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 struct ldlm_async_args {
  47         struct lustre_handle lock_handle;
  48 };
  49
  50 int ldlm_expired_completion_wait(void *data)
  51 {
  52         struct lock_wait_data *lwd = data;
  53         struct ldlm_lock *lock = lwd->lwd_lock;
  54         struct obd_import *imp;
  55         struct obd_device *obd;
  56
  57         ENTRY;
  58         if (lock->l_conn_export == NULL) {
  59                 static cfs_time_t next_dump = 0, last_dump = 0;
  60
  61                 if (ptlrpc_check_suspend())
  62                         RETURN(0);
  63
  64                 LDLM_ERROR(lock, "lock timed out (enqueued at %lu, %lus ago); "
  65                            "not entering recovery in server code, just going "
  66                            "back to sleep", lock->l_enqueued_time.tv_sec,
  67                            CURRENT_SECONDS - lock->l_enqueued_time.tv_sec);
  68                 if (cfs_time_after(cfs_time_current(), next_dump)) {
  69                         last_dump = next_dump;
  70                         next_dump = cfs_time_shift(300);
  71                         ldlm_namespace_dump(D_DLMTRACE,
  72                                             lock->l_resource->lr_namespace);
  73                         if (last_dump == 0)
  74                                 libcfs_debug_dumplog();
  75                 }
  76                 RETURN(0);
  77         }
  78
  79         obd = lock->l_conn_export->exp_obd;
  80         imp = obd->u.cli.cl_import;
  81         ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
  82         LDLM_ERROR(lock, "lock timed out (enqueued at %lu, %lus ago), entering "
  83                    "recovery for %s@%s", lock->l_enqueued_time.tv_sec,
  84                    CURRENT_SECONDS - lock->l_enqueued_time.tv_sec,
  85                    obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
  86
  87         RETURN(0);
  88 }
  89
  90 int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data)
  91 {
  92         /* XXX ALLOCATE - 160 bytes */
  93         struct lock_wait_data lwd;
  94         struct obd_device *obd;
  95         struct obd_import *imp = NULL;
  96         struct l_wait_info lwi;
  97         int rc = 0;
  98         ENTRY;
  99
 100         if (flags == LDLM_FL_WAIT_NOREPROC) {
 101                 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
 102                 goto noreproc;
 103         }
 104
 105         if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
 106                        LDLM_FL_BLOCK_CONV))) {
 107                 cfs_waitq_signal(&lock->l_waitq);
 108                 RETURN(0);
 109         }
 110
 111         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
 112                    "sleeping");
 113         ldlm_lock_dump(D_OTHER, lock, 0);
 114         ldlm_reprocess_all(lock->l_resource);
 115
 116 noreproc:
 117
 118         obd = class_exp2obd(lock->l_conn_export);
 119
 120         /* if this is a local lock, then there is no import */
 121         if (obd != NULL)
 122                 imp = obd->u.cli.cl_import;
 123
 124         lwd.lwd_lock = lock;
 125
 126         if (lock->l_flags & LDLM_FL_NO_TIMEOUT) {
 127                 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
 128                 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
 129         } else {
 130                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
 131                                        ldlm_expired_completion_wait,
 132                                        interrupted_completion_wait, &lwd);
 133         }
 134
 135         if (imp != NULL) {
 136                 spin_lock(&imp->imp_lock);
 137                 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
 138                 spin_unlock(&imp->imp_lock);
 139         }
 140
 141         /* Go to sleep until the lock is granted or cancelled. */
 142         rc = l_wait_event(lock->l_waitq,
 143                           ((lock->l_req_mode == lock->l_granted_mode) ||
 144                            (lock->l_flags & LDLM_FL_FAILED)), &lwi);
 145
 146         if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED) {
 147                 LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed");
 148                 RETURN(-EIO);
 149         }
 150
 151         if (rc) {
 152                 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
 153                            rc);
 154                 RETURN(rc);
 155         }
 156
 157         LDLM_DEBUG(lock, "client-side enqueue waking up: granted");
 158         RETURN(0);
 159 }
 160
 161 /*
 162  * ->l_blocking_ast() callback for LDLM locks acquired by server-side OBDs.
 163  */
 164 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 165                       void *data, int flag)
 166 {
 167         int do_ast;
 168         ENTRY;
 169
 170         if (flag == LDLM_CB_CANCELING) {
 171                 /* Don't need to do anything here. */
 172                 RETURN(0);
 173         }
 174
 175         lock_res_and_lock(lock);
 176         /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
 177          * that ldlm_blocking_ast is called just before intent_policy method
 178          * takes the ns_lock, then by the time we get the lock, we might not
 179          * be the correct blocking function anymore.  So check, and return
 180          * early, if so. */
 181         if (lock->l_blocking_ast != ldlm_blocking_ast) {
 182                 unlock_res_and_lock(lock);
 183                 RETURN(0);
 184         }
 185
 186         lock->l_flags |= LDLM_FL_CBPENDING;
 187         do_ast = (!lock->l_readers && !lock->l_writers);
 188         unlock_res_and_lock(lock);
 189
 190         if (do_ast) {
 191                 struct lustre_handle lockh;
 192                 int rc;
 193
 194                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
 195                 ldlm_lock2handle(lock, &lockh);
 196                 rc = ldlm_cli_cancel(&lockh);
 197                 if (rc < 0)
 198                         CERROR("ldlm_cli_cancel: %d\n", rc);
 199         } else {
 200                 LDLM_DEBUG(lock, "Lock still has references, will be "
 201                            "cancelled later");
 202         }
 203         RETURN(0);
 204 }
 205
 206 /*
 207  * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See
 208  * comment in filter_intent_policy() on why you may need this.
 209  */
 210 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
 211 {
 212         /*
 213          * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for
 214          * that is rather subtle: with OST-side locking, it may so happen that
 215          * _all_ extent locks are held by the OST. If client wants to obtain
 216          * current file size it calls ll{,u}_glimpse_size(), and (as locks are
 217          * on the server), dummy glimpse callback fires and does
 218          * nothing. Client still receives correct file size due to the
 219          * following fragment in filter_intent_policy():
 220          *
 221          * rc = l->l_glimpse_ast(l, NULL); // this will update the LVB
 222          * if (rc != 0 && res->lr_namespace->ns_lvbo &&
 223          *     res->lr_namespace->ns_lvbo->lvbo_update) {
 224          *         res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1);
 225          * }
 226          *
 227          * that is, after glimpse_ast() fails, filter_lvbo_update() runs, and
 228          * returns correct file size to the client.
 229          */
 230         return -ELDLM_NO_LOCK_DATA;
 231 }
 232
 233 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
 234                            const struct ldlm_res_id *res_id,
 235                            ldlm_type_t type, ldlm_policy_data_t *policy,
 236                            ldlm_mode_t mode, int *flags,
 237                            ldlm_blocking_callback blocking,
 238                            ldlm_completion_callback completion,
 239                            ldlm_glimpse_callback glimpse,
 240                            void *data, __u32 lvb_len, void *lvb_swabber,
 241                            struct lustre_handle *lockh)
 242 {
 243         struct ldlm_lock *lock;
 244         int err;
 245         ENTRY;
 246
 247         LASSERT(!(*flags & LDLM_FL_REPLAY));
 248         if (unlikely(ns->ns_client)) {
 249                 CERROR("Trying to enqueue local lock in a shadow namespace\n");
 250                 LBUG();
 251         }
 252
 253         lock = ldlm_lock_create(ns, res_id, type, mode, blocking,
 254                                 completion, glimpse, data, lvb_len);
 255         if (unlikely(!lock))
 256                 GOTO(out_nolock, err = -ENOMEM);
 257         LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
 258
 259         ldlm_lock_addref_internal(lock, mode);
 260         ldlm_lock2handle(lock, lockh);
 261         lock_res_and_lock(lock);
 262         lock->l_flags |= LDLM_FL_LOCAL;
 263         if (*flags & LDLM_FL_ATOMIC_CB)
 264                 lock->l_flags |= LDLM_FL_ATOMIC_CB;
 265         lock->l_lvb_swabber = lvb_swabber;
 266         unlock_res_and_lock(lock);
 267         if (policy != NULL)
 268                 lock->l_policy_data = *policy;
 269         if (type == LDLM_EXTENT)
 270                 lock->l_req_extent = policy->l_extent;
 271
 272         err = ldlm_lock_enqueue(ns, &lock, policy, flags);
 273         if (unlikely(err != ELDLM_OK))
 274                 GOTO(out, err);
 275
 276         if (policy != NULL)
 277                 *policy = lock->l_policy_data;
 278
 279         LDLM_DEBUG_NOLOCK("client-side local enqueue handler END (lock %p)",
 280                           lock);
 281
 282         if (lock->l_completion_ast)
 283                 lock->l_completion_ast(lock, *flags, NULL);
 284
 285         LDLM_DEBUG(lock, "client-side local enqueue END");
 286         EXIT;
 287  out:
 288         LDLM_LOCK_PUT(lock);
 289  out_nolock:
 290         return err;
 291 }
 292
 293 static void failed_lock_cleanup(struct ldlm_namespace *ns,
 294                                 struct ldlm_lock *lock,
 295                                 struct lustre_handle *lockh, int mode)
 296 {
 297         /* Set a flag to prevent us from sending a CANCEL (bug 407) */
 298         lock_res_and_lock(lock);
 299         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 300         unlock_res_and_lock(lock);
 301         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 302
 303         ldlm_lock_decref_and_cancel(lockh, mode);
 304
 305         /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
 306          *       from llite/file.c/ll_file_flock(). */
 307         if (lock->l_resource->lr_type == LDLM_FLOCK) {
 308                 ldlm_lock_destroy(lock);
 309         }
 310 }
 311
 312 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
 313                           ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
 314                           int *flags, void *lvb, __u32 lvb_len,
 315                           void *lvb_swabber, struct lustre_handle *lockh,int rc)
 316 {
 317         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 318         int is_replay = *flags & LDLM_FL_REPLAY;
 319         struct ldlm_lock *lock;
 320         struct ldlm_reply *reply;
 321         int cleanup_phase = 1;
 322         ENTRY;
 323
 324         lock = ldlm_handle2lock(lockh);
 325         /* ldlm_cli_enqueue is holding a reference on this lock. */
 326         if (!lock) {
 327                 LASSERT(type == LDLM_FLOCK);
 328                 RETURN(-ENOLCK);
 329         }
 330
 331         if (rc != ELDLM_OK) {
 332                 LASSERT(!is_replay);
 333                 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
 334                            rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
 335                 if (rc == ELDLM_LOCK_ABORTED) {
 336                         /* Before we return, swab the reply */
 337                         reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF,
 338                                                    sizeof(*reply),
 339                                                    lustre_swab_ldlm_reply);
 340                         if (reply == NULL) {
 341                                 CERROR("Can't unpack ldlm_reply\n");
 342                                 rc = -EPROTO;
 343                         }
 344                         if (lvb_len) {
 345                                 void *tmplvb;
 346                                 tmplvb = lustre_swab_repbuf(req,
 347                                                             DLM_REPLY_REC_OFF,
 348                                                             lvb_len,
 349                                                             lvb_swabber);
 350                                 if (tmplvb == NULL)
 351                                         GOTO(cleanup, rc = -EPROTO);
 352                                 if (lvb != NULL)
 353                                         memcpy(lvb, tmplvb, lvb_len);
 354                         }
 355                 }
 356                 GOTO(cleanup, rc);
 357         }
 358
 359         reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF, sizeof(*reply),
 360                                    lustre_swab_ldlm_reply);
 361         if (reply == NULL) {
 362                 CERROR("Can't unpack ldlm_reply\n");
 363                 GOTO(cleanup, rc = -EPROTO);
 364         }
 365
 366         /* lock enqueued on the server */
 367         cleanup_phase = 0;
 368
 369         lock_res_and_lock(lock);
 370         lock->l_remote_handle = reply->lock_handle;
 371         *flags = reply->lock_flags;
 372         lock->l_flags |= reply->lock_flags & LDLM_INHERIT_FLAGS;
 373         /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match()
 374          * to wait with no timeout as well */
 375         lock->l_flags |= reply->lock_flags & LDLM_FL_NO_TIMEOUT;
 376         unlock_res_and_lock(lock);
 377
 378         CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%x\n",
 379                lock, reply->lock_handle.cookie, *flags);
 380
 381         /* If enqueue returned a blocked lock but the completion handler has
 382          * already run, then it fixed up the resource and we don't need to do it
 383          * again. */
 384         if ((*flags) & LDLM_FL_LOCK_CHANGED) {
 385                 int newmode = reply->lock_desc.l_req_mode;
 386                 LASSERT(!is_replay);
 387                 if (newmode && newmode != lock->l_req_mode) {
 388                         LDLM_DEBUG(lock, "server returned different mode %s",
 389                                    ldlm_lockname[newmode]);
 390                         lock->l_req_mode = newmode;
 391                 }
 392
 393                 if (memcmp(reply->lock_desc.l_resource.lr_name.name,
 394                           lock->l_resource->lr_name.name,
 395                           sizeof(struct ldlm_res_id))) {
 396                         CDEBUG(D_INFO, "remote intent success, locking "
 397                                         "(%ld,%ld,%ld) instead of "
 398                                         "(%ld,%ld,%ld)\n",
 399                               (long)reply->lock_desc.l_resource.lr_name.name[0],
 400                               (long)reply->lock_desc.l_resource.lr_name.name[1],
 401                               (long)reply->lock_desc.l_resource.lr_name.name[2],
 402                               (long)lock->l_resource->lr_name.name[0],
 403                               (long)lock->l_resource->lr_name.name[1],
 404                               (long)lock->l_resource->lr_name.name[2]);
 405
 406                         ldlm_lock_change_resource(ns, lock,
 407                                           &reply->lock_desc.l_resource.lr_name);
 408                         if (lock->l_resource == NULL) {
 409                                 LBUG();
 410                                 GOTO(cleanup, rc = -ENOMEM);
 411                         }
 412                         LDLM_DEBUG(lock, "client-side enqueue, new resource");
 413                 }
 414                 if (with_policy)
 415                         if (!(type == LDLM_IBITS && !(exp->exp_connect_flags &
 416                                                     OBD_CONNECT_IBITS)))
 417                                 lock->l_policy_data =
 418                                                  reply->lock_desc.l_policy_data;
 419                 if (type != LDLM_PLAIN)
 420                         LDLM_DEBUG(lock,"client-side enqueue, new policy data");
 421         }
 422
 423         if ((*flags) & LDLM_FL_AST_SENT ||
 424             /* Cancel extent locks as soon as possible on a liblustre client,
 425              * because it cannot handle asynchronous ASTs robustly (see
 426              * bug 7311). */
 427             (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) {
 428                 lock_res_and_lock(lock);
 429                 lock->l_flags |= LDLM_FL_CBPENDING;
 430                 unlock_res_and_lock(lock);
 431                 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
 432         }
 433
 434         /* If the lock has already been granted by a completion AST, don't
 435          * clobber the LVB with an older one. */
 436         if (lvb_len && (lock->l_req_mode != lock->l_granted_mode)) {
 437                 void *tmplvb;
 438                 tmplvb = lustre_swab_repbuf(req, DLM_REPLY_REC_OFF, lvb_len,
 439                                             lvb_swabber);
 440                 if (tmplvb == NULL)
 441                         GOTO(cleanup, rc = -EPROTO);
 442                 memcpy(lock->l_lvb_data, tmplvb, lvb_len);
 443         }
 444
 445         if (!is_replay) {
 446                 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
 447                 if (lock->l_completion_ast != NULL) {
 448                         int err = lock->l_completion_ast(lock, *flags, NULL);
 449                         if (!rc)
 450                                 rc = err;
 451                         if (rc && type != LDLM_FLOCK) /* bug 9425, bug 10250 */
 452                                 cleanup_phase = 1;
 453                 }
 454         }
 455
 456         if (lvb_len && lvb != NULL) {
 457                 /* Copy the LVB here, and not earlier, because the completion
 458                  * AST (if any) can override what we got in the reply */
 459                 memcpy(lvb, lock->l_lvb_data, lvb_len);
 460         }
 461
 462         LDLM_DEBUG(lock, "client-side enqueue END");
 463         EXIT;
 464 cleanup:
 465         if (cleanup_phase == 1 && rc)
 466                 failed_lock_cleanup(ns, lock, lockh, mode);
 467         /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
 468         LDLM_LOCK_PUT(lock);
 469         LDLM_LOCK_PUT(lock);
 470         return rc;
 471 }
 472
 473 /* PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
 474  * a single page on the send/receive side. XXX: 512 should be changed
 475  * to more adequate value. */
 476 static inline int ldlm_req_handles_avail(struct obd_export *exp,
 477                                          int *size, int bufcount, int off)
 478 {
 479         int avail = min_t(int, LDLM_MAXREQSIZE, PAGE_SIZE - 512);
 480         int old_size = size[DLM_LOCKREQ_OFF];
 481
 482         size[DLM_LOCKREQ_OFF] = sizeof(struct ldlm_request);
 483         avail -= lustre_msg_size(class_exp2cliimp(exp)->imp_msg_magic,
 484                                  bufcount, size);
 485         avail /= sizeof(struct lustre_handle);
 486         avail += LDLM_LOCKREQ_HANDLES - off;
 487         size[DLM_LOCKREQ_OFF] = old_size;
 488
 489         return avail;
 490 }
 491
 492 static inline int ldlm_cancel_handles_avail(struct obd_export *exp)
 493 {
 494         int size[2] = { sizeof(struct ptlrpc_body),
 495                         sizeof(struct ldlm_request) };
 496         return ldlm_req_handles_avail(exp, size, 2, 0);
 497 }
 498
 499 /* Cancel lru locks and pack them into the enqueue request. Pack there the given
 500  * @count locks in @cancels. */
 501 struct ptlrpc_request *ldlm_prep_enqueue_req(struct obd_export *exp,
 502                                              int bufcount, int *size,
 503                                              struct list_head *cancels,
 504                                              int count)
 505 {
 506         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 507         struct ldlm_request *dlm = NULL;
 508         struct ptlrpc_request *req;
 509         CFS_LIST_HEAD(head);
 510         ENTRY;
 511
 512         if (cancels == NULL)
 513                 cancels = &head;
 514         if (exp_connect_cancelset(exp)) {
 515                 /* Estimate the amount of available space in the request. */
 516                 int avail = ldlm_req_handles_avail(exp, size, bufcount,
 517                                                    LDLM_ENQUEUE_CANCEL_OFF);
 518                 LASSERT(avail >= count);
 519
 520                 /* Cancel lru locks here _only_ if the server supports
 521                  * EARLY_CANCEL. Otherwise we have to send extra CANCEL
 522                  * rpc right on enqueue, what will make it slower, vs.
 523                  * asynchronous rpc in blocking thread. */
 524                 count += ldlm_cancel_lru_local(ns, cancels, 1, avail - count,
 525                                                LDLM_CANCEL_AGED);
 526                 size[DLM_LOCKREQ_OFF] =
 527                         ldlm_request_bufsize(count, LDLM_ENQUEUE);
 528         }
 529         req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
 530                               LDLM_ENQUEUE, bufcount, size, NULL);
 531         if (exp_connect_cancelset(exp) && req) {
 532                 dlm = lustre_msg_buf(req->rq_reqmsg,
 533                                      DLM_LOCKREQ_OFF, sizeof(*dlm));
 534                 /* Skip first lock handler in ldlm_request_pack(), this method
 535                  * will incrment @lock_count according to the lock handle amount
 536                  * actually written to the buffer. */
 537                 dlm->lock_count = LDLM_ENQUEUE_CANCEL_OFF;
 538                 ldlm_cli_cancel_list(cancels, count, req, DLM_LOCKREQ_OFF, 0);
 539         } else {
 540                 ldlm_lock_list_put(cancels, l_bl_ast, count);
 541         }
 542         RETURN(req);
 543 }
 544
 545 /* If a request has some specific initialisation it is passed in @reqp,
 546  * otherwise it is created in ldlm_cli_enqueue.
 547  *
 548  * Supports sync and async requests, pass @async flag accordingly. If a
 549  * request was created in ldlm_cli_enqueue and it is the async request,
 550  * pass it to the caller in @reqp. */
 551 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
 552                      struct ldlm_enqueue_info *einfo,
 553                      const struct ldlm_res_id *res_id,
 554                      ldlm_policy_data_t *policy, int *flags,
 555                      void *lvb, __u32 lvb_len, void *lvb_swabber,
 556                      struct lustre_handle *lockh, int async)
 557 {
 558         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 559         struct ldlm_lock *lock;
 560         struct ldlm_request *body;
 561         struct ldlm_reply *reply;
 562         int size[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
 563                         [DLM_LOCKREQ_OFF]     = sizeof(*body),
 564                         [DLM_REPLY_REC_OFF]   = lvb_len };
 565         int is_replay = *flags & LDLM_FL_REPLAY;
 566         int req_passed_in = 1, rc, err;
 567         struct ptlrpc_request *req;
 568         ENTRY;
 569
 570         LASSERT(exp != NULL);
 571
 572         /* If we're replaying this lock, just check some invariants.
 573          * If we're creating a new lock, get everything all setup nice. */
 574         if (is_replay) {
 575                 lock = ldlm_handle2lock(lockh);
 576                 LASSERT(lock != NULL);
 577                 LDLM_DEBUG(lock, "client-side enqueue START");
 578                 LASSERT(exp == lock->l_conn_export);
 579         } else {
 580                 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
 581                                         einfo->ei_mode, einfo->ei_cb_bl,
 582                                         einfo->ei_cb_cp, einfo->ei_cb_gl,
 583                                         einfo->ei_cbdata, lvb_len);
 584                 if (lock == NULL)
 585                         RETURN(-ENOMEM);
 586                 /* for the local lock, add the reference */
 587                 ldlm_lock_addref_internal(lock, einfo->ei_mode);
 588                 ldlm_lock2handle(lock, lockh);
 589                 lock->l_lvb_swabber = lvb_swabber;
 590                 if (policy != NULL) {
 591                         /* INODEBITS_INTEROP: If the server does not support
 592                          * inodebits, we will request a plain lock in the
 593                          * descriptor (ldlm_lock2desc() below) but use an
 594                          * inodebits lock internally with both bits set.
 595                          */
 596                         if (einfo->ei_type == LDLM_IBITS &&
 597                             !(exp->exp_connect_flags & OBD_CONNECT_IBITS))
 598                                 lock->l_policy_data.l_inodebits.bits =
 599                                         MDS_INODELOCK_LOOKUP |
 600                                         MDS_INODELOCK_UPDATE;
 601                         else
 602                                 lock->l_policy_data = *policy;
 603                 }
 604
 605                 if (einfo->ei_type == LDLM_EXTENT)
 606                         lock->l_req_extent = policy->l_extent;
 607                 LDLM_DEBUG(lock, "client-side enqueue START");
 608         }
 609
 610         /* lock not sent to server yet */
 611
 612         if (reqp == NULL || *reqp == NULL) {
 613                 req = ldlm_prep_enqueue_req(exp, 2, size, NULL, 0);
 614                 if (req == NULL) {
 615                         failed_lock_cleanup(ns, lock, lockh, einfo->ei_mode);
 616                         LDLM_LOCK_PUT(lock);
 617                         RETURN(-ENOMEM);
 618                 }
 619                 req_passed_in = 0;
 620                 if (reqp)
 621                         *reqp = req;
 622         } else {
 623                 req = *reqp;
 624                 LASSERTF(lustre_msg_buflen(req->rq_reqmsg, DLM_LOCKREQ_OFF) >=
 625                          sizeof(*body), "buflen[%d] = %d, not "LPSZ"\n",
 626                          DLM_LOCKREQ_OFF,
 627                          lustre_msg_buflen(req->rq_reqmsg, DLM_LOCKREQ_OFF),
 628                          sizeof(*body));
 629         }
 630
 631         lock->l_conn_export = exp;
 632         lock->l_export = NULL;
 633         lock->l_blocking_ast = einfo->ei_cb_bl;
 634
 635         /* Dump lock data into the request buffer */
 636         body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, sizeof(*body));
 637         ldlm_lock2desc(lock, &body->lock_desc);
 638         body->lock_flags = *flags;
 639         body->lock_handle[0] = *lockh;
 640
 641         /* Continue as normal. */
 642         if (!req_passed_in) {
 643                 size[DLM_LOCKREPLY_OFF] = sizeof(*reply);
 644                 ptlrpc_req_set_repsize(req, 2 + (lvb_len > 0), size);
 645         }
 646
 647         /*
 648          * Liblustre client doesn't get extent locks, except for O_APPEND case
 649          * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
 650          * [i_size, OBD_OBJECT_EOF] lock is taken.
 651          */
 652         LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT ||
 653                      policy->l_extent.end == OBD_OBJECT_EOF));
 654
 655         if (async) {
 656                 LASSERT(reqp != NULL);
 657                 RETURN(0);
 658         }
 659
 660         LDLM_DEBUG(lock, "sending request");
 661         rc = ptlrpc_queue_wait(req);
 662         err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
 663                                     einfo->ei_mode, flags, lvb, lvb_len,
 664                                     lvb_swabber, lockh, rc);
 665
 666         /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
 667          * one reference that we took */
 668         if (err == -ENOLCK)
 669                 LDLM_LOCK_PUT(lock);
 670         else
 671                 rc = err;
 672
 673         if (!req_passed_in && req != NULL) {
 674                 ptlrpc_req_finished(req);
 675                 if (reqp)
 676                         *reqp = NULL;
 677         }
 678
 679         RETURN(rc);
 680 }
 681
 682 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
 683                                   int *flags)
 684 {
 685         struct ldlm_resource *res;
 686         int rc;
 687         ENTRY;
 688         if (lock->l_resource->lr_namespace->ns_client) {
 689                 CERROR("Trying to cancel local lock\n");
 690                 LBUG();
 691         }
 692         LDLM_DEBUG(lock, "client-side local convert");
 693
 694         res = ldlm_lock_convert(lock, new_mode, flags);
 695         if (res) {
 696                 ldlm_reprocess_all(res);
 697                 rc = 0;
 698         } else {
 699                 rc = EDEADLOCK;
 700         }
 701         LDLM_DEBUG(lock, "client-side local convert handler END");
 702         LDLM_LOCK_PUT(lock);
 703         RETURN(rc);
 704 }
 705
 706 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
 707  * conversion of locks which are on the waiting or converting queue */
 708 /* Caller of this code is supposed to take care of lock readers/writers
 709    accounting */
 710 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, int *flags)
 711 {
 712         struct ldlm_request *body;
 713         struct ldlm_reply *reply;
 714         struct ldlm_lock *lock;
 715         struct ldlm_resource *res;
 716         struct ptlrpc_request *req;
 717         int size[2] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
 718                         [DLM_LOCKREQ_OFF]     = sizeof(*body) };
 719         int rc;
 720         ENTRY;
 721
 722         lock = ldlm_handle2lock(lockh);
 723         if (!lock) {
 724                 LBUG();
 725                 RETURN(-EINVAL);
 726         }
 727         *flags = 0;
 728
 729         if (lock->l_conn_export == NULL)
 730                 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
 731
 732         LDLM_DEBUG(lock, "client-side convert");
 733
 734         req = ptlrpc_prep_req(class_exp2cliimp(lock->l_conn_export),
 735                               LUSTRE_DLM_VERSION, LDLM_CONVERT, 2, size, NULL);
 736         if (!req)
 737                 GOTO(out, rc = -ENOMEM);
 738
 739         body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, sizeof(*body));
 740         body->lock_handle[0] = lock->l_remote_handle;
 741
 742         body->lock_desc.l_req_mode = new_mode;
 743         body->lock_flags = *flags;
 744
 745         size[DLM_LOCKREPLY_OFF] = sizeof(*reply);
 746         ptlrpc_req_set_repsize(req, 2, size);
 747
 748         rc = ptlrpc_queue_wait(req);
 749         if (rc != ELDLM_OK)
 750                 GOTO(out, rc);
 751
 752         reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF, sizeof(*reply),
 753                                    lustre_swab_ldlm_reply);
 754         if (reply == NULL) {
 755                 CERROR ("Can't unpack ldlm_reply\n");
 756                 GOTO (out, rc = -EPROTO);
 757         }
 758
 759         if (req->rq_status)
 760                 GOTO(out, rc = req->rq_status);
 761
 762         res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
 763         if (res != NULL) {
 764                 ldlm_reprocess_all(res);
 765                 /* Go to sleep until the lock is granted. */
 766                 /* FIXME: or cancelled. */
 767                 if (lock->l_completion_ast) {
 768                         rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
 769                                                     NULL);
 770                         if (rc)
 771                                 GOTO(out, rc);
 772                 }
 773         } else {
 774                 rc = EDEADLOCK;
 775         }
 776         EXIT;
 777  out:
 778         LDLM_LOCK_PUT(lock);
 779         ptlrpc_req_finished(req);
 780         return rc;
 781 }
 782
 783 /* Cancel locks locally.
 784  * Returns:
 785  * LDLM_FL_LOCAL_ONLY if tere is no need in a CANCEL rpc to the server;
 786  * LDLM_FL_CANCELING otherwise;
 787  * LDLM_FL_BL_AST if there is a need in a separate CANCEL rpc. */
 788 static int ldlm_cli_cancel_local(struct ldlm_lock *lock)
 789 {
 790         int rc = LDLM_FL_LOCAL_ONLY;
 791         ENTRY;
 792
 793         if (lock->l_conn_export) {
 794                 int local_only;
 795
 796                 LDLM_DEBUG(lock, "client-side cancel");
 797                 /* Set this flag to prevent others from getting new references*/
 798                 lock_res_and_lock(lock);
 799                 lock->l_flags |= LDLM_FL_CBPENDING;
 800                 local_only = (lock->l_flags &
 801                               (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
 802                 ldlm_cancel_callback(lock);
 803                 rc = (lock->l_flags & LDLM_FL_BL_AST) ?
 804                         LDLM_FL_BL_AST : LDLM_FL_CANCELING;
 805                 unlock_res_and_lock(lock);
 806
 807                 if (local_only) {
 808                         CDEBUG(D_DLMTRACE, "not sending request (at caller's "
 809                                "instruction)\n");
 810                         rc = LDLM_FL_LOCAL_ONLY;
 811                 }
 812                 ldlm_lock_cancel(lock);
 813         } else {
 814                 if (lock->l_resource->lr_namespace->ns_client) {
 815                         LDLM_ERROR(lock, "Trying to cancel local lock");
 816                         LBUG();
 817                 }
 818                 LDLM_DEBUG(lock, "server-side local cancel");
 819                 ldlm_lock_cancel(lock);
 820                 ldlm_reprocess_all(lock->l_resource);
 821                 LDLM_DEBUG(lock, "server-side local cancel handler END");
 822         }
 823
 824         RETURN(rc);
 825 }
 826
 827 /* Pack @count locks in @head into ldlm_request buffer at the offset @off,
 828    of the request @req. */
 829 static void ldlm_cancel_pack(struct ptlrpc_request *req, int off,
 830                              struct list_head *head, int count)
 831 {
 832         struct ldlm_request *dlm;
 833         struct ldlm_lock *lock;
 834         int max, packed = 0;
 835         ENTRY;
 836
 837         dlm = lustre_msg_buf(req->rq_reqmsg, off, sizeof(*dlm));
 838         LASSERT(dlm != NULL);
 839
 840         /* Check the room in the request buffer. */
 841         max = lustre_msg_buflen(req->rq_reqmsg, off) -
 842                 sizeof(struct ldlm_request);
 843         max /= sizeof(struct lustre_handle);
 844         max += LDLM_LOCKREQ_HANDLES;
 845         LASSERT(max >= dlm->lock_count + count);
 846
 847         /* XXX: it would be better to pack lock handles grouped by resource.
 848          * so that the server cancel would call filter_lvbo_update() less
 849          * frequently. */
 850         list_for_each_entry(lock, head, l_bl_ast) {
 851                 if (!count--)
 852                         break;
 853                 LASSERT(lock->l_conn_export);
 854                 /* Pack the lock handle to the given request buffer. */
 855                 LDLM_DEBUG(lock, "packing");
 856                 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
 857                 packed++;
 858         }
 859         CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
 860         EXIT;
 861 }
 862
 863 /* Prepare and send a batched cancel rpc, it will include count lock handles
 864  * of locks given in @head. */
 865 int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
 866                         int count, int flags)
 867 {
 868         struct ptlrpc_request *req = NULL;
 869         struct ldlm_request *body;
 870         int size[2] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
 871                 [DLM_LOCKREQ_OFF]     = sizeof(*body) };
 872         struct obd_import *imp;
 873         int free, sent = 0;
 874         int rc = 0;
 875         ENTRY;
 876
 877         LASSERT(exp != NULL);
 878         LASSERT(count > 0);
 879
 880         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
 881                 RETURN(count);
 882
 883         free = ldlm_req_handles_avail(exp, size, 2, 0);
 884         if (count > free)
 885                 count = free;
 886
 887         size[DLM_LOCKREQ_OFF] = ldlm_request_bufsize(count, LDLM_CANCEL);
 888         while (1) {
 889                 imp = class_exp2cliimp(exp);
 890                 if (imp == NULL || imp->imp_invalid) {
 891                         CDEBUG(D_HA, "skipping cancel on invalid import %p\n",
 892                                imp);
 893                         break;
 894                 }
 895
 896                 req = ptlrpc_prep_req(imp, LUSTRE_DLM_VERSION, LDLM_CANCEL, 2,
 897                                       size, NULL);
 898                 if (!req)
 899                         GOTO(out, rc = -ENOMEM);
 900
 901                 req->rq_no_resend = 1;
 902                 req->rq_no_delay = 1;
 903
 904                 /* XXX FIXME bug 249 */
 905                 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
 906                 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
 907
 908                 body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF,
 909                                       sizeof(*body));
 910                 ldlm_cancel_pack(req, DLM_LOCKREQ_OFF, cancels, count);
 911
 912                 ptlrpc_req_set_repsize(req, 1, NULL);
 913                 if (flags & LDLM_FL_ASYNC) {
 914                         ptlrpcd_add_req(req);
 915                         sent = count;
 916                         GOTO(out, 0);
 917                 } else {
 918                         rc = ptlrpc_queue_wait(req);
 919                 }
 920                 if (rc == ESTALE) {
 921                         CDEBUG(D_DLMTRACE, "client/server (nid %s) "
 922                                "out of sync -- not fatal\n",
 923                                libcfs_nid2str(req->rq_import->
 924                                               imp_connection->c_peer.nid));
 925                         rc = 0;
 926                 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
 927                            req->rq_import_generation == imp->imp_generation) {
 928                         ptlrpc_req_finished(req);
 929                         continue;
 930                 } else if (rc != ELDLM_OK) {
 931                         CERROR("Got rc %d from cancel RPC: canceling "
 932                                "anyway\n", rc);
 933                         break;
 934                 }
 935                 sent = count;
 936                 break;
 937         }
 938
 939         ptlrpc_req_finished(req);
 940         EXIT;
 941 out:
 942         return sent ? sent : rc;
 943 }
 944
 945 int ldlm_cli_cancel(struct lustre_handle *lockh)
 946 {
 947         struct ldlm_lock *lock;
 948         CFS_LIST_HEAD(cancels);
 949         int rc = 0;
 950         ENTRY;
 951
 952         /* concurrent cancels on the same handle can happen */
 953         lock = __ldlm_handle2lock(lockh, LDLM_FL_CANCELING);
 954         if (lock == NULL) {
 955                 LDLM_DEBUG_NOLOCK("lock is already being destroyed\n");
 956                 RETURN(0);
 957         }
 958
 959         rc = ldlm_cli_cancel_local(lock);
 960         list_add(&lock->l_bl_ast, &cancels);
 961
 962         if (rc == LDLM_FL_BL_AST) {
 963                 rc = ldlm_cli_cancel_req(lock->l_conn_export, &cancels, 1, 0);
 964         } else if (rc == LDLM_FL_CANCELING) {
 965                 int avail = ldlm_cancel_handles_avail(lock->l_conn_export);
 966                 int count = 1;
 967                 LASSERT(avail > 0);
 968                 count += ldlm_cancel_lru_local(lock->l_resource->lr_namespace,
 969                                                &cancels, 0, avail - 1,
 970                                                LDLM_CANCEL_AGED);
 971                 ldlm_cli_cancel_list(&cancels, count, NULL, 0, 0);
 972         }
 973         if (rc != LDLM_FL_CANCELING)
 974                 LDLM_LOCK_PUT(lock);
 975         RETURN(rc < 0 ? rc : 0);
 976 }
 977
 978 /* - Free space in lru for @count new locks,
 979  *   redundant unused locks are canceled locally;
 980  * - also cancel locally unused aged locks;
 981  * - do not cancel more than @max locks;
 982  * - GET the found locks and add them into the @cancels list.
 983  *
 984  * A client lock can be added to the l_bl_ast list only when it is
 985  * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing CANCEL.
 986  * There are the following use cases: ldlm_cancel_resource_local(),
 987  * ldlm_cancel_lru_local() and ldlm_cli_cancel(), which check&set this
 988  * flag properly. As any attempt to cancel a lock rely on this flag,
 989  * l_bl_ast list is accessed later without any special locking. */
 990 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
 991                           int count, int max, int flags)
 992 {
 993         cfs_time_t cur = cfs_time_current();
 994         struct ldlm_lock *lock, *next;
 995         int rc, added = 0, left;
 996         ENTRY;
 997
 998         spin_lock(&ns->ns_unused_lock);
 999         count += ns->ns_nr_unused - ns->ns_max_unused;
1000         while (!list_empty(&ns->ns_unused_list)) {
1001                 if (max && added >= max)
1002                         break;
1003
1004                 list_for_each_entry(lock, &ns->ns_unused_list, l_lru) {
1005                         /* somebody is already doing CANCEL or there is a
1006                          * blocking request will send cancel. */
1007                         if (!(lock->l_flags & LDLM_FL_CANCELING) &&
1008                             !(lock->l_flags & LDLM_FL_BL_AST))
1009                                 break;
1010                 }
1011                 if (&lock->l_lru == &ns->ns_unused_list)
1012                         break;
1013
1014                 if ((added >= count) &&
1015                     (!(flags & LDLM_CANCEL_AGED) ||
1016                      cfs_time_before_64(cur, (__u64)ns->ns_max_age +
1017                                         lock->l_last_used)))
1018                         break;
1019
1020                 LDLM_LOCK_GET(lock); /* dropped by bl thread */
1021                 spin_unlock(&ns->ns_unused_lock);
1022
1023                 lock_res_and_lock(lock);
1024                 /* Check flags again under the lock. */
1025                 if ((lock->l_flags & LDLM_FL_CANCELING) ||
1026                     (lock->l_flags & LDLM_FL_BL_AST) ||
1027                     (ldlm_lock_remove_from_lru(lock) == 0)) {
1028                         /* other thread is removing lock from lru or
1029                          * somebody is already doing CANCEL or
1030                          * there is a blocking request which will send
1031                          * cancel by itseft. */
1032                         unlock_res_and_lock(lock);
1033                         LDLM_LOCK_PUT(lock);
1034                         spin_lock(&ns->ns_unused_lock);
1035                         continue;
1036                 }
1037                 LASSERT(!lock->l_readers && !lock->l_writers);
1038
1039                 /* If we have chosen to canecl this lock voluntarily, we better
1040                    send cancel notification to server, so that it frees
1041                    appropriate state. This might lead to a race where while
1042                    we are doing cancel here, server is also silently
1043                    cancelling this lock. */
1044                 lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK;
1045
1046                 /* Setting the CBPENDING flag is a little misleading, but
1047                  * prevents an important race; namely, once CBPENDING is set,
1048                  * the lock can accumulate no more readers/writers.  Since
1049                  * readers and writers are already zero here, ldlm_lock_decref
1050                  * won't see this flag and call l_blocking_ast */
1051                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1052                 /* We can't re-add to l_lru as it confuses the refcounting in
1053                  * ldlm_lock_remove_from_lru() if an AST arrives after we drop
1054                  * ns_lock below. We use l_bl_ast and can't use l_pending_chain
1055                  * as it is used both on server and client nevertheles bug 5666
1056                  * says it is used only on server. --umka */
1057
1058                 LASSERT(list_empty(&lock->l_bl_ast));
1059                 list_add(&lock->l_bl_ast, cancels);
1060                 unlock_res_and_lock(lock);
1061                 spin_lock(&ns->ns_unused_lock);
1062                 added++;
1063         }
1064         spin_unlock(&ns->ns_unused_lock);
1065
1066         /* Handle only @added inserted locks. */
1067         left = added;
1068         list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1069                 if (left-- == 0)
1070                         break;
1071
1072                 rc = ldlm_cli_cancel_local(lock);
1073                 if (rc == LDLM_FL_BL_AST) {
1074                         CFS_LIST_HEAD(head);
1075
1076                         LDLM_DEBUG(lock, "Cancel lock separately");
1077                         list_del_init(&lock->l_bl_ast);
1078                         list_add(&lock->l_bl_ast, &head);
1079                         ldlm_cli_cancel_req(lock->l_conn_export, &head, 1, 0);
1080                         rc = LDLM_FL_LOCAL_ONLY;
1081                 }
1082                 if (rc == LDLM_FL_LOCAL_ONLY) {
1083                         /* CANCEL RPC should not be sent to server. */
1084                         list_del_init(&lock->l_bl_ast);
1085                         LDLM_LOCK_PUT(lock);
1086                         added--;
1087                 }
1088
1089         }
1090         RETURN(added);
1091 }
1092
1093 /* when called with LDLM_ASYNC the blocking callback will be handled
1094  * in a thread and this function will return after the thread has been
1095  * asked to call the callback.  when called with LDLM_SYNC the blocking
1096  * callback will be performed in this function. */
1097 int ldlm_cancel_lru(struct ldlm_namespace *ns, ldlm_sync_t sync)
1098 {
1099         CFS_LIST_HEAD(cancels);
1100         int count, rc;
1101         ENTRY;
1102
1103 #ifndef __KERNEL__
1104         sync = LDLM_SYNC; /* force to be sync in user space */
1105 #endif
1106         count = ldlm_cancel_lru_local(ns, &cancels, 0, 0, 0);
1107         if (sync == LDLM_ASYNC) {
1108                 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count);
1109                 if (rc == 0)
1110                         RETURN(count);
1111         }
1112
1113         /* If an error occured in ASYNC mode, or
1114          * this is SYNC mode, cancel the list. */
1115         ldlm_cli_cancel_list(&cancels, count, NULL, 0, 0);
1116         RETURN(0);
1117 }
1118
1119 /* Find and cancel locally unused locks found on resource, matched to the
1120  * given policy, mode. GET the found locks and add them into the @cancels
1121  * list. */
1122 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1123                                struct list_head *cancels,
1124                                ldlm_policy_data_t *policy,
1125                                ldlm_mode_t mode, int lock_flags,
1126                                int flags, void *opaque)
1127 {
1128         struct ldlm_lock *lock, *next;
1129         int count = 0, left;
1130         ENTRY;
1131
1132         lock_res(res);
1133         list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1134                 if (opaque != NULL && lock->l_ast_data != opaque) {
1135                         LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1136                                    lock->l_ast_data, opaque);
1137                         //LBUG();
1138                         continue;
1139                 }
1140
1141                 if (lock->l_readers || lock->l_writers) {
1142                         if (flags & LDLM_FL_WARN) {
1143                                 LDLM_ERROR(lock, "lock in use");
1144                                 //LBUG();
1145                         }
1146                         continue;
1147                 }
1148
1149                 /* If somebody is already doing CANCEL, or blocking ast came,
1150                  * skip this lock. */
1151                 if (lock->l_flags & LDLM_FL_BL_AST ||
1152                     lock->l_flags & LDLM_FL_CANCELING)
1153                         continue;
1154
1155                 if (lockmode_compat(lock->l_granted_mode, mode))
1156                         continue;
1157
1158                 /* If policy is given and this is IBITS lock, add to list only
1159                  * those locks that match by policy. */
1160                 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1161                     !(lock->l_policy_data.l_inodebits.bits &
1162                       policy->l_inodebits.bits))
1163                         continue;
1164
1165                 /* See CBPENDING comment in ldlm_cancel_lru */
1166                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1167                         lock_flags;
1168
1169                 LASSERT(list_empty(&lock->l_bl_ast));
1170                 list_add(&lock->l_bl_ast, cancels);
1171                 LDLM_LOCK_GET(lock);
1172                 count++;
1173         }
1174         unlock_res(res);
1175
1176         /* Handle only @count inserted locks. */
1177         left = count;
1178         list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1179                 int rc = LDLM_FL_LOCAL_ONLY;
1180
1181                 if (left-- == 0)
1182                         break;
1183                 if (flags & LDLM_FL_LOCAL_ONLY)
1184                         ldlm_lock_cancel(lock);
1185                 else
1186                         rc = ldlm_cli_cancel_local(lock);
1187
1188                 if (rc == LDLM_FL_BL_AST) {
1189                         CFS_LIST_HEAD(head);
1190
1191                         LDLM_DEBUG(lock, "Cancel lock separately");
1192                         list_del_init(&lock->l_bl_ast);
1193                         list_add(&lock->l_bl_ast, &head);
1194                         ldlm_cli_cancel_req(lock->l_conn_export, &head, 1, 0);
1195                         rc = LDLM_FL_LOCAL_ONLY;
1196                 }
1197                 if (rc == LDLM_FL_LOCAL_ONLY) {
1198                         /* CANCEL RPC should not be sent to server. */
1199                         list_del_init(&lock->l_bl_ast);
1200                         LDLM_LOCK_PUT(lock);
1201                         count--;
1202                 }
1203         }
1204         RETURN(count);
1205 }
1206
1207 /* If @req is NULL, send CANCEL request to server with handles of locks
1208  * in the @cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1209  * separately per lock.
1210  * If @req is not NULL, put handles of locks in @cancels into the request
1211  * buffer at the offset @off.
1212  * Destroy @cancels at the end. */
1213 int ldlm_cli_cancel_list(struct list_head *cancels, int count,
1214                          struct ptlrpc_request *req, int off, int flags)
1215 {
1216         struct ldlm_lock *lock;
1217         int res = 0;
1218         ENTRY;
1219
1220         if (list_empty(cancels) || count == 0)
1221                 RETURN(0);
1222
1223         /* XXX: requests (both batched and not) could be sent in parallel.
1224          * Usually it is enough to have just 1 RPC, but it is possible that
1225          * there are to many locks to be cancelled in LRU or on a resource.
1226          * It would also speed up the case when the server does not support
1227          * the feature. */
1228         while (count > 0) {
1229                 LASSERT(!list_empty(cancels));
1230                 lock = list_entry(cancels->next, struct ldlm_lock, l_bl_ast);
1231                 LASSERT(lock->l_conn_export);
1232
1233                 if (exp_connect_cancelset(lock->l_conn_export)) {
1234                         res = count;
1235                         if (req)
1236                                 ldlm_cancel_pack(req, off, cancels, count);
1237                         else
1238                                 res = ldlm_cli_cancel_req(lock->l_conn_export,
1239                                                           cancels, count, flags);
1240                 } else {
1241                         res = ldlm_cli_cancel_req(lock->l_conn_export,
1242                                                   cancels, 1, flags);
1243                 }
1244
1245                 if (res < 0) {
1246                         CERROR("ldlm_cli_cancel_list: %d\n", res);
1247                         res = count;
1248                 }
1249
1250                 count -= res;
1251                 ldlm_lock_list_put(cancels, l_bl_ast, res);
1252         }
1253         LASSERT(list_empty(cancels));
1254         LASSERT(count == 0);
1255         RETURN(0);
1256 }
1257
1258 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1259                                     const struct ldlm_res_id *res_id,
1260                                     ldlm_policy_data_t *policy,
1261                                     ldlm_mode_t mode, int flags, void *opaque)
1262 {
1263         struct ldlm_resource *res;
1264         CFS_LIST_HEAD(cancels);
1265         int count;
1266         int rc;
1267         ENTRY;
1268
1269         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1270         if (res == NULL) {
1271                 /* This is not a problem. */
1272                 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id->name[0]);
1273                 RETURN(0);
1274         }
1275
1276         count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
1277                                            0, flags, opaque);
1278         rc = ldlm_cli_cancel_list(&cancels, count, NULL, 0, flags);
1279         if (rc != ELDLM_OK)
1280                 CERROR("ldlm_cli_cancel_unused_resource: %d\n", rc);
1281
1282         ldlm_resource_putref(res);
1283         RETURN(0);
1284 }
1285
1286 static inline int have_no_nsresource(struct ldlm_namespace *ns)
1287 {
1288         int no_resource = 0;
1289
1290         spin_lock(&ns->ns_hash_lock);
1291         if (ns->ns_resources == 0)
1292                 no_resource = 1;
1293         spin_unlock(&ns->ns_hash_lock);
1294
1295         RETURN(no_resource);
1296 }
1297
1298 /* Cancel all locks on a namespace (or a specific resource, if given)
1299  * that have 0 readers/writers.
1300  *
1301  * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
1302  * to notify the server. */
1303 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1304                            const struct ldlm_res_id *res_id,
1305                            int flags, void *opaque)
1306 {
1307         int i;
1308         ENTRY;
1309
1310         if (ns == NULL)
1311                 RETURN(ELDLM_OK);
1312
1313         if (res_id)
1314                 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
1315                                                        LCK_MINMODE, flags,
1316                                                        opaque));
1317
1318         spin_lock(&ns->ns_hash_lock);
1319         for (i = 0; i < RES_HASH_SIZE; i++) {
1320                 struct list_head *tmp;
1321                 tmp = ns->ns_hash[i].next;
1322                 while (tmp != &(ns->ns_hash[i])) {
1323                         struct ldlm_resource *res;
1324                         int rc;
1325
1326                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
1327                         ldlm_resource_getref(res);
1328                         spin_unlock(&ns->ns_hash_lock);
1329
1330                         rc = ldlm_cli_cancel_unused_resource(ns, &res->lr_name,
1331                                                              NULL, LCK_MINMODE,
1332                                                              flags, opaque);
1333
1334                         if (rc)
1335                                 CERROR("ldlm_cli_cancel_unused ("LPU64"): %d\n",
1336                                        res->lr_name.name[0], rc);
1337
1338                         spin_lock(&ns->ns_hash_lock);
1339                         tmp = tmp->next;
1340                         ldlm_resource_putref_locked(res);
1341                 }
1342         }
1343         spin_unlock(&ns->ns_hash_lock);
1344
1345         RETURN(ELDLM_OK);
1346 }
1347
1348 /* join/split resource locks to/from lru list */
1349 int ldlm_cli_join_lru(struct ldlm_namespace *ns,
1350                       const struct ldlm_res_id *res_id, int join)
1351 {
1352         struct ldlm_resource *res;
1353         struct ldlm_lock *lock, *n;
1354         int count = 0;
1355         ENTRY;
1356
1357         LASSERT(ns->ns_client == LDLM_NAMESPACE_CLIENT);
1358
1359         res = ldlm_resource_get(ns, NULL, res_id, LDLM_EXTENT, 0);
1360         if (res == NULL)
1361                 RETURN(count);
1362         LASSERT(res->lr_type == LDLM_EXTENT);
1363
1364         lock_res(res);
1365         if (!join)
1366                 goto split;
1367
1368         list_for_each_entry_safe (lock, n, &res->lr_granted, l_res_link) {
1369                 if (list_empty(&lock->l_lru) &&
1370                     !lock->l_readers && !lock->l_writers &&
1371                     !(lock->l_flags & LDLM_FL_LOCAL) &&
1372                     !(lock->l_flags & LDLM_FL_CBPENDING)) {
1373                         lock->l_last_used = cfs_time_current();
1374                         spin_lock(&ns->ns_unused_lock);
1375                         LASSERT(ns->ns_nr_unused >= 0);
1376                         list_add_tail(&lock->l_lru, &ns->ns_unused_list);
1377                         ns->ns_nr_unused++;
1378                         spin_unlock(&ns->ns_unused_lock);
1379                         lock->l_flags &= ~LDLM_FL_NO_LRU;
1380                         LDLM_DEBUG(lock, "join lock to lru");
1381                         count++;
1382                 }
1383         }
1384         goto unlock;
1385 split:
1386         spin_lock(&ns->ns_unused_lock);
1387         list_for_each_entry_safe (lock, n, &ns->ns_unused_list, l_lru) {
1388                 if (lock->l_resource == res) {
1389                         ldlm_lock_remove_from_lru_nolock(lock);
1390                         lock->l_flags |= LDLM_FL_NO_LRU;
1391                         LDLM_DEBUG(lock, "split lock from lru");
1392                         count++;
1393                 }
1394         }
1395         spin_unlock(&ns->ns_unused_lock);
1396 unlock:
1397         unlock_res(res);
1398         ldlm_resource_putref(res);
1399         RETURN(count);
1400 }
1401
1402 /* Lock iterators. */
1403
1404 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
1405                           void *closure)
1406 {
1407         struct list_head *tmp, *next;
1408         struct ldlm_lock *lock;
1409         int rc = LDLM_ITER_CONTINUE;
1410
1411         ENTRY;
1412
1413         if (!res)
1414                 RETURN(LDLM_ITER_CONTINUE);
1415
1416         lock_res(res);
1417         list_for_each_safe(tmp, next, &res->lr_granted) {
1418                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1419
1420                 if (iter(lock, closure) == LDLM_ITER_STOP)
1421                         GOTO(out, rc = LDLM_ITER_STOP);
1422         }
1423
1424         list_for_each_safe(tmp, next, &res->lr_converting) {
1425                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1426
1427                 if (iter(lock, closure) == LDLM_ITER_STOP)
1428                         GOTO(out, rc = LDLM_ITER_STOP);
1429         }
1430
1431         list_for_each_safe(tmp, next, &res->lr_waiting) {
1432                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1433
1434                 if (iter(lock, closure) == LDLM_ITER_STOP)
1435                         GOTO(out, rc = LDLM_ITER_STOP);
1436         }
1437  out:
1438         unlock_res(res);
1439         RETURN(rc);
1440 }
1441
1442 struct iter_helper_data {
1443         ldlm_iterator_t iter;
1444         void *closure;
1445 };
1446
1447 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
1448 {
1449         struct iter_helper_data *helper = closure;
1450         return helper->iter(lock, helper->closure);
1451 }
1452
1453 static int ldlm_res_iter_helper(struct ldlm_resource *res, void *closure)
1454 {
1455         return ldlm_resource_foreach(res, ldlm_iter_helper, closure);
1456 }
1457
1458 int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
1459                            void *closure)
1460 {
1461         struct iter_helper_data helper = { iter: iter, closure: closure };
1462         return ldlm_namespace_foreach_res(ns, ldlm_res_iter_helper, &helper);
1463 }
1464
1465 int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
1466                                ldlm_res_iterator_t iter, void *closure)
1467 {
1468         int i, rc = LDLM_ITER_CONTINUE;
1469         struct ldlm_resource *res;
1470         struct list_head *tmp;
1471
1472         ENTRY;
1473         spin_lock(&ns->ns_hash_lock);
1474         for (i = 0; i < RES_HASH_SIZE; i++) {
1475                 tmp = ns->ns_hash[i].next;
1476                 while (tmp != &(ns->ns_hash[i])) {
1477                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
1478                         ldlm_resource_getref(res);
1479                         spin_unlock(&ns->ns_hash_lock);
1480
1481                         rc = iter(res, closure);
1482
1483                         spin_lock(&ns->ns_hash_lock);
1484                         tmp = tmp->next;
1485                         ldlm_resource_putref_locked(res);
1486                         if (rc == LDLM_ITER_STOP)
1487                                 GOTO(out, rc);
1488                 }
1489         }
1490  out:
1491         spin_unlock(&ns->ns_hash_lock);
1492         RETURN(rc);
1493 }
1494
1495 /* non-blocking function to manipulate a lock whose cb_data is being put away.*/
1496 void ldlm_resource_iterate(struct ldlm_namespace *ns,
1497                            const struct ldlm_res_id *res_id,
1498                            ldlm_iterator_t iter, void *data)
1499 {
1500         struct ldlm_resource *res;
1501         ENTRY;
1502
1503         if (ns == NULL) {
1504                 CERROR("must pass in namespace\n");
1505                 LBUG();
1506         }
1507
1508         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1509         if (res == NULL) {
1510                 EXIT;
1511                 return;
1512         }
1513
1514         ldlm_resource_foreach(res, iter, data);
1515         ldlm_resource_putref(res);
1516         EXIT;
1517 }
1518
1519 /* Lock replay */
1520
1521 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
1522 {
1523         struct list_head *list = closure;
1524
1525         /* we use l_pending_chain here, because it's unused on clients. */
1526         LASSERTF(list_empty(&lock->l_pending_chain),"lock %p next %p prev %p\n",
1527                  lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
1528         /* bug 9573: don't replay locks left after eviction */
1529         if (!(lock->l_flags & LDLM_FL_FAILED))
1530                 list_add(&lock->l_pending_chain, list);
1531         return LDLM_ITER_CONTINUE;
1532 }
1533
1534 static int replay_lock_interpret(struct ptlrpc_request *req,
1535                                  struct ldlm_async_args *aa, int rc)
1536 {
1537         struct ldlm_lock *lock;
1538         struct ldlm_reply *reply;
1539
1540         ENTRY;
1541         atomic_dec(&req->rq_import->imp_replay_inflight);
1542         if (rc != ELDLM_OK)
1543                 GOTO(out, rc);
1544
1545
1546         reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF, sizeof(*reply),
1547                                    lustre_swab_ldlm_reply);
1548         if (reply == NULL) {
1549                 CERROR("Can't unpack ldlm_reply\n");
1550                 GOTO (out, rc = -EPROTO);
1551         }
1552
1553         lock = ldlm_handle2lock(&aa->lock_handle);
1554         if (!lock) {
1555                 CERROR("received replay ack for unknown local cookie "LPX64
1556                        " remote cookie "LPX64 " from server %s id %s\n",
1557                        aa->lock_handle.cookie, reply->lock_handle.cookie,
1558                        req->rq_export->exp_client_uuid.uuid,
1559                        libcfs_id2str(req->rq_peer));
1560                 GOTO(out, rc = -ESTALE);
1561         }
1562
1563         lock->l_remote_handle = reply->lock_handle;
1564         LDLM_DEBUG(lock, "replayed lock:");
1565         ptlrpc_import_recovery_state_machine(req->rq_import);
1566         LDLM_LOCK_PUT(lock);
1567 out:
1568         if (rc != ELDLM_OK)
1569                 ptlrpc_connect_import(req->rq_import, NULL);
1570
1571
1572         RETURN(rc);
1573 }
1574
1575 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
1576 {
1577         struct ptlrpc_request *req;
1578         struct ldlm_request *body;
1579         struct ldlm_reply *reply;
1580         struct ldlm_async_args *aa;
1581         int buffers = 2;
1582         int size[3] = { sizeof(struct ptlrpc_body) };
1583         int flags;
1584         ENTRY;
1585
1586
1587         /* Bug 11974: Do not replay a lock which is actively being canceled */
1588         if (lock->l_flags & LDLM_FL_CANCELING) {
1589                 LDLM_DEBUG(lock, "Not replaying canceled lock:");
1590                 RETURN(0);
1591         }
1592
1593         /* If this is reply-less callback lock, we cannot replay it, since
1594          * server might have long dropped it, but notification of that event was
1595          * lost by network. (and server granted conflicting lock already) */
1596         if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1597                 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
1598                 ldlm_lock_cancel(lock);
1599                 RETURN(0);
1600         }
1601         /*
1602          * If granted mode matches the requested mode, this lock is granted.
1603          *
1604          * If they differ, but we have a granted mode, then we were granted
1605          * one mode and now want another: ergo, converting.
1606          *
1607          * If we haven't been granted anything and are on a resource list,
1608          * then we're blocked/waiting.
1609          *
1610          * If we haven't been granted anything and we're NOT on a resource list,
1611          * then we haven't got a reply yet and don't have a known disposition.
1612          * This happens whenever a lock enqueue is the request that triggers
1613          * recovery.
1614          */
1615         if (lock->l_granted_mode == lock->l_req_mode)
1616                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
1617         else if (lock->l_granted_mode)
1618                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
1619         else if (!list_empty(&lock->l_res_link))
1620                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
1621         else
1622                 flags = LDLM_FL_REPLAY;
1623
1624         size[DLM_LOCKREQ_OFF] = sizeof(*body);
1625         req = ptlrpc_prep_req(imp, LUSTRE_DLM_VERSION, LDLM_ENQUEUE, 2, size,
1626                               NULL);
1627         if (!req)
1628                 RETURN(-ENOMEM);
1629
1630         /* We're part of recovery, so don't wait for it. */
1631         req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
1632
1633         body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, sizeof(*body));
1634         ldlm_lock2desc(lock, &body->lock_desc);
1635         body->lock_flags = flags;
1636
1637         ldlm_lock2handle(lock, &body->lock_handle[0]);
1638         size[DLM_LOCKREPLY_OFF] = sizeof(*reply);
1639         if (lock->l_lvb_len != 0) {
1640                 buffers = 3;
1641                 size[DLM_REPLY_REC_OFF] = lock->l_lvb_len;
1642         }
1643         ptlrpc_req_set_repsize(req, buffers, size);
1644         /* notify the server we've replayed all requests.
1645          * also, we mark the request to be put on a dedicated
1646          * queue to be processed after all request replayes.
1647          * bug 6063 */
1648         lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
1649
1650         LDLM_DEBUG(lock, "replaying lock:");
1651
1652         atomic_inc(&req->rq_import->imp_replay_inflight);
1653         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1654         aa = (struct ldlm_async_args *)&req->rq_async_args;
1655         aa->lock_handle = body->lock_handle[0];
1656         req->rq_interpret_reply = replay_lock_interpret;
1657         ptlrpcd_add_req(req);
1658
1659         RETURN(0);
1660 }
1661
1662 int ldlm_replay_locks(struct obd_import *imp)
1663 {
1664         struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
1665         struct list_head list;
1666         struct ldlm_lock *lock, *next;
1667         int rc = 0;
1668
1669         ENTRY;
1670         CFS_INIT_LIST_HEAD(&list);
1671
1672         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1673
1674         /* ensure this doesn't fall to 0 before all have been queued */
1675         atomic_inc(&imp->imp_replay_inflight);
1676
1677         (void)ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
1678
1679         list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
1680                 list_del_init(&lock->l_pending_chain);
1681                 if (rc)
1682                         continue; /* or try to do the rest? */
1683                 rc = replay_one_lock(imp, lock);
1684         }
1685
1686         atomic_dec(&imp->imp_replay_inflight);
1687
1688         RETURN(rc);
1689 }