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