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 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
  24 #include <linux/lustre_dlm.h>
  25 #include <linux/obd_class.h>
  26 #include <linux/obd.h>
  27
  28 static int interrupted_completion_wait(void *data)
  29 {
  30         RETURN(1);
  31 }
  32
  33 static int expired_completion_wait(void *data)
  34 {
  35         struct ldlm_lock *lock = data;
  36         struct ptlrpc_connection *conn;
  37         struct obd_device *obd;
  38
  39         if (!lock)
  40                 CERROR("NULL lock\n");
  41         else if (!lock->l_connh)
  42                 CERROR("lock %p has NULL connh\n", lock);
  43         else if (!(obd = class_conn2obd(lock->l_connh)))
  44                 CERROR("lock %p has NULL obd\n", lock);
  45         else if (!(conn = obd->u.cli.cl_import.imp_connection))
  46                 CERROR("lock %p has NULL connection\n", lock);
  47         else {
  48                 class_signal_connection_failure(conn);
  49         }
  50         RETURN(0);
  51 }
  52
  53 #if 0
  54 static int expired_completion_wait(void *data)
  55 {
  56         struct ldlm_lock *lock = data;
  57         struct ptlrpc_connection *conn =
  58                 class_conn2cliimp(lock->l_connh)->imp_connection;
  59
  60         if (!conn) {
  61                 CERROR("lock %p has NULL import connection\n", lock);
  62                 RETURN(1);
  63         }
  64
  65         class_signal_connection_failure(conn);
  66         RETURN(0);
  67 }
  68 #endif
  69
  70 int ldlm_completion_ast(struct ldlm_lock *lock, int flags)
  71 {
  72         struct l_wait_info lwi =
  73                 LWI_TIMEOUT_INTR(obd_timeout * HZ, expired_completion_wait,
  74                                  interrupted_completion_wait, lock);
  75         int rc = 0;
  76         ENTRY;
  77
  78         if (flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
  79                       LDLM_FL_BLOCK_CONV)) {
  80                 /* Go to sleep until the lock is granted. */
  81                 /* FIXME: or cancelled. */
  82                 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock,"
  83                            " sleeping");
  84                 ldlm_lock_dump(lock);
  85                 ldlm_reprocess_all(lock->l_resource);
  86                 rc = l_wait_event(lock->l_waitq,
  87                                   (lock->l_req_mode == lock->l_granted_mode),
  88                                   &lwi);
  89                 if (rc) {
  90                         LDLM_DEBUG(lock,
  91                                    "client-side enqueue waking up: failed (%d)",
  92                                    rc);
  93                 } else {
  94                         LDLM_DEBUG(lock,
  95                                    "client-side enqueue waking up: granted");
  96                 }
  97         } else if (flags == LDLM_FL_WAIT_NOREPROC) {
  98                 rc = l_wait_event(lock->l_waitq,
  99                                   (lock->l_req_mode == lock->l_granted_mode),
 100                                   &lwi);
 101         } else if (flags == 0) {
 102                 wake_up(&lock->l_waitq);
 103         }
 104
 105         RETURN(rc);
 106 }
 107
 108 static int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
 109                                   struct lustre_handle *parent_lockh,
 110                                   __u64 *res_id,
 111                                   __u32 type,
 112                                   void *cookie, int cookielen,
 113                                   ldlm_mode_t mode,
 114                                   int *flags,
 115                                   ldlm_completion_callback completion,
 116                                   ldlm_blocking_callback blocking,
 117                                   void *data,
 118                                   __u32 data_len,
 119                                   struct lustre_handle *lockh)
 120 {
 121         struct ldlm_lock *lock;
 122         int err;
 123
 124         if (ns->ns_client) {
 125                 CERROR("Trying to cancel local lock\n");
 126                 LBUG();
 127         }
 128
 129         lock = ldlm_lock_create(ns, parent_lockh, res_id, type, mode, data, data_len);
 130         if (!lock)
 131                 GOTO(out_nolock, err = -ENOMEM);
 132         LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
 133
 134         ldlm_lock_addref_internal(lock, mode);
 135         ldlm_lock2handle(lock, lockh);
 136         lock->l_connh = NULL;
 137
 138         err = ldlm_lock_enqueue(lock, cookie, cookielen, flags, completion,
 139                                 blocking);
 140         if (err != ELDLM_OK)
 141                 GOTO(out, err);
 142
 143         if (type == LDLM_EXTENT)
 144                 memcpy(cookie, &lock->l_extent, sizeof(lock->l_extent));
 145         if ((*flags) & LDLM_FL_LOCK_CHANGED)
 146                 memcpy(res_id, lock->l_resource->lr_name, sizeof(*res_id));
 147
 148         LDLM_DEBUG_NOLOCK("client-side local enqueue handler END (lock %p)",
 149                           lock);
 150
 151         if (lock->l_completion_ast)
 152                 lock->l_completion_ast(lock, *flags);
 153
 154         LDLM_DEBUG(lock, "client-side local enqueue END");
 155         EXIT;
 156  out:
 157         LDLM_LOCK_PUT(lock);
 158  out_nolock:
 159         return err;
 160 }
 161
 162 int ldlm_cli_enqueue(struct lustre_handle *connh,
 163                      struct ptlrpc_request *req,
 164                      struct ldlm_namespace *ns,
 165                      struct lustre_handle *parent_lock_handle,
 166                      __u64 *res_id,
 167                      __u32 type,
 168                      void *cookie, int cookielen,
 169                      ldlm_mode_t mode,
 170                      int *flags,
 171                      ldlm_completion_callback completion,
 172                      ldlm_blocking_callback blocking,
 173                      void *data,
 174                      __u32 data_len,
 175                      struct lustre_handle *lockh)
 176 {
 177         struct ldlm_lock *lock;
 178         struct ldlm_request *body;
 179         struct ldlm_reply *reply;
 180         int rc, size = sizeof(*body), req_passed_in = 1;
 181         ENTRY;
 182
 183         if (connh == NULL)
 184                 return ldlm_cli_enqueue_local(ns, parent_lock_handle, res_id,
 185                                               type, cookie, cookielen, mode,
 186                                               flags, completion, blocking, data,
 187                                               data_len, lockh);
 188
 189         *flags = 0;
 190         lock = ldlm_lock_create(ns, parent_lock_handle, res_id, type, mode,
 191                                 data, data_len);
 192         if (lock == NULL)
 193                 GOTO(out_nolock, rc = -ENOMEM);
 194         LDLM_DEBUG(lock, "client-side enqueue START");
 195         /* for the local lock, add the reference */
 196         ldlm_lock_addref_internal(lock, mode);
 197         ldlm_lock2handle(lock, lockh);
 198
 199         if (req == NULL) {
 200                 req = ptlrpc_prep_req(class_conn2cliimp(connh), LDLM_ENQUEUE, 1,
 201                                       &size, NULL);
 202                 if (!req)
 203                         GOTO(out, rc = -ENOMEM);
 204                 req_passed_in = 0;
 205         } else if (req->rq_reqmsg->buflens[0] != sizeof(*body))
 206                 LBUG();
 207
 208         /* Dump all of this data into the request buffer */
 209         body = lustre_msg_buf(req->rq_reqmsg, 0);
 210         ldlm_lock2desc(lock, &body->lock_desc);
 211         /* Phil: make this part of ldlm_lock2desc */
 212         if (type == LDLM_EXTENT)
 213                 memcpy(&body->lock_desc.l_extent, cookie,
 214                        sizeof(body->lock_desc.l_extent));
 215         body->lock_flags = *flags;
 216
 217         memcpy(&body->lock_handle1, lockh, sizeof(*lockh));
 218         if (parent_lock_handle)
 219                 memcpy(&body->lock_handle2, parent_lock_handle,
 220                        sizeof(body->lock_handle2));
 221
 222         /* Continue as normal. */
 223         if (!req_passed_in) {
 224                 size = sizeof(*reply);
 225                 req->rq_replen = lustre_msg_size(1, &size);
 226         }
 227         lock->l_connh = connh;
 228         lock->l_export = NULL;
 229
 230         rc = ptlrpc_queue_wait(req);
 231         /* FIXME: status check here? */
 232         rc = ptlrpc_check_status(req, rc);
 233
 234         if (rc != ELDLM_OK) {
 235                 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
 236                            rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
 237                 ldlm_lock_decref(lockh, mode);
 238                 /* FIXME: if we've already received a completion AST, this will
 239                  * LBUG! */
 240                 ldlm_lock_destroy(lock);
 241                 GOTO(out, rc);
 242         }
 243
 244         reply = lustre_msg_buf(req->rq_repmsg, 0);
 245         memcpy(&lock->l_remote_handle, &reply->lock_handle,
 246                sizeof(lock->l_remote_handle));
 247         if (type == LDLM_EXTENT)
 248                 memcpy(cookie, &reply->lock_extent, sizeof(reply->lock_extent));
 249         *flags = reply->lock_flags;
 250
 251         CDEBUG(D_INFO, "remote handle: %p, flags: %d\n",
 252                (void *)(unsigned long)reply->lock_handle.addr, *flags);
 253         CDEBUG(D_INFO, "extent: %Lu -> %Lu\n",
 254                (unsigned long long)reply->lock_extent.start,
 255                (unsigned long long)reply->lock_extent.end);
 256
 257         /* If enqueue returned a blocked lock but the completion handler has
 258          * already run, then it fixed up the resource and we don't need to do it
 259          * again. */
 260         if ((*flags) & LDLM_FL_LOCK_CHANGED) {
 261                 int newmode = reply->lock_mode;
 262                 if (newmode && newmode != lock->l_req_mode) {
 263                         LDLM_DEBUG(lock, "server returned different mode %s",
 264                                    ldlm_lockname[newmode]);
 265                         lock->l_req_mode = newmode;
 266                 }
 267
 268                 if (reply->lock_resource_name[0] !=
 269                     lock->l_resource->lr_name[0]) {
 270                         CDEBUG(D_INFO, "remote intent success, locking %ld "
 271                                "instead of %ld\n",
 272                                (long)reply->lock_resource_name[0],
 273                                (long)lock->l_resource->lr_name[0]);
 274
 275                         ldlm_lock_change_resource(lock,
 276                                                   reply->lock_resource_name);
 277                         if (lock->l_resource == NULL) {
 278                                 LBUG();
 279                                 RETURN(-ENOMEM);
 280                         }
 281                         LDLM_DEBUG(lock, "client-side enqueue, new resource");
 282                 }
 283         }
 284
 285         if (!req_passed_in)
 286                 ptlrpc_free_req(req);
 287
 288         rc = ldlm_lock_enqueue(lock, cookie, cookielen, flags, completion,
 289                                blocking);
 290         if (lock->l_completion_ast)
 291                 lock->l_completion_ast(lock, *flags);
 292
 293         LDLM_DEBUG(lock, "client-side enqueue END");
 294         EXIT;
 295  out:
 296         LDLM_LOCK_PUT(lock);
 297  out_nolock:
 298         return rc;
 299 }
 300
 301 int ldlm_match_or_enqueue(struct lustre_handle *connh,
 302                           struct ptlrpc_request *req,
 303                           struct ldlm_namespace *ns,
 304                           struct lustre_handle *parent_lock_handle,
 305                           __u64 *res_id,
 306                           __u32 type,
 307                           void *cookie, int cookielen,
 308                           ldlm_mode_t mode,
 309                           int *flags,
 310                           ldlm_completion_callback completion,
 311                           ldlm_blocking_callback blocking,
 312                           void *data,
 313                           __u32 data_len,
 314                           struct lustre_handle *lockh)
 315 {
 316         int rc;
 317         ENTRY;
 318         rc = ldlm_lock_match(ns, res_id, type, cookie, cookielen, mode, lockh);
 319         if (rc == 0) {
 320                 rc = ldlm_cli_enqueue(connh, req, ns,
 321                                       parent_lock_handle, res_id, type, cookie,
 322                                       cookielen, mode, flags, completion,
 323                                       blocking, data, data_len, lockh);
 324                 if (rc != ELDLM_OK)
 325                         CERROR("ldlm_cli_enqueue: err: %d\n", rc);
 326                 RETURN(rc);
 327         } else
 328                 RETURN(0);
 329 }
 330
 331 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
 332                                   int *flags)
 333 {
 334
 335         if (lock->l_resource->lr_namespace->ns_client) {
 336                 CERROR("Trying to cancel local lock\n");
 337                 LBUG();
 338         }
 339         LDLM_DEBUG(lock, "client-side local convert");
 340
 341         ldlm_lock_convert(lock, new_mode, flags);
 342         ldlm_reprocess_all(lock->l_resource);
 343
 344         LDLM_DEBUG(lock, "client-side local convert handler END");
 345         LDLM_LOCK_PUT(lock);
 346         RETURN(0);
 347 }
 348
 349 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
 350  * conversion of locks which are on the waiting or converting queue */
 351 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, int *flags)
 352 {
 353         struct ldlm_request *body;
 354         struct lustre_handle *connh;
 355         struct ldlm_reply *reply;
 356         struct ldlm_lock *lock;
 357         struct ldlm_resource *res;
 358         struct ptlrpc_request *req;
 359         int rc, size = sizeof(*body);
 360         ENTRY;
 361
 362         lock = ldlm_handle2lock(lockh);
 363         if (!lock) {
 364                 LBUG();
 365                 RETURN(-EINVAL);
 366         }
 367         *flags = 0;
 368         connh = lock->l_connh;
 369
 370         if (!connh)
 371                 return ldlm_cli_convert_local(lock, new_mode, flags);
 372
 373         LDLM_DEBUG(lock, "client-side convert");
 374
 375         req = ptlrpc_prep_req(class_conn2cliimp(connh), LDLM_CONVERT, 1, &size,
 376                               NULL);
 377         if (!req)
 378                 GOTO(out, rc = -ENOMEM);
 379
 380         body = lustre_msg_buf(req->rq_reqmsg, 0);
 381         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 382                sizeof(body->lock_handle1));
 383
 384         body->lock_desc.l_req_mode = new_mode;
 385         body->lock_flags = *flags;
 386
 387         size = sizeof(*reply);
 388         req->rq_replen = lustre_msg_size(1, &size);
 389
 390         rc = ptlrpc_queue_wait(req);
 391         rc = ptlrpc_check_status(req, rc);
 392         if (rc != ELDLM_OK)
 393                 GOTO(out, rc);
 394
 395         reply = lustre_msg_buf(req->rq_repmsg, 0);
 396         res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
 397         if (res != NULL)
 398                 ldlm_reprocess_all(res);
 399         /* Go to sleep until the lock is granted. */
 400         /* FIXME: or cancelled. */
 401         if (lock->l_completion_ast)
 402                 lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC);
 403         EXIT;
 404  out:
 405         LDLM_LOCK_PUT(lock);
 406         ptlrpc_free_req(req);
 407         return rc;
 408 }
 409
 410 int ldlm_cli_cancel(struct lustre_handle *lockh)
 411 {
 412         struct ptlrpc_request *req;
 413         struct ldlm_lock *lock;
 414         struct ldlm_request *body;
 415         int rc = 0, size = sizeof(*body);
 416         ENTRY;
 417
 418         lock = ldlm_handle2lock(lockh);
 419         if (!lock) {
 420                 /* It's possible that the decref that we did just before this
 421                  * cancel was the last reader/writer, and caused a cancel before
 422                  * we could call this function.  If we want to make this
 423                  * impossible (by adding a dec_and_cancel() or similar), then
 424                  * we can put the LBUG back. */
 425                 //LBUG();
 426                 RETURN(-EINVAL);
 427         }
 428
 429         if (lock->l_connh) {
 430                 LDLM_DEBUG(lock, "client-side cancel");
 431                 /* Set this flag to prevent others from getting new references*/
 432                 l_lock(&lock->l_resource->lr_namespace->ns_lock);
 433                 lock->l_flags |= LDLM_FL_CBPENDING;
 434                 ldlm_cancel_callback(lock);
 435                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 436
 437                 req = ptlrpc_prep_req(class_conn2cliimp(lock->l_connh),
 438                                       LDLM_CANCEL, 1, &size, NULL);
 439                 if (!req)
 440                         GOTO(out, rc = -ENOMEM);
 441
 442                 body = lustre_msg_buf(req->rq_reqmsg, 0);
 443                 memcpy(&body->lock_handle1, &lock->l_remote_handle,
 444                        sizeof(body->lock_handle1));
 445
 446                 req->rq_replen = lustre_msg_size(0, NULL);
 447
 448                 rc = ptlrpc_queue_wait(req);
 449                 rc = ptlrpc_check_status(req, rc);
 450                 ptlrpc_free_req(req);
 451                 if (rc != ELDLM_OK)
 452                         GOTO(out, rc);
 453
 454                 ldlm_lock_cancel(lock);
 455         } else {
 456                 LDLM_DEBUG(lock, "client-side local cancel");
 457                 if (lock->l_resource->lr_namespace->ns_client) {
 458                         CERROR("Trying to cancel local lock\n");
 459                         LBUG();
 460                 }
 461                 ldlm_lock_cancel(lock);
 462                 ldlm_reprocess_all(lock->l_resource);
 463                 LDLM_DEBUG(lock, "client-side local cancel handler END");
 464         }
 465
 466         EXIT;
 467  out:
 468         LDLM_LOCK_PUT(lock);
 469         return rc;
 470 }
 471
 472 /* Cancel all locks on a given resource that have 0 readers/writers.
 473  *
 474  * If 'local_only' is true, throw the locks away without trying to notify the
 475  * server. */
 476 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns, __u64 *res_id,
 477                            int local_only)
 478 {
 479         struct ldlm_resource *res;
 480         struct list_head *tmp, *next, list = LIST_HEAD_INIT(list);
 481         struct ldlm_ast_work *w;
 482         ENTRY;
 483
 484         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
 485         if (res == NULL)
 486                 RETURN(-EINVAL);
 487
 488         l_lock(&ns->ns_lock);
 489         list_for_each(tmp, &res->lr_granted) {
 490                 struct ldlm_lock *lock;
 491                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 492
 493                 if (lock->l_readers || lock->l_writers)
 494                         continue;
 495
 496                 /* Setting the CBPENDING flag is a little misleading, but
 497                  * prevents an important race; namely, once CBPENDING is set,
 498                  * the lock can accumulate no more readers/writers.  Since
 499                  * readers and writers are already zero here, ldlm_lock_decref
 500                  * won't see this flag and call l_blocking_ast */
 501                 lock->l_flags |= LDLM_FL_CBPENDING;
 502
 503                 OBD_ALLOC(w, sizeof(*w));
 504                 LASSERT(w);
 505
 506                 w->w_lock = LDLM_LOCK_GET(lock);
 507                 list_add(&w->w_list, &list);
 508         }
 509         l_unlock(&ns->ns_lock);
 510
 511         list_for_each_safe(tmp, next, &list) {
 512                 struct lustre_handle lockh;
 513                 int rc;
 514                 w = list_entry(tmp, struct ldlm_ast_work, w_list);
 515
 516                 if (local_only)
 517                         ldlm_lock_cancel(w->w_lock);
 518                 else {
 519                         ldlm_lock2handle(w->w_lock, &lockh);
 520                         rc = ldlm_cli_cancel(&lockh);
 521                         if (rc != ELDLM_OK)
 522                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 523                 }
 524                 LDLM_LOCK_PUT(w->w_lock);
 525                 list_del(&w->w_list);
 526                 OBD_FREE(w, sizeof(*w));
 527         }
 528
 529         ldlm_resource_put(res);
 530
 531         RETURN(0);
 532 }