drivers/md/md-cluster.c

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