1 // SPDX-License-Identifier: GPL-2.0-only 2 /****************************************************************************** 3 ******************************************************************************* 4 ** 5 ** Copyright (C) 2005-2011 Red Hat, Inc. All rights reserved. 6 ** 7 ** 8 ******************************************************************************* 9 ******************************************************************************/ 10 11 #include "dlm_internal.h" 12 #include "lockspace.h" 13 #include "member.h" 14 #include "recoverd.h" 15 #include "recover.h" 16 #include "rcom.h" 17 #include "config.h" 18 #include "lowcomms.h" 19 20 int dlm_slots_version(struct dlm_header *h) 21 { 22 if ((h->h_version & 0x0000FFFF) < DLM_HEADER_SLOTS) 23 return 0; 24 return 1; 25 } 26 27 void dlm_slot_save(struct dlm_ls *ls, struct dlm_rcom *rc, 28 struct dlm_member *memb) 29 { 30 struct rcom_config *rf = (struct rcom_config *)rc->rc_buf; 31 32 if (!dlm_slots_version(&rc->rc_header)) 33 return; 34 35 memb->slot = le16_to_cpu(rf->rf_our_slot); 36 memb->generation = le32_to_cpu(rf->rf_generation); 37 } 38 39 void dlm_slots_copy_out(struct dlm_ls *ls, struct dlm_rcom *rc) 40 { 41 struct dlm_slot *slot; 42 struct rcom_slot *ro; 43 int i; 44 45 ro = (struct rcom_slot *)(rc->rc_buf + sizeof(struct rcom_config)); 46 47 /* ls_slots array is sparse, but not rcom_slots */ 48 49 for (i = 0; i < ls->ls_slots_size; i++) { 50 slot = &ls->ls_slots[i]; 51 if (!slot->nodeid) 52 continue; 53 ro->ro_nodeid = cpu_to_le32(slot->nodeid); 54 ro->ro_slot = cpu_to_le16(slot->slot); 55 ro++; 56 } 57 } 58 59 #define SLOT_DEBUG_LINE 128 60 61 static void log_slots(struct dlm_ls *ls, uint32_t gen, int num_slots, 62 struct rcom_slot *ro0, struct dlm_slot *array, 63 int array_size) 64 { 65 char line[SLOT_DEBUG_LINE]; 66 int len = SLOT_DEBUG_LINE - 1; 67 int pos = 0; 68 int ret, i; 69 70 memset(line, 0, sizeof(line)); 71 72 if (array) { 73 for (i = 0; i < array_size; i++) { 74 if (!array[i].nodeid) 75 continue; 76 77 ret = snprintf(line + pos, len - pos, " %d:%d", 78 array[i].slot, array[i].nodeid); 79 if (ret >= len - pos) 80 break; 81 pos += ret; 82 } 83 } else if (ro0) { 84 for (i = 0; i < num_slots; i++) { 85 ret = snprintf(line + pos, len - pos, " %d:%d", 86 ro0[i].ro_slot, ro0[i].ro_nodeid); 87 if (ret >= len - pos) 88 break; 89 pos += ret; 90 } 91 } 92 93 log_rinfo(ls, "generation %u slots %d%s", gen, num_slots, line); 94 } 95 96 int dlm_slots_copy_in(struct dlm_ls *ls) 97 { 98 struct dlm_member *memb; 99 struct dlm_rcom *rc = ls->ls_recover_buf; 100 struct rcom_config *rf = (struct rcom_config *)rc->rc_buf; 101 struct rcom_slot *ro0, *ro; 102 int our_nodeid = dlm_our_nodeid(); 103 int i, num_slots; 104 uint32_t gen; 105 106 if (!dlm_slots_version(&rc->rc_header)) 107 return -1; 108 109 gen = le32_to_cpu(rf->rf_generation); 110 if (gen <= ls->ls_generation) { 111 log_error(ls, "dlm_slots_copy_in gen %u old %u", 112 gen, ls->ls_generation); 113 } 114 ls->ls_generation = gen; 115 116 num_slots = le16_to_cpu(rf->rf_num_slots); 117 if (!num_slots) 118 return -1; 119 120 ro0 = (struct rcom_slot *)(rc->rc_buf + sizeof(struct rcom_config)); 121 122 for (i = 0, ro = ro0; i < num_slots; i++, ro++) { 123 ro->ro_nodeid = le32_to_cpu(ro->ro_nodeid); 124 ro->ro_slot = le16_to_cpu(ro->ro_slot); 125 } 126 127 log_slots(ls, gen, num_slots, ro0, NULL, 0); 128 129 list_for_each_entry(memb, &ls->ls_nodes, list) { 130 for (i = 0, ro = ro0; i < num_slots; i++, ro++) { 131 if (ro->ro_nodeid != memb->nodeid) 132 continue; 133 memb->slot = ro->ro_slot; 134 memb->slot_prev = memb->slot; 135 break; 136 } 137 138 if (memb->nodeid == our_nodeid) { 139 if (ls->ls_slot && ls->ls_slot != memb->slot) { 140 log_error(ls, "dlm_slots_copy_in our slot " 141 "changed %d %d", ls->ls_slot, 142 memb->slot); 143 return -1; 144 } 145 146 if (!ls->ls_slot) 147 ls->ls_slot = memb->slot; 148 } 149 150 if (!memb->slot) { 151 log_error(ls, "dlm_slots_copy_in nodeid %d no slot", 152 memb->nodeid); 153 return -1; 154 } 155 } 156 157 return 0; 158 } 159 160 /* for any nodes that do not support slots, we will not have set memb->slot 161 in wait_status_all(), so memb->slot will remain -1, and we will not 162 assign slots or set ls_num_slots here */ 163 164 int dlm_slots_assign(struct dlm_ls *ls, int *num_slots, int *slots_size, 165 struct dlm_slot **slots_out, uint32_t *gen_out) 166 { 167 struct dlm_member *memb; 168 struct dlm_slot *array; 169 int our_nodeid = dlm_our_nodeid(); 170 int array_size, max_slots, i; 171 int need = 0; 172 int max = 0; 173 int num = 0; 174 uint32_t gen = 0; 175 176 /* our own memb struct will have slot -1 gen 0 */ 177 178 list_for_each_entry(memb, &ls->ls_nodes, list) { 179 if (memb->nodeid == our_nodeid) { 180 memb->slot = ls->ls_slot; 181 memb->generation = ls->ls_generation; 182 break; 183 } 184 } 185 186 list_for_each_entry(memb, &ls->ls_nodes, list) { 187 if (memb->generation > gen) 188 gen = memb->generation; 189 190 /* node doesn't support slots */ 191 192 if (memb->slot == -1) 193 return -1; 194 195 /* node needs a slot assigned */ 196 197 if (!memb->slot) 198 need++; 199 200 /* node has a slot assigned */ 201 202 num++; 203 204 if (!max || max < memb->slot) 205 max = memb->slot; 206 207 /* sanity check, once slot is assigned it shouldn't change */ 208 209 if (memb->slot_prev && memb->slot && memb->slot_prev != memb->slot) { 210 log_error(ls, "nodeid %d slot changed %d %d", 211 memb->nodeid, memb->slot_prev, memb->slot); 212 return -1; 213 } 214 memb->slot_prev = memb->slot; 215 } 216 217 array_size = max + need; 218 array = kcalloc(array_size, sizeof(*array), GFP_NOFS); 219 if (!array) 220 return -ENOMEM; 221 222 num = 0; 223 224 /* fill in slots (offsets) that are used */ 225 226 list_for_each_entry(memb, &ls->ls_nodes, list) { 227 if (!memb->slot) 228 continue; 229 230 if (memb->slot > array_size) { 231 log_error(ls, "invalid slot number %d", memb->slot); 232 kfree(array); 233 return -1; 234 } 235 236 array[memb->slot - 1].nodeid = memb->nodeid; 237 array[memb->slot - 1].slot = memb->slot; 238 num++; 239 } 240 241 /* assign new slots from unused offsets */ 242 243 list_for_each_entry(memb, &ls->ls_nodes, list) { 244 if (memb->slot) 245 continue; 246 247 for (i = 0; i < array_size; i++) { 248 if (array[i].nodeid) 249 continue; 250 251 memb->slot = i + 1; 252 memb->slot_prev = memb->slot; 253 array[i].nodeid = memb->nodeid; 254 array[i].slot = memb->slot; 255 num++; 256 257 if (!ls->ls_slot && memb->nodeid == our_nodeid) 258 ls->ls_slot = memb->slot; 259 break; 260 } 261 262 if (!memb->slot) { 263 log_error(ls, "no free slot found"); 264 kfree(array); 265 return -1; 266 } 267 } 268 269 gen++; 270 271 log_slots(ls, gen, num, NULL, array, array_size); 272 273 max_slots = (dlm_config.ci_buffer_size - sizeof(struct dlm_rcom) - 274 sizeof(struct rcom_config)) / sizeof(struct rcom_slot); 275 276 if (num > max_slots) { 277 log_error(ls, "num_slots %d exceeds max_slots %d", 278 num, max_slots); 279 kfree(array); 280 return -1; 281 } 282 283 *gen_out = gen; 284 *slots_out = array; 285 *slots_size = array_size; 286 *num_slots = num; 287 return 0; 288 } 289 290 static void add_ordered_member(struct dlm_ls *ls, struct dlm_member *new) 291 { 292 struct dlm_member *memb = NULL; 293 struct list_head *tmp; 294 struct list_head *newlist = &new->list; 295 struct list_head *head = &ls->ls_nodes; 296 297 list_for_each(tmp, head) { 298 memb = list_entry(tmp, struct dlm_member, list); 299 if (new->nodeid < memb->nodeid) 300 break; 301 } 302 303 if (!memb) 304 list_add_tail(newlist, head); 305 else { 306 /* FIXME: can use list macro here */ 307 newlist->prev = tmp->prev; 308 newlist->next = tmp; 309 tmp->prev->next = newlist; 310 tmp->prev = newlist; 311 } 312 } 313 314 static int dlm_add_member(struct dlm_ls *ls, struct dlm_config_node *node) 315 { 316 struct dlm_member *memb; 317 int error; 318 319 memb = kzalloc(sizeof(*memb), GFP_NOFS); 320 if (!memb) 321 return -ENOMEM; 322 323 error = dlm_lowcomms_connect_node(node->nodeid); 324 if (error < 0) { 325 kfree(memb); 326 return error; 327 } 328 329 memb->nodeid = node->nodeid; 330 memb->weight = node->weight; 331 memb->comm_seq = node->comm_seq; 332 add_ordered_member(ls, memb); 333 ls->ls_num_nodes++; 334 return 0; 335 } 336 337 static struct dlm_member *find_memb(struct list_head *head, int nodeid) 338 { 339 struct dlm_member *memb; 340 341 list_for_each_entry(memb, head, list) { 342 if (memb->nodeid == nodeid) 343 return memb; 344 } 345 return NULL; 346 } 347 348 int dlm_is_member(struct dlm_ls *ls, int nodeid) 349 { 350 if (find_memb(&ls->ls_nodes, nodeid)) 351 return 1; 352 return 0; 353 } 354 355 int dlm_is_removed(struct dlm_ls *ls, int nodeid) 356 { 357 if (find_memb(&ls->ls_nodes_gone, nodeid)) 358 return 1; 359 return 0; 360 } 361 362 static void clear_memb_list(struct list_head *head) 363 { 364 struct dlm_member *memb; 365 366 while (!list_empty(head)) { 367 memb = list_entry(head->next, struct dlm_member, list); 368 list_del(&memb->list); 369 kfree(memb); 370 } 371 } 372 373 void dlm_clear_members(struct dlm_ls *ls) 374 { 375 clear_memb_list(&ls->ls_nodes); 376 ls->ls_num_nodes = 0; 377 } 378 379 void dlm_clear_members_gone(struct dlm_ls *ls) 380 { 381 clear_memb_list(&ls->ls_nodes_gone); 382 } 383 384 static void make_member_array(struct dlm_ls *ls) 385 { 386 struct dlm_member *memb; 387 int i, w, x = 0, total = 0, all_zero = 0, *array; 388 389 kfree(ls->ls_node_array); 390 ls->ls_node_array = NULL; 391 392 list_for_each_entry(memb, &ls->ls_nodes, list) { 393 if (memb->weight) 394 total += memb->weight; 395 } 396 397 /* all nodes revert to weight of 1 if all have weight 0 */ 398 399 if (!total) { 400 total = ls->ls_num_nodes; 401 all_zero = 1; 402 } 403 404 ls->ls_total_weight = total; 405 array = kmalloc_array(total, sizeof(*array), GFP_NOFS); 406 if (!array) 407 return; 408 409 list_for_each_entry(memb, &ls->ls_nodes, list) { 410 if (!all_zero && !memb->weight) 411 continue; 412 413 if (all_zero) 414 w = 1; 415 else 416 w = memb->weight; 417 418 DLM_ASSERT(x < total, printk("total %d x %d\n", total, x);); 419 420 for (i = 0; i < w; i++) 421 array[x++] = memb->nodeid; 422 } 423 424 ls->ls_node_array = array; 425 } 426 427 /* send a status request to all members just to establish comms connections */ 428 429 static int ping_members(struct dlm_ls *ls) 430 { 431 struct dlm_member *memb; 432 int error = 0; 433 434 list_for_each_entry(memb, &ls->ls_nodes, list) { 435 error = dlm_recovery_stopped(ls); 436 if (error) 437 break; 438 error = dlm_rcom_status(ls, memb->nodeid, 0); 439 if (error) 440 break; 441 } 442 if (error) 443 log_rinfo(ls, "ping_members aborted %d last nodeid %d", 444 error, ls->ls_recover_nodeid); 445 return error; 446 } 447 448 static void dlm_lsop_recover_prep(struct dlm_ls *ls) 449 { 450 if (!ls->ls_ops || !ls->ls_ops->recover_prep) 451 return; 452 ls->ls_ops->recover_prep(ls->ls_ops_arg); 453 } 454 455 static void dlm_lsop_recover_slot(struct dlm_ls *ls, struct dlm_member *memb) 456 { 457 struct dlm_slot slot; 458 uint32_t seq; 459 int error; 460 461 if (!ls->ls_ops || !ls->ls_ops->recover_slot) 462 return; 463 464 /* if there is no comms connection with this node 465 or the present comms connection is newer 466 than the one when this member was added, then 467 we consider the node to have failed (versus 468 being removed due to dlm_release_lockspace) */ 469 470 error = dlm_comm_seq(memb->nodeid, &seq); 471 472 if (!error && seq == memb->comm_seq) 473 return; 474 475 slot.nodeid = memb->nodeid; 476 slot.slot = memb->slot; 477 478 ls->ls_ops->recover_slot(ls->ls_ops_arg, &slot); 479 } 480 481 void dlm_lsop_recover_done(struct dlm_ls *ls) 482 { 483 struct dlm_member *memb; 484 struct dlm_slot *slots; 485 int i, num; 486 487 if (!ls->ls_ops || !ls->ls_ops->recover_done) 488 return; 489 490 num = ls->ls_num_nodes; 491 slots = kcalloc(num, sizeof(*slots), GFP_KERNEL); 492 if (!slots) 493 return; 494 495 i = 0; 496 list_for_each_entry(memb, &ls->ls_nodes, list) { 497 if (i == num) { 498 log_error(ls, "dlm_lsop_recover_done bad num %d", num); 499 goto out; 500 } 501 slots[i].nodeid = memb->nodeid; 502 slots[i].slot = memb->slot; 503 i++; 504 } 505 506 ls->ls_ops->recover_done(ls->ls_ops_arg, slots, num, 507 ls->ls_slot, ls->ls_generation); 508 out: 509 kfree(slots); 510 } 511 512 static struct dlm_config_node *find_config_node(struct dlm_recover *rv, 513 int nodeid) 514 { 515 int i; 516 517 for (i = 0; i < rv->nodes_count; i++) { 518 if (rv->nodes[i].nodeid == nodeid) 519 return &rv->nodes[i]; 520 } 521 return NULL; 522 } 523 524 int dlm_recover_members(struct dlm_ls *ls, struct dlm_recover *rv, int *neg_out) 525 { 526 struct dlm_member *memb, *safe; 527 struct dlm_config_node *node; 528 int i, error, neg = 0, low = -1; 529 530 /* previously removed members that we've not finished removing need to 531 count as a negative change so the "neg" recovery steps will happen */ 532 533 list_for_each_entry(memb, &ls->ls_nodes_gone, list) { 534 log_rinfo(ls, "prev removed member %d", memb->nodeid); 535 neg++; 536 } 537 538 /* move departed members from ls_nodes to ls_nodes_gone */ 539 540 list_for_each_entry_safe(memb, safe, &ls->ls_nodes, list) { 541 node = find_config_node(rv, memb->nodeid); 542 if (node && !node->new) 543 continue; 544 545 if (!node) { 546 log_rinfo(ls, "remove member %d", memb->nodeid); 547 } else { 548 /* removed and re-added */ 549 log_rinfo(ls, "remove member %d comm_seq %u %u", 550 memb->nodeid, memb->comm_seq, node->comm_seq); 551 } 552 553 neg++; 554 list_move(&memb->list, &ls->ls_nodes_gone); 555 ls->ls_num_nodes--; 556 dlm_lsop_recover_slot(ls, memb); 557 } 558 559 /* add new members to ls_nodes */ 560 561 for (i = 0; i < rv->nodes_count; i++) { 562 node = &rv->nodes[i]; 563 if (dlm_is_member(ls, node->nodeid)) 564 continue; 565 dlm_add_member(ls, node); 566 log_rinfo(ls, "add member %d", node->nodeid); 567 } 568 569 list_for_each_entry(memb, &ls->ls_nodes, list) { 570 if (low == -1 || memb->nodeid < low) 571 low = memb->nodeid; 572 } 573 ls->ls_low_nodeid = low; 574 575 make_member_array(ls); 576 *neg_out = neg; 577 578 error = ping_members(ls); 579 if (!error || error == -EPROTO) { 580 /* new_lockspace() may be waiting to know if the config 581 is good or bad */ 582 ls->ls_members_result = error; 583 complete(&ls->ls_members_done); 584 } 585 586 log_rinfo(ls, "dlm_recover_members %d nodes", ls->ls_num_nodes); 587 return error; 588 } 589 590 /* Userspace guarantees that dlm_ls_stop() has completed on all nodes before 591 dlm_ls_start() is called on any of them to start the new recovery. */ 592 593 int dlm_ls_stop(struct dlm_ls *ls) 594 { 595 int new; 596 597 /* 598 * Prevent dlm_recv from being in the middle of something when we do 599 * the stop. This includes ensuring dlm_recv isn't processing a 600 * recovery message (rcom), while dlm_recoverd is aborting and 601 * resetting things from an in-progress recovery. i.e. we want 602 * dlm_recoverd to abort its recovery without worrying about dlm_recv 603 * processing an rcom at the same time. Stopping dlm_recv also makes 604 * it easy for dlm_receive_message() to check locking stopped and add a 605 * message to the requestqueue without races. 606 */ 607 608 down_write(&ls->ls_recv_active); 609 610 /* 611 * Abort any recovery that's in progress (see RECOVER_STOP, 612 * dlm_recovery_stopped()) and tell any other threads running in the 613 * dlm to quit any processing (see RUNNING, dlm_locking_stopped()). 614 */ 615 616 spin_lock(&ls->ls_recover_lock); 617 set_bit(LSFL_RECOVER_STOP, &ls->ls_flags); 618 new = test_and_clear_bit(LSFL_RUNNING, &ls->ls_flags); 619 ls->ls_recover_seq++; 620 spin_unlock(&ls->ls_recover_lock); 621 622 /* 623 * Let dlm_recv run again, now any normal messages will be saved on the 624 * requestqueue for later. 625 */ 626 627 up_write(&ls->ls_recv_active); 628 629 /* 630 * This in_recovery lock does two things: 631 * 1) Keeps this function from returning until all threads are out 632 * of locking routines and locking is truly stopped. 633 * 2) Keeps any new requests from being processed until it's unlocked 634 * when recovery is complete. 635 */ 636 637 if (new) { 638 set_bit(LSFL_RECOVER_DOWN, &ls->ls_flags); 639 wake_up_process(ls->ls_recoverd_task); 640 wait_event(ls->ls_recover_lock_wait, 641 test_bit(LSFL_RECOVER_LOCK, &ls->ls_flags)); 642 } 643 644 /* 645 * The recoverd suspend/resume makes sure that dlm_recoverd (if 646 * running) has noticed RECOVER_STOP above and quit processing the 647 * previous recovery. 648 */ 649 650 dlm_recoverd_suspend(ls); 651 652 spin_lock(&ls->ls_recover_lock); 653 kfree(ls->ls_slots); 654 ls->ls_slots = NULL; 655 ls->ls_num_slots = 0; 656 ls->ls_slots_size = 0; 657 ls->ls_recover_status = 0; 658 spin_unlock(&ls->ls_recover_lock); 659 660 dlm_recoverd_resume(ls); 661 662 if (!ls->ls_recover_begin) 663 ls->ls_recover_begin = jiffies; 664 665 dlm_lsop_recover_prep(ls); 666 return 0; 667 } 668 669 int dlm_ls_start(struct dlm_ls *ls) 670 { 671 struct dlm_recover *rv, *rv_old; 672 struct dlm_config_node *nodes = NULL; 673 int error, count; 674 675 rv = kzalloc(sizeof(*rv), GFP_NOFS); 676 if (!rv) 677 return -ENOMEM; 678 679 error = dlm_config_nodes(ls->ls_name, &nodes, &count); 680 if (error < 0) 681 goto fail_rv; 682 683 spin_lock(&ls->ls_recover_lock); 684 685 /* the lockspace needs to be stopped before it can be started */ 686 687 if (!dlm_locking_stopped(ls)) { 688 spin_unlock(&ls->ls_recover_lock); 689 log_error(ls, "start ignored: lockspace running"); 690 error = -EINVAL; 691 goto fail; 692 } 693 694 rv->nodes = nodes; 695 rv->nodes_count = count; 696 rv->seq = ++ls->ls_recover_seq; 697 rv_old = ls->ls_recover_args; 698 ls->ls_recover_args = rv; 699 spin_unlock(&ls->ls_recover_lock); 700 701 if (rv_old) { 702 log_error(ls, "unused recovery %llx %d", 703 (unsigned long long)rv_old->seq, rv_old->nodes_count); 704 kfree(rv_old->nodes); 705 kfree(rv_old); 706 } 707 708 set_bit(LSFL_RECOVER_WORK, &ls->ls_flags); 709 wake_up_process(ls->ls_recoverd_task); 710 return 0; 711 712 fail: 713 kfree(nodes); 714 fail_rv: 715 kfree(rv); 716 return error; 717 } 718 719