lustre/obdclass/genops.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (c) 2001-2003 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  * These are the only exported functions, they provide some generic
  22  * infrastructure for managing object devices
  23  */
  24
  25 #define DEBUG_SUBSYSTEM S_CLASS
  26 #ifdef __KERNEL__
  27 #include <linux/kmod.h>   /* for request_module() */
  28 #include <linux/module.h>
  29 #include <linux/obd_class.h>
  30 #include <linux/random.h>
  31 #include <linux/slab.h>
  32 #include <linux/pagemap.h>
  33 #else
  34 #include <liblustre.h>
  35 #include <linux/obd_class.h>
  36 #include <linux/obd.h>
  37 #endif
  38 #include <linux/lprocfs_status.h>
  39
  40 extern struct list_head obd_types;
  41 static spinlock_t obd_types_lock = SPIN_LOCK_UNLOCKED;
  42 kmem_cache_t *obdo_cachep = NULL;
  43 kmem_cache_t *import_cachep = NULL;
  44
  45 int (*ptlrpc_put_connection_superhack)(struct ptlrpc_connection *c);
  46 void (*ptlrpc_abort_inflight_superhack)(struct obd_import *imp);
  47
  48 /*
  49  * support functions: we could use inter-module communication, but this
  50  * is more portable to other OS's
  51  */
  52 static struct obd_type *class_search_type(char *name)
  53 {
  54         struct list_head *tmp;
  55         struct obd_type *type;
  56
  57         spin_lock(&obd_types_lock);
  58         list_for_each(tmp, &obd_types) {
  59                 type = list_entry(tmp, struct obd_type, typ_chain);
  60                 if (strlen(type->typ_name) == strlen(name) &&
  61                     strcmp(type->typ_name, name) == 0) {
  62                         spin_unlock(&obd_types_lock);
  63                         return type;
  64                 }
  65         }
  66         spin_unlock(&obd_types_lock);
  67         return NULL;
  68 }
  69
  70 struct obd_type *class_get_type(char *name)
  71 {
  72         struct obd_type *type = class_search_type(name);
  73
  74 #ifdef CONFIG_KMOD
  75         if (!type) {
  76                 if (!request_module(name)) {
  77                         CDEBUG(D_INFO, "Loaded module '%s'\n", name);
  78                         type = class_search_type(name);
  79                 } else
  80                         CDEBUG(D_INFO, "Can't load module '%s'\n", name);
  81         }
  82 #endif
  83         if (type)
  84                 try_module_get(type->typ_ops->o_owner);
  85         return type;
  86 }
  87
  88 void class_put_type(struct obd_type *type)
  89 {
  90         LASSERT(type);
  91         module_put(type->typ_ops->o_owner);
  92 }
  93
  94 int class_register_type(struct obd_ops *ops, struct md_ops *md_ops,
  95                         struct lprocfs_vars *vars, char *name)
  96 {
  97         struct obd_type *type;
  98         int rc = 0;
  99         ENTRY;
 100
 101         LASSERT(strnlen(name, 1024) < 1024);    /* sanity check */
 102
 103         if (class_search_type(name)) {
 104                 CDEBUG(D_IOCTL, "Type %s already registered\n", name);
 105                 RETURN(-EEXIST);
 106         }
 107
 108         rc = -ENOMEM;
 109         OBD_ALLOC(type, sizeof(*type));
 110         if (type == NULL)
 111                 RETURN(rc);
 112
 113         OBD_ALLOC(type->typ_ops, sizeof(*type->typ_ops));
 114         OBD_ALLOC(type->typ_name, strlen(name) + 1);
 115         if (md_ops)
 116                 OBD_ALLOC(type->typ_md_ops, sizeof(*type->typ_md_ops));
 117         if (type->typ_ops == NULL || type->typ_name == NULL ||
 118                         (md_ops && type->typ_md_ops == NULL))
 119                 GOTO (failed, rc);
 120
 121         *(type->typ_ops) = *ops;
 122         if (md_ops)
 123                 *(type->typ_md_ops) = *md_ops;
 124         else
 125                 type->typ_md_ops = NULL;
 126         strcpy(type->typ_name, name);
 127
 128 #ifdef LPROCFS
 129         type->typ_procroot = lprocfs_register(type->typ_name, proc_lustre_root,
 130                                               vars, type);
 131 #endif
 132         if (IS_ERR(type->typ_procroot)) {
 133                 rc = PTR_ERR(type->typ_procroot);
 134                 type->typ_procroot = NULL;
 135                 GOTO (failed, rc);
 136         }
 137
 138         spin_lock(&obd_types_lock);
 139         list_add(&type->typ_chain, &obd_types);
 140         spin_unlock(&obd_types_lock);
 141
 142         RETURN (0);
 143
 144  failed:
 145         if (type->typ_name != NULL)
 146                 OBD_FREE(type->typ_name, strlen(name) + 1);
 147         if (type->typ_ops != NULL)
 148                 OBD_FREE (type->typ_ops, sizeof (*type->typ_ops));
 149         if (type->typ_md_ops != NULL)
 150                 OBD_FREE (type->typ_md_ops, sizeof (*type->typ_md_ops));
 151         OBD_FREE(type, sizeof(*type));
 152         RETURN(rc);
 153 }
 154
 155 int class_unregister_type(char *name)
 156 {
 157         struct obd_type *type = class_search_type(name);
 158         ENTRY;
 159
 160         if (!type) {
 161                 CERROR("unknown obd type\n");
 162                 RETURN(-EINVAL);
 163         }
 164
 165         if (type->typ_refcnt) {
 166                 CERROR("type %s has refcount (%d)\n", name, type->typ_refcnt);
 167                 /* This is a bad situation, let's make the best of it */
 168                 /* Remove ops, but leave the name for debugging */
 169                 OBD_FREE(type->typ_ops, sizeof(*type->typ_ops));
 170                 RETURN(-EBUSY);
 171         }
 172
 173         if (type->typ_procroot) {
 174                 lprocfs_remove(type->typ_procroot);
 175                 type->typ_procroot = NULL;
 176         }
 177
 178         spin_lock(&obd_types_lock);
 179         list_del(&type->typ_chain);
 180         spin_unlock(&obd_types_lock);
 181         OBD_FREE(type->typ_name, strlen(name) + 1);
 182         if (type->typ_ops != NULL)
 183                 OBD_FREE(type->typ_ops, sizeof(*type->typ_ops));
 184         if (type->typ_md_ops != NULL)
 185                 OBD_FREE (type->typ_md_ops, sizeof (*type->typ_md_ops));
 186         OBD_FREE(type, sizeof(*type));
 187         RETURN(0);
 188 } /* class_unregister_type */
 189
 190 struct obd_device *class_newdev(int *dev)
 191 {
 192         struct obd_device *result = NULL;
 193         int i;
 194
 195         for (i = 0 ; i < MAX_OBD_DEVICES ; i++) {
 196                 struct obd_device *obd = &obd_dev[i];
 197                 if (!obd->obd_type) {
 198                         result = obd;
 199                         if (dev)
 200                                 *dev = i;
 201                         break;
 202                 }
 203         }
 204         return result;
 205 }
 206
 207 int class_name2dev(char *name)
 208 {
 209         int i;
 210
 211         if (!name)
 212                 return -1;
 213
 214         for (i = 0; i < MAX_OBD_DEVICES; i++) {
 215                 struct obd_device *obd = &obd_dev[i];
 216                 if (obd->obd_name && strcmp(name, obd->obd_name) == 0)
 217                         return i;
 218         }
 219
 220         return -1;
 221 }
 222
 223 struct obd_device *class_name2obd(char *name)
 224 {
 225         int dev = class_name2dev(name);
 226         if (dev < 0)
 227                 return NULL;
 228         return &obd_dev[dev];
 229 }
 230
 231 int class_uuid2dev(struct obd_uuid *uuid)
 232 {
 233         int i;
 234
 235         for (i = 0; i < MAX_OBD_DEVICES; i++) {
 236                 struct obd_device *obd = &obd_dev[i];
 237                 if (obd_uuid_equals(uuid, &obd->obd_uuid))
 238                         return i;
 239         }
 240
 241         return -1;
 242 }
 243
 244 struct obd_device *class_uuid2obd(struct obd_uuid *uuid)
 245 {
 246         int dev = class_uuid2dev(uuid);
 247         if (dev < 0)
 248                 return NULL;
 249         return &obd_dev[dev];
 250 }
 251
 252 /* Search for a client OBD connected to tgt_uuid.  If grp_uuid is
 253    specified, then only the client with that uuid is returned,
 254    otherwise any client connected to the tgt is returned. */
 255 struct obd_device * class_find_client_obd(struct obd_uuid *tgt_uuid,
 256                                           char * typ_name,
 257                                           struct obd_uuid *grp_uuid)
 258 {
 259         int i;
 260
 261         for (i = 0; i < MAX_OBD_DEVICES; i++) {
 262                 struct obd_device *obd = &obd_dev[i];
 263                 if (obd->obd_type == NULL)
 264                         continue;
 265                 if ((strncmp(obd->obd_type->typ_name, typ_name,
 266                              strlen(typ_name)) == 0)) {
 267                         struct client_obd *cli = &obd->u.cli;
 268                         struct obd_import *imp = cli->cl_import;
 269                         if (obd_uuid_equals(tgt_uuid, &imp->imp_target_uuid) &&
 270                             ((grp_uuid)? obd_uuid_equals(grp_uuid,
 271                                                          &obd->obd_uuid) : 1)) {
 272                                 return obd;
 273                         }
 274                 }
 275         }
 276
 277         return NULL;
 278 }
 279
 280 /* Iterate the obd_device list looking devices have grp_uuid. Start
 281    searching at *next, and if a device is found, the next index to look
 282    it is saved in *next. If next is NULL, then the first matching device
 283    will always be returned. */
 284 struct obd_device * class_devices_in_group(struct obd_uuid *grp_uuid, int *next)
 285 {
 286         int i;
 287         if (next == NULL)
 288                 i = 0;
 289         else if (*next >= 0 && *next < MAX_OBD_DEVICES)
 290                 i = *next;
 291         else
 292                 return NULL;
 293
 294         for (; i < MAX_OBD_DEVICES; i++) {
 295                 struct obd_device *obd = &obd_dev[i];
 296                 if (obd->obd_type == NULL)
 297                         continue;
 298                 if (obd_uuid_equals(grp_uuid, &obd->obd_uuid)) {
 299                         if (next != NULL)
 300                                 *next = i+1;
 301                         return obd;
 302                 }
 303         }
 304
 305         return NULL;
 306 }
 307
 308
 309 void obd_cleanup_caches(void)
 310 {
 311         int rc;
 312         ENTRY;
 313         if (obdo_cachep) {
 314                 rc = kmem_cache_destroy(obdo_cachep);
 315                 if (rc)
 316                         CERROR("Cannot destory ll_obdo_cache\n");
 317                 obdo_cachep = NULL;
 318         }
 319         if (import_cachep) {
 320                 rc = kmem_cache_destroy(import_cachep);
 321                 if (rc)
 322                         CERROR("Cannot destory ll_import_cache\n");
 323                 import_cachep = NULL;
 324         }
 325         EXIT;
 326 }
 327
 328 int obd_init_caches(void)
 329 {
 330         ENTRY;
 331         LASSERT(obdo_cachep == NULL);
 332         obdo_cachep = kmem_cache_create("ll_obdo_cache", sizeof(struct obdo),
 333                                         0, 0, NULL, NULL);
 334         if (!obdo_cachep)
 335                 GOTO(out, -ENOMEM);
 336
 337         LASSERT(import_cachep == NULL);
 338         import_cachep = kmem_cache_create("ll_import_cache",
 339                                           sizeof(struct obd_import),
 340                                           0, 0, NULL, NULL);
 341         if (!import_cachep)
 342                 GOTO(out, -ENOMEM);
 343
 344         RETURN(0);
 345  out:
 346         obd_cleanup_caches();
 347         RETURN(-ENOMEM);
 348
 349 }
 350
 351 /* map connection to client */
 352 struct obd_export *class_conn2export(struct lustre_handle *conn)
 353 {
 354         struct obd_export *export;
 355         ENTRY;
 356
 357         if (!conn) {
 358                 CDEBUG(D_CACHE, "looking for null handle\n");
 359                 RETURN(NULL);
 360         }
 361
 362         if (conn->cookie == -1) {  /* this means assign a new connection */
 363                 CDEBUG(D_CACHE, "want a new connection\n");
 364                 RETURN(NULL);
 365         }
 366
 367         CDEBUG(D_IOCTL, "looking for export cookie "LPX64"\n", conn->cookie);
 368         export = class_handle2object(conn->cookie);
 369         RETURN(export);
 370 }
 371
 372 struct obd_device *class_exp2obd(struct obd_export *exp)
 373 {
 374         if (exp)
 375                 return exp->exp_obd;
 376         return NULL;
 377 }
 378
 379 struct obd_device *class_conn2obd(struct lustre_handle *conn)
 380 {
 381         struct obd_export *export;
 382         export = class_conn2export(conn);
 383         if (export) {
 384                 struct obd_device *obd = export->exp_obd;
 385                 class_export_put(export);
 386                 return obd;
 387         }
 388         return NULL;
 389 }
 390
 391 struct obd_import *class_exp2cliimp(struct obd_export *exp)
 392 {
 393         struct obd_device *obd = exp->exp_obd;
 394         if (obd == NULL)
 395                 return NULL;
 396         return obd->u.cli.cl_import;
 397 }
 398
 399 struct obd_import *class_conn2cliimp(struct lustre_handle *conn)
 400 {
 401         struct obd_device *obd = class_conn2obd(conn);
 402         if (obd == NULL)
 403                 return NULL;
 404         return obd->u.cli.cl_import;
 405 }
 406
 407 /* Export management functions */
 408 static void export_handle_addref(void *export)
 409 {
 410         class_export_get(export);
 411 }
 412
 413 void __class_export_put(struct obd_export *exp)
 414 {
 415         if (atomic_dec_and_test(&exp->exp_refcount)) {
 416                 struct obd_device *obd = exp->exp_obd;
 417                 CDEBUG(D_IOCTL, "destroying export %p/%s\n", exp,
 418                        exp->exp_client_uuid.uuid);
 419
 420                 LASSERT(obd != NULL);
 421
 422                 /* "Local" exports (lctl, LOV->{mdc,osc}) have no connection. */
 423                 if (exp->exp_connection)
 424                         ptlrpc_put_connection_superhack(exp->exp_connection);
 425
 426                 LASSERT(list_empty(&exp->exp_outstanding_replies));
 427                 LASSERT(list_empty(&exp->exp_handle.h_link));
 428                 obd_destroy_export(exp);
 429
 430                 OBD_FREE(exp, sizeof(*exp));
 431                 if (obd->obd_set_up) {
 432                         atomic_dec(&obd->obd_refcount);
 433                         wake_up(&obd->obd_refcount_waitq);
 434                 }
 435         }
 436 }
 437
 438 /* Creates a new export, adds it to the hash table, and returns a
 439  * pointer to it. The refcount is 2: one for the hash reference, and
 440  * one for the pointer returned by this function. */
 441 struct obd_export *class_new_export(struct obd_device *obd)
 442 {
 443         struct obd_export *export;
 444
 445         OBD_ALLOC(export, sizeof(*export));
 446         if (!export) {
 447                 CERROR("no memory! (minor %d)\n", obd->obd_minor);
 448                 return NULL;
 449         }
 450
 451         export->exp_conn_cnt = 0;
 452         atomic_set(&export->exp_refcount, 2);
 453         export->exp_obd = obd;
 454         INIT_LIST_HEAD(&export->exp_outstanding_replies);
 455         /* XXX this should be in LDLM init */
 456         INIT_LIST_HEAD(&export->exp_ldlm_data.led_held_locks);
 457
 458         INIT_LIST_HEAD(&export->exp_handle.h_link);
 459         class_handle_hash(&export->exp_handle, export_handle_addref);
 460         spin_lock_init(&export->exp_lock);
 461
 462         spin_lock(&obd->obd_dev_lock);
 463         LASSERT(!obd->obd_stopping); /* shouldn't happen, but might race */
 464         atomic_inc(&obd->obd_refcount);
 465         list_add(&export->exp_obd_chain, &export->exp_obd->obd_exports);
 466         export->exp_obd->obd_num_exports++;
 467         spin_unlock(&obd->obd_dev_lock);
 468         obd_init_export(export);
 469         return export;
 470 }
 471
 472 void class_unlink_export(struct obd_export *exp)
 473 {
 474         class_handle_unhash(&exp->exp_handle);
 475
 476         spin_lock(&exp->exp_obd->obd_dev_lock);
 477         list_del_init(&exp->exp_obd_chain);
 478         exp->exp_obd->obd_num_exports--;
 479         spin_unlock(&exp->exp_obd->obd_dev_lock);
 480
 481         class_export_put(exp);
 482 }
 483
 484 /* Import management functions */
 485 static void import_handle_addref(void *import)
 486 {
 487         class_import_get(import);
 488 }
 489
 490 struct obd_import *class_import_get(struct obd_import *import)
 491 {
 492         atomic_inc(&import->imp_refcount);
 493         CDEBUG(D_IOCTL, "import %p refcount=%d\n", import,
 494                atomic_read(&import->imp_refcount));
 495         return import;
 496 }
 497
 498 void class_import_put(struct obd_import *import)
 499 {
 500         ENTRY;
 501
 502         CDEBUG(D_IOCTL, "import %p refcount=%d\n", import,
 503                atomic_read(&import->imp_refcount) - 1);
 504
 505         LASSERT(atomic_read(&import->imp_refcount) > 0);
 506         LASSERT(atomic_read(&import->imp_refcount) < 0x5a5a5a);
 507         if (!atomic_dec_and_test(&import->imp_refcount)) {
 508                 EXIT;
 509                 return;
 510         }
 511
 512         CDEBUG(D_IOCTL, "destroying import %p\n", import);
 513
 514         ptlrpc_put_connection_superhack(import->imp_connection);
 515
 516         LASSERT(list_empty(&import->imp_handle.h_link));
 517         OBD_FREE(import, sizeof(*import));
 518         EXIT;
 519 }
 520
 521 struct obd_import *class_new_import(void)
 522 {
 523         struct obd_import *imp;
 524
 525         OBD_ALLOC(imp, sizeof(*imp));
 526         if (imp == NULL)
 527                 return NULL;
 528
 529         INIT_LIST_HEAD(&imp->imp_replay_list);
 530         INIT_LIST_HEAD(&imp->imp_sending_list);
 531         INIT_LIST_HEAD(&imp->imp_delayed_list);
 532         spin_lock_init(&imp->imp_lock);
 533         imp->imp_conn_cnt = 0;
 534         imp->imp_max_transno = 0;
 535         imp->imp_peer_committed_transno = 0;
 536         imp->imp_state = LUSTRE_IMP_NEW;
 537         init_waitqueue_head(&imp->imp_recovery_waitq);
 538
 539         atomic_set(&imp->imp_refcount, 2);
 540         atomic_set(&imp->imp_inflight, 0);
 541         atomic_set(&imp->imp_replay_inflight, 0);
 542         INIT_LIST_HEAD(&imp->imp_handle.h_link);
 543         class_handle_hash(&imp->imp_handle, import_handle_addref);
 544
 545         return imp;
 546 }
 547
 548 void class_destroy_import(struct obd_import *import)
 549 {
 550         LASSERT(import != NULL);
 551         LASSERT((unsigned long)import != 0x5a5a5a5a);
 552
 553         class_handle_unhash(&import->imp_handle);
 554
 555         /* Abort any inflight DLM requests and NULL out their (about to be
 556          * freed) import. */
 557         /* Invalidate all requests on import, would be better to call
 558            ptlrpc_set_import_active(imp, 0); */
 559         import->imp_generation++;
 560         ptlrpc_abort_inflight_superhack(import);
 561
 562         class_import_put(import);
 563 }
 564
 565 /* A connection defines an export context in which preallocation can
 566    be managed. This releases the export pointer reference, and returns
 567    the export handle, so the export refcount is 1 when this function
 568    returns. */
 569 int class_connect(struct lustre_handle *conn, struct obd_device *obd,
 570                   struct obd_uuid *cluuid)
 571 {
 572         struct obd_export *export;
 573         LASSERT(conn != NULL);
 574         LASSERT(obd != NULL);
 575         LASSERT(cluuid != NULL);
 576         ENTRY;
 577
 578         export = class_new_export(obd);
 579         if (export == NULL)
 580                 RETURN(-ENOMEM);
 581
 582         conn->cookie = export->exp_handle.h_cookie;
 583         memcpy(&export->exp_client_uuid, cluuid,
 584                sizeof(export->exp_client_uuid));
 585         class_export_put(export);
 586
 587         CDEBUG(D_IOCTL, "connect: client %s, cookie "LPX64"\n",
 588                cluuid->uuid, conn->cookie);
 589         RETURN(0);
 590 }
 591
 592 /* This function removes two references from the export: one for the
 593  * hash entry and one for the export pointer passed in.  The export
 594  * pointer passed to this function is destroyed should not be used
 595  * again. */
 596 int class_disconnect(struct obd_export *export, int flags)
 597 {
 598         ENTRY;
 599
 600         if (export == NULL) {
 601                 fixme();
 602                 CDEBUG(D_IOCTL, "attempting to free NULL export %p\n", export);
 603                 RETURN(-EINVAL);
 604         }
 605
 606         /* XXX this shouldn't have to be here, but double-disconnect will crash
 607          * otherwise, and sometimes double-disconnect happens.  abort_recovery,
 608          * for example. */
 609         if (list_empty(&export->exp_handle.h_link))
 610                 RETURN(0);
 611
 612         CDEBUG(D_IOCTL, "disconnect: cookie "LPX64"\n",
 613                export->exp_handle.h_cookie);
 614
 615         class_unlink_export(export);
 616         class_export_put(export);
 617         RETURN(0);
 618 }
 619
 620 void class_disconnect_exports(struct obd_device *obd, int flags)
 621 {
 622         int rc;
 623         struct list_head *tmp, *n, work_list;
 624         struct lustre_handle fake_conn;
 625         struct obd_export *fake_exp, *exp;
 626         ENTRY;
 627
 628         /* Move all of the exports from obd_exports to a work list, en masse. */
 629         spin_lock(&obd->obd_dev_lock);
 630         list_add(&work_list, &obd->obd_exports);
 631         list_del_init(&obd->obd_exports);
 632         spin_unlock(&obd->obd_dev_lock);
 633
 634         CDEBUG(D_HA, "OBD device %d (%p) has exports, "
 635                "disconnecting them\n", obd->obd_minor, obd);
 636         list_for_each_safe(tmp, n, &work_list) {
 637                 exp = list_entry(tmp, struct obd_export, exp_obd_chain);
 638                 class_export_get(exp);
 639
 640                 if (obd_uuid_equals(&exp->exp_client_uuid,
 641                                     &exp->exp_obd->obd_uuid)) {
 642                         CDEBUG(D_HA,
 643                                "exp %p export uuid == obd uuid, don't discon\n",
 644                                exp);
 645                         /* Need to delete this now so we don't end up pointing
 646                          * to work_list later when this export is cleaned up. */
 647                         list_del_init(&exp->exp_obd_chain);
 648                         class_export_put(exp);
 649                         continue;
 650                 }
 651
 652                 fake_conn.cookie = exp->exp_handle.h_cookie;
 653                 fake_exp = class_conn2export(&fake_conn);
 654                 if (!fake_exp) {
 655                         class_export_put(exp);
 656                         continue;
 657                 }
 658                 rc = obd_disconnect(fake_exp, flags);
 659                 class_export_put(exp);
 660                 if (rc) {
 661                         CDEBUG(D_HA, "disconnecting export %p failed: %d\n",
 662                                exp, rc);
 663                 } else {
 664                         CDEBUG(D_HA, "export %p disconnected\n", exp);
 665                 }
 666         }
 667         EXIT;
 668 }
 669
 670 int oig_init(struct obd_io_group **oig_out)
 671 {
 672         struct obd_io_group *oig;
 673         ENTRY;
 674
 675         OBD_ALLOC(oig, sizeof(*oig));
 676         if (oig == NULL)
 677                 RETURN(-ENOMEM);
 678
 679         spin_lock_init(&oig->oig_lock);
 680         oig->oig_rc = 0;
 681         oig->oig_pending = 0;
 682         atomic_set(&oig->oig_refcount, 1);
 683         init_waitqueue_head(&oig->oig_waitq);
 684         INIT_LIST_HEAD(&oig->oig_occ_list);
 685
 686         *oig_out = oig;
 687         RETURN(0);
 688 };
 689
 690 static inline void oig_grab(struct obd_io_group *oig)
 691 {
 692         atomic_inc(&oig->oig_refcount);
 693 }
 694 void oig_release(struct obd_io_group *oig)
 695 {
 696         if (atomic_dec_and_test(&oig->oig_refcount))
 697                 OBD_FREE(oig, sizeof(*oig));
 698 }
 699
 700 void oig_add_one(struct obd_io_group *oig,
 701                   struct oig_callback_context *occ)
 702 {
 703         unsigned long flags;
 704         CDEBUG(D_CACHE, "oig %p ready to roll\n", oig);
 705         spin_lock_irqsave(&oig->oig_lock, flags);
 706         oig->oig_pending++;
 707         if (occ != NULL)
 708                 list_add_tail(&occ->occ_oig_item, &oig->oig_occ_list);
 709         spin_unlock_irqrestore(&oig->oig_lock, flags);
 710         oig_grab(oig);
 711 }
 712
 713 void oig_complete_one(struct obd_io_group *oig,
 714                       struct oig_callback_context *occ, int rc)
 715 {
 716         unsigned long flags;
 717         wait_queue_head_t *wake = NULL;
 718         int old_rc;
 719
 720         spin_lock_irqsave(&oig->oig_lock, flags);
 721
 722         if (occ != NULL)
 723                 list_del_init(&occ->occ_oig_item);
 724
 725         old_rc = oig->oig_rc;
 726         if (oig->oig_rc == 0 && rc != 0)
 727                 oig->oig_rc = rc;
 728
 729         if (--oig->oig_pending <= 0)
 730                 wake = &oig->oig_waitq;
 731
 732         spin_unlock_irqrestore(&oig->oig_lock, flags);
 733
 734         CDEBUG(D_CACHE, "oig %p completed, rc %d -> %d via %d, %d now "
 735                         "pending (racey)\n", oig, old_rc, oig->oig_rc, rc,
 736                         oig->oig_pending);
 737         if (wake)
 738                 wake_up(wake);
 739         oig_release(oig);
 740 }
 741
 742 static int oig_done(struct obd_io_group *oig)
 743 {
 744         unsigned long flags;
 745         int rc = 0;
 746         spin_lock_irqsave(&oig->oig_lock, flags);
 747         if (oig->oig_pending <= 0)
 748                 rc = 1;
 749         spin_unlock_irqrestore(&oig->oig_lock, flags);
 750         return rc;
 751 }
 752
 753 static void interrupted_oig(void *data)
 754 {
 755         struct obd_io_group *oig = data;
 756         struct list_head *pos;
 757         struct oig_callback_context *occ;
 758         unsigned long flags;
 759
 760         spin_lock_irqsave(&oig->oig_lock, flags);
 761         list_for_each(pos, &oig->oig_occ_list) {
 762                 occ = list_entry(pos, struct oig_callback_context,
 763                                  occ_oig_item);
 764                 occ->occ_interrupted(occ);
 765         }
 766         spin_unlock_irqrestore(&oig->oig_lock, flags);
 767 }
 768
 769 int oig_wait(struct obd_io_group *oig)
 770 {
 771         struct l_wait_info lwi = LWI_INTR(interrupted_oig, oig);
 772         int rc;
 773
 774         CDEBUG(D_CACHE, "waiting for oig %p\n", oig);
 775
 776         do {
 777                 rc = l_wait_event(oig->oig_waitq, oig_done(oig), &lwi);
 778                 LASSERTF(rc == 0 || rc == -EINTR, "rc: %d\n", rc);
 779                 /* we can't continue until the oig has emptied and stopped
 780                  * referencing state that the caller will free upon return */
 781                 if (rc == -EINTR)
 782                         lwi = (struct l_wait_info){ 0, };
 783         } while (rc == -EINTR);
 784
 785         LASSERTF(oig->oig_pending == 0,
 786                  "exiting oig_wait(oig = %p) with %d pending\n", oig,
 787                  oig->oig_pending);
 788
 789         CDEBUG(D_CACHE, "done waiting on oig %p rc %d\n", oig, oig->oig_rc);
 790         return oig->oig_rc;
 791 }