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 the Lustre file system, http://www.lustre.org
   7  *   Lustre is a trademark of Cluster File Systems, Inc.
   8  *
   9  *   You may have signed or agreed to another license before downloading
  10  *   this software.  If so, you are bound by the terms and conditions
  11  *   of that agreement, and the following does not apply to you.  See the
  12  *   LICENSE file included with this distribution for more information.
  13  *
  14  *   If you did not agree to a different license, then this copy of Lustre
  15  *   is open source software; you can redistribute it and/or modify it
  16  *   under the terms of version 2 of the GNU General Public License as
  17  *   published by the Free Software Foundation.
  18  *
  19  *   In either case, Lustre is distributed in the hope that it will be
  20  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  21  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  22  *   license text for more details.
  23  */
  24
  25 #define DEBUG_SUBSYSTEM S_LDLM
  26 #ifndef __KERNEL__
  27 #include <signal.h>
  28 #include <liblustre.h>
  29 #endif
  30
  31 #include <lustre_dlm.h>
  32 #include <obd_class.h>
  33 #include <obd.h>
  34
  35 #include "ldlm_internal.h"
  36
  37 static void interrupted_completion_wait(void *data)
  38 {
  39 }
  40
  41 struct lock_wait_data {
  42         struct ldlm_lock *lwd_lock;
  43         __u32             lwd_conn_cnt;
  44 };
  45
  46 int ldlm_expired_completion_wait(void *data)
  47 {
  48         struct lock_wait_data *lwd = data;
  49         struct ldlm_lock *lock = lwd->lwd_lock;
  50         struct obd_import *imp;
  51         struct obd_device *obd;
  52
  53         ENTRY;
  54         if (lock->l_conn_export == NULL) {
  55                 static cfs_time_t next_dump = 0, last_dump = 0;
  56
  57                 if (ptlrpc_check_suspend())
  58                         RETURN(0);
  59
  60                 LDLM_ERROR(lock, "lock timed out (enqueued at %lu, %lus ago); "
  61                            "not entering recovery in server code, just going "
  62                            "back to sleep", lock->l_enqueued_time.tv_sec,
  63                            CURRENT_SECONDS - lock->l_enqueued_time.tv_sec);
  64                 if (cfs_time_after(cfs_time_current(), next_dump)) {
  65                         last_dump = next_dump;
  66                         next_dump = cfs_time_shift(300);
  67                         ldlm_namespace_dump(D_DLMTRACE,
  68                                             lock->l_resource->lr_namespace);
  69                         if (last_dump == 0)
  70                                 libcfs_debug_dumplog();
  71                 }
  72                 RETURN(0);
  73         }
  74
  75         obd = lock->l_conn_export->exp_obd;
  76         imp = obd->u.cli.cl_import;
  77         ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
  78         LDLM_ERROR(lock, "lock timed out (enqueued at %lu, %lus ago), entering "
  79                    "recovery for %s@%s", lock->l_enqueued_time.tv_sec,
  80                    CURRENT_SECONDS - lock->l_enqueued_time.tv_sec,
  81                    obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
  82
  83         RETURN(0);
  84 }
  85
  86 int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data)
  87 {
  88         /* XXX ALLOCATE - 160 bytes */
  89         struct lock_wait_data lwd;
  90         struct obd_device *obd;
  91         struct obd_import *imp = NULL;
  92         struct l_wait_info lwi;
  93         int rc = 0;
  94         ENTRY;
  95
  96         if (flags == LDLM_FL_WAIT_NOREPROC) {
  97                 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
  98                 goto noreproc;
  99         }
 100
 101         if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
 102                        LDLM_FL_BLOCK_CONV))) {
 103                 cfs_waitq_signal(&lock->l_waitq);
 104                 RETURN(0);
 105         }
 106
 107         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
 108                    "sleeping");
 109         ldlm_lock_dump(D_OTHER, lock, 0);
 110         ldlm_reprocess_all(lock->l_resource);
 111
 112 noreproc:
 113
 114         obd = class_exp2obd(lock->l_conn_export);
 115
 116         /* if this is a local lock, then there is no import */
 117         if (obd != NULL)
 118                 imp = obd->u.cli.cl_import;
 119
 120         lwd.lwd_lock = lock;
 121
 122         if (lock->l_flags & LDLM_FL_NO_TIMEOUT) {
 123                 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
 124                 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
 125         } else {
 126                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
 127                                        ldlm_expired_completion_wait,
 128                                        interrupted_completion_wait, &lwd);
 129         }
 130
 131         if (imp != NULL) {
 132                 spin_lock(&imp->imp_lock);
 133                 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
 134                 spin_unlock(&imp->imp_lock);
 135         }
 136
 137         /* Go to sleep until the lock is granted or cancelled. */
 138         rc = l_wait_event(lock->l_waitq,
 139                           ((lock->l_req_mode == lock->l_granted_mode) ||
 140                            (lock->l_flags & LDLM_FL_FAILED)), &lwi);
 141
 142         if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED) {
 143                 LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed");
 144                 RETURN(-EIO);
 145         }
 146
 147         if (rc) {
 148                 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
 149                            rc);
 150                 RETURN(rc);
 151         }
 152
 153         LDLM_DEBUG(lock, "client-side enqueue waking up: granted");
 154         RETURN(0);
 155 }
 156
 157 /*
 158  * ->l_blocking_ast() callback for LDLM locks acquired by server-side OBDs.
 159  */
 160 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 161                       void *data, int flag)
 162 {
 163         int do_ast;
 164         ENTRY;
 165
 166         if (flag == LDLM_CB_CANCELING) {
 167                 /* Don't need to do anything here. */
 168                 RETURN(0);
 169         }
 170
 171         lock_res_and_lock(lock);
 172         /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
 173          * that ldlm_blocking_ast is called just before intent_policy method
 174          * takes the ns_lock, then by the time we get the lock, we might not
 175          * be the correct blocking function anymore.  So check, and return
 176          * early, if so. */
 177         if (lock->l_blocking_ast != ldlm_blocking_ast) {
 178                 unlock_res_and_lock(lock);
 179                 RETURN(0);
 180         }
 181
 182         lock->l_flags |= LDLM_FL_CBPENDING;
 183         do_ast = (!lock->l_readers && !lock->l_writers);
 184         unlock_res_and_lock(lock);
 185
 186         if (do_ast) {
 187                 struct lustre_handle lockh;
 188                 int rc;
 189
 190                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
 191                 ldlm_lock2handle(lock, &lockh);
 192                 rc = ldlm_cli_cancel(&lockh);
 193                 if (rc < 0)
 194                         CERROR("ldlm_cli_cancel: %d\n", rc);
 195         } else {
 196                 LDLM_DEBUG(lock, "Lock still has references, will be "
 197                            "cancelled later");
 198         }
 199         RETURN(0);
 200 }
 201
 202 /*
 203  * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See
 204  * comment in filter_intent_policy() on why you may need this.
 205  */
 206 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
 207 {
 208         /*
 209          * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for
 210          * that is rather subtle: with OST-side locking, it may so happen that
 211          * _all_ extent locks are held by the OST. If client wants to obtain
 212          * current file size it calls ll{,u}_glimpse_size(), and (as locks are
 213          * on the server), dummy glimpse callback fires and does
 214          * nothing. Client still receives correct file size due to the
 215          * following fragment in filter_intent_policy():
 216          *
 217          * rc = l->l_glimpse_ast(l, NULL); // this will update the LVB
 218          * if (rc != 0 && res->lr_namespace->ns_lvbo &&
 219          *     res->lr_namespace->ns_lvbo->lvbo_update) {
 220          *         res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1);
 221          * }
 222          *
 223          * that is, after glimpse_ast() fails, filter_lvbo_update() runs, and
 224          * returns correct file size to the client.
 225          */
 226         return -ELDLM_NO_LOCK_DATA;
 227 }
 228
 229 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
 230                            const struct ldlm_res_id *res_id,
 231                            ldlm_type_t type, ldlm_policy_data_t *policy,
 232                            ldlm_mode_t mode, int *flags,
 233                            ldlm_blocking_callback blocking,
 234                            ldlm_completion_callback completion,
 235                            ldlm_glimpse_callback glimpse,
 236                            void *data, __u32 lvb_len, void *lvb_swabber,
 237                            struct lustre_handle *lockh)
 238 {
 239         struct ldlm_lock *lock;
 240         int err;
 241         ENTRY;
 242
 243         LASSERT(!(*flags & LDLM_FL_REPLAY));
 244         if (unlikely(ns->ns_client)) {
 245                 CERROR("Trying to enqueue local lock in a shadow namespace\n");
 246                 LBUG();
 247         }
 248
 249         lock = ldlm_lock_create(ns, NULL, res_id, type, mode, blocking,
 250                                 completion, glimpse, data, lvb_len);
 251         if (unlikely(!lock))
 252                 GOTO(out_nolock, err = -ENOMEM);
 253         LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
 254
 255         ldlm_lock_addref_internal(lock, mode);
 256         ldlm_lock2handle(lock, lockh);
 257         lock_res_and_lock(lock);
 258         lock->l_flags |= LDLM_FL_LOCAL;
 259         if (*flags & LDLM_FL_ATOMIC_CB)
 260                 lock->l_flags |= LDLM_FL_ATOMIC_CB;
 261         lock->l_lvb_swabber = lvb_swabber;
 262         unlock_res_and_lock(lock);
 263         if (policy != NULL)
 264                 lock->l_policy_data = *policy;
 265         if (type == LDLM_EXTENT)
 266                 lock->l_req_extent = policy->l_extent;
 267
 268         err = ldlm_lock_enqueue(ns, &lock, policy, flags);
 269         if (unlikely(err != ELDLM_OK))
 270                 GOTO(out, err);
 271
 272         if (policy != NULL)
 273                 *policy = lock->l_policy_data;
 274
 275         LDLM_DEBUG_NOLOCK("client-side local enqueue handler END (lock %p)",
 276                           lock);
 277
 278         if (lock->l_completion_ast)
 279                 lock->l_completion_ast(lock, *flags, NULL);
 280
 281         LDLM_DEBUG(lock, "client-side local enqueue END");
 282         EXIT;
 283  out:
 284         LDLM_LOCK_PUT(lock);
 285  out_nolock:
 286         return err;
 287 }
 288
 289 static void failed_lock_cleanup(struct ldlm_namespace *ns,
 290                                 struct ldlm_lock *lock,
 291                                 struct lustre_handle *lockh, int mode)
 292 {
 293         /* Set a flag to prevent us from sending a CANCEL (bug 407) */
 294         lock_res_and_lock(lock);
 295         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 296         unlock_res_and_lock(lock);
 297         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 298
 299         ldlm_lock_decref_and_cancel(lockh, mode);
 300
 301         /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
 302          *       from llite/file.c/ll_file_flock(). */
 303         if (lock->l_resource->lr_type == LDLM_FLOCK) {
 304                 ldlm_lock_destroy(lock);
 305         }
 306 }
 307
 308 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
 309                           ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
 310                           int *flags, void *lvb, __u32 lvb_len,
 311                           void *lvb_swabber, struct lustre_handle *lockh,int rc)
 312 {
 313         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 314         int is_replay = *flags & LDLM_FL_REPLAY;
 315         struct ldlm_lock *lock;
 316         struct ldlm_reply *reply;
 317         int cleanup_phase = 1;
 318         ENTRY;
 319
 320         lock = ldlm_handle2lock(lockh);
 321         /* ldlm_cli_enqueue is holding a reference on this lock. */
 322         LASSERT(lock != NULL);
 323         if (rc != ELDLM_OK) {
 324                 LASSERT(!is_replay);
 325                 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
 326                            rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
 327                 if (rc == ELDLM_LOCK_ABORTED) {
 328                         /* Before we return, swab the reply */
 329                         reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF,
 330                                                    sizeof(*reply),
 331                                                    lustre_swab_ldlm_reply);
 332                         if (reply == NULL) {
 333                                 CERROR("Can't unpack ldlm_reply\n");
 334                                 rc = -EPROTO;
 335                         }
 336                         if (lvb_len) {
 337                                 void *tmplvb;
 338                                 tmplvb = lustre_swab_repbuf(req,
 339                                                             DLM_REPLY_REC_OFF,
 340                                                             lvb_len,
 341                                                             lvb_swabber);
 342                                 if (tmplvb == NULL)
 343                                         GOTO(cleanup, rc = -EPROTO);
 344                                 if (lvb != NULL)
 345                                         memcpy(lvb, tmplvb, lvb_len);
 346                         }
 347                 }
 348                 GOTO(cleanup, rc);
 349         }
 350
 351         reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF, sizeof(*reply),
 352                                    lustre_swab_ldlm_reply);
 353         if (reply == NULL) {
 354                 CERROR("Can't unpack ldlm_reply\n");
 355                 GOTO(cleanup, rc = -EPROTO);
 356         }
 357
 358         /* lock enqueued on the server */
 359         cleanup_phase = 0;
 360
 361         lock_res_and_lock(lock);
 362         lock->l_remote_handle = reply->lock_handle;
 363         *flags = reply->lock_flags;
 364         lock->l_flags |= reply->lock_flags & LDLM_INHERIT_FLAGS;
 365         /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match()
 366          * to wait with no timeout as well */
 367         lock->l_flags |= reply->lock_flags & LDLM_FL_NO_TIMEOUT;
 368         unlock_res_and_lock(lock);
 369
 370         CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%x\n",
 371                lock, reply->lock_handle.cookie, *flags);
 372
 373         /* If enqueue returned a blocked lock but the completion handler has
 374          * already run, then it fixed up the resource and we don't need to do it
 375          * again. */
 376         if ((*flags) & LDLM_FL_LOCK_CHANGED) {
 377                 int newmode = reply->lock_desc.l_req_mode;
 378                 LASSERT(!is_replay);
 379                 if (newmode && newmode != lock->l_req_mode) {
 380                         LDLM_DEBUG(lock, "server returned different mode %s",
 381                                    ldlm_lockname[newmode]);
 382                         lock->l_req_mode = newmode;
 383                 }
 384
 385                 if (memcmp(reply->lock_desc.l_resource.lr_name.name,
 386                           lock->l_resource->lr_name.name,
 387                           sizeof(struct ldlm_res_id))) {
 388                         CDEBUG(D_INFO, "remote intent success, locking "
 389                                         "(%ld,%ld,%ld) instead of "
 390                                         "(%ld,%ld,%ld)\n",
 391                               (long)reply->lock_desc.l_resource.lr_name.name[0],
 392                               (long)reply->lock_desc.l_resource.lr_name.name[1],
 393                               (long)reply->lock_desc.l_resource.lr_name.name[2],
 394                               (long)lock->l_resource->lr_name.name[0],
 395                               (long)lock->l_resource->lr_name.name[1],
 396                               (long)lock->l_resource->lr_name.name[2]);
 397
 398                         ldlm_lock_change_resource(ns, lock,
 399                                           &reply->lock_desc.l_resource.lr_name);
 400                         if (lock->l_resource == NULL) {
 401                                 LBUG();
 402                                 GOTO(cleanup, rc = -ENOMEM);
 403                         }
 404                         LDLM_DEBUG(lock, "client-side enqueue, new resource");
 405                 }
 406                 if (with_policy)
 407                         if (!(type == LDLM_IBITS && !(exp->exp_connect_flags &
 408                                                     OBD_CONNECT_IBITS)))
 409                                 lock->l_policy_data =
 410                                                  reply->lock_desc.l_policy_data;
 411                 if (type != LDLM_PLAIN)
 412                         LDLM_DEBUG(lock,"client-side enqueue, new policy data");
 413         }
 414
 415         if ((*flags) & LDLM_FL_AST_SENT ||
 416             /* Cancel extent locks as soon as possible on a liblustre client,
 417              * because it cannot handle asynchronous ASTs robustly (see
 418              * bug 7311). */
 419             (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) {
 420                 lock_res_and_lock(lock);
 421                 lock->l_flags |= LDLM_FL_CBPENDING;
 422                 unlock_res_and_lock(lock);
 423                 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
 424         }
 425
 426         /* If the lock has already been granted by a completion AST, don't
 427          * clobber the LVB with an older one. */
 428         if (lvb_len && (lock->l_req_mode != lock->l_granted_mode)) {
 429                 void *tmplvb;
 430                 tmplvb = lustre_swab_repbuf(req, DLM_REPLY_REC_OFF, lvb_len,
 431                                             lvb_swabber);
 432                 if (tmplvb == NULL)
 433                         GOTO(cleanup, rc = -EPROTO);
 434                 memcpy(lock->l_lvb_data, tmplvb, lvb_len);
 435         }
 436
 437         if (!is_replay) {
 438                 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
 439                 if (lock->l_completion_ast != NULL) {
 440                         int err = lock->l_completion_ast(lock, *flags, NULL);
 441                         if (!rc)
 442                                 rc = err;
 443                         if (rc && type != LDLM_FLOCK) /* bug 9425, bug 10250 */
 444                                 cleanup_phase = 1;
 445                 }
 446         }
 447
 448         if (lvb_len && lvb != NULL) {
 449                 /* Copy the LVB here, and not earlier, because the completion
 450                  * AST (if any) can override what we got in the reply */
 451                 memcpy(lvb, lock->l_lvb_data, lvb_len);
 452         }
 453
 454         LDLM_DEBUG(lock, "client-side enqueue END");
 455         EXIT;
 456 cleanup:
 457         if (cleanup_phase == 1 && rc)
 458                 failed_lock_cleanup(ns, lock, lockh, mode);
 459         /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
 460         LDLM_LOCK_PUT(lock);
 461         LDLM_LOCK_PUT(lock);
 462         return rc;
 463 }
 464
 465 /* If a request has some specific initialisation it is passed in @reqp,
 466  * otherwise it is created in ldlm_cli_enqueue.
 467  *
 468  * Supports sync and async requests, pass @async flag accordingly. If a
 469  * request was created in ldlm_cli_enqueue and it is the async request,
 470  * pass it to the caller in @reqp. */
 471 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
 472                      const struct ldlm_res_id *res_id,
 473                      ldlm_type_t type, ldlm_policy_data_t *policy,
 474                      ldlm_mode_t mode, int *flags,
 475                      ldlm_blocking_callback blocking,
 476                      ldlm_completion_callback completion,
 477                      ldlm_glimpse_callback glimpse,
 478                      void *data, void *lvb, __u32 lvb_len, void *lvb_swabber,
 479                      struct lustre_handle *lockh, int async)
 480 {
 481         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 482         struct ldlm_lock *lock;
 483         struct ldlm_request *body;
 484         struct ldlm_reply *reply;
 485         int size[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
 486                         [DLM_LOCKREQ_OFF]     = sizeof(*body),
 487                         [DLM_REPLY_REC_OFF]   = lvb_len };
 488         int is_replay = *flags & LDLM_FL_REPLAY;
 489         int req_passed_in = 1, rc;
 490         struct ptlrpc_request *req;
 491         ENTRY;
 492
 493         LASSERT(exp != NULL);
 494
 495         /* If we're replaying this lock, just check some invariants.
 496          * If we're creating a new lock, get everything all setup nice. */
 497         if (is_replay) {
 498                 lock = ldlm_handle2lock(lockh);
 499                 LASSERT(lock != NULL);
 500                 LDLM_DEBUG(lock, "client-side enqueue START");
 501                 LASSERT(exp == lock->l_conn_export);
 502         } else {
 503                 lock = ldlm_lock_create(ns, NULL, res_id, type, mode, blocking,
 504                                         completion, glimpse, data, lvb_len);
 505                 if (lock == NULL)
 506                         RETURN(-ENOMEM);
 507                 /* for the local lock, add the reference */
 508                 ldlm_lock_addref_internal(lock, mode);
 509                 ldlm_lock2handle(lock, lockh);
 510                 lock->l_lvb_swabber = lvb_swabber;
 511                 if (policy != NULL) {
 512                         /* INODEBITS_INTEROP: If the server does not support
 513                          * inodebits, we will request a plain lock in the
 514                          * descriptor (ldlm_lock2desc() below) but use an
 515                          * inodebits lock internally with both bits set.
 516                          */
 517                         if (type == LDLM_IBITS && !(exp->exp_connect_flags &
 518                                                     OBD_CONNECT_IBITS))
 519                                 lock->l_policy_data.l_inodebits.bits =
 520                                         MDS_INODELOCK_LOOKUP |
 521                                         MDS_INODELOCK_UPDATE;
 522                         else
 523                                 lock->l_policy_data = *policy;
 524                 }
 525
 526                 if (type == LDLM_EXTENT)
 527                         lock->l_req_extent = policy->l_extent;
 528                 LDLM_DEBUG(lock, "client-side enqueue START");
 529         }
 530
 531         /* lock not sent to server yet */
 532
 533         if (reqp == NULL || *reqp == NULL) {
 534                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
 535                                       LDLM_ENQUEUE, 2, size, NULL);
 536                 if (req == NULL) {
 537                         failed_lock_cleanup(ns, lock, lockh, mode);
 538                         LDLM_LOCK_PUT(lock);
 539                         RETURN(-ENOMEM);
 540                 }
 541                 req_passed_in = 0;
 542                 if (reqp)
 543                         *reqp = req;
 544         } else {
 545                 req = *reqp;
 546                 LASSERTF(lustre_msg_buflen(req->rq_reqmsg, DLM_LOCKREQ_OFF) ==
 547                          sizeof(*body), "buflen[%d] = %d, not "LPSZ"\n",
 548                          DLM_LOCKREQ_OFF,
 549                          lustre_msg_buflen(req->rq_reqmsg, DLM_LOCKREQ_OFF),
 550                          sizeof(*body));
 551         }
 552
 553         lock->l_conn_export = exp;
 554         lock->l_export = NULL;
 555         lock->l_blocking_ast = blocking;
 556
 557         /* Dump lock data into the request buffer */
 558         body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, sizeof(*body));
 559         ldlm_lock2desc(lock, &body->lock_desc);
 560         body->lock_flags = *flags;
 561         body->lock_handle1 = *lockh;
 562
 563         /* Continue as normal. */
 564         if (!req_passed_in) {
 565                 size[DLM_LOCKREPLY_OFF] = sizeof(*reply);
 566                 ptlrpc_req_set_repsize(req, 2 + (lvb_len > 0), size);
 567         }
 568
 569         /*
 570          * Liblustre client doesn't get extent locks, except for O_APPEND case
 571          * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
 572          * [i_size, OBD_OBJECT_EOF] lock is taken.
 573          */
 574         LASSERT(ergo(LIBLUSTRE_CLIENT, type != LDLM_EXTENT ||
 575                      policy->l_extent.end == OBD_OBJECT_EOF));
 576
 577         if (async) {
 578                 LASSERT(reqp != NULL);
 579                 RETURN(0);
 580         }
 581
 582         LDLM_DEBUG(lock, "sending request");
 583         rc = ptlrpc_queue_wait(req);
 584         rc = ldlm_cli_enqueue_fini(exp, req, type, policy ? 1 : 0,
 585                                    mode, flags, lvb, lvb_len, lvb_swabber,
 586                                    lockh, rc);
 587
 588         if (!req_passed_in && req != NULL) {
 589                 ptlrpc_req_finished(req);
 590                 if (reqp)
 591                         *reqp = NULL;
 592         }
 593
 594         RETURN(rc);
 595 }
 596
 597 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
 598                                   int *flags)
 599 {
 600         struct ldlm_resource *res;
 601         int rc;
 602         ENTRY;
 603         if (lock->l_resource->lr_namespace->ns_client) {
 604                 CERROR("Trying to cancel local lock\n");
 605                 LBUG();
 606         }
 607         LDLM_DEBUG(lock, "client-side local convert");
 608
 609         res = ldlm_lock_convert(lock, new_mode, flags);
 610         if (res) {
 611                 ldlm_reprocess_all(res);
 612                 rc = 0;
 613         } else {
 614                 rc = EDEADLOCK;
 615         }
 616         LDLM_DEBUG(lock, "client-side local convert handler END");
 617         LDLM_LOCK_PUT(lock);
 618         RETURN(rc);
 619 }
 620
 621 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
 622  * conversion of locks which are on the waiting or converting queue */
 623 /* Caller of this code is supposed to take care of lock readers/writers
 624    accounting */
 625 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, int *flags)
 626 {
 627         struct ldlm_request *body;
 628         struct ldlm_reply *reply;
 629         struct ldlm_lock *lock;
 630         struct ldlm_resource *res;
 631         struct ptlrpc_request *req;
 632         int size[2] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
 633                         [DLM_LOCKREQ_OFF]     = sizeof(*body) };
 634         int rc;
 635         ENTRY;
 636
 637         lock = ldlm_handle2lock(lockh);
 638         if (!lock) {
 639                 LBUG();
 640                 RETURN(-EINVAL);
 641         }
 642         *flags = 0;
 643
 644         if (lock->l_conn_export == NULL)
 645                 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
 646
 647         LDLM_DEBUG(lock, "client-side convert");
 648
 649         req = ptlrpc_prep_req(class_exp2cliimp(lock->l_conn_export),
 650                               LUSTRE_DLM_VERSION, LDLM_CONVERT, 2, size, NULL);
 651         if (!req)
 652                 GOTO(out, rc = -ENOMEM);
 653
 654         body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, sizeof(*body));
 655         body->lock_handle1 = lock->l_remote_handle;
 656
 657         body->lock_desc.l_req_mode = new_mode;
 658         body->lock_flags = *flags;
 659
 660         size[DLM_LOCKREPLY_OFF] = sizeof(*reply);
 661         ptlrpc_req_set_repsize(req, 2, size);
 662
 663         rc = ptlrpc_queue_wait(req);
 664         if (rc != ELDLM_OK)
 665                 GOTO(out, rc);
 666
 667         reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF, sizeof(*reply),
 668                                    lustre_swab_ldlm_reply);
 669         if (reply == NULL) {
 670                 CERROR ("Can't unpack ldlm_reply\n");
 671                 GOTO (out, rc = -EPROTO);
 672         }
 673
 674         if (req->rq_status)
 675                 GOTO(out, rc = req->rq_status);
 676
 677         res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
 678         if (res != NULL) {
 679                 ldlm_reprocess_all(res);
 680                 /* Go to sleep until the lock is granted. */
 681                 /* FIXME: or cancelled. */
 682                 if (lock->l_completion_ast) {
 683                         rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
 684                                                     NULL);
 685                         if (rc)
 686                                 GOTO(out, rc);
 687                 }
 688         } else {
 689                 rc = EDEADLOCK;
 690         }
 691         EXIT;
 692  out:
 693         LDLM_LOCK_PUT(lock);
 694         ptlrpc_req_finished(req);
 695         return rc;
 696 }
 697
 698 int ldlm_cli_cancel(struct lustre_handle *lockh)
 699 {
 700         struct ptlrpc_request *req;
 701         struct ldlm_lock *lock;
 702         struct ldlm_request *body;
 703         int size[2] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
 704                         [DLM_LOCKREQ_OFF]     = sizeof(*body) };
 705         int rc = 0;
 706         ENTRY;
 707
 708         /* concurrent cancels on the same handle can happen */
 709         lock = __ldlm_handle2lock(lockh, LDLM_FL_CANCELING);
 710         if (lock == NULL)
 711                 RETURN(0);
 712
 713         if (lock->l_conn_export) {
 714                 int local_only;
 715                 struct obd_import *imp;
 716
 717                 LDLM_DEBUG(lock, "client-side cancel");
 718                 /* Set this flag to prevent others from getting new references*/
 719                 lock_res_and_lock(lock);
 720                 lock->l_flags |= LDLM_FL_CBPENDING;
 721                 local_only = (lock->l_flags &
 722                               (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
 723                 ldlm_cancel_callback(lock);
 724                 unlock_res_and_lock(lock);
 725
 726                 if (local_only) {
 727                         CDEBUG(D_INFO, "not sending request (at caller's "
 728                                "instruction)\n");
 729                         goto local_cancel;
 730                 }
 731
 732         restart:
 733                 imp = class_exp2cliimp(lock->l_conn_export);
 734                 if (imp == NULL || imp->imp_invalid) {
 735                         CDEBUG(D_HA, "skipping cancel on invalid import %p\n",
 736                                imp);
 737                         goto local_cancel;
 738                 }
 739
 740                 req = ptlrpc_prep_req(imp, LUSTRE_DLM_VERSION, LDLM_CANCEL, 2,
 741                                       size, NULL);
 742                 if (!req)
 743                         GOTO(out, rc = -ENOMEM);
 744                 req->rq_no_resend = 1;
 745
 746                 /* XXX FIXME bug 249 */
 747                 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
 748                 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
 749
 750                 body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF,
 751                                       sizeof(*body));
 752                 body->lock_handle1 = lock->l_remote_handle;
 753
 754                 ptlrpc_req_set_repsize(req, 1, NULL);
 755                 rc = ptlrpc_queue_wait(req);
 756
 757                 if (rc == ESTALE) {
 758                         CDEBUG(D_DLMTRACE, "client/server (nid %s) out of sync "
 759                                "-- not fatal, flags %x\n",
 760                                libcfs_nid2str(req->rq_import->
 761                                               imp_connection->c_peer.nid),
 762                                lock->l_flags);
 763                 } else if (rc == -ETIMEDOUT) {
 764                         ptlrpc_req_finished(req);
 765                         GOTO(restart, rc);
 766                 } else if (rc != ELDLM_OK) {
 767                         CERROR("Got rc %d from cancel RPC: canceling "
 768                                "anyway\n", rc);
 769                 }
 770
 771                 ptlrpc_req_finished(req);
 772         local_cancel:
 773                 ldlm_lock_cancel(lock);
 774         } else {
 775                 if (lock->l_resource->lr_namespace->ns_client) {
 776                         LDLM_ERROR(lock, "Trying to cancel local lock");
 777                         LBUG();
 778                 }
 779                 LDLM_DEBUG(lock, "client-side local cancel");
 780                 ldlm_lock_cancel(lock);
 781                 ldlm_reprocess_all(lock->l_resource);
 782                 LDLM_DEBUG(lock, "client-side local cancel handler END");
 783         }
 784
 785         EXIT;
 786  out:
 787         LDLM_LOCK_PUT(lock);
 788         return rc;
 789 }
 790
 791 /* when called with LDLM_ASYNC the blocking callback will be handled
 792  * in a thread and this function will return after the thread has been
 793  * asked to call the callback.  when called with LDLM_SYNC the blocking
 794  * callback will be performed in this function. */
 795 int ldlm_cancel_lru(struct ldlm_namespace *ns, ldlm_sync_t sync)
 796 {
 797         struct ldlm_lock *lock, *next;
 798         int count, rc = 0;
 799         CFS_LIST_HEAD(cblist);
 800         ENTRY;
 801
 802 #ifndef __KERNEL__
 803         sync = LDLM_SYNC; /* force to be sync in user space */
 804 #endif
 805
 806         spin_lock(&ns->ns_unused_lock);
 807         count = ns->ns_nr_unused - ns->ns_max_unused;
 808
 809         if (count <= 0) {
 810                 spin_unlock(&ns->ns_unused_lock);
 811                 RETURN(0);
 812         }
 813
 814         while (!list_empty(&ns->ns_unused_list)) {
 815                 struct list_head *tmp = ns->ns_unused_list.next;
 816                 lock = list_entry(tmp, struct ldlm_lock, l_lru);
 817                 LASSERT(!lock->l_readers && !lock->l_writers);
 818
 819                 LDLM_LOCK_GET(lock); /* dropped by bl thread */
 820                 spin_unlock(&ns->ns_unused_lock);
 821
 822                 lock_res_and_lock(lock);
 823                 if (ldlm_lock_remove_from_lru(lock) == 0) {
 824                         /* other thread is removing lock from lru */
 825                         unlock_res_and_lock(lock);
 826                         LDLM_LOCK_PUT(lock);
 827                         spin_lock(&ns->ns_unused_lock);
 828                         continue;
 829
 830                 }
 831
 832                 /* If we have chosen to canecl this lock voluntarily, we better
 833                    send cancel notification to server, so that it frees
 834                    appropriate state. This might lead to a race where while
 835                    we are doing cancel here, server is also silently
 836                    cancelling this lock. */
 837                 lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK;
 838
 839                 /* Setting the CBPENDING flag is a little misleading, but
 840                  * prevents an important race; namely, once CBPENDING is set,
 841                  * the lock can accumulate no more readers/writers.  Since
 842                  * readers and writers are already zero here, ldlm_lock_decref
 843                  * won't see this flag and call l_blocking_ast */
 844                 lock->l_flags |= LDLM_FL_CBPENDING;
 845
 846                 /* We can't re-add to l_lru as it confuses the refcounting in
 847                  * ldlm_lock_remove_from_lru() if an AST arrives after we drop
 848                  * ns_lock below. We use l_tmp and can't use l_pending_chain as
 849                  * it is used both on server and client nevertheles bug 5666
 850                  * says it is used only on server. --umka */
 851                 list_add(&lock->l_tmp, &cblist);
 852                 unlock_res_and_lock(lock);
 853
 854                 LDLM_LOCK_GET(lock); /* to hold lock after bl thread */
 855                 if (sync == LDLM_ASYNC && (ldlm_bl_to_thread(ns, NULL, lock) == 0)) {
 856                         lock_res_and_lock(lock);
 857                         list_del_init(&lock->l_tmp);
 858                         unlock_res_and_lock(lock);
 859                 }
 860                 LDLM_LOCK_PUT(lock);
 861
 862                 spin_lock(&ns->ns_unused_lock);
 863
 864                 if (--count == 0)
 865                         break;
 866         }
 867         spin_unlock(&ns->ns_unused_lock);
 868
 869         list_for_each_entry_safe(lock, next, &cblist, l_tmp) {
 870                 list_del_init(&lock->l_tmp);
 871                 ldlm_handle_bl_callback(ns, NULL, lock);
 872         }
 873
 874         RETURN(rc);
 875 }
 876
 877 static int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
 878                                            const struct ldlm_res_id *res_id,
 879                                            int flags, void *opaque)
 880 {
 881         struct list_head *tmp, *next, list = CFS_LIST_HEAD_INIT(list);
 882         struct ldlm_resource *res;
 883         struct ldlm_lock *lock;
 884         ENTRY;
 885
 886         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
 887         if (res == NULL) {
 888                 /* This is not a problem. */
 889                 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id->name[0]);
 890                 RETURN(0);
 891         }
 892
 893         lock_res(res);
 894         list_for_each(tmp, &res->lr_granted) {
 895                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 896
 897                 if (opaque != NULL && lock->l_ast_data != opaque) {
 898                         LDLM_ERROR(lock, "data %p doesn't match opaque %p",
 899                                    lock->l_ast_data, opaque);
 900                         //LBUG();
 901                         continue;
 902                 }
 903
 904                 if (lock->l_readers || lock->l_writers) {
 905                         if (flags & LDLM_FL_WARN) {
 906                                 LDLM_ERROR(lock, "lock in use");
 907                                 //LBUG();
 908                         }
 909                         continue;
 910                 }
 911
 912                 /* See CBPENDING comment in ldlm_cancel_lru */
 913                 lock->l_flags |= LDLM_FL_CBPENDING;
 914
 915                 LASSERT(list_empty(&lock->l_bl_ast));
 916                 list_add(&lock->l_bl_ast, &list);
 917                 LDLM_LOCK_GET(lock);
 918         }
 919         unlock_res(res);
 920
 921         list_for_each_safe(tmp, next, &list) {
 922                 struct lustre_handle lockh;
 923                 int rc;
 924                 lock = list_entry(tmp, struct ldlm_lock, l_bl_ast);
 925
 926                 if (flags & LDLM_FL_LOCAL_ONLY) {
 927                         ldlm_lock_cancel(lock);
 928                 } else {
 929                         ldlm_lock2handle(lock, &lockh);
 930                         rc = ldlm_cli_cancel(&lockh);
 931                         if (rc != ELDLM_OK)
 932                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 933                 }
 934                 list_del_init(&lock->l_bl_ast);
 935                 LDLM_LOCK_PUT(lock);
 936         }
 937
 938         ldlm_resource_putref(res);
 939
 940         RETURN(0);
 941 }
 942
 943 static inline int have_no_nsresource(struct ldlm_namespace *ns)
 944 {
 945         int no_resource = 0;
 946
 947         spin_lock(&ns->ns_hash_lock);
 948         if (ns->ns_resources == 0)
 949                 no_resource = 1;
 950         spin_unlock(&ns->ns_hash_lock);
 951
 952         RETURN(no_resource);
 953 }
 954
 955 /* Cancel all locks on a namespace (or a specific resource, if given)
 956  * that have 0 readers/writers.
 957  *
 958  * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
 959  * to notify the server.
 960  * If flags & LDLM_FL_WARN, print a warning if some locks are still in use. */
 961 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
 962                            const struct ldlm_res_id *res_id,
 963                            int flags, void *opaque)
 964 {
 965         int i;
 966         ENTRY;
 967
 968         if (ns == NULL)
 969                 RETURN(ELDLM_OK);
 970
 971         if (res_id)
 972                 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, flags,
 973                                                        opaque));
 974
 975         spin_lock(&ns->ns_hash_lock);
 976         for (i = 0; i < RES_HASH_SIZE; i++) {
 977                 struct list_head *tmp;
 978                 tmp = ns->ns_hash[i].next;
 979                 while (tmp != &(ns->ns_hash[i])) {
 980                         struct ldlm_resource *res;
 981                         int rc;
 982
 983                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
 984                         ldlm_resource_getref(res);
 985                         spin_unlock(&ns->ns_hash_lock);
 986
 987                         rc = ldlm_cli_cancel_unused_resource(ns, &res->lr_name,
 988                                                              flags, opaque);
 989
 990                         if (rc)
 991                                 CERROR("cancel_unused_res ("LPU64"): %d\n",
 992                                        res->lr_name.name[0], rc);
 993
 994                         spin_lock(&ns->ns_hash_lock);
 995                         tmp = tmp->next;
 996                         ldlm_resource_putref_locked(res);
 997                 }
 998         }
 999         spin_unlock(&ns->ns_hash_lock);
1000
1001         RETURN(ELDLM_OK);
1002 }
1003
1004 /* join/split resource locks to/from lru list */
1005 int ldlm_cli_join_lru(struct ldlm_namespace *ns,
1006                       const struct ldlm_res_id *res_id, int join)
1007 {
1008         struct ldlm_resource *res;
1009         struct ldlm_lock *lock, *n;
1010         int count = 0;
1011         ENTRY;
1012
1013         LASSERT(ns->ns_client == LDLM_NAMESPACE_CLIENT);
1014
1015         res = ldlm_resource_get(ns, NULL, res_id, LDLM_EXTENT, 0);
1016         if (res == NULL)
1017                 RETURN(count);
1018         LASSERT(res->lr_type == LDLM_EXTENT);
1019
1020         lock_res(res);
1021         if (!join)
1022                 goto split;
1023
1024         list_for_each_entry_safe (lock, n, &res->lr_granted, l_res_link) {
1025                 if (list_empty(&lock->l_lru) &&
1026                     !lock->l_readers && !lock->l_writers &&
1027                     !(lock->l_flags & LDLM_FL_LOCAL) &&
1028                     !(lock->l_flags & LDLM_FL_CBPENDING)) {
1029                         spin_lock(&ns->ns_unused_lock);
1030                         LASSERT(ns->ns_nr_unused >= 0);
1031                         list_add_tail(&lock->l_lru, &ns->ns_unused_list);
1032                         ns->ns_nr_unused++;
1033                         spin_unlock(&ns->ns_unused_lock);
1034                         lock->l_flags &= ~LDLM_FL_NO_LRU;
1035                         LDLM_DEBUG(lock, "join lock to lru");
1036                         count++;
1037                 }
1038         }
1039         goto unlock;
1040 split:
1041         spin_lock(&ns->ns_unused_lock);
1042         list_for_each_entry_safe (lock, n, &ns->ns_unused_list, l_lru) {
1043                 if (lock->l_resource == res) {
1044                         ldlm_lock_remove_from_lru_nolock(lock);
1045                         lock->l_flags |= LDLM_FL_NO_LRU;
1046                         LDLM_DEBUG(lock, "split lock from lru");
1047                         count++;
1048                 }
1049         }
1050         spin_unlock(&ns->ns_unused_lock);
1051 unlock:
1052         unlock_res(res);
1053         ldlm_resource_putref(res);
1054         RETURN(count);
1055 }
1056
1057 /* Lock iterators. */
1058
1059 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
1060                           void *closure)
1061 {
1062         struct list_head *tmp, *next;
1063         struct ldlm_lock *lock;
1064         int rc = LDLM_ITER_CONTINUE;
1065
1066         ENTRY;
1067
1068         if (!res)
1069                 RETURN(LDLM_ITER_CONTINUE);
1070
1071         lock_res(res);
1072         list_for_each_safe(tmp, next, &res->lr_granted) {
1073                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1074
1075                 if (iter(lock, closure) == LDLM_ITER_STOP)
1076                         GOTO(out, rc = LDLM_ITER_STOP);
1077         }
1078
1079         list_for_each_safe(tmp, next, &res->lr_converting) {
1080                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1081
1082                 if (iter(lock, closure) == LDLM_ITER_STOP)
1083                         GOTO(out, rc = LDLM_ITER_STOP);
1084         }
1085
1086         list_for_each_safe(tmp, next, &res->lr_waiting) {
1087                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1088
1089                 if (iter(lock, closure) == LDLM_ITER_STOP)
1090                         GOTO(out, rc = LDLM_ITER_STOP);
1091         }
1092  out:
1093         unlock_res(res);
1094         RETURN(rc);
1095 }
1096
1097 struct iter_helper_data {
1098         ldlm_iterator_t iter;
1099         void *closure;
1100 };
1101
1102 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
1103 {
1104         struct iter_helper_data *helper = closure;
1105         return helper->iter(lock, helper->closure);
1106 }
1107
1108 static int ldlm_res_iter_helper(struct ldlm_resource *res, void *closure)
1109 {
1110         return ldlm_resource_foreach(res, ldlm_iter_helper, closure);
1111 }
1112
1113 int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
1114                            void *closure)
1115 {
1116         struct iter_helper_data helper = { iter: iter, closure: closure };
1117         return ldlm_namespace_foreach_res(ns, ldlm_res_iter_helper, &helper);
1118 }
1119
1120 int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
1121                                ldlm_res_iterator_t iter, void *closure)
1122 {
1123         int i, rc = LDLM_ITER_CONTINUE;
1124         struct ldlm_resource *res;
1125         struct list_head *tmp;
1126
1127         ENTRY;
1128         spin_lock(&ns->ns_hash_lock);
1129         for (i = 0; i < RES_HASH_SIZE; i++) {
1130                 tmp = ns->ns_hash[i].next;
1131                 while (tmp != &(ns->ns_hash[i])) {
1132                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
1133                         ldlm_resource_getref(res);
1134                         spin_unlock(&ns->ns_hash_lock);
1135
1136                         rc = iter(res, closure);
1137
1138                         spin_lock(&ns->ns_hash_lock);
1139                         tmp = tmp->next;
1140                         ldlm_resource_putref_locked(res);
1141                         if (rc == LDLM_ITER_STOP)
1142                                 GOTO(out, rc);
1143                 }
1144         }
1145  out:
1146         spin_unlock(&ns->ns_hash_lock);
1147         RETURN(rc);
1148 }
1149
1150 /* non-blocking function to manipulate a lock whose cb_data is being put away.*/
1151 void ldlm_resource_iterate(struct ldlm_namespace *ns,
1152                            const struct ldlm_res_id *res_id,
1153                            ldlm_iterator_t iter, void *data)
1154 {
1155         struct ldlm_resource *res;
1156         ENTRY;
1157
1158         if (ns == NULL) {
1159                 CERROR("must pass in namespace\n");
1160                 LBUG();
1161         }
1162
1163         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1164         if (res == NULL) {
1165                 EXIT;
1166                 return;
1167         }
1168
1169         ldlm_resource_foreach(res, iter, data);
1170         ldlm_resource_putref(res);
1171         EXIT;
1172 }
1173
1174 /* Lock replay */
1175
1176 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
1177 {
1178         struct list_head *list = closure;
1179
1180         /* we use l_pending_chain here, because it's unused on clients. */
1181         LASSERTF(list_empty(&lock->l_pending_chain),"lock %p next %p prev %p\n",
1182                  lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
1183         /* bug 9573: don't replay locks left after eviction */
1184         if (!(lock->l_flags & LDLM_FL_FAILED))
1185                 list_add(&lock->l_pending_chain, list);
1186         return LDLM_ITER_CONTINUE;
1187 }
1188
1189 static int replay_lock_interpret(struct ptlrpc_request *req,
1190                                     void * data, int rc)
1191 {
1192         struct ldlm_lock *lock;
1193         struct ldlm_reply *reply;
1194
1195         ENTRY;
1196         atomic_dec(&req->rq_import->imp_replay_inflight);
1197         if (rc != ELDLM_OK)
1198                 GOTO(out, rc);
1199
1200         lock = req->rq_async_args.pointer_arg[0];
1201         LASSERT(lock != NULL);
1202
1203         reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF, sizeof(*reply),
1204                                    lustre_swab_ldlm_reply);
1205         if (reply == NULL) {
1206                 CERROR("Can't unpack ldlm_reply\n");
1207                 GOTO (out, rc = -EPROTO);
1208         }
1209
1210         lock->l_remote_handle = reply->lock_handle;
1211         LDLM_DEBUG(lock, "replayed lock:");
1212         ptlrpc_import_recovery_state_machine(req->rq_import);
1213  out:
1214         if (rc != ELDLM_OK)
1215                 ptlrpc_connect_import(req->rq_import, NULL);
1216
1217
1218         RETURN(rc);
1219 }
1220
1221 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
1222 {
1223         struct ptlrpc_request *req;
1224         struct ldlm_request *body;
1225         struct ldlm_reply *reply;
1226         int buffers = 2;
1227         int size[3] = { sizeof(struct ptlrpc_body) };
1228         int flags;
1229         ENTRY;
1230
1231         /* If this is reply-less callback lock, we cannot replay it, since
1232          * server might have long dropped it, but notification of that event was
1233          * lost by network. (and server granted conflicting lock already) */
1234         if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1235                 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
1236                 ldlm_lock_cancel(lock);
1237                 RETURN(0);
1238         }
1239         /*
1240          * If granted mode matches the requested mode, this lock is granted.
1241          *
1242          * If they differ, but we have a granted mode, then we were granted
1243          * one mode and now want another: ergo, converting.
1244          *
1245          * If we haven't been granted anything and are on a resource list,
1246          * then we're blocked/waiting.
1247          *
1248          * If we haven't been granted anything and we're NOT on a resource list,
1249          * then we haven't got a reply yet and don't have a known disposition.
1250          * This happens whenever a lock enqueue is the request that triggers
1251          * recovery.
1252          */
1253         if (lock->l_granted_mode == lock->l_req_mode)
1254                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
1255         else if (lock->l_granted_mode)
1256                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
1257         else if (!list_empty(&lock->l_res_link))
1258                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
1259         else
1260                 flags = LDLM_FL_REPLAY;
1261
1262         size[DLM_LOCKREQ_OFF] = sizeof(*body);
1263         req = ptlrpc_prep_req(imp, LUSTRE_DLM_VERSION, LDLM_ENQUEUE, 2, size,
1264                               NULL);
1265         if (!req)
1266                 RETURN(-ENOMEM);
1267
1268         /* We're part of recovery, so don't wait for it. */
1269         req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
1270
1271         body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, sizeof(*body));
1272         ldlm_lock2desc(lock, &body->lock_desc);
1273         body->lock_flags = flags;
1274
1275         ldlm_lock2handle(lock, &body->lock_handle1);
1276         size[DLM_LOCKREPLY_OFF] = sizeof(*reply);
1277         if (lock->l_lvb_len != 0) {
1278                 buffers = 3;
1279                 size[DLM_REPLY_REC_OFF] = lock->l_lvb_len;
1280         }
1281         ptlrpc_req_set_repsize(req, buffers, size);
1282         /* notify the server we've replayed all requests.
1283          * also, we mark the request to be put on a dedicated
1284          * queue to be processed after all request replayes.
1285          * bug 6063 */
1286         lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
1287
1288         LDLM_DEBUG(lock, "replaying lock:");
1289
1290         atomic_inc(&req->rq_import->imp_replay_inflight);
1291         req->rq_async_args.pointer_arg[0] = lock;
1292         req->rq_interpret_reply = replay_lock_interpret;
1293         ptlrpcd_add_req(req);
1294
1295         RETURN(0);
1296 }
1297
1298 int ldlm_replay_locks(struct obd_import *imp)
1299 {
1300         struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
1301         struct list_head list;
1302         struct ldlm_lock *lock, *next;
1303         int rc = 0;
1304
1305         ENTRY;
1306         CFS_INIT_LIST_HEAD(&list);
1307
1308         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1309
1310         /* ensure this doesn't fall to 0 before all have been queued */
1311         atomic_inc(&imp->imp_replay_inflight);
1312
1313         (void)ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
1314
1315         list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
1316                 list_del_init(&lock->l_pending_chain);
1317                 if (rc)
1318                         continue; /* or try to do the rest? */
1319                 rc = replay_one_lock(imp, lock);
1320         }
1321
1322         atomic_dec(&imp->imp_replay_inflight);
1323
1324         RETURN(rc);
1325 }