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         struct lock_wait_data lwd;
  64         unsigned long irqflags;
  65         struct obd_device *obd;
  66         struct obd_import *imp = NULL;
  67         int rc = 0;
  68         struct l_wait_info lwi;
  69
  70         obd = class_conn2obd(lock->l_connh);
  71
  72         /* if this is a local lock, then there is no import */
  73         if (obd != NULL)
  74                 imp = obd->u.cli.cl_import;
  75
  76         lwd.lwd_lock = lock;
  77
  78         lwi = LWI_TIMEOUT_INTR(obd_timeout * HZ, ldlm_expired_completion_wait,
  79                                interrupted_completion_wait, &lwd);
  80         ENTRY;
  81
  82         if (flags == LDLM_FL_WAIT_NOREPROC)
  83                 goto noreproc;
  84
  85         if (flags == 0) {
  86                 wake_up(&lock->l_waitq);
  87                 RETURN(0);
  88         }
  89
  90         if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
  91                        LDLM_FL_BLOCK_CONV)))
  92                 RETURN(0);
  93
  94         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
  95                    "sleeping");
  96         ldlm_lock_dump(D_OTHER, lock);
  97         ldlm_reprocess_all(lock->l_resource);
  98
  99  noreproc:
 100         if (imp != NULL) {
 101                 spin_lock_irqsave(&imp->imp_lock, irqflags);
 102                 lwd.lwd_generation = imp->imp_generation;
 103                 spin_unlock_irqrestore(&imp->imp_lock, irqflags);
 104         }
 105
 106         /* Go to sleep until the lock is granted or cancelled. */
 107         rc = l_wait_event(lock->l_waitq,
 108                           ((lock->l_req_mode == lock->l_granted_mode) ||
 109                            lock->l_destroyed), &lwi);
 110
 111         if (lock->l_destroyed) {
 112                 LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed");
 113                 RETURN(-EIO);
 114         }
 115
 116         if (rc) {
 117                 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
 118                            rc);
 119                 RETURN(rc);
 120         }
 121
 122         LDLM_DEBUG(lock, "client-side enqueue waking up: granted");
 123         RETURN(0);
 124 }
 125
 126 static int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
 127                                   struct lustre_handle *parent_lockh,
 128                                   struct ldlm_res_id res_id,
 129                                   __u32 type,
 130                                   void *cookie, int cookielen,
 131                                   ldlm_mode_t mode,
 132                                   int *flags,
 133                                   ldlm_completion_callback completion,
 134                                   ldlm_blocking_callback blocking,
 135                                   void *data,
 136                                   struct lustre_handle *lockh)
 137 {
 138         struct ldlm_lock *lock;
 139         int err;
 140         ENTRY;
 141
 142         if (ns->ns_client) {
 143                 CERROR("Trying to enqueue local lock in a shadow namespace\n");
 144                 LBUG();
 145         }
 146
 147         lock = ldlm_lock_create(ns, parent_lockh, res_id, type, mode,
 148                                 blocking, data);
 149         if (!lock)
 150                 GOTO(out_nolock, err = -ENOMEM);
 151         LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
 152
 153         ldlm_lock_addref_internal(lock, mode);
 154         ldlm_lock2handle(lock, lockh);
 155         lock->l_flags |= LDLM_FL_LOCAL;
 156
 157         err = ldlm_lock_enqueue(ns, &lock, cookie, cookielen, flags,
 158                                 completion);
 159         if (err != ELDLM_OK)
 160                 GOTO(out, err);
 161
 162         if (type == LDLM_EXTENT)
 163                 memcpy(cookie, &lock->l_extent, sizeof(lock->l_extent));
 164         if ((*flags) & LDLM_FL_LOCK_CHANGED)
 165                 memcpy(&res_id, &lock->l_resource->lr_name, sizeof(res_id));
 166
 167         LDLM_DEBUG_NOLOCK("client-side local enqueue handler END (lock %p)",
 168                           lock);
 169
 170         if (lock->l_completion_ast)
 171                 lock->l_completion_ast(lock, *flags, NULL);
 172
 173         LDLM_DEBUG(lock, "client-side local enqueue END");
 174         EXIT;
 175  out:
 176         LDLM_LOCK_PUT(lock);
 177  out_nolock:
 178         return err;
 179 }
 180
 181 int ldlm_cli_enqueue(struct lustre_handle *connh,
 182                      struct ptlrpc_request *req,
 183                      struct ldlm_namespace *ns,
 184                      struct lustre_handle *parent_lock_handle,
 185                      struct ldlm_res_id res_id,
 186                      __u32 type,
 187                      void *cookie, int cookielen,
 188                      ldlm_mode_t mode,
 189                      int *flags,
 190                      ldlm_completion_callback completion,
 191                      ldlm_blocking_callback blocking,
 192                      void *data,
 193                      struct lustre_handle *lockh)
 194 {
 195         struct ldlm_lock *lock;
 196         struct ldlm_request *body;
 197         struct ldlm_reply *reply;
 198         int rc, size = sizeof(*body), req_passed_in = 1, is_replay;
 199         ENTRY;
 200
 201         is_replay = *flags & LDLM_FL_REPLAY;
 202         LASSERT(connh != NULL || !is_replay);
 203
 204         if (connh == NULL) {
 205                 rc = ldlm_cli_enqueue_local(ns, parent_lock_handle, res_id,
 206                                             type, cookie, cookielen, mode,
 207                                             flags, completion, blocking, data,
 208                                             lockh);
 209                 RETURN(rc);
 210         }
 211
 212         /* If we're replaying this lock, just check some invariants.
 213          * If we're creating a new lock, get everything all setup nice. */
 214         if (is_replay) {
 215                 lock = ldlm_handle2lock(lockh);
 216                 LDLM_DEBUG(lock, "client-side enqueue START");
 217                 LASSERT(connh == lock->l_connh);
 218         } else {
 219                 lock = ldlm_lock_create(ns, parent_lock_handle, res_id, type,
 220                                         mode, blocking, data);
 221                 if (lock == NULL)
 222                         GOTO(out_nolock, rc = -ENOMEM);
 223                 /* ugh.  I set this early (instead of waiting for _enqueue)
 224                  * because the completion AST might arrive early, and we need
 225                  * (in just this one case) to run the completion_cb even if it
 226                  * arrives before the reply. */
 227                 lock->l_completion_ast = completion;
 228                 LDLM_DEBUG(lock, "client-side enqueue START");
 229                 /* for the local lock, add the reference */
 230                 ldlm_lock_addref_internal(lock, mode);
 231                 ldlm_lock2handle(lock, lockh);
 232                 if (type == LDLM_EXTENT)
 233                         memcpy(&lock->l_extent, cookie,
 234                                sizeof(body->lock_desc.l_extent));
 235         }
 236
 237         if (req == NULL) {
 238                 req = ptlrpc_prep_req(class_conn2cliimp(connh), LDLM_ENQUEUE, 1,
 239                                       &size, NULL);
 240                 if (!req)
 241                         GOTO(out, rc = -ENOMEM);
 242                 req_passed_in = 0;
 243         } else if (req->rq_reqmsg->buflens[0] != sizeof(*body))
 244                 LBUG();
 245
 246         /* Dump lock data into the request buffer */
 247         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 248         ldlm_lock2desc(lock, &body->lock_desc);
 249         body->lock_flags = *flags;
 250
 251         memcpy(&body->lock_handle1, lockh, sizeof(*lockh));
 252         if (parent_lock_handle)
 253                 memcpy(&body->lock_handle2, parent_lock_handle,
 254                        sizeof(body->lock_handle2));
 255
 256         /* Continue as normal. */
 257         if (!req_passed_in) {
 258                 size = sizeof(*reply);
 259                 req->rq_replen = lustre_msg_size(1, &size);
 260         }
 261         lock->l_connh = connh;
 262         lock->l_export = NULL;
 263         lock->l_blocking_ast = blocking;
 264
 265         LDLM_DEBUG(lock, "sending request");
 266         rc = ptlrpc_queue_wait(req);
 267
 268         if (rc != ELDLM_OK) {
 269                 LASSERT(!is_replay);
 270                 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
 271                            rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
 272                 /* Set a flag to prevent us from sending a CANCEL (bug 407) */
 273                 l_lock(&ns->ns_lock);
 274                 lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 275                 l_unlock(&ns->ns_lock);
 276
 277                 ldlm_lock_decref_and_cancel(lockh, mode);
 278
 279                 if (rc == ELDLM_LOCK_ABORTED) {
 280                         /* caller expects reply buffer 0 to have been swabbed */
 281                         reply = lustre_swab_repbuf(req, 0, sizeof (*reply),
 282                                                    lustre_swab_ldlm_reply);
 283                         if (reply == NULL) {
 284                                 CERROR ("Can't unpack ldlm_reply\n");
 285                                 GOTO (out_req, rc = -EPROTO);
 286                         }
 287                 }
 288                 GOTO(out_req, rc);
 289         }
 290
 291         reply = lustre_swab_repbuf(req, 0, sizeof (*reply),
 292                                    lustre_swab_ldlm_reply);
 293         if (reply == NULL) {
 294                 CERROR ("Can't unpack ldlm_reply\n");
 295                 GOTO (out_req, rc = -EPROTO);
 296         }
 297
 298         memcpy(&lock->l_remote_handle, &reply->lock_handle,
 299                sizeof(lock->l_remote_handle));
 300         *flags = reply->lock_flags;
 301
 302         CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: %d\n", lock,
 303                reply->lock_handle.cookie, *flags);
 304         if (type == LDLM_EXTENT) {
 305                 CDEBUG(D_INFO, "requested extent: "LPU64" -> "LPU64", got "
 306                        "extent "LPU64" -> "LPU64"\n",
 307                        body->lock_desc.l_extent.start,
 308                        body->lock_desc.l_extent.end,
 309                        reply->lock_extent.start, reply->lock_extent.end);
 310
 311                 if ((reply->lock_extent.end & ~PAGE_MASK) != ~PAGE_MASK) {
 312                         /* XXX Old versions of BA OST code have a fencepost bug
 313                          * which will cause them to grant a lock that's one
 314                          * byte too large.  This can be safely removed after BA
 315                          * ships their next release -phik (02 Apr 2003) */
 316                         reply->lock_extent.end--;
 317                 } else if ((reply->lock_extent.start & ~PAGE_MASK) ==
 318                            ~PAGE_MASK) {
 319                         reply->lock_extent.start++;
 320                 }
 321
 322                 cookie = &reply->lock_extent; /* FIXME bug 267 */
 323                 cookielen = sizeof(reply->lock_extent);
 324         }
 325
 326         /* If enqueue returned a blocked lock but the completion handler has
 327          * already run, then it fixed up the resource and we don't need to do it
 328          * again. */
 329         if ((*flags) & LDLM_FL_LOCK_CHANGED) {
 330                 int newmode = reply->lock_mode;
 331                 LASSERT(!is_replay);
 332                 if (newmode && newmode != lock->l_req_mode) {
 333                         LDLM_DEBUG(lock, "server returned different mode %s",
 334                                    ldlm_lockname[newmode]);
 335                         lock->l_req_mode = newmode;
 336                 }
 337
 338                 if (reply->lock_resource_name.name[0] !=
 339                     lock->l_resource->lr_name.name[0]) {
 340                         CDEBUG(D_INFO, "remote intent success, locking %ld "
 341                                "instead of %ld\n",
 342                                (long)reply->lock_resource_name.name[0],
 343                                (long)lock->l_resource->lr_name.name[0]);
 344
 345                         ldlm_lock_change_resource(ns, lock,
 346                                                   reply->lock_resource_name);
 347                         if (lock->l_resource == NULL) {
 348                                 LBUG();
 349                                 GOTO(out_req, rc = -ENOMEM);
 350                         }
 351                         LDLM_DEBUG(lock, "client-side enqueue, new resource");
 352                 }
 353         }
 354
 355         if (!is_replay) {
 356                 l_lock(&ns->ns_lock);
 357                 lock->l_completion_ast = NULL;
 358                 rc = ldlm_lock_enqueue(ns, &lock, cookie, cookielen, flags,
 359                                        completion);
 360                 l_unlock(&ns->ns_lock);
 361                 if (lock->l_completion_ast)
 362                         lock->l_completion_ast(lock, *flags, NULL);
 363         }
 364
 365         LDLM_DEBUG(lock, "client-side enqueue END");
 366         EXIT;
 367  out_req:
 368         if (!req_passed_in)
 369                 ptlrpc_req_finished(req);
 370  out:
 371         LDLM_LOCK_PUT(lock);
 372  out_nolock:
 373         return rc;
 374 }
 375
 376 int ldlm_match_or_enqueue(struct lustre_handle *connh,
 377                           struct ptlrpc_request *req,
 378                           struct ldlm_namespace *ns,
 379                           struct lustre_handle *parent_lock_handle,
 380                           struct ldlm_res_id res_id,
 381                           __u32 type,
 382                           void *cookie, int cookielen,
 383                           ldlm_mode_t mode,
 384                           int *flags,
 385                           ldlm_completion_callback completion,
 386                           ldlm_blocking_callback blocking,
 387                           void *data,
 388                           struct lustre_handle *lockh)
 389 {
 390         int rc;
 391         ENTRY;
 392         if (connh == NULL) {
 393                 /* Just to make sure that I understand things --phil */
 394                 LASSERT(*flags & LDLM_FL_LOCAL_ONLY);
 395         }
 396
 397         LDLM_DEBUG_NOLOCK("resource "LPU64"/"LPU64, res_id.name[0],
 398                           res_id.name[1]);
 399         rc = ldlm_lock_match(ns, *flags, &res_id, type, cookie, cookielen, mode,
 400                              lockh);
 401         if (rc == 0) {
 402                 rc = ldlm_cli_enqueue(connh, req, ns, parent_lock_handle,
 403                                       res_id, type, cookie, cookielen, mode,
 404                                       flags, completion, blocking, data,
 405                                       lockh);
 406                 if (rc != ELDLM_OK)
 407                         CERROR("ldlm_cli_enqueue: err: %d\n", rc);
 408                 RETURN(rc);
 409         }
 410         RETURN(0);
 411 }
 412
 413 int ldlm_cli_replay_enqueue(struct ldlm_lock *lock)
 414 {
 415         struct lustre_handle lockh;
 416         struct ldlm_res_id junk;
 417         int flags = LDLM_FL_REPLAY;
 418         ldlm_lock2handle(lock, &lockh);
 419         return ldlm_cli_enqueue(lock->l_connh, NULL, NULL, NULL, junk,
 420                                 lock->l_resource->lr_type, NULL, 0, -1, &flags,
 421                                 NULL, NULL, NULL, &lockh);
 422 }
 423
 424 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
 425                                   int *flags)
 426 {
 427         ENTRY;
 428         if (lock->l_resource->lr_namespace->ns_client) {
 429                 CERROR("Trying to cancel local lock\n");
 430                 LBUG();
 431         }
 432         LDLM_DEBUG(lock, "client-side local convert");
 433
 434         ldlm_lock_convert(lock, new_mode, flags);
 435         ldlm_reprocess_all(lock->l_resource);
 436
 437         LDLM_DEBUG(lock, "client-side local convert handler END");
 438         LDLM_LOCK_PUT(lock);
 439         RETURN(0);
 440 }
 441
 442 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
 443  * conversion of locks which are on the waiting or converting queue */
 444 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, int *flags)
 445 {
 446         struct ldlm_request *body;
 447         struct lustre_handle *connh;
 448         struct ldlm_reply *reply;
 449         struct ldlm_lock *lock;
 450         struct ldlm_resource *res;
 451         struct ptlrpc_request *req;
 452         int rc, size = sizeof(*body);
 453         ENTRY;
 454
 455         lock = ldlm_handle2lock(lockh);
 456         if (!lock) {
 457                 LBUG();
 458                 RETURN(-EINVAL);
 459         }
 460         *flags = 0;
 461         connh = lock->l_connh;
 462
 463         if (!connh)
 464                 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
 465
 466         LDLM_DEBUG(lock, "client-side convert");
 467
 468         req = ptlrpc_prep_req(class_conn2cliimp(connh), LDLM_CONVERT, 1, &size,
 469                               NULL);
 470         if (!req)
 471                 GOTO(out, rc = -ENOMEM);
 472
 473         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 474         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 475                sizeof(body->lock_handle1));
 476
 477         body->lock_desc.l_req_mode = new_mode;
 478         body->lock_flags = *flags;
 479
 480         size = sizeof(*reply);
 481         req->rq_replen = lustre_msg_size(1, &size);
 482
 483         rc = ptlrpc_queue_wait(req);
 484         if (rc != ELDLM_OK)
 485                 GOTO(out, rc);
 486
 487         reply = lustre_swab_repbuf(req, 0, sizeof (*reply),
 488                                    lustre_swab_ldlm_reply);
 489         if (reply == NULL) {
 490                 CERROR ("Can't unpack ldlm_reply\n");
 491                 GOTO (out, rc = -EPROTO);
 492         }
 493
 494         res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
 495         if (res != NULL)
 496                 ldlm_reprocess_all(res);
 497         /* Go to sleep until the lock is granted. */
 498         /* FIXME: or cancelled. */
 499         if (lock->l_completion_ast)
 500                 lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC, NULL);
 501         EXIT;
 502  out:
 503         LDLM_LOCK_PUT(lock);
 504         ptlrpc_req_finished(req);
 505         return rc;
 506 }
 507
 508 int ldlm_cli_cancel(struct lustre_handle *lockh)
 509 {
 510         struct ptlrpc_request *req;
 511         struct ldlm_lock *lock;
 512         struct ldlm_request *body;
 513         int rc = 0, size = sizeof(*body);
 514         ENTRY;
 515
 516         /* concurrent cancels on the same handle can happen */
 517         lock = __ldlm_handle2lock(lockh, LDLM_FL_CANCELING);
 518         if (lock == NULL)
 519                 RETURN(0);
 520
 521         if (lock->l_connh) {
 522                 int local_only;
 523                 struct obd_import *imp;
 524
 525                 LDLM_DEBUG(lock, "client-side cancel");
 526                 /* Set this flag to prevent others from getting new references*/
 527                 l_lock(&lock->l_resource->lr_namespace->ns_lock);
 528                 lock->l_flags |= LDLM_FL_CBPENDING;
 529                 local_only = (lock->l_flags & LDLM_FL_LOCAL_ONLY);
 530                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 531                 ldlm_cancel_callback(lock);
 532
 533                 if (local_only) {
 534                         CDEBUG(D_INFO, "not sending request (at caller's "
 535                                "instruction)\n");
 536                         goto local_cancel;
 537                 }
 538
 539                 imp = class_conn2cliimp(lock->l_connh);
 540                 if (imp == NULL || imp->imp_invalid) {
 541                         CDEBUG(D_HA, "skipping cancel on invalid import %p\n",
 542                                imp);
 543                         goto local_cancel;
 544                 }
 545
 546                 req = ptlrpc_prep_req(imp, LDLM_CANCEL, 1, &size, NULL);
 547                 if (!req)
 548                         GOTO(out, rc = -ENOMEM);
 549
 550                 /* XXX FIXME bug 249 */
 551                 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
 552                 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
 553
 554                 body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 555                 memcpy(&body->lock_handle1, &lock->l_remote_handle,
 556                        sizeof(body->lock_handle1));
 557
 558                 req->rq_replen = lustre_msg_size(0, NULL);
 559
 560                 rc = ptlrpc_queue_wait(req);
 561
 562                 if (rc == ESTALE)
 563                         CERROR("client/server (nid "LPU64") out of sync--not "
 564                                "fatal\n",
 565                                req->rq_import->imp_connection->c_peer.peer_nid);
 566                 else if (rc != ELDLM_OK)
 567                         CERROR("Got rc %d from cancel RPC: canceling "
 568                                "anyway\n", rc);
 569
 570                 ptlrpc_req_finished(req);
 571         local_cancel:
 572                 ldlm_lock_cancel(lock);
 573         } else {
 574                 LDLM_DEBUG(lock, "client-side local cancel");
 575                 if (lock->l_resource->lr_namespace->ns_client) {
 576                         CERROR("Trying to cancel local lock\n");
 577                         LBUG();
 578                 }
 579                 ldlm_lock_cancel(lock);
 580                 ldlm_reprocess_all(lock->l_resource);
 581                 LDLM_DEBUG(lock, "client-side local cancel handler END");
 582         }
 583
 584         EXIT;
 585  out:
 586         LDLM_LOCK_PUT(lock);
 587         return rc;
 588 }
 589
 590 int ldlm_cancel_lru(struct ldlm_namespace *ns)
 591 {
 592         struct list_head *tmp, *next, list = LIST_HEAD_INIT(list);
 593         int count, rc = 0;
 594         struct ldlm_ast_work *w;
 595         ENTRY;
 596
 597         l_lock(&ns->ns_lock);
 598         count = ns->ns_nr_unused - ns->ns_max_unused;
 599
 600         if (count <= 0) {
 601                 l_unlock(&ns->ns_lock);
 602                 RETURN(0);
 603         }
 604
 605         list_for_each_safe(tmp, next, &ns->ns_unused_list) {
 606                 struct ldlm_lock *lock;
 607                 lock = list_entry(tmp, struct ldlm_lock, l_lru);
 608
 609                 LASSERT(!lock->l_readers && !lock->l_writers);
 610
 611                 /* Setting the CBPENDING flag is a little misleading, but
 612                  * prevents an important race; namely, once CBPENDING is set,
 613                  * the lock can accumulate no more readers/writers.  Since
 614                  * readers and writers are already zero here, ldlm_lock_decref
 615                  * won't see this flag and call l_blocking_ast */
 616                 lock->l_flags |= LDLM_FL_CBPENDING;
 617
 618                 OBD_ALLOC(w, sizeof(*w));
 619                 LASSERT(w);
 620
 621                 w->w_lock = LDLM_LOCK_GET(lock);
 622                 list_add(&w->w_list, &list);
 623                 ldlm_lock_remove_from_lru(lock);
 624
 625                 if (--count == 0)
 626                         break;
 627         }
 628         l_unlock(&ns->ns_lock);
 629
 630         list_for_each_safe(tmp, next, &list) {
 631                 struct lustre_handle lockh;
 632                 int rc;
 633                 w = list_entry(tmp, struct ldlm_ast_work, w_list);
 634
 635                 ldlm_lock2handle(w->w_lock, &lockh);
 636                 rc = ldlm_cli_cancel(&lockh);
 637                 if (rc != ELDLM_OK)
 638                         CDEBUG(D_INFO, "ldlm_cli_cancel: %d\n", rc);
 639
 640                 list_del(&w->w_list);
 641                 LDLM_LOCK_PUT(w->w_lock);
 642                 OBD_FREE(w, sizeof(*w));
 643         }
 644
 645         RETURN(rc);
 646 }
 647
 648 static int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
 649                                            struct ldlm_res_id res_id, int flags,
 650                                            void *opaque)
 651 {
 652         struct ldlm_resource *res;
 653         struct list_head *tmp, *next, list = LIST_HEAD_INIT(list);
 654         struct ldlm_ast_work *w;
 655         ENTRY;
 656
 657         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
 658         if (res == NULL) {
 659                 /* This is not a problem. */
 660                 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id.name[0]);
 661                 RETURN(0);
 662         }
 663
 664         l_lock(&ns->ns_lock);
 665         list_for_each(tmp, &res->lr_granted) {
 666                 struct ldlm_lock *lock;
 667                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 668
 669                 if (lock->l_readers || lock->l_writers) {
 670                         if (flags & LDLM_FL_WARN) {
 671                                 LDLM_ERROR(lock, "lock in use");
 672                                 LBUG();
 673                         }
 674                 }
 675                 if (opaque != NULL && lock->l_data != opaque) {
 676                         LDLM_ERROR(lock, "data %p doesn't match opaque %p",
 677                                    lock->l_data, opaque);
 678                         LBUG();
 679                 }
 680
 681                 /* See CBPENDING comment in ldlm_cancel_lru */
 682                 lock->l_flags |= LDLM_FL_CBPENDING;
 683
 684                 OBD_ALLOC(w, sizeof(*w));
 685                 LASSERT(w);
 686
 687                 w->w_lock = LDLM_LOCK_GET(lock);
 688
 689                 /* Prevent the cancel callback from being called by setting
 690                  * LDLM_FL_CANCEL in the lock.  Very sneaky. -p */
 691                 if (flags & LDLM_FL_NO_CALLBACK)
 692                         w->w_lock->l_flags |= LDLM_FL_CANCEL;
 693
 694                 list_add(&w->w_list, &list);
 695         }
 696         l_unlock(&ns->ns_lock);
 697
 698         list_for_each_safe(tmp, next, &list) {
 699                 struct lustre_handle lockh;
 700                 int rc;
 701                 w = list_entry(tmp, struct ldlm_ast_work, w_list);
 702
 703                 if (flags & LDLM_FL_LOCAL_ONLY) {
 704                         ldlm_lock_cancel(w->w_lock);
 705                 } else {
 706                         ldlm_lock2handle(w->w_lock, &lockh);
 707                         rc = ldlm_cli_cancel(&lockh);
 708                         if (rc != ELDLM_OK)
 709                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 710                 }
 711                 list_del(&w->w_list);
 712                 LDLM_LOCK_PUT(w->w_lock);
 713                 OBD_FREE(w, sizeof(*w));
 714         }
 715
 716         ldlm_resource_putref(res);
 717
 718         RETURN(0);
 719 }
 720
 721 /* Cancel all locks on a namespace (or a specific resource, if given)
 722  * that have 0 readers/writers.
 723  *
 724  * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
 725  * to notify the server.
 726  * If flags & LDLM_FL_NO_CALLBACK, don't run the cancel callback.
 727  * If flags & LDLM_FL_WARN, print a warning if some locks are still in use. */
 728 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
 729                            struct ldlm_res_id *res_id, int flags, void *opaque)
 730 {
 731         int i;
 732         ENTRY;
 733
 734         if (ns == NULL)
 735                 RETURN(ELDLM_OK);
 736
 737         if (res_id)
 738                 RETURN(ldlm_cli_cancel_unused_resource(ns, *res_id, flags,
 739                                                        opaque));
 740
 741         l_lock(&ns->ns_lock);
 742         for (i = 0; i < RES_HASH_SIZE; i++) {
 743                 struct list_head *tmp, *pos;
 744                 list_for_each_safe(tmp, pos, &(ns->ns_hash[i])) {
 745                         int rc;
 746                         struct ldlm_resource *res;
 747                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
 748                         ldlm_resource_getref(res);
 749
 750                         rc = ldlm_cli_cancel_unused_resource(ns, res->lr_name,
 751                                                              flags, opaque);
 752
 753                         if (rc)
 754                                 CERROR("cancel_unused_res ("LPU64"): %d\n",
 755                                        res->lr_name.name[0], rc);
 756                         ldlm_resource_putref(res);
 757                 }
 758         }
 759         l_unlock(&ns->ns_lock);
 760
 761         RETURN(ELDLM_OK);
 762 }
 763
 764 /* Lock iterators. */
 765
 766 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
 767                           void *closure)
 768 {
 769         struct list_head *tmp, *next;
 770         struct ldlm_lock *lock;
 771         int rc = LDLM_ITER_CONTINUE;
 772         struct ldlm_namespace *ns = res->lr_namespace;
 773
 774         ENTRY;
 775
 776         if (!res)
 777                 RETURN(LDLM_ITER_CONTINUE);
 778
 779         l_lock(&ns->ns_lock);
 780         list_for_each_safe(tmp, next, &res->lr_granted) {
 781                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 782
 783                 if (iter(lock, closure) == LDLM_ITER_STOP)
 784                         GOTO(out, rc = LDLM_ITER_STOP);
 785         }
 786
 787         list_for_each_safe(tmp, next, &res->lr_converting) {
 788                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 789
 790                 if (iter(lock, closure) == LDLM_ITER_STOP)
 791                         GOTO(out, rc = LDLM_ITER_STOP);
 792         }
 793
 794         list_for_each_safe(tmp, next, &res->lr_waiting) {
 795                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 796
 797                 if (iter(lock, closure) == LDLM_ITER_STOP)
 798                         GOTO(out, rc = LDLM_ITER_STOP);
 799         }
 800  out:
 801         l_unlock(&ns->ns_lock);
 802         RETURN(rc);
 803 }
 804
 805 struct iter_helper_data {
 806         ldlm_iterator_t iter;
 807         void *closure;
 808 };
 809
 810 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
 811 {
 812         struct iter_helper_data *helper = closure;
 813         return helper->iter(lock, helper->closure);
 814 }
 815
 816 static int ldlm_res_iter_helper(struct ldlm_resource *res, void *closure)
 817 {
 818         return ldlm_resource_foreach(res, ldlm_iter_helper, closure);
 819 }
 820
 821 int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
 822                            void *closure)
 823 {
 824         struct iter_helper_data helper = { iter: iter, closure: closure };
 825         return ldlm_namespace_foreach_res(ns, ldlm_res_iter_helper, &helper);
 826 }
 827
 828 int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
 829                                ldlm_res_iterator_t iter, void *closure)
 830 {
 831         int i, rc = LDLM_ITER_CONTINUE;
 832
 833         l_lock(&ns->ns_lock);
 834         for (i = 0; i < RES_HASH_SIZE; i++) {
 835                 struct list_head *tmp, *next;
 836                 list_for_each_safe(tmp, next, &(ns->ns_hash[i])) {
 837                         struct ldlm_resource *res =
 838                                 list_entry(tmp, struct ldlm_resource, lr_hash);
 839
 840                         ldlm_resource_getref(res);
 841                         rc = iter(res, closure);
 842                         ldlm_resource_putref(res);
 843                         if (rc == LDLM_ITER_STOP)
 844                                 GOTO(out, rc);
 845                 }
 846         }
 847  out:
 848         l_unlock(&ns->ns_lock);
 849         RETURN(rc);
 850 }
 851
 852 /* Lock replay */
 853
 854 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
 855 {
 856         struct list_head *list = closure;
 857
 858         /* we use l_pending_chain here, because it's unused on clients. */
 859         list_add(&lock->l_pending_chain, list);
 860         return LDLM_ITER_CONTINUE;
 861 }
 862
 863 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
 864 {
 865         struct ptlrpc_request *req;
 866         struct ldlm_request *body;
 867         struct ldlm_reply *reply;
 868         int rc, size;
 869         int flags;
 870
 871         /*
 872          * If granted mode matches the requested mode, this lock is granted.
 873          *
 874          * If they differ, but we have a granted mode, then we were granted
 875          * one mode and now want another: ergo, converting.
 876          *
 877          * If we haven't been granted anything and are on a resource list,
 878          * then we're blocked/waiting.
 879          *
 880          * If we haven't been granted anything and we're NOT on a resource list,
 881          * then we haven't got a reply yet and don't have a known disposition.
 882          * This happens whenever a lock enqueue is the request that triggers
 883          * recovery.
 884          */
 885         if (lock->l_granted_mode == lock->l_req_mode)
 886                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
 887         else if (lock->l_granted_mode)
 888                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
 889         else if (!list_empty(&lock->l_res_link))
 890                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
 891         else
 892                 flags = LDLM_FL_REPLAY;
 893
 894         size = sizeof(*body);
 895         req = ptlrpc_prep_req(imp, LDLM_ENQUEUE, 1, &size, NULL);
 896         if (!req)
 897                 RETURN(-ENOMEM);
 898
 899         /* We're part of recovery, so don't wait for it. */
 900         req->rq_level = LUSTRE_CONN_RECOVD;
 901
 902         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 903         ldlm_lock2desc(lock, &body->lock_desc);
 904         body->lock_flags = flags;
 905
 906         ldlm_lock2handle(lock, &body->lock_handle1);
 907         size = sizeof(*reply);
 908         req->rq_replen = lustre_msg_size(1, &size);
 909
 910         LDLM_DEBUG(lock, "replaying lock:");
 911         rc = ptlrpc_queue_wait(req);
 912         if (rc != ELDLM_OK)
 913                 GOTO(out, rc);
 914
 915         reply = lustre_swab_repbuf(req, 0, sizeof (*reply),
 916                                    lustre_swab_ldlm_reply);
 917         if (reply == NULL) {
 918                 CERROR("Can't unpack ldlm_reply\n");
 919                 GOTO (out, rc = -EPROTO);
 920         }
 921
 922         memcpy(&lock->l_remote_handle, &reply->lock_handle,
 923                sizeof(lock->l_remote_handle));
 924         LDLM_DEBUG(lock, "replayed lock:");
 925  out:
 926         ptlrpc_req_finished(req);
 927         RETURN(rc);
 928 }
 929
 930 int ldlm_replay_locks(struct obd_import *imp)
 931 {
 932         struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
 933         struct list_head list, *pos, *next;
 934         struct ldlm_lock *lock;
 935         int rc = 0;
 936
 937         ENTRY;
 938         INIT_LIST_HEAD(&list);
 939
 940         l_lock(&ns->ns_lock);
 941         (void)ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
 942
 943         list_for_each_safe(pos, next, &list) {
 944                 lock = list_entry(pos, struct ldlm_lock, l_pending_chain);
 945                 rc = replay_one_lock(imp, lock);
 946                 if (rc)
 947                         break; /* or try to do the rest? */
 948         }
 949         l_unlock(&ns->ns_lock);
 950         RETURN(rc);
 951 }