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                 if (ptlrpc_check_suspend())
  54                         RETURN(0);
  55
  56                 LDLM_ERROR(lock, "lock timed out; not entering recovery in "
  57                            "server code, just going back to sleep");
  58                 if (time_after(jiffies, next_dump)) {
  59                         last_dump = next_dump;
  60                         next_dump = jiffies + 300 * HZ;
  61                         ldlm_namespace_dump(D_DLMTRACE,
  62                                             lock->l_resource->lr_namespace);
  63                         if (last_dump == 0)
  64                                 portals_debug_dumplog();
  65                 }
  66                 RETURN(0);
  67         }
  68
  69         obd = lock->l_conn_export->exp_obd;
  70         imp = obd->u.cli.cl_import;
  71         ptlrpc_fail_import(imp, lwd->lwd_generation);
  72         LDLM_ERROR(lock, "lock timed out, entering recovery for %s@%s",
  73                    imp->imp_target_uuid.uuid,
  74                    imp->imp_connection->c_remote_uuid.uuid);
  75
  76         RETURN(0);
  77 }
  78
  79 int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data)
  80 {
  81         /* XXX ALLOCATE - 160 bytes */
  82         struct lock_wait_data lwd;
  83         unsigned long irqflags;
  84         struct obd_device *obd;
  85         struct obd_import *imp = NULL;
  86         struct l_wait_info lwi;
  87         int rc = 0;
  88         ENTRY;
  89
  90         if (flags == LDLM_FL_WAIT_NOREPROC)
  91                 goto noreproc;
  92
  93         if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
  94                        LDLM_FL_BLOCK_CONV))) {
  95                 wake_up(&lock->l_waitq);
  96                 RETURN(0);
  97         }
  98
  99         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
 100                    "sleeping");
 101         ldlm_lock_dump(D_OTHER, lock, 0);
 102         ldlm_reprocess_all(lock->l_resource);
 103
 104 noreproc:
 105
 106         obd = class_exp2obd(lock->l_conn_export);
 107
 108         /* if this is a local lock, then there is no import */
 109         if (obd != NULL)
 110                 imp = obd->u.cli.cl_import;
 111
 112         lwd.lwd_lock = lock;
 113
 114         if (flags & LDLM_FL_NO_TIMEOUT) {
 115                 LDLM_DEBUG(lock, "waiting indefinitely for group lock\n");
 116                 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
 117         } else {
 118                 lwi = LWI_TIMEOUT_INTR(obd_timeout * HZ,
 119                                        ldlm_expired_completion_wait,
 120                                        interrupted_completion_wait, &lwd);
 121         }
 122
 123         if (imp != NULL) {
 124                 spin_lock_irqsave(&imp->imp_lock, irqflags);
 125                 lwd.lwd_generation = imp->imp_generation;
 126                 spin_unlock_irqrestore(&imp->imp_lock, irqflags);
 127         }
 128
 129         /* Go to sleep until the lock is granted or cancelled. */
 130         rc = l_wait_event(lock->l_waitq,
 131                           ((lock->l_req_mode == lock->l_granted_mode) ||
 132                            (lock->l_flags & LDLM_FL_FAILED)), &lwi);
 133
 134         if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED) {
 135                 LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed");
 136                 RETURN(-EIO);
 137         }
 138
 139         if (rc) {
 140                 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
 141                            rc);
 142                 RETURN(rc);
 143         }
 144
 145         LDLM_DEBUG(lock, "client-side enqueue waking up: granted");
 146         RETURN(0);
 147 }
 148
 149 static int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
 150                                   struct ldlm_res_id res_id,
 151                                   __u32 type,
 152                                   ldlm_policy_data_t *policy,
 153                                   ldlm_mode_t mode,
 154                                   int *flags,
 155                                   ldlm_blocking_callback blocking,
 156                                   ldlm_completion_callback completion,
 157                                   ldlm_glimpse_callback glimpse,
 158                                   void *data, __u32 lvb_len,
 159                                   void *lvb_swabber,
 160                                   struct lustre_handle *lockh)
 161 {
 162         struct ldlm_lock *lock;
 163         int err;
 164         ENTRY;
 165
 166         if (ns->ns_client) {
 167                 CERROR("Trying to enqueue local lock in a shadow namespace\n");
 168                 LBUG();
 169         }
 170
 171         lock = ldlm_lock_create(ns, NULL, res_id, type, mode, blocking,
 172                                 completion, glimpse, data, lvb_len);
 173         if (!lock)
 174                 GOTO(out_nolock, err = -ENOMEM);
 175         LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
 176
 177         ldlm_lock_addref_internal(lock, mode);
 178         ldlm_lock2handle(lock, lockh);
 179         lock->l_flags |= LDLM_FL_LOCAL;
 180         lock->l_lvb_swabber = lvb_swabber;
 181         if (policy != NULL)
 182                 memcpy(&lock->l_policy_data, policy, sizeof(*policy));
 183         if (type == LDLM_EXTENT)
 184                 memcpy(&lock->l_req_extent, &policy->l_extent,
 185                        sizeof(policy->l_extent));
 186
 187         err = ldlm_lock_enqueue(ns, &lock, policy, flags);
 188         if (err != ELDLM_OK)
 189                 GOTO(out, err);
 190
 191         if (policy != NULL)
 192                 memcpy(policy, &lock->l_policy_data, sizeof(*policy));
 193         if ((*flags) & LDLM_FL_LOCK_CHANGED)
 194                 memcpy(&res_id, &lock->l_resource->lr_name, sizeof(res_id));
 195
 196         LDLM_DEBUG_NOLOCK("client-side local enqueue handler END (lock %p)",
 197                           lock);
 198
 199         if (lock->l_completion_ast)
 200                 lock->l_completion_ast(lock, *flags, NULL);
 201
 202         LDLM_DEBUG(lock, "client-side local enqueue END");
 203         EXIT;
 204  out:
 205         LDLM_LOCK_PUT(lock);
 206  out_nolock:
 207         return err;
 208 }
 209
 210 static void failed_lock_cleanup(struct ldlm_namespace *ns,
 211                                 struct ldlm_lock *lock,
 212                                 struct lustre_handle *lockh, int mode)
 213 {
 214         /* Set a flag to prevent us from sending a CANCEL (bug 407) */
 215         l_lock(&ns->ns_lock);
 216         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 217         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 218         l_unlock(&ns->ns_lock);
 219
 220         ldlm_lock_decref_and_cancel(lockh, mode);
 221
 222         /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
 223          *       from llite/file.c/ll_file_flock(). */
 224         if (lock->l_resource->lr_type == LDLM_FLOCK) {
 225                 ldlm_lock_destroy(lock);
 226         }
 227 }
 228
 229 int ldlm_cli_enqueue(struct obd_export *exp,
 230                      struct ptlrpc_request *req,
 231                      struct ldlm_namespace *ns,
 232                      struct ldlm_res_id res_id,
 233                      __u32 type,
 234                      ldlm_policy_data_t *policy,
 235                      ldlm_mode_t mode,
 236                      int *flags,
 237                      ldlm_blocking_callback blocking,
 238                      ldlm_completion_callback completion,
 239                      ldlm_glimpse_callback glimpse,
 240                      void *data,
 241                      void *lvb,
 242                      __u32 lvb_len,
 243                      void *lvb_swabber,
 244                      struct lustre_handle *lockh)
 245 {
 246         struct ldlm_lock *lock;
 247         struct ldlm_request *body;
 248         struct ldlm_reply *reply;
 249         int rc, size[3] = {0, sizeof(*body), lvb_len}, req_passed_in = 1;
 250         int is_replay = *flags & LDLM_FL_REPLAY;
 251         int cleanup_phase = 0;
 252         ENTRY;
 253
 254         if (exp == NULL) {
 255                 LASSERT(!is_replay);
 256                 rc = ldlm_cli_enqueue_local(ns, res_id, type, policy, mode,
 257                                             flags, blocking, completion,
 258                                             glimpse, data, lvb_len, lvb_swabber,
 259                                             lockh);
 260                 RETURN(rc);
 261         }
 262
 263         /* If we're replaying this lock, just check some invariants.
 264          * If we're creating a new lock, get everything all setup nice. */
 265         if (is_replay) {
 266                 lock = ldlm_handle2lock(lockh);
 267                 LDLM_DEBUG(lock, "client-side enqueue START");
 268                 LASSERT(exp == lock->l_conn_export);
 269         } else {
 270                 lock = ldlm_lock_create(ns, NULL, res_id, type, mode, blocking,
 271                                         completion, glimpse, data, lvb_len);
 272                 if (lock == NULL)
 273                         RETURN(-ENOMEM);
 274                 /* for the local lock, add the reference */
 275                 ldlm_lock_addref_internal(lock, mode);
 276                 ldlm_lock2handle(lock, lockh);
 277                 lock->l_lvb_swabber = lvb_swabber;
 278                 if (policy != NULL)
 279                         memcpy(&lock->l_policy_data, policy, sizeof(*policy));
 280                 if (type == LDLM_EXTENT)
 281                         memcpy(&lock->l_req_extent, &policy->l_extent,
 282                                sizeof(policy->l_extent));
 283                 LDLM_DEBUG(lock, "client-side enqueue START");
 284         }
 285
 286         /* lock not sent to server yet */
 287         cleanup_phase = 2;
 288
 289         if (req == NULL) {
 290                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
 291                                       LDLM_ENQUEUE, 2, size, NULL);
 292                 if (req == NULL)
 293                         GOTO(cleanup, rc = -ENOMEM);
 294                 req_passed_in = 0;
 295         }
 296
 297         LASSERTF(req->rq_reqmsg->buflens[MDS_REQ_INTENT_LOCKREQ_OFF] ==
 298                  sizeof(*body), "buflen[%d] = %d, not %d\n",
 299                  MDS_REQ_INTENT_LOCKREQ_OFF,
 300                  req->rq_reqmsg->buflens[MDS_REQ_INTENT_LOCKREQ_OFF],
 301                  sizeof(*body));
 302
 303         /* Dump lock data into the request buffer */
 304         body = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_INTENT_LOCKREQ_OFF,
 305                               sizeof (*body));
 306         ldlm_lock2desc(lock, &body->lock_desc);
 307         body->lock_flags = *flags;
 308
 309         memcpy(&body->lock_handle1, lockh, sizeof(*lockh));
 310
 311         /* Continue as normal. */
 312         if (!req_passed_in) {
 313                 size[0] = sizeof(*reply);
 314                 req->rq_replen = lustre_msg_size(1 + (lvb_len > 0), size);
 315         }
 316         lock->l_conn_export = exp;
 317         lock->l_export = NULL;
 318         lock->l_blocking_ast = blocking;
 319
 320         LDLM_DEBUG(lock, "sending request");
 321         rc = ptlrpc_queue_wait(req);
 322
 323         if (rc != ELDLM_OK) {
 324                 LASSERT(!is_replay);
 325                 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
 326                            rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
 327                 if (rc == ELDLM_LOCK_ABORTED) {
 328                         /* Before we return, swab the reply */
 329                         reply = lustre_swab_repbuf(req, 0, sizeof(*reply),
 330                                                    lustre_swab_ldlm_reply);
 331                         if (reply == NULL) {
 332                                 CERROR("Can't unpack ldlm_reply\n");
 333                                 rc = -EPROTO;
 334                         }
 335                         if (lvb_len) {
 336                                 void *tmplvb;
 337                                 tmplvb = lustre_swab_repbuf(req, 1, lvb_len,
 338                                                             lvb_swabber);
 339                                 if (tmplvb == NULL)
 340                                         GOTO(cleanup, rc = -EPROTO);
 341                                 if (lvb != NULL)
 342                                         memcpy(lvb, tmplvb, lvb_len);
 343                         }
 344                 }
 345                 GOTO(cleanup, rc);
 346         }
 347
 348         reply = lustre_swab_repbuf(req, 0, sizeof(*reply),
 349                                    lustre_swab_ldlm_reply);
 350         if (reply == NULL) {
 351                 CERROR("Can't unpack ldlm_reply\n");
 352                 GOTO(cleanup, rc = -EPROTO);
 353         }
 354
 355         /* lock enqueued on the server */
 356         cleanup_phase = 1;
 357
 358         memcpy(&lock->l_remote_handle, &reply->lock_handle,
 359                sizeof(lock->l_remote_handle));
 360         *flags = reply->lock_flags;
 361
 362         CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%x\n",
 363                lock, reply->lock_handle.cookie, *flags);
 364
 365         /* If enqueue returned a blocked lock but the completion handler has
 366          * already run, then it fixed up the resource and we don't need to do it
 367          * again. */
 368         if ((*flags) & LDLM_FL_LOCK_CHANGED) {
 369                 int newmode = reply->lock_desc.l_req_mode;
 370                 LASSERT(!is_replay);
 371                 if (newmode && newmode != lock->l_req_mode) {
 372                         LDLM_DEBUG(lock, "server returned different mode %s",
 373                                    ldlm_lockname[newmode]);
 374                         lock->l_req_mode = newmode;
 375                 }
 376
 377                 if (reply->lock_desc.l_resource.lr_name.name[0] !=
 378                     lock->l_resource->lr_name.name[0] ||
 379                    reply->lock_desc.l_resource.lr_name.name[1] !=
 380                     lock->l_resource->lr_name.name[1]) {
 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                         cleanup_phase = 2;
 416                         GOTO(cleanup, rc = -EPROTO);
 417                 }
 418                 memcpy(lock->l_lvb_data, tmplvb, lvb_len);
 419         }
 420
 421         if (!is_replay) {
 422                 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
 423                 if (lock->l_completion_ast != NULL) {
 424                         int err = lock->l_completion_ast(lock, *flags, NULL);
 425                         if (!rc)
 426                                 rc = err;
 427                         if (rc)
 428                                 cleanup_phase = 2;
 429                 }
 430         }
 431
 432         if (lvb_len && lvb != NULL) {
 433                 /* Copy the LVB here, and not earlier, because the completion
 434                  * AST (if any) can override what we got in the reply */
 435                 memcpy(lvb, lock->l_lvb_data, lvb_len);
 436         }
 437
 438         LDLM_DEBUG(lock, "client-side enqueue END");
 439         EXIT;
 440 cleanup:
 441         switch (cleanup_phase) {
 442         case 2:
 443                 if (rc)
 444                         failed_lock_cleanup(ns, lock, lockh, mode);
 445         case 1:
 446                 if (!req_passed_in && req != NULL)
 447                         ptlrpc_req_finished(req);
 448         }
 449
 450         LDLM_LOCK_PUT(lock);
 451         return rc;
 452 }
 453
 454 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
 455                                   int *flags)
 456 {
 457         struct ldlm_resource *res;
 458         int rc;
 459         ENTRY;
 460         if (lock->l_resource->lr_namespace->ns_client) {
 461                 CERROR("Trying to cancel local lock\n");
 462                 LBUG();
 463         }
 464         LDLM_DEBUG(lock, "client-side local convert");
 465
 466         res = ldlm_lock_convert(lock, new_mode, flags);
 467         if (res) {
 468                 ldlm_reprocess_all(res);
 469                 rc = 0;
 470         } else {
 471                 rc = EDEADLOCK;
 472         }
 473         LDLM_DEBUG(lock, "client-side local convert handler END");
 474         LDLM_LOCK_PUT(lock);
 475         RETURN(rc);
 476 }
 477
 478 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
 479  * conversion of locks which are on the waiting or converting queue */
 480 /* Caller of this code is supposed to take care of lock readers/writers
 481    accounting */
 482 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, int *flags)
 483 {
 484         struct ldlm_request *body;
 485         struct ldlm_reply *reply;
 486         struct ldlm_lock *lock;
 487         struct ldlm_resource *res;
 488         struct ptlrpc_request *req = NULL;
 489         int rc, size = sizeof(*body);
 490         ENTRY;
 491
 492         lock = ldlm_handle2lock(lockh);
 493         if (!lock) {
 494                 LBUG();
 495                 RETURN(-EINVAL);
 496         }
 497         *flags = 0;
 498
 499         if (lock->l_conn_export == NULL)
 500                 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
 501
 502         LDLM_DEBUG(lock, "client-side convert");
 503
 504         req = ptlrpc_prep_req(class_exp2cliimp(lock->l_conn_export),
 505                               LUSTRE_DLM_VERSION, LDLM_CONVERT, 1, &size, NULL);
 506         if (!req)
 507                 GOTO(out, rc = -ENOMEM);
 508
 509         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 510         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 511                sizeof(body->lock_handle1));
 512
 513         body->lock_desc.l_req_mode = new_mode;
 514         body->lock_flags = *flags;
 515
 516         size = sizeof(*reply);
 517         req->rq_replen = lustre_msg_size(1, &size);
 518
 519         rc = ptlrpc_queue_wait(req);
 520         if (rc != ELDLM_OK)
 521                 GOTO(out, rc);
 522
 523         reply = lustre_swab_repbuf(req, 0, sizeof (*reply),
 524                                    lustre_swab_ldlm_reply);
 525         if (reply == NULL) {
 526                 CERROR ("Can't unpack ldlm_reply\n");
 527                 GOTO (out, rc = -EPROTO);
 528         }
 529
 530         if (req->rq_status)
 531                 GOTO(out, rc = req->rq_status);
 532
 533         res = ldlm_lock_convert(lock, new_mode,
 534                                 (int *)&reply->lock_flags);
 535         if (res != NULL) {
 536                 ldlm_reprocess_all(res);
 537                 /* Go to sleep until the lock is granted. */
 538                 /* FIXME: or cancelled. */
 539                 if (lock->l_completion_ast) {
 540                         rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
 541                                                     NULL);
 542                         if (rc)
 543                                 GOTO(out, rc);
 544                 }
 545         } else {
 546                 rc = EDEADLOCK;
 547         }
 548         EXIT;
 549  out:
 550         LDLM_LOCK_PUT(lock);
 551         ptlrpc_req_finished(req);
 552         return rc;
 553 }
 554
 555 int ldlm_cli_cancel(struct lustre_handle *lockh)
 556 {
 557         struct ptlrpc_request *req;
 558         struct ldlm_lock *lock;
 559         struct ldlm_request *body;
 560         int rc = 0, size = sizeof(*body);
 561         ENTRY;
 562
 563         /* concurrent cancels on the same handle can happen */
 564         lock = __ldlm_handle2lock(lockh, LDLM_FL_CANCELING);
 565         if (lock == NULL)
 566                 RETURN(0);
 567
 568         if (lock->l_conn_export) {
 569                 int local_only;
 570                 struct obd_import *imp;
 571
 572                 LDLM_DEBUG(lock, "client-side cancel");
 573                 /* Set this flag to prevent others from getting new references*/
 574                 l_lock(&lock->l_resource->lr_namespace->ns_lock);
 575                 lock->l_flags |= LDLM_FL_CBPENDING;
 576                 local_only = lock->l_flags & LDLM_FL_LOCAL_ONLY;
 577                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 578                 ldlm_cancel_callback(lock);
 579
 580                 if (local_only) {
 581                         CDEBUG(D_INFO, "not sending request (at caller's "
 582                                "instruction)\n");
 583                         goto local_cancel;
 584                 }
 585
 586         restart:
 587                 imp = class_exp2cliimp(lock->l_conn_export);
 588                 if (imp == NULL || imp->imp_invalid) {
 589                         CDEBUG(D_HA, "skipping cancel on invalid import %p\n",
 590                                imp);
 591                         goto local_cancel;
 592                 }
 593
 594                 req = ptlrpc_prep_req(imp, LUSTRE_DLM_VERSION, LDLM_CANCEL,
 595                                       1, &size, NULL);
 596                 if (!req)
 597                         GOTO(out, rc = -ENOMEM);
 598                 req->rq_no_resend = 1;
 599
 600                 /* XXX FIXME bug 249 */
 601                 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
 602                 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
 603
 604                 body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 605                 memcpy(&body->lock_handle1, &lock->l_remote_handle,
 606                        sizeof(body->lock_handle1));
 607
 608                 req->rq_replen = lustre_msg_size(0, NULL);
 609
 610                 rc = ptlrpc_queue_wait(req);
 611
 612                 if (rc == ESTALE) {
 613                         char str[PTL_NALFMT_SIZE];
 614                         CERROR("client/server (nid %s) out of sync"
 615                                " -- not fatal\n",
 616                                ptlrpc_peernid2str(&req->rq_import->
 617                                                   imp_connection->c_peer, str));
 618                 } else if (rc == -ETIMEDOUT) {
 619                         ptlrpc_req_finished(req);
 620                         GOTO(restart, rc);
 621                 } else if (rc != ELDLM_OK) {
 622                         CERROR("Got rc %d from cancel RPC: canceling "
 623                                "anyway\n", rc);
 624                 }
 625
 626                 ptlrpc_req_finished(req);
 627         local_cancel:
 628                 ldlm_lock_cancel(lock);
 629         } else {
 630                 if (lock->l_resource->lr_namespace->ns_client) {
 631                         LDLM_ERROR(lock, "Trying to cancel local lock\n");
 632                         LBUG();
 633                 }
 634                 LDLM_DEBUG(lock, "client-side local cancel");
 635                 ldlm_lock_cancel(lock);
 636                 ldlm_reprocess_all(lock->l_resource);
 637                 LDLM_DEBUG(lock, "client-side local cancel handler END");
 638         }
 639
 640         EXIT;
 641  out:
 642         LDLM_LOCK_PUT(lock);
 643         return rc;
 644 }
 645
 646 /* when called with LDLM_ASYNC the blocking callback will be handled
 647  * in a thread and this function will return after the thread has been
 648  * asked to call the callback.  when called with LDLM_SYNC the blocking
 649  * callback will be performed in this function. */
 650 int ldlm_cancel_lru(struct ldlm_namespace *ns, ldlm_sync_t sync)
 651 {
 652         struct ldlm_lock *lock, *next;
 653         int count, rc = 0;
 654         LIST_HEAD(cblist);
 655         ENTRY;
 656
 657 #ifndef __KERNEL__
 658         sync = LDLM_SYNC; /* force to be sync in user space */
 659 #endif
 660
 661         l_lock(&ns->ns_lock);
 662         count = ns->ns_nr_unused - ns->ns_max_unused;
 663
 664         if (count <= 0) {
 665                 l_unlock(&ns->ns_lock);
 666                 RETURN(0);
 667         }
 668
 669         list_for_each_entry_safe(lock, next, &ns->ns_unused_list, l_lru) {
 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
 682                 /* We can't re-add to l_lru as it confuses the refcounting in
 683                  * ldlm_lock_remove_from_lru() if an AST arrives after we drop
 684                  * ns_lock below. We use l_tmp and can't use l_pending_chain as
 685                  * it is used both on server and client nevertheles bug 5666
 686                  * says it is used only on server. --umka */
 687                 if (sync != LDLM_ASYNC || ldlm_bl_to_thread(ns, NULL, lock))
 688                         list_add(&lock->l_tmp, &cblist);
 689
 690                 if (--count == 0)
 691                         break;
 692         }
 693         l_unlock(&ns->ns_lock);
 694
 695         list_for_each_entry_safe(lock, next, &cblist, l_tmp) {
 696                 list_del_init(&lock->l_tmp);
 697                 ldlm_handle_bl_callback(ns, NULL, lock);
 698         }
 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                         continue;
 728                 }
 729
 730                 if (lock->l_readers || lock->l_writers) {
 731                         if (flags & LDLM_FL_CONFIG_CHANGE)
 732                                 lock->l_flags |= LDLM_FL_CBPENDING;
 733                         else if (flags & LDLM_FL_WARN)
 734                                 LDLM_ERROR(lock, "lock in use");
 735                         continue;
 736                 }
 737
 738                 /* See CBPENDING comment in ldlm_cancel_lru */
 739                 lock->l_flags |= LDLM_FL_CBPENDING;
 740
 741                 OBD_ALLOC(w, sizeof(*w));
 742                 LASSERT(w);
 743
 744                 w->w_lock = LDLM_LOCK_GET(lock);
 745
 746                 list_add(&w->w_list, &list);
 747         }
 748         l_unlock(&ns->ns_lock);
 749
 750         list_for_each_safe(tmp, next, &list) {
 751                 struct lustre_handle lockh;
 752                 int rc;
 753                 w = list_entry(tmp, struct ldlm_ast_work, w_list);
 754
 755                 if (flags & LDLM_FL_LOCAL_ONLY) {
 756                         ldlm_lock_cancel(w->w_lock);
 757                 } else {
 758                         ldlm_lock2handle(w->w_lock, &lockh);
 759                         rc = ldlm_cli_cancel(&lockh);
 760                         if (rc != ELDLM_OK)
 761                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 762                 }
 763                 list_del(&w->w_list);
 764                 LDLM_LOCK_PUT(w->w_lock);
 765                 OBD_FREE(w, sizeof(*w));
 766         }
 767
 768         ldlm_resource_putref(res);
 769
 770         RETURN(0);
 771 }
 772
 773 static inline int have_no_nsresource(struct ldlm_namespace *ns)
 774 {
 775         int no_resource = 0;
 776
 777         spin_lock(&ns->ns_counter_lock);
 778         if (ns->ns_resources == 0)
 779                 no_resource = 1;
 780         spin_unlock(&ns->ns_counter_lock);
 781
 782         RETURN(no_resource);
 783 }
 784
 785 /* Cancel all locks on a namespace (or a specific resource, if given)
 786  * that have 0 readers/writers.
 787  *
 788  * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
 789  * to notify the server.
 790  * If flags & LDLM_FL_NO_CALLBACK, don't run the cancel callback.
 791  * If flags & LDLM_FL_WARN, print a warning if some locks are still in use.
 792  * If flags & LDLM_FL_CONFIG_CHANGE, mark all locks as having a pending callback
 793  */
 794 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
 795                            struct ldlm_res_id *res_id, int flags, void *opaque)
 796 {
 797         int i;
 798         struct l_wait_info lwi = { 0 };
 799         ENTRY;
 800
 801         if (ns == NULL)
 802                 RETURN(ELDLM_OK);
 803
 804         if (res_id)
 805                 RETURN(ldlm_cli_cancel_unused_resource(ns, *res_id, flags,
 806                                                        opaque));
 807
 808         l_lock(&ns->ns_lock);
 809         for (i = 0; i < RES_HASH_SIZE; i++) {
 810                 struct list_head *tmp, *next;
 811                 list_for_each_safe(tmp, next, &(ns->ns_hash[i])) {
 812                         int rc;
 813                         struct ldlm_resource *res;
 814                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
 815                         ldlm_resource_getref(res);
 816                         l_unlock(&ns->ns_lock);
 817
 818                         rc = ldlm_cli_cancel_unused_resource(ns, res->lr_name,
 819                                                              flags, opaque);
 820                         if (rc)
 821                                 CERROR("cancel_unused_res ("LPU64"): %d\n",
 822                                        res->lr_name.name[0], rc);
 823
 824                         l_lock(&ns->ns_lock);
 825                         next = tmp->next;
 826                         ldlm_resource_putref(res);
 827                 }
 828         }
 829         l_unlock(&ns->ns_lock);
 830         if (flags & LDLM_FL_CONFIG_CHANGE)
 831                 l_wait_event(ns->ns_waitq, have_no_nsresource(ns), &lwi);
 832
 833         RETURN(ELDLM_OK);
 834 }
 835
 836 /* Lock iterators. */
 837
 838 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
 839                           void *closure)
 840 {
 841         struct list_head *tmp, *next;
 842         struct ldlm_lock *lock;
 843         int rc = LDLM_ITER_CONTINUE;
 844         struct ldlm_namespace *ns = res->lr_namespace;
 845
 846         ENTRY;
 847
 848         if (!res)
 849                 RETURN(LDLM_ITER_CONTINUE);
 850
 851         l_lock(&ns->ns_lock);
 852         list_for_each_safe(tmp, next, &res->lr_granted) {
 853                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 854
 855                 if (iter(lock, closure) == LDLM_ITER_STOP)
 856                         GOTO(out, rc = LDLM_ITER_STOP);
 857         }
 858
 859         list_for_each_safe(tmp, next, &res->lr_converting) {
 860                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 861
 862                 if (iter(lock, closure) == LDLM_ITER_STOP)
 863                         GOTO(out, rc = LDLM_ITER_STOP);
 864         }
 865
 866         list_for_each_safe(tmp, next, &res->lr_waiting) {
 867                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 868
 869                 if (iter(lock, closure) == LDLM_ITER_STOP)
 870                         GOTO(out, rc = LDLM_ITER_STOP);
 871         }
 872  out:
 873         l_unlock(&ns->ns_lock);
 874         RETURN(rc);
 875 }
 876
 877 struct iter_helper_data {
 878         ldlm_iterator_t iter;
 879         void *closure;
 880 };
 881
 882 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
 883 {
 884         struct iter_helper_data *helper = closure;
 885         return helper->iter(lock, helper->closure);
 886 }
 887
 888 static int ldlm_res_iter_helper(struct ldlm_resource *res, void *closure)
 889 {
 890         return ldlm_resource_foreach(res, ldlm_iter_helper, closure);
 891 }
 892
 893 int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
 894                            void *closure)
 895 {
 896         struct iter_helper_data helper = { iter: iter, closure: closure };
 897         return ldlm_namespace_foreach_res(ns, ldlm_res_iter_helper, &helper);
 898 }
 899
 900 int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
 901                                ldlm_res_iterator_t iter, void *closure)
 902 {
 903         int i, rc = LDLM_ITER_CONTINUE;
 904
 905         l_lock(&ns->ns_lock);
 906         for (i = 0; i < RES_HASH_SIZE; i++) {
 907                 struct list_head *tmp, *next;
 908                 list_for_each_safe(tmp, next, &(ns->ns_hash[i])) {
 909                         struct ldlm_resource *res =
 910                                 list_entry(tmp, struct ldlm_resource, lr_hash);
 911
 912                         ldlm_resource_getref(res);
 913                         rc = iter(res, closure);
 914                         ldlm_resource_putref(res);
 915                         if (rc == LDLM_ITER_STOP)
 916                                 GOTO(out, rc);
 917                 }
 918         }
 919  out:
 920         l_unlock(&ns->ns_lock);
 921         RETURN(rc);
 922 }
 923
 924 /* non-blocking function to manipulate a lock whose cb_data is being put away.*/
 925 void ldlm_change_cbdata(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
 926                         ldlm_iterator_t iter, void *data)
 927 {
 928         struct ldlm_resource *res;
 929         ENTRY;
 930
 931         if (ns == NULL) {
 932                 CERROR("must pass in namespace\n");
 933                 LBUG();
 934         }
 935
 936         res = ldlm_resource_get(ns, NULL, *res_id, 0, 0);
 937         if (res == NULL) {
 938                 EXIT;
 939                 return;
 940         }
 941
 942         l_lock(&ns->ns_lock);
 943         ldlm_resource_foreach(res, iter, data);
 944         l_unlock(&ns->ns_lock);
 945         ldlm_resource_putref(res);
 946         EXIT;
 947 }
 948
 949 /* Lock replay */
 950
 951 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
 952 {
 953         struct list_head *list = closure;
 954
 955         /* we use l_pending_chain here, because it's unused on clients. */
 956         list_add(&lock->l_pending_chain, list);
 957         return LDLM_ITER_CONTINUE;
 958 }
 959
 960 static int replay_lock_interpret(struct ptlrpc_request *req,
 961                                     void * data, int rc)
 962 {
 963         struct ldlm_lock *lock;
 964         struct ldlm_reply *reply;
 965
 966         atomic_dec(&req->rq_import->imp_replay_inflight);
 967         if (rc != ELDLM_OK)
 968                 GOTO(out, rc);
 969
 970         lock = req->rq_async_args.pointer_arg[0];
 971         LASSERT(lock != NULL);
 972
 973         reply = lustre_swab_repbuf(req, 0, sizeof (*reply),
 974                                    lustre_swab_ldlm_reply);
 975         if (reply == NULL) {
 976                 CERROR("Can't unpack ldlm_reply\n");
 977                 GOTO (out, rc = -EPROTO);
 978         }
 979
 980         memcpy(&lock->l_remote_handle, &reply->lock_handle,
 981                sizeof(lock->l_remote_handle));
 982         LDLM_DEBUG(lock, "replayed lock:");
 983         ptlrpc_import_recovery_state_machine(req->rq_import);
 984 out:
 985         if (rc != ELDLM_OK)
 986                 ptlrpc_connect_import(req->rq_import, NULL);
 987
 988         RETURN(rc);
 989 }
 990
 991 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
 992 {
 993         struct ptlrpc_request *req;
 994         struct ldlm_request *body;
 995         struct ldlm_reply *reply;
 996         int buffers = 1;
 997         int size[2] = {0, sizeof(*body)};
 998         int flags;
 999
1000         /*
1001          * If granted mode matches the requested mode, this lock is granted.
1002          *
1003          * If they differ, but we have a granted mode, then we were granted
1004          * one mode and now want another: ergo, converting.
1005          *
1006          * If we haven't been granted anything and are on a resource list,
1007          * then we're blocked/waiting.
1008          *
1009          * If we haven't been granted anything and we're NOT on a resource list,
1010          * then we haven't got a reply yet and don't have a known disposition.
1011          * This happens whenever a lock enqueue is the request that triggers
1012          * recovery.
1013          */
1014         if (lock->l_granted_mode == lock->l_req_mode)
1015                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
1016         else if (lock->l_granted_mode)
1017                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
1018         else if (!list_empty(&lock->l_res_link))
1019                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
1020         else
1021                 flags = LDLM_FL_REPLAY;
1022
1023         req = ptlrpc_prep_req(imp, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
1024                               2, size, NULL);
1025         if (!req)
1026                 RETURN(-ENOMEM);
1027
1028         /* We're part of recovery, so don't wait for it. */
1029         req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
1030
1031         body = lustre_msg_buf(req->rq_reqmsg, 1, sizeof (*body));
1032         ldlm_lock2desc(lock, &body->lock_desc);
1033         body->lock_flags = flags;
1034
1035         ldlm_lock2handle(lock, &body->lock_handle1);
1036         size[0] = sizeof(*reply);
1037         if (lock->l_lvb_len != 0) {
1038                 buffers = 2;
1039                 size[1] = lock->l_lvb_len;
1040         }
1041         req->rq_replen = lustre_msg_size(buffers, size);
1042         /* notify the server we've replayed all requests.
1043          * also, we mark the request to be put on a dedicated
1044          * queue to be processed after all request replayes.
1045          * bug 6063 */
1046         lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
1047
1048         LDLM_DEBUG(lock, "replaying lock:");
1049
1050         imp->imp_locks_replayed++;
1051         atomic_inc(&req->rq_import->imp_replay_inflight);
1052         req->rq_async_args.pointer_arg[0] = lock;
1053         req->rq_interpret_reply = replay_lock_interpret;
1054         ptlrpcd_add_req(req);
1055
1056         RETURN(0);
1057 }
1058
1059 int ldlm_replay_locks(struct obd_import *imp)
1060 {
1061         struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
1062         struct list_head list, *pos, *next;
1063         struct ldlm_lock *lock;
1064         int rc = 0;
1065
1066         ENTRY;
1067         INIT_LIST_HEAD(&list);
1068
1069         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1070         LASSERT(ns != NULL);
1071
1072         /* ensure this doesn't fall to 0 before all have been queued */
1073         atomic_inc(&imp->imp_replay_inflight);
1074
1075         l_lock(&ns->ns_lock);
1076         (void)ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
1077
1078         list_for_each_safe(pos, next, &list) {
1079                 lock = list_entry(pos, struct ldlm_lock, l_pending_chain);
1080                 rc = replay_one_lock(imp, lock);
1081                 if (rc)
1082                         break; /* or try to do the rest? */
1083         }
1084         l_unlock(&ns->ns_lock);
1085
1086         atomic_dec(&imp->imp_replay_inflight);
1087
1088         RETURN(rc);
1089 }