lustre/ldlm/ldlm_request.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (C) 2002, 2003 Cluster File Systems, Inc.
   5  *
   6  *   This file is part of Lustre, http://www.lustre.org.
   7  *
   8  *   Lustre is free software; you can redistribute it and/or
   9  *   modify it under the terms of version 2 of the GNU General Public
  10  *   License as published by the Free Software Foundation.
  11  *
  12  *   Lustre is distributed in the hope that it will be useful,
  13  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  *   GNU General Public License for more details.
  16  *
  17  *   You should have received a copy of the GNU General Public License
  18  *   along with Lustre; if not, write to the Free Software
  19  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  20  */
  21
  22 #define DEBUG_SUBSYSTEM S_LDLM
  23 #ifndef __KERNEL__
  24 #include <signal.h>
  25 #include <liblustre.h>
  26 #endif
  27
  28 #include <linux/lustre_dlm.h>
  29 #include <linux/obd_class.h>
  30 #include <linux/obd.h>
  31
  32 #include "ldlm_internal.h"
  33
  34 static void interrupted_completion_wait(void *data)
  35 {
  36 }
  37
  38 struct lock_wait_data {
  39         struct ldlm_lock *lwd_lock;
  40         int               lwd_generation;
  41 };
  42
  43 int ldlm_expired_completion_wait(void *data)
  44 {
  45         static unsigned long next_dump = 0;
  46         struct lock_wait_data *lwd = data;
  47         struct ldlm_lock *lock = lwd->lwd_lock;
  48         struct obd_import *imp;
  49         struct obd_device *obd;
  50
  51         if (lock->l_conn_export == NULL) {
  52                 LDLM_ERROR(lock, "lock timed out; not entering recovery in "
  53                            "server code, just going back to sleep");
  54                 if (time_after(jiffies, next_dump)) {
  55                         unsigned int debug = portal_debug;
  56                         next_dump = jiffies + 300 * HZ;
  57                         portal_debug |= D_OTHER;
  58                         ldlm_namespace_dump(lock->l_resource->lr_namespace);
  59                         portal_debug = debug;
  60                 }
  61                 RETURN(0);
  62         }
  63
  64         obd = lock->l_conn_export->exp_obd;
  65         imp = obd->u.cli.cl_import;
  66         ptlrpc_fail_import(imp, lwd->lwd_generation);
  67         LDLM_ERROR(lock, "lock timed out, entering recovery for %s@%s",
  68                    imp->imp_target_uuid.uuid,
  69                    imp->imp_connection->c_remote_uuid.uuid);
  70
  71         RETURN(0);
  72 }
  73
  74 int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data)
  75 {
  76         /* XXX ALLOCATE - 160 mytes */
  77         struct lock_wait_data lwd;
  78         unsigned long irqflags;
  79         struct obd_device *obd;
  80         struct obd_import *imp = NULL;
  81         int rc = 0;
  82         struct l_wait_info lwi;
  83         ENTRY;
  84
  85         if (flags == LDLM_FL_WAIT_NOREPROC)
  86                 goto noreproc;
  87
  88         if (flags == 0) {
  89                 wake_up(&lock->l_waitq);
  90                 RETURN(0);
  91         }
  92
  93         if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
  94                        LDLM_FL_BLOCK_CONV)))
  95                 RETURN(0);
  96
  97         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
  98                    "sleeping");
  99         ldlm_lock_dump(D_OTHER, lock, 0);
 100         ldlm_reprocess_all(lock->l_resource);
 101
 102 noreproc:
 103
 104         obd = class_exp2obd(lock->l_conn_export);
 105
 106         /* if this is a local lock, then there is no import */
 107         if (obd != NULL)
 108                 imp = obd->u.cli.cl_import;
 109
 110         lwd.lwd_lock = lock;
 111
 112         lwi = LWI_TIMEOUT_INTR(obd_timeout * HZ, ldlm_expired_completion_wait,
 113                                interrupted_completion_wait, &lwd);
 114         if (imp != NULL) {
 115                 spin_lock_irqsave(&imp->imp_lock, irqflags);
 116                 lwd.lwd_generation = imp->imp_generation;
 117                 spin_unlock_irqrestore(&imp->imp_lock, irqflags);
 118         }
 119
 120         /* Go to sleep until the lock is granted or cancelled. */
 121         rc = l_wait_event(lock->l_waitq,
 122                           ((lock->l_req_mode == lock->l_granted_mode) ||
 123                            (lock->l_flags & LDLM_FL_CANCEL)), &lwi);
 124
 125         if (lock->l_destroyed) {
 126                 LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed");
 127                 RETURN(-EIO);
 128         }
 129
 130         if (rc) {
 131                 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
 132                            rc);
 133                 RETURN(rc);
 134         }
 135
 136         LDLM_DEBUG(lock, "client-side enqueue waking up: granted");
 137         RETURN(0);
 138 }
 139
 140 static int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
 141                                   struct lustre_handle *parent_lockh,
 142                                   struct ldlm_res_id res_id,
 143                                   __u32 type,
 144                                   void *cookie, int cookielen,
 145                                   ldlm_mode_t mode,
 146                                   int *flags,
 147                                   ldlm_completion_callback completion,
 148                                   ldlm_blocking_callback blocking,
 149                                   void *data,
 150                                   struct lustre_handle *lockh)
 151 {
 152         struct ldlm_lock *lock;
 153         int err;
 154         ENTRY;
 155
 156         if (ns->ns_client) {
 157                 CERROR("Trying to enqueue local lock in a shadow namespace\n");
 158                 LBUG();
 159         }
 160
 161         lock = ldlm_lock_create(ns, parent_lockh, res_id, type, mode,
 162                                 blocking, completion, data);
 163         if (!lock)
 164                 GOTO(out_nolock, err = -ENOMEM);
 165         LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
 166
 167         ldlm_lock_addref_internal(lock, mode);
 168         ldlm_lock2handle(lock, lockh);
 169         lock->l_flags |= LDLM_FL_LOCAL;
 170
 171         err = ldlm_lock_enqueue(ns, &lock, cookie, cookielen, flags);
 172         if (err != ELDLM_OK)
 173                 GOTO(out, err);
 174
 175         if (type != LDLM_PLAIN)
 176                 memcpy(cookie, &lock->l_policy_data, cookielen);
 177         if ((*flags) & LDLM_FL_LOCK_CHANGED)
 178                 memcpy(&res_id, &lock->l_resource->lr_name, sizeof(res_id));
 179
 180         LDLM_DEBUG_NOLOCK("client-side local enqueue handler END (lock %p)",
 181                           lock);
 182
 183         if (lock->l_completion_ast)
 184                 lock->l_completion_ast(lock, *flags, NULL);
 185
 186         LDLM_DEBUG(lock, "client-side local enqueue END");
 187         EXIT;
 188  out:
 189         LDLM_LOCK_PUT(lock);
 190  out_nolock:
 191         return err;
 192 }
 193
 194 static void failed_lock_cleanup(struct ldlm_namespace *ns,
 195                                 struct ldlm_lock *lock,
 196                                 struct lustre_handle *lockh, int mode)
 197 {
 198         /* Set a flag to prevent us from sending a CANCEL (bug 407) */
 199         l_lock(&ns->ns_lock);
 200         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 201         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 202         l_unlock(&ns->ns_lock);
 203
 204         ldlm_lock_decref_and_cancel(lockh, mode);
 205 }
 206
 207 int ldlm_cli_enqueue(struct obd_export *exp,
 208                      struct ptlrpc_request *req,
 209                      struct ldlm_namespace *ns,
 210                      struct lustre_handle *parent_lock_handle,
 211                      struct ldlm_res_id res_id,
 212                      __u32 type,
 213                      void *cookie, int cookielen,
 214                      ldlm_mode_t mode,
 215                      int *flags,
 216                      ldlm_completion_callback completion,
 217                      ldlm_blocking_callback blocking,
 218                      void *data,
 219                      struct lustre_handle *lockh)
 220 {
 221         struct ldlm_lock *lock;
 222         struct ldlm_request *body;
 223         struct ldlm_reply *reply;
 224         int rc, size = sizeof(*body), req_passed_in = 1, is_replay;
 225         ENTRY;
 226
 227         is_replay = *flags & LDLM_FL_REPLAY;
 228         LASSERT(exp != NULL || !is_replay);
 229
 230         if (exp == NULL) {
 231                 rc = ldlm_cli_enqueue_local(ns, parent_lock_handle, res_id,
 232                                             type, cookie, cookielen, mode,
 233                                             flags, completion, blocking, data,
 234                                             lockh);
 235                 RETURN(rc);
 236         }
 237
 238         /* If we're replaying this lock, just check some invariants.
 239          * If we're creating a new lock, get everything all setup nice. */
 240         if (is_replay) {
 241                 lock = ldlm_handle2lock(lockh);
 242                 LDLM_DEBUG(lock, "client-side enqueue START");
 243                 LASSERT(exp == lock->l_conn_export);
 244         } else {
 245                 lock = ldlm_lock_create(ns, parent_lock_handle, res_id, type,
 246                                         mode, blocking, completion, data);
 247                 if (lock == NULL)
 248                         GOTO(out_nolock, rc = -ENOMEM);
 249                 /* for the local lock, add the reference */
 250                 ldlm_lock_addref_internal(lock, mode);
 251                 ldlm_lock2handle(lock, lockh);
 252                 if (type != LDLM_PLAIN)
 253                         memcpy(&lock->l_policy_data, cookie, cookielen);
 254                 LDLM_DEBUG(lock, "client-side enqueue START");
 255         }
 256
 257         if (req == NULL) {
 258                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LDLM_ENQUEUE, 1,
 259                                       &size, NULL);
 260                 if (!req)
 261                         GOTO(out, rc = -ENOMEM);
 262                 req_passed_in = 0;
 263         } else if (req->rq_reqmsg->buflens[0] != sizeof(*body))
 264                 LBUG();
 265
 266         /* Dump lock data into the request buffer */
 267         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 268         ldlm_lock2desc(lock, &body->lock_desc);
 269         body->lock_flags = *flags;
 270
 271         memcpy(&body->lock_handle1, lockh, sizeof(*lockh));
 272         if (parent_lock_handle)
 273                 memcpy(&body->lock_handle2, parent_lock_handle,
 274                        sizeof(body->lock_handle2));
 275
 276         /* Continue as normal. */
 277         if (!req_passed_in) {
 278                 size = sizeof(*reply);
 279                 req->rq_replen = lustre_msg_size(1, &size);
 280         }
 281         lock->l_conn_export = exp;
 282         lock->l_export = NULL;
 283         lock->l_blocking_ast = blocking;
 284
 285         LDLM_DEBUG(lock, "sending request");
 286         rc = ptlrpc_queue_wait(req);
 287
 288         if (rc != ELDLM_OK) {
 289                 LASSERT(!is_replay);
 290                 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
 291                            rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
 292                 failed_lock_cleanup(ns, lock, lockh, mode);
 293                 if (rc == ELDLM_LOCK_ABORTED) {
 294                         /* Before we return, swab the reply */
 295                         reply = lustre_swab_repbuf(req, 0, sizeof(*reply),
 296                                                    lustre_swab_ldlm_reply);
 297                         if (reply == NULL) {
 298                                 CERROR("Can't unpack ldlm_reply\n");
 299                                 GOTO(out_req, rc = -EPROTO);
 300                         }
 301                 }
 302                 GOTO(out_req, rc);
 303         }
 304
 305         reply = lustre_swab_repbuf(req, 0, sizeof(*reply),
 306                                    lustre_swab_ldlm_reply);
 307         if (reply == NULL) {
 308                 CERROR("Can't unpack ldlm_reply\n");
 309                 GOTO(out_req, rc = -EPROTO);
 310         }
 311
 312         memcpy(&lock->l_remote_handle, &reply->lock_handle,
 313                sizeof(lock->l_remote_handle));
 314         *flags = reply->lock_flags;
 315
 316         CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%x\n",
 317                lock, reply->lock_handle.cookie, *flags);
 318         if (type == LDLM_EXTENT) {
 319                 CDEBUG(D_INFO, "requested extent: "LPU64" -> "LPU64", got "
 320                        "extent "LPU64" -> "LPU64"\n",
 321                        body->lock_desc.l_policy_data.l_extent.start,
 322                        body->lock_desc.l_policy_data.l_extent.end,
 323                        reply->lock_policy_data.l_extent.start,
 324                        reply->lock_policy_data.l_extent.end);
 325
 326                 cookie = &reply->lock_policy_data; /* FIXME bug 267 */
 327                 cookielen = sizeof(struct ldlm_extent);
 328         } else if (type == LDLM_FLOCK) {
 329                 cookie = &reply->lock_policy_data;
 330                 cookielen = sizeof(struct ldlm_flock);
 331         }
 332
 333         /* If enqueue returned a blocked lock but the completion handler has
 334          * already run, then it fixed up the resource and we don't need to do it
 335          * again. */
 336         if ((*flags) & LDLM_FL_LOCK_CHANGED) {
 337                 int newmode = reply->lock_mode;
 338                 LASSERT(!is_replay);
 339                 if (newmode && newmode != lock->l_req_mode) {
 340                         LDLM_DEBUG(lock, "server returned different mode %s",
 341                                    ldlm_lockname[newmode]);
 342                         lock->l_req_mode = newmode;
 343                 }
 344
 345                 if (reply->lock_resource_name.name[0] !=
 346                     lock->l_resource->lr_name.name[0]) {
 347                         CDEBUG(D_INFO, "remote intent success, locking %ld "
 348                                "instead of %ld\n",
 349                                (long)reply->lock_resource_name.name[0],
 350                                (long)lock->l_resource->lr_name.name[0]);
 351
 352                         ldlm_lock_change_resource(ns, lock,
 353                                                   reply->lock_resource_name);
 354                         if (lock->l_resource == NULL) {
 355                                 LBUG();
 356                                 GOTO(out_req, rc = -ENOMEM);
 357                         }
 358                         LDLM_DEBUG(lock, "client-side enqueue, new resource");
 359                 }
 360         }
 361         if ((*flags) & LDLM_FL_AST_SENT) {
 362                 l_lock(&ns->ns_lock);
 363                 lock->l_flags |= LDLM_FL_CBPENDING;
 364                 l_unlock(&ns->ns_lock);
 365                 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
 366         }
 367
 368         if (!is_replay) {
 369                 rc = ldlm_lock_enqueue(ns, &lock, cookie, cookielen, flags);
 370                 if (lock->l_completion_ast != NULL) {
 371                         int err = lock->l_completion_ast(lock, *flags, NULL);
 372                         if (err)
 373                                 failed_lock_cleanup(ns, lock, lockh, mode);
 374                         if (!rc)
 375                                 rc = err;
 376                 }
 377         }
 378
 379         LDLM_DEBUG(lock, "client-side enqueue END");
 380         EXIT;
 381  out_req:
 382         if (!req_passed_in)
 383                 ptlrpc_req_finished(req);
 384  out:
 385         LDLM_LOCK_PUT(lock);
 386  out_nolock:
 387         return rc;
 388 }
 389
 390 int ldlm_cli_replay_enqueue(struct ldlm_lock *lock)
 391 {
 392         struct lustre_handle lockh;
 393         struct ldlm_res_id junk;
 394         int flags = LDLM_FL_REPLAY;
 395         ldlm_lock2handle(lock, &lockh);
 396         return ldlm_cli_enqueue(lock->l_conn_export, NULL, NULL, NULL, junk,
 397                                 lock->l_resource->lr_type, NULL, 0, -1, &flags,
 398                                 NULL, NULL, NULL, &lockh);
 399 }
 400
 401 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
 402                                   int *flags)
 403 {
 404         ENTRY;
 405         if (lock->l_resource->lr_namespace->ns_client) {
 406                 CERROR("Trying to cancel local lock\n");
 407                 LBUG();
 408         }
 409         LDLM_DEBUG(lock, "client-side local convert");
 410
 411         ldlm_lock_convert(lock, new_mode, flags);
 412         ldlm_reprocess_all(lock->l_resource);
 413
 414         LDLM_DEBUG(lock, "client-side local convert handler END");
 415         LDLM_LOCK_PUT(lock);
 416         RETURN(0);
 417 }
 418
 419 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
 420  * conversion of locks which are on the waiting or converting queue */
 421 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, int *flags)
 422 {
 423         struct ldlm_request *body;
 424         struct ldlm_reply *reply;
 425         struct ldlm_lock *lock;
 426         struct ldlm_resource *res;
 427         struct ptlrpc_request *req;
 428         int rc, size = sizeof(*body);
 429         ENTRY;
 430
 431         lock = ldlm_handle2lock(lockh);
 432         if (!lock) {
 433                 LBUG();
 434                 RETURN(-EINVAL);
 435         }
 436         *flags = 0;
 437
 438         if (lock->l_conn_export == NULL)
 439                 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
 440
 441         LDLM_DEBUG(lock, "client-side convert");
 442
 443         req = ptlrpc_prep_req(class_exp2cliimp(lock->l_conn_export),
 444                               LDLM_CONVERT, 1, &size, NULL);
 445         if (!req)
 446                 GOTO(out, rc = -ENOMEM);
 447
 448         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 449         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 450                sizeof(body->lock_handle1));
 451
 452         body->lock_desc.l_req_mode = new_mode;
 453         body->lock_flags = *flags;
 454
 455         size = sizeof(*reply);
 456         req->rq_replen = lustre_msg_size(1, &size);
 457
 458         rc = ptlrpc_queue_wait(req);
 459         if (rc != ELDLM_OK)
 460                 GOTO(out, rc);
 461
 462         reply = lustre_swab_repbuf(req, 0, sizeof (*reply),
 463                                    lustre_swab_ldlm_reply);
 464         if (reply == NULL) {
 465                 CERROR ("Can't unpack ldlm_reply\n");
 466                 GOTO (out, rc = -EPROTO);
 467         }
 468
 469         res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
 470         if (res != NULL)
 471                 ldlm_reprocess_all(res);
 472         /* Go to sleep until the lock is granted. */
 473         /* FIXME: or cancelled. */
 474         if (lock->l_completion_ast)
 475                 lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC, NULL);
 476         EXIT;
 477  out:
 478         LDLM_LOCK_PUT(lock);
 479         ptlrpc_req_finished(req);
 480         return rc;
 481 }
 482
 483 int ldlm_cli_cancel(struct lustre_handle *lockh)
 484 {
 485         struct ptlrpc_request *req;
 486         struct ldlm_lock *lock;
 487         struct ldlm_request *body;
 488         int rc = 0, size = sizeof(*body);
 489         ENTRY;
 490
 491         /* concurrent cancels on the same handle can happen */
 492         lock = __ldlm_handle2lock(lockh, LDLM_FL_CANCELING);
 493         if (lock == NULL)
 494                 RETURN(0);
 495
 496         if (lock->l_conn_export) {
 497                 int local_only;
 498                 struct obd_import *imp;
 499
 500                 LDLM_DEBUG(lock, "client-side cancel");
 501                 /* Set this flag to prevent others from getting new references*/
 502                 l_lock(&lock->l_resource->lr_namespace->ns_lock);
 503                 lock->l_flags |= LDLM_FL_CBPENDING;
 504                 local_only = (lock->l_flags & LDLM_FL_LOCAL_ONLY);
 505                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 506                 ldlm_cancel_callback(lock);
 507
 508                 if (local_only) {
 509                         CDEBUG(D_INFO, "not sending request (at caller's "
 510                                "instruction)\n");
 511                         goto local_cancel;
 512                 }
 513
 514         restart:
 515                 imp = class_exp2cliimp(lock->l_conn_export);
 516                 if (imp == NULL || imp->imp_invalid) {
 517                         CDEBUG(D_HA, "skipping cancel on invalid import %p\n",
 518                                imp);
 519                         goto local_cancel;
 520                 }
 521
 522                 req = ptlrpc_prep_req(imp, LDLM_CANCEL, 1, &size, NULL);
 523                 if (!req)
 524                         GOTO(out, rc = -ENOMEM);
 525                 req->rq_no_resend = 1;
 526
 527                 /* XXX FIXME bug 249 */
 528                 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
 529                 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
 530
 531                 body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 532                 memcpy(&body->lock_handle1, &lock->l_remote_handle,
 533                        sizeof(body->lock_handle1));
 534
 535                 req->rq_replen = lustre_msg_size(0, NULL);
 536
 537                 rc = ptlrpc_queue_wait(req);
 538
 539                 if (rc == ESTALE) {
 540                         CERROR("client/server (nid "LPU64") out of sync--not "
 541                                "fatal\n",
 542                                req->rq_import->imp_connection->c_peer.peer_nid);
 543                 } else if (rc == -ETIMEDOUT) {
 544                         ptlrpc_req_finished(req);
 545                         GOTO(restart, rc);
 546                 } else if (rc != ELDLM_OK) {
 547                         CERROR("Got rc %d from cancel RPC: canceling "
 548                                "anyway\n", rc);
 549                 }
 550
 551                 ptlrpc_req_finished(req);
 552         local_cancel:
 553                 ldlm_lock_cancel(lock);
 554         } else {
 555                 if (lock->l_resource->lr_namespace->ns_client) {
 556                         LDLM_ERROR(lock, "Trying to cancel local lock\n");
 557                         LBUG();
 558                 }
 559                 LDLM_DEBUG(lock, "client-side local cancel");
 560                 ldlm_lock_cancel(lock);
 561                 ldlm_reprocess_all(lock->l_resource);
 562                 LDLM_DEBUG(lock, "client-side local cancel handler END");
 563         }
 564
 565         EXIT;
 566  out:
 567         LDLM_LOCK_PUT(lock);
 568         return rc;
 569 }
 570
 571 int ldlm_cancel_lru(struct ldlm_namespace *ns)
 572 {
 573         struct list_head *tmp, *next, list = LIST_HEAD_INIT(list);
 574         int count, rc = 0;
 575         struct ldlm_ast_work *w;
 576         ENTRY;
 577
 578         l_lock(&ns->ns_lock);
 579         count = ns->ns_nr_unused - ns->ns_max_unused;
 580
 581         if (count <= 0) {
 582                 l_unlock(&ns->ns_lock);
 583                 RETURN(0);
 584         }
 585
 586         list_for_each_safe(tmp, next, &ns->ns_unused_list) {
 587                 struct ldlm_lock *lock;
 588                 lock = list_entry(tmp, struct ldlm_lock, l_lru);
 589
 590                 LASSERT(!lock->l_readers && !lock->l_writers);
 591
 592                 /* Setting the CBPENDING flag is a little misleading, but
 593                  * prevents an important race; namely, once CBPENDING is set,
 594                  * the lock can accumulate no more readers/writers.  Since
 595                  * readers and writers are already zero here, ldlm_lock_decref
 596                  * won't see this flag and call l_blocking_ast */
 597                 lock->l_flags |= LDLM_FL_CBPENDING;
 598
 599                 OBD_ALLOC(w, sizeof(*w));
 600                 LASSERT(w);
 601
 602                 w->w_lock = LDLM_LOCK_GET(lock);
 603                 list_add(&w->w_list, &list);
 604                 ldlm_lock_remove_from_lru(lock);
 605
 606                 if (--count == 0)
 607                         break;
 608         }
 609         l_unlock(&ns->ns_lock);
 610
 611         list_for_each_safe(tmp, next, &list) {
 612                 struct lustre_handle lockh;
 613                 int rc;
 614                 w = list_entry(tmp, struct ldlm_ast_work, w_list);
 615
 616                 ldlm_lock2handle(w->w_lock, &lockh);
 617                 rc = ldlm_cli_cancel(&lockh);
 618                 if (rc != ELDLM_OK)
 619                         CDEBUG(D_INFO, "ldlm_cli_cancel: %d\n", rc);
 620
 621                 list_del(&w->w_list);
 622                 LDLM_LOCK_PUT(w->w_lock);
 623                 OBD_FREE(w, sizeof(*w));
 624         }
 625
 626         RETURN(rc);
 627 }
 628
 629 static int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
 630                                            struct ldlm_res_id res_id, int flags,
 631                                            void *opaque)
 632 {
 633         struct ldlm_resource *res;
 634         struct list_head *tmp, *next, list = LIST_HEAD_INIT(list);
 635         struct ldlm_ast_work *w;
 636         ENTRY;
 637
 638         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
 639         if (res == NULL) {
 640                 /* This is not a problem. */
 641                 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id.name[0]);
 642                 RETURN(0);
 643         }
 644
 645         l_lock(&ns->ns_lock);
 646         list_for_each(tmp, &res->lr_granted) {
 647                 struct ldlm_lock *lock;
 648                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 649
 650                 if (opaque != NULL && lock->l_ast_data != opaque) {
 651                         LDLM_ERROR(lock, "data %p doesn't match opaque %p",
 652                                    lock->l_ast_data, opaque);
 653                         //LBUG();
 654                         continue;
 655                 }
 656
 657                 if (lock->l_readers || lock->l_writers) {
 658                         if (flags & LDLM_FL_WARN) {
 659                                 LDLM_ERROR(lock, "lock in use");
 660                                 //LBUG();
 661                         }
 662                         continue;
 663                 }
 664
 665                 /* See CBPENDING comment in ldlm_cancel_lru */
 666                 lock->l_flags |= LDLM_FL_CBPENDING;
 667
 668                 OBD_ALLOC(w, sizeof(*w));
 669                 LASSERT(w);
 670
 671                 w->w_lock = LDLM_LOCK_GET(lock);
 672
 673                 /* Prevent the cancel callback from being called by setting
 674                  * LDLM_FL_CANCEL in the lock.  Very sneaky. -p */
 675                 if (flags & LDLM_FL_NO_CALLBACK)
 676                         w->w_lock->l_flags |= LDLM_FL_CANCEL;
 677
 678                 list_add(&w->w_list, &list);
 679         }
 680         l_unlock(&ns->ns_lock);
 681
 682         list_for_each_safe(tmp, next, &list) {
 683                 struct lustre_handle lockh;
 684                 int rc;
 685                 w = list_entry(tmp, struct ldlm_ast_work, w_list);
 686
 687                 if (flags & LDLM_FL_LOCAL_ONLY) {
 688                         ldlm_lock_cancel(w->w_lock);
 689                 } else {
 690                         ldlm_lock2handle(w->w_lock, &lockh);
 691                         rc = ldlm_cli_cancel(&lockh);
 692                         if (rc != ELDLM_OK)
 693                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 694                 }
 695                 list_del(&w->w_list);
 696                 LDLM_LOCK_PUT(w->w_lock);
 697                 OBD_FREE(w, sizeof(*w));
 698         }
 699
 700         ldlm_resource_putref(res);
 701
 702         RETURN(0);
 703 }
 704
 705 /* Cancel all locks on a namespace (or a specific resource, if given)
 706  * that have 0 readers/writers.
 707  *
 708  * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
 709  * to notify the server.
 710  * If flags & LDLM_FL_NO_CALLBACK, don't run the cancel callback.
 711  * If flags & LDLM_FL_WARN, print a warning if some locks are still in use. */
 712 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
 713                            struct ldlm_res_id *res_id, int flags, void *opaque)
 714 {
 715         int i;
 716         ENTRY;
 717
 718         if (ns == NULL)
 719                 RETURN(ELDLM_OK);
 720
 721         if (res_id)
 722                 RETURN(ldlm_cli_cancel_unused_resource(ns, *res_id, flags,
 723                                                        opaque));
 724
 725         l_lock(&ns->ns_lock);
 726         for (i = 0; i < RES_HASH_SIZE; i++) {
 727                 struct list_head *tmp, *pos;
 728                 list_for_each_safe(tmp, pos, &(ns->ns_hash[i])) {
 729                         int rc;
 730                         struct ldlm_resource *res;
 731                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
 732                         ldlm_resource_getref(res);
 733
 734                         rc = ldlm_cli_cancel_unused_resource(ns, res->lr_name,
 735                                                              flags, opaque);
 736
 737                         if (rc)
 738                                 CERROR("cancel_unused_res ("LPU64"): %d\n",
 739                                        res->lr_name.name[0], rc);
 740                         ldlm_resource_putref(res);
 741                 }
 742         }
 743         l_unlock(&ns->ns_lock);
 744
 745         RETURN(ELDLM_OK);
 746 }
 747
 748 /* Lock iterators. */
 749
 750 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
 751                           void *closure)
 752 {
 753         struct list_head *tmp, *next;
 754         struct ldlm_lock *lock;
 755         int rc = LDLM_ITER_CONTINUE;
 756         struct ldlm_namespace *ns = res->lr_namespace;
 757
 758         ENTRY;
 759
 760         if (!res)
 761                 RETURN(LDLM_ITER_CONTINUE);
 762
 763         l_lock(&ns->ns_lock);
 764         list_for_each_safe(tmp, next, &res->lr_granted) {
 765                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 766
 767                 if (iter(lock, closure) == LDLM_ITER_STOP)
 768                         GOTO(out, rc = LDLM_ITER_STOP);
 769         }
 770
 771         list_for_each_safe(tmp, next, &res->lr_converting) {
 772                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 773
 774                 if (iter(lock, closure) == LDLM_ITER_STOP)
 775                         GOTO(out, rc = LDLM_ITER_STOP);
 776         }
 777
 778         list_for_each_safe(tmp, next, &res->lr_waiting) {
 779                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 780
 781                 if (iter(lock, closure) == LDLM_ITER_STOP)
 782                         GOTO(out, rc = LDLM_ITER_STOP);
 783         }
 784  out:
 785         l_unlock(&ns->ns_lock);
 786         RETURN(rc);
 787 }
 788
 789 struct iter_helper_data {
 790         ldlm_iterator_t iter;
 791         void *closure;
 792 };
 793
 794 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
 795 {
 796         struct iter_helper_data *helper = closure;
 797         return helper->iter(lock, helper->closure);
 798 }
 799
 800 static int ldlm_res_iter_helper(struct ldlm_resource *res, void *closure)
 801 {
 802         return ldlm_resource_foreach(res, ldlm_iter_helper, closure);
 803 }
 804
 805 int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
 806                            void *closure)
 807 {
 808         struct iter_helper_data helper = { iter: iter, closure: closure };
 809         return ldlm_namespace_foreach_res(ns, ldlm_res_iter_helper, &helper);
 810 }
 811
 812 int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
 813                                ldlm_res_iterator_t iter, void *closure)
 814 {
 815         int i, rc = LDLM_ITER_CONTINUE;
 816
 817         l_lock(&ns->ns_lock);
 818         for (i = 0; i < RES_HASH_SIZE; i++) {
 819                 struct list_head *tmp, *next;
 820                 list_for_each_safe(tmp, next, &(ns->ns_hash[i])) {
 821                         struct ldlm_resource *res =
 822                                 list_entry(tmp, struct ldlm_resource, lr_hash);
 823
 824                         ldlm_resource_getref(res);
 825                         rc = iter(res, closure);
 826                         ldlm_resource_putref(res);
 827                         if (rc == LDLM_ITER_STOP)
 828                                 GOTO(out, rc);
 829                 }
 830         }
 831  out:
 832         l_unlock(&ns->ns_lock);
 833         RETURN(rc);
 834 }
 835
 836 /* non-blocking function to manipulate a lock whose cb_data is being put away.*/
 837 void ldlm_change_cbdata(struct ldlm_namespace *ns,
 838                        struct ldlm_res_id *res_id,
 839                        ldlm_iterator_t iter,
 840                        void *data)
 841 {
 842         struct ldlm_resource *res;
 843         ENTRY;
 844
 845         if (ns == NULL) {
 846                 CERROR("must pass in namespace");
 847                 LBUG();
 848         }
 849
 850         res = ldlm_resource_get(ns, NULL, *res_id, 0, 0);
 851         if (res == NULL) {
 852                 EXIT;
 853                 return;
 854         }
 855
 856         l_lock(&ns->ns_lock);
 857         ldlm_resource_foreach(res, iter, data);
 858         l_unlock(&ns->ns_lock);
 859         ldlm_resource_putref(res);
 860         EXIT;
 861 }
 862
 863 /* Lock replay */
 864
 865 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
 866 {
 867         struct list_head *list = closure;
 868
 869         /* we use l_pending_chain here, because it's unused on clients. */
 870         list_add(&lock->l_pending_chain, list);
 871         return LDLM_ITER_CONTINUE;
 872 }
 873
 874 static int replay_lock_interpret(struct ptlrpc_request *req,
 875                                     void * data, int rc)
 876 {
 877         struct ldlm_lock *lock;
 878         struct ldlm_reply *reply;
 879
 880         atomic_dec(&req->rq_import->imp_replay_inflight);
 881         if (rc != ELDLM_OK)
 882                 GOTO(out, rc);
 883
 884         lock = req->rq_async_args.pointer_arg[0];
 885         LASSERT(lock != NULL);
 886
 887         reply = lustre_swab_repbuf(req, 0, sizeof (*reply),
 888                                    lustre_swab_ldlm_reply);
 889         if (reply == NULL) {
 890                 CERROR("Can't unpack ldlm_reply\n");
 891                 GOTO (out, rc = -EPROTO);
 892         }
 893
 894         memcpy(&lock->l_remote_handle, &reply->lock_handle,
 895                sizeof(lock->l_remote_handle));
 896         LDLM_DEBUG(lock, "replayed lock:");
 897         ptlrpc_import_recovery_state_machine(req->rq_import);
 898  out:
 899         RETURN(rc);
 900 }
 901
 902 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
 903 {
 904         struct ptlrpc_request *req;
 905         struct ldlm_request *body;
 906         struct ldlm_reply *reply;
 907         int size;
 908         int flags;
 909
 910         /*
 911          * If granted mode matches the requested mode, this lock is granted.
 912          *
 913          * If they differ, but we have a granted mode, then we were granted
 914          * one mode and now want another: ergo, converting.
 915          *
 916          * If we haven't been granted anything and are on a resource list,
 917          * then we're blocked/waiting.
 918          *
 919          * If we haven't been granted anything and we're NOT on a resource list,
 920          * then we haven't got a reply yet and don't have a known disposition.
 921          * This happens whenever a lock enqueue is the request that triggers
 922          * recovery.
 923          */
 924         if (lock->l_granted_mode == lock->l_req_mode)
 925                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
 926         else if (lock->l_granted_mode)
 927                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
 928         else if (!list_empty(&lock->l_res_link))
 929                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
 930         else
 931                 flags = LDLM_FL_REPLAY;
 932
 933         size = sizeof(*body);
 934         req = ptlrpc_prep_req(imp, LDLM_ENQUEUE, 1, &size, NULL);
 935         if (!req)
 936                 RETURN(-ENOMEM);
 937
 938         /* We're part of recovery, so don't wait for it. */
 939         req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
 940
 941         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 942         ldlm_lock2desc(lock, &body->lock_desc);
 943         body->lock_flags = flags;
 944
 945         ldlm_lock2handle(lock, &body->lock_handle1);
 946         size = sizeof(*reply);
 947         req->rq_replen = lustre_msg_size(1, &size);
 948
 949         LDLM_DEBUG(lock, "replaying lock:");
 950
 951         atomic_inc(&req->rq_import->imp_replay_inflight);
 952         req->rq_async_args.pointer_arg[0] = lock;
 953         req->rq_interpret_reply = replay_lock_interpret;
 954         ptlrpcd_add_req(req);
 955
 956         RETURN(0);
 957 }
 958
 959 int ldlm_replay_locks(struct obd_import *imp)
 960 {
 961         struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
 962         struct list_head list, *pos, *next;
 963         struct ldlm_lock *lock;
 964         int rc = 0;
 965
 966         ENTRY;
 967         INIT_LIST_HEAD(&list);
 968
 969         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
 970
 971         /* ensure this doesn't fall to 0 before all have been queued */
 972         atomic_inc(&imp->imp_replay_inflight);
 973
 974         l_lock(&ns->ns_lock);
 975         (void)ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
 976
 977         list_for_each_safe(pos, next, &list) {
 978                 lock = list_entry(pos, struct ldlm_lock, l_pending_chain);
 979                 rc = replay_one_lock(imp, lock);
 980                 if (rc)
 981                         break; /* or try to do the rest? */
 982         }
 983         l_unlock(&ns->ns_lock);
 984
 985         atomic_dec(&imp->imp_replay_inflight);
 986
 987         RETURN(rc);
 988 }