GNU Linux-libre 4.19.264-gnu1
[releases.git] / drivers / md / md-cluster.c
1 /*
2  * Copyright (C) 2015, SUSE
3  *
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)
7  * any later version.
8  *
9  */
10
11
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>
17 #include "md.h"
18 #include "md-bitmap.h"
19 #include "md-cluster.h"
20
21 #define LVB_SIZE        64
22 #define NEW_DEV_TIMEOUT 5000
23
24 struct dlm_lock_resource {
25         dlm_lockspace_t *ls;
26         struct dlm_lksb lksb;
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. */
33         int mode;
34 };
35
36 struct suspend_info {
37         int slot;
38         sector_t lo;
39         sector_t hi;
40         struct list_head list;
41 };
42
43 struct resync_info {
44         __le64 lo;
45         __le64 hi;
46 };
47
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
52
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.
56  */
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
63  * the lock.
64  */
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
71
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;
76         int slot_number;
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;
94         unsigned long state;
95         /* record the region in RESYNCING message */
96         sector_t sync_low;
97         sector_t sync_hi;
98 };
99
100 enum msg_type {
101         METADATA_UPDATED = 0,
102         RESYNCING,
103         NEWDISK,
104         REMOVE,
105         RE_ADD,
106         BITMAP_NEEDS_SYNC,
107         CHANGE_CAPACITY,
108 };
109
110 struct cluster_msg {
111         __le32 type;
112         __le32 slot;
113         /* TODO: Unionize this for smaller footprint */
114         __le64 low;
115         __le64 high;
116         char uuid[16];
117         __le32 raid_slot;
118 };
119
120 static void sync_ast(void *arg)
121 {
122         struct dlm_lock_resource *res;
123
124         res = arg;
125         res->sync_locking_done = true;
126         wake_up(&res->sync_locking);
127 }
128
129 static int dlm_lock_sync(struct dlm_lock_resource *res, int mode)
130 {
131         int ret = 0;
132
133         ret = dlm_lock(res->ls, mode, &res->lksb,
134                         res->flags, res->name, strlen(res->name),
135                         0, sync_ast, res, res->bast);
136         if (ret)
137                 return ret;
138         wait_event(res->sync_locking, res->sync_locking_done);
139         res->sync_locking_done = false;
140         if (res->lksb.sb_status == 0)
141                 res->mode = mode;
142         return res->lksb.sb_status;
143 }
144
145 static int dlm_unlock_sync(struct dlm_lock_resource *res)
146 {
147         return dlm_lock_sync(res, DLM_LOCK_NL);
148 }
149
150 /*
151  * An variation of dlm_lock_sync, which make lock request could
152  * be interrupted
153  */
154 static int dlm_lock_sync_interruptible(struct dlm_lock_resource *res, int mode,
155                                        struct mddev *mddev)
156 {
157         int ret = 0;
158
159         ret = dlm_lock(res->ls, mode, &res->lksb,
160                         res->flags, res->name, strlen(res->name),
161                         0, sync_ast, res, res->bast);
162         if (ret)
163                 return ret;
164
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) {
169                 /*
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
173                  */
174                 ret = dlm_unlock(res->ls, res->lksb.sb_lkid, DLM_LKF_CANCEL,
175                         &res->lksb, res);
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);
180                 return -EPERM;
181         } else
182                 res->sync_locking_done = false;
183         if (res->lksb.sb_status == 0)
184                 res->mode = mode;
185         return res->lksb.sb_status;
186 }
187
188 static struct dlm_lock_resource *lockres_init(struct mddev *mddev,
189                 char *name, void (*bastfn)(void *arg, int mode), int with_lvb)
190 {
191         struct dlm_lock_resource *res = NULL;
192         int ret, namelen;
193         struct md_cluster_info *cinfo = mddev->cluster_info;
194
195         res = kzalloc(sizeof(struct dlm_lock_resource), GFP_KERNEL);
196         if (!res)
197                 return NULL;
198         init_waitqueue_head(&res->sync_locking);
199         res->sync_locking_done = false;
200         res->ls = cinfo->lockspace;
201         res->mddev = mddev;
202         res->mode = DLM_LOCK_IV;
203         namelen = strlen(name);
204         res->name = kzalloc(namelen + 1, GFP_KERNEL);
205         if (!res->name) {
206                 pr_err("md-cluster: Unable to allocate resource name for resource %s\n", name);
207                 goto out_err;
208         }
209         strlcpy(res->name, name, namelen + 1);
210         if (with_lvb) {
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);
214                         goto out_err;
215                 }
216                 res->flags = DLM_LKF_VALBLK;
217         }
218
219         if (bastfn)
220                 res->bast = bastfn;
221
222         res->flags |= DLM_LKF_EXPEDITE;
223
224         ret = dlm_lock_sync(res, DLM_LOCK_NL);
225         if (ret) {
226                 pr_err("md-cluster: Unable to lock NL on new lock resource %s\n", name);
227                 goto out_err;
228         }
229         res->flags &= ~DLM_LKF_EXPEDITE;
230         res->flags |= DLM_LKF_CONVERT;
231
232         return res;
233 out_err:
234         kfree(res->lksb.sb_lvbptr);
235         kfree(res->name);
236         kfree(res);
237         return NULL;
238 }
239
240 static void lockres_free(struct dlm_lock_resource *res)
241 {
242         int ret = 0;
243
244         if (!res)
245                 return;
246
247         /*
248          * use FORCEUNLOCK flag, so we can unlock even the lock is on the
249          * waiting or convert queue
250          */
251         ret = dlm_unlock(res->ls, res->lksb.sb_lkid, DLM_LKF_FORCEUNLOCK,
252                 &res->lksb, res);
253         if (unlikely(ret != 0))
254                 pr_err("failed to unlock %s return %d\n", res->name, ret);
255         else
256                 wait_event(res->sync_locking, res->sync_locking_done);
257
258         kfree(res->name);
259         kfree(res->lksb.sb_lvbptr);
260         kfree(res);
261 }
262
263 static void add_resync_info(struct dlm_lock_resource *lockres,
264                             sector_t lo, sector_t hi)
265 {
266         struct resync_info *ri;
267
268         ri = (struct resync_info *)lockres->lksb.sb_lvbptr;
269         ri->lo = cpu_to_le64(lo);
270         ri->hi = cpu_to_le64(hi);
271 }
272
273 static struct suspend_info *read_resync_info(struct mddev *mddev, struct dlm_lock_resource *lockres)
274 {
275         struct resync_info ri;
276         struct suspend_info *s = NULL;
277         sector_t hi = 0;
278
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);
282         if (hi > 0) {
283                 s = kzalloc(sizeof(struct suspend_info), GFP_KERNEL);
284                 if (!s)
285                         goto out;
286                 s->hi = hi;
287                 s->lo = le64_to_cpu(ri.lo);
288         }
289         dlm_unlock_sync(lockres);
290 out:
291         return s;
292 }
293
294 static void recover_bitmaps(struct md_thread *thread)
295 {
296         struct mddev *mddev = thread->mddev;
297         struct md_cluster_info *cinfo = mddev->cluster_info;
298         struct dlm_lock_resource *bm_lockres;
299         char str[64];
300         int slot, ret;
301         struct suspend_info *s, *tmp;
302         sector_t lo, hi;
303
304         while (cinfo->recovery_map) {
305                 slot = fls64((u64)cinfo->recovery_map) - 1;
306
307                 snprintf(str, 64, "bitmap%04d", slot);
308                 bm_lockres = lockres_init(mddev, str, NULL, 1);
309                 if (!bm_lockres) {
310                         pr_err("md-cluster: Cannot initialize bitmaps\n");
311                         goto clear_bit;
312                 }
313
314                 ret = dlm_lock_sync_interruptible(bm_lockres, DLM_LOCK_PW, mddev);
315                 if (ret) {
316                         pr_err("md-cluster: Could not DLM lock %s: %d\n",
317                                         str, ret);
318                         goto clear_bit;
319                 }
320                 ret = md_bitmap_copy_from_slot(mddev, slot, &lo, &hi, true);
321                 if (ret) {
322                         pr_err("md-cluster: Could not copy data from bitmap %d\n", slot);
323                         goto clear_bit;
324                 }
325
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) {
330                                 list_del(&s->list);
331                                 kfree(s);
332                         }
333                 spin_unlock_irq(&cinfo->suspend_lock);
334
335                 if (hi > 0) {
336                         if (lo < mddev->recovery_cp)
337                                 mddev->recovery_cp = lo;
338                         /* wake up thread to continue resync in case resync
339                          * is not finished */
340                         if (mddev->recovery_cp != MaxSector) {
341                                 /*
342                                  * clear the REMOTE flag since we will launch
343                                  * resync thread in current node.
344                                  */
345                                 clear_bit(MD_RESYNCING_REMOTE,
346                                           &mddev->recovery);
347                                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
348                                 md_wakeup_thread(mddev->thread);
349                         }
350                 }
351 clear_bit:
352                 lockres_free(bm_lockres);
353                 clear_bit(slot, &cinfo->recovery_map);
354         }
355 }
356
357 static void recover_prep(void *arg)
358 {
359         struct mddev *mddev = arg;
360         struct md_cluster_info *cinfo = mddev->cluster_info;
361         set_bit(MD_CLUSTER_SUSPEND_READ_BALANCING, &cinfo->state);
362 }
363
364 static void __recover_slot(struct mddev *mddev, int slot)
365 {
366         struct md_cluster_info *cinfo = mddev->cluster_info;
367
368         set_bit(slot, &cinfo->recovery_map);
369         if (!cinfo->recovery_thread) {
370                 cinfo->recovery_thread = md_register_thread(recover_bitmaps,
371                                 mddev, "recover");
372                 if (!cinfo->recovery_thread) {
373                         pr_warn("md-cluster: Could not create recovery thread\n");
374                         return;
375                 }
376         }
377         md_wakeup_thread(cinfo->recovery_thread);
378 }
379
380 static void recover_slot(void *arg, struct dlm_slot *slot)
381 {
382         struct mddev *mddev = arg;
383         struct md_cluster_info *cinfo = mddev->cluster_info;
384
385         pr_info("md-cluster: %s Node %d/%d down. My slot: %d. Initiating recovery.\n",
386                         mddev->bitmap_info.cluster_name,
387                         slot->nodeid, slot->slot,
388                         cinfo->slot_number);
389         /* deduct one since dlm slot starts from one while the num of
390          * cluster-md begins with 0 */
391         __recover_slot(mddev, slot->slot - 1);
392 }
393
394 static void recover_done(void *arg, struct dlm_slot *slots,
395                 int num_slots, int our_slot,
396                 uint32_t generation)
397 {
398         struct mddev *mddev = arg;
399         struct md_cluster_info *cinfo = mddev->cluster_info;
400
401         cinfo->slot_number = our_slot;
402         /* completion is only need to be complete when node join cluster,
403          * it doesn't need to run during another node's failure */
404         if (test_bit(MD_CLUSTER_BEGIN_JOIN_CLUSTER, &cinfo->state)) {
405                 complete(&cinfo->completion);
406                 clear_bit(MD_CLUSTER_BEGIN_JOIN_CLUSTER, &cinfo->state);
407         }
408         clear_bit(MD_CLUSTER_SUSPEND_READ_BALANCING, &cinfo->state);
409 }
410
411 /* the ops is called when node join the cluster, and do lock recovery
412  * if node failure occurs */
413 static const struct dlm_lockspace_ops md_ls_ops = {
414         .recover_prep = recover_prep,
415         .recover_slot = recover_slot,
416         .recover_done = recover_done,
417 };
418
419 /*
420  * The BAST function for the ack lock resource
421  * This function wakes up the receive thread in
422  * order to receive and process the message.
423  */
424 static void ack_bast(void *arg, int mode)
425 {
426         struct dlm_lock_resource *res = arg;
427         struct md_cluster_info *cinfo = res->mddev->cluster_info;
428
429         if (mode == DLM_LOCK_EX) {
430                 if (test_bit(MD_CLUSTER_ALREADY_IN_CLUSTER, &cinfo->state))
431                         md_wakeup_thread(cinfo->recv_thread);
432                 else
433                         set_bit(MD_CLUSTER_PENDING_RECV_EVENT, &cinfo->state);
434         }
435 }
436
437 static void __remove_suspend_info(struct md_cluster_info *cinfo, int slot)
438 {
439         struct suspend_info *s, *tmp;
440
441         list_for_each_entry_safe(s, tmp, &cinfo->suspend_list, list)
442                 if (slot == s->slot) {
443                         list_del(&s->list);
444                         kfree(s);
445                         break;
446                 }
447 }
448
449 static void remove_suspend_info(struct mddev *mddev, int slot)
450 {
451         struct md_cluster_info *cinfo = mddev->cluster_info;
452         mddev->pers->quiesce(mddev, 1);
453         spin_lock_irq(&cinfo->suspend_lock);
454         __remove_suspend_info(cinfo, slot);
455         spin_unlock_irq(&cinfo->suspend_lock);
456         mddev->pers->quiesce(mddev, 0);
457 }
458
459
460 static void process_suspend_info(struct mddev *mddev,
461                 int slot, sector_t lo, sector_t hi)
462 {
463         struct md_cluster_info *cinfo = mddev->cluster_info;
464         struct suspend_info *s;
465
466         if (!hi) {
467                 /*
468                  * clear the REMOTE flag since resync or recovery is finished
469                  * in remote node.
470                  */
471                 clear_bit(MD_RESYNCING_REMOTE, &mddev->recovery);
472                 remove_suspend_info(mddev, slot);
473                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
474                 md_wakeup_thread(mddev->thread);
475                 return;
476         }
477
478         /*
479          * The bitmaps are not same for different nodes
480          * if RESYNCING is happening in one node, then
481          * the node which received the RESYNCING message
482          * probably will perform resync with the region
483          * [lo, hi] again, so we could reduce resync time
484          * a lot if we can ensure that the bitmaps among
485          * different nodes are match up well.
486          *
487          * sync_low/hi is used to record the region which
488          * arrived in the previous RESYNCING message,
489          *
490          * Call bitmap_sync_with_cluster to clear
491          * NEEDED_MASK and set RESYNC_MASK since
492          * resync thread is running in another node,
493          * so we don't need to do the resync again
494          * with the same section */
495         md_bitmap_sync_with_cluster(mddev, cinfo->sync_low, cinfo->sync_hi, lo, hi);
496         cinfo->sync_low = lo;
497         cinfo->sync_hi = hi;
498
499         s = kzalloc(sizeof(struct suspend_info), GFP_KERNEL);
500         if (!s)
501                 return;
502         s->slot = slot;
503         s->lo = lo;
504         s->hi = hi;
505         mddev->pers->quiesce(mddev, 1);
506         spin_lock_irq(&cinfo->suspend_lock);
507         /* Remove existing entry (if exists) before adding */
508         __remove_suspend_info(cinfo, slot);
509         list_add(&s->list, &cinfo->suspend_list);
510         spin_unlock_irq(&cinfo->suspend_lock);
511         mddev->pers->quiesce(mddev, 0);
512 }
513
514 static void process_add_new_disk(struct mddev *mddev, struct cluster_msg *cmsg)
515 {
516         char disk_uuid[64];
517         struct md_cluster_info *cinfo = mddev->cluster_info;
518         char event_name[] = "EVENT=ADD_DEVICE";
519         char raid_slot[16];
520         char *envp[] = {event_name, disk_uuid, raid_slot, NULL};
521         int len;
522
523         len = snprintf(disk_uuid, 64, "DEVICE_UUID=");
524         sprintf(disk_uuid + len, "%pU", cmsg->uuid);
525         snprintf(raid_slot, 16, "RAID_DISK=%d", le32_to_cpu(cmsg->raid_slot));
526         pr_info("%s:%d Sending kobject change with %s and %s\n", __func__, __LINE__, disk_uuid, raid_slot);
527         init_completion(&cinfo->newdisk_completion);
528         set_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state);
529         kobject_uevent_env(&disk_to_dev(mddev->gendisk)->kobj, KOBJ_CHANGE, envp);
530         wait_for_completion_timeout(&cinfo->newdisk_completion,
531                         NEW_DEV_TIMEOUT);
532         clear_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state);
533 }
534
535
536 static void process_metadata_update(struct mddev *mddev, struct cluster_msg *msg)
537 {
538         int got_lock = 0;
539         struct md_cluster_info *cinfo = mddev->cluster_info;
540         mddev->good_device_nr = le32_to_cpu(msg->raid_slot);
541
542         dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR);
543         wait_event(mddev->thread->wqueue,
544                    (got_lock = mddev_trylock(mddev)) ||
545                     test_bit(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state));
546         md_reload_sb(mddev, mddev->good_device_nr);
547         if (got_lock)
548                 mddev_unlock(mddev);
549 }
550
551 static void process_remove_disk(struct mddev *mddev, struct cluster_msg *msg)
552 {
553         struct md_rdev *rdev;
554
555         rcu_read_lock();
556         rdev = md_find_rdev_nr_rcu(mddev, le32_to_cpu(msg->raid_slot));
557         if (rdev) {
558                 set_bit(ClusterRemove, &rdev->flags);
559                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
560                 md_wakeup_thread(mddev->thread);
561         }
562         else
563                 pr_warn("%s: %d Could not find disk(%d) to REMOVE\n",
564                         __func__, __LINE__, le32_to_cpu(msg->raid_slot));
565         rcu_read_unlock();
566 }
567
568 static void process_readd_disk(struct mddev *mddev, struct cluster_msg *msg)
569 {
570         struct md_rdev *rdev;
571
572         rcu_read_lock();
573         rdev = md_find_rdev_nr_rcu(mddev, le32_to_cpu(msg->raid_slot));
574         if (rdev && test_bit(Faulty, &rdev->flags))
575                 clear_bit(Faulty, &rdev->flags);
576         else
577                 pr_warn("%s: %d Could not find disk(%d) which is faulty",
578                         __func__, __LINE__, le32_to_cpu(msg->raid_slot));
579         rcu_read_unlock();
580 }
581
582 static int process_recvd_msg(struct mddev *mddev, struct cluster_msg *msg)
583 {
584         int ret = 0;
585
586         if (WARN(mddev->cluster_info->slot_number - 1 == le32_to_cpu(msg->slot),
587                 "node %d received it's own msg\n", le32_to_cpu(msg->slot)))
588                 return -1;
589         switch (le32_to_cpu(msg->type)) {
590         case METADATA_UPDATED:
591                 process_metadata_update(mddev, msg);
592                 break;
593         case CHANGE_CAPACITY:
594                 set_capacity(mddev->gendisk, mddev->array_sectors);
595                 revalidate_disk(mddev->gendisk);
596                 break;
597         case RESYNCING:
598                 set_bit(MD_RESYNCING_REMOTE, &mddev->recovery);
599                 process_suspend_info(mddev, le32_to_cpu(msg->slot),
600                                      le64_to_cpu(msg->low),
601                                      le64_to_cpu(msg->high));
602                 break;
603         case NEWDISK:
604                 process_add_new_disk(mddev, msg);
605                 break;
606         case REMOVE:
607                 process_remove_disk(mddev, msg);
608                 break;
609         case RE_ADD:
610                 process_readd_disk(mddev, msg);
611                 break;
612         case BITMAP_NEEDS_SYNC:
613                 __recover_slot(mddev, le32_to_cpu(msg->slot));
614                 break;
615         default:
616                 ret = -1;
617                 pr_warn("%s:%d Received unknown message from %d\n",
618                         __func__, __LINE__, msg->slot);
619         }
620         return ret;
621 }
622
623 /*
624  * thread for receiving message
625  */
626 static void recv_daemon(struct md_thread *thread)
627 {
628         struct md_cluster_info *cinfo = thread->mddev->cluster_info;
629         struct dlm_lock_resource *ack_lockres = cinfo->ack_lockres;
630         struct dlm_lock_resource *message_lockres = cinfo->message_lockres;
631         struct cluster_msg msg;
632         int ret;
633
634         mutex_lock(&cinfo->recv_mutex);
635         /*get CR on Message*/
636         if (dlm_lock_sync(message_lockres, DLM_LOCK_CR)) {
637                 pr_err("md/raid1:failed to get CR on MESSAGE\n");
638                 mutex_unlock(&cinfo->recv_mutex);
639                 return;
640         }
641
642         /* read lvb and wake up thread to process this message_lockres */
643         memcpy(&msg, message_lockres->lksb.sb_lvbptr, sizeof(struct cluster_msg));
644         ret = process_recvd_msg(thread->mddev, &msg);
645         if (ret)
646                 goto out;
647
648         /*release CR on ack_lockres*/
649         ret = dlm_unlock_sync(ack_lockres);
650         if (unlikely(ret != 0))
651                 pr_info("unlock ack failed return %d\n", ret);
652         /*up-convert to PR on message_lockres*/
653         ret = dlm_lock_sync(message_lockres, DLM_LOCK_PR);
654         if (unlikely(ret != 0))
655                 pr_info("lock PR on msg failed return %d\n", ret);
656         /*get CR on ack_lockres again*/
657         ret = dlm_lock_sync(ack_lockres, DLM_LOCK_CR);
658         if (unlikely(ret != 0))
659                 pr_info("lock CR on ack failed return %d\n", ret);
660 out:
661         /*release CR on message_lockres*/
662         ret = dlm_unlock_sync(message_lockres);
663         if (unlikely(ret != 0))
664                 pr_info("unlock msg failed return %d\n", ret);
665         mutex_unlock(&cinfo->recv_mutex);
666 }
667
668 /* lock_token()
669  * Takes the lock on the TOKEN lock resource so no other
670  * node can communicate while the operation is underway.
671  */
672 static int lock_token(struct md_cluster_info *cinfo)
673 {
674         int error;
675
676         error = dlm_lock_sync(cinfo->token_lockres, DLM_LOCK_EX);
677         if (error) {
678                 pr_err("md-cluster(%s:%d): failed to get EX on TOKEN (%d)\n",
679                                 __func__, __LINE__, error);
680         } else {
681                 /* Lock the receive sequence */
682                 mutex_lock(&cinfo->recv_mutex);
683         }
684         return error;
685 }
686
687 /* lock_comm()
688  * Sets the MD_CLUSTER_SEND_LOCK bit to lock the send channel.
689  */
690 static int lock_comm(struct md_cluster_info *cinfo, bool mddev_locked)
691 {
692         int rv, set_bit = 0;
693         struct mddev *mddev = cinfo->mddev;
694
695         /*
696          * If resync thread run after raid1d thread, then process_metadata_update
697          * could not continue if raid1d held reconfig_mutex (and raid1d is blocked
698          * since another node already got EX on Token and waitting the EX of Ack),
699          * so let resync wake up thread in case flag is set.
700          */
701         if (mddev_locked && !test_bit(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD,
702                                       &cinfo->state)) {
703                 rv = test_and_set_bit_lock(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD,
704                                               &cinfo->state);
705                 WARN_ON_ONCE(rv);
706                 md_wakeup_thread(mddev->thread);
707                 set_bit = 1;
708         }
709
710         wait_event(cinfo->wait,
711                    !test_and_set_bit(MD_CLUSTER_SEND_LOCK, &cinfo->state));
712         rv = lock_token(cinfo);
713         if (set_bit)
714                 clear_bit_unlock(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state);
715         return rv;
716 }
717
718 static void unlock_comm(struct md_cluster_info *cinfo)
719 {
720         WARN_ON(cinfo->token_lockres->mode != DLM_LOCK_EX);
721         mutex_unlock(&cinfo->recv_mutex);
722         dlm_unlock_sync(cinfo->token_lockres);
723         clear_bit(MD_CLUSTER_SEND_LOCK, &cinfo->state);
724         wake_up(&cinfo->wait);
725 }
726
727 /* __sendmsg()
728  * This function performs the actual sending of the message. This function is
729  * usually called after performing the encompassing operation
730  * The function:
731  * 1. Grabs the message lockresource in EX mode
732  * 2. Copies the message to the message LVB
733  * 3. Downconverts message lockresource to CW
734  * 4. Upconverts ack lock resource from CR to EX. This forces the BAST on other nodes
735  *    and the other nodes read the message. The thread will wait here until all other
736  *    nodes have released ack lock resource.
737  * 5. Downconvert ack lockresource to CR
738  */
739 static int __sendmsg(struct md_cluster_info *cinfo, struct cluster_msg *cmsg)
740 {
741         int error;
742         int slot = cinfo->slot_number - 1;
743
744         cmsg->slot = cpu_to_le32(slot);
745         /*get EX on Message*/
746         error = dlm_lock_sync(cinfo->message_lockres, DLM_LOCK_EX);
747         if (error) {
748                 pr_err("md-cluster: failed to get EX on MESSAGE (%d)\n", error);
749                 goto failed_message;
750         }
751
752         memcpy(cinfo->message_lockres->lksb.sb_lvbptr, (void *)cmsg,
753                         sizeof(struct cluster_msg));
754         /*down-convert EX to CW on Message*/
755         error = dlm_lock_sync(cinfo->message_lockres, DLM_LOCK_CW);
756         if (error) {
757                 pr_err("md-cluster: failed to convert EX to CW on MESSAGE(%d)\n",
758                                 error);
759                 goto failed_ack;
760         }
761
762         /*up-convert CR to EX on Ack*/
763         error = dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_EX);
764         if (error) {
765                 pr_err("md-cluster: failed to convert CR to EX on ACK(%d)\n",
766                                 error);
767                 goto failed_ack;
768         }
769
770         /*down-convert EX to CR on Ack*/
771         error = dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_CR);
772         if (error) {
773                 pr_err("md-cluster: failed to convert EX to CR on ACK(%d)\n",
774                                 error);
775                 goto failed_ack;
776         }
777
778 failed_ack:
779         error = dlm_unlock_sync(cinfo->message_lockres);
780         if (unlikely(error != 0)) {
781                 pr_err("md-cluster: failed convert to NL on MESSAGE(%d)\n",
782                         error);
783                 /* in case the message can't be released due to some reason */
784                 goto failed_ack;
785         }
786 failed_message:
787         return error;
788 }
789
790 static int sendmsg(struct md_cluster_info *cinfo, struct cluster_msg *cmsg,
791                    bool mddev_locked)
792 {
793         int ret;
794
795         ret = lock_comm(cinfo, mddev_locked);
796         if (!ret) {
797                 ret = __sendmsg(cinfo, cmsg);
798                 unlock_comm(cinfo);
799         }
800         return ret;
801 }
802
803 static int gather_all_resync_info(struct mddev *mddev, int total_slots)
804 {
805         struct md_cluster_info *cinfo = mddev->cluster_info;
806         int i, ret = 0;
807         struct dlm_lock_resource *bm_lockres;
808         struct suspend_info *s;
809         char str[64];
810         sector_t lo, hi;
811
812
813         for (i = 0; i < total_slots; i++) {
814                 memset(str, '\0', 64);
815                 snprintf(str, 64, "bitmap%04d", i);
816                 bm_lockres = lockres_init(mddev, str, NULL, 1);
817                 if (!bm_lockres)
818                         return -ENOMEM;
819                 if (i == (cinfo->slot_number - 1)) {
820                         lockres_free(bm_lockres);
821                         continue;
822                 }
823
824                 bm_lockres->flags |= DLM_LKF_NOQUEUE;
825                 ret = dlm_lock_sync(bm_lockres, DLM_LOCK_PW);
826                 if (ret == -EAGAIN) {
827                         s = read_resync_info(mddev, bm_lockres);
828                         if (s) {
829                                 pr_info("%s:%d Resync[%llu..%llu] in progress on %d\n",
830                                                 __func__, __LINE__,
831                                                 (unsigned long long) s->lo,
832                                                 (unsigned long long) s->hi, i);
833                                 spin_lock_irq(&cinfo->suspend_lock);
834                                 s->slot = i;
835                                 list_add(&s->list, &cinfo->suspend_list);
836                                 spin_unlock_irq(&cinfo->suspend_lock);
837                         }
838                         ret = 0;
839                         lockres_free(bm_lockres);
840                         continue;
841                 }
842                 if (ret) {
843                         lockres_free(bm_lockres);
844                         goto out;
845                 }
846
847                 /* Read the disk bitmap sb and check if it needs recovery */
848                 ret = md_bitmap_copy_from_slot(mddev, i, &lo, &hi, false);
849                 if (ret) {
850                         pr_warn("md-cluster: Could not gather bitmaps from slot %d", i);
851                         lockres_free(bm_lockres);
852                         continue;
853                 }
854                 if ((hi > 0) && (lo < mddev->recovery_cp)) {
855                         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
856                         mddev->recovery_cp = lo;
857                         md_check_recovery(mddev);
858                 }
859
860                 lockres_free(bm_lockres);
861         }
862 out:
863         return ret;
864 }
865
866 static int join(struct mddev *mddev, int nodes)
867 {
868         struct md_cluster_info *cinfo;
869         int ret, ops_rv;
870         char str[64];
871
872         cinfo = kzalloc(sizeof(struct md_cluster_info), GFP_KERNEL);
873         if (!cinfo)
874                 return -ENOMEM;
875
876         INIT_LIST_HEAD(&cinfo->suspend_list);
877         spin_lock_init(&cinfo->suspend_lock);
878         init_completion(&cinfo->completion);
879         set_bit(MD_CLUSTER_BEGIN_JOIN_CLUSTER, &cinfo->state);
880         init_waitqueue_head(&cinfo->wait);
881         mutex_init(&cinfo->recv_mutex);
882
883         mddev->cluster_info = cinfo;
884         cinfo->mddev = mddev;
885
886         memset(str, 0, 64);
887         sprintf(str, "%pU", mddev->uuid);
888         ret = dlm_new_lockspace(str, mddev->bitmap_info.cluster_name,
889                                 DLM_LSFL_FS, LVB_SIZE,
890                                 &md_ls_ops, mddev, &ops_rv, &cinfo->lockspace);
891         if (ret)
892                 goto err;
893         wait_for_completion(&cinfo->completion);
894         if (nodes < cinfo->slot_number) {
895                 pr_err("md-cluster: Slot allotted(%d) is greater than available slots(%d).",
896                         cinfo->slot_number, nodes);
897                 ret = -ERANGE;
898                 goto err;
899         }
900         /* Initiate the communication resources */
901         ret = -ENOMEM;
902         cinfo->recv_thread = md_register_thread(recv_daemon, mddev, "cluster_recv");
903         if (!cinfo->recv_thread) {
904                 pr_err("md-cluster: cannot allocate memory for recv_thread!\n");
905                 goto err;
906         }
907         cinfo->message_lockres = lockres_init(mddev, "message", NULL, 1);
908         if (!cinfo->message_lockres)
909                 goto err;
910         cinfo->token_lockres = lockres_init(mddev, "token", NULL, 0);
911         if (!cinfo->token_lockres)
912                 goto err;
913         cinfo->no_new_dev_lockres = lockres_init(mddev, "no-new-dev", NULL, 0);
914         if (!cinfo->no_new_dev_lockres)
915                 goto err;
916
917         ret = dlm_lock_sync(cinfo->token_lockres, DLM_LOCK_EX);
918         if (ret) {
919                 ret = -EAGAIN;
920                 pr_err("md-cluster: can't join cluster to avoid lock issue\n");
921                 goto err;
922         }
923         cinfo->ack_lockres = lockres_init(mddev, "ack", ack_bast, 0);
924         if (!cinfo->ack_lockres) {
925                 ret = -ENOMEM;
926                 goto err;
927         }
928         /* get sync CR lock on ACK. */
929         if (dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_CR))
930                 pr_err("md-cluster: failed to get a sync CR lock on ACK!(%d)\n",
931                                 ret);
932         dlm_unlock_sync(cinfo->token_lockres);
933         /* get sync CR lock on no-new-dev. */
934         if (dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR))
935                 pr_err("md-cluster: failed to get a sync CR lock on no-new-dev!(%d)\n", ret);
936
937
938         pr_info("md-cluster: Joined cluster %s slot %d\n", str, cinfo->slot_number);
939         snprintf(str, 64, "bitmap%04d", cinfo->slot_number - 1);
940         cinfo->bitmap_lockres = lockres_init(mddev, str, NULL, 1);
941         if (!cinfo->bitmap_lockres) {
942                 ret = -ENOMEM;
943                 goto err;
944         }
945         if (dlm_lock_sync(cinfo->bitmap_lockres, DLM_LOCK_PW)) {
946                 pr_err("Failed to get bitmap lock\n");
947                 ret = -EINVAL;
948                 goto err;
949         }
950
951         cinfo->resync_lockres = lockres_init(mddev, "resync", NULL, 0);
952         if (!cinfo->resync_lockres) {
953                 ret = -ENOMEM;
954                 goto err;
955         }
956
957         return 0;
958 err:
959         set_bit(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state);
960         md_unregister_thread(&cinfo->recovery_thread);
961         md_unregister_thread(&cinfo->recv_thread);
962         lockres_free(cinfo->message_lockres);
963         lockres_free(cinfo->token_lockres);
964         lockres_free(cinfo->ack_lockres);
965         lockres_free(cinfo->no_new_dev_lockres);
966         lockres_free(cinfo->resync_lockres);
967         lockres_free(cinfo->bitmap_lockres);
968         if (cinfo->lockspace)
969                 dlm_release_lockspace(cinfo->lockspace, 2);
970         mddev->cluster_info = NULL;
971         kfree(cinfo);
972         return ret;
973 }
974
975 static void load_bitmaps(struct mddev *mddev, int total_slots)
976 {
977         struct md_cluster_info *cinfo = mddev->cluster_info;
978
979         /* load all the node's bitmap info for resync */
980         if (gather_all_resync_info(mddev, total_slots))
981                 pr_err("md-cluster: failed to gather all resyn infos\n");
982         set_bit(MD_CLUSTER_ALREADY_IN_CLUSTER, &cinfo->state);
983         /* wake up recv thread in case something need to be handled */
984         if (test_and_clear_bit(MD_CLUSTER_PENDING_RECV_EVENT, &cinfo->state))
985                 md_wakeup_thread(cinfo->recv_thread);
986 }
987
988 static void resync_bitmap(struct mddev *mddev)
989 {
990         struct md_cluster_info *cinfo = mddev->cluster_info;
991         struct cluster_msg cmsg = {0};
992         int err;
993
994         cmsg.type = cpu_to_le32(BITMAP_NEEDS_SYNC);
995         err = sendmsg(cinfo, &cmsg, 1);
996         if (err)
997                 pr_err("%s:%d: failed to send BITMAP_NEEDS_SYNC message (%d)\n",
998                         __func__, __LINE__, err);
999 }
1000
1001 static void unlock_all_bitmaps(struct mddev *mddev);
1002 static int leave(struct mddev *mddev)
1003 {
1004         struct md_cluster_info *cinfo = mddev->cluster_info;
1005
1006         if (!cinfo)
1007                 return 0;
1008
1009         /* BITMAP_NEEDS_SYNC message should be sent when node
1010          * is leaving the cluster with dirty bitmap, also we
1011          * can only deliver it when dlm connection is available */
1012         if (cinfo->slot_number > 0 && mddev->recovery_cp != MaxSector)
1013                 resync_bitmap(mddev);
1014
1015         set_bit(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state);
1016         md_unregister_thread(&cinfo->recovery_thread);
1017         md_unregister_thread(&cinfo->recv_thread);
1018         lockres_free(cinfo->message_lockres);
1019         lockres_free(cinfo->token_lockres);
1020         lockres_free(cinfo->ack_lockres);
1021         lockres_free(cinfo->no_new_dev_lockres);
1022         lockres_free(cinfo->resync_lockres);
1023         lockres_free(cinfo->bitmap_lockres);
1024         unlock_all_bitmaps(mddev);
1025         dlm_release_lockspace(cinfo->lockspace, 2);
1026         kfree(cinfo);
1027         return 0;
1028 }
1029
1030 /* slot_number(): Returns the MD slot number to use
1031  * DLM starts the slot numbers from 1, wheras cluster-md
1032  * wants the number to be from zero, so we deduct one
1033  */
1034 static int slot_number(struct mddev *mddev)
1035 {
1036         struct md_cluster_info *cinfo = mddev->cluster_info;
1037
1038         return cinfo->slot_number - 1;
1039 }
1040
1041 /*
1042  * Check if the communication is already locked, else lock the communication
1043  * channel.
1044  * If it is already locked, token is in EX mode, and hence lock_token()
1045  * should not be called.
1046  */
1047 static int metadata_update_start(struct mddev *mddev)
1048 {
1049         struct md_cluster_info *cinfo = mddev->cluster_info;
1050         int ret;
1051
1052         /*
1053          * metadata_update_start is always called with the protection of
1054          * reconfig_mutex, so set WAITING_FOR_TOKEN here.
1055          */
1056         ret = test_and_set_bit_lock(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD,
1057                                     &cinfo->state);
1058         WARN_ON_ONCE(ret);
1059         md_wakeup_thread(mddev->thread);
1060
1061         wait_event(cinfo->wait,
1062                    !test_and_set_bit(MD_CLUSTER_SEND_LOCK, &cinfo->state) ||
1063                    test_and_clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state));
1064
1065         /* If token is already locked, return 0 */
1066         if (cinfo->token_lockres->mode == DLM_LOCK_EX) {
1067                 clear_bit_unlock(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state);
1068                 return 0;
1069         }
1070
1071         ret = lock_token(cinfo);
1072         clear_bit_unlock(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state);
1073         return ret;
1074 }
1075
1076 static int metadata_update_finish(struct mddev *mddev)
1077 {
1078         struct md_cluster_info *cinfo = mddev->cluster_info;
1079         struct cluster_msg cmsg;
1080         struct md_rdev *rdev;
1081         int ret = 0;
1082         int raid_slot = -1;
1083
1084         memset(&cmsg, 0, sizeof(cmsg));
1085         cmsg.type = cpu_to_le32(METADATA_UPDATED);
1086         /* Pick up a good active device number to send.
1087          */
1088         rdev_for_each(rdev, mddev)
1089                 if (rdev->raid_disk > -1 && !test_bit(Faulty, &rdev->flags)) {
1090                         raid_slot = rdev->desc_nr;
1091                         break;
1092                 }
1093         if (raid_slot >= 0) {
1094                 cmsg.raid_slot = cpu_to_le32(raid_slot);
1095                 ret = __sendmsg(cinfo, &cmsg);
1096         } else
1097                 pr_warn("md-cluster: No good device id found to send\n");
1098         clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1099         unlock_comm(cinfo);
1100         return ret;
1101 }
1102
1103 static void metadata_update_cancel(struct mddev *mddev)
1104 {
1105         struct md_cluster_info *cinfo = mddev->cluster_info;
1106         clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1107         unlock_comm(cinfo);
1108 }
1109
1110 /*
1111  * return 0 if all the bitmaps have the same sync_size
1112  */
1113 static int cluster_check_sync_size(struct mddev *mddev)
1114 {
1115         int i, rv;
1116         bitmap_super_t *sb;
1117         unsigned long my_sync_size, sync_size = 0;
1118         int node_num = mddev->bitmap_info.nodes;
1119         int current_slot = md_cluster_ops->slot_number(mddev);
1120         struct bitmap *bitmap = mddev->bitmap;
1121         char str[64];
1122         struct dlm_lock_resource *bm_lockres;
1123
1124         sb = kmap_atomic(bitmap->storage.sb_page);
1125         my_sync_size = sb->sync_size;
1126         kunmap_atomic(sb);
1127
1128         for (i = 0; i < node_num; i++) {
1129                 if (i == current_slot)
1130                         continue;
1131
1132                 bitmap = get_bitmap_from_slot(mddev, i);
1133                 if (IS_ERR(bitmap)) {
1134                         pr_err("can't get bitmap from slot %d\n", i);
1135                         return -1;
1136                 }
1137
1138                 /*
1139                  * If we can hold the bitmap lock of one node then
1140                  * the slot is not occupied, update the sb.
1141                  */
1142                 snprintf(str, 64, "bitmap%04d", i);
1143                 bm_lockres = lockres_init(mddev, str, NULL, 1);
1144                 if (!bm_lockres) {
1145                         pr_err("md-cluster: Cannot initialize %s\n", str);
1146                         md_bitmap_free(bitmap);
1147                         return -1;
1148                 }
1149                 bm_lockres->flags |= DLM_LKF_NOQUEUE;
1150                 rv = dlm_lock_sync(bm_lockres, DLM_LOCK_PW);
1151                 if (!rv)
1152                         md_bitmap_update_sb(bitmap);
1153                 lockres_free(bm_lockres);
1154
1155                 sb = kmap_atomic(bitmap->storage.sb_page);
1156                 if (sync_size == 0)
1157                         sync_size = sb->sync_size;
1158                 else if (sync_size != sb->sync_size) {
1159                         kunmap_atomic(sb);
1160                         md_bitmap_free(bitmap);
1161                         return -1;
1162                 }
1163                 kunmap_atomic(sb);
1164                 md_bitmap_free(bitmap);
1165         }
1166
1167         return (my_sync_size == sync_size) ? 0 : -1;
1168 }
1169
1170 /*
1171  * Update the size for cluster raid is a little more complex, we perform it
1172  * by the steps:
1173  * 1. hold token lock and update superblock in initiator node.
1174  * 2. send METADATA_UPDATED msg to other nodes.
1175  * 3. The initiator node continues to check each bitmap's sync_size, if all
1176  *    bitmaps have the same value of sync_size, then we can set capacity and
1177  *    let other nodes to perform it. If one node can't update sync_size
1178  *    accordingly, we need to revert to previous value.
1179  */
1180 static void update_size(struct mddev *mddev, sector_t old_dev_sectors)
1181 {
1182         struct md_cluster_info *cinfo = mddev->cluster_info;
1183         struct cluster_msg cmsg;
1184         struct md_rdev *rdev;
1185         int ret = 0;
1186         int raid_slot = -1;
1187
1188         md_update_sb(mddev, 1);
1189         if (lock_comm(cinfo, 1)) {
1190                 pr_err("%s: lock_comm failed\n", __func__);
1191                 return;
1192         }
1193
1194         memset(&cmsg, 0, sizeof(cmsg));
1195         cmsg.type = cpu_to_le32(METADATA_UPDATED);
1196         rdev_for_each(rdev, mddev)
1197                 if (rdev->raid_disk >= 0 && !test_bit(Faulty, &rdev->flags)) {
1198                         raid_slot = rdev->desc_nr;
1199                         break;
1200                 }
1201         if (raid_slot >= 0) {
1202                 cmsg.raid_slot = cpu_to_le32(raid_slot);
1203                 /*
1204                  * We can only change capiticy after all the nodes can do it,
1205                  * so need to wait after other nodes already received the msg
1206                  * and handled the change
1207                  */
1208                 ret = __sendmsg(cinfo, &cmsg);
1209                 if (ret) {
1210                         pr_err("%s:%d: failed to send METADATA_UPDATED msg\n",
1211                                __func__, __LINE__);
1212                         unlock_comm(cinfo);
1213                         return;
1214                 }
1215         } else {
1216                 pr_err("md-cluster: No good device id found to send\n");
1217                 unlock_comm(cinfo);
1218                 return;
1219         }
1220
1221         /*
1222          * check the sync_size from other node's bitmap, if sync_size
1223          * have already updated in other nodes as expected, send an
1224          * empty metadata msg to permit the change of capacity
1225          */
1226         if (cluster_check_sync_size(mddev) == 0) {
1227                 memset(&cmsg, 0, sizeof(cmsg));
1228                 cmsg.type = cpu_to_le32(CHANGE_CAPACITY);
1229                 ret = __sendmsg(cinfo, &cmsg);
1230                 if (ret)
1231                         pr_err("%s:%d: failed to send CHANGE_CAPACITY msg\n",
1232                                __func__, __LINE__);
1233                 set_capacity(mddev->gendisk, mddev->array_sectors);
1234                 revalidate_disk(mddev->gendisk);
1235         } else {
1236                 /* revert to previous sectors */
1237                 ret = mddev->pers->resize(mddev, old_dev_sectors);
1238                 if (!ret)
1239                         revalidate_disk(mddev->gendisk);
1240                 ret = __sendmsg(cinfo, &cmsg);
1241                 if (ret)
1242                         pr_err("%s:%d: failed to send METADATA_UPDATED msg\n",
1243                                __func__, __LINE__);
1244         }
1245         unlock_comm(cinfo);
1246 }
1247
1248 static int resync_start(struct mddev *mddev)
1249 {
1250         struct md_cluster_info *cinfo = mddev->cluster_info;
1251         return dlm_lock_sync_interruptible(cinfo->resync_lockres, DLM_LOCK_EX, mddev);
1252 }
1253
1254 static int resync_info_update(struct mddev *mddev, sector_t lo, sector_t hi)
1255 {
1256         struct md_cluster_info *cinfo = mddev->cluster_info;
1257         struct resync_info ri;
1258         struct cluster_msg cmsg = {0};
1259
1260         /* do not send zero again, if we have sent before */
1261         if (hi == 0) {
1262                 memcpy(&ri, cinfo->bitmap_lockres->lksb.sb_lvbptr, sizeof(struct resync_info));
1263                 if (le64_to_cpu(ri.hi) == 0)
1264                         return 0;
1265         }
1266
1267         add_resync_info(cinfo->bitmap_lockres, lo, hi);
1268         /* Re-acquire the lock to refresh LVB */
1269         dlm_lock_sync(cinfo->bitmap_lockres, DLM_LOCK_PW);
1270         cmsg.type = cpu_to_le32(RESYNCING);
1271         cmsg.low = cpu_to_le64(lo);
1272         cmsg.high = cpu_to_le64(hi);
1273
1274         /*
1275          * mddev_lock is held if resync_info_update is called from
1276          * resync_finish (md_reap_sync_thread -> resync_finish)
1277          */
1278         if (lo == 0 && hi == 0)
1279                 return sendmsg(cinfo, &cmsg, 1);
1280         else
1281                 return sendmsg(cinfo, &cmsg, 0);
1282 }
1283
1284 static int resync_finish(struct mddev *mddev)
1285 {
1286         struct md_cluster_info *cinfo = mddev->cluster_info;
1287         int ret = 0;
1288
1289         clear_bit(MD_RESYNCING_REMOTE, &mddev->recovery);
1290
1291         /*
1292          * If resync thread is interrupted so we can't say resync is finished,
1293          * another node will launch resync thread to continue.
1294          */
1295         if (!test_bit(MD_CLOSING, &mddev->flags))
1296                 ret = resync_info_update(mddev, 0, 0);
1297         dlm_unlock_sync(cinfo->resync_lockres);
1298         return ret;
1299 }
1300
1301 static int area_resyncing(struct mddev *mddev, int direction,
1302                 sector_t lo, sector_t hi)
1303 {
1304         struct md_cluster_info *cinfo = mddev->cluster_info;
1305         int ret = 0;
1306         struct suspend_info *s;
1307
1308         if ((direction == READ) &&
1309                 test_bit(MD_CLUSTER_SUSPEND_READ_BALANCING, &cinfo->state))
1310                 return 1;
1311
1312         spin_lock_irq(&cinfo->suspend_lock);
1313         if (list_empty(&cinfo->suspend_list))
1314                 goto out;
1315         list_for_each_entry(s, &cinfo->suspend_list, list)
1316                 if (hi > s->lo && lo < s->hi) {
1317                         ret = 1;
1318                         break;
1319                 }
1320 out:
1321         spin_unlock_irq(&cinfo->suspend_lock);
1322         return ret;
1323 }
1324
1325 /* add_new_disk() - initiates a disk add
1326  * However, if this fails before writing md_update_sb(),
1327  * add_new_disk_cancel() must be called to release token lock
1328  */
1329 static int add_new_disk(struct mddev *mddev, struct md_rdev *rdev)
1330 {
1331         struct md_cluster_info *cinfo = mddev->cluster_info;
1332         struct cluster_msg cmsg;
1333         int ret = 0;
1334         struct mdp_superblock_1 *sb = page_address(rdev->sb_page);
1335         char *uuid = sb->device_uuid;
1336
1337         memset(&cmsg, 0, sizeof(cmsg));
1338         cmsg.type = cpu_to_le32(NEWDISK);
1339         memcpy(cmsg.uuid, uuid, 16);
1340         cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
1341         if (lock_comm(cinfo, 1))
1342                 return -EAGAIN;
1343         ret = __sendmsg(cinfo, &cmsg);
1344         if (ret) {
1345                 unlock_comm(cinfo);
1346                 return ret;
1347         }
1348         cinfo->no_new_dev_lockres->flags |= DLM_LKF_NOQUEUE;
1349         ret = dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_EX);
1350         cinfo->no_new_dev_lockres->flags &= ~DLM_LKF_NOQUEUE;
1351         /* Some node does not "see" the device */
1352         if (ret == -EAGAIN)
1353                 ret = -ENOENT;
1354         if (ret)
1355                 unlock_comm(cinfo);
1356         else {
1357                 dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR);
1358                 /* Since MD_CHANGE_DEVS will be set in add_bound_rdev which
1359                  * will run soon after add_new_disk, the below path will be
1360                  * invoked:
1361                  *   md_wakeup_thread(mddev->thread)
1362                  *      -> conf->thread (raid1d)
1363                  *      -> md_check_recovery -> md_update_sb
1364                  *      -> metadata_update_start/finish
1365                  * MD_CLUSTER_SEND_LOCKED_ALREADY will be cleared eventually.
1366                  *
1367                  * For other failure cases, metadata_update_cancel and
1368                  * add_new_disk_cancel also clear below bit as well.
1369                  * */
1370                 set_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1371                 wake_up(&cinfo->wait);
1372         }
1373         return ret;
1374 }
1375
1376 static void add_new_disk_cancel(struct mddev *mddev)
1377 {
1378         struct md_cluster_info *cinfo = mddev->cluster_info;
1379         clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1380         unlock_comm(cinfo);
1381 }
1382
1383 static int new_disk_ack(struct mddev *mddev, bool ack)
1384 {
1385         struct md_cluster_info *cinfo = mddev->cluster_info;
1386
1387         if (!test_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state)) {
1388                 pr_warn("md-cluster(%s): Spurious cluster confirmation\n", mdname(mddev));
1389                 return -EINVAL;
1390         }
1391
1392         if (ack)
1393                 dlm_unlock_sync(cinfo->no_new_dev_lockres);
1394         complete(&cinfo->newdisk_completion);
1395         return 0;
1396 }
1397
1398 static int remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1399 {
1400         struct cluster_msg cmsg = {0};
1401         struct md_cluster_info *cinfo = mddev->cluster_info;
1402         cmsg.type = cpu_to_le32(REMOVE);
1403         cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
1404         return sendmsg(cinfo, &cmsg, 1);
1405 }
1406
1407 static int lock_all_bitmaps(struct mddev *mddev)
1408 {
1409         int slot, my_slot, ret, held = 1, i = 0;
1410         char str[64];
1411         struct md_cluster_info *cinfo = mddev->cluster_info;
1412
1413         cinfo->other_bitmap_lockres =
1414                 kcalloc(mddev->bitmap_info.nodes - 1,
1415                         sizeof(struct dlm_lock_resource *), GFP_KERNEL);
1416         if (!cinfo->other_bitmap_lockres) {
1417                 pr_err("md: can't alloc mem for other bitmap locks\n");
1418                 return 0;
1419         }
1420
1421         my_slot = slot_number(mddev);
1422         for (slot = 0; slot < mddev->bitmap_info.nodes; slot++) {
1423                 if (slot == my_slot)
1424                         continue;
1425
1426                 memset(str, '\0', 64);
1427                 snprintf(str, 64, "bitmap%04d", slot);
1428                 cinfo->other_bitmap_lockres[i] = lockres_init(mddev, str, NULL, 1);
1429                 if (!cinfo->other_bitmap_lockres[i])
1430                         return -ENOMEM;
1431
1432                 cinfo->other_bitmap_lockres[i]->flags |= DLM_LKF_NOQUEUE;
1433                 ret = dlm_lock_sync(cinfo->other_bitmap_lockres[i], DLM_LOCK_PW);
1434                 if (ret)
1435                         held = -1;
1436                 i++;
1437         }
1438
1439         return held;
1440 }
1441
1442 static void unlock_all_bitmaps(struct mddev *mddev)
1443 {
1444         struct md_cluster_info *cinfo = mddev->cluster_info;
1445         int i;
1446
1447         /* release other node's bitmap lock if they are existed */
1448         if (cinfo->other_bitmap_lockres) {
1449                 for (i = 0; i < mddev->bitmap_info.nodes - 1; i++) {
1450                         if (cinfo->other_bitmap_lockres[i]) {
1451                                 lockres_free(cinfo->other_bitmap_lockres[i]);
1452                         }
1453                 }
1454                 kfree(cinfo->other_bitmap_lockres);
1455                 cinfo->other_bitmap_lockres = NULL;
1456         }
1457 }
1458
1459 static int gather_bitmaps(struct md_rdev *rdev)
1460 {
1461         int sn, err;
1462         sector_t lo, hi;
1463         struct cluster_msg cmsg = {0};
1464         struct mddev *mddev = rdev->mddev;
1465         struct md_cluster_info *cinfo = mddev->cluster_info;
1466
1467         cmsg.type = cpu_to_le32(RE_ADD);
1468         cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
1469         err = sendmsg(cinfo, &cmsg, 1);
1470         if (err)
1471                 goto out;
1472
1473         for (sn = 0; sn < mddev->bitmap_info.nodes; sn++) {
1474                 if (sn == (cinfo->slot_number - 1))
1475                         continue;
1476                 err = md_bitmap_copy_from_slot(mddev, sn, &lo, &hi, false);
1477                 if (err) {
1478                         pr_warn("md-cluster: Could not gather bitmaps from slot %d", sn);
1479                         goto out;
1480                 }
1481                 if ((hi > 0) && (lo < mddev->recovery_cp))
1482                         mddev->recovery_cp = lo;
1483         }
1484 out:
1485         return err;
1486 }
1487
1488 static struct md_cluster_operations cluster_ops = {
1489         .join   = join,
1490         .leave  = leave,
1491         .slot_number = slot_number,
1492         .resync_start = resync_start,
1493         .resync_finish = resync_finish,
1494         .resync_info_update = resync_info_update,
1495         .metadata_update_start = metadata_update_start,
1496         .metadata_update_finish = metadata_update_finish,
1497         .metadata_update_cancel = metadata_update_cancel,
1498         .area_resyncing = area_resyncing,
1499         .add_new_disk = add_new_disk,
1500         .add_new_disk_cancel = add_new_disk_cancel,
1501         .new_disk_ack = new_disk_ack,
1502         .remove_disk = remove_disk,
1503         .load_bitmaps = load_bitmaps,
1504         .gather_bitmaps = gather_bitmaps,
1505         .lock_all_bitmaps = lock_all_bitmaps,
1506         .unlock_all_bitmaps = unlock_all_bitmaps,
1507         .update_size = update_size,
1508 };
1509
1510 static int __init cluster_init(void)
1511 {
1512         pr_warn("md-cluster: support raid1 and raid10 (limited support)\n");
1513         pr_info("Registering Cluster MD functions\n");
1514         register_md_cluster_operations(&cluster_ops, THIS_MODULE);
1515         return 0;
1516 }
1517
1518 static void cluster_exit(void)
1519 {
1520         unregister_md_cluster_operations();
1521 }
1522
1523 module_init(cluster_init);
1524 module_exit(cluster_exit);
1525 MODULE_AUTHOR("SUSE");
1526 MODULE_LICENSE("GPL");
1527 MODULE_DESCRIPTION("Clustering support for MD");