drivers/md/md-cluster.c

   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
  19 struct 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
  31 struct 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  * accommodate 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
  60 struct 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 __rcu *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 __rcu *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
  94 enum 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
 105 struct 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
 115 static 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
 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_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
 140 static 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  */
 149 static 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
 183 static 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         strscpy(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;
 228 out_err:
 229         kfree(res->lksb.sb_lvbptr);
 230         kfree(res->name);
 231         kfree(res);
 232         return NULL;
 233 }
 234
 235 static 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
 258 static 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
 268 static 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
 286 static 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                 }
 346 clear_bit:
 347                 lockres_free(bm_lockres);
 348                 clear_bit(slot, &cinfo->recovery_map);
 349         }
 350 }
 351
 352 static 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
 359 static 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                 rcu_assign_pointer(cinfo->recovery_thread,
 366                         md_register_thread(recover_bitmaps, 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
 375 static 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
 389 static 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 */
 408 static 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  */
 419 static 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
 432 static 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
 443 static 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
 504 static int 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         int res = 0;
 513
 514         len = snprintf(disk_uuid, 64, "DEVICE_UUID=");
 515         sprintf(disk_uuid + len, "%pU", cmsg->uuid);
 516         snprintf(raid_slot, 16, "RAID_DISK=%d", le32_to_cpu(cmsg->raid_slot));
 517         pr_info("%s:%d Sending kobject change with %s and %s\n", __func__, __LINE__, disk_uuid, raid_slot);
 518         init_completion(&cinfo->newdisk_completion);
 519         set_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state);
 520         kobject_uevent_env(&disk_to_dev(mddev->gendisk)->kobj, KOBJ_CHANGE, envp);
 521         if (!wait_for_completion_timeout(&cinfo->newdisk_completion,
 522                                         NEW_DEV_TIMEOUT)) {
 523                 pr_err("md-cluster(%s:%d): timeout on a new disk adding\n",
 524                         __func__, __LINE__);
 525                 res = -1;
 526         }
 527         clear_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state);
 528         return res;
 529 }
 530
 531
 532 static void process_metadata_update(struct mddev *mddev, struct cluster_msg *msg)
 533 {
 534         int got_lock = 0;
 535         struct md_thread *thread;
 536         struct md_cluster_info *cinfo = mddev->cluster_info;
 537         mddev->good_device_nr = le32_to_cpu(msg->raid_slot);
 538
 539         dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR);
 540
 541         /* daemaon thread must exist */
 542         thread = rcu_dereference_protected(mddev->thread, true);
 543         wait_event(thread->wqueue,
 544                    (got_lock = mddev_trylock(mddev)) ||
 545                     test_bit(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state));
 546         md_reload_sb(mddev, mddev->good_device_nr);
 547         if (got_lock)
 548                 mddev_unlock(mddev);
 549 }
 550
 551 static void process_remove_disk(struct mddev *mddev, struct cluster_msg *msg)
 552 {
 553         struct md_rdev *rdev;
 554
 555         rcu_read_lock();
 556         rdev = md_find_rdev_nr_rcu(mddev, le32_to_cpu(msg->raid_slot));
 557         if (rdev) {
 558                 set_bit(ClusterRemove, &rdev->flags);
 559                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
 560                 md_wakeup_thread(mddev->thread);
 561         }
 562         else
 563                 pr_warn("%s: %d Could not find disk(%d) to REMOVE\n",
 564                         __func__, __LINE__, le32_to_cpu(msg->raid_slot));
 565         rcu_read_unlock();
 566 }
 567
 568 static void process_readd_disk(struct mddev *mddev, struct cluster_msg *msg)
 569 {
 570         struct md_rdev *rdev;
 571
 572         rcu_read_lock();
 573         rdev = md_find_rdev_nr_rcu(mddev, le32_to_cpu(msg->raid_slot));
 574         if (rdev && test_bit(Faulty, &rdev->flags))
 575                 clear_bit(Faulty, &rdev->flags);
 576         else
 577                 pr_warn("%s: %d Could not find disk(%d) which is faulty",
 578                         __func__, __LINE__, le32_to_cpu(msg->raid_slot));
 579         rcu_read_unlock();
 580 }
 581
 582 static int process_recvd_msg(struct mddev *mddev, struct cluster_msg *msg)
 583 {
 584         int ret = 0;
 585
 586         if (WARN(mddev->cluster_info->slot_number - 1 == le32_to_cpu(msg->slot),
 587                 "node %d received its own msg\n", le32_to_cpu(msg->slot)))
 588                 return -1;
 589         switch (le32_to_cpu(msg->type)) {
 590         case METADATA_UPDATED:
 591                 process_metadata_update(mddev, msg);
 592                 break;
 593         case CHANGE_CAPACITY:
 594                 set_capacity_and_notify(mddev->gendisk, mddev->array_sectors);
 595                 break;
 596         case RESYNCING:
 597                 set_bit(MD_RESYNCING_REMOTE, &mddev->recovery);
 598                 process_suspend_info(mddev, le32_to_cpu(msg->slot),
 599                                      le64_to_cpu(msg->low),
 600                                      le64_to_cpu(msg->high));
 601                 break;
 602         case NEWDISK:
 603                 if (process_add_new_disk(mddev, msg))
 604                         ret = -1;
 605                 break;
 606         case REMOVE:
 607                 process_remove_disk(mddev, msg);
 608                 break;
 609         case RE_ADD:
 610                 process_readd_disk(mddev, msg);
 611                 break;
 612         case BITMAP_NEEDS_SYNC:
 613                 __recover_slot(mddev, le32_to_cpu(msg->slot));
 614                 break;
 615         case BITMAP_RESIZE:
 616                 if (le64_to_cpu(msg->high) != mddev->pers->size(mddev, 0, 0))
 617                         ret = md_bitmap_resize(mddev->bitmap,
 618                                             le64_to_cpu(msg->high), 0, 0);
 619                 break;
 620         default:
 621                 ret = -1;
 622                 pr_warn("%s:%d Received unknown message from %d\n",
 623                         __func__, __LINE__, msg->slot);
 624         }
 625         return ret;
 626 }
 627
 628 /*
 629  * thread for receiving message
 630  */
 631 static void recv_daemon(struct md_thread *thread)
 632 {
 633         struct md_cluster_info *cinfo = thread->mddev->cluster_info;
 634         struct dlm_lock_resource *ack_lockres = cinfo->ack_lockres;
 635         struct dlm_lock_resource *message_lockres = cinfo->message_lockres;
 636         struct cluster_msg msg;
 637         int ret;
 638
 639         mutex_lock(&cinfo->recv_mutex);
 640         /*get CR on Message*/
 641         if (dlm_lock_sync(message_lockres, DLM_LOCK_CR)) {
 642                 pr_err("md/raid1:failed to get CR on MESSAGE\n");
 643                 mutex_unlock(&cinfo->recv_mutex);
 644                 return;
 645         }
 646
 647         /* read lvb and wake up thread to process this message_lockres */
 648         memcpy(&msg, message_lockres->lksb.sb_lvbptr, sizeof(struct cluster_msg));
 649         ret = process_recvd_msg(thread->mddev, &msg);
 650         if (ret)
 651                 goto out;
 652
 653         /*release CR on ack_lockres*/
 654         ret = dlm_unlock_sync(ack_lockres);
 655         if (unlikely(ret != 0))
 656                 pr_info("unlock ack failed return %d\n", ret);
 657         /*up-convert to PR on message_lockres*/
 658         ret = dlm_lock_sync(message_lockres, DLM_LOCK_PR);
 659         if (unlikely(ret != 0))
 660                 pr_info("lock PR on msg failed return %d\n", ret);
 661         /*get CR on ack_lockres again*/
 662         ret = dlm_lock_sync(ack_lockres, DLM_LOCK_CR);
 663         if (unlikely(ret != 0))
 664                 pr_info("lock CR on ack failed return %d\n", ret);
 665 out:
 666         /*release CR on message_lockres*/
 667         ret = dlm_unlock_sync(message_lockres);
 668         if (unlikely(ret != 0))
 669                 pr_info("unlock msg failed return %d\n", ret);
 670         mutex_unlock(&cinfo->recv_mutex);
 671 }
 672
 673 /* lock_token()
 674  * Takes the lock on the TOKEN lock resource so no other
 675  * node can communicate while the operation is underway.
 676  */
 677 static int lock_token(struct md_cluster_info *cinfo)
 678 {
 679         int error;
 680
 681         error = dlm_lock_sync(cinfo->token_lockres, DLM_LOCK_EX);
 682         if (error) {
 683                 pr_err("md-cluster(%s:%d): failed to get EX on TOKEN (%d)\n",
 684                                 __func__, __LINE__, error);
 685         } else {
 686                 /* Lock the receive sequence */
 687                 mutex_lock(&cinfo->recv_mutex);
 688         }
 689         return error;
 690 }
 691
 692 /* lock_comm()
 693  * Sets the MD_CLUSTER_SEND_LOCK bit to lock the send channel.
 694  */
 695 static int lock_comm(struct md_cluster_info *cinfo, bool mddev_locked)
 696 {
 697         int rv, set_bit = 0;
 698         struct mddev *mddev = cinfo->mddev;
 699
 700         /*
 701          * If resync thread run after raid1d thread, then process_metadata_update
 702          * could not continue if raid1d held reconfig_mutex (and raid1d is blocked
 703          * since another node already got EX on Token and waiting the EX of Ack),
 704          * so let resync wake up thread in case flag is set.
 705          */
 706         if (mddev_locked && !test_bit(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD,
 707                                       &cinfo->state)) {
 708                 rv = test_and_set_bit_lock(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD,
 709                                               &cinfo->state);
 710                 WARN_ON_ONCE(rv);
 711                 md_wakeup_thread(mddev->thread);
 712                 set_bit = 1;
 713         }
 714
 715         wait_event(cinfo->wait,
 716                    !test_and_set_bit(MD_CLUSTER_SEND_LOCK, &cinfo->state));
 717         rv = lock_token(cinfo);
 718         if (set_bit)
 719                 clear_bit_unlock(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state);
 720         return rv;
 721 }
 722
 723 static void unlock_comm(struct md_cluster_info *cinfo)
 724 {
 725         WARN_ON(cinfo->token_lockres->mode != DLM_LOCK_EX);
 726         mutex_unlock(&cinfo->recv_mutex);
 727         dlm_unlock_sync(cinfo->token_lockres);
 728         clear_bit(MD_CLUSTER_SEND_LOCK, &cinfo->state);
 729         wake_up(&cinfo->wait);
 730 }
 731
 732 /* __sendmsg()
 733  * This function performs the actual sending of the message. This function is
 734  * usually called after performing the encompassing operation
 735  * The function:
 736  * 1. Grabs the message lockresource in EX mode
 737  * 2. Copies the message to the message LVB
 738  * 3. Downconverts message lockresource to CW
 739  * 4. Upconverts ack lock resource from CR to EX. This forces the BAST on other nodes
 740  *    and the other nodes read the message. The thread will wait here until all other
 741  *    nodes have released ack lock resource.
 742  * 5. Downconvert ack lockresource to CR
 743  */
 744 static int __sendmsg(struct md_cluster_info *cinfo, struct cluster_msg *cmsg)
 745 {
 746         int error;
 747         int slot = cinfo->slot_number - 1;
 748
 749         cmsg->slot = cpu_to_le32(slot);
 750         /*get EX on Message*/
 751         error = dlm_lock_sync(cinfo->message_lockres, DLM_LOCK_EX);
 752         if (error) {
 753                 pr_err("md-cluster: failed to get EX on MESSAGE (%d)\n", error);
 754                 goto failed_message;
 755         }
 756
 757         memcpy(cinfo->message_lockres->lksb.sb_lvbptr, (void *)cmsg,
 758                         sizeof(struct cluster_msg));
 759         /*down-convert EX to CW on Message*/
 760         error = dlm_lock_sync(cinfo->message_lockres, DLM_LOCK_CW);
 761         if (error) {
 762                 pr_err("md-cluster: failed to convert EX to CW on MESSAGE(%d)\n",
 763                                 error);
 764                 goto failed_ack;
 765         }
 766
 767         /*up-convert CR to EX on Ack*/
 768         error = dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_EX);
 769         if (error) {
 770                 pr_err("md-cluster: failed to convert CR to EX on ACK(%d)\n",
 771                                 error);
 772                 goto failed_ack;
 773         }
 774
 775         /*down-convert EX to CR on Ack*/
 776         error = dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_CR);
 777         if (error) {
 778                 pr_err("md-cluster: failed to convert EX to CR on ACK(%d)\n",
 779                                 error);
 780                 goto failed_ack;
 781         }
 782
 783 failed_ack:
 784         error = dlm_unlock_sync(cinfo->message_lockres);
 785         if (unlikely(error != 0)) {
 786                 pr_err("md-cluster: failed convert to NL on MESSAGE(%d)\n",
 787                         error);
 788                 /* in case the message can't be released due to some reason */
 789                 goto failed_ack;
 790         }
 791 failed_message:
 792         return error;
 793 }
 794
 795 static int sendmsg(struct md_cluster_info *cinfo, struct cluster_msg *cmsg,
 796                    bool mddev_locked)
 797 {
 798         int ret;
 799
 800         ret = lock_comm(cinfo, mddev_locked);
 801         if (!ret) {
 802                 ret = __sendmsg(cinfo, cmsg);
 803                 unlock_comm(cinfo);
 804         }
 805         return ret;
 806 }
 807
 808 static int gather_all_resync_info(struct mddev *mddev, int total_slots)
 809 {
 810         struct md_cluster_info *cinfo = mddev->cluster_info;
 811         int i, ret = 0;
 812         struct dlm_lock_resource *bm_lockres;
 813         char str[64];
 814         sector_t lo, hi;
 815
 816
 817         for (i = 0; i < total_slots; i++) {
 818                 memset(str, '\0', 64);
 819                 snprintf(str, 64, "bitmap%04d", i);
 820                 bm_lockres = lockres_init(mddev, str, NULL, 1);
 821                 if (!bm_lockres)
 822                         return -ENOMEM;
 823                 if (i == (cinfo->slot_number - 1)) {
 824                         lockres_free(bm_lockres);
 825                         continue;
 826                 }
 827
 828                 bm_lockres->flags |= DLM_LKF_NOQUEUE;
 829                 ret = dlm_lock_sync(bm_lockres, DLM_LOCK_PW);
 830                 if (ret == -EAGAIN) {
 831                         if (read_resync_info(mddev, bm_lockres)) {
 832                                 pr_info("%s:%d Resync[%llu..%llu] in progress on %d\n",
 833                                                 __func__, __LINE__,
 834                                         (unsigned long long) cinfo->suspend_lo,
 835                                         (unsigned long long) cinfo->suspend_hi,
 836                                         i);
 837                                 cinfo->suspend_from = i;
 838                         }
 839                         ret = 0;
 840                         lockres_free(bm_lockres);
 841                         continue;
 842                 }
 843                 if (ret) {
 844                         lockres_free(bm_lockres);
 845                         goto out;
 846                 }
 847
 848                 /* Read the disk bitmap sb and check if it needs recovery */
 849                 ret = md_bitmap_copy_from_slot(mddev, i, &lo, &hi, false);
 850                 if (ret) {
 851                         pr_warn("md-cluster: Could not gather bitmaps from slot %d", i);
 852                         lockres_free(bm_lockres);
 853                         continue;
 854                 }
 855                 if ((hi > 0) && (lo < mddev->recovery_cp)) {
 856                         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
 857                         mddev->recovery_cp = lo;
 858                         md_check_recovery(mddev);
 859                 }
 860
 861                 lockres_free(bm_lockres);
 862         }
 863 out:
 864         return ret;
 865 }
 866
 867 static int join(struct mddev *mddev, int nodes)
 868 {
 869         struct md_cluster_info *cinfo;
 870         int ret, ops_rv;
 871         char str[64];
 872
 873         cinfo = kzalloc(sizeof(struct md_cluster_info), GFP_KERNEL);
 874         if (!cinfo)
 875                 return -ENOMEM;
 876
 877         INIT_LIST_HEAD(&cinfo->suspend_list);
 878         spin_lock_init(&cinfo->suspend_lock);
 879         init_completion(&cinfo->completion);
 880         set_bit(MD_CLUSTER_BEGIN_JOIN_CLUSTER, &cinfo->state);
 881         init_waitqueue_head(&cinfo->wait);
 882         mutex_init(&cinfo->recv_mutex);
 883
 884         mddev->cluster_info = cinfo;
 885         cinfo->mddev = mddev;
 886
 887         memset(str, 0, 64);
 888         sprintf(str, "%pU", mddev->uuid);
 889         ret = dlm_new_lockspace(str, mddev->bitmap_info.cluster_name,
 890                                 0, LVB_SIZE, &md_ls_ops, mddev,
 891                                 &ops_rv, &cinfo->lockspace);
 892         if (ret)
 893                 goto err;
 894         wait_for_completion(&cinfo->completion);
 895         if (nodes < cinfo->slot_number) {
 896                 pr_err("md-cluster: Slot allotted(%d) is greater than available slots(%d).",
 897                         cinfo->slot_number, nodes);
 898                 ret = -ERANGE;
 899                 goto err;
 900         }
 901         /* Initiate the communication resources */
 902         ret = -ENOMEM;
 903         rcu_assign_pointer(cinfo->recv_thread,
 904                         md_register_thread(recv_daemon, mddev, "cluster_recv"));
 905         if (!cinfo->recv_thread) {
 906                 pr_err("md-cluster: cannot allocate memory for recv_thread!\n");
 907                 goto err;
 908         }
 909         cinfo->message_lockres = lockres_init(mddev, "message", NULL, 1);
 910         if (!cinfo->message_lockres)
 911                 goto err;
 912         cinfo->token_lockres = lockres_init(mddev, "token", NULL, 0);
 913         if (!cinfo->token_lockres)
 914                 goto err;
 915         cinfo->no_new_dev_lockres = lockres_init(mddev, "no-new-dev", NULL, 0);
 916         if (!cinfo->no_new_dev_lockres)
 917                 goto err;
 918
 919         ret = dlm_lock_sync(cinfo->token_lockres, DLM_LOCK_EX);
 920         if (ret) {
 921                 ret = -EAGAIN;
 922                 pr_err("md-cluster: can't join cluster to avoid lock issue\n");
 923                 goto err;
 924         }
 925         cinfo->ack_lockres = lockres_init(mddev, "ack", ack_bast, 0);
 926         if (!cinfo->ack_lockres) {
 927                 ret = -ENOMEM;
 928                 goto err;
 929         }
 930         /* get sync CR lock on ACK. */
 931         if (dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_CR))
 932                 pr_err("md-cluster: failed to get a sync CR lock on ACK!(%d)\n",
 933                                 ret);
 934         dlm_unlock_sync(cinfo->token_lockres);
 935         /* get sync CR lock on no-new-dev. */
 936         if (dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR))
 937                 pr_err("md-cluster: failed to get a sync CR lock on no-new-dev!(%d)\n", ret);
 938
 939
 940         pr_info("md-cluster: Joined cluster %s slot %d\n", str, cinfo->slot_number);
 941         snprintf(str, 64, "bitmap%04d", cinfo->slot_number - 1);
 942         cinfo->bitmap_lockres = lockres_init(mddev, str, NULL, 1);
 943         if (!cinfo->bitmap_lockres) {
 944                 ret = -ENOMEM;
 945                 goto err;
 946         }
 947         if (dlm_lock_sync(cinfo->bitmap_lockres, DLM_LOCK_PW)) {
 948                 pr_err("Failed to get bitmap lock\n");
 949                 ret = -EINVAL;
 950                 goto err;
 951         }
 952
 953         cinfo->resync_lockres = lockres_init(mddev, "resync", NULL, 0);
 954         if (!cinfo->resync_lockres) {
 955                 ret = -ENOMEM;
 956                 goto err;
 957         }
 958
 959         return 0;
 960 err:
 961         set_bit(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state);
 962         md_unregister_thread(mddev, &cinfo->recovery_thread);
 963         md_unregister_thread(mddev, &cinfo->recv_thread);
 964         lockres_free(cinfo->message_lockres);
 965         lockres_free(cinfo->token_lockres);
 966         lockres_free(cinfo->ack_lockres);
 967         lockres_free(cinfo->no_new_dev_lockres);
 968         lockres_free(cinfo->resync_lockres);
 969         lockres_free(cinfo->bitmap_lockres);
 970         if (cinfo->lockspace)
 971                 dlm_release_lockspace(cinfo->lockspace, 2);
 972         mddev->cluster_info = NULL;
 973         kfree(cinfo);
 974         return ret;
 975 }
 976
 977 static void load_bitmaps(struct mddev *mddev, int total_slots)
 978 {
 979         struct md_cluster_info *cinfo = mddev->cluster_info;
 980
 981         /* load all the node's bitmap info for resync */
 982         if (gather_all_resync_info(mddev, total_slots))
 983                 pr_err("md-cluster: failed to gather all resyn infos\n");
 984         set_bit(MD_CLUSTER_ALREADY_IN_CLUSTER, &cinfo->state);
 985         /* wake up recv thread in case something need to be handled */
 986         if (test_and_clear_bit(MD_CLUSTER_PENDING_RECV_EVENT, &cinfo->state))
 987                 md_wakeup_thread(cinfo->recv_thread);
 988 }
 989
 990 static void resync_bitmap(struct mddev *mddev)
 991 {
 992         struct md_cluster_info *cinfo = mddev->cluster_info;
 993         struct cluster_msg cmsg = {0};
 994         int err;
 995
 996         cmsg.type = cpu_to_le32(BITMAP_NEEDS_SYNC);
 997         err = sendmsg(cinfo, &cmsg, 1);
 998         if (err)
 999                 pr_err("%s:%d: failed to send BITMAP_NEEDS_SYNC message (%d)\n",
1000                         __func__, __LINE__, err);
1001 }
1002
1003 static void unlock_all_bitmaps(struct mddev *mddev);
1004 static int leave(struct mddev *mddev)
1005 {
1006         struct md_cluster_info *cinfo = mddev->cluster_info;
1007
1008         if (!cinfo)
1009                 return 0;
1010
1011         /*
1012          * BITMAP_NEEDS_SYNC message should be sent when node
1013          * is leaving the cluster with dirty bitmap, also we
1014          * can only deliver it when dlm connection is available.
1015          *
1016          * Also, we should send BITMAP_NEEDS_SYNC message in
1017          * case reshaping is interrupted.
1018          */
1019         if ((cinfo->slot_number > 0 && mddev->recovery_cp != MaxSector) ||
1020             (mddev->reshape_position != MaxSector &&
1021              test_bit(MD_CLOSING, &mddev->flags)))
1022                 resync_bitmap(mddev);
1023
1024         set_bit(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state);
1025         md_unregister_thread(mddev, &cinfo->recovery_thread);
1026         md_unregister_thread(mddev, &cinfo->recv_thread);
1027         lockres_free(cinfo->message_lockres);
1028         lockres_free(cinfo->token_lockres);
1029         lockres_free(cinfo->ack_lockres);
1030         lockres_free(cinfo->no_new_dev_lockres);
1031         lockres_free(cinfo->resync_lockres);
1032         lockres_free(cinfo->bitmap_lockres);
1033         unlock_all_bitmaps(mddev);
1034         dlm_release_lockspace(cinfo->lockspace, 2);
1035         kfree(cinfo);
1036         return 0;
1037 }
1038
1039 /* slot_number(): Returns the MD slot number to use
1040  * DLM starts the slot numbers from 1, wheras cluster-md
1041  * wants the number to be from zero, so we deduct one
1042  */
1043 static int slot_number(struct mddev *mddev)
1044 {
1045         struct md_cluster_info *cinfo = mddev->cluster_info;
1046
1047         return cinfo->slot_number - 1;
1048 }
1049
1050 /*
1051  * Check if the communication is already locked, else lock the communication
1052  * channel.
1053  * If it is already locked, token is in EX mode, and hence lock_token()
1054  * should not be called.
1055  */
1056 static int metadata_update_start(struct mddev *mddev)
1057 {
1058         struct md_cluster_info *cinfo = mddev->cluster_info;
1059         int ret;
1060
1061         /*
1062          * metadata_update_start is always called with the protection of
1063          * reconfig_mutex, so set WAITING_FOR_TOKEN here.
1064          */
1065         ret = test_and_set_bit_lock(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD,
1066                                     &cinfo->state);
1067         WARN_ON_ONCE(ret);
1068         md_wakeup_thread(mddev->thread);
1069
1070         wait_event(cinfo->wait,
1071                    !test_and_set_bit(MD_CLUSTER_SEND_LOCK, &cinfo->state) ||
1072                    test_and_clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state));
1073
1074         /* If token is already locked, return 0 */
1075         if (cinfo->token_lockres->mode == DLM_LOCK_EX) {
1076                 clear_bit_unlock(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state);
1077                 return 0;
1078         }
1079
1080         ret = lock_token(cinfo);
1081         clear_bit_unlock(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state);
1082         return ret;
1083 }
1084
1085 static int metadata_update_finish(struct mddev *mddev)
1086 {
1087         struct md_cluster_info *cinfo = mddev->cluster_info;
1088         struct cluster_msg cmsg;
1089         struct md_rdev *rdev;
1090         int ret = 0;
1091         int raid_slot = -1;
1092
1093         memset(&cmsg, 0, sizeof(cmsg));
1094         cmsg.type = cpu_to_le32(METADATA_UPDATED);
1095         /* Pick up a good active device number to send.
1096          */
1097         rdev_for_each(rdev, mddev)
1098                 if (rdev->raid_disk > -1 && !test_bit(Faulty, &rdev->flags)) {
1099                         raid_slot = rdev->desc_nr;
1100                         break;
1101                 }
1102         if (raid_slot >= 0) {
1103                 cmsg.raid_slot = cpu_to_le32(raid_slot);
1104                 ret = __sendmsg(cinfo, &cmsg);
1105         } else
1106                 pr_warn("md-cluster: No good device id found to send\n");
1107         clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1108         unlock_comm(cinfo);
1109         return ret;
1110 }
1111
1112 static void metadata_update_cancel(struct mddev *mddev)
1113 {
1114         struct md_cluster_info *cinfo = mddev->cluster_info;
1115         clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1116         unlock_comm(cinfo);
1117 }
1118
1119 static int update_bitmap_size(struct mddev *mddev, sector_t size)
1120 {
1121         struct md_cluster_info *cinfo = mddev->cluster_info;
1122         struct cluster_msg cmsg = {0};
1123         int ret;
1124
1125         cmsg.type = cpu_to_le32(BITMAP_RESIZE);
1126         cmsg.high = cpu_to_le64(size);
1127         ret = sendmsg(cinfo, &cmsg, 0);
1128         if (ret)
1129                 pr_err("%s:%d: failed to send BITMAP_RESIZE message (%d)\n",
1130                         __func__, __LINE__, ret);
1131         return ret;
1132 }
1133
1134 static int resize_bitmaps(struct mddev *mddev, sector_t newsize, sector_t oldsize)
1135 {
1136         struct bitmap_counts *counts;
1137         char str[64];
1138         struct dlm_lock_resource *bm_lockres;
1139         struct bitmap *bitmap = mddev->bitmap;
1140         unsigned long my_pages = bitmap->counts.pages;
1141         int i, rv;
1142
1143         /*
1144          * We need to ensure all the nodes can grow to a larger
1145          * bitmap size before make the reshaping.
1146          */
1147         rv = update_bitmap_size(mddev, newsize);
1148         if (rv)
1149                 return rv;
1150
1151         for (i = 0; i < mddev->bitmap_info.nodes; i++) {
1152                 if (i == md_cluster_ops->slot_number(mddev))
1153                         continue;
1154
1155                 bitmap = get_bitmap_from_slot(mddev, i);
1156                 if (IS_ERR(bitmap)) {
1157                         pr_err("can't get bitmap from slot %d\n", i);
1158                         bitmap = NULL;
1159                         goto out;
1160                 }
1161                 counts = &bitmap->counts;
1162
1163                 /*
1164                  * If we can hold the bitmap lock of one node then
1165                  * the slot is not occupied, update the pages.
1166                  */
1167                 snprintf(str, 64, "bitmap%04d", i);
1168                 bm_lockres = lockres_init(mddev, str, NULL, 1);
1169                 if (!bm_lockres) {
1170                         pr_err("Cannot initialize %s lock\n", str);
1171                         goto out;
1172                 }
1173                 bm_lockres->flags |= DLM_LKF_NOQUEUE;
1174                 rv = dlm_lock_sync(bm_lockres, DLM_LOCK_PW);
1175                 if (!rv)
1176                         counts->pages = my_pages;
1177                 lockres_free(bm_lockres);
1178
1179                 if (my_pages != counts->pages)
1180                         /*
1181                          * Let's revert the bitmap size if one node
1182                          * can't resize bitmap
1183                          */
1184                         goto out;
1185                 md_bitmap_free(bitmap);
1186         }
1187
1188         return 0;
1189 out:
1190         md_bitmap_free(bitmap);
1191         update_bitmap_size(mddev, oldsize);
1192         return -1;
1193 }
1194
1195 /*
1196  * return 0 if all the bitmaps have the same sync_size
1197  */
1198 static int cluster_check_sync_size(struct mddev *mddev)
1199 {
1200         int i, rv;
1201         bitmap_super_t *sb;
1202         unsigned long my_sync_size, sync_size = 0;
1203         int node_num = mddev->bitmap_info.nodes;
1204         int current_slot = md_cluster_ops->slot_number(mddev);
1205         struct bitmap *bitmap = mddev->bitmap;
1206         char str[64];
1207         struct dlm_lock_resource *bm_lockres;
1208
1209         sb = kmap_atomic(bitmap->storage.sb_page);
1210         my_sync_size = sb->sync_size;
1211         kunmap_atomic(sb);
1212
1213         for (i = 0; i < node_num; i++) {
1214                 if (i == current_slot)
1215                         continue;
1216
1217                 bitmap = get_bitmap_from_slot(mddev, i);
1218                 if (IS_ERR(bitmap)) {
1219                         pr_err("can't get bitmap from slot %d\n", i);
1220                         return -1;
1221                 }
1222
1223                 /*
1224                  * If we can hold the bitmap lock of one node then
1225                  * the slot is not occupied, update the sb.
1226                  */
1227                 snprintf(str, 64, "bitmap%04d", i);
1228                 bm_lockres = lockres_init(mddev, str, NULL, 1);
1229                 if (!bm_lockres) {
1230                         pr_err("md-cluster: Cannot initialize %s\n", str);
1231                         md_bitmap_free(bitmap);
1232                         return -1;
1233                 }
1234                 bm_lockres->flags |= DLM_LKF_NOQUEUE;
1235                 rv = dlm_lock_sync(bm_lockres, DLM_LOCK_PW);
1236                 if (!rv)
1237                         md_bitmap_update_sb(bitmap);
1238                 lockres_free(bm_lockres);
1239
1240                 sb = kmap_atomic(bitmap->storage.sb_page);
1241                 if (sync_size == 0)
1242                         sync_size = sb->sync_size;
1243                 else if (sync_size != sb->sync_size) {
1244                         kunmap_atomic(sb);
1245                         md_bitmap_free(bitmap);
1246                         return -1;
1247                 }
1248                 kunmap_atomic(sb);
1249                 md_bitmap_free(bitmap);
1250         }
1251
1252         return (my_sync_size == sync_size) ? 0 : -1;
1253 }
1254
1255 /*
1256  * Update the size for cluster raid is a little more complex, we perform it
1257  * by the steps:
1258  * 1. hold token lock and update superblock in initiator node.
1259  * 2. send METADATA_UPDATED msg to other nodes.
1260  * 3. The initiator node continues to check each bitmap's sync_size, if all
1261  *    bitmaps have the same value of sync_size, then we can set capacity and
1262  *    let other nodes to perform it. If one node can't update sync_size
1263  *    accordingly, we need to revert to previous value.
1264  */
1265 static void update_size(struct mddev *mddev, sector_t old_dev_sectors)
1266 {
1267         struct md_cluster_info *cinfo = mddev->cluster_info;
1268         struct cluster_msg cmsg;
1269         struct md_rdev *rdev;
1270         int ret = 0;
1271         int raid_slot = -1;
1272
1273         md_update_sb(mddev, 1);
1274         if (lock_comm(cinfo, 1)) {
1275                 pr_err("%s: lock_comm failed\n", __func__);
1276                 return;
1277         }
1278
1279         memset(&cmsg, 0, sizeof(cmsg));
1280         cmsg.type = cpu_to_le32(METADATA_UPDATED);
1281         rdev_for_each(rdev, mddev)
1282                 if (rdev->raid_disk >= 0 && !test_bit(Faulty, &rdev->flags)) {
1283                         raid_slot = rdev->desc_nr;
1284                         break;
1285                 }
1286         if (raid_slot >= 0) {
1287                 cmsg.raid_slot = cpu_to_le32(raid_slot);
1288                 /*
1289                  * We can only change capiticy after all the nodes can do it,
1290                  * so need to wait after other nodes already received the msg
1291                  * and handled the change
1292                  */
1293                 ret = __sendmsg(cinfo, &cmsg);
1294                 if (ret) {
1295                         pr_err("%s:%d: failed to send METADATA_UPDATED msg\n",
1296                                __func__, __LINE__);
1297                         unlock_comm(cinfo);
1298                         return;
1299                 }
1300         } else {
1301                 pr_err("md-cluster: No good device id found to send\n");
1302                 unlock_comm(cinfo);
1303                 return;
1304         }
1305
1306         /*
1307          * check the sync_size from other node's bitmap, if sync_size
1308          * have already updated in other nodes as expected, send an
1309          * empty metadata msg to permit the change of capacity
1310          */
1311         if (cluster_check_sync_size(mddev) == 0) {
1312                 memset(&cmsg, 0, sizeof(cmsg));
1313                 cmsg.type = cpu_to_le32(CHANGE_CAPACITY);
1314                 ret = __sendmsg(cinfo, &cmsg);
1315                 if (ret)
1316                         pr_err("%s:%d: failed to send CHANGE_CAPACITY msg\n",
1317                                __func__, __LINE__);
1318                 set_capacity_and_notify(mddev->gendisk, mddev->array_sectors);
1319         } else {
1320                 /* revert to previous sectors */
1321                 ret = mddev->pers->resize(mddev, old_dev_sectors);
1322                 ret = __sendmsg(cinfo, &cmsg);
1323                 if (ret)
1324                         pr_err("%s:%d: failed to send METADATA_UPDATED msg\n",
1325                                __func__, __LINE__);
1326         }
1327         unlock_comm(cinfo);
1328 }
1329
1330 static int resync_start(struct mddev *mddev)
1331 {
1332         struct md_cluster_info *cinfo = mddev->cluster_info;
1333         return dlm_lock_sync_interruptible(cinfo->resync_lockres, DLM_LOCK_EX, mddev);
1334 }
1335
1336 static void resync_info_get(struct mddev *mddev, sector_t *lo, sector_t *hi)
1337 {
1338         struct md_cluster_info *cinfo = mddev->cluster_info;
1339
1340         spin_lock_irq(&cinfo->suspend_lock);
1341         *lo = cinfo->suspend_lo;
1342         *hi = cinfo->suspend_hi;
1343         spin_unlock_irq(&cinfo->suspend_lock);
1344 }
1345
1346 static int resync_info_update(struct mddev *mddev, sector_t lo, sector_t hi)
1347 {
1348         struct md_cluster_info *cinfo = mddev->cluster_info;
1349         struct resync_info ri;
1350         struct cluster_msg cmsg = {0};
1351
1352         /* do not send zero again, if we have sent before */
1353         if (hi == 0) {
1354                 memcpy(&ri, cinfo->bitmap_lockres->lksb.sb_lvbptr, sizeof(struct resync_info));
1355                 if (le64_to_cpu(ri.hi) == 0)
1356                         return 0;
1357         }
1358
1359         add_resync_info(cinfo->bitmap_lockres, lo, hi);
1360         /* Re-acquire the lock to refresh LVB */
1361         dlm_lock_sync(cinfo->bitmap_lockres, DLM_LOCK_PW);
1362         cmsg.type = cpu_to_le32(RESYNCING);
1363         cmsg.low = cpu_to_le64(lo);
1364         cmsg.high = cpu_to_le64(hi);
1365
1366         /*
1367          * mddev_lock is held if resync_info_update is called from
1368          * resync_finish (md_reap_sync_thread -> resync_finish)
1369          */
1370         if (lo == 0 && hi == 0)
1371                 return sendmsg(cinfo, &cmsg, 1);
1372         else
1373                 return sendmsg(cinfo, &cmsg, 0);
1374 }
1375
1376 static int resync_finish(struct mddev *mddev)
1377 {
1378         struct md_cluster_info *cinfo = mddev->cluster_info;
1379         int ret = 0;
1380
1381         clear_bit(MD_RESYNCING_REMOTE, &mddev->recovery);
1382
1383         /*
1384          * If resync thread is interrupted so we can't say resync is finished,
1385          * another node will launch resync thread to continue.
1386          */
1387         if (!test_bit(MD_CLOSING, &mddev->flags))
1388                 ret = resync_info_update(mddev, 0, 0);
1389         dlm_unlock_sync(cinfo->resync_lockres);
1390         return ret;
1391 }
1392
1393 static int area_resyncing(struct mddev *mddev, int direction,
1394                 sector_t lo, sector_t hi)
1395 {
1396         struct md_cluster_info *cinfo = mddev->cluster_info;
1397         int ret = 0;
1398
1399         if ((direction == READ) &&
1400                 test_bit(MD_CLUSTER_SUSPEND_READ_BALANCING, &cinfo->state))
1401                 return 1;
1402
1403         spin_lock_irq(&cinfo->suspend_lock);
1404         if (hi > cinfo->suspend_lo && lo < cinfo->suspend_hi)
1405                 ret = 1;
1406         spin_unlock_irq(&cinfo->suspend_lock);
1407         return ret;
1408 }
1409
1410 /* add_new_disk() - initiates a disk add
1411  * However, if this fails before writing md_update_sb(),
1412  * add_new_disk_cancel() must be called to release token lock
1413  */
1414 static int add_new_disk(struct mddev *mddev, struct md_rdev *rdev)
1415 {
1416         struct md_cluster_info *cinfo = mddev->cluster_info;
1417         struct cluster_msg cmsg;
1418         int ret = 0;
1419         struct mdp_superblock_1 *sb = page_address(rdev->sb_page);
1420         char *uuid = sb->device_uuid;
1421
1422         memset(&cmsg, 0, sizeof(cmsg));
1423         cmsg.type = cpu_to_le32(NEWDISK);
1424         memcpy(cmsg.uuid, uuid, 16);
1425         cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
1426         if (lock_comm(cinfo, 1))
1427                 return -EAGAIN;
1428         ret = __sendmsg(cinfo, &cmsg);
1429         if (ret) {
1430                 unlock_comm(cinfo);
1431                 return ret;
1432         }
1433         cinfo->no_new_dev_lockres->flags |= DLM_LKF_NOQUEUE;
1434         ret = dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_EX);
1435         cinfo->no_new_dev_lockres->flags &= ~DLM_LKF_NOQUEUE;
1436         /* Some node does not "see" the device */
1437         if (ret == -EAGAIN)
1438                 ret = -ENOENT;
1439         if (ret)
1440                 unlock_comm(cinfo);
1441         else {
1442                 dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR);
1443                 /* Since MD_CHANGE_DEVS will be set in add_bound_rdev which
1444                  * will run soon after add_new_disk, the below path will be
1445                  * invoked:
1446                  *   md_wakeup_thread(mddev->thread)
1447                  *      -> conf->thread (raid1d)
1448                  *      -> md_check_recovery -> md_update_sb
1449                  *      -> metadata_update_start/finish
1450                  * MD_CLUSTER_SEND_LOCKED_ALREADY will be cleared eventually.
1451                  *
1452                  * For other failure cases, metadata_update_cancel and
1453                  * add_new_disk_cancel also clear below bit as well.
1454                  * */
1455                 set_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1456                 wake_up(&cinfo->wait);
1457         }
1458         return ret;
1459 }
1460
1461 static void add_new_disk_cancel(struct mddev *mddev)
1462 {
1463         struct md_cluster_info *cinfo = mddev->cluster_info;
1464         clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1465         unlock_comm(cinfo);
1466 }
1467
1468 static int new_disk_ack(struct mddev *mddev, bool ack)
1469 {
1470         struct md_cluster_info *cinfo = mddev->cluster_info;
1471
1472         if (!test_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state)) {
1473                 pr_warn("md-cluster(%s): Spurious cluster confirmation\n", mdname(mddev));
1474                 return -EINVAL;
1475         }
1476
1477         if (ack)
1478                 dlm_unlock_sync(cinfo->no_new_dev_lockres);
1479         complete(&cinfo->newdisk_completion);
1480         return 0;
1481 }
1482
1483 static int remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1484 {
1485         struct cluster_msg cmsg = {0};
1486         struct md_cluster_info *cinfo = mddev->cluster_info;
1487         cmsg.type = cpu_to_le32(REMOVE);
1488         cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
1489         return sendmsg(cinfo, &cmsg, 1);
1490 }
1491
1492 static int lock_all_bitmaps(struct mddev *mddev)
1493 {
1494         int slot, my_slot, ret, held = 1, i = 0;
1495         char str[64];
1496         struct md_cluster_info *cinfo = mddev->cluster_info;
1497
1498         cinfo->other_bitmap_lockres =
1499                 kcalloc(mddev->bitmap_info.nodes - 1,
1500                         sizeof(struct dlm_lock_resource *), GFP_KERNEL);
1501         if (!cinfo->other_bitmap_lockres) {
1502                 pr_err("md: can't alloc mem for other bitmap locks\n");
1503                 return 0;
1504         }
1505
1506         my_slot = slot_number(mddev);
1507         for (slot = 0; slot < mddev->bitmap_info.nodes; slot++) {
1508                 if (slot == my_slot)
1509                         continue;
1510
1511                 memset(str, '\0', 64);
1512                 snprintf(str, 64, "bitmap%04d", slot);
1513                 cinfo->other_bitmap_lockres[i] = lockres_init(mddev, str, NULL, 1);
1514                 if (!cinfo->other_bitmap_lockres[i])
1515                         return -ENOMEM;
1516
1517                 cinfo->other_bitmap_lockres[i]->flags |= DLM_LKF_NOQUEUE;
1518                 ret = dlm_lock_sync(cinfo->other_bitmap_lockres[i], DLM_LOCK_PW);
1519                 if (ret)
1520                         held = -1;
1521                 i++;
1522         }
1523
1524         return held;
1525 }
1526
1527 static void unlock_all_bitmaps(struct mddev *mddev)
1528 {
1529         struct md_cluster_info *cinfo = mddev->cluster_info;
1530         int i;
1531
1532         /* release other node's bitmap lock if they are existed */
1533         if (cinfo->other_bitmap_lockres) {
1534                 for (i = 0; i < mddev->bitmap_info.nodes - 1; i++) {
1535                         if (cinfo->other_bitmap_lockres[i]) {
1536                                 lockres_free(cinfo->other_bitmap_lockres[i]);
1537                         }
1538                 }
1539                 kfree(cinfo->other_bitmap_lockres);
1540                 cinfo->other_bitmap_lockres = NULL;
1541         }
1542 }
1543
1544 static int gather_bitmaps(struct md_rdev *rdev)
1545 {
1546         int sn, err;
1547         sector_t lo, hi;
1548         struct cluster_msg cmsg = {0};
1549         struct mddev *mddev = rdev->mddev;
1550         struct md_cluster_info *cinfo = mddev->cluster_info;
1551
1552         cmsg.type = cpu_to_le32(RE_ADD);
1553         cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
1554         err = sendmsg(cinfo, &cmsg, 1);
1555         if (err)
1556                 goto out;
1557
1558         for (sn = 0; sn < mddev->bitmap_info.nodes; sn++) {
1559                 if (sn == (cinfo->slot_number - 1))
1560                         continue;
1561                 err = md_bitmap_copy_from_slot(mddev, sn, &lo, &hi, false);
1562                 if (err) {
1563                         pr_warn("md-cluster: Could not gather bitmaps from slot %d", sn);
1564                         goto out;
1565                 }
1566                 if ((hi > 0) && (lo < mddev->recovery_cp))
1567                         mddev->recovery_cp = lo;
1568         }
1569 out:
1570         return err;
1571 }
1572
1573 static struct md_cluster_operations cluster_ops = {
1574         .join   = join,
1575         .leave  = leave,
1576         .slot_number = slot_number,
1577         .resync_start = resync_start,
1578         .resync_finish = resync_finish,
1579         .resync_info_update = resync_info_update,
1580         .resync_info_get = resync_info_get,
1581         .metadata_update_start = metadata_update_start,
1582         .metadata_update_finish = metadata_update_finish,
1583         .metadata_update_cancel = metadata_update_cancel,
1584         .area_resyncing = area_resyncing,
1585         .add_new_disk = add_new_disk,
1586         .add_new_disk_cancel = add_new_disk_cancel,
1587         .new_disk_ack = new_disk_ack,
1588         .remove_disk = remove_disk,
1589         .load_bitmaps = load_bitmaps,
1590         .gather_bitmaps = gather_bitmaps,
1591         .resize_bitmaps = resize_bitmaps,
1592         .lock_all_bitmaps = lock_all_bitmaps,
1593         .unlock_all_bitmaps = unlock_all_bitmaps,
1594         .update_size = update_size,
1595 };
1596
1597 static int __init cluster_init(void)
1598 {
1599         pr_warn("md-cluster: support raid1 and raid10 (limited support)\n");
1600         pr_info("Registering Cluster MD functions\n");
1601         register_md_cluster_operations(&cluster_ops, THIS_MODULE);
1602         return 0;
1603 }
1604
1605 static void cluster_exit(void)
1606 {
1607         unregister_md_cluster_operations();
1608 }
1609
1610 module_init(cluster_init);
1611 module_exit(cluster_exit);
1612 MODULE_AUTHOR("SUSE");
1613 MODULE_LICENSE("GPL");
1614 MODULE_DESCRIPTION("Clustering support for MD");