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 int ldlm_expired_completion_wait(void *data)
  47 {
  48         struct lock_wait_data *lwd = data;
  49         struct ldlm_lock *lock = lwd->lwd_lock;
  50         struct obd_import *imp;
  51         struct obd_device *obd;
  52
  53         ENTRY;
  54         if (lock->l_conn_export == NULL) {
  55                 static cfs_time_t next_dump = 0, last_dump = 0;
  56
  57                 LDLM_ERROR(lock, "lock timed out (enqueued at %lu, %lus ago); "
  58                            "not entering recovery in server code, just going "
  59                            "back to sleep", lock->l_enqueued_time.tv_sec,
  60                            CURRENT_SECONDS - lock->l_enqueued_time.tv_sec);
  61                 if (cfs_time_after(cfs_time_current(), next_dump)) {
  62                         last_dump = next_dump;
  63                         next_dump = cfs_time_shift(300);
  64                         ldlm_namespace_dump(D_DLMTRACE,
  65                                             lock->l_resource->lr_namespace);
  66                         if (last_dump == 0)
  67                                 libcfs_debug_dumplog();
  68                 }
  69                 RETURN(0);
  70         }
  71
  72         obd = lock->l_conn_export->exp_obd;
  73         imp = obd->u.cli.cl_import;
  74         ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
  75         LDLM_ERROR(lock, "lock timed out (enqueued at %lu, %lus ago), entering "
  76                    "recovery for %s@%s", lock->l_enqueued_time.tv_sec,
  77                    CURRENT_SECONDS - lock->l_enqueued_time.tv_sec,
  78                    obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
  79
  80         RETURN(0);
  81 }
  82
  83 int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data)
  84 {
  85         /* XXX ALLOCATE - 160 bytes */
  86         struct lock_wait_data lwd;
  87         struct obd_device *obd;
  88         struct obd_import *imp = NULL;
  89         struct l_wait_info lwi;
  90         int rc = 0;
  91         ENTRY;
  92
  93         if (flags == LDLM_FL_WAIT_NOREPROC) {
  94                 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
  95                 goto noreproc;
  96         }
  97
  98         if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
  99                        LDLM_FL_BLOCK_CONV))) {
 100                 cfs_waitq_signal(&lock->l_waitq);
 101                 RETURN(0);
 102         }
 103
 104         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
 105                    "sleeping");
 106         ldlm_lock_dump(D_OTHER, lock, 0);
 107         ldlm_reprocess_all(lock->l_resource);
 108
 109 noreproc:
 110
 111         obd = class_exp2obd(lock->l_conn_export);
 112
 113         /* if this is a local lock, then there is no import */
 114         if (obd != NULL)
 115                 imp = obd->u.cli.cl_import;
 116
 117         lwd.lwd_lock = lock;
 118
 119         if (lock->l_flags & LDLM_FL_NO_TIMEOUT) {
 120                 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
 121                 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
 122         } else {
 123                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
 124                                        ldlm_expired_completion_wait,
 125                                        interrupted_completion_wait, &lwd);
 126         }
 127
 128         if (imp != NULL) {
 129                 spin_lock(&imp->imp_lock);
 130                 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
 131                 spin_unlock(&imp->imp_lock);
 132         }
 133
 134         /* Go to sleep until the lock is granted or cancelled. */
 135         rc = l_wait_event(lock->l_waitq,
 136                           ((lock->l_req_mode == lock->l_granted_mode) ||
 137                            (lock->l_flags & LDLM_FL_FAILED)), &lwi);
 138
 139         if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED) {
 140                 LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed");
 141                 RETURN(-EIO);
 142         }
 143
 144         if (rc) {
 145                 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
 146                            rc);
 147                 RETURN(rc);
 148         }
 149
 150         LDLM_DEBUG(lock, "client-side enqueue waking up: granted");
 151         RETURN(0);
 152 }
 153
 154 /*
 155  * ->l_blocking_ast() callback for LDLM locks acquired by server-side OBDs.
 156  */
 157 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 158                       void *data, int flag)
 159 {
 160         int do_ast;
 161         ENTRY;
 162
 163         if (flag == LDLM_CB_CANCELING) {
 164                 /* Don't need to do anything here. */
 165                 RETURN(0);
 166         }
 167
 168         lock_res_and_lock(lock);
 169         /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
 170          * that ldlm_blocking_ast is called just before intent_policy method
 171          * takes the ns_lock, then by the time we get the lock, we might not
 172          * be the correct blocking function anymore.  So check, and return
 173          * early, if so. */
 174         if (lock->l_blocking_ast != ldlm_blocking_ast) {
 175                 unlock_res_and_lock(lock);
 176                 RETURN(0);
 177         }
 178
 179         lock->l_flags |= LDLM_FL_CBPENDING;
 180         do_ast = (!lock->l_readers && !lock->l_writers);
 181         unlock_res_and_lock(lock);
 182
 183         if (do_ast) {
 184                 struct lustre_handle lockh;
 185                 int rc;
 186
 187                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
 188                 ldlm_lock2handle(lock, &lockh);
 189                 rc = ldlm_cli_cancel(&lockh);
 190                 if (rc < 0)
 191                         CERROR("ldlm_cli_cancel: %d\n", rc);
 192         } else {
 193                 LDLM_DEBUG(lock, "Lock still has references, will be "
 194                            "cancelled later");
 195         }
 196         RETURN(0);
 197 }
 198
 199 /*
 200  * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See
 201  * comment in filter_intent_policy() on why you may need this.
 202  */
 203 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
 204 {
 205         /*
 206          * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for
 207          * that is rather subtle: with OST-side locking, it may so happen that
 208          * _all_ extent locks are held by the OST. If client wants to obtain
 209          * current file size it calls ll{,u}_glimpse_size(), and (as locks are
 210          * on the server), dummy glimpse callback fires and does
 211          * nothing. Client still receives correct file size due to the
 212          * following fragment in filter_intent_policy():
 213          *
 214          * rc = l->l_glimpse_ast(l, NULL); // this will update the LVB
 215          * if (rc != 0 && res->lr_namespace->ns_lvbo &&
 216          *     res->lr_namespace->ns_lvbo->lvbo_update) {
 217          *         res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1);
 218          * }
 219          *
 220          * that is, after glimpse_ast() fails, filter_lvbo_update() runs, and
 221          * returns correct file size to the client.
 222          */
 223         return -ELDLM_NO_LOCK_DATA;
 224 }
 225
 226 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns, struct ldlm_res_id res_id,
 227                            ldlm_type_t type, ldlm_policy_data_t *policy,
 228                            ldlm_mode_t mode, int *flags,
 229                            ldlm_blocking_callback blocking,
 230                            ldlm_completion_callback completion,
 231                            ldlm_glimpse_callback glimpse,
 232                            void *data, __u32 lvb_len, void *lvb_swabber,
 233                            struct lustre_handle *lockh)
 234 {
 235         struct ldlm_lock *lock;
 236         int err;
 237         ENTRY;
 238
 239         LASSERT(!(*flags & LDLM_FL_REPLAY));
 240         if (ns->ns_client) {
 241                 CERROR("Trying to enqueue local lock in a shadow namespace\n");
 242                 LBUG();
 243         }
 244
 245         lock = ldlm_lock_create(ns, NULL, res_id, type, mode, blocking,
 246                                 completion, glimpse, data, lvb_len);
 247         if (!lock)
 248                 GOTO(out_nolock, err = -ENOMEM);
 249         LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
 250
 251         ldlm_lock_addref_internal(lock, mode);
 252         ldlm_lock2handle(lock, lockh);
 253         lock_res_and_lock(lock);
 254         lock->l_flags |= LDLM_FL_LOCAL;
 255         if (*flags & LDLM_FL_ATOMIC_CB)
 256                 lock->l_flags |= LDLM_FL_ATOMIC_CB;
 257         lock->l_lvb_swabber = lvb_swabber;
 258         unlock_res_and_lock(lock);
 259         if (policy != NULL)
 260                 lock->l_policy_data = *policy;
 261         if (type == LDLM_EXTENT)
 262                 lock->l_req_extent = policy->l_extent;
 263
 264         err = ldlm_lock_enqueue(ns, &lock, policy, flags);
 265         if (err != ELDLM_OK)
 266                 GOTO(out, err);
 267
 268         if (policy != NULL)
 269                 *policy = lock->l_policy_data;
 270         if ((*flags) & LDLM_FL_LOCK_CHANGED)
 271                 res_id = lock->l_resource->lr_name;
 272
 273         LDLM_DEBUG_NOLOCK("client-side local enqueue handler END (lock %p)",
 274                           lock);
 275
 276         if (lock->l_completion_ast)
 277                 lock->l_completion_ast(lock, *flags, NULL);
 278
 279         LDLM_DEBUG(lock, "client-side local enqueue END");
 280         EXIT;
 281  out:
 282         LDLM_LOCK_PUT(lock);
 283  out_nolock:
 284         return err;
 285 }
 286
 287 static void failed_lock_cleanup(struct ldlm_namespace *ns,
 288                                 struct ldlm_lock *lock,
 289                                 struct lustre_handle *lockh, int mode)
 290 {
 291         /* Set a flag to prevent us from sending a CANCEL (bug 407) */
 292         lock_res_and_lock(lock);
 293         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 294         unlock_res_and_lock(lock);
 295         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 296
 297         ldlm_lock_decref_and_cancel(lockh, mode);
 298
 299         /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
 300          *       from llite/file.c/ll_file_flock(). */
 301         if (lock->l_resource->lr_type == LDLM_FLOCK) {
 302                 ldlm_lock_destroy(lock);
 303         }
 304 }
 305
 306 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
 307                           ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
 308                           int *flags, void *lvb, __u32 lvb_len,
 309                           void *lvb_swabber, struct lustre_handle *lockh,int rc)
 310 {
 311         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 312         int is_replay = *flags & LDLM_FL_REPLAY;
 313         struct ldlm_lock *lock;
 314         struct ldlm_reply *reply;
 315         int cleanup_phase = 1;
 316         ENTRY;
 317
 318         lock = ldlm_handle2lock(lockh);
 319         /* ldlm_cli_enqueue is holding a reference on this lock. */
 320         LASSERT(lock != NULL);
 321         if (rc != ELDLM_OK) {
 322                 LASSERT(!is_replay);
 323                 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
 324                            rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
 325                 if (rc == ELDLM_LOCK_ABORTED) {
 326                         /* Before we return, swab the reply */
 327                         reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF,
 328                                                    sizeof(*reply),
 329                                                    lustre_swab_ldlm_reply);
 330                         if (reply == NULL) {
 331                                 CERROR("Can't unpack ldlm_reply\n");
 332                                 rc = -EPROTO;
 333                         }
 334                         if (lvb_len) {
 335                                 void *tmplvb;
 336                                 tmplvb = lustre_swab_repbuf(req,
 337                                                             DLM_REPLY_REC_OFF,
 338                                                             lvb_len,
 339                                                             lvb_swabber);
 340                                 if (tmplvb == NULL)
 341                                         GOTO(cleanup, rc = -EPROTO);
 342                                 if (lvb != NULL)
 343                                         memcpy(lvb, tmplvb, lvb_len);
 344                         }
 345                 }
 346                 GOTO(cleanup, rc);
 347         }
 348
 349         reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF, sizeof(*reply),
 350                                    lustre_swab_ldlm_reply);
 351         if (reply == NULL) {
 352                 CERROR("Can't unpack ldlm_reply\n");
 353                 GOTO(cleanup, rc = -EPROTO);
 354         }
 355
 356         /* lock enqueued on the server */
 357         cleanup_phase = 0;
 358
 359         lock_res_and_lock(lock);
 360         lock->l_remote_handle = reply->lock_handle;
 361         *flags = reply->lock_flags;
 362         lock->l_flags |= reply->lock_flags & LDLM_INHERIT_FLAGS;
 363         /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match()
 364          * to wait with no timeout as well */
 365         lock->l_flags |= reply->lock_flags & LDLM_FL_NO_TIMEOUT;
 366         unlock_res_and_lock(lock);
 367
 368         CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%x\n",
 369                lock, reply->lock_handle.cookie, *flags);
 370
 371         /* If enqueue returned a blocked lock but the completion handler has
 372          * already run, then it fixed up the resource and we don't need to do it
 373          * again. */
 374         if ((*flags) & LDLM_FL_LOCK_CHANGED) {
 375                 int newmode = reply->lock_desc.l_req_mode;
 376                 LASSERT(!is_replay);
 377                 if (newmode && newmode != lock->l_req_mode) {
 378                         LDLM_DEBUG(lock, "server returned different mode %s",
 379                                    ldlm_lockname[newmode]);
 380                         lock->l_req_mode = newmode;
 381                 }
 382
 383                 if (reply->lock_desc.l_resource.lr_name.name[0] !=
 384                     lock->l_resource->lr_name.name[0]) {
 385                         CDEBUG(D_INFO, "remote intent success, locking %ld "
 386                                "instead of %ld\n",
 387                               (long)reply->lock_desc.l_resource.lr_name.name[0],
 388                                (long)lock->l_resource->lr_name.name[0]);
 389
 390                         ldlm_lock_change_resource(ns, lock,
 391                                            reply->lock_desc.l_resource.lr_name);
 392                         if (lock->l_resource == NULL) {
 393                                 LBUG();
 394                                 GOTO(cleanup, rc = -ENOMEM);
 395                         }
 396                         LDLM_DEBUG(lock, "client-side enqueue, new resource");
 397                 }
 398                 if (with_policy)
 399                         if (!(type == LDLM_IBITS && !(exp->exp_connect_flags &
 400                                                     OBD_CONNECT_IBITS)))
 401                                 lock->l_policy_data =
 402                                                  reply->lock_desc.l_policy_data;
 403                 if (type != LDLM_PLAIN)
 404                         LDLM_DEBUG(lock,"client-side enqueue, new policy data");
 405         }
 406
 407         if ((*flags) & LDLM_FL_AST_SENT ||
 408             /* Cancel extent locks as soon as possible on a liblustre client,
 409              * because it cannot handle asynchronous ASTs robustly (see
 410              * bug 7311). */
 411             (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) {
 412                 lock_res_and_lock(lock);
 413                 lock->l_flags |= LDLM_FL_CBPENDING;
 414                 unlock_res_and_lock(lock);
 415                 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
 416         }
 417
 418         /* If the lock has already been granted by a completion AST, don't
 419          * clobber the LVB with an older one. */
 420         if (lvb_len && (lock->l_req_mode != lock->l_granted_mode)) {
 421                 void *tmplvb;
 422                 tmplvb = lustre_swab_repbuf(req, DLM_REPLY_REC_OFF, lvb_len,
 423                                             lvb_swabber);
 424                 if (tmplvb == NULL)
 425                         GOTO(cleanup, rc = -EPROTO);
 426                 memcpy(lock->l_lvb_data, tmplvb, lvb_len);
 427         }
 428
 429         if (!is_replay) {
 430                 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
 431                 if (lock->l_completion_ast != NULL) {
 432                         int err = lock->l_completion_ast(lock, *flags, NULL);
 433                         if (!rc)
 434                                 rc = err;
 435                         if (rc && type != LDLM_FLOCK) /* bug 9425, bug 10250 */
 436                                 cleanup_phase = 1;
 437                 }
 438         }
 439
 440         if (lvb_len && lvb != NULL) {
 441                 /* Copy the LVB here, and not earlier, because the completion
 442                  * AST (if any) can override what we got in the reply */
 443                 memcpy(lvb, lock->l_lvb_data, lvb_len);
 444         }
 445
 446         LDLM_DEBUG(lock, "client-side enqueue END");
 447         EXIT;
 448 cleanup:
 449         if (cleanup_phase == 1 && rc)
 450                 failed_lock_cleanup(ns, lock, lockh, mode);
 451         /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
 452         LDLM_LOCK_PUT(lock);
 453         LDLM_LOCK_PUT(lock);
 454         return rc;
 455 }
 456
 457 /* If a request has some specific initialisation it is passed in @reqp,
 458  * otherwise it is created in ldlm_cli_enqueue.
 459  *
 460  * Supports sync and async requests, pass @async flag accordingly. If a
 461  * request was created in ldlm_cli_enqueue and it is the async request,
 462  * pass it to the caller in @reqp. */
 463 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
 464                      struct ldlm_res_id res_id, ldlm_type_t type,
 465                      ldlm_policy_data_t *policy, ldlm_mode_t mode, int *flags,
 466                      ldlm_blocking_callback blocking,
 467                      ldlm_completion_callback completion,
 468                      ldlm_glimpse_callback glimpse,
 469                      void *data, void *lvb, __u32 lvb_len, void *lvb_swabber,
 470                      struct lustre_handle *lockh, int async)
 471 {
 472         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 473         struct ldlm_lock *lock;
 474         struct ldlm_request *body;
 475         struct ldlm_reply *reply;
 476         int size[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
 477                         [DLM_LOCKREQ_OFF]     = sizeof(*body),
 478                         [DLM_REPLY_REC_OFF]   = lvb_len };
 479         int is_replay = *flags & LDLM_FL_REPLAY;
 480         int req_passed_in = 1, rc;
 481         struct ptlrpc_request *req;
 482         ENTRY;
 483
 484         LASSERT(exp != NULL);
 485
 486         /* If we're replaying this lock, just check some invariants.
 487          * If we're creating a new lock, get everything all setup nice. */
 488         if (is_replay) {
 489                 lock = ldlm_handle2lock(lockh);
 490                 LASSERT(lock != NULL);
 491                 LDLM_DEBUG(lock, "client-side enqueue START");
 492                 LASSERT(exp == lock->l_conn_export);
 493         } else {
 494                 lock = ldlm_lock_create(ns, NULL, res_id, type, mode, blocking,
 495                                         completion, glimpse, data, lvb_len);
 496                 if (lock == NULL)
 497                         RETURN(-ENOMEM);
 498                 /* for the local lock, add the reference */
 499                 ldlm_lock_addref_internal(lock, mode);
 500                 ldlm_lock2handle(lock, lockh);
 501                 lock->l_lvb_swabber = lvb_swabber;
 502                 if (policy != NULL) {
 503                         /* INODEBITS_INTEROP: If the server does not support
 504                          * inodebits, we will request a plain lock in the
 505                          * descriptor (ldlm_lock2desc() below) but use an
 506                          * inodebits lock internally with both bits set.
 507                          */
 508                         if (type == LDLM_IBITS && !(exp->exp_connect_flags &
 509                                                     OBD_CONNECT_IBITS))
 510                                 lock->l_policy_data.l_inodebits.bits =
 511                                         MDS_INODELOCK_LOOKUP |
 512                                         MDS_INODELOCK_UPDATE;
 513                         else
 514                                 lock->l_policy_data = *policy;
 515                 }
 516
 517                 if (type == LDLM_EXTENT)
 518                         lock->l_req_extent = policy->l_extent;
 519                 LDLM_DEBUG(lock, "client-side enqueue START");
 520         }
 521
 522         /* lock not sent to server yet */
 523
 524         if (reqp == NULL || *reqp == NULL) {
 525                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
 526                                       LDLM_ENQUEUE, 2, size, NULL);
 527                 if (req == NULL) {
 528                         failed_lock_cleanup(ns, lock, lockh, mode);
 529                         LDLM_LOCK_PUT(lock);
 530                         RETURN(-ENOMEM);
 531                 }
 532                 req_passed_in = 0;
 533                 if (reqp)
 534                         *reqp = req;
 535         } else {
 536                 req = *reqp;
 537                 LASSERTF(lustre_msg_buflen(req->rq_reqmsg, DLM_LOCKREQ_OFF) ==
 538                          sizeof(*body), "buflen[%d] = %d, not "LPSZ"\n",
 539                          DLM_LOCKREQ_OFF,
 540                          lustre_msg_buflen(req->rq_reqmsg, DLM_LOCKREQ_OFF),
 541                          sizeof(*body));
 542         }
 543
 544         lock->l_conn_export = exp;
 545         lock->l_export = NULL;
 546         lock->l_blocking_ast = blocking;
 547
 548         /* Dump lock data into the request buffer */
 549         body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, sizeof(*body));
 550         ldlm_lock2desc(lock, &body->lock_desc);
 551         body->lock_flags = *flags;
 552         body->lock_handle1 = *lockh;
 553
 554         /* Continue as normal. */
 555         if (!req_passed_in) {
 556                 size[DLM_LOCKREPLY_OFF] = sizeof(*reply);
 557                 ptlrpc_req_set_repsize(req, 2 + (lvb_len > 0), size);
 558         }
 559
 560         /*
 561          * Liblustre client doesn't get extent locks, except for O_APPEND case
 562          * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
 563          * [i_size, OBD_OBJECT_EOF] lock is taken.
 564          */
 565         LASSERT(ergo(LIBLUSTRE_CLIENT, type != LDLM_EXTENT ||
 566                      policy->l_extent.end == OBD_OBJECT_EOF));
 567
 568         if (async) {
 569                 LASSERT(reqp != NULL);
 570                 RETURN(0);
 571         }
 572
 573         LDLM_DEBUG(lock, "sending request");
 574         rc = ptlrpc_queue_wait(req);
 575         rc = ldlm_cli_enqueue_fini(exp, req, type, policy ? 1 : 0,
 576                                    mode, flags, lvb, lvb_len, lvb_swabber,
 577                                    lockh, rc);
 578
 579         if (!req_passed_in && req != NULL) {
 580                 ptlrpc_req_finished(req);
 581                 if (reqp)
 582                         *reqp = NULL;
 583         }
 584
 585         RETURN(rc);
 586 }
 587
 588 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
 589                                   int *flags)
 590 {
 591         struct ldlm_resource *res;
 592         int rc;
 593         ENTRY;
 594         if (lock->l_resource->lr_namespace->ns_client) {
 595                 CERROR("Trying to cancel local lock\n");
 596                 LBUG();
 597         }
 598         LDLM_DEBUG(lock, "client-side local convert");
 599
 600         res = ldlm_lock_convert(lock, new_mode, flags);
 601         if (res) {
 602                 ldlm_reprocess_all(res);
 603                 rc = 0;
 604         } else {
 605                 rc = EDEADLOCK;
 606         }
 607         LDLM_DEBUG(lock, "client-side local convert handler END");
 608         LDLM_LOCK_PUT(lock);
 609         RETURN(rc);
 610 }
 611
 612 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
 613  * conversion of locks which are on the waiting or converting queue */
 614 /* Caller of this code is supposed to take care of lock readers/writers
 615    accounting */
 616 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, int *flags)
 617 {
 618         struct ldlm_request *body;
 619         struct ldlm_reply *reply;
 620         struct ldlm_lock *lock;
 621         struct ldlm_resource *res;
 622         struct ptlrpc_request *req;
 623         int size[2] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
 624                         [DLM_LOCKREQ_OFF]     = sizeof(*body) };
 625         int rc;
 626         ENTRY;
 627
 628         lock = ldlm_handle2lock(lockh);
 629         if (!lock) {
 630                 LBUG();
 631                 RETURN(-EINVAL);
 632         }
 633         *flags = 0;
 634
 635         if (lock->l_conn_export == NULL)
 636                 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
 637
 638         LDLM_DEBUG(lock, "client-side convert");
 639
 640         req = ptlrpc_prep_req(class_exp2cliimp(lock->l_conn_export),
 641                               LUSTRE_DLM_VERSION, LDLM_CONVERT, 2, size, NULL);
 642         if (!req)
 643                 GOTO(out, rc = -ENOMEM);
 644
 645         body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, sizeof(*body));
 646         body->lock_handle1 = lock->l_remote_handle;
 647
 648         body->lock_desc.l_req_mode = new_mode;
 649         body->lock_flags = *flags;
 650
 651         size[DLM_LOCKREPLY_OFF] = sizeof(*reply);
 652         ptlrpc_req_set_repsize(req, 2, size);
 653
 654         rc = ptlrpc_queue_wait(req);
 655         if (rc != ELDLM_OK)
 656                 GOTO(out, rc);
 657
 658         reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF, sizeof(*reply),
 659                                    lustre_swab_ldlm_reply);
 660         if (reply == NULL) {
 661                 CERROR ("Can't unpack ldlm_reply\n");
 662                 GOTO (out, rc = -EPROTO);
 663         }
 664
 665         if (req->rq_status)
 666                 GOTO(out, rc = req->rq_status);
 667
 668         res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
 669         if (res != NULL) {
 670                 ldlm_reprocess_all(res);
 671                 /* Go to sleep until the lock is granted. */
 672                 /* FIXME: or cancelled. */
 673                 if (lock->l_completion_ast) {
 674                         rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
 675                                                     NULL);
 676                         if (rc)
 677                                 GOTO(out, rc);
 678                 }
 679         } else {
 680                 rc = EDEADLOCK;
 681         }
 682         EXIT;
 683  out:
 684         LDLM_LOCK_PUT(lock);
 685         ptlrpc_req_finished(req);
 686         return rc;
 687 }
 688
 689 int ldlm_cli_cancel(struct lustre_handle *lockh)
 690 {
 691         struct ptlrpc_request *req;
 692         struct ldlm_lock *lock;
 693         struct ldlm_request *body;
 694         int size[2] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
 695                         [DLM_LOCKREQ_OFF]     = sizeof(*body) };
 696         int rc = 0;
 697         ENTRY;
 698
 699         /* concurrent cancels on the same handle can happen */
 700         lock = __ldlm_handle2lock(lockh, LDLM_FL_CANCELING);
 701         if (lock == NULL)
 702                 RETURN(0);
 703
 704         if (lock->l_conn_export) {
 705                 int local_only;
 706                 struct obd_import *imp;
 707
 708                 LDLM_DEBUG(lock, "client-side cancel");
 709                 /* Set this flag to prevent others from getting new references*/
 710                 lock_res_and_lock(lock);
 711                 lock->l_flags |= LDLM_FL_CBPENDING;
 712                 local_only = (lock->l_flags &
 713                               (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
 714                 ldlm_cancel_callback(lock);
 715                 unlock_res_and_lock(lock);
 716
 717                 if (local_only) {
 718                         CDEBUG(D_INFO, "not sending request (at caller's "
 719                                "instruction)\n");
 720                         goto local_cancel;
 721                 }
 722
 723         restart:
 724                 imp = class_exp2cliimp(lock->l_conn_export);
 725                 if (imp == NULL || imp->imp_invalid) {
 726                         CDEBUG(D_HA, "skipping cancel on invalid import %p\n",
 727                                imp);
 728                         goto local_cancel;
 729                 }
 730
 731                 req = ptlrpc_prep_req(imp, LUSTRE_DLM_VERSION, LDLM_CANCEL, 2,
 732                                       size, NULL);
 733                 if (!req)
 734                         GOTO(out, rc = -ENOMEM);
 735                 req->rq_no_resend = 1;
 736
 737                 /* XXX FIXME bug 249 */
 738                 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
 739                 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
 740
 741                 body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF,
 742                                       sizeof(*body));
 743                 body->lock_handle1 = lock->l_remote_handle;
 744
 745                 ptlrpc_req_set_repsize(req, 1, NULL);
 746                 rc = ptlrpc_queue_wait(req);
 747
 748                 if (rc == ESTALE) {
 749                         CDEBUG(D_DLMTRACE, "client/server (nid %s) out of sync "
 750                                "-- not fatal, flags %x\n",
 751                                libcfs_nid2str(req->rq_import->
 752                                               imp_connection->c_peer.nid),
 753                                lock->l_flags);
 754                 } else if (rc == -ETIMEDOUT) {
 755                         ptlrpc_req_finished(req);
 756                         GOTO(restart, rc);
 757                 } else if (rc != ELDLM_OK) {
 758                         CERROR("Got rc %d from cancel RPC: canceling "
 759                                "anyway\n", rc);
 760                 }
 761
 762                 ptlrpc_req_finished(req);
 763         local_cancel:
 764                 ldlm_lock_cancel(lock);
 765         } else {
 766                 if (lock->l_resource->lr_namespace->ns_client) {
 767                         LDLM_ERROR(lock, "Trying to cancel local lock");
 768                         LBUG();
 769                 }
 770                 LDLM_DEBUG(lock, "client-side local cancel");
 771                 ldlm_lock_cancel(lock);
 772                 ldlm_reprocess_all(lock->l_resource);
 773                 LDLM_DEBUG(lock, "client-side local cancel handler END");
 774         }
 775
 776         EXIT;
 777  out:
 778         LDLM_LOCK_PUT(lock);
 779         return rc;
 780 }
 781
 782 /* when called with LDLM_ASYNC the blocking callback will be handled
 783  * in a thread and this function will return after the thread has been
 784  * asked to call the callback.  when called with LDLM_SYNC the blocking
 785  * callback will be performed in this function. */
 786 int ldlm_cancel_lru(struct ldlm_namespace *ns, ldlm_sync_t sync)
 787 {
 788         struct ldlm_lock *lock, *next;
 789         int count, rc = 0;
 790         CFS_LIST_HEAD(cblist);
 791         ENTRY;
 792
 793 #ifndef __KERNEL__
 794         sync = LDLM_SYNC; /* force to be sync in user space */
 795 #endif
 796
 797         spin_lock(&ns->ns_unused_lock);
 798         count = ns->ns_nr_unused - ns->ns_max_unused;
 799
 800         if (count <= 0) {
 801                 spin_unlock(&ns->ns_unused_lock);
 802                 RETURN(0);
 803         }
 804
 805         while (!list_empty(&ns->ns_unused_list)) {
 806                 struct list_head *tmp = ns->ns_unused_list.next;
 807                 lock = list_entry(tmp, struct ldlm_lock, l_lru);
 808                 LASSERT(!lock->l_readers && !lock->l_writers);
 809
 810                 LDLM_LOCK_GET(lock); /* dropped by bl thread */
 811                 spin_unlock(&ns->ns_unused_lock);
 812
 813                 lock_res_and_lock(lock);
 814                 if (ldlm_lock_remove_from_lru(lock) == 0) {
 815                         /* other thread is removing lock from lru */
 816                         unlock_res_and_lock(lock);
 817                         LDLM_LOCK_PUT(lock);
 818                         spin_lock(&ns->ns_unused_lock);
 819                         continue;
 820
 821                 }
 822
 823                 /* If we have chosen to canecl this lock voluntarily, we better
 824                    send cancel notification to server, so that it frees
 825                    appropriate state. This might lead to a race where while
 826                    we are doing cancel here, server is also silently
 827                    cancelling this lock. */
 828                 lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK;
 829
 830                 /* Setting the CBPENDING flag is a little misleading, but
 831                  * prevents an important race; namely, once CBPENDING is set,
 832                  * the lock can accumulate no more readers/writers.  Since
 833                  * readers and writers are already zero here, ldlm_lock_decref
 834                  * won't see this flag and call l_blocking_ast */
 835                 lock->l_flags |= LDLM_FL_CBPENDING;
 836
 837                 /* We can't re-add to l_lru as it confuses the refcounting in
 838                  * ldlm_lock_remove_from_lru() if an AST arrives after we drop
 839                  * ns_lock below. We use l_tmp and can't use l_pending_chain as
 840                  * it is used both on server and client nevertheles bug 5666
 841                  * says it is used only on server. --umka */
 842                 list_add(&lock->l_tmp, &cblist);
 843                 unlock_res_and_lock(lock);
 844
 845                 LDLM_LOCK_GET(lock); /* to hold lock after bl thread */
 846                 if (sync == LDLM_ASYNC && (ldlm_bl_to_thread(ns, NULL, lock) == 0)) {
 847                         lock_res_and_lock(lock);
 848                         list_del_init(&lock->l_tmp);
 849                         unlock_res_and_lock(lock);
 850                 }
 851                 LDLM_LOCK_PUT(lock);
 852
 853                 spin_lock(&ns->ns_unused_lock);
 854
 855                 if (--count == 0)
 856                         break;
 857         }
 858         spin_unlock(&ns->ns_unused_lock);
 859
 860         list_for_each_entry_safe(lock, next, &cblist, l_tmp) {
 861                 list_del_init(&lock->l_tmp);
 862                 ldlm_handle_bl_callback(ns, NULL, lock);
 863         }
 864
 865         RETURN(rc);
 866 }
 867
 868 static int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
 869                                            struct ldlm_res_id res_id, int flags,
 870                                            void *opaque)
 871 {
 872         struct list_head *tmp, *next, list = CFS_LIST_HEAD_INIT(list);
 873         struct ldlm_resource *res;
 874         struct ldlm_lock *lock;
 875         ENTRY;
 876
 877         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
 878         if (res == NULL) {
 879                 /* This is not a problem. */
 880                 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id.name[0]);
 881                 RETURN(0);
 882         }
 883
 884         lock_res(res);
 885         list_for_each(tmp, &res->lr_granted) {
 886                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 887
 888                 if (opaque != NULL && lock->l_ast_data != opaque) {
 889                         LDLM_ERROR(lock, "data %p doesn't match opaque %p",
 890                                    lock->l_ast_data, opaque);
 891                         //LBUG();
 892                         continue;
 893                 }
 894
 895                 if (lock->l_readers || lock->l_writers) {
 896                         if (flags & LDLM_FL_WARN) {
 897                                 LDLM_ERROR(lock, "lock in use");
 898                                 //LBUG();
 899                         }
 900                         continue;
 901                 }
 902
 903                 /* See CBPENDING comment in ldlm_cancel_lru */
 904                 lock->l_flags |= LDLM_FL_CBPENDING;
 905
 906                 LASSERT(list_empty(&lock->l_bl_ast));
 907                 list_add(&lock->l_bl_ast, &list);
 908                 LDLM_LOCK_GET(lock);
 909         }
 910         unlock_res(res);
 911
 912         list_for_each_safe(tmp, next, &list) {
 913                 struct lustre_handle lockh;
 914                 int rc;
 915                 lock = list_entry(tmp, struct ldlm_lock, l_bl_ast);
 916
 917                 if (flags & LDLM_FL_LOCAL_ONLY) {
 918                         ldlm_lock_cancel(lock);
 919                 } else {
 920                         ldlm_lock2handle(lock, &lockh);
 921                         rc = ldlm_cli_cancel(&lockh);
 922                         if (rc != ELDLM_OK)
 923                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 924                 }
 925                 list_del_init(&lock->l_bl_ast);
 926                 LDLM_LOCK_PUT(lock);
 927         }
 928
 929         ldlm_resource_putref(res);
 930
 931         RETURN(0);
 932 }
 933
 934 static inline int have_no_nsresource(struct ldlm_namespace *ns)
 935 {
 936         int no_resource = 0;
 937
 938         spin_lock(&ns->ns_hash_lock);
 939         if (ns->ns_resources == 0)
 940                 no_resource = 1;
 941         spin_unlock(&ns->ns_hash_lock);
 942
 943         RETURN(no_resource);
 944 }
 945
 946 /* Cancel all locks on a namespace (or a specific resource, if given)
 947  * that have 0 readers/writers.
 948  *
 949  * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
 950  * to notify the server.
 951  * If flags & LDLM_FL_WARN, print a warning if some locks are still in use. */
 952 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
 953                            struct ldlm_res_id *res_id, int flags, void *opaque)
 954 {
 955         int i;
 956         ENTRY;
 957
 958         if (ns == NULL)
 959                 RETURN(ELDLM_OK);
 960
 961         if (res_id)
 962                 RETURN(ldlm_cli_cancel_unused_resource(ns, *res_id, flags,
 963                                                        opaque));
 964
 965         spin_lock(&ns->ns_hash_lock);
 966         for (i = 0; i < RES_HASH_SIZE; i++) {
 967                 struct list_head *tmp;
 968                 tmp = ns->ns_hash[i].next;
 969                 while (tmp != &(ns->ns_hash[i])) {
 970                         struct ldlm_resource *res;
 971                         int rc;
 972
 973                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
 974                         ldlm_resource_getref(res);
 975                         spin_unlock(&ns->ns_hash_lock);
 976
 977                         rc = ldlm_cli_cancel_unused_resource(ns, res->lr_name,
 978                                                              flags, opaque);
 979
 980                         if (rc)
 981                                 CERROR("cancel_unused_res ("LPU64"): %d\n",
 982                                        res->lr_name.name[0], rc);
 983
 984                         spin_lock(&ns->ns_hash_lock);
 985                         tmp = tmp->next;
 986                         ldlm_resource_putref_locked(res);
 987                 }
 988         }
 989         spin_unlock(&ns->ns_hash_lock);
 990
 991         RETURN(ELDLM_OK);
 992 }
 993
 994 /* join/split resource locks to/from lru list */
 995 int ldlm_cli_join_lru(struct ldlm_namespace *ns,
 996                       struct ldlm_res_id *res_id, int join)
 997 {
 998         struct ldlm_resource *res;
 999         struct ldlm_lock *lock, *n;
1000         int count = 0;
1001         ENTRY;
1002
1003         LASSERT(ns->ns_client == LDLM_NAMESPACE_CLIENT);
1004
1005         res = ldlm_resource_get(ns, NULL, *res_id, LDLM_EXTENT, 0);
1006         if (res == NULL)
1007                 RETURN(count);
1008         LASSERT(res->lr_type == LDLM_EXTENT);
1009
1010         lock_res(res);
1011         if (!join)
1012                 goto split;
1013
1014         list_for_each_entry_safe (lock, n, &res->lr_granted, l_res_link) {
1015                 if (list_empty(&lock->l_lru) &&
1016                     !lock->l_readers && !lock->l_writers &&
1017                     !(lock->l_flags & LDLM_FL_LOCAL) &&
1018                     !(lock->l_flags & LDLM_FL_CBPENDING)) {
1019                         spin_lock(&ns->ns_unused_lock);
1020                         LASSERT(ns->ns_nr_unused >= 0);
1021                         list_add_tail(&lock->l_lru, &ns->ns_unused_list);
1022                         ns->ns_nr_unused++;
1023                         spin_unlock(&ns->ns_unused_lock);
1024                         lock->l_flags &= ~LDLM_FL_NO_LRU;
1025                         LDLM_DEBUG(lock, "join lock to lru");
1026                         count++;
1027                 }
1028         }
1029         goto unlock;
1030 split:
1031         spin_lock(&ns->ns_unused_lock);
1032         list_for_each_entry_safe (lock, n, &ns->ns_unused_list, l_lru) {
1033                 if (lock->l_resource == res) {
1034                         ldlm_lock_remove_from_lru_nolock(lock);
1035                         lock->l_flags |= LDLM_FL_NO_LRU;
1036                         LDLM_DEBUG(lock, "split lock from lru");
1037                         count++;
1038                 }
1039         }
1040         spin_unlock(&ns->ns_unused_lock);
1041 unlock:
1042         unlock_res(res);
1043         ldlm_resource_putref(res);
1044         RETURN(count);
1045 }
1046
1047 /* Lock iterators. */
1048
1049 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
1050                           void *closure)
1051 {
1052         struct list_head *tmp, *next;
1053         struct ldlm_lock *lock;
1054         int rc = LDLM_ITER_CONTINUE;
1055
1056         ENTRY;
1057
1058         if (!res)
1059                 RETURN(LDLM_ITER_CONTINUE);
1060
1061         lock_res(res);
1062         list_for_each_safe(tmp, next, &res->lr_granted) {
1063                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1064
1065                 if (iter(lock, closure) == LDLM_ITER_STOP)
1066                         GOTO(out, rc = LDLM_ITER_STOP);
1067         }
1068
1069         list_for_each_safe(tmp, next, &res->lr_converting) {
1070                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1071
1072                 if (iter(lock, closure) == LDLM_ITER_STOP)
1073                         GOTO(out, rc = LDLM_ITER_STOP);
1074         }
1075
1076         list_for_each_safe(tmp, next, &res->lr_waiting) {
1077                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1078
1079                 if (iter(lock, closure) == LDLM_ITER_STOP)
1080                         GOTO(out, rc = LDLM_ITER_STOP);
1081         }
1082  out:
1083         unlock_res(res);
1084         RETURN(rc);
1085 }
1086
1087 struct iter_helper_data {
1088         ldlm_iterator_t iter;
1089         void *closure;
1090 };
1091
1092 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
1093 {
1094         struct iter_helper_data *helper = closure;
1095         return helper->iter(lock, helper->closure);
1096 }
1097
1098 static int ldlm_res_iter_helper(struct ldlm_resource *res, void *closure)
1099 {
1100         return ldlm_resource_foreach(res, ldlm_iter_helper, closure);
1101 }
1102
1103 int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
1104                            void *closure)
1105 {
1106         struct iter_helper_data helper = { iter: iter, closure: closure };
1107         return ldlm_namespace_foreach_res(ns, ldlm_res_iter_helper, &helper);
1108 }
1109
1110 int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
1111                                ldlm_res_iterator_t iter, void *closure)
1112 {
1113         int i, rc = LDLM_ITER_CONTINUE;
1114         struct ldlm_resource *res;
1115         struct list_head *tmp;
1116
1117         ENTRY;
1118         spin_lock(&ns->ns_hash_lock);
1119         for (i = 0; i < RES_HASH_SIZE; i++) {
1120                 tmp = ns->ns_hash[i].next;
1121                 while (tmp != &(ns->ns_hash[i])) {
1122                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
1123                         ldlm_resource_getref(res);
1124                         spin_unlock(&ns->ns_hash_lock);
1125
1126                         rc = iter(res, closure);
1127
1128                         spin_lock(&ns->ns_hash_lock);
1129                         tmp = tmp->next;
1130                         ldlm_resource_putref_locked(res);
1131                         if (rc == LDLM_ITER_STOP)
1132                                 GOTO(out, rc);
1133                 }
1134         }
1135  out:
1136         spin_unlock(&ns->ns_hash_lock);
1137         RETURN(rc);
1138 }
1139
1140 /* non-blocking function to manipulate a lock whose cb_data is being put away.*/
1141 void ldlm_resource_iterate(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
1142                            ldlm_iterator_t iter, void *data)
1143 {
1144         struct ldlm_resource *res;
1145         ENTRY;
1146
1147         if (ns == NULL) {
1148                 CERROR("must pass in namespace\n");
1149                 LBUG();
1150         }
1151
1152         res = ldlm_resource_get(ns, NULL, *res_id, 0, 0);
1153         if (res == NULL) {
1154                 EXIT;
1155                 return;
1156         }
1157
1158         ldlm_resource_foreach(res, iter, data);
1159         ldlm_resource_putref(res);
1160         EXIT;
1161 }
1162
1163 /* Lock replay */
1164
1165 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
1166 {
1167         struct list_head *list = closure;
1168
1169         /* we use l_pending_chain here, because it's unused on clients. */
1170         LASSERTF(list_empty(&lock->l_pending_chain),"lock %p next %p prev %p\n",
1171                  lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
1172         /* bug 9573: don't replay locks left after eviction */
1173         if (!(lock->l_flags & LDLM_FL_FAILED))
1174                 list_add(&lock->l_pending_chain, list);
1175         return LDLM_ITER_CONTINUE;
1176 }
1177
1178 static int replay_lock_interpret(struct ptlrpc_request *req,
1179                                     void * data, int rc)
1180 {
1181         struct ldlm_lock *lock;
1182         struct ldlm_reply *reply;
1183
1184         ENTRY;
1185         atomic_dec(&req->rq_import->imp_replay_inflight);
1186         if (rc != ELDLM_OK)
1187                 GOTO(out, rc);
1188
1189         lock = req->rq_async_args.pointer_arg[0];
1190         LASSERT(lock != NULL);
1191
1192         reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF, sizeof(*reply),
1193                                    lustre_swab_ldlm_reply);
1194         if (reply == NULL) {
1195                 CERROR("Can't unpack ldlm_reply\n");
1196                 GOTO (out, rc = -EPROTO);
1197         }
1198
1199         lock->l_remote_handle = reply->lock_handle;
1200         LDLM_DEBUG(lock, "replayed lock:");
1201         ptlrpc_import_recovery_state_machine(req->rq_import);
1202  out:
1203         if (rc != ELDLM_OK)
1204                 ptlrpc_connect_import(req->rq_import, NULL);
1205
1206
1207         RETURN(rc);
1208 }
1209
1210 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
1211 {
1212         struct ptlrpc_request *req;
1213         struct ldlm_request *body;
1214         struct ldlm_reply *reply;
1215         int buffers = 2;
1216         int size[3] = { sizeof(struct ptlrpc_body) };
1217         int flags;
1218         ENTRY;
1219
1220         /* If this is reply-less callback lock, we cannot replay it, since
1221          * server might have long dropped it, but notification of that event was
1222          * lost by network. (and server granted conflicting lock already) */
1223         if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1224                 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
1225                 ldlm_lock_cancel(lock);
1226                 RETURN(0);
1227         }
1228         /*
1229          * If granted mode matches the requested mode, this lock is granted.
1230          *
1231          * If they differ, but we have a granted mode, then we were granted
1232          * one mode and now want another: ergo, converting.
1233          *
1234          * If we haven't been granted anything and are on a resource list,
1235          * then we're blocked/waiting.
1236          *
1237          * If we haven't been granted anything and we're NOT on a resource list,
1238          * then we haven't got a reply yet and don't have a known disposition.
1239          * This happens whenever a lock enqueue is the request that triggers
1240          * recovery.
1241          */
1242         if (lock->l_granted_mode == lock->l_req_mode)
1243                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
1244         else if (lock->l_granted_mode)
1245                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
1246         else if (!list_empty(&lock->l_res_link))
1247                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
1248         else
1249                 flags = LDLM_FL_REPLAY;
1250
1251         size[DLM_LOCKREQ_OFF] = sizeof(*body);
1252         req = ptlrpc_prep_req(imp, LUSTRE_DLM_VERSION, LDLM_ENQUEUE, 2, size,
1253                               NULL);
1254         if (!req)
1255                 RETURN(-ENOMEM);
1256
1257         /* We're part of recovery, so don't wait for it. */
1258         req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
1259
1260         body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, sizeof(*body));
1261         ldlm_lock2desc(lock, &body->lock_desc);
1262         body->lock_flags = flags;
1263
1264         ldlm_lock2handle(lock, &body->lock_handle1);
1265         size[DLM_LOCKREPLY_OFF] = sizeof(*reply);
1266         if (lock->l_lvb_len != 0) {
1267                 buffers = 3;
1268                 size[DLM_REPLY_REC_OFF] = lock->l_lvb_len;
1269         }
1270         ptlrpc_req_set_repsize(req, buffers, size);
1271
1272         LDLM_DEBUG(lock, "replaying lock:");
1273
1274         atomic_inc(&req->rq_import->imp_replay_inflight);
1275         req->rq_async_args.pointer_arg[0] = lock;
1276         req->rq_interpret_reply = replay_lock_interpret;
1277         ptlrpcd_add_req(req);
1278
1279         RETURN(0);
1280 }
1281
1282 int ldlm_replay_locks(struct obd_import *imp)
1283 {
1284         struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
1285         struct list_head list;
1286         struct ldlm_lock *lock, *next;
1287         int rc = 0;
1288
1289         ENTRY;
1290         CFS_INIT_LIST_HEAD(&list);
1291
1292         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1293
1294         /* ensure this doesn't fall to 0 before all have been queued */
1295         atomic_inc(&imp->imp_replay_inflight);
1296
1297         (void)ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
1298
1299         list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
1300                 list_del_init(&lock->l_pending_chain);
1301                 if (rc)
1302                         continue; /* or try to do the rest? */
1303                 rc = replay_one_lock(imp, lock);
1304         }
1305
1306         atomic_dec(&imp->imp_replay_inflight);
1307
1308         RETURN(rc);
1309 }