lustre/obdclass/genops.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 1999, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2015, Intel Corporation.
  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
  44 #include <linux/kthread.h>
  45 #include <obd_class.h>
  46 #include <lprocfs_status.h>
  47 #include <lustre_disk.h>
  48 #include <lustre_kernelcomm.h>
  49
  50 spinlock_t obd_types_lock;
  51
  52 static struct kmem_cache *obd_device_cachep;
  53 struct kmem_cache *obdo_cachep;
  54 EXPORT_SYMBOL(obdo_cachep);
  55 static struct kmem_cache *import_cachep;
  56
  57 static struct list_head obd_zombie_imports;
  58 static struct list_head obd_zombie_exports;
  59 static spinlock_t  obd_zombie_impexp_lock;
  60
  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 struct list_head obd_stale_exports;
  68 spinlock_t       obd_stale_export_lock;
  69 atomic_t         obd_stale_export_num;
  70
  71 int (*ptlrpc_put_connection_superhack)(struct ptlrpc_connection *c);
  72 EXPORT_SYMBOL(ptlrpc_put_connection_superhack);
  73
  74 /*
  75  * support functions: we could use inter-module communication, but this
  76  * is more portable to other OS's
  77  */
  78 static struct obd_device *obd_device_alloc(void)
  79 {
  80         struct obd_device *obd;
  81
  82         OBD_SLAB_ALLOC_PTR_GFP(obd, obd_device_cachep, GFP_NOFS);
  83         if (obd != NULL) {
  84                 obd->obd_magic = OBD_DEVICE_MAGIC;
  85         }
  86         return obd;
  87 }
  88
  89 static void obd_device_free(struct obd_device *obd)
  90 {
  91         LASSERT(obd != NULL);
  92         LASSERTF(obd->obd_magic == OBD_DEVICE_MAGIC, "obd %p obd_magic %08x != %08x\n",
  93                  obd, obd->obd_magic, OBD_DEVICE_MAGIC);
  94         if (obd->obd_namespace != NULL) {
  95                 CERROR("obd %p: namespace %p was not properly cleaned up (obd_force=%d)!\n",
  96                        obd, obd->obd_namespace, obd->obd_force);
  97                 LBUG();
  98         }
  99         lu_ref_fini(&obd->obd_reference);
 100         OBD_SLAB_FREE_PTR(obd, obd_device_cachep);
 101 }
 102
 103 struct obd_type *class_search_type(const char *name)
 104 {
 105         struct list_head *tmp;
 106         struct obd_type *type;
 107
 108         spin_lock(&obd_types_lock);
 109         list_for_each(tmp, &obd_types) {
 110                 type = list_entry(tmp, struct obd_type, typ_chain);
 111                 if (strcmp(type->typ_name, name) == 0) {
 112                         spin_unlock(&obd_types_lock);
 113                         return type;
 114                 }
 115         }
 116         spin_unlock(&obd_types_lock);
 117         return NULL;
 118 }
 119 EXPORT_SYMBOL(class_search_type);
 120
 121 struct obd_type *class_get_type(const char *name)
 122 {
 123         struct obd_type *type = class_search_type(name);
 124
 125 #ifdef HAVE_MODULE_LOADING_SUPPORT
 126         if (!type) {
 127                 const char *modname = name;
 128
 129                 if (strcmp(modname, "obdfilter") == 0)
 130                         modname = "ofd";
 131
 132                 if (strcmp(modname, LUSTRE_LWP_NAME) == 0)
 133                         modname = LUSTRE_OSP_NAME;
 134
 135                 if (!strncmp(modname, LUSTRE_MDS_NAME, strlen(LUSTRE_MDS_NAME)))
 136                         modname = LUSTRE_MDT_NAME;
 137
 138                 if (!request_module("%s", modname)) {
 139                         CDEBUG(D_INFO, "Loaded module '%s'\n", modname);
 140                         type = class_search_type(name);
 141                 } else {
 142                         LCONSOLE_ERROR_MSG(0x158, "Can't load module '%s'\n",
 143                                            modname);
 144                 }
 145         }
 146 #endif
 147         if (type) {
 148                 spin_lock(&type->obd_type_lock);
 149                 type->typ_refcnt++;
 150                 try_module_get(type->typ_dt_ops->o_owner);
 151                 spin_unlock(&type->obd_type_lock);
 152         }
 153         return type;
 154 }
 155
 156 void class_put_type(struct obd_type *type)
 157 {
 158         LASSERT(type);
 159         spin_lock(&type->obd_type_lock);
 160         type->typ_refcnt--;
 161         module_put(type->typ_dt_ops->o_owner);
 162         spin_unlock(&type->obd_type_lock);
 163 }
 164
 165 #define CLASS_MAX_NAME 1024
 166
 167 int class_register_type(struct obd_ops *dt_ops, struct md_ops *md_ops,
 168                         bool enable_proc, struct lprocfs_vars *vars,
 169                         const char *name, struct lu_device_type *ldt)
 170 {
 171         struct obd_type *type;
 172         int rc = 0;
 173         ENTRY;
 174
 175         /* sanity check */
 176         LASSERT(strnlen(name, CLASS_MAX_NAME) < CLASS_MAX_NAME);
 177
 178         if (class_search_type(name)) {
 179                 CDEBUG(D_IOCTL, "Type %s already registered\n", name);
 180                 RETURN(-EEXIST);
 181         }
 182
 183         rc = -ENOMEM;
 184         OBD_ALLOC(type, sizeof(*type));
 185         if (type == NULL)
 186                 RETURN(rc);
 187
 188         OBD_ALLOC_PTR(type->typ_dt_ops);
 189         OBD_ALLOC_PTR(type->typ_md_ops);
 190         OBD_ALLOC(type->typ_name, strlen(name) + 1);
 191
 192         if (type->typ_dt_ops == NULL ||
 193             type->typ_md_ops == NULL ||
 194             type->typ_name == NULL)
 195                 GOTO (failed, rc);
 196
 197         *(type->typ_dt_ops) = *dt_ops;
 198         /* md_ops is optional */
 199         if (md_ops)
 200                 *(type->typ_md_ops) = *md_ops;
 201         strcpy(type->typ_name, name);
 202         spin_lock_init(&type->obd_type_lock);
 203
 204 #ifdef CONFIG_PROC_FS
 205         if (enable_proc) {
 206                 type->typ_procroot = lprocfs_register(type->typ_name,
 207                                                       proc_lustre_root,
 208                                                       vars, type);
 209                 if (IS_ERR(type->typ_procroot)) {
 210                         rc = PTR_ERR(type->typ_procroot);
 211                         type->typ_procroot = NULL;
 212                         GOTO(failed, rc);
 213                 }
 214         }
 215 #endif
 216         if (ldt != NULL) {
 217                 type->typ_lu = ldt;
 218                 rc = lu_device_type_init(ldt);
 219                 if (rc != 0)
 220                         GOTO (failed, rc);
 221         }
 222
 223         spin_lock(&obd_types_lock);
 224         list_add(&type->typ_chain, &obd_types);
 225         spin_unlock(&obd_types_lock);
 226
 227         RETURN (0);
 228
 229 failed:
 230         if (type->typ_name != NULL) {
 231 #ifdef CONFIG_PROC_FS
 232                 if (type->typ_procroot != NULL)
 233                         remove_proc_subtree(type->typ_name, proc_lustre_root);
 234 #endif
 235                 OBD_FREE(type->typ_name, strlen(name) + 1);
 236         }
 237         if (type->typ_md_ops != NULL)
 238                 OBD_FREE_PTR(type->typ_md_ops);
 239         if (type->typ_dt_ops != NULL)
 240                 OBD_FREE_PTR(type->typ_dt_ops);
 241         OBD_FREE(type, sizeof(*type));
 242         RETURN(rc);
 243 }
 244 EXPORT_SYMBOL(class_register_type);
 245
 246 int class_unregister_type(const char *name)
 247 {
 248         struct obd_type *type = class_search_type(name);
 249         ENTRY;
 250
 251         if (!type) {
 252                 CERROR("unknown obd type\n");
 253                 RETURN(-EINVAL);
 254         }
 255
 256         if (type->typ_refcnt) {
 257                 CERROR("type %s has refcount (%d)\n", name, type->typ_refcnt);
 258                 /* This is a bad situation, let's make the best of it */
 259                 /* Remove ops, but leave the name for debugging */
 260                 OBD_FREE_PTR(type->typ_dt_ops);
 261                 OBD_FREE_PTR(type->typ_md_ops);
 262                 RETURN(-EBUSY);
 263         }
 264
 265         /* we do not use type->typ_procroot as for compatibility purposes
 266          * other modules can share names (i.e. lod can use lov entry). so
 267          * we can't reference pointer as it can get invalided when another
 268          * module removes the entry */
 269 #ifdef CONFIG_PROC_FS
 270         if (type->typ_procroot != NULL)
 271                 remove_proc_subtree(type->typ_name, proc_lustre_root);
 272         if (type->typ_procsym != NULL)
 273                 lprocfs_remove(&type->typ_procsym);
 274 #endif
 275         if (type->typ_lu)
 276                 lu_device_type_fini(type->typ_lu);
 277
 278         spin_lock(&obd_types_lock);
 279         list_del(&type->typ_chain);
 280         spin_unlock(&obd_types_lock);
 281         OBD_FREE(type->typ_name, strlen(name) + 1);
 282         if (type->typ_dt_ops != NULL)
 283                 OBD_FREE_PTR(type->typ_dt_ops);
 284         if (type->typ_md_ops != NULL)
 285                 OBD_FREE_PTR(type->typ_md_ops);
 286         OBD_FREE(type, sizeof(*type));
 287         RETURN(0);
 288 } /* class_unregister_type */
 289 EXPORT_SYMBOL(class_unregister_type);
 290
 291 /**
 292  * Create a new obd device.
 293  *
 294  * Find an empty slot in ::obd_devs[], create a new obd device in it.
 295  *
 296  * \param[in] type_name obd device type string.
 297  * \param[in] name      obd device name.
 298  *
 299  * \retval NULL if create fails, otherwise return the obd device
 300  *         pointer created.
 301  */
 302 struct obd_device *class_newdev(const char *type_name, const char *name)
 303 {
 304         struct obd_device *result = NULL;
 305         struct obd_device *newdev;
 306         struct obd_type *type = NULL;
 307         int i;
 308         int new_obd_minor = 0;
 309         ENTRY;
 310
 311         if (strlen(name) >= MAX_OBD_NAME) {
 312                 CERROR("name/uuid must be < %u bytes long\n", MAX_OBD_NAME);
 313                 RETURN(ERR_PTR(-EINVAL));
 314         }
 315
 316         type = class_get_type(type_name);
 317         if (type == NULL){
 318                 CERROR("OBD: unknown type: %s\n", type_name);
 319                 RETURN(ERR_PTR(-ENODEV));
 320         }
 321
 322         newdev = obd_device_alloc();
 323         if (newdev == NULL)
 324                 GOTO(out_type, result = ERR_PTR(-ENOMEM));
 325
 326         LASSERT(newdev->obd_magic == OBD_DEVICE_MAGIC);
 327
 328         write_lock(&obd_dev_lock);
 329         for (i = 0; i < class_devno_max(); i++) {
 330                 struct obd_device *obd = class_num2obd(i);
 331
 332                 if (obd && (strcmp(name, obd->obd_name) == 0)) {
 333                         CERROR("Device %s already exists at %d, won't add\n",
 334                                name, i);
 335                         if (result) {
 336                                 LASSERTF(result->obd_magic == OBD_DEVICE_MAGIC,
 337                                          "%p obd_magic %08x != %08x\n", result,
 338                                          result->obd_magic, OBD_DEVICE_MAGIC);
 339                                 LASSERTF(result->obd_minor == new_obd_minor,
 340                                          "%p obd_minor %d != %d\n", result,
 341                                          result->obd_minor, new_obd_minor);
 342
 343                                 obd_devs[result->obd_minor] = NULL;
 344                                 result->obd_name[0]='\0';
 345                          }
 346                         result = ERR_PTR(-EEXIST);
 347                         break;
 348                 }
 349                 if (!result && !obd) {
 350                         result = newdev;
 351                         result->obd_minor = i;
 352                         new_obd_minor = i;
 353                         result->obd_type = type;
 354                         strncpy(result->obd_name, name,
 355                                 sizeof(result->obd_name) - 1);
 356                         obd_devs[i] = result;
 357                 }
 358         }
 359         write_unlock(&obd_dev_lock);
 360
 361         if (result == NULL && i >= class_devno_max()) {
 362                 CERROR("all %u OBD devices used, increase MAX_OBD_DEVICES\n",
 363                        class_devno_max());
 364                 GOTO(out, result = ERR_PTR(-EOVERFLOW));
 365         }
 366
 367         if (IS_ERR(result))
 368                 GOTO(out, result);
 369
 370         CDEBUG(D_IOCTL, "Adding new device %s (%p)\n",
 371                result->obd_name, result);
 372
 373         RETURN(result);
 374 out:
 375         obd_device_free(newdev);
 376 out_type:
 377         class_put_type(type);
 378         return result;
 379 }
 380
 381 void class_release_dev(struct obd_device *obd)
 382 {
 383         struct obd_type *obd_type = obd->obd_type;
 384
 385         LASSERTF(obd->obd_magic == OBD_DEVICE_MAGIC, "%p obd_magic %08x != %08x\n",
 386                  obd, obd->obd_magic, OBD_DEVICE_MAGIC);
 387         LASSERTF(obd == obd_devs[obd->obd_minor], "obd %p != obd_devs[%d] %p\n",
 388                  obd, obd->obd_minor, obd_devs[obd->obd_minor]);
 389         LASSERT(obd_type != NULL);
 390
 391         CDEBUG(D_INFO, "Release obd device %s at %d obd_type name =%s\n",
 392                obd->obd_name, obd->obd_minor, obd->obd_type->typ_name);
 393
 394         write_lock(&obd_dev_lock);
 395         obd_devs[obd->obd_minor] = NULL;
 396         write_unlock(&obd_dev_lock);
 397         obd_device_free(obd);
 398
 399         class_put_type(obd_type);
 400 }
 401
 402 int class_name2dev(const char *name)
 403 {
 404         int i;
 405
 406         if (!name)
 407                 return -1;
 408
 409         read_lock(&obd_dev_lock);
 410         for (i = 0; i < class_devno_max(); i++) {
 411                 struct obd_device *obd = class_num2obd(i);
 412
 413                 if (obd && strcmp(name, obd->obd_name) == 0) {
 414                         /* Make sure we finished attaching before we give
 415                            out any references */
 416                         LASSERT(obd->obd_magic == OBD_DEVICE_MAGIC);
 417                         if (obd->obd_attached) {
 418                                 read_unlock(&obd_dev_lock);
 419                                 return i;
 420                         }
 421                         break;
 422                 }
 423         }
 424         read_unlock(&obd_dev_lock);
 425
 426         return -1;
 427 }
 428
 429 struct obd_device *class_name2obd(const char *name)
 430 {
 431         int dev = class_name2dev(name);
 432
 433         if (dev < 0 || dev > class_devno_max())
 434                 return NULL;
 435         return class_num2obd(dev);
 436 }
 437 EXPORT_SYMBOL(class_name2obd);
 438
 439 int class_uuid2dev(struct obd_uuid *uuid)
 440 {
 441         int i;
 442
 443         read_lock(&obd_dev_lock);
 444         for (i = 0; i < class_devno_max(); i++) {
 445                 struct obd_device *obd = class_num2obd(i);
 446
 447                 if (obd && obd_uuid_equals(uuid, &obd->obd_uuid)) {
 448                         LASSERT(obd->obd_magic == OBD_DEVICE_MAGIC);
 449                         read_unlock(&obd_dev_lock);
 450                         return i;
 451                 }
 452         }
 453         read_unlock(&obd_dev_lock);
 454
 455         return -1;
 456 }
 457
 458 struct obd_device *class_uuid2obd(struct obd_uuid *uuid)
 459 {
 460         int dev = class_uuid2dev(uuid);
 461         if (dev < 0)
 462                 return NULL;
 463         return class_num2obd(dev);
 464 }
 465 EXPORT_SYMBOL(class_uuid2obd);
 466
 467 /**
 468  * Get obd device from ::obd_devs[]
 469  *
 470  * \param num [in] array index
 471  *
 472  * \retval NULL if ::obd_devs[\a num] does not contains an obd device
 473  *         otherwise return the obd device there.
 474  */
 475 struct obd_device *class_num2obd(int num)
 476 {
 477         struct obd_device *obd = NULL;
 478
 479         if (num < class_devno_max()) {
 480                 obd = obd_devs[num];
 481                 if (obd == NULL)
 482                         return NULL;
 483
 484                 LASSERTF(obd->obd_magic == OBD_DEVICE_MAGIC,
 485                          "%p obd_magic %08x != %08x\n",
 486                          obd, obd->obd_magic, OBD_DEVICE_MAGIC);
 487                 LASSERTF(obd->obd_minor == num,
 488                          "%p obd_minor %0d != %0d\n",
 489                          obd, obd->obd_minor, num);
 490         }
 491
 492         return obd;
 493 }
 494
 495 /**
 496  * Get obd devices count. Device in any
 497  *    state are counted
 498  * \retval obd device count
 499  */
 500 int get_devices_count(void)
 501 {
 502         int index, max_index = class_devno_max(), dev_count = 0;
 503
 504         read_lock(&obd_dev_lock);
 505         for (index = 0; index <= max_index; index++) {
 506                 struct obd_device *obd = class_num2obd(index);
 507                 if (obd != NULL)
 508                         dev_count++;
 509         }
 510         read_unlock(&obd_dev_lock);
 511
 512         return dev_count;
 513 }
 514 EXPORT_SYMBOL(get_devices_count);
 515
 516 void class_obd_list(void)
 517 {
 518         char *status;
 519         int i;
 520
 521         read_lock(&obd_dev_lock);
 522         for (i = 0; i < class_devno_max(); i++) {
 523                 struct obd_device *obd = class_num2obd(i);
 524
 525                 if (obd == NULL)
 526                         continue;
 527                 if (obd->obd_stopping)
 528                         status = "ST";
 529                 else if (obd->obd_set_up)
 530                         status = "UP";
 531                 else if (obd->obd_attached)
 532                         status = "AT";
 533                 else
 534                         status = "--";
 535                 LCONSOLE(D_CONFIG, "%3d %s %s %s %s %d\n",
 536                          i, status, obd->obd_type->typ_name,
 537                          obd->obd_name, obd->obd_uuid.uuid,
 538                          atomic_read(&obd->obd_refcount));
 539         }
 540         read_unlock(&obd_dev_lock);
 541         return;
 542 }
 543
 544 /* Search for a client OBD connected to tgt_uuid.  If grp_uuid is
 545    specified, then only the client with that uuid is returned,
 546    otherwise any client connected to the tgt is returned. */
 547 struct obd_device * class_find_client_obd(struct obd_uuid *tgt_uuid,
 548                                           const char * typ_name,
 549                                           struct obd_uuid *grp_uuid)
 550 {
 551         int i;
 552
 553         read_lock(&obd_dev_lock);
 554         for (i = 0; i < class_devno_max(); i++) {
 555                 struct obd_device *obd = class_num2obd(i);
 556
 557                 if (obd == NULL)
 558                         continue;
 559                 if ((strncmp(obd->obd_type->typ_name, typ_name,
 560                              strlen(typ_name)) == 0)) {
 561                         if (obd_uuid_equals(tgt_uuid,
 562                                             &obd->u.cli.cl_target_uuid) &&
 563                             ((grp_uuid)? obd_uuid_equals(grp_uuid,
 564                                                          &obd->obd_uuid) : 1)) {
 565                                 read_unlock(&obd_dev_lock);
 566                                 return obd;
 567                         }
 568                 }
 569         }
 570         read_unlock(&obd_dev_lock);
 571
 572         return NULL;
 573 }
 574 EXPORT_SYMBOL(class_find_client_obd);
 575
 576 /* Iterate the obd_device list looking devices have grp_uuid. Start
 577    searching at *next, and if a device is found, the next index to look
 578    at is saved in *next. If next is NULL, then the first matching device
 579    will always be returned. */
 580 struct obd_device * class_devices_in_group(struct obd_uuid *grp_uuid, int *next)
 581 {
 582         int i;
 583
 584         if (next == NULL)
 585                 i = 0;
 586         else if (*next >= 0 && *next < class_devno_max())
 587                 i = *next;
 588         else
 589                 return NULL;
 590
 591         read_lock(&obd_dev_lock);
 592         for (; i < class_devno_max(); i++) {
 593                 struct obd_device *obd = class_num2obd(i);
 594
 595                 if (obd == NULL)
 596                         continue;
 597                 if (obd_uuid_equals(grp_uuid, &obd->obd_uuid)) {
 598                         if (next != NULL)
 599                                 *next = i+1;
 600                         read_unlock(&obd_dev_lock);
 601                         return obd;
 602                 }
 603         }
 604         read_unlock(&obd_dev_lock);
 605
 606         return NULL;
 607 }
 608 EXPORT_SYMBOL(class_devices_in_group);
 609
 610 /**
 611  * to notify sptlrpc log for \a fsname has changed, let every relevant OBD
 612  * adjust sptlrpc settings accordingly.
 613  */
 614 int class_notify_sptlrpc_conf(const char *fsname, int namelen)
 615 {
 616         struct obd_device  *obd;
 617         const char         *type;
 618         int                 i, rc = 0, rc2;
 619
 620         LASSERT(namelen > 0);
 621
 622         read_lock(&obd_dev_lock);
 623         for (i = 0; i < class_devno_max(); i++) {
 624                 obd = class_num2obd(i);
 625
 626                 if (obd == NULL || obd->obd_set_up == 0 || obd->obd_stopping)
 627                         continue;
 628
 629                 /* only notify mdc, osc, osp, lwp, mdt, ost
 630                  * because only these have a -sptlrpc llog */
 631                 type = obd->obd_type->typ_name;
 632                 if (strcmp(type, LUSTRE_MDC_NAME) != 0 &&
 633                     strcmp(type, LUSTRE_OSC_NAME) != 0 &&
 634                     strcmp(type, LUSTRE_OSP_NAME) != 0 &&
 635                     strcmp(type, LUSTRE_LWP_NAME) != 0 &&
 636                     strcmp(type, LUSTRE_MDT_NAME) != 0 &&
 637                     strcmp(type, LUSTRE_OST_NAME) != 0)
 638                         continue;
 639
 640                 if (strncmp(obd->obd_name, fsname, namelen))
 641                         continue;
 642
 643                 class_incref(obd, __FUNCTION__, obd);
 644                 read_unlock(&obd_dev_lock);
 645                 rc2 = obd_set_info_async(NULL, obd->obd_self_export,
 646                                          sizeof(KEY_SPTLRPC_CONF),
 647                                          KEY_SPTLRPC_CONF, 0, NULL, NULL);
 648                 rc = rc ? rc : rc2;
 649                 class_decref(obd, __FUNCTION__, obd);
 650                 read_lock(&obd_dev_lock);
 651         }
 652         read_unlock(&obd_dev_lock);
 653         return rc;
 654 }
 655 EXPORT_SYMBOL(class_notify_sptlrpc_conf);
 656
 657 void obd_cleanup_caches(void)
 658 {
 659         ENTRY;
 660         if (obd_device_cachep) {
 661                 kmem_cache_destroy(obd_device_cachep);
 662                 obd_device_cachep = NULL;
 663         }
 664         if (obdo_cachep) {
 665                 kmem_cache_destroy(obdo_cachep);
 666                 obdo_cachep = NULL;
 667         }
 668         if (import_cachep) {
 669                 kmem_cache_destroy(import_cachep);
 670                 import_cachep = NULL;
 671         }
 672
 673         EXIT;
 674 }
 675
 676 int obd_init_caches(void)
 677 {
 678         int rc;
 679         ENTRY;
 680
 681         LASSERT(obd_device_cachep == NULL);
 682         obd_device_cachep = kmem_cache_create("ll_obd_dev_cache",
 683                                               sizeof(struct obd_device),
 684                                               0, 0, NULL);
 685         if (!obd_device_cachep)
 686                 GOTO(out, rc = -ENOMEM);
 687
 688         LASSERT(obdo_cachep == NULL);
 689         obdo_cachep = kmem_cache_create("ll_obdo_cache", sizeof(struct obdo),
 690                                         0, 0, NULL);
 691         if (!obdo_cachep)
 692                 GOTO(out, rc = -ENOMEM);
 693
 694         LASSERT(import_cachep == NULL);
 695         import_cachep = kmem_cache_create("ll_import_cache",
 696                                           sizeof(struct obd_import),
 697                                           0, 0, NULL);
 698         if (!import_cachep)
 699                 GOTO(out, rc = -ENOMEM);
 700
 701         RETURN(0);
 702 out:
 703         obd_cleanup_caches();
 704         RETURN(rc);
 705 }
 706
 707 /* map connection to client */
 708 struct obd_export *class_conn2export(struct lustre_handle *conn)
 709 {
 710         struct obd_export *export;
 711         ENTRY;
 712
 713         if (!conn) {
 714                 CDEBUG(D_CACHE, "looking for null handle\n");
 715                 RETURN(NULL);
 716         }
 717
 718         if (conn->cookie == -1) {  /* this means assign a new connection */
 719                 CDEBUG(D_CACHE, "want a new connection\n");
 720                 RETURN(NULL);
 721         }
 722
 723         CDEBUG(D_INFO, "looking for export cookie "LPX64"\n", conn->cookie);
 724         export = class_handle2object(conn->cookie, NULL);
 725         RETURN(export);
 726 }
 727 EXPORT_SYMBOL(class_conn2export);
 728
 729 struct obd_device *class_exp2obd(struct obd_export *exp)
 730 {
 731         if (exp)
 732                 return exp->exp_obd;
 733         return NULL;
 734 }
 735 EXPORT_SYMBOL(class_exp2obd);
 736
 737 struct obd_device *class_conn2obd(struct lustre_handle *conn)
 738 {
 739         struct obd_export *export;
 740         export = class_conn2export(conn);
 741         if (export) {
 742                 struct obd_device *obd = export->exp_obd;
 743                 class_export_put(export);
 744                 return obd;
 745         }
 746         return NULL;
 747 }
 748
 749 struct obd_import *class_exp2cliimp(struct obd_export *exp)
 750 {
 751         struct obd_device *obd = exp->exp_obd;
 752         if (obd == NULL)
 753                 return NULL;
 754         return obd->u.cli.cl_import;
 755 }
 756 EXPORT_SYMBOL(class_exp2cliimp);
 757
 758 struct obd_import *class_conn2cliimp(struct lustre_handle *conn)
 759 {
 760         struct obd_device *obd = class_conn2obd(conn);
 761         if (obd == NULL)
 762                 return NULL;
 763         return obd->u.cli.cl_import;
 764 }
 765
 766 /* Export management functions */
 767 static void class_export_destroy(struct obd_export *exp)
 768 {
 769         struct obd_device *obd = exp->exp_obd;
 770         ENTRY;
 771
 772         LASSERT_ATOMIC_ZERO(&exp->exp_refcount);
 773         LASSERT(obd != NULL);
 774
 775         CDEBUG(D_IOCTL, "destroying export %p/%s for %s\n", exp,
 776                exp->exp_client_uuid.uuid, obd->obd_name);
 777
 778         /* "Local" exports (lctl, LOV->{mdc,osc}) have no connection. */
 779         if (exp->exp_connection)
 780                 ptlrpc_put_connection_superhack(exp->exp_connection);
 781
 782         LASSERT(list_empty(&exp->exp_outstanding_replies));
 783         LASSERT(list_empty(&exp->exp_uncommitted_replies));
 784         LASSERT(list_empty(&exp->exp_req_replay_queue));
 785         LASSERT(list_empty(&exp->exp_hp_rpcs));
 786         obd_destroy_export(exp);
 787         class_decref(obd, "export", exp);
 788
 789         OBD_FREE_RCU(exp, sizeof(*exp), &exp->exp_handle);
 790         EXIT;
 791 }
 792
 793 static void export_handle_addref(void *export)
 794 {
 795         class_export_get(export);
 796 }
 797
 798 static struct portals_handle_ops export_handle_ops = {
 799         .hop_addref = export_handle_addref,
 800         .hop_free   = NULL,
 801 };
 802
 803 struct obd_export *class_export_get(struct obd_export *exp)
 804 {
 805         atomic_inc(&exp->exp_refcount);
 806         CDEBUG(D_INFO, "GETting export %p : new refcount %d\n", exp,
 807                atomic_read(&exp->exp_refcount));
 808         return exp;
 809 }
 810 EXPORT_SYMBOL(class_export_get);
 811
 812 void class_export_put(struct obd_export *exp)
 813 {
 814         LASSERT(exp != NULL);
 815         LASSERT_ATOMIC_GT_LT(&exp->exp_refcount, 0, LI_POISON);
 816         CDEBUG(D_INFO, "PUTting export %p : new refcount %d\n", exp,
 817                atomic_read(&exp->exp_refcount) - 1);
 818
 819         if (atomic_dec_and_test(&exp->exp_refcount)) {
 820                 LASSERT(!list_empty(&exp->exp_obd_chain));
 821                 LASSERT(list_empty(&exp->exp_stale_list));
 822                 CDEBUG(D_IOCTL, "final put %p/%s\n",
 823                        exp, exp->exp_client_uuid.uuid);
 824
 825                 /* release nid stat refererence */
 826                 lprocfs_exp_cleanup(exp);
 827
 828                 obd_zombie_export_add(exp);
 829         }
 830 }
 831 EXPORT_SYMBOL(class_export_put);
 832
 833 /* Creates a new export, adds it to the hash table, and returns a
 834  * pointer to it. The refcount is 2: one for the hash reference, and
 835  * one for the pointer returned by this function. */
 836 struct obd_export *class_new_export(struct obd_device *obd,
 837                                     struct obd_uuid *cluuid)
 838 {
 839         struct obd_export *export;
 840         struct cfs_hash *hash = NULL;
 841         int rc = 0;
 842         ENTRY;
 843
 844         OBD_ALLOC_PTR(export);
 845         if (!export)
 846                 return ERR_PTR(-ENOMEM);
 847
 848         export->exp_conn_cnt = 0;
 849         export->exp_lock_hash = NULL;
 850         export->exp_flock_hash = NULL;
 851         atomic_set(&export->exp_refcount, 2);
 852         atomic_set(&export->exp_rpc_count, 0);
 853         atomic_set(&export->exp_cb_count, 0);
 854         atomic_set(&export->exp_locks_count, 0);
 855 #if LUSTRE_TRACKS_LOCK_EXP_REFS
 856         INIT_LIST_HEAD(&export->exp_locks_list);
 857         spin_lock_init(&export->exp_locks_list_guard);
 858 #endif
 859         atomic_set(&export->exp_replay_count, 0);
 860         export->exp_obd = obd;
 861         INIT_LIST_HEAD(&export->exp_outstanding_replies);
 862         spin_lock_init(&export->exp_uncommitted_replies_lock);
 863         INIT_LIST_HEAD(&export->exp_uncommitted_replies);
 864         INIT_LIST_HEAD(&export->exp_req_replay_queue);
 865         INIT_LIST_HEAD(&export->exp_handle.h_link);
 866         INIT_LIST_HEAD(&export->exp_hp_rpcs);
 867         INIT_LIST_HEAD(&export->exp_reg_rpcs);
 868         class_handle_hash(&export->exp_handle, &export_handle_ops);
 869         export->exp_last_request_time = cfs_time_current_sec();
 870         spin_lock_init(&export->exp_lock);
 871         spin_lock_init(&export->exp_rpc_lock);
 872         INIT_HLIST_NODE(&export->exp_uuid_hash);
 873         INIT_HLIST_NODE(&export->exp_nid_hash);
 874         INIT_HLIST_NODE(&export->exp_gen_hash);
 875         spin_lock_init(&export->exp_bl_list_lock);
 876         INIT_LIST_HEAD(&export->exp_bl_list);
 877         INIT_LIST_HEAD(&export->exp_stale_list);
 878
 879         export->exp_sp_peer = LUSTRE_SP_ANY;
 880         export->exp_flvr.sf_rpc = SPTLRPC_FLVR_INVALID;
 881         export->exp_client_uuid = *cluuid;
 882         obd_init_export(export);
 883
 884         spin_lock(&obd->obd_dev_lock);
 885         /* shouldn't happen, but might race */
 886         if (obd->obd_stopping)
 887                 GOTO(exit_unlock, rc = -ENODEV);
 888
 889         hash = cfs_hash_getref(obd->obd_uuid_hash);
 890         if (hash == NULL)
 891                 GOTO(exit_unlock, rc = -ENODEV);
 892         spin_unlock(&obd->obd_dev_lock);
 893
 894         if (!obd_uuid_equals(cluuid, &obd->obd_uuid)) {
 895                 rc = cfs_hash_add_unique(hash, cluuid, &export->exp_uuid_hash);
 896                 if (rc != 0) {
 897                         LCONSOLE_WARN("%s: denying duplicate export for %s, %d\n",
 898                                       obd->obd_name, cluuid->uuid, rc);
 899                         GOTO(exit_err, rc = -EALREADY);
 900                 }
 901         }
 902
 903         at_init(&export->exp_bl_lock_at, obd_timeout, 0);
 904         spin_lock(&obd->obd_dev_lock);
 905         if (obd->obd_stopping) {
 906                 cfs_hash_del(hash, cluuid, &export->exp_uuid_hash);
 907                 GOTO(exit_unlock, rc = -ENODEV);
 908         }
 909
 910         class_incref(obd, "export", export);
 911         list_add(&export->exp_obd_chain, &export->exp_obd->obd_exports);
 912         list_add_tail(&export->exp_obd_chain_timed,
 913                       &export->exp_obd->obd_exports_timed);
 914         export->exp_obd->obd_num_exports++;
 915         spin_unlock(&obd->obd_dev_lock);
 916         cfs_hash_putref(hash);
 917         RETURN(export);
 918
 919 exit_unlock:
 920         spin_unlock(&obd->obd_dev_lock);
 921 exit_err:
 922         if (hash)
 923                 cfs_hash_putref(hash);
 924         class_handle_unhash(&export->exp_handle);
 925         LASSERT(hlist_unhashed(&export->exp_uuid_hash));
 926         obd_destroy_export(export);
 927         OBD_FREE_PTR(export);
 928         return ERR_PTR(rc);
 929 }
 930 EXPORT_SYMBOL(class_new_export);
 931
 932 void class_unlink_export(struct obd_export *exp)
 933 {
 934         class_handle_unhash(&exp->exp_handle);
 935
 936         spin_lock(&exp->exp_obd->obd_dev_lock);
 937         /* delete an uuid-export hashitem from hashtables */
 938         if (!hlist_unhashed(&exp->exp_uuid_hash))
 939                 cfs_hash_del(exp->exp_obd->obd_uuid_hash,
 940                              &exp->exp_client_uuid,
 941                              &exp->exp_uuid_hash);
 942
 943         if (!hlist_unhashed(&exp->exp_gen_hash)) {
 944                 struct tg_export_data   *ted = &exp->exp_target_data;
 945                 struct cfs_hash         *hash;
 946
 947                 /* Because obd_gen_hash will not be released until
 948                  * class_cleanup(), so hash should never be NULL here */
 949                 hash = cfs_hash_getref(exp->exp_obd->obd_gen_hash);
 950                 LASSERT(hash != NULL);
 951                 cfs_hash_del(hash, &ted->ted_lcd->lcd_generation,
 952                              &exp->exp_gen_hash);
 953                 cfs_hash_putref(hash);
 954         }
 955
 956         list_move(&exp->exp_obd_chain, &exp->exp_obd->obd_unlinked_exports);
 957         list_del_init(&exp->exp_obd_chain_timed);
 958         exp->exp_obd->obd_num_exports--;
 959         spin_unlock(&exp->exp_obd->obd_dev_lock);
 960         atomic_inc(&obd_stale_export_num);
 961
 962         /* A reference is kept by obd_stale_exports list */
 963         obd_stale_export_put(exp);
 964 }
 965 EXPORT_SYMBOL(class_unlink_export);
 966
 967 /* Import management functions */
 968 static void class_import_destroy(struct obd_import *imp)
 969 {
 970         ENTRY;
 971
 972         CDEBUG(D_IOCTL, "destroying import %p for %s\n", imp,
 973                 imp->imp_obd->obd_name);
 974
 975         LASSERT_ATOMIC_ZERO(&imp->imp_refcount);
 976
 977         ptlrpc_put_connection_superhack(imp->imp_connection);
 978
 979         while (!list_empty(&imp->imp_conn_list)) {
 980                 struct obd_import_conn *imp_conn;
 981
 982                 imp_conn = list_entry(imp->imp_conn_list.next,
 983                                       struct obd_import_conn, oic_item);
 984                 list_del_init(&imp_conn->oic_item);
 985                 ptlrpc_put_connection_superhack(imp_conn->oic_conn);
 986                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 987         }
 988
 989         LASSERT(imp->imp_sec == NULL);
 990         class_decref(imp->imp_obd, "import", imp);
 991         OBD_FREE_RCU(imp, sizeof(*imp), &imp->imp_handle);
 992         EXIT;
 993 }
 994
 995 static void import_handle_addref(void *import)
 996 {
 997         class_import_get(import);
 998 }
 999
1000 static struct portals_handle_ops import_handle_ops = {
1001         .hop_addref = import_handle_addref,
1002         .hop_free   = NULL,
1003 };
1004
1005 struct obd_import *class_import_get(struct obd_import *import)
1006 {
1007         atomic_inc(&import->imp_refcount);
1008         CDEBUG(D_INFO, "import %p refcount=%d obd=%s\n", import,
1009                atomic_read(&import->imp_refcount),
1010                import->imp_obd->obd_name);
1011         return import;
1012 }
1013 EXPORT_SYMBOL(class_import_get);
1014
1015 void class_import_put(struct obd_import *imp)
1016 {
1017         ENTRY;
1018
1019         LASSERT(list_empty(&imp->imp_zombie_chain));
1020         LASSERT_ATOMIC_GT_LT(&imp->imp_refcount, 0, LI_POISON);
1021
1022         CDEBUG(D_INFO, "import %p refcount=%d obd=%s\n", imp,
1023                atomic_read(&imp->imp_refcount) - 1,
1024                imp->imp_obd->obd_name);
1025
1026         if (atomic_dec_and_test(&imp->imp_refcount)) {
1027                 CDEBUG(D_INFO, "final put import %p\n", imp);
1028                 obd_zombie_import_add(imp);
1029         }
1030
1031         /* catch possible import put race */
1032         LASSERT_ATOMIC_GE_LT(&imp->imp_refcount, 0, LI_POISON);
1033         EXIT;
1034 }
1035 EXPORT_SYMBOL(class_import_put);
1036
1037 static void init_imp_at(struct imp_at *at) {
1038         int i;
1039         at_init(&at->iat_net_latency, 0, 0);
1040         for (i = 0; i < IMP_AT_MAX_PORTALS; i++) {
1041                 /* max service estimates are tracked on the server side, so
1042                    don't use the AT history here, just use the last reported
1043                    val. (But keep hist for proc histogram, worst_ever) */
1044                 at_init(&at->iat_service_estimate[i], INITIAL_CONNECT_TIMEOUT,
1045                         AT_FLG_NOHIST);
1046         }
1047 }
1048
1049 struct obd_import *class_new_import(struct obd_device *obd)
1050 {
1051         struct obd_import *imp;
1052
1053         OBD_ALLOC(imp, sizeof(*imp));
1054         if (imp == NULL)
1055                 return NULL;
1056
1057         INIT_LIST_HEAD(&imp->imp_pinger_chain);
1058         INIT_LIST_HEAD(&imp->imp_zombie_chain);
1059         INIT_LIST_HEAD(&imp->imp_replay_list);
1060         INIT_LIST_HEAD(&imp->imp_sending_list);
1061         INIT_LIST_HEAD(&imp->imp_delayed_list);
1062         INIT_LIST_HEAD(&imp->imp_committed_list);
1063         imp->imp_replay_cursor = &imp->imp_committed_list;
1064         spin_lock_init(&imp->imp_lock);
1065         imp->imp_last_success_conn = 0;
1066         imp->imp_state = LUSTRE_IMP_NEW;
1067         imp->imp_obd = class_incref(obd, "import", imp);
1068         mutex_init(&imp->imp_sec_mutex);
1069         init_waitqueue_head(&imp->imp_recovery_waitq);
1070
1071         atomic_set(&imp->imp_refcount, 2);
1072         atomic_set(&imp->imp_unregistering, 0);
1073         atomic_set(&imp->imp_inflight, 0);
1074         atomic_set(&imp->imp_replay_inflight, 0);
1075         atomic_set(&imp->imp_inval_count, 0);
1076         INIT_LIST_HEAD(&imp->imp_conn_list);
1077         INIT_LIST_HEAD(&imp->imp_handle.h_link);
1078         class_handle_hash(&imp->imp_handle, &import_handle_ops);
1079         init_imp_at(&imp->imp_at);
1080
1081         /* the default magic is V2, will be used in connect RPC, and
1082          * then adjusted according to the flags in request/reply. */
1083         imp->imp_msg_magic = LUSTRE_MSG_MAGIC_V2;
1084
1085         return imp;
1086 }
1087 EXPORT_SYMBOL(class_new_import);
1088
1089 void class_destroy_import(struct obd_import *import)
1090 {
1091         LASSERT(import != NULL);
1092         LASSERT(import != LP_POISON);
1093
1094         class_handle_unhash(&import->imp_handle);
1095
1096         spin_lock(&import->imp_lock);
1097         import->imp_generation++;
1098         spin_unlock(&import->imp_lock);
1099         class_import_put(import);
1100 }
1101 EXPORT_SYMBOL(class_destroy_import);
1102
1103 #if LUSTRE_TRACKS_LOCK_EXP_REFS
1104
1105 void __class_export_add_lock_ref(struct obd_export *exp, struct ldlm_lock *lock)
1106 {
1107         spin_lock(&exp->exp_locks_list_guard);
1108
1109         LASSERT(lock->l_exp_refs_nr >= 0);
1110
1111         if (lock->l_exp_refs_target != NULL &&
1112             lock->l_exp_refs_target != exp) {
1113                 LCONSOLE_WARN("setting export %p for lock %p which already has export %p\n",
1114                               exp, lock, lock->l_exp_refs_target);
1115         }
1116         if ((lock->l_exp_refs_nr ++) == 0) {
1117                 list_add(&lock->l_exp_refs_link, &exp->exp_locks_list);
1118                 lock->l_exp_refs_target = exp;
1119         }
1120         CDEBUG(D_INFO, "lock = %p, export = %p, refs = %u\n",
1121                lock, exp, lock->l_exp_refs_nr);
1122         spin_unlock(&exp->exp_locks_list_guard);
1123 }
1124
1125 void __class_export_del_lock_ref(struct obd_export *exp, struct ldlm_lock *lock)
1126 {
1127         spin_lock(&exp->exp_locks_list_guard);
1128         LASSERT(lock->l_exp_refs_nr > 0);
1129         if (lock->l_exp_refs_target != exp) {
1130                 LCONSOLE_WARN("lock %p, "
1131                               "mismatching export pointers: %p, %p\n",
1132                               lock, lock->l_exp_refs_target, exp);
1133         }
1134         if (-- lock->l_exp_refs_nr == 0) {
1135                 list_del_init(&lock->l_exp_refs_link);
1136                 lock->l_exp_refs_target = NULL;
1137         }
1138         CDEBUG(D_INFO, "lock = %p, export = %p, refs = %u\n",
1139                lock, exp, lock->l_exp_refs_nr);
1140         spin_unlock(&exp->exp_locks_list_guard);
1141 }
1142 #endif
1143
1144 /* A connection defines an export context in which preallocation can
1145    be managed. This releases the export pointer reference, and returns
1146    the export handle, so the export refcount is 1 when this function
1147    returns. */
1148 int class_connect(struct lustre_handle *conn, struct obd_device *obd,
1149                   struct obd_uuid *cluuid)
1150 {
1151         struct obd_export *export;
1152         LASSERT(conn != NULL);
1153         LASSERT(obd != NULL);
1154         LASSERT(cluuid != NULL);
1155         ENTRY;
1156
1157         export = class_new_export(obd, cluuid);
1158         if (IS_ERR(export))
1159                 RETURN(PTR_ERR(export));
1160
1161         conn->cookie = export->exp_handle.h_cookie;
1162         class_export_put(export);
1163
1164         CDEBUG(D_IOCTL, "connect: client %s, cookie "LPX64"\n",
1165                cluuid->uuid, conn->cookie);
1166         RETURN(0);
1167 }
1168 EXPORT_SYMBOL(class_connect);
1169
1170 /* if export is involved in recovery then clean up related things */
1171 static void class_export_recovery_cleanup(struct obd_export *exp)
1172 {
1173         struct obd_device *obd = exp->exp_obd;
1174
1175         spin_lock(&obd->obd_recovery_task_lock);
1176         if (obd->obd_recovering) {
1177                 if (exp->exp_in_recovery) {
1178                         spin_lock(&exp->exp_lock);
1179                         exp->exp_in_recovery = 0;
1180                         spin_unlock(&exp->exp_lock);
1181                         LASSERT_ATOMIC_POS(&obd->obd_connected_clients);
1182                         atomic_dec(&obd->obd_connected_clients);
1183                 }
1184
1185                 /* if called during recovery then should update
1186                  * obd_stale_clients counter,
1187                  * lightweight exports are not counted */
1188                 if ((exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT) == 0)
1189                         exp->exp_obd->obd_stale_clients++;
1190         }
1191         spin_unlock(&obd->obd_recovery_task_lock);
1192
1193         spin_lock(&exp->exp_lock);
1194         /** Cleanup req replay fields */
1195         if (exp->exp_req_replay_needed) {
1196                 exp->exp_req_replay_needed = 0;
1197
1198                 LASSERT(atomic_read(&obd->obd_req_replay_clients));
1199                 atomic_dec(&obd->obd_req_replay_clients);
1200         }
1201
1202         /** Cleanup lock replay data */
1203         if (exp->exp_lock_replay_needed) {
1204                 exp->exp_lock_replay_needed = 0;
1205
1206                 LASSERT(atomic_read(&obd->obd_lock_replay_clients));
1207                 atomic_dec(&obd->obd_lock_replay_clients);
1208         }
1209         spin_unlock(&exp->exp_lock);
1210 }
1211
1212 /* This function removes 1-3 references from the export:
1213  * 1 - for export pointer passed
1214  * and if disconnect really need
1215  * 2 - removing from hash
1216  * 3 - in client_unlink_export
1217  * The export pointer passed to this function can destroyed */
1218 int class_disconnect(struct obd_export *export)
1219 {
1220         int already_disconnected;
1221         ENTRY;
1222
1223         if (export == NULL) {
1224                 CWARN("attempting to free NULL export %p\n", export);
1225                 RETURN(-EINVAL);
1226         }
1227
1228         spin_lock(&export->exp_lock);
1229         already_disconnected = export->exp_disconnected;
1230         export->exp_disconnected = 1;
1231         spin_unlock(&export->exp_lock);
1232
1233         /* class_cleanup(), abort_recovery(), and class_fail_export()
1234          * all end up in here, and if any of them race we shouldn't
1235          * call extra class_export_puts(). */
1236         if (already_disconnected) {
1237                 LASSERT(hlist_unhashed(&export->exp_nid_hash));
1238                 GOTO(no_disconn, already_disconnected);
1239         }
1240
1241         CDEBUG(D_IOCTL, "disconnect: cookie "LPX64"\n",
1242                export->exp_handle.h_cookie);
1243
1244         if (!hlist_unhashed(&export->exp_nid_hash))
1245                 cfs_hash_del(export->exp_obd->obd_nid_hash,
1246                              &export->exp_connection->c_peer.nid,
1247                              &export->exp_nid_hash);
1248
1249         class_export_recovery_cleanup(export);
1250         class_unlink_export(export);
1251 no_disconn:
1252         class_export_put(export);
1253         RETURN(0);
1254 }
1255 EXPORT_SYMBOL(class_disconnect);
1256
1257 /* Return non-zero for a fully connected export */
1258 int class_connected_export(struct obd_export *exp)
1259 {
1260         int connected = 0;
1261
1262         if (exp) {
1263                 spin_lock(&exp->exp_lock);
1264                 connected = (exp->exp_conn_cnt > 0) && !exp->exp_failed;
1265                 spin_unlock(&exp->exp_lock);
1266         }
1267         return connected;
1268 }
1269 EXPORT_SYMBOL(class_connected_export);
1270
1271 static void class_disconnect_export_list(struct list_head *list,
1272                                          enum obd_option flags)
1273 {
1274         int rc;
1275         struct obd_export *exp;
1276         ENTRY;
1277
1278         /* It's possible that an export may disconnect itself, but
1279          * nothing else will be added to this list. */
1280         while (!list_empty(list)) {
1281                 exp = list_entry(list->next, struct obd_export,
1282                                  exp_obd_chain);
1283                 /* need for safe call CDEBUG after obd_disconnect */
1284                 class_export_get(exp);
1285
1286                 spin_lock(&exp->exp_lock);
1287                 exp->exp_flags = flags;
1288                 spin_unlock(&exp->exp_lock);
1289
1290                 if (obd_uuid_equals(&exp->exp_client_uuid,
1291                                     &exp->exp_obd->obd_uuid)) {
1292                         CDEBUG(D_HA,
1293                                "exp %p export uuid == obd uuid, don't discon\n",
1294                                exp);
1295                         /* Need to delete this now so we don't end up pointing
1296                          * to work_list later when this export is cleaned up. */
1297                         list_del_init(&exp->exp_obd_chain);
1298                         class_export_put(exp);
1299                         continue;
1300                 }
1301
1302                 class_export_get(exp);
1303                 CDEBUG(D_HA, "%s: disconnecting export at %s (%p), "
1304                        "last request at "CFS_TIME_T"\n",
1305                        exp->exp_obd->obd_name, obd_export_nid2str(exp),
1306                        exp, exp->exp_last_request_time);
1307                 /* release one export reference anyway */
1308                 rc = obd_disconnect(exp);
1309
1310                 CDEBUG(D_HA, "disconnected export at %s (%p): rc %d\n",
1311                        obd_export_nid2str(exp), exp, rc);
1312                 class_export_put(exp);
1313         }
1314         EXIT;
1315 }
1316
1317 void class_disconnect_exports(struct obd_device *obd)
1318 {
1319         struct list_head work_list;
1320         ENTRY;
1321
1322         /* Move all of the exports from obd_exports to a work list, en masse. */
1323         INIT_LIST_HEAD(&work_list);
1324         spin_lock(&obd->obd_dev_lock);
1325         list_splice_init(&obd->obd_exports, &work_list);
1326         list_splice_init(&obd->obd_delayed_exports, &work_list);
1327         spin_unlock(&obd->obd_dev_lock);
1328
1329         if (!list_empty(&work_list)) {
1330                 CDEBUG(D_HA, "OBD device %d (%p) has exports, "
1331                        "disconnecting them\n", obd->obd_minor, obd);
1332                 class_disconnect_export_list(&work_list,
1333                                              exp_flags_from_obd(obd));
1334         } else
1335                 CDEBUG(D_HA, "OBD device %d (%p) has no exports\n",
1336                        obd->obd_minor, obd);
1337         EXIT;
1338 }
1339 EXPORT_SYMBOL(class_disconnect_exports);
1340
1341 /* Remove exports that have not completed recovery.
1342  */
1343 void class_disconnect_stale_exports(struct obd_device *obd,
1344                                     int (*test_export)(struct obd_export *))
1345 {
1346         struct list_head work_list;
1347         struct obd_export *exp, *n;
1348         int evicted = 0;
1349         ENTRY;
1350
1351         INIT_LIST_HEAD(&work_list);
1352         spin_lock(&obd->obd_dev_lock);
1353         list_for_each_entry_safe(exp, n, &obd->obd_exports,
1354                                  exp_obd_chain) {
1355                 /* don't count self-export as client */
1356                 if (obd_uuid_equals(&exp->exp_client_uuid,
1357                                     &exp->exp_obd->obd_uuid))
1358                         continue;
1359
1360                 /* don't evict clients which have no slot in last_rcvd
1361                  * (e.g. lightweight connection) */
1362                 if (exp->exp_target_data.ted_lr_idx == -1)
1363                         continue;
1364
1365                 spin_lock(&exp->exp_lock);
1366                 if (exp->exp_failed || test_export(exp)) {
1367                         spin_unlock(&exp->exp_lock);
1368                         continue;
1369                 }
1370                 exp->exp_failed = 1;
1371                 spin_unlock(&exp->exp_lock);
1372
1373                 list_move(&exp->exp_obd_chain, &work_list);
1374                 evicted++;
1375                 CDEBUG(D_HA, "%s: disconnect stale client %s@%s\n",
1376                        obd->obd_name, exp->exp_client_uuid.uuid,
1377                        exp->exp_connection == NULL ? "<unknown>" :
1378                        libcfs_nid2str(exp->exp_connection->c_peer.nid));
1379                 print_export_data(exp, "EVICTING", 0);
1380         }
1381         spin_unlock(&obd->obd_dev_lock);
1382
1383         if (evicted)
1384                 LCONSOLE_WARN("%s: disconnecting %d stale clients\n",
1385                               obd->obd_name, evicted);
1386
1387         class_disconnect_export_list(&work_list, exp_flags_from_obd(obd) |
1388                                                  OBD_OPT_ABORT_RECOV);
1389         EXIT;
1390 }
1391 EXPORT_SYMBOL(class_disconnect_stale_exports);
1392
1393 void class_fail_export(struct obd_export *exp)
1394 {
1395         int rc, already_failed;
1396
1397         spin_lock(&exp->exp_lock);
1398         already_failed = exp->exp_failed;
1399         exp->exp_failed = 1;
1400         spin_unlock(&exp->exp_lock);
1401
1402         if (already_failed) {
1403                 CDEBUG(D_HA, "disconnecting dead export %p/%s; skipping\n",
1404                        exp, exp->exp_client_uuid.uuid);
1405                 return;
1406         }
1407
1408         CDEBUG(D_HA, "disconnecting export %p/%s\n",
1409                exp, exp->exp_client_uuid.uuid);
1410
1411         if (obd_dump_on_timeout)
1412                 libcfs_debug_dumplog();
1413
1414         /* need for safe call CDEBUG after obd_disconnect */
1415         class_export_get(exp);
1416
1417         /* Most callers into obd_disconnect are removing their own reference
1418          * (request, for example) in addition to the one from the hash table.
1419          * We don't have such a reference here, so make one. */
1420         class_export_get(exp);
1421         rc = obd_disconnect(exp);
1422         if (rc)
1423                 CERROR("disconnecting export %p failed: %d\n", exp, rc);
1424         else
1425                 CDEBUG(D_HA, "disconnected export %p/%s\n",
1426                        exp, exp->exp_client_uuid.uuid);
1427         class_export_put(exp);
1428 }
1429 EXPORT_SYMBOL(class_fail_export);
1430
1431 char *obd_export_nid2str(struct obd_export *exp)
1432 {
1433         if (exp->exp_connection != NULL)
1434                 return libcfs_nid2str(exp->exp_connection->c_peer.nid);
1435
1436         return "(no nid)";
1437 }
1438 EXPORT_SYMBOL(obd_export_nid2str);
1439
1440 int obd_export_evict_by_nid(struct obd_device *obd, const char *nid)
1441 {
1442         struct cfs_hash *nid_hash;
1443         struct obd_export *doomed_exp = NULL;
1444         int exports_evicted = 0;
1445
1446         lnet_nid_t nid_key = libcfs_str2nid((char *)nid);
1447
1448         spin_lock(&obd->obd_dev_lock);
1449         /* umount has run already, so evict thread should leave
1450          * its task to umount thread now */
1451         if (obd->obd_stopping) {
1452                 spin_unlock(&obd->obd_dev_lock);
1453                 return exports_evicted;
1454         }
1455         nid_hash = obd->obd_nid_hash;
1456         cfs_hash_getref(nid_hash);
1457         spin_unlock(&obd->obd_dev_lock);
1458
1459         do {
1460                 doomed_exp = cfs_hash_lookup(nid_hash, &nid_key);
1461                 if (doomed_exp == NULL)
1462                         break;
1463
1464                 LASSERTF(doomed_exp->exp_connection->c_peer.nid == nid_key,
1465                          "nid %s found, wanted nid %s, requested nid %s\n",
1466                          obd_export_nid2str(doomed_exp),
1467                          libcfs_nid2str(nid_key), nid);
1468                 LASSERTF(doomed_exp != obd->obd_self_export,
1469                          "self-export is hashed by NID?\n");
1470                 exports_evicted++;
1471                 LCONSOLE_WARN("%s: evicting %s (at %s) by administrative "
1472                               "request\n", obd->obd_name,
1473                               obd_uuid2str(&doomed_exp->exp_client_uuid),
1474                               obd_export_nid2str(doomed_exp));
1475                 class_fail_export(doomed_exp);
1476                 class_export_put(doomed_exp);
1477         } while (1);
1478
1479         cfs_hash_putref(nid_hash);
1480
1481         if (!exports_evicted)
1482                 CDEBUG(D_HA,"%s: can't disconnect NID '%s': no exports found\n",
1483                        obd->obd_name, nid);
1484         return exports_evicted;
1485 }
1486 EXPORT_SYMBOL(obd_export_evict_by_nid);
1487
1488 int obd_export_evict_by_uuid(struct obd_device *obd, const char *uuid)
1489 {
1490         struct cfs_hash *uuid_hash;
1491         struct obd_export *doomed_exp = NULL;
1492         struct obd_uuid doomed_uuid;
1493         int exports_evicted = 0;
1494
1495         spin_lock(&obd->obd_dev_lock);
1496         if (obd->obd_stopping) {
1497                 spin_unlock(&obd->obd_dev_lock);
1498                 return exports_evicted;
1499         }
1500         uuid_hash = obd->obd_uuid_hash;
1501         cfs_hash_getref(uuid_hash);
1502         spin_unlock(&obd->obd_dev_lock);
1503
1504         obd_str2uuid(&doomed_uuid, uuid);
1505         if (obd_uuid_equals(&doomed_uuid, &obd->obd_uuid)) {
1506                 CERROR("%s: can't evict myself\n", obd->obd_name);
1507                 cfs_hash_putref(uuid_hash);
1508                 return exports_evicted;
1509         }
1510
1511         doomed_exp = cfs_hash_lookup(uuid_hash, &doomed_uuid);
1512
1513         if (doomed_exp == NULL) {
1514                 CERROR("%s: can't disconnect %s: no exports found\n",
1515                        obd->obd_name, uuid);
1516         } else {
1517                 CWARN("%s: evicting %s at adminstrative request\n",
1518                        obd->obd_name, doomed_exp->exp_client_uuid.uuid);
1519                 class_fail_export(doomed_exp);
1520                 class_export_put(doomed_exp);
1521                 exports_evicted++;
1522         }
1523         cfs_hash_putref(uuid_hash);
1524
1525         return exports_evicted;
1526 }
1527
1528 #if LUSTRE_TRACKS_LOCK_EXP_REFS
1529 void (*class_export_dump_hook)(struct obd_export*) = NULL;
1530 #endif
1531
1532 static void print_export_data(struct obd_export *exp, const char *status,
1533                               int locks)
1534 {
1535         struct ptlrpc_reply_state *rs;
1536         struct ptlrpc_reply_state *first_reply = NULL;
1537         int nreplies = 0;
1538
1539         spin_lock(&exp->exp_lock);
1540         list_for_each_entry(rs, &exp->exp_outstanding_replies,
1541                             rs_exp_list) {
1542                 if (nreplies == 0)
1543                         first_reply = rs;
1544                 nreplies++;
1545         }
1546         spin_unlock(&exp->exp_lock);
1547
1548         CDEBUG(D_HA, "%s: %s %p %s %s %d (%d %d %d) %d %d %d %d: %p %s "LPU64"\n",
1549                exp->exp_obd->obd_name, status, exp, exp->exp_client_uuid.uuid,
1550                obd_export_nid2str(exp), atomic_read(&exp->exp_refcount),
1551                atomic_read(&exp->exp_rpc_count),
1552                atomic_read(&exp->exp_cb_count),
1553                atomic_read(&exp->exp_locks_count),
1554                exp->exp_disconnected, exp->exp_delayed, exp->exp_failed,
1555                nreplies, first_reply, nreplies > 3 ? "..." : "",
1556                exp->exp_last_committed);
1557 #if LUSTRE_TRACKS_LOCK_EXP_REFS
1558         if (locks && class_export_dump_hook != NULL)
1559                 class_export_dump_hook(exp);
1560 #endif
1561 }
1562
1563 void dump_exports(struct obd_device *obd, int locks)
1564 {
1565         struct obd_export *exp;
1566
1567         spin_lock(&obd->obd_dev_lock);
1568         list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain)
1569                 print_export_data(exp, "ACTIVE", locks);
1570         list_for_each_entry(exp, &obd->obd_unlinked_exports, exp_obd_chain)
1571                 print_export_data(exp, "UNLINKED", locks);
1572         list_for_each_entry(exp, &obd->obd_delayed_exports, exp_obd_chain)
1573                 print_export_data(exp, "DELAYED", locks);
1574         spin_unlock(&obd->obd_dev_lock);
1575         spin_lock(&obd_zombie_impexp_lock);
1576         list_for_each_entry(exp, &obd_zombie_exports, exp_obd_chain)
1577                 print_export_data(exp, "ZOMBIE", locks);
1578         spin_unlock(&obd_zombie_impexp_lock);
1579 }
1580
1581 void obd_exports_barrier(struct obd_device *obd)
1582 {
1583         int waited = 2;
1584         LASSERT(list_empty(&obd->obd_exports));
1585         spin_lock(&obd->obd_dev_lock);
1586         while (!list_empty(&obd->obd_unlinked_exports)) {
1587                 spin_unlock(&obd->obd_dev_lock);
1588                 set_current_state(TASK_UNINTERRUPTIBLE);
1589                 schedule_timeout(cfs_time_seconds(waited));
1590                 if (waited > 5 && IS_PO2(waited)) {
1591                         LCONSOLE_WARN("%s is waiting for obd_unlinked_exports "
1592                                       "more than %d seconds. "
1593                                       "The obd refcount = %d. Is it stuck?\n",
1594                                       obd->obd_name, waited,
1595                                       atomic_read(&obd->obd_refcount));
1596                         dump_exports(obd, 1);
1597                 }
1598                 waited *= 2;
1599                 spin_lock(&obd->obd_dev_lock);
1600         }
1601         spin_unlock(&obd->obd_dev_lock);
1602 }
1603 EXPORT_SYMBOL(obd_exports_barrier);
1604
1605 /* Total amount of zombies to be destroyed */
1606 static int zombies_count = 0;
1607
1608 /**
1609  * kill zombie imports and exports
1610  */
1611 void obd_zombie_impexp_cull(void)
1612 {
1613         struct obd_import *import;
1614         struct obd_export *export;
1615         ENTRY;
1616
1617         do {
1618                 spin_lock(&obd_zombie_impexp_lock);
1619
1620                 import = NULL;
1621                 if (!list_empty(&obd_zombie_imports)) {
1622                         import = list_entry(obd_zombie_imports.next,
1623                                             struct obd_import,
1624                                             imp_zombie_chain);
1625                         list_del_init(&import->imp_zombie_chain);
1626                 }
1627
1628                 export = NULL;
1629                 if (!list_empty(&obd_zombie_exports)) {
1630                         export = list_entry(obd_zombie_exports.next,
1631                                             struct obd_export,
1632                                             exp_obd_chain);
1633                         list_del_init(&export->exp_obd_chain);
1634                 }
1635
1636                 spin_unlock(&obd_zombie_impexp_lock);
1637
1638                 if (import != NULL) {
1639                         class_import_destroy(import);
1640                         spin_lock(&obd_zombie_impexp_lock);
1641                         zombies_count--;
1642                         spin_unlock(&obd_zombie_impexp_lock);
1643                 }
1644
1645                 if (export != NULL) {
1646                         class_export_destroy(export);
1647                         spin_lock(&obd_zombie_impexp_lock);
1648                         zombies_count--;
1649                         spin_unlock(&obd_zombie_impexp_lock);
1650                 }
1651
1652                 cond_resched();
1653         } while (import != NULL || export != NULL);
1654         EXIT;
1655 }
1656
1657 static struct completion        obd_zombie_start;
1658 static struct completion        obd_zombie_stop;
1659 static unsigned long            obd_zombie_flags;
1660 static wait_queue_head_t        obd_zombie_waitq;
1661 static pid_t                    obd_zombie_pid;
1662
1663 enum {
1664         OBD_ZOMBIE_STOP         = 0x0001,
1665 };
1666
1667 /**
1668  * check for work for kill zombie import/export thread.
1669  */
1670 static int obd_zombie_impexp_check(void *arg)
1671 {
1672         int rc;
1673
1674         spin_lock(&obd_zombie_impexp_lock);
1675         rc = (zombies_count == 0) &&
1676              !test_bit(OBD_ZOMBIE_STOP, &obd_zombie_flags);
1677         spin_unlock(&obd_zombie_impexp_lock);
1678
1679         RETURN(rc);
1680 }
1681
1682 /**
1683  * Add export to the obd_zombe thread and notify it.
1684  */
1685 static void obd_zombie_export_add(struct obd_export *exp) {
1686         atomic_dec(&obd_stale_export_num);
1687         spin_lock(&exp->exp_obd->obd_dev_lock);
1688         LASSERT(!list_empty(&exp->exp_obd_chain));
1689         list_del_init(&exp->exp_obd_chain);
1690         spin_unlock(&exp->exp_obd->obd_dev_lock);
1691         spin_lock(&obd_zombie_impexp_lock);
1692         zombies_count++;
1693         list_add(&exp->exp_obd_chain, &obd_zombie_exports);
1694         spin_unlock(&obd_zombie_impexp_lock);
1695
1696         obd_zombie_impexp_notify();
1697 }
1698
1699 /**
1700  * Add import to the obd_zombe thread and notify it.
1701  */
1702 static void obd_zombie_import_add(struct obd_import *imp) {
1703         LASSERT(imp->imp_sec == NULL);
1704         spin_lock(&obd_zombie_impexp_lock);
1705         LASSERT(list_empty(&imp->imp_zombie_chain));
1706         zombies_count++;
1707         list_add(&imp->imp_zombie_chain, &obd_zombie_imports);
1708         spin_unlock(&obd_zombie_impexp_lock);
1709
1710         obd_zombie_impexp_notify();
1711 }
1712
1713 /**
1714  * notify import/export destroy thread about new zombie.
1715  */
1716 static void obd_zombie_impexp_notify(void)
1717 {
1718         /*
1719          * Make sure obd_zomebie_impexp_thread get this notification.
1720          * It is possible this signal only get by obd_zombie_barrier, and
1721          * barrier gulps this notification and sleeps away and hangs ensues
1722          */
1723         wake_up_all(&obd_zombie_waitq);
1724 }
1725
1726 /**
1727  * check whether obd_zombie is idle
1728  */
1729 static int obd_zombie_is_idle(void)
1730 {
1731         int rc;
1732
1733         LASSERT(!test_bit(OBD_ZOMBIE_STOP, &obd_zombie_flags));
1734         spin_lock(&obd_zombie_impexp_lock);
1735         rc = (zombies_count == 0);
1736         spin_unlock(&obd_zombie_impexp_lock);
1737         return rc;
1738 }
1739
1740 /**
1741  * wait when obd_zombie import/export queues become empty
1742  */
1743 void obd_zombie_barrier(void)
1744 {
1745         struct l_wait_info lwi = { 0 };
1746
1747         if (obd_zombie_pid == current_pid())
1748                 /* don't wait for myself */
1749                 return;
1750         l_wait_event(obd_zombie_waitq, obd_zombie_is_idle(), &lwi);
1751 }
1752 EXPORT_SYMBOL(obd_zombie_barrier);
1753
1754
1755 struct obd_export *obd_stale_export_get(void)
1756 {
1757         struct obd_export *exp = NULL;
1758         ENTRY;
1759
1760         spin_lock(&obd_stale_export_lock);
1761         if (!list_empty(&obd_stale_exports)) {
1762                 exp = list_entry(obd_stale_exports.next,
1763                                  struct obd_export, exp_stale_list);
1764                 list_del_init(&exp->exp_stale_list);
1765         }
1766         spin_unlock(&obd_stale_export_lock);
1767
1768         if (exp) {
1769                 CDEBUG(D_DLMTRACE, "Get export %p: total %d\n", exp,
1770                        atomic_read(&obd_stale_export_num));
1771         }
1772         RETURN(exp);
1773 }
1774 EXPORT_SYMBOL(obd_stale_export_get);
1775
1776 void obd_stale_export_put(struct obd_export *exp)
1777 {
1778         ENTRY;
1779
1780         LASSERT(list_empty(&exp->exp_stale_list));
1781         if (exp->exp_lock_hash &&
1782             atomic_read(&exp->exp_lock_hash->hs_count)) {
1783                 CDEBUG(D_DLMTRACE, "Put export %p: total %d\n", exp,
1784                        atomic_read(&obd_stale_export_num));
1785
1786                 spin_lock_bh(&exp->exp_bl_list_lock);
1787                 spin_lock(&obd_stale_export_lock);
1788                 /* Add to the tail if there is no blocked locks,
1789                  * to the head otherwise. */
1790                 if (list_empty(&exp->exp_bl_list))
1791                         list_add_tail(&exp->exp_stale_list,
1792                                       &obd_stale_exports);
1793                 else
1794                         list_add(&exp->exp_stale_list,
1795                                  &obd_stale_exports);
1796
1797                 spin_unlock(&obd_stale_export_lock);
1798                 spin_unlock_bh(&exp->exp_bl_list_lock);
1799         } else {
1800                 class_export_put(exp);
1801         }
1802         EXIT;
1803 }
1804 EXPORT_SYMBOL(obd_stale_export_put);
1805
1806 /**
1807  * Adjust the position of the export in the stale list,
1808  * i.e. move to the head of the list if is needed.
1809  **/
1810 void obd_stale_export_adjust(struct obd_export *exp)
1811 {
1812         LASSERT(exp != NULL);
1813         spin_lock_bh(&exp->exp_bl_list_lock);
1814         spin_lock(&obd_stale_export_lock);
1815
1816         if (!list_empty(&exp->exp_stale_list) &&
1817             !list_empty(&exp->exp_bl_list))
1818                 list_move(&exp->exp_stale_list, &obd_stale_exports);
1819
1820         spin_unlock(&obd_stale_export_lock);
1821         spin_unlock_bh(&exp->exp_bl_list_lock);
1822 }
1823 EXPORT_SYMBOL(obd_stale_export_adjust);
1824
1825 /**
1826  * destroy zombie export/import thread.
1827  */
1828 static int obd_zombie_impexp_thread(void *unused)
1829 {
1830         unshare_fs_struct();
1831         complete(&obd_zombie_start);
1832
1833         obd_zombie_pid = current_pid();
1834
1835         while (!test_bit(OBD_ZOMBIE_STOP, &obd_zombie_flags)) {
1836                 struct l_wait_info lwi = { 0 };
1837
1838                 l_wait_event(obd_zombie_waitq,
1839                              !obd_zombie_impexp_check(NULL), &lwi);
1840                 obd_zombie_impexp_cull();
1841
1842                 /*
1843                  * Notify obd_zombie_barrier callers that queues
1844                  * may be empty.
1845                  */
1846                 wake_up(&obd_zombie_waitq);
1847         }
1848
1849         complete(&obd_zombie_stop);
1850
1851         RETURN(0);
1852 }
1853
1854
1855 /**
1856  * start destroy zombie import/export thread
1857  */
1858 int obd_zombie_impexp_init(void)
1859 {
1860         struct task_struct *task;
1861
1862         INIT_LIST_HEAD(&obd_zombie_imports);
1863
1864         INIT_LIST_HEAD(&obd_zombie_exports);
1865         spin_lock_init(&obd_zombie_impexp_lock);
1866         init_completion(&obd_zombie_start);
1867         init_completion(&obd_zombie_stop);
1868         init_waitqueue_head(&obd_zombie_waitq);
1869         obd_zombie_pid = 0;
1870
1871         task = kthread_run(obd_zombie_impexp_thread, NULL, "obd_zombid");
1872         if (IS_ERR(task))
1873                 RETURN(PTR_ERR(task));
1874
1875         wait_for_completion(&obd_zombie_start);
1876         RETURN(0);
1877 }
1878 /**
1879  * stop destroy zombie import/export thread
1880  */
1881 void obd_zombie_impexp_stop(void)
1882 {
1883         set_bit(OBD_ZOMBIE_STOP, &obd_zombie_flags);
1884         obd_zombie_impexp_notify();
1885         wait_for_completion(&obd_zombie_stop);
1886 }
1887
1888 /***** Kernel-userspace comm helpers *******/
1889
1890 /* Get length of entire message, including header */
1891 int kuc_len(int payload_len)
1892 {
1893         return sizeof(struct kuc_hdr) + payload_len;
1894 }
1895 EXPORT_SYMBOL(kuc_len);
1896
1897 /* Get a pointer to kuc header, given a ptr to the payload
1898  * @param p Pointer to payload area
1899  * @returns Pointer to kuc header
1900  */
1901 struct kuc_hdr * kuc_ptr(void *p)
1902 {
1903         struct kuc_hdr *lh = ((struct kuc_hdr *)p) - 1;
1904         LASSERT(lh->kuc_magic == KUC_MAGIC);
1905         return lh;
1906 }
1907 EXPORT_SYMBOL(kuc_ptr);
1908
1909 /* Test if payload is part of kuc message
1910  * @param p Pointer to payload area
1911  * @returns boolean
1912  */
1913 int kuc_ispayload(void *p)
1914 {
1915         struct kuc_hdr *kh = ((struct kuc_hdr *)p) - 1;
1916
1917         if (kh->kuc_magic == KUC_MAGIC)
1918                 return 1;
1919         else
1920                 return 0;
1921 }
1922 EXPORT_SYMBOL(kuc_ispayload);
1923
1924 /* Alloc space for a message, and fill in header
1925  * @return Pointer to payload area
1926  */
1927 void *kuc_alloc(int payload_len, int transport, int type)
1928 {
1929         struct kuc_hdr *lh;
1930         int len = kuc_len(payload_len);
1931
1932         OBD_ALLOC(lh, len);
1933         if (lh == NULL)
1934                 return ERR_PTR(-ENOMEM);
1935
1936         lh->kuc_magic = KUC_MAGIC;
1937         lh->kuc_transport = transport;
1938         lh->kuc_msgtype = type;
1939         lh->kuc_msglen = len;
1940
1941         return (void *)(lh + 1);
1942 }
1943 EXPORT_SYMBOL(kuc_alloc);
1944
1945 /* Takes pointer to payload area */
1946 inline void kuc_free(void *p, int payload_len)
1947 {
1948         struct kuc_hdr *lh = kuc_ptr(p);
1949         OBD_FREE(lh, kuc_len(payload_len));
1950 }
1951 EXPORT_SYMBOL(kuc_free);
1952
1953 struct obd_request_slot_waiter {
1954         struct list_head        orsw_entry;
1955         wait_queue_head_t       orsw_waitq;
1956         bool                    orsw_signaled;
1957 };
1958
1959 static bool obd_request_slot_avail(struct client_obd *cli,
1960                                    struct obd_request_slot_waiter *orsw)
1961 {
1962         bool avail;
1963
1964         spin_lock(&cli->cl_loi_list_lock);
1965         avail = !!list_empty(&orsw->orsw_entry);
1966         spin_unlock(&cli->cl_loi_list_lock);
1967
1968         return avail;
1969 };
1970
1971 /*
1972  * For network flow control, the RPC sponsor needs to acquire a credit
1973  * before sending the RPC. The credits count for a connection is defined
1974  * by the "cl_max_rpcs_in_flight". If all the credits are occpuied, then
1975  * the subsequent RPC sponsors need to wait until others released their
1976  * credits, or the administrator increased the "cl_max_rpcs_in_flight".
1977  */
1978 int obd_get_request_slot(struct client_obd *cli)
1979 {
1980         struct obd_request_slot_waiter   orsw;
1981         struct l_wait_info               lwi;
1982         int                              rc;
1983
1984         spin_lock(&cli->cl_loi_list_lock);
1985         if (cli->cl_r_in_flight < cli->cl_max_rpcs_in_flight) {
1986                 cli->cl_r_in_flight++;
1987                 spin_unlock(&cli->cl_loi_list_lock);
1988                 return 0;
1989         }
1990
1991         init_waitqueue_head(&orsw.orsw_waitq);
1992         list_add_tail(&orsw.orsw_entry, &cli->cl_loi_read_list);
1993         orsw.orsw_signaled = false;
1994         spin_unlock(&cli->cl_loi_list_lock);
1995
1996         lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
1997         rc = l_wait_event(orsw.orsw_waitq,
1998                           obd_request_slot_avail(cli, &orsw) ||
1999                           orsw.orsw_signaled,
2000                           &lwi);
2001
2002         /* Here, we must take the lock to avoid the on-stack 'orsw' to be
2003          * freed but other (such as obd_put_request_slot) is using it. */
2004         spin_lock(&cli->cl_loi_list_lock);
2005         if (rc != 0) {
2006                 if (!orsw.orsw_signaled) {
2007                         if (list_empty(&orsw.orsw_entry))
2008                                 cli->cl_r_in_flight--;
2009                         else
2010                                 list_del(&orsw.orsw_entry);
2011                 }
2012         }
2013
2014         if (orsw.orsw_signaled) {
2015                 LASSERT(list_empty(&orsw.orsw_entry));
2016
2017                 rc = -EINTR;
2018         }
2019         spin_unlock(&cli->cl_loi_list_lock);
2020
2021         return rc;
2022 }
2023 EXPORT_SYMBOL(obd_get_request_slot);
2024
2025 void obd_put_request_slot(struct client_obd *cli)
2026 {
2027         struct obd_request_slot_waiter *orsw;
2028
2029         spin_lock(&cli->cl_loi_list_lock);
2030         cli->cl_r_in_flight--;
2031
2032         /* If there is free slot, wakeup the first waiter. */
2033         if (!list_empty(&cli->cl_loi_read_list) &&
2034             likely(cli->cl_r_in_flight < cli->cl_max_rpcs_in_flight)) {
2035                 orsw = list_entry(cli->cl_loi_read_list.next,
2036                                   struct obd_request_slot_waiter, orsw_entry);
2037                 list_del_init(&orsw->orsw_entry);
2038                 cli->cl_r_in_flight++;
2039                 wake_up(&orsw->orsw_waitq);
2040         }
2041         spin_unlock(&cli->cl_loi_list_lock);
2042 }
2043 EXPORT_SYMBOL(obd_put_request_slot);
2044
2045 __u32 obd_get_max_rpcs_in_flight(struct client_obd *cli)
2046 {
2047         return cli->cl_max_rpcs_in_flight;
2048 }
2049 EXPORT_SYMBOL(obd_get_max_rpcs_in_flight);
2050
2051 int obd_set_max_rpcs_in_flight(struct client_obd *cli, __u32 max)
2052 {
2053         struct obd_request_slot_waiter *orsw;
2054         __u32                           old;
2055         int                             diff;
2056         int                             i;
2057         char                            *typ_name;
2058         int                             rc;
2059
2060         if (max > OBD_MAX_RIF_MAX || max < 1)
2061                 return -ERANGE;
2062
2063         typ_name = cli->cl_import->imp_obd->obd_type->typ_name;
2064         if (strcmp(typ_name, LUSTRE_MDC_NAME) == 0) {
2065                 /* adjust max_mod_rpcs_in_flight to ensure it is always
2066                  * strictly lower that max_rpcs_in_flight */
2067                 if (max < 2) {
2068                         CERROR("%s: cannot set max_rpcs_in_flight to 1 "
2069                                "because it must be higher than "
2070                                "max_mod_rpcs_in_flight value",
2071                                cli->cl_import->imp_obd->obd_name);
2072                         return -ERANGE;
2073                 }
2074                 if (max <= cli->cl_max_mod_rpcs_in_flight) {
2075                         rc = obd_set_max_mod_rpcs_in_flight(cli, max - 1);
2076                         if (rc != 0)
2077                                 return rc;
2078                 }
2079         }
2080
2081         spin_lock(&cli->cl_loi_list_lock);
2082         old = cli->cl_max_rpcs_in_flight;
2083         cli->cl_max_rpcs_in_flight = max;
2084         diff = max - old;
2085
2086         /* We increase the max_rpcs_in_flight, then wakeup some waiters. */
2087         for (i = 0; i < diff; i++) {
2088                 if (list_empty(&cli->cl_loi_read_list))
2089                         break;
2090
2091                 orsw = list_entry(cli->cl_loi_read_list.next,
2092                                   struct obd_request_slot_waiter, orsw_entry);
2093                 list_del_init(&orsw->orsw_entry);
2094                 cli->cl_r_in_flight++;
2095                 wake_up(&orsw->orsw_waitq);
2096         }
2097         spin_unlock(&cli->cl_loi_list_lock);
2098
2099         return 0;
2100 }
2101 EXPORT_SYMBOL(obd_set_max_rpcs_in_flight);
2102
2103 __u16 obd_get_max_mod_rpcs_in_flight(struct client_obd *cli)
2104 {
2105         return cli->cl_max_mod_rpcs_in_flight;
2106 }
2107 EXPORT_SYMBOL(obd_get_max_mod_rpcs_in_flight);
2108
2109 int obd_set_max_mod_rpcs_in_flight(struct client_obd *cli, __u16 max)
2110 {
2111         struct obd_connect_data *ocd;
2112         __u16 maxmodrpcs;
2113         __u16 prev;
2114
2115         if (max > OBD_MAX_RIF_MAX || max < 1)
2116                 return -ERANGE;
2117
2118         /* cannot exceed or equal max_rpcs_in_flight */
2119         if (max >= cli->cl_max_rpcs_in_flight) {
2120                 CERROR("%s: can't set max_mod_rpcs_in_flight to a value (%hu) "
2121                        "higher or equal to max_rpcs_in_flight value (%u)\n",
2122                        cli->cl_import->imp_obd->obd_name,
2123                        max, cli->cl_max_rpcs_in_flight);
2124                 return -ERANGE;
2125         }
2126
2127         /* cannot exceed max modify RPCs in flight supported by the server */
2128         ocd = &cli->cl_import->imp_connect_data;
2129         if (ocd->ocd_connect_flags & OBD_CONNECT_MULTIMODRPCS)
2130                 maxmodrpcs = ocd->ocd_maxmodrpcs;
2131         else
2132                 maxmodrpcs = 1;
2133         if (max > maxmodrpcs) {
2134                 CERROR("%s: can't set max_mod_rpcs_in_flight to a value (%hu) "
2135                        "higher than max_mod_rpcs_per_client value (%hu) "
2136                        "returned by the server at connection\n",
2137                        cli->cl_import->imp_obd->obd_name,
2138                        max, maxmodrpcs);
2139                 return -ERANGE;
2140         }
2141
2142         spin_lock(&cli->cl_mod_rpcs_lock);
2143
2144         prev = cli->cl_max_mod_rpcs_in_flight;
2145         cli->cl_max_mod_rpcs_in_flight = max;
2146
2147         /* wakeup waiters if limit has been increased */
2148         if (cli->cl_max_mod_rpcs_in_flight > prev)
2149                 wake_up(&cli->cl_mod_rpcs_waitq);
2150
2151         spin_unlock(&cli->cl_mod_rpcs_lock);
2152
2153         return 0;
2154 }
2155 EXPORT_SYMBOL(obd_set_max_mod_rpcs_in_flight);
2156
2157
2158 #define pct(a, b) (b ? a * 100 / b : 0)
2159 int obd_mod_rpc_stats_seq_show(struct client_obd *cli,
2160                                struct seq_file *seq)
2161 {
2162         struct timeval now;
2163         unsigned long mod_tot = 0, mod_cum;
2164         int i;
2165
2166         do_gettimeofday(&now);
2167
2168         spin_lock(&cli->cl_mod_rpcs_lock);
2169
2170         seq_printf(seq, "snapshot_time:         %lu.%lu (secs.usecs)\n",
2171                    now.tv_sec, now.tv_usec);
2172         seq_printf(seq, "modify_RPCs_in_flight:  %hu\n",
2173                    cli->cl_mod_rpcs_in_flight);
2174
2175         seq_printf(seq, "\n\t\t\tmodify\n");
2176         seq_printf(seq, "rpcs in flight        rpcs   %% cum %%\n");
2177
2178         mod_tot = lprocfs_oh_sum(&cli->cl_mod_rpcs_hist);
2179
2180         mod_cum = 0;
2181         for (i = 0; i < OBD_HIST_MAX; i++) {
2182                 unsigned long mod = cli->cl_mod_rpcs_hist.oh_buckets[i];
2183                 mod_cum += mod;
2184                 seq_printf(seq, "%d:\t\t%10lu %3lu %3lu\n",
2185                                  i, mod, pct(mod, mod_tot),
2186                                  pct(mod_cum, mod_tot));
2187                 if (mod_cum == mod_tot)
2188                         break;
2189         }
2190
2191         spin_unlock(&cli->cl_mod_rpcs_lock);
2192
2193         return 0;
2194 }
2195 EXPORT_SYMBOL(obd_mod_rpc_stats_seq_show);
2196 #undef pct
2197
2198
2199 /* The number of modify RPCs sent in parallel is limited
2200  * because the server has a finite number of slots per client to
2201  * store request result and ensure reply reconstruction when needed.
2202  * On the client, this limit is stored in cl_max_mod_rpcs_in_flight
2203  * that takes into account server limit and cl_max_rpcs_in_flight
2204  * value.
2205  * On the MDC client, to avoid a potential deadlock (see Bugzilla 3462),
2206  * one close request is allowed above the maximum.
2207  */
2208 static inline bool obd_mod_rpc_slot_avail_locked(struct client_obd *cli,
2209                                                  bool close_req)
2210 {
2211         bool avail;
2212
2213         /* A slot is available if
2214          * - number of modify RPCs in flight is less than the max
2215          * - it's a close RPC and no other close request is in flight
2216          */
2217         avail = cli->cl_mod_rpcs_in_flight < cli->cl_max_mod_rpcs_in_flight ||
2218                 (close_req && cli->cl_close_rpcs_in_flight == 0);
2219
2220         return avail;
2221 }
2222
2223 static inline bool obd_mod_rpc_slot_avail(struct client_obd *cli,
2224                                          bool close_req)
2225 {
2226         bool avail;
2227
2228         spin_lock(&cli->cl_mod_rpcs_lock);
2229         avail = obd_mod_rpc_slot_avail_locked(cli, close_req);
2230         spin_unlock(&cli->cl_mod_rpcs_lock);
2231         return avail;
2232 }
2233
2234 /* Get a modify RPC slot from the obd client @cli according
2235  * to the kind of operation @opc that is going to be sent
2236  * and the intent @it of the operation if it applies.
2237  * If the maximum number of modify RPCs in flight is reached
2238  * the thread is put to sleep.
2239  * Returns the tag to be set in the request message. Tag 0
2240  * is reserved for non-modifying requests.
2241  */
2242 __u16 obd_get_mod_rpc_slot(struct client_obd *cli, __u32 opc,
2243                            struct lookup_intent *it)
2244 {
2245         struct l_wait_info      lwi = LWI_INTR(NULL, NULL);
2246         bool                    close_req = false;
2247         __u16                   i, max;
2248
2249         /* read-only metadata RPCs don't consume a slot on MDT
2250          * for reply reconstruction
2251          */
2252         if (it != NULL && (it->it_op == IT_GETATTR || it->it_op == IT_LOOKUP ||
2253                            it->it_op == IT_LAYOUT || it->it_op == IT_READDIR))
2254                 return 0;
2255
2256         if (opc == MDS_CLOSE)
2257                 close_req = true;
2258
2259         do {
2260                 spin_lock(&cli->cl_mod_rpcs_lock);
2261                 max = cli->cl_max_mod_rpcs_in_flight;
2262                 if (obd_mod_rpc_slot_avail_locked(cli, close_req)) {
2263                         /* there is a slot available */
2264                         cli->cl_mod_rpcs_in_flight++;
2265                         if (close_req)
2266                                 cli->cl_close_rpcs_in_flight++;
2267                         lprocfs_oh_tally(&cli->cl_mod_rpcs_hist,
2268                                          cli->cl_mod_rpcs_in_flight);
2269                         /* find a free tag */
2270                         i = find_first_zero_bit(cli->cl_mod_tag_bitmap,
2271                                                 max + 1);
2272                         LASSERT(i < OBD_MAX_RIF_MAX);
2273                         LASSERT(!test_and_set_bit(i, cli->cl_mod_tag_bitmap));
2274                         spin_unlock(&cli->cl_mod_rpcs_lock);
2275                         /* tag 0 is reserved for non-modify RPCs */
2276                         return i + 1;
2277                 }
2278                 spin_unlock(&cli->cl_mod_rpcs_lock);
2279
2280                 CDEBUG(D_RPCTRACE, "%s: sleeping for a modify RPC slot "
2281                        "opc %u, max %hu\n",
2282                        cli->cl_import->imp_obd->obd_name, opc, max);
2283
2284                 l_wait_event(cli->cl_mod_rpcs_waitq,
2285                              obd_mod_rpc_slot_avail(cli, close_req), &lwi);
2286         } while (true);
2287 }
2288 EXPORT_SYMBOL(obd_get_mod_rpc_slot);
2289
2290 /* Put a modify RPC slot from the obd client @cli according
2291  * to the kind of operation @opc that has been sent and the
2292  * intent @it of the operation if it applies.
2293  */
2294 void obd_put_mod_rpc_slot(struct client_obd *cli, __u32 opc,
2295                           struct lookup_intent *it, __u16 tag)
2296 {
2297         bool                    close_req = false;
2298
2299         if (it != NULL && (it->it_op == IT_GETATTR || it->it_op == IT_LOOKUP ||
2300                            it->it_op == IT_LAYOUT || it->it_op == IT_READDIR))
2301                 return;
2302
2303         if (opc == MDS_CLOSE)
2304                 close_req = true;
2305
2306         spin_lock(&cli->cl_mod_rpcs_lock);
2307         cli->cl_mod_rpcs_in_flight--;
2308         if (close_req)
2309                 cli->cl_close_rpcs_in_flight--;
2310         /* release the tag in the bitmap */
2311         LASSERT(tag - 1 < OBD_MAX_RIF_MAX);
2312         LASSERT(test_and_clear_bit(tag - 1, cli->cl_mod_tag_bitmap) != 0);
2313         spin_unlock(&cli->cl_mod_rpcs_lock);
2314         wake_up(&cli->cl_mod_rpcs_waitq);
2315 }
2316 EXPORT_SYMBOL(obd_put_mod_rpc_slot);
2317