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 "midcomms.h" 19 #include "lowcomms.h" 20 21 int dlm_slots_version(struct dlm_header *h) 22 { 23 if ((le32_to_cpu(h->h_version) & 0x0000FFFF) < DLM_HEADER_SLOTS) 24 return 0; 25 return 1; 26 } 27 28 void dlm_slot_save(struct dlm_ls *ls, struct dlm_rcom *rc, 29 struct dlm_member *memb) 30 { 31 struct rcom_config *rf = (struct rcom_config *)rc->rc_buf; 32 33 if (!dlm_slots_version(&rc->rc_header)) 34 return; 35 36 memb->slot = le16_to_cpu(rf->rf_our_slot); 37 memb->generation = le32_to_cpu(rf->rf_generation); 38 } 39 40 void dlm_slots_copy_out(struct dlm_ls *ls, struct dlm_rcom *rc) 41 { 42 struct dlm_slot *slot; 43 struct rcom_slot *ro; 44 int i; 45 46 ro = (struct rcom_slot *)(rc->rc_buf + sizeof(struct rcom_config)); 47 48 /* ls_slots array is sparse, but not rcom_slots */ 49 50 for (i = 0; i < ls->ls_slots_size; i++) { 51 slot = &ls->ls_slots[i]; 52 if (!slot->nodeid) 53 continue; 54 ro->ro_nodeid = cpu_to_le32(slot->nodeid); 55 ro->ro_slot = cpu_to_le16(slot->slot); 56 ro++; 57 } 58 } 59 60 #define SLOT_DEBUG_LINE 128 61 62 static void log_slots(struct dlm_ls *ls, uint32_t gen, int num_slots, 63 struct rcom_slot *ro0, struct dlm_slot *array, 64 int array_size) 65 { 66 char line[SLOT_DEBUG_LINE]; 67 int len = SLOT_DEBUG_LINE - 1; 68 int pos = 0; 69 int ret, i; 70 71 memset(line, 0, sizeof(line)); 72 73 if (array) { 74 for (i = 0; i < array_size; i++) { 75 if (!array[i].nodeid) 76 continue; 77 78 ret = snprintf(line + pos, len - pos, " %d:%d", 79 array[i].slot, array[i].nodeid); 80 if (ret >= len - pos) 81 break; 82 pos += ret; 83 } 84 } else if (ro0) { 85 for (i = 0; i < num_slots; i++) { 86 ret = snprintf(line + pos, len - pos, " %d:%d", 87 ro0[i].ro_slot, ro0[i].ro_nodeid); 88 if (ret >= len - pos) 89 break; 90 pos += ret; 91 } 92 } 93 94 log_rinfo(ls, "generation %u slots %d%s", gen, num_slots, line); 95 } 96 97 int dlm_slots_copy_in(struct dlm_ls *ls) 98 { 99 struct dlm_member *memb; 100 struct dlm_rcom *rc = ls->ls_recover_buf; 101 struct rcom_config *rf = (struct rcom_config *)rc->rc_buf; 102 struct rcom_slot *ro0, *ro; 103 int our_nodeid = dlm_our_nodeid(); 104 int i, num_slots; 105 uint32_t gen; 106 107 if (!dlm_slots_version(&rc->rc_header)) 108 return -1; 109 110 gen = le32_to_cpu(rf->rf_generation); 111 if (gen <= ls->ls_generation) { 112 log_error(ls, "dlm_slots_copy_in gen %u old %u", 113 gen, ls->ls_generation); 114 } 115 ls->ls_generation = gen; 116 117 num_slots = le16_to_cpu(rf->rf_num_slots); 118 if (!num_slots) 119 return -1; 120 121 ro0 = (struct rcom_slot *)(rc->rc_buf + sizeof(struct rcom_config)); 122 123 log_slots(ls, gen, num_slots, ro0, NULL, 0); 124 125 list_for_each_entry(memb, &ls->ls_nodes, list) { 126 for (i = 0, ro = ro0; i < num_slots; i++, ro++) { 127 if (le32_to_cpu(ro->ro_nodeid) != memb->nodeid) 128 continue; 129 memb->slot = le16_to_cpu(ro->ro_slot); 130 memb->slot_prev = memb->slot; 131 break; 132 } 133 134 if (memb->nodeid == our_nodeid) { 135 if (ls->ls_slot && ls->ls_slot != memb->slot) { 136 log_error(ls, "dlm_slots_copy_in our slot " 137 "changed %d %d", ls->ls_slot, 138 memb->slot); 139 return -1; 140 } 141 142 if (!ls->ls_slot) 143 ls->ls_slot = memb->slot; 144 } 145 146 if (!memb->slot) { 147 log_error(ls, "dlm_slots_copy_in nodeid %d no slot", 148 memb->nodeid); 149 return -1; 150 } 151 } 152 153 return 0; 154 } 155 156 /* for any nodes that do not support slots, we will not have set memb->slot 157 in wait_status_all(), so memb->slot will remain -1, and we will not 158 assign slots or set ls_num_slots here */ 159 160 int dlm_slots_assign(struct dlm_ls *ls, int *num_slots, int *slots_size, 161 struct dlm_slot **slots_out, uint32_t *gen_out) 162 { 163 struct dlm_member *memb; 164 struct dlm_slot *array; 165 int our_nodeid = dlm_our_nodeid(); 166 int array_size, max_slots, i; 167 int need = 0; 168 int max = 0; 169 int num = 0; 170 uint32_t gen = 0; 171 172 /* our own memb struct will have slot -1 gen 0 */ 173 174 list_for_each_entry(memb, &ls->ls_nodes, list) { 175 if (memb->nodeid == our_nodeid) { 176 memb->slot = ls->ls_slot; 177 memb->generation = ls->ls_generation; 178 break; 179 } 180 } 181 182 list_for_each_entry(memb, &ls->ls_nodes, list) { 183 if (memb->generation > gen) 184 gen = memb->generation; 185 186 /* node doesn't support slots */ 187 188 if (memb->slot == -1) 189 return -1; 190 191 /* node needs a slot assigned */ 192 193 if (!memb->slot) 194 need++; 195 196 /* node has a slot assigned */ 197 198 num++; 199 200 if (!max || max < memb->slot) 201 max = memb->slot; 202 203 /* sanity check, once slot is assigned it shouldn't change */ 204 205 if (memb->slot_prev && memb->slot && memb->slot_prev != memb->slot) { 206 log_error(ls, "nodeid %d slot changed %d %d", 207 memb->nodeid, memb->slot_prev, memb->slot); 208 return -1; 209 } 210 memb->slot_prev = memb->slot; 211 } 212 213 array_size = max + need; 214 array = kcalloc(array_size, sizeof(*array), GFP_NOFS); 215 if (!array) 216 return -ENOMEM; 217 218 num = 0; 219 220 /* fill in slots (offsets) that are used */ 221 222 list_for_each_entry(memb, &ls->ls_nodes, list) { 223 if (!memb->slot) 224 continue; 225 226 if (memb->slot > array_size) { 227 log_error(ls, "invalid slot number %d", memb->slot); 228 kfree(array); 229 return -1; 230 } 231 232 array[memb->slot - 1].nodeid = memb->nodeid; 233 array[memb->slot - 1].slot = memb->slot; 234 num++; 235 } 236 237 /* assign new slots from unused offsets */ 238 239 list_for_each_entry(memb, &ls->ls_nodes, list) { 240 if (memb->slot) 241 continue; 242 243 for (i = 0; i < array_size; i++) { 244 if (array[i].nodeid) 245 continue; 246 247 memb->slot = i + 1; 248 memb->slot_prev = memb->slot; 249 array[i].nodeid = memb->nodeid; 250 array[i].slot = memb->slot; 251 num++; 252 253 if (!ls->ls_slot && memb->nodeid == our_nodeid) 254 ls->ls_slot = memb->slot; 255 break; 256 } 257 258 if (!memb->slot) { 259 log_error(ls, "no free slot found"); 260 kfree(array); 261 return -1; 262 } 263 } 264 265 gen++; 266 267 log_slots(ls, gen, num, NULL, array, array_size); 268 269 max_slots = (DLM_MAX_APP_BUFSIZE - sizeof(struct dlm_rcom) - 270 sizeof(struct rcom_config)) / sizeof(struct rcom_slot); 271 272 if (num > max_slots) { 273 log_error(ls, "num_slots %d exceeds max_slots %d", 274 num, max_slots); 275 kfree(array); 276 return -1; 277 } 278 279 *gen_out = gen; 280 *slots_out = array; 281 *slots_size = array_size; 282 *num_slots = num; 283 return 0; 284 } 285 286 static void add_ordered_member(struct dlm_ls *ls, struct dlm_member *new) 287 { 288 struct dlm_member *memb = NULL; 289 struct list_head *tmp; 290 struct list_head *newlist = &new->list; 291 struct list_head *head = &ls->ls_nodes; 292 293 list_for_each(tmp, head) { 294 memb = list_entry(tmp, struct dlm_member, list); 295 if (new->nodeid < memb->nodeid) 296 break; 297 } 298 299 if (!memb) 300 list_add_tail(newlist, head); 301 else { 302 /* FIXME: can use list macro here */ 303 newlist->prev = tmp->prev; 304 newlist->next = tmp; 305 tmp->prev->next = newlist; 306 tmp->prev = newlist; 307 } 308 } 309 310 static int dlm_add_member(struct dlm_ls *ls, struct dlm_config_node *node) 311 { 312 struct dlm_member *memb; 313 int error; 314 315 memb = kzalloc(sizeof(*memb), GFP_NOFS); 316 if (!memb) 317 return -ENOMEM; 318 319 error = dlm_lowcomms_connect_node(node->nodeid); 320 if (error < 0) { 321 kfree(memb); 322 return error; 323 } 324 325 memb->nodeid = node->nodeid; 326 memb->weight = node->weight; 327 memb->comm_seq = node->comm_seq; 328 dlm_midcomms_add_member(node->nodeid); 329 add_ordered_member(ls, memb); 330 ls->ls_num_nodes++; 331 return 0; 332 } 333 334 static struct dlm_member *find_memb(struct list_head *head, int nodeid) 335 { 336 struct dlm_member *memb; 337 338 list_for_each_entry(memb, head, list) { 339 if (memb->nodeid == nodeid) 340 return memb; 341 } 342 return NULL; 343 } 344 345 int dlm_is_member(struct dlm_ls *ls, int nodeid) 346 { 347 if (find_memb(&ls->ls_nodes, nodeid)) 348 return 1; 349 return 0; 350 } 351 352 int dlm_is_removed(struct dlm_ls *ls, int nodeid) 353 { 354 if (find_memb(&ls->ls_nodes_gone, nodeid)) 355 return 1; 356 return 0; 357 } 358 359 static void clear_memb_list(struct list_head *head, 360 void (*after_del)(int nodeid)) 361 { 362 struct dlm_member *memb; 363 364 while (!list_empty(head)) { 365 memb = list_entry(head->next, struct dlm_member, list); 366 list_del(&memb->list); 367 if (after_del) 368 after_del(memb->nodeid); 369 kfree(memb); 370 } 371 } 372 373 static void clear_members_cb(int nodeid) 374 { 375 dlm_midcomms_remove_member(nodeid); 376 } 377 378 void dlm_clear_members(struct dlm_ls *ls) 379 { 380 clear_memb_list(&ls->ls_nodes, clear_members_cb); 381 ls->ls_num_nodes = 0; 382 } 383 384 void dlm_clear_members_gone(struct dlm_ls *ls) 385 { 386 clear_memb_list(&ls->ls_nodes_gone, NULL); 387 } 388 389 static void make_member_array(struct dlm_ls *ls) 390 { 391 struct dlm_member *memb; 392 int i, w, x = 0, total = 0, all_zero = 0, *array; 393 394 kfree(ls->ls_node_array); 395 ls->ls_node_array = NULL; 396 397 list_for_each_entry(memb, &ls->ls_nodes, list) { 398 if (memb->weight) 399 total += memb->weight; 400 } 401 402 /* all nodes revert to weight of 1 if all have weight 0 */ 403 404 if (!total) { 405 total = ls->ls_num_nodes; 406 all_zero = 1; 407 } 408 409 ls->ls_total_weight = total; 410 array = kmalloc_array(total, sizeof(*array), GFP_NOFS); 411 if (!array) 412 return; 413 414 list_for_each_entry(memb, &ls->ls_nodes, list) { 415 if (!all_zero && !memb->weight) 416 continue; 417 418 if (all_zero) 419 w = 1; 420 else 421 w = memb->weight; 422 423 DLM_ASSERT(x < total, printk("total %d x %d\n", total, x);); 424 425 for (i = 0; i < w; i++) 426 array[x++] = memb->nodeid; 427 } 428 429 ls->ls_node_array = array; 430 } 431 432 /* send a status request to all members just to establish comms connections */ 433 434 static int ping_members(struct dlm_ls *ls) 435 { 436 struct dlm_member *memb; 437 int error = 0; 438 439 list_for_each_entry(memb, &ls->ls_nodes, list) { 440 if (dlm_recovery_stopped(ls)) { 441 error = -EINTR; 442 break; 443 } 444 error = dlm_rcom_status(ls, memb->nodeid, 0); 445 if (error) 446 break; 447 } 448 if (error) 449 log_rinfo(ls, "ping_members aborted %d last nodeid %d", 450 error, ls->ls_recover_nodeid); 451 return error; 452 } 453 454 static void dlm_lsop_recover_prep(struct dlm_ls *ls) 455 { 456 if (!ls->ls_ops || !ls->ls_ops->recover_prep) 457 return; 458 ls->ls_ops->recover_prep(ls->ls_ops_arg); 459 } 460 461 static void dlm_lsop_recover_slot(struct dlm_ls *ls, struct dlm_member *memb) 462 { 463 struct dlm_slot slot; 464 uint32_t seq; 465 int error; 466 467 if (!ls->ls_ops || !ls->ls_ops->recover_slot) 468 return; 469 470 /* if there is no comms connection with this node 471 or the present comms connection is newer 472 than the one when this member was added, then 473 we consider the node to have failed (versus 474 being removed due to dlm_release_lockspace) */ 475 476 error = dlm_comm_seq(memb->nodeid, &seq); 477 478 if (!error && seq == memb->comm_seq) 479 return; 480 481 slot.nodeid = memb->nodeid; 482 slot.slot = memb->slot; 483 484 ls->ls_ops->recover_slot(ls->ls_ops_arg, &slot); 485 } 486 487 void dlm_lsop_recover_done(struct dlm_ls *ls) 488 { 489 struct dlm_member *memb; 490 struct dlm_slot *slots; 491 int i, num; 492 493 if (!ls->ls_ops || !ls->ls_ops->recover_done) 494 return; 495 496 num = ls->ls_num_nodes; 497 slots = kcalloc(num, sizeof(*slots), GFP_KERNEL); 498 if (!slots) 499 return; 500 501 i = 0; 502 list_for_each_entry(memb, &ls->ls_nodes, list) { 503 if (i == num) { 504 log_error(ls, "dlm_lsop_recover_done bad num %d", num); 505 goto out; 506 } 507 slots[i].nodeid = memb->nodeid; 508 slots[i].slot = memb->slot; 509 i++; 510 } 511 512 ls->ls_ops->recover_done(ls->ls_ops_arg, slots, num, 513 ls->ls_slot, ls->ls_generation); 514 out: 515 kfree(slots); 516 } 517 518 static struct dlm_config_node *find_config_node(struct dlm_recover *rv, 519 int nodeid) 520 { 521 int i; 522 523 for (i = 0; i < rv->nodes_count; i++) { 524 if (rv->nodes[i].nodeid == nodeid) 525 return &rv->nodes[i]; 526 } 527 return NULL; 528 } 529 530 int dlm_recover_members(struct dlm_ls *ls, struct dlm_recover *rv, int *neg_out) 531 { 532 struct dlm_member *memb, *safe; 533 struct dlm_config_node *node; 534 int i, error, neg = 0, low = -1; 535 536 /* previously removed members that we've not finished removing need to 537 * count as a negative change so the "neg" recovery steps will happen 538 * 539 * This functionality must report all member changes to lsops or 540 * midcomms layer and must never return before. 541 */ 542 543 list_for_each_entry(memb, &ls->ls_nodes_gone, list) { 544 log_rinfo(ls, "prev removed member %d", memb->nodeid); 545 neg++; 546 } 547 548 /* move departed members from ls_nodes to ls_nodes_gone */ 549 550 list_for_each_entry_safe(memb, safe, &ls->ls_nodes, list) { 551 node = find_config_node(rv, memb->nodeid); 552 if (node && !node->new) 553 continue; 554 555 if (!node) { 556 log_rinfo(ls, "remove member %d", memb->nodeid); 557 } else { 558 /* removed and re-added */ 559 log_rinfo(ls, "remove member %d comm_seq %u %u", 560 memb->nodeid, memb->comm_seq, node->comm_seq); 561 } 562 563 neg++; 564 list_move(&memb->list, &ls->ls_nodes_gone); 565 dlm_midcomms_remove_member(memb->nodeid); 566 ls->ls_num_nodes--; 567 dlm_lsop_recover_slot(ls, memb); 568 } 569 570 /* add new members to ls_nodes */ 571 572 for (i = 0; i < rv->nodes_count; i++) { 573 node = &rv->nodes[i]; 574 if (dlm_is_member(ls, node->nodeid)) 575 continue; 576 dlm_add_member(ls, node); 577 log_rinfo(ls, "add member %d", node->nodeid); 578 } 579 580 list_for_each_entry(memb, &ls->ls_nodes, list) { 581 if (low == -1 || memb->nodeid < low) 582 low = memb->nodeid; 583 } 584 ls->ls_low_nodeid = low; 585 586 make_member_array(ls); 587 *neg_out = neg; 588 589 error = ping_members(ls); 590 log_rinfo(ls, "dlm_recover_members %d nodes", ls->ls_num_nodes); 591 return error; 592 } 593 594 /* Userspace guarantees that dlm_ls_stop() has completed on all nodes before 595 dlm_ls_start() is called on any of them to start the new recovery. */ 596 597 int dlm_ls_stop(struct dlm_ls *ls) 598 { 599 int new; 600 601 /* 602 * Prevent dlm_recv from being in the middle of something when we do 603 * the stop. This includes ensuring dlm_recv isn't processing a 604 * recovery message (rcom), while dlm_recoverd is aborting and 605 * resetting things from an in-progress recovery. i.e. we want 606 * dlm_recoverd to abort its recovery without worrying about dlm_recv 607 * processing an rcom at the same time. Stopping dlm_recv also makes 608 * it easy for dlm_receive_message() to check locking stopped and add a 609 * message to the requestqueue without races. 610 */ 611 612 down_write(&ls->ls_recv_active); 613 614 /* 615 * Abort any recovery that's in progress (see RECOVER_STOP, 616 * dlm_recovery_stopped()) and tell any other threads running in the 617 * dlm to quit any processing (see RUNNING, dlm_locking_stopped()). 618 */ 619 620 spin_lock(&ls->ls_recover_lock); 621 set_bit(LSFL_RECOVER_STOP, &ls->ls_flags); 622 new = test_and_clear_bit(LSFL_RUNNING, &ls->ls_flags); 623 ls->ls_recover_seq++; 624 spin_unlock(&ls->ls_recover_lock); 625 626 /* 627 * Let dlm_recv run again, now any normal messages will be saved on the 628 * requestqueue for later. 629 */ 630 631 up_write(&ls->ls_recv_active); 632 633 /* 634 * This in_recovery lock does two things: 635 * 1) Keeps this function from returning until all threads are out 636 * of locking routines and locking is truly stopped. 637 * 2) Keeps any new requests from being processed until it's unlocked 638 * when recovery is complete. 639 */ 640 641 if (new) { 642 set_bit(LSFL_RECOVER_DOWN, &ls->ls_flags); 643 wake_up_process(ls->ls_recoverd_task); 644 wait_event(ls->ls_recover_lock_wait, 645 test_bit(LSFL_RECOVER_LOCK, &ls->ls_flags)); 646 } 647 648 /* 649 * The recoverd suspend/resume makes sure that dlm_recoverd (if 650 * running) has noticed RECOVER_STOP above and quit processing the 651 * previous recovery. 652 */ 653 654 dlm_recoverd_suspend(ls); 655 656 spin_lock(&ls->ls_recover_lock); 657 kfree(ls->ls_slots); 658 ls->ls_slots = NULL; 659 ls->ls_num_slots = 0; 660 ls->ls_slots_size = 0; 661 ls->ls_recover_status = 0; 662 spin_unlock(&ls->ls_recover_lock); 663 664 dlm_recoverd_resume(ls); 665 666 if (!ls->ls_recover_begin) 667 ls->ls_recover_begin = jiffies; 668 669 /* call recover_prep ops only once and not multiple times 670 * for each possible dlm_ls_stop() when recovery is already 671 * stopped. 672 * 673 * If we successful was able to clear LSFL_RUNNING bit and 674 * it was set we know it is the first dlm_ls_stop() call. 675 */ 676 if (new) 677 dlm_lsop_recover_prep(ls); 678 679 return 0; 680 } 681 682 int dlm_ls_start(struct dlm_ls *ls) 683 { 684 struct dlm_recover *rv, *rv_old; 685 struct dlm_config_node *nodes = NULL; 686 int error, count; 687 688 rv = kzalloc(sizeof(*rv), GFP_NOFS); 689 if (!rv) 690 return -ENOMEM; 691 692 error = dlm_config_nodes(ls->ls_name, &nodes, &count); 693 if (error < 0) 694 goto fail_rv; 695 696 spin_lock(&ls->ls_recover_lock); 697 698 /* the lockspace needs to be stopped before it can be started */ 699 700 if (!dlm_locking_stopped(ls)) { 701 spin_unlock(&ls->ls_recover_lock); 702 log_error(ls, "start ignored: lockspace running"); 703 error = -EINVAL; 704 goto fail; 705 } 706 707 rv->nodes = nodes; 708 rv->nodes_count = count; 709 rv->seq = ++ls->ls_recover_seq; 710 rv_old = ls->ls_recover_args; 711 ls->ls_recover_args = rv; 712 spin_unlock(&ls->ls_recover_lock); 713 714 if (rv_old) { 715 log_error(ls, "unused recovery %llx %d", 716 (unsigned long long)rv_old->seq, rv_old->nodes_count); 717 kfree(rv_old->nodes); 718 kfree(rv_old); 719 } 720 721 set_bit(LSFL_RECOVER_WORK, &ls->ls_flags); 722 wake_up_process(ls->ls_recoverd_task); 723 return 0; 724 725 fail: 726 kfree(nodes); 727 fail_rv: 728 kfree(rv); 729 return error; 730 } 731 732