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         if (*flags & LDLM_FL_ATOMIC_CB)
 181                 lock->l_flags |= LDLM_FL_ATOMIC_CB;
 182         lock->l_lvb_swabber = lvb_swabber;
 183         if (policy != NULL)
 184                 memcpy(&lock->l_policy_data, policy, sizeof(*policy));
 185         if (type == LDLM_EXTENT)
 186                 memcpy(&lock->l_req_extent, &policy->l_extent,
 187                        sizeof(policy->l_extent));
 188
 189         err = ldlm_lock_enqueue(ns, &lock, policy, flags);
 190         if (err != ELDLM_OK)
 191                 GOTO(out, err);
 192
 193         if (policy != NULL)
 194                 memcpy(policy, &lock->l_policy_data, sizeof(*policy));
 195         if ((*flags) & LDLM_FL_LOCK_CHANGED)
 196                 memcpy(&res_id, &lock->l_resource->lr_name, sizeof(res_id));
 197
 198         LDLM_DEBUG_NOLOCK("client-side local enqueue handler END (lock %p)",
 199                           lock);
 200
 201         if (lock->l_completion_ast)
 202                 lock->l_completion_ast(lock, *flags, NULL);
 203
 204         LDLM_DEBUG(lock, "client-side local enqueue END");
 205         EXIT;
 206  out:
 207         LDLM_LOCK_PUT(lock);
 208  out_nolock:
 209         return err;
 210 }
 211
 212 static void failed_lock_cleanup(struct ldlm_namespace *ns,
 213                                 struct ldlm_lock *lock,
 214                                 struct lustre_handle *lockh, int mode)
 215 {
 216         /* Set a flag to prevent us from sending a CANCEL (bug 407) */
 217         lock_res_and_lock(lock);
 218         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 219         unlock_res_and_lock(lock);
 220         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 221
 222         ldlm_lock_decref_and_cancel(lockh, mode);
 223
 224         /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
 225          *       from llite/file.c/ll_file_flock(). */
 226         if (lock->l_resource->lr_type == LDLM_FLOCK) {
 227                 ldlm_lock_destroy(lock);
 228         }
 229 }
 230
 231 int ldlm_cli_enqueue(struct obd_export *exp,
 232                      struct ptlrpc_request *req,
 233                      struct ldlm_namespace *ns,
 234                      struct ldlm_res_id res_id,
 235                      __u32 type,
 236                      ldlm_policy_data_t *policy,
 237                      ldlm_mode_t mode,
 238                      int *flags,
 239                      ldlm_blocking_callback blocking,
 240                      ldlm_completion_callback completion,
 241                      ldlm_glimpse_callback glimpse,
 242                      void *data,
 243                      void *lvb,
 244                      __u32 lvb_len,
 245                      void *lvb_swabber,
 246                      struct lustre_handle *lockh)
 247 {
 248         struct ldlm_lock *lock;
 249         struct ldlm_request *body;
 250         struct ldlm_reply *reply;
 251         int rc, size[3] = {0, sizeof(*body), lvb_len}, req_passed_in = 1;
 252         int is_replay = *flags & LDLM_FL_REPLAY;
 253         int cleanup_phase = 0;
 254         ENTRY;
 255
 256         if (exp == NULL) {
 257                 LASSERT(!is_replay);
 258                 rc = ldlm_cli_enqueue_local(ns, res_id, type, policy, mode,
 259                                             flags, blocking, completion,
 260                                             glimpse, data, lvb_len, lvb_swabber,
 261                                             lockh);
 262                 RETURN(rc);
 263         }
 264
 265         /* If we're replaying this lock, just check some invariants.
 266          * If we're creating a new lock, get everything all setup nice. */
 267         if (is_replay) {
 268                 lock = ldlm_handle2lock(lockh);
 269                 LDLM_DEBUG(lock, "client-side enqueue START");
 270                 LASSERT(exp == lock->l_conn_export);
 271         } else {
 272                 lock = ldlm_lock_create(ns, NULL, res_id, type, mode, blocking,
 273                                         completion, glimpse, data, lvb_len);
 274                 if (lock == NULL)
 275                         RETURN(-ENOMEM);
 276                 /* for the local lock, add the reference */
 277                 ldlm_lock_addref_internal(lock, mode);
 278                 ldlm_lock2handle(lock, lockh);
 279                 lock->l_lvb_swabber = lvb_swabber;
 280                 if (policy != NULL)
 281                         memcpy(&lock->l_policy_data, policy, sizeof(*policy));
 282                 if (type == LDLM_EXTENT)
 283                         memcpy(&lock->l_req_extent, &policy->l_extent,
 284                                sizeof(policy->l_extent));
 285                 LDLM_DEBUG(lock, "client-side enqueue START");
 286         }
 287
 288         /* lock not sent to server yet */
 289         cleanup_phase = 2;
 290
 291         if (req == NULL) {
 292                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
 293                                       LDLM_ENQUEUE, 2, size, NULL);
 294                 if (req == NULL)
 295                         GOTO(cleanup, rc = -ENOMEM);
 296                 req_passed_in = 0;
 297         }
 298
 299         LASSERTF(req->rq_reqmsg->buflens[MDS_REQ_INTENT_LOCKREQ_OFF] ==
 300                  sizeof(*body), "buflen[%d] = %d, not %d\n",
 301                  MDS_REQ_INTENT_LOCKREQ_OFF,
 302                  req->rq_reqmsg->buflens[MDS_REQ_INTENT_LOCKREQ_OFF],
 303                  sizeof(*body));
 304
 305         /* Dump lock data into the request buffer */
 306         body = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_INTENT_LOCKREQ_OFF,
 307                               sizeof (*body));
 308         ldlm_lock2desc(lock, &body->lock_desc);
 309         body->lock_flags = *flags;
 310
 311         memcpy(&body->lock_handle1, lockh, sizeof(*lockh));
 312
 313         /* Continue as normal. */
 314         if (!req_passed_in) {
 315                 size[0] = sizeof(*reply);
 316                 req->rq_replen = lustre_msg_size(1 + (lvb_len > 0), size);
 317         }
 318         lock->l_conn_export = exp;
 319         lock->l_export = NULL;
 320         lock->l_blocking_ast = blocking;
 321
 322         LDLM_DEBUG(lock, "sending request");
 323         rc = ptlrpc_queue_wait(req);
 324
 325         if (rc != ELDLM_OK) {
 326                 LASSERT(!is_replay);
 327                 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
 328                            rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
 329                 if (rc == ELDLM_LOCK_ABORTED) {
 330                         /* Before we return, swab the reply */
 331                         reply = lustre_swab_repbuf(req, 0, sizeof(*reply),
 332                                                    lustre_swab_ldlm_reply);
 333                         if (reply == NULL) {
 334                                 CERROR("Can't unpack ldlm_reply\n");
 335                                 rc = -EPROTO;
 336                         }
 337                         if (lvb_len) {
 338                                 void *tmplvb;
 339                                 tmplvb = lustre_swab_repbuf(req, 1, lvb_len,
 340                                                             lvb_swabber);
 341                                 if (tmplvb == NULL)
 342                                         GOTO(cleanup, rc = -EPROTO);
 343                                 if (lvb != NULL)
 344                                         memcpy(lvb, tmplvb, lvb_len);
 345                         }
 346                 }
 347                 GOTO(cleanup, rc);
 348         }
 349
 350         reply = lustre_swab_repbuf(req, 0, sizeof(*reply),
 351                                    lustre_swab_ldlm_reply);
 352         if (reply == NULL) {
 353                 CERROR("Can't unpack ldlm_reply\n");
 354                 GOTO(cleanup, rc = -EPROTO);
 355         }
 356
 357         /* lock enqueued on the server */
 358         cleanup_phase = 1;
 359
 360         do_gettimeofday(&lock->l_enqueued_time);
 361         memcpy(&lock->l_remote_handle, &reply->lock_handle,
 362                sizeof(lock->l_remote_handle));
 363         *flags = reply->lock_flags;
 364
 365         CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%x\n",
 366                lock, reply->lock_handle.cookie, *flags);
 367
 368         /* If enqueue returned a blocked lock but the completion handler has
 369          * already run, then it fixed up the resource and we don't need to do it
 370          * again. */
 371         if ((*flags) & LDLM_FL_LOCK_CHANGED) {
 372                 int newmode = reply->lock_desc.l_req_mode;
 373                 LASSERT(!is_replay);
 374                 if (newmode && newmode != lock->l_req_mode) {
 375                         LDLM_DEBUG(lock, "server returned different mode %s",
 376                                    ldlm_lockname[newmode]);
 377                         lock->l_req_mode = newmode;
 378                 }
 379
 380                 if (reply->lock_desc.l_resource.lr_name.name[0] !=
 381                     lock->l_resource->lr_name.name[0] ||
 382                    reply->lock_desc.l_resource.lr_name.name[1] !=
 383                     lock->l_resource->lr_name.name[1]) {
 384                         CDEBUG(D_INFO, "remote intent success, locking %ld "
 385                                "instead of %ld\n",
 386                               (long)reply->lock_desc.l_resource.lr_name.name[0],
 387                                (long)lock->l_resource->lr_name.name[0]);
 388
 389                         ldlm_lock_change_resource(ns, lock,
 390                                            reply->lock_desc.l_resource.lr_name);
 391                         if (lock->l_resource == NULL) {
 392                                 LBUG();
 393                                 GOTO(cleanup, rc = -ENOMEM);
 394                         }
 395                         LDLM_DEBUG(lock, "client-side enqueue, new resource");
 396                 }
 397                 if (policy != NULL)
 398                         memcpy(&lock->l_policy_data,
 399                                &reply->lock_desc.l_policy_data,
 400                                sizeof(reply->lock_desc.l_policy_data));
 401                 if (type != LDLM_PLAIN)
 402                         LDLM_DEBUG(lock,"client-side enqueue, new policy data");
 403         }
 404
 405         if ((*flags) & LDLM_FL_AST_SENT) {
 406                 lock_res_and_lock(lock);
 407                 lock->l_flags |= LDLM_FL_CBPENDING;
 408                 unlock_res_and_lock(lock);
 409                 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
 410         }
 411
 412         /* If the lock has already been granted by a completion AST, don't
 413          * clobber the LVB with an older one. */
 414         if (lvb_len && (lock->l_req_mode != lock->l_granted_mode)) {
 415                 void *tmplvb;
 416                 tmplvb = lustre_swab_repbuf(req, 1, lvb_len, lvb_swabber);
 417                 if (tmplvb == NULL) {
 418                         cleanup_phase = 2;
 419                         GOTO(cleanup, rc = -EPROTO);
 420                 }
 421                 memcpy(lock->l_lvb_data, tmplvb, lvb_len);
 422         }
 423
 424         if (!is_replay) {
 425                 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
 426                 if (lock->l_completion_ast != NULL) {
 427                         /* since the lock made it to the server at this point
 428                          * it's the completion AST's responsibilty to cleanup
 429                          * the lock if the completion processing fails since
 430                          * it's no longer a simple local lock cancel. */
 431                         int err = lock->l_completion_ast(lock, *flags, NULL);
 432                         if (!rc)
 433                                 rc = err;
 434                 }
 435         }
 436
 437         if (lvb_len && lvb != NULL) {
 438                 /* Copy the LVB here, and not earlier, because the completion
 439                  * AST (if any) can override what we got in the reply */
 440                 memcpy(lvb, lock->l_lvb_data, lvb_len);
 441         }
 442
 443         LDLM_DEBUG(lock, "client-side enqueue END");
 444         EXIT;
 445 cleanup:
 446         switch (cleanup_phase) {
 447         case 2:
 448                 if (rc)
 449                         failed_lock_cleanup(ns, lock, lockh, mode);
 450         case 1:
 451                 if (!req_passed_in && req != NULL)
 452                         ptlrpc_req_finished(req);
 453         }
 454
 455         LDLM_LOCK_PUT(lock);
 456         return rc;
 457 }
 458
 459 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
 460                                   int *flags)
 461 {
 462         struct ldlm_resource *res;
 463         int rc;
 464         ENTRY;
 465         if (lock->l_resource->lr_namespace->ns_client) {
 466                 CERROR("Trying to cancel local lock\n");
 467                 LBUG();
 468         }
 469         LDLM_DEBUG(lock, "client-side local convert");
 470
 471         res = ldlm_lock_convert(lock, new_mode, flags);
 472         if (res) {
 473                 ldlm_reprocess_all(res);
 474                 rc = 0;
 475         } else {
 476                 rc = EDEADLOCK;
 477         }
 478         LDLM_DEBUG(lock, "client-side local convert handler END");
 479         LDLM_LOCK_PUT(lock);
 480         RETURN(rc);
 481 }
 482
 483 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
 484  * conversion of locks which are on the waiting or converting queue */
 485 /* Caller of this code is supposed to take care of lock readers/writers
 486    accounting */
 487 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, int *flags)
 488 {
 489         struct ldlm_request *body;
 490         struct ldlm_reply *reply;
 491         struct ldlm_lock *lock;
 492         struct ldlm_resource *res;
 493         struct ptlrpc_request *req = NULL;
 494         int rc, size = sizeof(*body);
 495         ENTRY;
 496
 497         lock = ldlm_handle2lock(lockh);
 498         if (!lock) {
 499                 LBUG();
 500                 RETURN(-EINVAL);
 501         }
 502         *flags = 0;
 503
 504         if (lock->l_conn_export == NULL)
 505                 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
 506
 507         LDLM_DEBUG(lock, "client-side convert");
 508
 509         req = ptlrpc_prep_req(class_exp2cliimp(lock->l_conn_export),
 510                               LUSTRE_DLM_VERSION, LDLM_CONVERT, 1, &size, NULL);
 511         if (!req)
 512                 GOTO(out, rc = -ENOMEM);
 513
 514         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 515         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 516                sizeof(body->lock_handle1));
 517
 518         body->lock_desc.l_req_mode = new_mode;
 519         body->lock_flags = *flags;
 520
 521         size = sizeof(*reply);
 522         req->rq_replen = lustre_msg_size(1, &size);
 523
 524         rc = ptlrpc_queue_wait(req);
 525         if (rc != ELDLM_OK)
 526                 GOTO(out, rc);
 527
 528         reply = lustre_swab_repbuf(req, 0, sizeof (*reply),
 529                                    lustre_swab_ldlm_reply);
 530         if (reply == NULL) {
 531                 CERROR ("Can't unpack ldlm_reply\n");
 532                 GOTO (out, rc = -EPROTO);
 533         }
 534
 535         if (req->rq_status)
 536                 GOTO(out, rc = req->rq_status);
 537
 538         res = ldlm_lock_convert(lock, new_mode,
 539                                 (int *)&reply->lock_flags);
 540         if (res != NULL) {
 541                 ldlm_reprocess_all(res);
 542                 /* Go to sleep until the lock is granted. */
 543                 /* FIXME: or cancelled. */
 544                 if (lock->l_completion_ast) {
 545                         rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
 546                                                     NULL);
 547                         if (rc)
 548                                 GOTO(out, rc);
 549                 }
 550         } else {
 551                 rc = EDEADLOCK;
 552         }
 553         EXIT;
 554  out:
 555         LDLM_LOCK_PUT(lock);
 556         ptlrpc_req_finished(req);
 557         return rc;
 558 }
 559
 560 int ldlm_cli_cancel(struct lustre_handle *lockh)
 561 {
 562         struct ptlrpc_request *req;
 563         struct ldlm_lock *lock;
 564         struct ldlm_request *body;
 565         int rc = 0, size = sizeof(*body);
 566         ENTRY;
 567
 568         /* concurrent cancels on the same handle can happen */
 569         lock = __ldlm_handle2lock(lockh, LDLM_FL_CANCELING);
 570         if (lock == NULL)
 571                 RETURN(0);
 572
 573         if (lock->l_conn_export) {
 574                 int local_only;
 575                 struct obd_import *imp;
 576
 577                 LDLM_DEBUG(lock, "client-side cancel");
 578                 /* Set this flag to prevent others from getting new references*/
 579                 lock_res_and_lock(lock);
 580                 lock->l_flags |= LDLM_FL_CBPENDING;
 581                 local_only = lock->l_flags & LDLM_FL_LOCAL_ONLY;
 582                 ldlm_cancel_callback(lock);
 583                 unlock_res_and_lock(lock);
 584
 585                 if (local_only) {
 586                         CDEBUG(D_INFO, "not sending request (at caller's "
 587                                "instruction)\n");
 588                         goto local_cancel;
 589                 }
 590
 591         restart:
 592                 imp = class_exp2cliimp(lock->l_conn_export);
 593                 if (imp == NULL || imp->imp_invalid) {
 594                         CDEBUG(D_HA, "skipping cancel on invalid import %p\n",
 595                                imp);
 596                         goto local_cancel;
 597                 }
 598
 599                 req = ptlrpc_prep_req(imp, LUSTRE_DLM_VERSION, LDLM_CANCEL,
 600                                       1, &size, NULL);
 601                 if (!req)
 602                         GOTO(out, rc = -ENOMEM);
 603                 req->rq_no_resend = 1;
 604
 605                 /* XXX FIXME bug 249 */
 606                 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
 607                 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
 608
 609                 body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 610                 memcpy(&body->lock_handle1, &lock->l_remote_handle,
 611                        sizeof(body->lock_handle1));
 612
 613                 req->rq_replen = lustre_msg_size(0, NULL);
 614
 615                 rc = ptlrpc_queue_wait(req);
 616
 617                 if (rc == ESTALE) {
 618                         char str[PTL_NALFMT_SIZE];
 619                         CERROR("client/server (nid %s) out of sync"
 620                                " -- not fatal\n",
 621                                ptlrpc_peernid2str(&req->rq_import->
 622                                                   imp_connection->c_peer, str));
 623                 } else if (rc == -ETIMEDOUT) {
 624                         ptlrpc_req_finished(req);
 625                         GOTO(restart, rc);
 626                 } else if (rc != ELDLM_OK) {
 627                         CERROR("Got rc %d from cancel RPC: canceling "
 628                                "anyway\n", rc);
 629                 }
 630
 631                 ptlrpc_req_finished(req);
 632         local_cancel:
 633                 ldlm_lock_cancel(lock);
 634         } else {
 635                 if (lock->l_resource->lr_namespace->ns_client) {
 636                         LDLM_ERROR(lock, "Trying to cancel local lock\n");
 637                         LBUG();
 638                 }
 639                 LDLM_DEBUG(lock, "client-side local cancel");
 640                 ldlm_lock_cancel(lock);
 641                 ldlm_reprocess_all(lock->l_resource);
 642                 LDLM_DEBUG(lock, "client-side local cancel handler END");
 643         }
 644
 645         EXIT;
 646  out:
 647         LDLM_LOCK_PUT(lock);
 648         return rc;
 649 }
 650
 651 /* when called with LDLM_ASYNC the blocking callback will be handled
 652  * in a thread and this function will return after the thread has been
 653  * asked to call the callback.  when called with LDLM_SYNC the blocking
 654  * callback will be performed in this function. */
 655 int ldlm_cancel_lru(struct ldlm_namespace *ns, ldlm_sync_t sync)
 656 {
 657         struct ldlm_lock *lock, *next;
 658         int count, rc = 0;
 659         LIST_HEAD(cblist);
 660         ENTRY;
 661
 662 #ifndef __KERNEL__
 663         sync = LDLM_SYNC; /* force to be sync in user space */
 664 #endif
 665
 666         spin_lock(&ns->ns_unused_lock);
 667         count = ns->ns_nr_unused - ns->ns_max_unused;
 668
 669         if (count <= 0) {
 670                 spin_unlock(&ns->ns_unused_lock);
 671                 RETURN(0);
 672         }
 673
 674         while (!list_empty(&ns->ns_unused_list)) {
 675                 struct list_head *tmp = ns->ns_unused_list.next;
 676                 lock = list_entry(tmp, struct ldlm_lock, l_lru);
 677                 LASSERT(!lock->l_readers && !lock->l_writers);
 678
 679                 LDLM_LOCK_GET(lock); /* dropped by bl thread */
 680                 spin_unlock(&ns->ns_unused_lock);
 681
 682                 lock_res_and_lock(lock);
 683                 ldlm_lock_remove_from_lru(lock);
 684
 685                 /* Setting the CBPENDING flag is a little misleading, but
 686                  * prevents an important race; namely, once CBPENDING is set,
 687                  * the lock can accumulate no more readers/writers.  Since
 688                  * readers and writers are already zero here, ldlm_lock_decref
 689                  * won't see this flag and call l_blocking_ast */
 690                 lock->l_flags |= LDLM_FL_CBPENDING;
 691
 692                 /* We can't re-add to l_lru as it confuses the refcounting in
 693                  * ldlm_lock_remove_from_lru() if an AST arrives after we drop
 694                  * ns_lock below. We use l_tmp and can't use l_pending_chain as
 695                  * it is used both on server and client nevertheles bug 5666
 696                  * says it is used only on server. --umka */
 697                 if (sync != LDLM_ASYNC || ldlm_bl_to_thread(ns, NULL, lock))
 698                         list_add(&lock->l_tmp, &cblist);
 699
 700                 unlock_res_and_lock(lock);
 701
 702                 spin_lock(&ns->ns_unused_lock);
 703
 704                 if (--count == 0)
 705                         break;
 706         }
 707         spin_unlock(&ns->ns_unused_lock);
 708
 709         list_for_each_entry_safe(lock, next, &cblist, l_tmp) {
 710                 list_del_init(&lock->l_tmp);
 711                 ldlm_handle_bl_callback(ns, NULL, lock);
 712         }
 713
 714         RETURN(rc);
 715 }
 716
 717 static int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
 718                                            struct ldlm_res_id res_id, int flags,
 719                                            void *opaque)
 720 {
 721         struct list_head *tmp, *next, list = LIST_HEAD_INIT(list);
 722         struct ldlm_resource *res;
 723         struct ldlm_lock *lock;
 724         ENTRY;
 725
 726         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
 727         if (res == NULL) {
 728                 /* This is not a problem. */
 729                 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id.name[0]);
 730                 RETURN(0);
 731         }
 732
 733         lock_res(res);
 734         list_for_each(tmp, &res->lr_granted) {
 735                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 736
 737                 if (opaque != NULL && lock->l_ast_data != opaque) {
 738                         LDLM_ERROR(lock, "data %p doesn't match opaque %p",
 739                                    lock->l_ast_data, opaque);
 740                         continue;
 741                 }
 742
 743                 if (lock->l_readers || lock->l_writers) {
 744                         if (flags & LDLM_FL_CONFIG_CHANGE)
 745                                 lock->l_flags |= LDLM_FL_CBPENDING;
 746                         else if (flags & LDLM_FL_WARN)
 747                                 LDLM_ERROR(lock, "lock in use");
 748                         continue;
 749                 }
 750
 751                 /* See CBPENDING comment in ldlm_cancel_lru */
 752                 lock->l_flags |= LDLM_FL_CBPENDING;
 753
 754                 LASSERT(list_empty(&lock->l_bl_ast));
 755                 list_add(&lock->l_bl_ast, &list);
 756                 LDLM_LOCK_GET(lock);
 757         }
 758         unlock_res(res);
 759
 760         list_for_each_safe(tmp, next, &list) {
 761                 struct lustre_handle lockh;
 762                 int rc;
 763                 lock = list_entry(tmp, struct ldlm_lock, l_bl_ast);
 764
 765                 if (flags & LDLM_FL_LOCAL_ONLY) {
 766                         ldlm_lock_cancel(lock);
 767                 } else {
 768                         ldlm_lock2handle(lock, &lockh);
 769                         rc = ldlm_cli_cancel(&lockh);
 770                         if (rc != ELDLM_OK)
 771                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 772                 }
 773                 list_del_init(&lock->l_bl_ast);
 774                 LDLM_LOCK_PUT(lock);
 775         }
 776
 777         ldlm_resource_putref(res);
 778
 779         RETURN(0);
 780 }
 781
 782 static inline int have_no_nsresource(struct ldlm_namespace *ns)
 783 {
 784         int no_resource = 0;
 785
 786         spin_lock(&ns->ns_hash_lock);
 787         if (ns->ns_resources == 0)
 788                 no_resource = 1;
 789         spin_unlock(&ns->ns_hash_lock);
 790
 791         RETURN(no_resource);
 792 }
 793
 794 /* Cancel all locks on a namespace (or a specific resource, if given)
 795  * that have 0 readers/writers.
 796  *
 797  * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
 798  * to notify the server.
 799  * If flags & LDLM_FL_NO_CALLBACK, don't run the cancel callback.
 800  * If flags & LDLM_FL_WARN, print a warning if some locks are still in use.
 801  * If flags & LDLM_FL_CONFIG_CHANGE, mark all locks as having a pending callback
 802  */
 803 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
 804                            struct ldlm_res_id *res_id, int flags, void *opaque)
 805 {
 806         int i;
 807         struct l_wait_info lwi = { 0 };
 808         ENTRY;
 809
 810         if (ns == NULL)
 811                 RETURN(ELDLM_OK);
 812
 813         if (res_id)
 814                 RETURN(ldlm_cli_cancel_unused_resource(ns, *res_id, flags,
 815                                                        opaque));
 816
 817         spin_lock(&ns->ns_hash_lock);
 818         for (i = 0; i < RES_HASH_SIZE; i++) {
 819                 struct list_head *tmp;
 820                 tmp = ns->ns_hash[i].next;
 821                 while (tmp != &(ns->ns_hash[i])) {
 822                         struct ldlm_resource *res;
 823                         int rc;
 824
 825                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
 826                         ldlm_resource_getref(res);
 827                         spin_unlock(&ns->ns_hash_lock);
 828
 829                         rc = ldlm_cli_cancel_unused_resource(ns, res->lr_name,
 830                                                              flags, opaque);
 831                         if (rc)
 832                                 CERROR("cancel_unused_res ("LPU64"): %d\n",
 833                                        res->lr_name.name[0], rc);
 834
 835                         spin_lock(&ns->ns_hash_lock);
 836                         tmp = tmp->next;
 837                         ldlm_resource_putref_locked(res);
 838                 }
 839         }
 840         spin_unlock(&ns->ns_hash_lock);
 841
 842         if (flags & LDLM_FL_CONFIG_CHANGE)
 843                 l_wait_event(ns->ns_waitq, have_no_nsresource(ns), &lwi);
 844
 845         RETURN(ELDLM_OK);
 846 }
 847
 848 /* Lock iterators. */
 849
 850 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
 851                           void *closure)
 852 {
 853         struct list_head *tmp, *next;
 854         struct ldlm_lock *lock;
 855         int rc = LDLM_ITER_CONTINUE;
 856
 857         ENTRY;
 858
 859         if (!res)
 860                 RETURN(LDLM_ITER_CONTINUE);
 861
 862         lock_res(res);
 863         list_for_each_safe(tmp, next, &res->lr_granted) {
 864                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 865
 866                 if (iter(lock, closure) == LDLM_ITER_STOP)
 867                         GOTO(out, rc = LDLM_ITER_STOP);
 868         }
 869
 870         list_for_each_safe(tmp, next, &res->lr_converting) {
 871                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 872
 873                 if (iter(lock, closure) == LDLM_ITER_STOP)
 874                         GOTO(out, rc = LDLM_ITER_STOP);
 875         }
 876
 877         list_for_each_safe(tmp, next, &res->lr_waiting) {
 878                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 879
 880                 if (iter(lock, closure) == LDLM_ITER_STOP)
 881                         GOTO(out, rc = LDLM_ITER_STOP);
 882         }
 883  out:
 884         unlock_res(res);
 885         RETURN(rc);
 886 }
 887
 888 struct iter_helper_data {
 889         ldlm_iterator_t iter;
 890         void *closure;
 891 };
 892
 893 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
 894 {
 895         struct iter_helper_data *helper = closure;
 896         return helper->iter(lock, helper->closure);
 897 }
 898
 899 static int ldlm_res_iter_helper(struct ldlm_resource *res, void *closure)
 900 {
 901         return ldlm_resource_foreach(res, ldlm_iter_helper, closure);
 902 }
 903
 904 int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
 905                            void *closure)
 906 {
 907         struct iter_helper_data helper = { iter: iter, closure: closure };
 908         return ldlm_namespace_foreach_res(ns, ldlm_res_iter_helper, &helper);
 909 }
 910
 911 int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
 912                                ldlm_res_iterator_t iter, void *closure)
 913 {
 914         int i, rc = LDLM_ITER_CONTINUE;
 915         struct ldlm_resource *res;
 916         struct list_head *tmp;
 917
 918         spin_lock(&ns->ns_hash_lock);
 919         for (i = 0; i < RES_HASH_SIZE; i++) {
 920                 tmp = ns->ns_hash[i].next;
 921                 while (tmp != &(ns->ns_hash[i])) {
 922                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
 923                         ldlm_resource_getref(res);
 924                         spin_unlock(&ns->ns_hash_lock);
 925
 926                         rc = iter(res, closure);
 927
 928                         spin_lock(&ns->ns_hash_lock);
 929                         tmp = tmp->next;
 930                         ldlm_resource_putref_locked(res);
 931                         if (rc == LDLM_ITER_STOP)
 932                                 GOTO(out, rc);
 933                 }
 934         }
 935  out:
 936         spin_unlock(&ns->ns_hash_lock);
 937         RETURN(rc);
 938 }
 939
 940 /* non-blocking function to manipulate a lock whose cb_data is being put away.*/
 941 void ldlm_change_cbdata(struct ldlm_namespace *ns,
 942                         struct ldlm_res_id *res_id,
 943                         ldlm_iterator_t iter,
 944                         void *data)
 945 {
 946         struct ldlm_resource *res;
 947         ENTRY;
 948
 949         if (ns == NULL) {
 950                 CERROR("must pass in namespace\n");
 951                 LBUG();
 952         }
 953
 954         res = ldlm_resource_get(ns, NULL, *res_id, 0, 0);
 955         if (res == NULL) {
 956                 EXIT;
 957                 return;
 958         }
 959
 960         ldlm_resource_foreach(res, iter, data);
 961         ldlm_resource_putref(res);
 962         EXIT;
 963 }
 964
 965 /* Lock replay */
 966
 967 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
 968 {
 969         struct list_head *list = closure;
 970
 971         /* we use l_pending_chain here, because it's unused on clients. */
 972         list_add(&lock->l_pending_chain, list);
 973         return LDLM_ITER_CONTINUE;
 974 }
 975
 976 static int replay_lock_interpret(struct ptlrpc_request *req,
 977                                     void * data, int rc)
 978 {
 979         struct ldlm_lock *lock;
 980         struct ldlm_reply *reply;
 981
 982         atomic_dec(&req->rq_import->imp_replay_inflight);
 983         if (rc != ELDLM_OK)
 984                 GOTO(out, rc);
 985
 986         lock = req->rq_async_args.pointer_arg[0];
 987         LASSERT(lock != NULL);
 988
 989         reply = lustre_swab_repbuf(req, 0, sizeof (*reply),
 990                                    lustre_swab_ldlm_reply);
 991         if (reply == NULL) {
 992                 CERROR("Can't unpack ldlm_reply\n");
 993                 GOTO (out, rc = -EPROTO);
 994         }
 995
 996         memcpy(&lock->l_remote_handle, &reply->lock_handle,
 997                sizeof(lock->l_remote_handle));
 998         LDLM_DEBUG(lock, "replayed lock:");
 999         ptlrpc_import_recovery_state_machine(req->rq_import);
1000 out:
1001         if (rc != ELDLM_OK)
1002                 ptlrpc_connect_import(req->rq_import, NULL);
1003
1004         RETURN(rc);
1005 }
1006
1007 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
1008 {
1009         struct ptlrpc_request *req;
1010         struct ldlm_request *body;
1011         struct ldlm_reply *reply;
1012         int buffers = 1;
1013         int size[2] = {0, sizeof(*body)};
1014         int flags;
1015
1016         /*
1017          * If granted mode matches the requested mode, this lock is granted.
1018          *
1019          * If they differ, but we have a granted mode, then we were granted
1020          * one mode and now want another: ergo, converting.
1021          *
1022          * If we haven't been granted anything and are on a resource list,
1023          * then we're blocked/waiting.
1024          *
1025          * If we haven't been granted anything and we're NOT on a resource list,
1026          * then we haven't got a reply yet and don't have a known disposition.
1027          * This happens whenever a lock enqueue is the request that triggers
1028          * recovery.
1029          */
1030         if (lock->l_granted_mode == lock->l_req_mode)
1031                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
1032         else if (lock->l_granted_mode)
1033                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
1034         else if (!list_empty(&lock->l_res_link))
1035                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
1036         else
1037                 flags = LDLM_FL_REPLAY;
1038
1039         req = ptlrpc_prep_req(imp, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
1040                               2, size, NULL);
1041         if (!req)
1042                 RETURN(-ENOMEM);
1043
1044         /* We're part of recovery, so don't wait for it. */
1045         req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
1046
1047         body = lustre_msg_buf(req->rq_reqmsg, 1, sizeof (*body));
1048         ldlm_lock2desc(lock, &body->lock_desc);
1049         body->lock_flags = flags;
1050
1051         ldlm_lock2handle(lock, &body->lock_handle1);
1052         size[0] = sizeof(*reply);
1053         if (lock->l_lvb_len != 0) {
1054                 buffers = 2;
1055                 size[1] = lock->l_lvb_len;
1056         }
1057         req->rq_replen = lustre_msg_size(buffers, size);
1058         /* notify the server we've replayed all requests.
1059          * also, we mark the request to be put on a dedicated
1060          * queue to be processed after all request replayes.
1061          * bug 6063 */
1062         lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
1063
1064         LDLM_DEBUG(lock, "replaying lock:");
1065
1066         imp->imp_locks_replayed++;
1067         atomic_inc(&req->rq_import->imp_replay_inflight);
1068         req->rq_async_args.pointer_arg[0] = lock;
1069         req->rq_interpret_reply = replay_lock_interpret;
1070         ptlrpcd_add_req(req);
1071
1072         RETURN(0);
1073 }
1074
1075 int ldlm_replay_locks(struct obd_import *imp)
1076 {
1077         struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
1078         struct list_head list, *pos, *next;
1079         struct ldlm_lock *lock;
1080         int rc = 0;
1081
1082         ENTRY;
1083         INIT_LIST_HEAD(&list);
1084
1085         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1086         LASSERT(ns != NULL);
1087
1088         /* ensure this doesn't fall to 0 before all have been queued */
1089         atomic_inc(&imp->imp_replay_inflight);
1090
1091         (void)ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
1092
1093         list_for_each_safe(pos, next, &list) {
1094                 lock = list_entry(pos, struct ldlm_lock, l_pending_chain);
1095                 rc = replay_one_lock(imp, lock);
1096                 if (rc)
1097                         break; /* or try to do the rest? */
1098         }
1099
1100         atomic_dec(&imp->imp_replay_inflight);
1101
1102         RETURN(rc);
1103 }