index : cgit

/*
 * The entry point every request passes through. It reads cgitrc and the query
 * string into the global ctx, works out how long the answer may be cached,
 * and then hands over to the cache, which either serves a copy it already has
 * or calls back here to render one. Rendering means authenticating the
 * request, resolving the repository and the branch it names, and dispatching
 * to the page handler cmd.c maps the page name to. Run from a shell rather
 * than from a web server the same binary instead prints what it was built
 * with, or scans a directory tree and writes out the repositories it found in
 * cgitrc form.
 */

#define USE_THE_REPOSITORY_VARIABLE

#include "cache.h"
#include "cgit.h"
#include "cmd.h"
#include "config.h"
#include "filter.h"
#include "html.h"
#include "parsing.h"
#include "scan-tree.h"
#include "shared.h"
#include "ui-blob.h"
#include "ui-diff.h"
#include "ui-empty.h"
#include "ui-shared.h"
#include "ui-snapshot.h"
#include "ui-stats.h"

// Deliberately small. The whole body is read at once and handed to the auth
// filter, so a filter never has to defend itself against a huge or a
// dribbled-out POST.
#define MAX_AUTHENTICATION_POST_BYTES 4096

// The filter bar matches the search against every item a page lists, so this
// bounds both the work one request can ask for and the cache key the query
// lands in.
#define MAX_SEARCH_LEN 512

// Furthest into a listing a request may ask to start. A walk has to step over
// every row it skips, so this bounds the work an offset alone can buy.
#define MAX_QUERY_OFFSET 100000

// The first branch found is the fallback, so a repository whose default branch
// does not exist still has something to show.
struct refmatch {
	char *wanted;
	char *first;
	int found;
};

const char *cgit_version = CGIT_VERSION;

/*
 * Isolate git from the calling user's configuration, which could otherwise
 * break once HOME is unset below. It runs as a constructor because git reads
 * the global config for its trace settings before cmd_main is reached, and
 * cmd_main calls it again for a compiler that drops the attribute, which
 * git-compat-util.h defines away where it is unsupported.
 */
__attribute__((constructor))
static void isolate_git_environment(void)
{
	setenv("GIT_CONFIG_NOSYSTEM", "1", 1);
	setenv("GIT_CONFIG_GLOBAL", "/dev/null", 1);
	setenv("GIT_ATTR_NOSYSTEM", "1", 1);
	unsetenv("HOME");
	unsetenv("XDG_CONFIG_HOME");
}

/*
 * Turn a die from anywhere inside git into a rendered page, since a CGI that
 * produced no output leaves the visitor with whatever the web server makes of
 * it. Git's message names files and objects on the server, so it goes to the
 * log and the visitor gets a fixed page.
 */
static NORETURN void die_routine(const char *msg, va_list params)
{
	fputs("[cgit] ", stderr);
	vfprintf(stderr, msg, params);
	fputc('\n', stderr);
	cgit_abort_filters();
	cgit_print_error_page(500, "Internal Server Error", "Unable to complete the request");
	exit(0);
}

static char *ensure_end(const char *str, char c)
{
	size_t len = strlen(str);

	return len && str[len - 1] == c ? xstrdup(str) : xstrfmt("%s%c", str, c);
}

static void prepare_context(void)
{
	const char *script_name = getenv("SCRIPT_NAME");

	memset(&ctx, 0, sizeof(ctx));
	ctx.cfg.agefile = "info/web/last-modified";
	ctx.cfg.cache_root = CGIT_CACHE_ROOT;
	ctx.cfg.cache_about_ttl = 15;
	ctx.cfg.cache_snapshot_ttl = 5;
	ctx.cfg.cache_summary_ttl = 5;
	ctx.cfg.cache_index_ttl = 5;
	ctx.cfg.cache_scan_ttl = 15;
	ctx.cfg.cache_dynamic_ttl = 5;
	ctx.cfg.cache_static_ttl = -1;
	// Counted in kilobytes, so sixty-four megabytes, past which a response
	// is served but not kept.
	ctx.cfg.cache_max_slot_size = 64 * 1024;
	ctx.cfg.case_sensitive_sort = 1;
	ctx.cfg.logo = "/cgit.png";
	ctx.cfg.favicon = "/favicon.ico";
	ctx.cfg.date_mode = date_mode_from_type(DATE_SHORT);
	ctx.cfg.enable_header = 1;
	ctx.cfg.enable_mailmap = 1;
	ctx.cfg.enable_plain_email = 1;
	ctx.cfg.enable_relative_dates = 1;
	ctx.cfg.enable_http_clone = 1;
	ctx.cfg.enable_index_owner = 1;
	ctx.cfg.enable_tree_linenumbers = 1;
	ctx.cfg.max_repo_count = 50;
	ctx.cfg.max_commit_count = 50;
	ctx.cfg.max_patch_count = 50;
	ctx.cfg.max_diff_files = 200;
	ctx.cfg.max_diff_lines = 1000;
	ctx.cfg.max_msg_len = 80;
	ctx.cfg.max_ref_count = 200;
	ctx.cfg.max_repodesc_len = 80;
	// Counted in kilobytes, so ten megabytes, which bounds the memory one
	// request can be made to allocate for a blob.
	ctx.cfg.max_blob_size = 10 * 1024;
	ctx.cfg.renamelimit = -1;
	ctx.cfg.robots = "index, nofollow";
	ctx.cfg.root_title = "Git repository browser";
	ctx.cfg.root_desc = "a fast webinterface for the git dscm";
	ctx.cfg.section = "";
	ctx.cfg.repository_sort = "name";
	ctx.cfg.summary_branches = 10;
	ctx.cfg.summary_log = 10;
	ctx.cfg.summary_tags = 10;
	ctx.cfg.max_atom_items = 10;
	ctx.env.cgit_config = getenv("CGIT_CONFIG");
	ctx.env.http_host = getenv("HTTP_HOST");
	ctx.env.https = getenv("HTTPS");
	ctx.env.no_http = getenv("NO_HTTP");
	ctx.env.path_info = getenv("PATH_INFO");
	ctx.env.query_string = getenv("QUERY_STRING");
	ctx.env.request_method = getenv("REQUEST_METHOD");
	ctx.env.server_name = getenv("SERVER_NAME");
	ctx.env.server_port = getenv("SERVER_PORT");
	ctx.env.http_cookie = getenv("HTTP_COOKIE");
	ctx.env.http_referer = getenv("HTTP_REFERER");
	ctx.env.content_length =
		getenv("CONTENT_LENGTH") ? strtoul(getenv("CONTENT_LENGTH"), NULL, 10) : 0;
	ctx.page.mimetype = "text/html";
	ctx.page.charset = PAGE_ENCODING;
	string_list_init_dup(&ctx.cfg.mimetypes);
	ctx.cfg.virtual_root = ensure_end(script_name ? script_name : CGIT_SCRIPT_NAME, '/');
	if (ctx.env.query_string)
		ctx.qry.raw = xstrdup(ctx.env.query_string);
	if (!ctx.env.cgit_config)
		ctx.env.cgit_config = CGIT_CONFIG;
}

static void print_version(void)
{
	printf("cgit %s | https://github.com/brycekwon/cgit\n\nCompiled in features:\n", CGIT_VERSION);
#ifdef NO_LUA
	printf("[-] ");
#else
	printf("[+] ");
#endif
	printf("Lua scripting\n");
#ifndef HAVE_LINUX_SENDFILE
	printf("[-] ");
#else
	printf("[+] ");
#endif
	printf("Linux sendfile() usage\n");
}

static int cmp_repos(const void *a, const void *b)
{
	const struct cgit_repo *repo_a = a, *repo_b = b;

	return strcmp(repo_a->url, repo_b->url);
}

static char *build_snapshot_setting(int mask)
{
	const struct cgit_snapshot_format *format;
	struct strbuf result = STRBUF_INIT;

	for (format = cgit_snapshot_formats; format->suffix; format++) {
		if (cgit_snapshot_format_bit(format) & mask) {
			if (result.len)
				strbuf_addch(&result, ' ');
			strbuf_addstr(&result, format->suffix);
		}
	}
	return strbuf_detach(&result, NULL);
}

/*
 * Only a filter the repository set itself is written, since one it kept from
 * the global settings is copied in again when the list is read back. An empty
 * value stands for a repository that turned the global filter off.
 */
static void print_filter(FILE *f, const char *key, struct cgit_filter *filter,
	struct cgit_filter *global)
{
	if (filter == global)
		return;
	if (filter)
		cgit_fprintf_filter(filter, f, key);
	else
		fprintf(f, "%s\n", key);
}

static void print_repo(FILE *f, struct cgit_repo *repo)
{
	struct string_list_item *item;

	fprintf(f, "repo.url=%s\n", repo->url);
	fprintf(f, "repo.name=%s\n", repo->name);
	if (repo->path)
		fprintf(f, "repo.path=%s\n", repo->path);
	if (repo->owner)
		fprintf(f, "repo.owner=%s\n", repo->owner);
	// Written out, the placeholder would read back as a description of
	// its own and describe the repository's pages.
	if (repo->desc && repo->desc != cgit_default_repo_desc)
		fprintf(f, "repo.desc=%s\n", repo->desc);
	for_each_string_list_item(item, &repo->readme) {
		if (item->util)
			fprintf(f, "repo.readme=%s:%s\n", (char *)item->util, item->string);
		else
			fprintf(f, "repo.readme=%s\n", item->string);
	}
	if (repo->defbranch)
		fprintf(f, "repo.defbranch=%s\n", repo->defbranch);
	if (repo->head_content)
		fprintf(f, "repo.head-content=%s\n", repo->head_content);
	if (repo->module_link)
		fprintf(f, "repo.module-link=%s\n", repo->module_link);
	for_each_string_list_item(item, &repo->submodules)
		fprintf(f, "repo.module-link.%s=%s\n", item->string, (char *)item->util);
	if (repo->section)
		fprintf(f, "repo.section=%s\n", repo->section);
	if (repo->clone_url)
		fprintf(f, "repo.clone-url=%s\n", repo->clone_url);
	fprintf(f, "repo.enable-blame=%d\n", repo->enable_blame);
	fprintf(f, "repo.enable-commit-graph=%d\n", repo->enable_commit_graph);
	fprintf(f, "repo.enable-follow-links=%d\n", repo->enable_follow_links);
	fprintf(f, "repo.enable-gitmodules-links=%d\n", repo->enable_gitmodules_links);
	fprintf(f, "repo.enable-log-filecount=%d\n", repo->enable_log_filecount);
	fprintf(f, "repo.enable-log-linecount=%d\n", repo->enable_log_linecount);
	fprintf(f, "repo.enable-mailmap=%d\n", repo->enable_mailmap);
	print_filter(f, "repo.about-filter=", repo->about_filter, ctx.cfg.about_filter);
	print_filter(f, "repo.commit-filter=", repo->commit_filter, ctx.cfg.commit_filter);
	print_filter(f, "repo.source-filter=", repo->source_filter, ctx.cfg.source_filter);
	print_filter(f, "repo.email-filter=", repo->email_filter, ctx.cfg.email_filter);
	print_filter(f, "repo.trailer-filter=", repo->trailer_filter, ctx.cfg.trailer_filter);
	if (repo->snapshots != ctx.cfg.snapshots) {
		char *formats = build_snapshot_setting(repo->snapshots);
		fprintf(f, "repo.snapshots=%s\n", formats);
		free(formats);
	}
	if (repo->snapshot_prefix)
		fprintf(f, "repo.snapshot-prefix=%s\n", repo->snapshot_prefix);
	if (repo->max_stats != ctx.cfg.max_stats)
		fprintf(f, "repo.max-stats=%s\n", cgit_find_stats_periodname(repo->max_stats));
	if (repo->logo)
		fprintf(f, "repo.logo=%s\n", repo->logo);
	if (repo->logo_link)
		fprintf(f, "repo.logo-link=%s\n", repo->logo_link);
	fprintf(f, "repo.enable-remote-branches=%d\n", repo->enable_remote_branches);
	fprintf(f, "repo.enable-subject-links=%d\n", repo->enable_subject_links);
	fprintf(f, "repo.enable-commit-trailers=%d\n", repo->enable_commit_trailers);
	fprintf(f, "repo.enable-html-serving=%d\n", repo->enable_html_serving);
	if (repo->branch_sort != ctx.cfg.branch_sort)
		fprintf(f, "repo.branch-sort=%s\n", repo->branch_sort ? "age" : "name");
	if (repo->commit_sort == 1)
		fprintf(f, "repo.commit-sort=date\n");
	else if (repo->commit_sort == 2)
		fprintf(f, "repo.commit-sort=topo\n");
	fprintf(f, "repo.hide=%d\n", repo->hide);
	fprintf(f, "repo.ignore=%d\n", repo->ignore);
	fprintf(f, "\n");
}

static void print_repolist(FILE *f, struct cgit_repolist *list, int start)
{
	int i;

	for (i = start; i < list->count; i++)
		print_repo(f, &list->repos[i]);
}

static void parse_args(int argc, const char **argv)
{
	int i;
	const char *arg;
	int scanned = 0;

	for (i = 1; i < argc; i++) {
		if (!strcmp(argv[i], "--version")) {
			print_version();
			exit(0);
		}
		if (skip_prefix(argv[i], "--cache=", &arg)) {
			ctx.cfg.cache_root = xstrdup(arg);
		} else if (!strcmp(argv[i], "--nohttp")) {
			ctx.env.no_http = "1";
		} else if (skip_prefix(argv[i], "--query=", &arg)) {
			ctx.qry.raw = xstrdup(arg);
		} else if (
			skip_prefix(argv[i], "--scan-tree=", &arg) ||
			skip_prefix(argv[i], "--scan-path=", &arg)
		) {
			// A repository's snapshots setting is masked with the
			// global one, and cgitrc has not been read here, so an
			// empty mask would discard what the repository set.
			ctx.cfg.snapshots = INT_MAX;
			scanned++;
			scan_tree(arg);
		}
	}
	if (scanned) {
		QSORT(cgit_repolist.repos, cgit_repolist.count, cmp_repos);
		print_repolist(stdout, &cgit_repolist, 0);
		exit(0);
	}
}

/*
 * The lock is an flock rather than the mere existence of the lock file,
 * because a lock the kernel drops with its process cannot outlive a scan
 * killed mid-run. A leftover file would count as a scan forever in progress.
 */
static int generate_cached_repolist(const char *path, const char *cached_rc)
{
	struct strbuf locked_rc = STRBUF_INIT;
	int err = 0;
	int fd;
	int first;
	FILE *f;

	strbuf_addf(&locked_rc, "%s.lock", cached_rc);
	err = cache_lock_file(locked_rc.buf, &fd);
	if (err)
		goto out;
	if (!(f = fdopen(fd, "w"))) {
		err = errno;
		fprintf(stderr, "[cgit] Unable to write %s: %s (%d)\n", locked_rc.buf, strerror(err), err);
		unlink(locked_rc.buf);
		close(fd);
		goto out;
	}
	first = cgit_repolist.count;
	scan_tree(path);
	print_repolist(f, &cgit_repolist, first);
	// Flushed before the rename, because print_repolist writes through
	// stdio and a rename of the file would otherwise publish a repolist
	// that stops wherever the buffer happened to end.
	if (fflush(f) || ferror(f)) {
		err = errno;
		fprintf(stderr, "[cgit] Unable to write %s: %s (%d)\n", locked_rc.buf, strerror(err), err);
		unlink(locked_rc.buf);
		fclose(f);
		goto out;
	}
	if (rename(locked_rc.buf, cached_rc)) {
		err = errno;
		fprintf(stderr, "[cgit] Unable to rename %s to %s: %s (%d)\n",
			locked_rc.buf, cached_rc, strerror(err), err);
		unlink(locked_rc.buf);
	}
	// Closed after the rename so the lock is held until the fresh list is
	// in place, and nobody can truncate the file being renamed.
	fclose(f);
out:
	strbuf_release(&locked_rc);
	return err;
}

/*
 * A cached repolist is itself a config file, so these two call each other.
 */
static void apply_config(const char *name, const char *value);

/*
 * The cached list is read back as trusted config, so it is kept under a name
 * drawn from every setting that changes what the scan writes, from the scan
 * itself to the defaults cgit_add_repo hands each repository. Two sites
 * sharing a cache-root, or one that changed a setting, would otherwise read
 * a list scanned under other settings until it expired.
 */
static void build_scan_key(struct strbuf *key, const char *path)
{
	const char *texts[] = {
		path,
		ctx.cfg.project_list,
		ctx.cfg.strict_export,
		ctx.cfg.section,
		ctx.cfg.module_link,
		ctx.cfg.clone_url,
	};
	const int numbers[] = {
		ctx.cfg.scan_hidden_path,
		ctx.cfg.case_sensitive_sort,
		ctx.cfg.remove_suffix,
		ctx.cfg.section_from_path,
		ctx.cfg.enable_git_config,
		ctx.cfg.trust_scan_config,
		ctx.cfg.snapshots,
		ctx.cfg.max_stats,
		ctx.cfg.branch_sort,
		ctx.cfg.commit_sort,
		ctx.cfg.enable_blame,
		ctx.cfg.enable_commit_graph,
		ctx.cfg.enable_commit_trailers,
		ctx.cfg.enable_follow_links,
		ctx.cfg.enable_gitmodules_links,
		ctx.cfg.enable_html_serving,
		ctx.cfg.enable_log_filecount,
		ctx.cfg.enable_log_linecount,
		ctx.cfg.enable_mailmap,
		ctx.cfg.enable_remote_branches,
		ctx.cfg.enable_subject_links,
		// A filter kept from the global settings follows cgitrc when
		// the list is read, but whether one is set decides what is
		// written for a repository that turned it off.
		ctx.cfg.about_filter != NULL,
		ctx.cfg.commit_filter != NULL,
		ctx.cfg.email_filter != NULL,
		ctx.cfg.source_filter != NULL,
		ctx.cfg.trailer_filter != NULL,
	};
	size_t i;

	// Each text goes behind its length so no value can run into the next,
	// and an unset one is told apart from an empty one.
	for (i = 0; i < ARRAY_SIZE(texts); i++) {
		if (texts[i])
			strbuf_addf(key, "%zu|%s", strlen(texts[i]), texts[i]);
		else
			strbuf_addch(key, '-');
	}
	for (i = 0; i < ARRAY_SIZE(numbers); i++)
		strbuf_addf(key, "%d|", numbers[i]);
}

static void process_cached_repolist(const char *path)
{
	struct stat st;
	struct strbuf key = STRBUF_INIT;
	struct strbuf cached_rc = STRBUF_INIT;
	time_t age;
	int devnull, first, i;

	build_scan_key(&key, path);
	strbuf_addf(&cached_rc, "%s/rc-%8lx", ctx.cfg.cache_root, cache_hash_str(key.buf));
	strbuf_release(&key);

	if (stat(cached_rc.buf, &st)) {
		// Nothing is cached yet, so this request scans in its own
		// process, leaving no copy behind when it cannot take the lock.
		if (generate_cached_repolist(path, cached_rc.buf))
			scan_tree(path);
		goto out;
	}

	first = cgit_repolist.count;
	config_file_parse(cached_rc.buf, apply_config);
	// The scan that wrote the list read each description file already, so
	// a repository it left without one is not read again on every request.
	for (i = first; i < cgit_repolist.count; i++)
		cgit_repolist.repos[i].desc_read = 1;

	// A negative ttl never expires, the same as for a cached page.
	age = time(NULL) - st.st_mtime;
	if (ctx.cfg.cache_scan_ttl < 0 || age <= (ctx.cfg.cache_scan_ttl * 60))
		goto out;

	// The list just parsed is stale but usable, so a child rebuilds it
	// while this request answers from what it already has.
	if (fork())
		goto out;

	// The child inherits the request's descriptors, and the web server
	// reads stdout until every holder is gone, so left in place they
	// would keep the visitor waiting on the scan and let its output
	// land on the response.
	devnull = open("/dev/null", O_RDWR);
	if (devnull >= 0) {
		dup2(devnull, STDIN_FILENO);
		dup2(devnull, STDOUT_FILENO);
		dup2(devnull, STDERR_FILENO);
		if (devnull > STDERR_FILENO)
			close(devnull);
	}
	// _exit rather than exit, so the handlers the request registered do
	// not run a second time in the child.
	_exit(generate_cached_repolist(path, cached_rc.buf));
out:
	strbuf_release(&cached_rc);
}

/*
 * Unlike the cgitrc a scanned repository may carry, cgitrc and the files it
 * includes are meant to be there, and one that cannot be read would otherwise
 * leave nothing but an empty index to go on.
 */
static void parse_cgitrc(const char *filename)
{
	// Copied, since the name usually sits in the buffer macro expansion
	// reuses, which a line in the file can overwrite.
	char *file = xstrdup(filename);
	int err = config_file_parse(file, apply_config);

	if (err)
		fprintf(stderr, "[cgit] Unable to read config file %s: %s (%d)\n",
			file, strerror(err), err);
	free(file);
}

static void add_mimetype(const char *name, const char *value)
{
	struct string_list_item *item;

	item = string_list_insert(&ctx.cfg.mimetypes, name);
	item->util = xstrdup(value);
}

/*
 * Accept only the date formats cgit documents, leaving mode unchanged for
 * anything else. Git's parse_date_format dies on anything unknown, which would
 * turn a cgitrc typo into a failed request.
 */
static void parse_date_setting(const char *format, struct date_mode *mode)
{
	static const char * const names[] = {
		"default", "human", "iso", "iso-strict", "iso8601",
		"iso8601-strict", "raw", "relative", "rfc",
		"rfc2822", "short", "unix",
	};
	const char *rest;
	size_t i;

	if (skip_prefix(format, "format:", &rest)) {
		// An empty strftime format would render every date blank.
		if (*rest)
			parse_date_format(format, mode);
		return;
	}

	// git reads local only as a whole alias of default-local, and dies on
	// it with a suffix of its own.
	if (!strcmp(format, "local")) {
		parse_date_format(format, mode);
		return;
	}

	for (i = 0; i < ARRAY_SIZE(names); i++) {
		if (!skip_prefix(format, names[i], &rest))
			continue;
		// "iso" also prefixes "iso8601", so a partial hit is skipped.
		if (*rest && strcmp(rest, "-local"))
			continue;
		parse_date_format(format, mode);
		return;
	}
}

/*
 * A max-stats value that names no period turns the statistics page off, the
 * same as leaving it unset, so a typo is logged with where it was written.
 * An empty value and "0" turn the page off on purpose, and a cached repolist
 * writes the empty one, so those stay quiet.
 */
static int config_stats_period(const char *value)
{
	int period = cgit_find_stats_period(value, NULL);
	int len;

	if (period || !*value || !strcmp(value, "0"))
		return period;
	// The first line alone, since a git config value can hold a newline
	// that would forge a log line of its own.
	len = (int)strcspn(value, "\n");
	if (config_origin.file)
		fprintf(stderr, "[cgit] Turning statistics off for max-stats=%.*s in %s line %d, since it names no period\n",
			len, value, config_origin.file, config_origin.line);
	else
		fprintf(stderr, "[cgit] Turning statistics off for max-stats=%.*s, since it names no period\n",
			len, value);
	return period;
}

static void apply_config(const char *name, const char *value)
{
	const char *arg;

	if (!strcmp(name, "section"))
		ctx.cfg.section = cgit_strdup_first_line(value);
	else if (!strcmp(name, "repo.url"))
		ctx.repo = cgit_add_repo(value);
	else if (ctx.repo && !strcmp(name, "repo.path"))
		ctx.repo->path = cgit_trim_end(value, '/');
	else if (ctx.repo && skip_prefix(name, "repo.", &arg))
		cgit_repo_config(ctx.repo, arg, value);
	else if (!strcmp(name, "readme"))
		string_list_append(&ctx.cfg.readme, cgit_strdup_first_line(value));
	else if (!strcmp(name, "root-title"))
		ctx.cfg.root_title = cgit_strdup_first_line(value);
	else if (!strcmp(name, "root-desc"))
		ctx.cfg.root_desc = cgit_strdup_first_line(value);
	else if (!strcmp(name, "root-readme"))
		ctx.cfg.root_readme = cgit_strdup_first_line(value);
	else if (!strcmp(name, "css"))
		string_list_append(&ctx.cfg.css, cgit_strdup_first_line(value));
	else if (!strcmp(name, "js"))
		string_list_append(&ctx.cfg.js, cgit_strdup_first_line(value));
	else if (!strcmp(name, "favicon"))
		ctx.cfg.favicon = cgit_strdup_first_line(value);
	else if (!strcmp(name, "footer"))
		ctx.cfg.footer = cgit_strdup_first_line(value);
	else if (!strcmp(name, "head-include"))
		ctx.cfg.head_include = cgit_strdup_first_line(value);
	else if (!strcmp(name, "header"))
		ctx.cfg.header = cgit_strdup_first_line(value);
	else if (!strcmp(name, "logo"))
		ctx.cfg.logo = cgit_strdup_first_line(value);
	else if (!strcmp(name, "logo-link"))
		ctx.cfg.logo_link = cgit_strdup_first_line(value);
	else if (!strcmp(name, "module-link"))
		ctx.cfg.module_link = cgit_strdup_first_line(value);
	else if (!strcmp(name, "strict-export"))
		ctx.cfg.strict_export = cgit_strdup_first_line(value);
	else if (!strcmp(name, "virtual-root"))
		ctx.cfg.virtual_root = ensure_end(value, '/');
	else if (!strcmp(name, "enable-plain-email"))
		ctx.cfg.enable_plain_email = config_int(name, value);
	else if (!strcmp(name, "enable-header"))
		ctx.cfg.enable_header = config_int(name, value);
	else if (!strcmp(name, "snapshots"))
		ctx.cfg.snapshots = cgit_parse_snapshots_mask(value);
	else if (!strcmp(name, "trust-scan-config"))
		ctx.cfg.trust_scan_config = config_int(name, value);
	else if (!strcmp(name, "enable-follow-links"))
		ctx.cfg.enable_follow_links = config_int(name, value);
	else if (!strcmp(name, "enable-http-clone"))
		ctx.cfg.enable_http_clone = config_int(name, value);
	else if (!strcmp(name, "enable-index-links"))
		ctx.cfg.enable_index_links = config_int(name, value);
	else if (!strcmp(name, "enable-index-owner"))
		ctx.cfg.enable_index_owner = config_int(name, value);
	else if (!strcmp(name, "enable-blame"))
		ctx.cfg.enable_blame = config_int(name, value);
	else if (!strcmp(name, "enable-commit-graph"))
		ctx.cfg.enable_commit_graph = config_int(name, value);
	else if (!strcmp(name, "enable-gitmodules-links"))
		ctx.cfg.enable_gitmodules_links = config_int(name, value);
	else if (!strcmp(name, "enable-log-filecount"))
		ctx.cfg.enable_log_filecount = config_int(name, value);
	else if (!strcmp(name, "enable-log-linecount"))
		ctx.cfg.enable_log_linecount = config_int(name, value);
	else if (!strcmp(name, "enable-mailmap"))
		ctx.cfg.enable_mailmap = config_int(name, value);
	else if (!strcmp(name, "enable-relative-dates"))
		ctx.cfg.enable_relative_dates = config_int(name, value);
	else if (!strcmp(name, "enable-remote-branches"))
		ctx.cfg.enable_remote_branches = config_int(name, value);
	else if (!strcmp(name, "enable-subject-links"))
		ctx.cfg.enable_subject_links = config_int(name, value);
	else if (!strcmp(name, "enable-commit-trailers"))
		ctx.cfg.enable_commit_trailers = config_int(name, value);
	else if (!strcmp(name, "enable-html-serving"))
		ctx.cfg.enable_html_serving = config_int(name, value);
	else if (!strcmp(name, "enable-tree-linenumbers"))
		ctx.cfg.enable_tree_linenumbers = config_int(name, value);
	else if (!strcmp(name, "enable-version-info"))
		ctx.cfg.enable_version_info = config_int(name, value);
	else if (!strcmp(name, "enable-tree-group-dirs"))
		ctx.cfg.enable_tree_group_dirs = config_int(name, value);
	else if (!strcmp(name, "enable-git-config"))
		ctx.cfg.enable_git_config = config_int(name, value);
	else if (!strcmp(name, "enable-cache-list"))
		ctx.cfg.enable_cache_list = config_int(name, value);
	else if (!strcmp(name, "max-stats"))
		ctx.cfg.max_stats = config_stats_period(value);
	else if (!strcmp(name, "cache-size"))
		ctx.cfg.cache_size = config_int(name, value);
	else if (!strcmp(name, "cache-root"))
		ctx.cfg.cache_root = cgit_strdup_first_line(cgit_expand_macros(value));
	else if (!strcmp(name, "cache-index-ttl"))
		ctx.cfg.cache_index_ttl = config_int(name, value);
	else if (!strcmp(name, "cache-summary-ttl"))
		ctx.cfg.cache_summary_ttl = config_int(name, value);
	else if (!strcmp(name, "cache-scan-ttl"))
		ctx.cfg.cache_scan_ttl = config_int(name, value);
	else if (!strcmp(name, "cache-static-ttl"))
		ctx.cfg.cache_static_ttl = config_int(name, value);
	else if (!strcmp(name, "cache-dynamic-ttl"))
		ctx.cfg.cache_dynamic_ttl = config_int(name, value);
	else if (!strcmp(name, "cache-about-ttl"))
		ctx.cfg.cache_about_ttl = config_int(name, value);
	else if (!strcmp(name, "cache-snapshot-ttl"))
		ctx.cfg.cache_snapshot_ttl = config_int(name, value);
	else if (!strcmp(name, "cache-max-slot-size"))
		ctx.cfg.cache_max_slot_size = config_int(name, value);
	else if (!strcmp(name, "case-sensitive-sort"))
		ctx.cfg.case_sensitive_sort = config_int(name, value);
	else if (!strcmp(name, "about-filter"))
		ctx.cfg.about_filter = cgit_new_filter(value, ABOUT);
	else if (!strcmp(name, "commit-filter"))
		ctx.cfg.commit_filter = cgit_new_filter(value, COMMIT);
	else if (!strcmp(name, "email-filter"))
		ctx.cfg.email_filter = cgit_new_filter(value, EMAIL);
	else if (!strcmp(name, "trailer-filter"))
		ctx.cfg.trailer_filter = cgit_new_filter(value, TRAILER);
	else if (!strcmp(name, "auth-filter"))
		ctx.cfg.auth_filter = cgit_new_filter(value, AUTH);
	else if (!strcmp(name, "embedded"))
		ctx.cfg.embedded = config_int(name, value);
	else if (!strcmp(name, "max-atom-items"))
		ctx.cfg.max_atom_items = config_int(name, value);
	else if (!strcmp(name, "max-message-length"))
		ctx.cfg.max_msg_len = config_int(name, value);
	else if (!strcmp(name, "max-repodesc-length"))
		ctx.cfg.max_repodesc_len = config_int(name, value);
	else if (!strcmp(name, "max-blob-size"))
		ctx.cfg.max_blob_size = config_int(name, value);
	else if (!strcmp(name, "max-diff-files"))
		ctx.cfg.max_diff_files = config_int(name, value);
	else if (!strcmp(name, "max-diff-lines"))
		ctx.cfg.max_diff_lines = config_int(name, value);
	else if (!strcmp(name, "max-ref-count"))
		ctx.cfg.max_ref_count = config_int(name, value);
	else if (!strcmp(name, "max-repo-count")) {
		ctx.cfg.max_repo_count = config_int(name, value);
		if (ctx.cfg.max_repo_count <= 0)
			ctx.cfg.max_repo_count = INT_MAX;
	} else if (!strcmp(name, "max-commit-count"))
		ctx.cfg.max_commit_count = config_int(name, value);
	else if (!strcmp(name, "max-patch-count"))
		ctx.cfg.max_patch_count = config_int(name, value);
	else if (!strcmp(name, "project-list"))
		ctx.cfg.project_list = cgit_strdup_first_line(cgit_expand_macros(value));
	else if (!strcmp(name, "scan-path")) {
		if (ctx.cfg.cache_size > 0)
			process_cached_repolist(cgit_expand_macros(value));
		else
			scan_tree(cgit_expand_macros(value));
		// The scan grows the repository array, so a record taken before
		// it may have moved, and a repo key after it belongs to nothing.
		ctx.repo = NULL;
	} else if (!strcmp(name, "scan-hidden-path"))
		ctx.cfg.scan_hidden_path = config_int(name, value);
	else if (!strcmp(name, "section-from-path"))
		ctx.cfg.section_from_path = config_int(name, value);
	else if (!strcmp(name, "repository-sort"))
		ctx.cfg.repository_sort = cgit_strdup_first_line(value);
	else if (!strcmp(name, "section-sort"))
		ctx.cfg.section_sort = config_int(name, value);
	else if (!strcmp(name, "source-filter"))
		ctx.cfg.source_filter = cgit_new_filter(value, SOURCE);
	else if (!strcmp(name, "summary-log"))
		ctx.cfg.summary_log = config_int(name, value);
	else if (!strcmp(name, "summary-branches"))
		ctx.cfg.summary_branches = config_int(name, value);
	else if (!strcmp(name, "summary-tags"))
		ctx.cfg.summary_tags = config_int(name, value);
	else if (!strcmp(name, "side-by-side-diffs"))
		ctx.cfg.difftype = config_int(name, value) ? DIFF_SSDIFF : DIFF_UNIFIED;
	else if (!strcmp(name, "age-file"))
		ctx.cfg.agefile = cgit_strdup_first_line(value);
	else if (!strcmp(name, "mimetype-file"))
		ctx.cfg.mimetype_file = cgit_strdup_first_line(value);
	else if (!strcmp(name, "rename-limit"))
		ctx.cfg.renamelimit = config_int(name, value);
	else if (!strcmp(name, "remove-suffix"))
		ctx.cfg.remove_suffix = config_int(name, value);
	else if (!strcmp(name, "robots"))
		ctx.cfg.robots = cgit_strdup_first_line(value);
	else if (!strcmp(name, "clone-prefix"))
		ctx.cfg.clone_prefix = cgit_strdup_first_line(value);
	else if (!strcmp(name, "clone-url"))
		ctx.cfg.clone_url = cgit_strdup_first_line(value);
	else if (!strcmp(name, "local-time"))
		ctx.cfg.local_time = config_int(name, value);
	else if (!strcmp(name, "date-format"))
		parse_date_setting(value, &ctx.cfg.date_mode);
	else if (!strcmp(name, "commit-sort")) {
		if (!strcmp(value, "date"))
			ctx.cfg.commit_sort = 1;
		if (!strcmp(value, "topo"))
			ctx.cfg.commit_sort = 2;
	} else if (!strcmp(name, "branch-sort")) {
		if (!strcmp(value, "age"))
			ctx.cfg.branch_sort = 1;
		if (!strcmp(value, "name"))
			ctx.cfg.branch_sort = 0;
	} else if (skip_prefix(name, "mimetype.", &arg))
		add_mimetype(arg, value);
	else if (!strcmp(name, "include"))
		parse_cgitrc(cgit_expand_macros(value));
	else if (skip_prefix(name, "repo.", &arg))
		fprintf(stderr, "[cgit] Ignoring %s before any repo.url\n", name);
	else
		fprintf(stderr, "[cgit] Unknown config key: %s\n", name);
}

/*
 * Read a whole number a request supplied, clamped into the range the caller
 * accepts. strtol rather than atoi, because atoi is undefined on overflow
 * and every value here arrives straight from the query string.
 */
static int query_int(const char *value, int min, int max)
{
	long number = strtol(value, NULL, 10);

	if (number < min)
		return min;
	if (number > max)
		return max;
	return number;
}

static void apply_query_param(const char *name, const char *value)
{
	if (!strcmp(name, "r")) {
		ctx.qry.repo = xstrdup(value);
		ctx.repo = cgit_get_repoinfo(value);
	} else if (!strcmp(name, "p")) {
		ctx.qry.page = xstrdup(value);
	} else if (!strcmp(name, "url")) {
		// Every leading slash goes, not just one, so a value like
		// //example.com cannot survive as /example.com and make the
		// virtual root join a scheme-relative link to another host.
		while (*value == '/')
			value++;
		ctx.qry.url = xstrdup(value);
		cgit_parse_url(value);
	} else if (!strcmp(name, "qt")) {
		ctx.qry.grep = xstrdup(value);
	} else if (!strcmp(name, "q")) {
		ctx.qry.search = xstrndup(value, MAX_SEARCH_LEN);
	} else if (!strcmp(name, "h")) {
		ctx.qry.head = xstrdup(value);
	} else if (!strcmp(name, "id")) {
		ctx.qry.oid = xstrdup(value);
		ctx.qry.has_oid = 1;
	} else if (!strcmp(name, "id2")) {
		ctx.qry.oid2 = xstrdup(value);
		ctx.qry.has_oid = 1;
	} else if (!strcmp(name, "ofs")) {
		// Bounded above so a crafted value cannot force a walk over
		// the whole history. The floor is -1 rather than 0 because
		// the stats page submits -1 for all authors.
		ctx.qry.ofs = query_int(value, -1, MAX_QUERY_OFFSET);
	} else if (!strcmp(name, "path")) {
		ctx.qry.path = cgit_trim_end(value, '/');
	} else if (!strcmp(name, "s")) {
		ctx.qry.sort = xstrdup(value);
	} else if (!strcmp(name, "showmsg")) {
		ctx.qry.showmsg = query_int(value, INT_MIN, INT_MAX);
	} else if (!strcmp(name, "period")) {
		ctx.qry.period = xstrdup(value);
	} else if (!strcmp(name, "dt")) {
		ctx.qry.difftype = query_int(value, INT_MIN, INT_MAX);
		ctx.qry.has_difftype = 1;
	} else if (!strcmp(name, "ss")) {
		// No longer generated, but old links still carry it.
		ctx.qry.difftype = query_int(value, INT_MIN, INT_MAX) ? DIFF_SSDIFF : DIFF_UNIFIED;
		ctx.qry.has_difftype = 1;
	} else if (!strcmp(name, "all")) {
		ctx.qry.show_all = query_int(value, INT_MIN, INT_MAX);
	} else if (!strcmp(name, "context")) {
		// Context lines are not counted against max-diff-lines, so an
		// unbounded width turns a whole blob into context and renders
		// it in full however small the change was.
		ctx.qry.context = query_int(value, 0, MAX_DIFF_CONTEXT_LINES);
	} else if (!strcmp(name, "ignorews")) {
		ctx.qry.ignorews = query_int(value, INT_MIN, INT_MAX);
	} else if (!strcmp(name, "follow")) {
		ctx.qry.follow = query_int(value, INT_MIN, INT_MAX);
	}
}

static void http_parse_querystring(const char *txt, void (*fn)(const char *name, const char *value))
{
	const char *p = txt;

	while (p && *p) {
		char *name = url_decode_parameter_name(&p);
		// A name ends at an equals sign or an ampersand, and only the
		// first is followed by a value.
		if (*name) {
			char *value = p[-1] == '=' ? url_decode_parameter_value(&p) : xstrdup("");
			fn(name, value);
			free(value);
		}
		free(name);
	}
}

static void open_auth_filter(const char *action)
{
	cgit_open_filter(ctx.cfg.auth_filter, action,
		ctx.env.http_cookie ? ctx.env.http_cookie : "",
		ctx.env.request_method ? ctx.env.request_method : "",
		ctx.env.query_string ? ctx.env.query_string : "",
		ctx.env.http_referer ? ctx.env.http_referer : "",
		ctx.env.path_info ? ctx.env.path_info : "",
		ctx.env.http_host ? ctx.env.http_host : "",
		ctx.env.https ? ctx.env.https : "",
		ctx.qry.repo ? ctx.qry.repo : "",
		ctx.qry.page ? ctx.qry.page : "",
		cgit_currentfullurl(),
		cgit_loginurl());
}

/*
 * The filter answers the login POST itself, writing the status line and every
 * header, so nothing here prints any and the process ends before this returns.
 */
static NORETURN void authenticate_post(void)
{
	char buffer[MAX_AUTHENTICATION_POST_BYTES];
	size_t len;
	ssize_t got;

	open_auth_filter("authenticate-post");
	len = ctx.env.content_length;
	if (len > MAX_AUTHENTICATION_POST_BYTES)
		len = MAX_AUTHENTICATION_POST_BYTES;
	// The body can arrive in more than one write, so read until it is
	// all there or the server closes the stream.
	if ((got = read_in_full(STDIN_FILENO, buffer, len)) < 0)
		die_errno("Unable to read the POST body");
	if (write_in_full(STDOUT_FILENO, buffer, got) < 0)
		die_errno("Unable to pass the POST body to the auth filter");
	cgit_close_filter(ctx.cfg.auth_filter);
	exit(0);
}

static void authenticate_cookie(void)
{
	if (!ctx.cfg.auth_filter) {
		ctx.env.authenticated = 1;
		return;
	}

	if (
		ctx.env.request_method                  &&
		ctx.qry.page                            &&
		!ctx.repo                               &&
		!strcmp(ctx.env.request_method, "POST") &&
		!strcmp(ctx.qry.page, "login")
	)
		authenticate_post();

	open_auth_filter("authenticate-cookie");
	ctx.env.authenticated = cgit_close_filter(ctx.cfg.auth_filter);
}

/*
 * Only a full object id names content that can never change. The id
 * parameter accepts anything git can resolve, so a page pinned to a ref
 * name or abbreviation must not be cached under the static ttl.
 */
static int is_full_oid(const char *rev)
{
	size_t len = strlen(rev);

	if (len != GIT_SHA1_HEXSZ && len != GIT_SHA256_HEXSZ)
		return 0;
	for (; *rev; rev++) {
		if (!isxdigit((unsigned char)*rev))
			return 0;
	}
	return 1;
}

/*
 * Every cache-*-ttl setting is written in minutes, and so is this.
 */
static int calc_ttl(void)
{
	const struct cgit_cmd *cmd;

	if (!ctx.repo)
		return ctx.cfg.cache_index_ttl;

	if (!ctx.qry.page || !strcmp(ctx.qry.page, "summary"))
		return ctx.cfg.cache_summary_ttl;

	// The dumb transport serves files that already sit on the disk, so a
	// copy in a slot would cost that disk twice and gain nothing.
	cmd = cgit_find_cmd(ctx.qry.page);
	if (cmd && cmd->is_clone)
		return 0;

	if (!strcmp(ctx.qry.page, "about"))
		return ctx.cfg.cache_about_ttl;

	// Checked ahead of the snapshot ttl, because a tarball pinned to an
	// object id can never come out differently and is the most expensive
	// page to rebuild.
	if (
		ctx.qry.has_oid                            &&
		cmd                                        &&
		cmd->is_static                             &&
		(!ctx.qry.oid || is_full_oid(ctx.qry.oid)) &&
		(!ctx.qry.oid2 || is_full_oid(ctx.qry.oid2))
	)
		return ctx.cfg.cache_static_ttl;

	if (!strcmp(ctx.qry.page, "snapshot"))
		return ctx.cfg.cache_snapshot_ttl;

	return ctx.cfg.cache_dynamic_ttl;
}

/*
 * The scheme and host are folded in because the page's absolute urls, its
 * clone urls and atom links, are built from them, so a spoofed Host must
 * not poison the page served on the real one.
 */
static void build_cache_key(struct strbuf *key)
{
	char *hosturl = cgit_hosturl();
	const char *parts[] = {
		cgit_httpscheme(),
		hosturl ? hosturl : "",
		ctx.env.path_info ? ctx.env.path_info : "",
		ctx.env.query_string ? ctx.env.query_string : "",
	};
	size_t i;

	// Each part is written behind its own length so no value can make two
	// requests spell one key, and the path and query come from the
	// environment because cgit's rebuilt query string folds them together.
	for (i = 0; i < ARRAY_SIZE(parts); i++)
		strbuf_addf(key, "%zu|%s", strlen(parts[i]), parts[i]);
	free(hosturl);
}

/*
 * Returning non-zero ends git's walk, so the search stops at the first hit.
 */
static int find_current_ref(const struct reference *ref, void *data)
{
	struct refmatch *match = data;

	if (!strcmp(ref->name, match->wanted))
		match->found = 1;
	// The fallback has to pass the check every request makes on its head,
	// or one branch named with a leading dash takes the repository offline.
	if (!match->first && ref->name[0] != '-')
		match->first = xstrdup(ref->name);
	return match->found;
}

static char *find_default_branch(struct cgit_repo *repo)
{
	struct refmatch match = { .wanted = repo->defbranch };

	refs_for_each_branch_ref(get_main_ref_store(the_repository), find_current_ref, &match);
	if (!match.found)
		return match.first;
	free(match.first);
	return xstrdup(match.wanted);
}

static char *guess_defbranch(void)
{
	const char *ref, *refname;
	struct object_id oid;

	ref = refs_resolve_ref_unsafe(get_main_ref_store(the_repository), "HEAD", 0, &oid, NULL);
	if (!ref || !skip_prefix(ref, "refs/heads/", &refname))
		return xstrdup("master");
	return xstrdup(refname);
}

/*
 * Split one readme setting into the file it names and the ref that file is
 * read from, leaving the ref NULL for a file on disk. The caller frees both.
 */
static void parse_readme(const char *readme, char **filename, char **ref, struct cgit_repo *repo)
{
	const char *colon;

	*filename = NULL;
	*ref = NULL;

	if (!readme || !readme[0])
		return;

	if (cgit_readme_is_tracked(readme)) {
		colon = strchr(readme, ':');
		if (colon == readme && ctx.qry.head)
			*ref = xstrdup(ctx.qry.head);
		else if (colon == readme && repo->defbranch)
			*ref = xstrdup(repo->defbranch);
		else
			*ref = xstrndup(readme, colon - readme);
		readme = colon + 1;
	}

	if (!(*ref) && readme[0] != '/')
		*filename = xstrfmt("%s/%s", repo->path, readme);
	else
		*filename = xstrdup(readme);
}

static void choose_readme(struct cgit_repo *repo)
{
	int found;
	char *filename = NULL, *ref = NULL;
	struct string_list *list;
	struct string_list_item *entry;

	// A repository without readme settings of its own follows the global
	// ones, which it must not free.
	list = repo->readme.nr ? &repo->readme : &ctx.cfg.readme;
	if (!list->nr)
		return;

	found = 0;
	for_each_string_list_item(entry, list) {
		parse_readme(entry->string, &filename, &ref, repo);
		if (!filename) {
			free(ref);
			continue;
		}
		if (ref) {
			if (cgit_ref_path_exists(filename, ref)) {
				found = 1;
				break;
			}
		} else if (!access(filename, R_OK)) {
			found = 1;
			break;
		}
		free(filename);
		free(ref);
	}
	repo->readme.strdup_strings = 1;
	string_list_clear(&repo->readme, 0);
	repo->readme.strdup_strings = 0;
	if (found)
		string_list_append(&repo->readme, filename)->util = ref;
}

/*
 * Export the CGIT_REPO_* variables that filters and hooks read.
 */
static void prepare_repo_env(struct cgit_repo *repo)
{
	const struct {
		const char *name;
		const char *value;
	} vars[] = {
		{ .name = "CGIT_REPO_URL", .value = repo->url },
		{ .name = "CGIT_REPO_NAME", .value = repo->name },
		{ .name = "CGIT_REPO_PATH", .value = repo->path },
		{ .name = "CGIT_REPO_OWNER", .value = repo->owner },
		{ .name = "CGIT_REPO_DEFBRANCH", .value = repo->defbranch },
		{ .name = "CGIT_REPO_SECTION", .value = repo->section },
		{ .name = "CGIT_REPO_CLONE_URL", .value = repo->clone_url }
	};
	size_t i;

	for (i = 0; i < ARRAY_SIZE(vars); i++)
		if (vars[i].value && setenv(vars[i].name, vars[i].value, 1))
			fprintf(stderr, "[cgit] Unable to set %s=%s: %s (%d)\n",
				vars[i].name, vars[i].value, strerror(errno), errno);
}

static void open_repo(int *nongit, int *err)
{
	static int reported;

	// A repository configured without a path has nothing to open.
	if (!ctx.repo->path) {
		*nongit = 1;
		*err = 0;
		return;
	}
	setenv("GIT_DIR", ctx.repo->path, 1);

	errno = 0;
	setup_git_directory_gently(the_repository, nongit);
	*err = errno;
	// The error page names the repository alone, so the path goes to the
	// log, and only once, since a failed cache fill renders a page again.
	if (*nongit && !reported) {
		reported = 1;
		fprintf(stderr, "[cgit] Unable to open repository %s at %s: %s\n", ctx.repo->url,
			ctx.repo->path, *err ? strerror(*err) : "Not a valid git repository");
	}
	// Notes come out of the object store, which a repository that failed
	// to open has none of. The commit and log pages always ask for them,
	// so they are set up either way, but a default ref that names no tree
	// is left out, since git dies loading one.
	if (!*nongit) {
		struct display_notes_opt opt;
		char *ref = default_notes_ref(the_repository);

		init_display_notes(&opt);
		opt.use_default_notes = cgit_notes_ref_usable(ref);
		load_display_notes(&opt);
		release_display_notes(&opt);
		free(ref);
	}
}

/*
 * Returns non-zero once it has written a complete response of its own, in
 * which case the caller must not render a page over the top of it.
 */
static int prepare_repo_cmd(int nongit, int err, int clone)
{
	// Static, since the pages replace the title without freeing it. Reset
	// on entry, since a failed cache fill renders the page a second time.
	static struct strbuf title = STRBUF_INIT;
	struct object_id oid;

	strbuf_reset(&title);
	if (nongit) {
		const char *name = ctx.repo->name;
		ctx.repo = NULL;
		cgit_print_error_page(
			404, "Not Found", "Failed to open %s: %s", name,
			err ? strerror(err) : "Not a valid git repository"
		);
		return 1;
	}
	strbuf_addf(&title, "%s - %s", ctx.repo->name, ctx.cfg.root_title);
	ctx.page.title = title.buf;
	cgit_read_repo_desc(ctx.repo);

	if (!ctx.repo->defbranch)
		ctx.repo->defbranch = guess_defbranch();

	if (!ctx.qry.head) {
		ctx.qry.nohead = 1;
		ctx.qry.head = find_default_branch(ctx.repo);
	}

	if (!ctx.qry.head) {
		// The dumb transport serves refs and objects off the disk and
		// asks nothing of the head, and an empty repository clones.
		if (clone) {
			prepare_repo_env(ctx.repo);
			return 0;
		}
		ctx.empty_repo = 1;
		// Before the document starts, since the <head> carries
		// <link rel='vcs-git'> and those clone urls expand macros such
		// as $CGIT_REPO_URL out of this environment.
		prepare_repo_env(ctx.repo);
		// A feed reader cannot read the page below, so the feed goes
		// out with no entries.
		if (!strcmp(ctx.qry.page, "atom"))
			return 0;
		cgit_print_layout_start();
		cgit_print_empty_repo();
		cgit_print_layout_end();
		return 1;
	}

	// Every revision the request names is checked before git sees it,
	// see cgit_valid_rev, and the head has to resolve as well.
	if (!cgit_valid_rev(ctx.qry.head) || repo_get_oid(the_repository, ctx.qry.head, &oid)) {
		char *old_head = ctx.qry.head;
		ctx.qry.head = xstrdup(ctx.repo->defbranch);
		cgit_print_error_page(404, "Not Found", "Invalid branch: %s", old_head);
		free(old_head);
		return 1;
	}
	if (
		(ctx.qry.oid && !cgit_valid_rev(ctx.qry.oid)) ||
		(ctx.qry.oid2 && !cgit_valid_rev(ctx.qry.oid2))
	) {
		cgit_print_error_page(400, "Bad Request", "Invalid revision");
		return 1;
	}
	string_list_sort(&ctx.repo->submodules);
	prepare_repo_env(ctx.repo);
	choose_readme(ctx.repo);
	return 0;
}

static void process_request(void)
{
	const struct cgit_cmd *cmd;
	int nongit = 0, err = 0;

	// An unauthenticated request is answered with the filter's own body
	// whatever page it asked for.
	if (!ctx.env.authenticated) {
		ctx.page.title = "Authentication Required";
		cgit_print_layout_start();
		open_auth_filter("body");
		cgit_close_filter(ctx.cfg.auth_filter);
		cgit_print_layout_end();
		return;
	}

	if (ctx.repo)
		open_repo(&nongit, &err);

	cmd = cgit_get_cmd();
	if (!cmd || (!ctx.cfg.enable_http_clone && cmd->is_clone)) {
		// The error page carries the repository's own header, whose
		// branch switcher needs the head resolved first.
		if (ctx.repo && prepare_repo_cmd(nongit, err, 0))
			return;
		cgit_print_error_page(404, "Not Found", "Invalid request");
		return;
	}

	if (cmd->want_repo && !ctx.repo) {
		cgit_print_error_page(400, "Bad Request", "No repository selected");
		return;
	}

	ctx.qry.vpath = cmd->want_vpath ? ctx.qry.path : NULL;

	if (ctx.repo && prepare_repo_cmd(nongit, err, cmd->is_clone))
		return;

	cmd->fn();
}

void cgit_repo_config(struct cgit_repo *repo, const char *name, const char *value)
{
	const char *path;
	struct string_list_item *item;

	if (!strcmp(name, "name"))
		repo->name = cgit_strdup_first_line(value);
	else if (!strcmp(name, "clone-url"))
		repo->clone_url = cgit_strdup_first_line(value);
	else if (!strcmp(name, "desc"))
		repo->desc = cgit_strdup_first_line(value);
	else if (!strcmp(name, "owner"))
		repo->owner = cgit_strdup_first_line(value);
	else if (!strcmp(name, "defbranch"))
		repo->defbranch = cgit_strdup_first_line(value);
	else if (!strcmp(name, "head-content"))
		repo->head_content = cgit_strdup_first_line(value);
	else if (!strcmp(name, "snapshots"))
		repo->snapshots = ctx.cfg.snapshots & cgit_parse_snapshots_mask(value);
	else if (!strcmp(name, "enable-blame"))
		repo->enable_blame = config_int(name, value);
	else if (!strcmp(name, "enable-commit-graph"))
		repo->enable_commit_graph = config_int(name, value);
	else if (!strcmp(name, "enable-follow-links"))
		repo->enable_follow_links = config_int(name, value);
	else if (!strcmp(name, "enable-gitmodules-links"))
		repo->enable_gitmodules_links = config_int(name, value);
	else if (!strcmp(name, "enable-log-filecount"))
		repo->enable_log_filecount = config_int(name, value);
	else if (!strcmp(name, "enable-log-linecount"))
		repo->enable_log_linecount = config_int(name, value);
	else if (!strcmp(name, "enable-mailmap"))
		repo->enable_mailmap = config_int(name, value);
	else if (!strcmp(name, "enable-remote-branches"))
		repo->enable_remote_branches = config_int(name, value);
	else if (!strcmp(name, "enable-subject-links"))
		repo->enable_subject_links = config_int(name, value);
	else if (!strcmp(name, "enable-commit-trailers"))
		repo->enable_commit_trailers = config_int(name, value);
	else if (!strcmp(name, "enable-html-serving"))
		repo->enable_html_serving = config_int(name, value);
	else if (!strcmp(name, "branch-sort")) {
		if (!strcmp(value, "age"))
			repo->branch_sort = 1;
		if (!strcmp(value, "name"))
			repo->branch_sort = 0;
	} else if (!strcmp(name, "commit-sort")) {
		if (!strcmp(value, "date"))
			repo->commit_sort = 1;
		if (!strcmp(value, "topo"))
			repo->commit_sort = 2;
	} else if (!strcmp(name, "max-stats"))
		repo->max_stats = config_stats_period(value);
	else if (!strcmp(name, "module-link"))
		repo->module_link = cgit_strdup_first_line(value);
	else if (skip_prefix(name, "module-link.", &path)) {
		item = string_list_append(&repo->submodules, cgit_strdup_first_line(path));
		item->util = cgit_strdup_first_line(value);
	} else if (!strcmp(name, "section"))
		repo->section = cgit_strdup_first_line(value);
	else if (!strcmp(name, "snapshot-prefix"))
		repo->snapshot_prefix = cgit_strdup_first_line(value);
	else if (!strcmp(name, "readme"))
		string_list_append(&repo->readme, cgit_strdup_first_line(value));
	else if (!strcmp(name, "logo"))
		repo->logo = cgit_strdup_first_line(value);
	else if (!strcmp(name, "logo-link"))
		repo->logo_link = cgit_strdup_first_line(value);
	else if (!strcmp(name, "hide"))
		repo->hide = config_int(name, value);
	else if (!strcmp(name, "ignore"))
		repo->ignore = config_int(name, value);
	else if (!strcmp(name, "about-filter"))
		repo->about_filter = cgit_new_filter(value, ABOUT);
	else if (!strcmp(name, "commit-filter"))
		repo->commit_filter = cgit_new_filter(value, COMMIT);
	else if (!strcmp(name, "source-filter"))
		repo->source_filter = cgit_new_filter(value, SOURCE);
	else if (!strcmp(name, "email-filter"))
		repo->email_filter = cgit_new_filter(value, EMAIL);
	else if (!strcmp(name, "trailer-filter"))
		repo->trailer_filter = cgit_new_filter(value, TRAILER);
	else
		fprintf(stderr, "[cgit] Unknown repo config key: %s\n", name);
}

int cmd_main(int argc, const char **argv)
{
	struct strbuf cache_key = STRBUF_INIT;
	const char *path;
	size_t max_bytes = 0;
	int err, ttl;

	isolate_git_environment();
	// Ignored, a filter that exits early makes the write fail with EPIPE
	// and the error page reach the visitor, instead of ending cgit.
	signal(SIGPIPE, SIG_IGN);
	cgit_init_filters();
	atexit(cgit_cleanup_filters);
	// Registered second so it runs first, since exit is reached from error
	// paths and from the HEAD shortcut with a page still buffered.
	atexit(html_flush);
	set_die_routine(die_routine);

	prepare_context();

	parse_args(argc, argv);
	parse_cgitrc(cgit_expand_macros(ctx.env.cgit_config));
	ctx.repo = NULL;
	http_parse_querystring(ctx.qry.raw, apply_query_param);

	// Falling back to PATH_INFO lets cgit serve virtual urls without a
	// rewrite rule in the web server, and folding it into the raw query
	// string keeps it part of the cache key.
	path = ctx.env.path_info;
	if (!ctx.qry.url && path) {
		// Stripped like the url parameter and for the same reason.
		while (*path == '/')
			path++;
		ctx.qry.url = xstrdup(path);
		if (ctx.qry.raw) {
			char *path_and_query = xstrfmt("%s?%s", path, ctx.qry.raw);
			free(ctx.qry.raw);
			ctx.qry.raw = path_and_query;
		} else {
			ctx.qry.raw = xstrdup(ctx.qry.url);
		}
		cgit_parse_url(ctx.qry.url);
	}

	authenticate_cookie();

	ttl = calc_ttl();
	// An unauthenticated request gets a body meant for one visitor, and a
	// HEAD request stops after the headers.
	if (!ctx.env.authenticated || (ctx.env.request_method && !strcmp(ctx.env.request_method, "HEAD")))
		ctx.cfg.cache_size = 0;

	// Written in kilobytes in cgitrc, where zero and below lift the bound.
	if (ctx.cfg.cache_max_slot_size > 0)
		max_bytes = (size_t)ctx.cfg.cache_max_slot_size * 1024;
	build_cache_key(&cache_key);
	err = cache_process(ctx.cfg.cache_size, ctx.cfg.cache_root, cache_key.buf, ttl,
		max_bytes, process_request);
	strbuf_release(&cache_key);

	// A slot that could not be sent because the client went away is not
	// worth an error page nobody will read.
	if (err && err != EPIPE)
		cgit_print_error("Error processing page: %s (%d)", strerror(err), err);
	return err;
}