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 "CFS_TIME_T", "
  65                            CFS_DURATION_T"s ago); not entering recovery in "
  66                            "server code, just going back to sleep",
  67                            lock->l_enqueued_time.tv_sec,
  68                            cfs_time_current_sec() - lock->l_enqueued_time.tv_sec);
  69                 if (cfs_time_after(cfs_time_current(), next_dump)) {
  70                         last_dump = next_dump;
  71                         next_dump = cfs_time_shift(300);
  72                         ldlm_namespace_dump(D_DLMTRACE,
  73                                             lock->l_resource->lr_namespace);
  74                         if (last_dump == 0)
  75                                 libcfs_debug_dumplog();
  76                 }
  77                 RETURN(0);
  78         }
  79
  80         obd = lock->l_conn_export->exp_obd;
  81         imp = obd->u.cli.cl_import;
  82         ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
  83         LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", "
  84                   CFS_DURATION_T"s ago), entering recovery for %s@%s",
  85                   lock->l_enqueued_time.tv_sec,
  86                   cfs_time_current_sec() - lock->l_enqueued_time.tv_sec,
  87                   obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
  88
  89         RETURN(0);
  90 }
  91
  92 static int is_granted_or_cancelled(struct ldlm_lock *lock)
  93 {
  94         int ret = 0;
  95
  96         lock_res_and_lock(lock);
  97         if (((lock->l_req_mode == lock->l_granted_mode) &&
  98              !(lock->l_flags & LDLM_FL_CP_REQD)) ||
  99             (lock->l_flags & LDLM_FL_FAILED))
 100                 ret = 1;
 101         unlock_res_and_lock(lock);
 102
 103         return ret;
 104 }
 105
 106 int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data)
 107 {
 108         /* XXX ALLOCATE - 160 bytes */
 109         struct lock_wait_data lwd;
 110         struct obd_device *obd;
 111         struct obd_import *imp = NULL;
 112         struct l_wait_info lwi;
 113         int rc = 0;
 114         ENTRY;
 115
 116         if (flags == LDLM_FL_WAIT_NOREPROC) {
 117                 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
 118                 goto noreproc;
 119         }
 120
 121         if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
 122                        LDLM_FL_BLOCK_CONV))) {
 123                 cfs_waitq_signal(&lock->l_waitq);
 124                 RETURN(0);
 125         }
 126
 127         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
 128                    "sleeping");
 129         ldlm_lock_dump(D_OTHER, lock, 0);
 130         ldlm_reprocess_all(lock->l_resource);
 131
 132 noreproc:
 133
 134         obd = class_exp2obd(lock->l_conn_export);
 135
 136         /* if this is a local lock, then there is no import */
 137         if (obd != NULL)
 138                 imp = obd->u.cli.cl_import;
 139
 140         lwd.lwd_lock = lock;
 141
 142         if (lock->l_flags & LDLM_FL_NO_TIMEOUT) {
 143                 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
 144                 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
 145         } else {
 146                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
 147                                        ldlm_expired_completion_wait,
 148                                        interrupted_completion_wait, &lwd);
 149         }
 150
 151         if (imp != NULL) {
 152                 spin_lock(&imp->imp_lock);
 153                 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
 154                 spin_unlock(&imp->imp_lock);
 155         }
 156
 157         /* Go to sleep until the lock is granted or cancelled. */
 158         rc = l_wait_event(lock->l_waitq, is_granted_or_cancelled(lock), &lwi);
 159
 160         if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED) {
 161                 LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed");
 162                 RETURN(-EIO);
 163         }
 164
 165         if (rc) {
 166                 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
 167                            rc);
 168                 RETURN(rc);
 169         }
 170
 171         LDLM_DEBUG(lock, "client-side enqueue waking up: granted");
 172         RETURN(0);
 173 }
 174
 175 /*
 176  * ->l_blocking_ast() callback for LDLM locks acquired by server-side OBDs.
 177  */
 178 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 179                       void *data, int flag)
 180 {
 181         int do_ast;
 182         ENTRY;
 183
 184         if (flag == LDLM_CB_CANCELING) {
 185                 /* Don't need to do anything here. */
 186                 RETURN(0);
 187         }
 188
 189         lock_res_and_lock(lock);
 190         /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
 191          * that ldlm_blocking_ast is called just before intent_policy method
 192          * takes the ns_lock, then by the time we get the lock, we might not
 193          * be the correct blocking function anymore.  So check, and return
 194          * early, if so. */
 195         if (lock->l_blocking_ast != ldlm_blocking_ast) {
 196                 unlock_res_and_lock(lock);
 197                 RETURN(0);
 198         }
 199
 200         lock->l_flags |= LDLM_FL_CBPENDING;
 201         do_ast = (!lock->l_readers && !lock->l_writers);
 202         unlock_res_and_lock(lock);
 203
 204         if (do_ast) {
 205                 struct lustre_handle lockh;
 206                 int rc;
 207
 208                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
 209                 ldlm_lock2handle(lock, &lockh);
 210                 rc = ldlm_cli_cancel(&lockh);
 211                 if (rc < 0)
 212                         CERROR("ldlm_cli_cancel: %d\n", rc);
 213         } else {
 214                 LDLM_DEBUG(lock, "Lock still has references, will be "
 215                            "cancelled later");
 216         }
 217         RETURN(0);
 218 }
 219
 220 /*
 221  * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See
 222  * comment in filter_intent_policy() on why you may need this.
 223  */
 224 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
 225 {
 226         /*
 227          * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for
 228          * that is rather subtle: with OST-side locking, it may so happen that
 229          * _all_ extent locks are held by the OST. If client wants to obtain
 230          * current file size it calls ll{,u}_glimpse_size(), and (as locks are
 231          * on the server), dummy glimpse callback fires and does
 232          * nothing. Client still receives correct file size due to the
 233          * following fragment in filter_intent_policy():
 234          *
 235          * rc = l->l_glimpse_ast(l, NULL); // this will update the LVB
 236          * if (rc != 0 && res->lr_namespace->ns_lvbo &&
 237          *     res->lr_namespace->ns_lvbo->lvbo_update) {
 238          *         res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1);
 239          * }
 240          *
 241          * that is, after glimpse_ast() fails, filter_lvbo_update() runs, and
 242          * returns correct file size to the client.
 243          */
 244         return -ELDLM_NO_LOCK_DATA;
 245 }
 246
 247 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
 248                            const struct ldlm_res_id *res_id,
 249                            ldlm_type_t type, ldlm_policy_data_t *policy,
 250                            ldlm_mode_t mode, int *flags,
 251                            ldlm_blocking_callback blocking,
 252                            ldlm_completion_callback completion,
 253                            ldlm_glimpse_callback glimpse,
 254                            void *data, __u32 lvb_len, void *lvb_swabber,
 255                            struct lustre_handle *lockh)
 256 {
 257         struct ldlm_lock *lock;
 258         int err;
 259         ENTRY;
 260
 261         LASSERT(!(*flags & LDLM_FL_REPLAY));
 262         if (unlikely(ns_is_client(ns))) {
 263                 CERROR("Trying to enqueue local lock in a shadow namespace\n");
 264                 LBUG();
 265         }
 266
 267         lock = ldlm_lock_create(ns, res_id, type, mode, blocking,
 268                                 completion, glimpse, data, lvb_len);
 269         if (unlikely(!lock))
 270                 GOTO(out_nolock, err = -ENOMEM);
 271         LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
 272
 273         ldlm_lock_addref_internal(lock, mode);
 274         ldlm_lock2handle(lock, lockh);
 275         lock_res_and_lock(lock);
 276         lock->l_flags |= LDLM_FL_LOCAL;
 277         if (*flags & LDLM_FL_ATOMIC_CB)
 278                 lock->l_flags |= LDLM_FL_ATOMIC_CB;
 279         lock->l_lvb_swabber = lvb_swabber;
 280         unlock_res_and_lock(lock);
 281         if (policy != NULL)
 282                 lock->l_policy_data = *policy;
 283         if (type == LDLM_EXTENT)
 284                 lock->l_req_extent = policy->l_extent;
 285
 286         err = ldlm_lock_enqueue(ns, &lock, policy, flags);
 287         if (unlikely(err != ELDLM_OK))
 288                 GOTO(out, err);
 289
 290         if (policy != NULL)
 291                 *policy = lock->l_policy_data;
 292
 293         LDLM_DEBUG_NOLOCK("client-side local enqueue handler END (lock %p)",
 294                           lock);
 295
 296         if (lock->l_completion_ast)
 297                 lock->l_completion_ast(lock, *flags, NULL);
 298
 299         LDLM_DEBUG(lock, "client-side local enqueue END");
 300         EXIT;
 301  out:
 302         LDLM_LOCK_PUT(lock);
 303  out_nolock:
 304         return err;
 305 }
 306
 307 static void failed_lock_cleanup(struct ldlm_namespace *ns,
 308                                 struct ldlm_lock *lock,
 309                                 struct lustre_handle *lockh, int mode)
 310 {
 311         /* Set a flag to prevent us from sending a CANCEL (bug 407) */
 312         lock_res_and_lock(lock);
 313         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 314         unlock_res_and_lock(lock);
 315         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 316
 317         ldlm_lock_decref_and_cancel(lockh, mode);
 318
 319         /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
 320          *       from llite/file.c/ll_file_flock(). */
 321         if (lock->l_resource->lr_type == LDLM_FLOCK) {
 322                 ldlm_lock_destroy(lock);
 323         }
 324 }
 325
 326 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
 327                           ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
 328                           int *flags, void *lvb, __u32 lvb_len,
 329                           void *lvb_swabber, struct lustre_handle *lockh,int rc)
 330 {
 331         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 332         int is_replay = *flags & LDLM_FL_REPLAY;
 333         struct ldlm_lock *lock;
 334         struct ldlm_reply *reply;
 335         int cleanup_phase = 1;
 336         ENTRY;
 337
 338         lock = ldlm_handle2lock(lockh);
 339         /* ldlm_cli_enqueue is holding a reference on this lock. */
 340         if (!lock) {
 341                 LASSERT(type == LDLM_FLOCK);
 342                 RETURN(-ENOLCK);
 343         }
 344
 345         if (rc != ELDLM_OK) {
 346                 LASSERT(!is_replay);
 347                 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
 348                            rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
 349                 if (rc == ELDLM_LOCK_ABORTED) {
 350                         /* Before we return, swab the reply */
 351                         reply = req_capsule_server_get(&req->rq_pill,
 352                                                        &RMF_DLM_REP);
 353                         if (reply == NULL)
 354                                 rc = -EPROTO;
 355                         if (lvb_len) {
 356                                 struct ost_lvb *tmplvb;
 357
 358                                 req_capsule_set_size(&req->rq_pill,
 359                                                      &RMF_DLM_LVB, RCL_SERVER,
 360                                                      lvb_len);
 361                             tmplvb = req_capsule_server_swab_get(&req->rq_pill,
 362                                                                  &RMF_DLM_LVB,
 363                                                                  lvb_swabber);
 364                                 if (tmplvb == NULL)
 365                                         GOTO(cleanup, rc = -EPROTO);
 366                                 if (lvb != NULL)
 367                                         memcpy(lvb, tmplvb, lvb_len);
 368                         }
 369                 }
 370                 GOTO(cleanup, rc);
 371         }
 372
 373         reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
 374         if (reply == NULL)
 375                 GOTO(cleanup, rc = -EPROTO);
 376
 377         /* lock enqueued on the server */
 378         cleanup_phase = 0;
 379
 380         lock_res_and_lock(lock);
 381         lock->l_remote_handle = reply->lock_handle;
 382         *flags = reply->lock_flags;
 383         lock->l_flags |= reply->lock_flags & LDLM_INHERIT_FLAGS;
 384         /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match()
 385          * to wait with no timeout as well */
 386         lock->l_flags |= reply->lock_flags & LDLM_FL_NO_TIMEOUT;
 387         unlock_res_and_lock(lock);
 388
 389         CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%x\n",
 390                lock, reply->lock_handle.cookie, *flags);
 391
 392         /* If enqueue returned a blocked lock but the completion handler has
 393          * already run, then it fixed up the resource and we don't need to do it
 394          * again. */
 395         if ((*flags) & LDLM_FL_LOCK_CHANGED) {
 396                 int newmode = reply->lock_desc.l_req_mode;
 397                 LASSERT(!is_replay);
 398                 if (newmode && newmode != lock->l_req_mode) {
 399                         LDLM_DEBUG(lock, "server returned different mode %s",
 400                                    ldlm_lockname[newmode]);
 401                         lock->l_req_mode = newmode;
 402                 }
 403
 404                 if (memcmp(reply->lock_desc.l_resource.lr_name.name,
 405                           lock->l_resource->lr_name.name,
 406                           sizeof(struct ldlm_res_id))) {
 407                         CDEBUG(D_INFO, "remote intent success, locking "
 408                                         "(%ld,%ld,%ld) instead of "
 409                                         "(%ld,%ld,%ld)\n",
 410                               (long)reply->lock_desc.l_resource.lr_name.name[0],
 411                               (long)reply->lock_desc.l_resource.lr_name.name[1],
 412                               (long)reply->lock_desc.l_resource.lr_name.name[2],
 413                               (long)lock->l_resource->lr_name.name[0],
 414                               (long)lock->l_resource->lr_name.name[1],
 415                               (long)lock->l_resource->lr_name.name[2]);
 416
 417                         rc = ldlm_lock_change_resource(ns, lock,
 418                                         &reply->lock_desc.l_resource.lr_name);
 419                         if (rc || lock->l_resource == NULL)
 420                                 GOTO(cleanup, rc = -ENOMEM);
 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(int req_size, int off)
 490 {
 491         int avail;
 492
 493         avail = min_t(int, LDLM_MAXREQSIZE, CFS_PAGE_SIZE - 512) - req_size;
 494         avail /= sizeof(struct lustre_handle);
 495         avail += LDLM_LOCKREQ_HANDLES - off;
 496
 497         return avail;
 498 }
 499
 500 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
 501                                              enum req_location loc,
 502                                              int off)
 503 {
 504         int size = req_capsule_msg_size(pill, loc);
 505         return ldlm_req_handles_avail(size, off);
 506 }
 507
 508 static inline int ldlm_format_handles_avail(struct obd_import *imp,
 509                                             const struct req_format *fmt,
 510                                             enum req_location loc, int off)
 511 {
 512         int size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
 513         return ldlm_req_handles_avail(size, off);
 514 }
 515
 516 /* Cancel lru locks and pack them into the enqueue request. Pack there the given
 517  * @count locks in @cancels. */
 518 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
 519                       int version, int opc, int canceloff,
 520                       struct list_head *cancels, int count)
 521 {
 522         struct ldlm_namespace   *ns = exp->exp_obd->obd_namespace;
 523         struct req_capsule      *pill = &req->rq_pill;
 524         struct ldlm_request     *dlm = NULL;
 525         int flags, avail, to_free, bufcount, pack = 0;
 526         CFS_LIST_HEAD(head);
 527         int rc;
 528         ENTRY;
 529
 530         if (cancels == NULL)
 531                 cancels = &head;
 532         if (exp_connect_cancelset(exp)) {
 533                 /* Estimate the amount of available space in the request. */
 534                 bufcount = req_capsule_filled_sizes(pill, RCL_CLIENT);
 535                 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
 536
 537                 flags = ns_connect_lru_resize(ns) ?
 538                         LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
 539                 to_free = !ns_connect_lru_resize(ns) &&
 540                           opc == LDLM_ENQUEUE ? 1 : 0;
 541
 542                 /* Cancel lru locks here _only_ if the server supports
 543                  * EARLY_CANCEL. Otherwise we have to send extra CANCEL
 544                  * rpc, what will make us slower. */
 545                 if (avail > count)
 546                         count += ldlm_cancel_lru_local(ns, cancels, to_free,
 547                                                        avail - count, 0, flags);
 548                 if (avail > count)
 549                         pack = count;
 550                 else
 551                         pack = avail;
 552                 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
 553                                      ldlm_request_bufsize(pack, opc));
 554         }
 555
 556         rc = ptlrpc_request_pack(req, version, opc);
 557         if (rc) {
 558                 ldlm_lock_list_put(cancels, l_bl_ast, count);
 559                 RETURN(rc);
 560         }
 561
 562         if (exp_connect_cancelset(exp)) {
 563                 if (canceloff) {
 564                         dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
 565                         LASSERT(dlm);
 566                         /* Skip first lock handler in ldlm_request_pack(),
 567                          * this method will incrment @lock_count according
 568                          * to the lock handle amount actually written to
 569                          * the buffer. */
 570                         dlm->lock_count = canceloff;
 571                 }
 572                 /* Pack into the request @pack lock handles. */
 573                 ldlm_cli_cancel_list(cancels, pack, req, 0);
 574                 /* Prepare and send separate cancel rpc for others. */
 575                 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
 576         } else {
 577                 ldlm_lock_list_put(cancels, l_bl_ast, count);
 578         }
 579         RETURN(0);
 580 }
 581
 582 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
 583                           struct list_head *cancels, int count)
 584 {
 585         return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
 586                                  LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
 587 }
 588
 589 /* If a request has some specific initialisation it is passed in @reqp,
 590  * otherwise it is created in ldlm_cli_enqueue.
 591  *
 592  * Supports sync and async requests, pass @async flag accordingly. If a
 593  * request was created in ldlm_cli_enqueue and it is the async request,
 594  * pass it to the caller in @reqp. */
 595 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
 596                      struct ldlm_enqueue_info *einfo,
 597                      const struct ldlm_res_id *res_id,
 598                      ldlm_policy_data_t *policy, int *flags,
 599                      void *lvb, __u32 lvb_len, void *lvb_swabber,
 600                      struct lustre_handle *lockh, int async)
 601 {
 602         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 603         struct ldlm_lock      *lock;
 604         struct ldlm_request   *body;
 605         int                    is_replay = *flags & LDLM_FL_REPLAY;
 606         int                    req_passed_in = 1;
 607         int                    rc, err;
 608         struct ptlrpc_request *req;
 609         ENTRY;
 610
 611         LASSERT(exp != NULL);
 612
 613         /* If we're replaying this lock, just check some invariants.
 614          * If we're creating a new lock, get everything all setup nice. */
 615         if (is_replay) {
 616                 lock = ldlm_handle2lock(lockh);
 617                 LASSERT(lock != NULL);
 618                 LDLM_DEBUG(lock, "client-side enqueue START");
 619                 LASSERT(exp == lock->l_conn_export);
 620         } else {
 621                 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
 622                                         einfo->ei_mode, einfo->ei_cb_bl,
 623                                         einfo->ei_cb_cp, einfo->ei_cb_gl,
 624                                         einfo->ei_cbdata, lvb_len);
 625                 if (lock == NULL)
 626                         RETURN(-ENOMEM);
 627                 /* for the local lock, add the reference */
 628                 ldlm_lock_addref_internal(lock, einfo->ei_mode);
 629                 ldlm_lock2handle(lock, lockh);
 630                 lock->l_lvb_swabber = lvb_swabber;
 631                 if (policy != NULL) {
 632                         /* INODEBITS_INTEROP: If the server does not support
 633                          * inodebits, we will request a plain lock in the
 634                          * descriptor (ldlm_lock2desc() below) but use an
 635                          * inodebits lock internally with both bits set.
 636                          */
 637                         if (einfo->ei_type == LDLM_IBITS &&
 638                             !(exp->exp_connect_flags & OBD_CONNECT_IBITS))
 639                                 lock->l_policy_data.l_inodebits.bits =
 640                                         MDS_INODELOCK_LOOKUP |
 641                                         MDS_INODELOCK_UPDATE;
 642                         else
 643                                 lock->l_policy_data = *policy;
 644                 }
 645
 646                 if (einfo->ei_type == LDLM_EXTENT)
 647                         lock->l_req_extent = policy->l_extent;
 648                 LDLM_DEBUG(lock, "client-side enqueue START");
 649         }
 650
 651         /* lock not sent to server yet */
 652
 653         if (reqp == NULL || *reqp == NULL) {
 654                 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
 655                                                 &RQF_LDLM_ENQUEUE,
 656                                                 LUSTRE_DLM_VERSION,
 657                                                 LDLM_ENQUEUE);
 658                 if (req == NULL) {
 659                         failed_lock_cleanup(ns, lock, lockh, einfo->ei_mode);
 660                         LDLM_LOCK_PUT(lock);
 661                         RETURN(-ENOMEM);
 662                 }
 663                 req_passed_in = 0;
 664                 if (reqp)
 665                         *reqp = req;
 666         } else {
 667                 int len;
 668
 669                 req = *reqp;
 670                 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
 671                                            RCL_CLIENT);
 672                 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
 673                          DLM_LOCKREQ_OFF, len, sizeof(*body));
 674         }
 675
 676         lock->l_conn_export = exp;
 677         lock->l_export = NULL;
 678         lock->l_blocking_ast = einfo->ei_cb_bl;
 679
 680         /* Dump lock data into the request buffer */
 681         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
 682         ldlm_lock2desc(lock, &body->lock_desc);
 683         body->lock_flags = *flags;
 684         body->lock_handle[0] = *lockh;
 685
 686         /* Continue as normal. */
 687         if (!req_passed_in) {
 688                 if (lvb_len > 0) {
 689                         req_capsule_extend(&req->rq_pill,
 690                                            &RQF_LDLM_ENQUEUE_LVB);
 691                         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
 692                                              RCL_SERVER, lvb_len);
 693                 }
 694                 ptlrpc_request_set_replen(req);
 695         }
 696
 697         /*
 698          * Liblustre client doesn't get extent locks, except for O_APPEND case
 699          * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
 700          * [i_size, OBD_OBJECT_EOF] lock is taken.
 701          */
 702         LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT ||
 703                      policy->l_extent.end == OBD_OBJECT_EOF));
 704
 705         if (async) {
 706                 LASSERT(reqp != NULL);
 707                 RETURN(0);
 708         }
 709
 710         LDLM_DEBUG(lock, "sending request");
 711         rc = ptlrpc_queue_wait(req);
 712         err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
 713                                     einfo->ei_mode, flags, lvb, lvb_len,
 714                                     lvb_swabber, lockh, rc);
 715
 716         /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
 717          * one reference that we took */
 718         if (err == -ENOLCK)
 719                 LDLM_LOCK_PUT(lock);
 720         else
 721                 rc = err;
 722
 723         if (!req_passed_in && req != NULL) {
 724                 ptlrpc_req_finished(req);
 725                 if (reqp)
 726                         *reqp = NULL;
 727         }
 728
 729         RETURN(rc);
 730 }
 731
 732 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
 733                                   __u32 *flags)
 734 {
 735         struct ldlm_resource *res;
 736         int rc;
 737         ENTRY;
 738         if (ns_is_client(lock->l_resource->lr_namespace)) {
 739                 CERROR("Trying to cancel local lock\n");
 740                 LBUG();
 741         }
 742         LDLM_DEBUG(lock, "client-side local convert");
 743
 744         res = ldlm_lock_convert(lock, new_mode, flags);
 745         if (res) {
 746                 ldlm_reprocess_all(res);
 747                 rc = 0;
 748         } else {
 749                 rc = EDEADLOCK;
 750         }
 751         LDLM_DEBUG(lock, "client-side local convert handler END");
 752         LDLM_LOCK_PUT(lock);
 753         RETURN(rc);
 754 }
 755
 756 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
 757  * conversion of locks which are on the waiting or converting queue */
 758 /* Caller of this code is supposed to take care of lock readers/writers
 759    accounting */
 760 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags)
 761 {
 762         struct ldlm_request   *body;
 763         struct ldlm_reply     *reply;
 764         struct ldlm_lock      *lock;
 765         struct ldlm_resource  *res;
 766         struct ptlrpc_request *req;
 767         int                    rc;
 768         ENTRY;
 769
 770         lock = ldlm_handle2lock(lockh);
 771         if (!lock) {
 772                 LBUG();
 773                 RETURN(-EINVAL);
 774         }
 775         *flags = 0;
 776
 777         if (lock->l_conn_export == NULL)
 778                 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
 779
 780         LDLM_DEBUG(lock, "client-side convert");
 781
 782         req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export),
 783                                         &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
 784                                         LDLM_CONVERT);
 785         if (req == NULL) {
 786                 LDLM_LOCK_PUT(lock);
 787                 RETURN(-ENOMEM);
 788         }
 789
 790         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
 791         body->lock_handle[0] = lock->l_remote_handle;
 792
 793         body->lock_desc.l_req_mode = new_mode;
 794         body->lock_flags = *flags;
 795
 796
 797         ptlrpc_request_set_replen(req);
 798         rc = ptlrpc_queue_wait(req);
 799         if (rc != ELDLM_OK)
 800                 GOTO(out, rc);
 801
 802         reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
 803         if (reply == NULL)
 804                 GOTO(out, rc = -EPROTO);
 805
 806         if (req->rq_status)
 807                 GOTO(out, rc = req->rq_status);
 808
 809         res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
 810         if (res != NULL) {
 811                 ldlm_reprocess_all(res);
 812                 /* Go to sleep until the lock is granted. */
 813                 /* FIXME: or cancelled. */
 814                 if (lock->l_completion_ast) {
 815                         rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
 816                                                     NULL);
 817                         if (rc)
 818                                 GOTO(out, rc);
 819                 }
 820         } else {
 821                 rc = EDEADLOCK;
 822         }
 823         EXIT;
 824  out:
 825         LDLM_LOCK_PUT(lock);
 826         ptlrpc_req_finished(req);
 827         return rc;
 828 }
 829
 830 /* Cancel locks locally.
 831  * Returns:
 832  * LDLM_FL_LOCAL_ONLY if tere is no need in a CANCEL rpc to the server;
 833  * LDLM_FL_CANCELING otherwise;
 834  * LDLM_FL_BL_AST if there is a need in a separate CANCEL rpc. */
 835 static int ldlm_cli_cancel_local(struct ldlm_lock *lock)
 836 {
 837         int rc = LDLM_FL_LOCAL_ONLY;
 838         ENTRY;
 839
 840         if (lock->l_conn_export) {
 841                 int local_only;
 842
 843                 LDLM_DEBUG(lock, "client-side cancel");
 844                 /* Set this flag to prevent others from getting new references*/
 845                 lock_res_and_lock(lock);
 846                 lock->l_flags |= LDLM_FL_CBPENDING;
 847                 local_only = (lock->l_flags &
 848                               (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
 849                 ldlm_cancel_callback(lock);
 850                 rc = (lock->l_flags & LDLM_FL_BL_AST) ?
 851                         LDLM_FL_BL_AST : LDLM_FL_CANCELING;
 852                 unlock_res_and_lock(lock);
 853
 854                 if (local_only) {
 855                         CDEBUG(D_DLMTRACE, "not sending request (at caller's "
 856                                "instruction)\n");
 857                         rc = LDLM_FL_LOCAL_ONLY;
 858                 }
 859                 ldlm_lock_cancel(lock);
 860         } else {
 861                 if (ns_is_client(lock->l_resource->lr_namespace)) {
 862                         LDLM_ERROR(lock, "Trying to cancel local lock");
 863                         LBUG();
 864                 }
 865                 LDLM_DEBUG(lock, "server-side local cancel");
 866                 ldlm_lock_cancel(lock);
 867                 ldlm_reprocess_all(lock->l_resource);
 868                 LDLM_DEBUG(lock, "server-side local cancel handler END");
 869         }
 870
 871         RETURN(rc);
 872 }
 873
 874 /* Pack @count locks in @head into ldlm_request buffer at the offset @off,
 875    of the request @req. */
 876 static void ldlm_cancel_pack(struct ptlrpc_request *req,
 877                              struct list_head *head, int count)
 878 {
 879         struct ldlm_request *dlm;
 880         struct ldlm_lock *lock;
 881         int max, packed = 0;
 882         ENTRY;
 883
 884         dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
 885         LASSERT(dlm != NULL);
 886
 887         /* Check the room in the request buffer. */
 888         max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
 889                 sizeof(struct ldlm_request);
 890         max /= sizeof(struct lustre_handle);
 891         max += LDLM_LOCKREQ_HANDLES;
 892         LASSERT(max >= dlm->lock_count + count);
 893
 894         /* XXX: it would be better to pack lock handles grouped by resource.
 895          * so that the server cancel would call filter_lvbo_update() less
 896          * frequently. */
 897         list_for_each_entry(lock, head, l_bl_ast) {
 898                 if (!count--)
 899                         break;
 900                 LASSERT(lock->l_conn_export);
 901                 /* Pack the lock handle to the given request buffer. */
 902                 LDLM_DEBUG(lock, "packing");
 903                 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
 904                 packed++;
 905         }
 906         CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
 907         EXIT;
 908 }
 909
 910 /* Prepare and send a batched cancel rpc, it will include count lock handles
 911  * of locks given in @head. */
 912 int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
 913                         int count, int flags)
 914 {
 915         struct ptlrpc_request *req = NULL;
 916         struct obd_import *imp;
 917         int free, sent = 0;
 918         int rc = 0;
 919         ENTRY;
 920
 921         LASSERT(exp != NULL);
 922         LASSERT(count > 0);
 923
 924         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
 925                 RETURN(count);
 926
 927         free = ldlm_format_handles_avail(class_exp2cliimp(exp),
 928                                          &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
 929         if (count > free)
 930                 count = free;
 931
 932         while (1) {
 933                 int bufcount;
 934
 935                 imp = class_exp2cliimp(exp);
 936                 if (imp == NULL || imp->imp_invalid) {
 937                         CDEBUG(D_DLMTRACE,
 938                                "skipping cancel on invalid import %p\n", imp);
 939                         RETURN(count);
 940                 }
 941
 942                 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
 943                 if (req == NULL)
 944                         GOTO(out, rc = -ENOMEM);
 945
 946                 bufcount = req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
 947                 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
 948                                      ldlm_request_bufsize(count, LDLM_CANCEL));
 949
 950                 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
 951                 if (rc) {
 952                         ptlrpc_request_free(req);
 953                         GOTO(out, rc);
 954                 }
 955                 req->rq_no_resend = 1;
 956                 req->rq_no_delay = 1;
 957
 958                 /* XXX FIXME bug 249 */
 959                 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
 960                 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
 961
 962                 ldlm_cancel_pack(req, cancels, count);
 963
 964                 ptlrpc_request_set_replen(req);
 965                 if (flags & LDLM_FL_ASYNC) {
 966                         ptlrpcd_add_req(req);
 967                         sent = count;
 968                         GOTO(out, 0);
 969                 } else {
 970                         rc = ptlrpc_queue_wait(req);
 971                 }
 972                 if (rc == ESTALE) {
 973                         CDEBUG(D_DLMTRACE, "client/server (nid %s) "
 974                                "out of sync -- not fatal\n",
 975                                libcfs_nid2str(req->rq_import->
 976                                               imp_connection->c_peer.nid));
 977                         rc = 0;
 978                 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
 979                            req->rq_import_generation == imp->imp_generation) {
 980                         ptlrpc_req_finished(req);
 981                         continue;
 982                 } else if (rc != ELDLM_OK) {
 983                         CERROR("Got rc %d from cancel RPC: canceling "
 984                                "anyway\n", rc);
 985                         break;
 986                 }
 987                 sent = count;
 988                 break;
 989         }
 990
 991         ptlrpc_req_finished(req);
 992         EXIT;
 993 out:
 994         return sent ? sent : rc;
 995 }
 996
 997 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
 998 {
 999         LASSERT(imp != NULL);
1000         return &imp->imp_obd->obd_namespace->ns_pool;
1001 }
1002
1003 /**
1004  * Update client's obd pool related fields with new SLV and Limit from \a req.
1005  */
1006 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1007 {
1008         struct obd_device *obd;
1009         __u64 old_slv, new_slv;
1010         __u32 new_limit;
1011         ENTRY;
1012
1013         if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1014                      !imp_connect_lru_resize(req->rq_import)))
1015         {
1016                 /*
1017                  * Do nothing for corner cases.
1018                  */
1019                 RETURN(0);
1020         }
1021
1022         /*
1023          * In some cases RPC may contain slv and limit zeroed out. This is
1024          * the case when server does not support lru resize feature. This is
1025          * also possible in some recovery cases when server side reqs have no
1026          * ref to obd export and thus access to server side namespace is no
1027          * possible.
1028          */
1029         if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1030             lustre_msg_get_limit(req->rq_repmsg) == 0) {
1031                 DEBUG_REQ(D_HA, req, "Zero SLV or Limit found "
1032                           "(SLV: "LPU64", Limit: %u)",
1033                           lustre_msg_get_slv(req->rq_repmsg),
1034                           lustre_msg_get_limit(req->rq_repmsg));
1035                 RETURN(0);
1036         }
1037
1038         new_limit = lustre_msg_get_limit(req->rq_repmsg);
1039         new_slv = lustre_msg_get_slv(req->rq_repmsg);
1040         obd = req->rq_import->imp_obd;
1041
1042         /*
1043          * Set new SLV and Limit to obd fields to make accessible for pool
1044          * thread. We do not access obd_namespace and pool directly here
1045          * as there is no reliable way to make sure that they are still
1046          * alive in cleanup time. Evil races are possible which may cause
1047          * oops in that time.
1048          */
1049         write_lock(&obd->obd_pool_lock);
1050         old_slv = obd->obd_pool_slv;
1051         obd->obd_pool_slv = new_slv;
1052         obd->obd_pool_limit = new_limit;
1053         write_unlock(&obd->obd_pool_lock);
1054
1055         /*
1056          * Check if we need to wakeup pools thread for fast SLV change.
1057          * This is only done when threads period is noticably long like
1058          * 10s or more.
1059          */
1060 #if defined(__KERNEL__) && (LDLM_POOLS_THREAD_PERIOD >= 10)
1061         if (old_slv > 0) {
1062                 __u64 fast_change = old_slv * LDLM_POOLS_FAST_SLV_CHANGE;
1063                 do_div(fast_change, 100);
1064
1065                 /*
1066                  * Wake up pools thread only if SLV has changed more than
1067                  * 50% since last update. In this case we want to react asap.
1068                  * Otherwise it is no sense to wake up pools as they are
1069                  * re-calculated every LDLM_POOLS_THREAD_PERIOD anyways.
1070                  */
1071                 if (old_slv > new_slv && old_slv - new_slv > fast_change)
1072                         ldlm_pools_wakeup();
1073         }
1074 #endif
1075         RETURN(0);
1076 }
1077 EXPORT_SYMBOL(ldlm_cli_update_pool);
1078
1079 int ldlm_cli_cancel(struct lustre_handle *lockh)
1080 {
1081         struct obd_export *exp;
1082         int avail, flags, count = 1, rc = 0;
1083         struct ldlm_namespace *ns;
1084         struct ldlm_lock *lock;
1085         CFS_LIST_HEAD(cancels);
1086         ENTRY;
1087
1088         /* concurrent cancels on the same handle can happen */
1089         lock = __ldlm_handle2lock(lockh, LDLM_FL_CANCELING);
1090         if (lock == NULL) {
1091                 LDLM_DEBUG_NOLOCK("lock is already being destroyed\n");
1092                 RETURN(0);
1093         }
1094
1095         rc = ldlm_cli_cancel_local(lock);
1096         if (rc < 0 || rc == LDLM_FL_LOCAL_ONLY) {
1097                 LDLM_LOCK_PUT(lock);
1098                 RETURN(rc < 0 ? rc : 0);
1099         }
1100         /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1101          * rpc which goes to canceld portal, so we can cancel other lru locks
1102          * here and send them all as one LDLM_CANCEL rpc. */
1103         LASSERT(list_empty(&lock->l_bl_ast));
1104         list_add(&lock->l_bl_ast, &cancels);
1105
1106         exp = lock->l_conn_export;
1107         if (exp_connect_cancelset(exp)) {
1108                 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1109                                                   &RQF_LDLM_CANCEL,
1110                                                   RCL_CLIENT, 0);
1111                 LASSERT(avail > 0);
1112
1113                 ns = lock->l_resource->lr_namespace;
1114                 flags = ns_connect_lru_resize(ns) ?
1115                         LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
1116                 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1117                                                LDLM_FL_BL_AST, flags);
1118         }
1119         ldlm_cli_cancel_list(&cancels, count, NULL, 0);
1120         RETURN(0);
1121 }
1122
1123 /* XXX until we will have compound requests and can cut cancels from generic rpc
1124  * we need send cancels with LDLM_FL_BL_AST flag as separate rpc */
1125 static int ldlm_cancel_list(struct list_head *cancels, int count, int flags)
1126 {
1127         CFS_LIST_HEAD(head);
1128         struct ldlm_lock *lock, *next;
1129         int left = 0, bl_ast = 0, rc;
1130
1131         left = count;
1132         list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1133                 if (left-- == 0)
1134                         break;
1135
1136                 if (flags & LDLM_FL_LOCAL_ONLY) {
1137                         rc = LDLM_FL_LOCAL_ONLY;
1138                         ldlm_lock_cancel(lock);
1139                 } else {
1140                         rc = ldlm_cli_cancel_local(lock);
1141                 }
1142                 if (!(flags & LDLM_FL_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1143                         LDLM_DEBUG(lock, "Cancel lock separately");
1144                         list_del_init(&lock->l_bl_ast);
1145                         list_add(&lock->l_bl_ast, &head);
1146                         bl_ast ++;
1147                         continue;
1148                 }
1149                 if (rc == LDLM_FL_LOCAL_ONLY) {
1150                         /* CANCEL RPC should not be sent to server. */
1151                         list_del_init(&lock->l_bl_ast);
1152                         LDLM_LOCK_PUT(lock);
1153                         count--;
1154                 }
1155
1156         }
1157         if (bl_ast > 0) {
1158                 count -= bl_ast;
1159                 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1160         }
1161
1162         RETURN(count);
1163 }
1164
1165 /* Return 1 to stop lru processing and keep current lock cached. Return zero
1166  * otherwise. */
1167 static ldlm_policy_res_t ldlm_cancel_shrink_policy(struct ldlm_namespace *ns,
1168                                                    struct ldlm_lock *lock,
1169                                                    int unused, int added,
1170                                                    int count)
1171 {
1172         int lock_cost;
1173         __u64 page_nr;
1174
1175         /* Stop lru processing when we reached passed @count or checked all
1176          * locks in lru. */
1177         if (count && added >= count)
1178                 return LDLM_POLICY_KEEP_LOCK;
1179
1180         if (lock->l_resource->lr_type == LDLM_EXTENT) {
1181                 struct ldlm_extent *l_extent;
1182
1183                 /* For all extent locks cost is 1 + number of pages in
1184                  * their extent. */
1185                 l_extent = &lock->l_policy_data.l_extent;
1186                 page_nr = (l_extent->end - l_extent->start);
1187                 do_div(page_nr, CFS_PAGE_SIZE);
1188
1189 #ifdef __KERNEL__
1190                 /* XXX: In fact this is evil hack, we can't access inode
1191                  * here. For doing it right we need somehow to have number
1192                  * of covered by lock. This should be fixed later when 10718
1193                  * is landed. */
1194                 if (lock->l_ast_data != NULL) {
1195                         struct inode *inode = lock->l_ast_data;
1196                         if (page_nr > inode->i_mapping->nrpages)
1197                                 page_nr = inode->i_mapping->nrpages;
1198                 }
1199 #endif
1200                 lock_cost = 1 + page_nr;
1201         } else {
1202                 /* For all locks which are not extent ones cost is 1 */
1203                 lock_cost = 1;
1204         }
1205
1206         /* Keep all expensive locks in lru for the memory pressure time
1207          * cancel policy. They anyways may be canceled by lru resize
1208          * pplicy if they have not small enough CLV. */
1209         return lock_cost > ns->ns_shrink_thumb ?
1210                 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1211 }
1212
1213 /* Return 1 to stop lru processing and keep current lock cached. Return zero
1214  * otherwise. */
1215 static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1216                                                  struct ldlm_lock *lock,
1217                                                  int unused, int added,
1218                                                  int count)
1219 {
1220         cfs_time_t cur = cfs_time_current();
1221         struct ldlm_pool *pl = &ns->ns_pool;
1222         __u64 slv, lvf, lv;
1223         cfs_time_t la;
1224
1225         /* Stop lru processing when we reached passed @count or checked all
1226          * locks in lru. */
1227         if (count && added >= count)
1228                 return LDLM_POLICY_KEEP_LOCK;
1229
1230         slv = ldlm_pool_get_slv(pl);
1231         lvf = ldlm_pool_get_lvf(pl);
1232         la = cfs_duration_sec(cfs_time_sub(cur,
1233                               lock->l_last_used));
1234
1235         /* Stop when slv is not yet come from server or
1236          * lv is smaller than it is. */
1237         lv = lvf * la * unused;
1238
1239         /* Inform pool about current CLV to see it via proc. */
1240         ldlm_pool_set_clv(pl, lv);
1241         return (slv == 1 || lv < slv) ?
1242                 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1243 }
1244
1245 /* Return 1 to stop lru processing and keep current lock cached. Return zero
1246  * otherwise. */
1247 static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1248                                                    struct ldlm_lock *lock,
1249                                                    int unused, int added,
1250                                                    int count)
1251 {
1252         /* Stop lru processing when we reached passed @count or checked all
1253          * locks in lru. */
1254         return (added >= count) ?
1255                 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1256 }
1257
1258 /* Return 1 to stop lru processing and keep current lock cached. Return zero
1259  * otherwise. */
1260 static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1261                                                  struct ldlm_lock *lock,
1262                                                  int unused, int added,
1263                                                  int count)
1264 {
1265         /* Stop lru processing if young lock is found and we reached passed
1266          * @count. */
1267         return ((added >= count) &&
1268                 cfs_time_before(cfs_time_current(),
1269                                 cfs_time_add(lock->l_last_used,
1270                                              ns->ns_max_age))) ?
1271                 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1272 }
1273
1274 /* Return 1 to stop lru processing and keep current lock cached. Return zero
1275  * otherwise. */
1276 static ldlm_policy_res_t ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1277                                                     struct ldlm_lock *lock,
1278                                                     int unused, int added,
1279                                                     int count)
1280 {
1281         /* Stop lru processing when we reached passed @count or checked all
1282          * locks in lru. */
1283         return (added >= count) ?
1284                 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1285 }
1286
1287 typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *,
1288                                                       struct ldlm_lock *, int,
1289                                                       int, int);
1290
1291 static ldlm_cancel_lru_policy_t
1292 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
1293 {
1294         if (ns_connect_lru_resize(ns)) {
1295                 if (flags & LDLM_CANCEL_SHRINK)
1296                         return ldlm_cancel_shrink_policy;
1297                 else if (flags & LDLM_CANCEL_LRUR)
1298                         return ldlm_cancel_lrur_policy;
1299                 else if (flags & LDLM_CANCEL_PASSED)
1300                         return ldlm_cancel_passed_policy;
1301         } else {
1302                 if (flags & LDLM_CANCEL_AGED)
1303                         return ldlm_cancel_aged_policy;
1304         }
1305
1306         return ldlm_cancel_default_policy;
1307 }
1308
1309 /* - Free space in lru for @count new locks,
1310  *   redundant unused locks are canceled locally;
1311  * - also cancel locally unused aged locks;
1312  * - do not cancel more than @max locks;
1313  * - GET the found locks and add them into the @cancels list.
1314  *
1315  * A client lock can be added to the l_bl_ast list only when it is
1316  * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing CANCEL.
1317  * There are the following use cases: ldlm_cancel_resource_local(),
1318  * ldlm_cancel_lru_local() and ldlm_cli_cancel(), which check&set this
1319  * flag properly. As any attempt to cancel a lock rely on this flag,
1320  * l_bl_ast list is accessed later without any special locking.
1321  *
1322  * Calling policies for enabled lru resize:
1323  * ----------------------------------------
1324  * flags & LDLM_CANCEL_LRUR - use lru resize policy (SLV from server) to
1325  *                            cancel not more than @count locks;
1326  *
1327  * flags & LDLM_CANCEL_PASSED - cancel @count number of old locks (located at
1328  *                              the beginning of lru list);
1329  *
1330  * flags & LDLM_CANCEL_SHRINK - cancel not more than @count locks according to
1331  *                              memory pressre policy function;
1332  *
1333  * flags & LDLM_CANCEL_AGED -   cancel alocks according to "aged policy".
1334  */
1335 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
1336                           int count, int max, int cancel_flags, int flags)
1337 {
1338         ldlm_cancel_lru_policy_t pf;
1339         struct ldlm_lock *lock, *next;
1340         int added = 0, unused;
1341         ENTRY;
1342
1343         spin_lock(&ns->ns_unused_lock);
1344         unused = ns->ns_nr_unused;
1345
1346         if (!ns_connect_lru_resize(ns))
1347                 count += unused - ns->ns_max_unused;
1348
1349         pf = ldlm_cancel_lru_policy(ns, flags);
1350         LASSERT(pf != NULL);
1351
1352         while (!list_empty(&ns->ns_unused_list)) {
1353                 /* For any flags, stop scanning if @max is reached. */
1354                 if (max && added >= max)
1355                         break;
1356
1357                 list_for_each_entry_safe(lock, next, &ns->ns_unused_list, l_lru){
1358                         /* No locks which got blocking requests. */
1359                         LASSERT(!(lock->l_flags & LDLM_FL_BL_AST));
1360
1361                         /* Somebody is already doing CANCEL. No need in this
1362                          * lock in lru, do not traverse it again. */
1363                         if (!(lock->l_flags & LDLM_FL_CANCELING))
1364                                 break;
1365
1366                         ldlm_lock_remove_from_lru_nolock(lock);
1367                 }
1368                 if (&lock->l_lru == &ns->ns_unused_list)
1369                         break;
1370
1371                 /* Pass the lock through the policy filter and see if it
1372                  * should stay in lru.
1373                  *
1374                  * Even for shrinker policy we stop scanning if
1375                  * we find a lock that should stay in the cache.
1376                  * We should take into account lock age anyway
1377                  * as new lock even if it is small of weight is
1378                  * valuable resource.
1379                  *
1380                  * That is, for shrinker policy we drop only
1381                  * old locks, but additionally chose them by
1382                  * their weight. Big extent locks will stay in
1383                  * the cache. */
1384                 if (pf(ns, lock, unused, added, count) == LDLM_POLICY_KEEP_LOCK)
1385                         break;
1386
1387                 LDLM_LOCK_GET(lock); /* dropped by bl thread */
1388                 spin_unlock(&ns->ns_unused_lock);
1389
1390                 lock_res_and_lock(lock);
1391                 /* Check flags again under the lock. */
1392                 if ((lock->l_flags & LDLM_FL_CANCELING) ||
1393                     (ldlm_lock_remove_from_lru(lock) == 0)) {
1394                         /* other thread is removing lock from lru or
1395                          * somebody is already doing CANCEL or
1396                          * there is a blocking request which will send
1397                          * cancel by itseft or the lock is matched
1398                          * is already not unused. */
1399                         unlock_res_and_lock(lock);
1400                         LDLM_LOCK_PUT(lock);
1401                         spin_lock(&ns->ns_unused_lock);
1402                         continue;
1403                 }
1404                 LASSERT(!lock->l_readers && !lock->l_writers);
1405
1406                 /* If we have chosen to cancel this lock voluntarily, we
1407                  * better send cancel notification to server, so that it
1408                  * frees appropriate state. This might lead to a race
1409                  * where while we are doing cancel here, server is also
1410                  * silently cancelling this lock. */
1411                 lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK;
1412
1413                 /* Setting the CBPENDING flag is a little misleading,
1414                  * but prevents an important race; namely, once
1415                  * CBPENDING is set, the lock can accumulate no more
1416                  * readers/writers. Since readers and writers are
1417                  * already zero here, ldlm_lock_decref() won't see
1418                  * this flag and call l_blocking_ast */
1419                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1420
1421                 /* We can't re-add to l_lru as it confuses the
1422                  * refcounting in ldlm_lock_remove_from_lru() if an AST
1423                  * arrives after we drop ns_lock below. We use l_bl_ast
1424                  * and can't use l_pending_chain as it is used both on
1425                  * server and client nevertheless bug 5666 says it is
1426                  * used only on server */
1427                 LASSERT(list_empty(&lock->l_bl_ast));
1428                 list_add(&lock->l_bl_ast, cancels);
1429                 unlock_res_and_lock(lock);
1430                 spin_lock(&ns->ns_unused_lock);
1431                 added++;
1432                 unused--;
1433         }
1434         spin_unlock(&ns->ns_unused_lock);
1435         RETURN(ldlm_cancel_list(cancels, added, cancel_flags));
1436 }
1437
1438 /* Returns number of locks which could be canceled next time when
1439  * ldlm_cancel_lru() is called. Used from locks pool shrinker. */
1440 int ldlm_cancel_lru_estimate(struct ldlm_namespace *ns,
1441                              int count, int max, int flags)
1442 {
1443         ldlm_cancel_lru_policy_t pf;
1444         struct ldlm_lock *lock;
1445         int added = 0, unused;
1446         ENTRY;
1447
1448         pf = ldlm_cancel_lru_policy(ns, flags);
1449         LASSERT(pf != NULL);
1450         spin_lock(&ns->ns_unused_lock);
1451         unused = ns->ns_nr_unused;
1452
1453         list_for_each_entry(lock, &ns->ns_unused_list, l_lru) {
1454                 /* For any flags, stop scanning if @max is reached. */
1455                 if (max && added >= max)
1456                         break;
1457
1458                 /* Somebody is already doing CANCEL or there is a
1459                  * blocking request will send cancel. Let's not count
1460                  * this lock. */
1461                 if ((lock->l_flags & LDLM_FL_CANCELING) ||
1462                     (lock->l_flags & LDLM_FL_BL_AST))
1463                         continue;
1464
1465                 /* Pass the lock through the policy filter and see if it
1466                  * should stay in lru. */
1467                 if (pf(ns, lock, unused, added, count) == LDLM_POLICY_KEEP_LOCK)
1468                         break;
1469
1470                 added++;
1471                 unused--;
1472         }
1473         spin_unlock(&ns->ns_unused_lock);
1474         RETURN(added);
1475 }
1476
1477 /* when called with LDLM_ASYNC the blocking callback will be handled
1478  * in a thread and this function will return after the thread has been
1479  * asked to call the callback.  when called with LDLM_SYNC the blocking
1480  * callback will be performed in this function. */
1481 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr, ldlm_sync_t sync,
1482                     int flags)
1483 {
1484         CFS_LIST_HEAD(cancels);
1485         int count, rc;
1486         ENTRY;
1487
1488 #ifndef __KERNEL__
1489         sync = LDLM_SYNC; /* force to be sync in user space */
1490 #endif
1491         count = ldlm_cancel_lru_local(ns, &cancels, nr, 0, 0, flags);
1492         if (sync == LDLM_ASYNC) {
1493                 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count);
1494                 if (rc == 0)
1495                         RETURN(count);
1496         }
1497
1498         /* If an error occured in ASYNC mode, or
1499          * this is SYNC mode, cancel the list. */
1500         ldlm_cli_cancel_list(&cancels, count, NULL, 0);
1501         RETURN(count);
1502 }
1503
1504 /* Find and cancel locally unused locks found on resource, matched to the
1505  * given policy, mode. GET the found locks and add them into the @cancels
1506  * list. */
1507 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1508                                struct list_head *cancels,
1509                                ldlm_policy_data_t *policy,
1510                                ldlm_mode_t mode, int lock_flags,
1511                                int cancel_flags, void *opaque)
1512 {
1513         struct ldlm_lock *lock;
1514         int count = 0;
1515         ENTRY;
1516
1517         lock_res(res);
1518         list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1519                 if (opaque != NULL && lock->l_ast_data != opaque) {
1520                         LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1521                                    lock->l_ast_data, opaque);
1522                         //LBUG();
1523                         continue;
1524                 }
1525
1526                 if (lock->l_readers || lock->l_writers) {
1527                         if (cancel_flags & LDLM_FL_WARN) {
1528                                 LDLM_ERROR(lock, "lock in use");
1529                                 //LBUG();
1530                         }
1531                         continue;
1532                 }
1533
1534                 /* If somebody is already doing CANCEL, or blocking ast came,
1535                  * skip this lock. */
1536                 if (lock->l_flags & LDLM_FL_BL_AST ||
1537                     lock->l_flags & LDLM_FL_CANCELING)
1538                         continue;
1539
1540                 if (lockmode_compat(lock->l_granted_mode, mode))
1541                         continue;
1542
1543                 /* If policy is given and this is IBITS lock, add to list only
1544                  * those locks that match by policy. */
1545                 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1546                     !(lock->l_policy_data.l_inodebits.bits &
1547                       policy->l_inodebits.bits))
1548                         continue;
1549
1550                 /* See CBPENDING comment in ldlm_cancel_lru */
1551                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1552                                  lock_flags;
1553
1554                 LASSERT(list_empty(&lock->l_bl_ast));
1555                 list_add(&lock->l_bl_ast, cancels);
1556                 LDLM_LOCK_GET(lock);
1557                 count++;
1558         }
1559         unlock_res(res);
1560
1561         RETURN(ldlm_cancel_list(cancels, count, cancel_flags));
1562 }
1563
1564 /* If @req is NULL, send CANCEL request to server with handles of locks
1565  * in the @cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1566  * separately per lock.
1567  * If @req is not NULL, put handles of locks in @cancels into the request
1568  * buffer at the offset @off.
1569  * Destroy @cancels at the end. */
1570 int ldlm_cli_cancel_list(struct list_head *cancels, int count,
1571                          struct ptlrpc_request *req, int flags)
1572 {
1573         struct ldlm_lock *lock;
1574         int res = 0;
1575         ENTRY;
1576
1577         if (list_empty(cancels) || count == 0)
1578                 RETURN(0);
1579
1580         /* XXX: requests (both batched and not) could be sent in parallel.
1581          * Usually it is enough to have just 1 RPC, but it is possible that
1582          * there are to many locks to be cancelled in LRU or on a resource.
1583          * It would also speed up the case when the server does not support
1584          * the feature. */
1585         while (count > 0) {
1586                 LASSERT(!list_empty(cancels));
1587                 lock = list_entry(cancels->next, struct ldlm_lock, l_bl_ast);
1588                 LASSERT(lock->l_conn_export);
1589
1590                 if (exp_connect_cancelset(lock->l_conn_export)) {
1591                         res = count;
1592                         if (req)
1593                                 ldlm_cancel_pack(req, cancels, count);
1594                         else
1595                                 res = ldlm_cli_cancel_req(lock->l_conn_export,
1596                                                           cancels, count,
1597                                                           flags);
1598                 } else {
1599                         res = ldlm_cli_cancel_req(lock->l_conn_export,
1600                                                   cancels, 1, flags);
1601                 }
1602
1603                 if (res < 0) {
1604                         CERROR("ldlm_cli_cancel_list: %d\n", res);
1605                         res = count;
1606                 }
1607
1608                 count -= res;
1609                 ldlm_lock_list_put(cancels, l_bl_ast, res);
1610         }
1611         LASSERT(count == 0);
1612         RETURN(0);
1613 }
1614
1615 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1616                                     const struct ldlm_res_id *res_id,
1617                                     ldlm_policy_data_t *policy,
1618                                     ldlm_mode_t mode, int flags, void *opaque)
1619 {
1620         struct ldlm_resource *res;
1621         CFS_LIST_HEAD(cancels);
1622         int count;
1623         int rc;
1624         ENTRY;
1625
1626         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1627         if (res == NULL) {
1628                 /* This is not a problem. */
1629                 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id->name[0]);
1630                 RETURN(0);
1631         }
1632
1633         count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
1634                                            0, flags, opaque);
1635         rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
1636         if (rc != ELDLM_OK)
1637                 CERROR("ldlm_cli_cancel_unused_resource: %d\n", rc);
1638
1639         ldlm_resource_putref(res);
1640         RETURN(0);
1641 }
1642
1643 static inline int have_no_nsresource(struct ldlm_namespace *ns)
1644 {
1645         int no_resource = 0;
1646
1647         spin_lock(&ns->ns_hash_lock);
1648         if (ns->ns_resources == 0)
1649                 no_resource = 1;
1650         spin_unlock(&ns->ns_hash_lock);
1651
1652         RETURN(no_resource);
1653 }
1654
1655 /* Cancel all locks on a namespace (or a specific resource, if given)
1656  * that have 0 readers/writers.
1657  *
1658  * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
1659  * to notify the server. */
1660 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1661                            const struct ldlm_res_id *res_id,
1662                            int flags, void *opaque)
1663 {
1664         int i;
1665         ENTRY;
1666
1667         if (ns == NULL)
1668                 RETURN(ELDLM_OK);
1669
1670         if (res_id)
1671                 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
1672                                                        LCK_MINMODE, flags,
1673                                                        opaque));
1674
1675         spin_lock(&ns->ns_hash_lock);
1676         for (i = 0; i < RES_HASH_SIZE; i++) {
1677                 struct list_head *tmp;
1678                 tmp = ns->ns_hash[i].next;
1679                 while (tmp != &(ns->ns_hash[i])) {
1680                         struct ldlm_resource *res;
1681                         int rc;
1682
1683                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
1684                         ldlm_resource_getref(res);
1685                         spin_unlock(&ns->ns_hash_lock);
1686
1687                         rc = ldlm_cli_cancel_unused_resource(ns, &res->lr_name,
1688                                                              NULL, LCK_MINMODE,
1689                                                              flags, opaque);
1690
1691                         if (rc)
1692                                 CERROR("ldlm_cli_cancel_unused ("LPU64"): %d\n",
1693                                        res->lr_name.name[0], rc);
1694
1695                         spin_lock(&ns->ns_hash_lock);
1696                         tmp = tmp->next;
1697                         ldlm_resource_putref_locked(res);
1698                 }
1699         }
1700         spin_unlock(&ns->ns_hash_lock);
1701
1702         RETURN(ELDLM_OK);
1703 }
1704
1705 /* join/split resource locks to/from lru list */
1706 int ldlm_cli_join_lru(struct ldlm_namespace *ns,
1707                       const struct ldlm_res_id *res_id, int join)
1708 {
1709         struct ldlm_resource *res;
1710         struct ldlm_lock *lock, *n;
1711         int count = 0;
1712         ENTRY;
1713
1714         LASSERT(ns_is_client(ns));
1715
1716         res = ldlm_resource_get(ns, NULL, res_id, LDLM_EXTENT, 0);
1717         if (res == NULL)
1718                 RETURN(count);
1719         LASSERT(res->lr_type == LDLM_EXTENT);
1720
1721         lock_res(res);
1722         if (!join)
1723                 goto split;
1724
1725         list_for_each_entry_safe (lock, n, &res->lr_granted, l_res_link) {
1726                 if (list_empty(&lock->l_lru) &&
1727                     !lock->l_readers && !lock->l_writers &&
1728                     !(lock->l_flags & LDLM_FL_LOCAL) &&
1729                     !(lock->l_flags & LDLM_FL_CBPENDING) &&
1730                     !(lock->l_flags & LDLM_FL_BL_AST)) {
1731                         ldlm_lock_add_to_lru(lock);
1732                         lock->l_flags &= ~LDLM_FL_NO_LRU;
1733                         LDLM_DEBUG(lock, "join lock to lru");
1734                         count++;
1735                 }
1736         }
1737         goto unlock;
1738 split:
1739         spin_lock(&ns->ns_unused_lock);
1740         list_for_each_entry_safe (lock, n, &ns->ns_unused_list, l_lru) {
1741                 if (lock->l_resource == res) {
1742                         ldlm_lock_remove_from_lru_nolock(lock);
1743                         lock->l_flags |= LDLM_FL_NO_LRU;
1744                         LDLM_DEBUG(lock, "split lock from lru");
1745                         count++;
1746                 }
1747         }
1748         spin_unlock(&ns->ns_unused_lock);
1749 unlock:
1750         unlock_res(res);
1751         ldlm_resource_putref(res);
1752         RETURN(count);
1753 }
1754
1755 /* Lock iterators. */
1756
1757 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
1758                           void *closure)
1759 {
1760         struct list_head *tmp, *next;
1761         struct ldlm_lock *lock;
1762         int rc = LDLM_ITER_CONTINUE;
1763
1764         ENTRY;
1765
1766         if (!res)
1767                 RETURN(LDLM_ITER_CONTINUE);
1768
1769         lock_res(res);
1770         list_for_each_safe(tmp, next, &res->lr_granted) {
1771                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1772
1773                 if (iter(lock, closure) == LDLM_ITER_STOP)
1774                         GOTO(out, rc = LDLM_ITER_STOP);
1775         }
1776
1777         list_for_each_safe(tmp, next, &res->lr_converting) {
1778                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1779
1780                 if (iter(lock, closure) == LDLM_ITER_STOP)
1781                         GOTO(out, rc = LDLM_ITER_STOP);
1782         }
1783
1784         list_for_each_safe(tmp, next, &res->lr_waiting) {
1785                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1786
1787                 if (iter(lock, closure) == LDLM_ITER_STOP)
1788                         GOTO(out, rc = LDLM_ITER_STOP);
1789         }
1790  out:
1791         unlock_res(res);
1792         RETURN(rc);
1793 }
1794
1795 struct iter_helper_data {
1796         ldlm_iterator_t iter;
1797         void *closure;
1798 };
1799
1800 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
1801 {
1802         struct iter_helper_data *helper = closure;
1803         return helper->iter(lock, helper->closure);
1804 }
1805
1806 static int ldlm_res_iter_helper(struct ldlm_resource *res, void *closure)
1807 {
1808         return ldlm_resource_foreach(res, ldlm_iter_helper, closure);
1809 }
1810
1811 int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
1812                            void *closure)
1813 {
1814         struct iter_helper_data helper = { iter: iter, closure: closure };
1815         return ldlm_namespace_foreach_res(ns, ldlm_res_iter_helper, &helper);
1816 }
1817
1818 int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
1819                                ldlm_res_iterator_t iter, void *closure)
1820 {
1821         int i, rc = LDLM_ITER_CONTINUE;
1822         struct ldlm_resource *res;
1823         struct list_head *tmp;
1824
1825         ENTRY;
1826         spin_lock(&ns->ns_hash_lock);
1827         for (i = 0; i < RES_HASH_SIZE; i++) {
1828                 tmp = ns->ns_hash[i].next;
1829                 while (tmp != &(ns->ns_hash[i])) {
1830                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
1831                         ldlm_resource_getref(res);
1832                         spin_unlock(&ns->ns_hash_lock);
1833
1834                         rc = iter(res, closure);
1835
1836                         spin_lock(&ns->ns_hash_lock);
1837                         tmp = tmp->next;
1838                         ldlm_resource_putref_locked(res);
1839                         if (rc == LDLM_ITER_STOP)
1840                                 GOTO(out, rc);
1841                 }
1842         }
1843  out:
1844         spin_unlock(&ns->ns_hash_lock);
1845         RETURN(rc);
1846 }
1847
1848 /* non-blocking function to manipulate a lock whose cb_data is being put away.*/
1849 void ldlm_resource_iterate(struct ldlm_namespace *ns,
1850                            const struct ldlm_res_id *res_id,
1851                            ldlm_iterator_t iter, void *data)
1852 {
1853         struct ldlm_resource *res;
1854         ENTRY;
1855
1856         if (ns == NULL) {
1857                 CERROR("must pass in namespace\n");
1858                 LBUG();
1859         }
1860
1861         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1862         if (res == NULL) {
1863                 EXIT;
1864                 return;
1865         }
1866
1867         ldlm_resource_foreach(res, iter, data);
1868         ldlm_resource_putref(res);
1869         EXIT;
1870 }
1871
1872 /* Lock replay */
1873
1874 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
1875 {
1876         struct list_head *list = closure;
1877
1878         /* we use l_pending_chain here, because it's unused on clients. */
1879         LASSERTF(list_empty(&lock->l_pending_chain),"lock %p next %p prev %p\n",
1880                  lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
1881         /* bug 9573: don't replay locks left after eviction */
1882         if (!(lock->l_flags & LDLM_FL_FAILED))
1883                 list_add(&lock->l_pending_chain, list);
1884         return LDLM_ITER_CONTINUE;
1885 }
1886
1887 static int replay_lock_interpret(struct ptlrpc_request *req,
1888                                  struct ldlm_async_args *aa, int rc)
1889 {
1890         struct ldlm_lock  *lock;
1891         struct ldlm_reply *reply;
1892
1893         ENTRY;
1894         atomic_dec(&req->rq_import->imp_replay_inflight);
1895         if (rc != ELDLM_OK)
1896                 GOTO(out, rc);
1897
1898
1899         reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1900         if (reply == NULL)
1901                 GOTO(out, rc = -EPROTO);
1902
1903         lock = ldlm_handle2lock(&aa->lock_handle);
1904         if (!lock) {
1905                 CERROR("received replay ack for unknown local cookie "LPX64
1906                        " remote cookie "LPX64 " from server %s id %s\n",
1907                        aa->lock_handle.cookie, reply->lock_handle.cookie,
1908                        req->rq_export->exp_client_uuid.uuid,
1909                        libcfs_id2str(req->rq_peer));
1910                 GOTO(out, rc = -ESTALE);
1911         }
1912
1913         lock->l_remote_handle = reply->lock_handle;
1914         LDLM_DEBUG(lock, "replayed lock:");
1915         ptlrpc_import_recovery_state_machine(req->rq_import);
1916         LDLM_LOCK_PUT(lock);
1917 out:
1918         if (rc != ELDLM_OK)
1919                 ptlrpc_connect_import(req->rq_import, NULL);
1920
1921
1922         RETURN(rc);
1923 }
1924
1925 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
1926 {
1927         struct ptlrpc_request *req;
1928         struct ldlm_async_args *aa;
1929         struct ldlm_request   *body;
1930         int flags;
1931         ENTRY;
1932
1933
1934         /* Bug 11974: Do not replay a lock which is actively being canceled */
1935         if (lock->l_flags & LDLM_FL_CANCELING) {
1936                 LDLM_DEBUG(lock, "Not replaying canceled lock:");
1937                 RETURN(0);
1938         }
1939
1940         /* If this is reply-less callback lock, we cannot replay it, since
1941          * server might have long dropped it, but notification of that event was
1942          * lost by network. (and server granted conflicting lock already) */
1943         if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1944                 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
1945                 ldlm_lock_cancel(lock);
1946                 RETURN(0);
1947         }
1948         /*
1949          * If granted mode matches the requested mode, this lock is granted.
1950          *
1951          * If they differ, but we have a granted mode, then we were granted
1952          * one mode and now want another: ergo, converting.
1953          *
1954          * If we haven't been granted anything and are on a resource list,
1955          * then we're blocked/waiting.
1956          *
1957          * If we haven't been granted anything and we're NOT on a resource list,
1958          * then we haven't got a reply yet and don't have a known disposition.
1959          * This happens whenever a lock enqueue is the request that triggers
1960          * recovery.
1961          */
1962         if (lock->l_granted_mode == lock->l_req_mode)
1963                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
1964         else if (lock->l_granted_mode)
1965                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
1966         else if (!list_empty(&lock->l_res_link))
1967                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
1968         else
1969                 flags = LDLM_FL_REPLAY;
1970
1971         req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
1972                                         LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
1973         if (req == NULL)
1974                 RETURN(-ENOMEM);
1975
1976         /* We're part of recovery, so don't wait for it. */
1977         req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
1978
1979         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1980         ldlm_lock2desc(lock, &body->lock_desc);
1981         body->lock_flags = flags;
1982
1983         ldlm_lock2handle(lock, &body->lock_handle[0]);
1984         if (lock->l_lvb_len != 0) {
1985                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
1986                 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1987                                      lock->l_lvb_len);
1988         }
1989         ptlrpc_request_set_replen(req);
1990         /* notify the server we've replayed all requests.
1991          * also, we mark the request to be put on a dedicated
1992          * queue to be processed after all request replayes.
1993          * bug 6063 */
1994         lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
1995
1996         LDLM_DEBUG(lock, "replaying lock:");
1997
1998         atomic_inc(&req->rq_import->imp_replay_inflight);
1999         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2000         aa = (struct ldlm_async_args *)&req->rq_async_args;
2001         aa->lock_handle = body->lock_handle[0];
2002         req->rq_interpret_reply = replay_lock_interpret;
2003         ptlrpcd_add_req(req);
2004
2005         RETURN(0);
2006 }
2007
2008 int ldlm_replay_locks(struct obd_import *imp)
2009 {
2010         struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2011         CFS_LIST_HEAD(list);
2012         struct ldlm_lock *lock, *next;
2013         int rc = 0;
2014
2015         ENTRY;
2016
2017         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
2018
2019         /* ensure this doesn't fall to 0 before all have been queued */
2020         atomic_inc(&imp->imp_replay_inflight);
2021
2022         (void)ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2023
2024         list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2025                 list_del_init(&lock->l_pending_chain);
2026                 if (rc)
2027                         continue; /* or try to do the rest? */
2028                 rc = replay_one_lock(imp, lock);
2029         }
2030
2031         atomic_dec(&imp->imp_replay_inflight);
2032
2033         RETURN(rc);
2034 }