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 static int dlm_do_recovery(struct dlm_ctxt *dlm); 59 60 static int dlm_pick_recovery_master(struct dlm_ctxt *dlm); 61 static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node); 62 static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node); 63 static int dlm_request_all_locks(struct dlm_ctxt *dlm, 64 u8 request_from, u8 dead_node); 65 static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm, u8 dead_node); 66 67 static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res); 68 static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres, 69 const char *lockname, int namelen, 70 int total_locks, u64 cookie, 71 u8 flags, u8 master); 72 static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm, 73 struct dlm_migratable_lockres *mres, 74 u8 send_to, 75 struct dlm_lock_resource *res, 76 int total_locks); 77 static int dlm_process_recovery_data(struct dlm_ctxt *dlm, 78 struct dlm_lock_resource *res, 79 struct dlm_migratable_lockres *mres); 80 static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm); 81 static int dlm_send_all_done_msg(struct dlm_ctxt *dlm, 82 u8 dead_node, u8 send_to); 83 static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node); 84 static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm, 85 struct list_head *list, u8 dead_node); 86 static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm, 87 u8 dead_node, u8 new_master); 88 static void dlm_reco_ast(void *astdata); 89 static void dlm_reco_bast(void *astdata, int blocked_type); 90 static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st); 91 static void dlm_request_all_locks_worker(struct dlm_work_item *item, 92 void *data); 93 static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data); 94 static int dlm_lockres_master_requery(struct dlm_ctxt *dlm, 95 struct dlm_lock_resource *res, 96 u8 *real_master); 97 98 static u64 dlm_get_next_mig_cookie(void); 99 100 static DEFINE_SPINLOCK(dlm_reco_state_lock); 101 static DEFINE_SPINLOCK(dlm_mig_cookie_lock); 102 static u64 dlm_mig_cookie = 1; 103 104 static u64 dlm_get_next_mig_cookie(void) 105 { 106 u64 c; 107 spin_lock(&dlm_mig_cookie_lock); 108 c = dlm_mig_cookie; 109 if (dlm_mig_cookie == (~0ULL)) 110 dlm_mig_cookie = 1; 111 else 112 dlm_mig_cookie++; 113 spin_unlock(&dlm_mig_cookie_lock); 114 return c; 115 } 116 117 static inline void dlm_set_reco_dead_node(struct dlm_ctxt *dlm, 118 u8 dead_node) 119 { 120 assert_spin_locked(&dlm->spinlock); 121 if (dlm->reco.dead_node != dead_node) 122 mlog(0, "%s: changing dead_node from %u to %u\n", 123 dlm->name, dlm->reco.dead_node, dead_node); 124 dlm->reco.dead_node = dead_node; 125 } 126 127 static inline void dlm_set_reco_master(struct dlm_ctxt *dlm, 128 u8 master) 129 { 130 assert_spin_locked(&dlm->spinlock); 131 mlog(0, "%s: changing new_master from %u to %u\n", 132 dlm->name, dlm->reco.new_master, master); 133 dlm->reco.new_master = master; 134 } 135 136 static inline void __dlm_reset_recovery(struct dlm_ctxt *dlm) 137 { 138 assert_spin_locked(&dlm->spinlock); 139 clear_bit(dlm->reco.dead_node, dlm->recovery_map); 140 dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM); 141 dlm_set_reco_master(dlm, O2NM_INVALID_NODE_NUM); 142 } 143 144 static inline void dlm_reset_recovery(struct dlm_ctxt *dlm) 145 { 146 spin_lock(&dlm->spinlock); 147 __dlm_reset_recovery(dlm); 148 spin_unlock(&dlm->spinlock); 149 } 150 151 /* Worker function used during recovery. */ 152 void dlm_dispatch_work(struct work_struct *work) 153 { 154 struct dlm_ctxt *dlm = 155 container_of(work, struct dlm_ctxt, dispatched_work); 156 LIST_HEAD(tmp_list); 157 struct dlm_work_item *item, *next; 158 dlm_workfunc_t *workfunc; 159 int tot=0; 160 161 spin_lock(&dlm->work_lock); 162 list_splice_init(&dlm->work_list, &tmp_list); 163 spin_unlock(&dlm->work_lock); 164 165 list_for_each_entry(item, &tmp_list, list) { 166 tot++; 167 } 168 mlog(0, "%s: work thread has %d work items\n", dlm->name, tot); 169 170 list_for_each_entry_safe(item, next, &tmp_list, list) { 171 workfunc = item->func; 172 list_del_init(&item->list); 173 174 /* already have ref on dlm to avoid having 175 * it disappear. just double-check. */ 176 BUG_ON(item->dlm != dlm); 177 178 /* this is allowed to sleep and 179 * call network stuff */ 180 workfunc(item, item->data); 181 182 dlm_put(dlm); 183 kfree(item); 184 } 185 } 186 187 /* 188 * RECOVERY THREAD 189 */ 190 191 void dlm_kick_recovery_thread(struct dlm_ctxt *dlm) 192 { 193 /* wake the recovery thread 194 * this will wake the reco thread in one of three places 195 * 1) sleeping with no recovery happening 196 * 2) sleeping with recovery mastered elsewhere 197 * 3) recovery mastered here, waiting on reco data */ 198 199 wake_up(&dlm->dlm_reco_thread_wq); 200 } 201 202 /* Launch the recovery thread */ 203 int dlm_launch_recovery_thread(struct dlm_ctxt *dlm) 204 { 205 mlog(0, "starting dlm recovery thread...\n"); 206 207 dlm->dlm_reco_thread_task = kthread_run(dlm_recovery_thread, dlm, 208 "dlm_reco_thread"); 209 if (IS_ERR(dlm->dlm_reco_thread_task)) { 210 mlog_errno(PTR_ERR(dlm->dlm_reco_thread_task)); 211 dlm->dlm_reco_thread_task = NULL; 212 return -EINVAL; 213 } 214 215 return 0; 216 } 217 218 void dlm_complete_recovery_thread(struct dlm_ctxt *dlm) 219 { 220 if (dlm->dlm_reco_thread_task) { 221 mlog(0, "waiting for dlm recovery thread to exit\n"); 222 kthread_stop(dlm->dlm_reco_thread_task); 223 dlm->dlm_reco_thread_task = NULL; 224 } 225 } 226 227 228 229 /* 230 * this is lame, but here's how recovery works... 231 * 1) all recovery threads cluster wide will work on recovering 232 * ONE node at a time 233 * 2) negotiate who will take over all the locks for the dead node. 234 * thats right... ALL the locks. 235 * 3) once a new master is chosen, everyone scans all locks 236 * and moves aside those mastered by the dead guy 237 * 4) each of these locks should be locked until recovery is done 238 * 5) the new master collects up all of secondary lock queue info 239 * one lock at a time, forcing each node to communicate back 240 * before continuing 241 * 6) each secondary lock queue responds with the full known lock info 242 * 7) once the new master has run all its locks, it sends a ALLDONE! 243 * message to everyone 244 * 8) upon receiving this message, the secondary queue node unlocks 245 * and responds to the ALLDONE 246 * 9) once the new master gets responses from everyone, he unlocks 247 * everything and recovery for this dead node is done 248 *10) go back to 2) while there are still dead nodes 249 * 250 */ 251 252 static void dlm_print_reco_node_status(struct dlm_ctxt *dlm) 253 { 254 struct dlm_reco_node_data *ndata; 255 struct dlm_lock_resource *res; 256 257 mlog(ML_NOTICE, "%s(%d): recovery info, state=%s, dead=%u, master=%u\n", 258 dlm->name, task_pid_nr(dlm->dlm_reco_thread_task), 259 dlm->reco.state & DLM_RECO_STATE_ACTIVE ? "ACTIVE" : "inactive", 260 dlm->reco.dead_node, dlm->reco.new_master); 261 262 list_for_each_entry(ndata, &dlm->reco.node_data, list) { 263 char *st = "unknown"; 264 switch (ndata->state) { 265 case DLM_RECO_NODE_DATA_INIT: 266 st = "init"; 267 break; 268 case DLM_RECO_NODE_DATA_REQUESTING: 269 st = "requesting"; 270 break; 271 case DLM_RECO_NODE_DATA_DEAD: 272 st = "dead"; 273 break; 274 case DLM_RECO_NODE_DATA_RECEIVING: 275 st = "receiving"; 276 break; 277 case DLM_RECO_NODE_DATA_REQUESTED: 278 st = "requested"; 279 break; 280 case DLM_RECO_NODE_DATA_DONE: 281 st = "done"; 282 break; 283 case DLM_RECO_NODE_DATA_FINALIZE_SENT: 284 st = "finalize-sent"; 285 break; 286 default: 287 st = "bad"; 288 break; 289 } 290 mlog(ML_NOTICE, "%s: reco state, node %u, state=%s\n", 291 dlm->name, ndata->node_num, st); 292 } 293 list_for_each_entry(res, &dlm->reco.resources, recovering) { 294 mlog(ML_NOTICE, "%s: lockres %.*s on recovering list\n", 295 dlm->name, res->lockname.len, res->lockname.name); 296 } 297 } 298 299 #define DLM_RECO_THREAD_TIMEOUT_MS (5 * 1000) 300 301 static int dlm_recovery_thread(void *data) 302 { 303 int status; 304 struct dlm_ctxt *dlm = data; 305 unsigned long timeout = msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS); 306 307 mlog(0, "dlm thread running for %s...\n", dlm->name); 308 309 while (!kthread_should_stop()) { 310 if (dlm_domain_fully_joined(dlm)) { 311 status = dlm_do_recovery(dlm); 312 if (status == -EAGAIN) { 313 /* do not sleep, recheck immediately. */ 314 continue; 315 } 316 if (status < 0) 317 mlog_errno(status); 318 } 319 320 wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq, 321 kthread_should_stop(), 322 timeout); 323 } 324 325 mlog(0, "quitting DLM recovery thread\n"); 326 return 0; 327 } 328 329 /* returns true when the recovery master has contacted us */ 330 static int dlm_reco_master_ready(struct dlm_ctxt *dlm) 331 { 332 int ready; 333 spin_lock(&dlm->spinlock); 334 ready = (dlm->reco.new_master != O2NM_INVALID_NODE_NUM); 335 spin_unlock(&dlm->spinlock); 336 return ready; 337 } 338 339 /* returns true if node is no longer in the domain 340 * could be dead or just not joined */ 341 int dlm_is_node_dead(struct dlm_ctxt *dlm, u8 node) 342 { 343 int dead; 344 spin_lock(&dlm->spinlock); 345 dead = !test_bit(node, dlm->domain_map); 346 spin_unlock(&dlm->spinlock); 347 return dead; 348 } 349 350 /* returns true if node is no longer in the domain 351 * could be dead or just not joined */ 352 static int dlm_is_node_recovered(struct dlm_ctxt *dlm, u8 node) 353 { 354 int recovered; 355 spin_lock(&dlm->spinlock); 356 recovered = !test_bit(node, dlm->recovery_map); 357 spin_unlock(&dlm->spinlock); 358 return recovered; 359 } 360 361 362 void dlm_wait_for_node_death(struct dlm_ctxt *dlm, u8 node, int timeout) 363 { 364 if (dlm_is_node_dead(dlm, node)) 365 return; 366 367 printk(KERN_NOTICE "o2dlm: Waiting on the death of node %u in " 368 "domain %s\n", node, dlm->name); 369 370 if (timeout) 371 wait_event_timeout(dlm->dlm_reco_thread_wq, 372 dlm_is_node_dead(dlm, node), 373 msecs_to_jiffies(timeout)); 374 else 375 wait_event(dlm->dlm_reco_thread_wq, 376 dlm_is_node_dead(dlm, node)); 377 } 378 379 void dlm_wait_for_node_recovery(struct dlm_ctxt *dlm, u8 node, int timeout) 380 { 381 if (dlm_is_node_recovered(dlm, node)) 382 return; 383 384 printk(KERN_NOTICE "o2dlm: Waiting on the recovery of node %u in " 385 "domain %s\n", node, dlm->name); 386 387 if (timeout) 388 wait_event_timeout(dlm->dlm_reco_thread_wq, 389 dlm_is_node_recovered(dlm, node), 390 msecs_to_jiffies(timeout)); 391 else 392 wait_event(dlm->dlm_reco_thread_wq, 393 dlm_is_node_recovered(dlm, node)); 394 } 395 396 /* callers of the top-level api calls (dlmlock/dlmunlock) should 397 * block on the dlm->reco.event when recovery is in progress. 398 * the dlm recovery thread will set this state when it begins 399 * recovering a dead node (as the new master or not) and clear 400 * the state and wake as soon as all affected lock resources have 401 * been marked with the RECOVERY flag */ 402 static int dlm_in_recovery(struct dlm_ctxt *dlm) 403 { 404 int in_recovery; 405 spin_lock(&dlm->spinlock); 406 in_recovery = !!(dlm->reco.state & DLM_RECO_STATE_ACTIVE); 407 spin_unlock(&dlm->spinlock); 408 return in_recovery; 409 } 410 411 412 void dlm_wait_for_recovery(struct dlm_ctxt *dlm) 413 { 414 if (dlm_in_recovery(dlm)) { 415 mlog(0, "%s: reco thread %d in recovery: " 416 "state=%d, master=%u, dead=%u\n", 417 dlm->name, task_pid_nr(dlm->dlm_reco_thread_task), 418 dlm->reco.state, dlm->reco.new_master, 419 dlm->reco.dead_node); 420 } 421 wait_event(dlm->reco.event, !dlm_in_recovery(dlm)); 422 } 423 424 static void dlm_begin_recovery(struct dlm_ctxt *dlm) 425 { 426 spin_lock(&dlm->spinlock); 427 BUG_ON(dlm->reco.state & DLM_RECO_STATE_ACTIVE); 428 printk(KERN_NOTICE "o2dlm: Begin recovery on domain %s for node %u\n", 429 dlm->name, dlm->reco.dead_node); 430 dlm->reco.state |= DLM_RECO_STATE_ACTIVE; 431 spin_unlock(&dlm->spinlock); 432 } 433 434 static void dlm_end_recovery(struct dlm_ctxt *dlm) 435 { 436 spin_lock(&dlm->spinlock); 437 BUG_ON(!(dlm->reco.state & DLM_RECO_STATE_ACTIVE)); 438 dlm->reco.state &= ~DLM_RECO_STATE_ACTIVE; 439 spin_unlock(&dlm->spinlock); 440 printk(KERN_NOTICE "o2dlm: End recovery on domain %s\n", dlm->name); 441 wake_up(&dlm->reco.event); 442 } 443 444 static void dlm_print_recovery_master(struct dlm_ctxt *dlm) 445 { 446 printk(KERN_NOTICE "o2dlm: Node %u (%s) is the Recovery Master for the " 447 "dead node %u in domain %s\n", dlm->reco.new_master, 448 (dlm->node_num == dlm->reco.new_master ? "me" : "he"), 449 dlm->reco.dead_node, dlm->name); 450 } 451 452 static int dlm_do_recovery(struct dlm_ctxt *dlm) 453 { 454 int status = 0; 455 int ret; 456 457 spin_lock(&dlm->spinlock); 458 459 /* check to see if the new master has died */ 460 if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM && 461 test_bit(dlm->reco.new_master, dlm->recovery_map)) { 462 mlog(0, "new master %u died while recovering %u!\n", 463 dlm->reco.new_master, dlm->reco.dead_node); 464 /* unset the new_master, leave dead_node */ 465 dlm_set_reco_master(dlm, O2NM_INVALID_NODE_NUM); 466 } 467 468 /* select a target to recover */ 469 if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) { 470 int bit; 471 472 bit = find_next_bit (dlm->recovery_map, O2NM_MAX_NODES, 0); 473 if (bit >= O2NM_MAX_NODES || bit < 0) 474 dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM); 475 else 476 dlm_set_reco_dead_node(dlm, bit); 477 } else if (!test_bit(dlm->reco.dead_node, dlm->recovery_map)) { 478 /* BUG? */ 479 mlog(ML_ERROR, "dead_node %u no longer in recovery map!\n", 480 dlm->reco.dead_node); 481 dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM); 482 } 483 484 if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) { 485 // mlog(0, "nothing to recover! sleeping now!\n"); 486 spin_unlock(&dlm->spinlock); 487 /* return to main thread loop and sleep. */ 488 return 0; 489 } 490 mlog(0, "%s(%d):recovery thread found node %u in the recovery map!\n", 491 dlm->name, task_pid_nr(dlm->dlm_reco_thread_task), 492 dlm->reco.dead_node); 493 spin_unlock(&dlm->spinlock); 494 495 /* take write barrier */ 496 /* (stops the list reshuffling thread, proxy ast handling) */ 497 dlm_begin_recovery(dlm); 498 499 if (dlm->reco.new_master == dlm->node_num) 500 goto master_here; 501 502 if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) { 503 /* choose a new master, returns 0 if this node 504 * is the master, -EEXIST if it's another node. 505 * this does not return until a new master is chosen 506 * or recovery completes entirely. */ 507 ret = dlm_pick_recovery_master(dlm); 508 if (!ret) { 509 /* already notified everyone. go. */ 510 goto master_here; 511 } 512 mlog(0, "another node will master this recovery session.\n"); 513 } 514 515 dlm_print_recovery_master(dlm); 516 517 /* it is safe to start everything back up here 518 * because all of the dead node's lock resources 519 * have been marked as in-recovery */ 520 dlm_end_recovery(dlm); 521 522 /* sleep out in main dlm_recovery_thread loop. */ 523 return 0; 524 525 master_here: 526 dlm_print_recovery_master(dlm); 527 528 status = dlm_remaster_locks(dlm, dlm->reco.dead_node); 529 if (status < 0) { 530 /* we should never hit this anymore */ 531 mlog(ML_ERROR, "%s: Error %d remastering locks for node %u, " 532 "retrying.\n", dlm->name, status, dlm->reco.dead_node); 533 /* yield a bit to allow any final network messages 534 * to get handled on remaining nodes */ 535 msleep(100); 536 } else { 537 /* success! see if any other nodes need recovery */ 538 mlog(0, "DONE mastering recovery of %s:%u here(this=%u)!\n", 539 dlm->name, dlm->reco.dead_node, dlm->node_num); 540 spin_lock(&dlm->spinlock); 541 __dlm_reset_recovery(dlm); 542 dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE; 543 spin_unlock(&dlm->spinlock); 544 } 545 dlm_end_recovery(dlm); 546 547 /* continue and look for another dead node */ 548 return -EAGAIN; 549 } 550 551 static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node) 552 { 553 int status = 0; 554 struct dlm_reco_node_data *ndata; 555 int all_nodes_done; 556 int destroy = 0; 557 int pass = 0; 558 559 do { 560 /* we have become recovery master. there is no escaping 561 * this, so just keep trying until we get it. */ 562 status = dlm_init_recovery_area(dlm, dead_node); 563 if (status < 0) { 564 mlog(ML_ERROR, "%s: failed to alloc recovery area, " 565 "retrying\n", dlm->name); 566 msleep(1000); 567 } 568 } while (status != 0); 569 570 /* safe to access the node data list without a lock, since this 571 * process is the only one to change the list */ 572 list_for_each_entry(ndata, &dlm->reco.node_data, list) { 573 BUG_ON(ndata->state != DLM_RECO_NODE_DATA_INIT); 574 ndata->state = DLM_RECO_NODE_DATA_REQUESTING; 575 576 mlog(0, "%s: Requesting lock info from node %u\n", dlm->name, 577 ndata->node_num); 578 579 if (ndata->node_num == dlm->node_num) { 580 ndata->state = DLM_RECO_NODE_DATA_DONE; 581 continue; 582 } 583 584 do { 585 status = dlm_request_all_locks(dlm, ndata->node_num, 586 dead_node); 587 if (status < 0) { 588 mlog_errno(status); 589 if (dlm_is_host_down(status)) { 590 /* node died, ignore it for recovery */ 591 status = 0; 592 ndata->state = DLM_RECO_NODE_DATA_DEAD; 593 /* wait for the domain map to catch up 594 * with the network state. */ 595 wait_event_timeout(dlm->dlm_reco_thread_wq, 596 dlm_is_node_dead(dlm, 597 ndata->node_num), 598 msecs_to_jiffies(1000)); 599 mlog(0, "waited 1 sec for %u, " 600 "dead? %s\n", ndata->node_num, 601 dlm_is_node_dead(dlm, ndata->node_num) ? 602 "yes" : "no"); 603 } else { 604 /* -ENOMEM on the other node */ 605 mlog(0, "%s: node %u returned " 606 "%d during recovery, retrying " 607 "after a short wait\n", 608 dlm->name, ndata->node_num, 609 status); 610 msleep(100); 611 } 612 } 613 } while (status != 0); 614 615 spin_lock(&dlm_reco_state_lock); 616 switch (ndata->state) { 617 case DLM_RECO_NODE_DATA_INIT: 618 case DLM_RECO_NODE_DATA_FINALIZE_SENT: 619 case DLM_RECO_NODE_DATA_REQUESTED: 620 BUG(); 621 break; 622 case DLM_RECO_NODE_DATA_DEAD: 623 mlog(0, "node %u died after requesting " 624 "recovery info for node %u\n", 625 ndata->node_num, dead_node); 626 /* fine. don't need this node's info. 627 * continue without it. */ 628 break; 629 case DLM_RECO_NODE_DATA_REQUESTING: 630 ndata->state = DLM_RECO_NODE_DATA_REQUESTED; 631 mlog(0, "now receiving recovery data from " 632 "node %u for dead node %u\n", 633 ndata->node_num, dead_node); 634 break; 635 case DLM_RECO_NODE_DATA_RECEIVING: 636 mlog(0, "already receiving recovery data from " 637 "node %u for dead node %u\n", 638 ndata->node_num, dead_node); 639 break; 640 case DLM_RECO_NODE_DATA_DONE: 641 mlog(0, "already DONE receiving recovery data " 642 "from node %u for dead node %u\n", 643 ndata->node_num, dead_node); 644 break; 645 } 646 spin_unlock(&dlm_reco_state_lock); 647 } 648 649 mlog(0, "%s: Done requesting all lock info\n", dlm->name); 650 651 /* nodes should be sending reco data now 652 * just need to wait */ 653 654 while (1) { 655 /* check all the nodes now to see if we are 656 * done, or if anyone died */ 657 all_nodes_done = 1; 658 spin_lock(&dlm_reco_state_lock); 659 list_for_each_entry(ndata, &dlm->reco.node_data, list) { 660 mlog(0, "checking recovery state of node %u\n", 661 ndata->node_num); 662 switch (ndata->state) { 663 case DLM_RECO_NODE_DATA_INIT: 664 case DLM_RECO_NODE_DATA_REQUESTING: 665 mlog(ML_ERROR, "bad ndata state for " 666 "node %u: state=%d\n", 667 ndata->node_num, ndata->state); 668 BUG(); 669 break; 670 case DLM_RECO_NODE_DATA_DEAD: 671 mlog(0, "node %u died after " 672 "requesting recovery info for " 673 "node %u\n", ndata->node_num, 674 dead_node); 675 break; 676 case DLM_RECO_NODE_DATA_RECEIVING: 677 case DLM_RECO_NODE_DATA_REQUESTED: 678 mlog(0, "%s: node %u still in state %s\n", 679 dlm->name, ndata->node_num, 680 ndata->state==DLM_RECO_NODE_DATA_RECEIVING ? 681 "receiving" : "requested"); 682 all_nodes_done = 0; 683 break; 684 case DLM_RECO_NODE_DATA_DONE: 685 mlog(0, "%s: node %u state is done\n", 686 dlm->name, ndata->node_num); 687 break; 688 case DLM_RECO_NODE_DATA_FINALIZE_SENT: 689 mlog(0, "%s: node %u state is finalize\n", 690 dlm->name, ndata->node_num); 691 break; 692 } 693 } 694 spin_unlock(&dlm_reco_state_lock); 695 696 mlog(0, "pass #%d, all_nodes_done?: %s\n", ++pass, 697 all_nodes_done?"yes":"no"); 698 if (all_nodes_done) { 699 int ret; 700 701 /* Set this flag on recovery master to avoid 702 * a new recovery for another dead node start 703 * before the recovery is not done. That may 704 * cause recovery hung.*/ 705 spin_lock(&dlm->spinlock); 706 dlm->reco.state |= DLM_RECO_STATE_FINALIZE; 707 spin_unlock(&dlm->spinlock); 708 709 /* all nodes are now in DLM_RECO_NODE_DATA_DONE state 710 * just send a finalize message to everyone and 711 * clean up */ 712 mlog(0, "all nodes are done! send finalize\n"); 713 ret = dlm_send_finalize_reco_message(dlm); 714 if (ret < 0) 715 mlog_errno(ret); 716 717 spin_lock(&dlm->spinlock); 718 dlm_finish_local_lockres_recovery(dlm, dead_node, 719 dlm->node_num); 720 spin_unlock(&dlm->spinlock); 721 mlog(0, "should be done with recovery!\n"); 722 723 mlog(0, "finishing recovery of %s at %lu, " 724 "dead=%u, this=%u, new=%u\n", dlm->name, 725 jiffies, dlm->reco.dead_node, 726 dlm->node_num, dlm->reco.new_master); 727 destroy = 1; 728 status = 0; 729 /* rescan everything marked dirty along the way */ 730 dlm_kick_thread(dlm, NULL); 731 break; 732 } 733 /* wait to be signalled, with periodic timeout 734 * to check for node death */ 735 wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq, 736 kthread_should_stop(), 737 msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS)); 738 739 } 740 741 if (destroy) 742 dlm_destroy_recovery_area(dlm, dead_node); 743 744 return status; 745 } 746 747 static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node) 748 { 749 int num=0; 750 struct dlm_reco_node_data *ndata; 751 752 spin_lock(&dlm->spinlock); 753 memcpy(dlm->reco.node_map, dlm->domain_map, sizeof(dlm->domain_map)); 754 /* nodes can only be removed (by dying) after dropping 755 * this lock, and death will be trapped later, so this should do */ 756 spin_unlock(&dlm->spinlock); 757 758 while (1) { 759 num = find_next_bit (dlm->reco.node_map, O2NM_MAX_NODES, num); 760 if (num >= O2NM_MAX_NODES) { 761 break; 762 } 763 BUG_ON(num == dead_node); 764 765 ndata = kzalloc(sizeof(*ndata), GFP_NOFS); 766 if (!ndata) { 767 dlm_destroy_recovery_area(dlm, dead_node); 768 return -ENOMEM; 769 } 770 ndata->node_num = num; 771 ndata->state = DLM_RECO_NODE_DATA_INIT; 772 spin_lock(&dlm_reco_state_lock); 773 list_add_tail(&ndata->list, &dlm->reco.node_data); 774 spin_unlock(&dlm_reco_state_lock); 775 num++; 776 } 777 778 return 0; 779 } 780 781 static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm, u8 dead_node) 782 { 783 struct dlm_reco_node_data *ndata, *next; 784 LIST_HEAD(tmplist); 785 786 spin_lock(&dlm_reco_state_lock); 787 list_splice_init(&dlm->reco.node_data, &tmplist); 788 spin_unlock(&dlm_reco_state_lock); 789 790 list_for_each_entry_safe(ndata, next, &tmplist, list) { 791 list_del_init(&ndata->list); 792 kfree(ndata); 793 } 794 } 795 796 static int dlm_request_all_locks(struct dlm_ctxt *dlm, u8 request_from, 797 u8 dead_node) 798 { 799 struct dlm_lock_request lr; 800 int ret; 801 int status; 802 803 mlog(0, "\n"); 804 805 806 mlog(0, "dlm_request_all_locks: dead node is %u, sending request " 807 "to %u\n", dead_node, request_from); 808 809 memset(&lr, 0, sizeof(lr)); 810 lr.node_idx = dlm->node_num; 811 lr.dead_node = dead_node; 812 813 // send message 814 ret = o2net_send_message(DLM_LOCK_REQUEST_MSG, dlm->key, 815 &lr, sizeof(lr), request_from, &status); 816 817 /* negative status is handled by caller */ 818 if (ret < 0) 819 mlog(ML_ERROR, "%s: Error %d send LOCK_REQUEST to node %u " 820 "to recover dead node %u\n", dlm->name, ret, 821 request_from, dead_node); 822 else 823 ret = status; 824 // return from here, then 825 // sleep until all received or error 826 return ret; 827 828 } 829 830 int dlm_request_all_locks_handler(struct o2net_msg *msg, u32 len, void *data, 831 void **ret_data) 832 { 833 struct dlm_ctxt *dlm = data; 834 struct dlm_lock_request *lr = (struct dlm_lock_request *)msg->buf; 835 char *buf = NULL; 836 struct dlm_work_item *item = NULL; 837 838 if (!dlm_grab(dlm)) 839 return -EINVAL; 840 841 if (lr->dead_node != dlm->reco.dead_node) { 842 mlog(ML_ERROR, "%s: node %u sent dead_node=%u, but local " 843 "dead_node is %u\n", dlm->name, lr->node_idx, 844 lr->dead_node, dlm->reco.dead_node); 845 dlm_print_reco_node_status(dlm); 846 /* this is a hack */ 847 dlm_put(dlm); 848 return -ENOMEM; 849 } 850 BUG_ON(lr->dead_node != dlm->reco.dead_node); 851 852 item = kzalloc(sizeof(*item), GFP_NOFS); 853 if (!item) { 854 dlm_put(dlm); 855 return -ENOMEM; 856 } 857 858 /* this will get freed by dlm_request_all_locks_worker */ 859 buf = (char *) __get_free_page(GFP_NOFS); 860 if (!buf) { 861 kfree(item); 862 dlm_put(dlm); 863 return -ENOMEM; 864 } 865 866 /* queue up work for dlm_request_all_locks_worker */ 867 dlm_grab(dlm); /* get an extra ref for the work item */ 868 dlm_init_work_item(dlm, item, dlm_request_all_locks_worker, buf); 869 item->u.ral.reco_master = lr->node_idx; 870 item->u.ral.dead_node = lr->dead_node; 871 spin_lock(&dlm->work_lock); 872 list_add_tail(&item->list, &dlm->work_list); 873 spin_unlock(&dlm->work_lock); 874 queue_work(dlm->dlm_worker, &dlm->dispatched_work); 875 876 dlm_put(dlm); 877 return 0; 878 } 879 880 static void dlm_request_all_locks_worker(struct dlm_work_item *item, void *data) 881 { 882 struct dlm_migratable_lockres *mres; 883 struct dlm_lock_resource *res; 884 struct dlm_ctxt *dlm; 885 LIST_HEAD(resources); 886 int ret; 887 u8 dead_node, reco_master; 888 int skip_all_done = 0; 889 890 dlm = item->dlm; 891 dead_node = item->u.ral.dead_node; 892 reco_master = item->u.ral.reco_master; 893 mres = (struct dlm_migratable_lockres *)data; 894 895 mlog(0, "%s: recovery worker started, dead=%u, master=%u\n", 896 dlm->name, dead_node, reco_master); 897 898 if (dead_node != dlm->reco.dead_node || 899 reco_master != dlm->reco.new_master) { 900 /* worker could have been created before the recovery master 901 * died. if so, do not continue, but do not error. */ 902 if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) { 903 mlog(ML_NOTICE, "%s: will not send recovery state, " 904 "recovery master %u died, thread=(dead=%u,mas=%u)" 905 " current=(dead=%u,mas=%u)\n", dlm->name, 906 reco_master, dead_node, reco_master, 907 dlm->reco.dead_node, dlm->reco.new_master); 908 } else { 909 mlog(ML_NOTICE, "%s: reco state invalid: reco(dead=%u, " 910 "master=%u), request(dead=%u, master=%u)\n", 911 dlm->name, dlm->reco.dead_node, 912 dlm->reco.new_master, dead_node, reco_master); 913 } 914 goto leave; 915 } 916 917 /* lock resources should have already been moved to the 918 * dlm->reco.resources list. now move items from that list 919 * to a temp list if the dead owner matches. note that the 920 * whole cluster recovers only one node at a time, so we 921 * can safely move UNKNOWN lock resources for each recovery 922 * session. */ 923 dlm_move_reco_locks_to_list(dlm, &resources, dead_node); 924 925 /* now we can begin blasting lockreses without the dlm lock */ 926 927 /* any errors returned will be due to the new_master dying, 928 * the dlm_reco_thread should detect this */ 929 list_for_each_entry(res, &resources, recovering) { 930 ret = dlm_send_one_lockres(dlm, res, mres, reco_master, 931 DLM_MRES_RECOVERY); 932 if (ret < 0) { 933 mlog(ML_ERROR, "%s: node %u went down while sending " 934 "recovery state for dead node %u, ret=%d\n", dlm->name, 935 reco_master, dead_node, ret); 936 skip_all_done = 1; 937 break; 938 } 939 } 940 941 /* move the resources back to the list */ 942 spin_lock(&dlm->spinlock); 943 list_splice_init(&resources, &dlm->reco.resources); 944 spin_unlock(&dlm->spinlock); 945 946 if (!skip_all_done) { 947 ret = dlm_send_all_done_msg(dlm, dead_node, reco_master); 948 if (ret < 0) { 949 mlog(ML_ERROR, "%s: node %u went down while sending " 950 "recovery all-done for dead node %u, ret=%d\n", 951 dlm->name, reco_master, dead_node, ret); 952 } 953 } 954 leave: 955 free_page((unsigned long)data); 956 } 957 958 959 static int dlm_send_all_done_msg(struct dlm_ctxt *dlm, u8 dead_node, u8 send_to) 960 { 961 int ret, tmpret; 962 struct dlm_reco_data_done done_msg; 963 964 memset(&done_msg, 0, sizeof(done_msg)); 965 done_msg.node_idx = dlm->node_num; 966 done_msg.dead_node = dead_node; 967 mlog(0, "sending DATA DONE message to %u, " 968 "my node=%u, dead node=%u\n", send_to, done_msg.node_idx, 969 done_msg.dead_node); 970 971 ret = o2net_send_message(DLM_RECO_DATA_DONE_MSG, dlm->key, &done_msg, 972 sizeof(done_msg), send_to, &tmpret); 973 if (ret < 0) { 974 mlog(ML_ERROR, "%s: Error %d send RECO_DATA_DONE to node %u " 975 "to recover dead node %u\n", dlm->name, ret, send_to, 976 dead_node); 977 if (!dlm_is_host_down(ret)) { 978 BUG(); 979 } 980 } else 981 ret = tmpret; 982 return ret; 983 } 984 985 986 int dlm_reco_data_done_handler(struct o2net_msg *msg, u32 len, void *data, 987 void **ret_data) 988 { 989 struct dlm_ctxt *dlm = data; 990 struct dlm_reco_data_done *done = (struct dlm_reco_data_done *)msg->buf; 991 struct dlm_reco_node_data *ndata = NULL; 992 int ret = -EINVAL; 993 994 if (!dlm_grab(dlm)) 995 return -EINVAL; 996 997 mlog(0, "got DATA DONE: dead_node=%u, reco.dead_node=%u, " 998 "node_idx=%u, this node=%u\n", done->dead_node, 999 dlm->reco.dead_node, done->node_idx, dlm->node_num); 1000 1001 mlog_bug_on_msg((done->dead_node != dlm->reco.dead_node), 1002 "Got DATA DONE: dead_node=%u, reco.dead_node=%u, " 1003 "node_idx=%u, this node=%u\n", done->dead_node, 1004 dlm->reco.dead_node, done->node_idx, dlm->node_num); 1005 1006 spin_lock(&dlm_reco_state_lock); 1007 list_for_each_entry(ndata, &dlm->reco.node_data, list) { 1008 if (ndata->node_num != done->node_idx) 1009 continue; 1010 1011 switch (ndata->state) { 1012 /* should have moved beyond INIT but not to FINALIZE yet */ 1013 case DLM_RECO_NODE_DATA_INIT: 1014 case DLM_RECO_NODE_DATA_DEAD: 1015 case DLM_RECO_NODE_DATA_FINALIZE_SENT: 1016 mlog(ML_ERROR, "bad ndata state for node %u:" 1017 " state=%d\n", ndata->node_num, 1018 ndata->state); 1019 BUG(); 1020 break; 1021 /* these states are possible at this point, anywhere along 1022 * the line of recovery */ 1023 case DLM_RECO_NODE_DATA_DONE: 1024 case DLM_RECO_NODE_DATA_RECEIVING: 1025 case DLM_RECO_NODE_DATA_REQUESTED: 1026 case DLM_RECO_NODE_DATA_REQUESTING: 1027 mlog(0, "node %u is DONE sending " 1028 "recovery data!\n", 1029 ndata->node_num); 1030 1031 ndata->state = DLM_RECO_NODE_DATA_DONE; 1032 ret = 0; 1033 break; 1034 } 1035 } 1036 spin_unlock(&dlm_reco_state_lock); 1037 1038 /* wake the recovery thread, some node is done */ 1039 if (!ret) 1040 dlm_kick_recovery_thread(dlm); 1041 1042 if (ret < 0) 1043 mlog(ML_ERROR, "failed to find recovery node data for node " 1044 "%u\n", done->node_idx); 1045 dlm_put(dlm); 1046 1047 mlog(0, "leaving reco data done handler, ret=%d\n", ret); 1048 return ret; 1049 } 1050 1051 static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm, 1052 struct list_head *list, 1053 u8 dead_node) 1054 { 1055 struct dlm_lock_resource *res, *next; 1056 struct dlm_lock *lock; 1057 1058 spin_lock(&dlm->spinlock); 1059 list_for_each_entry_safe(res, next, &dlm->reco.resources, recovering) { 1060 /* always prune any $RECOVERY entries for dead nodes, 1061 * otherwise hangs can occur during later recovery */ 1062 if (dlm_is_recovery_lock(res->lockname.name, 1063 res->lockname.len)) { 1064 spin_lock(&res->spinlock); 1065 list_for_each_entry(lock, &res->granted, list) { 1066 if (lock->ml.node == dead_node) { 1067 mlog(0, "AHA! there was " 1068 "a $RECOVERY lock for dead " 1069 "node %u (%s)!\n", 1070 dead_node, dlm->name); 1071 list_del_init(&lock->list); 1072 dlm_lock_put(lock); 1073 break; 1074 } 1075 } 1076 spin_unlock(&res->spinlock); 1077 continue; 1078 } 1079 1080 if (res->owner == dead_node) { 1081 mlog(0, "found lockres owned by dead node while " 1082 "doing recovery for node %u. sending it.\n", 1083 dead_node); 1084 list_move_tail(&res->recovering, list); 1085 } else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) { 1086 mlog(0, "found UNKNOWN owner while doing recovery " 1087 "for node %u. sending it.\n", dead_node); 1088 list_move_tail(&res->recovering, list); 1089 } 1090 } 1091 spin_unlock(&dlm->spinlock); 1092 } 1093 1094 static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res) 1095 { 1096 int total_locks = 0; 1097 struct list_head *iter, *queue = &res->granted; 1098 int i; 1099 1100 for (i=0; i<3; i++) { 1101 list_for_each(iter, queue) 1102 total_locks++; 1103 queue++; 1104 } 1105 return total_locks; 1106 } 1107 1108 1109 static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm, 1110 struct dlm_migratable_lockres *mres, 1111 u8 send_to, 1112 struct dlm_lock_resource *res, 1113 int total_locks) 1114 { 1115 u64 mig_cookie = be64_to_cpu(mres->mig_cookie); 1116 int mres_total_locks = be32_to_cpu(mres->total_locks); 1117 int sz, ret = 0, status = 0; 1118 u8 orig_flags = mres->flags, 1119 orig_master = mres->master; 1120 1121 BUG_ON(mres->num_locks > DLM_MAX_MIGRATABLE_LOCKS); 1122 if (!mres->num_locks) 1123 return 0; 1124 1125 sz = sizeof(struct dlm_migratable_lockres) + 1126 (mres->num_locks * sizeof(struct dlm_migratable_lock)); 1127 1128 /* add an all-done flag if we reached the last lock */ 1129 orig_flags = mres->flags; 1130 BUG_ON(total_locks > mres_total_locks); 1131 if (total_locks == mres_total_locks) 1132 mres->flags |= DLM_MRES_ALL_DONE; 1133 1134 mlog(0, "%s:%.*s: sending mig lockres (%s) to %u\n", 1135 dlm->name, res->lockname.len, res->lockname.name, 1136 orig_flags & DLM_MRES_MIGRATION ? "migration" : "recovery", 1137 send_to); 1138 1139 /* send it */ 1140 ret = o2net_send_message(DLM_MIG_LOCKRES_MSG, dlm->key, mres, 1141 sz, send_to, &status); 1142 if (ret < 0) { 1143 /* XXX: negative status is not handled. 1144 * this will end up killing this node. */ 1145 mlog(ML_ERROR, "%s: res %.*s, Error %d send MIG_LOCKRES to " 1146 "node %u (%s)\n", dlm->name, mres->lockname_len, 1147 mres->lockname, ret, send_to, 1148 (orig_flags & DLM_MRES_MIGRATION ? 1149 "migration" : "recovery")); 1150 } else { 1151 /* might get an -ENOMEM back here */ 1152 ret = status; 1153 if (ret < 0) { 1154 mlog_errno(ret); 1155 1156 if (ret == -EFAULT) { 1157 mlog(ML_ERROR, "node %u told me to kill " 1158 "myself!\n", send_to); 1159 BUG(); 1160 } 1161 } 1162 } 1163 1164 /* zero and reinit the message buffer */ 1165 dlm_init_migratable_lockres(mres, res->lockname.name, 1166 res->lockname.len, mres_total_locks, 1167 mig_cookie, orig_flags, orig_master); 1168 return ret; 1169 } 1170 1171 static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres, 1172 const char *lockname, int namelen, 1173 int total_locks, u64 cookie, 1174 u8 flags, u8 master) 1175 { 1176 /* mres here is one full page */ 1177 clear_page(mres); 1178 mres->lockname_len = namelen; 1179 memcpy(mres->lockname, lockname, namelen); 1180 mres->num_locks = 0; 1181 mres->total_locks = cpu_to_be32(total_locks); 1182 mres->mig_cookie = cpu_to_be64(cookie); 1183 mres->flags = flags; 1184 mres->master = master; 1185 } 1186 1187 static void dlm_prepare_lvb_for_migration(struct dlm_lock *lock, 1188 struct dlm_migratable_lockres *mres, 1189 int queue) 1190 { 1191 if (!lock->lksb) 1192 return; 1193 1194 /* Ignore lvb in all locks in the blocked list */ 1195 if (queue == DLM_BLOCKED_LIST) 1196 return; 1197 1198 /* Only consider lvbs in locks with granted EX or PR lock levels */ 1199 if (lock->ml.type != LKM_EXMODE && lock->ml.type != LKM_PRMODE) 1200 return; 1201 1202 if (dlm_lvb_is_empty(mres->lvb)) { 1203 memcpy(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN); 1204 return; 1205 } 1206 1207 /* Ensure the lvb copied for migration matches in other valid locks */ 1208 if (!memcmp(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN)) 1209 return; 1210 1211 mlog(ML_ERROR, "Mismatched lvb in lock cookie=%u:%llu, name=%.*s, " 1212 "node=%u\n", 1213 dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)), 1214 dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)), 1215 lock->lockres->lockname.len, lock->lockres->lockname.name, 1216 lock->ml.node); 1217 dlm_print_one_lock_resource(lock->lockres); 1218 BUG(); 1219 } 1220 1221 /* returns 1 if this lock fills the network structure, 1222 * 0 otherwise */ 1223 static int dlm_add_lock_to_array(struct dlm_lock *lock, 1224 struct dlm_migratable_lockres *mres, int queue) 1225 { 1226 struct dlm_migratable_lock *ml; 1227 int lock_num = mres->num_locks; 1228 1229 ml = &(mres->ml[lock_num]); 1230 ml->cookie = lock->ml.cookie; 1231 ml->type = lock->ml.type; 1232 ml->convert_type = lock->ml.convert_type; 1233 ml->highest_blocked = lock->ml.highest_blocked; 1234 ml->list = queue; 1235 if (lock->lksb) { 1236 ml->flags = lock->lksb->flags; 1237 dlm_prepare_lvb_for_migration(lock, mres, queue); 1238 } 1239 ml->node = lock->ml.node; 1240 mres->num_locks++; 1241 /* we reached the max, send this network message */ 1242 if (mres->num_locks == DLM_MAX_MIGRATABLE_LOCKS) 1243 return 1; 1244 return 0; 1245 } 1246 1247 static void dlm_add_dummy_lock(struct dlm_ctxt *dlm, 1248 struct dlm_migratable_lockres *mres) 1249 { 1250 struct dlm_lock dummy; 1251 memset(&dummy, 0, sizeof(dummy)); 1252 dummy.ml.cookie = 0; 1253 dummy.ml.type = LKM_IVMODE; 1254 dummy.ml.convert_type = LKM_IVMODE; 1255 dummy.ml.highest_blocked = LKM_IVMODE; 1256 dummy.lksb = NULL; 1257 dummy.ml.node = dlm->node_num; 1258 dlm_add_lock_to_array(&dummy, mres, DLM_BLOCKED_LIST); 1259 } 1260 1261 static inline int dlm_is_dummy_lock(struct dlm_ctxt *dlm, 1262 struct dlm_migratable_lock *ml, 1263 u8 *nodenum) 1264 { 1265 if (unlikely(ml->cookie == 0 && 1266 ml->type == LKM_IVMODE && 1267 ml->convert_type == LKM_IVMODE && 1268 ml->highest_blocked == LKM_IVMODE && 1269 ml->list == DLM_BLOCKED_LIST)) { 1270 *nodenum = ml->node; 1271 return 1; 1272 } 1273 return 0; 1274 } 1275 1276 int dlm_send_one_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res, 1277 struct dlm_migratable_lockres *mres, 1278 u8 send_to, u8 flags) 1279 { 1280 struct list_head *queue; 1281 int total_locks, i; 1282 u64 mig_cookie = 0; 1283 struct dlm_lock *lock; 1284 int ret = 0; 1285 1286 BUG_ON(!(flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION))); 1287 1288 mlog(0, "sending to %u\n", send_to); 1289 1290 total_locks = dlm_num_locks_in_lockres(res); 1291 if (total_locks > DLM_MAX_MIGRATABLE_LOCKS) { 1292 /* rare, but possible */ 1293 mlog(0, "argh. lockres has %d locks. this will " 1294 "require more than one network packet to " 1295 "migrate\n", total_locks); 1296 mig_cookie = dlm_get_next_mig_cookie(); 1297 } 1298 1299 dlm_init_migratable_lockres(mres, res->lockname.name, 1300 res->lockname.len, total_locks, 1301 mig_cookie, flags, res->owner); 1302 1303 total_locks = 0; 1304 for (i=DLM_GRANTED_LIST; i<=DLM_BLOCKED_LIST; i++) { 1305 queue = dlm_list_idx_to_ptr(res, i); 1306 list_for_each_entry(lock, queue, list) { 1307 /* add another lock. */ 1308 total_locks++; 1309 if (!dlm_add_lock_to_array(lock, mres, i)) 1310 continue; 1311 1312 /* this filled the lock message, 1313 * we must send it immediately. */ 1314 ret = dlm_send_mig_lockres_msg(dlm, mres, send_to, 1315 res, total_locks); 1316 if (ret < 0) 1317 goto error; 1318 } 1319 } 1320 if (total_locks == 0) { 1321 /* send a dummy lock to indicate a mastery reference only */ 1322 mlog(0, "%s:%.*s: sending dummy lock to %u, %s\n", 1323 dlm->name, res->lockname.len, res->lockname.name, 1324 send_to, flags & DLM_MRES_RECOVERY ? "recovery" : 1325 "migration"); 1326 dlm_add_dummy_lock(dlm, mres); 1327 } 1328 /* flush any remaining locks */ 1329 ret = dlm_send_mig_lockres_msg(dlm, mres, send_to, res, total_locks); 1330 if (ret < 0) 1331 goto error; 1332 return ret; 1333 1334 error: 1335 mlog(ML_ERROR, "%s: dlm_send_mig_lockres_msg returned %d\n", 1336 dlm->name, ret); 1337 if (!dlm_is_host_down(ret)) 1338 BUG(); 1339 mlog(0, "%s: node %u went down while sending %s " 1340 "lockres %.*s\n", dlm->name, send_to, 1341 flags & DLM_MRES_RECOVERY ? "recovery" : "migration", 1342 res->lockname.len, res->lockname.name); 1343 return ret; 1344 } 1345 1346 1347 1348 /* 1349 * this message will contain no more than one page worth of 1350 * recovery data, and it will work on only one lockres. 1351 * there may be many locks in this page, and we may need to wait 1352 * for additional packets to complete all the locks (rare, but 1353 * possible). 1354 */ 1355 /* 1356 * NOTE: the allocation error cases here are scary 1357 * we really cannot afford to fail an alloc in recovery 1358 * do we spin? returning an error only delays the problem really 1359 */ 1360 1361 int dlm_mig_lockres_handler(struct o2net_msg *msg, u32 len, void *data, 1362 void **ret_data) 1363 { 1364 struct dlm_ctxt *dlm = data; 1365 struct dlm_migratable_lockres *mres = 1366 (struct dlm_migratable_lockres *)msg->buf; 1367 int ret = 0; 1368 u8 real_master; 1369 u8 extra_refs = 0; 1370 char *buf = NULL; 1371 struct dlm_work_item *item = NULL; 1372 struct dlm_lock_resource *res = NULL; 1373 1374 if (!dlm_grab(dlm)) 1375 return -EINVAL; 1376 1377 BUG_ON(!(mres->flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION))); 1378 1379 real_master = mres->master; 1380 if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) { 1381 /* cannot migrate a lockres with no master */ 1382 BUG_ON(!(mres->flags & DLM_MRES_RECOVERY)); 1383 } 1384 1385 mlog(0, "%s message received from node %u\n", 1386 (mres->flags & DLM_MRES_RECOVERY) ? 1387 "recovery" : "migration", mres->master); 1388 if (mres->flags & DLM_MRES_ALL_DONE) 1389 mlog(0, "all done flag. all lockres data received!\n"); 1390 1391 ret = -ENOMEM; 1392 buf = kmalloc(be16_to_cpu(msg->data_len), GFP_NOFS); 1393 item = kzalloc(sizeof(*item), GFP_NOFS); 1394 if (!buf || !item) 1395 goto leave; 1396 1397 /* lookup the lock to see if we have a secondary queue for this 1398 * already... just add the locks in and this will have its owner 1399 * and RECOVERY flag changed when it completes. */ 1400 res = dlm_lookup_lockres(dlm, mres->lockname, mres->lockname_len); 1401 if (res) { 1402 /* this will get a ref on res */ 1403 /* mark it as recovering/migrating and hash it */ 1404 spin_lock(&res->spinlock); 1405 if (mres->flags & DLM_MRES_RECOVERY) { 1406 res->state |= DLM_LOCK_RES_RECOVERING; 1407 } else { 1408 if (res->state & DLM_LOCK_RES_MIGRATING) { 1409 /* this is at least the second 1410 * lockres message */ 1411 mlog(0, "lock %.*s is already migrating\n", 1412 mres->lockname_len, 1413 mres->lockname); 1414 } else if (res->state & DLM_LOCK_RES_RECOVERING) { 1415 /* caller should BUG */ 1416 mlog(ML_ERROR, "node is attempting to migrate " 1417 "lock %.*s, but marked as recovering!\n", 1418 mres->lockname_len, mres->lockname); 1419 ret = -EFAULT; 1420 spin_unlock(&res->spinlock); 1421 dlm_lockres_put(res); 1422 goto leave; 1423 } 1424 res->state |= DLM_LOCK_RES_MIGRATING; 1425 } 1426 spin_unlock(&res->spinlock); 1427 } else { 1428 /* need to allocate, just like if it was 1429 * mastered here normally */ 1430 res = dlm_new_lockres(dlm, mres->lockname, mres->lockname_len); 1431 if (!res) 1432 goto leave; 1433 1434 /* to match the ref that we would have gotten if 1435 * dlm_lookup_lockres had succeeded */ 1436 dlm_lockres_get(res); 1437 1438 /* mark it as recovering/migrating and hash it */ 1439 if (mres->flags & DLM_MRES_RECOVERY) 1440 res->state |= DLM_LOCK_RES_RECOVERING; 1441 else 1442 res->state |= DLM_LOCK_RES_MIGRATING; 1443 1444 spin_lock(&dlm->spinlock); 1445 __dlm_insert_lockres(dlm, res); 1446 spin_unlock(&dlm->spinlock); 1447 1448 /* Add an extra ref for this lock-less lockres lest the 1449 * dlm_thread purges it before we get the chance to add 1450 * locks to it */ 1451 dlm_lockres_get(res); 1452 1453 /* There are three refs that need to be put. 1454 * 1. Taken above. 1455 * 2. kref_init in dlm_new_lockres()->dlm_init_lockres(). 1456 * 3. dlm_lookup_lockres() 1457 * The first one is handled at the end of this function. The 1458 * other two are handled in the worker thread after locks have 1459 * been attached. Yes, we don't wait for purge time to match 1460 * kref_init. The lockres will still have atleast one ref 1461 * added because it is in the hash __dlm_insert_lockres() */ 1462 extra_refs++; 1463 1464 /* now that the new lockres is inserted, 1465 * make it usable by other processes */ 1466 spin_lock(&res->spinlock); 1467 res->state &= ~DLM_LOCK_RES_IN_PROGRESS; 1468 spin_unlock(&res->spinlock); 1469 wake_up(&res->wq); 1470 } 1471 1472 /* at this point we have allocated everything we need, 1473 * and we have a hashed lockres with an extra ref and 1474 * the proper res->state flags. */ 1475 ret = 0; 1476 spin_lock(&res->spinlock); 1477 /* drop this either when master requery finds a different master 1478 * or when a lock is added by the recovery worker */ 1479 dlm_lockres_grab_inflight_ref(dlm, res); 1480 if (mres->master == DLM_LOCK_RES_OWNER_UNKNOWN) { 1481 /* migration cannot have an unknown master */ 1482 BUG_ON(!(mres->flags & DLM_MRES_RECOVERY)); 1483 mlog(0, "recovery has passed me a lockres with an " 1484 "unknown owner.. will need to requery: " 1485 "%.*s\n", mres->lockname_len, mres->lockname); 1486 } else { 1487 /* take a reference now to pin the lockres, drop it 1488 * when locks are added in the worker */ 1489 dlm_change_lockres_owner(dlm, res, dlm->node_num); 1490 } 1491 spin_unlock(&res->spinlock); 1492 1493 /* queue up work for dlm_mig_lockres_worker */ 1494 dlm_grab(dlm); /* get an extra ref for the work item */ 1495 memcpy(buf, msg->buf, be16_to_cpu(msg->data_len)); /* copy the whole message */ 1496 dlm_init_work_item(dlm, item, dlm_mig_lockres_worker, buf); 1497 item->u.ml.lockres = res; /* already have a ref */ 1498 item->u.ml.real_master = real_master; 1499 item->u.ml.extra_ref = extra_refs; 1500 spin_lock(&dlm->work_lock); 1501 list_add_tail(&item->list, &dlm->work_list); 1502 spin_unlock(&dlm->work_lock); 1503 queue_work(dlm->dlm_worker, &dlm->dispatched_work); 1504 1505 leave: 1506 /* One extra ref taken needs to be put here */ 1507 if (extra_refs) 1508 dlm_lockres_put(res); 1509 1510 dlm_put(dlm); 1511 if (ret < 0) { 1512 kfree(buf); 1513 kfree(item); 1514 mlog_errno(ret); 1515 } 1516 1517 return ret; 1518 } 1519 1520 1521 static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data) 1522 { 1523 struct dlm_ctxt *dlm; 1524 struct dlm_migratable_lockres *mres; 1525 int ret = 0; 1526 struct dlm_lock_resource *res; 1527 u8 real_master; 1528 u8 extra_ref; 1529 1530 dlm = item->dlm; 1531 mres = (struct dlm_migratable_lockres *)data; 1532 1533 res = item->u.ml.lockres; 1534 real_master = item->u.ml.real_master; 1535 extra_ref = item->u.ml.extra_ref; 1536 1537 if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) { 1538 /* this case is super-rare. only occurs if 1539 * node death happens during migration. */ 1540 again: 1541 ret = dlm_lockres_master_requery(dlm, res, &real_master); 1542 if (ret < 0) { 1543 mlog(0, "dlm_lockres_master_requery ret=%d\n", 1544 ret); 1545 goto again; 1546 } 1547 if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) { 1548 mlog(0, "lockres %.*s not claimed. " 1549 "this node will take it.\n", 1550 res->lockname.len, res->lockname.name); 1551 } else { 1552 spin_lock(&res->spinlock); 1553 dlm_lockres_drop_inflight_ref(dlm, res); 1554 spin_unlock(&res->spinlock); 1555 mlog(0, "master needs to respond to sender " 1556 "that node %u still owns %.*s\n", 1557 real_master, res->lockname.len, 1558 res->lockname.name); 1559 /* cannot touch this lockres */ 1560 goto leave; 1561 } 1562 } 1563 1564 ret = dlm_process_recovery_data(dlm, res, mres); 1565 if (ret < 0) 1566 mlog(0, "dlm_process_recovery_data returned %d\n", ret); 1567 else 1568 mlog(0, "dlm_process_recovery_data succeeded\n"); 1569 1570 if ((mres->flags & (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) == 1571 (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) { 1572 ret = dlm_finish_migration(dlm, res, mres->master); 1573 if (ret < 0) 1574 mlog_errno(ret); 1575 } 1576 1577 leave: 1578 /* See comment in dlm_mig_lockres_handler() */ 1579 if (res) { 1580 if (extra_ref) 1581 dlm_lockres_put(res); 1582 dlm_lockres_put(res); 1583 } 1584 kfree(data); 1585 } 1586 1587 1588 1589 static int dlm_lockres_master_requery(struct dlm_ctxt *dlm, 1590 struct dlm_lock_resource *res, 1591 u8 *real_master) 1592 { 1593 struct dlm_node_iter iter; 1594 int nodenum; 1595 int ret = 0; 1596 1597 *real_master = DLM_LOCK_RES_OWNER_UNKNOWN; 1598 1599 /* we only reach here if one of the two nodes in a 1600 * migration died while the migration was in progress. 1601 * at this point we need to requery the master. we 1602 * know that the new_master got as far as creating 1603 * an mle on at least one node, but we do not know 1604 * if any nodes had actually cleared the mle and set 1605 * the master to the new_master. the old master 1606 * is supposed to set the owner to UNKNOWN in the 1607 * event of a new_master death, so the only possible 1608 * responses that we can get from nodes here are 1609 * that the master is new_master, or that the master 1610 * is UNKNOWN. 1611 * if all nodes come back with UNKNOWN then we know 1612 * the lock needs remastering here. 1613 * if any node comes back with a valid master, check 1614 * to see if that master is the one that we are 1615 * recovering. if so, then the new_master died and 1616 * we need to remaster this lock. if not, then the 1617 * new_master survived and that node will respond to 1618 * other nodes about the owner. 1619 * if there is an owner, this node needs to dump this 1620 * lockres and alert the sender that this lockres 1621 * was rejected. */ 1622 spin_lock(&dlm->spinlock); 1623 dlm_node_iter_init(dlm->domain_map, &iter); 1624 spin_unlock(&dlm->spinlock); 1625 1626 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) { 1627 /* do not send to self */ 1628 if (nodenum == dlm->node_num) 1629 continue; 1630 ret = dlm_do_master_requery(dlm, res, nodenum, real_master); 1631 if (ret < 0) { 1632 mlog_errno(ret); 1633 if (!dlm_is_host_down(ret)) 1634 BUG(); 1635 /* host is down, so answer for that node would be 1636 * DLM_LOCK_RES_OWNER_UNKNOWN. continue. */ 1637 } 1638 if (*real_master != DLM_LOCK_RES_OWNER_UNKNOWN) { 1639 mlog(0, "lock master is %u\n", *real_master); 1640 break; 1641 } 1642 } 1643 return ret; 1644 } 1645 1646 1647 int dlm_do_master_requery(struct dlm_ctxt *dlm, struct dlm_lock_resource *res, 1648 u8 nodenum, u8 *real_master) 1649 { 1650 int ret = -EINVAL; 1651 struct dlm_master_requery req; 1652 int status = DLM_LOCK_RES_OWNER_UNKNOWN; 1653 1654 memset(&req, 0, sizeof(req)); 1655 req.node_idx = dlm->node_num; 1656 req.namelen = res->lockname.len; 1657 memcpy(req.name, res->lockname.name, res->lockname.len); 1658 1659 ret = o2net_send_message(DLM_MASTER_REQUERY_MSG, dlm->key, 1660 &req, sizeof(req), nodenum, &status); 1661 /* XXX: negative status not handled properly here. */ 1662 if (ret < 0) 1663 mlog(ML_ERROR, "Error %d when sending message %u (key " 1664 "0x%x) to node %u\n", ret, DLM_MASTER_REQUERY_MSG, 1665 dlm->key, nodenum); 1666 else { 1667 BUG_ON(status < 0); 1668 BUG_ON(status > DLM_LOCK_RES_OWNER_UNKNOWN); 1669 *real_master = (u8) (status & 0xff); 1670 mlog(0, "node %u responded to master requery with %u\n", 1671 nodenum, *real_master); 1672 ret = 0; 1673 } 1674 return ret; 1675 } 1676 1677 1678 /* this function cannot error, so unless the sending 1679 * or receiving of the message failed, the owner can 1680 * be trusted */ 1681 int dlm_master_requery_handler(struct o2net_msg *msg, u32 len, void *data, 1682 void **ret_data) 1683 { 1684 struct dlm_ctxt *dlm = data; 1685 struct dlm_master_requery *req = (struct dlm_master_requery *)msg->buf; 1686 struct dlm_lock_resource *res = NULL; 1687 unsigned int hash; 1688 int master = DLM_LOCK_RES_OWNER_UNKNOWN; 1689 u32 flags = DLM_ASSERT_MASTER_REQUERY; 1690 1691 if (!dlm_grab(dlm)) { 1692 /* since the domain has gone away on this 1693 * node, the proper response is UNKNOWN */ 1694 return master; 1695 } 1696 1697 hash = dlm_lockid_hash(req->name, req->namelen); 1698 1699 spin_lock(&dlm->spinlock); 1700 res = __dlm_lookup_lockres(dlm, req->name, req->namelen, hash); 1701 if (res) { 1702 spin_lock(&res->spinlock); 1703 master = res->owner; 1704 if (master == dlm->node_num) { 1705 int ret = dlm_dispatch_assert_master(dlm, res, 1706 0, 0, flags); 1707 if (ret < 0) { 1708 mlog_errno(-ENOMEM); 1709 /* retry!? */ 1710 BUG(); 1711 } else 1712 __dlm_lockres_grab_inflight_worker(dlm, res); 1713 spin_unlock(&res->spinlock); 1714 } else { 1715 /* put.. incase we are not the master */ 1716 spin_unlock(&res->spinlock); 1717 dlm_lockres_put(res); 1718 } 1719 } 1720 spin_unlock(&dlm->spinlock); 1721 1722 dlm_put(dlm); 1723 return master; 1724 } 1725 1726 static inline struct list_head * 1727 dlm_list_num_to_pointer(struct dlm_lock_resource *res, int list_num) 1728 { 1729 struct list_head *ret; 1730 BUG_ON(list_num < 0); 1731 BUG_ON(list_num > 2); 1732 ret = &(res->granted); 1733 ret += list_num; 1734 return ret; 1735 } 1736 /* TODO: do ast flush business 1737 * TODO: do MIGRATING and RECOVERING spinning 1738 */ 1739 1740 /* 1741 * NOTE about in-flight requests during migration: 1742 * 1743 * Before attempting the migrate, the master has marked the lockres as 1744 * MIGRATING and then flushed all of its pending ASTS. So any in-flight 1745 * requests either got queued before the MIGRATING flag got set, in which 1746 * case the lock data will reflect the change and a return message is on 1747 * the way, or the request failed to get in before MIGRATING got set. In 1748 * this case, the caller will be told to spin and wait for the MIGRATING 1749 * flag to be dropped, then recheck the master. 1750 * This holds true for the convert, cancel and unlock cases, and since lvb 1751 * updates are tied to these same messages, it applies to lvb updates as 1752 * well. For the lock case, there is no way a lock can be on the master 1753 * queue and not be on the secondary queue since the lock is always added 1754 * locally first. This means that the new target node will never be sent 1755 * a lock that he doesn't already have on the list. 1756 * In total, this means that the local lock is correct and should not be 1757 * updated to match the one sent by the master. Any messages sent back 1758 * from the master before the MIGRATING flag will bring the lock properly 1759 * up-to-date, and the change will be ordered properly for the waiter. 1760 * We will *not* attempt to modify the lock underneath the waiter. 1761 */ 1762 1763 static int dlm_process_recovery_data(struct dlm_ctxt *dlm, 1764 struct dlm_lock_resource *res, 1765 struct dlm_migratable_lockres *mres) 1766 { 1767 struct dlm_migratable_lock *ml; 1768 struct list_head *queue, *iter; 1769 struct list_head *tmpq = NULL; 1770 struct dlm_lock *newlock = NULL; 1771 struct dlm_lockstatus *lksb = NULL; 1772 int ret = 0; 1773 int i, j, bad; 1774 struct dlm_lock *lock; 1775 u8 from = O2NM_MAX_NODES; 1776 unsigned int added = 0; 1777 __be64 c; 1778 1779 mlog(0, "running %d locks for this lockres\n", mres->num_locks); 1780 for (i=0; i<mres->num_locks; i++) { 1781 ml = &(mres->ml[i]); 1782 1783 if (dlm_is_dummy_lock(dlm, ml, &from)) { 1784 /* placeholder, just need to set the refmap bit */ 1785 BUG_ON(mres->num_locks != 1); 1786 mlog(0, "%s:%.*s: dummy lock for %u\n", 1787 dlm->name, mres->lockname_len, mres->lockname, 1788 from); 1789 spin_lock(&res->spinlock); 1790 dlm_lockres_set_refmap_bit(dlm, res, from); 1791 spin_unlock(&res->spinlock); 1792 added++; 1793 break; 1794 } 1795 BUG_ON(ml->highest_blocked != LKM_IVMODE); 1796 newlock = NULL; 1797 lksb = NULL; 1798 1799 queue = dlm_list_num_to_pointer(res, ml->list); 1800 tmpq = NULL; 1801 1802 /* if the lock is for the local node it needs to 1803 * be moved to the proper location within the queue. 1804 * do not allocate a new lock structure. */ 1805 if (ml->node == dlm->node_num) { 1806 /* MIGRATION ONLY! */ 1807 BUG_ON(!(mres->flags & DLM_MRES_MIGRATION)); 1808 1809 lock = NULL; 1810 spin_lock(&res->spinlock); 1811 for (j = DLM_GRANTED_LIST; j <= DLM_BLOCKED_LIST; j++) { 1812 tmpq = dlm_list_idx_to_ptr(res, j); 1813 list_for_each(iter, tmpq) { 1814 lock = list_entry(iter, 1815 struct dlm_lock, list); 1816 if (lock->ml.cookie == ml->cookie) 1817 break; 1818 lock = NULL; 1819 } 1820 if (lock) 1821 break; 1822 } 1823 1824 /* lock is always created locally first, and 1825 * destroyed locally last. it must be on the list */ 1826 if (!lock) { 1827 c = ml->cookie; 1828 mlog(ML_ERROR, "Could not find local lock " 1829 "with cookie %u:%llu, node %u, " 1830 "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 (lock->ml.node != ml->node) { 1841 c = lock->ml.cookie; 1842 mlog(ML_ERROR, "Mismatched node# in lock " 1843 "cookie %u:%llu, name %.*s, node %u\n", 1844 dlm_get_lock_cookie_node(be64_to_cpu(c)), 1845 dlm_get_lock_cookie_seq(be64_to_cpu(c)), 1846 res->lockname.len, res->lockname.name, 1847 lock->ml.node); 1848 c = ml->cookie; 1849 mlog(ML_ERROR, "Migrate lock cookie %u:%llu, " 1850 "node %u, list %u, flags 0x%x, type %d, " 1851 "conv %d, highest blocked %d\n", 1852 dlm_get_lock_cookie_node(be64_to_cpu(c)), 1853 dlm_get_lock_cookie_seq(be64_to_cpu(c)), 1854 ml->node, ml->list, ml->flags, ml->type, 1855 ml->convert_type, ml->highest_blocked); 1856 __dlm_print_one_lock_resource(res); 1857 BUG(); 1858 } 1859 1860 if (tmpq != queue) { 1861 c = ml->cookie; 1862 mlog(0, "Lock cookie %u:%llu was on list %u " 1863 "instead of list %u for %.*s\n", 1864 dlm_get_lock_cookie_node(be64_to_cpu(c)), 1865 dlm_get_lock_cookie_seq(be64_to_cpu(c)), 1866 j, ml->list, res->lockname.len, 1867 res->lockname.name); 1868 __dlm_print_one_lock_resource(res); 1869 spin_unlock(&res->spinlock); 1870 continue; 1871 } 1872 1873 /* see NOTE above about why we do not update 1874 * to match the master here */ 1875 1876 /* move the lock to its proper place */ 1877 /* do not alter lock refcount. switching lists. */ 1878 list_move_tail(&lock->list, queue); 1879 spin_unlock(&res->spinlock); 1880 added++; 1881 1882 mlog(0, "just reordered a local lock!\n"); 1883 continue; 1884 } 1885 1886 /* lock is for another node. */ 1887 newlock = dlm_new_lock(ml->type, ml->node, 1888 be64_to_cpu(ml->cookie), NULL); 1889 if (!newlock) { 1890 ret = -ENOMEM; 1891 goto leave; 1892 } 1893 lksb = newlock->lksb; 1894 dlm_lock_attach_lockres(newlock, res); 1895 1896 if (ml->convert_type != LKM_IVMODE) { 1897 BUG_ON(queue != &res->converting); 1898 newlock->ml.convert_type = ml->convert_type; 1899 } 1900 lksb->flags |= (ml->flags & 1901 (DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB)); 1902 1903 if (ml->type == LKM_NLMODE) 1904 goto skip_lvb; 1905 1906 /* 1907 * If the lock is in the blocked list it can't have a valid lvb, 1908 * so skip it 1909 */ 1910 if (ml->list == DLM_BLOCKED_LIST) 1911 goto skip_lvb; 1912 1913 if (!dlm_lvb_is_empty(mres->lvb)) { 1914 if (lksb->flags & DLM_LKSB_PUT_LVB) { 1915 /* other node was trying to update 1916 * lvb when node died. recreate the 1917 * lksb with the updated lvb. */ 1918 memcpy(lksb->lvb, mres->lvb, DLM_LVB_LEN); 1919 /* the lock resource lvb update must happen 1920 * NOW, before the spinlock is dropped. 1921 * we no longer wait for the AST to update 1922 * the lvb. */ 1923 memcpy(res->lvb, mres->lvb, DLM_LVB_LEN); 1924 } else { 1925 /* otherwise, the node is sending its 1926 * most recent valid lvb info */ 1927 BUG_ON(ml->type != LKM_EXMODE && 1928 ml->type != LKM_PRMODE); 1929 if (!dlm_lvb_is_empty(res->lvb) && 1930 (ml->type == LKM_EXMODE || 1931 memcmp(res->lvb, mres->lvb, DLM_LVB_LEN))) { 1932 int i; 1933 mlog(ML_ERROR, "%s:%.*s: received bad " 1934 "lvb! type=%d\n", dlm->name, 1935 res->lockname.len, 1936 res->lockname.name, ml->type); 1937 printk("lockres lvb=["); 1938 for (i=0; i<DLM_LVB_LEN; i++) 1939 printk("%02x", res->lvb[i]); 1940 printk("]\nmigrated lvb=["); 1941 for (i=0; i<DLM_LVB_LEN; i++) 1942 printk("%02x", mres->lvb[i]); 1943 printk("]\n"); 1944 dlm_print_one_lock_resource(res); 1945 BUG(); 1946 } 1947 memcpy(res->lvb, mres->lvb, DLM_LVB_LEN); 1948 } 1949 } 1950 skip_lvb: 1951 1952 /* NOTE: 1953 * wrt lock queue ordering and recovery: 1954 * 1. order of locks on granted queue is 1955 * meaningless. 1956 * 2. order of locks on converting queue is 1957 * LOST with the node death. sorry charlie. 1958 * 3. order of locks on the blocked queue is 1959 * also LOST. 1960 * order of locks does not affect integrity, it 1961 * just means that a lock request may get pushed 1962 * back in line as a result of the node death. 1963 * also note that for a given node the lock order 1964 * for its secondary queue locks is preserved 1965 * relative to each other, but clearly *not* 1966 * preserved relative to locks from other nodes. 1967 */ 1968 bad = 0; 1969 spin_lock(&res->spinlock); 1970 list_for_each_entry(lock, queue, list) { 1971 if (lock->ml.cookie == ml->cookie) { 1972 c = lock->ml.cookie; 1973 mlog(ML_ERROR, "%s:%.*s: %u:%llu: lock already " 1974 "exists on this lockres!\n", dlm->name, 1975 res->lockname.len, res->lockname.name, 1976 dlm_get_lock_cookie_node(be64_to_cpu(c)), 1977 dlm_get_lock_cookie_seq(be64_to_cpu(c))); 1978 1979 mlog(ML_NOTICE, "sent lock: type=%d, conv=%d, " 1980 "node=%u, cookie=%u:%llu, queue=%d\n", 1981 ml->type, ml->convert_type, ml->node, 1982 dlm_get_lock_cookie_node(be64_to_cpu(ml->cookie)), 1983 dlm_get_lock_cookie_seq(be64_to_cpu(ml->cookie)), 1984 ml->list); 1985 1986 __dlm_print_one_lock_resource(res); 1987 bad = 1; 1988 break; 1989 } 1990 } 1991 if (!bad) { 1992 dlm_lock_get(newlock); 1993 if (mres->flags & DLM_MRES_RECOVERY && 1994 ml->list == DLM_CONVERTING_LIST && 1995 newlock->ml.type > 1996 newlock->ml.convert_type) { 1997 /* newlock is doing downconvert, add it to the 1998 * head of converting list */ 1999 list_add(&newlock->list, queue); 2000 } else 2001 list_add_tail(&newlock->list, queue); 2002 mlog(0, "%s:%.*s: added lock for node %u, " 2003 "setting refmap bit\n", dlm->name, 2004 res->lockname.len, res->lockname.name, ml->node); 2005 dlm_lockres_set_refmap_bit(dlm, res, ml->node); 2006 added++; 2007 } 2008 spin_unlock(&res->spinlock); 2009 } 2010 mlog(0, "done running all the locks\n"); 2011 2012 leave: 2013 /* balance the ref taken when the work was queued */ 2014 spin_lock(&res->spinlock); 2015 dlm_lockres_drop_inflight_ref(dlm, res); 2016 spin_unlock(&res->spinlock); 2017 2018 if (ret < 0) { 2019 mlog_errno(ret); 2020 if (newlock) 2021 dlm_lock_put(newlock); 2022 } 2023 2024 return ret; 2025 } 2026 2027 void dlm_move_lockres_to_recovery_list(struct dlm_ctxt *dlm, 2028 struct dlm_lock_resource *res) 2029 { 2030 int i; 2031 struct list_head *queue; 2032 struct dlm_lock *lock, *next; 2033 2034 assert_spin_locked(&dlm->spinlock); 2035 assert_spin_locked(&res->spinlock); 2036 res->state |= DLM_LOCK_RES_RECOVERING; 2037 if (!list_empty(&res->recovering)) { 2038 mlog(0, 2039 "Recovering res %s:%.*s, is already on recovery list!\n", 2040 dlm->name, res->lockname.len, res->lockname.name); 2041 list_del_init(&res->recovering); 2042 dlm_lockres_put(res); 2043 } 2044 /* We need to hold a reference while on the recovery list */ 2045 dlm_lockres_get(res); 2046 list_add_tail(&res->recovering, &dlm->reco.resources); 2047 2048 /* find any pending locks and put them back on proper list */ 2049 for (i=DLM_BLOCKED_LIST; i>=DLM_GRANTED_LIST; i--) { 2050 queue = dlm_list_idx_to_ptr(res, i); 2051 list_for_each_entry_safe(lock, next, queue, list) { 2052 dlm_lock_get(lock); 2053 if (lock->convert_pending) { 2054 /* move converting lock back to granted */ 2055 BUG_ON(i != DLM_CONVERTING_LIST); 2056 mlog(0, "node died with convert pending " 2057 "on %.*s. move back to granted list.\n", 2058 res->lockname.len, res->lockname.name); 2059 dlm_revert_pending_convert(res, lock); 2060 lock->convert_pending = 0; 2061 } else if (lock->lock_pending) { 2062 /* remove pending lock requests completely */ 2063 BUG_ON(i != DLM_BLOCKED_LIST); 2064 mlog(0, "node died with lock pending " 2065 "on %.*s. remove from blocked list and skip.\n", 2066 res->lockname.len, res->lockname.name); 2067 /* lock will be floating until ref in 2068 * dlmlock_remote is freed after the network 2069 * call returns. ok for it to not be on any 2070 * list since no ast can be called 2071 * (the master is dead). */ 2072 dlm_revert_pending_lock(res, lock); 2073 lock->lock_pending = 0; 2074 } else if (lock->unlock_pending) { 2075 /* if an unlock was in progress, treat as 2076 * if this had completed successfully 2077 * before sending this lock state to the 2078 * new master. note that the dlm_unlock 2079 * call is still responsible for calling 2080 * the unlockast. that will happen after 2081 * the network call times out. for now, 2082 * just move lists to prepare the new 2083 * recovery master. */ 2084 BUG_ON(i != DLM_GRANTED_LIST); 2085 mlog(0, "node died with unlock pending " 2086 "on %.*s. remove from blocked list and skip.\n", 2087 res->lockname.len, res->lockname.name); 2088 dlm_commit_pending_unlock(res, lock); 2089 lock->unlock_pending = 0; 2090 } else if (lock->cancel_pending) { 2091 /* if a cancel was in progress, treat as 2092 * if this had completed successfully 2093 * before sending this lock state to the 2094 * new master */ 2095 BUG_ON(i != DLM_CONVERTING_LIST); 2096 mlog(0, "node died with cancel pending " 2097 "on %.*s. move back to granted list.\n", 2098 res->lockname.len, res->lockname.name); 2099 dlm_commit_pending_cancel(res, lock); 2100 lock->cancel_pending = 0; 2101 } 2102 dlm_lock_put(lock); 2103 } 2104 } 2105 } 2106 2107 2108 2109 /* removes all recovered locks from the recovery list. 2110 * sets the res->owner to the new master. 2111 * unsets the RECOVERY flag and wakes waiters. */ 2112 static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm, 2113 u8 dead_node, u8 new_master) 2114 { 2115 int i; 2116 struct hlist_head *bucket; 2117 struct dlm_lock_resource *res, *next; 2118 2119 assert_spin_locked(&dlm->spinlock); 2120 2121 list_for_each_entry_safe(res, next, &dlm->reco.resources, recovering) { 2122 if (res->owner == dead_node) { 2123 mlog(0, "%s: res %.*s, Changing owner from %u to %u\n", 2124 dlm->name, res->lockname.len, res->lockname.name, 2125 res->owner, new_master); 2126 list_del_init(&res->recovering); 2127 spin_lock(&res->spinlock); 2128 /* new_master has our reference from 2129 * the lock state sent during recovery */ 2130 dlm_change_lockres_owner(dlm, res, new_master); 2131 res->state &= ~DLM_LOCK_RES_RECOVERING; 2132 if (__dlm_lockres_has_locks(res)) 2133 __dlm_dirty_lockres(dlm, res); 2134 spin_unlock(&res->spinlock); 2135 wake_up(&res->wq); 2136 dlm_lockres_put(res); 2137 } 2138 } 2139 2140 /* this will become unnecessary eventually, but 2141 * for now we need to run the whole hash, clear 2142 * the RECOVERING state and set the owner 2143 * if necessary */ 2144 for (i = 0; i < DLM_HASH_BUCKETS; i++) { 2145 bucket = dlm_lockres_hash(dlm, i); 2146 hlist_for_each_entry(res, bucket, hash_node) { 2147 if (!(res->state & DLM_LOCK_RES_RECOVERING)) 2148 continue; 2149 2150 if (res->owner != dead_node && 2151 res->owner != dlm->node_num) 2152 continue; 2153 2154 if (!list_empty(&res->recovering)) { 2155 list_del_init(&res->recovering); 2156 dlm_lockres_put(res); 2157 } 2158 2159 /* new_master has our reference from 2160 * the lock state sent during recovery */ 2161 mlog(0, "%s: res %.*s, Changing owner from %u to %u\n", 2162 dlm->name, res->lockname.len, res->lockname.name, 2163 res->owner, new_master); 2164 spin_lock(&res->spinlock); 2165 dlm_change_lockres_owner(dlm, res, new_master); 2166 res->state &= ~DLM_LOCK_RES_RECOVERING; 2167 if (__dlm_lockres_has_locks(res)) 2168 __dlm_dirty_lockres(dlm, res); 2169 spin_unlock(&res->spinlock); 2170 wake_up(&res->wq); 2171 } 2172 } 2173 } 2174 2175 static inline int dlm_lvb_needs_invalidation(struct dlm_lock *lock, int local) 2176 { 2177 if (local) { 2178 if (lock->ml.type != LKM_EXMODE && 2179 lock->ml.type != LKM_PRMODE) 2180 return 1; 2181 } else if (lock->ml.type == LKM_EXMODE) 2182 return 1; 2183 return 0; 2184 } 2185 2186 static void dlm_revalidate_lvb(struct dlm_ctxt *dlm, 2187 struct dlm_lock_resource *res, u8 dead_node) 2188 { 2189 struct list_head *queue; 2190 struct dlm_lock *lock; 2191 int blank_lvb = 0, local = 0; 2192 int i; 2193 u8 search_node; 2194 2195 assert_spin_locked(&dlm->spinlock); 2196 assert_spin_locked(&res->spinlock); 2197 2198 if (res->owner == dlm->node_num) 2199 /* if this node owned the lockres, and if the dead node 2200 * had an EX when he died, blank out the lvb */ 2201 search_node = dead_node; 2202 else { 2203 /* if this is a secondary lockres, and we had no EX or PR 2204 * locks granted, we can no longer trust the lvb */ 2205 search_node = dlm->node_num; 2206 local = 1; /* check local state for valid lvb */ 2207 } 2208 2209 for (i=DLM_GRANTED_LIST; i<=DLM_CONVERTING_LIST; i++) { 2210 queue = dlm_list_idx_to_ptr(res, i); 2211 list_for_each_entry(lock, queue, list) { 2212 if (lock->ml.node == search_node) { 2213 if (dlm_lvb_needs_invalidation(lock, local)) { 2214 /* zero the lksb lvb and lockres lvb */ 2215 blank_lvb = 1; 2216 memset(lock->lksb->lvb, 0, DLM_LVB_LEN); 2217 } 2218 } 2219 } 2220 } 2221 2222 if (blank_lvb) { 2223 mlog(0, "clearing %.*s lvb, dead node %u had EX\n", 2224 res->lockname.len, res->lockname.name, dead_node); 2225 memset(res->lvb, 0, DLM_LVB_LEN); 2226 } 2227 } 2228 2229 static void dlm_free_dead_locks(struct dlm_ctxt *dlm, 2230 struct dlm_lock_resource *res, u8 dead_node) 2231 { 2232 struct dlm_lock *lock, *next; 2233 unsigned int freed = 0; 2234 2235 /* this node is the lockres master: 2236 * 1) remove any stale locks for the dead node 2237 * 2) if the dead node had an EX when he died, blank out the lvb 2238 */ 2239 assert_spin_locked(&dlm->spinlock); 2240 assert_spin_locked(&res->spinlock); 2241 2242 /* We do two dlm_lock_put(). One for removing from list and the other is 2243 * to force the DLM_UNLOCK_FREE_LOCK action so as to free the locks */ 2244 2245 /* TODO: check pending_asts, pending_basts here */ 2246 list_for_each_entry_safe(lock, next, &res->granted, list) { 2247 if (lock->ml.node == dead_node) { 2248 list_del_init(&lock->list); 2249 dlm_lock_put(lock); 2250 /* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */ 2251 dlm_lock_put(lock); 2252 freed++; 2253 } 2254 } 2255 list_for_each_entry_safe(lock, next, &res->converting, list) { 2256 if (lock->ml.node == dead_node) { 2257 list_del_init(&lock->list); 2258 dlm_lock_put(lock); 2259 /* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */ 2260 dlm_lock_put(lock); 2261 freed++; 2262 } 2263 } 2264 list_for_each_entry_safe(lock, next, &res->blocked, list) { 2265 if (lock->ml.node == dead_node) { 2266 list_del_init(&lock->list); 2267 dlm_lock_put(lock); 2268 /* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */ 2269 dlm_lock_put(lock); 2270 freed++; 2271 } 2272 } 2273 2274 if (freed) { 2275 mlog(0, "%s:%.*s: freed %u locks for dead node %u, " 2276 "dropping ref from lockres\n", dlm->name, 2277 res->lockname.len, res->lockname.name, freed, dead_node); 2278 if(!test_bit(dead_node, res->refmap)) { 2279 mlog(ML_ERROR, "%s:%.*s: freed %u locks for dead node %u, " 2280 "but ref was not set\n", dlm->name, 2281 res->lockname.len, res->lockname.name, freed, dead_node); 2282 __dlm_print_one_lock_resource(res); 2283 } 2284 dlm_lockres_clear_refmap_bit(dlm, res, dead_node); 2285 } else if (test_bit(dead_node, res->refmap)) { 2286 mlog(0, "%s:%.*s: dead node %u had a ref, but had " 2287 "no locks and had not purged before dying\n", dlm->name, 2288 res->lockname.len, res->lockname.name, dead_node); 2289 dlm_lockres_clear_refmap_bit(dlm, res, dead_node); 2290 } 2291 2292 /* do not kick thread yet */ 2293 __dlm_dirty_lockres(dlm, res); 2294 } 2295 2296 /* if this node is the recovery master, and there are no 2297 * locks for a given lockres owned by this node that are in 2298 * either PR or EX mode, zero out the lvb before requesting. 2299 * 2300 */ 2301 2302 2303 static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node) 2304 { 2305 struct dlm_lock_resource *res; 2306 int i; 2307 struct hlist_head *bucket; 2308 struct dlm_lock *lock; 2309 2310 2311 /* purge any stale mles */ 2312 dlm_clean_master_list(dlm, dead_node); 2313 2314 /* 2315 * now clean up all lock resources. there are two rules: 2316 * 2317 * 1) if the dead node was the master, move the lockres 2318 * to the recovering list. set the RECOVERING flag. 2319 * this lockres needs to be cleaned up before it can 2320 * be used further. 2321 * 2322 * 2) if this node was the master, remove all locks from 2323 * each of the lockres queues that were owned by the 2324 * dead node. once recovery finishes, the dlm thread 2325 * can be kicked again to see if any ASTs or BASTs 2326 * need to be fired as a result. 2327 */ 2328 for (i = 0; i < DLM_HASH_BUCKETS; i++) { 2329 bucket = dlm_lockres_hash(dlm, i); 2330 hlist_for_each_entry(res, bucket, hash_node) { 2331 /* always prune any $RECOVERY entries for dead nodes, 2332 * otherwise hangs can occur during later recovery */ 2333 if (dlm_is_recovery_lock(res->lockname.name, 2334 res->lockname.len)) { 2335 spin_lock(&res->spinlock); 2336 list_for_each_entry(lock, &res->granted, list) { 2337 if (lock->ml.node == dead_node) { 2338 mlog(0, "AHA! there was " 2339 "a $RECOVERY lock for dead " 2340 "node %u (%s)!\n", 2341 dead_node, dlm->name); 2342 list_del_init(&lock->list); 2343 dlm_lock_put(lock); 2344 break; 2345 } 2346 } 2347 spin_unlock(&res->spinlock); 2348 continue; 2349 } 2350 spin_lock(&res->spinlock); 2351 /* zero the lvb if necessary */ 2352 dlm_revalidate_lvb(dlm, res, dead_node); 2353 if (res->owner == dead_node) { 2354 if (res->state & DLM_LOCK_RES_DROPPING_REF) { 2355 mlog(ML_NOTICE, "%s: res %.*s, Skip " 2356 "recovery as it is being freed\n", 2357 dlm->name, res->lockname.len, 2358 res->lockname.name); 2359 } else 2360 dlm_move_lockres_to_recovery_list(dlm, 2361 res); 2362 2363 } else if (res->owner == dlm->node_num) { 2364 dlm_free_dead_locks(dlm, res, dead_node); 2365 __dlm_lockres_calc_usage(dlm, res); 2366 } else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) { 2367 if (test_bit(dead_node, res->refmap)) { 2368 mlog(0, "%s:%.*s: dead node %u had a ref, but had " 2369 "no locks and had not purged before dying\n", 2370 dlm->name, res->lockname.len, 2371 res->lockname.name, dead_node); 2372 dlm_lockres_clear_refmap_bit(dlm, res, dead_node); 2373 } 2374 } 2375 spin_unlock(&res->spinlock); 2376 } 2377 } 2378 2379 } 2380 2381 static void __dlm_hb_node_down(struct dlm_ctxt *dlm, int idx) 2382 { 2383 assert_spin_locked(&dlm->spinlock); 2384 2385 if (dlm->reco.new_master == idx) { 2386 mlog(0, "%s: recovery master %d just died\n", 2387 dlm->name, idx); 2388 if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) { 2389 /* finalize1 was reached, so it is safe to clear 2390 * the new_master and dead_node. that recovery 2391 * is complete. */ 2392 mlog(0, "%s: dead master %d had reached " 2393 "finalize1 state, clearing\n", dlm->name, idx); 2394 dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE; 2395 __dlm_reset_recovery(dlm); 2396 } 2397 } 2398 2399 /* Clean up join state on node death. */ 2400 if (dlm->joining_node == idx) { 2401 mlog(0, "Clearing join state for node %u\n", idx); 2402 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN); 2403 } 2404 2405 /* check to see if the node is already considered dead */ 2406 if (!test_bit(idx, dlm->live_nodes_map)) { 2407 mlog(0, "for domain %s, node %d is already dead. " 2408 "another node likely did recovery already.\n", 2409 dlm->name, idx); 2410 return; 2411 } 2412 2413 /* check to see if we do not care about this node */ 2414 if (!test_bit(idx, dlm->domain_map)) { 2415 /* This also catches the case that we get a node down 2416 * but haven't joined the domain yet. */ 2417 mlog(0, "node %u already removed from domain!\n", idx); 2418 return; 2419 } 2420 2421 clear_bit(idx, dlm->live_nodes_map); 2422 2423 /* make sure local cleanup occurs before the heartbeat events */ 2424 if (!test_bit(idx, dlm->recovery_map)) 2425 dlm_do_local_recovery_cleanup(dlm, idx); 2426 2427 /* notify anything attached to the heartbeat events */ 2428 dlm_hb_event_notify_attached(dlm, idx, 0); 2429 2430 mlog(0, "node %u being removed from domain map!\n", idx); 2431 clear_bit(idx, dlm->domain_map); 2432 clear_bit(idx, dlm->exit_domain_map); 2433 /* wake up migration waiters if a node goes down. 2434 * perhaps later we can genericize this for other waiters. */ 2435 wake_up(&dlm->migration_wq); 2436 2437 if (test_bit(idx, dlm->recovery_map)) 2438 mlog(0, "domain %s, node %u already added " 2439 "to recovery map!\n", dlm->name, idx); 2440 else 2441 set_bit(idx, dlm->recovery_map); 2442 } 2443 2444 void dlm_hb_node_down_cb(struct o2nm_node *node, int idx, void *data) 2445 { 2446 struct dlm_ctxt *dlm = data; 2447 2448 if (!dlm_grab(dlm)) 2449 return; 2450 2451 /* 2452 * This will notify any dlm users that a node in our domain 2453 * went away without notifying us first. 2454 */ 2455 if (test_bit(idx, dlm->domain_map)) 2456 dlm_fire_domain_eviction_callbacks(dlm, idx); 2457 2458 spin_lock(&dlm->spinlock); 2459 __dlm_hb_node_down(dlm, idx); 2460 spin_unlock(&dlm->spinlock); 2461 2462 dlm_put(dlm); 2463 } 2464 2465 void dlm_hb_node_up_cb(struct o2nm_node *node, int idx, void *data) 2466 { 2467 struct dlm_ctxt *dlm = data; 2468 2469 if (!dlm_grab(dlm)) 2470 return; 2471 2472 spin_lock(&dlm->spinlock); 2473 set_bit(idx, dlm->live_nodes_map); 2474 /* do NOT notify mle attached to the heartbeat events. 2475 * new nodes are not interesting in mastery until joined. */ 2476 spin_unlock(&dlm->spinlock); 2477 2478 dlm_put(dlm); 2479 } 2480 2481 static void dlm_reco_ast(void *astdata) 2482 { 2483 struct dlm_ctxt *dlm = astdata; 2484 mlog(0, "ast for recovery lock fired!, this=%u, dlm=%s\n", 2485 dlm->node_num, dlm->name); 2486 } 2487 static void dlm_reco_bast(void *astdata, int blocked_type) 2488 { 2489 struct dlm_ctxt *dlm = astdata; 2490 mlog(0, "bast for recovery lock fired!, this=%u, dlm=%s\n", 2491 dlm->node_num, dlm->name); 2492 } 2493 static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st) 2494 { 2495 mlog(0, "unlockast for recovery lock fired!\n"); 2496 } 2497 2498 /* 2499 * dlm_pick_recovery_master will continually attempt to use 2500 * dlmlock() on the special "$RECOVERY" lockres with the 2501 * LKM_NOQUEUE flag to get an EX. every thread that enters 2502 * this function on each node racing to become the recovery 2503 * master will not stop attempting this until either: 2504 * a) this node gets the EX (and becomes the recovery master), 2505 * or b) dlm->reco.new_master gets set to some nodenum 2506 * != O2NM_INVALID_NODE_NUM (another node will do the reco). 2507 * so each time a recovery master is needed, the entire cluster 2508 * will sync at this point. if the new master dies, that will 2509 * be detected in dlm_do_recovery */ 2510 static int dlm_pick_recovery_master(struct dlm_ctxt *dlm) 2511 { 2512 enum dlm_status ret; 2513 struct dlm_lockstatus lksb; 2514 int status = -EINVAL; 2515 2516 mlog(0, "starting recovery of %s at %lu, dead=%u, this=%u\n", 2517 dlm->name, jiffies, dlm->reco.dead_node, dlm->node_num); 2518 again: 2519 memset(&lksb, 0, sizeof(lksb)); 2520 2521 ret = dlmlock(dlm, LKM_EXMODE, &lksb, LKM_NOQUEUE|LKM_RECOVERY, 2522 DLM_RECOVERY_LOCK_NAME, DLM_RECOVERY_LOCK_NAME_LEN, 2523 dlm_reco_ast, dlm, dlm_reco_bast); 2524 2525 mlog(0, "%s: dlmlock($RECOVERY) returned %d, lksb=%d\n", 2526 dlm->name, ret, lksb.status); 2527 2528 if (ret == DLM_NORMAL) { 2529 mlog(0, "dlm=%s dlmlock says I got it (this=%u)\n", 2530 dlm->name, dlm->node_num); 2531 2532 /* got the EX lock. check to see if another node 2533 * just became the reco master */ 2534 if (dlm_reco_master_ready(dlm)) { 2535 mlog(0, "%s: got reco EX lock, but %u will " 2536 "do the recovery\n", dlm->name, 2537 dlm->reco.new_master); 2538 status = -EEXIST; 2539 } else { 2540 status = 0; 2541 2542 /* see if recovery was already finished elsewhere */ 2543 spin_lock(&dlm->spinlock); 2544 if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) { 2545 status = -EINVAL; 2546 mlog(0, "%s: got reco EX lock, but " 2547 "node got recovered already\n", dlm->name); 2548 if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) { 2549 mlog(ML_ERROR, "%s: new master is %u " 2550 "but no dead node!\n", 2551 dlm->name, dlm->reco.new_master); 2552 BUG(); 2553 } 2554 } 2555 spin_unlock(&dlm->spinlock); 2556 } 2557 2558 /* if this node has actually become the recovery master, 2559 * set the master and send the messages to begin recovery */ 2560 if (!status) { 2561 mlog(0, "%s: dead=%u, this=%u, sending " 2562 "begin_reco now\n", dlm->name, 2563 dlm->reco.dead_node, dlm->node_num); 2564 status = dlm_send_begin_reco_message(dlm, 2565 dlm->reco.dead_node); 2566 /* this always succeeds */ 2567 BUG_ON(status); 2568 2569 /* set the new_master to this node */ 2570 spin_lock(&dlm->spinlock); 2571 dlm_set_reco_master(dlm, dlm->node_num); 2572 spin_unlock(&dlm->spinlock); 2573 } 2574 2575 /* recovery lock is a special case. ast will not get fired, 2576 * so just go ahead and unlock it. */ 2577 ret = dlmunlock(dlm, &lksb, 0, dlm_reco_unlock_ast, dlm); 2578 if (ret == DLM_DENIED) { 2579 mlog(0, "got DLM_DENIED, trying LKM_CANCEL\n"); 2580 ret = dlmunlock(dlm, &lksb, LKM_CANCEL, dlm_reco_unlock_ast, dlm); 2581 } 2582 if (ret != DLM_NORMAL) { 2583 /* this would really suck. this could only happen 2584 * if there was a network error during the unlock 2585 * because of node death. this means the unlock 2586 * is actually "done" and the lock structure is 2587 * even freed. we can continue, but only 2588 * because this specific lock name is special. */ 2589 mlog(ML_ERROR, "dlmunlock returned %d\n", ret); 2590 } 2591 } else if (ret == DLM_NOTQUEUED) { 2592 mlog(0, "dlm=%s dlmlock says another node got it (this=%u)\n", 2593 dlm->name, dlm->node_num); 2594 /* another node is master. wait on 2595 * reco.new_master != O2NM_INVALID_NODE_NUM 2596 * for at most one second */ 2597 wait_event_timeout(dlm->dlm_reco_thread_wq, 2598 dlm_reco_master_ready(dlm), 2599 msecs_to_jiffies(1000)); 2600 if (!dlm_reco_master_ready(dlm)) { 2601 mlog(0, "%s: reco master taking awhile\n", 2602 dlm->name); 2603 goto again; 2604 } 2605 /* another node has informed this one that it is reco master */ 2606 mlog(0, "%s: reco master %u is ready to recover %u\n", 2607 dlm->name, dlm->reco.new_master, dlm->reco.dead_node); 2608 status = -EEXIST; 2609 } else if (ret == DLM_RECOVERING) { 2610 mlog(0, "dlm=%s dlmlock says master node died (this=%u)\n", 2611 dlm->name, dlm->node_num); 2612 goto again; 2613 } else { 2614 struct dlm_lock_resource *res; 2615 2616 /* dlmlock returned something other than NOTQUEUED or NORMAL */ 2617 mlog(ML_ERROR, "%s: got %s from dlmlock($RECOVERY), " 2618 "lksb.status=%s\n", dlm->name, dlm_errname(ret), 2619 dlm_errname(lksb.status)); 2620 res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME, 2621 DLM_RECOVERY_LOCK_NAME_LEN); 2622 if (res) { 2623 dlm_print_one_lock_resource(res); 2624 dlm_lockres_put(res); 2625 } else { 2626 mlog(ML_ERROR, "recovery lock not found\n"); 2627 } 2628 BUG(); 2629 } 2630 2631 return status; 2632 } 2633 2634 static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node) 2635 { 2636 struct dlm_begin_reco br; 2637 int ret = 0; 2638 struct dlm_node_iter iter; 2639 int nodenum; 2640 int status; 2641 2642 mlog(0, "%s: dead node is %u\n", dlm->name, dead_node); 2643 2644 spin_lock(&dlm->spinlock); 2645 dlm_node_iter_init(dlm->domain_map, &iter); 2646 spin_unlock(&dlm->spinlock); 2647 2648 clear_bit(dead_node, iter.node_map); 2649 2650 memset(&br, 0, sizeof(br)); 2651 br.node_idx = dlm->node_num; 2652 br.dead_node = dead_node; 2653 2654 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) { 2655 ret = 0; 2656 if (nodenum == dead_node) { 2657 mlog(0, "not sending begin reco to dead node " 2658 "%u\n", dead_node); 2659 continue; 2660 } 2661 if (nodenum == dlm->node_num) { 2662 mlog(0, "not sending begin reco to self\n"); 2663 continue; 2664 } 2665 retry: 2666 ret = -EINVAL; 2667 mlog(0, "attempting to send begin reco msg to %d\n", 2668 nodenum); 2669 ret = o2net_send_message(DLM_BEGIN_RECO_MSG, dlm->key, 2670 &br, sizeof(br), nodenum, &status); 2671 /* negative status is handled ok by caller here */ 2672 if (ret >= 0) 2673 ret = status; 2674 if (dlm_is_host_down(ret)) { 2675 /* node is down. not involved in recovery 2676 * so just keep going */ 2677 mlog(ML_NOTICE, "%s: node %u was down when sending " 2678 "begin reco msg (%d)\n", dlm->name, nodenum, ret); 2679 ret = 0; 2680 } 2681 2682 /* 2683 * Prior to commit aad1b15310b9bcd59fa81ab8f2b1513b59553ea8, 2684 * dlm_begin_reco_handler() returned EAGAIN and not -EAGAIN. 2685 * We are handling both for compatibility reasons. 2686 */ 2687 if (ret == -EAGAIN || ret == EAGAIN) { 2688 mlog(0, "%s: trying to start recovery of node " 2689 "%u, but node %u is waiting for last recovery " 2690 "to complete, backoff for a bit\n", dlm->name, 2691 dead_node, nodenum); 2692 msleep(100); 2693 goto retry; 2694 } 2695 if (ret < 0) { 2696 struct dlm_lock_resource *res; 2697 2698 /* this is now a serious problem, possibly ENOMEM 2699 * in the network stack. must retry */ 2700 mlog_errno(ret); 2701 mlog(ML_ERROR, "begin reco of dlm %s to node %u " 2702 "returned %d\n", dlm->name, nodenum, ret); 2703 res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME, 2704 DLM_RECOVERY_LOCK_NAME_LEN); 2705 if (res) { 2706 dlm_print_one_lock_resource(res); 2707 dlm_lockres_put(res); 2708 } else { 2709 mlog(ML_ERROR, "recovery lock not found\n"); 2710 } 2711 /* sleep for a bit in hopes that we can avoid 2712 * another ENOMEM */ 2713 msleep(100); 2714 goto retry; 2715 } 2716 } 2717 2718 return ret; 2719 } 2720 2721 int dlm_begin_reco_handler(struct o2net_msg *msg, u32 len, void *data, 2722 void **ret_data) 2723 { 2724 struct dlm_ctxt *dlm = data; 2725 struct dlm_begin_reco *br = (struct dlm_begin_reco *)msg->buf; 2726 2727 /* ok to return 0, domain has gone away */ 2728 if (!dlm_grab(dlm)) 2729 return 0; 2730 2731 spin_lock(&dlm->spinlock); 2732 if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) { 2733 mlog(0, "%s: node %u wants to recover node %u (%u:%u) " 2734 "but this node is in finalize state, waiting on finalize2\n", 2735 dlm->name, br->node_idx, br->dead_node, 2736 dlm->reco.dead_node, dlm->reco.new_master); 2737 spin_unlock(&dlm->spinlock); 2738 dlm_put(dlm); 2739 return -EAGAIN; 2740 } 2741 spin_unlock(&dlm->spinlock); 2742 2743 mlog(0, "%s: node %u wants to recover node %u (%u:%u)\n", 2744 dlm->name, br->node_idx, br->dead_node, 2745 dlm->reco.dead_node, dlm->reco.new_master); 2746 2747 dlm_fire_domain_eviction_callbacks(dlm, br->dead_node); 2748 2749 spin_lock(&dlm->spinlock); 2750 if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) { 2751 if (test_bit(dlm->reco.new_master, dlm->recovery_map)) { 2752 mlog(0, "%s: new_master %u died, changing " 2753 "to %u\n", dlm->name, dlm->reco.new_master, 2754 br->node_idx); 2755 } else { 2756 mlog(0, "%s: new_master %u NOT DEAD, changing " 2757 "to %u\n", dlm->name, dlm->reco.new_master, 2758 br->node_idx); 2759 /* may not have seen the new master as dead yet */ 2760 } 2761 } 2762 if (dlm->reco.dead_node != O2NM_INVALID_NODE_NUM) { 2763 mlog(ML_NOTICE, "%s: dead_node previously set to %u, " 2764 "node %u changing it to %u\n", dlm->name, 2765 dlm->reco.dead_node, br->node_idx, br->dead_node); 2766 } 2767 dlm_set_reco_master(dlm, br->node_idx); 2768 dlm_set_reco_dead_node(dlm, br->dead_node); 2769 if (!test_bit(br->dead_node, dlm->recovery_map)) { 2770 mlog(0, "recovery master %u sees %u as dead, but this " 2771 "node has not yet. marking %u as dead\n", 2772 br->node_idx, br->dead_node, br->dead_node); 2773 if (!test_bit(br->dead_node, dlm->domain_map) || 2774 !test_bit(br->dead_node, dlm->live_nodes_map)) 2775 mlog(0, "%u not in domain/live_nodes map " 2776 "so setting it in reco map manually\n", 2777 br->dead_node); 2778 /* force the recovery cleanup in __dlm_hb_node_down 2779 * both of these will be cleared in a moment */ 2780 set_bit(br->dead_node, dlm->domain_map); 2781 set_bit(br->dead_node, dlm->live_nodes_map); 2782 __dlm_hb_node_down(dlm, br->dead_node); 2783 } 2784 spin_unlock(&dlm->spinlock); 2785 2786 dlm_kick_recovery_thread(dlm); 2787 2788 mlog(0, "%s: recovery started by node %u, for %u (%u:%u)\n", 2789 dlm->name, br->node_idx, br->dead_node, 2790 dlm->reco.dead_node, dlm->reco.new_master); 2791 2792 dlm_put(dlm); 2793 return 0; 2794 } 2795 2796 #define DLM_FINALIZE_STAGE2 0x01 2797 static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm) 2798 { 2799 int ret = 0; 2800 struct dlm_finalize_reco fr; 2801 struct dlm_node_iter iter; 2802 int nodenum; 2803 int status; 2804 int stage = 1; 2805 2806 mlog(0, "finishing recovery for node %s:%u, " 2807 "stage %d\n", dlm->name, dlm->reco.dead_node, stage); 2808 2809 spin_lock(&dlm->spinlock); 2810 dlm_node_iter_init(dlm->domain_map, &iter); 2811 spin_unlock(&dlm->spinlock); 2812 2813 stage2: 2814 memset(&fr, 0, sizeof(fr)); 2815 fr.node_idx = dlm->node_num; 2816 fr.dead_node = dlm->reco.dead_node; 2817 if (stage == 2) 2818 fr.flags |= DLM_FINALIZE_STAGE2; 2819 2820 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) { 2821 if (nodenum == dlm->node_num) 2822 continue; 2823 ret = o2net_send_message(DLM_FINALIZE_RECO_MSG, dlm->key, 2824 &fr, sizeof(fr), nodenum, &status); 2825 if (ret >= 0) 2826 ret = status; 2827 if (ret < 0) { 2828 mlog(ML_ERROR, "Error %d when sending message %u (key " 2829 "0x%x) to node %u\n", ret, DLM_FINALIZE_RECO_MSG, 2830 dlm->key, nodenum); 2831 if (dlm_is_host_down(ret)) { 2832 /* this has no effect on this recovery 2833 * session, so set the status to zero to 2834 * finish out the last recovery */ 2835 mlog(ML_ERROR, "node %u went down after this " 2836 "node finished recovery.\n", nodenum); 2837 ret = 0; 2838 continue; 2839 } 2840 break; 2841 } 2842 } 2843 if (stage == 1) { 2844 /* reset the node_iter back to the top and send finalize2 */ 2845 iter.curnode = -1; 2846 stage = 2; 2847 goto stage2; 2848 } 2849 2850 return ret; 2851 } 2852 2853 int dlm_finalize_reco_handler(struct o2net_msg *msg, u32 len, void *data, 2854 void **ret_data) 2855 { 2856 struct dlm_ctxt *dlm = data; 2857 struct dlm_finalize_reco *fr = (struct dlm_finalize_reco *)msg->buf; 2858 int stage = 1; 2859 2860 /* ok to return 0, domain has gone away */ 2861 if (!dlm_grab(dlm)) 2862 return 0; 2863 2864 if (fr->flags & DLM_FINALIZE_STAGE2) 2865 stage = 2; 2866 2867 mlog(0, "%s: node %u finalizing recovery stage%d of " 2868 "node %u (%u:%u)\n", dlm->name, fr->node_idx, stage, 2869 fr->dead_node, dlm->reco.dead_node, dlm->reco.new_master); 2870 2871 spin_lock(&dlm->spinlock); 2872 2873 if (dlm->reco.new_master != fr->node_idx) { 2874 mlog(ML_ERROR, "node %u sent recovery finalize msg, but node " 2875 "%u is supposed to be the new master, dead=%u\n", 2876 fr->node_idx, dlm->reco.new_master, fr->dead_node); 2877 BUG(); 2878 } 2879 if (dlm->reco.dead_node != fr->dead_node) { 2880 mlog(ML_ERROR, "node %u sent recovery finalize msg for dead " 2881 "node %u, but node %u is supposed to be dead\n", 2882 fr->node_idx, fr->dead_node, dlm->reco.dead_node); 2883 BUG(); 2884 } 2885 2886 switch (stage) { 2887 case 1: 2888 dlm_finish_local_lockres_recovery(dlm, fr->dead_node, fr->node_idx); 2889 if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) { 2890 mlog(ML_ERROR, "%s: received finalize1 from " 2891 "new master %u for dead node %u, but " 2892 "this node has already received it!\n", 2893 dlm->name, fr->node_idx, fr->dead_node); 2894 dlm_print_reco_node_status(dlm); 2895 BUG(); 2896 } 2897 dlm->reco.state |= DLM_RECO_STATE_FINALIZE; 2898 spin_unlock(&dlm->spinlock); 2899 break; 2900 case 2: 2901 if (!(dlm->reco.state & DLM_RECO_STATE_FINALIZE)) { 2902 mlog(ML_ERROR, "%s: received finalize2 from " 2903 "new master %u for dead node %u, but " 2904 "this node did not have finalize1!\n", 2905 dlm->name, fr->node_idx, fr->dead_node); 2906 dlm_print_reco_node_status(dlm); 2907 BUG(); 2908 } 2909 dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE; 2910 __dlm_reset_recovery(dlm); 2911 spin_unlock(&dlm->spinlock); 2912 dlm_kick_recovery_thread(dlm); 2913 break; 2914 default: 2915 BUG(); 2916 } 2917 2918 mlog(0, "%s: recovery done, reco master was %u, dead now %u, master now %u\n", 2919 dlm->name, fr->node_idx, dlm->reco.dead_node, dlm->reco.new_master); 2920 2921 dlm_put(dlm); 2922 return 0; 2923 } 2924