1 /****************************************************************************** 2 ******************************************************************************* 3 ** 4 ** Copyright (C) 2005-2011 Red Hat, Inc. All rights reserved. 5 ** 6 ** This copyrighted material is made available to anyone wishing to use, 7 ** modify, copy, or redistribute it subject to the terms and conditions 8 ** of the GNU General Public License v.2. 9 ** 10 ******************************************************************************* 11 ******************************************************************************/ 12 13 #include "dlm_internal.h" 14 #include "lockspace.h" 15 #include "member.h" 16 #include "recoverd.h" 17 #include "recover.h" 18 #include "rcom.h" 19 #include "config.h" 20 #include "lowcomms.h" 21 22 int dlm_slots_version(struct dlm_header *h) 23 { 24 if ((h->h_version & 0x0000FFFF) < DLM_HEADER_SLOTS) 25 return 0; 26 return 1; 27 } 28 29 void dlm_slot_save(struct dlm_ls *ls, struct dlm_rcom *rc, 30 struct dlm_member *memb) 31 { 32 struct rcom_config *rf = (struct rcom_config *)rc->rc_buf; 33 34 if (!dlm_slots_version(&rc->rc_header)) 35 return; 36 37 memb->slot = le16_to_cpu(rf->rf_our_slot); 38 memb->generation = le32_to_cpu(rf->rf_generation); 39 } 40 41 void dlm_slots_copy_out(struct dlm_ls *ls, struct dlm_rcom *rc) 42 { 43 struct dlm_slot *slot; 44 struct rcom_slot *ro; 45 int i; 46 47 ro = (struct rcom_slot *)(rc->rc_buf + sizeof(struct rcom_config)); 48 49 /* ls_slots array is sparse, but not rcom_slots */ 50 51 for (i = 0; i < ls->ls_slots_size; i++) { 52 slot = &ls->ls_slots[i]; 53 if (!slot->nodeid) 54 continue; 55 ro->ro_nodeid = cpu_to_le32(slot->nodeid); 56 ro->ro_slot = cpu_to_le16(slot->slot); 57 ro++; 58 } 59 } 60 61 #define SLOT_DEBUG_LINE 128 62 63 static void log_slots(struct dlm_ls *ls, uint32_t gen, int num_slots, 64 struct rcom_slot *ro0, struct dlm_slot *array, 65 int array_size) 66 { 67 char line[SLOT_DEBUG_LINE]; 68 int len = SLOT_DEBUG_LINE - 1; 69 int pos = 0; 70 int ret, i; 71 72 memset(line, 0, sizeof(line)); 73 74 if (array) { 75 for (i = 0; i < array_size; i++) { 76 if (!array[i].nodeid) 77 continue; 78 79 ret = snprintf(line + pos, len - pos, " %d:%d", 80 array[i].slot, array[i].nodeid); 81 if (ret >= len - pos) 82 break; 83 pos += ret; 84 } 85 } else if (ro0) { 86 for (i = 0; i < num_slots; i++) { 87 ret = snprintf(line + pos, len - pos, " %d:%d", 88 ro0[i].ro_slot, ro0[i].ro_nodeid); 89 if (ret >= len - pos) 90 break; 91 pos += ret; 92 } 93 } 94 95 log_rinfo(ls, "generation %u slots %d%s", gen, num_slots, line); 96 } 97 98 int dlm_slots_copy_in(struct dlm_ls *ls) 99 { 100 struct dlm_member *memb; 101 struct dlm_rcom *rc = ls->ls_recover_buf; 102 struct rcom_config *rf = (struct rcom_config *)rc->rc_buf; 103 struct rcom_slot *ro0, *ro; 104 int our_nodeid = dlm_our_nodeid(); 105 int i, num_slots; 106 uint32_t gen; 107 108 if (!dlm_slots_version(&rc->rc_header)) 109 return -1; 110 111 gen = le32_to_cpu(rf->rf_generation); 112 if (gen <= ls->ls_generation) { 113 log_error(ls, "dlm_slots_copy_in gen %u old %u", 114 gen, ls->ls_generation); 115 } 116 ls->ls_generation = gen; 117 118 num_slots = le16_to_cpu(rf->rf_num_slots); 119 if (!num_slots) 120 return -1; 121 122 ro0 = (struct rcom_slot *)(rc->rc_buf + sizeof(struct rcom_config)); 123 124 for (i = 0, ro = ro0; i < num_slots; i++, ro++) { 125 ro->ro_nodeid = le32_to_cpu(ro->ro_nodeid); 126 ro->ro_slot = le16_to_cpu(ro->ro_slot); 127 } 128 129 log_slots(ls, gen, num_slots, ro0, NULL, 0); 130 131 list_for_each_entry(memb, &ls->ls_nodes, list) { 132 for (i = 0, ro = ro0; i < num_slots; i++, ro++) { 133 if (ro->ro_nodeid != memb->nodeid) 134 continue; 135 memb->slot = ro->ro_slot; 136 memb->slot_prev = memb->slot; 137 break; 138 } 139 140 if (memb->nodeid == our_nodeid) { 141 if (ls->ls_slot && ls->ls_slot != memb->slot) { 142 log_error(ls, "dlm_slots_copy_in our slot " 143 "changed %d %d", ls->ls_slot, 144 memb->slot); 145 return -1; 146 } 147 148 if (!ls->ls_slot) 149 ls->ls_slot = memb->slot; 150 } 151 152 if (!memb->slot) { 153 log_error(ls, "dlm_slots_copy_in nodeid %d no slot", 154 memb->nodeid); 155 return -1; 156 } 157 } 158 159 return 0; 160 } 161 162 /* for any nodes that do not support slots, we will not have set memb->slot 163 in wait_status_all(), so memb->slot will remain -1, and we will not 164 assign slots or set ls_num_slots here */ 165 166 int dlm_slots_assign(struct dlm_ls *ls, int *num_slots, int *slots_size, 167 struct dlm_slot **slots_out, uint32_t *gen_out) 168 { 169 struct dlm_member *memb; 170 struct dlm_slot *array; 171 int our_nodeid = dlm_our_nodeid(); 172 int array_size, max_slots, i; 173 int need = 0; 174 int max = 0; 175 int num = 0; 176 uint32_t gen = 0; 177 178 /* our own memb struct will have slot -1 gen 0 */ 179 180 list_for_each_entry(memb, &ls->ls_nodes, list) { 181 if (memb->nodeid == our_nodeid) { 182 memb->slot = ls->ls_slot; 183 memb->generation = ls->ls_generation; 184 break; 185 } 186 } 187 188 list_for_each_entry(memb, &ls->ls_nodes, list) { 189 if (memb->generation > gen) 190 gen = memb->generation; 191 192 /* node doesn't support slots */ 193 194 if (memb->slot == -1) 195 return -1; 196 197 /* node needs a slot assigned */ 198 199 if (!memb->slot) 200 need++; 201 202 /* node has a slot assigned */ 203 204 num++; 205 206 if (!max || max < memb->slot) 207 max = memb->slot; 208 209 /* sanity check, once slot is assigned it shouldn't change */ 210 211 if (memb->slot_prev && memb->slot && memb->slot_prev != memb->slot) { 212 log_error(ls, "nodeid %d slot changed %d %d", 213 memb->nodeid, memb->slot_prev, memb->slot); 214 return -1; 215 } 216 memb->slot_prev = memb->slot; 217 } 218 219 array_size = max + need; 220 array = kcalloc(array_size, sizeof(*array), GFP_NOFS); 221 if (!array) 222 return -ENOMEM; 223 224 num = 0; 225 226 /* fill in slots (offsets) that are used */ 227 228 list_for_each_entry(memb, &ls->ls_nodes, list) { 229 if (!memb->slot) 230 continue; 231 232 if (memb->slot > array_size) { 233 log_error(ls, "invalid slot number %d", memb->slot); 234 kfree(array); 235 return -1; 236 } 237 238 array[memb->slot - 1].nodeid = memb->nodeid; 239 array[memb->slot - 1].slot = memb->slot; 240 num++; 241 } 242 243 /* assign new slots from unused offsets */ 244 245 list_for_each_entry(memb, &ls->ls_nodes, list) { 246 if (memb->slot) 247 continue; 248 249 for (i = 0; i < array_size; i++) { 250 if (array[i].nodeid) 251 continue; 252 253 memb->slot = i + 1; 254 memb->slot_prev = memb->slot; 255 array[i].nodeid = memb->nodeid; 256 array[i].slot = memb->slot; 257 num++; 258 259 if (!ls->ls_slot && memb->nodeid == our_nodeid) 260 ls->ls_slot = memb->slot; 261 break; 262 } 263 264 if (!memb->slot) { 265 log_error(ls, "no free slot found"); 266 kfree(array); 267 return -1; 268 } 269 } 270 271 gen++; 272 273 log_slots(ls, gen, num, NULL, array, array_size); 274 275 max_slots = (dlm_config.ci_buffer_size - sizeof(struct dlm_rcom) - 276 sizeof(struct rcom_config)) / sizeof(struct rcom_slot); 277 278 if (num > max_slots) { 279 log_error(ls, "num_slots %d exceeds max_slots %d", 280 num, max_slots); 281 kfree(array); 282 return -1; 283 } 284 285 *gen_out = gen; 286 *slots_out = array; 287 *slots_size = array_size; 288 *num_slots = num; 289 return 0; 290 } 291 292 static void add_ordered_member(struct dlm_ls *ls, struct dlm_member *new) 293 { 294 struct dlm_member *memb = NULL; 295 struct list_head *tmp; 296 struct list_head *newlist = &new->list; 297 struct list_head *head = &ls->ls_nodes; 298 299 list_for_each(tmp, head) { 300 memb = list_entry(tmp, struct dlm_member, list); 301 if (new->nodeid < memb->nodeid) 302 break; 303 } 304 305 if (!memb) 306 list_add_tail(newlist, head); 307 else { 308 /* FIXME: can use list macro here */ 309 newlist->prev = tmp->prev; 310 newlist->next = tmp; 311 tmp->prev->next = newlist; 312 tmp->prev = newlist; 313 } 314 } 315 316 static int dlm_add_member(struct dlm_ls *ls, struct dlm_config_node *node) 317 { 318 struct dlm_member *memb; 319 int error; 320 321 memb = kzalloc(sizeof(*memb), GFP_NOFS); 322 if (!memb) 323 return -ENOMEM; 324 325 error = dlm_lowcomms_connect_node(node->nodeid); 326 if (error < 0) { 327 kfree(memb); 328 return error; 329 } 330 331 memb->nodeid = node->nodeid; 332 memb->weight = node->weight; 333 memb->comm_seq = node->comm_seq; 334 add_ordered_member(ls, memb); 335 ls->ls_num_nodes++; 336 return 0; 337 } 338 339 static struct dlm_member *find_memb(struct list_head *head, int nodeid) 340 { 341 struct dlm_member *memb; 342 343 list_for_each_entry(memb, head, list) { 344 if (memb->nodeid == nodeid) 345 return memb; 346 } 347 return NULL; 348 } 349 350 int dlm_is_member(struct dlm_ls *ls, int nodeid) 351 { 352 if (find_memb(&ls->ls_nodes, nodeid)) 353 return 1; 354 return 0; 355 } 356 357 int dlm_is_removed(struct dlm_ls *ls, int nodeid) 358 { 359 if (find_memb(&ls->ls_nodes_gone, nodeid)) 360 return 1; 361 return 0; 362 } 363 364 static void clear_memb_list(struct list_head *head) 365 { 366 struct dlm_member *memb; 367 368 while (!list_empty(head)) { 369 memb = list_entry(head->next, struct dlm_member, list); 370 list_del(&memb->list); 371 kfree(memb); 372 } 373 } 374 375 void dlm_clear_members(struct dlm_ls *ls) 376 { 377 clear_memb_list(&ls->ls_nodes); 378 ls->ls_num_nodes = 0; 379 } 380 381 void dlm_clear_members_gone(struct dlm_ls *ls) 382 { 383 clear_memb_list(&ls->ls_nodes_gone); 384 } 385 386 static void make_member_array(struct dlm_ls *ls) 387 { 388 struct dlm_member *memb; 389 int i, w, x = 0, total = 0, all_zero = 0, *array; 390 391 kfree(ls->ls_node_array); 392 ls->ls_node_array = NULL; 393 394 list_for_each_entry(memb, &ls->ls_nodes, list) { 395 if (memb->weight) 396 total += memb->weight; 397 } 398 399 /* all nodes revert to weight of 1 if all have weight 0 */ 400 401 if (!total) { 402 total = ls->ls_num_nodes; 403 all_zero = 1; 404 } 405 406 ls->ls_total_weight = total; 407 array = kmalloc_array(total, sizeof(*array), GFP_NOFS); 408 if (!array) 409 return; 410 411 list_for_each_entry(memb, &ls->ls_nodes, list) { 412 if (!all_zero && !memb->weight) 413 continue; 414 415 if (all_zero) 416 w = 1; 417 else 418 w = memb->weight; 419 420 DLM_ASSERT(x < total, printk("total %d x %d\n", total, x);); 421 422 for (i = 0; i < w; i++) 423 array[x++] = memb->nodeid; 424 } 425 426 ls->ls_node_array = array; 427 } 428 429 /* send a status request to all members just to establish comms connections */ 430 431 static int ping_members(struct dlm_ls *ls) 432 { 433 struct dlm_member *memb; 434 int error = 0; 435 436 list_for_each_entry(memb, &ls->ls_nodes, list) { 437 error = dlm_recovery_stopped(ls); 438 if (error) 439 break; 440 error = dlm_rcom_status(ls, memb->nodeid, 0); 441 if (error) 442 break; 443 } 444 if (error) 445 log_rinfo(ls, "ping_members aborted %d last nodeid %d", 446 error, ls->ls_recover_nodeid); 447 return error; 448 } 449 450 static void dlm_lsop_recover_prep(struct dlm_ls *ls) 451 { 452 if (!ls->ls_ops || !ls->ls_ops->recover_prep) 453 return; 454 ls->ls_ops->recover_prep(ls->ls_ops_arg); 455 } 456 457 static void dlm_lsop_recover_slot(struct dlm_ls *ls, struct dlm_member *memb) 458 { 459 struct dlm_slot slot; 460 uint32_t seq; 461 int error; 462 463 if (!ls->ls_ops || !ls->ls_ops->recover_slot) 464 return; 465 466 /* if there is no comms connection with this node 467 or the present comms connection is newer 468 than the one when this member was added, then 469 we consider the node to have failed (versus 470 being removed due to dlm_release_lockspace) */ 471 472 error = dlm_comm_seq(memb->nodeid, &seq); 473 474 if (!error && seq == memb->comm_seq) 475 return; 476 477 slot.nodeid = memb->nodeid; 478 slot.slot = memb->slot; 479 480 ls->ls_ops->recover_slot(ls->ls_ops_arg, &slot); 481 } 482 483 void dlm_lsop_recover_done(struct dlm_ls *ls) 484 { 485 struct dlm_member *memb; 486 struct dlm_slot *slots; 487 int i, num; 488 489 if (!ls->ls_ops || !ls->ls_ops->recover_done) 490 return; 491 492 num = ls->ls_num_nodes; 493 slots = kcalloc(num, sizeof(*slots), GFP_KERNEL); 494 if (!slots) 495 return; 496 497 i = 0; 498 list_for_each_entry(memb, &ls->ls_nodes, list) { 499 if (i == num) { 500 log_error(ls, "dlm_lsop_recover_done bad num %d", num); 501 goto out; 502 } 503 slots[i].nodeid = memb->nodeid; 504 slots[i].slot = memb->slot; 505 i++; 506 } 507 508 ls->ls_ops->recover_done(ls->ls_ops_arg, slots, num, 509 ls->ls_slot, ls->ls_generation); 510 out: 511 kfree(slots); 512 } 513 514 static struct dlm_config_node *find_config_node(struct dlm_recover *rv, 515 int nodeid) 516 { 517 int i; 518 519 for (i = 0; i < rv->nodes_count; i++) { 520 if (rv->nodes[i].nodeid == nodeid) 521 return &rv->nodes[i]; 522 } 523 return NULL; 524 } 525 526 int dlm_recover_members(struct dlm_ls *ls, struct dlm_recover *rv, int *neg_out) 527 { 528 struct dlm_member *memb, *safe; 529 struct dlm_config_node *node; 530 int i, error, neg = 0, low = -1; 531 532 /* previously removed members that we've not finished removing need to 533 count as a negative change so the "neg" recovery steps will happen */ 534 535 list_for_each_entry(memb, &ls->ls_nodes_gone, list) { 536 log_rinfo(ls, "prev removed member %d", memb->nodeid); 537 neg++; 538 } 539 540 /* move departed members from ls_nodes to ls_nodes_gone */ 541 542 list_for_each_entry_safe(memb, safe, &ls->ls_nodes, list) { 543 node = find_config_node(rv, memb->nodeid); 544 if (node && !node->new) 545 continue; 546 547 if (!node) { 548 log_rinfo(ls, "remove member %d", memb->nodeid); 549 } else { 550 /* removed and re-added */ 551 log_rinfo(ls, "remove member %d comm_seq %u %u", 552 memb->nodeid, memb->comm_seq, node->comm_seq); 553 } 554 555 neg++; 556 list_move(&memb->list, &ls->ls_nodes_gone); 557 ls->ls_num_nodes--; 558 dlm_lsop_recover_slot(ls, memb); 559 } 560 561 /* add new members to ls_nodes */ 562 563 for (i = 0; i < rv->nodes_count; i++) { 564 node = &rv->nodes[i]; 565 if (dlm_is_member(ls, node->nodeid)) 566 continue; 567 dlm_add_member(ls, node); 568 log_rinfo(ls, "add member %d", node->nodeid); 569 } 570 571 list_for_each_entry(memb, &ls->ls_nodes, list) { 572 if (low == -1 || memb->nodeid < low) 573 low = memb->nodeid; 574 } 575 ls->ls_low_nodeid = low; 576 577 make_member_array(ls); 578 *neg_out = neg; 579 580 error = ping_members(ls); 581 if (!error || error == -EPROTO) { 582 /* new_lockspace() may be waiting to know if the config 583 is good or bad */ 584 ls->ls_members_result = error; 585 complete(&ls->ls_members_done); 586 } 587 588 log_rinfo(ls, "dlm_recover_members %d nodes", ls->ls_num_nodes); 589 return error; 590 } 591 592 /* Userspace guarantees that dlm_ls_stop() has completed on all nodes before 593 dlm_ls_start() is called on any of them to start the new recovery. */ 594 595 int dlm_ls_stop(struct dlm_ls *ls) 596 { 597 int new; 598 599 /* 600 * Prevent dlm_recv from being in the middle of something when we do 601 * the stop. This includes ensuring dlm_recv isn't processing a 602 * recovery message (rcom), while dlm_recoverd is aborting and 603 * resetting things from an in-progress recovery. i.e. we want 604 * dlm_recoverd to abort its recovery without worrying about dlm_recv 605 * processing an rcom at the same time. Stopping dlm_recv also makes 606 * it easy for dlm_receive_message() to check locking stopped and add a 607 * message to the requestqueue without races. 608 */ 609 610 down_write(&ls->ls_recv_active); 611 612 /* 613 * Abort any recovery that's in progress (see RECOVER_STOP, 614 * dlm_recovery_stopped()) and tell any other threads running in the 615 * dlm to quit any processing (see RUNNING, dlm_locking_stopped()). 616 */ 617 618 spin_lock(&ls->ls_recover_lock); 619 set_bit(LSFL_RECOVER_STOP, &ls->ls_flags); 620 new = test_and_clear_bit(LSFL_RUNNING, &ls->ls_flags); 621 ls->ls_recover_seq++; 622 spin_unlock(&ls->ls_recover_lock); 623 624 /* 625 * Let dlm_recv run again, now any normal messages will be saved on the 626 * requestqueue for later. 627 */ 628 629 up_write(&ls->ls_recv_active); 630 631 /* 632 * This in_recovery lock does two things: 633 * 1) Keeps this function from returning until all threads are out 634 * of locking routines and locking is truly stopped. 635 * 2) Keeps any new requests from being processed until it's unlocked 636 * when recovery is complete. 637 */ 638 639 if (new) { 640 set_bit(LSFL_RECOVER_DOWN, &ls->ls_flags); 641 wake_up_process(ls->ls_recoverd_task); 642 wait_event(ls->ls_recover_lock_wait, 643 test_bit(LSFL_RECOVER_LOCK, &ls->ls_flags)); 644 } 645 646 /* 647 * The recoverd suspend/resume makes sure that dlm_recoverd (if 648 * running) has noticed RECOVER_STOP above and quit processing the 649 * previous recovery. 650 */ 651 652 dlm_recoverd_suspend(ls); 653 654 spin_lock(&ls->ls_recover_lock); 655 kfree(ls->ls_slots); 656 ls->ls_slots = NULL; 657 ls->ls_num_slots = 0; 658 ls->ls_slots_size = 0; 659 ls->ls_recover_status = 0; 660 spin_unlock(&ls->ls_recover_lock); 661 662 dlm_recoverd_resume(ls); 663 664 if (!ls->ls_recover_begin) 665 ls->ls_recover_begin = jiffies; 666 667 dlm_lsop_recover_prep(ls); 668 return 0; 669 } 670 671 int dlm_ls_start(struct dlm_ls *ls) 672 { 673 struct dlm_recover *rv, *rv_old; 674 struct dlm_config_node *nodes = NULL; 675 int error, count; 676 677 rv = kzalloc(sizeof(*rv), GFP_NOFS); 678 if (!rv) 679 return -ENOMEM; 680 681 error = dlm_config_nodes(ls->ls_name, &nodes, &count); 682 if (error < 0) 683 goto fail_rv; 684 685 spin_lock(&ls->ls_recover_lock); 686 687 /* the lockspace needs to be stopped before it can be started */ 688 689 if (!dlm_locking_stopped(ls)) { 690 spin_unlock(&ls->ls_recover_lock); 691 log_error(ls, "start ignored: lockspace running"); 692 error = -EINVAL; 693 goto fail; 694 } 695 696 rv->nodes = nodes; 697 rv->nodes_count = count; 698 rv->seq = ++ls->ls_recover_seq; 699 rv_old = ls->ls_recover_args; 700 ls->ls_recover_args = rv; 701 spin_unlock(&ls->ls_recover_lock); 702 703 if (rv_old) { 704 log_error(ls, "unused recovery %llx %d", 705 (unsigned long long)rv_old->seq, rv_old->nodes_count); 706 kfree(rv_old->nodes); 707 kfree(rv_old); 708 } 709 710 set_bit(LSFL_RECOVER_WORK, &ls->ls_flags); 711 wake_up_process(ls->ls_recoverd_task); 712 return 0; 713 714 fail: 715 kfree(nodes); 716 fail_rv: 717 kfree(rv); 718 return error; 719 } 720 721