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