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         ENTRY;
 124
 125         if (ns->ns_client) {
 126                 CERROR("Trying to enqueue local lock in a shadow namespace\n");
 127                 LBUG();
 128         }
 129
 130         lock = ldlm_lock_create(ns, parent_lockh, res_id, type, mode, data,
 131                                 data_len);
 132         if (!lock)
 133                 GOTO(out_nolock, err = -ENOMEM);
 134         LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
 135
 136         ldlm_lock_addref_internal(lock, mode);
 137         ldlm_lock2handle(lock, lockh);
 138         lock->l_connh = NULL;
 139
 140         err = ldlm_lock_enqueue(lock, cookie, cookielen, flags, completion,
 141                                 blocking);
 142         if (err != ELDLM_OK)
 143                 GOTO(out, err);
 144
 145         if (type == LDLM_EXTENT)
 146                 memcpy(cookie, &lock->l_extent, sizeof(lock->l_extent));
 147         if ((*flags) & LDLM_FL_LOCK_CHANGED)
 148                 memcpy(res_id, lock->l_resource->lr_name, sizeof(*res_id));
 149
 150         LDLM_DEBUG_NOLOCK("client-side local enqueue handler END (lock %p)",
 151                           lock);
 152
 153         if (lock->l_completion_ast)
 154                 lock->l_completion_ast(lock, *flags);
 155
 156         LDLM_DEBUG(lock, "client-side local enqueue END");
 157         EXIT;
 158  out:
 159         LDLM_LOCK_PUT(lock);
 160  out_nolock:
 161         return err;
 162 }
 163
 164 int ldlm_cli_enqueue(struct lustre_handle *connh,
 165                      struct ptlrpc_request *req,
 166                      struct ldlm_namespace *ns,
 167                      struct lustre_handle *parent_lock_handle,
 168                      __u64 *res_id,
 169                      __u32 type,
 170                      void *cookie, int cookielen,
 171                      ldlm_mode_t mode,
 172                      int *flags,
 173                      ldlm_completion_callback completion,
 174                      ldlm_blocking_callback blocking,
 175                      void *data,
 176                      __u32 data_len,
 177                      struct lustre_handle *lockh)
 178 {
 179         struct ldlm_lock *lock;
 180         struct ldlm_request *body;
 181         struct ldlm_reply *reply;
 182         int rc, size = sizeof(*body), req_passed_in = 1;
 183         ENTRY;
 184
 185         if (connh == NULL)
 186                 return ldlm_cli_enqueue_local(ns, parent_lock_handle, res_id,
 187                                               type, cookie, cookielen, mode,
 188                                               flags, completion, blocking, data,
 189                                               data_len, lockh);
 190
 191         *flags = 0;
 192         lock = ldlm_lock_create(ns, parent_lock_handle, res_id, type, mode,
 193                                 data, data_len);
 194         if (lock == NULL)
 195                 GOTO(out_nolock, rc = -ENOMEM);
 196         LDLM_DEBUG(lock, "client-side enqueue START");
 197         /* for the local lock, add the reference */
 198         ldlm_lock_addref_internal(lock, mode);
 199         ldlm_lock2handle(lock, lockh);
 200
 201         if (req == NULL) {
 202                 req = ptlrpc_prep_req(class_conn2cliimp(connh), LDLM_ENQUEUE, 1,
 203                                       &size, NULL);
 204                 if (!req)
 205                         GOTO(out, rc = -ENOMEM);
 206                 req_passed_in = 0;
 207         } else if (req->rq_reqmsg->buflens[0] != sizeof(*body))
 208                 LBUG();
 209
 210         /* Dump all of this data into the request buffer */
 211         body = lustre_msg_buf(req->rq_reqmsg, 0);
 212         ldlm_lock2desc(lock, &body->lock_desc);
 213         /* Phil: make this part of ldlm_lock2desc */
 214         if (type == LDLM_EXTENT)
 215                 memcpy(&body->lock_desc.l_extent, cookie,
 216                        sizeof(body->lock_desc.l_extent));
 217         body->lock_flags = *flags;
 218
 219         memcpy(&body->lock_handle1, lockh, sizeof(*lockh));
 220         if (parent_lock_handle)
 221                 memcpy(&body->lock_handle2, parent_lock_handle,
 222                        sizeof(body->lock_handle2));
 223
 224         /* Continue as normal. */
 225         if (!req_passed_in) {
 226                 size = sizeof(*reply);
 227                 req->rq_replen = lustre_msg_size(1, &size);
 228         }
 229         lock->l_connh = connh;
 230         lock->l_export = NULL;
 231
 232         rc = ptlrpc_queue_wait(req);
 233         /* FIXME: status check here? */
 234         rc = ptlrpc_check_status(req, rc);
 235
 236         if (rc != ELDLM_OK) {
 237                 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
 238                            rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
 239                 ldlm_lock_decref(lockh, mode);
 240                 /* FIXME: if we've already received a completion AST, this will
 241                  * LBUG! */
 242                 ldlm_lock_destroy(lock);
 243                 GOTO(out, rc);
 244         }
 245
 246         reply = lustre_msg_buf(req->rq_repmsg, 0);
 247         memcpy(&lock->l_remote_handle, &reply->lock_handle,
 248                sizeof(lock->l_remote_handle));
 249         if (type == LDLM_EXTENT)
 250                 memcpy(cookie, &reply->lock_extent, sizeof(reply->lock_extent));
 251         *flags = reply->lock_flags;
 252
 253         CDEBUG(D_INFO, "remote handle: %p, flags: %d\n",
 254                (void *)(unsigned long)reply->lock_handle.addr, *flags);
 255         CDEBUG(D_INFO, "requested extent: "LPU64" -> "LPU64", got extent "
 256                LPU64" -> "LPU64"\n",
 257                body->lock_desc.l_extent.start, body->lock_desc.l_extent.end,
 258                reply->lock_extent.start, reply->lock_extent.end);
 259
 260         /* If enqueue returned a blocked lock but the completion handler has
 261          * already run, then it fixed up the resource and we don't need to do it
 262          * again. */
 263         if ((*flags) & LDLM_FL_LOCK_CHANGED) {
 264                 int newmode = reply->lock_mode;
 265                 if (newmode && newmode != lock->l_req_mode) {
 266                         LDLM_DEBUG(lock, "server returned different mode %s",
 267                                    ldlm_lockname[newmode]);
 268                         lock->l_req_mode = newmode;
 269                 }
 270
 271                 if (reply->lock_resource_name[0] !=
 272                     lock->l_resource->lr_name[0]) {
 273                         CDEBUG(D_INFO, "remote intent success, locking %ld "
 274                                "instead of %ld\n",
 275                                (long)reply->lock_resource_name[0],
 276                                (long)lock->l_resource->lr_name[0]);
 277
 278                         ldlm_lock_change_resource(lock,
 279                                                   reply->lock_resource_name);
 280                         if (lock->l_resource == NULL) {
 281                                 LBUG();
 282                                 RETURN(-ENOMEM);
 283                         }
 284                         LDLM_DEBUG(lock, "client-side enqueue, new resource");
 285                 }
 286         }
 287
 288         if (!req_passed_in)
 289                 ptlrpc_free_req(req);
 290
 291         rc = ldlm_lock_enqueue(lock, cookie, cookielen, flags, completion,
 292                                blocking);
 293         if (lock->l_completion_ast)
 294                 lock->l_completion_ast(lock, *flags);
 295
 296         LDLM_DEBUG(lock, "client-side enqueue END");
 297         EXIT;
 298  out:
 299         LDLM_LOCK_PUT(lock);
 300  out_nolock:
 301         return rc;
 302 }
 303
 304 int ldlm_match_or_enqueue(struct lustre_handle *connh,
 305                           struct ptlrpc_request *req,
 306                           struct ldlm_namespace *ns,
 307                           struct lustre_handle *parent_lock_handle,
 308                           __u64 *res_id,
 309                           __u32 type,
 310                           void *cookie, int cookielen,
 311                           ldlm_mode_t mode,
 312                           int *flags,
 313                           ldlm_completion_callback completion,
 314                           ldlm_blocking_callback blocking,
 315                           void *data,
 316                           __u32 data_len,
 317                           struct lustre_handle *lockh)
 318 {
 319         int rc;
 320         ENTRY;
 321         rc = ldlm_lock_match(ns, res_id, type, cookie, cookielen, mode, lockh);
 322         if (rc == 0) {
 323                 rc = ldlm_cli_enqueue(connh, req, ns,
 324                                       parent_lock_handle, res_id, type, cookie,
 325                                       cookielen, mode, flags, completion,
 326                                       blocking, data, data_len, lockh);
 327                 if (rc != ELDLM_OK)
 328                         CERROR("ldlm_cli_enqueue: err: %d\n", rc);
 329                 RETURN(rc);
 330         } else
 331                 RETURN(0);
 332 }
 333
 334 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
 335                                   int *flags)
 336 {
 337
 338         if (lock->l_resource->lr_namespace->ns_client) {
 339                 CERROR("Trying to cancel local lock\n");
 340                 LBUG();
 341         }
 342         LDLM_DEBUG(lock, "client-side local convert");
 343
 344         ldlm_lock_convert(lock, new_mode, flags);
 345         ldlm_reprocess_all(lock->l_resource);
 346
 347         LDLM_DEBUG(lock, "client-side local convert handler END");
 348         LDLM_LOCK_PUT(lock);
 349         RETURN(0);
 350 }
 351
 352 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
 353  * conversion of locks which are on the waiting or converting queue */
 354 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, int *flags)
 355 {
 356         struct ldlm_request *body;
 357         struct lustre_handle *connh;
 358         struct ldlm_reply *reply;
 359         struct ldlm_lock *lock;
 360         struct ldlm_resource *res;
 361         struct ptlrpc_request *req;
 362         int rc, size = sizeof(*body);
 363         ENTRY;
 364
 365         lock = ldlm_handle2lock(lockh);
 366         if (!lock) {
 367                 LBUG();
 368                 RETURN(-EINVAL);
 369         }
 370         *flags = 0;
 371         connh = lock->l_connh;
 372
 373         if (!connh)
 374                 return ldlm_cli_convert_local(lock, new_mode, flags);
 375
 376         LDLM_DEBUG(lock, "client-side convert");
 377
 378         req = ptlrpc_prep_req(class_conn2cliimp(connh), LDLM_CONVERT, 1, &size,
 379                               NULL);
 380         if (!req)
 381                 GOTO(out, rc = -ENOMEM);
 382
 383         body = lustre_msg_buf(req->rq_reqmsg, 0);
 384         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 385                sizeof(body->lock_handle1));
 386
 387         body->lock_desc.l_req_mode = new_mode;
 388         body->lock_flags = *flags;
 389
 390         size = sizeof(*reply);
 391         req->rq_replen = lustre_msg_size(1, &size);
 392
 393         rc = ptlrpc_queue_wait(req);
 394         rc = ptlrpc_check_status(req, rc);
 395         if (rc != ELDLM_OK)
 396                 GOTO(out, rc);
 397
 398         reply = lustre_msg_buf(req->rq_repmsg, 0);
 399         res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
 400         if (res != NULL)
 401                 ldlm_reprocess_all(res);
 402         /* Go to sleep until the lock is granted. */
 403         /* FIXME: or cancelled. */
 404         if (lock->l_completion_ast)
 405                 lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC);
 406         EXIT;
 407  out:
 408         LDLM_LOCK_PUT(lock);
 409         ptlrpc_free_req(req);
 410         return rc;
 411 }
 412
 413 int ldlm_cli_cancel(struct lustre_handle *lockh)
 414 {
 415         struct ptlrpc_request *req;
 416         struct ldlm_lock *lock;
 417         struct ldlm_request *body;
 418         int rc = 0, size = sizeof(*body);
 419         ENTRY;
 420
 421         lock = ldlm_handle2lock(lockh);
 422         if (!lock) {
 423                 /* It's possible that the decref that we did just before this
 424                  * cancel was the last reader/writer, and caused a cancel before
 425                  * we could call this function.  If we want to make this
 426                  * impossible (by adding a dec_and_cancel() or similar), then
 427                  * we can put the LBUG back. */
 428                 //LBUG();
 429                 RETURN(-EINVAL);
 430         }
 431
 432         if (lock->l_connh) {
 433                 LDLM_DEBUG(lock, "client-side cancel");
 434                 /* Set this flag to prevent others from getting new references*/
 435                 l_lock(&lock->l_resource->lr_namespace->ns_lock);
 436                 lock->l_flags |= LDLM_FL_CBPENDING;
 437                 ldlm_cancel_callback(lock);
 438                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 439
 440                 req = ptlrpc_prep_req(class_conn2cliimp(lock->l_connh),
 441                                       LDLM_CANCEL, 1, &size, NULL);
 442                 if (!req)
 443                         GOTO(out, rc = -ENOMEM);
 444
 445                 body = lustre_msg_buf(req->rq_reqmsg, 0);
 446                 memcpy(&body->lock_handle1, &lock->l_remote_handle,
 447                        sizeof(body->lock_handle1));
 448
 449                 req->rq_replen = lustre_msg_size(0, NULL);
 450
 451                 rc = ptlrpc_queue_wait(req);
 452                 rc = ptlrpc_check_status(req, rc);
 453                 ptlrpc_free_req(req);
 454                 if (rc != ELDLM_OK)
 455                         GOTO(out, rc);
 456
 457                 ldlm_lock_cancel(lock);
 458         } else {
 459                 LDLM_DEBUG(lock, "client-side local cancel");
 460                 if (lock->l_resource->lr_namespace->ns_client) {
 461                         CERROR("Trying to cancel local lock\n");
 462                         LBUG();
 463                 }
 464                 ldlm_lock_cancel(lock);
 465                 ldlm_reprocess_all(lock->l_resource);
 466                 LDLM_DEBUG(lock, "client-side local cancel handler END");
 467         }
 468
 469         EXIT;
 470  out:
 471         LDLM_LOCK_PUT(lock);
 472         return rc;
 473 }
 474
 475 /* Cancel all locks on a given resource that have 0 readers/writers.
 476  *
 477  * If 'local_only' is true, throw the locks away without trying to notify the
 478  * server. */
 479 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns, __u64 *res_id,
 480                            int local_only)
 481 {
 482         struct ldlm_resource *res;
 483         struct list_head *tmp, *next, list = LIST_HEAD_INIT(list);
 484         struct ldlm_ast_work *w;
 485         ENTRY;
 486
 487         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
 488         if (res == NULL) {
 489                 /* This is not a problem. */
 490                 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id[0]);
 491                 RETURN(0);
 492         }
 493
 494         l_lock(&ns->ns_lock);
 495         list_for_each(tmp, &res->lr_granted) {
 496                 struct ldlm_lock *lock;
 497                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 498
 499                 if (lock->l_readers || lock->l_writers)
 500                         continue;
 501
 502                 /* Setting the CBPENDING flag is a little misleading, but
 503                  * prevents an important race; namely, once CBPENDING is set,
 504                  * the lock can accumulate no more readers/writers.  Since
 505                  * readers and writers are already zero here, ldlm_lock_decref
 506                  * won't see this flag and call l_blocking_ast */
 507                 lock->l_flags |= LDLM_FL_CBPENDING;
 508
 509                 OBD_ALLOC(w, sizeof(*w));
 510                 LASSERT(w);
 511
 512                 w->w_lock = LDLM_LOCK_GET(lock);
 513                 list_add(&w->w_list, &list);
 514         }
 515         l_unlock(&ns->ns_lock);
 516
 517         list_for_each_safe(tmp, next, &list) {
 518                 struct lustre_handle lockh;
 519                 int rc;
 520                 w = list_entry(tmp, struct ldlm_ast_work, w_list);
 521
 522                 if (local_only)
 523                         ldlm_lock_cancel(w->w_lock);
 524                 else {
 525                         ldlm_lock2handle(w->w_lock, &lockh);
 526                         rc = ldlm_cli_cancel(&lockh);
 527                         if (rc != ELDLM_OK)
 528                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 529                 }
 530                 LDLM_LOCK_PUT(w->w_lock);
 531                 list_del(&w->w_list);
 532                 OBD_FREE(w, sizeof(*w));
 533         }
 534
 535         ldlm_resource_put(res);
 536
 537         RETURN(0);
 538 }