lustre/ldlm/ldlm_request.c

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