Loading...
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * Copyright (C) 2015, SUSE
4 */
5
6
7#include <linux/module.h>
8#include <linux/kthread.h>
9#include <linux/dlm.h>
10#include <linux/sched.h>
11#include <linux/raid/md_p.h>
12#include "md.h"
13#include "md-bitmap.h"
14#include "md-cluster.h"
15
16#define LVB_SIZE 64
17#define NEW_DEV_TIMEOUT 5000
18#define WAIT_DLM_LOCK_TIMEOUT (30 * HZ)
19
20struct dlm_lock_resource {
21 dlm_lockspace_t *ls;
22 struct dlm_lksb lksb;
23 char *name; /* lock name. */
24 uint32_t flags; /* flags to pass to dlm_lock() */
25 wait_queue_head_t sync_locking; /* wait queue for synchronized locking */
26 bool sync_locking_done;
27 void (*bast)(void *arg, int mode); /* blocking AST function pointer*/
28 struct mddev *mddev; /* pointing back to mddev. */
29 int mode;
30};
31
32struct resync_info {
33 __le64 lo;
34 __le64 hi;
35};
36
37/* md_cluster_info flags */
38#define MD_CLUSTER_WAITING_FOR_NEWDISK 1
39#define MD_CLUSTER_SUSPEND_READ_BALANCING 2
40#define MD_CLUSTER_BEGIN_JOIN_CLUSTER 3
41
42/* Lock the send communication. This is done through
43 * bit manipulation as opposed to a mutex in order to
44 * accommodate lock and hold. See next comment.
45 */
46#define MD_CLUSTER_SEND_LOCK 4
47/* If cluster operations (such as adding a disk) must lock the
48 * communication channel, so as to perform extra operations
49 * (update metadata) and no other operation is allowed on the
50 * MD. Token needs to be locked and held until the operation
51 * completes witha md_update_sb(), which would eventually release
52 * the lock.
53 */
54#define MD_CLUSTER_SEND_LOCKED_ALREADY 5
55/* We should receive message after node joined cluster and
56 * set up all the related infos such as bitmap and personality */
57#define MD_CLUSTER_ALREADY_IN_CLUSTER 6
58#define MD_CLUSTER_PENDING_RECV_EVENT 7
59#define MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD 8
60#define MD_CLUSTER_WAITING_FOR_SYNC 9
61
62struct md_cluster_info {
63 struct mddev *mddev; /* the md device which md_cluster_info belongs to */
64 /* dlm lock space and resources for clustered raid. */
65 dlm_lockspace_t *lockspace;
66 int slot_number;
67 struct completion completion;
68 struct mutex recv_mutex;
69 struct dlm_lock_resource *bitmap_lockres;
70 struct dlm_lock_resource **other_bitmap_lockres;
71 struct dlm_lock_resource *resync_lockres;
72 struct list_head suspend_list;
73
74 spinlock_t suspend_lock;
75 /* record the region which write should be suspended */
76 sector_t suspend_lo;
77 sector_t suspend_hi;
78 int suspend_from; /* the slot which broadcast suspend_lo/hi */
79
80 struct md_thread __rcu *recovery_thread;
81 unsigned long recovery_map;
82 /* communication loc resources */
83 struct dlm_lock_resource *ack_lockres;
84 struct dlm_lock_resource *message_lockres;
85 struct dlm_lock_resource *token_lockres;
86 struct dlm_lock_resource *no_new_dev_lockres;
87 struct md_thread __rcu *recv_thread;
88 struct completion newdisk_completion;
89 wait_queue_head_t wait;
90 unsigned long state;
91 /* record the region in RESYNCING message */
92 sector_t sync_low;
93 sector_t sync_hi;
94};
95
96/* For compatibility, add the new msg_type at the end. */
97enum msg_type {
98 METADATA_UPDATED = 0,
99 RESYNCING,
100 NEWDISK,
101 REMOVE,
102 RE_ADD,
103 BITMAP_NEEDS_SYNC,
104 CHANGE_CAPACITY,
105 BITMAP_RESIZE,
106 RESYNCING_START,
107};
108
109struct cluster_msg {
110 __le32 type;
111 __le32 slot;
112 /* TODO: Unionize this for smaller footprint */
113 __le64 low;
114 __le64 high;
115 char uuid[16];
116 __le32 raid_slot;
117};
118
119static void sync_ast(void *arg)
120{
121 struct dlm_lock_resource *res;
122
123 res = arg;
124 res->sync_locking_done = true;
125 wake_up(&res->sync_locking);
126}
127
128static int dlm_lock_sync(struct dlm_lock_resource *res, int mode)
129{
130 int ret = 0;
131
132 ret = dlm_lock(res->ls, mode, &res->lksb,
133 res->flags, res->name, strlen(res->name),
134 0, sync_ast, res, res->bast);
135 if (ret)
136 return ret;
137 ret = wait_event_timeout(res->sync_locking, res->sync_locking_done,
138 WAIT_DLM_LOCK_TIMEOUT);
139 res->sync_locking_done = false;
140 if (!ret) {
141 pr_err("locking DLM '%s' timeout!\n", res->name);
142 return -EBUSY;
143 }
144 if (res->lksb.sb_status == 0)
145 res->mode = mode;
146 return res->lksb.sb_status;
147}
148
149static int dlm_unlock_sync(struct dlm_lock_resource *res)
150{
151 return dlm_lock_sync(res, DLM_LOCK_NL);
152}
153
154/*
155 * An variation of dlm_lock_sync, which make lock request could
156 * be interrupted
157 */
158static int dlm_lock_sync_interruptible(struct dlm_lock_resource *res, int mode,
159 struct mddev *mddev)
160{
161 int ret = 0;
162
163 ret = dlm_lock(res->ls, mode, &res->lksb,
164 res->flags, res->name, strlen(res->name),
165 0, sync_ast, res, res->bast);
166 if (ret)
167 return ret;
168
169 wait_event(res->sync_locking, res->sync_locking_done
170 || kthread_should_stop()
171 || test_bit(MD_CLOSING, &mddev->flags));
172 if (!res->sync_locking_done) {
173 /*
174 * the convert queue contains the lock request when request is
175 * interrupted, and sync_ast could still be run, so need to
176 * cancel the request and reset completion
177 */
178 ret = dlm_unlock(res->ls, res->lksb.sb_lkid, DLM_LKF_CANCEL,
179 &res->lksb, res);
180 res->sync_locking_done = false;
181 if (unlikely(ret != 0))
182 pr_info("failed to cancel previous lock request "
183 "%s return %d\n", res->name, ret);
184 return -EPERM;
185 } else
186 res->sync_locking_done = false;
187 if (res->lksb.sb_status == 0)
188 res->mode = mode;
189 return res->lksb.sb_status;
190}
191
192static struct dlm_lock_resource *lockres_init(struct mddev *mddev,
193 char *name, void (*bastfn)(void *arg, int mode), int with_lvb)
194{
195 struct dlm_lock_resource *res = NULL;
196 int ret, namelen;
197 struct md_cluster_info *cinfo = mddev->cluster_info;
198
199 res = kzalloc(sizeof(struct dlm_lock_resource), GFP_KERNEL);
200 if (!res)
201 return NULL;
202 init_waitqueue_head(&res->sync_locking);
203 res->sync_locking_done = false;
204 res->ls = cinfo->lockspace;
205 res->mddev = mddev;
206 res->mode = DLM_LOCK_IV;
207 namelen = strlen(name);
208 res->name = kzalloc(namelen + 1, GFP_KERNEL);
209 if (!res->name) {
210 pr_err("md-cluster: Unable to allocate resource name for resource %s\n", name);
211 goto out_err;
212 }
213 strscpy(res->name, name, namelen + 1);
214 if (with_lvb) {
215 res->lksb.sb_lvbptr = kzalloc(LVB_SIZE, GFP_KERNEL);
216 if (!res->lksb.sb_lvbptr) {
217 pr_err("md-cluster: Unable to allocate LVB for resource %s\n", name);
218 goto out_err;
219 }
220 res->flags = DLM_LKF_VALBLK;
221 }
222
223 if (bastfn)
224 res->bast = bastfn;
225
226 res->flags |= DLM_LKF_EXPEDITE;
227
228 ret = dlm_lock_sync(res, DLM_LOCK_NL);
229 if (ret) {
230 pr_err("md-cluster: Unable to lock NL on new lock resource %s\n", name);
231 goto out_err;
232 }
233 res->flags &= ~DLM_LKF_EXPEDITE;
234 res->flags |= DLM_LKF_CONVERT;
235
236 return res;
237out_err:
238 kfree(res->lksb.sb_lvbptr);
239 kfree(res->name);
240 kfree(res);
241 return NULL;
242}
243
244static void lockres_free(struct dlm_lock_resource *res)
245{
246 int ret = 0;
247
248 if (!res)
249 return;
250
251 /*
252 * use FORCEUNLOCK flag, so we can unlock even the lock is on the
253 * waiting or convert queue
254 */
255 ret = dlm_unlock(res->ls, res->lksb.sb_lkid, DLM_LKF_FORCEUNLOCK,
256 &res->lksb, res);
257 if (unlikely(ret != 0))
258 pr_err("failed to unlock %s return %d\n", res->name, ret);
259 else
260 wait_event(res->sync_locking, res->sync_locking_done);
261
262 kfree(res->name);
263 kfree(res->lksb.sb_lvbptr);
264 kfree(res);
265}
266
267static void add_resync_info(struct dlm_lock_resource *lockres,
268 sector_t lo, sector_t hi)
269{
270 struct resync_info *ri;
271
272 ri = (struct resync_info *)lockres->lksb.sb_lvbptr;
273 ri->lo = cpu_to_le64(lo);
274 ri->hi = cpu_to_le64(hi);
275}
276
277static int read_resync_info(struct mddev *mddev,
278 struct dlm_lock_resource *lockres)
279{
280 struct resync_info ri;
281 struct md_cluster_info *cinfo = mddev->cluster_info;
282 int ret = 0;
283
284 dlm_lock_sync(lockres, DLM_LOCK_CR);
285 memcpy(&ri, lockres->lksb.sb_lvbptr, sizeof(struct resync_info));
286 if (le64_to_cpu(ri.hi) > 0) {
287 cinfo->suspend_hi = le64_to_cpu(ri.hi);
288 cinfo->suspend_lo = le64_to_cpu(ri.lo);
289 ret = 1;
290 }
291 dlm_unlock_sync(lockres);
292 return ret;
293}
294
295static void recover_bitmaps(struct md_thread *thread)
296{
297 struct mddev *mddev = thread->mddev;
298 struct md_cluster_info *cinfo = mddev->cluster_info;
299 struct dlm_lock_resource *bm_lockres;
300 char str[64];
301 int slot, ret;
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 = mddev->bitmap_ops->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 cinfo->suspend_hi = 0;
329 cinfo->suspend_lo = 0;
330 cinfo->suspend_from = -1;
331 spin_unlock_irq(&cinfo->suspend_lock);
332
333 /* Kick off a reshape if needed */
334 if (test_bit(MD_RESYNCING_REMOTE, &mddev->recovery) &&
335 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
336 mddev->reshape_position != MaxSector)
337 md_wakeup_thread(mddev->sync_thread);
338
339 if (hi > 0) {
340 if (lo < mddev->recovery_cp)
341 mddev->recovery_cp = lo;
342 /* wake up thread to continue resync in case resync
343 * is not finished */
344 if (mddev->recovery_cp != MaxSector) {
345 /*
346 * clear the REMOTE flag since we will launch
347 * resync thread in current node.
348 */
349 clear_bit(MD_RESYNCING_REMOTE,
350 &mddev->recovery);
351 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
352 md_wakeup_thread(mddev->thread);
353 }
354 }
355clear_bit:
356 lockres_free(bm_lockres);
357 clear_bit(slot, &cinfo->recovery_map);
358 }
359}
360
361static void recover_prep(void *arg)
362{
363 struct mddev *mddev = arg;
364 struct md_cluster_info *cinfo = mddev->cluster_info;
365 set_bit(MD_CLUSTER_SUSPEND_READ_BALANCING, &cinfo->state);
366}
367
368static void __recover_slot(struct mddev *mddev, int slot)
369{
370 struct md_cluster_info *cinfo = mddev->cluster_info;
371
372 set_bit(slot, &cinfo->recovery_map);
373 if (!cinfo->recovery_thread) {
374 rcu_assign_pointer(cinfo->recovery_thread,
375 md_register_thread(recover_bitmaps, mddev, "recover"));
376 if (!cinfo->recovery_thread) {
377 pr_warn("md-cluster: Could not create recovery thread\n");
378 return;
379 }
380 }
381 md_wakeup_thread(cinfo->recovery_thread);
382}
383
384static void recover_slot(void *arg, struct dlm_slot *slot)
385{
386 struct mddev *mddev = arg;
387 struct md_cluster_info *cinfo = mddev->cluster_info;
388
389 pr_info("md-cluster: %s Node %d/%d down. My slot: %d. Initiating recovery.\n",
390 mddev->bitmap_info.cluster_name,
391 slot->nodeid, slot->slot,
392 cinfo->slot_number);
393 /* deduct one since dlm slot starts from one while the num of
394 * cluster-md begins with 0 */
395 __recover_slot(mddev, slot->slot - 1);
396}
397
398static void recover_done(void *arg, struct dlm_slot *slots,
399 int num_slots, int our_slot,
400 uint32_t generation)
401{
402 struct mddev *mddev = arg;
403 struct md_cluster_info *cinfo = mddev->cluster_info;
404
405 cinfo->slot_number = our_slot;
406 /* completion is only need to be complete when node join cluster,
407 * it doesn't need to run during another node's failure */
408 if (test_bit(MD_CLUSTER_BEGIN_JOIN_CLUSTER, &cinfo->state)) {
409 complete(&cinfo->completion);
410 clear_bit(MD_CLUSTER_BEGIN_JOIN_CLUSTER, &cinfo->state);
411 }
412 clear_bit(MD_CLUSTER_SUSPEND_READ_BALANCING, &cinfo->state);
413}
414
415/* the ops is called when node join the cluster, and do lock recovery
416 * if node failure occurs */
417static const struct dlm_lockspace_ops md_ls_ops = {
418 .recover_prep = recover_prep,
419 .recover_slot = recover_slot,
420 .recover_done = recover_done,
421};
422
423/*
424 * The BAST function for the ack lock resource
425 * This function wakes up the receive thread in
426 * order to receive and process the message.
427 */
428static void ack_bast(void *arg, int mode)
429{
430 struct dlm_lock_resource *res = arg;
431 struct md_cluster_info *cinfo = res->mddev->cluster_info;
432
433 if (mode == DLM_LOCK_EX) {
434 if (test_bit(MD_CLUSTER_ALREADY_IN_CLUSTER, &cinfo->state))
435 md_wakeup_thread(cinfo->recv_thread);
436 else
437 set_bit(MD_CLUSTER_PENDING_RECV_EVENT, &cinfo->state);
438 }
439}
440
441static void remove_suspend_info(struct mddev *mddev, int slot)
442{
443 struct md_cluster_info *cinfo = mddev->cluster_info;
444 mddev->pers->quiesce(mddev, 1);
445 spin_lock_irq(&cinfo->suspend_lock);
446 cinfo->suspend_hi = 0;
447 cinfo->suspend_lo = 0;
448 spin_unlock_irq(&cinfo->suspend_lock);
449 mddev->pers->quiesce(mddev, 0);
450}
451
452static void process_suspend_info(struct mddev *mddev,
453 int slot, sector_t lo, sector_t hi)
454{
455 struct md_cluster_info *cinfo = mddev->cluster_info;
456 struct mdp_superblock_1 *sb = NULL;
457 struct md_rdev *rdev;
458
459 if (!hi) {
460 /*
461 * clear the REMOTE flag since resync or recovery is finished
462 * in remote node.
463 */
464 clear_bit(MD_RESYNCING_REMOTE, &mddev->recovery);
465 remove_suspend_info(mddev, slot);
466 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
467 clear_bit(MD_CLUSTER_WAITING_FOR_SYNC, &cinfo->state);
468 md_wakeup_thread(mddev->thread);
469 return;
470 }
471
472 rdev_for_each(rdev, mddev)
473 if (rdev->raid_disk > -1 && !test_bit(Faulty, &rdev->flags)) {
474 sb = page_address(rdev->sb_page);
475 break;
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 md_bitmap_sync_with_cluster to clear NEEDED_MASK
491 * and set RESYNC_MASK since resync thread is running
492 * in another node, so we don't need to do the resync
493 * again with the same section.
494 *
495 * Skip md_bitmap_sync_with_cluster in case reshape
496 * happening, because reshaping region is small and
497 * we don't want to trigger lots of WARN.
498 */
499 if (sb && !(le32_to_cpu(sb->feature_map) & MD_FEATURE_RESHAPE_ACTIVE))
500 mddev->bitmap_ops->sync_with_cluster(mddev, cinfo->sync_low,
501 cinfo->sync_hi, lo, hi);
502 cinfo->sync_low = lo;
503 cinfo->sync_hi = hi;
504
505 mddev->pers->quiesce(mddev, 1);
506 spin_lock_irq(&cinfo->suspend_lock);
507 cinfo->suspend_from = slot;
508 cinfo->suspend_lo = lo;
509 cinfo->suspend_hi = hi;
510 spin_unlock_irq(&cinfo->suspend_lock);
511 mddev->pers->quiesce(mddev, 0);
512}
513
514static int 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 int res = 0;
523
524 len = snprintf(disk_uuid, 64, "DEVICE_UUID=");
525 sprintf(disk_uuid + len, "%pU", cmsg->uuid);
526 snprintf(raid_slot, 16, "RAID_DISK=%d", le32_to_cpu(cmsg->raid_slot));
527 pr_info("%s:%d Sending kobject change with %s and %s\n", __func__, __LINE__, disk_uuid, raid_slot);
528 init_completion(&cinfo->newdisk_completion);
529 set_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state);
530 kobject_uevent_env(&disk_to_dev(mddev->gendisk)->kobj, KOBJ_CHANGE, envp);
531 if (!wait_for_completion_timeout(&cinfo->newdisk_completion,
532 NEW_DEV_TIMEOUT)) {
533 pr_err("md-cluster(%s:%d): timeout on a new disk adding\n",
534 __func__, __LINE__);
535 res = -1;
536 }
537 clear_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state);
538 set_bit(MD_CLUSTER_WAITING_FOR_SYNC, &cinfo->state);
539 return res;
540}
541
542
543static void process_metadata_update(struct mddev *mddev, struct cluster_msg *msg)
544{
545 int got_lock = 0;
546 struct md_thread *thread;
547 struct md_cluster_info *cinfo = mddev->cluster_info;
548 mddev->good_device_nr = le32_to_cpu(msg->raid_slot);
549
550 dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR);
551
552 /* daemaon thread must exist */
553 thread = rcu_dereference_protected(mddev->thread, true);
554 wait_event(thread->wqueue,
555 (got_lock = mddev_trylock(mddev)) ||
556 test_bit(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state));
557 md_reload_sb(mddev, mddev->good_device_nr);
558 if (got_lock)
559 mddev_unlock(mddev);
560}
561
562static void process_remove_disk(struct mddev *mddev, struct cluster_msg *msg)
563{
564 struct md_rdev *rdev;
565
566 rcu_read_lock();
567 rdev = md_find_rdev_nr_rcu(mddev, le32_to_cpu(msg->raid_slot));
568 if (rdev) {
569 set_bit(ClusterRemove, &rdev->flags);
570 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
571 md_wakeup_thread(mddev->thread);
572 }
573 else
574 pr_warn("%s: %d Could not find disk(%d) to REMOVE\n",
575 __func__, __LINE__, le32_to_cpu(msg->raid_slot));
576 rcu_read_unlock();
577}
578
579static void process_readd_disk(struct mddev *mddev, struct cluster_msg *msg)
580{
581 struct md_rdev *rdev;
582
583 rcu_read_lock();
584 rdev = md_find_rdev_nr_rcu(mddev, le32_to_cpu(msg->raid_slot));
585 if (rdev && test_bit(Faulty, &rdev->flags))
586 clear_bit(Faulty, &rdev->flags);
587 else
588 pr_warn("%s: %d Could not find disk(%d) which is faulty",
589 __func__, __LINE__, le32_to_cpu(msg->raid_slot));
590 rcu_read_unlock();
591}
592
593static int process_recvd_msg(struct mddev *mddev, struct cluster_msg *msg)
594{
595 int ret = 0;
596
597 if (WARN(mddev->cluster_info->slot_number - 1 == le32_to_cpu(msg->slot),
598 "node %d received its own msg\n", le32_to_cpu(msg->slot)))
599 return -1;
600 switch (le32_to_cpu(msg->type)) {
601 case METADATA_UPDATED:
602 process_metadata_update(mddev, msg);
603 break;
604 case CHANGE_CAPACITY:
605 set_capacity_and_notify(mddev->gendisk, mddev->array_sectors);
606 break;
607 case RESYNCING_START:
608 clear_bit(MD_CLUSTER_WAITING_FOR_SYNC, &mddev->cluster_info->state);
609 break;
610 case RESYNCING:
611 set_bit(MD_RESYNCING_REMOTE, &mddev->recovery);
612 process_suspend_info(mddev, le32_to_cpu(msg->slot),
613 le64_to_cpu(msg->low),
614 le64_to_cpu(msg->high));
615 break;
616 case NEWDISK:
617 if (process_add_new_disk(mddev, msg))
618 ret = -1;
619 break;
620 case REMOVE:
621 process_remove_disk(mddev, msg);
622 break;
623 case RE_ADD:
624 process_readd_disk(mddev, msg);
625 break;
626 case BITMAP_NEEDS_SYNC:
627 __recover_slot(mddev, le32_to_cpu(msg->slot));
628 break;
629 case BITMAP_RESIZE:
630 if (le64_to_cpu(msg->high) != mddev->pers->size(mddev, 0, 0))
631 ret = mddev->bitmap_ops->resize(mddev,
632 le64_to_cpu(msg->high),
633 0, false);
634 break;
635 default:
636 ret = -1;
637 pr_warn("%s:%d Received unknown message from %d\n",
638 __func__, __LINE__, msg->slot);
639 }
640 return ret;
641}
642
643/*
644 * thread for receiving message
645 */
646static void recv_daemon(struct md_thread *thread)
647{
648 struct md_cluster_info *cinfo = thread->mddev->cluster_info;
649 struct dlm_lock_resource *ack_lockres = cinfo->ack_lockres;
650 struct dlm_lock_resource *message_lockres = cinfo->message_lockres;
651 struct cluster_msg msg;
652 int ret;
653
654 mutex_lock(&cinfo->recv_mutex);
655 /*get CR on Message*/
656 if (dlm_lock_sync(message_lockres, DLM_LOCK_CR)) {
657 pr_err("md/raid1:failed to get CR on MESSAGE\n");
658 mutex_unlock(&cinfo->recv_mutex);
659 return;
660 }
661
662 /* read lvb and wake up thread to process this message_lockres */
663 memcpy(&msg, message_lockres->lksb.sb_lvbptr, sizeof(struct cluster_msg));
664 ret = process_recvd_msg(thread->mddev, &msg);
665 if (ret)
666 goto out;
667
668 /*release CR on ack_lockres*/
669 ret = dlm_unlock_sync(ack_lockres);
670 if (unlikely(ret != 0))
671 pr_info("unlock ack failed return %d\n", ret);
672 /*up-convert to PR on message_lockres*/
673 ret = dlm_lock_sync(message_lockres, DLM_LOCK_PR);
674 if (unlikely(ret != 0))
675 pr_info("lock PR on msg failed return %d\n", ret);
676 /*get CR on ack_lockres again*/
677 ret = dlm_lock_sync(ack_lockres, DLM_LOCK_CR);
678 if (unlikely(ret != 0))
679 pr_info("lock CR on ack failed return %d\n", ret);
680out:
681 /*release CR on message_lockres*/
682 ret = dlm_unlock_sync(message_lockres);
683 if (unlikely(ret != 0))
684 pr_info("unlock msg failed return %d\n", ret);
685 mutex_unlock(&cinfo->recv_mutex);
686}
687
688/* lock_token()
689 * Takes the lock on the TOKEN lock resource so no other
690 * node can communicate while the operation is underway.
691 */
692static int lock_token(struct md_cluster_info *cinfo)
693{
694 int error;
695
696 error = dlm_lock_sync(cinfo->token_lockres, DLM_LOCK_EX);
697 if (error) {
698 pr_err("md-cluster(%s:%d): failed to get EX on TOKEN (%d)\n",
699 __func__, __LINE__, error);
700 } else {
701 /* Lock the receive sequence */
702 mutex_lock(&cinfo->recv_mutex);
703 }
704 return error;
705}
706
707/* lock_comm()
708 * Sets the MD_CLUSTER_SEND_LOCK bit to lock the send channel.
709 */
710static int lock_comm(struct md_cluster_info *cinfo, bool mddev_locked)
711{
712 int rv, set_bit = 0;
713 struct mddev *mddev = cinfo->mddev;
714
715 /*
716 * If resync thread run after raid1d thread, then process_metadata_update
717 * could not continue if raid1d held reconfig_mutex (and raid1d is blocked
718 * since another node already got EX on Token and waiting the EX of Ack),
719 * so let resync wake up thread in case flag is set.
720 */
721 if (mddev_locked && !test_bit(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD,
722 &cinfo->state)) {
723 rv = test_and_set_bit_lock(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD,
724 &cinfo->state);
725 WARN_ON_ONCE(rv);
726 md_wakeup_thread(mddev->thread);
727 set_bit = 1;
728 }
729
730 wait_event(cinfo->wait,
731 !test_and_set_bit(MD_CLUSTER_SEND_LOCK, &cinfo->state));
732 rv = lock_token(cinfo);
733 if (set_bit)
734 clear_bit_unlock(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state);
735 return rv;
736}
737
738static void unlock_comm(struct md_cluster_info *cinfo)
739{
740 WARN_ON(cinfo->token_lockres->mode != DLM_LOCK_EX);
741 mutex_unlock(&cinfo->recv_mutex);
742 dlm_unlock_sync(cinfo->token_lockres);
743 clear_bit(MD_CLUSTER_SEND_LOCK, &cinfo->state);
744 wake_up(&cinfo->wait);
745}
746
747/* __sendmsg()
748 * This function performs the actual sending of the message. This function is
749 * usually called after performing the encompassing operation
750 * The function:
751 * 1. Grabs the message lockresource in EX mode
752 * 2. Copies the message to the message LVB
753 * 3. Downconverts message lockresource to CW
754 * 4. Upconverts ack lock resource from CR to EX. This forces the BAST on other nodes
755 * and the other nodes read the message. The thread will wait here until all other
756 * nodes have released ack lock resource.
757 * 5. Downconvert ack lockresource to CR
758 */
759static int __sendmsg(struct md_cluster_info *cinfo, struct cluster_msg *cmsg)
760{
761 int error, unlock_error;
762 int slot = cinfo->slot_number - 1;
763
764 cmsg->slot = cpu_to_le32(slot);
765 /*get EX on Message*/
766 error = dlm_lock_sync(cinfo->message_lockres, DLM_LOCK_EX);
767 if (error) {
768 pr_err("md-cluster: failed to get EX on MESSAGE (%d)\n", error);
769 return error;
770 }
771
772 memcpy(cinfo->message_lockres->lksb.sb_lvbptr, (void *)cmsg,
773 sizeof(struct cluster_msg));
774 /*down-convert EX to CW on Message*/
775 error = dlm_lock_sync(cinfo->message_lockres, DLM_LOCK_CW);
776 if (error) {
777 pr_err("md-cluster: failed to convert EX to CW on MESSAGE(%d)\n",
778 error);
779 goto failed_ack;
780 }
781
782 /*up-convert CR to EX on Ack*/
783 error = dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_EX);
784 if (error) {
785 pr_err("md-cluster: failed to convert CR to EX on ACK(%d)\n",
786 error);
787 goto failed_ack;
788 }
789
790 /*down-convert EX to CR on Ack*/
791 error = dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_CR);
792 if (error) {
793 pr_err("md-cluster: failed to convert EX to CR on ACK(%d)\n",
794 error);
795 goto failed_ack;
796 }
797
798failed_ack:
799 while ((unlock_error = dlm_unlock_sync(cinfo->message_lockres)))
800 pr_err("md-cluster: failed convert to NL on MESSAGE(%d)\n",
801 unlock_error);
802
803 return error;
804}
805
806static int sendmsg(struct md_cluster_info *cinfo, struct cluster_msg *cmsg,
807 bool mddev_locked)
808{
809 int ret;
810
811 ret = lock_comm(cinfo, mddev_locked);
812 if (!ret) {
813 ret = __sendmsg(cinfo, cmsg);
814 unlock_comm(cinfo);
815 }
816 return ret;
817}
818
819static int gather_all_resync_info(struct mddev *mddev, int total_slots)
820{
821 struct md_cluster_info *cinfo = mddev->cluster_info;
822 int i, ret = 0;
823 struct dlm_lock_resource *bm_lockres;
824 char str[64];
825 sector_t lo, hi;
826
827
828 for (i = 0; i < total_slots; i++) {
829 memset(str, '\0', 64);
830 snprintf(str, 64, "bitmap%04d", i);
831 bm_lockres = lockres_init(mddev, str, NULL, 1);
832 if (!bm_lockres)
833 return -ENOMEM;
834 if (i == (cinfo->slot_number - 1)) {
835 lockres_free(bm_lockres);
836 continue;
837 }
838
839 bm_lockres->flags |= DLM_LKF_NOQUEUE;
840 ret = dlm_lock_sync(bm_lockres, DLM_LOCK_PW);
841 if (ret == -EAGAIN) {
842 if (read_resync_info(mddev, bm_lockres)) {
843 pr_info("%s:%d Resync[%llu..%llu] in progress on %d\n",
844 __func__, __LINE__,
845 (unsigned long long) cinfo->suspend_lo,
846 (unsigned long long) cinfo->suspend_hi,
847 i);
848 cinfo->suspend_from = i;
849 }
850 ret = 0;
851 lockres_free(bm_lockres);
852 continue;
853 }
854 if (ret) {
855 lockres_free(bm_lockres);
856 goto out;
857 }
858
859 /* Read the disk bitmap sb and check if it needs recovery */
860 ret = mddev->bitmap_ops->copy_from_slot(mddev, i, &lo, &hi, false);
861 if (ret) {
862 pr_warn("md-cluster: Could not gather bitmaps from slot %d", i);
863 lockres_free(bm_lockres);
864 continue;
865 }
866 if ((hi > 0) && (lo < mddev->recovery_cp)) {
867 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
868 mddev->recovery_cp = lo;
869 md_check_recovery(mddev);
870 }
871
872 lockres_free(bm_lockres);
873 }
874out:
875 return ret;
876}
877
878static int join(struct mddev *mddev, int nodes)
879{
880 struct md_cluster_info *cinfo;
881 int ret, ops_rv;
882 char str[64];
883
884 cinfo = kzalloc(sizeof(struct md_cluster_info), GFP_KERNEL);
885 if (!cinfo)
886 return -ENOMEM;
887
888 INIT_LIST_HEAD(&cinfo->suspend_list);
889 spin_lock_init(&cinfo->suspend_lock);
890 init_completion(&cinfo->completion);
891 set_bit(MD_CLUSTER_BEGIN_JOIN_CLUSTER, &cinfo->state);
892 init_waitqueue_head(&cinfo->wait);
893 mutex_init(&cinfo->recv_mutex);
894
895 mddev->cluster_info = cinfo;
896 cinfo->mddev = mddev;
897
898 memset(str, 0, 64);
899 sprintf(str, "%pU", mddev->uuid);
900 ret = dlm_new_lockspace(str, mddev->bitmap_info.cluster_name,
901 DLM_LSFL_SOFTIRQ, LVB_SIZE, &md_ls_ops, mddev,
902 &ops_rv, &cinfo->lockspace);
903 if (ret)
904 goto err;
905 wait_for_completion(&cinfo->completion);
906 if (nodes < cinfo->slot_number) {
907 pr_err("md-cluster: Slot allotted(%d) is greater than available slots(%d).",
908 cinfo->slot_number, nodes);
909 ret = -ERANGE;
910 goto err;
911 }
912 /* Initiate the communication resources */
913 ret = -ENOMEM;
914 rcu_assign_pointer(cinfo->recv_thread,
915 md_register_thread(recv_daemon, mddev, "cluster_recv"));
916 if (!cinfo->recv_thread) {
917 pr_err("md-cluster: cannot allocate memory for recv_thread!\n");
918 goto err;
919 }
920 cinfo->message_lockres = lockres_init(mddev, "message", NULL, 1);
921 if (!cinfo->message_lockres)
922 goto err;
923 cinfo->token_lockres = lockres_init(mddev, "token", NULL, 0);
924 if (!cinfo->token_lockres)
925 goto err;
926 cinfo->no_new_dev_lockres = lockres_init(mddev, "no-new-dev", NULL, 0);
927 if (!cinfo->no_new_dev_lockres)
928 goto err;
929
930 ret = dlm_lock_sync(cinfo->token_lockres, DLM_LOCK_EX);
931 if (ret) {
932 ret = -EAGAIN;
933 pr_err("md-cluster: can't join cluster to avoid lock issue\n");
934 goto err;
935 }
936 cinfo->ack_lockres = lockres_init(mddev, "ack", ack_bast, 0);
937 if (!cinfo->ack_lockres) {
938 ret = -ENOMEM;
939 goto err;
940 }
941 /* get sync CR lock on ACK. */
942 if (dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_CR))
943 pr_err("md-cluster: failed to get a sync CR lock on ACK!(%d)\n",
944 ret);
945 dlm_unlock_sync(cinfo->token_lockres);
946 /* get sync CR lock on no-new-dev. */
947 if (dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR))
948 pr_err("md-cluster: failed to get a sync CR lock on no-new-dev!(%d)\n", ret);
949
950
951 pr_info("md-cluster: Joined cluster %s slot %d\n", str, cinfo->slot_number);
952 snprintf(str, 64, "bitmap%04d", cinfo->slot_number - 1);
953 cinfo->bitmap_lockres = lockres_init(mddev, str, NULL, 1);
954 if (!cinfo->bitmap_lockres) {
955 ret = -ENOMEM;
956 goto err;
957 }
958 if (dlm_lock_sync(cinfo->bitmap_lockres, DLM_LOCK_PW)) {
959 pr_err("Failed to get bitmap lock\n");
960 ret = -EINVAL;
961 goto err;
962 }
963
964 cinfo->resync_lockres = lockres_init(mddev, "resync", NULL, 0);
965 if (!cinfo->resync_lockres) {
966 ret = -ENOMEM;
967 goto err;
968 }
969
970 return 0;
971err:
972 set_bit(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state);
973 md_unregister_thread(mddev, &cinfo->recovery_thread);
974 md_unregister_thread(mddev, &cinfo->recv_thread);
975 lockres_free(cinfo->message_lockres);
976 lockres_free(cinfo->token_lockres);
977 lockres_free(cinfo->ack_lockres);
978 lockres_free(cinfo->no_new_dev_lockres);
979 lockres_free(cinfo->resync_lockres);
980 lockres_free(cinfo->bitmap_lockres);
981 if (cinfo->lockspace)
982 dlm_release_lockspace(cinfo->lockspace, 2);
983 mddev->cluster_info = NULL;
984 kfree(cinfo);
985 return ret;
986}
987
988static void load_bitmaps(struct mddev *mddev, int total_slots)
989{
990 struct md_cluster_info *cinfo = mddev->cluster_info;
991
992 /* load all the node's bitmap info for resync */
993 if (gather_all_resync_info(mddev, total_slots))
994 pr_err("md-cluster: failed to gather all resyn infos\n");
995 set_bit(MD_CLUSTER_ALREADY_IN_CLUSTER, &cinfo->state);
996 /* wake up recv thread in case something need to be handled */
997 if (test_and_clear_bit(MD_CLUSTER_PENDING_RECV_EVENT, &cinfo->state))
998 md_wakeup_thread(cinfo->recv_thread);
999}
1000
1001static void resync_bitmap(struct mddev *mddev)
1002{
1003 struct md_cluster_info *cinfo = mddev->cluster_info;
1004 struct cluster_msg cmsg = {0};
1005 int err;
1006
1007 cmsg.type = cpu_to_le32(BITMAP_NEEDS_SYNC);
1008 err = sendmsg(cinfo, &cmsg, 1);
1009 if (err)
1010 pr_err("%s:%d: failed to send BITMAP_NEEDS_SYNC message (%d)\n",
1011 __func__, __LINE__, err);
1012}
1013
1014static void unlock_all_bitmaps(struct mddev *mddev);
1015static int leave(struct mddev *mddev)
1016{
1017 struct md_cluster_info *cinfo = mddev->cluster_info;
1018
1019 if (!cinfo)
1020 return 0;
1021
1022 /*
1023 * BITMAP_NEEDS_SYNC message should be sent when node
1024 * is leaving the cluster with dirty bitmap, also we
1025 * can only deliver it when dlm connection is available.
1026 *
1027 * Also, we should send BITMAP_NEEDS_SYNC message in
1028 * case reshaping is interrupted.
1029 */
1030 if ((cinfo->slot_number > 0 && mddev->recovery_cp != MaxSector) ||
1031 (mddev->reshape_position != MaxSector &&
1032 test_bit(MD_CLOSING, &mddev->flags)))
1033 resync_bitmap(mddev);
1034
1035 set_bit(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state);
1036 md_unregister_thread(mddev, &cinfo->recovery_thread);
1037 md_unregister_thread(mddev, &cinfo->recv_thread);
1038 lockres_free(cinfo->message_lockres);
1039 lockres_free(cinfo->token_lockres);
1040 lockres_free(cinfo->ack_lockres);
1041 lockres_free(cinfo->no_new_dev_lockres);
1042 lockres_free(cinfo->resync_lockres);
1043 lockres_free(cinfo->bitmap_lockres);
1044 unlock_all_bitmaps(mddev);
1045 dlm_release_lockspace(cinfo->lockspace, 2);
1046 kfree(cinfo);
1047 return 0;
1048}
1049
1050/* slot_number(): Returns the MD slot number to use
1051 * DLM starts the slot numbers from 1, wheras cluster-md
1052 * wants the number to be from zero, so we deduct one
1053 */
1054static int slot_number(struct mddev *mddev)
1055{
1056 struct md_cluster_info *cinfo = mddev->cluster_info;
1057
1058 return cinfo->slot_number - 1;
1059}
1060
1061/*
1062 * Check if the communication is already locked, else lock the communication
1063 * channel.
1064 * If it is already locked, token is in EX mode, and hence lock_token()
1065 * should not be called.
1066 */
1067static int metadata_update_start(struct mddev *mddev)
1068{
1069 struct md_cluster_info *cinfo = mddev->cluster_info;
1070 int ret;
1071
1072 /*
1073 * metadata_update_start is always called with the protection of
1074 * reconfig_mutex, so set WAITING_FOR_TOKEN here.
1075 */
1076 ret = test_and_set_bit_lock(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD,
1077 &cinfo->state);
1078 WARN_ON_ONCE(ret);
1079 md_wakeup_thread(mddev->thread);
1080
1081 wait_event(cinfo->wait,
1082 !test_and_set_bit(MD_CLUSTER_SEND_LOCK, &cinfo->state) ||
1083 test_and_clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state));
1084
1085 /* If token is already locked, return 0 */
1086 if (cinfo->token_lockres->mode == DLM_LOCK_EX) {
1087 clear_bit_unlock(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state);
1088 return 0;
1089 }
1090
1091 ret = lock_token(cinfo);
1092 clear_bit_unlock(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state);
1093 return ret;
1094}
1095
1096static int metadata_update_finish(struct mddev *mddev)
1097{
1098 struct md_cluster_info *cinfo = mddev->cluster_info;
1099 struct cluster_msg cmsg;
1100 struct md_rdev *rdev;
1101 int ret = 0;
1102 int raid_slot = -1;
1103
1104 memset(&cmsg, 0, sizeof(cmsg));
1105 cmsg.type = cpu_to_le32(METADATA_UPDATED);
1106 /* Pick up a good active device number to send.
1107 */
1108 rdev_for_each(rdev, mddev)
1109 if (rdev->raid_disk > -1 && !test_bit(Faulty, &rdev->flags)) {
1110 raid_slot = rdev->desc_nr;
1111 break;
1112 }
1113 if (raid_slot >= 0) {
1114 cmsg.raid_slot = cpu_to_le32(raid_slot);
1115 ret = __sendmsg(cinfo, &cmsg);
1116 } else
1117 pr_warn("md-cluster: No good device id found to send\n");
1118 clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1119 unlock_comm(cinfo);
1120 return ret;
1121}
1122
1123static void metadata_update_cancel(struct mddev *mddev)
1124{
1125 struct md_cluster_info *cinfo = mddev->cluster_info;
1126 clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1127 unlock_comm(cinfo);
1128}
1129
1130static int update_bitmap_size(struct mddev *mddev, sector_t size)
1131{
1132 struct md_cluster_info *cinfo = mddev->cluster_info;
1133 struct cluster_msg cmsg = {0};
1134 int ret;
1135
1136 cmsg.type = cpu_to_le32(BITMAP_RESIZE);
1137 cmsg.high = cpu_to_le64(size);
1138 ret = sendmsg(cinfo, &cmsg, 0);
1139 if (ret)
1140 pr_err("%s:%d: failed to send BITMAP_RESIZE message (%d)\n",
1141 __func__, __LINE__, ret);
1142 return ret;
1143}
1144
1145static int resize_bitmaps(struct mddev *mddev, sector_t newsize, sector_t oldsize)
1146{
1147 void *bitmap = mddev->bitmap;
1148 struct md_bitmap_stats stats;
1149 unsigned long my_pages;
1150 int i, rv;
1151
1152 rv = mddev->bitmap_ops->get_stats(bitmap, &stats);
1153 if (rv)
1154 return rv;
1155
1156 my_pages = stats.pages;
1157 /*
1158 * We need to ensure all the nodes can grow to a larger
1159 * bitmap size before make the reshaping.
1160 */
1161 rv = update_bitmap_size(mddev, newsize);
1162 if (rv)
1163 return rv;
1164
1165 for (i = 0; i < mddev->bitmap_info.nodes; i++) {
1166 struct dlm_lock_resource *bm_lockres;
1167 char str[64];
1168
1169 if (i == md_cluster_ops->slot_number(mddev))
1170 continue;
1171
1172 bitmap = mddev->bitmap_ops->get_from_slot(mddev, i);
1173 if (IS_ERR(bitmap)) {
1174 pr_err("can't get bitmap from slot %d\n", i);
1175 bitmap = NULL;
1176 goto out;
1177 }
1178
1179 rv = mddev->bitmap_ops->get_stats(bitmap, &stats);
1180 if (rv)
1181 goto out;
1182 /*
1183 * If we can hold the bitmap lock of one node then
1184 * the slot is not occupied, update the pages.
1185 */
1186 snprintf(str, 64, "bitmap%04d", i);
1187 bm_lockres = lockres_init(mddev, str, NULL, 1);
1188 if (!bm_lockres) {
1189 pr_err("Cannot initialize %s lock\n", str);
1190 goto out;
1191 }
1192 bm_lockres->flags |= DLM_LKF_NOQUEUE;
1193 rv = dlm_lock_sync(bm_lockres, DLM_LOCK_PW);
1194 if (!rv)
1195 mddev->bitmap_ops->set_pages(bitmap, my_pages);
1196 lockres_free(bm_lockres);
1197
1198 if (my_pages != stats.pages)
1199 /*
1200 * Let's revert the bitmap size if one node
1201 * can't resize bitmap
1202 */
1203 goto out;
1204 mddev->bitmap_ops->free(bitmap);
1205 }
1206
1207 return 0;
1208out:
1209 mddev->bitmap_ops->free(bitmap);
1210 update_bitmap_size(mddev, oldsize);
1211 return -1;
1212}
1213
1214/*
1215 * return 0 if all the bitmaps have the same sync_size
1216 */
1217static int cluster_check_sync_size(struct mddev *mddev)
1218{
1219 int current_slot = md_cluster_ops->slot_number(mddev);
1220 int node_num = mddev->bitmap_info.nodes;
1221 struct dlm_lock_resource *bm_lockres;
1222 struct md_bitmap_stats stats;
1223 void *bitmap = mddev->bitmap;
1224 unsigned long sync_size = 0;
1225 unsigned long my_sync_size;
1226 char str[64];
1227 int i, rv;
1228
1229 rv = mddev->bitmap_ops->get_stats(bitmap, &stats);
1230 if (rv)
1231 return rv;
1232
1233 my_sync_size = stats.sync_size;
1234
1235 for (i = 0; i < node_num; i++) {
1236 if (i == current_slot)
1237 continue;
1238
1239 bitmap = mddev->bitmap_ops->get_from_slot(mddev, i);
1240 if (IS_ERR(bitmap)) {
1241 pr_err("can't get bitmap from slot %d\n", i);
1242 return -1;
1243 }
1244
1245 /*
1246 * If we can hold the bitmap lock of one node then
1247 * the slot is not occupied, update the sb.
1248 */
1249 snprintf(str, 64, "bitmap%04d", i);
1250 bm_lockres = lockres_init(mddev, str, NULL, 1);
1251 if (!bm_lockres) {
1252 pr_err("md-cluster: Cannot initialize %s\n", str);
1253 mddev->bitmap_ops->free(bitmap);
1254 return -1;
1255 }
1256 bm_lockres->flags |= DLM_LKF_NOQUEUE;
1257 rv = dlm_lock_sync(bm_lockres, DLM_LOCK_PW);
1258 if (!rv)
1259 mddev->bitmap_ops->update_sb(bitmap);
1260 lockres_free(bm_lockres);
1261
1262 rv = mddev->bitmap_ops->get_stats(bitmap, &stats);
1263 if (rv) {
1264 mddev->bitmap_ops->free(bitmap);
1265 return rv;
1266 }
1267
1268 if (sync_size == 0) {
1269 sync_size = stats.sync_size;
1270 } else if (sync_size != stats.sync_size) {
1271 mddev->bitmap_ops->free(bitmap);
1272 return -1;
1273 }
1274 mddev->bitmap_ops->free(bitmap);
1275 }
1276
1277 return (my_sync_size == sync_size) ? 0 : -1;
1278}
1279
1280/*
1281 * Update the size for cluster raid is a little more complex, we perform it
1282 * by the steps:
1283 * 1. hold token lock and update superblock in initiator node.
1284 * 2. send METADATA_UPDATED msg to other nodes.
1285 * 3. The initiator node continues to check each bitmap's sync_size, if all
1286 * bitmaps have the same value of sync_size, then we can set capacity and
1287 * let other nodes to perform it. If one node can't update sync_size
1288 * accordingly, we need to revert to previous value.
1289 */
1290static void update_size(struct mddev *mddev, sector_t old_dev_sectors)
1291{
1292 struct md_cluster_info *cinfo = mddev->cluster_info;
1293 struct cluster_msg cmsg;
1294 struct md_rdev *rdev;
1295 int ret = 0;
1296 int raid_slot = -1;
1297
1298 md_update_sb(mddev, 1);
1299 if (lock_comm(cinfo, 1)) {
1300 pr_err("%s: lock_comm failed\n", __func__);
1301 return;
1302 }
1303
1304 memset(&cmsg, 0, sizeof(cmsg));
1305 cmsg.type = cpu_to_le32(METADATA_UPDATED);
1306 rdev_for_each(rdev, mddev)
1307 if (rdev->raid_disk >= 0 && !test_bit(Faulty, &rdev->flags)) {
1308 raid_slot = rdev->desc_nr;
1309 break;
1310 }
1311 if (raid_slot >= 0) {
1312 cmsg.raid_slot = cpu_to_le32(raid_slot);
1313 /*
1314 * We can only change capiticy after all the nodes can do it,
1315 * so need to wait after other nodes already received the msg
1316 * and handled the change
1317 */
1318 ret = __sendmsg(cinfo, &cmsg);
1319 if (ret) {
1320 pr_err("%s:%d: failed to send METADATA_UPDATED msg\n",
1321 __func__, __LINE__);
1322 unlock_comm(cinfo);
1323 return;
1324 }
1325 } else {
1326 pr_err("md-cluster: No good device id found to send\n");
1327 unlock_comm(cinfo);
1328 return;
1329 }
1330
1331 /*
1332 * check the sync_size from other node's bitmap, if sync_size
1333 * have already updated in other nodes as expected, send an
1334 * empty metadata msg to permit the change of capacity
1335 */
1336 if (cluster_check_sync_size(mddev) == 0) {
1337 memset(&cmsg, 0, sizeof(cmsg));
1338 cmsg.type = cpu_to_le32(CHANGE_CAPACITY);
1339 ret = __sendmsg(cinfo, &cmsg);
1340 if (ret)
1341 pr_err("%s:%d: failed to send CHANGE_CAPACITY msg\n",
1342 __func__, __LINE__);
1343 set_capacity_and_notify(mddev->gendisk, mddev->array_sectors);
1344 } else {
1345 /* revert to previous sectors */
1346 ret = mddev->pers->resize(mddev, old_dev_sectors);
1347 ret = __sendmsg(cinfo, &cmsg);
1348 if (ret)
1349 pr_err("%s:%d: failed to send METADATA_UPDATED msg\n",
1350 __func__, __LINE__);
1351 }
1352 unlock_comm(cinfo);
1353}
1354
1355static int resync_start(struct mddev *mddev)
1356{
1357 struct md_cluster_info *cinfo = mddev->cluster_info;
1358 return dlm_lock_sync_interruptible(cinfo->resync_lockres, DLM_LOCK_EX, mddev);
1359}
1360
1361static void resync_info_get(struct mddev *mddev, sector_t *lo, sector_t *hi)
1362{
1363 struct md_cluster_info *cinfo = mddev->cluster_info;
1364
1365 spin_lock_irq(&cinfo->suspend_lock);
1366 *lo = cinfo->suspend_lo;
1367 *hi = cinfo->suspend_hi;
1368 spin_unlock_irq(&cinfo->suspend_lock);
1369}
1370
1371static int resync_status_get(struct mddev *mddev)
1372{
1373 struct md_cluster_info *cinfo = mddev->cluster_info;
1374
1375 return test_bit(MD_CLUSTER_WAITING_FOR_SYNC, &cinfo->state);
1376}
1377
1378static int resync_start_notify(struct mddev *mddev)
1379{
1380 struct md_cluster_info *cinfo = mddev->cluster_info;
1381 struct cluster_msg cmsg = {0};
1382
1383 cmsg.type = cpu_to_le32(RESYNCING_START);
1384
1385 return sendmsg(cinfo, &cmsg, 0);
1386}
1387
1388static int resync_info_update(struct mddev *mddev, sector_t lo, sector_t hi)
1389{
1390 struct md_cluster_info *cinfo = mddev->cluster_info;
1391 struct resync_info ri;
1392 struct cluster_msg cmsg = {0};
1393
1394 /* do not send zero again, if we have sent before */
1395 if (hi == 0) {
1396 memcpy(&ri, cinfo->bitmap_lockres->lksb.sb_lvbptr, sizeof(struct resync_info));
1397 if (le64_to_cpu(ri.hi) == 0)
1398 return 0;
1399 }
1400
1401 add_resync_info(cinfo->bitmap_lockres, lo, hi);
1402 /* Re-acquire the lock to refresh LVB */
1403 dlm_lock_sync(cinfo->bitmap_lockres, DLM_LOCK_PW);
1404 cmsg.type = cpu_to_le32(RESYNCING);
1405 cmsg.low = cpu_to_le64(lo);
1406 cmsg.high = cpu_to_le64(hi);
1407
1408 /*
1409 * mddev_lock is held if resync_info_update is called from
1410 * resync_finish (md_reap_sync_thread -> resync_finish)
1411 */
1412 if (lo == 0 && hi == 0)
1413 return sendmsg(cinfo, &cmsg, 1);
1414 else
1415 return sendmsg(cinfo, &cmsg, 0);
1416}
1417
1418static int resync_finish(struct mddev *mddev)
1419{
1420 struct md_cluster_info *cinfo = mddev->cluster_info;
1421 int ret = 0;
1422
1423 clear_bit(MD_RESYNCING_REMOTE, &mddev->recovery);
1424
1425 /*
1426 * If resync thread is interrupted so we can't say resync is finished,
1427 * another node will launch resync thread to continue.
1428 */
1429 if (!test_bit(MD_CLOSING, &mddev->flags))
1430 ret = resync_info_update(mddev, 0, 0);
1431 dlm_unlock_sync(cinfo->resync_lockres);
1432 return ret;
1433}
1434
1435static int area_resyncing(struct mddev *mddev, int direction,
1436 sector_t lo, sector_t hi)
1437{
1438 struct md_cluster_info *cinfo = mddev->cluster_info;
1439 int ret = 0;
1440
1441 if ((direction == READ) &&
1442 test_bit(MD_CLUSTER_SUSPEND_READ_BALANCING, &cinfo->state))
1443 return 1;
1444
1445 spin_lock_irq(&cinfo->suspend_lock);
1446 if (hi > cinfo->suspend_lo && lo < cinfo->suspend_hi)
1447 ret = 1;
1448 spin_unlock_irq(&cinfo->suspend_lock);
1449 return ret;
1450}
1451
1452/* add_new_disk() - initiates a disk add
1453 * However, if this fails before writing md_update_sb(),
1454 * add_new_disk_cancel() must be called to release token lock
1455 */
1456static int add_new_disk(struct mddev *mddev, struct md_rdev *rdev)
1457{
1458 struct md_cluster_info *cinfo = mddev->cluster_info;
1459 struct cluster_msg cmsg;
1460 int ret = 0;
1461 struct mdp_superblock_1 *sb = page_address(rdev->sb_page);
1462 char *uuid = sb->device_uuid;
1463
1464 memset(&cmsg, 0, sizeof(cmsg));
1465 cmsg.type = cpu_to_le32(NEWDISK);
1466 memcpy(cmsg.uuid, uuid, 16);
1467 cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
1468 if (lock_comm(cinfo, 1))
1469 return -EAGAIN;
1470 ret = __sendmsg(cinfo, &cmsg);
1471 if (ret) {
1472 unlock_comm(cinfo);
1473 return ret;
1474 }
1475 cinfo->no_new_dev_lockres->flags |= DLM_LKF_NOQUEUE;
1476 ret = dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_EX);
1477 cinfo->no_new_dev_lockres->flags &= ~DLM_LKF_NOQUEUE;
1478 /* Some node does not "see" the device */
1479 if (ret == -EAGAIN)
1480 ret = -ENOENT;
1481 if (ret)
1482 unlock_comm(cinfo);
1483 else {
1484 dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR);
1485 /* Since MD_CHANGE_DEVS will be set in add_bound_rdev which
1486 * will run soon after add_new_disk, the below path will be
1487 * invoked:
1488 * md_wakeup_thread(mddev->thread)
1489 * -> conf->thread (raid1d)
1490 * -> md_check_recovery -> md_update_sb
1491 * -> metadata_update_start/finish
1492 * MD_CLUSTER_SEND_LOCKED_ALREADY will be cleared eventually.
1493 *
1494 * For other failure cases, metadata_update_cancel and
1495 * add_new_disk_cancel also clear below bit as well.
1496 * */
1497 set_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1498 wake_up(&cinfo->wait);
1499 }
1500 return ret;
1501}
1502
1503static void add_new_disk_cancel(struct mddev *mddev)
1504{
1505 struct md_cluster_info *cinfo = mddev->cluster_info;
1506 clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1507 unlock_comm(cinfo);
1508}
1509
1510static int new_disk_ack(struct mddev *mddev, bool ack)
1511{
1512 struct md_cluster_info *cinfo = mddev->cluster_info;
1513
1514 if (!test_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state)) {
1515 pr_warn("md-cluster(%s): Spurious cluster confirmation\n", mdname(mddev));
1516 return -EINVAL;
1517 }
1518
1519 if (ack)
1520 dlm_unlock_sync(cinfo->no_new_dev_lockres);
1521 complete(&cinfo->newdisk_completion);
1522 return 0;
1523}
1524
1525static int remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1526{
1527 struct cluster_msg cmsg = {0};
1528 struct md_cluster_info *cinfo = mddev->cluster_info;
1529 cmsg.type = cpu_to_le32(REMOVE);
1530 cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
1531 return sendmsg(cinfo, &cmsg, 1);
1532}
1533
1534static int lock_all_bitmaps(struct mddev *mddev)
1535{
1536 int slot, my_slot, ret, held = 1, i = 0;
1537 char str[64];
1538 struct md_cluster_info *cinfo = mddev->cluster_info;
1539
1540 cinfo->other_bitmap_lockres =
1541 kcalloc(mddev->bitmap_info.nodes - 1,
1542 sizeof(struct dlm_lock_resource *), GFP_KERNEL);
1543 if (!cinfo->other_bitmap_lockres) {
1544 pr_err("md: can't alloc mem for other bitmap locks\n");
1545 return 0;
1546 }
1547
1548 my_slot = slot_number(mddev);
1549 for (slot = 0; slot < mddev->bitmap_info.nodes; slot++) {
1550 if (slot == my_slot)
1551 continue;
1552
1553 memset(str, '\0', 64);
1554 snprintf(str, 64, "bitmap%04d", slot);
1555 cinfo->other_bitmap_lockres[i] = lockres_init(mddev, str, NULL, 1);
1556 if (!cinfo->other_bitmap_lockres[i])
1557 return -ENOMEM;
1558
1559 cinfo->other_bitmap_lockres[i]->flags |= DLM_LKF_NOQUEUE;
1560 ret = dlm_lock_sync(cinfo->other_bitmap_lockres[i], DLM_LOCK_PW);
1561 if (ret)
1562 held = -1;
1563 i++;
1564 }
1565
1566 return held;
1567}
1568
1569static void unlock_all_bitmaps(struct mddev *mddev)
1570{
1571 struct md_cluster_info *cinfo = mddev->cluster_info;
1572 int i;
1573
1574 /* release other node's bitmap lock if they are existed */
1575 if (cinfo->other_bitmap_lockres) {
1576 for (i = 0; i < mddev->bitmap_info.nodes - 1; i++) {
1577 if (cinfo->other_bitmap_lockres[i]) {
1578 lockres_free(cinfo->other_bitmap_lockres[i]);
1579 }
1580 }
1581 kfree(cinfo->other_bitmap_lockres);
1582 cinfo->other_bitmap_lockres = NULL;
1583 }
1584}
1585
1586static int gather_bitmaps(struct md_rdev *rdev)
1587{
1588 int sn, err;
1589 sector_t lo, hi;
1590 struct cluster_msg cmsg = {0};
1591 struct mddev *mddev = rdev->mddev;
1592 struct md_cluster_info *cinfo = mddev->cluster_info;
1593
1594 cmsg.type = cpu_to_le32(RE_ADD);
1595 cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
1596 err = sendmsg(cinfo, &cmsg, 1);
1597 if (err)
1598 goto out;
1599
1600 for (sn = 0; sn < mddev->bitmap_info.nodes; sn++) {
1601 if (sn == (cinfo->slot_number - 1))
1602 continue;
1603 err = mddev->bitmap_ops->copy_from_slot(mddev, sn, &lo, &hi, false);
1604 if (err) {
1605 pr_warn("md-cluster: Could not gather bitmaps from slot %d", sn);
1606 goto out;
1607 }
1608 if ((hi > 0) && (lo < mddev->recovery_cp))
1609 mddev->recovery_cp = lo;
1610 }
1611out:
1612 return err;
1613}
1614
1615static const struct md_cluster_operations cluster_ops = {
1616 .join = join,
1617 .leave = leave,
1618 .slot_number = slot_number,
1619 .resync_start = resync_start,
1620 .resync_finish = resync_finish,
1621 .resync_info_update = resync_info_update,
1622 .resync_start_notify = resync_start_notify,
1623 .resync_status_get = resync_status_get,
1624 .resync_info_get = resync_info_get,
1625 .metadata_update_start = metadata_update_start,
1626 .metadata_update_finish = metadata_update_finish,
1627 .metadata_update_cancel = metadata_update_cancel,
1628 .area_resyncing = area_resyncing,
1629 .add_new_disk = add_new_disk,
1630 .add_new_disk_cancel = add_new_disk_cancel,
1631 .new_disk_ack = new_disk_ack,
1632 .remove_disk = remove_disk,
1633 .load_bitmaps = load_bitmaps,
1634 .gather_bitmaps = gather_bitmaps,
1635 .resize_bitmaps = resize_bitmaps,
1636 .lock_all_bitmaps = lock_all_bitmaps,
1637 .unlock_all_bitmaps = unlock_all_bitmaps,
1638 .update_size = update_size,
1639};
1640
1641static int __init cluster_init(void)
1642{
1643 pr_warn("md-cluster: support raid1 and raid10 (limited support)\n");
1644 pr_info("Registering Cluster MD functions\n");
1645 register_md_cluster_operations(&cluster_ops, THIS_MODULE);
1646 return 0;
1647}
1648
1649static void cluster_exit(void)
1650{
1651 unregister_md_cluster_operations();
1652}
1653
1654module_init(cluster_init);
1655module_exit(cluster_exit);
1656MODULE_AUTHOR("SUSE");
1657MODULE_LICENSE("GPL");
1658MODULE_DESCRIPTION("Clustering support for MD");
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 "bitmap.h"
19#include "md-cluster.h"
20
21#define LVB_SIZE 64
22#define NEW_DEV_TIMEOUT 5000
23
24struct 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
36struct suspend_info {
37 int slot;
38 sector_t lo;
39 sector_t hi;
40 struct list_head list;
41};
42
43struct 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
71
72struct md_cluster_info {
73 /* dlm lock space and resources for clustered raid. */
74 dlm_lockspace_t *lockspace;
75 int slot_number;
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;
93 unsigned long state;
94 /* record the region in RESYNCING message */
95 sector_t sync_low;
96 sector_t sync_hi;
97};
98
99enum msg_type {
100 METADATA_UPDATED = 0,
101 RESYNCING,
102 NEWDISK,
103 REMOVE,
104 RE_ADD,
105 BITMAP_NEEDS_SYNC,
106};
107
108struct cluster_msg {
109 __le32 type;
110 __le32 slot;
111 /* TODO: Unionize this for smaller footprint */
112 __le64 low;
113 __le64 high;
114 char uuid[16];
115 __le32 raid_slot;
116};
117
118static void sync_ast(void *arg)
119{
120 struct dlm_lock_resource *res;
121
122 res = arg;
123 res->sync_locking_done = true;
124 wake_up(&res->sync_locking);
125}
126
127static int dlm_lock_sync(struct dlm_lock_resource *res, int mode)
128{
129 int ret = 0;
130
131 ret = dlm_lock(res->ls, mode, &res->lksb,
132 res->flags, res->name, strlen(res->name),
133 0, sync_ast, res, res->bast);
134 if (ret)
135 return ret;
136 wait_event(res->sync_locking, res->sync_locking_done);
137 res->sync_locking_done = false;
138 if (res->lksb.sb_status == 0)
139 res->mode = mode;
140 return res->lksb.sb_status;
141}
142
143static int dlm_unlock_sync(struct dlm_lock_resource *res)
144{
145 return dlm_lock_sync(res, DLM_LOCK_NL);
146}
147
148/*
149 * An variation of dlm_lock_sync, which make lock request could
150 * be interrupted
151 */
152static int dlm_lock_sync_interruptible(struct dlm_lock_resource *res, int mode,
153 struct mddev *mddev)
154{
155 int ret = 0;
156
157 ret = dlm_lock(res->ls, mode, &res->lksb,
158 res->flags, res->name, strlen(res->name),
159 0, sync_ast, res, res->bast);
160 if (ret)
161 return ret;
162
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) {
167 /*
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
171 */
172 ret = dlm_unlock(res->ls, res->lksb.sb_lkid, DLM_LKF_CANCEL,
173 &res->lksb, res);
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);
178 return -EPERM;
179 } else
180 res->sync_locking_done = false;
181 if (res->lksb.sb_status == 0)
182 res->mode = mode;
183 return res->lksb.sb_status;
184}
185
186static struct dlm_lock_resource *lockres_init(struct mddev *mddev,
187 char *name, void (*bastfn)(void *arg, int mode), int with_lvb)
188{
189 struct dlm_lock_resource *res = NULL;
190 int ret, namelen;
191 struct md_cluster_info *cinfo = mddev->cluster_info;
192
193 res = kzalloc(sizeof(struct dlm_lock_resource), GFP_KERNEL);
194 if (!res)
195 return NULL;
196 init_waitqueue_head(&res->sync_locking);
197 res->sync_locking_done = false;
198 res->ls = cinfo->lockspace;
199 res->mddev = mddev;
200 res->mode = DLM_LOCK_IV;
201 namelen = strlen(name);
202 res->name = kzalloc(namelen + 1, GFP_KERNEL);
203 if (!res->name) {
204 pr_err("md-cluster: Unable to allocate resource name for resource %s\n", name);
205 goto out_err;
206 }
207 strlcpy(res->name, name, namelen + 1);
208 if (with_lvb) {
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);
212 goto out_err;
213 }
214 res->flags = DLM_LKF_VALBLK;
215 }
216
217 if (bastfn)
218 res->bast = bastfn;
219
220 res->flags |= DLM_LKF_EXPEDITE;
221
222 ret = dlm_lock_sync(res, DLM_LOCK_NL);
223 if (ret) {
224 pr_err("md-cluster: Unable to lock NL on new lock resource %s\n", name);
225 goto out_err;
226 }
227 res->flags &= ~DLM_LKF_EXPEDITE;
228 res->flags |= DLM_LKF_CONVERT;
229
230 return res;
231out_err:
232 kfree(res->lksb.sb_lvbptr);
233 kfree(res->name);
234 kfree(res);
235 return NULL;
236}
237
238static void lockres_free(struct dlm_lock_resource *res)
239{
240 int ret = 0;
241
242 if (!res)
243 return;
244
245 /*
246 * use FORCEUNLOCK flag, so we can unlock even the lock is on the
247 * waiting or convert queue
248 */
249 ret = dlm_unlock(res->ls, res->lksb.sb_lkid, DLM_LKF_FORCEUNLOCK,
250 &res->lksb, res);
251 if (unlikely(ret != 0))
252 pr_err("failed to unlock %s return %d\n", res->name, ret);
253 else
254 wait_event(res->sync_locking, res->sync_locking_done);
255
256 kfree(res->name);
257 kfree(res->lksb.sb_lvbptr);
258 kfree(res);
259}
260
261static void add_resync_info(struct dlm_lock_resource *lockres,
262 sector_t lo, sector_t hi)
263{
264 struct resync_info *ri;
265
266 ri = (struct resync_info *)lockres->lksb.sb_lvbptr;
267 ri->lo = cpu_to_le64(lo);
268 ri->hi = cpu_to_le64(hi);
269}
270
271static struct suspend_info *read_resync_info(struct mddev *mddev, struct dlm_lock_resource *lockres)
272{
273 struct resync_info ri;
274 struct suspend_info *s = NULL;
275 sector_t hi = 0;
276
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);
280 if (hi > 0) {
281 s = kzalloc(sizeof(struct suspend_info), GFP_KERNEL);
282 if (!s)
283 goto out;
284 s->hi = hi;
285 s->lo = le64_to_cpu(ri.lo);
286 }
287 dlm_unlock_sync(lockres);
288out:
289 return s;
290}
291
292static void recover_bitmaps(struct md_thread *thread)
293{
294 struct mddev *mddev = thread->mddev;
295 struct md_cluster_info *cinfo = mddev->cluster_info;
296 struct dlm_lock_resource *bm_lockres;
297 char str[64];
298 int slot, ret;
299 struct suspend_info *s, *tmp;
300 sector_t lo, hi;
301
302 while (cinfo->recovery_map) {
303 slot = fls64((u64)cinfo->recovery_map) - 1;
304
305 /* Clear suspend_area associated with the bitmap */
306 spin_lock_irq(&cinfo->suspend_lock);
307 list_for_each_entry_safe(s, tmp, &cinfo->suspend_list, list)
308 if (slot == s->slot) {
309 list_del(&s->list);
310 kfree(s);
311 }
312 spin_unlock_irq(&cinfo->suspend_lock);
313
314 snprintf(str, 64, "bitmap%04d", slot);
315 bm_lockres = lockres_init(mddev, str, NULL, 1);
316 if (!bm_lockres) {
317 pr_err("md-cluster: Cannot initialize bitmaps\n");
318 goto clear_bit;
319 }
320
321 ret = dlm_lock_sync_interruptible(bm_lockres, DLM_LOCK_PW, mddev);
322 if (ret) {
323 pr_err("md-cluster: Could not DLM lock %s: %d\n",
324 str, ret);
325 goto clear_bit;
326 }
327 ret = bitmap_copy_from_slot(mddev, slot, &lo, &hi, true);
328 if (ret) {
329 pr_err("md-cluster: Could not copy data from bitmap %d\n", slot);
330 goto clear_bit;
331 }
332 if (hi > 0) {
333 if (lo < mddev->recovery_cp)
334 mddev->recovery_cp = lo;
335 /* wake up thread to continue resync in case resync
336 * is not finished */
337 if (mddev->recovery_cp != MaxSector) {
338 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
339 md_wakeup_thread(mddev->thread);
340 }
341 }
342clear_bit:
343 lockres_free(bm_lockres);
344 clear_bit(slot, &cinfo->recovery_map);
345 }
346}
347
348static void recover_prep(void *arg)
349{
350 struct mddev *mddev = arg;
351 struct md_cluster_info *cinfo = mddev->cluster_info;
352 set_bit(MD_CLUSTER_SUSPEND_READ_BALANCING, &cinfo->state);
353}
354
355static void __recover_slot(struct mddev *mddev, int slot)
356{
357 struct md_cluster_info *cinfo = mddev->cluster_info;
358
359 set_bit(slot, &cinfo->recovery_map);
360 if (!cinfo->recovery_thread) {
361 cinfo->recovery_thread = md_register_thread(recover_bitmaps,
362 mddev, "recover");
363 if (!cinfo->recovery_thread) {
364 pr_warn("md-cluster: Could not create recovery thread\n");
365 return;
366 }
367 }
368 md_wakeup_thread(cinfo->recovery_thread);
369}
370
371static void recover_slot(void *arg, struct dlm_slot *slot)
372{
373 struct mddev *mddev = arg;
374 struct md_cluster_info *cinfo = mddev->cluster_info;
375
376 pr_info("md-cluster: %s Node %d/%d down. My slot: %d. Initiating recovery.\n",
377 mddev->bitmap_info.cluster_name,
378 slot->nodeid, slot->slot,
379 cinfo->slot_number);
380 /* deduct one since dlm slot starts from one while the num of
381 * cluster-md begins with 0 */
382 __recover_slot(mddev, slot->slot - 1);
383}
384
385static void recover_done(void *arg, struct dlm_slot *slots,
386 int num_slots, int our_slot,
387 uint32_t generation)
388{
389 struct mddev *mddev = arg;
390 struct md_cluster_info *cinfo = mddev->cluster_info;
391
392 cinfo->slot_number = our_slot;
393 /* completion is only need to be complete when node join cluster,
394 * it doesn't need to run during another node's failure */
395 if (test_bit(MD_CLUSTER_BEGIN_JOIN_CLUSTER, &cinfo->state)) {
396 complete(&cinfo->completion);
397 clear_bit(MD_CLUSTER_BEGIN_JOIN_CLUSTER, &cinfo->state);
398 }
399 clear_bit(MD_CLUSTER_SUSPEND_READ_BALANCING, &cinfo->state);
400}
401
402/* the ops is called when node join the cluster, and do lock recovery
403 * if node failure occurs */
404static const struct dlm_lockspace_ops md_ls_ops = {
405 .recover_prep = recover_prep,
406 .recover_slot = recover_slot,
407 .recover_done = recover_done,
408};
409
410/*
411 * The BAST function for the ack lock resource
412 * This function wakes up the receive thread in
413 * order to receive and process the message.
414 */
415static void ack_bast(void *arg, int mode)
416{
417 struct dlm_lock_resource *res = arg;
418 struct md_cluster_info *cinfo = res->mddev->cluster_info;
419
420 if (mode == DLM_LOCK_EX) {
421 if (test_bit(MD_CLUSTER_ALREADY_IN_CLUSTER, &cinfo->state))
422 md_wakeup_thread(cinfo->recv_thread);
423 else
424 set_bit(MD_CLUSTER_PENDING_RECV_EVENT, &cinfo->state);
425 }
426}
427
428static void __remove_suspend_info(struct md_cluster_info *cinfo, int slot)
429{
430 struct suspend_info *s, *tmp;
431
432 list_for_each_entry_safe(s, tmp, &cinfo->suspend_list, list)
433 if (slot == s->slot) {
434 list_del(&s->list);
435 kfree(s);
436 break;
437 }
438}
439
440static void remove_suspend_info(struct mddev *mddev, int slot)
441{
442 struct md_cluster_info *cinfo = mddev->cluster_info;
443 spin_lock_irq(&cinfo->suspend_lock);
444 __remove_suspend_info(cinfo, slot);
445 spin_unlock_irq(&cinfo->suspend_lock);
446 mddev->pers->quiesce(mddev, 2);
447}
448
449
450static void process_suspend_info(struct mddev *mddev,
451 int slot, sector_t lo, sector_t hi)
452{
453 struct md_cluster_info *cinfo = mddev->cluster_info;
454 struct suspend_info *s;
455
456 if (!hi) {
457 remove_suspend_info(mddev, slot);
458 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
459 md_wakeup_thread(mddev->thread);
460 return;
461 }
462
463 /*
464 * The bitmaps are not same for different nodes
465 * if RESYNCING is happening in one node, then
466 * the node which received the RESYNCING message
467 * probably will perform resync with the region
468 * [lo, hi] again, so we could reduce resync time
469 * a lot if we can ensure that the bitmaps among
470 * different nodes are match up well.
471 *
472 * sync_low/hi is used to record the region which
473 * arrived in the previous RESYNCING message,
474 *
475 * Call bitmap_sync_with_cluster to clear
476 * NEEDED_MASK and set RESYNC_MASK since
477 * resync thread is running in another node,
478 * so we don't need to do the resync again
479 * with the same section */
480 bitmap_sync_with_cluster(mddev, cinfo->sync_low,
481 cinfo->sync_hi,
482 lo, hi);
483 cinfo->sync_low = lo;
484 cinfo->sync_hi = hi;
485
486 s = kzalloc(sizeof(struct suspend_info), GFP_KERNEL);
487 if (!s)
488 return;
489 s->slot = slot;
490 s->lo = lo;
491 s->hi = hi;
492 mddev->pers->quiesce(mddev, 1);
493 mddev->pers->quiesce(mddev, 0);
494 spin_lock_irq(&cinfo->suspend_lock);
495 /* Remove existing entry (if exists) before adding */
496 __remove_suspend_info(cinfo, slot);
497 list_add(&s->list, &cinfo->suspend_list);
498 spin_unlock_irq(&cinfo->suspend_lock);
499 mddev->pers->quiesce(mddev, 2);
500}
501
502static void process_add_new_disk(struct mddev *mddev, struct cluster_msg *cmsg)
503{
504 char disk_uuid[64];
505 struct md_cluster_info *cinfo = mddev->cluster_info;
506 char event_name[] = "EVENT=ADD_DEVICE";
507 char raid_slot[16];
508 char *envp[] = {event_name, disk_uuid, raid_slot, NULL};
509 int len;
510
511 len = snprintf(disk_uuid, 64, "DEVICE_UUID=");
512 sprintf(disk_uuid + len, "%pU", cmsg->uuid);
513 snprintf(raid_slot, 16, "RAID_DISK=%d", le32_to_cpu(cmsg->raid_slot));
514 pr_info("%s:%d Sending kobject change with %s and %s\n", __func__, __LINE__, disk_uuid, raid_slot);
515 init_completion(&cinfo->newdisk_completion);
516 set_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state);
517 kobject_uevent_env(&disk_to_dev(mddev->gendisk)->kobj, KOBJ_CHANGE, envp);
518 wait_for_completion_timeout(&cinfo->newdisk_completion,
519 NEW_DEV_TIMEOUT);
520 clear_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state);
521}
522
523
524static void process_metadata_update(struct mddev *mddev, struct cluster_msg *msg)
525{
526 struct md_cluster_info *cinfo = mddev->cluster_info;
527 mddev->good_device_nr = le32_to_cpu(msg->raid_slot);
528 set_bit(MD_RELOAD_SB, &mddev->flags);
529 dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR);
530 md_wakeup_thread(mddev->thread);
531}
532
533static void process_remove_disk(struct mddev *mddev, struct cluster_msg *msg)
534{
535 struct md_rdev *rdev;
536
537 rcu_read_lock();
538 rdev = md_find_rdev_nr_rcu(mddev, le32_to_cpu(msg->raid_slot));
539 if (rdev) {
540 set_bit(ClusterRemove, &rdev->flags);
541 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
542 md_wakeup_thread(mddev->thread);
543 }
544 else
545 pr_warn("%s: %d Could not find disk(%d) to REMOVE\n",
546 __func__, __LINE__, le32_to_cpu(msg->raid_slot));
547 rcu_read_unlock();
548}
549
550static void process_readd_disk(struct mddev *mddev, struct cluster_msg *msg)
551{
552 struct md_rdev *rdev;
553
554 rcu_read_lock();
555 rdev = md_find_rdev_nr_rcu(mddev, le32_to_cpu(msg->raid_slot));
556 if (rdev && test_bit(Faulty, &rdev->flags))
557 clear_bit(Faulty, &rdev->flags);
558 else
559 pr_warn("%s: %d Could not find disk(%d) which is faulty",
560 __func__, __LINE__, le32_to_cpu(msg->raid_slot));
561 rcu_read_unlock();
562}
563
564static int process_recvd_msg(struct mddev *mddev, struct cluster_msg *msg)
565{
566 int ret = 0;
567
568 if (WARN(mddev->cluster_info->slot_number - 1 == le32_to_cpu(msg->slot),
569 "node %d received it's own msg\n", le32_to_cpu(msg->slot)))
570 return -1;
571 switch (le32_to_cpu(msg->type)) {
572 case METADATA_UPDATED:
573 process_metadata_update(mddev, msg);
574 break;
575 case RESYNCING:
576 process_suspend_info(mddev, le32_to_cpu(msg->slot),
577 le64_to_cpu(msg->low),
578 le64_to_cpu(msg->high));
579 break;
580 case NEWDISK:
581 process_add_new_disk(mddev, msg);
582 break;
583 case REMOVE:
584 process_remove_disk(mddev, msg);
585 break;
586 case RE_ADD:
587 process_readd_disk(mddev, msg);
588 break;
589 case BITMAP_NEEDS_SYNC:
590 __recover_slot(mddev, le32_to_cpu(msg->slot));
591 break;
592 default:
593 ret = -1;
594 pr_warn("%s:%d Received unknown message from %d\n",
595 __func__, __LINE__, msg->slot);
596 }
597 return ret;
598}
599
600/*
601 * thread for receiving message
602 */
603static void recv_daemon(struct md_thread *thread)
604{
605 struct md_cluster_info *cinfo = thread->mddev->cluster_info;
606 struct dlm_lock_resource *ack_lockres = cinfo->ack_lockres;
607 struct dlm_lock_resource *message_lockres = cinfo->message_lockres;
608 struct cluster_msg msg;
609 int ret;
610
611 mutex_lock(&cinfo->recv_mutex);
612 /*get CR on Message*/
613 if (dlm_lock_sync(message_lockres, DLM_LOCK_CR)) {
614 pr_err("md/raid1:failed to get CR on MESSAGE\n");
615 mutex_unlock(&cinfo->recv_mutex);
616 return;
617 }
618
619 /* read lvb and wake up thread to process this message_lockres */
620 memcpy(&msg, message_lockres->lksb.sb_lvbptr, sizeof(struct cluster_msg));
621 ret = process_recvd_msg(thread->mddev, &msg);
622 if (ret)
623 goto out;
624
625 /*release CR on ack_lockres*/
626 ret = dlm_unlock_sync(ack_lockres);
627 if (unlikely(ret != 0))
628 pr_info("unlock ack failed return %d\n", ret);
629 /*up-convert to PR on message_lockres*/
630 ret = dlm_lock_sync(message_lockres, DLM_LOCK_PR);
631 if (unlikely(ret != 0))
632 pr_info("lock PR on msg failed return %d\n", ret);
633 /*get CR on ack_lockres again*/
634 ret = dlm_lock_sync(ack_lockres, DLM_LOCK_CR);
635 if (unlikely(ret != 0))
636 pr_info("lock CR on ack failed return %d\n", ret);
637out:
638 /*release CR on message_lockres*/
639 ret = dlm_unlock_sync(message_lockres);
640 if (unlikely(ret != 0))
641 pr_info("unlock msg failed return %d\n", ret);
642 mutex_unlock(&cinfo->recv_mutex);
643}
644
645/* lock_token()
646 * Takes the lock on the TOKEN lock resource so no other
647 * node can communicate while the operation is underway.
648 */
649static int lock_token(struct md_cluster_info *cinfo)
650{
651 int error;
652
653 error = dlm_lock_sync(cinfo->token_lockres, DLM_LOCK_EX);
654 if (error)
655 pr_err("md-cluster(%s:%d): failed to get EX on TOKEN (%d)\n",
656 __func__, __LINE__, error);
657
658 /* Lock the receive sequence */
659 mutex_lock(&cinfo->recv_mutex);
660 return error;
661}
662
663/* lock_comm()
664 * Sets the MD_CLUSTER_SEND_LOCK bit to lock the send channel.
665 */
666static int lock_comm(struct md_cluster_info *cinfo)
667{
668 wait_event(cinfo->wait,
669 !test_and_set_bit(MD_CLUSTER_SEND_LOCK, &cinfo->state));
670
671 return lock_token(cinfo);
672}
673
674static void unlock_comm(struct md_cluster_info *cinfo)
675{
676 WARN_ON(cinfo->token_lockres->mode != DLM_LOCK_EX);
677 mutex_unlock(&cinfo->recv_mutex);
678 dlm_unlock_sync(cinfo->token_lockres);
679 clear_bit(MD_CLUSTER_SEND_LOCK, &cinfo->state);
680 wake_up(&cinfo->wait);
681}
682
683/* __sendmsg()
684 * This function performs the actual sending of the message. This function is
685 * usually called after performing the encompassing operation
686 * The function:
687 * 1. Grabs the message lockresource in EX mode
688 * 2. Copies the message to the message LVB
689 * 3. Downconverts message lockresource to CW
690 * 4. Upconverts ack lock resource from CR to EX. This forces the BAST on other nodes
691 * and the other nodes read the message. The thread will wait here until all other
692 * nodes have released ack lock resource.
693 * 5. Downconvert ack lockresource to CR
694 */
695static int __sendmsg(struct md_cluster_info *cinfo, struct cluster_msg *cmsg)
696{
697 int error;
698 int slot = cinfo->slot_number - 1;
699
700 cmsg->slot = cpu_to_le32(slot);
701 /*get EX on Message*/
702 error = dlm_lock_sync(cinfo->message_lockres, DLM_LOCK_EX);
703 if (error) {
704 pr_err("md-cluster: failed to get EX on MESSAGE (%d)\n", error);
705 goto failed_message;
706 }
707
708 memcpy(cinfo->message_lockres->lksb.sb_lvbptr, (void *)cmsg,
709 sizeof(struct cluster_msg));
710 /*down-convert EX to CW on Message*/
711 error = dlm_lock_sync(cinfo->message_lockres, DLM_LOCK_CW);
712 if (error) {
713 pr_err("md-cluster: failed to convert EX to CW on MESSAGE(%d)\n",
714 error);
715 goto failed_ack;
716 }
717
718 /*up-convert CR to EX on Ack*/
719 error = dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_EX);
720 if (error) {
721 pr_err("md-cluster: failed to convert CR to EX on ACK(%d)\n",
722 error);
723 goto failed_ack;
724 }
725
726 /*down-convert EX to CR on Ack*/
727 error = dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_CR);
728 if (error) {
729 pr_err("md-cluster: failed to convert EX to CR on ACK(%d)\n",
730 error);
731 goto failed_ack;
732 }
733
734failed_ack:
735 error = dlm_unlock_sync(cinfo->message_lockres);
736 if (unlikely(error != 0)) {
737 pr_err("md-cluster: failed convert to NL on MESSAGE(%d)\n",
738 error);
739 /* in case the message can't be released due to some reason */
740 goto failed_ack;
741 }
742failed_message:
743 return error;
744}
745
746static int sendmsg(struct md_cluster_info *cinfo, struct cluster_msg *cmsg)
747{
748 int ret;
749
750 lock_comm(cinfo);
751 ret = __sendmsg(cinfo, cmsg);
752 unlock_comm(cinfo);
753 return ret;
754}
755
756static int gather_all_resync_info(struct mddev *mddev, int total_slots)
757{
758 struct md_cluster_info *cinfo = mddev->cluster_info;
759 int i, ret = 0;
760 struct dlm_lock_resource *bm_lockres;
761 struct suspend_info *s;
762 char str[64];
763 sector_t lo, hi;
764
765
766 for (i = 0; i < total_slots; i++) {
767 memset(str, '\0', 64);
768 snprintf(str, 64, "bitmap%04d", i);
769 bm_lockres = lockres_init(mddev, str, NULL, 1);
770 if (!bm_lockres)
771 return -ENOMEM;
772 if (i == (cinfo->slot_number - 1)) {
773 lockres_free(bm_lockres);
774 continue;
775 }
776
777 bm_lockres->flags |= DLM_LKF_NOQUEUE;
778 ret = dlm_lock_sync(bm_lockres, DLM_LOCK_PW);
779 if (ret == -EAGAIN) {
780 memset(bm_lockres->lksb.sb_lvbptr, '\0', LVB_SIZE);
781 s = read_resync_info(mddev, bm_lockres);
782 if (s) {
783 pr_info("%s:%d Resync[%llu..%llu] in progress on %d\n",
784 __func__, __LINE__,
785 (unsigned long long) s->lo,
786 (unsigned long long) s->hi, i);
787 spin_lock_irq(&cinfo->suspend_lock);
788 s->slot = i;
789 list_add(&s->list, &cinfo->suspend_list);
790 spin_unlock_irq(&cinfo->suspend_lock);
791 }
792 ret = 0;
793 lockres_free(bm_lockres);
794 continue;
795 }
796 if (ret) {
797 lockres_free(bm_lockres);
798 goto out;
799 }
800
801 /* Read the disk bitmap sb and check if it needs recovery */
802 ret = bitmap_copy_from_slot(mddev, i, &lo, &hi, false);
803 if (ret) {
804 pr_warn("md-cluster: Could not gather bitmaps from slot %d", i);
805 lockres_free(bm_lockres);
806 continue;
807 }
808 if ((hi > 0) && (lo < mddev->recovery_cp)) {
809 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
810 mddev->recovery_cp = lo;
811 md_check_recovery(mddev);
812 }
813
814 lockres_free(bm_lockres);
815 }
816out:
817 return ret;
818}
819
820static int join(struct mddev *mddev, int nodes)
821{
822 struct md_cluster_info *cinfo;
823 int ret, ops_rv;
824 char str[64];
825
826 cinfo = kzalloc(sizeof(struct md_cluster_info), GFP_KERNEL);
827 if (!cinfo)
828 return -ENOMEM;
829
830 INIT_LIST_HEAD(&cinfo->suspend_list);
831 spin_lock_init(&cinfo->suspend_lock);
832 init_completion(&cinfo->completion);
833 set_bit(MD_CLUSTER_BEGIN_JOIN_CLUSTER, &cinfo->state);
834 init_waitqueue_head(&cinfo->wait);
835 mutex_init(&cinfo->recv_mutex);
836
837 mddev->cluster_info = cinfo;
838
839 memset(str, 0, 64);
840 sprintf(str, "%pU", mddev->uuid);
841 ret = dlm_new_lockspace(str, mddev->bitmap_info.cluster_name,
842 DLM_LSFL_FS, LVB_SIZE,
843 &md_ls_ops, mddev, &ops_rv, &cinfo->lockspace);
844 if (ret)
845 goto err;
846 wait_for_completion(&cinfo->completion);
847 if (nodes < cinfo->slot_number) {
848 pr_err("md-cluster: Slot allotted(%d) is greater than available slots(%d).",
849 cinfo->slot_number, nodes);
850 ret = -ERANGE;
851 goto err;
852 }
853 /* Initiate the communication resources */
854 ret = -ENOMEM;
855 cinfo->recv_thread = md_register_thread(recv_daemon, mddev, "cluster_recv");
856 if (!cinfo->recv_thread) {
857 pr_err("md-cluster: cannot allocate memory for recv_thread!\n");
858 goto err;
859 }
860 cinfo->message_lockres = lockres_init(mddev, "message", NULL, 1);
861 if (!cinfo->message_lockres)
862 goto err;
863 cinfo->token_lockres = lockres_init(mddev, "token", NULL, 0);
864 if (!cinfo->token_lockres)
865 goto err;
866 cinfo->no_new_dev_lockres = lockres_init(mddev, "no-new-dev", NULL, 0);
867 if (!cinfo->no_new_dev_lockres)
868 goto err;
869
870 ret = dlm_lock_sync(cinfo->token_lockres, DLM_LOCK_EX);
871 if (ret) {
872 ret = -EAGAIN;
873 pr_err("md-cluster: can't join cluster to avoid lock issue\n");
874 goto err;
875 }
876 cinfo->ack_lockres = lockres_init(mddev, "ack", ack_bast, 0);
877 if (!cinfo->ack_lockres) {
878 ret = -ENOMEM;
879 goto err;
880 }
881 /* get sync CR lock on ACK. */
882 if (dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_CR))
883 pr_err("md-cluster: failed to get a sync CR lock on ACK!(%d)\n",
884 ret);
885 dlm_unlock_sync(cinfo->token_lockres);
886 /* get sync CR lock on no-new-dev. */
887 if (dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR))
888 pr_err("md-cluster: failed to get a sync CR lock on no-new-dev!(%d)\n", ret);
889
890
891 pr_info("md-cluster: Joined cluster %s slot %d\n", str, cinfo->slot_number);
892 snprintf(str, 64, "bitmap%04d", cinfo->slot_number - 1);
893 cinfo->bitmap_lockres = lockres_init(mddev, str, NULL, 1);
894 if (!cinfo->bitmap_lockres) {
895 ret = -ENOMEM;
896 goto err;
897 }
898 if (dlm_lock_sync(cinfo->bitmap_lockres, DLM_LOCK_PW)) {
899 pr_err("Failed to get bitmap lock\n");
900 ret = -EINVAL;
901 goto err;
902 }
903
904 cinfo->resync_lockres = lockres_init(mddev, "resync", NULL, 0);
905 if (!cinfo->resync_lockres) {
906 ret = -ENOMEM;
907 goto err;
908 }
909
910 return 0;
911err:
912 md_unregister_thread(&cinfo->recovery_thread);
913 md_unregister_thread(&cinfo->recv_thread);
914 lockres_free(cinfo->message_lockres);
915 lockres_free(cinfo->token_lockres);
916 lockres_free(cinfo->ack_lockres);
917 lockres_free(cinfo->no_new_dev_lockres);
918 lockres_free(cinfo->resync_lockres);
919 lockres_free(cinfo->bitmap_lockres);
920 if (cinfo->lockspace)
921 dlm_release_lockspace(cinfo->lockspace, 2);
922 mddev->cluster_info = NULL;
923 kfree(cinfo);
924 return ret;
925}
926
927static void load_bitmaps(struct mddev *mddev, int total_slots)
928{
929 struct md_cluster_info *cinfo = mddev->cluster_info;
930
931 /* load all the node's bitmap info for resync */
932 if (gather_all_resync_info(mddev, total_slots))
933 pr_err("md-cluster: failed to gather all resyn infos\n");
934 set_bit(MD_CLUSTER_ALREADY_IN_CLUSTER, &cinfo->state);
935 /* wake up recv thread in case something need to be handled */
936 if (test_and_clear_bit(MD_CLUSTER_PENDING_RECV_EVENT, &cinfo->state))
937 md_wakeup_thread(cinfo->recv_thread);
938}
939
940static void resync_bitmap(struct mddev *mddev)
941{
942 struct md_cluster_info *cinfo = mddev->cluster_info;
943 struct cluster_msg cmsg = {0};
944 int err;
945
946 cmsg.type = cpu_to_le32(BITMAP_NEEDS_SYNC);
947 err = sendmsg(cinfo, &cmsg);
948 if (err)
949 pr_err("%s:%d: failed to send BITMAP_NEEDS_SYNC message (%d)\n",
950 __func__, __LINE__, err);
951}
952
953static void unlock_all_bitmaps(struct mddev *mddev);
954static int leave(struct mddev *mddev)
955{
956 struct md_cluster_info *cinfo = mddev->cluster_info;
957
958 if (!cinfo)
959 return 0;
960
961 /* BITMAP_NEEDS_SYNC message should be sent when node
962 * is leaving the cluster with dirty bitmap, also we
963 * can only deliver it when dlm connection is available */
964 if (cinfo->slot_number > 0 && mddev->recovery_cp != MaxSector)
965 resync_bitmap(mddev);
966
967 md_unregister_thread(&cinfo->recovery_thread);
968 md_unregister_thread(&cinfo->recv_thread);
969 lockres_free(cinfo->message_lockres);
970 lockres_free(cinfo->token_lockres);
971 lockres_free(cinfo->ack_lockres);
972 lockres_free(cinfo->no_new_dev_lockres);
973 lockres_free(cinfo->resync_lockres);
974 lockres_free(cinfo->bitmap_lockres);
975 unlock_all_bitmaps(mddev);
976 dlm_release_lockspace(cinfo->lockspace, 2);
977 return 0;
978}
979
980/* slot_number(): Returns the MD slot number to use
981 * DLM starts the slot numbers from 1, wheras cluster-md
982 * wants the number to be from zero, so we deduct one
983 */
984static int slot_number(struct mddev *mddev)
985{
986 struct md_cluster_info *cinfo = mddev->cluster_info;
987
988 return cinfo->slot_number - 1;
989}
990
991/*
992 * Check if the communication is already locked, else lock the communication
993 * channel.
994 * If it is already locked, token is in EX mode, and hence lock_token()
995 * should not be called.
996 */
997static int metadata_update_start(struct mddev *mddev)
998{
999 struct md_cluster_info *cinfo = mddev->cluster_info;
1000
1001 wait_event(cinfo->wait,
1002 !test_and_set_bit(MD_CLUSTER_SEND_LOCK, &cinfo->state) ||
1003 test_and_clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state));
1004
1005 /* If token is already locked, return 0 */
1006 if (cinfo->token_lockres->mode == DLM_LOCK_EX)
1007 return 0;
1008
1009 return lock_token(cinfo);
1010}
1011
1012static int metadata_update_finish(struct mddev *mddev)
1013{
1014 struct md_cluster_info *cinfo = mddev->cluster_info;
1015 struct cluster_msg cmsg;
1016 struct md_rdev *rdev;
1017 int ret = 0;
1018 int raid_slot = -1;
1019
1020 memset(&cmsg, 0, sizeof(cmsg));
1021 cmsg.type = cpu_to_le32(METADATA_UPDATED);
1022 /* Pick up a good active device number to send.
1023 */
1024 rdev_for_each(rdev, mddev)
1025 if (rdev->raid_disk > -1 && !test_bit(Faulty, &rdev->flags)) {
1026 raid_slot = rdev->desc_nr;
1027 break;
1028 }
1029 if (raid_slot >= 0) {
1030 cmsg.raid_slot = cpu_to_le32(raid_slot);
1031 ret = __sendmsg(cinfo, &cmsg);
1032 } else
1033 pr_warn("md-cluster: No good device id found to send\n");
1034 clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1035 unlock_comm(cinfo);
1036 return ret;
1037}
1038
1039static void metadata_update_cancel(struct mddev *mddev)
1040{
1041 struct md_cluster_info *cinfo = mddev->cluster_info;
1042 clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1043 unlock_comm(cinfo);
1044}
1045
1046static int resync_start(struct mddev *mddev)
1047{
1048 struct md_cluster_info *cinfo = mddev->cluster_info;
1049 return dlm_lock_sync_interruptible(cinfo->resync_lockres, DLM_LOCK_EX, mddev);
1050}
1051
1052static int resync_info_update(struct mddev *mddev, sector_t lo, sector_t hi)
1053{
1054 struct md_cluster_info *cinfo = mddev->cluster_info;
1055 struct resync_info ri;
1056 struct cluster_msg cmsg = {0};
1057
1058 /* do not send zero again, if we have sent before */
1059 if (hi == 0) {
1060 memcpy(&ri, cinfo->bitmap_lockres->lksb.sb_lvbptr, sizeof(struct resync_info));
1061 if (le64_to_cpu(ri.hi) == 0)
1062 return 0;
1063 }
1064
1065 add_resync_info(cinfo->bitmap_lockres, lo, hi);
1066 /* Re-acquire the lock to refresh LVB */
1067 dlm_lock_sync(cinfo->bitmap_lockres, DLM_LOCK_PW);
1068 cmsg.type = cpu_to_le32(RESYNCING);
1069 cmsg.low = cpu_to_le64(lo);
1070 cmsg.high = cpu_to_le64(hi);
1071
1072 return sendmsg(cinfo, &cmsg);
1073}
1074
1075static int resync_finish(struct mddev *mddev)
1076{
1077 struct md_cluster_info *cinfo = mddev->cluster_info;
1078 dlm_unlock_sync(cinfo->resync_lockres);
1079 return resync_info_update(mddev, 0, 0);
1080}
1081
1082static int area_resyncing(struct mddev *mddev, int direction,
1083 sector_t lo, sector_t hi)
1084{
1085 struct md_cluster_info *cinfo = mddev->cluster_info;
1086 int ret = 0;
1087 struct suspend_info *s;
1088
1089 if ((direction == READ) &&
1090 test_bit(MD_CLUSTER_SUSPEND_READ_BALANCING, &cinfo->state))
1091 return 1;
1092
1093 spin_lock_irq(&cinfo->suspend_lock);
1094 if (list_empty(&cinfo->suspend_list))
1095 goto out;
1096 list_for_each_entry(s, &cinfo->suspend_list, list)
1097 if (hi > s->lo && lo < s->hi) {
1098 ret = 1;
1099 break;
1100 }
1101out:
1102 spin_unlock_irq(&cinfo->suspend_lock);
1103 return ret;
1104}
1105
1106/* add_new_disk() - initiates a disk add
1107 * However, if this fails before writing md_update_sb(),
1108 * add_new_disk_cancel() must be called to release token lock
1109 */
1110static int add_new_disk(struct mddev *mddev, struct md_rdev *rdev)
1111{
1112 struct md_cluster_info *cinfo = mddev->cluster_info;
1113 struct cluster_msg cmsg;
1114 int ret = 0;
1115 struct mdp_superblock_1 *sb = page_address(rdev->sb_page);
1116 char *uuid = sb->device_uuid;
1117
1118 memset(&cmsg, 0, sizeof(cmsg));
1119 cmsg.type = cpu_to_le32(NEWDISK);
1120 memcpy(cmsg.uuid, uuid, 16);
1121 cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
1122 lock_comm(cinfo);
1123 ret = __sendmsg(cinfo, &cmsg);
1124 if (ret)
1125 return ret;
1126 cinfo->no_new_dev_lockres->flags |= DLM_LKF_NOQUEUE;
1127 ret = dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_EX);
1128 cinfo->no_new_dev_lockres->flags &= ~DLM_LKF_NOQUEUE;
1129 /* Some node does not "see" the device */
1130 if (ret == -EAGAIN)
1131 ret = -ENOENT;
1132 if (ret)
1133 unlock_comm(cinfo);
1134 else {
1135 dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR);
1136 /* Since MD_CHANGE_DEVS will be set in add_bound_rdev which
1137 * will run soon after add_new_disk, the below path will be
1138 * invoked:
1139 * md_wakeup_thread(mddev->thread)
1140 * -> conf->thread (raid1d)
1141 * -> md_check_recovery -> md_update_sb
1142 * -> metadata_update_start/finish
1143 * MD_CLUSTER_SEND_LOCKED_ALREADY will be cleared eventually.
1144 *
1145 * For other failure cases, metadata_update_cancel and
1146 * add_new_disk_cancel also clear below bit as well.
1147 * */
1148 set_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1149 wake_up(&cinfo->wait);
1150 }
1151 return ret;
1152}
1153
1154static void add_new_disk_cancel(struct mddev *mddev)
1155{
1156 struct md_cluster_info *cinfo = mddev->cluster_info;
1157 clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1158 unlock_comm(cinfo);
1159}
1160
1161static int new_disk_ack(struct mddev *mddev, bool ack)
1162{
1163 struct md_cluster_info *cinfo = mddev->cluster_info;
1164
1165 if (!test_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state)) {
1166 pr_warn("md-cluster(%s): Spurious cluster confirmation\n", mdname(mddev));
1167 return -EINVAL;
1168 }
1169
1170 if (ack)
1171 dlm_unlock_sync(cinfo->no_new_dev_lockres);
1172 complete(&cinfo->newdisk_completion);
1173 return 0;
1174}
1175
1176static int remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1177{
1178 struct cluster_msg cmsg = {0};
1179 struct md_cluster_info *cinfo = mddev->cluster_info;
1180 cmsg.type = cpu_to_le32(REMOVE);
1181 cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
1182 return sendmsg(cinfo, &cmsg);
1183}
1184
1185static int lock_all_bitmaps(struct mddev *mddev)
1186{
1187 int slot, my_slot, ret, held = 1, i = 0;
1188 char str[64];
1189 struct md_cluster_info *cinfo = mddev->cluster_info;
1190
1191 cinfo->other_bitmap_lockres = kzalloc((mddev->bitmap_info.nodes - 1) *
1192 sizeof(struct dlm_lock_resource *),
1193 GFP_KERNEL);
1194 if (!cinfo->other_bitmap_lockres) {
1195 pr_err("md: can't alloc mem for other bitmap locks\n");
1196 return 0;
1197 }
1198
1199 my_slot = slot_number(mddev);
1200 for (slot = 0; slot < mddev->bitmap_info.nodes; slot++) {
1201 if (slot == my_slot)
1202 continue;
1203
1204 memset(str, '\0', 64);
1205 snprintf(str, 64, "bitmap%04d", slot);
1206 cinfo->other_bitmap_lockres[i] = lockres_init(mddev, str, NULL, 1);
1207 if (!cinfo->other_bitmap_lockres[i])
1208 return -ENOMEM;
1209
1210 cinfo->other_bitmap_lockres[i]->flags |= DLM_LKF_NOQUEUE;
1211 ret = dlm_lock_sync(cinfo->other_bitmap_lockres[i], DLM_LOCK_PW);
1212 if (ret)
1213 held = -1;
1214 i++;
1215 }
1216
1217 return held;
1218}
1219
1220static void unlock_all_bitmaps(struct mddev *mddev)
1221{
1222 struct md_cluster_info *cinfo = mddev->cluster_info;
1223 int i;
1224
1225 /* release other node's bitmap lock if they are existed */
1226 if (cinfo->other_bitmap_lockres) {
1227 for (i = 0; i < mddev->bitmap_info.nodes - 1; i++) {
1228 if (cinfo->other_bitmap_lockres[i]) {
1229 lockres_free(cinfo->other_bitmap_lockres[i]);
1230 }
1231 }
1232 kfree(cinfo->other_bitmap_lockres);
1233 }
1234}
1235
1236static int gather_bitmaps(struct md_rdev *rdev)
1237{
1238 int sn, err;
1239 sector_t lo, hi;
1240 struct cluster_msg cmsg = {0};
1241 struct mddev *mddev = rdev->mddev;
1242 struct md_cluster_info *cinfo = mddev->cluster_info;
1243
1244 cmsg.type = cpu_to_le32(RE_ADD);
1245 cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
1246 err = sendmsg(cinfo, &cmsg);
1247 if (err)
1248 goto out;
1249
1250 for (sn = 0; sn < mddev->bitmap_info.nodes; sn++) {
1251 if (sn == (cinfo->slot_number - 1))
1252 continue;
1253 err = bitmap_copy_from_slot(mddev, sn, &lo, &hi, false);
1254 if (err) {
1255 pr_warn("md-cluster: Could not gather bitmaps from slot %d", sn);
1256 goto out;
1257 }
1258 if ((hi > 0) && (lo < mddev->recovery_cp))
1259 mddev->recovery_cp = lo;
1260 }
1261out:
1262 return err;
1263}
1264
1265static struct md_cluster_operations cluster_ops = {
1266 .join = join,
1267 .leave = leave,
1268 .slot_number = slot_number,
1269 .resync_start = resync_start,
1270 .resync_finish = resync_finish,
1271 .resync_info_update = resync_info_update,
1272 .metadata_update_start = metadata_update_start,
1273 .metadata_update_finish = metadata_update_finish,
1274 .metadata_update_cancel = metadata_update_cancel,
1275 .area_resyncing = area_resyncing,
1276 .add_new_disk = add_new_disk,
1277 .add_new_disk_cancel = add_new_disk_cancel,
1278 .new_disk_ack = new_disk_ack,
1279 .remove_disk = remove_disk,
1280 .load_bitmaps = load_bitmaps,
1281 .gather_bitmaps = gather_bitmaps,
1282 .lock_all_bitmaps = lock_all_bitmaps,
1283 .unlock_all_bitmaps = unlock_all_bitmaps,
1284};
1285
1286static int __init cluster_init(void)
1287{
1288 pr_warn("md-cluster: EXPERIMENTAL. Use with caution\n");
1289 pr_info("Registering Cluster MD functions\n");
1290 register_md_cluster_operations(&cluster_ops, THIS_MODULE);
1291 return 0;
1292}
1293
1294static void cluster_exit(void)
1295{
1296 unregister_md_cluster_operations();
1297}
1298
1299module_init(cluster_init);
1300module_exit(cluster_exit);
1301MODULE_AUTHOR("SUSE");
1302MODULE_LICENSE("GPL");
1303MODULE_DESCRIPTION("Clustering support for MD");