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 the Lustre file system, http://www.lustre.org
   7  *   Lustre is a trademark of Cluster File Systems, Inc.
   8  *
   9  *   You may have signed or agreed to another license before downloading
  10  *   this software.  If so, you are bound by the terms and conditions
  11  *   of that agreement, and the following does not apply to you.  See the
  12  *   LICENSE file included with this distribution for more information.
  13  *
  14  *   If you did not agree to a different license, then this copy of Lustre
  15  *   is open source software; you can redistribute it and/or modify it
  16  *   under the terms of version 2 of the GNU General Public License as
  17  *   published by the Free Software Foundation.
  18  *
  19  *   In either case, Lustre is distributed in the hope that it will be
  20  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  21  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  22  *   license text for more details.
  23  */
  24
  25 #define DEBUG_SUBSYSTEM S_LDLM
  26 #ifndef __KERNEL__
  27 #include <signal.h>
  28 #include <liblustre.h>
  29 #endif
  30
  31 #include <lustre_dlm.h>
  32 #include <obd_class.h>
  33 #include <obd.h>
  34
  35 #include "ldlm_internal.h"
  36
  37 static void interrupted_completion_wait(void *data)
  38 {
  39 }
  40
  41 struct lock_wait_data {
  42         struct ldlm_lock *lwd_lock;
  43         __u32             lwd_conn_cnt;
  44 };
  45
  46 int ldlm_expired_completion_wait(void *data)
  47 {
  48         struct lock_wait_data *lwd = data;
  49         struct ldlm_lock *lock = lwd->lwd_lock;
  50         struct obd_import *imp;
  51         struct obd_device *obd;
  52
  53         ENTRY;
  54         if (lock->l_conn_export == NULL) {
  55                 static cfs_time_t next_dump = 0, last_dump = 0;
  56
  57                 if (ptlrpc_check_suspend())
  58                         RETURN(0);
  59
  60                 LDLM_ERROR(lock, "lock timed out (enqueued at %lu, %lus ago); "
  61                            "not entering recovery in server code, just going "
  62                            "back to sleep", lock->l_enqueued_time.tv_sec,
  63                            CURRENT_SECONDS - lock->l_enqueued_time.tv_sec);
  64                 if (cfs_time_after(cfs_time_current(), next_dump)) {
  65                         last_dump = next_dump;
  66                         next_dump = cfs_time_shift(300);
  67                         ldlm_namespace_dump(D_DLMTRACE,
  68                                             lock->l_resource->lr_namespace);
  69                         if (last_dump == 0)
  70                                 libcfs_debug_dumplog();
  71                 }
  72                 RETURN(0);
  73         }
  74
  75         obd = lock->l_conn_export->exp_obd;
  76         imp = obd->u.cli.cl_import;
  77         ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
  78         LDLM_ERROR(lock, "lock timed out (enqueued at %lu, %lus ago), entering "
  79                    "recovery for %s@%s", lock->l_enqueued_time.tv_sec,
  80                    CURRENT_SECONDS - lock->l_enqueued_time.tv_sec,
  81                    obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
  82
  83         RETURN(0);
  84 }
  85
  86 int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data)
  87 {
  88         /* XXX ALLOCATE - 160 bytes */
  89         struct lock_wait_data lwd;
  90         struct obd_device *obd;
  91         struct obd_import *imp = NULL;
  92         struct l_wait_info lwi;
  93         int rc = 0;
  94         ENTRY;
  95
  96         if (flags == LDLM_FL_WAIT_NOREPROC) {
  97                 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
  98                 goto noreproc;
  99         }
 100
 101         if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
 102                        LDLM_FL_BLOCK_CONV))) {
 103                 cfs_waitq_signal(&lock->l_waitq);
 104                 RETURN(0);
 105         }
 106
 107         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
 108                    "sleeping");
 109         ldlm_lock_dump(D_OTHER, lock, 0);
 110         ldlm_reprocess_all(lock->l_resource);
 111
 112 noreproc:
 113
 114         obd = class_exp2obd(lock->l_conn_export);
 115
 116         /* if this is a local lock, then there is no import */
 117         if (obd != NULL)
 118                 imp = obd->u.cli.cl_import;
 119
 120         lwd.lwd_lock = lock;
 121
 122         if (lock->l_flags & LDLM_FL_NO_TIMEOUT) {
 123                 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
 124                 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
 125         } else {
 126                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
 127                                        ldlm_expired_completion_wait,
 128                                        interrupted_completion_wait, &lwd);
 129         }
 130
 131         if (imp != NULL) {
 132                 spin_lock(&imp->imp_lock);
 133                 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
 134                 spin_unlock(&imp->imp_lock);
 135         }
 136
 137         /* Go to sleep until the lock is granted or cancelled. */
 138         rc = l_wait_event(lock->l_waitq,
 139                           ((lock->l_req_mode == lock->l_granted_mode) ||
 140                            (lock->l_flags & LDLM_FL_FAILED)), &lwi);
 141
 142         if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED) {
 143                 LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed");
 144                 RETURN(-EIO);
 145         }
 146
 147         if (rc) {
 148                 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
 149                            rc);
 150                 RETURN(rc);
 151         }
 152
 153         LDLM_DEBUG(lock, "client-side enqueue waking up: granted");
 154         RETURN(0);
 155 }
 156
 157 /*
 158  * ->l_blocking_ast() callback for LDLM locks acquired by server-side OBDs.
 159  */
 160 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 161                       void *data, int flag)
 162 {
 163         int do_ast;
 164         ENTRY;
 165
 166         if (flag == LDLM_CB_CANCELING) {
 167                 /* Don't need to do anything here. */
 168                 RETURN(0);
 169         }
 170
 171         lock_res_and_lock(lock);
 172         /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
 173          * that ldlm_blocking_ast is called just before intent_policy method
 174          * takes the ns_lock, then by the time we get the lock, we might not
 175          * be the correct blocking function anymore.  So check, and return
 176          * early, if so. */
 177         if (lock->l_blocking_ast != ldlm_blocking_ast) {
 178                 unlock_res_and_lock(lock);
 179                 RETURN(0);
 180         }
 181
 182         lock->l_flags |= LDLM_FL_CBPENDING;
 183         do_ast = (!lock->l_readers && !lock->l_writers);
 184         unlock_res_and_lock(lock);
 185
 186         if (do_ast) {
 187                 struct lustre_handle lockh;
 188                 int rc;
 189
 190                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
 191                 ldlm_lock2handle(lock, &lockh);
 192                 rc = ldlm_cli_cancel(&lockh);
 193                 if (rc < 0)
 194                         CERROR("ldlm_cli_cancel: %d\n", rc);
 195         } else {
 196                 LDLM_DEBUG(lock, "Lock still has references, will be "
 197                            "cancelled later");
 198         }
 199         RETURN(0);
 200 }
 201
 202 /*
 203  * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See
 204  * comment in filter_intent_policy() on why you may need this.
 205  */
 206 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
 207 {
 208         /*
 209          * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for
 210          * that is rather subtle: with OST-side locking, it may so happen that
 211          * _all_ extent locks are held by the OST. If client wants to obtain
 212          * current file size it calls ll{,u}_glimpse_size(), and (as locks are
 213          * on the server), dummy glimpse callback fires and does
 214          * nothing. Client still receives correct file size due to the
 215          * following fragment in filter_intent_policy():
 216          *
 217          * rc = l->l_glimpse_ast(l, NULL); // this will update the LVB
 218          * if (rc != 0 && res->lr_namespace->ns_lvbo &&
 219          *     res->lr_namespace->ns_lvbo->lvbo_update) {
 220          *         res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1);
 221          * }
 222          *
 223          * that is, after glimpse_ast() fails, filter_lvbo_update() runs, and
 224          * returns correct file size to the client.
 225          */
 226         return -ELDLM_NO_LOCK_DATA;
 227 }
 228
 229 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
 230                            const struct ldlm_res_id *res_id,
 231                            ldlm_type_t type, ldlm_policy_data_t *policy,
 232                            ldlm_mode_t mode, int *flags,
 233                            ldlm_blocking_callback blocking,
 234                            ldlm_completion_callback completion,
 235                            ldlm_glimpse_callback glimpse,
 236                            void *data, __u32 lvb_len, void *lvb_swabber,
 237                            struct lustre_handle *lockh)
 238 {
 239         struct ldlm_lock *lock;
 240         int err;
 241         ENTRY;
 242
 243         LASSERT(!(*flags & LDLM_FL_REPLAY));
 244         if (unlikely(ns->ns_client)) {
 245                 CERROR("Trying to enqueue local lock in a shadow namespace\n");
 246                 LBUG();
 247         }
 248
 249         lock = ldlm_lock_create(ns, res_id, type, mode, blocking,
 250                                 completion, glimpse, data, lvb_len);
 251         if (unlikely(!lock))
 252                 GOTO(out_nolock, err = -ENOMEM);
 253         LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
 254
 255         ldlm_lock_addref_internal(lock, mode);
 256         ldlm_lock2handle(lock, lockh);
 257         lock_res_and_lock(lock);
 258         lock->l_flags |= LDLM_FL_LOCAL;
 259         if (*flags & LDLM_FL_ATOMIC_CB)
 260                 lock->l_flags |= LDLM_FL_ATOMIC_CB;
 261         lock->l_lvb_swabber = lvb_swabber;
 262         unlock_res_and_lock(lock);
 263         if (policy != NULL)
 264                 lock->l_policy_data = *policy;
 265         if (type == LDLM_EXTENT)
 266                 lock->l_req_extent = policy->l_extent;
 267
 268         err = ldlm_lock_enqueue(ns, &lock, policy, flags);
 269         if (unlikely(err != ELDLM_OK))
 270                 GOTO(out, err);
 271
 272         if (policy != NULL)
 273                 *policy = lock->l_policy_data;
 274
 275         LDLM_DEBUG_NOLOCK("client-side local enqueue handler END (lock %p)",
 276                           lock);
 277
 278         if (lock->l_completion_ast)
 279                 lock->l_completion_ast(lock, *flags, NULL);
 280
 281         LDLM_DEBUG(lock, "client-side local enqueue END");
 282         EXIT;
 283  out:
 284         LDLM_LOCK_PUT(lock);
 285  out_nolock:
 286         return err;
 287 }
 288
 289 static void failed_lock_cleanup(struct ldlm_namespace *ns,
 290                                 struct ldlm_lock *lock,
 291                                 struct lustre_handle *lockh, int mode)
 292 {
 293         /* Set a flag to prevent us from sending a CANCEL (bug 407) */
 294         lock_res_and_lock(lock);
 295         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 296         unlock_res_and_lock(lock);
 297         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 298
 299         ldlm_lock_decref_and_cancel(lockh, mode);
 300
 301         /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
 302          *       from llite/file.c/ll_file_flock(). */
 303         if (lock->l_resource->lr_type == LDLM_FLOCK) {
 304                 ldlm_lock_destroy(lock);
 305         }
 306 }
 307
 308 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
 309                           ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
 310                           int *flags, void *lvb, __u32 lvb_len,
 311                           void *lvb_swabber, struct lustre_handle *lockh,int rc)
 312 {
 313         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 314         int is_replay = *flags & LDLM_FL_REPLAY;
 315         struct ldlm_lock *lock;
 316         struct ldlm_reply *reply;
 317         int cleanup_phase = 1;
 318         ENTRY;
 319
 320         lock = ldlm_handle2lock(lockh);
 321         /* ldlm_cli_enqueue is holding a reference on this lock. */
 322         LASSERT(lock != NULL);
 323         if (rc != ELDLM_OK) {
 324                 LASSERT(!is_replay);
 325                 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
 326                            rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
 327                 if (rc == ELDLM_LOCK_ABORTED) {
 328                         /* Before we return, swab the reply */
 329                         reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF,
 330                                                    sizeof(*reply),
 331                                                    lustre_swab_ldlm_reply);
 332                         if (reply == NULL) {
 333                                 CERROR("Can't unpack ldlm_reply\n");
 334                                 rc = -EPROTO;
 335                         }
 336                         if (lvb_len) {
 337                                 void *tmplvb;
 338                                 tmplvb = lustre_swab_repbuf(req,
 339                                                             DLM_REPLY_REC_OFF,
 340                                                             lvb_len,
 341                                                             lvb_swabber);
 342                                 if (tmplvb == NULL)
 343                                         GOTO(cleanup, rc = -EPROTO);
 344                                 if (lvb != NULL)
 345                                         memcpy(lvb, tmplvb, lvb_len);
 346                         }
 347                 }
 348                 GOTO(cleanup, rc);
 349         }
 350
 351         reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF, sizeof(*reply),
 352                                    lustre_swab_ldlm_reply);
 353         if (reply == NULL) {
 354                 CERROR("Can't unpack ldlm_reply\n");
 355                 GOTO(cleanup, rc = -EPROTO);
 356         }
 357
 358         /* lock enqueued on the server */
 359         cleanup_phase = 0;
 360
 361         lock_res_and_lock(lock);
 362         lock->l_remote_handle = reply->lock_handle;
 363         *flags = reply->lock_flags;
 364         lock->l_flags |= reply->lock_flags & LDLM_INHERIT_FLAGS;
 365         /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match()
 366          * to wait with no timeout as well */
 367         lock->l_flags |= reply->lock_flags & LDLM_FL_NO_TIMEOUT;
 368         unlock_res_and_lock(lock);
 369
 370         CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%x\n",
 371                lock, reply->lock_handle.cookie, *flags);
 372
 373         /* If enqueue returned a blocked lock but the completion handler has
 374          * already run, then it fixed up the resource and we don't need to do it
 375          * again. */
 376         if ((*flags) & LDLM_FL_LOCK_CHANGED) {
 377                 int newmode = reply->lock_desc.l_req_mode;
 378                 LASSERT(!is_replay);
 379                 if (newmode && newmode != lock->l_req_mode) {
 380                         LDLM_DEBUG(lock, "server returned different mode %s",
 381                                    ldlm_lockname[newmode]);
 382                         lock->l_req_mode = newmode;
 383                 }
 384
 385                 if (memcmp(reply->lock_desc.l_resource.lr_name.name,
 386                           lock->l_resource->lr_name.name,
 387                           sizeof(struct ldlm_res_id))) {
 388                         CDEBUG(D_INFO, "remote intent success, locking "
 389                                         "(%ld,%ld,%ld) instead of "
 390                                         "(%ld,%ld,%ld)\n",
 391                               (long)reply->lock_desc.l_resource.lr_name.name[0],
 392                               (long)reply->lock_desc.l_resource.lr_name.name[1],
 393                               (long)reply->lock_desc.l_resource.lr_name.name[2],
 394                               (long)lock->l_resource->lr_name.name[0],
 395                               (long)lock->l_resource->lr_name.name[1],
 396                               (long)lock->l_resource->lr_name.name[2]);
 397
 398                         ldlm_lock_change_resource(ns, lock,
 399                                           &reply->lock_desc.l_resource.lr_name);
 400                         if (lock->l_resource == NULL) {
 401                                 LBUG();
 402                                 GOTO(cleanup, rc = -ENOMEM);
 403                         }
 404                         LDLM_DEBUG(lock, "client-side enqueue, new resource");
 405                 }
 406                 if (with_policy)
 407                         if (!(type == LDLM_IBITS && !(exp->exp_connect_flags &
 408                                                     OBD_CONNECT_IBITS)))
 409                                 lock->l_policy_data =
 410                                                  reply->lock_desc.l_policy_data;
 411                 if (type != LDLM_PLAIN)
 412                         LDLM_DEBUG(lock,"client-side enqueue, new policy data");
 413         }
 414
 415         if ((*flags) & LDLM_FL_AST_SENT ||
 416             /* Cancel extent locks as soon as possible on a liblustre client,
 417              * because it cannot handle asynchronous ASTs robustly (see
 418              * bug 7311). */
 419             (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) {
 420                 lock_res_and_lock(lock);
 421                 lock->l_flags |= LDLM_FL_CBPENDING;
 422                 unlock_res_and_lock(lock);
 423                 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
 424         }
 425
 426         /* If the lock has already been granted by a completion AST, don't
 427          * clobber the LVB with an older one. */
 428         if (lvb_len && (lock->l_req_mode != lock->l_granted_mode)) {
 429                 void *tmplvb;
 430                 tmplvb = lustre_swab_repbuf(req, DLM_REPLY_REC_OFF, lvb_len,
 431                                             lvb_swabber);
 432                 if (tmplvb == NULL)
 433                         GOTO(cleanup, rc = -EPROTO);
 434                 memcpy(lock->l_lvb_data, tmplvb, lvb_len);
 435         }
 436
 437         if (!is_replay) {
 438                 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
 439                 if (lock->l_completion_ast != NULL) {
 440                         int err = lock->l_completion_ast(lock, *flags, NULL);
 441                         if (!rc)
 442                                 rc = err;
 443                         if (rc && type != LDLM_FLOCK) /* bug 9425, bug 10250 */
 444                                 cleanup_phase = 1;
 445                 }
 446         }
 447
 448         if (lvb_len && lvb != NULL) {
 449                 /* Copy the LVB here, and not earlier, because the completion
 450                  * AST (if any) can override what we got in the reply */
 451                 memcpy(lvb, lock->l_lvb_data, lvb_len);
 452         }
 453
 454         LDLM_DEBUG(lock, "client-side enqueue END");
 455         EXIT;
 456 cleanup:
 457         if (cleanup_phase == 1 && rc)
 458                 failed_lock_cleanup(ns, lock, lockh, mode);
 459         /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
 460         LDLM_LOCK_PUT(lock);
 461         LDLM_LOCK_PUT(lock);
 462         return rc;
 463 }
 464
 465 /* PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
 466  * a single page on the send/receive side. XXX: 512 should be changed
 467  * to more adequate value. */
 468 #define ldlm_req_handles_avail(exp, size, bufcount, off)                \
 469 ({                                                                      \
 470         int _avail = min_t(int, LDLM_MAXREQSIZE, PAGE_SIZE - 512);      \
 471         int _s = size[DLM_LOCKREQ_OFF];                                 \
 472         size[DLM_LOCKREQ_OFF] = sizeof(struct ldlm_request);            \
 473         _avail -= lustre_msg_size(class_exp2cliimp(exp)->imp_msg_magic, \
 474                                 bufcount, size);                        \
 475         _avail /= sizeof(struct lustre_handle);                         \
 476         _avail += LDLM_LOCKREQ_HANDLES - off;                           \
 477         size[DLM_LOCKREQ_OFF] = _s;                                     \
 478         _avail;                                                         \
 479 })
 480
 481 /* Cancel lru locks and pack them into the enqueue request. Pack there the given
 482  * @count locks in @cancel. */
 483 struct ptlrpc_request *ldlm_prep_enqueue_req(struct obd_export *exp,
 484                                              int bufcount, int *size,
 485                                              struct list_head *cancels,
 486                                              int count)
 487 {
 488         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 489         struct ldlm_request *dlm = NULL;
 490         struct ptlrpc_request *req;
 491         CFS_LIST_HEAD(head);
 492         ENTRY;
 493
 494         if (cancels == NULL)
 495                 cancels = &head;
 496         if (exp_connect_cancelset(exp)) {
 497                 /* Estimate the amount of available space in the request. */
 498                 int avail = ldlm_req_handles_avail(exp, size, bufcount,
 499                                                    LDLM_ENQUEUE_CANCEL_OFF);
 500                 LASSERT(avail >= count);
 501
 502                 /* Cancel lru locks here _only_ if the server supports
 503                  * EARLY_CANCEL. Otherwise we have to send extra CANCEL
 504                  * rpc right on enqueue, what will make it slower, vs.
 505                  * asynchronous rpc in blocking thread. */
 506                 count += ldlm_cancel_lru_local(ns, cancels, 1, avail - count,
 507                                                LDLM_CANCEL_AGED);
 508                 size[DLM_LOCKREQ_OFF] =
 509                         ldlm_request_bufsize(count, LDLM_ENQUEUE);
 510         }
 511         req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
 512                               LDLM_ENQUEUE, bufcount, size, NULL);
 513         if (exp_connect_cancelset(exp) && req) {
 514                 dlm = lustre_msg_buf(req->rq_reqmsg,
 515                                      DLM_LOCKREQ_OFF, sizeof(*dlm));
 516                 /* Skip first lock handler in ldlm_request_pack(), this method
 517                  * will incrment @lock_count according to the lock handle amount
 518                  * actually written to the buffer. */
 519                 dlm->lock_count = LDLM_ENQUEUE_CANCEL_OFF;
 520                 ldlm_cli_cancel_list(cancels, count, req, DLM_LOCKREQ_OFF, 0);
 521         } else {
 522                 ldlm_lock_list_put(cancels, l_bl_ast, count);
 523         }
 524         RETURN(req);
 525 }
 526
 527 /* If a request has some specific initialisation it is passed in @reqp,
 528  * otherwise it is created in ldlm_cli_enqueue.
 529  *
 530  * Supports sync and async requests, pass @async flag accordingly. If a
 531  * request was created in ldlm_cli_enqueue and it is the async request,
 532  * pass it to the caller in @reqp. */
 533 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
 534                      const struct ldlm_res_id *res_id,
 535                      ldlm_type_t type, ldlm_policy_data_t *policy,
 536                      ldlm_mode_t mode, int *flags,
 537                      ldlm_blocking_callback blocking,
 538                      ldlm_completion_callback completion,
 539                      ldlm_glimpse_callback glimpse,
 540                      void *data, void *lvb, __u32 lvb_len, void *lvb_swabber,
 541                      struct lustre_handle *lockh, int async)
 542 {
 543         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 544         struct ldlm_lock *lock;
 545         struct ldlm_request *body;
 546         struct ldlm_reply *reply;
 547         int size[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
 548                         [DLM_LOCKREQ_OFF]     = sizeof(*body),
 549                         [DLM_REPLY_REC_OFF]   = lvb_len };
 550         int is_replay = *flags & LDLM_FL_REPLAY;
 551         int req_passed_in = 1, rc;
 552         struct ptlrpc_request *req;
 553         ENTRY;
 554
 555         LASSERT(exp != NULL);
 556
 557         /* If we're replaying this lock, just check some invariants.
 558          * If we're creating a new lock, get everything all setup nice. */
 559         if (is_replay) {
 560                 lock = ldlm_handle2lock(lockh);
 561                 LASSERT(lock != NULL);
 562                 LDLM_DEBUG(lock, "client-side enqueue START");
 563                 LASSERT(exp == lock->l_conn_export);
 564         } else {
 565                 lock = ldlm_lock_create(ns, res_id, type, mode, blocking,
 566                                         completion, glimpse, data, lvb_len);
 567                 if (lock == NULL)
 568                         RETURN(-ENOMEM);
 569                 /* for the local lock, add the reference */
 570                 ldlm_lock_addref_internal(lock, mode);
 571                 ldlm_lock2handle(lock, lockh);
 572                 lock->l_lvb_swabber = lvb_swabber;
 573                 if (policy != NULL) {
 574                         /* INODEBITS_INTEROP: If the server does not support
 575                          * inodebits, we will request a plain lock in the
 576                          * descriptor (ldlm_lock2desc() below) but use an
 577                          * inodebits lock internally with both bits set.
 578                          */
 579                         if (type == LDLM_IBITS && !(exp->exp_connect_flags &
 580                                                     OBD_CONNECT_IBITS))
 581                                 lock->l_policy_data.l_inodebits.bits =
 582                                         MDS_INODELOCK_LOOKUP |
 583                                         MDS_INODELOCK_UPDATE;
 584                         else
 585                                 lock->l_policy_data = *policy;
 586                 }
 587
 588                 if (type == LDLM_EXTENT)
 589                         lock->l_req_extent = policy->l_extent;
 590                 LDLM_DEBUG(lock, "client-side enqueue START");
 591         }
 592
 593         /* lock not sent to server yet */
 594
 595         if (reqp == NULL || *reqp == NULL) {
 596                 req = ldlm_prep_enqueue_req(exp, 2, size, NULL, 0);
 597                 if (req == NULL) {
 598                         failed_lock_cleanup(ns, lock, lockh, mode);
 599                         LDLM_LOCK_PUT(lock);
 600                         RETURN(-ENOMEM);
 601                 }
 602                 req_passed_in = 0;
 603                 if (reqp)
 604                         *reqp = req;
 605         } else {
 606                 req = *reqp;
 607                 LASSERTF(lustre_msg_buflen(req->rq_reqmsg, DLM_LOCKREQ_OFF) >=
 608                          sizeof(*body), "buflen[%d] = %d, not "LPSZ"\n",
 609                          DLM_LOCKREQ_OFF,
 610                          lustre_msg_buflen(req->rq_reqmsg, DLM_LOCKREQ_OFF),
 611                          sizeof(*body));
 612         }
 613
 614         lock->l_conn_export = exp;
 615         lock->l_export = NULL;
 616         lock->l_blocking_ast = blocking;
 617
 618         /* Dump lock data into the request buffer */
 619         body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, sizeof(*body));
 620         ldlm_lock2desc(lock, &body->lock_desc);
 621         body->lock_flags = *flags;
 622         body->lock_handle[0] = *lockh;
 623
 624         /* Continue as normal. */
 625         if (!req_passed_in) {
 626                 size[DLM_LOCKREPLY_OFF] = sizeof(*reply);
 627                 ptlrpc_req_set_repsize(req, 2 + (lvb_len > 0), size);
 628         }
 629
 630         /*
 631          * Liblustre client doesn't get extent locks, except for O_APPEND case
 632          * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
 633          * [i_size, OBD_OBJECT_EOF] lock is taken.
 634          */
 635         LASSERT(ergo(LIBLUSTRE_CLIENT, type != LDLM_EXTENT ||
 636                      policy->l_extent.end == OBD_OBJECT_EOF));
 637
 638         if (async) {
 639                 LASSERT(reqp != NULL);
 640                 RETURN(0);
 641         }
 642
 643         LDLM_DEBUG(lock, "sending request");
 644         rc = ptlrpc_queue_wait(req);
 645         rc = ldlm_cli_enqueue_fini(exp, req, type, policy ? 1 : 0,
 646                                    mode, flags, lvb, lvb_len, lvb_swabber,
 647                                    lockh, rc);
 648
 649         if (!req_passed_in && req != NULL) {
 650                 ptlrpc_req_finished(req);
 651                 if (reqp)
 652                         *reqp = NULL;
 653         }
 654
 655         RETURN(rc);
 656 }
 657
 658 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
 659                                   int *flags)
 660 {
 661         struct ldlm_resource *res;
 662         int rc;
 663         ENTRY;
 664         if (lock->l_resource->lr_namespace->ns_client) {
 665                 CERROR("Trying to cancel local lock\n");
 666                 LBUG();
 667         }
 668         LDLM_DEBUG(lock, "client-side local convert");
 669
 670         res = ldlm_lock_convert(lock, new_mode, flags);
 671         if (res) {
 672                 ldlm_reprocess_all(res);
 673                 rc = 0;
 674         } else {
 675                 rc = EDEADLOCK;
 676         }
 677         LDLM_DEBUG(lock, "client-side local convert handler END");
 678         LDLM_LOCK_PUT(lock);
 679         RETURN(rc);
 680 }
 681
 682 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
 683  * conversion of locks which are on the waiting or converting queue */
 684 /* Caller of this code is supposed to take care of lock readers/writers
 685    accounting */
 686 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, int *flags)
 687 {
 688         struct ldlm_request *body;
 689         struct ldlm_reply *reply;
 690         struct ldlm_lock *lock;
 691         struct ldlm_resource *res;
 692         struct ptlrpc_request *req;
 693         int size[2] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
 694                         [DLM_LOCKREQ_OFF]     = sizeof(*body) };
 695         int rc;
 696         ENTRY;
 697
 698         lock = ldlm_handle2lock(lockh);
 699         if (!lock) {
 700                 LBUG();
 701                 RETURN(-EINVAL);
 702         }
 703         *flags = 0;
 704
 705         if (lock->l_conn_export == NULL)
 706                 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
 707
 708         LDLM_DEBUG(lock, "client-side convert");
 709
 710         req = ptlrpc_prep_req(class_exp2cliimp(lock->l_conn_export),
 711                               LUSTRE_DLM_VERSION, LDLM_CONVERT, 2, size, NULL);
 712         if (!req)
 713                 GOTO(out, rc = -ENOMEM);
 714
 715         body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, sizeof(*body));
 716         body->lock_handle[0] = lock->l_remote_handle;
 717
 718         body->lock_desc.l_req_mode = new_mode;
 719         body->lock_flags = *flags;
 720
 721         size[DLM_LOCKREPLY_OFF] = sizeof(*reply);
 722         ptlrpc_req_set_repsize(req, 2, size);
 723
 724         rc = ptlrpc_queue_wait(req);
 725         if (rc != ELDLM_OK)
 726                 GOTO(out, rc);
 727
 728         reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF, sizeof(*reply),
 729                                    lustre_swab_ldlm_reply);
 730         if (reply == NULL) {
 731                 CERROR ("Can't unpack ldlm_reply\n");
 732                 GOTO (out, rc = -EPROTO);
 733         }
 734
 735         if (req->rq_status)
 736                 GOTO(out, rc = req->rq_status);
 737
 738         res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
 739         if (res != NULL) {
 740                 ldlm_reprocess_all(res);
 741                 /* Go to sleep until the lock is granted. */
 742                 /* FIXME: or cancelled. */
 743                 if (lock->l_completion_ast) {
 744                         rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
 745                                                     NULL);
 746                         if (rc)
 747                                 GOTO(out, rc);
 748                 }
 749         } else {
 750                 rc = EDEADLOCK;
 751         }
 752         EXIT;
 753  out:
 754         LDLM_LOCK_PUT(lock);
 755         ptlrpc_req_finished(req);
 756         return rc;
 757 }
 758
 759 /* Cancel locks locally.
 760  * Returns: 1 if there is a need to send a cancel RPC to server. 0 otherwise. */
 761 static int ldlm_cli_cancel_local(struct ldlm_lock *lock)
 762 {
 763         int rc = 0;
 764         ENTRY;
 765
 766         if (lock->l_conn_export) {
 767                 int local_only;
 768
 769                 LDLM_DEBUG(lock, "client-side cancel");
 770                 /* Set this flag to prevent others from getting new references*/
 771                 lock_res_and_lock(lock);
 772                 lock->l_flags |= LDLM_FL_CBPENDING;
 773                 local_only = (lock->l_flags &
 774                               (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
 775                 ldlm_cancel_callback(lock);
 776                 unlock_res_and_lock(lock);
 777
 778                 if (local_only)
 779                         CDEBUG(D_INFO, "not sending request (at caller's "
 780                                "instruction)\n");
 781                 else
 782                         rc = 1;
 783
 784                 ldlm_lock_cancel(lock);
 785         } else {
 786                 if (lock->l_resource->lr_namespace->ns_client) {
 787                         LDLM_ERROR(lock, "Trying to cancel local lock");
 788                         LBUG();
 789                 }
 790                 LDLM_DEBUG(lock, "server-side local cancel");
 791                 ldlm_lock_cancel(lock);
 792                 ldlm_reprocess_all(lock->l_resource);
 793                 LDLM_DEBUG(lock, "server-side local cancel handler END");
 794         }
 795
 796         RETURN(rc);
 797 }
 798
 799 /* Pack @count locks in @head into ldlm_request buffer at the offset @off,
 800    of the request @req. */
 801 static void ldlm_cancel_pack(struct ptlrpc_request *req, int off,
 802                              struct list_head *head, int count)
 803 {
 804         struct ldlm_request *dlm;
 805         struct ldlm_lock *lock;
 806         int max;
 807         ENTRY;
 808
 809         dlm = lustre_msg_buf(req->rq_reqmsg, off, sizeof(*dlm));
 810         LASSERT(dlm != NULL);
 811
 812         /* Check the room in the request buffer. */
 813         max = lustre_msg_buflen(req->rq_reqmsg, off) -
 814                 sizeof(struct ldlm_request);
 815         max /= sizeof(struct lustre_handle);
 816         max += LDLM_LOCKREQ_HANDLES;
 817         LASSERT(max >= dlm->lock_count + count);
 818
 819         /* XXX: it would be better to pack lock handles grouped by resource.
 820          * so that the server cancel would call filter_lvbo_update() less
 821          * frequently. */
 822         list_for_each_entry(lock, head, l_bl_ast) {
 823                 if (!count--)
 824                         break;
 825                 /* Pack the lock handle to the given request buffer. */
 826                 LASSERT(lock->l_conn_export);
 827                 /* Cannot be set on a lock in a resource granted list.*/
 828                 LASSERT(!(lock->l_flags &
 829                           (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK)));
 830                 /* If @lock is marked CANCEL_ON_BLOCK, cancel
 831                  * will not be sent in ldlm_cli_cancel(). It
 832                  * is used for liblustre clients, no cancel on
 833                  * block requests. However, even for liblustre
 834                  * clients, when the flag is set, batched cancel
 835                  * should be sent (what if no block rpc has
 836                  * come). To not send another separated rpc in
 837                  * this case, the caller pass CANCEL_ON_BLOCK
 838                  * flag to ldlm_cli_cancel_unused_resource(). */
 839                 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
 840         }
 841         EXIT;
 842 }
 843
 844 /* Prepare and send a batched cancel rpc, it will include count lock handles
 845  * of locks given in @head. */
 846 int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
 847                         int count, int flags)
 848 {
 849         struct ptlrpc_request *req = NULL;
 850         struct ldlm_request *body;
 851         int size[2] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
 852                 [DLM_LOCKREQ_OFF]     = sizeof(*body) };
 853         struct obd_import *imp;
 854         int free, sent = 0;
 855         int rc = 0;
 856         ENTRY;
 857
 858         LASSERT(exp != NULL);
 859         LASSERT(count > 0);
 860
 861         free = ldlm_req_handles_avail(exp, size, 2, 0);
 862         if (count > free)
 863                 count = free;
 864
 865         size[DLM_LOCKREQ_OFF] = ldlm_request_bufsize(count, LDLM_CANCEL);
 866         while (1) {
 867                 imp = class_exp2cliimp(exp);
 868                 if (imp == NULL || imp->imp_invalid) {
 869                         CDEBUG(D_HA, "skipping cancel on invalid import %p\n",
 870                                imp);
 871                         break;
 872                 }
 873
 874                 req = ptlrpc_prep_req(imp, LUSTRE_DLM_VERSION, LDLM_CANCEL, 2,
 875                                       size, NULL);
 876                 if (!req)
 877                         GOTO(out, rc = -ENOMEM);
 878
 879                 req->rq_no_resend = 1;
 880
 881                 /* XXX FIXME bug 249 */
 882                 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
 883                 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
 884
 885                 body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF,
 886                                       sizeof(*body));
 887                 ldlm_cancel_pack(req, DLM_LOCKREQ_OFF, cancels, count);
 888
 889                 ptlrpc_req_set_repsize(req, 1, NULL);
 890                 if (flags & LDLM_FL_ASYNC) {
 891                         ptlrpcd_add_req(req);
 892                         sent = count;
 893                         GOTO(out, 0);
 894                 } else {
 895                         rc = ptlrpc_queue_wait(req);
 896                 }
 897                 if (rc == ESTALE) {
 898                         CDEBUG(D_DLMTRACE, "client/server (nid %s) "
 899                                "out of sync -- not fatal\n",
 900                                libcfs_nid2str(req->rq_import->
 901                                               imp_connection->c_peer.nid));
 902                 } else if (rc == -ETIMEDOUT) {
 903                         ptlrpc_req_finished(req);
 904                         continue;
 905                 } else if (rc != ELDLM_OK) {
 906                         CERROR("Got rc %d from cancel RPC: canceling "
 907                                "anyway\n", rc);
 908                         break;
 909                 }
 910                 sent = count;
 911                 break;
 912         }
 913
 914         ptlrpc_req_finished(req);
 915         EXIT;
 916 out:
 917         return sent ? sent : rc;
 918 }
 919
 920 int ldlm_cli_cancel(struct lustre_handle *lockh)
 921 {
 922         struct ldlm_lock *lock;
 923         CFS_LIST_HEAD(head);
 924         int rc = 0;
 925         ENTRY;
 926
 927         /* concurrent cancels on the same handle can happen */
 928         lock = __ldlm_handle2lock(lockh, LDLM_FL_CANCELING);
 929         if (lock == NULL)
 930                 RETURN(0);
 931
 932         rc = ldlm_cli_cancel_local(lock);
 933         if (rc <= 0)
 934                 GOTO(out, rc);
 935
 936         list_add(&lock->l_bl_ast, &head);
 937         rc = ldlm_cli_cancel_req(lock->l_conn_export, &head, 1, 0);
 938         EXIT;
 939 out:
 940         LDLM_LOCK_PUT(lock);
 941         return rc < 0 ? rc : 0;
 942 }
 943
 944 /* - Free space in lru for @count new locks,
 945  *   redundant unused locks are canceled locally;
 946  * - also cancel locally unused aged locks;
 947  * - do not cancel more than @max locks;
 948  * - GET the found locks and add them into the @cancels list.
 949  *
 950  * A client lock can be added to the l_bl_ast list only when it is
 951  * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing CANCEL.
 952  * There are the following use cases: ldlm_cancel_resource_local(),
 953  * ldlm_cancel_lru_local() and ldlm_cli_cancel(), which check&set this
 954  * flag properly. As any attempt to cancel a lock rely on this flag,
 955  * l_bl_ast list is accessed later without any special locking. */
 956 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
 957                           int count, int max, int flags)
 958 {
 959         cfs_time_t cur = cfs_time_current();
 960         struct ldlm_lock *lock, *next;
 961         int rc, added = 0, left;
 962         ENTRY;
 963
 964         spin_lock(&ns->ns_unused_lock);
 965         count += ns->ns_nr_unused - ns->ns_max_unused;
 966         while (!list_empty(&ns->ns_unused_list)) {
 967                 struct list_head *tmp = ns->ns_unused_list.next;
 968                 lock = list_entry(tmp, struct ldlm_lock, l_lru);
 969
 970                 if (max && added >= max)
 971                         break;
 972
 973                 if ((added >= count) &&
 974                     (!(flags & LDLM_CANCEL_AGED) ||
 975                      cfs_time_before_64(cur, ns->ns_max_age +
 976                                         lock->l_last_used)))
 977                         break;
 978
 979                 LDLM_LOCK_GET(lock); /* dropped by bl thread */
 980                 spin_unlock(&ns->ns_unused_lock);
 981
 982                 lock_res_and_lock(lock);
 983                 if ((ldlm_lock_remove_from_lru(lock) == 0) ||
 984                     (lock->l_flags & LDLM_FL_CANCELING)) {
 985                         /* other thread is removing lock from lru or
 986                          * somebody is already doing CANCEL. */
 987                         unlock_res_and_lock(lock);
 988                         LDLM_LOCK_PUT(lock);
 989                         spin_lock(&ns->ns_unused_lock);
 990                         continue;
 991                 }
 992                 LASSERT(!lock->l_readers && !lock->l_writers);
 993
 994                 /* If we have chosen to canecl this lock voluntarily, we better
 995                    send cancel notification to server, so that it frees
 996                    appropriate state. This might lead to a race where while
 997                    we are doing cancel here, server is also silently
 998                    cancelling this lock. */
 999                 lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK;
1000
1001                 /* Setting the CBPENDING flag is a little misleading, but
1002                  * prevents an important race; namely, once CBPENDING is set,
1003                  * the lock can accumulate no more readers/writers.  Since
1004                  * readers and writers are already zero here, ldlm_lock_decref
1005                  * won't see this flag and call l_blocking_ast */
1006                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1007                 /* We can't re-add to l_lru as it confuses the refcounting in
1008                  * ldlm_lock_remove_from_lru() if an AST arrives after we drop
1009                  * ns_lock below. We use l_bl_ast and can't use l_pending_chain
1010                  * as it is used both on server and client nevertheles bug 5666
1011                  * says it is used only on server. --umka */
1012
1013                 LASSERT(list_empty(&lock->l_bl_ast));
1014                 list_add(&lock->l_bl_ast, cancels);
1015                 unlock_res_and_lock(lock);
1016                 spin_lock(&ns->ns_unused_lock);
1017                 added++;
1018         }
1019         spin_unlock(&ns->ns_unused_lock);
1020
1021         /* Handle only @added inserted locks. */
1022         left = added;
1023         list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1024                 if (left-- == 0)
1025                         break;
1026                 rc = ldlm_cli_cancel_local(lock);
1027                 if (rc == 0) {
1028                         /* CANCEL RPC should not be sent to server. */
1029                         list_del_init(&lock->l_bl_ast);
1030                         LDLM_LOCK_PUT(lock);
1031                         added--;
1032                 }
1033         }
1034         RETURN(added);
1035 }
1036
1037 /* when called with LDLM_ASYNC the blocking callback will be handled
1038  * in a thread and this function will return after the thread has been
1039  * asked to call the callback.  when called with LDLM_SYNC the blocking
1040  * callback will be performed in this function. */
1041 int ldlm_cancel_lru(struct ldlm_namespace *ns, ldlm_sync_t sync)
1042 {
1043         CFS_LIST_HEAD(cancels);
1044         int count, rc;
1045         ENTRY;
1046
1047 #ifndef __KERNEL__
1048         sync = LDLM_SYNC; /* force to be sync in user space */
1049 #endif
1050         count = ldlm_cancel_lru_local(ns, &cancels, 0, 0, 0);
1051         if (sync == LDLM_ASYNC) {
1052                 struct ldlm_lock *lock, *next;
1053                 list_for_each_entry_safe(lock, next, &cancels, l_bl_ast) {
1054                         /* Remove from the list to allow blocking thread to
1055                          * re-use l_bl_ast. */
1056                         list_del_init(&lock->l_bl_ast);
1057                         rc = ldlm_bl_to_thread(ns, NULL, lock,
1058                                                LDLM_FL_CANCELING);
1059                         if (rc)
1060                                 list_add_tail(&lock->l_bl_ast, &next->l_bl_ast);
1061                 }
1062         }
1063
1064         /* If some locks are left in the list in ASYNC mode, or
1065          * this is SYNC mode, cancel the list. */
1066         ldlm_cli_cancel_list(&cancels, count, NULL, DLM_LOCKREQ_OFF, 0);
1067         RETURN(0);
1068 }
1069
1070 /* Find and cancel locally unused locks found on resource, matched to the
1071  * given policy, mode. GET the found locks and add them into the @cancels
1072  * list. */
1073 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1074                                struct list_head *cancels,
1075                                ldlm_policy_data_t *policy,
1076                                ldlm_mode_t mode, int lock_flags,
1077                                int flags, void *opaque)
1078 {
1079         struct ldlm_lock *lock, *next;
1080         int count = 0, left;
1081         ENTRY;
1082
1083         lock_res(res);
1084         list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1085                 if (opaque != NULL && lock->l_ast_data != opaque) {
1086                         LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1087                                    lock->l_ast_data, opaque);
1088                         //LBUG();
1089                         continue;
1090                 }
1091
1092                 if (lock->l_readers || lock->l_writers) {
1093                         if (flags & LDLM_FL_WARN) {
1094                                 LDLM_ERROR(lock, "lock in use");
1095                                 //LBUG();
1096                         }
1097                         continue;
1098                 }
1099
1100                 if (lockmode_compat(lock->l_granted_mode, mode))
1101                         continue;
1102
1103                 /* If policy is given and this is IBITS lock, add to list only
1104                  * those locks that match by policy. */
1105                 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1106                     !(lock->l_policy_data.l_inodebits.bits &
1107                       policy->l_inodebits.bits))
1108                         continue;
1109
1110                 /* If somebody is already doing CANCEL, skip it. */
1111                 if (lock->l_flags & LDLM_FL_CANCELING)
1112                         continue;
1113
1114                 /* See CBPENDING comment in ldlm_cancel_lru */
1115                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1116                         lock_flags;
1117
1118                 LASSERT(list_empty(&lock->l_bl_ast));
1119                 list_add(&lock->l_bl_ast, cancels);
1120                 LDLM_LOCK_GET(lock);
1121                 count++;
1122         }
1123         unlock_res(res);
1124
1125         /* Handle only @count inserted locks. */
1126         left = count;
1127         list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1128                 int rc = 0;
1129
1130                 if (left-- == 0)
1131                         break;
1132                 if (flags & LDLM_FL_LOCAL_ONLY)
1133                         ldlm_lock_cancel(lock);
1134                 else
1135                         rc = ldlm_cli_cancel_local(lock);
1136
1137                 if (rc == 0) {
1138                         /* CANCEL RPC should not be sent to server. */
1139                         list_del_init(&lock->l_bl_ast);
1140                         LDLM_LOCK_PUT(lock);
1141                         count--;
1142                 }
1143         }
1144         RETURN(count);
1145 }
1146
1147 /* If @req is NULL, send CANCEL request to server with handles of locks
1148  * in the @cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1149  * separately per lock.
1150  * If @req is not NULL, put handles of locks in @cancels into the request
1151  * buffer at the offset @off.
1152  * Destroy @cancels at the end. */
1153 int ldlm_cli_cancel_list(struct list_head *cancels, int count,
1154                          struct ptlrpc_request *req, int off, int flags)
1155 {
1156         struct ldlm_lock *lock;
1157         int res = 0;
1158         ENTRY;
1159
1160         if (list_empty(cancels) || count == 0)
1161                 RETURN(0);
1162
1163         /* XXX: requests (both batched and not) could be sent in parallel.
1164          * Usually it is enough to have just 1 RPC, but it is possible that
1165          * there are to many locks to be cancelled in LRU or on a resource.
1166          * It would also speed up the case when the server does not support
1167          * the feature. */
1168         while (count > 0) {
1169                 LASSERT(!list_empty(cancels));
1170                 lock = list_entry(cancels->next, struct ldlm_lock, l_bl_ast);
1171                 LASSERT(lock->l_conn_export);
1172
1173                 if (exp_connect_cancelset(lock->l_conn_export)) {
1174                         res = count;
1175                         if (req)
1176                                 ldlm_cancel_pack(req, off, cancels, count);
1177                         else
1178                                 res = ldlm_cli_cancel_req(lock->l_conn_export,
1179                                                           cancels, count, flags);
1180                 } else {
1181                         res = ldlm_cli_cancel_req(lock->l_conn_export,
1182                                                   cancels, 1, flags);
1183                 }
1184
1185                 if (res < 0) {
1186                         CERROR("ldlm_cli_cancel_list: %d\n", res);
1187                         res = count;
1188                 }
1189
1190                 count -= res;
1191                 ldlm_lock_list_put(cancels, l_bl_ast, res);
1192         }
1193         LASSERT(list_empty(cancels));
1194         LASSERT(count == 0);
1195         RETURN(0);
1196 }
1197
1198 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1199                                     const struct ldlm_res_id *res_id,
1200                                     ldlm_policy_data_t *policy,
1201                                     int mode, int flags, void *opaque)
1202 {
1203         struct ldlm_resource *res;
1204         CFS_LIST_HEAD(cancels);
1205         int count;
1206         int rc;
1207         ENTRY;
1208
1209         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1210         if (res == NULL) {
1211                 /* This is not a problem. */
1212                 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id->name[0]);
1213                 RETURN(0);
1214         }
1215
1216         count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
1217                                            0, flags, opaque);
1218         rc = ldlm_cli_cancel_list(&cancels, count, NULL,
1219                                   DLM_LOCKREQ_OFF, flags);
1220         if (rc != ELDLM_OK)
1221                 CERROR("ldlm_cli_cancel_unused_resource: %d\n", rc);
1222
1223         ldlm_resource_putref(res);
1224         RETURN(0);
1225 }
1226
1227 static inline int have_no_nsresource(struct ldlm_namespace *ns)
1228 {
1229         int no_resource = 0;
1230
1231         spin_lock(&ns->ns_hash_lock);
1232         if (ns->ns_resources == 0)
1233                 no_resource = 1;
1234         spin_unlock(&ns->ns_hash_lock);
1235
1236         RETURN(no_resource);
1237 }
1238
1239 /* Cancel all locks on a namespace (or a specific resource, if given)
1240  * that have 0 readers/writers.
1241  *
1242  * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
1243  * to notify the server. */
1244 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1245                            const struct ldlm_res_id *res_id,
1246                            int flags, void *opaque)
1247 {
1248         int i;
1249         ENTRY;
1250
1251         if (ns == NULL)
1252                 RETURN(ELDLM_OK);
1253
1254         if (res_id)
1255                 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
1256                                                        LCK_MINMODE, flags,
1257                                                        opaque));
1258
1259         spin_lock(&ns->ns_hash_lock);
1260         for (i = 0; i < RES_HASH_SIZE; i++) {
1261                 struct list_head *tmp;
1262                 tmp = ns->ns_hash[i].next;
1263                 while (tmp != &(ns->ns_hash[i])) {
1264                         struct ldlm_resource *res;
1265                         int rc;
1266
1267                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
1268                         ldlm_resource_getref(res);
1269                         spin_unlock(&ns->ns_hash_lock);
1270
1271                         rc = ldlm_cli_cancel_unused_resource(ns, &res->lr_name,
1272                                                              NULL, LCK_MINMODE,
1273                                                              flags, opaque);
1274
1275                         if (rc)
1276                                 CERROR("ldlm_cli_cancel_unused ("LPU64"): %d\n",
1277                                        res->lr_name.name[0], rc);
1278
1279                         spin_lock(&ns->ns_hash_lock);
1280                         tmp = tmp->next;
1281                         ldlm_resource_putref_locked(res);
1282                 }
1283         }
1284         spin_unlock(&ns->ns_hash_lock);
1285
1286         RETURN(ELDLM_OK);
1287 }
1288
1289 /* join/split resource locks to/from lru list */
1290 int ldlm_cli_join_lru(struct ldlm_namespace *ns,
1291                       const struct ldlm_res_id *res_id, int join)
1292 {
1293         struct ldlm_resource *res;
1294         struct ldlm_lock *lock, *n;
1295         int count = 0;
1296         ENTRY;
1297
1298         LASSERT(ns->ns_client == LDLM_NAMESPACE_CLIENT);
1299
1300         res = ldlm_resource_get(ns, NULL, res_id, LDLM_EXTENT, 0);
1301         if (res == NULL)
1302                 RETURN(count);
1303         LASSERT(res->lr_type == LDLM_EXTENT);
1304
1305         lock_res(res);
1306         if (!join)
1307                 goto split;
1308
1309         list_for_each_entry_safe (lock, n, &res->lr_granted, l_res_link) {
1310                 if (list_empty(&lock->l_lru) &&
1311                     !lock->l_readers && !lock->l_writers &&
1312                     !(lock->l_flags & LDLM_FL_LOCAL) &&
1313                     !(lock->l_flags & LDLM_FL_CBPENDING)) {
1314                         lock->l_last_used = cfs_time_current();
1315                         spin_lock(&ns->ns_unused_lock);
1316                         LASSERT(ns->ns_nr_unused >= 0);
1317                         list_add_tail(&lock->l_lru, &ns->ns_unused_list);
1318                         ns->ns_nr_unused++;
1319                         spin_unlock(&ns->ns_unused_lock);
1320                         lock->l_flags &= ~LDLM_FL_NO_LRU;
1321                         LDLM_DEBUG(lock, "join lock to lru");
1322                         count++;
1323                 }
1324         }
1325         goto unlock;
1326 split:
1327         spin_lock(&ns->ns_unused_lock);
1328         list_for_each_entry_safe (lock, n, &ns->ns_unused_list, l_lru) {
1329                 if (lock->l_resource == res) {
1330                         ldlm_lock_remove_from_lru_nolock(lock);
1331                         lock->l_flags |= LDLM_FL_NO_LRU;
1332                         LDLM_DEBUG(lock, "split lock from lru");
1333                         count++;
1334                 }
1335         }
1336         spin_unlock(&ns->ns_unused_lock);
1337 unlock:
1338         unlock_res(res);
1339         ldlm_resource_putref(res);
1340         RETURN(count);
1341 }
1342
1343 /* Lock iterators. */
1344
1345 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
1346                           void *closure)
1347 {
1348         struct list_head *tmp, *next;
1349         struct ldlm_lock *lock;
1350         int rc = LDLM_ITER_CONTINUE;
1351
1352         ENTRY;
1353
1354         if (!res)
1355                 RETURN(LDLM_ITER_CONTINUE);
1356
1357         lock_res(res);
1358         list_for_each_safe(tmp, next, &res->lr_granted) {
1359                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1360
1361                 if (iter(lock, closure) == LDLM_ITER_STOP)
1362                         GOTO(out, rc = LDLM_ITER_STOP);
1363         }
1364
1365         list_for_each_safe(tmp, next, &res->lr_converting) {
1366                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1367
1368                 if (iter(lock, closure) == LDLM_ITER_STOP)
1369                         GOTO(out, rc = LDLM_ITER_STOP);
1370         }
1371
1372         list_for_each_safe(tmp, next, &res->lr_waiting) {
1373                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1374
1375                 if (iter(lock, closure) == LDLM_ITER_STOP)
1376                         GOTO(out, rc = LDLM_ITER_STOP);
1377         }
1378  out:
1379         unlock_res(res);
1380         RETURN(rc);
1381 }
1382
1383 struct iter_helper_data {
1384         ldlm_iterator_t iter;
1385         void *closure;
1386 };
1387
1388 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
1389 {
1390         struct iter_helper_data *helper = closure;
1391         return helper->iter(lock, helper->closure);
1392 }
1393
1394 static int ldlm_res_iter_helper(struct ldlm_resource *res, void *closure)
1395 {
1396         return ldlm_resource_foreach(res, ldlm_iter_helper, closure);
1397 }
1398
1399 int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
1400                            void *closure)
1401 {
1402         struct iter_helper_data helper = { iter: iter, closure: closure };
1403         return ldlm_namespace_foreach_res(ns, ldlm_res_iter_helper, &helper);
1404 }
1405
1406 int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
1407                                ldlm_res_iterator_t iter, void *closure)
1408 {
1409         int i, rc = LDLM_ITER_CONTINUE;
1410         struct ldlm_resource *res;
1411         struct list_head *tmp;
1412
1413         ENTRY;
1414         spin_lock(&ns->ns_hash_lock);
1415         for (i = 0; i < RES_HASH_SIZE; i++) {
1416                 tmp = ns->ns_hash[i].next;
1417                 while (tmp != &(ns->ns_hash[i])) {
1418                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
1419                         ldlm_resource_getref(res);
1420                         spin_unlock(&ns->ns_hash_lock);
1421
1422                         rc = iter(res, closure);
1423
1424                         spin_lock(&ns->ns_hash_lock);
1425                         tmp = tmp->next;
1426                         ldlm_resource_putref_locked(res);
1427                         if (rc == LDLM_ITER_STOP)
1428                                 GOTO(out, rc);
1429                 }
1430         }
1431  out:
1432         spin_unlock(&ns->ns_hash_lock);
1433         RETURN(rc);
1434 }
1435
1436 /* non-blocking function to manipulate a lock whose cb_data is being put away.*/
1437 void ldlm_resource_iterate(struct ldlm_namespace *ns,
1438                            const struct ldlm_res_id *res_id,
1439                            ldlm_iterator_t iter, void *data)
1440 {
1441         struct ldlm_resource *res;
1442         ENTRY;
1443
1444         if (ns == NULL) {
1445                 CERROR("must pass in namespace\n");
1446                 LBUG();
1447         }
1448
1449         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1450         if (res == NULL) {
1451                 EXIT;
1452                 return;
1453         }
1454
1455         ldlm_resource_foreach(res, iter, data);
1456         ldlm_resource_putref(res);
1457         EXIT;
1458 }
1459
1460 /* Lock replay */
1461
1462 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
1463 {
1464         struct list_head *list = closure;
1465
1466         /* we use l_pending_chain here, because it's unused on clients. */
1467         LASSERTF(list_empty(&lock->l_pending_chain),"lock %p next %p prev %p\n",
1468                  lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
1469         /* bug 9573: don't replay locks left after eviction */
1470         if (!(lock->l_flags & LDLM_FL_FAILED))
1471                 list_add(&lock->l_pending_chain, list);
1472         return LDLM_ITER_CONTINUE;
1473 }
1474
1475 static int replay_lock_interpret(struct ptlrpc_request *req,
1476                                     void * data, int rc)
1477 {
1478         struct ldlm_lock *lock;
1479         struct ldlm_reply *reply;
1480
1481         ENTRY;
1482         atomic_dec(&req->rq_import->imp_replay_inflight);
1483         if (rc != ELDLM_OK)
1484                 GOTO(out, rc);
1485
1486         lock = req->rq_async_args.pointer_arg[0];
1487         LASSERT(lock != NULL);
1488
1489         reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF, sizeof(*reply),
1490                                    lustre_swab_ldlm_reply);
1491         if (reply == NULL) {
1492                 CERROR("Can't unpack ldlm_reply\n");
1493                 GOTO (out, rc = -EPROTO);
1494         }
1495
1496         lock->l_remote_handle = reply->lock_handle;
1497         LDLM_DEBUG(lock, "replayed lock:");
1498         ptlrpc_import_recovery_state_machine(req->rq_import);
1499  out:
1500         if (rc != ELDLM_OK)
1501                 ptlrpc_connect_import(req->rq_import, NULL);
1502
1503
1504         RETURN(rc);
1505 }
1506
1507 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
1508 {
1509         struct ptlrpc_request *req;
1510         struct ldlm_request *body;
1511         struct ldlm_reply *reply;
1512         int buffers = 2;
1513         int size[3] = { sizeof(struct ptlrpc_body) };
1514         int flags;
1515         ENTRY;
1516
1517         /* If this is reply-less callback lock, we cannot replay it, since
1518          * server might have long dropped it, but notification of that event was
1519          * lost by network. (and server granted conflicting lock already) */
1520         if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1521                 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
1522                 ldlm_lock_cancel(lock);
1523                 RETURN(0);
1524         }
1525         /*
1526          * If granted mode matches the requested mode, this lock is granted.
1527          *
1528          * If they differ, but we have a granted mode, then we were granted
1529          * one mode and now want another: ergo, converting.
1530          *
1531          * If we haven't been granted anything and are on a resource list,
1532          * then we're blocked/waiting.
1533          *
1534          * If we haven't been granted anything and we're NOT on a resource list,
1535          * then we haven't got a reply yet and don't have a known disposition.
1536          * This happens whenever a lock enqueue is the request that triggers
1537          * recovery.
1538          */
1539         if (lock->l_granted_mode == lock->l_req_mode)
1540                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
1541         else if (lock->l_granted_mode)
1542                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
1543         else if (!list_empty(&lock->l_res_link))
1544                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
1545         else
1546                 flags = LDLM_FL_REPLAY;
1547
1548         size[DLM_LOCKREQ_OFF] = sizeof(*body);
1549         req = ptlrpc_prep_req(imp, LUSTRE_DLM_VERSION, LDLM_ENQUEUE, 2, size,
1550                               NULL);
1551         if (!req)
1552                 RETURN(-ENOMEM);
1553
1554         /* We're part of recovery, so don't wait for it. */
1555         req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
1556
1557         body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, sizeof(*body));
1558         ldlm_lock2desc(lock, &body->lock_desc);
1559         body->lock_flags = flags;
1560
1561         ldlm_lock2handle(lock, &body->lock_handle[0]);
1562         size[DLM_LOCKREPLY_OFF] = sizeof(*reply);
1563         if (lock->l_lvb_len != 0) {
1564                 buffers = 3;
1565                 size[DLM_REPLY_REC_OFF] = lock->l_lvb_len;
1566         }
1567         ptlrpc_req_set_repsize(req, buffers, size);
1568         /* notify the server we've replayed all requests.
1569          * also, we mark the request to be put on a dedicated
1570          * queue to be processed after all request replayes.
1571          * bug 6063 */
1572         lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
1573
1574         LDLM_DEBUG(lock, "replaying lock:");
1575
1576         atomic_inc(&req->rq_import->imp_replay_inflight);
1577         req->rq_async_args.pointer_arg[0] = lock;
1578         req->rq_interpret_reply = replay_lock_interpret;
1579         ptlrpcd_add_req(req);
1580
1581         RETURN(0);
1582 }
1583
1584 int ldlm_replay_locks(struct obd_import *imp)
1585 {
1586         struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
1587         struct list_head list;
1588         struct ldlm_lock *lock, *next;
1589         int rc = 0;
1590
1591         ENTRY;
1592         CFS_INIT_LIST_HEAD(&list);
1593
1594         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1595
1596         /* ensure this doesn't fall to 0 before all have been queued */
1597         atomic_inc(&imp->imp_replay_inflight);
1598
1599         (void)ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
1600
1601         list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
1602                 list_del_init(&lock->l_pending_chain);
1603                 if (rc)
1604                         continue; /* or try to do the rest? */
1605                 rc = replay_one_lock(imp, lock);
1606         }
1607
1608         atomic_dec(&imp->imp_replay_inflight);
1609
1610         RETURN(rc);
1611 }