#include "mdmon.h"
#include <sys/select.h>
-
+#include <signal.h>
static char *array_states[] = {
"clear", "inactive", "suspended", "readonly", "read-auto",
return n;
}
-static int get_sync_pos(struct active_array *a)
-{
- char buf[30];
- int n;
-
- n = read_attr(buf, 30, a->sync_pos_fd);
- if (n <= 0)
- return n;
-
- if (strncmp(buf, "max", 3) == 0) {
- a->sync_pos = ~(unsigned long long)0;
- return 1;
- }
- a->sync_pos = strtoull(buf, NULL, 10);
- return 1;
-}
static int get_resync_start(struct active_array *a)
{
return (enum sync_action) match_word(buf, sync_actions);
}
-#define DS_FAULTY 1
-#define DS_INSYNC 2
-#define DS_WRITE_MOSTLY 4
-#define DS_SPARE 8
-#define DS_REMOVE 1024
-
int read_dev_state(int fd)
{
char buf[60];
cp = buf;
while (cp) {
- if (attr_match("faulty", cp))
+ if (attr_match(cp, "faulty"))
rv |= DS_FAULTY;
- if (attr_match("in_sync", cp))
+ if (attr_match(cp, "in_sync"))
rv |= DS_INSYNC;
- if (attr_match("write_mostly", cp))
+ if (attr_match(cp, "write_mostly"))
rv |= DS_WRITE_MOSTLY;
- if (attr_match("spare", cp))
+ if (attr_match(cp, "spare"))
rv |= DS_SPARE;
+ if (attr_match(cp, "blocked"))
+ rv |= DS_BLOCKED;
cp = strchr(cp, ',');
if (cp)
cp++;
return rv;
}
+static void signal_manager(void)
+{
+ kill(getpid(), SIGUSR1);
+}
/* Monitor a set of active md arrays - all of which share the
* same metadata - and respond to events that require
*
* device fails
* detected by rd-N/state reporting "faulty"
- * mark device as 'failed' in metadata, the remove device
- * by writing 'remove' to rd/state.
+ * mark device as 'failed' in metadata, let the kernel release the
+ * device by writing '-blocked' to rd/state, and finally write 'remove' to
+ * rd/state. Before a disk can be replaced it must be failed and removed
+ * from all container members, this will be preemptive for the other
+ * arrays... safe?
*
* sync completes
* sync_action was 'resync' and becomes 'idle' and resync_start becomes
* MaxSector
* Notify metadata that sync is complete.
- * "Deal with Degraded"
*
* recovery completes
* sync_action changes from 'recover' to 'idle'
* Check each device state and mark metadata if 'faulty' or 'in_sync'.
- * "Deal with Degraded"
- *
- * deal with degraded array
- * We only do this when first noticing the array is degraded.
- * This can be when we first see the array, when sync completes or
- * when recovery completes.
- *
- * Check if number of failed devices suggests recovery is needed, and
- * skip if not.
- * Ask metadata for a spare device
- * Add device as not in_sync and give a role
- * Update metadata.
- * Start recovery.
*
* deal with resync
* This only happens on finding a new array... mdadm will have set
static int read_and_act(struct active_array *a)
{
- int check_degraded;
+ int check_degraded = 0;
+ int deactivate = 0;
struct mdinfo *mdi;
a->next_state = bad_word;
a->curr_action = read_action(a->action_fd);
for (mdi = a->info.devs; mdi ; mdi = mdi->next) {
mdi->next_state = 0;
- mdi->curr_state = read_dev_state(mdi->state_fd);
+ if (mdi->state_fd >= 0)
+ mdi->curr_state = read_dev_state(mdi->state_fd);
}
if (a->curr_state <= inactive &&
a->prev_state > inactive) {
/* array has been stopped */
- get_sync_pos(a);
- a->container->ss->mark_clean(a, a->sync_pos);
+ a->container->ss->set_array_state(a, 1);
a->next_state = clear;
+ deactivate = 1;
}
if (a->curr_state == write_pending) {
- a->container->ss->mark_dirty(a);
+ get_resync_start(a);
+ a->container->ss->set_array_state(a, 0);
a->next_state = active;
}
if (a->curr_state == active_idle) {
/* Set array to 'clean' FIRST, then
- * a->ss->mark_clean(a);
+ * a->ss->mark_clean(a, ~0ULL);
* just ignore for now.
*/
}
* readonly ???
*/
get_resync_start(a);
+ printf("Found a readonly array at %llu\n", a->resync_start);
if (a->resync_start == ~0ULL)
a->next_state = read_auto; /* array is clean */
else {
- a->container->ss->mark_dirty(a);
+ a->container->ss->set_array_state(a, 0);
a->next_state = active;
}
}
if (a->curr_action == idle &&
a->prev_action == resync) {
- /* check resync_start to see if it is 'max' */
+ /* A resync has finished. The endpoint is recorded in
+ * 'sync_start'. We don't update the metadata
+ * until the array goes inactive or readonly though.
+ * Just check if we need to fiddle spares.
+ */
get_resync_start(a);
- a->container->ss->mark_sync(a, a->resync_start);
+ a->container->ss->set_array_state(a, 0);
check_degraded = 1;
}
if (a->curr_action == idle &&
a->prev_action == recover) {
for (mdi = a->info.devs ; mdi ; mdi = mdi->next) {
- a->container->ss->set_disk(a, mdi->disk.raid_disk);
+ a->container->ss->set_disk(a, mdi->disk.raid_disk,
+ mdi->curr_state);
if (! (mdi->curr_state & DS_INSYNC))
check_degraded = 1;
}
}
-
for (mdi = a->info.devs ; mdi ; mdi = mdi->next) {
if (mdi->curr_state & DS_FAULTY) {
- a->container->ss->set_disk(a, mdi->disk.raid_disk);
+ a->container->ss->set_disk(a, mdi->disk.raid_disk,
+ mdi->curr_state);
check_degraded = 1;
mdi->next_state = DS_REMOVE;
}
}
- if (check_degraded) {
- // FIXME;
- }
-
- a->container->ss->sync_metadata(a);
+ a->container->ss->sync_metadata(a->container);
/* Effect state changes in the array */
if (a->next_state != bad_word)
if (a->next_action != bad_action)
write_attr(sync_actions[a->next_action], a->action_fd);
for (mdi = a->info.devs; mdi ; mdi = mdi->next) {
- if (mdi->next_state == DS_REMOVE)
- write_attr("remove", mdi->state_fd);
+ if (mdi->next_state == DS_REMOVE && mdi->state_fd >= 0) {
+ int remove_result;
+
+ write_attr("-blocked", mdi->state_fd);
+ /* the kernel may not be able to immediately remove the
+ * disk, we can simply wait until the next event to try
+ * again.
+ */
+ remove_result = write_attr("remove", mdi->state_fd);
+ if (remove_result > 0) {
+ close(mdi->state_fd);
+ mdi->state_fd = -1;
+ }
+ }
if (mdi->next_state & DS_INSYNC)
write_attr("+in_sync", mdi->state_fd);
}
mdi->next_state = 0;
}
+ if (check_degraded) {
+ /* manager will do the actual check */
+ a->check_degraded = 1;
+ signal_manager();
+ }
+
+ if (deactivate)
+ a->container = NULL;
+
return 1;
}
-static int wait_and_act(struct active_array *aa, int pfd, int nowait)
+static struct mdinfo *
+find_device(struct active_array *a, int major, int minor)
+{
+ struct mdinfo *mdi;
+
+ for (mdi = a->info.devs ; mdi ; mdi = mdi->next)
+ if (mdi->disk.major == major && mdi->disk.minor == minor)
+ return mdi;
+
+ return NULL;
+}
+
+static void reconcile_failed(struct active_array *aa, struct mdinfo *failed)
+{
+ struct active_array *a;
+ struct mdinfo *victim;
+
+ for (a = aa; a; a = a->next) {
+ if (!a->container)
+ continue;
+ victim = find_device(a, failed->disk.major, failed->disk.minor);
+ if (!victim)
+ continue;
+
+ if (!(victim->curr_state & DS_FAULTY))
+ write_attr("faulty", victim->state_fd);
+ }
+}
+
+static int handle_remove_device(struct md_remove_device_cmd *cmd, struct active_array *aa)
+{
+ struct active_array *a;
+ struct mdinfo *victim;
+ int rv;
+
+ /* scan all arrays for the given device, if ->state_fd is closed (-1)
+ * in all cases then mark the disk as removed in the metadata.
+ * Otherwise reply that it is busy.
+ */
+
+ /* pass1 check that it is not in use anywhere */
+ /* note: we are safe from re-adds as long as the device exists in the
+ * container
+ */
+ for (a = aa; a; a = a->next) {
+ if (!a->container)
+ continue;
+ victim = find_device(a, major(cmd->rdev), minor(cmd->rdev));
+ if (!victim)
+ continue;
+ if (victim->state_fd > 0)
+ return -EBUSY;
+ }
+
+ /* pass2 schedule and process removal per array */
+ for (a = aa; a; a = a->next) {
+ if (!a->container)
+ continue;
+ victim = find_device(a, major(cmd->rdev), minor(cmd->rdev));
+ if (!victim)
+ continue;
+ victim->curr_state |= DS_REMOVE;
+ rv = read_and_act(a);
+ if (rv < 0)
+ return rv;
+ }
+
+ return 0;
+}
+
+static int handle_pipe(struct md_generic_cmd *cmd, struct active_array *aa)
+{
+ switch (cmd->action) {
+ case md_action_ping_monitor:
+ return 0;
+ case md_action_remove_device:
+ return handle_remove_device((void *) cmd, aa);
+ }
+
+ return -1;
+}
+
+static int wait_and_act(struct supertype *container, int pfd,
+ int monfd, int nowait)
{
fd_set rfds;
int maxfd = 0;
- struct active_array *a;
+ struct active_array **aap = &container->arrays;
+ struct active_array *a, **ap;
int rv;
+ struct mdinfo *mdi;
FD_ZERO(&rfds);
add_fd(&rfds, &maxfd, pfd);
- for (a = aa ; a ; a = a->next) {
- struct mdinfo *mdi;
+ for (ap = aap ; *ap ;) {
+ a = *ap;
+ /* once an array has been deactivated we want to
+ * ask the manager to discard it.
+ */
+ if (!a->container) {
+ if (discard_this) {
+ ap = &(*ap)->next;
+ continue;
+ }
+ *ap = a->next;
+ a->next = NULL;
+ discard_this = a;
+ signal_manager();
+ continue;
+ }
add_fd(&rfds, &maxfd, a->info.state_fd);
add_fd(&rfds, &maxfd, a->action_fd);
for (mdi = a->info.devs ; mdi ; mdi = mdi->next)
add_fd(&rfds, &maxfd, mdi->state_fd);
+
+ ap = &(*ap)->next;
+ }
+
+ if (manager_ready && *aap == NULL) {
+ /* No interesting arrays. Lets see about exiting.
+ * Note that blocking at this point is not a problem
+ * as there are no active arrays, there is nothing that
+ * we need to be ready to do.
+ */
+ int fd = open(container->device_name, O_RDONLY|O_EXCL);
+ if (fd >= 0 || errno != EBUSY) {
+ /* OK, we are safe to leave */
+ exit_now = 1;
+ signal_manager();
+ remove_pidfile(container->devname);
+ exit(0);
+ }
}
if (!nowait) {
return rv;
if (FD_ISSET(pfd, &rfds)) {
- char buf[4];
- read(pfd, buf, 4);
- ; // FIXME read from the pipe
+ int err = -1;
+
+ if (read(pfd, &err, 1) > 0)
+ err = handle_pipe(active_cmd, *aap);
+ write(monfd, &err, 1);
}
}
- for (a = aa; a ; a = a->next) {
- if (a->replaces) {
+ if (update_queue) {
+ struct metadata_update *this;
+
+ for (this = update_queue; this ; this = this->next)
+ container->ss->process_update(container, this);
+
+ update_queue_handled = update_queue;
+ update_queue = NULL;
+ signal_manager();
+ container->ss->sync_metadata(container);
+ }
+
+ for (a = *aap; a ; a = a->next) {
+ if (a->replaces && !discard_this) {
struct active_array **ap;
for (ap = &a->next; *ap && *ap != a->replaces;
ap = & (*ap)->next)
*ap = (*ap)->next;
discard_this = a->replaces;
a->replaces = NULL;
+ /* FIXME check if device->state_fd need to be cleared?*/
+ signal_manager();
}
- rv += read_and_act(a);
+ if (a->container)
+ rv += read_and_act(a);
}
+
+ /* propagate failures across container members */
+ for (a = *aap; a ; a = a->next) {
+ if (!a->container)
+ continue;
+ for (mdi = a->info.devs ; mdi ; mdi = mdi->next)
+ if (mdi->curr_state & DS_FAULTY)
+ reconcile_failed(*aap, mdi);
+ }
+
return rv;
}
int rv;
int first = 1;
do {
- rv = wait_and_act(container->arrays, container->pipe[0], first);
+ rv = wait_and_act(container, container->mgr_pipe[0],
+ container->mon_pipe[1], first);
first = 0;
} while (rv >= 0);
}