fs/ocfs2/dlm/dlmast.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * dlmast.c
   5  *
   6  * AST and BAST functionality for local and remote nodes
   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 #define MLOG_MASK_PREFIX ML_DLM
  49 #include "cluster/masklog.h"
  50
  51 static void dlm_update_lvb(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
  52                            struct dlm_lock *lock);
  53 static int dlm_should_cancel_bast(struct dlm_ctxt *dlm, struct dlm_lock *lock);
  54
  55 /* Should be called as an ast gets queued to see if the new
  56  * lock level will obsolete a pending bast.
  57  * For example, if dlm_thread queued a bast for an EX lock that
  58  * was blocking another EX, but before sending the bast the
  59  * lock owner downconverted to NL, the bast is now obsolete.
  60  * Only the ast should be sent.
  61  * This is needed because the lock and convert paths can queue
  62  * asts out-of-band (not waiting for dlm_thread) in order to
  63  * allow for LKM_NOQUEUE to get immediate responses. */
  64 static int dlm_should_cancel_bast(struct dlm_ctxt *dlm, struct dlm_lock *lock)
  65 {
  66         assert_spin_locked(&dlm->ast_lock);
  67         assert_spin_locked(&lock->spinlock);
  68
  69         if (lock->ml.highest_blocked == LKM_IVMODE)
  70                 return 0;
  71         BUG_ON(lock->ml.highest_blocked == LKM_NLMODE);
  72
  73         if (lock->bast_pending &&
  74             list_empty(&lock->bast_list))
  75                 /* old bast already sent, ok */
  76                 return 0;
  77
  78         if (lock->ml.type == LKM_EXMODE)
  79                 /* EX blocks anything left, any bast still valid */
  80                 return 0;
  81         else if (lock->ml.type == LKM_NLMODE)
  82                 /* NL blocks nothing, no reason to send any bast, cancel it */
  83                 return 1;
  84         else if (lock->ml.highest_blocked != LKM_EXMODE)
  85                 /* PR only blocks EX */
  86                 return 1;
  87
  88         return 0;
  89 }
  90
  91 static void __dlm_queue_ast(struct dlm_ctxt *dlm, struct dlm_lock *lock)
  92 {
  93         mlog_entry_void();
  94
  95         BUG_ON(!dlm);
  96         BUG_ON(!lock);
  97
  98         assert_spin_locked(&dlm->ast_lock);
  99         if (!list_empty(&lock->ast_list)) {
 100                 mlog(ML_ERROR, "ast list not empty!!  pending=%d, newlevel=%d\n",
 101                      lock->ast_pending, lock->ml.type);
 102                 BUG();
 103         }
 104         if (lock->ast_pending)
 105                 mlog(0, "lock has an ast getting flushed right now\n");
 106
 107         /* putting lock on list, add a ref */
 108         dlm_lock_get(lock);
 109         spin_lock(&lock->spinlock);
 110
 111         /* check to see if this ast obsoletes the bast */
 112         if (dlm_should_cancel_bast(dlm, lock)) {
 113                 struct dlm_lock_resource *res = lock->lockres;
 114                 mlog(0, "%s: cancelling bast for %.*s\n",
 115                      dlm->name, res->lockname.len, res->lockname.name);
 116                 lock->bast_pending = 0;
 117                 list_del_init(&lock->bast_list);
 118                 lock->ml.highest_blocked = LKM_IVMODE;
 119                 /* removing lock from list, remove a ref.  guaranteed
 120                  * this won't be the last ref because of the get above,
 121                  * so res->spinlock will not be taken here */
 122                 dlm_lock_put(lock);
 123                 /* free up the reserved bast that we are cancelling.
 124                  * guaranteed that this will not be the last reserved
 125                  * ast because *both* an ast and a bast were reserved
 126                  * to get to this point.  the res->spinlock will not be
 127                  * taken here */
 128                 dlm_lockres_release_ast(dlm, res);
 129         }
 130         list_add_tail(&lock->ast_list, &dlm->pending_asts);
 131         lock->ast_pending = 1;
 132         spin_unlock(&lock->spinlock);
 133 }
 134
 135 void dlm_queue_ast(struct dlm_ctxt *dlm, struct dlm_lock *lock)
 136 {
 137         mlog_entry_void();
 138
 139         BUG_ON(!dlm);
 140         BUG_ON(!lock);
 141
 142         spin_lock(&dlm->ast_lock);
 143         __dlm_queue_ast(dlm, lock);
 144         spin_unlock(&dlm->ast_lock);
 145 }
 146
 147
 148 static void __dlm_queue_bast(struct dlm_ctxt *dlm, struct dlm_lock *lock)
 149 {
 150         mlog_entry_void();
 151
 152         BUG_ON(!dlm);
 153         BUG_ON(!lock);
 154         assert_spin_locked(&dlm->ast_lock);
 155
 156         BUG_ON(!list_empty(&lock->bast_list));
 157         if (lock->bast_pending)
 158                 mlog(0, "lock has a bast getting flushed right now\n");
 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 void dlm_queue_bast(struct dlm_ctxt *dlm, struct dlm_lock *lock)
 169 {
 170         mlog_entry_void();
 171
 172         BUG_ON(!dlm);
 173         BUG_ON(!lock);
 174
 175         spin_lock(&dlm->ast_lock);
 176         __dlm_queue_bast(dlm, lock);
 177         spin_unlock(&dlm->ast_lock);
 178 }
 179
 180 static void dlm_update_lvb(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
 181                            struct dlm_lock *lock)
 182 {
 183         struct dlm_lockstatus *lksb = lock->lksb;
 184         BUG_ON(!lksb);
 185
 186         /* only updates if this node masters the lockres */
 187         if (res->owner == dlm->node_num) {
 188
 189                 spin_lock(&res->spinlock);
 190                 /* check the lksb flags for the direction */
 191                 if (lksb->flags & DLM_LKSB_GET_LVB) {
 192                         mlog(0, "getting lvb from lockres for %s node\n",
 193                                   lock->ml.node == dlm->node_num ? "master" :
 194                                   "remote");
 195                         memcpy(lksb->lvb, res->lvb, DLM_LVB_LEN);
 196                 }
 197                 /* Do nothing for lvb put requests - they should be done in
 198                  * place when the lock is downconverted - otherwise we risk
 199                  * racing gets and puts which could result in old lvb data
 200                  * being propagated. We leave the put flag set and clear it
 201                  * here. In the future we might want to clear it at the time
 202                  * the put is actually done.
 203                  */
 204                 spin_unlock(&res->spinlock);
 205         }
 206
 207         /* reset any lvb flags on the lksb */
 208         lksb->flags &= ~(DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB);
 209 }
 210
 211 void dlm_do_local_ast(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
 212                       struct dlm_lock *lock)
 213 {
 214         dlm_astlockfunc_t *fn;
 215         struct dlm_lockstatus *lksb;
 216
 217         mlog_entry_void();
 218
 219         lksb = lock->lksb;
 220         fn = lock->ast;
 221         BUG_ON(lock->ml.node != dlm->node_num);
 222
 223         dlm_update_lvb(dlm, res, lock);
 224         (*fn)(lock->astdata);
 225 }
 226
 227
 228 int dlm_do_remote_ast(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
 229                       struct dlm_lock *lock)
 230 {
 231         int ret;
 232         struct dlm_lockstatus *lksb;
 233         int lksbflags;
 234
 235         mlog_entry_void();
 236
 237         lksb = lock->lksb;
 238         BUG_ON(lock->ml.node == dlm->node_num);
 239
 240         lksbflags = lksb->flags;
 241         dlm_update_lvb(dlm, res, lock);
 242
 243         /* lock request came from another node
 244          * go do the ast over there */
 245         ret = dlm_send_proxy_ast(dlm, res, lock, lksbflags);
 246         return ret;
 247 }
 248
 249 void dlm_do_local_bast(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
 250                        struct dlm_lock *lock, int blocked_type)
 251 {
 252         dlm_bastlockfunc_t *fn = lock->bast;
 253
 254         mlog_entry_void();
 255         BUG_ON(lock->ml.node != dlm->node_num);
 256
 257         (*fn)(lock->astdata, blocked_type);
 258 }
 259
 260
 261
 262 int dlm_proxy_ast_handler(struct o2net_msg *msg, u32 len, void *data,
 263                           void **ret_data)
 264 {
 265         int ret;
 266         unsigned int locklen;
 267         struct dlm_ctxt *dlm = data;
 268         struct dlm_lock_resource *res = NULL;
 269         struct dlm_lock *lock = NULL;
 270         struct dlm_proxy_ast *past = (struct dlm_proxy_ast *) msg->buf;
 271         char *name;
 272         struct list_head *iter, *head=NULL;
 273         u64 cookie;
 274         u32 flags;
 275         u8 node;
 276
 277         if (!dlm_grab(dlm)) {
 278                 dlm_error(DLM_REJECTED);
 279                 return DLM_REJECTED;
 280         }
 281
 282         mlog_bug_on_msg(!dlm_domain_fully_joined(dlm),
 283                         "Domain %s not fully joined!\n", dlm->name);
 284
 285         name = past->name;
 286         locklen = past->namelen;
 287         cookie = past->cookie;
 288         flags = be32_to_cpu(past->flags);
 289         node = past->node_idx;
 290
 291         if (locklen > DLM_LOCKID_NAME_MAX) {
 292                 ret = DLM_IVBUFLEN;
 293                 mlog(ML_ERROR, "Invalid name length (%d) in proxy ast "
 294                      "handler!\n", locklen);
 295                 goto leave;
 296         }
 297
 298         if ((flags & (LKM_PUT_LVB|LKM_GET_LVB)) ==
 299              (LKM_PUT_LVB|LKM_GET_LVB)) {
 300                 mlog(ML_ERROR, "Both PUT and GET lvb specified, (0x%x)\n",
 301                      flags);
 302                 ret = DLM_BADARGS;
 303                 goto leave;
 304         }
 305
 306         mlog(0, "lvb: %s\n", flags & LKM_PUT_LVB ? "put lvb" :
 307                   (flags & LKM_GET_LVB ? "get lvb" : "none"));
 308
 309         mlog(0, "type=%d, blocked_type=%d\n", past->type, past->blocked_type);
 310
 311         if (past->type != DLM_AST &&
 312             past->type != DLM_BAST) {
 313                 mlog(ML_ERROR, "Unknown ast type! %d, cookie=%u:%llu"
 314                      "name=%.*s, node=%u\n", past->type,
 315                      dlm_get_lock_cookie_node(be64_to_cpu(cookie)),
 316                      dlm_get_lock_cookie_seq(be64_to_cpu(cookie)),
 317                      locklen, name, node);
 318                 ret = DLM_IVLOCKID;
 319                 goto leave;
 320         }
 321
 322         res = dlm_lookup_lockres(dlm, name, locklen);
 323         if (!res) {
 324                 mlog(0, "Got %sast for unknown lockres! cookie=%u:%llu, "
 325                      "name=%.*s, node=%u\n", (past->type == DLM_AST ? "" : "b"),
 326                      dlm_get_lock_cookie_node(be64_to_cpu(cookie)),
 327                      dlm_get_lock_cookie_seq(be64_to_cpu(cookie)),
 328                      locklen, name, node);
 329                 ret = DLM_IVLOCKID;
 330                 goto leave;
 331         }
 332
 333         /* cannot get a proxy ast message if this node owns it */
 334         BUG_ON(res->owner == dlm->node_num);
 335
 336         mlog(0, "lockres %.*s\n", res->lockname.len, res->lockname.name);
 337
 338         spin_lock(&res->spinlock);
 339         if (res->state & DLM_LOCK_RES_RECOVERING) {
 340                 mlog(0, "Responding with DLM_RECOVERING!\n");
 341                 ret = DLM_RECOVERING;
 342                 goto unlock_out;
 343         }
 344         if (res->state & DLM_LOCK_RES_MIGRATING) {
 345                 mlog(0, "Responding with DLM_MIGRATING!\n");
 346                 ret = DLM_MIGRATING;
 347                 goto unlock_out;
 348         }
 349         /* try convert queue for both ast/bast */
 350         head = &res->converting;
 351         lock = NULL;
 352         list_for_each(iter, head) {
 353                 lock = list_entry (iter, struct dlm_lock, list);
 354                 if (lock->ml.cookie == cookie)
 355                         goto do_ast;
 356         }
 357
 358         /* if not on convert, try blocked for ast, granted for bast */
 359         if (past->type == DLM_AST)
 360                 head = &res->blocked;
 361         else
 362                 head = &res->granted;
 363
 364         list_for_each(iter, head) {
 365                 lock = list_entry (iter, struct dlm_lock, list);
 366                 if (lock->ml.cookie == cookie)
 367                         goto do_ast;
 368         }
 369
 370         mlog(0, "Got %sast for unknown lock! cookie=%u:%llu, name=%.*s, "
 371              "node=%u\n", past->type == DLM_AST ? "" : "b",
 372              dlm_get_lock_cookie_node(be64_to_cpu(cookie)),
 373              dlm_get_lock_cookie_seq(be64_to_cpu(cookie)),
 374              locklen, name, node);
 375
 376         ret = DLM_NORMAL;
 377 unlock_out:
 378         spin_unlock(&res->spinlock);
 379         goto leave;
 380
 381 do_ast:
 382         ret = DLM_NORMAL;
 383         if (past->type == DLM_AST) {
 384                 /* do not alter lock refcount.  switching lists. */
 385                 list_move_tail(&lock->list, &res->granted);
 386                 mlog(0, "ast: Adding to granted list... type=%d, "
 387                      "convert_type=%d\n", lock->ml.type, lock->ml.convert_type);
 388                 if (lock->ml.convert_type != LKM_IVMODE) {
 389                         lock->ml.type = lock->ml.convert_type;
 390                         lock->ml.convert_type = LKM_IVMODE;
 391                 } else {
 392                         // should already be there....
 393                 }
 394
 395                 lock->lksb->status = DLM_NORMAL;
 396
 397                 /* if we requested the lvb, fetch it into our lksb now */
 398                 if (flags & LKM_GET_LVB) {
 399                         BUG_ON(!(lock->lksb->flags & DLM_LKSB_GET_LVB));
 400                         memcpy(lock->lksb->lvb, past->lvb, DLM_LVB_LEN);
 401                 }
 402         }
 403         spin_unlock(&res->spinlock);
 404
 405         if (past->type == DLM_AST)
 406                 dlm_do_local_ast(dlm, res, lock);
 407         else
 408                 dlm_do_local_bast(dlm, res, lock, past->blocked_type);
 409
 410 leave:
 411         if (res)
 412                 dlm_lockres_put(res);
 413
 414         dlm_put(dlm);
 415         return ret;
 416 }
 417
 418
 419
 420 int dlm_send_proxy_ast_msg(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
 421                            struct dlm_lock *lock, int msg_type,
 422                            int blocked_type, int flags)
 423 {
 424         int ret = 0;
 425         struct dlm_proxy_ast past;
 426         struct kvec vec[2];
 427         size_t veclen = 1;
 428         int status;
 429
 430         mlog_entry("res %.*s, to=%u, type=%d, blocked_type=%d\n",
 431                    res->lockname.len, res->lockname.name, lock->ml.node,
 432                    msg_type, blocked_type);
 433
 434         memset(&past, 0, sizeof(struct dlm_proxy_ast));
 435         past.node_idx = dlm->node_num;
 436         past.type = msg_type;
 437         past.blocked_type = blocked_type;
 438         past.namelen = res->lockname.len;
 439         memcpy(past.name, res->lockname.name, past.namelen);
 440         past.cookie = lock->ml.cookie;
 441
 442         vec[0].iov_len = sizeof(struct dlm_proxy_ast);
 443         vec[0].iov_base = &past;
 444         if (flags & DLM_LKSB_GET_LVB) {
 445                 mlog(0, "returning requested LVB data\n");
 446                 be32_add_cpu(&past.flags, LKM_GET_LVB);
 447                 vec[1].iov_len = DLM_LVB_LEN;
 448                 vec[1].iov_base = lock->lksb->lvb;
 449                 veclen++;
 450         }
 451
 452         ret = o2net_send_message_vec(DLM_PROXY_AST_MSG, dlm->key, vec, veclen,
 453                                      lock->ml.node, &status);
 454         if (ret < 0)
 455                 mlog_errno(ret);
 456         else {
 457                 if (status == DLM_RECOVERING) {
 458                         mlog(ML_ERROR, "sent AST to node %u, it thinks this "
 459                              "node is dead!\n", lock->ml.node);
 460                         BUG();
 461                 } else if (status == DLM_MIGRATING) {
 462                         mlog(ML_ERROR, "sent AST to node %u, it returned "
 463                              "DLM_MIGRATING!\n", lock->ml.node);
 464                         BUG();
 465                 } else if (status != DLM_NORMAL && status != DLM_IVLOCKID) {
 466                         mlog(ML_ERROR, "AST to node %u returned %d!\n",
 467                              lock->ml.node, status);
 468                         /* ignore it */
 469                 }
 470                 ret = 0;
 471         }
 472         return ret;
 473 }