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 static void interrupted_completion_wait(void *data)
  38 {
  39 }
  40
  41 struct lock_wait_data {
  42         struct ldlm_lock *lwd_lock;
  43         __u32             lwd_conn_cnt;
  44 };
  45
  46 struct ldlm_async_args {
  47         struct lustre_handle lock_handle;
  48 };
  49
  50 int ldlm_expired_completion_wait(void *data)
  51 {
  52         struct lock_wait_data *lwd = data;
  53         struct ldlm_lock *lock = lwd->lwd_lock;
  54         struct obd_import *imp;
  55         struct obd_device *obd;
  56
  57         ENTRY;
  58         if (lock->l_conn_export == NULL) {
  59                 static cfs_time_t next_dump = 0, last_dump = 0;
  60
  61                 if (ptlrpc_check_suspend())
  62                         RETURN(0);
  63
  64                 LDLM_ERROR(lock, "lock timed out (enqueued at %lu, %lus ago); "
  65                            "not entering recovery in server code, just going "
  66                            "back to sleep", lock->l_enqueued_time.tv_sec,
  67                            CURRENT_SECONDS - lock->l_enqueued_time.tv_sec);
  68                 if (cfs_time_after(cfs_time_current(), next_dump)) {
  69                         last_dump = next_dump;
  70                         next_dump = cfs_time_shift(300);
  71                         ldlm_namespace_dump(D_DLMTRACE,
  72                                             lock->l_resource->lr_namespace);
  73                         if (last_dump == 0)
  74                                 libcfs_debug_dumplog();
  75                 }
  76                 RETURN(0);
  77         }
  78
  79         obd = lock->l_conn_export->exp_obd;
  80         imp = obd->u.cli.cl_import;
  81         ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
  82         LDLM_ERROR(lock, "lock timed out (enqueued at %lu, %lus ago), entering "
  83                    "recovery for %s@%s", lock->l_enqueued_time.tv_sec,
  84                    CURRENT_SECONDS - lock->l_enqueued_time.tv_sec,
  85                    obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
  86
  87         RETURN(0);
  88 }
  89
  90 static int is_granted_or_cancelled(struct ldlm_lock *lock)
  91 {
  92         int ret = 0;
  93
  94         lock_res_and_lock(lock);
  95         if (((lock->l_req_mode == lock->l_granted_mode) &&
  96              !(lock->l_flags & LDLM_FL_CP_REQD)) ||
  97             (lock->l_flags & LDLM_FL_FAILED))
  98                 ret = 1;
  99         unlock_res_and_lock(lock);
 100
 101         return ret;
 102 }
 103
 104 int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data)
 105 {
 106         /* XXX ALLOCATE - 160 bytes */
 107         struct lock_wait_data lwd;
 108         struct obd_device *obd;
 109         struct obd_import *imp = NULL;
 110         struct l_wait_info lwi;
 111         int rc = 0;
 112         ENTRY;
 113
 114         if (flags == LDLM_FL_WAIT_NOREPROC) {
 115                 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
 116                 goto noreproc;
 117         }
 118
 119         if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
 120                        LDLM_FL_BLOCK_CONV))) {
 121                 cfs_waitq_signal(&lock->l_waitq);
 122                 RETURN(0);
 123         }
 124
 125         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
 126                    "sleeping");
 127         ldlm_lock_dump(D_OTHER, lock, 0);
 128         ldlm_reprocess_all(lock->l_resource);
 129
 130 noreproc:
 131
 132         obd = class_exp2obd(lock->l_conn_export);
 133
 134         /* if this is a local lock, then there is no import */
 135         if (obd != NULL)
 136                 imp = obd->u.cli.cl_import;
 137
 138         lwd.lwd_lock = lock;
 139
 140         if (lock->l_flags & LDLM_FL_NO_TIMEOUT) {
 141                 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
 142                 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
 143         } else {
 144                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
 145                                        ldlm_expired_completion_wait,
 146                                        interrupted_completion_wait, &lwd);
 147         }
 148
 149         if (imp != NULL) {
 150                 spin_lock(&imp->imp_lock);
 151                 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
 152                 spin_unlock(&imp->imp_lock);
 153         }
 154
 155         /* Go to sleep until the lock is granted or cancelled. */
 156         rc = l_wait_event(lock->l_waitq, is_granted_or_cancelled(lock), &lwi);
 157
 158         if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED) {
 159                 LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed");
 160                 RETURN(-EIO);
 161         }
 162
 163         if (rc) {
 164                 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
 165                            rc);
 166                 RETURN(rc);
 167         }
 168
 169         LDLM_DEBUG(lock, "client-side enqueue waking up: granted");
 170         RETURN(0);
 171 }
 172
 173 /*
 174  * ->l_blocking_ast() callback for LDLM locks acquired by server-side OBDs.
 175  */
 176 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 177                       void *data, int flag)
 178 {
 179         int do_ast;
 180         ENTRY;
 181
 182         if (flag == LDLM_CB_CANCELING) {
 183                 /* Don't need to do anything here. */
 184                 RETURN(0);
 185         }
 186
 187         lock_res_and_lock(lock);
 188         /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
 189          * that ldlm_blocking_ast is called just before intent_policy method
 190          * takes the ns_lock, then by the time we get the lock, we might not
 191          * be the correct blocking function anymore.  So check, and return
 192          * early, if so. */
 193         if (lock->l_blocking_ast != ldlm_blocking_ast) {
 194                 unlock_res_and_lock(lock);
 195                 RETURN(0);
 196         }
 197
 198         lock->l_flags |= LDLM_FL_CBPENDING;
 199         do_ast = (!lock->l_readers && !lock->l_writers);
 200         unlock_res_and_lock(lock);
 201
 202         if (do_ast) {
 203                 struct lustre_handle lockh;
 204                 int rc;
 205
 206                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
 207                 ldlm_lock2handle(lock, &lockh);
 208                 rc = ldlm_cli_cancel(&lockh);
 209                 if (rc < 0)
 210                         CERROR("ldlm_cli_cancel: %d\n", rc);
 211         } else {
 212                 LDLM_DEBUG(lock, "Lock still has references, will be "
 213                            "cancelled later");
 214         }
 215         RETURN(0);
 216 }
 217
 218 /*
 219  * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See
 220  * comment in filter_intent_policy() on why you may need this.
 221  */
 222 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
 223 {
 224         /*
 225          * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for
 226          * that is rather subtle: with OST-side locking, it may so happen that
 227          * _all_ extent locks are held by the OST. If client wants to obtain
 228          * current file size it calls ll{,u}_glimpse_size(), and (as locks are
 229          * on the server), dummy glimpse callback fires and does
 230          * nothing. Client still receives correct file size due to the
 231          * following fragment in filter_intent_policy():
 232          *
 233          * rc = l->l_glimpse_ast(l, NULL); // this will update the LVB
 234          * if (rc != 0 && res->lr_namespace->ns_lvbo &&
 235          *     res->lr_namespace->ns_lvbo->lvbo_update) {
 236          *         res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1);
 237          * }
 238          *
 239          * that is, after glimpse_ast() fails, filter_lvbo_update() runs, and
 240          * returns correct file size to the client.
 241          */
 242         return -ELDLM_NO_LOCK_DATA;
 243 }
 244
 245 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
 246                            const struct ldlm_res_id *res_id,
 247                            ldlm_type_t type, ldlm_policy_data_t *policy,
 248                            ldlm_mode_t mode, int *flags,
 249                            ldlm_blocking_callback blocking,
 250                            ldlm_completion_callback completion,
 251                            ldlm_glimpse_callback glimpse,
 252                            void *data, __u32 lvb_len, void *lvb_swabber,
 253                            struct lustre_handle *lockh)
 254 {
 255         struct ldlm_lock *lock;
 256         int err;
 257         ENTRY;
 258
 259         LASSERT(!(*flags & LDLM_FL_REPLAY));
 260         if (unlikely(ns_is_client(ns))) {
 261                 CERROR("Trying to enqueue local lock in a shadow namespace\n");
 262                 LBUG();
 263         }
 264
 265         lock = ldlm_lock_create(ns, res_id, type, mode, blocking,
 266                                 completion, glimpse, data, lvb_len);
 267         if (unlikely(!lock))
 268                 GOTO(out_nolock, err = -ENOMEM);
 269         LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
 270
 271         ldlm_lock_addref_internal(lock, mode);
 272         ldlm_lock2handle(lock, lockh);
 273         lock_res_and_lock(lock);
 274         lock->l_flags |= LDLM_FL_LOCAL;
 275         if (*flags & LDLM_FL_ATOMIC_CB)
 276                 lock->l_flags |= LDLM_FL_ATOMIC_CB;
 277         lock->l_lvb_swabber = lvb_swabber;
 278         unlock_res_and_lock(lock);
 279         if (policy != NULL)
 280                 lock->l_policy_data = *policy;
 281         if (type == LDLM_EXTENT)
 282                 lock->l_req_extent = policy->l_extent;
 283
 284         err = ldlm_lock_enqueue(ns, &lock, policy, flags);
 285         if (unlikely(err != ELDLM_OK))
 286                 GOTO(out, err);
 287
 288         if (policy != NULL)
 289                 *policy = lock->l_policy_data;
 290
 291         LDLM_DEBUG_NOLOCK("client-side local enqueue handler END (lock %p)",
 292                           lock);
 293
 294         if (lock->l_completion_ast)
 295                 lock->l_completion_ast(lock, *flags, NULL);
 296
 297         LDLM_DEBUG(lock, "client-side local enqueue END");
 298         EXIT;
 299  out:
 300         LDLM_LOCK_PUT(lock);
 301  out_nolock:
 302         return err;
 303 }
 304
 305 static void failed_lock_cleanup(struct ldlm_namespace *ns,
 306                                 struct ldlm_lock *lock,
 307                                 struct lustre_handle *lockh, int mode)
 308 {
 309         /* Set a flag to prevent us from sending a CANCEL (bug 407) */
 310         lock_res_and_lock(lock);
 311         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 312         unlock_res_and_lock(lock);
 313         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 314
 315         ldlm_lock_decref_and_cancel(lockh, mode);
 316
 317         /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
 318          *       from llite/file.c/ll_file_flock(). */
 319         if (lock->l_resource->lr_type == LDLM_FLOCK) {
 320                 ldlm_lock_destroy(lock);
 321         }
 322 }
 323
 324 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
 325                           ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
 326                           int *flags, void *lvb, __u32 lvb_len,
 327                           void *lvb_swabber, struct lustre_handle *lockh,int rc)
 328 {
 329         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 330         int is_replay = *flags & LDLM_FL_REPLAY;
 331         struct ldlm_lock *lock;
 332         struct ldlm_reply *reply;
 333         int cleanup_phase = 1;
 334         ENTRY;
 335
 336         lock = ldlm_handle2lock(lockh);
 337         /* ldlm_cli_enqueue is holding a reference on this lock. */
 338         if (!lock) {
 339                 LASSERT(type == LDLM_FLOCK);
 340                 RETURN(-ENOLCK);
 341         }
 342
 343         if (rc != ELDLM_OK) {
 344                 LASSERT(!is_replay);
 345                 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
 346                            rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
 347                 if (rc == ELDLM_LOCK_ABORTED) {
 348                         /* Before we return, swab the reply */
 349                         reply = req_capsule_server_get(&req->rq_pill,
 350                                                        &RMF_DLM_REP);
 351                         if (reply == NULL)
 352                                 rc = -EPROTO;
 353                         if (lvb_len) {
 354                                 struct ost_lvb *tmplvb;
 355
 356                                 req_capsule_set_size(&req->rq_pill,
 357                                                      &RMF_DLM_LVB, RCL_SERVER,
 358                                                      lvb_len);
 359                             tmplvb = req_capsule_server_swab_get(&req->rq_pill,
 360                                                                  &RMF_DLM_LVB,
 361                                                                  lvb_swabber);
 362                                 if (tmplvb == NULL)
 363                                         GOTO(cleanup, rc = -EPROTO);
 364                                 if (lvb != NULL)
 365                                         memcpy(lvb, tmplvb, lvb_len);
 366                         }
 367                 }
 368                 GOTO(cleanup, rc);
 369         }
 370
 371         reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
 372         if (reply == NULL)
 373                 GOTO(cleanup, rc = -EPROTO);
 374
 375         /* lock enqueued on the server */
 376         cleanup_phase = 0;
 377
 378         lock_res_and_lock(lock);
 379         lock->l_remote_handle = reply->lock_handle;
 380         *flags = reply->lock_flags;
 381         lock->l_flags |= reply->lock_flags & LDLM_INHERIT_FLAGS;
 382         /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match()
 383          * to wait with no timeout as well */
 384         lock->l_flags |= reply->lock_flags & LDLM_FL_NO_TIMEOUT;
 385         unlock_res_and_lock(lock);
 386
 387         CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%x\n",
 388                lock, reply->lock_handle.cookie, *flags);
 389
 390         /* If enqueue returned a blocked lock but the completion handler has
 391          * already run, then it fixed up the resource and we don't need to do it
 392          * again. */
 393         if ((*flags) & LDLM_FL_LOCK_CHANGED) {
 394                 int newmode = reply->lock_desc.l_req_mode;
 395                 LASSERT(!is_replay);
 396                 if (newmode && newmode != lock->l_req_mode) {
 397                         LDLM_DEBUG(lock, "server returned different mode %s",
 398                                    ldlm_lockname[newmode]);
 399                         lock->l_req_mode = newmode;
 400                 }
 401
 402                 if (memcmp(reply->lock_desc.l_resource.lr_name.name,
 403                           lock->l_resource->lr_name.name,
 404                           sizeof(struct ldlm_res_id))) {
 405                         CDEBUG(D_INFO, "remote intent success, locking "
 406                                         "(%ld,%ld,%ld) instead of "
 407                                         "(%ld,%ld,%ld)\n",
 408                               (long)reply->lock_desc.l_resource.lr_name.name[0],
 409                               (long)reply->lock_desc.l_resource.lr_name.name[1],
 410                               (long)reply->lock_desc.l_resource.lr_name.name[2],
 411                               (long)lock->l_resource->lr_name.name[0],
 412                               (long)lock->l_resource->lr_name.name[1],
 413                               (long)lock->l_resource->lr_name.name[2]);
 414
 415                         ldlm_lock_change_resource(ns, lock,
 416                                           &reply->lock_desc.l_resource.lr_name);
 417                         if (lock->l_resource == NULL) {
 418                                 LBUG();
 419                                 GOTO(cleanup, rc = -ENOMEM);
 420                         }
 421                         LDLM_DEBUG(lock, "client-side enqueue, new resource");
 422                 }
 423                 if (with_policy)
 424                         if (!(type == LDLM_IBITS && !(exp->exp_connect_flags &
 425                                                     OBD_CONNECT_IBITS)))
 426                                 lock->l_policy_data =
 427                                                  reply->lock_desc.l_policy_data;
 428                 if (type != LDLM_PLAIN)
 429                         LDLM_DEBUG(lock,"client-side enqueue, new policy data");
 430         }
 431
 432         if ((*flags) & LDLM_FL_AST_SENT ||
 433             /* Cancel extent locks as soon as possible on a liblustre client,
 434              * because it cannot handle asynchronous ASTs robustly (see
 435              * bug 7311). */
 436             (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) {
 437                 lock_res_and_lock(lock);
 438                 lock->l_flags |= LDLM_FL_CBPENDING |  LDLM_FL_BL_AST;
 439                 unlock_res_and_lock(lock);
 440                 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
 441         }
 442
 443         /* If the lock has already been granted by a completion AST, don't
 444          * clobber the LVB with an older one. */
 445         if (lvb_len && (lock->l_req_mode != lock->l_granted_mode)) {
 446                 void *tmplvb;
 447
 448                 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
 449                                      lvb_len);
 450                 tmplvb = req_capsule_server_swab_get(&req->rq_pill,
 451                                                      &RMF_DLM_LVB,
 452                                                      lvb_swabber);
 453                 if (tmplvb == NULL)
 454                         GOTO(cleanup, rc = -EPROTO);
 455                 memcpy(lock->l_lvb_data, tmplvb, lvb_len);
 456         }
 457
 458         if (!is_replay) {
 459                 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
 460                 if (lock->l_completion_ast != NULL) {
 461                         int err = lock->l_completion_ast(lock, *flags, NULL);
 462                         if (!rc)
 463                                 rc = err;
 464                         if (rc && type != LDLM_FLOCK) /* bug 9425, bug 10250 */
 465                                 cleanup_phase = 1;
 466                 }
 467         }
 468
 469         if (lvb_len && lvb != NULL) {
 470                 /* Copy the LVB here, and not earlier, because the completion
 471                  * AST (if any) can override what we got in the reply */
 472                 memcpy(lvb, lock->l_lvb_data, lvb_len);
 473         }
 474
 475         LDLM_DEBUG(lock, "client-side enqueue END");
 476         EXIT;
 477 cleanup:
 478         if (cleanup_phase == 1 && rc)
 479                 failed_lock_cleanup(ns, lock, lockh, mode);
 480         /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
 481         LDLM_LOCK_PUT(lock);
 482         LDLM_LOCK_PUT(lock);
 483         return rc;
 484 }
 485
 486 /* PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
 487  * a single page on the send/receive side. XXX: 512 should be changed
 488  * to more adequate value. */
 489 static inline int ldlm_req_handles_avail(int req_size, int off)
 490 {
 491         int avail;
 492
 493         avail = min_t(int, LDLM_MAXREQSIZE, PAGE_SIZE - 512) - req_size;
 494         avail /= sizeof(struct lustre_handle);
 495         avail += LDLM_LOCKREQ_HANDLES - off;
 496
 497         return avail;
 498 }
 499
 500 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
 501                                              enum req_location loc,
 502                                              int off)
 503 {
 504         int size = req_capsule_msg_size(pill, loc);
 505         return ldlm_req_handles_avail(size, off);
 506 }
 507
 508 static inline int ldlm_format_handles_avail(struct obd_import *imp,
 509                                             const struct req_format *fmt,
 510                                             enum req_location loc, int off)
 511 {
 512         int size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
 513         return ldlm_req_handles_avail(size, off);
 514 }
 515
 516 /* Cancel lru locks and pack them into the enqueue request. Pack there the given
 517  * @count locks in @cancels. */
 518 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
 519                       int version, int opc, int canceloff,
 520                       struct list_head *cancels, int count)
 521 {
 522         struct ldlm_namespace   *ns = exp->exp_obd->obd_namespace;
 523         struct req_capsule      *pill = &req->rq_pill;
 524         struct ldlm_request     *dlm = NULL;
 525         int flags, avail, to_free, bufcount, pack = 0;
 526         CFS_LIST_HEAD(head);
 527         int rc;
 528         ENTRY;
 529
 530         if (cancels == NULL)
 531                 cancels = &head;
 532         if (exp_connect_cancelset(exp)) {
 533                 /* Estimate the amount of available space in the request. */
 534                 bufcount = req_capsule_filled_sizes(pill, RCL_CLIENT);
 535                 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
 536
 537                 flags = ns_connect_lru_resize(ns) ?
 538                         LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
 539                 to_free = !ns_connect_lru_resize(ns) &&
 540                           opc == LDLM_ENQUEUE ? 1 : 0;
 541
 542                 /* Cancel lru locks here _only_ if the server supports
 543                  * EARLY_CANCEL. Otherwise we have to send extra CANCEL
 544                  * rpc, what will make us slower. */
 545                 if (avail > count)
 546                         count += ldlm_cancel_lru_local(ns, cancels, to_free,
 547                                                        avail - count, 0, flags);
 548                 if (avail > count)
 549                         pack = count;
 550                 else
 551                         pack = avail;
 552                 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
 553                                      ldlm_request_bufsize(pack, opc));
 554         }
 555
 556         rc = ptlrpc_request_pack(req, version, opc);
 557         if (rc) {
 558                 ldlm_lock_list_put(cancels, l_bl_ast, count);
 559                 RETURN(rc);
 560         }
 561
 562         if (exp_connect_cancelset(exp)) {
 563                 if (canceloff) {
 564                         dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
 565                         LASSERT(dlm);
 566                         /* Skip first lock handler in ldlm_request_pack(),
 567                          * this method will incrment @lock_count according
 568                          * to the lock handle amount actually written to
 569                          * the buffer. */
 570                         dlm->lock_count = canceloff;
 571                 }
 572                 /* Pack into the request @pack lock handles. */
 573                 ldlm_cli_cancel_list(cancels, pack, req, 0);
 574                 /* Prepare and send separate cancel rpc for others. */
 575                 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
 576         } else {
 577                 ldlm_lock_list_put(cancels, l_bl_ast, count);
 578         }
 579         RETURN(0);
 580 }
 581
 582 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
 583                           struct list_head *cancels, int count)
 584 {
 585         return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
 586                                  LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
 587 }
 588
 589 /* If a request has some specific initialisation it is passed in @reqp,
 590  * otherwise it is created in ldlm_cli_enqueue.
 591  *
 592  * Supports sync and async requests, pass @async flag accordingly. If a
 593  * request was created in ldlm_cli_enqueue and it is the async request,
 594  * pass it to the caller in @reqp. */
 595 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
 596                      struct ldlm_enqueue_info *einfo,
 597                      const struct ldlm_res_id *res_id,
 598                      ldlm_policy_data_t *policy, int *flags,
 599                      void *lvb, __u32 lvb_len, void *lvb_swabber,
 600                      struct lustre_handle *lockh, int async)
 601 {
 602         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 603         struct ldlm_lock      *lock;
 604         struct ldlm_request   *body;
 605         int                    is_replay = *flags & LDLM_FL_REPLAY;
 606         int                    req_passed_in = 1;
 607         int                    rc, err;
 608         struct ptlrpc_request *req;
 609         ENTRY;
 610
 611         LASSERT(exp != NULL);
 612
 613         /* If we're replaying this lock, just check some invariants.
 614          * If we're creating a new lock, get everything all setup nice. */
 615         if (is_replay) {
 616                 lock = ldlm_handle2lock(lockh);
 617                 LASSERT(lock != NULL);
 618                 LDLM_DEBUG(lock, "client-side enqueue START");
 619                 LASSERT(exp == lock->l_conn_export);
 620         } else {
 621                 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
 622                                         einfo->ei_mode, einfo->ei_cb_bl,
 623                                         einfo->ei_cb_cp, einfo->ei_cb_gl,
 624                                         einfo->ei_cbdata, lvb_len);
 625                 if (lock == NULL)
 626                         RETURN(-ENOMEM);
 627                 /* for the local lock, add the reference */
 628                 ldlm_lock_addref_internal(lock, einfo->ei_mode);
 629                 ldlm_lock2handle(lock, lockh);
 630                 lock->l_lvb_swabber = lvb_swabber;
 631                 if (policy != NULL) {
 632                         /* INODEBITS_INTEROP: If the server does not support
 633                          * inodebits, we will request a plain lock in the
 634                          * descriptor (ldlm_lock2desc() below) but use an
 635                          * inodebits lock internally with both bits set.
 636                          */
 637                         if (einfo->ei_type == LDLM_IBITS &&
 638                             !(exp->exp_connect_flags & OBD_CONNECT_IBITS))
 639                                 lock->l_policy_data.l_inodebits.bits =
 640                                         MDS_INODELOCK_LOOKUP |
 641                                         MDS_INODELOCK_UPDATE;
 642                         else
 643                                 lock->l_policy_data = *policy;
 644                 }
 645
 646                 if (einfo->ei_type == LDLM_EXTENT)
 647                         lock->l_req_extent = policy->l_extent;
 648                 LDLM_DEBUG(lock, "client-side enqueue START");
 649         }
 650
 651         /* lock not sent to server yet */
 652
 653         if (reqp == NULL || *reqp == NULL) {
 654                 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
 655                                                 &RQF_LDLM_ENQUEUE,
 656                                                 LUSTRE_DLM_VERSION,
 657                                                 LDLM_ENQUEUE);
 658                 if (req == NULL) {
 659                         failed_lock_cleanup(ns, lock, lockh, einfo->ei_mode);
 660                         LDLM_LOCK_PUT(lock);
 661                         RETURN(-ENOMEM);
 662                 }
 663                 req_passed_in = 0;
 664                 if (reqp)
 665                         *reqp = req;
 666         } else {
 667                 int len;
 668
 669                 req = *reqp;
 670                 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
 671                                            RCL_CLIENT);
 672                 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
 673                          DLM_LOCKREQ_OFF, len, sizeof(*body));
 674         }
 675
 676         lock->l_conn_export = exp;
 677         lock->l_export = NULL;
 678         lock->l_blocking_ast = einfo->ei_cb_bl;
 679
 680         /* Dump lock data into the request buffer */
 681         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
 682         ldlm_lock2desc(lock, &body->lock_desc);
 683         body->lock_flags = *flags;
 684         body->lock_handle[0] = *lockh;
 685
 686         /* Continue as normal. */
 687         if (!req_passed_in) {
 688                 if (lvb_len > 0) {
 689                         req_capsule_extend(&req->rq_pill,
 690                                            &RQF_LDLM_ENQUEUE_LVB);
 691                         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
 692                                              RCL_SERVER, lvb_len);
 693                 }
 694                 ptlrpc_request_set_replen(req);
 695         }
 696
 697         /*
 698          * Liblustre client doesn't get extent locks, except for O_APPEND case
 699          * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
 700          * [i_size, OBD_OBJECT_EOF] lock is taken.
 701          */
 702         LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT ||
 703                      policy->l_extent.end == OBD_OBJECT_EOF));
 704
 705         if (async) {
 706                 LASSERT(reqp != NULL);
 707                 RETURN(0);
 708         }
 709
 710         LDLM_DEBUG(lock, "sending request");
 711         rc = ptlrpc_queue_wait(req);
 712         err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
 713                                     einfo->ei_mode, flags, lvb, lvb_len,
 714                                     lvb_swabber, lockh, rc);
 715
 716         /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
 717          * one reference that we took */
 718         if (err == -ENOLCK)
 719                 LDLM_LOCK_PUT(lock);
 720         else
 721                 rc = err;
 722
 723         if (!req_passed_in && req != NULL) {
 724                 ptlrpc_req_finished(req);
 725                 if (reqp)
 726                         *reqp = NULL;
 727         }
 728
 729         RETURN(rc);
 730 }
 731
 732 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
 733                                   __u32 *flags)
 734 {
 735         struct ldlm_resource *res;
 736         int rc;
 737         ENTRY;
 738         if (ns_is_client(lock->l_resource->lr_namespace)) {
 739                 CERROR("Trying to cancel local lock\n");
 740                 LBUG();
 741         }
 742         LDLM_DEBUG(lock, "client-side local convert");
 743
 744         res = ldlm_lock_convert(lock, new_mode, flags);
 745         if (res) {
 746                 ldlm_reprocess_all(res);
 747                 rc = 0;
 748         } else {
 749                 rc = EDEADLOCK;
 750         }
 751         LDLM_DEBUG(lock, "client-side local convert handler END");
 752         LDLM_LOCK_PUT(lock);
 753         RETURN(rc);
 754 }
 755
 756 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
 757  * conversion of locks which are on the waiting or converting queue */
 758 /* Caller of this code is supposed to take care of lock readers/writers
 759    accounting */
 760 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags)
 761 {
 762         struct ldlm_request   *body;
 763         struct ldlm_reply     *reply;
 764         struct ldlm_lock      *lock;
 765         struct ldlm_resource  *res;
 766         struct ptlrpc_request *req;
 767         int                    rc;
 768         ENTRY;
 769
 770         lock = ldlm_handle2lock(lockh);
 771         if (!lock) {
 772                 LBUG();
 773                 RETURN(-EINVAL);
 774         }
 775         *flags = 0;
 776
 777         if (lock->l_conn_export == NULL)
 778                 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
 779
 780         LDLM_DEBUG(lock, "client-side convert");
 781
 782         req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export),
 783                                         &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
 784                                         LDLM_CONVERT);
 785         if (req == NULL) {
 786                 LDLM_LOCK_PUT(lock);
 787                 RETURN(-ENOMEM);
 788         }
 789
 790         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
 791         body->lock_handle[0] = lock->l_remote_handle;
 792
 793         body->lock_desc.l_req_mode = new_mode;
 794         body->lock_flags = *flags;
 795
 796
 797         ptlrpc_request_set_replen(req);
 798         rc = ptlrpc_queue_wait(req);
 799         if (rc != ELDLM_OK)
 800                 GOTO(out, rc);
 801
 802         reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
 803         if (reply == NULL)
 804                 GOTO(out, rc = -EPROTO);
 805
 806         if (req->rq_status)
 807                 GOTO(out, rc = req->rq_status);
 808
 809         res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
 810         if (res != NULL) {
 811                 ldlm_reprocess_all(res);
 812                 /* Go to sleep until the lock is granted. */
 813                 /* FIXME: or cancelled. */
 814                 if (lock->l_completion_ast) {
 815                         rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
 816                                                     NULL);
 817                         if (rc)
 818                                 GOTO(out, rc);
 819                 }
 820         } else {
 821                 rc = EDEADLOCK;
 822         }
 823         EXIT;
 824  out:
 825         LDLM_LOCK_PUT(lock);
 826         ptlrpc_req_finished(req);
 827         return rc;
 828 }
 829
 830 /* Cancel locks locally.
 831  * Returns:
 832  * LDLM_FL_LOCAL_ONLY if tere is no need in a CANCEL rpc to the server;
 833  * LDLM_FL_CANCELING otherwise;
 834  * LDLM_FL_BL_AST if there is a need in a separate CANCEL rpc. */
 835 static int ldlm_cli_cancel_local(struct ldlm_lock *lock)
 836 {
 837         int rc = LDLM_FL_LOCAL_ONLY;
 838         ENTRY;
 839
 840         if (lock->l_conn_export) {
 841                 int local_only;
 842
 843                 LDLM_DEBUG(lock, "client-side cancel");
 844                 /* Set this flag to prevent others from getting new references*/
 845                 lock_res_and_lock(lock);
 846                 lock->l_flags |= LDLM_FL_CBPENDING;
 847                 local_only = (lock->l_flags &
 848                               (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
 849                 ldlm_cancel_callback(lock);
 850                 rc = (lock->l_flags & LDLM_FL_BL_AST) ?
 851                         LDLM_FL_BL_AST : LDLM_FL_CANCELING;
 852                 unlock_res_and_lock(lock);
 853
 854                 if (local_only) {
 855                         CDEBUG(D_DLMTRACE, "not sending request (at caller's "
 856                                "instruction)\n");
 857                         rc = LDLM_FL_LOCAL_ONLY;
 858                 }
 859                 ldlm_lock_cancel(lock);
 860         } else {
 861                 if (ns_is_client(lock->l_resource->lr_namespace)) {
 862                         LDLM_ERROR(lock, "Trying to cancel local lock");
 863                         LBUG();
 864                 }
 865                 LDLM_DEBUG(lock, "server-side local cancel");
 866                 ldlm_lock_cancel(lock);
 867                 ldlm_reprocess_all(lock->l_resource);
 868                 LDLM_DEBUG(lock, "server-side local cancel handler END");
 869         }
 870
 871         RETURN(rc);
 872 }
 873
 874 /* Pack @count locks in @head into ldlm_request buffer at the offset @off,
 875    of the request @req. */
 876 static void ldlm_cancel_pack(struct ptlrpc_request *req,
 877                              struct list_head *head, int count)
 878 {
 879         struct ldlm_request *dlm;
 880         struct ldlm_lock *lock;
 881         int max, packed = 0;
 882         ENTRY;
 883
 884         dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
 885         LASSERT(dlm != NULL);
 886
 887         /* Check the room in the request buffer. */
 888         max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
 889                 sizeof(struct ldlm_request);
 890         max /= sizeof(struct lustre_handle);
 891         max += LDLM_LOCKREQ_HANDLES;
 892         LASSERT(max >= dlm->lock_count + count);
 893
 894         /* XXX: it would be better to pack lock handles grouped by resource.
 895          * so that the server cancel would call filter_lvbo_update() less
 896          * frequently. */
 897         list_for_each_entry(lock, head, l_bl_ast) {
 898                 if (!count--)
 899                         break;
 900                 LASSERT(lock->l_conn_export);
 901                 /* Pack the lock handle to the given request buffer. */
 902                 LDLM_DEBUG(lock, "packing");
 903                 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
 904                 packed++;
 905         }
 906         CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
 907         EXIT;
 908 }
 909
 910 /* Prepare and send a batched cancel rpc, it will include count lock handles
 911  * of locks given in @head. */
 912 int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
 913                         int count, int flags)
 914 {
 915         struct ptlrpc_request *req = NULL;
 916         struct obd_import *imp;
 917         int free, sent = 0;
 918         int rc = 0;
 919         ENTRY;
 920
 921         LASSERT(exp != NULL);
 922         LASSERT(count > 0);
 923
 924         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
 925                 RETURN(count);
 926
 927         free = ldlm_format_handles_avail(class_exp2cliimp(exp),
 928                                          &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
 929         if (count > free)
 930                 count = free;
 931
 932         while (1) {
 933                 int bufcount;
 934
 935                 imp = class_exp2cliimp(exp);
 936                 if (imp == NULL || imp->imp_invalid) {
 937                         CDEBUG(D_DLMTRACE,
 938                                "skipping cancel on invalid import %p\n", imp);
 939                         RETURN(count);
 940                 }
 941
 942                 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
 943                 if (req == NULL)
 944                         GOTO(out, rc = -ENOMEM);
 945
 946                 bufcount = req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
 947                 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
 948                                      ldlm_request_bufsize(count, LDLM_CANCEL));
 949
 950                 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
 951                 if (rc) {
 952                         ptlrpc_request_free(req);
 953                         GOTO(out, rc);
 954                 }
 955                 req->rq_no_resend = 1;
 956                 req->rq_no_delay = 1;
 957
 958                 /* XXX FIXME bug 249 */
 959                 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
 960                 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
 961
 962                 ldlm_cancel_pack(req, cancels, count);
 963
 964                 ptlrpc_request_set_replen(req);
 965                 if (flags & LDLM_FL_ASYNC) {
 966                         ptlrpcd_add_req(req);
 967                         sent = count;
 968                         GOTO(out, 0);
 969                 } else {
 970                         rc = ptlrpc_queue_wait(req);
 971                 }
 972                 if (rc == ESTALE) {
 973                         CDEBUG(D_DLMTRACE, "client/server (nid %s) "
 974                                "out of sync -- not fatal\n",
 975                                libcfs_nid2str(req->rq_import->
 976                                               imp_connection->c_peer.nid));
 977                         rc = 0;
 978                 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
 979                            req->rq_import_generation == imp->imp_generation) {
 980                         ptlrpc_req_finished(req);
 981                         continue;
 982                 } else if (rc != ELDLM_OK) {
 983                         CERROR("Got rc %d from cancel RPC: canceling "
 984                                "anyway\n", rc);
 985                         break;
 986                 }
 987                 sent = count;
 988                 break;
 989         }
 990
 991         ptlrpc_req_finished(req);
 992         EXIT;
 993 out:
 994         return sent ? sent : rc;
 995 }
 996
 997 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
 998 {
 999         LASSERT(imp != NULL);
1000         return &imp->imp_obd->obd_namespace->ns_pool;
1001 }
1002
1003 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1004 {
1005         __u64 old_slv, new_slv;
1006         struct ldlm_pool *pl;
1007         __u32 new_limit;
1008         ENTRY;
1009
1010         if (!imp_connect_lru_resize(req->rq_import))
1011                 RETURN(0);
1012
1013         /* In some cases RPC may contain slv and limit zeroed out. This is
1014          * the case when server does not support lru resize feature. This is
1015          * also possible in some recovery cases when server side reqs have no
1016          * ref to obd export and thus access to server side namespace is no
1017          * possible. */
1018         if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1019             lustre_msg_get_limit(req->rq_repmsg) == 0) {
1020                 DEBUG_REQ(D_HA, req, "zero SLV or Limit found "
1021                           "(SLV: "LPU64", Limit: %u)",
1022                           lustre_msg_get_slv(req->rq_repmsg),
1023                           lustre_msg_get_limit(req->rq_repmsg));
1024                 RETURN(0);
1025         }
1026
1027         new_limit = lustre_msg_get_limit(req->rq_repmsg);
1028         new_slv = lustre_msg_get_slv(req->rq_repmsg);
1029         pl = ldlm_imp2pl(req->rq_import);
1030
1031         spin_lock(&pl->pl_lock);
1032         old_slv = ldlm_pool_get_slv(pl);
1033         ldlm_pool_set_slv(pl, new_slv);
1034         ldlm_pool_set_limit(pl, new_limit);
1035
1036         /* Check if we need to wakeup pools thread for fast SLV change.
1037          * This is only done when threads period is noticably long like
1038          * 10s or more. */
1039 #if defined(__KERNEL__) && (LDLM_POOLS_THREAD_PERIOD >= 10)
1040         {
1041                 __u64 fast_change = old_slv * LDLM_POOLS_FAST_SLV_CHANGE;
1042                 do_div(fast_change, 100);
1043
1044                 /* Wake up pools thread only if SLV has changed more than
1045                  * 50% since last update. In this case we want to react asap.
1046                  * Otherwise it is no sense to wake up pools as they are
1047                  * re-calculated every LDLM_POOLS_THREAD_PERIOD anyways. */
1048                 if (old_slv > new_slv && old_slv - new_slv > fast_change)
1049                         ldlm_pools_wakeup();
1050         }
1051 #endif
1052         spin_unlock(&pl->pl_lock);
1053         RETURN(0);
1054 }
1055 EXPORT_SYMBOL(ldlm_cli_update_pool);
1056
1057 int ldlm_cli_cancel(struct lustre_handle *lockh)
1058 {
1059         struct obd_export *exp;
1060         int avail, flags, count = 1, rc = 0;
1061         struct ldlm_namespace *ns;
1062         struct ldlm_lock *lock;
1063         CFS_LIST_HEAD(cancels);
1064         ENTRY;
1065
1066         /* concurrent cancels on the same handle can happen */
1067         lock = __ldlm_handle2lock(lockh, LDLM_FL_CANCELING);
1068         if (lock == NULL) {
1069                 LDLM_DEBUG_NOLOCK("lock is already being destroyed\n");
1070                 RETURN(0);
1071         }
1072
1073         rc = ldlm_cli_cancel_local(lock);
1074         if (rc < 0 || rc == LDLM_FL_LOCAL_ONLY) {
1075                 LDLM_LOCK_PUT(lock);
1076                 RETURN(rc < 0 ? rc : 0);
1077         }
1078         /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1079          * rpc which goes to canceld portal, so we can cancel other lru locks
1080          * here and send them all as one LDLM_CANCEL rpc. */
1081         LASSERT(list_empty(&lock->l_bl_ast));
1082         list_add(&lock->l_bl_ast, &cancels);
1083
1084         exp = lock->l_conn_export;
1085         if (exp_connect_cancelset(exp)) {
1086                 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1087                                                   &RQF_LDLM_CANCEL,
1088                                                   RCL_CLIENT, 0);
1089                 LASSERT(avail > 0);
1090
1091                 ns = lock->l_resource->lr_namespace;
1092                 flags = ns_connect_lru_resize(ns) ?
1093                         LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
1094                 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1095                                                LDLM_FL_BL_AST, flags);
1096         }
1097         ldlm_cli_cancel_list(&cancels, count, NULL, 0);
1098         RETURN(0);
1099 }
1100
1101 /* XXX until we will have compound requests and can cut cancels from generic rpc
1102  * we need send cancels with LDLM_FL_BL_AST flag as separate rpc */
1103 static int ldlm_cancel_list(struct list_head *cancels, int count, int flags)
1104 {
1105         CFS_LIST_HEAD(head);
1106         struct ldlm_lock *lock, *next;
1107         int left = 0, bl_ast = 0, rc;
1108
1109         left = count;
1110         list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1111                 if (left-- == 0)
1112                         break;
1113
1114                 if (flags & LDLM_FL_LOCAL_ONLY) {
1115                         rc = LDLM_FL_LOCAL_ONLY;
1116                         ldlm_lock_cancel(lock);
1117                 } else {
1118                         rc = ldlm_cli_cancel_local(lock);
1119                 }
1120                 if (!(flags & LDLM_FL_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1121                         LDLM_DEBUG(lock, "Cancel lock separately");
1122                         list_del_init(&lock->l_bl_ast);
1123                         list_add(&lock->l_bl_ast, &head);
1124                         bl_ast ++;
1125                         continue;
1126                 }
1127                 if (rc == LDLM_FL_LOCAL_ONLY) {
1128                         /* CANCEL RPC should not be sent to server. */
1129                         list_del_init(&lock->l_bl_ast);
1130                         LDLM_LOCK_PUT(lock);
1131                         count--;
1132                 }
1133
1134         }
1135         if (bl_ast > 0) {
1136                 count -= bl_ast;
1137                 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1138         }
1139
1140         RETURN(count);
1141 }
1142
1143 /* Return 1 to stop lru processing and keep current lock cached. Return zero
1144  * otherwise. */
1145 static ldlm_policy_res_t ldlm_cancel_shrink_policy(struct ldlm_namespace *ns,
1146                                                    struct ldlm_lock *lock,
1147                                                    int unused, int added,
1148                                                    int count)
1149 {
1150         int lock_cost;
1151         __u64 page_nr;
1152
1153         /* Stop lru processing when we reached passed @count or checked all
1154          * locks in lru. */
1155         if (count && added >= count)
1156                 return LDLM_POLICY_KEEP_LOCK;
1157
1158         if (lock->l_resource->lr_type == LDLM_EXTENT) {
1159                 struct ldlm_extent *l_extent;
1160
1161                 /* For all extent locks cost is 1 + number of pages in
1162                  * their extent. */
1163                 l_extent = &lock->l_policy_data.l_extent;
1164                 page_nr = (l_extent->end - l_extent->start);
1165                 do_div(page_nr, CFS_PAGE_SIZE);
1166
1167 #ifdef __KERNEL__
1168                 /* XXX: In fact this is evil hack, we can't access inode
1169                  * here. For doing it right we need somehow to have number
1170                  * of covered by lock. This should be fixed later when 10718
1171                  * is landed. */
1172                 if (lock->l_ast_data != NULL) {
1173                         struct inode *inode = lock->l_ast_data;
1174                         if (page_nr > inode->i_mapping->nrpages)
1175                                 page_nr = inode->i_mapping->nrpages;
1176                 }
1177 #endif
1178                 lock_cost = 1 + page_nr;
1179         } else {
1180                 /* For all locks which are not extent ones cost is 1 */
1181                 lock_cost = 1;
1182         }
1183
1184         /* Keep all expensive locks in lru for the memory pressure time
1185          * cancel policy. They anyways may be canceled by lru resize
1186          * pplicy if they have not small enough CLV. */
1187         return lock_cost > ns->ns_shrink_thumb ?
1188                 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1189 }
1190
1191 /* Return 1 to stop lru processing and keep current lock cached. Return zero
1192  * otherwise. */
1193 static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1194                                                  struct ldlm_lock *lock,
1195                                                  int unused, int added,
1196                                                  int count)
1197 {
1198         cfs_time_t cur = cfs_time_current();
1199         struct ldlm_pool *pl = &ns->ns_pool;
1200         __u64 slv, lvf, lv;
1201         cfs_time_t la;
1202
1203         /* Stop lru processing when we reached passed @count or checked all
1204          * locks in lru. */
1205         if (count && added >= count)
1206                 return LDLM_POLICY_KEEP_LOCK;
1207
1208         spin_lock(&pl->pl_lock);
1209         slv = ldlm_pool_get_slv(pl);
1210         lvf = atomic_read(&pl->pl_lock_volume_factor);
1211         spin_unlock(&pl->pl_lock);
1212
1213         la = cfs_duration_sec(cfs_time_sub(cur,
1214                               lock->l_last_used));
1215
1216         /* Stop when slv is not yet come from server or
1217          * lv is smaller than it is. */
1218         lv = lvf * la * unused;
1219         return (slv == 1 || lv < slv) ?
1220                 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1221 }
1222
1223 /* Return 1 to stop lru processing and keep current lock cached. Return zero
1224  * otherwise. */
1225 static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1226                                                    struct ldlm_lock *lock,
1227                                                    int unused, int added,
1228                                                    int count)
1229 {
1230         /* Stop lru processing when we reached passed @count or checked all
1231          * locks in lru. */
1232         return (added >= count) ?
1233                 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1234 }
1235
1236 /* Return 1 to stop lru processing and keep current lock cached. Return zero
1237  * otherwise. */
1238 static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1239                                                  struct ldlm_lock *lock,
1240                                                  int unused, int added,
1241                                                  int count)
1242 {
1243         /* Stop lru processing if young lock is found and we reached passed
1244          * @count. */
1245         return ((added >= count) &&
1246                 cfs_time_before(cfs_time_current(),
1247                                 cfs_time_add(lock->l_last_used,
1248                                              ns->ns_max_age))) ?
1249                 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1250 }
1251
1252 /* Return 1 to stop lru processing and keep current lock cached. Return zero
1253  * otherwise. */
1254 static ldlm_policy_res_t ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1255                                                     struct ldlm_lock *lock,
1256                                                     int unused, int added,
1257                                                     int count)
1258 {
1259         /* Stop lru processing when we reached passed @count or checked all
1260          * locks in lru. */
1261         return (added >= count) ?
1262                 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1263 }
1264
1265 typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *,
1266                                                       struct ldlm_lock *, int,
1267                                                       int, int);
1268
1269 static ldlm_cancel_lru_policy_t
1270 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
1271 {
1272         if (ns_connect_lru_resize(ns)) {
1273                 if (flags & LDLM_CANCEL_SHRINK)
1274                         return ldlm_cancel_shrink_policy;
1275                 else if (flags & LDLM_CANCEL_LRUR)
1276                         return ldlm_cancel_lrur_policy;
1277                 else if (flags & LDLM_CANCEL_PASSED)
1278                         return ldlm_cancel_passed_policy;
1279         } else {
1280                 if (flags & LDLM_CANCEL_AGED)
1281                         return ldlm_cancel_aged_policy;
1282         }
1283
1284         return ldlm_cancel_default_policy;
1285 }
1286
1287 /* - Free space in lru for @count new locks,
1288  *   redundant unused locks are canceled locally;
1289  * - also cancel locally unused aged locks;
1290  * - do not cancel more than @max locks;
1291  * - GET the found locks and add them into the @cancels list.
1292  *
1293  * A client lock can be added to the l_bl_ast list only when it is
1294  * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing CANCEL.
1295  * There are the following use cases: ldlm_cancel_resource_local(),
1296  * ldlm_cancel_lru_local() and ldlm_cli_cancel(), which check&set this
1297  * flag properly. As any attempt to cancel a lock rely on this flag,
1298  * l_bl_ast list is accessed later without any special locking.
1299  *
1300  * Calling policies for enabled lru resize:
1301  * ----------------------------------------
1302  * flags & LDLM_CANCEL_LRUR - use lru resize policy (SLV from server) to
1303  *                            cancel not more than @count locks;
1304  *
1305  * flags & LDLM_CANCEL_PASSED - cancel @count number of old locks (located at
1306  *                              the beginning of lru list);
1307  *
1308  * flags & LDLM_CANCEL_SHRINK - cancel not more than @count locks according to
1309  *                              memory pressre policy function;
1310  *
1311  * flags & LDLM_CANCEL_AGED -   cancel alocks according to "aged policy".
1312  */
1313 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
1314                           int count, int max, int cancel_flags, int flags)
1315 {
1316         ldlm_cancel_lru_policy_t pf;
1317         struct ldlm_lock *lock;
1318         int added = 0, unused;
1319         ENTRY;
1320
1321         spin_lock(&ns->ns_unused_lock);
1322         unused = ns->ns_nr_unused;
1323
1324         if (!ns_connect_lru_resize(ns))
1325                 count += unused - ns->ns_max_unused;
1326
1327         pf = ldlm_cancel_lru_policy(ns, flags);
1328         LASSERT(pf != NULL);
1329
1330         while (!list_empty(&ns->ns_unused_list)) {
1331                 /* For any flags, stop scanning if @max is reached. */
1332                 if (max && added >= max)
1333                         break;
1334
1335                 list_for_each_entry(lock, &ns->ns_unused_list, l_lru) {
1336                         /* Somebody is already doing CANCEL or there is a
1337                          * blocking request will send cancel. */
1338                         if (!(lock->l_flags & LDLM_FL_CANCELING) &&
1339                             !(lock->l_flags & LDLM_FL_BL_AST))
1340                                 break;
1341                 }
1342                 if (&lock->l_lru == &ns->ns_unused_list)
1343                         break;
1344
1345                 /* Pass the lock through the policy filter and see if it
1346                  * should stay in lru.
1347                  *
1348                  * Even for shrinker policy we stop scanning if
1349                  * we find a lock that should stay in the cache.
1350                  * We should take into account lock age anyway
1351                  * as new lock even if it is small of weight is
1352                  * valuable resource.
1353                  *
1354                  * That is, for shrinker policy we drop only
1355                  * old locks, but additionally chose them by
1356                  * their weight. Big extent locks will stay in
1357                  * the cache. */
1358                 if (pf(ns, lock, unused, added, count) == LDLM_POLICY_KEEP_LOCK)
1359                         break;
1360
1361                 LDLM_LOCK_GET(lock); /* dropped by bl thread */
1362                 spin_unlock(&ns->ns_unused_lock);
1363
1364                 lock_res_and_lock(lock);
1365                 /* Check flags again under the lock. */
1366                 if ((lock->l_flags & LDLM_FL_CANCELING) ||
1367                     (lock->l_flags & LDLM_FL_BL_AST) ||
1368                     (ldlm_lock_remove_from_lru(lock) == 0)) {
1369                         /* other thread is removing lock from lru or
1370                          * somebody is already doing CANCEL or
1371                          * there is a blocking request which will send
1372                          * cancel by itseft. */
1373                         unlock_res_and_lock(lock);
1374                         LDLM_LOCK_PUT(lock);
1375                         spin_lock(&ns->ns_unused_lock);
1376                         continue;
1377                 }
1378                 LASSERT(!lock->l_readers && !lock->l_writers);
1379
1380                 /* If we have chosen to cancel this lock voluntarily, we
1381                  * better send cancel notification to server, so that it
1382                  * frees appropriate state. This might lead to a race
1383                  * where while we are doing cancel here, server is also
1384                  * silently cancelling this lock. */
1385                 lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK;
1386
1387                 /* Setting the CBPENDING flag is a little misleading,
1388                  * but prevents an important race; namely, once
1389                  * CBPENDING is set, the lock can accumulate no more
1390                  * readers/writers. Since readers and writers are
1391                  * already zero here, ldlm_lock_decref() won't see
1392                  * this flag and call l_blocking_ast */
1393                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1394
1395                 /* We can't re-add to l_lru as it confuses the
1396                  * refcounting in ldlm_lock_remove_from_lru() if an AST
1397                  * arrives after we drop ns_lock below. We use l_bl_ast
1398                  * and can't use l_pending_chain as it is used both on
1399                  * server and client nevertheless bug 5666 says it is
1400                  * used only on server */
1401                 LASSERT(list_empty(&lock->l_bl_ast));
1402                 list_add(&lock->l_bl_ast, cancels);
1403                 unlock_res_and_lock(lock);
1404                 spin_lock(&ns->ns_unused_lock);
1405                 added++;
1406                 unused--;
1407         }
1408         spin_unlock(&ns->ns_unused_lock);
1409         RETURN(ldlm_cancel_list(cancels, added, cancel_flags));
1410 }
1411
1412 /* Returns number of locks which could be canceled next time when
1413  * ldlm_cancel_lru() is called. Used from locks pool shrinker. */
1414 int ldlm_cancel_lru_estimate(struct ldlm_namespace *ns,
1415                              int count, int max, int flags)
1416 {
1417         ldlm_cancel_lru_policy_t pf;
1418         struct ldlm_lock *lock;
1419         int added = 0, unused;
1420         ENTRY;
1421
1422         pf = ldlm_cancel_lru_policy(ns, flags);
1423         LASSERT(pf != NULL);
1424         spin_lock(&ns->ns_unused_lock);
1425         unused = ns->ns_nr_unused;
1426
1427         list_for_each_entry(lock, &ns->ns_unused_list, l_lru) {
1428                 /* For any flags, stop scanning if @max is reached. */
1429                 if (max && added >= max)
1430                         break;
1431
1432                 /* Somebody is already doing CANCEL or there is a
1433                  * blocking request will send cancel. Let's not count
1434                  * this lock. */
1435                 if ((lock->l_flags & LDLM_FL_CANCELING) ||
1436                     (lock->l_flags & LDLM_FL_BL_AST))
1437                         continue;
1438
1439                 /* Pass the lock through the policy filter and see if it
1440                  * should stay in lru. */
1441                 if (pf(ns, lock, unused, added, count) == LDLM_POLICY_KEEP_LOCK)
1442                         break;
1443
1444                 added++;
1445                 unused--;
1446         }
1447         spin_unlock(&ns->ns_unused_lock);
1448         RETURN(added);
1449 }
1450
1451 /* when called with LDLM_ASYNC the blocking callback will be handled
1452  * in a thread and this function will return after the thread has been
1453  * asked to call the callback.  when called with LDLM_SYNC the blocking
1454  * callback will be performed in this function. */
1455 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr, ldlm_sync_t sync,
1456                     int flags)
1457 {
1458         CFS_LIST_HEAD(cancels);
1459         int count, rc;
1460         ENTRY;
1461
1462 #ifndef __KERNEL__
1463         sync = LDLM_SYNC; /* force to be sync in user space */
1464 #endif
1465         count = ldlm_cancel_lru_local(ns, &cancels, nr, 0, 0, flags);
1466         if (sync == LDLM_ASYNC) {
1467                 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count);
1468                 if (rc == 0)
1469                         RETURN(count);
1470         }
1471
1472         /* If an error occured in ASYNC mode, or
1473          * this is SYNC mode, cancel the list. */
1474         ldlm_cli_cancel_list(&cancels, count, NULL, 0);
1475         RETURN(count);
1476 }
1477
1478 /* Find and cancel locally unused locks found on resource, matched to the
1479  * given policy, mode. GET the found locks and add them into the @cancels
1480  * list. */
1481 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1482                                struct list_head *cancels,
1483                                ldlm_policy_data_t *policy,
1484                                ldlm_mode_t mode, int lock_flags,
1485                                int cancel_flags, void *opaque)
1486 {
1487         struct ldlm_lock *lock;
1488         int count = 0;
1489         ENTRY;
1490
1491         lock_res(res);
1492         list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1493                 if (opaque != NULL && lock->l_ast_data != opaque) {
1494                         LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1495                                    lock->l_ast_data, opaque);
1496                         //LBUG();
1497                         continue;
1498                 }
1499
1500                 if (lock->l_readers || lock->l_writers) {
1501                         if (cancel_flags & LDLM_FL_WARN) {
1502                                 LDLM_ERROR(lock, "lock in use");
1503                                 //LBUG();
1504                         }
1505                         continue;
1506                 }
1507
1508                 /* If somebody is already doing CANCEL, or blocking ast came,
1509                  * skip this lock. */
1510                 if (lock->l_flags & LDLM_FL_BL_AST ||
1511                     lock->l_flags & LDLM_FL_CANCELING)
1512                         continue;
1513
1514                 if (lockmode_compat(lock->l_granted_mode, mode))
1515                         continue;
1516
1517                 /* If policy is given and this is IBITS lock, add to list only
1518                  * those locks that match by policy. */
1519                 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1520                     !(lock->l_policy_data.l_inodebits.bits &
1521                       policy->l_inodebits.bits))
1522                         continue;
1523
1524                 /* See CBPENDING comment in ldlm_cancel_lru */
1525                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1526                                  lock_flags;
1527
1528                 LASSERT(list_empty(&lock->l_bl_ast));
1529                 list_add(&lock->l_bl_ast, cancels);
1530                 LDLM_LOCK_GET(lock);
1531                 count++;
1532         }
1533         unlock_res(res);
1534
1535         RETURN(ldlm_cancel_list(cancels, count, cancel_flags));
1536 }
1537
1538 /* If @req is NULL, send CANCEL request to server with handles of locks
1539  * in the @cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1540  * separately per lock.
1541  * If @req is not NULL, put handles of locks in @cancels into the request
1542  * buffer at the offset @off.
1543  * Destroy @cancels at the end. */
1544 int ldlm_cli_cancel_list(struct list_head *cancels, int count,
1545                          struct ptlrpc_request *req, int flags)
1546 {
1547         struct ldlm_lock *lock;
1548         int res = 0;
1549         ENTRY;
1550
1551         if (list_empty(cancels) || count == 0)
1552                 RETURN(0);
1553
1554         /* XXX: requests (both batched and not) could be sent in parallel.
1555          * Usually it is enough to have just 1 RPC, but it is possible that
1556          * there are to many locks to be cancelled in LRU or on a resource.
1557          * It would also speed up the case when the server does not support
1558          * the feature. */
1559         while (count > 0) {
1560                 LASSERT(!list_empty(cancels));
1561                 lock = list_entry(cancels->next, struct ldlm_lock, l_bl_ast);
1562                 LASSERT(lock->l_conn_export);
1563
1564                 if (exp_connect_cancelset(lock->l_conn_export)) {
1565                         res = count;
1566                         if (req)
1567                                 ldlm_cancel_pack(req, cancels, count);
1568                         else
1569                                 res = ldlm_cli_cancel_req(lock->l_conn_export,
1570                                                           cancels, count,
1571                                                           flags);
1572                 } else {
1573                         res = ldlm_cli_cancel_req(lock->l_conn_export,
1574                                                   cancels, 1, flags);
1575                 }
1576
1577                 if (res < 0) {
1578                         CERROR("ldlm_cli_cancel_list: %d\n", res);
1579                         res = count;
1580                 }
1581
1582                 count -= res;
1583                 ldlm_lock_list_put(cancels, l_bl_ast, res);
1584         }
1585         LASSERT(count == 0);
1586         RETURN(0);
1587 }
1588
1589 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1590                                     const struct ldlm_res_id *res_id,
1591                                     ldlm_policy_data_t *policy,
1592                                     ldlm_mode_t mode, int flags, void *opaque)
1593 {
1594         struct ldlm_resource *res;
1595         CFS_LIST_HEAD(cancels);
1596         int count;
1597         int rc;
1598         ENTRY;
1599
1600         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1601         if (res == NULL) {
1602                 /* This is not a problem. */
1603                 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id->name[0]);
1604                 RETURN(0);
1605         }
1606
1607         count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
1608                                            0, flags, opaque);
1609         rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
1610         if (rc != ELDLM_OK)
1611                 CERROR("ldlm_cli_cancel_unused_resource: %d\n", rc);
1612
1613         ldlm_resource_putref(res);
1614         RETURN(0);
1615 }
1616
1617 static inline int have_no_nsresource(struct ldlm_namespace *ns)
1618 {
1619         int no_resource = 0;
1620
1621         spin_lock(&ns->ns_hash_lock);
1622         if (ns->ns_resources == 0)
1623                 no_resource = 1;
1624         spin_unlock(&ns->ns_hash_lock);
1625
1626         RETURN(no_resource);
1627 }
1628
1629 /* Cancel all locks on a namespace (or a specific resource, if given)
1630  * that have 0 readers/writers.
1631  *
1632  * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
1633  * to notify the server. */
1634 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1635                            const struct ldlm_res_id *res_id,
1636                            int flags, void *opaque)
1637 {
1638         int i;
1639         ENTRY;
1640
1641         if (ns == NULL)
1642                 RETURN(ELDLM_OK);
1643
1644         if (res_id)
1645                 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
1646                                                        LCK_MINMODE, flags,
1647                                                        opaque));
1648
1649         spin_lock(&ns->ns_hash_lock);
1650         for (i = 0; i < RES_HASH_SIZE; i++) {
1651                 struct list_head *tmp;
1652                 tmp = ns->ns_hash[i].next;
1653                 while (tmp != &(ns->ns_hash[i])) {
1654                         struct ldlm_resource *res;
1655                         int rc;
1656
1657                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
1658                         ldlm_resource_getref(res);
1659                         spin_unlock(&ns->ns_hash_lock);
1660
1661                         rc = ldlm_cli_cancel_unused_resource(ns, &res->lr_name,
1662                                                              NULL, LCK_MINMODE,
1663                                                              flags, opaque);
1664
1665                         if (rc)
1666                                 CERROR("ldlm_cli_cancel_unused ("LPU64"): %d\n",
1667                                        res->lr_name.name[0], rc);
1668
1669                         spin_lock(&ns->ns_hash_lock);
1670                         tmp = tmp->next;
1671                         ldlm_resource_putref_locked(res);
1672                 }
1673         }
1674         spin_unlock(&ns->ns_hash_lock);
1675
1676         RETURN(ELDLM_OK);
1677 }
1678
1679 /* join/split resource locks to/from lru list */
1680 int ldlm_cli_join_lru(struct ldlm_namespace *ns,
1681                       const struct ldlm_res_id *res_id, int join)
1682 {
1683         struct ldlm_resource *res;
1684         struct ldlm_lock *lock, *n;
1685         int count = 0;
1686         ENTRY;
1687
1688         LASSERT(ns_is_client(ns));
1689
1690         res = ldlm_resource_get(ns, NULL, res_id, LDLM_EXTENT, 0);
1691         if (res == NULL)
1692                 RETURN(count);
1693         LASSERT(res->lr_type == LDLM_EXTENT);
1694
1695         lock_res(res);
1696         if (!join)
1697                 goto split;
1698
1699         list_for_each_entry_safe (lock, n, &res->lr_granted, l_res_link) {
1700                 if (list_empty(&lock->l_lru) &&
1701                     !lock->l_readers && !lock->l_writers &&
1702                     !(lock->l_flags & LDLM_FL_LOCAL) &&
1703                     !(lock->l_flags & LDLM_FL_CBPENDING)) {
1704                         ldlm_lock_add_to_lru(lock);
1705                         lock->l_flags &= ~LDLM_FL_NO_LRU;
1706                         LDLM_DEBUG(lock, "join lock to lru");
1707                         count++;
1708                 }
1709         }
1710         goto unlock;
1711 split:
1712         spin_lock(&ns->ns_unused_lock);
1713         list_for_each_entry_safe (lock, n, &ns->ns_unused_list, l_lru) {
1714                 if (lock->l_resource == res) {
1715                         ldlm_lock_remove_from_lru_nolock(lock);
1716                         lock->l_flags |= LDLM_FL_NO_LRU;
1717                         LDLM_DEBUG(lock, "split lock from lru");
1718                         count++;
1719                 }
1720         }
1721         spin_unlock(&ns->ns_unused_lock);
1722 unlock:
1723         unlock_res(res);
1724         ldlm_resource_putref(res);
1725         RETURN(count);
1726 }
1727
1728 /* Lock iterators. */
1729
1730 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
1731                           void *closure)
1732 {
1733         struct list_head *tmp, *next;
1734         struct ldlm_lock *lock;
1735         int rc = LDLM_ITER_CONTINUE;
1736
1737         ENTRY;
1738
1739         if (!res)
1740                 RETURN(LDLM_ITER_CONTINUE);
1741
1742         lock_res(res);
1743         list_for_each_safe(tmp, next, &res->lr_granted) {
1744                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1745
1746                 if (iter(lock, closure) == LDLM_ITER_STOP)
1747                         GOTO(out, rc = LDLM_ITER_STOP);
1748         }
1749
1750         list_for_each_safe(tmp, next, &res->lr_converting) {
1751                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1752
1753                 if (iter(lock, closure) == LDLM_ITER_STOP)
1754                         GOTO(out, rc = LDLM_ITER_STOP);
1755         }
1756
1757         list_for_each_safe(tmp, next, &res->lr_waiting) {
1758                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1759
1760                 if (iter(lock, closure) == LDLM_ITER_STOP)
1761                         GOTO(out, rc = LDLM_ITER_STOP);
1762         }
1763  out:
1764         unlock_res(res);
1765         RETURN(rc);
1766 }
1767
1768 struct iter_helper_data {
1769         ldlm_iterator_t iter;
1770         void *closure;
1771 };
1772
1773 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
1774 {
1775         struct iter_helper_data *helper = closure;
1776         return helper->iter(lock, helper->closure);
1777 }
1778
1779 static int ldlm_res_iter_helper(struct ldlm_resource *res, void *closure)
1780 {
1781         return ldlm_resource_foreach(res, ldlm_iter_helper, closure);
1782 }
1783
1784 int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
1785                            void *closure)
1786 {
1787         struct iter_helper_data helper = { iter: iter, closure: closure };
1788         return ldlm_namespace_foreach_res(ns, ldlm_res_iter_helper, &helper);
1789 }
1790
1791 int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
1792                                ldlm_res_iterator_t iter, void *closure)
1793 {
1794         int i, rc = LDLM_ITER_CONTINUE;
1795         struct ldlm_resource *res;
1796         struct list_head *tmp;
1797
1798         ENTRY;
1799         spin_lock(&ns->ns_hash_lock);
1800         for (i = 0; i < RES_HASH_SIZE; i++) {
1801                 tmp = ns->ns_hash[i].next;
1802                 while (tmp != &(ns->ns_hash[i])) {
1803                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
1804                         ldlm_resource_getref(res);
1805                         spin_unlock(&ns->ns_hash_lock);
1806
1807                         rc = iter(res, closure);
1808
1809                         spin_lock(&ns->ns_hash_lock);
1810                         tmp = tmp->next;
1811                         ldlm_resource_putref_locked(res);
1812                         if (rc == LDLM_ITER_STOP)
1813                                 GOTO(out, rc);
1814                 }
1815         }
1816  out:
1817         spin_unlock(&ns->ns_hash_lock);
1818         RETURN(rc);
1819 }
1820
1821 /* non-blocking function to manipulate a lock whose cb_data is being put away.*/
1822 void ldlm_resource_iterate(struct ldlm_namespace *ns,
1823                            const struct ldlm_res_id *res_id,
1824                            ldlm_iterator_t iter, void *data)
1825 {
1826         struct ldlm_resource *res;
1827         ENTRY;
1828
1829         if (ns == NULL) {
1830                 CERROR("must pass in namespace\n");
1831                 LBUG();
1832         }
1833
1834         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1835         if (res == NULL) {
1836                 EXIT;
1837                 return;
1838         }
1839
1840         ldlm_resource_foreach(res, iter, data);
1841         ldlm_resource_putref(res);
1842         EXIT;
1843 }
1844
1845 /* Lock replay */
1846
1847 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
1848 {
1849         struct list_head *list = closure;
1850
1851         /* we use l_pending_chain here, because it's unused on clients. */
1852         LASSERTF(list_empty(&lock->l_pending_chain),"lock %p next %p prev %p\n",
1853                  lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
1854         /* bug 9573: don't replay locks left after eviction */
1855         if (!(lock->l_flags & LDLM_FL_FAILED))
1856                 list_add(&lock->l_pending_chain, list);
1857         return LDLM_ITER_CONTINUE;
1858 }
1859
1860 static int replay_lock_interpret(struct ptlrpc_request *req,
1861                                  struct ldlm_async_args *aa, int rc)
1862 {
1863         struct ldlm_lock  *lock;
1864         struct ldlm_reply *reply;
1865
1866         ENTRY;
1867         atomic_dec(&req->rq_import->imp_replay_inflight);
1868         if (rc != ELDLM_OK)
1869                 GOTO(out, rc);
1870
1871
1872         reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1873         if (reply == NULL)
1874                 GOTO(out, rc = -EPROTO);
1875
1876         lock = ldlm_handle2lock(&aa->lock_handle);
1877         if (!lock) {
1878                 CERROR("received replay ack for unknown local cookie "LPX64
1879                        " remote cookie "LPX64 " from server %s id %s\n",
1880                        aa->lock_handle.cookie, reply->lock_handle.cookie,
1881                        req->rq_export->exp_client_uuid.uuid,
1882                        libcfs_id2str(req->rq_peer));
1883                 GOTO(out, rc = -ESTALE);
1884         }
1885
1886         lock->l_remote_handle = reply->lock_handle;
1887         LDLM_DEBUG(lock, "replayed lock:");
1888         ptlrpc_import_recovery_state_machine(req->rq_import);
1889         LDLM_LOCK_PUT(lock);
1890 out:
1891         if (rc != ELDLM_OK)
1892                 ptlrpc_connect_import(req->rq_import, NULL);
1893
1894
1895         RETURN(rc);
1896 }
1897
1898 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
1899 {
1900         struct ptlrpc_request *req;
1901         struct ldlm_async_args *aa;
1902         struct ldlm_request   *body;
1903         int flags;
1904         ENTRY;
1905
1906
1907         /* Bug 11974: Do not replay a lock which is actively being canceled */
1908         if (lock->l_flags & LDLM_FL_CANCELING) {
1909                 LDLM_DEBUG(lock, "Not replaying canceled lock:");
1910                 RETURN(0);
1911         }
1912
1913         /* If this is reply-less callback lock, we cannot replay it, since
1914          * server might have long dropped it, but notification of that event was
1915          * lost by network. (and server granted conflicting lock already) */
1916         if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1917                 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
1918                 ldlm_lock_cancel(lock);
1919                 RETURN(0);
1920         }
1921         /*
1922          * If granted mode matches the requested mode, this lock is granted.
1923          *
1924          * If they differ, but we have a granted mode, then we were granted
1925          * one mode and now want another: ergo, converting.
1926          *
1927          * If we haven't been granted anything and are on a resource list,
1928          * then we're blocked/waiting.
1929          *
1930          * If we haven't been granted anything and we're NOT on a resource list,
1931          * then we haven't got a reply yet and don't have a known disposition.
1932          * This happens whenever a lock enqueue is the request that triggers
1933          * recovery.
1934          */
1935         if (lock->l_granted_mode == lock->l_req_mode)
1936                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
1937         else if (lock->l_granted_mode)
1938                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
1939         else if (!list_empty(&lock->l_res_link))
1940                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
1941         else
1942                 flags = LDLM_FL_REPLAY;
1943
1944         req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
1945                                         LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
1946         if (req == NULL)
1947                 RETURN(-ENOMEM);
1948
1949         /* We're part of recovery, so don't wait for it. */
1950         req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
1951
1952         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1953         ldlm_lock2desc(lock, &body->lock_desc);
1954         body->lock_flags = flags;
1955
1956         ldlm_lock2handle(lock, &body->lock_handle[0]);
1957         if (lock->l_lvb_len != 0) {
1958                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
1959                 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1960                                      lock->l_lvb_len);
1961         }
1962         ptlrpc_request_set_replen(req);
1963         /* notify the server we've replayed all requests.
1964          * also, we mark the request to be put on a dedicated
1965          * queue to be processed after all request replayes.
1966          * bug 6063 */
1967         lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
1968
1969         LDLM_DEBUG(lock, "replaying lock:");
1970
1971         atomic_inc(&req->rq_import->imp_replay_inflight);
1972         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1973         aa = (struct ldlm_async_args *)&req->rq_async_args;
1974         aa->lock_handle = body->lock_handle[0];
1975         req->rq_interpret_reply = replay_lock_interpret;
1976         ptlrpcd_add_req(req);
1977
1978         RETURN(0);
1979 }
1980
1981 int ldlm_replay_locks(struct obd_import *imp)
1982 {
1983         struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
1984         struct list_head list;
1985         struct ldlm_lock *lock, *next;
1986         int rc = 0;
1987
1988         ENTRY;
1989         CFS_INIT_LIST_HEAD(&list);
1990
1991         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1992
1993         /* ensure this doesn't fall to 0 before all have been queued */
1994         atomic_inc(&imp->imp_replay_inflight);
1995
1996         (void)ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
1997
1998         list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
1999                 list_del_init(&lock->l_pending_chain);
2000                 if (rc)
2001                         continue; /* or try to do the rest? */
2002                 rc = replay_one_lock(imp, lock);
2003         }
2004
2005         atomic_dec(&imp->imp_replay_inflight);
2006
2007         RETURN(rc);
2008 }