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 <linux/lustre_dlm.h>
  32 #include <linux/obd_class.h>
  33 #include <linux/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         int               lwd_generation;
  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         if (lock->l_conn_export == NULL) {
  54                 static unsigned long next_dump = 0, last_dump = 0;
  55
  56                 LDLM_ERROR(lock, "lock timed out (enq %lus ago); not entering "
  57                            "recovery in server code, just going back to sleep",
  58                            lock->l_enqueued_time.tv_sec);
  59                 if (time_after(jiffies, next_dump)) {
  60                         last_dump = next_dump;
  61                         next_dump = jiffies + 300 * HZ;
  62                         ldlm_namespace_dump(D_DLMTRACE,
  63                                             lock->l_resource->lr_namespace);
  64                         if (last_dump == 0)
  65                                 libcfs_debug_dumplog();
  66                 }
  67                 RETURN(0);
  68         }
  69
  70         obd = lock->l_conn_export->exp_obd;
  71         imp = obd->u.cli.cl_import;
  72         ptlrpc_fail_import(imp, lwd->lwd_generation);
  73         LDLM_ERROR(lock, "lock timed out (enqueued %lus ago), entering "
  74                    "recovery for %s@%s", lock->l_enqueued_time.tv_sec,
  75                    imp->imp_target_uuid.uuid,
  76                    imp->imp_connection->c_remote_uuid.uuid);
  77
  78         RETURN(0);
  79 }
  80
  81 int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data)
  82 {
  83         /* XXX ALLOCATE - 160 bytes */
  84         struct lock_wait_data lwd;
  85         unsigned long irqflags;
  86         struct obd_device *obd;
  87         struct obd_import *imp = NULL;
  88         struct l_wait_info lwi;
  89         int rc = 0;
  90         ENTRY;
  91
  92         if (flags == LDLM_FL_WAIT_NOREPROC) {
  93                 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
  94                 goto noreproc;
  95         }
  96
  97         if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
  98                        LDLM_FL_BLOCK_CONV))) {
  99                 wake_up(&lock->l_waitq);
 100                 RETURN(0);
 101         }
 102
 103         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
 104                    "sleeping");
 105         ldlm_lock_dump(D_OTHER, lock, 0);
 106         ldlm_reprocess_all(lock->l_resource);
 107
 108 noreproc:
 109
 110         obd = class_exp2obd(lock->l_conn_export);
 111
 112         /* if this is a local lock, then there is no import */
 113         if (obd != NULL)
 114                 imp = obd->u.cli.cl_import;
 115
 116         lwd.lwd_lock = lock;
 117
 118         if (unlikely(flags & LDLM_FL_NO_TIMEOUT)) {
 119                 LDLM_DEBUG(lock, "waiting indefinitely because CW lock was"
 120                            " met\n");
 121                 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
 122         } else {
 123                 lwi = LWI_TIMEOUT_INTR(obd_timeout * HZ,
 124                                        ldlm_expired_completion_wait,
 125                                        interrupted_completion_wait, &lwd);
 126         }
 127
 128         if (imp != NULL) {
 129                 spin_lock_irqsave(&imp->imp_lock, irqflags);
 130                 lwd.lwd_generation = imp->imp_generation;
 131                 spin_unlock_irqrestore(&imp->imp_lock, irqflags);
 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 static int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
 155                                   struct ldlm_res_id res_id,
 156                                   __u32 type,
 157                                   ldlm_policy_data_t *policy,
 158                                   ldlm_mode_t mode,
 159                                   int *flags,
 160                                   ldlm_blocking_callback blocking,
 161                                   ldlm_completion_callback completion,
 162                                   ldlm_glimpse_callback glimpse,
 163                                   void *data, __u32 lvb_len,
 164                                   void *lvb_swabber,
 165                                   struct lustre_handle *lockh)
 166 {
 167         struct ldlm_lock *lock;
 168         int err;
 169         ENTRY;
 170
 171         if (ns->ns_client) {
 172                 CERROR("Trying to enqueue local lock in a shadow namespace\n");
 173                 LBUG();
 174         }
 175
 176         lock = ldlm_lock_create(ns, NULL, res_id, type, mode, blocking,
 177                                 completion, glimpse, data, lvb_len);
 178         if (!lock)
 179                 GOTO(out_nolock, err = -ENOMEM);
 180         LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
 181
 182         ldlm_lock_addref_internal(lock, mode);
 183         ldlm_lock2handle(lock, lockh);
 184         lock->l_flags |= LDLM_FL_LOCAL;
 185         lock->l_lvb_swabber = lvb_swabber;
 186         if (policy != NULL)
 187                 memcpy(&lock->l_policy_data, policy, sizeof(*policy));
 188         if (type == LDLM_EXTENT)
 189                 memcpy(&lock->l_req_extent, &policy->l_extent,
 190                        sizeof(policy->l_extent));
 191
 192         err = ldlm_lock_enqueue(ns, &lock, policy, flags);
 193         if (err != ELDLM_OK)
 194                 GOTO(out, err);
 195
 196         if (policy != NULL)
 197                 memcpy(policy, &lock->l_policy_data, sizeof(*policy));
 198         if ((*flags) & LDLM_FL_LOCK_CHANGED)
 199                 memcpy(&res_id, &lock->l_resource->lr_name, sizeof(res_id));
 200
 201         LDLM_DEBUG_NOLOCK("client-side local enqueue handler END (lock %p)",
 202                           lock);
 203
 204         if (lock->l_completion_ast)
 205                 lock->l_completion_ast(lock, *flags, NULL);
 206
 207         LDLM_DEBUG(lock, "client-side local enqueue END");
 208         EXIT;
 209  out:
 210         LDLM_LOCK_PUT(lock);
 211  out_nolock:
 212         return err;
 213 }
 214
 215 static void failed_lock_cleanup(struct ldlm_namespace *ns,
 216                                 struct ldlm_lock *lock,
 217                                 struct lustre_handle *lockh, int mode)
 218 {
 219         /* Set a flag to prevent us from sending a CANCEL (bug 407) */
 220         l_lock(&ns->ns_lock);
 221         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 222         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 223         l_unlock(&ns->ns_lock);
 224
 225         ldlm_lock_decref_and_cancel(lockh, mode);
 226
 227         /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
 228          *       from llite/file.c/ll_file_flock(). */
 229         if (lock->l_resource->lr_type == LDLM_FLOCK) {
 230                 ldlm_lock_destroy(lock);
 231         }
 232 }
 233
 234 int ldlm_cli_enqueue(struct obd_export *exp,
 235                      struct ptlrpc_request *req,
 236                      struct ldlm_namespace *ns,
 237                      struct ldlm_res_id res_id,
 238                      __u32 type,
 239                      ldlm_policy_data_t *policy,
 240                      ldlm_mode_t mode,
 241                      int *flags,
 242                      ldlm_blocking_callback blocking,
 243                      ldlm_completion_callback completion,
 244                      ldlm_glimpse_callback glimpse,
 245                      void *data,
 246                      void *lvb,
 247                      __u32 lvb_len,
 248                      void *lvb_swabber,
 249                      struct lustre_handle *lockh)
 250 {
 251         struct ldlm_lock *lock;
 252         struct ldlm_request *body;
 253         struct ldlm_reply *reply;
 254         int rc, size[2] = {sizeof(*body), lvb_len}, req_passed_in = 1;
 255         int is_replay = *flags & LDLM_FL_REPLAY;
 256         int cleanup_phase = 0;
 257         ENTRY;
 258
 259         if (exp == NULL) {
 260                 LASSERT(!is_replay);
 261                 rc = ldlm_cli_enqueue_local(ns, res_id, type, policy, mode,
 262                                             flags, blocking, completion,
 263                                             glimpse, data, lvb_len, lvb_swabber,
 264                                             lockh);
 265                 RETURN(rc);
 266         }
 267
 268         /* If we're replaying this lock, just check some invariants.
 269          * If we're creating a new lock, get everything all setup nice. */
 270         if (is_replay) {
 271                 lock = ldlm_handle2lock(lockh);
 272                 LDLM_DEBUG(lock, "client-side enqueue START");
 273                 LASSERT(exp == lock->l_conn_export);
 274         } else {
 275                 lock = ldlm_lock_create(ns, NULL, res_id, type, mode, blocking,
 276                                         completion, glimpse, data, lvb_len);
 277                 if (lock == NULL)
 278                         RETURN(-ENOMEM);
 279                 /* for the local lock, add the reference */
 280                 ldlm_lock_addref_internal(lock, mode);
 281                 ldlm_lock2handle(lock, lockh);
 282                 lock->l_lvb_swabber = lvb_swabber;
 283                 if (policy != NULL)
 284                         memcpy(&lock->l_policy_data, policy, sizeof(*policy));
 285                 if (type == LDLM_EXTENT)
 286                         memcpy(&lock->l_req_extent, &policy->l_extent,
 287                                sizeof(policy->l_extent));
 288                 LDLM_DEBUG(lock, "client-side enqueue START");
 289         }
 290
 291         /* lock not sent to server yet */
 292         cleanup_phase = 2;
 293
 294         if (req == NULL) {
 295                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LDLM_ENQUEUE, 1,
 296                                       size, NULL);
 297                 if (req == NULL)
 298                         GOTO(cleanup, rc = -ENOMEM);
 299                 req_passed_in = 0;
 300         } else if (req->rq_reqmsg->buflens[0] != sizeof(*body))
 301                 LBUG();
 302
 303         /* Dump lock data into the request buffer */
 304         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 305         ldlm_lock2desc(lock, &body->lock_desc);
 306         body->lock_flags = *flags;
 307
 308         memcpy(&body->lock_handle1, lockh, sizeof(*lockh));
 309
 310         /* Continue as normal. */
 311         if (!req_passed_in) {
 312                 int buffers = 1;
 313                 if (lvb_len > 0)
 314                         buffers = 2;
 315                 size[0] = sizeof(*reply);
 316                 req->rq_replen = lustre_msg_size(buffers, size);
 317         }
 318         lock->l_conn_export = exp;
 319         lock->l_export = NULL;
 320         lock->l_blocking_ast = blocking;
 321
 322         LDLM_DEBUG(lock, "sending request");
 323         rc = ptlrpc_queue_wait(req);
 324
 325         if (rc != ELDLM_OK) {
 326                 LASSERT(!is_replay);
 327                 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
 328                            rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
 329                 if (rc == ELDLM_LOCK_ABORTED) {
 330                         /* Before we return, swab the reply */
 331                         reply = lustre_swab_repbuf(req, 0, sizeof(*reply),
 332                                                    lustre_swab_ldlm_reply);
 333                         if (reply == NULL) {
 334                                 CERROR("Can't unpack ldlm_reply\n");
 335                                 rc = -EPROTO;
 336                         }
 337                         if (lvb_len) {
 338                                 void *tmplvb;
 339                                 tmplvb = lustre_swab_repbuf(req, 1, lvb_len,
 340                                                             lvb_swabber);
 341                                 if (tmplvb == NULL)
 342                                         GOTO(cleanup, rc = -EPROTO);
 343                                 if (lvb != NULL)
 344                                         memcpy(lvb, tmplvb, lvb_len);
 345                         }
 346                 }
 347                 GOTO(cleanup, rc);
 348         }
 349
 350         reply = lustre_swab_repbuf(req, 0, sizeof(*reply),
 351                                    lustre_swab_ldlm_reply);
 352         if (reply == NULL) {
 353                 CERROR("Can't unpack ldlm_reply\n");
 354                 GOTO(cleanup, rc = -EPROTO);
 355         }
 356
 357         /* lock enqueued on the server */
 358         cleanup_phase = 1;
 359
 360         memcpy(&lock->l_remote_handle, &reply->lock_handle,
 361                sizeof(lock->l_remote_handle));
 362         *flags = reply->lock_flags;
 363
 364         CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%x\n",
 365                lock, reply->lock_handle.cookie, *flags);
 366
 367         /* If enqueue returned a blocked lock but the completion handler has
 368          * already run, then it fixed up the resource and we don't need to do it
 369          * again. */
 370         if ((*flags) & LDLM_FL_LOCK_CHANGED) {
 371                 int newmode = reply->lock_desc.l_req_mode;
 372                 LASSERT(!is_replay);
 373                 if (newmode && newmode != lock->l_req_mode) {
 374                         LDLM_DEBUG(lock, "server returned different mode %s",
 375                                    ldlm_lockname[newmode]);
 376                         lock->l_req_mode = newmode;
 377                 }
 378
 379                 if (reply->lock_desc.l_resource.lr_name.name[0] !=
 380                     lock->l_resource->lr_name.name[0]) {
 381                         CDEBUG(D_INFO, "remote intent success, locking %ld "
 382                                "instead of %ld\n",
 383                               (long)reply->lock_desc.l_resource.lr_name.name[0],
 384                                (long)lock->l_resource->lr_name.name[0]);
 385
 386                         ldlm_lock_change_resource(ns, lock,
 387                                            reply->lock_desc.l_resource.lr_name);
 388                         if (lock->l_resource == NULL) {
 389                                 LBUG();
 390                                 GOTO(cleanup, rc = -ENOMEM);
 391                         }
 392                         LDLM_DEBUG(lock, "client-side enqueue, new resource");
 393                 }
 394                 if (policy != NULL)
 395                         memcpy(&lock->l_policy_data,
 396                                &reply->lock_desc.l_policy_data,
 397                                sizeof(reply->lock_desc.l_policy_data));
 398                 if (type != LDLM_PLAIN)
 399                         LDLM_DEBUG(lock,"client-side enqueue, new policy data");
 400         }
 401
 402         if ((*flags) & LDLM_FL_AST_SENT) {
 403                 l_lock(&ns->ns_lock);
 404                 lock->l_flags |= LDLM_FL_CBPENDING;
 405                 l_unlock(&ns->ns_lock);
 406                 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
 407         }
 408
 409         /* If the lock has already been granted by a completion AST, don't
 410          * clobber the LVB with an older one. */
 411         if (lvb_len && (lock->l_req_mode != lock->l_granted_mode)) {
 412                 void *tmplvb;
 413                 tmplvb = lustre_swab_repbuf(req, 1, lvb_len, lvb_swabber);
 414                 if (tmplvb == NULL)
 415                         GOTO(cleanup, rc = -EPROTO);
 416                 memcpy(lock->l_lvb_data, tmplvb, lvb_len);
 417         }
 418
 419         if (!is_replay) {
 420                 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
 421                 if (lock->l_completion_ast != NULL) {
 422                         int err = lock->l_completion_ast(lock, *flags, NULL);
 423                         if (!rc)
 424                                 rc = err;
 425                         if (rc)
 426                                 cleanup_phase = 2;
 427                 }
 428         }
 429
 430         if (lvb_len && lvb != NULL) {
 431                 /* Copy the LVB here, and not earlier, because the completion
 432                  * AST (if any) can override what we got in the reply */
 433                 memcpy(lvb, lock->l_lvb_data, lvb_len);
 434         }
 435
 436         LDLM_DEBUG(lock, "client-side enqueue END");
 437         EXIT;
 438 cleanup:
 439         switch (cleanup_phase) {
 440         case 2:
 441                 if (rc)
 442                         failed_lock_cleanup(ns, lock, lockh, mode);
 443         case 1:
 444                 if (!req_passed_in && req != NULL)
 445                         ptlrpc_req_finished(req);
 446         }
 447
 448         LDLM_LOCK_PUT(lock);
 449         return rc;
 450 }
 451
 452 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
 453                                   int *flags)
 454 {
 455         struct ldlm_resource *res;
 456         int rc;
 457         ENTRY;
 458         if (lock->l_resource->lr_namespace->ns_client) {
 459                 CERROR("Trying to cancel local lock\n");
 460                 LBUG();
 461         }
 462         LDLM_DEBUG(lock, "client-side local convert");
 463
 464         res = ldlm_lock_convert(lock, new_mode, flags);
 465         if (res) {
 466                 ldlm_reprocess_all(res);
 467                 rc = 0;
 468         } else {
 469                 rc = EDEADLOCK;
 470         }
 471         LDLM_DEBUG(lock, "client-side local convert handler END");
 472         LDLM_LOCK_PUT(lock);
 473         RETURN(rc);
 474 }
 475
 476 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
 477  * conversion of locks which are on the waiting or converting queue */
 478 /* Caller of this code is supposed to take care of lock readers/writers
 479    accounting */
 480 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, int *flags)
 481 {
 482         struct ldlm_request *body;
 483         struct ldlm_reply *reply;
 484         struct ldlm_lock *lock;
 485         struct ldlm_resource *res;
 486         struct ptlrpc_request *req;
 487         int rc, size = sizeof(*body);
 488         ENTRY;
 489
 490         lock = ldlm_handle2lock(lockh);
 491         if (!lock) {
 492                 LBUG();
 493                 RETURN(-EINVAL);
 494         }
 495         *flags = 0;
 496
 497         if (lock->l_conn_export == NULL)
 498                 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
 499
 500         LDLM_DEBUG(lock, "client-side convert");
 501
 502         req = ptlrpc_prep_req(class_exp2cliimp(lock->l_conn_export),
 503                               LDLM_CONVERT, 1, &size, NULL);
 504         if (!req)
 505                 GOTO(out, rc = -ENOMEM);
 506
 507         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 508         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 509                sizeof(body->lock_handle1));
 510
 511         body->lock_desc.l_req_mode = new_mode;
 512         body->lock_flags = *flags;
 513
 514         size = sizeof(*reply);
 515         req->rq_replen = lustre_msg_size(1, &size);
 516
 517         rc = ptlrpc_queue_wait(req);
 518         if (rc != ELDLM_OK)
 519                 GOTO(out, rc);
 520
 521         reply = lustre_swab_repbuf(req, 0, sizeof (*reply),
 522                                    lustre_swab_ldlm_reply);
 523         if (reply == NULL) {
 524                 CERROR ("Can't unpack ldlm_reply\n");
 525                 GOTO (out, rc = -EPROTO);
 526         }
 527
 528         if (req->rq_status)
 529                 GOTO(out, rc = req->rq_status);
 530
 531         res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
 532         if (res != NULL) {
 533                 ldlm_reprocess_all(res);
 534                 /* Go to sleep until the lock is granted. */
 535                 /* FIXME: or cancelled. */
 536                 if (lock->l_completion_ast) {
 537                         rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
 538                                                     NULL);
 539                         if (rc)
 540                                 GOTO(out, rc);
 541                 }
 542         } else {
 543                 rc = EDEADLOCK;
 544         }
 545         EXIT;
 546  out:
 547         LDLM_LOCK_PUT(lock);
 548         ptlrpc_req_finished(req);
 549         return rc;
 550 }
 551
 552 int ldlm_cli_cancel(struct lustre_handle *lockh)
 553 {
 554         struct ptlrpc_request *req;
 555         struct ldlm_lock *lock;
 556         struct ldlm_request *body;
 557         int rc = 0, size = sizeof(*body);
 558         ENTRY;
 559
 560         /* concurrent cancels on the same handle can happen */
 561         lock = __ldlm_handle2lock(lockh, LDLM_FL_CANCELING);
 562         if (lock == NULL)
 563                 RETURN(0);
 564
 565         if (lock->l_conn_export) {
 566                 int local_only;
 567                 struct obd_import *imp;
 568
 569                 LDLM_DEBUG(lock, "client-side cancel");
 570                 /* Set this flag to prevent others from getting new references*/
 571                 l_lock(&lock->l_resource->lr_namespace->ns_lock);
 572                 lock->l_flags |= LDLM_FL_CBPENDING;
 573                 local_only = (lock->l_flags & LDLM_FL_LOCAL_ONLY);
 574                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 575                 ldlm_cancel_callback(lock);
 576
 577                 if (local_only) {
 578                         CDEBUG(D_INFO, "not sending request (at caller's "
 579                                "instruction)\n");
 580                         goto local_cancel;
 581                 }
 582
 583         restart:
 584                 imp = class_exp2cliimp(lock->l_conn_export);
 585                 if (imp == NULL || imp->imp_invalid) {
 586                         CDEBUG(D_HA, "skipping cancel on invalid import %p\n",
 587                                imp);
 588                         goto local_cancel;
 589                 }
 590
 591                 req = ptlrpc_prep_req(imp, LDLM_CANCEL, 1, &size, NULL);
 592                 if (!req)
 593                         GOTO(out, rc = -ENOMEM);
 594                 req->rq_no_resend = 1;
 595
 596                 /* XXX FIXME bug 249 */
 597                 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
 598                 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
 599
 600                 body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 601                 memcpy(&body->lock_handle1, &lock->l_remote_handle,
 602                        sizeof(body->lock_handle1));
 603
 604                 req->rq_replen = lustre_msg_size(0, NULL);
 605
 606                 rc = ptlrpc_queue_wait(req);
 607
 608                 if (rc == ESTALE) {
 609                         CERROR("client/server (nid %s) out of sync"
 610                                " -- not fatal\n",
 611                                libcfs_nid2str(req->rq_import->
 612                                               imp_connection->c_peer.nid));
 613                 } else if (rc == -ETIMEDOUT) {
 614                         ptlrpc_req_finished(req);
 615                         GOTO(restart, rc);
 616                 } else if (rc != ELDLM_OK) {
 617                         CERROR("Got rc %d from cancel RPC: canceling "
 618                                "anyway\n", rc);
 619                 }
 620
 621                 ptlrpc_req_finished(req);
 622         local_cancel:
 623                 ldlm_lock_cancel(lock);
 624         } else {
 625                 if (lock->l_resource->lr_namespace->ns_client) {
 626                         LDLM_ERROR(lock, "Trying to cancel local lock");
 627                         LBUG();
 628                 }
 629                 LDLM_DEBUG(lock, "client-side local cancel");
 630                 ldlm_lock_cancel(lock);
 631                 ldlm_reprocess_all(lock->l_resource);
 632                 LDLM_DEBUG(lock, "client-side local cancel handler END");
 633         }
 634
 635         EXIT;
 636  out:
 637         LDLM_LOCK_PUT(lock);
 638         return rc;
 639 }
 640
 641 /* when called with LDLM_ASYNC the blocking callback will be handled
 642  * in a thread and this function will return after the thread has been
 643  * asked to call the callback.  when called with LDLM_SYNC the blocking
 644  * callback will be performed in this function. */
 645 int ldlm_cancel_lru(struct ldlm_namespace *ns, ldlm_sync_t sync)
 646 {
 647         struct ldlm_lock *lock, *next;
 648         int count, rc = 0;
 649         LIST_HEAD(cblist);
 650         ENTRY;
 651
 652 #ifndef __KERNEL__
 653         sync = LDLM_SYNC; /* force to be sync in user space */
 654 #endif
 655
 656         l_lock(&ns->ns_lock);
 657         count = ns->ns_nr_unused - ns->ns_max_unused;
 658
 659         if (count <= 0) {
 660                 l_unlock(&ns->ns_lock);
 661                 RETURN(0);
 662         }
 663
 664         list_for_each_entry_safe(lock, next, &ns->ns_unused_list, l_lru) {
 665                 LASSERT(!lock->l_readers && !lock->l_writers);
 666
 667                 /* Setting the CBPENDING flag is a little misleading, but
 668                  * prevents an important race; namely, once CBPENDING is set,
 669                  * the lock can accumulate no more readers/writers.  Since
 670                  * readers and writers are already zero here, ldlm_lock_decref
 671                  * won't see this flag and call l_blocking_ast */
 672                 lock->l_flags |= LDLM_FL_CBPENDING;
 673
 674                 LDLM_LOCK_GET(lock); /* dropped by bl thread */
 675                 ldlm_lock_remove_from_lru(lock);
 676
 677                 /* We can't re-add to l_lru as it confuses the refcounting in
 678                  * ldlm_lock_remove_from_lru() if an AST arrives after we drop
 679                  * ns_lock below.  Use l_export_chain as that is unused on
 680                  * client, and lru is client-only (l_pending_chain is used by
 681                  * ldlm_chain_lock_for_replay() on client).  bug 5666 */
 682                 if (sync != LDLM_ASYNC || ldlm_bl_to_thread(ns, NULL, lock)) {
 683                         LASSERTF(list_empty(&lock->l_export_chain),
 684                                  "lock %p next %p prev %p\n",
 685                                  lock, &lock->l_export_chain.next,
 686                                  &lock->l_export_chain.prev);
 687                         __LDLM_DEBUG(D_INFO, lock, "adding to LRU clear list");
 688                         list_add(&lock->l_export_chain, &cblist);
 689                 }
 690
 691                 if (--count == 0)
 692                         break;
 693         }
 694         l_unlock(&ns->ns_lock);
 695
 696         list_for_each_entry_safe(lock, next, &cblist, l_export_chain) {
 697                 list_del_init(&lock->l_export_chain);
 698                 ldlm_handle_bl_callback(ns, NULL, lock);
 699         }
 700         RETURN(rc);
 701 }
 702
 703 static int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
 704                                            struct ldlm_res_id res_id, int flags,
 705                                            void *opaque)
 706 {
 707         struct ldlm_resource *res;
 708         struct list_head *tmp, *next, list = LIST_HEAD_INIT(list);
 709         struct ldlm_ast_work *w;
 710         ENTRY;
 711
 712         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
 713         if (res == NULL) {
 714                 /* This is not a problem. */
 715                 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id.name[0]);
 716                 RETURN(0);
 717         }
 718
 719         l_lock(&ns->ns_lock);
 720         list_for_each(tmp, &res->lr_granted) {
 721                 struct ldlm_lock *lock;
 722                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 723
 724                 if (opaque != NULL && lock->l_ast_data != opaque) {
 725                         LDLM_ERROR(lock, "data %p doesn't match opaque %p",
 726                                    lock->l_ast_data, opaque);
 727                         //LBUG();
 728                         continue;
 729                 }
 730
 731                 if (lock->l_readers || lock->l_writers) {
 732                         if (flags & LDLM_FL_WARN) {
 733                                 LDLM_ERROR(lock, "lock in use");
 734                                 //LBUG();
 735                         }
 736                         continue;
 737                 }
 738
 739                 /* See CBPENDING comment in ldlm_cancel_lru */
 740                 lock->l_flags |= LDLM_FL_CBPENDING;
 741
 742                 OBD_ALLOC(w, sizeof(*w));
 743                 LASSERT(w);
 744
 745                 w->w_lock = LDLM_LOCK_GET(lock);
 746
 747                 list_add(&w->w_list, &list);
 748         }
 749         l_unlock(&ns->ns_lock);
 750
 751         list_for_each_safe(tmp, next, &list) {
 752                 struct lustre_handle lockh;
 753                 int rc;
 754                 w = list_entry(tmp, struct ldlm_ast_work, w_list);
 755
 756                 if (flags & LDLM_FL_LOCAL_ONLY) {
 757                         ldlm_lock_cancel(w->w_lock);
 758                 } else {
 759                         ldlm_lock2handle(w->w_lock, &lockh);
 760                         rc = ldlm_cli_cancel(&lockh);
 761                         if (rc != ELDLM_OK)
 762                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 763                 }
 764                 list_del(&w->w_list);
 765                 LDLM_LOCK_PUT(w->w_lock);
 766                 OBD_FREE(w, sizeof(*w));
 767         }
 768
 769         ldlm_resource_putref(res);
 770
 771         RETURN(0);
 772 }
 773
 774 /* Cancel all locks on a namespace (or a specific resource, if given)
 775  * that have 0 readers/writers.
 776  *
 777  * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
 778  * to notify the server.
 779  * If flags & LDLM_FL_WARN, print a warning if some locks are still in use. */
 780 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
 781                            struct ldlm_res_id *res_id, int flags, void *opaque)
 782 {
 783         int i;
 784         ENTRY;
 785
 786         if (ns == NULL)
 787                 RETURN(ELDLM_OK);
 788
 789         if (res_id)
 790                 RETURN(ldlm_cli_cancel_unused_resource(ns, *res_id, flags,
 791                                                        opaque));
 792
 793         l_lock(&ns->ns_lock);
 794         for (i = 0; i < RES_HASH_SIZE; i++) {
 795                 struct list_head *tmp, *pos;
 796                 list_for_each_safe(tmp, pos, &(ns->ns_hash[i])) {
 797                         int rc;
 798                         struct ldlm_resource *res;
 799                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
 800                         ldlm_resource_getref(res);
 801
 802                         rc = ldlm_cli_cancel_unused_resource(ns, res->lr_name,
 803                                                              flags, opaque);
 804
 805                         if (rc)
 806                                 CERROR("cancel_unused_res ("LPU64"): %d\n",
 807                                        res->lr_name.name[0], rc);
 808                         ldlm_resource_putref(res);
 809                 }
 810         }
 811         l_unlock(&ns->ns_lock);
 812
 813         RETURN(ELDLM_OK);
 814 }
 815
 816 /* join/split resource locks to/from lru list */
 817 int ldlm_cli_join_lru(struct ldlm_namespace *ns,
 818                       struct ldlm_res_id *res_id, int join)
 819 {
 820         struct ldlm_resource *res;
 821         struct ldlm_lock *lock, *n;
 822         int count = 0;
 823         ENTRY;
 824
 825         LASSERT(ns->ns_client == LDLM_NAMESPACE_CLIENT);
 826
 827         res = ldlm_resource_get(ns, NULL, *res_id, LDLM_EXTENT, 0);
 828         if (res == NULL)
 829                 RETURN(count);
 830         LASSERT(res->lr_type == LDLM_EXTENT);
 831
 832         l_lock(&ns->ns_lock);
 833         if (!join)
 834                 goto split;
 835
 836         list_for_each_entry_safe (lock, n, &res->lr_granted, l_res_link) {
 837                 if (list_empty(&lock->l_lru) &&
 838                     !lock->l_readers && !lock->l_writers &&
 839                     !(lock->l_flags & LDLM_FL_LOCAL) &&
 840                     !(lock->l_flags & LDLM_FL_CBPENDING)) {
 841                         LASSERT(ns->ns_nr_unused >= 0);
 842                         list_add_tail(&lock->l_lru, &ns->ns_unused_list);
 843                         ns->ns_nr_unused++;
 844                         lock->l_flags &= ~LDLM_FL_NO_LRU;
 845                         LDLM_DEBUG(lock, "join lock to lru");
 846                         count++;
 847                 }
 848         }
 849         goto unlock;
 850 split:
 851         list_for_each_entry_safe (lock, n, &ns->ns_unused_list, l_lru) {
 852                 if (lock->l_resource == res) {
 853                         ldlm_lock_remove_from_lru(lock);
 854                         lock->l_flags |= LDLM_FL_NO_LRU;
 855                         LDLM_DEBUG(lock, "split lock from lru");
 856                         count++;
 857                 }
 858         }
 859 unlock:
 860         l_unlock(&ns->ns_lock);
 861         ldlm_resource_putref(res);
 862         RETURN(count);
 863 }
 864
 865 /* Lock iterators. */
 866
 867 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
 868                           void *closure)
 869 {
 870         struct list_head *tmp, *next;
 871         struct ldlm_lock *lock;
 872         int rc = LDLM_ITER_CONTINUE;
 873         struct ldlm_namespace *ns = res->lr_namespace;
 874
 875         ENTRY;
 876
 877         if (!res)
 878                 RETURN(LDLM_ITER_CONTINUE);
 879
 880         l_lock(&ns->ns_lock);
 881         list_for_each_safe(tmp, next, &res->lr_granted) {
 882                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 883
 884                 if (iter(lock, closure) == LDLM_ITER_STOP)
 885                         GOTO(out, rc = LDLM_ITER_STOP);
 886         }
 887
 888         list_for_each_safe(tmp, next, &res->lr_converting) {
 889                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 890
 891                 if (iter(lock, closure) == LDLM_ITER_STOP)
 892                         GOTO(out, rc = LDLM_ITER_STOP);
 893         }
 894
 895         list_for_each_safe(tmp, next, &res->lr_waiting) {
 896                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 897
 898                 if (iter(lock, closure) == LDLM_ITER_STOP)
 899                         GOTO(out, rc = LDLM_ITER_STOP);
 900         }
 901  out:
 902         l_unlock(&ns->ns_lock);
 903         RETURN(rc);
 904 }
 905
 906 struct iter_helper_data {
 907         ldlm_iterator_t iter;
 908         void *closure;
 909 };
 910
 911 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
 912 {
 913         struct iter_helper_data *helper = closure;
 914         return helper->iter(lock, helper->closure);
 915 }
 916
 917 static int ldlm_res_iter_helper(struct ldlm_resource *res, void *closure)
 918 {
 919         return ldlm_resource_foreach(res, ldlm_iter_helper, closure);
 920 }
 921
 922 int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
 923                            void *closure)
 924 {
 925         struct iter_helper_data helper = { iter: iter, closure: closure };
 926         return ldlm_namespace_foreach_res(ns, ldlm_res_iter_helper, &helper);
 927 }
 928
 929 int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
 930                                ldlm_res_iterator_t iter, void *closure)
 931 {
 932         int i, rc = LDLM_ITER_CONTINUE;
 933
 934         l_lock(&ns->ns_lock);
 935         for (i = 0; i < RES_HASH_SIZE; i++) {
 936                 struct list_head *tmp, *next;
 937                 list_for_each_safe(tmp, next, &(ns->ns_hash[i])) {
 938                         struct ldlm_resource *res =
 939                                 list_entry(tmp, struct ldlm_resource, lr_hash);
 940
 941                         ldlm_resource_getref(res);
 942                         rc = iter(res, closure);
 943                         ldlm_resource_putref(res);
 944                         if (rc == LDLM_ITER_STOP)
 945                                 GOTO(out, rc);
 946                 }
 947         }
 948  out:
 949         l_unlock(&ns->ns_lock);
 950         RETURN(rc);
 951 }
 952
 953 /* non-blocking function to manipulate a lock whose cb_data is being put away.*/
 954 void ldlm_change_cbdata(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
 955                         ldlm_iterator_t iter, void *data)
 956 {
 957         struct ldlm_resource *res;
 958         ENTRY;
 959
 960         if (ns == NULL) {
 961                 CERROR("must pass in namespace\n");
 962                 LBUG();
 963         }
 964
 965         res = ldlm_resource_get(ns, NULL, *res_id, 0, 0);
 966         if (res == NULL) {
 967                 EXIT;
 968                 return;
 969         }
 970
 971         l_lock(&ns->ns_lock);
 972         ldlm_resource_foreach(res, iter, data);
 973         l_unlock(&ns->ns_lock);
 974         ldlm_resource_putref(res);
 975         EXIT;
 976 }
 977
 978 /* Lock replay */
 979
 980 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
 981 {
 982         struct list_head *list = closure;
 983
 984         /* we use l_pending_chain here, because it's unused on clients. */
 985         LASSERTF(list_empty(&lock->l_pending_chain),"lock %p next %p prev %p\n",
 986                  lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
 987         list_add(&lock->l_pending_chain, list);
 988         return LDLM_ITER_CONTINUE;
 989 }
 990
 991 static int replay_lock_interpret(struct ptlrpc_request *req,
 992                                     void * data, int rc)
 993 {
 994         struct ldlm_lock *lock;
 995         struct ldlm_reply *reply;
 996
 997         atomic_dec(&req->rq_import->imp_replay_inflight);
 998         if (rc != ELDLM_OK)
 999                 GOTO(out, rc);
1000
1001         lock = req->rq_async_args.pointer_arg[0];
1002         LASSERT(lock != NULL);
1003
1004         reply = lustre_swab_repbuf(req, 0, sizeof (*reply),
1005                                    lustre_swab_ldlm_reply);
1006         if (reply == NULL) {
1007                 CERROR("Can't unpack ldlm_reply\n");
1008                 GOTO (out, rc = -EPROTO);
1009         }
1010
1011         memcpy(&lock->l_remote_handle, &reply->lock_handle,
1012                sizeof(lock->l_remote_handle));
1013         LDLM_DEBUG(lock, "replayed lock:");
1014         ptlrpc_import_recovery_state_machine(req->rq_import);
1015  out:
1016         if (rc != ELDLM_OK)
1017                 ptlrpc_connect_import(req->rq_import, NULL);
1018
1019
1020         RETURN(rc);
1021 }
1022
1023 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
1024 {
1025         struct ptlrpc_request *req;
1026         struct ldlm_request *body;
1027         struct ldlm_reply *reply;
1028         int buffers = 1;
1029         int size[2];
1030         int flags;
1031
1032         /*
1033          * If granted mode matches the requested mode, this lock is granted.
1034          *
1035          * If they differ, but we have a granted mode, then we were granted
1036          * one mode and now want another: ergo, converting.
1037          *
1038          * If we haven't been granted anything and are on a resource list,
1039          * then we're blocked/waiting.
1040          *
1041          * If we haven't been granted anything and we're NOT on a resource list,
1042          * then we haven't got a reply yet and don't have a known disposition.
1043          * This happens whenever a lock enqueue is the request that triggers
1044          * recovery.
1045          */
1046         if (lock->l_granted_mode == lock->l_req_mode)
1047                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
1048         else if (lock->l_granted_mode)
1049                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
1050         else if (!list_empty(&lock->l_res_link))
1051                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
1052         else
1053                 flags = LDLM_FL_REPLAY;
1054
1055         size[0] = sizeof(*body);
1056         req = ptlrpc_prep_req(imp, LDLM_ENQUEUE, 1, size, NULL);
1057         if (!req)
1058                 RETURN(-ENOMEM);
1059
1060         /* We're part of recovery, so don't wait for it. */
1061         req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
1062
1063         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
1064         ldlm_lock2desc(lock, &body->lock_desc);
1065         body->lock_flags = flags;
1066
1067         ldlm_lock2handle(lock, &body->lock_handle1);
1068         size[0] = sizeof(*reply);
1069         if (lock->l_lvb_len != 0) {
1070                 buffers = 2;
1071                 size[1] = lock->l_lvb_len;
1072         }
1073         req->rq_replen = lustre_msg_size(buffers, size);
1074
1075         LDLM_DEBUG(lock, "replaying lock:");
1076
1077         atomic_inc(&req->rq_import->imp_replay_inflight);
1078         req->rq_async_args.pointer_arg[0] = lock;
1079         req->rq_interpret_reply = replay_lock_interpret;
1080         ptlrpcd_add_req(req);
1081
1082         RETURN(0);
1083 }
1084
1085 int ldlm_replay_locks(struct obd_import *imp)
1086 {
1087         struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
1088         struct list_head list;
1089         struct ldlm_lock *lock, *next;
1090         int rc = 0;
1091
1092         ENTRY;
1093         INIT_LIST_HEAD(&list);
1094
1095         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1096
1097         /* ensure this doesn't fall to 0 before all have been queued */
1098         atomic_inc(&imp->imp_replay_inflight);
1099
1100         l_lock(&ns->ns_lock);
1101         (void)ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
1102
1103         list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
1104                 list_del_init(&lock->l_pending_chain);
1105                 if (rc)
1106                         continue; /* or try to do the rest? */
1107                 rc = replay_one_lock(imp, lock);
1108         }
1109         l_unlock(&ns->ns_lock);
1110
1111         atomic_dec(&imp->imp_replay_inflight);
1112
1113         RETURN(rc);
1114 }