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 /* Lock the send communication. This is done through
  52  * bit manipulation as opposed to a mutex in order to
  53  * accomodate lock and hold. See next comment.
  54  */
  55 #define         MD_CLUSTER_SEND_LOCK                    4
  56 /* If cluster operations must lock the communication channel,
  57  * so as to perform extra operations (and no other operation
  58  * is allowed on the MD, such as adding a disk. Token needs
  59  * to be locked and held until the operation completes with
  60  * a md_update_sb(), which would eventually release the lock.
  61  */
  62 #define         MD_CLUSTER_SEND_LOCKED_ALREADY          5
  63
  64
  65 struct md_cluster_info {
  66         /* dlm lock space and resources for clustered raid. */
  67         dlm_lockspace_t *lockspace;
  68         int slot_number;
  69         struct completion completion;
  70         struct mutex recv_mutex;
  71         struct dlm_lock_resource *bitmap_lockres;
  72         struct dlm_lock_resource **other_bitmap_lockres;
  73         struct dlm_lock_resource *resync_lockres;
  74         struct list_head suspend_list;
  75         spinlock_t suspend_lock;
  76         struct md_thread *recovery_thread;
  77         unsigned long recovery_map;
  78         /* communication loc resources */
  79         struct dlm_lock_resource *ack_lockres;
  80         struct dlm_lock_resource *message_lockres;
  81         struct dlm_lock_resource *token_lockres;
  82         struct dlm_lock_resource *no_new_dev_lockres;
  83         struct md_thread *recv_thread;
  84         struct completion newdisk_completion;
  85         wait_queue_head_t wait;
  86         unsigned long state;
  87 };
  88
  89 enum msg_type {
  90         METADATA_UPDATED = 0,
  91         RESYNCING,
  92         NEWDISK,
  93         REMOVE,
  94         RE_ADD,
  95         BITMAP_NEEDS_SYNC,
  96 };
  97
  98 struct cluster_msg {
  99         __le32 type;
 100         __le32 slot;
 101         /* TODO: Unionize this for smaller footprint */
 102         __le64 low;
 103         __le64 high;
 104         char uuid[16];
 105         __le32 raid_slot;
 106 };
 107
 108 static void sync_ast(void *arg)
 109 {
 110         struct dlm_lock_resource *res;
 111
 112         res = arg;
 113         complete(&res->completion);
 114 }
 115
 116 static int dlm_lock_sync(struct dlm_lock_resource *res, int mode)
 117 {
 118         int ret = 0;
 119
 120         ret = dlm_lock(res->ls, mode, &res->lksb,
 121                         res->flags, res->name, strlen(res->name),
 122                         0, sync_ast, res, res->bast);
 123         if (ret)
 124                 return ret;
 125         wait_for_completion(&res->completion);
 126         if (res->lksb.sb_status == 0)
 127                 res->mode = mode;
 128         return res->lksb.sb_status;
 129 }
 130
 131 static int dlm_unlock_sync(struct dlm_lock_resource *res)
 132 {
 133         return dlm_lock_sync(res, DLM_LOCK_NL);
 134 }
 135
 136 static struct dlm_lock_resource *lockres_init(struct mddev *mddev,
 137                 char *name, void (*bastfn)(void *arg, int mode), int with_lvb)
 138 {
 139         struct dlm_lock_resource *res = NULL;
 140         int ret, namelen;
 141         struct md_cluster_info *cinfo = mddev->cluster_info;
 142
 143         res = kzalloc(sizeof(struct dlm_lock_resource), GFP_KERNEL);
 144         if (!res)
 145                 return NULL;
 146         init_completion(&res->completion);
 147         res->ls = cinfo->lockspace;
 148         res->mddev = mddev;
 149         res->mode = DLM_LOCK_IV;
 150         namelen = strlen(name);
 151         res->name = kzalloc(namelen + 1, GFP_KERNEL);
 152         if (!res->name) {
 153                 pr_err("md-cluster: Unable to allocate resource name for resource %s\n", name);
 154                 goto out_err;
 155         }
 156         strlcpy(res->name, name, namelen + 1);
 157         if (with_lvb) {
 158                 res->lksb.sb_lvbptr = kzalloc(LVB_SIZE, GFP_KERNEL);
 159                 if (!res->lksb.sb_lvbptr) {
 160                         pr_err("md-cluster: Unable to allocate LVB for resource %s\n", name);
 161                         goto out_err;
 162                 }
 163                 res->flags = DLM_LKF_VALBLK;
 164         }
 165
 166         if (bastfn)
 167                 res->bast = bastfn;
 168
 169         res->flags |= DLM_LKF_EXPEDITE;
 170
 171         ret = dlm_lock_sync(res, DLM_LOCK_NL);
 172         if (ret) {
 173                 pr_err("md-cluster: Unable to lock NL on new lock resource %s\n", name);
 174                 goto out_err;
 175         }
 176         res->flags &= ~DLM_LKF_EXPEDITE;
 177         res->flags |= DLM_LKF_CONVERT;
 178
 179         return res;
 180 out_err:
 181         kfree(res->lksb.sb_lvbptr);
 182         kfree(res->name);
 183         kfree(res);
 184         return NULL;
 185 }
 186
 187 static void lockres_free(struct dlm_lock_resource *res)
 188 {
 189         int ret;
 190
 191         if (!res)
 192                 return;
 193
 194         /* cancel a lock request or a conversion request that is blocked */
 195         res->flags |= DLM_LKF_CANCEL;
 196 retry:
 197         ret = dlm_unlock(res->ls, res->lksb.sb_lkid, 0, &res->lksb, res);
 198         if (unlikely(ret != 0)) {
 199                 pr_info("%s: failed to unlock %s return %d\n", __func__, res->name, ret);
 200
 201                 /* if a lock conversion is cancelled, then the lock is put
 202                  * back to grant queue, need to ensure it is unlocked */
 203                 if (ret == -DLM_ECANCEL)
 204                         goto retry;
 205         }
 206         res->flags &= ~DLM_LKF_CANCEL;
 207         wait_for_completion(&res->completion);
 208
 209         kfree(res->name);
 210         kfree(res->lksb.sb_lvbptr);
 211         kfree(res);
 212 }
 213
 214 static void add_resync_info(struct dlm_lock_resource *lockres,
 215                             sector_t lo, sector_t hi)
 216 {
 217         struct resync_info *ri;
 218
 219         ri = (struct resync_info *)lockres->lksb.sb_lvbptr;
 220         ri->lo = cpu_to_le64(lo);
 221         ri->hi = cpu_to_le64(hi);
 222 }
 223
 224 static struct suspend_info *read_resync_info(struct mddev *mddev, struct dlm_lock_resource *lockres)
 225 {
 226         struct resync_info ri;
 227         struct suspend_info *s = NULL;
 228         sector_t hi = 0;
 229
 230         dlm_lock_sync(lockres, DLM_LOCK_CR);
 231         memcpy(&ri, lockres->lksb.sb_lvbptr, sizeof(struct resync_info));
 232         hi = le64_to_cpu(ri.hi);
 233         if (hi > 0) {
 234                 s = kzalloc(sizeof(struct suspend_info), GFP_KERNEL);
 235                 if (!s)
 236                         goto out;
 237                 s->hi = hi;
 238                 s->lo = le64_to_cpu(ri.lo);
 239         }
 240         dlm_unlock_sync(lockres);
 241 out:
 242         return s;
 243 }
 244
 245 static void recover_bitmaps(struct md_thread *thread)
 246 {
 247         struct mddev *mddev = thread->mddev;
 248         struct md_cluster_info *cinfo = mddev->cluster_info;
 249         struct dlm_lock_resource *bm_lockres;
 250         char str[64];
 251         int slot, ret;
 252         struct suspend_info *s, *tmp;
 253         sector_t lo, hi;
 254
 255         while (cinfo->recovery_map) {
 256                 slot = fls64((u64)cinfo->recovery_map) - 1;
 257
 258                 /* Clear suspend_area associated with the bitmap */
 259                 spin_lock_irq(&cinfo->suspend_lock);
 260                 list_for_each_entry_safe(s, tmp, &cinfo->suspend_list, list)
 261                         if (slot == s->slot) {
 262                                 list_del(&s->list);
 263                                 kfree(s);
 264                         }
 265                 spin_unlock_irq(&cinfo->suspend_lock);
 266
 267                 snprintf(str, 64, "bitmap%04d", slot);
 268                 bm_lockres = lockres_init(mddev, str, NULL, 1);
 269                 if (!bm_lockres) {
 270                         pr_err("md-cluster: Cannot initialize bitmaps\n");
 271                         goto clear_bit;
 272                 }
 273
 274                 ret = dlm_lock_sync(bm_lockres, DLM_LOCK_PW);
 275                 if (ret) {
 276                         pr_err("md-cluster: Could not DLM lock %s: %d\n",
 277                                         str, ret);
 278                         goto clear_bit;
 279                 }
 280                 ret = bitmap_copy_from_slot(mddev, slot, &lo, &hi, true);
 281                 if (ret) {
 282                         pr_err("md-cluster: Could not copy data from bitmap %d\n", slot);
 283                         goto dlm_unlock;
 284                 }
 285                 if (hi > 0) {
 286                         /* TODO:Wait for current resync to get over */
 287                         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
 288                         if (lo < mddev->recovery_cp)
 289                                 mddev->recovery_cp = lo;
 290                         md_check_recovery(mddev);
 291                 }
 292 dlm_unlock:
 293                 dlm_unlock_sync(bm_lockres);
 294 clear_bit:
 295                 clear_bit(slot, &cinfo->recovery_map);
 296         }
 297 }
 298
 299 static void recover_prep(void *arg)
 300 {
 301         struct mddev *mddev = arg;
 302         struct md_cluster_info *cinfo = mddev->cluster_info;
 303         set_bit(MD_CLUSTER_SUSPEND_READ_BALANCING, &cinfo->state);
 304 }
 305
 306 static void __recover_slot(struct mddev *mddev, int slot)
 307 {
 308         struct md_cluster_info *cinfo = mddev->cluster_info;
 309
 310         set_bit(slot, &cinfo->recovery_map);
 311         if (!cinfo->recovery_thread) {
 312                 cinfo->recovery_thread = md_register_thread(recover_bitmaps,
 313                                 mddev, "recover");
 314                 if (!cinfo->recovery_thread) {
 315                         pr_warn("md-cluster: Could not create recovery thread\n");
 316                         return;
 317                 }
 318         }
 319         md_wakeup_thread(cinfo->recovery_thread);
 320 }
 321
 322 static void recover_slot(void *arg, struct dlm_slot *slot)
 323 {
 324         struct mddev *mddev = arg;
 325         struct md_cluster_info *cinfo = mddev->cluster_info;
 326
 327         pr_info("md-cluster: %s Node %d/%d down. My slot: %d. Initiating recovery.\n",
 328                         mddev->bitmap_info.cluster_name,
 329                         slot->nodeid, slot->slot,
 330                         cinfo->slot_number);
 331         /* deduct one since dlm slot starts from one while the num of
 332          * cluster-md begins with 0 */
 333         __recover_slot(mddev, slot->slot - 1);
 334 }
 335
 336 static void recover_done(void *arg, struct dlm_slot *slots,
 337                 int num_slots, int our_slot,
 338                 uint32_t generation)
 339 {
 340         struct mddev *mddev = arg;
 341         struct md_cluster_info *cinfo = mddev->cluster_info;
 342
 343         cinfo->slot_number = our_slot;
 344         /* completion is only need to be complete when node join cluster,
 345          * it doesn't need to run during another node's failure */
 346         if (test_bit(MD_CLUSTER_BEGIN_JOIN_CLUSTER, &cinfo->state)) {
 347                 complete(&cinfo->completion);
 348                 clear_bit(MD_CLUSTER_BEGIN_JOIN_CLUSTER, &cinfo->state);
 349         }
 350         clear_bit(MD_CLUSTER_SUSPEND_READ_BALANCING, &cinfo->state);
 351 }
 352
 353 /* the ops is called when node join the cluster, and do lock recovery
 354  * if node failure occurs */
 355 static const struct dlm_lockspace_ops md_ls_ops = {
 356         .recover_prep = recover_prep,
 357         .recover_slot = recover_slot,
 358         .recover_done = recover_done,
 359 };
 360
 361 /*
 362  * The BAST function for the ack lock resource
 363  * This function wakes up the receive thread in
 364  * order to receive and process the message.
 365  */
 366 static void ack_bast(void *arg, int mode)
 367 {
 368         struct dlm_lock_resource *res = arg;
 369         struct md_cluster_info *cinfo = res->mddev->cluster_info;
 370
 371         if (mode == DLM_LOCK_EX)
 372                 md_wakeup_thread(cinfo->recv_thread);
 373 }
 374
 375 static void __remove_suspend_info(struct md_cluster_info *cinfo, int slot)
 376 {
 377         struct suspend_info *s, *tmp;
 378
 379         list_for_each_entry_safe(s, tmp, &cinfo->suspend_list, list)
 380                 if (slot == s->slot) {
 381                         list_del(&s->list);
 382                         kfree(s);
 383                         break;
 384                 }
 385 }
 386
 387 static void remove_suspend_info(struct mddev *mddev, int slot)
 388 {
 389         struct md_cluster_info *cinfo = mddev->cluster_info;
 390         spin_lock_irq(&cinfo->suspend_lock);
 391         __remove_suspend_info(cinfo, slot);
 392         spin_unlock_irq(&cinfo->suspend_lock);
 393         mddev->pers->quiesce(mddev, 2);
 394 }
 395
 396
 397 static void process_suspend_info(struct mddev *mddev,
 398                 int slot, sector_t lo, sector_t hi)
 399 {
 400         struct md_cluster_info *cinfo = mddev->cluster_info;
 401         struct suspend_info *s;
 402
 403         if (!hi) {
 404                 remove_suspend_info(mddev, slot);
 405                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
 406                 md_wakeup_thread(mddev->thread);
 407                 return;
 408         }
 409         s = kzalloc(sizeof(struct suspend_info), GFP_KERNEL);
 410         if (!s)
 411                 return;
 412         s->slot = slot;
 413         s->lo = lo;
 414         s->hi = hi;
 415         mddev->pers->quiesce(mddev, 1);
 416         mddev->pers->quiesce(mddev, 0);
 417         spin_lock_irq(&cinfo->suspend_lock);
 418         /* Remove existing entry (if exists) before adding */
 419         __remove_suspend_info(cinfo, slot);
 420         list_add(&s->list, &cinfo->suspend_list);
 421         spin_unlock_irq(&cinfo->suspend_lock);
 422         mddev->pers->quiesce(mddev, 2);
 423 }
 424
 425 static void process_add_new_disk(struct mddev *mddev, struct cluster_msg *cmsg)
 426 {
 427         char disk_uuid[64];
 428         struct md_cluster_info *cinfo = mddev->cluster_info;
 429         char event_name[] = "EVENT=ADD_DEVICE";
 430         char raid_slot[16];
 431         char *envp[] = {event_name, disk_uuid, raid_slot, NULL};
 432         int len;
 433
 434         len = snprintf(disk_uuid, 64, "DEVICE_UUID=");
 435         sprintf(disk_uuid + len, "%pU", cmsg->uuid);
 436         snprintf(raid_slot, 16, "RAID_DISK=%d", le32_to_cpu(cmsg->raid_slot));
 437         pr_info("%s:%d Sending kobject change with %s and %s\n", __func__, __LINE__, disk_uuid, raid_slot);
 438         init_completion(&cinfo->newdisk_completion);
 439         set_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state);
 440         kobject_uevent_env(&disk_to_dev(mddev->gendisk)->kobj, KOBJ_CHANGE, envp);
 441         wait_for_completion_timeout(&cinfo->newdisk_completion,
 442                         NEW_DEV_TIMEOUT);
 443         clear_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state);
 444 }
 445
 446
 447 static void process_metadata_update(struct mddev *mddev, struct cluster_msg *msg)
 448 {
 449         struct md_cluster_info *cinfo = mddev->cluster_info;
 450         mddev->good_device_nr = le32_to_cpu(msg->raid_slot);
 451         set_bit(MD_RELOAD_SB, &mddev->flags);
 452         dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR);
 453         md_wakeup_thread(mddev->thread);
 454 }
 455
 456 static void process_remove_disk(struct mddev *mddev, struct cluster_msg *msg)
 457 {
 458         struct md_rdev *rdev = md_find_rdev_nr_rcu(mddev,
 459                                                    le32_to_cpu(msg->raid_slot));
 460
 461         if (rdev) {
 462                 set_bit(ClusterRemove, &rdev->flags);
 463                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
 464                 md_wakeup_thread(mddev->thread);
 465         }
 466         else
 467                 pr_warn("%s: %d Could not find disk(%d) to REMOVE\n",
 468                         __func__, __LINE__, le32_to_cpu(msg->raid_slot));
 469 }
 470
 471 static void process_readd_disk(struct mddev *mddev, struct cluster_msg *msg)
 472 {
 473         struct md_rdev *rdev = md_find_rdev_nr_rcu(mddev,
 474                                                    le32_to_cpu(msg->raid_slot));
 475
 476         if (rdev && test_bit(Faulty, &rdev->flags))
 477                 clear_bit(Faulty, &rdev->flags);
 478         else
 479                 pr_warn("%s: %d Could not find disk(%d) which is faulty",
 480                         __func__, __LINE__, le32_to_cpu(msg->raid_slot));
 481 }
 482
 483 static void process_recvd_msg(struct mddev *mddev, struct cluster_msg *msg)
 484 {
 485         if (WARN(mddev->cluster_info->slot_number - 1 == le32_to_cpu(msg->slot),
 486                 "node %d received it's own msg\n", le32_to_cpu(msg->slot)))
 487                 return;
 488         switch (le32_to_cpu(msg->type)) {
 489         case METADATA_UPDATED:
 490                 process_metadata_update(mddev, msg);
 491                 break;
 492         case RESYNCING:
 493                 process_suspend_info(mddev, le32_to_cpu(msg->slot),
 494                                      le64_to_cpu(msg->low),
 495                                      le64_to_cpu(msg->high));
 496                 break;
 497         case NEWDISK:
 498                 process_add_new_disk(mddev, msg);
 499                 break;
 500         case REMOVE:
 501                 process_remove_disk(mddev, msg);
 502                 break;
 503         case RE_ADD:
 504                 process_readd_disk(mddev, msg);
 505                 break;
 506         case BITMAP_NEEDS_SYNC:
 507                 __recover_slot(mddev, le32_to_cpu(msg->slot));
 508                 break;
 509         default:
 510                 pr_warn("%s:%d Received unknown message from %d\n",
 511                         __func__, __LINE__, msg->slot);
 512         }
 513 }
 514
 515 /*
 516  * thread for receiving message
 517  */
 518 static void recv_daemon(struct md_thread *thread)
 519 {
 520         struct md_cluster_info *cinfo = thread->mddev->cluster_info;
 521         struct dlm_lock_resource *ack_lockres = cinfo->ack_lockres;
 522         struct dlm_lock_resource *message_lockres = cinfo->message_lockres;
 523         struct cluster_msg msg;
 524         int ret;
 525
 526         mutex_lock(&cinfo->recv_mutex);
 527         /*get CR on Message*/
 528         if (dlm_lock_sync(message_lockres, DLM_LOCK_CR)) {
 529                 pr_err("md/raid1:failed to get CR on MESSAGE\n");
 530                 mutex_unlock(&cinfo->recv_mutex);
 531                 return;
 532         }
 533
 534         /* read lvb and wake up thread to process this message_lockres */
 535         memcpy(&msg, message_lockres->lksb.sb_lvbptr, sizeof(struct cluster_msg));
 536         process_recvd_msg(thread->mddev, &msg);
 537
 538         /*release CR on ack_lockres*/
 539         ret = dlm_unlock_sync(ack_lockres);
 540         if (unlikely(ret != 0))
 541                 pr_info("unlock ack failed return %d\n", ret);
 542         /*up-convert to PR on message_lockres*/
 543         ret = dlm_lock_sync(message_lockres, DLM_LOCK_PR);
 544         if (unlikely(ret != 0))
 545                 pr_info("lock PR on msg failed return %d\n", ret);
 546         /*get CR on ack_lockres again*/
 547         ret = dlm_lock_sync(ack_lockres, DLM_LOCK_CR);
 548         if (unlikely(ret != 0))
 549                 pr_info("lock CR on ack failed return %d\n", ret);
 550         /*release CR on message_lockres*/
 551         ret = dlm_unlock_sync(message_lockres);
 552         if (unlikely(ret != 0))
 553                 pr_info("unlock msg failed return %d\n", ret);
 554         mutex_unlock(&cinfo->recv_mutex);
 555 }
 556
 557 /* lock_token()
 558  * Takes the lock on the TOKEN lock resource so no other
 559  * node can communicate while the operation is underway.
 560  */
 561 static int lock_token(struct md_cluster_info *cinfo)
 562 {
 563         int error;
 564
 565         error = dlm_lock_sync(cinfo->token_lockres, DLM_LOCK_EX);
 566         if (error)
 567                 pr_err("md-cluster(%s:%d): failed to get EX on TOKEN (%d)\n",
 568                                 __func__, __LINE__, error);
 569
 570         /* Lock the receive sequence */
 571         mutex_lock(&cinfo->recv_mutex);
 572         return error;
 573 }
 574
 575 /* lock_comm()
 576  * Sets the MD_CLUSTER_SEND_LOCK bit to lock the send channel.
 577  */
 578 static int lock_comm(struct md_cluster_info *cinfo)
 579 {
 580         wait_event(cinfo->wait,
 581                    !test_and_set_bit(MD_CLUSTER_SEND_LOCK, &cinfo->state));
 582
 583         return lock_token(cinfo);
 584 }
 585
 586 static void unlock_comm(struct md_cluster_info *cinfo)
 587 {
 588         WARN_ON(cinfo->token_lockres->mode != DLM_LOCK_EX);
 589         mutex_unlock(&cinfo->recv_mutex);
 590         dlm_unlock_sync(cinfo->token_lockres);
 591         clear_bit(MD_CLUSTER_SEND_LOCK, &cinfo->state);
 592         wake_up(&cinfo->wait);
 593 }
 594
 595 /* __sendmsg()
 596  * This function performs the actual sending of the message. This function is
 597  * usually called after performing the encompassing operation
 598  * The function:
 599  * 1. Grabs the message lockresource in EX mode
 600  * 2. Copies the message to the message LVB
 601  * 3. Downconverts message lockresource to CW
 602  * 4. Upconverts ack lock resource from CR to EX. This forces the BAST on other nodes
 603  *    and the other nodes read the message. The thread will wait here until all other
 604  *    nodes have released ack lock resource.
 605  * 5. Downconvert ack lockresource to CR
 606  */
 607 static int __sendmsg(struct md_cluster_info *cinfo, struct cluster_msg *cmsg)
 608 {
 609         int error;
 610         int slot = cinfo->slot_number - 1;
 611
 612         cmsg->slot = cpu_to_le32(slot);
 613         /*get EX on Message*/
 614         error = dlm_lock_sync(cinfo->message_lockres, DLM_LOCK_EX);
 615         if (error) {
 616                 pr_err("md-cluster: failed to get EX on MESSAGE (%d)\n", error);
 617                 goto failed_message;
 618         }
 619
 620         memcpy(cinfo->message_lockres->lksb.sb_lvbptr, (void *)cmsg,
 621                         sizeof(struct cluster_msg));
 622         /*down-convert EX to CW on Message*/
 623         error = dlm_lock_sync(cinfo->message_lockres, DLM_LOCK_CW);
 624         if (error) {
 625                 pr_err("md-cluster: failed to convert EX to CW on MESSAGE(%d)\n",
 626                                 error);
 627                 goto failed_ack;
 628         }
 629
 630         /*up-convert CR to EX on Ack*/
 631         error = dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_EX);
 632         if (error) {
 633                 pr_err("md-cluster: failed to convert CR to EX on ACK(%d)\n",
 634                                 error);
 635                 goto failed_ack;
 636         }
 637
 638         /*down-convert EX to CR on Ack*/
 639         error = dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_CR);
 640         if (error) {
 641                 pr_err("md-cluster: failed to convert EX to CR on ACK(%d)\n",
 642                                 error);
 643                 goto failed_ack;
 644         }
 645
 646 failed_ack:
 647         error = dlm_unlock_sync(cinfo->message_lockres);
 648         if (unlikely(error != 0)) {
 649                 pr_err("md-cluster: failed convert to NL on MESSAGE(%d)\n",
 650                         error);
 651                 /* in case the message can't be released due to some reason */
 652                 goto failed_ack;
 653         }
 654 failed_message:
 655         return error;
 656 }
 657
 658 static int sendmsg(struct md_cluster_info *cinfo, struct cluster_msg *cmsg)
 659 {
 660         int ret;
 661
 662         lock_comm(cinfo);
 663         ret = __sendmsg(cinfo, cmsg);
 664         unlock_comm(cinfo);
 665         return ret;
 666 }
 667
 668 static int gather_all_resync_info(struct mddev *mddev, int total_slots)
 669 {
 670         struct md_cluster_info *cinfo = mddev->cluster_info;
 671         int i, ret = 0;
 672         struct dlm_lock_resource *bm_lockres;
 673         struct suspend_info *s;
 674         char str[64];
 675         sector_t lo, hi;
 676
 677
 678         for (i = 0; i < total_slots; i++) {
 679                 memset(str, '\0', 64);
 680                 snprintf(str, 64, "bitmap%04d", i);
 681                 bm_lockres = lockres_init(mddev, str, NULL, 1);
 682                 if (!bm_lockres)
 683                         return -ENOMEM;
 684                 if (i == (cinfo->slot_number - 1))
 685                         continue;
 686
 687                 bm_lockres->flags |= DLM_LKF_NOQUEUE;
 688                 ret = dlm_lock_sync(bm_lockres, DLM_LOCK_PW);
 689                 if (ret == -EAGAIN) {
 690                         memset(bm_lockres->lksb.sb_lvbptr, '\0', LVB_SIZE);
 691                         s = read_resync_info(mddev, bm_lockres);
 692                         if (s) {
 693                                 pr_info("%s:%d Resync[%llu..%llu] in progress on %d\n",
 694                                                 __func__, __LINE__,
 695                                                 (unsigned long long) s->lo,
 696                                                 (unsigned long long) s->hi, i);
 697                                 spin_lock_irq(&cinfo->suspend_lock);
 698                                 s->slot = i;
 699                                 list_add(&s->list, &cinfo->suspend_list);
 700                                 spin_unlock_irq(&cinfo->suspend_lock);
 701                         }
 702                         ret = 0;
 703                         lockres_free(bm_lockres);
 704                         continue;
 705                 }
 706                 if (ret) {
 707                         lockres_free(bm_lockres);
 708                         goto out;
 709                 }
 710
 711                 /* Read the disk bitmap sb and check if it needs recovery */
 712                 ret = bitmap_copy_from_slot(mddev, i, &lo, &hi, false);
 713                 if (ret) {
 714                         pr_warn("md-cluster: Could not gather bitmaps from slot %d", i);
 715                         lockres_free(bm_lockres);
 716                         continue;
 717                 }
 718                 if ((hi > 0) && (lo < mddev->recovery_cp)) {
 719                         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
 720                         mddev->recovery_cp = lo;
 721                         md_check_recovery(mddev);
 722                 }
 723
 724                 dlm_unlock_sync(bm_lockres);
 725                 lockres_free(bm_lockres);
 726         }
 727 out:
 728         return ret;
 729 }
 730
 731 static int join(struct mddev *mddev, int nodes)
 732 {
 733         struct md_cluster_info *cinfo;
 734         int ret, ops_rv;
 735         char str[64];
 736
 737         cinfo = kzalloc(sizeof(struct md_cluster_info), GFP_KERNEL);
 738         if (!cinfo)
 739                 return -ENOMEM;
 740
 741         INIT_LIST_HEAD(&cinfo->suspend_list);
 742         spin_lock_init(&cinfo->suspend_lock);
 743         init_completion(&cinfo->completion);
 744         set_bit(MD_CLUSTER_BEGIN_JOIN_CLUSTER, &cinfo->state);
 745         init_waitqueue_head(&cinfo->wait);
 746         mutex_init(&cinfo->recv_mutex);
 747
 748         mddev->cluster_info = cinfo;
 749
 750         memset(str, 0, 64);
 751         sprintf(str, "%pU", mddev->uuid);
 752         ret = dlm_new_lockspace(str, mddev->bitmap_info.cluster_name,
 753                                 DLM_LSFL_FS, LVB_SIZE,
 754                                 &md_ls_ops, mddev, &ops_rv, &cinfo->lockspace);
 755         if (ret)
 756                 goto err;
 757         wait_for_completion(&cinfo->completion);
 758         if (nodes < cinfo->slot_number) {
 759                 pr_err("md-cluster: Slot allotted(%d) is greater than available slots(%d).",
 760                         cinfo->slot_number, nodes);
 761                 ret = -ERANGE;
 762                 goto err;
 763         }
 764         /* Initiate the communication resources */
 765         ret = -ENOMEM;
 766         cinfo->recv_thread = md_register_thread(recv_daemon, mddev, "cluster_recv");
 767         if (!cinfo->recv_thread) {
 768                 pr_err("md-cluster: cannot allocate memory for recv_thread!\n");
 769                 goto err;
 770         }
 771         cinfo->message_lockres = lockres_init(mddev, "message", NULL, 1);
 772         if (!cinfo->message_lockres)
 773                 goto err;
 774         cinfo->token_lockres = lockres_init(mddev, "token", NULL, 0);
 775         if (!cinfo->token_lockres)
 776                 goto err;
 777         cinfo->ack_lockres = lockres_init(mddev, "ack", ack_bast, 0);
 778         if (!cinfo->ack_lockres)
 779                 goto err;
 780         cinfo->no_new_dev_lockres = lockres_init(mddev, "no-new-dev", NULL, 0);
 781         if (!cinfo->no_new_dev_lockres)
 782                 goto err;
 783
 784         /* get sync CR lock on ACK. */
 785         if (dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_CR))
 786                 pr_err("md-cluster: failed to get a sync CR lock on ACK!(%d)\n",
 787                                 ret);
 788         /* get sync CR lock on no-new-dev. */
 789         if (dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR))
 790                 pr_err("md-cluster: failed to get a sync CR lock on no-new-dev!(%d)\n", ret);
 791
 792
 793         pr_info("md-cluster: Joined cluster %s slot %d\n", str, cinfo->slot_number);
 794         snprintf(str, 64, "bitmap%04d", cinfo->slot_number - 1);
 795         cinfo->bitmap_lockres = lockres_init(mddev, str, NULL, 1);
 796         if (!cinfo->bitmap_lockres)
 797                 goto err;
 798         if (dlm_lock_sync(cinfo->bitmap_lockres, DLM_LOCK_PW)) {
 799                 pr_err("Failed to get bitmap lock\n");
 800                 ret = -EINVAL;
 801                 goto err;
 802         }
 803
 804         cinfo->resync_lockres = lockres_init(mddev, "resync", NULL, 0);
 805         if (!cinfo->resync_lockres)
 806                 goto err;
 807
 808         ret = gather_all_resync_info(mddev, nodes);
 809         if (ret)
 810                 goto err;
 811
 812         return 0;
 813 err:
 814         lockres_free(cinfo->message_lockres);
 815         lockres_free(cinfo->token_lockres);
 816         lockres_free(cinfo->ack_lockres);
 817         lockres_free(cinfo->no_new_dev_lockres);
 818         lockres_free(cinfo->resync_lockres);
 819         lockres_free(cinfo->bitmap_lockres);
 820         if (cinfo->lockspace)
 821                 dlm_release_lockspace(cinfo->lockspace, 2);
 822         mddev->cluster_info = NULL;
 823         kfree(cinfo);
 824         return ret;
 825 }
 826
 827 static void resync_bitmap(struct mddev *mddev)
 828 {
 829         struct md_cluster_info *cinfo = mddev->cluster_info;
 830         struct cluster_msg cmsg = {0};
 831         int err;
 832
 833         cmsg.type = cpu_to_le32(BITMAP_NEEDS_SYNC);
 834         err = sendmsg(cinfo, &cmsg);
 835         if (err)
 836                 pr_err("%s:%d: failed to send BITMAP_NEEDS_SYNC message (%d)\n",
 837                         __func__, __LINE__, err);
 838 }
 839
 840 static void unlock_all_bitmaps(struct mddev *mddev);
 841 static int leave(struct mddev *mddev)
 842 {
 843         struct md_cluster_info *cinfo = mddev->cluster_info;
 844
 845         if (!cinfo)
 846                 return 0;
 847
 848         /* BITMAP_NEEDS_SYNC message should be sent when node
 849          * is leaving the cluster with dirty bitmap, also we
 850          * can only deliver it when dlm connection is available */
 851         if (cinfo->slot_number > 0 && mddev->recovery_cp != MaxSector)
 852                 resync_bitmap(mddev);
 853
 854         md_unregister_thread(&cinfo->recovery_thread);
 855         md_unregister_thread(&cinfo->recv_thread);
 856         lockres_free(cinfo->message_lockres);
 857         lockres_free(cinfo->token_lockres);
 858         lockres_free(cinfo->ack_lockres);
 859         lockres_free(cinfo->no_new_dev_lockres);
 860         lockres_free(cinfo->bitmap_lockres);
 861         unlock_all_bitmaps(mddev);
 862         dlm_release_lockspace(cinfo->lockspace, 2);
 863         return 0;
 864 }
 865
 866 /* slot_number(): Returns the MD slot number to use
 867  * DLM starts the slot numbers from 1, wheras cluster-md
 868  * wants the number to be from zero, so we deduct one
 869  */
 870 static int slot_number(struct mddev *mddev)
 871 {
 872         struct md_cluster_info *cinfo = mddev->cluster_info;
 873
 874         return cinfo->slot_number - 1;
 875 }
 876
 877 /*
 878  * Check if the communication is already locked, else lock the communication
 879  * channel.
 880  * If it is already locked, token is in EX mode, and hence lock_token()
 881  * should not be called.
 882  */
 883 static int metadata_update_start(struct mddev *mddev)
 884 {
 885         struct md_cluster_info *cinfo = mddev->cluster_info;
 886
 887         wait_event(cinfo->wait,
 888                    !test_and_set_bit(MD_CLUSTER_SEND_LOCK, &cinfo->state) ||
 889                    test_and_clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state));
 890
 891         /* If token is already locked, return 0 */
 892         if (cinfo->token_lockres->mode == DLM_LOCK_EX)
 893                 return 0;
 894
 895         return lock_token(cinfo);
 896 }
 897
 898 static int metadata_update_finish(struct mddev *mddev)
 899 {
 900         struct md_cluster_info *cinfo = mddev->cluster_info;
 901         struct cluster_msg cmsg;
 902         struct md_rdev *rdev;
 903         int ret = 0;
 904         int raid_slot = -1;
 905
 906         memset(&cmsg, 0, sizeof(cmsg));
 907         cmsg.type = cpu_to_le32(METADATA_UPDATED);
 908         /* Pick up a good active device number to send.
 909          */
 910         rdev_for_each(rdev, mddev)
 911                 if (rdev->raid_disk > -1 && !test_bit(Faulty, &rdev->flags)) {
 912                         raid_slot = rdev->desc_nr;
 913                         break;
 914                 }
 915         if (raid_slot >= 0) {
 916                 cmsg.raid_slot = cpu_to_le32(raid_slot);
 917                 ret = __sendmsg(cinfo, &cmsg);
 918         } else
 919                 pr_warn("md-cluster: No good device id found to send\n");
 920         clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
 921         unlock_comm(cinfo);
 922         return ret;
 923 }
 924
 925 static void metadata_update_cancel(struct mddev *mddev)
 926 {
 927         struct md_cluster_info *cinfo = mddev->cluster_info;
 928         clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
 929         unlock_comm(cinfo);
 930 }
 931
 932 static int resync_start(struct mddev *mddev)
 933 {
 934         struct md_cluster_info *cinfo = mddev->cluster_info;
 935         cinfo->resync_lockres->flags |= DLM_LKF_NOQUEUE;
 936         return dlm_lock_sync(cinfo->resync_lockres, DLM_LOCK_EX);
 937 }
 938
 939 static int resync_info_update(struct mddev *mddev, sector_t lo, sector_t hi)
 940 {
 941         struct md_cluster_info *cinfo = mddev->cluster_info;
 942         struct resync_info ri;
 943         struct cluster_msg cmsg = {0};
 944
 945         /* do not send zero again, if we have sent before */
 946         if (hi == 0) {
 947                 memcpy(&ri, cinfo->bitmap_lockres->lksb.sb_lvbptr, sizeof(struct resync_info));
 948                 if (le64_to_cpu(ri.hi) == 0)
 949                         return 0;
 950         }
 951
 952         add_resync_info(cinfo->bitmap_lockres, lo, hi);
 953         /* Re-acquire the lock to refresh LVB */
 954         dlm_lock_sync(cinfo->bitmap_lockres, DLM_LOCK_PW);
 955         cmsg.type = cpu_to_le32(RESYNCING);
 956         cmsg.low = cpu_to_le64(lo);
 957         cmsg.high = cpu_to_le64(hi);
 958
 959         return sendmsg(cinfo, &cmsg);
 960 }
 961
 962 static int resync_finish(struct mddev *mddev)
 963 {
 964         struct md_cluster_info *cinfo = mddev->cluster_info;
 965         cinfo->resync_lockres->flags &= ~DLM_LKF_NOQUEUE;
 966         dlm_unlock_sync(cinfo->resync_lockres);
 967         return resync_info_update(mddev, 0, 0);
 968 }
 969
 970 static int area_resyncing(struct mddev *mddev, int direction,
 971                 sector_t lo, sector_t hi)
 972 {
 973         struct md_cluster_info *cinfo = mddev->cluster_info;
 974         int ret = 0;
 975         struct suspend_info *s;
 976
 977         if ((direction == READ) &&
 978                 test_bit(MD_CLUSTER_SUSPEND_READ_BALANCING, &cinfo->state))
 979                 return 1;
 980
 981         spin_lock_irq(&cinfo->suspend_lock);
 982         if (list_empty(&cinfo->suspend_list))
 983                 goto out;
 984         list_for_each_entry(s, &cinfo->suspend_list, list)
 985                 if (hi > s->lo && lo < s->hi) {
 986                         ret = 1;
 987                         break;
 988                 }
 989 out:
 990         spin_unlock_irq(&cinfo->suspend_lock);
 991         return ret;
 992 }
 993
 994 /* add_new_disk() - initiates a disk add
 995  * However, if this fails before writing md_update_sb(),
 996  * add_new_disk_cancel() must be called to release token lock
 997  */
 998 static int add_new_disk(struct mddev *mddev, struct md_rdev *rdev)
 999 {
1000         struct md_cluster_info *cinfo = mddev->cluster_info;
1001         struct cluster_msg cmsg;
1002         int ret = 0;
1003         struct mdp_superblock_1 *sb = page_address(rdev->sb_page);
1004         char *uuid = sb->device_uuid;
1005
1006         memset(&cmsg, 0, sizeof(cmsg));
1007         cmsg.type = cpu_to_le32(NEWDISK);
1008         memcpy(cmsg.uuid, uuid, 16);
1009         cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
1010         lock_comm(cinfo);
1011         ret = __sendmsg(cinfo, &cmsg);
1012         if (ret)
1013                 return ret;
1014         cinfo->no_new_dev_lockres->flags |= DLM_LKF_NOQUEUE;
1015         ret = dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_EX);
1016         cinfo->no_new_dev_lockres->flags &= ~DLM_LKF_NOQUEUE;
1017         /* Some node does not "see" the device */
1018         if (ret == -EAGAIN)
1019                 ret = -ENOENT;
1020         if (ret)
1021                 unlock_comm(cinfo);
1022         else {
1023                 dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR);
1024                 set_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1025                 wake_up(&cinfo->wait);
1026         }
1027         return ret;
1028 }
1029
1030 static void add_new_disk_cancel(struct mddev *mddev)
1031 {
1032         struct md_cluster_info *cinfo = mddev->cluster_info;
1033         clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1034         unlock_comm(cinfo);
1035 }
1036
1037 static int new_disk_ack(struct mddev *mddev, bool ack)
1038 {
1039         struct md_cluster_info *cinfo = mddev->cluster_info;
1040
1041         if (!test_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state)) {
1042                 pr_warn("md-cluster(%s): Spurious cluster confirmation\n", mdname(mddev));
1043                 return -EINVAL;
1044         }
1045
1046         if (ack)
1047                 dlm_unlock_sync(cinfo->no_new_dev_lockres);
1048         complete(&cinfo->newdisk_completion);
1049         return 0;
1050 }
1051
1052 static int remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1053 {
1054         struct cluster_msg cmsg = {0};
1055         struct md_cluster_info *cinfo = mddev->cluster_info;
1056         cmsg.type = cpu_to_le32(REMOVE);
1057         cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
1058         return sendmsg(cinfo, &cmsg);
1059 }
1060
1061 static int lock_all_bitmaps(struct mddev *mddev)
1062 {
1063         int slot, my_slot, ret, held = 1, i = 0;
1064         char str[64];
1065         struct md_cluster_info *cinfo = mddev->cluster_info;
1066
1067         cinfo->other_bitmap_lockres = kzalloc((mddev->bitmap_info.nodes - 1) *
1068                                              sizeof(struct dlm_lock_resource *),
1069                                              GFP_KERNEL);
1070         if (!cinfo->other_bitmap_lockres) {
1071                 pr_err("md: can't alloc mem for other bitmap locks\n");
1072                 return 0;
1073         }
1074
1075         my_slot = slot_number(mddev);
1076         for (slot = 0; slot < mddev->bitmap_info.nodes; slot++) {
1077                 if (slot == my_slot)
1078                         continue;
1079
1080                 memset(str, '\0', 64);
1081                 snprintf(str, 64, "bitmap%04d", slot);
1082                 cinfo->other_bitmap_lockres[i] = lockres_init(mddev, str, NULL, 1);
1083                 if (!cinfo->other_bitmap_lockres[i])
1084                         return -ENOMEM;
1085
1086                 cinfo->other_bitmap_lockres[i]->flags |= DLM_LKF_NOQUEUE;
1087                 ret = dlm_lock_sync(cinfo->other_bitmap_lockres[i], DLM_LOCK_PW);
1088                 if (ret)
1089                         held = -1;
1090                 i++;
1091         }
1092
1093         return held;
1094 }
1095
1096 static void unlock_all_bitmaps(struct mddev *mddev)
1097 {
1098         struct md_cluster_info *cinfo = mddev->cluster_info;
1099         int i;
1100
1101         /* release other node's bitmap lock if they are existed */
1102         if (cinfo->other_bitmap_lockres) {
1103                 for (i = 0; i < mddev->bitmap_info.nodes - 1; i++) {
1104                         if (cinfo->other_bitmap_lockres[i]) {
1105                                 dlm_unlock_sync(cinfo->other_bitmap_lockres[i]);
1106                                 lockres_free(cinfo->other_bitmap_lockres[i]);
1107                         }
1108                 }
1109                 kfree(cinfo->other_bitmap_lockres);
1110         }
1111 }
1112
1113 static int gather_bitmaps(struct md_rdev *rdev)
1114 {
1115         int sn, err;
1116         sector_t lo, hi;
1117         struct cluster_msg cmsg = {0};
1118         struct mddev *mddev = rdev->mddev;
1119         struct md_cluster_info *cinfo = mddev->cluster_info;
1120
1121         cmsg.type = cpu_to_le32(RE_ADD);
1122         cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
1123         err = sendmsg(cinfo, &cmsg);
1124         if (err)
1125                 goto out;
1126
1127         for (sn = 0; sn < mddev->bitmap_info.nodes; sn++) {
1128                 if (sn == (cinfo->slot_number - 1))
1129                         continue;
1130                 err = bitmap_copy_from_slot(mddev, sn, &lo, &hi, false);
1131                 if (err) {
1132                         pr_warn("md-cluster: Could not gather bitmaps from slot %d", sn);
1133                         goto out;
1134                 }
1135                 if ((hi > 0) && (lo < mddev->recovery_cp))
1136                         mddev->recovery_cp = lo;
1137         }
1138 out:
1139         return err;
1140 }
1141
1142 static struct md_cluster_operations cluster_ops = {
1143         .join   = join,
1144         .leave  = leave,
1145         .slot_number = slot_number,
1146         .resync_start = resync_start,
1147         .resync_finish = resync_finish,
1148         .resync_info_update = resync_info_update,
1149         .metadata_update_start = metadata_update_start,
1150         .metadata_update_finish = metadata_update_finish,
1151         .metadata_update_cancel = metadata_update_cancel,
1152         .area_resyncing = area_resyncing,
1153         .add_new_disk = add_new_disk,
1154         .add_new_disk_cancel = add_new_disk_cancel,
1155         .new_disk_ack = new_disk_ack,
1156         .remove_disk = remove_disk,
1157         .gather_bitmaps = gather_bitmaps,
1158         .lock_all_bitmaps = lock_all_bitmaps,
1159         .unlock_all_bitmaps = unlock_all_bitmaps,
1160 };
1161
1162 static int __init cluster_init(void)
1163 {
1164         pr_warn("md-cluster: EXPERIMENTAL. Use with caution\n");
1165         pr_info("Registering Cluster MD functions\n");
1166         register_md_cluster_operations(&cluster_ops, THIS_MODULE);
1167         return 0;
1168 }
1169
1170 static void cluster_exit(void)
1171 {
1172         unregister_md_cluster_operations();
1173 }
1174
1175 module_init(cluster_init);
1176 module_exit(cluster_exit);
1177 MODULE_AUTHOR("SUSE");
1178 MODULE_LICENSE("GPL");
1179 MODULE_DESCRIPTION("Clustering support for MD");