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 void process_add_new_disk(struct mddev *mddev, struct cluster_msg *cmsg)
 505 {
 506         char disk_uuid[64];
 507         struct md_cluster_info *cinfo = mddev->cluster_info;
 508         char event_name[] = "EVENT=ADD_DEVICE";
 509         char raid_slot[16];
 510         char *envp[] = {event_name, disk_uuid, raid_slot, NULL};
 511         int len;
 512
 513         len = snprintf(disk_uuid, 64, "DEVICE_UUID=");
 514         sprintf(disk_uuid + len, "%pU", cmsg->uuid);
 515         snprintf(raid_slot, 16, "RAID_DISK=%d", le32_to_cpu(cmsg->raid_slot));
 516         pr_info("%s:%d Sending kobject change with %s and %s\n", __func__, __LINE__, disk_uuid, raid_slot);
 517         init_completion(&cinfo->newdisk_completion);
 518         set_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state);
 519         kobject_uevent_env(&disk_to_dev(mddev->gendisk)->kobj, KOBJ_CHANGE, envp);
 520         wait_for_completion_timeout(&cinfo->newdisk_completion,
 521                         NEW_DEV_TIMEOUT);
 522         clear_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state);
 523 }
 524
 525
 526 static void process_metadata_update(struct mddev *mddev, struct cluster_msg *msg)
 527 {
 528         int got_lock = 0;
 529         struct md_thread *thread;
 530         struct md_cluster_info *cinfo = mddev->cluster_info;
 531         mddev->good_device_nr = le32_to_cpu(msg->raid_slot);
 532
 533         dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR);
 534
 535         /* daemaon thread must exist */
 536         thread = rcu_dereference_protected(mddev->thread, true);
 537         wait_event(thread->wqueue,
 538                    (got_lock = mddev_trylock(mddev)) ||
 539                     test_bit(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state));
 540         md_reload_sb(mddev, mddev->good_device_nr);
 541         if (got_lock)
 542                 mddev_unlock(mddev);
 543 }
 544
 545 static void process_remove_disk(struct mddev *mddev, struct cluster_msg *msg)
 546 {
 547         struct md_rdev *rdev;
 548
 549         rcu_read_lock();
 550         rdev = md_find_rdev_nr_rcu(mddev, le32_to_cpu(msg->raid_slot));
 551         if (rdev) {
 552                 set_bit(ClusterRemove, &rdev->flags);
 553                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
 554                 md_wakeup_thread(mddev->thread);
 555         }
 556         else
 557                 pr_warn("%s: %d Could not find disk(%d) to REMOVE\n",
 558                         __func__, __LINE__, le32_to_cpu(msg->raid_slot));
 559         rcu_read_unlock();
 560 }
 561
 562 static void process_readd_disk(struct mddev *mddev, struct cluster_msg *msg)
 563 {
 564         struct md_rdev *rdev;
 565
 566         rcu_read_lock();
 567         rdev = md_find_rdev_nr_rcu(mddev, le32_to_cpu(msg->raid_slot));
 568         if (rdev && test_bit(Faulty, &rdev->flags))
 569                 clear_bit(Faulty, &rdev->flags);
 570         else
 571                 pr_warn("%s: %d Could not find disk(%d) which is faulty",
 572                         __func__, __LINE__, le32_to_cpu(msg->raid_slot));
 573         rcu_read_unlock();
 574 }
 575
 576 static int process_recvd_msg(struct mddev *mddev, struct cluster_msg *msg)
 577 {
 578         int ret = 0;
 579
 580         if (WARN(mddev->cluster_info->slot_number - 1 == le32_to_cpu(msg->slot),
 581                 "node %d received its own msg\n", le32_to_cpu(msg->slot)))
 582                 return -1;
 583         switch (le32_to_cpu(msg->type)) {
 584         case METADATA_UPDATED:
 585                 process_metadata_update(mddev, msg);
 586                 break;
 587         case CHANGE_CAPACITY:
 588                 set_capacity_and_notify(mddev->gendisk, mddev->array_sectors);
 589                 break;
 590         case RESYNCING:
 591                 set_bit(MD_RESYNCING_REMOTE, &mddev->recovery);
 592                 process_suspend_info(mddev, le32_to_cpu(msg->slot),
 593                                      le64_to_cpu(msg->low),
 594                                      le64_to_cpu(msg->high));
 595                 break;
 596         case NEWDISK:
 597                 process_add_new_disk(mddev, msg);
 598                 break;
 599         case REMOVE:
 600                 process_remove_disk(mddev, msg);
 601                 break;
 602         case RE_ADD:
 603                 process_readd_disk(mddev, msg);
 604                 break;
 605         case BITMAP_NEEDS_SYNC:
 606                 __recover_slot(mddev, le32_to_cpu(msg->slot));
 607                 break;
 608         case BITMAP_RESIZE:
 609                 if (le64_to_cpu(msg->high) != mddev->pers->size(mddev, 0, 0))
 610                         ret = md_bitmap_resize(mddev->bitmap,
 611                                             le64_to_cpu(msg->high), 0, 0);
 612                 break;
 613         default:
 614                 ret = -1;
 615                 pr_warn("%s:%d Received unknown message from %d\n",
 616                         __func__, __LINE__, msg->slot);
 617         }
 618         return ret;
 619 }
 620
 621 /*
 622  * thread for receiving message
 623  */
 624 static void recv_daemon(struct md_thread *thread)
 625 {
 626         struct md_cluster_info *cinfo = thread->mddev->cluster_info;
 627         struct dlm_lock_resource *ack_lockres = cinfo->ack_lockres;
 628         struct dlm_lock_resource *message_lockres = cinfo->message_lockres;
 629         struct cluster_msg msg;
 630         int ret;
 631
 632         mutex_lock(&cinfo->recv_mutex);
 633         /*get CR on Message*/
 634         if (dlm_lock_sync(message_lockres, DLM_LOCK_CR)) {
 635                 pr_err("md/raid1:failed to get CR on MESSAGE\n");
 636                 mutex_unlock(&cinfo->recv_mutex);
 637                 return;
 638         }
 639
 640         /* read lvb and wake up thread to process this message_lockres */
 641         memcpy(&msg, message_lockres->lksb.sb_lvbptr, sizeof(struct cluster_msg));
 642         ret = process_recvd_msg(thread->mddev, &msg);
 643         if (ret)
 644                 goto out;
 645
 646         /*release CR on ack_lockres*/
 647         ret = dlm_unlock_sync(ack_lockres);
 648         if (unlikely(ret != 0))
 649                 pr_info("unlock ack failed return %d\n", ret);
 650         /*up-convert to PR on message_lockres*/
 651         ret = dlm_lock_sync(message_lockres, DLM_LOCK_PR);
 652         if (unlikely(ret != 0))
 653                 pr_info("lock PR on msg failed return %d\n", ret);
 654         /*get CR on ack_lockres again*/
 655         ret = dlm_lock_sync(ack_lockres, DLM_LOCK_CR);
 656         if (unlikely(ret != 0))
 657                 pr_info("lock CR on ack failed return %d\n", ret);
 658 out:
 659         /*release CR on message_lockres*/
 660         ret = dlm_unlock_sync(message_lockres);
 661         if (unlikely(ret != 0))
 662                 pr_info("unlock msg failed return %d\n", ret);
 663         mutex_unlock(&cinfo->recv_mutex);
 664 }
 665
 666 /* lock_token()
 667  * Takes the lock on the TOKEN lock resource so no other
 668  * node can communicate while the operation is underway.
 669  */
 670 static int lock_token(struct md_cluster_info *cinfo)
 671 {
 672         int error;
 673
 674         error = dlm_lock_sync(cinfo->token_lockres, DLM_LOCK_EX);
 675         if (error) {
 676                 pr_err("md-cluster(%s:%d): failed to get EX on TOKEN (%d)\n",
 677                                 __func__, __LINE__, error);
 678         } else {
 679                 /* Lock the receive sequence */
 680                 mutex_lock(&cinfo->recv_mutex);
 681         }
 682         return error;
 683 }
 684
 685 /* lock_comm()
 686  * Sets the MD_CLUSTER_SEND_LOCK bit to lock the send channel.
 687  */
 688 static int lock_comm(struct md_cluster_info *cinfo, bool mddev_locked)
 689 {
 690         int rv, set_bit = 0;
 691         struct mddev *mddev = cinfo->mddev;
 692
 693         /*
 694          * If resync thread run after raid1d thread, then process_metadata_update
 695          * could not continue if raid1d held reconfig_mutex (and raid1d is blocked
 696          * since another node already got EX on Token and waiting the EX of Ack),
 697          * so let resync wake up thread in case flag is set.
 698          */
 699         if (mddev_locked && !test_bit(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD,
 700                                       &cinfo->state)) {
 701                 rv = test_and_set_bit_lock(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD,
 702                                               &cinfo->state);
 703                 WARN_ON_ONCE(rv);
 704                 md_wakeup_thread(mddev->thread);
 705                 set_bit = 1;
 706         }
 707
 708         wait_event(cinfo->wait,
 709                    !test_and_set_bit(MD_CLUSTER_SEND_LOCK, &cinfo->state));
 710         rv = lock_token(cinfo);
 711         if (set_bit)
 712                 clear_bit_unlock(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state);
 713         return rv;
 714 }
 715
 716 static void unlock_comm(struct md_cluster_info *cinfo)
 717 {
 718         WARN_ON(cinfo->token_lockres->mode != DLM_LOCK_EX);
 719         mutex_unlock(&cinfo->recv_mutex);
 720         dlm_unlock_sync(cinfo->token_lockres);
 721         clear_bit(MD_CLUSTER_SEND_LOCK, &cinfo->state);
 722         wake_up(&cinfo->wait);
 723 }
 724
 725 /* __sendmsg()
 726  * This function performs the actual sending of the message. This function is
 727  * usually called after performing the encompassing operation
 728  * The function:
 729  * 1. Grabs the message lockresource in EX mode
 730  * 2. Copies the message to the message LVB
 731  * 3. Downconverts message lockresource to CW
 732  * 4. Upconverts ack lock resource from CR to EX. This forces the BAST on other nodes
 733  *    and the other nodes read the message. The thread will wait here until all other
 734  *    nodes have released ack lock resource.
 735  * 5. Downconvert ack lockresource to CR
 736  */
 737 static int __sendmsg(struct md_cluster_info *cinfo, struct cluster_msg *cmsg)
 738 {
 739         int error;
 740         int slot = cinfo->slot_number - 1;
 741
 742         cmsg->slot = cpu_to_le32(slot);
 743         /*get EX on Message*/
 744         error = dlm_lock_sync(cinfo->message_lockres, DLM_LOCK_EX);
 745         if (error) {
 746                 pr_err("md-cluster: failed to get EX on MESSAGE (%d)\n", error);
 747                 goto failed_message;
 748         }
 749
 750         memcpy(cinfo->message_lockres->lksb.sb_lvbptr, (void *)cmsg,
 751                         sizeof(struct cluster_msg));
 752         /*down-convert EX to CW on Message*/
 753         error = dlm_lock_sync(cinfo->message_lockres, DLM_LOCK_CW);
 754         if (error) {
 755                 pr_err("md-cluster: failed to convert EX to CW on MESSAGE(%d)\n",
 756                                 error);
 757                 goto failed_ack;
 758         }
 759
 760         /*up-convert CR to EX on Ack*/
 761         error = dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_EX);
 762         if (error) {
 763                 pr_err("md-cluster: failed to convert CR to EX on ACK(%d)\n",
 764                                 error);
 765                 goto failed_ack;
 766         }
 767
 768         /*down-convert EX to CR on Ack*/
 769         error = dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_CR);
 770         if (error) {
 771                 pr_err("md-cluster: failed to convert EX to CR on ACK(%d)\n",
 772                                 error);
 773                 goto failed_ack;
 774         }
 775
 776 failed_ack:
 777         error = dlm_unlock_sync(cinfo->message_lockres);
 778         if (unlikely(error != 0)) {
 779                 pr_err("md-cluster: failed convert to NL on MESSAGE(%d)\n",
 780                         error);
 781                 /* in case the message can't be released due to some reason */
 782                 goto failed_ack;
 783         }
 784 failed_message:
 785         return error;
 786 }
 787
 788 static int sendmsg(struct md_cluster_info *cinfo, struct cluster_msg *cmsg,
 789                    bool mddev_locked)
 790 {
 791         int ret;
 792
 793         ret = lock_comm(cinfo, mddev_locked);
 794         if (!ret) {
 795                 ret = __sendmsg(cinfo, cmsg);
 796                 unlock_comm(cinfo);
 797         }
 798         return ret;
 799 }
 800
 801 static int gather_all_resync_info(struct mddev *mddev, int total_slots)
 802 {
 803         struct md_cluster_info *cinfo = mddev->cluster_info;
 804         int i, ret = 0;
 805         struct dlm_lock_resource *bm_lockres;
 806         char str[64];
 807         sector_t lo, hi;
 808
 809
 810         for (i = 0; i < total_slots; i++) {
 811                 memset(str, '\0', 64);
 812                 snprintf(str, 64, "bitmap%04d", i);
 813                 bm_lockres = lockres_init(mddev, str, NULL, 1);
 814                 if (!bm_lockres)
 815                         return -ENOMEM;
 816                 if (i == (cinfo->slot_number - 1)) {
 817                         lockres_free(bm_lockres);
 818                         continue;
 819                 }
 820
 821                 bm_lockres->flags |= DLM_LKF_NOQUEUE;
 822                 ret = dlm_lock_sync(bm_lockres, DLM_LOCK_PW);
 823                 if (ret == -EAGAIN) {
 824                         if (read_resync_info(mddev, bm_lockres)) {
 825                                 pr_info("%s:%d Resync[%llu..%llu] in progress on %d\n",
 826                                                 __func__, __LINE__,
 827                                         (unsigned long long) cinfo->suspend_lo,
 828                                         (unsigned long long) cinfo->suspend_hi,
 829                                         i);
 830                                 cinfo->suspend_from = i;
 831                         }
 832                         ret = 0;
 833                         lockres_free(bm_lockres);
 834                         continue;
 835                 }
 836                 if (ret) {
 837                         lockres_free(bm_lockres);
 838                         goto out;
 839                 }
 840
 841                 /* Read the disk bitmap sb and check if it needs recovery */
 842                 ret = md_bitmap_copy_from_slot(mddev, i, &lo, &hi, false);
 843                 if (ret) {
 844                         pr_warn("md-cluster: Could not gather bitmaps from slot %d", i);
 845                         lockres_free(bm_lockres);
 846                         continue;
 847                 }
 848                 if ((hi > 0) && (lo < mddev->recovery_cp)) {
 849                         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
 850                         mddev->recovery_cp = lo;
 851                         md_check_recovery(mddev);
 852                 }
 853
 854                 lockres_free(bm_lockres);
 855         }
 856 out:
 857         return ret;
 858 }
 859
 860 static int join(struct mddev *mddev, int nodes)
 861 {
 862         struct md_cluster_info *cinfo;
 863         int ret, ops_rv;
 864         char str[64];
 865
 866         cinfo = kzalloc(sizeof(struct md_cluster_info), GFP_KERNEL);
 867         if (!cinfo)
 868                 return -ENOMEM;
 869
 870         INIT_LIST_HEAD(&cinfo->suspend_list);
 871         spin_lock_init(&cinfo->suspend_lock);
 872         init_completion(&cinfo->completion);
 873         set_bit(MD_CLUSTER_BEGIN_JOIN_CLUSTER, &cinfo->state);
 874         init_waitqueue_head(&cinfo->wait);
 875         mutex_init(&cinfo->recv_mutex);
 876
 877         mddev->cluster_info = cinfo;
 878         cinfo->mddev = mddev;
 879
 880         memset(str, 0, 64);
 881         sprintf(str, "%pU", mddev->uuid);
 882         ret = dlm_new_lockspace(str, mddev->bitmap_info.cluster_name,
 883                                 0, LVB_SIZE, &md_ls_ops, mddev,
 884                                 &ops_rv, &cinfo->lockspace);
 885         if (ret)
 886                 goto err;
 887         wait_for_completion(&cinfo->completion);
 888         if (nodes < cinfo->slot_number) {
 889                 pr_err("md-cluster: Slot allotted(%d) is greater than available slots(%d).",
 890                         cinfo->slot_number, nodes);
 891                 ret = -ERANGE;
 892                 goto err;
 893         }
 894         /* Initiate the communication resources */
 895         ret = -ENOMEM;
 896         rcu_assign_pointer(cinfo->recv_thread,
 897                         md_register_thread(recv_daemon, mddev, "cluster_recv"));
 898         if (!cinfo->recv_thread) {
 899                 pr_err("md-cluster: cannot allocate memory for recv_thread!\n");
 900                 goto err;
 901         }
 902         cinfo->message_lockres = lockres_init(mddev, "message", NULL, 1);
 903         if (!cinfo->message_lockres)
 904                 goto err;
 905         cinfo->token_lockres = lockres_init(mddev, "token", NULL, 0);
 906         if (!cinfo->token_lockres)
 907                 goto err;
 908         cinfo->no_new_dev_lockres = lockres_init(mddev, "no-new-dev", NULL, 0);
 909         if (!cinfo->no_new_dev_lockres)
 910                 goto err;
 911
 912         ret = dlm_lock_sync(cinfo->token_lockres, DLM_LOCK_EX);
 913         if (ret) {
 914                 ret = -EAGAIN;
 915                 pr_err("md-cluster: can't join cluster to avoid lock issue\n");
 916                 goto err;
 917         }
 918         cinfo->ack_lockres = lockres_init(mddev, "ack", ack_bast, 0);
 919         if (!cinfo->ack_lockres) {
 920                 ret = -ENOMEM;
 921                 goto err;
 922         }
 923         /* get sync CR lock on ACK. */
 924         if (dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_CR))
 925                 pr_err("md-cluster: failed to get a sync CR lock on ACK!(%d)\n",
 926                                 ret);
 927         dlm_unlock_sync(cinfo->token_lockres);
 928         /* get sync CR lock on no-new-dev. */
 929         if (dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR))
 930                 pr_err("md-cluster: failed to get a sync CR lock on no-new-dev!(%d)\n", ret);
 931
 932
 933         pr_info("md-cluster: Joined cluster %s slot %d\n", str, cinfo->slot_number);
 934         snprintf(str, 64, "bitmap%04d", cinfo->slot_number - 1);
 935         cinfo->bitmap_lockres = lockres_init(mddev, str, NULL, 1);
 936         if (!cinfo->bitmap_lockres) {
 937                 ret = -ENOMEM;
 938                 goto err;
 939         }
 940         if (dlm_lock_sync(cinfo->bitmap_lockres, DLM_LOCK_PW)) {
 941                 pr_err("Failed to get bitmap lock\n");
 942                 ret = -EINVAL;
 943                 goto err;
 944         }
 945
 946         cinfo->resync_lockres = lockres_init(mddev, "resync", NULL, 0);
 947         if (!cinfo->resync_lockres) {
 948                 ret = -ENOMEM;
 949                 goto err;
 950         }
 951
 952         return 0;
 953 err:
 954         set_bit(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state);
 955         md_unregister_thread(mddev, &cinfo->recovery_thread);
 956         md_unregister_thread(mddev, &cinfo->recv_thread);
 957         lockres_free(cinfo->message_lockres);
 958         lockres_free(cinfo->token_lockres);
 959         lockres_free(cinfo->ack_lockres);
 960         lockres_free(cinfo->no_new_dev_lockres);
 961         lockres_free(cinfo->resync_lockres);
 962         lockres_free(cinfo->bitmap_lockres);
 963         if (cinfo->lockspace)
 964                 dlm_release_lockspace(cinfo->lockspace, 2);
 965         mddev->cluster_info = NULL;
 966         kfree(cinfo);
 967         return ret;
 968 }
 969
 970 static void load_bitmaps(struct mddev *mddev, int total_slots)
 971 {
 972         struct md_cluster_info *cinfo = mddev->cluster_info;
 973
 974         /* load all the node's bitmap info for resync */
 975         if (gather_all_resync_info(mddev, total_slots))
 976                 pr_err("md-cluster: failed to gather all resyn infos\n");
 977         set_bit(MD_CLUSTER_ALREADY_IN_CLUSTER, &cinfo->state);
 978         /* wake up recv thread in case something need to be handled */
 979         if (test_and_clear_bit(MD_CLUSTER_PENDING_RECV_EVENT, &cinfo->state))
 980                 md_wakeup_thread(cinfo->recv_thread);
 981 }
 982
 983 static void resync_bitmap(struct mddev *mddev)
 984 {
 985         struct md_cluster_info *cinfo = mddev->cluster_info;
 986         struct cluster_msg cmsg = {0};
 987         int err;
 988
 989         cmsg.type = cpu_to_le32(BITMAP_NEEDS_SYNC);
 990         err = sendmsg(cinfo, &cmsg, 1);
 991         if (err)
 992                 pr_err("%s:%d: failed to send BITMAP_NEEDS_SYNC message (%d)\n",
 993                         __func__, __LINE__, err);
 994 }
 995
 996 static void unlock_all_bitmaps(struct mddev *mddev);
 997 static int leave(struct mddev *mddev)
 998 {
 999         struct md_cluster_info *cinfo = mddev->cluster_info;
1000
1001         if (!cinfo)
1002                 return 0;
1003
1004         /*
1005          * BITMAP_NEEDS_SYNC message should be sent when node
1006          * is leaving the cluster with dirty bitmap, also we
1007          * can only deliver it when dlm connection is available.
1008          *
1009          * Also, we should send BITMAP_NEEDS_SYNC message in
1010          * case reshaping is interrupted.
1011          */
1012         if ((cinfo->slot_number > 0 && mddev->recovery_cp != MaxSector) ||
1013             (mddev->reshape_position != MaxSector &&
1014              test_bit(MD_CLOSING, &mddev->flags)))
1015                 resync_bitmap(mddev);
1016
1017         set_bit(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state);
1018         md_unregister_thread(mddev, &cinfo->recovery_thread);
1019         md_unregister_thread(mddev, &cinfo->recv_thread);
1020         lockres_free(cinfo->message_lockres);
1021         lockres_free(cinfo->token_lockres);
1022         lockres_free(cinfo->ack_lockres);
1023         lockres_free(cinfo->no_new_dev_lockres);
1024         lockres_free(cinfo->resync_lockres);
1025         lockres_free(cinfo->bitmap_lockres);
1026         unlock_all_bitmaps(mddev);
1027         dlm_release_lockspace(cinfo->lockspace, 2);
1028         kfree(cinfo);
1029         return 0;
1030 }
1031
1032 /* slot_number(): Returns the MD slot number to use
1033  * DLM starts the slot numbers from 1, wheras cluster-md
1034  * wants the number to be from zero, so we deduct one
1035  */
1036 static int slot_number(struct mddev *mddev)
1037 {
1038         struct md_cluster_info *cinfo = mddev->cluster_info;
1039
1040         return cinfo->slot_number - 1;
1041 }
1042
1043 /*
1044  * Check if the communication is already locked, else lock the communication
1045  * channel.
1046  * If it is already locked, token is in EX mode, and hence lock_token()
1047  * should not be called.
1048  */
1049 static int metadata_update_start(struct mddev *mddev)
1050 {
1051         struct md_cluster_info *cinfo = mddev->cluster_info;
1052         int ret;
1053
1054         /*
1055          * metadata_update_start is always called with the protection of
1056          * reconfig_mutex, so set WAITING_FOR_TOKEN here.
1057          */
1058         ret = test_and_set_bit_lock(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD,
1059                                     &cinfo->state);
1060         WARN_ON_ONCE(ret);
1061         md_wakeup_thread(mddev->thread);
1062
1063         wait_event(cinfo->wait,
1064                    !test_and_set_bit(MD_CLUSTER_SEND_LOCK, &cinfo->state) ||
1065                    test_and_clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state));
1066
1067         /* If token is already locked, return 0 */
1068         if (cinfo->token_lockres->mode == DLM_LOCK_EX) {
1069                 clear_bit_unlock(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state);
1070                 return 0;
1071         }
1072
1073         ret = lock_token(cinfo);
1074         clear_bit_unlock(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state);
1075         return ret;
1076 }
1077
1078 static int metadata_update_finish(struct mddev *mddev)
1079 {
1080         struct md_cluster_info *cinfo = mddev->cluster_info;
1081         struct cluster_msg cmsg;
1082         struct md_rdev *rdev;
1083         int ret = 0;
1084         int raid_slot = -1;
1085
1086         memset(&cmsg, 0, sizeof(cmsg));
1087         cmsg.type = cpu_to_le32(METADATA_UPDATED);
1088         /* Pick up a good active device number to send.
1089          */
1090         rdev_for_each(rdev, mddev)
1091                 if (rdev->raid_disk > -1 && !test_bit(Faulty, &rdev->flags)) {
1092                         raid_slot = rdev->desc_nr;
1093                         break;
1094                 }
1095         if (raid_slot >= 0) {
1096                 cmsg.raid_slot = cpu_to_le32(raid_slot);
1097                 ret = __sendmsg(cinfo, &cmsg);
1098         } else
1099                 pr_warn("md-cluster: No good device id found to send\n");
1100         clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1101         unlock_comm(cinfo);
1102         return ret;
1103 }
1104
1105 static void metadata_update_cancel(struct mddev *mddev)
1106 {
1107         struct md_cluster_info *cinfo = mddev->cluster_info;
1108         clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1109         unlock_comm(cinfo);
1110 }
1111
1112 static int update_bitmap_size(struct mddev *mddev, sector_t size)
1113 {
1114         struct md_cluster_info *cinfo = mddev->cluster_info;
1115         struct cluster_msg cmsg = {0};
1116         int ret;
1117
1118         cmsg.type = cpu_to_le32(BITMAP_RESIZE);
1119         cmsg.high = cpu_to_le64(size);
1120         ret = sendmsg(cinfo, &cmsg, 0);
1121         if (ret)
1122                 pr_err("%s:%d: failed to send BITMAP_RESIZE message (%d)\n",
1123                         __func__, __LINE__, ret);
1124         return ret;
1125 }
1126
1127 static int resize_bitmaps(struct mddev *mddev, sector_t newsize, sector_t oldsize)
1128 {
1129         struct bitmap_counts *counts;
1130         char str[64];
1131         struct dlm_lock_resource *bm_lockres;
1132         struct bitmap *bitmap = mddev->bitmap;
1133         unsigned long my_pages = bitmap->counts.pages;
1134         int i, rv;
1135
1136         /*
1137          * We need to ensure all the nodes can grow to a larger
1138          * bitmap size before make the reshaping.
1139          */
1140         rv = update_bitmap_size(mddev, newsize);
1141         if (rv)
1142                 return rv;
1143
1144         for (i = 0; i < mddev->bitmap_info.nodes; i++) {
1145                 if (i == md_cluster_ops->slot_number(mddev))
1146                         continue;
1147
1148                 bitmap = get_bitmap_from_slot(mddev, i);
1149                 if (IS_ERR(bitmap)) {
1150                         pr_err("can't get bitmap from slot %d\n", i);
1151                         bitmap = NULL;
1152                         goto out;
1153                 }
1154                 counts = &bitmap->counts;
1155
1156                 /*
1157                  * If we can hold the bitmap lock of one node then
1158                  * the slot is not occupied, update the pages.
1159                  */
1160                 snprintf(str, 64, "bitmap%04d", i);
1161                 bm_lockres = lockres_init(mddev, str, NULL, 1);
1162                 if (!bm_lockres) {
1163                         pr_err("Cannot initialize %s lock\n", str);
1164                         goto out;
1165                 }
1166                 bm_lockres->flags |= DLM_LKF_NOQUEUE;
1167                 rv = dlm_lock_sync(bm_lockres, DLM_LOCK_PW);
1168                 if (!rv)
1169                         counts->pages = my_pages;
1170                 lockres_free(bm_lockres);
1171
1172                 if (my_pages != counts->pages)
1173                         /*
1174                          * Let's revert the bitmap size if one node
1175                          * can't resize bitmap
1176                          */
1177                         goto out;
1178                 md_bitmap_free(bitmap);
1179         }
1180
1181         return 0;
1182 out:
1183         md_bitmap_free(bitmap);
1184         update_bitmap_size(mddev, oldsize);
1185         return -1;
1186 }
1187
1188 /*
1189  * return 0 if all the bitmaps have the same sync_size
1190  */
1191 static int cluster_check_sync_size(struct mddev *mddev)
1192 {
1193         int i, rv;
1194         bitmap_super_t *sb;
1195         unsigned long my_sync_size, sync_size = 0;
1196         int node_num = mddev->bitmap_info.nodes;
1197         int current_slot = md_cluster_ops->slot_number(mddev);
1198         struct bitmap *bitmap = mddev->bitmap;
1199         char str[64];
1200         struct dlm_lock_resource *bm_lockres;
1201
1202         sb = kmap_atomic(bitmap->storage.sb_page);
1203         my_sync_size = sb->sync_size;
1204         kunmap_atomic(sb);
1205
1206         for (i = 0; i < node_num; i++) {
1207                 if (i == current_slot)
1208                         continue;
1209
1210                 bitmap = get_bitmap_from_slot(mddev, i);
1211                 if (IS_ERR(bitmap)) {
1212                         pr_err("can't get bitmap from slot %d\n", i);
1213                         return -1;
1214                 }
1215
1216                 /*
1217                  * If we can hold the bitmap lock of one node then
1218                  * the slot is not occupied, update the sb.
1219                  */
1220                 snprintf(str, 64, "bitmap%04d", i);
1221                 bm_lockres = lockres_init(mddev, str, NULL, 1);
1222                 if (!bm_lockres) {
1223                         pr_err("md-cluster: Cannot initialize %s\n", str);
1224                         md_bitmap_free(bitmap);
1225                         return -1;
1226                 }
1227                 bm_lockres->flags |= DLM_LKF_NOQUEUE;
1228                 rv = dlm_lock_sync(bm_lockres, DLM_LOCK_PW);
1229                 if (!rv)
1230                         md_bitmap_update_sb(bitmap);
1231                 lockres_free(bm_lockres);
1232
1233                 sb = kmap_atomic(bitmap->storage.sb_page);
1234                 if (sync_size == 0)
1235                         sync_size = sb->sync_size;
1236                 else if (sync_size != sb->sync_size) {
1237                         kunmap_atomic(sb);
1238                         md_bitmap_free(bitmap);
1239                         return -1;
1240                 }
1241                 kunmap_atomic(sb);
1242                 md_bitmap_free(bitmap);
1243         }
1244
1245         return (my_sync_size == sync_size) ? 0 : -1;
1246 }
1247
1248 /*
1249  * Update the size for cluster raid is a little more complex, we perform it
1250  * by the steps:
1251  * 1. hold token lock and update superblock in initiator node.
1252  * 2. send METADATA_UPDATED msg to other nodes.
1253  * 3. The initiator node continues to check each bitmap's sync_size, if all
1254  *    bitmaps have the same value of sync_size, then we can set capacity and
1255  *    let other nodes to perform it. If one node can't update sync_size
1256  *    accordingly, we need to revert to previous value.
1257  */
1258 static void update_size(struct mddev *mddev, sector_t old_dev_sectors)
1259 {
1260         struct md_cluster_info *cinfo = mddev->cluster_info;
1261         struct cluster_msg cmsg;
1262         struct md_rdev *rdev;
1263         int ret = 0;
1264         int raid_slot = -1;
1265
1266         md_update_sb(mddev, 1);
1267         if (lock_comm(cinfo, 1)) {
1268                 pr_err("%s: lock_comm failed\n", __func__);
1269                 return;
1270         }
1271
1272         memset(&cmsg, 0, sizeof(cmsg));
1273         cmsg.type = cpu_to_le32(METADATA_UPDATED);
1274         rdev_for_each(rdev, mddev)
1275                 if (rdev->raid_disk >= 0 && !test_bit(Faulty, &rdev->flags)) {
1276                         raid_slot = rdev->desc_nr;
1277                         break;
1278                 }
1279         if (raid_slot >= 0) {
1280                 cmsg.raid_slot = cpu_to_le32(raid_slot);
1281                 /*
1282                  * We can only change capiticy after all the nodes can do it,
1283                  * so need to wait after other nodes already received the msg
1284                  * and handled the change
1285                  */
1286                 ret = __sendmsg(cinfo, &cmsg);
1287                 if (ret) {
1288                         pr_err("%s:%d: failed to send METADATA_UPDATED msg\n",
1289                                __func__, __LINE__);
1290                         unlock_comm(cinfo);
1291                         return;
1292                 }
1293         } else {
1294                 pr_err("md-cluster: No good device id found to send\n");
1295                 unlock_comm(cinfo);
1296                 return;
1297         }
1298
1299         /*
1300          * check the sync_size from other node's bitmap, if sync_size
1301          * have already updated in other nodes as expected, send an
1302          * empty metadata msg to permit the change of capacity
1303          */
1304         if (cluster_check_sync_size(mddev) == 0) {
1305                 memset(&cmsg, 0, sizeof(cmsg));
1306                 cmsg.type = cpu_to_le32(CHANGE_CAPACITY);
1307                 ret = __sendmsg(cinfo, &cmsg);
1308                 if (ret)
1309                         pr_err("%s:%d: failed to send CHANGE_CAPACITY msg\n",
1310                                __func__, __LINE__);
1311                 set_capacity_and_notify(mddev->gendisk, mddev->array_sectors);
1312         } else {
1313                 /* revert to previous sectors */
1314                 ret = mddev->pers->resize(mddev, old_dev_sectors);
1315                 ret = __sendmsg(cinfo, &cmsg);
1316                 if (ret)
1317                         pr_err("%s:%d: failed to send METADATA_UPDATED msg\n",
1318                                __func__, __LINE__);
1319         }
1320         unlock_comm(cinfo);
1321 }
1322
1323 static int resync_start(struct mddev *mddev)
1324 {
1325         struct md_cluster_info *cinfo = mddev->cluster_info;
1326         return dlm_lock_sync_interruptible(cinfo->resync_lockres, DLM_LOCK_EX, mddev);
1327 }
1328
1329 static void resync_info_get(struct mddev *mddev, sector_t *lo, sector_t *hi)
1330 {
1331         struct md_cluster_info *cinfo = mddev->cluster_info;
1332
1333         spin_lock_irq(&cinfo->suspend_lock);
1334         *lo = cinfo->suspend_lo;
1335         *hi = cinfo->suspend_hi;
1336         spin_unlock_irq(&cinfo->suspend_lock);
1337 }
1338
1339 static int resync_info_update(struct mddev *mddev, sector_t lo, sector_t hi)
1340 {
1341         struct md_cluster_info *cinfo = mddev->cluster_info;
1342         struct resync_info ri;
1343         struct cluster_msg cmsg = {0};
1344
1345         /* do not send zero again, if we have sent before */
1346         if (hi == 0) {
1347                 memcpy(&ri, cinfo->bitmap_lockres->lksb.sb_lvbptr, sizeof(struct resync_info));
1348                 if (le64_to_cpu(ri.hi) == 0)
1349                         return 0;
1350         }
1351
1352         add_resync_info(cinfo->bitmap_lockres, lo, hi);
1353         /* Re-acquire the lock to refresh LVB */
1354         dlm_lock_sync(cinfo->bitmap_lockres, DLM_LOCK_PW);
1355         cmsg.type = cpu_to_le32(RESYNCING);
1356         cmsg.low = cpu_to_le64(lo);
1357         cmsg.high = cpu_to_le64(hi);
1358
1359         /*
1360          * mddev_lock is held if resync_info_update is called from
1361          * resync_finish (md_reap_sync_thread -> resync_finish)
1362          */
1363         if (lo == 0 && hi == 0)
1364                 return sendmsg(cinfo, &cmsg, 1);
1365         else
1366                 return sendmsg(cinfo, &cmsg, 0);
1367 }
1368
1369 static int resync_finish(struct mddev *mddev)
1370 {
1371         struct md_cluster_info *cinfo = mddev->cluster_info;
1372         int ret = 0;
1373
1374         clear_bit(MD_RESYNCING_REMOTE, &mddev->recovery);
1375
1376         /*
1377          * If resync thread is interrupted so we can't say resync is finished,
1378          * another node will launch resync thread to continue.
1379          */
1380         if (!test_bit(MD_CLOSING, &mddev->flags))
1381                 ret = resync_info_update(mddev, 0, 0);
1382         dlm_unlock_sync(cinfo->resync_lockres);
1383         return ret;
1384 }
1385
1386 static int area_resyncing(struct mddev *mddev, int direction,
1387                 sector_t lo, sector_t hi)
1388 {
1389         struct md_cluster_info *cinfo = mddev->cluster_info;
1390         int ret = 0;
1391
1392         if ((direction == READ) &&
1393                 test_bit(MD_CLUSTER_SUSPEND_READ_BALANCING, &cinfo->state))
1394                 return 1;
1395
1396         spin_lock_irq(&cinfo->suspend_lock);
1397         if (hi > cinfo->suspend_lo && lo < cinfo->suspend_hi)
1398                 ret = 1;
1399         spin_unlock_irq(&cinfo->suspend_lock);
1400         return ret;
1401 }
1402
1403 /* add_new_disk() - initiates a disk add
1404  * However, if this fails before writing md_update_sb(),
1405  * add_new_disk_cancel() must be called to release token lock
1406  */
1407 static int add_new_disk(struct mddev *mddev, struct md_rdev *rdev)
1408 {
1409         struct md_cluster_info *cinfo = mddev->cluster_info;
1410         struct cluster_msg cmsg;
1411         int ret = 0;
1412         struct mdp_superblock_1 *sb = page_address(rdev->sb_page);
1413         char *uuid = sb->device_uuid;
1414
1415         memset(&cmsg, 0, sizeof(cmsg));
1416         cmsg.type = cpu_to_le32(NEWDISK);
1417         memcpy(cmsg.uuid, uuid, 16);
1418         cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
1419         if (lock_comm(cinfo, 1))
1420                 return -EAGAIN;
1421         ret = __sendmsg(cinfo, &cmsg);
1422         if (ret) {
1423                 unlock_comm(cinfo);
1424                 return ret;
1425         }
1426         cinfo->no_new_dev_lockres->flags |= DLM_LKF_NOQUEUE;
1427         ret = dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_EX);
1428         cinfo->no_new_dev_lockres->flags &= ~DLM_LKF_NOQUEUE;
1429         /* Some node does not "see" the device */
1430         if (ret == -EAGAIN)
1431                 ret = -ENOENT;
1432         if (ret)
1433                 unlock_comm(cinfo);
1434         else {
1435                 dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR);
1436                 /* Since MD_CHANGE_DEVS will be set in add_bound_rdev which
1437                  * will run soon after add_new_disk, the below path will be
1438                  * invoked:
1439                  *   md_wakeup_thread(mddev->thread)
1440                  *      -> conf->thread (raid1d)
1441                  *      -> md_check_recovery -> md_update_sb
1442                  *      -> metadata_update_start/finish
1443                  * MD_CLUSTER_SEND_LOCKED_ALREADY will be cleared eventually.
1444                  *
1445                  * For other failure cases, metadata_update_cancel and
1446                  * add_new_disk_cancel also clear below bit as well.
1447                  * */
1448                 set_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1449                 wake_up(&cinfo->wait);
1450         }
1451         return ret;
1452 }
1453
1454 static void add_new_disk_cancel(struct mddev *mddev)
1455 {
1456         struct md_cluster_info *cinfo = mddev->cluster_info;
1457         clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1458         unlock_comm(cinfo);
1459 }
1460
1461 static int new_disk_ack(struct mddev *mddev, bool ack)
1462 {
1463         struct md_cluster_info *cinfo = mddev->cluster_info;
1464
1465         if (!test_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state)) {
1466                 pr_warn("md-cluster(%s): Spurious cluster confirmation\n", mdname(mddev));
1467                 return -EINVAL;
1468         }
1469
1470         if (ack)
1471                 dlm_unlock_sync(cinfo->no_new_dev_lockres);
1472         complete(&cinfo->newdisk_completion);
1473         return 0;
1474 }
1475
1476 static int remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1477 {
1478         struct cluster_msg cmsg = {0};
1479         struct md_cluster_info *cinfo = mddev->cluster_info;
1480         cmsg.type = cpu_to_le32(REMOVE);
1481         cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
1482         return sendmsg(cinfo, &cmsg, 1);
1483 }
1484
1485 static int lock_all_bitmaps(struct mddev *mddev)
1486 {
1487         int slot, my_slot, ret, held = 1, i = 0;
1488         char str[64];
1489         struct md_cluster_info *cinfo = mddev->cluster_info;
1490
1491         cinfo->other_bitmap_lockres =
1492                 kcalloc(mddev->bitmap_info.nodes - 1,
1493                         sizeof(struct dlm_lock_resource *), GFP_KERNEL);
1494         if (!cinfo->other_bitmap_lockres) {
1495                 pr_err("md: can't alloc mem for other bitmap locks\n");
1496                 return 0;
1497         }
1498
1499         my_slot = slot_number(mddev);
1500         for (slot = 0; slot < mddev->bitmap_info.nodes; slot++) {
1501                 if (slot == my_slot)
1502                         continue;
1503
1504                 memset(str, '\0', 64);
1505                 snprintf(str, 64, "bitmap%04d", slot);
1506                 cinfo->other_bitmap_lockres[i] = lockres_init(mddev, str, NULL, 1);
1507                 if (!cinfo->other_bitmap_lockres[i])
1508                         return -ENOMEM;
1509
1510                 cinfo->other_bitmap_lockres[i]->flags |= DLM_LKF_NOQUEUE;
1511                 ret = dlm_lock_sync(cinfo->other_bitmap_lockres[i], DLM_LOCK_PW);
1512                 if (ret)
1513                         held = -1;
1514                 i++;
1515         }
1516
1517         return held;
1518 }
1519
1520 static void unlock_all_bitmaps(struct mddev *mddev)
1521 {
1522         struct md_cluster_info *cinfo = mddev->cluster_info;
1523         int i;
1524
1525         /* release other node's bitmap lock if they are existed */
1526         if (cinfo->other_bitmap_lockres) {
1527                 for (i = 0; i < mddev->bitmap_info.nodes - 1; i++) {
1528                         if (cinfo->other_bitmap_lockres[i]) {
1529                                 lockres_free(cinfo->other_bitmap_lockres[i]);
1530                         }
1531                 }
1532                 kfree(cinfo->other_bitmap_lockres);
1533                 cinfo->other_bitmap_lockres = NULL;
1534         }
1535 }
1536
1537 static int gather_bitmaps(struct md_rdev *rdev)
1538 {
1539         int sn, err;
1540         sector_t lo, hi;
1541         struct cluster_msg cmsg = {0};
1542         struct mddev *mddev = rdev->mddev;
1543         struct md_cluster_info *cinfo = mddev->cluster_info;
1544
1545         cmsg.type = cpu_to_le32(RE_ADD);
1546         cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
1547         err = sendmsg(cinfo, &cmsg, 1);
1548         if (err)
1549                 goto out;
1550
1551         for (sn = 0; sn < mddev->bitmap_info.nodes; sn++) {
1552                 if (sn == (cinfo->slot_number - 1))
1553                         continue;
1554                 err = md_bitmap_copy_from_slot(mddev, sn, &lo, &hi, false);
1555                 if (err) {
1556                         pr_warn("md-cluster: Could not gather bitmaps from slot %d", sn);
1557                         goto out;
1558                 }
1559                 if ((hi > 0) && (lo < mddev->recovery_cp))
1560                         mddev->recovery_cp = lo;
1561         }
1562 out:
1563         return err;
1564 }
1565
1566 static struct md_cluster_operations cluster_ops = {
1567         .join   = join,
1568         .leave  = leave,
1569         .slot_number = slot_number,
1570         .resync_start = resync_start,
1571         .resync_finish = resync_finish,
1572         .resync_info_update = resync_info_update,
1573         .resync_info_get = resync_info_get,
1574         .metadata_update_start = metadata_update_start,
1575         .metadata_update_finish = metadata_update_finish,
1576         .metadata_update_cancel = metadata_update_cancel,
1577         .area_resyncing = area_resyncing,
1578         .add_new_disk = add_new_disk,
1579         .add_new_disk_cancel = add_new_disk_cancel,
1580         .new_disk_ack = new_disk_ack,
1581         .remove_disk = remove_disk,
1582         .load_bitmaps = load_bitmaps,
1583         .gather_bitmaps = gather_bitmaps,
1584         .resize_bitmaps = resize_bitmaps,
1585         .lock_all_bitmaps = lock_all_bitmaps,
1586         .unlock_all_bitmaps = unlock_all_bitmaps,
1587         .update_size = update_size,
1588 };
1589
1590 static int __init cluster_init(void)
1591 {
1592         pr_warn("md-cluster: support raid1 and raid10 (limited support)\n");
1593         pr_info("Registering Cluster MD functions\n");
1594         register_md_cluster_operations(&cluster_ops, THIS_MODULE);
1595         return 0;
1596 }
1597
1598 static void cluster_exit(void)
1599 {
1600         unregister_md_cluster_operations();
1601 }
1602
1603 module_init(cluster_init);
1604 module_exit(cluster_exit);
1605 MODULE_AUTHOR("SUSE");
1606 MODULE_LICENSE("GPL");
1607 MODULE_DESCRIPTION("Clustering support for MD");