drivers/md/md-cluster.c

   1 /*
   2  * Copyright (C) 2015, SUSE
   3  *
   4  * This program is free software; you can redistribute it and/or modify
   5  * it under the terms of the GNU General Public License as published by
   6  * the Free Software Foundation; either version 2, or (at your option)
   7  * any later version.
   8  *
   9  */
  10
  11
  12 #include <linux/module.h>
  13 #include <linux/dlm.h>
  14 #include <linux/sched.h>
  15 #include <linux/raid/md_p.h>
  16 #include "md.h"
  17 #include "bitmap.h"
  18 #include "md-cluster.h"
  19
  20 #define LVB_SIZE        64
  21 #define NEW_DEV_TIMEOUT 5000
  22
  23 struct dlm_lock_resource {
  24         dlm_lockspace_t *ls;
  25         struct dlm_lksb lksb;
  26         char *name; /* lock name. */
  27         uint32_t flags; /* flags to pass to dlm_lock() */
  28         struct completion completion; /* completion for synchronized locking */
  29         void (*bast)(void *arg, int mode); /* blocking AST function pointer*/
  30         struct mddev *mddev; /* pointing back to mddev. */
  31         int mode;
  32 };
  33
  34 struct suspend_info {
  35         int slot;
  36         sector_t lo;
  37         sector_t hi;
  38         struct list_head list;
  39 };
  40
  41 struct resync_info {
  42         __le64 lo;
  43         __le64 hi;
  44 };
  45
  46 /* md_cluster_info flags */
  47 #define         MD_CLUSTER_WAITING_FOR_NEWDISK          1
  48 #define         MD_CLUSTER_SUSPEND_READ_BALANCING       2
  49 #define         MD_CLUSTER_BEGIN_JOIN_CLUSTER           3
  50
  51
  52 struct md_cluster_info {
  53         /* dlm lock space and resources for clustered raid. */
  54         dlm_lockspace_t *lockspace;
  55         int slot_number;
  56         struct completion completion;
  57         struct mutex sb_mutex;
  58         struct dlm_lock_resource *bitmap_lockres;
  59         struct dlm_lock_resource *resync_lockres;
  60         struct list_head suspend_list;
  61         spinlock_t suspend_lock;
  62         struct md_thread *recovery_thread;
  63         unsigned long recovery_map;
  64         /* communication loc resources */
  65         struct dlm_lock_resource *ack_lockres;
  66         struct dlm_lock_resource *message_lockres;
  67         struct dlm_lock_resource *token_lockres;
  68         struct dlm_lock_resource *no_new_dev_lockres;
  69         struct md_thread *recv_thread;
  70         struct completion newdisk_completion;
  71         unsigned long state;
  72 };
  73
  74 enum msg_type {
  75         METADATA_UPDATED = 0,
  76         RESYNCING,
  77         NEWDISK,
  78         REMOVE,
  79         RE_ADD,
  80         BITMAP_NEEDS_SYNC,
  81 };
  82
  83 struct cluster_msg {
  84         int type;
  85         int slot;
  86         /* TODO: Unionize this for smaller footprint */
  87         sector_t low;
  88         sector_t high;
  89         char uuid[16];
  90         int raid_slot;
  91 };
  92
  93 static void sync_ast(void *arg)
  94 {
  95         struct dlm_lock_resource *res;
  96
  97         res = (struct dlm_lock_resource *) arg;
  98         complete(&res->completion);
  99 }
 100
 101 static int dlm_lock_sync(struct dlm_lock_resource *res, int mode)
 102 {
 103         int ret = 0;
 104
 105         ret = dlm_lock(res->ls, mode, &res->lksb,
 106                         res->flags, res->name, strlen(res->name),
 107                         0, sync_ast, res, res->bast);
 108         if (ret)
 109                 return ret;
 110         wait_for_completion(&res->completion);
 111         if (res->lksb.sb_status == 0)
 112                 res->mode = mode;
 113         return res->lksb.sb_status;
 114 }
 115
 116 static int dlm_unlock_sync(struct dlm_lock_resource *res)
 117 {
 118         return dlm_lock_sync(res, DLM_LOCK_NL);
 119 }
 120
 121 static struct dlm_lock_resource *lockres_init(struct mddev *mddev,
 122                 char *name, void (*bastfn)(void *arg, int mode), int with_lvb)
 123 {
 124         struct dlm_lock_resource *res = NULL;
 125         int ret, namelen;
 126         struct md_cluster_info *cinfo = mddev->cluster_info;
 127
 128         res = kzalloc(sizeof(struct dlm_lock_resource), GFP_KERNEL);
 129         if (!res)
 130                 return NULL;
 131         init_completion(&res->completion);
 132         res->ls = cinfo->lockspace;
 133         res->mddev = mddev;
 134         res->mode = DLM_LOCK_IV;
 135         namelen = strlen(name);
 136         res->name = kzalloc(namelen + 1, GFP_KERNEL);
 137         if (!res->name) {
 138                 pr_err("md-cluster: Unable to allocate resource name for resource %s\n", name);
 139                 goto out_err;
 140         }
 141         strlcpy(res->name, name, namelen + 1);
 142         if (with_lvb) {
 143                 res->lksb.sb_lvbptr = kzalloc(LVB_SIZE, GFP_KERNEL);
 144                 if (!res->lksb.sb_lvbptr) {
 145                         pr_err("md-cluster: Unable to allocate LVB for resource %s\n", name);
 146                         goto out_err;
 147                 }
 148                 res->flags = DLM_LKF_VALBLK;
 149         }
 150
 151         if (bastfn)
 152                 res->bast = bastfn;
 153
 154         res->flags |= DLM_LKF_EXPEDITE;
 155
 156         ret = dlm_lock_sync(res, DLM_LOCK_NL);
 157         if (ret) {
 158                 pr_err("md-cluster: Unable to lock NL on new lock resource %s\n", name);
 159                 goto out_err;
 160         }
 161         res->flags &= ~DLM_LKF_EXPEDITE;
 162         res->flags |= DLM_LKF_CONVERT;
 163
 164         return res;
 165 out_err:
 166         kfree(res->lksb.sb_lvbptr);
 167         kfree(res->name);
 168         kfree(res);
 169         return NULL;
 170 }
 171
 172 static void lockres_free(struct dlm_lock_resource *res)
 173 {
 174         int ret;
 175
 176         if (!res)
 177                 return;
 178
 179         /* cancel a lock request or a conversion request that is blocked */
 180         res->flags |= DLM_LKF_CANCEL;
 181 retry:
 182         ret = dlm_unlock(res->ls, res->lksb.sb_lkid, 0, &res->lksb, res);
 183         if (unlikely(ret != 0)) {
 184                 pr_info("%s: failed to unlock %s return %d\n", __func__, res->name, ret);
 185
 186                 /* if a lock conversion is cancelled, then the lock is put
 187                  * back to grant queue, need to ensure it is unlocked */
 188                 if (ret == -DLM_ECANCEL)
 189                         goto retry;
 190         }
 191         res->flags &= ~DLM_LKF_CANCEL;
 192         wait_for_completion(&res->completion);
 193
 194         kfree(res->name);
 195         kfree(res->lksb.sb_lvbptr);
 196         kfree(res);
 197 }
 198
 199 static void add_resync_info(struct mddev *mddev, struct dlm_lock_resource *lockres,
 200                 sector_t lo, sector_t hi)
 201 {
 202         struct resync_info *ri;
 203
 204         ri = (struct resync_info *)lockres->lksb.sb_lvbptr;
 205         ri->lo = cpu_to_le64(lo);
 206         ri->hi = cpu_to_le64(hi);
 207 }
 208
 209 static struct suspend_info *read_resync_info(struct mddev *mddev, struct dlm_lock_resource *lockres)
 210 {
 211         struct resync_info ri;
 212         struct suspend_info *s = NULL;
 213         sector_t hi = 0;
 214
 215         dlm_lock_sync(lockres, DLM_LOCK_CR);
 216         memcpy(&ri, lockres->lksb.sb_lvbptr, sizeof(struct resync_info));
 217         hi = le64_to_cpu(ri.hi);
 218         if (ri.hi > 0) {
 219                 s = kzalloc(sizeof(struct suspend_info), GFP_KERNEL);
 220                 if (!s)
 221                         goto out;
 222                 s->hi = hi;
 223                 s->lo = le64_to_cpu(ri.lo);
 224         }
 225         dlm_unlock_sync(lockres);
 226 out:
 227         return s;
 228 }
 229
 230 static void recover_bitmaps(struct md_thread *thread)
 231 {
 232         struct mddev *mddev = thread->mddev;
 233         struct md_cluster_info *cinfo = mddev->cluster_info;
 234         struct dlm_lock_resource *bm_lockres;
 235         char str[64];
 236         int slot, ret;
 237         struct suspend_info *s, *tmp;
 238         sector_t lo, hi;
 239
 240         while (cinfo->recovery_map) {
 241                 slot = fls64((u64)cinfo->recovery_map) - 1;
 242
 243                 /* Clear suspend_area associated with the bitmap */
 244                 spin_lock_irq(&cinfo->suspend_lock);
 245                 list_for_each_entry_safe(s, tmp, &cinfo->suspend_list, list)
 246                         if (slot == s->slot) {
 247                                 list_del(&s->list);
 248                                 kfree(s);
 249                         }
 250                 spin_unlock_irq(&cinfo->suspend_lock);
 251
 252                 snprintf(str, 64, "bitmap%04d", slot);
 253                 bm_lockres = lockres_init(mddev, str, NULL, 1);
 254                 if (!bm_lockres) {
 255                         pr_err("md-cluster: Cannot initialize bitmaps\n");
 256                         goto clear_bit;
 257                 }
 258
 259                 ret = dlm_lock_sync(bm_lockres, DLM_LOCK_PW);
 260                 if (ret) {
 261                         pr_err("md-cluster: Could not DLM lock %s: %d\n",
 262                                         str, ret);
 263                         goto clear_bit;
 264                 }
 265                 ret = bitmap_copy_from_slot(mddev, slot, &lo, &hi, true);
 266                 if (ret) {
 267                         pr_err("md-cluster: Could not copy data from bitmap %d\n", slot);
 268                         goto dlm_unlock;
 269                 }
 270                 if (hi > 0) {
 271                         /* TODO:Wait for current resync to get over */
 272                         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
 273                         if (lo < mddev->recovery_cp)
 274                                 mddev->recovery_cp = lo;
 275                         md_check_recovery(mddev);
 276                 }
 277 dlm_unlock:
 278                 dlm_unlock_sync(bm_lockres);
 279 clear_bit:
 280                 clear_bit(slot, &cinfo->recovery_map);
 281         }
 282 }
 283
 284 static void recover_prep(void *arg)
 285 {
 286         struct mddev *mddev = arg;
 287         struct md_cluster_info *cinfo = mddev->cluster_info;
 288         set_bit(MD_CLUSTER_SUSPEND_READ_BALANCING, &cinfo->state);
 289 }
 290
 291 static void __recover_slot(struct mddev *mddev, int slot)
 292 {
 293         struct md_cluster_info *cinfo = mddev->cluster_info;
 294
 295         set_bit(slot, &cinfo->recovery_map);
 296         if (!cinfo->recovery_thread) {
 297                 cinfo->recovery_thread = md_register_thread(recover_bitmaps,
 298                                 mddev, "recover");
 299                 if (!cinfo->recovery_thread) {
 300                         pr_warn("md-cluster: Could not create recovery thread\n");
 301                         return;
 302                 }
 303         }
 304         md_wakeup_thread(cinfo->recovery_thread);
 305 }
 306
 307 static void recover_slot(void *arg, struct dlm_slot *slot)
 308 {
 309         struct mddev *mddev = arg;
 310         struct md_cluster_info *cinfo = mddev->cluster_info;
 311
 312         pr_info("md-cluster: %s Node %d/%d down. My slot: %d. Initiating recovery.\n",
 313                         mddev->bitmap_info.cluster_name,
 314                         slot->nodeid, slot->slot,
 315                         cinfo->slot_number);
 316         /* deduct one since dlm slot starts from one while the num of
 317          * cluster-md begins with 0 */
 318         __recover_slot(mddev, slot->slot - 1);
 319 }
 320
 321 static void recover_done(void *arg, struct dlm_slot *slots,
 322                 int num_slots, int our_slot,
 323                 uint32_t generation)
 324 {
 325         struct mddev *mddev = arg;
 326         struct md_cluster_info *cinfo = mddev->cluster_info;
 327
 328         cinfo->slot_number = our_slot;
 329         /* completion is only need to be complete when node join cluster,
 330          * it doesn't need to run during another node's failure */
 331         if (test_bit(MD_CLUSTER_BEGIN_JOIN_CLUSTER, &cinfo->state)) {
 332                 complete(&cinfo->completion);
 333                 clear_bit(MD_CLUSTER_BEGIN_JOIN_CLUSTER, &cinfo->state);
 334         }
 335         clear_bit(MD_CLUSTER_SUSPEND_READ_BALANCING, &cinfo->state);
 336 }
 337
 338 /* the ops is called when node join the cluster, and do lock recovery
 339  * if node failure occurs */
 340 static const struct dlm_lockspace_ops md_ls_ops = {
 341         .recover_prep = recover_prep,
 342         .recover_slot = recover_slot,
 343         .recover_done = recover_done,
 344 };
 345
 346 /*
 347  * The BAST function for the ack lock resource
 348  * This function wakes up the receive thread in
 349  * order to receive and process the message.
 350  */
 351 static void ack_bast(void *arg, int mode)
 352 {
 353         struct dlm_lock_resource *res = (struct dlm_lock_resource *)arg;
 354         struct md_cluster_info *cinfo = res->mddev->cluster_info;
 355
 356         if (mode == DLM_LOCK_EX)
 357                 md_wakeup_thread(cinfo->recv_thread);
 358 }
 359
 360 static void __remove_suspend_info(struct md_cluster_info *cinfo, int slot)
 361 {
 362         struct suspend_info *s, *tmp;
 363
 364         list_for_each_entry_safe(s, tmp, &cinfo->suspend_list, list)
 365                 if (slot == s->slot) {
 366                         list_del(&s->list);
 367                         kfree(s);
 368                         break;
 369                 }
 370 }
 371
 372 static void remove_suspend_info(struct mddev *mddev, int slot)
 373 {
 374         struct md_cluster_info *cinfo = mddev->cluster_info;
 375         spin_lock_irq(&cinfo->suspend_lock);
 376         __remove_suspend_info(cinfo, slot);
 377         spin_unlock_irq(&cinfo->suspend_lock);
 378         mddev->pers->quiesce(mddev, 2);
 379 }
 380
 381
 382 static void process_suspend_info(struct mddev *mddev,
 383                 int slot, sector_t lo, sector_t hi)
 384 {
 385         struct md_cluster_info *cinfo = mddev->cluster_info;
 386         struct suspend_info *s;
 387
 388         if (!hi) {
 389                 remove_suspend_info(mddev, slot);
 390                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
 391                 md_wakeup_thread(mddev->thread);
 392                 return;
 393         }
 394         s = kzalloc(sizeof(struct suspend_info), GFP_KERNEL);
 395         if (!s)
 396                 return;
 397         s->slot = slot;
 398         s->lo = lo;
 399         s->hi = hi;
 400         mddev->pers->quiesce(mddev, 1);
 401         mddev->pers->quiesce(mddev, 0);
 402         spin_lock_irq(&cinfo->suspend_lock);
 403         /* Remove existing entry (if exists) before adding */
 404         __remove_suspend_info(cinfo, slot);
 405         list_add(&s->list, &cinfo->suspend_list);
 406         spin_unlock_irq(&cinfo->suspend_lock);
 407         mddev->pers->quiesce(mddev, 2);
 408 }
 409
 410 static void process_add_new_disk(struct mddev *mddev, struct cluster_msg *cmsg)
 411 {
 412         char disk_uuid[64];
 413         struct md_cluster_info *cinfo = mddev->cluster_info;
 414         char event_name[] = "EVENT=ADD_DEVICE";
 415         char raid_slot[16];
 416         char *envp[] = {event_name, disk_uuid, raid_slot, NULL};
 417         int len;
 418
 419         len = snprintf(disk_uuid, 64, "DEVICE_UUID=");
 420         sprintf(disk_uuid + len, "%pU", cmsg->uuid);
 421         snprintf(raid_slot, 16, "RAID_DISK=%d", le32_to_cpu(cmsg->raid_slot));
 422         pr_info("%s:%d Sending kobject change with %s and %s\n", __func__, __LINE__, disk_uuid, raid_slot);
 423         init_completion(&cinfo->newdisk_completion);
 424         set_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state);
 425         kobject_uevent_env(&disk_to_dev(mddev->gendisk)->kobj, KOBJ_CHANGE, envp);
 426         wait_for_completion_timeout(&cinfo->newdisk_completion,
 427                         NEW_DEV_TIMEOUT);
 428         clear_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state);
 429 }
 430
 431
 432 static void process_metadata_update(struct mddev *mddev, struct cluster_msg *msg)
 433 {
 434         struct md_cluster_info *cinfo = mddev->cluster_info;
 435         md_reload_sb(mddev, le32_to_cpu(msg->raid_slot));
 436         dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR);
 437 }
 438
 439 static void process_remove_disk(struct mddev *mddev, struct cluster_msg *msg)
 440 {
 441         struct md_rdev *rdev = md_find_rdev_nr_rcu(mddev,
 442                                                    le32_to_cpu(msg->raid_slot));
 443
 444         if (rdev)
 445                 md_kick_rdev_from_array(rdev);
 446         else
 447                 pr_warn("%s: %d Could not find disk(%d) to REMOVE\n",
 448                         __func__, __LINE__, le32_to_cpu(msg->raid_slot));
 449 }
 450
 451 static void process_readd_disk(struct mddev *mddev, struct cluster_msg *msg)
 452 {
 453         struct md_rdev *rdev = md_find_rdev_nr_rcu(mddev,
 454                                                    le32_to_cpu(msg->raid_slot));
 455
 456         if (rdev && test_bit(Faulty, &rdev->flags))
 457                 clear_bit(Faulty, &rdev->flags);
 458         else
 459                 pr_warn("%s: %d Could not find disk(%d) which is faulty",
 460                         __func__, __LINE__, le32_to_cpu(msg->raid_slot));
 461 }
 462
 463 static void process_recvd_msg(struct mddev *mddev, struct cluster_msg *msg)
 464 {
 465         switch (msg->type) {
 466         case METADATA_UPDATED:
 467                 process_metadata_update(mddev, msg);
 468                 break;
 469         case RESYNCING:
 470                 process_suspend_info(mddev, msg->slot,
 471                                 msg->low, msg->high);
 472                 break;
 473         case NEWDISK:
 474                 process_add_new_disk(mddev, msg);
 475                 break;
 476         case REMOVE:
 477                 process_remove_disk(mddev, msg);
 478                 break;
 479         case RE_ADD:
 480                 process_readd_disk(mddev, msg);
 481                 break;
 482         case BITMAP_NEEDS_SYNC:
 483                 __recover_slot(mddev, msg->slot);
 484                 break;
 485         default:
 486                 pr_warn("%s:%d Received unknown message from %d\n",
 487                         __func__, __LINE__, msg->slot);
 488         }
 489 }
 490
 491 /*
 492  * thread for receiving message
 493  */
 494 static void recv_daemon(struct md_thread *thread)
 495 {
 496         struct md_cluster_info *cinfo = thread->mddev->cluster_info;
 497         struct dlm_lock_resource *ack_lockres = cinfo->ack_lockres;
 498         struct dlm_lock_resource *message_lockres = cinfo->message_lockres;
 499         struct cluster_msg msg;
 500         int ret;
 501
 502         /*get CR on Message*/
 503         if (dlm_lock_sync(message_lockres, DLM_LOCK_CR)) {
 504                 pr_err("md/raid1:failed to get CR on MESSAGE\n");
 505                 return;
 506         }
 507
 508         /* read lvb and wake up thread to process this message_lockres */
 509         memcpy(&msg, message_lockres->lksb.sb_lvbptr, sizeof(struct cluster_msg));
 510         process_recvd_msg(thread->mddev, &msg);
 511
 512         /*release CR on ack_lockres*/
 513         ret = dlm_unlock_sync(ack_lockres);
 514         if (unlikely(ret != 0))
 515                 pr_info("unlock ack failed return %d\n", ret);
 516         /*up-convert to PR on message_lockres*/
 517         ret = dlm_lock_sync(message_lockres, DLM_LOCK_PR);
 518         if (unlikely(ret != 0))
 519                 pr_info("lock PR on msg failed return %d\n", ret);
 520         /*get CR on ack_lockres again*/
 521         ret = dlm_lock_sync(ack_lockres, DLM_LOCK_CR);
 522         if (unlikely(ret != 0))
 523                 pr_info("lock CR on ack failed return %d\n", ret);
 524         /*release CR on message_lockres*/
 525         ret = dlm_unlock_sync(message_lockres);
 526         if (unlikely(ret != 0))
 527                 pr_info("unlock msg failed return %d\n", ret);
 528 }
 529
 530 /* lock_comm()
 531  * Takes the lock on the TOKEN lock resource so no other
 532  * node can communicate while the operation is underway.
 533  * If called again, and the TOKEN lock is alread in EX mode
 534  * return success. However, care must be taken that unlock_comm()
 535  * is called only once.
 536  */
 537 static int lock_comm(struct md_cluster_info *cinfo)
 538 {
 539         int error;
 540
 541         if (cinfo->token_lockres->mode == DLM_LOCK_EX)
 542                 return 0;
 543
 544         error = dlm_lock_sync(cinfo->token_lockres, DLM_LOCK_EX);
 545         if (error)
 546                 pr_err("md-cluster(%s:%d): failed to get EX on TOKEN (%d)\n",
 547                                 __func__, __LINE__, error);
 548         return error;
 549 }
 550
 551 static void unlock_comm(struct md_cluster_info *cinfo)
 552 {
 553         WARN_ON(cinfo->token_lockres->mode != DLM_LOCK_EX);
 554         dlm_unlock_sync(cinfo->token_lockres);
 555 }
 556
 557 /* __sendmsg()
 558  * This function performs the actual sending of the message. This function is
 559  * usually called after performing the encompassing operation
 560  * The function:
 561  * 1. Grabs the message lockresource in EX mode
 562  * 2. Copies the message to the message LVB
 563  * 3. Downconverts message lockresource to CW
 564  * 4. Upconverts ack lock resource from CR to EX. This forces the BAST on other nodes
 565  *    and the other nodes read the message. The thread will wait here until all other
 566  *    nodes have released ack lock resource.
 567  * 5. Downconvert ack lockresource to CR
 568  */
 569 static int __sendmsg(struct md_cluster_info *cinfo, struct cluster_msg *cmsg)
 570 {
 571         int error;
 572         int slot = cinfo->slot_number - 1;
 573
 574         cmsg->slot = cpu_to_le32(slot);
 575         /*get EX on Message*/
 576         error = dlm_lock_sync(cinfo->message_lockres, DLM_LOCK_EX);
 577         if (error) {
 578                 pr_err("md-cluster: failed to get EX on MESSAGE (%d)\n", error);
 579                 goto failed_message;
 580         }
 581
 582         memcpy(cinfo->message_lockres->lksb.sb_lvbptr, (void *)cmsg,
 583                         sizeof(struct cluster_msg));
 584         /*down-convert EX to CW on Message*/
 585         error = dlm_lock_sync(cinfo->message_lockres, DLM_LOCK_CW);
 586         if (error) {
 587                 pr_err("md-cluster: failed to convert EX to CW on MESSAGE(%d)\n",
 588                                 error);
 589                 goto failed_ack;
 590         }
 591
 592         /*up-convert CR to EX on Ack*/
 593         error = dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_EX);
 594         if (error) {
 595                 pr_err("md-cluster: failed to convert CR to EX on ACK(%d)\n",
 596                                 error);
 597                 goto failed_ack;
 598         }
 599
 600         /*down-convert EX to CR on Ack*/
 601         error = dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_CR);
 602         if (error) {
 603                 pr_err("md-cluster: failed to convert EX to CR on ACK(%d)\n",
 604                                 error);
 605                 goto failed_ack;
 606         }
 607
 608 failed_ack:
 609         error = dlm_unlock_sync(cinfo->message_lockres);
 610         if (unlikely(error != 0)) {
 611                 pr_err("md-cluster: failed convert to NL on MESSAGE(%d)\n",
 612                         error);
 613                 /* in case the message can't be released due to some reason */
 614                 goto failed_ack;
 615         }
 616 failed_message:
 617         return error;
 618 }
 619
 620 static int sendmsg(struct md_cluster_info *cinfo, struct cluster_msg *cmsg)
 621 {
 622         int ret;
 623
 624         lock_comm(cinfo);
 625         ret = __sendmsg(cinfo, cmsg);
 626         unlock_comm(cinfo);
 627         return ret;
 628 }
 629
 630 static int gather_all_resync_info(struct mddev *mddev, int total_slots)
 631 {
 632         struct md_cluster_info *cinfo = mddev->cluster_info;
 633         int i, ret = 0;
 634         struct dlm_lock_resource *bm_lockres;
 635         struct suspend_info *s;
 636         char str[64];
 637         sector_t lo, hi;
 638
 639
 640         for (i = 0; i < total_slots; i++) {
 641                 memset(str, '\0', 64);
 642                 snprintf(str, 64, "bitmap%04d", i);
 643                 bm_lockres = lockres_init(mddev, str, NULL, 1);
 644                 if (!bm_lockres)
 645                         return -ENOMEM;
 646                 if (i == (cinfo->slot_number - 1))
 647                         continue;
 648
 649                 bm_lockres->flags |= DLM_LKF_NOQUEUE;
 650                 ret = dlm_lock_sync(bm_lockres, DLM_LOCK_PW);
 651                 if (ret == -EAGAIN) {
 652                         memset(bm_lockres->lksb.sb_lvbptr, '\0', LVB_SIZE);
 653                         s = read_resync_info(mddev, bm_lockres);
 654                         if (s) {
 655                                 pr_info("%s:%d Resync[%llu..%llu] in progress on %d\n",
 656                                                 __func__, __LINE__,
 657                                                 (unsigned long long) s->lo,
 658                                                 (unsigned long long) s->hi, i);
 659                                 spin_lock_irq(&cinfo->suspend_lock);
 660                                 s->slot = i;
 661                                 list_add(&s->list, &cinfo->suspend_list);
 662                                 spin_unlock_irq(&cinfo->suspend_lock);
 663                         }
 664                         ret = 0;
 665                         lockres_free(bm_lockres);
 666                         continue;
 667                 }
 668                 if (ret) {
 669                         lockres_free(bm_lockres);
 670                         goto out;
 671                 }
 672
 673                 /* Read the disk bitmap sb and check if it needs recovery */
 674                 ret = bitmap_copy_from_slot(mddev, i, &lo, &hi, false);
 675                 if (ret) {
 676                         pr_warn("md-cluster: Could not gather bitmaps from slot %d", i);
 677                         lockres_free(bm_lockres);
 678                         continue;
 679                 }
 680                 if ((hi > 0) && (lo < mddev->recovery_cp)) {
 681                         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
 682                         mddev->recovery_cp = lo;
 683                         md_check_recovery(mddev);
 684                 }
 685
 686                 dlm_unlock_sync(bm_lockres);
 687                 lockres_free(bm_lockres);
 688         }
 689 out:
 690         return ret;
 691 }
 692
 693 static int join(struct mddev *mddev, int nodes)
 694 {
 695         struct md_cluster_info *cinfo;
 696         int ret, ops_rv;
 697         char str[64];
 698
 699         cinfo = kzalloc(sizeof(struct md_cluster_info), GFP_KERNEL);
 700         if (!cinfo)
 701                 return -ENOMEM;
 702
 703         INIT_LIST_HEAD(&cinfo->suspend_list);
 704         spin_lock_init(&cinfo->suspend_lock);
 705         init_completion(&cinfo->completion);
 706         set_bit(MD_CLUSTER_BEGIN_JOIN_CLUSTER, &cinfo->state);
 707
 708         mutex_init(&cinfo->sb_mutex);
 709         mddev->cluster_info = cinfo;
 710
 711         memset(str, 0, 64);
 712         sprintf(str, "%pU", mddev->uuid);
 713         ret = dlm_new_lockspace(str, mddev->bitmap_info.cluster_name,
 714                                 DLM_LSFL_FS, LVB_SIZE,
 715                                 &md_ls_ops, mddev, &ops_rv, &cinfo->lockspace);
 716         if (ret)
 717                 goto err;
 718         wait_for_completion(&cinfo->completion);
 719         if (nodes < cinfo->slot_number) {
 720                 pr_err("md-cluster: Slot allotted(%d) is greater than available slots(%d).",
 721                         cinfo->slot_number, nodes);
 722                 ret = -ERANGE;
 723                 goto err;
 724         }
 725         /* Initiate the communication resources */
 726         ret = -ENOMEM;
 727         cinfo->recv_thread = md_register_thread(recv_daemon, mddev, "cluster_recv");
 728         if (!cinfo->recv_thread) {
 729                 pr_err("md-cluster: cannot allocate memory for recv_thread!\n");
 730                 goto err;
 731         }
 732         cinfo->message_lockres = lockres_init(mddev, "message", NULL, 1);
 733         if (!cinfo->message_lockres)
 734                 goto err;
 735         cinfo->token_lockres = lockres_init(mddev, "token", NULL, 0);
 736         if (!cinfo->token_lockres)
 737                 goto err;
 738         cinfo->ack_lockres = lockres_init(mddev, "ack", ack_bast, 0);
 739         if (!cinfo->ack_lockres)
 740                 goto err;
 741         cinfo->no_new_dev_lockres = lockres_init(mddev, "no-new-dev", NULL, 0);
 742         if (!cinfo->no_new_dev_lockres)
 743                 goto err;
 744
 745         /* get sync CR lock on ACK. */
 746         if (dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_CR))
 747                 pr_err("md-cluster: failed to get a sync CR lock on ACK!(%d)\n",
 748                                 ret);
 749         /* get sync CR lock on no-new-dev. */
 750         if (dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR))
 751                 pr_err("md-cluster: failed to get a sync CR lock on no-new-dev!(%d)\n", ret);
 752
 753
 754         pr_info("md-cluster: Joined cluster %s slot %d\n", str, cinfo->slot_number);
 755         snprintf(str, 64, "bitmap%04d", cinfo->slot_number - 1);
 756         cinfo->bitmap_lockres = lockres_init(mddev, str, NULL, 1);
 757         if (!cinfo->bitmap_lockres)
 758                 goto err;
 759         if (dlm_lock_sync(cinfo->bitmap_lockres, DLM_LOCK_PW)) {
 760                 pr_err("Failed to get bitmap lock\n");
 761                 ret = -EINVAL;
 762                 goto err;
 763         }
 764
 765         cinfo->resync_lockres = lockres_init(mddev, "resync", NULL, 0);
 766         if (!cinfo->resync_lockres)
 767                 goto err;
 768
 769         ret = gather_all_resync_info(mddev, nodes);
 770         if (ret)
 771                 goto err;
 772
 773         return 0;
 774 err:
 775         lockres_free(cinfo->message_lockres);
 776         lockres_free(cinfo->token_lockres);
 777         lockres_free(cinfo->ack_lockres);
 778         lockres_free(cinfo->no_new_dev_lockres);
 779         lockres_free(cinfo->resync_lockres);
 780         lockres_free(cinfo->bitmap_lockres);
 781         if (cinfo->lockspace)
 782                 dlm_release_lockspace(cinfo->lockspace, 2);
 783         mddev->cluster_info = NULL;
 784         kfree(cinfo);
 785         return ret;
 786 }
 787
 788 static void resync_bitmap(struct mddev *mddev)
 789 {
 790         struct md_cluster_info *cinfo = mddev->cluster_info;
 791         struct cluster_msg cmsg = {0};
 792         int err;
 793
 794         cmsg.type = cpu_to_le32(BITMAP_NEEDS_SYNC);
 795         err = sendmsg(cinfo, &cmsg);
 796         if (err)
 797                 pr_err("%s:%d: failed to send BITMAP_NEEDS_SYNC message (%d)\n",
 798                         __func__, __LINE__, err);
 799 }
 800
 801 static int leave(struct mddev *mddev)
 802 {
 803         struct md_cluster_info *cinfo = mddev->cluster_info;
 804
 805         if (!cinfo)
 806                 return 0;
 807
 808         /* BITMAP_NEEDS_SYNC message should be sent when node
 809          * is leaving the cluster with dirty bitmap, also we
 810          * can only deliver it when dlm connection is available */
 811         if (cinfo->slot_number > 0 && mddev->recovery_cp != MaxSector)
 812                 resync_bitmap(mddev);
 813
 814         md_unregister_thread(&cinfo->recovery_thread);
 815         md_unregister_thread(&cinfo->recv_thread);
 816         lockres_free(cinfo->message_lockres);
 817         lockres_free(cinfo->token_lockres);
 818         lockres_free(cinfo->ack_lockres);
 819         lockres_free(cinfo->no_new_dev_lockres);
 820         lockres_free(cinfo->bitmap_lockres);
 821         dlm_release_lockspace(cinfo->lockspace, 2);
 822         return 0;
 823 }
 824
 825 /* slot_number(): Returns the MD slot number to use
 826  * DLM starts the slot numbers from 1, wheras cluster-md
 827  * wants the number to be from zero, so we deduct one
 828  */
 829 static int slot_number(struct mddev *mddev)
 830 {
 831         struct md_cluster_info *cinfo = mddev->cluster_info;
 832
 833         return cinfo->slot_number - 1;
 834 }
 835
 836 static int metadata_update_start(struct mddev *mddev)
 837 {
 838         return lock_comm(mddev->cluster_info);
 839 }
 840
 841 static int metadata_update_finish(struct mddev *mddev)
 842 {
 843         struct md_cluster_info *cinfo = mddev->cluster_info;
 844         struct cluster_msg cmsg;
 845         struct md_rdev *rdev;
 846         int ret = 0;
 847
 848         memset(&cmsg, 0, sizeof(cmsg));
 849         cmsg.type = cpu_to_le32(METADATA_UPDATED);
 850         cmsg.raid_slot = -1;
 851         /* Pick up a good active device number to send.
 852          */
 853         rdev_for_each(rdev, mddev)
 854                 if (rdev->raid_disk > -1 && !test_bit(Faulty, &rdev->flags)) {
 855                         cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
 856                         break;
 857                 }
 858         if (cmsg.raid_slot >= 0)
 859                 ret = __sendmsg(cinfo, &cmsg);
 860         else
 861                 pr_warn("md-cluster: No good device id found to send\n");
 862         unlock_comm(cinfo);
 863         return ret;
 864 }
 865
 866 static void metadata_update_cancel(struct mddev *mddev)
 867 {
 868         struct md_cluster_info *cinfo = mddev->cluster_info;
 869         unlock_comm(cinfo);
 870 }
 871
 872 static int resync_start(struct mddev *mddev)
 873 {
 874         struct md_cluster_info *cinfo = mddev->cluster_info;
 875         cinfo->resync_lockres->flags |= DLM_LKF_NOQUEUE;
 876         return dlm_lock_sync(cinfo->resync_lockres, DLM_LOCK_EX);
 877 }
 878
 879 static int resync_info_update(struct mddev *mddev, sector_t lo, sector_t hi)
 880 {
 881         struct md_cluster_info *cinfo = mddev->cluster_info;
 882         struct cluster_msg cmsg;
 883
 884         add_resync_info(mddev, cinfo->bitmap_lockres, lo, hi);
 885         /* Re-acquire the lock to refresh LVB */
 886         dlm_lock_sync(cinfo->bitmap_lockres, DLM_LOCK_PW);
 887         cmsg.type = cpu_to_le32(RESYNCING);
 888         cmsg.low = cpu_to_le64(lo);
 889         cmsg.high = cpu_to_le64(hi);
 890
 891         return sendmsg(cinfo, &cmsg);
 892 }
 893
 894 static int resync_finish(struct mddev *mddev)
 895 {
 896         struct md_cluster_info *cinfo = mddev->cluster_info;
 897         cinfo->resync_lockres->flags &= ~DLM_LKF_NOQUEUE;
 898         dlm_unlock_sync(cinfo->resync_lockres);
 899         return resync_info_update(mddev, 0, 0);
 900 }
 901
 902 static int area_resyncing(struct mddev *mddev, int direction,
 903                 sector_t lo, sector_t hi)
 904 {
 905         struct md_cluster_info *cinfo = mddev->cluster_info;
 906         int ret = 0;
 907         struct suspend_info *s;
 908
 909         if ((direction == READ) &&
 910                 test_bit(MD_CLUSTER_SUSPEND_READ_BALANCING, &cinfo->state))
 911                 return 1;
 912
 913         spin_lock_irq(&cinfo->suspend_lock);
 914         if (list_empty(&cinfo->suspend_list))
 915                 goto out;
 916         list_for_each_entry(s, &cinfo->suspend_list, list)
 917                 if (hi > s->lo && lo < s->hi) {
 918                         ret = 1;
 919                         break;
 920                 }
 921 out:
 922         spin_unlock_irq(&cinfo->suspend_lock);
 923         return ret;
 924 }
 925
 926 /* add_new_disk() - initiates a disk add
 927  * However, if this fails before writing md_update_sb(),
 928  * add_new_disk_cancel() must be called to release token lock
 929  */
 930 static int add_new_disk(struct mddev *mddev, struct md_rdev *rdev)
 931 {
 932         struct md_cluster_info *cinfo = mddev->cluster_info;
 933         struct cluster_msg cmsg;
 934         int ret = 0;
 935         struct mdp_superblock_1 *sb = page_address(rdev->sb_page);
 936         char *uuid = sb->device_uuid;
 937
 938         memset(&cmsg, 0, sizeof(cmsg));
 939         cmsg.type = cpu_to_le32(NEWDISK);
 940         memcpy(cmsg.uuid, uuid, 16);
 941         cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
 942         lock_comm(cinfo);
 943         ret = __sendmsg(cinfo, &cmsg);
 944         if (ret)
 945                 return ret;
 946         cinfo->no_new_dev_lockres->flags |= DLM_LKF_NOQUEUE;
 947         ret = dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_EX);
 948         cinfo->no_new_dev_lockres->flags &= ~DLM_LKF_NOQUEUE;
 949         /* Some node does not "see" the device */
 950         if (ret == -EAGAIN)
 951                 ret = -ENOENT;
 952         if (ret)
 953                 unlock_comm(cinfo);
 954         else
 955                 dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR);
 956         return ret;
 957 }
 958
 959 static void add_new_disk_cancel(struct mddev *mddev)
 960 {
 961         struct md_cluster_info *cinfo = mddev->cluster_info;
 962         unlock_comm(cinfo);
 963 }
 964
 965 static int new_disk_ack(struct mddev *mddev, bool ack)
 966 {
 967         struct md_cluster_info *cinfo = mddev->cluster_info;
 968
 969         if (!test_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state)) {
 970                 pr_warn("md-cluster(%s): Spurious cluster confirmation\n", mdname(mddev));
 971                 return -EINVAL;
 972         }
 973
 974         if (ack)
 975                 dlm_unlock_sync(cinfo->no_new_dev_lockres);
 976         complete(&cinfo->newdisk_completion);
 977         return 0;
 978 }
 979
 980 static int remove_disk(struct mddev *mddev, struct md_rdev *rdev)
 981 {
 982         struct cluster_msg cmsg;
 983         struct md_cluster_info *cinfo = mddev->cluster_info;
 984         cmsg.type = cpu_to_le32(REMOVE);
 985         cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
 986         return __sendmsg(cinfo, &cmsg);
 987 }
 988
 989 static int gather_bitmaps(struct md_rdev *rdev)
 990 {
 991         int sn, err;
 992         sector_t lo, hi;
 993         struct cluster_msg cmsg;
 994         struct mddev *mddev = rdev->mddev;
 995         struct md_cluster_info *cinfo = mddev->cluster_info;
 996
 997         cmsg.type = cpu_to_le32(RE_ADD);
 998         cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
 999         err = sendmsg(cinfo, &cmsg);
1000         if (err)
1001                 goto out;
1002
1003         for (sn = 0; sn < mddev->bitmap_info.nodes; sn++) {
1004                 if (sn == (cinfo->slot_number - 1))
1005                         continue;
1006                 err = bitmap_copy_from_slot(mddev, sn, &lo, &hi, false);
1007                 if (err) {
1008                         pr_warn("md-cluster: Could not gather bitmaps from slot %d", sn);
1009                         goto out;
1010                 }
1011                 if ((hi > 0) && (lo < mddev->recovery_cp))
1012                         mddev->recovery_cp = lo;
1013         }
1014 out:
1015         return err;
1016 }
1017
1018 static struct md_cluster_operations cluster_ops = {
1019         .join   = join,
1020         .leave  = leave,
1021         .slot_number = slot_number,
1022         .resync_start = resync_start,
1023         .resync_finish = resync_finish,
1024         .resync_info_update = resync_info_update,
1025         .metadata_update_start = metadata_update_start,
1026         .metadata_update_finish = metadata_update_finish,
1027         .metadata_update_cancel = metadata_update_cancel,
1028         .area_resyncing = area_resyncing,
1029         .add_new_disk = add_new_disk,
1030         .add_new_disk_cancel = add_new_disk_cancel,
1031         .new_disk_ack = new_disk_ack,
1032         .remove_disk = remove_disk,
1033         .gather_bitmaps = gather_bitmaps,
1034 };
1035
1036 static int __init cluster_init(void)
1037 {
1038         pr_warn("md-cluster: EXPERIMENTAL. Use with caution\n");
1039         pr_info("Registering Cluster MD functions\n");
1040         register_md_cluster_operations(&cluster_ops, THIS_MODULE);
1041         return 0;
1042 }
1043
1044 static void cluster_exit(void)
1045 {
1046         unregister_md_cluster_operations();
1047 }
1048
1049 module_init(cluster_init);
1050 module_exit(cluster_exit);
1051 MODULE_LICENSE("GPL");
1052 MODULE_DESCRIPTION("Clustering support for MD");