xref: /linux/drivers/md/md-cluster.c (revision 26bb0d3f38a764b743a3ad5c8b6e5b5044d7ceb4)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * Copyright (C) 2015, SUSE
4  */
5 
6 
7 #include <linux/module.h>
8 #include <linux/kthread.h>
9 #include <linux/dlm.h>
10 #include <linux/sched.h>
11 #include <linux/raid/md_p.h>
12 #include "md.h"
13 #include "md-bitmap.h"
14 #include "md-cluster.h"
15 
16 #define LVB_SIZE	64
17 #define NEW_DEV_TIMEOUT 5000
18 #define WAIT_DLM_LOCK_TIMEOUT (30 * HZ)
19 
20 struct dlm_lock_resource {
21 	dlm_lockspace_t *ls;
22 	struct dlm_lksb lksb;
23 	char *name; /* lock name. */
24 	uint32_t flags; /* flags to pass to dlm_lock() */
25 	wait_queue_head_t sync_locking; /* wait queue for synchronized locking */
26 	bool sync_locking_done;
27 	void (*bast)(void *arg, int mode); /* blocking AST function pointer*/
28 	struct mddev *mddev; /* pointing back to mddev. */
29 	int mode;
30 };
31 
32 struct resync_info {
33 	__le64 lo;
34 	__le64 hi;
35 };
36 
37 /* md_cluster_info flags */
38 #define		MD_CLUSTER_WAITING_FOR_NEWDISK		1
39 #define		MD_CLUSTER_SUSPEND_READ_BALANCING	2
40 #define		MD_CLUSTER_BEGIN_JOIN_CLUSTER		3
41 
42 /* Lock the send communication. This is done through
43  * bit manipulation as opposed to a mutex in order to
44  * accommodate lock and hold. See next comment.
45  */
46 #define		MD_CLUSTER_SEND_LOCK			4
47 /* If cluster operations (such as adding a disk) must lock the
48  * communication channel, so as to perform extra operations
49  * (update metadata) and no other operation is allowed on the
50  * MD. Token needs to be locked and held until the operation
51  * completes witha md_update_sb(), which would eventually release
52  * the lock.
53  */
54 #define		MD_CLUSTER_SEND_LOCKED_ALREADY		5
55 /* We should receive message after node joined cluster and
56  * set up all the related infos such as bitmap and personality */
57 #define		MD_CLUSTER_ALREADY_IN_CLUSTER		6
58 #define		MD_CLUSTER_PENDING_RECV_EVENT		7
59 #define 	MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD		8
60 #define		MD_CLUSTER_WAITING_FOR_SYNC		9
61 
62 struct md_cluster_info {
63 	struct mddev *mddev; /* the md device which md_cluster_info belongs to */
64 	/* dlm lock space and resources for clustered raid. */
65 	dlm_lockspace_t *lockspace;
66 	int slot_number;
67 	struct completion completion;
68 	struct mutex recv_mutex;
69 	struct dlm_lock_resource *bitmap_lockres;
70 	struct dlm_lock_resource **other_bitmap_lockres;
71 	struct dlm_lock_resource *resync_lockres;
72 	struct list_head suspend_list;
73 
74 	spinlock_t suspend_lock;
75 	/* record the region which write should be suspended */
76 	sector_t suspend_lo;
77 	sector_t suspend_hi;
78 	int suspend_from; /* the slot which broadcast suspend_lo/hi */
79 
80 	struct md_thread __rcu *recovery_thread;
81 	unsigned long recovery_map;
82 	/* communication loc resources */
83 	struct dlm_lock_resource *ack_lockres;
84 	struct dlm_lock_resource *message_lockres;
85 	struct dlm_lock_resource *token_lockres;
86 	struct dlm_lock_resource *no_new_dev_lockres;
87 	struct md_thread __rcu *recv_thread;
88 	struct completion newdisk_completion;
89 	wait_queue_head_t wait;
90 	unsigned long state;
91 	/* record the region in RESYNCING message */
92 	sector_t sync_low;
93 	sector_t sync_hi;
94 };
95 
96 /* For compatibility, add the new msg_type at the end. */
97 enum msg_type {
98 	METADATA_UPDATED = 0,
99 	RESYNCING,
100 	NEWDISK,
101 	REMOVE,
102 	RE_ADD,
103 	BITMAP_NEEDS_SYNC,
104 	CHANGE_CAPACITY,
105 	BITMAP_RESIZE,
106 	RESYNCING_START,
107 };
108 
109 struct cluster_msg {
110 	__le32 type;
111 	__le32 slot;
112 	/* TODO: Unionize this for smaller footprint */
113 	__le64 low;
114 	__le64 high;
115 	char uuid[16];
116 	__le32 raid_slot;
117 };
118 
sync_ast(void * arg)119 static void sync_ast(void *arg)
120 {
121 	struct dlm_lock_resource *res;
122 
123 	res = arg;
124 	res->sync_locking_done = true;
125 	wake_up(&res->sync_locking);
126 }
127 
dlm_lock_sync(struct dlm_lock_resource * res,int mode)128 static int dlm_lock_sync(struct dlm_lock_resource *res, int mode)
129 {
130 	int ret = 0;
131 
132 	ret = dlm_lock(res->ls, mode, &res->lksb,
133 			res->flags, res->name, strlen(res->name),
134 			0, sync_ast, res, res->bast);
135 	if (ret)
136 		return ret;
137 	ret = wait_event_timeout(res->sync_locking, res->sync_locking_done,
138 				WAIT_DLM_LOCK_TIMEOUT);
139 	res->sync_locking_done = false;
140 	if (!ret) {
141 		pr_err("locking DLM '%s' timeout!\n", res->name);
142 		return -EBUSY;
143 	}
144 	if (res->lksb.sb_status == 0)
145 		res->mode = mode;
146 	return res->lksb.sb_status;
147 }
148 
dlm_unlock_sync(struct dlm_lock_resource * res)149 static int dlm_unlock_sync(struct dlm_lock_resource *res)
150 {
151 	return dlm_lock_sync(res, DLM_LOCK_NL);
152 }
153 
154 /*
155  * An variation of dlm_lock_sync, which make lock request could
156  * be interrupted
157  */
dlm_lock_sync_interruptible(struct dlm_lock_resource * res,int mode,struct mddev * mddev)158 static int dlm_lock_sync_interruptible(struct dlm_lock_resource *res, int mode,
159 				       struct mddev *mddev)
160 {
161 	int ret = 0;
162 
163 	ret = dlm_lock(res->ls, mode, &res->lksb,
164 			res->flags, res->name, strlen(res->name),
165 			0, sync_ast, res, res->bast);
166 	if (ret)
167 		return ret;
168 
169 	wait_event(res->sync_locking, res->sync_locking_done
170 				      || kthread_should_stop()
171 				      || test_bit(MD_CLOSING, &mddev->flags));
172 	if (!res->sync_locking_done) {
173 		/*
174 		 * the convert queue contains the lock request when request is
175 		 * interrupted, and sync_ast could still be run, so need to
176 		 * cancel the request and reset completion
177 		 */
178 		ret = dlm_unlock(res->ls, res->lksb.sb_lkid, DLM_LKF_CANCEL,
179 			&res->lksb, res);
180 		res->sync_locking_done = false;
181 		if (unlikely(ret != 0))
182 			pr_info("failed to cancel previous lock request "
183 				 "%s return %d\n", res->name, ret);
184 		return -EPERM;
185 	} else
186 		res->sync_locking_done = false;
187 	if (res->lksb.sb_status == 0)
188 		res->mode = mode;
189 	return res->lksb.sb_status;
190 }
191 
lockres_init(struct mddev * mddev,char * name,void (* bastfn)(void * arg,int mode),int with_lvb)192 static struct dlm_lock_resource *lockres_init(struct mddev *mddev,
193 		char *name, void (*bastfn)(void *arg, int mode), int with_lvb)
194 {
195 	struct dlm_lock_resource *res = NULL;
196 	int ret, namelen;
197 	struct md_cluster_info *cinfo = mddev->cluster_info;
198 
199 	res = kzalloc(sizeof(struct dlm_lock_resource), GFP_KERNEL);
200 	if (!res)
201 		return NULL;
202 	init_waitqueue_head(&res->sync_locking);
203 	res->sync_locking_done = false;
204 	res->ls = cinfo->lockspace;
205 	res->mddev = mddev;
206 	res->mode = DLM_LOCK_IV;
207 	namelen = strlen(name);
208 	res->name = kzalloc(namelen + 1, GFP_KERNEL);
209 	if (!res->name) {
210 		pr_err("md-cluster: Unable to allocate resource name for resource %s\n", name);
211 		goto out_err;
212 	}
213 	strscpy(res->name, name, namelen + 1);
214 	if (with_lvb) {
215 		res->lksb.sb_lvbptr = kzalloc(LVB_SIZE, GFP_KERNEL);
216 		if (!res->lksb.sb_lvbptr) {
217 			pr_err("md-cluster: Unable to allocate LVB for resource %s\n", name);
218 			goto out_err;
219 		}
220 		res->flags = DLM_LKF_VALBLK;
221 	}
222 
223 	if (bastfn)
224 		res->bast = bastfn;
225 
226 	res->flags |= DLM_LKF_EXPEDITE;
227 
228 	ret = dlm_lock_sync(res, DLM_LOCK_NL);
229 	if (ret) {
230 		pr_err("md-cluster: Unable to lock NL on new lock resource %s\n", name);
231 		goto out_err;
232 	}
233 	res->flags &= ~DLM_LKF_EXPEDITE;
234 	res->flags |= DLM_LKF_CONVERT;
235 
236 	return res;
237 out_err:
238 	kfree(res->lksb.sb_lvbptr);
239 	kfree(res->name);
240 	kfree(res);
241 	return NULL;
242 }
243 
lockres_free(struct dlm_lock_resource * res)244 static void lockres_free(struct dlm_lock_resource *res)
245 {
246 	int ret = 0;
247 
248 	if (!res)
249 		return;
250 
251 	/*
252 	 * use FORCEUNLOCK flag, so we can unlock even the lock is on the
253 	 * waiting or convert queue
254 	 */
255 	ret = dlm_unlock(res->ls, res->lksb.sb_lkid, DLM_LKF_FORCEUNLOCK,
256 		&res->lksb, res);
257 	if (unlikely(ret != 0))
258 		pr_err("failed to unlock %s return %d\n", res->name, ret);
259 	else
260 		wait_event(res->sync_locking, res->sync_locking_done);
261 
262 	kfree(res->name);
263 	kfree(res->lksb.sb_lvbptr);
264 	kfree(res);
265 }
266 
add_resync_info(struct dlm_lock_resource * lockres,sector_t lo,sector_t hi)267 static void add_resync_info(struct dlm_lock_resource *lockres,
268 			    sector_t lo, sector_t hi)
269 {
270 	struct resync_info *ri;
271 
272 	ri = (struct resync_info *)lockres->lksb.sb_lvbptr;
273 	ri->lo = cpu_to_le64(lo);
274 	ri->hi = cpu_to_le64(hi);
275 }
276 
read_resync_info(struct mddev * mddev,struct dlm_lock_resource * lockres)277 static int read_resync_info(struct mddev *mddev,
278 			    struct dlm_lock_resource *lockres)
279 {
280 	struct resync_info ri;
281 	struct md_cluster_info *cinfo = mddev->cluster_info;
282 	int ret = 0;
283 
284 	dlm_lock_sync(lockres, DLM_LOCK_CR);
285 	memcpy(&ri, lockres->lksb.sb_lvbptr, sizeof(struct resync_info));
286 	if (le64_to_cpu(ri.hi) > 0) {
287 		cinfo->suspend_hi = le64_to_cpu(ri.hi);
288 		cinfo->suspend_lo = le64_to_cpu(ri.lo);
289 		ret = 1;
290 	}
291 	dlm_unlock_sync(lockres);
292 	return ret;
293 }
294 
recover_bitmaps(struct md_thread * thread)295 static void recover_bitmaps(struct md_thread *thread)
296 {
297 	struct mddev *mddev = thread->mddev;
298 	struct md_cluster_info *cinfo = mddev->cluster_info;
299 	struct dlm_lock_resource *bm_lockres;
300 	char str[64];
301 	int slot, ret;
302 	sector_t lo, hi;
303 
304 	while (cinfo->recovery_map) {
305 		slot = fls64((u64)cinfo->recovery_map) - 1;
306 
307 		snprintf(str, 64, "bitmap%04d", slot);
308 		bm_lockres = lockres_init(mddev, str, NULL, 1);
309 		if (!bm_lockres) {
310 			pr_err("md-cluster: Cannot initialize bitmaps\n");
311 			goto clear_bit;
312 		}
313 
314 		ret = dlm_lock_sync_interruptible(bm_lockres, DLM_LOCK_PW, mddev);
315 		if (ret) {
316 			pr_err("md-cluster: Could not DLM lock %s: %d\n",
317 					str, ret);
318 			goto clear_bit;
319 		}
320 		ret = mddev->bitmap_ops->copy_from_slot(mddev, slot, &lo, &hi, true);
321 		if (ret) {
322 			pr_err("md-cluster: Could not copy data from bitmap %d\n", slot);
323 			goto clear_bit;
324 		}
325 
326 		/* Clear suspend_area associated with the bitmap */
327 		spin_lock_irq(&cinfo->suspend_lock);
328 		cinfo->suspend_hi = 0;
329 		cinfo->suspend_lo = 0;
330 		cinfo->suspend_from = -1;
331 		spin_unlock_irq(&cinfo->suspend_lock);
332 
333 		/* Kick off a reshape if needed */
334 		if (test_bit(MD_RESYNCING_REMOTE, &mddev->recovery) &&
335 		    test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
336 		    mddev->reshape_position != MaxSector)
337 			md_wakeup_thread(mddev->sync_thread);
338 
339 		if (hi > 0) {
340 			if (lo < mddev->recovery_cp)
341 				mddev->recovery_cp = lo;
342 			/* wake up thread to continue resync in case resync
343 			 * is not finished */
344 			if (mddev->recovery_cp != MaxSector) {
345 				/*
346 				 * clear the REMOTE flag since we will launch
347 				 * resync thread in current node.
348 				 */
349 				clear_bit(MD_RESYNCING_REMOTE,
350 					  &mddev->recovery);
351 				set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
352 				md_wakeup_thread(mddev->thread);
353 			}
354 		}
355 clear_bit:
356 		lockres_free(bm_lockres);
357 		clear_bit(slot, &cinfo->recovery_map);
358 	}
359 }
360 
recover_prep(void * arg)361 static void recover_prep(void *arg)
362 {
363 	struct mddev *mddev = arg;
364 	struct md_cluster_info *cinfo = mddev->cluster_info;
365 	set_bit(MD_CLUSTER_SUSPEND_READ_BALANCING, &cinfo->state);
366 }
367 
__recover_slot(struct mddev * mddev,int slot)368 static void __recover_slot(struct mddev *mddev, int slot)
369 {
370 	struct md_cluster_info *cinfo = mddev->cluster_info;
371 
372 	set_bit(slot, &cinfo->recovery_map);
373 	if (!cinfo->recovery_thread) {
374 		rcu_assign_pointer(cinfo->recovery_thread,
375 			md_register_thread(recover_bitmaps, mddev, "recover"));
376 		if (!cinfo->recovery_thread) {
377 			pr_warn("md-cluster: Could not create recovery thread\n");
378 			return;
379 		}
380 	}
381 	md_wakeup_thread(cinfo->recovery_thread);
382 }
383 
recover_slot(void * arg,struct dlm_slot * slot)384 static void recover_slot(void *arg, struct dlm_slot *slot)
385 {
386 	struct mddev *mddev = arg;
387 	struct md_cluster_info *cinfo = mddev->cluster_info;
388 
389 	pr_info("md-cluster: %s Node %d/%d down. My slot: %d. Initiating recovery.\n",
390 			mddev->bitmap_info.cluster_name,
391 			slot->nodeid, slot->slot,
392 			cinfo->slot_number);
393 	/* deduct one since dlm slot starts from one while the num of
394 	 * cluster-md begins with 0 */
395 	__recover_slot(mddev, slot->slot - 1);
396 }
397 
recover_done(void * arg,struct dlm_slot * slots,int num_slots,int our_slot,uint32_t generation)398 static void recover_done(void *arg, struct dlm_slot *slots,
399 		int num_slots, int our_slot,
400 		uint32_t generation)
401 {
402 	struct mddev *mddev = arg;
403 	struct md_cluster_info *cinfo = mddev->cluster_info;
404 
405 	cinfo->slot_number = our_slot;
406 	/* completion is only need to be complete when node join cluster,
407 	 * it doesn't need to run during another node's failure */
408 	if (test_bit(MD_CLUSTER_BEGIN_JOIN_CLUSTER, &cinfo->state)) {
409 		complete(&cinfo->completion);
410 		clear_bit(MD_CLUSTER_BEGIN_JOIN_CLUSTER, &cinfo->state);
411 	}
412 	clear_bit(MD_CLUSTER_SUSPEND_READ_BALANCING, &cinfo->state);
413 }
414 
415 /* the ops is called when node join the cluster, and do lock recovery
416  * if node failure occurs */
417 static const struct dlm_lockspace_ops md_ls_ops = {
418 	.recover_prep = recover_prep,
419 	.recover_slot = recover_slot,
420 	.recover_done = recover_done,
421 };
422 
423 /*
424  * The BAST function for the ack lock resource
425  * This function wakes up the receive thread in
426  * order to receive and process the message.
427  */
ack_bast(void * arg,int mode)428 static void ack_bast(void *arg, int mode)
429 {
430 	struct dlm_lock_resource *res = arg;
431 	struct md_cluster_info *cinfo = res->mddev->cluster_info;
432 
433 	if (mode == DLM_LOCK_EX) {
434 		if (test_bit(MD_CLUSTER_ALREADY_IN_CLUSTER, &cinfo->state))
435 			md_wakeup_thread(cinfo->recv_thread);
436 		else
437 			set_bit(MD_CLUSTER_PENDING_RECV_EVENT, &cinfo->state);
438 	}
439 }
440 
remove_suspend_info(struct mddev * mddev,int slot)441 static void remove_suspend_info(struct mddev *mddev, int slot)
442 {
443 	struct md_cluster_info *cinfo = mddev->cluster_info;
444 	mddev->pers->quiesce(mddev, 1);
445 	spin_lock_irq(&cinfo->suspend_lock);
446 	cinfo->suspend_hi = 0;
447 	cinfo->suspend_lo = 0;
448 	spin_unlock_irq(&cinfo->suspend_lock);
449 	mddev->pers->quiesce(mddev, 0);
450 }
451 
process_suspend_info(struct mddev * mddev,int slot,sector_t lo,sector_t hi)452 static void process_suspend_info(struct mddev *mddev,
453 		int slot, sector_t lo, sector_t hi)
454 {
455 	struct md_cluster_info *cinfo = mddev->cluster_info;
456 	struct mdp_superblock_1 *sb = NULL;
457 	struct md_rdev *rdev;
458 
459 	if (!hi) {
460 		/*
461 		 * clear the REMOTE flag since resync or recovery is finished
462 		 * in remote node.
463 		 */
464 		clear_bit(MD_RESYNCING_REMOTE, &mddev->recovery);
465 		remove_suspend_info(mddev, slot);
466 		set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
467 		clear_bit(MD_CLUSTER_WAITING_FOR_SYNC, &cinfo->state);
468 		md_wakeup_thread(mddev->thread);
469 		return;
470 	}
471 
472 	rdev_for_each(rdev, mddev)
473 		if (rdev->raid_disk > -1 && !test_bit(Faulty, &rdev->flags)) {
474 			sb = page_address(rdev->sb_page);
475 			break;
476 		}
477 
478 	/*
479 	 * The bitmaps are not same for different nodes
480 	 * if RESYNCING is happening in one node, then
481 	 * the node which received the RESYNCING message
482 	 * probably will perform resync with the region
483 	 * [lo, hi] again, so we could reduce resync time
484 	 * a lot if we can ensure that the bitmaps among
485 	 * different nodes are match up well.
486 	 *
487 	 * sync_low/hi is used to record the region which
488 	 * arrived in the previous RESYNCING message,
489 	 *
490 	 * Call md_bitmap_sync_with_cluster to clear NEEDED_MASK
491 	 * and set RESYNC_MASK since  resync thread is running
492 	 * in another node, so we don't need to do the resync
493 	 * again with the same section.
494 	 *
495 	 * Skip md_bitmap_sync_with_cluster in case reshape
496 	 * happening, because reshaping region is small and
497 	 * we don't want to trigger lots of WARN.
498 	 */
499 	if (sb && !(le32_to_cpu(sb->feature_map) & MD_FEATURE_RESHAPE_ACTIVE))
500 		mddev->bitmap_ops->sync_with_cluster(mddev, cinfo->sync_low,
501 						     cinfo->sync_hi, lo, hi);
502 	cinfo->sync_low = lo;
503 	cinfo->sync_hi = hi;
504 
505 	mddev->pers->quiesce(mddev, 1);
506 	spin_lock_irq(&cinfo->suspend_lock);
507 	cinfo->suspend_from = slot;
508 	cinfo->suspend_lo = lo;
509 	cinfo->suspend_hi = hi;
510 	spin_unlock_irq(&cinfo->suspend_lock);
511 	mddev->pers->quiesce(mddev, 0);
512 }
513 
process_add_new_disk(struct mddev * mddev,struct cluster_msg * cmsg)514 static int process_add_new_disk(struct mddev *mddev, struct cluster_msg *cmsg)
515 {
516 	char disk_uuid[64];
517 	struct md_cluster_info *cinfo = mddev->cluster_info;
518 	char event_name[] = "EVENT=ADD_DEVICE";
519 	char raid_slot[16];
520 	char *envp[] = {event_name, disk_uuid, raid_slot, NULL};
521 	int len;
522 	int res = 0;
523 
524 	len = snprintf(disk_uuid, 64, "DEVICE_UUID=");
525 	sprintf(disk_uuid + len, "%pU", cmsg->uuid);
526 	snprintf(raid_slot, 16, "RAID_DISK=%d", le32_to_cpu(cmsg->raid_slot));
527 	pr_info("%s:%d Sending kobject change with %s and %s\n", __func__, __LINE__, disk_uuid, raid_slot);
528 	init_completion(&cinfo->newdisk_completion);
529 	set_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state);
530 	kobject_uevent_env(&disk_to_dev(mddev->gendisk)->kobj, KOBJ_CHANGE, envp);
531 	if (!wait_for_completion_timeout(&cinfo->newdisk_completion,
532 					NEW_DEV_TIMEOUT)) {
533 		pr_err("md-cluster(%s:%d): timeout on a new disk adding\n",
534 			__func__, __LINE__);
535 		res = -1;
536 	}
537 	clear_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state);
538 	set_bit(MD_CLUSTER_WAITING_FOR_SYNC, &cinfo->state);
539 	return res;
540 }
541 
542 
process_metadata_update(struct mddev * mddev,struct cluster_msg * msg)543 static void process_metadata_update(struct mddev *mddev, struct cluster_msg *msg)
544 {
545 	int got_lock = 0;
546 	struct md_thread *thread;
547 	struct md_cluster_info *cinfo = mddev->cluster_info;
548 	mddev->good_device_nr = le32_to_cpu(msg->raid_slot);
549 
550 	dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR);
551 
552 	/* daemaon thread must exist */
553 	thread = rcu_dereference_protected(mddev->thread, true);
554 	wait_event(thread->wqueue,
555 		   (got_lock = mddev_trylock(mddev)) ||
556 		    test_bit(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state));
557 	md_reload_sb(mddev, mddev->good_device_nr);
558 	if (got_lock)
559 		mddev_unlock(mddev);
560 }
561 
process_remove_disk(struct mddev * mddev,struct cluster_msg * msg)562 static void process_remove_disk(struct mddev *mddev, struct cluster_msg *msg)
563 {
564 	struct md_rdev *rdev;
565 
566 	rcu_read_lock();
567 	rdev = md_find_rdev_nr_rcu(mddev, le32_to_cpu(msg->raid_slot));
568 	if (rdev) {
569 		set_bit(ClusterRemove, &rdev->flags);
570 		set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
571 		md_wakeup_thread(mddev->thread);
572 	}
573 	else
574 		pr_warn("%s: %d Could not find disk(%d) to REMOVE\n",
575 			__func__, __LINE__, le32_to_cpu(msg->raid_slot));
576 	rcu_read_unlock();
577 }
578 
process_readd_disk(struct mddev * mddev,struct cluster_msg * msg)579 static void process_readd_disk(struct mddev *mddev, struct cluster_msg *msg)
580 {
581 	struct md_rdev *rdev;
582 
583 	rcu_read_lock();
584 	rdev = md_find_rdev_nr_rcu(mddev, le32_to_cpu(msg->raid_slot));
585 	if (rdev && test_bit(Faulty, &rdev->flags))
586 		clear_bit(Faulty, &rdev->flags);
587 	else
588 		pr_warn("%s: %d Could not find disk(%d) which is faulty",
589 			__func__, __LINE__, le32_to_cpu(msg->raid_slot));
590 	rcu_read_unlock();
591 }
592 
process_recvd_msg(struct mddev * mddev,struct cluster_msg * msg)593 static int process_recvd_msg(struct mddev *mddev, struct cluster_msg *msg)
594 {
595 	int ret = 0;
596 
597 	if (WARN(mddev->cluster_info->slot_number - 1 == le32_to_cpu(msg->slot),
598 		"node %d received its own msg\n", le32_to_cpu(msg->slot)))
599 		return -1;
600 	switch (le32_to_cpu(msg->type)) {
601 	case METADATA_UPDATED:
602 		process_metadata_update(mddev, msg);
603 		break;
604 	case CHANGE_CAPACITY:
605 		set_capacity_and_notify(mddev->gendisk, mddev->array_sectors);
606 		break;
607 	case RESYNCING_START:
608 		clear_bit(MD_CLUSTER_WAITING_FOR_SYNC, &mddev->cluster_info->state);
609 		break;
610 	case RESYNCING:
611 		set_bit(MD_RESYNCING_REMOTE, &mddev->recovery);
612 		process_suspend_info(mddev, le32_to_cpu(msg->slot),
613 				     le64_to_cpu(msg->low),
614 				     le64_to_cpu(msg->high));
615 		break;
616 	case NEWDISK:
617 		if (process_add_new_disk(mddev, msg))
618 			ret = -1;
619 		break;
620 	case REMOVE:
621 		process_remove_disk(mddev, msg);
622 		break;
623 	case RE_ADD:
624 		process_readd_disk(mddev, msg);
625 		break;
626 	case BITMAP_NEEDS_SYNC:
627 		__recover_slot(mddev, le32_to_cpu(msg->slot));
628 		break;
629 	case BITMAP_RESIZE:
630 		if (le64_to_cpu(msg->high) != mddev->pers->size(mddev, 0, 0))
631 			ret = mddev->bitmap_ops->resize(mddev,
632 							le64_to_cpu(msg->high),
633 							0, false);
634 		break;
635 	default:
636 		ret = -1;
637 		pr_warn("%s:%d Received unknown message from %d\n",
638 			__func__, __LINE__, msg->slot);
639 	}
640 	return ret;
641 }
642 
643 /*
644  * thread for receiving message
645  */
recv_daemon(struct md_thread * thread)646 static void recv_daemon(struct md_thread *thread)
647 {
648 	struct md_cluster_info *cinfo = thread->mddev->cluster_info;
649 	struct dlm_lock_resource *ack_lockres = cinfo->ack_lockres;
650 	struct dlm_lock_resource *message_lockres = cinfo->message_lockres;
651 	struct cluster_msg msg;
652 	int ret;
653 
654 	mutex_lock(&cinfo->recv_mutex);
655 	/*get CR on Message*/
656 	if (dlm_lock_sync(message_lockres, DLM_LOCK_CR)) {
657 		pr_err("md/raid1:failed to get CR on MESSAGE\n");
658 		mutex_unlock(&cinfo->recv_mutex);
659 		return;
660 	}
661 
662 	/* read lvb and wake up thread to process this message_lockres */
663 	memcpy(&msg, message_lockres->lksb.sb_lvbptr, sizeof(struct cluster_msg));
664 	ret = process_recvd_msg(thread->mddev, &msg);
665 	if (ret)
666 		goto out;
667 
668 	/*release CR on ack_lockres*/
669 	ret = dlm_unlock_sync(ack_lockres);
670 	if (unlikely(ret != 0))
671 		pr_info("unlock ack failed return %d\n", ret);
672 	/*up-convert to PR on message_lockres*/
673 	ret = dlm_lock_sync(message_lockres, DLM_LOCK_PR);
674 	if (unlikely(ret != 0))
675 		pr_info("lock PR on msg failed return %d\n", ret);
676 	/*get CR on ack_lockres again*/
677 	ret = dlm_lock_sync(ack_lockres, DLM_LOCK_CR);
678 	if (unlikely(ret != 0))
679 		pr_info("lock CR on ack failed return %d\n", ret);
680 out:
681 	/*release CR on message_lockres*/
682 	ret = dlm_unlock_sync(message_lockres);
683 	if (unlikely(ret != 0))
684 		pr_info("unlock msg failed return %d\n", ret);
685 	mutex_unlock(&cinfo->recv_mutex);
686 }
687 
688 /* lock_token()
689  * Takes the lock on the TOKEN lock resource so no other
690  * node can communicate while the operation is underway.
691  */
lock_token(struct md_cluster_info * cinfo)692 static int lock_token(struct md_cluster_info *cinfo)
693 {
694 	int error;
695 
696 	error = dlm_lock_sync(cinfo->token_lockres, DLM_LOCK_EX);
697 	if (error) {
698 		pr_err("md-cluster(%s:%d): failed to get EX on TOKEN (%d)\n",
699 				__func__, __LINE__, error);
700 	} else {
701 		/* Lock the receive sequence */
702 		mutex_lock(&cinfo->recv_mutex);
703 	}
704 	return error;
705 }
706 
707 /* lock_comm()
708  * Sets the MD_CLUSTER_SEND_LOCK bit to lock the send channel.
709  */
lock_comm(struct md_cluster_info * cinfo,bool mddev_locked)710 static int lock_comm(struct md_cluster_info *cinfo, bool mddev_locked)
711 {
712 	int rv, set_bit = 0;
713 	struct mddev *mddev = cinfo->mddev;
714 
715 	/*
716 	 * If resync thread run after raid1d thread, then process_metadata_update
717 	 * could not continue if raid1d held reconfig_mutex (and raid1d is blocked
718 	 * since another node already got EX on Token and waiting the EX of Ack),
719 	 * so let resync wake up thread in case flag is set.
720 	 */
721 	if (mddev_locked && !test_bit(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD,
722 				      &cinfo->state)) {
723 		rv = test_and_set_bit_lock(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD,
724 					      &cinfo->state);
725 		WARN_ON_ONCE(rv);
726 		md_wakeup_thread(mddev->thread);
727 		set_bit = 1;
728 	}
729 
730 	wait_event(cinfo->wait,
731 		   !test_and_set_bit(MD_CLUSTER_SEND_LOCK, &cinfo->state));
732 	rv = lock_token(cinfo);
733 	if (set_bit)
734 		clear_bit_unlock(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state);
735 	return rv;
736 }
737 
unlock_comm(struct md_cluster_info * cinfo)738 static void unlock_comm(struct md_cluster_info *cinfo)
739 {
740 	WARN_ON(cinfo->token_lockres->mode != DLM_LOCK_EX);
741 	mutex_unlock(&cinfo->recv_mutex);
742 	dlm_unlock_sync(cinfo->token_lockres);
743 	clear_bit(MD_CLUSTER_SEND_LOCK, &cinfo->state);
744 	wake_up(&cinfo->wait);
745 }
746 
747 /* __sendmsg()
748  * This function performs the actual sending of the message. This function is
749  * usually called after performing the encompassing operation
750  * The function:
751  * 1. Grabs the message lockresource in EX mode
752  * 2. Copies the message to the message LVB
753  * 3. Downconverts message lockresource to CW
754  * 4. Upconverts ack lock resource from CR to EX. This forces the BAST on other nodes
755  *    and the other nodes read the message. The thread will wait here until all other
756  *    nodes have released ack lock resource.
757  * 5. Downconvert ack lockresource to CR
758  */
__sendmsg(struct md_cluster_info * cinfo,struct cluster_msg * cmsg)759 static int __sendmsg(struct md_cluster_info *cinfo, struct cluster_msg *cmsg)
760 {
761 	int error, unlock_error;
762 	int slot = cinfo->slot_number - 1;
763 
764 	cmsg->slot = cpu_to_le32(slot);
765 	/*get EX on Message*/
766 	error = dlm_lock_sync(cinfo->message_lockres, DLM_LOCK_EX);
767 	if (error) {
768 		pr_err("md-cluster: failed to get EX on MESSAGE (%d)\n", error);
769 		return error;
770 	}
771 
772 	memcpy(cinfo->message_lockres->lksb.sb_lvbptr, (void *)cmsg,
773 			sizeof(struct cluster_msg));
774 	/*down-convert EX to CW on Message*/
775 	error = dlm_lock_sync(cinfo->message_lockres, DLM_LOCK_CW);
776 	if (error) {
777 		pr_err("md-cluster: failed to convert EX to CW on MESSAGE(%d)\n",
778 				error);
779 		goto failed_ack;
780 	}
781 
782 	/*up-convert CR to EX on Ack*/
783 	error = dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_EX);
784 	if (error) {
785 		pr_err("md-cluster: failed to convert CR to EX on ACK(%d)\n",
786 				error);
787 		goto failed_ack;
788 	}
789 
790 	/*down-convert EX to CR on Ack*/
791 	error = dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_CR);
792 	if (error) {
793 		pr_err("md-cluster: failed to convert EX to CR on ACK(%d)\n",
794 				error);
795 		goto failed_ack;
796 	}
797 
798 failed_ack:
799 	while ((unlock_error = dlm_unlock_sync(cinfo->message_lockres)))
800 		pr_err("md-cluster: failed convert to NL on MESSAGE(%d)\n",
801 			unlock_error);
802 
803 	return error;
804 }
805 
sendmsg(struct md_cluster_info * cinfo,struct cluster_msg * cmsg,bool mddev_locked)806 static int sendmsg(struct md_cluster_info *cinfo, struct cluster_msg *cmsg,
807 		   bool mddev_locked)
808 {
809 	int ret;
810 
811 	ret = lock_comm(cinfo, mddev_locked);
812 	if (!ret) {
813 		ret = __sendmsg(cinfo, cmsg);
814 		unlock_comm(cinfo);
815 	}
816 	return ret;
817 }
818 
gather_all_resync_info(struct mddev * mddev,int total_slots)819 static int gather_all_resync_info(struct mddev *mddev, int total_slots)
820 {
821 	struct md_cluster_info *cinfo = mddev->cluster_info;
822 	int i, ret = 0;
823 	struct dlm_lock_resource *bm_lockres;
824 	char str[64];
825 	sector_t lo, hi;
826 
827 
828 	for (i = 0; i < total_slots; i++) {
829 		memset(str, '\0', 64);
830 		snprintf(str, 64, "bitmap%04d", i);
831 		bm_lockres = lockres_init(mddev, str, NULL, 1);
832 		if (!bm_lockres)
833 			return -ENOMEM;
834 		if (i == (cinfo->slot_number - 1)) {
835 			lockres_free(bm_lockres);
836 			continue;
837 		}
838 
839 		bm_lockres->flags |= DLM_LKF_NOQUEUE;
840 		ret = dlm_lock_sync(bm_lockres, DLM_LOCK_PW);
841 		if (ret == -EAGAIN) {
842 			if (read_resync_info(mddev, bm_lockres)) {
843 				pr_info("%s:%d Resync[%llu..%llu] in progress on %d\n",
844 						__func__, __LINE__,
845 					(unsigned long long) cinfo->suspend_lo,
846 					(unsigned long long) cinfo->suspend_hi,
847 					i);
848 				cinfo->suspend_from = i;
849 			}
850 			ret = 0;
851 			lockres_free(bm_lockres);
852 			continue;
853 		}
854 		if (ret) {
855 			lockres_free(bm_lockres);
856 			goto out;
857 		}
858 
859 		/* Read the disk bitmap sb and check if it needs recovery */
860 		ret = mddev->bitmap_ops->copy_from_slot(mddev, i, &lo, &hi, false);
861 		if (ret) {
862 			pr_warn("md-cluster: Could not gather bitmaps from slot %d", i);
863 			lockres_free(bm_lockres);
864 			continue;
865 		}
866 		if ((hi > 0) && (lo < mddev->recovery_cp)) {
867 			set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
868 			mddev->recovery_cp = lo;
869 			md_check_recovery(mddev);
870 		}
871 
872 		lockres_free(bm_lockres);
873 	}
874 out:
875 	return ret;
876 }
877 
join(struct mddev * mddev,int nodes)878 static int join(struct mddev *mddev, int nodes)
879 {
880 	struct md_cluster_info *cinfo;
881 	int ret, ops_rv;
882 	char str[64];
883 
884 	cinfo = kzalloc(sizeof(struct md_cluster_info), GFP_KERNEL);
885 	if (!cinfo)
886 		return -ENOMEM;
887 
888 	INIT_LIST_HEAD(&cinfo->suspend_list);
889 	spin_lock_init(&cinfo->suspend_lock);
890 	init_completion(&cinfo->completion);
891 	set_bit(MD_CLUSTER_BEGIN_JOIN_CLUSTER, &cinfo->state);
892 	init_waitqueue_head(&cinfo->wait);
893 	mutex_init(&cinfo->recv_mutex);
894 
895 	mddev->cluster_info = cinfo;
896 	cinfo->mddev = mddev;
897 
898 	memset(str, 0, 64);
899 	sprintf(str, "%pU", mddev->uuid);
900 	ret = dlm_new_lockspace(str, mddev->bitmap_info.cluster_name,
901 				DLM_LSFL_SOFTIRQ, LVB_SIZE, &md_ls_ops, mddev,
902 				&ops_rv, &cinfo->lockspace);
903 	if (ret)
904 		goto err;
905 	wait_for_completion(&cinfo->completion);
906 	if (nodes < cinfo->slot_number) {
907 		pr_err("md-cluster: Slot allotted(%d) is greater than available slots(%d).",
908 			cinfo->slot_number, nodes);
909 		ret = -ERANGE;
910 		goto err;
911 	}
912 	/* Initiate the communication resources */
913 	ret = -ENOMEM;
914 	rcu_assign_pointer(cinfo->recv_thread,
915 			md_register_thread(recv_daemon, mddev, "cluster_recv"));
916 	if (!cinfo->recv_thread) {
917 		pr_err("md-cluster: cannot allocate memory for recv_thread!\n");
918 		goto err;
919 	}
920 	cinfo->message_lockres = lockres_init(mddev, "message", NULL, 1);
921 	if (!cinfo->message_lockres)
922 		goto err;
923 	cinfo->token_lockres = lockres_init(mddev, "token", NULL, 0);
924 	if (!cinfo->token_lockres)
925 		goto err;
926 	cinfo->no_new_dev_lockres = lockres_init(mddev, "no-new-dev", NULL, 0);
927 	if (!cinfo->no_new_dev_lockres)
928 		goto err;
929 
930 	ret = dlm_lock_sync(cinfo->token_lockres, DLM_LOCK_EX);
931 	if (ret) {
932 		ret = -EAGAIN;
933 		pr_err("md-cluster: can't join cluster to avoid lock issue\n");
934 		goto err;
935 	}
936 	cinfo->ack_lockres = lockres_init(mddev, "ack", ack_bast, 0);
937 	if (!cinfo->ack_lockres) {
938 		ret = -ENOMEM;
939 		goto err;
940 	}
941 	/* get sync CR lock on ACK. */
942 	if (dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_CR))
943 		pr_err("md-cluster: failed to get a sync CR lock on ACK!(%d)\n",
944 				ret);
945 	dlm_unlock_sync(cinfo->token_lockres);
946 	/* get sync CR lock on no-new-dev. */
947 	if (dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR))
948 		pr_err("md-cluster: failed to get a sync CR lock on no-new-dev!(%d)\n", ret);
949 
950 
951 	pr_info("md-cluster: Joined cluster %s slot %d\n", str, cinfo->slot_number);
952 	snprintf(str, 64, "bitmap%04d", cinfo->slot_number - 1);
953 	cinfo->bitmap_lockres = lockres_init(mddev, str, NULL, 1);
954 	if (!cinfo->bitmap_lockres) {
955 		ret = -ENOMEM;
956 		goto err;
957 	}
958 	if (dlm_lock_sync(cinfo->bitmap_lockres, DLM_LOCK_PW)) {
959 		pr_err("Failed to get bitmap lock\n");
960 		ret = -EINVAL;
961 		goto err;
962 	}
963 
964 	cinfo->resync_lockres = lockres_init(mddev, "resync", NULL, 0);
965 	if (!cinfo->resync_lockres) {
966 		ret = -ENOMEM;
967 		goto err;
968 	}
969 
970 	return 0;
971 err:
972 	set_bit(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state);
973 	md_unregister_thread(mddev, &cinfo->recovery_thread);
974 	md_unregister_thread(mddev, &cinfo->recv_thread);
975 	lockres_free(cinfo->message_lockres);
976 	lockres_free(cinfo->token_lockres);
977 	lockres_free(cinfo->ack_lockres);
978 	lockres_free(cinfo->no_new_dev_lockres);
979 	lockres_free(cinfo->resync_lockres);
980 	lockres_free(cinfo->bitmap_lockres);
981 	if (cinfo->lockspace)
982 		dlm_release_lockspace(cinfo->lockspace, 2);
983 	mddev->cluster_info = NULL;
984 	kfree(cinfo);
985 	return ret;
986 }
987 
load_bitmaps(struct mddev * mddev,int total_slots)988 static void load_bitmaps(struct mddev *mddev, int total_slots)
989 {
990 	struct md_cluster_info *cinfo = mddev->cluster_info;
991 
992 	/* load all the node's bitmap info for resync */
993 	if (gather_all_resync_info(mddev, total_slots))
994 		pr_err("md-cluster: failed to gather all resyn infos\n");
995 	set_bit(MD_CLUSTER_ALREADY_IN_CLUSTER, &cinfo->state);
996 	/* wake up recv thread in case something need to be handled */
997 	if (test_and_clear_bit(MD_CLUSTER_PENDING_RECV_EVENT, &cinfo->state))
998 		md_wakeup_thread(cinfo->recv_thread);
999 }
1000 
resync_bitmap(struct mddev * mddev)1001 static void resync_bitmap(struct mddev *mddev)
1002 {
1003 	struct md_cluster_info *cinfo = mddev->cluster_info;
1004 	struct cluster_msg cmsg = {0};
1005 	int err;
1006 
1007 	cmsg.type = cpu_to_le32(BITMAP_NEEDS_SYNC);
1008 	err = sendmsg(cinfo, &cmsg, 1);
1009 	if (err)
1010 		pr_err("%s:%d: failed to send BITMAP_NEEDS_SYNC message (%d)\n",
1011 			__func__, __LINE__, err);
1012 }
1013 
1014 static void unlock_all_bitmaps(struct mddev *mddev);
leave(struct mddev * mddev)1015 static int leave(struct mddev *mddev)
1016 {
1017 	struct md_cluster_info *cinfo = mddev->cluster_info;
1018 
1019 	if (!cinfo)
1020 		return 0;
1021 
1022 	/*
1023 	 * BITMAP_NEEDS_SYNC message should be sent when node
1024 	 * is leaving the cluster with dirty bitmap, also we
1025 	 * can only deliver it when dlm connection is available.
1026 	 *
1027 	 * Also, we should send BITMAP_NEEDS_SYNC message in
1028 	 * case reshaping is interrupted.
1029 	 */
1030 	if ((cinfo->slot_number > 0 && mddev->recovery_cp != MaxSector) ||
1031 	    (mddev->reshape_position != MaxSector &&
1032 	     test_bit(MD_CLOSING, &mddev->flags)))
1033 		resync_bitmap(mddev);
1034 
1035 	set_bit(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state);
1036 	md_unregister_thread(mddev, &cinfo->recovery_thread);
1037 	md_unregister_thread(mddev, &cinfo->recv_thread);
1038 	lockres_free(cinfo->message_lockres);
1039 	lockres_free(cinfo->token_lockres);
1040 	lockres_free(cinfo->ack_lockres);
1041 	lockres_free(cinfo->no_new_dev_lockres);
1042 	lockres_free(cinfo->resync_lockres);
1043 	lockres_free(cinfo->bitmap_lockres);
1044 	unlock_all_bitmaps(mddev);
1045 	dlm_release_lockspace(cinfo->lockspace, 2);
1046 	kfree(cinfo);
1047 	return 0;
1048 }
1049 
1050 /* slot_number(): Returns the MD slot number to use
1051  * DLM starts the slot numbers from 1, wheras cluster-md
1052  * wants the number to be from zero, so we deduct one
1053  */
slot_number(struct mddev * mddev)1054 static int slot_number(struct mddev *mddev)
1055 {
1056 	struct md_cluster_info *cinfo = mddev->cluster_info;
1057 
1058 	return cinfo->slot_number - 1;
1059 }
1060 
1061 /*
1062  * Check if the communication is already locked, else lock the communication
1063  * channel.
1064  * If it is already locked, token is in EX mode, and hence lock_token()
1065  * should not be called.
1066  */
metadata_update_start(struct mddev * mddev)1067 static int metadata_update_start(struct mddev *mddev)
1068 {
1069 	struct md_cluster_info *cinfo = mddev->cluster_info;
1070 	int ret;
1071 
1072 	/*
1073 	 * metadata_update_start is always called with the protection of
1074 	 * reconfig_mutex, so set WAITING_FOR_TOKEN here.
1075 	 */
1076 	ret = test_and_set_bit_lock(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD,
1077 				    &cinfo->state);
1078 	WARN_ON_ONCE(ret);
1079 	md_wakeup_thread(mddev->thread);
1080 
1081 	wait_event(cinfo->wait,
1082 		   !test_and_set_bit(MD_CLUSTER_SEND_LOCK, &cinfo->state) ||
1083 		   test_and_clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state));
1084 
1085 	/* If token is already locked, return 0 */
1086 	if (cinfo->token_lockres->mode == DLM_LOCK_EX) {
1087 		clear_bit_unlock(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state);
1088 		return 0;
1089 	}
1090 
1091 	ret = lock_token(cinfo);
1092 	clear_bit_unlock(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state);
1093 	return ret;
1094 }
1095 
metadata_update_finish(struct mddev * mddev)1096 static int metadata_update_finish(struct mddev *mddev)
1097 {
1098 	struct md_cluster_info *cinfo = mddev->cluster_info;
1099 	struct cluster_msg cmsg;
1100 	struct md_rdev *rdev;
1101 	int ret = 0;
1102 	int raid_slot = -1;
1103 
1104 	memset(&cmsg, 0, sizeof(cmsg));
1105 	cmsg.type = cpu_to_le32(METADATA_UPDATED);
1106 	/* Pick up a good active device number to send.
1107 	 */
1108 	rdev_for_each(rdev, mddev)
1109 		if (rdev->raid_disk > -1 && !test_bit(Faulty, &rdev->flags)) {
1110 			raid_slot = rdev->desc_nr;
1111 			break;
1112 		}
1113 	if (raid_slot >= 0) {
1114 		cmsg.raid_slot = cpu_to_le32(raid_slot);
1115 		ret = __sendmsg(cinfo, &cmsg);
1116 	} else
1117 		pr_warn("md-cluster: No good device id found to send\n");
1118 	clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1119 	unlock_comm(cinfo);
1120 	return ret;
1121 }
1122 
metadata_update_cancel(struct mddev * mddev)1123 static void metadata_update_cancel(struct mddev *mddev)
1124 {
1125 	struct md_cluster_info *cinfo = mddev->cluster_info;
1126 	clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1127 	unlock_comm(cinfo);
1128 }
1129 
update_bitmap_size(struct mddev * mddev,sector_t size)1130 static int update_bitmap_size(struct mddev *mddev, sector_t size)
1131 {
1132 	struct md_cluster_info *cinfo = mddev->cluster_info;
1133 	struct cluster_msg cmsg = {0};
1134 	int ret;
1135 
1136 	cmsg.type = cpu_to_le32(BITMAP_RESIZE);
1137 	cmsg.high = cpu_to_le64(size);
1138 	ret = sendmsg(cinfo, &cmsg, 0);
1139 	if (ret)
1140 		pr_err("%s:%d: failed to send BITMAP_RESIZE message (%d)\n",
1141 			__func__, __LINE__, ret);
1142 	return ret;
1143 }
1144 
resize_bitmaps(struct mddev * mddev,sector_t newsize,sector_t oldsize)1145 static int resize_bitmaps(struct mddev *mddev, sector_t newsize, sector_t oldsize)
1146 {
1147 	void *bitmap = mddev->bitmap;
1148 	struct md_bitmap_stats stats;
1149 	unsigned long my_pages;
1150 	int i, rv;
1151 
1152 	rv = mddev->bitmap_ops->get_stats(bitmap, &stats);
1153 	if (rv)
1154 		return rv;
1155 
1156 	my_pages = stats.pages;
1157 	/*
1158 	 * We need to ensure all the nodes can grow to a larger
1159 	 * bitmap size before make the reshaping.
1160 	 */
1161 	rv = update_bitmap_size(mddev, newsize);
1162 	if (rv)
1163 		return rv;
1164 
1165 	for (i = 0; i < mddev->bitmap_info.nodes; i++) {
1166 		struct dlm_lock_resource *bm_lockres;
1167 		char str[64];
1168 
1169 		if (i == md_cluster_ops->slot_number(mddev))
1170 			continue;
1171 
1172 		bitmap = mddev->bitmap_ops->get_from_slot(mddev, i);
1173 		if (IS_ERR(bitmap)) {
1174 			pr_err("can't get bitmap from slot %d\n", i);
1175 			bitmap = NULL;
1176 			goto out;
1177 		}
1178 
1179 		rv = mddev->bitmap_ops->get_stats(bitmap, &stats);
1180 		if (rv)
1181 			goto out;
1182 		/*
1183 		 * If we can hold the bitmap lock of one node then
1184 		 * the slot is not occupied, update the pages.
1185 		 */
1186 		snprintf(str, 64, "bitmap%04d", i);
1187 		bm_lockres = lockres_init(mddev, str, NULL, 1);
1188 		if (!bm_lockres) {
1189 			pr_err("Cannot initialize %s lock\n", str);
1190 			goto out;
1191 		}
1192 		bm_lockres->flags |= DLM_LKF_NOQUEUE;
1193 		rv = dlm_lock_sync(bm_lockres, DLM_LOCK_PW);
1194 		if (!rv)
1195 			mddev->bitmap_ops->set_pages(bitmap, my_pages);
1196 		lockres_free(bm_lockres);
1197 
1198 		if (my_pages != stats.pages)
1199 			/*
1200 			 * Let's revert the bitmap size if one node
1201 			 * can't resize bitmap
1202 			 */
1203 			goto out;
1204 		mddev->bitmap_ops->free(bitmap);
1205 	}
1206 
1207 	return 0;
1208 out:
1209 	mddev->bitmap_ops->free(bitmap);
1210 	update_bitmap_size(mddev, oldsize);
1211 	return -1;
1212 }
1213 
1214 /*
1215  * return 0 if all the bitmaps have the same sync_size
1216  */
cluster_check_sync_size(struct mddev * mddev)1217 static int cluster_check_sync_size(struct mddev *mddev)
1218 {
1219 	int current_slot = md_cluster_ops->slot_number(mddev);
1220 	int node_num = mddev->bitmap_info.nodes;
1221 	struct dlm_lock_resource *bm_lockres;
1222 	struct md_bitmap_stats stats;
1223 	void *bitmap = mddev->bitmap;
1224 	unsigned long sync_size = 0;
1225 	unsigned long my_sync_size;
1226 	char str[64];
1227 	int i, rv;
1228 
1229 	rv = mddev->bitmap_ops->get_stats(bitmap, &stats);
1230 	if (rv)
1231 		return rv;
1232 
1233 	my_sync_size = stats.sync_size;
1234 
1235 	for (i = 0; i < node_num; i++) {
1236 		if (i == current_slot)
1237 			continue;
1238 
1239 		bitmap = mddev->bitmap_ops->get_from_slot(mddev, i);
1240 		if (IS_ERR(bitmap)) {
1241 			pr_err("can't get bitmap from slot %d\n", i);
1242 			return -1;
1243 		}
1244 
1245 		/*
1246 		 * If we can hold the bitmap lock of one node then
1247 		 * the slot is not occupied, update the sb.
1248 		 */
1249 		snprintf(str, 64, "bitmap%04d", i);
1250 		bm_lockres = lockres_init(mddev, str, NULL, 1);
1251 		if (!bm_lockres) {
1252 			pr_err("md-cluster: Cannot initialize %s\n", str);
1253 			mddev->bitmap_ops->free(bitmap);
1254 			return -1;
1255 		}
1256 		bm_lockres->flags |= DLM_LKF_NOQUEUE;
1257 		rv = dlm_lock_sync(bm_lockres, DLM_LOCK_PW);
1258 		if (!rv)
1259 			mddev->bitmap_ops->update_sb(bitmap);
1260 		lockres_free(bm_lockres);
1261 
1262 		rv = mddev->bitmap_ops->get_stats(bitmap, &stats);
1263 		if (rv) {
1264 			mddev->bitmap_ops->free(bitmap);
1265 			return rv;
1266 		}
1267 
1268 		if (sync_size == 0) {
1269 			sync_size = stats.sync_size;
1270 		} else if (sync_size != stats.sync_size) {
1271 			mddev->bitmap_ops->free(bitmap);
1272 			return -1;
1273 		}
1274 		mddev->bitmap_ops->free(bitmap);
1275 	}
1276 
1277 	return (my_sync_size == sync_size) ? 0 : -1;
1278 }
1279 
1280 /*
1281  * Update the size for cluster raid is a little more complex, we perform it
1282  * by the steps:
1283  * 1. hold token lock and update superblock in initiator node.
1284  * 2. send METADATA_UPDATED msg to other nodes.
1285  * 3. The initiator node continues to check each bitmap's sync_size, if all
1286  *    bitmaps have the same value of sync_size, then we can set capacity and
1287  *    let other nodes to perform it. If one node can't update sync_size
1288  *    accordingly, we need to revert to previous value.
1289  */
update_size(struct mddev * mddev,sector_t old_dev_sectors)1290 static void update_size(struct mddev *mddev, sector_t old_dev_sectors)
1291 {
1292 	struct md_cluster_info *cinfo = mddev->cluster_info;
1293 	struct cluster_msg cmsg;
1294 	struct md_rdev *rdev;
1295 	int ret = 0;
1296 	int raid_slot = -1;
1297 
1298 	md_update_sb(mddev, 1);
1299 	if (lock_comm(cinfo, 1)) {
1300 		pr_err("%s: lock_comm failed\n", __func__);
1301 		return;
1302 	}
1303 
1304 	memset(&cmsg, 0, sizeof(cmsg));
1305 	cmsg.type = cpu_to_le32(METADATA_UPDATED);
1306 	rdev_for_each(rdev, mddev)
1307 		if (rdev->raid_disk >= 0 && !test_bit(Faulty, &rdev->flags)) {
1308 			raid_slot = rdev->desc_nr;
1309 			break;
1310 		}
1311 	if (raid_slot >= 0) {
1312 		cmsg.raid_slot = cpu_to_le32(raid_slot);
1313 		/*
1314 		 * We can only change capiticy after all the nodes can do it,
1315 		 * so need to wait after other nodes already received the msg
1316 		 * and handled the change
1317 		 */
1318 		ret = __sendmsg(cinfo, &cmsg);
1319 		if (ret) {
1320 			pr_err("%s:%d: failed to send METADATA_UPDATED msg\n",
1321 			       __func__, __LINE__);
1322 			unlock_comm(cinfo);
1323 			return;
1324 		}
1325 	} else {
1326 		pr_err("md-cluster: No good device id found to send\n");
1327 		unlock_comm(cinfo);
1328 		return;
1329 	}
1330 
1331 	/*
1332 	 * check the sync_size from other node's bitmap, if sync_size
1333 	 * have already updated in other nodes as expected, send an
1334 	 * empty metadata msg to permit the change of capacity
1335 	 */
1336 	if (cluster_check_sync_size(mddev) == 0) {
1337 		memset(&cmsg, 0, sizeof(cmsg));
1338 		cmsg.type = cpu_to_le32(CHANGE_CAPACITY);
1339 		ret = __sendmsg(cinfo, &cmsg);
1340 		if (ret)
1341 			pr_err("%s:%d: failed to send CHANGE_CAPACITY msg\n",
1342 			       __func__, __LINE__);
1343 		set_capacity_and_notify(mddev->gendisk, mddev->array_sectors);
1344 	} else {
1345 		/* revert to previous sectors */
1346 		ret = mddev->pers->resize(mddev, old_dev_sectors);
1347 		ret = __sendmsg(cinfo, &cmsg);
1348 		if (ret)
1349 			pr_err("%s:%d: failed to send METADATA_UPDATED msg\n",
1350 			       __func__, __LINE__);
1351 	}
1352 	unlock_comm(cinfo);
1353 }
1354 
resync_start(struct mddev * mddev)1355 static int resync_start(struct mddev *mddev)
1356 {
1357 	struct md_cluster_info *cinfo = mddev->cluster_info;
1358 	return dlm_lock_sync_interruptible(cinfo->resync_lockres, DLM_LOCK_EX, mddev);
1359 }
1360 
resync_info_get(struct mddev * mddev,sector_t * lo,sector_t * hi)1361 static void resync_info_get(struct mddev *mddev, sector_t *lo, sector_t *hi)
1362 {
1363 	struct md_cluster_info *cinfo = mddev->cluster_info;
1364 
1365 	spin_lock_irq(&cinfo->suspend_lock);
1366 	*lo = cinfo->suspend_lo;
1367 	*hi = cinfo->suspend_hi;
1368 	spin_unlock_irq(&cinfo->suspend_lock);
1369 }
1370 
resync_status_get(struct mddev * mddev)1371 static int resync_status_get(struct mddev *mddev)
1372 {
1373 	struct md_cluster_info *cinfo = mddev->cluster_info;
1374 
1375 	return test_bit(MD_CLUSTER_WAITING_FOR_SYNC, &cinfo->state);
1376 }
1377 
resync_start_notify(struct mddev * mddev)1378 static int resync_start_notify(struct mddev *mddev)
1379 {
1380 	struct md_cluster_info *cinfo = mddev->cluster_info;
1381 	struct cluster_msg cmsg = {0};
1382 
1383 	cmsg.type = cpu_to_le32(RESYNCING_START);
1384 
1385 	return sendmsg(cinfo, &cmsg, 0);
1386 }
1387 
resync_info_update(struct mddev * mddev,sector_t lo,sector_t hi)1388 static int resync_info_update(struct mddev *mddev, sector_t lo, sector_t hi)
1389 {
1390 	struct md_cluster_info *cinfo = mddev->cluster_info;
1391 	struct resync_info ri;
1392 	struct cluster_msg cmsg = {0};
1393 
1394 	/* do not send zero again, if we have sent before */
1395 	if (hi == 0) {
1396 		memcpy(&ri, cinfo->bitmap_lockres->lksb.sb_lvbptr, sizeof(struct resync_info));
1397 		if (le64_to_cpu(ri.hi) == 0)
1398 			return 0;
1399 	}
1400 
1401 	add_resync_info(cinfo->bitmap_lockres, lo, hi);
1402 	/* Re-acquire the lock to refresh LVB */
1403 	dlm_lock_sync(cinfo->bitmap_lockres, DLM_LOCK_PW);
1404 	cmsg.type = cpu_to_le32(RESYNCING);
1405 	cmsg.low = cpu_to_le64(lo);
1406 	cmsg.high = cpu_to_le64(hi);
1407 
1408 	/*
1409 	 * mddev_lock is held if resync_info_update is called from
1410 	 * resync_finish (md_reap_sync_thread -> resync_finish)
1411 	 */
1412 	if (lo == 0 && hi == 0)
1413 		return sendmsg(cinfo, &cmsg, 1);
1414 	else
1415 		return sendmsg(cinfo, &cmsg, 0);
1416 }
1417 
resync_finish(struct mddev * mddev)1418 static int resync_finish(struct mddev *mddev)
1419 {
1420 	struct md_cluster_info *cinfo = mddev->cluster_info;
1421 	int ret = 0;
1422 
1423 	clear_bit(MD_RESYNCING_REMOTE, &mddev->recovery);
1424 
1425 	/*
1426 	 * If resync thread is interrupted so we can't say resync is finished,
1427 	 * another node will launch resync thread to continue.
1428 	 */
1429 	if (!test_bit(MD_CLOSING, &mddev->flags))
1430 		ret = resync_info_update(mddev, 0, 0);
1431 	dlm_unlock_sync(cinfo->resync_lockres);
1432 	return ret;
1433 }
1434 
area_resyncing(struct mddev * mddev,int direction,sector_t lo,sector_t hi)1435 static int area_resyncing(struct mddev *mddev, int direction,
1436 		sector_t lo, sector_t hi)
1437 {
1438 	struct md_cluster_info *cinfo = mddev->cluster_info;
1439 	int ret = 0;
1440 
1441 	if ((direction == READ) &&
1442 		test_bit(MD_CLUSTER_SUSPEND_READ_BALANCING, &cinfo->state))
1443 		return 1;
1444 
1445 	spin_lock_irq(&cinfo->suspend_lock);
1446 	if (hi > cinfo->suspend_lo && lo < cinfo->suspend_hi)
1447 		ret = 1;
1448 	spin_unlock_irq(&cinfo->suspend_lock);
1449 	return ret;
1450 }
1451 
1452 /* add_new_disk() - initiates a disk add
1453  * However, if this fails before writing md_update_sb(),
1454  * add_new_disk_cancel() must be called to release token lock
1455  */
add_new_disk(struct mddev * mddev,struct md_rdev * rdev)1456 static int add_new_disk(struct mddev *mddev, struct md_rdev *rdev)
1457 {
1458 	struct md_cluster_info *cinfo = mddev->cluster_info;
1459 	struct cluster_msg cmsg;
1460 	int ret = 0;
1461 	struct mdp_superblock_1 *sb = page_address(rdev->sb_page);
1462 	char *uuid = sb->device_uuid;
1463 
1464 	memset(&cmsg, 0, sizeof(cmsg));
1465 	cmsg.type = cpu_to_le32(NEWDISK);
1466 	memcpy(cmsg.uuid, uuid, 16);
1467 	cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
1468 	if (lock_comm(cinfo, 1))
1469 		return -EAGAIN;
1470 	ret = __sendmsg(cinfo, &cmsg);
1471 	if (ret) {
1472 		unlock_comm(cinfo);
1473 		return ret;
1474 	}
1475 	cinfo->no_new_dev_lockres->flags |= DLM_LKF_NOQUEUE;
1476 	ret = dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_EX);
1477 	cinfo->no_new_dev_lockres->flags &= ~DLM_LKF_NOQUEUE;
1478 	/* Some node does not "see" the device */
1479 	if (ret == -EAGAIN)
1480 		ret = -ENOENT;
1481 	if (ret)
1482 		unlock_comm(cinfo);
1483 	else {
1484 		dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR);
1485 		/* Since MD_CHANGE_DEVS will be set in add_bound_rdev which
1486 		 * will run soon after add_new_disk, the below path will be
1487 		 * invoked:
1488 		 *   md_wakeup_thread(mddev->thread)
1489 		 *	-> conf->thread (raid1d)
1490 		 *	-> md_check_recovery -> md_update_sb
1491 		 *	-> metadata_update_start/finish
1492 		 * MD_CLUSTER_SEND_LOCKED_ALREADY will be cleared eventually.
1493 		 *
1494 		 * For other failure cases, metadata_update_cancel and
1495 		 * add_new_disk_cancel also clear below bit as well.
1496 		 * */
1497 		set_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1498 		wake_up(&cinfo->wait);
1499 	}
1500 	return ret;
1501 }
1502 
add_new_disk_cancel(struct mddev * mddev)1503 static void add_new_disk_cancel(struct mddev *mddev)
1504 {
1505 	struct md_cluster_info *cinfo = mddev->cluster_info;
1506 	clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1507 	unlock_comm(cinfo);
1508 }
1509 
new_disk_ack(struct mddev * mddev,bool ack)1510 static int new_disk_ack(struct mddev *mddev, bool ack)
1511 {
1512 	struct md_cluster_info *cinfo = mddev->cluster_info;
1513 
1514 	if (!test_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state)) {
1515 		pr_warn("md-cluster(%s): Spurious cluster confirmation\n", mdname(mddev));
1516 		return -EINVAL;
1517 	}
1518 
1519 	if (ack)
1520 		dlm_unlock_sync(cinfo->no_new_dev_lockres);
1521 	complete(&cinfo->newdisk_completion);
1522 	return 0;
1523 }
1524 
remove_disk(struct mddev * mddev,struct md_rdev * rdev)1525 static int remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1526 {
1527 	struct cluster_msg cmsg = {0};
1528 	struct md_cluster_info *cinfo = mddev->cluster_info;
1529 	cmsg.type = cpu_to_le32(REMOVE);
1530 	cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
1531 	return sendmsg(cinfo, &cmsg, 1);
1532 }
1533 
lock_all_bitmaps(struct mddev * mddev)1534 static int lock_all_bitmaps(struct mddev *mddev)
1535 {
1536 	int slot, my_slot, ret, held = 1, i = 0;
1537 	char str[64];
1538 	struct md_cluster_info *cinfo = mddev->cluster_info;
1539 
1540 	cinfo->other_bitmap_lockres =
1541 		kcalloc(mddev->bitmap_info.nodes - 1,
1542 			sizeof(struct dlm_lock_resource *), GFP_KERNEL);
1543 	if (!cinfo->other_bitmap_lockres) {
1544 		pr_err("md: can't alloc mem for other bitmap locks\n");
1545 		return 0;
1546 	}
1547 
1548 	my_slot = slot_number(mddev);
1549 	for (slot = 0; slot < mddev->bitmap_info.nodes; slot++) {
1550 		if (slot == my_slot)
1551 			continue;
1552 
1553 		memset(str, '\0', 64);
1554 		snprintf(str, 64, "bitmap%04d", slot);
1555 		cinfo->other_bitmap_lockres[i] = lockres_init(mddev, str, NULL, 1);
1556 		if (!cinfo->other_bitmap_lockres[i])
1557 			return -ENOMEM;
1558 
1559 		cinfo->other_bitmap_lockres[i]->flags |= DLM_LKF_NOQUEUE;
1560 		ret = dlm_lock_sync(cinfo->other_bitmap_lockres[i], DLM_LOCK_PW);
1561 		if (ret)
1562 			held = -1;
1563 		i++;
1564 	}
1565 
1566 	return held;
1567 }
1568 
unlock_all_bitmaps(struct mddev * mddev)1569 static void unlock_all_bitmaps(struct mddev *mddev)
1570 {
1571 	struct md_cluster_info *cinfo = mddev->cluster_info;
1572 	int i;
1573 
1574 	/* release other node's bitmap lock if they are existed */
1575 	if (cinfo->other_bitmap_lockres) {
1576 		for (i = 0; i < mddev->bitmap_info.nodes - 1; i++) {
1577 			if (cinfo->other_bitmap_lockres[i]) {
1578 				lockres_free(cinfo->other_bitmap_lockres[i]);
1579 			}
1580 		}
1581 		kfree(cinfo->other_bitmap_lockres);
1582 		cinfo->other_bitmap_lockres = NULL;
1583 	}
1584 }
1585 
gather_bitmaps(struct md_rdev * rdev)1586 static int gather_bitmaps(struct md_rdev *rdev)
1587 {
1588 	int sn, err;
1589 	sector_t lo, hi;
1590 	struct cluster_msg cmsg = {0};
1591 	struct mddev *mddev = rdev->mddev;
1592 	struct md_cluster_info *cinfo = mddev->cluster_info;
1593 
1594 	cmsg.type = cpu_to_le32(RE_ADD);
1595 	cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
1596 	err = sendmsg(cinfo, &cmsg, 1);
1597 	if (err)
1598 		goto out;
1599 
1600 	for (sn = 0; sn < mddev->bitmap_info.nodes; sn++) {
1601 		if (sn == (cinfo->slot_number - 1))
1602 			continue;
1603 		err = mddev->bitmap_ops->copy_from_slot(mddev, sn, &lo, &hi, false);
1604 		if (err) {
1605 			pr_warn("md-cluster: Could not gather bitmaps from slot %d", sn);
1606 			goto out;
1607 		}
1608 		if ((hi > 0) && (lo < mddev->recovery_cp))
1609 			mddev->recovery_cp = lo;
1610 	}
1611 out:
1612 	return err;
1613 }
1614 
1615 static const struct md_cluster_operations cluster_ops = {
1616 	.join   = join,
1617 	.leave  = leave,
1618 	.slot_number = slot_number,
1619 	.resync_start = resync_start,
1620 	.resync_finish = resync_finish,
1621 	.resync_info_update = resync_info_update,
1622 	.resync_start_notify = resync_start_notify,
1623 	.resync_status_get = resync_status_get,
1624 	.resync_info_get = resync_info_get,
1625 	.metadata_update_start = metadata_update_start,
1626 	.metadata_update_finish = metadata_update_finish,
1627 	.metadata_update_cancel = metadata_update_cancel,
1628 	.area_resyncing = area_resyncing,
1629 	.add_new_disk = add_new_disk,
1630 	.add_new_disk_cancel = add_new_disk_cancel,
1631 	.new_disk_ack = new_disk_ack,
1632 	.remove_disk = remove_disk,
1633 	.load_bitmaps = load_bitmaps,
1634 	.gather_bitmaps = gather_bitmaps,
1635 	.resize_bitmaps = resize_bitmaps,
1636 	.lock_all_bitmaps = lock_all_bitmaps,
1637 	.unlock_all_bitmaps = unlock_all_bitmaps,
1638 	.update_size = update_size,
1639 };
1640 
cluster_init(void)1641 static int __init cluster_init(void)
1642 {
1643 	pr_warn("md-cluster: support raid1 and raid10 (limited support)\n");
1644 	pr_info("Registering Cluster MD functions\n");
1645 	register_md_cluster_operations(&cluster_ops, THIS_MODULE);
1646 	return 0;
1647 }
1648 
cluster_exit(void)1649 static void cluster_exit(void)
1650 {
1651 	unregister_md_cluster_operations();
1652 }
1653 
1654 module_init(cluster_init);
1655 module_exit(cluster_exit);
1656 MODULE_AUTHOR("SUSE");
1657 MODULE_LICENSE("GPL");
1658 MODULE_DESCRIPTION("Clustering support for MD");
1659