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
72 struct md_cluster_info {
73 /* dlm lock space and resources for clustered raid. */
74 dlm_lockspace_t *lockspace;
76 struct completion completion;
77 struct mutex recv_mutex;
78 struct dlm_lock_resource *bitmap_lockres;
79 struct dlm_lock_resource **other_bitmap_lockres;
80 struct dlm_lock_resource *resync_lockres;
81 struct list_head suspend_list;
82 spinlock_t suspend_lock;
83 struct md_thread *recovery_thread;
84 unsigned long recovery_map;
85 /* communication loc resources */
86 struct dlm_lock_resource *ack_lockres;
87 struct dlm_lock_resource *message_lockres;
88 struct dlm_lock_resource *token_lockres;
89 struct dlm_lock_resource *no_new_dev_lockres;
90 struct md_thread *recv_thread;
91 struct completion newdisk_completion;
92 wait_queue_head_t wait;
94 /* record the region in RESYNCING message */
100 METADATA_UPDATED = 0,
111 /* TODO: Unionize this for smaller footprint */
118 static void sync_ast(void *arg)
120 struct dlm_lock_resource *res;
123 res->sync_locking_done = true;
124 wake_up(&res->sync_locking);
127 static int dlm_lock_sync(struct dlm_lock_resource *res, int mode)
131 ret = dlm_lock(res->ls, mode, &res->lksb,
132 res->flags, res->name, strlen(res->name),
133 0, sync_ast, res, res->bast);
136 wait_event(res->sync_locking, res->sync_locking_done);
137 res->sync_locking_done = false;
138 if (res->lksb.sb_status == 0)
140 return res->lksb.sb_status;
143 static int dlm_unlock_sync(struct dlm_lock_resource *res)
145 return dlm_lock_sync(res, DLM_LOCK_NL);
149 * An variation of dlm_lock_sync, which make lock request could
152 static int dlm_lock_sync_interruptible(struct dlm_lock_resource *res, int mode,
157 ret = dlm_lock(res->ls, mode, &res->lksb,
158 res->flags, res->name, strlen(res->name),
159 0, sync_ast, res, res->bast);
163 wait_event(res->sync_locking, res->sync_locking_done
164 || kthread_should_stop()
165 || test_bit(MD_CLOSING, &mddev->flags));
166 if (!res->sync_locking_done) {
168 * the convert queue contains the lock request when request is
169 * interrupted, and sync_ast could still be run, so need to
170 * cancel the request and reset completion
172 ret = dlm_unlock(res->ls, res->lksb.sb_lkid, DLM_LKF_CANCEL,
174 res->sync_locking_done = false;
175 if (unlikely(ret != 0))
176 pr_info("failed to cancel previous lock request "
177 "%s return %d\n", res->name, ret);
180 res->sync_locking_done = false;
181 if (res->lksb.sb_status == 0)
183 return res->lksb.sb_status;
186 static struct dlm_lock_resource *lockres_init(struct mddev *mddev,
187 char *name, void (*bastfn)(void *arg, int mode), int with_lvb)
189 struct dlm_lock_resource *res = NULL;
191 struct md_cluster_info *cinfo = mddev->cluster_info;
193 res = kzalloc(sizeof(struct dlm_lock_resource), GFP_KERNEL);
196 init_waitqueue_head(&res->sync_locking);
197 res->sync_locking_done = false;
198 res->ls = cinfo->lockspace;
200 res->mode = DLM_LOCK_IV;
201 namelen = strlen(name);
202 res->name = kzalloc(namelen + 1, GFP_KERNEL);
204 pr_err("md-cluster: Unable to allocate resource name for resource %s\n", name);
207 strlcpy(res->name, name, namelen + 1);
209 res->lksb.sb_lvbptr = kzalloc(LVB_SIZE, GFP_KERNEL);
210 if (!res->lksb.sb_lvbptr) {
211 pr_err("md-cluster: Unable to allocate LVB for resource %s\n", name);
214 res->flags = DLM_LKF_VALBLK;
220 res->flags |= DLM_LKF_EXPEDITE;
222 ret = dlm_lock_sync(res, DLM_LOCK_NL);
224 pr_err("md-cluster: Unable to lock NL on new lock resource %s\n", name);
227 res->flags &= ~DLM_LKF_EXPEDITE;
228 res->flags |= DLM_LKF_CONVERT;
232 kfree(res->lksb.sb_lvbptr);
238 static void lockres_free(struct dlm_lock_resource *res)
246 * use FORCEUNLOCK flag, so we can unlock even the lock is on the
247 * waiting or convert queue
249 ret = dlm_unlock(res->ls, res->lksb.sb_lkid, DLM_LKF_FORCEUNLOCK,
251 if (unlikely(ret != 0))
252 pr_err("failed to unlock %s return %d\n", res->name, ret);
254 wait_event(res->sync_locking, res->sync_locking_done);
257 kfree(res->lksb.sb_lvbptr);
261 static void add_resync_info(struct dlm_lock_resource *lockres,
262 sector_t lo, sector_t hi)
264 struct resync_info *ri;
266 ri = (struct resync_info *)lockres->lksb.sb_lvbptr;
267 ri->lo = cpu_to_le64(lo);
268 ri->hi = cpu_to_le64(hi);
271 static struct suspend_info *read_resync_info(struct mddev *mddev, struct dlm_lock_resource *lockres)
273 struct resync_info ri;
274 struct suspend_info *s = NULL;
277 dlm_lock_sync(lockres, DLM_LOCK_CR);
278 memcpy(&ri, lockres->lksb.sb_lvbptr, sizeof(struct resync_info));
279 hi = le64_to_cpu(ri.hi);
281 s = kzalloc(sizeof(struct suspend_info), GFP_KERNEL);
285 s->lo = le64_to_cpu(ri.lo);
287 dlm_unlock_sync(lockres);
292 static void recover_bitmaps(struct md_thread *thread)
294 struct mddev *mddev = thread->mddev;
295 struct md_cluster_info *cinfo = mddev->cluster_info;
296 struct dlm_lock_resource *bm_lockres;
299 struct suspend_info *s, *tmp;
302 while (cinfo->recovery_map) {
303 slot = fls64((u64)cinfo->recovery_map) - 1;
305 snprintf(str, 64, "bitmap%04d", slot);
306 bm_lockres = lockres_init(mddev, str, NULL, 1);
308 pr_err("md-cluster: Cannot initialize bitmaps\n");
312 ret = dlm_lock_sync_interruptible(bm_lockres, DLM_LOCK_PW, mddev);
314 pr_err("md-cluster: Could not DLM lock %s: %d\n",
318 ret = bitmap_copy_from_slot(mddev, slot, &lo, &hi, true);
320 pr_err("md-cluster: Could not copy data from bitmap %d\n", slot);
324 /* Clear suspend_area associated with the bitmap */
325 spin_lock_irq(&cinfo->suspend_lock);
326 list_for_each_entry_safe(s, tmp, &cinfo->suspend_list, list)
327 if (slot == s->slot) {
331 spin_unlock_irq(&cinfo->suspend_lock);
334 if (lo < mddev->recovery_cp)
335 mddev->recovery_cp = lo;
336 /* wake up thread to continue resync in case resync
338 if (mddev->recovery_cp != MaxSector) {
339 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
340 md_wakeup_thread(mddev->thread);
344 lockres_free(bm_lockres);
345 clear_bit(slot, &cinfo->recovery_map);
349 static void recover_prep(void *arg)
351 struct mddev *mddev = arg;
352 struct md_cluster_info *cinfo = mddev->cluster_info;
353 set_bit(MD_CLUSTER_SUSPEND_READ_BALANCING, &cinfo->state);
356 static void __recover_slot(struct mddev *mddev, int slot)
358 struct md_cluster_info *cinfo = mddev->cluster_info;
360 set_bit(slot, &cinfo->recovery_map);
361 if (!cinfo->recovery_thread) {
362 cinfo->recovery_thread = md_register_thread(recover_bitmaps,
364 if (!cinfo->recovery_thread) {
365 pr_warn("md-cluster: Could not create recovery thread\n");
369 md_wakeup_thread(cinfo->recovery_thread);
372 static void recover_slot(void *arg, struct dlm_slot *slot)
374 struct mddev *mddev = arg;
375 struct md_cluster_info *cinfo = mddev->cluster_info;
377 pr_info("md-cluster: %s Node %d/%d down. My slot: %d. Initiating recovery.\n",
378 mddev->bitmap_info.cluster_name,
379 slot->nodeid, slot->slot,
381 /* deduct one since dlm slot starts from one while the num of
382 * cluster-md begins with 0 */
383 __recover_slot(mddev, slot->slot - 1);
386 static void recover_done(void *arg, struct dlm_slot *slots,
387 int num_slots, int our_slot,
390 struct mddev *mddev = arg;
391 struct md_cluster_info *cinfo = mddev->cluster_info;
393 cinfo->slot_number = our_slot;
394 /* completion is only need to be complete when node join cluster,
395 * it doesn't need to run during another node's failure */
396 if (test_bit(MD_CLUSTER_BEGIN_JOIN_CLUSTER, &cinfo->state)) {
397 complete(&cinfo->completion);
398 clear_bit(MD_CLUSTER_BEGIN_JOIN_CLUSTER, &cinfo->state);
400 clear_bit(MD_CLUSTER_SUSPEND_READ_BALANCING, &cinfo->state);
403 /* the ops is called when node join the cluster, and do lock recovery
404 * if node failure occurs */
405 static const struct dlm_lockspace_ops md_ls_ops = {
406 .recover_prep = recover_prep,
407 .recover_slot = recover_slot,
408 .recover_done = recover_done,
412 * The BAST function for the ack lock resource
413 * This function wakes up the receive thread in
414 * order to receive and process the message.
416 static void ack_bast(void *arg, int mode)
418 struct dlm_lock_resource *res = arg;
419 struct md_cluster_info *cinfo = res->mddev->cluster_info;
421 if (mode == DLM_LOCK_EX) {
422 if (test_bit(MD_CLUSTER_ALREADY_IN_CLUSTER, &cinfo->state))
423 md_wakeup_thread(cinfo->recv_thread);
425 set_bit(MD_CLUSTER_PENDING_RECV_EVENT, &cinfo->state);
429 static void __remove_suspend_info(struct md_cluster_info *cinfo, int slot)
431 struct suspend_info *s, *tmp;
433 list_for_each_entry_safe(s, tmp, &cinfo->suspend_list, list)
434 if (slot == s->slot) {
441 static void remove_suspend_info(struct mddev *mddev, int slot)
443 struct md_cluster_info *cinfo = mddev->cluster_info;
444 spin_lock_irq(&cinfo->suspend_lock);
445 __remove_suspend_info(cinfo, slot);
446 spin_unlock_irq(&cinfo->suspend_lock);
447 mddev->pers->quiesce(mddev, 2);
451 static void process_suspend_info(struct mddev *mddev,
452 int slot, sector_t lo, sector_t hi)
454 struct md_cluster_info *cinfo = mddev->cluster_info;
455 struct suspend_info *s;
458 remove_suspend_info(mddev, slot);
459 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
460 md_wakeup_thread(mddev->thread);
465 * The bitmaps are not same for different nodes
466 * if RESYNCING is happening in one node, then
467 * the node which received the RESYNCING message
468 * probably will perform resync with the region
469 * [lo, hi] again, so we could reduce resync time
470 * a lot if we can ensure that the bitmaps among
471 * different nodes are match up well.
473 * sync_low/hi is used to record the region which
474 * arrived in the previous RESYNCING message,
476 * Call bitmap_sync_with_cluster to clear
477 * NEEDED_MASK and set RESYNC_MASK since
478 * resync thread is running in another node,
479 * so we don't need to do the resync again
480 * with the same section */
481 bitmap_sync_with_cluster(mddev, cinfo->sync_low,
484 cinfo->sync_low = lo;
487 s = kzalloc(sizeof(struct suspend_info), GFP_KERNEL);
493 mddev->pers->quiesce(mddev, 1);
494 mddev->pers->quiesce(mddev, 0);
495 spin_lock_irq(&cinfo->suspend_lock);
496 /* Remove existing entry (if exists) before adding */
497 __remove_suspend_info(cinfo, slot);
498 list_add(&s->list, &cinfo->suspend_list);
499 spin_unlock_irq(&cinfo->suspend_lock);
500 mddev->pers->quiesce(mddev, 2);
503 static void process_add_new_disk(struct mddev *mddev, struct cluster_msg *cmsg)
506 struct md_cluster_info *cinfo = mddev->cluster_info;
507 char event_name[] = "EVENT=ADD_DEVICE";
509 char *envp[] = {event_name, disk_uuid, raid_slot, NULL};
512 len = snprintf(disk_uuid, 64, "DEVICE_UUID=");
513 sprintf(disk_uuid + len, "%pU", cmsg->uuid);
514 snprintf(raid_slot, 16, "RAID_DISK=%d", le32_to_cpu(cmsg->raid_slot));
515 pr_info("%s:%d Sending kobject change with %s and %s\n", __func__, __LINE__, disk_uuid, raid_slot);
516 init_completion(&cinfo->newdisk_completion);
517 set_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state);
518 kobject_uevent_env(&disk_to_dev(mddev->gendisk)->kobj, KOBJ_CHANGE, envp);
519 wait_for_completion_timeout(&cinfo->newdisk_completion,
521 clear_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state);
525 static void process_metadata_update(struct mddev *mddev, struct cluster_msg *msg)
527 struct md_cluster_info *cinfo = mddev->cluster_info;
528 mddev->good_device_nr = le32_to_cpu(msg->raid_slot);
529 set_bit(MD_RELOAD_SB, &mddev->flags);
530 dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR);
531 md_wakeup_thread(mddev->thread);
534 static void process_remove_disk(struct mddev *mddev, struct cluster_msg *msg)
536 struct md_rdev *rdev;
539 rdev = md_find_rdev_nr_rcu(mddev, le32_to_cpu(msg->raid_slot));
541 set_bit(ClusterRemove, &rdev->flags);
542 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
543 md_wakeup_thread(mddev->thread);
546 pr_warn("%s: %d Could not find disk(%d) to REMOVE\n",
547 __func__, __LINE__, le32_to_cpu(msg->raid_slot));
551 static void process_readd_disk(struct mddev *mddev, struct cluster_msg *msg)
553 struct md_rdev *rdev;
556 rdev = md_find_rdev_nr_rcu(mddev, le32_to_cpu(msg->raid_slot));
557 if (rdev && test_bit(Faulty, &rdev->flags))
558 clear_bit(Faulty, &rdev->flags);
560 pr_warn("%s: %d Could not find disk(%d) which is faulty",
561 __func__, __LINE__, le32_to_cpu(msg->raid_slot));
565 static int process_recvd_msg(struct mddev *mddev, struct cluster_msg *msg)
569 if (WARN(mddev->cluster_info->slot_number - 1 == le32_to_cpu(msg->slot),
570 "node %d received it's own msg\n", le32_to_cpu(msg->slot)))
572 switch (le32_to_cpu(msg->type)) {
573 case METADATA_UPDATED:
574 process_metadata_update(mddev, msg);
577 process_suspend_info(mddev, le32_to_cpu(msg->slot),
578 le64_to_cpu(msg->low),
579 le64_to_cpu(msg->high));
582 process_add_new_disk(mddev, msg);
585 process_remove_disk(mddev, msg);
588 process_readd_disk(mddev, msg);
590 case BITMAP_NEEDS_SYNC:
591 __recover_slot(mddev, le32_to_cpu(msg->slot));
595 pr_warn("%s:%d Received unknown message from %d\n",
596 __func__, __LINE__, msg->slot);
602 * thread for receiving message
604 static void recv_daemon(struct md_thread *thread)
606 struct md_cluster_info *cinfo = thread->mddev->cluster_info;
607 struct dlm_lock_resource *ack_lockres = cinfo->ack_lockres;
608 struct dlm_lock_resource *message_lockres = cinfo->message_lockres;
609 struct cluster_msg msg;
612 mutex_lock(&cinfo->recv_mutex);
613 /*get CR on Message*/
614 if (dlm_lock_sync(message_lockres, DLM_LOCK_CR)) {
615 pr_err("md/raid1:failed to get CR on MESSAGE\n");
616 mutex_unlock(&cinfo->recv_mutex);
620 /* read lvb and wake up thread to process this message_lockres */
621 memcpy(&msg, message_lockres->lksb.sb_lvbptr, sizeof(struct cluster_msg));
622 ret = process_recvd_msg(thread->mddev, &msg);
626 /*release CR on ack_lockres*/
627 ret = dlm_unlock_sync(ack_lockres);
628 if (unlikely(ret != 0))
629 pr_info("unlock ack failed return %d\n", ret);
630 /*up-convert to PR on message_lockres*/
631 ret = dlm_lock_sync(message_lockres, DLM_LOCK_PR);
632 if (unlikely(ret != 0))
633 pr_info("lock PR on msg failed return %d\n", ret);
634 /*get CR on ack_lockres again*/
635 ret = dlm_lock_sync(ack_lockres, DLM_LOCK_CR);
636 if (unlikely(ret != 0))
637 pr_info("lock CR on ack failed return %d\n", ret);
639 /*release CR on message_lockres*/
640 ret = dlm_unlock_sync(message_lockres);
641 if (unlikely(ret != 0))
642 pr_info("unlock msg failed return %d\n", ret);
643 mutex_unlock(&cinfo->recv_mutex);
647 * Takes the lock on the TOKEN lock resource so no other
648 * node can communicate while the operation is underway.
650 static int lock_token(struct md_cluster_info *cinfo)
654 error = dlm_lock_sync(cinfo->token_lockres, DLM_LOCK_EX);
656 pr_err("md-cluster(%s:%d): failed to get EX on TOKEN (%d)\n",
657 __func__, __LINE__, error);
659 /* Lock the receive sequence */
660 mutex_lock(&cinfo->recv_mutex);
665 * Sets the MD_CLUSTER_SEND_LOCK bit to lock the send channel.
667 static int lock_comm(struct md_cluster_info *cinfo)
669 wait_event(cinfo->wait,
670 !test_and_set_bit(MD_CLUSTER_SEND_LOCK, &cinfo->state));
672 return lock_token(cinfo);
675 static void unlock_comm(struct md_cluster_info *cinfo)
677 WARN_ON(cinfo->token_lockres->mode != DLM_LOCK_EX);
678 mutex_unlock(&cinfo->recv_mutex);
679 dlm_unlock_sync(cinfo->token_lockres);
680 clear_bit(MD_CLUSTER_SEND_LOCK, &cinfo->state);
681 wake_up(&cinfo->wait);
685 * This function performs the actual sending of the message. This function is
686 * usually called after performing the encompassing operation
688 * 1. Grabs the message lockresource in EX mode
689 * 2. Copies the message to the message LVB
690 * 3. Downconverts message lockresource to CW
691 * 4. Upconverts ack lock resource from CR to EX. This forces the BAST on other nodes
692 * and the other nodes read the message. The thread will wait here until all other
693 * nodes have released ack lock resource.
694 * 5. Downconvert ack lockresource to CR
696 static int __sendmsg(struct md_cluster_info *cinfo, struct cluster_msg *cmsg)
699 int slot = cinfo->slot_number - 1;
701 cmsg->slot = cpu_to_le32(slot);
702 /*get EX on Message*/
703 error = dlm_lock_sync(cinfo->message_lockres, DLM_LOCK_EX);
705 pr_err("md-cluster: failed to get EX on MESSAGE (%d)\n", error);
709 memcpy(cinfo->message_lockres->lksb.sb_lvbptr, (void *)cmsg,
710 sizeof(struct cluster_msg));
711 /*down-convert EX to CW on Message*/
712 error = dlm_lock_sync(cinfo->message_lockres, DLM_LOCK_CW);
714 pr_err("md-cluster: failed to convert EX to CW on MESSAGE(%d)\n",
719 /*up-convert CR to EX on Ack*/
720 error = dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_EX);
722 pr_err("md-cluster: failed to convert CR to EX on ACK(%d)\n",
727 /*down-convert EX to CR on Ack*/
728 error = dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_CR);
730 pr_err("md-cluster: failed to convert EX to CR on ACK(%d)\n",
736 error = dlm_unlock_sync(cinfo->message_lockres);
737 if (unlikely(error != 0)) {
738 pr_err("md-cluster: failed convert to NL on MESSAGE(%d)\n",
740 /* in case the message can't be released due to some reason */
747 static int sendmsg(struct md_cluster_info *cinfo, struct cluster_msg *cmsg)
752 ret = __sendmsg(cinfo, cmsg);
757 static int gather_all_resync_info(struct mddev *mddev, int total_slots)
759 struct md_cluster_info *cinfo = mddev->cluster_info;
761 struct dlm_lock_resource *bm_lockres;
762 struct suspend_info *s;
767 for (i = 0; i < total_slots; i++) {
768 memset(str, '\0', 64);
769 snprintf(str, 64, "bitmap%04d", i);
770 bm_lockres = lockres_init(mddev, str, NULL, 1);
773 if (i == (cinfo->slot_number - 1)) {
774 lockres_free(bm_lockres);
778 bm_lockres->flags |= DLM_LKF_NOQUEUE;
779 ret = dlm_lock_sync(bm_lockres, DLM_LOCK_PW);
780 if (ret == -EAGAIN) {
781 memset(bm_lockres->lksb.sb_lvbptr, '\0', LVB_SIZE);
782 s = read_resync_info(mddev, bm_lockres);
784 pr_info("%s:%d Resync[%llu..%llu] in progress on %d\n",
786 (unsigned long long) s->lo,
787 (unsigned long long) s->hi, i);
788 spin_lock_irq(&cinfo->suspend_lock);
790 list_add(&s->list, &cinfo->suspend_list);
791 spin_unlock_irq(&cinfo->suspend_lock);
794 lockres_free(bm_lockres);
798 lockres_free(bm_lockres);
802 /* Read the disk bitmap sb and check if it needs recovery */
803 ret = bitmap_copy_from_slot(mddev, i, &lo, &hi, false);
805 pr_warn("md-cluster: Could not gather bitmaps from slot %d", i);
806 lockres_free(bm_lockres);
809 if ((hi > 0) && (lo < mddev->recovery_cp)) {
810 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
811 mddev->recovery_cp = lo;
812 md_check_recovery(mddev);
815 lockres_free(bm_lockres);
821 static int join(struct mddev *mddev, int nodes)
823 struct md_cluster_info *cinfo;
827 cinfo = kzalloc(sizeof(struct md_cluster_info), GFP_KERNEL);
831 INIT_LIST_HEAD(&cinfo->suspend_list);
832 spin_lock_init(&cinfo->suspend_lock);
833 init_completion(&cinfo->completion);
834 set_bit(MD_CLUSTER_BEGIN_JOIN_CLUSTER, &cinfo->state);
835 init_waitqueue_head(&cinfo->wait);
836 mutex_init(&cinfo->recv_mutex);
838 mddev->cluster_info = cinfo;
841 sprintf(str, "%pU", mddev->uuid);
842 ret = dlm_new_lockspace(str, mddev->bitmap_info.cluster_name,
843 DLM_LSFL_FS, LVB_SIZE,
844 &md_ls_ops, mddev, &ops_rv, &cinfo->lockspace);
847 wait_for_completion(&cinfo->completion);
848 if (nodes < cinfo->slot_number) {
849 pr_err("md-cluster: Slot allotted(%d) is greater than available slots(%d).",
850 cinfo->slot_number, nodes);
854 /* Initiate the communication resources */
856 cinfo->recv_thread = md_register_thread(recv_daemon, mddev, "cluster_recv");
857 if (!cinfo->recv_thread) {
858 pr_err("md-cluster: cannot allocate memory for recv_thread!\n");
861 cinfo->message_lockres = lockres_init(mddev, "message", NULL, 1);
862 if (!cinfo->message_lockres)
864 cinfo->token_lockres = lockres_init(mddev, "token", NULL, 0);
865 if (!cinfo->token_lockres)
867 cinfo->no_new_dev_lockres = lockres_init(mddev, "no-new-dev", NULL, 0);
868 if (!cinfo->no_new_dev_lockres)
871 ret = dlm_lock_sync(cinfo->token_lockres, DLM_LOCK_EX);
874 pr_err("md-cluster: can't join cluster to avoid lock issue\n");
877 cinfo->ack_lockres = lockres_init(mddev, "ack", ack_bast, 0);
878 if (!cinfo->ack_lockres) {
882 /* get sync CR lock on ACK. */
883 if (dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_CR))
884 pr_err("md-cluster: failed to get a sync CR lock on ACK!(%d)\n",
886 dlm_unlock_sync(cinfo->token_lockres);
887 /* get sync CR lock on no-new-dev. */
888 if (dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR))
889 pr_err("md-cluster: failed to get a sync CR lock on no-new-dev!(%d)\n", ret);
892 pr_info("md-cluster: Joined cluster %s slot %d\n", str, cinfo->slot_number);
893 snprintf(str, 64, "bitmap%04d", cinfo->slot_number - 1);
894 cinfo->bitmap_lockres = lockres_init(mddev, str, NULL, 1);
895 if (!cinfo->bitmap_lockres) {
899 if (dlm_lock_sync(cinfo->bitmap_lockres, DLM_LOCK_PW)) {
900 pr_err("Failed to get bitmap lock\n");
905 cinfo->resync_lockres = lockres_init(mddev, "resync", NULL, 0);
906 if (!cinfo->resync_lockres) {
913 md_unregister_thread(&cinfo->recovery_thread);
914 md_unregister_thread(&cinfo->recv_thread);
915 lockres_free(cinfo->message_lockres);
916 lockres_free(cinfo->token_lockres);
917 lockres_free(cinfo->ack_lockres);
918 lockres_free(cinfo->no_new_dev_lockres);
919 lockres_free(cinfo->resync_lockres);
920 lockres_free(cinfo->bitmap_lockres);
921 if (cinfo->lockspace)
922 dlm_release_lockspace(cinfo->lockspace, 2);
923 mddev->cluster_info = NULL;
928 static void load_bitmaps(struct mddev *mddev, int total_slots)
930 struct md_cluster_info *cinfo = mddev->cluster_info;
932 /* load all the node's bitmap info for resync */
933 if (gather_all_resync_info(mddev, total_slots))
934 pr_err("md-cluster: failed to gather all resyn infos\n");
935 set_bit(MD_CLUSTER_ALREADY_IN_CLUSTER, &cinfo->state);
936 /* wake up recv thread in case something need to be handled */
937 if (test_and_clear_bit(MD_CLUSTER_PENDING_RECV_EVENT, &cinfo->state))
938 md_wakeup_thread(cinfo->recv_thread);
941 static void resync_bitmap(struct mddev *mddev)
943 struct md_cluster_info *cinfo = mddev->cluster_info;
944 struct cluster_msg cmsg = {0};
947 cmsg.type = cpu_to_le32(BITMAP_NEEDS_SYNC);
948 err = sendmsg(cinfo, &cmsg);
950 pr_err("%s:%d: failed to send BITMAP_NEEDS_SYNC message (%d)\n",
951 __func__, __LINE__, err);
954 static void unlock_all_bitmaps(struct mddev *mddev);
955 static int leave(struct mddev *mddev)
957 struct md_cluster_info *cinfo = mddev->cluster_info;
962 /* BITMAP_NEEDS_SYNC message should be sent when node
963 * is leaving the cluster with dirty bitmap, also we
964 * can only deliver it when dlm connection is available */
965 if (cinfo->slot_number > 0 && mddev->recovery_cp != MaxSector)
966 resync_bitmap(mddev);
968 md_unregister_thread(&cinfo->recovery_thread);
969 md_unregister_thread(&cinfo->recv_thread);
970 lockres_free(cinfo->message_lockres);
971 lockres_free(cinfo->token_lockres);
972 lockres_free(cinfo->ack_lockres);
973 lockres_free(cinfo->no_new_dev_lockres);
974 lockres_free(cinfo->resync_lockres);
975 lockres_free(cinfo->bitmap_lockres);
976 unlock_all_bitmaps(mddev);
977 dlm_release_lockspace(cinfo->lockspace, 2);
982 /* slot_number(): Returns the MD slot number to use
983 * DLM starts the slot numbers from 1, wheras cluster-md
984 * wants the number to be from zero, so we deduct one
986 static int slot_number(struct mddev *mddev)
988 struct md_cluster_info *cinfo = mddev->cluster_info;
990 return cinfo->slot_number - 1;
994 * Check if the communication is already locked, else lock the communication
996 * If it is already locked, token is in EX mode, and hence lock_token()
997 * should not be called.
999 static int metadata_update_start(struct mddev *mddev)
1001 struct md_cluster_info *cinfo = mddev->cluster_info;
1003 wait_event(cinfo->wait,
1004 !test_and_set_bit(MD_CLUSTER_SEND_LOCK, &cinfo->state) ||
1005 test_and_clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state));
1007 /* If token is already locked, return 0 */
1008 if (cinfo->token_lockres->mode == DLM_LOCK_EX)
1011 return lock_token(cinfo);
1014 static int metadata_update_finish(struct mddev *mddev)
1016 struct md_cluster_info *cinfo = mddev->cluster_info;
1017 struct cluster_msg cmsg;
1018 struct md_rdev *rdev;
1022 memset(&cmsg, 0, sizeof(cmsg));
1023 cmsg.type = cpu_to_le32(METADATA_UPDATED);
1024 /* Pick up a good active device number to send.
1026 rdev_for_each(rdev, mddev)
1027 if (rdev->raid_disk > -1 && !test_bit(Faulty, &rdev->flags)) {
1028 raid_slot = rdev->desc_nr;
1031 if (raid_slot >= 0) {
1032 cmsg.raid_slot = cpu_to_le32(raid_slot);
1033 ret = __sendmsg(cinfo, &cmsg);
1035 pr_warn("md-cluster: No good device id found to send\n");
1036 clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1041 static void metadata_update_cancel(struct mddev *mddev)
1043 struct md_cluster_info *cinfo = mddev->cluster_info;
1044 clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1048 static int resync_start(struct mddev *mddev)
1050 struct md_cluster_info *cinfo = mddev->cluster_info;
1051 return dlm_lock_sync_interruptible(cinfo->resync_lockres, DLM_LOCK_EX, mddev);
1054 static int resync_info_update(struct mddev *mddev, sector_t lo, sector_t hi)
1056 struct md_cluster_info *cinfo = mddev->cluster_info;
1057 struct resync_info ri;
1058 struct cluster_msg cmsg = {0};
1060 /* do not send zero again, if we have sent before */
1062 memcpy(&ri, cinfo->bitmap_lockres->lksb.sb_lvbptr, sizeof(struct resync_info));
1063 if (le64_to_cpu(ri.hi) == 0)
1067 add_resync_info(cinfo->bitmap_lockres, lo, hi);
1068 /* Re-acquire the lock to refresh LVB */
1069 dlm_lock_sync(cinfo->bitmap_lockres, DLM_LOCK_PW);
1070 cmsg.type = cpu_to_le32(RESYNCING);
1071 cmsg.low = cpu_to_le64(lo);
1072 cmsg.high = cpu_to_le64(hi);
1074 return sendmsg(cinfo, &cmsg);
1077 static int resync_finish(struct mddev *mddev)
1079 struct md_cluster_info *cinfo = mddev->cluster_info;
1080 dlm_unlock_sync(cinfo->resync_lockres);
1081 return resync_info_update(mddev, 0, 0);
1084 static int area_resyncing(struct mddev *mddev, int direction,
1085 sector_t lo, sector_t hi)
1087 struct md_cluster_info *cinfo = mddev->cluster_info;
1089 struct suspend_info *s;
1091 if ((direction == READ) &&
1092 test_bit(MD_CLUSTER_SUSPEND_READ_BALANCING, &cinfo->state))
1095 spin_lock_irq(&cinfo->suspend_lock);
1096 if (list_empty(&cinfo->suspend_list))
1098 list_for_each_entry(s, &cinfo->suspend_list, list)
1099 if (hi > s->lo && lo < s->hi) {
1104 spin_unlock_irq(&cinfo->suspend_lock);
1108 /* add_new_disk() - initiates a disk add
1109 * However, if this fails before writing md_update_sb(),
1110 * add_new_disk_cancel() must be called to release token lock
1112 static int add_new_disk(struct mddev *mddev, struct md_rdev *rdev)
1114 struct md_cluster_info *cinfo = mddev->cluster_info;
1115 struct cluster_msg cmsg;
1117 struct mdp_superblock_1 *sb = page_address(rdev->sb_page);
1118 char *uuid = sb->device_uuid;
1120 memset(&cmsg, 0, sizeof(cmsg));
1121 cmsg.type = cpu_to_le32(NEWDISK);
1122 memcpy(cmsg.uuid, uuid, 16);
1123 cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
1125 ret = __sendmsg(cinfo, &cmsg);
1130 cinfo->no_new_dev_lockres->flags |= DLM_LKF_NOQUEUE;
1131 ret = dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_EX);
1132 cinfo->no_new_dev_lockres->flags &= ~DLM_LKF_NOQUEUE;
1133 /* Some node does not "see" the device */
1139 dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR);
1140 /* Since MD_CHANGE_DEVS will be set in add_bound_rdev which
1141 * will run soon after add_new_disk, the below path will be
1143 * md_wakeup_thread(mddev->thread)
1144 * -> conf->thread (raid1d)
1145 * -> md_check_recovery -> md_update_sb
1146 * -> metadata_update_start/finish
1147 * MD_CLUSTER_SEND_LOCKED_ALREADY will be cleared eventually.
1149 * For other failure cases, metadata_update_cancel and
1150 * add_new_disk_cancel also clear below bit as well.
1152 set_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1153 wake_up(&cinfo->wait);
1158 static void add_new_disk_cancel(struct mddev *mddev)
1160 struct md_cluster_info *cinfo = mddev->cluster_info;
1161 clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1165 static int new_disk_ack(struct mddev *mddev, bool ack)
1167 struct md_cluster_info *cinfo = mddev->cluster_info;
1169 if (!test_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state)) {
1170 pr_warn("md-cluster(%s): Spurious cluster confirmation\n", mdname(mddev));
1175 dlm_unlock_sync(cinfo->no_new_dev_lockres);
1176 complete(&cinfo->newdisk_completion);
1180 static int remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1182 struct cluster_msg cmsg = {0};
1183 struct md_cluster_info *cinfo = mddev->cluster_info;
1184 cmsg.type = cpu_to_le32(REMOVE);
1185 cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
1186 return sendmsg(cinfo, &cmsg);
1189 static int lock_all_bitmaps(struct mddev *mddev)
1191 int slot, my_slot, ret, held = 1, i = 0;
1193 struct md_cluster_info *cinfo = mddev->cluster_info;
1195 cinfo->other_bitmap_lockres = kzalloc((mddev->bitmap_info.nodes - 1) *
1196 sizeof(struct dlm_lock_resource *),
1198 if (!cinfo->other_bitmap_lockres) {
1199 pr_err("md: can't alloc mem for other bitmap locks\n");
1203 my_slot = slot_number(mddev);
1204 for (slot = 0; slot < mddev->bitmap_info.nodes; slot++) {
1205 if (slot == my_slot)
1208 memset(str, '\0', 64);
1209 snprintf(str, 64, "bitmap%04d", slot);
1210 cinfo->other_bitmap_lockres[i] = lockres_init(mddev, str, NULL, 1);
1211 if (!cinfo->other_bitmap_lockres[i])
1214 cinfo->other_bitmap_lockres[i]->flags |= DLM_LKF_NOQUEUE;
1215 ret = dlm_lock_sync(cinfo->other_bitmap_lockres[i], DLM_LOCK_PW);
1224 static void unlock_all_bitmaps(struct mddev *mddev)
1226 struct md_cluster_info *cinfo = mddev->cluster_info;
1229 /* release other node's bitmap lock if they are existed */
1230 if (cinfo->other_bitmap_lockres) {
1231 for (i = 0; i < mddev->bitmap_info.nodes - 1; i++) {
1232 if (cinfo->other_bitmap_lockres[i]) {
1233 lockres_free(cinfo->other_bitmap_lockres[i]);
1236 kfree(cinfo->other_bitmap_lockres);
1237 cinfo->other_bitmap_lockres = NULL;
1241 static int gather_bitmaps(struct md_rdev *rdev)
1245 struct cluster_msg cmsg = {0};
1246 struct mddev *mddev = rdev->mddev;
1247 struct md_cluster_info *cinfo = mddev->cluster_info;
1249 cmsg.type = cpu_to_le32(RE_ADD);
1250 cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
1251 err = sendmsg(cinfo, &cmsg);
1255 for (sn = 0; sn < mddev->bitmap_info.nodes; sn++) {
1256 if (sn == (cinfo->slot_number - 1))
1258 err = bitmap_copy_from_slot(mddev, sn, &lo, &hi, false);
1260 pr_warn("md-cluster: Could not gather bitmaps from slot %d", sn);
1263 if ((hi > 0) && (lo < mddev->recovery_cp))
1264 mddev->recovery_cp = lo;
1270 static struct md_cluster_operations cluster_ops = {
1273 .slot_number = slot_number,
1274 .resync_start = resync_start,
1275 .resync_finish = resync_finish,
1276 .resync_info_update = resync_info_update,
1277 .metadata_update_start = metadata_update_start,
1278 .metadata_update_finish = metadata_update_finish,
1279 .metadata_update_cancel = metadata_update_cancel,
1280 .area_resyncing = area_resyncing,
1281 .add_new_disk = add_new_disk,
1282 .add_new_disk_cancel = add_new_disk_cancel,
1283 .new_disk_ack = new_disk_ack,
1284 .remove_disk = remove_disk,
1285 .load_bitmaps = load_bitmaps,
1286 .gather_bitmaps = gather_bitmaps,
1287 .lock_all_bitmaps = lock_all_bitmaps,
1288 .unlock_all_bitmaps = unlock_all_bitmaps,
1291 static int __init cluster_init(void)
1293 pr_warn("md-cluster: EXPERIMENTAL. Use with caution\n");
1294 pr_info("Registering Cluster MD functions\n");
1295 register_md_cluster_operations(&cluster_ops, THIS_MODULE);
1299 static void cluster_exit(void)
1301 unregister_md_cluster_operations();
1304 module_init(cluster_init);
1305 module_exit(cluster_exit);
1306 MODULE_AUTHOR("SUSE");
1307 MODULE_LICENSE("GPL");
1308 MODULE_DESCRIPTION("Clustering support for MD");