "[html.c] string truncated: %s\n", format);
 		exit(1);
 	}
-	return buf[bufidx];
+	return buf[slot];
 }
 
 char *cgit_fmtalloc(const char *format, ...)
@@ -81,77 +101,44 @@ char *cgit_fmtalloc(const char *format, ...)
 	return strbuf_detach(&sb, NULL);
 }
 
-/*
- * Page output is collected here and written out in whole buffers. A page is
- * built from a great many small fragments, and writing each one cost a syscall
- * apiece: a tree listing spent more time entering the kernel than generating
- * anything.
- *
- * cgit does not own stdout by itself, so the buffer has to be emptied before
- * anything else writes there. Those points are a filter taking over stdout,
- * the header block that precedes output produced by git itself, the cache
- * slot being measured, and process exit.
- */
-static void html_write(const char *data, size_t size)
-{
-	// A blob, snapshot or patch reaches this with a size well past what one
-	// write can move onto a pipe, so a short write is ordinary rather than
-	// an error and has to be resumed instead of reported.
-	if (write_in_full(STDOUT_FILENO, data, size) < 0)
-		die_errno("write error on html output");
-}
-
 void html_flush(void)
 {
-	size_t len = html_buflen;
+	size_t len = out_len;
 
 	if (!len)
 		return;
-	// Clear the length first: html_write can die, and the error page it
-	// produces would otherwise try to flush the same bytes again.
-	html_buflen = 0;
-	html_write(html_buf, len);
+	// Clear the length first, because write_out can die and the error page
+	// it produces would otherwise try to flush the same bytes again.
+	out_len = 0;
+	write_out(out_buf, len);
 }
 
-/*
- * While a capture is in effect, page output is collected into the caller's
- * buffer instead of being written. The diff view uses this to render a file's
- * body at the point it already has the file open, rather than walking the whole
- * tree a second time to produce what the diffstat above it has to be printed
- * before.
- *
- * A capture must not span anything that writes to stdout by another route, a
- * filter in particular, since that output would escape the capture.
- */
-static struct strbuf *html_capture;
-
 void html_capture_begin(struct strbuf *sb)
 {
 	html_flush();
-	html_capture = sb;
+	capture = sb;
 }
 
 void html_capture_end(void)
 {
-	html_capture = NULL;
+	capture = NULL;
 }
 
 void html_raw(const char *data, size_t size)
 {
-	if (html_capture) {
-		strbuf_add(html_capture, data, size);
+	if (capture) {
+		strbuf_add(capture, data, size);
 		return;
 	}
 	if (size >= HTML_WRITE_BUFSIZE) {
-		// Nothing is gained by copying a blob through the buffer.
 		html_flush();
-		html_write(data, size);
+		write_out(data, size);
 		return;
 	}
-	if (html_buflen + size > HTML_WRITE_BUFSIZE)
+	if (out_len + size > HTML_WRITE_BUFSIZE)
 		html_flush();
-	memcpy(html_buf + html_buflen, data, size);
-	html_buflen += size;
+	memcpy(out_buf + out_len, data, size);
+	out_len += size;
 }
 
 void html(const char *txt)
@@ -162,13 +149,13 @@ void html(const char *txt)
 void htmlf(const char *format, ...)
 {
 	va_list args;
-	struct strbuf buf = STRBUF_INIT;
+	struct strbuf sb = STRBUF_INIT;
 
 	va_start(args, format);
-	strbuf_vaddf(&buf, format, args);
+	strbuf_vaddf(&sb, format, args);
 	va_end(args);
-	html(buf.buf);
-	strbuf_release(&buf);
+	html(sb.buf);
+	strbuf_release(&sb);
 }
 
 void html_txtf(const char *format, ...)
@@ -182,14 +169,14 @@ void html_txtf(const char *format, ...)
 
 void html_vtxtf(const char *format, va_list ap)
 {
-	va_list cp;
-	struct strbuf buf = STRBUF_INIT;
-
-	va_copy(cp, ap);
-	strbuf_vaddf(&buf, format, cp);
-	va_end(cp);
-	html_txt(buf.buf);
-	strbuf_release(&buf);
+	va_list copy;
+	struct strbuf sb = STRBUF_INIT;
+
+	va_copy(copy, ap);
+	strbuf_vaddf(&sb, format, copy);
+	va_end(copy);
+	html_txt(sb.buf);
+	strbuf_release(&sb);
 }
 
 void html_txt(const char *txt)
@@ -200,40 +187,40 @@ void html_txt(const char *txt)
 
 ssize_t html_ntxt(const char *txt, size_t len)
 {
-	const char *t = txt;
-	ssize_t slen;
+	const char *p = txt;
+	ssize_t left;
 
 	if (len > SSIZE_MAX)
 		return -1;
 
-	slen = (ssize_t) len;
-	while (t && *t && slen--) {
-		int c = *t;
+	left = (ssize_t) len;
+	while (p && *p && left--) {
+		int c = *p;
 		if (c == '<' || c == '>' || c == '&') {
-			html_raw(txt, t - txt);
+			html_raw(txt, p - txt);
 			if (c == '>')
 				html("&gt;");
 			else if (c == '<')
 				html("&lt;");
 			else if (c == '&')
 				html("&amp;");
-			txt = t + 1;
+			txt = p + 1;
 		}
-		t++;
+		p++;
 	}
-	if (t != txt)
-		html_raw(txt, t - txt);
-	return slen;
+	if (p != txt)
+		html_raw(txt, p - txt);
+	return left;
 }
 
-void html_attrf(const char *fmt, ...)
+void html_attrf(const char *format, ...)
 {
-	va_list ap;
+	va_list args;
 	struct strbuf sb = STRBUF_INIT;
 
-	va_start(ap, fmt);
-	strbuf_vaddf(&sb, fmt, ap);
-	va_end(ap);
+	va_start(args, format);
+	strbuf_vaddf(&sb, format, args);
+	va_end(args);
 
 	html_attr(sb.buf);
 	strbuf_release(&sb);
@@ -241,11 +228,11 @@ void html_attrf(const char *fmt, ...)
 
 void html_attr(const char *txt)
 {
-	const char *t = txt;
-	while (t && *t) {
-		int c = *t;
+	const char *p = txt;
+	while (p && *p) {
+		int c = *p;
 		if (c == '<' || c == '>' || c == '\'' || c == '\"' || c == '&') {
-			html_raw(txt, t - txt);
+			html_raw(txt, p - txt);
 			if (c == '>')
 				html("&gt;");
 			else if (c == '<')
@@ -256,74 +243,76 @@ void html_attr(const char *txt)
 				html("&quot;");
 			else if (c == '&')
 				html("&amp;");
-			txt = t + 1;
+			txt = p + 1;
 		}
-		t++;
+		p++;
 	}
-	if (t != txt)
+	if (p != txt)
 		html(txt);
 }
 
 void html_url_path(const char *txt)
 {
-	const char *t = txt;
-	while (t && *t) {
-		unsigned char c = *t;
-		const char *e = url_escape_table[c];
-		if (e && c != '+' && c != '&') {
-			html_raw(txt, t - txt);
-			html(e);
-			txt = t + 1;
+	const char *p = txt;
+	while (p && *p) {
+		unsigned char c = *p;
+		const char *esc = url_escape_table[c];
+		// The table is shared with the query string case, where a plus
+		// means a space and an ampersand separates parameters. Neither
+		// carries that meaning in a path.
+		if (esc && c != '+' && c != '&') {
+			html_raw(txt, p - txt);
+			html(esc);
+			txt = p + 1;
 		}
-		t++;
+		p++;
 	}
-	if (t != txt)
+	if (p != txt)
 		html(txt);
 }
 
 void html_url_arg(const char *txt)
 {
-	const char *t = txt;
-	while (t && *t) {
-		unsigned char c = *t;
-		const char *e = url_escape_table[c];
+	const char *p = txt;
+	while (p && *p) {
+		unsigned char c = *p;
+		const char *esc = url_escape_table[c];
 		if (c == ' ')
-			e = "+";
-		if (e) {
-			html_raw(txt, t - txt);
-			html(e);
-			txt = t + 1;
+			esc = "+";
+		if (esc) {
+			html_raw(txt, p - txt);
+			html(esc);
+			txt = p + 1;
 		}
-		t++;
+		p++;
 	}
-	if (t != txt)
+	if (p != txt)
 		html(txt);
 }
 
 void html_header_arg_in_quotes(const char *txt)
 {
-	const char *t = txt;
-	while (t && *t) {
-		unsigned char c = *t;
-		const char *e = NULL;
+	const char *p = txt;
+	while (p && *p) {
+		unsigned char c = *p;
+		const char *esc = NULL;
 		if (c == '\\')
-			e = "\\\\";
+			esc = "\\\\";
 		else if (c == '\r')
-			e = "\\r";
+			esc = "\\r";
 		else if (c == '\n')
-			e = "\\n";
+			esc = "\\n";
 		else if (c == '"')
-			e = "\\\"";
-		if (e) {
-			html_raw(txt, t - txt);
-			html(e);
-			txt = t + 1;
+			esc = "\\\"";
+		if (esc) {
+			html_raw(txt, p - txt);
+			html(esc);
+			txt = p + 1;
 		}
-		t++;
+		p++;
 	}
-	if (t != txt)
+	if (p != txt)
 		html(txt);
-
 }
 
 void html_hidden(const char *name, const char *value)
@@ -335,7 +324,8 @@ void html_hidden(const char *name, const char *value)
 	html("'/>");
 }
 
-void html_option(const char *value, const char *text, const char *selected_value)
+void html_option(const char *value, const char *text,
+		 const char *selected_value)
 {
 	html("<option value='");
 	html_attr(value);
@@ -375,6 +365,10 @@ void html_link_close(void)
 	html("</a>");
 }
 
+/*
+ * Render one permission triplet, so a caller prints a whole mode by passing it
+ * shifted right by six, then by three, then unshifted.
+ */
 void html_fileperm(unsigned short mode)
 {
 	htmlf("%c%c%c", (mode & 4 ? 'r' : '-'),
@@ -392,20 +386,21 @@ int html_include(const char *filename)
 			filename, strerror(errno), errno);
 		return -1;
 	}
-	while ((len = fread(buf, 1, 4096, f)) > 0)
+	while ((len = fread(buf, 1, sizeof(buf), f)) > 0)
 		html_raw(buf, len);
 	fclose(f);
 	return 0;
 }
 
-void http_parse_querystring(const char *txt, void (*fn)(const char *name, const char *value))
+void http_parse_querystring(const char *txt,
+			    void (*fn)(const char *name, const char *value))
 {
-	const char *t = txt;
+	const char *p = txt;
 
-	while (t && *t) {
-		char *name = url_decode_parameter_name(&t);
+	while (p && *p) {
+		char *name = url_decode_parameter_name(&p);
 		if (*name) {
-			char *value = url_decode_parameter_value(&t);
+			char *value = url_decode_parameter_value(&p);
 			fn(name, value);
 			free(value);
 		}
diff --git a/source/html.h b/source/html.h
index 0818ab5..ecc7141 100644
--- a/source/html.h
+++ b/source/html.h
@@ -1,53 +1,83 @@
-#ifndef HTML_H
-#define HTML_H
+/*
+ * The printing side of cgit, which every page is written through. Text is
+ * handed in as plain strings and the call a renderer picks decides how it is
+ * escaped for where it lands, whether that is page text, an attribute value, a
+ * URL path, a query argument or a quoted HTTP header. Output is buffered
+ * rather than written as it is produced, so the flush and capture calls below
+ * are part of the contract for anything that reaches stdout by another route.
+ */
+
+#ifndef CGIT_HTML_H
+#define CGIT_HTML_H
 
 #include "cgit.h"
 
+// How much of a generated run to build before handing it to html_raw. Growing
+// past this gains nothing, since html_raw gathers what it is given into a
+// buffer of its own, and a run built whole would instead be sized by the file
+// it came from, which the blob limits do not bound.
+#define HTML_BATCH (64 * 1024)
+
 extern void html_raw(const char *txt, size_t size);
 extern void html(const char *txt);
 
-/* How much of a generated run to build before handing it to html_raw. Batching
- * is what removes the write per line, and html_raw already gathers what it is
- * given into one buffer, so nothing is gained by growing past this. A run built
- * whole would instead be sized by the file it was generated from, which the
- * blob limits do not bound. */
-#define HTML_BATCH (64 * 1024)
-
-/* Write out whatever page output is still buffered. Call before anything
- * other than html_raw writes to stdout. */
+/*
+ * Write out whatever page output is still buffered. Call this before anything
+ * other than html_raw writes to stdout.
+ */
 extern void html_flush(void);
 
-/* Collect page output into sb rather than writing it, until html_capture_end.
- * Must not span a filter or anything else that writes to stdout directly. */
+/*
+ * Collect page output into sb rather than writing it, until html_capture_end.
+ * A capture must not span a filter or anything else that writes to stdout
+ * directly, since that output would escape the capture.
+ */
 extern void html_capture_begin(struct strbuf *sb);
 extern void html_capture_end(void);
 
 __attribute__((format (printf,1,2)))
-extern void htmlf(const char *format,...);
+extern void htmlf(const char *format, ...);
 
 __attribute__((format (printf,1,2)))
-extern void html_txtf(const char *format,...);
+extern void html_txtf(const char *format, ...);
 
 __attribute__((format (printf,1,0)))
 extern void html_vtxtf(const char *format, va_list ap);
 
 __attribute__((format (printf,1,2)))
-extern void html_attrf(const char *format,...);
+extern void html_attrf(const char *format, ...);
 
 extern void html_txt(const char *txt);
+
+/*
+ * Escape at most len bytes of txt. The return is what is left of that budget,
+ * and a caller prints an ellipsis when it comes back negative.
+ */
 extern ssize_t html_ntxt(const char *txt, size_t len);
+
 extern void html_attr(const char *txt);
 extern void html_url_path(const char *txt);
 extern void html_url_arg(const char *txt);
 extern void html_header_arg_in_quotes(const char *txt);
 extern void html_hidden(const char *name, const char *value);
-extern void html_option(const char *value, const char *text, const char *selected_value);
+
+__attribute__((format (printf,1,2)))
+extern char *cgit_fmt(const char *format, ...);
+
+__attribute__((format (printf,1,2)))
+extern char *cgit_fmtalloc(const char *format, ...);
+
+extern void html_option(const char *value, const char *text,
+			const char *selected_value);
 extern void html_intoption(int value, const char *text, int selected_value);
-extern void html_link_open(const char *url, const char *title, const char *class);
+extern void html_link_open(const char *url, const char *title,
+			   const char *class);
 extern void html_link_close(void);
 extern void html_fileperm(unsigned short mode);
 extern int html_include(const char *filename);
 
-extern void http_parse_querystring(const char *txt, void (*fn)(const char *name, const char *value));
+extern void http_parse_querystring(const char *txt,
+				   void (*fn)(const char *name,
+					      const char *value));
 
-#endif /* HTML_H */
+#endif // CGIT_HTML_H
diff --git a/source/parsing.c b/source/parsing.c
index 0d63b51..de2798a 100644
--- a/source/parsing.c
+++ b/source/parsing.c
@@ -1,127 +1,61 @@
-/* parsing.c: parsing of config files
- *
- * Copyright (C) 2006-2014 cgit Development Team <cgit@lists.zx2c4.com>
- *
- * Licensed under GNU General Public License v2
- *   (see LICENSE.txt for full license text)
+/*
+ * The two kinds of raw text cgit reads before it can render a page, the path
+ * of the incoming request and the body of a commit or a tag object. Reading
+ * the request path settles which repository and which page the request works
+ * on. Commit and tag objects arrive as header lines, a blank line, and then
+ * the message, so the readers here walk the headers, copy out the fields the
+ * pages need, and re-encode the message into the page encoding.
  */
 
 #define USE_THE_REPOSITORY_VARIABLE
 
 #include "cgit.h"
+#include "parsing.h"
+#include "shared.h"
 
-/*
- * url syntax: [repo ['/' cmd [ '/' path]]]
- *   repo: any valid repo url, may contain '/'
- *   cmd:  log | commit | diff | tree | view | blob | snapshot
- *   path: any valid path, may contain '/'
- *
- */
-void cgit_parse_url(const char *url)
-{
-	char *c, *cmd, *p, *buf;
-	struct cgit_repo *repo;
-
-	if (!url || url[0] == '\0')
-		return;
-
-	ctx.qry.page = NULL;
-	ctx.repo = cgit_get_repoinfo(url);
-	if (ctx.repo) {
-		ctx.qry.repo = ctx.repo->url;
-		return;
-	}
-
-	buf = xstrdup(url);
-	cmd = NULL;
-	c = strchr(buf, '/');
-	while (c) {
-		c[0] = '\0';
-		repo = cgit_get_repoinfo(buf);
-		if (repo) {
-			ctx.repo = repo;
-			cmd = c;
-		}
-		c[0] = '/';
-		c = strchr(c + 1, '/');
-	}
-
-	if (ctx.repo) {
-		ctx.qry.repo = ctx.repo->url;
-		p = strchr(cmd + 1, '/');
-		if (p) {
-			p[0] = '\0';
-			if (p[1])
-				ctx.qry.path = cgit_trim_end(p + 1, '/');
-		}
-		if (cmd[1])
-			ctx.qry.page = xstrdup(cmd + 1);
-	}
-	free(buf);
-}
-
-static char *substr(const char *head, const char *tail)
+static char *substr(const char *start, const char *end)
 {
 	size_t len;
 	char *buf;
 
-	if (tail < head)
+	if (end < start)
 		return xstrdup("");
-	// head points into the commit buffer, so strlcpy would measure the
-	// whole remaining commit just to copy a name or a subject off the front
-	// of it.
-	len = tail - head;
+	// start points into the object buffer, so strlcpy would measure the
+	// whole rest of the object to copy a name off the front of it.
+	len = end - start;
 	buf = xmalloc(len + 1);
-	memcpy(buf, head, len);
+	memcpy(buf, start, len);
 	buf[len] = '\0';
 	return buf;
 }
 
-static void parse_user(const char *t, char **name, char **email, unsigned long *date, int *tz)
+static void parse_user(const char *line, char **name, char **email,
+		       timestamp_t *date, int *tz)
 {
 	struct ident_split ident;
-	unsigned email_len;
+	struct strbuf address = STRBUF_INIT;
+	ptrdiff_t email_len;
 
-	if (!split_ident_line(&ident, t, strchrnul(t, '\n') - t)) {
+	if (!split_ident_line(&ident, line, strchrnul(line, '\n') - line)) {
 		*name = substr(ident.name_begin, ident.name_end);
 
+		// Assembled rather than formatted. The length is a pointer
+		// difference, and the precision of a %.*s conversion has to be
+		// an int, which cannot hold one on a 64 bit host.
 		email_len = ident.mail_end - ident.mail_begin;
-		*email = xmalloc(strlen("<") + email_len + strlen(">") + 1);
-		xsnprintf(*email, email_len + 3, "<%.*s>", email_len, ident.mail_begin);
+		strbuf_addch(&address, '<');
+		if (email_len > 0)
+			strbuf_add(&address, ident.mail_begin, email_len);
+		strbuf_addch(&address, '>');
+		*email = strbuf_detach(&address, NULL);
 
 		if (ident.date_begin)
-			*date = strtoul(ident.date_begin, NULL, 10);
+			*date = parse_timestamp(ident.date_begin, NULL, 10);
 		if (ident.tz_begin)
 			*tz = atoi(ident.tz_begin);
 	}
 }
 
-#ifdef NO_ICONV
-#define reencode(a, b, c)
-#else
-static const char *reencode(char **txt, const char *src_enc, const char *dst_enc)
-{
-	char *tmp;
-
-	if (!txt)
-		return NULL;
-
-	if (!*txt || !src_enc || !dst_enc)
-		return *txt;
-
-	/* no encoding needed if src_enc equals dst_enc */
-	if (!strcasecmp(src_enc, dst_enc))
-		return *txt;
-
-	tmp = reencode_string(*txt, dst_enc, src_enc);
-	if (tmp) {
-		free(*txt);
-		*txt = tmp;
-	}
-	return *txt;
-}
-#endif
-
 static const char *next_header_line(const char *p)
 {
 	p = strchr(p, '\n');
@@ -135,17 +69,89 @@ static int end_of_header(const char *p)
 	return !p || (*p == '\n');
 }
 
+/*
+ * A git built without iconv still offers reencode_string, where it always
+ * fails, so the field is left in whatever encoding it arrived in.
+ */
+static const char *reencode(char **text, const char *from, const char *to)
+{
+	char *converted;
+
+	if (!text)
+		return NULL;
+
+	if (!*text || !from || !to)
+		return *text;
+
+	if (!strcasecmp(from, to))
+		return *text;
+
+	converted = reencode_string(*text, to, from);
+	if (converted) {
+		free(*text);
+		*text = converted;
+	}
+	return *text;
+}
+
+/*
+ * A repository url may itself contain slashes, so every slash-separated
+ * prefix is looked up and the longest one that names a repository wins.
+ */
+void cgit_parse_url(const char *url)
+{
+	char *buf, *slash, *repo_end, *page_end;
+	struct cgit_repo *repo;
+
+	if (!url || url[0] == '\0')
+		return;
+
+	ctx.qry.page = NULL;
+	ctx.repo = cgit_get_repoinfo(url);
+	if (ctx.repo) {
+		ctx.qry.repo = ctx.repo->url;
+		return;
+	}
+
+	buf = xstrdup(url);
+	repo_end = NULL;
+	slash = strchr(buf, '/');
+	while (slash) {
+		slash[0] = '\0';
+		repo = cgit_get_repoinfo(buf);
+		if (repo) {
+			ctx.repo = repo;
+			repo_end = slash;
+		}
+		slash[0] = '/';
+		slash = strchr(slash + 1, '/');
+	}
+
+	if (ctx.repo) {
+		ctx.qry.repo = ctx.repo->url;
+		page_end = strchr(repo_end + 1, '/');
+		if (page_end) {
+			page_end[0] = '\0';
+			if (page_end[1])
+				ctx.qry.path = cgit_trim_end(page_end + 1, '/');
+		}
+		if (repo_end[1])
+			ctx.qry.page = xstrdup(repo_end + 1);
+	}
+	free(buf);
+}
+
 struct commitinfo *cgit_parse_commit(struct commit *commit)
 {
-	struct commitinfo *ret;
+	struct commitinfo *info;
 	const char *p = repo_get_commit_buffer(the_repository, commit, NULL);
-	const char *t;
+	const char *eol;
 
-	ret = xcalloc(1, sizeof(struct commitinfo));
-	ret->commit = commit;
+	info = xcalloc(1, sizeof(struct commitinfo));
+	info->commit = commit;
 
 	if (!p)
-		return ret;
+		return info;
 
 	if (!skip_prefix(p, "tree ", &p))
 		die("Bad commit: %s", oid_to_hex(&commit->object.oid));
@@ -155,27 +161,29 @@ struct commitinfo *cgit_parse_commit(struct commit *commit)
 		p += the_hash_algo->hexsz + 1;
 
 	if (p && skip_prefix(p, "author ", &p)) {
-		parse_user(p, &ret->author, &ret->author_email,
-			&ret->author_date, &ret->author_tz);
+		parse_user(p, &info->author, &info->author_email,
+			&info->author_date, &info->author_tz);
 		p = next_header_line(p);
 	}
 
 	if (p && skip_prefix(p, "committer ", &p)) {
-		parse_user(p, &ret->committer, &ret->committer_email,
-			&ret->committer_date, &ret->committer_tz);
+		parse_user(p, &info->committer, &info->committer_email,
+			&info->committer_date, &info->committer_tz);
 		p = next_header_line(p);
 	}
 
 	if (p && skip_prefix(p, "encoding ", &p)) {
-		t = strchr(p, '\n');
-		if (t) {
-			ret->msg_encoding = substr(p, t + 1);
-			p = t + 1;
+		eol = strchr(p, '\n');
+		if (eol) {
+			info->msg_encoding = substr(p, eol + 1);
+			p = eol + 1;
 		}
 	}
 
-	if (!ret->msg_encoding)
-		ret->msg_encoding = xstrdup("UTF-8");
+	// Git only writes the header when the message is in something other
+	// than UTF-8, so its absence means UTF-8.
+	if (!info->msg_encoding)
+		info->msg_encoding = xstrdup("UTF-8");
 
 	while (!end_of_header(p))
 		p = next_header_line(p);
@@ -183,27 +191,28 @@ struct commitinfo *cgit_parse_commit(struct commit *commit)
 		p++;
 
 	if (p) {
-		t = strchrnul(p, '\n');
-		ret->subject = substr(p, t);
-		while (*t == '\n')
-			t++;
-		ret->msg = xstrdup(t);
+		eol = strchrnul(p, '\n');
+		info->subject = substr(p, eol);
+		while (*eol == '\n')
+			eol++;
+		info->msg = xstrdup(eol);
 	} else {
-		// A crafted commit can end right after its headers with no
-		// message. Keep subject and msg as empty strings so callers
-		// can treat them as text unconditionally.
-		ret->subject = xstrdup("");
-		ret->msg = xstrdup("");
+		// Reached when an object is truncated mid header, which
+		// leaves nothing at all after them. Callers render subject
+		// and msg as text without checking, so they get empty
+		// strings rather than NULL.
+		info->subject = xstrdup("");
+		info->msg = xstrdup("");
 	}
 
-	reencode(&ret->author, ret->msg_encoding, PAGE_ENCODING);
-	reencode(&ret->author_email, ret->msg_encoding, PAGE_ENCODING);
-	reencode(&ret->committer, ret->msg_encoding, PAGE_ENCODING);
-	reencode(&ret->committer_email, ret->msg_encoding, PAGE_ENCODING);
-	reencode(&ret->subject, ret->msg_encoding, PAGE_ENCODING);
-	reencode(&ret->msg, ret->msg_encoding, PAGE_ENCODING);
+	reencode(&info->author, info->msg_encoding, PAGE_ENCODING);
+	reencode(&info->author_email, info->msg_encoding, PAGE_ENCODING);
+	reencode(&info->committer, info->msg_encoding, PAGE_ENCODING);
+	reencode(&info->committer_email, info->msg_encoding, PAGE_ENCODING);
+	reencode(&info->subject, info->msg_encoding, PAGE_ENCODING);
+	reencode(&info->msg, info->msg_encoding, PAGE_ENCODING);
 
-	return ret;
+	return info;
 }
 
 struct taginfo *cgit_parse_tag(struct tag *tag)
@@ -212,18 +221,19 @@ struct taginfo *cgit_parse_tag(struct tag *tag)
 	enum object_type type;
 	unsigned long size;
 	const char *p;
-	struct taginfo *ret = NULL;
+	struct taginfo *info = NULL;
 
-	data = odb_read_object(the_repository->objects, &tag->object.oid, &type, &size);
+	data = odb_read_object(the_repository->objects, &tag->object.oid,
+			       &type, &size);
 	if (!data || type != OBJ_TAG)
 		goto cleanup;
 
-	ret = xcalloc(1, sizeof(struct taginfo));
+	info = xcalloc(1, sizeof(struct taginfo));
 
 	for (p = data; !end_of_header(p); p = next_header_line(p)) {
 		if (skip_prefix(p, "tagger ", &p)) {
-			parse_user(p, &ret->tagger, &ret->tagger_email,
-				&ret->tagger_date, &ret->tagger_tz);
+			parse_user(p, &info->tagger, &info->tagger_email,
+				&info->tagger_date, &info->tagger_tz);
 		}
 	}
 
@@ -231,9 +241,9 @@ struct taginfo *cgit_parse_tag(struct tag *tag)
 		p++;
 
 	if (p && *p)
-		ret->msg = xstrdup(p);
+		info->msg = xstrdup(p);
 
 cleanup:
 	free(data);
-	return ret;
+	return info;
 }
diff --git a/source/parsing.h b/source/parsing.h
new file mode 100644
index 0000000..2230eaa
--- /dev/null
+++ b/source/parsing.h
@@ -0,0 +1,20 @@
+/*
+ * Readers that turn raw git object bodies and the incoming request URL into
+ * the records cgit renders from. Commit and tag objects arrive as flat text,
+ * so the work here is splitting off the headers and re-encoding the remaining
+ * free text into the page encoding.
+ */
+
+#ifndef CGIT_PARSING_H
+#define CGIT_PARSING_H
+
+#include "cgit.h"
+
+// Split PATH_INFO into the repository and the page it names, storing both in
+// ctx.qry.
+extern void cgit_parse_url(const char *url);
+
+extern struct commitinfo *cgit_parse_commit(struct commit *commit);
+extern struct taginfo *cgit_parse_tag(struct tag *tag);
+
+#endif // CGIT_PARSING_H
diff --git a/source/scan-tree.c b/source/scan-tree.c
index 89f65c9..4a1444f 100644
--- a/source/scan-tree.c
+++ b/source/scan-tree.c
@@ -1,86 +1,120 @@
-/* scan-tree.c
- *
- * Copyright (C) 2006-2014 cgit Development Team <cgit@lists.zx2c4.com>
- *
- * Licensed under GNU General Public License v2
- *   (see LICENSE.txt for full license text)
+/*
+ * Discovery of repositories on disk, so that an administrator can point cgit
+ * at a directory instead of naming every repository in cgitrc. The tree below
+ * that directory is walked for bare repositories and working trees, or when a
+ * project list is configured only the paths it names are visited. Every
+ * repository found is registered under its path relative to the root of the
+ * scan, with owner, description and section taken from the files git keeps
+ * beside it, and a cgitrc in the repository has the last word over all of them.
  */
 
 #include "cgit.h"
+#include "config.h"
 #include "scan-tree.h"
-#include "configfile.h"
-#include "html.h"
-#include <config.h>
+#include "shared.h"
 
-// The description git puts in every freshly created repository.
+// Git writes this line into the description file of every repository it
+// creates, so a repository still carrying it gets cgit's own default instead.
 static const char *default_git_desc = "Unnamed repository; edit this file 'description' to name the repository.";
 
-static struct cgit_repo *repo;
+// The config callbacks are handed only a name and a value, so the repository
+// being filled in waits here for the length of one add_repo call.
+static struct cgit_repo *current_repo;
 
-/* return 1 if path contains a objects/ directory and a HEAD file */
-static int is_git_dir(const char *path)
+static int stat_entry(const char *dir, const char *name, struct stat *st)
 {
-	struct stat st;
-	struct strbuf pathbuf = STRBUF_INIT;
-	int result = 0;
-
-	strbuf_addf(&pathbuf, "%s/objects", path);
-	if (stat(pathbuf.buf, &st)) {
-		if (errno != ENOENT)
-			fprintf(stderr, "Error checking path %s: %s (%d)\n",
-				path, strerror(errno), errno);
-		goto out;
-	}
-	if (!S_ISDIR(st.st_mode))
-		goto out;
-
-	strbuf_reset(&pathbuf);
-	strbuf_addf(&pathbuf, "%s/HEAD", path);
-	if (stat(pathbuf.buf, &st)) {
-		if (errno != ENOENT)
-			fprintf(stderr, "Error checking path %s: %s (%d)\n",
-				path, strerror(errno), errno);
-		goto out;
-	}
-	if (!S_ISREG(st.st_mode))
-		goto out;
+	struct strbuf path = STRBUF_INIT;
+	int err;
 
-	result = 1;
-out:
-	strbuf_release(&pathbuf);
-	return result;
+	strbuf_addf(&path, "%s/%s", dir, name);
+	err = stat(path.buf, st);
+	// A missing entry is the ordinary answer for a directory that is not a
+	// repository, so only some other failure is worth reporting.
+	if (err && errno != ENOENT)
+		fprintf(stderr, "Error checking path %s: %s (%d)\n",
+			dir, strerror(errno), errno);
+	strbuf_release(&path);
+	return err;
 }
 
-static void scan_tree_repo_config(const char *name, const char *value)
+static int is_git_dir(const char *path)
 {
-	cgit_repo_config(repo, name, value);
+	struct stat st;
+
+	if (stat_entry(path, "objects", &st) || !S_ISDIR(st.st_mode))
+		return 0;
+	if (stat_entry(path, "HEAD", &st) || !S_ISREG(st.st_mode))
+		return 0;
+	return 1;
 }
 
-static int gitconfig_config(const char *key, const char *value,
+static int apply_gitconfig(const char *key, const char *value,
 		const __attribute__((unused)) struct config_context *cfg_ctx,
 		void *cb)
 {
 	const char *name;
 
 	if (!strcmp(key, "gitweb.owner"))
-		cgit_repo_config(repo, "owner", value);
+		cgit_repo_config(current_repo, "owner", value);
 	else if (!strcmp(key, "gitweb.description"))
-		cgit_repo_config(repo, "desc", value);
+		cgit_repo_config(current_repo, "desc", value);
 	else if (!strcmp(key, "gitweb.category"))
-		cgit_repo_config(repo, "section", value);
+		cgit_repo_config(current_repo, "section", value);
 	else if (!strcmp(key, "gitweb.homepage"))
-		cgit_repo_config(repo, "homepage", value);
+		cgit_repo_config(current_repo, "homepage", value);
 	else if (skip_prefix(key, "cgit.", &name))
-		cgit_repo_config(repo, name, value);
+		cgit_repo_config(current_repo, name, value);
 
 	return 0;
 }
 
-static char *xstrrchr(char *s, char *from, int c)
+static void apply_cgitrc(const char *name, const char *value)
 {
-	while (from >= s && *from != c)
+	cgit_repo_config(current_repo, name, value);
+}
+
+static char *find_char_back(char *start, char *from, int c)
+{
+	while (from >= start && *from != c)
 		from--;
-	return from < s ? NULL : from;
+	return from < start ? NULL : from;
+}
+
+/*
+ * A positive depth counts separators from the left of the path and a negative
+ * one counts back from the right.
+ */
+static char *section_slash(struct strbuf *relpath, int depth)
+{
+	char *slash;
+
+	if (depth > 0) {
+		slash = relpath->buf - 1;
+		while (slash && depth && (slash = strchr(slash + 1, '/')))
+			depth--;
+	} else {
+		slash = relpath->buf + relpath->len;
+		while (slash && depth &&
+		       (slash = find_char_back(relpath->buf, slash - 1, '/')))
+			depth++;
+	}
+	return slash && !depth ? slash : NULL;
+}
+
+static void set_section_from_path(struct strbuf *relpath, int depth)
+{
+	char *slash = section_slash(relpath, depth);
+
+	if (!slash)
+		return;
+	*slash = '\0';
+	current_repo->section = cgit_strdup_first_line(relpath->buf);
+	*slash = '/';
+	if (starts_with(current_repo->name, current_repo->section)) {
+		current_repo->name += strlen(current_repo->section);
+		if (*current_repo->name == '/')
+			current_repo->name++;
+	}
 }
 
 static void add_repo(const char *base, struct strbuf *path)
@@ -88,10 +122,9 @@ static void add_repo(const char *base, struct strbuf *path)
 	struct stat st;
 	struct passwd *pwd;
 	size_t pathlen;
-	struct strbuf rel = STRBUF_INIT;
-	char *p, *slash;
-	int n;
-	size_t size;
+	struct strbuf relpath = STRBUF_INIT;
+	char *comma;
+	size_t desc_size;
 
 	if (stat(path->buf, &st)) {
 		fprintf(stderr, "Error accessing %s: %s (%d)\n",
@@ -104,7 +137,7 @@ static void add_repo(const char *base, struct strbuf *path)
 
 	if (ctx.cfg.strict_export) {
 		strbuf_addstr(path, ctx.cfg.strict_export);
-		if(stat(path->buf, &st))
+		if (stat(path->buf, &st))
 			return;
 		strbuf_setlen(path, pathlen);
 	}
@@ -115,86 +148,79 @@ static void add_repo(const char *base, struct strbuf *path)
 	strbuf_setlen(path, pathlen);
 
 	if (!starts_with(path->buf, base))
-		strbuf_addbuf(&rel, path);
+		strbuf_addbuf(&relpath, path);
 	else
-		strbuf_addstr(&rel, path->buf + strlen(base) + 1);
+		strbuf_addstr(&relpath, path->buf + strlen(base) + 1);
 
 	// Drop the trailing slash added above before testing for "/.git", since
 	// with it still attached the suffix never matches and every ordinary
 	// working tree is named "repo/.git" instead of "repo".
-	if (rel.len && rel.buf[rel.len - 1] == '/')
-		strbuf_setlen(&rel, rel.len - 1);
-	if (rel.len >= 5 && !strcmp(rel.buf + rel.len - 5, "/.git"))
-		strbuf_setlen(&rel, rel.len - 5);
+	if (relpath.len && relpath.buf[relpath.len - 1] == '/')
+		strbuf_setlen(&relpath, relpath.len - 1);
+	if (relpath.len >= 5 && !strcmp(relpath.buf + relpath.len - 5, "/.git"))
+		strbuf_setlen(&relpath, relpath.len - 5);
 
-	repo = cgit_add_repo(rel.buf);
+	current_repo = cgit_add_repo(relpath.buf);
 	if (ctx.cfg.enable_git_config) {
 		strbuf_addstr(path, "config");
-		git_config_from_file(gitconfig_config, path->buf, NULL);
+		git_config_from_file(apply_gitconfig, path->buf, NULL);
 		strbuf_setlen(path, pathlen);
 	}
 
 	if (ctx.cfg.remove_suffix) {
 		size_t urllen;
-		strip_suffix(repo->url, ".git", &urllen);
-		strip_suffix_mem(repo->url, &urllen, "/");
-		repo->url[urllen] = '\0';
+		strip_suffix(current_repo->url, ".git", &urllen);
+		strip_suffix_mem(current_repo->url, &urllen, "/");
+		current_repo->url[urllen] = '\0';
 	}
-	repo->path = cgit_strdup_first_line(path->buf);
-	while (!repo->owner) {
+	current_repo->path = cgit_strdup_first_line(path->buf);
+	while (!current_repo->owner) {
 		if ((pwd = getpwuid(st.st_uid)) == NULL) {
 			fprintf(stderr, "Error reading owner-info for %s: %s (%d)\n",
 				path->buf, strerror(errno), errno);
 			break;
 		}
+		// A gecos field puts the owner's name in front of a comma
+		// separated list of office and phone details.
 		if (pwd->pw_gecos)
-			if ((p = strchr(pwd->pw_gecos, ',')))
-				*p = '\0';
-		repo->owner = cgit_strdup_first_line(pwd->pw_gecos ? pwd->pw_gecos : pwd->pw_name);
+			if ((comma = strchr(pwd->pw_gecos, ',')))
+				*comma = '\0';
+		current_repo->owner = cgit_strdup_first_line(
+			pwd->pw_gecos ? pwd->pw_gecos : pwd->pw_name);
 	}
 
-	if (repo->desc == cgit_default_repo_desc || !repo->desc) {
+	if (current_repo->desc == cgit_default_repo_desc || !current_repo->desc) {
 		strbuf_addstr(path, "description");
 		if (!stat(path->buf, &st))
-			cgit_read_first_line(path->buf, &repo->desc, &size);
+			cgit_read_first_line(path->buf, &current_repo->desc,
+					     &desc_size);
 		strbuf_setlen(path, pathlen);
-		// Git writes this line into every repository it creates, so it
-		// describes nothing. Treat it as no description at all rather
-		// than repeating it down the whole index.
-		if (repo->desc && !strcmp(repo->desc, default_git_desc)) {
-			free(repo->desc);
-			repo->desc = cgit_default_repo_desc;
+		if (current_repo->desc &&
+		    !strcmp(current_repo->desc, default_git_desc)) {
+			free(current_repo->desc);
+			current_repo->desc = cgit_default_repo_desc;
 		}
 	}
 
-	if (ctx.cfg.section_from_path) {
-		n = ctx.cfg.section_from_path;
-		if (n > 0) {
-			slash = rel.buf - 1;
-			while (slash && n && (slash = strchr(slash + 1, '/')))
-				n--;
-		} else {
-			slash = rel.buf + rel.len;
-			while (slash && n && (slash = xstrrchr(rel.buf, slash - 1, '/')))
-				n++;
-		}
-		if (slash && !n) {
-			*slash = '\0';
-			repo->section = cgit_strdup_first_line(rel.buf);
-			*slash = '/';
-			if (starts_with(repo->name, repo->section)) {
-				repo->name += strlen(repo->section);
-				if (*repo->name == '/')
-					repo->name++;
-			}
-		}
-	}
+	if (ctx.cfg.section_from_path)
+		set_section_from_path(&relpath, ctx.cfg.section_from_path);
 
 	strbuf_addstr(path, "cgitrc");
 	if (!stat(path->buf, &st))
-		parse_configfile(path->buf, &scan_tree_repo_config);
+		config_file_parse(path->buf, &apply_cgitrc);
 
-	strbuf_release(&rel);
+	strbuf_release(&relpath);
+}
+
+static int should_scan(const struct dirent *ent)
+{
+	if (ent->d_name[0] != '.')
+		return 1;
+	if (ent->d_name[1] == '\0')
+		return 0;
+	if (ent->d_name[1] == '.' && ent->d_name[2] == '\0')
+		return 0;
+	return ctx.cfg.scan_hidden_path;
 }
 
 static void scan_path(const char *base, const char *path)
@@ -211,7 +237,7 @@ static void scan_path(const char *base, const char *path)
 		return;
 	}
 
-	strbuf_add(&pathbuf, path, strlen(path));
+	strbuf_add(&pathbuf, path, pathlen);
 	if (is_git_dir(pathbuf.buf)) {
 		add_repo(base, &pathbuf);
 		goto end;
@@ -221,20 +247,12 @@ static void scan_path(const char *base, const char *path)
 		add_repo(base, &pathbuf);
 		goto end;
 	}
-	/*
-	 * Add one because we don't want to lose the trailing '/' when we
-	 * reset the length of pathbuf in the loop below.
-	 */
+	// Take in the '/' that "/.git" left in the buffer, since the loop below
+	// truncates to this length and then appends an entry name straight on.
 	pathlen++;
 	while ((ent = readdir(dir)) != NULL) {
-		if (ent->d_name[0] == '.') {
-			if (ent->d_name[1] == '\0')
-				continue;
-			if (ent->d_name[1] == '.' && ent->d_name[2] == '\0')
-				continue;
-			if (!ctx.cfg.scan_hidden_path)
-				continue;
-		}
+		if (!should_scan(ent))
+			continue;
 		strbuf_setlen(&pathbuf, pathlen);
 		strbuf_addstr(&pathbuf, ent->d_name);
 		if (stat(pathbuf.buf, &st)) {
diff --git a/source/scan-tree.h b/source/scan-tree.h
index def0a7a..515037a 100644
--- a/source/scan-tree.h
+++ b/source/scan-tree.h
@@ -1,2 +1,16 @@
+/*
+ * Discovery of repositories on disk, so that an administrator can point cgit
+ * at a directory instead of naming every repository in cgitrc. The tree below
+ * that directory is walked for bare repositories and working trees, or when a
+ * project list is configured only the paths it names are visited. Every
+ * repository found is added to the shared repository list under its path
+ * relative to the root of the scan.
+ */
+
+#ifndef CGIT_SCAN_TREE_H
+#define CGIT_SCAN_TREE_H
+
 extern void scan_projects(const char *path, const char *projectsfile);
 extern void scan_tree(const char *path);
+
+#endif // CGIT_SCAN_TREE_H
diff --git a/source/shared.c b/source/shared.c
index 259802a..e6dd011 100644
--- a/source/shared.c
+++ b/source/shared.c
This diff is too large to be rendered inline. View it on its own page.
diff --git a/source/shared.h b/source/shared.h
new file mode 100644
index 0000000..9f4cba3
--- /dev/null
+++ b/source/shared.h
@@ -0,0 +1,92 @@
+/*
+ * The helpers that belong to no single page. Repository registration, the
+ * lifetimes of the reference and commit records the renderers build, thin
+ * wrappers over git's diff machinery, and a handful of string, date and
+ * mimetype utilities all live here because several renderers need them. The
+ * two process-wide globals, ctx and cgit_repolist, are defined here as well.
+ */
+
+#ifndef CGIT_SHARED_H
+#define CGIT_SHARED_H
+
+#include "cgit.h"
+
+extern char *cgit_default_repo_desc;
+
+typedef void (*filepair_fn)(struct diff_filepair *pair);
+typedef void (*linediff_fn)(char *line, int len);
+
+/*
+ * Each of these returns its argument untouched when the syscall it wraps
+ * succeeded, and dies with the message and errno otherwise.
+ */
+extern int cgit_die_unless_zero(int result, const char *msg);
+extern int cgit_die_unless_positive(int result, const char *msg);
+extern int cgit_die_unless_non_negative(int result, const char *msg);
+
+/*
+ * Append a repository to cgit_repolist and return it, seeded from the current
+ * global configuration. The returned pointer is invalidated by the next call,
+ * because the list grows by reallocation.
+ */
+extern struct cgit_repo *cgit_add_repo(const char *url);
+extern struct cgit_repo *cgit_get_repoinfo(const char *url);
+
+// Export the CGIT_REPO_* variables that filters and hooks read.
+extern void cgit_prepare_repo_env(struct cgit_repo *repo);
+
+extern void cgit_free_reflist_inner(struct reflist *list);
+
+// Matches git's reference iteration callback, and appends each ref it is
+// given to the struct reflist passed as cb_data.
+extern int cgit_refs_cb(const struct reference *ref, void *cb_data);
+
+extern void cgit_free_commitinfo(struct commitinfo *info);
+extern void cgit_free_taginfo(struct taginfo *info);
+
+extern void cgit_diff_tree_cb(struct diff_queue_struct *q,
+			      struct diff_options *options, void *data);
+
+/*
+ * Diff two blobs and hand fn one line at a time. Returns non-zero when either
+ * blob could not be read. A blob that is binary or larger than max-blob-size
+ * sets binary instead, and produces no lines, so the caller can offer a link
+ * in place of the content.
+ */
+extern int cgit_diff_files(const struct object_id *old_oid,
+			   const struct object_id *new_oid,
+			   unsigned long *old_size, unsigned long *new_size,
+			   int *binary, int context, int ignorews,
+			   linediff_fn fn);
+
+extern void cgit_diff_tree(const struct object_id *old_oid,
+			   const struct object_id *new_oid,
+			   filepair_fn fn, const char *prefix, int ignorews);
+extern void cgit_diff_commit(struct commit *commit, filepair_fn fn,
+			     const char *prefix);
+
+// Accept only the date formats cgit documents, leaving mode unchanged for
+// anything else.
+extern void cgit_parse_date_format(const char *format, struct date_mode *mode);
+
+// Copy str without its trailing runs of c, returning NULL if nothing is left.
+extern char *cgit_trim_end(const char *str, char c);
+extern char *cgit_ensure_end(const char *str, char c);
+extern void strbuf_ensure_end(struct strbuf *sb, char c);
+
+/*
+ * Read the first line of a file into a newly allocated, zero terminated
+ * buffer. Returns 0 on success and an errno value otherwise.
+ */
+extern int cgit_read_first_line(const char *path, char **buf, size_t *size);
+extern char *cgit_strdup_first_line(const char *text);
+
+/*
+ * Replace every $token in text with the matching environment variable. The
+ * result is a static buffer, so a caller that needs to keep it must copy it.
+ */
+extern char *cgit_expand_macros(const char *text);
+
+extern char *cgit_get_mimetype_for_filename(const char *filename);
+
+#endif // CGIT_SHARED_H
diff --git a/source/ui-atom.c b/source/ui-atom.c
index f968e24..aa9237a 100644
--- a/source/ui-atom.c
+++ b/source/ui-atom.c
@@ -1,24 +1,57 @@
-/* ui-atom.c: functions for atom feeds
- *
- * Copyright (C) 2006-2014 cgit Development Team <cgit@lists.zx2c4.com>
- *
- * Licensed under GNU General Public License v2
- *   (see LICENSE.txt for full license text)
+/*
+ * The Atom feed for a repository, which lists recent commits so a reader can
+ * follow the project from a feed reader instead of the browsable pages. The
+ * response is XML rather than a page, so this file writes its own HTTP headers
+ * and never goes through the shared HTML layout.
  */
 
 #define USE_THE_REPOSITORY_VARIABLE
 
 #include "cgit.h"
-#include "ui-atom.h"
 #include "html.h"
+#include "parsing.h"
+#include "shared.h"
+#include "ui-atom.h"
 #include "ui-shared.h"
 
-static void add_entry(struct commit *commit, const char *host)
+/*
+ * Atom timestamps have to be RFC 3339, so a feed ignores the date-format and
+ * local-time settings the browsable pages honour. The zero is the timezone
+ * offset, which pins every feed date to UTC.
+ */
+static const char *feed_date(timestamp_t when)
+{
+	return show_date(when, 0, date_mode_from_type(DATE_ISO8601_STRICT));
+}
+
+/*
+ * cgit keeps an address with the angle brackets it was written with, while
+ * Atom wants the bare address.
+ */
+static void print_email(const char *email)
+{
+	char *copy = xstrdup(email);
+	char *start, *end;
+
+	start = strchr(copy, '<');
+	if (start)
+		start++;
+	else
+		start = copy;
+	end = strchr(start, '>');
+	if (end)
+		*end = '\0';
+
+	html("<email>");
+	html_txt(start);
+	html("</email>\n");
+	free(copy);
+}
+
+static void print_entry(struct commit *commit, const char *host)
 {
-	char delim = '&';
-	char *hex;
-	char *mail, *t, *t2;
 	struct commitinfo *info;
+	char *hex;
 
 	info = cgit_parse_commit(commit);
 	hex = oid_to_hex(&commit->object.oid);
@@ -27,8 +60,7 @@ static void add_entry(struct commit *commit, const char *host)
 	html_txt(info->subject);
 	html("</title>\n");
 	html("<updated>");
-	html_txt(show_date(info->committer_date, 0,
-                    date_mode_from_type(DATE_ISO8601_STRICT)));
+	html_txt(feed_date(info->committer_date));
 	html("</updated>\n");
 	html("<author>\n");
 	if (info->author) {
@@ -36,33 +68,24 @@ static void add_entry(struct commit *commit, const char *host)
 		html_txt(info->author);
 		html("</name>\n");
 	}
-	if (info->author_email && !ctx.cfg.noplainemail) {
-		mail = xstrdup(info->author_email);
-		t = strchr(mail, '<');
-		if (t)
-			t++;
-		else
-			t = mail;
-		t2 = strchr(t, '>');
-		if (t2)
-			*t2 = '\0';
-		html("<email>");
-		html_txt(t);
-		html("</email>\n");
-		free(mail);
-	}
+	if (info->author_email && !ctx.cfg.noplainemail)
+		print_email(info->author_email);
 	html("</author>\n");
 	html("<published>");
-	html_txt(show_date(info->author_date, 0,
-                    date_mode_from_type(DATE_ISO8601_STRICT)));
+	html_txt(feed_date(info->author_date));
 	html("</published>\n");
 	if (host) {
 		char *pageurl;
+		char delim = '&';
+
 		html("<link rel='alternate' type='text/html' href='");
 		html(cgit_httpscheme());
 		html_attr(host);
 		pageurl = cgit_pageurl(ctx.repo->url, "commit", NULL);
 		html_attr(pageurl);
+		// Without a virtual root the page url is already a query
+		// string, so the commit id continues it instead of opening
+		// one.
 		if (ctx.cfg.virtual_root)
 			delim = '?';
 		html_attrf("%cid=%s", delim, hex);
@@ -79,15 +102,16 @@ static void add_entry(struct commit *commit, const char *host)
 	cgit_free_commitinfo(info);
 }
 
-
 void cgit_print_atom(char *tip, const char *path, int max_count)
 {
 	char *host;
+	// setup_revisions reads a command line, so the first slot is the
+	// unused program name and parsing starts at the second.
 	const char *argv[] = {NULL, tip, NULL, NULL, NULL};
 	struct commit *commit;
 	struct rev_info rev;
 	int argc = 2;
-	bool first = true;
+	bool need_updated = true;
 
 	if (ctx.qry.show_all)
 		argv[1] = "--all";
@@ -143,22 +167,23 @@ void cgit_print_atom(char *tip, const char *path, int max_count)
 		free(repourl);
 	}
 	while ((commit = get_revision(&rev)) != NULL) {
-		if (first) {
+		if (need_updated) {
 			html("<updated>");
-			html_txt(show_date(commit->date, 0,
-				date_mode_from_type(DATE_ISO8601_STRICT)));
+			html_txt(feed_date(commit->date));
 			html("</updated>\n");
-			first = false;
+			need_updated = false;
 		}
-		add_entry(commit, host);
+		print_entry(commit, host);
+		// release_commit_memory frees the parent list without clearing
+		// the pointer to it, so drop the dangling reference here.
 		release_commit_memory(the_repository->parsed_objects, commit);
 		commit->parents = NULL;
 	}
-	if (first) {
-		/* An empty feed still needs one feed-level <updated>. */
+	if (need_updated) {
+		// Atom makes a feed level updated mandatory, and an empty feed
+		// has no commit to take one from.
 		html("<updated>");
-		html_txt(show_date(time(NULL), 0,
-			date_mode_from_type(DATE_ISO8601_STRICT)));
+		html_txt(feed_date(time(NULL)));
 		html("</updated>\n");
 	}
 	html("</feed>\n");
diff --git a/source/ui-atom.h b/source/ui-atom.h
index dda953b..71549e8 100644
--- a/source/ui-atom.h
+++ b/source/ui-atom.h
@@ -1,6 +1,11 @@
-#ifndef UI_ATOM_H
-#define UI_ATOM_H
+/*
+ * The Atom feed page, which serves a repository's recent commits as XML for a
+ * feed reader rather than as one of the browsable pages.
+ */
+
+#ifndef CGIT_UI_ATOM_H
+#define CGIT_UI_ATOM_H
 
 extern void cgit_print_atom(char *tip, const char *path, int max_count);
 
-#endif
+#endif // CGIT_UI_ATOM_H
diff --git a/source/ui-blame.c b/source/ui-blame.c
index 132121b..80512a0 100644
--- a/source/ui-blame.c
+++ b/source/ui-blame.c
@@ -1,25 +1,42 @@
-/* ui-blame.c: functions for blame output
- *
- * Copyright (C) 2006-2017 cgit Development Team <cgit@lists.zx2c4.com>
- *
- * Licensed under GNU General Public License v2
- *   (see LICENSE.txt for full license text)
+/*
+ * The blame page, which shows a file next to the commit that last touched
+ * each of its lines. git returns blame as runs of neighbouring lines that
+ * share a commit, and the page turns every run into one block in each column
+ * so the hashes, the line numbers and the striped background all line up with
+ * the source. Only a regular file can be blamed, so a path naming a folder is
+ * turned away.
  */
 
 #define USE_THE_REPOSITORY_VARIABLE
 
 #include "cgit.h"
-#include "ui-blame.h"
+#include "filter.h"
 #include "html.h"
+#include "parsing.h"
+#include "shared.h"
+#include "ui-blame.h"
 #include "ui-shared.h"
-#include "strvec.h"
-#include "blame.h"
 
+// A tab in the rendered source runs on to the next multiple of this. The
+// stylesheet leaves tab-size alone, so the measurement has to match what the
+// browser does on its own rather than anything cgit picks.
+#define TAB_WIDTH 8
 
-/* Blame coalesces neighbouring lines from one commit, but a commit that
- * touched several separate parts of the file still comes back once per part.
- * Each of those repeats a full commit parse, so the rendered detail is kept
- * and looked up by object id. */
+enum blame_target {
+	TARGET_MISSING,
+	TARGET_FILE,
+	TARGET_FOLDER,
+};
+
+struct walk_tree_context {
+	char *rev;
+	int match_baselen;
+	enum blame_target found;
+};
+
+// A commit that touched several separate parts of the file comes back once
+// per part, and each repeat would parse the commit again, so the rendered
+// detail is cached here and looked up by object id.
 static struct string_list suspect_details = STRING_LIST_INIT_DUP;
 
 static void free_suspect_details(void)
@@ -31,8 +48,40 @@ static void free_suspect_details(void)
 	string_list_clear(&suspect_details, 0);
 }
 
-/* Write out a run of one repeated character. A single blame entry can cover
- * the whole file, so the run is handed on in batches rather than built whole. */
+/*
+ * The scoreboard keeps a pointer to revs, so the caller owns both and has to
+ * keep them alive together.
+ */
+static void run_blame(struct blame_scoreboard *sb, struct rev_info *revs,
+		      const char *path, const char *rev)
+{
+	struct strvec argv = STRVEC_INIT;
+	struct blame_origin *origin;
+
+	strvec_push(&argv, "blame");
+	strvec_push(&argv, rev);
+	repo_init_revisions(the_repository, revs, NULL);
+	revs->diffopt.flags.allow_textconv = 1;
+	setup_revisions(argv.nr, argv.v, revs, NULL);
+	init_scoreboard(sb);
+	sb->revs = revs;
+	sb->repo = the_repository;
+	sb->path = path;
+	setup_scoreboard(sb, &origin);
+	origin->suspects = blame_entry_prepend(NULL, 0, sb->num_lines, origin);
+	prio_queue_put(&sb->commits, origin->commit);
+	blame_origin_decref(origin);
+	sb->ent = NULL;
+	sb->path = path;
+	assign_blame(sb, 0);
+	blame_sort_final(sb);
+	blame_coalesce(sb);
+}
+
+/*
+ * A single blame entry can cover the whole file, so the run is handed on in
+ * batches rather than built whole.
+ */
 static void emit_chars(char ch, unsigned long count)
 {
 	struct strbuf run = STRBUF_INIT;
@@ -48,7 +97,10 @@ static void emit_chars(char ch, unsigned long count)
 	strbuf_release(&run);
 }
 
-static char *emit_suspect_detail(struct blame_origin *suspect)
+/*
+ * The returned string belongs to suspect_details, not the caller.
+ */
+static char *suspect_detail(struct blame_origin *suspect)
 {
 	struct commitinfo *info;
 	struct strbuf detail = STRBUF_INIT;
@@ -87,18 +139,21 @@ static char *emit_suspect_detail(struct blame_origin *suspect)
 	return cached->util;
 }
 
-static void emit_blame_entry_hash(struct blame_entry *ent)
+static void emit_entry_hash(struct blame_entry *ent)
 {
 	struct blame_origin *suspect = ent->suspect;
 	struct object_id *oid = &suspect->commit->object.oid;
+	const char *detail = suspect_detail(suspect);
 
-	const char *detail = emit_suspect_detail(suspect);
 	html("<span class='oid'>");
-	cgit_commit_link(repo_find_unique_abbrev(the_repository, oid, DEFAULT_ABBREV), detail,
-			 NULL, ctx.qry.head, oid_to_hex(oid), suspect->path);
+	cgit_commit_link(repo_find_unique_abbrev(the_repository, oid,
+						 DEFAULT_ABBREV),
+			 detail, NULL, ctx.qry.head, oid_to_hex(oid),
+			 suspect->path);
 	html("</span>");
 
-	if (!repo_parse_commit(the_repository, suspect->commit) && suspect->commit->parents) {
+	if (!repo_parse_commit(the_repository, suspect->commit) &&
+	    suspect->commit->parents) {
 		struct commit *parent = suspect->commit->parents->item;
 
 		html(" ");
@@ -107,17 +162,30 @@ static void emit_blame_entry_hash(struct blame_entry *ent)
 				suspect->path);
 	}
 
-	// Batched rather than one write per line of the entry.
+	// The stripes only line up across the columns if each column gives an
+	// entry the same height, so pad out to the lines the entry covers.
 	emit_chars('\n', ent->num_lines);
 }
 
-static void emit_blame_entry_linenumber(struct blame_entry *ent)
+static void emit_hashes(struct blame_scoreboard *sb)
+{
+	struct blame_entry *ent;
+
+	html("<td class='hashes'>");
+	for (ent = sb->ent; ent; ent = ent->next) {
+		html("<div class='alt'><pre>");
+		emit_entry_hash(ent);
+		html("</pre></div>");
+	}
+	html("</td>\n");
+}
+
+static void emit_entry_linenumbers(struct blame_entry *ent)
 {
 	const char *numberfmt = "<a id='n%1$d' href='#n%1$d'>%1$d</a>\n";
 	struct strbuf numbers = STRBUF_INIT;
 	int lineno = ent->lno;
 
-	// Batched rather than a formatted write per line of the entry.
 	while (lineno < ent->lno + ent->num_lines) {
 		strbuf_addf(&numbers, numberfmt, ++lineno);
 		if (numbers.len >= HTML_BATCH) {
@@ -129,50 +197,84 @@ static void emit_blame_entry_linenumber(struct blame_entry *ent)
 	strbuf_release(&numbers);
 }
 
-static void emit_blame_entry_line_background(struct blame_scoreboard *sb,
-					     struct blame_entry *ent)
+static void emit_linenumbers(struct blame_scoreboard *sb)
 {
+	struct blame_entry *ent;
+
+	html("<td class='linenumbers'>");
+	for (ent = sb->ent; ent; ent = ent->next) {
+		html("<div class='alt'><pre>");
+		emit_entry_linenumbers(ent);
+		html("</pre></div>");
+	}
+	html("</td>\n");
+}
+
+static size_t line_width(struct blame_scoreboard *sb, int line)
+{
+	const char *pos = blame_nth_line(sb, line);
+	const char *end = blame_nth_line(sb, line + 1);
+	size_t width = 0;
+
+	while (pos < end) {
+		width++;
+		if (*pos++ == '\t')
+			width = (width + TAB_WIDTH - 1) & ~(TAB_WIDTH - 1);
+	}
+	return width;
+}
+
+/*
+ * The stylesheet takes the source pre out of flow and positions it over these
+ * blocks, so nothing else gives the cell a size and each block has to be
+ * padded to the height and the width of the lines it stands behind.
+ */
+static void emit_entry_background(struct blame_scoreboard *sb,
+				  struct blame_entry *ent)
+{
+	size_t widest = 2;
 	int line;
-	size_t len, maxlen = 2;
-	const char* pos, *endpos;
 
 	for (line = ent->lno; line < ent->lno + ent->num_lines; line++) {
-		pos = blame_nth_line(sb, line);
-		endpos = blame_nth_line(sb, line + 1);
-		len = 0;
-		while (pos < endpos) {
-			len++;
-			if (*pos++ == '\t')
-				len = (len + 7) & ~7;
-		}
-		if (len > maxlen)
-			maxlen = len;
+		size_t width = line_width(sb, line);
+
+		if (width > widest)
+			widest = width;
 	}
 
-	// The widest line decides the padding, so the entry has to be measured
-	// before any of it is written. The newlines do not depend on that
-	// measurement, so both runs go out batched once it is known.
 	emit_chars('\n', ent->num_lines);
-	emit_chars(' ', maxlen - 1);
+	emit_chars(' ', widest - 1);
 }
 
-struct walk_tree_context {
-	char *curr_rev;
-	int match_baselen;
-	int state;
-};
+/*
+ * Frees each entry on the way past, so this has to be the last pass over the
+ * scoreboard's entries.
+ */
+static void emit_line_backgrounds(struct blame_scoreboard *sb)
+{
+	struct blame_entry *ent = sb->ent;
 
-static void print_object(const struct object_id *oid, const char *path,
-			 const char *basename, const char *rev)
+	html("<div>");
+	while (ent) {
+		struct blame_entry *next = ent->next;
+
+		html("<div class='alt'><pre>");
+		emit_entry_background(sb, ent);
+		html("</pre></div>");
+		free(ent);
+		ent = next;
+	}
+	html("</div>");
+}
+
+static void print_blame_page(const struct object_id *oid, const char *path,
+			     const char *filename, const char *rev)
 {
 	enum object_type type;
 	char *buf;
 	unsigned long size;
-	struct strvec rev_argv = STRVEC_INIT;
 	struct rev_info revs;
 	struct blame_scoreboard sb;
-	struct blame_origin *o;
-	struct blame_entry *ent = NULL;
 
 	type = odb_read_object_info(the_repository->objects, oid, &size);
 	if (type == OBJ_BAD) {
@@ -194,24 +296,7 @@ static void print_object(const struct object_id *oid, const char *path,
 		return;
 	}
 
-	strvec_push(&rev_argv, "blame");
-	strvec_push(&rev_argv, rev);
-	repo_init_revisions(the_repository, &revs, NULL);
-	revs.diffopt.flags.allow_textconv = 1;
-	setup_revisions(rev_argv.nr, rev_argv.v, &revs, NULL);
-	init_scoreboard(&sb);
-	sb.revs = &revs;
-	sb.repo = the_repository;
-	sb.path = path;
-	setup_scoreboard(&sb, &o);
-	o->suspects = blame_entry_prepend(NULL, 0, sb.num_lines, o);
-	prio_queue_put(&sb.commits, o->commit);
-	blame_origin_decref(o);
-	sb.ent = NULL;
-	sb.path = path;
-	assign_blame(&sb, 0);
-	blame_sort_final(&sb);
-	blame_coalesce(&sb);
+	run_blame(&sb, &revs, path, rev);
 
 	cgit_set_title_from_path(path);
 
@@ -228,46 +313,19 @@ static void print_object(const struct object_id *oid, const char *path,
 	}
 	html("<table class='blame blob'>\n<tr>\n");
 
-	/* Commit hashes */
-	html("<td class='hashes'>");
-	for (ent = sb.ent; ent; ent = ent->next) {
-		html("<div class='alt'><pre>");
-		emit_blame_entry_hash(ent);
-		html("</pre></div>");
-	}
-	html("</td>\n");
+	emit_hashes(&sb);
 
-	/* Line numbers */
-	if (ctx.cfg.enable_tree_linenumbers) {
-		html("<td class='linenumbers'>");
-		for (ent = sb.ent; ent; ent = ent->next) {
-			html("<div class='alt'><pre>");
-			emit_blame_entry_linenumber(ent);
-			html("</pre></div>");
-		}
-		html("</td>\n");
-	}
+	if (ctx.cfg.enable_tree_linenumbers)
+		emit_linenumbers(&sb);
 
 	html("<td class='lines'><div>");
-
-	/* Colored bars behind lines */
-	html("<div>");
-	for (ent = sb.ent; ent; ) {
-		struct blame_entry *e = ent->next;
-		html("<div class='alt'><pre>");
-		emit_blame_entry_line_background(&sb, ent);
-		html("</pre></div>");
-		free(ent);
-		ent = e;
-	}
-	html("</div>");
+	emit_line_backgrounds(&sb);
 
 	free((void *)sb.final_buf);
 
-	/* Lines */
 	html("<pre><code>");
 	if (ctx.repo->source_filter) {
-		char *filter_arg = xstrdup(basename);
+		char *filter_arg = xstrdup(filename);
 		cgit_open_filter(ctx.repo->source_filter, filter_arg);
 		html_raw(buf, size);
 		cgit_close_filter(ctx.repo->source_filter);
@@ -282,35 +340,34 @@ static void print_object(const struct object_id *oid, const char *path,
 	html("</tr>\n</table>\n");
 
 cleanup:
-	/* The binary and oversized branches jump here with the layout still
-	 * open, so close it on every path rather than only the normal one. */
 	cgit_print_layout_end();
 	free_suspect_details();
 	free(buf);
 }
 
 static int walk_tree(const struct object_id *oid, struct strbuf *base,
-		     const char *pathname, unsigned mode, void *cbdata)
+		     const char *pathname, unsigned mode, void *data)
 {
-	struct walk_tree_context *walk_tree_ctx = cbdata;
+	struct walk_tree_context *walk = data;
 
 	// match_baselen is -1 when no path was given, which no length equals.
-	if (walk_tree_ctx->match_baselen >= 0 &&
-	    base->len == (size_t)walk_tree_ctx->match_baselen) {
+	if (walk->match_baselen >= 0 &&
+	    base->len == (size_t)walk->match_baselen) {
 		if (S_ISREG(mode)) {
-			struct strbuf buffer = STRBUF_INIT;
-			strbuf_addbuf(&buffer, base);
-			strbuf_addstr(&buffer, pathname);
-			print_object(oid, buffer.buf, pathname,
-				     walk_tree_ctx->curr_rev);
-			strbuf_release(&buffer);
-			walk_tree_ctx->state = 1;
+			struct strbuf fullpath = STRBUF_INIT;
+
+			strbuf_addbuf(&fullpath, base);
+			strbuf_addstr(&fullpath, pathname);
+			print_blame_page(oid, fullpath.buf, pathname,
+					 walk->rev);
+			strbuf_release(&fullpath);
+			walk->found = TARGET_FILE;
 		} else if (S_ISDIR(mode)) {
-			walk_tree_ctx->state = 2;
+			walk->found = TARGET_FOLDER;
 		}
 	} else if (base->len < INT_MAX
-			&& (int)base->len > walk_tree_ctx->match_baselen) {
-		walk_tree_ctx->state = 2;
+			&& (int)base->len > walk->match_baselen) {
+		walk->found = TARGET_FOLDER;
 	} else if (S_ISDIR(mode)) {
 		return READ_TREE_RECURSIVE;
 	}
@@ -319,9 +376,10 @@ static int walk_tree(const struct object_id *oid, struct strbuf *base,
 
 static int basedir_len(const char *path)
 {
-	const char *p = strrchr(path, '/');
-	if (p)
-		return p - path + 1;
+	const char *slash = strrchr(path, '/');
+
+	if (slash)
+		return slash - path + 1;
 	return 0;
 }
 
@@ -330,16 +388,20 @@ void cgit_print_blame(void)
 	const char *rev = ctx.qry.oid;
 	struct object_id oid;
 	struct commit *commit;
+	int path_len = ctx.qry.path ? strlen(ctx.qry.path) : 0;
+	// nowildcard_len matches len so git treats the path as literal rather
+	// than as a glob, which is what a request naming one file means.
 	struct pathspec_item path_items = {
 		.match = ctx.qry.path,
-		.len = ctx.qry.path ? strlen(ctx.qry.path) : 0
+		.len = path_len,
+		.nowildcard_len = path_len
 	};
 	struct pathspec paths = {
 		.nr = 1,
 		.items = &path_items
 	};
-	struct walk_tree_context walk_tree_ctx = {
-		.state = 0
+	struct walk_tree_context walk = {
+		.found = TARGET_MISSING
 	};
 
 	if (!rev)
@@ -357,17 +419,17 @@ void cgit_print_blame(void)
 		return;
 	}
 
-	walk_tree_ctx.curr_rev = xstrdup(rev);
-	walk_tree_ctx.match_baselen = (path_items.match) ?
-				       basedir_len(path_items.match) : -1;
+	walk.rev = xstrdup(rev);
+	walk.match_baselen = path_items.match ?
+			     basedir_len(path_items.match) : -1;
 
 	read_tree(the_repository, repo_get_commit_tree(the_repository, commit),
-		  &paths, walk_tree, &walk_tree_ctx);
-	if (!walk_tree_ctx.state)
+		  &paths, walk_tree, &walk);
+	if (walk.found == TARGET_MISSING)
 		cgit_print_error_page(404, "Not found", "Not found");
-	else if (walk_tree_ctx.state == 2)
+	else if (walk.found == TARGET_FOLDER)
 		cgit_print_error_page(404, "No blame for folders",
 			"Blame is not available for folders.");
 
-	free(walk_tree_ctx.curr_rev);
+	free(walk.rev);
 }
diff --git a/source/ui-blame.h b/source/ui-blame.h
index 5b97e03..f406698 100644
--- a/source/ui-blame.h
+++ b/source/ui-blame.h
@@ -1,6 +1,13 @@
-#ifndef UI_BLAME_H
-#define UI_BLAME_H
+/*
+ * The blame page, which pairs every line of a file with the commit that last
+ * changed it. It is one repository command among those in cmd.c, reached with
+ * a revision and a path in the query string, and a repository only offers it
+ * when enable-blame is set.
+ */
+
+#ifndef CGIT_UI_BLAME_H
+#define CGIT_UI_BLAME_H
 
 extern void cgit_print_blame(void);
 
-#endif /* UI_BLAME_H */
+#endif // CGIT_UI_BLAME_H
diff --git a/source/ui-blob.c b/source/ui-blob.c
index 92edcf1..beb968b 100644
--- a/source/ui-blob.c
+++ b/source/ui-blob.c
@@ -1,16 +1,17 @@
-/* ui-blob.c: show blob content
- *
- * Copyright (C) 2006-2014 cgit Development Team <cgit@lists.zx2c4.com>
- *
- * Licensed under GNU General Public License v2
- *   (see LICENSE.txt for full license text)
+/*
+ * Reads a blob out of the object database and writes its bytes to the client.
+ * The blob is named either directly by object id or by a path walked out of a
+ * commit's tree. The blob page serves those bytes as the whole response, under
+ * headers that stop a browser treating repository content as markup, while the
+ * summary page instead drops a readme's contents into a page it is already
+ * building.
  */
 
 #define USE_THE_REPOSITORY_VARIABLE
 
 #include "cgit.h"
-#include "ui-blob.h"
 #include "html.h"
+#include "ui-blob.h"
 #include "ui-shared.h"
 
 struct walk_tree_context {
@@ -20,93 +21,110 @@ struct walk_tree_context {
 	unsigned int file_only:1;
 };
 
+/*
+ * read_tree reads the return value as a direction rather than a status, so
+ * READ_TREE_RECURSIVE means step into this entry and zero means step over it.
+ */
 static int walk_tree(const struct object_id *oid, struct strbuf *base,
-		const char *pathname, unsigned mode, void *cbdata)
+		const char *pathname, unsigned mode, void *context)
 {
-	struct walk_tree_context *walk_tree_ctx = cbdata;
+	struct walk_tree_context *walk = context;
 
-	if (walk_tree_ctx->file_only && !S_ISREG(mode))
+	if (walk->file_only && !S_ISREG(mode))
 		return READ_TREE_RECURSIVE;
-	if (strncmp(base->buf, walk_tree_ctx->match_path, base->len)
-		|| strcmp(walk_tree_ctx->match_path + base->len, pathname))
+	if (strncmp(base->buf, walk->match_path, base->len)
+		|| strcmp(walk->match_path + base->len, pathname))
 		return READ_TREE_RECURSIVE;
-	oidcpy(walk_tree_ctx->matched_oid, oid);
-	walk_tree_ctx->found_path = 1;
+	oidcpy(walk->matched_oid, oid);
+	walk->found_path = 1;
 	return 0;
 }
 
-int cgit_ref_path_exists(const char *path, const char *ref, int file_only)
+/*
+ * oid comes in naming a commit and goes out naming the blob found at path,
+ * which both callers rely on. The path has to be writable because a pathspec
+ * item does not hold a const string.
+ */
+static int find_path_oid(struct object_id *oid, char *path, int file_only)
 {
-	struct object_id oid;
-	unsigned long size;
-	struct pathspec_item path_items = {
-		.match = xstrdup(path),
-		.len = strlen(path)
+	struct commit *commit = lookup_commit_reference(the_repository, oid);
+	// nowildcard_len matching len makes git treat the path as literal
+	// rather than as a glob.
+	struct pathspec_item item = {
+		.match = path,
+		.len = strlen(path),
+		.nowildcard_len = strlen(path)
 	};
 	struct pathspec paths = {
 		.nr = 1,
-		.items = &path_items
+		.items = &item
 	};
-	struct walk_tree_context walk_tree_ctx = {
+	struct walk_tree_context walk = {
 		.match_path = path,
-		.matched_oid = &oid,
+		.matched_oid = oid,
 		.found_path = 0,
 		.file_only = file_only
 	};
 
+	read_tree(the_repository, repo_get_commit_tree(the_repository, commit),
+		  &paths, walk_tree, &walk);
+	return walk.found_path;
+}
+
+/*
+ * Callers ask before reading the object, because the point of the limit is to
+ * keep a huge blob out of memory rather than to notice it once it is already
+ * there.
+ */
+static int over_size_limit(unsigned long size)
+{
+	return ctx.cfg.max_blob_size &&
+		size / 1024 > (unsigned long)ctx.cfg.max_blob_size;
+}
+
+int cgit_ref_path_exists(const char *path, const char *ref, int file_only)
+{
+	struct object_id oid;
+	unsigned long size;
+	char *path_copy = xstrdup(path);
+	int found = 0;
+
 	if (repo_get_oid(the_repository, ref, &oid))
 		goto done;
 	if (odb_read_object_info(the_repository->objects, &oid, &size) != OBJ_COMMIT)
 		goto done;
-	read_tree(the_repository,
-		  repo_get_commit_tree(the_repository, lookup_commit_reference(the_repository, &oid)),
-		  &paths, walk_tree, &walk_tree_ctx);
+	found = find_path_oid(&oid, path_copy, file_only);
 
 done:
-	free(path_items.match);
-	return walk_tree_ctx.found_path;
+	free(path_copy);
+	return found;
 }
 
 int cgit_print_file(char *path, const char *head, int file_only, int html_escape)
 {
 	struct object_id oid;
 	enum object_type type;
-	char *buf;
 	unsigned long size;
-	struct commit *commit;
-	struct pathspec_item path_items = {
-		.match = path,
-		.len = strlen(path)
-	};
-	struct pathspec paths = {
-		.nr = 1,
-		.items = &path_items
-	};
-	struct walk_tree_context walk_tree_ctx = {
-		.match_path = path,
-		.matched_oid = &oid,
-		.found_path = 0,
-		.file_only = file_only
-	};
+	char *buf;
 
 	if (repo_get_oid(the_repository, head, &oid))
 		return -1;
 	type = odb_read_object_info(the_repository->objects, &oid, &size);
 	if (type == OBJ_COMMIT) {
-		commit = lookup_commit_reference(the_repository, &oid);
-		read_tree(the_repository, repo_get_commit_tree(the_repository, commit),
-			  &paths, walk_tree, &walk_tree_ctx);
-		if (!walk_tree_ctx.found_path)
+		if (!find_path_oid(&oid, path, file_only))
 			return -1;
 		type = odb_read_object_info(the_repository->objects, &oid, &size);
 	}
 	if (type == OBJ_BAD)
 		return -1;
-	if (ctx.cfg.max_blob_size && size / 1024 > (unsigned long)ctx.cfg.max_blob_size)
+	if (over_size_limit(size))
 		return -1;
 	buf = odb_read_object(the_repository->objects, &oid, &type, &size);
 	if (!buf)
 		return -1;
+
+	// html_txt wants a terminated string, and git leaves a spare byte
+	// past every object it reads, so this write stays in the allocation.
 	buf[size] = '\0';
 	if (html_escape)
 		html_txt(buf);
@@ -120,23 +138,8 @@ void cgit_print_blob(const char *hex, char *path, const char *head, int file_onl
 {
 	struct object_id oid;
 	enum object_type type;
-	char *buf;
 	unsigned long size;
-	struct commit *commit;
-	struct pathspec_item path_items = {
-		.match = path,
-		.len = path ? strlen(path) : 0
-	};
-	struct pathspec paths = {
-		.nr = 1,
-		.items = &path_items
-	};
-	struct walk_tree_context walk_tree_ctx = {
-		.match_path = path,
-		.matched_oid = &oid,
-		.found_path = 0,
-		.file_only = file_only
-	};
+	char *buf;
 
 	if (hex) {
 		if (get_oid_hex(hex, &oid)) {
@@ -154,11 +157,8 @@ void cgit_print_blob(const char *hex, char *path, const char *head, int file_onl
 
 	type = odb_read_object_info(the_repository->objects, &oid, &size);
 
-	if ((!hex) && type == OBJ_COMMIT && path) {
-		commit = lookup_commit_reference(the_repository, &oid);
-		read_tree(the_repository, repo_get_commit_tree(the_repository, commit),
-			  &paths, walk_tree, &walk_tree_ctx);
-		if (!walk_tree_ctx.found_path) {
+	if (!hex && type == OBJ_COMMIT && path) {
+		if (!find_path_oid(&oid, path, file_only)) {
 			cgit_print_error_page(404, "Not found",
 					"Path not found: %s", path);
 			return;
@@ -172,8 +172,7 @@ void cgit_print_blob(const char *hex, char *path, const char *head, int file_onl
 		return;
 	}
 
-	/* Reject an oversized object before reading it whole into memory. */
-	if (ctx.cfg.max_blob_size && size / 1024 > (unsigned long)ctx.cfg.max_blob_size) {
+	if (over_size_limit(size)) {
 		cgit_print_error_page(413, "Too large",
 				"Object size (%luKB) exceeds limit (%dKB)",
 				size / 1024, ctx.cfg.max_blob_size);
@@ -194,6 +193,10 @@ void cgit_print_blob(const char *hex, char *path, const char *head, int file_onl
 		ctx.page.mimetype = "text/plain";
 	ctx.page.filename = path;
 
+	// The bytes are whatever the repository holds, so the browser is told
+	// not to guess a type of its own from them and not to load anything
+	// they reference. Both must go out before cgit_print_http_headers,
+	// which closes the header block.
 	html("X-Content-Type-Options: nosniff\n");
 	html("Content-Security-Policy: default-src 'none'\n");
 	cgit_print_http_headers();
diff --git a/source/ui-blob.h b/source/ui-blob.h
index efbc94e..eb3dbd1 100644
--- a/source/ui-blob.h
+++ b/source/ui-blob.h
@@ -1,8 +1,18 @@
-#ifndef UI_BLOB_H
-#define UI_BLOB_H
+/*
+ * Declarations for reading a blob out of the object database and writing it to
+ * the client. A caller either hands the whole response over to the blob's own
+ * bytes and headers, or drops a file's contents into a page cgit is already
+ * building, as the summary page does for a readme.
+ */
 
-extern int cgit_ref_path_exists(const char *path, const char *ref, int file_only);
-extern int cgit_print_file(char *path, const char *head, int file_only, int html_escape);
-extern void cgit_print_blob(const char *hex, char *path, const char *head, int file_only);
+#ifndef CGIT_UI_BLOB_H
+#define CGIT_UI_BLOB_H
 
-#endif /* UI_BLOB_H */
+extern int cgit_ref_path_exists(const char *path, const char *ref,
+				int file_only);
+extern int cgit_print_file(char *path, const char *head, int file_only,
+			   int html_escape);
+extern void cgit_print_blob(const char *hex, char *path, const char *head,
+			    int file_only);
+
+#endif // CGIT_UI_BLOB_H
diff --git a/source/ui-clone.c b/source/ui-clone.c
index 71bb2ed..70e4446 100644
--- a/source/ui-clone.c
+++ b/source/ui-clone.c
@@ -1,40 +1,64 @@
-/* ui-clone.c: functions for http cloning, based on
- * git's http-backend.c by Shawn O. Pearce
- *
- * Copyright (C) 2006-2014 cgit Development Team <cgit@lists.zx2c4.com>
- *
- * Licensed under GNU General Public License v2
- *   (see LICENSE.txt for full license text)
+/*
+ * The dumb HTTP transport, which lets a client clone by fetching plain files
+ * rather than by talking to a server side helper. It answers the requests
+ * such a client makes, the ref listing under info, the loose objects and pack
+ * files under objects, and HEAD, each written as raw bytes rather than as a
+ * page. The two listings a clone starts from are built per request, so a
+ * repository serves without anyone having run git update-server-info over it.
+ * The shape of it all follows git's own http-backend.
  */
 
 #define USE_THE_REPOSITORY_VARIABLE
 
 #include "cgit.h"
-#include "ui-clone.h"
 #include "html.h"
+#include "ui-clone.h"
 #include "ui-shared.h"
-#include "packfile.h"
 
-static int print_ref_info(const struct reference *ref, void *cb_data)
+/*
+ * A client reading this listing expects an annotated tag to be followed by a
+ * second line, marked with ^{}, naming the object the tag peels to, the
+ * format git update-server-info writes into info/refs.
+ */
+static int print_ref(const struct reference *ref, void *cb_data)
 {
 	struct object *obj;
 
-	if (!(obj = parse_object(the_repository, ref->oid)))
+	obj = parse_object(the_repository, ref->oid);
+	if (!obj)
 		return 0;
 
 	htmlf("%s\t%s\n", oid_to_hex(ref->oid), ref->name);
 	if (obj->type == OBJ_TAG) {
-		if (!(obj = deref_tag(the_repository, obj, ref->name, 0)))
+		obj = deref_tag(the_repository, obj, ref->name, 0);
+		if (!obj)
 			return 0;
 		htmlf("%s\t%s^{}\n", oid_to_hex(&obj->oid), ref->name);
 	}
 	return 0;
 }
 
+/*
+ * A path ending in a slash is handed back whole, where git's own
+ * pack_basename would return the empty string.
+ */
+static const char *last_path_component(const char *path)
+{
+	const char *slash = strrchr(path, '/');
+
+	if (slash && slash[1] != '\0')
+		return slash + 1;
+	return path;
+}
+
+/*
+ * Only the packs this repository holds itself are listed, because one
+ * borrowed from an alternate is not reachable below this URL and a client
+ * told about it would come back for a file that is not there.
+ */
 static void print_pack_info(void)
 {
 	struct odb_source *source;
-	char *offset;
 
 	ctx.page.mimetype = "text/plain";
 	ctx.page.filename = "objects/info/packs";
@@ -42,21 +66,38 @@ static void print_pack_info(void)
 	odb_reprepare(the_repository->objects);
 	for (source = the_repository->objects->sources; source; source = source->next) {
 		struct odb_source_files *files = odb_source_files_downcast(source);
-		struct packfile_list_entry *e;
-		for (e = files->packed->packs.head; e; e = e->next) {
-			struct packed_git *p = e->pack;
-			if (p->pack_local) {
-				offset = strrchr(p->pack_name, '/');
-				if (offset && offset[1] != '\0')
-					++offset;
-				else
-					offset = p->pack_name;
-				htmlf("P %s\n", offset);
-			}
+		struct packfile_list_entry *entry;
+		// Asked for through the accessor rather than read off the list,
+		// because a pack the multi-pack-index already covers joins that
+		// list only when the accessor loads it, and reading the field
+		// directly leaves those packs unfindable.
+		for (entry = packfile_store_get_packs(files->packed); entry;
+		     entry = entry->next) {
+			struct packed_git *pack = entry->pack;
+			if (pack->pack_local)
+				htmlf("P %s\n",
+				      last_path_component(pack->pack_name));
 		}
 	}
 }
 
+/*
+ * Beyond the obvious directory traversal, the strict character set heads off
+ * other funny business, for example the file name quirks of the Cygwin port.
+ */
+static int path_is_safe(const char *path)
+{
+	const char *p;
+
+	for (p = path; *p; ++p) {
+		if (*p == '.' && *(p + 1) == '.')
+			return 0;
+		if (!isalnum((unsigned char)*p) && *p != '/' && *p != '.' && *p != '-')
+			return 0;
+	}
+	return 1;
+}
+
 static void send_file(const char *path)
 {
 	struct stat st;
@@ -75,6 +116,8 @@ static void send_file(const char *path)
 		return;
 	}
 	ctx.page.mimetype = "application/octet-stream";
+	// Offer the file under its path inside the repository, so the layout
+	// of the server's disk stays out of the download name.
 	ctx.page.filename = path;
 	skip_prefix(path, ctx.repo->path, &ctx.page.filename);
 	skip_prefix(ctx.page.filename, "/", &ctx.page.filename);
@@ -93,12 +136,12 @@ void cgit_clone_info(void)
 	ctx.page.filename = "info/refs";
 	cgit_print_http_headers();
 	refs_for_each_ref(get_main_ref_store(the_repository),
-			  print_ref_info, NULL);
+			  print_ref, NULL);
 }
 
 void cgit_clone_objects(void)
 {
-	char *p, *path;
+	char *path;
 
 	if (!ctx.qry.path)
 		goto err;
@@ -108,16 +151,8 @@ void cgit_clone_objects(void)
 		return;
 	}
 
-	/* Avoid directory traversal by forbidding "..", but also work around
-	 * other funny business by just specifying a fairly strict format. For
-	 * example, now we don't have to stress out about the Cygwin port.
-	 */
-	for (p = ctx.qry.path; *p; ++p) {
-		if (*p == '.' && *(p + 1) == '.')
-			goto err;
-		if (!isalnum((unsigned char)*p) && *p != '/' && *p != '.' && *p != '-')
-			goto err;
-	}
+	if (!path_is_safe(ctx.qry.path))
+		goto err;
 
 	path = repo_git_path(the_repository, "objects/%s", ctx.qry.path);
 	send_file(path);
diff --git a/source/ui-clone.h b/source/ui-clone.h
index 3e460a3..71deff2 100644
--- a/source/ui-clone.h
+++ b/source/ui-clone.h
@@ -1,8 +1,15 @@
-#ifndef UI_CLONE_H
-#define UI_CLONE_H
+/*
+ * The entry points that serve a dumb HTTP clone, one for each kind of file
+ * such a client fetches. Each writes raw bytes rather than a page, and cmd.c
+ * flags them as clone requests so a site with http cloning turned off can
+ * refuse all of them in one place.
+ */
 
-void cgit_clone_info(void);
-void cgit_clone_objects(void);
-void cgit_clone_head(void);
+#ifndef CGIT_UI_CLONE_H
+#define CGIT_UI_CLONE_H
 
-#endif /* UI_CLONE_H */
+extern void cgit_clone_info(void);
+extern void cgit_clone_objects(void);
+extern void cgit_clone_head(void);
+
+#endif // CGIT_UI_CLONE_H
diff --git a/source/ui-commit.c b/source/ui-commit.c
index 4d6d4da..f9051c1 100644
--- a/source/ui-commit.c
+++ b/source/ui-commit.c
@@ -1,29 +1,93 @@
-/* ui-commit.c: generate commit view
- *
- * Copyright (C) 2006-2014 cgit Development Team <cgit@lists.zx2c4.com>
- *
- * Licensed under GNU General Public License v2
- *   (see LICENSE.txt for full license text)
+/*
+ * The commit page, which shows one commit on its own and is where the log and
+ * the ref listings link. It renders the idents, the object ids, the message
+ * and any note, then hands off to ui-diff.c for the diff against the first
+ * parent. Free text and addresses are written through the repository's commit
+ * and email filters, so a site can rewrite either on the way out.
  */
 
 #define USE_THE_REPOSITORY_VARIABLE
 
 #include "cgit.h"
-#include "ui-commit.h"
+#include "filter.h"
 #include "html.h"
-#include "ui-shared.h"
+#include "parsing.h"
+#include "shared.h"
+#include "ui-commit.h"
 #include "ui-diff.h"
 #include "ui-log.h"
+#include "ui-shared.h"
 
-void cgit_print_commit(char *hex, const char *prefix)
+// The diff below the message is always taken against the first parent alone,
+// which says little about a merge of many branches, so a commit with this
+// many parents or more is shown without a diff at all.
+#define OCTOPUS_PARENTS 3
+
+static void print_ident_row(const char *role, const char *name,
+			    const char *email, timestamp_t date, int tz)
+{
+	htmlf("<tr><th>%s</th><td>", role);
+	cgit_open_filter(ctx.repo->email_filter, email, "commit");
+	html_txt(name);
+	if (!ctx.cfg.noplainemail) {
+		html(" ");
+		html_txt(email);
+	}
+	cgit_close_filter(ctx.repo->email_filter);
+	html("</td><td class='right'>");
+	html_txt(show_date(date, tz, cgit_date_mode(DATE_ISO8601)));
+	html("</td></tr>\n");
+}
+
+static int print_parent_rows(struct commit *commit, const char *rev,
+			     const char *prefix)
 {
-	struct commit *commit, *parent;
-	struct commitinfo *info, *parent_info;
 	struct commit_list *p;
+	struct commit *parent;
+	const char *parent_hex, *label;
+	int parents = 0;
+
+	for (p = commit->parents; p; p = p->next) {
+		parent = lookup_commit_reference(the_repository,
+						 &p->item->object.oid);
+		if (!parent) {
+			html("<tr><td colspan='3'>");
+			cgit_print_error("Error reading parent commit");
+			html("</td></tr>");
+			continue;
+		}
+		html("<tr><th>parent</th>"
+		     "<td colspan='2' class='oid'>");
+		parent_hex = label = oid_to_hex(&p->item->object.oid);
+		if (ctx.repo->enable_subject_links)
+			label = cgit_parse_commit(parent)->subject;
+		cgit_commit_link(label, NULL, NULL, ctx.qry.head, parent_hex,
+				 prefix);
+		html(" (");
+		cgit_diff_link("diff", NULL, NULL, ctx.qry.head, rev,
+			       oid_to_hex(&p->item->object.oid), prefix);
+		html(")</td></tr>");
+		parents++;
+	}
+	return parents;
+}
+
+static void print_filtered_text(const char *text)
+{
+	cgit_open_filter(ctx.repo->commit_filter);
+	html_txt(text);
+	cgit_close_filter(ctx.repo->commit_filter);
+}
+
+void cgit_print_commit(char *hex, const char *prefix)
+{
+	struct commit *commit;
+	struct commitinfo *info;
 	struct strbuf notes = STRBUF_INIT;
 	struct object_id oid;
-	char *tmp, *tmp2;
-	int parents = 0;
+	const char *commit_hex, *first_parent;
+	char *tree_rev;
+	int parents;
 
 	if (!hex)
 		hex = ctx.qry.head;
@@ -48,102 +112,66 @@ void cgit_print_commit(char *hex, const char *prefix)
 	ctx.page.title = cgit_fmtalloc("%s - %s", info->subject, ctx.page.title);
 	cgit_print_layout_start();
 	cgit_print_diff_ctrls();
+
 	html("<table summary='commit info' class='commit-info'>\n");
-	html("<tr><th>author</th><td>");
-	cgit_open_filter(ctx.repo->email_filter, info->author_email, "commit");
-	html_txt(info->author);
-	if (!ctx.cfg.noplainemail) {
-		html(" ");
-		html_txt(info->author_email);
-	}
-	cgit_close_filter(ctx.repo->email_filter);
-	html("</td><td class='right'>");
-	html_txt(show_date(info->author_date, info->author_tz,
-				cgit_date_mode(DATE_ISO8601)));
-	html("</td></tr>\n");
-	html("<tr><th>committer</th><td>");
-	cgit_open_filter(ctx.repo->email_filter, info->committer_email, "commit");
-	html_txt(info->committer);
-	if (!ctx.cfg.noplainemail) {
-		html(" ");
-		html_txt(info->committer_email);
-	}
-	cgit_close_filter(ctx.repo->email_filter);
-	html("</td><td class='right'>");
-	html_txt(show_date(info->committer_date, info->committer_tz,
-				cgit_date_mode(DATE_ISO8601)));
-	html("</td></tr>\n");
+	print_ident_row("author", info->author, info->author_email,
+			info->author_date, info->author_tz);
+	print_ident_row("committer", info->committer, info->committer_email,
+			info->committer_date, info->committer_tz);
+
 	html("<tr><th>commit</th><td colspan='2' class='oid'>");
-	tmp = oid_to_hex(&commit->object.oid);
-	cgit_commit_link(tmp, NULL, NULL, ctx.qry.head, tmp, prefix);
+	commit_hex = oid_to_hex(&commit->object.oid);
+	cgit_commit_link(commit_hex, NULL, NULL, ctx.qry.head, commit_hex,
+			 prefix);
 	html(" (");
-	cgit_patch_link("patch", NULL, NULL, NULL, tmp, prefix);
+	cgit_patch_link("patch", NULL, NULL, NULL, commit_hex, prefix);
 	html(")</td></tr>\n");
+
 	html("<tr><th>tree</th><td colspan='2' class='oid'>");
-	tmp = xstrdup(hex);
+	tree_rev = xstrdup(hex);
 	cgit_tree_link(oid_to_hex(get_commit_tree_oid(commit)), NULL, NULL,
-		       ctx.qry.head, tmp, NULL);
+		       ctx.qry.head, tree_rev, NULL);
 	if (prefix) {
 		html(" /");
-		cgit_tree_link(prefix, NULL, NULL, ctx.qry.head, tmp, prefix);
+		cgit_tree_link(prefix, NULL, NULL, ctx.qry.head, tree_rev,
+			       prefix);
 	}
-	free(tmp);
+	free(tree_rev);
 	html("</td></tr>\n");
-	for (p = commit->parents; p; p = p->next) {
-		parent = lookup_commit_reference(the_repository, &p->item->object.oid);
-		if (!parent) {
-			html("<tr><td colspan='3'>");
-			cgit_print_error("Error reading parent commit");
-			html("</td></tr>");
-			continue;
-		}
-		html("<tr><th>parent</th>"
-		     "<td colspan='2' class='oid'>");
-		tmp = tmp2 = oid_to_hex(&p->item->object.oid);
-		if (ctx.repo->enable_subject_links) {
-			parent_info = cgit_parse_commit(parent);
-			tmp2 = parent_info->subject;
-		}
-		cgit_commit_link(tmp2, NULL, NULL, ctx.qry.head, tmp, prefix);
-		html(" (");
-		cgit_diff_link("diff", NULL, NULL, ctx.qry.head, hex,
-			       oid_to_hex(&p->item->object.oid), prefix);
-		html(")</td></tr>");
-		parents++;
-	}
+
+	parents = print_parent_rows(commit, hex, prefix);
+
 	if (ctx.repo->snapshots) {
 		html("<tr><th>download</th><td colspan='2' class='oid'>");
 		cgit_print_snapshot_links(ctx.repo, hex, "<br/>");
 		html("</td></tr>");
 	}
 	html("</table>\n");
+
 	html("<div class='commit-subject'>");
-	cgit_open_filter(ctx.repo->commit_filter);
-	html_txt(info->subject);
-	cgit_close_filter(ctx.repo->commit_filter);
+	print_filtered_text(info->subject);
 	cgit_print_commit_decorations(commit);
 	html("</div>");
 	html("<div class='commit-msg'>");
-	cgit_open_filter(ctx.repo->commit_filter);
-	html_txt(info->msg);
-	cgit_close_filter(ctx.repo->commit_filter);
+	print_filtered_text(info->msg);
 	html("</div>");
 	if (notes.len != 0) {
 		html("<div class='notes-header'>Notes</div>");
 		html("<div class='notes'>");
-		cgit_open_filter(ctx.repo->commit_filter);
-		html_txt(notes.buf);
-		cgit_close_filter(ctx.repo->commit_filter);
+		print_filtered_text(notes.buf);
 		html("</div>");
 		html("<div class='notes-footer'></div>");
 	}
-	if (parents < 3) {
+
+	if (parents < OCTOPUS_PARENTS) {
 		if (parents)
-			tmp = oid_to_hex(&commit->parents->item->object.oid);
+			first_parent =
+				oid_to_hex(&commit->parents->item->object.oid);
 		else
-			tmp = NULL;
-		cgit_print_diff(ctx.qry.oid, tmp, prefix, 0, 0);
+			first_parent = NULL;
+		cgit_print_diff(ctx.qry.oid, first_parent, prefix, 0, 0);
 	}
+
 	strbuf_release(&notes);
 	cgit_free_commitinfo(info);
 	cgit_print_layout_end();
diff --git a/source/ui-commit.h b/source/ui-commit.h
index 8198b4b..0cc0cb0 100644
--- a/source/ui-commit.h
+++ b/source/ui-commit.h
@@ -1,6 +1,12 @@
-#ifndef UI_COMMIT_H
-#define UI_COMMIT_H
+/*
+ * The commit page, which shows one commit on its own and is where the log and
+ * the ref listings link. It takes the commit to show plus an optional path
+ * that narrows the diff printed below the message.
+ */
+
+#ifndef CGIT_UI_COMMIT_H
+#define CGIT_UI_COMMIT_H
 
 extern void cgit_print_commit(char *hex, const char *prefix);
 
-#endif /* UI_COMMIT_H */
+#endif // CGIT_UI_COMMIT_H
diff --git a/source/ui-diff.c b/source/ui-diff.c
index a0790f7..09e2790 100644
--- a/source/ui-diff.c
+++ b/source/ui-diff.c
This diff is too large to be rendered inline. View it on its own page.
diff --git a/source/ui-diff.h b/source/ui-diff.h
index 39264a1..67d33ba 100644
--- a/source/ui-diff.h
+++ b/source/ui-diff.h
@@ -1,5 +1,19 @@
-#ifndef UI_DIFF_H
-#define UI_DIFF_H
+/*
+ * The page that shows what one revision changed against another, and the
+ * options panel that sits above it. The side by side renderer in ui-ssdiff
+ * builds a link on every line number, so the two revisions being compared and
+ * the file pair being rendered are published here for it rather than kept
+ * private to the page.
+ */
+
+#ifndef CGIT_UI_DIFF_H
+#define CGIT_UI_DIFF_H
+
+#include "cgit.h"
+
+// Widest context the options panel offers, and so the widest a request may
+// ask for, since the query string is clamped against it before the page runs.
+#define MAX_DIFF_CONTEXT_LINES 40
 
 extern void cgit_print_diff_ctrls(void);
 
@@ -12,4 +26,4 @@ extern struct diff_filespec *cgit_get_current_new_file(void);
 extern struct object_id old_rev_oid[1];
 extern struct object_id new_rev_oid[1];
 
-#endif /* UI_DIFF_H */
+#endif // CGIT_UI_DIFF_H
diff --git a/source/ui-empty.c b/source/ui-empty.c
index cb91fbc..56ca6e1 100644
--- a/source/ui-empty.c
+++ b/source/ui-empty.c
@@ -1,15 +1,14 @@
 /*
- * ui-empty.c: page shown for a repository that holds no commits
- *
- * Copyright (C) 2006-2014 cgit Development Team <cgit@lists.zx2c4.com>
- *
- * Licensed under GNU General Public License v2
- *   (see LICENSE.txt for full license text)
+ * The page a repository with no commits falls back to, an error line and then
+ * the clone urls someone needs in order to push the first commit. cgit.c
+ * prints this in place of whatever page was asked for as soon as it finds
+ * there is no branch to resolve, so none of the ref driven pages ever run
+ * against an empty repository.
  */
 
 #include "cgit.h"
-#include "ui-empty.h"
 #include "html.h"
+#include "ui-empty.h"
 #include "ui-shared.h"
 
 static void print_clone_url(const char *url)
@@ -31,6 +30,8 @@ void cgit_print_empty_repo(void)
 {
 	cgit_print_error("Repository seems to be empty");
 
+	// cgit_add_clone_urls draws on just these two, so with neither set the
+	// table below would be a Clone heading with no rows.
 	if (!ctx.repo->clone_url && !ctx.cfg.clone_prefix)
 		return;
 
diff --git a/source/ui-empty.h b/source/ui-empty.h
index dda303a..25ae049 100644
--- a/source/ui-empty.h
+++ b/source/ui-empty.h
@@ -1,6 +1,13 @@
-#ifndef UI_EMPTY_H
-#define UI_EMPTY_H
+/*
+ * The page shown for a repository that holds no commits, an error line saying
+ * it looks empty followed by the clone urls a first push would use. It is not
+ * one of the commands in cmd.c, because cgit.c reaches it directly when a
+ * repository has no branch for the requested page to work from.
+ */
+
+#ifndef CGIT_UI_EMPTY_H
+#define CGIT_UI_EMPTY_H
 
 extern void cgit_print_empty_repo(void);
 
-#endif /* UI_EMPTY_H */
+#endif // CGIT_UI_EMPTY_H
diff --git a/source/ui-log.c b/source/ui-log.c
index 5eb9a74..f80ed16 100644
--- a/source/ui-log.c
+++ b/source/ui-log.c
This diff is too large to be rendered inline. View it on its own page.
diff --git a/source/ui-log.h b/source/ui-log.h
index 562524b..ecfe85d 100644
--- a/source/ui-log.h
+++ b/source/ui-log.h
@@ -1,9 +1,16 @@
-#ifndef UI_LOG_H
-#define UI_LOG_H
+/*
+ * The listing of commits, shown by the log page and by the block of recent
+ * commits on a repository's summary page. The commit page borrows the
+ * decorations alone, so that branch and tag labels look the same wherever a
+ * commit is named.
+ */
+
+#ifndef CGIT_UI_LOG_H
+#define CGIT_UI_LOG_H
 
 extern void cgit_print_log(const char *tip, int ofs, int cnt, char *grep,
 			   char *pattern, const char *path, int pager,
 			   int commit_graph, int commit_sort);
 extern void cgit_print_commit_decorations(struct commit *commit);
 
-#endif /* UI_LOG_H */
+#endif // CGIT_UI_LOG_H
diff --git a/source/ui-patch.c b/source/ui-patch.c
index 29432ac..6f75b6a 100644
--- a/source/ui-patch.c
+++ b/source/ui-patch.c
@@ -1,83 +1,107 @@
-/* ui-patch.c: generate patch view
- *
- * Copyright (C) 2006-2014 cgit Development Team <cgit@lists.zx2c4.com>
- *
- * Licensed under GNU General Public License v2
- *   (see LICENSE.txt for full license text)
+/*
+ * The patch page, which serves a commit or a range of commits as plain text
+ * in the mail format git format-patch writes, so that a change read in cgit
+ * can be fed straight to git am. It is one of the repository commands in
+ * cmd.c, taking the newer revision from id, the older one from id2, and an
+ * optional path that narrows the diff. A merge carries no single patch and so
+ * drops out of a range, and max-patch-count bounds how many commits one
+ * request may emit.
  */
 
 #define USE_THE_REPOSITORY_VARIABLE
 
 #include "cgit.h"
-#include "ui-patch.h"
 #include "html.h"
+#include "ui-diff.h"
+#include "ui-patch.h"
 #include "ui-shared.h"
 
-/* two commit hashes with two dots in between and termination */
-#define REV_RANGE_LEN 2 * GIT_MAX_HEXSZ + 3
+// Room for two hex object ids, the two dots between them, and the null byte.
+#define REV_RANGE_LEN (2 * GIT_MAX_HEXSZ + 3)
 
-void cgit_print_patch(const char *new_rev, const char *old_rev,
-		      const char *prefix)
+// Where the format argument sits in the walk arguments below.
+#define FORMAT_ARG 2
+
+/*
+ * A root commit has no parent, so the old end is left null and the caller asks
+ * for that one commit instead.
+ */
+static int resolve_range(const char *new_rev, const char *old_rev,
+			 struct object_id *new_oid, struct object_id *old_oid)
 {
-	struct rev_info rev;
 	struct commit *commit;
-	struct object_id new_rev_oid, old_rev_oid;
-	char rev_range[REV_RANGE_LEN];
-	const char *rev_argv[] = { NULL, "--reverse", "--format=email", rev_range, "--", prefix, NULL };
-	int rev_argc = ARRAY_SIZE(rev_argv) - 1;
-	char *patchname;
 
-	if (!prefix)
-		rev_argc--;
-
-	if (!new_rev)
-		new_rev = ctx.qry.head;
-
-	if (repo_get_oid(the_repository, new_rev, &new_rev_oid)) {
+	if (repo_get_oid(the_repository, new_rev, new_oid)) {
 		cgit_print_error_page(404, "Not found",
 				"Bad object id: %s", new_rev);
-		return;
+		return -1;
 	}
-	commit = lookup_commit_reference(the_repository, &new_rev_oid);
+	commit = lookup_commit_reference(the_repository, new_oid);
 	if (!commit) {
 		cgit_print_error_page(404, "Not found",
 				"Bad commit reference: %s", new_rev);
-		return;
+		return -1;
 	}
 
 	if (old_rev) {
-		if (repo_get_oid(the_repository, old_rev, &old_rev_oid)) {
+		if (repo_get_oid(the_repository, old_rev, old_oid)) {
 			cgit_print_error_page(404, "Not found",
 					"Bad object id: %s", old_rev);
-			return;
+			return -1;
 		}
-		if (!lookup_commit_reference(the_repository, &old_rev_oid)) {
+		if (!lookup_commit_reference(the_repository, old_oid)) {
 			cgit_print_error_page(404, "Not found",
 					"Bad commit reference: %s", old_rev);
-			return;
+			return -1;
 		}
 	} else if (commit->parents && commit->parents->item) {
-		oidcpy(&old_rev_oid, &commit->parents->item->object.oid);
+		oidcpy(old_oid, &commit->parents->item->object.oid);
 	} else {
-		oidclr(&old_rev_oid, the_repository->hash_algo);
+		oidclr(old_oid, the_repository->hash_algo);
 	}
+	return 0;
+}
+
+void cgit_print_patch(const char *new_rev, const char *old_rev,
+		      const char *prefix)
+{
+	struct rev_info rev;
+	struct commit *commit;
+	struct object_id new_oid, old_oid;
+	char rev_range[REV_RANGE_LEN];
+	// setup_revisions reads these the way git reads a command line, so the
+	// first entry stands in for the program name and is skipped, and the
+	// array has to stay null terminated because the path after the double
+	// dash is picked up past the count.
+	const char *rev_argv[] = { NULL, "--reverse", "--format=email",
+				   rev_range, "--", prefix, NULL };
+	int rev_argc = ARRAY_SIZE(rev_argv) - 1;
+
+	if (!prefix)
+		rev_argc--;
+
+	if (!new_rev)
+		new_rev = ctx.qry.head;
+
+	if (resolve_range(new_rev, old_rev, &new_oid, &old_oid))
+		return;
 
-	if (is_null_oid(&old_rev_oid)) {
-		memcpy(rev_range, oid_to_hex(&new_rev_oid), the_hash_algo->hexsz + 1);
+	if (is_null_oid(&old_oid)) {
+		memcpy(rev_range, oid_to_hex(&new_oid), the_hash_algo->hexsz + 1);
 	} else {
-		xsnprintf(rev_range, REV_RANGE_LEN, "%s..%s", oid_to_hex(&old_rev_oid),
-			oid_to_hex(&new_rev_oid));
+		xsnprintf(rev_range, REV_RANGE_LEN, "%s..%s",
+			oid_to_hex(&old_oid), oid_to_hex(&new_oid));
 	}
 
-	patchname = cgit_fmt("%s.patch", rev_range);
 	ctx.page.mimetype = "text/plain";
-	ctx.page.filename = patchname;
+	ctx.page.filename = cgit_fmt("%s.patch", rev_range);
 	cgit_print_http_headers();
 
 	if (ctx.cfg.noplainemail) {
-		rev_argv[2] = "--format=format:From %H Mon Sep 17 00:00:00 "
-			      "2001%nFrom: %an%nDate: %aD%n%w(78,0,1)Subject: "
-			      "%s%n%n%w(0)%b";
+		rev_argv[FORMAT_ARG] =
+			"--format=format:From %H Mon Sep 17 00:00:00 "
+			"2001%nFrom: %an%nDate: %aD%n%w(78,0,1)Subject: "
+			"%s%n%n%w(0)%b";
 	}
 
 	repo_init_revisions(the_repository, &rev, NULL);
@@ -89,17 +113,25 @@ void cgit_print_patch(const char *new_rev, const char *old_rev,
 	rev.diffopt.output_format |= DIFF_FORMAT_DIFFSTAT |
 			DIFF_FORMAT_PATCH | DIFF_FORMAT_SUMMARY;
 	if (prefix)
-		rev.diffopt.stat_sep = cgit_fmt("(limited to '%s')\n\n", prefix);
+		// Allocated rather than formatted into cgit_fmt's fixed
+		// buffer, because the path comes from the request and a long
+		// one would abort the process here, with the headers for a
+		// successful response already on the wire.
+		rev.diffopt.stat_sep = cgit_fmtalloc("(limited to '%s')\n\n",
+						     prefix);
 	setup_revisions(rev_argc, rev_argv, &rev, NULL);
-	// A single commit resolves to a parent..commit range, so this only
-	// bounds an explicit id/id2 range and keeps one request from
-	// emitting a patch for the entire history.
+	// A single commit resolves to a range starting at its parent, so this
+	// only ever cuts an explicit range short and keeps one request from
+	// emitting a patch for the whole history.
 	if (ctx.cfg.max_patch_count > 0)
 		rev.max_count = ctx.cfg.max_patch_count;
 	prepare_revision_walk(&rev);
 
 	while ((commit = get_revision(&rev)) != NULL) {
 		log_tree_commit(&rev, commit);
+		// Two dashes and a space is the mail signature separator, so
+		// git am and mail readers drop the version note below rather
+		// than carrying it into the commit message.
 		printf("-- \ncgit %s\n\n", cgit_version);
 	}
 }
diff --git a/source/ui-patch.h b/source/ui-patch.h
index 7a6cacd..4dfe8bc 100644
--- a/source/ui-patch.h
+++ b/source/ui-patch.h
@@ -1,7 +1,14 @@
-#ifndef UI_PATCH_H
-#define UI_PATCH_H
+/*
+ * The patch page, which hands back a commit or a range of commits as a plain
+ * text mail patch that git am can apply. Any of the newer revision, the older
+ * one, and the path that limits the diff may be null, meaning the current
+ * head, the first parent, and the whole tree.
+ */
+
+#ifndef CGIT_UI_PATCH_H
+#define CGIT_UI_PATCH_H
 
 extern void cgit_print_patch(const char *new_rev, const char *old_rev,
 			     const char *prefix);
 
-#endif /* UI_PATCH_H */
+#endif // CGIT_UI_PATCH_H
diff --git a/source/ui-plain.c b/source/ui-plain.c
index 8085382..9ce25b3 100644
--- a/source/ui-plain.c
+++ b/source/ui-plain.c
@@ -1,23 +1,53 @@
-/* ui-plain.c: functions for output of plain blobs by path
- *
- * Copyright (C) 2006-2014 cgit Development Team <cgit@lists.zx2c4.com>
- *
- * Licensed under GNU General Public License v2
- *   (see LICENSE.txt for full license text)
+/*
+ * The plain page, which hands over a repository's own bytes rather than
+ * rendering a view of them. A file is written out whole under a content type
+ * guessed from its name, though a repository that has not enabled html serving
+ * keeps only the types a browser will not act on. A directory, or a request
+ * carrying no path, is answered with a bare document of links to the entries
+ * below it rather than with one of cgit's themed pages.
  */
 
 #define USE_THE_REPOSITORY_VARIABLE
 
 #include "cgit.h"
-#include "ui-plain.h"
 #include "html.h"
+#include "shared.h"
+#include "ui-plain.h"
 #include "ui-shared.h"
 
+/*
+ * A listing is opened by the entry that matched and closed only once the walk
+ * is over, so the end of the page has to tell the three cases apart.
+ */
+enum response {
+	RESPONSE_NONE,
+	RESPONSE_BLOB,
+	RESPONSE_LISTING
+};
+
 struct walk_tree_context {
-	int match_baselen;
-	int match;
+	// Length of the directory part of the requested path, slash included,
+	// and -1 when no path was requested so that no base length can equal
+	// it.
+	int dir_len;
+	enum response response;
 };
 
+/*
+ * Everything below text/ and application/ can carry markup or script that a
+ * browser would run against the site, so only PDF is let back through.
+ */
+static int is_unsafe_type(const char *mimetype)
+{
+	return (starts_with(mimetype, "text/") ||
+		starts_with(mimetype, "application/")) &&
+		strcmp(mimetype, "application/pdf");
+}
+
+/*
+ * A nonzero return says the response has been written, error pages included,
+ * so the walk does not go on to report the path as missing.
+ */
 static int print_object(const struct object_id *oid, const char *path)
 {
 	enum object_type type;
@@ -30,8 +60,10 @@ static int print_object(const struct object_id *oid, const char *path)
 		return 1;
 	}
 
-	/* Reject an oversized object before reading it whole into memory. */
-	if (ctx.cfg.max_blob_size && size / 1024 > (unsigned long)ctx.cfg.max_blob_size) {
+	// The limit counts kilobytes and is checked before the read, so a huge
+	// blob is kept out of memory rather than noticed once it is there.
+	if (ctx.cfg.max_blob_size &&
+	    size / 1024 > (unsigned long)ctx.cfg.max_blob_size) {
 		cgit_print_error_page(413, "Too large",
 			"Object size (%luKB) exceeds limit (%dKB)",
 			size / 1024, ctx.cfg.max_blob_size);
@@ -48,13 +80,14 @@ static int print_object(const struct object_id *oid, const char *path)
 	ctx.page.mimetype = mimetype;
 
 	if (!ctx.repo->enable_html_serving) {
+		// The bytes are whatever the repository holds, so the browser
+		// is told not to guess a type of its own and not to load
+		// anything they reference. Both lines must go out before
+		// cgit_print_http_headers, which closes the header block.
 		html("X-Content-Type-Options: nosniff\n");
 		html("Content-Security-Policy: default-src 'none'\n");
-		if (mimetype) {
-			/* Built-in white list allows PDF and everything that isn't text/ and application/ */
-			if ((!strncmp(mimetype, "text/", 5) || !strncmp(mimetype, "application/", 12)) && strcmp(mimetype, "application/pdf"))
-				ctx.page.mimetype = NULL;
-		}
+		if (mimetype && is_unsafe_type(mimetype))
+			ctx.page.mimetype = NULL;
 	}
 
 	if (!ctx.page.mimetype) {
@@ -75,7 +108,7 @@ static int print_object(const struct object_id *oid, const char *path)
 	return 1;
 }
 
-static char *buildpath(const char *base, int baselen, const char *path)
+static char *build_path(const char *base, int baselen, const char *path)
 {
 	if (path[0])
 		return cgit_fmtalloc("%.*s%s/", baselen, base, path);
@@ -86,20 +119,25 @@ static char *buildpath(const char *base, int baselen, const char *path)
 static void print_dir(const struct object_id *oid, const char *base,
 		      int baselen, const char *path)
 {
-	char *fullpath, *slash;
+	char *fullpath;
+	const char *leading_slash;
 	size_t len;
 
-	fullpath = buildpath(base, baselen, path);
-	slash = (fullpath[0] == '/' ? "" : "/");
+	fullpath = build_path(base, baselen, path);
+	leading_slash = (fullpath[0] == '/' ? "" : "/");
 	ctx.page.etag = oid_to_hex(oid);
 	cgit_print_http_headers();
-	htmlf("<html><head><title>%s", slash);
+	htmlf("<html><head><title>%s", leading_slash);
 	html_txt(fullpath);
-	htmlf("</title></head>\n<body>\n<h2>%s", slash);
+	htmlf("</title></head>\n<body>\n<h2>%s", leading_slash);
 	html_txt(fullpath);
 	html("</h2>\n<ul>\n");
 	len = strlen(fullpath);
 	if (len > 1) {
+		char *slash;
+
+		// Nothing left to drop means the parent is the root, which
+		// cgit_plain_link is asked for with a null path.
 		fullpath[len - 1] = 0;
 		slash = strrchr(fullpath, '/');
 		if (slash)
@@ -121,13 +159,13 @@ static void print_dir_entry(const struct object_id *oid, const char *base,
 {
 	char *fullpath;
 
-	fullpath = buildpath(base, baselen, path);
+	fullpath = build_path(base, baselen, path);
 	if (!S_ISDIR(mode) && !S_ISGITLINK(mode))
 		fullpath[strlen(fullpath) - 1] = 0;
 	html("  <li>");
-	if (S_ISGITLINK(mode)) {
+	if (S_ISGITLINK(mode))
 		cgit_submodule_link(NULL, fullpath, oid_to_hex(oid));
-	} else
+	else
 		cgit_plain_link(path, NULL, NULL, ctx.qry.head, ctx.qry.oid,
 				fullpath);
 	html("</li>\n");
@@ -139,25 +177,27 @@ static void print_dir_tail(void)
 	html(" </ul>\n</body></html>\n");
 }
 
+/*
+ * read_tree reads the return value as a direction rather than a status, so
+ * READ_TREE_RECURSIVE means step into this entry and zero means step over it.
+ */
 static int walk_tree(const struct object_id *oid, struct strbuf *base,
-		const char *pathname, unsigned mode, void *cbdata)
+		const char *pathname, unsigned mode, void *context)
 {
-	struct walk_tree_context *walk_tree_ctx = cbdata;
+	struct walk_tree_context *walk = context;
 
-	// match_baselen is -1 when no path was given, which no length equals.
-	if (walk_tree_ctx->match_baselen >= 0 &&
-	    base->len == (size_t)walk_tree_ctx->match_baselen) {
+	if (walk->dir_len >= 0 && base->len == (size_t)walk->dir_len) {
 		if (S_ISREG(mode) || S_ISLNK(mode)) {
 			if (print_object(oid, pathname))
-				walk_tree_ctx->match = 1;
+				walk->response = RESPONSE_BLOB;
 		} else if (S_ISDIR(mode)) {
 			print_dir(oid, base->buf, base->len, pathname);
-			walk_tree_ctx->match = 2;
+			walk->response = RESPONSE_LISTING;
 			return READ_TREE_RECURSIVE;
 		}
-	} else if (base->len < INT_MAX && (int)base->len > walk_tree_ctx->match_baselen) {
+	} else if (base->len < INT_MAX && (int)base->len > walk->dir_len) {
 		print_dir_entry(oid, base->buf, base->len, pathname, mode);
-		walk_tree_ctx->match = 2;
+		walk->response = RESPONSE_LISTING;
 	} else if (S_ISDIR(mode)) {
 		return READ_TREE_RECURSIVE;
 	}
@@ -165,11 +205,11 @@ static int walk_tree(const struct object_id *oid, struct strbuf *base,
 	return 0;
 }
 
-static int basedir_len(const char *path)
+static int dir_prefix_len(const char *path)
 {
-	const char *p = strrchr(path, '/');
-	if (p)
-		return p - path + 1;
+	const char *slash = strrchr(path, '/');
+	if (slash)
+		return slash - path + 1;
 	return 0;
 }
 
@@ -178,16 +218,22 @@ void cgit_print_plain(void)
 	const char *rev = ctx.qry.oid;
 	struct object_id oid;
 	struct commit *commit;
+	int path_len = ctx.qry.path ? strlen(ctx.qry.path) : 0;
+	// A hand built pathspec leaves nowildcard_len at zero, which tells git
+	// the match may be a glob. It would then hand this walk every entry a
+	// pattern like * matches, and each one would be answered with its own
+	// set of HTTP headers inside the body of the first.
 	struct pathspec_item path_items = {
 		.match = ctx.qry.path,
-		.len = ctx.qry.path ? strlen(ctx.qry.path) : 0
+		.len = path_len,
+		.nowildcard_len = path_len
 	};
 	struct pathspec paths = {
 		.nr = 1,
 		.items = &path_items
 	};
-	struct walk_tree_context walk_tree_ctx = {
-		.match = 0
+	struct walk_tree_context walk = {
+		.response = RESPONSE_NONE
 	};
 
 	if (!rev)
@@ -203,17 +249,19 @@ void cgit_print_plain(void)
 		return;
 	}
 	if (!path_items.match) {
+		// The walk is never handed an entry for the top of the tree
+		// itself, so the listing it would have opened is opened here.
 		path_items.match = "";
-		walk_tree_ctx.match_baselen = -1;
+		walk.dir_len = -1;
 		print_dir(get_commit_tree_oid(commit), "", 0, "");
-		walk_tree_ctx.match = 2;
+		walk.response = RESPONSE_LISTING;
+	} else {
+		walk.dir_len = dir_prefix_len(path_items.match);
 	}
-	else
-		walk_tree_ctx.match_baselen = basedir_len(path_items.match);
 	read_tree(the_repository, repo_get_commit_tree(the_repository, commit),
-		  &paths, walk_tree, &walk_tree_ctx);
-	if (!walk_tree_ctx.match)
+		  &paths, walk_tree, &walk);
+	if (walk.response == RESPONSE_NONE)
 		cgit_print_error_page(404, "Not found", "Not found");
-	else if (walk_tree_ctx.match == 2)
+	else if (walk.response == RESPONSE_LISTING)
 		print_dir_tail();
 }
diff --git a/source/ui-plain.h b/source/ui-plain.h
index 5bff07b..ec1eb7d 100644
--- a/source/ui-plain.h
+++ b/source/ui-plain.h
@@ -1,6 +1,12 @@
-#ifndef UI_PLAIN_H
-#define UI_PLAIN_H
+/*
+ * The plain page, which serves a path out of a repository as the bytes the
+ * repository holds, or as a bare listing of links when the path names a
+ * directory. The tree and blame views link to it for the raw file.
+ */
+
+#ifndef CGIT_UI_PLAIN_H
+#define CGIT_UI_PLAIN_H
 
 extern void cgit_print_plain(void);
 
-#endif /* UI_PLAIN_H */
+#endif // CGIT_UI_PLAIN_H
diff --git a/source/ui-refs.c b/source/ui-refs.c
index 84b929c..d179cdf 100644
--- a/source/ui-refs.c
+++ b/source/ui-refs.c
@@ -1,33 +1,38 @@
-/* ui-refs.c: browse symbolic refs
- *
- * Copyright (C) 2006-2014 cgit Development Team <cgit@lists.zx2c4.com>
- *
- * Licensed under GNU General Public License v2
- *   (see LICENSE.txt for full license text)
+/*
+ * The branch and tag listings, shown as sections of the summary page and as
+ * the whole of a repository's refs page. Every row is one ref beside
+ * the commit or tag object it points at, ordered newest first, with branches
+ * then reordered by name unless branch-sort asks for age. A section longer
+ * than max-ref-count is cut short and ends in a link to a dedicated heads or
+ * tags page, which walks the same list one offset at a time.
  */
 
 #define USE_THE_REPOSITORY_VARIABLE
 
 #include "cgit.h"
-#include "ui-refs.h"
+#include "filter.h"
 #include "html.h"
+#include "shared.h"
+#include "ui-refs.h"
 #include "ui-shared.h"
 
-static inline int cmp_age(int age1, int age2)
-{
-	/* age1 and age2 are assumed to be non-negative */
-	return age2 - age1;
-}
-
-static int cmp_ref_name(const void *a, const void *b)
-{
-	struct refinfo *r1 = *(struct refinfo **)a;
-	struct refinfo *r2 = *(struct refinfo **)b;
-
-	return strcmp(r1->refname, r2->refname);
-}
-
-static int get_ref_age(struct refinfo *ref)
+/*
+ * The slice of a sorted ref list that one page shows, with end one past the
+ * last row. size is what a full page holds, so it also decides whether the
+ * list needs a pager.
+ */
+struct ref_page {
+	int size;
+	int start;
+	int end;
+};
+
+/*
+ * The tagger date and the committer date live in a union in struct refinfo and
+ * only the member matching the object type is ever filled, so assuming a ref
+ * points at a commit reads past the end of the smaller struct.
+ */
+static timestamp_t ref_date(struct refinfo *ref)
 {
 	if (!ref->object)
 		return 0;
@@ -40,22 +45,53 @@ static int get_ref_age(struct refinfo *ref)
 	return 0;
 }
 
-// tag and commit share a union in struct refinfo and only the member matching
-// the object type is ever filled, so the date has to be reached through
-// get_ref_age rather than by assuming a branch points at a commit. Reading the
-// wrong member ran off the end of the smaller struct.
-static int cmp_ref_age(const void *a, const void *b)
+static int cmp_date(const void *a, const void *b)
+{
+	struct refinfo *ref1 = *(struct refinfo **)a;
+	struct refinfo *ref2 = *(struct refinfo **)b;
+	timestamp_t date1 = ref_date(ref1), date2 = ref_date(ref2);
+
+	// Compared rather than subtracted, at the width the dates are stored
+	// at. A commit may carry any timestamp, so a difference that does not
+	// fit an int would leave qsort with a contradictory ordering.
+	if (date1 < date2)
+		return 1;
+	if (date1 > date2)
+		return -1;
+	return 0;
+}
+
+static int cmp_name(const void *a, const void *b)
 {
-	struct refinfo *r1 = *(struct refinfo **)a;
-	struct refinfo *r2 = *(struct refinfo **)b;
+	struct refinfo *ref1 = *(struct refinfo **)a;
+	struct refinfo *ref2 = *(struct refinfo **)b;
+
+	return strcmp(ref1->refname, ref2->refname);
+}
+
+static void collect_branches(struct reflist *list)
+{
+	list->refs = NULL;
+	list->alloc = list->count = 0;
+	refs_for_each_branch_ref(get_main_ref_store(the_repository),
+				 cgit_refs_cb, list);
+	if (ctx.repo->enable_remote_branches)
+		refs_for_each_remote_ref(get_main_ref_store(the_repository),
+					 cgit_refs_cb, list);
+}
 
-	return cmp_age(get_ref_age(r1), get_ref_age(r2));
+static void print_branch_header(void)
+{
+	html("<tr class='nohover'><th class='left'>Branch</th>"
+	     "<th class='left'>Commit message</th>"
+	     "<th class='left col-author'>Author</th>"
+	     "<th class='left' colspan='2'>Age</th></tr>\n");
 }
 
 static int print_branch(struct refinfo *ref)
 {
 	struct commitinfo *info = ref->commit;
-	char *name = (char *)ref->refname;
+	const char *name = ref->refname;
 
 	if (!info)
 		return 1;
@@ -80,6 +116,14 @@ static int print_branch(struct refinfo *ref)
 	return 0;
 }
 
+static void collect_tags(struct reflist *list)
+{
+	list->refs = NULL;
+	list->alloc = list->count = 0;
+	refs_for_each_tag_ref(get_main_ref_store(the_repository),
+			      cgit_refs_cb, list);
+}
+
 static void print_tag_header(void)
 {
 	html("<tr class='nohover'><th class='left'>Tag</th>"
@@ -90,13 +134,15 @@ static void print_tag_header(void)
 
 static int print_tag(struct refinfo *ref)
 {
-	struct tag *tag = NULL;
 	struct taginfo *info = NULL;
-	char *name = (char *)ref->refname;
+	const char *name = ref->refname;
 	struct object *obj = ref->object;
 
+	// A lightweight tag has no tag object, so the author and age columns
+	// below fall back to the commit.
 	if (obj->type == OBJ_TAG) {
-		tag = (struct tag *)obj;
+		struct tag *tag = (struct tag *)obj;
+
 		obj = tag->tagged;
 		info = ref->tag;
 		if (!info)
@@ -141,16 +187,16 @@ static void print_refs_link(const char *path)
 	html("</td></tr>");
 }
 
-/* Prev/next links for the dedicated branch and tag pages, so each
- * category pages independently instead of sharing one endless page. */
 static void print_ref_pager(int ofs, int pagesize, int count, const char *path)
 {
 	char *url;
 
 	html("<tr class='nohover'><td colspan='5' class='refs-pager'>");
 	if (ofs > 0) {
+		int prev_ofs = ofs > pagesize ? ofs - pagesize : 0;
+
 		url = cgit_pageurl(ctx.qry.repo, cgit_fmt("refs/%s", path),
-				   cgit_fmt("ofs=%d", ofs > pagesize ? ofs - pagesize : 0));
+				   cgit_fmt("ofs=%d", prev_ofs));
 		html("<a href='");
 		html_attr(url);
 		html("'>[prev]</a> ");
@@ -169,132 +215,116 @@ static void print_ref_pager(int ofs, int pagesize, int count, const char *path)
 	html("</td></tr>");
 }
 
-static void collect_branches(struct reflist *list)
-{
-	list->refs = NULL;
-	list->alloc = list->count = 0;
-	refs_for_each_branch_ref(get_main_ref_store(the_repository),
-				 cgit_refs_cb, list);
-	if (ctx.repo->enable_remote_branches)
-		refs_for_each_remote_ref(get_main_ref_store(the_repository),
-					 cgit_refs_cb, list);
-}
-
-static void print_branch_header(void)
+static struct ref_page page_bounds(int pagesize, int count)
 {
-	html("<tr class='nohover'><th class='left'>Branch</th>"
-	     "<th class='left'>Commit message</th>"
-	     "<th class='left col-author'>Author</th>"
-	     "<th class='left' colspan='2'>Age</th></tr>\n");
+	struct ref_page page;
+
+	page.size = (pagesize <= 0 || pagesize > count) ? count : pagesize;
+	page.start = ctx.qry.ofs > 0 ? ctx.qry.ofs : 0;
+	if (page.start > count)
+		page.start = count;
+	page.end = page.start + page.size < count ?
+		   page.start + page.size : count;
+	return page;
 }
 
-void cgit_print_branches(int maxcount)
+/*
+ * Unlike the capped sections, the whole list is sorted before a page is cut
+ * out of it, so a branch keeps its place no matter which page it lands on.
+ */
+static void print_branches_page(int pagesize)
 {
 	struct reflist list;
+	struct ref_page page;
 	int i;
 
 	print_branch_header();
 	collect_branches(&list);
 
-	if (maxcount == 0 || maxcount > list.count)
-		maxcount = list.count;
-
-	qsort(list.refs, list.count, sizeof(*list.refs), cmp_ref_age);
+	qsort(list.refs, list.count, sizeof(*list.refs), cmp_date);
 	if (ctx.repo->branch_sort == 0)
-		qsort(list.refs, maxcount, sizeof(*list.refs), cmp_ref_name);
+		qsort(list.refs, list.count, sizeof(*list.refs), cmp_name);
 
-	for (i = 0; i < maxcount; i++)
+	page = page_bounds(pagesize, list.count);
+
+	for (i = page.start; i < page.end; i++)
 		print_branch(list.refs[i]);
 
-	if (maxcount < list.count)
-		print_refs_link("heads");
+	if (page.size < list.count)
+		print_ref_pager(page.start, page.size, list.count, "heads");
 
 	cgit_free_reflist_inner(&list);
 }
 
-void cgit_print_tags(int maxcount)
+static void print_tags_page(int pagesize)
 {
 	struct reflist list;
+	struct ref_page page;
 	int i;
 
-	list.refs = NULL;
-	list.alloc = list.count = 0;
-	refs_for_each_tag_ref(get_main_ref_store(the_repository),
-			      cgit_refs_cb, &list);
+	collect_tags(&list);
 	if (list.count == 0)
 		return;
-	qsort(list.refs, list.count, sizeof(*list.refs), cmp_ref_age);
-	if (!maxcount)
-		maxcount = list.count;
-	else if (maxcount > list.count)
-		maxcount = list.count;
+	qsort(list.refs, list.count, sizeof(*list.refs), cmp_date);
+
+	page = page_bounds(pagesize, list.count);
+
 	print_tag_header();
-	for (i = 0; i < maxcount; i++)
+	for (i = page.start; i < page.end; i++)
 		print_tag(list.refs[i]);
 
-	if (maxcount < list.count)
-		print_refs_link("tags");
+	if (page.size < list.count)
+		print_ref_pager(page.start, page.size, list.count, "tags");
 
 	cgit_free_reflist_inner(&list);
 }
 
-/* The dedicated branch page lists everything, a page at a time. The
- * whole list is name-sorted (or age-sorted per branch-sort) so the
- * order is stable across pages. */
-static void print_branches_page(int pagesize)
+void cgit_print_branches(int maxcount)
 {
 	struct reflist list;
-	int i, ofs, end;
+	int i;
 
 	print_branch_header();
 	collect_branches(&list);
 
-	qsort(list.refs, list.count, sizeof(*list.refs), cmp_ref_age);
-	if (ctx.repo->branch_sort == 0)
-		qsort(list.refs, list.count, sizeof(*list.refs), cmp_ref_name);
+	if (maxcount == 0 || maxcount > list.count)
+		maxcount = list.count;
 
-	if (pagesize <= 0 || pagesize > list.count)
-		pagesize = list.count;
-	ofs = ctx.qry.ofs > 0 ? ctx.qry.ofs : 0;
-	if (ofs > list.count)
-		ofs = list.count;
-	end = ofs + pagesize < list.count ? ofs + pagesize : list.count;
+	// The date sort covers the list, the name sort only the rows about to
+	// be shown, so the section holds the newest branches rather than the
+	// first ones by name.
+	qsort(list.refs, list.count, sizeof(*list.refs), cmp_date);
+	if (ctx.repo->branch_sort == 0)
+		qsort(list.refs, maxcount, sizeof(*list.refs), cmp_name);
 
-	for (i = ofs; i < end; i++)
+	for (i = 0; i < maxcount; i++)
 		print_branch(list.refs[i]);
 
-	if (pagesize < list.count)
-		print_ref_pager(ofs, pagesize, list.count, "heads");
+	if (maxcount < list.count)
+		print_refs_link("heads");
 
 	cgit_free_reflist_inner(&list);
 }
 
-static void print_tags_page(int pagesize)
+void cgit_print_tags(int maxcount)
 {
 	struct reflist list;
-	int i, ofs, end;
+	int i;
 
-	list.refs = NULL;
-	list.alloc = list.count = 0;
-	refs_for_each_tag_ref(get_main_ref_store(the_repository),
-			      cgit_refs_cb, &list);
+	collect_tags(&list);
 	if (list.count == 0)
 		return;
-	qsort(list.refs, list.count, sizeof(*list.refs), cmp_ref_age);
-
-	if (pagesize <= 0 || pagesize > list.count)
-		pagesize = list.count;
-	ofs = ctx.qry.ofs > 0 ? ctx.qry.ofs : 0;
-	if (ofs > list.count)
-		ofs = list.count;
-	end = ofs + pagesize < list.count ? ofs + pagesize : list.count;
-
+	qsort(list.refs, list.count, sizeof(*list.refs), cmp_date);
+	if (!maxcount)
+		maxcount = list.count;
+	else if (maxcount > list.count)
+		maxcount = list.count;
 	print_tag_header();
-	for (i = ofs; i < end; i++)
+	for (i = 0; i < maxcount; i++)
 		print_tag(list.refs[i]);
 
-	if (pagesize < list.count)
-		print_ref_pager(ofs, pagesize, list.count, "tags");
+	if (maxcount < list.count)
+		print_refs_link("tags");
 
 	cgit_free_reflist_inner(&list);
 }
@@ -309,8 +339,6 @@ void cgit_print_refs(void)
 	else if (ctx.qry.path && starts_with(ctx.qry.path, "tags"))
 		print_tags_page(ctx.cfg.max_ref_count);
 	else {
-		/* The combined page caps each section, with the [...] rows
-		 * leading to the dedicated pages above. */
 		cgit_print_branches(ctx.cfg.max_ref_count);
 		html("<tr class='nohover'><td colspan='5'>&nbsp;</td></tr>");
 		cgit_print_tags(ctx.cfg.max_ref_count);
diff --git a/source/ui-refs.h b/source/ui-refs.h
index 1d4a54a..c66050b 100644
--- a/source/ui-refs.h
+++ b/source/ui-refs.h
@@ -1,8 +1,25 @@
-#ifndef UI_REFS_H
-#define UI_REFS_H
+/*
+ * The branch and tag listings of a repository. The two section printers write
+ * rows into a table the caller has already opened, which is how the summary
+ * page carries a short version of each above its recent log.
+ */
 
+#ifndef CGIT_UI_REFS_H
+#define CGIT_UI_REFS_H
+
+/*
+ * Rows for the newest branches or tags, at most maxcount of them, where a
+ * maxcount of zero means every one. A section that leaves refs out ends in a
+ * link to the matching form of the refs page.
+ */
 extern void cgit_print_branches(int maxcount);
 extern void cgit_print_tags(int maxcount);
+
+/*
+ * The whole refs page. Without a path it holds a capped section of each kind,
+ * and with a path of heads or tags it becomes a page of that one kind alone,
+ * walked an offset at a time.
+ */
 extern void cgit_print_refs(void);
 
-#endif /* UI_REFS_H */
+#endif // CGIT_UI_REFS_H
diff --git a/source/ui-repolist.c b/source/ui-repolist.c
index e2c4abf..3bdae37 100644
--- a/source/ui-repolist.c
+++ b/source/ui-repolist.c
@@ -1,113 +1,137 @@
-/* ui-repolist.c: functions for generating the repolist page
- *
- * Copyright (C) 2006-2014 cgit Development Team <cgit@lists.zx2c4.com>
- *
- * Licensed under GNU General Public License v2
- *   (see LICENSE.txt for full license text)
+/*
+ * The repository index, which is the page a cgit site opens on, and the site
+ * readme the about page falls back to when a request names no repository.
+ * Every repository cgitrc registered is a candidate row, cut down to those
+ * matching the search terms and the url prefix the request carried, ordered
+ * either by the column the reader asked for or by section, and split into
+ * pages of max-repo-count rows.
  */
 
 #include "cgit.h"
-#include "ui-repolist.h"
+#include "filter.h"
 #include "html.h"
+#include "shared.h"
+#include "ui-repolist.h"
 #include "ui-shared.h"
 
+// A section heading spans the whole table, so this has to stay in step with
+// the headings print_header_row emits. Name, description and idle are always
+// there, owner and links are added when cgitrc asks for them.
+#define BASE_COLUMNS 3
+
+struct sort_column {
+	const char *name;
+	int (*cmp)(const void *a, const void *b);
+};
+
 static time_t read_agefile(const char *path)
 {
 	time_t result;
 	size_t size;
 	char *buf = NULL;
-	struct strbuf date_buf = STRBUF_INIT;
+	struct strbuf date = STRBUF_INIT;
 
 	if (cgit_read_first_line(path, &buf, &size)) {
 		free(buf);
 		return 0;
 	}
 
-	if (parse_date(buf, &date_buf) == 0)
-		result = strtoul(date_buf.buf, NULL, 10);
+	// parse_date writes the timestamp followed by its timezone offset, so
+	// only the seconds at the front are wanted here.
+	if (parse_date(buf, &date) == 0)
+		result = strtoul(date.buf, NULL, 10);
 	else
 		result = 0;
 	free(buf);
-	strbuf_release(&date_buf);
+	strbuf_release(&date);
 	return result;
 }
 
+/*
+ * The returned name can point into head, which the caller has to keep alive
+ * for as long as it uses it.
+ */
+static const char *tip_branch(const struct cgit_repo *repo, struct strbuf *head)
+{
+	struct strbuf path = STRBUF_INIT;
+	const char *branch = repo->defbranch;
+
+	if (branch)
+		return branch;
+
+	strbuf_addf(&path, "%s/HEAD", repo->path);
+	if (strbuf_read_file(head, path.buf, 0) > 0) {
+		strbuf_rtrim(head);
+		if (!skip_prefix(head->buf, "ref: refs/heads/", &branch))
+			branch = NULL;
+		// HEAD belongs to the repository, so a crafted target such as
+		// "ref: refs/heads/../../.." must not let the caller's stat
+		// walk outside it. Git forbids ".." in a ref name anyway.
+		if (branch && strstr(branch, ".."))
+			branch = NULL;
+	}
+	strbuf_release(&path);
+
+	return branch ? branch : "master";
+}
+
 static int get_repo_modtime(const struct cgit_repo *repo, time_t *mtime)
 {
 	struct strbuf path = STRBUF_INIT;
 	struct strbuf head = STRBUF_INIT;
-	struct stat s;
-	struct cgit_repo *r = (struct cgit_repo *)repo;
+	struct stat st;
 	const char *branch;
+	// The comparators below are handed const repositories, but the answer
+	// is kept in the repository itself so one page stats it only once.
+	struct cgit_repo *writable = (struct cgit_repo *)repo;
 
 	if (repo->mtime != -1) {
 		*mtime = repo->mtime;
 		return 1;
 	}
 	strbuf_addf(&path, "%s/%s", repo->path, ctx.cfg.agefile);
-	if (stat(path.buf, &s) == 0) {
+	if (stat(path.buf, &st) == 0) {
 		*mtime = read_agefile(path.buf);
 		if (*mtime) {
-			r->mtime = *mtime;
+			writable->mtime = *mtime;
 			goto end;
 		}
 	}
 
-	/* Stat the tip of the default branch. Prefer a configured defbranch,
-	 * otherwise read HEAD so a repo on "main" (or any branch name) is
-	 * handled, not only "master".
-	 */
-	branch = repo->defbranch;
-	if (!branch) {
-		strbuf_reset(&path);
-		strbuf_addf(&path, "%s/HEAD", repo->path);
-		if (strbuf_read_file(&head, path.buf, 0) > 0) {
-			strbuf_rtrim(&head);
-			if (!skip_prefix(head.buf, "ref: refs/heads/", &branch))
-				branch = NULL;
-			/* HEAD is repo-controlled, so a crafted target such as
-			 * "ref: refs/heads/../../.." must not let the stat()
-			 * below walk outside the repository. Git forbids ".."
-			 * in ref names anyway. */
-			if (branch && strstr(branch, ".."))
-				branch = NULL;
-		}
-		if (!branch)
-			branch = "master";
-	}
-
+	branch = tip_branch(repo, &head);
 	strbuf_reset(&path);
 	strbuf_addf(&path, "%s/refs/heads/%s", repo->path, branch);
-	if (stat(path.buf, &s) == 0) {
-		*mtime = s.st_mtime;
-		r->mtime = *mtime;
+	if (stat(path.buf, &st) == 0) {
+		*mtime = st.st_mtime;
+		writable->mtime = *mtime;
 		goto end;
 	}
 
 	strbuf_reset(&path);
-	strbuf_addf(&path, "%s/%s", repo->path, "packed-refs");
-	if (stat(path.buf, &s) == 0) {
-		*mtime = s.st_mtime;
-		r->mtime = *mtime;
+	strbuf_addf(&path, "%s/packed-refs", repo->path);
+	if (stat(path.buf, &st) == 0) {
+		*mtime = st.st_mtime;
+		writable->mtime = *mtime;
 		goto end;
 	}
 
 	*mtime = 0;
-	r->mtime = *mtime;
+	writable->mtime = *mtime;
 end:
 	strbuf_release(&path);
 	strbuf_release(&head);
-	return (r->mtime != 0);
+	return (writable->mtime != 0);
 }
 
 static void print_modtime(struct cgit_repo *repo)
 {
-	time_t t;
-	if (get_repo_modtime(repo, &t))
-		cgit_print_age(t, 0, -1);
+	time_t mtime;
+
+	if (get_repo_modtime(repo, &mtime))
+		cgit_print_age(mtime, 0, -1);
 }
 
-static int is_match(struct cgit_repo *repo)
+static int matches_search(struct cgit_repo *repo)
 {
 	if (!ctx.qry.search)
 		return 1;
@@ -122,7 +146,7 @@ static int is_match(struct cgit_repo *repo)
 	return 0;
 }
 
-static int is_in_url(struct cgit_repo *repo)
+static int matches_url(struct cgit_repo *repo)
 {
 	if (!ctx.qry.url)
 		return 1;
@@ -135,7 +159,7 @@ static int is_visible(struct cgit_repo *repo)
 {
 	if (repo->hide || repo->ignore)
 		return 0;
-	if (!(is_match(repo) && is_in_url(repo)))
+	if (!(matches_search(repo) && matches_url(repo)))
 		return 0;
 	return 1;
 }
@@ -151,14 +175,14 @@ static int any_repos_visible(void)
 	return 0;
 }
 
-// The index url is the same for every heading and every row, so the caller
-// works it out once rather than building and freeing one per cell.
-static void print_sort_header(const char *title, const char *sort,
-			      const char *currenturl)
+// currenturl is passed in because it is the same for every heading and every
+// row, and working it out here would mean an allocation and a free per cell.
+static void print_column_header(const char *title, const char *column,
+				const char *currenturl)
 {
-	htmlf("<th class='left col-%s'><a href='", sort);
+	htmlf("<th class='left col-%s'><a href='", column);
 	html_attr(currenturl);
-	htmlf("?s=%s", sort);
+	htmlf("?s=%s", column);
 	if (ctx.qry.search) {
 		html("&amp;q=");
 		html_url_arg(ctx.qry.search);
@@ -166,26 +190,83 @@ static void print_sort_header(const char *title, const char *sort,
 	htmlf("'>%s</a></th>", title);
 }
 
-static void print_header(const char *currenturl)
+static void print_header_row(const char *currenturl)
 {
 	html("<tr class='nohover'>");
-	print_sort_header("Name", "name", currenturl);
-	print_sort_header("Description", "desc", currenturl);
+	print_column_header("Name", "name", currenturl);
+	print_column_header("Description", "desc", currenturl);
 	if (ctx.cfg.enable_index_owner)
-		print_sort_header("Owner", "owner", currenturl);
-	print_sort_header("Idle", "idle", currenturl);
+		print_column_header("Owner", "owner", currenturl);
+	print_column_header("Idle", "idle", currenturl);
 	if (ctx.cfg.enable_index_links)
 		html("<th class='left col-links'>Links</th>");
 	html("</tr>\n");
 }
 
+static int section_changed(const char *section, const char *last)
+{
+	if (!section && !last)
+		return 0;
+	if (!section || !last)
+		return 1;
+	return strcmp(section, last) != 0;
+}
+
+static void print_section_row(const char *section, int columns)
+{
+	htmlf("<tr class='nohover-highlight'><td colspan='%d' class='reposection'>",
+	      columns);
+	html_txt(section);
+	html("</td></tr>");
+}
 
-static void print_pager(int items, int pagelen, char *search, char *sort)
+static void print_repo_row(const char *currenturl, int sublevel)
+{
+	char *repourl;
+
+	htmlf("<tr><td class='col-name %s'>",
+	      sublevel ? "sublevel-repo" : "toplevel-repo");
+	cgit_summary_link(ctx.repo->name, NULL, NULL, NULL);
+	html("</td><td class='col-desc'>");
+	repourl = cgit_repourl(ctx.repo->url);
+	html_link_open(repourl, NULL, NULL);
+	free(repourl);
+	if (html_ntxt(ctx.repo->desc, ctx.cfg.max_repodesc_len) < 0)
+		html("...");
+	html_link_close();
+	html("</td>");
+	if (ctx.cfg.enable_index_owner) {
+		html("<td class='col-owner'>");
+		html("<a href='");
+		html_attr(currenturl);
+		html("?q=");
+		html_url_arg(ctx.repo->owner);
+		html("'>");
+		html_txt(ctx.repo->owner);
+		html("</a>");
+		html("</td>");
+	}
+	html("<td class='col-idle'>");
+	print_modtime(ctx.repo);
+	html("</td>");
+	if (ctx.cfg.enable_index_links) {
+		html("<td class='col-links'>");
+		cgit_summary_link("summary", NULL, "button", NULL);
+		cgit_log_link("log", NULL, "button", NULL, NULL, NULL,
+			      0, NULL, NULL, ctx.qry.showmsg, 0);
+		cgit_tree_link("tree", NULL, "button", NULL, NULL, NULL);
+		html("</td>");
+	}
+	html("</tr>\n");
+}
+
+static void print_pager(int total, int pagelen, char *search, char *sort)
 {
 	int i, ofs;
 	char *class = NULL;
+
 	html("<ul class='pager'>");
-	for (i = 0, ofs = 0; ofs < items; i++, ofs = i * pagelen) {
+	for (i = 0, ofs = 0; ofs < total; i++, ofs = i * pagelen) {
 		class = (ctx.qry.ofs == ofs) ? "current" : NULL;
 		html("<li>");
 		cgit_index_link(cgit_fmt("[%d]", i + 1), cgit_fmt("Page %d", i + 1),
@@ -195,7 +276,7 @@ static void print_pager(int items, int pagelen, char *search, char *sort)
 	html("</ul>");
 }
 
-static int cmp(const char *s1, const char *s2)
+static int cmp_str(const char *s1, const char *s2)
 {
 	if (s1 && s2) {
 		if (ctx.cfg.case_sensitive_sort)
@@ -210,44 +291,31 @@ static int cmp(const char *s1, const char *s2)
 	return 0;
 }
 
-static int sort_name(const void *a, const void *b)
+static int cmp_name(const void *a, const void *b)
 {
 	const struct cgit_repo *r1 = a;
 	const struct cgit_repo *r2 = b;
 
-	return cmp(r1->name, r2->name);
+	return cmp_str(r1->name, r2->name);
 }
 
-static int sort_desc(const void *a, const void *b)
+static int cmp_desc(const void *a, const void *b)
 {
 	const struct cgit_repo *r1 = a;
 	const struct cgit_repo *r2 = b;
 
-	return cmp(r1->desc, r2->desc);
+	return cmp_str(r1->desc, r2->desc);
 }
 
-static int sort_owner(const void *a, const void *b)
+static int cmp_owner(const void *a, const void *b)
 {
 	const struct cgit_repo *r1 = a;
 	const struct cgit_repo *r2 = b;
 
-	return cmp(r1->owner, r2->owner);
-}
-
-/* Resolve every repository's modification time up front. get_repo_modtime
- * caches into the repo it is given, but qsort moves those structs around while
- * it sorts, so a comparator that fills the cache loses most of what it stored
- * and stats the same repository again and again. */
-static void resolve_modtimes(void)
-{
-	time_t t;
-	int i;
-
-	for (i = 0; i < cgit_repolist.count; i++)
-		get_repo_modtime(&cgit_repolist.repos[i], &t);
+	return cmp_str(r1->owner, r2->owner);
 }
 
-static int sort_idle(const void *a, const void *b)
+static int cmp_idle(const void *a, const void *b)
 {
 	const struct cgit_repo *r1 = a;
 	const struct cgit_repo *r2 = b;
@@ -256,8 +324,9 @@ static int sort_idle(const void *a, const void *b)
 	t1 = t2 = 0;
 	get_repo_modtime(r1, &t1);
 	get_repo_modtime(r2, &t2);
-	/* Return the sign only; a truncated 64-bit time_t difference could
-	 * flip and make the comparator inconsistent. */
+	// Only the sign is returned, because a 64-bit difference truncated
+	// into an int could come back with the wrong sign and leave the
+	// ordering inconsistent.
 	if (t2 > t1)
 		return 1;
 	if (t2 < t1)
@@ -265,61 +334,70 @@ static int sort_idle(const void *a, const void *b)
 	return 0;
 }
 
-static int sort_section(const void *a, const void *b)
+static int cmp_section(const void *a, const void *b)
 {
 	const struct cgit_repo *r1 = a;
 	const struct cgit_repo *r2 = b;
 	int result;
 
-	result = cmp(r1->section, r2->section);
+	result = cmp_str(r1->section, r2->section);
 	if (!result) {
 		if (!strcmp(ctx.cfg.repository_sort, "age"))
-			result = sort_idle(r1, r2);
+			result = cmp_idle(r1, r2);
 		if (!result)
-			result = cmp(r1->name, r2->name);
+			result = cmp_str(r1->name, r2->name);
 	}
 	return result;
 }
 
-struct sortcolumn {
-	const char *name;
-	int (*fn)(const void *a, const void *b);
-};
+/*
+ * get_repo_modtime caches into the repository it is handed, but qsort moves
+ * those structs around as it works, so a comparator left to fill the cache
+ * loses most of what it stored and stats the same repository over and over.
+ */
+static void resolve_modtimes(void)
+{
+	time_t t;
+	int i;
+
+	for (i = 0; i < cgit_repolist.count; i++)
+		get_repo_modtime(&cgit_repolist.repos[i], &t);
+}
 
-static const struct sortcolumn sortcolumn[] = {
-	{"section", sort_section},
-	{"name", sort_name},
-	{"desc", sort_desc},
-	{"owner", sort_owner},
-	{"idle", sort_idle},
+static const struct sort_column sort_columns[] = {
+	{"section", cmp_section},
+	{"name", cmp_name},
+	{"desc", cmp_desc},
+	{"owner", cmp_owner},
+	{"idle", cmp_idle},
 	{NULL, NULL}
 };
 
 static int sort_repolist(char *field)
 {
-	const struct sortcolumn *column;
+	const struct sort_column *column;
 
-	for (column = &sortcolumn[0]; column->name; column++) {
+	for (column = &sort_columns[0]; column->name; column++) {
 		if (strcmp(field, column->name))
 			continue;
-		if (column->fn == sort_idle || column->fn == sort_section)
+		if (column->cmp == cmp_idle || column->cmp == cmp_section)
 			resolve_modtimes();
 		qsort(cgit_repolist.repos, cgit_repolist.count,
-			sizeof(struct cgit_repo), column->fn);
+			sizeof(struct cgit_repo), column->cmp);
 		return 1;
 	}
 	return 0;
 }
 
-
 void cgit_print_repolist(void)
 {
-	int i, columns = 3, hits = 0, header = 0;
 	char *last_section = NULL;
-	char *section;
-	char *repourl;
 	char *currenturl;
-	int sorted = 0;
+	int columns = BASE_COLUMNS;
+	int column_sorted = 0;
+	int hits = 0;
+	int shown = 0;
+	int i;
 
 	if (!any_repos_visible()) {
 		cgit_print_error_page(404, "Not found", "No repositories found");
@@ -337,71 +415,38 @@ void cgit_print_repolist(void)
 	cgit_print_pageheader();
 
 	if (ctx.qry.sort)
-		sorted = sort_repolist(ctx.qry.sort);
+		column_sorted = sort_repolist(ctx.qry.sort);
 	else if (ctx.cfg.section_sort)
 		sort_repolist("section");
 
 	currenturl = cgit_currenturl();
 	html("<table class='list nowrap repolist'>");
 	for (i = 0; i < cgit_repolist.count; i++) {
+		char *section;
+
 		ctx.repo = &cgit_repolist.repos[i];
 		if (!is_visible(ctx.repo))
 			continue;
 		hits++;
+		// Rows outside the page are stepped over rather than broken
+		// out of, because the pager below is sized from the total.
 		if (hits <= ctx.qry.ofs)
 			continue;
-		if (hits > ctx.qry.ofs + ctx.cfg.max_repo_count)
+		// Written as a subtraction because max-repo-count of zero
+		// means unlimited and is held as INT_MAX, which any positive
+		// offset would overflow if it were added to instead.
+		if (hits - ctx.qry.ofs > ctx.cfg.max_repo_count)
 			continue;
-		if (!header++)
-			print_header(currenturl);
+		if (!shown++)
+			print_header_row(currenturl);
 		section = ctx.repo->section;
 		if (section && !strcmp(section, ""))
 			section = NULL;
-		if (!sorted &&
-		    ((last_section == NULL && section != NULL) ||
-		    (last_section != NULL && section == NULL) ||
-		    (last_section != NULL && section != NULL &&
-		     strcmp(section, last_section)))) {
-			htmlf("<tr class='nohover-highlight'><td colspan='%d' class='reposection'>",
-			      columns);
-			html_txt(section);
-			html("</td></tr>");
+		if (!column_sorted && section_changed(section, last_section)) {
+			print_section_row(section, columns);
 			last_section = section;
 		}
-		htmlf("<tr><td class='col-name %s'>",
-		      !sorted && section ? "sublevel-repo" : "toplevel-repo");
-		cgit_summary_link(ctx.repo->name, NULL, NULL, NULL);
-		html("</td><td class='col-desc'>");
-		repourl = cgit_repourl(ctx.repo->url);
-		html_link_open(repourl, NULL, NULL);
-		free(repourl);
-		if (html_ntxt(ctx.repo->desc, ctx.cfg.max_repodesc_len) < 0)
-			html("...");
-		html_link_close();
-		html("</td>");
-		if (ctx.cfg.enable_index_owner) {
-			html("<td class='col-owner'>");
-			html("<a href='");
-			html_attr(currenturl);
-			html("?q=");
-			html_url_arg(ctx.repo->owner);
-			html("'>");
-			html_txt(ctx.repo->owner);
-			html("</a>");
-			html("</td>");
-		}
-		html("<td class='col-idle'>");
-		print_modtime(ctx.repo);
-		html("</td>");
-		if (ctx.cfg.enable_index_links) {
-			html("<td class='col-links'>");
-			cgit_summary_link("summary", NULL, "button", NULL);
-			cgit_log_link("log", NULL, "button", NULL, NULL, NULL,
-				      0, NULL, NULL, ctx.qry.showmsg, 0);
-			cgit_tree_link("tree", NULL, "button", NULL, NULL, NULL);
-			html("</td>");
-		}
-		html("</tr>\n");
+		print_repo_row(currenturl, !column_sorted && section);
 	}
 	html("</table>");
 	if (hits > ctx.cfg.max_repo_count)
diff --git a/source/ui-repolist.h b/source/ui-repolist.h
index 1b6b322..294e13f 100644
--- a/source/ui-repolist.h
+++ b/source/ui-repolist.h
@@ -1,7 +1,14 @@
-#ifndef UI_REPOLIST_H
-#define UI_REPOLIST_H
+/*
+ * The two pages cgit renders without a repository in hand. One is the index
+ * of every repository the configuration knows about, which is where a request
+ * naming neither a repository nor a page lands. The other is the site wide
+ * readme, which the about page shows when no repository was selected.
+ */
+
+#ifndef CGIT_UI_REPOLIST_H
+#define CGIT_UI_REPOLIST_H
 
 extern void cgit_print_repolist(void);
 extern void cgit_print_site_readme(void);
 
-#endif /* UI_REPOLIST_H */
+#endif // CGIT_UI_REPOLIST_H
diff --git a/source/ui-shared.c b/source/ui-shared.c
index 23da286..938f292 100644
--- a/source/ui-shared.c
+++ b/source/ui-shared.c
This diff is too large to be rendered inline. View it on its own page.
diff --git a/source/ui-shared.h b/source/ui-shared.h
index b6f2797..c266951 100644
--- a/source/ui-shared.h
+++ b/source/ui-shared.h
@@ -1,9 +1,25 @@
-#ifndef UI_SHARED_H
-#define UI_SHARED_H
+/*
+ * The page furniture and the link builders every renderer shares. A page
+ * handler renders its own body and calls in here for all that surrounds it,
+ * which is the HTTP headers, the document head and footer, the header block
+ * with its tabs and search form, the error page, and an anchor to any other
+ * cgit page. Whether a site addresses a request as a path below a virtual root
+ * or as a query string is settled behind these calls, so that nothing else has
+ * to know which of the two it is looking at.
+ */
 
+#ifndef CGIT_UI_SHARED_H
+#define CGIT_UI_SHARED_H
+
+#include "cgit.h"
 extern const char *cgit_httpscheme(void);
 extern char *cgit_hosturl(void);
-extern const char *cgit_rooturl(void);
+
+/*
+ * The URL of the request being answered, which cgit_currentfullurl gives with
+ * the query string on the end as well, minus the url argument that the path
+ * was taken from.
+ */
 extern char *cgit_currenturl(void);
 extern char *cgit_currentfullurl(void);
 extern const char *cgit_loginurl(void);
@@ -13,10 +29,17 @@ extern char *cgit_fileurl(const char *reponame, const char *pagename,
 extern char *cgit_pageurl(const char *reponame, const char *pagename,
 			  const char *query);
 
+// Call fn once for every URL this repository can be cloned from.
 extern void cgit_add_clone_urls(void (*fn)(const char *));
 
+/*
+ * Each of these writes one anchor to the page and nothing around it, with name
+ * as the text a reader sees. A NULL title or class leaves that attribute out,
+ * and what follows them is the state the target page should open in.
+ */
 extern void cgit_index_link(const char *name, const char *title,
-			    const char *class, const char *pattern, const char *sort, int ofs, int always_root);
+			    const char *class, const char *pattern,
+			    const char *sort, int ofs, int always_root);
 extern void cgit_summary_link(const char *name, const char *title,
 			      const char *class, const char *head);
 extern void cgit_tag_link(const char *name, const char *title,
@@ -43,9 +66,6 @@ extern void cgit_patch_link(const char *name, const char *title,
 extern void cgit_refs_link(const char *name, const char *title,
 			   const char *class, const char *head,
 			   const char *rev, const char *path);
-extern void cgit_snapshot_link(const char *name, const char *title,
-			       const char *class, const char *head,
-			       const char *rev, const char *archivename);
 extern void cgit_diff_link(const char *name, const char *title,
 			   const char *class, const char *head,
 			   const char *new_rev, const char *old_rev,
@@ -63,26 +83,42 @@ extern void cgit_print_layout_end(void);
 
 __attribute__((format (printf,1,2)))
 extern void cgit_print_error(const char *fmt, ...);
-__attribute__((format (printf,1,0)))
-extern void cgit_vprint_error(const char *fmt, va_list ap);
 extern struct date_mode cgit_date_mode(enum date_mode_type type);
+
+/*
+ * Print the time t, which was recorded in the timezone tz, as an age. An age
+ * further back than max_relative is given as a calendar date instead, and a
+ * negative max_relative stays relative however old it is. A site that turns
+ * relative dates off gets the date either way.
+ */
 extern void cgit_print_age(time_t t, int tz, time_t max_relative);
 extern void cgit_print_http_headers(void);
 extern void cgit_redirect(const char *url, bool permanent);
 extern void cgit_print_docstart(void);
 extern void cgit_print_docend(void);
 __attribute__((format (printf,3,4)))
-extern void cgit_print_error_page(int code, const char *msg, const char *fmt, ...);
-extern void cgit_vprint_error_page(int code, const char *msg, const char *fmt, va_list ap);
+extern void cgit_print_error_page(int code, const char *msg, const char *fmt,
+				  ...);
+extern void cgit_vprint_error_page(int code, const char *msg, const char *fmt,
+				   va_list ap);
 extern void cgit_print_pageheader(void);
 extern void cgit_print_filemode(unsigned short mode);
-extern void cgit_compose_snapshot_prefix(struct strbuf *filename,
-					 const char *base, const char *ref);
 extern void cgit_print_snapshot_links(const struct cgit_repo *repo,
 				      const char *ref, const char *separator);
+
+// The name a snapshot of this repository is downloaded under, which is the
+// configured prefix or the last component of the repository URL.
 extern const char *cgit_snapshot_prefix(const struct cgit_repo *repo);
+
+/*
+ * Carry the current request into a form as hidden inputs, so that submitting
+ * it changes only what the form itself asks about. The branch and the search
+ * terms are carried only when the caller asks for them, since a form with a
+ * field of its own for either would otherwise submit that name twice.
+ */
 extern void cgit_add_hidden_formfields(int incl_head, int incl_search,
 				       const char *page);
 
+// Put path in front of the page title, its last component first.
 extern void cgit_set_title_from_path(const char *path);
-#endif /* UI_SHARED_H */
+#endif // CGIT_UI_SHARED_H
diff --git a/source/ui-snapshot.c b/source/ui-snapshot.c
index 97472c5..0bab8ea 100644
--- a/source/ui-snapshot.c
+++ b/source/ui-snapshot.c
@@ -1,27 +1,37 @@
-/* ui-snapshot.c: generate snapshot of a commit
- *
- * Copyright (C) 2006-2014 cgit Development Team <cgit@lists.zx2c4.com>
- *
- * Licensed under GNU General Public License v2
- *   (see LICENSE.txt for full license text)
+/*
+ * Serves a commit as a downloadable archive. Writing the archive is git's
+ * work, so each format here is a thin wrapper that either calls write_archive
+ * directly or pipes its tar output through an external compressor. The table
+ * of formats also lives here, and since a repository's snapshots mask is one
+ * bit per table position, that order is part of what cgitrc means. When the
+ * request carries no revision the file name is worked backwards to find one,
+ * so that a link to cgit-1.2.tar.gz can serve the tag it was named after. A
+ * name ending in .asc asks for the detached signature kept under refs/notes
+ * instead of the archive itself.
  */
 
 #define USE_THE_REPOSITORY_VARIABLE
 
 #include "cgit.h"
-#include "ui-snapshot.h"
+#include "filter.h"
 #include "html.h"
 #include "ui-shared.h"
+#include "ui-snapshot.h"
+
+#define SIG_SUFFIX ".asc"
 
-static int write_archive_type(const char *format, const char *hex, const char *prefix)
+static int write_archive_format(const char *format_arg, const char *hex,
+				const char *prefix)
 {
 	struct strvec argv = STRVEC_INIT;
-	const char **nargv;
+	const char **args;
 	int result;
+
 	strvec_push(&argv, "snapshot");
-	strvec_push(&argv, format);
+	strvec_push(&argv, format_arg);
 	if (prefix) {
 		struct strbuf buf = STRBUF_INIT;
+
 		strbuf_addstr(&buf, prefix);
 		strbuf_addch(&buf, '/');
 		strvec_push(&argv, "--prefix");
@@ -29,44 +39,40 @@ static int write_archive_type(const char *format, const char *hex, const char *p
 		strbuf_release(&buf);
 	}
 	strvec_push(&argv, hex);
-	/*
-	 * Now we need to copy the pointers to arguments into a new
-	 * structure because write_archive will rearrange its arguments
-	 * which may result in duplicated/missing entries causing leaks
-	 * or double-frees in strvec_clear.
-	 */
-	nargv = xmalloc(sizeof(char *) * (argv.nr + 1));
-	/* strvec guarantees a trailing NULL entry. */
-	memcpy(nargv, argv.v, sizeof(char *) * (argv.nr + 1));
-
-	result = write_archive(argv.nr, nargv, NULL, the_repository, NULL, 0);
+
+	// write_archive rearranges the argv it is handed, which would leave
+	// strvec_clear leaking or double freeing, so it gets a copy of the
+	// pointers, one entry longer than argv.nr for the trailing NULL.
+	args = xmalloc(sizeof(char *) * (argv.nr + 1));
+	memcpy(args, argv.v, sizeof(char *) * (argv.nr + 1));
+
+	result = write_archive(argv.nr, args, NULL, the_repository, NULL, 0);
 	strvec_clear(&argv);
-	free(nargv);
+	free(args);
 	return result;
 }
 
 static int write_tar_archive(const char *hex, const char *prefix)
 {
-	return write_archive_type("--format=tar", hex, prefix);
+	return write_archive_format("--format=tar", hex, prefix);
 }
 
 static int write_zip_archive(const char *hex, const char *prefix)
 {
-	return write_archive_type("--format=zip", hex, prefix);
+	return write_archive_format("--format=zip", hex, prefix);
 }
 
-static int write_compressed_tar_archive(const char *hex,
-					const char *prefix,
-					char *filter_argv[])
+static int write_compressed_tar_archive(const char *hex, const char *prefix,
+					char *argv[])
 {
-	int rv;
-	struct cgit_exec_filter f;
-	cgit_exec_filter_init(&f, filter_argv[0], filter_argv);
-
-	cgit_open_filter(&f.base);
-	rv = write_tar_archive(hex, prefix);
-	cgit_close_filter(&f.base);
-	return rv;
+	struct cgit_exec_filter filter;
+	int result;
+
+	cgit_exec_filter_init(&filter, argv[0], argv);
+	cgit_open_filter(&filter.base);
+	result = write_tar_archive(hex, prefix);
+	cgit_close_filter(&filter.base);
+	return result;
 }
 
 static int write_tar_gzip_archive(const char *hex, const char *prefix)
@@ -95,14 +101,15 @@ static int write_tar_xz_archive(const char *hex, const char *prefix)
 
 static int write_tar_zstd_archive(const char *hex, const char *prefix)
 {
-	// Stay single-threaded like the other compressors so one request
-	// cannot pin every core. -T0 would fan out across all of them.
+	// Single-threaded like the other compressors, since -T0 would let one
+	// request pin every core.
 	char *argv[] = { "zstd", NULL };
 	return write_compressed_tar_archive(hex, prefix, argv);
 }
 
+// ui-shared.c reads entry zero as the tar whose signature stands in for every
+// tar variant, so this order cannot change.
 const struct cgit_snapshot_format cgit_snapshot_formats[] = {
-	/* .tar must remain the 0 index */
 	{ ".tar",	"application/x-tar",	write_tar_archive	},
 	{ ".tar.gz",	"application/x-gzip",	write_tar_gzip_archive	},
 	{ ".tar.bz2",	"application/x-bzip2",	write_tar_bzip2_archive	},
@@ -113,48 +120,77 @@ const struct cgit_snapshot_format cgit_snapshot_formats[] = {
 	{ NULL }
 };
 
-static struct notes_tree snapshot_sig_notes[ARRAY_SIZE(cgit_snapshot_formats)];
+// Each tree is read the first time a signature for that format is asked for,
+// then kept for the life of the process.
+static struct notes_tree sig_notes[ARRAY_SIZE(cgit_snapshot_formats)];
 
-const struct object_id *cgit_snapshot_get_sig(const char *ref,
-					      const struct cgit_snapshot_format *f)
+static size_t format_index(const struct cgit_snapshot_format *f)
 {
-	struct notes_tree *tree;
-	struct object_id oid;
-
-	if (repo_get_oid(the_repository, ref, &oid))
-		return NULL;
-
-	tree = &snapshot_sig_notes[f - &cgit_snapshot_formats[0]];
-	if (!tree->initialized) {
-		struct strbuf notes_ref = STRBUF_INIT;
+	return f - cgit_snapshot_formats;
+}
 
-		strbuf_addf(&notes_ref, "refs/notes/signatures/%s",
-			    f->suffix + 1);
+static const struct cgit_snapshot_format *find_format(const char *filename)
+{
+	const struct cgit_snapshot_format *f;
 
-		init_notes(tree, notes_ref.buf, combine_notes_ignore, 0);
-		strbuf_release(&notes_ref);
+	for (f = cgit_snapshot_formats; f->suffix; f++) {
+		if (ends_with(filename, f->suffix))
+			return f;
 	}
-
-	return get_note(tree, &oid);
+	return NULL;
 }
 
-static const struct cgit_snapshot_format *get_format(const char *filename)
+static int resolves(const char *rev)
 {
-	const struct cgit_snapshot_format *fmt;
+	struct object_id oid;
 
-	for (fmt = cgit_snapshot_formats; fmt->suffix; fmt++) {
-		if (ends_with(filename, fmt->suffix))
-			return fmt;
-	}
-	return NULL;
+	return repo_get_oid(the_repository, rev, &oid) == 0;
 }
 
-unsigned cgit_snapshot_format_bit(const struct cgit_snapshot_format *f)
+static const char *ref_from_filename(const struct cgit_repo *repo,
+				     const char *filename,
+				     const struct cgit_snapshot_format *format)
 {
-	return BIT(f - &cgit_snapshot_formats[0]);
+	struct strbuf rev = STRBUF_INIT;
+	const char *repo_prefix;
+	int found = 1;
+
+	strbuf_addstr(&rev, filename);
+	strbuf_setlen(&rev, rev.len - strlen(format->suffix));
+
+	if (resolves(rev.buf))
+		goto out;
+
+	repo_prefix = cgit_snapshot_prefix(repo);
+	if (starts_with(rev.buf, repo_prefix)) {
+		const char *rest = rev.buf + strlen(repo_prefix);
+
+		while (rest && (*rest == '-' || *rest == '_'))
+			rest++;
+		strbuf_splice(&rev, 0, rest - rev.buf, "", 0);
+	}
+
+	if (resolves(rev.buf))
+		goto out;
+
+	// A tag is often written with a leading v while the file named after it
+	// is not, so that is tried last.
+	strbuf_insert(&rev, 0, "v", 1);
+	if (resolves(rev.buf))
+		goto out;
+
+	strbuf_splice(&rev, 0, 1, "V", 1);
+	if (resolves(rev.buf))
+		goto out;
+
+	found = 0;
+	strbuf_release(&rev);
+
+out:
+	return found ? strbuf_detach(&rev, NULL) : NULL;
 }
 
-static int make_snapshot(const struct cgit_snapshot_format *format,
+static int send_snapshot(const struct cgit_snapshot_format *format,
 			 const char *hex, const char *prefix,
 			 const char *filename)
 {
@@ -179,9 +215,9 @@ static int make_snapshot(const struct cgit_snapshot_format *format,
 	return 0;
 }
 
-static int write_sig(const struct cgit_snapshot_format *format,
-		     const char *hex, const char *archive,
-		     const char *filename)
+static int send_sig(const struct cgit_snapshot_format *format,
+		    const char *hex, const char *archive_name,
+		    const char *sig_filename)
 {
 	const struct object_id *note = cgit_snapshot_get_sig(hex, format);
 	enum object_type type;
@@ -190,7 +226,7 @@ static int write_sig(const struct cgit_snapshot_format *format,
 
 	if (!note) {
 		cgit_print_error_page(404, "Not found",
-				"No signature for %s", archive);
+				"No signature for %s", archive_name);
 		return 0;
 	}
 
@@ -200,11 +236,14 @@ static int write_sig(const struct cgit_snapshot_format *format,
 		return 0;
 	}
 
+	// The body is whatever bytes the note holds, so these go out ahead of
+	// the usual headers to stop a browser sniffing it into a type it will
+	// act on.
 	html("X-Content-Type-Options: nosniff\n");
 	html("Content-Security-Policy: default-src 'none'\n");
 	ctx.page.etag = oid_to_hex(note);
 	ctx.page.mimetype = xstrdup("application/pgp-signature");
-	ctx.page.filename = xstrdup(filename);
+	ctx.page.filename = xstrdup(sig_filename);
 	cgit_print_http_headers();
 
 	html_raw(buf, size);
@@ -212,64 +251,74 @@ static int write_sig(const struct cgit_snapshot_format *format,
 	return 0;
 }
 
-/* Try to guess the requested revision from the requested snapshot name.
- * First the format extension is stripped, e.g. "cgit-0.7.2.tar.gz" become
- * "cgit-0.7.2". If this is a valid commit object name we've got a winner.
- * Otherwise, if the snapshot name has a prefix matching the result from
- * repo_basename(), we strip the basename and any following '-' and '_'
- * characters ("cgit-0.7.2" -> "0.7.2") and check the resulting name once
- * more. If this still isn't a valid commit object name, we check if pre-
- * pending a 'v' or a 'V' to the remaining snapshot name ("0.7.2" ->
- * "v0.7.2") gives us something valid.
- */
-static const char *get_ref_from_filename(const struct cgit_repo *repo,
-					 const char *filename,
-					 const struct cgit_snapshot_format *format)
+const struct object_id *cgit_snapshot_get_sig(const char *ref,
+					      const struct cgit_snapshot_format *f)
 {
-	const char *reponame;
+	struct notes_tree *tree;
 	struct object_id oid;
-	struct strbuf snapshot = STRBUF_INIT;
-	int result = 1;
 
-	strbuf_addstr(&snapshot, filename);
-	strbuf_setlen(&snapshot, snapshot.len - strlen(format->suffix));
+	if (repo_get_oid(the_repository, ref, &oid))
+		return NULL;
 
-	if (repo_get_oid(the_repository, snapshot.buf, &oid) == 0)
-		goto out;
+	tree = &sig_notes[format_index(f)];
+	if (!tree->initialized) {
+		struct strbuf notes_ref = STRBUF_INIT;
 
-	reponame = cgit_snapshot_prefix(repo);
-	if (starts_with(snapshot.buf, reponame)) {
-		const char *new_start = snapshot.buf;
-		new_start += strlen(reponame);
-		while (new_start && (*new_start == '-' || *new_start == '_'))
-			new_start++;
-		strbuf_splice(&snapshot, 0, new_start - snapshot.buf, "", 0);
+		// Signatures live under the format suffix with the leading dot
+		// dropped, so plain tar is refs/notes/signatures/tar.
+		strbuf_addf(&notes_ref, "refs/notes/signatures/%s",
+			    f->suffix + 1);
+
+		init_notes(tree, notes_ref.buf, combine_notes_ignore, 0);
+		strbuf_release(&notes_ref);
 	}
 
-	if (repo_get_oid(the_repository, snapshot.buf, &oid) == 0)
-		goto out;
+	return get_note(tree, &oid);
+}
 
-	strbuf_insert(&snapshot, 0, "v", 1);
-	if (repo_get_oid(the_repository, snapshot.buf, &oid) == 0)
-		goto out;
+unsigned cgit_snapshot_format_bit(const struct cgit_snapshot_format *f)
+{
+	return 1U << format_index(f);
+}
 
-	strbuf_splice(&snapshot, 0, 1, "V", 1);
-	if (repo_get_oid(the_repository, snapshot.buf, &oid) == 0)
-		goto out;
+int cgit_parse_snapshots_mask(const char *str)
+{
+	struct string_list tokens = STRING_LIST_INIT_DUP;
+	struct string_list_item *item;
+	const struct cgit_snapshot_format *f;
+	int mask = 0;
+
+	// A number is the legacy form of this setting and is still taken
+	// ahead of the named forms below.
+	if (atoi(str))
+		return 1;
 
-	result = 0;
-	strbuf_release(&snapshot);
+	if (strcmp(str, "all") == 0)
+		return INT_MAX;
 
-out:
-	return result ? strbuf_detach(&snapshot, NULL) : NULL;
+	string_list_split(&tokens, str, " ", -1);
+	string_list_remove_empty_items(&tokens, 0);
+
+	for_each_string_list_item(item, &tokens) {
+		for (f = cgit_snapshot_formats; f->suffix; f++) {
+			if (!strcmp(item->string, f->suffix) ||
+			    !strcmp(item->string, f->suffix + 1)) {
+				mask |= cgit_snapshot_format_bit(f);
+				break;
+			}
+		}
+	}
+
+	string_list_clear(&tokens, 0);
+	return mask;
 }
 
 void cgit_print_snapshot(const char *head, const char *hex,
 			 const char *filename, int dwim)
 {
-	const struct cgit_snapshot_format* f;
+	const struct cgit_snapshot_format *f;
 	const char *sig_filename = NULL;
-	char *adj_filename = NULL;
+	char *archive_name = NULL;
 	char *prefix = NULL;
 
 	if (!filename) {
@@ -278,25 +327,24 @@ void cgit_print_snapshot(const char *head, const char *hex,
 		return;
 	}
 
-	if (ends_with(filename, ".asc")) {
+	if (ends_with(filename, SIG_SUFFIX)) {
 		sig_filename = filename;
-
-		/* Strip ".asc" from filename for common format processing */
-		adj_filename = xstrdup(filename);
-		adj_filename[strlen(adj_filename) - 4] = '\0';
-		filename = adj_filename;
+		archive_name = xstrdup(filename);
+		archive_name[strlen(archive_name) - strlen(SIG_SUFFIX)] = '\0';
+		filename = archive_name;
 	}
 
-	f = get_format(filename);
-	if (!f || (!sig_filename && !(ctx.repo->snapshots & cgit_snapshot_format_bit(f)))) {
+	f = find_format(filename);
+	if (!f || (!sig_filename &&
+		   !(ctx.repo->snapshots & cgit_snapshot_format_bit(f)))) {
 		cgit_print_error_page(400, "Bad request",
 				"Unsupported snapshot format: %s", filename);
 		return;
 	}
 
 	if (!hex && dwim) {
-		hex = get_ref_from_filename(ctx.repo, filename, f);
-		if (hex == NULL) {
+		hex = ref_from_filename(ctx.repo, filename, f);
+		if (!hex) {
 			cgit_print_error_page(404, "Not found", "Not found");
 			return;
 		}
@@ -311,10 +359,10 @@ void cgit_print_snapshot(const char *head, const char *hex,
 		prefix = xstrdup(cgit_snapshot_prefix(ctx.repo));
 
 	if (sig_filename)
-		write_sig(f, hex, filename, sig_filename);
+		send_sig(f, hex, filename, sig_filename);
 	else
-		make_snapshot(f, hex, prefix, filename);
+		send_snapshot(f, hex, prefix, filename);
 
 	free(prefix);
-	free(adj_filename);
+	free(archive_name);
 }
diff --git a/source/ui-snapshot.h b/source/ui-snapshot.h
index a8deec3..9f1ec02 100644
--- a/source/ui-snapshot.h
+++ b/source/ui-snapshot.h
@@ -1,7 +1,41 @@
-#ifndef UI_SNAPSHOT_H
-#define UI_SNAPSHOT_H
+/*
+ * The snapshot page, which serves a commit as a downloadable archive. It owns
+ * the table of formats cgit knows how to write, so the config reader and the
+ * link builders both reach through here to ask which formats a repository
+ * offers. Each format has a bit in the per-repository snapshots mask, taken
+ * from its position in that table.
+ */
+
+#ifndef CGIT_UI_SNAPSHOT_H
+#define CGIT_UI_SNAPSHOT_H
+
+#include "cgit.h"
+
+typedef int (*write_archive_fn_t)(const char *hex, const char *prefix);
+
+struct cgit_snapshot_format {
+	const char *suffix;
+	const char *mimetype;
+	write_archive_fn_t write_func;
+};
+
+// Terminated by an entry with a NULL suffix.
+extern const struct cgit_snapshot_format cgit_snapshot_formats[];
+
+extern unsigned cgit_snapshot_format_bit(const struct cgit_snapshot_format *f);
+
+/*
+ * Turn a cgitrc snapshots value into a mask over cgit_snapshot_formats. A
+ * plain number is the legacy form meaning every format, as is the word all,
+ * otherwise the value is a space separated list of suffixes.
+ */
+extern int cgit_parse_snapshots_mask(const char *str);
+
+// The detached signature stored for this format under refs/notes, if any.
+extern const struct object_id *cgit_snapshot_get_sig(
+	const char *ref, const struct cgit_snapshot_format *f);
 
 extern void cgit_print_snapshot(const char *head, const char *hex,
 				const char *filename, int dwim);
 
-#endif /* UI_SNAPSHOT_H */
+#endif // CGIT_UI_SNAPSHOT_H
diff --git a/source/ui-ssdiff.c b/source/ui-ssdiff.c
index e4472c4..2ea7021 100644
--- a/source/ui-ssdiff.c
+++ b/source/ui-ssdiff.c
This diff is too large to be rendered inline. View it on its own page.
diff --git a/source/ui-ssdiff.h b/source/ui-ssdiff.h
index 11f2714..f6489c8 100644
--- a/source/ui-ssdiff.h
+++ b/source/ui-ssdiff.h
@@ -1,9 +1,16 @@
-#ifndef UI_SSDIFF_H
-#define UI_SSDIFF_H
-
 /*
- * ssdiff line limits
+ * The side by side diff view, which ui-diff.c renders through when a reader
+ * asks for that layout rather than the unified one. Every call below is made
+ * as ui-diff.c walks a diff, so nothing here is a page of its own.
  */
+
+#ifndef CGIT_UI_SSDIFF_H
+#define CGIT_UI_SSDIFF_H
+
+// The longest old line and the longest new line the view will compare
+// character by character. The comparison fills a cell for every pair of
+// characters, so its cost is the product of the two, and a line past the limit
+// is shown whole instead.
 #ifndef MAX_SSDIFF_M
 #define MAX_SSDIFF_M 128
 #endif
@@ -13,8 +20,6 @@
 #endif
 #define MAX_SSDIFF_SIZE ((MAX_SSDIFF_M) * (MAX_SSDIFF_N))
 
-extern void cgit_ssdiff_print_deferred_lines(void);
-
 extern void cgit_ssdiff_line_cb(char *line, int len);
 
 extern void cgit_ssdiff_header_begin(void);
@@ -22,4 +27,4 @@ extern void cgit_ssdiff_header_end(void);
 
 extern void cgit_ssdiff_footer(void);
 
-#endif /* UI_SSDIFF_H */
+#endif // CGIT_UI_SSDIFF_H
diff --git a/source/ui-stats.c b/source/ui-stats.c
index 20e9a30..5436907 100644
--- a/source/ui-stats.c
+++ b/source/ui-stats.c
This diff is too large to be rendered inline. View it on its own page.
diff --git a/source/ui-stats.h b/source/ui-stats.h
index 0e61b03..718e335 100644
--- a/source/ui-stats.h
+++ b/source/ui-stats.h
@@ -1,5 +1,13 @@
-#ifndef UI_STATS_H
-#define UI_STATS_H
+/*
+ * The statistics page, which counts commits per author over a window of
+ * recent weeks, months, quarters or years and sizes the tip of the branch by
+ * language. The windows are described here rather than inside the page,
+ * because cgit.c resolves a repository's max-stats setting against the same
+ * set.
+ */
+
+#ifndef CGIT_UI_STATS_H
+#define CGIT_UI_STATS_H
 
 #include "cgit.h"
 
@@ -7,22 +15,36 @@ struct cgit_period {
 	const char code;
 	const char *name;
 	int max_periods;
+
+	// How many periods a page shows side by side.
 	int count;
 
-	/* Convert a tm value to the first day in the period */
+	// Move a tm back to the first day of the period it falls in.
 	void (*trunc)(struct tm *tm);
 
-	/* Update tm value to start of next/previous period */
+	// Step a truncated tm one whole period back or forward.
 	void (*dec)(struct tm *tm);
 	void (*inc)(struct tm *tm);
 
-	/* Pretty-print a tm value */
+	// Label the period a tm falls in. The buffer is reused, so a caller
+	// that keeps a label copies it.
 	char *(*pretty)(struct tm *tm);
 };
 
-extern int cgit_find_stats_period(const char *expr, const struct cgit_period **period);
+/*
+ * Look a period up by its one letter code or by its full name, storing the
+ * match through period when that is not NULL. The return is a one based index
+ * into the table of periods, or 0 when nothing matches, and because that
+ * table runs from the finest window to the coarsest, the index is what
+ * max-stats is compared against.
+ */
+extern int cgit_find_stats_period(const char *expr,
+				  const struct cgit_period **period);
+
+// The name of the period at a one based index, or an empty string when the
+// index names no period.
 extern const char *cgit_find_stats_periodname(int idx);
 
 extern void cgit_show_stats(void);
 
-#endif /* UI_STATS_H */
+#endif // CGIT_UI_STATS_H
diff --git a/source/ui-summary.c b/source/ui-summary.c
index 2169048..ada6b1b 100644
--- a/source/ui-summary.c
+++ b/source/ui-summary.c
@@ -1,32 +1,49 @@
-/* ui-summary.c: functions for generating repo summary page
- *
- * Copyright (C) 2006-2014 cgit Development Team <cgit@lists.zx2c4.com>
- *
- * Licensed under GNU General Public License v2
- *   (see LICENSE.txt for full license text)
+/*
+ * The repository summary page and the about page. The summary stacks the
+ * branch list, the tag list and the head of the log into one table and ends
+ * with the clone urls, reusing the listings ui-refs.c and ui-log.c draw
+ * elsewhere. The about page renders the readme that config parsing picked for
+ * the repository, either through the configured about filter or escaped as
+ * plain text, and it can serve a file sitting beside that readme so links
+ * inside the readme resolve.
  */
 
 #include "cgit.h"
-#include "ui-summary.h"
+#include "filter.h"
 #include "html.h"
+#include "shared.h"
 #include "ui-blob.h"
 #include "ui-log.h"
 #include "ui-plain.h"
 #include "ui-refs.h"
 #include "ui-shared.h"
+#include "ui-summary.h"
+
+// Age, commit message and author, the three columns a log row always has.
+// ui-log.c adds one more for each of the two optional counts, and the rows
+// this page stretches across the table have to match that width.
+#define LOG_BASE_COLUMNS 3
 
-static int urls;
+static int clone_urls_printed;
 
-static void print_url(const char *url)
+static int log_columns(void)
 {
-	int columns = 3;
+	int columns = LOG_BASE_COLUMNS;
 
 	if (ctx.repo->enable_log_filecount)
 		columns++;
 	if (ctx.repo->enable_log_linecount)
 		columns++;
+	return columns;
+}
+
+static void print_clone_url(const char *url)
+{
+	int columns = log_columns();
 
-	if (urls++ == 0) {
+	// cgit_add_clone_urls may call back no times at all, so the heading
+	// waits for a first url rather than being printed ahead of the walk.
+	if (clone_urls_printed++ == 0) {
 		htmlf("<tr class='nohover'><td colspan='%d'>&nbsp;</td></tr>", columns);
 		htmlf("<tr class='nohover'><th class='left' colspan='%d'>Clone</th></tr>\n", columns);
 	}
@@ -40,43 +57,38 @@ static void print_url(const char *url)
 	html("</a></td></tr>\n");
 }
 
-void cgit_print_summary(void)
+/*
+ * Without the separator boundary a sibling directory that merely shares the
+ * base as a name prefix, such as repo.git-backup beside repo.git, would pass.
+ */
+static int path_within(const char *base, const char *path)
 {
-	int columns = 3;
-
-	if (ctx.repo->enable_log_filecount)
-		columns++;
-	if (ctx.repo->enable_log_linecount)
-		columns++;
+	size_t len = strlen(base);
 
-	cgit_print_layout_start();
-	html("<table summary='repository info' class='list nowrap'>");
-	cgit_print_branches(ctx.cfg.summary_branches);
-	htmlf("<tr class='nohover'><td colspan='%d'>&nbsp;</td></tr>", columns);
-	cgit_print_tags(ctx.cfg.summary_tags);
-	if (ctx.cfg.summary_log > 0) {
-		htmlf("<tr class='nohover'><td colspan='%d'>&nbsp;</td></tr>", columns);
-		cgit_print_log(ctx.qry.head, 0, ctx.cfg.summary_log, NULL,
-			       NULL, NULL, 0, 0, 0);
-	}
-	urls = 0;
-	cgit_add_clone_urls(print_url);
-	html("</table>");
-	cgit_print_layout_end();
+	return starts_with(path, base) &&
+	       (path[len] == '\0' || path[len] == '/');
 }
 
-/* The caller must free the return value. */
-static char* append_readme_path(const char *filename, const char *ref, const char *path)
+/*
+ * Returns a path the caller must free, or NULL when the request cannot be
+ * served. A null ref means the readme is a file on the server's disk rather
+ * than a path inside a ref, and such a readme is confined to its own
+ * directory, so one named without a directory has nothing to confine it to
+ * and is refused.
+ */
+static char *resolve_about_path(const char *filename, const char *ref,
+				const char *path)
 {
-	char *file, *base_dir, *full_path, *resolved_base = NULL, *resolved_full = NULL;
-	/* If a subpath is specified for the about page, make it relative
-	 * to the directory containing the configured readme. */
+	char *copy, *base_dir, *full_path;
+	char *resolved_base = NULL, *resolved_full = NULL;
 
-	file = xstrdup(filename);
-	base_dir = dirname(file);
+	// dirname is allowed to write into its argument and to return a
+	// pointer into it, so base_dir borrows from a copy we keep alive.
+	copy = xstrdup(filename);
+	base_dir = dirname(copy);
 	if (!strcmp(base_dir, ".") || !strcmp(base_dir, "..")) {
 		if (!ref) {
-			free(file);
+			free(copy);
 			return NULL;
 		}
 		full_path = xstrdup(path);
@@ -86,32 +98,48 @@ static char* append_readme_path(const char *filename, const char *ref, const cha
 	if (!ref) {
 		resolved_base = realpath(base_dir, NULL);
 		resolved_full = realpath(full_path, NULL);
-		/* Require a path-separator boundary after the base so a sibling
-		 * directory that merely shares the base as a name prefix (say
-		 * repo.git-backup next to repo.git) cannot pass the check. */
 		if (!resolved_base || !resolved_full ||
-		    !starts_with(resolved_full, resolved_base) ||
-		    (resolved_full[strlen(resolved_base)] != '\0' &&
-		     resolved_full[strlen(resolved_base)] != '/')) {
+		    !path_within(resolved_base, resolved_full)) {
 			free(full_path);
 			full_path = NULL;
 		}
 	}
 
-	free(file);
+	free(copy);
 	free(resolved_base);
 	free(resolved_full);
 
 	return full_path;
 }
 
+void cgit_print_summary(void)
+{
+	int columns = log_columns();
+
+	cgit_print_layout_start();
+	html("<table summary='repository info' class='list nowrap'>");
+	cgit_print_branches(ctx.cfg.summary_branches);
+	htmlf("<tr class='nohover'><td colspan='%d'>&nbsp;</td></tr>", columns);
+	cgit_print_tags(ctx.cfg.summary_tags);
+	if (ctx.cfg.summary_log > 0) {
+		htmlf("<tr class='nohover'><td colspan='%d'>&nbsp;</td></tr>", columns);
+		cgit_print_log(ctx.qry.head, 0, ctx.cfg.summary_log, NULL,
+			       NULL, NULL, 0, 0, 0);
+	}
+	clone_urls_printed = 0;
+	cgit_add_clone_urls(print_clone_url);
+	html("</table>");
+	cgit_print_layout_end();
+}
+
 void cgit_print_repo_readme(const char *path)
 {
 	char *filename, *ref, *mimetype;
 	int free_filename = 0;
 
 	mimetype = cgit_get_mimetype_for_filename(path);
-	if (mimetype && (!strncmp(mimetype, "image/", 6) || !strncmp(mimetype, "video/", 6))) {
+	if (mimetype && (starts_with(mimetype, "image/") ||
+			 starts_with(mimetype, "video/"))) {
 		ctx.page.mimetype = mimetype;
 		ctx.page.charset = NULL;
 		cgit_print_plain();
@@ -129,21 +157,17 @@ void cgit_print_repo_readme(const char *path)
 
 	if (path) {
 		free_filename = 1;
-		filename = append_readme_path(filename, ref, path);
+		filename = resolve_about_path(filename, ref, path);
 		if (!filename)
 			goto done;
 	}
 
 	html("<div id='summary'>");
 	if (!ctx.repo->about_filter) {
-		/* No about-filter is configured, so there is nothing to turn
-		 * the readme source into safe HTML. Escape it rather than serve
-		 * repo content raw, which would let an untrusted repository
-		 * inject script into the about page. The pre keeps the line
-		 * structure of the text, which bare escaped output loses. Point
-		 * about-filter at the bundled about-render.lua to render a
-		 * markdown or man readme instead, see cgitrc.5.txt.
-		 */
+		// With no about-filter configured there is nothing to turn the
+		// readme source into safe HTML, so it is escaped rather than
+		// served raw, which would let an untrusted repository put
+		// script on this page.
 		html("<pre class='plaintext'>");
 		if (ref) {
 			cgit_print_file(filename, ref, 1, 1);
@@ -155,9 +179,8 @@ void cgit_print_repo_readme(const char *path)
 		}
 		html("</pre>");
 	} else {
-		/* An about-filter is configured and is responsible for turning
-		 * the source into safe HTML, so pass it through the filter raw.
-		 */
+		// The filter is what makes the source safe here, so it is fed
+		// through unescaped.
 		cgit_open_filter(ctx.repo->about_filter, filename);
 		if (ref)
 			cgit_print_file(filename, ref, 1, 0);
diff --git a/source/ui-summary.h b/source/ui-summary.h
index cba696a..676cc47 100644
--- a/source/ui-summary.h
+++ b/source/ui-summary.h
@@ -1,7 +1,14 @@
-#ifndef UI_SUMMARY_H
-#define UI_SUMMARY_H
+/*
+ * Declarations for the two pages a repository opens with, the summary of its
+ * refs and recent commits and the about page built from its readme. cmd.c
+ * reaches both, and it hands the about page whatever path follows the page
+ * name in the url, so a readme can link to a file beside it.
+ */
+
+#ifndef CGIT_UI_SUMMARY_H
+#define CGIT_UI_SUMMARY_H
 
 extern void cgit_print_summary(void);
 extern void cgit_print_repo_readme(const char *path);
 
-#endif /* UI_SUMMARY_H */
+#endif // CGIT_UI_SUMMARY_H
diff --git a/source/ui-tag.c b/source/ui-tag.c
index 7d01c5a..d09ff2b 100644
--- a/source/ui-tag.c
+++ b/source/ui-tag.c
@@ -1,43 +1,118 @@
-/* ui-tag.c: display a tag
- *
- * Copyright (C) 2006-2014 cgit Development Team <cgit@lists.zx2c4.com>
- *
- * Licensed under GNU General Public License v2
- *   (see LICENSE.txt for full license text)
+/*
+ * The tag page, which shows a single ref under refs/tags and is where the
+ * branch and tag listings link. An annotated tag is a git object in its own
+ * right, so its page carries the tagger, the date and the message, while a
+ * lightweight tag is only a name for a commit and gets a shorter table. Both
+ * forms link the object the tag points at and, when the repository enables
+ * snapshots, offer the tree at that tag for download as an archive.
  */
 
 #define USE_THE_REPOSITORY_VARIABLE
 
 #include "cgit.h"
-#include "ui-tag.h"
+#include "filter.h"
 #include "html.h"
+#include "parsing.h"
+#include "shared.h"
 #include "ui-shared.h"
+#include "ui-tag.h"
+
+static void print_object_row(struct object *obj)
+{
+	html("<tr><td>tagged object</td><td class='oid'>");
+	cgit_object_link(obj);
+	html("</td></tr>\n");
+}
+
+static void print_download_row(const char *revname)
+{
+	html("<tr><th>download</th><td class='oid'>");
+	cgit_print_snapshot_links(ctx.repo, revname, "<br/>");
+	html("</td></tr>");
+}
 
-static void print_tag_content(char *buf)
+/*
+ * Cutting the subject out terminates the message in place, which is safe only
+ * because the caller frees it straight after.
+ */
+static void print_message(char *msg)
 {
-	char *p;
+	char *newline;
 
-	if (!buf)
+	if (!msg)
 		return;
 
 	html("<div class='commit-subject'>");
-	p = strchr(buf, '\n');
-	if (p)
-		*p = '\0';
-	html_txt(buf);
+	newline = strchr(msg, '\n');
+	if (newline)
+		*newline = '\0';
+	html_txt(msg);
 	html("</div>");
-	if (p) {
+	if (newline) {
 		html("<div class='commit-msg'>");
-		html_txt(++p);
+		html_txt(newline + 1);
 		html("</div>");
 	}
 }
 
-static void print_download_links(char *revname)
+static void print_annotated_tag(const char *revname,
+				const struct object_id *oid)
 {
-	html("<tr><th>download</th><td class='oid'>");
-	cgit_print_snapshot_links(ctx.repo, revname, "<br/>");
-	html("</td></tr>");
+	struct tag *tag;
+	struct taginfo *info;
+
+	tag = lookup_tag(the_repository, oid);
+	if (!tag || parse_tag(the_repository, tag) ||
+	    !(info = cgit_parse_tag(tag))) {
+		cgit_print_error_page(500, "Internal server error",
+			"Bad tag object: %s", revname);
+		return;
+	}
+
+	cgit_print_layout_start();
+	html("<table class='commit-info'>\n");
+	html("<tr><td>tag name</td><td>");
+	html_txt(revname);
+	htmlf(" (%s)</td></tr>\n", oid_to_hex(oid));
+	if (info->tagger_date > 0) {
+		html("<tr><td>tag date</td><td>");
+		html_txt(show_date(info->tagger_date, info->tagger_tz,
+					cgit_date_mode(DATE_ISO8601)));
+		html("</td></tr>\n");
+	}
+	if (info->tagger) {
+		html("<tr><td>tagged by</td><td>");
+		cgit_open_filter(ctx.repo->email_filter, info->tagger_email,
+				 "tag");
+		html_txt(info->tagger);
+		if (info->tagger_email && !ctx.cfg.noplainemail) {
+			html(" ");
+			html_txt(info->tagger_email);
+		}
+		cgit_close_filter(ctx.repo->email_filter);
+		html("</td></tr>\n");
+	}
+	print_object_row(tag->tagged);
+	if (ctx.repo->snapshots)
+		print_download_row(revname);
+	html("</table>\n");
+	print_message(info->msg);
+	cgit_print_layout_end();
+	cgit_free_taginfo(info);
+}
+
+static void print_lightweight_tag(const char *revname, struct object *obj)
+{
+	cgit_print_layout_start();
+	html("<table class='commit-info'>\n");
+	html("<tr><td>tag name</td><td>");
+	html_txt(revname);
+	html("</td></tr>\n");
+	print_object_row(obj);
+	if (ctx.repo->snapshots)
+		print_download_row(revname);
+	html("</table>\n");
+	cgit_print_layout_end();
 }
 
 void cgit_print_tag(char *revname)
@@ -61,61 +136,10 @@ void cgit_print_tag(char *revname)
 			"Bad object id: %s", oid_to_hex(&oid));
 		goto cleanup;
 	}
-	if (obj->type == OBJ_TAG) {
-		struct tag *tag;
-		struct taginfo *info;
-
-		tag = lookup_tag(the_repository, &oid);
-		if (!tag || parse_tag(the_repository, tag) || !(info = cgit_parse_tag(tag))) {
-			cgit_print_error_page(500, "Internal server error",
-				"Bad tag object: %s", revname);
-			goto cleanup;
-		}
-		cgit_print_layout_start();
-		html("<table class='commit-info'>\n");
-		html("<tr><td>tag name</td><td>");
-		html_txt(revname);
-		htmlf(" (%s)</td></tr>\n", oid_to_hex(&oid));
-		if (info->tagger_date > 0) {
-			html("<tr><td>tag date</td><td>");
-			html_txt(show_date(info->tagger_date, info->tagger_tz,
-						cgit_date_mode(DATE_ISO8601)));
-			html("</td></tr>\n");
-		}
-		if (info->tagger) {
-			html("<tr><td>tagged by</td><td>");
-			cgit_open_filter(ctx.repo->email_filter, info->tagger_email, "tag");
-			html_txt(info->tagger);
-			if (info->tagger_email && !ctx.cfg.noplainemail) {
-				html(" ");
-				html_txt(info->tagger_email);
-			}
-			cgit_close_filter(ctx.repo->email_filter);
-			html("</td></tr>\n");
-		}
-		html("<tr><td>tagged object</td><td class='oid'>");
-		cgit_object_link(tag->tagged);
-		html("</td></tr>\n");
-		if (ctx.repo->snapshots)
-			print_download_links(revname);
-		html("</table>\n");
-		print_tag_content(info->msg);
-		cgit_print_layout_end();
-		cgit_free_taginfo(info);
-	} else {
-		cgit_print_layout_start();
-		html("<table class='commit-info'>\n");
-		html("<tr><td>tag name</td><td>");
-		html_txt(revname);
-		html("</td></tr>\n");
-		html("<tr><td>tagged object</td><td class='oid'>");
-		cgit_object_link(obj);
-		html("</td></tr>\n");
-		if (ctx.repo->snapshots)
-			print_download_links(revname);
-		html("</table>\n");
-		cgit_print_layout_end();
-	}
+	if (obj->type == OBJ_TAG)
+		print_annotated_tag(revname, &oid);
+	else
+		print_lightweight_tag(revname, obj);
 
 cleanup:
 	strbuf_release(&fullref);
diff --git a/source/ui-tag.h b/source/ui-tag.h
index d295cdc..864d29c 100644
--- a/source/ui-tag.h
+++ b/source/ui-tag.h
@@ -1,6 +1,12 @@
-#ifndef UI_TAG_H
-#define UI_TAG_H
+/*
+ * The tag page, which shows a single ref under refs/tags on its own. It is
+ * one repository command among those in cmd.c, reached with the tag name in
+ * the id query string parameter, and the branch and tag listings link to it.
+ */
+
+#ifndef CGIT_UI_TAG_H
+#define CGIT_UI_TAG_H
 
 extern void cgit_print_tag(char *revname);
 
-#endif /* UI_TAG_H */
+#endif // CGIT_UI_TAG_H
diff --git a/source/ui-tree.c b/source/ui-tree.c
index 64be0b6..0054b0c 100644
--- a/source/ui-tree.c
+++ b/source/ui-tree.c
@@ -1,76 +1,105 @@
-/* ui-tree.c: functions for tree output
- *
- * Copyright (C) 2006-2017 cgit Development Team <cgit@lists.zx2c4.com>
- *
- * Licensed under GNU General Public License v2
- *   (see LICENSE.txt for full license text)
+/*
+ * The tree page, which renders one level of a repository at a revision as a
+ * table of the folders and files in it, or the contents of the file when the
+ * path names one. A file is shown as numbered source, passed through the
+ * repository's source filter when it has one, or as a hex dump when its bytes
+ * look binary. A folder whose only entry is another folder is followed, and
+ * both are named in the same row, so a long chain of single folders does not
+ * cost a page each.
  */
 
- #define USE_THE_REPOSITORY_VARIABLE
+#define USE_THE_REPOSITORY_VARIABLE
 
 #include "cgit.h"
-#include "ui-tree.h"
+#include "filter.h"
 #include "html.h"
 #include "ui-shared.h"
+#include "ui-tree.h"
 
-/* Bytes shown per row of the binary hex dump. */
 #define HEXDUMP_ROW_BYTES 32
+#define HEXDUMP_ROW_HALF (HEXDUMP_ROW_BYTES / 2)
+#define HEXDUMP_GAP_WIDTH 4
+
+enum walk_state {
+	WALK_LOOKING,
+	WALK_LISTING,
+	WALK_BLOB_SHOWN,
+	WALK_ERROR_SHOWN,
+};
 
-struct tree_ls_entry {
+struct ls_entry {
 	struct object_id oid;
 	char *name;
 	unsigned mode;
 };
 
 struct walk_tree_context {
-	char *curr_rev;
+	char *rev;
 	char *match_path;
-	int state;
-	struct tree_ls_entry *entries;
+	enum walk_state state;
+	struct ls_entry *entries;
 	size_t entries_nr, entries_alloc;
 };
 
-static void print_text_buffer(const char *name, char *buf, unsigned long size)
+/*
+ * The count runs past one as soon as a second entry or a file turns up, which
+ * ends the descent.
+ */
+struct only_child {
+	struct strbuf *path;
+	struct object_id oid;
+	char *name;
+	size_t count;
+};
+
+/*
+ * A formatted write per line meant a syscall and a temporary buffer for every
+ * line of the file, so the anchors are handed over in batches. Building the
+ * column whole was rejected because it would come to several times the size of
+ * the blob.
+ */
+static void print_linenumbers(const char *buf, unsigned long size)
 {
-	unsigned long lineno, idx;
 	const char *numberfmt = "<a id='n%1$d' href='#n%1$d'>%1$d</a>\n";
+	struct strbuf numbers = STRBUF_INIT;
+	unsigned long lineno = 0, idx = 0;
+
+	if (size) {
+		strbuf_addf(&numbers, numberfmt, ++lineno);
+
+		// The newline that ends the last line must not open a line of
+		// its own, so the final byte is left out of the scan.
+		while (idx < size - 1) {
+			if (buf[idx] == '\n') {
+				strbuf_addf(&numbers, numberfmt, ++lineno);
+				if (numbers.len >= HTML_BATCH) {
+					html_raw(numbers.buf, numbers.len);
+					strbuf_reset(&numbers);
+				}
+			}
+			idx++;
+		}
+		html_raw(numbers.buf, numbers.len);
+	}
+	strbuf_release(&numbers);
+}
 
+static void print_text_buffer(const char *filename, char *buf,
+			      unsigned long size)
+{
 	html("<table summary='blob content' class='blob'>\n");
 
 	if (ctx.cfg.enable_tree_linenumbers) {
-		struct strbuf numbers = STRBUF_INIT;
-
 		html("<tr><td class='linenumbers'><pre>");
-		idx = 0;
-		lineno = 0;
-
-		// Build the column in batches. A formatted write per line meant
-		// a syscall and a temporary buffer for every line of the file,
-		// which is most of what rendering a large blob cost, and the
-		// whole column at once would come to several times the blob.
-		if (size) {
-			strbuf_addf(&numbers, numberfmt, ++lineno);
-			while (idx < size - 1) { // skip absolute last newline
-				if (buf[idx] == '\n') {
-					strbuf_addf(&numbers, numberfmt, ++lineno);
-					if (numbers.len >= HTML_BATCH) {
-						html_raw(numbers.buf, numbers.len);
-						strbuf_reset(&numbers);
-					}
-				}
-				idx++;
-			}
-			html_raw(numbers.buf, numbers.len);
-		}
-		strbuf_release(&numbers);
+		print_linenumbers(buf, size);
 		html("</pre></td>\n");
-	}
-	else {
+	} else {
 		html("<tr>\n");
 	}
 
 	if (ctx.repo->source_filter) {
-		char *filter_arg = xstrdup(name);
+		char *filter_arg = xstrdup(filename);
+
 		html("<td class='lines'><pre><code>");
 		cgit_open_filter(ctx.repo->source_filter, filter_arg);
 		html_raw(buf, size);
@@ -80,38 +109,47 @@ static void print_text_buffer(const char *name, char *buf, unsigned long size)
 		return;
 	}
 
-	/* No source filter is configured, so serve the text plain. Syntax
-	 * highlighting ships as an optional source filter in custom/extensions/,
-	 * keeping language knowledge out of the core. */
+	// Syntax highlighting ships as one of the filters under
+	// custom/extensions, which keeps knowledge of languages out of cgit.
 	html("<td class='lines'><pre><code>");
 	html_txt(buf);
 	html("</code></pre></td></tr></table>\n");
 }
 
+/*
+ * git's ctype macros are locale free, but there is no isgraph among them, so
+ * the dump works one out for itself.
+ */
+static int is_graphic(unsigned char ch)
+{
+	return isprint(ch) && !isspace(ch);
+}
+
 static void print_binary_buffer(char *buf, unsigned long size)
 {
-	unsigned long ofs, idx;
+	unsigned long offset, idx;
 	char ascii[HEXDUMP_ROW_BYTES + 1];
 	struct strbuf row = STRBUF_INIT;
 
 	html("<table summary='blob content' class='bin-blob'>\n");
 	html("<tr><th>ofs</th><th>hex dump</th><th>ascii</th></tr>");
-	for (ofs = 0; ofs < size; ofs += HEXDUMP_ROW_BYTES, buf += HEXDUMP_ROW_BYTES) {
-		// One write per row rather than one per byte. At the default
-		// blob limit the per-byte form spent almost all of its time in
-		// the kernel, which made a single request for a large binary
-		// blob far more expensive than the page it produced.
+
+	// At the default blob size limit a write per byte spent almost all of
+	// its time in the kernel, so a row goes out in one write.
+	for (offset = 0; offset < size;
+	     offset += HEXDUMP_ROW_BYTES, buf += HEXDUMP_ROW_BYTES) {
 		strbuf_reset(&row);
-		strbuf_addf(&row, "<tr><td class='right'>%04lx</td><td class='hex'>", ofs);
-		for (idx = 0; idx < HEXDUMP_ROW_BYTES && ofs + idx < size; idx++)
-			strbuf_addf(&row, "%*s%02x",
-				    idx == 16 ? 4 : 1, "",
-				    buf[idx] & 0xff);
+		strbuf_addf(&row, "<tr><td class='right'>%04lx</td><td class='hex'>", offset);
+		for (idx = 0; idx < HEXDUMP_ROW_BYTES && offset + idx < size; idx++) {
+			int gap = idx == HEXDUMP_ROW_HALF ? HEXDUMP_GAP_WIDTH : 1;
+
+			strbuf_addf(&row, "%*s%02x", gap, "", buf[idx] & 0xff);
+		}
 		strbuf_addstr(&row, " </td><td class='hex'>");
 		html_raw(row.buf, row.len);
 
-		for (idx = 0; idx < HEXDUMP_ROW_BYTES && ofs + idx < size; idx++)
-			ascii[idx] = isgraph((unsigned char)buf[idx]) ? buf[idx] : '.';
+		for (idx = 0; idx < HEXDUMP_ROW_BYTES && offset + idx < size; idx++)
+			ascii[idx] = is_graphic(buf[idx]) ? buf[idx] : '.';
 		ascii[idx] = '\0';
 		html_txt(ascii);
 		html("</td></tr>\n");
@@ -120,9 +158,13 @@ static void print_binary_buffer(char *buf, unsigned long size)
 	html("</table>\n");
 }
 
-/* Returns 1 if it opened the page layout for the caller to close, or 0 if
- * it emitted a complete standalone error page. */
-static int print_object(const struct object_id *oid, const char *path, const char *basename, const char *rev)
+/*
+ * Answer whether the page layout was left open for the caller to close. A
+ * false means a complete error page went out in place of the blob, so nothing
+ * may be added to it.
+ */
+static bool print_object(const struct object_id *oid, const char *path,
+			 const char *filename, const char *rev)
 {
 	enum object_type type;
 	char *buf;
@@ -133,22 +175,21 @@ static int print_object(const struct object_id *oid, const char *path, const cha
 	if (type == OBJ_BAD) {
 		cgit_print_error_page(404, "Not found",
 			"Bad object name: %s", oid_to_hex(oid));
-		return 0;
+		return false;
 	}
 
-	/* Reject an oversized object before reading it whole into memory. */
 	if (ctx.cfg.max_blob_size && size / 1024 > (unsigned long)ctx.cfg.max_blob_size) {
 		cgit_print_error_page(413, "Too large",
 			"blob size (%luKB) exceeds display size limit (%dKB)",
 			size / 1024, ctx.cfg.max_blob_size);
-		return 0;
+		return false;
 	}
 
 	buf = odb_read_object(the_repository->objects, oid, &type, &size);
 	if (!buf) {
 		cgit_print_error_page(500, "Internal server error",
 			"Error reading object %s", oid_to_hex(oid));
-		return 0;
+		return false;
 	}
 	is_binary = buffer_is_binary(buf, size);
 
@@ -168,46 +209,37 @@ static int print_object(const struct object_id *oid, const char *path, const cha
 	if (is_binary)
 		print_binary_buffer(buf, size);
 	else
-		print_text_buffer(basename, buf, size);
+		print_text_buffer(filename, buf, size);
 
 	free(buf);
-	return 1;
+	return true;
 }
 
-struct single_tree_ctx {
-	struct strbuf *path;
-	struct object_id oid;
-	char *name;
-	size_t count;
-};
-
-static int single_tree_cb(const struct object_id *oid, struct strbuf *base,
-			  const char *pathname, unsigned mode, void *cbdata)
+static int only_child_cb(const struct object_id *oid, struct strbuf *base,
+			 const char *pathname, unsigned mode, void *data)
 {
-	// Not named ctx: that is the global request context, and shadowing it
-	// here would hide it from anything added to this function later.
-	struct single_tree_ctx *tree_ctx = cbdata;
+	struct only_child *child = data;
 
-	if (++tree_ctx->count > 1)
+	if (++child->count > 1)
 		return -1;
 
 	if (!S_ISDIR(mode)) {
-		tree_ctx->count = 2;
+		child->count = 2;
 		return -1;
 	}
 
-	tree_ctx->name = xstrdup(pathname);
-	oidcpy(&tree_ctx->oid, oid);
-	strbuf_addf(tree_ctx->path, "/%s", pathname);
+	child->name = xstrdup(pathname);
+	oidcpy(&child->oid, oid);
+	strbuf_addf(child->path, "/%s", pathname);
 	return 0;
 }
 
-static void write_tree_link(const struct object_id *oid, char *name,
+static void print_dir_chain(const struct object_id *oid, char *name,
 			    char *rev, struct strbuf *fullpath)
 {
 	size_t initial_length = fullpath->len;
 	struct tree *tree;
-	struct single_tree_ctx tree_ctx = {
+	struct only_child child = {
 		.path = fullpath,
 		.count = 1,
 	};
@@ -215,34 +247,34 @@ static void write_tree_link(const struct object_id *oid, char *name,
 		.nr = 0
 	};
 
-	oidcpy(&tree_ctx.oid, oid);
+	oidcpy(&child.oid, oid);
 
-	while (tree_ctx.count == 1) {
+	while (child.count == 1) {
 		cgit_tree_link(name, NULL, "ls-dir", ctx.qry.head, rev,
 			       fullpath->buf);
 
-		tree = lookup_tree(the_repository, &tree_ctx.oid);
+		tree = lookup_tree(the_repository, &child.oid);
 		if (!tree)
 			return;
 
-		free(tree_ctx.name);
-		tree_ctx.name = NULL;
-		tree_ctx.count = 0;
+		free(child.name);
+		child.name = NULL;
+		child.count = 0;
 
-		read_tree(the_repository, tree, &paths, single_tree_cb, &tree_ctx);
+		read_tree(the_repository, tree, &paths, only_child_cb, &child);
 
-		if (tree_ctx.count != 1)
+		if (child.count != 1)
 			break;
 
 		html(" / ");
-		name = tree_ctx.name;
+		name = child.name;
 	}
 
 	strbuf_setlen(fullpath, initial_length);
 }
 
-static void render_ls_item(const struct object_id *oid, const char *pathname,
-			   unsigned mode, struct walk_tree_context *walk_tree_ctx)
+static void print_ls_row(const struct object_id *oid, const char *pathname,
+			 unsigned mode, struct walk_tree_context *walk)
 {
 	char *name;
 	struct strbuf fullpath = STRBUF_INIT;
@@ -259,9 +291,11 @@ static void render_ls_item(const struct object_id *oid, const char *pathname,
 	if (!S_ISGITLINK(mode)) {
 		type = odb_read_object_info(the_repository->objects, oid, &size);
 		if (type == OBJ_BAD) {
-			htmlf("<tr><td colspan='3'>Bad object: %s %s</td></tr>",
-			      name,
-			      oid_to_hex(oid));
+			// The name comes from the tree, so it can hold
+			// anything a commit was allowed to record.
+			html("<tr><td colspan='3'>Bad object: ");
+			html_txt(name);
+			htmlf(" %s</td></tr>", oid_to_hex(oid));
 			goto cleanup;
 		}
 	}
@@ -272,15 +306,14 @@ static void render_ls_item(const struct object_id *oid, const char *pathname,
 	if (S_ISGITLINK(mode)) {
 		cgit_submodule_link("ls-mod", fullpath.buf, oid_to_hex(oid));
 	} else if (S_ISDIR(mode)) {
-		write_tree_link(oid, name, walk_tree_ctx->curr_rev,
-				&fullpath);
+		print_dir_chain(oid, name, walk->rev, &fullpath);
 	} else {
 		char *ext = strrchr(name, '.');
 		strbuf_addstr(&class, "ls-blob");
 		if (ext)
 			strbuf_addf(&class, " %s", ext + 1);
 		cgit_tree_link(name, NULL, class.buf, ctx.qry.head,
-			       walk_tree_ctx->curr_rev, fullpath.buf);
+			       walk->rev, fullpath.buf);
 	}
 	if (S_ISLNK(mode)) {
 		html(" -> ");
@@ -293,7 +326,7 @@ static void render_ls_item(const struct object_id *oid, const char *pathname,
 		strbuf_addf(&linkpath, "/../%s", buf);
 		strbuf_normalize_path(&linkpath);
 		cgit_tree_link(buf, NULL, class.buf, ctx.qry.head,
-			walk_tree_ctx->curr_rev, linkpath.buf);
+			walk->rev, linkpath.buf);
 		free(buf);
 		strbuf_release(&linkpath);
 	}
@@ -301,17 +334,17 @@ static void render_ls_item(const struct object_id *oid, const char *pathname,
 
 	html("<td>");
 	cgit_log_link("log", NULL, "button", ctx.qry.head,
-		      walk_tree_ctx->curr_rev, fullpath.buf, 0, NULL, NULL,
+		      walk->rev, fullpath.buf, 0, NULL, NULL,
 		      ctx.qry.showmsg, 0);
 	if (ctx.repo->enable_stats)
 		cgit_stats_link("stats", NULL, "button", ctx.qry.head,
 				fullpath.buf);
 	if (!S_ISGITLINK(mode))
 		cgit_plain_link("plain", NULL, "button", ctx.qry.head,
-				walk_tree_ctx->curr_rev, fullpath.buf);
+				walk->rev, fullpath.buf);
 	if (!S_ISDIR(mode) && ctx.repo->enable_blame)
 		cgit_blame_link("blame", NULL, "button", ctx.qry.head,
-				walk_tree_ctx->curr_rev, fullpath.buf);
+				walk->rev, fullpath.buf);
 	html("</td></tr>\n");
 
 cleanup:
@@ -321,50 +354,50 @@ cleanup:
 }
 
 static int ls_item(const struct object_id *oid, struct strbuf *base,
-		const char *pathname, unsigned mode, void *cbdata)
+		const char *pathname, unsigned mode, void *data)
 {
-	struct walk_tree_context *walk_tree_ctx = cbdata;
+	struct walk_tree_context *walk = data;
 
-	/* When grouping directories first, collect the entries now and render
-	 * them once the whole level has been read (see ls_flush). */
+	// With folders grouped first, a row cannot go out as it arrives,
+	// because git hands the level over in name order.
 	if (ctx.cfg.enable_tree_group_dirs) {
-		struct tree_ls_entry *e;
-		ALLOC_GROW(walk_tree_ctx->entries, walk_tree_ctx->entries_nr + 1,
-			   walk_tree_ctx->entries_alloc);
-		e = &walk_tree_ctx->entries[walk_tree_ctx->entries_nr++];
-		oidcpy(&e->oid, oid);
-		e->name = xstrdup(pathname);
-		e->mode = mode;
+		struct ls_entry *entry;
+
+		ALLOC_GROW(walk->entries, walk->entries_nr + 1,
+			   walk->entries_alloc);
+		entry = &walk->entries[walk->entries_nr++];
+		oidcpy(&entry->oid, oid);
+		entry->name = xstrdup(pathname);
+		entry->mode = mode;
 		return 0;
 	}
 
-	render_ls_item(oid, pathname, mode, walk_tree_ctx);
+	print_ls_row(oid, pathname, mode, walk);
 	return 0;
 }
 
-/* Render any entries collected by ls_item, directories first and then the
- * rest, keeping git's ordering within each group. A no-op unless directory
- * grouping is enabled, in which case nothing was collected. */
-static void ls_flush(struct walk_tree_context *walk_tree_ctx)
+static void ls_flush(struct walk_tree_context *walk)
 {
 	size_t i;
 
-	for (i = 0; i < walk_tree_ctx->entries_nr; i++) {
-		struct tree_ls_entry *e = &walk_tree_ctx->entries[i];
-		if (S_ISDIR(e->mode))
-			render_ls_item(&e->oid, e->name, e->mode, walk_tree_ctx);
+	for (i = 0; i < walk->entries_nr; i++) {
+		struct ls_entry *entry = &walk->entries[i];
+
+		if (S_ISDIR(entry->mode))
+			print_ls_row(&entry->oid, entry->name, entry->mode, walk);
 	}
-	for (i = 0; i < walk_tree_ctx->entries_nr; i++) {
-		struct tree_ls_entry *e = &walk_tree_ctx->entries[i];
-		if (!S_ISDIR(e->mode))
-			render_ls_item(&e->oid, e->name, e->mode, walk_tree_ctx);
+	for (i = 0; i < walk->entries_nr; i++) {
+		struct ls_entry *entry = &walk->entries[i];
+
+		if (!S_ISDIR(entry->mode))
+			print_ls_row(&entry->oid, entry->name, entry->mode, walk);
 	}
 
-	for (i = 0; i < walk_tree_ctx->entries_nr; i++)
-		free(walk_tree_ctx->entries[i].name);
-	free(walk_tree_ctx->entries);
-	walk_tree_ctx->entries = NULL;
-	walk_tree_ctx->entries_nr = walk_tree_ctx->entries_alloc = 0;
+	for (i = 0; i < walk->entries_nr; i++)
+		free(walk->entries[i].name);
+	free(walk->entries);
+	walk->entries = NULL;
+	walk->entries_nr = walk->entries_alloc = 0;
 }
 
 static void ls_head(void)
@@ -385,7 +418,8 @@ static void ls_tail(void)
 	cgit_print_layout_end();
 }
 
-static void ls_tree(const struct object_id *oid, const char *path, struct walk_tree_context *walk_tree_ctx)
+static void ls_tree(const struct object_id *oid, const char *path,
+		    struct walk_tree_context *walk)
 {
 	struct tree *tree;
 	struct pathspec paths = {
@@ -400,64 +434,64 @@ static void ls_tree(const struct object_id *oid, const char *path, struct walk_t
 	}
 
 	ls_head();
-	read_tree(the_repository, tree, &paths, ls_item, walk_tree_ctx);
-	ls_flush(walk_tree_ctx);
+	read_tree(the_repository, tree, &paths, ls_item, walk);
+	ls_flush(walk);
 	ls_tail();
 }
 
 static int walk_tree(const struct object_id *oid, struct strbuf *base,
-		const char *pathname, unsigned mode, void *cbdata)
+		const char *pathname, unsigned mode, void *data)
 {
-	struct walk_tree_context *walk_tree_ctx = cbdata;
+	struct walk_tree_context *walk = data;
 
-	if (walk_tree_ctx->state == 0) {
-		struct strbuf buffer = STRBUF_INIT;
+	if (walk->state == WALK_LOOKING) {
+		struct strbuf fullpath = STRBUF_INIT;
 
-		strbuf_addbuf(&buffer, base);
-		strbuf_addstr(&buffer, pathname);
-		if (strcmp(walk_tree_ctx->match_path, buffer.buf))
+		strbuf_addbuf(&fullpath, base);
+		strbuf_addstr(&fullpath, pathname);
+		if (strcmp(walk->match_path, fullpath.buf))
 			return READ_TREE_RECURSIVE;
 
 		if (S_ISDIR(mode)) {
-			walk_tree_ctx->state = 1;
-			cgit_set_title_from_path(buffer.buf);
-			strbuf_release(&buffer);
+			walk->state = WALK_LISTING;
+			cgit_set_title_from_path(fullpath.buf);
+			strbuf_release(&fullpath);
 			ls_head();
 			return READ_TREE_RECURSIVE;
 		} else {
-			/* state 2: layout left open for us to close; state 3:
-			 * print_object already emitted a standalone error page. */
-			walk_tree_ctx->state =
-				print_object(oid, buffer.buf, pathname,
-					     walk_tree_ctx->curr_rev) ? 2 : 3;
-			strbuf_release(&buffer);
+			bool shown = print_object(oid, fullpath.buf,
+						  pathname, walk->rev);
+
+			walk->state = shown ? WALK_BLOB_SHOWN : WALK_ERROR_SHOWN;
+			strbuf_release(&fullpath);
 			return 0;
 		}
 	}
-	ls_item(oid, base, pathname, mode, walk_tree_ctx);
+	ls_item(oid, base, pathname, mode, walk);
 	return 0;
 }
 
-/*
- * Show a tree or a blob
- *   rev:  the commit pointing at the root tree object
- *   path: path to tree or blob
- */
+// Either argument may be null, in which case the head of the current query
+// stands in for the revision and the listing starts at the root of the tree.
 void cgit_print_tree(const char *rev, char *path)
 {
 	struct object_id oid;
 	struct commit *commit;
+	int path_len = path ? strlen(path) : 0;
+	// nowildcard_len matches len so git treats the path as literal rather
+	// than as a glob.
 	struct pathspec_item path_items = {
 		.match = path,
-		.len = path ? strlen(path) : 0
+		.len = path_len,
+		.nowildcard_len = path_len
 	};
 	struct pathspec paths = {
 		.nr = path ? 1 : 0,
 		.items = &path_items
 	};
-	struct walk_tree_context walk_tree_ctx = {
+	struct walk_tree_context walk = {
 		.match_path = path,
-		.state = 0
+		.state = WALK_LOOKING
 	};
 
 	if (!rev)
@@ -475,24 +509,25 @@ void cgit_print_tree(const char *rev, char *path)
 		return;
 	}
 
-	walk_tree_ctx.curr_rev = xstrdup(rev);
+	walk.rev = xstrdup(rev);
 
 	if (path == NULL) {
-		ls_tree(get_commit_tree_oid(commit), NULL, &walk_tree_ctx);
+		ls_tree(get_commit_tree_oid(commit), NULL, &walk);
 		goto cleanup;
 	}
 
 	read_tree(the_repository, repo_get_commit_tree(the_repository, commit),
-		  &paths, walk_tree, &walk_tree_ctx);
-	if (walk_tree_ctx.state == 1) {
-		ls_flush(&walk_tree_ctx);
+		  &paths, walk_tree, &walk);
+	if (walk.state == WALK_LISTING) {
+		ls_flush(&walk);
 		ls_tail();
-	} else if (walk_tree_ctx.state == 2)
+	} else if (walk.state == WALK_BLOB_SHOWN)
 		cgit_print_layout_end();
-	else if (walk_tree_ctx.state == 0)
+	else if (walk.state == WALK_LOOKING)
 		cgit_print_error_page(404, "Not found", "Path not found");
-	/* state 3: print_object already emitted a complete error page */
+	// WALK_ERROR_SHOWN is left alone, since the error page print_object
+	// put out is already complete.
 
 cleanup:
-	free(walk_tree_ctx.curr_rev);
+	free(walk.rev);
 }
diff --git a/source/ui-tree.h b/source/ui-tree.h
index bbd34e3..f99e085 100644
--- a/source/ui-tree.h
+++ b/source/ui-tree.h
@@ -1,6 +1,13 @@
-#ifndef UI_TREE_H
-#define UI_TREE_H
+/*
+ * The tree page, which browses a repository at a revision, listing the folder
+ * a path selects or showing the contents of the file it names. It is one of
+ * the repository commands in cmd.c, and it is where the plain, blame, log and
+ * stats views of a path are linked from.
+ */
+
+#ifndef CGIT_UI_TREE_H
+#define CGIT_UI_TREE_H
 
 extern void cgit_print_tree(const char *rev, char *path);
 
-#endif /* UI_TREE_H */
+#endif // CGIT_UI_TREE_H
diff --git a/tools/release-build.sh b/tools/release-build.sh
index b55fc43..12f4260 100755
--- a/tools/release-build.sh
+++ b/tools/release-build.sh
@@ -1,16 +1,11 @@
 #!/bin/sh
-# Build cgit for a release with Linux hardening flags.
-#
-# These are ELF and GCC/Clang specific, so this targets a Linux deploy
-# rather than local macOS development, where a plain make is enough. The
-# flags add a stack protector, fortified libc calls, a position independent
-# executable, and full RELRO.
-#
-# By default Lua is pinned off so the binary needs no Lua at runtime. Pass
-# "lua" as the first argument to link the lua: filter backend instead, which
-# needs a Lua dev package installed.
-#
-# Usage: ./tools/release-build.sh [lua]
+# Build cgit for a release with Linux hardening flags. These are ELF and GCC
+# or Clang specific, so this targets a Linux deploy rather than local macOS
+# development, where a plain make is enough. The flags add a stack protector,
+# fortified libc calls, a position independent executable, and full RELRO. By
+# default Lua is pinned off so the binary needs no Lua at runtime, and running
+# it as ./tools/release-build.sh lua links in the backend behind the "lua:"
+# filter prefix instead, which needs a Lua dev package installed.
 
 set -eu
 
@@ -35,14 +30,16 @@ cd "$(dirname "$0")/.."
 CC=${CC:-cc}
 
 # Not every hardening flag exists on every toolchain, and an unknown one would
-# fail the build rather than be ignored, so each is compiled before it is used.
+# fail the build rather than be ignored, so the ones that might be missing are
+# compiled before they are used. The argument is left unquoted so a caller can
+# probe several flags at once.
 supports() {
 	printf 'int main(void){return 0;}\n' >"$probe.c"
 	$CC $1 -o "$probe.out" "$probe.c" >/dev/null 2>&1
 }
 
-# An explicit template rather than -t, whose handling of a prefix differs
-# between the GNU and BSD versions.
+# The template is spelled out rather than passed with -t, whose handling of a
+# prefix differs between the GNU and BSD versions.
 probe=$(mktemp "${TMPDIR:-/tmp}/cgit-probe.XXXXXX")
 trap 'rm -f "$probe" "$probe.c" "$probe.out"' EXIT
 
@@ -59,7 +56,8 @@ else
 	CFLAGS="$CFLAGS -U_FORTIFY_SOURCE -D_FORTIFY_SOURCE=2"
 fi
 
-# Makes a large stack frame touch each guard page rather than step over it.
+# Without this a large stack frame can step over a guard page rather than
+# touch it.
 if supports "-fstack-clash-protection"; then
 	CFLAGS="$CFLAGS -fstack-clash-protection"
 fi
@@ -70,20 +68,20 @@ LDFLAGS="-pie \
 
 # These reach only the cgit objects, so git's own sources are not held to them.
 # -Wformat-security is an error because a non-literal format with no arguments
-# is never intentional.
-#
-# The shape that let a repository supply its own format string through
-# module-link was a non-literal *with* arguments, which only -Wformat-nonliteral
-# reports. It is left out here because it also fires on forwarding a va_list and
-# on local format constants, so it cannot be an error without false positives.
-# Worth running by hand when touching anything that formats:
+# is never intentional. The shape that let a repository supply its own format
+# string through module-link was the other one, a non-literal that does take
+# arguments, and only -Wformat-nonliteral reports that, and that one is left out because it also fires
+# on forwarding a va_list and on local format constants, so it cannot be an
+# error without false positives. It is still worth running by hand when
+# touching anything that formats.
 #
 #     make cgit CGIT_EXTRA_CFLAGS=-Wformat-nonliteral
 #
 CGIT_EXTRA_CFLAGS="-Wformat -Wformat-security -Werror=format-security"
 
-# A full rebuild so the bundled git objects pick up the same flags.
-# cleanall also descends into vendor/git, which plain clean does not.
+# The build starts from clean so the bundled git objects pick up the same
+# flags, and cleanall rather than clean because only cleanall descends into
+# vendor/git.
 make cleanall
 exec make "$@" CFLAGS="$CFLAGS" LDFLAGS="$LDFLAGS" \
 	CGIT_EXTRA_CFLAGS="$CGIT_EXTRA_CFLAGS"
author	Bryce Kwon <bryce@brycekwon.com>	2026-08-10 19:29:44 -1000
committer	Bryce Kwon <bryce@brycekwon.com>	2026-08-10 19:29:44 -1000
commit	6c62bdc268ce8a8d98ab999846fcabd972507ad6 (patch)
parent	b8c11770bc2d7aa8638d6f1a99074e2ce626f2b2 (diff)
tree	5e6dc6aa5bd640cb3fe2975586573fdafc10261d
download	cgit-6c62bdc268ce8a8d98ab999846fcabd972507ad6.tar.gz cgit-6c62bdc268ce8a8d98ab999846fcabd972507ad6.zip
-rw-r--r--	Makefile	65
-rw-r--r--	source/cache.c	464
-rw-r--r--	source/cache.h	41
-rw-r--r--	source/cgit.c	1761
-rw-r--r--	source/cgit.h	186
-rw-r--r--	source/cgit.mk	61
-rw-r--r--	source/cmd.c	164
-rw-r--r--	source/cmd.h	16
-rw-r--r--	source/config.c	113
-rw-r--r--	source/config.h	16
-rw-r--r--	source/configfile.c	100
-rw-r--r--	source/configfile.h	10
-rw-r--r--	source/filter.c	324
-rw-r--r--	source/filter.h	54
-rwxr-xr-x	source/gen-version.sh	34
-rw-r--r--	source/html.c	277
-rw-r--r--	source/html.h	70
-rw-r--r--	source/parsing.c	278
-rw-r--r--	source/parsing.h	20
-rw-r--r--	source/scan-tree.c	254
-rw-r--r--	source/scan-tree.h	14
-rw-r--r--	source/shared.c	620
-rw-r--r--	source/shared.h	92
-rw-r--r--	source/ui-atom.c	107
-rw-r--r--	source/ui-atom.h	11
-rw-r--r--	source/ui-blame.c	324
-rw-r--r--	source/ui-blame.h	13
-rw-r--r--	source/ui-blob.c	147
-rw-r--r--	source/ui-blob.h	22
-rw-r--r--	source/ui-clone.c	107
-rw-r--r--	source/ui-clone.h	19
-rw-r--r--	source/ui-commit.c	186
-rw-r--r--	source/ui-commit.h	12
-rw-r--r--	source/ui-diff.c	629
-rw-r--r--	source/ui-diff.h	20
-rw-r--r--	source/ui-empty.c	15
-rw-r--r--	source/ui-empty.h	13
-rw-r--r--	source/ui-log.c	507
-rw-r--r--	source/ui-log.h	13
-rw-r--r--	source/ui-patch.c	124
-rw-r--r--	source/ui-patch.h	13
-rw-r--r--	source/ui-plain.c	144
-rw-r--r--	source/ui-plain.h	12
-rw-r--r--	source/ui-refs.c	252
-rw-r--r--	source/ui-refs.h	23
-rw-r--r--	source/ui-repolist.c	363
-rw-r--r--	source/ui-repolist.h	13
-rw-r--r--	source/ui-shared.c	1525
-rw-r--r--	source/ui-shared.h	64
-rw-r--r--	source/ui-snapshot.c	298
-rw-r--r--	source/ui-snapshot.h	40
-rw-r--r--	source/ui-ssdiff.c	501
-rw-r--r--	source/ui-ssdiff.h	19
-rw-r--r--	source/ui-stats.c	574
-rw-r--r--	source/ui-stats.h	36
-rw-r--r--	source/ui-summary.c	143
-rw-r--r--	source/ui-summary.h	13
-rw-r--r--	source/ui-tag.c	174
-rw-r--r--	source/ui-tag.h	12
-rw-r--r--	source/ui-tree.c	381
-rw-r--r--	source/ui-tree.h	13
-rwxr-xr-x	tools/release-build.sh	48