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