lustre/ldlm/ldlm_request.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  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 = 0;
  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                 LDLM_ERROR(lock, "lock timed out (enqueued at %lu, %lus ago); "
  81                            "not entering recovery in server code, just going "
  82                            "back to sleep", lock->l_last_activity,
  83                            cfs_time_current_sec() - lock->l_last_activity);
  84                 if (cfs_time_after(cfs_time_current(), next_dump)) {
  85                         last_dump = next_dump;
  86                         next_dump = cfs_time_shift(300);
  87                         ldlm_namespace_dump(D_DLMTRACE,
  88                                             lock->l_resource->lr_namespace);
  89                         if (last_dump == 0)
  90                                 libcfs_debug_dumplog();
  91                 }
  92                 RETURN(0);
  93         }
  94
  95         obd = lock->l_conn_export->exp_obd;
  96         imp = obd->u.cli.cl_import;
  97         ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
  98         LDLM_ERROR(lock, "lock timed out (enqueued at %lu, %lus ago), entering "
  99                    "recovery for %s@%s", lock->l_last_activity,
 100                    cfs_time_current_sec() - lock->l_last_activity,
 101                    obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
 102
 103         RETURN(0);
 104 }
 105
 106 /* We use the same basis for both server side and client side functions
 107    from a single node. */
 108 int ldlm_get_enq_timeout(struct ldlm_lock *lock)
 109 {
 110         int timeout = at_get(&lock->l_resource->lr_namespace->ns_at_estimate);
 111         if (AT_OFF)
 112                 return obd_timeout / 2;
 113         /* Since these are non-updating timeouts, we should be conservative.
 114            It would be nice to have some kind of "early reply" mechanism for
 115            lock callbacks too... */
 116         timeout = min_t(int, at_max, timeout + (timeout >> 1)); /* 150% */
 117         return max(timeout, ldlm_enqueue_min);
 118 }
 119 EXPORT_SYMBOL(ldlm_get_enq_timeout);
 120
 121 int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data)
 122 {
 123         /* XXX ALLOCATE - 160 bytes */
 124         struct lock_wait_data lwd;
 125         struct obd_device *obd;
 126         struct obd_import *imp = NULL;
 127         struct l_wait_info lwi;
 128         __u32 timeout;
 129         int rc = 0;
 130         ENTRY;
 131
 132         if (flags == LDLM_FL_WAIT_NOREPROC) {
 133                 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
 134                 goto noreproc;
 135         }
 136
 137         if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
 138                        LDLM_FL_BLOCK_CONV))) {
 139                 cfs_waitq_signal(&lock->l_waitq);
 140                 RETURN(0);
 141         }
 142
 143         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
 144                    "sleeping");
 145         ldlm_lock_dump(D_OTHER, lock, 0);
 146         ldlm_reprocess_all(lock->l_resource);
 147
 148 noreproc:
 149
 150         obd = class_exp2obd(lock->l_conn_export);
 151
 152         /* if this is a local lock, then there is no import */
 153         if (obd != NULL) {
 154                 imp = obd->u.cli.cl_import;
 155         }
 156
 157         /* Wait a long time for enqueue - server may have to callback a
 158            lock from another client.  Server will evict the other client if it
 159            doesn't respond reasonably, and then give us the lock. */
 160         timeout = ldlm_get_enq_timeout(lock) * 2;
 161
 162         lwd.lwd_lock = lock;
 163
 164         if (lock->l_flags & LDLM_FL_NO_TIMEOUT) {
 165                 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
 166                 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
 167         } else {
 168                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout),
 169                                        ldlm_expired_completion_wait,
 170                                        interrupted_completion_wait, &lwd);
 171         }
 172
 173         if (imp != NULL) {
 174                 spin_lock(&imp->imp_lock);
 175                 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
 176                 spin_unlock(&imp->imp_lock);
 177         }
 178
 179         if (ns_is_client(lock->l_resource->lr_namespace) &&
 180             OBD_FAIL_CHECK(OBD_FAIL_LDLM_INTR_CP_AST | OBD_FAIL_ONCE)) {
 181                 obd_fail_loc = OBD_FAIL_LDLM_CP_BL_RACE | OBD_FAIL_ONCE;
 182                 lock->l_flags |= LDLM_FL_FAIL_LOC;
 183                 rc = -EINTR;
 184         } else {
 185                 /* Go to sleep until the lock is granted or cancelled. */
 186                 rc = l_wait_event(lock->l_waitq,
 187                                   is_granted_or_cancelled(lock), &lwi);
 188         }
 189
 190         if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED) {
 191                 LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed");
 192                 RETURN(-EIO);
 193         }
 194
 195         if (rc) {
 196                 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
 197                            rc);
 198                 RETURN(rc);
 199         }
 200
 201         LDLM_DEBUG(lock, "client-side enqueue waking up: granted after %lds",
 202                    cfs_time_current_sec() - lock->l_last_activity);
 203
 204         /* Update our time estimate */
 205         at_measured(&lock->l_resource->lr_namespace->ns_at_estimate,
 206                     cfs_time_current_sec() - lock->l_last_activity);
 207
 208         RETURN(0);
 209 }
 210
 211 /*
 212  * ->l_blocking_ast() callback for LDLM locks acquired by server-side OBDs.
 213  */
 214 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 215                       void *data, int flag)
 216 {
 217         int do_ast;
 218         ENTRY;
 219
 220         if (flag == LDLM_CB_CANCELING) {
 221                 /* Don't need to do anything here. */
 222                 RETURN(0);
 223         }
 224
 225         lock_res_and_lock(lock);
 226         /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
 227          * that ldlm_blocking_ast is called just before intent_policy method
 228          * takes the ns_lock, then by the time we get the lock, we might not
 229          * be the correct blocking function anymore.  So check, and return
 230          * early, if so. */
 231         if (lock->l_blocking_ast != ldlm_blocking_ast) {
 232                 unlock_res_and_lock(lock);
 233                 RETURN(0);
 234         }
 235
 236         lock->l_flags |= LDLM_FL_CBPENDING;
 237         do_ast = (!lock->l_readers && !lock->l_writers);
 238         unlock_res_and_lock(lock);
 239
 240         if (do_ast) {
 241                 struct lustre_handle lockh;
 242                 int rc;
 243
 244                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
 245                 ldlm_lock2handle(lock, &lockh);
 246                 rc = ldlm_cli_cancel(&lockh);
 247                 if (rc < 0)
 248                         CERROR("ldlm_cli_cancel: %d\n", rc);
 249         } else {
 250                 LDLM_DEBUG(lock, "Lock still has references, will be "
 251                            "cancelled later");
 252         }
 253         RETURN(0);
 254 }
 255
 256 /*
 257  * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See
 258  * comment in filter_intent_policy() on why you may need this.
 259  */
 260 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
 261 {
 262         /*
 263          * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for
 264          * that is rather subtle: with OST-side locking, it may so happen that
 265          * _all_ extent locks are held by the OST. If client wants to obtain
 266          * current file size it calls ll{,u}_glimpse_size(), and (as locks are
 267          * on the server), dummy glimpse callback fires and does
 268          * nothing. Client still receives correct file size due to the
 269          * following fragment in filter_intent_policy():
 270          *
 271          * rc = l->l_glimpse_ast(l, NULL); // this will update the LVB
 272          * if (rc != 0 && res->lr_namespace->ns_lvbo &&
 273          *     res->lr_namespace->ns_lvbo->lvbo_update) {
 274          *         res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1);
 275          * }
 276          *
 277          * that is, after glimpse_ast() fails, filter_lvbo_update() runs, and
 278          * returns correct file size to the client.
 279          */
 280         return -ELDLM_NO_LOCK_DATA;
 281 }
 282
 283 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
 284                            struct ldlm_res_id *res_id,
 285                            ldlm_type_t type, ldlm_policy_data_t *policy,
 286                            ldlm_mode_t mode, int *flags,
 287                            ldlm_blocking_callback blocking,
 288                            ldlm_completion_callback completion,
 289                            ldlm_glimpse_callback glimpse,
 290                            void *data, __u32 lvb_len, void *lvb_swabber,
 291                            struct lustre_handle *lockh)
 292 {
 293         struct ldlm_lock *lock;
 294         int err;
 295         ENTRY;
 296
 297         LASSERT(!(*flags & LDLM_FL_REPLAY));
 298         if (ns_is_client(ns)) {
 299                 CERROR("Trying to enqueue local lock in a shadow namespace\n");
 300                 LBUG();
 301         }
 302
 303         lock = ldlm_lock_create(ns, *res_id, type, mode, blocking,
 304                                 completion, glimpse, data, lvb_len);
 305         if (!lock)
 306                 GOTO(out_nolock, err = -ENOMEM);
 307
 308         ldlm_lock_addref_internal(lock, mode);
 309         ldlm_lock2handle(lock, lockh);
 310         lock_res_and_lock(lock);
 311         lock->l_flags |= LDLM_FL_LOCAL;
 312         if (*flags & LDLM_FL_ATOMIC_CB)
 313                 lock->l_flags |= LDLM_FL_ATOMIC_CB;
 314         lock->l_lvb_swabber = lvb_swabber;
 315         unlock_res_and_lock(lock);
 316         if (policy != NULL)
 317                 lock->l_policy_data = *policy;
 318         if (type == LDLM_EXTENT)
 319                 lock->l_req_extent = policy->l_extent;
 320
 321         err = ldlm_lock_enqueue(ns, &lock, policy, flags);
 322         if (err != ELDLM_OK)
 323                 GOTO(out, err);
 324
 325         if (policy != NULL)
 326                 *policy = lock->l_policy_data;
 327         if ((*flags) & LDLM_FL_LOCK_CHANGED)
 328                 *res_id = lock->l_resource->lr_name;
 329
 330         if (lock->l_completion_ast)
 331                 lock->l_completion_ast(lock, *flags, NULL);
 332
 333         LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
 334         EXIT;
 335  out:
 336         LDLM_LOCK_PUT(lock);
 337  out_nolock:
 338         return err;
 339 }
 340
 341 static void failed_lock_cleanup(struct ldlm_namespace *ns,
 342                                 struct ldlm_lock *lock,
 343                                 struct lustre_handle *lockh, int mode)
 344 {
 345         int need_cancel = 0;
 346
 347         /* Set a flag to prevent us from sending a CANCEL (bug 407) */
 348         lock_res_and_lock(lock);
 349         /* Check that lock is not granted or failed, we might race. */
 350         if ((lock->l_req_mode != lock->l_granted_mode) &&
 351             !(lock->l_flags & LDLM_FL_FAILED)) {
 352                 /* Make sure that this lock will not be found by raced
 353                  * bl_ast and -EINVAL reply is sent to server anyways.
 354                  * bug 17645 */
 355                 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED |
 356                                  LDLM_FL_ATOMIC_CB;
 357                 need_cancel = 1;
 358         }
 359         unlock_res_and_lock(lock);
 360
 361         if (need_cancel) {
 362                 LDLM_DEBUG(lock,
 363                            "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | "
 364                            "LDLM_FL_ATOMIC_CB");
 365                 ldlm_lock_decref_and_cancel(lockh, mode);
 366         } else {
 367                 LDLM_DEBUG(lock, "lock was granted or failed in race");
 368                 ldlm_lock_decref(lockh, mode);
 369         }
 370
 371         /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
 372          *       from llite/file.c/ll_file_flock(). */
 373         if (lock->l_resource->lr_type == LDLM_FLOCK) {
 374                 ldlm_lock_destroy(lock);
 375         }
 376 }
 377
 378 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
 379                           ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
 380                           int *flags, void *lvb, __u32 lvb_len,
 381                           void *lvb_swabber, struct lustre_handle *lockh,int rc)
 382 {
 383         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 384         int is_replay = *flags & LDLM_FL_REPLAY;
 385         struct lustre_handle old_hash_key;
 386         struct ldlm_lock *lock;
 387         struct ldlm_reply *reply;
 388         int cleanup_phase = 1;
 389         ENTRY;
 390
 391         lock = ldlm_handle2lock(lockh);
 392         /* ldlm_cli_enqueue is holding a reference on this lock. */
 393         if (!lock) {
 394                 LASSERT(type == LDLM_FLOCK);
 395                 RETURN(-ENOLCK);
 396         }
 397
 398         if (rc != ELDLM_OK) {
 399                 LASSERT(!is_replay);
 400                 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
 401                            rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
 402                 if (rc == ELDLM_LOCK_ABORTED) {
 403                         /* Before we return, swab the reply */
 404                         reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF,
 405                                                    sizeof(*reply),
 406                                                    lustre_swab_ldlm_reply);
 407                         if (reply == NULL) {
 408                                 CERROR("Can't unpack ldlm_reply\n");
 409                                 rc = -EPROTO;
 410                         }
 411                         if (lvb_len) {
 412                                 void *tmplvb;
 413                                 tmplvb = lustre_swab_repbuf(req,
 414                                                             DLM_REPLY_REC_OFF,
 415                                                             lvb_len,
 416                                                             lvb_swabber);
 417                                 if (tmplvb == NULL)
 418                                         GOTO(cleanup, rc = -EPROTO);
 419                                 if (lvb != NULL)
 420                                         memcpy(lvb, tmplvb, lvb_len);
 421                         }
 422                 }
 423                 GOTO(cleanup, rc);
 424         }
 425
 426         reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF, sizeof(*reply),
 427                                    lustre_swab_ldlm_reply);
 428         if (reply == NULL) {
 429                 CERROR("Can't unpack ldlm_reply\n");
 430                 GOTO(cleanup, rc = -EPROTO);
 431         }
 432
 433         /* lock enqueued on the server */
 434         cleanup_phase = 0;
 435
 436         lock_res_and_lock(lock);
 437         old_hash_key = lock->l_remote_handle;
 438         lock->l_remote_handle = reply->lock_handle;
 439
 440         /* Key change rehash lock in per-export hash with new key */
 441         if (exp->exp_lock_hash)
 442                 lustre_hash_rehash_key(exp->exp_lock_hash, &old_hash_key,
 443                                        &lock->l_remote_handle,
 444                                        &lock->l_exp_hash);
 445
 446         *flags = reply->lock_flags;
 447         lock->l_flags |= reply->lock_flags & LDLM_INHERIT_FLAGS;
 448         /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match()
 449          * to wait with no timeout as well */
 450         lock->l_flags |= reply->lock_flags & LDLM_FL_NO_TIMEOUT;
 451         unlock_res_and_lock(lock);
 452
 453         CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%x\n",
 454                lock, reply->lock_handle.cookie, *flags);
 455
 456         /* If enqueue returned a blocked lock but the completion handler has
 457          * already run, then it fixed up the resource and we don't need to do it
 458          * again. */
 459         if ((*flags) & LDLM_FL_LOCK_CHANGED) {
 460                 int newmode = reply->lock_desc.l_req_mode;
 461                 LASSERT(!is_replay);
 462                 if (newmode && newmode != lock->l_req_mode) {
 463                         LDLM_DEBUG(lock, "server returned different mode %s",
 464                                    ldlm_lockname[newmode]);
 465                         lock->l_req_mode = newmode;
 466                 }
 467
 468                 if (memcmp(reply->lock_desc.l_resource.lr_name.name,
 469                           lock->l_resource->lr_name.name,
 470                           sizeof(struct ldlm_res_id))) {
 471                         CDEBUG(D_INFO, "remote intent success, locking "
 472                                         "("LPU64"/"LPU64"/"LPU64") instead of "
 473                                         "("LPU64"/"LPU64"/"LPU64")\n",
 474                                reply->lock_desc.l_resource.lr_name.name[0],
 475                                reply->lock_desc.l_resource.lr_name.name[1],
 476                                reply->lock_desc.l_resource.lr_name.name[2],
 477                                lock->l_resource->lr_name.name[0],
 478                                lock->l_resource->lr_name.name[1],
 479                                lock->l_resource->lr_name.name[2]);
 480
 481                         rc = ldlm_lock_change_resource(ns, lock,
 482                                            reply->lock_desc.l_resource.lr_name);
 483                         if (rc || lock->l_resource == NULL)
 484                                 GOTO(cleanup, rc = -ENOMEM);
 485                         LDLM_DEBUG(lock, "client-side enqueue, new resource");
 486                 }
 487                 if (with_policy)
 488                         if (!(type == LDLM_IBITS && !(exp->exp_connect_flags &
 489                                                     OBD_CONNECT_IBITS)))
 490                                 lock->l_policy_data =
 491                                                  reply->lock_desc.l_policy_data;
 492                 if (type != LDLM_PLAIN)
 493                         LDLM_DEBUG(lock,"client-side enqueue, new policy data");
 494         }
 495
 496         if ((*flags) & LDLM_FL_AST_SENT ||
 497             /* Cancel extent locks as soon as possible on a liblustre client,
 498              * because it cannot handle asynchronous ASTs robustly (see
 499              * bug 7311). */
 500             (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) {
 501                 lock_res_and_lock(lock);
 502                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
 503                 unlock_res_and_lock(lock);
 504                 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
 505         }
 506
 507         /* If the lock has already been granted by a completion AST, don't
 508          * clobber the LVB with an older one. */
 509         if (lvb_len && (lock->l_req_mode != lock->l_granted_mode)) {
 510                 void *tmplvb;
 511                 tmplvb = lustre_swab_repbuf(req, DLM_REPLY_REC_OFF, lvb_len,
 512                                             lvb_swabber);
 513                 if (tmplvb == NULL)
 514                         GOTO(cleanup, rc = -EPROTO);
 515                 memcpy(lock->l_lvb_data, tmplvb, lvb_len);
 516         }
 517
 518         if (!is_replay) {
 519                 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
 520                 if (lock->l_completion_ast != NULL) {
 521                         int err = lock->l_completion_ast(lock, *flags, NULL);
 522                         if (!rc)
 523                                 rc = err;
 524                         if (rc && type != LDLM_FLOCK) /* bug 9425, bug 10250 */
 525                                 cleanup_phase = 1;
 526                 }
 527         }
 528
 529         if (lvb_len && lvb != NULL) {
 530                 /* Copy the LVB here, and not earlier, because the completion
 531                  * AST (if any) can override what we got in the reply */
 532                 memcpy(lvb, lock->l_lvb_data, lvb_len);
 533         }
 534
 535         EXIT;
 536 cleanup:
 537         if (cleanup_phase == 1 && rc)
 538                 failed_lock_cleanup(ns, lock, lockh, mode);
 539         /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
 540         LDLM_LOCK_PUT(lock);
 541         LDLM_LOCK_PUT(lock);
 542         return rc;
 543 }
 544
 545 /* PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
 546  * a single page on the send/receive side. XXX: 512 should be changed
 547  * to more adequate value. */
 548 static inline int ldlm_req_handles_avail(struct obd_export *exp,
 549                                          __u32 *size, int bufcount, int off)
 550 {
 551         int avail = min_t(int, LDLM_MAXREQSIZE, CFS_PAGE_SIZE - 512);
 552
 553         avail -= lustre_msg_size(class_exp2cliimp(exp)->imp_msg_magic,
 554                                  bufcount, size);
 555         if (likely(avail >= 0))
 556                 avail /= (int)sizeof(struct lustre_handle);
 557         else
 558                 avail = 0;
 559         avail += LDLM_LOCKREQ_HANDLES - off;
 560
 561         return avail;
 562 }
 563
 564 static inline int ldlm_cancel_handles_avail(struct obd_export *exp)
 565 {
 566         __u32 size[2] = { sizeof(struct ptlrpc_body),
 567                         sizeof(struct ldlm_request) };
 568         return ldlm_req_handles_avail(exp, size, 2, 0);
 569 }
 570
 571 /* Cancel lru locks and pack them into the enqueue request. Pack there the given
 572  * @count locks in @cancels. */
 573 struct ptlrpc_request *ldlm_prep_elc_req(struct obd_export *exp, int version,
 574                                          int opc, int bufcount, __u32 *size,
 575                                          int bufoff, int canceloff,
 576                                          struct list_head *cancels, int count)
 577 {
 578         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 579         int flags, avail, to_free, pack = 0;
 580         struct ldlm_request *dlm = NULL;
 581         struct ptlrpc_request *req;
 582         CFS_LIST_HEAD(head);
 583         ENTRY;
 584
 585         if (cancels == NULL)
 586                 cancels = &head;
 587         if (exp_connect_cancelset(exp)) {
 588                 /* Estimate the amount of free space in the request. */
 589                 LASSERT(bufoff < bufcount);
 590
 591                 avail = ldlm_req_handles_avail(exp, size, bufcount, canceloff);
 592                 flags = ns_connect_lru_resize(ns) ?
 593                         LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
 594                 to_free = !ns_connect_lru_resize(ns) &&
 595                           opc == LDLM_ENQUEUE ? 1 : 0;
 596
 597                 /* Cancel lru locks here _only_ if the server supports
 598                  * EARLY_CANCEL. Otherwise we have to send extra CANCEL
 599                  * rpc, what will make us slower. */
 600                 if (avail > count)
 601                         count += ldlm_cancel_lru_local(ns, cancels, to_free,
 602                                                        avail - count, 0, flags);
 603                 if (avail > count)
 604                         pack = count;
 605                 else
 606                         pack = avail;
 607                 size[bufoff] = ldlm_request_bufsize(pack, opc);
 608         }
 609
 610         req = ptlrpc_prep_req(class_exp2cliimp(exp), version,
 611                               opc, bufcount, size, NULL);
 612         req->rq_export = class_export_get(exp);
 613         if (exp_connect_cancelset(exp) && req) {
 614                 if (canceloff) {
 615                         dlm = lustre_msg_buf(req->rq_reqmsg, bufoff,
 616                                              sizeof(*dlm));
 617                         /* Skip first lock handler in ldlm_request_pack(),
 618                          * this method will incrment @lock_count according
 619                          * to the lock handle amount actually written to
 620                          * the buffer. */
 621                         dlm->lock_count = canceloff;
 622                 }
 623                 /* Pack into the request @pack lock handles. */
 624                 ldlm_cli_cancel_list(cancels, pack, req, bufoff);
 625                 /* Prepare and send separate cancel rpc for others. */
 626                 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
 627         } else {
 628                 ldlm_lock_list_put(cancels, l_bl_ast, count);
 629         }
 630         RETURN(req);
 631 }
 632
 633 struct ptlrpc_request *ldlm_prep_enqueue_req(struct obd_export *exp,
 634                                              int bufcount, __u32 *size,
 635                                              struct list_head *cancels,
 636                                              int count)
 637 {
 638         return ldlm_prep_elc_req(exp, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
 639                                  bufcount, size, DLM_LOCKREQ_OFF,
 640                                  LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
 641 }
 642
 643 /* If a request has some specific initialisation it is passed in @reqp,
 644  * otherwise it is created in ldlm_cli_enqueue.
 645  *
 646  * Supports sync and async requests, pass @async flag accordingly. If a
 647  * request was created in ldlm_cli_enqueue and it is the async request,
 648  * pass it to the caller in @reqp. */
 649 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
 650                      struct ldlm_enqueue_info *einfo, struct ldlm_res_id res_id,
 651                      ldlm_policy_data_t *policy, int *flags,
 652                      void *lvb, __u32 lvb_len, void *lvb_swabber,
 653                      struct lustre_handle *lockh, int async)
 654 {
 655         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 656         struct ldlm_lock *lock;
 657         struct ldlm_request *body;
 658         struct ldlm_reply *reply;
 659         __u32 size[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
 660                         [DLM_LOCKREQ_OFF]     = sizeof(*body),
 661                         [DLM_REPLY_REC_OFF]   = lvb_len ? lvb_len :
 662                                                 sizeof(struct ost_lvb) };
 663         int is_replay = *flags & LDLM_FL_REPLAY;
 664         int req_passed_in = 1, rc, err;
 665         struct ptlrpc_request *req;
 666         ENTRY;
 667
 668         LASSERT(exp != NULL);
 669
 670         /* If we're replaying this lock, just check some invariants.
 671          * If we're creating a new lock, get everything all setup nice. */
 672         if (is_replay) {
 673                 lock = ldlm_handle2lock(lockh);
 674                 LASSERT(lock != NULL);
 675                 LASSERT(exp == lock->l_conn_export);
 676         } else {
 677                 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
 678                                         einfo->ei_mode, einfo->ei_cb_bl,
 679                                         einfo->ei_cb_cp, einfo->ei_cb_gl,
 680                                         einfo->ei_cbdata, lvb_len);
 681                 if (lock == NULL)
 682                         RETURN(-ENOMEM);
 683                 /* for the local lock, add the reference */
 684                 ldlm_lock_addref_internal(lock, einfo->ei_mode);
 685                 ldlm_lock2handle(lock, lockh);
 686                 lock->l_lvb_swabber = lvb_swabber;
 687                 if (policy != NULL) {
 688                         /* INODEBITS_INTEROP: If the server does not support
 689                          * inodebits, we will request a plain lock in the
 690                          * descriptor (ldlm_lock2desc() below) but use an
 691                          * inodebits lock internally with both bits set.
 692                          */
 693                         if (einfo->ei_type == LDLM_IBITS &&
 694                             !(exp->exp_connect_flags & OBD_CONNECT_IBITS))
 695                                 lock->l_policy_data.l_inodebits.bits =
 696                                         MDS_INODELOCK_LOOKUP |
 697                                         MDS_INODELOCK_UPDATE;
 698                         else
 699                                 lock->l_policy_data = *policy;
 700                 }
 701
 702                 if (einfo->ei_type == LDLM_EXTENT)
 703                         lock->l_req_extent = policy->l_extent;
 704                 LDLM_DEBUG(lock, "client-side enqueue START");
 705         }
 706
 707         /* lock not sent to server yet */
 708
 709         if (reqp == NULL || *reqp == NULL) {
 710                 req = ldlm_prep_enqueue_req(exp, 2, size, NULL, 0);
 711                 if (req == NULL) {
 712                         failed_lock_cleanup(ns, lock, lockh, einfo->ei_mode);
 713                         LDLM_LOCK_PUT(lock);
 714                         RETURN(-ENOMEM);
 715                 }
 716                 req_passed_in = 0;
 717                 if (reqp)
 718                         *reqp = req;
 719         } else {
 720                 req = *reqp;
 721                 LASSERTF(lustre_msg_buflen(req->rq_reqmsg, DLM_LOCKREQ_OFF) >=
 722                          sizeof(*body), "buflen[%d] = %d, not %d\n",
 723                          DLM_LOCKREQ_OFF,
 724                          lustre_msg_buflen(req->rq_reqmsg, DLM_LOCKREQ_OFF),
 725                          (int)sizeof(*body));
 726         }
 727
 728         lock->l_conn_export = exp;
 729         lock->l_export = NULL;
 730         lock->l_blocking_ast = einfo->ei_cb_bl;
 731
 732         /* Dump lock data into the request buffer */
 733         body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, sizeof(*body));
 734         ldlm_lock2desc(lock, &body->lock_desc);
 735         body->lock_flags = *flags;
 736         body->lock_handle[0] = *lockh;
 737
 738         /* Continue as normal. */
 739         if (!req_passed_in) {
 740                 size[DLM_LOCKREPLY_OFF] = sizeof(*reply);
 741                 ptlrpc_req_set_repsize(req, 3, size);
 742         }
 743
 744         /*
 745          * Liblustre client doesn't get extent locks, except for O_APPEND case
 746          * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
 747          * [i_size, OBD_OBJECT_EOF] lock is taken.
 748          */
 749         LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT ||
 750                      policy->l_extent.end == OBD_OBJECT_EOF));
 751
 752         if (async) {
 753                 LASSERT(reqp != NULL);
 754                 RETURN(0);
 755         }
 756
 757         rc = ptlrpc_queue_wait(req);
 758         err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
 759                                     einfo->ei_mode, flags, lvb, lvb_len,
 760                                     lvb_swabber, lockh, rc);
 761
 762         /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
 763          * one reference that we took */
 764         if (err == -ENOLCK)
 765                 LDLM_LOCK_PUT(lock);
 766         else
 767                 rc = err;
 768
 769         if (!req_passed_in && req != NULL) {
 770                 ptlrpc_req_finished(req);
 771                 if (reqp)
 772                         *reqp = NULL;
 773         }
 774
 775         RETURN(rc);
 776 }
 777
 778 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
 779                                   __u32 *flags)
 780 {
 781         struct ldlm_resource *res;
 782         int rc;
 783         ENTRY;
 784         if (ns_is_client(lock->l_resource->lr_namespace)) {
 785                 CERROR("Trying to cancel local lock\n");
 786                 LBUG();
 787         }
 788         LDLM_DEBUG(lock, "client-side local convert");
 789
 790         res = ldlm_lock_convert(lock, new_mode, flags);
 791         if (res) {
 792                 ldlm_reprocess_all(res);
 793                 rc = 0;
 794         } else {
 795                 rc = EDEADLOCK;
 796         }
 797         LDLM_DEBUG(lock, "client-side local convert handler END");
 798         LDLM_LOCK_PUT(lock);
 799         RETURN(rc);
 800 }
 801
 802 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
 803  * conversion of locks which are on the waiting or converting queue */
 804 /* Caller of this code is supposed to take care of lock readers/writers
 805    accounting */
 806 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags)
 807 {
 808         struct ldlm_request *body;
 809         struct ldlm_reply *reply;
 810         struct ldlm_lock *lock;
 811         struct ldlm_resource *res;
 812         struct ptlrpc_request *req;
 813         __u32 size[2] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
 814                         [DLM_LOCKREQ_OFF]     = sizeof(*body) };
 815         int rc;
 816         ENTRY;
 817
 818         lock = ldlm_handle2lock(lockh);
 819         if (!lock) {
 820                 LBUG();
 821                 RETURN(-EINVAL);
 822         }
 823         *flags = 0;
 824
 825         if (lock->l_conn_export == NULL)
 826                 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
 827
 828         LDLM_DEBUG(lock, "client-side convert");
 829
 830         req = ptlrpc_prep_req(class_exp2cliimp(lock->l_conn_export),
 831                               LUSTRE_DLM_VERSION, LDLM_CONVERT, 2, size, NULL);
 832         if (!req)
 833                 GOTO(out, rc = -ENOMEM);
 834
 835         body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, sizeof(*body));
 836         body->lock_handle[0] = lock->l_remote_handle;
 837
 838         body->lock_desc.l_req_mode = new_mode;
 839         body->lock_flags = *flags;
 840
 841         size[DLM_LOCKREPLY_OFF] = sizeof(*reply);
 842         ptlrpc_req_set_repsize(req, 2, size);
 843
 844         rc = ptlrpc_queue_wait(req);
 845         if (rc != ELDLM_OK)
 846                 GOTO(out, rc);
 847
 848         reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF, sizeof(*reply),
 849                                    lustre_swab_ldlm_reply);
 850         if (reply == NULL) {
 851                 CERROR ("Can't unpack ldlm_reply\n");
 852                 GOTO (out, rc = -EPROTO);
 853         }
 854
 855         if (req->rq_status)
 856                 GOTO(out, rc = req->rq_status);
 857
 858         res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
 859         if (res != NULL) {
 860                 ldlm_reprocess_all(res);
 861                 /* Go to sleep until the lock is granted. */
 862                 /* FIXME: or cancelled. */
 863                 if (lock->l_completion_ast) {
 864                         rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
 865                                                     NULL);
 866                         if (rc)
 867                                 GOTO(out, rc);
 868                 }
 869         } else {
 870                 rc = EDEADLOCK;
 871         }
 872         EXIT;
 873  out:
 874         LDLM_LOCK_PUT(lock);
 875         ptlrpc_req_finished(req);
 876         return rc;
 877 }
 878
 879 /* Cancel locks locally.
 880  * Returns:
 881  * LDLM_FL_LOCAL_ONLY if tere is no need in a CANCEL rpc to the server;
 882  * LDLM_FL_CANCELING otherwise;
 883  * LDLM_FL_BL_AST if there is a need in a separate CANCEL rpc. */
 884 static int ldlm_cli_cancel_local(struct ldlm_lock *lock)
 885 {
 886         int rc = LDLM_FL_LOCAL_ONLY;
 887         ENTRY;
 888
 889         if (lock->l_conn_export) {
 890                 int local_only;
 891
 892                 LDLM_DEBUG(lock, "client-side cancel");
 893                 /* Set this flag to prevent others from getting new references*/
 894                 lock_res_and_lock(lock);
 895                 lock->l_flags |= LDLM_FL_CBPENDING;
 896                 local_only = (lock->l_flags &
 897                               (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
 898                 ldlm_cancel_callback(lock);
 899                 rc = (lock->l_flags & LDLM_FL_BL_AST) ?
 900                         LDLM_FL_BL_AST : LDLM_FL_CANCELING;
 901                 unlock_res_and_lock(lock);
 902
 903                 if (local_only) {
 904                         CDEBUG(D_DLMTRACE, "not sending request (at caller's "
 905                                "instruction)\n");
 906                         rc = LDLM_FL_LOCAL_ONLY;
 907                 }
 908                 ldlm_lock_cancel(lock);
 909         } else {
 910                 if (ns_is_client(lock->l_resource->lr_namespace)) {
 911                         LDLM_ERROR(lock, "Trying to cancel local lock");
 912                         LBUG();
 913                 }
 914                 LDLM_DEBUG(lock, "server-side local cancel");
 915                 ldlm_lock_cancel(lock);
 916                 ldlm_reprocess_all(lock->l_resource);
 917         }
 918
 919         RETURN(rc);
 920 }
 921
 922 /* Pack @count locks in @head into ldlm_request buffer at the offset @off,
 923    of the request @req. */
 924 static void ldlm_cancel_pack(struct ptlrpc_request *req, int off,
 925                              struct list_head *head, int count)
 926 {
 927         struct ldlm_request *dlm;
 928         struct ldlm_lock *lock;
 929         int max, packed = 0;
 930         ENTRY;
 931
 932         dlm = lustre_msg_buf(req->rq_reqmsg, off, sizeof(*dlm));
 933         LASSERT(dlm != NULL);
 934
 935         /* Check the room in the request buffer. */
 936         max = lustre_msg_buflen(req->rq_reqmsg, off) -
 937                 sizeof(struct ldlm_request);
 938         max /= sizeof(struct lustre_handle);
 939         max += LDLM_LOCKREQ_HANDLES;
 940         LASSERT(max >= dlm->lock_count + count);
 941
 942         /* XXX: it would be better to pack lock handles grouped by resource.
 943          * so that the server cancel would call filter_lvbo_update() less
 944          * frequently. */
 945         list_for_each_entry(lock, head, l_bl_ast) {
 946                 if (!count--)
 947                         break;
 948                 LASSERT(lock->l_conn_export);
 949                 /* Pack the lock handle to the given request buffer. */
 950                 LDLM_DEBUG(lock, "packing");
 951                 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
 952                 packed++;
 953         }
 954         CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
 955         EXIT;
 956 }
 957
 958 /* Prepare and send a batched cancel rpc, it will include count lock handles
 959  * of locks given in @head. */
 960 int ldlm_cli_cancel_req(struct obd_export *exp,
 961                         struct list_head *cancels, int count)
 962 {
 963         struct ptlrpc_request *req = NULL;
 964         __u32 size[2] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
 965                         [DLM_LOCKREQ_OFF]     = sizeof(struct ldlm_request) };
 966         struct obd_import *imp;
 967         int free, sent = 0;
 968         int rc = 0;
 969         ENTRY;
 970
 971         LASSERT(exp != NULL);
 972         LASSERT(count > 0);
 973
 974         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, obd_fail_val);
 975
 976         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
 977                 RETURN(count);
 978
 979         free = ldlm_req_handles_avail(exp, size, 2, 0);
 980         if (count > free)
 981                 count = free;
 982
 983         size[DLM_LOCKREQ_OFF] = ldlm_request_bufsize(count, LDLM_CANCEL);
 984         while (1) {
 985                 imp = class_exp2cliimp(exp);
 986                 if (imp == NULL || imp->imp_invalid) {
 987                         CDEBUG(D_DLMTRACE,
 988                                "skipping cancel on invalid import %p\n", imp);
 989                         RETURN(count);
 990                 }
 991
 992                 req = ptlrpc_prep_req(imp, LUSTRE_DLM_VERSION, LDLM_CANCEL, 2,
 993                                       size, NULL);
 994                 if (!req)
 995                         GOTO(out, rc = -ENOMEM);
 996
 997                 req->rq_no_resend = 1;
 998                 req->rq_no_delay = 1;
 999
1000                 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1001                 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1002                 ptlrpc_at_set_req_timeout(req);
1003
1004                 /*
1005                 body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF,
1006                                       sizeof(*body));
1007                 */
1008                 ldlm_cancel_pack(req, DLM_LOCKREQ_OFF, cancels, count);
1009
1010                 ptlrpc_req_set_repsize(req, 1, NULL);
1011                 rc = ptlrpc_queue_wait(req);
1012
1013                 if (rc == ESTALE) {
1014                         CDEBUG(D_DLMTRACE, "client/server (nid %s) "
1015                                "out of sync -- not fatal\n",
1016                                libcfs_nid2str(req->rq_import->
1017                                               imp_connection->c_peer.nid));
1018                         rc = 0;
1019                 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1020                            req->rq_import_generation == imp->imp_generation) {
1021                         ptlrpc_req_finished(req);
1022                         continue;
1023                 } else if (rc != ELDLM_OK) {
1024                         CERROR("Got rc %d from cancel RPC: canceling "
1025                                "anyway\n", rc);
1026                         break;
1027                 }
1028                 sent = count;
1029                 break;
1030         }
1031
1032         ptlrpc_req_finished(req);
1033         EXIT;
1034 out:
1035         return sent ? sent : rc;
1036 }
1037
1038 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1039 {
1040         LASSERT(imp != NULL);
1041         return &imp->imp_obd->obd_namespace->ns_pool;
1042 }
1043
1044 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1045 {
1046         struct obd_device *obd;
1047         __u64 old_slv, new_slv;
1048         __u32 new_limit;
1049         ENTRY;
1050
1051         if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1052                      !imp_connect_lru_resize(req->rq_import)))
1053         {
1054                 /*
1055                  * Do nothing for corner cases.
1056                  */
1057                 RETURN(0);
1058         }
1059
1060         /*
1061          * In some cases RPC may contain slv and limit zeroed out. This is
1062          * the case when server does not support lru resize feature. This is
1063          * also possible in some recovery cases when server side reqs have no
1064          * ref to obd export and thus access to server side namespace is no
1065          * possible.
1066          */
1067         if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1068             lustre_msg_get_limit(req->rq_repmsg) == 0) {
1069                 DEBUG_REQ(D_HA, req, "Zero SLV or Limit found "
1070                           "(SLV: "LPU64", Limit: %u)",
1071                           lustre_msg_get_slv(req->rq_repmsg),
1072                           lustre_msg_get_limit(req->rq_repmsg));
1073                 RETURN(0);
1074         }
1075
1076         new_limit = lustre_msg_get_limit(req->rq_repmsg);
1077         new_slv = lustre_msg_get_slv(req->rq_repmsg);
1078         obd = req->rq_import->imp_obd;
1079
1080         /*
1081          * Set new SLV and Limit to obd fields to make accessible for pool
1082          * thread. We do not access obd_namespace and pool directly here
1083          * as there is no reliable way to make sure that they are still
1084          * alive in cleanup time. Evil races are possible which may cause
1085          * oops in that time.
1086          */
1087         write_lock(&obd->obd_pool_lock);
1088         old_slv = obd->obd_pool_slv;
1089         obd->obd_pool_slv = new_slv;
1090         obd->obd_pool_limit = new_limit;
1091         write_unlock(&obd->obd_pool_lock);
1092
1093         RETURN(0);
1094 }
1095 EXPORT_SYMBOL(ldlm_cli_update_pool);
1096
1097 int ldlm_cli_cancel(struct lustre_handle *lockh)
1098 {
1099         struct ldlm_namespace *ns;
1100         int avail, flags, count = 1, rc = 0;
1101         struct ldlm_lock *lock;
1102         CFS_LIST_HEAD(cancels);
1103         ENTRY;
1104
1105         /* concurrent cancels on the same handle can happen */
1106         lock = __ldlm_handle2lock(lockh, LDLM_FL_CANCELING);
1107         if (lock == NULL) {
1108                 LDLM_DEBUG_NOLOCK("lock is already being destroyed\n");
1109                 RETURN(0);
1110         }
1111
1112         rc = ldlm_cli_cancel_local(lock);
1113         if (rc < 0 || rc == LDLM_FL_LOCAL_ONLY) {
1114                 LDLM_LOCK_PUT(lock);
1115                 RETURN(rc < 0 ? rc : 0);
1116         }
1117         /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1118          * rpc which goes to canceld portal, so we can cancel other lru locks
1119          * here and send them all as one LDLM_CANCEL rpc. */
1120         LASSERT(list_empty(&lock->l_bl_ast));
1121         list_add(&lock->l_bl_ast, &cancels);
1122
1123         if (exp_connect_cancelset(lock->l_conn_export)) {
1124                 avail = ldlm_cancel_handles_avail(lock->l_conn_export);
1125                 LASSERT(avail > 0);
1126
1127                 ns = lock->l_resource->lr_namespace;
1128                 flags = ns_connect_lru_resize(ns) ?
1129                         LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
1130                 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1131                                                LDLM_FL_BL_AST, flags);
1132         }
1133         ldlm_cli_cancel_list(&cancels, count, NULL, 0);
1134         RETURN(0);
1135 }
1136
1137 /* XXX until we will have compound requests and can cut cancels from generic rpc
1138  * we need send cancels with LDLM_FL_BL_AST flag as separate rpc */
1139 static int ldlm_cancel_list(struct list_head *cancels, int count, int flags)
1140 {
1141         CFS_LIST_HEAD(head);
1142         struct ldlm_lock *lock, *next;
1143         int left = 0, bl_ast = 0, rc;
1144
1145         left = count;
1146         list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1147                 if (left-- == 0)
1148                         break;
1149
1150                 if (flags & LDLM_FL_LOCAL_ONLY) {
1151                         rc = LDLM_FL_LOCAL_ONLY;
1152                         ldlm_lock_cancel(lock);
1153                 } else {
1154                         rc = ldlm_cli_cancel_local(lock);
1155                 }
1156                 if (!(flags & LDLM_FL_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1157                         LDLM_DEBUG(lock, "Cancel lock separately");
1158                         list_del_init(&lock->l_bl_ast);
1159                         list_add(&lock->l_bl_ast, &head);
1160                         bl_ast ++;
1161                         continue;
1162                 }
1163                 if (rc == LDLM_FL_LOCAL_ONLY) {
1164                         /* CANCEL RPC should not be sent to server. */
1165                         list_del_init(&lock->l_bl_ast);
1166                         LDLM_LOCK_PUT(lock);
1167                         count--;
1168                 }
1169
1170         }
1171         if (bl_ast > 0) {
1172                 count -= bl_ast;
1173                 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1174         }
1175
1176         RETURN(count);
1177 }
1178
1179 /* cancel as many locks as possible w/o sending any rpcs (e.g. to write back
1180  * dirty data, to close a file, ...) or waiting for any rpcs in-flight (e.g.
1181  * readahead requests, ...) */
1182 static ldlm_policy_res_t ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns,
1183                                                     struct ldlm_lock *lock,
1184                                                     int unused, int added,
1185                                                     int count)
1186 {
1187         ldlm_policy_res_t result = LDLM_POLICY_CANCEL_LOCK;
1188         ldlm_cancel_for_recovery cb = ns->ns_cancel_for_recovery;
1189         check_res_locked(lock->l_resource);
1190
1191         /* don't check added & count since we want to process all locks
1192          * from unused list */
1193         switch (lock->l_resource->lr_type) {
1194                 case LDLM_EXTENT:
1195                 case LDLM_IBITS:
1196                         if (cb && cb(lock))
1197                                 break;
1198                 default:
1199                         result = LDLM_POLICY_SKIP_LOCK;
1200                         lock->l_flags |= LDLM_FL_SKIPPED;
1201                         break;
1202         }
1203
1204         RETURN(result);
1205 }
1206
1207 /* Return 1 to stop lru processing and keep current lock cached. Return zero
1208  * otherwise. */
1209 static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1210                                                  struct ldlm_lock *lock,
1211                                                  int unused, int added,
1212                                                  int count)
1213 {
1214         cfs_time_t cur = cfs_time_current();
1215         struct ldlm_pool *pl = &ns->ns_pool;
1216         __u64 slv, lvf, lv;
1217         cfs_time_t la;
1218
1219         /* Stop lru processing when we reached passed @count or checked all
1220          * locks in lru. */
1221         if (count && added >= count)
1222                 return LDLM_POLICY_KEEP_LOCK;
1223
1224         slv = ldlm_pool_get_slv(pl);
1225         lvf = ldlm_pool_get_lvf(pl);
1226
1227         la = cfs_duration_sec(cfs_time_sub(cur,
1228                               lock->l_last_used));
1229
1230         /* Stop when slv is not yet come from server or
1231          * lv is smaller than it is. */
1232         lv = lvf * la * unused;
1233
1234         /* Inform pool about current CLV to see it via proc. */
1235         ldlm_pool_set_clv(pl, lv);
1236         return (slv == 1 || lv < slv) ?
1237                 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1238 }
1239
1240 /* Return 1 to stop lru processing and keep current lock cached. Return zero
1241  * otherwise. */
1242 static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1243                                                    struct ldlm_lock *lock,
1244                                                    int unused, int added,
1245                                                    int count)
1246 {
1247         /* Stop lru processing when we reached passed @count or checked all
1248          * locks in lru. */
1249         return (added >= count) ?
1250                 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1251 }
1252
1253 /* Return 1 to stop lru processing and keep current lock cached. Return zero
1254  * otherwise. */
1255 static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1256                                                  struct ldlm_lock *lock,
1257                                                  int unused, int added,
1258                                                  int count)
1259 {
1260         /* Stop lru processing if young lock is found and we reached passed
1261          * @count. */
1262         return ((added >= count) &&
1263                 cfs_time_before(cfs_time_current(),
1264                                 cfs_time_add(lock->l_last_used,
1265                                              ns->ns_max_age))) ?
1266                 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1267 }
1268
1269 /* Return 1 to stop lru processing and keep current lock cached. Return zero
1270  * otherwise. */
1271 static ldlm_policy_res_t ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1272                                                     struct ldlm_lock *lock,
1273                                                     int unused, int added,
1274                                                     int count)
1275 {
1276         /* Stop lru processing when we reached passed @count or checked all
1277          * locks in lru. */
1278         return (added >= count) ?
1279                 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1280 }
1281
1282 typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *,
1283                                                       struct ldlm_lock *, int,
1284                                                       int, int);
1285
1286 static ldlm_cancel_lru_policy_t
1287 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
1288 {
1289         if (flags & LDLM_CANCEL_NO_WAIT)
1290                 return ldlm_cancel_no_wait_policy;
1291
1292         if (ns_connect_lru_resize(ns)) {
1293                 if (flags & LDLM_CANCEL_SHRINK)
1294                         /* We kill passed number of old locks. */
1295                         return ldlm_cancel_passed_policy;
1296                 else if (flags & LDLM_CANCEL_LRUR)
1297                         return ldlm_cancel_lrur_policy;
1298                 else if (flags & LDLM_CANCEL_PASSED)
1299                         return ldlm_cancel_passed_policy;
1300         } else {
1301                 if (flags & LDLM_CANCEL_AGED)
1302                         return ldlm_cancel_aged_policy;
1303         }
1304
1305         return ldlm_cancel_default_policy;
1306 }
1307
1308 /* - Free space in lru for @count new locks,
1309  *   redundant unused locks are canceled locally;
1310  * - also cancel locally unused aged locks;
1311  * - do not cancel more than @max locks;
1312  * - GET the found locks and add them into the @cancels list.
1313  *
1314  * A client lock can be added to the l_bl_ast list only when it is
1315  * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing CANCEL.
1316  * There are the following use cases: ldlm_cancel_resource_local(),
1317  * ldlm_cancel_lru_local() and ldlm_cli_cancel(), which check&set this
1318  * flag properly. As any attempt to cancel a lock rely on this flag,
1319  * l_bl_ast list is accessed later without any special locking.
1320  *
1321  * Calling policies for enabled lru resize:
1322  * ----------------------------------------
1323  * flags & LDLM_CANCEL_LRUR - use lru resize policy (SLV from server) to
1324  *                            cancel not more than @count locks;
1325  *
1326  * flags & LDLM_CANCEL_PASSED - cancel @count number of old locks (located at
1327  *                              the beginning of lru list);
1328  *
1329  * flags & LDLM_CANCEL_SHRINK - cancel not more than @count locks according to
1330  *                              memory pressre policy function;
1331  *
1332  * flags & LDLM_CANCEL_AGED -   cancel locks according to "aged policy".
1333  *
1334  * flags & LDLM_CANCEL_NO_WAIT - cancel as many unused locks as possible
1335  *                               (typically before replaying locks) w/o
1336  *                               sending any rpcs or waiting for any
1337  *                               outstanding rpc to complete
1338  */
1339 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
1340                           int count, int max, int cancel_flags, int flags)
1341 {
1342         ldlm_cancel_lru_policy_t pf;
1343         struct ldlm_lock *lock, *next;
1344         int added = 0, unused, remained;
1345         ENTRY;
1346
1347         spin_lock(&ns->ns_unused_lock);
1348         unused = ns->ns_nr_unused;
1349         remained = unused;
1350
1351         if (!ns_connect_lru_resize(ns))
1352                 count += unused - ns->ns_max_unused;
1353
1354         pf = ldlm_cancel_lru_policy(ns, flags);
1355         LASSERT(pf != NULL);
1356
1357         while (!list_empty(&ns->ns_unused_list)) {
1358                 ldlm_policy_res_t result;
1359
1360                 /* Are all unused locks checked? */
1361                 if (remained-- <= 0)
1362                         break;
1363
1364                 /* For any flags, stop scanning if @max is reached. */
1365                 if (max && added >= max)
1366                         break;
1367
1368                 list_for_each_entry_safe(lock, next, &ns->ns_unused_list, l_lru){
1369                         /* No locks which got blocking requests. */
1370                         LASSERT(!(lock->l_flags & LDLM_FL_BL_AST));
1371
1372                         if (flags & LDLM_CANCEL_NO_WAIT &&
1373                             lock->l_flags & LDLM_FL_SKIPPED)
1374                                 /* already processed */
1375                                 continue;
1376
1377                         /* Somebody is already doing CANCEL. No need in this
1378                          * lock in lru, do not traverse it again. */
1379                         if (!(lock->l_flags & LDLM_FL_CANCELING))
1380                                 break;
1381
1382                         ldlm_lock_remove_from_lru_nolock(lock);
1383                 }
1384                 if (&lock->l_lru == &ns->ns_unused_list)
1385                         break;
1386
1387                 LDLM_LOCK_GET(lock); /* dropped by bl thread */
1388                 spin_unlock(&ns->ns_unused_lock);
1389
1390                 lock_res_and_lock(lock);
1391                 /* Pass the lock through the policy filter and see if it
1392                  * should stay in lru.
1393                  *
1394                  * Even for shrinker policy we stop scanning if
1395                  * we find a lock that should stay in the cache.
1396                  * We should take into account lock age anyway
1397                  * as new lock even if it is small of weight is
1398                  * valuable resource.
1399                  *
1400                  * That is, for shrinker policy we drop only
1401                  * old locks, but additionally chose them by
1402                  * their weight. Big extent locks will stay in
1403                  * the cache. */
1404                 result = pf(ns, lock, unused, added, count);
1405                 if (result == LDLM_POLICY_KEEP_LOCK) {
1406                         unlock_res_and_lock(lock);
1407                         LDLM_LOCK_PUT(lock);
1408                         spin_lock(&ns->ns_unused_lock);
1409                         break;
1410                 }
1411                 if (result == LDLM_POLICY_SKIP_LOCK) {
1412                         unlock_res_and_lock(lock);
1413                         LDLM_LOCK_PUT(lock);
1414                         spin_lock(&ns->ns_unused_lock);
1415                         continue;
1416                 }
1417
1418                 /* Check flags again under the lock. */
1419                 if ((lock->l_flags & LDLM_FL_CANCELING) ||
1420                     (ldlm_lock_remove_from_lru(lock) == 0)) {
1421                         /* other thread is removing lock from lru or
1422                          * somebody is already doing CANCEL or
1423                          * there is a blocking request which will send
1424                          * cancel by itseft or the lock is matched
1425                          * is already not unused. */
1426                         unlock_res_and_lock(lock);
1427                         LDLM_LOCK_PUT(lock);
1428                         spin_lock(&ns->ns_unused_lock);
1429                         continue;
1430                 }
1431                 LASSERT(!lock->l_readers && !lock->l_writers);
1432
1433                 /* If we have chosen to cancel this lock voluntarily, we
1434                  * better send cancel notification to server, so that it
1435                  * frees appropriate state. This might lead to a race
1436                  * where while we are doing cancel here, server is also
1437                  * silently cancelling this lock. */
1438                 lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK;
1439
1440                 /* Setting the CBPENDING flag is a little misleading,
1441                  * but prevents an important race; namely, once
1442                  * CBPENDING is set, the lock can accumulate no more
1443                  * readers/writers. Since readers and writers are
1444                  * already zero here, ldlm_lock_decref() won't see
1445                  * this flag and call l_blocking_ast */
1446                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1447
1448                 /* We can't re-add to l_lru as it confuses the
1449                  * refcounting in ldlm_lock_remove_from_lru() if an AST
1450                  * arrives after we drop ns_lock below. We use l_bl_ast
1451                  * and can't use l_pending_chain as it is used both on
1452                  * server and client nevertheless bug 5666 says it is
1453                  * used only on server */
1454                 LASSERT(list_empty(&lock->l_bl_ast));
1455                 list_add(&lock->l_bl_ast, cancels);
1456                 unlock_res_and_lock(lock);
1457                 spin_lock(&ns->ns_unused_lock);
1458                 added++;
1459                 unused--;
1460         }
1461         spin_unlock(&ns->ns_unused_lock);
1462         RETURN(ldlm_cancel_list(cancels, added, cancel_flags));
1463 }
1464
1465 /* when called with LDLM_ASYNC the blocking callback will be handled
1466  * in a thread and this function will return after the thread has been
1467  * asked to call the callback.  when called with LDLM_SYNC the blocking
1468  * callback will be performed in this function. */
1469 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr, ldlm_sync_t sync,
1470                     int flags)
1471 {
1472         CFS_LIST_HEAD(cancels);
1473         int count, rc;
1474         ENTRY;
1475
1476 #ifndef __KERNEL__
1477         sync = LDLM_SYNC; /* force to be sync in user space */
1478 #endif
1479         count = ldlm_cancel_lru_local(ns, &cancels, nr, 0, 0, flags);
1480         if (sync == LDLM_ASYNC) {
1481                 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count);
1482                 if (rc == 0)
1483                         RETURN(count);
1484         }
1485
1486         /* If an error occured in ASYNC mode, or this is SYNC mode,
1487          * cancel the list. */
1488         ldlm_cli_cancel_list(&cancels, count, NULL, 0);
1489         RETURN(count);
1490 }
1491
1492 /* Find and cancel locally unused locks found on resource, matched to the
1493  * given policy, mode. GET the found locks and add them into the @cancels
1494  * list. */
1495 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1496                                struct list_head *cancels,
1497                                ldlm_policy_data_t *policy,
1498                                ldlm_mode_t mode, int lock_flags,
1499                                int cancel_flags, void *opaque)
1500 {
1501         struct ldlm_lock *lock;
1502         int count = 0;
1503         ENTRY;
1504
1505         lock_res(res);
1506         list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1507                 if (opaque != NULL && lock->l_ast_data != opaque) {
1508                         LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1509                                    lock->l_ast_data, opaque);
1510                         //LBUG();
1511                         continue;
1512                 }
1513
1514                 if (lock->l_readers || lock->l_writers) {
1515                         if (cancel_flags & LDLM_FL_WARN) {
1516                                 LDLM_ERROR(lock, "lock in use");
1517                                 //LBUG();
1518                         }
1519                         continue;
1520                 }
1521
1522                 /* If somebody is already doing CANCEL, or blocking ast came,
1523                  * skip this lock. */
1524                 if (lock->l_flags & LDLM_FL_BL_AST ||
1525                     lock->l_flags & LDLM_FL_CANCELING)
1526                         continue;
1527
1528                 if (lockmode_compat(lock->l_granted_mode, mode))
1529                         continue;
1530
1531                 /* If policy is given and this is IBITS lock, add to list only
1532                  * those locks that match by policy. */
1533                 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1534                     !(lock->l_policy_data.l_inodebits.bits &
1535                       policy->l_inodebits.bits))
1536                         continue;
1537
1538                 /* See CBPENDING comment in ldlm_cancel_lru */
1539                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1540                         lock_flags;
1541
1542                 LASSERT(list_empty(&lock->l_bl_ast));
1543                 list_add(&lock->l_bl_ast, cancels);
1544                 LDLM_LOCK_GET(lock);
1545                 count++;
1546         }
1547         unlock_res(res);
1548
1549         /* Handle only @count inserted locks. */
1550         RETURN(ldlm_cancel_list(cancels, count, cancel_flags));
1551 }
1552
1553 /* If @req is NULL, send CANCEL request to server with handles of locks
1554  * in the @cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1555  * separately per lock.
1556  * If @req is not NULL, put handles of locks in @cancels into the request
1557  * buffer at the offset @off.
1558  * Destroy @cancels at the end. */
1559 int ldlm_cli_cancel_list(struct list_head *cancels, int count,
1560                          struct ptlrpc_request *req, int off)
1561 {
1562         struct ldlm_lock *lock;
1563         int res = 0;
1564         ENTRY;
1565
1566         if (list_empty(cancels) || count == 0)
1567                 RETURN(0);
1568
1569         while (count) {
1570                 LASSERT(!list_empty(cancels));
1571                 lock = list_entry(cancels->next, struct ldlm_lock, l_bl_ast);
1572                 LASSERT(lock->l_conn_export);
1573
1574                 if (exp_connect_cancelset(lock->l_conn_export)) {
1575                         res = count;
1576                         if (req)
1577                                 ldlm_cancel_pack(req, off, cancels, count);
1578                         else
1579                                 res = ldlm_cli_cancel_req(lock->l_conn_export,
1580                                                           cancels, count);
1581                 } else {
1582                         res = ldlm_cli_cancel_req(lock->l_conn_export,
1583                                                   cancels, 1);
1584                 }
1585
1586                 if (res < 0) {
1587                         CERROR("ldlm_cli_cancel_list: %d\n", res);
1588                         res = count;
1589                 }
1590
1591                 count -= res;
1592                 ldlm_lock_list_put(cancels, l_bl_ast, res);
1593         }
1594         RETURN(0);
1595 }
1596
1597 static int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1598                                            struct ldlm_res_id res_id,
1599                                            int flags, void *opaque)
1600 {
1601         struct ldlm_resource *res;
1602         CFS_LIST_HEAD(cancels);
1603         int count;
1604         int rc;
1605         ENTRY;
1606
1607         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1608         if (res == NULL) {
1609                 /* This is not a problem. */
1610                 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id.name[0]);
1611                 RETURN(0);
1612         }
1613
1614         count = ldlm_cancel_resource_local(res, &cancels, NULL, LCK_MINMODE,
1615                                            0, flags, opaque);
1616         rc = ldlm_cli_cancel_list(&cancels, count, NULL, 0);
1617         if (rc != ELDLM_OK)
1618                 CERROR("ldlm_cli_cancel_unused_resource: %d\n", rc);
1619
1620         ldlm_resource_putref(res);
1621         RETURN(0);
1622 }
1623
1624 static inline int have_no_nsresource(struct ldlm_namespace *ns)
1625 {
1626         int no_resource = 0;
1627
1628         spin_lock(&ns->ns_hash_lock);
1629         if (ns->ns_resources == 0)
1630                 no_resource = 1;
1631         spin_unlock(&ns->ns_hash_lock);
1632
1633         RETURN(no_resource);
1634 }
1635
1636 /* Cancel all locks on a namespace (or a specific resource, if given)
1637  * that have 0 readers/writers.
1638  *
1639  * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
1640  * to notify the server. */
1641 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1642                            struct ldlm_res_id *res_id, int flags, void *opaque)
1643 {
1644         int i;
1645         ENTRY;
1646
1647         if (ns == NULL)
1648                 RETURN(ELDLM_OK);
1649
1650         if (res_id)
1651                 RETURN(ldlm_cli_cancel_unused_resource(ns, *res_id, flags,
1652                                                        opaque));
1653
1654         spin_lock(&ns->ns_hash_lock);
1655         for (i = 0; i < RES_HASH_SIZE; i++) {
1656                 struct list_head *tmp;
1657                 tmp = ns->ns_hash[i].next;
1658                 while (tmp != &(ns->ns_hash[i])) {
1659                         struct ldlm_resource *res;
1660                         int rc;
1661
1662                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
1663                         ldlm_resource_getref(res);
1664                         spin_unlock(&ns->ns_hash_lock);
1665
1666                         rc = ldlm_cli_cancel_unused_resource(ns, res->lr_name,
1667                                                              flags, opaque);
1668
1669                         if (rc)
1670                                 CERROR("ldlm_cli_cancel_unused ("LPU64"): %d\n",
1671                                        res->lr_name.name[0], rc);
1672
1673                         spin_lock(&ns->ns_hash_lock);
1674                         tmp = tmp->next;
1675                         ldlm_resource_putref_locked(res);
1676                 }
1677         }
1678         spin_unlock(&ns->ns_hash_lock);
1679
1680         RETURN(ELDLM_OK);
1681 }
1682
1683 /* join/split resource locks to/from lru list */
1684 int ldlm_cli_join_lru(struct ldlm_namespace *ns,
1685                       struct ldlm_res_id *res_id, int join)
1686 {
1687         struct ldlm_resource *res;
1688         struct ldlm_lock *lock, *n;
1689         int count = 0;
1690         ENTRY;
1691
1692         LASSERT(ns_is_client(ns));
1693
1694         res = ldlm_resource_get(ns, NULL, *res_id, LDLM_EXTENT, 0);
1695         if (res == NULL)
1696                 RETURN(count);
1697         LASSERT(res->lr_type == LDLM_EXTENT);
1698
1699         lock_res(res);
1700         if (!join)
1701                 goto split;
1702
1703         list_for_each_entry_safe (lock, n, &res->lr_granted, l_res_link) {
1704                 if (list_empty(&lock->l_lru) &&
1705                     !lock->l_readers && !lock->l_writers &&
1706                     !(lock->l_flags & LDLM_FL_LOCAL) &&
1707                     !(lock->l_flags & LDLM_FL_CBPENDING) &&
1708                     !(lock->l_flags & LDLM_FL_BL_AST)) {
1709                         ldlm_lock_add_to_lru(lock);
1710                         lock->l_flags &= ~LDLM_FL_NO_LRU;
1711                         LDLM_DEBUG(lock, "join lock to lru");
1712                         count++;
1713                 }
1714         }
1715         goto unlock;
1716 split:
1717         spin_lock(&ns->ns_unused_lock);
1718         list_for_each_entry_safe (lock, n, &ns->ns_unused_list, l_lru) {
1719                 if (lock->l_resource == res) {
1720                         ldlm_lock_remove_from_lru_nolock(lock);
1721                         lock->l_flags |= LDLM_FL_NO_LRU;
1722                         LDLM_DEBUG(lock, "split lock from lru");
1723                         count++;
1724                 }
1725         }
1726         spin_unlock(&ns->ns_unused_lock);
1727 unlock:
1728         unlock_res(res);
1729         ldlm_resource_putref(res);
1730         RETURN(count);
1731 }
1732
1733 /* Lock iterators. */
1734
1735 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
1736                           void *closure)
1737 {
1738         struct list_head *tmp, *next;
1739         struct ldlm_lock *lock;
1740         int rc = LDLM_ITER_CONTINUE;
1741
1742         ENTRY;
1743
1744         if (!res)
1745                 RETURN(LDLM_ITER_CONTINUE);
1746
1747         lock_res(res);
1748         list_for_each_safe(tmp, next, &res->lr_granted) {
1749                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1750
1751                 if (iter(lock, closure) == LDLM_ITER_STOP)
1752                         GOTO(out, rc = LDLM_ITER_STOP);
1753         }
1754
1755         list_for_each_safe(tmp, next, &res->lr_converting) {
1756                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1757
1758                 if (iter(lock, closure) == LDLM_ITER_STOP)
1759                         GOTO(out, rc = LDLM_ITER_STOP);
1760         }
1761
1762         list_for_each_safe(tmp, next, &res->lr_waiting) {
1763                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1764
1765                 if (iter(lock, closure) == LDLM_ITER_STOP)
1766                         GOTO(out, rc = LDLM_ITER_STOP);
1767         }
1768  out:
1769         unlock_res(res);
1770         RETURN(rc);
1771 }
1772
1773 struct iter_helper_data {
1774         ldlm_iterator_t iter;
1775         void *closure;
1776 };
1777
1778 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
1779 {
1780         struct iter_helper_data *helper = closure;
1781         return helper->iter(lock, helper->closure);
1782 }
1783
1784 static int ldlm_res_iter_helper(struct ldlm_resource *res, void *closure)
1785 {
1786         return ldlm_resource_foreach(res, ldlm_iter_helper, closure);
1787 }
1788
1789 int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
1790                            void *closure)
1791 {
1792         struct iter_helper_data helper = { iter: iter, closure: closure };
1793         return ldlm_namespace_foreach_res(ns, ldlm_res_iter_helper, &helper);
1794 }
1795
1796 int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
1797                                ldlm_res_iterator_t iter, void *closure)
1798 {
1799         int i, rc = LDLM_ITER_CONTINUE;
1800         struct ldlm_resource *res;
1801         struct list_head *tmp;
1802
1803         ENTRY;
1804         spin_lock(&ns->ns_hash_lock);
1805         for (i = 0; i < RES_HASH_SIZE; i++) {
1806                 tmp = ns->ns_hash[i].next;
1807                 while (tmp != &(ns->ns_hash[i])) {
1808                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
1809                         ldlm_resource_getref(res);
1810                         spin_unlock(&ns->ns_hash_lock);
1811
1812                         rc = iter(res, closure);
1813
1814                         spin_lock(&ns->ns_hash_lock);
1815                         tmp = tmp->next;
1816                         ldlm_resource_putref_locked(res);
1817                         if (rc == LDLM_ITER_STOP)
1818                                 GOTO(out, rc);
1819                 }
1820         }
1821  out:
1822         spin_unlock(&ns->ns_hash_lock);
1823         RETURN(rc);
1824 }
1825
1826 /* non-blocking function to manipulate a lock whose cb_data is being put away.*/
1827 void ldlm_resource_iterate(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
1828                            ldlm_iterator_t iter, void *data)
1829 {
1830         struct ldlm_resource *res;
1831         ENTRY;
1832
1833         if (ns == NULL) {
1834                 CERROR("must pass in namespace\n");
1835                 LBUG();
1836         }
1837
1838         res = ldlm_resource_get(ns, NULL, *res_id, 0, 0);
1839         if (res == NULL) {
1840                 EXIT;
1841                 return;
1842         }
1843
1844         ldlm_resource_foreach(res, iter, data);
1845         ldlm_resource_putref(res);
1846         EXIT;
1847 }
1848
1849 /* Lock replay */
1850
1851 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
1852 {
1853         struct list_head *list = closure;
1854
1855         /* we use l_pending_chain here, because it's unused on clients. */
1856         LASSERTF(list_empty(&lock->l_pending_chain),"lock %p next %p prev %p\n",
1857                  lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
1858         /* bug 9573: don't replay locks left after eviction, or
1859          * bug 17614: locks being actively cancelled. Get a reference
1860          * on a lock so that it does not disapear under us (e.g. due to cancel)
1861          */
1862         if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_CANCELING))) {
1863                 list_add(&lock->l_pending_chain, list);
1864                 LDLM_LOCK_GET(lock);
1865         }
1866
1867         return LDLM_ITER_CONTINUE;
1868 }
1869
1870 static int replay_lock_interpret(struct ptlrpc_request *req,
1871                                  void *data, int rc)
1872 {
1873         struct ldlm_async_args *aa = data;
1874         struct lustre_handle  old_hash_key;
1875         struct ldlm_lock     *lock;
1876         struct ldlm_reply    *reply;
1877         struct obd_export    *exp;
1878
1879         ENTRY;
1880         atomic_dec(&req->rq_import->imp_replay_inflight);
1881         if (rc != ELDLM_OK)
1882                 GOTO(out, rc);
1883
1884         reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF, sizeof(*reply),
1885                                    lustre_swab_ldlm_reply);
1886         if (reply == NULL) {
1887                 CERROR("Can't unpack ldlm_reply\n");
1888                 GOTO (out, rc = -EPROTO);
1889         }
1890
1891         lock = ldlm_handle2lock(&aa->lock_handle);
1892         if (!lock) {
1893                 CERROR("received replay ack for unknown local cookie "LPX64
1894                        " remote cookie "LPX64 " from server %s id %s\n",
1895                        aa->lock_handle.cookie, reply->lock_handle.cookie,
1896                        req->rq_export->exp_client_uuid.uuid,
1897                        libcfs_id2str(req->rq_peer));
1898                 GOTO(out, rc = -ESTALE);
1899         }
1900
1901         old_hash_key = lock->l_remote_handle;
1902         lock->l_remote_handle = reply->lock_handle;
1903
1904         /* Key change rehash lock in per-export hash with new key */
1905         exp = req->rq_export;
1906         if (exp && exp->exp_lock_hash)
1907                 lustre_hash_rehash_key(exp->exp_lock_hash, &old_hash_key,
1908                                        &lock->l_remote_handle,
1909                                        &lock->l_exp_hash);
1910
1911         LDLM_DEBUG(lock, "replayed lock:");
1912         ptlrpc_import_recovery_state_machine(req->rq_import);
1913         LDLM_LOCK_PUT(lock);
1914 out:
1915         if (rc != ELDLM_OK)
1916                 ptlrpc_connect_import(req->rq_import, NULL);
1917
1918
1919         RETURN(rc);
1920 }
1921
1922 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
1923 {
1924         struct ptlrpc_request *req;
1925         struct ldlm_request *body;
1926         struct ldlm_reply *reply;
1927         struct ldlm_async_args *aa;
1928         int buffers = 2;
1929         __u32 size[3] = { sizeof(struct ptlrpc_body) };
1930         int flags;
1931         ENTRY;
1932
1933
1934         /* Bug 11974: Do not replay a lock which is actively being canceled */
1935         if (lock->l_flags & LDLM_FL_CANCELING) {
1936                 LDLM_DEBUG(lock, "Not replaying canceled lock:");
1937                 RETURN(0);
1938         }
1939
1940         /* If this is reply-less callback lock, we cannot replay it, since
1941          * server might have long dropped it, but notification of that event was
1942          * lost by network. (and server granted conflicting lock already) */
1943         if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1944                 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
1945                 ldlm_lock_cancel(lock);
1946                 RETURN(0);
1947         }
1948
1949         /*
1950          * If granted mode matches the requested mode, this lock is granted.
1951          *
1952          * If they differ, but we have a granted mode, then we were granted
1953          * one mode and now want another: ergo, converting.
1954          *
1955          * If we haven't been granted anything and are on a resource list,
1956          * then we're blocked/waiting.
1957          *
1958          * If we haven't been granted anything and we're NOT on a resource list,
1959          * then we haven't got a reply yet and don't have a known disposition.
1960          * This happens whenever a lock enqueue is the request that triggers
1961          * recovery.
1962          */
1963         if (lock->l_granted_mode == lock->l_req_mode)
1964                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
1965         else if (lock->l_granted_mode)
1966                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
1967         else if (!list_empty(&lock->l_res_link))
1968                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
1969         else
1970                 flags = LDLM_FL_REPLAY;
1971
1972         size[DLM_LOCKREQ_OFF] = sizeof(*body);
1973         req = ptlrpc_prep_req(imp, LUSTRE_DLM_VERSION, LDLM_ENQUEUE, 2, size,
1974                               NULL);
1975         if (!req)
1976                 RETURN(-ENOMEM);
1977
1978         /* We're part of recovery, so don't wait for it. */
1979         req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
1980
1981         body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, sizeof(*body));
1982         ldlm_lock2desc(lock, &body->lock_desc);
1983         body->lock_flags = flags;
1984
1985         ldlm_lock2handle(lock, &body->lock_handle[0]);
1986         size[DLM_LOCKREPLY_OFF] = sizeof(*reply);
1987         buffers = 3;
1988         if (lock->l_lvb_len != 0) {
1989                 size[DLM_REPLY_REC_OFF] = lock->l_lvb_len;
1990         } else
1991                 size[DLM_REPLY_REC_OFF] = sizeof (struct ost_lvb);
1992         ptlrpc_req_set_repsize(req, buffers, size);
1993
1994         LDLM_DEBUG(lock, "replaying lock:");
1995
1996         atomic_inc(&req->rq_import->imp_replay_inflight);
1997         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1998         aa = ptlrpc_req_async_args(req);
1999         aa->lock_handle = body->lock_handle[0];
2000         req->rq_interpret_reply = replay_lock_interpret;
2001         ptlrpcd_add_req(req);
2002
2003         RETURN(0);
2004 }
2005
2006 /* cancel as many unused locks as possible before replay. since we are
2007  * in recovery, we can't wait for any outstanding RPCs ro send any RPC
2008  * to the server.
2009  *
2010  * callced only in recovery before replaying locks. there is no need to
2011  * replay locks that are unused. since the clients may hold thousands of
2012  * cached unused locks, dropping the unused locks can greatly reduce the
2013  * load on the servers at recovery time.
2014  */
2015 static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns)
2016 {
2017         int canceled;
2018         CFS_LIST_HEAD(cancels);
2019
2020         CDEBUG(D_DLMTRACE, "Dropping as many as unused locks as possible before"                           "replay for namespace %s (%d)\n", ns->ns_name,
2021                            ns->ns_nr_unused);
2022
2023         /* we don't need to care whether or not LRU resize is enabled
2024          * because the LDLM_CANCLE_NO_WAIT policy doesn't use the
2025          * count parameter */
2026         canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0,
2027                                          LDLM_FL_LOCAL_ONLY,
2028                                          LDLM_CANCEL_NO_WAIT);
2029
2030         CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n",
2031                            canceled, ns->ns_name);
2032 }
2033
2034 int ldlm_replay_locks(struct obd_import *imp)
2035 {
2036         struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2037         struct list_head list;
2038         struct ldlm_lock *lock, *next;
2039         int rc = 0;
2040
2041         ENTRY;
2042         CFS_INIT_LIST_HEAD(&list);
2043
2044         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
2045
2046         /* ensure this doesn't fall to 0 before all have been queued */
2047         atomic_inc(&imp->imp_replay_inflight);
2048
2049         if (ldlm_cancel_unused_locks_before_replay)
2050                 ldlm_cancel_unused_locks_for_replay(ns);
2051
2052         if (imp->imp_no_lock_replay) {
2053                 /* VBR: locks should be cancelled here */
2054                 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2055                 spin_lock(&imp->imp_lock);
2056                 imp->imp_no_lock_replay = 0;
2057                 spin_unlock(&imp->imp_lock);
2058         } else {
2059                 (void)ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay,
2060                                              &list);
2061                 list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2062                         list_del_init(&lock->l_pending_chain);
2063                         if (rc) {
2064                                 LDLM_LOCK_PUT(lock);
2065                                 continue; /* or try to do the rest? */
2066                         }
2067                         rc = replay_one_lock(imp, lock);
2068                         LDLM_LOCK_PUT(lock);
2069                 }
2070         }
2071         atomic_dec(&imp->imp_replay_inflight);
2072
2073         RETURN(rc);
2074 }