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++;
*
* 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
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 &&
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 (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_err;
+
+ 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_err = write_attr("remove", mdi->state_fd);
+ if (!remove_err) {
+ close(mdi->state_fd);
+ mdi->state_fd = -1;
+ }
+ }
if (mdi->next_state & DS_INSYNC)
write_attr("+in_sync", mdi->state_fd);
}
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 active_array *aa, int pfd, int monfd, int nowait)
{
fd_set rfds;
int maxfd = 0;
struct active_array *a;
int rv;
+ struct mdinfo *mdi;
FD_ZERO(&rfds);
add_fd(&rfds, &maxfd, pfd);
for (a = aa ; a ; a = a->next) {
- struct mdinfo *mdi;
-
/* once an array has been deactivated only the manager
* thread can make us care about it again
*/
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, aa);
+ write(monfd, &err, 1);
}
}
if (a->container)
rv += read_and_act(a);
}
+
+ /* propagate failures across container members */
+ for (a = aa; 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(aa, mdi);
+ }
+
return rv;
}
int rv;
int first = 1;
do {
- rv = wait_and_act(container->arrays, container->pipe[0], first);
+ rv = wait_and_act(container->arrays, container->mgr_pipe[0],
+ container->mon_pipe[1], first);
first = 0;
} while (rv >= 0);
}