Add support for configuring cgroups and their settings on services

This patch adds support for modifying settings for the default cgroups;
init, user, and system, as well as adding up to a total of eight groups
for the system.

Services can now be assigned to a cgroup, with optional extra settings
for that particular process group.  The syntax is slightly contrivied
but follows the overall Finit syntax of prop:value,prop':value', e.g.

   cgroup maint cpu.weight:123,mem.max:10000

Starting with the introduction of rlimits, a group of services sharing
the same .conf file can share the same (locally "global") rlimits, and
now also the same cgroup, e.g.

    cgroup.maint
    service foo
    service bar cgroup:mem.max:1000

This puts foo and bar in the same top-level cgroup 'maint', with an
extra memory restriction on bar for max 1000 bytes memory.

NOTE: 'mem.' is a Finit extension, a shorthand for cgroups2 'memory.'

Signed-off-by: Joachim Wiberg <troglobit@gmail.com>

contrib/finit.conf

@@ -8,6 +8,18 @@ module evdev
 module loop
 module psmouse
 
+# Top-level cgroups and their default settings.  All groups mandatory
+# but more can be added, max 8 groups in total currently.  The cgroup
+# 'root' is also available, reserved for RT processes.  Settings are
+# as-is, only one shorthand 'mem.' exists, other than that it's the
+# cgroup v2 controller default names.
+cgroup init   cpu.weight:100
+cgroup user   cpu.weight:100
+cgroup system cpu.weight:9700 cpu.max:50000
+
+# Example extra cgroup
+cgroup maint  cpu.weight:100
+
 # Runlevel to start after bootstrap, runlevel 'S'
 runlevel 2
 
@@ -16,7 +28,7 @@ network service networking start
 
 # Max file size for each log file: 100 kiB, rotate max 4 copies:
 # log => log.1 => log.2.gz => log.3.gz => log.4.gz
-log size=100k count=4
+log size:100k count:4
 
 # Virtual consoles to start getty on
 tty /dev/tty1
@@ -30,9 +42,24 @@ tty /dev/tty3
 #run [2] /etc/init.d/networking start      -- Start networking
 
 # Services to be monitored and respawned as needed
+
+# klgod and syslogd are placed in the maint cgroup
+cgroup.maint
 service [2345] <pid/syslogd> klogd -n     -- Kernel logging server
 service [2345] syslogd -n                 -- Syslog server
-service [3]    gdm                        -- GNOME Display Manager
+
+# gdm in the system cgroup (default when read from it's own .conf file)
+cgroup.system
+service [3] cgroup:cpu.weight:250,mem.max:10M gdm -- GNOME Display Manager
+
+# Start SSH daemon with opts from /etc/default/ssh (if available), move
+# to user cgroup with cpu.weight:250.  Default for services is system
+# group.  Default weight in user group is 100 (above).  Max length of
+# cgroup argument is (currently) 63 chars.
+#
+# Notice alternative syntax where current cgroup is only set for this
+# particular service.
+service [2345] cgroup.user:cpu.weight:250,mem.max:10M env:-/etc/default/ssh /usr/sbin/sshd -D $SSHD_OPTS -- OpenSSH daemon
 
 # The BusyBox ntpd does not use syslog when running in the foreground
 # So we use this trick to redirect stdout/stderr to a log file.  The

src/cgroup.c

@@ -28,6 +28,7 @@
 #include <sys/mount.h>
 #include <sys/sysinfo.h>		/* get_nprocs_conf() */
 
+#include "cgroup.h"
 #include "finit.h"
 #include "iwatch.h"
 #include "log.h"
@@ -39,8 +40,52 @@ static struct iwatch iw_cgroup;
 static uev_t cgw;
 
 
-static void group_init(char *path, int leaf, int weight)
+static void cgset(const char *path, char *ctrl, char *prop)
 {
+	char *val;
+
+	if (!path || !ctrl) {
+		_e("Missing path or controller, skipping!");
+		return;
+	}
+
+	if (!prop) {
+		prop = strchr(ctrl, '.');
+		if (!prop) {
+			_e("Invalid cgroup ctrl syntax: %s", ctrl);
+			return;
+		}
+
+		*prop++ = 0;
+	}
+
+	val = strchr(prop, ':');
+	if (!val) {
+		_e("Missing cgroup ctrl value, prop %s", prop);
+		return;
+	}
+	*val++ = 0;
+
+	/* disallow sneaky relative paths */
+	if (strstr(ctrl, "..") || strstr(prop, "..")) {
+		_e("Possible security violation; '..' not allowed in cgroup config!");
+		return;
+	}
+
+	_d("%s/%s.%s <= %s", path, ctrl, prop, val);
+	if (fnwrite(val, "%s/%s.%s", path, ctrl, prop))
+		_pe("Failed setting %s/%s.%s = %s", path, ctrl, prop, val);
+}
+
+/*
+ * Settings for a cgroup are on the form: cpu.weight:1234,mem.max:4321,...
+ * Finit supports the short-form 'mem.', replacing it with 'memory.' when
+ * writing the setting to the file system.
+ */
+static void group_init(char *path, int leaf, const char *cfg)
+{
+	char *ptr, *s;
+
 	if (mkdir(path, 0755) && EEXIST != errno) {
 		_pe("Failed creating cgroup %s", path);
 		return;
@@ -48,14 +93,33 @@ static void group_init(char *path, int leaf, int weight)
 
 	/* enable detected controllers on domain groups */
 	if (!leaf && fnwrite(controllers, "%s/cgroup.subtree_control", path))
-		_pe("Failed %s for %s", controllers, path);
+		_pe("Failed enabling %s for %s", controllers, path);
 
-	/* set initial group weight */
-	if (weight && fnwrite(str("%d", weight), "%s/cpu.weight", path))
-		_pe("Failed setting cpu.weight %d for %s", weight, path);
+	if (!cfg)
+		return;
+
+	s = strdup(cfg);
+	if (!s) {
+		_pe("Failed activating cgroup cfg for %s", path);
+		return;
+	}
+
+	_d("%s <=> %s", path, s);
+	ptr = strtok(s, ",");
+	while (ptr) {
+		_d("ptr: %s", ptr);
+		if (!strncmp("mem.", ptr, 4))
+			cgset(path, "memory", &ptr[4]);
+		else
+			cgset(path, ptr, NULL);
+
+		ptr = strtok(NULL, ",");
+	}
+
+	free(s);
 }
 
-static int cgroup_leaf_init(char *group, char *name, int pid)
+static int cgroup_leaf_init(char *group, char *name, int pid, const char *cfg)
 {
 	char path[256];
 
@@ -66,11 +130,11 @@ static int cgroup_leaf_init(char *group, char *name, int pid)
 
 	/* create and initialize new group */
 	snprintf(path, sizeof(path), "/sys/fs/cgroup/%s/%s", group, name);
-	group_init(path, 1, 0);
+	group_init(path, 1, cfg);
 
 	/* move process to new group */
 	if (fnwrite(str("%d", pid), "%s/cgroup.procs", path))
-		_pe("Failed moving pid %d to %s", pid, path);
+		_pe("Failed moving pid %d to group %s", pid, path);
 
 	strlcat(path, "/cgroup.events", sizeof(path));
 
@@ -79,12 +143,22 @@ static int cgroup_leaf_init(char *group, char *name, int pid)
 
 int cgroup_user(char *name, int pid)
 {
-	return cgroup_leaf_init("user", name, pid);
+	return cgroup_leaf_init("user", name, pid, NULL);
 }
 
-int cgroup_service(char *name, int pid)
+int cgroup_service(char *name, int pid, struct cgroup *cg)
 {
-	return cgroup_leaf_init("system", name, pid);
+	char *group = "system";
+
+	if (cg && cg->name[0]) {
+		char path[256];
+
+		snprintf(path, sizeof(path), "/sys/fs/cgroup/%s", cg->name);
+		if (fisdir(path))
+			group = cg->name;
+	}
+
+	return cgroup_leaf_init(group, name, pid, cg->cfg);
 }
 
 static void append_ctrl(char *ctrl)
@@ -214,14 +288,14 @@ void cgroup_init(uev_ctx_t *ctx)
 
 	/* Enable all controllers */
 	if (fnwrite(controllers, FINIT_CGPATH "/cgroup.subtree_control"))
-		_pe("Failed %s for %s", controllers, FINIT_CGPATH "/cgroup.subtree_control");
+		_pe("Failed enabling %s for %s", controllers, FINIT_CGPATH "/cgroup.subtree_control");
 
 	/* Default groups, PID 1, services, and user/login processes */
-	group_init(FINIT_CGPATH "/init",   1,  100);
-	group_init(FINIT_CGPATH "/user",   0,  100);
-	group_init(FINIT_CGPATH "/system", 0, 9800);
+	group_init(FINIT_CGPATH "/init",   1, "cpu.weight:100");
+	group_init(FINIT_CGPATH "/user",   0, "cpu.weight:100");
+	group_init(FINIT_CGPATH "/system", 0, "cpu.weight:9800");
 
-	/* Move ourselves to init */
+	/* Move ourselves to init (best effort, otherwise run in 'root' group */
 	fnwrite("1", FINIT_CGPATH "/init/cgroup.procs");
 
 	/* prepare cgroup.events watcher */
@@ -232,6 +306,19 @@ void cgroup_init(uev_ctx_t *ctx)
 	}
 }
 
+/* the top-level init cgroup is a leaf, that's ensured in cgroup_init() */
+void cgroup_config(size_t num, struct cgroup cg[])
+{
+	size_t i;
+
+	for (i = 0; i < num; i++) {
+		char path[256];
+
+		snprintf(path, sizeof(path), "%s/%s", FINIT_CGPATH, cg[i].name);
+		group_init(path, 0, cg[i].cfg);
+	}
+}
+
 /**
  * Local Variables:
  *  indent-tabs-mode: t

src/cgroup.h

@@ -26,9 +26,15 @@
 
 #include <uev/uev.h>
 
+struct cgroup {
+	char name[16];
+	char cfg[128];
+};
+
 void cgroup_init    (uev_ctx_t *ctx);
+void cgroup_config  (size_t num, struct cgroup cg[]);
 
 int  cgroup_user    (char *name, int pid);
-int  cgroup_service (char *name, int pid);
+int  cgroup_service (char *name, int pid, struct cgroup *cg);
 
 #endif /* FINIT_CGROUP_H_ */

src/conf.c

@@ -23,6 +23,7 @@
 
 #include "config.h"		/* Generated by configure script */
 
+#include <ctype.h>
 #include <dirent.h>
 #include <string.h>
 #include <sys/inotify.h>
@@ -50,6 +51,10 @@ int logfile_count_max = 5;
 struct rlimit initial_rlimit[RLIMIT_NLIMITS];
 struct rlimit global_rlimit[RLIMIT_NLIMITS];
 
+struct cgroup cgroups[8];
+size_t cgroups_num;
+char cgroup_current[16]; /* cgroup.NAME sets current cgroup for a set of services */
+
 struct conf_change {
 	TAILQ_ENTRY(conf_change) link;
 	char *name;
@@ -423,6 +428,67 @@ error:
 	logit(LOG_WARNING, "rlimit: parse error");
 }
 
+/* reset defaults */
+static void init_cgroups(void)
+{
+	struct cgroup init[3] = {
+		{ .name = "init",   .cfg = "cpu.weight:100"  },
+		{ .name = "user",   .cfg = "cpu.weight:100"  },
+		{ .name = "system", .cfg = "cpu.weight:9800" },
+	};
+	size_t i;
+
+	for (i = 0; i < NELEMS(init); i++)
+		memcpy(&cgroups[i], &init[i], sizeof(struct cgroup));
+
+	cgroups_num = i;
+}
+
+struct cgroup *conf_cgfind(char *name)
+{
+	size_t i;
+
+	for (i = 0; i < cgroups_num; i++) {
+		if (strcmp(cgroups[i].name, name))
+			continue;
+
+		return &cgroups[i];
+	}
+
+	return NULL;
+}
+
+/* cgroup NAME ctrl.prop:value,ctrl.prop:value ... */
+static void conf_parse_cgroup(char *line)
+{
+	struct cgroup *cg;
+	char *ptr, *name;
+
+	name = strtok(line, " \t");
+	if (!name)
+		return;
+
+	cg = conf_cgfind(name);
+	if (!cg) {
+		if (strstr(name, "..") || strchr(name, '/'))
+			return;		/* illegal */
+		if (cgroups_num + 1 == NELEMS(cgroups))
+			return;		/* exhausted */
+
+		cg = &cgroups[cgroups_num++];
+		strlcpy(cg->name, name, sizeof(cg->name));
+	}
+
+	cg->cfg[0] = 0;
+	while ((ptr = strtok(NULL, " \t"))) {
+		if (cg->cfg[0])
+			strlcat(cg->cfg, ",", sizeof(cg->cfg));
+		strlcat(cg->cfg, ptr, sizeof(cg->cfg));
+	}
+
+	_d("cg->cfg: %s", cg->cfg);
+}
+
 static void parse_static(char *line, int is_rcsd)
 {
 	char cmd[CMD_SIZE];
@@ -552,6 +618,18 @@ static void parse_dynamic(char *line, struct rlimit rlimit[], char *file)
 		return;
 	}
 
+	/* Read contrl group limits */
+	if (MATCH_CMD(line, "cgroup ", x)) {
+		conf_parse_cgroup(x);
+		return;
+	}
+
+	/* Set current cgroup for the following services/run/tasks */
+	if (MATCH_CMD(line, "cgroup.", x)) {
+		strlcpy(cgroup_current, x, sizeof(cgroup_current));
+		return;
+	}
+
 	/* Regular or serial TTYs to run getty */
 	if (MATCH_CMD(line, "tty ", x)) {
 		tty_register(strip_line(x), rlimit, file);
@@ -580,9 +658,11 @@ static int parse_conf(char *file, int is_rcsd)
 	if (!fp)
 		return 1;
 
-	/* Prepare default limits for each service in /etc/finit.d/ */
-	if (is_rcsd)
+	/* Prepare default limits and group for each service in /etc/finit.d/ */
+	if (is_rcsd) {
 		memcpy(rlimit, global_rlimit, sizeof(rlimit));
+		memset(cgroup_current, 0, sizeof(cgroup_current));
+	}
 
 	_d("*** Parsing %s", file);
 	while (!feof(fp)) {
@@ -630,6 +710,9 @@ int conf_reload(void)
 	 */
 	memcpy(global_rlimit, initial_rlimit, sizeof(global_rlimit));
 
+	/* Initialize default cgroups */
+	init_cgroups();
+
 	if (rescue) {
 		int rc;
 		char line[80] = "tty [12345] @console noclear nologin";
@@ -699,6 +782,9 @@ int conf_reload(void)
 
 	/* Mark any reverse deps as chenaged. */
 	service_update_rdeps();
+
+	/* Set up top-level cgroups */
+	cgroup_config(cgroups_num, cgroups);
 done:
 	/* Drop record of all .conf changes */
 	drop_changes();
@@ -888,7 +974,7 @@ int conf_monitor(void)
 }
 
 /*
- * Prepare .conf parser and load all .conf files
+ * Prepare .conf parser and load /etc/finit.conf for global settings
  */
 int conf_init(uev_ctx_t *ctx)
 {
@@ -911,6 +997,12 @@ int conf_init(uev_ctx_t *ctx)
 	/* Initialize global rlimits, e.g. for built-in services */
 	memcpy(global_rlimit, initial_rlimit, sizeof(global_rlimit));
 
+	/* Initialize default cgroups */
+	init_cgroups();
+
+	/* Read global rlimits and global cgroup setup from /etc/finit.conf */
+	parse_conf(FINIT_CONF, 0);
+
 	/* prepare /etc watcher */
 	fd = iwatch_init(&iw_conf);
 	if (fd < 0)

src/conf.h

@@ -24,16 +24,22 @@
 #ifndef FINIT_CONF_H_
 #define FINIT_CONF_H_
 
+#include "cgroup.h"
 #include "svc.h"
 
 extern int logfile_size_max;
 extern int logfile_count_max;
 
 extern struct rlimit global_rlimit[];
+extern struct cgroup cgroups[];
+extern size_t cgroup_num;
+extern char cgroup_current[];
 
 int   str2rlim(char *str);
 char *rlim2str(int rlim);
 
+struct cgroup *conf_cgfind(char *name);
+
 int  conf_init            (uev_ctx_t *ctx);
 void conf_reload          (void);
 int  conf_any_change      (void);

src/service.c

@@ -570,7 +570,7 @@ static int service_start(svc_t *svc)
 		_d("Starting %s %s", svc->cmd, buf);
 	}
 
-	cgroup_service(group_name(svc, grnam, sizeof(grnam)), pid);
+	cgroup_service(group_name(svc, grnam, sizeof(grnam)), pid, &svc->cgroup);
 
 	logit(LOG_CONSOLE | LOG_NOTICE, "Starting %s[%d]", svc_ident(svc, NULL, 0), pid);
 
@@ -857,6 +857,30 @@ static void parse_env(svc_t *svc, char *env)
 	strlcpy(svc->env, env, sizeof(svc->env));
 }
 
+static void parse_cgroup(svc_t *svc, char *cgroup)
+{
+	char *ptr = cgroup;
+
+	if (!cgroup)
+		return;
+
+	if (cgroup[0] == '.') {
+		ptr = strchr(cgroup, ':');
+		if (ptr)
+			*ptr++ = 0;
+		strlcpy(svc->cgroup.name, &cgroup[1], sizeof(svc->cgroup.name));
+		if (!ptr)
+			return;
+	}
+
+	if (strlen(ptr) >= sizeof(svc->cgroup)) {
+		_e("%s: cgroup settings too long (>%d chars)", svc->cmd, sizeof(svc->cgroup));
+		return;
+	}
+
+	strlcpy(svc->cgroup.cfg, ptr, sizeof(svc->cgroup.cfg));
+}
+
 static void parse_sighalt(svc_t *svc, char *arg)
 {
 	int signo;
@@ -1026,7 +1050,7 @@ int service_register(int type, char *cfg, struct rlimit rlimit[], char *file)
 	char *cmd, *desc, *runlevels = NULL, *cond = NULL;
 	char *username = NULL, *log = NULL, *pid = NULL;
 	char *name = NULL, *halt = NULL, *delay = NULL;
-	char *id = NULL, *env = NULL;
+	char *id = NULL, *env = NULL, *cgroup = NULL;
 	int levels = 0;
 	int manual = 0;
 	char *line;
@@ -1090,6 +1114,10 @@ int service_register(int type, char *cfg, struct rlimit rlimit[], char *file)
 			delay = &cmd[5];
 		else if (!strncasecmp(cmd, "env:", 4))
 			env = &cmd[4];
+		else if (!strncasecmp(cmd, "cgroup:", 7))
+			cgroup = &cmd[7]; /* only settings */
+		else if (!strncasecmp(cmd, "cgroup.", 7))
+			cgroup = &cmd[6]; /* with group */
 		else
 			break;
 
@@ -1148,7 +1176,7 @@ int service_register(int type, char *cfg, struct rlimit rlimit[], char *file)
 	parse_cmdline_args(svc, cmd);
 
 	svc->runlevels = levels;
-	_d("Service %s runlevel 0x%2x", svc->cmd, svc->runlevels);
+	_d("Service %s runlevel 0x%02x", svc->cmd, svc->runlevels);
 
 	conf_parse_cond(svc, cond);
 
@@ -1169,6 +1197,11 @@ int service_register(int type, char *cfg, struct rlimit rlimit[], char *file)
 	/* Set configured limits */
 	memcpy(svc->rlimit, rlimit, sizeof(svc->rlimit));
 
+	/* Seed with currently active group, may be empty */
+	strlcpy(svc->cgroup.name, cgroup_current, sizeof(svc->cgroup.name));
+	if (cgroup)
+		parse_cgroup(svc, cgroup);
+
 	/* New, recently modified or unchanged ... used on reload. */
 	if ((file && conf_changed(file)) || conf_changed(svc_getenv(svc)))
 		svc_mark_dirty(svc);

src/svc.h

@@ -32,6 +32,7 @@
 #include <lite/queue.h>		/* BSD sys/queue.h API */
 #include <uev/uev.h>
 
+#include "cgroup.h"
 #include "helpers.h"
 
 typedef int svc_cmd_t;
@@ -94,6 +95,7 @@ typedef struct svc {
 
 	/* Limits and scoping */
 	struct rlimit  rlimit[RLIMIT_NLIMITS];
+	struct cgroup  cgroup;
 
 	/* Service details */
 	int            sighalt;        /* Signal to stop prorcess, default: SIGTERM */

src/util.c

@@ -88,9 +88,10 @@ int fnwrite(char *value, char *fmt, ...)
 		return -1;
 
 	/* echo(1) always adds a newline */
-	fputs(value, fp);
-	fputs("\n", fp);
-	fclose(fp);
+	if (fputs(value, fp) == EOF ||
+	    fputs("\n", fp) == EOF  ||
+	    fclose(fp))
+		return -1;
 
 	return 0;
 }