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, 2011, Whamcloud, Inc.
  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 #define DEBUG_SUBSYSTEM S_LDLM
  38 #ifndef __KERNEL__
  39 #include <signal.h>
  40 #include <liblustre.h>
  41 #endif
  42
  43 #include <lustre_dlm.h>
  44 #include <obd_class.h>
  45 #include <obd.h>
  46
  47 #include "ldlm_internal.h"
  48
  49 int ldlm_enqueue_min = OBD_TIMEOUT_DEFAULT;
  50 CFS_MODULE_PARM(ldlm_enqueue_min, "i", int, 0644,
  51                 "lock enqueue timeout minimum");
  52
  53 /* in client side, whether the cached locks will be canceled before replay */
  54 unsigned int ldlm_cancel_unused_locks_before_replay = 1;
  55
  56 static void interrupted_completion_wait(void *data)
  57 {
  58 }
  59
  60 struct lock_wait_data {
  61         struct ldlm_lock *lwd_lock;
  62         __u32             lwd_conn_cnt;
  63 };
  64
  65 struct ldlm_async_args {
  66         struct lustre_handle lock_handle;
  67 };
  68
  69 int ldlm_expired_completion_wait(void *data)
  70 {
  71         struct lock_wait_data *lwd = data;
  72         struct ldlm_lock *lock = lwd->lwd_lock;
  73         struct obd_import *imp;
  74         struct obd_device *obd;
  75
  76         ENTRY;
  77         if (lock->l_conn_export == NULL) {
  78                 static cfs_time_t next_dump = 0, last_dump = 0;
  79
  80                 if (ptlrpc_check_suspend())
  81                         RETURN(0);
  82
  83                 LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", "
  84                            CFS_DURATION_T"s ago); not entering recovery in "
  85                            "server code, just going back to sleep",
  86                            lock->l_last_activity,
  87                            cfs_time_sub(cfs_time_current_sec(),
  88                            lock->l_last_activity));
  89                 if (cfs_time_after(cfs_time_current(), next_dump)) {
  90                         last_dump = next_dump;
  91                         next_dump = cfs_time_shift(300);
  92                         ldlm_namespace_dump(D_DLMTRACE,
  93                                             ldlm_lock_to_ns(lock));
  94                         if (last_dump == 0)
  95                                 libcfs_debug_dumplog();
  96                 }
  97                 RETURN(0);
  98         }
  99
 100         obd = lock->l_conn_export->exp_obd;
 101         imp = obd->u.cli.cl_import;
 102         ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
 103         LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", "
 104                   CFS_DURATION_T"s ago), entering recovery for %s@%s",
 105                   lock->l_last_activity,
 106                   cfs_time_sub(cfs_time_current_sec(), lock->l_last_activity),
 107                   obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
 108
 109         RETURN(0);
 110 }
 111
 112 /* We use the same basis for both server side and client side functions
 113    from a single node. */
 114 int ldlm_get_enq_timeout(struct ldlm_lock *lock)
 115 {
 116         int timeout = at_get(ldlm_lock_to_ns_at(lock));
 117         if (AT_OFF)
 118                 return obd_timeout / 2;
 119         /* Since these are non-updating timeouts, we should be conservative.
 120            It would be nice to have some kind of "early reply" mechanism for
 121            lock callbacks too... */
 122         timeout = min_t(int, at_max, timeout + (timeout >> 1)); /* 150% */
 123         return max(timeout, ldlm_enqueue_min);
 124 }
 125 EXPORT_SYMBOL(ldlm_get_enq_timeout);
 126
 127 /**
 128  * Helper function for ldlm_completion_ast(), updating timings when lock is
 129  * actually granted.
 130  */
 131 static int ldlm_completion_tail(struct ldlm_lock *lock)
 132 {
 133         long delay;
 134         int  result;
 135
 136         if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED) {
 137                 LDLM_DEBUG(lock, "client-side enqueue: destroyed");
 138                 result = -EIO;
 139         } else {
 140                 delay = cfs_time_sub(cfs_time_current_sec(),
 141                                      lock->l_last_activity);
 142                 LDLM_DEBUG(lock, "client-side enqueue: granted after "
 143                            CFS_DURATION_T"s", delay);
 144
 145                 /* Update our time estimate */
 146                 at_measured(ldlm_lock_to_ns_at(lock),
 147                             delay);
 148                 result = 0;
 149         }
 150         return result;
 151 }
 152
 153 /**
 154  * Implementation of ->l_completion_ast() for a client, that doesn't wait
 155  * until lock is granted. Suitable for locks enqueued through ptlrpcd, of
 156  * other threads that cannot block for long.
 157  */
 158 int ldlm_completion_ast_async(struct ldlm_lock *lock, int flags, void *data)
 159 {
 160         ENTRY;
 161
 162         if (flags == LDLM_FL_WAIT_NOREPROC) {
 163                 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
 164                 RETURN(0);
 165         }
 166
 167         if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
 168                        LDLM_FL_BLOCK_CONV))) {
 169                 cfs_waitq_signal(&lock->l_waitq);
 170                 RETURN(ldlm_completion_tail(lock));
 171         }
 172
 173         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
 174                    "going forward");
 175         ldlm_reprocess_all(lock->l_resource);
 176         RETURN(0);
 177 }
 178
 179 /**
 180  * Client side LDLM "completion" AST. This is called in several cases:
 181  *
 182  *     - when a reply to an ENQUEUE rpc is received from the server
 183  *       (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at
 184  *       this point (determined by flags);
 185  *
 186  *     - when LDLM_CP_CALLBACK rpc comes to client to notify it that lock has
 187  *       been granted;
 188  *
 189  *     - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock
 190  *       gets correct lvb;
 191  *
 192  *     - to force all locks when resource is destroyed (cleanup_resource());
 193  *
 194  *     - during lock conversion (not used currently).
 195  *
 196  * If lock is not granted in the first case, this function waits until second
 197  * or penultimate cases happen in some other thread.
 198  *
 199  */
 200 int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data)
 201 {
 202         /* XXX ALLOCATE - 160 bytes */
 203         struct lock_wait_data lwd;
 204         struct obd_device *obd;
 205         struct obd_import *imp = NULL;
 206         struct l_wait_info lwi;
 207         __u32 timeout;
 208         int rc = 0;
 209         ENTRY;
 210
 211         if (flags == LDLM_FL_WAIT_NOREPROC) {
 212                 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
 213                 goto noreproc;
 214         }
 215
 216         if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
 217                        LDLM_FL_BLOCK_CONV))) {
 218                 cfs_waitq_signal(&lock->l_waitq);
 219                 RETURN(0);
 220         }
 221
 222         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
 223                    "sleeping");
 224
 225 noreproc:
 226
 227         obd = class_exp2obd(lock->l_conn_export);
 228
 229         /* if this is a local lock, then there is no import */
 230         if (obd != NULL) {
 231                 imp = obd->u.cli.cl_import;
 232         }
 233
 234         /* Wait a long time for enqueue - server may have to callback a
 235            lock from another client.  Server will evict the other client if it
 236            doesn't respond reasonably, and then give us the lock. */
 237         timeout = ldlm_get_enq_timeout(lock) * 2;
 238
 239         lwd.lwd_lock = lock;
 240
 241         if (lock->l_flags & LDLM_FL_NO_TIMEOUT) {
 242                 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
 243                 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
 244         } else {
 245                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout),
 246                                        ldlm_expired_completion_wait,
 247                                        interrupted_completion_wait, &lwd);
 248         }
 249
 250         if (imp != NULL) {
 251                 cfs_spin_lock(&imp->imp_lock);
 252                 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
 253                 cfs_spin_unlock(&imp->imp_lock);
 254         }
 255
 256         if (ns_is_client(ldlm_lock_to_ns(lock)) &&
 257             OBD_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST,
 258                                  OBD_FAIL_LDLM_CP_BL_RACE | OBD_FAIL_ONCE)) {
 259                 lock->l_flags |= LDLM_FL_FAIL_LOC;
 260                 rc = -EINTR;
 261         } else {
 262                 /* Go to sleep until the lock is granted or cancelled. */
 263                 rc = l_wait_event(lock->l_waitq,
 264                                   is_granted_or_cancelled(lock), &lwi);
 265         }
 266
 267         if (rc) {
 268                 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
 269                            rc);
 270                 RETURN(rc);
 271         }
 272
 273         RETURN(ldlm_completion_tail(lock));
 274 }
 275
 276 /**
 277  * A helper to build a blocking ast function
 278  *
 279  * Perform a common operation for blocking asts:
 280  * defferred lock cancellation.
 281  *
 282  * \param lock the lock blocking or canceling ast was called on
 283  * \retval 0
 284  * \see mdt_blocking_ast
 285  * \see ldlm_blocking_ast
 286  */
 287 int ldlm_blocking_ast_nocheck(struct ldlm_lock *lock)
 288 {
 289         int do_ast;
 290         ENTRY;
 291
 292         lock->l_flags |= LDLM_FL_CBPENDING;
 293         do_ast = (!lock->l_readers && !lock->l_writers);
 294         unlock_res_and_lock(lock);
 295
 296         if (do_ast) {
 297                 struct lustre_handle lockh;
 298                 int rc;
 299
 300                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
 301                 ldlm_lock2handle(lock, &lockh);
 302                 rc = ldlm_cli_cancel(&lockh);
 303                 if (rc < 0)
 304                         CERROR("ldlm_cli_cancel: %d\n", rc);
 305         } else {
 306                 LDLM_DEBUG(lock, "Lock still has references, will be "
 307                            "cancelled later");
 308         }
 309         RETURN(0);
 310 }
 311
 312 /**
 313  * Server blocking AST
 314  *
 315  * ->l_blocking_ast() callback for LDLM locks acquired by server-side
 316  * OBDs.
 317  *
 318  * \param lock the lock which blocks a request or cancelling lock
 319  * \param desc unused
 320  * \param data unused
 321  * \param flag indicates whether this cancelling or blocking callback
 322  * \retval 0
 323  * \see ldlm_blocking_ast_nocheck
 324  */
 325 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 326                       void *data, int flag)
 327 {
 328         ENTRY;
 329
 330         if (flag == LDLM_CB_CANCELING) {
 331                 /* Don't need to do anything here. */
 332                 RETURN(0);
 333         }
 334
 335         lock_res_and_lock(lock);
 336         /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
 337          * that ldlm_blocking_ast is called just before intent_policy method
 338          * takes the lr_lock, then by the time we get the lock, we might not
 339          * be the correct blocking function anymore.  So check, and return
 340          * early, if so. */
 341         if (lock->l_blocking_ast != ldlm_blocking_ast) {
 342                 unlock_res_and_lock(lock);
 343                 RETURN(0);
 344         }
 345         RETURN(ldlm_blocking_ast_nocheck(lock));
 346 }
 347
 348 /*
 349  * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See
 350  * comment in filter_intent_policy() on why you may need this.
 351  */
 352 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
 353 {
 354         /*
 355          * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for
 356          * that is rather subtle: with OST-side locking, it may so happen that
 357          * _all_ extent locks are held by the OST. If client wants to obtain
 358          * current file size it calls ll{,u}_glimpse_size(), and (as locks are
 359          * on the server), dummy glimpse callback fires and does
 360          * nothing. Client still receives correct file size due to the
 361          * following fragment in filter_intent_policy():
 362          *
 363          * rc = l->l_glimpse_ast(l, NULL); // this will update the LVB
 364          * if (rc != 0 && res->lr_namespace->ns_lvbo &&
 365          *     res->lr_namespace->ns_lvbo->lvbo_update) {
 366          *         res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1);
 367          * }
 368          *
 369          * that is, after glimpse_ast() fails, filter_lvbo_update() runs, and
 370          * returns correct file size to the client.
 371          */
 372         return -ELDLM_NO_LOCK_DATA;
 373 }
 374
 375 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
 376                            const struct ldlm_res_id *res_id,
 377                            ldlm_type_t type, ldlm_policy_data_t *policy,
 378                            ldlm_mode_t mode, int *flags,
 379                            ldlm_blocking_callback blocking,
 380                            ldlm_completion_callback completion,
 381                            ldlm_glimpse_callback glimpse,
 382                            void *data, __u32 lvb_len,
 383                            const __u64 *client_cookie,
 384                            struct lustre_handle *lockh)
 385 {
 386         struct ldlm_lock *lock;
 387         int err;
 388         const struct ldlm_callback_suite cbs = { .lcs_completion = completion,
 389                                                  .lcs_blocking   = blocking,
 390                                                  .lcs_glimpse    = glimpse,
 391         };
 392         ENTRY;
 393
 394         LASSERT(!(*flags & LDLM_FL_REPLAY));
 395         if (unlikely(ns_is_client(ns))) {
 396                 CERROR("Trying to enqueue local lock in a shadow namespace\n");
 397                 LBUG();
 398         }
 399
 400         lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len);
 401         if (unlikely(!lock))
 402                 GOTO(out_nolock, err = -ENOMEM);
 403
 404         ldlm_lock2handle(lock, lockh);
 405
 406         /* NB: we don't have any lock now (lock_res_and_lock)
 407          * because it's a new lock */
 408         ldlm_lock_addref_internal_nolock(lock, mode);
 409         lock->l_flags |= LDLM_FL_LOCAL;
 410         if (*flags & LDLM_FL_ATOMIC_CB)
 411                 lock->l_flags |= LDLM_FL_ATOMIC_CB;
 412
 413         if (policy != NULL)
 414                 lock->l_policy_data = *policy;
 415         if (client_cookie != NULL)
 416                 lock->l_client_cookie = *client_cookie;
 417         if (type == LDLM_EXTENT)
 418                 lock->l_req_extent = policy->l_extent;
 419
 420         err = ldlm_lock_enqueue(ns, &lock, policy, flags);
 421         if (unlikely(err != ELDLM_OK))
 422                 GOTO(out, err);
 423
 424         if (policy != NULL)
 425                 *policy = lock->l_policy_data;
 426
 427         if (lock->l_completion_ast)
 428                 lock->l_completion_ast(lock, *flags, NULL);
 429
 430         LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
 431         EXIT;
 432  out:
 433         LDLM_LOCK_RELEASE(lock);
 434  out_nolock:
 435         return err;
 436 }
 437
 438 static void failed_lock_cleanup(struct ldlm_namespace *ns,
 439                                 struct ldlm_lock *lock, int mode)
 440 {
 441         int need_cancel = 0;
 442
 443         /* Set a flag to prevent us from sending a CANCEL (bug 407) */
 444         lock_res_and_lock(lock);
 445         /* Check that lock is not granted or failed, we might race. */
 446         if ((lock->l_req_mode != lock->l_granted_mode) &&
 447             !(lock->l_flags & LDLM_FL_FAILED)) {
 448                 /* Make sure that this lock will not be found by raced
 449                  * bl_ast and -EINVAL reply is sent to server anyways.
 450                  * bug 17645 */
 451                 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED |
 452                                  LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING;
 453                 need_cancel = 1;
 454         }
 455         unlock_res_and_lock(lock);
 456
 457         if (need_cancel)
 458                 LDLM_DEBUG(lock,
 459                            "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | "
 460                            "LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING");
 461         else
 462                 LDLM_DEBUG(lock, "lock was granted or failed in race");
 463
 464         ldlm_lock_decref_internal(lock, mode);
 465
 466         /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
 467          *       from llite/file.c/ll_file_flock(). */
 468         /* This code makes for the fact that we do not have blocking handler on
 469          * a client for flock locks. As such this is the place where we must
 470          * completely kill failed locks. (interrupted and those that
 471          * were waiting to be granted when server evicted us. */
 472         if (lock->l_resource->lr_type == LDLM_FLOCK) {
 473                 lock_res_and_lock(lock);
 474                 ldlm_resource_unlink_lock(lock);
 475                 ldlm_lock_destroy_nolock(lock);
 476                 unlock_res_and_lock(lock);
 477         }
 478 }
 479
 480 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
 481                           ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
 482                           int *flags, void *lvb, __u32 lvb_len,
 483                           struct lustre_handle *lockh,int rc)
 484 {
 485         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 486         int is_replay = *flags & LDLM_FL_REPLAY;
 487         struct ldlm_lock *lock;
 488         struct ldlm_reply *reply;
 489         struct ost_lvb *tmplvb;
 490         int cleanup_phase = 1;
 491         ENTRY;
 492
 493         lock = ldlm_handle2lock(lockh);
 494         /* ldlm_cli_enqueue is holding a reference on this lock. */
 495         if (!lock) {
 496                 LASSERT(type == LDLM_FLOCK);
 497                 RETURN(-ENOLCK);
 498         }
 499
 500         if (rc != ELDLM_OK) {
 501                 LASSERT(!is_replay);
 502                 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
 503                            rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
 504                 if (rc == ELDLM_LOCK_ABORTED) {
 505                         /* Before we return, swab the reply */
 506                         reply = req_capsule_server_get(&req->rq_pill,
 507                                                        &RMF_DLM_REP);
 508                         if (reply == NULL)
 509                                 rc = -EPROTO;
 510                         if (lvb_len) {
 511
 512                                 req_capsule_set_size(&req->rq_pill,
 513                                                      &RMF_DLM_LVB, RCL_SERVER,
 514                                                      lvb_len);
 515                                 tmplvb = req_capsule_server_get(&req->rq_pill,
 516                                                                  &RMF_DLM_LVB);
 517                                 if (tmplvb == NULL)
 518                                         GOTO(cleanup, rc = -EPROTO);
 519                                 if (lvb != NULL)
 520                                         memcpy(lvb, tmplvb, lvb_len);
 521                         }
 522                 }
 523                 GOTO(cleanup, rc);
 524         }
 525
 526         reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
 527         if (reply == NULL)
 528                 GOTO(cleanup, rc = -EPROTO);
 529
 530         /* lock enqueued on the server */
 531         cleanup_phase = 0;
 532
 533         lock_res_and_lock(lock);
 534         /* Key change rehash lock in per-export hash with new key */
 535         if (exp->exp_lock_hash) {
 536                 cfs_hash_rehash_key(exp->exp_lock_hash,
 537                                     &lock->l_remote_handle,
 538                                     &reply->lock_handle,
 539                                     &lock->l_exp_hash);
 540         } else {
 541                 lock->l_remote_handle = reply->lock_handle;
 542         }
 543
 544         *flags = reply->lock_flags;
 545         lock->l_flags |= reply->lock_flags & LDLM_INHERIT_FLAGS;
 546         /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match()
 547          * to wait with no timeout as well */
 548         lock->l_flags |= reply->lock_flags & LDLM_FL_NO_TIMEOUT;
 549         unlock_res_and_lock(lock);
 550
 551         CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%x\n",
 552                lock, reply->lock_handle.cookie, *flags);
 553
 554         /* If enqueue returned a blocked lock but the completion handler has
 555          * already run, then it fixed up the resource and we don't need to do it
 556          * again. */
 557         if ((*flags) & LDLM_FL_LOCK_CHANGED) {
 558                 int newmode = reply->lock_desc.l_req_mode;
 559                 LASSERT(!is_replay);
 560                 if (newmode && newmode != lock->l_req_mode) {
 561                         LDLM_DEBUG(lock, "server returned different mode %s",
 562                                    ldlm_lockname[newmode]);
 563                         lock->l_req_mode = newmode;
 564                 }
 565
 566                 if (memcmp(reply->lock_desc.l_resource.lr_name.name,
 567                           lock->l_resource->lr_name.name,
 568                           sizeof(struct ldlm_res_id))) {
 569                         CDEBUG(D_INFO, "remote intent success, locking "
 570                                         "(%ld,%ld,%ld) instead of "
 571                                         "(%ld,%ld,%ld)\n",
 572                               (long)reply->lock_desc.l_resource.lr_name.name[0],
 573                               (long)reply->lock_desc.l_resource.lr_name.name[1],
 574                               (long)reply->lock_desc.l_resource.lr_name.name[2],
 575                               (long)lock->l_resource->lr_name.name[0],
 576                               (long)lock->l_resource->lr_name.name[1],
 577                               (long)lock->l_resource->lr_name.name[2]);
 578
 579                         rc = ldlm_lock_change_resource(ns, lock,
 580                                         &reply->lock_desc.l_resource.lr_name);
 581                         if (rc || lock->l_resource == NULL)
 582                                 GOTO(cleanup, rc = -ENOMEM);
 583                         LDLM_DEBUG(lock, "client-side enqueue, new resource");
 584                 }
 585                 if (with_policy)
 586                         if (!(type == LDLM_IBITS && !(exp->exp_connect_flags &
 587                                                     OBD_CONNECT_IBITS)))
 588                                 /* We assume lock type cannot change on server*/
 589                                 ldlm_convert_policy_to_local(exp,
 590                                                 lock->l_resource->lr_type,
 591                                                 &reply->lock_desc.l_policy_data,
 592                                                 &lock->l_policy_data);
 593                 if (type != LDLM_PLAIN)
 594                         LDLM_DEBUG(lock,"client-side enqueue, new policy data");
 595         }
 596
 597         if ((*flags) & LDLM_FL_AST_SENT ||
 598             /* Cancel extent locks as soon as possible on a liblustre client,
 599              * because it cannot handle asynchronous ASTs robustly (see
 600              * bug 7311). */
 601             (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) {
 602                 lock_res_and_lock(lock);
 603                 lock->l_flags |= LDLM_FL_CBPENDING |  LDLM_FL_BL_AST;
 604                 unlock_res_and_lock(lock);
 605                 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
 606         }
 607
 608         /* If the lock has already been granted by a completion AST, don't
 609          * clobber the LVB with an older one. */
 610         if (lvb_len) {
 611                 /* We must lock or a racing completion might update lvb
 612                    without letting us know and we'll clobber the correct value.
 613                    Cannot unlock after the check either, a that still leaves
 614                    a tiny window for completion to get in */
 615                 lock_res_and_lock(lock);
 616                 if (lock->l_req_mode != lock->l_granted_mode) {
 617
 618                         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
 619                                              RCL_SERVER, lvb_len);
 620                         tmplvb = req_capsule_server_get(&req->rq_pill,
 621                                                              &RMF_DLM_LVB);
 622                         if (tmplvb == NULL) {
 623                                 unlock_res_and_lock(lock);
 624                                 GOTO(cleanup, rc = -EPROTO);
 625                         }
 626                         memcpy(lock->l_lvb_data, tmplvb, lvb_len);
 627                 }
 628                 unlock_res_and_lock(lock);
 629         }
 630
 631         if (!is_replay) {
 632                 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
 633                 if (lock->l_completion_ast != NULL) {
 634                         int err = lock->l_completion_ast(lock, *flags, NULL);
 635                         if (!rc)
 636                                 rc = err;
 637                         if (rc)
 638                                 cleanup_phase = 1;
 639                 }
 640         }
 641
 642         if (lvb_len && lvb != NULL) {
 643                 /* Copy the LVB here, and not earlier, because the completion
 644                  * AST (if any) can override what we got in the reply */
 645                 memcpy(lvb, lock->l_lvb_data, lvb_len);
 646         }
 647
 648         LDLM_DEBUG(lock, "client-side enqueue END");
 649         EXIT;
 650 cleanup:
 651         if (cleanup_phase == 1 && rc)
 652                 failed_lock_cleanup(ns, lock, mode);
 653         /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
 654         LDLM_LOCK_PUT(lock);
 655         LDLM_LOCK_RELEASE(lock);
 656         return rc;
 657 }
 658
 659 /* PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
 660  * a single page on the send/receive side. XXX: 512 should be changed
 661  * to more adequate value. */
 662 static inline int ldlm_req_handles_avail(int req_size, int off)
 663 {
 664         int avail;
 665
 666         avail = min_t(int, LDLM_MAXREQSIZE, CFS_PAGE_SIZE - 512) - req_size;
 667         if (likely(avail >= 0))
 668                 avail /= (int)sizeof(struct lustre_handle);
 669         else
 670                 avail = 0;
 671         avail += LDLM_LOCKREQ_HANDLES - off;
 672
 673         return avail;
 674 }
 675
 676 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
 677                                              enum req_location loc,
 678                                              int off)
 679 {
 680         int size = req_capsule_msg_size(pill, loc);
 681         return ldlm_req_handles_avail(size, off);
 682 }
 683
 684 static inline int ldlm_format_handles_avail(struct obd_import *imp,
 685                                             const struct req_format *fmt,
 686                                             enum req_location loc, int off)
 687 {
 688         int size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
 689         return ldlm_req_handles_avail(size, off);
 690 }
 691
 692 /* Cancel lru locks and pack them into the enqueue request. Pack there the given
 693  * @count locks in @cancels. */
 694 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
 695                       int version, int opc, int canceloff,
 696                       cfs_list_t *cancels, int count)
 697 {
 698         struct ldlm_namespace   *ns = exp->exp_obd->obd_namespace;
 699         struct req_capsule      *pill = &req->rq_pill;
 700         struct ldlm_request     *dlm = NULL;
 701         int flags, avail, to_free, pack = 0;
 702         CFS_LIST_HEAD(head);
 703         int rc;
 704         ENTRY;
 705
 706         if (cancels == NULL)
 707                 cancels = &head;
 708         if (exp_connect_cancelset(exp)) {
 709                 /* Estimate the amount of available space in the request. */
 710                 req_capsule_filled_sizes(pill, RCL_CLIENT);
 711                 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
 712
 713                 flags = ns_connect_lru_resize(ns) ?
 714                         LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
 715                 to_free = !ns_connect_lru_resize(ns) &&
 716                           opc == LDLM_ENQUEUE ? 1 : 0;
 717
 718                 /* Cancel lru locks here _only_ if the server supports
 719                  * EARLY_CANCEL. Otherwise we have to send extra CANCEL
 720                  * rpc, what will make us slower. */
 721                 if (avail > count)
 722                         count += ldlm_cancel_lru_local(ns, cancels, to_free,
 723                                                        avail - count, 0, flags);
 724                 if (avail > count)
 725                         pack = count;
 726                 else
 727                         pack = avail;
 728                 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
 729                                      ldlm_request_bufsize(pack, opc));
 730         }
 731
 732         rc = ptlrpc_request_pack(req, version, opc);
 733         if (rc) {
 734                 ldlm_lock_list_put(cancels, l_bl_ast, count);
 735                 RETURN(rc);
 736         }
 737
 738         if (exp_connect_cancelset(exp)) {
 739                 if (canceloff) {
 740                         dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
 741                         LASSERT(dlm);
 742                         /* Skip first lock handler in ldlm_request_pack(),
 743                          * this method will incrment @lock_count according
 744                          * to the lock handle amount actually written to
 745                          * the buffer. */
 746                         dlm->lock_count = canceloff;
 747                 }
 748                 /* Pack into the request @pack lock handles. */
 749                 ldlm_cli_cancel_list(cancels, pack, req, 0);
 750                 /* Prepare and send separate cancel rpc for others. */
 751                 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
 752         } else {
 753                 ldlm_lock_list_put(cancels, l_bl_ast, count);
 754         }
 755         RETURN(0);
 756 }
 757
 758 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
 759                           cfs_list_t *cancels, int count)
 760 {
 761         return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
 762                                  LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
 763 }
 764
 765 /* If a request has some specific initialisation it is passed in @reqp,
 766  * otherwise it is created in ldlm_cli_enqueue.
 767  *
 768  * Supports sync and async requests, pass @async flag accordingly. If a
 769  * request was created in ldlm_cli_enqueue and it is the async request,
 770  * pass it to the caller in @reqp. */
 771 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
 772                      struct ldlm_enqueue_info *einfo,
 773                      const struct ldlm_res_id *res_id,
 774                      ldlm_policy_data_t const *policy, int *flags,
 775                      void *lvb, __u32 lvb_len, struct lustre_handle *lockh,
 776                      int async)
 777 {
 778         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 779         struct ldlm_lock      *lock;
 780         struct ldlm_request   *body;
 781         int                    is_replay = *flags & LDLM_FL_REPLAY;
 782         int                    req_passed_in = 1;
 783         int                    rc, err;
 784         struct ptlrpc_request *req;
 785         ENTRY;
 786
 787         LASSERT(exp != NULL);
 788
 789         /* If we're replaying this lock, just check some invariants.
 790          * If we're creating a new lock, get everything all setup nice. */
 791         if (is_replay) {
 792                 lock = ldlm_handle2lock_long(lockh, 0);
 793                 LASSERT(lock != NULL);
 794                 LDLM_DEBUG(lock, "client-side enqueue START");
 795                 LASSERT(exp == lock->l_conn_export);
 796         } else {
 797                 const struct ldlm_callback_suite cbs = {
 798                         .lcs_completion = einfo->ei_cb_cp,
 799                         .lcs_blocking   = einfo->ei_cb_bl,
 800                         .lcs_glimpse    = einfo->ei_cb_gl,
 801                         .lcs_weigh      = einfo->ei_cb_wg
 802                 };
 803                 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
 804                                         einfo->ei_mode, &cbs, einfo->ei_cbdata,
 805                                         lvb_len);
 806                 if (lock == NULL)
 807                         RETURN(-ENOMEM);
 808                 /* for the local lock, add the reference */
 809                 ldlm_lock_addref_internal(lock, einfo->ei_mode);
 810                 ldlm_lock2handle(lock, lockh);
 811                 if (policy != NULL) {
 812                         /* INODEBITS_INTEROP: If the server does not support
 813                          * inodebits, we will request a plain lock in the
 814                          * descriptor (ldlm_lock2desc() below) but use an
 815                          * inodebits lock internally with both bits set.
 816                          */
 817                         if (einfo->ei_type == LDLM_IBITS &&
 818                             !(exp->exp_connect_flags & OBD_CONNECT_IBITS))
 819                                 lock->l_policy_data.l_inodebits.bits =
 820                                         MDS_INODELOCK_LOOKUP |
 821                                         MDS_INODELOCK_UPDATE;
 822                         else
 823                                 lock->l_policy_data = *policy;
 824                 }
 825
 826                 if (einfo->ei_type == LDLM_EXTENT)
 827                         lock->l_req_extent = policy->l_extent;
 828                 LDLM_DEBUG(lock, "client-side enqueue START");
 829         }
 830
 831         lock->l_conn_export = exp;
 832         lock->l_export = NULL;
 833         lock->l_blocking_ast = einfo->ei_cb_bl;
 834         lock->l_flags |= (*flags & LDLM_FL_NO_LRU);
 835
 836         /* lock not sent to server yet */
 837
 838         if (reqp == NULL || *reqp == NULL) {
 839                 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
 840                                                 &RQF_LDLM_ENQUEUE,
 841                                                 LUSTRE_DLM_VERSION,
 842                                                 LDLM_ENQUEUE);
 843                 if (req == NULL) {
 844                         failed_lock_cleanup(ns, lock, einfo->ei_mode);
 845                         LDLM_LOCK_RELEASE(lock);
 846                         RETURN(-ENOMEM);
 847                 }
 848                 req_passed_in = 0;
 849                 if (reqp)
 850                         *reqp = req;
 851         } else {
 852                 int len;
 853
 854                 req = *reqp;
 855                 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
 856                                            RCL_CLIENT);
 857                 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
 858                          DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
 859         }
 860
 861         /* Dump lock data into the request buffer */
 862         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
 863         ldlm_lock2desc(lock, &body->lock_desc);
 864         body->lock_flags = *flags;
 865         body->lock_handle[0] = *lockh;
 866
 867         /* Continue as normal. */
 868         if (!req_passed_in) {
 869                 if (lvb_len > 0) {
 870                         req_capsule_extend(&req->rq_pill,
 871                                            &RQF_LDLM_ENQUEUE_LVB);
 872                         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
 873                                              RCL_SERVER, lvb_len);
 874                 }
 875                 ptlrpc_request_set_replen(req);
 876         }
 877
 878         /*
 879          * Liblustre client doesn't get extent locks, except for O_APPEND case
 880          * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
 881          * [i_size, OBD_OBJECT_EOF] lock is taken.
 882          */
 883         LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT ||
 884                      policy->l_extent.end == OBD_OBJECT_EOF));
 885
 886         if (async) {
 887                 LASSERT(reqp != NULL);
 888                 RETURN(0);
 889         }
 890
 891         LDLM_DEBUG(lock, "sending request");
 892
 893         rc = ptlrpc_queue_wait(req);
 894
 895         err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
 896                                     einfo->ei_mode, flags, lvb, lvb_len,
 897                                     lockh, rc);
 898
 899         /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
 900          * one reference that we took */
 901         if (err == -ENOLCK)
 902                 LDLM_LOCK_RELEASE(lock);
 903         else
 904                 rc = err;
 905
 906         if (!req_passed_in && req != NULL) {
 907                 ptlrpc_req_finished(req);
 908                 if (reqp)
 909                         *reqp = NULL;
 910         }
 911
 912         RETURN(rc);
 913 }
 914
 915 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
 916                                   __u32 *flags)
 917 {
 918         struct ldlm_resource *res;
 919         int rc;
 920         ENTRY;
 921         if (ns_is_client(ldlm_lock_to_ns(lock))) {
 922                 CERROR("Trying to cancel local lock\n");
 923                 LBUG();
 924         }
 925         LDLM_DEBUG(lock, "client-side local convert");
 926
 927         res = ldlm_lock_convert(lock, new_mode, flags);
 928         if (res) {
 929                 ldlm_reprocess_all(res);
 930                 rc = 0;
 931         } else {
 932                 rc = EDEADLOCK;
 933         }
 934         LDLM_DEBUG(lock, "client-side local convert handler END");
 935         LDLM_LOCK_PUT(lock);
 936         RETURN(rc);
 937 }
 938
 939 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
 940  * conversion of locks which are on the waiting or converting queue */
 941 /* Caller of this code is supposed to take care of lock readers/writers
 942    accounting */
 943 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags)
 944 {
 945         struct ldlm_request   *body;
 946         struct ldlm_reply     *reply;
 947         struct ldlm_lock      *lock;
 948         struct ldlm_resource  *res;
 949         struct ptlrpc_request *req;
 950         int                    rc;
 951         ENTRY;
 952
 953         lock = ldlm_handle2lock(lockh);
 954         if (!lock) {
 955                 LBUG();
 956                 RETURN(-EINVAL);
 957         }
 958         *flags = 0;
 959
 960         if (lock->l_conn_export == NULL)
 961                 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
 962
 963         LDLM_DEBUG(lock, "client-side convert");
 964
 965         req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export),
 966                                         &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
 967                                         LDLM_CONVERT);
 968         if (req == NULL) {
 969                 LDLM_LOCK_PUT(lock);
 970                 RETURN(-ENOMEM);
 971         }
 972
 973         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
 974         body->lock_handle[0] = lock->l_remote_handle;
 975
 976         body->lock_desc.l_req_mode = new_mode;
 977         body->lock_flags = *flags;
 978
 979
 980         ptlrpc_request_set_replen(req);
 981         rc = ptlrpc_queue_wait(req);
 982         if (rc != ELDLM_OK)
 983                 GOTO(out, rc);
 984
 985         reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
 986         if (reply == NULL)
 987                 GOTO(out, rc = -EPROTO);
 988
 989         if (req->rq_status)
 990                 GOTO(out, rc = req->rq_status);
 991
 992         res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
 993         if (res != NULL) {
 994                 ldlm_reprocess_all(res);
 995                 /* Go to sleep until the lock is granted. */
 996                 /* FIXME: or cancelled. */
 997                 if (lock->l_completion_ast) {
 998                         rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
 999                                                     NULL);
1000                         if (rc)
1001                                 GOTO(out, rc);
1002                 }
1003         } else {
1004                 rc = EDEADLOCK;
1005         }
1006         EXIT;
1007  out:
1008         LDLM_LOCK_PUT(lock);
1009         ptlrpc_req_finished(req);
1010         return rc;
1011 }
1012
1013 /* Cancel locks locally.
1014  * Returns:
1015  * LDLM_FL_LOCAL_ONLY if tere is no need in a CANCEL rpc to the server;
1016  * LDLM_FL_CANCELING otherwise;
1017  * LDLM_FL_BL_AST if there is a need in a separate CANCEL rpc. */
1018 static int ldlm_cli_cancel_local(struct ldlm_lock *lock)
1019 {
1020         int rc = LDLM_FL_LOCAL_ONLY;
1021         ENTRY;
1022
1023         if (lock->l_conn_export) {
1024                 int local_only;
1025
1026                 LDLM_DEBUG(lock, "client-side cancel");
1027                 /* Set this flag to prevent others from getting new references*/
1028                 lock_res_and_lock(lock);
1029                 lock->l_flags |= LDLM_FL_CBPENDING;
1030                 local_only = (lock->l_flags &
1031                               (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
1032                 ldlm_cancel_callback(lock);
1033                 rc = (lock->l_flags & LDLM_FL_BL_AST) ?
1034                         LDLM_FL_BL_AST : LDLM_FL_CANCELING;
1035                 unlock_res_and_lock(lock);
1036
1037                 if (local_only) {
1038                         CDEBUG(D_DLMTRACE, "not sending request (at caller's "
1039                                "instruction)\n");
1040                         rc = LDLM_FL_LOCAL_ONLY;
1041                 }
1042                 ldlm_lock_cancel(lock);
1043         } else {
1044                 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1045                         LDLM_ERROR(lock, "Trying to cancel local lock");
1046                         LBUG();
1047                 }
1048                 LDLM_DEBUG(lock, "server-side local cancel");
1049                 ldlm_lock_cancel(lock);
1050                 ldlm_reprocess_all(lock->l_resource);
1051         }
1052
1053         RETURN(rc);
1054 }
1055
1056 /* Pack @count locks in @head into ldlm_request buffer at the offset @off,
1057    of the request @req. */
1058 static void ldlm_cancel_pack(struct ptlrpc_request *req,
1059                              cfs_list_t *head, int count)
1060 {
1061         struct ldlm_request *dlm;
1062         struct ldlm_lock *lock;
1063         int max, packed = 0;
1064         ENTRY;
1065
1066         dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1067         LASSERT(dlm != NULL);
1068
1069         /* Check the room in the request buffer. */
1070         max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
1071                 sizeof(struct ldlm_request);
1072         max /= sizeof(struct lustre_handle);
1073         max += LDLM_LOCKREQ_HANDLES;
1074         LASSERT(max >= dlm->lock_count + count);
1075
1076         /* XXX: it would be better to pack lock handles grouped by resource.
1077          * so that the server cancel would call filter_lvbo_update() less
1078          * frequently. */
1079         cfs_list_for_each_entry(lock, head, l_bl_ast) {
1080                 if (!count--)
1081                         break;
1082                 LASSERT(lock->l_conn_export);
1083                 /* Pack the lock handle to the given request buffer. */
1084                 LDLM_DEBUG(lock, "packing");
1085                 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
1086                 packed++;
1087         }
1088         CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
1089         EXIT;
1090 }
1091
1092 /* Prepare and send a batched cancel rpc, it will include count lock handles
1093  * of locks given in @head. */
1094 int ldlm_cli_cancel_req(struct obd_export *exp, cfs_list_t *cancels,
1095                         int count, ldlm_cancel_flags_t flags)
1096 {
1097         struct ptlrpc_request *req = NULL;
1098         struct obd_import *imp;
1099         int free, sent = 0;
1100         int rc = 0;
1101         ENTRY;
1102
1103         LASSERT(exp != NULL);
1104         LASSERT(count > 0);
1105
1106         CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, cfs_fail_val);
1107
1108         if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
1109                 RETURN(count);
1110
1111         free = ldlm_format_handles_avail(class_exp2cliimp(exp),
1112                                          &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
1113         if (count > free)
1114                 count = free;
1115
1116         while (1) {
1117                 imp = class_exp2cliimp(exp);
1118                 if (imp == NULL || imp->imp_invalid) {
1119                         CDEBUG(D_DLMTRACE,
1120                                "skipping cancel on invalid import %p\n", imp);
1121                         RETURN(count);
1122                 }
1123
1124                 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
1125                 if (req == NULL)
1126                         GOTO(out, rc = -ENOMEM);
1127
1128                 req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
1129                 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
1130                                      ldlm_request_bufsize(count, LDLM_CANCEL));
1131
1132                 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
1133                 if (rc) {
1134                         ptlrpc_request_free(req);
1135                         GOTO(out, rc);
1136                 }
1137                 req->rq_no_resend = 1;
1138                 req->rq_no_delay = 1;
1139
1140                 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1141                 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1142                 ptlrpc_at_set_req_timeout(req);
1143
1144                 ldlm_cancel_pack(req, cancels, count);
1145
1146                 ptlrpc_request_set_replen(req);
1147                 if (flags & LCF_ASYNC) {
1148                         ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
1149                         sent = count;
1150                         GOTO(out, 0);
1151                 } else {
1152                         rc = ptlrpc_queue_wait(req);
1153                 }
1154                 if (rc == ESTALE) {
1155                         CDEBUG(D_DLMTRACE, "client/server (nid %s) "
1156                                "out of sync -- not fatal\n",
1157                                libcfs_nid2str(req->rq_import->
1158                                               imp_connection->c_peer.nid));
1159                         rc = 0;
1160                 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1161                            req->rq_import_generation == imp->imp_generation) {
1162                         ptlrpc_req_finished(req);
1163                         continue;
1164                 } else if (rc != ELDLM_OK) {
1165                         CERROR("Got rc %d from cancel RPC: canceling "
1166                                "anyway\n", rc);
1167                         break;
1168                 }
1169                 sent = count;
1170                 break;
1171         }
1172
1173         ptlrpc_req_finished(req);
1174         EXIT;
1175 out:
1176         return sent ? sent : rc;
1177 }
1178
1179 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1180 {
1181         LASSERT(imp != NULL);
1182         return &imp->imp_obd->obd_namespace->ns_pool;
1183 }
1184
1185 /**
1186  * Update client's obd pool related fields with new SLV and Limit from \a req.
1187  */
1188 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1189 {
1190         struct obd_device *obd;
1191         __u64 new_slv;
1192         __u32 new_limit;
1193         ENTRY;
1194         if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1195                      !imp_connect_lru_resize(req->rq_import)))
1196         {
1197                 /*
1198                  * Do nothing for corner cases.
1199                  */
1200                 RETURN(0);
1201         }
1202
1203         /*
1204          * In some cases RPC may contain slv and limit zeroed out. This is
1205          * the case when server does not support lru resize feature. This is
1206          * also possible in some recovery cases when server side reqs have no
1207          * ref to obd export and thus access to server side namespace is no
1208          * possible.
1209          */
1210         if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1211             lustre_msg_get_limit(req->rq_repmsg) == 0) {
1212                 DEBUG_REQ(D_HA, req, "Zero SLV or Limit found "
1213                           "(SLV: "LPU64", Limit: %u)",
1214                           lustre_msg_get_slv(req->rq_repmsg),
1215                           lustre_msg_get_limit(req->rq_repmsg));
1216                 RETURN(0);
1217         }
1218
1219         new_limit = lustre_msg_get_limit(req->rq_repmsg);
1220         new_slv = lustre_msg_get_slv(req->rq_repmsg);
1221         obd = req->rq_import->imp_obd;
1222
1223         /*
1224          * Set new SLV and Limit to obd fields to make accessible for pool
1225          * thread. We do not access obd_namespace and pool directly here
1226          * as there is no reliable way to make sure that they are still
1227          * alive in cleanup time. Evil races are possible which may cause
1228          * oops in that time.
1229          */
1230         cfs_write_lock(&obd->obd_pool_lock);
1231         obd->obd_pool_slv = new_slv;
1232         obd->obd_pool_limit = new_limit;
1233         cfs_write_unlock(&obd->obd_pool_lock);
1234
1235         RETURN(0);
1236 }
1237 EXPORT_SYMBOL(ldlm_cli_update_pool);
1238
1239 int ldlm_cli_cancel(struct lustre_handle *lockh)
1240 {
1241         struct obd_export *exp;
1242         int avail, flags, count = 1, rc = 0;
1243         struct ldlm_namespace *ns;
1244         struct ldlm_lock *lock;
1245         CFS_LIST_HEAD(cancels);
1246         ENTRY;
1247
1248         /* concurrent cancels on the same handle can happen */
1249         lock = ldlm_handle2lock_long(lockh, LDLM_FL_CANCELING);
1250         if (lock == NULL) {
1251                 LDLM_DEBUG_NOLOCK("lock is already being destroyed\n");
1252                 RETURN(0);
1253         }
1254
1255         rc = ldlm_cli_cancel_local(lock);
1256         if (rc < 0 || rc == LDLM_FL_LOCAL_ONLY) {
1257                 LDLM_LOCK_RELEASE(lock);
1258                 RETURN(rc < 0 ? rc : 0);
1259         }
1260         /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1261          * rpc which goes to canceld portal, so we can cancel other lru locks
1262          * here and send them all as one LDLM_CANCEL rpc. */
1263         LASSERT(cfs_list_empty(&lock->l_bl_ast));
1264         cfs_list_add(&lock->l_bl_ast, &cancels);
1265
1266         exp = lock->l_conn_export;
1267         if (exp_connect_cancelset(exp)) {
1268                 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1269                                                   &RQF_LDLM_CANCEL,
1270                                                   RCL_CLIENT, 0);
1271                 LASSERT(avail > 0);
1272
1273                 ns = ldlm_lock_to_ns(lock);
1274                 flags = ns_connect_lru_resize(ns) ?
1275                         LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
1276                 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1277                                                LCF_BL_AST, flags);
1278         }
1279         ldlm_cli_cancel_list(&cancels, count, NULL, 0);
1280         RETURN(0);
1281 }
1282
1283 /* XXX until we will have compound requests and can cut cancels from generic rpc
1284  * we need send cancels with LDLM_FL_BL_AST flag as separate rpc */
1285 int ldlm_cli_cancel_list_local(cfs_list_t *cancels, int count,
1286                                ldlm_cancel_flags_t flags)
1287 {
1288         CFS_LIST_HEAD(head);
1289         struct ldlm_lock *lock, *next;
1290         int left = 0, bl_ast = 0, rc;
1291
1292         left = count;
1293         cfs_list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1294                 if (left-- == 0)
1295                         break;
1296
1297                 if (flags & LCF_LOCAL) {
1298                         rc = LDLM_FL_LOCAL_ONLY;
1299                         ldlm_lock_cancel(lock);
1300                 } else {
1301                         rc = ldlm_cli_cancel_local(lock);
1302                 }
1303                 if (!(flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1304                         LDLM_DEBUG(lock, "Cancel lock separately");
1305                         cfs_list_del_init(&lock->l_bl_ast);
1306                         cfs_list_add(&lock->l_bl_ast, &head);
1307                         bl_ast ++;
1308                         continue;
1309                 }
1310                 if (rc == LDLM_FL_LOCAL_ONLY) {
1311                         /* CANCEL RPC should not be sent to server. */
1312                         cfs_list_del_init(&lock->l_bl_ast);
1313                         LDLM_LOCK_RELEASE(lock);
1314                         count--;
1315                 }
1316
1317         }
1318         if (bl_ast > 0) {
1319                 count -= bl_ast;
1320                 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1321         }
1322
1323         RETURN(count);
1324 }
1325
1326 /**
1327  * Cancel as many locks as possible w/o sending any rpcs (e.g. to write back
1328  * dirty data, to close a file, ...) or waiting for any rpcs in-flight (e.g.
1329  * readahead requests, ...)
1330  */
1331 static ldlm_policy_res_t ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns,
1332                                                     struct ldlm_lock *lock,
1333                                                     int unused, int added,
1334                                                     int count)
1335 {
1336         ldlm_policy_res_t result = LDLM_POLICY_CANCEL_LOCK;
1337         ldlm_cancel_for_recovery cb = ns->ns_cancel_for_recovery;
1338         lock_res_and_lock(lock);
1339
1340         /* don't check added & count since we want to process all locks
1341          * from unused list */
1342         switch (lock->l_resource->lr_type) {
1343                 case LDLM_EXTENT:
1344                 case LDLM_IBITS:
1345                         if (cb && cb(lock))
1346                                 break;
1347                 default:
1348                         result = LDLM_POLICY_SKIP_LOCK;
1349                         lock->l_flags |= LDLM_FL_SKIPPED;
1350                         break;
1351         }
1352
1353         unlock_res_and_lock(lock);
1354         RETURN(result);
1355 }
1356
1357 /**
1358  * Callback function for lru-resize policy. Makes decision whether to keep
1359  * \a lock in LRU for current \a LRU size \a unused, added in current scan
1360  * \a added and number of locks to be preferably canceled \a count.
1361  *
1362  * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1363  *
1364  * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1365  */
1366 static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1367                                                  struct ldlm_lock *lock,
1368                                                  int unused, int added,
1369                                                  int count)
1370 {
1371         cfs_time_t cur = cfs_time_current();
1372         struct ldlm_pool *pl = &ns->ns_pool;
1373         __u64 slv, lvf, lv;
1374         cfs_time_t la;
1375
1376         /*
1377          * Stop lru processing when we reached passed @count or checked all
1378          * locks in lru.
1379          */
1380         if (count && added >= count)
1381                 return LDLM_POLICY_KEEP_LOCK;
1382
1383         slv = ldlm_pool_get_slv(pl);
1384         lvf = ldlm_pool_get_lvf(pl);
1385         la = cfs_duration_sec(cfs_time_sub(cur,
1386                               lock->l_last_used));
1387
1388         /*
1389          * Stop when slv is not yet come from server or lv is smaller than
1390          * it is.
1391          */
1392         lv = lvf * la * unused;
1393
1394         /*
1395          * Inform pool about current CLV to see it via proc.
1396          */
1397         ldlm_pool_set_clv(pl, lv);
1398         return (slv == 0 || lv < slv) ?
1399                 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1400 }
1401
1402 /**
1403  * Callback function for proc used policy. Makes decision whether to keep
1404  * \a lock in LRU for current \a LRU size \a unused, added in current scan
1405  * \a added and number of locks to be preferably canceled \a count.
1406  *
1407  * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1408  *
1409  * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1410  */
1411 static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1412                                                    struct ldlm_lock *lock,
1413                                                    int unused, int added,
1414                                                    int count)
1415 {
1416         /*
1417          * Stop lru processing when we reached passed @count or checked all
1418          * locks in lru.
1419          */
1420         return (added >= count) ?
1421                 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1422 }
1423
1424 /**
1425  * Callback function for aged policy. Makes decision whether to keep
1426  * \a lock in LRU for current \a LRU size \a unused, added in current scan
1427  * \a added and number of locks to be preferably canceled \a count.
1428  *
1429  * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1430  *
1431  * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1432  */
1433 static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1434                                                  struct ldlm_lock *lock,
1435                                                  int unused, int added,
1436                                                  int count)
1437 {
1438         /*
1439          * Stop lru processing if young lock is found and we reached passed
1440          * @count.
1441          */
1442         return ((added >= count) &&
1443                 cfs_time_before(cfs_time_current(),
1444                                 cfs_time_add(lock->l_last_used,
1445                                              ns->ns_max_age))) ?
1446                 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1447 }
1448
1449 /**
1450  * Callback function for default policy. Makes decision whether to keep
1451  * \a lock in LRU for current \a LRU size \a unused, added in current scan
1452  * \a added and number of locks to be preferably canceled \a count.
1453  *
1454  * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1455  *
1456  * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1457  */
1458 static ldlm_policy_res_t ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1459                                                     struct ldlm_lock *lock,
1460                                                     int unused, int added,
1461                                                     int count)
1462 {
1463         /*
1464          * Stop lru processing when we reached passed @count or checked all
1465          * locks in lru.
1466          */
1467         return (added >= count) ?
1468                 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1469 }
1470
1471 typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *,
1472                                                       struct ldlm_lock *, int,
1473                                                       int, int);
1474
1475 static ldlm_cancel_lru_policy_t
1476 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
1477 {
1478         if (flags & LDLM_CANCEL_NO_WAIT)
1479                 return ldlm_cancel_no_wait_policy;
1480
1481         if (ns_connect_lru_resize(ns)) {
1482                 if (flags & LDLM_CANCEL_SHRINK)
1483                         /* We kill passed number of old locks. */
1484                         return ldlm_cancel_passed_policy;
1485                 else if (flags & LDLM_CANCEL_LRUR)
1486                         return ldlm_cancel_lrur_policy;
1487                 else if (flags & LDLM_CANCEL_PASSED)
1488                         return ldlm_cancel_passed_policy;
1489         } else {
1490                 if (flags & LDLM_CANCEL_AGED)
1491                         return ldlm_cancel_aged_policy;
1492         }
1493
1494         return ldlm_cancel_default_policy;
1495 }
1496
1497 /* - Free space in lru for @count new locks,
1498  *   redundant unused locks are canceled locally;
1499  * - also cancel locally unused aged locks;
1500  * - do not cancel more than @max locks;
1501  * - GET the found locks and add them into the @cancels list.
1502  *
1503  * A client lock can be added to the l_bl_ast list only when it is
1504  * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing CANCEL.
1505  * There are the following use cases: ldlm_cancel_resource_local(),
1506  * ldlm_cancel_lru_local() and ldlm_cli_cancel(), which check&set this
1507  * flag properly. As any attempt to cancel a lock rely on this flag,
1508  * l_bl_ast list is accessed later without any special locking.
1509  *
1510  * Calling policies for enabled lru resize:
1511  * ----------------------------------------
1512  * flags & LDLM_CANCEL_LRUR - use lru resize policy (SLV from server) to
1513  *                            cancel not more than @count locks;
1514  *
1515  * flags & LDLM_CANCEL_PASSED - cancel @count number of old locks (located at
1516  *                              the beginning of lru list);
1517  *
1518  * flags & LDLM_CANCEL_SHRINK - cancel not more than @count locks according to
1519  *                              memory pressre policy function;
1520  *
1521  * flags & LDLM_CANCEL_AGED -   cancel alocks according to "aged policy".
1522  *
1523  * flags & LDLM_CANCEL_NO_WAIT - cancel as many unused locks as possible
1524  *                               (typically before replaying locks) w/o
1525  *                               sending any rpcs or waiting for any
1526  *                               outstanding rpc to complete.
1527  */
1528 static int ldlm_prepare_lru_list(struct ldlm_namespace *ns, cfs_list_t *cancels,
1529                                  int count, int max, int flags)
1530 {
1531         ldlm_cancel_lru_policy_t pf;
1532         struct ldlm_lock *lock, *next;
1533         int added = 0, unused, remained;
1534         ENTRY;
1535
1536         cfs_spin_lock(&ns->ns_lock);
1537         unused = ns->ns_nr_unused;
1538         remained = unused;
1539
1540         if (!ns_connect_lru_resize(ns))
1541                 count += unused - ns->ns_max_unused;
1542
1543         pf = ldlm_cancel_lru_policy(ns, flags);
1544         LASSERT(pf != NULL);
1545
1546         while (!cfs_list_empty(&ns->ns_unused_list)) {
1547                 ldlm_policy_res_t result;
1548
1549                 /* all unused locks */
1550                 if (remained-- <= 0)
1551                         break;
1552
1553                 /* For any flags, stop scanning if @max is reached. */
1554                 if (max && added >= max)
1555                         break;
1556
1557                 cfs_list_for_each_entry_safe(lock, next, &ns->ns_unused_list,
1558                                              l_lru){
1559                         /* No locks which got blocking requests. */
1560                         LASSERT(!(lock->l_flags & LDLM_FL_BL_AST));
1561
1562                         if (flags & LDLM_CANCEL_NO_WAIT &&
1563                             lock->l_flags & LDLM_FL_SKIPPED)
1564                                 /* already processed */
1565                                 continue;
1566
1567                         /* Somebody is already doing CANCEL. No need in this
1568                          * lock in lru, do not traverse it again. */
1569                         if (!(lock->l_flags & LDLM_FL_CANCELING))
1570                                 break;
1571
1572                         ldlm_lock_remove_from_lru_nolock(lock);
1573                 }
1574                 if (&lock->l_lru == &ns->ns_unused_list)
1575                         break;
1576
1577                 LDLM_LOCK_GET(lock);
1578                 cfs_spin_unlock(&ns->ns_lock);
1579                 lu_ref_add(&lock->l_reference, __FUNCTION__, cfs_current());
1580
1581                 /* Pass the lock through the policy filter and see if it
1582                  * should stay in lru.
1583                  *
1584                  * Even for shrinker policy we stop scanning if
1585                  * we find a lock that should stay in the cache.
1586                  * We should take into account lock age anyway
1587                  * as new lock even if it is small of weight is
1588                  * valuable resource.
1589                  *
1590                  * That is, for shrinker policy we drop only
1591                  * old locks, but additionally chose them by
1592                  * their weight. Big extent locks will stay in
1593                  * the cache. */
1594                 result = pf(ns, lock, unused, added, count);
1595                 if (result == LDLM_POLICY_KEEP_LOCK) {
1596                         lu_ref_del(&lock->l_reference,
1597                                    __FUNCTION__, cfs_current());
1598                         LDLM_LOCK_RELEASE(lock);
1599                         cfs_spin_lock(&ns->ns_lock);
1600                         break;
1601                 }
1602                 if (result == LDLM_POLICY_SKIP_LOCK) {
1603                         lu_ref_del(&lock->l_reference,
1604                                    __FUNCTION__, cfs_current());
1605                         LDLM_LOCK_RELEASE(lock);
1606                         cfs_spin_lock(&ns->ns_lock);
1607                         continue;
1608                 }
1609
1610                 lock_res_and_lock(lock);
1611                 /* Check flags again under the lock. */
1612                 if ((lock->l_flags & LDLM_FL_CANCELING) ||
1613                     (ldlm_lock_remove_from_lru(lock) == 0)) {
1614                         /* other thread is removing lock from lru or
1615                          * somebody is already doing CANCEL or
1616                          * there is a blocking request which will send
1617                          * cancel by itseft or the lock is matched
1618                          * is already not unused. */
1619                         unlock_res_and_lock(lock);
1620                         lu_ref_del(&lock->l_reference,
1621                                    __FUNCTION__, cfs_current());
1622                         LDLM_LOCK_RELEASE(lock);
1623                         cfs_spin_lock(&ns->ns_lock);
1624                         continue;
1625                 }
1626                 LASSERT(!lock->l_readers && !lock->l_writers);
1627
1628                 /* If we have chosen to cancel this lock voluntarily, we
1629                  * better send cancel notification to server, so that it
1630                  * frees appropriate state. This might lead to a race
1631                  * where while we are doing cancel here, server is also
1632                  * silently cancelling this lock. */
1633                 lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK;
1634
1635                 /* Setting the CBPENDING flag is a little misleading,
1636                  * but prevents an important race; namely, once
1637                  * CBPENDING is set, the lock can accumulate no more
1638                  * readers/writers. Since readers and writers are
1639                  * already zero here, ldlm_lock_decref() won't see
1640                  * this flag and call l_blocking_ast */
1641                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1642
1643                 /* We can't re-add to l_lru as it confuses the
1644                  * refcounting in ldlm_lock_remove_from_lru() if an AST
1645                  * arrives after we drop lr_lock below. We use l_bl_ast
1646                  * and can't use l_pending_chain as it is used both on
1647                  * server and client nevertheless bug 5666 says it is
1648                  * used only on server */
1649                 LASSERT(cfs_list_empty(&lock->l_bl_ast));
1650                 cfs_list_add(&lock->l_bl_ast, cancels);
1651                 unlock_res_and_lock(lock);
1652                 lu_ref_del(&lock->l_reference, __FUNCTION__, cfs_current());
1653                 cfs_spin_lock(&ns->ns_lock);
1654                 added++;
1655                 unused--;
1656         }
1657         cfs_spin_unlock(&ns->ns_lock);
1658         RETURN(added);
1659 }
1660
1661 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, cfs_list_t *cancels,
1662                           int count, int max, ldlm_cancel_flags_t cancel_flags,
1663                           int flags)
1664 {
1665         int added;
1666         added = ldlm_prepare_lru_list(ns, cancels, count, max, flags);
1667         if (added <= 0)
1668                 return added;
1669         return ldlm_cli_cancel_list_local(cancels, added, cancel_flags);
1670 }
1671
1672 /* when called with LDLM_ASYNC the blocking callback will be handled
1673  * in a thread and this function will return after the thread has been
1674  * asked to call the callback.  when called with LDLM_SYNC the blocking
1675  * callback will be performed in this function. */
1676 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr, ldlm_sync_t mode,
1677                     int flags)
1678 {
1679         CFS_LIST_HEAD(cancels);
1680         int count, rc;
1681         ENTRY;
1682
1683 #ifndef __KERNEL__
1684         mode = LDLM_SYNC; /* force to be sync in user space */
1685 #endif
1686         /* Just prepare the list of locks, do not actually cancel them yet.
1687          * Locks are cancelled later in a separate thread. */
1688         count = ldlm_prepare_lru_list(ns, &cancels, nr, 0, flags);
1689         rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, mode);
1690         if (rc == 0)
1691                 RETURN(count);
1692
1693         RETURN(0);
1694 }
1695
1696 /* Find and cancel locally unused locks found on resource, matched to the
1697  * given policy, mode. GET the found locks and add them into the @cancels
1698  * list. */
1699 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1700                                cfs_list_t *cancels,
1701                                ldlm_policy_data_t *policy,
1702                                ldlm_mode_t mode, int lock_flags,
1703                                ldlm_cancel_flags_t cancel_flags, void *opaque)
1704 {
1705         struct ldlm_lock *lock;
1706         int count = 0;
1707         ENTRY;
1708
1709         lock_res(res);
1710         cfs_list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1711                 if (opaque != NULL && lock->l_ast_data != opaque) {
1712                         LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1713                                    lock->l_ast_data, opaque);
1714                         //LBUG();
1715                         continue;
1716                 }
1717
1718                 if (lock->l_readers || lock->l_writers)
1719                         continue;
1720
1721                 /* If somebody is already doing CANCEL, or blocking ast came,
1722                  * skip this lock. */
1723                 if (lock->l_flags & LDLM_FL_BL_AST ||
1724                     lock->l_flags & LDLM_FL_CANCELING)
1725                         continue;
1726
1727                 if (lockmode_compat(lock->l_granted_mode, mode))
1728                         continue;
1729
1730                 /* If policy is given and this is IBITS lock, add to list only
1731                  * those locks that match by policy. */
1732                 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1733                     !(lock->l_policy_data.l_inodebits.bits &
1734                       policy->l_inodebits.bits))
1735                         continue;
1736
1737                 /* See CBPENDING comment in ldlm_cancel_lru */
1738                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1739                                  lock_flags;
1740
1741                 LASSERT(cfs_list_empty(&lock->l_bl_ast));
1742                 cfs_list_add(&lock->l_bl_ast, cancels);
1743                 LDLM_LOCK_GET(lock);
1744                 count++;
1745         }
1746         unlock_res(res);
1747
1748         RETURN(ldlm_cli_cancel_list_local(cancels, count, cancel_flags));
1749 }
1750
1751 /* If @req is NULL, send CANCEL request to server with handles of locks
1752  * in the @cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1753  * separately per lock.
1754  * If @req is not NULL, put handles of locks in @cancels into the request
1755  * buffer at the offset @off.
1756  * Destroy @cancels at the end. */
1757 int ldlm_cli_cancel_list(cfs_list_t *cancels, int count,
1758                          struct ptlrpc_request *req, ldlm_cancel_flags_t flags)
1759 {
1760         struct ldlm_lock *lock;
1761         int res = 0;
1762         ENTRY;
1763
1764         if (cfs_list_empty(cancels) || count == 0)
1765                 RETURN(0);
1766
1767         /* XXX: requests (both batched and not) could be sent in parallel.
1768          * Usually it is enough to have just 1 RPC, but it is possible that
1769          * there are to many locks to be cancelled in LRU or on a resource.
1770          * It would also speed up the case when the server does not support
1771          * the feature. */
1772         while (count > 0) {
1773                 LASSERT(!cfs_list_empty(cancels));
1774                 lock = cfs_list_entry(cancels->next, struct ldlm_lock,
1775                                       l_bl_ast);
1776                 LASSERT(lock->l_conn_export);
1777
1778                 if (exp_connect_cancelset(lock->l_conn_export)) {
1779                         res = count;
1780                         if (req)
1781                                 ldlm_cancel_pack(req, cancels, count);
1782                         else
1783                                 res = ldlm_cli_cancel_req(lock->l_conn_export,
1784                                                           cancels, count,
1785                                                           flags);
1786                 } else {
1787                         res = ldlm_cli_cancel_req(lock->l_conn_export,
1788                                                   cancels, 1, flags);
1789                 }
1790
1791                 if (res < 0) {
1792                         CERROR("ldlm_cli_cancel_list: %d\n", res);
1793                         res = count;
1794                 }
1795
1796                 count -= res;
1797                 ldlm_lock_list_put(cancels, l_bl_ast, res);
1798         }
1799         LASSERT(count == 0);
1800         RETURN(0);
1801 }
1802
1803 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1804                                     const struct ldlm_res_id *res_id,
1805                                     ldlm_policy_data_t *policy,
1806                                     ldlm_mode_t mode,
1807                                     ldlm_cancel_flags_t flags,
1808                                     void *opaque)
1809 {
1810         struct ldlm_resource *res;
1811         CFS_LIST_HEAD(cancels);
1812         int count;
1813         int rc;
1814         ENTRY;
1815
1816         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1817         if (res == NULL) {
1818                 /* This is not a problem. */
1819                 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id->name[0]);
1820                 RETURN(0);
1821         }
1822
1823         LDLM_RESOURCE_ADDREF(res);
1824         count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
1825                                            0, flags | LCF_BL_AST, opaque);
1826         rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
1827         if (rc != ELDLM_OK)
1828                 CERROR("ldlm_cli_cancel_unused_resource: %d\n", rc);
1829
1830         LDLM_RESOURCE_DELREF(res);
1831         ldlm_resource_putref(res);
1832         RETURN(0);
1833 }
1834
1835 struct ldlm_cli_cancel_arg {
1836         int     lc_flags;
1837         void   *lc_opaque;
1838 };
1839
1840 static int ldlm_cli_hash_cancel_unused(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1841                                        cfs_hlist_node_t *hnode, void *arg)
1842 {
1843         struct ldlm_resource           *res = cfs_hash_object(hs, hnode);
1844         struct ldlm_cli_cancel_arg     *lc = arg;
1845         int                             rc;
1846
1847         rc = ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name,
1848                                              NULL, LCK_MINMODE,
1849                                              lc->lc_flags, lc->lc_opaque);
1850         if (rc != 0) {
1851                 CERROR("ldlm_cli_cancel_unused ("LPU64"): %d\n",
1852                        res->lr_name.name[0], rc);
1853         }
1854         /* must return 0 for hash iteration */
1855         return 0;
1856 }
1857
1858 /* Cancel all locks on a namespace (or a specific resource, if given)
1859  * that have 0 readers/writers.
1860  *
1861  * If flags & LCF_LOCAL, throw the locks away without trying
1862  * to notify the server. */
1863 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1864                            const struct ldlm_res_id *res_id,
1865                            ldlm_cancel_flags_t flags, void *opaque)
1866 {
1867         struct ldlm_cli_cancel_arg arg = {
1868                 .lc_flags       = flags,
1869                 .lc_opaque      = opaque,
1870         };
1871
1872         ENTRY;
1873
1874         if (ns == NULL)
1875                 RETURN(ELDLM_OK);
1876
1877         if (res_id != NULL) {
1878                 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
1879                                                        LCK_MINMODE, flags,
1880                                                        opaque));
1881         } else {
1882                 cfs_hash_for_each_nolock(ns->ns_rs_hash,
1883                                          ldlm_cli_hash_cancel_unused, &arg);
1884                 RETURN(ELDLM_OK);
1885         }
1886 }
1887
1888 /* Lock iterators. */
1889
1890 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
1891                           void *closure)
1892 {
1893         cfs_list_t *tmp, *next;
1894         struct ldlm_lock *lock;
1895         int rc = LDLM_ITER_CONTINUE;
1896
1897         ENTRY;
1898
1899         if (!res)
1900                 RETURN(LDLM_ITER_CONTINUE);
1901
1902         lock_res(res);
1903         cfs_list_for_each_safe(tmp, next, &res->lr_granted) {
1904                 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1905
1906                 if (iter(lock, closure) == LDLM_ITER_STOP)
1907                         GOTO(out, rc = LDLM_ITER_STOP);
1908         }
1909
1910         cfs_list_for_each_safe(tmp, next, &res->lr_converting) {
1911                 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1912
1913                 if (iter(lock, closure) == LDLM_ITER_STOP)
1914                         GOTO(out, rc = LDLM_ITER_STOP);
1915         }
1916
1917         cfs_list_for_each_safe(tmp, next, &res->lr_waiting) {
1918                 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1919
1920                 if (iter(lock, closure) == LDLM_ITER_STOP)
1921                         GOTO(out, rc = LDLM_ITER_STOP);
1922         }
1923  out:
1924         unlock_res(res);
1925         RETURN(rc);
1926 }
1927
1928 struct iter_helper_data {
1929         ldlm_iterator_t iter;
1930         void *closure;
1931 };
1932
1933 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
1934 {
1935         struct iter_helper_data *helper = closure;
1936         return helper->iter(lock, helper->closure);
1937 }
1938
1939 static int ldlm_res_iter_helper(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1940                                 cfs_hlist_node_t *hnode, void *arg)
1941
1942 {
1943         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1944
1945         return ldlm_resource_foreach(res, ldlm_iter_helper, arg) ==
1946                LDLM_ITER_STOP;
1947 }
1948
1949 void ldlm_namespace_foreach(struct ldlm_namespace *ns,
1950                             ldlm_iterator_t iter, void *closure)
1951
1952 {
1953         struct iter_helper_data helper = { iter: iter, closure: closure };
1954
1955         cfs_hash_for_each_nolock(ns->ns_rs_hash,
1956                                  ldlm_res_iter_helper, &helper);
1957
1958 }
1959
1960 /* non-blocking function to manipulate a lock whose cb_data is being put away.
1961  * return  0:  find no resource
1962  *       > 0:  must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE.
1963  *       < 0:  errors
1964  */
1965 int ldlm_resource_iterate(struct ldlm_namespace *ns,
1966                           const struct ldlm_res_id *res_id,
1967                           ldlm_iterator_t iter, void *data)
1968 {
1969         struct ldlm_resource *res;
1970         int rc;
1971         ENTRY;
1972
1973         if (ns == NULL) {
1974                 CERROR("must pass in namespace\n");
1975                 LBUG();
1976         }
1977
1978         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1979         if (res == NULL)
1980                 RETURN(0);
1981
1982         LDLM_RESOURCE_ADDREF(res);
1983         rc = ldlm_resource_foreach(res, iter, data);
1984         LDLM_RESOURCE_DELREF(res);
1985         ldlm_resource_putref(res);
1986         RETURN(rc);
1987 }
1988
1989 /* Lock replay */
1990
1991 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
1992 {
1993         cfs_list_t *list = closure;
1994
1995         /* we use l_pending_chain here, because it's unused on clients. */
1996         LASSERTF(cfs_list_empty(&lock->l_pending_chain),
1997                  "lock %p next %p prev %p\n",
1998                  lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
1999         /* bug 9573: don't replay locks left after eviction, or
2000          * bug 17614: locks being actively cancelled. Get a reference
2001          * on a lock so that it does not disapear under us (e.g. due to cancel)
2002          */
2003         if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_CANCELING))) {
2004                 cfs_list_add(&lock->l_pending_chain, list);
2005                 LDLM_LOCK_GET(lock);
2006         }
2007
2008         return LDLM_ITER_CONTINUE;
2009 }
2010
2011 static int replay_lock_interpret(const struct lu_env *env,
2012                                  struct ptlrpc_request *req,
2013                                  struct ldlm_async_args *aa, int rc)
2014 {
2015         struct ldlm_lock     *lock;
2016         struct ldlm_reply    *reply;
2017         struct obd_export    *exp;
2018
2019         ENTRY;
2020         cfs_atomic_dec(&req->rq_import->imp_replay_inflight);
2021         if (rc != ELDLM_OK)
2022                 GOTO(out, rc);
2023
2024
2025         reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2026         if (reply == NULL)
2027                 GOTO(out, rc = -EPROTO);
2028
2029         lock = ldlm_handle2lock(&aa->lock_handle);
2030         if (!lock) {
2031                 CERROR("received replay ack for unknown local cookie "LPX64
2032                        " remote cookie "LPX64 " from server %s id %s\n",
2033                        aa->lock_handle.cookie, reply->lock_handle.cookie,
2034                        req->rq_export->exp_client_uuid.uuid,
2035                        libcfs_id2str(req->rq_peer));
2036                 GOTO(out, rc = -ESTALE);
2037         }
2038
2039         /* Key change rehash lock in per-export hash with new key */
2040         exp = req->rq_export;
2041         if (exp && exp->exp_lock_hash) {
2042                 cfs_hash_rehash_key(exp->exp_lock_hash,
2043                                     &lock->l_remote_handle,
2044                                     &reply->lock_handle,
2045                                     &lock->l_exp_hash);
2046         } else {
2047                 lock->l_remote_handle = reply->lock_handle;
2048         }
2049
2050         LDLM_DEBUG(lock, "replayed lock:");
2051         ptlrpc_import_recovery_state_machine(req->rq_import);
2052         LDLM_LOCK_PUT(lock);
2053 out:
2054         if (rc != ELDLM_OK)
2055                 ptlrpc_connect_import(req->rq_import);
2056
2057         RETURN(rc);
2058 }
2059
2060 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2061 {
2062         struct ptlrpc_request *req;
2063         struct ldlm_async_args *aa;
2064         struct ldlm_request   *body;
2065         int flags;
2066         ENTRY;
2067
2068
2069         /* Bug 11974: Do not replay a lock which is actively being canceled */
2070         if (lock->l_flags & LDLM_FL_CANCELING) {
2071                 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2072                 RETURN(0);
2073         }
2074
2075         /* If this is reply-less callback lock, we cannot replay it, since
2076          * server might have long dropped it, but notification of that event was
2077          * lost by network. (and server granted conflicting lock already) */
2078         if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
2079                 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2080                 ldlm_lock_cancel(lock);
2081                 RETURN(0);
2082         }
2083         /*
2084          * If granted mode matches the requested mode, this lock is granted.
2085          *
2086          * If they differ, but we have a granted mode, then we were granted
2087          * one mode and now want another: ergo, converting.
2088          *
2089          * If we haven't been granted anything and are on a resource list,
2090          * then we're blocked/waiting.
2091          *
2092          * If we haven't been granted anything and we're NOT on a resource list,
2093          * then we haven't got a reply yet and don't have a known disposition.
2094          * This happens whenever a lock enqueue is the request that triggers
2095          * recovery.
2096          */
2097         if (lock->l_granted_mode == lock->l_req_mode)
2098                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2099         else if (lock->l_granted_mode)
2100                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
2101         else if (!cfs_list_empty(&lock->l_res_link))
2102                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2103         else
2104                 flags = LDLM_FL_REPLAY;
2105
2106         req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2107                                         LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2108         if (req == NULL)
2109                 RETURN(-ENOMEM);
2110
2111         /* We're part of recovery, so don't wait for it. */
2112         req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2113
2114         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2115         ldlm_lock2desc(lock, &body->lock_desc);
2116         body->lock_flags = flags;
2117
2118         ldlm_lock2handle(lock, &body->lock_handle[0]);
2119         if (lock->l_lvb_len != 0) {
2120                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2121                 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2122                                      lock->l_lvb_len);
2123         }
2124         ptlrpc_request_set_replen(req);
2125         /* notify the server we've replayed all requests.
2126          * also, we mark the request to be put on a dedicated
2127          * queue to be processed after all request replayes.
2128          * bug 6063 */
2129         lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2130
2131         LDLM_DEBUG(lock, "replaying lock:");
2132
2133         cfs_atomic_inc(&req->rq_import->imp_replay_inflight);
2134         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2135         aa = ptlrpc_req_async_args(req);
2136         aa->lock_handle = body->lock_handle[0];
2137         req->rq_interpret_reply = (ptlrpc_interpterer_t)replay_lock_interpret;
2138         ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
2139
2140         RETURN(0);
2141 }
2142
2143 /**
2144  * Cancel as many unused locks as possible before replay. since we are
2145  * in recovery, we can't wait for any outstanding RPCs to send any RPC
2146  * to the server.
2147  *
2148  * Called only in recovery before replaying locks. there is no need to
2149  * replay locks that are unused. since the clients may hold thousands of
2150  * cached unused locks, dropping the unused locks can greatly reduce the
2151  * load on the servers at recovery time.
2152  */
2153 static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns)
2154 {
2155         int canceled;
2156         CFS_LIST_HEAD(cancels);
2157
2158         CDEBUG(D_DLMTRACE, "Dropping as many unused locks as possible before"
2159                            "replay for namespace %s (%d)\n",
2160                            ldlm_ns_name(ns), ns->ns_nr_unused);
2161
2162         /* We don't need to care whether or not LRU resize is enabled
2163          * because the LDLM_CANCEL_NO_WAIT policy doesn't use the
2164          * count parameter */
2165         canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0,
2166                                          LCF_LOCAL, LDLM_CANCEL_NO_WAIT);
2167
2168         CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n",
2169                            canceled, ldlm_ns_name(ns));
2170 }
2171
2172 int ldlm_replay_locks(struct obd_import *imp)
2173 {
2174         struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2175         CFS_LIST_HEAD(list);
2176         struct ldlm_lock *lock, *next;
2177         int rc = 0;
2178
2179         ENTRY;
2180
2181         LASSERT(cfs_atomic_read(&imp->imp_replay_inflight) == 0);
2182
2183         /* don't replay locks if import failed recovery */
2184         if (imp->imp_vbr_failed)
2185                 RETURN(0);
2186
2187         /* ensure this doesn't fall to 0 before all have been queued */
2188         cfs_atomic_inc(&imp->imp_replay_inflight);
2189
2190         if (ldlm_cancel_unused_locks_before_replay)
2191                 ldlm_cancel_unused_locks_for_replay(ns);
2192
2193         ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2194
2195         cfs_list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2196                 cfs_list_del_init(&lock->l_pending_chain);
2197                 if (rc) {
2198                         LDLM_LOCK_RELEASE(lock);
2199                         continue; /* or try to do the rest? */
2200                 }
2201                 rc = replay_one_lock(imp, lock);
2202                 LDLM_LOCK_RELEASE(lock);
2203         }
2204
2205         cfs_atomic_dec(&imp->imp_replay_inflight);
2206
2207         RETURN(rc);
2208 }