[thirdparty/mdadm.git] / mdmon.c


/*
 * md array manager.
 * When md arrays have user-space managed metadata, this is the program
 * that does the managing.
 *
 * Given one argument: the name of the array (e.g. /dev/md0) that is
 * the container.
 * We fork off a helper that runs high priority and mlocked.  It responds to
 * device failures and other events that might stop writeout, or that are
 * trivial to deal with.
 * The main thread then watches for new arrays being created in the container
 * and starts monitoring them too ... along with a few other tasks.
 *
 * The main thread communicates with the priority thread by writing over
 * a pipe.
 * Separate programs can communicate with the main thread via Unix-domain
 * socket.
 * The two threads share address space and open file table.
 *
 */

#ifndef _GNU_SOURCE
#define _GNU_SOURCE
#endif

#include	<unistd.h>
#include	<stdlib.h>
#include	<sys/types.h>
#include	<sys/stat.h>
#include	<sys/socket.h>
#include	<sys/un.h>
#include	<sys/mman.h>
#include	<sys/syscall.h>
#include	<sys/wait.h>
#include	<stdio.h>
#include	<errno.h>
#include	<string.h>
#include	<fcntl.h>
#include	<signal.h>

#include	<sched.h>

#include	"mdadm.h"
#include	"mdmon.h"

struct active_array *discard_this;
struct active_array *pending_discard;

int mon_tid, mgr_tid;

int run_child(void *v)
{
	struct supertype *c = v;

	do_monitor(c);
	return 0;
}

int clone_monitor(struct supertype *container)
{
	static char stack[4096];

	mon_tid = clone(run_child, stack+4096-64,
		   CLONE_FS|CLONE_FILES|CLONE_VM|CLONE_SIGHAND|CLONE_THREAD,
		   container);

	mgr_tid = syscall(SYS_gettid);

	return mon_tid;
}

static struct superswitch *find_metadata_methods(char *vers)
{
	if (strcmp(vers, "ddf") == 0)
		return &super_ddf;
	if (strcmp(vers, "imsm") == 0)
		return &super_imsm;
	return NULL;
}


int make_pidfile(char *devname, int o_excl)
{
	char path[100];
	char pid[10];
	int fd;
	sprintf(path, "/var/run/mdadm/%s.pid", devname);

	fd = open(path, O_RDWR|O_CREAT|o_excl, 0600);
	if (fd < 0)
		return -errno;
	sprintf(pid, "%d\n", getpid());
	write(fd, pid, strlen(pid));
	close(fd);
	return 0;
}

static void try_kill_monitor(char *devname)
{
	char buf[100];
	int fd;
	pid_t pid;

	sprintf(buf, "/var/run/mdadm/%s.pid", devname);
	fd = open(buf, O_RDONLY);
	if (fd < 0)
		return;

	if (read(fd, buf, sizeof(buf)) < 0) {
		close(fd);
		return;
	}

	close(fd);
	pid = strtoul(buf, NULL, 10);

	/* kill this process if it is mdmon */
	sprintf(buf, "/proc/%lu/cmdline", (unsigned long) pid);
	fd = open(buf, O_RDONLY);
	if (fd < 0)
		return;

	if (read(fd, buf, sizeof(buf)) < 0) {
		close(fd);
		return;
	}

	if (strstr(buf, "mdmon") != NULL)
		kill(pid, SIGTERM);
}

void remove_pidfile(char *devname)
{
	char buf[100];

	sprintf(buf, "/var/run/mdadm/%s.pid", devname);
	unlink(buf);
}

int make_control_sock(char *devname)
{
	char path[100];
	int sfd;
	long fl;
	struct sockaddr_un addr;

	sprintf(path, "/var/run/mdadm/%s.sock", devname);
	unlink(path);
	sfd = socket(PF_LOCAL, SOCK_STREAM, 0);
	if (sfd < 0)
		return -1;

	addr.sun_family = PF_LOCAL;
	strcpy(addr.sun_path, path);
	if (bind(sfd, &addr, sizeof(addr)) < 0) {
		close(sfd);
		return -1;
	}
	listen(sfd, 10);
	fl = fcntl(sfd, F_GETFL, 0);
	fl |= O_NONBLOCK;
	fcntl(sfd, F_SETFL, fl);
	return sfd;
}

int socket_hup_requested;
static void hup(int sig)
{
	socket_hup_requested = 1;
}

static void wake_me(int sig)
{

}

/* if we are debugging and starting mdmon by hand then don't fork */
static int do_fork(void)
{
	#ifdef DEBUG
	if (env_no_mdmon())
		return 0;
	#endif

	return 1;
}


int main(int argc, char *argv[])
{
	int mdfd;
	struct mdinfo *mdi, *di;
	struct supertype *container;
	sigset_t set;
	struct sigaction act;
	int pfd[2];
	int status;

	if (argc != 2) {
		fprintf(stderr, "Usage: md-manage /device/name/for/container\n");
		exit(2);
	}
	mdfd = open(argv[1], O_RDWR);
	if (mdfd < 0) {
		fprintf(stderr, "md-manage: %s: %s\n", argv[1],
			strerror(errno));
		exit(1);
	}
	if (md_get_version(mdfd) < 0) {
		fprintf(stderr, "md-manage: %s: Not an md device\n",
			argv[1]);
		exit(1);
	}

	/* Fork, and have the child tell us when they are ready */
	if (do_fork()) {
		pipe(pfd);
		switch(fork()) {
		case -1:
			fprintf(stderr, "mdmon: failed to fork: %s\n",
				strerror(errno));
			exit(1);
		case 0: /* child */
			close(pfd[0]);
			break;
		default: /* parent */
			close(pfd[1]);
			if (read(pfd[0], &status, sizeof(status)) != sizeof(status)) {
				wait(&status);
				status = WEXITSTATUS(status);
			}
			exit(status);
		}
	} else
		pfd[0] = pfd[1] = -1;
	/* hopefully it is a container - we'll check later */

	container = malloc(sizeof(*container));
	container->devnum = fd2devnum(mdfd);
	container->devname = devnum2devname(container->devnum);
	container->device_name = argv[1];

	/* If this fails, we hope it already exists */
	mkdir("/var/run/mdadm", 0600);
	/* pid file lives in /var/run/mdadm/mdXX.pid */
	if (make_pidfile(container->devname, O_EXCL) < 0) {
		if (ping_monitor(container->devname) == 0) {
			fprintf(stderr, "mdmon: %s already managed\n",
				container->devname);
			exit(3);
		} else {
			int err;

			/* cleanup the old monitor, this one is taking over */
			try_kill_monitor(container->devname);
			err = make_pidfile(container->devname, 0);
			if (err < 0) {
				fprintf(stderr, "mdmon: %s Cannot create pidfile\n",
					container->devname);
				if (err == -EROFS) {
					/* FIXME implement a mechanism to
					 * prevent duplicate monitor instances
					 */
					fprintf(stderr,
						"mdmon: continuing on read-only file system\n");
				} else
					exit(3);
			}
		}
	}

	container->sock = make_control_sock(container->devname);
	if (container->sock < 0)
		fprintf(stderr, "mdmon: Cannot create socket in /var/run/mdadm\n");
	container->arrays = NULL;

	mdi = sysfs_read(mdfd, container->devnum,
			 GET_VERSION|GET_LEVEL|GET_DEVS);

	if (!mdi) {
		fprintf(stderr, "mdmon: failed to load sysfs info for %s\n",
			container->devname);
		exit(3);
	}
	if (mdi->array.level != UnSet) {
		fprintf(stderr, "mdmon: %s is not a container - cannot monitor\n",
			argv[1]);
		exit(3);
	}
	if (mdi->array.major_version != -1 ||
	    mdi->array.minor_version != -2) {
		fprintf(stderr, "mdmon: %s does not use external metadata - cannot monitor\n",
			argv[1]);
		exit(3);
	}

	container->ss = find_metadata_methods(mdi->text_version);
	if (container->ss == NULL) {
		fprintf(stderr, "mdmon: %s uses unknown metadata: %s\n",
			argv[1], mdi->text_version);
		exit(3);
	}

	container->devs = NULL;
	for (di = mdi->devs; di; di = di->next) {
		struct mdinfo *cd = malloc(sizeof(*cd));
		*cd = *di;
		cd->next = container->devs;
		container->devs = cd;
	}
	sysfs_free(mdi);


	if (container->ss->load_super(container, mdfd, argv[1])) {
		fprintf(stderr, "mdmon: Cannot load metadata for %s\n",
			argv[1]);
		exit(3);
	}

	/* Ok, this is close enough.  We can say goodbye to our parent now.
	 */
	status = 0;
	write(pfd[1], &status, sizeof(status));
	close(pfd[1]);

	chdir("/");
	setsid();
	close(0);
	open("/dev/null", O_RDWR);
	close(1);
	dup(0);
#ifndef DEBUG
	close(2);
	dup(0);
#endif

	mlockall(MCL_FUTURE);

	/* SIGUSR is sent between parent and child.  So both block it
	 * and enable it only with pselect.
	 */
	sigemptyset(&set);
	sigaddset(&set, SIGUSR1);
	sigaddset(&set, SIGHUP);
	sigprocmask(SIG_BLOCK, &set, NULL);
	act.sa_handler = wake_me;
	act.sa_flags = 0;
	sigaction(SIGUSR1, &act, NULL);
	act.sa_handler = hup;
	sigaction(SIGHUP, &act, NULL);
	act.sa_handler = SIG_IGN;
	sigaction(SIGPIPE, &act, NULL);

	if (clone_monitor(container) < 0) {
		fprintf(stderr, "mdmon: failed to start monitor process: %s\n",
			strerror(errno));
		exit(2);
	}

	do_manager(container);

	exit(0);
}
Commit	Line	Data
549e9569 NB	1
	2	/*
	3	* md array manager.
	4	* When md arrays have user-space managed metadata, this is the program
	5	* that does the managing.
	6	*
	7	* Given one argument: the name of the array (e.g. /dev/md0) that is
	8	* the container.
	9	* We fork off a helper that runs high priority and mlocked. It responds to
	10	* device failures and other events that might stop writeout, or that are
	11	* trivial to deal with.
	12	* The main thread then watches for new arrays being created in the container
	13	* and starts monitoring them too ... along with a few other tasks.
	14	*
	15	* The main thread communicates with the priority thread by writing over
	16	* a pipe.
	17	* Separate programs can communicate with the main thread via Unix-domain
	18	* socket.
	19	* The two threads share address space and open file table.
	20	*
	21	*/
	22
	23	#ifndef _GNU_SOURCE
	24	#define _GNU_SOURCE
	25	#endif
	26
	27	#include <unistd.h>
	28	#include <stdlib.h>
4d43913c	29	#include <sys/types.h>
549e9569 NB	30	#include <sys/stat.h>
	31	#include <sys/socket.h>
	32	#include <sys/un.h>
	33	#include <sys/mman.h>
4d43913c	34	#include <sys/syscall.h>
9fe32043	35	#include <sys/wait.h>
549e9569 NB	36	#include <stdio.h>
	37	#include <errno.h>
	38	#include <string.h>
	39	#include <fcntl.h>
b109d928	40	#include <signal.h>
549e9569 NB	41
	42	#include <sched.h>
	43
	44	#include "mdadm.h"
	45	#include "mdmon.h"
	46
549e9569 NB	47	struct active_array *discard_this;
549e9569 NB	48	struct active_array *pending_discard;
4d43913c NB	49
4d43913c NB	50	int mon_tid, mgr_tid;
549e9569 NB	51
	52	int run_child(void *v)
	53	{
	54	struct supertype *c = v;
1ed3f387	55
549e9569 NB	56	do_monitor(c);
	57	return 0;
	58	}
	59
	60	int clone_monitor(struct supertype *container)
	61	{
549e9569	62	static char stack[4096];
549e9569	63
2cc98f9e	64	mon_tid = clone(run_child, stack+4096-64,
549e9569 NB	65	CLONE_FS\|CLONE_FILES\|CLONE_VM\|CLONE_SIGHAND\|CLONE_THREAD,
549e9569 NB	66	container);
3e70c845	67
4d43913c	68	mgr_tid = syscall(SYS_gettid);
2cc98f9e DW	69
2cc98f9e DW	70	return mon_tid;
549e9569 NB	71	}
	72
	73	static struct superswitch find_metadata_methods(char vers)
	74	{
	75	if (strcmp(vers, "ddf") == 0)
	76	return &super_ddf;
5b65005f DW	77	if (strcmp(vers, "imsm") == 0)
5b65005f DW	78	return &super_imsm;
549e9569 NB	79	return NULL;
	80	}
	81
	82
295646b3	83	int make_pidfile(char *devname, int o_excl)
549e9569 NB	84	{
	85	char path[100];
	86	char pid[10];
	87	int fd;
	88	sprintf(path, "/var/run/mdadm/%s.pid", devname);
	89
b109d928	90	fd = open(path, O_RDWR\|O_CREAT\|o_excl, 0600);
549e9569	91	if (fd < 0)
295646b3	92	return -errno;
549e9569 NB	93	sprintf(pid, "%d\n", getpid());
	94	write(fd, pid, strlen(pid));
	95	close(fd);
	96	return 0;
	97	}
	98
b109d928 DW	99	static void try_kill_monitor(char *devname)
	100	{
	101	char buf[100];
	102	int fd;
	103	pid_t pid;
	104
	105	sprintf(buf, "/var/run/mdadm/%s.pid", devname);
	106	fd = open(buf, O_RDONLY);
	107	if (fd < 0)
	108	return;
	109
	110	if (read(fd, buf, sizeof(buf)) < 0) {
	111	close(fd);
	112	return;
	113	}
	114
	115	close(fd);
	116	pid = strtoul(buf, NULL, 10);
	117
	118	/* kill this process if it is mdmon */
	119	sprintf(buf, "/proc/%lu/cmdline", (unsigned long) pid);
	120	fd = open(buf, O_RDONLY);
	121	if (fd < 0)
	122	return;
	123
	124	if (read(fd, buf, sizeof(buf)) < 0) {
	125	close(fd);
	126	return;
	127	}
	128
	129	if (strstr(buf, "mdmon") != NULL)
	130	kill(pid, SIGTERM);
	131	}
	132
e0d6609f NB	133	void remove_pidfile(char *devname)
	134	{
	135	char buf[100];
	136
	137	sprintf(buf, "/var/run/mdadm/%s.pid", devname);
	138	unlink(buf);
	139	}
	140
295646b3	141	int make_control_sock(char *devname)
549e9569 NB	142	{
	143	char path[100];
	144	int sfd;
	145	long fl;
	146	struct sockaddr_un addr;
	147
	148	sprintf(path, "/var/run/mdadm/%s.sock", devname);
	149	unlink(path);
	150	sfd = socket(PF_LOCAL, SOCK_STREAM, 0);
	151	if (sfd < 0)
	152	return -1;
	153
	154	addr.sun_family = PF_LOCAL;
	155	strcpy(addr.sun_path, path);
	156	if (bind(sfd, &addr, sizeof(addr)) < 0) {
	157	close(sfd);
	158	return -1;
	159	}
	160	listen(sfd, 10);
	161	fl = fcntl(sfd, F_GETFL, 0);
	162	fl \|= O_NONBLOCK;
	163	fcntl(sfd, F_SETFL, fl);
	164	return sfd;
	165	}
	166
295646b3 DW	167	int socket_hup_requested;
	168	static void hup(int sig)
	169	{
	170	socket_hup_requested = 1;
	171	}
	172
4d43913c NB	173	static void wake_me(int sig)
	174	{
	175
	176	}
	177
16ddab0d DW	178	/* if we are debugging and starting mdmon by hand then don't fork */
	179	static int do_fork(void)
	180	{
	181	#ifdef DEBUG
	182	if (env_no_mdmon())
	183	return 0;
	184	#endif
	185
	186	return 1;
	187	}
	188
	189
	190
549e9569 NB	191	int main(int argc, char *argv[])
	192	{
	193	int mdfd;
549e9569 NB	194	struct mdinfo mdi, di;
549e9569 NB	195	struct supertype *container;
4d43913c	196	sigset_t set;
bfa44e2e	197	struct sigaction act;
9fe32043 N	198	int pfd[2];
9fe32043 N	199	int status;
4d43913c	200
549e9569 NB	201	if (argc != 2) {
	202	fprintf(stderr, "Usage: md-manage /device/name/for/container\n");
	203	exit(2);
	204	}
	205	mdfd = open(argv[1], O_RDWR);
	206	if (mdfd < 0) {
	207	fprintf(stderr, "md-manage: %s: %s\n", argv[1],
	208	strerror(errno));
	209	exit(1);
	210	}
	211	if (md_get_version(mdfd) < 0) {
	212	fprintf(stderr, "md-manage: %s: Not an md device\n",
	213	argv[1]);
	214	exit(1);
	215	}
	216
9fe32043	217	/* Fork, and have the child tell us when they are ready */
16ddab0d DW	218	if (do_fork()) {
	219	pipe(pfd);
	220	switch(fork()) {
	221	case -1:
	222	fprintf(stderr, "mdmon: failed to fork: %s\n",
	223	strerror(errno));
	224	exit(1);
	225	case 0: /* child */
	226	close(pfd[0]);
	227	break;
	228	default: /* parent */
	229	close(pfd[1]);
	230	if (read(pfd[0], &status, sizeof(status)) != sizeof(status)) {
	231	wait(&status);
	232	status = WEXITSTATUS(status);
	233	}
	234	exit(status);
9fe32043	235	}
16ddab0d DW	236	} else
16ddab0d DW	237	pfd[0] = pfd[1] = -1;
549e9569 NB	238	/* hopefully it is a container - we'll check later */
	239
	240	container = malloc(sizeof(*container));
549e9569 NB	241	container->devnum = fd2devnum(mdfd);
549e9569 NB	242	container->devname = devnum2devname(container->devnum);
e0d6609f	243	container->device_name = argv[1];
549e9569 NB	244
	245	/* If this fails, we hope it already exists */
	246	mkdir("/var/run/mdadm", 0600);
	247	/* pid file lives in /var/run/mdadm/mdXX.pid */
b109d928 DW	248	if (make_pidfile(container->devname, O_EXCL) < 0) {
	249	if (ping_monitor(container->devname) == 0) {
	250	fprintf(stderr, "mdmon: %s already managed\n",
	251	container->devname);
	252	exit(3);
	253	} else {
295646b3 DW	254	int err;
295646b3 DW	255
b109d928 DW	256	/* cleanup the old monitor, this one is taking over */
b109d928 DW	257	try_kill_monitor(container->devname);
295646b3 DW	258	err = make_pidfile(container->devname, 0);
295646b3 DW	259	if (err < 0) {
b109d928 DW	260	fprintf(stderr, "mdmon: %s Cannot create pidfile\n",
b109d928 DW	261	container->devname);
295646b3 DW	262	if (err == -EROFS) {
	263	/* FIXME implement a mechanism to
	264	* prevent duplicate monitor instances
	265	*/
	266	fprintf(stderr,
	267	"mdmon: continuing on read-only file system\n");
	268	} else
	269	exit(3);
b109d928 DW	270	}
b109d928 DW	271	}
549e9569 NB	272	}
	273
	274	container->sock = make_control_sock(container->devname);
295646b3	275	if (container->sock < 0)
549e9569	276	fprintf(stderr, "mdmon: Cannot create socket in /var/run/mdadm\n");
549e9569 NB	277	container->arrays = NULL;
	278
	279	mdi = sysfs_read(mdfd, container->devnum,
	280	GET_VERSION\|GET_LEVEL\|GET_DEVS);
	281
	282	if (!mdi) {
	283	fprintf(stderr, "mdmon: failed to load sysfs info for %s\n",
	284	container->devname);
	285	exit(3);
	286	}
	287	if (mdi->array.level != UnSet) {
	288	fprintf(stderr, "mdmon: %s is not a container - cannot monitor\n",
	289	argv[1]);
	290	exit(3);
	291	}
	292	if (mdi->array.major_version != -1 \|\|
	293	mdi->array.minor_version != -2) {
	294	fprintf(stderr, "mdmon: %s does not use external metadata - cannot monitor\n",
	295	argv[1]);
	296	exit(3);
	297	}
	298
	299	container->ss = find_metadata_methods(mdi->text_version);
	300	if (container->ss == NULL) {
	301	fprintf(stderr, "mdmon: %s uses unknown metadata: %s\n",
	302	argv[1], mdi->text_version);
	303	exit(3);
	304	}
	305
	306	container->devs = NULL;
	307	for (di = mdi->devs; di; di = di->next) {
	308	struct mdinfo cd = malloc(sizeof(cd));
7bc1962f	309	cd = di;
549e9569 NB	310	cd->next = container->devs;
	311	container->devs = cd;
	312	}
	313	sysfs_free(mdi);
	314
	315
	316	if (container->ss->load_super(container, mdfd, argv[1])) {
	317	fprintf(stderr, "mdmon: Cannot load metadata for %s\n",
	318	argv[1]);
	319	exit(3);
	320	}
549e9569	321
9fe32043 N	322	/* Ok, this is close enough. We can say goodbye to our parent now.
	323	*/
	324	status = 0;
	325	write(pfd[1], &status, sizeof(status));
	326	close(pfd[1]);
	327
	328	chdir("/");
	329	setsid();
	330	close(0);
	331	open("/dev/null", O_RDWR);
	332	close(1);
	333	dup(0);
	334	#ifndef DEBUG
	335	close(2);
	336	dup(0);
	337	#endif
	338
549e9569 NB	339	mlockall(MCL_FUTURE);
549e9569 NB	340
4d43913c NB	341	/* SIGUSR is sent between parent and child. So both block it
	342	* and enable it only with pselect.
	343	*/
	344	sigemptyset(&set);
	345	sigaddset(&set, SIGUSR1);
295646b3	346	sigaddset(&set, SIGHUP);
4d43913c	347	sigprocmask(SIG_BLOCK, &set, NULL);
bfa44e2e NB	348	act.sa_handler = wake_me;
	349	act.sa_flags = 0;
	350	sigaction(SIGUSR1, &act, NULL);
295646b3 DW	351	act.sa_handler = hup;
295646b3 DW	352	sigaction(SIGHUP, &act, NULL);
bfa44e2e NB	353	act.sa_handler = SIG_IGN;
bfa44e2e NB	354	sigaction(SIGPIPE, &act, NULL);
4d43913c	355
3e70c845	356	if (clone_monitor(container) < 0) {
295646b3	357	fprintf(stderr, "mdmon: failed to start monitor process: %s\n",
549e9569 NB	358	strerror(errno));
	359	exit(2);
	360	}
	361
	362	do_manager(container);
	363
	364	exit(0);
	365	}