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