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 (such as adding a disk) must lock the 57 * communication channel, so as to perform extra operations 58 * (update metadata) and no other operation is allowed on the 59 * MD. Token needs to be locked and held until the operation 60 * completes witha md_update_sb(), which would eventually release 61 * the lock. 62 */ 63 #define MD_CLUSTER_SEND_LOCKED_ALREADY 5 64 65 66 struct md_cluster_info { 67 /* dlm lock space and resources for clustered raid. */ 68 dlm_lockspace_t *lockspace; 69 int slot_number; 70 struct completion completion; 71 struct mutex recv_mutex; 72 struct dlm_lock_resource *bitmap_lockres; 73 struct dlm_lock_resource **other_bitmap_lockres; 74 struct dlm_lock_resource *resync_lockres; 75 struct list_head suspend_list; 76 spinlock_t suspend_lock; 77 struct md_thread *recovery_thread; 78 unsigned long recovery_map; 79 /* communication loc resources */ 80 struct dlm_lock_resource *ack_lockres; 81 struct dlm_lock_resource *message_lockres; 82 struct dlm_lock_resource *token_lockres; 83 struct dlm_lock_resource *no_new_dev_lockres; 84 struct md_thread *recv_thread; 85 struct completion newdisk_completion; 86 wait_queue_head_t wait; 87 unsigned long state; 88 }; 89 90 enum msg_type { 91 METADATA_UPDATED = 0, 92 RESYNCING, 93 NEWDISK, 94 REMOVE, 95 RE_ADD, 96 BITMAP_NEEDS_SYNC, 97 }; 98 99 struct cluster_msg { 100 __le32 type; 101 __le32 slot; 102 /* TODO: Unionize this for smaller footprint */ 103 __le64 low; 104 __le64 high; 105 char uuid[16]; 106 __le32 raid_slot; 107 }; 108 109 static void sync_ast(void *arg) 110 { 111 struct dlm_lock_resource *res; 112 113 res = arg; 114 complete(&res->completion); 115 } 116 117 static int dlm_lock_sync(struct dlm_lock_resource *res, int mode) 118 { 119 int ret = 0; 120 121 ret = dlm_lock(res->ls, mode, &res->lksb, 122 res->flags, res->name, strlen(res->name), 123 0, sync_ast, res, res->bast); 124 if (ret) 125 return ret; 126 wait_for_completion(&res->completion); 127 if (res->lksb.sb_status == 0) 128 res->mode = mode; 129 return res->lksb.sb_status; 130 } 131 132 static int dlm_unlock_sync(struct dlm_lock_resource *res) 133 { 134 return dlm_lock_sync(res, DLM_LOCK_NL); 135 } 136 137 static struct dlm_lock_resource *lockres_init(struct mddev *mddev, 138 char *name, void (*bastfn)(void *arg, int mode), int with_lvb) 139 { 140 struct dlm_lock_resource *res = NULL; 141 int ret, namelen; 142 struct md_cluster_info *cinfo = mddev->cluster_info; 143 144 res = kzalloc(sizeof(struct dlm_lock_resource), GFP_KERNEL); 145 if (!res) 146 return NULL; 147 init_completion(&res->completion); 148 res->ls = cinfo->lockspace; 149 res->mddev = mddev; 150 res->mode = DLM_LOCK_IV; 151 namelen = strlen(name); 152 res->name = kzalloc(namelen + 1, GFP_KERNEL); 153 if (!res->name) { 154 pr_err("md-cluster: Unable to allocate resource name for resource %s\n", name); 155 goto out_err; 156 } 157 strlcpy(res->name, name, namelen + 1); 158 if (with_lvb) { 159 res->lksb.sb_lvbptr = kzalloc(LVB_SIZE, GFP_KERNEL); 160 if (!res->lksb.sb_lvbptr) { 161 pr_err("md-cluster: Unable to allocate LVB for resource %s\n", name); 162 goto out_err; 163 } 164 res->flags = DLM_LKF_VALBLK; 165 } 166 167 if (bastfn) 168 res->bast = bastfn; 169 170 res->flags |= DLM_LKF_EXPEDITE; 171 172 ret = dlm_lock_sync(res, DLM_LOCK_NL); 173 if (ret) { 174 pr_err("md-cluster: Unable to lock NL on new lock resource %s\n", name); 175 goto out_err; 176 } 177 res->flags &= ~DLM_LKF_EXPEDITE; 178 res->flags |= DLM_LKF_CONVERT; 179 180 return res; 181 out_err: 182 kfree(res->lksb.sb_lvbptr); 183 kfree(res->name); 184 kfree(res); 185 return NULL; 186 } 187 188 static void lockres_free(struct dlm_lock_resource *res) 189 { 190 int ret; 191 192 if (!res) 193 return; 194 195 /* cancel a lock request or a conversion request that is blocked */ 196 res->flags |= DLM_LKF_CANCEL; 197 retry: 198 ret = dlm_unlock(res->ls, res->lksb.sb_lkid, 0, &res->lksb, res); 199 if (unlikely(ret != 0)) { 200 pr_info("%s: failed to unlock %s return %d\n", __func__, res->name, ret); 201 202 /* if a lock conversion is cancelled, then the lock is put 203 * back to grant queue, need to ensure it is unlocked */ 204 if (ret == -DLM_ECANCEL) 205 goto retry; 206 } 207 res->flags &= ~DLM_LKF_CANCEL; 208 wait_for_completion(&res->completion); 209 210 kfree(res->name); 211 kfree(res->lksb.sb_lvbptr); 212 kfree(res); 213 } 214 215 static void add_resync_info(struct dlm_lock_resource *lockres, 216 sector_t lo, sector_t hi) 217 { 218 struct resync_info *ri; 219 220 ri = (struct resync_info *)lockres->lksb.sb_lvbptr; 221 ri->lo = cpu_to_le64(lo); 222 ri->hi = cpu_to_le64(hi); 223 } 224 225 static struct suspend_info *read_resync_info(struct mddev *mddev, struct dlm_lock_resource *lockres) 226 { 227 struct resync_info ri; 228 struct suspend_info *s = NULL; 229 sector_t hi = 0; 230 231 dlm_lock_sync(lockres, DLM_LOCK_CR); 232 memcpy(&ri, lockres->lksb.sb_lvbptr, sizeof(struct resync_info)); 233 hi = le64_to_cpu(ri.hi); 234 if (hi > 0) { 235 s = kzalloc(sizeof(struct suspend_info), GFP_KERNEL); 236 if (!s) 237 goto out; 238 s->hi = hi; 239 s->lo = le64_to_cpu(ri.lo); 240 } 241 dlm_unlock_sync(lockres); 242 out: 243 return s; 244 } 245 246 static void recover_bitmaps(struct md_thread *thread) 247 { 248 struct mddev *mddev = thread->mddev; 249 struct md_cluster_info *cinfo = mddev->cluster_info; 250 struct dlm_lock_resource *bm_lockres; 251 char str[64]; 252 int slot, ret; 253 struct suspend_info *s, *tmp; 254 sector_t lo, hi; 255 256 while (cinfo->recovery_map) { 257 slot = fls64((u64)cinfo->recovery_map) - 1; 258 259 /* Clear suspend_area associated with the bitmap */ 260 spin_lock_irq(&cinfo->suspend_lock); 261 list_for_each_entry_safe(s, tmp, &cinfo->suspend_list, list) 262 if (slot == s->slot) { 263 list_del(&s->list); 264 kfree(s); 265 } 266 spin_unlock_irq(&cinfo->suspend_lock); 267 268 snprintf(str, 64, "bitmap%04d", slot); 269 bm_lockres = lockres_init(mddev, str, NULL, 1); 270 if (!bm_lockres) { 271 pr_err("md-cluster: Cannot initialize bitmaps\n"); 272 goto clear_bit; 273 } 274 275 ret = dlm_lock_sync(bm_lockres, DLM_LOCK_PW); 276 if (ret) { 277 pr_err("md-cluster: Could not DLM lock %s: %d\n", 278 str, ret); 279 goto clear_bit; 280 } 281 ret = bitmap_copy_from_slot(mddev, slot, &lo, &hi, true); 282 if (ret) { 283 pr_err("md-cluster: Could not copy data from bitmap %d\n", slot); 284 goto dlm_unlock; 285 } 286 if (hi > 0) { 287 /* TODO:Wait for current resync to get over */ 288 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery); 289 if (lo < mddev->recovery_cp) 290 mddev->recovery_cp = lo; 291 md_check_recovery(mddev); 292 } 293 dlm_unlock: 294 dlm_unlock_sync(bm_lockres); 295 clear_bit: 296 lockres_free(bm_lockres); 297 clear_bit(slot, &cinfo->recovery_map); 298 } 299 } 300 301 static void recover_prep(void *arg) 302 { 303 struct mddev *mddev = arg; 304 struct md_cluster_info *cinfo = mddev->cluster_info; 305 set_bit(MD_CLUSTER_SUSPEND_READ_BALANCING, &cinfo->state); 306 } 307 308 static void __recover_slot(struct mddev *mddev, int slot) 309 { 310 struct md_cluster_info *cinfo = mddev->cluster_info; 311 312 set_bit(slot, &cinfo->recovery_map); 313 if (!cinfo->recovery_thread) { 314 cinfo->recovery_thread = md_register_thread(recover_bitmaps, 315 mddev, "recover"); 316 if (!cinfo->recovery_thread) { 317 pr_warn("md-cluster: Could not create recovery thread\n"); 318 return; 319 } 320 } 321 md_wakeup_thread(cinfo->recovery_thread); 322 } 323 324 static void recover_slot(void *arg, struct dlm_slot *slot) 325 { 326 struct mddev *mddev = arg; 327 struct md_cluster_info *cinfo = mddev->cluster_info; 328 329 pr_info("md-cluster: %s Node %d/%d down. My slot: %d. Initiating recovery.\n", 330 mddev->bitmap_info.cluster_name, 331 slot->nodeid, slot->slot, 332 cinfo->slot_number); 333 /* deduct one since dlm slot starts from one while the num of 334 * cluster-md begins with 0 */ 335 __recover_slot(mddev, slot->slot - 1); 336 } 337 338 static void recover_done(void *arg, struct dlm_slot *slots, 339 int num_slots, int our_slot, 340 uint32_t generation) 341 { 342 struct mddev *mddev = arg; 343 struct md_cluster_info *cinfo = mddev->cluster_info; 344 345 cinfo->slot_number = our_slot; 346 /* completion is only need to be complete when node join cluster, 347 * it doesn't need to run during another node's failure */ 348 if (test_bit(MD_CLUSTER_BEGIN_JOIN_CLUSTER, &cinfo->state)) { 349 complete(&cinfo->completion); 350 clear_bit(MD_CLUSTER_BEGIN_JOIN_CLUSTER, &cinfo->state); 351 } 352 clear_bit(MD_CLUSTER_SUSPEND_READ_BALANCING, &cinfo->state); 353 } 354 355 /* the ops is called when node join the cluster, and do lock recovery 356 * if node failure occurs */ 357 static const struct dlm_lockspace_ops md_ls_ops = { 358 .recover_prep = recover_prep, 359 .recover_slot = recover_slot, 360 .recover_done = recover_done, 361 }; 362 363 /* 364 * The BAST function for the ack lock resource 365 * This function wakes up the receive thread in 366 * order to receive and process the message. 367 */ 368 static void ack_bast(void *arg, int mode) 369 { 370 struct dlm_lock_resource *res = arg; 371 struct md_cluster_info *cinfo = res->mddev->cluster_info; 372 373 if (mode == DLM_LOCK_EX) 374 md_wakeup_thread(cinfo->recv_thread); 375 } 376 377 static void __remove_suspend_info(struct md_cluster_info *cinfo, int slot) 378 { 379 struct suspend_info *s, *tmp; 380 381 list_for_each_entry_safe(s, tmp, &cinfo->suspend_list, list) 382 if (slot == s->slot) { 383 list_del(&s->list); 384 kfree(s); 385 break; 386 } 387 } 388 389 static void remove_suspend_info(struct mddev *mddev, int slot) 390 { 391 struct md_cluster_info *cinfo = mddev->cluster_info; 392 spin_lock_irq(&cinfo->suspend_lock); 393 __remove_suspend_info(cinfo, slot); 394 spin_unlock_irq(&cinfo->suspend_lock); 395 mddev->pers->quiesce(mddev, 2); 396 } 397 398 399 static void process_suspend_info(struct mddev *mddev, 400 int slot, sector_t lo, sector_t hi) 401 { 402 struct md_cluster_info *cinfo = mddev->cluster_info; 403 struct suspend_info *s; 404 405 if (!hi) { 406 remove_suspend_info(mddev, slot); 407 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery); 408 md_wakeup_thread(mddev->thread); 409 return; 410 } 411 s = kzalloc(sizeof(struct suspend_info), GFP_KERNEL); 412 if (!s) 413 return; 414 s->slot = slot; 415 s->lo = lo; 416 s->hi = hi; 417 mddev->pers->quiesce(mddev, 1); 418 mddev->pers->quiesce(mddev, 0); 419 spin_lock_irq(&cinfo->suspend_lock); 420 /* Remove existing entry (if exists) before adding */ 421 __remove_suspend_info(cinfo, slot); 422 list_add(&s->list, &cinfo->suspend_list); 423 spin_unlock_irq(&cinfo->suspend_lock); 424 mddev->pers->quiesce(mddev, 2); 425 } 426 427 static void process_add_new_disk(struct mddev *mddev, struct cluster_msg *cmsg) 428 { 429 char disk_uuid[64]; 430 struct md_cluster_info *cinfo = mddev->cluster_info; 431 char event_name[] = "EVENT=ADD_DEVICE"; 432 char raid_slot[16]; 433 char *envp[] = {event_name, disk_uuid, raid_slot, NULL}; 434 int len; 435 436 len = snprintf(disk_uuid, 64, "DEVICE_UUID="); 437 sprintf(disk_uuid + len, "%pU", cmsg->uuid); 438 snprintf(raid_slot, 16, "RAID_DISK=%d", le32_to_cpu(cmsg->raid_slot)); 439 pr_info("%s:%d Sending kobject change with %s and %s\n", __func__, __LINE__, disk_uuid, raid_slot); 440 init_completion(&cinfo->newdisk_completion); 441 set_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state); 442 kobject_uevent_env(&disk_to_dev(mddev->gendisk)->kobj, KOBJ_CHANGE, envp); 443 wait_for_completion_timeout(&cinfo->newdisk_completion, 444 NEW_DEV_TIMEOUT); 445 clear_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state); 446 } 447 448 449 static void process_metadata_update(struct mddev *mddev, struct cluster_msg *msg) 450 { 451 struct md_cluster_info *cinfo = mddev->cluster_info; 452 mddev->good_device_nr = le32_to_cpu(msg->raid_slot); 453 set_bit(MD_RELOAD_SB, &mddev->flags); 454 dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR); 455 md_wakeup_thread(mddev->thread); 456 } 457 458 static void process_remove_disk(struct mddev *mddev, struct cluster_msg *msg) 459 { 460 struct md_rdev *rdev = md_find_rdev_nr_rcu(mddev, 461 le32_to_cpu(msg->raid_slot)); 462 463 if (rdev) { 464 set_bit(ClusterRemove, &rdev->flags); 465 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery); 466 md_wakeup_thread(mddev->thread); 467 } 468 else 469 pr_warn("%s: %d Could not find disk(%d) to REMOVE\n", 470 __func__, __LINE__, le32_to_cpu(msg->raid_slot)); 471 } 472 473 static void process_readd_disk(struct mddev *mddev, struct cluster_msg *msg) 474 { 475 struct md_rdev *rdev = md_find_rdev_nr_rcu(mddev, 476 le32_to_cpu(msg->raid_slot)); 477 478 if (rdev && test_bit(Faulty, &rdev->flags)) 479 clear_bit(Faulty, &rdev->flags); 480 else 481 pr_warn("%s: %d Could not find disk(%d) which is faulty", 482 __func__, __LINE__, le32_to_cpu(msg->raid_slot)); 483 } 484 485 static void process_recvd_msg(struct mddev *mddev, struct cluster_msg *msg) 486 { 487 if (WARN(mddev->cluster_info->slot_number - 1 == le32_to_cpu(msg->slot), 488 "node %d received it's own msg\n", le32_to_cpu(msg->slot))) 489 return; 490 switch (le32_to_cpu(msg->type)) { 491 case METADATA_UPDATED: 492 process_metadata_update(mddev, msg); 493 break; 494 case RESYNCING: 495 process_suspend_info(mddev, le32_to_cpu(msg->slot), 496 le64_to_cpu(msg->low), 497 le64_to_cpu(msg->high)); 498 break; 499 case NEWDISK: 500 process_add_new_disk(mddev, msg); 501 break; 502 case REMOVE: 503 process_remove_disk(mddev, msg); 504 break; 505 case RE_ADD: 506 process_readd_disk(mddev, msg); 507 break; 508 case BITMAP_NEEDS_SYNC: 509 __recover_slot(mddev, le32_to_cpu(msg->slot)); 510 break; 511 default: 512 pr_warn("%s:%d Received unknown message from %d\n", 513 __func__, __LINE__, msg->slot); 514 } 515 } 516 517 /* 518 * thread for receiving message 519 */ 520 static void recv_daemon(struct md_thread *thread) 521 { 522 struct md_cluster_info *cinfo = thread->mddev->cluster_info; 523 struct dlm_lock_resource *ack_lockres = cinfo->ack_lockres; 524 struct dlm_lock_resource *message_lockres = cinfo->message_lockres; 525 struct cluster_msg msg; 526 int ret; 527 528 mutex_lock(&cinfo->recv_mutex); 529 /*get CR on Message*/ 530 if (dlm_lock_sync(message_lockres, DLM_LOCK_CR)) { 531 pr_err("md/raid1:failed to get CR on MESSAGE\n"); 532 mutex_unlock(&cinfo->recv_mutex); 533 return; 534 } 535 536 /* read lvb and wake up thread to process this message_lockres */ 537 memcpy(&msg, message_lockres->lksb.sb_lvbptr, sizeof(struct cluster_msg)); 538 process_recvd_msg(thread->mddev, &msg); 539 540 /*release CR on ack_lockres*/ 541 ret = dlm_unlock_sync(ack_lockres); 542 if (unlikely(ret != 0)) 543 pr_info("unlock ack failed return %d\n", ret); 544 /*up-convert to PR on message_lockres*/ 545 ret = dlm_lock_sync(message_lockres, DLM_LOCK_PR); 546 if (unlikely(ret != 0)) 547 pr_info("lock PR on msg failed return %d\n", ret); 548 /*get CR on ack_lockres again*/ 549 ret = dlm_lock_sync(ack_lockres, DLM_LOCK_CR); 550 if (unlikely(ret != 0)) 551 pr_info("lock CR on ack failed return %d\n", ret); 552 /*release CR on message_lockres*/ 553 ret = dlm_unlock_sync(message_lockres); 554 if (unlikely(ret != 0)) 555 pr_info("unlock msg failed return %d\n", ret); 556 mutex_unlock(&cinfo->recv_mutex); 557 } 558 559 /* lock_token() 560 * Takes the lock on the TOKEN lock resource so no other 561 * node can communicate while the operation is underway. 562 */ 563 static int lock_token(struct md_cluster_info *cinfo) 564 { 565 int error; 566 567 error = dlm_lock_sync(cinfo->token_lockres, DLM_LOCK_EX); 568 if (error) 569 pr_err("md-cluster(%s:%d): failed to get EX on TOKEN (%d)\n", 570 __func__, __LINE__, error); 571 572 /* Lock the receive sequence */ 573 mutex_lock(&cinfo->recv_mutex); 574 return error; 575 } 576 577 /* lock_comm() 578 * Sets the MD_CLUSTER_SEND_LOCK bit to lock the send channel. 579 */ 580 static int lock_comm(struct md_cluster_info *cinfo) 581 { 582 wait_event(cinfo->wait, 583 !test_and_set_bit(MD_CLUSTER_SEND_LOCK, &cinfo->state)); 584 585 return lock_token(cinfo); 586 } 587 588 static void unlock_comm(struct md_cluster_info *cinfo) 589 { 590 WARN_ON(cinfo->token_lockres->mode != DLM_LOCK_EX); 591 mutex_unlock(&cinfo->recv_mutex); 592 dlm_unlock_sync(cinfo->token_lockres); 593 clear_bit(MD_CLUSTER_SEND_LOCK, &cinfo->state); 594 wake_up(&cinfo->wait); 595 } 596 597 /* __sendmsg() 598 * This function performs the actual sending of the message. This function is 599 * usually called after performing the encompassing operation 600 * The function: 601 * 1. Grabs the message lockresource in EX mode 602 * 2. Copies the message to the message LVB 603 * 3. Downconverts message lockresource to CW 604 * 4. Upconverts ack lock resource from CR to EX. This forces the BAST on other nodes 605 * and the other nodes read the message. The thread will wait here until all other 606 * nodes have released ack lock resource. 607 * 5. Downconvert ack lockresource to CR 608 */ 609 static int __sendmsg(struct md_cluster_info *cinfo, struct cluster_msg *cmsg) 610 { 611 int error; 612 int slot = cinfo->slot_number - 1; 613 614 cmsg->slot = cpu_to_le32(slot); 615 /*get EX on Message*/ 616 error = dlm_lock_sync(cinfo->message_lockres, DLM_LOCK_EX); 617 if (error) { 618 pr_err("md-cluster: failed to get EX on MESSAGE (%d)\n", error); 619 goto failed_message; 620 } 621 622 memcpy(cinfo->message_lockres->lksb.sb_lvbptr, (void *)cmsg, 623 sizeof(struct cluster_msg)); 624 /*down-convert EX to CW on Message*/ 625 error = dlm_lock_sync(cinfo->message_lockres, DLM_LOCK_CW); 626 if (error) { 627 pr_err("md-cluster: failed to convert EX to CW on MESSAGE(%d)\n", 628 error); 629 goto failed_ack; 630 } 631 632 /*up-convert CR to EX on Ack*/ 633 error = dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_EX); 634 if (error) { 635 pr_err("md-cluster: failed to convert CR to EX on ACK(%d)\n", 636 error); 637 goto failed_ack; 638 } 639 640 /*down-convert EX to CR on Ack*/ 641 error = dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_CR); 642 if (error) { 643 pr_err("md-cluster: failed to convert EX to CR on ACK(%d)\n", 644 error); 645 goto failed_ack; 646 } 647 648 failed_ack: 649 error = dlm_unlock_sync(cinfo->message_lockres); 650 if (unlikely(error != 0)) { 651 pr_err("md-cluster: failed convert to NL on MESSAGE(%d)\n", 652 error); 653 /* in case the message can't be released due to some reason */ 654 goto failed_ack; 655 } 656 failed_message: 657 return error; 658 } 659 660 static int sendmsg(struct md_cluster_info *cinfo, struct cluster_msg *cmsg) 661 { 662 int ret; 663 664 lock_comm(cinfo); 665 ret = __sendmsg(cinfo, cmsg); 666 unlock_comm(cinfo); 667 return ret; 668 } 669 670 static int gather_all_resync_info(struct mddev *mddev, int total_slots) 671 { 672 struct md_cluster_info *cinfo = mddev->cluster_info; 673 int i, ret = 0; 674 struct dlm_lock_resource *bm_lockres; 675 struct suspend_info *s; 676 char str[64]; 677 sector_t lo, hi; 678 679 680 for (i = 0; i < total_slots; i++) { 681 memset(str, '\0', 64); 682 snprintf(str, 64, "bitmap%04d", i); 683 bm_lockres = lockres_init(mddev, str, NULL, 1); 684 if (!bm_lockres) 685 return -ENOMEM; 686 if (i == (cinfo->slot_number - 1)) { 687 lockres_free(bm_lockres); 688 continue; 689 } 690 691 bm_lockres->flags |= DLM_LKF_NOQUEUE; 692 ret = dlm_lock_sync(bm_lockres, DLM_LOCK_PW); 693 if (ret == -EAGAIN) { 694 memset(bm_lockres->lksb.sb_lvbptr, '\0', LVB_SIZE); 695 s = read_resync_info(mddev, bm_lockres); 696 if (s) { 697 pr_info("%s:%d Resync[%llu..%llu] in progress on %d\n", 698 __func__, __LINE__, 699 (unsigned long long) s->lo, 700 (unsigned long long) s->hi, i); 701 spin_lock_irq(&cinfo->suspend_lock); 702 s->slot = i; 703 list_add(&s->list, &cinfo->suspend_list); 704 spin_unlock_irq(&cinfo->suspend_lock); 705 } 706 ret = 0; 707 lockres_free(bm_lockres); 708 continue; 709 } 710 if (ret) { 711 lockres_free(bm_lockres); 712 goto out; 713 } 714 715 /* Read the disk bitmap sb and check if it needs recovery */ 716 ret = bitmap_copy_from_slot(mddev, i, &lo, &hi, false); 717 if (ret) { 718 pr_warn("md-cluster: Could not gather bitmaps from slot %d", i); 719 lockres_free(bm_lockres); 720 continue; 721 } 722 if ((hi > 0) && (lo < mddev->recovery_cp)) { 723 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery); 724 mddev->recovery_cp = lo; 725 md_check_recovery(mddev); 726 } 727 728 dlm_unlock_sync(bm_lockres); 729 lockres_free(bm_lockres); 730 } 731 out: 732 return ret; 733 } 734 735 static int join(struct mddev *mddev, int nodes) 736 { 737 struct md_cluster_info *cinfo; 738 int ret, ops_rv; 739 char str[64]; 740 741 cinfo = kzalloc(sizeof(struct md_cluster_info), GFP_KERNEL); 742 if (!cinfo) 743 return -ENOMEM; 744 745 INIT_LIST_HEAD(&cinfo->suspend_list); 746 spin_lock_init(&cinfo->suspend_lock); 747 init_completion(&cinfo->completion); 748 set_bit(MD_CLUSTER_BEGIN_JOIN_CLUSTER, &cinfo->state); 749 init_waitqueue_head(&cinfo->wait); 750 mutex_init(&cinfo->recv_mutex); 751 752 mddev->cluster_info = cinfo; 753 754 memset(str, 0, 64); 755 sprintf(str, "%pU", mddev->uuid); 756 ret = dlm_new_lockspace(str, mddev->bitmap_info.cluster_name, 757 DLM_LSFL_FS, LVB_SIZE, 758 &md_ls_ops, mddev, &ops_rv, &cinfo->lockspace); 759 if (ret) 760 goto err; 761 wait_for_completion(&cinfo->completion); 762 if (nodes < cinfo->slot_number) { 763 pr_err("md-cluster: Slot allotted(%d) is greater than available slots(%d).", 764 cinfo->slot_number, nodes); 765 ret = -ERANGE; 766 goto err; 767 } 768 /* Initiate the communication resources */ 769 ret = -ENOMEM; 770 cinfo->recv_thread = md_register_thread(recv_daemon, mddev, "cluster_recv"); 771 if (!cinfo->recv_thread) { 772 pr_err("md-cluster: cannot allocate memory for recv_thread!\n"); 773 goto err; 774 } 775 cinfo->message_lockres = lockres_init(mddev, "message", NULL, 1); 776 if (!cinfo->message_lockres) 777 goto err; 778 cinfo->token_lockres = lockres_init(mddev, "token", NULL, 0); 779 if (!cinfo->token_lockres) 780 goto err; 781 cinfo->ack_lockres = lockres_init(mddev, "ack", ack_bast, 0); 782 if (!cinfo->ack_lockres) 783 goto err; 784 cinfo->no_new_dev_lockres = lockres_init(mddev, "no-new-dev", NULL, 0); 785 if (!cinfo->no_new_dev_lockres) 786 goto err; 787 788 /* get sync CR lock on ACK. */ 789 if (dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_CR)) 790 pr_err("md-cluster: failed to get a sync CR lock on ACK!(%d)\n", 791 ret); 792 /* get sync CR lock on no-new-dev. */ 793 if (dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR)) 794 pr_err("md-cluster: failed to get a sync CR lock on no-new-dev!(%d)\n", ret); 795 796 797 pr_info("md-cluster: Joined cluster %s slot %d\n", str, cinfo->slot_number); 798 snprintf(str, 64, "bitmap%04d", cinfo->slot_number - 1); 799 cinfo->bitmap_lockres = lockres_init(mddev, str, NULL, 1); 800 if (!cinfo->bitmap_lockres) 801 goto err; 802 if (dlm_lock_sync(cinfo->bitmap_lockres, DLM_LOCK_PW)) { 803 pr_err("Failed to get bitmap lock\n"); 804 ret = -EINVAL; 805 goto err; 806 } 807 808 cinfo->resync_lockres = lockres_init(mddev, "resync", NULL, 0); 809 if (!cinfo->resync_lockres) 810 goto err; 811 812 ret = gather_all_resync_info(mddev, nodes); 813 if (ret) 814 goto err; 815 816 return 0; 817 err: 818 lockres_free(cinfo->message_lockres); 819 lockres_free(cinfo->token_lockres); 820 lockres_free(cinfo->ack_lockres); 821 lockres_free(cinfo->no_new_dev_lockres); 822 lockres_free(cinfo->resync_lockres); 823 lockres_free(cinfo->bitmap_lockres); 824 if (cinfo->lockspace) 825 dlm_release_lockspace(cinfo->lockspace, 2); 826 mddev->cluster_info = NULL; 827 kfree(cinfo); 828 return ret; 829 } 830 831 static void resync_bitmap(struct mddev *mddev) 832 { 833 struct md_cluster_info *cinfo = mddev->cluster_info; 834 struct cluster_msg cmsg = {0}; 835 int err; 836 837 cmsg.type = cpu_to_le32(BITMAP_NEEDS_SYNC); 838 err = sendmsg(cinfo, &cmsg); 839 if (err) 840 pr_err("%s:%d: failed to send BITMAP_NEEDS_SYNC message (%d)\n", 841 __func__, __LINE__, err); 842 } 843 844 static void unlock_all_bitmaps(struct mddev *mddev); 845 static int leave(struct mddev *mddev) 846 { 847 struct md_cluster_info *cinfo = mddev->cluster_info; 848 849 if (!cinfo) 850 return 0; 851 852 /* BITMAP_NEEDS_SYNC message should be sent when node 853 * is leaving the cluster with dirty bitmap, also we 854 * can only deliver it when dlm connection is available */ 855 if (cinfo->slot_number > 0 && mddev->recovery_cp != MaxSector) 856 resync_bitmap(mddev); 857 858 md_unregister_thread(&cinfo->recovery_thread); 859 md_unregister_thread(&cinfo->recv_thread); 860 lockres_free(cinfo->message_lockres); 861 lockres_free(cinfo->token_lockres); 862 lockres_free(cinfo->ack_lockres); 863 lockres_free(cinfo->no_new_dev_lockres); 864 lockres_free(cinfo->resync_lockres); 865 lockres_free(cinfo->bitmap_lockres); 866 unlock_all_bitmaps(mddev); 867 dlm_release_lockspace(cinfo->lockspace, 2); 868 return 0; 869 } 870 871 /* slot_number(): Returns the MD slot number to use 872 * DLM starts the slot numbers from 1, wheras cluster-md 873 * wants the number to be from zero, so we deduct one 874 */ 875 static int slot_number(struct mddev *mddev) 876 { 877 struct md_cluster_info *cinfo = mddev->cluster_info; 878 879 return cinfo->slot_number - 1; 880 } 881 882 /* 883 * Check if the communication is already locked, else lock the communication 884 * channel. 885 * If it is already locked, token is in EX mode, and hence lock_token() 886 * should not be called. 887 */ 888 static int metadata_update_start(struct mddev *mddev) 889 { 890 struct md_cluster_info *cinfo = mddev->cluster_info; 891 892 wait_event(cinfo->wait, 893 !test_and_set_bit(MD_CLUSTER_SEND_LOCK, &cinfo->state) || 894 test_and_clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state)); 895 896 /* If token is already locked, return 0 */ 897 if (cinfo->token_lockres->mode == DLM_LOCK_EX) 898 return 0; 899 900 return lock_token(cinfo); 901 } 902 903 static int metadata_update_finish(struct mddev *mddev) 904 { 905 struct md_cluster_info *cinfo = mddev->cluster_info; 906 struct cluster_msg cmsg; 907 struct md_rdev *rdev; 908 int ret = 0; 909 int raid_slot = -1; 910 911 memset(&cmsg, 0, sizeof(cmsg)); 912 cmsg.type = cpu_to_le32(METADATA_UPDATED); 913 /* Pick up a good active device number to send. 914 */ 915 rdev_for_each(rdev, mddev) 916 if (rdev->raid_disk > -1 && !test_bit(Faulty, &rdev->flags)) { 917 raid_slot = rdev->desc_nr; 918 break; 919 } 920 if (raid_slot >= 0) { 921 cmsg.raid_slot = cpu_to_le32(raid_slot); 922 ret = __sendmsg(cinfo, &cmsg); 923 } else 924 pr_warn("md-cluster: No good device id found to send\n"); 925 clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state); 926 unlock_comm(cinfo); 927 return ret; 928 } 929 930 static void metadata_update_cancel(struct mddev *mddev) 931 { 932 struct md_cluster_info *cinfo = mddev->cluster_info; 933 clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state); 934 unlock_comm(cinfo); 935 } 936 937 static int resync_start(struct mddev *mddev) 938 { 939 struct md_cluster_info *cinfo = mddev->cluster_info; 940 cinfo->resync_lockres->flags |= DLM_LKF_NOQUEUE; 941 return dlm_lock_sync(cinfo->resync_lockres, DLM_LOCK_EX); 942 } 943 944 static int resync_info_update(struct mddev *mddev, sector_t lo, sector_t hi) 945 { 946 struct md_cluster_info *cinfo = mddev->cluster_info; 947 struct resync_info ri; 948 struct cluster_msg cmsg = {0}; 949 950 /* do not send zero again, if we have sent before */ 951 if (hi == 0) { 952 memcpy(&ri, cinfo->bitmap_lockres->lksb.sb_lvbptr, sizeof(struct resync_info)); 953 if (le64_to_cpu(ri.hi) == 0) 954 return 0; 955 } 956 957 add_resync_info(cinfo->bitmap_lockres, lo, hi); 958 /* Re-acquire the lock to refresh LVB */ 959 dlm_lock_sync(cinfo->bitmap_lockres, DLM_LOCK_PW); 960 cmsg.type = cpu_to_le32(RESYNCING); 961 cmsg.low = cpu_to_le64(lo); 962 cmsg.high = cpu_to_le64(hi); 963 964 return sendmsg(cinfo, &cmsg); 965 } 966 967 static int resync_finish(struct mddev *mddev) 968 { 969 struct md_cluster_info *cinfo = mddev->cluster_info; 970 cinfo->resync_lockres->flags &= ~DLM_LKF_NOQUEUE; 971 dlm_unlock_sync(cinfo->resync_lockres); 972 return resync_info_update(mddev, 0, 0); 973 } 974 975 static int area_resyncing(struct mddev *mddev, int direction, 976 sector_t lo, sector_t hi) 977 { 978 struct md_cluster_info *cinfo = mddev->cluster_info; 979 int ret = 0; 980 struct suspend_info *s; 981 982 if ((direction == READ) && 983 test_bit(MD_CLUSTER_SUSPEND_READ_BALANCING, &cinfo->state)) 984 return 1; 985 986 spin_lock_irq(&cinfo->suspend_lock); 987 if (list_empty(&cinfo->suspend_list)) 988 goto out; 989 list_for_each_entry(s, &cinfo->suspend_list, list) 990 if (hi > s->lo && lo < s->hi) { 991 ret = 1; 992 break; 993 } 994 out: 995 spin_unlock_irq(&cinfo->suspend_lock); 996 return ret; 997 } 998 999 /* add_new_disk() - initiates a disk add 1000 * However, if this fails before writing md_update_sb(), 1001 * add_new_disk_cancel() must be called to release token lock 1002 */ 1003 static int add_new_disk(struct mddev *mddev, struct md_rdev *rdev) 1004 { 1005 struct md_cluster_info *cinfo = mddev->cluster_info; 1006 struct cluster_msg cmsg; 1007 int ret = 0; 1008 struct mdp_superblock_1 *sb = page_address(rdev->sb_page); 1009 char *uuid = sb->device_uuid; 1010 1011 memset(&cmsg, 0, sizeof(cmsg)); 1012 cmsg.type = cpu_to_le32(NEWDISK); 1013 memcpy(cmsg.uuid, uuid, 16); 1014 cmsg.raid_slot = cpu_to_le32(rdev->desc_nr); 1015 lock_comm(cinfo); 1016 ret = __sendmsg(cinfo, &cmsg); 1017 if (ret) 1018 return ret; 1019 cinfo->no_new_dev_lockres->flags |= DLM_LKF_NOQUEUE; 1020 ret = dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_EX); 1021 cinfo->no_new_dev_lockres->flags &= ~DLM_LKF_NOQUEUE; 1022 /* Some node does not "see" the device */ 1023 if (ret == -EAGAIN) 1024 ret = -ENOENT; 1025 if (ret) 1026 unlock_comm(cinfo); 1027 else { 1028 dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR); 1029 /* Since MD_CHANGE_DEVS will be set in add_bound_rdev which 1030 * will run soon after add_new_disk, the below path will be 1031 * invoked: 1032 * md_wakeup_thread(mddev->thread) 1033 * -> conf->thread (raid1d) 1034 * -> md_check_recovery -> md_update_sb 1035 * -> metadata_update_start/finish 1036 * MD_CLUSTER_SEND_LOCKED_ALREADY will be cleared eventually. 1037 * 1038 * For other failure cases, metadata_update_cancel and 1039 * add_new_disk_cancel also clear below bit as well. 1040 * */ 1041 set_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state); 1042 wake_up(&cinfo->wait); 1043 } 1044 return ret; 1045 } 1046 1047 static void add_new_disk_cancel(struct mddev *mddev) 1048 { 1049 struct md_cluster_info *cinfo = mddev->cluster_info; 1050 clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state); 1051 unlock_comm(cinfo); 1052 } 1053 1054 static int new_disk_ack(struct mddev *mddev, bool ack) 1055 { 1056 struct md_cluster_info *cinfo = mddev->cluster_info; 1057 1058 if (!test_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state)) { 1059 pr_warn("md-cluster(%s): Spurious cluster confirmation\n", mdname(mddev)); 1060 return -EINVAL; 1061 } 1062 1063 if (ack) 1064 dlm_unlock_sync(cinfo->no_new_dev_lockres); 1065 complete(&cinfo->newdisk_completion); 1066 return 0; 1067 } 1068 1069 static int remove_disk(struct mddev *mddev, struct md_rdev *rdev) 1070 { 1071 struct cluster_msg cmsg = {0}; 1072 struct md_cluster_info *cinfo = mddev->cluster_info; 1073 cmsg.type = cpu_to_le32(REMOVE); 1074 cmsg.raid_slot = cpu_to_le32(rdev->desc_nr); 1075 return sendmsg(cinfo, &cmsg); 1076 } 1077 1078 static int lock_all_bitmaps(struct mddev *mddev) 1079 { 1080 int slot, my_slot, ret, held = 1, i = 0; 1081 char str[64]; 1082 struct md_cluster_info *cinfo = mddev->cluster_info; 1083 1084 cinfo->other_bitmap_lockres = kzalloc((mddev->bitmap_info.nodes - 1) * 1085 sizeof(struct dlm_lock_resource *), 1086 GFP_KERNEL); 1087 if (!cinfo->other_bitmap_lockres) { 1088 pr_err("md: can't alloc mem for other bitmap locks\n"); 1089 return 0; 1090 } 1091 1092 my_slot = slot_number(mddev); 1093 for (slot = 0; slot < mddev->bitmap_info.nodes; slot++) { 1094 if (slot == my_slot) 1095 continue; 1096 1097 memset(str, '\0', 64); 1098 snprintf(str, 64, "bitmap%04d", slot); 1099 cinfo->other_bitmap_lockres[i] = lockres_init(mddev, str, NULL, 1); 1100 if (!cinfo->other_bitmap_lockres[i]) 1101 return -ENOMEM; 1102 1103 cinfo->other_bitmap_lockres[i]->flags |= DLM_LKF_NOQUEUE; 1104 ret = dlm_lock_sync(cinfo->other_bitmap_lockres[i], DLM_LOCK_PW); 1105 if (ret) 1106 held = -1; 1107 i++; 1108 } 1109 1110 return held; 1111 } 1112 1113 static void unlock_all_bitmaps(struct mddev *mddev) 1114 { 1115 struct md_cluster_info *cinfo = mddev->cluster_info; 1116 int i; 1117 1118 /* release other node's bitmap lock if they are existed */ 1119 if (cinfo->other_bitmap_lockres) { 1120 for (i = 0; i < mddev->bitmap_info.nodes - 1; i++) { 1121 if (cinfo->other_bitmap_lockres[i]) { 1122 dlm_unlock_sync(cinfo->other_bitmap_lockres[i]); 1123 lockres_free(cinfo->other_bitmap_lockres[i]); 1124 } 1125 } 1126 kfree(cinfo->other_bitmap_lockres); 1127 } 1128 } 1129 1130 static int gather_bitmaps(struct md_rdev *rdev) 1131 { 1132 int sn, err; 1133 sector_t lo, hi; 1134 struct cluster_msg cmsg = {0}; 1135 struct mddev *mddev = rdev->mddev; 1136 struct md_cluster_info *cinfo = mddev->cluster_info; 1137 1138 cmsg.type = cpu_to_le32(RE_ADD); 1139 cmsg.raid_slot = cpu_to_le32(rdev->desc_nr); 1140 err = sendmsg(cinfo, &cmsg); 1141 if (err) 1142 goto out; 1143 1144 for (sn = 0; sn < mddev->bitmap_info.nodes; sn++) { 1145 if (sn == (cinfo->slot_number - 1)) 1146 continue; 1147 err = bitmap_copy_from_slot(mddev, sn, &lo, &hi, false); 1148 if (err) { 1149 pr_warn("md-cluster: Could not gather bitmaps from slot %d", sn); 1150 goto out; 1151 } 1152 if ((hi > 0) && (lo < mddev->recovery_cp)) 1153 mddev->recovery_cp = lo; 1154 } 1155 out: 1156 return err; 1157 } 1158 1159 static struct md_cluster_operations cluster_ops = { 1160 .join = join, 1161 .leave = leave, 1162 .slot_number = slot_number, 1163 .resync_start = resync_start, 1164 .resync_finish = resync_finish, 1165 .resync_info_update = resync_info_update, 1166 .metadata_update_start = metadata_update_start, 1167 .metadata_update_finish = metadata_update_finish, 1168 .metadata_update_cancel = metadata_update_cancel, 1169 .area_resyncing = area_resyncing, 1170 .add_new_disk = add_new_disk, 1171 .add_new_disk_cancel = add_new_disk_cancel, 1172 .new_disk_ack = new_disk_ack, 1173 .remove_disk = remove_disk, 1174 .gather_bitmaps = gather_bitmaps, 1175 .lock_all_bitmaps = lock_all_bitmaps, 1176 .unlock_all_bitmaps = unlock_all_bitmaps, 1177 }; 1178 1179 static int __init cluster_init(void) 1180 { 1181 pr_warn("md-cluster: EXPERIMENTAL. Use with caution\n"); 1182 pr_info("Registering Cluster MD functions\n"); 1183 register_md_cluster_operations(&cluster_ops, THIS_MODULE); 1184 return 0; 1185 } 1186 1187 static void cluster_exit(void) 1188 { 1189 unregister_md_cluster_operations(); 1190 } 1191 1192 module_init(cluster_init); 1193 module_exit(cluster_exit); 1194 MODULE_AUTHOR("SUSE"); 1195 MODULE_LICENSE("GPL"); 1196 MODULE_DESCRIPTION("Clustering support for MD"); 1197