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_device *obd = class_conn2obd(lock->l_connh);
  48
  49         if (obd == NULL) {
  50                 LDLM_ERROR(lock, "lock timed out; mot entering recovery in "
  51                            "server code, just going back to sleep");
  52         } else {
  53                 struct obd_import *imp = obd->u.cli.cl_import;
  54                 ptlrpc_fail_import(imp, lwd->lwd_generation);
  55                 LDLM_ERROR(lock, "lock timed out, entering recovery for %s@%s",
  56                            imp->imp_target_uuid.uuid,
  57                            imp->imp_connection->c_remote_uuid.uuid);
  58         }
  59
  60         RETURN(0);
  61 }
  62
  63 int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data)
  64 {
  65         /* XXX ALLOCATE - 160 mytes */
  66         struct lock_wait_data lwd;
  67         unsigned long irqflags;
  68         struct obd_device *obd;
  69         struct obd_import *imp = NULL;
  70         int rc = 0;
  71         struct l_wait_info lwi;
  72
  73         obd = class_conn2obd(lock->l_connh);
  74
  75         /* if this is a local lock, then there is no import */
  76         if (obd != NULL)
  77                 imp = obd->u.cli.cl_import;
  78
  79         lwd.lwd_lock = lock;
  80
  81         lwi = LWI_TIMEOUT_INTR(obd_timeout * HZ, ldlm_expired_completion_wait,
  82                                interrupted_completion_wait, &lwd);
  83         ENTRY;
  84
  85         if (flags == LDLM_FL_WAIT_NOREPROC)
  86                 goto noreproc;
  87
  88         if (flags == 0) {
  89                 wake_up(&lock->l_waitq);
  90                 RETURN(0);
  91         }
  92
  93         if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
  94                        LDLM_FL_BLOCK_CONV)))
  95                 RETURN(0);
  96
  97         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
  98                    "sleeping");
  99         ldlm_lock_dump(D_OTHER, lock, 0);
 100         ldlm_reprocess_all(lock->l_resource);
 101
 102  noreproc:
 103         if (imp != NULL) {
 104                 spin_lock_irqsave(&imp->imp_lock, irqflags);
 105                 lwd.lwd_generation = imp->imp_generation;
 106                 spin_unlock_irqrestore(&imp->imp_lock, irqflags);
 107         }
 108
 109         /* Go to sleep until the lock is granted or cancelled. */
 110         rc = l_wait_event(lock->l_waitq,
 111                           ((lock->l_req_mode == lock->l_granted_mode) ||
 112                            lock->l_destroyed), &lwi);
 113
 114         if (lock->l_destroyed) {
 115                 LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed");
 116                 RETURN(-EIO);
 117         }
 118
 119         if (rc) {
 120                 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
 121                            rc);
 122                 RETURN(rc);
 123         }
 124
 125         LDLM_DEBUG(lock, "client-side enqueue waking up: granted");
 126         RETURN(0);
 127 }
 128
 129 static int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
 130                                   struct lustre_handle *parent_lockh,
 131                                   struct ldlm_res_id res_id,
 132                                   __u32 type,
 133                                   void *cookie, int cookielen,
 134                                   ldlm_mode_t mode,
 135                                   int *flags,
 136                                   ldlm_completion_callback completion,
 137                                   ldlm_blocking_callback blocking,
 138                                   void *data,
 139                                   struct lustre_handle *lockh)
 140 {
 141         struct ldlm_lock *lock;
 142         int err;
 143         ENTRY;
 144
 145         if (ns->ns_client) {
 146                 CERROR("Trying to enqueue local lock in a shadow namespace\n");
 147                 LBUG();
 148         }
 149
 150         lock = ldlm_lock_create(ns, parent_lockh, res_id, type, mode,
 151                                 blocking, completion, data);
 152         if (!lock)
 153                 GOTO(out_nolock, err = -ENOMEM);
 154         LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
 155
 156         ldlm_lock_addref_internal(lock, mode);
 157         ldlm_lock2handle(lock, lockh);
 158         lock->l_flags |= LDLM_FL_LOCAL;
 159
 160         err = ldlm_lock_enqueue(ns, &lock, cookie, cookielen, flags);
 161         if (err != ELDLM_OK)
 162                 GOTO(out, err);
 163
 164         if (type == LDLM_EXTENT)
 165                 memcpy(cookie, &lock->l_extent, sizeof(lock->l_extent));
 166         if ((*flags) & LDLM_FL_LOCK_CHANGED)
 167                 memcpy(&res_id, &lock->l_resource->lr_name, sizeof(res_id));
 168
 169         LDLM_DEBUG_NOLOCK("client-side local enqueue handler END (lock %p)",
 170                           lock);
 171
 172         if (lock->l_completion_ast)
 173                 lock->l_completion_ast(lock, *flags, NULL);
 174
 175         LDLM_DEBUG(lock, "client-side local enqueue END");
 176         EXIT;
 177  out:
 178         LDLM_LOCK_PUT(lock);
 179  out_nolock:
 180         return err;
 181 }
 182
 183 int ldlm_cli_enqueue(struct lustre_handle *connh,
 184                      struct ptlrpc_request *req,
 185                      struct ldlm_namespace *ns,
 186                      struct lustre_handle *parent_lock_handle,
 187                      struct ldlm_res_id res_id,
 188                      __u32 type,
 189                      void *cookie, int cookielen,
 190                      ldlm_mode_t mode,
 191                      int *flags,
 192                      ldlm_completion_callback completion,
 193                      ldlm_blocking_callback blocking,
 194                      void *data,
 195                      struct lustre_handle *lockh)
 196 {
 197         struct ldlm_lock *lock;
 198         struct ldlm_request *body;
 199         struct ldlm_reply *reply;
 200         int rc, size = sizeof(*body), req_passed_in = 1, is_replay;
 201         ENTRY;
 202
 203         is_replay = *flags & LDLM_FL_REPLAY;
 204         LASSERT(connh != NULL || !is_replay);
 205
 206         if (connh == NULL) {
 207                 rc = ldlm_cli_enqueue_local(ns, parent_lock_handle, res_id,
 208                                             type, cookie, cookielen, mode,
 209                                             flags, completion, blocking, data,
 210                                             lockh);
 211                 RETURN(rc);
 212         }
 213
 214         /* If we're replaying this lock, just check some invariants.
 215          * If we're creating a new lock, get everything all setup nice. */
 216         if (is_replay) {
 217                 lock = ldlm_handle2lock(lockh);
 218                 LDLM_DEBUG(lock, "client-side enqueue START");
 219                 LASSERT(connh == lock->l_connh);
 220         } else {
 221                 lock = ldlm_lock_create(ns, parent_lock_handle, res_id, type,
 222                                         mode, blocking, completion, data);
 223                 if (lock == NULL)
 224                         GOTO(out_nolock, rc = -ENOMEM);
 225                 LDLM_DEBUG(lock, "client-side enqueue START");
 226                 /* for the local lock, add the reference */
 227                 ldlm_lock_addref_internal(lock, mode);
 228                 ldlm_lock2handle(lock, lockh);
 229                 if (type == LDLM_EXTENT)
 230                         memcpy(&lock->l_extent, cookie,
 231                                sizeof(body->lock_desc.l_extent));
 232         }
 233
 234         if (req == NULL) {
 235                 req = ptlrpc_prep_req(class_conn2cliimp(connh), LDLM_ENQUEUE, 1,
 236                                       &size, NULL);
 237                 if (!req)
 238                         GOTO(out, rc = -ENOMEM);
 239                 req_passed_in = 0;
 240         } else if (req->rq_reqmsg->buflens[0] != sizeof(*body))
 241                 LBUG();
 242
 243         /* Dump lock data into the request buffer */
 244         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 245         ldlm_lock2desc(lock, &body->lock_desc);
 246         body->lock_flags = *flags;
 247
 248         memcpy(&body->lock_handle1, lockh, sizeof(*lockh));
 249         if (parent_lock_handle)
 250                 memcpy(&body->lock_handle2, parent_lock_handle,
 251                        sizeof(body->lock_handle2));
 252
 253         /* Continue as normal. */
 254         if (!req_passed_in) {
 255                 size = sizeof(*reply);
 256                 req->rq_replen = lustre_msg_size(1, &size);
 257         }
 258         lock->l_connh = connh;
 259         lock->l_export = NULL;
 260         lock->l_blocking_ast = blocking;
 261
 262         LDLM_DEBUG(lock, "sending request");
 263         rc = ptlrpc_queue_wait(req);
 264
 265         if (rc != ELDLM_OK) {
 266                 LASSERT(!is_replay);
 267                 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
 268                            rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
 269                 /* Set a flag to prevent us from sending a CANCEL (bug 407) */
 270                 l_lock(&ns->ns_lock);
 271                 lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 272                 LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 273                 l_unlock(&ns->ns_lock);
 274
 275                 ldlm_lock_decref_and_cancel(lockh, mode);
 276
 277                 if (rc == ELDLM_LOCK_ABORTED) {
 278                         /* caller expects reply buffer 0 to have been swabbed */
 279                         reply = lustre_swab_repbuf(req, 0, sizeof (*reply),
 280                                                    lustre_swab_ldlm_reply);
 281                         if (reply == NULL) {
 282                                 CERROR ("Can't unpack ldlm_reply\n");
 283                                 GOTO (out_req, rc = -EPROTO);
 284                         }
 285                 }
 286                 GOTO(out_req, rc);
 287         }
 288
 289         reply = lustre_swab_repbuf(req, 0, sizeof (*reply),
 290                                    lustre_swab_ldlm_reply);
 291         if (reply == NULL) {
 292                 CERROR ("Can't unpack ldlm_reply\n");
 293                 GOTO (out_req, rc = -EPROTO);
 294         }
 295
 296         memcpy(&lock->l_remote_handle, &reply->lock_handle,
 297                sizeof(lock->l_remote_handle));
 298         *flags = reply->lock_flags;
 299
 300         CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: %d\n", lock,
 301                reply->lock_handle.cookie, *flags);
 302         if (type == LDLM_EXTENT) {
 303                 CDEBUG(D_INFO, "requested extent: "LPU64" -> "LPU64", got "
 304                        "extent "LPU64" -> "LPU64"\n",
 305                        body->lock_desc.l_extent.start,
 306                        body->lock_desc.l_extent.end,
 307                        reply->lock_extent.start, reply->lock_extent.end);
 308
 309                 cookie = &reply->lock_extent; /* FIXME bug 267 */
 310                 cookielen = sizeof(reply->lock_extent);
 311         }
 312
 313         /* If enqueue returned a blocked lock but the completion handler has
 314          * already run, then it fixed up the resource and we don't need to do it
 315          * again. */
 316         if ((*flags) & LDLM_FL_LOCK_CHANGED) {
 317                 int newmode = reply->lock_mode;
 318                 LASSERT(!is_replay);
 319                 if (newmode && newmode != lock->l_req_mode) {
 320                         LDLM_DEBUG(lock, "server returned different mode %s",
 321                                    ldlm_lockname[newmode]);
 322                         lock->l_req_mode = newmode;
 323                 }
 324
 325                 if (reply->lock_resource_name.name[0] !=
 326                     lock->l_resource->lr_name.name[0]) {
 327                         CDEBUG(D_INFO, "remote intent success, locking %ld "
 328                                "instead of %ld\n",
 329                                (long)reply->lock_resource_name.name[0],
 330                                (long)lock->l_resource->lr_name.name[0]);
 331
 332                         ldlm_lock_change_resource(ns, lock,
 333                                                   reply->lock_resource_name);
 334                         if (lock->l_resource == NULL) {
 335                                 LBUG();
 336                                 GOTO(out_req, rc = -ENOMEM);
 337                         }
 338                         LDLM_DEBUG(lock, "client-side enqueue, new resource");
 339                 }
 340         }
 341
 342         if (!is_replay) {
 343                 rc = ldlm_lock_enqueue(ns, &lock, cookie, cookielen, flags);
 344                 if (lock->l_completion_ast)
 345                         lock->l_completion_ast(lock, *flags, NULL);
 346         }
 347
 348         LDLM_DEBUG(lock, "client-side enqueue END");
 349         EXIT;
 350  out_req:
 351         if (!req_passed_in)
 352                 ptlrpc_req_finished(req);
 353  out:
 354         LDLM_LOCK_PUT(lock);
 355  out_nolock:
 356         return rc;
 357 }
 358
 359 int ldlm_cli_replay_enqueue(struct ldlm_lock *lock)
 360 {
 361         struct lustre_handle lockh;
 362         struct ldlm_res_id junk;
 363         int flags = LDLM_FL_REPLAY;
 364         ldlm_lock2handle(lock, &lockh);
 365         return ldlm_cli_enqueue(lock->l_connh, NULL, NULL, NULL, junk,
 366                                 lock->l_resource->lr_type, NULL, 0, -1, &flags,
 367                                 NULL, NULL, NULL, &lockh);
 368 }
 369
 370 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
 371                                   int *flags)
 372 {
 373         ENTRY;
 374         if (lock->l_resource->lr_namespace->ns_client) {
 375                 CERROR("Trying to cancel local lock\n");
 376                 LBUG();
 377         }
 378         LDLM_DEBUG(lock, "client-side local convert");
 379
 380         ldlm_lock_convert(lock, new_mode, flags);
 381         ldlm_reprocess_all(lock->l_resource);
 382
 383         LDLM_DEBUG(lock, "client-side local convert handler END");
 384         LDLM_LOCK_PUT(lock);
 385         RETURN(0);
 386 }
 387
 388 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
 389  * conversion of locks which are on the waiting or converting queue */
 390 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, int *flags)
 391 {
 392         struct ldlm_request *body;
 393         struct lustre_handle *connh;
 394         struct ldlm_reply *reply;
 395         struct ldlm_lock *lock;
 396         struct ldlm_resource *res;
 397         struct ptlrpc_request *req;
 398         int rc, size = sizeof(*body);
 399         ENTRY;
 400
 401         lock = ldlm_handle2lock(lockh);
 402         if (!lock) {
 403                 LBUG();
 404                 RETURN(-EINVAL);
 405         }
 406         *flags = 0;
 407         connh = lock->l_connh;
 408
 409         if (!connh)
 410                 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
 411
 412         LDLM_DEBUG(lock, "client-side convert");
 413
 414         req = ptlrpc_prep_req(class_conn2cliimp(connh), LDLM_CONVERT, 1, &size,
 415                               NULL);
 416         if (!req)
 417                 GOTO(out, rc = -ENOMEM);
 418
 419         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 420         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 421                sizeof(body->lock_handle1));
 422
 423         body->lock_desc.l_req_mode = new_mode;
 424         body->lock_flags = *flags;
 425
 426         size = sizeof(*reply);
 427         req->rq_replen = lustre_msg_size(1, &size);
 428
 429         rc = ptlrpc_queue_wait(req);
 430         if (rc != ELDLM_OK)
 431                 GOTO(out, rc);
 432
 433         reply = lustre_swab_repbuf(req, 0, sizeof (*reply),
 434                                    lustre_swab_ldlm_reply);
 435         if (reply == NULL) {
 436                 CERROR ("Can't unpack ldlm_reply\n");
 437                 GOTO (out, rc = -EPROTO);
 438         }
 439
 440         res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
 441         if (res != NULL)
 442                 ldlm_reprocess_all(res);
 443         /* Go to sleep until the lock is granted. */
 444         /* FIXME: or cancelled. */
 445         if (lock->l_completion_ast)
 446                 lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC, NULL);
 447         EXIT;
 448  out:
 449         LDLM_LOCK_PUT(lock);
 450         ptlrpc_req_finished(req);
 451         return rc;
 452 }
 453
 454 int ldlm_cli_cancel(struct lustre_handle *lockh)
 455 {
 456         struct ptlrpc_request *req;
 457         struct ldlm_lock *lock;
 458         struct ldlm_request *body;
 459         int rc = 0, size = sizeof(*body);
 460         ENTRY;
 461
 462         /* concurrent cancels on the same handle can happen */
 463         lock = __ldlm_handle2lock(lockh, LDLM_FL_CANCELING);
 464         if (lock == NULL)
 465                 RETURN(0);
 466
 467         if (lock->l_connh) {
 468                 int local_only;
 469                 struct obd_import *imp;
 470
 471                 LDLM_DEBUG(lock, "client-side cancel");
 472                 /* Set this flag to prevent others from getting new references*/
 473                 l_lock(&lock->l_resource->lr_namespace->ns_lock);
 474                 lock->l_flags |= LDLM_FL_CBPENDING;
 475                 local_only = (lock->l_flags & LDLM_FL_LOCAL_ONLY);
 476                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 477                 ldlm_cancel_callback(lock);
 478
 479                 if (local_only) {
 480                         CDEBUG(D_INFO, "not sending request (at caller's "
 481                                "instruction)\n");
 482                         goto local_cancel;
 483                 }
 484
 485                 imp = class_conn2cliimp(lock->l_connh);
 486                 if (imp == NULL || imp->imp_invalid) {
 487                         CDEBUG(D_HA, "skipping cancel on invalid import %p\n",
 488                                imp);
 489                         goto local_cancel;
 490                 }
 491
 492                 req = ptlrpc_prep_req(imp, LDLM_CANCEL, 1, &size, NULL);
 493                 if (!req)
 494                         GOTO(out, rc = -ENOMEM);
 495
 496                 /* XXX FIXME bug 249 */
 497                 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
 498                 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
 499
 500                 body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 501                 memcpy(&body->lock_handle1, &lock->l_remote_handle,
 502                        sizeof(body->lock_handle1));
 503
 504                 req->rq_replen = lustre_msg_size(0, NULL);
 505
 506                 rc = ptlrpc_queue_wait(req);
 507
 508                 if (rc == ESTALE)
 509                         CERROR("client/server (nid "LPU64") out of sync--not "
 510                                "fatal\n",
 511                                req->rq_import->imp_connection->c_peer.peer_nid);
 512                 else if (rc != ELDLM_OK)
 513                         CERROR("Got rc %d from cancel RPC: canceling "
 514                                "anyway\n", rc);
 515
 516                 ptlrpc_req_finished(req);
 517         local_cancel:
 518                 ldlm_lock_cancel(lock);
 519         } else {
 520                 if (lock->l_resource->lr_namespace->ns_client) {
 521                         LDLM_ERROR(lock, "Trying to cancel local lock\n");
 522                         LBUG();
 523                 }
 524                 LDLM_DEBUG(lock, "client-side local cancel");
 525                 ldlm_lock_cancel(lock);
 526                 ldlm_reprocess_all(lock->l_resource);
 527                 LDLM_DEBUG(lock, "client-side local cancel handler END");
 528         }
 529
 530         EXIT;
 531  out:
 532         LDLM_LOCK_PUT(lock);
 533         return rc;
 534 }
 535
 536 int ldlm_cancel_lru(struct ldlm_namespace *ns)
 537 {
 538         struct list_head *tmp, *next, list = LIST_HEAD_INIT(list);
 539         int count, rc = 0;
 540         struct ldlm_ast_work *w;
 541         ENTRY;
 542
 543         l_lock(&ns->ns_lock);
 544         count = ns->ns_nr_unused - ns->ns_max_unused;
 545
 546         if (count <= 0) {
 547                 l_unlock(&ns->ns_lock);
 548                 RETURN(0);
 549         }
 550
 551         list_for_each_safe(tmp, next, &ns->ns_unused_list) {
 552                 struct ldlm_lock *lock;
 553                 lock = list_entry(tmp, struct ldlm_lock, l_lru);
 554
 555                 LASSERT(!lock->l_readers && !lock->l_writers);
 556
 557                 /* Setting the CBPENDING flag is a little misleading, but
 558                  * prevents an important race; namely, once CBPENDING is set,
 559                  * the lock can accumulate no more readers/writers.  Since
 560                  * readers and writers are already zero here, ldlm_lock_decref
 561                  * won't see this flag and call l_blocking_ast */
 562                 lock->l_flags |= LDLM_FL_CBPENDING;
 563
 564                 OBD_ALLOC(w, sizeof(*w));
 565                 LASSERT(w);
 566
 567                 w->w_lock = LDLM_LOCK_GET(lock);
 568                 list_add(&w->w_list, &list);
 569                 ldlm_lock_remove_from_lru(lock);
 570
 571                 if (--count == 0)
 572                         break;
 573         }
 574         l_unlock(&ns->ns_lock);
 575
 576         list_for_each_safe(tmp, next, &list) {
 577                 struct lustre_handle lockh;
 578                 int rc;
 579                 w = list_entry(tmp, struct ldlm_ast_work, w_list);
 580
 581                 ldlm_lock2handle(w->w_lock, &lockh);
 582                 rc = ldlm_cli_cancel(&lockh);
 583                 if (rc != ELDLM_OK)
 584                         CDEBUG(D_INFO, "ldlm_cli_cancel: %d\n", rc);
 585
 586                 list_del(&w->w_list);
 587                 LDLM_LOCK_PUT(w->w_lock);
 588                 OBD_FREE(w, sizeof(*w));
 589         }
 590
 591         RETURN(rc);
 592 }
 593
 594 static int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
 595                                            struct ldlm_res_id res_id, int flags,
 596                                            void *opaque)
 597 {
 598         struct ldlm_resource *res;
 599         struct list_head *tmp, *next, list = LIST_HEAD_INIT(list);
 600         struct ldlm_ast_work *w;
 601         ENTRY;
 602
 603         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
 604         if (res == NULL) {
 605                 /* This is not a problem. */
 606                 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id.name[0]);
 607                 RETURN(0);
 608         }
 609
 610         l_lock(&ns->ns_lock);
 611         list_for_each(tmp, &res->lr_granted) {
 612                 struct ldlm_lock *lock;
 613                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 614
 615                 if (opaque != NULL && lock->l_data != opaque) {
 616                         LDLM_ERROR(lock, "data %p doesn't match opaque %p",
 617                                    lock->l_data, opaque);
 618                         //LBUG();
 619                         continue;
 620                 }
 621
 622                 if (lock->l_readers || lock->l_writers) {
 623                         if (flags & LDLM_FL_WARN) {
 624                                 LDLM_ERROR(lock, "lock in use");
 625                                 //LBUG();
 626                         }
 627                         continue;
 628                 }
 629
 630                 /* See CBPENDING comment in ldlm_cancel_lru */
 631                 lock->l_flags |= LDLM_FL_CBPENDING;
 632
 633                 OBD_ALLOC(w, sizeof(*w));
 634                 LASSERT(w);
 635
 636                 w->w_lock = LDLM_LOCK_GET(lock);
 637
 638                 /* Prevent the cancel callback from being called by setting
 639                  * LDLM_FL_CANCEL in the lock.  Very sneaky. -p */
 640                 if (flags & LDLM_FL_NO_CALLBACK)
 641                         w->w_lock->l_flags |= LDLM_FL_CANCEL;
 642
 643                 list_add(&w->w_list, &list);
 644         }
 645         l_unlock(&ns->ns_lock);
 646
 647         list_for_each_safe(tmp, next, &list) {
 648                 struct lustre_handle lockh;
 649                 int rc;
 650                 w = list_entry(tmp, struct ldlm_ast_work, w_list);
 651
 652                 if (flags & LDLM_FL_LOCAL_ONLY) {
 653                         ldlm_lock_cancel(w->w_lock);
 654                 } else {
 655                         ldlm_lock2handle(w->w_lock, &lockh);
 656                         rc = ldlm_cli_cancel(&lockh);
 657                         if (rc != ELDLM_OK)
 658                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 659                 }
 660                 list_del(&w->w_list);
 661                 LDLM_LOCK_PUT(w->w_lock);
 662                 OBD_FREE(w, sizeof(*w));
 663         }
 664
 665         ldlm_resource_putref(res);
 666
 667         RETURN(0);
 668 }
 669
 670 /* Cancel all locks on a namespace (or a specific resource, if given)
 671  * that have 0 readers/writers.
 672  *
 673  * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
 674  * to notify the server.
 675  * If flags & LDLM_FL_NO_CALLBACK, don't run the cancel callback.
 676  * If flags & LDLM_FL_WARN, print a warning if some locks are still in use. */
 677 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
 678                            struct ldlm_res_id *res_id, int flags, void *opaque)
 679 {
 680         int i;
 681         ENTRY;
 682
 683         if (ns == NULL)
 684                 RETURN(ELDLM_OK);
 685
 686         if (res_id)
 687                 RETURN(ldlm_cli_cancel_unused_resource(ns, *res_id, flags,
 688                                                        opaque));
 689
 690         l_lock(&ns->ns_lock);
 691         for (i = 0; i < RES_HASH_SIZE; i++) {
 692                 struct list_head *tmp, *pos;
 693                 list_for_each_safe(tmp, pos, &(ns->ns_hash[i])) {
 694                         int rc;
 695                         struct ldlm_resource *res;
 696                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
 697                         ldlm_resource_getref(res);
 698
 699                         rc = ldlm_cli_cancel_unused_resource(ns, res->lr_name,
 700                                                              flags, opaque);
 701
 702                         if (rc)
 703                                 CERROR("cancel_unused_res ("LPU64"): %d\n",
 704                                        res->lr_name.name[0], rc);
 705                         ldlm_resource_putref(res);
 706                 }
 707         }
 708         l_unlock(&ns->ns_lock);
 709
 710         RETURN(ELDLM_OK);
 711 }
 712
 713 /* Lock iterators. */
 714
 715 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
 716                           void *closure)
 717 {
 718         struct list_head *tmp, *next;
 719         struct ldlm_lock *lock;
 720         int rc = LDLM_ITER_CONTINUE;
 721         struct ldlm_namespace *ns = res->lr_namespace;
 722
 723         ENTRY;
 724
 725         if (!res)
 726                 RETURN(LDLM_ITER_CONTINUE);
 727
 728         l_lock(&ns->ns_lock);
 729         list_for_each_safe(tmp, next, &res->lr_granted) {
 730                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 731
 732                 if (iter(lock, closure) == LDLM_ITER_STOP)
 733                         GOTO(out, rc = LDLM_ITER_STOP);
 734         }
 735
 736         list_for_each_safe(tmp, next, &res->lr_converting) {
 737                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 738
 739                 if (iter(lock, closure) == LDLM_ITER_STOP)
 740                         GOTO(out, rc = LDLM_ITER_STOP);
 741         }
 742
 743         list_for_each_safe(tmp, next, &res->lr_waiting) {
 744                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 745
 746                 if (iter(lock, closure) == LDLM_ITER_STOP)
 747                         GOTO(out, rc = LDLM_ITER_STOP);
 748         }
 749  out:
 750         l_unlock(&ns->ns_lock);
 751         RETURN(rc);
 752 }
 753
 754 struct iter_helper_data {
 755         ldlm_iterator_t iter;
 756         void *closure;
 757 };
 758
 759 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
 760 {
 761         struct iter_helper_data *helper = closure;
 762         return helper->iter(lock, helper->closure);
 763 }
 764
 765 static int ldlm_res_iter_helper(struct ldlm_resource *res, void *closure)
 766 {
 767         return ldlm_resource_foreach(res, ldlm_iter_helper, closure);
 768 }
 769
 770 int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
 771                            void *closure)
 772 {
 773         struct iter_helper_data helper = { iter: iter, closure: closure };
 774         return ldlm_namespace_foreach_res(ns, ldlm_res_iter_helper, &helper);
 775 }
 776
 777 int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
 778                                ldlm_res_iterator_t iter, void *closure)
 779 {
 780         int i, rc = LDLM_ITER_CONTINUE;
 781
 782         l_lock(&ns->ns_lock);
 783         for (i = 0; i < RES_HASH_SIZE; i++) {
 784                 struct list_head *tmp, *next;
 785                 list_for_each_safe(tmp, next, &(ns->ns_hash[i])) {
 786                         struct ldlm_resource *res =
 787                                 list_entry(tmp, struct ldlm_resource, lr_hash);
 788
 789                         ldlm_resource_getref(res);
 790                         rc = iter(res, closure);
 791                         ldlm_resource_putref(res);
 792                         if (rc == LDLM_ITER_STOP)
 793                                 GOTO(out, rc);
 794                 }
 795         }
 796  out:
 797         l_unlock(&ns->ns_lock);
 798         RETURN(rc);
 799 }
 800
 801 /* non-blocking function to manipulate a lock whose cb_data is being put away.*/
 802 void ldlm_change_cbdata(struct ldlm_namespace *ns,
 803                        struct ldlm_res_id *res_id,
 804                        ldlm_iterator_t iter,
 805                        void *data)
 806 {
 807         struct ldlm_resource *res;
 808         ENTRY;
 809
 810         if (ns == NULL) {
 811                 CERROR("must pass in namespace");
 812                 LBUG();
 813         }
 814
 815         res = ldlm_resource_get(ns, NULL, *res_id, 0, 0);
 816         if (res == NULL) {
 817                 EXIT;
 818                 return;
 819         }
 820
 821         l_lock(&ns->ns_lock);
 822         ldlm_resource_foreach(res, iter, data);
 823         l_unlock(&ns->ns_lock);
 824         ldlm_resource_putref(res);
 825         EXIT;
 826 }
 827
 828 /* Lock replay */
 829
 830 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
 831 {
 832         struct list_head *list = closure;
 833
 834         /* we use l_pending_chain here, because it's unused on clients. */
 835         list_add(&lock->l_pending_chain, list);
 836         return LDLM_ITER_CONTINUE;
 837 }
 838
 839 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
 840 {
 841         struct ptlrpc_request *req;
 842         struct ldlm_request *body;
 843         struct ldlm_reply *reply;
 844         int rc, size;
 845         int flags;
 846
 847         /*
 848          * If granted mode matches the requested mode, this lock is granted.
 849          *
 850          * If they differ, but we have a granted mode, then we were granted
 851          * one mode and now want another: ergo, converting.
 852          *
 853          * If we haven't been granted anything and are on a resource list,
 854          * then we're blocked/waiting.
 855          *
 856          * If we haven't been granted anything and we're NOT on a resource list,
 857          * then we haven't got a reply yet and don't have a known disposition.
 858          * This happens whenever a lock enqueue is the request that triggers
 859          * recovery.
 860          */
 861         if (lock->l_granted_mode == lock->l_req_mode)
 862                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
 863         else if (lock->l_granted_mode)
 864                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
 865         else if (!list_empty(&lock->l_res_link))
 866                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
 867         else
 868                 flags = LDLM_FL_REPLAY;
 869
 870         size = sizeof(*body);
 871         req = ptlrpc_prep_req(imp, LDLM_ENQUEUE, 1, &size, NULL);
 872         if (!req)
 873                 RETURN(-ENOMEM);
 874
 875         /* We're part of recovery, so don't wait for it. */
 876         req->rq_level = LUSTRE_CONN_RECOVER;
 877
 878         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 879         ldlm_lock2desc(lock, &body->lock_desc);
 880         body->lock_flags = flags;
 881
 882         ldlm_lock2handle(lock, &body->lock_handle1);
 883         size = sizeof(*reply);
 884         req->rq_replen = lustre_msg_size(1, &size);
 885
 886         LDLM_DEBUG(lock, "replaying lock:");
 887         rc = ptlrpc_queue_wait(req);
 888         if (rc != ELDLM_OK)
 889                 GOTO(out, rc);
 890
 891         reply = lustre_swab_repbuf(req, 0, sizeof (*reply),
 892                                    lustre_swab_ldlm_reply);
 893         if (reply == NULL) {
 894                 CERROR("Can't unpack ldlm_reply\n");
 895                 GOTO (out, rc = -EPROTO);
 896         }
 897
 898         memcpy(&lock->l_remote_handle, &reply->lock_handle,
 899                sizeof(lock->l_remote_handle));
 900         LDLM_DEBUG(lock, "replayed lock:");
 901  out:
 902         ptlrpc_req_finished(req);
 903         RETURN(rc);
 904 }
 905
 906 int ldlm_replay_locks(struct obd_import *imp)
 907 {
 908         struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
 909         struct list_head list, *pos, *next;
 910         struct ldlm_lock *lock;
 911         int rc = 0;
 912
 913         ENTRY;
 914         INIT_LIST_HEAD(&list);
 915
 916         l_lock(&ns->ns_lock);
 917         (void)ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
 918
 919         list_for_each_safe(pos, next, &list) {
 920                 lock = list_entry(pos, struct ldlm_lock, l_pending_chain);
 921                 rc = replay_one_lock(imp, lock);
 922                 if (rc)
 923                         break; /* or try to do the rest? */
 924         }
 925         l_unlock(&ns->ns_lock);
 926         RETURN(rc);
 927 }