lustre/ldlm/ldlm_request.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (C) 2002, 2003 Cluster File Systems, Inc.
   5  *
   6  *   This file is part of the Lustre file system, http://www.lustre.org
   7  *   Lustre is a trademark of Cluster File Systems, Inc.
   8  *
   9  *   You may have signed or agreed to another license before downloading
  10  *   this software.  If so, you are bound by the terms and conditions
  11  *   of that agreement, and the following does not apply to you.  See the
  12  *   LICENSE file included with this distribution for more information.
  13  *
  14  *   If you did not agree to a different license, then this copy of Lustre
  15  *   is open source software; you can redistribute it and/or modify it
  16  *   under the terms of version 2 of the GNU General Public License as
  17  *   published by the Free Software Foundation.
  18  *
  19  *   In either case, Lustre is distributed in the hope that it will be
  20  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  21  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  22  *   license text for more details.
  23  */
  24
  25 #define DEBUG_SUBSYSTEM S_LDLM
  26 #ifndef __KERNEL__
  27 #include <signal.h>
  28 #include <liblustre.h>
  29 #endif
  30
  31 #include <lustre_dlm.h>
  32 #include <obd_class.h>
  33 #include <obd.h>
  34
  35 #include "ldlm_internal.h"
  36
  37 int ldlm_enqueue_min = OBD_TIMEOUT_DEFAULT;
  38 CFS_MODULE_PARM(ldlm_enqueue_min, "i", int, 0644,
  39                 "lock enqueue timeout minimum");
  40
  41 static void interrupted_completion_wait(void *data)
  42 {
  43 }
  44
  45 struct lock_wait_data {
  46         struct ldlm_lock *lwd_lock;
  47         __u32             lwd_conn_cnt;
  48 };
  49
  50 struct ldlm_async_args {
  51         struct lustre_handle lock_handle;
  52 };
  53
  54 int ldlm_expired_completion_wait(void *data)
  55 {
  56         struct lock_wait_data *lwd = data;
  57         struct ldlm_lock *lock = lwd->lwd_lock;
  58         struct obd_import *imp;
  59         struct obd_device *obd;
  60
  61         ENTRY;
  62         if (lock->l_conn_export == NULL) {
  63                 static cfs_time_t next_dump = 0, last_dump = 0;
  64
  65                 LDLM_ERROR(lock, "lock timed out (enqueued at %lu, %lus ago); "
  66                            "not entering recovery in server code, just going "
  67                            "back to sleep", lock->l_enqueued_time.tv_sec,
  68                            cfs_time_current_sec() -
  69                            lock->l_enqueued_time.tv_sec);
  70                 if (cfs_time_after(cfs_time_current(), next_dump)) {
  71                         last_dump = next_dump;
  72                         next_dump = cfs_time_shift(300);
  73                         ldlm_namespace_dump(D_DLMTRACE,
  74                                             lock->l_resource->lr_namespace);
  75                         if (last_dump == 0)
  76                                 libcfs_debug_dumplog();
  77                 }
  78                 RETURN(0);
  79         }
  80
  81         obd = lock->l_conn_export->exp_obd;
  82         imp = obd->u.cli.cl_import;
  83         ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
  84         LDLM_ERROR(lock, "lock timed out (enqueued at %lu, %lus ago), entering "
  85                    "recovery for %s@%s", lock->l_enqueued_time.tv_sec,
  86                    CURRENT_SECONDS - lock->l_enqueued_time.tv_sec,
  87                    obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
  88
  89         RETURN(0);
  90 }
  91
  92 /* We use the same basis for both server side and client side functions
  93    from a single node. */
  94 int ldlm_get_enq_timeout(struct ldlm_lock *lock)
  95 {
  96         int timeout = at_get(&lock->l_resource->lr_namespace->ns_at_estimate);
  97         if (AT_OFF)
  98                 return obd_timeout / 2;
  99         /* Since these are non-updating timeouts, we should be conservative.
 100            It would be nice to have some kind of "early reply" mechanism for
 101            lock callbacks too... */
 102         timeout = timeout + (timeout >> 1); /* 150% */
 103         return max(timeout, ldlm_enqueue_min);
 104 }
 105
 106 int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data)
 107 {
 108         /* XXX ALLOCATE - 160 bytes */
 109         struct lock_wait_data lwd;
 110         struct obd_device *obd;
 111         struct obd_import *imp = NULL;
 112         struct l_wait_info lwi;
 113         __u32 timeout;
 114         int rc = 0;
 115         ENTRY;
 116
 117         if (flags == LDLM_FL_WAIT_NOREPROC) {
 118                 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
 119                 goto noreproc;
 120         }
 121
 122         if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
 123                        LDLM_FL_BLOCK_CONV))) {
 124                 cfs_waitq_signal(&lock->l_waitq);
 125                 RETURN(0);
 126         }
 127
 128         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
 129                    "sleeping");
 130         ldlm_lock_dump(D_OTHER, lock, 0);
 131         ldlm_reprocess_all(lock->l_resource);
 132
 133 noreproc:
 134
 135         obd = class_exp2obd(lock->l_conn_export);
 136
 137         /* if this is a local lock, then there is no import */
 138         if (obd != NULL) {
 139                 imp = obd->u.cli.cl_import;
 140         }
 141
 142         /* Wait a long time for enqueue - server may have to callback a
 143            lock from another client.  Server will evict the other client if it
 144            doesn't respond reasonably, and then give us the lock. */
 145         timeout = ldlm_get_enq_timeout(lock) * 2;
 146
 147         lwd.lwd_lock = lock;
 148
 149         if (lock->l_flags & LDLM_FL_NO_TIMEOUT) {
 150                 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
 151                 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
 152         } else {
 153                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout),
 154                                        ldlm_expired_completion_wait,
 155                                        interrupted_completion_wait, &lwd);
 156         }
 157
 158         if (imp != NULL) {
 159                 spin_lock(&imp->imp_lock);
 160                 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
 161                 spin_unlock(&imp->imp_lock);
 162         }
 163
 164         /* Go to sleep until the lock is granted or cancelled. */
 165         rc = l_wait_event(lock->l_waitq,
 166                           ((lock->l_req_mode == lock->l_granted_mode) ||
 167                            (lock->l_flags & LDLM_FL_FAILED)), &lwi);
 168
 169         if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED) {
 170                 LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed");
 171                 RETURN(-EIO);
 172         }
 173
 174         if (rc) {
 175                 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
 176                            rc);
 177                 RETURN(rc);
 178         }
 179
 180         LDLM_DEBUG(lock, "client-side enqueue waking up: granted after %lds",
 181                    cfs_time_current_sec() - lock->l_enqueued_time.tv_sec);
 182
 183         /* Update our time estimate */
 184         at_add(&lock->l_resource->lr_namespace->ns_at_estimate,
 185                cfs_time_current_sec() - lock->l_enqueued_time.tv_sec);
 186
 187         RETURN(0);
 188 }
 189
 190 /*
 191  * ->l_blocking_ast() callback for LDLM locks acquired by server-side OBDs.
 192  */
 193 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 194                       void *data, int flag)
 195 {
 196         int do_ast;
 197         ENTRY;
 198
 199         if (flag == LDLM_CB_CANCELING) {
 200                 /* Don't need to do anything here. */
 201                 RETURN(0);
 202         }
 203
 204         lock_res_and_lock(lock);
 205         /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
 206          * that ldlm_blocking_ast is called just before intent_policy method
 207          * takes the ns_lock, then by the time we get the lock, we might not
 208          * be the correct blocking function anymore.  So check, and return
 209          * early, if so. */
 210         if (lock->l_blocking_ast != ldlm_blocking_ast) {
 211                 unlock_res_and_lock(lock);
 212                 RETURN(0);
 213         }
 214
 215         lock->l_flags |= LDLM_FL_CBPENDING;
 216         do_ast = (!lock->l_readers && !lock->l_writers);
 217         unlock_res_and_lock(lock);
 218
 219         if (do_ast) {
 220                 struct lustre_handle lockh;
 221                 int rc;
 222
 223                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
 224                 ldlm_lock2handle(lock, &lockh);
 225                 rc = ldlm_cli_cancel(&lockh);
 226                 if (rc < 0)
 227                         CERROR("ldlm_cli_cancel: %d\n", rc);
 228         } else {
 229                 LDLM_DEBUG(lock, "Lock still has references, will be "
 230                            "cancelled later");
 231         }
 232         RETURN(0);
 233 }
 234
 235 /*
 236  * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See
 237  * comment in filter_intent_policy() on why you may need this.
 238  */
 239 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
 240 {
 241         /*
 242          * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for
 243          * that is rather subtle: with OST-side locking, it may so happen that
 244          * _all_ extent locks are held by the OST. If client wants to obtain
 245          * current file size it calls ll{,u}_glimpse_size(), and (as locks are
 246          * on the server), dummy glimpse callback fires and does
 247          * nothing. Client still receives correct file size due to the
 248          * following fragment in filter_intent_policy():
 249          *
 250          * rc = l->l_glimpse_ast(l, NULL); // this will update the LVB
 251          * if (rc != 0 && res->lr_namespace->ns_lvbo &&
 252          *     res->lr_namespace->ns_lvbo->lvbo_update) {
 253          *         res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1);
 254          * }
 255          *
 256          * that is, after glimpse_ast() fails, filter_lvbo_update() runs, and
 257          * returns correct file size to the client.
 258          */
 259         return -ELDLM_NO_LOCK_DATA;
 260 }
 261
 262 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns, struct ldlm_res_id res_id,
 263                            ldlm_type_t type, ldlm_policy_data_t *policy,
 264                            ldlm_mode_t mode, int *flags,
 265                            ldlm_blocking_callback blocking,
 266                            ldlm_completion_callback completion,
 267                            ldlm_glimpse_callback glimpse,
 268                            void *data, __u32 lvb_len, void *lvb_swabber,
 269                            struct lustre_handle *lockh)
 270 {
 271         struct ldlm_lock *lock;
 272         int err;
 273         ENTRY;
 274
 275         LASSERT(!(*flags & LDLM_FL_REPLAY));
 276         if (ns_is_client(ns)) {
 277                 CERROR("Trying to enqueue local lock in a shadow namespace\n");
 278                 LBUG();
 279         }
 280
 281         lock = ldlm_lock_create(ns, res_id, type, mode, blocking,
 282                                 completion, glimpse, data, lvb_len);
 283         if (!lock)
 284                 GOTO(out_nolock, err = -ENOMEM);
 285         LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
 286
 287         ldlm_lock_addref_internal(lock, mode);
 288         ldlm_lock2handle(lock, lockh);
 289         lock_res_and_lock(lock);
 290         lock->l_flags |= LDLM_FL_LOCAL;
 291         if (*flags & LDLM_FL_ATOMIC_CB)
 292                 lock->l_flags |= LDLM_FL_ATOMIC_CB;
 293         lock->l_lvb_swabber = lvb_swabber;
 294         unlock_res_and_lock(lock);
 295         if (policy != NULL)
 296                 lock->l_policy_data = *policy;
 297         if (type == LDLM_EXTENT)
 298                 lock->l_req_extent = policy->l_extent;
 299
 300         err = ldlm_lock_enqueue(ns, &lock, policy, flags);
 301         if (err != ELDLM_OK)
 302                 GOTO(out, err);
 303
 304         if (policy != NULL)
 305                 *policy = lock->l_policy_data;
 306         if ((*flags) & LDLM_FL_LOCK_CHANGED)
 307                 res_id = lock->l_resource->lr_name;
 308
 309         LDLM_DEBUG_NOLOCK("client-side local enqueue handler END (lock %p)",
 310                           lock);
 311
 312         if (lock->l_completion_ast)
 313                 lock->l_completion_ast(lock, *flags, NULL);
 314
 315         LDLM_DEBUG(lock, "client-side local enqueue END");
 316         EXIT;
 317  out:
 318         LDLM_LOCK_PUT(lock);
 319  out_nolock:
 320         return err;
 321 }
 322
 323 static void failed_lock_cleanup(struct ldlm_namespace *ns,
 324                                 struct ldlm_lock *lock,
 325                                 struct lustre_handle *lockh, int mode)
 326 {
 327         /* Set a flag to prevent us from sending a CANCEL (bug 407) */
 328         lock_res_and_lock(lock);
 329         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 330         unlock_res_and_lock(lock);
 331         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 332
 333         ldlm_lock_decref_and_cancel(lockh, mode);
 334
 335         /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
 336          *       from llite/file.c/ll_file_flock(). */
 337         if (lock->l_resource->lr_type == LDLM_FLOCK) {
 338                 ldlm_lock_destroy(lock);
 339         }
 340 }
 341
 342 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
 343                           ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
 344                           int *flags, void *lvb, __u32 lvb_len,
 345                           void *lvb_swabber, struct lustre_handle *lockh,int rc)
 346 {
 347         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 348         int is_replay = *flags & LDLM_FL_REPLAY;
 349         struct ldlm_lock *lock;
 350         struct ldlm_reply *reply;
 351         int cleanup_phase = 1;
 352         ENTRY;
 353
 354         lock = ldlm_handle2lock(lockh);
 355         /* ldlm_cli_enqueue is holding a reference on this lock. */
 356         if (!lock) {
 357                 LASSERT(type == LDLM_FLOCK);
 358                 RETURN(-ENOLCK);
 359         }
 360
 361         if (rc != ELDLM_OK) {
 362                 LASSERT(!is_replay);
 363                 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
 364                            rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
 365                 if (rc == ELDLM_LOCK_ABORTED) {
 366                         /* Before we return, swab the reply */
 367                         reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF,
 368                                                    sizeof(*reply),
 369                                                    lustre_swab_ldlm_reply);
 370                         if (reply == NULL) {
 371                                 CERROR("Can't unpack ldlm_reply\n");
 372                                 rc = -EPROTO;
 373                         }
 374                         if (lvb_len) {
 375                                 void *tmplvb;
 376                                 tmplvb = lustre_swab_repbuf(req,
 377                                                             DLM_REPLY_REC_OFF,
 378                                                             lvb_len,
 379                                                             lvb_swabber);
 380                                 if (tmplvb == NULL)
 381                                         GOTO(cleanup, rc = -EPROTO);
 382                                 if (lvb != NULL)
 383                                         memcpy(lvb, tmplvb, lvb_len);
 384                         }
 385                 }
 386                 GOTO(cleanup, rc);
 387         }
 388
 389         reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF, sizeof(*reply),
 390                                    lustre_swab_ldlm_reply);
 391         if (reply == NULL) {
 392                 CERROR("Can't unpack ldlm_reply\n");
 393                 GOTO(cleanup, rc = -EPROTO);
 394         }
 395
 396         /* lock enqueued on the server */
 397         cleanup_phase = 0;
 398
 399         lock_res_and_lock(lock);
 400         lock->l_remote_handle = reply->lock_handle;
 401         *flags = reply->lock_flags;
 402         lock->l_flags |= reply->lock_flags & LDLM_INHERIT_FLAGS;
 403         /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match()
 404          * to wait with no timeout as well */
 405         lock->l_flags |= reply->lock_flags & LDLM_FL_NO_TIMEOUT;
 406         unlock_res_and_lock(lock);
 407
 408         CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%x\n",
 409                lock, reply->lock_handle.cookie, *flags);
 410
 411         /* If enqueue returned a blocked lock but the completion handler has
 412          * already run, then it fixed up the resource and we don't need to do it
 413          * again. */
 414         if ((*flags) & LDLM_FL_LOCK_CHANGED) {
 415                 int newmode = reply->lock_desc.l_req_mode;
 416                 LASSERT(!is_replay);
 417                 if (newmode && newmode != lock->l_req_mode) {
 418                         LDLM_DEBUG(lock, "server returned different mode %s",
 419                                    ldlm_lockname[newmode]);
 420                         lock->l_req_mode = newmode;
 421                 }
 422
 423                 if (reply->lock_desc.l_resource.lr_name.name[0] !=
 424                     lock->l_resource->lr_name.name[0]) {
 425                         CDEBUG(D_INFO, "remote intent success, locking %ld "
 426                                "instead of %ld\n",
 427                               (long)reply->lock_desc.l_resource.lr_name.name[0],
 428                                (long)lock->l_resource->lr_name.name[0]);
 429
 430                         ldlm_lock_change_resource(ns, lock,
 431                                            reply->lock_desc.l_resource.lr_name);
 432                         if (lock->l_resource == NULL) {
 433                                 LBUG();
 434                                 GOTO(cleanup, rc = -ENOMEM);
 435                         }
 436                         LDLM_DEBUG(lock, "client-side enqueue, new resource");
 437                 }
 438                 if (with_policy)
 439                         if (!(type == LDLM_IBITS && !(exp->exp_connect_flags &
 440                                                     OBD_CONNECT_IBITS)))
 441                                 lock->l_policy_data =
 442                                                  reply->lock_desc.l_policy_data;
 443                 if (type != LDLM_PLAIN)
 444                         LDLM_DEBUG(lock,"client-side enqueue, new policy data");
 445         }
 446
 447         if ((*flags) & LDLM_FL_AST_SENT ||
 448             /* Cancel extent locks as soon as possible on a liblustre client,
 449              * because it cannot handle asynchronous ASTs robustly (see
 450              * bug 7311). */
 451             (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) {
 452                 lock_res_and_lock(lock);
 453                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
 454                 unlock_res_and_lock(lock);
 455                 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
 456         }
 457
 458         /* If the lock has already been granted by a completion AST, don't
 459          * clobber the LVB with an older one. */
 460         if (lvb_len && (lock->l_req_mode != lock->l_granted_mode)) {
 461                 void *tmplvb;
 462                 tmplvb = lustre_swab_repbuf(req, DLM_REPLY_REC_OFF, lvb_len,
 463                                             lvb_swabber);
 464                 if (tmplvb == NULL)
 465                         GOTO(cleanup, rc = -EPROTO);
 466                 memcpy(lock->l_lvb_data, tmplvb, lvb_len);
 467         }
 468
 469         if (!is_replay) {
 470                 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
 471                 if (lock->l_completion_ast != NULL) {
 472                         int err = lock->l_completion_ast(lock, *flags, NULL);
 473                         if (!rc)
 474                                 rc = err;
 475                         if (rc && type != LDLM_FLOCK) /* bug 9425, bug 10250 */
 476                                 cleanup_phase = 1;
 477                 }
 478         }
 479
 480         if (lvb_len && lvb != NULL) {
 481                 /* Copy the LVB here, and not earlier, because the completion
 482                  * AST (if any) can override what we got in the reply */
 483                 memcpy(lvb, lock->l_lvb_data, lvb_len);
 484         }
 485
 486         LDLM_DEBUG(lock, "client-side enqueue END");
 487         EXIT;
 488 cleanup:
 489         if (cleanup_phase == 1 && rc)
 490                 failed_lock_cleanup(ns, lock, lockh, mode);
 491         /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
 492         LDLM_LOCK_PUT(lock);
 493         LDLM_LOCK_PUT(lock);
 494         return rc;
 495 }
 496
 497 /* PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
 498  * a single page on the send/receive side. XXX: 512 should be changed
 499  * to more adequate value. */
 500 #define ldlm_req_handles_avail(exp, size, bufcount, off)                \
 501 ({                                                                      \
 502         int _avail = min_t(int, LDLM_MAXREQSIZE, PAGE_SIZE - 512);      \
 503         int _s = size[DLM_LOCKREQ_OFF];                                 \
 504         size[DLM_LOCKREQ_OFF] = sizeof(struct ldlm_request);            \
 505         _avail -= lustre_msg_size(class_exp2cliimp(exp)->imp_msg_magic, \
 506                                   bufcount, size);                      \
 507         _avail /= sizeof(struct lustre_handle);                         \
 508         _avail += LDLM_LOCKREQ_HANDLES - off;                           \
 509         size[DLM_LOCKREQ_OFF] = _s;                                     \
 510         _avail;                                                         \
 511 })
 512
 513 /* Cancel lru locks and pack them into the enqueue request. Pack there the given
 514  * @count locks in @cancels. */
 515 struct ptlrpc_request *ldlm_prep_enqueue_req(struct obd_export *exp,
 516                                              int bufcount, int *size,
 517                                              struct list_head *cancels,
 518                                              int count)
 519 {
 520         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 521         struct ldlm_request *dlm = NULL;
 522         struct ptlrpc_request *req;
 523         CFS_LIST_HEAD(head);
 524         ENTRY;
 525
 526         if (cancels == NULL)
 527                 cancels = &head;
 528         if (exp_connect_cancelset(exp)) {
 529                 /* Estimate the amount of free space in the request. */
 530                 int avail = ldlm_req_handles_avail(exp, size, bufcount,
 531                                                    LDLM_ENQUEUE_CANCEL_OFF);
 532                 int flags, cancel;
 533
 534                 LASSERT(avail >= count);
 535
 536                 flags = ns_connect_lru_resize(ns) ?
 537                         LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
 538                 cancel = ns_connect_lru_resize(ns) ? 0 : 1;
 539
 540                 /* Cancel lru locks here _only_ if the server supports
 541                  * EARLY_CANCEL. Otherwise we have to send extra CANCEL
 542                  * rpc right on enqueue, what will make it slower, vs.
 543                  * asynchronous rpc in blocking thread. */
 544                 count += ldlm_cancel_lru_local(ns, cancels, cancel,
 545                                                avail - count, flags);
 546                 size[DLM_LOCKREQ_OFF] =
 547                         ldlm_request_bufsize(count, LDLM_ENQUEUE);
 548         }
 549         req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
 550                               LDLM_ENQUEUE, bufcount, size, NULL);
 551         if (exp_connect_cancelset(exp) && req) {
 552                 dlm = lustre_msg_buf(req->rq_reqmsg,
 553                                      DLM_LOCKREQ_OFF, sizeof(*dlm));
 554                 /* Skip first lock handler in ldlm_request_pack(), this method
 555                  * will incrment @lock_count according to the lock handle amount
 556                  * actually written to the buffer. */
 557                 dlm->lock_count = LDLM_ENQUEUE_CANCEL_OFF;
 558                 ldlm_cli_cancel_list(cancels, count, req, DLM_LOCKREQ_OFF);
 559         } else {
 560                 ldlm_lock_list_put(cancels, l_bl_ast, count);
 561         }
 562         RETURN(req);
 563 }
 564
 565 /* If a request has some specific initialisation it is passed in @reqp,
 566  * otherwise it is created in ldlm_cli_enqueue.
 567  *
 568  * Supports sync and async requests, pass @async flag accordingly. If a
 569  * request was created in ldlm_cli_enqueue and it is the async request,
 570  * pass it to the caller in @reqp. */
 571 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
 572                      struct ldlm_enqueue_info *einfo, struct ldlm_res_id res_id,
 573                      ldlm_policy_data_t *policy, int *flags,
 574                      void *lvb, __u32 lvb_len, void *lvb_swabber,
 575                      struct lustre_handle *lockh, int async)
 576 {
 577         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 578         struct ldlm_lock *lock;
 579         struct ldlm_request *body;
 580         struct ldlm_reply *reply;
 581         int size[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
 582                         [DLM_LOCKREQ_OFF]     = sizeof(*body),
 583                         [DLM_REPLY_REC_OFF]   = lvb_len };
 584         int is_replay = *flags & LDLM_FL_REPLAY;
 585         int req_passed_in = 1, rc, err;
 586         struct ptlrpc_request *req;
 587         ENTRY;
 588
 589         LASSERT(exp != NULL);
 590
 591         /* If we're replaying this lock, just check some invariants.
 592          * If we're creating a new lock, get everything all setup nice. */
 593         if (is_replay) {
 594                 lock = ldlm_handle2lock(lockh);
 595                 LASSERT(lock != NULL);
 596                 LDLM_DEBUG(lock, "client-side enqueue START");
 597                 LASSERT(exp == lock->l_conn_export);
 598         } else {
 599                 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
 600                                         einfo->ei_mode, einfo->ei_cb_bl,
 601                                         einfo->ei_cb_cp, einfo->ei_cb_gl,
 602                                         einfo->ei_cbdata, lvb_len);
 603                 if (lock == NULL)
 604                         RETURN(-ENOMEM);
 605                 /* for the local lock, add the reference */
 606                 ldlm_lock_addref_internal(lock, einfo->ei_mode);
 607                 ldlm_lock2handle(lock, lockh);
 608                 lock->l_lvb_swabber = lvb_swabber;
 609                 if (policy != NULL) {
 610                         /* INODEBITS_INTEROP: If the server does not support
 611                          * inodebits, we will request a plain lock in the
 612                          * descriptor (ldlm_lock2desc() below) but use an
 613                          * inodebits lock internally with both bits set.
 614                          */
 615                         if (einfo->ei_type == LDLM_IBITS &&
 616                             !(exp->exp_connect_flags & OBD_CONNECT_IBITS))
 617                                 lock->l_policy_data.l_inodebits.bits =
 618                                         MDS_INODELOCK_LOOKUP |
 619                                         MDS_INODELOCK_UPDATE;
 620                         else
 621                                 lock->l_policy_data = *policy;
 622                 }
 623
 624                 if (einfo->ei_type == LDLM_EXTENT)
 625                         lock->l_req_extent = policy->l_extent;
 626                 LDLM_DEBUG(lock, "client-side enqueue START");
 627         }
 628
 629         /* lock not sent to server yet */
 630
 631         if (reqp == NULL || *reqp == NULL) {
 632                 req = ldlm_prep_enqueue_req(exp, 2, size, NULL, 0);
 633                 if (req == NULL) {
 634                         failed_lock_cleanup(ns, lock, lockh, einfo->ei_mode);
 635                         LDLM_LOCK_PUT(lock);
 636                         RETURN(-ENOMEM);
 637                 }
 638                 req_passed_in = 0;
 639                 if (reqp)
 640                         *reqp = req;
 641         } else {
 642                 req = *reqp;
 643                 LASSERTF(lustre_msg_buflen(req->rq_reqmsg, DLM_LOCKREQ_OFF) >=
 644                          sizeof(*body), "buflen[%d] = %d, not "LPSZ"\n",
 645                          DLM_LOCKREQ_OFF,
 646                          lustre_msg_buflen(req->rq_reqmsg, DLM_LOCKREQ_OFF),
 647                          sizeof(*body));
 648         }
 649
 650         lock->l_conn_export = exp;
 651         lock->l_export = NULL;
 652         lock->l_blocking_ast = einfo->ei_cb_bl;
 653
 654         /* Dump lock data into the request buffer */
 655         body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, sizeof(*body));
 656         ldlm_lock2desc(lock, &body->lock_desc);
 657         body->lock_flags = *flags;
 658         body->lock_handle[0] = *lockh;
 659
 660         /* Continue as normal. */
 661         if (!req_passed_in) {
 662                 size[DLM_LOCKREPLY_OFF] = sizeof(*reply);
 663                 ptlrpc_req_set_repsize(req, 2 + (lvb_len > 0), size);
 664         }
 665
 666         /*
 667          * Liblustre client doesn't get extent locks, except for O_APPEND case
 668          * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
 669          * [i_size, OBD_OBJECT_EOF] lock is taken.
 670          */
 671         LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT ||
 672                      policy->l_extent.end == OBD_OBJECT_EOF));
 673
 674         if (async) {
 675                 LASSERT(reqp != NULL);
 676                 RETURN(0);
 677         }
 678
 679         LDLM_DEBUG(lock, "sending request");
 680         rc = ptlrpc_queue_wait(req);
 681         err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
 682                                     einfo->ei_mode, flags, lvb, lvb_len,
 683                                     lvb_swabber, lockh, rc);
 684
 685         /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
 686          * one reference that we took */
 687         if (err == -ENOLCK)
 688                 LDLM_LOCK_PUT(lock);
 689         else
 690                 rc = err;
 691
 692         if (!req_passed_in && req != NULL) {
 693                 ptlrpc_req_finished(req);
 694                 if (reqp)
 695                         *reqp = NULL;
 696         }
 697
 698         RETURN(rc);
 699 }
 700
 701 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
 702                                   int *flags)
 703 {
 704         struct ldlm_resource *res;
 705         int rc;
 706         ENTRY;
 707         if (ns_is_client(lock->l_resource->lr_namespace)) {
 708                 CERROR("Trying to cancel local lock\n");
 709                 LBUG();
 710         }
 711         LDLM_DEBUG(lock, "client-side local convert");
 712
 713         res = ldlm_lock_convert(lock, new_mode, flags);
 714         if (res) {
 715                 ldlm_reprocess_all(res);
 716                 rc = 0;
 717         } else {
 718                 rc = EDEADLOCK;
 719         }
 720         LDLM_DEBUG(lock, "client-side local convert handler END");
 721         LDLM_LOCK_PUT(lock);
 722         RETURN(rc);
 723 }
 724
 725 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
 726  * conversion of locks which are on the waiting or converting queue */
 727 /* Caller of this code is supposed to take care of lock readers/writers
 728    accounting */
 729 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, int *flags)
 730 {
 731         struct ldlm_request *body;
 732         struct ldlm_reply *reply;
 733         struct ldlm_lock *lock;
 734         struct ldlm_resource *res;
 735         struct ptlrpc_request *req;
 736         int size[2] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
 737                         [DLM_LOCKREQ_OFF]     = sizeof(*body) };
 738         int rc;
 739         ENTRY;
 740
 741         lock = ldlm_handle2lock(lockh);
 742         if (!lock) {
 743                 LBUG();
 744                 RETURN(-EINVAL);
 745         }
 746         *flags = 0;
 747
 748         if (lock->l_conn_export == NULL)
 749                 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
 750
 751         LDLM_DEBUG(lock, "client-side convert");
 752
 753         req = ptlrpc_prep_req(class_exp2cliimp(lock->l_conn_export),
 754                               LUSTRE_DLM_VERSION, LDLM_CONVERT, 2, size, NULL);
 755         if (!req)
 756                 GOTO(out, rc = -ENOMEM);
 757
 758         body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, sizeof(*body));
 759         body->lock_handle[0] = lock->l_remote_handle;
 760
 761         body->lock_desc.l_req_mode = new_mode;
 762         body->lock_flags = *flags;
 763
 764         size[DLM_LOCKREPLY_OFF] = sizeof(*reply);
 765         ptlrpc_req_set_repsize(req, 2, size);
 766
 767         rc = ptlrpc_queue_wait(req);
 768         if (rc != ELDLM_OK)
 769                 GOTO(out, rc);
 770
 771         reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF, sizeof(*reply),
 772                                    lustre_swab_ldlm_reply);
 773         if (reply == NULL) {
 774                 CERROR ("Can't unpack ldlm_reply\n");
 775                 GOTO (out, rc = -EPROTO);
 776         }
 777
 778         if (req->rq_status)
 779                 GOTO(out, rc = req->rq_status);
 780
 781         res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
 782         if (res != NULL) {
 783                 ldlm_reprocess_all(res);
 784                 /* Go to sleep until the lock is granted. */
 785                 /* FIXME: or cancelled. */
 786                 if (lock->l_completion_ast) {
 787                         rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
 788                                                     NULL);
 789                         if (rc)
 790                                 GOTO(out, rc);
 791                 }
 792         } else {
 793                 rc = EDEADLOCK;
 794         }
 795         EXIT;
 796  out:
 797         LDLM_LOCK_PUT(lock);
 798         ptlrpc_req_finished(req);
 799         return rc;
 800 }
 801
 802 /* Cancel locks locally.
 803  * Returns:
 804  * LDLM_FL_LOCAL_ONLY if tere is no need in a CANCEL rpc to the server;
 805  * LDLM_FL_CANCELING otherwise;
 806  * LDLM_FL_BL_AST if there is a need in a separate CANCEL rpc. */
 807 static int ldlm_cli_cancel_local(struct ldlm_lock *lock)
 808 {
 809         int rc = LDLM_FL_LOCAL_ONLY;
 810         ENTRY;
 811
 812         if (lock->l_conn_export) {
 813                 int local_only;
 814
 815                 LDLM_DEBUG(lock, "client-side cancel");
 816                 /* Set this flag to prevent others from getting new references*/
 817                 lock_res_and_lock(lock);
 818                 lock->l_flags |= LDLM_FL_CBPENDING;
 819                 local_only = (lock->l_flags &
 820                               (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
 821                 ldlm_cancel_callback(lock);
 822                 rc = (lock->l_flags & LDLM_FL_BL_AST) ?
 823                         LDLM_FL_BL_AST : LDLM_FL_CANCELING;
 824                 unlock_res_and_lock(lock);
 825
 826                 if (local_only) {
 827                         CDEBUG(D_DLMTRACE, "not sending request (at caller's "
 828                                "instruction)\n");
 829                         rc = LDLM_FL_LOCAL_ONLY;
 830                 }
 831                 ldlm_lock_cancel(lock);
 832         } else {
 833                 if (ns_is_client(lock->l_resource->lr_namespace)) {
 834                         LDLM_ERROR(lock, "Trying to cancel local lock");
 835                         LBUG();
 836                 }
 837                 LDLM_DEBUG(lock, "server-side local cancel");
 838                 ldlm_lock_cancel(lock);
 839                 ldlm_reprocess_all(lock->l_resource);
 840                 LDLM_DEBUG(lock, "server-side local cancel handler END");
 841         }
 842
 843         RETURN(rc);
 844 }
 845
 846 /* Pack @count locks in @head into ldlm_request buffer at the offset @off,
 847    of the request @req. */
 848 static void ldlm_cancel_pack(struct ptlrpc_request *req, int off,
 849                              struct list_head *head, int count)
 850 {
 851         struct ldlm_request *dlm;
 852         struct ldlm_lock *lock;
 853         int max, packed = 0;
 854         ENTRY;
 855
 856         dlm = lustre_msg_buf(req->rq_reqmsg, off, sizeof(*dlm));
 857         LASSERT(dlm != NULL);
 858
 859         /* Check the room in the request buffer. */
 860         max = lustre_msg_buflen(req->rq_reqmsg, off) -
 861                 sizeof(struct ldlm_request);
 862         max /= sizeof(struct lustre_handle);
 863         max += LDLM_LOCKREQ_HANDLES;
 864         LASSERT(max >= dlm->lock_count + count);
 865
 866         /* XXX: it would be better to pack lock handles grouped by resource.
 867          * so that the server cancel would call filter_lvbo_update() less
 868          * frequently. */
 869         list_for_each_entry(lock, head, l_bl_ast) {
 870                 if (!count--)
 871                         break;
 872                 LASSERT(lock->l_conn_export);
 873                 /* Pack the lock handle to the given request buffer. */
 874                 LDLM_DEBUG(lock, "packing");
 875                 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
 876                 packed++;
 877         }
 878         CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
 879         EXIT;
 880 }
 881
 882 /* Prepare and send a batched cancel rpc, it will include count lock handles
 883  * of locks given in @head. */
 884 int ldlm_cli_cancel_req(struct obd_export *exp,
 885                         struct list_head *cancels, int count)
 886 {
 887         struct ptlrpc_request *req = NULL;
 888         struct ldlm_request *body;
 889         int size[2] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
 890                         [DLM_LOCKREQ_OFF]     = sizeof(*body) };
 891         struct obd_import *imp;
 892         int free, sent = 0;
 893         int rc = 0;
 894         ENTRY;
 895
 896         LASSERT(exp != NULL);
 897         LASSERT(count > 0);
 898
 899         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, obd_fail_val);
 900
 901         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
 902                 RETURN(count);
 903
 904         free = ldlm_req_handles_avail(exp, size, 2, 0);
 905         if (count > free)
 906                 count = free;
 907
 908         size[DLM_LOCKREQ_OFF] = ldlm_request_bufsize(count, LDLM_CANCEL);
 909         while (1) {
 910                 imp = class_exp2cliimp(exp);
 911                 if (imp == NULL || imp->imp_invalid) {
 912                         CDEBUG(D_DLMTRACE,
 913                                "skipping cancel on invalid import %p\n", imp);
 914                         RETURN(count);
 915                 }
 916
 917                 req = ptlrpc_prep_req(imp, LUSTRE_DLM_VERSION, LDLM_CANCEL, 2,
 918                                       size, NULL);
 919                 if (!req)
 920                         GOTO(out, rc = -ENOMEM);
 921
 922                 req->rq_no_resend = 1;
 923                 req->rq_no_delay = 1;
 924
 925                 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
 926                 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
 927                 ptlrpc_at_set_req_timeout(req);
 928
 929                 body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF,
 930                                       sizeof(*body));
 931                 ldlm_cancel_pack(req, DLM_LOCKREQ_OFF, cancels, count);
 932
 933                 ptlrpc_req_set_repsize(req, 1, NULL);
 934                 rc = ptlrpc_queue_wait(req);
 935
 936                 if (rc == ESTALE) {
 937                         CDEBUG(D_DLMTRACE, "client/server (nid %s) "
 938                                "out of sync -- not fatal\n",
 939                                libcfs_nid2str(req->rq_import->
 940                                               imp_connection->c_peer.nid));
 941                 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
 942                            req->rq_import_generation == imp->imp_generation) {
 943                         ptlrpc_req_finished(req);
 944                         continue;
 945                 } else if (rc != ELDLM_OK) {
 946                         CERROR("Got rc %d from cancel RPC: canceling "
 947                                "anyway\n", rc);
 948                         break;
 949                 }
 950                 sent = count;
 951                 break;
 952         }
 953
 954         ptlrpc_req_finished(req);
 955         EXIT;
 956 out:
 957         return sent ? sent : rc;
 958 }
 959
 960 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
 961 {
 962         LASSERT(imp != NULL);
 963         return &imp->imp_obd->obd_namespace->ns_pool;
 964 }
 965
 966 int ldlm_cli_update_pool(struct ptlrpc_request *req)
 967 {
 968         struct ldlm_pool *pl;
 969         ENTRY;
 970
 971         if (!imp_connect_lru_resize(req->rq_import))
 972                 RETURN(0);
 973
 974         pl = ldlm_imp2pl(req->rq_import);
 975
 976         spin_lock(&pl->pl_lock);
 977
 978         /* Check if we need to wakeup pools thread for fast SLV change.
 979          * This is only done when threads period is noticably long like
 980          * 10s or more. */
 981 #if defined(__KERNEL__) && (LDLM_POOLS_THREAD_PERIOD >= 10)
 982         {
 983                 __u64 old_slv, new_slv, fast_change;
 984
 985                 old_slv = ldlm_pool_get_slv(pl);
 986                 new_slv = lustre_msg_get_slv(req->rq_repmsg);
 987                 fast_change = old_slv * LDLM_POOLS_FAST_SLV_CHANGE;
 988                 do_div(fast_change, 100);
 989
 990                 /* Wake up pools thread only if SLV has changed more than
 991                  * 50% since last update. In this case we want to react asap.
 992                  * Otherwise it is no sense to wake up pools as they are
 993                  * re-calculated every LDLM_POOLS_THREAD_PERIOD anyways. */
 994                 if (old_slv > new_slv && old_slv - new_slv > fast_change)
 995                         ldlm_pools_wakeup();
 996         }
 997 #endif
 998         /* In some cases RPC may contain slv and limit zeroed out. This is
 999          * the case when server does not support lru resize feature. This is
1000          * also possible in some recovery cases when server side reqs have no
1001          * ref to obd export and thus access to server side namespace is no
1002          * possible. */
1003         if (lustre_msg_get_slv(req->rq_repmsg) != 0 &&
1004             lustre_msg_get_limit(req->rq_repmsg) != 0) {
1005                 ldlm_pool_set_slv(pl, lustre_msg_get_slv(req->rq_repmsg));
1006                 ldlm_pool_set_limit(pl, lustre_msg_get_limit(req->rq_repmsg));
1007         } else {
1008                 DEBUG_REQ(D_HA, req, "zero SLV or Limit found "
1009                           "(SLV: "LPU64", Limit: %u)",
1010                           lustre_msg_get_slv(req->rq_repmsg),
1011                           lustre_msg_get_limit(req->rq_repmsg));
1012         }
1013         spin_unlock(&pl->pl_lock);
1014
1015         RETURN(0);
1016 }
1017 EXPORT_SYMBOL(ldlm_cli_update_pool);
1018
1019 int ldlm_cli_cancel(struct lustre_handle *lockh)
1020 {
1021         struct ldlm_lock *lock;
1022         CFS_LIST_HEAD(cancels);
1023         int rc = 0;
1024         ENTRY;
1025
1026         /* concurrent cancels on the same handle can happen */
1027         lock = __ldlm_handle2lock(lockh, LDLM_FL_CANCELING);
1028         if (lock == NULL) {
1029                 LDLM_DEBUG_NOLOCK("lock is already being destroyed\n");
1030                 RETURN(0);
1031         }
1032
1033         rc = ldlm_cli_cancel_local(lock);
1034
1035         if (rc < 0 || rc == LDLM_FL_LOCAL_ONLY)
1036                 GOTO(out, rc);
1037
1038         list_add(&lock->l_bl_ast, &cancels);
1039         rc = ldlm_cli_cancel_req(lock->l_conn_export, &cancels, 1);
1040         EXIT;
1041  out:
1042         LDLM_LOCK_PUT(lock);
1043         return rc < 0 ? rc : 0;
1044
1045 }
1046
1047 /* XXX until we will have compound requests and can cut cancels from generic rpc
1048  * we need send cancels with LDLM_FL_BL_AST flag as separate rpc */
1049 static int ldlm_cancel_list(struct list_head *cancels, int count)
1050 {
1051         CFS_LIST_HEAD(head);
1052         struct ldlm_lock *lock, *next;
1053         int left = 0, bl_ast = 0, rc;
1054
1055         left = count;
1056         list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1057                 if (left-- == 0)
1058                         break;
1059
1060                 rc = ldlm_cli_cancel_local(lock);
1061                 if (rc == LDLM_FL_BL_AST) {
1062                         LDLM_DEBUG(lock, "Cancel lock separately");
1063                         list_del_init(&lock->l_bl_ast);
1064                         list_add(&lock->l_bl_ast, &head);
1065                         bl_ast ++;
1066                         continue;
1067                 }
1068                 if (rc == LDLM_FL_LOCAL_ONLY) {
1069                         /* CANCEL RPC should not be sent to server. */
1070                         list_del_init(&lock->l_bl_ast);
1071                         LDLM_LOCK_PUT(lock);
1072                         count--;
1073                 }
1074
1075         }
1076         if(bl_ast > 0) {
1077                 count -= bl_ast;
1078                 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1079         }
1080
1081         RETURN(count);
1082 }
1083
1084 /* cancel lock list without sending rpc to server*/
1085 static int ldlm_cancel_list_local(struct list_head *cancels, int count)
1086 {
1087         struct ldlm_lock *lock, *next;
1088         int left = 0;
1089
1090         left = count;
1091         list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1092                 if (left-- == 0)
1093                         break;
1094                 ldlm_lock_cancel(lock);
1095                 /* CANCEL RPC should not be sent to server. */
1096                 list_del_init(&lock->l_bl_ast);
1097                 LDLM_LOCK_PUT(lock);
1098                 count--;
1099         }
1100         RETURN(count);
1101 }
1102
1103 /* Return 1 if @lock should be canceled according to shrinker policy.
1104  * Return zero otherwise. */
1105 static int ldlm_cancel_shrink_policy(struct ldlm_namespace *ns,
1106                                      struct ldlm_lock *lock,
1107                                      int unused, int added,
1108                                      int asked)
1109 {
1110         int lock_cost;
1111         __u64 page_nr;
1112
1113         if (lock->l_resource->lr_type == LDLM_EXTENT) {
1114                 struct ldlm_extent *l_extent;
1115
1116                 /* For all extent locks cost is 1 + number of pages in
1117                  * their extent. */
1118                 l_extent = &lock->l_policy_data.l_extent;
1119                 page_nr = (l_extent->end - l_extent->start);
1120                 do_div(page_nr, CFS_PAGE_SIZE);
1121
1122 #ifdef __KERNEL__
1123                 /* XXX: In fact this is evil hack, we can't access inode
1124                  * here. For doing it right we need somehow to have number
1125                  * of covered by lock. This should be fixed later when 10718
1126                  * is landed. */
1127                 if (lock->l_ast_data != NULL) {
1128                         struct inode *inode = lock->l_ast_data;
1129                         if (page_nr > inode->i_mapping->nrpages)
1130                                 page_nr = inode->i_mapping->nrpages;
1131                 }
1132 #endif
1133                 lock_cost = 1 + page_nr;
1134         } else {
1135                 /* For all locks which are not extent ones cost is 1 */
1136                 lock_cost = 1;
1137         }
1138
1139         /* Keep all expensive locks in lru for the memory pressure time
1140          * cancel policy. They anyways may be canceled by lru resize
1141          * pplicy if they have not small enough CLV. */
1142         return (lock_cost <= ns->ns_shrink_thumb);
1143 }
1144
1145 /* Return 1 if @lock should be canceled according to lru resize policy.
1146  * Return zero otherwise. */
1147 static int ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1148                                    struct ldlm_lock *lock,
1149                                    int unused, int added,
1150                                    int asked)
1151 {
1152         cfs_time_t cur = cfs_time_current();
1153         struct ldlm_pool *pl = &ns->ns_pool;
1154         __u64 slv, lvf, lv;
1155         cfs_time_t la;
1156
1157         spin_lock(&pl->pl_lock);
1158         slv = ldlm_pool_get_slv(pl);
1159         lvf = atomic_read(&pl->pl_lock_volume_factor);
1160         spin_unlock(&pl->pl_lock);
1161
1162         la = cfs_duration_sec(cfs_time_sub(cur,
1163                               lock->l_last_used));
1164
1165         /* Stop when slv is not yet come from server or
1166          * lv is smaller than it is. */
1167         lv = lvf * la * unused;
1168         return (slv > 1 && lv >= slv);
1169 }
1170
1171 /* Return 1 if @lock should be canceled according to passed policy.
1172  * Return zero otherwise. */
1173 static int ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1174                                      struct ldlm_lock *lock,
1175                                      int unused, int added,
1176                                      int asked)
1177 {
1178         /* Do nothing here, we allow canceling all locks which
1179          * are passed here from upper layer logic. So that locks
1180          * number to be canceled will be limited by @count and
1181          * @max in ldlm_cancel_lru_local(). */
1182         return 1;
1183 }
1184
1185 /* Return 1 if @lock should be canceled according to aged policy.
1186  * Return zero otherwise. */
1187 static int ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1188                                    struct ldlm_lock *lock,
1189                                    int unused, int added,
1190                                    int asked)
1191 {
1192         /* Cancel old locks if reached asked limit. */
1193         return !((added >= asked) &&
1194                  cfs_time_before_64(cfs_time_current(),
1195                                     cfs_time_add(lock->l_last_used,
1196                                                  ns->ns_max_age)));
1197 }
1198
1199 typedef int (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *,
1200                                         struct ldlm_lock *, int,
1201                                         int, int);
1202
1203 static ldlm_cancel_lru_policy_t
1204 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
1205 {
1206         if (ns_connect_lru_resize(ns)) {
1207                 if (flags & LDLM_CANCEL_SHRINK)
1208                         return ldlm_cancel_shrink_policy;
1209                 else if (flags & LDLM_CANCEL_LRUR)
1210                         return ldlm_cancel_lrur_policy;
1211                 else if (flags & LDLM_CANCEL_PASSED)
1212                         return ldlm_cancel_passed_policy;
1213         } else {
1214                 if (flags & LDLM_CANCEL_AGED)
1215                         return ldlm_cancel_aged_policy;
1216         }
1217         return NULL;
1218 }
1219
1220 /* - Free space in lru for @count new locks,
1221  *   redundant unused locks are canceled locally;
1222  * - also cancel locally unused aged locks;
1223  * - do not cancel more than @max locks;
1224  * - GET the found locks and add them into the @cancels list.
1225  *
1226  * A client lock can be added to the l_bl_ast list only when it is
1227  * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing CANCEL.
1228  * There are the following use cases: ldlm_cancel_resource_local(),
1229  * ldlm_cancel_lru_local() and ldlm_cli_cancel(), which check&set this
1230  * flag properly. As any attempt to cancel a lock rely on this flag,
1231  * l_bl_ast list is accessed later without any special locking.
1232  *
1233  * Calling policies for enabled lru resize:
1234  * ----------------------------------------
1235  * flags & LDLM_CANCEL_LRUR - use lru resize policy (SLV from server) to
1236  *                            cancel not more than @count locks;
1237  *
1238  * flags & LDLM_CANCEL_PASSED - cancel @count number of old locks (located at
1239  *                              the beginning of lru list);
1240  *
1241  * flags & LDLM_CANCEL_SHRINK - cancel not more than @count locks according to
1242  *                              memory pressre policy function.
1243  */
1244 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
1245                           int count, int max, int flags)
1246 {
1247         ldlm_cancel_lru_policy_t cancel_lru_policy_func;
1248         int added = 0, unused, cancel;
1249         struct ldlm_lock *lock, *next;
1250         ENTRY;
1251
1252         spin_lock(&ns->ns_unused_lock);
1253         unused = ns->ns_nr_unused;
1254
1255         if (!ns_connect_lru_resize(ns))
1256                 count += unused - ns->ns_max_unused;
1257
1258         cancel_lru_policy_func = ldlm_cancel_lru_policy(ns, flags);
1259
1260         list_for_each_entry_safe(lock, next, &ns->ns_unused_list, l_lru) {
1261                 /* Make sure that we skip locks being already in cancel. */
1262                 if ((lock->l_flags & LDLM_FL_CANCELING) ||
1263                     (lock->l_flags & LDLM_FL_BL_AST))
1264                         continue;
1265
1266                 /* For any flags, stop scanning if @max or passed @count is
1267                  * reached. */
1268                 if ((max && added >= max) || (count && added >= count))
1269                         break;
1270
1271                 /* Pass the lock through the policy filter and see if it
1272                  * should stay in lru. */
1273                 if (cancel_lru_policy_func != NULL) {
1274                         cancel = cancel_lru_policy_func(ns, lock, unused,
1275                                                         added, count);
1276
1277                         /* Take next lock for shrink policy, we need to check
1278                          * whole list. Stop scanning for other policies. */
1279                         if ((flags & LDLM_CANCEL_SHRINK) && !cancel)
1280                                 continue;
1281                         else if (!cancel)
1282                                 break;
1283                 }
1284
1285                 if (cancels != NULL) {
1286                         LDLM_LOCK_GET(lock); /* dropped by bl thread */
1287                         spin_unlock(&ns->ns_unused_lock);
1288
1289                         lock_res_and_lock(lock);
1290                         /* Check flags again under the lock. */
1291                         if ((lock->l_flags & LDLM_FL_CANCELING) ||
1292                             (lock->l_flags & LDLM_FL_BL_AST) ||
1293                             (ldlm_lock_remove_from_lru(lock) == 0)) {
1294                                 /* other thread is removing lock from lru or
1295                                  * somebody is already doing CANCEL or
1296                                  * there is a blocking request which will send
1297                                  * cancel by itseft. */
1298                                 unlock_res_and_lock(lock);
1299                                 LDLM_LOCK_PUT(lock);
1300                                 spin_lock(&ns->ns_unused_lock);
1301                                 continue;
1302                         }
1303                         LASSERT(!lock->l_readers && !lock->l_writers);
1304
1305                         /* If we have chosen to cancel this lock voluntarily, we
1306                          * better send cancel notification to server, so that it
1307                          * frees appropriate state. This might lead to a race
1308                          * where while we are doing cancel here, server is also
1309                          * silently cancelling this lock. */
1310                         lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK;
1311
1312                         /* Setting the CBPENDING flag is a little misleading, but
1313                          * prevents an important race; namely, once CBPENDING is
1314                          * set, the lock can accumulate no more readers/writers.
1315                          * Since readers and writers are already zero here,
1316                          * ldlm_lock_decref() won't see this flag and call
1317                          * l_blocking_ast */
1318                         lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1319
1320                         /* We can't re-add to l_lru as it confuses the refcounting
1321                          * in ldlm_lock_remove_from_lru() if an AST arrives after
1322                          * we drop ns_lock below. We use l_bl_ast and can't use
1323                          * l_pending_chain as it is used both on server and client
1324                          * nevertheless bug 5666 says it is used only on server */
1325                         LASSERT(list_empty(&lock->l_bl_ast));
1326                         list_add(&lock->l_bl_ast, cancels);
1327                         unlock_res_and_lock(lock);
1328                         spin_lock(&ns->ns_unused_lock);
1329                 }
1330                 added++;
1331                 unused--;
1332         }
1333         spin_unlock(&ns->ns_unused_lock);
1334
1335         if (cancels == NULL)
1336                 RETURN(added);
1337
1338         RETURN(ldlm_cancel_list(cancels, added));
1339 }
1340
1341 /* when called with LDLM_ASYNC the blocking callback will be handled
1342  * in a thread and this function will return after the thread has been
1343  * asked to call the callback.  when called with LDLM_SYNC the blocking
1344  * callback will be performed in this function. */
1345 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr, ldlm_sync_t sync,
1346                     int flags)
1347 {
1348         CFS_LIST_HEAD(cancels);
1349         int count, rc;
1350         ENTRY;
1351
1352 #ifndef __KERNEL__
1353         sync = LDLM_SYNC; /* force to be sync in user space */
1354 #endif
1355         count = ldlm_cancel_lru_local(ns, &cancels, nr, 0, flags);
1356         if (sync == LDLM_ASYNC) {
1357                 struct ldlm_lock *lock, *next;
1358                 list_for_each_entry_safe(lock, next, &cancels, l_bl_ast) {
1359                         /* Remove from the list to allow blocking thread to
1360                          * re-use l_bl_ast. */
1361                         list_del_init(&lock->l_bl_ast);
1362                         rc = ldlm_bl_to_thread(ns, NULL, lock,
1363                                                LDLM_FL_CANCELING);
1364                         if (rc)
1365                                 list_add_tail(&lock->l_bl_ast, &next->l_bl_ast);
1366                 }
1367         }
1368
1369         /* If some locks are left in the list in ASYNC mode, or
1370          * this is SYNC mode, cancel the list. */
1371         ldlm_cli_cancel_list(&cancels, count, NULL, DLM_LOCKREQ_OFF);
1372         RETURN(count);
1373 }
1374
1375 /* Find and cancel locally unused locks found on resource, matched to the
1376  * given policy, mode. GET the found locks and add them into the @cancels
1377  * list. */
1378 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1379                                struct list_head *cancels,
1380                                ldlm_policy_data_t *policy,
1381                                ldlm_mode_t mode, int lock_flags,
1382                                int flags, void *opaque)
1383 {
1384         struct ldlm_lock *lock;
1385         int count = 0;
1386         ENTRY;
1387
1388         lock_res(res);
1389         list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1390                 if (opaque != NULL && lock->l_ast_data != opaque) {
1391                         LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1392                                    lock->l_ast_data, opaque);
1393                         //LBUG();
1394                         continue;
1395                 }
1396
1397                 if (lock->l_readers || lock->l_writers) {
1398                         if (flags & LDLM_FL_WARN) {
1399                                 LDLM_ERROR(lock, "lock in use");
1400                                 //LBUG();
1401                         }
1402                         continue;
1403                 }
1404
1405                 /* If somebody is already doing CANCEL, or blocking ast came,
1406                  * skip this lock. */
1407                 if (lock->l_flags & LDLM_FL_BL_AST ||
1408                     lock->l_flags & LDLM_FL_CANCELING)
1409                         continue;
1410
1411                 if (lockmode_compat(lock->l_granted_mode, mode))
1412                         continue;
1413
1414                 /* If policy is given and this is IBITS lock, add to list only
1415                  * those locks that match by policy. */
1416                 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1417                     !(lock->l_policy_data.l_inodebits.bits &
1418                       policy->l_inodebits.bits))
1419                         continue;
1420
1421                 /* See CBPENDING comment in ldlm_cancel_lru */
1422                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1423                         lock_flags;
1424
1425                 LASSERT(list_empty(&lock->l_bl_ast));
1426                 list_add(&lock->l_bl_ast, cancels);
1427                 LDLM_LOCK_GET(lock);
1428                 count++;
1429         }
1430         unlock_res(res);
1431
1432         /* Handle only @count inserted locks. */
1433         if ((flags & LDLM_FL_LOCAL_ONLY))
1434                 RETURN(ldlm_cancel_list_local(cancels, count));
1435
1436         RETURN(ldlm_cancel_list(cancels, count));
1437 }
1438
1439 /* If @req is NULL, send CANCEL request to server with handles of locks
1440  * in the @cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1441  * separately per lock.
1442  * If @req is not NULL, put handles of locks in @cancels into the request
1443  * buffer at the offset @off.
1444  * Destroy @cancels at the end. */
1445 int ldlm_cli_cancel_list(struct list_head *cancels, int count,
1446                          struct ptlrpc_request *req, int off)
1447 {
1448         struct ldlm_lock *lock;
1449         int res = 0;
1450         ENTRY;
1451
1452         if (list_empty(cancels) || count == 0)
1453                 RETURN(0);
1454
1455         while (count) {
1456                 LASSERT(!list_empty(cancels));
1457                 lock = list_entry(cancels->next, struct ldlm_lock, l_bl_ast);
1458                 LASSERT(lock->l_conn_export);
1459
1460                 if (exp_connect_cancelset(lock->l_conn_export)) {
1461                         res = count;
1462                         if (req)
1463                                 ldlm_cancel_pack(req, off, cancels, count);
1464                         else
1465                                 res = ldlm_cli_cancel_req(lock->l_conn_export,
1466                                                           cancels, count);
1467                 } else {
1468                         res = ldlm_cli_cancel_req(lock->l_conn_export,
1469                                                   cancels, 1);
1470                 }
1471
1472                 if (res < 0) {
1473                         CERROR("ldlm_cli_cancel_list: %d\n", res);
1474                         res = count;
1475                 }
1476
1477                 count -= res;
1478                 ldlm_lock_list_put(cancels, l_bl_ast, res);
1479         }
1480         LASSERT(list_empty(cancels));
1481         RETURN(0);
1482 }
1483
1484 static int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1485                                            struct ldlm_res_id res_id,
1486                                            int flags, void *opaque)
1487 {
1488         struct ldlm_resource *res;
1489         CFS_LIST_HEAD(cancels);
1490         int count;
1491         int rc;
1492         ENTRY;
1493
1494         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1495         if (res == NULL) {
1496                 /* This is not a problem. */
1497                 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id.name[0]);
1498                 RETURN(0);
1499         }
1500
1501         count = ldlm_cancel_resource_local(res, &cancels, NULL, LCK_MINMODE,
1502                                            0, flags, opaque);
1503         rc = ldlm_cli_cancel_list(&cancels, count, NULL, DLM_LOCKREQ_OFF);
1504         if (rc != ELDLM_OK)
1505                 CERROR("ldlm_cli_cancel_unused_resource: %d\n", rc);
1506
1507         ldlm_resource_putref(res);
1508         RETURN(0);
1509 }
1510
1511 static inline int have_no_nsresource(struct ldlm_namespace *ns)
1512 {
1513         int no_resource = 0;
1514
1515         spin_lock(&ns->ns_hash_lock);
1516         if (ns->ns_resources == 0)
1517                 no_resource = 1;
1518         spin_unlock(&ns->ns_hash_lock);
1519
1520         RETURN(no_resource);
1521 }
1522
1523 /* Cancel all locks on a namespace (or a specific resource, if given)
1524  * that have 0 readers/writers.
1525  *
1526  * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
1527  * to notify the server. */
1528 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1529                            struct ldlm_res_id *res_id, int flags, void *opaque)
1530 {
1531         int i;
1532         ENTRY;
1533
1534         if (ns == NULL)
1535                 RETURN(ELDLM_OK);
1536
1537         if (res_id)
1538                 RETURN(ldlm_cli_cancel_unused_resource(ns, *res_id, flags,
1539                                                        opaque));
1540
1541         spin_lock(&ns->ns_hash_lock);
1542         for (i = 0; i < RES_HASH_SIZE; i++) {
1543                 struct list_head *tmp;
1544                 tmp = ns->ns_hash[i].next;
1545                 while (tmp != &(ns->ns_hash[i])) {
1546                         struct ldlm_resource *res;
1547                         int rc;
1548
1549                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
1550                         ldlm_resource_getref(res);
1551                         spin_unlock(&ns->ns_hash_lock);
1552
1553                         rc = ldlm_cli_cancel_unused_resource(ns, res->lr_name,
1554                                                              flags, opaque);
1555
1556                         if (rc)
1557                                 CERROR("ldlm_cli_cancel_unused ("LPU64"): %d\n",
1558                                        res->lr_name.name[0], rc);
1559
1560                         spin_lock(&ns->ns_hash_lock);
1561                         tmp = tmp->next;
1562                         ldlm_resource_putref_locked(res);
1563                 }
1564         }
1565         spin_unlock(&ns->ns_hash_lock);
1566
1567         RETURN(ELDLM_OK);
1568 }
1569
1570 /* join/split resource locks to/from lru list */
1571 int ldlm_cli_join_lru(struct ldlm_namespace *ns,
1572                       struct ldlm_res_id *res_id, int join)
1573 {
1574         struct ldlm_resource *res;
1575         struct ldlm_lock *lock, *n;
1576         int count = 0;
1577         ENTRY;
1578
1579         LASSERT(ns_is_client(ns));
1580
1581         res = ldlm_resource_get(ns, NULL, *res_id, LDLM_EXTENT, 0);
1582         if (res == NULL)
1583                 RETURN(count);
1584         LASSERT(res->lr_type == LDLM_EXTENT);
1585
1586         lock_res(res);
1587         if (!join)
1588                 goto split;
1589
1590         list_for_each_entry_safe (lock, n, &res->lr_granted, l_res_link) {
1591                 if (list_empty(&lock->l_lru) &&
1592                     !lock->l_readers && !lock->l_writers &&
1593                     !(lock->l_flags & LDLM_FL_LOCAL) &&
1594                     !(lock->l_flags & LDLM_FL_CBPENDING)) {
1595                         ldlm_lock_add_to_lru(lock);
1596                         lock->l_flags &= ~LDLM_FL_NO_LRU;
1597                         LDLM_DEBUG(lock, "join lock to lru");
1598                         count++;
1599                 }
1600         }
1601         goto unlock;
1602 split:
1603         spin_lock(&ns->ns_unused_lock);
1604         list_for_each_entry_safe (lock, n, &ns->ns_unused_list, l_lru) {
1605                 if (lock->l_resource == res) {
1606                         ldlm_lock_remove_from_lru_nolock(lock);
1607                         lock->l_flags |= LDLM_FL_NO_LRU;
1608                         LDLM_DEBUG(lock, "split lock from lru");
1609                         count++;
1610                 }
1611         }
1612         spin_unlock(&ns->ns_unused_lock);
1613 unlock:
1614         unlock_res(res);
1615         ldlm_resource_putref(res);
1616         RETURN(count);
1617 }
1618
1619 /* Lock iterators. */
1620
1621 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
1622                           void *closure)
1623 {
1624         struct list_head *tmp, *next;
1625         struct ldlm_lock *lock;
1626         int rc = LDLM_ITER_CONTINUE;
1627
1628         ENTRY;
1629
1630         if (!res)
1631                 RETURN(LDLM_ITER_CONTINUE);
1632
1633         lock_res(res);
1634         list_for_each_safe(tmp, next, &res->lr_granted) {
1635                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1636
1637                 if (iter(lock, closure) == LDLM_ITER_STOP)
1638                         GOTO(out, rc = LDLM_ITER_STOP);
1639         }
1640
1641         list_for_each_safe(tmp, next, &res->lr_converting) {
1642                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1643
1644                 if (iter(lock, closure) == LDLM_ITER_STOP)
1645                         GOTO(out, rc = LDLM_ITER_STOP);
1646         }
1647
1648         list_for_each_safe(tmp, next, &res->lr_waiting) {
1649                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1650
1651                 if (iter(lock, closure) == LDLM_ITER_STOP)
1652                         GOTO(out, rc = LDLM_ITER_STOP);
1653         }
1654  out:
1655         unlock_res(res);
1656         RETURN(rc);
1657 }
1658
1659 struct iter_helper_data {
1660         ldlm_iterator_t iter;
1661         void *closure;
1662 };
1663
1664 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
1665 {
1666         struct iter_helper_data *helper = closure;
1667         return helper->iter(lock, helper->closure);
1668 }
1669
1670 static int ldlm_res_iter_helper(struct ldlm_resource *res, void *closure)
1671 {
1672         return ldlm_resource_foreach(res, ldlm_iter_helper, closure);
1673 }
1674
1675 int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
1676                            void *closure)
1677 {
1678         struct iter_helper_data helper = { iter: iter, closure: closure };
1679         return ldlm_namespace_foreach_res(ns, ldlm_res_iter_helper, &helper);
1680 }
1681
1682 int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
1683                                ldlm_res_iterator_t iter, void *closure)
1684 {
1685         int i, rc = LDLM_ITER_CONTINUE;
1686         struct ldlm_resource *res;
1687         struct list_head *tmp;
1688
1689         ENTRY;
1690         spin_lock(&ns->ns_hash_lock);
1691         for (i = 0; i < RES_HASH_SIZE; i++) {
1692                 tmp = ns->ns_hash[i].next;
1693                 while (tmp != &(ns->ns_hash[i])) {
1694                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
1695                         ldlm_resource_getref(res);
1696                         spin_unlock(&ns->ns_hash_lock);
1697
1698                         rc = iter(res, closure);
1699
1700                         spin_lock(&ns->ns_hash_lock);
1701                         tmp = tmp->next;
1702                         ldlm_resource_putref_locked(res);
1703                         if (rc == LDLM_ITER_STOP)
1704                                 GOTO(out, rc);
1705                 }
1706         }
1707  out:
1708         spin_unlock(&ns->ns_hash_lock);
1709         RETURN(rc);
1710 }
1711
1712 /* non-blocking function to manipulate a lock whose cb_data is being put away.*/
1713 void ldlm_resource_iterate(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
1714                            ldlm_iterator_t iter, void *data)
1715 {
1716         struct ldlm_resource *res;
1717         ENTRY;
1718
1719         if (ns == NULL) {
1720                 CERROR("must pass in namespace\n");
1721                 LBUG();
1722         }
1723
1724         res = ldlm_resource_get(ns, NULL, *res_id, 0, 0);
1725         if (res == NULL) {
1726                 EXIT;
1727                 return;
1728         }
1729
1730         ldlm_resource_foreach(res, iter, data);
1731         ldlm_resource_putref(res);
1732         EXIT;
1733 }
1734
1735 /* Lock replay */
1736
1737 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
1738 {
1739         struct list_head *list = closure;
1740
1741         /* we use l_pending_chain here, because it's unused on clients. */
1742         LASSERTF(list_empty(&lock->l_pending_chain),"lock %p next %p prev %p\n",
1743                  lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
1744         /* bug 9573: don't replay locks left after eviction */
1745         if (!(lock->l_flags & LDLM_FL_FAILED))
1746                 list_add(&lock->l_pending_chain, list);
1747         return LDLM_ITER_CONTINUE;
1748 }
1749
1750 static int replay_lock_interpret(struct ptlrpc_request *req,
1751                                  struct ldlm_async_args *aa, int rc)
1752 {
1753         struct ldlm_lock *lock;
1754         struct ldlm_reply *reply;
1755
1756         ENTRY;
1757         atomic_dec(&req->rq_import->imp_replay_inflight);
1758         if (rc != ELDLM_OK)
1759                 GOTO(out, rc);
1760
1761
1762         reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF, sizeof(*reply),
1763                                    lustre_swab_ldlm_reply);
1764         if (reply == NULL) {
1765                 CERROR("Can't unpack ldlm_reply\n");
1766                 GOTO (out, rc = -EPROTO);
1767         }
1768
1769         lock = ldlm_handle2lock(&aa->lock_handle);
1770         if (!lock) {
1771                 CERROR("received replay ack for unknown local cookie "LPX64
1772                        " remote cookie "LPX64 " from server %s id %s\n",
1773                        aa->lock_handle.cookie, reply->lock_handle.cookie,
1774                        req->rq_export->exp_client_uuid.uuid,
1775                        libcfs_id2str(req->rq_peer));
1776                 GOTO(out, rc = -ESTALE);
1777         }
1778
1779         lock->l_remote_handle = reply->lock_handle;
1780         LDLM_DEBUG(lock, "replayed lock:");
1781         ptlrpc_import_recovery_state_machine(req->rq_import);
1782         LDLM_LOCK_PUT(lock);
1783 out:
1784         if (rc != ELDLM_OK)
1785                 ptlrpc_connect_import(req->rq_import, NULL);
1786
1787
1788         RETURN(rc);
1789 }
1790
1791 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
1792 {
1793         struct ptlrpc_request *req;
1794         struct ldlm_request *body;
1795         struct ldlm_reply *reply;
1796         struct ldlm_async_args *aa;
1797         int buffers = 2;
1798         int size[3] = { sizeof(struct ptlrpc_body) };
1799         int flags;
1800         ENTRY;
1801
1802
1803         /* Bug 11974: Do not replay a lock which is actively being canceled */
1804         if (lock->l_flags & LDLM_FL_CANCELING) {
1805                 LDLM_DEBUG(lock, "Not replaying canceled lock:");
1806                 RETURN(0);
1807         }
1808
1809         /* If this is reply-less callback lock, we cannot replay it, since
1810          * server might have long dropped it, but notification of that event was
1811          * lost by network. (and server granted conflicting lock already) */
1812         if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1813                 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
1814                 ldlm_lock_cancel(lock);
1815                 RETURN(0);
1816         }
1817         /*
1818          * If granted mode matches the requested mode, this lock is granted.
1819          *
1820          * If they differ, but we have a granted mode, then we were granted
1821          * one mode and now want another: ergo, converting.
1822          *
1823          * If we haven't been granted anything and are on a resource list,
1824          * then we're blocked/waiting.
1825          *
1826          * If we haven't been granted anything and we're NOT on a resource list,
1827          * then we haven't got a reply yet and don't have a known disposition.
1828          * This happens whenever a lock enqueue is the request that triggers
1829          * recovery.
1830          */
1831         if (lock->l_granted_mode == lock->l_req_mode)
1832                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
1833         else if (lock->l_granted_mode)
1834                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
1835         else if (!list_empty(&lock->l_res_link))
1836                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
1837         else
1838                 flags = LDLM_FL_REPLAY;
1839
1840         size[DLM_LOCKREQ_OFF] = sizeof(*body);
1841         req = ptlrpc_prep_req(imp, LUSTRE_DLM_VERSION, LDLM_ENQUEUE, 2, size,
1842                               NULL);
1843         if (!req)
1844                 RETURN(-ENOMEM);
1845
1846         /* We're part of recovery, so don't wait for it. */
1847         req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
1848
1849         body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, sizeof(*body));
1850         ldlm_lock2desc(lock, &body->lock_desc);
1851         body->lock_flags = flags;
1852
1853         ldlm_lock2handle(lock, &body->lock_handle[0]);
1854         size[DLM_LOCKREPLY_OFF] = sizeof(*reply);
1855         if (lock->l_lvb_len != 0) {
1856                 buffers = 3;
1857                 size[DLM_REPLY_REC_OFF] = lock->l_lvb_len;
1858         }
1859         ptlrpc_req_set_repsize(req, buffers, size);
1860
1861         LDLM_DEBUG(lock, "replaying lock:");
1862
1863         atomic_inc(&req->rq_import->imp_replay_inflight);
1864         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1865         aa = (struct ldlm_async_args *)&req->rq_async_args;
1866         aa->lock_handle = body->lock_handle[0];
1867         req->rq_interpret_reply = replay_lock_interpret;
1868         ptlrpcd_add_req(req);
1869
1870         RETURN(0);
1871 }
1872
1873 int ldlm_replay_locks(struct obd_import *imp)
1874 {
1875         struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
1876         struct list_head list;
1877         struct ldlm_lock *lock, *next;
1878         int rc = 0;
1879
1880         ENTRY;
1881         CFS_INIT_LIST_HEAD(&list);
1882
1883         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1884
1885         /* ensure this doesn't fall to 0 before all have been queued */
1886         atomic_inc(&imp->imp_replay_inflight);
1887
1888         (void)ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
1889
1890         list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
1891                 list_del_init(&lock->l_pending_chain);
1892                 if (rc)
1893                         continue; /* or try to do the rest? */
1894                 rc = replay_one_lock(imp, lock);
1895         }
1896
1897         atomic_dec(&imp->imp_replay_inflight);
1898
1899         RETURN(rc);
1900 }