2 * Copyright (C) 2015, SUSE
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2, or (at your option)
12 #include <linux/module.h>
13 #include <linux/kthread.h>
14 #include <linux/dlm.h>
15 #include <linux/sched.h>
16 #include <linux/raid/md_p.h>
19 #include "md-cluster.h"
22 #define NEW_DEV_TIMEOUT 5000
24 struct dlm_lock_resource {
27 char *name; /* lock name. */
28 uint32_t flags; /* flags to pass to dlm_lock() */
29 wait_queue_head_t sync_locking; /* wait queue for synchronized locking */
30 bool sync_locking_done;
31 void (*bast)(void *arg, int mode); /* blocking AST function pointer*/
32 struct mddev *mddev; /* pointing back to mddev. */
40 struct list_head list;
48 /* md_cluster_info flags */
49 #define MD_CLUSTER_WAITING_FOR_NEWDISK 1
50 #define MD_CLUSTER_SUSPEND_READ_BALANCING 2
51 #define MD_CLUSTER_BEGIN_JOIN_CLUSTER 3
53 /* Lock the send communication. This is done through
54 * bit manipulation as opposed to a mutex in order to
55 * accomodate lock and hold. See next comment.
57 #define MD_CLUSTER_SEND_LOCK 4
58 /* If cluster operations (such as adding a disk) must lock the
59 * communication channel, so as to perform extra operations
60 * (update metadata) and no other operation is allowed on the
61 * MD. Token needs to be locked and held until the operation
62 * completes witha md_update_sb(), which would eventually release
65 #define MD_CLUSTER_SEND_LOCKED_ALREADY 5
66 /* We should receive message after node joined cluster and
67 * set up all the related infos such as bitmap and personality */
68 #define MD_CLUSTER_ALREADY_IN_CLUSTER 6
69 #define MD_CLUSTER_PENDING_RECV_EVENT 7
70 #define MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD 8
72 struct md_cluster_info {
73 struct mddev *mddev; /* the md device which md_cluster_info belongs to */
74 /* dlm lock space and resources for clustered raid. */
75 dlm_lockspace_t *lockspace;
77 struct completion completion;
78 struct mutex recv_mutex;
79 struct dlm_lock_resource *bitmap_lockres;
80 struct dlm_lock_resource **other_bitmap_lockres;
81 struct dlm_lock_resource *resync_lockres;
82 struct list_head suspend_list;
83 spinlock_t suspend_lock;
84 struct md_thread *recovery_thread;
85 unsigned long recovery_map;
86 /* communication loc resources */
87 struct dlm_lock_resource *ack_lockres;
88 struct dlm_lock_resource *message_lockres;
89 struct dlm_lock_resource *token_lockres;
90 struct dlm_lock_resource *no_new_dev_lockres;
91 struct md_thread *recv_thread;
92 struct completion newdisk_completion;
93 wait_queue_head_t wait;
95 /* record the region in RESYNCING message */
101 METADATA_UPDATED = 0,
113 /* TODO: Unionize this for smaller footprint */
120 static void sync_ast(void *arg)
122 struct dlm_lock_resource *res;
125 res->sync_locking_done = true;
126 wake_up(&res->sync_locking);
129 static int dlm_lock_sync(struct dlm_lock_resource *res, int mode)
133 ret = dlm_lock(res->ls, mode, &res->lksb,
134 res->flags, res->name, strlen(res->name),
135 0, sync_ast, res, res->bast);
138 wait_event(res->sync_locking, res->sync_locking_done);
139 res->sync_locking_done = false;
140 if (res->lksb.sb_status == 0)
142 return res->lksb.sb_status;
145 static int dlm_unlock_sync(struct dlm_lock_resource *res)
147 return dlm_lock_sync(res, DLM_LOCK_NL);
151 * An variation of dlm_lock_sync, which make lock request could
154 static int dlm_lock_sync_interruptible(struct dlm_lock_resource *res, int mode,
159 ret = dlm_lock(res->ls, mode, &res->lksb,
160 res->flags, res->name, strlen(res->name),
161 0, sync_ast, res, res->bast);
165 wait_event(res->sync_locking, res->sync_locking_done
166 || kthread_should_stop()
167 || test_bit(MD_CLOSING, &mddev->flags));
168 if (!res->sync_locking_done) {
170 * the convert queue contains the lock request when request is
171 * interrupted, and sync_ast could still be run, so need to
172 * cancel the request and reset completion
174 ret = dlm_unlock(res->ls, res->lksb.sb_lkid, DLM_LKF_CANCEL,
176 res->sync_locking_done = false;
177 if (unlikely(ret != 0))
178 pr_info("failed to cancel previous lock request "
179 "%s return %d\n", res->name, ret);
182 res->sync_locking_done = false;
183 if (res->lksb.sb_status == 0)
185 return res->lksb.sb_status;
188 static struct dlm_lock_resource *lockres_init(struct mddev *mddev,
189 char *name, void (*bastfn)(void *arg, int mode), int with_lvb)
191 struct dlm_lock_resource *res = NULL;
193 struct md_cluster_info *cinfo = mddev->cluster_info;
195 res = kzalloc(sizeof(struct dlm_lock_resource), GFP_KERNEL);
198 init_waitqueue_head(&res->sync_locking);
199 res->sync_locking_done = false;
200 res->ls = cinfo->lockspace;
202 res->mode = DLM_LOCK_IV;
203 namelen = strlen(name);
204 res->name = kzalloc(namelen + 1, GFP_KERNEL);
206 pr_err("md-cluster: Unable to allocate resource name for resource %s\n", name);
209 strlcpy(res->name, name, namelen + 1);
211 res->lksb.sb_lvbptr = kzalloc(LVB_SIZE, GFP_KERNEL);
212 if (!res->lksb.sb_lvbptr) {
213 pr_err("md-cluster: Unable to allocate LVB for resource %s\n", name);
216 res->flags = DLM_LKF_VALBLK;
222 res->flags |= DLM_LKF_EXPEDITE;
224 ret = dlm_lock_sync(res, DLM_LOCK_NL);
226 pr_err("md-cluster: Unable to lock NL on new lock resource %s\n", name);
229 res->flags &= ~DLM_LKF_EXPEDITE;
230 res->flags |= DLM_LKF_CONVERT;
234 kfree(res->lksb.sb_lvbptr);
240 static void lockres_free(struct dlm_lock_resource *res)
248 * use FORCEUNLOCK flag, so we can unlock even the lock is on the
249 * waiting or convert queue
251 ret = dlm_unlock(res->ls, res->lksb.sb_lkid, DLM_LKF_FORCEUNLOCK,
253 if (unlikely(ret != 0))
254 pr_err("failed to unlock %s return %d\n", res->name, ret);
256 wait_event(res->sync_locking, res->sync_locking_done);
259 kfree(res->lksb.sb_lvbptr);
263 static void add_resync_info(struct dlm_lock_resource *lockres,
264 sector_t lo, sector_t hi)
266 struct resync_info *ri;
268 ri = (struct resync_info *)lockres->lksb.sb_lvbptr;
269 ri->lo = cpu_to_le64(lo);
270 ri->hi = cpu_to_le64(hi);
273 static struct suspend_info *read_resync_info(struct mddev *mddev, struct dlm_lock_resource *lockres)
275 struct resync_info ri;
276 struct suspend_info *s = NULL;
279 dlm_lock_sync(lockres, DLM_LOCK_CR);
280 memcpy(&ri, lockres->lksb.sb_lvbptr, sizeof(struct resync_info));
281 hi = le64_to_cpu(ri.hi);
283 s = kzalloc(sizeof(struct suspend_info), GFP_KERNEL);
287 s->lo = le64_to_cpu(ri.lo);
289 dlm_unlock_sync(lockres);
294 static void recover_bitmaps(struct md_thread *thread)
296 struct mddev *mddev = thread->mddev;
297 struct md_cluster_info *cinfo = mddev->cluster_info;
298 struct dlm_lock_resource *bm_lockres;
301 struct suspend_info *s, *tmp;
304 while (cinfo->recovery_map) {
305 slot = fls64((u64)cinfo->recovery_map) - 1;
307 snprintf(str, 64, "bitmap%04d", slot);
308 bm_lockres = lockres_init(mddev, str, NULL, 1);
310 pr_err("md-cluster: Cannot initialize bitmaps\n");
314 ret = dlm_lock_sync_interruptible(bm_lockres, DLM_LOCK_PW, mddev);
316 pr_err("md-cluster: Could not DLM lock %s: %d\n",
320 ret = bitmap_copy_from_slot(mddev, slot, &lo, &hi, true);
322 pr_err("md-cluster: Could not copy data from bitmap %d\n", slot);
326 /* Clear suspend_area associated with the bitmap */
327 spin_lock_irq(&cinfo->suspend_lock);
328 list_for_each_entry_safe(s, tmp, &cinfo->suspend_list, list)
329 if (slot == s->slot) {
333 spin_unlock_irq(&cinfo->suspend_lock);
336 if (lo < mddev->recovery_cp)
337 mddev->recovery_cp = lo;
338 /* wake up thread to continue resync in case resync
340 if (mddev->recovery_cp != MaxSector) {
341 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
342 md_wakeup_thread(mddev->thread);
346 lockres_free(bm_lockres);
347 clear_bit(slot, &cinfo->recovery_map);
351 static void recover_prep(void *arg)
353 struct mddev *mddev = arg;
354 struct md_cluster_info *cinfo = mddev->cluster_info;
355 set_bit(MD_CLUSTER_SUSPEND_READ_BALANCING, &cinfo->state);
358 static void __recover_slot(struct mddev *mddev, int slot)
360 struct md_cluster_info *cinfo = mddev->cluster_info;
362 set_bit(slot, &cinfo->recovery_map);
363 if (!cinfo->recovery_thread) {
364 cinfo->recovery_thread = md_register_thread(recover_bitmaps,
366 if (!cinfo->recovery_thread) {
367 pr_warn("md-cluster: Could not create recovery thread\n");
371 md_wakeup_thread(cinfo->recovery_thread);
374 static void recover_slot(void *arg, struct dlm_slot *slot)
376 struct mddev *mddev = arg;
377 struct md_cluster_info *cinfo = mddev->cluster_info;
379 pr_info("md-cluster: %s Node %d/%d down. My slot: %d. Initiating recovery.\n",
380 mddev->bitmap_info.cluster_name,
381 slot->nodeid, slot->slot,
383 /* deduct one since dlm slot starts from one while the num of
384 * cluster-md begins with 0 */
385 __recover_slot(mddev, slot->slot - 1);
388 static void recover_done(void *arg, struct dlm_slot *slots,
389 int num_slots, int our_slot,
392 struct mddev *mddev = arg;
393 struct md_cluster_info *cinfo = mddev->cluster_info;
395 cinfo->slot_number = our_slot;
396 /* completion is only need to be complete when node join cluster,
397 * it doesn't need to run during another node's failure */
398 if (test_bit(MD_CLUSTER_BEGIN_JOIN_CLUSTER, &cinfo->state)) {
399 complete(&cinfo->completion);
400 clear_bit(MD_CLUSTER_BEGIN_JOIN_CLUSTER, &cinfo->state);
402 clear_bit(MD_CLUSTER_SUSPEND_READ_BALANCING, &cinfo->state);
405 /* the ops is called when node join the cluster, and do lock recovery
406 * if node failure occurs */
407 static const struct dlm_lockspace_ops md_ls_ops = {
408 .recover_prep = recover_prep,
409 .recover_slot = recover_slot,
410 .recover_done = recover_done,
414 * The BAST function for the ack lock resource
415 * This function wakes up the receive thread in
416 * order to receive and process the message.
418 static void ack_bast(void *arg, int mode)
420 struct dlm_lock_resource *res = arg;
421 struct md_cluster_info *cinfo = res->mddev->cluster_info;
423 if (mode == DLM_LOCK_EX) {
424 if (test_bit(MD_CLUSTER_ALREADY_IN_CLUSTER, &cinfo->state))
425 md_wakeup_thread(cinfo->recv_thread);
427 set_bit(MD_CLUSTER_PENDING_RECV_EVENT, &cinfo->state);
431 static void __remove_suspend_info(struct md_cluster_info *cinfo, int slot)
433 struct suspend_info *s, *tmp;
435 list_for_each_entry_safe(s, tmp, &cinfo->suspend_list, list)
436 if (slot == s->slot) {
443 static void remove_suspend_info(struct mddev *mddev, int slot)
445 struct md_cluster_info *cinfo = mddev->cluster_info;
446 mddev->pers->quiesce(mddev, 1);
447 spin_lock_irq(&cinfo->suspend_lock);
448 __remove_suspend_info(cinfo, slot);
449 spin_unlock_irq(&cinfo->suspend_lock);
450 mddev->pers->quiesce(mddev, 0);
454 static void process_suspend_info(struct mddev *mddev,
455 int slot, sector_t lo, sector_t hi)
457 struct md_cluster_info *cinfo = mddev->cluster_info;
458 struct suspend_info *s;
461 remove_suspend_info(mddev, slot);
462 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
463 md_wakeup_thread(mddev->thread);
468 * The bitmaps are not same for different nodes
469 * if RESYNCING is happening in one node, then
470 * the node which received the RESYNCING message
471 * probably will perform resync with the region
472 * [lo, hi] again, so we could reduce resync time
473 * a lot if we can ensure that the bitmaps among
474 * different nodes are match up well.
476 * sync_low/hi is used to record the region which
477 * arrived in the previous RESYNCING message,
479 * Call bitmap_sync_with_cluster to clear
480 * NEEDED_MASK and set RESYNC_MASK since
481 * resync thread is running in another node,
482 * so we don't need to do the resync again
483 * with the same section */
484 bitmap_sync_with_cluster(mddev, cinfo->sync_low,
487 cinfo->sync_low = lo;
490 s = kzalloc(sizeof(struct suspend_info), GFP_KERNEL);
496 mddev->pers->quiesce(mddev, 1);
497 spin_lock_irq(&cinfo->suspend_lock);
498 /* Remove existing entry (if exists) before adding */
499 __remove_suspend_info(cinfo, slot);
500 list_add(&s->list, &cinfo->suspend_list);
501 spin_unlock_irq(&cinfo->suspend_lock);
502 mddev->pers->quiesce(mddev, 0);
505 static void process_add_new_disk(struct mddev *mddev, struct cluster_msg *cmsg)
508 struct md_cluster_info *cinfo = mddev->cluster_info;
509 char event_name[] = "EVENT=ADD_DEVICE";
511 char *envp[] = {event_name, disk_uuid, raid_slot, NULL};
514 len = snprintf(disk_uuid, 64, "DEVICE_UUID=");
515 sprintf(disk_uuid + len, "%pU", cmsg->uuid);
516 snprintf(raid_slot, 16, "RAID_DISK=%d", le32_to_cpu(cmsg->raid_slot));
517 pr_info("%s:%d Sending kobject change with %s and %s\n", __func__, __LINE__, disk_uuid, raid_slot);
518 init_completion(&cinfo->newdisk_completion);
519 set_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state);
520 kobject_uevent_env(&disk_to_dev(mddev->gendisk)->kobj, KOBJ_CHANGE, envp);
521 wait_for_completion_timeout(&cinfo->newdisk_completion,
523 clear_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state);
527 static void process_metadata_update(struct mddev *mddev, struct cluster_msg *msg)
530 struct md_cluster_info *cinfo = mddev->cluster_info;
531 mddev->good_device_nr = le32_to_cpu(msg->raid_slot);
533 dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR);
534 wait_event(mddev->thread->wqueue,
535 (got_lock = mddev_trylock(mddev)) ||
536 test_bit(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state));
537 md_reload_sb(mddev, mddev->good_device_nr);
542 static void process_remove_disk(struct mddev *mddev, struct cluster_msg *msg)
544 struct md_rdev *rdev;
547 rdev = md_find_rdev_nr_rcu(mddev, le32_to_cpu(msg->raid_slot));
549 set_bit(ClusterRemove, &rdev->flags);
550 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
551 md_wakeup_thread(mddev->thread);
554 pr_warn("%s: %d Could not find disk(%d) to REMOVE\n",
555 __func__, __LINE__, le32_to_cpu(msg->raid_slot));
559 static void process_readd_disk(struct mddev *mddev, struct cluster_msg *msg)
561 struct md_rdev *rdev;
564 rdev = md_find_rdev_nr_rcu(mddev, le32_to_cpu(msg->raid_slot));
565 if (rdev && test_bit(Faulty, &rdev->flags))
566 clear_bit(Faulty, &rdev->flags);
568 pr_warn("%s: %d Could not find disk(%d) which is faulty",
569 __func__, __LINE__, le32_to_cpu(msg->raid_slot));
573 static int process_recvd_msg(struct mddev *mddev, struct cluster_msg *msg)
577 if (WARN(mddev->cluster_info->slot_number - 1 == le32_to_cpu(msg->slot),
578 "node %d received it's own msg\n", le32_to_cpu(msg->slot)))
580 switch (le32_to_cpu(msg->type)) {
581 case METADATA_UPDATED:
582 process_metadata_update(mddev, msg);
584 case CHANGE_CAPACITY:
585 set_capacity(mddev->gendisk, mddev->array_sectors);
586 revalidate_disk(mddev->gendisk);
589 process_suspend_info(mddev, le32_to_cpu(msg->slot),
590 le64_to_cpu(msg->low),
591 le64_to_cpu(msg->high));
594 process_add_new_disk(mddev, msg);
597 process_remove_disk(mddev, msg);
600 process_readd_disk(mddev, msg);
602 case BITMAP_NEEDS_SYNC:
603 __recover_slot(mddev, le32_to_cpu(msg->slot));
607 pr_warn("%s:%d Received unknown message from %d\n",
608 __func__, __LINE__, msg->slot);
614 * thread for receiving message
616 static void recv_daemon(struct md_thread *thread)
618 struct md_cluster_info *cinfo = thread->mddev->cluster_info;
619 struct dlm_lock_resource *ack_lockres = cinfo->ack_lockres;
620 struct dlm_lock_resource *message_lockres = cinfo->message_lockres;
621 struct cluster_msg msg;
624 mutex_lock(&cinfo->recv_mutex);
625 /*get CR on Message*/
626 if (dlm_lock_sync(message_lockres, DLM_LOCK_CR)) {
627 pr_err("md/raid1:failed to get CR on MESSAGE\n");
628 mutex_unlock(&cinfo->recv_mutex);
632 /* read lvb and wake up thread to process this message_lockres */
633 memcpy(&msg, message_lockres->lksb.sb_lvbptr, sizeof(struct cluster_msg));
634 ret = process_recvd_msg(thread->mddev, &msg);
638 /*release CR on ack_lockres*/
639 ret = dlm_unlock_sync(ack_lockres);
640 if (unlikely(ret != 0))
641 pr_info("unlock ack failed return %d\n", ret);
642 /*up-convert to PR on message_lockres*/
643 ret = dlm_lock_sync(message_lockres, DLM_LOCK_PR);
644 if (unlikely(ret != 0))
645 pr_info("lock PR on msg failed return %d\n", ret);
646 /*get CR on ack_lockres again*/
647 ret = dlm_lock_sync(ack_lockres, DLM_LOCK_CR);
648 if (unlikely(ret != 0))
649 pr_info("lock CR on ack failed return %d\n", ret);
651 /*release CR on message_lockres*/
652 ret = dlm_unlock_sync(message_lockres);
653 if (unlikely(ret != 0))
654 pr_info("unlock msg failed return %d\n", ret);
655 mutex_unlock(&cinfo->recv_mutex);
659 * Takes the lock on the TOKEN lock resource so no other
660 * node can communicate while the operation is underway.
662 static int lock_token(struct md_cluster_info *cinfo)
666 error = dlm_lock_sync(cinfo->token_lockres, DLM_LOCK_EX);
668 pr_err("md-cluster(%s:%d): failed to get EX on TOKEN (%d)\n",
669 __func__, __LINE__, error);
671 /* Lock the receive sequence */
672 mutex_lock(&cinfo->recv_mutex);
678 * Sets the MD_CLUSTER_SEND_LOCK bit to lock the send channel.
680 static int lock_comm(struct md_cluster_info *cinfo, bool mddev_locked)
683 struct mddev *mddev = cinfo->mddev;
686 * If resync thread run after raid1d thread, then process_metadata_update
687 * could not continue if raid1d held reconfig_mutex (and raid1d is blocked
688 * since another node already got EX on Token and waitting the EX of Ack),
689 * so let resync wake up thread in case flag is set.
691 if (mddev_locked && !test_bit(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD,
693 rv = test_and_set_bit_lock(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD,
696 md_wakeup_thread(mddev->thread);
700 wait_event(cinfo->wait,
701 !test_and_set_bit(MD_CLUSTER_SEND_LOCK, &cinfo->state));
702 rv = lock_token(cinfo);
704 clear_bit_unlock(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state);
708 static void unlock_comm(struct md_cluster_info *cinfo)
710 WARN_ON(cinfo->token_lockres->mode != DLM_LOCK_EX);
711 mutex_unlock(&cinfo->recv_mutex);
712 dlm_unlock_sync(cinfo->token_lockres);
713 clear_bit(MD_CLUSTER_SEND_LOCK, &cinfo->state);
714 wake_up(&cinfo->wait);
718 * This function performs the actual sending of the message. This function is
719 * usually called after performing the encompassing operation
721 * 1. Grabs the message lockresource in EX mode
722 * 2. Copies the message to the message LVB
723 * 3. Downconverts message lockresource to CW
724 * 4. Upconverts ack lock resource from CR to EX. This forces the BAST on other nodes
725 * and the other nodes read the message. The thread will wait here until all other
726 * nodes have released ack lock resource.
727 * 5. Downconvert ack lockresource to CR
729 static int __sendmsg(struct md_cluster_info *cinfo, struct cluster_msg *cmsg)
732 int slot = cinfo->slot_number - 1;
734 cmsg->slot = cpu_to_le32(slot);
735 /*get EX on Message*/
736 error = dlm_lock_sync(cinfo->message_lockres, DLM_LOCK_EX);
738 pr_err("md-cluster: failed to get EX on MESSAGE (%d)\n", error);
742 memcpy(cinfo->message_lockres->lksb.sb_lvbptr, (void *)cmsg,
743 sizeof(struct cluster_msg));
744 /*down-convert EX to CW on Message*/
745 error = dlm_lock_sync(cinfo->message_lockres, DLM_LOCK_CW);
747 pr_err("md-cluster: failed to convert EX to CW on MESSAGE(%d)\n",
752 /*up-convert CR to EX on Ack*/
753 error = dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_EX);
755 pr_err("md-cluster: failed to convert CR to EX on ACK(%d)\n",
760 /*down-convert EX to CR on Ack*/
761 error = dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_CR);
763 pr_err("md-cluster: failed to convert EX to CR on ACK(%d)\n",
769 error = dlm_unlock_sync(cinfo->message_lockres);
770 if (unlikely(error != 0)) {
771 pr_err("md-cluster: failed convert to NL on MESSAGE(%d)\n",
773 /* in case the message can't be released due to some reason */
780 static int sendmsg(struct md_cluster_info *cinfo, struct cluster_msg *cmsg,
785 ret = lock_comm(cinfo, mddev_locked);
787 ret = __sendmsg(cinfo, cmsg);
793 static int gather_all_resync_info(struct mddev *mddev, int total_slots)
795 struct md_cluster_info *cinfo = mddev->cluster_info;
797 struct dlm_lock_resource *bm_lockres;
798 struct suspend_info *s;
803 for (i = 0; i < total_slots; i++) {
804 memset(str, '\0', 64);
805 snprintf(str, 64, "bitmap%04d", i);
806 bm_lockres = lockres_init(mddev, str, NULL, 1);
809 if (i == (cinfo->slot_number - 1)) {
810 lockres_free(bm_lockres);
814 bm_lockres->flags |= DLM_LKF_NOQUEUE;
815 ret = dlm_lock_sync(bm_lockres, DLM_LOCK_PW);
816 if (ret == -EAGAIN) {
817 s = read_resync_info(mddev, bm_lockres);
819 pr_info("%s:%d Resync[%llu..%llu] in progress on %d\n",
821 (unsigned long long) s->lo,
822 (unsigned long long) s->hi, i);
823 spin_lock_irq(&cinfo->suspend_lock);
825 list_add(&s->list, &cinfo->suspend_list);
826 spin_unlock_irq(&cinfo->suspend_lock);
829 lockres_free(bm_lockres);
833 lockres_free(bm_lockres);
837 /* Read the disk bitmap sb and check if it needs recovery */
838 ret = bitmap_copy_from_slot(mddev, i, &lo, &hi, false);
840 pr_warn("md-cluster: Could not gather bitmaps from slot %d", i);
841 lockres_free(bm_lockres);
844 if ((hi > 0) && (lo < mddev->recovery_cp)) {
845 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
846 mddev->recovery_cp = lo;
847 md_check_recovery(mddev);
850 lockres_free(bm_lockres);
856 static int join(struct mddev *mddev, int nodes)
858 struct md_cluster_info *cinfo;
862 cinfo = kzalloc(sizeof(struct md_cluster_info), GFP_KERNEL);
866 INIT_LIST_HEAD(&cinfo->suspend_list);
867 spin_lock_init(&cinfo->suspend_lock);
868 init_completion(&cinfo->completion);
869 set_bit(MD_CLUSTER_BEGIN_JOIN_CLUSTER, &cinfo->state);
870 init_waitqueue_head(&cinfo->wait);
871 mutex_init(&cinfo->recv_mutex);
873 mddev->cluster_info = cinfo;
874 cinfo->mddev = mddev;
877 sprintf(str, "%pU", mddev->uuid);
878 ret = dlm_new_lockspace(str, mddev->bitmap_info.cluster_name,
879 DLM_LSFL_FS, LVB_SIZE,
880 &md_ls_ops, mddev, &ops_rv, &cinfo->lockspace);
883 wait_for_completion(&cinfo->completion);
884 if (nodes < cinfo->slot_number) {
885 pr_err("md-cluster: Slot allotted(%d) is greater than available slots(%d).",
886 cinfo->slot_number, nodes);
890 /* Initiate the communication resources */
892 cinfo->recv_thread = md_register_thread(recv_daemon, mddev, "cluster_recv");
893 if (!cinfo->recv_thread) {
894 pr_err("md-cluster: cannot allocate memory for recv_thread!\n");
897 cinfo->message_lockres = lockres_init(mddev, "message", NULL, 1);
898 if (!cinfo->message_lockres)
900 cinfo->token_lockres = lockres_init(mddev, "token", NULL, 0);
901 if (!cinfo->token_lockres)
903 cinfo->no_new_dev_lockres = lockres_init(mddev, "no-new-dev", NULL, 0);
904 if (!cinfo->no_new_dev_lockres)
907 ret = dlm_lock_sync(cinfo->token_lockres, DLM_LOCK_EX);
910 pr_err("md-cluster: can't join cluster to avoid lock issue\n");
913 cinfo->ack_lockres = lockres_init(mddev, "ack", ack_bast, 0);
914 if (!cinfo->ack_lockres) {
918 /* get sync CR lock on ACK. */
919 if (dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_CR))
920 pr_err("md-cluster: failed to get a sync CR lock on ACK!(%d)\n",
922 dlm_unlock_sync(cinfo->token_lockres);
923 /* get sync CR lock on no-new-dev. */
924 if (dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR))
925 pr_err("md-cluster: failed to get a sync CR lock on no-new-dev!(%d)\n", ret);
928 pr_info("md-cluster: Joined cluster %s slot %d\n", str, cinfo->slot_number);
929 snprintf(str, 64, "bitmap%04d", cinfo->slot_number - 1);
930 cinfo->bitmap_lockres = lockres_init(mddev, str, NULL, 1);
931 if (!cinfo->bitmap_lockres) {
935 if (dlm_lock_sync(cinfo->bitmap_lockres, DLM_LOCK_PW)) {
936 pr_err("Failed to get bitmap lock\n");
941 cinfo->resync_lockres = lockres_init(mddev, "resync", NULL, 0);
942 if (!cinfo->resync_lockres) {
949 set_bit(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state);
950 md_unregister_thread(&cinfo->recovery_thread);
951 md_unregister_thread(&cinfo->recv_thread);
952 lockres_free(cinfo->message_lockres);
953 lockres_free(cinfo->token_lockres);
954 lockres_free(cinfo->ack_lockres);
955 lockres_free(cinfo->no_new_dev_lockres);
956 lockres_free(cinfo->resync_lockres);
957 lockres_free(cinfo->bitmap_lockres);
958 if (cinfo->lockspace)
959 dlm_release_lockspace(cinfo->lockspace, 2);
960 mddev->cluster_info = NULL;
965 static void load_bitmaps(struct mddev *mddev, int total_slots)
967 struct md_cluster_info *cinfo = mddev->cluster_info;
969 /* load all the node's bitmap info for resync */
970 if (gather_all_resync_info(mddev, total_slots))
971 pr_err("md-cluster: failed to gather all resyn infos\n");
972 set_bit(MD_CLUSTER_ALREADY_IN_CLUSTER, &cinfo->state);
973 /* wake up recv thread in case something need to be handled */
974 if (test_and_clear_bit(MD_CLUSTER_PENDING_RECV_EVENT, &cinfo->state))
975 md_wakeup_thread(cinfo->recv_thread);
978 static void resync_bitmap(struct mddev *mddev)
980 struct md_cluster_info *cinfo = mddev->cluster_info;
981 struct cluster_msg cmsg = {0};
984 cmsg.type = cpu_to_le32(BITMAP_NEEDS_SYNC);
985 err = sendmsg(cinfo, &cmsg, 1);
987 pr_err("%s:%d: failed to send BITMAP_NEEDS_SYNC message (%d)\n",
988 __func__, __LINE__, err);
991 static void unlock_all_bitmaps(struct mddev *mddev);
992 static int leave(struct mddev *mddev)
994 struct md_cluster_info *cinfo = mddev->cluster_info;
999 /* BITMAP_NEEDS_SYNC message should be sent when node
1000 * is leaving the cluster with dirty bitmap, also we
1001 * can only deliver it when dlm connection is available */
1002 if (cinfo->slot_number > 0 && mddev->recovery_cp != MaxSector)
1003 resync_bitmap(mddev);
1005 set_bit(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state);
1006 md_unregister_thread(&cinfo->recovery_thread);
1007 md_unregister_thread(&cinfo->recv_thread);
1008 lockres_free(cinfo->message_lockres);
1009 lockres_free(cinfo->token_lockres);
1010 lockres_free(cinfo->ack_lockres);
1011 lockres_free(cinfo->no_new_dev_lockres);
1012 lockres_free(cinfo->resync_lockres);
1013 lockres_free(cinfo->bitmap_lockres);
1014 unlock_all_bitmaps(mddev);
1015 dlm_release_lockspace(cinfo->lockspace, 2);
1020 /* slot_number(): Returns the MD slot number to use
1021 * DLM starts the slot numbers from 1, wheras cluster-md
1022 * wants the number to be from zero, so we deduct one
1024 static int slot_number(struct mddev *mddev)
1026 struct md_cluster_info *cinfo = mddev->cluster_info;
1028 return cinfo->slot_number - 1;
1032 * Check if the communication is already locked, else lock the communication
1034 * If it is already locked, token is in EX mode, and hence lock_token()
1035 * should not be called.
1037 static int metadata_update_start(struct mddev *mddev)
1039 struct md_cluster_info *cinfo = mddev->cluster_info;
1043 * metadata_update_start is always called with the protection of
1044 * reconfig_mutex, so set WAITING_FOR_TOKEN here.
1046 ret = test_and_set_bit_lock(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD,
1049 md_wakeup_thread(mddev->thread);
1051 wait_event(cinfo->wait,
1052 !test_and_set_bit(MD_CLUSTER_SEND_LOCK, &cinfo->state) ||
1053 test_and_clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state));
1055 /* If token is already locked, return 0 */
1056 if (cinfo->token_lockres->mode == DLM_LOCK_EX) {
1057 clear_bit_unlock(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state);
1061 ret = lock_token(cinfo);
1062 clear_bit_unlock(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state);
1066 static int metadata_update_finish(struct mddev *mddev)
1068 struct md_cluster_info *cinfo = mddev->cluster_info;
1069 struct cluster_msg cmsg;
1070 struct md_rdev *rdev;
1074 memset(&cmsg, 0, sizeof(cmsg));
1075 cmsg.type = cpu_to_le32(METADATA_UPDATED);
1076 /* Pick up a good active device number to send.
1078 rdev_for_each(rdev, mddev)
1079 if (rdev->raid_disk > -1 && !test_bit(Faulty, &rdev->flags)) {
1080 raid_slot = rdev->desc_nr;
1083 if (raid_slot >= 0) {
1084 cmsg.raid_slot = cpu_to_le32(raid_slot);
1085 ret = __sendmsg(cinfo, &cmsg);
1087 pr_warn("md-cluster: No good device id found to send\n");
1088 clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1093 static void metadata_update_cancel(struct mddev *mddev)
1095 struct md_cluster_info *cinfo = mddev->cluster_info;
1096 clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1101 * return 0 if all the bitmaps have the same sync_size
1103 int cluster_check_sync_size(struct mddev *mddev)
1107 unsigned long my_sync_size, sync_size = 0;
1108 int node_num = mddev->bitmap_info.nodes;
1109 int current_slot = md_cluster_ops->slot_number(mddev);
1110 struct bitmap *bitmap = mddev->bitmap;
1112 struct dlm_lock_resource *bm_lockres;
1114 sb = kmap_atomic(bitmap->storage.sb_page);
1115 my_sync_size = sb->sync_size;
1118 for (i = 0; i < node_num; i++) {
1119 if (i == current_slot)
1122 bitmap = get_bitmap_from_slot(mddev, i);
1123 if (IS_ERR(bitmap)) {
1124 pr_err("can't get bitmap from slot %d\n", i);
1129 * If we can hold the bitmap lock of one node then
1130 * the slot is not occupied, update the sb.
1132 snprintf(str, 64, "bitmap%04d", i);
1133 bm_lockres = lockres_init(mddev, str, NULL, 1);
1135 pr_err("md-cluster: Cannot initialize %s\n", str);
1136 md_bitmap_free(bitmap);
1139 bm_lockres->flags |= DLM_LKF_NOQUEUE;
1140 rv = dlm_lock_sync(bm_lockres, DLM_LOCK_PW);
1142 bitmap_update_sb(bitmap);
1143 lockres_free(bm_lockres);
1145 sb = kmap_atomic(bitmap->storage.sb_page);
1147 sync_size = sb->sync_size;
1148 else if (sync_size != sb->sync_size) {
1150 md_bitmap_free(bitmap);
1154 md_bitmap_free(bitmap);
1157 return (my_sync_size == sync_size) ? 0 : -1;
1161 * Update the size for cluster raid is a little more complex, we perform it
1163 * 1. hold token lock and update superblock in initiator node.
1164 * 2. send METADATA_UPDATED msg to other nodes.
1165 * 3. The initiator node continues to check each bitmap's sync_size, if all
1166 * bitmaps have the same value of sync_size, then we can set capacity and
1167 * let other nodes to perform it. If one node can't update sync_size
1168 * accordingly, we need to revert to previous value.
1170 static void update_size(struct mddev *mddev, sector_t old_dev_sectors)
1172 struct md_cluster_info *cinfo = mddev->cluster_info;
1173 struct cluster_msg cmsg;
1174 struct md_rdev *rdev;
1178 md_update_sb(mddev, 1);
1179 if (lock_comm(cinfo, 1)) {
1180 pr_err("%s: lock_comm failed\n", __func__);
1184 memset(&cmsg, 0, sizeof(cmsg));
1185 cmsg.type = cpu_to_le32(METADATA_UPDATED);
1186 rdev_for_each(rdev, mddev)
1187 if (rdev->raid_disk >= 0 && !test_bit(Faulty, &rdev->flags)) {
1188 raid_slot = rdev->desc_nr;
1191 if (raid_slot >= 0) {
1192 cmsg.raid_slot = cpu_to_le32(raid_slot);
1194 * We can only change capiticy after all the nodes can do it,
1195 * so need to wait after other nodes already received the msg
1196 * and handled the change
1198 ret = __sendmsg(cinfo, &cmsg);
1200 pr_err("%s:%d: failed to send METADATA_UPDATED msg\n",
1201 __func__, __LINE__);
1206 pr_err("md-cluster: No good device id found to send\n");
1212 * check the sync_size from other node's bitmap, if sync_size
1213 * have already updated in other nodes as expected, send an
1214 * empty metadata msg to permit the change of capacity
1216 if (cluster_check_sync_size(mddev) == 0) {
1217 memset(&cmsg, 0, sizeof(cmsg));
1218 cmsg.type = cpu_to_le32(CHANGE_CAPACITY);
1219 ret = __sendmsg(cinfo, &cmsg);
1221 pr_err("%s:%d: failed to send CHANGE_CAPACITY msg\n",
1222 __func__, __LINE__);
1223 set_capacity(mddev->gendisk, mddev->array_sectors);
1224 revalidate_disk(mddev->gendisk);
1226 /* revert to previous sectors */
1227 ret = mddev->pers->resize(mddev, old_dev_sectors);
1229 revalidate_disk(mddev->gendisk);
1230 ret = __sendmsg(cinfo, &cmsg);
1232 pr_err("%s:%d: failed to send METADATA_UPDATED msg\n",
1233 __func__, __LINE__);
1238 static int resync_start(struct mddev *mddev)
1240 struct md_cluster_info *cinfo = mddev->cluster_info;
1241 return dlm_lock_sync_interruptible(cinfo->resync_lockres, DLM_LOCK_EX, mddev);
1244 static int resync_info_update(struct mddev *mddev, sector_t lo, sector_t hi)
1246 struct md_cluster_info *cinfo = mddev->cluster_info;
1247 struct resync_info ri;
1248 struct cluster_msg cmsg = {0};
1250 /* do not send zero again, if we have sent before */
1252 memcpy(&ri, cinfo->bitmap_lockres->lksb.sb_lvbptr, sizeof(struct resync_info));
1253 if (le64_to_cpu(ri.hi) == 0)
1257 add_resync_info(cinfo->bitmap_lockres, lo, hi);
1258 /* Re-acquire the lock to refresh LVB */
1259 dlm_lock_sync(cinfo->bitmap_lockres, DLM_LOCK_PW);
1260 cmsg.type = cpu_to_le32(RESYNCING);
1261 cmsg.low = cpu_to_le64(lo);
1262 cmsg.high = cpu_to_le64(hi);
1265 * mddev_lock is held if resync_info_update is called from
1266 * resync_finish (md_reap_sync_thread -> resync_finish)
1268 if (lo == 0 && hi == 0)
1269 return sendmsg(cinfo, &cmsg, 1);
1271 return sendmsg(cinfo, &cmsg, 0);
1274 static int resync_finish(struct mddev *mddev)
1276 struct md_cluster_info *cinfo = mddev->cluster_info;
1277 dlm_unlock_sync(cinfo->resync_lockres);
1278 return resync_info_update(mddev, 0, 0);
1281 static int area_resyncing(struct mddev *mddev, int direction,
1282 sector_t lo, sector_t hi)
1284 struct md_cluster_info *cinfo = mddev->cluster_info;
1286 struct suspend_info *s;
1288 if ((direction == READ) &&
1289 test_bit(MD_CLUSTER_SUSPEND_READ_BALANCING, &cinfo->state))
1292 spin_lock_irq(&cinfo->suspend_lock);
1293 if (list_empty(&cinfo->suspend_list))
1295 list_for_each_entry(s, &cinfo->suspend_list, list)
1296 if (hi > s->lo && lo < s->hi) {
1301 spin_unlock_irq(&cinfo->suspend_lock);
1305 /* add_new_disk() - initiates a disk add
1306 * However, if this fails before writing md_update_sb(),
1307 * add_new_disk_cancel() must be called to release token lock
1309 static int add_new_disk(struct mddev *mddev, struct md_rdev *rdev)
1311 struct md_cluster_info *cinfo = mddev->cluster_info;
1312 struct cluster_msg cmsg;
1314 struct mdp_superblock_1 *sb = page_address(rdev->sb_page);
1315 char *uuid = sb->device_uuid;
1317 memset(&cmsg, 0, sizeof(cmsg));
1318 cmsg.type = cpu_to_le32(NEWDISK);
1319 memcpy(cmsg.uuid, uuid, 16);
1320 cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
1321 if (lock_comm(cinfo, 1))
1323 ret = __sendmsg(cinfo, &cmsg);
1328 cinfo->no_new_dev_lockres->flags |= DLM_LKF_NOQUEUE;
1329 ret = dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_EX);
1330 cinfo->no_new_dev_lockres->flags &= ~DLM_LKF_NOQUEUE;
1331 /* Some node does not "see" the device */
1337 dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR);
1338 /* Since MD_CHANGE_DEVS will be set in add_bound_rdev which
1339 * will run soon after add_new_disk, the below path will be
1341 * md_wakeup_thread(mddev->thread)
1342 * -> conf->thread (raid1d)
1343 * -> md_check_recovery -> md_update_sb
1344 * -> metadata_update_start/finish
1345 * MD_CLUSTER_SEND_LOCKED_ALREADY will be cleared eventually.
1347 * For other failure cases, metadata_update_cancel and
1348 * add_new_disk_cancel also clear below bit as well.
1350 set_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1351 wake_up(&cinfo->wait);
1356 static void add_new_disk_cancel(struct mddev *mddev)
1358 struct md_cluster_info *cinfo = mddev->cluster_info;
1359 clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1363 static int new_disk_ack(struct mddev *mddev, bool ack)
1365 struct md_cluster_info *cinfo = mddev->cluster_info;
1367 if (!test_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state)) {
1368 pr_warn("md-cluster(%s): Spurious cluster confirmation\n", mdname(mddev));
1373 dlm_unlock_sync(cinfo->no_new_dev_lockres);
1374 complete(&cinfo->newdisk_completion);
1378 static int remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1380 struct cluster_msg cmsg = {0};
1381 struct md_cluster_info *cinfo = mddev->cluster_info;
1382 cmsg.type = cpu_to_le32(REMOVE);
1383 cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
1384 return sendmsg(cinfo, &cmsg, 1);
1387 static int lock_all_bitmaps(struct mddev *mddev)
1389 int slot, my_slot, ret, held = 1, i = 0;
1391 struct md_cluster_info *cinfo = mddev->cluster_info;
1393 cinfo->other_bitmap_lockres = kzalloc((mddev->bitmap_info.nodes - 1) *
1394 sizeof(struct dlm_lock_resource *),
1396 if (!cinfo->other_bitmap_lockres) {
1397 pr_err("md: can't alloc mem for other bitmap locks\n");
1401 my_slot = slot_number(mddev);
1402 for (slot = 0; slot < mddev->bitmap_info.nodes; slot++) {
1403 if (slot == my_slot)
1406 memset(str, '\0', 64);
1407 snprintf(str, 64, "bitmap%04d", slot);
1408 cinfo->other_bitmap_lockres[i] = lockres_init(mddev, str, NULL, 1);
1409 if (!cinfo->other_bitmap_lockres[i])
1412 cinfo->other_bitmap_lockres[i]->flags |= DLM_LKF_NOQUEUE;
1413 ret = dlm_lock_sync(cinfo->other_bitmap_lockres[i], DLM_LOCK_PW);
1422 static void unlock_all_bitmaps(struct mddev *mddev)
1424 struct md_cluster_info *cinfo = mddev->cluster_info;
1427 /* release other node's bitmap lock if they are existed */
1428 if (cinfo->other_bitmap_lockres) {
1429 for (i = 0; i < mddev->bitmap_info.nodes - 1; i++) {
1430 if (cinfo->other_bitmap_lockres[i]) {
1431 lockres_free(cinfo->other_bitmap_lockres[i]);
1434 kfree(cinfo->other_bitmap_lockres);
1435 cinfo->other_bitmap_lockres = NULL;
1439 static int gather_bitmaps(struct md_rdev *rdev)
1443 struct cluster_msg cmsg = {0};
1444 struct mddev *mddev = rdev->mddev;
1445 struct md_cluster_info *cinfo = mddev->cluster_info;
1447 cmsg.type = cpu_to_le32(RE_ADD);
1448 cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
1449 err = sendmsg(cinfo, &cmsg, 1);
1453 for (sn = 0; sn < mddev->bitmap_info.nodes; sn++) {
1454 if (sn == (cinfo->slot_number - 1))
1456 err = bitmap_copy_from_slot(mddev, sn, &lo, &hi, false);
1458 pr_warn("md-cluster: Could not gather bitmaps from slot %d", sn);
1461 if ((hi > 0) && (lo < mddev->recovery_cp))
1462 mddev->recovery_cp = lo;
1468 static struct md_cluster_operations cluster_ops = {
1471 .slot_number = slot_number,
1472 .resync_start = resync_start,
1473 .resync_finish = resync_finish,
1474 .resync_info_update = resync_info_update,
1475 .metadata_update_start = metadata_update_start,
1476 .metadata_update_finish = metadata_update_finish,
1477 .metadata_update_cancel = metadata_update_cancel,
1478 .area_resyncing = area_resyncing,
1479 .add_new_disk = add_new_disk,
1480 .add_new_disk_cancel = add_new_disk_cancel,
1481 .new_disk_ack = new_disk_ack,
1482 .remove_disk = remove_disk,
1483 .load_bitmaps = load_bitmaps,
1484 .gather_bitmaps = gather_bitmaps,
1485 .lock_all_bitmaps = lock_all_bitmaps,
1486 .unlock_all_bitmaps = unlock_all_bitmaps,
1487 .update_size = update_size,
1490 static int __init cluster_init(void)
1492 pr_warn("md-cluster: EXPERIMENTAL. Use with caution\n");
1493 pr_info("Registering Cluster MD functions\n");
1494 register_md_cluster_operations(&cluster_ops, THIS_MODULE);
1498 static void cluster_exit(void)
1500 unregister_md_cluster_operations();
1503 module_init(cluster_init);
1504 module_exit(cluster_exit);
1505 MODULE_AUTHOR("SUSE");
1506 MODULE_LICENSE("GPL");
1507 MODULE_DESCRIPTION("Clustering support for MD");