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 #define WAIT_DLM_LOCK_TIMEOUT (30 * HZ)
  19
  20 struct dlm_lock_resource {
  21         dlm_lockspace_t *ls;
  22         struct dlm_lksb lksb;
  23         char *name; /* lock name. */
  24         uint32_t flags; /* flags to pass to dlm_lock() */
  25         wait_queue_head_t sync_locking; /* wait queue for synchronized locking */
  26         bool sync_locking_done;
  27         void (*bast)(void *arg, int mode); /* blocking AST function pointer*/
  28         struct mddev *mddev; /* pointing back to mddev. */
  29         int mode;
  30 };
  31
  32 struct resync_info {
  33         __le64 lo;
  34         __le64 hi;
  35 };
  36
  37 /* md_cluster_info flags */
  38 #define         MD_CLUSTER_WAITING_FOR_NEWDISK          1
  39 #define         MD_CLUSTER_SUSPEND_READ_BALANCING       2
  40 #define         MD_CLUSTER_BEGIN_JOIN_CLUSTER           3
  41
  42 /* Lock the send communication. This is done through
  43  * bit manipulation as opposed to a mutex in order to
  44  * accommodate lock and hold. See next comment.
  45  */
  46 #define         MD_CLUSTER_SEND_LOCK                    4
  47 /* If cluster operations (such as adding a disk) must lock the
  48  * communication channel, so as to perform extra operations
  49  * (update metadata) and no other operation is allowed on the
  50  * MD. Token needs to be locked and held until the operation
  51  * completes witha md_update_sb(), which would eventually release
  52  * the lock.
  53  */
  54 #define         MD_CLUSTER_SEND_LOCKED_ALREADY          5
  55 /* We should receive message after node joined cluster and
  56  * set up all the related infos such as bitmap and personality */
  57 #define         MD_CLUSTER_ALREADY_IN_CLUSTER           6
  58 #define         MD_CLUSTER_PENDING_RECV_EVENT           7
  59 #define         MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD              8
  60 #define         MD_CLUSTER_WAITING_FOR_SYNC             9
  61
  62 struct md_cluster_info {
  63         struct mddev *mddev; /* the md device which md_cluster_info belongs to */
  64         /* dlm lock space and resources for clustered raid. */
  65         dlm_lockspace_t *lockspace;
  66         int slot_number;
  67         struct completion completion;
  68         struct mutex recv_mutex;
  69         struct dlm_lock_resource *bitmap_lockres;
  70         struct dlm_lock_resource **other_bitmap_lockres;
  71         struct dlm_lock_resource *resync_lockres;
  72         struct list_head suspend_list;
  73
  74         spinlock_t suspend_lock;
  75         /* record the region which write should be suspended */
  76         sector_t suspend_lo;
  77         sector_t suspend_hi;
  78         int suspend_from; /* the slot which broadcast suspend_lo/hi */
  79
  80         struct md_thread __rcu *recovery_thread;
  81         unsigned long recovery_map;
  82         /* communication loc resources */
  83         struct dlm_lock_resource *ack_lockres;
  84         struct dlm_lock_resource *message_lockres;
  85         struct dlm_lock_resource *token_lockres;
  86         struct dlm_lock_resource *no_new_dev_lockres;
  87         struct md_thread __rcu *recv_thread;
  88         struct completion newdisk_completion;
  89         wait_queue_head_t wait;
  90         unsigned long state;
  91         /* record the region in RESYNCING message */
  92         sector_t sync_low;
  93         sector_t sync_hi;
  94 };
  95
  96 /* For compatibility, add the new msg_type at the end. */
  97 enum msg_type {
  98         METADATA_UPDATED = 0,
  99         RESYNCING,
 100         NEWDISK,
 101         REMOVE,
 102         RE_ADD,
 103         BITMAP_NEEDS_SYNC,
 104         CHANGE_CAPACITY,
 105         BITMAP_RESIZE,
 106         RESYNCING_START,
 107 };
 108
 109 struct cluster_msg {
 110         __le32 type;
 111         __le32 slot;
 112         /* TODO: Unionize this for smaller footprint */
 113         __le64 low;
 114         __le64 high;
 115         char uuid[16];
 116         __le32 raid_slot;
 117 };
 118
 119 static void sync_ast(void *arg)
 120 {
 121         struct dlm_lock_resource *res;
 122
 123         res = arg;
 124         res->sync_locking_done = true;
 125         wake_up(&res->sync_locking);
 126 }
 127
 128 static int dlm_lock_sync(struct dlm_lock_resource *res, int mode)
 129 {
 130         int ret = 0;
 131
 132         ret = dlm_lock(res->ls, mode, &res->lksb,
 133                         res->flags, res->name, strlen(res->name),
 134                         0, sync_ast, res, res->bast);
 135         if (ret)
 136                 return ret;
 137         ret = wait_event_timeout(res->sync_locking, res->sync_locking_done,
 138                                 WAIT_DLM_LOCK_TIMEOUT);
 139         res->sync_locking_done = false;
 140         if (!ret) {
 141                 pr_err("locking DLM '%s' timeout!\n", res->name);
 142                 return -EBUSY;
 143         }
 144         if (res->lksb.sb_status == 0)
 145                 res->mode = mode;
 146         return res->lksb.sb_status;
 147 }
 148
 149 static int dlm_unlock_sync(struct dlm_lock_resource *res)
 150 {
 151         return dlm_lock_sync(res, DLM_LOCK_NL);
 152 }
 153
 154 /*
 155  * An variation of dlm_lock_sync, which make lock request could
 156  * be interrupted
 157  */
 158 static int dlm_lock_sync_interruptible(struct dlm_lock_resource *res, int mode,
 159                                        struct mddev *mddev)
 160 {
 161         int ret = 0;
 162
 163         ret = dlm_lock(res->ls, mode, &res->lksb,
 164                         res->flags, res->name, strlen(res->name),
 165                         0, sync_ast, res, res->bast);
 166         if (ret)
 167                 return ret;
 168
 169         wait_event(res->sync_locking, res->sync_locking_done
 170                                       || kthread_should_stop()
 171                                       || test_bit(MD_CLOSING, &mddev->flags));
 172         if (!res->sync_locking_done) {
 173                 /*
 174                  * the convert queue contains the lock request when request is
 175                  * interrupted, and sync_ast could still be run, so need to
 176                  * cancel the request and reset completion
 177                  */
 178                 ret = dlm_unlock(res->ls, res->lksb.sb_lkid, DLM_LKF_CANCEL,
 179                         &res->lksb, res);
 180                 res->sync_locking_done = false;
 181                 if (unlikely(ret != 0))
 182                         pr_info("failed to cancel previous lock request "
 183                                  "%s return %d\n", res->name, ret);
 184                 return -EPERM;
 185         } else
 186                 res->sync_locking_done = false;
 187         if (res->lksb.sb_status == 0)
 188                 res->mode = mode;
 189         return res->lksb.sb_status;
 190 }
 191
 192 static struct dlm_lock_resource *lockres_init(struct mddev *mddev,
 193                 char *name, void (*bastfn)(void *arg, int mode), int with_lvb)
 194 {
 195         struct dlm_lock_resource *res = NULL;
 196         int ret, namelen;
 197         struct md_cluster_info *cinfo = mddev->cluster_info;
 198
 199         res = kzalloc(sizeof(struct dlm_lock_resource), GFP_KERNEL);
 200         if (!res)
 201                 return NULL;
 202         init_waitqueue_head(&res->sync_locking);
 203         res->sync_locking_done = false;
 204         res->ls = cinfo->lockspace;
 205         res->mddev = mddev;
 206         res->mode = DLM_LOCK_IV;
 207         namelen = strlen(name);
 208         res->name = kzalloc(namelen + 1, GFP_KERNEL);
 209         if (!res->name) {
 210                 pr_err("md-cluster: Unable to allocate resource name for resource %s\n", name);
 211                 goto out_err;
 212         }
 213         strscpy(res->name, name, namelen + 1);
 214         if (with_lvb) {
 215                 res->lksb.sb_lvbptr = kzalloc(LVB_SIZE, GFP_KERNEL);
 216                 if (!res->lksb.sb_lvbptr) {
 217                         pr_err("md-cluster: Unable to allocate LVB for resource %s\n", name);
 218                         goto out_err;
 219                 }
 220                 res->flags = DLM_LKF_VALBLK;
 221         }
 222
 223         if (bastfn)
 224                 res->bast = bastfn;
 225
 226         res->flags |= DLM_LKF_EXPEDITE;
 227
 228         ret = dlm_lock_sync(res, DLM_LOCK_NL);
 229         if (ret) {
 230                 pr_err("md-cluster: Unable to lock NL on new lock resource %s\n", name);
 231                 goto out_err;
 232         }
 233         res->flags &= ~DLM_LKF_EXPEDITE;
 234         res->flags |= DLM_LKF_CONVERT;
 235
 236         return res;
 237 out_err:
 238         kfree(res->lksb.sb_lvbptr);
 239         kfree(res->name);
 240         kfree(res);
 241         return NULL;
 242 }
 243
 244 static void lockres_free(struct dlm_lock_resource *res)
 245 {
 246         int ret = 0;
 247
 248         if (!res)
 249                 return;
 250
 251         /*
 252          * use FORCEUNLOCK flag, so we can unlock even the lock is on the
 253          * waiting or convert queue
 254          */
 255         ret = dlm_unlock(res->ls, res->lksb.sb_lkid, DLM_LKF_FORCEUNLOCK,
 256                 &res->lksb, res);
 257         if (unlikely(ret != 0))
 258                 pr_err("failed to unlock %s return %d\n", res->name, ret);
 259         else
 260                 wait_event(res->sync_locking, res->sync_locking_done);
 261
 262         kfree(res->name);
 263         kfree(res->lksb.sb_lvbptr);
 264         kfree(res);
 265 }
 266
 267 static void add_resync_info(struct dlm_lock_resource *lockres,
 268                             sector_t lo, sector_t hi)
 269 {
 270         struct resync_info *ri;
 271
 272         ri = (struct resync_info *)lockres->lksb.sb_lvbptr;
 273         ri->lo = cpu_to_le64(lo);
 274         ri->hi = cpu_to_le64(hi);
 275 }
 276
 277 static int read_resync_info(struct mddev *mddev,
 278                             struct dlm_lock_resource *lockres)
 279 {
 280         struct resync_info ri;
 281         struct md_cluster_info *cinfo = mddev->cluster_info;
 282         int ret = 0;
 283
 284         dlm_lock_sync(lockres, DLM_LOCK_CR);
 285         memcpy(&ri, lockres->lksb.sb_lvbptr, sizeof(struct resync_info));
 286         if (le64_to_cpu(ri.hi) > 0) {
 287                 cinfo->suspend_hi = le64_to_cpu(ri.hi);
 288                 cinfo->suspend_lo = le64_to_cpu(ri.lo);
 289                 ret = 1;
 290         }
 291         dlm_unlock_sync(lockres);
 292         return ret;
 293 }
 294
 295 static void recover_bitmaps(struct md_thread *thread)
 296 {
 297         struct mddev *mddev = thread->mddev;
 298         struct md_cluster_info *cinfo = mddev->cluster_info;
 299         struct dlm_lock_resource *bm_lockres;
 300         char str[64];
 301         int slot, ret;
 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                 cinfo->suspend_hi = 0;
 329                 cinfo->suspend_lo = 0;
 330                 cinfo->suspend_from = -1;
 331                 spin_unlock_irq(&cinfo->suspend_lock);
 332
 333                 /* Kick off a reshape if needed */
 334                 if (test_bit(MD_RESYNCING_REMOTE, &mddev->recovery) &&
 335                     test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
 336                     mddev->reshape_position != MaxSector)
 337                         md_wakeup_thread(mddev->sync_thread);
 338
 339                 if (hi > 0) {
 340                         if (lo < mddev->recovery_cp)
 341                                 mddev->recovery_cp = lo;
 342                         /* wake up thread to continue resync in case resync
 343                          * is not finished */
 344                         if (mddev->recovery_cp != MaxSector) {
 345                                 /*
 346                                  * clear the REMOTE flag since we will launch
 347                                  * resync thread in current node.
 348                                  */
 349                                 clear_bit(MD_RESYNCING_REMOTE,
 350                                           &mddev->recovery);
 351                                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
 352                                 md_wakeup_thread(mddev->thread);
 353                         }
 354                 }
 355 clear_bit:
 356                 lockres_free(bm_lockres);
 357                 clear_bit(slot, &cinfo->recovery_map);
 358         }
 359 }
 360
 361 static void recover_prep(void *arg)
 362 {
 363         struct mddev *mddev = arg;
 364         struct md_cluster_info *cinfo = mddev->cluster_info;
 365         set_bit(MD_CLUSTER_SUSPEND_READ_BALANCING, &cinfo->state);
 366 }
 367
 368 static void __recover_slot(struct mddev *mddev, int slot)
 369 {
 370         struct md_cluster_info *cinfo = mddev->cluster_info;
 371
 372         set_bit(slot, &cinfo->recovery_map);
 373         if (!cinfo->recovery_thread) {
 374                 rcu_assign_pointer(cinfo->recovery_thread,
 375                         md_register_thread(recover_bitmaps, mddev, "recover"));
 376                 if (!cinfo->recovery_thread) {
 377                         pr_warn("md-cluster: Could not create recovery thread\n");
 378                         return;
 379                 }
 380         }
 381         md_wakeup_thread(cinfo->recovery_thread);
 382 }
 383
 384 static void recover_slot(void *arg, struct dlm_slot *slot)
 385 {
 386         struct mddev *mddev = arg;
 387         struct md_cluster_info *cinfo = mddev->cluster_info;
 388
 389         pr_info("md-cluster: %s Node %d/%d down. My slot: %d. Initiating recovery.\n",
 390                         mddev->bitmap_info.cluster_name,
 391                         slot->nodeid, slot->slot,
 392                         cinfo->slot_number);
 393         /* deduct one since dlm slot starts from one while the num of
 394          * cluster-md begins with 0 */
 395         __recover_slot(mddev, slot->slot - 1);
 396 }
 397
 398 static void recover_done(void *arg, struct dlm_slot *slots,
 399                 int num_slots, int our_slot,
 400                 uint32_t generation)
 401 {
 402         struct mddev *mddev = arg;
 403         struct md_cluster_info *cinfo = mddev->cluster_info;
 404
 405         cinfo->slot_number = our_slot;
 406         /* completion is only need to be complete when node join cluster,
 407          * it doesn't need to run during another node's failure */
 408         if (test_bit(MD_CLUSTER_BEGIN_JOIN_CLUSTER, &cinfo->state)) {
 409                 complete(&cinfo->completion);
 410                 clear_bit(MD_CLUSTER_BEGIN_JOIN_CLUSTER, &cinfo->state);
 411         }
 412         clear_bit(MD_CLUSTER_SUSPEND_READ_BALANCING, &cinfo->state);
 413 }
 414
 415 /* the ops is called when node join the cluster, and do lock recovery
 416  * if node failure occurs */
 417 static const struct dlm_lockspace_ops md_ls_ops = {
 418         .recover_prep = recover_prep,
 419         .recover_slot = recover_slot,
 420         .recover_done = recover_done,
 421 };
 422
 423 /*
 424  * The BAST function for the ack lock resource
 425  * This function wakes up the receive thread in
 426  * order to receive and process the message.
 427  */
 428 static void ack_bast(void *arg, int mode)
 429 {
 430         struct dlm_lock_resource *res = arg;
 431         struct md_cluster_info *cinfo = res->mddev->cluster_info;
 432
 433         if (mode == DLM_LOCK_EX) {
 434                 if (test_bit(MD_CLUSTER_ALREADY_IN_CLUSTER, &cinfo->state))
 435                         md_wakeup_thread(cinfo->recv_thread);
 436                 else
 437                         set_bit(MD_CLUSTER_PENDING_RECV_EVENT, &cinfo->state);
 438         }
 439 }
 440
 441 static void remove_suspend_info(struct mddev *mddev, int slot)
 442 {
 443         struct md_cluster_info *cinfo = mddev->cluster_info;
 444         mddev->pers->quiesce(mddev, 1);
 445         spin_lock_irq(&cinfo->suspend_lock);
 446         cinfo->suspend_hi = 0;
 447         cinfo->suspend_lo = 0;
 448         spin_unlock_irq(&cinfo->suspend_lock);
 449         mddev->pers->quiesce(mddev, 0);
 450 }
 451
 452 static void process_suspend_info(struct mddev *mddev,
 453                 int slot, sector_t lo, sector_t hi)
 454 {
 455         struct md_cluster_info *cinfo = mddev->cluster_info;
 456         struct mdp_superblock_1 *sb = NULL;
 457         struct md_rdev *rdev;
 458
 459         if (!hi) {
 460                 /*
 461                  * clear the REMOTE flag since resync or recovery is finished
 462                  * in remote node.
 463                  */
 464                 clear_bit(MD_RESYNCING_REMOTE, &mddev->recovery);
 465                 remove_suspend_info(mddev, slot);
 466                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
 467                 clear_bit(MD_CLUSTER_WAITING_FOR_SYNC, &cinfo->state);
 468                 md_wakeup_thread(mddev->thread);
 469                 return;
 470         }
 471
 472         rdev_for_each(rdev, mddev)
 473                 if (rdev->raid_disk > -1 && !test_bit(Faulty, &rdev->flags)) {
 474                         sb = page_address(rdev->sb_page);
 475                         break;
 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 md_bitmap_sync_with_cluster to clear NEEDED_MASK
 491          * and set RESYNC_MASK since  resync thread is running
 492          * in another node, so we don't need to do the resync
 493          * again with the same section.
 494          *
 495          * Skip md_bitmap_sync_with_cluster in case reshape
 496          * happening, because reshaping region is small and
 497          * we don't want to trigger lots of WARN.
 498          */
 499         if (sb && !(le32_to_cpu(sb->feature_map) & MD_FEATURE_RESHAPE_ACTIVE))
 500                 md_bitmap_sync_with_cluster(mddev, cinfo->sync_low,
 501                                             cinfo->sync_hi, lo, hi);
 502         cinfo->sync_low = lo;
 503         cinfo->sync_hi = hi;
 504
 505         mddev->pers->quiesce(mddev, 1);
 506         spin_lock_irq(&cinfo->suspend_lock);
 507         cinfo->suspend_from = slot;
 508         cinfo->suspend_lo = lo;
 509         cinfo->suspend_hi = hi;
 510         spin_unlock_irq(&cinfo->suspend_lock);
 511         mddev->pers->quiesce(mddev, 0);
 512 }
 513
 514 static int 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         int res = 0;
 523
 524         len = snprintf(disk_uuid, 64, "DEVICE_UUID=");
 525         sprintf(disk_uuid + len, "%pU", cmsg->uuid);
 526         snprintf(raid_slot, 16, "RAID_DISK=%d", le32_to_cpu(cmsg->raid_slot));
 527         pr_info("%s:%d Sending kobject change with %s and %s\n", __func__, __LINE__, disk_uuid, raid_slot);
 528         init_completion(&cinfo->newdisk_completion);
 529         set_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state);
 530         kobject_uevent_env(&disk_to_dev(mddev->gendisk)->kobj, KOBJ_CHANGE, envp);
 531         if (!wait_for_completion_timeout(&cinfo->newdisk_completion,
 532                                         NEW_DEV_TIMEOUT)) {
 533                 pr_err("md-cluster(%s:%d): timeout on a new disk adding\n",
 534                         __func__, __LINE__);
 535                 res = -1;
 536         }
 537         clear_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state);
 538         set_bit(MD_CLUSTER_WAITING_FOR_SYNC, &cinfo->state);
 539         return res;
 540 }
 541
 542
 543 static void process_metadata_update(struct mddev *mddev, struct cluster_msg *msg)
 544 {
 545         int got_lock = 0;
 546         struct md_thread *thread;
 547         struct md_cluster_info *cinfo = mddev->cluster_info;
 548         mddev->good_device_nr = le32_to_cpu(msg->raid_slot);
 549
 550         dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR);
 551
 552         /* daemaon thread must exist */
 553         thread = rcu_dereference_protected(mddev->thread, true);
 554         wait_event(thread->wqueue,
 555                    (got_lock = mddev_trylock(mddev)) ||
 556                     test_bit(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state));
 557         md_reload_sb(mddev, mddev->good_device_nr);
 558         if (got_lock)
 559                 mddev_unlock(mddev);
 560 }
 561
 562 static void process_remove_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) {
 569                 set_bit(ClusterRemove, &rdev->flags);
 570                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
 571                 md_wakeup_thread(mddev->thread);
 572         }
 573         else
 574                 pr_warn("%s: %d Could not find disk(%d) to REMOVE\n",
 575                         __func__, __LINE__, le32_to_cpu(msg->raid_slot));
 576         rcu_read_unlock();
 577 }
 578
 579 static void process_readd_disk(struct mddev *mddev, struct cluster_msg *msg)
 580 {
 581         struct md_rdev *rdev;
 582
 583         rcu_read_lock();
 584         rdev = md_find_rdev_nr_rcu(mddev, le32_to_cpu(msg->raid_slot));
 585         if (rdev && test_bit(Faulty, &rdev->flags))
 586                 clear_bit(Faulty, &rdev->flags);
 587         else
 588                 pr_warn("%s: %d Could not find disk(%d) which is faulty",
 589                         __func__, __LINE__, le32_to_cpu(msg->raid_slot));
 590         rcu_read_unlock();
 591 }
 592
 593 static int process_recvd_msg(struct mddev *mddev, struct cluster_msg *msg)
 594 {
 595         int ret = 0;
 596
 597         if (WARN(mddev->cluster_info->slot_number - 1 == le32_to_cpu(msg->slot),
 598                 "node %d received its own msg\n", le32_to_cpu(msg->slot)))
 599                 return -1;
 600         switch (le32_to_cpu(msg->type)) {
 601         case METADATA_UPDATED:
 602                 process_metadata_update(mddev, msg);
 603                 break;
 604         case CHANGE_CAPACITY:
 605                 set_capacity_and_notify(mddev->gendisk, mddev->array_sectors);
 606                 break;
 607         case RESYNCING_START:
 608                 clear_bit(MD_CLUSTER_WAITING_FOR_SYNC, &mddev->cluster_info->state);
 609                 break;
 610         case RESYNCING:
 611                 set_bit(MD_RESYNCING_REMOTE, &mddev->recovery);
 612                 process_suspend_info(mddev, le32_to_cpu(msg->slot),
 613                                      le64_to_cpu(msg->low),
 614                                      le64_to_cpu(msg->high));
 615                 break;
 616         case NEWDISK:
 617                 if (process_add_new_disk(mddev, msg))
 618                         ret = -1;
 619                 break;
 620         case REMOVE:
 621                 process_remove_disk(mddev, msg);
 622                 break;
 623         case RE_ADD:
 624                 process_readd_disk(mddev, msg);
 625                 break;
 626         case BITMAP_NEEDS_SYNC:
 627                 __recover_slot(mddev, le32_to_cpu(msg->slot));
 628                 break;
 629         case BITMAP_RESIZE:
 630                 if (le64_to_cpu(msg->high) != mddev->pers->size(mddev, 0, 0))
 631                         ret = md_bitmap_resize(mddev->bitmap,
 632                                             le64_to_cpu(msg->high), 0, 0);
 633                 break;
 634         default:
 635                 ret = -1;
 636                 pr_warn("%s:%d Received unknown message from %d\n",
 637                         __func__, __LINE__, msg->slot);
 638         }
 639         return ret;
 640 }
 641
 642 /*
 643  * thread for receiving message
 644  */
 645 static void recv_daemon(struct md_thread *thread)
 646 {
 647         struct md_cluster_info *cinfo = thread->mddev->cluster_info;
 648         struct dlm_lock_resource *ack_lockres = cinfo->ack_lockres;
 649         struct dlm_lock_resource *message_lockres = cinfo->message_lockres;
 650         struct cluster_msg msg;
 651         int ret;
 652
 653         mutex_lock(&cinfo->recv_mutex);
 654         /*get CR on Message*/
 655         if (dlm_lock_sync(message_lockres, DLM_LOCK_CR)) {
 656                 pr_err("md/raid1:failed to get CR on MESSAGE\n");
 657                 mutex_unlock(&cinfo->recv_mutex);
 658                 return;
 659         }
 660
 661         /* read lvb and wake up thread to process this message_lockres */
 662         memcpy(&msg, message_lockres->lksb.sb_lvbptr, sizeof(struct cluster_msg));
 663         ret = process_recvd_msg(thread->mddev, &msg);
 664         if (ret)
 665                 goto out;
 666
 667         /*release CR on ack_lockres*/
 668         ret = dlm_unlock_sync(ack_lockres);
 669         if (unlikely(ret != 0))
 670                 pr_info("unlock ack failed return %d\n", ret);
 671         /*up-convert to PR on message_lockres*/
 672         ret = dlm_lock_sync(message_lockres, DLM_LOCK_PR);
 673         if (unlikely(ret != 0))
 674                 pr_info("lock PR on msg failed return %d\n", ret);
 675         /*get CR on ack_lockres again*/
 676         ret = dlm_lock_sync(ack_lockres, DLM_LOCK_CR);
 677         if (unlikely(ret != 0))
 678                 pr_info("lock CR on ack failed return %d\n", ret);
 679 out:
 680         /*release CR on message_lockres*/
 681         ret = dlm_unlock_sync(message_lockres);
 682         if (unlikely(ret != 0))
 683                 pr_info("unlock msg failed return %d\n", ret);
 684         mutex_unlock(&cinfo->recv_mutex);
 685 }
 686
 687 /* lock_token()
 688  * Takes the lock on the TOKEN lock resource so no other
 689  * node can communicate while the operation is underway.
 690  */
 691 static int lock_token(struct md_cluster_info *cinfo)
 692 {
 693         int error;
 694
 695         error = dlm_lock_sync(cinfo->token_lockres, DLM_LOCK_EX);
 696         if (error) {
 697                 pr_err("md-cluster(%s:%d): failed to get EX on TOKEN (%d)\n",
 698                                 __func__, __LINE__, error);
 699         } else {
 700                 /* Lock the receive sequence */
 701                 mutex_lock(&cinfo->recv_mutex);
 702         }
 703         return error;
 704 }
 705
 706 /* lock_comm()
 707  * Sets the MD_CLUSTER_SEND_LOCK bit to lock the send channel.
 708  */
 709 static int lock_comm(struct md_cluster_info *cinfo, bool mddev_locked)
 710 {
 711         int rv, set_bit = 0;
 712         struct mddev *mddev = cinfo->mddev;
 713
 714         /*
 715          * If resync thread run after raid1d thread, then process_metadata_update
 716          * could not continue if raid1d held reconfig_mutex (and raid1d is blocked
 717          * since another node already got EX on Token and waiting the EX of Ack),
 718          * so let resync wake up thread in case flag is set.
 719          */
 720         if (mddev_locked && !test_bit(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD,
 721                                       &cinfo->state)) {
 722                 rv = test_and_set_bit_lock(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD,
 723                                               &cinfo->state);
 724                 WARN_ON_ONCE(rv);
 725                 md_wakeup_thread(mddev->thread);
 726                 set_bit = 1;
 727         }
 728
 729         wait_event(cinfo->wait,
 730                    !test_and_set_bit(MD_CLUSTER_SEND_LOCK, &cinfo->state));
 731         rv = lock_token(cinfo);
 732         if (set_bit)
 733                 clear_bit_unlock(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state);
 734         return rv;
 735 }
 736
 737 static void unlock_comm(struct md_cluster_info *cinfo)
 738 {
 739         WARN_ON(cinfo->token_lockres->mode != DLM_LOCK_EX);
 740         mutex_unlock(&cinfo->recv_mutex);
 741         dlm_unlock_sync(cinfo->token_lockres);
 742         clear_bit(MD_CLUSTER_SEND_LOCK, &cinfo->state);
 743         wake_up(&cinfo->wait);
 744 }
 745
 746 /* __sendmsg()
 747  * This function performs the actual sending of the message. This function is
 748  * usually called after performing the encompassing operation
 749  * The function:
 750  * 1. Grabs the message lockresource in EX mode
 751  * 2. Copies the message to the message LVB
 752  * 3. Downconverts message lockresource to CW
 753  * 4. Upconverts ack lock resource from CR to EX. This forces the BAST on other nodes
 754  *    and the other nodes read the message. The thread will wait here until all other
 755  *    nodes have released ack lock resource.
 756  * 5. Downconvert ack lockresource to CR
 757  */
 758 static int __sendmsg(struct md_cluster_info *cinfo, struct cluster_msg *cmsg)
 759 {
 760         int error, unlock_error;
 761         int slot = cinfo->slot_number - 1;
 762
 763         cmsg->slot = cpu_to_le32(slot);
 764         /*get EX on Message*/
 765         error = dlm_lock_sync(cinfo->message_lockres, DLM_LOCK_EX);
 766         if (error) {
 767                 pr_err("md-cluster: failed to get EX on MESSAGE (%d)\n", error);
 768                 return error;
 769         }
 770
 771         memcpy(cinfo->message_lockres->lksb.sb_lvbptr, (void *)cmsg,
 772                         sizeof(struct cluster_msg));
 773         /*down-convert EX to CW on Message*/
 774         error = dlm_lock_sync(cinfo->message_lockres, DLM_LOCK_CW);
 775         if (error) {
 776                 pr_err("md-cluster: failed to convert EX to CW on MESSAGE(%d)\n",
 777                                 error);
 778                 goto failed_ack;
 779         }
 780
 781         /*up-convert CR to EX on Ack*/
 782         error = dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_EX);
 783         if (error) {
 784                 pr_err("md-cluster: failed to convert CR to EX on ACK(%d)\n",
 785                                 error);
 786                 goto failed_ack;
 787         }
 788
 789         /*down-convert EX to CR on Ack*/
 790         error = dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_CR);
 791         if (error) {
 792                 pr_err("md-cluster: failed to convert EX to CR on ACK(%d)\n",
 793                                 error);
 794                 goto failed_ack;
 795         }
 796
 797 failed_ack:
 798         while ((unlock_error = dlm_unlock_sync(cinfo->message_lockres)))
 799                 pr_err("md-cluster: failed convert to NL on MESSAGE(%d)\n",
 800                         unlock_error);
 801
 802         return error;
 803 }
 804
 805 static int sendmsg(struct md_cluster_info *cinfo, struct cluster_msg *cmsg,
 806                    bool mddev_locked)
 807 {
 808         int ret;
 809
 810         ret = lock_comm(cinfo, mddev_locked);
 811         if (!ret) {
 812                 ret = __sendmsg(cinfo, cmsg);
 813                 unlock_comm(cinfo);
 814         }
 815         return ret;
 816 }
 817
 818 static int gather_all_resync_info(struct mddev *mddev, int total_slots)
 819 {
 820         struct md_cluster_info *cinfo = mddev->cluster_info;
 821         int i, ret = 0;
 822         struct dlm_lock_resource *bm_lockres;
 823         char str[64];
 824         sector_t lo, hi;
 825
 826
 827         for (i = 0; i < total_slots; i++) {
 828                 memset(str, '\0', 64);
 829                 snprintf(str, 64, "bitmap%04d", i);
 830                 bm_lockres = lockres_init(mddev, str, NULL, 1);
 831                 if (!bm_lockres)
 832                         return -ENOMEM;
 833                 if (i == (cinfo->slot_number - 1)) {
 834                         lockres_free(bm_lockres);
 835                         continue;
 836                 }
 837
 838                 bm_lockres->flags |= DLM_LKF_NOQUEUE;
 839                 ret = dlm_lock_sync(bm_lockres, DLM_LOCK_PW);
 840                 if (ret == -EAGAIN) {
 841                         if (read_resync_info(mddev, bm_lockres)) {
 842                                 pr_info("%s:%d Resync[%llu..%llu] in progress on %d\n",
 843                                                 __func__, __LINE__,
 844                                         (unsigned long long) cinfo->suspend_lo,
 845                                         (unsigned long long) cinfo->suspend_hi,
 846                                         i);
 847                                 cinfo->suspend_from = i;
 848                         }
 849                         ret = 0;
 850                         lockres_free(bm_lockres);
 851                         continue;
 852                 }
 853                 if (ret) {
 854                         lockres_free(bm_lockres);
 855                         goto out;
 856                 }
 857
 858                 /* Read the disk bitmap sb and check if it needs recovery */
 859                 ret = md_bitmap_copy_from_slot(mddev, i, &lo, &hi, false);
 860                 if (ret) {
 861                         pr_warn("md-cluster: Could not gather bitmaps from slot %d", i);
 862                         lockres_free(bm_lockres);
 863                         continue;
 864                 }
 865                 if ((hi > 0) && (lo < mddev->recovery_cp)) {
 866                         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
 867                         mddev->recovery_cp = lo;
 868                         md_check_recovery(mddev);
 869                 }
 870
 871                 lockres_free(bm_lockres);
 872         }
 873 out:
 874         return ret;
 875 }
 876
 877 static int join(struct mddev *mddev, int nodes)
 878 {
 879         struct md_cluster_info *cinfo;
 880         int ret, ops_rv;
 881         char str[64];
 882
 883         cinfo = kzalloc(sizeof(struct md_cluster_info), GFP_KERNEL);
 884         if (!cinfo)
 885                 return -ENOMEM;
 886
 887         INIT_LIST_HEAD(&cinfo->suspend_list);
 888         spin_lock_init(&cinfo->suspend_lock);
 889         init_completion(&cinfo->completion);
 890         set_bit(MD_CLUSTER_BEGIN_JOIN_CLUSTER, &cinfo->state);
 891         init_waitqueue_head(&cinfo->wait);
 892         mutex_init(&cinfo->recv_mutex);
 893
 894         mddev->cluster_info = cinfo;
 895         cinfo->mddev = mddev;
 896
 897         memset(str, 0, 64);
 898         sprintf(str, "%pU", mddev->uuid);
 899         ret = dlm_new_lockspace(str, mddev->bitmap_info.cluster_name,
 900                                 DLM_LSFL_SOFTIRQ, LVB_SIZE, &md_ls_ops, mddev,
 901                                 &ops_rv, &cinfo->lockspace);
 902         if (ret)
 903                 goto err;
 904         wait_for_completion(&cinfo->completion);
 905         if (nodes < cinfo->slot_number) {
 906                 pr_err("md-cluster: Slot allotted(%d) is greater than available slots(%d).",
 907                         cinfo->slot_number, nodes);
 908                 ret = -ERANGE;
 909                 goto err;
 910         }
 911         /* Initiate the communication resources */
 912         ret = -ENOMEM;
 913         rcu_assign_pointer(cinfo->recv_thread,
 914                         md_register_thread(recv_daemon, mddev, "cluster_recv"));
 915         if (!cinfo->recv_thread) {
 916                 pr_err("md-cluster: cannot allocate memory for recv_thread!\n");
 917                 goto err;
 918         }
 919         cinfo->message_lockres = lockres_init(mddev, "message", NULL, 1);
 920         if (!cinfo->message_lockres)
 921                 goto err;
 922         cinfo->token_lockres = lockres_init(mddev, "token", NULL, 0);
 923         if (!cinfo->token_lockres)
 924                 goto err;
 925         cinfo->no_new_dev_lockres = lockres_init(mddev, "no-new-dev", NULL, 0);
 926         if (!cinfo->no_new_dev_lockres)
 927                 goto err;
 928
 929         ret = dlm_lock_sync(cinfo->token_lockres, DLM_LOCK_EX);
 930         if (ret) {
 931                 ret = -EAGAIN;
 932                 pr_err("md-cluster: can't join cluster to avoid lock issue\n");
 933                 goto err;
 934         }
 935         cinfo->ack_lockres = lockres_init(mddev, "ack", ack_bast, 0);
 936         if (!cinfo->ack_lockres) {
 937                 ret = -ENOMEM;
 938                 goto err;
 939         }
 940         /* get sync CR lock on ACK. */
 941         if (dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_CR))
 942                 pr_err("md-cluster: failed to get a sync CR lock on ACK!(%d)\n",
 943                                 ret);
 944         dlm_unlock_sync(cinfo->token_lockres);
 945         /* get sync CR lock on no-new-dev. */
 946         if (dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR))
 947                 pr_err("md-cluster: failed to get a sync CR lock on no-new-dev!(%d)\n", ret);
 948
 949
 950         pr_info("md-cluster: Joined cluster %s slot %d\n", str, cinfo->slot_number);
 951         snprintf(str, 64, "bitmap%04d", cinfo->slot_number - 1);
 952         cinfo->bitmap_lockres = lockres_init(mddev, str, NULL, 1);
 953         if (!cinfo->bitmap_lockres) {
 954                 ret = -ENOMEM;
 955                 goto err;
 956         }
 957         if (dlm_lock_sync(cinfo->bitmap_lockres, DLM_LOCK_PW)) {
 958                 pr_err("Failed to get bitmap lock\n");
 959                 ret = -EINVAL;
 960                 goto err;
 961         }
 962
 963         cinfo->resync_lockres = lockres_init(mddev, "resync", NULL, 0);
 964         if (!cinfo->resync_lockres) {
 965                 ret = -ENOMEM;
 966                 goto err;
 967         }
 968
 969         return 0;
 970 err:
 971         set_bit(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state);
 972         md_unregister_thread(mddev, &cinfo->recovery_thread);
 973         md_unregister_thread(mddev, &cinfo->recv_thread);
 974         lockres_free(cinfo->message_lockres);
 975         lockres_free(cinfo->token_lockres);
 976         lockres_free(cinfo->ack_lockres);
 977         lockres_free(cinfo->no_new_dev_lockres);
 978         lockres_free(cinfo->resync_lockres);
 979         lockres_free(cinfo->bitmap_lockres);
 980         if (cinfo->lockspace)
 981                 dlm_release_lockspace(cinfo->lockspace, 2);
 982         mddev->cluster_info = NULL;
 983         kfree(cinfo);
 984         return ret;
 985 }
 986
 987 static void load_bitmaps(struct mddev *mddev, int total_slots)
 988 {
 989         struct md_cluster_info *cinfo = mddev->cluster_info;
 990
 991         /* load all the node's bitmap info for resync */
 992         if (gather_all_resync_info(mddev, total_slots))
 993                 pr_err("md-cluster: failed to gather all resyn infos\n");
 994         set_bit(MD_CLUSTER_ALREADY_IN_CLUSTER, &cinfo->state);
 995         /* wake up recv thread in case something need to be handled */
 996         if (test_and_clear_bit(MD_CLUSTER_PENDING_RECV_EVENT, &cinfo->state))
 997                 md_wakeup_thread(cinfo->recv_thread);
 998 }
 999
1000 static void resync_bitmap(struct mddev *mddev)
1001 {
1002         struct md_cluster_info *cinfo = mddev->cluster_info;
1003         struct cluster_msg cmsg = {0};
1004         int err;
1005
1006         cmsg.type = cpu_to_le32(BITMAP_NEEDS_SYNC);
1007         err = sendmsg(cinfo, &cmsg, 1);
1008         if (err)
1009                 pr_err("%s:%d: failed to send BITMAP_NEEDS_SYNC message (%d)\n",
1010                         __func__, __LINE__, err);
1011 }
1012
1013 static void unlock_all_bitmaps(struct mddev *mddev);
1014 static int leave(struct mddev *mddev)
1015 {
1016         struct md_cluster_info *cinfo = mddev->cluster_info;
1017
1018         if (!cinfo)
1019                 return 0;
1020
1021         /*
1022          * BITMAP_NEEDS_SYNC message should be sent when node
1023          * is leaving the cluster with dirty bitmap, also we
1024          * can only deliver it when dlm connection is available.
1025          *
1026          * Also, we should send BITMAP_NEEDS_SYNC message in
1027          * case reshaping is interrupted.
1028          */
1029         if ((cinfo->slot_number > 0 && mddev->recovery_cp != MaxSector) ||
1030             (mddev->reshape_position != MaxSector &&
1031              test_bit(MD_CLOSING, &mddev->flags)))
1032                 resync_bitmap(mddev);
1033
1034         set_bit(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state);
1035         md_unregister_thread(mddev, &cinfo->recovery_thread);
1036         md_unregister_thread(mddev, &cinfo->recv_thread);
1037         lockres_free(cinfo->message_lockres);
1038         lockres_free(cinfo->token_lockres);
1039         lockres_free(cinfo->ack_lockres);
1040         lockres_free(cinfo->no_new_dev_lockres);
1041         lockres_free(cinfo->resync_lockres);
1042         lockres_free(cinfo->bitmap_lockres);
1043         unlock_all_bitmaps(mddev);
1044         dlm_release_lockspace(cinfo->lockspace, 2);
1045         kfree(cinfo);
1046         return 0;
1047 }
1048
1049 /* slot_number(): Returns the MD slot number to use
1050  * DLM starts the slot numbers from 1, wheras cluster-md
1051  * wants the number to be from zero, so we deduct one
1052  */
1053 static int slot_number(struct mddev *mddev)
1054 {
1055         struct md_cluster_info *cinfo = mddev->cluster_info;
1056
1057         return cinfo->slot_number - 1;
1058 }
1059
1060 /*
1061  * Check if the communication is already locked, else lock the communication
1062  * channel.
1063  * If it is already locked, token is in EX mode, and hence lock_token()
1064  * should not be called.
1065  */
1066 static int metadata_update_start(struct mddev *mddev)
1067 {
1068         struct md_cluster_info *cinfo = mddev->cluster_info;
1069         int ret;
1070
1071         /*
1072          * metadata_update_start is always called with the protection of
1073          * reconfig_mutex, so set WAITING_FOR_TOKEN here.
1074          */
1075         ret = test_and_set_bit_lock(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD,
1076                                     &cinfo->state);
1077         WARN_ON_ONCE(ret);
1078         md_wakeup_thread(mddev->thread);
1079
1080         wait_event(cinfo->wait,
1081                    !test_and_set_bit(MD_CLUSTER_SEND_LOCK, &cinfo->state) ||
1082                    test_and_clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state));
1083
1084         /* If token is already locked, return 0 */
1085         if (cinfo->token_lockres->mode == DLM_LOCK_EX) {
1086                 clear_bit_unlock(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state);
1087                 return 0;
1088         }
1089
1090         ret = lock_token(cinfo);
1091         clear_bit_unlock(MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD, &cinfo->state);
1092         return ret;
1093 }
1094
1095 static int metadata_update_finish(struct mddev *mddev)
1096 {
1097         struct md_cluster_info *cinfo = mddev->cluster_info;
1098         struct cluster_msg cmsg;
1099         struct md_rdev *rdev;
1100         int ret = 0;
1101         int raid_slot = -1;
1102
1103         memset(&cmsg, 0, sizeof(cmsg));
1104         cmsg.type = cpu_to_le32(METADATA_UPDATED);
1105         /* Pick up a good active device number to send.
1106          */
1107         rdev_for_each(rdev, mddev)
1108                 if (rdev->raid_disk > -1 && !test_bit(Faulty, &rdev->flags)) {
1109                         raid_slot = rdev->desc_nr;
1110                         break;
1111                 }
1112         if (raid_slot >= 0) {
1113                 cmsg.raid_slot = cpu_to_le32(raid_slot);
1114                 ret = __sendmsg(cinfo, &cmsg);
1115         } else
1116                 pr_warn("md-cluster: No good device id found to send\n");
1117         clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1118         unlock_comm(cinfo);
1119         return ret;
1120 }
1121
1122 static void metadata_update_cancel(struct mddev *mddev)
1123 {
1124         struct md_cluster_info *cinfo = mddev->cluster_info;
1125         clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1126         unlock_comm(cinfo);
1127 }
1128
1129 static int update_bitmap_size(struct mddev *mddev, sector_t size)
1130 {
1131         struct md_cluster_info *cinfo = mddev->cluster_info;
1132         struct cluster_msg cmsg = {0};
1133         int ret;
1134
1135         cmsg.type = cpu_to_le32(BITMAP_RESIZE);
1136         cmsg.high = cpu_to_le64(size);
1137         ret = sendmsg(cinfo, &cmsg, 0);
1138         if (ret)
1139                 pr_err("%s:%d: failed to send BITMAP_RESIZE message (%d)\n",
1140                         __func__, __LINE__, ret);
1141         return ret;
1142 }
1143
1144 static int resize_bitmaps(struct mddev *mddev, sector_t newsize, sector_t oldsize)
1145 {
1146         struct bitmap_counts *counts;
1147         char str[64];
1148         struct dlm_lock_resource *bm_lockres;
1149         struct bitmap *bitmap = mddev->bitmap;
1150         unsigned long my_pages = bitmap->counts.pages;
1151         int i, rv;
1152
1153         /*
1154          * We need to ensure all the nodes can grow to a larger
1155          * bitmap size before make the reshaping.
1156          */
1157         rv = update_bitmap_size(mddev, newsize);
1158         if (rv)
1159                 return rv;
1160
1161         for (i = 0; i < mddev->bitmap_info.nodes; i++) {
1162                 if (i == md_cluster_ops->slot_number(mddev))
1163                         continue;
1164
1165                 bitmap = get_bitmap_from_slot(mddev, i);
1166                 if (IS_ERR(bitmap)) {
1167                         pr_err("can't get bitmap from slot %d\n", i);
1168                         bitmap = NULL;
1169                         goto out;
1170                 }
1171                 counts = &bitmap->counts;
1172
1173                 /*
1174                  * If we can hold the bitmap lock of one node then
1175                  * the slot is not occupied, update the pages.
1176                  */
1177                 snprintf(str, 64, "bitmap%04d", i);
1178                 bm_lockres = lockres_init(mddev, str, NULL, 1);
1179                 if (!bm_lockres) {
1180                         pr_err("Cannot initialize %s lock\n", str);
1181                         goto out;
1182                 }
1183                 bm_lockres->flags |= DLM_LKF_NOQUEUE;
1184                 rv = dlm_lock_sync(bm_lockres, DLM_LOCK_PW);
1185                 if (!rv)
1186                         counts->pages = my_pages;
1187                 lockres_free(bm_lockres);
1188
1189                 if (my_pages != counts->pages)
1190                         /*
1191                          * Let's revert the bitmap size if one node
1192                          * can't resize bitmap
1193                          */
1194                         goto out;
1195                 md_bitmap_free(bitmap);
1196         }
1197
1198         return 0;
1199 out:
1200         md_bitmap_free(bitmap);
1201         update_bitmap_size(mddev, oldsize);
1202         return -1;
1203 }
1204
1205 /*
1206  * return 0 if all the bitmaps have the same sync_size
1207  */
1208 static int cluster_check_sync_size(struct mddev *mddev)
1209 {
1210         int i, rv;
1211         bitmap_super_t *sb;
1212         unsigned long my_sync_size, sync_size = 0;
1213         int node_num = mddev->bitmap_info.nodes;
1214         int current_slot = md_cluster_ops->slot_number(mddev);
1215         struct bitmap *bitmap = mddev->bitmap;
1216         char str[64];
1217         struct dlm_lock_resource *bm_lockres;
1218
1219         sb = kmap_atomic(bitmap->storage.sb_page);
1220         my_sync_size = sb->sync_size;
1221         kunmap_atomic(sb);
1222
1223         for (i = 0; i < node_num; i++) {
1224                 if (i == current_slot)
1225                         continue;
1226
1227                 bitmap = get_bitmap_from_slot(mddev, i);
1228                 if (IS_ERR(bitmap)) {
1229                         pr_err("can't get bitmap from slot %d\n", i);
1230                         return -1;
1231                 }
1232
1233                 /*
1234                  * If we can hold the bitmap lock of one node then
1235                  * the slot is not occupied, update the sb.
1236                  */
1237                 snprintf(str, 64, "bitmap%04d", i);
1238                 bm_lockres = lockres_init(mddev, str, NULL, 1);
1239                 if (!bm_lockres) {
1240                         pr_err("md-cluster: Cannot initialize %s\n", str);
1241                         md_bitmap_free(bitmap);
1242                         return -1;
1243                 }
1244                 bm_lockres->flags |= DLM_LKF_NOQUEUE;
1245                 rv = dlm_lock_sync(bm_lockres, DLM_LOCK_PW);
1246                 if (!rv)
1247                         md_bitmap_update_sb(bitmap);
1248                 lockres_free(bm_lockres);
1249
1250                 sb = kmap_atomic(bitmap->storage.sb_page);
1251                 if (sync_size == 0)
1252                         sync_size = sb->sync_size;
1253                 else if (sync_size != sb->sync_size) {
1254                         kunmap_atomic(sb);
1255                         md_bitmap_free(bitmap);
1256                         return -1;
1257                 }
1258                 kunmap_atomic(sb);
1259                 md_bitmap_free(bitmap);
1260         }
1261
1262         return (my_sync_size == sync_size) ? 0 : -1;
1263 }
1264
1265 /*
1266  * Update the size for cluster raid is a little more complex, we perform it
1267  * by the steps:
1268  * 1. hold token lock and update superblock in initiator node.
1269  * 2. send METADATA_UPDATED msg to other nodes.
1270  * 3. The initiator node continues to check each bitmap's sync_size, if all
1271  *    bitmaps have the same value of sync_size, then we can set capacity and
1272  *    let other nodes to perform it. If one node can't update sync_size
1273  *    accordingly, we need to revert to previous value.
1274  */
1275 static void update_size(struct mddev *mddev, sector_t old_dev_sectors)
1276 {
1277         struct md_cluster_info *cinfo = mddev->cluster_info;
1278         struct cluster_msg cmsg;
1279         struct md_rdev *rdev;
1280         int ret = 0;
1281         int raid_slot = -1;
1282
1283         md_update_sb(mddev, 1);
1284         if (lock_comm(cinfo, 1)) {
1285                 pr_err("%s: lock_comm failed\n", __func__);
1286                 return;
1287         }
1288
1289         memset(&cmsg, 0, sizeof(cmsg));
1290         cmsg.type = cpu_to_le32(METADATA_UPDATED);
1291         rdev_for_each(rdev, mddev)
1292                 if (rdev->raid_disk >= 0 && !test_bit(Faulty, &rdev->flags)) {
1293                         raid_slot = rdev->desc_nr;
1294                         break;
1295                 }
1296         if (raid_slot >= 0) {
1297                 cmsg.raid_slot = cpu_to_le32(raid_slot);
1298                 /*
1299                  * We can only change capiticy after all the nodes can do it,
1300                  * so need to wait after other nodes already received the msg
1301                  * and handled the change
1302                  */
1303                 ret = __sendmsg(cinfo, &cmsg);
1304                 if (ret) {
1305                         pr_err("%s:%d: failed to send METADATA_UPDATED msg\n",
1306                                __func__, __LINE__);
1307                         unlock_comm(cinfo);
1308                         return;
1309                 }
1310         } else {
1311                 pr_err("md-cluster: No good device id found to send\n");
1312                 unlock_comm(cinfo);
1313                 return;
1314         }
1315
1316         /*
1317          * check the sync_size from other node's bitmap, if sync_size
1318          * have already updated in other nodes as expected, send an
1319          * empty metadata msg to permit the change of capacity
1320          */
1321         if (cluster_check_sync_size(mddev) == 0) {
1322                 memset(&cmsg, 0, sizeof(cmsg));
1323                 cmsg.type = cpu_to_le32(CHANGE_CAPACITY);
1324                 ret = __sendmsg(cinfo, &cmsg);
1325                 if (ret)
1326                         pr_err("%s:%d: failed to send CHANGE_CAPACITY msg\n",
1327                                __func__, __LINE__);
1328                 set_capacity_and_notify(mddev->gendisk, mddev->array_sectors);
1329         } else {
1330                 /* revert to previous sectors */
1331                 ret = mddev->pers->resize(mddev, old_dev_sectors);
1332                 ret = __sendmsg(cinfo, &cmsg);
1333                 if (ret)
1334                         pr_err("%s:%d: failed to send METADATA_UPDATED msg\n",
1335                                __func__, __LINE__);
1336         }
1337         unlock_comm(cinfo);
1338 }
1339
1340 static int resync_start(struct mddev *mddev)
1341 {
1342         struct md_cluster_info *cinfo = mddev->cluster_info;
1343         return dlm_lock_sync_interruptible(cinfo->resync_lockres, DLM_LOCK_EX, mddev);
1344 }
1345
1346 static void resync_info_get(struct mddev *mddev, sector_t *lo, sector_t *hi)
1347 {
1348         struct md_cluster_info *cinfo = mddev->cluster_info;
1349
1350         spin_lock_irq(&cinfo->suspend_lock);
1351         *lo = cinfo->suspend_lo;
1352         *hi = cinfo->suspend_hi;
1353         spin_unlock_irq(&cinfo->suspend_lock);
1354 }
1355
1356 static int resync_status_get(struct mddev *mddev)
1357 {
1358         struct md_cluster_info *cinfo = mddev->cluster_info;
1359
1360         return test_bit(MD_CLUSTER_WAITING_FOR_SYNC, &cinfo->state);
1361 }
1362
1363 static int resync_start_notify(struct mddev *mddev)
1364 {
1365         struct md_cluster_info *cinfo = mddev->cluster_info;
1366         struct cluster_msg cmsg = {0};
1367
1368         cmsg.type = cpu_to_le32(RESYNCING_START);
1369
1370         return sendmsg(cinfo, &cmsg, 0);
1371 }
1372
1373 static int resync_info_update(struct mddev *mddev, sector_t lo, sector_t hi)
1374 {
1375         struct md_cluster_info *cinfo = mddev->cluster_info;
1376         struct resync_info ri;
1377         struct cluster_msg cmsg = {0};
1378
1379         /* do not send zero again, if we have sent before */
1380         if (hi == 0) {
1381                 memcpy(&ri, cinfo->bitmap_lockres->lksb.sb_lvbptr, sizeof(struct resync_info));
1382                 if (le64_to_cpu(ri.hi) == 0)
1383                         return 0;
1384         }
1385
1386         add_resync_info(cinfo->bitmap_lockres, lo, hi);
1387         /* Re-acquire the lock to refresh LVB */
1388         dlm_lock_sync(cinfo->bitmap_lockres, DLM_LOCK_PW);
1389         cmsg.type = cpu_to_le32(RESYNCING);
1390         cmsg.low = cpu_to_le64(lo);
1391         cmsg.high = cpu_to_le64(hi);
1392
1393         /*
1394          * mddev_lock is held if resync_info_update is called from
1395          * resync_finish (md_reap_sync_thread -> resync_finish)
1396          */
1397         if (lo == 0 && hi == 0)
1398                 return sendmsg(cinfo, &cmsg, 1);
1399         else
1400                 return sendmsg(cinfo, &cmsg, 0);
1401 }
1402
1403 static int resync_finish(struct mddev *mddev)
1404 {
1405         struct md_cluster_info *cinfo = mddev->cluster_info;
1406         int ret = 0;
1407
1408         clear_bit(MD_RESYNCING_REMOTE, &mddev->recovery);
1409
1410         /*
1411          * If resync thread is interrupted so we can't say resync is finished,
1412          * another node will launch resync thread to continue.
1413          */
1414         if (!test_bit(MD_CLOSING, &mddev->flags))
1415                 ret = resync_info_update(mddev, 0, 0);
1416         dlm_unlock_sync(cinfo->resync_lockres);
1417         return ret;
1418 }
1419
1420 static int area_resyncing(struct mddev *mddev, int direction,
1421                 sector_t lo, sector_t hi)
1422 {
1423         struct md_cluster_info *cinfo = mddev->cluster_info;
1424         int ret = 0;
1425
1426         if ((direction == READ) &&
1427                 test_bit(MD_CLUSTER_SUSPEND_READ_BALANCING, &cinfo->state))
1428                 return 1;
1429
1430         spin_lock_irq(&cinfo->suspend_lock);
1431         if (hi > cinfo->suspend_lo && lo < cinfo->suspend_hi)
1432                 ret = 1;
1433         spin_unlock_irq(&cinfo->suspend_lock);
1434         return ret;
1435 }
1436
1437 /* add_new_disk() - initiates a disk add
1438  * However, if this fails before writing md_update_sb(),
1439  * add_new_disk_cancel() must be called to release token lock
1440  */
1441 static int add_new_disk(struct mddev *mddev, struct md_rdev *rdev)
1442 {
1443         struct md_cluster_info *cinfo = mddev->cluster_info;
1444         struct cluster_msg cmsg;
1445         int ret = 0;
1446         struct mdp_superblock_1 *sb = page_address(rdev->sb_page);
1447         char *uuid = sb->device_uuid;
1448
1449         memset(&cmsg, 0, sizeof(cmsg));
1450         cmsg.type = cpu_to_le32(NEWDISK);
1451         memcpy(cmsg.uuid, uuid, 16);
1452         cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
1453         if (lock_comm(cinfo, 1))
1454                 return -EAGAIN;
1455         ret = __sendmsg(cinfo, &cmsg);
1456         if (ret) {
1457                 unlock_comm(cinfo);
1458                 return ret;
1459         }
1460         cinfo->no_new_dev_lockres->flags |= DLM_LKF_NOQUEUE;
1461         ret = dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_EX);
1462         cinfo->no_new_dev_lockres->flags &= ~DLM_LKF_NOQUEUE;
1463         /* Some node does not "see" the device */
1464         if (ret == -EAGAIN)
1465                 ret = -ENOENT;
1466         if (ret)
1467                 unlock_comm(cinfo);
1468         else {
1469                 dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR);
1470                 /* Since MD_CHANGE_DEVS will be set in add_bound_rdev which
1471                  * will run soon after add_new_disk, the below path will be
1472                  * invoked:
1473                  *   md_wakeup_thread(mddev->thread)
1474                  *      -> conf->thread (raid1d)
1475                  *      -> md_check_recovery -> md_update_sb
1476                  *      -> metadata_update_start/finish
1477                  * MD_CLUSTER_SEND_LOCKED_ALREADY will be cleared eventually.
1478                  *
1479                  * For other failure cases, metadata_update_cancel and
1480                  * add_new_disk_cancel also clear below bit as well.
1481                  * */
1482                 set_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1483                 wake_up(&cinfo->wait);
1484         }
1485         return ret;
1486 }
1487
1488 static void add_new_disk_cancel(struct mddev *mddev)
1489 {
1490         struct md_cluster_info *cinfo = mddev->cluster_info;
1491         clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1492         unlock_comm(cinfo);
1493 }
1494
1495 static int new_disk_ack(struct mddev *mddev, bool ack)
1496 {
1497         struct md_cluster_info *cinfo = mddev->cluster_info;
1498
1499         if (!test_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state)) {
1500                 pr_warn("md-cluster(%s): Spurious cluster confirmation\n", mdname(mddev));
1501                 return -EINVAL;
1502         }
1503
1504         if (ack)
1505                 dlm_unlock_sync(cinfo->no_new_dev_lockres);
1506         complete(&cinfo->newdisk_completion);
1507         return 0;
1508 }
1509
1510 static int remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1511 {
1512         struct cluster_msg cmsg = {0};
1513         struct md_cluster_info *cinfo = mddev->cluster_info;
1514         cmsg.type = cpu_to_le32(REMOVE);
1515         cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
1516         return sendmsg(cinfo, &cmsg, 1);
1517 }
1518
1519 static int lock_all_bitmaps(struct mddev *mddev)
1520 {
1521         int slot, my_slot, ret, held = 1, i = 0;
1522         char str[64];
1523         struct md_cluster_info *cinfo = mddev->cluster_info;
1524
1525         cinfo->other_bitmap_lockres =
1526                 kcalloc(mddev->bitmap_info.nodes - 1,
1527                         sizeof(struct dlm_lock_resource *), GFP_KERNEL);
1528         if (!cinfo->other_bitmap_lockres) {
1529                 pr_err("md: can't alloc mem for other bitmap locks\n");
1530                 return 0;
1531         }
1532
1533         my_slot = slot_number(mddev);
1534         for (slot = 0; slot < mddev->bitmap_info.nodes; slot++) {
1535                 if (slot == my_slot)
1536                         continue;
1537
1538                 memset(str, '\0', 64);
1539                 snprintf(str, 64, "bitmap%04d", slot);
1540                 cinfo->other_bitmap_lockres[i] = lockres_init(mddev, str, NULL, 1);
1541                 if (!cinfo->other_bitmap_lockres[i])
1542                         return -ENOMEM;
1543
1544                 cinfo->other_bitmap_lockres[i]->flags |= DLM_LKF_NOQUEUE;
1545                 ret = dlm_lock_sync(cinfo->other_bitmap_lockres[i], DLM_LOCK_PW);
1546                 if (ret)
1547                         held = -1;
1548                 i++;
1549         }
1550
1551         return held;
1552 }
1553
1554 static void unlock_all_bitmaps(struct mddev *mddev)
1555 {
1556         struct md_cluster_info *cinfo = mddev->cluster_info;
1557         int i;
1558
1559         /* release other node's bitmap lock if they are existed */
1560         if (cinfo->other_bitmap_lockres) {
1561                 for (i = 0; i < mddev->bitmap_info.nodes - 1; i++) {
1562                         if (cinfo->other_bitmap_lockres[i]) {
1563                                 lockres_free(cinfo->other_bitmap_lockres[i]);
1564                         }
1565                 }
1566                 kfree(cinfo->other_bitmap_lockres);
1567                 cinfo->other_bitmap_lockres = NULL;
1568         }
1569 }
1570
1571 static int gather_bitmaps(struct md_rdev *rdev)
1572 {
1573         int sn, err;
1574         sector_t lo, hi;
1575         struct cluster_msg cmsg = {0};
1576         struct mddev *mddev = rdev->mddev;
1577         struct md_cluster_info *cinfo = mddev->cluster_info;
1578
1579         cmsg.type = cpu_to_le32(RE_ADD);
1580         cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
1581         err = sendmsg(cinfo, &cmsg, 1);
1582         if (err)
1583                 goto out;
1584
1585         for (sn = 0; sn < mddev->bitmap_info.nodes; sn++) {
1586                 if (sn == (cinfo->slot_number - 1))
1587                         continue;
1588                 err = md_bitmap_copy_from_slot(mddev, sn, &lo, &hi, false);
1589                 if (err) {
1590                         pr_warn("md-cluster: Could not gather bitmaps from slot %d", sn);
1591                         goto out;
1592                 }
1593                 if ((hi > 0) && (lo < mddev->recovery_cp))
1594                         mddev->recovery_cp = lo;
1595         }
1596 out:
1597         return err;
1598 }
1599
1600 static const struct md_cluster_operations cluster_ops = {
1601         .join   = join,
1602         .leave  = leave,
1603         .slot_number = slot_number,
1604         .resync_start = resync_start,
1605         .resync_finish = resync_finish,
1606         .resync_info_update = resync_info_update,
1607         .resync_start_notify = resync_start_notify,
1608         .resync_status_get = resync_status_get,
1609         .resync_info_get = resync_info_get,
1610         .metadata_update_start = metadata_update_start,
1611         .metadata_update_finish = metadata_update_finish,
1612         .metadata_update_cancel = metadata_update_cancel,
1613         .area_resyncing = area_resyncing,
1614         .add_new_disk = add_new_disk,
1615         .add_new_disk_cancel = add_new_disk_cancel,
1616         .new_disk_ack = new_disk_ack,
1617         .remove_disk = remove_disk,
1618         .load_bitmaps = load_bitmaps,
1619         .gather_bitmaps = gather_bitmaps,
1620         .resize_bitmaps = resize_bitmaps,
1621         .lock_all_bitmaps = lock_all_bitmaps,
1622         .unlock_all_bitmaps = unlock_all_bitmaps,
1623         .update_size = update_size,
1624 };
1625
1626 static int __init cluster_init(void)
1627 {
1628         pr_warn("md-cluster: support raid1 and raid10 (limited support)\n");
1629         pr_info("Registering Cluster MD functions\n");
1630         register_md_cluster_operations(&cluster_ops, THIS_MODULE);
1631         return 0;
1632 }
1633
1634 static void cluster_exit(void)
1635 {
1636         unregister_md_cluster_operations();
1637 }
1638
1639 module_init(cluster_init);
1640 module_exit(cluster_exit);
1641 MODULE_AUTHOR("SUSE");
1642 MODULE_LICENSE("GPL");
1643 MODULE_DESCRIPTION("Clustering support for MD");