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;
  52
  53                 LDLM_ERROR(lock, "lock timed out; not entering recovery in "
  54                            "server code, just going back to sleep");
  55                 if (time_after(jiffies, next_dump)) {
  56                         ldlm_namespace_dump(lock->l_resource->lr_namespace);
  57                         if (next_dump == 0)
  58                                 portals_debug_dumplog();
  59                         next_dump = jiffies + 300 * HZ;
  60                 }
  61                 RETURN(0);
  62         }
  63
  64         obd = lock->l_conn_export->exp_obd;
  65         imp = obd->u.cli.cl_import;
  66         ptlrpc_fail_import(imp, lwd->lwd_generation);
  67         LDLM_ERROR(lock, "lock timed out, entering recovery for %s@%s",
  68                    imp->imp_target_uuid.uuid,
  69                    imp->imp_connection->c_remote_uuid.uuid);
  70
  71         RETURN(0);
  72 }
  73
  74 int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data)
  75 {
  76         /* XXX ALLOCATE - 160 bytes */
  77         struct lock_wait_data lwd;
  78         unsigned long irqflags;
  79         struct obd_device *obd;
  80         struct obd_import *imp = NULL;
  81         struct l_wait_info lwi;
  82         int rc = 0;
  83         ENTRY;
  84
  85         if (flags == LDLM_FL_WAIT_NOREPROC)
  86                 goto noreproc;
  87
  88         if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
  89                        LDLM_FL_BLOCK_CONV))) {
  90                 wake_up(&lock->l_waitq);
  91                 RETURN(0);
  92         }
  93
  94         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
  95                    "sleeping");
  96         ldlm_lock_dump(D_OTHER, lock, 0);
  97         ldlm_reprocess_all(lock->l_resource);
  98
  99 noreproc:
 100
 101         obd = class_exp2obd(lock->l_conn_export);
 102
 103         /* if this is a local lock, then there is no import */
 104         if (obd != NULL)
 105                 imp = obd->u.cli.cl_import;
 106
 107         lwd.lwd_lock = lock;
 108
 109         if (flags & LDLM_FL_NO_TIMEOUT) {
 110                 LDLM_DEBUG(lock, "waiting indefinitely for group lock\n");
 111                 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
 112         } else {
 113                 lwi = LWI_TIMEOUT_INTR(obd_timeout * HZ,
 114                                        ldlm_expired_completion_wait,
 115                                        interrupted_completion_wait, &lwd);
 116         }
 117
 118         if (imp != NULL) {
 119                 spin_lock_irqsave(&imp->imp_lock, irqflags);
 120                 lwd.lwd_generation = imp->imp_generation;
 121                 spin_unlock_irqrestore(&imp->imp_lock, irqflags);
 122         }
 123
 124         /* Go to sleep until the lock is granted or cancelled. */
 125         rc = l_wait_event(lock->l_waitq,
 126                           ((lock->l_req_mode == lock->l_granted_mode) ||
 127                            (lock->l_flags & LDLM_FL_FAILED)), &lwi);
 128
 129         if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED) {
 130                 LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed");
 131                 RETURN(-EIO);
 132         }
 133
 134         if (rc) {
 135                 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
 136                            rc);
 137                 RETURN(rc);
 138         }
 139
 140         LDLM_DEBUG(lock, "client-side enqueue waking up: granted");
 141         RETURN(0);
 142 }
 143
 144 static int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
 145                                   struct ldlm_res_id res_id,
 146                                   __u32 type,
 147                                   ldlm_policy_data_t *policy,
 148                                   ldlm_mode_t mode,
 149                                   int *flags,
 150                                   ldlm_blocking_callback blocking,
 151                                   ldlm_completion_callback completion,
 152                                   ldlm_glimpse_callback glimpse,
 153                                   void *data, __u32 lvb_len,
 154                                   void *lvb_swabber,
 155                                   struct lustre_handle *lockh)
 156 {
 157         struct ldlm_lock *lock;
 158         int err;
 159         ENTRY;
 160
 161         if (ns->ns_client) {
 162                 CERROR("Trying to enqueue local lock in a shadow namespace\n");
 163                 LBUG();
 164         }
 165
 166         lock = ldlm_lock_create(ns, NULL, res_id, type, mode, blocking,
 167                                 completion, glimpse, data, lvb_len);
 168         if (!lock)
 169                 GOTO(out_nolock, err = -ENOMEM);
 170         LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
 171
 172         ldlm_lock_addref_internal(lock, mode);
 173         ldlm_lock2handle(lock, lockh);
 174         lock->l_flags |= LDLM_FL_LOCAL;
 175         lock->l_lvb_swabber = lvb_swabber;
 176         if (policy != NULL)
 177                 memcpy(&lock->l_policy_data, policy, sizeof(*policy));
 178         if (type == LDLM_EXTENT)
 179                 memcpy(&lock->l_req_extent, &policy->l_extent,
 180                        sizeof(policy->l_extent));
 181
 182         err = ldlm_lock_enqueue(ns, &lock, policy, flags);
 183         if (err != ELDLM_OK)
 184                 GOTO(out, err);
 185
 186         if (policy != NULL)
 187                 memcpy(policy, &lock->l_policy_data, sizeof(*policy));
 188         if ((*flags) & LDLM_FL_LOCK_CHANGED)
 189                 memcpy(&res_id, &lock->l_resource->lr_name, sizeof(res_id));
 190
 191         LDLM_DEBUG_NOLOCK("client-side local enqueue handler END (lock %p)",
 192                           lock);
 193
 194         if (lock->l_completion_ast)
 195                 lock->l_completion_ast(lock, *flags, NULL);
 196
 197         LDLM_DEBUG(lock, "client-side local enqueue END");
 198         EXIT;
 199  out:
 200         LDLM_LOCK_PUT(lock);
 201  out_nolock:
 202         return err;
 203 }
 204
 205 static void failed_lock_cleanup(struct ldlm_namespace *ns,
 206                                 struct ldlm_lock *lock,
 207                                 struct lustre_handle *lockh, int mode)
 208 {
 209         /* Set a flag to prevent us from sending a CANCEL (bug 407) */
 210         l_lock(&ns->ns_lock);
 211         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 212         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 213         l_unlock(&ns->ns_lock);
 214
 215         ldlm_lock_decref_and_cancel(lockh, mode);
 216 }
 217
 218 int ldlm_cli_enqueue(struct obd_export *exp,
 219                      struct ptlrpc_request *req,
 220                      struct ldlm_namespace *ns,
 221                      struct ldlm_res_id res_id,
 222                      __u32 type,
 223                      ldlm_policy_data_t *policy,
 224                      ldlm_mode_t mode,
 225                      int *flags,
 226                      ldlm_blocking_callback blocking,
 227                      ldlm_completion_callback completion,
 228                      ldlm_glimpse_callback glimpse,
 229                      void *data,
 230                      void *lvb,
 231                      __u32 lvb_len,
 232                      void *lvb_swabber,
 233                      struct lustre_handle *lockh)
 234 {
 235         struct ldlm_lock *lock;
 236         struct ldlm_request *body;
 237         struct ldlm_reply *reply;
 238         int rc, size[3] = {0, sizeof(*body), lvb_len}, req_passed_in = 1;
 239         int is_replay = *flags & LDLM_FL_REPLAY;
 240         int cleanup_phase = 0;
 241         ENTRY;
 242
 243         if (exp == NULL) {
 244                 LASSERT(!is_replay);
 245                 rc = ldlm_cli_enqueue_local(ns, res_id, type, policy, mode,
 246                                             flags, blocking, completion,
 247                                             glimpse, data, lvb_len, lvb_swabber,
 248                                             lockh);
 249                 RETURN(rc);
 250         }
 251
 252         /* If we're replaying this lock, just check some invariants.
 253          * If we're creating a new lock, get everything all setup nice. */
 254         if (is_replay) {
 255                 lock = ldlm_handle2lock(lockh);
 256                 LDLM_DEBUG(lock, "client-side enqueue START");
 257                 LASSERT(exp == lock->l_conn_export);
 258         } else {
 259                 lock = ldlm_lock_create(ns, NULL, res_id, type, mode, blocking,
 260                                         completion, glimpse, data, lvb_len);
 261                 if (lock == NULL)
 262                         RETURN(-ENOMEM);
 263                 /* for the local lock, add the reference */
 264                 ldlm_lock_addref_internal(lock, mode);
 265                 ldlm_lock2handle(lock, lockh);
 266                 lock->l_lvb_swabber = lvb_swabber;
 267                 if (policy != NULL)
 268                         memcpy(&lock->l_policy_data, policy, sizeof(*policy));
 269                 if (type == LDLM_EXTENT)
 270                         memcpy(&lock->l_req_extent, &policy->l_extent,
 271                                sizeof(policy->l_extent));
 272                 LDLM_DEBUG(lock, "client-side enqueue START");
 273         }
 274
 275         /* lock not sent to server yet */
 276         cleanup_phase = 2;
 277
 278         if (req == NULL) {
 279                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
 280                                       LDLM_ENQUEUE, 2, size, NULL);
 281                 if (req == NULL)
 282                         GOTO(cleanup, rc = -ENOMEM);
 283                 req_passed_in = 0;
 284         }
 285
 286         LASSERTF(req->rq_reqmsg->buflens[MDS_REQ_INTENT_LOCKREQ_OFF] ==
 287                  sizeof(*body), "buflen[%d] = %d, not %d\n",
 288                  MDS_REQ_INTENT_LOCKREQ_OFF,
 289                  req->rq_reqmsg->buflens[MDS_REQ_INTENT_LOCKREQ_OFF],
 290                  sizeof(*body));
 291
 292         /* Dump lock data into the request buffer */
 293         body = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_INTENT_LOCKREQ_OFF,
 294                               sizeof (*body));
 295         ldlm_lock2desc(lock, &body->lock_desc);
 296         body->lock_flags = *flags;
 297
 298         memcpy(&body->lock_handle1, lockh, sizeof(*lockh));
 299
 300         /* Continue as normal. */
 301         if (!req_passed_in) {
 302                 size[0] = sizeof(*reply);
 303                 req->rq_replen = lustre_msg_size(1 + (lvb_len > 0), size);
 304         }
 305         lock->l_conn_export = exp;
 306         lock->l_export = NULL;
 307         lock->l_blocking_ast = blocking;
 308
 309         LDLM_DEBUG(lock, "sending request");
 310         rc = ptlrpc_queue_wait(req);
 311
 312         if (rc != ELDLM_OK) {
 313                 LASSERT(!is_replay);
 314                 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
 315                            rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
 316                 if (rc == ELDLM_LOCK_ABORTED) {
 317                         /* Before we return, swab the reply */
 318                         reply = lustre_swab_repbuf(req, 0, sizeof(*reply),
 319                                                    lustre_swab_ldlm_reply);
 320                         if (reply == NULL) {
 321                                 CERROR("Can't unpack ldlm_reply\n");
 322                                 rc = -EPROTO;
 323                         }
 324                         if (lvb_len) {
 325                                 void *tmplvb;
 326                                 tmplvb = lustre_swab_repbuf(req, 1, lvb_len,
 327                                                             lvb_swabber);
 328                                 if (tmplvb == NULL)
 329                                         GOTO(cleanup, rc = -EPROTO);
 330                                 if (lvb != NULL)
 331                                         memcpy(lvb, tmplvb, lvb_len);
 332                         }
 333                 }
 334                 GOTO(cleanup, rc);
 335         }
 336
 337         reply = lustre_swab_repbuf(req, 0, sizeof(*reply),
 338                                    lustre_swab_ldlm_reply);
 339         if (reply == NULL) {
 340                 CERROR("Can't unpack ldlm_reply\n");
 341                 GOTO(cleanup, rc = -EPROTO);
 342         }
 343
 344         /* XXX - Phil, wasn't sure if this should go before or after the
 345          * lustre_swab_repbuf() ? If we can't unpack the reply then we
 346          * don't know what occurred on the server so I think the safest
 347          * bet is to cleanup the lock as if it didn't make it ? */
 348
 349         /* lock enqueued on the server */
 350         cleanup_phase = 1;
 351
 352         memcpy(&lock->l_remote_handle, &reply->lock_handle,
 353                sizeof(lock->l_remote_handle));
 354         *flags = reply->lock_flags;
 355
 356         CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%x\n",
 357                lock, reply->lock_handle.cookie, *flags);
 358
 359         /* If enqueue returned a blocked lock but the completion handler has
 360          * already run, then it fixed up the resource and we don't need to do it
 361          * again. */
 362         if ((*flags) & LDLM_FL_LOCK_CHANGED) {
 363                 int newmode = reply->lock_desc.l_req_mode;
 364                 LASSERT(!is_replay);
 365                 if (newmode && newmode != lock->l_req_mode) {
 366                         LDLM_DEBUG(lock, "server returned different mode %s",
 367                                    ldlm_lockname[newmode]);
 368                         lock->l_req_mode = newmode;
 369                 }
 370
 371                 if (reply->lock_desc.l_resource.lr_name.name[0] !=
 372                     lock->l_resource->lr_name.name[0] ||
 373                    reply->lock_desc.l_resource.lr_name.name[1] !=
 374                     lock->l_resource->lr_name.name[1]) {
 375                         CDEBUG(D_INFO, "remote intent success, locking %ld "
 376                                "instead of %ld\n",
 377                               (long)reply->lock_desc.l_resource.lr_name.name[0],
 378                                (long)lock->l_resource->lr_name.name[0]);
 379
 380                         ldlm_lock_change_resource(ns, lock,
 381                                            reply->lock_desc.l_resource.lr_name);
 382                         if (lock->l_resource == NULL) {
 383                                 LBUG();
 384                                 GOTO(cleanup, rc = -ENOMEM);
 385                         }
 386                         LDLM_DEBUG(lock, "client-side enqueue, new resource");
 387                 }
 388                 if (policy != NULL)
 389                         memcpy(&lock->l_policy_data,
 390                                &reply->lock_desc.l_policy_data,
 391                                sizeof(reply->lock_desc.l_policy_data));
 392                 if (type != LDLM_PLAIN)
 393                         LDLM_DEBUG(lock,"client-side enqueue, new policy data");
 394         }
 395
 396         if ((*flags) & LDLM_FL_AST_SENT) {
 397                 l_lock(&ns->ns_lock);
 398                 lock->l_flags |= LDLM_FL_CBPENDING;
 399                 l_unlock(&ns->ns_lock);
 400                 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
 401         }
 402
 403         /* If the lock has already been granted by a completion AST, don't
 404          * clobber the LVB with an older one. */
 405         if (lvb_len && (lock->l_req_mode != lock->l_granted_mode)) {
 406                 void *tmplvb;
 407                 tmplvb = lustre_swab_repbuf(req, 1, lvb_len, lvb_swabber);
 408                 if (tmplvb == NULL)
 409                         GOTO(cleanup, rc = -EPROTO);
 410                 memcpy(lock->l_lvb_data, tmplvb, lvb_len);
 411         }
 412
 413         if (!is_replay) {
 414                 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
 415                 if (lock->l_completion_ast != NULL) {
 416                         int err = lock->l_completion_ast(lock, *flags, NULL);
 417                         if (!rc)
 418                                 rc = err;
 419                 }
 420         }
 421
 422         if (lvb_len && lvb != NULL) {
 423                 /* Copy the LVB here, and not earlier, because the completion
 424                  * AST (if any) can override what we got in the reply */
 425                 memcpy(lvb, lock->l_lvb_data, lvb_len);
 426         }
 427
 428         LDLM_DEBUG(lock, "client-side enqueue END");
 429         EXIT;
 430 cleanup:
 431         switch (cleanup_phase) {
 432         case 2:
 433                 if (rc)
 434                         failed_lock_cleanup(ns, lock, lockh, mode);
 435         case 1:
 436                 if (!req_passed_in && req != NULL)
 437                         ptlrpc_req_finished(req);
 438         }
 439
 440         LDLM_LOCK_PUT(lock);
 441         return rc;
 442 }
 443
 444 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
 445                                   int *flags)
 446 {
 447         ENTRY;
 448         if (lock->l_resource->lr_namespace->ns_client) {
 449                 CERROR("Trying to cancel local lock\n");
 450                 LBUG();
 451         }
 452         LDLM_DEBUG(lock, "client-side local convert");
 453
 454         ldlm_lock_convert(lock, new_mode, flags);
 455         ldlm_reprocess_all(lock->l_resource);
 456
 457         LDLM_DEBUG(lock, "client-side local convert handler END");
 458         LDLM_LOCK_PUT(lock);
 459         RETURN(0);
 460 }
 461
 462 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
 463  * conversion of locks which are on the waiting or converting queue */
 464 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, int *flags)
 465 {
 466         struct ldlm_request *body;
 467         struct ldlm_reply *reply;
 468         struct ldlm_lock *lock;
 469         struct ldlm_resource *res;
 470         struct ptlrpc_request *req;
 471         int rc, size = sizeof(*body);
 472         ENTRY;
 473
 474         lock = ldlm_handle2lock(lockh);
 475         if (!lock) {
 476                 LBUG();
 477                 RETURN(-EINVAL);
 478         }
 479         *flags = 0;
 480
 481         if (lock->l_conn_export == NULL)
 482                 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
 483
 484         LDLM_DEBUG(lock, "client-side convert");
 485
 486         req = ptlrpc_prep_req(class_exp2cliimp(lock->l_conn_export),
 487                               LUSTRE_DLM_VERSION, LDLM_CONVERT, 1, &size, NULL);
 488         if (!req)
 489                 GOTO(out, rc = -ENOMEM);
 490
 491         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 492         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 493                sizeof(body->lock_handle1));
 494
 495         body->lock_desc.l_req_mode = new_mode;
 496         body->lock_flags = *flags;
 497
 498         size = sizeof(*reply);
 499         req->rq_replen = lustre_msg_size(1, &size);
 500
 501         rc = ptlrpc_queue_wait(req);
 502         if (rc != ELDLM_OK)
 503                 GOTO(out, rc);
 504
 505         reply = lustre_swab_repbuf(req, 0, sizeof (*reply),
 506                                    lustre_swab_ldlm_reply);
 507         if (reply == NULL) {
 508                 CERROR ("Can't unpack ldlm_reply\n");
 509                 GOTO (out, rc = -EPROTO);
 510         }
 511
 512         res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
 513         if (res != NULL)
 514                 ldlm_reprocess_all(res);
 515         /* Go to sleep until the lock is granted. */
 516         /* FIXME: or cancelled. */
 517         if (lock->l_completion_ast)
 518                 lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC, NULL);
 519         EXIT;
 520  out:
 521         LDLM_LOCK_PUT(lock);
 522         ptlrpc_req_finished(req);
 523         return rc;
 524 }
 525
 526 int ldlm_cli_cancel(struct lustre_handle *lockh)
 527 {
 528         struct ptlrpc_request *req;
 529         struct ldlm_lock *lock;
 530         struct ldlm_request *body;
 531         int rc = 0, size = sizeof(*body);
 532         ENTRY;
 533
 534         /* concurrent cancels on the same handle can happen */
 535         lock = __ldlm_handle2lock(lockh, LDLM_FL_CANCELING);
 536         if (lock == NULL)
 537                 RETURN(0);
 538
 539         if (lock->l_conn_export) {
 540                 int local_only;
 541                 struct obd_import *imp;
 542
 543                 LDLM_DEBUG(lock, "client-side cancel");
 544                 /* Set this flag to prevent others from getting new references*/
 545                 l_lock(&lock->l_resource->lr_namespace->ns_lock);
 546                 lock->l_flags |= LDLM_FL_CBPENDING;
 547                 local_only = (lock->l_flags & LDLM_FL_LOCAL_ONLY);
 548                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 549                 ldlm_cancel_callback(lock);
 550
 551                 if (local_only) {
 552                         CDEBUG(D_INFO, "not sending request (at caller's "
 553                                "instruction)\n");
 554                         goto local_cancel;
 555                 }
 556
 557         restart:
 558                 imp = class_exp2cliimp(lock->l_conn_export);
 559                 if (imp == NULL || imp->imp_invalid) {
 560                         CDEBUG(D_HA, "skipping cancel on invalid import %p\n",
 561                                imp);
 562                         goto local_cancel;
 563                 }
 564
 565                 req = ptlrpc_prep_req(imp, LUSTRE_DLM_VERSION, LDLM_CANCEL,
 566                                       1, &size, NULL);
 567                 if (!req)
 568                         GOTO(out, rc = -ENOMEM);
 569                 req->rq_no_resend = 1;
 570
 571                 /* XXX FIXME bug 249 */
 572                 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
 573                 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
 574
 575                 body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 576                 memcpy(&body->lock_handle1, &lock->l_remote_handle,
 577                        sizeof(body->lock_handle1));
 578
 579                 req->rq_replen = lustre_msg_size(0, NULL);
 580
 581                 rc = ptlrpc_queue_wait(req);
 582
 583                 if (rc == ESTALE) {
 584                         char str[PTL_NALFMT_SIZE];
 585                         CERROR("client/server (nid %s) out of sync"
 586                                " -- not fatal\n",
 587                                ptlrpc_peernid2str(&req->rq_import->
 588                                                   imp_connection->c_peer, str));
 589                 } else if (rc == -ETIMEDOUT) {
 590                         ptlrpc_req_finished(req);
 591                         GOTO(restart, rc);
 592                 } else if (rc != ELDLM_OK) {
 593                         CERROR("Got rc %d from cancel RPC: canceling "
 594                                "anyway\n", rc);
 595                 }
 596
 597                 ptlrpc_req_finished(req);
 598         local_cancel:
 599                 ldlm_lock_cancel(lock);
 600         } else {
 601                 if (lock->l_resource->lr_namespace->ns_client) {
 602                         LDLM_ERROR(lock, "Trying to cancel local lock\n");
 603                         LBUG();
 604                 }
 605                 LDLM_DEBUG(lock, "client-side local cancel");
 606                 ldlm_lock_cancel(lock);
 607                 ldlm_reprocess_all(lock->l_resource);
 608                 LDLM_DEBUG(lock, "client-side local cancel handler END");
 609         }
 610
 611         EXIT;
 612  out:
 613         LDLM_LOCK_PUT(lock);
 614         return rc;
 615 }
 616
 617 /* when called with LDLM_ASYNC the blocking callback will be handled
 618  * in a thread and this function will return after the thread has been
 619  * asked to call the callback.  when called with LDLM_SYNC the blocking
 620  * callback will be performed in this function. */
 621 int ldlm_cancel_lru(struct ldlm_namespace *ns, ldlm_sync_t sync)
 622 {
 623         struct list_head *tmp, *next;
 624         struct ldlm_lock *lock;
 625         int count, rc = 0;
 626         LIST_HEAD(cblist);
 627         ENTRY;
 628
 629         l_lock(&ns->ns_lock);
 630         count = ns->ns_nr_unused - ns->ns_max_unused;
 631
 632         if (count <= 0) {
 633                 l_unlock(&ns->ns_lock);
 634                 RETURN(0);
 635         }
 636
 637         list_for_each_safe(tmp, next, &ns->ns_unused_list) {
 638
 639                 lock = list_entry(tmp, struct ldlm_lock, l_lru);
 640
 641                 LASSERT(!lock->l_readers && !lock->l_writers);
 642
 643                 /* Setting the CBPENDING flag is a little misleading, but
 644                  * prevents an important race; namely, once CBPENDING is set,
 645                  * the lock can accumulate no more readers/writers.  Since
 646                  * readers and writers are already zero here, ldlm_lock_decref
 647                  * won't see this flag and call l_blocking_ast */
 648                 lock->l_flags |= LDLM_FL_CBPENDING;
 649
 650                 LDLM_LOCK_GET(lock); /* dropped by bl thread */
 651                 ldlm_lock_remove_from_lru(lock);
 652                 if (sync == LDLM_ASYNC)
 653                         ldlm_bl_to_thread(ns, NULL, lock);
 654                 else
 655                         list_add(&lock->l_lru, &cblist);
 656
 657                 if (--count == 0)
 658                         break;
 659         }
 660         l_unlock(&ns->ns_lock);
 661
 662         list_for_each_safe(tmp, next, &cblist) {
 663                 lock = list_entry(tmp, struct ldlm_lock, l_lru);
 664                 list_del_init(&lock->l_lru);
 665                 ldlm_handle_bl_callback(ns, NULL, lock);
 666         }
 667         RETURN(rc);
 668 }
 669
 670 static int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
 671                                            struct ldlm_res_id res_id, int flags,
 672                                            void *opaque)
 673 {
 674         struct ldlm_resource *res;
 675         struct list_head *tmp, *next, list = LIST_HEAD_INIT(list);
 676         struct ldlm_ast_work *w;
 677         ENTRY;
 678
 679         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
 680         if (res == NULL) {
 681                 /* This is not a problem. */
 682                 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id.name[0]);
 683                 RETURN(0);
 684         }
 685
 686         l_lock(&ns->ns_lock);
 687         list_for_each(tmp, &res->lr_granted) {
 688                 struct ldlm_lock *lock;
 689                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 690
 691                 if (opaque != NULL && lock->l_ast_data != opaque) {
 692                         LDLM_ERROR(lock, "data %p doesn't match opaque %p",
 693                                    lock->l_ast_data, opaque);
 694                         continue;
 695                 }
 696
 697                 if (lock->l_readers || lock->l_writers) {
 698                         if (flags & LDLM_FL_CONFIG_CHANGE)
 699                                 lock->l_flags |= LDLM_FL_CBPENDING;
 700                         else if (flags & LDLM_FL_WARN)
 701                                 LDLM_ERROR(lock, "lock in use");
 702                         continue;
 703                 }
 704
 705                 /* See CBPENDING comment in ldlm_cancel_lru */
 706                 lock->l_flags |= LDLM_FL_CBPENDING;
 707
 708                 OBD_ALLOC(w, sizeof(*w));
 709                 LASSERT(w);
 710
 711                 w->w_lock = LDLM_LOCK_GET(lock);
 712
 713                 list_add(&w->w_list, &list);
 714         }
 715         l_unlock(&ns->ns_lock);
 716
 717         list_for_each_safe(tmp, next, &list) {
 718                 struct lustre_handle lockh;
 719                 int rc;
 720                 w = list_entry(tmp, struct ldlm_ast_work, w_list);
 721
 722                 if (flags & LDLM_FL_LOCAL_ONLY) {
 723                         ldlm_lock_cancel(w->w_lock);
 724                 } else {
 725                         ldlm_lock2handle(w->w_lock, &lockh);
 726                         rc = ldlm_cli_cancel(&lockh);
 727                         if (rc != ELDLM_OK)
 728                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 729                 }
 730                 list_del(&w->w_list);
 731                 LDLM_LOCK_PUT(w->w_lock);
 732                 OBD_FREE(w, sizeof(*w));
 733         }
 734
 735         ldlm_resource_putref(res);
 736
 737         RETURN(0);
 738 }
 739
 740 static inline int have_no_nsresource(struct ldlm_namespace *ns)
 741 {
 742         int no_resource = 0;
 743
 744         spin_lock(&ns->ns_counter_lock);
 745         if (ns->ns_resources == 0)
 746                 no_resource = 1;
 747         spin_unlock(&ns->ns_counter_lock);
 748
 749         RETURN(no_resource);
 750 }
 751
 752 /* Cancel all locks on a namespace (or a specific resource, if given)
 753  * that have 0 readers/writers.
 754  *
 755  * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
 756  * to notify the server.
 757  * If flags & LDLM_FL_NO_CALLBACK, don't run the cancel callback.
 758  * If flags & LDLM_FL_WARN, print a warning if some locks are still in use.
 759  * If flags & LDLM_FL_CONFIG_CHANGE, mark all locks as having a pending callback
 760  */
 761 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
 762                            struct ldlm_res_id *res_id, int flags, void *opaque)
 763 {
 764         int i;
 765         struct l_wait_info lwi = { 0 };
 766         ENTRY;
 767
 768         if (ns == NULL)
 769                 RETURN(ELDLM_OK);
 770
 771         if (res_id)
 772                 RETURN(ldlm_cli_cancel_unused_resource(ns, *res_id, flags,
 773                                                        opaque));
 774
 775         l_lock(&ns->ns_lock);
 776         for (i = 0; i < RES_HASH_SIZE; i++) {
 777                 struct list_head *tmp, *next;
 778                 list_for_each_safe(tmp, next, &(ns->ns_hash[i])) {
 779                         int rc;
 780                         struct ldlm_resource *res;
 781                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
 782                         ldlm_resource_getref(res);
 783                         l_unlock(&ns->ns_lock);
 784
 785                         rc = ldlm_cli_cancel_unused_resource(ns, res->lr_name,
 786                                                              flags, opaque);
 787                         if (rc)
 788                                 CERROR("cancel_unused_res ("LPU64"): %d\n",
 789                                        res->lr_name.name[0], rc);
 790
 791                         l_lock(&ns->ns_lock);
 792                         next = tmp->next;
 793                         ldlm_resource_putref(res);
 794                 }
 795         }
 796         l_unlock(&ns->ns_lock);
 797         if (flags & LDLM_FL_CONFIG_CHANGE)
 798                 l_wait_event(ns->ns_waitq, have_no_nsresource(ns), &lwi);
 799
 800         RETURN(ELDLM_OK);
 801 }
 802
 803 /* Lock iterators. */
 804
 805 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
 806                           void *closure)
 807 {
 808         struct list_head *tmp, *next;
 809         struct ldlm_lock *lock;
 810         int rc = LDLM_ITER_CONTINUE;
 811         struct ldlm_namespace *ns = res->lr_namespace;
 812
 813         ENTRY;
 814
 815         if (!res)
 816                 RETURN(LDLM_ITER_CONTINUE);
 817
 818         l_lock(&ns->ns_lock);
 819         list_for_each_safe(tmp, next, &res->lr_granted) {
 820                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 821
 822                 if (iter(lock, closure) == LDLM_ITER_STOP)
 823                         GOTO(out, rc = LDLM_ITER_STOP);
 824         }
 825
 826         list_for_each_safe(tmp, next, &res->lr_converting) {
 827                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 828
 829                 if (iter(lock, closure) == LDLM_ITER_STOP)
 830                         GOTO(out, rc = LDLM_ITER_STOP);
 831         }
 832
 833         list_for_each_safe(tmp, next, &res->lr_waiting) {
 834                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 835
 836                 if (iter(lock, closure) == LDLM_ITER_STOP)
 837                         GOTO(out, rc = LDLM_ITER_STOP);
 838         }
 839  out:
 840         l_unlock(&ns->ns_lock);
 841         RETURN(rc);
 842 }
 843
 844 struct iter_helper_data {
 845         ldlm_iterator_t iter;
 846         void *closure;
 847 };
 848
 849 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
 850 {
 851         struct iter_helper_data *helper = closure;
 852         return helper->iter(lock, helper->closure);
 853 }
 854
 855 static int ldlm_res_iter_helper(struct ldlm_resource *res, void *closure)
 856 {
 857         return ldlm_resource_foreach(res, ldlm_iter_helper, closure);
 858 }
 859
 860 int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
 861                            void *closure)
 862 {
 863         struct iter_helper_data helper = { iter: iter, closure: closure };
 864         return ldlm_namespace_foreach_res(ns, ldlm_res_iter_helper, &helper);
 865 }
 866
 867 int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
 868                                ldlm_res_iterator_t iter, void *closure)
 869 {
 870         int i, rc = LDLM_ITER_CONTINUE;
 871
 872         l_lock(&ns->ns_lock);
 873         for (i = 0; i < RES_HASH_SIZE; i++) {
 874                 struct list_head *tmp, *next;
 875                 list_for_each_safe(tmp, next, &(ns->ns_hash[i])) {
 876                         struct ldlm_resource *res =
 877                                 list_entry(tmp, struct ldlm_resource, lr_hash);
 878
 879                         ldlm_resource_getref(res);
 880                         rc = iter(res, closure);
 881                         ldlm_resource_putref(res);
 882                         if (rc == LDLM_ITER_STOP)
 883                                 GOTO(out, rc);
 884                 }
 885         }
 886  out:
 887         l_unlock(&ns->ns_lock);
 888         RETURN(rc);
 889 }
 890
 891 /* non-blocking function to manipulate a lock whose cb_data is being put away.*/
 892 void ldlm_change_cbdata(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
 893                         ldlm_iterator_t iter, void *data)
 894 {
 895         struct ldlm_resource *res;
 896         ENTRY;
 897
 898         if (ns == NULL) {
 899                 CERROR("must pass in namespace");
 900                 LBUG();
 901         }
 902
 903         res = ldlm_resource_get(ns, NULL, *res_id, 0, 0);
 904         if (res == NULL) {
 905                 EXIT;
 906                 return;
 907         }
 908
 909         l_lock(&ns->ns_lock);
 910         ldlm_resource_foreach(res, iter, data);
 911         l_unlock(&ns->ns_lock);
 912         ldlm_resource_putref(res);
 913         EXIT;
 914 }
 915
 916 /* Lock replay */
 917
 918 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
 919 {
 920         struct list_head *list = closure;
 921
 922         /* we use l_pending_chain here, because it's unused on clients. */
 923         list_add(&lock->l_pending_chain, list);
 924         return LDLM_ITER_CONTINUE;
 925 }
 926
 927 static int replay_lock_interpret(struct ptlrpc_request *req,
 928                                     void * data, int rc)
 929 {
 930         struct ldlm_lock *lock;
 931         struct ldlm_reply *reply;
 932
 933         atomic_dec(&req->rq_import->imp_replay_inflight);
 934         if (rc != ELDLM_OK)
 935                 GOTO(out, rc);
 936
 937         lock = req->rq_async_args.pointer_arg[0];
 938         LASSERT(lock != NULL);
 939
 940         reply = lustre_swab_repbuf(req, 0, sizeof (*reply),
 941                                    lustre_swab_ldlm_reply);
 942         if (reply == NULL) {
 943                 CERROR("Can't unpack ldlm_reply\n");
 944                 GOTO (out, rc = -EPROTO);
 945         }
 946
 947         memcpy(&lock->l_remote_handle, &reply->lock_handle,
 948                sizeof(lock->l_remote_handle));
 949         LDLM_DEBUG(lock, "replayed lock:");
 950         ptlrpc_import_recovery_state_machine(req->rq_import);
 951  out:
 952         RETURN(rc);
 953 }
 954
 955 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
 956 {
 957         struct ptlrpc_request *req;
 958         struct ldlm_request *body;
 959         struct ldlm_reply *reply;
 960         int buffers = 1;
 961         int size[2] = {0, sizeof(*body)};
 962         int flags;
 963
 964         /*
 965          * If granted mode matches the requested mode, this lock is granted.
 966          *
 967          * If they differ, but we have a granted mode, then we were granted
 968          * one mode and now want another: ergo, converting.
 969          *
 970          * If we haven't been granted anything and are on a resource list,
 971          * then we're blocked/waiting.
 972          *
 973          * If we haven't been granted anything and we're NOT on a resource list,
 974          * then we haven't got a reply yet and don't have a known disposition.
 975          * This happens whenever a lock enqueue is the request that triggers
 976          * recovery.
 977          */
 978         if (lock->l_granted_mode == lock->l_req_mode)
 979                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
 980         else if (lock->l_granted_mode)
 981                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
 982         else if (!list_empty(&lock->l_res_link))
 983                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
 984         else
 985                 flags = LDLM_FL_REPLAY;
 986
 987         req = ptlrpc_prep_req(imp, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
 988                               2, size, NULL);
 989         if (!req)
 990                 RETURN(-ENOMEM);
 991
 992         /* We're part of recovery, so don't wait for it. */
 993         req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
 994
 995         body = lustre_msg_buf(req->rq_reqmsg, 1, sizeof (*body));
 996         ldlm_lock2desc(lock, &body->lock_desc);
 997         body->lock_flags = flags;
 998
 999         ldlm_lock2handle(lock, &body->lock_handle1);
1000         size[0] = sizeof(*reply);
1001         if (lock->l_lvb_len != 0) {
1002                 buffers = 2;
1003                 size[1] = lock->l_lvb_len;
1004         }
1005         req->rq_replen = lustre_msg_size(buffers, size);
1006
1007         LDLM_DEBUG(lock, "replaying lock:");
1008
1009         atomic_inc(&req->rq_import->imp_replay_inflight);
1010         req->rq_async_args.pointer_arg[0] = lock;
1011         req->rq_interpret_reply = replay_lock_interpret;
1012         ptlrpcd_add_req(req);
1013
1014         RETURN(0);
1015 }
1016
1017 int ldlm_replay_locks(struct obd_import *imp)
1018 {
1019         struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
1020         struct list_head list, *pos, *next;
1021         struct ldlm_lock *lock;
1022         int rc = 0;
1023
1024         ENTRY;
1025         INIT_LIST_HEAD(&list);
1026
1027         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1028         LASSERT(ns != NULL);
1029
1030         /* ensure this doesn't fall to 0 before all have been queued */
1031         atomic_inc(&imp->imp_replay_inflight);
1032
1033         l_lock(&ns->ns_lock);
1034         (void)ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
1035
1036         list_for_each_safe(pos, next, &list) {
1037                 lock = list_entry(pos, struct ldlm_lock, l_pending_chain);
1038                 rc = replay_one_lock(imp, lock);
1039                 if (rc)
1040                         break; /* or try to do the rest? */
1041         }
1042         l_unlock(&ns->ns_lock);
1043
1044         atomic_dec(&imp->imp_replay_inflight);
1045
1046         RETURN(rc);
1047 }