Linux Audio

Check our new training course

Loading...
v5.9
   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");
v4.10.11
 
   1/*
   2 * Copyright (C) 2015, SUSE
   3 *
   4 * This program is free software; you can redistribute it and/or modify
   5 * it under the terms of the GNU General Public License as published by
   6 * the Free Software Foundation; either version 2, or (at your option)
   7 * any later version.
   8 *
   9 */
  10
  11
  12#include <linux/module.h>
  13#include <linux/kthread.h>
  14#include <linux/dlm.h>
  15#include <linux/sched.h>
  16#include <linux/raid/md_p.h>
  17#include "md.h"
  18#include "bitmap.h"
  19#include "md-cluster.h"
  20
  21#define LVB_SIZE	64
  22#define NEW_DEV_TIMEOUT 5000
  23
  24struct dlm_lock_resource {
  25	dlm_lockspace_t *ls;
  26	struct dlm_lksb lksb;
  27	char *name; /* lock name. */
  28	uint32_t flags; /* flags to pass to dlm_lock() */
  29	wait_queue_head_t sync_locking; /* wait queue for synchronized locking */
  30	bool sync_locking_done;
  31	void (*bast)(void *arg, int mode); /* blocking AST function pointer*/
  32	struct mddev *mddev; /* pointing back to mddev. */
  33	int mode;
  34};
  35
  36struct suspend_info {
  37	int slot;
  38	sector_t lo;
  39	sector_t hi;
  40	struct list_head list;
  41};
  42
  43struct resync_info {
  44	__le64 lo;
  45	__le64 hi;
  46};
  47
  48/* md_cluster_info flags */
  49#define		MD_CLUSTER_WAITING_FOR_NEWDISK		1
  50#define		MD_CLUSTER_SUSPEND_READ_BALANCING	2
  51#define		MD_CLUSTER_BEGIN_JOIN_CLUSTER		3
  52
  53/* Lock the send communication. This is done through
  54 * bit manipulation as opposed to a mutex in order to
  55 * accomodate lock and hold. See next comment.
  56 */
  57#define		MD_CLUSTER_SEND_LOCK			4
  58/* If cluster operations (such as adding a disk) must lock the
  59 * communication channel, so as to perform extra operations
  60 * (update metadata) and no other operation is allowed on the
  61 * MD. Token needs to be locked and held until the operation
  62 * completes witha md_update_sb(), which would eventually release
  63 * the lock.
  64 */
  65#define		MD_CLUSTER_SEND_LOCKED_ALREADY		5
  66/* We should receive message after node joined cluster and
  67 * set up all the related infos such as bitmap and personality */
  68#define		MD_CLUSTER_ALREADY_IN_CLUSTER		6
  69#define		MD_CLUSTER_PENDING_RECV_EVENT		7
  70
  71
  72struct md_cluster_info {
 
  73	/* dlm lock space and resources for clustered raid. */
  74	dlm_lockspace_t *lockspace;
  75	int slot_number;
  76	struct completion completion;
  77	struct mutex recv_mutex;
  78	struct dlm_lock_resource *bitmap_lockres;
  79	struct dlm_lock_resource **other_bitmap_lockres;
  80	struct dlm_lock_resource *resync_lockres;
  81	struct list_head suspend_list;
 
  82	spinlock_t suspend_lock;
 
 
 
 
 
  83	struct md_thread *recovery_thread;
  84	unsigned long recovery_map;
  85	/* communication loc resources */
  86	struct dlm_lock_resource *ack_lockres;
  87	struct dlm_lock_resource *message_lockres;
  88	struct dlm_lock_resource *token_lockres;
  89	struct dlm_lock_resource *no_new_dev_lockres;
  90	struct md_thread *recv_thread;
  91	struct completion newdisk_completion;
  92	wait_queue_head_t wait;
  93	unsigned long state;
  94	/* record the region in RESYNCING message */
  95	sector_t sync_low;
  96	sector_t sync_hi;
  97};
  98
  99enum msg_type {
 100	METADATA_UPDATED = 0,
 101	RESYNCING,
 102	NEWDISK,
 103	REMOVE,
 104	RE_ADD,
 105	BITMAP_NEEDS_SYNC,
 
 
 106};
 107
 108struct cluster_msg {
 109	__le32 type;
 110	__le32 slot;
 111	/* TODO: Unionize this for smaller footprint */
 112	__le64 low;
 113	__le64 high;
 114	char uuid[16];
 115	__le32 raid_slot;
 116};
 117
 118static void sync_ast(void *arg)
 119{
 120	struct dlm_lock_resource *res;
 121
 122	res = arg;
 123	res->sync_locking_done = true;
 124	wake_up(&res->sync_locking);
 125}
 126
 127static int dlm_lock_sync(struct dlm_lock_resource *res, int mode)
 128{
 129	int ret = 0;
 130
 131	ret = dlm_lock(res->ls, mode, &res->lksb,
 132			res->flags, res->name, strlen(res->name),
 133			0, sync_ast, res, res->bast);
 134	if (ret)
 135		return ret;
 136	wait_event(res->sync_locking, res->sync_locking_done);
 137	res->sync_locking_done = false;
 138	if (res->lksb.sb_status == 0)
 139		res->mode = mode;
 140	return res->lksb.sb_status;
 141}
 142
 143static int dlm_unlock_sync(struct dlm_lock_resource *res)
 144{
 145	return dlm_lock_sync(res, DLM_LOCK_NL);
 146}
 147
 148/*
 149 * An variation of dlm_lock_sync, which make lock request could
 150 * be interrupted
 151 */
 152static int dlm_lock_sync_interruptible(struct dlm_lock_resource *res, int mode,
 153				       struct mddev *mddev)
 154{
 155	int ret = 0;
 156
 157	ret = dlm_lock(res->ls, mode, &res->lksb,
 158			res->flags, res->name, strlen(res->name),
 159			0, sync_ast, res, res->bast);
 160	if (ret)
 161		return ret;
 162
 163	wait_event(res->sync_locking, res->sync_locking_done
 164				      || kthread_should_stop()
 165				      || test_bit(MD_CLOSING, &mddev->flags));
 166	if (!res->sync_locking_done) {
 167		/*
 168		 * the convert queue contains the lock request when request is
 169		 * interrupted, and sync_ast could still be run, so need to
 170		 * cancel the request and reset completion
 171		 */
 172		ret = dlm_unlock(res->ls, res->lksb.sb_lkid, DLM_LKF_CANCEL,
 173			&res->lksb, res);
 174		res->sync_locking_done = false;
 175		if (unlikely(ret != 0))
 176			pr_info("failed to cancel previous lock request "
 177				 "%s return %d\n", res->name, ret);
 178		return -EPERM;
 179	} else
 180		res->sync_locking_done = false;
 181	if (res->lksb.sb_status == 0)
 182		res->mode = mode;
 183	return res->lksb.sb_status;
 184}
 185
 186static struct dlm_lock_resource *lockres_init(struct mddev *mddev,
 187		char *name, void (*bastfn)(void *arg, int mode), int with_lvb)
 188{
 189	struct dlm_lock_resource *res = NULL;
 190	int ret, namelen;
 191	struct md_cluster_info *cinfo = mddev->cluster_info;
 192
 193	res = kzalloc(sizeof(struct dlm_lock_resource), GFP_KERNEL);
 194	if (!res)
 195		return NULL;
 196	init_waitqueue_head(&res->sync_locking);
 197	res->sync_locking_done = false;
 198	res->ls = cinfo->lockspace;
 199	res->mddev = mddev;
 200	res->mode = DLM_LOCK_IV;
 201	namelen = strlen(name);
 202	res->name = kzalloc(namelen + 1, GFP_KERNEL);
 203	if (!res->name) {
 204		pr_err("md-cluster: Unable to allocate resource name for resource %s\n", name);
 205		goto out_err;
 206	}
 207	strlcpy(res->name, name, namelen + 1);
 208	if (with_lvb) {
 209		res->lksb.sb_lvbptr = kzalloc(LVB_SIZE, GFP_KERNEL);
 210		if (!res->lksb.sb_lvbptr) {
 211			pr_err("md-cluster: Unable to allocate LVB for resource %s\n", name);
 212			goto out_err;
 213		}
 214		res->flags = DLM_LKF_VALBLK;
 215	}
 216
 217	if (bastfn)
 218		res->bast = bastfn;
 219
 220	res->flags |= DLM_LKF_EXPEDITE;
 221
 222	ret = dlm_lock_sync(res, DLM_LOCK_NL);
 223	if (ret) {
 224		pr_err("md-cluster: Unable to lock NL on new lock resource %s\n", name);
 225		goto out_err;
 226	}
 227	res->flags &= ~DLM_LKF_EXPEDITE;
 228	res->flags |= DLM_LKF_CONVERT;
 229
 230	return res;
 231out_err:
 232	kfree(res->lksb.sb_lvbptr);
 233	kfree(res->name);
 234	kfree(res);
 235	return NULL;
 236}
 237
 238static void lockres_free(struct dlm_lock_resource *res)
 239{
 240	int ret = 0;
 241
 242	if (!res)
 243		return;
 244
 245	/*
 246	 * use FORCEUNLOCK flag, so we can unlock even the lock is on the
 247	 * waiting or convert queue
 248	 */
 249	ret = dlm_unlock(res->ls, res->lksb.sb_lkid, DLM_LKF_FORCEUNLOCK,
 250		&res->lksb, res);
 251	if (unlikely(ret != 0))
 252		pr_err("failed to unlock %s return %d\n", res->name, ret);
 253	else
 254		wait_event(res->sync_locking, res->sync_locking_done);
 255
 256	kfree(res->name);
 257	kfree(res->lksb.sb_lvbptr);
 258	kfree(res);
 259}
 260
 261static void add_resync_info(struct dlm_lock_resource *lockres,
 262			    sector_t lo, sector_t hi)
 263{
 264	struct resync_info *ri;
 265
 266	ri = (struct resync_info *)lockres->lksb.sb_lvbptr;
 267	ri->lo = cpu_to_le64(lo);
 268	ri->hi = cpu_to_le64(hi);
 269}
 270
 271static struct suspend_info *read_resync_info(struct mddev *mddev, struct dlm_lock_resource *lockres)
 
 272{
 273	struct resync_info ri;
 274	struct suspend_info *s = NULL;
 275	sector_t hi = 0;
 276
 277	dlm_lock_sync(lockres, DLM_LOCK_CR);
 278	memcpy(&ri, lockres->lksb.sb_lvbptr, sizeof(struct resync_info));
 279	hi = le64_to_cpu(ri.hi);
 280	if (hi > 0) {
 281		s = kzalloc(sizeof(struct suspend_info), GFP_KERNEL);
 282		if (!s)
 283			goto out;
 284		s->hi = hi;
 285		s->lo = le64_to_cpu(ri.lo);
 286	}
 287	dlm_unlock_sync(lockres);
 288out:
 289	return s;
 290}
 291
 292static void recover_bitmaps(struct md_thread *thread)
 293{
 294	struct mddev *mddev = thread->mddev;
 295	struct md_cluster_info *cinfo = mddev->cluster_info;
 296	struct dlm_lock_resource *bm_lockres;
 297	char str[64];
 298	int slot, ret;
 299	struct suspend_info *s, *tmp;
 300	sector_t lo, hi;
 301
 302	while (cinfo->recovery_map) {
 303		slot = fls64((u64)cinfo->recovery_map) - 1;
 304
 305		/* Clear suspend_area associated with the bitmap */
 306		spin_lock_irq(&cinfo->suspend_lock);
 307		list_for_each_entry_safe(s, tmp, &cinfo->suspend_list, list)
 308			if (slot == s->slot) {
 309				list_del(&s->list);
 310				kfree(s);
 311			}
 312		spin_unlock_irq(&cinfo->suspend_lock);
 313
 314		snprintf(str, 64, "bitmap%04d", slot);
 315		bm_lockres = lockres_init(mddev, str, NULL, 1);
 316		if (!bm_lockres) {
 317			pr_err("md-cluster: Cannot initialize bitmaps\n");
 318			goto clear_bit;
 319		}
 320
 321		ret = dlm_lock_sync_interruptible(bm_lockres, DLM_LOCK_PW, mddev);
 322		if (ret) {
 323			pr_err("md-cluster: Could not DLM lock %s: %d\n",
 324					str, ret);
 325			goto clear_bit;
 326		}
 327		ret = bitmap_copy_from_slot(mddev, slot, &lo, &hi, true);
 328		if (ret) {
 329			pr_err("md-cluster: Could not copy data from bitmap %d\n", slot);
 330			goto clear_bit;
 331		}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 332		if (hi > 0) {
 333			if (lo < mddev->recovery_cp)
 334				mddev->recovery_cp = lo;
 335			/* wake up thread to continue resync in case resync
 336			 * is not finished */
 337			if (mddev->recovery_cp != MaxSector) {
 338			    set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
 339			    md_wakeup_thread(mddev->thread);
 
 
 
 
 
 
 340			}
 341		}
 342clear_bit:
 343		lockres_free(bm_lockres);
 344		clear_bit(slot, &cinfo->recovery_map);
 345	}
 346}
 347
 348static void recover_prep(void *arg)
 349{
 350	struct mddev *mddev = arg;
 351	struct md_cluster_info *cinfo = mddev->cluster_info;
 352	set_bit(MD_CLUSTER_SUSPEND_READ_BALANCING, &cinfo->state);
 353}
 354
 355static void __recover_slot(struct mddev *mddev, int slot)
 356{
 357	struct md_cluster_info *cinfo = mddev->cluster_info;
 358
 359	set_bit(slot, &cinfo->recovery_map);
 360	if (!cinfo->recovery_thread) {
 361		cinfo->recovery_thread = md_register_thread(recover_bitmaps,
 362				mddev, "recover");
 363		if (!cinfo->recovery_thread) {
 364			pr_warn("md-cluster: Could not create recovery thread\n");
 365			return;
 366		}
 367	}
 368	md_wakeup_thread(cinfo->recovery_thread);
 369}
 370
 371static void recover_slot(void *arg, struct dlm_slot *slot)
 372{
 373	struct mddev *mddev = arg;
 374	struct md_cluster_info *cinfo = mddev->cluster_info;
 375
 376	pr_info("md-cluster: %s Node %d/%d down. My slot: %d. Initiating recovery.\n",
 377			mddev->bitmap_info.cluster_name,
 378			slot->nodeid, slot->slot,
 379			cinfo->slot_number);
 380	/* deduct one since dlm slot starts from one while the num of
 381	 * cluster-md begins with 0 */
 382	__recover_slot(mddev, slot->slot - 1);
 383}
 384
 385static void recover_done(void *arg, struct dlm_slot *slots,
 386		int num_slots, int our_slot,
 387		uint32_t generation)
 388{
 389	struct mddev *mddev = arg;
 390	struct md_cluster_info *cinfo = mddev->cluster_info;
 391
 392	cinfo->slot_number = our_slot;
 393	/* completion is only need to be complete when node join cluster,
 394	 * it doesn't need to run during another node's failure */
 395	if (test_bit(MD_CLUSTER_BEGIN_JOIN_CLUSTER, &cinfo->state)) {
 396		complete(&cinfo->completion);
 397		clear_bit(MD_CLUSTER_BEGIN_JOIN_CLUSTER, &cinfo->state);
 398	}
 399	clear_bit(MD_CLUSTER_SUSPEND_READ_BALANCING, &cinfo->state);
 400}
 401
 402/* the ops is called when node join the cluster, and do lock recovery
 403 * if node failure occurs */
 404static const struct dlm_lockspace_ops md_ls_ops = {
 405	.recover_prep = recover_prep,
 406	.recover_slot = recover_slot,
 407	.recover_done = recover_done,
 408};
 409
 410/*
 411 * The BAST function for the ack lock resource
 412 * This function wakes up the receive thread in
 413 * order to receive and process the message.
 414 */
 415static void ack_bast(void *arg, int mode)
 416{
 417	struct dlm_lock_resource *res = arg;
 418	struct md_cluster_info *cinfo = res->mddev->cluster_info;
 419
 420	if (mode == DLM_LOCK_EX) {
 421		if (test_bit(MD_CLUSTER_ALREADY_IN_CLUSTER, &cinfo->state))
 422			md_wakeup_thread(cinfo->recv_thread);
 423		else
 424			set_bit(MD_CLUSTER_PENDING_RECV_EVENT, &cinfo->state);
 425	}
 426}
 427
 428static void __remove_suspend_info(struct md_cluster_info *cinfo, int slot)
 429{
 430	struct suspend_info *s, *tmp;
 431
 432	list_for_each_entry_safe(s, tmp, &cinfo->suspend_list, list)
 433		if (slot == s->slot) {
 434			list_del(&s->list);
 435			kfree(s);
 436			break;
 437		}
 438}
 439
 440static void remove_suspend_info(struct mddev *mddev, int slot)
 441{
 442	struct md_cluster_info *cinfo = mddev->cluster_info;
 
 443	spin_lock_irq(&cinfo->suspend_lock);
 444	__remove_suspend_info(cinfo, slot);
 
 445	spin_unlock_irq(&cinfo->suspend_lock);
 446	mddev->pers->quiesce(mddev, 2);
 447}
 448
 449
 450static void process_suspend_info(struct mddev *mddev,
 451		int slot, sector_t lo, sector_t hi)
 452{
 453	struct md_cluster_info *cinfo = mddev->cluster_info;
 454	struct suspend_info *s;
 
 455
 456	if (!hi) {
 
 
 
 
 
 457		remove_suspend_info(mddev, slot);
 458		set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
 459		md_wakeup_thread(mddev->thread);
 460		return;
 461	}
 462
 
 
 
 
 
 
 463	/*
 464	 * The bitmaps are not same for different nodes
 465	 * if RESYNCING is happening in one node, then
 466	 * the node which received the RESYNCING message
 467	 * probably will perform resync with the region
 468	 * [lo, hi] again, so we could reduce resync time
 469	 * a lot if we can ensure that the bitmaps among
 470	 * different nodes are match up well.
 471	 *
 472	 * sync_low/hi is used to record the region which
 473	 * arrived in the previous RESYNCING message,
 474	 *
 475	 * Call bitmap_sync_with_cluster to clear
 476	 * NEEDED_MASK and set RESYNC_MASK since
 477	 * resync thread is running in another node,
 478	 * so we don't need to do the resync again
 479	 * with the same section */
 480	bitmap_sync_with_cluster(mddev, cinfo->sync_low,
 481					cinfo->sync_hi,
 482					lo, hi);
 
 
 
 
 483	cinfo->sync_low = lo;
 484	cinfo->sync_hi = hi;
 485
 486	s = kzalloc(sizeof(struct suspend_info), GFP_KERNEL);
 487	if (!s)
 488		return;
 489	s->slot = slot;
 490	s->lo = lo;
 491	s->hi = hi;
 492	mddev->pers->quiesce(mddev, 1);
 493	mddev->pers->quiesce(mddev, 0);
 494	spin_lock_irq(&cinfo->suspend_lock);
 495	/* Remove existing entry (if exists) before adding */
 496	__remove_suspend_info(cinfo, slot);
 497	list_add(&s->list, &cinfo->suspend_list);
 498	spin_unlock_irq(&cinfo->suspend_lock);
 499	mddev->pers->quiesce(mddev, 2);
 500}
 501
 502static void process_add_new_disk(struct mddev *mddev, struct cluster_msg *cmsg)
 503{
 504	char disk_uuid[64];
 505	struct md_cluster_info *cinfo = mddev->cluster_info;
 506	char event_name[] = "EVENT=ADD_DEVICE";
 507	char raid_slot[16];
 508	char *envp[] = {event_name, disk_uuid, raid_slot, NULL};
 509	int len;
 510
 511	len = snprintf(disk_uuid, 64, "DEVICE_UUID=");
 512	sprintf(disk_uuid + len, "%pU", cmsg->uuid);
 513	snprintf(raid_slot, 16, "RAID_DISK=%d", le32_to_cpu(cmsg->raid_slot));
 514	pr_info("%s:%d Sending kobject change with %s and %s\n", __func__, __LINE__, disk_uuid, raid_slot);
 515	init_completion(&cinfo->newdisk_completion);
 516	set_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state);
 517	kobject_uevent_env(&disk_to_dev(mddev->gendisk)->kobj, KOBJ_CHANGE, envp);
 518	wait_for_completion_timeout(&cinfo->newdisk_completion,
 519			NEW_DEV_TIMEOUT);
 520	clear_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state);
 521}
 522
 523
 524static void process_metadata_update(struct mddev *mddev, struct cluster_msg *msg)
 525{
 
 526	struct md_cluster_info *cinfo = mddev->cluster_info;
 527	mddev->good_device_nr = le32_to_cpu(msg->raid_slot);
 528	set_bit(MD_RELOAD_SB, &mddev->flags);
 529	dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR);
 530	md_wakeup_thread(mddev->thread);
 
 
 
 
 
 531}
 532
 533static void process_remove_disk(struct mddev *mddev, struct cluster_msg *msg)
 534{
 535	struct md_rdev *rdev;
 536
 537	rcu_read_lock();
 538	rdev = md_find_rdev_nr_rcu(mddev, le32_to_cpu(msg->raid_slot));
 539	if (rdev) {
 540		set_bit(ClusterRemove, &rdev->flags);
 541		set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
 542		md_wakeup_thread(mddev->thread);
 543	}
 544	else
 545		pr_warn("%s: %d Could not find disk(%d) to REMOVE\n",
 546			__func__, __LINE__, le32_to_cpu(msg->raid_slot));
 547	rcu_read_unlock();
 548}
 549
 550static void process_readd_disk(struct mddev *mddev, struct cluster_msg *msg)
 551{
 552	struct md_rdev *rdev;
 553
 554	rcu_read_lock();
 555	rdev = md_find_rdev_nr_rcu(mddev, le32_to_cpu(msg->raid_slot));
 556	if (rdev && test_bit(Faulty, &rdev->flags))
 557		clear_bit(Faulty, &rdev->flags);
 558	else
 559		pr_warn("%s: %d Could not find disk(%d) which is faulty",
 560			__func__, __LINE__, le32_to_cpu(msg->raid_slot));
 561	rcu_read_unlock();
 562}
 563
 564static int process_recvd_msg(struct mddev *mddev, struct cluster_msg *msg)
 565{
 566	int ret = 0;
 567
 568	if (WARN(mddev->cluster_info->slot_number - 1 == le32_to_cpu(msg->slot),
 569		"node %d received it's own msg\n", le32_to_cpu(msg->slot)))
 570		return -1;
 571	switch (le32_to_cpu(msg->type)) {
 572	case METADATA_UPDATED:
 573		process_metadata_update(mddev, msg);
 574		break;
 
 
 
 
 575	case RESYNCING:
 
 576		process_suspend_info(mddev, le32_to_cpu(msg->slot),
 577				     le64_to_cpu(msg->low),
 578				     le64_to_cpu(msg->high));
 579		break;
 580	case NEWDISK:
 581		process_add_new_disk(mddev, msg);
 582		break;
 583	case REMOVE:
 584		process_remove_disk(mddev, msg);
 585		break;
 586	case RE_ADD:
 587		process_readd_disk(mddev, msg);
 588		break;
 589	case BITMAP_NEEDS_SYNC:
 590		__recover_slot(mddev, le32_to_cpu(msg->slot));
 591		break;
 
 
 
 
 
 592	default:
 593		ret = -1;
 594		pr_warn("%s:%d Received unknown message from %d\n",
 595			__func__, __LINE__, msg->slot);
 596	}
 597	return ret;
 598}
 599
 600/*
 601 * thread for receiving message
 602 */
 603static void recv_daemon(struct md_thread *thread)
 604{
 605	struct md_cluster_info *cinfo = thread->mddev->cluster_info;
 606	struct dlm_lock_resource *ack_lockres = cinfo->ack_lockres;
 607	struct dlm_lock_resource *message_lockres = cinfo->message_lockres;
 608	struct cluster_msg msg;
 609	int ret;
 610
 611	mutex_lock(&cinfo->recv_mutex);
 612	/*get CR on Message*/
 613	if (dlm_lock_sync(message_lockres, DLM_LOCK_CR)) {
 614		pr_err("md/raid1:failed to get CR on MESSAGE\n");
 615		mutex_unlock(&cinfo->recv_mutex);
 616		return;
 617	}
 618
 619	/* read lvb and wake up thread to process this message_lockres */
 620	memcpy(&msg, message_lockres->lksb.sb_lvbptr, sizeof(struct cluster_msg));
 621	ret = process_recvd_msg(thread->mddev, &msg);
 622	if (ret)
 623		goto out;
 624
 625	/*release CR on ack_lockres*/
 626	ret = dlm_unlock_sync(ack_lockres);
 627	if (unlikely(ret != 0))
 628		pr_info("unlock ack failed return %d\n", ret);
 629	/*up-convert to PR on message_lockres*/
 630	ret = dlm_lock_sync(message_lockres, DLM_LOCK_PR);
 631	if (unlikely(ret != 0))
 632		pr_info("lock PR on msg failed return %d\n", ret);
 633	/*get CR on ack_lockres again*/
 634	ret = dlm_lock_sync(ack_lockres, DLM_LOCK_CR);
 635	if (unlikely(ret != 0))
 636		pr_info("lock CR on ack failed return %d\n", ret);
 637out:
 638	/*release CR on message_lockres*/
 639	ret = dlm_unlock_sync(message_lockres);
 640	if (unlikely(ret != 0))
 641		pr_info("unlock msg failed return %d\n", ret);
 642	mutex_unlock(&cinfo->recv_mutex);
 643}
 644
 645/* lock_token()
 646 * Takes the lock on the TOKEN lock resource so no other
 647 * node can communicate while the operation is underway.
 648 */
 649static int lock_token(struct md_cluster_info *cinfo)
 650{
 651	int error;
 
 652
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 653	error = dlm_lock_sync(cinfo->token_lockres, DLM_LOCK_EX);
 
 
 
 654	if (error)
 655		pr_err("md-cluster(%s:%d): failed to get EX on TOKEN (%d)\n",
 656				__func__, __LINE__, error);
 657
 658	/* Lock the receive sequence */
 659	mutex_lock(&cinfo->recv_mutex);
 660	return error;
 661}
 662
 663/* lock_comm()
 664 * Sets the MD_CLUSTER_SEND_LOCK bit to lock the send channel.
 665 */
 666static int lock_comm(struct md_cluster_info *cinfo)
 667{
 668	wait_event(cinfo->wait,
 669		   !test_and_set_bit(MD_CLUSTER_SEND_LOCK, &cinfo->state));
 670
 671	return lock_token(cinfo);
 672}
 673
 674static void unlock_comm(struct md_cluster_info *cinfo)
 675{
 676	WARN_ON(cinfo->token_lockres->mode != DLM_LOCK_EX);
 677	mutex_unlock(&cinfo->recv_mutex);
 678	dlm_unlock_sync(cinfo->token_lockres);
 679	clear_bit(MD_CLUSTER_SEND_LOCK, &cinfo->state);
 680	wake_up(&cinfo->wait);
 681}
 682
 683/* __sendmsg()
 684 * This function performs the actual sending of the message. This function is
 685 * usually called after performing the encompassing operation
 686 * The function:
 687 * 1. Grabs the message lockresource in EX mode
 688 * 2. Copies the message to the message LVB
 689 * 3. Downconverts message lockresource to CW
 690 * 4. Upconverts ack lock resource from CR to EX. This forces the BAST on other nodes
 691 *    and the other nodes read the message. The thread will wait here until all other
 692 *    nodes have released ack lock resource.
 693 * 5. Downconvert ack lockresource to CR
 694 */
 695static int __sendmsg(struct md_cluster_info *cinfo, struct cluster_msg *cmsg)
 696{
 697	int error;
 698	int slot = cinfo->slot_number - 1;
 699
 700	cmsg->slot = cpu_to_le32(slot);
 701	/*get EX on Message*/
 702	error = dlm_lock_sync(cinfo->message_lockres, DLM_LOCK_EX);
 703	if (error) {
 704		pr_err("md-cluster: failed to get EX on MESSAGE (%d)\n", error);
 705		goto failed_message;
 706	}
 707
 708	memcpy(cinfo->message_lockres->lksb.sb_lvbptr, (void *)cmsg,
 709			sizeof(struct cluster_msg));
 710	/*down-convert EX to CW on Message*/
 711	error = dlm_lock_sync(cinfo->message_lockres, DLM_LOCK_CW);
 712	if (error) {
 713		pr_err("md-cluster: failed to convert EX to CW on MESSAGE(%d)\n",
 714				error);
 715		goto failed_ack;
 716	}
 717
 718	/*up-convert CR to EX on Ack*/
 719	error = dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_EX);
 720	if (error) {
 721		pr_err("md-cluster: failed to convert CR to EX on ACK(%d)\n",
 722				error);
 723		goto failed_ack;
 724	}
 725
 726	/*down-convert EX to CR on Ack*/
 727	error = dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_CR);
 728	if (error) {
 729		pr_err("md-cluster: failed to convert EX to CR on ACK(%d)\n",
 730				error);
 731		goto failed_ack;
 732	}
 733
 734failed_ack:
 735	error = dlm_unlock_sync(cinfo->message_lockres);
 736	if (unlikely(error != 0)) {
 737		pr_err("md-cluster: failed convert to NL on MESSAGE(%d)\n",
 738			error);
 739		/* in case the message can't be released due to some reason */
 740		goto failed_ack;
 741	}
 742failed_message:
 743	return error;
 744}
 745
 746static int sendmsg(struct md_cluster_info *cinfo, struct cluster_msg *cmsg)
 
 747{
 748	int ret;
 749
 750	lock_comm(cinfo);
 751	ret = __sendmsg(cinfo, cmsg);
 752	unlock_comm(cinfo);
 753	return ret;
 754}
 755
 756static int gather_all_resync_info(struct mddev *mddev, int total_slots)
 757{
 758	struct md_cluster_info *cinfo = mddev->cluster_info;
 759	int i, ret = 0;
 760	struct dlm_lock_resource *bm_lockres;
 761	struct suspend_info *s;
 762	char str[64];
 763	sector_t lo, hi;
 764
 765
 766	for (i = 0; i < total_slots; i++) {
 767		memset(str, '\0', 64);
 768		snprintf(str, 64, "bitmap%04d", i);
 769		bm_lockres = lockres_init(mddev, str, NULL, 1);
 770		if (!bm_lockres)
 771			return -ENOMEM;
 772		if (i == (cinfo->slot_number - 1)) {
 773			lockres_free(bm_lockres);
 774			continue;
 775		}
 776
 777		bm_lockres->flags |= DLM_LKF_NOQUEUE;
 778		ret = dlm_lock_sync(bm_lockres, DLM_LOCK_PW);
 779		if (ret == -EAGAIN) {
 780			memset(bm_lockres->lksb.sb_lvbptr, '\0', LVB_SIZE);
 781			s = read_resync_info(mddev, bm_lockres);
 782			if (s) {
 783				pr_info("%s:%d Resync[%llu..%llu] in progress on %d\n",
 784						__func__, __LINE__,
 785						(unsigned long long) s->lo,
 786						(unsigned long long) s->hi, i);
 787				spin_lock_irq(&cinfo->suspend_lock);
 788				s->slot = i;
 789				list_add(&s->list, &cinfo->suspend_list);
 790				spin_unlock_irq(&cinfo->suspend_lock);
 791			}
 792			ret = 0;
 793			lockres_free(bm_lockres);
 794			continue;
 795		}
 796		if (ret) {
 797			lockres_free(bm_lockres);
 798			goto out;
 799		}
 800
 801		/* Read the disk bitmap sb and check if it needs recovery */
 802		ret = bitmap_copy_from_slot(mddev, i, &lo, &hi, false);
 803		if (ret) {
 804			pr_warn("md-cluster: Could not gather bitmaps from slot %d", i);
 805			lockres_free(bm_lockres);
 806			continue;
 807		}
 808		if ((hi > 0) && (lo < mddev->recovery_cp)) {
 809			set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
 810			mddev->recovery_cp = lo;
 811			md_check_recovery(mddev);
 812		}
 813
 814		lockres_free(bm_lockres);
 815	}
 816out:
 817	return ret;
 818}
 819
 820static int join(struct mddev *mddev, int nodes)
 821{
 822	struct md_cluster_info *cinfo;
 823	int ret, ops_rv;
 824	char str[64];
 825
 826	cinfo = kzalloc(sizeof(struct md_cluster_info), GFP_KERNEL);
 827	if (!cinfo)
 828		return -ENOMEM;
 829
 830	INIT_LIST_HEAD(&cinfo->suspend_list);
 831	spin_lock_init(&cinfo->suspend_lock);
 832	init_completion(&cinfo->completion);
 833	set_bit(MD_CLUSTER_BEGIN_JOIN_CLUSTER, &cinfo->state);
 834	init_waitqueue_head(&cinfo->wait);
 835	mutex_init(&cinfo->recv_mutex);
 836
 837	mddev->cluster_info = cinfo;
 
 838
 839	memset(str, 0, 64);
 840	sprintf(str, "%pU", mddev->uuid);
 841	ret = dlm_new_lockspace(str, mddev->bitmap_info.cluster_name,
 842				DLM_LSFL_FS, LVB_SIZE,
 843				&md_ls_ops, mddev, &ops_rv, &cinfo->lockspace);
 844	if (ret)
 845		goto err;
 846	wait_for_completion(&cinfo->completion);
 847	if (nodes < cinfo->slot_number) {
 848		pr_err("md-cluster: Slot allotted(%d) is greater than available slots(%d).",
 849			cinfo->slot_number, nodes);
 850		ret = -ERANGE;
 851		goto err;
 852	}
 853	/* Initiate the communication resources */
 854	ret = -ENOMEM;
 855	cinfo->recv_thread = md_register_thread(recv_daemon, mddev, "cluster_recv");
 856	if (!cinfo->recv_thread) {
 857		pr_err("md-cluster: cannot allocate memory for recv_thread!\n");
 858		goto err;
 859	}
 860	cinfo->message_lockres = lockres_init(mddev, "message", NULL, 1);
 861	if (!cinfo->message_lockres)
 862		goto err;
 863	cinfo->token_lockres = lockres_init(mddev, "token", NULL, 0);
 864	if (!cinfo->token_lockres)
 865		goto err;
 866	cinfo->no_new_dev_lockres = lockres_init(mddev, "no-new-dev", NULL, 0);
 867	if (!cinfo->no_new_dev_lockres)
 868		goto err;
 869
 870	ret = dlm_lock_sync(cinfo->token_lockres, DLM_LOCK_EX);
 871	if (ret) {
 872		ret = -EAGAIN;
 873		pr_err("md-cluster: can't join cluster to avoid lock issue\n");
 874		goto err;
 875	}
 876	cinfo->ack_lockres = lockres_init(mddev, "ack", ack_bast, 0);
 877	if (!cinfo->ack_lockres) {
 878		ret = -ENOMEM;
 879		goto err;
 880	}
 881	/* get sync CR lock on ACK. */
 882	if (dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_CR))
 883		pr_err("md-cluster: failed to get a sync CR lock on ACK!(%d)\n",
 884				ret);
 885	dlm_unlock_sync(cinfo->token_lockres);
 886	/* get sync CR lock on no-new-dev. */
 887	if (dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR))
 888		pr_err("md-cluster: failed to get a sync CR lock on no-new-dev!(%d)\n", ret);
 889
 890
 891	pr_info("md-cluster: Joined cluster %s slot %d\n", str, cinfo->slot_number);
 892	snprintf(str, 64, "bitmap%04d", cinfo->slot_number - 1);
 893	cinfo->bitmap_lockres = lockres_init(mddev, str, NULL, 1);
 894	if (!cinfo->bitmap_lockres) {
 895		ret = -ENOMEM;
 896		goto err;
 897	}
 898	if (dlm_lock_sync(cinfo->bitmap_lockres, DLM_LOCK_PW)) {
 899		pr_err("Failed to get bitmap lock\n");
 900		ret = -EINVAL;
 901		goto err;
 902	}
 903
 904	cinfo->resync_lockres = lockres_init(mddev, "resync", NULL, 0);
 905	if (!cinfo->resync_lockres) {
 906		ret = -ENOMEM;
 907		goto err;
 908	}
 909
 910	return 0;
 911err:
 
 912	md_unregister_thread(&cinfo->recovery_thread);
 913	md_unregister_thread(&cinfo->recv_thread);
 914	lockres_free(cinfo->message_lockres);
 915	lockres_free(cinfo->token_lockres);
 916	lockres_free(cinfo->ack_lockres);
 917	lockres_free(cinfo->no_new_dev_lockres);
 918	lockres_free(cinfo->resync_lockres);
 919	lockres_free(cinfo->bitmap_lockres);
 920	if (cinfo->lockspace)
 921		dlm_release_lockspace(cinfo->lockspace, 2);
 922	mddev->cluster_info = NULL;
 923	kfree(cinfo);
 924	return ret;
 925}
 926
 927static void load_bitmaps(struct mddev *mddev, int total_slots)
 928{
 929	struct md_cluster_info *cinfo = mddev->cluster_info;
 930
 931	/* load all the node's bitmap info for resync */
 932	if (gather_all_resync_info(mddev, total_slots))
 933		pr_err("md-cluster: failed to gather all resyn infos\n");
 934	set_bit(MD_CLUSTER_ALREADY_IN_CLUSTER, &cinfo->state);
 935	/* wake up recv thread in case something need to be handled */
 936	if (test_and_clear_bit(MD_CLUSTER_PENDING_RECV_EVENT, &cinfo->state))
 937		md_wakeup_thread(cinfo->recv_thread);
 938}
 939
 940static void resync_bitmap(struct mddev *mddev)
 941{
 942	struct md_cluster_info *cinfo = mddev->cluster_info;
 943	struct cluster_msg cmsg = {0};
 944	int err;
 945
 946	cmsg.type = cpu_to_le32(BITMAP_NEEDS_SYNC);
 947	err = sendmsg(cinfo, &cmsg);
 948	if (err)
 949		pr_err("%s:%d: failed to send BITMAP_NEEDS_SYNC message (%d)\n",
 950			__func__, __LINE__, err);
 951}
 952
 953static void unlock_all_bitmaps(struct mddev *mddev);
 954static int leave(struct mddev *mddev)
 955{
 956	struct md_cluster_info *cinfo = mddev->cluster_info;
 957
 958	if (!cinfo)
 959		return 0;
 960
 961	/* BITMAP_NEEDS_SYNC message should be sent when node
 
 962	 * is leaving the cluster with dirty bitmap, also we
 963	 * can only deliver it when dlm connection is available */
 964	if (cinfo->slot_number > 0 && mddev->recovery_cp != MaxSector)
 
 
 
 
 
 
 965		resync_bitmap(mddev);
 966
 
 967	md_unregister_thread(&cinfo->recovery_thread);
 968	md_unregister_thread(&cinfo->recv_thread);
 969	lockres_free(cinfo->message_lockres);
 970	lockres_free(cinfo->token_lockres);
 971	lockres_free(cinfo->ack_lockres);
 972	lockres_free(cinfo->no_new_dev_lockres);
 973	lockres_free(cinfo->resync_lockres);
 974	lockres_free(cinfo->bitmap_lockres);
 975	unlock_all_bitmaps(mddev);
 976	dlm_release_lockspace(cinfo->lockspace, 2);
 
 977	return 0;
 978}
 979
 980/* slot_number(): Returns the MD slot number to use
 981 * DLM starts the slot numbers from 1, wheras cluster-md
 982 * wants the number to be from zero, so we deduct one
 983 */
 984static int slot_number(struct mddev *mddev)
 985{
 986	struct md_cluster_info *cinfo = mddev->cluster_info;
 987
 988	return cinfo->slot_number - 1;
 989}
 990
 991/*
 992 * Check if the communication is already locked, else lock the communication
 993 * channel.
 994 * If it is already locked, token is in EX mode, and hence lock_token()
 995 * should not be called.
 996 */
 997static int metadata_update_start(struct mddev *mddev)
 998{
 999	struct md_cluster_info *cinfo = mddev->cluster_info;
 
 
 
 
 
 
 
 
 
 
1000
1001	wait_event(cinfo->wait,
1002		   !test_and_set_bit(MD_CLUSTER_SEND_LOCK, &cinfo->state) ||
1003		   test_and_clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state));
1004
1005	/* If token is already locked, return 0 */
1006	if (cinfo->token_lockres->mode == DLM_LOCK_EX)
 
1007		return 0;
 
1008
1009	return lock_token(cinfo);
 
 
1010}
1011
1012static int metadata_update_finish(struct mddev *mddev)
1013{
1014	struct md_cluster_info *cinfo = mddev->cluster_info;
1015	struct cluster_msg cmsg;
1016	struct md_rdev *rdev;
1017	int ret = 0;
1018	int raid_slot = -1;
1019
1020	memset(&cmsg, 0, sizeof(cmsg));
1021	cmsg.type = cpu_to_le32(METADATA_UPDATED);
1022	/* Pick up a good active device number to send.
1023	 */
1024	rdev_for_each(rdev, mddev)
1025		if (rdev->raid_disk > -1 && !test_bit(Faulty, &rdev->flags)) {
1026			raid_slot = rdev->desc_nr;
1027			break;
1028		}
1029	if (raid_slot >= 0) {
1030		cmsg.raid_slot = cpu_to_le32(raid_slot);
1031		ret = __sendmsg(cinfo, &cmsg);
1032	} else
1033		pr_warn("md-cluster: No good device id found to send\n");
1034	clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1035	unlock_comm(cinfo);
1036	return ret;
1037}
1038
1039static void metadata_update_cancel(struct mddev *mddev)
1040{
1041	struct md_cluster_info *cinfo = mddev->cluster_info;
1042	clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1043	unlock_comm(cinfo);
1044}
1045
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1046static int resync_start(struct mddev *mddev)
1047{
1048	struct md_cluster_info *cinfo = mddev->cluster_info;
1049	return dlm_lock_sync_interruptible(cinfo->resync_lockres, DLM_LOCK_EX, mddev);
1050}
1051
 
 
 
 
 
 
 
 
 
 
1052static int resync_info_update(struct mddev *mddev, sector_t lo, sector_t hi)
1053{
1054	struct md_cluster_info *cinfo = mddev->cluster_info;
1055	struct resync_info ri;
1056	struct cluster_msg cmsg = {0};
1057
1058	/* do not send zero again, if we have sent before */
1059	if (hi == 0) {
1060		memcpy(&ri, cinfo->bitmap_lockres->lksb.sb_lvbptr, sizeof(struct resync_info));
1061		if (le64_to_cpu(ri.hi) == 0)
1062			return 0;
1063	}
1064
1065	add_resync_info(cinfo->bitmap_lockres, lo, hi);
1066	/* Re-acquire the lock to refresh LVB */
1067	dlm_lock_sync(cinfo->bitmap_lockres, DLM_LOCK_PW);
1068	cmsg.type = cpu_to_le32(RESYNCING);
1069	cmsg.low = cpu_to_le64(lo);
1070	cmsg.high = cpu_to_le64(hi);
1071
1072	return sendmsg(cinfo, &cmsg);
 
 
 
 
 
 
 
1073}
1074
1075static int resync_finish(struct mddev *mddev)
1076{
1077	struct md_cluster_info *cinfo = mddev->cluster_info;
 
 
 
 
 
 
 
 
 
 
1078	dlm_unlock_sync(cinfo->resync_lockres);
1079	return resync_info_update(mddev, 0, 0);
1080}
1081
1082static int area_resyncing(struct mddev *mddev, int direction,
1083		sector_t lo, sector_t hi)
1084{
1085	struct md_cluster_info *cinfo = mddev->cluster_info;
1086	int ret = 0;
1087	struct suspend_info *s;
1088
1089	if ((direction == READ) &&
1090		test_bit(MD_CLUSTER_SUSPEND_READ_BALANCING, &cinfo->state))
1091		return 1;
1092
1093	spin_lock_irq(&cinfo->suspend_lock);
1094	if (list_empty(&cinfo->suspend_list))
1095		goto out;
1096	list_for_each_entry(s, &cinfo->suspend_list, list)
1097		if (hi > s->lo && lo < s->hi) {
1098			ret = 1;
1099			break;
1100		}
1101out:
1102	spin_unlock_irq(&cinfo->suspend_lock);
1103	return ret;
1104}
1105
1106/* add_new_disk() - initiates a disk add
1107 * However, if this fails before writing md_update_sb(),
1108 * add_new_disk_cancel() must be called to release token lock
1109 */
1110static int add_new_disk(struct mddev *mddev, struct md_rdev *rdev)
1111{
1112	struct md_cluster_info *cinfo = mddev->cluster_info;
1113	struct cluster_msg cmsg;
1114	int ret = 0;
1115	struct mdp_superblock_1 *sb = page_address(rdev->sb_page);
1116	char *uuid = sb->device_uuid;
1117
1118	memset(&cmsg, 0, sizeof(cmsg));
1119	cmsg.type = cpu_to_le32(NEWDISK);
1120	memcpy(cmsg.uuid, uuid, 16);
1121	cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
1122	lock_comm(cinfo);
1123	ret = __sendmsg(cinfo, &cmsg);
1124	if (ret)
 
1125		return ret;
 
1126	cinfo->no_new_dev_lockres->flags |= DLM_LKF_NOQUEUE;
1127	ret = dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_EX);
1128	cinfo->no_new_dev_lockres->flags &= ~DLM_LKF_NOQUEUE;
1129	/* Some node does not "see" the device */
1130	if (ret == -EAGAIN)
1131		ret = -ENOENT;
1132	if (ret)
1133		unlock_comm(cinfo);
1134	else {
1135		dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR);
1136		/* Since MD_CHANGE_DEVS will be set in add_bound_rdev which
1137		 * will run soon after add_new_disk, the below path will be
1138		 * invoked:
1139		 *   md_wakeup_thread(mddev->thread)
1140		 *	-> conf->thread (raid1d)
1141		 *	-> md_check_recovery -> md_update_sb
1142		 *	-> metadata_update_start/finish
1143		 * MD_CLUSTER_SEND_LOCKED_ALREADY will be cleared eventually.
1144		 *
1145		 * For other failure cases, metadata_update_cancel and
1146		 * add_new_disk_cancel also clear below bit as well.
1147		 * */
1148		set_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1149		wake_up(&cinfo->wait);
1150	}
1151	return ret;
1152}
1153
1154static void add_new_disk_cancel(struct mddev *mddev)
1155{
1156	struct md_cluster_info *cinfo = mddev->cluster_info;
1157	clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1158	unlock_comm(cinfo);
1159}
1160
1161static int new_disk_ack(struct mddev *mddev, bool ack)
1162{
1163	struct md_cluster_info *cinfo = mddev->cluster_info;
1164
1165	if (!test_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state)) {
1166		pr_warn("md-cluster(%s): Spurious cluster confirmation\n", mdname(mddev));
1167		return -EINVAL;
1168	}
1169
1170	if (ack)
1171		dlm_unlock_sync(cinfo->no_new_dev_lockres);
1172	complete(&cinfo->newdisk_completion);
1173	return 0;
1174}
1175
1176static int remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1177{
1178	struct cluster_msg cmsg = {0};
1179	struct md_cluster_info *cinfo = mddev->cluster_info;
1180	cmsg.type = cpu_to_le32(REMOVE);
1181	cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
1182	return sendmsg(cinfo, &cmsg);
1183}
1184
1185static int lock_all_bitmaps(struct mddev *mddev)
1186{
1187	int slot, my_slot, ret, held = 1, i = 0;
1188	char str[64];
1189	struct md_cluster_info *cinfo = mddev->cluster_info;
1190
1191	cinfo->other_bitmap_lockres = kzalloc((mddev->bitmap_info.nodes - 1) *
1192					     sizeof(struct dlm_lock_resource *),
1193					     GFP_KERNEL);
1194	if (!cinfo->other_bitmap_lockres) {
1195		pr_err("md: can't alloc mem for other bitmap locks\n");
1196		return 0;
1197	}
1198
1199	my_slot = slot_number(mddev);
1200	for (slot = 0; slot < mddev->bitmap_info.nodes; slot++) {
1201		if (slot == my_slot)
1202			continue;
1203
1204		memset(str, '\0', 64);
1205		snprintf(str, 64, "bitmap%04d", slot);
1206		cinfo->other_bitmap_lockres[i] = lockres_init(mddev, str, NULL, 1);
1207		if (!cinfo->other_bitmap_lockres[i])
1208			return -ENOMEM;
1209
1210		cinfo->other_bitmap_lockres[i]->flags |= DLM_LKF_NOQUEUE;
1211		ret = dlm_lock_sync(cinfo->other_bitmap_lockres[i], DLM_LOCK_PW);
1212		if (ret)
1213			held = -1;
1214		i++;
1215	}
1216
1217	return held;
1218}
1219
1220static void unlock_all_bitmaps(struct mddev *mddev)
1221{
1222	struct md_cluster_info *cinfo = mddev->cluster_info;
1223	int i;
1224
1225	/* release other node's bitmap lock if they are existed */
1226	if (cinfo->other_bitmap_lockres) {
1227		for (i = 0; i < mddev->bitmap_info.nodes - 1; i++) {
1228			if (cinfo->other_bitmap_lockres[i]) {
1229				lockres_free(cinfo->other_bitmap_lockres[i]);
1230			}
1231		}
1232		kfree(cinfo->other_bitmap_lockres);
 
1233	}
1234}
1235
1236static int gather_bitmaps(struct md_rdev *rdev)
1237{
1238	int sn, err;
1239	sector_t lo, hi;
1240	struct cluster_msg cmsg = {0};
1241	struct mddev *mddev = rdev->mddev;
1242	struct md_cluster_info *cinfo = mddev->cluster_info;
1243
1244	cmsg.type = cpu_to_le32(RE_ADD);
1245	cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
1246	err = sendmsg(cinfo, &cmsg);
1247	if (err)
1248		goto out;
1249
1250	for (sn = 0; sn < mddev->bitmap_info.nodes; sn++) {
1251		if (sn == (cinfo->slot_number - 1))
1252			continue;
1253		err = bitmap_copy_from_slot(mddev, sn, &lo, &hi, false);
1254		if (err) {
1255			pr_warn("md-cluster: Could not gather bitmaps from slot %d", sn);
1256			goto out;
1257		}
1258		if ((hi > 0) && (lo < mddev->recovery_cp))
1259			mddev->recovery_cp = lo;
1260	}
1261out:
1262	return err;
1263}
1264
1265static struct md_cluster_operations cluster_ops = {
1266	.join   = join,
1267	.leave  = leave,
1268	.slot_number = slot_number,
1269	.resync_start = resync_start,
1270	.resync_finish = resync_finish,
1271	.resync_info_update = resync_info_update,
 
1272	.metadata_update_start = metadata_update_start,
1273	.metadata_update_finish = metadata_update_finish,
1274	.metadata_update_cancel = metadata_update_cancel,
1275	.area_resyncing = area_resyncing,
1276	.add_new_disk = add_new_disk,
1277	.add_new_disk_cancel = add_new_disk_cancel,
1278	.new_disk_ack = new_disk_ack,
1279	.remove_disk = remove_disk,
1280	.load_bitmaps = load_bitmaps,
1281	.gather_bitmaps = gather_bitmaps,
 
1282	.lock_all_bitmaps = lock_all_bitmaps,
1283	.unlock_all_bitmaps = unlock_all_bitmaps,
 
1284};
1285
1286static int __init cluster_init(void)
1287{
1288	pr_warn("md-cluster: EXPERIMENTAL. Use with caution\n");
1289	pr_info("Registering Cluster MD functions\n");
1290	register_md_cluster_operations(&cluster_ops, THIS_MODULE);
1291	return 0;
1292}
1293
1294static void cluster_exit(void)
1295{
1296	unregister_md_cluster_operations();
1297}
1298
1299module_init(cluster_init);
1300module_exit(cluster_exit);
1301MODULE_AUTHOR("SUSE");
1302MODULE_LICENSE("GPL");
1303MODULE_DESCRIPTION("Clustering support for MD");