lustre/ldlm/ldlm_request.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (C) 2002, 2003 Cluster File Systems, Inc.
   5  *
   6  *   This file is part of Lustre, http://www.lustre.org.
   7  *
   8  *   Lustre is free software; you can redistribute it and/or
   9  *   modify it under the terms of version 2 of the GNU General Public
  10  *   License as published by the Free Software Foundation.
  11  *
  12  *   Lustre is distributed in the hope that it will be useful,
  13  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  *   GNU General Public License for more details.
  16  *
  17  *   You should have received a copy of the GNU General Public License
  18  *   along with Lustre; if not, write to the Free Software
  19  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  20  */
  21
  22 #define DEBUG_SUBSYSTEM S_LDLM
  23 #ifndef __KERNEL__
  24 #include <signal.h>
  25 #include <liblustre.h>
  26 #endif
  27
  28 #include <linux/lustre_dlm.h>
  29 #include <linux/obd_class.h>
  30 #include <linux/obd.h>
  31
  32 static void interrupted_completion_wait(void *data)
  33 {
  34 }
  35
  36 struct lock_wait_data {
  37         struct ldlm_lock *lwd_lock;
  38         int               lwd_generation;
  39 };
  40
  41 int ldlm_expired_completion_wait(void *data)
  42 {
  43         struct lock_wait_data *lwd = data;
  44         struct ldlm_lock *lock = lwd->lwd_lock;
  45         struct obd_device *obd = class_conn2obd(lock->l_connh);
  46
  47         if (obd == NULL) {
  48                 LDLM_ERROR(lock, "lock timed out; mot entering recovery in "
  49                            "server code, just going back to sleep");
  50         } else {
  51                 struct obd_import *imp = obd->u.cli.cl_import;
  52                 ptlrpc_fail_import(imp, lwd->lwd_generation);
  53                 LDLM_ERROR(lock, "lock timed out, entering recovery for %s@%s",
  54                            imp->imp_target_uuid.uuid,
  55                            imp->imp_connection->c_remote_uuid.uuid);
  56         }
  57
  58         RETURN(0);
  59 }
  60
  61 int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data)
  62 {
  63         /* XXX ALLOCATE - 160 mytes */
  64         struct lock_wait_data lwd;
  65         unsigned long irqflags;
  66         struct obd_device *obd;
  67         struct obd_import *imp = NULL;
  68         int rc = 0;
  69         struct l_wait_info lwi;
  70
  71         obd = class_conn2obd(lock->l_connh);
  72
  73         /* if this is a local lock, then there is no import */
  74         if (obd != NULL)
  75                 imp = obd->u.cli.cl_import;
  76
  77         lwd.lwd_lock = lock;
  78
  79         lwi = LWI_TIMEOUT_INTR(obd_timeout * HZ, ldlm_expired_completion_wait,
  80                                interrupted_completion_wait, &lwd);
  81         ENTRY;
  82
  83         if (flags == LDLM_FL_WAIT_NOREPROC)
  84                 goto noreproc;
  85
  86         if (flags == 0) {
  87                 wake_up(&lock->l_waitq);
  88                 RETURN(0);
  89         }
  90
  91         if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
  92                        LDLM_FL_BLOCK_CONV)))
  93                 RETURN(0);
  94
  95         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
  96                    "sleeping");
  97         ldlm_lock_dump(D_OTHER, lock);
  98         ldlm_reprocess_all(lock->l_resource);
  99
 100  noreproc:
 101         if (imp != NULL) {
 102                 spin_lock_irqsave(&imp->imp_lock, irqflags);
 103                 lwd.lwd_generation = imp->imp_generation;
 104                 spin_unlock_irqrestore(&imp->imp_lock, irqflags);
 105         }
 106
 107         /* Go to sleep until the lock is granted or cancelled. */
 108         rc = l_wait_event(lock->l_waitq,
 109                           ((lock->l_req_mode == lock->l_granted_mode) ||
 110                            lock->l_destroyed), &lwi);
 111
 112         if (lock->l_destroyed) {
 113                 LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed");
 114                 RETURN(-EIO);
 115         }
 116
 117         if (rc) {
 118                 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
 119                            rc);
 120                 RETURN(rc);
 121         }
 122
 123         LDLM_DEBUG(lock, "client-side enqueue waking up: granted");
 124         RETURN(0);
 125 }
 126
 127 static int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
 128                                   struct lustre_handle *parent_lockh,
 129                                   struct ldlm_res_id res_id,
 130                                   __u32 type,
 131                                   void *cookie, int cookielen,
 132                                   ldlm_mode_t mode,
 133                                   int *flags,
 134                                   ldlm_completion_callback completion,
 135                                   ldlm_blocking_callback blocking,
 136                                   void *data,
 137                                   struct lustre_handle *lockh)
 138 {
 139         struct ldlm_lock *lock;
 140         int err;
 141         ENTRY;
 142
 143         if (ns->ns_client) {
 144                 CERROR("Trying to enqueue local lock in a shadow namespace\n");
 145                 LBUG();
 146         }
 147
 148         lock = ldlm_lock_create(ns, parent_lockh, res_id, type, mode,
 149                                 blocking, data);
 150         if (!lock)
 151                 GOTO(out_nolock, err = -ENOMEM);
 152         LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
 153
 154         ldlm_lock_addref_internal(lock, mode);
 155         ldlm_lock2handle(lock, lockh);
 156         lock->l_flags |= LDLM_FL_LOCAL;
 157
 158         err = ldlm_lock_enqueue(ns, &lock, cookie, cookielen, flags,
 159                                 completion);
 160         if (err != ELDLM_OK)
 161                 GOTO(out, err);
 162
 163         if (type == LDLM_EXTENT)
 164                 memcpy(cookie, &lock->l_extent, sizeof(lock->l_extent));
 165         if ((*flags) & LDLM_FL_LOCK_CHANGED)
 166                 memcpy(&res_id, &lock->l_resource->lr_name, sizeof(res_id));
 167
 168         LDLM_DEBUG_NOLOCK("client-side local enqueue handler END (lock %p)",
 169                           lock);
 170
 171         if (lock->l_completion_ast)
 172                 lock->l_completion_ast(lock, *flags, NULL);
 173
 174         LDLM_DEBUG(lock, "client-side local enqueue END");
 175         EXIT;
 176  out:
 177         LDLM_LOCK_PUT(lock);
 178  out_nolock:
 179         return err;
 180 }
 181
 182 int ldlm_cli_enqueue(struct lustre_handle *connh,
 183                      struct ptlrpc_request *req,
 184                      struct ldlm_namespace *ns,
 185                      struct lustre_handle *parent_lock_handle,
 186                      struct ldlm_res_id res_id,
 187                      __u32 type,
 188                      void *cookie, int cookielen,
 189                      ldlm_mode_t mode,
 190                      int *flags,
 191                      ldlm_completion_callback completion,
 192                      ldlm_blocking_callback blocking,
 193                      void *data,
 194                      struct lustre_handle *lockh)
 195 {
 196         struct ldlm_lock *lock;
 197         struct ldlm_request *body;
 198         struct ldlm_reply *reply;
 199         int rc, size = sizeof(*body), req_passed_in = 1, is_replay;
 200         ENTRY;
 201
 202         is_replay = *flags & LDLM_FL_REPLAY;
 203         LASSERT(connh != NULL || !is_replay);
 204
 205         if (connh == NULL) {
 206                 rc = ldlm_cli_enqueue_local(ns, parent_lock_handle, res_id,
 207                                             type, cookie, cookielen, mode,
 208                                             flags, completion, blocking, data,
 209                                             lockh);
 210                 RETURN(rc);
 211         }
 212
 213         /* If we're replaying this lock, just check some invariants.
 214          * If we're creating a new lock, get everything all setup nice. */
 215         if (is_replay) {
 216                 lock = ldlm_handle2lock(lockh);
 217                 LDLM_DEBUG(lock, "client-side enqueue START");
 218                 LASSERT(connh == lock->l_connh);
 219         } else {
 220                 lock = ldlm_lock_create(ns, parent_lock_handle, res_id, type,
 221                                         mode, blocking, data);
 222                 if (lock == NULL)
 223                         GOTO(out_nolock, rc = -ENOMEM);
 224                 /* ugh.  I set this early (instead of waiting for _enqueue)
 225                  * because the completion AST might arrive early, and we need
 226                  * (in just this one case) to run the completion_cb even if it
 227                  * arrives before the reply. */
 228                 lock->l_completion_ast = completion;
 229                 LDLM_DEBUG(lock, "client-side enqueue START");
 230                 /* for the local lock, add the reference */
 231                 ldlm_lock_addref_internal(lock, mode);
 232                 ldlm_lock2handle(lock, lockh);
 233                 if (type == LDLM_EXTENT)
 234                         memcpy(&lock->l_extent, cookie,
 235                                sizeof(body->lock_desc.l_extent));
 236         }
 237
 238         if (req == NULL) {
 239                 req = ptlrpc_prep_req(class_conn2cliimp(connh), LDLM_ENQUEUE, 1,
 240                                       &size, NULL);
 241                 if (!req)
 242                         GOTO(out, rc = -ENOMEM);
 243                 req_passed_in = 0;
 244         } else if (req->rq_reqmsg->buflens[0] != sizeof(*body))
 245                 LBUG();
 246
 247         /* Dump lock data into the request buffer */
 248         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 249         ldlm_lock2desc(lock, &body->lock_desc);
 250         body->lock_flags = *flags;
 251
 252         memcpy(&body->lock_handle1, lockh, sizeof(*lockh));
 253         if (parent_lock_handle)
 254                 memcpy(&body->lock_handle2, parent_lock_handle,
 255                        sizeof(body->lock_handle2));
 256
 257         /* Continue as normal. */
 258         if (!req_passed_in) {
 259                 size = sizeof(*reply);
 260                 req->rq_replen = lustre_msg_size(1, &size);
 261         }
 262         lock->l_connh = connh;
 263         lock->l_export = NULL;
 264         lock->l_blocking_ast = blocking;
 265
 266         LDLM_DEBUG(lock, "sending request");
 267         rc = ptlrpc_queue_wait(req);
 268
 269         if (rc != ELDLM_OK) {
 270                 LASSERT(!is_replay);
 271                 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
 272                            rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
 273                 /* Set a flag to prevent us from sending a CANCEL (bug 407) */
 274                 l_lock(&ns->ns_lock);
 275                 lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 276                 LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 277                 l_unlock(&ns->ns_lock);
 278
 279                 ldlm_lock_decref_and_cancel(lockh, mode);
 280
 281                 if (rc == ELDLM_LOCK_ABORTED) {
 282                         /* caller expects reply buffer 0 to have been swabbed */
 283                         reply = lustre_swab_repbuf(req, 0, sizeof (*reply),
 284                                                    lustre_swab_ldlm_reply);
 285                         if (reply == NULL) {
 286                                 CERROR ("Can't unpack ldlm_reply\n");
 287                                 GOTO (out_req, rc = -EPROTO);
 288                         }
 289                 }
 290                 GOTO(out_req, rc);
 291         }
 292
 293         reply = lustre_swab_repbuf(req, 0, sizeof (*reply),
 294                                    lustre_swab_ldlm_reply);
 295         if (reply == NULL) {
 296                 CERROR ("Can't unpack ldlm_reply\n");
 297                 GOTO (out_req, rc = -EPROTO);
 298         }
 299
 300         memcpy(&lock->l_remote_handle, &reply->lock_handle,
 301                sizeof(lock->l_remote_handle));
 302         *flags = reply->lock_flags;
 303
 304         CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: %d\n", lock,
 305                reply->lock_handle.cookie, *flags);
 306         if (type == LDLM_EXTENT) {
 307                 CDEBUG(D_INFO, "requested extent: "LPU64" -> "LPU64", got "
 308                        "extent "LPU64" -> "LPU64"\n",
 309                        body->lock_desc.l_extent.start,
 310                        body->lock_desc.l_extent.end,
 311                        reply->lock_extent.start, reply->lock_extent.end);
 312
 313                 cookie = &reply->lock_extent; /* FIXME bug 267 */
 314                 cookielen = sizeof(reply->lock_extent);
 315         }
 316
 317         /* If enqueue returned a blocked lock but the completion handler has
 318          * already run, then it fixed up the resource and we don't need to do it
 319          * again. */
 320         if ((*flags) & LDLM_FL_LOCK_CHANGED) {
 321                 int newmode = reply->lock_mode;
 322                 LASSERT(!is_replay);
 323                 if (newmode && newmode != lock->l_req_mode) {
 324                         LDLM_DEBUG(lock, "server returned different mode %s",
 325                                    ldlm_lockname[newmode]);
 326                         lock->l_req_mode = newmode;
 327                 }
 328
 329                 if (reply->lock_resource_name.name[0] !=
 330                     lock->l_resource->lr_name.name[0]) {
 331                         CDEBUG(D_INFO, "remote intent success, locking %ld "
 332                                "instead of %ld\n",
 333                                (long)reply->lock_resource_name.name[0],
 334                                (long)lock->l_resource->lr_name.name[0]);
 335
 336                         ldlm_lock_change_resource(ns, lock,
 337                                                   reply->lock_resource_name);
 338                         if (lock->l_resource == NULL) {
 339                                 LBUG();
 340                                 GOTO(out_req, rc = -ENOMEM);
 341                         }
 342                         LDLM_DEBUG(lock, "client-side enqueue, new resource");
 343                 }
 344         }
 345
 346         if (!is_replay) {
 347                 l_lock(&ns->ns_lock);
 348                 lock->l_completion_ast = NULL;
 349                 rc = ldlm_lock_enqueue(ns, &lock, cookie, cookielen, flags,
 350                                        completion);
 351                 l_unlock(&ns->ns_lock);
 352                 if (lock->l_completion_ast)
 353                         lock->l_completion_ast(lock, *flags, NULL);
 354         }
 355
 356         LDLM_DEBUG(lock, "client-side enqueue END");
 357         EXIT;
 358  out_req:
 359         if (!req_passed_in)
 360                 ptlrpc_req_finished(req);
 361  out:
 362         LDLM_LOCK_PUT(lock);
 363  out_nolock:
 364         return rc;
 365 }
 366
 367 int ldlm_cli_replay_enqueue(struct ldlm_lock *lock)
 368 {
 369         struct lustre_handle lockh;
 370         struct ldlm_res_id junk;
 371         int flags = LDLM_FL_REPLAY;
 372         ldlm_lock2handle(lock, &lockh);
 373         return ldlm_cli_enqueue(lock->l_connh, NULL, NULL, NULL, junk,
 374                                 lock->l_resource->lr_type, NULL, 0, -1, &flags,
 375                                 NULL, NULL, NULL, &lockh);
 376 }
 377
 378 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
 379                                   int *flags)
 380 {
 381         ENTRY;
 382         if (lock->l_resource->lr_namespace->ns_client) {
 383                 CERROR("Trying to cancel local lock\n");
 384                 LBUG();
 385         }
 386         LDLM_DEBUG(lock, "client-side local convert");
 387
 388         ldlm_lock_convert(lock, new_mode, flags);
 389         ldlm_reprocess_all(lock->l_resource);
 390
 391         LDLM_DEBUG(lock, "client-side local convert handler END");
 392         LDLM_LOCK_PUT(lock);
 393         RETURN(0);
 394 }
 395
 396 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
 397  * conversion of locks which are on the waiting or converting queue */
 398 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, int *flags)
 399 {
 400         struct ldlm_request *body;
 401         struct lustre_handle *connh;
 402         struct ldlm_reply *reply;
 403         struct ldlm_lock *lock;
 404         struct ldlm_resource *res;
 405         struct ptlrpc_request *req;
 406         int rc, size = sizeof(*body);
 407         ENTRY;
 408
 409         lock = ldlm_handle2lock(lockh);
 410         if (!lock) {
 411                 LBUG();
 412                 RETURN(-EINVAL);
 413         }
 414         *flags = 0;
 415         connh = lock->l_connh;
 416
 417         if (!connh)
 418                 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
 419
 420         LDLM_DEBUG(lock, "client-side convert");
 421
 422         req = ptlrpc_prep_req(class_conn2cliimp(connh), LDLM_CONVERT, 1, &size,
 423                               NULL);
 424         if (!req)
 425                 GOTO(out, rc = -ENOMEM);
 426
 427         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 428         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 429                sizeof(body->lock_handle1));
 430
 431         body->lock_desc.l_req_mode = new_mode;
 432         body->lock_flags = *flags;
 433
 434         size = sizeof(*reply);
 435         req->rq_replen = lustre_msg_size(1, &size);
 436
 437         rc = ptlrpc_queue_wait(req);
 438         if (rc != ELDLM_OK)
 439                 GOTO(out, rc);
 440
 441         reply = lustre_swab_repbuf(req, 0, sizeof (*reply),
 442                                    lustre_swab_ldlm_reply);
 443         if (reply == NULL) {
 444                 CERROR ("Can't unpack ldlm_reply\n");
 445                 GOTO (out, rc = -EPROTO);
 446         }
 447
 448         res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
 449         if (res != NULL)
 450                 ldlm_reprocess_all(res);
 451         /* Go to sleep until the lock is granted. */
 452         /* FIXME: or cancelled. */
 453         if (lock->l_completion_ast)
 454                 lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC, NULL);
 455         EXIT;
 456  out:
 457         LDLM_LOCK_PUT(lock);
 458         ptlrpc_req_finished(req);
 459         return rc;
 460 }
 461
 462 int ldlm_cli_cancel(struct lustre_handle *lockh)
 463 {
 464         struct ptlrpc_request *req;
 465         struct ldlm_lock *lock;
 466         struct ldlm_request *body;
 467         int rc = 0, size = sizeof(*body);
 468         ENTRY;
 469
 470         /* concurrent cancels on the same handle can happen */
 471         lock = __ldlm_handle2lock(lockh, LDLM_FL_CANCELING);
 472         if (lock == NULL)
 473                 RETURN(0);
 474
 475         if (lock->l_connh) {
 476                 int local_only;
 477                 struct obd_import *imp;
 478
 479                 LDLM_DEBUG(lock, "client-side cancel");
 480                 /* Set this flag to prevent others from getting new references*/
 481                 l_lock(&lock->l_resource->lr_namespace->ns_lock);
 482                 lock->l_flags |= LDLM_FL_CBPENDING;
 483                 local_only = (lock->l_flags & LDLM_FL_LOCAL_ONLY);
 484                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 485                 ldlm_cancel_callback(lock);
 486
 487                 if (local_only) {
 488                         CDEBUG(D_INFO, "not sending request (at caller's "
 489                                "instruction)\n");
 490                         goto local_cancel;
 491                 }
 492
 493                 imp = class_conn2cliimp(lock->l_connh);
 494                 if (imp == NULL || imp->imp_invalid) {
 495                         CDEBUG(D_HA, "skipping cancel on invalid import %p\n",
 496                                imp);
 497                         goto local_cancel;
 498                 }
 499
 500                 req = ptlrpc_prep_req(imp, LDLM_CANCEL, 1, &size, NULL);
 501                 if (!req)
 502                         GOTO(out, rc = -ENOMEM);
 503
 504                 /* XXX FIXME bug 249 */
 505                 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
 506                 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
 507
 508                 body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 509                 memcpy(&body->lock_handle1, &lock->l_remote_handle,
 510                        sizeof(body->lock_handle1));
 511
 512                 req->rq_replen = lustre_msg_size(0, NULL);
 513
 514                 rc = ptlrpc_queue_wait(req);
 515
 516                 if (rc == ESTALE)
 517                         CERROR("client/server (nid "LPU64") out of sync--not "
 518                                "fatal\n",
 519                                req->rq_import->imp_connection->c_peer.peer_nid);
 520                 else if (rc != ELDLM_OK)
 521                         CERROR("Got rc %d from cancel RPC: canceling "
 522                                "anyway\n", rc);
 523
 524                 ptlrpc_req_finished(req);
 525         local_cancel:
 526                 ldlm_lock_cancel(lock);
 527         } else {
 528                 if (lock->l_resource->lr_namespace->ns_client) {
 529                         LDLM_ERROR(lock, "Trying to cancel local lock\n");
 530                         LBUG();
 531                 }
 532                 LDLM_DEBUG(lock, "client-side local cancel");
 533                 ldlm_lock_cancel(lock);
 534                 ldlm_reprocess_all(lock->l_resource);
 535                 LDLM_DEBUG(lock, "client-side local cancel handler END");
 536         }
 537
 538         EXIT;
 539  out:
 540         LDLM_LOCK_PUT(lock);
 541         return rc;
 542 }
 543
 544 int ldlm_cancel_lru(struct ldlm_namespace *ns)
 545 {
 546         struct list_head *tmp, *next, list = LIST_HEAD_INIT(list);
 547         int count, rc = 0;
 548         struct ldlm_ast_work *w;
 549         ENTRY;
 550
 551         l_lock(&ns->ns_lock);
 552         count = ns->ns_nr_unused - ns->ns_max_unused;
 553
 554         if (count <= 0) {
 555                 l_unlock(&ns->ns_lock);
 556                 RETURN(0);
 557         }
 558
 559         list_for_each_safe(tmp, next, &ns->ns_unused_list) {
 560                 struct ldlm_lock *lock;
 561                 lock = list_entry(tmp, struct ldlm_lock, l_lru);
 562
 563                 LASSERT(!lock->l_readers && !lock->l_writers);
 564
 565                 /* Setting the CBPENDING flag is a little misleading, but
 566                  * prevents an important race; namely, once CBPENDING is set,
 567                  * the lock can accumulate no more readers/writers.  Since
 568                  * readers and writers are already zero here, ldlm_lock_decref
 569                  * won't see this flag and call l_blocking_ast */
 570                 lock->l_flags |= LDLM_FL_CBPENDING;
 571
 572                 OBD_ALLOC(w, sizeof(*w));
 573                 LASSERT(w);
 574
 575                 w->w_lock = LDLM_LOCK_GET(lock);
 576                 list_add(&w->w_list, &list);
 577                 ldlm_lock_remove_from_lru(lock);
 578
 579                 if (--count == 0)
 580                         break;
 581         }
 582         l_unlock(&ns->ns_lock);
 583
 584         list_for_each_safe(tmp, next, &list) {
 585                 struct lustre_handle lockh;
 586                 int rc;
 587                 w = list_entry(tmp, struct ldlm_ast_work, w_list);
 588
 589                 ldlm_lock2handle(w->w_lock, &lockh);
 590                 rc = ldlm_cli_cancel(&lockh);
 591                 if (rc != ELDLM_OK)
 592                         CDEBUG(D_INFO, "ldlm_cli_cancel: %d\n", rc);
 593
 594                 list_del(&w->w_list);
 595                 LDLM_LOCK_PUT(w->w_lock);
 596                 OBD_FREE(w, sizeof(*w));
 597         }
 598
 599         RETURN(rc);
 600 }
 601
 602 static int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
 603                                            struct ldlm_res_id res_id, int flags,
 604                                            void *opaque)
 605 {
 606         struct ldlm_resource *res;
 607         struct list_head *tmp, *next, list = LIST_HEAD_INIT(list);
 608         struct ldlm_ast_work *w;
 609         ENTRY;
 610
 611         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
 612         if (res == NULL) {
 613                 /* This is not a problem. */
 614                 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id.name[0]);
 615                 RETURN(0);
 616         }
 617
 618         l_lock(&ns->ns_lock);
 619         list_for_each(tmp, &res->lr_granted) {
 620                 struct ldlm_lock *lock;
 621                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 622
 623                 if (opaque != NULL && lock->l_data != opaque) {
 624                         LDLM_ERROR(lock, "data %p doesn't match opaque %p",
 625                                    lock->l_data, opaque);
 626                         //LBUG();
 627                         continue;
 628                 }
 629
 630                 if (lock->l_readers || lock->l_writers) {
 631                         if (flags & LDLM_FL_WARN) {
 632                                 LDLM_ERROR(lock, "lock in use");
 633                                 //LBUG();
 634                         }
 635                         continue;
 636                 }
 637
 638                 /* See CBPENDING comment in ldlm_cancel_lru */
 639                 lock->l_flags |= LDLM_FL_CBPENDING;
 640
 641                 OBD_ALLOC(w, sizeof(*w));
 642                 LASSERT(w);
 643
 644                 w->w_lock = LDLM_LOCK_GET(lock);
 645
 646                 /* Prevent the cancel callback from being called by setting
 647                  * LDLM_FL_CANCEL in the lock.  Very sneaky. -p */
 648                 if (flags & LDLM_FL_NO_CALLBACK)
 649                         w->w_lock->l_flags |= LDLM_FL_CANCEL;
 650
 651                 list_add(&w->w_list, &list);
 652         }
 653         l_unlock(&ns->ns_lock);
 654
 655         list_for_each_safe(tmp, next, &list) {
 656                 struct lustre_handle lockh;
 657                 int rc;
 658                 w = list_entry(tmp, struct ldlm_ast_work, w_list);
 659
 660                 if (flags & LDLM_FL_LOCAL_ONLY) {
 661                         ldlm_lock_cancel(w->w_lock);
 662                 } else {
 663                         ldlm_lock2handle(w->w_lock, &lockh);
 664                         rc = ldlm_cli_cancel(&lockh);
 665                         if (rc != ELDLM_OK)
 666                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 667                 }
 668                 list_del(&w->w_list);
 669                 LDLM_LOCK_PUT(w->w_lock);
 670                 OBD_FREE(w, sizeof(*w));
 671         }
 672
 673         ldlm_resource_putref(res);
 674
 675         RETURN(0);
 676 }
 677
 678 /* Cancel all locks on a namespace (or a specific resource, if given)
 679  * that have 0 readers/writers.
 680  *
 681  * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
 682  * to notify the server.
 683  * If flags & LDLM_FL_NO_CALLBACK, don't run the cancel callback.
 684  * If flags & LDLM_FL_WARN, print a warning if some locks are still in use. */
 685 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
 686                            struct ldlm_res_id *res_id, int flags, void *opaque)
 687 {
 688         int i;
 689         ENTRY;
 690
 691         if (ns == NULL)
 692                 RETURN(ELDLM_OK);
 693
 694         if (res_id)
 695                 RETURN(ldlm_cli_cancel_unused_resource(ns, *res_id, flags,
 696                                                        opaque));
 697
 698         l_lock(&ns->ns_lock);
 699         for (i = 0; i < RES_HASH_SIZE; i++) {
 700                 struct list_head *tmp, *pos;
 701                 list_for_each_safe(tmp, pos, &(ns->ns_hash[i])) {
 702                         int rc;
 703                         struct ldlm_resource *res;
 704                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
 705                         ldlm_resource_getref(res);
 706
 707                         rc = ldlm_cli_cancel_unused_resource(ns, res->lr_name,
 708                                                              flags, opaque);
 709
 710                         if (rc)
 711                                 CERROR("cancel_unused_res ("LPU64"): %d\n",
 712                                        res->lr_name.name[0], rc);
 713                         ldlm_resource_putref(res);
 714                 }
 715         }
 716         l_unlock(&ns->ns_lock);
 717
 718         RETURN(ELDLM_OK);
 719 }
 720
 721 /* Lock iterators. */
 722
 723 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
 724                           void *closure)
 725 {
 726         struct list_head *tmp, *next;
 727         struct ldlm_lock *lock;
 728         int rc = LDLM_ITER_CONTINUE;
 729         struct ldlm_namespace *ns = res->lr_namespace;
 730
 731         ENTRY;
 732
 733         if (!res)
 734                 RETURN(LDLM_ITER_CONTINUE);
 735
 736         l_lock(&ns->ns_lock);
 737         list_for_each_safe(tmp, next, &res->lr_granted) {
 738                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 739
 740                 if (iter(lock, closure) == LDLM_ITER_STOP)
 741                         GOTO(out, rc = LDLM_ITER_STOP);
 742         }
 743
 744         list_for_each_safe(tmp, next, &res->lr_converting) {
 745                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 746
 747                 if (iter(lock, closure) == LDLM_ITER_STOP)
 748                         GOTO(out, rc = LDLM_ITER_STOP);
 749         }
 750
 751         list_for_each_safe(tmp, next, &res->lr_waiting) {
 752                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 753
 754                 if (iter(lock, closure) == LDLM_ITER_STOP)
 755                         GOTO(out, rc = LDLM_ITER_STOP);
 756         }
 757  out:
 758         l_unlock(&ns->ns_lock);
 759         RETURN(rc);
 760 }
 761
 762 struct iter_helper_data {
 763         ldlm_iterator_t iter;
 764         void *closure;
 765 };
 766
 767 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
 768 {
 769         struct iter_helper_data *helper = closure;
 770         return helper->iter(lock, helper->closure);
 771 }
 772
 773 static int ldlm_res_iter_helper(struct ldlm_resource *res, void *closure)
 774 {
 775         return ldlm_resource_foreach(res, ldlm_iter_helper, closure);
 776 }
 777
 778 int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
 779                            void *closure)
 780 {
 781         struct iter_helper_data helper = { iter: iter, closure: closure };
 782         return ldlm_namespace_foreach_res(ns, ldlm_res_iter_helper, &helper);
 783 }
 784
 785 int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
 786                                ldlm_res_iterator_t iter, void *closure)
 787 {
 788         int i, rc = LDLM_ITER_CONTINUE;
 789
 790         l_lock(&ns->ns_lock);
 791         for (i = 0; i < RES_HASH_SIZE; i++) {
 792                 struct list_head *tmp, *next;
 793                 list_for_each_safe(tmp, next, &(ns->ns_hash[i])) {
 794                         struct ldlm_resource *res =
 795                                 list_entry(tmp, struct ldlm_resource, lr_hash);
 796
 797                         ldlm_resource_getref(res);
 798                         rc = iter(res, closure);
 799                         ldlm_resource_putref(res);
 800                         if (rc == LDLM_ITER_STOP)
 801                                 GOTO(out, rc);
 802                 }
 803         }
 804  out:
 805         l_unlock(&ns->ns_lock);
 806         RETURN(rc);
 807 }
 808
 809 /* non-blocking function to manipulate a lock whose cb_data is being put away.*/
 810 void ldlm_change_cbdata(struct ldlm_namespace *ns,
 811                        struct ldlm_res_id *res_id,
 812                        ldlm_iterator_t iter,
 813                        void *data)
 814 {
 815         struct ldlm_resource *res;
 816         int rc = 0;
 817         ENTRY;
 818
 819         if (ns == NULL) {
 820                 CERROR("must pass in namespace");
 821                 LBUG();
 822         }
 823
 824         res = ldlm_resource_get(ns, NULL, *res_id, 0, 0);
 825         if (res == NULL) {
 826                 EXIT;
 827                 return;
 828         }
 829
 830         l_lock(&ns->ns_lock);
 831         rc = ldlm_resource_foreach(res, iter, data);
 832         l_unlock(&ns->ns_lock);
 833         ldlm_resource_putref(res);
 834         EXIT;
 835 }
 836
 837 /* Lock replay */
 838
 839 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
 840 {
 841         struct list_head *list = closure;
 842
 843         /* we use l_pending_chain here, because it's unused on clients. */
 844         list_add(&lock->l_pending_chain, list);
 845         return LDLM_ITER_CONTINUE;
 846 }
 847
 848 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
 849 {
 850         struct ptlrpc_request *req;
 851         struct ldlm_request *body;
 852         struct ldlm_reply *reply;
 853         int rc, size;
 854         int flags;
 855
 856         /*
 857          * If granted mode matches the requested mode, this lock is granted.
 858          *
 859          * If they differ, but we have a granted mode, then we were granted
 860          * one mode and now want another: ergo, converting.
 861          *
 862          * If we haven't been granted anything and are on a resource list,
 863          * then we're blocked/waiting.
 864          *
 865          * If we haven't been granted anything and we're NOT on a resource list,
 866          * then we haven't got a reply yet and don't have a known disposition.
 867          * This happens whenever a lock enqueue is the request that triggers
 868          * recovery.
 869          */
 870         if (lock->l_granted_mode == lock->l_req_mode)
 871                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
 872         else if (lock->l_granted_mode)
 873                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
 874         else if (!list_empty(&lock->l_res_link))
 875                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
 876         else
 877                 flags = LDLM_FL_REPLAY;
 878
 879         size = sizeof(*body);
 880         req = ptlrpc_prep_req(imp, LDLM_ENQUEUE, 1, &size, NULL);
 881         if (!req)
 882                 RETURN(-ENOMEM);
 883
 884         /* We're part of recovery, so don't wait for it. */
 885         req->rq_level = LUSTRE_CONN_RECOVER;
 886
 887         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 888         ldlm_lock2desc(lock, &body->lock_desc);
 889         body->lock_flags = flags;
 890
 891         ldlm_lock2handle(lock, &body->lock_handle1);
 892         size = sizeof(*reply);
 893         req->rq_replen = lustre_msg_size(1, &size);
 894
 895         LDLM_DEBUG(lock, "replaying lock:");
 896         rc = ptlrpc_queue_wait(req);
 897         if (rc != ELDLM_OK)
 898                 GOTO(out, rc);
 899
 900         reply = lustre_swab_repbuf(req, 0, sizeof (*reply),
 901                                    lustre_swab_ldlm_reply);
 902         if (reply == NULL) {
 903                 CERROR("Can't unpack ldlm_reply\n");
 904                 GOTO (out, rc = -EPROTO);
 905         }
 906
 907         memcpy(&lock->l_remote_handle, &reply->lock_handle,
 908                sizeof(lock->l_remote_handle));
 909         LDLM_DEBUG(lock, "replayed lock:");
 910  out:
 911         ptlrpc_req_finished(req);
 912         RETURN(rc);
 913 }
 914
 915 int ldlm_replay_locks(struct obd_import *imp)
 916 {
 917         struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
 918         struct list_head list, *pos, *next;
 919         struct ldlm_lock *lock;
 920         int rc = 0;
 921
 922         ENTRY;
 923         INIT_LIST_HEAD(&list);
 924
 925         l_lock(&ns->ns_lock);
 926         (void)ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
 927
 928         list_for_each_safe(pos, next, &list) {
 929                 lock = list_entry(pos, struct ldlm_lock, l_pending_chain);
 930                 rc = replay_one_lock(imp, lock);
 931                 if (rc)
 932                         break; /* or try to do the rest? */
 933         }
 934         l_unlock(&ns->ns_lock);
 935         RETURN(rc);
 936 }