fs/ocfs2/dlm/dlmconvert.c

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