1 /* -*- mode: c; c-basic-offset: 8; -*- 2 * vim: noexpandtab sw=8 ts=8 sts=0: 3 * 4 * dlmrecovery.c 5 * 6 * recovery stuff 7 * 8 * Copyright (C) 2004 Oracle. All rights reserved. 9 * 10 * This program is free software; you can redistribute it and/or 11 * modify it under the terms of the GNU General Public 12 * License as published by the Free Software Foundation; either 13 * version 2 of the License, or (at your option) any later version. 14 * 15 * This program is distributed in the hope that it will be useful, 16 * but WITHOUT ANY WARRANTY; without even the implied warranty of 17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 18 * General Public License for more details. 19 * 20 * You should have received a copy of the GNU General Public 21 * License along with this program; if not, write to the 22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330, 23 * Boston, MA 021110-1307, USA. 24 * 25 */ 26 27 28 #include <linux/module.h> 29 #include <linux/fs.h> 30 #include <linux/types.h> 31 #include <linux/slab.h> 32 #include <linux/highmem.h> 33 #include <linux/init.h> 34 #include <linux/sysctl.h> 35 #include <linux/random.h> 36 #include <linux/blkdev.h> 37 #include <linux/socket.h> 38 #include <linux/inet.h> 39 #include <linux/timer.h> 40 #include <linux/kthread.h> 41 #include <linux/delay.h> 42 43 44 #include "cluster/heartbeat.h" 45 #include "cluster/nodemanager.h" 46 #include "cluster/tcp.h" 47 48 #include "dlmapi.h" 49 #include "dlmcommon.h" 50 #include "dlmdomain.h" 51 52 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_RECOVERY) 53 #include "cluster/masklog.h" 54 55 static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node); 56 57 static int dlm_recovery_thread(void *data); 58 void dlm_complete_recovery_thread(struct dlm_ctxt *dlm); 59 int dlm_launch_recovery_thread(struct dlm_ctxt *dlm); 60 void dlm_kick_recovery_thread(struct dlm_ctxt *dlm); 61 static int dlm_do_recovery(struct dlm_ctxt *dlm); 62 63 static int dlm_pick_recovery_master(struct dlm_ctxt *dlm); 64 static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node); 65 static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node); 66 static int dlm_request_all_locks(struct dlm_ctxt *dlm, 67 u8 request_from, u8 dead_node); 68 static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm, u8 dead_node); 69 70 static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res); 71 static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres, 72 const char *lockname, int namelen, 73 int total_locks, u64 cookie, 74 u8 flags, u8 master); 75 static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm, 76 struct dlm_migratable_lockres *mres, 77 u8 send_to, 78 struct dlm_lock_resource *res, 79 int total_locks); 80 static int dlm_process_recovery_data(struct dlm_ctxt *dlm, 81 struct dlm_lock_resource *res, 82 struct dlm_migratable_lockres *mres); 83 static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm); 84 static int dlm_send_all_done_msg(struct dlm_ctxt *dlm, 85 u8 dead_node, u8 send_to); 86 static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node); 87 static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm, 88 struct list_head *list, u8 dead_node); 89 static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm, 90 u8 dead_node, u8 new_master); 91 static void dlm_reco_ast(void *astdata); 92 static void dlm_reco_bast(void *astdata, int blocked_type); 93 static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st); 94 static void dlm_request_all_locks_worker(struct dlm_work_item *item, 95 void *data); 96 static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data); 97 static int dlm_lockres_master_requery(struct dlm_ctxt *dlm, 98 struct dlm_lock_resource *res, 99 u8 *real_master); 100 101 static u64 dlm_get_next_mig_cookie(void); 102 103 static DEFINE_SPINLOCK(dlm_reco_state_lock); 104 static DEFINE_SPINLOCK(dlm_mig_cookie_lock); 105 static u64 dlm_mig_cookie = 1; 106 107 static u64 dlm_get_next_mig_cookie(void) 108 { 109 u64 c; 110 spin_lock(&dlm_mig_cookie_lock); 111 c = dlm_mig_cookie; 112 if (dlm_mig_cookie == (~0ULL)) 113 dlm_mig_cookie = 1; 114 else 115 dlm_mig_cookie++; 116 spin_unlock(&dlm_mig_cookie_lock); 117 return c; 118 } 119 120 static inline void dlm_set_reco_dead_node(struct dlm_ctxt *dlm, 121 u8 dead_node) 122 { 123 assert_spin_locked(&dlm->spinlock); 124 if (dlm->reco.dead_node != dead_node) 125 mlog(0, "%s: changing dead_node from %u to %u\n", 126 dlm->name, dlm->reco.dead_node, dead_node); 127 dlm->reco.dead_node = dead_node; 128 } 129 130 static inline void dlm_set_reco_master(struct dlm_ctxt *dlm, 131 u8 master) 132 { 133 assert_spin_locked(&dlm->spinlock); 134 mlog(0, "%s: changing new_master from %u to %u\n", 135 dlm->name, dlm->reco.new_master, master); 136 dlm->reco.new_master = master; 137 } 138 139 static inline void __dlm_reset_recovery(struct dlm_ctxt *dlm) 140 { 141 assert_spin_locked(&dlm->spinlock); 142 clear_bit(dlm->reco.dead_node, dlm->recovery_map); 143 dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM); 144 dlm_set_reco_master(dlm, O2NM_INVALID_NODE_NUM); 145 } 146 147 static inline void dlm_reset_recovery(struct dlm_ctxt *dlm) 148 { 149 spin_lock(&dlm->spinlock); 150 __dlm_reset_recovery(dlm); 151 spin_unlock(&dlm->spinlock); 152 } 153 154 /* Worker function used during recovery. */ 155 void dlm_dispatch_work(struct work_struct *work) 156 { 157 struct dlm_ctxt *dlm = 158 container_of(work, struct dlm_ctxt, dispatched_work); 159 LIST_HEAD(tmp_list); 160 struct dlm_work_item *item, *next; 161 dlm_workfunc_t *workfunc; 162 int tot=0; 163 164 spin_lock(&dlm->work_lock); 165 list_splice_init(&dlm->work_list, &tmp_list); 166 spin_unlock(&dlm->work_lock); 167 168 list_for_each_entry(item, &tmp_list, list) { 169 tot++; 170 } 171 mlog(0, "%s: work thread has %d work items\n", dlm->name, tot); 172 173 list_for_each_entry_safe(item, next, &tmp_list, list) { 174 workfunc = item->func; 175 list_del_init(&item->list); 176 177 /* already have ref on dlm to avoid having 178 * it disappear. just double-check. */ 179 BUG_ON(item->dlm != dlm); 180 181 /* this is allowed to sleep and 182 * call network stuff */ 183 workfunc(item, item->data); 184 185 dlm_put(dlm); 186 kfree(item); 187 } 188 } 189 190 /* 191 * RECOVERY THREAD 192 */ 193 194 void dlm_kick_recovery_thread(struct dlm_ctxt *dlm) 195 { 196 /* wake the recovery thread 197 * this will wake the reco thread in one of three places 198 * 1) sleeping with no recovery happening 199 * 2) sleeping with recovery mastered elsewhere 200 * 3) recovery mastered here, waiting on reco data */ 201 202 wake_up(&dlm->dlm_reco_thread_wq); 203 } 204 205 /* Launch the recovery thread */ 206 int dlm_launch_recovery_thread(struct dlm_ctxt *dlm) 207 { 208 mlog(0, "starting dlm recovery thread...\n"); 209 210 dlm->dlm_reco_thread_task = kthread_run(dlm_recovery_thread, dlm, 211 "dlm_reco_thread"); 212 if (IS_ERR(dlm->dlm_reco_thread_task)) { 213 mlog_errno(PTR_ERR(dlm->dlm_reco_thread_task)); 214 dlm->dlm_reco_thread_task = NULL; 215 return -EINVAL; 216 } 217 218 return 0; 219 } 220 221 void dlm_complete_recovery_thread(struct dlm_ctxt *dlm) 222 { 223 if (dlm->dlm_reco_thread_task) { 224 mlog(0, "waiting for dlm recovery thread to exit\n"); 225 kthread_stop(dlm->dlm_reco_thread_task); 226 dlm->dlm_reco_thread_task = NULL; 227 } 228 } 229 230 231 232 /* 233 * this is lame, but here's how recovery works... 234 * 1) all recovery threads cluster wide will work on recovering 235 * ONE node at a time 236 * 2) negotiate who will take over all the locks for the dead node. 237 * thats right... ALL the locks. 238 * 3) once a new master is chosen, everyone scans all locks 239 * and moves aside those mastered by the dead guy 240 * 4) each of these locks should be locked until recovery is done 241 * 5) the new master collects up all of secondary lock queue info 242 * one lock at a time, forcing each node to communicate back 243 * before continuing 244 * 6) each secondary lock queue responds with the full known lock info 245 * 7) once the new master has run all its locks, it sends a ALLDONE! 246 * message to everyone 247 * 8) upon receiving this message, the secondary queue node unlocks 248 * and responds to the ALLDONE 249 * 9) once the new master gets responses from everyone, he unlocks 250 * everything and recovery for this dead node is done 251 *10) go back to 2) while there are still dead nodes 252 * 253 */ 254 255 static void dlm_print_reco_node_status(struct dlm_ctxt *dlm) 256 { 257 struct dlm_reco_node_data *ndata; 258 struct dlm_lock_resource *res; 259 260 mlog(ML_NOTICE, "%s(%d): recovery info, state=%s, dead=%u, master=%u\n", 261 dlm->name, task_pid_nr(dlm->dlm_reco_thread_task), 262 dlm->reco.state & DLM_RECO_STATE_ACTIVE ? "ACTIVE" : "inactive", 263 dlm->reco.dead_node, dlm->reco.new_master); 264 265 list_for_each_entry(ndata, &dlm->reco.node_data, list) { 266 char *st = "unknown"; 267 switch (ndata->state) { 268 case DLM_RECO_NODE_DATA_INIT: 269 st = "init"; 270 break; 271 case DLM_RECO_NODE_DATA_REQUESTING: 272 st = "requesting"; 273 break; 274 case DLM_RECO_NODE_DATA_DEAD: 275 st = "dead"; 276 break; 277 case DLM_RECO_NODE_DATA_RECEIVING: 278 st = "receiving"; 279 break; 280 case DLM_RECO_NODE_DATA_REQUESTED: 281 st = "requested"; 282 break; 283 case DLM_RECO_NODE_DATA_DONE: 284 st = "done"; 285 break; 286 case DLM_RECO_NODE_DATA_FINALIZE_SENT: 287 st = "finalize-sent"; 288 break; 289 default: 290 st = "bad"; 291 break; 292 } 293 mlog(ML_NOTICE, "%s: reco state, node %u, state=%s\n", 294 dlm->name, ndata->node_num, st); 295 } 296 list_for_each_entry(res, &dlm->reco.resources, recovering) { 297 mlog(ML_NOTICE, "%s: lockres %.*s on recovering list\n", 298 dlm->name, res->lockname.len, res->lockname.name); 299 } 300 } 301 302 #define DLM_RECO_THREAD_TIMEOUT_MS (5 * 1000) 303 304 static int dlm_recovery_thread(void *data) 305 { 306 int status; 307 struct dlm_ctxt *dlm = data; 308 unsigned long timeout = msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS); 309 310 mlog(0, "dlm thread running for %s...\n", dlm->name); 311 312 while (!kthread_should_stop()) { 313 if (dlm_domain_fully_joined(dlm)) { 314 status = dlm_do_recovery(dlm); 315 if (status == -EAGAIN) { 316 /* do not sleep, recheck immediately. */ 317 continue; 318 } 319 if (status < 0) 320 mlog_errno(status); 321 } 322 323 wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq, 324 kthread_should_stop(), 325 timeout); 326 } 327 328 mlog(0, "quitting DLM recovery thread\n"); 329 return 0; 330 } 331 332 /* returns true when the recovery master has contacted us */ 333 static int dlm_reco_master_ready(struct dlm_ctxt *dlm) 334 { 335 int ready; 336 spin_lock(&dlm->spinlock); 337 ready = (dlm->reco.new_master != O2NM_INVALID_NODE_NUM); 338 spin_unlock(&dlm->spinlock); 339 return ready; 340 } 341 342 /* returns true if node is no longer in the domain 343 * could be dead or just not joined */ 344 int dlm_is_node_dead(struct dlm_ctxt *dlm, u8 node) 345 { 346 int dead; 347 spin_lock(&dlm->spinlock); 348 dead = !test_bit(node, dlm->domain_map); 349 spin_unlock(&dlm->spinlock); 350 return dead; 351 } 352 353 /* returns true if node is no longer in the domain 354 * could be dead or just not joined */ 355 static int dlm_is_node_recovered(struct dlm_ctxt *dlm, u8 node) 356 { 357 int recovered; 358 spin_lock(&dlm->spinlock); 359 recovered = !test_bit(node, dlm->recovery_map); 360 spin_unlock(&dlm->spinlock); 361 return recovered; 362 } 363 364 365 int dlm_wait_for_node_death(struct dlm_ctxt *dlm, u8 node, int timeout) 366 { 367 if (timeout) { 368 mlog(ML_NOTICE, "%s: waiting %dms for notification of " 369 "death of node %u\n", dlm->name, timeout, node); 370 wait_event_timeout(dlm->dlm_reco_thread_wq, 371 dlm_is_node_dead(dlm, node), 372 msecs_to_jiffies(timeout)); 373 } else { 374 mlog(ML_NOTICE, "%s: waiting indefinitely for notification " 375 "of death of node %u\n", dlm->name, node); 376 wait_event(dlm->dlm_reco_thread_wq, 377 dlm_is_node_dead(dlm, node)); 378 } 379 /* for now, return 0 */ 380 return 0; 381 } 382 383 int dlm_wait_for_node_recovery(struct dlm_ctxt *dlm, u8 node, int timeout) 384 { 385 if (timeout) { 386 mlog(0, "%s: waiting %dms for notification of " 387 "recovery of node %u\n", dlm->name, timeout, node); 388 wait_event_timeout(dlm->dlm_reco_thread_wq, 389 dlm_is_node_recovered(dlm, node), 390 msecs_to_jiffies(timeout)); 391 } else { 392 mlog(0, "%s: waiting indefinitely for notification " 393 "of recovery of node %u\n", dlm->name, node); 394 wait_event(dlm->dlm_reco_thread_wq, 395 dlm_is_node_recovered(dlm, node)); 396 } 397 /* for now, return 0 */ 398 return 0; 399 } 400 401 /* callers of the top-level api calls (dlmlock/dlmunlock) should 402 * block on the dlm->reco.event when recovery is in progress. 403 * the dlm recovery thread will set this state when it begins 404 * recovering a dead node (as the new master or not) and clear 405 * the state and wake as soon as all affected lock resources have 406 * been marked with the RECOVERY flag */ 407 static int dlm_in_recovery(struct dlm_ctxt *dlm) 408 { 409 int in_recovery; 410 spin_lock(&dlm->spinlock); 411 in_recovery = !!(dlm->reco.state & DLM_RECO_STATE_ACTIVE); 412 spin_unlock(&dlm->spinlock); 413 return in_recovery; 414 } 415 416 417 void dlm_wait_for_recovery(struct dlm_ctxt *dlm) 418 { 419 if (dlm_in_recovery(dlm)) { 420 mlog(0, "%s: reco thread %d in recovery: " 421 "state=%d, master=%u, dead=%u\n", 422 dlm->name, task_pid_nr(dlm->dlm_reco_thread_task), 423 dlm->reco.state, dlm->reco.new_master, 424 dlm->reco.dead_node); 425 } 426 wait_event(dlm->reco.event, !dlm_in_recovery(dlm)); 427 } 428 429 static void dlm_begin_recovery(struct dlm_ctxt *dlm) 430 { 431 spin_lock(&dlm->spinlock); 432 BUG_ON(dlm->reco.state & DLM_RECO_STATE_ACTIVE); 433 dlm->reco.state |= DLM_RECO_STATE_ACTIVE; 434 spin_unlock(&dlm->spinlock); 435 } 436 437 static void dlm_end_recovery(struct dlm_ctxt *dlm) 438 { 439 spin_lock(&dlm->spinlock); 440 BUG_ON(!(dlm->reco.state & DLM_RECO_STATE_ACTIVE)); 441 dlm->reco.state &= ~DLM_RECO_STATE_ACTIVE; 442 spin_unlock(&dlm->spinlock); 443 wake_up(&dlm->reco.event); 444 } 445 446 static int dlm_do_recovery(struct dlm_ctxt *dlm) 447 { 448 int status = 0; 449 int ret; 450 451 spin_lock(&dlm->spinlock); 452 453 /* check to see if the new master has died */ 454 if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM && 455 test_bit(dlm->reco.new_master, dlm->recovery_map)) { 456 mlog(0, "new master %u died while recovering %u!\n", 457 dlm->reco.new_master, dlm->reco.dead_node); 458 /* unset the new_master, leave dead_node */ 459 dlm_set_reco_master(dlm, O2NM_INVALID_NODE_NUM); 460 } 461 462 /* select a target to recover */ 463 if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) { 464 int bit; 465 466 bit = find_next_bit (dlm->recovery_map, O2NM_MAX_NODES, 0); 467 if (bit >= O2NM_MAX_NODES || bit < 0) 468 dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM); 469 else 470 dlm_set_reco_dead_node(dlm, bit); 471 } else if (!test_bit(dlm->reco.dead_node, dlm->recovery_map)) { 472 /* BUG? */ 473 mlog(ML_ERROR, "dead_node %u no longer in recovery map!\n", 474 dlm->reco.dead_node); 475 dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM); 476 } 477 478 if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) { 479 // mlog(0, "nothing to recover! sleeping now!\n"); 480 spin_unlock(&dlm->spinlock); 481 /* return to main thread loop and sleep. */ 482 return 0; 483 } 484 mlog(0, "%s(%d):recovery thread found node %u in the recovery map!\n", 485 dlm->name, task_pid_nr(dlm->dlm_reco_thread_task), 486 dlm->reco.dead_node); 487 spin_unlock(&dlm->spinlock); 488 489 /* take write barrier */ 490 /* (stops the list reshuffling thread, proxy ast handling) */ 491 dlm_begin_recovery(dlm); 492 493 if (dlm->reco.new_master == dlm->node_num) 494 goto master_here; 495 496 if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) { 497 /* choose a new master, returns 0 if this node 498 * is the master, -EEXIST if it's another node. 499 * this does not return until a new master is chosen 500 * or recovery completes entirely. */ 501 ret = dlm_pick_recovery_master(dlm); 502 if (!ret) { 503 /* already notified everyone. go. */ 504 goto master_here; 505 } 506 mlog(0, "another node will master this recovery session.\n"); 507 } 508 mlog(0, "dlm=%s (%d), new_master=%u, this node=%u, dead_node=%u\n", 509 dlm->name, task_pid_nr(dlm->dlm_reco_thread_task), dlm->reco.new_master, 510 dlm->node_num, dlm->reco.dead_node); 511 512 /* it is safe to start everything back up here 513 * because all of the dead node's lock resources 514 * have been marked as in-recovery */ 515 dlm_end_recovery(dlm); 516 517 /* sleep out in main dlm_recovery_thread loop. */ 518 return 0; 519 520 master_here: 521 mlog(ML_NOTICE, "(%d) Node %u is the Recovery Master for the Dead Node " 522 "%u for Domain %s\n", task_pid_nr(dlm->dlm_reco_thread_task), 523 dlm->node_num, dlm->reco.dead_node, dlm->name); 524 525 status = dlm_remaster_locks(dlm, dlm->reco.dead_node); 526 if (status < 0) { 527 /* we should never hit this anymore */ 528 mlog(ML_ERROR, "error %d remastering locks for node %u, " 529 "retrying.\n", status, dlm->reco.dead_node); 530 /* yield a bit to allow any final network messages 531 * to get handled on remaining nodes */ 532 msleep(100); 533 } else { 534 /* success! see if any other nodes need recovery */ 535 mlog(0, "DONE mastering recovery of %s:%u here(this=%u)!\n", 536 dlm->name, dlm->reco.dead_node, dlm->node_num); 537 dlm_reset_recovery(dlm); 538 } 539 dlm_end_recovery(dlm); 540 541 /* continue and look for another dead node */ 542 return -EAGAIN; 543 } 544 545 static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node) 546 { 547 int status = 0; 548 struct dlm_reco_node_data *ndata; 549 int all_nodes_done; 550 int destroy = 0; 551 int pass = 0; 552 553 do { 554 /* we have become recovery master. there is no escaping 555 * this, so just keep trying until we get it. */ 556 status = dlm_init_recovery_area(dlm, dead_node); 557 if (status < 0) { 558 mlog(ML_ERROR, "%s: failed to alloc recovery area, " 559 "retrying\n", dlm->name); 560 msleep(1000); 561 } 562 } while (status != 0); 563 564 /* safe to access the node data list without a lock, since this 565 * process is the only one to change the list */ 566 list_for_each_entry(ndata, &dlm->reco.node_data, list) { 567 BUG_ON(ndata->state != DLM_RECO_NODE_DATA_INIT); 568 ndata->state = DLM_RECO_NODE_DATA_REQUESTING; 569 570 mlog(0, "requesting lock info from node %u\n", 571 ndata->node_num); 572 573 if (ndata->node_num == dlm->node_num) { 574 ndata->state = DLM_RECO_NODE_DATA_DONE; 575 continue; 576 } 577 578 do { 579 status = dlm_request_all_locks(dlm, ndata->node_num, 580 dead_node); 581 if (status < 0) { 582 mlog_errno(status); 583 if (dlm_is_host_down(status)) { 584 /* node died, ignore it for recovery */ 585 status = 0; 586 ndata->state = DLM_RECO_NODE_DATA_DEAD; 587 /* wait for the domain map to catch up 588 * with the network state. */ 589 wait_event_timeout(dlm->dlm_reco_thread_wq, 590 dlm_is_node_dead(dlm, 591 ndata->node_num), 592 msecs_to_jiffies(1000)); 593 mlog(0, "waited 1 sec for %u, " 594 "dead? %s\n", ndata->node_num, 595 dlm_is_node_dead(dlm, ndata->node_num) ? 596 "yes" : "no"); 597 } else { 598 /* -ENOMEM on the other node */ 599 mlog(0, "%s: node %u returned " 600 "%d during recovery, retrying " 601 "after a short wait\n", 602 dlm->name, ndata->node_num, 603 status); 604 msleep(100); 605 } 606 } 607 } while (status != 0); 608 609 spin_lock(&dlm_reco_state_lock); 610 switch (ndata->state) { 611 case DLM_RECO_NODE_DATA_INIT: 612 case DLM_RECO_NODE_DATA_FINALIZE_SENT: 613 case DLM_RECO_NODE_DATA_REQUESTED: 614 BUG(); 615 break; 616 case DLM_RECO_NODE_DATA_DEAD: 617 mlog(0, "node %u died after requesting " 618 "recovery info for node %u\n", 619 ndata->node_num, dead_node); 620 /* fine. don't need this node's info. 621 * continue without it. */ 622 break; 623 case DLM_RECO_NODE_DATA_REQUESTING: 624 ndata->state = DLM_RECO_NODE_DATA_REQUESTED; 625 mlog(0, "now receiving recovery data from " 626 "node %u for dead node %u\n", 627 ndata->node_num, dead_node); 628 break; 629 case DLM_RECO_NODE_DATA_RECEIVING: 630 mlog(0, "already receiving recovery data from " 631 "node %u for dead node %u\n", 632 ndata->node_num, dead_node); 633 break; 634 case DLM_RECO_NODE_DATA_DONE: 635 mlog(0, "already DONE receiving recovery data " 636 "from node %u for dead node %u\n", 637 ndata->node_num, dead_node); 638 break; 639 } 640 spin_unlock(&dlm_reco_state_lock); 641 } 642 643 mlog(0, "done requesting all lock info\n"); 644 645 /* nodes should be sending reco data now 646 * just need to wait */ 647 648 while (1) { 649 /* check all the nodes now to see if we are 650 * done, or if anyone died */ 651 all_nodes_done = 1; 652 spin_lock(&dlm_reco_state_lock); 653 list_for_each_entry(ndata, &dlm->reco.node_data, list) { 654 mlog(0, "checking recovery state of node %u\n", 655 ndata->node_num); 656 switch (ndata->state) { 657 case DLM_RECO_NODE_DATA_INIT: 658 case DLM_RECO_NODE_DATA_REQUESTING: 659 mlog(ML_ERROR, "bad ndata state for " 660 "node %u: state=%d\n", 661 ndata->node_num, ndata->state); 662 BUG(); 663 break; 664 case DLM_RECO_NODE_DATA_DEAD: 665 mlog(0, "node %u died after " 666 "requesting recovery info for " 667 "node %u\n", ndata->node_num, 668 dead_node); 669 break; 670 case DLM_RECO_NODE_DATA_RECEIVING: 671 case DLM_RECO_NODE_DATA_REQUESTED: 672 mlog(0, "%s: node %u still in state %s\n", 673 dlm->name, ndata->node_num, 674 ndata->state==DLM_RECO_NODE_DATA_RECEIVING ? 675 "receiving" : "requested"); 676 all_nodes_done = 0; 677 break; 678 case DLM_RECO_NODE_DATA_DONE: 679 mlog(0, "%s: node %u state is done\n", 680 dlm->name, ndata->node_num); 681 break; 682 case DLM_RECO_NODE_DATA_FINALIZE_SENT: 683 mlog(0, "%s: node %u state is finalize\n", 684 dlm->name, ndata->node_num); 685 break; 686 } 687 } 688 spin_unlock(&dlm_reco_state_lock); 689 690 mlog(0, "pass #%d, all_nodes_done?: %s\n", ++pass, 691 all_nodes_done?"yes":"no"); 692 if (all_nodes_done) { 693 int ret; 694 695 /* all nodes are now in DLM_RECO_NODE_DATA_DONE state 696 * just send a finalize message to everyone and 697 * clean up */ 698 mlog(0, "all nodes are done! send finalize\n"); 699 ret = dlm_send_finalize_reco_message(dlm); 700 if (ret < 0) 701 mlog_errno(ret); 702 703 spin_lock(&dlm->spinlock); 704 dlm_finish_local_lockres_recovery(dlm, dead_node, 705 dlm->node_num); 706 spin_unlock(&dlm->spinlock); 707 mlog(0, "should be done with recovery!\n"); 708 709 mlog(0, "finishing recovery of %s at %lu, " 710 "dead=%u, this=%u, new=%u\n", dlm->name, 711 jiffies, dlm->reco.dead_node, 712 dlm->node_num, dlm->reco.new_master); 713 destroy = 1; 714 status = 0; 715 /* rescan everything marked dirty along the way */ 716 dlm_kick_thread(dlm, NULL); 717 break; 718 } 719 /* wait to be signalled, with periodic timeout 720 * to check for node death */ 721 wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq, 722 kthread_should_stop(), 723 msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS)); 724 725 } 726 727 if (destroy) 728 dlm_destroy_recovery_area(dlm, dead_node); 729 730 return status; 731 } 732 733 static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node) 734 { 735 int num=0; 736 struct dlm_reco_node_data *ndata; 737 738 spin_lock(&dlm->spinlock); 739 memcpy(dlm->reco.node_map, dlm->domain_map, sizeof(dlm->domain_map)); 740 /* nodes can only be removed (by dying) after dropping 741 * this lock, and death will be trapped later, so this should do */ 742 spin_unlock(&dlm->spinlock); 743 744 while (1) { 745 num = find_next_bit (dlm->reco.node_map, O2NM_MAX_NODES, num); 746 if (num >= O2NM_MAX_NODES) { 747 break; 748 } 749 BUG_ON(num == dead_node); 750 751 ndata = kzalloc(sizeof(*ndata), GFP_NOFS); 752 if (!ndata) { 753 dlm_destroy_recovery_area(dlm, dead_node); 754 return -ENOMEM; 755 } 756 ndata->node_num = num; 757 ndata->state = DLM_RECO_NODE_DATA_INIT; 758 spin_lock(&dlm_reco_state_lock); 759 list_add_tail(&ndata->list, &dlm->reco.node_data); 760 spin_unlock(&dlm_reco_state_lock); 761 num++; 762 } 763 764 return 0; 765 } 766 767 static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm, u8 dead_node) 768 { 769 struct dlm_reco_node_data *ndata, *next; 770 LIST_HEAD(tmplist); 771 772 spin_lock(&dlm_reco_state_lock); 773 list_splice_init(&dlm->reco.node_data, &tmplist); 774 spin_unlock(&dlm_reco_state_lock); 775 776 list_for_each_entry_safe(ndata, next, &tmplist, list) { 777 list_del_init(&ndata->list); 778 kfree(ndata); 779 } 780 } 781 782 static int dlm_request_all_locks(struct dlm_ctxt *dlm, u8 request_from, 783 u8 dead_node) 784 { 785 struct dlm_lock_request lr; 786 enum dlm_status ret; 787 788 mlog(0, "\n"); 789 790 791 mlog(0, "dlm_request_all_locks: dead node is %u, sending request " 792 "to %u\n", dead_node, request_from); 793 794 memset(&lr, 0, sizeof(lr)); 795 lr.node_idx = dlm->node_num; 796 lr.dead_node = dead_node; 797 798 // send message 799 ret = DLM_NOLOCKMGR; 800 ret = o2net_send_message(DLM_LOCK_REQUEST_MSG, dlm->key, 801 &lr, sizeof(lr), request_from, NULL); 802 803 /* negative status is handled by caller */ 804 if (ret < 0) 805 mlog(ML_ERROR, "Error %d when sending message %u (key " 806 "0x%x) to node %u\n", ret, DLM_LOCK_REQUEST_MSG, 807 dlm->key, request_from); 808 809 // return from here, then 810 // sleep until all received or error 811 return ret; 812 813 } 814 815 int dlm_request_all_locks_handler(struct o2net_msg *msg, u32 len, void *data, 816 void **ret_data) 817 { 818 struct dlm_ctxt *dlm = data; 819 struct dlm_lock_request *lr = (struct dlm_lock_request *)msg->buf; 820 char *buf = NULL; 821 struct dlm_work_item *item = NULL; 822 823 if (!dlm_grab(dlm)) 824 return -EINVAL; 825 826 if (lr->dead_node != dlm->reco.dead_node) { 827 mlog(ML_ERROR, "%s: node %u sent dead_node=%u, but local " 828 "dead_node is %u\n", dlm->name, lr->node_idx, 829 lr->dead_node, dlm->reco.dead_node); 830 dlm_print_reco_node_status(dlm); 831 /* this is a hack */ 832 dlm_put(dlm); 833 return -ENOMEM; 834 } 835 BUG_ON(lr->dead_node != dlm->reco.dead_node); 836 837 item = kzalloc(sizeof(*item), GFP_NOFS); 838 if (!item) { 839 dlm_put(dlm); 840 return -ENOMEM; 841 } 842 843 /* this will get freed by dlm_request_all_locks_worker */ 844 buf = (char *) __get_free_page(GFP_NOFS); 845 if (!buf) { 846 kfree(item); 847 dlm_put(dlm); 848 return -ENOMEM; 849 } 850 851 /* queue up work for dlm_request_all_locks_worker */ 852 dlm_grab(dlm); /* get an extra ref for the work item */ 853 dlm_init_work_item(dlm, item, dlm_request_all_locks_worker, buf); 854 item->u.ral.reco_master = lr->node_idx; 855 item->u.ral.dead_node = lr->dead_node; 856 spin_lock(&dlm->work_lock); 857 list_add_tail(&item->list, &dlm->work_list); 858 spin_unlock(&dlm->work_lock); 859 queue_work(dlm->dlm_worker, &dlm->dispatched_work); 860 861 dlm_put(dlm); 862 return 0; 863 } 864 865 static void dlm_request_all_locks_worker(struct dlm_work_item *item, void *data) 866 { 867 struct dlm_migratable_lockres *mres; 868 struct dlm_lock_resource *res; 869 struct dlm_ctxt *dlm; 870 LIST_HEAD(resources); 871 int ret; 872 u8 dead_node, reco_master; 873 int skip_all_done = 0; 874 875 dlm = item->dlm; 876 dead_node = item->u.ral.dead_node; 877 reco_master = item->u.ral.reco_master; 878 mres = (struct dlm_migratable_lockres *)data; 879 880 mlog(0, "%s: recovery worker started, dead=%u, master=%u\n", 881 dlm->name, dead_node, reco_master); 882 883 if (dead_node != dlm->reco.dead_node || 884 reco_master != dlm->reco.new_master) { 885 /* worker could have been created before the recovery master 886 * died. if so, do not continue, but do not error. */ 887 if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) { 888 mlog(ML_NOTICE, "%s: will not send recovery state, " 889 "recovery master %u died, thread=(dead=%u,mas=%u)" 890 " current=(dead=%u,mas=%u)\n", dlm->name, 891 reco_master, dead_node, reco_master, 892 dlm->reco.dead_node, dlm->reco.new_master); 893 } else { 894 mlog(ML_NOTICE, "%s: reco state invalid: reco(dead=%u, " 895 "master=%u), request(dead=%u, master=%u)\n", 896 dlm->name, dlm->reco.dead_node, 897 dlm->reco.new_master, dead_node, reco_master); 898 } 899 goto leave; 900 } 901 902 /* lock resources should have already been moved to the 903 * dlm->reco.resources list. now move items from that list 904 * to a temp list if the dead owner matches. note that the 905 * whole cluster recovers only one node at a time, so we 906 * can safely move UNKNOWN lock resources for each recovery 907 * session. */ 908 dlm_move_reco_locks_to_list(dlm, &resources, dead_node); 909 910 /* now we can begin blasting lockreses without the dlm lock */ 911 912 /* any errors returned will be due to the new_master dying, 913 * the dlm_reco_thread should detect this */ 914 list_for_each_entry(res, &resources, recovering) { 915 ret = dlm_send_one_lockres(dlm, res, mres, reco_master, 916 DLM_MRES_RECOVERY); 917 if (ret < 0) { 918 mlog(ML_ERROR, "%s: node %u went down while sending " 919 "recovery state for dead node %u, ret=%d\n", dlm->name, 920 reco_master, dead_node, ret); 921 skip_all_done = 1; 922 break; 923 } 924 } 925 926 /* move the resources back to the list */ 927 spin_lock(&dlm->spinlock); 928 list_splice_init(&resources, &dlm->reco.resources); 929 spin_unlock(&dlm->spinlock); 930 931 if (!skip_all_done) { 932 ret = dlm_send_all_done_msg(dlm, dead_node, reco_master); 933 if (ret < 0) { 934 mlog(ML_ERROR, "%s: node %u went down while sending " 935 "recovery all-done for dead node %u, ret=%d\n", 936 dlm->name, reco_master, dead_node, ret); 937 } 938 } 939 leave: 940 free_page((unsigned long)data); 941 } 942 943 944 static int dlm_send_all_done_msg(struct dlm_ctxt *dlm, u8 dead_node, u8 send_to) 945 { 946 int ret, tmpret; 947 struct dlm_reco_data_done done_msg; 948 949 memset(&done_msg, 0, sizeof(done_msg)); 950 done_msg.node_idx = dlm->node_num; 951 done_msg.dead_node = dead_node; 952 mlog(0, "sending DATA DONE message to %u, " 953 "my node=%u, dead node=%u\n", send_to, done_msg.node_idx, 954 done_msg.dead_node); 955 956 ret = o2net_send_message(DLM_RECO_DATA_DONE_MSG, dlm->key, &done_msg, 957 sizeof(done_msg), send_to, &tmpret); 958 if (ret < 0) { 959 mlog(ML_ERROR, "Error %d when sending message %u (key " 960 "0x%x) to node %u\n", ret, DLM_RECO_DATA_DONE_MSG, 961 dlm->key, send_to); 962 if (!dlm_is_host_down(ret)) { 963 BUG(); 964 } 965 } else 966 ret = tmpret; 967 return ret; 968 } 969 970 971 int dlm_reco_data_done_handler(struct o2net_msg *msg, u32 len, void *data, 972 void **ret_data) 973 { 974 struct dlm_ctxt *dlm = data; 975 struct dlm_reco_data_done *done = (struct dlm_reco_data_done *)msg->buf; 976 struct dlm_reco_node_data *ndata = NULL; 977 int ret = -EINVAL; 978 979 if (!dlm_grab(dlm)) 980 return -EINVAL; 981 982 mlog(0, "got DATA DONE: dead_node=%u, reco.dead_node=%u, " 983 "node_idx=%u, this node=%u\n", done->dead_node, 984 dlm->reco.dead_node, done->node_idx, dlm->node_num); 985 986 mlog_bug_on_msg((done->dead_node != dlm->reco.dead_node), 987 "Got DATA DONE: dead_node=%u, reco.dead_node=%u, " 988 "node_idx=%u, this node=%u\n", done->dead_node, 989 dlm->reco.dead_node, done->node_idx, dlm->node_num); 990 991 spin_lock(&dlm_reco_state_lock); 992 list_for_each_entry(ndata, &dlm->reco.node_data, list) { 993 if (ndata->node_num != done->node_idx) 994 continue; 995 996 switch (ndata->state) { 997 /* should have moved beyond INIT but not to FINALIZE yet */ 998 case DLM_RECO_NODE_DATA_INIT: 999 case DLM_RECO_NODE_DATA_DEAD: 1000 case DLM_RECO_NODE_DATA_FINALIZE_SENT: 1001 mlog(ML_ERROR, "bad ndata state for node %u:" 1002 " state=%d\n", ndata->node_num, 1003 ndata->state); 1004 BUG(); 1005 break; 1006 /* these states are possible at this point, anywhere along 1007 * the line of recovery */ 1008 case DLM_RECO_NODE_DATA_DONE: 1009 case DLM_RECO_NODE_DATA_RECEIVING: 1010 case DLM_RECO_NODE_DATA_REQUESTED: 1011 case DLM_RECO_NODE_DATA_REQUESTING: 1012 mlog(0, "node %u is DONE sending " 1013 "recovery data!\n", 1014 ndata->node_num); 1015 1016 ndata->state = DLM_RECO_NODE_DATA_DONE; 1017 ret = 0; 1018 break; 1019 } 1020 } 1021 spin_unlock(&dlm_reco_state_lock); 1022 1023 /* wake the recovery thread, some node is done */ 1024 if (!ret) 1025 dlm_kick_recovery_thread(dlm); 1026 1027 if (ret < 0) 1028 mlog(ML_ERROR, "failed to find recovery node data for node " 1029 "%u\n", done->node_idx); 1030 dlm_put(dlm); 1031 1032 mlog(0, "leaving reco data done handler, ret=%d\n", ret); 1033 return ret; 1034 } 1035 1036 static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm, 1037 struct list_head *list, 1038 u8 dead_node) 1039 { 1040 struct dlm_lock_resource *res, *next; 1041 struct dlm_lock *lock; 1042 1043 spin_lock(&dlm->spinlock); 1044 list_for_each_entry_safe(res, next, &dlm->reco.resources, recovering) { 1045 /* always prune any $RECOVERY entries for dead nodes, 1046 * otherwise hangs can occur during later recovery */ 1047 if (dlm_is_recovery_lock(res->lockname.name, 1048 res->lockname.len)) { 1049 spin_lock(&res->spinlock); 1050 list_for_each_entry(lock, &res->granted, list) { 1051 if (lock->ml.node == dead_node) { 1052 mlog(0, "AHA! there was " 1053 "a $RECOVERY lock for dead " 1054 "node %u (%s)!\n", 1055 dead_node, dlm->name); 1056 list_del_init(&lock->list); 1057 dlm_lock_put(lock); 1058 break; 1059 } 1060 } 1061 spin_unlock(&res->spinlock); 1062 continue; 1063 } 1064 1065 if (res->owner == dead_node) { 1066 mlog(0, "found lockres owned by dead node while " 1067 "doing recovery for node %u. sending it.\n", 1068 dead_node); 1069 list_move_tail(&res->recovering, list); 1070 } else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) { 1071 mlog(0, "found UNKNOWN owner while doing recovery " 1072 "for node %u. sending it.\n", dead_node); 1073 list_move_tail(&res->recovering, list); 1074 } 1075 } 1076 spin_unlock(&dlm->spinlock); 1077 } 1078 1079 static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res) 1080 { 1081 int total_locks = 0; 1082 struct list_head *iter, *queue = &res->granted; 1083 int i; 1084 1085 for (i=0; i<3; i++) { 1086 list_for_each(iter, queue) 1087 total_locks++; 1088 queue++; 1089 } 1090 return total_locks; 1091 } 1092 1093 1094 static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm, 1095 struct dlm_migratable_lockres *mres, 1096 u8 send_to, 1097 struct dlm_lock_resource *res, 1098 int total_locks) 1099 { 1100 u64 mig_cookie = be64_to_cpu(mres->mig_cookie); 1101 int mres_total_locks = be32_to_cpu(mres->total_locks); 1102 int sz, ret = 0, status = 0; 1103 u8 orig_flags = mres->flags, 1104 orig_master = mres->master; 1105 1106 BUG_ON(mres->num_locks > DLM_MAX_MIGRATABLE_LOCKS); 1107 if (!mres->num_locks) 1108 return 0; 1109 1110 sz = sizeof(struct dlm_migratable_lockres) + 1111 (mres->num_locks * sizeof(struct dlm_migratable_lock)); 1112 1113 /* add an all-done flag if we reached the last lock */ 1114 orig_flags = mres->flags; 1115 BUG_ON(total_locks > mres_total_locks); 1116 if (total_locks == mres_total_locks) 1117 mres->flags |= DLM_MRES_ALL_DONE; 1118 1119 mlog(0, "%s:%.*s: sending mig lockres (%s) to %u\n", 1120 dlm->name, res->lockname.len, res->lockname.name, 1121 orig_flags & DLM_MRES_MIGRATION ? "migration" : "recovery", 1122 send_to); 1123 1124 /* send it */ 1125 ret = o2net_send_message(DLM_MIG_LOCKRES_MSG, dlm->key, mres, 1126 sz, send_to, &status); 1127 if (ret < 0) { 1128 /* XXX: negative status is not handled. 1129 * this will end up killing this node. */ 1130 mlog(ML_ERROR, "Error %d when sending message %u (key " 1131 "0x%x) to node %u\n", ret, DLM_MIG_LOCKRES_MSG, 1132 dlm->key, send_to); 1133 } else { 1134 /* might get an -ENOMEM back here */ 1135 ret = status; 1136 if (ret < 0) { 1137 mlog_errno(ret); 1138 1139 if (ret == -EFAULT) { 1140 mlog(ML_ERROR, "node %u told me to kill " 1141 "myself!\n", send_to); 1142 BUG(); 1143 } 1144 } 1145 } 1146 1147 /* zero and reinit the message buffer */ 1148 dlm_init_migratable_lockres(mres, res->lockname.name, 1149 res->lockname.len, mres_total_locks, 1150 mig_cookie, orig_flags, orig_master); 1151 return ret; 1152 } 1153 1154 static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres, 1155 const char *lockname, int namelen, 1156 int total_locks, u64 cookie, 1157 u8 flags, u8 master) 1158 { 1159 /* mres here is one full page */ 1160 clear_page(mres); 1161 mres->lockname_len = namelen; 1162 memcpy(mres->lockname, lockname, namelen); 1163 mres->num_locks = 0; 1164 mres->total_locks = cpu_to_be32(total_locks); 1165 mres->mig_cookie = cpu_to_be64(cookie); 1166 mres->flags = flags; 1167 mres->master = master; 1168 } 1169 1170 static void dlm_prepare_lvb_for_migration(struct dlm_lock *lock, 1171 struct dlm_migratable_lockres *mres, 1172 int queue) 1173 { 1174 if (!lock->lksb) 1175 return; 1176 1177 /* Ignore lvb in all locks in the blocked list */ 1178 if (queue == DLM_BLOCKED_LIST) 1179 return; 1180 1181 /* Only consider lvbs in locks with granted EX or PR lock levels */ 1182 if (lock->ml.type != LKM_EXMODE && lock->ml.type != LKM_PRMODE) 1183 return; 1184 1185 if (dlm_lvb_is_empty(mres->lvb)) { 1186 memcpy(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN); 1187 return; 1188 } 1189 1190 /* Ensure the lvb copied for migration matches in other valid locks */ 1191 if (!memcmp(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN)) 1192 return; 1193 1194 mlog(ML_ERROR, "Mismatched lvb in lock cookie=%u:%llu, name=%.*s, " 1195 "node=%u\n", 1196 dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)), 1197 dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)), 1198 lock->lockres->lockname.len, lock->lockres->lockname.name, 1199 lock->ml.node); 1200 dlm_print_one_lock_resource(lock->lockres); 1201 BUG(); 1202 } 1203 1204 /* returns 1 if this lock fills the network structure, 1205 * 0 otherwise */ 1206 static int dlm_add_lock_to_array(struct dlm_lock *lock, 1207 struct dlm_migratable_lockres *mres, int queue) 1208 { 1209 struct dlm_migratable_lock *ml; 1210 int lock_num = mres->num_locks; 1211 1212 ml = &(mres->ml[lock_num]); 1213 ml->cookie = lock->ml.cookie; 1214 ml->type = lock->ml.type; 1215 ml->convert_type = lock->ml.convert_type; 1216 ml->highest_blocked = lock->ml.highest_blocked; 1217 ml->list = queue; 1218 if (lock->lksb) { 1219 ml->flags = lock->lksb->flags; 1220 dlm_prepare_lvb_for_migration(lock, mres, queue); 1221 } 1222 ml->node = lock->ml.node; 1223 mres->num_locks++; 1224 /* we reached the max, send this network message */ 1225 if (mres->num_locks == DLM_MAX_MIGRATABLE_LOCKS) 1226 return 1; 1227 return 0; 1228 } 1229 1230 static void dlm_add_dummy_lock(struct dlm_ctxt *dlm, 1231 struct dlm_migratable_lockres *mres) 1232 { 1233 struct dlm_lock dummy; 1234 memset(&dummy, 0, sizeof(dummy)); 1235 dummy.ml.cookie = 0; 1236 dummy.ml.type = LKM_IVMODE; 1237 dummy.ml.convert_type = LKM_IVMODE; 1238 dummy.ml.highest_blocked = LKM_IVMODE; 1239 dummy.lksb = NULL; 1240 dummy.ml.node = dlm->node_num; 1241 dlm_add_lock_to_array(&dummy, mres, DLM_BLOCKED_LIST); 1242 } 1243 1244 static inline int dlm_is_dummy_lock(struct dlm_ctxt *dlm, 1245 struct dlm_migratable_lock *ml, 1246 u8 *nodenum) 1247 { 1248 if (unlikely(ml->cookie == 0 && 1249 ml->type == LKM_IVMODE && 1250 ml->convert_type == LKM_IVMODE && 1251 ml->highest_blocked == LKM_IVMODE && 1252 ml->list == DLM_BLOCKED_LIST)) { 1253 *nodenum = ml->node; 1254 return 1; 1255 } 1256 return 0; 1257 } 1258 1259 int dlm_send_one_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res, 1260 struct dlm_migratable_lockres *mres, 1261 u8 send_to, u8 flags) 1262 { 1263 struct list_head *queue; 1264 int total_locks, i; 1265 u64 mig_cookie = 0; 1266 struct dlm_lock *lock; 1267 int ret = 0; 1268 1269 BUG_ON(!(flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION))); 1270 1271 mlog(0, "sending to %u\n", send_to); 1272 1273 total_locks = dlm_num_locks_in_lockres(res); 1274 if (total_locks > DLM_MAX_MIGRATABLE_LOCKS) { 1275 /* rare, but possible */ 1276 mlog(0, "argh. lockres has %d locks. this will " 1277 "require more than one network packet to " 1278 "migrate\n", total_locks); 1279 mig_cookie = dlm_get_next_mig_cookie(); 1280 } 1281 1282 dlm_init_migratable_lockres(mres, res->lockname.name, 1283 res->lockname.len, total_locks, 1284 mig_cookie, flags, res->owner); 1285 1286 total_locks = 0; 1287 for (i=DLM_GRANTED_LIST; i<=DLM_BLOCKED_LIST; i++) { 1288 queue = dlm_list_idx_to_ptr(res, i); 1289 list_for_each_entry(lock, queue, list) { 1290 /* add another lock. */ 1291 total_locks++; 1292 if (!dlm_add_lock_to_array(lock, mres, i)) 1293 continue; 1294 1295 /* this filled the lock message, 1296 * we must send it immediately. */ 1297 ret = dlm_send_mig_lockres_msg(dlm, mres, send_to, 1298 res, total_locks); 1299 if (ret < 0) 1300 goto error; 1301 } 1302 } 1303 if (total_locks == 0) { 1304 /* send a dummy lock to indicate a mastery reference only */ 1305 mlog(0, "%s:%.*s: sending dummy lock to %u, %s\n", 1306 dlm->name, res->lockname.len, res->lockname.name, 1307 send_to, flags & DLM_MRES_RECOVERY ? "recovery" : 1308 "migration"); 1309 dlm_add_dummy_lock(dlm, mres); 1310 } 1311 /* flush any remaining locks */ 1312 ret = dlm_send_mig_lockres_msg(dlm, mres, send_to, res, total_locks); 1313 if (ret < 0) 1314 goto error; 1315 return ret; 1316 1317 error: 1318 mlog(ML_ERROR, "%s: dlm_send_mig_lockres_msg returned %d\n", 1319 dlm->name, ret); 1320 if (!dlm_is_host_down(ret)) 1321 BUG(); 1322 mlog(0, "%s: node %u went down while sending %s " 1323 "lockres %.*s\n", dlm->name, send_to, 1324 flags & DLM_MRES_RECOVERY ? "recovery" : "migration", 1325 res->lockname.len, res->lockname.name); 1326 return ret; 1327 } 1328 1329 1330 1331 /* 1332 * this message will contain no more than one page worth of 1333 * recovery data, and it will work on only one lockres. 1334 * there may be many locks in this page, and we may need to wait 1335 * for additional packets to complete all the locks (rare, but 1336 * possible). 1337 */ 1338 /* 1339 * NOTE: the allocation error cases here are scary 1340 * we really cannot afford to fail an alloc in recovery 1341 * do we spin? returning an error only delays the problem really 1342 */ 1343 1344 int dlm_mig_lockres_handler(struct o2net_msg *msg, u32 len, void *data, 1345 void **ret_data) 1346 { 1347 struct dlm_ctxt *dlm = data; 1348 struct dlm_migratable_lockres *mres = 1349 (struct dlm_migratable_lockres *)msg->buf; 1350 int ret = 0; 1351 u8 real_master; 1352 u8 extra_refs = 0; 1353 char *buf = NULL; 1354 struct dlm_work_item *item = NULL; 1355 struct dlm_lock_resource *res = NULL; 1356 1357 if (!dlm_grab(dlm)) 1358 return -EINVAL; 1359 1360 BUG_ON(!(mres->flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION))); 1361 1362 real_master = mres->master; 1363 if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) { 1364 /* cannot migrate a lockres with no master */ 1365 BUG_ON(!(mres->flags & DLM_MRES_RECOVERY)); 1366 } 1367 1368 mlog(0, "%s message received from node %u\n", 1369 (mres->flags & DLM_MRES_RECOVERY) ? 1370 "recovery" : "migration", mres->master); 1371 if (mres->flags & DLM_MRES_ALL_DONE) 1372 mlog(0, "all done flag. all lockres data received!\n"); 1373 1374 ret = -ENOMEM; 1375 buf = kmalloc(be16_to_cpu(msg->data_len), GFP_NOFS); 1376 item = kzalloc(sizeof(*item), GFP_NOFS); 1377 if (!buf || !item) 1378 goto leave; 1379 1380 /* lookup the lock to see if we have a secondary queue for this 1381 * already... just add the locks in and this will have its owner 1382 * and RECOVERY flag changed when it completes. */ 1383 res = dlm_lookup_lockres(dlm, mres->lockname, mres->lockname_len); 1384 if (res) { 1385 /* this will get a ref on res */ 1386 /* mark it as recovering/migrating and hash it */ 1387 spin_lock(&res->spinlock); 1388 if (mres->flags & DLM_MRES_RECOVERY) { 1389 res->state |= DLM_LOCK_RES_RECOVERING; 1390 } else { 1391 if (res->state & DLM_LOCK_RES_MIGRATING) { 1392 /* this is at least the second 1393 * lockres message */ 1394 mlog(0, "lock %.*s is already migrating\n", 1395 mres->lockname_len, 1396 mres->lockname); 1397 } else if (res->state & DLM_LOCK_RES_RECOVERING) { 1398 /* caller should BUG */ 1399 mlog(ML_ERROR, "node is attempting to migrate " 1400 "lock %.*s, but marked as recovering!\n", 1401 mres->lockname_len, mres->lockname); 1402 ret = -EFAULT; 1403 spin_unlock(&res->spinlock); 1404 goto leave; 1405 } 1406 res->state |= DLM_LOCK_RES_MIGRATING; 1407 } 1408 spin_unlock(&res->spinlock); 1409 } else { 1410 /* need to allocate, just like if it was 1411 * mastered here normally */ 1412 res = dlm_new_lockres(dlm, mres->lockname, mres->lockname_len); 1413 if (!res) 1414 goto leave; 1415 1416 /* to match the ref that we would have gotten if 1417 * dlm_lookup_lockres had succeeded */ 1418 dlm_lockres_get(res); 1419 1420 /* mark it as recovering/migrating and hash it */ 1421 if (mres->flags & DLM_MRES_RECOVERY) 1422 res->state |= DLM_LOCK_RES_RECOVERING; 1423 else 1424 res->state |= DLM_LOCK_RES_MIGRATING; 1425 1426 spin_lock(&dlm->spinlock); 1427 __dlm_insert_lockres(dlm, res); 1428 spin_unlock(&dlm->spinlock); 1429 1430 /* Add an extra ref for this lock-less lockres lest the 1431 * dlm_thread purges it before we get the chance to add 1432 * locks to it */ 1433 dlm_lockres_get(res); 1434 1435 /* There are three refs that need to be put. 1436 * 1. Taken above. 1437 * 2. kref_init in dlm_new_lockres()->dlm_init_lockres(). 1438 * 3. dlm_lookup_lockres() 1439 * The first one is handled at the end of this function. The 1440 * other two are handled in the worker thread after locks have 1441 * been attached. Yes, we don't wait for purge time to match 1442 * kref_init. The lockres will still have atleast one ref 1443 * added because it is in the hash __dlm_insert_lockres() */ 1444 extra_refs++; 1445 1446 /* now that the new lockres is inserted, 1447 * make it usable by other processes */ 1448 spin_lock(&res->spinlock); 1449 res->state &= ~DLM_LOCK_RES_IN_PROGRESS; 1450 spin_unlock(&res->spinlock); 1451 wake_up(&res->wq); 1452 } 1453 1454 /* at this point we have allocated everything we need, 1455 * and we have a hashed lockres with an extra ref and 1456 * the proper res->state flags. */ 1457 ret = 0; 1458 spin_lock(&res->spinlock); 1459 /* drop this either when master requery finds a different master 1460 * or when a lock is added by the recovery worker */ 1461 dlm_lockres_grab_inflight_ref(dlm, res); 1462 if (mres->master == DLM_LOCK_RES_OWNER_UNKNOWN) { 1463 /* migration cannot have an unknown master */ 1464 BUG_ON(!(mres->flags & DLM_MRES_RECOVERY)); 1465 mlog(0, "recovery has passed me a lockres with an " 1466 "unknown owner.. will need to requery: " 1467 "%.*s\n", mres->lockname_len, mres->lockname); 1468 } else { 1469 /* take a reference now to pin the lockres, drop it 1470 * when locks are added in the worker */ 1471 dlm_change_lockres_owner(dlm, res, dlm->node_num); 1472 } 1473 spin_unlock(&res->spinlock); 1474 1475 /* queue up work for dlm_mig_lockres_worker */ 1476 dlm_grab(dlm); /* get an extra ref for the work item */ 1477 memcpy(buf, msg->buf, be16_to_cpu(msg->data_len)); /* copy the whole message */ 1478 dlm_init_work_item(dlm, item, dlm_mig_lockres_worker, buf); 1479 item->u.ml.lockres = res; /* already have a ref */ 1480 item->u.ml.real_master = real_master; 1481 item->u.ml.extra_ref = extra_refs; 1482 spin_lock(&dlm->work_lock); 1483 list_add_tail(&item->list, &dlm->work_list); 1484 spin_unlock(&dlm->work_lock); 1485 queue_work(dlm->dlm_worker, &dlm->dispatched_work); 1486 1487 leave: 1488 /* One extra ref taken needs to be put here */ 1489 if (extra_refs) 1490 dlm_lockres_put(res); 1491 1492 dlm_put(dlm); 1493 if (ret < 0) { 1494 if (buf) 1495 kfree(buf); 1496 if (item) 1497 kfree(item); 1498 mlog_errno(ret); 1499 } 1500 1501 return ret; 1502 } 1503 1504 1505 static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data) 1506 { 1507 struct dlm_ctxt *dlm; 1508 struct dlm_migratable_lockres *mres; 1509 int ret = 0; 1510 struct dlm_lock_resource *res; 1511 u8 real_master; 1512 u8 extra_ref; 1513 1514 dlm = item->dlm; 1515 mres = (struct dlm_migratable_lockres *)data; 1516 1517 res = item->u.ml.lockres; 1518 real_master = item->u.ml.real_master; 1519 extra_ref = item->u.ml.extra_ref; 1520 1521 if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) { 1522 /* this case is super-rare. only occurs if 1523 * node death happens during migration. */ 1524 again: 1525 ret = dlm_lockres_master_requery(dlm, res, &real_master); 1526 if (ret < 0) { 1527 mlog(0, "dlm_lockres_master_requery ret=%d\n", 1528 ret); 1529 goto again; 1530 } 1531 if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) { 1532 mlog(0, "lockres %.*s not claimed. " 1533 "this node will take it.\n", 1534 res->lockname.len, res->lockname.name); 1535 } else { 1536 spin_lock(&res->spinlock); 1537 dlm_lockres_drop_inflight_ref(dlm, res); 1538 spin_unlock(&res->spinlock); 1539 mlog(0, "master needs to respond to sender " 1540 "that node %u still owns %.*s\n", 1541 real_master, res->lockname.len, 1542 res->lockname.name); 1543 /* cannot touch this lockres */ 1544 goto leave; 1545 } 1546 } 1547 1548 ret = dlm_process_recovery_data(dlm, res, mres); 1549 if (ret < 0) 1550 mlog(0, "dlm_process_recovery_data returned %d\n", ret); 1551 else 1552 mlog(0, "dlm_process_recovery_data succeeded\n"); 1553 1554 if ((mres->flags & (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) == 1555 (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) { 1556 ret = dlm_finish_migration(dlm, res, mres->master); 1557 if (ret < 0) 1558 mlog_errno(ret); 1559 } 1560 1561 leave: 1562 /* See comment in dlm_mig_lockres_handler() */ 1563 if (res) { 1564 if (extra_ref) 1565 dlm_lockres_put(res); 1566 dlm_lockres_put(res); 1567 } 1568 kfree(data); 1569 } 1570 1571 1572 1573 static int dlm_lockres_master_requery(struct dlm_ctxt *dlm, 1574 struct dlm_lock_resource *res, 1575 u8 *real_master) 1576 { 1577 struct dlm_node_iter iter; 1578 int nodenum; 1579 int ret = 0; 1580 1581 *real_master = DLM_LOCK_RES_OWNER_UNKNOWN; 1582 1583 /* we only reach here if one of the two nodes in a 1584 * migration died while the migration was in progress. 1585 * at this point we need to requery the master. we 1586 * know that the new_master got as far as creating 1587 * an mle on at least one node, but we do not know 1588 * if any nodes had actually cleared the mle and set 1589 * the master to the new_master. the old master 1590 * is supposed to set the owner to UNKNOWN in the 1591 * event of a new_master death, so the only possible 1592 * responses that we can get from nodes here are 1593 * that the master is new_master, or that the master 1594 * is UNKNOWN. 1595 * if all nodes come back with UNKNOWN then we know 1596 * the lock needs remastering here. 1597 * if any node comes back with a valid master, check 1598 * to see if that master is the one that we are 1599 * recovering. if so, then the new_master died and 1600 * we need to remaster this lock. if not, then the 1601 * new_master survived and that node will respond to 1602 * other nodes about the owner. 1603 * if there is an owner, this node needs to dump this 1604 * lockres and alert the sender that this lockres 1605 * was rejected. */ 1606 spin_lock(&dlm->spinlock); 1607 dlm_node_iter_init(dlm->domain_map, &iter); 1608 spin_unlock(&dlm->spinlock); 1609 1610 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) { 1611 /* do not send to self */ 1612 if (nodenum == dlm->node_num) 1613 continue; 1614 ret = dlm_do_master_requery(dlm, res, nodenum, real_master); 1615 if (ret < 0) { 1616 mlog_errno(ret); 1617 if (!dlm_is_host_down(ret)) 1618 BUG(); 1619 /* host is down, so answer for that node would be 1620 * DLM_LOCK_RES_OWNER_UNKNOWN. continue. */ 1621 } 1622 if (*real_master != DLM_LOCK_RES_OWNER_UNKNOWN) { 1623 mlog(0, "lock master is %u\n", *real_master); 1624 break; 1625 } 1626 } 1627 return ret; 1628 } 1629 1630 1631 int dlm_do_master_requery(struct dlm_ctxt *dlm, struct dlm_lock_resource *res, 1632 u8 nodenum, u8 *real_master) 1633 { 1634 int ret = -EINVAL; 1635 struct dlm_master_requery req; 1636 int status = DLM_LOCK_RES_OWNER_UNKNOWN; 1637 1638 memset(&req, 0, sizeof(req)); 1639 req.node_idx = dlm->node_num; 1640 req.namelen = res->lockname.len; 1641 memcpy(req.name, res->lockname.name, res->lockname.len); 1642 1643 ret = o2net_send_message(DLM_MASTER_REQUERY_MSG, dlm->key, 1644 &req, sizeof(req), nodenum, &status); 1645 /* XXX: negative status not handled properly here. */ 1646 if (ret < 0) 1647 mlog(ML_ERROR, "Error %d when sending message %u (key " 1648 "0x%x) to node %u\n", ret, DLM_MASTER_REQUERY_MSG, 1649 dlm->key, nodenum); 1650 else { 1651 BUG_ON(status < 0); 1652 BUG_ON(status > DLM_LOCK_RES_OWNER_UNKNOWN); 1653 *real_master = (u8) (status & 0xff); 1654 mlog(0, "node %u responded to master requery with %u\n", 1655 nodenum, *real_master); 1656 ret = 0; 1657 } 1658 return ret; 1659 } 1660 1661 1662 /* this function cannot error, so unless the sending 1663 * or receiving of the message failed, the owner can 1664 * be trusted */ 1665 int dlm_master_requery_handler(struct o2net_msg *msg, u32 len, void *data, 1666 void **ret_data) 1667 { 1668 struct dlm_ctxt *dlm = data; 1669 struct dlm_master_requery *req = (struct dlm_master_requery *)msg->buf; 1670 struct dlm_lock_resource *res = NULL; 1671 unsigned int hash; 1672 int master = DLM_LOCK_RES_OWNER_UNKNOWN; 1673 u32 flags = DLM_ASSERT_MASTER_REQUERY; 1674 1675 if (!dlm_grab(dlm)) { 1676 /* since the domain has gone away on this 1677 * node, the proper response is UNKNOWN */ 1678 return master; 1679 } 1680 1681 hash = dlm_lockid_hash(req->name, req->namelen); 1682 1683 spin_lock(&dlm->spinlock); 1684 res = __dlm_lookup_lockres(dlm, req->name, req->namelen, hash); 1685 if (res) { 1686 spin_lock(&res->spinlock); 1687 master = res->owner; 1688 if (master == dlm->node_num) { 1689 int ret = dlm_dispatch_assert_master(dlm, res, 1690 0, 0, flags); 1691 if (ret < 0) { 1692 mlog_errno(-ENOMEM); 1693 /* retry!? */ 1694 BUG(); 1695 } 1696 } else /* put.. incase we are not the master */ 1697 dlm_lockres_put(res); 1698 spin_unlock(&res->spinlock); 1699 } 1700 spin_unlock(&dlm->spinlock); 1701 1702 dlm_put(dlm); 1703 return master; 1704 } 1705 1706 static inline struct list_head * 1707 dlm_list_num_to_pointer(struct dlm_lock_resource *res, int list_num) 1708 { 1709 struct list_head *ret; 1710 BUG_ON(list_num < 0); 1711 BUG_ON(list_num > 2); 1712 ret = &(res->granted); 1713 ret += list_num; 1714 return ret; 1715 } 1716 /* TODO: do ast flush business 1717 * TODO: do MIGRATING and RECOVERING spinning 1718 */ 1719 1720 /* 1721 * NOTE about in-flight requests during migration: 1722 * 1723 * Before attempting the migrate, the master has marked the lockres as 1724 * MIGRATING and then flushed all of its pending ASTS. So any in-flight 1725 * requests either got queued before the MIGRATING flag got set, in which 1726 * case the lock data will reflect the change and a return message is on 1727 * the way, or the request failed to get in before MIGRATING got set. In 1728 * this case, the caller will be told to spin and wait for the MIGRATING 1729 * flag to be dropped, then recheck the master. 1730 * This holds true for the convert, cancel and unlock cases, and since lvb 1731 * updates are tied to these same messages, it applies to lvb updates as 1732 * well. For the lock case, there is no way a lock can be on the master 1733 * queue and not be on the secondary queue since the lock is always added 1734 * locally first. This means that the new target node will never be sent 1735 * a lock that he doesn't already have on the list. 1736 * In total, this means that the local lock is correct and should not be 1737 * updated to match the one sent by the master. Any messages sent back 1738 * from the master before the MIGRATING flag will bring the lock properly 1739 * up-to-date, and the change will be ordered properly for the waiter. 1740 * We will *not* attempt to modify the lock underneath the waiter. 1741 */ 1742 1743 static int dlm_process_recovery_data(struct dlm_ctxt *dlm, 1744 struct dlm_lock_resource *res, 1745 struct dlm_migratable_lockres *mres) 1746 { 1747 struct dlm_migratable_lock *ml; 1748 struct list_head *queue; 1749 struct list_head *tmpq = NULL; 1750 struct dlm_lock *newlock = NULL; 1751 struct dlm_lockstatus *lksb = NULL; 1752 int ret = 0; 1753 int i, j, bad; 1754 struct dlm_lock *lock = NULL; 1755 u8 from = O2NM_MAX_NODES; 1756 unsigned int added = 0; 1757 __be64 c; 1758 1759 mlog(0, "running %d locks for this lockres\n", mres->num_locks); 1760 for (i=0; i<mres->num_locks; i++) { 1761 ml = &(mres->ml[i]); 1762 1763 if (dlm_is_dummy_lock(dlm, ml, &from)) { 1764 /* placeholder, just need to set the refmap bit */ 1765 BUG_ON(mres->num_locks != 1); 1766 mlog(0, "%s:%.*s: dummy lock for %u\n", 1767 dlm->name, mres->lockname_len, mres->lockname, 1768 from); 1769 spin_lock(&res->spinlock); 1770 dlm_lockres_set_refmap_bit(from, res); 1771 spin_unlock(&res->spinlock); 1772 added++; 1773 break; 1774 } 1775 BUG_ON(ml->highest_blocked != LKM_IVMODE); 1776 newlock = NULL; 1777 lksb = NULL; 1778 1779 queue = dlm_list_num_to_pointer(res, ml->list); 1780 tmpq = NULL; 1781 1782 /* if the lock is for the local node it needs to 1783 * be moved to the proper location within the queue. 1784 * do not allocate a new lock structure. */ 1785 if (ml->node == dlm->node_num) { 1786 /* MIGRATION ONLY! */ 1787 BUG_ON(!(mres->flags & DLM_MRES_MIGRATION)); 1788 1789 spin_lock(&res->spinlock); 1790 for (j = DLM_GRANTED_LIST; j <= DLM_BLOCKED_LIST; j++) { 1791 tmpq = dlm_list_idx_to_ptr(res, j); 1792 list_for_each_entry(lock, tmpq, list) { 1793 if (lock->ml.cookie != ml->cookie) 1794 lock = NULL; 1795 else 1796 break; 1797 } 1798 if (lock) 1799 break; 1800 } 1801 1802 /* lock is always created locally first, and 1803 * destroyed locally last. it must be on the list */ 1804 if (!lock) { 1805 c = ml->cookie; 1806 mlog(ML_ERROR, "Could not find local lock " 1807 "with cookie %u:%llu, node %u, " 1808 "list %u, flags 0x%x, type %d, " 1809 "conv %d, highest blocked %d\n", 1810 dlm_get_lock_cookie_node(be64_to_cpu(c)), 1811 dlm_get_lock_cookie_seq(be64_to_cpu(c)), 1812 ml->node, ml->list, ml->flags, ml->type, 1813 ml->convert_type, ml->highest_blocked); 1814 __dlm_print_one_lock_resource(res); 1815 BUG(); 1816 } 1817 1818 if (lock->ml.node != ml->node) { 1819 c = lock->ml.cookie; 1820 mlog(ML_ERROR, "Mismatched node# in lock " 1821 "cookie %u:%llu, name %.*s, node %u\n", 1822 dlm_get_lock_cookie_node(be64_to_cpu(c)), 1823 dlm_get_lock_cookie_seq(be64_to_cpu(c)), 1824 res->lockname.len, res->lockname.name, 1825 lock->ml.node); 1826 c = ml->cookie; 1827 mlog(ML_ERROR, "Migrate lock cookie %u:%llu, " 1828 "node %u, list %u, flags 0x%x, type %d, " 1829 "conv %d, highest blocked %d\n", 1830 dlm_get_lock_cookie_node(be64_to_cpu(c)), 1831 dlm_get_lock_cookie_seq(be64_to_cpu(c)), 1832 ml->node, ml->list, ml->flags, ml->type, 1833 ml->convert_type, ml->highest_blocked); 1834 __dlm_print_one_lock_resource(res); 1835 BUG(); 1836 } 1837 1838 if (tmpq != queue) { 1839 c = ml->cookie; 1840 mlog(0, "Lock cookie %u:%llu was on list %u " 1841 "instead of list %u for %.*s\n", 1842 dlm_get_lock_cookie_node(be64_to_cpu(c)), 1843 dlm_get_lock_cookie_seq(be64_to_cpu(c)), 1844 j, ml->list, res->lockname.len, 1845 res->lockname.name); 1846 __dlm_print_one_lock_resource(res); 1847 spin_unlock(&res->spinlock); 1848 continue; 1849 } 1850 1851 /* see NOTE above about why we do not update 1852 * to match the master here */ 1853 1854 /* move the lock to its proper place */ 1855 /* do not alter lock refcount. switching lists. */ 1856 list_move_tail(&lock->list, queue); 1857 spin_unlock(&res->spinlock); 1858 added++; 1859 1860 mlog(0, "just reordered a local lock!\n"); 1861 continue; 1862 } 1863 1864 /* lock is for another node. */ 1865 newlock = dlm_new_lock(ml->type, ml->node, 1866 be64_to_cpu(ml->cookie), NULL); 1867 if (!newlock) { 1868 ret = -ENOMEM; 1869 goto leave; 1870 } 1871 lksb = newlock->lksb; 1872 dlm_lock_attach_lockres(newlock, res); 1873 1874 if (ml->convert_type != LKM_IVMODE) { 1875 BUG_ON(queue != &res->converting); 1876 newlock->ml.convert_type = ml->convert_type; 1877 } 1878 lksb->flags |= (ml->flags & 1879 (DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB)); 1880 1881 if (ml->type == LKM_NLMODE) 1882 goto skip_lvb; 1883 1884 if (!dlm_lvb_is_empty(mres->lvb)) { 1885 if (lksb->flags & DLM_LKSB_PUT_LVB) { 1886 /* other node was trying to update 1887 * lvb when node died. recreate the 1888 * lksb with the updated lvb. */ 1889 memcpy(lksb->lvb, mres->lvb, DLM_LVB_LEN); 1890 /* the lock resource lvb update must happen 1891 * NOW, before the spinlock is dropped. 1892 * we no longer wait for the AST to update 1893 * the lvb. */ 1894 memcpy(res->lvb, mres->lvb, DLM_LVB_LEN); 1895 } else { 1896 /* otherwise, the node is sending its 1897 * most recent valid lvb info */ 1898 BUG_ON(ml->type != LKM_EXMODE && 1899 ml->type != LKM_PRMODE); 1900 if (!dlm_lvb_is_empty(res->lvb) && 1901 (ml->type == LKM_EXMODE || 1902 memcmp(res->lvb, mres->lvb, DLM_LVB_LEN))) { 1903 int i; 1904 mlog(ML_ERROR, "%s:%.*s: received bad " 1905 "lvb! type=%d\n", dlm->name, 1906 res->lockname.len, 1907 res->lockname.name, ml->type); 1908 printk("lockres lvb=["); 1909 for (i=0; i<DLM_LVB_LEN; i++) 1910 printk("%02x", res->lvb[i]); 1911 printk("]\nmigrated lvb=["); 1912 for (i=0; i<DLM_LVB_LEN; i++) 1913 printk("%02x", mres->lvb[i]); 1914 printk("]\n"); 1915 dlm_print_one_lock_resource(res); 1916 BUG(); 1917 } 1918 memcpy(res->lvb, mres->lvb, DLM_LVB_LEN); 1919 } 1920 } 1921 skip_lvb: 1922 1923 /* NOTE: 1924 * wrt lock queue ordering and recovery: 1925 * 1. order of locks on granted queue is 1926 * meaningless. 1927 * 2. order of locks on converting queue is 1928 * LOST with the node death. sorry charlie. 1929 * 3. order of locks on the blocked queue is 1930 * also LOST. 1931 * order of locks does not affect integrity, it 1932 * just means that a lock request may get pushed 1933 * back in line as a result of the node death. 1934 * also note that for a given node the lock order 1935 * for its secondary queue locks is preserved 1936 * relative to each other, but clearly *not* 1937 * preserved relative to locks from other nodes. 1938 */ 1939 bad = 0; 1940 spin_lock(&res->spinlock); 1941 list_for_each_entry(lock, queue, list) { 1942 if (lock->ml.cookie == ml->cookie) { 1943 c = lock->ml.cookie; 1944 mlog(ML_ERROR, "%s:%.*s: %u:%llu: lock already " 1945 "exists on this lockres!\n", dlm->name, 1946 res->lockname.len, res->lockname.name, 1947 dlm_get_lock_cookie_node(be64_to_cpu(c)), 1948 dlm_get_lock_cookie_seq(be64_to_cpu(c))); 1949 1950 mlog(ML_NOTICE, "sent lock: type=%d, conv=%d, " 1951 "node=%u, cookie=%u:%llu, queue=%d\n", 1952 ml->type, ml->convert_type, ml->node, 1953 dlm_get_lock_cookie_node(be64_to_cpu(ml->cookie)), 1954 dlm_get_lock_cookie_seq(be64_to_cpu(ml->cookie)), 1955 ml->list); 1956 1957 __dlm_print_one_lock_resource(res); 1958 bad = 1; 1959 break; 1960 } 1961 } 1962 if (!bad) { 1963 dlm_lock_get(newlock); 1964 list_add_tail(&newlock->list, queue); 1965 mlog(0, "%s:%.*s: added lock for node %u, " 1966 "setting refmap bit\n", dlm->name, 1967 res->lockname.len, res->lockname.name, ml->node); 1968 dlm_lockres_set_refmap_bit(ml->node, res); 1969 added++; 1970 } 1971 spin_unlock(&res->spinlock); 1972 } 1973 mlog(0, "done running all the locks\n"); 1974 1975 leave: 1976 /* balance the ref taken when the work was queued */ 1977 spin_lock(&res->spinlock); 1978 dlm_lockres_drop_inflight_ref(dlm, res); 1979 spin_unlock(&res->spinlock); 1980 1981 if (ret < 0) { 1982 mlog_errno(ret); 1983 if (newlock) 1984 dlm_lock_put(newlock); 1985 } 1986 1987 return ret; 1988 } 1989 1990 void dlm_move_lockres_to_recovery_list(struct dlm_ctxt *dlm, 1991 struct dlm_lock_resource *res) 1992 { 1993 int i; 1994 struct list_head *queue; 1995 struct dlm_lock *lock, *next; 1996 1997 assert_spin_locked(&dlm->spinlock); 1998 assert_spin_locked(&res->spinlock); 1999 res->state |= DLM_LOCK_RES_RECOVERING; 2000 if (!list_empty(&res->recovering)) { 2001 mlog(0, 2002 "Recovering res %s:%.*s, is already on recovery list!\n", 2003 dlm->name, res->lockname.len, res->lockname.name); 2004 list_del_init(&res->recovering); 2005 dlm_lockres_put(res); 2006 } 2007 /* We need to hold a reference while on the recovery list */ 2008 dlm_lockres_get(res); 2009 list_add_tail(&res->recovering, &dlm->reco.resources); 2010 2011 /* find any pending locks and put them back on proper list */ 2012 for (i=DLM_BLOCKED_LIST; i>=DLM_GRANTED_LIST; i--) { 2013 queue = dlm_list_idx_to_ptr(res, i); 2014 list_for_each_entry_safe(lock, next, queue, list) { 2015 dlm_lock_get(lock); 2016 if (lock->convert_pending) { 2017 /* move converting lock back to granted */ 2018 BUG_ON(i != DLM_CONVERTING_LIST); 2019 mlog(0, "node died with convert pending " 2020 "on %.*s. move back to granted list.\n", 2021 res->lockname.len, res->lockname.name); 2022 dlm_revert_pending_convert(res, lock); 2023 lock->convert_pending = 0; 2024 } else if (lock->lock_pending) { 2025 /* remove pending lock requests completely */ 2026 BUG_ON(i != DLM_BLOCKED_LIST); 2027 mlog(0, "node died with lock pending " 2028 "on %.*s. remove from blocked list and skip.\n", 2029 res->lockname.len, res->lockname.name); 2030 /* lock will be floating until ref in 2031 * dlmlock_remote is freed after the network 2032 * call returns. ok for it to not be on any 2033 * list since no ast can be called 2034 * (the master is dead). */ 2035 dlm_revert_pending_lock(res, lock); 2036 lock->lock_pending = 0; 2037 } else if (lock->unlock_pending) { 2038 /* if an unlock was in progress, treat as 2039 * if this had completed successfully 2040 * before sending this lock state to the 2041 * new master. note that the dlm_unlock 2042 * call is still responsible for calling 2043 * the unlockast. that will happen after 2044 * the network call times out. for now, 2045 * just move lists to prepare the new 2046 * recovery master. */ 2047 BUG_ON(i != DLM_GRANTED_LIST); 2048 mlog(0, "node died with unlock pending " 2049 "on %.*s. remove from blocked list and skip.\n", 2050 res->lockname.len, res->lockname.name); 2051 dlm_commit_pending_unlock(res, lock); 2052 lock->unlock_pending = 0; 2053 } else if (lock->cancel_pending) { 2054 /* if a cancel was in progress, treat as 2055 * if this had completed successfully 2056 * before sending this lock state to the 2057 * new master */ 2058 BUG_ON(i != DLM_CONVERTING_LIST); 2059 mlog(0, "node died with cancel pending " 2060 "on %.*s. move back to granted list.\n", 2061 res->lockname.len, res->lockname.name); 2062 dlm_commit_pending_cancel(res, lock); 2063 lock->cancel_pending = 0; 2064 } 2065 dlm_lock_put(lock); 2066 } 2067 } 2068 } 2069 2070 2071 2072 /* removes all recovered locks from the recovery list. 2073 * sets the res->owner to the new master. 2074 * unsets the RECOVERY flag and wakes waiters. */ 2075 static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm, 2076 u8 dead_node, u8 new_master) 2077 { 2078 int i; 2079 struct hlist_node *hash_iter; 2080 struct hlist_head *bucket; 2081 struct dlm_lock_resource *res, *next; 2082 2083 assert_spin_locked(&dlm->spinlock); 2084 2085 list_for_each_entry_safe(res, next, &dlm->reco.resources, recovering) { 2086 if (res->owner == dead_node) { 2087 list_del_init(&res->recovering); 2088 spin_lock(&res->spinlock); 2089 /* new_master has our reference from 2090 * the lock state sent during recovery */ 2091 dlm_change_lockres_owner(dlm, res, new_master); 2092 res->state &= ~DLM_LOCK_RES_RECOVERING; 2093 if (__dlm_lockres_has_locks(res)) 2094 __dlm_dirty_lockres(dlm, res); 2095 spin_unlock(&res->spinlock); 2096 wake_up(&res->wq); 2097 dlm_lockres_put(res); 2098 } 2099 } 2100 2101 /* this will become unnecessary eventually, but 2102 * for now we need to run the whole hash, clear 2103 * the RECOVERING state and set the owner 2104 * if necessary */ 2105 for (i = 0; i < DLM_HASH_BUCKETS; i++) { 2106 bucket = dlm_lockres_hash(dlm, i); 2107 hlist_for_each_entry(res, hash_iter, bucket, hash_node) { 2108 if (res->state & DLM_LOCK_RES_RECOVERING) { 2109 if (res->owner == dead_node) { 2110 mlog(0, "(this=%u) res %.*s owner=%u " 2111 "was not on recovering list, but " 2112 "clearing state anyway\n", 2113 dlm->node_num, res->lockname.len, 2114 res->lockname.name, new_master); 2115 } else if (res->owner == dlm->node_num) { 2116 mlog(0, "(this=%u) res %.*s owner=%u " 2117 "was not on recovering list, " 2118 "owner is THIS node, clearing\n", 2119 dlm->node_num, res->lockname.len, 2120 res->lockname.name, new_master); 2121 } else 2122 continue; 2123 2124 if (!list_empty(&res->recovering)) { 2125 mlog(0, "%s:%.*s: lockres was " 2126 "marked RECOVERING, owner=%u\n", 2127 dlm->name, res->lockname.len, 2128 res->lockname.name, res->owner); 2129 list_del_init(&res->recovering); 2130 dlm_lockres_put(res); 2131 } 2132 spin_lock(&res->spinlock); 2133 /* new_master has our reference from 2134 * the lock state sent during recovery */ 2135 dlm_change_lockres_owner(dlm, res, new_master); 2136 res->state &= ~DLM_LOCK_RES_RECOVERING; 2137 if (__dlm_lockres_has_locks(res)) 2138 __dlm_dirty_lockres(dlm, res); 2139 spin_unlock(&res->spinlock); 2140 wake_up(&res->wq); 2141 } 2142 } 2143 } 2144 } 2145 2146 static inline int dlm_lvb_needs_invalidation(struct dlm_lock *lock, int local) 2147 { 2148 if (local) { 2149 if (lock->ml.type != LKM_EXMODE && 2150 lock->ml.type != LKM_PRMODE) 2151 return 1; 2152 } else if (lock->ml.type == LKM_EXMODE) 2153 return 1; 2154 return 0; 2155 } 2156 2157 static void dlm_revalidate_lvb(struct dlm_ctxt *dlm, 2158 struct dlm_lock_resource *res, u8 dead_node) 2159 { 2160 struct list_head *queue; 2161 struct dlm_lock *lock; 2162 int blank_lvb = 0, local = 0; 2163 int i; 2164 u8 search_node; 2165 2166 assert_spin_locked(&dlm->spinlock); 2167 assert_spin_locked(&res->spinlock); 2168 2169 if (res->owner == dlm->node_num) 2170 /* if this node owned the lockres, and if the dead node 2171 * had an EX when he died, blank out the lvb */ 2172 search_node = dead_node; 2173 else { 2174 /* if this is a secondary lockres, and we had no EX or PR 2175 * locks granted, we can no longer trust the lvb */ 2176 search_node = dlm->node_num; 2177 local = 1; /* check local state for valid lvb */ 2178 } 2179 2180 for (i=DLM_GRANTED_LIST; i<=DLM_CONVERTING_LIST; i++) { 2181 queue = dlm_list_idx_to_ptr(res, i); 2182 list_for_each_entry(lock, queue, list) { 2183 if (lock->ml.node == search_node) { 2184 if (dlm_lvb_needs_invalidation(lock, local)) { 2185 /* zero the lksb lvb and lockres lvb */ 2186 blank_lvb = 1; 2187 memset(lock->lksb->lvb, 0, DLM_LVB_LEN); 2188 } 2189 } 2190 } 2191 } 2192 2193 if (blank_lvb) { 2194 mlog(0, "clearing %.*s lvb, dead node %u had EX\n", 2195 res->lockname.len, res->lockname.name, dead_node); 2196 memset(res->lvb, 0, DLM_LVB_LEN); 2197 } 2198 } 2199 2200 static void dlm_free_dead_locks(struct dlm_ctxt *dlm, 2201 struct dlm_lock_resource *res, u8 dead_node) 2202 { 2203 struct dlm_lock *lock, *next; 2204 unsigned int freed = 0; 2205 2206 /* this node is the lockres master: 2207 * 1) remove any stale locks for the dead node 2208 * 2) if the dead node had an EX when he died, blank out the lvb 2209 */ 2210 assert_spin_locked(&dlm->spinlock); 2211 assert_spin_locked(&res->spinlock); 2212 2213 /* We do two dlm_lock_put(). One for removing from list and the other is 2214 * to force the DLM_UNLOCK_FREE_LOCK action so as to free the locks */ 2215 2216 /* TODO: check pending_asts, pending_basts here */ 2217 list_for_each_entry_safe(lock, next, &res->granted, list) { 2218 if (lock->ml.node == dead_node) { 2219 list_del_init(&lock->list); 2220 dlm_lock_put(lock); 2221 /* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */ 2222 dlm_lock_put(lock); 2223 freed++; 2224 } 2225 } 2226 list_for_each_entry_safe(lock, next, &res->converting, list) { 2227 if (lock->ml.node == dead_node) { 2228 list_del_init(&lock->list); 2229 dlm_lock_put(lock); 2230 /* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */ 2231 dlm_lock_put(lock); 2232 freed++; 2233 } 2234 } 2235 list_for_each_entry_safe(lock, next, &res->blocked, list) { 2236 if (lock->ml.node == dead_node) { 2237 list_del_init(&lock->list); 2238 dlm_lock_put(lock); 2239 /* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */ 2240 dlm_lock_put(lock); 2241 freed++; 2242 } 2243 } 2244 2245 if (freed) { 2246 mlog(0, "%s:%.*s: freed %u locks for dead node %u, " 2247 "dropping ref from lockres\n", dlm->name, 2248 res->lockname.len, res->lockname.name, freed, dead_node); 2249 if(!test_bit(dead_node, res->refmap)) { 2250 mlog(ML_ERROR, "%s:%.*s: freed %u locks for dead node %u, " 2251 "but ref was not set\n", dlm->name, 2252 res->lockname.len, res->lockname.name, freed, dead_node); 2253 __dlm_print_one_lock_resource(res); 2254 } 2255 dlm_lockres_clear_refmap_bit(dead_node, res); 2256 } else if (test_bit(dead_node, res->refmap)) { 2257 mlog(0, "%s:%.*s: dead node %u had a ref, but had " 2258 "no locks and had not purged before dying\n", dlm->name, 2259 res->lockname.len, res->lockname.name, dead_node); 2260 dlm_lockres_clear_refmap_bit(dead_node, res); 2261 } 2262 2263 /* do not kick thread yet */ 2264 __dlm_dirty_lockres(dlm, res); 2265 } 2266 2267 /* if this node is the recovery master, and there are no 2268 * locks for a given lockres owned by this node that are in 2269 * either PR or EX mode, zero out the lvb before requesting. 2270 * 2271 */ 2272 2273 2274 static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node) 2275 { 2276 struct hlist_node *iter; 2277 struct dlm_lock_resource *res; 2278 int i; 2279 struct hlist_head *bucket; 2280 struct dlm_lock *lock; 2281 2282 2283 /* purge any stale mles */ 2284 dlm_clean_master_list(dlm, dead_node); 2285 2286 /* 2287 * now clean up all lock resources. there are two rules: 2288 * 2289 * 1) if the dead node was the master, move the lockres 2290 * to the recovering list. set the RECOVERING flag. 2291 * this lockres needs to be cleaned up before it can 2292 * be used further. 2293 * 2294 * 2) if this node was the master, remove all locks from 2295 * each of the lockres queues that were owned by the 2296 * dead node. once recovery finishes, the dlm thread 2297 * can be kicked again to see if any ASTs or BASTs 2298 * need to be fired as a result. 2299 */ 2300 for (i = 0; i < DLM_HASH_BUCKETS; i++) { 2301 bucket = dlm_lockres_hash(dlm, i); 2302 hlist_for_each_entry(res, iter, bucket, hash_node) { 2303 /* always prune any $RECOVERY entries for dead nodes, 2304 * otherwise hangs can occur during later recovery */ 2305 if (dlm_is_recovery_lock(res->lockname.name, 2306 res->lockname.len)) { 2307 spin_lock(&res->spinlock); 2308 list_for_each_entry(lock, &res->granted, list) { 2309 if (lock->ml.node == dead_node) { 2310 mlog(0, "AHA! there was " 2311 "a $RECOVERY lock for dead " 2312 "node %u (%s)!\n", 2313 dead_node, dlm->name); 2314 list_del_init(&lock->list); 2315 dlm_lock_put(lock); 2316 break; 2317 } 2318 } 2319 spin_unlock(&res->spinlock); 2320 continue; 2321 } 2322 spin_lock(&res->spinlock); 2323 /* zero the lvb if necessary */ 2324 dlm_revalidate_lvb(dlm, res, dead_node); 2325 if (res->owner == dead_node) { 2326 if (res->state & DLM_LOCK_RES_DROPPING_REF) { 2327 mlog(ML_NOTICE, "Ignore %.*s for " 2328 "recovery as it is being freed\n", 2329 res->lockname.len, 2330 res->lockname.name); 2331 } else 2332 dlm_move_lockres_to_recovery_list(dlm, 2333 res); 2334 2335 } else if (res->owner == dlm->node_num) { 2336 dlm_free_dead_locks(dlm, res, dead_node); 2337 __dlm_lockres_calc_usage(dlm, res); 2338 } 2339 spin_unlock(&res->spinlock); 2340 } 2341 } 2342 2343 } 2344 2345 static void __dlm_hb_node_down(struct dlm_ctxt *dlm, int idx) 2346 { 2347 assert_spin_locked(&dlm->spinlock); 2348 2349 if (dlm->reco.new_master == idx) { 2350 mlog(0, "%s: recovery master %d just died\n", 2351 dlm->name, idx); 2352 if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) { 2353 /* finalize1 was reached, so it is safe to clear 2354 * the new_master and dead_node. that recovery 2355 * is complete. */ 2356 mlog(0, "%s: dead master %d had reached " 2357 "finalize1 state, clearing\n", dlm->name, idx); 2358 dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE; 2359 __dlm_reset_recovery(dlm); 2360 } 2361 } 2362 2363 /* Clean up join state on node death. */ 2364 if (dlm->joining_node == idx) { 2365 mlog(0, "Clearing join state for node %u\n", idx); 2366 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN); 2367 } 2368 2369 /* check to see if the node is already considered dead */ 2370 if (!test_bit(idx, dlm->live_nodes_map)) { 2371 mlog(0, "for domain %s, node %d is already dead. " 2372 "another node likely did recovery already.\n", 2373 dlm->name, idx); 2374 return; 2375 } 2376 2377 /* check to see if we do not care about this node */ 2378 if (!test_bit(idx, dlm->domain_map)) { 2379 /* This also catches the case that we get a node down 2380 * but haven't joined the domain yet. */ 2381 mlog(0, "node %u already removed from domain!\n", idx); 2382 return; 2383 } 2384 2385 clear_bit(idx, dlm->live_nodes_map); 2386 2387 /* make sure local cleanup occurs before the heartbeat events */ 2388 if (!test_bit(idx, dlm->recovery_map)) 2389 dlm_do_local_recovery_cleanup(dlm, idx); 2390 2391 /* notify anything attached to the heartbeat events */ 2392 dlm_hb_event_notify_attached(dlm, idx, 0); 2393 2394 mlog(0, "node %u being removed from domain map!\n", idx); 2395 clear_bit(idx, dlm->domain_map); 2396 clear_bit(idx, dlm->exit_domain_map); 2397 /* wake up migration waiters if a node goes down. 2398 * perhaps later we can genericize this for other waiters. */ 2399 wake_up(&dlm->migration_wq); 2400 2401 if (test_bit(idx, dlm->recovery_map)) 2402 mlog(0, "domain %s, node %u already added " 2403 "to recovery map!\n", dlm->name, idx); 2404 else 2405 set_bit(idx, dlm->recovery_map); 2406 } 2407 2408 void dlm_hb_node_down_cb(struct o2nm_node *node, int idx, void *data) 2409 { 2410 struct dlm_ctxt *dlm = data; 2411 2412 if (!dlm_grab(dlm)) 2413 return; 2414 2415 /* 2416 * This will notify any dlm users that a node in our domain 2417 * went away without notifying us first. 2418 */ 2419 if (test_bit(idx, dlm->domain_map)) 2420 dlm_fire_domain_eviction_callbacks(dlm, idx); 2421 2422 spin_lock(&dlm->spinlock); 2423 __dlm_hb_node_down(dlm, idx); 2424 spin_unlock(&dlm->spinlock); 2425 2426 dlm_put(dlm); 2427 } 2428 2429 void dlm_hb_node_up_cb(struct o2nm_node *node, int idx, void *data) 2430 { 2431 struct dlm_ctxt *dlm = data; 2432 2433 if (!dlm_grab(dlm)) 2434 return; 2435 2436 spin_lock(&dlm->spinlock); 2437 set_bit(idx, dlm->live_nodes_map); 2438 /* do NOT notify mle attached to the heartbeat events. 2439 * new nodes are not interesting in mastery until joined. */ 2440 spin_unlock(&dlm->spinlock); 2441 2442 dlm_put(dlm); 2443 } 2444 2445 static void dlm_reco_ast(void *astdata) 2446 { 2447 struct dlm_ctxt *dlm = astdata; 2448 mlog(0, "ast for recovery lock fired!, this=%u, dlm=%s\n", 2449 dlm->node_num, dlm->name); 2450 } 2451 static void dlm_reco_bast(void *astdata, int blocked_type) 2452 { 2453 struct dlm_ctxt *dlm = astdata; 2454 mlog(0, "bast for recovery lock fired!, this=%u, dlm=%s\n", 2455 dlm->node_num, dlm->name); 2456 } 2457 static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st) 2458 { 2459 mlog(0, "unlockast for recovery lock fired!\n"); 2460 } 2461 2462 /* 2463 * dlm_pick_recovery_master will continually attempt to use 2464 * dlmlock() on the special "$RECOVERY" lockres with the 2465 * LKM_NOQUEUE flag to get an EX. every thread that enters 2466 * this function on each node racing to become the recovery 2467 * master will not stop attempting this until either: 2468 * a) this node gets the EX (and becomes the recovery master), 2469 * or b) dlm->reco.new_master gets set to some nodenum 2470 * != O2NM_INVALID_NODE_NUM (another node will do the reco). 2471 * so each time a recovery master is needed, the entire cluster 2472 * will sync at this point. if the new master dies, that will 2473 * be detected in dlm_do_recovery */ 2474 static int dlm_pick_recovery_master(struct dlm_ctxt *dlm) 2475 { 2476 enum dlm_status ret; 2477 struct dlm_lockstatus lksb; 2478 int status = -EINVAL; 2479 2480 mlog(0, "starting recovery of %s at %lu, dead=%u, this=%u\n", 2481 dlm->name, jiffies, dlm->reco.dead_node, dlm->node_num); 2482 again: 2483 memset(&lksb, 0, sizeof(lksb)); 2484 2485 ret = dlmlock(dlm, LKM_EXMODE, &lksb, LKM_NOQUEUE|LKM_RECOVERY, 2486 DLM_RECOVERY_LOCK_NAME, DLM_RECOVERY_LOCK_NAME_LEN, 2487 dlm_reco_ast, dlm, dlm_reco_bast); 2488 2489 mlog(0, "%s: dlmlock($RECOVERY) returned %d, lksb=%d\n", 2490 dlm->name, ret, lksb.status); 2491 2492 if (ret == DLM_NORMAL) { 2493 mlog(0, "dlm=%s dlmlock says I got it (this=%u)\n", 2494 dlm->name, dlm->node_num); 2495 2496 /* got the EX lock. check to see if another node 2497 * just became the reco master */ 2498 if (dlm_reco_master_ready(dlm)) { 2499 mlog(0, "%s: got reco EX lock, but %u will " 2500 "do the recovery\n", dlm->name, 2501 dlm->reco.new_master); 2502 status = -EEXIST; 2503 } else { 2504 status = 0; 2505 2506 /* see if recovery was already finished elsewhere */ 2507 spin_lock(&dlm->spinlock); 2508 if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) { 2509 status = -EINVAL; 2510 mlog(0, "%s: got reco EX lock, but " 2511 "node got recovered already\n", dlm->name); 2512 if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) { 2513 mlog(ML_ERROR, "%s: new master is %u " 2514 "but no dead node!\n", 2515 dlm->name, dlm->reco.new_master); 2516 BUG(); 2517 } 2518 } 2519 spin_unlock(&dlm->spinlock); 2520 } 2521 2522 /* if this node has actually become the recovery master, 2523 * set the master and send the messages to begin recovery */ 2524 if (!status) { 2525 mlog(0, "%s: dead=%u, this=%u, sending " 2526 "begin_reco now\n", dlm->name, 2527 dlm->reco.dead_node, dlm->node_num); 2528 status = dlm_send_begin_reco_message(dlm, 2529 dlm->reco.dead_node); 2530 /* this always succeeds */ 2531 BUG_ON(status); 2532 2533 /* set the new_master to this node */ 2534 spin_lock(&dlm->spinlock); 2535 dlm_set_reco_master(dlm, dlm->node_num); 2536 spin_unlock(&dlm->spinlock); 2537 } 2538 2539 /* recovery lock is a special case. ast will not get fired, 2540 * so just go ahead and unlock it. */ 2541 ret = dlmunlock(dlm, &lksb, 0, dlm_reco_unlock_ast, dlm); 2542 if (ret == DLM_DENIED) { 2543 mlog(0, "got DLM_DENIED, trying LKM_CANCEL\n"); 2544 ret = dlmunlock(dlm, &lksb, LKM_CANCEL, dlm_reco_unlock_ast, dlm); 2545 } 2546 if (ret != DLM_NORMAL) { 2547 /* this would really suck. this could only happen 2548 * if there was a network error during the unlock 2549 * because of node death. this means the unlock 2550 * is actually "done" and the lock structure is 2551 * even freed. we can continue, but only 2552 * because this specific lock name is special. */ 2553 mlog(ML_ERROR, "dlmunlock returned %d\n", ret); 2554 } 2555 } else if (ret == DLM_NOTQUEUED) { 2556 mlog(0, "dlm=%s dlmlock says another node got it (this=%u)\n", 2557 dlm->name, dlm->node_num); 2558 /* another node is master. wait on 2559 * reco.new_master != O2NM_INVALID_NODE_NUM 2560 * for at most one second */ 2561 wait_event_timeout(dlm->dlm_reco_thread_wq, 2562 dlm_reco_master_ready(dlm), 2563 msecs_to_jiffies(1000)); 2564 if (!dlm_reco_master_ready(dlm)) { 2565 mlog(0, "%s: reco master taking awhile\n", 2566 dlm->name); 2567 goto again; 2568 } 2569 /* another node has informed this one that it is reco master */ 2570 mlog(0, "%s: reco master %u is ready to recover %u\n", 2571 dlm->name, dlm->reco.new_master, dlm->reco.dead_node); 2572 status = -EEXIST; 2573 } else if (ret == DLM_RECOVERING) { 2574 mlog(0, "dlm=%s dlmlock says master node died (this=%u)\n", 2575 dlm->name, dlm->node_num); 2576 goto again; 2577 } else { 2578 struct dlm_lock_resource *res; 2579 2580 /* dlmlock returned something other than NOTQUEUED or NORMAL */ 2581 mlog(ML_ERROR, "%s: got %s from dlmlock($RECOVERY), " 2582 "lksb.status=%s\n", dlm->name, dlm_errname(ret), 2583 dlm_errname(lksb.status)); 2584 res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME, 2585 DLM_RECOVERY_LOCK_NAME_LEN); 2586 if (res) { 2587 dlm_print_one_lock_resource(res); 2588 dlm_lockres_put(res); 2589 } else { 2590 mlog(ML_ERROR, "recovery lock not found\n"); 2591 } 2592 BUG(); 2593 } 2594 2595 return status; 2596 } 2597 2598 static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node) 2599 { 2600 struct dlm_begin_reco br; 2601 int ret = 0; 2602 struct dlm_node_iter iter; 2603 int nodenum; 2604 int status; 2605 2606 mlog(0, "%s: dead node is %u\n", dlm->name, dead_node); 2607 2608 spin_lock(&dlm->spinlock); 2609 dlm_node_iter_init(dlm->domain_map, &iter); 2610 spin_unlock(&dlm->spinlock); 2611 2612 clear_bit(dead_node, iter.node_map); 2613 2614 memset(&br, 0, sizeof(br)); 2615 br.node_idx = dlm->node_num; 2616 br.dead_node = dead_node; 2617 2618 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) { 2619 ret = 0; 2620 if (nodenum == dead_node) { 2621 mlog(0, "not sending begin reco to dead node " 2622 "%u\n", dead_node); 2623 continue; 2624 } 2625 if (nodenum == dlm->node_num) { 2626 mlog(0, "not sending begin reco to self\n"); 2627 continue; 2628 } 2629 retry: 2630 ret = -EINVAL; 2631 mlog(0, "attempting to send begin reco msg to %d\n", 2632 nodenum); 2633 ret = o2net_send_message(DLM_BEGIN_RECO_MSG, dlm->key, 2634 &br, sizeof(br), nodenum, &status); 2635 /* negative status is handled ok by caller here */ 2636 if (ret >= 0) 2637 ret = status; 2638 if (dlm_is_host_down(ret)) { 2639 /* node is down. not involved in recovery 2640 * so just keep going */ 2641 mlog(ML_NOTICE, "%s: node %u was down when sending " 2642 "begin reco msg (%d)\n", dlm->name, nodenum, ret); 2643 ret = 0; 2644 } 2645 2646 /* 2647 * Prior to commit aad1b15310b9bcd59fa81ab8f2b1513b59553ea8, 2648 * dlm_begin_reco_handler() returned EAGAIN and not -EAGAIN. 2649 * We are handling both for compatibility reasons. 2650 */ 2651 if (ret == -EAGAIN || ret == EAGAIN) { 2652 mlog(0, "%s: trying to start recovery of node " 2653 "%u, but node %u is waiting for last recovery " 2654 "to complete, backoff for a bit\n", dlm->name, 2655 dead_node, nodenum); 2656 msleep(100); 2657 goto retry; 2658 } 2659 if (ret < 0) { 2660 struct dlm_lock_resource *res; 2661 2662 /* this is now a serious problem, possibly ENOMEM 2663 * in the network stack. must retry */ 2664 mlog_errno(ret); 2665 mlog(ML_ERROR, "begin reco of dlm %s to node %u " 2666 "returned %d\n", dlm->name, nodenum, ret); 2667 res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME, 2668 DLM_RECOVERY_LOCK_NAME_LEN); 2669 if (res) { 2670 dlm_print_one_lock_resource(res); 2671 dlm_lockres_put(res); 2672 } else { 2673 mlog(ML_ERROR, "recovery lock not found\n"); 2674 } 2675 /* sleep for a bit in hopes that we can avoid 2676 * another ENOMEM */ 2677 msleep(100); 2678 goto retry; 2679 } 2680 } 2681 2682 return ret; 2683 } 2684 2685 int dlm_begin_reco_handler(struct o2net_msg *msg, u32 len, void *data, 2686 void **ret_data) 2687 { 2688 struct dlm_ctxt *dlm = data; 2689 struct dlm_begin_reco *br = (struct dlm_begin_reco *)msg->buf; 2690 2691 /* ok to return 0, domain has gone away */ 2692 if (!dlm_grab(dlm)) 2693 return 0; 2694 2695 spin_lock(&dlm->spinlock); 2696 if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) { 2697 mlog(0, "%s: node %u wants to recover node %u (%u:%u) " 2698 "but this node is in finalize state, waiting on finalize2\n", 2699 dlm->name, br->node_idx, br->dead_node, 2700 dlm->reco.dead_node, dlm->reco.new_master); 2701 spin_unlock(&dlm->spinlock); 2702 return -EAGAIN; 2703 } 2704 spin_unlock(&dlm->spinlock); 2705 2706 mlog(0, "%s: node %u wants to recover node %u (%u:%u)\n", 2707 dlm->name, br->node_idx, br->dead_node, 2708 dlm->reco.dead_node, dlm->reco.new_master); 2709 2710 dlm_fire_domain_eviction_callbacks(dlm, br->dead_node); 2711 2712 spin_lock(&dlm->spinlock); 2713 if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) { 2714 if (test_bit(dlm->reco.new_master, dlm->recovery_map)) { 2715 mlog(0, "%s: new_master %u died, changing " 2716 "to %u\n", dlm->name, dlm->reco.new_master, 2717 br->node_idx); 2718 } else { 2719 mlog(0, "%s: new_master %u NOT DEAD, changing " 2720 "to %u\n", dlm->name, dlm->reco.new_master, 2721 br->node_idx); 2722 /* may not have seen the new master as dead yet */ 2723 } 2724 } 2725 if (dlm->reco.dead_node != O2NM_INVALID_NODE_NUM) { 2726 mlog(ML_NOTICE, "%s: dead_node previously set to %u, " 2727 "node %u changing it to %u\n", dlm->name, 2728 dlm->reco.dead_node, br->node_idx, br->dead_node); 2729 } 2730 dlm_set_reco_master(dlm, br->node_idx); 2731 dlm_set_reco_dead_node(dlm, br->dead_node); 2732 if (!test_bit(br->dead_node, dlm->recovery_map)) { 2733 mlog(0, "recovery master %u sees %u as dead, but this " 2734 "node has not yet. marking %u as dead\n", 2735 br->node_idx, br->dead_node, br->dead_node); 2736 if (!test_bit(br->dead_node, dlm->domain_map) || 2737 !test_bit(br->dead_node, dlm->live_nodes_map)) 2738 mlog(0, "%u not in domain/live_nodes map " 2739 "so setting it in reco map manually\n", 2740 br->dead_node); 2741 /* force the recovery cleanup in __dlm_hb_node_down 2742 * both of these will be cleared in a moment */ 2743 set_bit(br->dead_node, dlm->domain_map); 2744 set_bit(br->dead_node, dlm->live_nodes_map); 2745 __dlm_hb_node_down(dlm, br->dead_node); 2746 } 2747 spin_unlock(&dlm->spinlock); 2748 2749 dlm_kick_recovery_thread(dlm); 2750 2751 mlog(0, "%s: recovery started by node %u, for %u (%u:%u)\n", 2752 dlm->name, br->node_idx, br->dead_node, 2753 dlm->reco.dead_node, dlm->reco.new_master); 2754 2755 dlm_put(dlm); 2756 return 0; 2757 } 2758 2759 #define DLM_FINALIZE_STAGE2 0x01 2760 static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm) 2761 { 2762 int ret = 0; 2763 struct dlm_finalize_reco fr; 2764 struct dlm_node_iter iter; 2765 int nodenum; 2766 int status; 2767 int stage = 1; 2768 2769 mlog(0, "finishing recovery for node %s:%u, " 2770 "stage %d\n", dlm->name, dlm->reco.dead_node, stage); 2771 2772 spin_lock(&dlm->spinlock); 2773 dlm_node_iter_init(dlm->domain_map, &iter); 2774 spin_unlock(&dlm->spinlock); 2775 2776 stage2: 2777 memset(&fr, 0, sizeof(fr)); 2778 fr.node_idx = dlm->node_num; 2779 fr.dead_node = dlm->reco.dead_node; 2780 if (stage == 2) 2781 fr.flags |= DLM_FINALIZE_STAGE2; 2782 2783 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) { 2784 if (nodenum == dlm->node_num) 2785 continue; 2786 ret = o2net_send_message(DLM_FINALIZE_RECO_MSG, dlm->key, 2787 &fr, sizeof(fr), nodenum, &status); 2788 if (ret >= 0) 2789 ret = status; 2790 if (ret < 0) { 2791 mlog(ML_ERROR, "Error %d when sending message %u (key " 2792 "0x%x) to node %u\n", ret, DLM_FINALIZE_RECO_MSG, 2793 dlm->key, nodenum); 2794 if (dlm_is_host_down(ret)) { 2795 /* this has no effect on this recovery 2796 * session, so set the status to zero to 2797 * finish out the last recovery */ 2798 mlog(ML_ERROR, "node %u went down after this " 2799 "node finished recovery.\n", nodenum); 2800 ret = 0; 2801 continue; 2802 } 2803 break; 2804 } 2805 } 2806 if (stage == 1) { 2807 /* reset the node_iter back to the top and send finalize2 */ 2808 iter.curnode = -1; 2809 stage = 2; 2810 goto stage2; 2811 } 2812 2813 return ret; 2814 } 2815 2816 int dlm_finalize_reco_handler(struct o2net_msg *msg, u32 len, void *data, 2817 void **ret_data) 2818 { 2819 struct dlm_ctxt *dlm = data; 2820 struct dlm_finalize_reco *fr = (struct dlm_finalize_reco *)msg->buf; 2821 int stage = 1; 2822 2823 /* ok to return 0, domain has gone away */ 2824 if (!dlm_grab(dlm)) 2825 return 0; 2826 2827 if (fr->flags & DLM_FINALIZE_STAGE2) 2828 stage = 2; 2829 2830 mlog(0, "%s: node %u finalizing recovery stage%d of " 2831 "node %u (%u:%u)\n", dlm->name, fr->node_idx, stage, 2832 fr->dead_node, dlm->reco.dead_node, dlm->reco.new_master); 2833 2834 spin_lock(&dlm->spinlock); 2835 2836 if (dlm->reco.new_master != fr->node_idx) { 2837 mlog(ML_ERROR, "node %u sent recovery finalize msg, but node " 2838 "%u is supposed to be the new master, dead=%u\n", 2839 fr->node_idx, dlm->reco.new_master, fr->dead_node); 2840 BUG(); 2841 } 2842 if (dlm->reco.dead_node != fr->dead_node) { 2843 mlog(ML_ERROR, "node %u sent recovery finalize msg for dead " 2844 "node %u, but node %u is supposed to be dead\n", 2845 fr->node_idx, fr->dead_node, dlm->reco.dead_node); 2846 BUG(); 2847 } 2848 2849 switch (stage) { 2850 case 1: 2851 dlm_finish_local_lockres_recovery(dlm, fr->dead_node, fr->node_idx); 2852 if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) { 2853 mlog(ML_ERROR, "%s: received finalize1 from " 2854 "new master %u for dead node %u, but " 2855 "this node has already received it!\n", 2856 dlm->name, fr->node_idx, fr->dead_node); 2857 dlm_print_reco_node_status(dlm); 2858 BUG(); 2859 } 2860 dlm->reco.state |= DLM_RECO_STATE_FINALIZE; 2861 spin_unlock(&dlm->spinlock); 2862 break; 2863 case 2: 2864 if (!(dlm->reco.state & DLM_RECO_STATE_FINALIZE)) { 2865 mlog(ML_ERROR, "%s: received finalize2 from " 2866 "new master %u for dead node %u, but " 2867 "this node did not have finalize1!\n", 2868 dlm->name, fr->node_idx, fr->dead_node); 2869 dlm_print_reco_node_status(dlm); 2870 BUG(); 2871 } 2872 dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE; 2873 spin_unlock(&dlm->spinlock); 2874 dlm_reset_recovery(dlm); 2875 dlm_kick_recovery_thread(dlm); 2876 break; 2877 default: 2878 BUG(); 2879 } 2880 2881 mlog(0, "%s: recovery done, reco master was %u, dead now %u, master now %u\n", 2882 dlm->name, fr->node_idx, dlm->reco.dead_node, dlm->reco.new_master); 2883 2884 dlm_put(dlm); 2885 return 0; 2886 } 2887