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 Lustre, http://www.lustre.org.
   7  *
   8  *   Lustre is free software; you can redistribute it and/or
   9  *   modify it under the terms of version 2 of the GNU General Public
  10  *   License as published by the Free Software Foundation.
  11  *
  12  *   Lustre is distributed in the hope that it will be useful,
  13  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  *   GNU General Public License for more details.
  16  *
  17  *   You should have received a copy of the GNU General Public License
  18  *   along with Lustre; if not, write to the Free Software
  19  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  20  */
  21
  22 #define DEBUG_SUBSYSTEM S_LDLM
  23 #ifndef __KERNEL__
  24 #include <signal.h>
  25 #include <liblustre.h>
  26 #endif
  27
  28 #include <linux/lustre_dlm.h>
  29 #include <linux/obd_class.h>
  30 #include <linux/obd.h>
  31
  32 #include "ldlm_internal.h"
  33
  34 static void interrupted_completion_wait(void *data)
  35 {
  36 }
  37
  38 struct lock_wait_data {
  39         struct ldlm_lock *lwd_lock;
  40         int               lwd_generation;
  41 };
  42
  43 int ldlm_expired_completion_wait(void *data)
  44 {
  45         struct lock_wait_data *lwd = data;
  46         struct ldlm_lock *lock = lwd->lwd_lock;
  47         struct obd_import *imp;
  48         struct obd_device *obd;
  49
  50         if (lock->l_conn_export == NULL) {
  51                 static unsigned long next_dump = 0, last_dump = 0;
  52
  53                 LDLM_ERROR(lock, "lock timed out; not entering recovery in "
  54                            "server code, just going back to sleep");
  55                 if (time_after(jiffies, next_dump)) {
  56                         last_dump = next_dump;
  57                         next_dump = jiffies + 300 * HZ;
  58                         ldlm_namespace_dump(D_DLMTRACE,
  59                                             lock->l_resource->lr_namespace);
  60                         if (last_dump == 0)
  61                                 portals_debug_dumplog();
  62                 }
  63                 RETURN(0);
  64         }
  65
  66         obd = lock->l_conn_export->exp_obd;
  67         imp = obd->u.cli.cl_import;
  68         ptlrpc_fail_import(imp, lwd->lwd_generation);
  69         LDLM_ERROR(lock, "lock timed out, entering recovery for %s@%s",
  70                    imp->imp_target_uuid.uuid,
  71                    imp->imp_connection->c_remote_uuid.uuid);
  72
  73         RETURN(0);
  74 }
  75
  76 int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data)
  77 {
  78         /* XXX ALLOCATE - 160 bytes */
  79         struct lock_wait_data lwd;
  80         unsigned long irqflags;
  81         struct obd_device *obd;
  82         struct obd_import *imp = NULL;
  83         struct l_wait_info lwi;
  84         int rc = 0;
  85         ENTRY;
  86
  87         if (flags == LDLM_FL_WAIT_NOREPROC)
  88                 goto noreproc;
  89
  90         if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
  91                        LDLM_FL_BLOCK_CONV))) {
  92                 wake_up(&lock->l_waitq);
  93                 RETURN(0);
  94         }
  95
  96         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
  97                    "sleeping");
  98         ldlm_lock_dump(D_OTHER, lock, 0);
  99         ldlm_reprocess_all(lock->l_resource);
 100
 101 noreproc:
 102
 103         obd = class_exp2obd(lock->l_conn_export);
 104
 105         /* if this is a local lock, then there is no import */
 106         if (obd != NULL)
 107                 imp = obd->u.cli.cl_import;
 108
 109         lwd.lwd_lock = lock;
 110
 111         if (flags & LDLM_FL_NO_TIMEOUT) {
 112                 LDLM_DEBUG(lock, "waiting indefinitely for group lock\n");
 113                 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
 114         } else {
 115                 lwi = LWI_TIMEOUT_INTR(obd_timeout * HZ,
 116                                        ldlm_expired_completion_wait,
 117                                        interrupted_completion_wait, &lwd);
 118         }
 119
 120         if (imp != NULL) {
 121                 spin_lock_irqsave(&imp->imp_lock, irqflags);
 122                 lwd.lwd_generation = imp->imp_generation;
 123                 spin_unlock_irqrestore(&imp->imp_lock, irqflags);
 124         }
 125
 126         /* Go to sleep until the lock is granted or cancelled. */
 127         rc = l_wait_event(lock->l_waitq,
 128                           ((lock->l_req_mode == lock->l_granted_mode) ||
 129                            (lock->l_flags & LDLM_FL_FAILED)), &lwi);
 130
 131         if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED) {
 132                 LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed");
 133                 RETURN(-EIO);
 134         }
 135
 136         if (rc) {
 137                 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
 138                            rc);
 139                 RETURN(rc);
 140         }
 141
 142         LDLM_DEBUG(lock, "client-side enqueue waking up: granted");
 143         RETURN(0);
 144 }
 145
 146 static int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
 147                                   struct ldlm_res_id res_id,
 148                                   __u32 type,
 149                                   ldlm_policy_data_t *policy,
 150                                   ldlm_mode_t mode,
 151                                   int *flags,
 152                                   ldlm_blocking_callback blocking,
 153                                   ldlm_completion_callback completion,
 154                                   ldlm_glimpse_callback glimpse,
 155                                   void *data, __u32 lvb_len,
 156                                   void *lvb_swabber,
 157                                   struct lustre_handle *lockh)
 158 {
 159         struct ldlm_lock *lock;
 160         int err;
 161         ENTRY;
 162
 163         if (ns->ns_client) {
 164                 CERROR("Trying to enqueue local lock in a shadow namespace\n");
 165                 LBUG();
 166         }
 167
 168         lock = ldlm_lock_create(ns, NULL, res_id, type, mode, blocking,
 169                                 completion, glimpse, data, lvb_len);
 170         if (!lock)
 171                 GOTO(out_nolock, err = -ENOMEM);
 172         LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
 173
 174         ldlm_lock_addref_internal(lock, mode);
 175         ldlm_lock2handle(lock, lockh);
 176         lock->l_flags |= LDLM_FL_LOCAL;
 177         lock->l_lvb_swabber = lvb_swabber;
 178         if (policy != NULL)
 179                 memcpy(&lock->l_policy_data, policy, sizeof(*policy));
 180         if (type == LDLM_EXTENT)
 181                 memcpy(&lock->l_req_extent, &policy->l_extent,
 182                        sizeof(policy->l_extent));
 183
 184         err = ldlm_lock_enqueue(ns, &lock, policy, flags);
 185         if (err != ELDLM_OK)
 186                 GOTO(out, err);
 187
 188         if (policy != NULL)
 189                 memcpy(policy, &lock->l_policy_data, sizeof(*policy));
 190         if ((*flags) & LDLM_FL_LOCK_CHANGED)
 191                 memcpy(&res_id, &lock->l_resource->lr_name, sizeof(res_id));
 192
 193         LDLM_DEBUG_NOLOCK("client-side local enqueue handler END (lock %p)",
 194                           lock);
 195
 196         if (lock->l_completion_ast)
 197                 lock->l_completion_ast(lock, *flags, NULL);
 198
 199         LDLM_DEBUG(lock, "client-side local enqueue END");
 200         EXIT;
 201  out:
 202         LDLM_LOCK_PUT(lock);
 203  out_nolock:
 204         return err;
 205 }
 206
 207 static void failed_lock_cleanup(struct ldlm_namespace *ns,
 208                                 struct ldlm_lock *lock,
 209                                 struct lustre_handle *lockh, int mode)
 210 {
 211         /* Set a flag to prevent us from sending a CANCEL (bug 407) */
 212         l_lock(&ns->ns_lock);
 213         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 214         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 215         l_unlock(&ns->ns_lock);
 216
 217         ldlm_lock_decref_and_cancel(lockh, mode);
 218
 219         /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
 220          *       from llite/file.c/ll_file_flock(). */
 221         if (lock->l_resource->lr_type == LDLM_FLOCK) {
 222                 ldlm_lock_destroy(lock);
 223         }
 224 }
 225
 226 int ldlm_cli_enqueue(struct obd_export *exp,
 227                      struct ptlrpc_request *req,
 228                      struct ldlm_namespace *ns,
 229                      struct ldlm_res_id res_id,
 230                      __u32 type,
 231                      ldlm_policy_data_t *policy,
 232                      ldlm_mode_t mode,
 233                      int *flags,
 234                      ldlm_blocking_callback blocking,
 235                      ldlm_completion_callback completion,
 236                      ldlm_glimpse_callback glimpse,
 237                      void *data,
 238                      void *lvb,
 239                      __u32 lvb_len,
 240                      void *lvb_swabber,
 241                      struct lustre_handle *lockh)
 242 {
 243         struct ldlm_lock *lock;
 244         struct ldlm_request *body;
 245         struct ldlm_reply *reply;
 246         int rc, size[3] = {0, sizeof(*body), lvb_len}, req_passed_in = 1;
 247         int is_replay = *flags & LDLM_FL_REPLAY;
 248         int cleanup_phase = 0;
 249         ENTRY;
 250
 251         if (exp == NULL) {
 252                 LASSERT(!is_replay);
 253                 rc = ldlm_cli_enqueue_local(ns, res_id, type, policy, mode,
 254                                             flags, blocking, completion,
 255                                             glimpse, data, lvb_len, lvb_swabber,
 256                                             lockh);
 257                 RETURN(rc);
 258         }
 259
 260         /* If we're replaying this lock, just check some invariants.
 261          * If we're creating a new lock, get everything all setup nice. */
 262         if (is_replay) {
 263                 lock = ldlm_handle2lock(lockh);
 264                 LDLM_DEBUG(lock, "client-side enqueue START");
 265                 LASSERT(exp == lock->l_conn_export);
 266         } else {
 267                 lock = ldlm_lock_create(ns, NULL, res_id, type, mode, blocking,
 268                                         completion, glimpse, data, lvb_len);
 269                 if (lock == NULL)
 270                         RETURN(-ENOMEM);
 271                 /* for the local lock, add the reference */
 272                 ldlm_lock_addref_internal(lock, mode);
 273                 ldlm_lock2handle(lock, lockh);
 274                 lock->l_lvb_swabber = lvb_swabber;
 275                 if (policy != NULL)
 276                         memcpy(&lock->l_policy_data, policy, sizeof(*policy));
 277                 if (type == LDLM_EXTENT)
 278                         memcpy(&lock->l_req_extent, &policy->l_extent,
 279                                sizeof(policy->l_extent));
 280                 LDLM_DEBUG(lock, "client-side enqueue START");
 281         }
 282
 283         /* lock not sent to server yet */
 284         cleanup_phase = 2;
 285
 286         if (req == NULL) {
 287                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
 288                                       LDLM_ENQUEUE, 2, size, NULL);
 289                 if (req == NULL)
 290                         GOTO(cleanup, rc = -ENOMEM);
 291                 req_passed_in = 0;
 292         }
 293
 294         LASSERTF(req->rq_reqmsg->buflens[MDS_REQ_INTENT_LOCKREQ_OFF] ==
 295                  sizeof(*body), "buflen[%d] = %d, not %d\n",
 296                  MDS_REQ_INTENT_LOCKREQ_OFF,
 297                  req->rq_reqmsg->buflens[MDS_REQ_INTENT_LOCKREQ_OFF],
 298                  sizeof(*body));
 299
 300         /* Dump lock data into the request buffer */
 301         body = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_INTENT_LOCKREQ_OFF,
 302                               sizeof (*body));
 303         ldlm_lock2desc(lock, &body->lock_desc);
 304         body->lock_flags = *flags;
 305
 306         memcpy(&body->lock_handle1, lockh, sizeof(*lockh));
 307
 308         /* Continue as normal. */
 309         if (!req_passed_in) {
 310                 size[0] = sizeof(*reply);
 311                 req->rq_replen = lustre_msg_size(1 + (lvb_len > 0), size);
 312         }
 313         lock->l_conn_export = exp;
 314         lock->l_export = NULL;
 315         lock->l_blocking_ast = blocking;
 316
 317         LDLM_DEBUG(lock, "sending request");
 318         rc = ptlrpc_queue_wait(req);
 319
 320         if (rc != ELDLM_OK) {
 321                 LASSERT(!is_replay);
 322                 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
 323                            rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
 324                 if (rc == ELDLM_LOCK_ABORTED) {
 325                         /* Before we return, swab the reply */
 326                         reply = lustre_swab_repbuf(req, 0, sizeof(*reply),
 327                                                    lustre_swab_ldlm_reply);
 328                         if (reply == NULL) {
 329                                 CERROR("Can't unpack ldlm_reply\n");
 330                                 rc = -EPROTO;
 331                         }
 332                         if (lvb_len) {
 333                                 void *tmplvb;
 334                                 tmplvb = lustre_swab_repbuf(req, 1, lvb_len,
 335                                                             lvb_swabber);
 336                                 if (tmplvb == NULL)
 337                                         GOTO(cleanup, rc = -EPROTO);
 338                                 if (lvb != NULL)
 339                                         memcpy(lvb, tmplvb, lvb_len);
 340                         }
 341                 }
 342                 GOTO(cleanup, rc);
 343         }
 344
 345         reply = lustre_swab_repbuf(req, 0, sizeof(*reply),
 346                                    lustre_swab_ldlm_reply);
 347         if (reply == NULL) {
 348                 CERROR("Can't unpack ldlm_reply\n");
 349                 GOTO(cleanup, rc = -EPROTO);
 350         }
 351
 352         /* lock enqueued on the server */
 353         cleanup_phase = 1;
 354
 355         memcpy(&lock->l_remote_handle, &reply->lock_handle,
 356                sizeof(lock->l_remote_handle));
 357         *flags = reply->lock_flags;
 358
 359         CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%x\n",
 360                lock, reply->lock_handle.cookie, *flags);
 361
 362         /* If enqueue returned a blocked lock but the completion handler has
 363          * already run, then it fixed up the resource and we don't need to do it
 364          * again. */
 365         if ((*flags) & LDLM_FL_LOCK_CHANGED) {
 366                 int newmode = reply->lock_desc.l_req_mode;
 367                 LASSERT(!is_replay);
 368                 if (newmode && newmode != lock->l_req_mode) {
 369                         LDLM_DEBUG(lock, "server returned different mode %s",
 370                                    ldlm_lockname[newmode]);
 371                         lock->l_req_mode = newmode;
 372                 }
 373
 374                 if (reply->lock_desc.l_resource.lr_name.name[0] !=
 375                     lock->l_resource->lr_name.name[0] ||
 376                    reply->lock_desc.l_resource.lr_name.name[1] !=
 377                     lock->l_resource->lr_name.name[1]) {
 378                         CDEBUG(D_INFO, "remote intent success, locking %ld "
 379                                "instead of %ld\n",
 380                               (long)reply->lock_desc.l_resource.lr_name.name[0],
 381                                (long)lock->l_resource->lr_name.name[0]);
 382
 383                         ldlm_lock_change_resource(ns, lock,
 384                                            reply->lock_desc.l_resource.lr_name);
 385                         if (lock->l_resource == NULL) {
 386                                 LBUG();
 387                                 GOTO(cleanup, rc = -ENOMEM);
 388                         }
 389                         LDLM_DEBUG(lock, "client-side enqueue, new resource");
 390                 }
 391                 if (policy != NULL)
 392                         memcpy(&lock->l_policy_data,
 393                                &reply->lock_desc.l_policy_data,
 394                                sizeof(reply->lock_desc.l_policy_data));
 395                 if (type != LDLM_PLAIN)
 396                         LDLM_DEBUG(lock,"client-side enqueue, new policy data");
 397         }
 398
 399         if ((*flags) & LDLM_FL_AST_SENT) {
 400                 l_lock(&ns->ns_lock);
 401                 lock->l_flags |= LDLM_FL_CBPENDING;
 402                 l_unlock(&ns->ns_lock);
 403                 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
 404         }
 405
 406         /* If the lock has already been granted by a completion AST, don't
 407          * clobber the LVB with an older one. */
 408         if (lvb_len && (lock->l_req_mode != lock->l_granted_mode)) {
 409                 void *tmplvb;
 410                 tmplvb = lustre_swab_repbuf(req, 1, lvb_len, lvb_swabber);
 411                 if (tmplvb == NULL) {
 412                         cleanup_phase = 2;
 413                         GOTO(cleanup, rc = -EPROTO);
 414                 }
 415                 memcpy(lock->l_lvb_data, tmplvb, lvb_len);
 416         }
 417
 418         if (!is_replay) {
 419                 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
 420                 if (lock->l_completion_ast != NULL) {
 421                         int err = lock->l_completion_ast(lock, *flags, NULL);
 422                         if (!rc)
 423                                 rc = err;
 424                         if (rc)
 425                                 cleanup_phase = 2;
 426                 }
 427         }
 428
 429         if (lvb_len && lvb != NULL) {
 430                 /* Copy the LVB here, and not earlier, because the completion
 431                  * AST (if any) can override what we got in the reply */
 432                 memcpy(lvb, lock->l_lvb_data, lvb_len);
 433         }
 434
 435         LDLM_DEBUG(lock, "client-side enqueue END");
 436         EXIT;
 437 cleanup:
 438         switch (cleanup_phase) {
 439         case 2:
 440                 if (rc)
 441                         failed_lock_cleanup(ns, lock, lockh, mode);
 442         case 1:
 443                 if (!req_passed_in && req != NULL)
 444                         ptlrpc_req_finished(req);
 445         }
 446
 447         LDLM_LOCK_PUT(lock);
 448         return rc;
 449 }
 450
 451 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
 452                                   int *flags)
 453 {
 454         struct ldlm_resource *res;
 455         int rc;
 456         ENTRY;
 457         if (lock->l_resource->lr_namespace->ns_client) {
 458                 CERROR("Trying to cancel local lock\n");
 459                 LBUG();
 460         }
 461         LDLM_DEBUG(lock, "client-side local convert");
 462
 463         res = ldlm_lock_convert(lock, new_mode, flags);
 464         if (res) {
 465                 ldlm_reprocess_all(res);
 466                 rc = 0;
 467         } else {
 468                 rc = EDEADLOCK;
 469         }
 470         LDLM_DEBUG(lock, "client-side local convert handler END");
 471         LDLM_LOCK_PUT(lock);
 472         RETURN(rc);
 473 }
 474
 475 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
 476  * conversion of locks which are on the waiting or converting queue */
 477 /* Caller of this code is supposed to take care of lock readers/writers
 478    accounting */
 479 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, int *flags)
 480 {
 481         struct ldlm_request *body;
 482         struct ldlm_reply *reply;
 483         struct ldlm_lock *lock;
 484         struct ldlm_resource *res;
 485         struct ptlrpc_request *req = NULL;
 486         int rc, size = sizeof(*body);
 487         ENTRY;
 488
 489         lock = ldlm_handle2lock(lockh);
 490         if (!lock) {
 491                 LBUG();
 492                 RETURN(-EINVAL);
 493         }
 494         *flags = 0;
 495
 496         if (lock->l_conn_export == NULL)
 497                 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
 498
 499         LDLM_DEBUG(lock, "client-side convert");
 500
 501         req = ptlrpc_prep_req(class_exp2cliimp(lock->l_conn_export),
 502                               LUSTRE_DLM_VERSION, LDLM_CONVERT, 1, &size, NULL);
 503         if (!req)
 504                 GOTO(out, rc = -ENOMEM);
 505
 506         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 507         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 508                sizeof(body->lock_handle1));
 509
 510         body->lock_desc.l_req_mode = new_mode;
 511         body->lock_flags = *flags;
 512
 513         size = sizeof(*reply);
 514         req->rq_replen = lustre_msg_size(1, &size);
 515
 516         rc = ptlrpc_queue_wait(req);
 517         if (rc != ELDLM_OK)
 518                 GOTO(out, rc);
 519
 520         reply = lustre_swab_repbuf(req, 0, sizeof (*reply),
 521                                    lustre_swab_ldlm_reply);
 522         if (reply == NULL) {
 523                 CERROR ("Can't unpack ldlm_reply\n");
 524                 GOTO (out, rc = -EPROTO);
 525         }
 526
 527         if (req->rq_status)
 528                 GOTO(out, rc = req->rq_status);
 529
 530         res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
 531         if (res != NULL) {
 532                 ldlm_reprocess_all(res);
 533                 /* Go to sleep until the lock is granted. */
 534                 /* FIXME: or cancelled. */
 535                 if (lock->l_completion_ast) {
 536                         rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
 537                                                     NULL);
 538                         if (rc)
 539                                 GOTO(out, rc);
 540                 }
 541         } else {
 542                 rc = EDEADLOCK;
 543         }
 544         EXIT;
 545  out:
 546         LDLM_LOCK_PUT(lock);
 547         ptlrpc_req_finished(req);
 548         return rc;
 549 }
 550
 551 int ldlm_cli_cancel(struct lustre_handle *lockh)
 552 {
 553         struct ptlrpc_request *req;
 554         struct ldlm_lock *lock;
 555         struct ldlm_request *body;
 556         int rc = 0, size = sizeof(*body);
 557         ENTRY;
 558
 559         /* concurrent cancels on the same handle can happen */
 560         lock = __ldlm_handle2lock(lockh, LDLM_FL_CANCELING);
 561         if (lock == NULL)
 562                 RETURN(0);
 563
 564         if (lock->l_conn_export) {
 565                 int local_only;
 566                 struct obd_import *imp;
 567
 568                 LDLM_DEBUG(lock, "client-side cancel");
 569                 /* Set this flag to prevent others from getting new references*/
 570                 l_lock(&lock->l_resource->lr_namespace->ns_lock);
 571                 lock->l_flags |= LDLM_FL_CBPENDING;
 572                 local_only = lock->l_flags & LDLM_FL_LOCAL_ONLY;
 573                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 574                 ldlm_cancel_callback(lock);
 575
 576                 if (local_only) {
 577                         CDEBUG(D_INFO, "not sending request (at caller's "
 578                                "instruction)\n");
 579                         goto local_cancel;
 580                 }
 581
 582         restart:
 583                 imp = class_exp2cliimp(lock->l_conn_export);
 584                 if (imp == NULL || imp->imp_invalid) {
 585                         CDEBUG(D_HA, "skipping cancel on invalid import %p\n",
 586                                imp);
 587                         goto local_cancel;
 588                 }
 589
 590                 req = ptlrpc_prep_req(imp, LUSTRE_DLM_VERSION, LDLM_CANCEL,
 591                                       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                         char str[PTL_NALFMT_SIZE];
 610                         CERROR("client/server (nid %s) out of sync"
 611                                " -- not fatal\n",
 612                                ptlrpc_peernid2str(&req->rq_import->
 613                                                   imp_connection->c_peer, str));
 614                 } else if (rc == -ETIMEDOUT) {
 615                         ptlrpc_req_finished(req);
 616                         GOTO(restart, rc);
 617                 } else if (rc != ELDLM_OK) {
 618                         CERROR("Got rc %d from cancel RPC: canceling "
 619                                "anyway\n", rc);
 620                 }
 621
 622                 ptlrpc_req_finished(req);
 623         local_cancel:
 624                 ldlm_lock_cancel(lock);
 625         } else {
 626                 if (lock->l_resource->lr_namespace->ns_client) {
 627                         LDLM_ERROR(lock, "Trying to cancel local lock\n");
 628                         LBUG();
 629                 }
 630                 LDLM_DEBUG(lock, "client-side local cancel");
 631                 ldlm_lock_cancel(lock);
 632                 ldlm_reprocess_all(lock->l_resource);
 633                 LDLM_DEBUG(lock, "client-side local cancel handler END");
 634         }
 635
 636         EXIT;
 637  out:
 638         LDLM_LOCK_PUT(lock);
 639         return rc;
 640 }
 641
 642 /* when called with LDLM_ASYNC the blocking callback will be handled
 643  * in a thread and this function will return after the thread has been
 644  * asked to call the callback.  when called with LDLM_SYNC the blocking
 645  * callback will be performed in this function. */
 646 int ldlm_cancel_lru(struct ldlm_namespace *ns, ldlm_sync_t sync)
 647 {
 648         struct list_head *tmp, *next;
 649         struct ldlm_lock *lock;
 650         int count, rc = 0;
 651         LIST_HEAD(cblist);
 652         ENTRY;
 653
 654 #ifndef __KERNEL__
 655         sync = LDLM_SYNC; /* force to be sync in user space */
 656 #endif
 657
 658         l_lock(&ns->ns_lock);
 659         count = ns->ns_nr_unused - ns->ns_max_unused;
 660
 661         if (count <= 0) {
 662                 l_unlock(&ns->ns_lock);
 663                 RETURN(0);
 664         }
 665
 666         list_for_each_safe(tmp, next, &ns->ns_unused_list) {
 667
 668                 lock = list_entry(tmp, struct ldlm_lock, l_lru);
 669
 670                 LASSERT(!lock->l_readers && !lock->l_writers);
 671
 672                 /* Setting the CBPENDING flag is a little misleading, but
 673                  * prevents an important race; namely, once CBPENDING is set,
 674                  * the lock can accumulate no more readers/writers.  Since
 675                  * readers and writers are already zero here, ldlm_lock_decref
 676                  * won't see this flag and call l_blocking_ast */
 677                 lock->l_flags |= LDLM_FL_CBPENDING;
 678
 679                 LDLM_LOCK_GET(lock); /* dropped by bl thread */
 680                 ldlm_lock_remove_from_lru(lock);
 681                 if (sync != LDLM_ASYNC || ldlm_bl_to_thread(ns, NULL, lock))
 682                         list_add(&lock->l_lru, &cblist);
 683
 684                 if (--count == 0)
 685                         break;
 686         }
 687         l_unlock(&ns->ns_lock);
 688
 689         list_for_each_safe(tmp, next, &cblist) {
 690                 lock = list_entry(tmp, struct ldlm_lock, l_lru);
 691                 list_del_init(&lock->l_lru);
 692                 ldlm_handle_bl_callback(ns, NULL, lock);
 693         }
 694         RETURN(rc);
 695 }
 696
 697 static int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
 698                                            struct ldlm_res_id res_id, int flags,
 699                                            void *opaque)
 700 {
 701         struct ldlm_resource *res;
 702         struct list_head *tmp, *next, list = LIST_HEAD_INIT(list);
 703         struct ldlm_ast_work *w;
 704         ENTRY;
 705
 706         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
 707         if (res == NULL) {
 708                 /* This is not a problem. */
 709                 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id.name[0]);
 710                 RETURN(0);
 711         }
 712
 713         l_lock(&ns->ns_lock);
 714         list_for_each(tmp, &res->lr_granted) {
 715                 struct ldlm_lock *lock;
 716                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 717
 718                 if (opaque != NULL && lock->l_ast_data != opaque) {
 719                         LDLM_ERROR(lock, "data %p doesn't match opaque %p",
 720                                    lock->l_ast_data, opaque);
 721                         continue;
 722                 }
 723
 724                 if (lock->l_readers || lock->l_writers) {
 725                         if (flags & LDLM_FL_CONFIG_CHANGE)
 726                                 lock->l_flags |= LDLM_FL_CBPENDING;
 727                         else if (flags & LDLM_FL_WARN)
 728                                 LDLM_ERROR(lock, "lock in use");
 729                         continue;
 730                 }
 731
 732                 /* See CBPENDING comment in ldlm_cancel_lru */
 733                 lock->l_flags |= LDLM_FL_CBPENDING;
 734
 735                 OBD_ALLOC(w, sizeof(*w));
 736                 LASSERT(w);
 737
 738                 w->w_lock = LDLM_LOCK_GET(lock);
 739
 740                 list_add(&w->w_list, &list);
 741         }
 742         l_unlock(&ns->ns_lock);
 743
 744         list_for_each_safe(tmp, next, &list) {
 745                 struct lustre_handle lockh;
 746                 int rc;
 747                 w = list_entry(tmp, struct ldlm_ast_work, w_list);
 748
 749                 if (flags & LDLM_FL_LOCAL_ONLY) {
 750                         ldlm_lock_cancel(w->w_lock);
 751                 } else {
 752                         ldlm_lock2handle(w->w_lock, &lockh);
 753                         rc = ldlm_cli_cancel(&lockh);
 754                         if (rc != ELDLM_OK)
 755                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 756                 }
 757                 list_del(&w->w_list);
 758                 LDLM_LOCK_PUT(w->w_lock);
 759                 OBD_FREE(w, sizeof(*w));
 760         }
 761
 762         ldlm_resource_putref(res);
 763
 764         RETURN(0);
 765 }
 766
 767 static inline int have_no_nsresource(struct ldlm_namespace *ns)
 768 {
 769         int no_resource = 0;
 770
 771         spin_lock(&ns->ns_counter_lock);
 772         if (ns->ns_resources == 0)
 773                 no_resource = 1;
 774         spin_unlock(&ns->ns_counter_lock);
 775
 776         RETURN(no_resource);
 777 }
 778
 779 /* Cancel all locks on a namespace (or a specific resource, if given)
 780  * that have 0 readers/writers.
 781  *
 782  * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
 783  * to notify the server.
 784  * If flags & LDLM_FL_NO_CALLBACK, don't run the cancel callback.
 785  * If flags & LDLM_FL_WARN, print a warning if some locks are still in use.
 786  * If flags & LDLM_FL_CONFIG_CHANGE, mark all locks as having a pending callback
 787  */
 788 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
 789                            struct ldlm_res_id *res_id, int flags, void *opaque)
 790 {
 791         int i;
 792         struct l_wait_info lwi = { 0 };
 793         ENTRY;
 794
 795         if (ns == NULL)
 796                 RETURN(ELDLM_OK);
 797
 798         if (res_id)
 799                 RETURN(ldlm_cli_cancel_unused_resource(ns, *res_id, flags,
 800                                                        opaque));
 801
 802         l_lock(&ns->ns_lock);
 803         for (i = 0; i < RES_HASH_SIZE; i++) {
 804                 struct list_head *tmp, *next;
 805                 list_for_each_safe(tmp, next, &(ns->ns_hash[i])) {
 806                         int rc;
 807                         struct ldlm_resource *res;
 808                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
 809                         ldlm_resource_getref(res);
 810                         l_unlock(&ns->ns_lock);
 811
 812                         rc = ldlm_cli_cancel_unused_resource(ns, res->lr_name,
 813                                                              flags, opaque);
 814                         if (rc)
 815                                 CERROR("cancel_unused_res ("LPU64"): %d\n",
 816                                        res->lr_name.name[0], rc);
 817
 818                         l_lock(&ns->ns_lock);
 819                         next = tmp->next;
 820                         ldlm_resource_putref(res);
 821                 }
 822         }
 823         l_unlock(&ns->ns_lock);
 824         if (flags & LDLM_FL_CONFIG_CHANGE)
 825                 l_wait_event(ns->ns_waitq, have_no_nsresource(ns), &lwi);
 826
 827         RETURN(ELDLM_OK);
 828 }
 829
 830 /* Lock iterators. */
 831
 832 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
 833                           void *closure)
 834 {
 835         struct list_head *tmp, *next;
 836         struct ldlm_lock *lock;
 837         int rc = LDLM_ITER_CONTINUE;
 838         struct ldlm_namespace *ns = res->lr_namespace;
 839
 840         ENTRY;
 841
 842         if (!res)
 843                 RETURN(LDLM_ITER_CONTINUE);
 844
 845         l_lock(&ns->ns_lock);
 846         list_for_each_safe(tmp, next, &res->lr_granted) {
 847                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 848
 849                 if (iter(lock, closure) == LDLM_ITER_STOP)
 850                         GOTO(out, rc = LDLM_ITER_STOP);
 851         }
 852
 853         list_for_each_safe(tmp, next, &res->lr_converting) {
 854                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 855
 856                 if (iter(lock, closure) == LDLM_ITER_STOP)
 857                         GOTO(out, rc = LDLM_ITER_STOP);
 858         }
 859
 860         list_for_each_safe(tmp, next, &res->lr_waiting) {
 861                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 862
 863                 if (iter(lock, closure) == LDLM_ITER_STOP)
 864                         GOTO(out, rc = LDLM_ITER_STOP);
 865         }
 866  out:
 867         l_unlock(&ns->ns_lock);
 868         RETURN(rc);
 869 }
 870
 871 struct iter_helper_data {
 872         ldlm_iterator_t iter;
 873         void *closure;
 874 };
 875
 876 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
 877 {
 878         struct iter_helper_data *helper = closure;
 879         return helper->iter(lock, helper->closure);
 880 }
 881
 882 static int ldlm_res_iter_helper(struct ldlm_resource *res, void *closure)
 883 {
 884         return ldlm_resource_foreach(res, ldlm_iter_helper, closure);
 885 }
 886
 887 int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
 888                            void *closure)
 889 {
 890         struct iter_helper_data helper = { iter: iter, closure: closure };
 891         return ldlm_namespace_foreach_res(ns, ldlm_res_iter_helper, &helper);
 892 }
 893
 894 int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
 895                                ldlm_res_iterator_t iter, void *closure)
 896 {
 897         int i, rc = LDLM_ITER_CONTINUE;
 898
 899         l_lock(&ns->ns_lock);
 900         for (i = 0; i < RES_HASH_SIZE; i++) {
 901                 struct list_head *tmp, *next;
 902                 list_for_each_safe(tmp, next, &(ns->ns_hash[i])) {
 903                         struct ldlm_resource *res =
 904                                 list_entry(tmp, struct ldlm_resource, lr_hash);
 905
 906                         ldlm_resource_getref(res);
 907                         rc = iter(res, closure);
 908                         ldlm_resource_putref(res);
 909                         if (rc == LDLM_ITER_STOP)
 910                                 GOTO(out, rc);
 911                 }
 912         }
 913  out:
 914         l_unlock(&ns->ns_lock);
 915         RETURN(rc);
 916 }
 917
 918 /* non-blocking function to manipulate a lock whose cb_data is being put away.*/
 919 void ldlm_change_cbdata(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
 920                         ldlm_iterator_t iter, void *data)
 921 {
 922         struct ldlm_resource *res;
 923         ENTRY;
 924
 925         if (ns == NULL) {
 926                 CERROR("must pass in namespace\n");
 927                 LBUG();
 928         }
 929
 930         res = ldlm_resource_get(ns, NULL, *res_id, 0, 0);
 931         if (res == NULL) {
 932                 EXIT;
 933                 return;
 934         }
 935
 936         l_lock(&ns->ns_lock);
 937         ldlm_resource_foreach(res, iter, data);
 938         l_unlock(&ns->ns_lock);
 939         ldlm_resource_putref(res);
 940         EXIT;
 941 }
 942
 943 /* Lock replay */
 944
 945 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
 946 {
 947         struct list_head *list = closure;
 948
 949         /* we use l_pending_chain here, because it's unused on clients. */
 950         list_add(&lock->l_pending_chain, list);
 951         return LDLM_ITER_CONTINUE;
 952 }
 953
 954 static int replay_lock_interpret(struct ptlrpc_request *req,
 955                                     void * data, int rc)
 956 {
 957         struct ldlm_lock *lock;
 958         struct ldlm_reply *reply;
 959
 960         atomic_dec(&req->rq_import->imp_replay_inflight);
 961         if (rc != ELDLM_OK)
 962                 GOTO(out, rc);
 963
 964         lock = req->rq_async_args.pointer_arg[0];
 965         LASSERT(lock != NULL);
 966
 967         reply = lustre_swab_repbuf(req, 0, sizeof (*reply),
 968                                    lustre_swab_ldlm_reply);
 969         if (reply == NULL) {
 970                 CERROR("Can't unpack ldlm_reply\n");
 971                 GOTO (out, rc = -EPROTO);
 972         }
 973
 974         memcpy(&lock->l_remote_handle, &reply->lock_handle,
 975                sizeof(lock->l_remote_handle));
 976         LDLM_DEBUG(lock, "replayed lock:");
 977         ptlrpc_import_recovery_state_machine(req->rq_import);
 978 out:
 979         if (rc != ELDLM_OK)
 980                 ptlrpc_connect_import(req->rq_import, NULL);
 981
 982         RETURN(rc);
 983 }
 984
 985 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
 986 {
 987         struct ptlrpc_request *req;
 988         struct ldlm_request *body;
 989         struct ldlm_reply *reply;
 990         int buffers = 1;
 991         int size[2] = {0, sizeof(*body)};
 992         int flags;
 993
 994         /*
 995          * If granted mode matches the requested mode, this lock is granted.
 996          *
 997          * If they differ, but we have a granted mode, then we were granted
 998          * one mode and now want another: ergo, converting.
 999          *
1000          * If we haven't been granted anything and are on a resource list,
1001          * then we're blocked/waiting.
1002          *
1003          * If we haven't been granted anything and we're NOT on a resource list,
1004          * then we haven't got a reply yet and don't have a known disposition.
1005          * This happens whenever a lock enqueue is the request that triggers
1006          * recovery.
1007          */
1008         if (lock->l_granted_mode == lock->l_req_mode)
1009                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
1010         else if (lock->l_granted_mode)
1011                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
1012         else if (!list_empty(&lock->l_res_link))
1013                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
1014         else
1015                 flags = LDLM_FL_REPLAY;
1016
1017         req = ptlrpc_prep_req(imp, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
1018                               2, size, NULL);
1019         if (!req)
1020                 RETURN(-ENOMEM);
1021
1022         /* We're part of recovery, so don't wait for it. */
1023         req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
1024
1025         body = lustre_msg_buf(req->rq_reqmsg, 1, sizeof (*body));
1026         ldlm_lock2desc(lock, &body->lock_desc);
1027         body->lock_flags = flags;
1028
1029         ldlm_lock2handle(lock, &body->lock_handle1);
1030         size[0] = sizeof(*reply);
1031         if (lock->l_lvb_len != 0) {
1032                 buffers = 2;
1033                 size[1] = lock->l_lvb_len;
1034         }
1035         req->rq_replen = lustre_msg_size(buffers, size);
1036
1037         LDLM_DEBUG(lock, "replaying lock:");
1038
1039         atomic_inc(&req->rq_import->imp_replay_inflight);
1040         req->rq_async_args.pointer_arg[0] = lock;
1041         req->rq_interpret_reply = replay_lock_interpret;
1042         ptlrpcd_add_req(req);
1043
1044         RETURN(0);
1045 }
1046
1047 int ldlm_replay_locks(struct obd_import *imp)
1048 {
1049         struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
1050         struct list_head list, *pos, *next;
1051         struct ldlm_lock *lock;
1052         int rc = 0;
1053
1054         ENTRY;
1055         INIT_LIST_HEAD(&list);
1056
1057         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1058         LASSERT(ns != NULL);
1059
1060         /* ensure this doesn't fall to 0 before all have been queued */
1061         atomic_inc(&imp->imp_replay_inflight);
1062
1063         l_lock(&ns->ns_lock);
1064         (void)ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
1065
1066         list_for_each_safe(pos, next, &list) {
1067                 lock = list_entry(pos, struct ldlm_lock, l_pending_chain);
1068                 rc = replay_one_lock(imp, lock);
1069                 if (rc)
1070                         break; /* or try to do the rest? */
1071         }
1072         l_unlock(&ns->ns_lock);
1073
1074         atomic_dec(&imp->imp_replay_inflight);
1075
1076         RETURN(rc);
1077 }