[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;
}


static 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 -1;
	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);
}

static 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;
}

static void wake_me(int sig)
{

}

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 */
	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);
	}
	/* 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 {
			/* cleanup the old monitor, this one is taking over */
			try_kill_monitor(container->devname);
			if (make_pidfile(container->devname, 0) < 0) {
				fprintf(stderr, "mdmon: %s Cannot create pidfile\n",
					container->devname);
				exit(3);
			}
		}
	}

	container->sock = make_control_sock(container->devname);
	if (container->sock < 0) {
		fprintf(stderr, "mdmon: Cannot create socket in /var/run/mdadm\n");
		exit(3);
	}
	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);
	sigprocmask(SIG_BLOCK, &set, NULL);
	act.sa_handler = wake_me;
	act.sa_flags = 0;
	sigaction(SIGUSR1, &act, NULL);
	act.sa_handler = SIG_IGN;
	sigaction(SIGPIPE, &act, NULL);

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

	do_manager(container);

	exit(0);
}
Commit	Line	Data
	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>
	29	#include <sys/types.h>
	30	#include <sys/stat.h>
	31	#include <sys/socket.h>
	32	#include <sys/un.h>
	33	#include <sys/mman.h>
	34	#include <sys/syscall.h>
	35	#include <sys/wait.h>
	36	#include <stdio.h>
	37	#include <errno.h>
	38	#include <string.h>
	39	#include <fcntl.h>
	40	#include <signal.h>
	41
	42	#include <sched.h>
	43
	44	#include "mdadm.h"
	45	#include "mdmon.h"
	46
	47	struct active_array *discard_this;
	48	struct active_array *pending_discard;
	49
	50	int mon_tid, mgr_tid;
	51
	52	int run_child(void *v)
	53	{
	54	struct supertype *c = v;
	55
	56	do_monitor(c);
	57	return 0;
	58	}
	59
	60	int clone_monitor(struct supertype *container)
	61	{
	62	static char stack[4096];
	63
	64	mon_tid = clone(run_child, stack+4096-64,
	65	CLONE_FS\|CLONE_FILES\|CLONE_VM\|CLONE_SIGHAND\|CLONE_THREAD,
	66	container);
	67
	68	mgr_tid = syscall(SYS_gettid);
	69
	70	return mon_tid;
	71	}
	72
	73	static struct superswitch find_metadata_methods(char vers)
	74	{
	75	if (strcmp(vers, "ddf") == 0)
	76	return &super_ddf;
	77	if (strcmp(vers, "imsm") == 0)
	78	return &super_imsm;
	79	return NULL;
	80	}
	81
	82
	83	static int make_pidfile(char *devname, int o_excl)
	84	{
	85	char path[100];
	86	char pid[10];
	87	int fd;
	88	sprintf(path, "/var/run/mdadm/%s.pid", devname);
	89
	90	fd = open(path, O_RDWR\|O_CREAT\|o_excl, 0600);
	91	if (fd < 0)
	92	return -1;
	93	sprintf(pid, "%d\n", getpid());
	94	write(fd, pid, strlen(pid));
	95	close(fd);
	96	return 0;
	97	}
	98
	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
	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
	141	static int make_control_sock(char *devname)
	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
	167	static void wake_me(int sig)
	168	{
	169
	170	}
	171
	172	int main(int argc, char *argv[])
	173	{
	174	int mdfd;
	175	struct mdinfo mdi, di;
	176	struct supertype *container;
	177	sigset_t set;
	178	struct sigaction act;
	179	int pfd[2];
	180	int status;
	181
	182	if (argc != 2) {
	183	fprintf(stderr, "Usage: md-manage /device/name/for/container\n");
	184	exit(2);
	185	}
	186	mdfd = open(argv[1], O_RDWR);
	187	if (mdfd < 0) {
	188	fprintf(stderr, "md-manage: %s: %s\n", argv[1],
	189	strerror(errno));
	190	exit(1);
	191	}
	192	if (md_get_version(mdfd) < 0) {
	193	fprintf(stderr, "md-manage: %s: Not an md device\n",
	194	argv[1]);
	195	exit(1);
	196	}
	197
	198	/* Fork, and have the child tell us when they are ready */
	199	pipe(pfd);
	200	switch(fork()){
	201	case -1:
	202	fprintf(stderr, "mdmon: failed to fork: %s\n",
	203	strerror(errno));
	204	exit(1);
	205	case 0: /* child */
	206	close(pfd[0]);
	207	break;
	208	default: /* parent */
	209	close(pfd[1]);
	210	if (read(pfd[0], &status, sizeof(status)) != sizeof(status)) {
	211	wait(&status);
	212	status = WEXITSTATUS(status);
	213	}
	214	exit(status);
	215	}
	216	/* hopefully it is a container - we'll check later */
	217
	218	container = malloc(sizeof(*container));
	219	container->devnum = fd2devnum(mdfd);
	220	container->devname = devnum2devname(container->devnum);
	221	container->device_name = argv[1];
	222
	223	/* If this fails, we hope it already exists */
	224	mkdir("/var/run/mdadm", 0600);
	225	/* pid file lives in /var/run/mdadm/mdXX.pid */
	226	if (make_pidfile(container->devname, O_EXCL) < 0) {
	227	if (ping_monitor(container->devname) == 0) {
	228	fprintf(stderr, "mdmon: %s already managed\n",
	229	container->devname);
	230	exit(3);
	231	} else {
	232	/* cleanup the old monitor, this one is taking over */
	233	try_kill_monitor(container->devname);
	234	if (make_pidfile(container->devname, 0) < 0) {
	235	fprintf(stderr, "mdmon: %s Cannot create pidfile\n",
	236	container->devname);
	237	exit(3);
	238	}
	239	}
	240	}
	241
	242	container->sock = make_control_sock(container->devname);
	243	if (container->sock < 0) {
	244	fprintf(stderr, "mdmon: Cannot create socket in /var/run/mdadm\n");
	245	exit(3);
	246	}
	247	container->arrays = NULL;
	248
	249	mdi = sysfs_read(mdfd, container->devnum,
	250	GET_VERSION\|GET_LEVEL\|GET_DEVS);
	251
	252	if (!mdi) {
	253	fprintf(stderr, "mdmon: failed to load sysfs info for %s\n",
	254	container->devname);
	255	exit(3);
	256	}
	257	if (mdi->array.level != UnSet) {
	258	fprintf(stderr, "mdmon: %s is not a container - cannot monitor\n",
	259	argv[1]);
	260	exit(3);
	261	}
	262	if (mdi->array.major_version != -1 \|\|
	263	mdi->array.minor_version != -2) {
	264	fprintf(stderr, "mdmon: %s does not use external metadata - cannot monitor\n",
	265	argv[1]);
	266	exit(3);
	267	}
	268
	269	container->ss = find_metadata_methods(mdi->text_version);
	270	if (container->ss == NULL) {
	271	fprintf(stderr, "mdmon: %s uses unknown metadata: %s\n",
	272	argv[1], mdi->text_version);
	273	exit(3);
	274	}
	275
	276	container->devs = NULL;
	277	for (di = mdi->devs; di; di = di->next) {
	278	struct mdinfo cd = malloc(sizeof(cd));
	279	cd = di;
	280	cd->next = container->devs;
	281	container->devs = cd;
	282	}
	283	sysfs_free(mdi);
	284
	285
	286	if (container->ss->load_super(container, mdfd, argv[1])) {
	287	fprintf(stderr, "mdmon: Cannot load metadata for %s\n",
	288	argv[1]);
	289	exit(3);
	290	}
	291
	292	/* Ok, this is close enough. We can say goodbye to our parent now.
	293	*/
	294	status = 0;
	295	write(pfd[1], &status, sizeof(status));
	296	close(pfd[1]);
	297
	298	chdir("/");
	299	setsid();
	300	close(0);
	301	open("/dev/null", O_RDWR);
	302	close(1);
	303	dup(0);
	304	#ifndef DEBUG
	305	close(2);
	306	dup(0);
	307	#endif
	308
	309	mlockall(MCL_FUTURE);
	310
	311	/* SIGUSR is sent between parent and child. So both block it
	312	* and enable it only with pselect.
	313	*/
	314	sigemptyset(&set);
	315	sigaddset(&set, SIGUSR1);
	316	sigprocmask(SIG_BLOCK, &set, NULL);
	317	act.sa_handler = wake_me;
	318	act.sa_flags = 0;
	319	sigaction(SIGUSR1, &act, NULL);
	320	act.sa_handler = SIG_IGN;
	321	sigaction(SIGPIPE, &act, NULL);
	322
	323	if (clone_monitor(container) < 0) {
	324	fprintf(stderr, "md-manage: failed to start monitor process: %s\n",
	325	strerror(errno));
	326	exit(2);
	327	}
	328
	329	do_manager(container);
	330
	331	exit(0);
	332	}