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 mlog_exit(status); 731 return status; 732 } 733 734 static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node) 735 { 736 int num=0; 737 struct dlm_reco_node_data *ndata; 738 739 spin_lock(&dlm->spinlock); 740 memcpy(dlm->reco.node_map, dlm->domain_map, sizeof(dlm->domain_map)); 741 /* nodes can only be removed (by dying) after dropping 742 * this lock, and death will be trapped later, so this should do */ 743 spin_unlock(&dlm->spinlock); 744 745 while (1) { 746 num = find_next_bit (dlm->reco.node_map, O2NM_MAX_NODES, num); 747 if (num >= O2NM_MAX_NODES) { 748 break; 749 } 750 BUG_ON(num == dead_node); 751 752 ndata = kzalloc(sizeof(*ndata), GFP_NOFS); 753 if (!ndata) { 754 dlm_destroy_recovery_area(dlm, dead_node); 755 return -ENOMEM; 756 } 757 ndata->node_num = num; 758 ndata->state = DLM_RECO_NODE_DATA_INIT; 759 spin_lock(&dlm_reco_state_lock); 760 list_add_tail(&ndata->list, &dlm->reco.node_data); 761 spin_unlock(&dlm_reco_state_lock); 762 num++; 763 } 764 765 return 0; 766 } 767 768 static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm, u8 dead_node) 769 { 770 struct dlm_reco_node_data *ndata, *next; 771 LIST_HEAD(tmplist); 772 773 spin_lock(&dlm_reco_state_lock); 774 list_splice_init(&dlm->reco.node_data, &tmplist); 775 spin_unlock(&dlm_reco_state_lock); 776 777 list_for_each_entry_safe(ndata, next, &tmplist, list) { 778 list_del_init(&ndata->list); 779 kfree(ndata); 780 } 781 } 782 783 static int dlm_request_all_locks(struct dlm_ctxt *dlm, u8 request_from, 784 u8 dead_node) 785 { 786 struct dlm_lock_request lr; 787 enum dlm_status ret; 788 789 mlog(0, "\n"); 790 791 792 mlog(0, "dlm_request_all_locks: dead node is %u, sending request " 793 "to %u\n", dead_node, request_from); 794 795 memset(&lr, 0, sizeof(lr)); 796 lr.node_idx = dlm->node_num; 797 lr.dead_node = dead_node; 798 799 // send message 800 ret = DLM_NOLOCKMGR; 801 ret = o2net_send_message(DLM_LOCK_REQUEST_MSG, dlm->key, 802 &lr, sizeof(lr), request_from, NULL); 803 804 /* negative status is handled by caller */ 805 if (ret < 0) 806 mlog(ML_ERROR, "Error %d when sending message %u (key " 807 "0x%x) to node %u\n", ret, DLM_LOCK_REQUEST_MSG, 808 dlm->key, request_from); 809 810 // return from here, then 811 // sleep until all received or error 812 return ret; 813 814 } 815 816 int dlm_request_all_locks_handler(struct o2net_msg *msg, u32 len, void *data, 817 void **ret_data) 818 { 819 struct dlm_ctxt *dlm = data; 820 struct dlm_lock_request *lr = (struct dlm_lock_request *)msg->buf; 821 char *buf = NULL; 822 struct dlm_work_item *item = NULL; 823 824 if (!dlm_grab(dlm)) 825 return -EINVAL; 826 827 if (lr->dead_node != dlm->reco.dead_node) { 828 mlog(ML_ERROR, "%s: node %u sent dead_node=%u, but local " 829 "dead_node is %u\n", dlm->name, lr->node_idx, 830 lr->dead_node, dlm->reco.dead_node); 831 dlm_print_reco_node_status(dlm); 832 /* this is a hack */ 833 dlm_put(dlm); 834 return -ENOMEM; 835 } 836 BUG_ON(lr->dead_node != dlm->reco.dead_node); 837 838 item = kzalloc(sizeof(*item), GFP_NOFS); 839 if (!item) { 840 dlm_put(dlm); 841 return -ENOMEM; 842 } 843 844 /* this will get freed by dlm_request_all_locks_worker */ 845 buf = (char *) __get_free_page(GFP_NOFS); 846 if (!buf) { 847 kfree(item); 848 dlm_put(dlm); 849 return -ENOMEM; 850 } 851 852 /* queue up work for dlm_request_all_locks_worker */ 853 dlm_grab(dlm); /* get an extra ref for the work item */ 854 dlm_init_work_item(dlm, item, dlm_request_all_locks_worker, buf); 855 item->u.ral.reco_master = lr->node_idx; 856 item->u.ral.dead_node = lr->dead_node; 857 spin_lock(&dlm->work_lock); 858 list_add_tail(&item->list, &dlm->work_list); 859 spin_unlock(&dlm->work_lock); 860 queue_work(dlm->dlm_worker, &dlm->dispatched_work); 861 862 dlm_put(dlm); 863 return 0; 864 } 865 866 static void dlm_request_all_locks_worker(struct dlm_work_item *item, void *data) 867 { 868 struct dlm_migratable_lockres *mres; 869 struct dlm_lock_resource *res; 870 struct dlm_ctxt *dlm; 871 LIST_HEAD(resources); 872 int ret; 873 u8 dead_node, reco_master; 874 int skip_all_done = 0; 875 876 dlm = item->dlm; 877 dead_node = item->u.ral.dead_node; 878 reco_master = item->u.ral.reco_master; 879 mres = (struct dlm_migratable_lockres *)data; 880 881 mlog(0, "%s: recovery worker started, dead=%u, master=%u\n", 882 dlm->name, dead_node, reco_master); 883 884 if (dead_node != dlm->reco.dead_node || 885 reco_master != dlm->reco.new_master) { 886 /* worker could have been created before the recovery master 887 * died. if so, do not continue, but do not error. */ 888 if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) { 889 mlog(ML_NOTICE, "%s: will not send recovery state, " 890 "recovery master %u died, thread=(dead=%u,mas=%u)" 891 " current=(dead=%u,mas=%u)\n", dlm->name, 892 reco_master, dead_node, reco_master, 893 dlm->reco.dead_node, dlm->reco.new_master); 894 } else { 895 mlog(ML_NOTICE, "%s: reco state invalid: reco(dead=%u, " 896 "master=%u), request(dead=%u, master=%u)\n", 897 dlm->name, dlm->reco.dead_node, 898 dlm->reco.new_master, dead_node, reco_master); 899 } 900 goto leave; 901 } 902 903 /* lock resources should have already been moved to the 904 * dlm->reco.resources list. now move items from that list 905 * to a temp list if the dead owner matches. note that the 906 * whole cluster recovers only one node at a time, so we 907 * can safely move UNKNOWN lock resources for each recovery 908 * session. */ 909 dlm_move_reco_locks_to_list(dlm, &resources, dead_node); 910 911 /* now we can begin blasting lockreses without the dlm lock */ 912 913 /* any errors returned will be due to the new_master dying, 914 * the dlm_reco_thread should detect this */ 915 list_for_each_entry(res, &resources, recovering) { 916 ret = dlm_send_one_lockres(dlm, res, mres, reco_master, 917 DLM_MRES_RECOVERY); 918 if (ret < 0) { 919 mlog(ML_ERROR, "%s: node %u went down while sending " 920 "recovery state for dead node %u, ret=%d\n", dlm->name, 921 reco_master, dead_node, ret); 922 skip_all_done = 1; 923 break; 924 } 925 } 926 927 /* move the resources back to the list */ 928 spin_lock(&dlm->spinlock); 929 list_splice_init(&resources, &dlm->reco.resources); 930 spin_unlock(&dlm->spinlock); 931 932 if (!skip_all_done) { 933 ret = dlm_send_all_done_msg(dlm, dead_node, reco_master); 934 if (ret < 0) { 935 mlog(ML_ERROR, "%s: node %u went down while sending " 936 "recovery all-done for dead node %u, ret=%d\n", 937 dlm->name, reco_master, dead_node, ret); 938 } 939 } 940 leave: 941 free_page((unsigned long)data); 942 } 943 944 945 static int dlm_send_all_done_msg(struct dlm_ctxt *dlm, u8 dead_node, u8 send_to) 946 { 947 int ret, tmpret; 948 struct dlm_reco_data_done done_msg; 949 950 memset(&done_msg, 0, sizeof(done_msg)); 951 done_msg.node_idx = dlm->node_num; 952 done_msg.dead_node = dead_node; 953 mlog(0, "sending DATA DONE message to %u, " 954 "my node=%u, dead node=%u\n", send_to, done_msg.node_idx, 955 done_msg.dead_node); 956 957 ret = o2net_send_message(DLM_RECO_DATA_DONE_MSG, dlm->key, &done_msg, 958 sizeof(done_msg), send_to, &tmpret); 959 if (ret < 0) { 960 mlog(ML_ERROR, "Error %d when sending message %u (key " 961 "0x%x) to node %u\n", ret, DLM_RECO_DATA_DONE_MSG, 962 dlm->key, send_to); 963 if (!dlm_is_host_down(ret)) { 964 BUG(); 965 } 966 } else 967 ret = tmpret; 968 return ret; 969 } 970 971 972 int dlm_reco_data_done_handler(struct o2net_msg *msg, u32 len, void *data, 973 void **ret_data) 974 { 975 struct dlm_ctxt *dlm = data; 976 struct dlm_reco_data_done *done = (struct dlm_reco_data_done *)msg->buf; 977 struct dlm_reco_node_data *ndata = NULL; 978 int ret = -EINVAL; 979 980 if (!dlm_grab(dlm)) 981 return -EINVAL; 982 983 mlog(0, "got DATA DONE: dead_node=%u, reco.dead_node=%u, " 984 "node_idx=%u, this node=%u\n", done->dead_node, 985 dlm->reco.dead_node, done->node_idx, dlm->node_num); 986 987 mlog_bug_on_msg((done->dead_node != dlm->reco.dead_node), 988 "Got DATA DONE: dead_node=%u, reco.dead_node=%u, " 989 "node_idx=%u, this node=%u\n", done->dead_node, 990 dlm->reco.dead_node, done->node_idx, dlm->node_num); 991 992 spin_lock(&dlm_reco_state_lock); 993 list_for_each_entry(ndata, &dlm->reco.node_data, list) { 994 if (ndata->node_num != done->node_idx) 995 continue; 996 997 switch (ndata->state) { 998 /* should have moved beyond INIT but not to FINALIZE yet */ 999 case DLM_RECO_NODE_DATA_INIT: 1000 case DLM_RECO_NODE_DATA_DEAD: 1001 case DLM_RECO_NODE_DATA_FINALIZE_SENT: 1002 mlog(ML_ERROR, "bad ndata state for node %u:" 1003 " state=%d\n", ndata->node_num, 1004 ndata->state); 1005 BUG(); 1006 break; 1007 /* these states are possible at this point, anywhere along 1008 * the line of recovery */ 1009 case DLM_RECO_NODE_DATA_DONE: 1010 case DLM_RECO_NODE_DATA_RECEIVING: 1011 case DLM_RECO_NODE_DATA_REQUESTED: 1012 case DLM_RECO_NODE_DATA_REQUESTING: 1013 mlog(0, "node %u is DONE sending " 1014 "recovery data!\n", 1015 ndata->node_num); 1016 1017 ndata->state = DLM_RECO_NODE_DATA_DONE; 1018 ret = 0; 1019 break; 1020 } 1021 } 1022 spin_unlock(&dlm_reco_state_lock); 1023 1024 /* wake the recovery thread, some node is done */ 1025 if (!ret) 1026 dlm_kick_recovery_thread(dlm); 1027 1028 if (ret < 0) 1029 mlog(ML_ERROR, "failed to find recovery node data for node " 1030 "%u\n", done->node_idx); 1031 dlm_put(dlm); 1032 1033 mlog(0, "leaving reco data done handler, ret=%d\n", ret); 1034 return ret; 1035 } 1036 1037 static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm, 1038 struct list_head *list, 1039 u8 dead_node) 1040 { 1041 struct dlm_lock_resource *res, *next; 1042 struct dlm_lock *lock; 1043 1044 spin_lock(&dlm->spinlock); 1045 list_for_each_entry_safe(res, next, &dlm->reco.resources, recovering) { 1046 /* always prune any $RECOVERY entries for dead nodes, 1047 * otherwise hangs can occur during later recovery */ 1048 if (dlm_is_recovery_lock(res->lockname.name, 1049 res->lockname.len)) { 1050 spin_lock(&res->spinlock); 1051 list_for_each_entry(lock, &res->granted, list) { 1052 if (lock->ml.node == dead_node) { 1053 mlog(0, "AHA! there was " 1054 "a $RECOVERY lock for dead " 1055 "node %u (%s)!\n", 1056 dead_node, dlm->name); 1057 list_del_init(&lock->list); 1058 dlm_lock_put(lock); 1059 break; 1060 } 1061 } 1062 spin_unlock(&res->spinlock); 1063 continue; 1064 } 1065 1066 if (res->owner == dead_node) { 1067 mlog(0, "found lockres owned by dead node while " 1068 "doing recovery for node %u. sending it.\n", 1069 dead_node); 1070 list_move_tail(&res->recovering, list); 1071 } else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) { 1072 mlog(0, "found UNKNOWN owner while doing recovery " 1073 "for node %u. sending it.\n", dead_node); 1074 list_move_tail(&res->recovering, list); 1075 } 1076 } 1077 spin_unlock(&dlm->spinlock); 1078 } 1079 1080 static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res) 1081 { 1082 int total_locks = 0; 1083 struct list_head *iter, *queue = &res->granted; 1084 int i; 1085 1086 for (i=0; i<3; i++) { 1087 list_for_each(iter, queue) 1088 total_locks++; 1089 queue++; 1090 } 1091 return total_locks; 1092 } 1093 1094 1095 static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm, 1096 struct dlm_migratable_lockres *mres, 1097 u8 send_to, 1098 struct dlm_lock_resource *res, 1099 int total_locks) 1100 { 1101 u64 mig_cookie = be64_to_cpu(mres->mig_cookie); 1102 int mres_total_locks = be32_to_cpu(mres->total_locks); 1103 int sz, ret = 0, status = 0; 1104 u8 orig_flags = mres->flags, 1105 orig_master = mres->master; 1106 1107 BUG_ON(mres->num_locks > DLM_MAX_MIGRATABLE_LOCKS); 1108 if (!mres->num_locks) 1109 return 0; 1110 1111 sz = sizeof(struct dlm_migratable_lockres) + 1112 (mres->num_locks * sizeof(struct dlm_migratable_lock)); 1113 1114 /* add an all-done flag if we reached the last lock */ 1115 orig_flags = mres->flags; 1116 BUG_ON(total_locks > mres_total_locks); 1117 if (total_locks == mres_total_locks) 1118 mres->flags |= DLM_MRES_ALL_DONE; 1119 1120 mlog(0, "%s:%.*s: sending mig lockres (%s) to %u\n", 1121 dlm->name, res->lockname.len, res->lockname.name, 1122 orig_flags & DLM_MRES_MIGRATION ? "migration" : "recovery", 1123 send_to); 1124 1125 /* send it */ 1126 ret = o2net_send_message(DLM_MIG_LOCKRES_MSG, dlm->key, mres, 1127 sz, send_to, &status); 1128 if (ret < 0) { 1129 /* XXX: negative status is not handled. 1130 * this will end up killing this node. */ 1131 mlog(ML_ERROR, "Error %d when sending message %u (key " 1132 "0x%x) to node %u\n", ret, DLM_MIG_LOCKRES_MSG, 1133 dlm->key, send_to); 1134 } else { 1135 /* might get an -ENOMEM back here */ 1136 ret = status; 1137 if (ret < 0) { 1138 mlog_errno(ret); 1139 1140 if (ret == -EFAULT) { 1141 mlog(ML_ERROR, "node %u told me to kill " 1142 "myself!\n", send_to); 1143 BUG(); 1144 } 1145 } 1146 } 1147 1148 /* zero and reinit the message buffer */ 1149 dlm_init_migratable_lockres(mres, res->lockname.name, 1150 res->lockname.len, mres_total_locks, 1151 mig_cookie, orig_flags, orig_master); 1152 return ret; 1153 } 1154 1155 static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres, 1156 const char *lockname, int namelen, 1157 int total_locks, u64 cookie, 1158 u8 flags, u8 master) 1159 { 1160 /* mres here is one full page */ 1161 clear_page(mres); 1162 mres->lockname_len = namelen; 1163 memcpy(mres->lockname, lockname, namelen); 1164 mres->num_locks = 0; 1165 mres->total_locks = cpu_to_be32(total_locks); 1166 mres->mig_cookie = cpu_to_be64(cookie); 1167 mres->flags = flags; 1168 mres->master = master; 1169 } 1170 1171 static void dlm_prepare_lvb_for_migration(struct dlm_lock *lock, 1172 struct dlm_migratable_lockres *mres, 1173 int queue) 1174 { 1175 if (!lock->lksb) 1176 return; 1177 1178 /* Ignore lvb in all locks in the blocked list */ 1179 if (queue == DLM_BLOCKED_LIST) 1180 return; 1181 1182 /* Only consider lvbs in locks with granted EX or PR lock levels */ 1183 if (lock->ml.type != LKM_EXMODE && lock->ml.type != LKM_PRMODE) 1184 return; 1185 1186 if (dlm_lvb_is_empty(mres->lvb)) { 1187 memcpy(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN); 1188 return; 1189 } 1190 1191 /* Ensure the lvb copied for migration matches in other valid locks */ 1192 if (!memcmp(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN)) 1193 return; 1194 1195 mlog(ML_ERROR, "Mismatched lvb in lock cookie=%u:%llu, name=%.*s, " 1196 "node=%u\n", 1197 dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)), 1198 dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)), 1199 lock->lockres->lockname.len, lock->lockres->lockname.name, 1200 lock->ml.node); 1201 dlm_print_one_lock_resource(lock->lockres); 1202 BUG(); 1203 } 1204 1205 /* returns 1 if this lock fills the network structure, 1206 * 0 otherwise */ 1207 static int dlm_add_lock_to_array(struct dlm_lock *lock, 1208 struct dlm_migratable_lockres *mres, int queue) 1209 { 1210 struct dlm_migratable_lock *ml; 1211 int lock_num = mres->num_locks; 1212 1213 ml = &(mres->ml[lock_num]); 1214 ml->cookie = lock->ml.cookie; 1215 ml->type = lock->ml.type; 1216 ml->convert_type = lock->ml.convert_type; 1217 ml->highest_blocked = lock->ml.highest_blocked; 1218 ml->list = queue; 1219 if (lock->lksb) { 1220 ml->flags = lock->lksb->flags; 1221 dlm_prepare_lvb_for_migration(lock, mres, queue); 1222 } 1223 ml->node = lock->ml.node; 1224 mres->num_locks++; 1225 /* we reached the max, send this network message */ 1226 if (mres->num_locks == DLM_MAX_MIGRATABLE_LOCKS) 1227 return 1; 1228 return 0; 1229 } 1230 1231 static void dlm_add_dummy_lock(struct dlm_ctxt *dlm, 1232 struct dlm_migratable_lockres *mres) 1233 { 1234 struct dlm_lock dummy; 1235 memset(&dummy, 0, sizeof(dummy)); 1236 dummy.ml.cookie = 0; 1237 dummy.ml.type = LKM_IVMODE; 1238 dummy.ml.convert_type = LKM_IVMODE; 1239 dummy.ml.highest_blocked = LKM_IVMODE; 1240 dummy.lksb = NULL; 1241 dummy.ml.node = dlm->node_num; 1242 dlm_add_lock_to_array(&dummy, mres, DLM_BLOCKED_LIST); 1243 } 1244 1245 static inline int dlm_is_dummy_lock(struct dlm_ctxt *dlm, 1246 struct dlm_migratable_lock *ml, 1247 u8 *nodenum) 1248 { 1249 if (unlikely(ml->cookie == 0 && 1250 ml->type == LKM_IVMODE && 1251 ml->convert_type == LKM_IVMODE && 1252 ml->highest_blocked == LKM_IVMODE && 1253 ml->list == DLM_BLOCKED_LIST)) { 1254 *nodenum = ml->node; 1255 return 1; 1256 } 1257 return 0; 1258 } 1259 1260 int dlm_send_one_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res, 1261 struct dlm_migratable_lockres *mres, 1262 u8 send_to, u8 flags) 1263 { 1264 struct list_head *queue; 1265 int total_locks, i; 1266 u64 mig_cookie = 0; 1267 struct dlm_lock *lock; 1268 int ret = 0; 1269 1270 BUG_ON(!(flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION))); 1271 1272 mlog(0, "sending to %u\n", send_to); 1273 1274 total_locks = dlm_num_locks_in_lockres(res); 1275 if (total_locks > DLM_MAX_MIGRATABLE_LOCKS) { 1276 /* rare, but possible */ 1277 mlog(0, "argh. lockres has %d locks. this will " 1278 "require more than one network packet to " 1279 "migrate\n", total_locks); 1280 mig_cookie = dlm_get_next_mig_cookie(); 1281 } 1282 1283 dlm_init_migratable_lockres(mres, res->lockname.name, 1284 res->lockname.len, total_locks, 1285 mig_cookie, flags, res->owner); 1286 1287 total_locks = 0; 1288 for (i=DLM_GRANTED_LIST; i<=DLM_BLOCKED_LIST; i++) { 1289 queue = dlm_list_idx_to_ptr(res, i); 1290 list_for_each_entry(lock, queue, list) { 1291 /* add another lock. */ 1292 total_locks++; 1293 if (!dlm_add_lock_to_array(lock, mres, i)) 1294 continue; 1295 1296 /* this filled the lock message, 1297 * we must send it immediately. */ 1298 ret = dlm_send_mig_lockres_msg(dlm, mres, send_to, 1299 res, total_locks); 1300 if (ret < 0) 1301 goto error; 1302 } 1303 } 1304 if (total_locks == 0) { 1305 /* send a dummy lock to indicate a mastery reference only */ 1306 mlog(0, "%s:%.*s: sending dummy lock to %u, %s\n", 1307 dlm->name, res->lockname.len, res->lockname.name, 1308 send_to, flags & DLM_MRES_RECOVERY ? "recovery" : 1309 "migration"); 1310 dlm_add_dummy_lock(dlm, mres); 1311 } 1312 /* flush any remaining locks */ 1313 ret = dlm_send_mig_lockres_msg(dlm, mres, send_to, res, total_locks); 1314 if (ret < 0) 1315 goto error; 1316 return ret; 1317 1318 error: 1319 mlog(ML_ERROR, "%s: dlm_send_mig_lockres_msg returned %d\n", 1320 dlm->name, ret); 1321 if (!dlm_is_host_down(ret)) 1322 BUG(); 1323 mlog(0, "%s: node %u went down while sending %s " 1324 "lockres %.*s\n", dlm->name, send_to, 1325 flags & DLM_MRES_RECOVERY ? "recovery" : "migration", 1326 res->lockname.len, res->lockname.name); 1327 return ret; 1328 } 1329 1330 1331 1332 /* 1333 * this message will contain no more than one page worth of 1334 * recovery data, and it will work on only one lockres. 1335 * there may be many locks in this page, and we may need to wait 1336 * for additional packets to complete all the locks (rare, but 1337 * possible). 1338 */ 1339 /* 1340 * NOTE: the allocation error cases here are scary 1341 * we really cannot afford to fail an alloc in recovery 1342 * do we spin? returning an error only delays the problem really 1343 */ 1344 1345 int dlm_mig_lockres_handler(struct o2net_msg *msg, u32 len, void *data, 1346 void **ret_data) 1347 { 1348 struct dlm_ctxt *dlm = data; 1349 struct dlm_migratable_lockres *mres = 1350 (struct dlm_migratable_lockres *)msg->buf; 1351 int ret = 0; 1352 u8 real_master; 1353 u8 extra_refs = 0; 1354 char *buf = NULL; 1355 struct dlm_work_item *item = NULL; 1356 struct dlm_lock_resource *res = NULL; 1357 1358 if (!dlm_grab(dlm)) 1359 return -EINVAL; 1360 1361 BUG_ON(!(mres->flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION))); 1362 1363 real_master = mres->master; 1364 if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) { 1365 /* cannot migrate a lockres with no master */ 1366 BUG_ON(!(mres->flags & DLM_MRES_RECOVERY)); 1367 } 1368 1369 mlog(0, "%s message received from node %u\n", 1370 (mres->flags & DLM_MRES_RECOVERY) ? 1371 "recovery" : "migration", mres->master); 1372 if (mres->flags & DLM_MRES_ALL_DONE) 1373 mlog(0, "all done flag. all lockres data received!\n"); 1374 1375 ret = -ENOMEM; 1376 buf = kmalloc(be16_to_cpu(msg->data_len), GFP_NOFS); 1377 item = kzalloc(sizeof(*item), GFP_NOFS); 1378 if (!buf || !item) 1379 goto leave; 1380 1381 /* lookup the lock to see if we have a secondary queue for this 1382 * already... just add the locks in and this will have its owner 1383 * and RECOVERY flag changed when it completes. */ 1384 res = dlm_lookup_lockres(dlm, mres->lockname, mres->lockname_len); 1385 if (res) { 1386 /* this will get a ref on res */ 1387 /* mark it as recovering/migrating and hash it */ 1388 spin_lock(&res->spinlock); 1389 if (mres->flags & DLM_MRES_RECOVERY) { 1390 res->state |= DLM_LOCK_RES_RECOVERING; 1391 } else { 1392 if (res->state & DLM_LOCK_RES_MIGRATING) { 1393 /* this is at least the second 1394 * lockres message */ 1395 mlog(0, "lock %.*s is already migrating\n", 1396 mres->lockname_len, 1397 mres->lockname); 1398 } else if (res->state & DLM_LOCK_RES_RECOVERING) { 1399 /* caller should BUG */ 1400 mlog(ML_ERROR, "node is attempting to migrate " 1401 "lock %.*s, but marked as recovering!\n", 1402 mres->lockname_len, mres->lockname); 1403 ret = -EFAULT; 1404 spin_unlock(&res->spinlock); 1405 goto leave; 1406 } 1407 res->state |= DLM_LOCK_RES_MIGRATING; 1408 } 1409 spin_unlock(&res->spinlock); 1410 } else { 1411 /* need to allocate, just like if it was 1412 * mastered here normally */ 1413 res = dlm_new_lockres(dlm, mres->lockname, mres->lockname_len); 1414 if (!res) 1415 goto leave; 1416 1417 /* to match the ref that we would have gotten if 1418 * dlm_lookup_lockres had succeeded */ 1419 dlm_lockres_get(res); 1420 1421 /* mark it as recovering/migrating and hash it */ 1422 if (mres->flags & DLM_MRES_RECOVERY) 1423 res->state |= DLM_LOCK_RES_RECOVERING; 1424 else 1425 res->state |= DLM_LOCK_RES_MIGRATING; 1426 1427 spin_lock(&dlm->spinlock); 1428 __dlm_insert_lockres(dlm, res); 1429 spin_unlock(&dlm->spinlock); 1430 1431 /* Add an extra ref for this lock-less lockres lest the 1432 * dlm_thread purges it before we get the chance to add 1433 * locks to it */ 1434 dlm_lockres_get(res); 1435 1436 /* There are three refs that need to be put. 1437 * 1. Taken above. 1438 * 2. kref_init in dlm_new_lockres()->dlm_init_lockres(). 1439 * 3. dlm_lookup_lockres() 1440 * The first one is handled at the end of this function. The 1441 * other two are handled in the worker thread after locks have 1442 * been attached. Yes, we don't wait for purge time to match 1443 * kref_init. The lockres will still have atleast one ref 1444 * added because it is in the hash __dlm_insert_lockres() */ 1445 extra_refs++; 1446 1447 /* now that the new lockres is inserted, 1448 * make it usable by other processes */ 1449 spin_lock(&res->spinlock); 1450 res->state &= ~DLM_LOCK_RES_IN_PROGRESS; 1451 spin_unlock(&res->spinlock); 1452 wake_up(&res->wq); 1453 } 1454 1455 /* at this point we have allocated everything we need, 1456 * and we have a hashed lockres with an extra ref and 1457 * the proper res->state flags. */ 1458 ret = 0; 1459 spin_lock(&res->spinlock); 1460 /* drop this either when master requery finds a different master 1461 * or when a lock is added by the recovery worker */ 1462 dlm_lockres_grab_inflight_ref(dlm, res); 1463 if (mres->master == DLM_LOCK_RES_OWNER_UNKNOWN) { 1464 /* migration cannot have an unknown master */ 1465 BUG_ON(!(mres->flags & DLM_MRES_RECOVERY)); 1466 mlog(0, "recovery has passed me a lockres with an " 1467 "unknown owner.. will need to requery: " 1468 "%.*s\n", mres->lockname_len, mres->lockname); 1469 } else { 1470 /* take a reference now to pin the lockres, drop it 1471 * when locks are added in the worker */ 1472 dlm_change_lockres_owner(dlm, res, dlm->node_num); 1473 } 1474 spin_unlock(&res->spinlock); 1475 1476 /* queue up work for dlm_mig_lockres_worker */ 1477 dlm_grab(dlm); /* get an extra ref for the work item */ 1478 memcpy(buf, msg->buf, be16_to_cpu(msg->data_len)); /* copy the whole message */ 1479 dlm_init_work_item(dlm, item, dlm_mig_lockres_worker, buf); 1480 item->u.ml.lockres = res; /* already have a ref */ 1481 item->u.ml.real_master = real_master; 1482 item->u.ml.extra_ref = extra_refs; 1483 spin_lock(&dlm->work_lock); 1484 list_add_tail(&item->list, &dlm->work_list); 1485 spin_unlock(&dlm->work_lock); 1486 queue_work(dlm->dlm_worker, &dlm->dispatched_work); 1487 1488 leave: 1489 /* One extra ref taken needs to be put here */ 1490 if (extra_refs) 1491 dlm_lockres_put(res); 1492 1493 dlm_put(dlm); 1494 if (ret < 0) { 1495 if (buf) 1496 kfree(buf); 1497 if (item) 1498 kfree(item); 1499 } 1500 1501 mlog_exit(ret); 1502 return ret; 1503 } 1504 1505 1506 static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data) 1507 { 1508 struct dlm_ctxt *dlm; 1509 struct dlm_migratable_lockres *mres; 1510 int ret = 0; 1511 struct dlm_lock_resource *res; 1512 u8 real_master; 1513 u8 extra_ref; 1514 1515 dlm = item->dlm; 1516 mres = (struct dlm_migratable_lockres *)data; 1517 1518 res = item->u.ml.lockres; 1519 real_master = item->u.ml.real_master; 1520 extra_ref = item->u.ml.extra_ref; 1521 1522 if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) { 1523 /* this case is super-rare. only occurs if 1524 * node death happens during migration. */ 1525 again: 1526 ret = dlm_lockres_master_requery(dlm, res, &real_master); 1527 if (ret < 0) { 1528 mlog(0, "dlm_lockres_master_requery ret=%d\n", 1529 ret); 1530 goto again; 1531 } 1532 if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) { 1533 mlog(0, "lockres %.*s not claimed. " 1534 "this node will take it.\n", 1535 res->lockname.len, res->lockname.name); 1536 } else { 1537 spin_lock(&res->spinlock); 1538 dlm_lockres_drop_inflight_ref(dlm, res); 1539 spin_unlock(&res->spinlock); 1540 mlog(0, "master needs to respond to sender " 1541 "that node %u still owns %.*s\n", 1542 real_master, res->lockname.len, 1543 res->lockname.name); 1544 /* cannot touch this lockres */ 1545 goto leave; 1546 } 1547 } 1548 1549 ret = dlm_process_recovery_data(dlm, res, mres); 1550 if (ret < 0) 1551 mlog(0, "dlm_process_recovery_data returned %d\n", ret); 1552 else 1553 mlog(0, "dlm_process_recovery_data succeeded\n"); 1554 1555 if ((mres->flags & (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) == 1556 (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) { 1557 ret = dlm_finish_migration(dlm, res, mres->master); 1558 if (ret < 0) 1559 mlog_errno(ret); 1560 } 1561 1562 leave: 1563 /* See comment in dlm_mig_lockres_handler() */ 1564 if (res) { 1565 if (extra_ref) 1566 dlm_lockres_put(res); 1567 dlm_lockres_put(res); 1568 } 1569 kfree(data); 1570 mlog_exit(ret); 1571 } 1572 1573 1574 1575 static int dlm_lockres_master_requery(struct dlm_ctxt *dlm, 1576 struct dlm_lock_resource *res, 1577 u8 *real_master) 1578 { 1579 struct dlm_node_iter iter; 1580 int nodenum; 1581 int ret = 0; 1582 1583 *real_master = DLM_LOCK_RES_OWNER_UNKNOWN; 1584 1585 /* we only reach here if one of the two nodes in a 1586 * migration died while the migration was in progress. 1587 * at this point we need to requery the master. we 1588 * know that the new_master got as far as creating 1589 * an mle on at least one node, but we do not know 1590 * if any nodes had actually cleared the mle and set 1591 * the master to the new_master. the old master 1592 * is supposed to set the owner to UNKNOWN in the 1593 * event of a new_master death, so the only possible 1594 * responses that we can get from nodes here are 1595 * that the master is new_master, or that the master 1596 * is UNKNOWN. 1597 * if all nodes come back with UNKNOWN then we know 1598 * the lock needs remastering here. 1599 * if any node comes back with a valid master, check 1600 * to see if that master is the one that we are 1601 * recovering. if so, then the new_master died and 1602 * we need to remaster this lock. if not, then the 1603 * new_master survived and that node will respond to 1604 * other nodes about the owner. 1605 * if there is an owner, this node needs to dump this 1606 * lockres and alert the sender that this lockres 1607 * was rejected. */ 1608 spin_lock(&dlm->spinlock); 1609 dlm_node_iter_init(dlm->domain_map, &iter); 1610 spin_unlock(&dlm->spinlock); 1611 1612 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) { 1613 /* do not send to self */ 1614 if (nodenum == dlm->node_num) 1615 continue; 1616 ret = dlm_do_master_requery(dlm, res, nodenum, real_master); 1617 if (ret < 0) { 1618 mlog_errno(ret); 1619 if (!dlm_is_host_down(ret)) 1620 BUG(); 1621 /* host is down, so answer for that node would be 1622 * DLM_LOCK_RES_OWNER_UNKNOWN. continue. */ 1623 } 1624 if (*real_master != DLM_LOCK_RES_OWNER_UNKNOWN) { 1625 mlog(0, "lock master is %u\n", *real_master); 1626 break; 1627 } 1628 } 1629 return ret; 1630 } 1631 1632 1633 int dlm_do_master_requery(struct dlm_ctxt *dlm, struct dlm_lock_resource *res, 1634 u8 nodenum, u8 *real_master) 1635 { 1636 int ret = -EINVAL; 1637 struct dlm_master_requery req; 1638 int status = DLM_LOCK_RES_OWNER_UNKNOWN; 1639 1640 memset(&req, 0, sizeof(req)); 1641 req.node_idx = dlm->node_num; 1642 req.namelen = res->lockname.len; 1643 memcpy(req.name, res->lockname.name, res->lockname.len); 1644 1645 ret = o2net_send_message(DLM_MASTER_REQUERY_MSG, dlm->key, 1646 &req, sizeof(req), nodenum, &status); 1647 /* XXX: negative status not handled properly here. */ 1648 if (ret < 0) 1649 mlog(ML_ERROR, "Error %d when sending message %u (key " 1650 "0x%x) to node %u\n", ret, DLM_MASTER_REQUERY_MSG, 1651 dlm->key, nodenum); 1652 else { 1653 BUG_ON(status < 0); 1654 BUG_ON(status > DLM_LOCK_RES_OWNER_UNKNOWN); 1655 *real_master = (u8) (status & 0xff); 1656 mlog(0, "node %u responded to master requery with %u\n", 1657 nodenum, *real_master); 1658 ret = 0; 1659 } 1660 return ret; 1661 } 1662 1663 1664 /* this function cannot error, so unless the sending 1665 * or receiving of the message failed, the owner can 1666 * be trusted */ 1667 int dlm_master_requery_handler(struct o2net_msg *msg, u32 len, void *data, 1668 void **ret_data) 1669 { 1670 struct dlm_ctxt *dlm = data; 1671 struct dlm_master_requery *req = (struct dlm_master_requery *)msg->buf; 1672 struct dlm_lock_resource *res = NULL; 1673 unsigned int hash; 1674 int master = DLM_LOCK_RES_OWNER_UNKNOWN; 1675 u32 flags = DLM_ASSERT_MASTER_REQUERY; 1676 1677 if (!dlm_grab(dlm)) { 1678 /* since the domain has gone away on this 1679 * node, the proper response is UNKNOWN */ 1680 return master; 1681 } 1682 1683 hash = dlm_lockid_hash(req->name, req->namelen); 1684 1685 spin_lock(&dlm->spinlock); 1686 res = __dlm_lookup_lockres(dlm, req->name, req->namelen, hash); 1687 if (res) { 1688 spin_lock(&res->spinlock); 1689 master = res->owner; 1690 if (master == dlm->node_num) { 1691 int ret = dlm_dispatch_assert_master(dlm, res, 1692 0, 0, flags); 1693 if (ret < 0) { 1694 mlog_errno(-ENOMEM); 1695 /* retry!? */ 1696 BUG(); 1697 } 1698 } else /* put.. incase we are not the master */ 1699 dlm_lockres_put(res); 1700 spin_unlock(&res->spinlock); 1701 } 1702 spin_unlock(&dlm->spinlock); 1703 1704 dlm_put(dlm); 1705 return master; 1706 } 1707 1708 static inline struct list_head * 1709 dlm_list_num_to_pointer(struct dlm_lock_resource *res, int list_num) 1710 { 1711 struct list_head *ret; 1712 BUG_ON(list_num < 0); 1713 BUG_ON(list_num > 2); 1714 ret = &(res->granted); 1715 ret += list_num; 1716 return ret; 1717 } 1718 /* TODO: do ast flush business 1719 * TODO: do MIGRATING and RECOVERING spinning 1720 */ 1721 1722 /* 1723 * NOTE about in-flight requests during migration: 1724 * 1725 * Before attempting the migrate, the master has marked the lockres as 1726 * MIGRATING and then flushed all of its pending ASTS. So any in-flight 1727 * requests either got queued before the MIGRATING flag got set, in which 1728 * case the lock data will reflect the change and a return message is on 1729 * the way, or the request failed to get in before MIGRATING got set. In 1730 * this case, the caller will be told to spin and wait for the MIGRATING 1731 * flag to be dropped, then recheck the master. 1732 * This holds true for the convert, cancel and unlock cases, and since lvb 1733 * updates are tied to these same messages, it applies to lvb updates as 1734 * well. For the lock case, there is no way a lock can be on the master 1735 * queue and not be on the secondary queue since the lock is always added 1736 * locally first. This means that the new target node will never be sent 1737 * a lock that he doesn't already have on the list. 1738 * In total, this means that the local lock is correct and should not be 1739 * updated to match the one sent by the master. Any messages sent back 1740 * from the master before the MIGRATING flag will bring the lock properly 1741 * up-to-date, and the change will be ordered properly for the waiter. 1742 * We will *not* attempt to modify the lock underneath the waiter. 1743 */ 1744 1745 static int dlm_process_recovery_data(struct dlm_ctxt *dlm, 1746 struct dlm_lock_resource *res, 1747 struct dlm_migratable_lockres *mres) 1748 { 1749 struct dlm_migratable_lock *ml; 1750 struct list_head *queue; 1751 struct list_head *tmpq = NULL; 1752 struct dlm_lock *newlock = NULL; 1753 struct dlm_lockstatus *lksb = NULL; 1754 int ret = 0; 1755 int i, j, bad; 1756 struct dlm_lock *lock = NULL; 1757 u8 from = O2NM_MAX_NODES; 1758 unsigned int added = 0; 1759 __be64 c; 1760 1761 mlog(0, "running %d locks for this lockres\n", mres->num_locks); 1762 for (i=0; i<mres->num_locks; i++) { 1763 ml = &(mres->ml[i]); 1764 1765 if (dlm_is_dummy_lock(dlm, ml, &from)) { 1766 /* placeholder, just need to set the refmap bit */ 1767 BUG_ON(mres->num_locks != 1); 1768 mlog(0, "%s:%.*s: dummy lock for %u\n", 1769 dlm->name, mres->lockname_len, mres->lockname, 1770 from); 1771 spin_lock(&res->spinlock); 1772 dlm_lockres_set_refmap_bit(from, res); 1773 spin_unlock(&res->spinlock); 1774 added++; 1775 break; 1776 } 1777 BUG_ON(ml->highest_blocked != LKM_IVMODE); 1778 newlock = NULL; 1779 lksb = NULL; 1780 1781 queue = dlm_list_num_to_pointer(res, ml->list); 1782 tmpq = NULL; 1783 1784 /* if the lock is for the local node it needs to 1785 * be moved to the proper location within the queue. 1786 * do not allocate a new lock structure. */ 1787 if (ml->node == dlm->node_num) { 1788 /* MIGRATION ONLY! */ 1789 BUG_ON(!(mres->flags & DLM_MRES_MIGRATION)); 1790 1791 spin_lock(&res->spinlock); 1792 for (j = DLM_GRANTED_LIST; j <= DLM_BLOCKED_LIST; j++) { 1793 tmpq = dlm_list_idx_to_ptr(res, j); 1794 list_for_each_entry(lock, tmpq, list) { 1795 if (lock->ml.cookie != ml->cookie) 1796 lock = NULL; 1797 else 1798 break; 1799 } 1800 if (lock) 1801 break; 1802 } 1803 1804 /* lock is always created locally first, and 1805 * destroyed locally last. it must be on the list */ 1806 if (!lock) { 1807 c = ml->cookie; 1808 mlog(ML_ERROR, "Could not find local lock " 1809 "with cookie %u:%llu, node %u, " 1810 "list %u, flags 0x%x, type %d, " 1811 "conv %d, highest blocked %d\n", 1812 dlm_get_lock_cookie_node(be64_to_cpu(c)), 1813 dlm_get_lock_cookie_seq(be64_to_cpu(c)), 1814 ml->node, ml->list, ml->flags, ml->type, 1815 ml->convert_type, ml->highest_blocked); 1816 __dlm_print_one_lock_resource(res); 1817 BUG(); 1818 } 1819 1820 if (lock->ml.node != ml->node) { 1821 c = lock->ml.cookie; 1822 mlog(ML_ERROR, "Mismatched node# in lock " 1823 "cookie %u:%llu, name %.*s, node %u\n", 1824 dlm_get_lock_cookie_node(be64_to_cpu(c)), 1825 dlm_get_lock_cookie_seq(be64_to_cpu(c)), 1826 res->lockname.len, res->lockname.name, 1827 lock->ml.node); 1828 c = ml->cookie; 1829 mlog(ML_ERROR, "Migrate lock cookie %u:%llu, " 1830 "node %u, list %u, flags 0x%x, type %d, " 1831 "conv %d, highest blocked %d\n", 1832 dlm_get_lock_cookie_node(be64_to_cpu(c)), 1833 dlm_get_lock_cookie_seq(be64_to_cpu(c)), 1834 ml->node, ml->list, ml->flags, ml->type, 1835 ml->convert_type, ml->highest_blocked); 1836 __dlm_print_one_lock_resource(res); 1837 BUG(); 1838 } 1839 1840 if (tmpq != queue) { 1841 c = ml->cookie; 1842 mlog(0, "Lock cookie %u:%llu was on list %u " 1843 "instead of list %u for %.*s\n", 1844 dlm_get_lock_cookie_node(be64_to_cpu(c)), 1845 dlm_get_lock_cookie_seq(be64_to_cpu(c)), 1846 j, ml->list, res->lockname.len, 1847 res->lockname.name); 1848 __dlm_print_one_lock_resource(res); 1849 spin_unlock(&res->spinlock); 1850 continue; 1851 } 1852 1853 /* see NOTE above about why we do not update 1854 * to match the master here */ 1855 1856 /* move the lock to its proper place */ 1857 /* do not alter lock refcount. switching lists. */ 1858 list_move_tail(&lock->list, queue); 1859 spin_unlock(&res->spinlock); 1860 added++; 1861 1862 mlog(0, "just reordered a local lock!\n"); 1863 continue; 1864 } 1865 1866 /* lock is for another node. */ 1867 newlock = dlm_new_lock(ml->type, ml->node, 1868 be64_to_cpu(ml->cookie), NULL); 1869 if (!newlock) { 1870 ret = -ENOMEM; 1871 goto leave; 1872 } 1873 lksb = newlock->lksb; 1874 dlm_lock_attach_lockres(newlock, res); 1875 1876 if (ml->convert_type != LKM_IVMODE) { 1877 BUG_ON(queue != &res->converting); 1878 newlock->ml.convert_type = ml->convert_type; 1879 } 1880 lksb->flags |= (ml->flags & 1881 (DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB)); 1882 1883 if (ml->type == LKM_NLMODE) 1884 goto skip_lvb; 1885 1886 if (!dlm_lvb_is_empty(mres->lvb)) { 1887 if (lksb->flags & DLM_LKSB_PUT_LVB) { 1888 /* other node was trying to update 1889 * lvb when node died. recreate the 1890 * lksb with the updated lvb. */ 1891 memcpy(lksb->lvb, mres->lvb, DLM_LVB_LEN); 1892 /* the lock resource lvb update must happen 1893 * NOW, before the spinlock is dropped. 1894 * we no longer wait for the AST to update 1895 * the lvb. */ 1896 memcpy(res->lvb, mres->lvb, DLM_LVB_LEN); 1897 } else { 1898 /* otherwise, the node is sending its 1899 * most recent valid lvb info */ 1900 BUG_ON(ml->type != LKM_EXMODE && 1901 ml->type != LKM_PRMODE); 1902 if (!dlm_lvb_is_empty(res->lvb) && 1903 (ml->type == LKM_EXMODE || 1904 memcmp(res->lvb, mres->lvb, DLM_LVB_LEN))) { 1905 int i; 1906 mlog(ML_ERROR, "%s:%.*s: received bad " 1907 "lvb! type=%d\n", dlm->name, 1908 res->lockname.len, 1909 res->lockname.name, ml->type); 1910 printk("lockres lvb=["); 1911 for (i=0; i<DLM_LVB_LEN; i++) 1912 printk("%02x", res->lvb[i]); 1913 printk("]\nmigrated lvb=["); 1914 for (i=0; i<DLM_LVB_LEN; i++) 1915 printk("%02x", mres->lvb[i]); 1916 printk("]\n"); 1917 dlm_print_one_lock_resource(res); 1918 BUG(); 1919 } 1920 memcpy(res->lvb, mres->lvb, DLM_LVB_LEN); 1921 } 1922 } 1923 skip_lvb: 1924 1925 /* NOTE: 1926 * wrt lock queue ordering and recovery: 1927 * 1. order of locks on granted queue is 1928 * meaningless. 1929 * 2. order of locks on converting queue is 1930 * LOST with the node death. sorry charlie. 1931 * 3. order of locks on the blocked queue is 1932 * also LOST. 1933 * order of locks does not affect integrity, it 1934 * just means that a lock request may get pushed 1935 * back in line as a result of the node death. 1936 * also note that for a given node the lock order 1937 * for its secondary queue locks is preserved 1938 * relative to each other, but clearly *not* 1939 * preserved relative to locks from other nodes. 1940 */ 1941 bad = 0; 1942 spin_lock(&res->spinlock); 1943 list_for_each_entry(lock, queue, list) { 1944 if (lock->ml.cookie == ml->cookie) { 1945 c = lock->ml.cookie; 1946 mlog(ML_ERROR, "%s:%.*s: %u:%llu: lock already " 1947 "exists on this lockres!\n", dlm->name, 1948 res->lockname.len, res->lockname.name, 1949 dlm_get_lock_cookie_node(be64_to_cpu(c)), 1950 dlm_get_lock_cookie_seq(be64_to_cpu(c))); 1951 1952 mlog(ML_NOTICE, "sent lock: type=%d, conv=%d, " 1953 "node=%u, cookie=%u:%llu, queue=%d\n", 1954 ml->type, ml->convert_type, ml->node, 1955 dlm_get_lock_cookie_node(be64_to_cpu(ml->cookie)), 1956 dlm_get_lock_cookie_seq(be64_to_cpu(ml->cookie)), 1957 ml->list); 1958 1959 __dlm_print_one_lock_resource(res); 1960 bad = 1; 1961 break; 1962 } 1963 } 1964 if (!bad) { 1965 dlm_lock_get(newlock); 1966 list_add_tail(&newlock->list, queue); 1967 mlog(0, "%s:%.*s: added lock for node %u, " 1968 "setting refmap bit\n", dlm->name, 1969 res->lockname.len, res->lockname.name, ml->node); 1970 dlm_lockres_set_refmap_bit(ml->node, res); 1971 added++; 1972 } 1973 spin_unlock(&res->spinlock); 1974 } 1975 mlog(0, "done running all the locks\n"); 1976 1977 leave: 1978 /* balance the ref taken when the work was queued */ 1979 spin_lock(&res->spinlock); 1980 dlm_lockres_drop_inflight_ref(dlm, res); 1981 spin_unlock(&res->spinlock); 1982 1983 if (ret < 0) { 1984 mlog_errno(ret); 1985 if (newlock) 1986 dlm_lock_put(newlock); 1987 } 1988 1989 mlog_exit(ret); 1990 return ret; 1991 } 1992 1993 void dlm_move_lockres_to_recovery_list(struct dlm_ctxt *dlm, 1994 struct dlm_lock_resource *res) 1995 { 1996 int i; 1997 struct list_head *queue; 1998 struct dlm_lock *lock, *next; 1999 2000 assert_spin_locked(&dlm->spinlock); 2001 assert_spin_locked(&res->spinlock); 2002 res->state |= DLM_LOCK_RES_RECOVERING; 2003 if (!list_empty(&res->recovering)) { 2004 mlog(0, 2005 "Recovering res %s:%.*s, is already on recovery list!\n", 2006 dlm->name, res->lockname.len, res->lockname.name); 2007 list_del_init(&res->recovering); 2008 dlm_lockres_put(res); 2009 } 2010 /* We need to hold a reference while on the recovery list */ 2011 dlm_lockres_get(res); 2012 list_add_tail(&res->recovering, &dlm->reco.resources); 2013 2014 /* find any pending locks and put them back on proper list */ 2015 for (i=DLM_BLOCKED_LIST; i>=DLM_GRANTED_LIST; i--) { 2016 queue = dlm_list_idx_to_ptr(res, i); 2017 list_for_each_entry_safe(lock, next, queue, list) { 2018 dlm_lock_get(lock); 2019 if (lock->convert_pending) { 2020 /* move converting lock back to granted */ 2021 BUG_ON(i != DLM_CONVERTING_LIST); 2022 mlog(0, "node died with convert pending " 2023 "on %.*s. move back to granted list.\n", 2024 res->lockname.len, res->lockname.name); 2025 dlm_revert_pending_convert(res, lock); 2026 lock->convert_pending = 0; 2027 } else if (lock->lock_pending) { 2028 /* remove pending lock requests completely */ 2029 BUG_ON(i != DLM_BLOCKED_LIST); 2030 mlog(0, "node died with lock pending " 2031 "on %.*s. remove from blocked list and skip.\n", 2032 res->lockname.len, res->lockname.name); 2033 /* lock will be floating until ref in 2034 * dlmlock_remote is freed after the network 2035 * call returns. ok for it to not be on any 2036 * list since no ast can be called 2037 * (the master is dead). */ 2038 dlm_revert_pending_lock(res, lock); 2039 lock->lock_pending = 0; 2040 } else if (lock->unlock_pending) { 2041 /* if an unlock was in progress, treat as 2042 * if this had completed successfully 2043 * before sending this lock state to the 2044 * new master. note that the dlm_unlock 2045 * call is still responsible for calling 2046 * the unlockast. that will happen after 2047 * the network call times out. for now, 2048 * just move lists to prepare the new 2049 * recovery master. */ 2050 BUG_ON(i != DLM_GRANTED_LIST); 2051 mlog(0, "node died with unlock pending " 2052 "on %.*s. remove from blocked list and skip.\n", 2053 res->lockname.len, res->lockname.name); 2054 dlm_commit_pending_unlock(res, lock); 2055 lock->unlock_pending = 0; 2056 } else if (lock->cancel_pending) { 2057 /* if a cancel was in progress, treat as 2058 * if this had completed successfully 2059 * before sending this lock state to the 2060 * new master */ 2061 BUG_ON(i != DLM_CONVERTING_LIST); 2062 mlog(0, "node died with cancel pending " 2063 "on %.*s. move back to granted list.\n", 2064 res->lockname.len, res->lockname.name); 2065 dlm_commit_pending_cancel(res, lock); 2066 lock->cancel_pending = 0; 2067 } 2068 dlm_lock_put(lock); 2069 } 2070 } 2071 } 2072 2073 2074 2075 /* removes all recovered locks from the recovery list. 2076 * sets the res->owner to the new master. 2077 * unsets the RECOVERY flag and wakes waiters. */ 2078 static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm, 2079 u8 dead_node, u8 new_master) 2080 { 2081 int i; 2082 struct hlist_node *hash_iter; 2083 struct hlist_head *bucket; 2084 struct dlm_lock_resource *res, *next; 2085 2086 mlog_entry_void(); 2087 2088 assert_spin_locked(&dlm->spinlock); 2089 2090 list_for_each_entry_safe(res, next, &dlm->reco.resources, recovering) { 2091 if (res->owner == dead_node) { 2092 list_del_init(&res->recovering); 2093 spin_lock(&res->spinlock); 2094 /* new_master has our reference from 2095 * the lock state sent during recovery */ 2096 dlm_change_lockres_owner(dlm, res, new_master); 2097 res->state &= ~DLM_LOCK_RES_RECOVERING; 2098 if (__dlm_lockres_has_locks(res)) 2099 __dlm_dirty_lockres(dlm, res); 2100 spin_unlock(&res->spinlock); 2101 wake_up(&res->wq); 2102 dlm_lockres_put(res); 2103 } 2104 } 2105 2106 /* this will become unnecessary eventually, but 2107 * for now we need to run the whole hash, clear 2108 * the RECOVERING state and set the owner 2109 * if necessary */ 2110 for (i = 0; i < DLM_HASH_BUCKETS; i++) { 2111 bucket = dlm_lockres_hash(dlm, i); 2112 hlist_for_each_entry(res, hash_iter, bucket, hash_node) { 2113 if (res->state & DLM_LOCK_RES_RECOVERING) { 2114 if (res->owner == dead_node) { 2115 mlog(0, "(this=%u) res %.*s owner=%u " 2116 "was not on recovering list, but " 2117 "clearing state anyway\n", 2118 dlm->node_num, res->lockname.len, 2119 res->lockname.name, new_master); 2120 } else if (res->owner == dlm->node_num) { 2121 mlog(0, "(this=%u) res %.*s owner=%u " 2122 "was not on recovering list, " 2123 "owner is THIS node, clearing\n", 2124 dlm->node_num, res->lockname.len, 2125 res->lockname.name, new_master); 2126 } else 2127 continue; 2128 2129 if (!list_empty(&res->recovering)) { 2130 mlog(0, "%s:%.*s: lockres was " 2131 "marked RECOVERING, owner=%u\n", 2132 dlm->name, res->lockname.len, 2133 res->lockname.name, res->owner); 2134 list_del_init(&res->recovering); 2135 dlm_lockres_put(res); 2136 } 2137 spin_lock(&res->spinlock); 2138 /* new_master has our reference from 2139 * the lock state sent during recovery */ 2140 dlm_change_lockres_owner(dlm, res, new_master); 2141 res->state &= ~DLM_LOCK_RES_RECOVERING; 2142 if (__dlm_lockres_has_locks(res)) 2143 __dlm_dirty_lockres(dlm, res); 2144 spin_unlock(&res->spinlock); 2145 wake_up(&res->wq); 2146 } 2147 } 2148 } 2149 } 2150 2151 static inline int dlm_lvb_needs_invalidation(struct dlm_lock *lock, int local) 2152 { 2153 if (local) { 2154 if (lock->ml.type != LKM_EXMODE && 2155 lock->ml.type != LKM_PRMODE) 2156 return 1; 2157 } else if (lock->ml.type == LKM_EXMODE) 2158 return 1; 2159 return 0; 2160 } 2161 2162 static void dlm_revalidate_lvb(struct dlm_ctxt *dlm, 2163 struct dlm_lock_resource *res, u8 dead_node) 2164 { 2165 struct list_head *queue; 2166 struct dlm_lock *lock; 2167 int blank_lvb = 0, local = 0; 2168 int i; 2169 u8 search_node; 2170 2171 assert_spin_locked(&dlm->spinlock); 2172 assert_spin_locked(&res->spinlock); 2173 2174 if (res->owner == dlm->node_num) 2175 /* if this node owned the lockres, and if the dead node 2176 * had an EX when he died, blank out the lvb */ 2177 search_node = dead_node; 2178 else { 2179 /* if this is a secondary lockres, and we had no EX or PR 2180 * locks granted, we can no longer trust the lvb */ 2181 search_node = dlm->node_num; 2182 local = 1; /* check local state for valid lvb */ 2183 } 2184 2185 for (i=DLM_GRANTED_LIST; i<=DLM_CONVERTING_LIST; i++) { 2186 queue = dlm_list_idx_to_ptr(res, i); 2187 list_for_each_entry(lock, queue, list) { 2188 if (lock->ml.node == search_node) { 2189 if (dlm_lvb_needs_invalidation(lock, local)) { 2190 /* zero the lksb lvb and lockres lvb */ 2191 blank_lvb = 1; 2192 memset(lock->lksb->lvb, 0, DLM_LVB_LEN); 2193 } 2194 } 2195 } 2196 } 2197 2198 if (blank_lvb) { 2199 mlog(0, "clearing %.*s lvb, dead node %u had EX\n", 2200 res->lockname.len, res->lockname.name, dead_node); 2201 memset(res->lvb, 0, DLM_LVB_LEN); 2202 } 2203 } 2204 2205 static void dlm_free_dead_locks(struct dlm_ctxt *dlm, 2206 struct dlm_lock_resource *res, u8 dead_node) 2207 { 2208 struct dlm_lock *lock, *next; 2209 unsigned int freed = 0; 2210 2211 /* this node is the lockres master: 2212 * 1) remove any stale locks for the dead node 2213 * 2) if the dead node had an EX when he died, blank out the lvb 2214 */ 2215 assert_spin_locked(&dlm->spinlock); 2216 assert_spin_locked(&res->spinlock); 2217 2218 /* We do two dlm_lock_put(). One for removing from list and the other is 2219 * to force the DLM_UNLOCK_FREE_LOCK action so as to free the locks */ 2220 2221 /* TODO: check pending_asts, pending_basts here */ 2222 list_for_each_entry_safe(lock, next, &res->granted, list) { 2223 if (lock->ml.node == dead_node) { 2224 list_del_init(&lock->list); 2225 dlm_lock_put(lock); 2226 /* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */ 2227 dlm_lock_put(lock); 2228 freed++; 2229 } 2230 } 2231 list_for_each_entry_safe(lock, next, &res->converting, list) { 2232 if (lock->ml.node == dead_node) { 2233 list_del_init(&lock->list); 2234 dlm_lock_put(lock); 2235 /* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */ 2236 dlm_lock_put(lock); 2237 freed++; 2238 } 2239 } 2240 list_for_each_entry_safe(lock, next, &res->blocked, list) { 2241 if (lock->ml.node == dead_node) { 2242 list_del_init(&lock->list); 2243 dlm_lock_put(lock); 2244 /* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */ 2245 dlm_lock_put(lock); 2246 freed++; 2247 } 2248 } 2249 2250 if (freed) { 2251 mlog(0, "%s:%.*s: freed %u locks for dead node %u, " 2252 "dropping ref from lockres\n", dlm->name, 2253 res->lockname.len, res->lockname.name, freed, dead_node); 2254 if(!test_bit(dead_node, res->refmap)) { 2255 mlog(ML_ERROR, "%s:%.*s: freed %u locks for dead node %u, " 2256 "but ref was not set\n", dlm->name, 2257 res->lockname.len, res->lockname.name, freed, dead_node); 2258 __dlm_print_one_lock_resource(res); 2259 } 2260 dlm_lockres_clear_refmap_bit(dead_node, res); 2261 } else if (test_bit(dead_node, res->refmap)) { 2262 mlog(0, "%s:%.*s: dead node %u had a ref, but had " 2263 "no locks and had not purged before dying\n", dlm->name, 2264 res->lockname.len, res->lockname.name, dead_node); 2265 dlm_lockres_clear_refmap_bit(dead_node, res); 2266 } 2267 2268 /* do not kick thread yet */ 2269 __dlm_dirty_lockres(dlm, res); 2270 } 2271 2272 /* if this node is the recovery master, and there are no 2273 * locks for a given lockres owned by this node that are in 2274 * either PR or EX mode, zero out the lvb before requesting. 2275 * 2276 */ 2277 2278 2279 static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node) 2280 { 2281 struct hlist_node *iter; 2282 struct dlm_lock_resource *res; 2283 int i; 2284 struct hlist_head *bucket; 2285 struct dlm_lock *lock; 2286 2287 2288 /* purge any stale mles */ 2289 dlm_clean_master_list(dlm, dead_node); 2290 2291 /* 2292 * now clean up all lock resources. there are two rules: 2293 * 2294 * 1) if the dead node was the master, move the lockres 2295 * to the recovering list. set the RECOVERING flag. 2296 * this lockres needs to be cleaned up before it can 2297 * be used further. 2298 * 2299 * 2) if this node was the master, remove all locks from 2300 * each of the lockres queues that were owned by the 2301 * dead node. once recovery finishes, the dlm thread 2302 * can be kicked again to see if any ASTs or BASTs 2303 * need to be fired as a result. 2304 */ 2305 for (i = 0; i < DLM_HASH_BUCKETS; i++) { 2306 bucket = dlm_lockres_hash(dlm, i); 2307 hlist_for_each_entry(res, iter, bucket, hash_node) { 2308 /* always prune any $RECOVERY entries for dead nodes, 2309 * otherwise hangs can occur during later recovery */ 2310 if (dlm_is_recovery_lock(res->lockname.name, 2311 res->lockname.len)) { 2312 spin_lock(&res->spinlock); 2313 list_for_each_entry(lock, &res->granted, list) { 2314 if (lock->ml.node == dead_node) { 2315 mlog(0, "AHA! there was " 2316 "a $RECOVERY lock for dead " 2317 "node %u (%s)!\n", 2318 dead_node, dlm->name); 2319 list_del_init(&lock->list); 2320 dlm_lock_put(lock); 2321 break; 2322 } 2323 } 2324 spin_unlock(&res->spinlock); 2325 continue; 2326 } 2327 spin_lock(&res->spinlock); 2328 /* zero the lvb if necessary */ 2329 dlm_revalidate_lvb(dlm, res, dead_node); 2330 if (res->owner == dead_node) { 2331 if (res->state & DLM_LOCK_RES_DROPPING_REF) { 2332 mlog(ML_NOTICE, "Ignore %.*s for " 2333 "recovery as it is being freed\n", 2334 res->lockname.len, 2335 res->lockname.name); 2336 } else 2337 dlm_move_lockres_to_recovery_list(dlm, 2338 res); 2339 2340 } else if (res->owner == dlm->node_num) { 2341 dlm_free_dead_locks(dlm, res, dead_node); 2342 __dlm_lockres_calc_usage(dlm, res); 2343 } 2344 spin_unlock(&res->spinlock); 2345 } 2346 } 2347 2348 } 2349 2350 static void __dlm_hb_node_down(struct dlm_ctxt *dlm, int idx) 2351 { 2352 assert_spin_locked(&dlm->spinlock); 2353 2354 if (dlm->reco.new_master == idx) { 2355 mlog(0, "%s: recovery master %d just died\n", 2356 dlm->name, idx); 2357 if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) { 2358 /* finalize1 was reached, so it is safe to clear 2359 * the new_master and dead_node. that recovery 2360 * is complete. */ 2361 mlog(0, "%s: dead master %d had reached " 2362 "finalize1 state, clearing\n", dlm->name, idx); 2363 dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE; 2364 __dlm_reset_recovery(dlm); 2365 } 2366 } 2367 2368 /* Clean up join state on node death. */ 2369 if (dlm->joining_node == idx) { 2370 mlog(0, "Clearing join state for node %u\n", idx); 2371 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN); 2372 } 2373 2374 /* check to see if the node is already considered dead */ 2375 if (!test_bit(idx, dlm->live_nodes_map)) { 2376 mlog(0, "for domain %s, node %d is already dead. " 2377 "another node likely did recovery already.\n", 2378 dlm->name, idx); 2379 return; 2380 } 2381 2382 /* check to see if we do not care about this node */ 2383 if (!test_bit(idx, dlm->domain_map)) { 2384 /* This also catches the case that we get a node down 2385 * but haven't joined the domain yet. */ 2386 mlog(0, "node %u already removed from domain!\n", idx); 2387 return; 2388 } 2389 2390 clear_bit(idx, dlm->live_nodes_map); 2391 2392 /* make sure local cleanup occurs before the heartbeat events */ 2393 if (!test_bit(idx, dlm->recovery_map)) 2394 dlm_do_local_recovery_cleanup(dlm, idx); 2395 2396 /* notify anything attached to the heartbeat events */ 2397 dlm_hb_event_notify_attached(dlm, idx, 0); 2398 2399 mlog(0, "node %u being removed from domain map!\n", idx); 2400 clear_bit(idx, dlm->domain_map); 2401 /* wake up migration waiters if a node goes down. 2402 * perhaps later we can genericize this for other waiters. */ 2403 wake_up(&dlm->migration_wq); 2404 2405 if (test_bit(idx, dlm->recovery_map)) 2406 mlog(0, "domain %s, node %u already added " 2407 "to recovery map!\n", dlm->name, idx); 2408 else 2409 set_bit(idx, dlm->recovery_map); 2410 } 2411 2412 void dlm_hb_node_down_cb(struct o2nm_node *node, int idx, void *data) 2413 { 2414 struct dlm_ctxt *dlm = data; 2415 2416 if (!dlm_grab(dlm)) 2417 return; 2418 2419 /* 2420 * This will notify any dlm users that a node in our domain 2421 * went away without notifying us first. 2422 */ 2423 if (test_bit(idx, dlm->domain_map)) 2424 dlm_fire_domain_eviction_callbacks(dlm, idx); 2425 2426 spin_lock(&dlm->spinlock); 2427 __dlm_hb_node_down(dlm, idx); 2428 spin_unlock(&dlm->spinlock); 2429 2430 dlm_put(dlm); 2431 } 2432 2433 void dlm_hb_node_up_cb(struct o2nm_node *node, int idx, void *data) 2434 { 2435 struct dlm_ctxt *dlm = data; 2436 2437 if (!dlm_grab(dlm)) 2438 return; 2439 2440 spin_lock(&dlm->spinlock); 2441 set_bit(idx, dlm->live_nodes_map); 2442 /* do NOT notify mle attached to the heartbeat events. 2443 * new nodes are not interesting in mastery until joined. */ 2444 spin_unlock(&dlm->spinlock); 2445 2446 dlm_put(dlm); 2447 } 2448 2449 static void dlm_reco_ast(void *astdata) 2450 { 2451 struct dlm_ctxt *dlm = astdata; 2452 mlog(0, "ast for recovery lock fired!, this=%u, dlm=%s\n", 2453 dlm->node_num, dlm->name); 2454 } 2455 static void dlm_reco_bast(void *astdata, int blocked_type) 2456 { 2457 struct dlm_ctxt *dlm = astdata; 2458 mlog(0, "bast for recovery lock fired!, this=%u, dlm=%s\n", 2459 dlm->node_num, dlm->name); 2460 } 2461 static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st) 2462 { 2463 mlog(0, "unlockast for recovery lock fired!\n"); 2464 } 2465 2466 /* 2467 * dlm_pick_recovery_master will continually attempt to use 2468 * dlmlock() on the special "$RECOVERY" lockres with the 2469 * LKM_NOQUEUE flag to get an EX. every thread that enters 2470 * this function on each node racing to become the recovery 2471 * master will not stop attempting this until either: 2472 * a) this node gets the EX (and becomes the recovery master), 2473 * or b) dlm->reco.new_master gets set to some nodenum 2474 * != O2NM_INVALID_NODE_NUM (another node will do the reco). 2475 * so each time a recovery master is needed, the entire cluster 2476 * will sync at this point. if the new master dies, that will 2477 * be detected in dlm_do_recovery */ 2478 static int dlm_pick_recovery_master(struct dlm_ctxt *dlm) 2479 { 2480 enum dlm_status ret; 2481 struct dlm_lockstatus lksb; 2482 int status = -EINVAL; 2483 2484 mlog(0, "starting recovery of %s at %lu, dead=%u, this=%u\n", 2485 dlm->name, jiffies, dlm->reco.dead_node, dlm->node_num); 2486 again: 2487 memset(&lksb, 0, sizeof(lksb)); 2488 2489 ret = dlmlock(dlm, LKM_EXMODE, &lksb, LKM_NOQUEUE|LKM_RECOVERY, 2490 DLM_RECOVERY_LOCK_NAME, DLM_RECOVERY_LOCK_NAME_LEN, 2491 dlm_reco_ast, dlm, dlm_reco_bast); 2492 2493 mlog(0, "%s: dlmlock($RECOVERY) returned %d, lksb=%d\n", 2494 dlm->name, ret, lksb.status); 2495 2496 if (ret == DLM_NORMAL) { 2497 mlog(0, "dlm=%s dlmlock says I got it (this=%u)\n", 2498 dlm->name, dlm->node_num); 2499 2500 /* got the EX lock. check to see if another node 2501 * just became the reco master */ 2502 if (dlm_reco_master_ready(dlm)) { 2503 mlog(0, "%s: got reco EX lock, but %u will " 2504 "do the recovery\n", dlm->name, 2505 dlm->reco.new_master); 2506 status = -EEXIST; 2507 } else { 2508 status = 0; 2509 2510 /* see if recovery was already finished elsewhere */ 2511 spin_lock(&dlm->spinlock); 2512 if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) { 2513 status = -EINVAL; 2514 mlog(0, "%s: got reco EX lock, but " 2515 "node got recovered already\n", dlm->name); 2516 if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) { 2517 mlog(ML_ERROR, "%s: new master is %u " 2518 "but no dead node!\n", 2519 dlm->name, dlm->reco.new_master); 2520 BUG(); 2521 } 2522 } 2523 spin_unlock(&dlm->spinlock); 2524 } 2525 2526 /* if this node has actually become the recovery master, 2527 * set the master and send the messages to begin recovery */ 2528 if (!status) { 2529 mlog(0, "%s: dead=%u, this=%u, sending " 2530 "begin_reco now\n", dlm->name, 2531 dlm->reco.dead_node, dlm->node_num); 2532 status = dlm_send_begin_reco_message(dlm, 2533 dlm->reco.dead_node); 2534 /* this always succeeds */ 2535 BUG_ON(status); 2536 2537 /* set the new_master to this node */ 2538 spin_lock(&dlm->spinlock); 2539 dlm_set_reco_master(dlm, dlm->node_num); 2540 spin_unlock(&dlm->spinlock); 2541 } 2542 2543 /* recovery lock is a special case. ast will not get fired, 2544 * so just go ahead and unlock it. */ 2545 ret = dlmunlock(dlm, &lksb, 0, dlm_reco_unlock_ast, dlm); 2546 if (ret == DLM_DENIED) { 2547 mlog(0, "got DLM_DENIED, trying LKM_CANCEL\n"); 2548 ret = dlmunlock(dlm, &lksb, LKM_CANCEL, dlm_reco_unlock_ast, dlm); 2549 } 2550 if (ret != DLM_NORMAL) { 2551 /* this would really suck. this could only happen 2552 * if there was a network error during the unlock 2553 * because of node death. this means the unlock 2554 * is actually "done" and the lock structure is 2555 * even freed. we can continue, but only 2556 * because this specific lock name is special. */ 2557 mlog(ML_ERROR, "dlmunlock returned %d\n", ret); 2558 } 2559 } else if (ret == DLM_NOTQUEUED) { 2560 mlog(0, "dlm=%s dlmlock says another node got it (this=%u)\n", 2561 dlm->name, dlm->node_num); 2562 /* another node is master. wait on 2563 * reco.new_master != O2NM_INVALID_NODE_NUM 2564 * for at most one second */ 2565 wait_event_timeout(dlm->dlm_reco_thread_wq, 2566 dlm_reco_master_ready(dlm), 2567 msecs_to_jiffies(1000)); 2568 if (!dlm_reco_master_ready(dlm)) { 2569 mlog(0, "%s: reco master taking awhile\n", 2570 dlm->name); 2571 goto again; 2572 } 2573 /* another node has informed this one that it is reco master */ 2574 mlog(0, "%s: reco master %u is ready to recover %u\n", 2575 dlm->name, dlm->reco.new_master, dlm->reco.dead_node); 2576 status = -EEXIST; 2577 } else if (ret == DLM_RECOVERING) { 2578 mlog(0, "dlm=%s dlmlock says master node died (this=%u)\n", 2579 dlm->name, dlm->node_num); 2580 goto again; 2581 } else { 2582 struct dlm_lock_resource *res; 2583 2584 /* dlmlock returned something other than NOTQUEUED or NORMAL */ 2585 mlog(ML_ERROR, "%s: got %s from dlmlock($RECOVERY), " 2586 "lksb.status=%s\n", dlm->name, dlm_errname(ret), 2587 dlm_errname(lksb.status)); 2588 res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME, 2589 DLM_RECOVERY_LOCK_NAME_LEN); 2590 if (res) { 2591 dlm_print_one_lock_resource(res); 2592 dlm_lockres_put(res); 2593 } else { 2594 mlog(ML_ERROR, "recovery lock not found\n"); 2595 } 2596 BUG(); 2597 } 2598 2599 return status; 2600 } 2601 2602 static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node) 2603 { 2604 struct dlm_begin_reco br; 2605 int ret = 0; 2606 struct dlm_node_iter iter; 2607 int nodenum; 2608 int status; 2609 2610 mlog_entry("%u\n", dead_node); 2611 2612 mlog(0, "%s: dead node is %u\n", dlm->name, dead_node); 2613 2614 spin_lock(&dlm->spinlock); 2615 dlm_node_iter_init(dlm->domain_map, &iter); 2616 spin_unlock(&dlm->spinlock); 2617 2618 clear_bit(dead_node, iter.node_map); 2619 2620 memset(&br, 0, sizeof(br)); 2621 br.node_idx = dlm->node_num; 2622 br.dead_node = dead_node; 2623 2624 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) { 2625 ret = 0; 2626 if (nodenum == dead_node) { 2627 mlog(0, "not sending begin reco to dead node " 2628 "%u\n", dead_node); 2629 continue; 2630 } 2631 if (nodenum == dlm->node_num) { 2632 mlog(0, "not sending begin reco to self\n"); 2633 continue; 2634 } 2635 retry: 2636 ret = -EINVAL; 2637 mlog(0, "attempting to send begin reco msg to %d\n", 2638 nodenum); 2639 ret = o2net_send_message(DLM_BEGIN_RECO_MSG, dlm->key, 2640 &br, sizeof(br), nodenum, &status); 2641 /* negative status is handled ok by caller here */ 2642 if (ret >= 0) 2643 ret = status; 2644 if (dlm_is_host_down(ret)) { 2645 /* node is down. not involved in recovery 2646 * so just keep going */ 2647 mlog(ML_NOTICE, "%s: node %u was down when sending " 2648 "begin reco msg (%d)\n", dlm->name, nodenum, ret); 2649 ret = 0; 2650 } 2651 2652 /* 2653 * Prior to commit aad1b15310b9bcd59fa81ab8f2b1513b59553ea8, 2654 * dlm_begin_reco_handler() returned EAGAIN and not -EAGAIN. 2655 * We are handling both for compatibility reasons. 2656 */ 2657 if (ret == -EAGAIN || ret == EAGAIN) { 2658 mlog(0, "%s: trying to start recovery of node " 2659 "%u, but node %u is waiting for last recovery " 2660 "to complete, backoff for a bit\n", dlm->name, 2661 dead_node, nodenum); 2662 msleep(100); 2663 goto retry; 2664 } 2665 if (ret < 0) { 2666 struct dlm_lock_resource *res; 2667 2668 /* this is now a serious problem, possibly ENOMEM 2669 * in the network stack. must retry */ 2670 mlog_errno(ret); 2671 mlog(ML_ERROR, "begin reco of dlm %s to node %u " 2672 "returned %d\n", dlm->name, nodenum, ret); 2673 res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME, 2674 DLM_RECOVERY_LOCK_NAME_LEN); 2675 if (res) { 2676 dlm_print_one_lock_resource(res); 2677 dlm_lockres_put(res); 2678 } else { 2679 mlog(ML_ERROR, "recovery lock not found\n"); 2680 } 2681 /* sleep for a bit in hopes that we can avoid 2682 * another ENOMEM */ 2683 msleep(100); 2684 goto retry; 2685 } 2686 } 2687 2688 return ret; 2689 } 2690 2691 int dlm_begin_reco_handler(struct o2net_msg *msg, u32 len, void *data, 2692 void **ret_data) 2693 { 2694 struct dlm_ctxt *dlm = data; 2695 struct dlm_begin_reco *br = (struct dlm_begin_reco *)msg->buf; 2696 2697 /* ok to return 0, domain has gone away */ 2698 if (!dlm_grab(dlm)) 2699 return 0; 2700 2701 spin_lock(&dlm->spinlock); 2702 if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) { 2703 mlog(0, "%s: node %u wants to recover node %u (%u:%u) " 2704 "but this node is in finalize state, waiting on finalize2\n", 2705 dlm->name, br->node_idx, br->dead_node, 2706 dlm->reco.dead_node, dlm->reco.new_master); 2707 spin_unlock(&dlm->spinlock); 2708 return -EAGAIN; 2709 } 2710 spin_unlock(&dlm->spinlock); 2711 2712 mlog(0, "%s: node %u wants to recover node %u (%u:%u)\n", 2713 dlm->name, br->node_idx, br->dead_node, 2714 dlm->reco.dead_node, dlm->reco.new_master); 2715 2716 dlm_fire_domain_eviction_callbacks(dlm, br->dead_node); 2717 2718 spin_lock(&dlm->spinlock); 2719 if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) { 2720 if (test_bit(dlm->reco.new_master, dlm->recovery_map)) { 2721 mlog(0, "%s: new_master %u died, changing " 2722 "to %u\n", dlm->name, dlm->reco.new_master, 2723 br->node_idx); 2724 } else { 2725 mlog(0, "%s: new_master %u NOT DEAD, changing " 2726 "to %u\n", dlm->name, dlm->reco.new_master, 2727 br->node_idx); 2728 /* may not have seen the new master as dead yet */ 2729 } 2730 } 2731 if (dlm->reco.dead_node != O2NM_INVALID_NODE_NUM) { 2732 mlog(ML_NOTICE, "%s: dead_node previously set to %u, " 2733 "node %u changing it to %u\n", dlm->name, 2734 dlm->reco.dead_node, br->node_idx, br->dead_node); 2735 } 2736 dlm_set_reco_master(dlm, br->node_idx); 2737 dlm_set_reco_dead_node(dlm, br->dead_node); 2738 if (!test_bit(br->dead_node, dlm->recovery_map)) { 2739 mlog(0, "recovery master %u sees %u as dead, but this " 2740 "node has not yet. marking %u as dead\n", 2741 br->node_idx, br->dead_node, br->dead_node); 2742 if (!test_bit(br->dead_node, dlm->domain_map) || 2743 !test_bit(br->dead_node, dlm->live_nodes_map)) 2744 mlog(0, "%u not in domain/live_nodes map " 2745 "so setting it in reco map manually\n", 2746 br->dead_node); 2747 /* force the recovery cleanup in __dlm_hb_node_down 2748 * both of these will be cleared in a moment */ 2749 set_bit(br->dead_node, dlm->domain_map); 2750 set_bit(br->dead_node, dlm->live_nodes_map); 2751 __dlm_hb_node_down(dlm, br->dead_node); 2752 } 2753 spin_unlock(&dlm->spinlock); 2754 2755 dlm_kick_recovery_thread(dlm); 2756 2757 mlog(0, "%s: recovery started by node %u, for %u (%u:%u)\n", 2758 dlm->name, br->node_idx, br->dead_node, 2759 dlm->reco.dead_node, dlm->reco.new_master); 2760 2761 dlm_put(dlm); 2762 return 0; 2763 } 2764 2765 #define DLM_FINALIZE_STAGE2 0x01 2766 static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm) 2767 { 2768 int ret = 0; 2769 struct dlm_finalize_reco fr; 2770 struct dlm_node_iter iter; 2771 int nodenum; 2772 int status; 2773 int stage = 1; 2774 2775 mlog(0, "finishing recovery for node %s:%u, " 2776 "stage %d\n", dlm->name, dlm->reco.dead_node, stage); 2777 2778 spin_lock(&dlm->spinlock); 2779 dlm_node_iter_init(dlm->domain_map, &iter); 2780 spin_unlock(&dlm->spinlock); 2781 2782 stage2: 2783 memset(&fr, 0, sizeof(fr)); 2784 fr.node_idx = dlm->node_num; 2785 fr.dead_node = dlm->reco.dead_node; 2786 if (stage == 2) 2787 fr.flags |= DLM_FINALIZE_STAGE2; 2788 2789 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) { 2790 if (nodenum == dlm->node_num) 2791 continue; 2792 ret = o2net_send_message(DLM_FINALIZE_RECO_MSG, dlm->key, 2793 &fr, sizeof(fr), nodenum, &status); 2794 if (ret >= 0) 2795 ret = status; 2796 if (ret < 0) { 2797 mlog(ML_ERROR, "Error %d when sending message %u (key " 2798 "0x%x) to node %u\n", ret, DLM_FINALIZE_RECO_MSG, 2799 dlm->key, nodenum); 2800 if (dlm_is_host_down(ret)) { 2801 /* this has no effect on this recovery 2802 * session, so set the status to zero to 2803 * finish out the last recovery */ 2804 mlog(ML_ERROR, "node %u went down after this " 2805 "node finished recovery.\n", nodenum); 2806 ret = 0; 2807 continue; 2808 } 2809 break; 2810 } 2811 } 2812 if (stage == 1) { 2813 /* reset the node_iter back to the top and send finalize2 */ 2814 iter.curnode = -1; 2815 stage = 2; 2816 goto stage2; 2817 } 2818 2819 return ret; 2820 } 2821 2822 int dlm_finalize_reco_handler(struct o2net_msg *msg, u32 len, void *data, 2823 void **ret_data) 2824 { 2825 struct dlm_ctxt *dlm = data; 2826 struct dlm_finalize_reco *fr = (struct dlm_finalize_reco *)msg->buf; 2827 int stage = 1; 2828 2829 /* ok to return 0, domain has gone away */ 2830 if (!dlm_grab(dlm)) 2831 return 0; 2832 2833 if (fr->flags & DLM_FINALIZE_STAGE2) 2834 stage = 2; 2835 2836 mlog(0, "%s: node %u finalizing recovery stage%d of " 2837 "node %u (%u:%u)\n", dlm->name, fr->node_idx, stage, 2838 fr->dead_node, dlm->reco.dead_node, dlm->reco.new_master); 2839 2840 spin_lock(&dlm->spinlock); 2841 2842 if (dlm->reco.new_master != fr->node_idx) { 2843 mlog(ML_ERROR, "node %u sent recovery finalize msg, but node " 2844 "%u is supposed to be the new master, dead=%u\n", 2845 fr->node_idx, dlm->reco.new_master, fr->dead_node); 2846 BUG(); 2847 } 2848 if (dlm->reco.dead_node != fr->dead_node) { 2849 mlog(ML_ERROR, "node %u sent recovery finalize msg for dead " 2850 "node %u, but node %u is supposed to be dead\n", 2851 fr->node_idx, fr->dead_node, dlm->reco.dead_node); 2852 BUG(); 2853 } 2854 2855 switch (stage) { 2856 case 1: 2857 dlm_finish_local_lockres_recovery(dlm, fr->dead_node, fr->node_idx); 2858 if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) { 2859 mlog(ML_ERROR, "%s: received finalize1 from " 2860 "new master %u for dead node %u, but " 2861 "this node has already received it!\n", 2862 dlm->name, fr->node_idx, fr->dead_node); 2863 dlm_print_reco_node_status(dlm); 2864 BUG(); 2865 } 2866 dlm->reco.state |= DLM_RECO_STATE_FINALIZE; 2867 spin_unlock(&dlm->spinlock); 2868 break; 2869 case 2: 2870 if (!(dlm->reco.state & DLM_RECO_STATE_FINALIZE)) { 2871 mlog(ML_ERROR, "%s: received finalize2 from " 2872 "new master %u for dead node %u, but " 2873 "this node did not have finalize1!\n", 2874 dlm->name, fr->node_idx, fr->dead_node); 2875 dlm_print_reco_node_status(dlm); 2876 BUG(); 2877 } 2878 dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE; 2879 spin_unlock(&dlm->spinlock); 2880 dlm_reset_recovery(dlm); 2881 dlm_kick_recovery_thread(dlm); 2882 break; 2883 default: 2884 BUG(); 2885 } 2886 2887 mlog(0, "%s: recovery done, reco master was %u, dead now %u, master now %u\n", 2888 dlm->name, fr->node_idx, dlm->reco.dead_node, dlm->reco.new_master); 2889 2890 dlm_put(dlm); 2891 return 0; 2892 } 2893