1/* -*- mode: c; c-basic-offset: 8; -*- 2 * vim: noexpandtab sw=8 ts=8 sts=0: 3 * 4 * dlmconvert.c 5 * 6 * underlying calls for lock conversion 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/highmem.h> 32#include <linux/init.h> 33#include <linux/sysctl.h> 34#include <linux/random.h> 35#include <linux/blkdev.h> 36#include <linux/socket.h> 37#include <linux/inet.h> 38#include <linux/spinlock.h> 39 40 41#include "cluster/heartbeat.h" 42#include "cluster/nodemanager.h" 43#include "cluster/tcp.h" 44 45#include "dlmapi.h" 46#include "dlmcommon.h" 47 48#include "dlmconvert.h" 49 50#define MLOG_MASK_PREFIX ML_DLM 51#include "cluster/masklog.h" 52 53/* NOTE: __dlmconvert_master is the only function in here that 54 * needs a spinlock held on entry (res->spinlock) and it is the 55 * only one that holds a lock on exit (res->spinlock). 56 * All other functions in here need no locks and drop all of 57 * the locks that they acquire. */ 58static enum dlm_status __dlmconvert_master(struct dlm_ctxt *dlm, 59 struct dlm_lock_resource *res, 60 struct dlm_lock *lock, int flags, 61 int type, int *call_ast, 62 int *kick_thread); 63static enum dlm_status dlm_send_remote_convert_request(struct dlm_ctxt *dlm, 64 struct dlm_lock_resource *res, 65 struct dlm_lock *lock, int flags, int type); 66 67/* 68 * this is only called directly by dlmlock(), and only when the 69 * local node is the owner of the lockres 70 * locking: 71 * caller needs: none 72 * taken: takes and drops res->spinlock 73 * held on exit: none 74 * returns: see __dlmconvert_master 75 */ 76enum dlm_status dlmconvert_master(struct dlm_ctxt *dlm, 77 struct dlm_lock_resource *res, 78 struct dlm_lock *lock, int flags, int type) 79{ 80 int call_ast = 0, kick_thread = 0; 81 enum dlm_status status; 82 83 spin_lock(&res->spinlock); 84 /* we are not in a network handler, this is fine */ 85 __dlm_wait_on_lockres(res); 86 __dlm_lockres_reserve_ast(res); 87 res->state |= DLM_LOCK_RES_IN_PROGRESS; 88 89 status = __dlmconvert_master(dlm, res, lock, flags, type, 90 &call_ast, &kick_thread); 91 92 res->state &= ~DLM_LOCK_RES_IN_PROGRESS; 93 spin_unlock(&res->spinlock); 94 wake_up(&res->wq); 95 if (status != DLM_NORMAL && status != DLM_NOTQUEUED) 96 dlm_error(status); 97 98 /* either queue the ast or release it */ 99 if (call_ast) 100 dlm_queue_ast(dlm, lock); 101 else 102 dlm_lockres_release_ast(dlm, res); 103 104 if (kick_thread) 105 dlm_kick_thread(dlm, res); 106 107 return status; 108} 109 110/* performs lock conversion at the lockres master site 111 * locking: 112 * caller needs: res->spinlock 113 * taken: takes and drops lock->spinlock 114 * held on exit: res->spinlock 115 * returns: DLM_NORMAL, DLM_NOTQUEUED, DLM_DENIED 116 * call_ast: whether ast should be called for this lock 117 * kick_thread: whether dlm_kick_thread should be called 118 */ 119static enum dlm_status __dlmconvert_master(struct dlm_ctxt *dlm, 120 struct dlm_lock_resource *res, 121 struct dlm_lock *lock, int flags, 122 int type, int *call_ast, 123 int *kick_thread) 124{ 125 enum dlm_status status = DLM_NORMAL; 126 struct dlm_lock *tmplock=NULL; 127 128 assert_spin_locked(&res->spinlock); 129 130 mlog(0, "type=%d, convert_type=%d, new convert_type=%d\n", 131 lock->ml.type, lock->ml.convert_type, type); 132 133 spin_lock(&lock->spinlock); 134 135 /* already converting? */ 136 if (lock->ml.convert_type != LKM_IVMODE) { 137 mlog(ML_ERROR, "attempted to convert a lock with a lock " 138 "conversion pending\n"); 139 status = DLM_DENIED; 140 goto unlock_exit; 141 } 142 143 /* must be on grant queue to convert */ 144 if (!dlm_lock_on_list(&res->granted, lock)) { 145 mlog(ML_ERROR, "attempted to convert a lock not on grant " 146 "queue\n"); 147 status = DLM_DENIED; 148 goto unlock_exit; 149 } 150 151 if (flags & LKM_VALBLK) { 152 switch (lock->ml.type) { 153 case LKM_EXMODE: 154 /* EX + LKM_VALBLK + convert == set lvb */ 155 mlog(0, "will set lvb: converting %s->%s\n", 156 dlm_lock_mode_name(lock->ml.type), 157 dlm_lock_mode_name(type)); 158 lock->lksb->flags |= DLM_LKSB_PUT_LVB; 159 break; 160 case LKM_PRMODE: 161 case LKM_NLMODE: 162 /* refetch if new level is not NL */ 163 if (type > LKM_NLMODE) { 164 mlog(0, "will fetch new value into " 165 "lvb: converting %s->%s\n", 166 dlm_lock_mode_name(lock->ml.type), 167 dlm_lock_mode_name(type)); 168 lock->lksb->flags |= DLM_LKSB_GET_LVB; 169 } else { 170 mlog(0, "will NOT fetch new value " 171 "into lvb: converting %s->%s\n", 172 dlm_lock_mode_name(lock->ml.type), 173 dlm_lock_mode_name(type)); 174 flags &= ~(LKM_VALBLK); 175 } 176 break; 177 } 178 } 179 180 181 /* in-place downconvert? */ 182 if (type <= lock->ml.type) 183 goto grant; 184 185 /* upconvert from here on */ 186 status = DLM_NORMAL; 187 list_for_each_entry(tmplock, &res->granted, list) { 188 if (tmplock == lock) 189 continue; 190 if (!dlm_lock_compatible(tmplock->ml.type, type)) 191 goto switch_queues; 192 } 193 194 list_for_each_entry(tmplock, &res->converting, list) { 195 if (!dlm_lock_compatible(tmplock->ml.type, type)) 196 goto switch_queues; 197 /* existing conversion requests take precedence */ 198 if (!dlm_lock_compatible(tmplock->ml.convert_type, type)) 199 goto switch_queues; 200 } 201 202 /* fall thru to grant */ 203 204grant: 205 mlog(0, "res %.*s, granting %s lock\n", res->lockname.len, 206 res->lockname.name, dlm_lock_mode_name(type)); 207 /* immediately grant the new lock type */ 208 lock->lksb->status = DLM_NORMAL; 209 if (lock->ml.node == dlm->node_num) 210 mlog(0, "doing in-place convert for nonlocal lock\n"); 211 lock->ml.type = type; 212 if (lock->lksb->flags & DLM_LKSB_PUT_LVB) 213 memcpy(res->lvb, lock->lksb->lvb, DLM_LVB_LEN); 214 215 status = DLM_NORMAL; 216 *call_ast = 1; 217 goto unlock_exit; 218 219switch_queues: 220 if (flags & LKM_NOQUEUE) { 221 mlog(0, "failed to convert NOQUEUE lock %.*s from " 222 "%d to %d...\n", res->lockname.len, res->lockname.name, 223 lock->ml.type, type); 224 status = DLM_NOTQUEUED; 225 goto unlock_exit; 226 } 227 mlog(0, "res %.*s, queueing...\n", res->lockname.len, 228 res->lockname.name); 229 230 lock->ml.convert_type = type; 231 /* do not alter lock refcount. switching lists. */ 232 list_move_tail(&lock->list, &res->converting); 233 234unlock_exit: 235 spin_unlock(&lock->spinlock); 236 if (status == DLM_DENIED) { 237 __dlm_print_one_lock_resource(res); 238 } 239 if (status == DLM_NORMAL) 240 *kick_thread = 1; 241 return status; 242} 243 244void dlm_revert_pending_convert(struct dlm_lock_resource *res, 245 struct dlm_lock *lock) 246{ 247 /* do not alter lock refcount. switching lists. */ 248 list_move_tail(&lock->list, &res->granted); 249 lock->ml.convert_type = LKM_IVMODE; 250 lock->lksb->flags &= ~(DLM_LKSB_GET_LVB|DLM_LKSB_PUT_LVB); 251} 252 253/* messages the master site to do lock conversion 254 * locking: 255 * caller needs: none 256 * taken: takes and drops res->spinlock, uses DLM_LOCK_RES_IN_PROGRESS 257 * held on exit: none 258 * returns: DLM_NORMAL, DLM_RECOVERING, status from remote node 259 */ 260enum dlm_status dlmconvert_remote(struct dlm_ctxt *dlm, 261 struct dlm_lock_resource *res, 262 struct dlm_lock *lock, int flags, int type) 263{ 264 enum dlm_status status; 265 u8 old_owner = res->owner; 266 267 mlog(0, "type=%d, convert_type=%d, busy=%d\n", lock->ml.type, 268 lock->ml.convert_type, res->state & DLM_LOCK_RES_IN_PROGRESS); 269 270 spin_lock(&res->spinlock); 271 if (res->state & DLM_LOCK_RES_RECOVERING) { 272 mlog(0, "bailing out early since res is RECOVERING " 273 "on secondary queue\n"); 274 /* __dlm_print_one_lock_resource(res); */ 275 status = DLM_RECOVERING; 276 goto bail; 277 } 278 /* will exit this call with spinlock held */ 279 __dlm_wait_on_lockres(res); 280 281 if (lock->ml.convert_type != LKM_IVMODE) { 282 __dlm_print_one_lock_resource(res); 283 mlog(ML_ERROR, "converting a remote lock that is already " 284 "converting! (cookie=%u:%llu, conv=%d)\n", 285 dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)), 286 dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)), 287 lock->ml.convert_type); 288 status = DLM_DENIED; 289 goto bail; 290 } 291 292 if (lock->ml.type == type && lock->ml.convert_type == LKM_IVMODE) { 293 mlog(0, "last convert request returned DLM_RECOVERING, but " 294 "owner has already queued and sent ast to me. res %.*s, " 295 "(cookie=%u:%llu, type=%d, conv=%d)\n", 296 res->lockname.len, res->lockname.name, 297 dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)), 298 dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)), 299 lock->ml.type, lock->ml.convert_type); 300 status = DLM_NORMAL; 301 goto bail; 302 } 303 304 res->state |= DLM_LOCK_RES_IN_PROGRESS; 305 /* move lock to local convert queue */ 306 /* do not alter lock refcount. switching lists. */ 307 list_move_tail(&lock->list, &res->converting); 308 lock->convert_pending = 1; 309 lock->ml.convert_type = type; 310 311 if (flags & LKM_VALBLK) { 312 if (lock->ml.type == LKM_EXMODE) { 313 flags |= LKM_PUT_LVB; 314 lock->lksb->flags |= DLM_LKSB_PUT_LVB; 315 } else { 316 if (lock->ml.convert_type == LKM_NLMODE) 317 flags &= ~LKM_VALBLK; 318 else { 319 flags |= LKM_GET_LVB; 320 lock->lksb->flags |= DLM_LKSB_GET_LVB; 321 } 322 } 323 } 324 spin_unlock(&res->spinlock); 325 326 /* no locks held here. 327 * need to wait for a reply as to whether it got queued or not. */ 328 status = dlm_send_remote_convert_request(dlm, res, lock, flags, type); 329 330 spin_lock(&res->spinlock); 331 res->state &= ~DLM_LOCK_RES_IN_PROGRESS; 332 lock->convert_pending = 0; 333 /* if it failed, move it back to granted queue. 334 * if master returns DLM_NORMAL and then down before sending ast, 335 * it may have already been moved to granted queue, reset to 336 * DLM_RECOVERING and retry convert */ 337 if (status != DLM_NORMAL) { 338 if (status != DLM_NOTQUEUED) 339 dlm_error(status); 340 dlm_revert_pending_convert(res, lock); 341 } else if ((res->state & DLM_LOCK_RES_RECOVERING) || 342 (old_owner != res->owner)) { 343 mlog(0, "res %.*s is in recovering or has been recovered.\n", 344 res->lockname.len, res->lockname.name); 345 status = DLM_RECOVERING; 346 } 347bail: 348 spin_unlock(&res->spinlock); 349 350 /* TODO: should this be a wake_one? */ 351 /* wake up any IN_PROGRESS waiters */ 352 wake_up(&res->wq); 353 354 return status; 355} 356 357/* sends DLM_CONVERT_LOCK_MSG to master site 358 * locking: 359 * caller needs: none 360 * taken: none 361 * held on exit: none 362 * returns: DLM_NOLOCKMGR, status from remote node 363 */ 364static enum dlm_status dlm_send_remote_convert_request(struct dlm_ctxt *dlm, 365 struct dlm_lock_resource *res, 366 struct dlm_lock *lock, int flags, int type) 367{ 368 struct dlm_convert_lock convert; 369 int tmpret; 370 enum dlm_status ret; 371 int status = 0; 372 struct kvec vec[2]; 373 size_t veclen = 1; 374 375 mlog(0, "%.*s\n", res->lockname.len, res->lockname.name); 376 377 memset(&convert, 0, sizeof(struct dlm_convert_lock)); 378 convert.node_idx = dlm->node_num; 379 convert.requested_type = type; 380 convert.cookie = lock->ml.cookie; 381 convert.namelen = res->lockname.len; 382 convert.flags = cpu_to_be32(flags); 383 memcpy(convert.name, res->lockname.name, convert.namelen); 384 385 vec[0].iov_len = sizeof(struct dlm_convert_lock); 386 vec[0].iov_base = &convert; 387 388 if (flags & LKM_PUT_LVB) { 389 /* extra data to send if we are updating lvb */ 390 vec[1].iov_len = DLM_LVB_LEN; 391 vec[1].iov_base = lock->lksb->lvb; 392 veclen++; 393 } 394 395 tmpret = o2net_send_message_vec(DLM_CONVERT_LOCK_MSG, dlm->key, 396 vec, veclen, res->owner, &status); 397 if (tmpret >= 0) { 398 // successfully sent and received 399 ret = status; // this is already a dlm_status 400 if (ret == DLM_RECOVERING) { 401 mlog(0, "node %u returned DLM_RECOVERING from convert " 402 "message!\n", res->owner); 403 } else if (ret == DLM_MIGRATING) { 404 mlog(0, "node %u returned DLM_MIGRATING from convert " 405 "message!\n", res->owner); 406 } else if (ret == DLM_FORWARD) { 407 mlog(0, "node %u returned DLM_FORWARD from convert " 408 "message!\n", res->owner); 409 } else if (ret != DLM_NORMAL && ret != DLM_NOTQUEUED) 410 dlm_error(ret); 411 } else { 412 mlog(ML_ERROR, "Error %d when sending message %u (key 0x%x) to " 413 "node %u\n", tmpret, DLM_CONVERT_LOCK_MSG, dlm->key, 414 res->owner); 415 if (dlm_is_host_down(tmpret)) { 416 /* instead of logging the same network error over 417 * and over, sleep here and wait for the heartbeat 418 * to notice the node is dead. times out after 5s. */ 419 dlm_wait_for_node_death(dlm, res->owner, 420 DLM_NODE_DEATH_WAIT_MAX); 421 ret = DLM_RECOVERING; 422 mlog(0, "node %u died so returning DLM_RECOVERING " 423 "from convert message!\n", res->owner); 424 } else { 425 ret = dlm_err_to_dlm_status(tmpret); 426 } 427 } 428 429 return ret; 430} 431 432/* handler for DLM_CONVERT_LOCK_MSG on master site 433 * locking: 434 * caller needs: none 435 * taken: takes and drop res->spinlock 436 * held on exit: none 437 * returns: DLM_NORMAL, DLM_IVLOCKID, DLM_BADARGS, 438 * status from __dlmconvert_master 439 */ 440int dlm_convert_lock_handler(struct o2net_msg *msg, u32 len, void *data, 441 void **ret_data) 442{ 443 struct dlm_ctxt *dlm = data; 444 struct dlm_convert_lock *cnv = (struct dlm_convert_lock *)msg->buf; 445 struct dlm_lock_resource *res = NULL; 446 struct dlm_lock *lock = NULL; 447 struct dlm_lock *tmp_lock; 448 struct dlm_lockstatus *lksb; 449 enum dlm_status status = DLM_NORMAL; 450 u32 flags; 451 int call_ast = 0, kick_thread = 0, ast_reserved = 0, wake = 0; 452 453 if (!dlm_grab(dlm)) { 454 dlm_error(DLM_REJECTED); 455 return DLM_REJECTED; 456 } 457 458 mlog_bug_on_msg(!dlm_domain_fully_joined(dlm), 459 "Domain %s not fully joined!\n", dlm->name); 460 461 if (cnv->namelen > DLM_LOCKID_NAME_MAX) { 462 status = DLM_IVBUFLEN; 463 dlm_error(status); 464 goto leave; 465 } 466 467 flags = be32_to_cpu(cnv->flags); 468 469 if ((flags & (LKM_PUT_LVB|LKM_GET_LVB)) == 470 (LKM_PUT_LVB|LKM_GET_LVB)) { 471 mlog(ML_ERROR, "both PUT and GET lvb specified\n"); 472 status = DLM_BADARGS; 473 goto leave; 474 } 475 476 mlog(0, "lvb: %s\n", flags & LKM_PUT_LVB ? "put lvb" : 477 (flags & LKM_GET_LVB ? "get lvb" : "none")); 478 479 status = DLM_IVLOCKID; 480 res = dlm_lookup_lockres(dlm, cnv->name, cnv->namelen); 481 if (!res) { 482 dlm_error(status); 483 goto leave; 484 } 485 486 spin_lock(&res->spinlock); 487 status = __dlm_lockres_state_to_status(res); 488 if (status != DLM_NORMAL) { 489 spin_unlock(&res->spinlock); 490 dlm_error(status); 491 goto leave; 492 } 493 list_for_each_entry(tmp_lock, &res->granted, list) { 494 if (tmp_lock->ml.cookie == cnv->cookie && 495 tmp_lock->ml.node == cnv->node_idx) { 496 lock = tmp_lock; 497 dlm_lock_get(lock); 498 break; 499 } 500 } 501 spin_unlock(&res->spinlock); 502 if (!lock) { 503 status = DLM_IVLOCKID; 504 mlog(ML_ERROR, "did not find lock to convert on grant queue! " 505 "cookie=%u:%llu\n", 506 dlm_get_lock_cookie_node(be64_to_cpu(cnv->cookie)), 507 dlm_get_lock_cookie_seq(be64_to_cpu(cnv->cookie))); 508 dlm_print_one_lock_resource(res); 509 goto leave; 510 } 511 512 /* found the lock */ 513 lksb = lock->lksb; 514 515 /* see if caller needed to get/put lvb */ 516 if (flags & LKM_PUT_LVB) { 517 BUG_ON(lksb->flags & (DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB)); 518 lksb->flags |= DLM_LKSB_PUT_LVB; 519 memcpy(&lksb->lvb[0], &cnv->lvb[0], DLM_LVB_LEN); 520 } else if (flags & LKM_GET_LVB) { 521 BUG_ON(lksb->flags & (DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB)); 522 lksb->flags |= DLM_LKSB_GET_LVB; 523 } 524 525 spin_lock(&res->spinlock); 526 status = __dlm_lockres_state_to_status(res); 527 if (status == DLM_NORMAL) { 528 __dlm_lockres_reserve_ast(res); 529 ast_reserved = 1; 530 res->state |= DLM_LOCK_RES_IN_PROGRESS; 531 status = __dlmconvert_master(dlm, res, lock, flags, 532 cnv->requested_type, 533 &call_ast, &kick_thread); 534 res->state &= ~DLM_LOCK_RES_IN_PROGRESS; 535 wake = 1; 536 } 537 spin_unlock(&res->spinlock); 538 if (wake) 539 wake_up(&res->wq); 540 541 if (status != DLM_NORMAL) { 542 if (status != DLM_NOTQUEUED) 543 dlm_error(status); 544 lksb->flags &= ~(DLM_LKSB_GET_LVB|DLM_LKSB_PUT_LVB); 545 } 546 547leave: 548 if (lock) 549 dlm_lock_put(lock); 550 551 /* either queue the ast or release it, if reserved */ 552 if (call_ast) 553 dlm_queue_ast(dlm, lock); 554 else if (ast_reserved) 555 dlm_lockres_release_ast(dlm, res); 556 557 if (kick_thread) 558 dlm_kick_thread(dlm, res); 559 560 if (res) 561 dlm_lockres_put(res); 562 563 dlm_put(dlm); 564 565 return status; 566} 567