[thirdparty/mdadm.git] / managemon.c


/*
 * The management thread for monitoring active md arrays.
 * This thread does things which might block such as memory
 * allocation.
 * In particular:
 *
 * - Find out about new arrays in this container.
 *   Allocate the data structures and open the files.
 *
 *   For this we watch /proc/mdstat and find new arrays with
 *   metadata type that confirms sharing. e.g. "md4"
 *   When we find a new array we slip it into the list of
 *   arrays and signal 'monitor' by writing to a pipe.
 *
 * - Respond to reshape requests by allocating new data structures
 *   and opening new files.
 *
 *   These come as a change to raid_disks.  We allocate a new
 *   version of the data structures and slip it into the list.
 *   'monitor' will notice and release the old version.
 *   Changes to level, chunksize, layout.. do not need re-allocation.
 *   Reductions in raid_disks don't really either, but we handle
 *   them the same way for consistency.
 *
 * - When a device is added to the container, we add it to the metadata
 *   as a spare.
 *
 * - assist with activating spares by opening relevant sysfs file.
 *
 * - Pass on metadata updates from external programs such as
 *   mdadm creating a new array.
 *
 *   This is most-messy.
 *   It might involve adding a new array or changing the status of
 *   a spare, or any reconfig that the kernel doesn't get involved in.
 *
 *   The required updates are received via a named pipe.  There will
 *   be one named pipe for each container. Each message contains a
 *   sync marker: 0x5a5aa5a5, A byte count, and the message.  This is
 *   passed to the metadata handler which will interpret and process it.
 *   For 'DDF' messages are internal data blocks with the leading
 *   'magic number' signifying what sort of data it is.
 *
 */

/*
 * We select on /proc/mdstat and the named pipe.
 * We create new arrays or updated version of arrays and slip
 * them into the head of the list, then signal 'monitor' via a pipe write.
 * 'monitor' will notice and place the old array on a return list.
 * Metadata updates are placed on a queue just like they arrive
 * from the named pipe.
 *
 * When new arrays are found based on correct metadata string, we
 * need to identify them with an entry in the metadata.  Maybe we require
 * the metadata to be mdX/NN  when NN is the index into an appropriate table.
 *
 */

/*
 * List of tasks:
 * - Watch for spares to be added to the container, and write updated
 *   metadata to them.
 * - Watch for new arrays using this container, confirm they match metadata
 *   and if so, start monitoring them
 * - Watch for spares being added to monitored arrays.  This shouldn't
 *   happen, as we should do all the adding.  Just remove them.
 * - Watch for change in raid-disks, chunk-size, etc.  Update metadata and
 *   start a reshape.
 */
#ifndef _GNU_SOURCE
#define _GNU_SOURCE
#endif
#include	"mdadm.h"
#include	"mdmon.h"
#include	<sys/socket.h>


static void close_aa(struct active_array *aa)
{
	struct mdinfo *d;

	for (d = aa->info.devs; d; d = d->next)
		close(d->state_fd);

	close(aa->action_fd);
	close(aa->info.state_fd);
	close(aa->resync_start_fd);
	close(aa->sync_pos_fd);
}

static void free_aa(struct active_array *aa)
{
	/* Note that this doesn't close fds if they are being used
	 * by a clone.  ->container will be set for a clone
	 */
	if (!aa->container)
		close_aa(aa);
	while (aa->info.devs) {
		struct mdinfo *d = aa->info.devs;
		aa->info.devs = d->next;
		free(d);
	}
	free(aa);
}

static void write_wakeup(struct supertype *c)
{
	write(c->pipe[1], "PING", 4);
}

static void replace_array(struct supertype *container,
			  struct active_array *old,
			  struct active_array *new)
{
	/* To replace an array, we add it to the top of the list
	 * marked with ->replaces to point to the original.
	 * 'monitor' will take the original out of the list
	 * and put it on 'discard_this'.  We take it from there
	 * and discard it.
	 */

	while (pending_discard) {
		while (discard_this == NULL)
			sleep(1);
		if (discard_this != pending_discard)
			abort();
		discard_this->next = NULL;
		free_aa(discard_this);
		discard_this = NULL;
		pending_discard = NULL;
	}
	pending_discard = old;
	new->replaces = old;
	new->next = container->arrays;
	container->arrays = new;
	write_wakeup(container);
}


static void manage_container(struct mdstat_ent *mdstat,
			     struct supertype *container)
{
	/* The only thing of interest here is if a new device
	 * has been added to the container.  We add it to the
	 * array ignoring any metadata on it.
	 * FIXME should we look for compatible metadata and take hints
	 * about spare assignment.... probably not.
	 *
	 */
	if (mdstat->devcnt != container->devcnt) {
		/* read /sys/block/NAME/md/dev-??/block/dev to find out
		 * what is there, and compare with container->info.devs
		 * To see what is removed and what is added.
		 * These need to be remove from, or added to, the array
		 */
		// FIXME
		container->devcnt = mdstat->devcnt;
	}
}

static void manage_member(struct mdstat_ent *mdstat,
			  struct active_array *a)
{
	/* Compare mdstat info with known state of member array.
	 * We do not need to look for device state changes here, that
	 * is dealt with by the monitor.
	 *
	 * We just look for changes which suggest that a reshape is
	 * being requested.
	 * Unfortunately decreases in raid_disks don't show up in
	 * mdstat until the reshape completes FIXME.
	 */
	// FIXME
	a->info.array.raid_disks = mdstat->raid_disks;
	a->info.array.chunk_size = mdstat->chunk_size;
	// MORE

}

static void manage_new(struct mdstat_ent *mdstat,
		       struct supertype *container,
		       struct active_array *victim)
{
	/* A new array has appeared in this container.
	 * Hopefully it is already recorded in the metadata.
	 * Check, then create the new array to report it to
	 * the monitor.
	 */

	struct active_array *new;
	struct mdinfo *mdi, *di;
	char *n;
	int inst;
	int i;

	new = malloc(sizeof(*new));

	new->devnum = mdstat->devnum;

	new->prev_state = new->curr_state = new->next_state = inactive;
	new->prev_action= new->curr_action= new->next_action= idle;

	new->container = container;

	n = &mdstat->metadata_version[10+strlen(container->devname)+1];
	inst = atoi(n);
	if (inst < 0)
		abort();//FIXME

	mdi = sysfs_read(-1, new->devnum,
			 GET_LEVEL|GET_CHUNK|GET_DISKS|
			 GET_DEVS|GET_OFFSET|GET_SIZE|GET_STATE);
	if (!mdi) {
		/* Eeek. Cannot monitor this array.
		 * Mark it to be ignored by setting container to NULL
		 */
		new->container = NULL;
		replace_array(container, victim, new);
		return;
	}

	new->info.array = mdi->array;

	for (i = 0; i < new->info.array.raid_disks; i++) {
		struct mdinfo *newd = malloc(sizeof(*newd));

		for (di = mdi->devs; di; di = di->next)
			if (i == di->disk.raid_disk)
				break;

		if (di) {
			memcpy(newd, di, sizeof(*newd));

			sprintf(newd->sys_name, "rd%d", i);

			newd->state_fd = sysfs_open(new->devnum,
						    newd->sys_name,
						    "state");

			newd->prev_state = read_dev_state(newd->state_fd);
			newd->curr_state = newd->curr_state;
		} else {
			newd->state_fd = -1;
		}
		newd->next = new->info.devs;
		new->info.devs = newd;
	}
	new->action_fd = sysfs_open(new->devnum, NULL, "sync_action");
	new->info.state_fd = sysfs_open(new->devnum, NULL, "array_state");
	new->resync_start_fd = sysfs_open(new->devnum, NULL, "resync_start");
	new->sync_pos_fd = sysfs_open(new->devnum, NULL, "sync_completed");
	new->sync_pos = 0;

	// finds and compares.
	if (container->ss->open_new(container, new, inst) < 0) {
		// FIXME close all those files
		new->container = NULL;
		replace_array(container, victim, new);
		return;
	}
	replace_array(container, victim, new);
	return;
}

void manage(struct mdstat_ent *mdstat, struct active_array *aa,
	    struct supertype *container)
{
	/* We have just read mdstat and need to compare it with
	 * the known active arrays.
	 * Arrays with the wrong metadata are ignored.
	 */

	for ( ; mdstat ; mdstat = mdstat->next) {
		struct active_array *a;
		if (mdstat->devnum == container->devnum) {
			manage_container(mdstat, container);
			continue;
		}
		if (mdstat->metadata_version == NULL ||
		    strncmp(mdstat->metadata_version, "external:/", 10) != 0 ||
		    strncmp(mdstat->metadata_version+10, container->devname,
			    strlen(container->devname)) != 0 ||
		    mdstat->metadata_version[10+strlen(container->devname)]
		      != '/')
			/* Not for this array */
			continue;
		/* Looks like a member of this container */
		for (a = aa; a; a = a->next) {
			if (mdstat->devnum == a->devnum) {
				if (a->container)
					manage_member(mdstat, a);
				break;
			}
		}
		if (a == NULL || !a->container)
			manage_new(mdstat, container, a);
	}
}

void read_sock(int pfd)
{
	int fd;

	// FIXME set non-blocking
	fd = accept(pfd, NULL, NULL);
	if (fd < 0)
		return;
	// FIXME do something useful
	close(fd);
}
void do_manager(struct supertype *container)
{
	struct mdstat_ent *mdstat;

	do {
		mdstat = mdstat_read(1, 0);

		manage(mdstat, array_list, container);

		read_sock(container->sock);

		mdstat_wait_fd(container->sock);
	} while(1);
}
Commit	Line	Data
549e9569 NB	1
	2	/*
	3	* The management thread for monitoring active md arrays.
	4	* This thread does things which might block such as memory
	5	* allocation.
	6	* In particular:
	7	*
	8	* - Find out about new arrays in this container.
	9	* Allocate the data structures and open the files.
	10	*
	11	* For this we watch /proc/mdstat and find new arrays with
	12	* metadata type that confirms sharing. e.g. "md4"
	13	* When we find a new array we slip it into the list of
	14	* arrays and signal 'monitor' by writing to a pipe.
	15	*
	16	* - Respond to reshape requests by allocating new data structures
	17	* and opening new files.
	18	*
	19	* These come as a change to raid_disks. We allocate a new
	20	* version of the data structures and slip it into the list.
	21	* 'monitor' will notice and release the old version.
	22	* Changes to level, chunksize, layout.. do not need re-allocation.
	23	* Reductions in raid_disks don't really either, but we handle
	24	* them the same way for consistency.
	25	*
	26	* - When a device is added to the container, we add it to the metadata
	27	* as a spare.
	28	*
	29	* - assist with activating spares by opening relevant sysfs file.
	30	*
	31	* - Pass on metadata updates from external programs such as
	32	* mdadm creating a new array.
	33	*
	34	* This is most-messy.
	35	* It might involve adding a new array or changing the status of
	36	* a spare, or any reconfig that the kernel doesn't get involved in.
	37	*
	38	* The required updates are received via a named pipe. There will
	39	* be one named pipe for each container. Each message contains a
	40	* sync marker: 0x5a5aa5a5, A byte count, and the message. This is
	41	* passed to the metadata handler which will interpret and process it.
	42	* For 'DDF' messages are internal data blocks with the leading
	43	* 'magic number' signifying what sort of data it is.
	44	*
	45	*/
	46
	47	/*
	48	* We select on /proc/mdstat and the named pipe.
	49	* We create new arrays or updated version of arrays and slip
	50	* them into the head of the list, then signal 'monitor' via a pipe write.
	51	* 'monitor' will notice and place the old array on a return list.
	52	* Metadata updates are placed on a queue just like they arrive
	53	* from the named pipe.
	54	*
	55	* When new arrays are found based on correct metadata string, we
	56	* need to identify them with an entry in the metadata. Maybe we require
	57	* the metadata to be mdX/NN when NN is the index into an appropriate table.
	58	*
	59	*/
	60
	61	/*
	62	* List of tasks:
	63	* - Watch for spares to be added to the container, and write updated
	64	* metadata to them.
65	* - Watch for new arrays using this container, confirm they match metadata
66	* and if so, start monitoring them
67	* - Watch for spares being added to monitored arrays. This shouldn't
68	* happen, as we should do all the adding. Just remove them.
69	* - Watch for change in raid-disks, chunk-size, etc. Update metadata and
70	* start a reshape.
71	*/
72	#ifndef _GNU_SOURCE
73	#define _GNU_SOURCE
74	#endif
75	#include "mdadm.h"
76	#include "mdmon.h"
77	#include <sys/socket.h>
78
79
2a0bb19e DW	80	static void close_aa(struct active_array *aa)
	81	{
	82	struct mdinfo *d;
	83
	84	for (d = aa->info.devs; d; d = d->next)
	85	close(d->state_fd);
	86
	87	close(aa->action_fd);
	88	close(aa->info.state_fd);
	89	close(aa->resync_start_fd);
	90	close(aa->sync_pos_fd);
	91	}
	92
549e9569 NB	93	static void free_aa(struct active_array *aa)
549e9569 NB	94	{
2a0bb19e DW	95	/* Note that this doesn't close fds if they are being used
2a0bb19e DW	96	* by a clone. ->container will be set for a clone
549e9569	97	*/
2a0bb19e DW	98	if (!aa->container)
2a0bb19e DW	99	close_aa(aa);
549e9569 NB	100	while (aa->info.devs) {
	101	struct mdinfo *d = aa->info.devs;
	102	aa->info.devs = d->next;
	103	free(d);
	104	}
	105	free(aa);
	106	}
	107
2a0bb19e DW	108	static void write_wakeup(struct supertype *c)
	109	{
	110	write(c->pipe[1], "PING", 4);
	111	}
	112
549e9569 NB	113	static void replace_array(struct supertype *container,
	114	struct active_array *old,
	115	struct active_array *new)
	116	{
	117	/* To replace an array, we add it to the top of the list
	118	* marked with ->replaces to point to the original.
	119	* 'monitor' will take the original out of the list
	120	* and put it on 'discard_this'. We take it from there
	121	* and discard it.
	122	*/
	123
	124	while (pending_discard) {
	125	while (discard_this == NULL)
	126	sleep(1);
	127	if (discard_this != pending_discard)
	128	abort();
	129	discard_this->next = NULL;
	130	free_aa(discard_this);
	131	discard_this = NULL;
	132	pending_discard = NULL;
	133	}
	134	pending_discard = old;
	135	new->replaces = old;
	136	new->next = container->arrays;
	137	container->arrays = new;
2a0bb19e	138	write_wakeup(container);
549e9569 NB	139	}
	140
	141
	142	static void manage_container(struct mdstat_ent *mdstat,
	143	struct supertype *container)
	144	{
	145	/* The only thing of interest here is if a new device
	146	* has been added to the container. We add it to the
	147	* array ignoring any metadata on it.
	148	* FIXME should we look for compatible metadata and take hints
	149	* about spare assignment.... probably not.
	150	*
	151	*/
	152	if (mdstat->devcnt != container->devcnt) {
	153	/* read /sys/block/NAME/md/dev-??/block/dev to find out
	154	* what is there, and compare with container->info.devs
	155	* To see what is removed and what is added.
	156	* These need to be remove from, or added to, the array
	157	*/
	158	// FIXME
	159	container->devcnt = mdstat->devcnt;
	160	}
	161	}
	162
	163	static void manage_member(struct mdstat_ent *mdstat,
	164	struct active_array *a)
	165	{
	166	/* Compare mdstat info with known state of member array.
	167	* We do not need to look for device state changes here, that
	168	* is dealt with by the monitor.
	169	*
	170	* We just look for changes which suggest that a reshape is
	171	* being requested.
	172	* Unfortunately decreases in raid_disks don't show up in
	173	* mdstat until the reshape completes FIXME.
	174	*/
	175	// FIXME
	176	a->info.array.raid_disks = mdstat->raid_disks;
	177	a->info.array.chunk_size = mdstat->chunk_size;
	178	// MORE
	179
	180	}
	181
549e9569	182	static void manage_new(struct mdstat_ent *mdstat,
2a0bb19e DW	183	struct supertype *container,
2a0bb19e DW	184	struct active_array *victim)
549e9569 NB	185	{
	186	/* A new array has appeared in this container.
	187	* Hopefully it is already recorded in the metadata.
	188	* Check, then create the new array to report it to
	189	* the monitor.
	190	*/
	191
	192	struct active_array *new;
	193	struct mdinfo mdi, di;
	194	char *n;
	195	int inst;
	196	int i;
	197
	198	new = malloc(sizeof(*new));
	199
	200	new->devnum = mdstat->devnum;
	201
	202	new->prev_state = new->curr_state = new->next_state = inactive;
	203	new->prev_action= new->curr_action= new->next_action= idle;
	204
	205	new->container = container;
	206
	207	n = &mdstat->metadata_version[10+strlen(container->devname)+1];
	208	inst = atoi(n);
	209	if (inst < 0)
	210	abort();//FIXME
	211
	212	mdi = sysfs_read(-1, new->devnum,
	213	GET_LEVEL\|GET_CHUNK\|GET_DISKS\|
	214	GET_DEVS\|GET_OFFSET\|GET_SIZE\|GET_STATE);
	215	if (!mdi) {
	216	/* Eeek. Cannot monitor this array.
	217	* Mark it to be ignored by setting container to NULL
	218	*/
	219	new->container = NULL;
2a0bb19e	220	replace_array(container, victim, new);
549e9569 NB	221	return;
	222	}
	223
	224	new->info.array = mdi->array;
	225
	226	for (i = 0; i < new->info.array.raid_disks; i++) {
	227	struct mdinfo newd = malloc(sizeof(newd));
	228
	229	for (di = mdi->devs; di; di = di->next)
	230	if (i == di->disk.raid_disk)
	231	break;
	232
	233	if (di) {
	234	memcpy(newd, di, sizeof(*newd));
	235
	236	sprintf(newd->sys_name, "rd%d", i);
	237
	238	newd->state_fd = sysfs_open(new->devnum,
	239	newd->sys_name,
	240	"state");
	241
	242	newd->prev_state = read_dev_state(newd->state_fd);
	243	newd->curr_state = newd->curr_state;
	244	} else {
	245	newd->state_fd = -1;
	246	}
	247	newd->next = new->info.devs;
	248	new->info.devs = newd;
	249	}
	250	new->action_fd = sysfs_open(new->devnum, NULL, "sync_action");
	251	new->info.state_fd = sysfs_open(new->devnum, NULL, "array_state");
c052ba30	252	new->resync_start_fd = sysfs_open(new->devnum, NULL, "resync_start");
549e9569 NB	253	new->sync_pos_fd = sysfs_open(new->devnum, NULL, "sync_completed");
	254	new->sync_pos = 0;
	255
	256	// finds and compares.
	257	if (container->ss->open_new(container, new, inst) < 0) {
	258	// FIXME close all those files
	259	new->container = NULL;
2a0bb19e	260	replace_array(container, victim, new);
549e9569 NB	261	return;
549e9569 NB	262	}
2a0bb19e	263	replace_array(container, victim, new);
549e9569 NB	264	return;
	265	}
	266
	267	void manage(struct mdstat_ent mdstat, struct active_array aa,
	268	struct supertype *container)
	269	{
	270	/* We have just read mdstat and need to compare it with
	271	* the known active arrays.
	272	* Arrays with the wrong metadata are ignored.
	273	*/
	274
	275	for ( ; mdstat ; mdstat = mdstat->next) {
	276	struct active_array *a;
	277	if (mdstat->devnum == container->devnum) {
	278	manage_container(mdstat, container);
	279	continue;
	280	}
	281	if (mdstat->metadata_version == NULL \|\|
	282	strncmp(mdstat->metadata_version, "external:/", 10) != 0 \|\|
	283	strncmp(mdstat->metadata_version+10, container->devname,
	284	strlen(container->devname)) != 0 \|\|
	285	mdstat->metadata_version[10+strlen(container->devname)]
	286	!= '/')
	287	/* Not for this array */
	288	continue;
	289	/* Looks like a member of this container */
	290	for (a = aa; a; a = a->next) {
	291	if (mdstat->devnum == a->devnum) {
	292	if (a->container)
	293	manage_member(mdstat, a);
	294	break;
	295	}
	296	}
2a0bb19e DW	297	if (a == NULL \|\| !a->container)
2a0bb19e DW	298	manage_new(mdstat, container, a);
549e9569 NB	299	}
	300	}
	301
	302	void read_sock(int pfd)
	303	{
	304	int fd;
	305
	306	// FIXME set non-blocking
	307	fd = accept(pfd, NULL, NULL);
	308	if (fd < 0)
	309	return;
	310	// FIXME do something useful
	311	close(fd);
	312	}
	313	void do_manager(struct supertype *container)
	314	{
	315	struct mdstat_ent *mdstat;
	316
	317	do {
	318	mdstat = mdstat_read(1, 0);
	319
	320	manage(mdstat, array_list, container);
	321
	322	read_sock(container->sock);
	323
	324	mdstat_wait_fd(container->sock);
	325	} while(1);
	326	}