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