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(struct obd_type *type)
 191 {
 192         struct obd_device *result = NULL;
 193         int i;
 194
 195         spin_lock(&obd_dev_lock);
 196         for (i = 0 ; i < MAX_OBD_DEVICES && result == NULL; i++) {
 197                 struct obd_device *obd = &obd_dev[i];
 198                 if (!obd->obd_type) {
 199                         LASSERT(obd->obd_minor == i);
 200                         memset(obd, 0, sizeof(*obd));
 201                         obd->obd_minor = i;
 202                         obd->obd_type = type;
 203                         result = obd;
 204                 }
 205         }
 206         spin_unlock(&obd_dev_lock);
 207         return result;
 208 }
 209
 210 void class_release_dev(struct obd_device *obd)
 211 {
 212         int minor = obd->obd_minor;
 213
 214         spin_lock(&obd_dev_lock);
 215         memset(obd, 0, sizeof(*obd));
 216         obd->obd_minor = minor;
 217         spin_unlock(&obd_dev_lock);
 218 }
 219
 220 int class_name2dev(char *name)
 221 {
 222         int i;
 223
 224         if (!name)
 225                 return -1;
 226
 227         spin_lock(&obd_dev_lock);
 228         for (i = 0; i < MAX_OBD_DEVICES; i++) {
 229                 struct obd_device *obd = &obd_dev[i];
 230                 if (obd->obd_name && strcmp(name, obd->obd_name) == 0) {
 231                         spin_unlock(&obd_dev_lock);
 232                         return i;
 233                 }
 234         }
 235         spin_unlock(&obd_dev_lock);
 236         return -1;
 237 }
 238
 239 struct obd_device *class_name2obd(char *name)
 240 {
 241         int dev = class_name2dev(name);
 242         if (dev < 0)
 243                 return NULL;
 244         return &obd_dev[dev];
 245 }
 246
 247 int class_uuid2dev(struct obd_uuid *uuid)
 248 {
 249         int i;
 250         spin_lock(&obd_dev_lock);
 251         for (i = 0; i < MAX_OBD_DEVICES; i++) {
 252                 struct obd_device *obd = &obd_dev[i];
 253                 if (obd_uuid_equals(uuid, &obd->obd_uuid)) {
 254                         spin_unlock(&obd_dev_lock);
 255                         return i;
 256                 }
 257         }
 258         spin_unlock(&obd_dev_lock);
 259         return -1;
 260 }
 261
 262 struct obd_device *class_uuid2obd(struct obd_uuid *uuid)
 263 {
 264         int dev = class_uuid2dev(uuid);
 265         if (dev < 0)
 266                 return NULL;
 267         return &obd_dev[dev];
 268 }
 269
 270 /* Search for a client OBD connected to tgt_uuid.  If grp_uuid is
 271    specified, then only the client with that uuid is returned,
 272    otherwise any client connected to the tgt is returned.
 273    If tgt_uuid is NULL, the lov with grp_uuid is returned. */
 274 struct obd_device * class_find_client_obd(struct obd_uuid *tgt_uuid,
 275                                           char * typ_name,
 276                                           struct obd_uuid *grp_uuid)
 277 {
 278         int i;
 279
 280         spin_lock(&obd_dev_lock);
 281         for (i = 0; i < MAX_OBD_DEVICES; i++) {
 282                 struct obd_device *obd = &obd_dev[i];
 283                 if (obd->obd_type == NULL)
 284                         continue;
 285                 if ((strncmp(obd->obd_type->typ_name, typ_name,
 286                              strlen(typ_name)) == 0)) {
 287                         struct client_obd *cli = &obd->u.cli;
 288                         struct obd_import *imp = cli->cl_import;
 289                         if (tgt_uuid == NULL) {
 290                                 LASSERT(grp_uuid);
 291                                 if (obd_uuid_equals(grp_uuid, &obd->obd_uuid))
 292                                         return obd;
 293                                 continue;
 294                         }
 295                         if (obd_uuid_equals(tgt_uuid, &imp->imp_target_uuid) &&
 296                             ((grp_uuid)? obd_uuid_equals(grp_uuid,
 297                                                          &obd->obd_uuid) : 1)) {
 298                                 spin_unlock(&obd_dev_lock);
 299                                 return obd;
 300                         }
 301                 }
 302         }
 303         spin_unlock(&obd_dev_lock);
 304         return NULL;
 305 }
 306
 307 /* Iterate the obd_device list looking devices have grp_uuid. Start
 308    searching at *next, and if a device is found, the next index to look
 309    it is saved in *next. If next is NULL, then the first matching device
 310    will always be returned. */
 311 struct obd_device * class_devices_in_group(struct obd_uuid *grp_uuid, int *next)
 312 {
 313         int i;
 314
 315         if (next == NULL)
 316                 i = 0;
 317         else if (*next >= 0 && *next < MAX_OBD_DEVICES)
 318                 i = *next;
 319         else
 320                 return NULL;
 321         spin_lock(&obd_dev_lock);
 322         for (; i < MAX_OBD_DEVICES; i++) {
 323                 struct obd_device *obd = &obd_dev[i];
 324                 if (obd->obd_type == NULL)
 325                         continue;
 326                 if (obd_uuid_equals(grp_uuid, &obd->obd_uuid)) {
 327                         if (next != NULL)
 328                                 *next = i+1;
 329                         spin_unlock(&obd_dev_lock);
 330                         return obd;
 331                 }
 332         }
 333
 334         spin_unlock(&obd_dev_lock);
 335         return NULL;
 336 }
 337
 338
 339 void obd_cleanup_caches(void)
 340 {
 341         ENTRY;
 342         if (obdo_cachep) {
 343                 LASSERTF(kmem_cache_destroy(obdo_cachep) == 0,
 344                          "Cannot destory ll_obdo_cache\n");
 345                 obdo_cachep = NULL;
 346         }
 347         if (import_cachep) {
 348                 LASSERTF(kmem_cache_destroy(import_cachep) == 0,
 349                          "Cannot destory ll_import_cache\n");
 350                 import_cachep = NULL;
 351         }
 352         EXIT;
 353 }
 354
 355 int obd_init_caches(void)
 356 {
 357         ENTRY;
 358         LASSERT(obdo_cachep == NULL);
 359         obdo_cachep = kmem_cache_create("ll_obdo_cache", sizeof(struct obdo),
 360                                         0, 0, NULL, NULL);
 361         if (!obdo_cachep)
 362                 GOTO(out, -ENOMEM);
 363
 364         LASSERT(import_cachep == NULL);
 365         import_cachep = kmem_cache_create("ll_import_cache",
 366                                           sizeof(struct obd_import),
 367                                           0, 0, NULL, NULL);
 368         if (!import_cachep)
 369                 GOTO(out, -ENOMEM);
 370
 371         RETURN(0);
 372  out:
 373         obd_cleanup_caches();
 374         RETURN(-ENOMEM);
 375
 376 }
 377
 378 /* map connection to client */
 379 struct obd_export *class_conn2export(struct lustre_handle *conn)
 380 {
 381         struct obd_export *export;
 382         ENTRY;
 383
 384         if (!conn) {
 385                 CDEBUG(D_CACHE, "looking for null handle\n");
 386                 RETURN(NULL);
 387         }
 388
 389         if (conn->cookie == -1) {  /* this means assign a new connection */
 390                 CDEBUG(D_CACHE, "want a new connection\n");
 391                 RETURN(NULL);
 392         }
 393
 394         CDEBUG(D_IOCTL, "looking for export cookie "LPX64"\n", conn->cookie);
 395         export = class_handle2object(conn->cookie);
 396         RETURN(export);
 397 }
 398
 399 struct obd_device *class_exp2obd(struct obd_export *exp)
 400 {
 401         if (exp)
 402                 return exp->exp_obd;
 403         return NULL;
 404 }
 405
 406 struct obd_device *class_conn2obd(struct lustre_handle *conn)
 407 {
 408         struct obd_export *export;
 409         export = class_conn2export(conn);
 410         if (export) {
 411                 struct obd_device *obd = export->exp_obd;
 412                 class_export_put(export);
 413                 return obd;
 414         }
 415         return NULL;
 416 }
 417
 418 struct obd_import *class_exp2cliimp(struct obd_export *exp)
 419 {
 420         struct obd_device *obd = exp->exp_obd;
 421         if (obd == NULL)
 422                 return NULL;
 423         return obd->u.cli.cl_import;
 424 }
 425
 426 struct obd_import *class_conn2cliimp(struct lustre_handle *conn)
 427 {
 428         struct obd_device *obd = class_conn2obd(conn);
 429         if (obd == NULL)
 430                 return NULL;
 431         return obd->u.cli.cl_import;
 432 }
 433
 434 /* Export management functions */
 435 static void export_handle_addref(void *export)
 436 {
 437         class_export_get(export);
 438 }
 439
 440 void __class_export_put(struct obd_export *exp)
 441 {
 442         if (atomic_dec_and_test(&exp->exp_refcount)) {
 443                 struct obd_device *obd = exp->exp_obd;
 444                 CDEBUG(D_IOCTL, "destroying export %p/%s\n", exp,
 445                        exp->exp_client_uuid.uuid);
 446
 447                 LASSERT(obd != NULL);
 448
 449                 /* "Local" exports (lctl, LOV->{mdc,osc}) have no connection. */
 450                 if (exp->exp_connection)
 451                         ptlrpc_put_connection_superhack(exp->exp_connection);
 452
 453                 LASSERT(list_empty(&exp->exp_outstanding_replies));
 454                 LASSERT(list_empty(&exp->exp_handle.h_link));
 455                 obd_destroy_export(exp);
 456
 457                 OBD_FREE(exp, sizeof(*exp));
 458                 if (obd->obd_set_up) {
 459                         atomic_dec(&obd->obd_refcount);
 460                         wake_up(&obd->obd_refcount_waitq);
 461                 } else {
 462                         CERROR("removing export %p from obd %s (%p) -- OBD "
 463                                "not set up (refcount = %d)\n", exp,
 464                                obd->obd_name, obd,
 465                                atomic_read(&obd->obd_refcount));
 466                 }
 467         }
 468 }
 469
 470 /* Creates a new export, adds it to the hash table, and returns a
 471  * pointer to it. The refcount is 2: one for the hash reference, and
 472  * one for the pointer returned by this function. */
 473 struct obd_export *class_new_export(struct obd_device *obd)
 474 {
 475         struct obd_export *export;
 476
 477         OBD_ALLOC(export, sizeof(*export));
 478         if (!export) {
 479                 CERROR("no memory! (minor %d)\n", obd->obd_minor);
 480                 return NULL;
 481         }
 482
 483         export->exp_conn_cnt = 0;
 484         atomic_set(&export->exp_refcount, 2);
 485         atomic_set(&export->exp_rpc_count, 0);
 486         export->exp_obd = obd;
 487         export->exp_flags = 0;
 488         INIT_LIST_HEAD(&export->exp_outstanding_replies);
 489         /* XXX this should be in LDLM init */
 490         INIT_LIST_HEAD(&export->exp_ldlm_data.led_held_locks);
 491
 492         INIT_LIST_HEAD(&export->exp_handle.h_link);
 493         class_handle_hash(&export->exp_handle, export_handle_addref);
 494         spin_lock_init(&export->exp_lock);
 495
 496         spin_lock(&obd->obd_dev_lock);
 497         LASSERT(!obd->obd_stopping); /* shouldn't happen, but might race */
 498         atomic_inc(&obd->obd_refcount);
 499         list_add(&export->exp_obd_chain, &export->exp_obd->obd_exports);
 500         export->exp_obd->obd_num_exports++;
 501         spin_unlock(&obd->obd_dev_lock);
 502         obd_init_export(export);
 503         return export;
 504 }
 505
 506 void class_unlink_export(struct obd_export *exp)
 507 {
 508         class_handle_unhash(&exp->exp_handle);
 509
 510         spin_lock(&exp->exp_obd->obd_dev_lock);
 511         list_del_init(&exp->exp_obd_chain);
 512         exp->exp_obd->obd_num_exports--;
 513         spin_unlock(&exp->exp_obd->obd_dev_lock);
 514
 515         class_export_put(exp);
 516 }
 517
 518 /* Import management functions */
 519 static void import_handle_addref(void *import)
 520 {
 521         class_import_get(import);
 522 }
 523
 524 struct obd_import *class_import_get(struct obd_import *import)
 525 {
 526         atomic_inc(&import->imp_refcount);
 527         CDEBUG(D_IOCTL, "import %p refcount=%d\n", import,
 528                atomic_read(&import->imp_refcount));
 529         return import;
 530 }
 531
 532 void class_import_put(struct obd_import *import)
 533 {
 534         ENTRY;
 535
 536         CDEBUG(D_IOCTL, "import %p refcount=%d\n", import,
 537                atomic_read(&import->imp_refcount) - 1);
 538
 539         LASSERT(atomic_read(&import->imp_refcount) > 0);
 540         LASSERT(atomic_read(&import->imp_refcount) < 0x5a5a5a);
 541
 542         if (!atomic_dec_and_test(&import->imp_refcount)) {
 543                 EXIT;
 544                 return;
 545         }
 546
 547         CDEBUG(D_IOCTL, "destroying import %p\n", import);
 548
 549         if (import->imp_connection)
 550                 ptlrpc_put_connection_superhack(import->imp_connection);
 551
 552         while (!list_empty(&import->imp_conn_list)) {
 553                 struct obd_import_conn *imp_conn;
 554
 555                 imp_conn = list_entry(import->imp_conn_list.next,
 556                                       struct obd_import_conn, oic_item);
 557                 list_del(&imp_conn->oic_item);
 558                 if (imp_conn->oic_conn)
 559                         ptlrpc_put_connection_superhack(imp_conn->oic_conn);
 560                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 561         }
 562
 563         LASSERT(list_empty(&import->imp_handle.h_link));
 564         OBD_FREE(import, sizeof(*import));
 565         EXIT;
 566 }
 567
 568 struct obd_import *class_new_import(void)
 569 {
 570         struct obd_import *imp;
 571
 572         OBD_ALLOC(imp, sizeof(*imp));
 573         if (imp == NULL)
 574                 return NULL;
 575
 576         INIT_LIST_HEAD(&imp->imp_replay_list);
 577         INIT_LIST_HEAD(&imp->imp_sending_list);
 578         INIT_LIST_HEAD(&imp->imp_delayed_list);
 579         spin_lock_init(&imp->imp_lock);
 580         imp->imp_conn_cnt = 0;
 581         imp->imp_max_transno = 0;
 582         imp->imp_peer_committed_transno = 0;
 583         imp->imp_state = LUSTRE_IMP_NEW;
 584         init_waitqueue_head(&imp->imp_recovery_waitq);
 585
 586         atomic_set(&imp->imp_refcount, 2);
 587         atomic_set(&imp->imp_inflight, 0);
 588         atomic_set(&imp->imp_replay_inflight, 0);
 589         INIT_LIST_HEAD(&imp->imp_conn_list);
 590         INIT_LIST_HEAD(&imp->imp_handle.h_link);
 591         class_handle_hash(&imp->imp_handle, import_handle_addref);
 592
 593         return imp;
 594 }
 595
 596 void class_destroy_import(struct obd_import *import)
 597 {
 598         LASSERT(import != NULL);
 599         LASSERT(import != LP_POISON);
 600
 601         class_handle_unhash(&import->imp_handle);
 602
 603         /* Abort any inflight DLM requests and NULL out their (about to be
 604          * freed) import. */
 605         /* Invalidate all requests on import, would be better to call
 606            ptlrpc_set_import_active(imp, 0); */
 607         import->imp_generation++;
 608         ptlrpc_abort_inflight_superhack(import);
 609
 610         class_import_put(import);
 611 }
 612
 613 /* A connection defines an export context in which preallocation can
 614    be managed. This releases the export pointer reference, and returns
 615    the export handle, so the export refcount is 1 when this function
 616    returns. */
 617 int class_connect(struct lustre_handle *conn, struct obd_device *obd,
 618                   struct obd_uuid *cluuid)
 619 {
 620         struct obd_export *export;
 621         LASSERT(conn != NULL);
 622         LASSERT(obd != NULL);
 623         LASSERT(cluuid != NULL);
 624         ENTRY;
 625
 626         export = class_new_export(obd);
 627         if (export == NULL)
 628                 RETURN(-ENOMEM);
 629
 630         conn->cookie = export->exp_handle.h_cookie;
 631         memcpy(&export->exp_client_uuid, cluuid,
 632                sizeof(export->exp_client_uuid));
 633         class_export_put(export);
 634
 635         CDEBUG(D_IOCTL, "connect: client %s, cookie "LPX64"\n",
 636                cluuid->uuid, conn->cookie);
 637         RETURN(0);
 638 }
 639
 640 /* This function removes two references from the export: one for the
 641  * hash entry and one for the export pointer passed in.  The export
 642  * pointer passed to this function is destroyed should not be used
 643  * again. */
 644 int class_disconnect(struct obd_export *export, int flags)
 645 {
 646         ENTRY;
 647
 648         if (export == NULL) {
 649                 fixme();
 650                 CDEBUG(D_IOCTL, "attempting to free NULL export %p\n", export);
 651                 RETURN(-EINVAL);
 652         }
 653
 654         /* XXX this shouldn't have to be here, but double-disconnect will crash
 655          * otherwise, and sometimes double-disconnect happens.  abort_recovery,
 656          * for example. */
 657         if (list_empty(&export->exp_handle.h_link))
 658                 RETURN(0);
 659
 660         CDEBUG(D_IOCTL, "disconnect: cookie "LPX64"\n",
 661                export->exp_handle.h_cookie);
 662
 663         if (export->exp_handle.h_cookie == LL_POISON) {
 664                 CERROR("disconnecting freed export %p, ignoring\n", export);
 665         } else {
 666                 class_unlink_export(export);
 667                 class_export_put(export);
 668         }
 669         RETURN(0);
 670 }
 671
 672 static void  class_disconnect_export_list(struct list_head *list, int flags)
 673 {
 674         struct obd_export *fake_exp, *exp;
 675         struct lustre_handle fake_conn;
 676         int rc;
 677         ENTRY;
 678
 679         /* Move all of the exports from obd_exports to a work list, en masse. */
 680         /* It's possible that an export may disconnect itself, but
 681          * nothing else will be added to this list. */
 682         while(!list_empty(list)) {
 683                 exp = list_entry(list->next, struct obd_export, exp_obd_chain);
 684                 class_export_get(exp);
 685
 686                 if (obd_uuid_equals(&exp->exp_client_uuid,
 687                                     &exp->exp_obd->obd_uuid)) {
 688                         CDEBUG(D_HA,
 689                                "exp %p export uuid == obd uuid, don't discon\n",
 690                                exp);
 691                         /*
 692                          * need to delete this now so we don't end up pointing
 693                          * to work_list later when this export is cleaned up.
 694                          */
 695                         list_del_init(&exp->exp_obd_chain);
 696                         class_export_put(exp);
 697                         continue;
 698                 }
 699
 700                 fake_conn.cookie = exp->exp_handle.h_cookie;
 701                 fake_exp = class_conn2export(&fake_conn);
 702                 if (!fake_exp) {
 703                         class_export_put(exp);
 704                         continue;
 705                 }
 706
 707                 rc = obd_disconnect(fake_exp, flags);
 708                 class_export_put(exp);
 709                 if (rc) {
 710                         CDEBUG(D_HA, "disconnecting export %p failed: %d\n",
 711                                exp, rc);
 712                 } else {
 713                         CDEBUG(D_HA, "export %p disconnected\n", exp);
 714                 }
 715         }
 716         EXIT;
 717 }
 718
 719 void class_disconnect_exports(struct obd_device *obd, int flags)
 720 {
 721         struct list_head work_list;
 722         ENTRY;
 723
 724         /* Move all of the exports from obd_exports to a work list, en masse. */
 725         spin_lock(&obd->obd_dev_lock);
 726         list_add(&work_list, &obd->obd_exports);
 727         list_del_init(&obd->obd_exports);
 728         spin_unlock(&obd->obd_dev_lock);
 729
 730         CDEBUG(D_HA, "OBD device %d (%p) has exports, "
 731                "disconnecting them\n", obd->obd_minor, obd);
 732         class_disconnect_export_list(&work_list, flags);
 733         EXIT;
 734 }
 735
 736 /* Remove exports that have not completed recovery.
 737  */
 738 void class_disconnect_stale_exports(struct obd_device *obd, int flags)
 739 {
 740         struct list_head work_list;
 741         struct list_head *pos, *n;
 742         struct obd_export *exp;
 743         int cnt = 0;
 744         ENTRY;
 745
 746         INIT_LIST_HEAD(&work_list);
 747         spin_lock(&obd->obd_dev_lock);
 748         list_for_each_safe(pos, n, &obd->obd_exports) {
 749                 exp = list_entry(pos, struct obd_export, exp_obd_chain);
 750                 if (exp->exp_replay_needed) {
 751                         list_del(&exp->exp_obd_chain);
 752                         list_add(&exp->exp_obd_chain, &work_list);
 753                         cnt++;
 754                 }
 755         }
 756         spin_unlock(&obd->obd_dev_lock);
 757
 758         CDEBUG(D_ERROR, "%s: disconnecting %d stale clients\n",
 759                obd->obd_name, cnt);
 760         class_disconnect_export_list(&work_list, flags);
 761         EXIT;
 762 }
 763
 764
 765
 766 int oig_init(struct obd_io_group **oig_out)
 767 {
 768         struct obd_io_group *oig;
 769         ENTRY;
 770
 771         OBD_ALLOC(oig, sizeof(*oig));
 772         if (oig == NULL)
 773                 RETURN(-ENOMEM);
 774
 775         spin_lock_init(&oig->oig_lock);
 776         oig->oig_rc = 0;
 777         oig->oig_pending = 0;
 778         atomic_set(&oig->oig_refcount, 1);
 779         init_waitqueue_head(&oig->oig_waitq);
 780         INIT_LIST_HEAD(&oig->oig_occ_list);
 781
 782         *oig_out = oig;
 783         RETURN(0);
 784 };
 785
 786 static inline void oig_grab(struct obd_io_group *oig)
 787 {
 788         atomic_inc(&oig->oig_refcount);
 789 }
 790 void oig_release(struct obd_io_group *oig)
 791 {
 792         if (atomic_dec_and_test(&oig->oig_refcount))
 793                 OBD_FREE(oig, sizeof(*oig));
 794 }
 795
 796 void oig_add_one(struct obd_io_group *oig, struct oig_callback_context *occ)
 797 {
 798         unsigned long flags;
 799         CDEBUG(D_CACHE, "oig %p ready to roll\n", oig);
 800         spin_lock_irqsave(&oig->oig_lock, flags);
 801         oig->oig_pending++;
 802         if (occ != NULL)
 803                 list_add_tail(&occ->occ_oig_item, &oig->oig_occ_list);
 804         spin_unlock_irqrestore(&oig->oig_lock, flags);
 805         oig_grab(oig);
 806 }
 807
 808 void oig_complete_one(struct obd_io_group *oig,
 809                       struct oig_callback_context *occ, int rc)
 810 {
 811         unsigned long flags;
 812         wait_queue_head_t *wake = NULL;
 813         int old_rc;
 814
 815         spin_lock_irqsave(&oig->oig_lock, flags);
 816
 817         if (occ != NULL)
 818                 list_del_init(&occ->occ_oig_item);
 819
 820         old_rc = oig->oig_rc;
 821         if (oig->oig_rc == 0 && rc != 0)
 822                 oig->oig_rc = rc;
 823
 824         if (--oig->oig_pending <= 0)
 825                 wake = &oig->oig_waitq;
 826
 827         spin_unlock_irqrestore(&oig->oig_lock, flags);
 828
 829         CDEBUG(D_CACHE, "oig %p completed, rc %d -> %d via %d, %d now "
 830                         "pending (racey)\n", oig, old_rc, oig->oig_rc, rc,
 831                         oig->oig_pending);
 832         if (wake)
 833                 wake_up(wake);
 834         oig_release(oig);
 835 }
 836
 837 static int oig_done(struct obd_io_group *oig)
 838 {
 839         unsigned long flags;
 840         int rc = 0;
 841         spin_lock_irqsave(&oig->oig_lock, flags);
 842         if (oig->oig_pending <= 0)
 843                 rc = 1;
 844         spin_unlock_irqrestore(&oig->oig_lock, flags);
 845         return rc;
 846 }
 847
 848 static void interrupted_oig(void *data)
 849 {
 850         struct obd_io_group *oig = data;
 851         struct oig_callback_context *occ;
 852         unsigned long flags;
 853
 854         spin_lock_irqsave(&oig->oig_lock, flags);
 855         /* We need to restart the processing each time we drop the lock, as
 856          * it is possible other threads called oig_complete_one() to remove
 857          * an entry elsewhere in the list while we dropped lock.  We need to
 858          * drop the lock because osc_ap_completion() calls oig_complete_one()
 859          * which re-gets this lock ;-) as well as a lock ordering issue. */
 860 restart:
 861         list_for_each_entry(occ, &oig->oig_occ_list, occ_oig_item) {
 862                 if (occ->interrupted)
 863                         continue;
 864                 occ->interrupted = 1;
 865                 spin_unlock_irqrestore(&oig->oig_lock, flags);
 866                 occ->occ_interrupted(occ);
 867                 spin_lock_irqsave(&oig->oig_lock, flags);
 868                 goto restart;
 869         }
 870         spin_unlock_irqrestore(&oig->oig_lock, flags);
 871 }
 872
 873 int oig_wait(struct obd_io_group *oig)
 874 {
 875         struct l_wait_info lwi = LWI_INTR(interrupted_oig, oig);
 876         int rc;
 877
 878         CDEBUG(D_CACHE, "waiting for oig %p\n", oig);
 879
 880         do {
 881                 rc = l_wait_event(oig->oig_waitq, oig_done(oig), &lwi);
 882                 LASSERTF(rc == 0 || rc == -EINTR, "rc: %d\n", rc);
 883                 /* we can't continue until the oig has emptied and stopped
 884                  * referencing state that the caller will free upon return */
 885                 if (rc == -EINTR)
 886                         lwi = (struct l_wait_info){ 0, };
 887         } while (rc == -EINTR);
 888
 889         LASSERTF(oig->oig_pending == 0,
 890                  "exiting oig_wait(oig = %p) with %d pending\n", oig,
 891                  oig->oig_pending);
 892
 893         CDEBUG(D_CACHE, "done waiting on oig %p rc %d\n", oig, oig->oig_rc);
 894         return oig->oig_rc;
 895 }