kernel/drivers/staging/lustre/lustre/ldlm/ldlm_request.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2010, 2012, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36 /**
  37  * This file contains Asynchronous System Trap (AST) handlers and related
  38  * LDLM request-processing routines.
  39  *
  40  * An AST is a callback issued on a lock when its state is changed. There are
  41  * several different types of ASTs (callbacks) registered for each lock:
  42  *
  43  * - completion AST: when a lock is enqueued by some process, but cannot be
  44  *   granted immediately due to other conflicting locks on the same resource,
  45  *   the completion AST is sent to notify the caller when the lock is
  46  *   eventually granted
  47  *
  48  * - blocking AST: when a lock is granted to some process, if another process
  49  *   enqueues a conflicting (blocking) lock on a resource, a blocking AST is
  50  *   sent to notify the holder(s) of the lock(s) of the conflicting lock
  51  *   request. The lock holder(s) must release their lock(s) on that resource in
  52  *   a timely manner or be evicted by the server.
  53  *
  54  * - glimpse AST: this is used when a process wants information about a lock
  55  *   (i.e. the lock value block (LVB)) but does not necessarily require holding
  56  *   the lock. If the resource is locked, the lock holder(s) are sent glimpse
  57  *   ASTs and the LVB is returned to the caller, and lock holder(s) may CANCEL
  58  *   their lock(s) if they are idle. If the resource is not locked, the server
  59  *   may grant the lock.
  60  */
  61
  62 #define DEBUG_SUBSYSTEM S_LDLM
  63
  64 #include "../include/lustre_dlm.h"
  65 #include "../include/obd_class.h"
  66 #include "../include/obd.h"
  67
  68 #include "ldlm_internal.h"
  69
  70 int ldlm_enqueue_min = OBD_TIMEOUT_DEFAULT;
  71 module_param(ldlm_enqueue_min, int, 0644);
  72 MODULE_PARM_DESC(ldlm_enqueue_min, "lock enqueue timeout minimum");
  73
  74 /* in client side, whether the cached locks will be canceled before replay */
  75 unsigned int ldlm_cancel_unused_locks_before_replay = 1;
  76
  77 static void interrupted_completion_wait(void *data)
  78 {
  79 }
  80
  81 struct lock_wait_data {
  82         struct ldlm_lock *lwd_lock;
  83         __u32        lwd_conn_cnt;
  84 };
  85
  86 struct ldlm_async_args {
  87         struct lustre_handle lock_handle;
  88 };
  89
  90 int ldlm_expired_completion_wait(void *data)
  91 {
  92         struct lock_wait_data *lwd = data;
  93         struct ldlm_lock *lock = lwd->lwd_lock;
  94         struct obd_import *imp;
  95         struct obd_device *obd;
  96
  97         if (lock->l_conn_export == NULL) {
  98                 static unsigned long next_dump, last_dump;
  99
 100                 LCONSOLE_WARN("lock timed out (enqueued at "CFS_TIME_T", "
 101                               CFS_DURATION_T"s ago)\n",
 102                               lock->l_last_activity,
 103                               cfs_time_sub(get_seconds(),
 104                                            lock->l_last_activity));
 105                 LDLM_DEBUG(lock, "lock timed out (enqueued at " CFS_TIME_T ", " CFS_DURATION_T "s ago); not entering recovery in server code, just going back to sleep",
 106                            lock->l_last_activity,
 107                            cfs_time_sub(get_seconds(),
 108                                         lock->l_last_activity));
 109                 if (cfs_time_after(cfs_time_current(), next_dump)) {
 110                         last_dump = next_dump;
 111                         next_dump = cfs_time_shift(300);
 112                         ldlm_namespace_dump(D_DLMTRACE,
 113                                             ldlm_lock_to_ns(lock));
 114                         if (last_dump == 0)
 115                                 libcfs_debug_dumplog();
 116                 }
 117                 return 0;
 118         }
 119
 120         obd = lock->l_conn_export->exp_obd;
 121         imp = obd->u.cli.cl_import;
 122         ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
 123         LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", "
 124                   CFS_DURATION_T"s ago), entering recovery for %s@%s",
 125                   lock->l_last_activity,
 126                   cfs_time_sub(get_seconds(), lock->l_last_activity),
 127                   obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
 128
 129         return 0;
 130 }
 131 EXPORT_SYMBOL(ldlm_expired_completion_wait);
 132
 133 /* We use the same basis for both server side and client side functions
 134    from a single node. */
 135 int ldlm_get_enq_timeout(struct ldlm_lock *lock)
 136 {
 137         int timeout = at_get(ldlm_lock_to_ns_at(lock));
 138
 139         if (AT_OFF)
 140                 return obd_timeout / 2;
 141         /* Since these are non-updating timeouts, we should be conservative.
 142            It would be nice to have some kind of "early reply" mechanism for
 143            lock callbacks too... */
 144         timeout = min_t(int, at_max, timeout + (timeout >> 1)); /* 150% */
 145         return max(timeout, ldlm_enqueue_min);
 146 }
 147 EXPORT_SYMBOL(ldlm_get_enq_timeout);
 148
 149 /**
 150  * Helper function for ldlm_completion_ast(), updating timings when lock is
 151  * actually granted.
 152  */
 153 static int ldlm_completion_tail(struct ldlm_lock *lock)
 154 {
 155         long delay;
 156         int  result;
 157
 158         if (lock->l_flags & (LDLM_FL_DESTROYED | LDLM_FL_FAILED)) {
 159                 LDLM_DEBUG(lock, "client-side enqueue: destroyed");
 160                 result = -EIO;
 161         } else {
 162                 delay = cfs_time_sub(get_seconds(),
 163                                      lock->l_last_activity);
 164                 LDLM_DEBUG(lock, "client-side enqueue: granted after "
 165                            CFS_DURATION_T"s", delay);
 166
 167                 /* Update our time estimate */
 168                 at_measured(ldlm_lock_to_ns_at(lock),
 169                             delay);
 170                 result = 0;
 171         }
 172         return result;
 173 }
 174
 175 /**
 176  * Implementation of ->l_completion_ast() for a client, that doesn't wait
 177  * until lock is granted. Suitable for locks enqueued through ptlrpcd, of
 178  * other threads that cannot block for long.
 179  */
 180 int ldlm_completion_ast_async(struct ldlm_lock *lock, __u64 flags, void *data)
 181 {
 182         if (flags == LDLM_FL_WAIT_NOREPROC) {
 183                 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
 184                 return 0;
 185         }
 186
 187         if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
 188                        LDLM_FL_BLOCK_CONV))) {
 189                 wake_up(&lock->l_waitq);
 190                 return ldlm_completion_tail(lock);
 191         }
 192
 193         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, going forward");
 194         ldlm_reprocess_all(lock->l_resource);
 195         return 0;
 196 }
 197 EXPORT_SYMBOL(ldlm_completion_ast_async);
 198
 199 /**
 200  * Generic LDLM "completion" AST. This is called in several cases:
 201  *
 202  *     - when a reply to an ENQUEUE RPC is received from the server
 203  *       (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at
 204  *       this point (determined by flags);
 205  *
 206  *     - when LDLM_CP_CALLBACK RPC comes to client to notify it that lock has
 207  *       been granted;
 208  *
 209  *     - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock
 210  *       gets correct lvb;
 211  *
 212  *     - to force all locks when resource is destroyed (cleanup_resource());
 213  *
 214  *     - during lock conversion (not used currently).
 215  *
 216  * If lock is not granted in the first case, this function waits until second
 217  * or penultimate cases happen in some other thread.
 218  *
 219  */
 220 int ldlm_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
 221 {
 222         /* XXX ALLOCATE - 160 bytes */
 223         struct lock_wait_data lwd;
 224         struct obd_device *obd;
 225         struct obd_import *imp = NULL;
 226         struct l_wait_info lwi;
 227         __u32 timeout;
 228         int rc = 0;
 229
 230         if (flags == LDLM_FL_WAIT_NOREPROC) {
 231                 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
 232                 goto noreproc;
 233         }
 234
 235         if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
 236                        LDLM_FL_BLOCK_CONV))) {
 237                 wake_up(&lock->l_waitq);
 238                 return 0;
 239         }
 240
 241         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, sleeping");
 242
 243 noreproc:
 244
 245         obd = class_exp2obd(lock->l_conn_export);
 246
 247         /* if this is a local lock, then there is no import */
 248         if (obd != NULL)
 249                 imp = obd->u.cli.cl_import;
 250
 251         /* Wait a long time for enqueue - server may have to callback a
 252            lock from another client.  Server will evict the other client if it
 253            doesn't respond reasonably, and then give us the lock. */
 254         timeout = ldlm_get_enq_timeout(lock) * 2;
 255
 256         lwd.lwd_lock = lock;
 257
 258         if (lock->l_flags & LDLM_FL_NO_TIMEOUT) {
 259                 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
 260                 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
 261         } else {
 262                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout),
 263                                        ldlm_expired_completion_wait,
 264                                        interrupted_completion_wait, &lwd);
 265         }
 266
 267         if (imp != NULL) {
 268                 spin_lock(&imp->imp_lock);
 269                 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
 270                 spin_unlock(&imp->imp_lock);
 271         }
 272
 273         if (ns_is_client(ldlm_lock_to_ns(lock)) &&
 274             OBD_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST,
 275                                  OBD_FAIL_LDLM_CP_BL_RACE | OBD_FAIL_ONCE)) {
 276                 lock->l_flags |= LDLM_FL_FAIL_LOC;
 277                 rc = -EINTR;
 278         } else {
 279                 /* Go to sleep until the lock is granted or cancelled. */
 280                 rc = l_wait_event(lock->l_waitq,
 281                                   is_granted_or_cancelled(lock), &lwi);
 282         }
 283
 284         if (rc) {
 285                 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
 286                            rc);
 287                 return rc;
 288         }
 289
 290         return ldlm_completion_tail(lock);
 291 }
 292 EXPORT_SYMBOL(ldlm_completion_ast);
 293
 294 /**
 295  * A helper to build a blocking AST function
 296  *
 297  * Perform a common operation for blocking ASTs:
 298  * deferred lock cancellation.
 299  *
 300  * \param lock the lock blocking or canceling AST was called on
 301  * \retval 0
 302  * \see mdt_blocking_ast
 303  * \see ldlm_blocking_ast
 304  */
 305 int ldlm_blocking_ast_nocheck(struct ldlm_lock *lock)
 306 {
 307         int do_ast;
 308
 309         lock->l_flags |= LDLM_FL_CBPENDING;
 310         do_ast = !lock->l_readers && !lock->l_writers;
 311         unlock_res_and_lock(lock);
 312
 313         if (do_ast) {
 314                 struct lustre_handle lockh;
 315                 int rc;
 316
 317                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
 318                 ldlm_lock2handle(lock, &lockh);
 319                 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
 320                 if (rc < 0)
 321                         CERROR("ldlm_cli_cancel: %d\n", rc);
 322         } else {
 323                 LDLM_DEBUG(lock, "Lock still has references, will be cancelled later");
 324         }
 325         return 0;
 326 }
 327 EXPORT_SYMBOL(ldlm_blocking_ast_nocheck);
 328
 329 /**
 330  * Server blocking AST
 331  *
 332  * ->l_blocking_ast() callback for LDLM locks acquired by server-side
 333  * OBDs.
 334  *
 335  * \param lock the lock which blocks a request or cancelling lock
 336  * \param desc unused
 337  * \param data unused
 338  * \param flag indicates whether this cancelling or blocking callback
 339  * \retval 0
 340  * \see ldlm_blocking_ast_nocheck
 341  */
 342 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 343                       void *data, int flag)
 344 {
 345         if (flag == LDLM_CB_CANCELING) {
 346                 /* Don't need to do anything here. */
 347                 return 0;
 348         }
 349
 350         lock_res_and_lock(lock);
 351         /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
 352          * that ldlm_blocking_ast is called just before intent_policy method
 353          * takes the lr_lock, then by the time we get the lock, we might not
 354          * be the correct blocking function anymore.  So check, and return
 355          * early, if so. */
 356         if (lock->l_blocking_ast != ldlm_blocking_ast) {
 357                 unlock_res_and_lock(lock);
 358                 return 0;
 359         }
 360         return ldlm_blocking_ast_nocheck(lock);
 361 }
 362 EXPORT_SYMBOL(ldlm_blocking_ast);
 363
 364 /**
 365  * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See
 366  * comment in filter_intent_policy() on why you may need this.
 367  */
 368 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
 369 {
 370         /*
 371          * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for
 372          * that is rather subtle: with OST-side locking, it may so happen that
 373          * _all_ extent locks are held by the OST. If client wants to obtain
 374          * current file size it calls ll{,u}_glimpse_size(), and (as locks are
 375          * on the server), dummy glimpse callback fires and does
 376          * nothing. Client still receives correct file size due to the
 377          * following fragment in filter_intent_policy():
 378          *
 379          * rc = l->l_glimpse_ast(l, NULL); // this will update the LVB
 380          * if (rc != 0 && res->lr_namespace->ns_lvbo &&
 381          *     res->lr_namespace->ns_lvbo->lvbo_update) {
 382          *       res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1);
 383          * }
 384          *
 385          * that is, after glimpse_ast() fails, filter_lvbo_update() runs, and
 386          * returns correct file size to the client.
 387          */
 388         return -ELDLM_NO_LOCK_DATA;
 389 }
 390 EXPORT_SYMBOL(ldlm_glimpse_ast);
 391
 392 /**
 393  * Enqueue a local lock (typically on a server).
 394  */
 395 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
 396                            const struct ldlm_res_id *res_id,
 397                            ldlm_type_t type, ldlm_policy_data_t *policy,
 398                            ldlm_mode_t mode, __u64 *flags,
 399                            ldlm_blocking_callback blocking,
 400                            ldlm_completion_callback completion,
 401                            ldlm_glimpse_callback glimpse,
 402                            void *data, __u32 lvb_len, enum lvb_type lvb_type,
 403                            const __u64 *client_cookie,
 404                            struct lustre_handle *lockh)
 405 {
 406         struct ldlm_lock *lock;
 407         int err;
 408         const struct ldlm_callback_suite cbs = { .lcs_completion = completion,
 409                                                  .lcs_blocking   = blocking,
 410                                                  .lcs_glimpse    = glimpse,
 411         };
 412
 413         LASSERT(!(*flags & LDLM_FL_REPLAY));
 414         if (unlikely(ns_is_client(ns))) {
 415                 CERROR("Trying to enqueue local lock in a shadow namespace\n");
 416                 LBUG();
 417         }
 418
 419         lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len,
 420                                 lvb_type);
 421         if (unlikely(!lock)) {
 422                 err = -ENOMEM;
 423                 goto out_nolock;
 424         }
 425
 426         ldlm_lock2handle(lock, lockh);
 427
 428         /* NB: we don't have any lock now (lock_res_and_lock)
 429          * because it's a new lock */
 430         ldlm_lock_addref_internal_nolock(lock, mode);
 431         lock->l_flags |= LDLM_FL_LOCAL;
 432         if (*flags & LDLM_FL_ATOMIC_CB)
 433                 lock->l_flags |= LDLM_FL_ATOMIC_CB;
 434
 435         if (policy != NULL)
 436                 lock->l_policy_data = *policy;
 437         if (client_cookie != NULL)
 438                 lock->l_client_cookie = *client_cookie;
 439         if (type == LDLM_EXTENT)
 440                 lock->l_req_extent = policy->l_extent;
 441
 442         err = ldlm_lock_enqueue(ns, &lock, policy, flags);
 443         if (unlikely(err != ELDLM_OK))
 444                 goto out;
 445
 446         if (policy != NULL)
 447                 *policy = lock->l_policy_data;
 448
 449         if (lock->l_completion_ast)
 450                 lock->l_completion_ast(lock, *flags, NULL);
 451
 452         LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
 453  out:
 454         LDLM_LOCK_RELEASE(lock);
 455  out_nolock:
 456         return err;
 457 }
 458 EXPORT_SYMBOL(ldlm_cli_enqueue_local);
 459
 460 static void failed_lock_cleanup(struct ldlm_namespace *ns,
 461                                 struct ldlm_lock *lock, int mode)
 462 {
 463         int need_cancel = 0;
 464
 465         /* Set a flag to prevent us from sending a CANCEL (bug 407) */
 466         lock_res_and_lock(lock);
 467         /* Check that lock is not granted or failed, we might race. */
 468         if ((lock->l_req_mode != lock->l_granted_mode) &&
 469             !(lock->l_flags & LDLM_FL_FAILED)) {
 470                 /* Make sure that this lock will not be found by raced
 471                  * bl_ast and -EINVAL reply is sent to server anyways.
 472                  * bug 17645 */
 473                 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED |
 474                                  LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING;
 475                 need_cancel = 1;
 476         }
 477         unlock_res_and_lock(lock);
 478
 479         if (need_cancel)
 480                 LDLM_DEBUG(lock,
 481                            "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING");
 482         else
 483                 LDLM_DEBUG(lock, "lock was granted or failed in race");
 484
 485         ldlm_lock_decref_internal(lock, mode);
 486
 487         /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
 488          *       from llite/file.c/ll_file_flock(). */
 489         /* This code makes for the fact that we do not have blocking handler on
 490          * a client for flock locks. As such this is the place where we must
 491          * completely kill failed locks. (interrupted and those that
 492          * were waiting to be granted when server evicted us. */
 493         if (lock->l_resource->lr_type == LDLM_FLOCK) {
 494                 lock_res_and_lock(lock);
 495                 ldlm_resource_unlink_lock(lock);
 496                 ldlm_lock_destroy_nolock(lock);
 497                 unlock_res_and_lock(lock);
 498         }
 499 }
 500
 501 /**
 502  * Finishing portion of client lock enqueue code.
 503  *
 504  * Called after receiving reply from server.
 505  */
 506 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
 507                           ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
 508                           __u64 *flags, void *lvb, __u32 lvb_len,
 509                           struct lustre_handle *lockh, int rc)
 510 {
 511         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 512         int is_replay = *flags & LDLM_FL_REPLAY;
 513         struct ldlm_lock *lock;
 514         struct ldlm_reply *reply;
 515         int cleanup_phase = 1;
 516         int size = 0;
 517
 518         lock = ldlm_handle2lock(lockh);
 519         /* ldlm_cli_enqueue is holding a reference on this lock. */
 520         if (!lock) {
 521                 LASSERT(type == LDLM_FLOCK);
 522                 return -ENOLCK;
 523         }
 524
 525         LASSERTF(ergo(lvb_len != 0, lvb_len == lock->l_lvb_len),
 526                  "lvb_len = %d, l_lvb_len = %d\n", lvb_len, lock->l_lvb_len);
 527
 528         if (rc != ELDLM_OK) {
 529                 LASSERT(!is_replay);
 530                 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
 531                            rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
 532
 533                 if (rc != ELDLM_LOCK_ABORTED)
 534                         goto cleanup;
 535         }
 536
 537         /* Before we return, swab the reply */
 538         reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
 539         if (reply == NULL) {
 540                 rc = -EPROTO;
 541                 goto cleanup;
 542         }
 543
 544         if (lvb_len != 0) {
 545                 LASSERT(lvb != NULL);
 546
 547                 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB,
 548                                             RCL_SERVER);
 549                 if (size < 0) {
 550                         LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", size);
 551                         rc = size;
 552                         goto cleanup;
 553                 } else if (unlikely(size > lvb_len)) {
 554                         LDLM_ERROR(lock, "Replied LVB is larger than expectation, expected = %d, replied = %d",
 555                                    lvb_len, size);
 556                         rc = -EINVAL;
 557                         goto cleanup;
 558                 }
 559         }
 560
 561         if (rc == ELDLM_LOCK_ABORTED) {
 562                 if (lvb_len != 0)
 563                         rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
 564                                            lvb, size);
 565                 if (rc == 0)
 566                         rc = ELDLM_LOCK_ABORTED;
 567                 goto cleanup;
 568         }
 569
 570         /* lock enqueued on the server */
 571         cleanup_phase = 0;
 572
 573         lock_res_and_lock(lock);
 574         /* Key change rehash lock in per-export hash with new key */
 575         if (exp->exp_lock_hash) {
 576                 /* In the function below, .hs_keycmp resolves to
 577                  * ldlm_export_lock_keycmp() */
 578                 /* coverity[overrun-buffer-val] */
 579                 cfs_hash_rehash_key(exp->exp_lock_hash,
 580                                     &lock->l_remote_handle,
 581                                     &reply->lock_handle,
 582                                     &lock->l_exp_hash);
 583         } else {
 584                 lock->l_remote_handle = reply->lock_handle;
 585         }
 586
 587         *flags = ldlm_flags_from_wire(reply->lock_flags);
 588         lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags &
 589                                               LDLM_INHERIT_FLAGS);
 590         /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match()
 591          * to wait with no timeout as well */
 592         lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags &
 593                                               LDLM_FL_NO_TIMEOUT);
 594         unlock_res_and_lock(lock);
 595
 596         CDEBUG(D_INFO, "local: %p, remote cookie: %#llx, flags: 0x%llx\n",
 597                lock, reply->lock_handle.cookie, *flags);
 598
 599         /* If enqueue returned a blocked lock but the completion handler has
 600          * already run, then it fixed up the resource and we don't need to do it
 601          * again. */
 602         if ((*flags) & LDLM_FL_LOCK_CHANGED) {
 603                 int newmode = reply->lock_desc.l_req_mode;
 604
 605                 LASSERT(!is_replay);
 606                 if (newmode && newmode != lock->l_req_mode) {
 607                         LDLM_DEBUG(lock, "server returned different mode %s",
 608                                    ldlm_lockname[newmode]);
 609                         lock->l_req_mode = newmode;
 610                 }
 611
 612                 if (!ldlm_res_eq(&reply->lock_desc.l_resource.lr_name,
 613                                  &lock->l_resource->lr_name)) {
 614                         CDEBUG(D_INFO, "remote intent success, locking "DLDLMRES
 615                                        " instead of "DLDLMRES"\n",
 616                                PLDLMRES(&reply->lock_desc.l_resource),
 617                                PLDLMRES(lock->l_resource));
 618
 619                         rc = ldlm_lock_change_resource(ns, lock,
 620                                         &reply->lock_desc.l_resource.lr_name);
 621                         if (rc || lock->l_resource == NULL) {
 622                                 rc = -ENOMEM;
 623                                 goto cleanup;
 624                         }
 625                         LDLM_DEBUG(lock, "client-side enqueue, new resource");
 626                 }
 627                 if (with_policy)
 628                         if (!(type == LDLM_IBITS &&
 629                               !(exp_connect_flags(exp) & OBD_CONNECT_IBITS)))
 630                                 /* We assume lock type cannot change on server*/
 631                                 ldlm_convert_policy_to_local(exp,
 632                                                 lock->l_resource->lr_type,
 633                                                 &reply->lock_desc.l_policy_data,
 634                                                 &lock->l_policy_data);
 635                 if (type != LDLM_PLAIN)
 636                         LDLM_DEBUG(lock,
 637                                    "client-side enqueue, new policy data");
 638         }
 639
 640         if ((*flags) & LDLM_FL_AST_SENT ||
 641             /* Cancel extent locks as soon as possible on a liblustre client,
 642              * because it cannot handle asynchronous ASTs robustly (see
 643              * bug 7311). */
 644             (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) {
 645                 lock_res_and_lock(lock);
 646                 lock->l_flags |= LDLM_FL_CBPENDING |  LDLM_FL_BL_AST;
 647                 unlock_res_and_lock(lock);
 648                 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
 649         }
 650
 651         /* If the lock has already been granted by a completion AST, don't
 652          * clobber the LVB with an older one. */
 653         if (lvb_len != 0) {
 654                 /* We must lock or a racing completion might update lvb without
 655                  * letting us know and we'll clobber the correct value.
 656                  * Cannot unlock after the check either, a that still leaves
 657                  * a tiny window for completion to get in */
 658                 lock_res_and_lock(lock);
 659                 if (lock->l_req_mode != lock->l_granted_mode)
 660                         rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
 661                                            lock->l_lvb_data, size);
 662                 unlock_res_and_lock(lock);
 663                 if (rc < 0) {
 664                         cleanup_phase = 1;
 665                         goto cleanup;
 666                 }
 667         }
 668
 669         if (!is_replay) {
 670                 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
 671                 if (lock->l_completion_ast != NULL) {
 672                         int err = lock->l_completion_ast(lock, *flags, NULL);
 673
 674                         if (!rc)
 675                                 rc = err;
 676                         if (rc)
 677                                 cleanup_phase = 1;
 678                 }
 679         }
 680
 681         if (lvb_len && lvb != NULL) {
 682                 /* Copy the LVB here, and not earlier, because the completion
 683                  * AST (if any) can override what we got in the reply */
 684                 memcpy(lvb, lock->l_lvb_data, lvb_len);
 685         }
 686
 687         LDLM_DEBUG(lock, "client-side enqueue END");
 688 cleanup:
 689         if (cleanup_phase == 1 && rc)
 690                 failed_lock_cleanup(ns, lock, mode);
 691         /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
 692         LDLM_LOCK_PUT(lock);
 693         LDLM_LOCK_RELEASE(lock);
 694         return rc;
 695 }
 696 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
 697
 698 /**
 699  * Estimate number of lock handles that would fit into request of given
 700  * size.  PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
 701  * a single page on the send/receive side. XXX: 512 should be changed to
 702  * more adequate value.
 703  */
 704 static inline int ldlm_req_handles_avail(int req_size, int off)
 705 {
 706         int avail;
 707
 708         avail = min_t(int, LDLM_MAXREQSIZE, PAGE_CACHE_SIZE - 512) - req_size;
 709         if (likely(avail >= 0))
 710                 avail /= (int)sizeof(struct lustre_handle);
 711         else
 712                 avail = 0;
 713         avail += LDLM_LOCKREQ_HANDLES - off;
 714
 715         return avail;
 716 }
 717
 718 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
 719                                              enum req_location loc,
 720                                              int off)
 721 {
 722         int size = req_capsule_msg_size(pill, loc);
 723
 724         return ldlm_req_handles_avail(size, off);
 725 }
 726
 727 static inline int ldlm_format_handles_avail(struct obd_import *imp,
 728                                             const struct req_format *fmt,
 729                                             enum req_location loc, int off)
 730 {
 731         int size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
 732
 733         return ldlm_req_handles_avail(size, off);
 734 }
 735
 736 /**
 737  * Cancel LRU locks and pack them into the enqueue request. Pack there the given
 738  * \a count locks in \a cancels.
 739  *
 740  * This is to be called by functions preparing their own requests that
 741  * might contain lists of locks to cancel in addition to actual operation
 742  * that needs to be performed.
 743  */
 744 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
 745                       int version, int opc, int canceloff,
 746                       struct list_head *cancels, int count)
 747 {
 748         struct ldlm_namespace   *ns = exp->exp_obd->obd_namespace;
 749         struct req_capsule      *pill = &req->rq_pill;
 750         struct ldlm_request     *dlm = NULL;
 751         int flags, avail, to_free, pack = 0;
 752         LIST_HEAD(head);
 753         int rc;
 754
 755         if (cancels == NULL)
 756                 cancels = &head;
 757         if (ns_connect_cancelset(ns)) {
 758                 /* Estimate the amount of available space in the request. */
 759                 req_capsule_filled_sizes(pill, RCL_CLIENT);
 760                 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
 761
 762                 flags = ns_connect_lru_resize(ns) ?
 763                         LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
 764                 to_free = !ns_connect_lru_resize(ns) &&
 765                           opc == LDLM_ENQUEUE ? 1 : 0;
 766
 767                 /* Cancel LRU locks here _only_ if the server supports
 768                  * EARLY_CANCEL. Otherwise we have to send extra CANCEL
 769                  * RPC, which will make us slower. */
 770                 if (avail > count)
 771                         count += ldlm_cancel_lru_local(ns, cancels, to_free,
 772                                                        avail - count, 0, flags);
 773                 if (avail > count)
 774                         pack = count;
 775                 else
 776                         pack = avail;
 777                 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
 778                                      ldlm_request_bufsize(pack, opc));
 779         }
 780
 781         rc = ptlrpc_request_pack(req, version, opc);
 782         if (rc) {
 783                 ldlm_lock_list_put(cancels, l_bl_ast, count);
 784                 return rc;
 785         }
 786
 787         if (ns_connect_cancelset(ns)) {
 788                 if (canceloff) {
 789                         dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
 790                         LASSERT(dlm);
 791                         /* Skip first lock handler in ldlm_request_pack(),
 792                          * this method will increment @lock_count according
 793                          * to the lock handle amount actually written to
 794                          * the buffer. */
 795                         dlm->lock_count = canceloff;
 796                 }
 797                 /* Pack into the request @pack lock handles. */
 798                 ldlm_cli_cancel_list(cancels, pack, req, 0);
 799                 /* Prepare and send separate cancel RPC for others. */
 800                 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
 801         } else {
 802                 ldlm_lock_list_put(cancels, l_bl_ast, count);
 803         }
 804         return 0;
 805 }
 806 EXPORT_SYMBOL(ldlm_prep_elc_req);
 807
 808 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
 809                           struct list_head *cancels, int count)
 810 {
 811         return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
 812                                  LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
 813 }
 814 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
 815
 816 struct ptlrpc_request *ldlm_enqueue_pack(struct obd_export *exp, int lvb_len)
 817 {
 818         struct ptlrpc_request *req;
 819         int rc;
 820
 821         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LDLM_ENQUEUE);
 822         if (req == NULL)
 823                 return ERR_PTR(-ENOMEM);
 824
 825         rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
 826         if (rc) {
 827                 ptlrpc_request_free(req);
 828                 return ERR_PTR(rc);
 829         }
 830
 831         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, lvb_len);
 832         ptlrpc_request_set_replen(req);
 833         return req;
 834 }
 835 EXPORT_SYMBOL(ldlm_enqueue_pack);
 836
 837 /**
 838  * Client-side lock enqueue.
 839  *
 840  * If a request has some specific initialisation it is passed in \a reqp,
 841  * otherwise it is created in ldlm_cli_enqueue.
 842  *
 843  * Supports sync and async requests, pass \a async flag accordingly. If a
 844  * request was created in ldlm_cli_enqueue and it is the async request,
 845  * pass it to the caller in \a reqp.
 846  */
 847 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
 848                      struct ldlm_enqueue_info *einfo,
 849                      const struct ldlm_res_id *res_id,
 850                      ldlm_policy_data_t const *policy, __u64 *flags,
 851                      void *lvb, __u32 lvb_len, enum lvb_type lvb_type,
 852                      struct lustre_handle *lockh, int async)
 853 {
 854         struct ldlm_namespace *ns;
 855         struct ldlm_lock      *lock;
 856         struct ldlm_request   *body;
 857         int                 is_replay = *flags & LDLM_FL_REPLAY;
 858         int                 req_passed_in = 1;
 859         int                 rc, err;
 860         struct ptlrpc_request *req;
 861
 862         LASSERT(exp != NULL);
 863
 864         ns = exp->exp_obd->obd_namespace;
 865
 866         /* If we're replaying this lock, just check some invariants.
 867          * If we're creating a new lock, get everything all setup nice. */
 868         if (is_replay) {
 869                 lock = ldlm_handle2lock_long(lockh, 0);
 870                 LASSERT(lock != NULL);
 871                 LDLM_DEBUG(lock, "client-side enqueue START");
 872                 LASSERT(exp == lock->l_conn_export);
 873         } else {
 874                 const struct ldlm_callback_suite cbs = {
 875                         .lcs_completion = einfo->ei_cb_cp,
 876                         .lcs_blocking   = einfo->ei_cb_bl,
 877                         .lcs_glimpse    = einfo->ei_cb_gl
 878                 };
 879                 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
 880                                         einfo->ei_mode, &cbs, einfo->ei_cbdata,
 881                                         lvb_len, lvb_type);
 882                 if (lock == NULL)
 883                         return -ENOMEM;
 884                 /* for the local lock, add the reference */
 885                 ldlm_lock_addref_internal(lock, einfo->ei_mode);
 886                 ldlm_lock2handle(lock, lockh);
 887                 if (policy != NULL)
 888                                 lock->l_policy_data = *policy;
 889
 890                 if (einfo->ei_type == LDLM_EXTENT)
 891                         lock->l_req_extent = policy->l_extent;
 892                 LDLM_DEBUG(lock, "client-side enqueue START, flags %llx\n",
 893                            *flags);
 894         }
 895
 896         lock->l_conn_export = exp;
 897         lock->l_export = NULL;
 898         lock->l_blocking_ast = einfo->ei_cb_bl;
 899         lock->l_flags |= (*flags & (LDLM_FL_NO_LRU | LDLM_FL_EXCL));
 900
 901         /* lock not sent to server yet */
 902
 903         if (reqp == NULL || *reqp == NULL) {
 904                 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
 905                                                 &RQF_LDLM_ENQUEUE,
 906                                                 LUSTRE_DLM_VERSION,
 907                                                 LDLM_ENQUEUE);
 908                 if (req == NULL) {
 909                         failed_lock_cleanup(ns, lock, einfo->ei_mode);
 910                         LDLM_LOCK_RELEASE(lock);
 911                         return -ENOMEM;
 912                 }
 913                 req_passed_in = 0;
 914                 if (reqp)
 915                         *reqp = req;
 916         } else {
 917                 int len;
 918
 919                 req = *reqp;
 920                 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
 921                                            RCL_CLIENT);
 922                 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
 923                          DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
 924         }
 925
 926         /* Dump lock data into the request buffer */
 927         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
 928         ldlm_lock2desc(lock, &body->lock_desc);
 929         body->lock_flags = ldlm_flags_to_wire(*flags);
 930         body->lock_handle[0] = *lockh;
 931
 932         /* Continue as normal. */
 933         if (!req_passed_in) {
 934                 if (lvb_len > 0)
 935                         req_capsule_extend(&req->rq_pill,
 936                                            &RQF_LDLM_ENQUEUE_LVB);
 937                 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
 938                                      lvb_len);
 939                 ptlrpc_request_set_replen(req);
 940         }
 941
 942         /*
 943          * Liblustre client doesn't get extent locks, except for O_APPEND case
 944          * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
 945          * [i_size, OBD_OBJECT_EOF] lock is taken.
 946          */
 947         LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT ||
 948                      policy->l_extent.end == OBD_OBJECT_EOF));
 949
 950         if (async) {
 951                 LASSERT(reqp != NULL);
 952                 return 0;
 953         }
 954
 955         LDLM_DEBUG(lock, "sending request");
 956
 957         rc = ptlrpc_queue_wait(req);
 958
 959         err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
 960                                     einfo->ei_mode, flags, lvb, lvb_len,
 961                                     lockh, rc);
 962
 963         /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
 964          * one reference that we took */
 965         if (err == -ENOLCK)
 966                 LDLM_LOCK_RELEASE(lock);
 967         else
 968                 rc = err;
 969
 970         if (!req_passed_in && req != NULL) {
 971                 ptlrpc_req_finished(req);
 972                 if (reqp)
 973                         *reqp = NULL;
 974         }
 975
 976         return rc;
 977 }
 978 EXPORT_SYMBOL(ldlm_cli_enqueue);
 979
 980 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
 981                                   __u32 *flags)
 982 {
 983         struct ldlm_resource *res;
 984         int rc;
 985
 986         if (ns_is_client(ldlm_lock_to_ns(lock))) {
 987                 CERROR("Trying to cancel local lock\n");
 988                 LBUG();
 989         }
 990         LDLM_DEBUG(lock, "client-side local convert");
 991
 992         res = ldlm_lock_convert(lock, new_mode, flags);
 993         if (res) {
 994                 ldlm_reprocess_all(res);
 995                 rc = 0;
 996         } else {
 997                 rc = LUSTRE_EDEADLK;
 998         }
 999         LDLM_DEBUG(lock, "client-side local convert handler END");
1000         LDLM_LOCK_PUT(lock);
1001         return rc;
1002 }
1003
1004 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
1005  * conversion of locks which are on the waiting or converting queue */
1006 /* Caller of this code is supposed to take care of lock readers/writers
1007    accounting */
1008 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags)
1009 {
1010         struct ldlm_request   *body;
1011         struct ldlm_reply     *reply;
1012         struct ldlm_lock      *lock;
1013         struct ldlm_resource  *res;
1014         struct ptlrpc_request *req;
1015         int                 rc;
1016
1017         lock = ldlm_handle2lock(lockh);
1018         if (!lock) {
1019                 LBUG();
1020                 return -EINVAL;
1021         }
1022         *flags = 0;
1023
1024         if (lock->l_conn_export == NULL)
1025                 return ldlm_cli_convert_local(lock, new_mode, flags);
1026
1027         LDLM_DEBUG(lock, "client-side convert");
1028
1029         req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export),
1030                                         &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
1031                                         LDLM_CONVERT);
1032         if (req == NULL) {
1033                 LDLM_LOCK_PUT(lock);
1034                 return -ENOMEM;
1035         }
1036
1037         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1038         body->lock_handle[0] = lock->l_remote_handle;
1039
1040         body->lock_desc.l_req_mode = new_mode;
1041         body->lock_flags = ldlm_flags_to_wire(*flags);
1042
1043
1044         ptlrpc_request_set_replen(req);
1045         rc = ptlrpc_queue_wait(req);
1046         if (rc != ELDLM_OK)
1047                 goto out;
1048
1049         reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1050         if (reply == NULL) {
1051                 rc = -EPROTO;
1052                 goto out;
1053         }
1054
1055         if (req->rq_status) {
1056                 rc = req->rq_status;
1057                 goto out;
1058         }
1059
1060         res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
1061         if (res != NULL) {
1062                 ldlm_reprocess_all(res);
1063                 /* Go to sleep until the lock is granted. */
1064                 /* FIXME: or cancelled. */
1065                 if (lock->l_completion_ast) {
1066                         rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
1067                                                     NULL);
1068                         if (rc)
1069                                 goto out;
1070                 }
1071         } else {
1072                 rc = LUSTRE_EDEADLK;
1073         }
1074  out:
1075         LDLM_LOCK_PUT(lock);
1076         ptlrpc_req_finished(req);
1077         return rc;
1078 }
1079 EXPORT_SYMBOL(ldlm_cli_convert);
1080
1081 /**
1082  * Cancel locks locally.
1083  * Returns:
1084  * \retval LDLM_FL_LOCAL_ONLY if there is no need for a CANCEL RPC to the server
1085  * \retval LDLM_FL_CANCELING otherwise;
1086  * \retval LDLM_FL_BL_AST if there is a need for a separate CANCEL RPC.
1087  */
1088 static __u64 ldlm_cli_cancel_local(struct ldlm_lock *lock)
1089 {
1090         __u64 rc = LDLM_FL_LOCAL_ONLY;
1091
1092         if (lock->l_conn_export) {
1093                 bool local_only;
1094
1095                 LDLM_DEBUG(lock, "client-side cancel");
1096                 /* Set this flag to prevent others from getting new references*/
1097                 lock_res_and_lock(lock);
1098                 lock->l_flags |= LDLM_FL_CBPENDING;
1099                 local_only = !!(lock->l_flags &
1100                                 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
1101                 ldlm_cancel_callback(lock);
1102                 rc = (lock->l_flags & LDLM_FL_BL_AST) ?
1103                         LDLM_FL_BL_AST : LDLM_FL_CANCELING;
1104                 unlock_res_and_lock(lock);
1105
1106                 if (local_only) {
1107                         CDEBUG(D_DLMTRACE, "not sending request (at caller's instruction)\n");
1108                         rc = LDLM_FL_LOCAL_ONLY;
1109                 }
1110                 ldlm_lock_cancel(lock);
1111         } else {
1112                 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1113                         LDLM_ERROR(lock, "Trying to cancel local lock");
1114                         LBUG();
1115                 }
1116                 LDLM_DEBUG(lock, "server-side local cancel");
1117                 ldlm_lock_cancel(lock);
1118                 ldlm_reprocess_all(lock->l_resource);
1119         }
1120
1121         return rc;
1122 }
1123
1124 /**
1125  * Pack \a count locks in \a head into ldlm_request buffer of request \a req.
1126  */
1127 static void ldlm_cancel_pack(struct ptlrpc_request *req,
1128                              struct list_head *head, int count)
1129 {
1130         struct ldlm_request *dlm;
1131         struct ldlm_lock *lock;
1132         int max, packed = 0;
1133
1134         dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1135         LASSERT(dlm != NULL);
1136
1137         /* Check the room in the request buffer. */
1138         max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
1139                 sizeof(struct ldlm_request);
1140         max /= sizeof(struct lustre_handle);
1141         max += LDLM_LOCKREQ_HANDLES;
1142         LASSERT(max >= dlm->lock_count + count);
1143
1144         /* XXX: it would be better to pack lock handles grouped by resource.
1145          * so that the server cancel would call filter_lvbo_update() less
1146          * frequently. */
1147         list_for_each_entry(lock, head, l_bl_ast) {
1148                 if (!count--)
1149                         break;
1150                 LASSERT(lock->l_conn_export);
1151                 /* Pack the lock handle to the given request buffer. */
1152                 LDLM_DEBUG(lock, "packing");
1153                 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
1154                 packed++;
1155         }
1156         CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
1157 }
1158
1159 /**
1160  * Prepare and send a batched cancel RPC. It will include \a count lock
1161  * handles of locks given in \a cancels list. */
1162 int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
1163                         int count, ldlm_cancel_flags_t flags)
1164 {
1165         struct ptlrpc_request *req = NULL;
1166         struct obd_import *imp;
1167         int free, sent = 0;
1168         int rc = 0;
1169
1170         LASSERT(exp != NULL);
1171         LASSERT(count > 0);
1172
1173         CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, cfs_fail_val);
1174
1175         if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
1176                 return count;
1177
1178         free = ldlm_format_handles_avail(class_exp2cliimp(exp),
1179                                          &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
1180         if (count > free)
1181                 count = free;
1182
1183         while (1) {
1184                 imp = class_exp2cliimp(exp);
1185                 if (imp == NULL || imp->imp_invalid) {
1186                         CDEBUG(D_DLMTRACE,
1187                                "skipping cancel on invalid import %p\n", imp);
1188                         return count;
1189                 }
1190
1191                 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
1192                 if (req == NULL) {
1193                         rc = -ENOMEM;
1194                         goto out;
1195                 }
1196
1197                 req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
1198                 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
1199                                      ldlm_request_bufsize(count, LDLM_CANCEL));
1200
1201                 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
1202                 if (rc) {
1203                         ptlrpc_request_free(req);
1204                         goto out;
1205                 }
1206
1207                 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1208                 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1209                 ptlrpc_at_set_req_timeout(req);
1210
1211                 ldlm_cancel_pack(req, cancels, count);
1212
1213                 ptlrpc_request_set_replen(req);
1214                 if (flags & LCF_ASYNC) {
1215                         ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
1216                         sent = count;
1217                         goto out;
1218                 } else {
1219                         rc = ptlrpc_queue_wait(req);
1220                 }
1221                 if (rc == LUSTRE_ESTALE) {
1222                         CDEBUG(D_DLMTRACE, "client/server (nid %s) out of sync -- not fatal\n",
1223                                libcfs_nid2str(req->rq_import->
1224                                               imp_connection->c_peer.nid));
1225                         rc = 0;
1226                 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1227                            req->rq_import_generation == imp->imp_generation) {
1228                         ptlrpc_req_finished(req);
1229                         continue;
1230                 } else if (rc != ELDLM_OK) {
1231                         /* -ESHUTDOWN is common on umount */
1232                         CDEBUG_LIMIT(rc == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1233                                      "Got rc %d from cancel RPC: canceling anyway\n",
1234                                      rc);
1235                         break;
1236                 }
1237                 sent = count;
1238                 break;
1239         }
1240
1241         ptlrpc_req_finished(req);
1242 out:
1243         return sent ? sent : rc;
1244 }
1245 EXPORT_SYMBOL(ldlm_cli_cancel_req);
1246
1247 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1248 {
1249         LASSERT(imp != NULL);
1250         return &imp->imp_obd->obd_namespace->ns_pool;
1251 }
1252
1253 /**
1254  * Update client's OBD pool related fields with new SLV and Limit from \a req.
1255  */
1256 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1257 {
1258         struct obd_device *obd;
1259         __u64 new_slv;
1260         __u32 new_limit;
1261
1262         if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1263                      !imp_connect_lru_resize(req->rq_import))) {
1264                 /*
1265                  * Do nothing for corner cases.
1266                  */
1267                 return 0;
1268         }
1269
1270         /* In some cases RPC may contain SLV and limit zeroed out. This
1271          * is the case when server does not support LRU resize feature.
1272          * This is also possible in some recovery cases when server-side
1273          * reqs have no reference to the OBD export and thus access to
1274          * server-side namespace is not possible. */
1275         if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1276             lustre_msg_get_limit(req->rq_repmsg) == 0) {
1277                 DEBUG_REQ(D_HA, req,
1278                           "Zero SLV or Limit found (SLV: %llu, Limit: %u)",
1279                           lustre_msg_get_slv(req->rq_repmsg),
1280                           lustre_msg_get_limit(req->rq_repmsg));
1281                 return 0;
1282         }
1283
1284         new_limit = lustre_msg_get_limit(req->rq_repmsg);
1285         new_slv = lustre_msg_get_slv(req->rq_repmsg);
1286         obd = req->rq_import->imp_obd;
1287
1288         /* Set new SLV and limit in OBD fields to make them accessible
1289          * to the pool thread. We do not access obd_namespace and pool
1290          * directly here as there is no reliable way to make sure that
1291          * they are still alive at cleanup time. Evil races are possible
1292          * which may cause Oops at that time. */
1293         write_lock(&obd->obd_pool_lock);
1294         obd->obd_pool_slv = new_slv;
1295         obd->obd_pool_limit = new_limit;
1296         write_unlock(&obd->obd_pool_lock);
1297
1298         return 0;
1299 }
1300 EXPORT_SYMBOL(ldlm_cli_update_pool);
1301
1302 /**
1303  * Client side lock cancel.
1304  *
1305  * Lock must not have any readers or writers by this time.
1306  */
1307 int ldlm_cli_cancel(struct lustre_handle *lockh,
1308                     ldlm_cancel_flags_t cancel_flags)
1309 {
1310         struct obd_export *exp;
1311         int avail, flags, count = 1;
1312         __u64 rc = 0;
1313         struct ldlm_namespace *ns;
1314         struct ldlm_lock *lock;
1315         LIST_HEAD(cancels);
1316
1317         /* concurrent cancels on the same handle can happen */
1318         lock = ldlm_handle2lock_long(lockh, LDLM_FL_CANCELING);
1319         if (lock == NULL) {
1320                 LDLM_DEBUG_NOLOCK("lock is already being destroyed\n");
1321                 return 0;
1322         }
1323
1324         rc = ldlm_cli_cancel_local(lock);
1325         if (rc == LDLM_FL_LOCAL_ONLY || cancel_flags & LCF_LOCAL) {
1326                 LDLM_LOCK_RELEASE(lock);
1327                 return 0;
1328         }
1329         /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1330          * RPC which goes to canceld portal, so we can cancel other LRU locks
1331          * here and send them all as one LDLM_CANCEL RPC. */
1332         LASSERT(list_empty(&lock->l_bl_ast));
1333         list_add(&lock->l_bl_ast, &cancels);
1334
1335         exp = lock->l_conn_export;
1336         if (exp_connect_cancelset(exp)) {
1337                 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1338                                                   &RQF_LDLM_CANCEL,
1339                                                   RCL_CLIENT, 0);
1340                 LASSERT(avail > 0);
1341
1342                 ns = ldlm_lock_to_ns(lock);
1343                 flags = ns_connect_lru_resize(ns) ?
1344                         LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
1345                 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1346                                                LCF_BL_AST, flags);
1347         }
1348         ldlm_cli_cancel_list(&cancels, count, NULL, cancel_flags);
1349         return 0;
1350 }
1351 EXPORT_SYMBOL(ldlm_cli_cancel);
1352
1353 /**
1354  * Locally cancel up to \a count locks in list \a cancels.
1355  * Return the number of cancelled locks.
1356  */
1357 int ldlm_cli_cancel_list_local(struct list_head *cancels, int count,
1358                                ldlm_cancel_flags_t flags)
1359 {
1360         LIST_HEAD(head);
1361         struct ldlm_lock *lock, *next;
1362         int left = 0, bl_ast = 0;
1363         __u64 rc;
1364
1365         left = count;
1366         list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1367                 if (left-- == 0)
1368                         break;
1369
1370                 if (flags & LCF_LOCAL) {
1371                         rc = LDLM_FL_LOCAL_ONLY;
1372                         ldlm_lock_cancel(lock);
1373                 } else {
1374                         rc = ldlm_cli_cancel_local(lock);
1375                 }
1376                 /* Until we have compound requests and can send LDLM_CANCEL
1377                  * requests batched with generic RPCs, we need to send cancels
1378                  * with the LDLM_FL_BL_AST flag in a separate RPC from
1379                  * the one being generated now. */
1380                 if (!(flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1381                         LDLM_DEBUG(lock, "Cancel lock separately");
1382                         list_del_init(&lock->l_bl_ast);
1383                         list_add(&lock->l_bl_ast, &head);
1384                         bl_ast++;
1385                         continue;
1386                 }
1387                 if (rc == LDLM_FL_LOCAL_ONLY) {
1388                         /* CANCEL RPC should not be sent to server. */
1389                         list_del_init(&lock->l_bl_ast);
1390                         LDLM_LOCK_RELEASE(lock);
1391                         count--;
1392                 }
1393         }
1394         if (bl_ast > 0) {
1395                 count -= bl_ast;
1396                 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1397         }
1398
1399         return count;
1400 }
1401 EXPORT_SYMBOL(ldlm_cli_cancel_list_local);
1402
1403 /**
1404  * Cancel as many locks as possible w/o sending any RPCs (e.g. to write back
1405  * dirty data, to close a file, ...) or waiting for any RPCs in-flight (e.g.
1406  * readahead requests, ...)
1407  */
1408 static ldlm_policy_res_t ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns,
1409                                                     struct ldlm_lock *lock,
1410                                                     int unused, int added,
1411                                                     int count)
1412 {
1413         ldlm_policy_res_t result = LDLM_POLICY_CANCEL_LOCK;
1414         ldlm_cancel_for_recovery cb = ns->ns_cancel_for_recovery;
1415
1416         lock_res_and_lock(lock);
1417
1418         /* don't check added & count since we want to process all locks
1419          * from unused list */
1420         switch (lock->l_resource->lr_type) {
1421         case LDLM_EXTENT:
1422         case LDLM_IBITS:
1423                 if (cb && cb(lock))
1424                         break;
1425         default:
1426                 result = LDLM_POLICY_SKIP_LOCK;
1427                 lock->l_flags |= LDLM_FL_SKIPPED;
1428                 break;
1429         }
1430
1431         unlock_res_and_lock(lock);
1432         return result;
1433 }
1434
1435 /**
1436  * Callback function for LRU-resize policy. Decides whether to keep
1437  * \a lock in LRU for current \a LRU size \a unused, added in current
1438  * scan \a added and number of locks to be preferably canceled \a count.
1439  *
1440  * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1441  *
1442  * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1443  */
1444 static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1445                                                  struct ldlm_lock *lock,
1446                                                  int unused, int added,
1447                                                  int count)
1448 {
1449         unsigned long cur = cfs_time_current();
1450         struct ldlm_pool *pl = &ns->ns_pool;
1451         __u64 slv, lvf, lv;
1452         unsigned long la;
1453
1454         /* Stop LRU processing when we reach past @count or have checked all
1455          * locks in LRU. */
1456         if (count && added >= count)
1457                 return LDLM_POLICY_KEEP_LOCK;
1458
1459         slv = ldlm_pool_get_slv(pl);
1460         lvf = ldlm_pool_get_lvf(pl);
1461         la = cfs_duration_sec(cfs_time_sub(cur,
1462                               lock->l_last_used));
1463         lv = lvf * la * unused;
1464
1465         /* Inform pool about current CLV to see it via proc. */
1466         ldlm_pool_set_clv(pl, lv);
1467
1468         /* Stop when SLV is not yet come from server or lv is smaller than
1469          * it is. */
1470         return (slv == 0 || lv < slv) ?
1471                 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1472 }
1473
1474 /**
1475  * Callback function for proc used policy. Makes decision whether to keep
1476  * \a lock in LRU for current \a LRU size \a unused, added in current scan \a
1477  * added and number of locks to be preferably canceled \a count.
1478  *
1479  * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1480  *
1481  * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1482  */
1483 static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1484                                                    struct ldlm_lock *lock,
1485                                                    int unused, int added,
1486                                                    int count)
1487 {
1488         /* Stop LRU processing when we reach past @count or have checked all
1489          * locks in LRU. */
1490         return (added >= count) ?
1491                 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1492 }
1493
1494 /**
1495  * Callback function for aged policy. Makes decision whether to keep \a lock in
1496  * LRU for current LRU size \a unused, added in current scan \a added and
1497  * number of locks to be preferably canceled \a count.
1498  *
1499  * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1500  *
1501  * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1502  */
1503 static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1504                                                  struct ldlm_lock *lock,
1505                                                  int unused, int added,
1506                                                  int count)
1507 {
1508         /* Stop LRU processing if young lock is found and we reach past count */
1509         return ((added >= count) &&
1510                 time_before(cfs_time_current(),
1511                             cfs_time_add(lock->l_last_used, ns->ns_max_age))) ?
1512                 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1513 }
1514
1515 /**
1516  * Callback function for default policy. Makes decision whether to keep \a lock
1517  * in LRU for current LRU size \a unused, added in current scan \a added and
1518  * number of locks to be preferably canceled \a count.
1519  *
1520  * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1521  *
1522  * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1523  */
1524 static ldlm_policy_res_t ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1525                                                     struct ldlm_lock *lock,
1526                                                     int unused, int added,
1527                                                     int count)
1528 {
1529         /* Stop LRU processing when we reach past count or have checked all
1530          * locks in LRU. */
1531         return (added >= count) ?
1532                 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1533 }
1534
1535 typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *,
1536                                                       struct ldlm_lock *, int,
1537                                                       int, int);
1538
1539 static ldlm_cancel_lru_policy_t
1540 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
1541 {
1542         if (flags & LDLM_CANCEL_NO_WAIT)
1543                 return ldlm_cancel_no_wait_policy;
1544
1545         if (ns_connect_lru_resize(ns)) {
1546                 if (flags & LDLM_CANCEL_SHRINK)
1547                         /* We kill passed number of old locks. */
1548                         return ldlm_cancel_passed_policy;
1549                 else if (flags & LDLM_CANCEL_LRUR)
1550                         return ldlm_cancel_lrur_policy;
1551                 else if (flags & LDLM_CANCEL_PASSED)
1552                         return ldlm_cancel_passed_policy;
1553         } else {
1554                 if (flags & LDLM_CANCEL_AGED)
1555                         return ldlm_cancel_aged_policy;
1556         }
1557
1558         return ldlm_cancel_default_policy;
1559 }
1560
1561 /**
1562  * - Free space in LRU for \a count new locks,
1563  *   redundant unused locks are canceled locally;
1564  * - also cancel locally unused aged locks;
1565  * - do not cancel more than \a max locks;
1566  * - GET the found locks and add them into the \a cancels list.
1567  *
1568  * A client lock can be added to the l_bl_ast list only when it is
1569  * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing
1570  * CANCEL.  There are the following use cases:
1571  * ldlm_cancel_resource_local(), ldlm_cancel_lru_local() and
1572  * ldlm_cli_cancel(), which check and set this flag properly. As any
1573  * attempt to cancel a lock rely on this flag, l_bl_ast list is accessed
1574  * later without any special locking.
1575  *
1576  * Calling policies for enabled LRU resize:
1577  * ----------------------------------------
1578  * flags & LDLM_CANCEL_LRUR - use LRU resize policy (SLV from server) to
1579  *                          cancel not more than \a count locks;
1580  *
1581  * flags & LDLM_CANCEL_PASSED - cancel \a count number of old locks (located at
1582  *                            the beginning of LRU list);
1583  *
1584  * flags & LDLM_CANCEL_SHRINK - cancel not more than \a count locks according to
1585  *                            memory pressure policy function;
1586  *
1587  * flags & LDLM_CANCEL_AGED - cancel \a count locks according to "aged policy".
1588  *
1589  * flags & LDLM_CANCEL_NO_WAIT - cancel as many unused locks as possible
1590  *                             (typically before replaying locks) w/o
1591  *                             sending any RPCs or waiting for any
1592  *                             outstanding RPC to complete.
1593  */
1594 static int ldlm_prepare_lru_list(struct ldlm_namespace *ns,
1595                                  struct list_head *cancels, int count, int max,
1596                                  int flags)
1597 {
1598         ldlm_cancel_lru_policy_t pf;
1599         struct ldlm_lock *lock, *next;
1600         int added = 0, unused, remained;
1601
1602         spin_lock(&ns->ns_lock);
1603         unused = ns->ns_nr_unused;
1604         remained = unused;
1605
1606         if (!ns_connect_lru_resize(ns))
1607                 count += unused - ns->ns_max_unused;
1608
1609         pf = ldlm_cancel_lru_policy(ns, flags);
1610         LASSERT(pf != NULL);
1611
1612         while (!list_empty(&ns->ns_unused_list)) {
1613                 ldlm_policy_res_t result;
1614
1615                 /* all unused locks */
1616                 if (remained-- <= 0)
1617                         break;
1618
1619                 /* For any flags, stop scanning if @max is reached. */
1620                 if (max && added >= max)
1621                         break;
1622
1623                 list_for_each_entry_safe(lock, next, &ns->ns_unused_list,
1624                                              l_lru) {
1625                         /* No locks which got blocking requests. */
1626                         LASSERT(!(lock->l_flags & LDLM_FL_BL_AST));
1627
1628                         if (flags & LDLM_CANCEL_NO_WAIT &&
1629                             lock->l_flags & LDLM_FL_SKIPPED)
1630                                 /* already processed */
1631                                 continue;
1632
1633                         /* Somebody is already doing CANCEL. No need for this
1634                          * lock in LRU, do not traverse it again. */
1635                         if (!(lock->l_flags & LDLM_FL_CANCELING))
1636                                 break;
1637
1638                         ldlm_lock_remove_from_lru_nolock(lock);
1639                 }
1640                 if (&lock->l_lru == &ns->ns_unused_list)
1641                         break;
1642
1643                 LDLM_LOCK_GET(lock);
1644                 spin_unlock(&ns->ns_lock);
1645                 lu_ref_add(&lock->l_reference, __func__, current);
1646
1647                 /* Pass the lock through the policy filter and see if it
1648                  * should stay in LRU.
1649                  *
1650                  * Even for shrinker policy we stop scanning if
1651                  * we find a lock that should stay in the cache.
1652                  * We should take into account lock age anyway
1653                  * as a new lock is a valuable resource even if
1654                  * it has a low weight.
1655                  *
1656                  * That is, for shrinker policy we drop only
1657                  * old locks, but additionally choose them by
1658                  * their weight. Big extent locks will stay in
1659                  * the cache. */
1660                 result = pf(ns, lock, unused, added, count);
1661                 if (result == LDLM_POLICY_KEEP_LOCK) {
1662                         lu_ref_del(&lock->l_reference,
1663                                    __func__, current);
1664                         LDLM_LOCK_RELEASE(lock);
1665                         spin_lock(&ns->ns_lock);
1666                         break;
1667                 }
1668                 if (result == LDLM_POLICY_SKIP_LOCK) {
1669                         lu_ref_del(&lock->l_reference,
1670                                    __func__, current);
1671                         LDLM_LOCK_RELEASE(lock);
1672                         spin_lock(&ns->ns_lock);
1673                         continue;
1674                 }
1675
1676                 lock_res_and_lock(lock);
1677                 /* Check flags again under the lock. */
1678                 if ((lock->l_flags & LDLM_FL_CANCELING) ||
1679                     (ldlm_lock_remove_from_lru(lock) == 0)) {
1680                         /* Another thread is removing lock from LRU, or
1681                          * somebody is already doing CANCEL, or there
1682                          * is a blocking request which will send cancel
1683                          * by itself, or the lock is no longer unused. */
1684                         unlock_res_and_lock(lock);
1685                         lu_ref_del(&lock->l_reference,
1686                                    __func__, current);
1687                         LDLM_LOCK_RELEASE(lock);
1688                         spin_lock(&ns->ns_lock);
1689                         continue;
1690                 }
1691                 LASSERT(!lock->l_readers && !lock->l_writers);
1692
1693                 /* If we have chosen to cancel this lock voluntarily, we
1694                  * better send cancel notification to server, so that it
1695                  * frees appropriate state. This might lead to a race
1696                  * where while we are doing cancel here, server is also
1697                  * silently cancelling this lock. */
1698                 lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK;
1699
1700                 /* Setting the CBPENDING flag is a little misleading,
1701                  * but prevents an important race; namely, once
1702                  * CBPENDING is set, the lock can accumulate no more
1703                  * readers/writers. Since readers and writers are
1704                  * already zero here, ldlm_lock_decref() won't see
1705                  * this flag and call l_blocking_ast */
1706                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1707
1708                 /* We can't re-add to l_lru as it confuses the
1709                  * refcounting in ldlm_lock_remove_from_lru() if an AST
1710                  * arrives after we drop lr_lock below. We use l_bl_ast
1711                  * and can't use l_pending_chain as it is used both on
1712                  * server and client nevertheless bug 5666 says it is
1713                  * used only on server */
1714                 LASSERT(list_empty(&lock->l_bl_ast));
1715                 list_add(&lock->l_bl_ast, cancels);
1716                 unlock_res_and_lock(lock);
1717                 lu_ref_del(&lock->l_reference, __func__, current);
1718                 spin_lock(&ns->ns_lock);
1719                 added++;
1720                 unused--;
1721         }
1722         spin_unlock(&ns->ns_lock);
1723         return added;
1724 }
1725
1726 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
1727                           int count, int max, ldlm_cancel_flags_t cancel_flags,
1728                           int flags)
1729 {
1730         int added;
1731
1732         added = ldlm_prepare_lru_list(ns, cancels, count, max, flags);
1733         if (added <= 0)
1734                 return added;
1735         return ldlm_cli_cancel_list_local(cancels, added, cancel_flags);
1736 }
1737
1738 /**
1739  * Cancel at least \a nr locks from given namespace LRU.
1740  *
1741  * When called with LCF_ASYNC the blocking callback will be handled
1742  * in a thread and this function will return after the thread has been
1743  * asked to call the callback.  When called with LCF_ASYNC the blocking
1744  * callback will be performed in this function.
1745  */
1746 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr,
1747                     ldlm_cancel_flags_t cancel_flags,
1748                     int flags)
1749 {
1750         LIST_HEAD(cancels);
1751         int count, rc;
1752
1753         /* Just prepare the list of locks, do not actually cancel them yet.
1754          * Locks are cancelled later in a separate thread. */
1755         count = ldlm_prepare_lru_list(ns, &cancels, nr, 0, flags);
1756         rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, cancel_flags);
1757         if (rc == 0)
1758                 return count;
1759
1760         return 0;
1761 }
1762
1763 /**
1764  * Find and cancel locally unused locks found on resource, matched to the
1765  * given policy, mode. GET the found locks and add them into the \a cancels
1766  * list.
1767  */
1768 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1769                                struct list_head *cancels,
1770                                ldlm_policy_data_t *policy,
1771                                ldlm_mode_t mode, __u64 lock_flags,
1772                                ldlm_cancel_flags_t cancel_flags, void *opaque)
1773 {
1774         struct ldlm_lock *lock;
1775         int count = 0;
1776
1777         lock_res(res);
1778         list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1779                 if (opaque != NULL && lock->l_ast_data != opaque) {
1780                         LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1781                                    lock->l_ast_data, opaque);
1782                         continue;
1783                 }
1784
1785                 if (lock->l_readers || lock->l_writers)
1786                         continue;
1787
1788                 /* If somebody is already doing CANCEL, or blocking AST came,
1789                  * skip this lock. */
1790                 if (lock->l_flags & LDLM_FL_BL_AST ||
1791                     lock->l_flags & LDLM_FL_CANCELING)
1792                         continue;
1793
1794                 if (lockmode_compat(lock->l_granted_mode, mode))
1795                         continue;
1796
1797                 /* If policy is given and this is IBITS lock, add to list only
1798                  * those locks that match by policy. */
1799                 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1800                     !(lock->l_policy_data.l_inodebits.bits &
1801                       policy->l_inodebits.bits))
1802                         continue;
1803
1804                 /* See CBPENDING comment in ldlm_cancel_lru */
1805                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1806                                  lock_flags;
1807
1808                 LASSERT(list_empty(&lock->l_bl_ast));
1809                 list_add(&lock->l_bl_ast, cancels);
1810                 LDLM_LOCK_GET(lock);
1811                 count++;
1812         }
1813         unlock_res(res);
1814
1815         return ldlm_cli_cancel_list_local(cancels, count, cancel_flags);
1816 }
1817 EXPORT_SYMBOL(ldlm_cancel_resource_local);
1818
1819 /**
1820  * Cancel client-side locks from a list and send/prepare cancel RPCs to the
1821  * server.
1822  * If \a req is NULL, send CANCEL request to server with handles of locks
1823  * in the \a cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1824  * separately per lock.
1825  * If \a req is not NULL, put handles of locks in \a cancels into the request
1826  * buffer at the offset \a off.
1827  * Destroy \a cancels at the end.
1828  */
1829 int ldlm_cli_cancel_list(struct list_head *cancels, int count,
1830                          struct ptlrpc_request *req, ldlm_cancel_flags_t flags)
1831 {
1832         struct ldlm_lock *lock;
1833         int res = 0;
1834
1835         if (list_empty(cancels) || count == 0)
1836                 return 0;
1837
1838         /* XXX: requests (both batched and not) could be sent in parallel.
1839          * Usually it is enough to have just 1 RPC, but it is possible that
1840          * there are too many locks to be cancelled in LRU or on a resource.
1841          * It would also speed up the case when the server does not support
1842          * the feature. */
1843         while (count > 0) {
1844                 LASSERT(!list_empty(cancels));
1845                 lock = list_entry(cancels->next, struct ldlm_lock,
1846                                       l_bl_ast);
1847                 LASSERT(lock->l_conn_export);
1848
1849                 if (exp_connect_cancelset(lock->l_conn_export)) {
1850                         res = count;
1851                         if (req)
1852                                 ldlm_cancel_pack(req, cancels, count);
1853                         else
1854                                 res = ldlm_cli_cancel_req(lock->l_conn_export,
1855                                                           cancels, count,
1856                                                           flags);
1857                 } else {
1858                         res = ldlm_cli_cancel_req(lock->l_conn_export,
1859                                                   cancels, 1, flags);
1860                 }
1861
1862                 if (res < 0) {
1863                         CDEBUG_LIMIT(res == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1864                                      "ldlm_cli_cancel_list: %d\n", res);
1865                         res = count;
1866                 }
1867
1868                 count -= res;
1869                 ldlm_lock_list_put(cancels, l_bl_ast, res);
1870         }
1871         LASSERT(count == 0);
1872         return 0;
1873 }
1874 EXPORT_SYMBOL(ldlm_cli_cancel_list);
1875
1876 /**
1877  * Cancel all locks on a resource that have 0 readers/writers.
1878  *
1879  * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
1880  * to notify the server. */
1881 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1882                                     const struct ldlm_res_id *res_id,
1883                                     ldlm_policy_data_t *policy,
1884                                     ldlm_mode_t mode,
1885                                     ldlm_cancel_flags_t flags,
1886                                     void *opaque)
1887 {
1888         struct ldlm_resource *res;
1889         LIST_HEAD(cancels);
1890         int count;
1891         int rc;
1892
1893         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1894         if (res == NULL) {
1895                 /* This is not a problem. */
1896                 CDEBUG(D_INFO, "No resource %llu\n", res_id->name[0]);
1897                 return 0;
1898         }
1899
1900         LDLM_RESOURCE_ADDREF(res);
1901         count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
1902                                            0, flags | LCF_BL_AST, opaque);
1903         rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
1904         if (rc != ELDLM_OK)
1905                 CERROR("canceling unused lock "DLDLMRES": rc = %d\n",
1906                        PLDLMRES(res), rc);
1907
1908         LDLM_RESOURCE_DELREF(res);
1909         ldlm_resource_putref(res);
1910         return 0;
1911 }
1912 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
1913
1914 struct ldlm_cli_cancel_arg {
1915         int     lc_flags;
1916         void   *lc_opaque;
1917 };
1918
1919 static int ldlm_cli_hash_cancel_unused(struct cfs_hash *hs,
1920                                        struct cfs_hash_bd *bd,
1921                                        struct hlist_node *hnode, void *arg)
1922 {
1923         struct ldlm_resource       *res = cfs_hash_object(hs, hnode);
1924         struct ldlm_cli_cancel_arg     *lc = arg;
1925
1926         ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name,
1927                                         NULL, LCK_MINMODE,
1928                                         lc->lc_flags, lc->lc_opaque);
1929         /* must return 0 for hash iteration */
1930         return 0;
1931 }
1932
1933 /**
1934  * Cancel all locks on a namespace (or a specific resource, if given)
1935  * that have 0 readers/writers.
1936  *
1937  * If flags & LCF_LOCAL, throw the locks away without trying
1938  * to notify the server. */
1939 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1940                            const struct ldlm_res_id *res_id,
1941                            ldlm_cancel_flags_t flags, void *opaque)
1942 {
1943         struct ldlm_cli_cancel_arg arg = {
1944                 .lc_flags       = flags,
1945                 .lc_opaque      = opaque,
1946         };
1947
1948         if (ns == NULL)
1949                 return ELDLM_OK;
1950
1951         if (res_id != NULL) {
1952                 return ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
1953                                                        LCK_MINMODE, flags,
1954                                                        opaque);
1955         } else {
1956                 cfs_hash_for_each_nolock(ns->ns_rs_hash,
1957                                          ldlm_cli_hash_cancel_unused, &arg);
1958                 return ELDLM_OK;
1959         }
1960 }
1961 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
1962
1963 /* Lock iterators. */
1964
1965 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
1966                           void *closure)
1967 {
1968         struct list_head *tmp, *next;
1969         struct ldlm_lock *lock;
1970         int rc = LDLM_ITER_CONTINUE;
1971
1972         if (!res)
1973                 return LDLM_ITER_CONTINUE;
1974
1975         lock_res(res);
1976         list_for_each_safe(tmp, next, &res->lr_granted) {
1977                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1978
1979                 if (iter(lock, closure) == LDLM_ITER_STOP) {
1980                         rc = LDLM_ITER_STOP;
1981                         goto out;
1982                 }
1983         }
1984
1985         list_for_each_safe(tmp, next, &res->lr_converting) {
1986                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1987
1988                 if (iter(lock, closure) == LDLM_ITER_STOP) {
1989                         rc = LDLM_ITER_STOP;
1990                         goto out;
1991                 }
1992         }
1993
1994         list_for_each_safe(tmp, next, &res->lr_waiting) {
1995                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1996
1997                 if (iter(lock, closure) == LDLM_ITER_STOP) {
1998                         rc = LDLM_ITER_STOP;
1999                         goto out;
2000                 }
2001         }
2002  out:
2003         unlock_res(res);
2004         return rc;
2005 }
2006 EXPORT_SYMBOL(ldlm_resource_foreach);
2007
2008 struct iter_helper_data {
2009         ldlm_iterator_t iter;
2010         void *closure;
2011 };
2012
2013 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
2014 {
2015         struct iter_helper_data *helper = closure;
2016
2017         return helper->iter(lock, helper->closure);
2018 }
2019
2020 static int ldlm_res_iter_helper(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2021                                 struct hlist_node *hnode, void *arg)
2022
2023 {
2024         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2025
2026         return ldlm_resource_foreach(res, ldlm_iter_helper, arg) ==
2027                LDLM_ITER_STOP;
2028 }
2029
2030 void ldlm_namespace_foreach(struct ldlm_namespace *ns,
2031                             ldlm_iterator_t iter, void *closure)
2032
2033 {
2034         struct iter_helper_data helper = {
2035                 .iter           = iter,
2036                 .closure        = closure,
2037         };
2038
2039         cfs_hash_for_each_nolock(ns->ns_rs_hash,
2040                                  ldlm_res_iter_helper, &helper);
2041
2042 }
2043 EXPORT_SYMBOL(ldlm_namespace_foreach);
2044
2045 /* non-blocking function to manipulate a lock whose cb_data is being put away.
2046  * return  0:  find no resource
2047  *       > 0:  must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE.
2048  *       < 0:  errors
2049  */
2050 int ldlm_resource_iterate(struct ldlm_namespace *ns,
2051                           const struct ldlm_res_id *res_id,
2052                           ldlm_iterator_t iter, void *data)
2053 {
2054         struct ldlm_resource *res;
2055         int rc;
2056
2057         if (ns == NULL) {
2058                 CERROR("must pass in namespace\n");
2059                 LBUG();
2060         }
2061
2062         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
2063         if (res == NULL)
2064                 return 0;
2065
2066         LDLM_RESOURCE_ADDREF(res);
2067         rc = ldlm_resource_foreach(res, iter, data);
2068         LDLM_RESOURCE_DELREF(res);
2069         ldlm_resource_putref(res);
2070         return rc;
2071 }
2072 EXPORT_SYMBOL(ldlm_resource_iterate);
2073
2074 /* Lock replay */
2075
2076 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
2077 {
2078         struct list_head *list = closure;
2079
2080         /* we use l_pending_chain here, because it's unused on clients. */
2081         LASSERTF(list_empty(&lock->l_pending_chain),
2082                  "lock %p next %p prev %p\n",
2083                  lock, &lock->l_pending_chain.next,
2084                  &lock->l_pending_chain.prev);
2085         /* bug 9573: don't replay locks left after eviction, or
2086          * bug 17614: locks being actively cancelled. Get a reference
2087          * on a lock so that it does not disappear under us (e.g. due to cancel)
2088          */
2089         if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_CANCELING))) {
2090                 list_add(&lock->l_pending_chain, list);
2091                 LDLM_LOCK_GET(lock);
2092         }
2093
2094         return LDLM_ITER_CONTINUE;
2095 }
2096
2097 static int replay_lock_interpret(const struct lu_env *env,
2098                                  struct ptlrpc_request *req,
2099                                  struct ldlm_async_args *aa, int rc)
2100 {
2101         struct ldlm_lock     *lock;
2102         struct ldlm_reply    *reply;
2103         struct obd_export    *exp;
2104
2105         atomic_dec(&req->rq_import->imp_replay_inflight);
2106         if (rc != ELDLM_OK)
2107                 goto out;
2108
2109
2110         reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2111         if (reply == NULL) {
2112                 rc = -EPROTO;
2113                 goto out;
2114         }
2115
2116         lock = ldlm_handle2lock(&aa->lock_handle);
2117         if (!lock) {
2118                 CERROR("received replay ack for unknown local cookie %#llx remote cookie %#llx from server %s id %s\n",
2119                        aa->lock_handle.cookie, reply->lock_handle.cookie,
2120                        req->rq_export->exp_client_uuid.uuid,
2121                        libcfs_id2str(req->rq_peer));
2122                 rc = -ESTALE;
2123                 goto out;
2124         }
2125
2126         /* Key change rehash lock in per-export hash with new key */
2127         exp = req->rq_export;
2128         if (exp && exp->exp_lock_hash) {
2129                 /* In the function below, .hs_keycmp resolves to
2130                  * ldlm_export_lock_keycmp() */
2131                 /* coverity[overrun-buffer-val] */
2132                 cfs_hash_rehash_key(exp->exp_lock_hash,
2133                                     &lock->l_remote_handle,
2134                                     &reply->lock_handle,
2135                                     &lock->l_exp_hash);
2136         } else {
2137                 lock->l_remote_handle = reply->lock_handle;
2138         }
2139
2140         LDLM_DEBUG(lock, "replayed lock:");
2141         ptlrpc_import_recovery_state_machine(req->rq_import);
2142         LDLM_LOCK_PUT(lock);
2143 out:
2144         if (rc != ELDLM_OK)
2145                 ptlrpc_connect_import(req->rq_import);
2146
2147         return rc;
2148 }
2149
2150 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2151 {
2152         struct ptlrpc_request *req;
2153         struct ldlm_async_args *aa;
2154         struct ldlm_request   *body;
2155         int flags;
2156
2157         /* Bug 11974: Do not replay a lock which is actively being canceled */
2158         if (lock->l_flags & LDLM_FL_CANCELING) {
2159                 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2160                 return 0;
2161         }
2162
2163         /* If this is reply-less callback lock, we cannot replay it, since
2164          * server might have long dropped it, but notification of that event was
2165          * lost by network. (and server granted conflicting lock already) */
2166         if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
2167                 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2168                 ldlm_lock_cancel(lock);
2169                 return 0;
2170         }
2171
2172         /*
2173          * If granted mode matches the requested mode, this lock is granted.
2174          *
2175          * If they differ, but we have a granted mode, then we were granted
2176          * one mode and now want another: ergo, converting.
2177          *
2178          * If we haven't been granted anything and are on a resource list,
2179          * then we're blocked/waiting.
2180          *
2181          * If we haven't been granted anything and we're NOT on a resource list,
2182          * then we haven't got a reply yet and don't have a known disposition.
2183          * This happens whenever a lock enqueue is the request that triggers
2184          * recovery.
2185          */
2186         if (lock->l_granted_mode == lock->l_req_mode)
2187                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2188         else if (lock->l_granted_mode)
2189                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
2190         else if (!list_empty(&lock->l_res_link))
2191                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2192         else
2193                 flags = LDLM_FL_REPLAY;
2194
2195         req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2196                                         LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2197         if (req == NULL)
2198                 return -ENOMEM;
2199
2200         /* We're part of recovery, so don't wait for it. */
2201         req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2202
2203         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2204         ldlm_lock2desc(lock, &body->lock_desc);
2205         body->lock_flags = ldlm_flags_to_wire(flags);
2206
2207         ldlm_lock2handle(lock, &body->lock_handle[0]);
2208         if (lock->l_lvb_len > 0)
2209                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2210         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2211                              lock->l_lvb_len);
2212         ptlrpc_request_set_replen(req);
2213         /* notify the server we've replayed all requests.
2214          * also, we mark the request to be put on a dedicated
2215          * queue to be processed after all request replayes.
2216          * bug 6063 */
2217         lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2218
2219         LDLM_DEBUG(lock, "replaying lock:");
2220
2221         atomic_inc(&req->rq_import->imp_replay_inflight);
2222         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2223         aa = ptlrpc_req_async_args(req);
2224         aa->lock_handle = body->lock_handle[0];
2225         req->rq_interpret_reply = (ptlrpc_interpterer_t)replay_lock_interpret;
2226         ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
2227
2228         return 0;
2229 }
2230
2231 /**
2232  * Cancel as many unused locks as possible before replay. since we are
2233  * in recovery, we can't wait for any outstanding RPCs to send any RPC
2234  * to the server.
2235  *
2236  * Called only in recovery before replaying locks. there is no need to
2237  * replay locks that are unused. since the clients may hold thousands of
2238  * cached unused locks, dropping the unused locks can greatly reduce the
2239  * load on the servers at recovery time.
2240  */
2241 static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns)
2242 {
2243         int canceled;
2244         LIST_HEAD(cancels);
2245
2246         CDEBUG(D_DLMTRACE, "Dropping as many unused locks as possible before replay for namespace %s (%d)\n",
2247                ldlm_ns_name(ns), ns->ns_nr_unused);
2248
2249         /* We don't need to care whether or not LRU resize is enabled
2250          * because the LDLM_CANCEL_NO_WAIT policy doesn't use the
2251          * count parameter */
2252         canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0,
2253                                          LCF_LOCAL, LDLM_CANCEL_NO_WAIT);
2254
2255         CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n",
2256                            canceled, ldlm_ns_name(ns));
2257 }
2258
2259 int ldlm_replay_locks(struct obd_import *imp)
2260 {
2261         struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2262         LIST_HEAD(list);
2263         struct ldlm_lock *lock, *next;
2264         int rc = 0;
2265
2266         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
2267
2268         /* don't replay locks if import failed recovery */
2269         if (imp->imp_vbr_failed)
2270                 return 0;
2271
2272         /* ensure this doesn't fall to 0 before all have been queued */
2273         atomic_inc(&imp->imp_replay_inflight);
2274
2275         if (ldlm_cancel_unused_locks_before_replay)
2276                 ldlm_cancel_unused_locks_for_replay(ns);
2277
2278         ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2279
2280         list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2281                 list_del_init(&lock->l_pending_chain);
2282                 if (rc) {
2283                         LDLM_LOCK_RELEASE(lock);
2284                         continue; /* or try to do the rest? */
2285                 }
2286                 rc = replay_one_lock(imp, lock);
2287                 LDLM_LOCK_RELEASE(lock);
2288         }
2289
2290         atomic_dec(&imp->imp_replay_inflight);
2291
2292         return rc;
2293 }
2294 EXPORT_SYMBOL(ldlm_replay_locks);