drivers/md/md-cluster.c

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