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 = 0, last_dump = 0;
  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
 252         /* Wait a long time for enqueue - server may have to callback a
 253            lock from another client.  Server will evict the other client if it
 254            doesn't respond reasonably, and then give us the lock. */
 255         timeout = ldlm_get_enq_timeout(lock) * 2;
 256
 257         lwd.lwd_lock = lock;
 258
 259         if (lock->l_flags & LDLM_FL_NO_TIMEOUT) {
 260                 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
 261                 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
 262         } else {
 263                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout),
 264                                        ldlm_expired_completion_wait,
 265                                        interrupted_completion_wait, &lwd);
 266         }
 267
 268         if (imp != NULL) {
 269                 spin_lock(&imp->imp_lock);
 270                 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
 271                 spin_unlock(&imp->imp_lock);
 272         }
 273
 274         if (ns_is_client(ldlm_lock_to_ns(lock)) &&
 275             OBD_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST,
 276                                  OBD_FAIL_LDLM_CP_BL_RACE | OBD_FAIL_ONCE)) {
 277                 lock->l_flags |= LDLM_FL_FAIL_LOC;
 278                 rc = -EINTR;
 279         } else {
 280                 /* Go to sleep until the lock is granted or cancelled. */
 281                 rc = l_wait_event(lock->l_waitq,
 282                                   is_granted_or_cancelled(lock), &lwi);
 283         }
 284
 285         if (rc) {
 286                 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
 287                            rc);
 288                 return rc;
 289         }
 290
 291         return ldlm_completion_tail(lock);
 292 }
 293 EXPORT_SYMBOL(ldlm_completion_ast);
 294
 295 /**
 296  * A helper to build a blocking AST function
 297  *
 298  * Perform a common operation for blocking ASTs:
 299  * deferred lock cancellation.
 300  *
 301  * \param lock the lock blocking or canceling AST was called on
 302  * \retval 0
 303  * \see mdt_blocking_ast
 304  * \see ldlm_blocking_ast
 305  */
 306 int ldlm_blocking_ast_nocheck(struct ldlm_lock *lock)
 307 {
 308         int do_ast;
 309
 310         lock->l_flags |= LDLM_FL_CBPENDING;
 311         do_ast = (!lock->l_readers && !lock->l_writers);
 312         unlock_res_and_lock(lock);
 313
 314         if (do_ast) {
 315                 struct lustre_handle lockh;
 316                 int rc;
 317
 318                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
 319                 ldlm_lock2handle(lock, &lockh);
 320                 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
 321                 if (rc < 0)
 322                         CERROR("ldlm_cli_cancel: %d\n", rc);
 323         } else {
 324                 LDLM_DEBUG(lock, "Lock still has references, will be cancelled later");
 325         }
 326         return 0;
 327 }
 328 EXPORT_SYMBOL(ldlm_blocking_ast_nocheck);
 329
 330 /**
 331  * Server blocking AST
 332  *
 333  * ->l_blocking_ast() callback for LDLM locks acquired by server-side
 334  * OBDs.
 335  *
 336  * \param lock the lock which blocks a request or cancelling lock
 337  * \param desc unused
 338  * \param data unused
 339  * \param flag indicates whether this cancelling or blocking callback
 340  * \retval 0
 341  * \see ldlm_blocking_ast_nocheck
 342  */
 343 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 344                       void *data, int flag)
 345 {
 346         if (flag == LDLM_CB_CANCELING) {
 347                 /* Don't need to do anything here. */
 348                 return 0;
 349         }
 350
 351         lock_res_and_lock(lock);
 352         /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
 353          * that ldlm_blocking_ast is called just before intent_policy method
 354          * takes the lr_lock, then by the time we get the lock, we might not
 355          * be the correct blocking function anymore.  So check, and return
 356          * early, if so. */
 357         if (lock->l_blocking_ast != ldlm_blocking_ast) {
 358                 unlock_res_and_lock(lock);
 359                 return 0;
 360         }
 361         return ldlm_blocking_ast_nocheck(lock);
 362 }
 363 EXPORT_SYMBOL(ldlm_blocking_ast);
 364
 365 /**
 366  * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See
 367  * comment in filter_intent_policy() on why you may need this.
 368  */
 369 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
 370 {
 371         /*
 372          * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for
 373          * that is rather subtle: with OST-side locking, it may so happen that
 374          * _all_ extent locks are held by the OST. If client wants to obtain
 375          * current file size it calls ll{,u}_glimpse_size(), and (as locks are
 376          * on the server), dummy glimpse callback fires and does
 377          * nothing. Client still receives correct file size due to the
 378          * following fragment in filter_intent_policy():
 379          *
 380          * rc = l->l_glimpse_ast(l, NULL); // this will update the LVB
 381          * if (rc != 0 && res->lr_namespace->ns_lvbo &&
 382          *     res->lr_namespace->ns_lvbo->lvbo_update) {
 383          *       res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1);
 384          * }
 385          *
 386          * that is, after glimpse_ast() fails, filter_lvbo_update() runs, and
 387          * returns correct file size to the client.
 388          */
 389         return -ELDLM_NO_LOCK_DATA;
 390 }
 391 EXPORT_SYMBOL(ldlm_glimpse_ast);
 392
 393 /**
 394  * Enqueue a local lock (typically on a server).
 395  */
 396 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
 397                            const struct ldlm_res_id *res_id,
 398                            ldlm_type_t type, ldlm_policy_data_t *policy,
 399                            ldlm_mode_t mode, __u64 *flags,
 400                            ldlm_blocking_callback blocking,
 401                            ldlm_completion_callback completion,
 402                            ldlm_glimpse_callback glimpse,
 403                            void *data, __u32 lvb_len, enum lvb_type lvb_type,
 404                            const __u64 *client_cookie,
 405                            struct lustre_handle *lockh)
 406 {
 407         struct ldlm_lock *lock;
 408         int err;
 409         const struct ldlm_callback_suite cbs = { .lcs_completion = completion,
 410                                                  .lcs_blocking   = blocking,
 411                                                  .lcs_glimpse    = glimpse,
 412         };
 413
 414         LASSERT(!(*flags & LDLM_FL_REPLAY));
 415         if (unlikely(ns_is_client(ns))) {
 416                 CERROR("Trying to enqueue local lock in a shadow namespace\n");
 417                 LBUG();
 418         }
 419
 420         lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len,
 421                                 lvb_type);
 422         if (unlikely(!lock)) {
 423                 err = -ENOMEM;
 424                 goto out_nolock;
 425         }
 426
 427         ldlm_lock2handle(lock, lockh);
 428
 429         /* NB: we don't have any lock now (lock_res_and_lock)
 430          * because it's a new lock */
 431         ldlm_lock_addref_internal_nolock(lock, mode);
 432         lock->l_flags |= LDLM_FL_LOCAL;
 433         if (*flags & LDLM_FL_ATOMIC_CB)
 434                 lock->l_flags |= LDLM_FL_ATOMIC_CB;
 435
 436         if (policy != NULL)
 437                 lock->l_policy_data = *policy;
 438         if (client_cookie != NULL)
 439                 lock->l_client_cookie = *client_cookie;
 440         if (type == LDLM_EXTENT)
 441                 lock->l_req_extent = policy->l_extent;
 442
 443         err = ldlm_lock_enqueue(ns, &lock, policy, flags);
 444         if (unlikely(err != ELDLM_OK))
 445                 goto out;
 446
 447         if (policy != NULL)
 448                 *policy = lock->l_policy_data;
 449
 450         if (lock->l_completion_ast)
 451                 lock->l_completion_ast(lock, *flags, NULL);
 452
 453         LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
 454  out:
 455         LDLM_LOCK_RELEASE(lock);
 456  out_nolock:
 457         return err;
 458 }
 459 EXPORT_SYMBOL(ldlm_cli_enqueue_local);
 460
 461 static void failed_lock_cleanup(struct ldlm_namespace *ns,
 462                                 struct ldlm_lock *lock, int mode)
 463 {
 464         int need_cancel = 0;
 465
 466         /* Set a flag to prevent us from sending a CANCEL (bug 407) */
 467         lock_res_and_lock(lock);
 468         /* Check that lock is not granted or failed, we might race. */
 469         if ((lock->l_req_mode != lock->l_granted_mode) &&
 470             !(lock->l_flags & LDLM_FL_FAILED)) {
 471                 /* Make sure that this lock will not be found by raced
 472                  * bl_ast and -EINVAL reply is sent to server anyways.
 473                  * bug 17645 */
 474                 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED |
 475                                  LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING;
 476                 need_cancel = 1;
 477         }
 478         unlock_res_and_lock(lock);
 479
 480         if (need_cancel)
 481                 LDLM_DEBUG(lock,
 482                            "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING");
 483         else
 484                 LDLM_DEBUG(lock, "lock was granted or failed in race");
 485
 486         ldlm_lock_decref_internal(lock, mode);
 487
 488         /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
 489          *       from llite/file.c/ll_file_flock(). */
 490         /* This code makes for the fact that we do not have blocking handler on
 491          * a client for flock locks. As such this is the place where we must
 492          * completely kill failed locks. (interrupted and those that
 493          * were waiting to be granted when server evicted us. */
 494         if (lock->l_resource->lr_type == LDLM_FLOCK) {
 495                 lock_res_and_lock(lock);
 496                 ldlm_resource_unlink_lock(lock);
 497                 ldlm_lock_destroy_nolock(lock);
 498                 unlock_res_and_lock(lock);
 499         }
 500 }
 501
 502 /**
 503  * Finishing portion of client lock enqueue code.
 504  *
 505  * Called after receiving reply from server.
 506  */
 507 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
 508                           ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
 509                           __u64 *flags, void *lvb, __u32 lvb_len,
 510                           struct lustre_handle *lockh, int rc)
 511 {
 512         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 513         int is_replay = *flags & LDLM_FL_REPLAY;
 514         struct ldlm_lock *lock;
 515         struct ldlm_reply *reply;
 516         int cleanup_phase = 1;
 517         int size = 0;
 518
 519         lock = ldlm_handle2lock(lockh);
 520         /* ldlm_cli_enqueue is holding a reference on this lock. */
 521         if (!lock) {
 522                 LASSERT(type == LDLM_FLOCK);
 523                 return -ENOLCK;
 524         }
 525
 526         LASSERTF(ergo(lvb_len != 0, lvb_len == lock->l_lvb_len),
 527                  "lvb_len = %d, l_lvb_len = %d\n", lvb_len, lock->l_lvb_len);
 528
 529         if (rc != ELDLM_OK) {
 530                 LASSERT(!is_replay);
 531                 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
 532                            rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
 533
 534                 if (rc != ELDLM_LOCK_ABORTED)
 535                         goto cleanup;
 536         }
 537
 538         /* Before we return, swab the reply */
 539         reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
 540         if (reply == NULL) {
 541                 rc = -EPROTO;
 542                 goto cleanup;
 543         }
 544
 545         if (lvb_len != 0) {
 546                 LASSERT(lvb != NULL);
 547
 548                 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB,
 549                                             RCL_SERVER);
 550                 if (size < 0) {
 551                         LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", size);
 552                         rc = size;
 553                         goto cleanup;
 554                 } else if (unlikely(size > lvb_len)) {
 555                         LDLM_ERROR(lock, "Replied LVB is larger than expectation, expected = %d, replied = %d",
 556                                    lvb_len, size);
 557                         rc = -EINVAL;
 558                         goto cleanup;
 559                 }
 560         }
 561
 562         if (rc == ELDLM_LOCK_ABORTED) {
 563                 if (lvb_len != 0)
 564                         rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
 565                                            lvb, size);
 566                 if (rc == 0)
 567                         rc = ELDLM_LOCK_ABORTED;
 568                 goto cleanup;
 569         }
 570
 571         /* lock enqueued on the server */
 572         cleanup_phase = 0;
 573
 574         lock_res_and_lock(lock);
 575         /* Key change rehash lock in per-export hash with new key */
 576         if (exp->exp_lock_hash) {
 577                 /* In the function below, .hs_keycmp resolves to
 578                  * ldlm_export_lock_keycmp() */
 579                 /* coverity[overrun-buffer-val] */
 580                 cfs_hash_rehash_key(exp->exp_lock_hash,
 581                                     &lock->l_remote_handle,
 582                                     &reply->lock_handle,
 583                                     &lock->l_exp_hash);
 584         } else {
 585                 lock->l_remote_handle = reply->lock_handle;
 586         }
 587
 588         *flags = ldlm_flags_from_wire(reply->lock_flags);
 589         lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags &
 590                                               LDLM_INHERIT_FLAGS);
 591         /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match()
 592          * to wait with no timeout as well */
 593         lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags &
 594                                               LDLM_FL_NO_TIMEOUT);
 595         unlock_res_and_lock(lock);
 596
 597         CDEBUG(D_INFO, "local: %p, remote cookie: %#llx, flags: 0x%llx\n",
 598                lock, reply->lock_handle.cookie, *flags);
 599
 600         /* If enqueue returned a blocked lock but the completion handler has
 601          * already run, then it fixed up the resource and we don't need to do it
 602          * again. */
 603         if ((*flags) & LDLM_FL_LOCK_CHANGED) {
 604                 int newmode = reply->lock_desc.l_req_mode;
 605
 606                 LASSERT(!is_replay);
 607                 if (newmode && newmode != lock->l_req_mode) {
 608                         LDLM_DEBUG(lock, "server returned different mode %s",
 609                                    ldlm_lockname[newmode]);
 610                         lock->l_req_mode = newmode;
 611                 }
 612
 613                 if (!ldlm_res_eq(&reply->lock_desc.l_resource.lr_name,
 614                                  &lock->l_resource->lr_name)) {
 615                         CDEBUG(D_INFO, "remote intent success, locking "DLDLMRES
 616                                        " instead of "DLDLMRES"\n",
 617                                PLDLMRES(&reply->lock_desc.l_resource),
 618                                PLDLMRES(lock->l_resource));
 619
 620                         rc = ldlm_lock_change_resource(ns, lock,
 621                                         &reply->lock_desc.l_resource.lr_name);
 622                         if (rc || lock->l_resource == NULL) {
 623                                 rc = -ENOMEM;
 624                                 goto cleanup;
 625                         }
 626                         LDLM_DEBUG(lock, "client-side enqueue, new resource");
 627                 }
 628                 if (with_policy)
 629                         if (!(type == LDLM_IBITS &&
 630                               !(exp_connect_flags(exp) & OBD_CONNECT_IBITS)))
 631                                 /* We assume lock type cannot change on server*/
 632                                 ldlm_convert_policy_to_local(exp,
 633                                                 lock->l_resource->lr_type,
 634                                                 &reply->lock_desc.l_policy_data,
 635                                                 &lock->l_policy_data);
 636                 if (type != LDLM_PLAIN)
 637                         LDLM_DEBUG(lock,
 638                                    "client-side enqueue, new policy data");
 639         }
 640
 641         if ((*flags) & LDLM_FL_AST_SENT ||
 642             /* Cancel extent locks as soon as possible on a liblustre client,
 643              * because it cannot handle asynchronous ASTs robustly (see
 644              * bug 7311). */
 645             (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) {
 646                 lock_res_and_lock(lock);
 647                 lock->l_flags |= LDLM_FL_CBPENDING |  LDLM_FL_BL_AST;
 648                 unlock_res_and_lock(lock);
 649                 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
 650         }
 651
 652         /* If the lock has already been granted by a completion AST, don't
 653          * clobber the LVB with an older one. */
 654         if (lvb_len != 0) {
 655                 /* We must lock or a racing completion might update lvb without
 656                  * letting us know and we'll clobber the correct value.
 657                  * Cannot unlock after the check either, a that still leaves
 658                  * a tiny window for completion to get in */
 659                 lock_res_and_lock(lock);
 660                 if (lock->l_req_mode != lock->l_granted_mode)
 661                         rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
 662                                            lock->l_lvb_data, size);
 663                 unlock_res_and_lock(lock);
 664                 if (rc < 0) {
 665                         cleanup_phase = 1;
 666                         goto cleanup;
 667                 }
 668         }
 669
 670         if (!is_replay) {
 671                 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
 672                 if (lock->l_completion_ast != NULL) {
 673                         int err = lock->l_completion_ast(lock, *flags, NULL);
 674
 675                         if (!rc)
 676                                 rc = err;
 677                         if (rc)
 678                                 cleanup_phase = 1;
 679                 }
 680         }
 681
 682         if (lvb_len && lvb != NULL) {
 683                 /* Copy the LVB here, and not earlier, because the completion
 684                  * AST (if any) can override what we got in the reply */
 685                 memcpy(lvb, lock->l_lvb_data, lvb_len);
 686         }
 687
 688         LDLM_DEBUG(lock, "client-side enqueue END");
 689 cleanup:
 690         if (cleanup_phase == 1 && rc)
 691                 failed_lock_cleanup(ns, lock, mode);
 692         /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
 693         LDLM_LOCK_PUT(lock);
 694         LDLM_LOCK_RELEASE(lock);
 695         return rc;
 696 }
 697 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
 698
 699 /**
 700  * Estimate number of lock handles that would fit into request of given
 701  * size.  PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
 702  * a single page on the send/receive side. XXX: 512 should be changed to
 703  * more adequate value.
 704  */
 705 static inline int ldlm_req_handles_avail(int req_size, int off)
 706 {
 707         int avail;
 708
 709         avail = min_t(int, LDLM_MAXREQSIZE, PAGE_CACHE_SIZE - 512) - req_size;
 710         if (likely(avail >= 0))
 711                 avail /= (int)sizeof(struct lustre_handle);
 712         else
 713                 avail = 0;
 714         avail += LDLM_LOCKREQ_HANDLES - off;
 715
 716         return avail;
 717 }
 718
 719 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
 720                                              enum req_location loc,
 721                                              int off)
 722 {
 723         int size = req_capsule_msg_size(pill, loc);
 724
 725         return ldlm_req_handles_avail(size, off);
 726 }
 727
 728 static inline int ldlm_format_handles_avail(struct obd_import *imp,
 729                                             const struct req_format *fmt,
 730                                             enum req_location loc, int off)
 731 {
 732         int size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
 733
 734         return ldlm_req_handles_avail(size, off);
 735 }
 736
 737 /**
 738  * Cancel LRU locks and pack them into the enqueue request. Pack there the given
 739  * \a count locks in \a cancels.
 740  *
 741  * This is to be called by functions preparing their own requests that
 742  * might contain lists of locks to cancel in addition to actual operation
 743  * that needs to be performed.
 744  */
 745 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
 746                       int version, int opc, int canceloff,
 747                       struct list_head *cancels, int count)
 748 {
 749         struct ldlm_namespace   *ns = exp->exp_obd->obd_namespace;
 750         struct req_capsule      *pill = &req->rq_pill;
 751         struct ldlm_request     *dlm = NULL;
 752         int flags, avail, to_free, pack = 0;
 753         LIST_HEAD(head);
 754         int rc;
 755
 756         if (cancels == NULL)
 757                 cancels = &head;
 758         if (ns_connect_cancelset(ns)) {
 759                 /* Estimate the amount of available space in the request. */
 760                 req_capsule_filled_sizes(pill, RCL_CLIENT);
 761                 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
 762
 763                 flags = ns_connect_lru_resize(ns) ?
 764                         LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
 765                 to_free = !ns_connect_lru_resize(ns) &&
 766                           opc == LDLM_ENQUEUE ? 1 : 0;
 767
 768                 /* Cancel LRU locks here _only_ if the server supports
 769                  * EARLY_CANCEL. Otherwise we have to send extra CANCEL
 770                  * RPC, which will make us slower. */
 771                 if (avail > count)
 772                         count += ldlm_cancel_lru_local(ns, cancels, to_free,
 773                                                        avail - count, 0, flags);
 774                 if (avail > count)
 775                         pack = count;
 776                 else
 777                         pack = avail;
 778                 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
 779                                      ldlm_request_bufsize(pack, opc));
 780         }
 781
 782         rc = ptlrpc_request_pack(req, version, opc);
 783         if (rc) {
 784                 ldlm_lock_list_put(cancels, l_bl_ast, count);
 785                 return rc;
 786         }
 787
 788         if (ns_connect_cancelset(ns)) {
 789                 if (canceloff) {
 790                         dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
 791                         LASSERT(dlm);
 792                         /* Skip first lock handler in ldlm_request_pack(),
 793                          * this method will increment @lock_count according
 794                          * to the lock handle amount actually written to
 795                          * the buffer. */
 796                         dlm->lock_count = canceloff;
 797                 }
 798                 /* Pack into the request @pack lock handles. */
 799                 ldlm_cli_cancel_list(cancels, pack, req, 0);
 800                 /* Prepare and send separate cancel RPC for others. */
 801                 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
 802         } else {
 803                 ldlm_lock_list_put(cancels, l_bl_ast, count);
 804         }
 805         return 0;
 806 }
 807 EXPORT_SYMBOL(ldlm_prep_elc_req);
 808
 809 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
 810                           struct list_head *cancels, int count)
 811 {
 812         return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
 813                                  LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
 814 }
 815 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
 816
 817 struct ptlrpc_request *ldlm_enqueue_pack(struct obd_export *exp, int lvb_len)
 818 {
 819         struct ptlrpc_request *req;
 820         int rc;
 821
 822         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LDLM_ENQUEUE);
 823         if (req == NULL)
 824                 return ERR_PTR(-ENOMEM);
 825
 826         rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
 827         if (rc) {
 828                 ptlrpc_request_free(req);
 829                 return ERR_PTR(rc);
 830         }
 831
 832         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, lvb_len);
 833         ptlrpc_request_set_replen(req);
 834         return req;
 835 }
 836 EXPORT_SYMBOL(ldlm_enqueue_pack);
 837
 838 /**
 839  * Client-side lock enqueue.
 840  *
 841  * If a request has some specific initialisation it is passed in \a reqp,
 842  * otherwise it is created in ldlm_cli_enqueue.
 843  *
 844  * Supports sync and async requests, pass \a async flag accordingly. If a
 845  * request was created in ldlm_cli_enqueue and it is the async request,
 846  * pass it to the caller in \a reqp.
 847  */
 848 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
 849                      struct ldlm_enqueue_info *einfo,
 850                      const struct ldlm_res_id *res_id,
 851                      ldlm_policy_data_t const *policy, __u64 *flags,
 852                      void *lvb, __u32 lvb_len, enum lvb_type lvb_type,
 853                      struct lustre_handle *lockh, int async)
 854 {
 855         struct ldlm_namespace *ns;
 856         struct ldlm_lock      *lock;
 857         struct ldlm_request   *body;
 858         int                 is_replay = *flags & LDLM_FL_REPLAY;
 859         int                 req_passed_in = 1;
 860         int                 rc, err;
 861         struct ptlrpc_request *req;
 862
 863         LASSERT(exp != NULL);
 864
 865         ns = exp->exp_obd->obd_namespace;
 866
 867         /* If we're replaying this lock, just check some invariants.
 868          * If we're creating a new lock, get everything all setup nice. */
 869         if (is_replay) {
 870                 lock = ldlm_handle2lock_long(lockh, 0);
 871                 LASSERT(lock != NULL);
 872                 LDLM_DEBUG(lock, "client-side enqueue START");
 873                 LASSERT(exp == lock->l_conn_export);
 874         } else {
 875                 const struct ldlm_callback_suite cbs = {
 876                         .lcs_completion = einfo->ei_cb_cp,
 877                         .lcs_blocking   = einfo->ei_cb_bl,
 878                         .lcs_glimpse    = einfo->ei_cb_gl
 879                 };
 880                 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
 881                                         einfo->ei_mode, &cbs, einfo->ei_cbdata,
 882                                         lvb_len, lvb_type);
 883                 if (lock == NULL)
 884                         return -ENOMEM;
 885                 /* for the local lock, add the reference */
 886                 ldlm_lock_addref_internal(lock, einfo->ei_mode);
 887                 ldlm_lock2handle(lock, lockh);
 888                 if (policy != NULL)
 889                                 lock->l_policy_data = *policy;
 890
 891                 if (einfo->ei_type == LDLM_EXTENT)
 892                         lock->l_req_extent = policy->l_extent;
 893                 LDLM_DEBUG(lock, "client-side enqueue START, flags %llx\n",
 894                            *flags);
 895         }
 896
 897         lock->l_conn_export = exp;
 898         lock->l_export = NULL;
 899         lock->l_blocking_ast = einfo->ei_cb_bl;
 900         lock->l_flags |= (*flags & (LDLM_FL_NO_LRU | LDLM_FL_EXCL));
 901
 902         /* lock not sent to server yet */
 903
 904         if (reqp == NULL || *reqp == NULL) {
 905                 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
 906                                                 &RQF_LDLM_ENQUEUE,
 907                                                 LUSTRE_DLM_VERSION,
 908                                                 LDLM_ENQUEUE);
 909                 if (req == NULL) {
 910                         failed_lock_cleanup(ns, lock, einfo->ei_mode);
 911                         LDLM_LOCK_RELEASE(lock);
 912                         return -ENOMEM;
 913                 }
 914                 req_passed_in = 0;
 915                 if (reqp)
 916                         *reqp = req;
 917         } else {
 918                 int len;
 919
 920                 req = *reqp;
 921                 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
 922                                            RCL_CLIENT);
 923                 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
 924                          DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
 925         }
 926
 927         /* Dump lock data into the request buffer */
 928         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
 929         ldlm_lock2desc(lock, &body->lock_desc);
 930         body->lock_flags = ldlm_flags_to_wire(*flags);
 931         body->lock_handle[0] = *lockh;
 932
 933         /* Continue as normal. */
 934         if (!req_passed_in) {
 935                 if (lvb_len > 0)
 936                         req_capsule_extend(&req->rq_pill,
 937                                            &RQF_LDLM_ENQUEUE_LVB);
 938                 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
 939                                      lvb_len);
 940                 ptlrpc_request_set_replen(req);
 941         }
 942
 943         /*
 944          * Liblustre client doesn't get extent locks, except for O_APPEND case
 945          * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
 946          * [i_size, OBD_OBJECT_EOF] lock is taken.
 947          */
 948         LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT ||
 949                      policy->l_extent.end == OBD_OBJECT_EOF));
 950
 951         if (async) {
 952                 LASSERT(reqp != NULL);
 953                 return 0;
 954         }
 955
 956         LDLM_DEBUG(lock, "sending request");
 957
 958         rc = ptlrpc_queue_wait(req);
 959
 960         err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
 961                                     einfo->ei_mode, flags, lvb, lvb_len,
 962                                     lockh, rc);
 963
 964         /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
 965          * one reference that we took */
 966         if (err == -ENOLCK)
 967                 LDLM_LOCK_RELEASE(lock);
 968         else
 969                 rc = err;
 970
 971         if (!req_passed_in && req != NULL) {
 972                 ptlrpc_req_finished(req);
 973                 if (reqp)
 974                         *reqp = NULL;
 975         }
 976
 977         return rc;
 978 }
 979 EXPORT_SYMBOL(ldlm_cli_enqueue);
 980
 981 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
 982                                   __u32 *flags)
 983 {
 984         struct ldlm_resource *res;
 985         int rc;
 986
 987         if (ns_is_client(ldlm_lock_to_ns(lock))) {
 988                 CERROR("Trying to cancel local lock\n");
 989                 LBUG();
 990         }
 991         LDLM_DEBUG(lock, "client-side local convert");
 992
 993         res = ldlm_lock_convert(lock, new_mode, flags);
 994         if (res) {
 995                 ldlm_reprocess_all(res);
 996                 rc = 0;
 997         } else {
 998                 rc = LUSTRE_EDEADLK;
 999         }
1000         LDLM_DEBUG(lock, "client-side local convert handler END");
1001         LDLM_LOCK_PUT(lock);
1002         return rc;
1003 }
1004
1005 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
1006  * conversion of locks which are on the waiting or converting queue */
1007 /* Caller of this code is supposed to take care of lock readers/writers
1008    accounting */
1009 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags)
1010 {
1011         struct ldlm_request   *body;
1012         struct ldlm_reply     *reply;
1013         struct ldlm_lock      *lock;
1014         struct ldlm_resource  *res;
1015         struct ptlrpc_request *req;
1016         int                 rc;
1017
1018         lock = ldlm_handle2lock(lockh);
1019         if (!lock) {
1020                 LBUG();
1021                 return -EINVAL;
1022         }
1023         *flags = 0;
1024
1025         if (lock->l_conn_export == NULL)
1026                 return ldlm_cli_convert_local(lock, new_mode, flags);
1027
1028         LDLM_DEBUG(lock, "client-side convert");
1029
1030         req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export),
1031                                         &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
1032                                         LDLM_CONVERT);
1033         if (req == NULL) {
1034                 LDLM_LOCK_PUT(lock);
1035                 return -ENOMEM;
1036         }
1037
1038         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1039         body->lock_handle[0] = lock->l_remote_handle;
1040
1041         body->lock_desc.l_req_mode = new_mode;
1042         body->lock_flags = ldlm_flags_to_wire(*flags);
1043
1044
1045         ptlrpc_request_set_replen(req);
1046         rc = ptlrpc_queue_wait(req);
1047         if (rc != ELDLM_OK)
1048                 goto out;
1049
1050         reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1051         if (reply == NULL) {
1052                 rc = -EPROTO;
1053                 goto out;
1054         }
1055
1056         if (req->rq_status) {
1057                 rc = req->rq_status;
1058                 goto out;
1059         }
1060
1061         res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
1062         if (res != NULL) {
1063                 ldlm_reprocess_all(res);
1064                 /* Go to sleep until the lock is granted. */
1065                 /* FIXME: or cancelled. */
1066                 if (lock->l_completion_ast) {
1067                         rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
1068                                                     NULL);
1069                         if (rc)
1070                                 goto out;
1071                 }
1072         } else {
1073                 rc = LUSTRE_EDEADLK;
1074         }
1075  out:
1076         LDLM_LOCK_PUT(lock);
1077         ptlrpc_req_finished(req);
1078         return rc;
1079 }
1080 EXPORT_SYMBOL(ldlm_cli_convert);
1081
1082 /**
1083  * Cancel locks locally.
1084  * Returns:
1085  * \retval LDLM_FL_LOCAL_ONLY if there is no need for a CANCEL RPC to the server
1086  * \retval LDLM_FL_CANCELING otherwise;
1087  * \retval LDLM_FL_BL_AST if there is a need for a separate CANCEL RPC.
1088  */
1089 static __u64 ldlm_cli_cancel_local(struct ldlm_lock *lock)
1090 {
1091         __u64 rc = LDLM_FL_LOCAL_ONLY;
1092
1093         if (lock->l_conn_export) {
1094                 bool local_only;
1095
1096                 LDLM_DEBUG(lock, "client-side cancel");
1097                 /* Set this flag to prevent others from getting new references*/
1098                 lock_res_and_lock(lock);
1099                 lock->l_flags |= LDLM_FL_CBPENDING;
1100                 local_only = !!(lock->l_flags &
1101                                 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
1102                 ldlm_cancel_callback(lock);
1103                 rc = (lock->l_flags & LDLM_FL_BL_AST) ?
1104                         LDLM_FL_BL_AST : LDLM_FL_CANCELING;
1105                 unlock_res_and_lock(lock);
1106
1107                 if (local_only) {
1108                         CDEBUG(D_DLMTRACE, "not sending request (at caller's instruction)\n");
1109                         rc = LDLM_FL_LOCAL_ONLY;
1110                 }
1111                 ldlm_lock_cancel(lock);
1112         } else {
1113                 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1114                         LDLM_ERROR(lock, "Trying to cancel local lock");
1115                         LBUG();
1116                 }
1117                 LDLM_DEBUG(lock, "server-side local cancel");
1118                 ldlm_lock_cancel(lock);
1119                 ldlm_reprocess_all(lock->l_resource);
1120         }
1121
1122         return rc;
1123 }
1124
1125 /**
1126  * Pack \a count locks in \a head into ldlm_request buffer of request \a req.
1127  */
1128 static void ldlm_cancel_pack(struct ptlrpc_request *req,
1129                              struct list_head *head, int count)
1130 {
1131         struct ldlm_request *dlm;
1132         struct ldlm_lock *lock;
1133         int max, packed = 0;
1134
1135         dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1136         LASSERT(dlm != NULL);
1137
1138         /* Check the room in the request buffer. */
1139         max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
1140                 sizeof(struct ldlm_request);
1141         max /= sizeof(struct lustre_handle);
1142         max += LDLM_LOCKREQ_HANDLES;
1143         LASSERT(max >= dlm->lock_count + count);
1144
1145         /* XXX: it would be better to pack lock handles grouped by resource.
1146          * so that the server cancel would call filter_lvbo_update() less
1147          * frequently. */
1148         list_for_each_entry(lock, head, l_bl_ast) {
1149                 if (!count--)
1150                         break;
1151                 LASSERT(lock->l_conn_export);
1152                 /* Pack the lock handle to the given request buffer. */
1153                 LDLM_DEBUG(lock, "packing");
1154                 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
1155                 packed++;
1156         }
1157         CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
1158 }
1159
1160 /**
1161  * Prepare and send a batched cancel RPC. It will include \a count lock
1162  * handles of locks given in \a cancels list. */
1163 int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
1164                         int count, ldlm_cancel_flags_t flags)
1165 {
1166         struct ptlrpc_request *req = NULL;
1167         struct obd_import *imp;
1168         int free, sent = 0;
1169         int rc = 0;
1170
1171         LASSERT(exp != NULL);
1172         LASSERT(count > 0);
1173
1174         CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, cfs_fail_val);
1175
1176         if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
1177                 return count;
1178
1179         free = ldlm_format_handles_avail(class_exp2cliimp(exp),
1180                                          &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
1181         if (count > free)
1182                 count = free;
1183
1184         while (1) {
1185                 imp = class_exp2cliimp(exp);
1186                 if (imp == NULL || imp->imp_invalid) {
1187                         CDEBUG(D_DLMTRACE,
1188                                "skipping cancel on invalid import %p\n", imp);
1189                         return count;
1190                 }
1191
1192                 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
1193                 if (req == NULL) {
1194                         rc = -ENOMEM;
1195                         goto out;
1196                 }
1197
1198                 req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
1199                 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
1200                                      ldlm_request_bufsize(count, LDLM_CANCEL));
1201
1202                 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
1203                 if (rc) {
1204                         ptlrpc_request_free(req);
1205                         goto out;
1206                 }
1207
1208                 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1209                 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1210                 ptlrpc_at_set_req_timeout(req);
1211
1212                 ldlm_cancel_pack(req, cancels, count);
1213
1214                 ptlrpc_request_set_replen(req);
1215                 if (flags & LCF_ASYNC) {
1216                         ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
1217                         sent = count;
1218                         goto out;
1219                 } else {
1220                         rc = ptlrpc_queue_wait(req);
1221                 }
1222                 if (rc == LUSTRE_ESTALE) {
1223                         CDEBUG(D_DLMTRACE, "client/server (nid %s) out of sync -- not fatal\n",
1224                                libcfs_nid2str(req->rq_import->
1225                                               imp_connection->c_peer.nid));
1226                         rc = 0;
1227                 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1228                            req->rq_import_generation == imp->imp_generation) {
1229                         ptlrpc_req_finished(req);
1230                         continue;
1231                 } else if (rc != ELDLM_OK) {
1232                         /* -ESHUTDOWN is common on umount */
1233                         CDEBUG_LIMIT(rc == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1234                                      "Got rc %d from cancel RPC: canceling anyway\n",
1235                                      rc);
1236                         break;
1237                 }
1238                 sent = count;
1239                 break;
1240         }
1241
1242         ptlrpc_req_finished(req);
1243 out:
1244         return sent ? sent : rc;
1245 }
1246 EXPORT_SYMBOL(ldlm_cli_cancel_req);
1247
1248 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1249 {
1250         LASSERT(imp != NULL);
1251         return &imp->imp_obd->obd_namespace->ns_pool;
1252 }
1253
1254 /**
1255  * Update client's OBD pool related fields with new SLV and Limit from \a req.
1256  */
1257 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1258 {
1259         struct obd_device *obd;
1260         __u64 new_slv;
1261         __u32 new_limit;
1262
1263         if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1264                      !imp_connect_lru_resize(req->rq_import))) {
1265                 /*
1266                  * Do nothing for corner cases.
1267                  */
1268                 return 0;
1269         }
1270
1271         /* In some cases RPC may contain SLV and limit zeroed out. This
1272          * is the case when server does not support LRU resize feature.
1273          * This is also possible in some recovery cases when server-side
1274          * reqs have no reference to the OBD export and thus access to
1275          * server-side namespace is not possible. */
1276         if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1277             lustre_msg_get_limit(req->rq_repmsg) == 0) {
1278                 DEBUG_REQ(D_HA, req, "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, struct list_head *cancels,
1595                                  int count, int max, int flags)
1596 {
1597         ldlm_cancel_lru_policy_t pf;
1598         struct ldlm_lock *lock, *next;
1599         int added = 0, unused, remained;
1600
1601         spin_lock(&ns->ns_lock);
1602         unused = ns->ns_nr_unused;
1603         remained = unused;
1604
1605         if (!ns_connect_lru_resize(ns))
1606                 count += unused - ns->ns_max_unused;
1607
1608         pf = ldlm_cancel_lru_policy(ns, flags);
1609         LASSERT(pf != NULL);
1610
1611         while (!list_empty(&ns->ns_unused_list)) {
1612                 ldlm_policy_res_t result;
1613
1614                 /* all unused locks */
1615                 if (remained-- <= 0)
1616                         break;
1617
1618                 /* For any flags, stop scanning if @max is reached. */
1619                 if (max && added >= max)
1620                         break;
1621
1622                 list_for_each_entry_safe(lock, next, &ns->ns_unused_list,
1623                                              l_lru) {
1624                         /* No locks which got blocking requests. */
1625                         LASSERT(!(lock->l_flags & LDLM_FL_BL_AST));
1626
1627                         if (flags & LDLM_CANCEL_NO_WAIT &&
1628                             lock->l_flags & LDLM_FL_SKIPPED)
1629                                 /* already processed */
1630                                 continue;
1631
1632                         /* Somebody is already doing CANCEL. No need for this
1633                          * lock in LRU, do not traverse it again. */
1634                         if (!(lock->l_flags & LDLM_FL_CANCELING))
1635                                 break;
1636
1637                         ldlm_lock_remove_from_lru_nolock(lock);
1638                 }
1639                 if (&lock->l_lru == &ns->ns_unused_list)
1640                         break;
1641
1642                 LDLM_LOCK_GET(lock);
1643                 spin_unlock(&ns->ns_lock);
1644                 lu_ref_add(&lock->l_reference, __func__, current);
1645
1646                 /* Pass the lock through the policy filter and see if it
1647                  * should stay in LRU.
1648                  *
1649                  * Even for shrinker policy we stop scanning if
1650                  * we find a lock that should stay in the cache.
1651                  * We should take into account lock age anyway
1652                  * as a new lock is a valuable resource even if
1653                  * it has a low weight.
1654                  *
1655                  * That is, for shrinker policy we drop only
1656                  * old locks, but additionally choose them by
1657                  * their weight. Big extent locks will stay in
1658                  * the cache. */
1659                 result = pf(ns, lock, unused, added, count);
1660                 if (result == LDLM_POLICY_KEEP_LOCK) {
1661                         lu_ref_del(&lock->l_reference,
1662                                    __func__, current);
1663                         LDLM_LOCK_RELEASE(lock);
1664                         spin_lock(&ns->ns_lock);
1665                         break;
1666                 }
1667                 if (result == LDLM_POLICY_SKIP_LOCK) {
1668                         lu_ref_del(&lock->l_reference,
1669                                    __func__, current);
1670                         LDLM_LOCK_RELEASE(lock);
1671                         spin_lock(&ns->ns_lock);
1672                         continue;
1673                 }
1674
1675                 lock_res_and_lock(lock);
1676                 /* Check flags again under the lock. */
1677                 if ((lock->l_flags & LDLM_FL_CANCELING) ||
1678                     (ldlm_lock_remove_from_lru(lock) == 0)) {
1679                         /* Another thread is removing lock from LRU, or
1680                          * somebody is already doing CANCEL, or there
1681                          * is a blocking request which will send cancel
1682                          * by itself, or the lock is no longer unused. */
1683                         unlock_res_and_lock(lock);
1684                         lu_ref_del(&lock->l_reference,
1685                                    __func__, current);
1686                         LDLM_LOCK_RELEASE(lock);
1687                         spin_lock(&ns->ns_lock);
1688                         continue;
1689                 }
1690                 LASSERT(!lock->l_readers && !lock->l_writers);
1691
1692                 /* If we have chosen to cancel this lock voluntarily, we
1693                  * better send cancel notification to server, so that it
1694                  * frees appropriate state. This might lead to a race
1695                  * where while we are doing cancel here, server is also
1696                  * silently cancelling this lock. */
1697                 lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK;
1698
1699                 /* Setting the CBPENDING flag is a little misleading,
1700                  * but prevents an important race; namely, once
1701                  * CBPENDING is set, the lock can accumulate no more
1702                  * readers/writers. Since readers and writers are
1703                  * already zero here, ldlm_lock_decref() won't see
1704                  * this flag and call l_blocking_ast */
1705                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1706
1707                 /* We can't re-add to l_lru as it confuses the
1708                  * refcounting in ldlm_lock_remove_from_lru() if an AST
1709                  * arrives after we drop lr_lock below. We use l_bl_ast
1710                  * and can't use l_pending_chain as it is used both on
1711                  * server and client nevertheless bug 5666 says it is
1712                  * used only on server */
1713                 LASSERT(list_empty(&lock->l_bl_ast));
1714                 list_add(&lock->l_bl_ast, cancels);
1715                 unlock_res_and_lock(lock);
1716                 lu_ref_del(&lock->l_reference, __func__, current);
1717                 spin_lock(&ns->ns_lock);
1718                 added++;
1719                 unused--;
1720         }
1721         spin_unlock(&ns->ns_lock);
1722         return added;
1723 }
1724
1725 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
1726                           int count, int max, ldlm_cancel_flags_t cancel_flags,
1727                           int flags)
1728 {
1729         int added;
1730
1731         added = ldlm_prepare_lru_list(ns, cancels, count, max, flags);
1732         if (added <= 0)
1733                 return added;
1734         return ldlm_cli_cancel_list_local(cancels, added, cancel_flags);
1735 }
1736
1737 /**
1738  * Cancel at least \a nr locks from given namespace LRU.
1739  *
1740  * When called with LCF_ASYNC the blocking callback will be handled
1741  * in a thread and this function will return after the thread has been
1742  * asked to call the callback.  When called with LCF_ASYNC the blocking
1743  * callback will be performed in this function.
1744  */
1745 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr,
1746                     ldlm_cancel_flags_t cancel_flags,
1747                     int flags)
1748 {
1749         LIST_HEAD(cancels);
1750         int count, rc;
1751
1752         /* Just prepare the list of locks, do not actually cancel them yet.
1753          * Locks are cancelled later in a separate thread. */
1754         count = ldlm_prepare_lru_list(ns, &cancels, nr, 0, flags);
1755         rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, cancel_flags);
1756         if (rc == 0)
1757                 return count;
1758
1759         return 0;
1760 }
1761
1762 /**
1763  * Find and cancel locally unused locks found on resource, matched to the
1764  * given policy, mode. GET the found locks and add them into the \a cancels
1765  * list.
1766  */
1767 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1768                                struct list_head *cancels,
1769                                ldlm_policy_data_t *policy,
1770                                ldlm_mode_t mode, __u64 lock_flags,
1771                                ldlm_cancel_flags_t cancel_flags, void *opaque)
1772 {
1773         struct ldlm_lock *lock;
1774         int count = 0;
1775
1776         lock_res(res);
1777         list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1778                 if (opaque != NULL && lock->l_ast_data != opaque) {
1779                         LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1780                                    lock->l_ast_data, opaque);
1781                         //LBUG();
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, struct cfs_hash_bd *bd,
1920                                        struct hlist_node *hnode, void *arg)
1921 {
1922         struct ldlm_resource       *res = cfs_hash_object(hs, hnode);
1923         struct ldlm_cli_cancel_arg     *lc = arg;
1924
1925         ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name,
1926                                         NULL, LCK_MINMODE,
1927                                         lc->lc_flags, lc->lc_opaque);
1928         /* must return 0 for hash iteration */
1929         return 0;
1930 }
1931
1932 /**
1933  * Cancel all locks on a namespace (or a specific resource, if given)
1934  * that have 0 readers/writers.
1935  *
1936  * If flags & LCF_LOCAL, throw the locks away without trying
1937  * to notify the server. */
1938 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1939                            const struct ldlm_res_id *res_id,
1940                            ldlm_cancel_flags_t flags, void *opaque)
1941 {
1942         struct ldlm_cli_cancel_arg arg = {
1943                 .lc_flags       = flags,
1944                 .lc_opaque      = opaque,
1945         };
1946
1947         if (ns == NULL)
1948                 return ELDLM_OK;
1949
1950         if (res_id != NULL) {
1951                 return ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
1952                                                        LCK_MINMODE, flags,
1953                                                        opaque);
1954         } else {
1955                 cfs_hash_for_each_nolock(ns->ns_rs_hash,
1956                                          ldlm_cli_hash_cancel_unused, &arg);
1957                 return ELDLM_OK;
1958         }
1959 }
1960 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
1961
1962 /* Lock iterators. */
1963
1964 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
1965                           void *closure)
1966 {
1967         struct list_head *tmp, *next;
1968         struct ldlm_lock *lock;
1969         int rc = LDLM_ITER_CONTINUE;
1970
1971         if (!res)
1972                 return LDLM_ITER_CONTINUE;
1973
1974         lock_res(res);
1975         list_for_each_safe(tmp, next, &res->lr_granted) {
1976                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1977
1978                 if (iter(lock, closure) == LDLM_ITER_STOP) {
1979                         rc = LDLM_ITER_STOP;
1980                         goto out;
1981                 }
1982         }
1983
1984         list_for_each_safe(tmp, next, &res->lr_converting) {
1985                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1986
1987                 if (iter(lock, closure) == LDLM_ITER_STOP) {
1988                         rc = LDLM_ITER_STOP;
1989                         goto out;
1990                 }
1991         }
1992
1993         list_for_each_safe(tmp, next, &res->lr_waiting) {
1994                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1995
1996                 if (iter(lock, closure) == LDLM_ITER_STOP) {
1997                         rc = LDLM_ITER_STOP;
1998                         goto out;
1999                 }
2000         }
2001  out:
2002         unlock_res(res);
2003         return rc;
2004 }
2005 EXPORT_SYMBOL(ldlm_resource_foreach);
2006
2007 struct iter_helper_data {
2008         ldlm_iterator_t iter;
2009         void *closure;
2010 };
2011
2012 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
2013 {
2014         struct iter_helper_data *helper = closure;
2015
2016         return helper->iter(lock, helper->closure);
2017 }
2018
2019 static int ldlm_res_iter_helper(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2020                                 struct hlist_node *hnode, void *arg)
2021
2022 {
2023         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2024
2025         return ldlm_resource_foreach(res, ldlm_iter_helper, arg) ==
2026                LDLM_ITER_STOP;
2027 }
2028
2029 void ldlm_namespace_foreach(struct ldlm_namespace *ns,
2030                             ldlm_iterator_t iter, void *closure)
2031
2032 {
2033         struct iter_helper_data helper = {
2034                 .iter           = iter,
2035                 .closure        = closure,
2036         };
2037
2038         cfs_hash_for_each_nolock(ns->ns_rs_hash,
2039                                  ldlm_res_iter_helper, &helper);
2040
2041 }
2042 EXPORT_SYMBOL(ldlm_namespace_foreach);
2043
2044 /* non-blocking function to manipulate a lock whose cb_data is being put away.
2045  * return  0:  find no resource
2046  *       > 0:  must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE.
2047  *       < 0:  errors
2048  */
2049 int ldlm_resource_iterate(struct ldlm_namespace *ns,
2050                           const struct ldlm_res_id *res_id,
2051                           ldlm_iterator_t iter, void *data)
2052 {
2053         struct ldlm_resource *res;
2054         int rc;
2055
2056         if (ns == NULL) {
2057                 CERROR("must pass in namespace\n");
2058                 LBUG();
2059         }
2060
2061         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
2062         if (res == NULL)
2063                 return 0;
2064
2065         LDLM_RESOURCE_ADDREF(res);
2066         rc = ldlm_resource_foreach(res, iter, data);
2067         LDLM_RESOURCE_DELREF(res);
2068         ldlm_resource_putref(res);
2069         return rc;
2070 }
2071 EXPORT_SYMBOL(ldlm_resource_iterate);
2072
2073 /* Lock replay */
2074
2075 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
2076 {
2077         struct list_head *list = closure;
2078
2079         /* we use l_pending_chain here, because it's unused on clients. */
2080         LASSERTF(list_empty(&lock->l_pending_chain),
2081                  "lock %p next %p prev %p\n",
2082                  lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
2083         /* bug 9573: don't replay locks left after eviction, or
2084          * bug 17614: locks being actively cancelled. Get a reference
2085          * on a lock so that it does not disappear under us (e.g. due to cancel)
2086          */
2087         if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_CANCELING))) {
2088                 list_add(&lock->l_pending_chain, list);
2089                 LDLM_LOCK_GET(lock);
2090         }
2091
2092         return LDLM_ITER_CONTINUE;
2093 }
2094
2095 static int replay_lock_interpret(const struct lu_env *env,
2096                                  struct ptlrpc_request *req,
2097                                  struct ldlm_async_args *aa, int rc)
2098 {
2099         struct ldlm_lock     *lock;
2100         struct ldlm_reply    *reply;
2101         struct obd_export    *exp;
2102
2103         atomic_dec(&req->rq_import->imp_replay_inflight);
2104         if (rc != ELDLM_OK)
2105                 goto out;
2106
2107
2108         reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2109         if (reply == NULL) {
2110                 rc = -EPROTO;
2111                 goto out;
2112         }
2113
2114         lock = ldlm_handle2lock(&aa->lock_handle);
2115         if (!lock) {
2116                 CERROR("received replay ack for unknown local cookie %#llx remote cookie %#llx from server %s id %s\n",
2117                        aa->lock_handle.cookie, reply->lock_handle.cookie,
2118                        req->rq_export->exp_client_uuid.uuid,
2119                        libcfs_id2str(req->rq_peer));
2120                 rc = -ESTALE;
2121                 goto out;
2122         }
2123
2124         /* Key change rehash lock in per-export hash with new key */
2125         exp = req->rq_export;
2126         if (exp && exp->exp_lock_hash) {
2127                 /* In the function below, .hs_keycmp resolves to
2128                  * ldlm_export_lock_keycmp() */
2129                 /* coverity[overrun-buffer-val] */
2130                 cfs_hash_rehash_key(exp->exp_lock_hash,
2131                                     &lock->l_remote_handle,
2132                                     &reply->lock_handle,
2133                                     &lock->l_exp_hash);
2134         } else {
2135                 lock->l_remote_handle = reply->lock_handle;
2136         }
2137
2138         LDLM_DEBUG(lock, "replayed lock:");
2139         ptlrpc_import_recovery_state_machine(req->rq_import);
2140         LDLM_LOCK_PUT(lock);
2141 out:
2142         if (rc != ELDLM_OK)
2143                 ptlrpc_connect_import(req->rq_import);
2144
2145         return rc;
2146 }
2147
2148 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2149 {
2150         struct ptlrpc_request *req;
2151         struct ldlm_async_args *aa;
2152         struct ldlm_request   *body;
2153         int flags;
2154
2155         /* Bug 11974: Do not replay a lock which is actively being canceled */
2156         if (lock->l_flags & LDLM_FL_CANCELING) {
2157                 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2158                 return 0;
2159         }
2160
2161         /* If this is reply-less callback lock, we cannot replay it, since
2162          * server might have long dropped it, but notification of that event was
2163          * lost by network. (and server granted conflicting lock already) */
2164         if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
2165                 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2166                 ldlm_lock_cancel(lock);
2167                 return 0;
2168         }
2169
2170         /*
2171          * If granted mode matches the requested mode, this lock is granted.
2172          *
2173          * If they differ, but we have a granted mode, then we were granted
2174          * one mode and now want another: ergo, converting.
2175          *
2176          * If we haven't been granted anything and are on a resource list,
2177          * then we're blocked/waiting.
2178          *
2179          * If we haven't been granted anything and we're NOT on a resource list,
2180          * then we haven't got a reply yet and don't have a known disposition.
2181          * This happens whenever a lock enqueue is the request that triggers
2182          * recovery.
2183          */
2184         if (lock->l_granted_mode == lock->l_req_mode)
2185                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2186         else if (lock->l_granted_mode)
2187                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
2188         else if (!list_empty(&lock->l_res_link))
2189                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2190         else
2191                 flags = LDLM_FL_REPLAY;
2192
2193         req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2194                                         LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2195         if (req == NULL)
2196                 return -ENOMEM;
2197
2198         /* We're part of recovery, so don't wait for it. */
2199         req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2200
2201         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2202         ldlm_lock2desc(lock, &body->lock_desc);
2203         body->lock_flags = ldlm_flags_to_wire(flags);
2204
2205         ldlm_lock2handle(lock, &body->lock_handle[0]);
2206         if (lock->l_lvb_len > 0)
2207                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2208         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2209                              lock->l_lvb_len);
2210         ptlrpc_request_set_replen(req);
2211         /* notify the server we've replayed all requests.
2212          * also, we mark the request to be put on a dedicated
2213          * queue to be processed after all request replayes.
2214          * bug 6063 */
2215         lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2216
2217         LDLM_DEBUG(lock, "replaying lock:");
2218
2219         atomic_inc(&req->rq_import->imp_replay_inflight);
2220         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2221         aa = ptlrpc_req_async_args(req);
2222         aa->lock_handle = body->lock_handle[0];
2223         req->rq_interpret_reply = (ptlrpc_interpterer_t)replay_lock_interpret;
2224         ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
2225
2226         return 0;
2227 }
2228
2229 /**
2230  * Cancel as many unused locks as possible before replay. since we are
2231  * in recovery, we can't wait for any outstanding RPCs to send any RPC
2232  * to the server.
2233  *
2234  * Called only in recovery before replaying locks. there is no need to
2235  * replay locks that are unused. since the clients may hold thousands of
2236  * cached unused locks, dropping the unused locks can greatly reduce the
2237  * load on the servers at recovery time.
2238  */
2239 static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns)
2240 {
2241         int canceled;
2242         LIST_HEAD(cancels);
2243
2244         CDEBUG(D_DLMTRACE, "Dropping as many unused locks as possible before replay for namespace %s (%d)\n",
2245                ldlm_ns_name(ns), ns->ns_nr_unused);
2246
2247         /* We don't need to care whether or not LRU resize is enabled
2248          * because the LDLM_CANCEL_NO_WAIT policy doesn't use the
2249          * count parameter */
2250         canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0,
2251                                          LCF_LOCAL, LDLM_CANCEL_NO_WAIT);
2252
2253         CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n",
2254                            canceled, ldlm_ns_name(ns));
2255 }
2256
2257 int ldlm_replay_locks(struct obd_import *imp)
2258 {
2259         struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2260         LIST_HEAD(list);
2261         struct ldlm_lock *lock, *next;
2262         int rc = 0;
2263
2264         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
2265
2266         /* don't replay locks if import failed recovery */
2267         if (imp->imp_vbr_failed)
2268                 return 0;
2269
2270         /* ensure this doesn't fall to 0 before all have been queued */
2271         atomic_inc(&imp->imp_replay_inflight);
2272
2273         if (ldlm_cancel_unused_locks_before_replay)
2274                 ldlm_cancel_unused_locks_for_replay(ns);
2275
2276         ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2277
2278         list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2279                 list_del_init(&lock->l_pending_chain);
2280                 if (rc) {
2281                         LDLM_LOCK_RELEASE(lock);
2282                         continue; /* or try to do the rest? */
2283                 }
2284                 rc = replay_one_lock(imp, lock);
2285                 LDLM_LOCK_RELEASE(lock);
2286         }
2287
2288         atomic_dec(&imp->imp_replay_inflight);
2289
2290         return rc;
2291 }
2292 EXPORT_SYMBOL(ldlm_replay_locks);