1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* -*- mode: c; c-basic-offset: 8; -*- 3 * vim: noexpandtab sw=8 ts=8 sts=0: 4 * 5 * dlmast.c 6 * 7 * AST and BAST functionality for local and remote nodes 8 * 9 * Copyright (C) 2004 Oracle. All rights reserved. 10 */ 11 12 13 #include <linux/module.h> 14 #include <linux/fs.h> 15 #include <linux/types.h> 16 #include <linux/highmem.h> 17 #include <linux/init.h> 18 #include <linux/sysctl.h> 19 #include <linux/random.h> 20 #include <linux/blkdev.h> 21 #include <linux/socket.h> 22 #include <linux/inet.h> 23 #include <linux/spinlock.h> 24 25 26 #include "../cluster/heartbeat.h" 27 #include "../cluster/nodemanager.h" 28 #include "../cluster/tcp.h" 29 30 #include "dlmapi.h" 31 #include "dlmcommon.h" 32 33 #define MLOG_MASK_PREFIX ML_DLM 34 #include "../cluster/masklog.h" 35 36 static void dlm_update_lvb(struct dlm_ctxt *dlm, struct dlm_lock_resource *res, 37 struct dlm_lock *lock); 38 static int dlm_should_cancel_bast(struct dlm_ctxt *dlm, struct dlm_lock *lock); 39 40 /* Should be called as an ast gets queued to see if the new 41 * lock level will obsolete a pending bast. 42 * For example, if dlm_thread queued a bast for an EX lock that 43 * was blocking another EX, but before sending the bast the 44 * lock owner downconverted to NL, the bast is now obsolete. 45 * Only the ast should be sent. 46 * This is needed because the lock and convert paths can queue 47 * asts out-of-band (not waiting for dlm_thread) in order to 48 * allow for LKM_NOQUEUE to get immediate responses. */ 49 static int dlm_should_cancel_bast(struct dlm_ctxt *dlm, struct dlm_lock *lock) 50 { 51 assert_spin_locked(&dlm->ast_lock); 52 assert_spin_locked(&lock->spinlock); 53 54 if (lock->ml.highest_blocked == LKM_IVMODE) 55 return 0; 56 BUG_ON(lock->ml.highest_blocked == LKM_NLMODE); 57 58 if (lock->bast_pending && 59 list_empty(&lock->bast_list)) 60 /* old bast already sent, ok */ 61 return 0; 62 63 if (lock->ml.type == LKM_EXMODE) 64 /* EX blocks anything left, any bast still valid */ 65 return 0; 66 else if (lock->ml.type == LKM_NLMODE) 67 /* NL blocks nothing, no reason to send any bast, cancel it */ 68 return 1; 69 else if (lock->ml.highest_blocked != LKM_EXMODE) 70 /* PR only blocks EX */ 71 return 1; 72 73 return 0; 74 } 75 76 void __dlm_queue_ast(struct dlm_ctxt *dlm, struct dlm_lock *lock) 77 { 78 struct dlm_lock_resource *res; 79 80 BUG_ON(!dlm); 81 BUG_ON(!lock); 82 83 res = lock->lockres; 84 85 assert_spin_locked(&dlm->ast_lock); 86 87 if (!list_empty(&lock->ast_list)) { 88 mlog(ML_ERROR, "%s: res %.*s, lock %u:%llu, " 89 "AST list not empty, pending %d, newlevel %d\n", 90 dlm->name, res->lockname.len, res->lockname.name, 91 dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)), 92 dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)), 93 lock->ast_pending, lock->ml.type); 94 BUG(); 95 } 96 if (lock->ast_pending) 97 mlog(0, "%s: res %.*s, lock %u:%llu, AST getting flushed\n", 98 dlm->name, res->lockname.len, res->lockname.name, 99 dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)), 100 dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie))); 101 102 /* putting lock on list, add a ref */ 103 dlm_lock_get(lock); 104 spin_lock(&lock->spinlock); 105 106 /* check to see if this ast obsoletes the bast */ 107 if (dlm_should_cancel_bast(dlm, lock)) { 108 mlog(0, "%s: res %.*s, lock %u:%llu, Cancelling BAST\n", 109 dlm->name, res->lockname.len, res->lockname.name, 110 dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)), 111 dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie))); 112 lock->bast_pending = 0; 113 list_del_init(&lock->bast_list); 114 lock->ml.highest_blocked = LKM_IVMODE; 115 /* removing lock from list, remove a ref. guaranteed 116 * this won't be the last ref because of the get above, 117 * so res->spinlock will not be taken here */ 118 dlm_lock_put(lock); 119 /* free up the reserved bast that we are cancelling. 120 * guaranteed that this will not be the last reserved 121 * ast because *both* an ast and a bast were reserved 122 * to get to this point. the res->spinlock will not be 123 * taken here */ 124 dlm_lockres_release_ast(dlm, res); 125 } 126 list_add_tail(&lock->ast_list, &dlm->pending_asts); 127 lock->ast_pending = 1; 128 spin_unlock(&lock->spinlock); 129 } 130 131 void dlm_queue_ast(struct dlm_ctxt *dlm, struct dlm_lock *lock) 132 { 133 BUG_ON(!dlm); 134 BUG_ON(!lock); 135 136 spin_lock(&dlm->ast_lock); 137 __dlm_queue_ast(dlm, lock); 138 spin_unlock(&dlm->ast_lock); 139 } 140 141 142 void __dlm_queue_bast(struct dlm_ctxt *dlm, struct dlm_lock *lock) 143 { 144 struct dlm_lock_resource *res; 145 146 BUG_ON(!dlm); 147 BUG_ON(!lock); 148 149 assert_spin_locked(&dlm->ast_lock); 150 151 res = lock->lockres; 152 153 BUG_ON(!list_empty(&lock->bast_list)); 154 if (lock->bast_pending) 155 mlog(0, "%s: res %.*s, lock %u:%llu, BAST getting flushed\n", 156 dlm->name, res->lockname.len, res->lockname.name, 157 dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)), 158 dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie))); 159 160 /* putting lock on list, add a ref */ 161 dlm_lock_get(lock); 162 spin_lock(&lock->spinlock); 163 list_add_tail(&lock->bast_list, &dlm->pending_basts); 164 lock->bast_pending = 1; 165 spin_unlock(&lock->spinlock); 166 } 167 168 static void dlm_update_lvb(struct dlm_ctxt *dlm, struct dlm_lock_resource *res, 169 struct dlm_lock *lock) 170 { 171 struct dlm_lockstatus *lksb = lock->lksb; 172 BUG_ON(!lksb); 173 174 /* only updates if this node masters the lockres */ 175 spin_lock(&res->spinlock); 176 if (res->owner == dlm->node_num) { 177 /* check the lksb flags for the direction */ 178 if (lksb->flags & DLM_LKSB_GET_LVB) { 179 mlog(0, "getting lvb from lockres for %s node\n", 180 lock->ml.node == dlm->node_num ? "master" : 181 "remote"); 182 memcpy(lksb->lvb, res->lvb, DLM_LVB_LEN); 183 } 184 /* Do nothing for lvb put requests - they should be done in 185 * place when the lock is downconverted - otherwise we risk 186 * racing gets and puts which could result in old lvb data 187 * being propagated. We leave the put flag set and clear it 188 * here. In the future we might want to clear it at the time 189 * the put is actually done. 190 */ 191 } 192 spin_unlock(&res->spinlock); 193 194 /* reset any lvb flags on the lksb */ 195 lksb->flags &= ~(DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB); 196 } 197 198 void dlm_do_local_ast(struct dlm_ctxt *dlm, struct dlm_lock_resource *res, 199 struct dlm_lock *lock) 200 { 201 dlm_astlockfunc_t *fn; 202 203 mlog(0, "%s: res %.*s, lock %u:%llu, Local AST\n", dlm->name, 204 res->lockname.len, res->lockname.name, 205 dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)), 206 dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie))); 207 208 fn = lock->ast; 209 BUG_ON(lock->ml.node != dlm->node_num); 210 211 dlm_update_lvb(dlm, res, lock); 212 (*fn)(lock->astdata); 213 } 214 215 216 int dlm_do_remote_ast(struct dlm_ctxt *dlm, struct dlm_lock_resource *res, 217 struct dlm_lock *lock) 218 { 219 int ret; 220 struct dlm_lockstatus *lksb; 221 int lksbflags; 222 223 mlog(0, "%s: res %.*s, lock %u:%llu, Remote AST\n", dlm->name, 224 res->lockname.len, res->lockname.name, 225 dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)), 226 dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie))); 227 228 lksb = lock->lksb; 229 BUG_ON(lock->ml.node == dlm->node_num); 230 231 lksbflags = lksb->flags; 232 dlm_update_lvb(dlm, res, lock); 233 234 /* lock request came from another node 235 * go do the ast over there */ 236 ret = dlm_send_proxy_ast(dlm, res, lock, lksbflags); 237 return ret; 238 } 239 240 void dlm_do_local_bast(struct dlm_ctxt *dlm, struct dlm_lock_resource *res, 241 struct dlm_lock *lock, int blocked_type) 242 { 243 dlm_bastlockfunc_t *fn = lock->bast; 244 245 BUG_ON(lock->ml.node != dlm->node_num); 246 247 mlog(0, "%s: res %.*s, lock %u:%llu, Local BAST, blocked %d\n", 248 dlm->name, res->lockname.len, res->lockname.name, 249 dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)), 250 dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)), 251 blocked_type); 252 253 (*fn)(lock->astdata, blocked_type); 254 } 255 256 257 258 int dlm_proxy_ast_handler(struct o2net_msg *msg, u32 len, void *data, 259 void **ret_data) 260 { 261 int ret; 262 unsigned int locklen; 263 struct dlm_ctxt *dlm = data; 264 struct dlm_lock_resource *res = NULL; 265 struct dlm_lock *lock = NULL; 266 struct dlm_proxy_ast *past = (struct dlm_proxy_ast *) msg->buf; 267 char *name; 268 struct list_head *head = NULL; 269 __be64 cookie; 270 u32 flags; 271 u8 node; 272 273 if (!dlm_grab(dlm)) { 274 dlm_error(DLM_REJECTED); 275 return DLM_REJECTED; 276 } 277 278 mlog_bug_on_msg(!dlm_domain_fully_joined(dlm), 279 "Domain %s not fully joined!\n", dlm->name); 280 281 name = past->name; 282 locklen = past->namelen; 283 cookie = past->cookie; 284 flags = be32_to_cpu(past->flags); 285 node = past->node_idx; 286 287 if (locklen > DLM_LOCKID_NAME_MAX) { 288 ret = DLM_IVBUFLEN; 289 mlog(ML_ERROR, "Invalid name length (%d) in proxy ast " 290 "handler!\n", locklen); 291 goto leave; 292 } 293 294 if ((flags & (LKM_PUT_LVB|LKM_GET_LVB)) == 295 (LKM_PUT_LVB|LKM_GET_LVB)) { 296 mlog(ML_ERROR, "Both PUT and GET lvb specified, (0x%x)\n", 297 flags); 298 ret = DLM_BADARGS; 299 goto leave; 300 } 301 302 mlog(0, "lvb: %s\n", flags & LKM_PUT_LVB ? "put lvb" : 303 (flags & LKM_GET_LVB ? "get lvb" : "none")); 304 305 mlog(0, "type=%d, blocked_type=%d\n", past->type, past->blocked_type); 306 307 if (past->type != DLM_AST && 308 past->type != DLM_BAST) { 309 mlog(ML_ERROR, "Unknown ast type! %d, cookie=%u:%llu" 310 "name=%.*s, node=%u\n", past->type, 311 dlm_get_lock_cookie_node(be64_to_cpu(cookie)), 312 dlm_get_lock_cookie_seq(be64_to_cpu(cookie)), 313 locklen, name, node); 314 ret = DLM_IVLOCKID; 315 goto leave; 316 } 317 318 res = dlm_lookup_lockres(dlm, name, locklen); 319 if (!res) { 320 mlog(0, "Got %sast for unknown lockres! cookie=%u:%llu, " 321 "name=%.*s, node=%u\n", (past->type == DLM_AST ? "" : "b"), 322 dlm_get_lock_cookie_node(be64_to_cpu(cookie)), 323 dlm_get_lock_cookie_seq(be64_to_cpu(cookie)), 324 locklen, name, node); 325 ret = DLM_IVLOCKID; 326 goto leave; 327 } 328 329 /* cannot get a proxy ast message if this node owns it */ 330 BUG_ON(res->owner == dlm->node_num); 331 332 mlog(0, "%s: res %.*s\n", dlm->name, res->lockname.len, 333 res->lockname.name); 334 335 spin_lock(&res->spinlock); 336 if (res->state & DLM_LOCK_RES_RECOVERING) { 337 mlog(0, "Responding with DLM_RECOVERING!\n"); 338 ret = DLM_RECOVERING; 339 goto unlock_out; 340 } 341 if (res->state & DLM_LOCK_RES_MIGRATING) { 342 mlog(0, "Responding with DLM_MIGRATING!\n"); 343 ret = DLM_MIGRATING; 344 goto unlock_out; 345 } 346 /* try convert queue for both ast/bast */ 347 head = &res->converting; 348 lock = NULL; 349 list_for_each_entry(lock, head, list) { 350 if (lock->ml.cookie == cookie) 351 goto do_ast; 352 } 353 354 /* if not on convert, try blocked for ast, granted for bast */ 355 if (past->type == DLM_AST) 356 head = &res->blocked; 357 else 358 head = &res->granted; 359 360 list_for_each_entry(lock, head, list) { 361 /* if lock is found but unlock is pending ignore the bast */ 362 if (lock->ml.cookie == cookie) { 363 if (lock->unlock_pending) 364 break; 365 goto do_ast; 366 } 367 } 368 369 mlog(0, "Got %sast for unknown lock! cookie=%u:%llu, name=%.*s, " 370 "node=%u\n", past->type == DLM_AST ? "" : "b", 371 dlm_get_lock_cookie_node(be64_to_cpu(cookie)), 372 dlm_get_lock_cookie_seq(be64_to_cpu(cookie)), 373 locklen, name, node); 374 375 ret = DLM_NORMAL; 376 unlock_out: 377 spin_unlock(&res->spinlock); 378 goto leave; 379 380 do_ast: 381 ret = DLM_NORMAL; 382 if (past->type == DLM_AST) { 383 /* do not alter lock refcount. switching lists. */ 384 list_move_tail(&lock->list, &res->granted); 385 mlog(0, "%s: res %.*s, lock %u:%llu, Granted type %d => %d\n", 386 dlm->name, res->lockname.len, res->lockname.name, 387 dlm_get_lock_cookie_node(be64_to_cpu(cookie)), 388 dlm_get_lock_cookie_seq(be64_to_cpu(cookie)), 389 lock->ml.type, lock->ml.convert_type); 390 391 if (lock->ml.convert_type != LKM_IVMODE) { 392 lock->ml.type = lock->ml.convert_type; 393 lock->ml.convert_type = LKM_IVMODE; 394 } else { 395 // should already be there.... 396 } 397 398 lock->lksb->status = DLM_NORMAL; 399 400 /* if we requested the lvb, fetch it into our lksb now */ 401 if (flags & LKM_GET_LVB) { 402 BUG_ON(!(lock->lksb->flags & DLM_LKSB_GET_LVB)); 403 memcpy(lock->lksb->lvb, past->lvb, DLM_LVB_LEN); 404 } 405 } 406 spin_unlock(&res->spinlock); 407 408 if (past->type == DLM_AST) 409 dlm_do_local_ast(dlm, res, lock); 410 else 411 dlm_do_local_bast(dlm, res, lock, past->blocked_type); 412 413 leave: 414 if (res) 415 dlm_lockres_put(res); 416 417 dlm_put(dlm); 418 return ret; 419 } 420 421 422 423 int dlm_send_proxy_ast_msg(struct dlm_ctxt *dlm, struct dlm_lock_resource *res, 424 struct dlm_lock *lock, int msg_type, 425 int blocked_type, int flags) 426 { 427 int ret = 0; 428 struct dlm_proxy_ast past; 429 struct kvec vec[2]; 430 size_t veclen = 1; 431 int status; 432 433 mlog(0, "%s: res %.*s, to %u, type %d, blocked_type %d\n", dlm->name, 434 res->lockname.len, res->lockname.name, lock->ml.node, msg_type, 435 blocked_type); 436 437 memset(&past, 0, sizeof(struct dlm_proxy_ast)); 438 past.node_idx = dlm->node_num; 439 past.type = msg_type; 440 past.blocked_type = blocked_type; 441 past.namelen = res->lockname.len; 442 memcpy(past.name, res->lockname.name, past.namelen); 443 past.cookie = lock->ml.cookie; 444 445 vec[0].iov_len = sizeof(struct dlm_proxy_ast); 446 vec[0].iov_base = &past; 447 if (flags & DLM_LKSB_GET_LVB) { 448 be32_add_cpu(&past.flags, LKM_GET_LVB); 449 vec[1].iov_len = DLM_LVB_LEN; 450 vec[1].iov_base = lock->lksb->lvb; 451 veclen++; 452 } 453 454 ret = o2net_send_message_vec(DLM_PROXY_AST_MSG, dlm->key, vec, veclen, 455 lock->ml.node, &status); 456 if (ret < 0) 457 mlog(ML_ERROR, "%s: res %.*s, error %d send AST to node %u\n", 458 dlm->name, res->lockname.len, res->lockname.name, ret, 459 lock->ml.node); 460 else { 461 if (status == DLM_RECOVERING) { 462 mlog(ML_ERROR, "sent AST to node %u, it thinks this " 463 "node is dead!\n", lock->ml.node); 464 BUG(); 465 } else if (status == DLM_MIGRATING) { 466 mlog(ML_ERROR, "sent AST to node %u, it returned " 467 "DLM_MIGRATING!\n", lock->ml.node); 468 BUG(); 469 } else if (status != DLM_NORMAL && status != DLM_IVLOCKID) { 470 mlog(ML_ERROR, "AST to node %u returned %d!\n", 471 lock->ml.node, status); 472 /* ignore it */ 473 } 474 ret = 0; 475 } 476 return ret; 477 } 478