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 struct ldlm_async_args {
  47         struct lustre_handle lock_handle;
  48 };
  49
  50 int ldlm_expired_completion_wait(void *data)
  51 {
  52         struct lock_wait_data *lwd = data;
  53         struct ldlm_lock *lock = lwd->lwd_lock;
  54         struct obd_import *imp;
  55         struct obd_device *obd;
  56
  57         ENTRY;
  58         if (lock->l_conn_export == NULL) {
  59                 static cfs_time_t next_dump = 0, last_dump = 0;
  60
  61                 if (ptlrpc_check_suspend())
  62                         RETURN(0);
  63
  64                 LDLM_ERROR(lock, "lock timed out (enqueued at %lu, %lus ago); "
  65                            "not entering recovery in server code, just going "
  66                            "back to sleep", lock->l_enqueued_time.tv_sec,
  67                            CURRENT_SECONDS - lock->l_enqueued_time.tv_sec);
  68                 if (cfs_time_after(cfs_time_current(), next_dump)) {
  69                         last_dump = next_dump;
  70                         next_dump = cfs_time_shift(300);
  71                         ldlm_namespace_dump(D_DLMTRACE,
  72                                             lock->l_resource->lr_namespace);
  73                         if (last_dump == 0)
  74                                 libcfs_debug_dumplog();
  75                 }
  76                 RETURN(0);
  77         }
  78
  79         obd = lock->l_conn_export->exp_obd;
  80         imp = obd->u.cli.cl_import;
  81         ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
  82         LDLM_ERROR(lock, "lock timed out (enqueued at %lu, %lus ago), entering "
  83                    "recovery for %s@%s", lock->l_enqueued_time.tv_sec,
  84                    CURRENT_SECONDS - lock->l_enqueued_time.tv_sec,
  85                    obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
  86
  87         RETURN(0);
  88 }
  89
  90 int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data)
  91 {
  92         /* XXX ALLOCATE - 160 bytes */
  93         struct lock_wait_data lwd;
  94         struct obd_device *obd;
  95         struct obd_import *imp = NULL;
  96         struct l_wait_info lwi;
  97         int rc = 0;
  98         ENTRY;
  99
 100         if (flags == LDLM_FL_WAIT_NOREPROC) {
 101                 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
 102                 goto noreproc;
 103         }
 104
 105         if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
 106                        LDLM_FL_BLOCK_CONV))) {
 107                 cfs_waitq_signal(&lock->l_waitq);
 108                 RETURN(0);
 109         }
 110
 111         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
 112                    "sleeping");
 113         ldlm_lock_dump(D_OTHER, lock, 0);
 114         ldlm_reprocess_all(lock->l_resource);
 115
 116 noreproc:
 117
 118         obd = class_exp2obd(lock->l_conn_export);
 119
 120         /* if this is a local lock, then there is no import */
 121         if (obd != NULL)
 122                 imp = obd->u.cli.cl_import;
 123
 124         lwd.lwd_lock = lock;
 125
 126         if (lock->l_flags & LDLM_FL_NO_TIMEOUT) {
 127                 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
 128                 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
 129         } else {
 130                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
 131                                        ldlm_expired_completion_wait,
 132                                        interrupted_completion_wait, &lwd);
 133         }
 134
 135         if (imp != NULL) {
 136                 spin_lock(&imp->imp_lock);
 137                 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
 138                 spin_unlock(&imp->imp_lock);
 139         }
 140
 141         /* Go to sleep until the lock is granted or cancelled. */
 142         rc = l_wait_event(lock->l_waitq,
 143                           ((lock->l_req_mode == lock->l_granted_mode) ||
 144                            (lock->l_flags & LDLM_FL_FAILED)), &lwi);
 145
 146         if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED) {
 147                 LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed");
 148                 RETURN(-EIO);
 149         }
 150
 151         if (rc) {
 152                 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
 153                            rc);
 154                 RETURN(rc);
 155         }
 156
 157         LDLM_DEBUG(lock, "client-side enqueue waking up: granted");
 158         RETURN(0);
 159 }
 160
 161 /*
 162  * ->l_blocking_ast() callback for LDLM locks acquired by server-side OBDs.
 163  */
 164 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 165                       void *data, int flag)
 166 {
 167         int do_ast;
 168         ENTRY;
 169
 170         if (flag == LDLM_CB_CANCELING) {
 171                 /* Don't need to do anything here. */
 172                 RETURN(0);
 173         }
 174
 175         lock_res_and_lock(lock);
 176         /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
 177          * that ldlm_blocking_ast is called just before intent_policy method
 178          * takes the ns_lock, then by the time we get the lock, we might not
 179          * be the correct blocking function anymore.  So check, and return
 180          * early, if so. */
 181         if (lock->l_blocking_ast != ldlm_blocking_ast) {
 182                 unlock_res_and_lock(lock);
 183                 RETURN(0);
 184         }
 185
 186         lock->l_flags |= LDLM_FL_CBPENDING;
 187         do_ast = (!lock->l_readers && !lock->l_writers);
 188         unlock_res_and_lock(lock);
 189
 190         if (do_ast) {
 191                 struct lustre_handle lockh;
 192                 int rc;
 193
 194                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
 195                 ldlm_lock2handle(lock, &lockh);
 196                 rc = ldlm_cli_cancel(&lockh);
 197                 if (rc < 0)
 198                         CERROR("ldlm_cli_cancel: %d\n", rc);
 199         } else {
 200                 LDLM_DEBUG(lock, "Lock still has references, will be "
 201                            "cancelled later");
 202         }
 203         RETURN(0);
 204 }
 205
 206 /*
 207  * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See
 208  * comment in filter_intent_policy() on why you may need this.
 209  */
 210 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
 211 {
 212         /*
 213          * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for
 214          * that is rather subtle: with OST-side locking, it may so happen that
 215          * _all_ extent locks are held by the OST. If client wants to obtain
 216          * current file size it calls ll{,u}_glimpse_size(), and (as locks are
 217          * on the server), dummy glimpse callback fires and does
 218          * nothing. Client still receives correct file size due to the
 219          * following fragment in filter_intent_policy():
 220          *
 221          * rc = l->l_glimpse_ast(l, NULL); // this will update the LVB
 222          * if (rc != 0 && res->lr_namespace->ns_lvbo &&
 223          *     res->lr_namespace->ns_lvbo->lvbo_update) {
 224          *         res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1);
 225          * }
 226          *
 227          * that is, after glimpse_ast() fails, filter_lvbo_update() runs, and
 228          * returns correct file size to the client.
 229          */
 230         return -ELDLM_NO_LOCK_DATA;
 231 }
 232
 233 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
 234                            const struct ldlm_res_id *res_id,
 235                            ldlm_type_t type, ldlm_policy_data_t *policy,
 236                            ldlm_mode_t mode, int *flags,
 237                            ldlm_blocking_callback blocking,
 238                            ldlm_completion_callback completion,
 239                            ldlm_glimpse_callback glimpse,
 240                            void *data, __u32 lvb_len, void *lvb_swabber,
 241                            struct lustre_handle *lockh)
 242 {
 243         struct ldlm_lock *lock;
 244         int err;
 245         ENTRY;
 246
 247         LASSERT(!(*flags & LDLM_FL_REPLAY));
 248         if (unlikely(ns->ns_client)) {
 249                 CERROR("Trying to enqueue local lock in a shadow namespace\n");
 250                 LBUG();
 251         }
 252
 253         lock = ldlm_lock_create(ns, res_id, type, mode, blocking,
 254                                 completion, glimpse, data, lvb_len);
 255         if (unlikely(!lock))
 256                 GOTO(out_nolock, err = -ENOMEM);
 257         LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
 258
 259         ldlm_lock_addref_internal(lock, mode);
 260         ldlm_lock2handle(lock, lockh);
 261         lock_res_and_lock(lock);
 262         lock->l_flags |= LDLM_FL_LOCAL;
 263         if (*flags & LDLM_FL_ATOMIC_CB)
 264                 lock->l_flags |= LDLM_FL_ATOMIC_CB;
 265         lock->l_lvb_swabber = lvb_swabber;
 266         unlock_res_and_lock(lock);
 267         if (policy != NULL)
 268                 lock->l_policy_data = *policy;
 269         if (type == LDLM_EXTENT)
 270                 lock->l_req_extent = policy->l_extent;
 271
 272         err = ldlm_lock_enqueue(ns, &lock, policy, flags);
 273         if (unlikely(err != ELDLM_OK))
 274                 GOTO(out, err);
 275
 276         if (policy != NULL)
 277                 *policy = lock->l_policy_data;
 278
 279         LDLM_DEBUG_NOLOCK("client-side local enqueue handler END (lock %p)",
 280                           lock);
 281
 282         if (lock->l_completion_ast)
 283                 lock->l_completion_ast(lock, *flags, NULL);
 284
 285         LDLM_DEBUG(lock, "client-side local enqueue END");
 286         EXIT;
 287  out:
 288         LDLM_LOCK_PUT(lock);
 289  out_nolock:
 290         return err;
 291 }
 292
 293 static void failed_lock_cleanup(struct ldlm_namespace *ns,
 294                                 struct ldlm_lock *lock,
 295                                 struct lustre_handle *lockh, int mode)
 296 {
 297         /* Set a flag to prevent us from sending a CANCEL (bug 407) */
 298         lock_res_and_lock(lock);
 299         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 300         unlock_res_and_lock(lock);
 301         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 302
 303         ldlm_lock_decref_and_cancel(lockh, mode);
 304
 305         /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
 306          *       from llite/file.c/ll_file_flock(). */
 307         if (lock->l_resource->lr_type == LDLM_FLOCK) {
 308                 ldlm_lock_destroy(lock);
 309         }
 310 }
 311
 312 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
 313                           ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
 314                           int *flags, void *lvb, __u32 lvb_len,
 315                           void *lvb_swabber, struct lustre_handle *lockh,int rc)
 316 {
 317         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 318         int is_replay = *flags & LDLM_FL_REPLAY;
 319         struct ldlm_lock *lock;
 320         struct ldlm_reply *reply;
 321         int cleanup_phase = 1;
 322         ENTRY;
 323
 324         lock = ldlm_handle2lock(lockh);
 325         /* ldlm_cli_enqueue is holding a reference on this lock. */
 326         if (!lock) {
 327                 LASSERT(type == LDLM_FLOCK);
 328                 RETURN(-ENOLCK);
 329         }
 330
 331         if (rc != ELDLM_OK) {
 332                 LASSERT(!is_replay);
 333                 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
 334                            rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
 335                 if (rc == ELDLM_LOCK_ABORTED) {
 336                         /* Before we return, swab the reply */
 337                         reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF,
 338                                                    sizeof(*reply),
 339                                                    lustre_swab_ldlm_reply);
 340                         if (reply == NULL) {
 341                                 CERROR("Can't unpack ldlm_reply\n");
 342                                 rc = -EPROTO;
 343                         }
 344                         if (lvb_len) {
 345                                 void *tmplvb;
 346                                 tmplvb = lustre_swab_repbuf(req,
 347                                                             DLM_REPLY_REC_OFF,
 348                                                             lvb_len,
 349                                                             lvb_swabber);
 350                                 if (tmplvb == NULL)
 351                                         GOTO(cleanup, rc = -EPROTO);
 352                                 if (lvb != NULL)
 353                                         memcpy(lvb, tmplvb, lvb_len);
 354                         }
 355                 }
 356                 GOTO(cleanup, rc);
 357         }
 358
 359         reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF, sizeof(*reply),
 360                                    lustre_swab_ldlm_reply);
 361         if (reply == NULL) {
 362                 CERROR("Can't unpack ldlm_reply\n");
 363                 GOTO(cleanup, rc = -EPROTO);
 364         }
 365
 366         /* lock enqueued on the server */
 367         cleanup_phase = 0;
 368
 369         lock_res_and_lock(lock);
 370         lock->l_remote_handle = reply->lock_handle;
 371         *flags = reply->lock_flags;
 372         lock->l_flags |= reply->lock_flags & LDLM_INHERIT_FLAGS;
 373         /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match()
 374          * to wait with no timeout as well */
 375         lock->l_flags |= reply->lock_flags & LDLM_FL_NO_TIMEOUT;
 376         unlock_res_and_lock(lock);
 377
 378         CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%x\n",
 379                lock, reply->lock_handle.cookie, *flags);
 380
 381         /* If enqueue returned a blocked lock but the completion handler has
 382          * already run, then it fixed up the resource and we don't need to do it
 383          * again. */
 384         if ((*flags) & LDLM_FL_LOCK_CHANGED) {
 385                 int newmode = reply->lock_desc.l_req_mode;
 386                 LASSERT(!is_replay);
 387                 if (newmode && newmode != lock->l_req_mode) {
 388                         LDLM_DEBUG(lock, "server returned different mode %s",
 389                                    ldlm_lockname[newmode]);
 390                         lock->l_req_mode = newmode;
 391                 }
 392
 393                 if (memcmp(reply->lock_desc.l_resource.lr_name.name,
 394                           lock->l_resource->lr_name.name,
 395                           sizeof(struct ldlm_res_id))) {
 396                         CDEBUG(D_INFO, "remote intent success, locking "
 397                                         "(%ld,%ld,%ld) instead of "
 398                                         "(%ld,%ld,%ld)\n",
 399                               (long)reply->lock_desc.l_resource.lr_name.name[0],
 400                               (long)reply->lock_desc.l_resource.lr_name.name[1],
 401                               (long)reply->lock_desc.l_resource.lr_name.name[2],
 402                               (long)lock->l_resource->lr_name.name[0],
 403                               (long)lock->l_resource->lr_name.name[1],
 404                               (long)lock->l_resource->lr_name.name[2]);
 405
 406                         ldlm_lock_change_resource(ns, lock,
 407                                           &reply->lock_desc.l_resource.lr_name);
 408                         if (lock->l_resource == NULL) {
 409                                 LBUG();
 410                                 GOTO(cleanup, rc = -ENOMEM);
 411                         }
 412                         LDLM_DEBUG(lock, "client-side enqueue, new resource");
 413                 }
 414                 if (with_policy)
 415                         if (!(type == LDLM_IBITS && !(exp->exp_connect_flags &
 416                                                     OBD_CONNECT_IBITS)))
 417                                 lock->l_policy_data =
 418                                                  reply->lock_desc.l_policy_data;
 419                 if (type != LDLM_PLAIN)
 420                         LDLM_DEBUG(lock,"client-side enqueue, new policy data");
 421         }
 422
 423         if ((*flags) & LDLM_FL_AST_SENT ||
 424             /* Cancel extent locks as soon as possible on a liblustre client,
 425              * because it cannot handle asynchronous ASTs robustly (see
 426              * bug 7311). */
 427             (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) {
 428                 lock_res_and_lock(lock);
 429                 lock->l_flags |= LDLM_FL_CBPENDING;
 430                 unlock_res_and_lock(lock);
 431                 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
 432         }
 433
 434         /* If the lock has already been granted by a completion AST, don't
 435          * clobber the LVB with an older one. */
 436         if (lvb_len && (lock->l_req_mode != lock->l_granted_mode)) {
 437                 void *tmplvb;
 438                 tmplvb = lustre_swab_repbuf(req, DLM_REPLY_REC_OFF, lvb_len,
 439                                             lvb_swabber);
 440                 if (tmplvb == NULL)
 441                         GOTO(cleanup, rc = -EPROTO);
 442                 memcpy(lock->l_lvb_data, tmplvb, lvb_len);
 443         }
 444
 445         if (!is_replay) {
 446                 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
 447                 if (lock->l_completion_ast != NULL) {
 448                         int err = lock->l_completion_ast(lock, *flags, NULL);
 449                         if (!rc)
 450                                 rc = err;
 451                         if (rc && type != LDLM_FLOCK) /* bug 9425, bug 10250 */
 452                                 cleanup_phase = 1;
 453                 }
 454         }
 455
 456         if (lvb_len && lvb != NULL) {
 457                 /* Copy the LVB here, and not earlier, because the completion
 458                  * AST (if any) can override what we got in the reply */
 459                 memcpy(lvb, lock->l_lvb_data, lvb_len);
 460         }
 461
 462         LDLM_DEBUG(lock, "client-side enqueue END");
 463         EXIT;
 464 cleanup:
 465         if (cleanup_phase == 1 && rc)
 466                 failed_lock_cleanup(ns, lock, lockh, mode);
 467         /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
 468         LDLM_LOCK_PUT(lock);
 469         LDLM_LOCK_PUT(lock);
 470         return rc;
 471 }
 472
 473 /* PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
 474  * a single page on the send/receive side. XXX: 512 should be changed
 475  * to more adequate value. */
 476 static inline int ldlm_req_handles_avail(struct obd_export *exp,
 477                                          int *size, int bufcount, int off)
 478 {
 479         int avail = min_t(int, LDLM_MAXREQSIZE, PAGE_SIZE - 512);
 480         int old_size = size[DLM_LOCKREQ_OFF];
 481
 482         size[DLM_LOCKREQ_OFF] = sizeof(struct ldlm_request);
 483         avail -= lustre_msg_size(class_exp2cliimp(exp)->imp_msg_magic,
 484                                  bufcount, size);
 485         avail /= sizeof(struct lustre_handle);
 486         avail += LDLM_LOCKREQ_HANDLES - off;
 487         size[DLM_LOCKREQ_OFF] = old_size;
 488
 489         return avail;
 490 }
 491
 492 static inline int ldlm_cancel_handles_avail(struct obd_export *exp)
 493 {
 494         int size[2] = { sizeof(struct ptlrpc_body),
 495                         sizeof(struct ldlm_request) };
 496         return ldlm_req_handles_avail(exp, size, 2, 0);
 497 }
 498
 499 /* Cancel lru locks and pack them into the enqueue request. Pack there the given
 500  * @count locks in @cancels. */
 501 struct ptlrpc_request *ldlm_prep_enqueue_req(struct obd_export *exp,
 502                                              int bufcount, int *size,
 503                                              struct list_head *cancels,
 504                                              int count)
 505 {
 506         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 507         struct ldlm_request *dlm = NULL;
 508         struct ptlrpc_request *req;
 509         CFS_LIST_HEAD(head);
 510         ENTRY;
 511
 512         if (cancels == NULL)
 513                 cancels = &head;
 514         if (exp_connect_cancelset(exp)) {
 515                 /* Estimate the amount of available space in the request. */
 516                 int avail = ldlm_req_handles_avail(exp, size, bufcount,
 517                                                    LDLM_ENQUEUE_CANCEL_OFF);
 518                 LASSERT(avail >= count);
 519
 520                 /* Cancel lru locks here _only_ if the server supports
 521                  * EARLY_CANCEL. Otherwise we have to send extra CANCEL
 522                  * rpc right on enqueue, what will make it slower, vs.
 523                  * asynchronous rpc in blocking thread. */
 524                 count += ldlm_cancel_lru_local(ns, cancels,
 525                                                exp_connect_lru_resize(exp) ? 0 : 1,
 526                                                avail - count, LDLM_CANCEL_AGED);
 527                 size[DLM_LOCKREQ_OFF] =
 528                         ldlm_request_bufsize(count, LDLM_ENQUEUE);
 529         }
 530         req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
 531                               LDLM_ENQUEUE, bufcount, size, NULL);
 532         if (exp_connect_cancelset(exp) && req) {
 533                 dlm = lustre_msg_buf(req->rq_reqmsg,
 534                                      DLM_LOCKREQ_OFF, sizeof(*dlm));
 535                 /* Skip first lock handler in ldlm_request_pack(), this method
 536                  * will incrment @lock_count according to the lock handle amount
 537                  * actually written to the buffer. */
 538                 dlm->lock_count = LDLM_ENQUEUE_CANCEL_OFF;
 539                 ldlm_cli_cancel_list(cancels, count, req, DLM_LOCKREQ_OFF, 0);
 540         } else {
 541                 ldlm_lock_list_put(cancels, l_bl_ast, count);
 542         }
 543         RETURN(req);
 544 }
 545
 546 /* If a request has some specific initialisation it is passed in @reqp,
 547  * otherwise it is created in ldlm_cli_enqueue.
 548  *
 549  * Supports sync and async requests, pass @async flag accordingly. If a
 550  * request was created in ldlm_cli_enqueue and it is the async request,
 551  * pass it to the caller in @reqp. */
 552 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
 553                      struct ldlm_enqueue_info *einfo,
 554                      const struct ldlm_res_id *res_id,
 555                      ldlm_policy_data_t *policy, int *flags,
 556                      void *lvb, __u32 lvb_len, void *lvb_swabber,
 557                      struct lustre_handle *lockh, int async)
 558 {
 559         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 560         struct ldlm_lock *lock;
 561         struct ldlm_request *body;
 562         struct ldlm_reply *reply;
 563         int size[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
 564                         [DLM_LOCKREQ_OFF]     = sizeof(*body),
 565                         [DLM_REPLY_REC_OFF]   = lvb_len };
 566         int is_replay = *flags & LDLM_FL_REPLAY;
 567         int req_passed_in = 1, rc, err;
 568         struct ptlrpc_request *req;
 569         ENTRY;
 570
 571         LASSERT(exp != NULL);
 572
 573         /* If we're replaying this lock, just check some invariants.
 574          * If we're creating a new lock, get everything all setup nice. */
 575         if (is_replay) {
 576                 lock = ldlm_handle2lock(lockh);
 577                 LASSERT(lock != NULL);
 578                 LDLM_DEBUG(lock, "client-side enqueue START");
 579                 LASSERT(exp == lock->l_conn_export);
 580         } else {
 581                 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
 582                                         einfo->ei_mode, einfo->ei_cb_bl,
 583                                         einfo->ei_cb_cp, einfo->ei_cb_gl,
 584                                         einfo->ei_cbdata, lvb_len);
 585                 if (lock == NULL)
 586                         RETURN(-ENOMEM);
 587                 /* for the local lock, add the reference */
 588                 ldlm_lock_addref_internal(lock, einfo->ei_mode);
 589                 ldlm_lock2handle(lock, lockh);
 590                 lock->l_lvb_swabber = lvb_swabber;
 591                 if (policy != NULL) {
 592                         /* INODEBITS_INTEROP: If the server does not support
 593                          * inodebits, we will request a plain lock in the
 594                          * descriptor (ldlm_lock2desc() below) but use an
 595                          * inodebits lock internally with both bits set.
 596                          */
 597                         if (einfo->ei_type == LDLM_IBITS &&
 598                             !(exp->exp_connect_flags & OBD_CONNECT_IBITS))
 599                                 lock->l_policy_data.l_inodebits.bits =
 600                                         MDS_INODELOCK_LOOKUP |
 601                                         MDS_INODELOCK_UPDATE;
 602                         else
 603                                 lock->l_policy_data = *policy;
 604                 }
 605
 606                 if (einfo->ei_type == LDLM_EXTENT)
 607                         lock->l_req_extent = policy->l_extent;
 608                 LDLM_DEBUG(lock, "client-side enqueue START");
 609         }
 610
 611         /* lock not sent to server yet */
 612
 613         if (reqp == NULL || *reqp == NULL) {
 614                 req = ldlm_prep_enqueue_req(exp, 2, size, NULL, 0);
 615                 if (req == NULL) {
 616                         failed_lock_cleanup(ns, lock, lockh, einfo->ei_mode);
 617                         LDLM_LOCK_PUT(lock);
 618                         RETURN(-ENOMEM);
 619                 }
 620                 req_passed_in = 0;
 621                 if (reqp)
 622                         *reqp = req;
 623         } else {
 624                 req = *reqp;
 625                 LASSERTF(lustre_msg_buflen(req->rq_reqmsg, DLM_LOCKREQ_OFF) >=
 626                          sizeof(*body), "buflen[%d] = %d, not "LPSZ"\n",
 627                          DLM_LOCKREQ_OFF,
 628                          lustre_msg_buflen(req->rq_reqmsg, DLM_LOCKREQ_OFF),
 629                          sizeof(*body));
 630         }
 631
 632         lock->l_conn_export = exp;
 633         lock->l_export = NULL;
 634         lock->l_blocking_ast = einfo->ei_cb_bl;
 635
 636         /* Dump lock data into the request buffer */
 637         body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, sizeof(*body));
 638         ldlm_lock2desc(lock, &body->lock_desc);
 639         body->lock_flags = *flags;
 640         body->lock_handle[0] = *lockh;
 641
 642         /* Continue as normal. */
 643         if (!req_passed_in) {
 644                 size[DLM_LOCKREPLY_OFF] = sizeof(*reply);
 645                 ptlrpc_req_set_repsize(req, 2 + (lvb_len > 0), size);
 646         }
 647
 648         /*
 649          * Liblustre client doesn't get extent locks, except for O_APPEND case
 650          * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
 651          * [i_size, OBD_OBJECT_EOF] lock is taken.
 652          */
 653         LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT ||
 654                      policy->l_extent.end == OBD_OBJECT_EOF));
 655
 656         if (async) {
 657                 LASSERT(reqp != NULL);
 658                 RETURN(0);
 659         }
 660
 661         LDLM_DEBUG(lock, "sending request");
 662         rc = ptlrpc_queue_wait(req);
 663         err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
 664                                     einfo->ei_mode, flags, lvb, lvb_len,
 665                                     lvb_swabber, lockh, rc);
 666
 667         /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
 668          * one reference that we took */
 669         if (err == -ENOLCK)
 670                 LDLM_LOCK_PUT(lock);
 671         else
 672                 rc = err;
 673
 674         if (!req_passed_in && req != NULL) {
 675                 ptlrpc_req_finished(req);
 676                 if (reqp)
 677                         *reqp = NULL;
 678         }
 679
 680         RETURN(rc);
 681 }
 682
 683 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
 684                                   int *flags)
 685 {
 686         struct ldlm_resource *res;
 687         int rc;
 688         ENTRY;
 689         if (lock->l_resource->lr_namespace->ns_client) {
 690                 CERROR("Trying to cancel local lock\n");
 691                 LBUG();
 692         }
 693         LDLM_DEBUG(lock, "client-side local convert");
 694
 695         res = ldlm_lock_convert(lock, new_mode, flags);
 696         if (res) {
 697                 ldlm_reprocess_all(res);
 698                 rc = 0;
 699         } else {
 700                 rc = EDEADLOCK;
 701         }
 702         LDLM_DEBUG(lock, "client-side local convert handler END");
 703         LDLM_LOCK_PUT(lock);
 704         RETURN(rc);
 705 }
 706
 707 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
 708  * conversion of locks which are on the waiting or converting queue */
 709 /* Caller of this code is supposed to take care of lock readers/writers
 710    accounting */
 711 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, int *flags)
 712 {
 713         struct ldlm_request *body;
 714         struct ldlm_reply *reply;
 715         struct ldlm_lock *lock;
 716         struct ldlm_resource *res;
 717         struct ptlrpc_request *req;
 718         int size[2] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
 719                         [DLM_LOCKREQ_OFF]     = sizeof(*body) };
 720         int rc;
 721         ENTRY;
 722
 723         lock = ldlm_handle2lock(lockh);
 724         if (!lock) {
 725                 LBUG();
 726                 RETURN(-EINVAL);
 727         }
 728         *flags = 0;
 729
 730         if (lock->l_conn_export == NULL)
 731                 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
 732
 733         LDLM_DEBUG(lock, "client-side convert");
 734
 735         req = ptlrpc_prep_req(class_exp2cliimp(lock->l_conn_export),
 736                               LUSTRE_DLM_VERSION, LDLM_CONVERT, 2, size, NULL);
 737         if (!req)
 738                 GOTO(out, rc = -ENOMEM);
 739
 740         body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, sizeof(*body));
 741         body->lock_handle[0] = lock->l_remote_handle;
 742
 743         body->lock_desc.l_req_mode = new_mode;
 744         body->lock_flags = *flags;
 745
 746         size[DLM_LOCKREPLY_OFF] = sizeof(*reply);
 747         ptlrpc_req_set_repsize(req, 2, size);
 748
 749         rc = ptlrpc_queue_wait(req);
 750         if (rc != ELDLM_OK)
 751                 GOTO(out, rc);
 752
 753         reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF, sizeof(*reply),
 754                                    lustre_swab_ldlm_reply);
 755         if (reply == NULL) {
 756                 CERROR ("Can't unpack ldlm_reply\n");
 757                 GOTO (out, rc = -EPROTO);
 758         }
 759
 760         if (req->rq_status)
 761                 GOTO(out, rc = req->rq_status);
 762
 763         res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
 764         if (res != NULL) {
 765                 ldlm_reprocess_all(res);
 766                 /* Go to sleep until the lock is granted. */
 767                 /* FIXME: or cancelled. */
 768                 if (lock->l_completion_ast) {
 769                         rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
 770                                                     NULL);
 771                         if (rc)
 772                                 GOTO(out, rc);
 773                 }
 774         } else {
 775                 rc = EDEADLOCK;
 776         }
 777         EXIT;
 778  out:
 779         LDLM_LOCK_PUT(lock);
 780         ptlrpc_req_finished(req);
 781         return rc;
 782 }
 783
 784 /* Cancel locks locally.
 785  * Returns:
 786  * LDLM_FL_LOCAL_ONLY if tere is no need in a CANCEL rpc to the server;
 787  * LDLM_FL_CANCELING otherwise;
 788  * LDLM_FL_BL_AST if there is a need in a separate CANCEL rpc. */
 789 static int ldlm_cli_cancel_local(struct ldlm_lock *lock)
 790 {
 791         int rc = LDLM_FL_LOCAL_ONLY;
 792         ENTRY;
 793
 794         if (lock->l_conn_export) {
 795                 int local_only;
 796
 797                 LDLM_DEBUG(lock, "client-side cancel");
 798                 /* Set this flag to prevent others from getting new references*/
 799                 lock_res_and_lock(lock);
 800                 lock->l_flags |= LDLM_FL_CBPENDING;
 801                 local_only = (lock->l_flags &
 802                               (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
 803                 ldlm_cancel_callback(lock);
 804                 rc = (lock->l_flags & LDLM_FL_BL_AST) ?
 805                         LDLM_FL_BL_AST : LDLM_FL_CANCELING;
 806                 unlock_res_and_lock(lock);
 807
 808                 if (local_only) {
 809                         CDEBUG(D_DLMTRACE, "not sending request (at caller's "
 810                                "instruction)\n");
 811                         rc = LDLM_FL_LOCAL_ONLY;
 812                 }
 813                 ldlm_lock_cancel(lock);
 814         } else {
 815                 if (lock->l_resource->lr_namespace->ns_client) {
 816                         LDLM_ERROR(lock, "Trying to cancel local lock");
 817                         LBUG();
 818                 }
 819                 LDLM_DEBUG(lock, "server-side local cancel");
 820                 ldlm_lock_cancel(lock);
 821                 ldlm_reprocess_all(lock->l_resource);
 822                 LDLM_DEBUG(lock, "server-side local cancel handler END");
 823         }
 824
 825         RETURN(rc);
 826 }
 827
 828 /* Pack @count locks in @head into ldlm_request buffer at the offset @off,
 829    of the request @req. */
 830 static void ldlm_cancel_pack(struct ptlrpc_request *req, int off,
 831                              struct list_head *head, int count)
 832 {
 833         struct ldlm_request *dlm;
 834         struct ldlm_lock *lock;
 835         int max, packed = 0;
 836         ENTRY;
 837
 838         dlm = lustre_msg_buf(req->rq_reqmsg, off, sizeof(*dlm));
 839         LASSERT(dlm != NULL);
 840
 841         /* Check the room in the request buffer. */
 842         max = lustre_msg_buflen(req->rq_reqmsg, off) -
 843                 sizeof(struct ldlm_request);
 844         max /= sizeof(struct lustre_handle);
 845         max += LDLM_LOCKREQ_HANDLES;
 846         LASSERT(max >= dlm->lock_count + count);
 847
 848         /* XXX: it would be better to pack lock handles grouped by resource.
 849          * so that the server cancel would call filter_lvbo_update() less
 850          * frequently. */
 851         list_for_each_entry(lock, head, l_bl_ast) {
 852                 if (!count--)
 853                         break;
 854                 LASSERT(lock->l_conn_export);
 855                 /* Pack the lock handle to the given request buffer. */
 856                 LDLM_DEBUG(lock, "packing");
 857                 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
 858                 packed++;
 859         }
 860         CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
 861         EXIT;
 862 }
 863
 864 /* Prepare and send a batched cancel rpc, it will include count lock handles
 865  * of locks given in @head. */
 866 int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
 867                         int count, int flags)
 868 {
 869         struct ptlrpc_request *req = NULL;
 870         struct ldlm_request *body;
 871         int size[2] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
 872                 [DLM_LOCKREQ_OFF]     = sizeof(*body) };
 873         struct obd_import *imp;
 874         int free, sent = 0;
 875         int rc = 0;
 876         ENTRY;
 877
 878         LASSERT(exp != NULL);
 879         LASSERT(count > 0);
 880
 881         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
 882                 RETURN(count);
 883
 884         free = ldlm_req_handles_avail(exp, size, 2, 0);
 885         if (count > free)
 886                 count = free;
 887
 888         size[DLM_LOCKREQ_OFF] = ldlm_request_bufsize(count, LDLM_CANCEL);
 889         while (1) {
 890                 imp = class_exp2cliimp(exp);
 891                 if (imp == NULL || imp->imp_invalid) {
 892                         CDEBUG(D_HA, "skipping cancel on invalid import %p\n",
 893                                imp);
 894                         RETURN(count);
 895                 }
 896
 897                 req = ptlrpc_prep_req(imp, LUSTRE_DLM_VERSION, LDLM_CANCEL, 2,
 898                                       size, NULL);
 899                 if (!req)
 900                         GOTO(out, rc = -ENOMEM);
 901
 902                 req->rq_no_resend = 1;
 903                 req->rq_no_delay = 1;
 904
 905                 /* XXX FIXME bug 249 */
 906                 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
 907                 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
 908
 909                 body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF,
 910                                       sizeof(*body));
 911                 ldlm_cancel_pack(req, DLM_LOCKREQ_OFF, cancels, count);
 912
 913                 ptlrpc_req_set_repsize(req, 1, NULL);
 914                 if (flags & LDLM_FL_ASYNC) {
 915                         ptlrpcd_add_req(req);
 916                         sent = count;
 917                         GOTO(out, 0);
 918                 } else {
 919                         rc = ptlrpc_queue_wait(req);
 920                 }
 921                 if (rc == ESTALE) {
 922                         CDEBUG(D_DLMTRACE, "client/server (nid %s) "
 923                                "out of sync -- not fatal\n",
 924                                libcfs_nid2str(req->rq_import->
 925                                               imp_connection->c_peer.nid));
 926                         rc = 0;
 927                 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
 928                            req->rq_import_generation == imp->imp_generation) {
 929                         ptlrpc_req_finished(req);
 930                         continue;
 931                 } else if (rc != ELDLM_OK) {
 932                         CERROR("Got rc %d from cancel RPC: canceling "
 933                                "anyway\n", rc);
 934                         break;
 935                 }
 936                 sent = count;
 937                 break;
 938         }
 939
 940         ptlrpc_req_finished(req);
 941         EXIT;
 942 out:
 943         return sent ? sent : rc;
 944 }
 945
 946 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
 947 {
 948         LASSERT(imp != NULL);
 949         return &imp->imp_obd->obd_namespace->ns_pool;
 950 }
 951
 952 int ldlm_cli_update_pool(struct ptlrpc_request *req)
 953 {
 954         struct ldlm_pool *pl;
 955         ENTRY;
 956
 957         if (!imp_connect_lru_resize(req->rq_import))
 958                 RETURN(0);
 959
 960         if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
 961             lustre_msg_get_limit(req->rq_repmsg) == 0)
 962                 RETURN(0);
 963
 964         pl = ldlm_imp2pl(req->rq_import);
 965
 966         spin_lock(&pl->pl_lock);
 967 #ifdef __KERNEL__
 968         {
 969                 __u64 old_slv, fast_slv_change;
 970
 971                 old_slv = ldlm_pool_get_slv(pl);
 972                 fast_slv_change = old_slv * LDLM_POOLS_FAST_SLV_CHANGE;
 973                 do_div(fast_slv_change, 100);
 974 #endif
 975                 pl->pl_update_time = cfs_time_current();
 976                 ldlm_pool_set_slv(pl, lustre_msg_get_slv(req->rq_repmsg));
 977                 ldlm_pool_set_limit(pl, lustre_msg_get_limit(req->rq_repmsg));
 978 #ifdef __KERNEL__
 979                 /* Wake up pools thread only if SLV has changed more than
 980                  * 5% since last update. In this case we want to react asap.
 981                  * Otherwise it is no sense to wake up pools as they are
 982                  * re-calculated every 1s anyways. */
 983                 if (old_slv > ldlm_pool_get_slv(pl) &&
 984                     old_slv - ldlm_pool_get_slv(pl) > fast_slv_change)
 985                         ldlm_pools_wakeup();
 986         }
 987 #endif
 988         spin_unlock(&pl->pl_lock);
 989
 990         RETURN(0);
 991 }
 992 EXPORT_SYMBOL(ldlm_cli_update_pool);
 993
 994 int ldlm_cli_cancel(struct lustre_handle *lockh)
 995 {
 996         struct ldlm_lock *lock;
 997         CFS_LIST_HEAD(cancels);
 998         int rc = 0;
 999         ENTRY;
1000
1001         /* concurrent cancels on the same handle can happen */
1002         lock = __ldlm_handle2lock(lockh, LDLM_FL_CANCELING);
1003         if (lock == NULL) {
1004                 LDLM_DEBUG_NOLOCK("lock is already being destroyed\n");
1005                 RETURN(0);
1006         }
1007
1008         rc = ldlm_cli_cancel_local(lock);
1009         list_add(&lock->l_bl_ast, &cancels);
1010
1011         if (rc == LDLM_FL_BL_AST) {
1012                 rc = ldlm_cli_cancel_req(lock->l_conn_export, &cancels, 1, 0);
1013         } else if (rc == LDLM_FL_CANCELING) {
1014                 int avail = ldlm_cancel_handles_avail(lock->l_conn_export);
1015                 int count = 1;
1016                 LASSERT(avail > 0);
1017                 count += ldlm_cancel_lru_local(lock->l_resource->lr_namespace,
1018                                                &cancels, 0, avail - 1,
1019                                                LDLM_CANCEL_AGED);
1020                 ldlm_cli_cancel_list(&cancels, count, NULL, 0, 0);
1021         }
1022         if (rc != LDLM_FL_CANCELING)
1023                 LDLM_LOCK_PUT(lock);
1024         RETURN(rc < 0 ? rc : 0);
1025 }
1026
1027 /* - Free space in lru for @count new locks,
1028  *   redundant unused locks are canceled locally;
1029  * - also cancel locally unused aged locks;
1030  * - do not cancel more than @max locks;
1031  * - GET the found locks and add them into the @cancels list.
1032  *
1033  * A client lock can be added to the l_bl_ast list only when it is
1034  * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing CANCEL.
1035  * There are the following use cases: ldlm_cancel_resource_local(),
1036  * ldlm_cancel_lru_local() and ldlm_cli_cancel(), which check&set this
1037  * flag properly. As any attempt to cancel a lock rely on this flag,
1038  * l_bl_ast list is accessed later without any special locking. */
1039 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
1040                           int count, int max, int flags)
1041 {
1042         cfs_time_t cur = cfs_time_current();
1043         int rc, added = 0, left, unused;
1044         struct ldlm_lock *lock, *next;
1045         __u64 slv, lvf, lv;
1046         ENTRY;
1047
1048         spin_lock(&ns->ns_unused_lock);
1049         unused = ns->ns_nr_unused;
1050
1051         if (!ns_connect_lru_resize(ns))
1052                 count += unused - ns->ns_max_unused;
1053
1054         while (!list_empty(&ns->ns_unused_list)) {
1055                 struct ldlm_pool *pl = &ns->ns_pool;
1056
1057                 if (max && added >= max)
1058                         break;
1059
1060                 list_for_each_entry(lock, &ns->ns_unused_list, l_lru) {
1061                         /* somebody is already doing CANCEL or there is a
1062                          * blocking request will send cancel. */
1063                         if (!(lock->l_flags & LDLM_FL_CANCELING) &&
1064                             !(lock->l_flags & LDLM_FL_BL_AST))
1065                                 break;
1066                 }
1067                 if (&lock->l_lru == &ns->ns_unused_list)
1068                         break;
1069
1070                 if (ns_connect_lru_resize(ns)) {
1071                         cfs_time_t la;
1072
1073                         /* Take into account SLV only if cpount == 0. */
1074                         if (count == 0) {
1075                                 /* Calculate lv for every lock. */
1076                                 spin_lock(&pl->pl_lock);
1077                                 slv = ldlm_pool_get_slv(pl);
1078                                 lvf = atomic_read(&pl->pl_lock_volume_factor);
1079                                 spin_unlock(&pl->pl_lock);
1080
1081                                 la = cfs_duration_sec(cfs_time_sub(cur,
1082                                                       lock->l_last_used));
1083                                 if (la == 0)
1084                                         la = 1;
1085
1086                                 /* Stop when slv is not yet come from server
1087                                  * or lv is smaller than it is. */
1088                                 lv = lvf * la * unused;
1089                                 if (slv == 1 || lv < slv)
1090                                         break;
1091                         } else {
1092                                 if (added >= count)
1093                                         break;
1094                         }
1095                 } else {
1096                         if ((added >= count) &&
1097                             (!(flags & LDLM_CANCEL_AGED) ||
1098                              cfs_time_before_64(cur, ns->ns_max_age +
1099                                                 lock->l_last_used)))
1100                                 break;
1101                 }
1102
1103                 LDLM_LOCK_GET(lock); /* dropped by bl thread */
1104                 spin_unlock(&ns->ns_unused_lock);
1105
1106                 lock_res_and_lock(lock);
1107                 /* Check flags again under the lock. */
1108                 if ((lock->l_flags & LDLM_FL_CANCELING) ||
1109                     (lock->l_flags & LDLM_FL_BL_AST) ||
1110                     (ldlm_lock_remove_from_lru(lock) == 0)) {
1111                         /* other thread is removing lock from lru or
1112                          * somebody is already doing CANCEL or
1113                          * there is a blocking request which will send
1114                          * cancel by itseft. */
1115                         unlock_res_and_lock(lock);
1116                         LDLM_LOCK_PUT(lock);
1117                         spin_lock(&ns->ns_unused_lock);
1118                         continue;
1119                 }
1120                 LASSERT(!lock->l_readers && !lock->l_writers);
1121
1122                 /* If we have chosen to canecl this lock voluntarily, we better
1123                    send cancel notification to server, so that it frees
1124                    appropriate state. This might lead to a race where while
1125                    we are doing cancel here, server is also silently
1126                    cancelling this lock. */
1127                 lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK;
1128
1129                 /* Setting the CBPENDING flag is a little misleading, but
1130                  * prevents an important race; namely, once CBPENDING is set,
1131                  * the lock can accumulate no more readers/writers.  Since
1132                  * readers and writers are already zero here, ldlm_lock_decref
1133                  * won't see this flag and call l_blocking_ast */
1134                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1135                 /* We can't re-add to l_lru as it confuses the refcounting in
1136                  * ldlm_lock_remove_from_lru() if an AST arrives after we drop
1137                  * ns_lock below. We use l_bl_ast and can't use l_pending_chain
1138                  * as it is used both on server and client nevertheles bug 5666
1139                  * says it is used only on server. --umka */
1140
1141                 LASSERT(list_empty(&lock->l_bl_ast));
1142                 list_add(&lock->l_bl_ast, cancels);
1143                 unlock_res_and_lock(lock);
1144                 spin_lock(&ns->ns_unused_lock);
1145                 added++;
1146                 unused--;
1147         }
1148         spin_unlock(&ns->ns_unused_lock);
1149
1150         /* Handle only @added inserted locks. */
1151         left = added;
1152         list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1153                 if (left-- == 0)
1154                         break;
1155
1156                 rc = ldlm_cli_cancel_local(lock);
1157                 if (rc == LDLM_FL_BL_AST) {
1158                         CFS_LIST_HEAD(head);
1159
1160                         LDLM_DEBUG(lock, "Cancel lock separately");
1161                         list_del_init(&lock->l_bl_ast);
1162                         list_add(&lock->l_bl_ast, &head);
1163                         ldlm_cli_cancel_req(lock->l_conn_export, &head, 1, 0);
1164                         rc = LDLM_FL_LOCAL_ONLY;
1165                 }
1166                 if (rc == LDLM_FL_LOCAL_ONLY) {
1167                         /* CANCEL RPC should not be sent to server. */
1168                         list_del_init(&lock->l_bl_ast);
1169                         LDLM_LOCK_PUT(lock);
1170                         added--;
1171                 }
1172
1173         }
1174         RETURN(added);
1175 }
1176
1177 /* when called with LDLM_ASYNC the blocking callback will be handled
1178  * in a thread and this function will return after the thread has been
1179  * asked to call the callback.  when called with LDLM_SYNC the blocking
1180  * callback will be performed in this function. */
1181 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr, ldlm_sync_t sync)
1182 {
1183         CFS_LIST_HEAD(cancels);
1184         int count, rc;
1185         ENTRY;
1186
1187 #ifndef __KERNEL__
1188         sync = LDLM_SYNC; /* force to be sync in user space */
1189 #endif
1190         count = ldlm_cancel_lru_local(ns, &cancels, nr, 0, 0);
1191         if (sync == LDLM_ASYNC) {
1192                 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count);
1193                 if (rc == 0)
1194                         RETURN(count);
1195         }
1196
1197         /* If an error occured in ASYNC mode, or
1198          * this is SYNC mode, cancel the list. */
1199         ldlm_cli_cancel_list(&cancels, count, NULL, 0, 0);
1200         RETURN(count);
1201 }
1202
1203 /* Find and cancel locally unused locks found on resource, matched to the
1204  * given policy, mode. GET the found locks and add them into the @cancels
1205  * list. */
1206 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1207                                struct list_head *cancels,
1208                                ldlm_policy_data_t *policy,
1209                                ldlm_mode_t mode, int lock_flags,
1210                                int flags, void *opaque)
1211 {
1212         struct ldlm_lock *lock, *next;
1213         int count = 0, left;
1214         ENTRY;
1215
1216         lock_res(res);
1217         list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1218                 if (opaque != NULL && lock->l_ast_data != opaque) {
1219                         LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1220                                    lock->l_ast_data, opaque);
1221                         //LBUG();
1222                         continue;
1223                 }
1224
1225                 if (lock->l_readers || lock->l_writers) {
1226                         if (flags & LDLM_FL_WARN) {
1227                                 LDLM_ERROR(lock, "lock in use");
1228                                 //LBUG();
1229                         }
1230                         continue;
1231                 }
1232
1233                 /* If somebody is already doing CANCEL, or blocking ast came,
1234                  * skip this lock. */
1235                 if (lock->l_flags & LDLM_FL_BL_AST ||
1236                     lock->l_flags & LDLM_FL_CANCELING)
1237                         continue;
1238
1239                 if (lockmode_compat(lock->l_granted_mode, mode))
1240                         continue;
1241
1242                 /* If policy is given and this is IBITS lock, add to list only
1243                  * those locks that match by policy. */
1244                 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1245                     !(lock->l_policy_data.l_inodebits.bits &
1246                       policy->l_inodebits.bits))
1247                         continue;
1248
1249                 /* See CBPENDING comment in ldlm_cancel_lru */
1250                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1251                         lock_flags;
1252
1253                 LASSERT(list_empty(&lock->l_bl_ast));
1254                 list_add(&lock->l_bl_ast, cancels);
1255                 LDLM_LOCK_GET(lock);
1256                 count++;
1257         }
1258         unlock_res(res);
1259
1260         /* Handle only @count inserted locks. */
1261         left = count;
1262         list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1263                 int rc = LDLM_FL_LOCAL_ONLY;
1264
1265                 if (left-- == 0)
1266                         break;
1267                 if (flags & LDLM_FL_LOCAL_ONLY)
1268                         ldlm_lock_cancel(lock);
1269                 else
1270                         rc = ldlm_cli_cancel_local(lock);
1271
1272                 if (rc == LDLM_FL_BL_AST) {
1273                         CFS_LIST_HEAD(head);
1274
1275                         LDLM_DEBUG(lock, "Cancel lock separately");
1276                         list_del_init(&lock->l_bl_ast);
1277                         list_add(&lock->l_bl_ast, &head);
1278                         ldlm_cli_cancel_req(lock->l_conn_export, &head, 1, 0);
1279                         rc = LDLM_FL_LOCAL_ONLY;
1280                 }
1281                 if (rc == LDLM_FL_LOCAL_ONLY) {
1282                         /* CANCEL RPC should not be sent to server. */
1283                         list_del_init(&lock->l_bl_ast);
1284                         LDLM_LOCK_PUT(lock);
1285                         count--;
1286                 }
1287         }
1288         RETURN(count);
1289 }
1290
1291 /* If @req is NULL, send CANCEL request to server with handles of locks
1292  * in the @cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1293  * separately per lock.
1294  * If @req is not NULL, put handles of locks in @cancels into the request
1295  * buffer at the offset @off.
1296  * Destroy @cancels at the end. */
1297 int ldlm_cli_cancel_list(struct list_head *cancels, int count,
1298                          struct ptlrpc_request *req, int off, int flags)
1299 {
1300         struct ldlm_lock *lock;
1301         int res = 0;
1302         ENTRY;
1303
1304         if (list_empty(cancels) || count == 0)
1305                 RETURN(0);
1306
1307         /* XXX: requests (both batched and not) could be sent in parallel.
1308          * Usually it is enough to have just 1 RPC, but it is possible that
1309          * there are to many locks to be cancelled in LRU or on a resource.
1310          * It would also speed up the case when the server does not support
1311          * the feature. */
1312         while (count > 0) {
1313                 LASSERT(!list_empty(cancels));
1314                 lock = list_entry(cancels->next, struct ldlm_lock, l_bl_ast);
1315                 LASSERT(lock->l_conn_export);
1316
1317                 if (exp_connect_cancelset(lock->l_conn_export)) {
1318                         res = count;
1319                         if (req)
1320                                 ldlm_cancel_pack(req, off, cancels, count);
1321                         else
1322                                 res = ldlm_cli_cancel_req(lock->l_conn_export,
1323                                                           cancels, count, flags);
1324                 } else {
1325                         res = ldlm_cli_cancel_req(lock->l_conn_export,
1326                                                   cancels, 1, flags);
1327                 }
1328
1329                 if (res < 0) {
1330                         CERROR("ldlm_cli_cancel_list: %d\n", res);
1331                         res = count;
1332                 }
1333
1334                 count -= res;
1335                 ldlm_lock_list_put(cancels, l_bl_ast, res);
1336         }
1337         LASSERT(list_empty(cancels));
1338         LASSERT(count == 0);
1339         RETURN(0);
1340 }
1341
1342 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1343                                     const struct ldlm_res_id *res_id,
1344                                     ldlm_policy_data_t *policy,
1345                                     ldlm_mode_t mode, int flags, void *opaque)
1346 {
1347         struct ldlm_resource *res;
1348         CFS_LIST_HEAD(cancels);
1349         int count;
1350         int rc;
1351         ENTRY;
1352
1353         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1354         if (res == NULL) {
1355                 /* This is not a problem. */
1356                 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id->name[0]);
1357                 RETURN(0);
1358         }
1359
1360         count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
1361                                            0, flags, opaque);
1362         rc = ldlm_cli_cancel_list(&cancels, count, NULL, 0, flags);
1363         if (rc != ELDLM_OK)
1364                 CERROR("ldlm_cli_cancel_unused_resource: %d\n", rc);
1365
1366         ldlm_resource_putref(res);
1367         RETURN(0);
1368 }
1369
1370 static inline int have_no_nsresource(struct ldlm_namespace *ns)
1371 {
1372         int no_resource = 0;
1373
1374         spin_lock(&ns->ns_hash_lock);
1375         if (ns->ns_resources == 0)
1376                 no_resource = 1;
1377         spin_unlock(&ns->ns_hash_lock);
1378
1379         RETURN(no_resource);
1380 }
1381
1382 /* Cancel all locks on a namespace (or a specific resource, if given)
1383  * that have 0 readers/writers.
1384  *
1385  * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
1386  * to notify the server. */
1387 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1388                            const struct ldlm_res_id *res_id,
1389                            int flags, void *opaque)
1390 {
1391         int i;
1392         ENTRY;
1393
1394         if (ns == NULL)
1395                 RETURN(ELDLM_OK);
1396
1397         if (res_id)
1398                 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
1399                                                        LCK_MINMODE, flags,
1400                                                        opaque));
1401
1402         spin_lock(&ns->ns_hash_lock);
1403         for (i = 0; i < RES_HASH_SIZE; i++) {
1404                 struct list_head *tmp;
1405                 tmp = ns->ns_hash[i].next;
1406                 while (tmp != &(ns->ns_hash[i])) {
1407                         struct ldlm_resource *res;
1408                         int rc;
1409
1410                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
1411                         ldlm_resource_getref(res);
1412                         spin_unlock(&ns->ns_hash_lock);
1413
1414                         rc = ldlm_cli_cancel_unused_resource(ns, &res->lr_name,
1415                                                              NULL, LCK_MINMODE,
1416                                                              flags, opaque);
1417
1418                         if (rc)
1419                                 CERROR("ldlm_cli_cancel_unused ("LPU64"): %d\n",
1420                                        res->lr_name.name[0], rc);
1421
1422                         spin_lock(&ns->ns_hash_lock);
1423                         tmp = tmp->next;
1424                         ldlm_resource_putref_locked(res);
1425                 }
1426         }
1427         spin_unlock(&ns->ns_hash_lock);
1428
1429         RETURN(ELDLM_OK);
1430 }
1431
1432 /* join/split resource locks to/from lru list */
1433 int ldlm_cli_join_lru(struct ldlm_namespace *ns,
1434                       const struct ldlm_res_id *res_id, int join)
1435 {
1436         struct ldlm_resource *res;
1437         struct ldlm_lock *lock, *n;
1438         int count = 0;
1439         ENTRY;
1440
1441         LASSERT(ns->ns_client == LDLM_NAMESPACE_CLIENT);
1442
1443         res = ldlm_resource_get(ns, NULL, res_id, LDLM_EXTENT, 0);
1444         if (res == NULL)
1445                 RETURN(count);
1446         LASSERT(res->lr_type == LDLM_EXTENT);
1447
1448         lock_res(res);
1449         if (!join)
1450                 goto split;
1451
1452         list_for_each_entry_safe (lock, n, &res->lr_granted, l_res_link) {
1453                 if (list_empty(&lock->l_lru) &&
1454                     !lock->l_readers && !lock->l_writers &&
1455                     !(lock->l_flags & LDLM_FL_LOCAL) &&
1456                     !(lock->l_flags & LDLM_FL_CBPENDING)) {
1457                         ldlm_lock_add_to_lru(lock);
1458                         lock->l_flags &= ~LDLM_FL_NO_LRU;
1459                         LDLM_DEBUG(lock, "join lock to lru");
1460                         count++;
1461                 }
1462         }
1463         goto unlock;
1464 split:
1465         spin_lock(&ns->ns_unused_lock);
1466         list_for_each_entry_safe (lock, n, &ns->ns_unused_list, l_lru) {
1467                 if (lock->l_resource == res) {
1468                         ldlm_lock_remove_from_lru_nolock(lock);
1469                         lock->l_flags |= LDLM_FL_NO_LRU;
1470                         LDLM_DEBUG(lock, "split lock from lru");
1471                         count++;
1472                 }
1473         }
1474         spin_unlock(&ns->ns_unused_lock);
1475 unlock:
1476         unlock_res(res);
1477         ldlm_resource_putref(res);
1478         RETURN(count);
1479 }
1480
1481 /* Lock iterators. */
1482
1483 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
1484                           void *closure)
1485 {
1486         struct list_head *tmp, *next;
1487         struct ldlm_lock *lock;
1488         int rc = LDLM_ITER_CONTINUE;
1489
1490         ENTRY;
1491
1492         if (!res)
1493                 RETURN(LDLM_ITER_CONTINUE);
1494
1495         lock_res(res);
1496         list_for_each_safe(tmp, next, &res->lr_granted) {
1497                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1498
1499                 if (iter(lock, closure) == LDLM_ITER_STOP)
1500                         GOTO(out, rc = LDLM_ITER_STOP);
1501         }
1502
1503         list_for_each_safe(tmp, next, &res->lr_converting) {
1504                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1505
1506                 if (iter(lock, closure) == LDLM_ITER_STOP)
1507                         GOTO(out, rc = LDLM_ITER_STOP);
1508         }
1509
1510         list_for_each_safe(tmp, next, &res->lr_waiting) {
1511                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1512
1513                 if (iter(lock, closure) == LDLM_ITER_STOP)
1514                         GOTO(out, rc = LDLM_ITER_STOP);
1515         }
1516  out:
1517         unlock_res(res);
1518         RETURN(rc);
1519 }
1520
1521 struct iter_helper_data {
1522         ldlm_iterator_t iter;
1523         void *closure;
1524 };
1525
1526 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
1527 {
1528         struct iter_helper_data *helper = closure;
1529         return helper->iter(lock, helper->closure);
1530 }
1531
1532 static int ldlm_res_iter_helper(struct ldlm_resource *res, void *closure)
1533 {
1534         return ldlm_resource_foreach(res, ldlm_iter_helper, closure);
1535 }
1536
1537 int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
1538                            void *closure)
1539 {
1540         struct iter_helper_data helper = { iter: iter, closure: closure };
1541         return ldlm_namespace_foreach_res(ns, ldlm_res_iter_helper, &helper);
1542 }
1543
1544 int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
1545                                ldlm_res_iterator_t iter, void *closure)
1546 {
1547         int i, rc = LDLM_ITER_CONTINUE;
1548         struct ldlm_resource *res;
1549         struct list_head *tmp;
1550
1551         ENTRY;
1552         spin_lock(&ns->ns_hash_lock);
1553         for (i = 0; i < RES_HASH_SIZE; i++) {
1554                 tmp = ns->ns_hash[i].next;
1555                 while (tmp != &(ns->ns_hash[i])) {
1556                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
1557                         ldlm_resource_getref(res);
1558                         spin_unlock(&ns->ns_hash_lock);
1559
1560                         rc = iter(res, closure);
1561
1562                         spin_lock(&ns->ns_hash_lock);
1563                         tmp = tmp->next;
1564                         ldlm_resource_putref_locked(res);
1565                         if (rc == LDLM_ITER_STOP)
1566                                 GOTO(out, rc);
1567                 }
1568         }
1569  out:
1570         spin_unlock(&ns->ns_hash_lock);
1571         RETURN(rc);
1572 }
1573
1574 /* non-blocking function to manipulate a lock whose cb_data is being put away.*/
1575 void ldlm_resource_iterate(struct ldlm_namespace *ns,
1576                            const struct ldlm_res_id *res_id,
1577                            ldlm_iterator_t iter, void *data)
1578 {
1579         struct ldlm_resource *res;
1580         ENTRY;
1581
1582         if (ns == NULL) {
1583                 CERROR("must pass in namespace\n");
1584                 LBUG();
1585         }
1586
1587         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1588         if (res == NULL) {
1589                 EXIT;
1590                 return;
1591         }
1592
1593         ldlm_resource_foreach(res, iter, data);
1594         ldlm_resource_putref(res);
1595         EXIT;
1596 }
1597
1598 /* Lock replay */
1599
1600 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
1601 {
1602         struct list_head *list = closure;
1603
1604         /* we use l_pending_chain here, because it's unused on clients. */
1605         LASSERTF(list_empty(&lock->l_pending_chain),"lock %p next %p prev %p\n",
1606                  lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
1607         /* bug 9573: don't replay locks left after eviction */
1608         if (!(lock->l_flags & LDLM_FL_FAILED))
1609                 list_add(&lock->l_pending_chain, list);
1610         return LDLM_ITER_CONTINUE;
1611 }
1612
1613 static int replay_lock_interpret(struct ptlrpc_request *req,
1614                                  struct ldlm_async_args *aa, int rc)
1615 {
1616         struct ldlm_lock *lock;
1617         struct ldlm_reply *reply;
1618
1619         ENTRY;
1620         atomic_dec(&req->rq_import->imp_replay_inflight);
1621         if (rc != ELDLM_OK)
1622                 GOTO(out, rc);
1623
1624
1625         reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF, sizeof(*reply),
1626                                    lustre_swab_ldlm_reply);
1627         if (reply == NULL) {
1628                 CERROR("Can't unpack ldlm_reply\n");
1629                 GOTO (out, rc = -EPROTO);
1630         }
1631
1632         lock = ldlm_handle2lock(&aa->lock_handle);
1633         if (!lock) {
1634                 CERROR("received replay ack for unknown local cookie "LPX64
1635                        " remote cookie "LPX64 " from server %s id %s\n",
1636                        aa->lock_handle.cookie, reply->lock_handle.cookie,
1637                        req->rq_export->exp_client_uuid.uuid,
1638                        libcfs_id2str(req->rq_peer));
1639                 GOTO(out, rc = -ESTALE);
1640         }
1641
1642         lock->l_remote_handle = reply->lock_handle;
1643         LDLM_DEBUG(lock, "replayed lock:");
1644         ptlrpc_import_recovery_state_machine(req->rq_import);
1645         LDLM_LOCK_PUT(lock);
1646 out:
1647         if (rc != ELDLM_OK)
1648                 ptlrpc_connect_import(req->rq_import, NULL);
1649
1650
1651         RETURN(rc);
1652 }
1653
1654 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
1655 {
1656         struct ptlrpc_request *req;
1657         struct ldlm_request *body;
1658         struct ldlm_reply *reply;
1659         struct ldlm_async_args *aa;
1660         int buffers = 2;
1661         int size[3] = { sizeof(struct ptlrpc_body) };
1662         int flags;
1663         ENTRY;
1664
1665
1666         /* Bug 11974: Do not replay a lock which is actively being canceled */
1667         if (lock->l_flags & LDLM_FL_CANCELING) {
1668                 LDLM_DEBUG(lock, "Not replaying canceled lock:");
1669                 RETURN(0);
1670         }
1671
1672         /* If this is reply-less callback lock, we cannot replay it, since
1673          * server might have long dropped it, but notification of that event was
1674          * lost by network. (and server granted conflicting lock already) */
1675         if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1676                 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
1677                 ldlm_lock_cancel(lock);
1678                 RETURN(0);
1679         }
1680         /*
1681          * If granted mode matches the requested mode, this lock is granted.
1682          *
1683          * If they differ, but we have a granted mode, then we were granted
1684          * one mode and now want another: ergo, converting.
1685          *
1686          * If we haven't been granted anything and are on a resource list,
1687          * then we're blocked/waiting.
1688          *
1689          * If we haven't been granted anything and we're NOT on a resource list,
1690          * then we haven't got a reply yet and don't have a known disposition.
1691          * This happens whenever a lock enqueue is the request that triggers
1692          * recovery.
1693          */
1694         if (lock->l_granted_mode == lock->l_req_mode)
1695                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
1696         else if (lock->l_granted_mode)
1697                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
1698         else if (!list_empty(&lock->l_res_link))
1699                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
1700         else
1701                 flags = LDLM_FL_REPLAY;
1702
1703         size[DLM_LOCKREQ_OFF] = sizeof(*body);
1704         req = ptlrpc_prep_req(imp, LUSTRE_DLM_VERSION, LDLM_ENQUEUE, 2, size,
1705                               NULL);
1706         if (!req)
1707                 RETURN(-ENOMEM);
1708
1709         /* We're part of recovery, so don't wait for it. */
1710         req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
1711
1712         body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, sizeof(*body));
1713         ldlm_lock2desc(lock, &body->lock_desc);
1714         body->lock_flags = flags;
1715
1716         ldlm_lock2handle(lock, &body->lock_handle[0]);
1717         size[DLM_LOCKREPLY_OFF] = sizeof(*reply);
1718         if (lock->l_lvb_len != 0) {
1719                 buffers = 3;
1720                 size[DLM_REPLY_REC_OFF] = lock->l_lvb_len;
1721         }
1722         ptlrpc_req_set_repsize(req, buffers, size);
1723         /* notify the server we've replayed all requests.
1724          * also, we mark the request to be put on a dedicated
1725          * queue to be processed after all request replayes.
1726          * bug 6063 */
1727         lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
1728
1729         LDLM_DEBUG(lock, "replaying lock:");
1730
1731         atomic_inc(&req->rq_import->imp_replay_inflight);
1732         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1733         aa = (struct ldlm_async_args *)&req->rq_async_args;
1734         aa->lock_handle = body->lock_handle[0];
1735         req->rq_interpret_reply = replay_lock_interpret;
1736         ptlrpcd_add_req(req);
1737
1738         RETURN(0);
1739 }
1740
1741 int ldlm_replay_locks(struct obd_import *imp)
1742 {
1743         struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
1744         struct list_head list;
1745         struct ldlm_lock *lock, *next;
1746         int rc = 0;
1747
1748         ENTRY;
1749         CFS_INIT_LIST_HEAD(&list);
1750
1751         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1752
1753         /* ensure this doesn't fall to 0 before all have been queued */
1754         atomic_inc(&imp->imp_replay_inflight);
1755
1756         (void)ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
1757
1758         list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
1759                 list_del_init(&lock->l_pending_chain);
1760                 if (rc)
1761                         continue; /* or try to do the rest? */
1762                 rc = replay_one_lock(imp, lock);
1763         }
1764
1765         atomic_dec(&imp->imp_replay_inflight);
1766
1767         RETURN(rc);
1768 }