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 static int is_granted_or_cancelled(struct ldlm_lock *lock)
  91 {
  92         int ret = 0;
  93
  94         lock_res_and_lock(lock);
  95         if (((lock->l_req_mode == lock->l_granted_mode) &&
  96              !(lock->l_flags & LDLM_FL_CP_REQD)) ||
  97             (lock->l_flags & LDLM_FL_FAILED))
  98                 ret = 1;
  99         unlock_res_and_lock(lock);
 100
 101         return ret;
 102 }
 103
 104 int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data)
 105 {
 106         /* XXX ALLOCATE - 160 bytes */
 107         struct lock_wait_data lwd;
 108         struct obd_device *obd;
 109         struct obd_import *imp = NULL;
 110         struct l_wait_info lwi;
 111         int rc = 0;
 112         ENTRY;
 113
 114         if (flags == LDLM_FL_WAIT_NOREPROC) {
 115                 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
 116                 goto noreproc;
 117         }
 118
 119         if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
 120                        LDLM_FL_BLOCK_CONV))) {
 121                 cfs_waitq_signal(&lock->l_waitq);
 122                 RETURN(0);
 123         }
 124
 125         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
 126                    "sleeping");
 127         ldlm_lock_dump(D_OTHER, lock, 0);
 128         ldlm_reprocess_all(lock->l_resource);
 129
 130 noreproc:
 131
 132         obd = class_exp2obd(lock->l_conn_export);
 133
 134         /* if this is a local lock, then there is no import */
 135         if (obd != NULL)
 136                 imp = obd->u.cli.cl_import;
 137
 138         lwd.lwd_lock = lock;
 139
 140         if (lock->l_flags & LDLM_FL_NO_TIMEOUT) {
 141                 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
 142                 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
 143         } else {
 144                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
 145                                        ldlm_expired_completion_wait,
 146                                        interrupted_completion_wait, &lwd);
 147         }
 148
 149         if (imp != NULL) {
 150                 spin_lock(&imp->imp_lock);
 151                 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
 152                 spin_unlock(&imp->imp_lock);
 153         }
 154
 155         /* Go to sleep until the lock is granted or cancelled. */
 156         rc = l_wait_event(lock->l_waitq, is_granted_or_cancelled(lock), &lwi);
 157
 158         if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED) {
 159                 LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed");
 160                 RETURN(-EIO);
 161         }
 162
 163         if (rc) {
 164                 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
 165                            rc);
 166                 RETURN(rc);
 167         }
 168
 169         LDLM_DEBUG(lock, "client-side enqueue waking up: granted");
 170         RETURN(0);
 171 }
 172
 173 /*
 174  * ->l_blocking_ast() callback for LDLM locks acquired by server-side OBDs.
 175  */
 176 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 177                       void *data, int flag)
 178 {
 179         int do_ast;
 180         ENTRY;
 181
 182         if (flag == LDLM_CB_CANCELING) {
 183                 /* Don't need to do anything here. */
 184                 RETURN(0);
 185         }
 186
 187         lock_res_and_lock(lock);
 188         /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
 189          * that ldlm_blocking_ast is called just before intent_policy method
 190          * takes the ns_lock, then by the time we get the lock, we might not
 191          * be the correct blocking function anymore.  So check, and return
 192          * early, if so. */
 193         if (lock->l_blocking_ast != ldlm_blocking_ast) {
 194                 unlock_res_and_lock(lock);
 195                 RETURN(0);
 196         }
 197
 198         lock->l_flags |= LDLM_FL_CBPENDING;
 199         do_ast = (!lock->l_readers && !lock->l_writers);
 200         unlock_res_and_lock(lock);
 201
 202         if (do_ast) {
 203                 struct lustre_handle lockh;
 204                 int rc;
 205
 206                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
 207                 ldlm_lock2handle(lock, &lockh);
 208                 rc = ldlm_cli_cancel(&lockh);
 209                 if (rc < 0)
 210                         CERROR("ldlm_cli_cancel: %d\n", rc);
 211         } else {
 212                 LDLM_DEBUG(lock, "Lock still has references, will be "
 213                            "cancelled later");
 214         }
 215         RETURN(0);
 216 }
 217
 218 /*
 219  * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See
 220  * comment in filter_intent_policy() on why you may need this.
 221  */
 222 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
 223 {
 224         /*
 225          * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for
 226          * that is rather subtle: with OST-side locking, it may so happen that
 227          * _all_ extent locks are held by the OST. If client wants to obtain
 228          * current file size it calls ll{,u}_glimpse_size(), and (as locks are
 229          * on the server), dummy glimpse callback fires and does
 230          * nothing. Client still receives correct file size due to the
 231          * following fragment in filter_intent_policy():
 232          *
 233          * rc = l->l_glimpse_ast(l, NULL); // this will update the LVB
 234          * if (rc != 0 && res->lr_namespace->ns_lvbo &&
 235          *     res->lr_namespace->ns_lvbo->lvbo_update) {
 236          *         res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1);
 237          * }
 238          *
 239          * that is, after glimpse_ast() fails, filter_lvbo_update() runs, and
 240          * returns correct file size to the client.
 241          */
 242         return -ELDLM_NO_LOCK_DATA;
 243 }
 244
 245 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
 246                            const struct ldlm_res_id *res_id,
 247                            ldlm_type_t type, ldlm_policy_data_t *policy,
 248                            ldlm_mode_t mode, int *flags,
 249                            ldlm_blocking_callback blocking,
 250                            ldlm_completion_callback completion,
 251                            ldlm_glimpse_callback glimpse,
 252                            void *data, __u32 lvb_len, void *lvb_swabber,
 253                            struct lustre_handle *lockh)
 254 {
 255         struct ldlm_lock *lock;
 256         int err;
 257         ENTRY;
 258
 259         LASSERT(!(*flags & LDLM_FL_REPLAY));
 260         if (unlikely(ns_is_client(ns))) {
 261                 CERROR("Trying to enqueue local lock in a shadow namespace\n");
 262                 LBUG();
 263         }
 264
 265         lock = ldlm_lock_create(ns, res_id, type, mode, blocking,
 266                                 completion, glimpse, data, lvb_len);
 267         if (unlikely(!lock))
 268                 GOTO(out_nolock, err = -ENOMEM);
 269         LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
 270
 271         ldlm_lock_addref_internal(lock, mode);
 272         ldlm_lock2handle(lock, lockh);
 273         lock_res_and_lock(lock);
 274         lock->l_flags |= LDLM_FL_LOCAL;
 275         if (*flags & LDLM_FL_ATOMIC_CB)
 276                 lock->l_flags |= LDLM_FL_ATOMIC_CB;
 277         lock->l_lvb_swabber = lvb_swabber;
 278         unlock_res_and_lock(lock);
 279         if (policy != NULL)
 280                 lock->l_policy_data = *policy;
 281         if (type == LDLM_EXTENT)
 282                 lock->l_req_extent = policy->l_extent;
 283
 284         err = ldlm_lock_enqueue(ns, &lock, policy, flags);
 285         if (unlikely(err != ELDLM_OK))
 286                 GOTO(out, err);
 287
 288         if (policy != NULL)
 289                 *policy = lock->l_policy_data;
 290
 291         LDLM_DEBUG_NOLOCK("client-side local enqueue handler END (lock %p)",
 292                           lock);
 293
 294         if (lock->l_completion_ast)
 295                 lock->l_completion_ast(lock, *flags, NULL);
 296
 297         LDLM_DEBUG(lock, "client-side local enqueue END");
 298         EXIT;
 299  out:
 300         LDLM_LOCK_PUT(lock);
 301  out_nolock:
 302         return err;
 303 }
 304
 305 static void failed_lock_cleanup(struct ldlm_namespace *ns,
 306                                 struct ldlm_lock *lock,
 307                                 struct lustre_handle *lockh, int mode)
 308 {
 309         /* Set a flag to prevent us from sending a CANCEL (bug 407) */
 310         lock_res_and_lock(lock);
 311         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 312         unlock_res_and_lock(lock);
 313         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 314
 315         ldlm_lock_decref_and_cancel(lockh, mode);
 316
 317         /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
 318          *       from llite/file.c/ll_file_flock(). */
 319         if (lock->l_resource->lr_type == LDLM_FLOCK) {
 320                 ldlm_lock_destroy(lock);
 321         }
 322 }
 323
 324 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
 325                           ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
 326                           int *flags, void *lvb, __u32 lvb_len,
 327                           void *lvb_swabber, struct lustre_handle *lockh,int rc)
 328 {
 329         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 330         int is_replay = *flags & LDLM_FL_REPLAY;
 331         struct ldlm_lock *lock;
 332         struct ldlm_reply *reply;
 333         int cleanup_phase = 1;
 334         ENTRY;
 335
 336         lock = ldlm_handle2lock(lockh);
 337         /* ldlm_cli_enqueue is holding a reference on this lock. */
 338         if (!lock) {
 339                 LASSERT(type == LDLM_FLOCK);
 340                 RETURN(-ENOLCK);
 341         }
 342
 343         if (rc != ELDLM_OK) {
 344                 LASSERT(!is_replay);
 345                 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
 346                            rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
 347                 if (rc == ELDLM_LOCK_ABORTED) {
 348                         /* Before we return, swab the reply */
 349                         reply = req_capsule_server_get(&req->rq_pill,
 350                                                        &RMF_DLM_REP);
 351                         if (reply == NULL)
 352                                 rc = -EPROTO;
 353                         if (lvb_len) {
 354                                 struct ost_lvb *tmplvb;
 355
 356                                 req_capsule_set_size(&req->rq_pill,
 357                                                      &RMF_DLM_LVB, RCL_SERVER,
 358                                                      lvb_len);
 359                             tmplvb = req_capsule_server_swab_get(&req->rq_pill,
 360                                                                  &RMF_DLM_LVB,
 361                                                                  lvb_swabber);
 362                                 if (tmplvb == NULL)
 363                                         GOTO(cleanup, rc = -EPROTO);
 364                                 if (lvb != NULL)
 365                                         memcpy(lvb, tmplvb, lvb_len);
 366                         }
 367                 }
 368                 GOTO(cleanup, rc);
 369         }
 370
 371         reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
 372         if (reply == NULL)
 373                 GOTO(cleanup, rc = -EPROTO);
 374
 375         /* lock enqueued on the server */
 376         cleanup_phase = 0;
 377
 378         lock_res_and_lock(lock);
 379         lock->l_remote_handle = reply->lock_handle;
 380         *flags = reply->lock_flags;
 381         lock->l_flags |= reply->lock_flags & LDLM_INHERIT_FLAGS;
 382         /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match()
 383          * to wait with no timeout as well */
 384         lock->l_flags |= reply->lock_flags & LDLM_FL_NO_TIMEOUT;
 385         unlock_res_and_lock(lock);
 386
 387         CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%x\n",
 388                lock, reply->lock_handle.cookie, *flags);
 389
 390         /* If enqueue returned a blocked lock but the completion handler has
 391          * already run, then it fixed up the resource and we don't need to do it
 392          * again. */
 393         if ((*flags) & LDLM_FL_LOCK_CHANGED) {
 394                 int newmode = reply->lock_desc.l_req_mode;
 395                 LASSERT(!is_replay);
 396                 if (newmode && newmode != lock->l_req_mode) {
 397                         LDLM_DEBUG(lock, "server returned different mode %s",
 398                                    ldlm_lockname[newmode]);
 399                         lock->l_req_mode = newmode;
 400                 }
 401
 402                 if (memcmp(reply->lock_desc.l_resource.lr_name.name,
 403                           lock->l_resource->lr_name.name,
 404                           sizeof(struct ldlm_res_id))) {
 405                         CDEBUG(D_INFO, "remote intent success, locking "
 406                                         "(%ld,%ld,%ld) instead of "
 407                                         "(%ld,%ld,%ld)\n",
 408                               (long)reply->lock_desc.l_resource.lr_name.name[0],
 409                               (long)reply->lock_desc.l_resource.lr_name.name[1],
 410                               (long)reply->lock_desc.l_resource.lr_name.name[2],
 411                               (long)lock->l_resource->lr_name.name[0],
 412                               (long)lock->l_resource->lr_name.name[1],
 413                               (long)lock->l_resource->lr_name.name[2]);
 414
 415                         ldlm_lock_change_resource(ns, lock,
 416                                           &reply->lock_desc.l_resource.lr_name);
 417                         if (lock->l_resource == NULL) {
 418                                 LBUG();
 419                                 GOTO(cleanup, rc = -ENOMEM);
 420                         }
 421                         LDLM_DEBUG(lock, "client-side enqueue, new resource");
 422                 }
 423                 if (with_policy)
 424                         if (!(type == LDLM_IBITS && !(exp->exp_connect_flags &
 425                                                     OBD_CONNECT_IBITS)))
 426                                 lock->l_policy_data =
 427                                                  reply->lock_desc.l_policy_data;
 428                 if (type != LDLM_PLAIN)
 429                         LDLM_DEBUG(lock,"client-side enqueue, new policy data");
 430         }
 431
 432         if ((*flags) & LDLM_FL_AST_SENT ||
 433             /* Cancel extent locks as soon as possible on a liblustre client,
 434              * because it cannot handle asynchronous ASTs robustly (see
 435              * bug 7311). */
 436             (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) {
 437                 lock_res_and_lock(lock);
 438                 lock->l_flags |= LDLM_FL_CBPENDING |  LDLM_FL_BL_AST;
 439                 unlock_res_and_lock(lock);
 440                 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
 441         }
 442
 443         /* If the lock has already been granted by a completion AST, don't
 444          * clobber the LVB with an older one. */
 445         if (lvb_len && (lock->l_req_mode != lock->l_granted_mode)) {
 446                 void *tmplvb;
 447
 448                 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
 449                                      lvb_len);
 450                 tmplvb = req_capsule_server_swab_get(&req->rq_pill,
 451                                                      &RMF_DLM_LVB,
 452                                                      lvb_swabber);
 453                 if (tmplvb == NULL)
 454                         GOTO(cleanup, rc = -EPROTO);
 455                 memcpy(lock->l_lvb_data, tmplvb, lvb_len);
 456         }
 457
 458         if (!is_replay) {
 459                 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
 460                 if (lock->l_completion_ast != NULL) {
 461                         int err = lock->l_completion_ast(lock, *flags, NULL);
 462                         if (!rc)
 463                                 rc = err;
 464                         if (rc && type != LDLM_FLOCK) /* bug 9425, bug 10250 */
 465                                 cleanup_phase = 1;
 466                 }
 467         }
 468
 469         if (lvb_len && lvb != NULL) {
 470                 /* Copy the LVB here, and not earlier, because the completion
 471                  * AST (if any) can override what we got in the reply */
 472                 memcpy(lvb, lock->l_lvb_data, lvb_len);
 473         }
 474
 475         LDLM_DEBUG(lock, "client-side enqueue END");
 476         EXIT;
 477 cleanup:
 478         if (cleanup_phase == 1 && rc)
 479                 failed_lock_cleanup(ns, lock, lockh, mode);
 480         /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
 481         LDLM_LOCK_PUT(lock);
 482         LDLM_LOCK_PUT(lock);
 483         return rc;
 484 }
 485
 486 /* PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
 487  * a single page on the send/receive side. XXX: 512 should be changed
 488  * to more adequate value. */
 489 static inline int ldlm_req_handles_avail(struct obd_export *exp,
 490                                          int *size, int bufcount,
 491                                          int bufoff, int off)
 492 {
 493         int avail = min_t(int, LDLM_MAXREQSIZE, PAGE_SIZE - 512);
 494         int old_size = size[bufoff];
 495
 496         size[bufoff] = sizeof(struct ldlm_request);
 497         avail -= lustre_msg_size(class_exp2cliimp(exp)->imp_msg_magic,
 498                                  bufcount, size);
 499         avail /= sizeof(struct lustre_handle);
 500         avail += LDLM_LOCKREQ_HANDLES - off;
 501         size[bufoff] = old_size;
 502
 503         return avail;
 504 }
 505
 506 static inline int ldlm_cancel_handles_avail(struct obd_export *exp)
 507 {
 508         int size[2] = { sizeof(struct ptlrpc_body),
 509                         sizeof(struct ldlm_request) };
 510         return ldlm_req_handles_avail(exp, size, 2, DLM_LOCKREQ_OFF, 0);
 511 }
 512
 513 /* Cancel lru locks and pack them into the enqueue request. Pack there the given
 514  * @count locks in @cancels. */
 515 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
 516                       int version, int opc, int canceloff,
 517                       struct list_head *cancels, int count)
 518 {
 519         struct ldlm_namespace   *ns = exp->exp_obd->obd_namespace;
 520         struct req_capsule      *pill = &req->rq_pill;
 521         struct ldlm_request     *dlm = NULL;
 522         int flags, avail, to_free, bufcount, pack = 0;
 523         int rc;
 524         ENTRY;
 525
 526
 527         LASSERT(cancels != NULL);
 528
 529         if (exp_connect_cancelset(exp)) {
 530                 /* Estimate the amount of available space in the request. */
 531                 bufcount = req_capsule_filled_sizes(pill, RCL_CLIENT);
 532                 avail = ldlm_req_handles_avail(exp, pill->rc_area[RCL_CLIENT],
 533                                                bufcount, bufcount - 1, canceloff);
 534                 flags = ns_connect_lru_resize(ns) ?
 535                         LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
 536                 to_free = !ns_connect_lru_resize(ns) &&
 537                           opc == LDLM_ENQUEUE ? 1 : 0;
 538
 539                 /* Cancel lru locks here _only_ if the server supports
 540                  * EARLY_CANCEL. Otherwise we have to send extra CANCEL
 541                  * rpc, what will make us slower. */
 542                 if (avail > count)
 543                         count += ldlm_cancel_lru_local(ns, cancels, to_free,
 544                                                        avail - count, 0, flags);
 545                 if (avail > count)
 546                         pack = count;
 547                 else
 548                         pack = avail;
 549                 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
 550                                      ldlm_request_bufsize(count, opc));
 551         }
 552
 553         rc = ptlrpc_request_pack(req, version, opc);
 554         if (rc) {
 555                 ldlm_lock_list_put(cancels, l_bl_ast, count);
 556                 RETURN(rc);
 557         }
 558
 559         if (exp_connect_cancelset(exp)) {
 560                 if (canceloff) {
 561                         dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
 562                         LASSERT(dlm);
 563                         /* Skip first lock handler in ldlm_request_pack(),
 564                          * this method will incrment @lock_count according
 565                          * to the lock handle amount actually written to
 566                          * the buffer. */
 567                         dlm->lock_count = canceloff;
 568                 }
 569                 /* Pack into the request @pack lock handles. */
 570                 ldlm_cli_cancel_list(cancels, pack, req, 0);
 571                 /* Prepare and send separate cancel rpc for others. */
 572                 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
 573         } else {
 574                 ldlm_lock_list_put(cancels, l_bl_ast, count);
 575         }
 576         RETURN(0);
 577 }
 578
 579 int ldlm_prep_enqueue_req(struct obd_export *exp,
 580                           struct ptlrpc_request *req,
 581                           struct list_head *cancels,
 582                           int count)
 583 {
 584         return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
 585                                  LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
 586 }
 587
 588 /* If a request has some specific initialisation it is passed in @reqp,
 589  * otherwise it is created in ldlm_cli_enqueue.
 590  *
 591  * Supports sync and async requests, pass @async flag accordingly. If a
 592  * request was created in ldlm_cli_enqueue and it is the async request,
 593  * pass it to the caller in @reqp. */
 594 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
 595                      struct ldlm_enqueue_info *einfo,
 596                      const struct ldlm_res_id *res_id,
 597                      ldlm_policy_data_t *policy, int *flags,
 598                      void *lvb, __u32 lvb_len, void *lvb_swabber,
 599                      struct lustre_handle *lockh, int async)
 600 {
 601         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 602         struct ldlm_lock      *lock;
 603         struct ldlm_request   *body;
 604         int                    is_replay = *flags & LDLM_FL_REPLAY;
 605         int                    req_passed_in = 1;
 606         int                    rc, err;
 607         struct ptlrpc_request *req;
 608         ENTRY;
 609
 610         LASSERT(exp != NULL);
 611
 612         /* If we're replaying this lock, just check some invariants.
 613          * If we're creating a new lock, get everything all setup nice. */
 614         if (is_replay) {
 615                 lock = ldlm_handle2lock(lockh);
 616                 LASSERT(lock != NULL);
 617                 LDLM_DEBUG(lock, "client-side enqueue START");
 618                 LASSERT(exp == lock->l_conn_export);
 619         } else {
 620                 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
 621                                         einfo->ei_mode, einfo->ei_cb_bl,
 622                                         einfo->ei_cb_cp, einfo->ei_cb_gl,
 623                                         einfo->ei_cbdata, lvb_len);
 624                 if (lock == NULL)
 625                         RETURN(-ENOMEM);
 626                 /* for the local lock, add the reference */
 627                 ldlm_lock_addref_internal(lock, einfo->ei_mode);
 628                 ldlm_lock2handle(lock, lockh);
 629                 lock->l_lvb_swabber = lvb_swabber;
 630                 if (policy != NULL) {
 631                         /* INODEBITS_INTEROP: If the server does not support
 632                          * inodebits, we will request a plain lock in the
 633                          * descriptor (ldlm_lock2desc() below) but use an
 634                          * inodebits lock internally with both bits set.
 635                          */
 636                         if (einfo->ei_type == LDLM_IBITS &&
 637                             !(exp->exp_connect_flags & OBD_CONNECT_IBITS))
 638                                 lock->l_policy_data.l_inodebits.bits =
 639                                         MDS_INODELOCK_LOOKUP |
 640                                         MDS_INODELOCK_UPDATE;
 641                         else
 642                                 lock->l_policy_data = *policy;
 643                 }
 644
 645                 if (einfo->ei_type == LDLM_EXTENT)
 646                         lock->l_req_extent = policy->l_extent;
 647                 LDLM_DEBUG(lock, "client-side enqueue START");
 648         }
 649
 650         /* lock not sent to server yet */
 651
 652         if (reqp == NULL || *reqp == NULL) {
 653                 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
 654                                                 &RQF_LDLM_ENQUEUE,
 655                                                 LUSTRE_DLM_VERSION,
 656                                                 LDLM_ENQUEUE);
 657                 if (req == NULL) {
 658                         failed_lock_cleanup(ns, lock, lockh, einfo->ei_mode);
 659                         LDLM_LOCK_PUT(lock);
 660                         RETURN(-ENOMEM);
 661                 }
 662                 req_passed_in = 0;
 663                 if (reqp)
 664                         *reqp = req;
 665         } else {
 666                 int len;
 667
 668                 req = *reqp;
 669                 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
 670                                            RCL_CLIENT);
 671                 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
 672                          DLM_LOCKREQ_OFF, len, sizeof(*body));
 673         }
 674
 675         lock->l_conn_export = exp;
 676         lock->l_export = NULL;
 677         lock->l_blocking_ast = einfo->ei_cb_bl;
 678
 679         /* Dump lock data into the request buffer */
 680         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
 681         ldlm_lock2desc(lock, &body->lock_desc);
 682         body->lock_flags = *flags;
 683         body->lock_handle[0] = *lockh;
 684
 685         /* Continue as normal. */
 686         if (!req_passed_in) {
 687                 if (lvb_len > 0) {
 688                         req_capsule_extend(&req->rq_pill,
 689                                            &RQF_LDLM_ENQUEUE_LVB);
 690                         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
 691                                              RCL_SERVER, lvb_len);
 692                 }
 693                 ptlrpc_request_set_replen(req);
 694         }
 695
 696         /*
 697          * Liblustre client doesn't get extent locks, except for O_APPEND case
 698          * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
 699          * [i_size, OBD_OBJECT_EOF] lock is taken.
 700          */
 701         LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT ||
 702                      policy->l_extent.end == OBD_OBJECT_EOF));
 703
 704         if (async) {
 705                 LASSERT(reqp != NULL);
 706                 RETURN(0);
 707         }
 708
 709         LDLM_DEBUG(lock, "sending request");
 710         rc = ptlrpc_queue_wait(req);
 711         err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
 712                                     einfo->ei_mode, flags, lvb, lvb_len,
 713                                     lvb_swabber, lockh, rc);
 714
 715         /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
 716          * one reference that we took */
 717         if (err == -ENOLCK)
 718                 LDLM_LOCK_PUT(lock);
 719         else
 720                 rc = err;
 721
 722         if (!req_passed_in && req != NULL) {
 723                 ptlrpc_req_finished(req);
 724                 if (reqp)
 725                         *reqp = NULL;
 726         }
 727
 728         RETURN(rc);
 729 }
 730
 731 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
 732                                   __u32 *flags)
 733 {
 734         struct ldlm_resource *res;
 735         int rc;
 736         ENTRY;
 737         if (ns_is_client(lock->l_resource->lr_namespace)) {
 738                 CERROR("Trying to cancel local lock\n");
 739                 LBUG();
 740         }
 741         LDLM_DEBUG(lock, "client-side local convert");
 742
 743         res = ldlm_lock_convert(lock, new_mode, flags);
 744         if (res) {
 745                 ldlm_reprocess_all(res);
 746                 rc = 0;
 747         } else {
 748                 rc = EDEADLOCK;
 749         }
 750         LDLM_DEBUG(lock, "client-side local convert handler END");
 751         LDLM_LOCK_PUT(lock);
 752         RETURN(rc);
 753 }
 754
 755 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
 756  * conversion of locks which are on the waiting or converting queue */
 757 /* Caller of this code is supposed to take care of lock readers/writers
 758    accounting */
 759 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags)
 760 {
 761         struct ldlm_request   *body;
 762         struct ldlm_reply     *reply;
 763         struct ldlm_lock      *lock;
 764         struct ldlm_resource  *res;
 765         struct ptlrpc_request *req;
 766         int                    rc;
 767         ENTRY;
 768
 769         lock = ldlm_handle2lock(lockh);
 770         if (!lock) {
 771                 LBUG();
 772                 RETURN(-EINVAL);
 773         }
 774         *flags = 0;
 775
 776         if (lock->l_conn_export == NULL)
 777                 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
 778
 779         LDLM_DEBUG(lock, "client-side convert");
 780
 781         req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export),
 782                                         &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
 783                                         LDLM_CONVERT);
 784         if (req == NULL) {
 785                 LDLM_LOCK_PUT(lock);
 786                 RETURN(-ENOMEM);
 787         }
 788
 789         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
 790         body->lock_handle[0] = lock->l_remote_handle;
 791
 792         body->lock_desc.l_req_mode = new_mode;
 793         body->lock_flags = *flags;
 794
 795
 796         ptlrpc_request_set_replen(req);
 797         rc = ptlrpc_queue_wait(req);
 798         if (rc != ELDLM_OK)
 799                 GOTO(out, rc);
 800
 801         reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
 802         if (reply == NULL)
 803                 GOTO(out, rc = -EPROTO);
 804
 805         if (req->rq_status)
 806                 GOTO(out, rc = req->rq_status);
 807
 808         res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
 809         if (res != NULL) {
 810                 ldlm_reprocess_all(res);
 811                 /* Go to sleep until the lock is granted. */
 812                 /* FIXME: or cancelled. */
 813                 if (lock->l_completion_ast) {
 814                         rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
 815                                                     NULL);
 816                         if (rc)
 817                                 GOTO(out, rc);
 818                 }
 819         } else {
 820                 rc = EDEADLOCK;
 821         }
 822         EXIT;
 823  out:
 824         LDLM_LOCK_PUT(lock);
 825         ptlrpc_req_finished(req);
 826         return rc;
 827 }
 828
 829 /* Cancel locks locally.
 830  * Returns:
 831  * LDLM_FL_LOCAL_ONLY if tere is no need in a CANCEL rpc to the server;
 832  * LDLM_FL_CANCELING otherwise;
 833  * LDLM_FL_BL_AST if there is a need in a separate CANCEL rpc. */
 834 static int ldlm_cli_cancel_local(struct ldlm_lock *lock)
 835 {
 836         int rc = LDLM_FL_LOCAL_ONLY;
 837         ENTRY;
 838
 839         if (lock->l_conn_export) {
 840                 int local_only;
 841
 842                 LDLM_DEBUG(lock, "client-side cancel");
 843                 /* Set this flag to prevent others from getting new references*/
 844                 lock_res_and_lock(lock);
 845                 lock->l_flags |= LDLM_FL_CBPENDING;
 846                 local_only = (lock->l_flags &
 847                               (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
 848                 ldlm_cancel_callback(lock);
 849                 rc = (lock->l_flags & LDLM_FL_BL_AST) ?
 850                         LDLM_FL_BL_AST : LDLM_FL_CANCELING;
 851                 unlock_res_and_lock(lock);
 852
 853                 if (local_only) {
 854                         CDEBUG(D_DLMTRACE, "not sending request (at caller's "
 855                                "instruction)\n");
 856                         rc = LDLM_FL_LOCAL_ONLY;
 857                 }
 858                 ldlm_lock_cancel(lock);
 859         } else {
 860                 if (ns_is_client(lock->l_resource->lr_namespace)) {
 861                         LDLM_ERROR(lock, "Trying to cancel local lock");
 862                         LBUG();
 863                 }
 864                 LDLM_DEBUG(lock, "server-side local cancel");
 865                 ldlm_lock_cancel(lock);
 866                 ldlm_reprocess_all(lock->l_resource);
 867                 LDLM_DEBUG(lock, "server-side local cancel handler END");
 868         }
 869
 870         RETURN(rc);
 871 }
 872
 873 /* Pack @count locks in @head into ldlm_request buffer at the offset @off,
 874    of the request @req. */
 875 static void ldlm_cancel_pack(struct ptlrpc_request *req,
 876                              struct list_head *head, int count)
 877 {
 878         struct ldlm_request *dlm;
 879         struct ldlm_lock *lock;
 880         int max, packed = 0;
 881         ENTRY;
 882
 883         dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
 884         LASSERT(dlm != NULL);
 885
 886         /* Check the room in the request buffer. */
 887         max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
 888                 sizeof(struct ldlm_request);
 889         max /= sizeof(struct lustre_handle);
 890         max += LDLM_LOCKREQ_HANDLES;
 891         LASSERT(max >= dlm->lock_count + count);
 892
 893         /* XXX: it would be better to pack lock handles grouped by resource.
 894          * so that the server cancel would call filter_lvbo_update() less
 895          * frequently. */
 896         list_for_each_entry(lock, head, l_bl_ast) {
 897                 if (!count--)
 898                         break;
 899                 LASSERT(lock->l_conn_export);
 900                 /* Pack the lock handle to the given request buffer. */
 901                 LDLM_DEBUG(lock, "packing");
 902                 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
 903                 packed++;
 904         }
 905         CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
 906         EXIT;
 907 }
 908
 909 /* Prepare and send a batched cancel rpc, it will include count lock handles
 910  * of locks given in @head. */
 911 int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
 912                         int count, int flags)
 913 {
 914         struct ptlrpc_request *req = NULL;
 915         struct obd_import *imp;
 916         int free, sent = 0;
 917         int rc = 0;
 918         ENTRY;
 919
 920         LASSERT(exp != NULL);
 921         LASSERT(count > 0);
 922
 923         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
 924                 RETURN(count);
 925
 926         while (1) {
 927                 int bufcount;
 928                 struct req_capsule *pill;
 929                 imp = class_exp2cliimp(exp);
 930                 if (imp == NULL || imp->imp_invalid) {
 931                         CDEBUG(D_DLMTRACE,
 932                                "skipping cancel on invalid import %p\n", imp);
 933                         RETURN(count);
 934                 }
 935
 936                 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
 937                 if (req == NULL)
 938                         GOTO(out, rc = -ENOMEM);
 939
 940                 pill = &req->rq_pill;
 941                 bufcount = req_capsule_filled_sizes(pill, RCL_CLIENT);
 942
 943                 free = ldlm_req_handles_avail(exp, pill->rc_area[RCL_CLIENT],
 944                                               bufcount, bufcount, 0);
 945                 if (count > free)
 946                         count = free;
 947
 948                 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
 949                                      ldlm_request_bufsize(count, LDLM_CANCEL));
 950
 951                 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
 952                 if (rc) {
 953                         ptlrpc_request_free(req);
 954                         GOTO(out, rc);
 955                 }
 956                 req->rq_no_resend = 1;
 957                 req->rq_no_delay = 1;
 958
 959                 /* XXX FIXME bug 249 */
 960                 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
 961                 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
 962
 963                 ldlm_cancel_pack(req, cancels, count);
 964
 965                 ptlrpc_request_set_replen(req);
 966                 if (flags & LDLM_FL_ASYNC) {
 967                         ptlrpcd_add_req(req);
 968                         sent = count;
 969                         GOTO(out, 0);
 970                 } else {
 971                         rc = ptlrpc_queue_wait(req);
 972                 }
 973                 if (rc == ESTALE) {
 974                         CDEBUG(D_DLMTRACE, "client/server (nid %s) "
 975                                "out of sync -- not fatal\n",
 976                                libcfs_nid2str(req->rq_import->
 977                                               imp_connection->c_peer.nid));
 978                         rc = 0;
 979                 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
 980                            req->rq_import_generation == imp->imp_generation) {
 981                         ptlrpc_req_finished(req);
 982                         continue;
 983                 } else if (rc != ELDLM_OK) {
 984                         CERROR("Got rc %d from cancel RPC: canceling "
 985                                "anyway\n", rc);
 986                         break;
 987                 }
 988                 sent = count;
 989                 break;
 990         }
 991
 992         ptlrpc_req_finished(req);
 993         EXIT;
 994 out:
 995         return sent ? sent : rc;
 996 }
 997
 998 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
 999 {
1000         LASSERT(imp != NULL);
1001         return &imp->imp_obd->obd_namespace->ns_pool;
1002 }
1003
1004 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1005 {
1006         struct ldlm_pool *pl;
1007         ENTRY;
1008
1009         if (!imp_connect_lru_resize(req->rq_import))
1010                 RETURN(0);
1011
1012         if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1013             lustre_msg_get_limit(req->rq_repmsg) == 0)
1014                 RETURN(0);
1015
1016         pl = ldlm_imp2pl(req->rq_import);
1017
1018         spin_lock(&pl->pl_lock);
1019
1020         /* Check if we need to wakeup pools thread for fast SLV change.
1021          * This is only done when threads period is noticably long like
1022          * 10s or more. */
1023 #if defined(__KERNEL__) && (LDLM_POOLS_THREAD_PERIOD >= 10)
1024         {
1025                 __u64 old_slv, new_slv, fast_change;
1026
1027                 old_slv = ldlm_pool_get_slv(pl);
1028                 new_slv = lustre_msg_get_slv(req->rq_repmsg);
1029                 fast_change = old_slv * LDLM_POOLS_FAST_SLV_CHANGE;
1030                 do_div(fast_change, 100);
1031
1032                 /* Wake up pools thread only if SLV has changed more than
1033                  * 50% since last update. In this case we want to react asap.
1034                  * Otherwise it is no sense to wake up pools as they are
1035                  * re-calculated every LDLM_POOLS_THREAD_PERIOD anyways. */
1036                 if (old_slv > new_slv && old_slv - new_slv > fast_change)
1037                         ldlm_pools_wakeup();
1038         }
1039 #endif
1040         /* In some cases RPC may contain slv and limit zeroed out. This is
1041          * the case when server does not support lru resize feature. This is
1042          * also possible in some recovery cases when server side reqs have no
1043          * ref to obd export and thus access to server side namespace is no
1044          * possible. */
1045         if (lustre_msg_get_slv(req->rq_repmsg) != 0 &&
1046             lustre_msg_get_limit(req->rq_repmsg) != 0) {
1047                 ldlm_pool_set_slv(pl, lustre_msg_get_slv(req->rq_repmsg));
1048                 ldlm_pool_set_limit(pl, lustre_msg_get_limit(req->rq_repmsg));
1049         } else {
1050                 DEBUG_REQ(D_HA, req, "zero SLV or Limit found "
1051                           "(SLV: "LPU64", Limit: %u)",
1052                           lustre_msg_get_slv(req->rq_repmsg),
1053                           lustre_msg_get_limit(req->rq_repmsg));
1054         }
1055         spin_unlock(&pl->pl_lock);
1056
1057         RETURN(0);
1058 }
1059 EXPORT_SYMBOL(ldlm_cli_update_pool);
1060
1061 int ldlm_cli_cancel(struct lustre_handle *lockh)
1062 {
1063         int avail, flags, count = 1, rc = 0;
1064         struct ldlm_namespace *ns;
1065         struct ldlm_lock *lock;
1066         CFS_LIST_HEAD(cancels);
1067         ENTRY;
1068
1069         /* concurrent cancels on the same handle can happen */
1070         lock = __ldlm_handle2lock(lockh, LDLM_FL_CANCELING);
1071         if (lock == NULL) {
1072                 LDLM_DEBUG_NOLOCK("lock is already being destroyed\n");
1073                 RETURN(0);
1074         }
1075
1076         rc = ldlm_cli_cancel_local(lock);
1077         if (rc < 0 || rc == LDLM_FL_LOCAL_ONLY) {
1078                 LDLM_LOCK_PUT(lock);
1079                 RETURN(rc < 0 ? rc : 0);
1080         }
1081         /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1082          * rpc which goes to canceld portal, so we can cancel other lru locks
1083          * here and send them all as one LDLM_CANCEL rpc. */
1084         LASSERT(list_empty(&lock->l_bl_ast));
1085         list_add(&lock->l_bl_ast, &cancels);
1086         if (exp_connect_cancelset(lock->l_conn_export)) {
1087                 avail = ldlm_cancel_handles_avail(lock->l_conn_export);
1088                 LASSERT(avail > 0);
1089
1090                 ns = lock->l_resource->lr_namespace;
1091                 flags = ns_connect_lru_resize(ns) ?
1092                         LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
1093                 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1094                                                LDLM_FL_BL_AST, flags);
1095         }
1096         ldlm_cli_cancel_list(&cancels, count, NULL, 0);
1097         RETURN(0);
1098 }
1099
1100 /* XXX until we will have compound requests and can cut cancels from generic rpc
1101  * we need send cancels with LDLM_FL_BL_AST flag as separate rpc */
1102 static int ldlm_cancel_list(struct list_head *cancels, int count, int flags)
1103 {
1104         CFS_LIST_HEAD(head);
1105         struct ldlm_lock *lock, *next;
1106         int left = 0, bl_ast = 0, rc;
1107
1108         left = count;
1109         list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1110                 if (left-- == 0)
1111                         break;
1112
1113                 if (flags & LDLM_FL_LOCAL_ONLY) {
1114                         rc = LDLM_FL_LOCAL_ONLY;
1115                         ldlm_lock_cancel(lock);
1116                 } else {
1117                         rc = ldlm_cli_cancel_local(lock);
1118                 }
1119                 if (!(flags & LDLM_FL_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1120                         LDLM_DEBUG(lock, "Cancel lock separately");
1121                         list_del_init(&lock->l_bl_ast);
1122                         list_add(&lock->l_bl_ast, &head);
1123                         bl_ast ++;
1124                         continue;
1125                 }
1126                 if (rc == LDLM_FL_LOCAL_ONLY) {
1127                         /* CANCEL RPC should not be sent to server. */
1128                         list_del_init(&lock->l_bl_ast);
1129                         LDLM_LOCK_PUT(lock);
1130                         count--;
1131                 }
1132
1133         }
1134         if (bl_ast > 0) {
1135                 count -= bl_ast;
1136                 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1137         }
1138
1139         RETURN(count);
1140 }
1141
1142 /* Return 1 if @lock should be canceled according to shrinker policy.
1143  * Return zero otherwise. */
1144 static int ldlm_cancel_shrink_policy(struct ldlm_namespace *ns,
1145                                      struct ldlm_lock *lock,
1146                                      int unused, int added,
1147                                      int asked)
1148 {
1149         int lock_cost;
1150         __u64 page_nr;
1151
1152         if (lock->l_resource->lr_type == LDLM_EXTENT) {
1153                 struct ldlm_extent *l_extent;
1154
1155                 /* For all extent locks cost is 1 + number of pages in
1156                  * their extent. */
1157                 l_extent = &lock->l_policy_data.l_extent;
1158                 page_nr = (l_extent->end - l_extent->start);
1159                 do_div(page_nr, CFS_PAGE_SIZE);
1160
1161 #ifdef __KERNEL__
1162                 /* XXX: In fact this is evil hack, we can't access inode
1163                  * here. For doing it right we need somehow to have number
1164                  * of covered by lock. This should be fixed later when 10718
1165                  * is landed. */
1166                 if (lock->l_ast_data != NULL) {
1167                         struct inode *inode = lock->l_ast_data;
1168                         if (page_nr > inode->i_mapping->nrpages)
1169                                 page_nr = inode->i_mapping->nrpages;
1170                 }
1171 #endif
1172                 lock_cost = 1 + page_nr;
1173         } else {
1174                 /* For all locks which are not extent ones cost is 1 */
1175                 lock_cost = 1;
1176         }
1177
1178         /* Keep all expensive locks in lru for the memory pressure time
1179          * cancel policy. They anyways may be canceled by lru resize
1180          * pplicy if they have not small enough CLV. */
1181         return (lock_cost <= ns->ns_shrink_thumb);
1182 }
1183
1184 /* Return 1 if @lock should be canceled according to lru resize policy.
1185  * Return zero otherwise. */
1186 static int ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1187                                    struct ldlm_lock *lock,
1188                                    int unused, int added,
1189                                    int asked)
1190 {
1191         cfs_time_t cur = cfs_time_current();
1192         struct ldlm_pool *pl = &ns->ns_pool;
1193         __u64 slv, lvf, lv;
1194         cfs_time_t la;
1195
1196         spin_lock(&pl->pl_lock);
1197         slv = ldlm_pool_get_slv(pl);
1198         lvf = atomic_read(&pl->pl_lock_volume_factor);
1199         spin_unlock(&pl->pl_lock);
1200
1201         la = cfs_duration_sec(cfs_time_sub(cur,
1202                               lock->l_last_used));
1203
1204         /* Stop when slv is not yet come from server or
1205          * lv is smaller than it is. */
1206         lv = lvf * la * unused;
1207         return (slv > 1 && lv >= slv);
1208 }
1209
1210 /* Return 1 if @lock should be canceled according to passed policy.
1211  * Return zero otherwise. */
1212 static int ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1213                                      struct ldlm_lock *lock,
1214                                      int unused, int added,
1215                                      int asked)
1216 {
1217         /* Do nothing here, we allow canceling all locks which
1218          * are passed here from upper layer logic. So that locks
1219          * number to be canceled will be limited by @count and
1220          * @max in ldlm_cancel_lru_local(). */
1221         return 1;
1222 }
1223
1224 /* Return 1 if @lock should be canceled according to aged policy.
1225  * Return zero otherwise. */
1226 static int ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1227                                    struct ldlm_lock *lock,
1228                                    int unused, int added,
1229                                    int asked)
1230 {
1231         /* Cancel old locks if reached asked limit. */
1232         return !((added >= asked) &&
1233                  cfs_time_before_64(cfs_time_current(),
1234                                     cfs_time_add(lock->l_last_used,
1235                                                  ns->ns_max_age)));
1236 }
1237
1238 typedef int (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *,
1239                                         struct ldlm_lock *, int,
1240                                         int, int);
1241
1242 static ldlm_cancel_lru_policy_t
1243 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
1244 {
1245         if (ns_connect_lru_resize(ns)) {
1246                 if (flags & LDLM_CANCEL_SHRINK)
1247                         return ldlm_cancel_shrink_policy;
1248                 else if (flags & LDLM_CANCEL_LRUR)
1249                         return ldlm_cancel_lrur_policy;
1250                 else if (flags & LDLM_CANCEL_PASSED)
1251                         return ldlm_cancel_passed_policy;
1252         } else {
1253                 if (flags & LDLM_CANCEL_AGED)
1254                         return ldlm_cancel_aged_policy;
1255         }
1256         return NULL;
1257 }
1258
1259 /* - Free space in lru for @count new locks,
1260  *   redundant unused locks are canceled locally;
1261  * - also cancel locally unused aged locks;
1262  * - do not cancel more than @max locks;
1263  * - GET the found locks and add them into the @cancels list.
1264  *
1265  * A client lock can be added to the l_bl_ast list only when it is
1266  * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing CANCEL.
1267  * There are the following use cases: ldlm_cancel_resource_local(),
1268  * ldlm_cancel_lru_local() and ldlm_cli_cancel(), which check&set this
1269  * flag properly. As any attempt to cancel a lock rely on this flag,
1270  * l_bl_ast list is accessed later without any special locking.
1271  *
1272  * Calling policies for enabled lru resize:
1273  * ----------------------------------------
1274  * flags & LDLM_CANCEL_LRUR - use lru resize policy (SLV from server) to
1275  *                            cancel not more than @count locks;
1276  *
1277  * flags & LDLM_CANCEL_PASSED - cancel @count number of old locks (located at
1278  *                              the beginning of lru list);
1279  *
1280  * flags & LDLM_CANCEL_SHRINK - cancel not more than @count locks according to
1281  *                              memory pressre policy function.
1282  */
1283 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
1284                           int count, int max, int cancel_flags, int flags)
1285 {
1286         ldlm_cancel_lru_policy_t cancel_lru_policy_func;
1287         int added = 0, unused, cancel;
1288         struct ldlm_lock *lock, *next;
1289         ENTRY;
1290
1291         spin_lock(&ns->ns_unused_lock);
1292         unused = ns->ns_nr_unused;
1293
1294         if (!ns_connect_lru_resize(ns))
1295                 count += unused - ns->ns_max_unused;
1296
1297         cancel_lru_policy_func = ldlm_cancel_lru_policy(ns, flags);
1298
1299         list_for_each_entry_safe(lock, next, &ns->ns_unused_list, l_lru) {
1300                 /* Make sure that we skip locks being already in cancel. */
1301                 if ((lock->l_flags & LDLM_FL_CANCELING) ||
1302                     (lock->l_flags & LDLM_FL_BL_AST))
1303                         continue;
1304
1305                 /* For any flags, stop scanning if @max or passed @count is
1306                  * reached. */
1307                 if ((max && added >= max) || (count && added >= count))
1308                         break;
1309
1310                 /* Pass the lock through the policy filter and see if it
1311                  * should stay in lru. */
1312                 if (cancel_lru_policy_func != NULL) {
1313                         cancel = cancel_lru_policy_func(ns, lock, unused,
1314                                                         added, count);
1315
1316                         /* Take next lock for shrink policy, we need to check
1317                          * whole list. Stop scanning for other policies. */
1318                         if ((flags & LDLM_CANCEL_SHRINK) && !cancel)
1319                                 continue;
1320                         else if (!cancel)
1321                                 break;
1322                 }
1323
1324                 if (cancels != NULL) {
1325                         LDLM_LOCK_GET(lock); /* dropped by bl thread */
1326                         spin_unlock(&ns->ns_unused_lock);
1327
1328                         lock_res_and_lock(lock);
1329                         /* Check flags again under the lock. */
1330                         if ((lock->l_flags & LDLM_FL_CANCELING) ||
1331                             (lock->l_flags & LDLM_FL_BL_AST) ||
1332                             (ldlm_lock_remove_from_lru(lock) == 0)) {
1333                                 /* other thread is removing lock from lru or
1334                                  * somebody is already doing CANCEL or
1335                                  * there is a blocking request which will send
1336                                  * cancel by itseft. */
1337                                 unlock_res_and_lock(lock);
1338                                 LDLM_LOCK_PUT(lock);
1339                                 spin_lock(&ns->ns_unused_lock);
1340                                 continue;
1341                         }
1342                         LASSERT(!lock->l_readers && !lock->l_writers);
1343
1344                         /* If we have chosen to cancel this lock voluntarily, we
1345                          * better send cancel notification to server, so that it
1346                          * frees appropriate state. This might lead to a race
1347                          * where while we are doing cancel here, server is also
1348                          * silently cancelling this lock. */
1349                         lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK;
1350
1351                         /* Setting the CBPENDING flag is a little misleading,
1352                          * but prevents an important race; namely, once
1353                          * CBPENDING is set, the lock can accumulate no more
1354                          * readers/writers. Since readers and writers are
1355                          * already zero here, ldlm_lock_decref() won't see
1356                          * this flag and call l_blocking_ast */
1357                         lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1358
1359                         /* We can't re-add to l_lru as it confuses the
1360                          * refcounting in ldlm_lock_remove_from_lru() if an AST
1361                          * arrives after we drop ns_lock below. We use l_bl_ast
1362                          * and can't use l_pending_chain as it is used both on
1363                          * server and client nevertheless bug 5666 says it is
1364                          * used only on server */
1365                         LASSERT(list_empty(&lock->l_bl_ast));
1366                         list_add(&lock->l_bl_ast, cancels);
1367                         unlock_res_and_lock(lock);
1368                         spin_lock(&ns->ns_unused_lock);
1369                 }
1370                 added++;
1371                 unused--;
1372         }
1373         spin_unlock(&ns->ns_unused_lock);
1374
1375         if (cancels == NULL)
1376                 RETURN(added);
1377
1378         RETURN(ldlm_cancel_list(cancels, added, cancel_flags));
1379 }
1380
1381 /* when called with LDLM_ASYNC the blocking callback will be handled
1382  * in a thread and this function will return after the thread has been
1383  * asked to call the callback.  when called with LDLM_SYNC the blocking
1384  * callback will be performed in this function. */
1385 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr, ldlm_sync_t sync,
1386                     int flags)
1387 {
1388         CFS_LIST_HEAD(cancels);
1389         int count, rc;
1390         ENTRY;
1391
1392 #ifndef __KERNEL__
1393         sync = LDLM_SYNC; /* force to be sync in user space */
1394 #endif
1395         count = ldlm_cancel_lru_local(ns, &cancels, nr, 0, 0, flags);
1396         if (sync == LDLM_ASYNC) {
1397                 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count);
1398                 if (rc == 0)
1399                         RETURN(count);
1400         }
1401
1402         /* If an error occured in ASYNC mode, or
1403          * this is SYNC mode, cancel the list. */
1404         ldlm_cli_cancel_list(&cancels, count, NULL, 0);
1405         RETURN(count);
1406 }
1407
1408 /* Find and cancel locally unused locks found on resource, matched to the
1409  * given policy, mode. GET the found locks and add them into the @cancels
1410  * list. */
1411 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1412                                struct list_head *cancels,
1413                                ldlm_policy_data_t *policy,
1414                                ldlm_mode_t mode, int lock_flags,
1415                                int cancel_flags, void *opaque)
1416 {
1417         struct ldlm_lock *lock;
1418         int count = 0;
1419         ENTRY;
1420
1421         lock_res(res);
1422         list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1423                 if (opaque != NULL && lock->l_ast_data != opaque) {
1424                         LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1425                                    lock->l_ast_data, opaque);
1426                         //LBUG();
1427                         continue;
1428                 }
1429
1430                 if (lock->l_readers || lock->l_writers) {
1431                         if (cancel_flags & LDLM_FL_WARN) {
1432                                 LDLM_ERROR(lock, "lock in use");
1433                                 //LBUG();
1434                         }
1435                         continue;
1436                 }
1437
1438                 /* If somebody is already doing CANCEL, or blocking ast came,
1439                  * skip this lock. */
1440                 if (lock->l_flags & LDLM_FL_BL_AST ||
1441                     lock->l_flags & LDLM_FL_CANCELING)
1442                         continue;
1443
1444                 if (lockmode_compat(lock->l_granted_mode, mode))
1445                         continue;
1446
1447                 /* If policy is given and this is IBITS lock, add to list only
1448                  * those locks that match by policy. */
1449                 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1450                     !(lock->l_policy_data.l_inodebits.bits &
1451                       policy->l_inodebits.bits))
1452                         continue;
1453
1454                 /* See CBPENDING comment in ldlm_cancel_lru */
1455                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1456                                  lock_flags;
1457
1458                 LASSERT(list_empty(&lock->l_bl_ast));
1459                 list_add(&lock->l_bl_ast, cancels);
1460                 LDLM_LOCK_GET(lock);
1461                 count++;
1462         }
1463         unlock_res(res);
1464
1465         RETURN(ldlm_cancel_list(cancels, count, cancel_flags));
1466 }
1467
1468 /* If @req is NULL, send CANCEL request to server with handles of locks
1469  * in the @cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1470  * separately per lock.
1471  * If @req is not NULL, put handles of locks in @cancels into the request
1472  * buffer at the offset @off.
1473  * Destroy @cancels at the end. */
1474 int ldlm_cli_cancel_list(struct list_head *cancels, int count,
1475                          struct ptlrpc_request *req, int flags)
1476 {
1477         struct ldlm_lock *lock;
1478         int res = 0;
1479         ENTRY;
1480
1481         if (list_empty(cancels) || count == 0)
1482                 RETURN(0);
1483
1484         /* XXX: requests (both batched and not) could be sent in parallel.
1485          * Usually it is enough to have just 1 RPC, but it is possible that
1486          * there are to many locks to be cancelled in LRU or on a resource.
1487          * It would also speed up the case when the server does not support
1488          * the feature. */
1489         while (count > 0) {
1490                 LASSERT(!list_empty(cancels));
1491                 lock = list_entry(cancels->next, struct ldlm_lock, l_bl_ast);
1492                 LASSERT(lock->l_conn_export);
1493
1494                 if (exp_connect_cancelset(lock->l_conn_export)) {
1495                         res = count;
1496                         if (req)
1497                                 ldlm_cancel_pack(req, cancels, count);
1498                         else
1499                                 res = ldlm_cli_cancel_req(lock->l_conn_export,
1500                                                           cancels, count,
1501                                                           flags);
1502                 } else {
1503                         res = ldlm_cli_cancel_req(lock->l_conn_export,
1504                                                   cancels, 1, flags);
1505                 }
1506
1507                 if (res < 0) {
1508                         CERROR("ldlm_cli_cancel_list: %d\n", res);
1509                         res = count;
1510                 }
1511
1512                 count -= res;
1513                 ldlm_lock_list_put(cancels, l_bl_ast, res);
1514         }
1515         LASSERT(count == 0);
1516         RETURN(0);
1517 }
1518
1519 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1520                                     const struct ldlm_res_id *res_id,
1521                                     ldlm_policy_data_t *policy,
1522                                     ldlm_mode_t mode, int flags, void *opaque)
1523 {
1524         struct ldlm_resource *res;
1525         CFS_LIST_HEAD(cancels);
1526         int count;
1527         int rc;
1528         ENTRY;
1529
1530         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1531         if (res == NULL) {
1532                 /* This is not a problem. */
1533                 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id->name[0]);
1534                 RETURN(0);
1535         }
1536
1537         count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
1538                                            0, flags, opaque);
1539         rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
1540         if (rc != ELDLM_OK)
1541                 CERROR("ldlm_cli_cancel_unused_resource: %d\n", rc);
1542
1543         ldlm_resource_putref(res);
1544         RETURN(0);
1545 }
1546
1547 static inline int have_no_nsresource(struct ldlm_namespace *ns)
1548 {
1549         int no_resource = 0;
1550
1551         spin_lock(&ns->ns_hash_lock);
1552         if (ns->ns_resources == 0)
1553                 no_resource = 1;
1554         spin_unlock(&ns->ns_hash_lock);
1555
1556         RETURN(no_resource);
1557 }
1558
1559 /* Cancel all locks on a namespace (or a specific resource, if given)
1560  * that have 0 readers/writers.
1561  *
1562  * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
1563  * to notify the server. */
1564 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1565                            const struct ldlm_res_id *res_id,
1566                            int flags, void *opaque)
1567 {
1568         int i;
1569         ENTRY;
1570
1571         if (ns == NULL)
1572                 RETURN(ELDLM_OK);
1573
1574         if (res_id)
1575                 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
1576                                                        LCK_MINMODE, flags,
1577                                                        opaque));
1578
1579         spin_lock(&ns->ns_hash_lock);
1580         for (i = 0; i < RES_HASH_SIZE; i++) {
1581                 struct list_head *tmp;
1582                 tmp = ns->ns_hash[i].next;
1583                 while (tmp != &(ns->ns_hash[i])) {
1584                         struct ldlm_resource *res;
1585                         int rc;
1586
1587                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
1588                         ldlm_resource_getref(res);
1589                         spin_unlock(&ns->ns_hash_lock);
1590
1591                         rc = ldlm_cli_cancel_unused_resource(ns, &res->lr_name,
1592                                                              NULL, LCK_MINMODE,
1593                                                              flags, opaque);
1594
1595                         if (rc)
1596                                 CERROR("ldlm_cli_cancel_unused ("LPU64"): %d\n",
1597                                        res->lr_name.name[0], rc);
1598
1599                         spin_lock(&ns->ns_hash_lock);
1600                         tmp = tmp->next;
1601                         ldlm_resource_putref_locked(res);
1602                 }
1603         }
1604         spin_unlock(&ns->ns_hash_lock);
1605
1606         RETURN(ELDLM_OK);
1607 }
1608
1609 /* join/split resource locks to/from lru list */
1610 int ldlm_cli_join_lru(struct ldlm_namespace *ns,
1611                       const struct ldlm_res_id *res_id, int join)
1612 {
1613         struct ldlm_resource *res;
1614         struct ldlm_lock *lock, *n;
1615         int count = 0;
1616         ENTRY;
1617
1618         LASSERT(ns_is_client(ns));
1619
1620         res = ldlm_resource_get(ns, NULL, res_id, LDLM_EXTENT, 0);
1621         if (res == NULL)
1622                 RETURN(count);
1623         LASSERT(res->lr_type == LDLM_EXTENT);
1624
1625         lock_res(res);
1626         if (!join)
1627                 goto split;
1628
1629         list_for_each_entry_safe (lock, n, &res->lr_granted, l_res_link) {
1630                 if (list_empty(&lock->l_lru) &&
1631                     !lock->l_readers && !lock->l_writers &&
1632                     !(lock->l_flags & LDLM_FL_LOCAL) &&
1633                     !(lock->l_flags & LDLM_FL_CBPENDING)) {
1634                         ldlm_lock_add_to_lru(lock);
1635                         lock->l_flags &= ~LDLM_FL_NO_LRU;
1636                         LDLM_DEBUG(lock, "join lock to lru");
1637                         count++;
1638                 }
1639         }
1640         goto unlock;
1641 split:
1642         spin_lock(&ns->ns_unused_lock);
1643         list_for_each_entry_safe (lock, n, &ns->ns_unused_list, l_lru) {
1644                 if (lock->l_resource == res) {
1645                         ldlm_lock_remove_from_lru_nolock(lock);
1646                         lock->l_flags |= LDLM_FL_NO_LRU;
1647                         LDLM_DEBUG(lock, "split lock from lru");
1648                         count++;
1649                 }
1650         }
1651         spin_unlock(&ns->ns_unused_lock);
1652 unlock:
1653         unlock_res(res);
1654         ldlm_resource_putref(res);
1655         RETURN(count);
1656 }
1657
1658 /* Lock iterators. */
1659
1660 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
1661                           void *closure)
1662 {
1663         struct list_head *tmp, *next;
1664         struct ldlm_lock *lock;
1665         int rc = LDLM_ITER_CONTINUE;
1666
1667         ENTRY;
1668
1669         if (!res)
1670                 RETURN(LDLM_ITER_CONTINUE);
1671
1672         lock_res(res);
1673         list_for_each_safe(tmp, next, &res->lr_granted) {
1674                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1675
1676                 if (iter(lock, closure) == LDLM_ITER_STOP)
1677                         GOTO(out, rc = LDLM_ITER_STOP);
1678         }
1679
1680         list_for_each_safe(tmp, next, &res->lr_converting) {
1681                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1682
1683                 if (iter(lock, closure) == LDLM_ITER_STOP)
1684                         GOTO(out, rc = LDLM_ITER_STOP);
1685         }
1686
1687         list_for_each_safe(tmp, next, &res->lr_waiting) {
1688                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1689
1690                 if (iter(lock, closure) == LDLM_ITER_STOP)
1691                         GOTO(out, rc = LDLM_ITER_STOP);
1692         }
1693  out:
1694         unlock_res(res);
1695         RETURN(rc);
1696 }
1697
1698 struct iter_helper_data {
1699         ldlm_iterator_t iter;
1700         void *closure;
1701 };
1702
1703 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
1704 {
1705         struct iter_helper_data *helper = closure;
1706         return helper->iter(lock, helper->closure);
1707 }
1708
1709 static int ldlm_res_iter_helper(struct ldlm_resource *res, void *closure)
1710 {
1711         return ldlm_resource_foreach(res, ldlm_iter_helper, closure);
1712 }
1713
1714 int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
1715                            void *closure)
1716 {
1717         struct iter_helper_data helper = { iter: iter, closure: closure };
1718         return ldlm_namespace_foreach_res(ns, ldlm_res_iter_helper, &helper);
1719 }
1720
1721 int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
1722                                ldlm_res_iterator_t iter, void *closure)
1723 {
1724         int i, rc = LDLM_ITER_CONTINUE;
1725         struct ldlm_resource *res;
1726         struct list_head *tmp;
1727
1728         ENTRY;
1729         spin_lock(&ns->ns_hash_lock);
1730         for (i = 0; i < RES_HASH_SIZE; i++) {
1731                 tmp = ns->ns_hash[i].next;
1732                 while (tmp != &(ns->ns_hash[i])) {
1733                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
1734                         ldlm_resource_getref(res);
1735                         spin_unlock(&ns->ns_hash_lock);
1736
1737                         rc = iter(res, closure);
1738
1739                         spin_lock(&ns->ns_hash_lock);
1740                         tmp = tmp->next;
1741                         ldlm_resource_putref_locked(res);
1742                         if (rc == LDLM_ITER_STOP)
1743                                 GOTO(out, rc);
1744                 }
1745         }
1746  out:
1747         spin_unlock(&ns->ns_hash_lock);
1748         RETURN(rc);
1749 }
1750
1751 /* non-blocking function to manipulate a lock whose cb_data is being put away.*/
1752 void ldlm_resource_iterate(struct ldlm_namespace *ns,
1753                            const struct ldlm_res_id *res_id,
1754                            ldlm_iterator_t iter, void *data)
1755 {
1756         struct ldlm_resource *res;
1757         ENTRY;
1758
1759         if (ns == NULL) {
1760                 CERROR("must pass in namespace\n");
1761                 LBUG();
1762         }
1763
1764         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1765         if (res == NULL) {
1766                 EXIT;
1767                 return;
1768         }
1769
1770         ldlm_resource_foreach(res, iter, data);
1771         ldlm_resource_putref(res);
1772         EXIT;
1773 }
1774
1775 /* Lock replay */
1776
1777 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
1778 {
1779         struct list_head *list = closure;
1780
1781         /* we use l_pending_chain here, because it's unused on clients. */
1782         LASSERTF(list_empty(&lock->l_pending_chain),"lock %p next %p prev %p\n",
1783                  lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
1784         /* bug 9573: don't replay locks left after eviction */
1785         if (!(lock->l_flags & LDLM_FL_FAILED))
1786                 list_add(&lock->l_pending_chain, list);
1787         return LDLM_ITER_CONTINUE;
1788 }
1789
1790 static int replay_lock_interpret(struct ptlrpc_request *req,
1791                                  struct ldlm_async_args *aa, int rc)
1792 {
1793         struct ldlm_lock  *lock;
1794         struct ldlm_reply *reply;
1795
1796         ENTRY;
1797         atomic_dec(&req->rq_import->imp_replay_inflight);
1798         if (rc != ELDLM_OK)
1799                 GOTO(out, rc);
1800
1801
1802         reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1803         if (reply == NULL)
1804                 GOTO(out, rc = -EPROTO);
1805
1806         lock = ldlm_handle2lock(&aa->lock_handle);
1807         if (!lock) {
1808                 CERROR("received replay ack for unknown local cookie "LPX64
1809                        " remote cookie "LPX64 " from server %s id %s\n",
1810                        aa->lock_handle.cookie, reply->lock_handle.cookie,
1811                        req->rq_export->exp_client_uuid.uuid,
1812                        libcfs_id2str(req->rq_peer));
1813                 GOTO(out, rc = -ESTALE);
1814         }
1815
1816         lock->l_remote_handle = reply->lock_handle;
1817         LDLM_DEBUG(lock, "replayed lock:");
1818         ptlrpc_import_recovery_state_machine(req->rq_import);
1819         LDLM_LOCK_PUT(lock);
1820 out:
1821         if (rc != ELDLM_OK)
1822                 ptlrpc_connect_import(req->rq_import, NULL);
1823
1824
1825         RETURN(rc);
1826 }
1827
1828 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
1829 {
1830         struct ptlrpc_request *req;
1831         struct ldlm_async_args *aa;
1832         struct ldlm_request   *body;
1833         int flags;
1834         ENTRY;
1835
1836
1837         /* Bug 11974: Do not replay a lock which is actively being canceled */
1838         if (lock->l_flags & LDLM_FL_CANCELING) {
1839                 LDLM_DEBUG(lock, "Not replaying canceled lock:");
1840                 RETURN(0);
1841         }
1842
1843         /* If this is reply-less callback lock, we cannot replay it, since
1844          * server might have long dropped it, but notification of that event was
1845          * lost by network. (and server granted conflicting lock already) */
1846         if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1847                 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
1848                 ldlm_lock_cancel(lock);
1849                 RETURN(0);
1850         }
1851         /*
1852          * If granted mode matches the requested mode, this lock is granted.
1853          *
1854          * If they differ, but we have a granted mode, then we were granted
1855          * one mode and now want another: ergo, converting.
1856          *
1857          * If we haven't been granted anything and are on a resource list,
1858          * then we're blocked/waiting.
1859          *
1860          * If we haven't been granted anything and we're NOT on a resource list,
1861          * then we haven't got a reply yet and don't have a known disposition.
1862          * This happens whenever a lock enqueue is the request that triggers
1863          * recovery.
1864          */
1865         if (lock->l_granted_mode == lock->l_req_mode)
1866                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
1867         else if (lock->l_granted_mode)
1868                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
1869         else if (!list_empty(&lock->l_res_link))
1870                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
1871         else
1872                 flags = LDLM_FL_REPLAY;
1873
1874         req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
1875                                         LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
1876         if (req == NULL)
1877                 RETURN(-ENOMEM);
1878
1879         /* We're part of recovery, so don't wait for it. */
1880         req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
1881
1882         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1883         ldlm_lock2desc(lock, &body->lock_desc);
1884         body->lock_flags = flags;
1885
1886         ldlm_lock2handle(lock, &body->lock_handle[0]);
1887         if (lock->l_lvb_len != 0) {
1888                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
1889                 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1890                                      lock->l_lvb_len);
1891         }
1892         ptlrpc_request_set_replen(req);
1893         /* notify the server we've replayed all requests.
1894          * also, we mark the request to be put on a dedicated
1895          * queue to be processed after all request replayes.
1896          * bug 6063 */
1897         lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
1898
1899         LDLM_DEBUG(lock, "replaying lock:");
1900
1901         atomic_inc(&req->rq_import->imp_replay_inflight);
1902         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1903         aa = (struct ldlm_async_args *)&req->rq_async_args;
1904         aa->lock_handle = body->lock_handle[0];
1905         req->rq_interpret_reply = replay_lock_interpret;
1906         ptlrpcd_add_req(req);
1907
1908         RETURN(0);
1909 }
1910
1911 int ldlm_replay_locks(struct obd_import *imp)
1912 {
1913         struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
1914         struct list_head list;
1915         struct ldlm_lock *lock, *next;
1916         int rc = 0;
1917
1918         ENTRY;
1919         CFS_INIT_LIST_HEAD(&list);
1920
1921         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1922
1923         /* ensure this doesn't fall to 0 before all have been queued */
1924         atomic_inc(&imp->imp_replay_inflight);
1925
1926         (void)ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
1927
1928         list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
1929                 list_del_init(&lock->l_pending_chain);
1930                 if (rc)
1931                         continue; /* or try to do the rest? */
1932                 rc = replay_one_lock(imp, lock);
1933         }
1934
1935         atomic_dec(&imp->imp_replay_inflight);
1936
1937         RETURN(rc);
1938 }