lustre/obdclass/genops.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright (c) 1999, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/obdclass/genops.c
  37  *
  38  * These are the only exported functions, they provide some generic
  39  * infrastructure for managing object devices
  40  */
  41
  42 #define DEBUG_SUBSYSTEM S_CLASS
  43 #ifndef __KERNEL__
  44 #include <liblustre.h>
  45 #endif
  46 #include <obd_ost.h>
  47 #include <obd_class.h>
  48 #include <lprocfs_status.h>
  49
  50 extern cfs_list_t obd_types;
  51 cfs_spinlock_t obd_types_lock;
  52
  53 cfs_mem_cache_t *obd_device_cachep;
  54 cfs_mem_cache_t *obdo_cachep;
  55 EXPORT_SYMBOL(obdo_cachep);
  56 cfs_mem_cache_t *import_cachep;
  57
  58 cfs_list_t      obd_zombie_imports;
  59 cfs_list_t      obd_zombie_exports;
  60 cfs_spinlock_t  obd_zombie_impexp_lock;
  61 static void obd_zombie_impexp_notify(void);
  62 static void obd_zombie_export_add(struct obd_export *exp);
  63 static void obd_zombie_import_add(struct obd_import *imp);
  64 static void print_export_data(struct obd_export *exp,
  65                               const char *status, int locks);
  66
  67 int (*ptlrpc_put_connection_superhack)(struct ptlrpc_connection *c);
  68
  69 /*
  70  * support functions: we could use inter-module communication, but this
  71  * is more portable to other OS's
  72  */
  73 static struct obd_device *obd_device_alloc(void)
  74 {
  75         struct obd_device *obd;
  76
  77         OBD_SLAB_ALLOC_PTR_GFP(obd, obd_device_cachep, CFS_ALLOC_IO);
  78         if (obd != NULL) {
  79                 obd->obd_magic = OBD_DEVICE_MAGIC;
  80         }
  81         return obd;
  82 }
  83 EXPORT_SYMBOL(obd_device_alloc);
  84
  85 static void obd_device_free(struct obd_device *obd)
  86 {
  87         LASSERT(obd != NULL);
  88         LASSERTF(obd->obd_magic == OBD_DEVICE_MAGIC, "obd %p obd_magic %08x != %08x\n",
  89                  obd, obd->obd_magic, OBD_DEVICE_MAGIC);
  90         if (obd->obd_namespace != NULL) {
  91                 CERROR("obd %p: namespace %p was not properly cleaned up (obd_force=%d)!\n",
  92                        obd, obd->obd_namespace, obd->obd_force);
  93                 LBUG();
  94         }
  95         lu_ref_fini(&obd->obd_reference);
  96         OBD_SLAB_FREE_PTR(obd, obd_device_cachep);
  97 }
  98
  99 struct obd_type *class_search_type(const char *name)
 100 {
 101         cfs_list_t *tmp;
 102         struct obd_type *type;
 103
 104         cfs_spin_lock(&obd_types_lock);
 105         cfs_list_for_each(tmp, &obd_types) {
 106                 type = cfs_list_entry(tmp, struct obd_type, typ_chain);
 107                 if (strcmp(type->typ_name, name) == 0) {
 108                         cfs_spin_unlock(&obd_types_lock);
 109                         return type;
 110                 }
 111         }
 112         cfs_spin_unlock(&obd_types_lock);
 113         return NULL;
 114 }
 115
 116 struct obd_type *class_get_type(const char *name)
 117 {
 118         struct obd_type *type = class_search_type(name);
 119
 120 #ifdef HAVE_MODULE_LOADING_SUPPORT
 121         if (!type) {
 122                 const char *modname = name;
 123                 if (!cfs_request_module("%s", modname)) {
 124                         CDEBUG(D_INFO, "Loaded module '%s'\n", modname);
 125                         type = class_search_type(name);
 126                 } else {
 127                         LCONSOLE_ERROR_MSG(0x158, "Can't load module '%s'\n",
 128                                            modname);
 129                 }
 130         }
 131 #endif
 132         if (type) {
 133                 cfs_spin_lock(&type->obd_type_lock);
 134                 type->typ_refcnt++;
 135                 cfs_try_module_get(type->typ_dt_ops->o_owner);
 136                 cfs_spin_unlock(&type->obd_type_lock);
 137         }
 138         return type;
 139 }
 140
 141 void class_put_type(struct obd_type *type)
 142 {
 143         LASSERT(type);
 144         cfs_spin_lock(&type->obd_type_lock);
 145         type->typ_refcnt--;
 146         cfs_module_put(type->typ_dt_ops->o_owner);
 147         cfs_spin_unlock(&type->obd_type_lock);
 148 }
 149
 150 #define CLASS_MAX_NAME 1024
 151
 152 int class_register_type(struct obd_ops *dt_ops, struct md_ops *md_ops,
 153                         struct lprocfs_vars *vars, const char *name,
 154                         struct lu_device_type *ldt)
 155 {
 156         struct obd_type *type;
 157         int rc = 0;
 158         ENTRY;
 159
 160         /* sanity check */
 161         LASSERT(strnlen(name, CLASS_MAX_NAME) < CLASS_MAX_NAME);
 162
 163         if (class_search_type(name)) {
 164                 CDEBUG(D_IOCTL, "Type %s already registered\n", name);
 165                 RETURN(-EEXIST);
 166         }
 167
 168         rc = -ENOMEM;
 169         OBD_ALLOC(type, sizeof(*type));
 170         if (type == NULL)
 171                 RETURN(rc);
 172
 173         OBD_ALLOC_PTR(type->typ_dt_ops);
 174         OBD_ALLOC_PTR(type->typ_md_ops);
 175         OBD_ALLOC(type->typ_name, strlen(name) + 1);
 176
 177         if (type->typ_dt_ops == NULL ||
 178             type->typ_md_ops == NULL ||
 179             type->typ_name == NULL)
 180                 GOTO (failed, rc);
 181
 182         *(type->typ_dt_ops) = *dt_ops;
 183         /* md_ops is optional */
 184         if (md_ops)
 185                 *(type->typ_md_ops) = *md_ops;
 186         strcpy(type->typ_name, name);
 187         cfs_spin_lock_init(&type->obd_type_lock);
 188
 189 #ifdef LPROCFS
 190         type->typ_procroot = lprocfs_register(type->typ_name, proc_lustre_root,
 191                                               vars, type);
 192         if (IS_ERR(type->typ_procroot)) {
 193                 rc = PTR_ERR(type->typ_procroot);
 194                 type->typ_procroot = NULL;
 195                 GOTO (failed, rc);
 196         }
 197 #endif
 198         if (ldt != NULL) {
 199                 type->typ_lu = ldt;
 200                 rc = lu_device_type_init(ldt);
 201                 if (rc != 0)
 202                         GOTO (failed, rc);
 203         }
 204
 205         cfs_spin_lock(&obd_types_lock);
 206         cfs_list_add(&type->typ_chain, &obd_types);
 207         cfs_spin_unlock(&obd_types_lock);
 208
 209         RETURN (0);
 210
 211  failed:
 212         if (type->typ_name != NULL)
 213                 OBD_FREE(type->typ_name, strlen(name) + 1);
 214         if (type->typ_md_ops != NULL)
 215                 OBD_FREE_PTR(type->typ_md_ops);
 216         if (type->typ_dt_ops != NULL)
 217                 OBD_FREE_PTR(type->typ_dt_ops);
 218         OBD_FREE(type, sizeof(*type));
 219         RETURN(rc);
 220 }
 221
 222 int class_unregister_type(const char *name)
 223 {
 224         struct obd_type *type = class_search_type(name);
 225         ENTRY;
 226
 227         if (!type) {
 228                 CERROR("unknown obd type\n");
 229                 RETURN(-EINVAL);
 230         }
 231
 232         if (type->typ_refcnt) {
 233                 CERROR("type %s has refcount (%d)\n", name, type->typ_refcnt);
 234                 /* This is a bad situation, let's make the best of it */
 235                 /* Remove ops, but leave the name for debugging */
 236                 OBD_FREE_PTR(type->typ_dt_ops);
 237                 OBD_FREE_PTR(type->typ_md_ops);
 238                 RETURN(-EBUSY);
 239         }
 240
 241         if (type->typ_procroot) {
 242                 lprocfs_remove(&type->typ_procroot);
 243         }
 244
 245         if (type->typ_lu)
 246                 lu_device_type_fini(type->typ_lu);
 247
 248         cfs_spin_lock(&obd_types_lock);
 249         cfs_list_del(&type->typ_chain);
 250         cfs_spin_unlock(&obd_types_lock);
 251         OBD_FREE(type->typ_name, strlen(name) + 1);
 252         if (type->typ_dt_ops != NULL)
 253                 OBD_FREE_PTR(type->typ_dt_ops);
 254         if (type->typ_md_ops != NULL)
 255                 OBD_FREE_PTR(type->typ_md_ops);
 256         OBD_FREE(type, sizeof(*type));
 257         RETURN(0);
 258 } /* class_unregister_type */
 259
 260 /**
 261  * Create a new obd device.
 262  *
 263  * Find an empty slot in ::obd_devs[], create a new obd device in it.
 264  *
 265  * \param[in] type_name obd device type string.
 266  * \param[in] name      obd device name.
 267  *
 268  * \retval NULL if create fails, otherwise return the obd device
 269  *         pointer created.
 270  */
 271 struct obd_device *class_newdev(const char *type_name, const char *name)
 272 {
 273         struct obd_device *result = NULL;
 274         struct obd_device *newdev;
 275         struct obd_type *type = NULL;
 276         int i;
 277         int new_obd_minor = 0;
 278
 279         if (strlen(name) >= MAX_OBD_NAME) {
 280                 CERROR("name/uuid must be < %u bytes long\n", MAX_OBD_NAME);
 281                 RETURN(ERR_PTR(-EINVAL));
 282         }
 283
 284         type = class_get_type(type_name);
 285         if (type == NULL){
 286                 CERROR("OBD: unknown type: %s\n", type_name);
 287                 RETURN(ERR_PTR(-ENODEV));
 288         }
 289
 290         newdev = obd_device_alloc();
 291         if (newdev == NULL) {
 292                 class_put_type(type);
 293                 RETURN(ERR_PTR(-ENOMEM));
 294         }
 295         LASSERT(newdev->obd_magic == OBD_DEVICE_MAGIC);
 296
 297         cfs_spin_lock(&obd_dev_lock);
 298         for (i = 0; i < class_devno_max(); i++) {
 299                 struct obd_device *obd = class_num2obd(i);
 300                 if (obd && obd->obd_name &&
 301                     (strcmp(name, obd->obd_name) == 0)) {
 302                         CERROR("Device %s already exists, won't add\n", name);
 303                         if (result) {
 304                                 LASSERTF(result->obd_magic == OBD_DEVICE_MAGIC,
 305                                          "%p obd_magic %08x != %08x\n", result,
 306                                          result->obd_magic, OBD_DEVICE_MAGIC);
 307                                 LASSERTF(result->obd_minor == new_obd_minor,
 308                                          "%p obd_minor %d != %d\n", result,
 309                                          result->obd_minor, new_obd_minor);
 310
 311                                 obd_devs[result->obd_minor] = NULL;
 312                                 result->obd_name[0]='\0';
 313                          }
 314                         result = ERR_PTR(-EEXIST);
 315                         break;
 316                 }
 317                 if (!result && !obd) {
 318                         result = newdev;
 319                         result->obd_minor = i;
 320                         new_obd_minor = i;
 321                         result->obd_type = type;
 322                         strncpy(result->obd_name, name,
 323                                 sizeof(result->obd_name) - 1);
 324                         obd_devs[i] = result;
 325                 }
 326         }
 327         cfs_spin_unlock(&obd_dev_lock);
 328
 329         if (result == NULL && i >= class_devno_max()) {
 330                 CERROR("all %u OBD devices used, increase MAX_OBD_DEVICES\n",
 331                        class_devno_max());
 332                 result = ERR_PTR(-EOVERFLOW);
 333         }
 334
 335         if (IS_ERR(result)) {
 336                 obd_device_free(newdev);
 337                 class_put_type(type);
 338         } else {
 339                 CDEBUG(D_IOCTL, "Adding new device %s (%p)\n",
 340                        result->obd_name, result);
 341         }
 342         return result;
 343 }
 344
 345 void class_release_dev(struct obd_device *obd)
 346 {
 347         struct obd_type *obd_type = obd->obd_type;
 348
 349         LASSERTF(obd->obd_magic == OBD_DEVICE_MAGIC, "%p obd_magic %08x != %08x\n",
 350                  obd, obd->obd_magic, OBD_DEVICE_MAGIC);
 351         LASSERTF(obd == obd_devs[obd->obd_minor], "obd %p != obd_devs[%d] %p\n",
 352                  obd, obd->obd_minor, obd_devs[obd->obd_minor]);
 353         LASSERT(obd_type != NULL);
 354
 355         CDEBUG(D_INFO, "Release obd device %s obd_type name =%s\n",
 356                obd->obd_name,obd->obd_type->typ_name);
 357
 358         cfs_spin_lock(&obd_dev_lock);
 359         obd_devs[obd->obd_minor] = NULL;
 360         cfs_spin_unlock(&obd_dev_lock);
 361         obd_device_free(obd);
 362
 363         class_put_type(obd_type);
 364 }
 365
 366 int class_name2dev(const char *name)
 367 {
 368         int i;
 369
 370         if (!name)
 371                 return -1;
 372
 373         cfs_spin_lock(&obd_dev_lock);
 374         for (i = 0; i < class_devno_max(); i++) {
 375                 struct obd_device *obd = class_num2obd(i);
 376                 if (obd && obd->obd_name && strcmp(name, obd->obd_name) == 0) {
 377                         /* Make sure we finished attaching before we give
 378                            out any references */
 379                         LASSERT(obd->obd_magic == OBD_DEVICE_MAGIC);
 380                         if (obd->obd_attached) {
 381                                 cfs_spin_unlock(&obd_dev_lock);
 382                                 return i;
 383                         }
 384                         break;
 385                 }
 386         }
 387         cfs_spin_unlock(&obd_dev_lock);
 388
 389         return -1;
 390 }
 391
 392 struct obd_device *class_name2obd(const char *name)
 393 {
 394         int dev = class_name2dev(name);
 395
 396         if (dev < 0 || dev > class_devno_max())
 397                 return NULL;
 398         return class_num2obd(dev);
 399 }
 400
 401 int class_uuid2dev(struct obd_uuid *uuid)
 402 {
 403         int i;
 404
 405         cfs_spin_lock(&obd_dev_lock);
 406         for (i = 0; i < class_devno_max(); i++) {
 407                 struct obd_device *obd = class_num2obd(i);
 408                 if (obd && obd_uuid_equals(uuid, &obd->obd_uuid)) {
 409                         LASSERT(obd->obd_magic == OBD_DEVICE_MAGIC);
 410                         cfs_spin_unlock(&obd_dev_lock);
 411                         return i;
 412                 }
 413         }
 414         cfs_spin_unlock(&obd_dev_lock);
 415
 416         return -1;
 417 }
 418
 419 struct obd_device *class_uuid2obd(struct obd_uuid *uuid)
 420 {
 421         int dev = class_uuid2dev(uuid);
 422         if (dev < 0)
 423                 return NULL;
 424         return class_num2obd(dev);
 425 }
 426
 427 /**
 428  * Get obd device from ::obd_devs[]
 429  *
 430  * \param num [in] array index
 431  *
 432  * \retval NULL if ::obd_devs[\a num] does not contains an obd device
 433  *         otherwise return the obd device there.
 434  */
 435 struct obd_device *class_num2obd(int num)
 436 {
 437         struct obd_device *obd = NULL;
 438
 439         if (num < class_devno_max()) {
 440                 obd = obd_devs[num];
 441                 if (obd == NULL)
 442                         return NULL;
 443
 444                 LASSERTF(obd->obd_magic == OBD_DEVICE_MAGIC,
 445                          "%p obd_magic %08x != %08x\n",
 446                          obd, obd->obd_magic, OBD_DEVICE_MAGIC);
 447                 LASSERTF(obd->obd_minor == num,
 448                          "%p obd_minor %0d != %0d\n",
 449                          obd, obd->obd_minor, num);
 450         }
 451
 452         return obd;
 453 }
 454
 455 void class_obd_list(void)
 456 {
 457         char *status;
 458         int i;
 459
 460         cfs_spin_lock(&obd_dev_lock);
 461         for (i = 0; i < class_devno_max(); i++) {
 462                 struct obd_device *obd = class_num2obd(i);
 463                 if (obd == NULL)
 464                         continue;
 465                 if (obd->obd_stopping)
 466                         status = "ST";
 467                 else if (obd->obd_set_up)
 468                         status = "UP";
 469                 else if (obd->obd_attached)
 470                         status = "AT";
 471                 else
 472                         status = "--";
 473                 LCONSOLE(D_CONFIG, "%3d %s %s %s %s %d\n",
 474                          i, status, obd->obd_type->typ_name,
 475                          obd->obd_name, obd->obd_uuid.uuid,
 476                          cfs_atomic_read(&obd->obd_refcount));
 477         }
 478         cfs_spin_unlock(&obd_dev_lock);
 479         return;
 480 }
 481
 482 /* Search for a client OBD connected to tgt_uuid.  If grp_uuid is
 483    specified, then only the client with that uuid is returned,
 484    otherwise any client connected to the tgt is returned. */
 485 struct obd_device * class_find_client_obd(struct obd_uuid *tgt_uuid,
 486                                           const char * typ_name,
 487                                           struct obd_uuid *grp_uuid)
 488 {
 489         int i;
 490
 491         cfs_spin_lock(&obd_dev_lock);
 492         for (i = 0; i < class_devno_max(); i++) {
 493                 struct obd_device *obd = class_num2obd(i);
 494                 if (obd == NULL)
 495                         continue;
 496                 if ((strncmp(obd->obd_type->typ_name, typ_name,
 497                              strlen(typ_name)) == 0)) {
 498                         if (obd_uuid_equals(tgt_uuid,
 499                                             &obd->u.cli.cl_target_uuid) &&
 500                             ((grp_uuid)? obd_uuid_equals(grp_uuid,
 501                                                          &obd->obd_uuid) : 1)) {
 502                                 cfs_spin_unlock(&obd_dev_lock);
 503                                 return obd;
 504                         }
 505                 }
 506         }
 507         cfs_spin_unlock(&obd_dev_lock);
 508
 509         return NULL;
 510 }
 511
 512 /* Iterate the obd_device list looking devices have grp_uuid. Start
 513    searching at *next, and if a device is found, the next index to look
 514    at is saved in *next. If next is NULL, then the first matching device
 515    will always be returned. */
 516 struct obd_device * class_devices_in_group(struct obd_uuid *grp_uuid, int *next)
 517 {
 518         int i;
 519
 520         if (next == NULL)
 521                 i = 0;
 522         else if (*next >= 0 && *next < class_devno_max())
 523                 i = *next;
 524         else
 525                 return NULL;
 526
 527         cfs_spin_lock(&obd_dev_lock);
 528         for (; i < class_devno_max(); i++) {
 529                 struct obd_device *obd = class_num2obd(i);
 530                 if (obd == NULL)
 531                         continue;
 532                 if (obd_uuid_equals(grp_uuid, &obd->obd_uuid)) {
 533                         if (next != NULL)
 534                                 *next = i+1;
 535                         cfs_spin_unlock(&obd_dev_lock);
 536                         return obd;
 537                 }
 538         }
 539         cfs_spin_unlock(&obd_dev_lock);
 540
 541         return NULL;
 542 }
 543
 544 /**
 545  * to notify sptlrpc log for \a fsname has changed, let every relevant OBD
 546  * adjust sptlrpc settings accordingly.
 547  */
 548 int class_notify_sptlrpc_conf(const char *fsname, int namelen)
 549 {
 550         struct obd_device  *obd;
 551         const char         *type;
 552         int                 i, rc = 0, rc2;
 553
 554         LASSERT(namelen > 0);
 555
 556         cfs_spin_lock(&obd_dev_lock);
 557         for (i = 0; i < class_devno_max(); i++) {
 558                 obd = class_num2obd(i);
 559
 560                 if (obd == NULL || obd->obd_set_up == 0 || obd->obd_stopping)
 561                         continue;
 562
 563                 /* only notify mdc, osc, mdt, ost */
 564                 type = obd->obd_type->typ_name;
 565                 if (strcmp(type, LUSTRE_MDC_NAME) != 0 &&
 566                     strcmp(type, LUSTRE_OSC_NAME) != 0 &&
 567                     strcmp(type, LUSTRE_MDT_NAME) != 0 &&
 568                     strcmp(type, LUSTRE_OST_NAME) != 0)
 569                         continue;
 570
 571                 if (strncmp(obd->obd_name, fsname, namelen))
 572                         continue;
 573
 574                 class_incref(obd, __FUNCTION__, obd);
 575                 cfs_spin_unlock(&obd_dev_lock);
 576                 rc2 = obd_set_info_async(obd->obd_self_export,
 577                                          sizeof(KEY_SPTLRPC_CONF),
 578                                          KEY_SPTLRPC_CONF, 0, NULL, NULL);
 579                 rc = rc ? rc : rc2;
 580                 class_decref(obd, __FUNCTION__, obd);
 581                 cfs_spin_lock(&obd_dev_lock);
 582         }
 583         cfs_spin_unlock(&obd_dev_lock);
 584         return rc;
 585 }
 586 EXPORT_SYMBOL(class_notify_sptlrpc_conf);
 587
 588 void obd_cleanup_caches(void)
 589 {
 590         int rc;
 591
 592         ENTRY;
 593         if (obd_device_cachep) {
 594                 rc = cfs_mem_cache_destroy(obd_device_cachep);
 595                 LASSERTF(rc == 0, "Cannot destropy ll_obd_device_cache: rc %d\n", rc);
 596                 obd_device_cachep = NULL;
 597         }
 598         if (obdo_cachep) {
 599                 rc = cfs_mem_cache_destroy(obdo_cachep);
 600                 LASSERTF(rc == 0, "Cannot destory ll_obdo_cache\n");
 601                 obdo_cachep = NULL;
 602         }
 603         if (import_cachep) {
 604                 rc = cfs_mem_cache_destroy(import_cachep);
 605                 LASSERTF(rc == 0, "Cannot destory ll_import_cache\n");
 606                 import_cachep = NULL;
 607         }
 608         if (capa_cachep) {
 609                 rc = cfs_mem_cache_destroy(capa_cachep);
 610                 LASSERTF(rc == 0, "Cannot destory capa_cache\n");
 611                 capa_cachep = NULL;
 612         }
 613         EXIT;
 614 }
 615
 616 int obd_init_caches(void)
 617 {
 618         ENTRY;
 619
 620         LASSERT(obd_device_cachep == NULL);
 621         obd_device_cachep = cfs_mem_cache_create("ll_obd_dev_cache",
 622                                                  sizeof(struct obd_device),
 623                                                  0, 0);
 624         if (!obd_device_cachep)
 625                 GOTO(out, -ENOMEM);
 626
 627         LASSERT(obdo_cachep == NULL);
 628         obdo_cachep = cfs_mem_cache_create("ll_obdo_cache", sizeof(struct obdo),
 629                                            0, 0);
 630         if (!obdo_cachep)
 631                 GOTO(out, -ENOMEM);
 632
 633         LASSERT(import_cachep == NULL);
 634         import_cachep = cfs_mem_cache_create("ll_import_cache",
 635                                              sizeof(struct obd_import),
 636                                              0, 0);
 637         if (!import_cachep)
 638                 GOTO(out, -ENOMEM);
 639
 640         LASSERT(capa_cachep == NULL);
 641         capa_cachep = cfs_mem_cache_create("capa_cache",
 642                                            sizeof(struct obd_capa), 0, 0);
 643         if (!capa_cachep)
 644                 GOTO(out, -ENOMEM);
 645
 646         RETURN(0);
 647  out:
 648         obd_cleanup_caches();
 649         RETURN(-ENOMEM);
 650
 651 }
 652
 653 /* map connection to client */
 654 struct obd_export *class_conn2export(struct lustre_handle *conn)
 655 {
 656         struct obd_export *export;
 657         ENTRY;
 658
 659         if (!conn) {
 660                 CDEBUG(D_CACHE, "looking for null handle\n");
 661                 RETURN(NULL);
 662         }
 663
 664         if (conn->cookie == -1) {  /* this means assign a new connection */
 665                 CDEBUG(D_CACHE, "want a new connection\n");
 666                 RETURN(NULL);
 667         }
 668
 669         CDEBUG(D_INFO, "looking for export cookie "LPX64"\n", conn->cookie);
 670         export = class_handle2object(conn->cookie);
 671         RETURN(export);
 672 }
 673
 674 struct obd_device *class_exp2obd(struct obd_export *exp)
 675 {
 676         if (exp)
 677                 return exp->exp_obd;
 678         return NULL;
 679 }
 680
 681 struct obd_device *class_conn2obd(struct lustre_handle *conn)
 682 {
 683         struct obd_export *export;
 684         export = class_conn2export(conn);
 685         if (export) {
 686                 struct obd_device *obd = export->exp_obd;
 687                 class_export_put(export);
 688                 return obd;
 689         }
 690         return NULL;
 691 }
 692
 693 struct obd_import *class_exp2cliimp(struct obd_export *exp)
 694 {
 695         struct obd_device *obd = exp->exp_obd;
 696         if (obd == NULL)
 697                 return NULL;
 698         return obd->u.cli.cl_import;
 699 }
 700
 701 struct obd_import *class_conn2cliimp(struct lustre_handle *conn)
 702 {
 703         struct obd_device *obd = class_conn2obd(conn);
 704         if (obd == NULL)
 705                 return NULL;
 706         return obd->u.cli.cl_import;
 707 }
 708
 709 /* Export management functions */
 710 static void class_export_destroy(struct obd_export *exp)
 711 {
 712         struct obd_device *obd = exp->exp_obd;
 713         ENTRY;
 714
 715         LASSERT (cfs_atomic_read(&exp->exp_refcount) == 0);
 716
 717         CDEBUG(D_IOCTL, "destroying export %p/%s for %s\n", exp,
 718                exp->exp_client_uuid.uuid, obd->obd_name);
 719
 720         LASSERT(obd != NULL);
 721
 722         /* "Local" exports (lctl, LOV->{mdc,osc}) have no connection. */
 723         if (exp->exp_connection)
 724                 ptlrpc_put_connection_superhack(exp->exp_connection);
 725
 726         LASSERT(cfs_list_empty(&exp->exp_outstanding_replies));
 727         LASSERT(cfs_list_empty(&exp->exp_uncommitted_replies));
 728         LASSERT(cfs_list_empty(&exp->exp_req_replay_queue));
 729         LASSERT(cfs_list_empty(&exp->exp_queued_rpc));
 730         obd_destroy_export(exp);
 731         class_decref(obd, "export", exp);
 732
 733         OBD_FREE_RCU(exp, sizeof(*exp), &exp->exp_handle);
 734         EXIT;
 735 }
 736
 737 static void export_handle_addref(void *export)
 738 {
 739         class_export_get(export);
 740 }
 741
 742 struct obd_export *class_export_get(struct obd_export *exp)
 743 {
 744         cfs_atomic_inc(&exp->exp_refcount);
 745         CDEBUG(D_INFO, "GETting export %p : new refcount %d\n", exp,
 746                cfs_atomic_read(&exp->exp_refcount));
 747         return exp;
 748 }
 749 EXPORT_SYMBOL(class_export_get);
 750
 751 void class_export_put(struct obd_export *exp)
 752 {
 753         LASSERT(exp != NULL);
 754         CDEBUG(D_INFO, "PUTting export %p : new refcount %d\n", exp,
 755                cfs_atomic_read(&exp->exp_refcount) - 1);
 756         LASSERT(cfs_atomic_read(&exp->exp_refcount) > 0);
 757         LASSERT(cfs_atomic_read(&exp->exp_refcount) < 0x5a5a5a);
 758
 759         if (cfs_atomic_dec_and_test(&exp->exp_refcount)) {
 760                 LASSERT(!cfs_list_empty(&exp->exp_obd_chain));
 761                 CDEBUG(D_IOCTL, "final put %p/%s\n",
 762                        exp, exp->exp_client_uuid.uuid);
 763                 obd_zombie_export_add(exp);
 764         }
 765 }
 766 EXPORT_SYMBOL(class_export_put);
 767
 768 /* Creates a new export, adds it to the hash table, and returns a
 769  * pointer to it. The refcount is 2: one for the hash reference, and
 770  * one for the pointer returned by this function. */
 771 struct obd_export *class_new_export(struct obd_device *obd,
 772                                     struct obd_uuid *cluuid)
 773 {
 774         struct obd_export *export;
 775         cfs_hash_t *hash = NULL;
 776         int rc = 0;
 777         ENTRY;
 778
 779         OBD_ALLOC_PTR(export);
 780         if (!export)
 781                 return ERR_PTR(-ENOMEM);
 782
 783         export->exp_conn_cnt = 0;
 784         export->exp_lock_hash = NULL;
 785         cfs_atomic_set(&export->exp_refcount, 2);
 786         cfs_atomic_set(&export->exp_rpc_count, 0);
 787         cfs_atomic_set(&export->exp_cb_count, 0);
 788         cfs_atomic_set(&export->exp_locks_count, 0);
 789 #if LUSTRE_TRACKS_LOCK_EXP_REFS
 790         CFS_INIT_LIST_HEAD(&export->exp_locks_list);
 791         cfs_spin_lock_init(&export->exp_locks_list_guard);
 792 #endif
 793         cfs_atomic_set(&export->exp_replay_count, 0);
 794         export->exp_obd = obd;
 795         CFS_INIT_LIST_HEAD(&export->exp_outstanding_replies);
 796         cfs_spin_lock_init(&export->exp_uncommitted_replies_lock);
 797         CFS_INIT_LIST_HEAD(&export->exp_uncommitted_replies);
 798         CFS_INIT_LIST_HEAD(&export->exp_req_replay_queue);
 799         CFS_INIT_LIST_HEAD(&export->exp_handle.h_link);
 800         CFS_INIT_LIST_HEAD(&export->exp_queued_rpc);
 801         class_handle_hash(&export->exp_handle, export_handle_addref);
 802         export->exp_last_request_time = cfs_time_current_sec();
 803         cfs_spin_lock_init(&export->exp_lock);
 804         CFS_INIT_HLIST_NODE(&export->exp_uuid_hash);
 805         CFS_INIT_HLIST_NODE(&export->exp_nid_hash);
 806
 807         export->exp_sp_peer = LUSTRE_SP_ANY;
 808         export->exp_flvr.sf_rpc = SPTLRPC_FLVR_INVALID;
 809         export->exp_client_uuid = *cluuid;
 810         obd_init_export(export);
 811
 812         cfs_spin_lock(&obd->obd_dev_lock);
 813          /* shouldn't happen, but might race */
 814         if (obd->obd_stopping)
 815                 GOTO(exit_unlock, rc = -ENODEV);
 816
 817         hash = cfs_hash_getref(obd->obd_uuid_hash);
 818         if (hash == NULL)
 819                 GOTO(exit_unlock, rc = -ENODEV);
 820         cfs_spin_unlock(&obd->obd_dev_lock);
 821
 822         if (!obd_uuid_equals(cluuid, &obd->obd_uuid)) {
 823                 rc = cfs_hash_add_unique(hash, cluuid, &export->exp_uuid_hash);
 824                 if (rc != 0) {
 825                         LCONSOLE_WARN("%s: denying duplicate export for %s, %d\n",
 826                                       obd->obd_name, cluuid->uuid, rc);
 827                         GOTO(exit_err, rc = -EALREADY);
 828                 }
 829         }
 830
 831         cfs_spin_lock(&obd->obd_dev_lock);
 832         if (obd->obd_stopping) {
 833                 cfs_hash_del(hash, cluuid, &export->exp_uuid_hash);
 834                 GOTO(exit_unlock, rc = -ENODEV);
 835         }
 836
 837         class_incref(obd, "export", export);
 838         cfs_list_add(&export->exp_obd_chain, &export->exp_obd->obd_exports);
 839         cfs_list_add_tail(&export->exp_obd_chain_timed,
 840                           &export->exp_obd->obd_exports_timed);
 841         export->exp_obd->obd_num_exports++;
 842         cfs_spin_unlock(&obd->obd_dev_lock);
 843         cfs_hash_putref(hash);
 844         RETURN(export);
 845
 846 exit_unlock:
 847         cfs_spin_unlock(&obd->obd_dev_lock);
 848 exit_err:
 849         if (hash)
 850                 cfs_hash_putref(hash);
 851         class_handle_unhash(&export->exp_handle);
 852         LASSERT(cfs_hlist_unhashed(&export->exp_uuid_hash));
 853         obd_destroy_export(export);
 854         OBD_FREE_PTR(export);
 855         return ERR_PTR(rc);
 856 }
 857 EXPORT_SYMBOL(class_new_export);
 858
 859 void class_unlink_export(struct obd_export *exp)
 860 {
 861         class_handle_unhash(&exp->exp_handle);
 862
 863         cfs_spin_lock(&exp->exp_obd->obd_dev_lock);
 864         /* delete an uuid-export hashitem from hashtables */
 865         if (!cfs_hlist_unhashed(&exp->exp_uuid_hash))
 866                 cfs_hash_del(exp->exp_obd->obd_uuid_hash,
 867                              &exp->exp_client_uuid,
 868                              &exp->exp_uuid_hash);
 869
 870         cfs_list_move(&exp->exp_obd_chain, &exp->exp_obd->obd_unlinked_exports);
 871         cfs_list_del_init(&exp->exp_obd_chain_timed);
 872         exp->exp_obd->obd_num_exports--;
 873         cfs_spin_unlock(&exp->exp_obd->obd_dev_lock);
 874         class_export_put(exp);
 875 }
 876 EXPORT_SYMBOL(class_unlink_export);
 877
 878 /* Import management functions */
 879 void class_import_destroy(struct obd_import *imp)
 880 {
 881         ENTRY;
 882
 883         CDEBUG(D_IOCTL, "destroying import %p for %s\n", imp,
 884                 imp->imp_obd->obd_name);
 885
 886         LASSERT(cfs_atomic_read(&imp->imp_refcount) == 0);
 887
 888         ptlrpc_put_connection_superhack(imp->imp_connection);
 889
 890         while (!cfs_list_empty(&imp->imp_conn_list)) {
 891                 struct obd_import_conn *imp_conn;
 892
 893                 imp_conn = cfs_list_entry(imp->imp_conn_list.next,
 894                                           struct obd_import_conn, oic_item);
 895                 cfs_list_del_init(&imp_conn->oic_item);
 896                 ptlrpc_put_connection_superhack(imp_conn->oic_conn);
 897                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 898         }
 899
 900         LASSERT(imp->imp_sec == NULL);
 901         class_decref(imp->imp_obd, "import", imp);
 902         OBD_FREE_RCU(imp, sizeof(*imp), &imp->imp_handle);
 903         EXIT;
 904 }
 905
 906 static void import_handle_addref(void *import)
 907 {
 908         class_import_get(import);
 909 }
 910
 911 struct obd_import *class_import_get(struct obd_import *import)
 912 {
 913         LASSERT(cfs_atomic_read(&import->imp_refcount) >= 0);
 914         LASSERT(cfs_atomic_read(&import->imp_refcount) < 0x5a5a5a);
 915         cfs_atomic_inc(&import->imp_refcount);
 916         CDEBUG(D_INFO, "import %p refcount=%d obd=%s\n", import,
 917                cfs_atomic_read(&import->imp_refcount),
 918                import->imp_obd->obd_name);
 919         return import;
 920 }
 921 EXPORT_SYMBOL(class_import_get);
 922
 923 void class_import_put(struct obd_import *imp)
 924 {
 925         ENTRY;
 926
 927         LASSERT(cfs_atomic_read(&imp->imp_refcount) > 0);
 928         LASSERT(cfs_atomic_read(&imp->imp_refcount) < 0x5a5a5a);
 929         LASSERT(cfs_list_empty(&imp->imp_zombie_chain));
 930
 931         CDEBUG(D_INFO, "import %p refcount=%d obd=%s\n", imp,
 932                cfs_atomic_read(&imp->imp_refcount) - 1,
 933                imp->imp_obd->obd_name);
 934
 935         if (cfs_atomic_dec_and_test(&imp->imp_refcount)) {
 936                 CDEBUG(D_INFO, "final put import %p\n", imp);
 937                 obd_zombie_import_add(imp);
 938         }
 939
 940         EXIT;
 941 }
 942 EXPORT_SYMBOL(class_import_put);
 943
 944 static void init_imp_at(struct imp_at *at) {
 945         int i;
 946         at_init(&at->iat_net_latency, 0, 0);
 947         for (i = 0; i < IMP_AT_MAX_PORTALS; i++) {
 948                 /* max service estimates are tracked on the server side, so
 949                    don't use the AT history here, just use the last reported
 950                    val. (But keep hist for proc histogram, worst_ever) */
 951                 at_init(&at->iat_service_estimate[i], INITIAL_CONNECT_TIMEOUT,
 952                         AT_FLG_NOHIST);
 953         }
 954 }
 955
 956 struct obd_import *class_new_import(struct obd_device *obd)
 957 {
 958         struct obd_import *imp;
 959
 960         OBD_ALLOC(imp, sizeof(*imp));
 961         if (imp == NULL)
 962                 return NULL;
 963
 964         CFS_INIT_LIST_HEAD(&imp->imp_zombie_chain);
 965         CFS_INIT_LIST_HEAD(&imp->imp_replay_list);
 966         CFS_INIT_LIST_HEAD(&imp->imp_sending_list);
 967         CFS_INIT_LIST_HEAD(&imp->imp_delayed_list);
 968         cfs_spin_lock_init(&imp->imp_lock);
 969         imp->imp_last_success_conn = 0;
 970         imp->imp_state = LUSTRE_IMP_NEW;
 971         imp->imp_obd = class_incref(obd, "import", imp);
 972         cfs_sema_init(&imp->imp_sec_mutex, 1);
 973         cfs_waitq_init(&imp->imp_recovery_waitq);
 974
 975         cfs_atomic_set(&imp->imp_refcount, 2);
 976         cfs_atomic_set(&imp->imp_unregistering, 0);
 977         cfs_atomic_set(&imp->imp_inflight, 0);
 978         cfs_atomic_set(&imp->imp_replay_inflight, 0);
 979         cfs_atomic_set(&imp->imp_inval_count, 0);
 980         CFS_INIT_LIST_HEAD(&imp->imp_conn_list);
 981         CFS_INIT_LIST_HEAD(&imp->imp_handle.h_link);
 982         class_handle_hash(&imp->imp_handle, import_handle_addref);
 983         init_imp_at(&imp->imp_at);
 984
 985         /* the default magic is V2, will be used in connect RPC, and
 986          * then adjusted according to the flags in request/reply. */
 987         imp->imp_msg_magic = LUSTRE_MSG_MAGIC_V2;
 988
 989         return imp;
 990 }
 991 EXPORT_SYMBOL(class_new_import);
 992
 993 void class_destroy_import(struct obd_import *import)
 994 {
 995         LASSERT(import != NULL);
 996         LASSERT(import != LP_POISON);
 997
 998         class_handle_unhash(&import->imp_handle);
 999
1000         cfs_spin_lock(&import->imp_lock);
1001         import->imp_generation++;
1002         cfs_spin_unlock(&import->imp_lock);
1003         class_import_put(import);
1004 }
1005 EXPORT_SYMBOL(class_destroy_import);
1006
1007 #if LUSTRE_TRACKS_LOCK_EXP_REFS
1008
1009 void __class_export_add_lock_ref(struct obd_export *exp, struct ldlm_lock *lock)
1010 {
1011         cfs_spin_lock(&exp->exp_locks_list_guard);
1012
1013         LASSERT(lock->l_exp_refs_nr >= 0);
1014
1015         if (lock->l_exp_refs_target != NULL &&
1016             lock->l_exp_refs_target != exp) {
1017                 LCONSOLE_WARN("setting export %p for lock %p which already has export %p\n",
1018                               exp, lock, lock->l_exp_refs_target);
1019         }
1020         if ((lock->l_exp_refs_nr ++) == 0) {
1021                 cfs_list_add(&lock->l_exp_refs_link, &exp->exp_locks_list);
1022                 lock->l_exp_refs_target = exp;
1023         }
1024         CDEBUG(D_INFO, "lock = %p, export = %p, refs = %u\n",
1025                lock, exp, lock->l_exp_refs_nr);
1026         cfs_spin_unlock(&exp->exp_locks_list_guard);
1027 }
1028 EXPORT_SYMBOL(__class_export_add_lock_ref);
1029
1030 void __class_export_del_lock_ref(struct obd_export *exp, struct ldlm_lock *lock)
1031 {
1032         cfs_spin_lock(&exp->exp_locks_list_guard);
1033         LASSERT(lock->l_exp_refs_nr > 0);
1034         if (lock->l_exp_refs_target != exp) {
1035                 LCONSOLE_WARN("lock %p, "
1036                               "mismatching export pointers: %p, %p\n",
1037                               lock, lock->l_exp_refs_target, exp);
1038         }
1039         if (-- lock->l_exp_refs_nr == 0) {
1040                 cfs_list_del_init(&lock->l_exp_refs_link);
1041                 lock->l_exp_refs_target = NULL;
1042         }
1043         CDEBUG(D_INFO, "lock = %p, export = %p, refs = %u\n",
1044                lock, exp, lock->l_exp_refs_nr);
1045         cfs_spin_unlock(&exp->exp_locks_list_guard);
1046 }
1047 EXPORT_SYMBOL(__class_export_del_lock_ref);
1048 #endif
1049
1050 /* A connection defines an export context in which preallocation can
1051    be managed. This releases the export pointer reference, and returns
1052    the export handle, so the export refcount is 1 when this function
1053    returns. */
1054 int class_connect(struct lustre_handle *conn, struct obd_device *obd,
1055                   struct obd_uuid *cluuid)
1056 {
1057         struct obd_export *export;
1058         LASSERT(conn != NULL);
1059         LASSERT(obd != NULL);
1060         LASSERT(cluuid != NULL);
1061         ENTRY;
1062
1063         export = class_new_export(obd, cluuid);
1064         if (IS_ERR(export))
1065                 RETURN(PTR_ERR(export));
1066
1067         conn->cookie = export->exp_handle.h_cookie;
1068         class_export_put(export);
1069
1070         CDEBUG(D_IOCTL, "connect: client %s, cookie "LPX64"\n",
1071                cluuid->uuid, conn->cookie);
1072         RETURN(0);
1073 }
1074 EXPORT_SYMBOL(class_connect);
1075
1076 /* if export is involved in recovery then clean up related things */
1077 void class_export_recovery_cleanup(struct obd_export *exp)
1078 {
1079         struct obd_device *obd = exp->exp_obd;
1080
1081         cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1082         if (exp->exp_delayed)
1083                 obd->obd_delayed_clients--;
1084         if (obd->obd_recovering && exp->exp_in_recovery) {
1085                 cfs_spin_lock(&exp->exp_lock);
1086                 exp->exp_in_recovery = 0;
1087                 cfs_spin_unlock(&exp->exp_lock);
1088                 LASSERT(obd->obd_connected_clients);
1089                 obd->obd_connected_clients--;
1090         }
1091         /** Cleanup req replay fields */
1092         if (exp->exp_req_replay_needed) {
1093                 cfs_spin_lock(&exp->exp_lock);
1094                 exp->exp_req_replay_needed = 0;
1095                 cfs_spin_unlock(&exp->exp_lock);
1096                 LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients));
1097                 cfs_atomic_dec(&obd->obd_req_replay_clients);
1098         }
1099         /** Cleanup lock replay data */
1100         if (exp->exp_lock_replay_needed) {
1101                 cfs_spin_lock(&exp->exp_lock);
1102                 exp->exp_lock_replay_needed = 0;
1103                 cfs_spin_unlock(&exp->exp_lock);
1104                 LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients));
1105                 cfs_atomic_dec(&obd->obd_lock_replay_clients);
1106         }
1107         cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1108 }
1109
1110 /* This function removes 1-3 references from the export:
1111  * 1 - for export pointer passed
1112  * and if disconnect really need
1113  * 2 - removing from hash
1114  * 3 - in client_unlink_export
1115  * The export pointer passed to this function can destroyed */
1116 int class_disconnect(struct obd_export *export)
1117 {
1118         int already_disconnected;
1119         ENTRY;
1120
1121         if (export == NULL) {
1122                 fixme();
1123                 CDEBUG(D_IOCTL, "attempting to free NULL export %p\n", export);
1124                 RETURN(-EINVAL);
1125         }
1126
1127         cfs_spin_lock(&export->exp_lock);
1128         already_disconnected = export->exp_disconnected;
1129         export->exp_disconnected = 1;
1130         cfs_spin_unlock(&export->exp_lock);
1131
1132         /* class_cleanup(), abort_recovery(), and class_fail_export()
1133          * all end up in here, and if any of them race we shouldn't
1134          * call extra class_export_puts(). */
1135         if (already_disconnected) {
1136                 LASSERT(cfs_hlist_unhashed(&export->exp_nid_hash));
1137                 GOTO(no_disconn, already_disconnected);
1138         }
1139
1140         CDEBUG(D_IOCTL, "disconnect: cookie "LPX64"\n",
1141                export->exp_handle.h_cookie);
1142
1143         if (!cfs_hlist_unhashed(&export->exp_nid_hash))
1144                 cfs_hash_del(export->exp_obd->obd_nid_hash,
1145                              &export->exp_connection->c_peer.nid,
1146                              &export->exp_nid_hash);
1147
1148         class_export_recovery_cleanup(export);
1149         class_unlink_export(export);
1150 no_disconn:
1151         class_export_put(export);
1152         RETURN(0);
1153 }
1154
1155 /* Return non-zero for a fully connected export */
1156 int class_connected_export(struct obd_export *exp)
1157 {
1158         if (exp) {
1159                 int connected;
1160                 cfs_spin_lock(&exp->exp_lock);
1161                 connected = (exp->exp_conn_cnt > 0);
1162                 cfs_spin_unlock(&exp->exp_lock);
1163                 return connected;
1164         }
1165         return 0;
1166 }
1167 EXPORT_SYMBOL(class_connected_export);
1168
1169 static void class_disconnect_export_list(cfs_list_t *list,
1170                                          enum obd_option flags)
1171 {
1172         int rc;
1173         struct obd_export *exp;
1174         ENTRY;
1175
1176         /* It's possible that an export may disconnect itself, but
1177          * nothing else will be added to this list. */
1178         while (!cfs_list_empty(list)) {
1179                 exp = cfs_list_entry(list->next, struct obd_export,
1180                                      exp_obd_chain);
1181                 /* need for safe call CDEBUG after obd_disconnect */
1182                 class_export_get(exp);
1183
1184                 cfs_spin_lock(&exp->exp_lock);
1185                 exp->exp_flags = flags;
1186                 cfs_spin_unlock(&exp->exp_lock);
1187
1188                 if (obd_uuid_equals(&exp->exp_client_uuid,
1189                                     &exp->exp_obd->obd_uuid)) {
1190                         CDEBUG(D_HA,
1191                                "exp %p export uuid == obd uuid, don't discon\n",
1192                                exp);
1193                         /* Need to delete this now so we don't end up pointing
1194                          * to work_list later when this export is cleaned up. */
1195                         cfs_list_del_init(&exp->exp_obd_chain);
1196                         class_export_put(exp);
1197                         continue;
1198                 }
1199
1200                 class_export_get(exp);
1201                 CDEBUG(D_HA, "%s: disconnecting export at %s (%p), "
1202                        "last request at "CFS_TIME_T"\n",
1203                        exp->exp_obd->obd_name, obd_export_nid2str(exp),
1204                        exp, exp->exp_last_request_time);
1205                 /* release one export reference anyway */
1206                 rc = obd_disconnect(exp);
1207
1208                 CDEBUG(D_HA, "disconnected export at %s (%p): rc %d\n",
1209                        obd_export_nid2str(exp), exp, rc);
1210                 class_export_put(exp);
1211         }
1212         EXIT;
1213 }
1214
1215 void class_disconnect_exports(struct obd_device *obd)
1216 {
1217         cfs_list_t work_list;
1218         ENTRY;
1219
1220         /* Move all of the exports from obd_exports to a work list, en masse. */
1221         CFS_INIT_LIST_HEAD(&work_list);
1222         cfs_spin_lock(&obd->obd_dev_lock);
1223         cfs_list_splice_init(&obd->obd_exports, &work_list);
1224         cfs_list_splice_init(&obd->obd_delayed_exports, &work_list);
1225         cfs_spin_unlock(&obd->obd_dev_lock);
1226
1227         if (!cfs_list_empty(&work_list)) {
1228                 CDEBUG(D_HA, "OBD device %d (%p) has exports, "
1229                        "disconnecting them\n", obd->obd_minor, obd);
1230                 class_disconnect_export_list(&work_list,
1231                                              exp_flags_from_obd(obd));
1232         } else
1233                 CDEBUG(D_HA, "OBD device %d (%p) has no exports\n",
1234                        obd->obd_minor, obd);
1235         EXIT;
1236 }
1237 EXPORT_SYMBOL(class_disconnect_exports);
1238
1239 /* Remove exports that have not completed recovery.
1240  */
1241 void class_disconnect_stale_exports(struct obd_device *obd,
1242                                     int (*test_export)(struct obd_export *))
1243 {
1244         cfs_list_t work_list;
1245         cfs_list_t *pos, *n;
1246         struct obd_export *exp;
1247         int evicted = 0;
1248         ENTRY;
1249
1250         CFS_INIT_LIST_HEAD(&work_list);
1251         cfs_spin_lock(&obd->obd_dev_lock);
1252         cfs_list_for_each_safe(pos, n, &obd->obd_exports) {
1253                 exp = cfs_list_entry(pos, struct obd_export, exp_obd_chain);
1254                 if (test_export(exp))
1255                         continue;
1256
1257                 /* don't count self-export as client */
1258                 if (obd_uuid_equals(&exp->exp_client_uuid,
1259                                     &exp->exp_obd->obd_uuid))
1260                         continue;
1261
1262                 cfs_list_move(&exp->exp_obd_chain, &work_list);
1263                 evicted++;
1264                 CDEBUG(D_ERROR, "%s: disconnect stale client %s@%s\n",
1265                        obd->obd_name, exp->exp_client_uuid.uuid,
1266                        exp->exp_connection == NULL ? "<unknown>" :
1267                        libcfs_nid2str(exp->exp_connection->c_peer.nid));
1268                 print_export_data(exp, "EVICTING", 0);
1269         }
1270         cfs_spin_unlock(&obd->obd_dev_lock);
1271
1272         if (evicted) {
1273                 CDEBUG(D_HA, "%s: disconnecting %d stale clients\n",
1274                        obd->obd_name, evicted);
1275                 obd->obd_stale_clients += evicted;
1276         }
1277         class_disconnect_export_list(&work_list, exp_flags_from_obd(obd) |
1278                                                  OBD_OPT_ABORT_RECOV);
1279         EXIT;
1280 }
1281 EXPORT_SYMBOL(class_disconnect_stale_exports);
1282
1283 void class_fail_export(struct obd_export *exp)
1284 {
1285         int rc, already_failed;
1286
1287         cfs_spin_lock(&exp->exp_lock);
1288         already_failed = exp->exp_failed;
1289         exp->exp_failed = 1;
1290         cfs_spin_unlock(&exp->exp_lock);
1291
1292         if (already_failed) {
1293                 CDEBUG(D_HA, "disconnecting dead export %p/%s; skipping\n",
1294                        exp, exp->exp_client_uuid.uuid);
1295                 return;
1296         }
1297
1298         CDEBUG(D_HA, "disconnecting export %p/%s\n",
1299                exp, exp->exp_client_uuid.uuid);
1300
1301         if (obd_dump_on_timeout)
1302                 libcfs_debug_dumplog();
1303
1304         /* Most callers into obd_disconnect are removing their own reference
1305          * (request, for example) in addition to the one from the hash table.
1306          * We don't have such a reference here, so make one. */
1307         class_export_get(exp);
1308         rc = obd_disconnect(exp);
1309         if (rc)
1310                 CERROR("disconnecting export %p failed: %d\n", exp, rc);
1311         else
1312                 CDEBUG(D_HA, "disconnected export %p/%s\n",
1313                        exp, exp->exp_client_uuid.uuid);
1314 }
1315 EXPORT_SYMBOL(class_fail_export);
1316
1317 char *obd_export_nid2str(struct obd_export *exp)
1318 {
1319         if (exp->exp_connection != NULL)
1320                 return libcfs_nid2str(exp->exp_connection->c_peer.nid);
1321
1322         return "(no nid)";
1323 }
1324 EXPORT_SYMBOL(obd_export_nid2str);
1325
1326 int obd_export_evict_by_nid(struct obd_device *obd, const char *nid)
1327 {
1328         struct obd_export *doomed_exp = NULL;
1329         int exports_evicted = 0;
1330
1331         lnet_nid_t nid_key = libcfs_str2nid((char *)nid);
1332
1333         do {
1334                 doomed_exp = cfs_hash_lookup(obd->obd_nid_hash, &nid_key);
1335                 if (doomed_exp == NULL)
1336                         break;
1337
1338                 LASSERTF(doomed_exp->exp_connection->c_peer.nid == nid_key,
1339                          "nid %s found, wanted nid %s, requested nid %s\n",
1340                          obd_export_nid2str(doomed_exp),
1341                          libcfs_nid2str(nid_key), nid);
1342                 LASSERTF(doomed_exp != obd->obd_self_export,
1343                          "self-export is hashed by NID?\n");
1344                 exports_evicted++;
1345                 CWARN("%s: evict NID '%s' (%s) #%d at adminstrative request\n",
1346                        obd->obd_name, nid, doomed_exp->exp_client_uuid.uuid,
1347                        exports_evicted);
1348                 class_fail_export(doomed_exp);
1349                 class_export_put(doomed_exp);
1350         } while (1);
1351
1352         if (!exports_evicted)
1353                 CDEBUG(D_HA,"%s: can't disconnect NID '%s': no exports found\n",
1354                        obd->obd_name, nid);
1355         return exports_evicted;
1356 }
1357 EXPORT_SYMBOL(obd_export_evict_by_nid);
1358
1359 int obd_export_evict_by_uuid(struct obd_device *obd, const char *uuid)
1360 {
1361         struct obd_export *doomed_exp = NULL;
1362         struct obd_uuid doomed_uuid;
1363         int exports_evicted = 0;
1364
1365         obd_str2uuid(&doomed_uuid, uuid);
1366         if (obd_uuid_equals(&doomed_uuid, &obd->obd_uuid)) {
1367                 CERROR("%s: can't evict myself\n", obd->obd_name);
1368                 return exports_evicted;
1369         }
1370
1371         doomed_exp = cfs_hash_lookup(obd->obd_uuid_hash, &doomed_uuid);
1372
1373         if (doomed_exp == NULL) {
1374                 CERROR("%s: can't disconnect %s: no exports found\n",
1375                        obd->obd_name, uuid);
1376         } else {
1377                 CWARN("%s: evicting %s at adminstrative request\n",
1378                        obd->obd_name, doomed_exp->exp_client_uuid.uuid);
1379                 class_fail_export(doomed_exp);
1380                 class_export_put(doomed_exp);
1381                 exports_evicted++;
1382         }
1383
1384         return exports_evicted;
1385 }
1386 EXPORT_SYMBOL(obd_export_evict_by_uuid);
1387
1388 #if LUSTRE_TRACKS_LOCK_EXP_REFS
1389 void (*class_export_dump_hook)(struct obd_export*) = NULL;
1390 EXPORT_SYMBOL(class_export_dump_hook);
1391 #endif
1392
1393 static void print_export_data(struct obd_export *exp, const char *status,
1394                               int locks)
1395 {
1396         struct ptlrpc_reply_state *rs;
1397         struct ptlrpc_reply_state *first_reply = NULL;
1398         int nreplies = 0;
1399
1400         cfs_spin_lock(&exp->exp_lock);
1401         cfs_list_for_each_entry(rs, &exp->exp_outstanding_replies,
1402                                 rs_exp_list) {
1403                 if (nreplies == 0)
1404                         first_reply = rs;
1405                 nreplies++;
1406         }
1407         cfs_spin_unlock(&exp->exp_lock);
1408
1409         CDEBUG(D_HA, "%s: %s %p %s %s %d (%d %d %d) %d %d %d %d: %p %s "LPU64"\n",
1410                exp->exp_obd->obd_name, status, exp, exp->exp_client_uuid.uuid,
1411                obd_export_nid2str(exp), cfs_atomic_read(&exp->exp_refcount),
1412                cfs_atomic_read(&exp->exp_rpc_count),
1413                cfs_atomic_read(&exp->exp_cb_count),
1414                cfs_atomic_read(&exp->exp_locks_count),
1415                exp->exp_disconnected, exp->exp_delayed, exp->exp_failed,
1416                nreplies, first_reply, nreplies > 3 ? "..." : "",
1417                exp->exp_last_committed);
1418 #if LUSTRE_TRACKS_LOCK_EXP_REFS
1419         if (locks && class_export_dump_hook != NULL)
1420                 class_export_dump_hook(exp);
1421 #endif
1422 }
1423
1424 void dump_exports(struct obd_device *obd, int locks)
1425 {
1426         struct obd_export *exp;
1427
1428         cfs_spin_lock(&obd->obd_dev_lock);
1429         cfs_list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain)
1430                 print_export_data(exp, "ACTIVE", locks);
1431         cfs_list_for_each_entry(exp, &obd->obd_unlinked_exports, exp_obd_chain)
1432                 print_export_data(exp, "UNLINKED", locks);
1433         cfs_list_for_each_entry(exp, &obd->obd_delayed_exports, exp_obd_chain)
1434                 print_export_data(exp, "DELAYED", locks);
1435         cfs_spin_unlock(&obd->obd_dev_lock);
1436         cfs_spin_lock(&obd_zombie_impexp_lock);
1437         cfs_list_for_each_entry(exp, &obd_zombie_exports, exp_obd_chain)
1438                 print_export_data(exp, "ZOMBIE", locks);
1439         cfs_spin_unlock(&obd_zombie_impexp_lock);
1440 }
1441 EXPORT_SYMBOL(dump_exports);
1442
1443 void obd_exports_barrier(struct obd_device *obd)
1444 {
1445         int waited = 2;
1446         LASSERT(cfs_list_empty(&obd->obd_exports));
1447         cfs_spin_lock(&obd->obd_dev_lock);
1448         while (!cfs_list_empty(&obd->obd_unlinked_exports)) {
1449                 cfs_spin_unlock(&obd->obd_dev_lock);
1450                 cfs_schedule_timeout_and_set_state(CFS_TASK_UNINT,
1451                                                    cfs_time_seconds(waited));
1452                 if (waited > 5 && IS_PO2(waited)) {
1453                         LCONSOLE_WARN("%s is waiting for obd_unlinked_exports "
1454                                       "more than %d seconds. "
1455                                       "The obd refcount = %d. Is it stuck?\n",
1456                                       obd->obd_name, waited,
1457                                       cfs_atomic_read(&obd->obd_refcount));
1458                         dump_exports(obd, 1);
1459                 }
1460                 waited *= 2;
1461                 cfs_spin_lock(&obd->obd_dev_lock);
1462         }
1463         cfs_spin_unlock(&obd->obd_dev_lock);
1464 }
1465 EXPORT_SYMBOL(obd_exports_barrier);
1466
1467 /**
1468  * kill zombie imports and exports
1469  */
1470 void obd_zombie_impexp_cull(void)
1471 {
1472         struct obd_import *import;
1473         struct obd_export *export;
1474         ENTRY;
1475
1476         do {
1477                 cfs_spin_lock(&obd_zombie_impexp_lock);
1478
1479                 import = NULL;
1480                 if (!cfs_list_empty(&obd_zombie_imports)) {
1481                         import = cfs_list_entry(obd_zombie_imports.next,
1482                                                 struct obd_import,
1483                                                 imp_zombie_chain);
1484                         cfs_list_del_init(&import->imp_zombie_chain);
1485                 }
1486
1487                 export = NULL;
1488                 if (!cfs_list_empty(&obd_zombie_exports)) {
1489                         export = cfs_list_entry(obd_zombie_exports.next,
1490                                                 struct obd_export,
1491                                                 exp_obd_chain);
1492                         cfs_list_del_init(&export->exp_obd_chain);
1493                 }
1494
1495                 cfs_spin_unlock(&obd_zombie_impexp_lock);
1496
1497                 if (import != NULL)
1498                         class_import_destroy(import);
1499
1500                 if (export != NULL)
1501                         class_export_destroy(export);
1502
1503                 cfs_cond_resched();
1504         } while (import != NULL || export != NULL);
1505         EXIT;
1506 }
1507
1508 static cfs_completion_t         obd_zombie_start;
1509 static cfs_completion_t         obd_zombie_stop;
1510 static unsigned long            obd_zombie_flags;
1511 static cfs_waitq_t              obd_zombie_waitq;
1512 static pid_t                    obd_zombie_pid;
1513
1514 enum {
1515         OBD_ZOMBIE_STOP   = 1 << 1
1516 };
1517
1518 /**
1519  * check for work for kill zombie import/export thread.
1520  */
1521 static int obd_zombie_impexp_check(void *arg)
1522 {
1523         int rc;
1524
1525         cfs_spin_lock(&obd_zombie_impexp_lock);
1526         rc = cfs_list_empty(&obd_zombie_imports) &&
1527              cfs_list_empty(&obd_zombie_exports) &&
1528              !cfs_test_bit(OBD_ZOMBIE_STOP, &obd_zombie_flags);
1529
1530         cfs_spin_unlock(&obd_zombie_impexp_lock);
1531
1532         RETURN(rc);
1533 }
1534
1535 /**
1536  * Add export to the obd_zombe thread and notify it.
1537  */
1538 static void obd_zombie_export_add(struct obd_export *exp) {
1539         cfs_spin_lock(&exp->exp_obd->obd_dev_lock);
1540         LASSERT(!cfs_list_empty(&exp->exp_obd_chain));
1541         cfs_list_del_init(&exp->exp_obd_chain);
1542         cfs_spin_unlock(&exp->exp_obd->obd_dev_lock);
1543         cfs_spin_lock(&obd_zombie_impexp_lock);
1544         cfs_list_add(&exp->exp_obd_chain, &obd_zombie_exports);
1545         cfs_spin_unlock(&obd_zombie_impexp_lock);
1546
1547         if (obd_zombie_impexp_notify != NULL)
1548                 obd_zombie_impexp_notify();
1549 }
1550
1551 /**
1552  * Add import to the obd_zombe thread and notify it.
1553  */
1554 static void obd_zombie_import_add(struct obd_import *imp) {
1555         LASSERT(imp->imp_sec == NULL);
1556         cfs_spin_lock(&obd_zombie_impexp_lock);
1557         LASSERT(cfs_list_empty(&imp->imp_zombie_chain));
1558         cfs_list_add(&imp->imp_zombie_chain, &obd_zombie_imports);
1559         cfs_spin_unlock(&obd_zombie_impexp_lock);
1560
1561         if (obd_zombie_impexp_notify != NULL)
1562                 obd_zombie_impexp_notify();
1563 }
1564
1565 /**
1566  * notify import/export destroy thread about new zombie.
1567  */
1568 static void obd_zombie_impexp_notify(void)
1569 {
1570         cfs_waitq_signal(&obd_zombie_waitq);
1571 }
1572
1573 /**
1574  * check whether obd_zombie is idle
1575  */
1576 static int obd_zombie_is_idle(void)
1577 {
1578         int rc;
1579
1580         LASSERT(!cfs_test_bit(OBD_ZOMBIE_STOP, &obd_zombie_flags));
1581         cfs_spin_lock(&obd_zombie_impexp_lock);
1582         rc = cfs_list_empty(&obd_zombie_imports) &&
1583              cfs_list_empty(&obd_zombie_exports);
1584         cfs_spin_unlock(&obd_zombie_impexp_lock);
1585         return rc;
1586 }
1587
1588 /**
1589  * wait when obd_zombie import/export queues become empty
1590  */
1591 void obd_zombie_barrier(void)
1592 {
1593         struct l_wait_info lwi = { 0 };
1594
1595         if (obd_zombie_pid == cfs_curproc_pid())
1596                 /* don't wait for myself */
1597                 return;
1598         l_wait_event(obd_zombie_waitq, obd_zombie_is_idle(), &lwi);
1599 }
1600 EXPORT_SYMBOL(obd_zombie_barrier);
1601
1602 #ifdef __KERNEL__
1603
1604 /**
1605  * destroy zombie export/import thread.
1606  */
1607 static int obd_zombie_impexp_thread(void *unused)
1608 {
1609         int rc;
1610
1611         if ((rc = cfs_daemonize_ctxt("obd_zombid"))) {
1612                 cfs_complete(&obd_zombie_start);
1613                 RETURN(rc);
1614         }
1615
1616         cfs_complete(&obd_zombie_start);
1617
1618         obd_zombie_pid = cfs_curproc_pid();
1619
1620         while(!cfs_test_bit(OBD_ZOMBIE_STOP, &obd_zombie_flags)) {
1621                 struct l_wait_info lwi = { 0 };
1622
1623                 l_wait_event(obd_zombie_waitq,
1624                              !obd_zombie_impexp_check(NULL), &lwi);
1625                 obd_zombie_impexp_cull();
1626
1627                 /*
1628                  * Notify obd_zombie_barrier callers that queues
1629                  * may be empty.
1630                  */
1631                 cfs_waitq_signal(&obd_zombie_waitq);
1632         }
1633
1634         cfs_complete(&obd_zombie_stop);
1635
1636         RETURN(0);
1637 }
1638
1639 #else /* ! KERNEL */
1640
1641 static cfs_atomic_t zombie_recur = CFS_ATOMIC_INIT(0);
1642 static void *obd_zombie_impexp_work_cb;
1643 static void *obd_zombie_impexp_idle_cb;
1644
1645 int obd_zombie_impexp_kill(void *arg)
1646 {
1647         int rc = 0;
1648
1649         if (cfs_atomic_inc_return(&zombie_recur) == 1) {
1650                 obd_zombie_impexp_cull();
1651                 rc = 1;
1652         }
1653         cfs_atomic_dec(&zombie_recur);
1654         return rc;
1655 }
1656
1657 #endif
1658
1659 /**
1660  * start destroy zombie import/export thread
1661  */
1662 int obd_zombie_impexp_init(void)
1663 {
1664         int rc;
1665
1666         CFS_INIT_LIST_HEAD(&obd_zombie_imports);
1667         CFS_INIT_LIST_HEAD(&obd_zombie_exports);
1668         cfs_spin_lock_init(&obd_zombie_impexp_lock);
1669         cfs_init_completion(&obd_zombie_start);
1670         cfs_init_completion(&obd_zombie_stop);
1671         cfs_waitq_init(&obd_zombie_waitq);
1672         obd_zombie_pid = 0;
1673
1674 #ifdef __KERNEL__
1675         rc = cfs_kernel_thread(obd_zombie_impexp_thread, NULL, 0);
1676         if (rc < 0)
1677                 RETURN(rc);
1678
1679         cfs_wait_for_completion(&obd_zombie_start);
1680 #else
1681
1682         obd_zombie_impexp_work_cb =
1683                 liblustre_register_wait_callback("obd_zombi_impexp_kill",
1684                                                  &obd_zombie_impexp_kill, NULL);
1685
1686         obd_zombie_impexp_idle_cb =
1687                 liblustre_register_idle_callback("obd_zombi_impexp_check",
1688                                                  &obd_zombie_impexp_check, NULL);
1689         rc = 0;
1690 #endif
1691         RETURN(rc);
1692 }
1693 /**
1694  * stop destroy zombie import/export thread
1695  */
1696 void obd_zombie_impexp_stop(void)
1697 {
1698         cfs_set_bit(OBD_ZOMBIE_STOP, &obd_zombie_flags);
1699         obd_zombie_impexp_notify();
1700 #ifdef __KERNEL__
1701         cfs_wait_for_completion(&obd_zombie_stop);
1702 #else
1703         liblustre_deregister_wait_callback(obd_zombie_impexp_work_cb);
1704         liblustre_deregister_idle_callback(obd_zombie_impexp_idle_cb);
1705 #endif
1706 }
1707