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 %#llx\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         INIT_LIST_HEAD(&imp->imp_unreplied_list);
1064         imp->imp_known_replied_xid = 0;
1065         imp->imp_replay_cursor = &imp->imp_committed_list;
1066         spin_lock_init(&imp->imp_lock);
1067         imp->imp_last_success_conn = 0;
1068         imp->imp_state = LUSTRE_IMP_NEW;
1069         imp->imp_obd = class_incref(obd, "import", imp);
1070         mutex_init(&imp->imp_sec_mutex);
1071         init_waitqueue_head(&imp->imp_recovery_waitq);
1072
1073         atomic_set(&imp->imp_refcount, 2);
1074         atomic_set(&imp->imp_unregistering, 0);
1075         atomic_set(&imp->imp_inflight, 0);
1076         atomic_set(&imp->imp_replay_inflight, 0);
1077         atomic_set(&imp->imp_inval_count, 0);
1078         INIT_LIST_HEAD(&imp->imp_conn_list);
1079         INIT_LIST_HEAD(&imp->imp_handle.h_link);
1080         class_handle_hash(&imp->imp_handle, &import_handle_ops);
1081         init_imp_at(&imp->imp_at);
1082
1083         /* the default magic is V2, will be used in connect RPC, and
1084          * then adjusted according to the flags in request/reply. */
1085         imp->imp_msg_magic = LUSTRE_MSG_MAGIC_V2;
1086
1087         return imp;
1088 }
1089 EXPORT_SYMBOL(class_new_import);
1090
1091 void class_destroy_import(struct obd_import *import)
1092 {
1093         LASSERT(import != NULL);
1094         LASSERT(import != LP_POISON);
1095
1096         class_handle_unhash(&import->imp_handle);
1097
1098         spin_lock(&import->imp_lock);
1099         import->imp_generation++;
1100         spin_unlock(&import->imp_lock);
1101         class_import_put(import);
1102 }
1103 EXPORT_SYMBOL(class_destroy_import);
1104
1105 #if LUSTRE_TRACKS_LOCK_EXP_REFS
1106
1107 void __class_export_add_lock_ref(struct obd_export *exp, struct ldlm_lock *lock)
1108 {
1109         spin_lock(&exp->exp_locks_list_guard);
1110
1111         LASSERT(lock->l_exp_refs_nr >= 0);
1112
1113         if (lock->l_exp_refs_target != NULL &&
1114             lock->l_exp_refs_target != exp) {
1115                 LCONSOLE_WARN("setting export %p for lock %p which already has export %p\n",
1116                               exp, lock, lock->l_exp_refs_target);
1117         }
1118         if ((lock->l_exp_refs_nr ++) == 0) {
1119                 list_add(&lock->l_exp_refs_link, &exp->exp_locks_list);
1120                 lock->l_exp_refs_target = exp;
1121         }
1122         CDEBUG(D_INFO, "lock = %p, export = %p, refs = %u\n",
1123                lock, exp, lock->l_exp_refs_nr);
1124         spin_unlock(&exp->exp_locks_list_guard);
1125 }
1126 EXPORT_SYMBOL(__class_export_add_lock_ref);
1127
1128 void __class_export_del_lock_ref(struct obd_export *exp, struct ldlm_lock *lock)
1129 {
1130         spin_lock(&exp->exp_locks_list_guard);
1131         LASSERT(lock->l_exp_refs_nr > 0);
1132         if (lock->l_exp_refs_target != exp) {
1133                 LCONSOLE_WARN("lock %p, "
1134                               "mismatching export pointers: %p, %p\n",
1135                               lock, lock->l_exp_refs_target, exp);
1136         }
1137         if (-- lock->l_exp_refs_nr == 0) {
1138                 list_del_init(&lock->l_exp_refs_link);
1139                 lock->l_exp_refs_target = NULL;
1140         }
1141         CDEBUG(D_INFO, "lock = %p, export = %p, refs = %u\n",
1142                lock, exp, lock->l_exp_refs_nr);
1143         spin_unlock(&exp->exp_locks_list_guard);
1144 }
1145 EXPORT_SYMBOL(__class_export_del_lock_ref);
1146 #endif
1147
1148 /* A connection defines an export context in which preallocation can
1149    be managed. This releases the export pointer reference, and returns
1150    the export handle, so the export refcount is 1 when this function
1151    returns. */
1152 int class_connect(struct lustre_handle *conn, struct obd_device *obd,
1153                   struct obd_uuid *cluuid)
1154 {
1155         struct obd_export *export;
1156         LASSERT(conn != NULL);
1157         LASSERT(obd != NULL);
1158         LASSERT(cluuid != NULL);
1159         ENTRY;
1160
1161         export = class_new_export(obd, cluuid);
1162         if (IS_ERR(export))
1163                 RETURN(PTR_ERR(export));
1164
1165         conn->cookie = export->exp_handle.h_cookie;
1166         class_export_put(export);
1167
1168         CDEBUG(D_IOCTL, "connect: client %s, cookie %#llx\n",
1169                cluuid->uuid, conn->cookie);
1170         RETURN(0);
1171 }
1172 EXPORT_SYMBOL(class_connect);
1173
1174 /* if export is involved in recovery then clean up related things */
1175 static void class_export_recovery_cleanup(struct obd_export *exp)
1176 {
1177         struct obd_device *obd = exp->exp_obd;
1178
1179         spin_lock(&obd->obd_recovery_task_lock);
1180         if (obd->obd_recovering) {
1181                 if (exp->exp_in_recovery) {
1182                         spin_lock(&exp->exp_lock);
1183                         exp->exp_in_recovery = 0;
1184                         spin_unlock(&exp->exp_lock);
1185                         LASSERT_ATOMIC_POS(&obd->obd_connected_clients);
1186                         atomic_dec(&obd->obd_connected_clients);
1187                 }
1188
1189                 /* if called during recovery then should update
1190                  * obd_stale_clients counter,
1191                  * lightweight exports are not counted */
1192                 if ((exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT) == 0)
1193                         exp->exp_obd->obd_stale_clients++;
1194         }
1195         spin_unlock(&obd->obd_recovery_task_lock);
1196
1197         spin_lock(&exp->exp_lock);
1198         /** Cleanup req replay fields */
1199         if (exp->exp_req_replay_needed) {
1200                 exp->exp_req_replay_needed = 0;
1201
1202                 LASSERT(atomic_read(&obd->obd_req_replay_clients));
1203                 atomic_dec(&obd->obd_req_replay_clients);
1204         }
1205
1206         /** Cleanup lock replay data */
1207         if (exp->exp_lock_replay_needed) {
1208                 exp->exp_lock_replay_needed = 0;
1209
1210                 LASSERT(atomic_read(&obd->obd_lock_replay_clients));
1211                 atomic_dec(&obd->obd_lock_replay_clients);
1212         }
1213         spin_unlock(&exp->exp_lock);
1214 }
1215
1216 /* This function removes 1-3 references from the export:
1217  * 1 - for export pointer passed
1218  * and if disconnect really need
1219  * 2 - removing from hash
1220  * 3 - in client_unlink_export
1221  * The export pointer passed to this function can destroyed */
1222 int class_disconnect(struct obd_export *export)
1223 {
1224         int already_disconnected;
1225         ENTRY;
1226
1227         if (export == NULL) {
1228                 CWARN("attempting to free NULL export %p\n", export);
1229                 RETURN(-EINVAL);
1230         }
1231
1232         spin_lock(&export->exp_lock);
1233         already_disconnected = export->exp_disconnected;
1234         export->exp_disconnected = 1;
1235         spin_unlock(&export->exp_lock);
1236
1237         /* class_cleanup(), abort_recovery(), and class_fail_export()
1238          * all end up in here, and if any of them race we shouldn't
1239          * call extra class_export_puts(). */
1240         if (already_disconnected) {
1241                 LASSERT(hlist_unhashed(&export->exp_nid_hash));
1242                 GOTO(no_disconn, already_disconnected);
1243         }
1244
1245         CDEBUG(D_IOCTL, "disconnect: cookie %#llx\n",
1246                export->exp_handle.h_cookie);
1247
1248         if (!hlist_unhashed(&export->exp_nid_hash))
1249                 cfs_hash_del(export->exp_obd->obd_nid_hash,
1250                              &export->exp_connection->c_peer.nid,
1251                              &export->exp_nid_hash);
1252
1253         class_export_recovery_cleanup(export);
1254         class_unlink_export(export);
1255 no_disconn:
1256         class_export_put(export);
1257         RETURN(0);
1258 }
1259 EXPORT_SYMBOL(class_disconnect);
1260
1261 /* Return non-zero for a fully connected export */
1262 int class_connected_export(struct obd_export *exp)
1263 {
1264         int connected = 0;
1265
1266         if (exp) {
1267                 spin_lock(&exp->exp_lock);
1268                 connected = (exp->exp_conn_cnt > 0) && !exp->exp_failed;
1269                 spin_unlock(&exp->exp_lock);
1270         }
1271         return connected;
1272 }
1273 EXPORT_SYMBOL(class_connected_export);
1274
1275 static void class_disconnect_export_list(struct list_head *list,
1276                                          enum obd_option flags)
1277 {
1278         int rc;
1279         struct obd_export *exp;
1280         ENTRY;
1281
1282         /* It's possible that an export may disconnect itself, but
1283          * nothing else will be added to this list. */
1284         while (!list_empty(list)) {
1285                 exp = list_entry(list->next, struct obd_export,
1286                                  exp_obd_chain);
1287                 /* need for safe call CDEBUG after obd_disconnect */
1288                 class_export_get(exp);
1289
1290                 spin_lock(&exp->exp_lock);
1291                 exp->exp_flags = flags;
1292                 spin_unlock(&exp->exp_lock);
1293
1294                 if (obd_uuid_equals(&exp->exp_client_uuid,
1295                                     &exp->exp_obd->obd_uuid)) {
1296                         CDEBUG(D_HA,
1297                                "exp %p export uuid == obd uuid, don't discon\n",
1298                                exp);
1299                         /* Need to delete this now so we don't end up pointing
1300                          * to work_list later when this export is cleaned up. */
1301                         list_del_init(&exp->exp_obd_chain);
1302                         class_export_put(exp);
1303                         continue;
1304                 }
1305
1306                 class_export_get(exp);
1307                 CDEBUG(D_HA, "%s: disconnecting export at %s (%p), "
1308                        "last request at "CFS_TIME_T"\n",
1309                        exp->exp_obd->obd_name, obd_export_nid2str(exp),
1310                        exp, exp->exp_last_request_time);
1311                 /* release one export reference anyway */
1312                 rc = obd_disconnect(exp);
1313
1314                 CDEBUG(D_HA, "disconnected export at %s (%p): rc %d\n",
1315                        obd_export_nid2str(exp), exp, rc);
1316                 class_export_put(exp);
1317         }
1318         EXIT;
1319 }
1320
1321 void class_disconnect_exports(struct obd_device *obd)
1322 {
1323         struct list_head work_list;
1324         ENTRY;
1325
1326         /* Move all of the exports from obd_exports to a work list, en masse. */
1327         INIT_LIST_HEAD(&work_list);
1328         spin_lock(&obd->obd_dev_lock);
1329         list_splice_init(&obd->obd_exports, &work_list);
1330         list_splice_init(&obd->obd_delayed_exports, &work_list);
1331         spin_unlock(&obd->obd_dev_lock);
1332
1333         if (!list_empty(&work_list)) {
1334                 CDEBUG(D_HA, "OBD device %d (%p) has exports, "
1335                        "disconnecting them\n", obd->obd_minor, obd);
1336                 class_disconnect_export_list(&work_list,
1337                                              exp_flags_from_obd(obd));
1338         } else
1339                 CDEBUG(D_HA, "OBD device %d (%p) has no exports\n",
1340                        obd->obd_minor, obd);
1341         EXIT;
1342 }
1343 EXPORT_SYMBOL(class_disconnect_exports);
1344
1345 /* Remove exports that have not completed recovery.
1346  */
1347 void class_disconnect_stale_exports(struct obd_device *obd,
1348                                     int (*test_export)(struct obd_export *))
1349 {
1350         struct list_head work_list;
1351         struct obd_export *exp, *n;
1352         int evicted = 0;
1353         ENTRY;
1354
1355         INIT_LIST_HEAD(&work_list);
1356         spin_lock(&obd->obd_dev_lock);
1357         list_for_each_entry_safe(exp, n, &obd->obd_exports,
1358                                  exp_obd_chain) {
1359                 /* don't count self-export as client */
1360                 if (obd_uuid_equals(&exp->exp_client_uuid,
1361                                     &exp->exp_obd->obd_uuid))
1362                         continue;
1363
1364                 /* don't evict clients which have no slot in last_rcvd
1365                  * (e.g. lightweight connection) */
1366                 if (exp->exp_target_data.ted_lr_idx == -1)
1367                         continue;
1368
1369                 spin_lock(&exp->exp_lock);
1370                 if (exp->exp_failed || test_export(exp)) {
1371                         spin_unlock(&exp->exp_lock);
1372                         continue;
1373                 }
1374                 exp->exp_failed = 1;
1375                 spin_unlock(&exp->exp_lock);
1376
1377                 list_move(&exp->exp_obd_chain, &work_list);
1378                 evicted++;
1379                 CDEBUG(D_HA, "%s: disconnect stale client %s@%s\n",
1380                        obd->obd_name, exp->exp_client_uuid.uuid,
1381                        exp->exp_connection == NULL ? "<unknown>" :
1382                        libcfs_nid2str(exp->exp_connection->c_peer.nid));
1383                 print_export_data(exp, "EVICTING", 0);
1384         }
1385         spin_unlock(&obd->obd_dev_lock);
1386
1387         if (evicted)
1388                 LCONSOLE_WARN("%s: disconnecting %d stale clients\n",
1389                               obd->obd_name, evicted);
1390
1391         class_disconnect_export_list(&work_list, exp_flags_from_obd(obd) |
1392                                                  OBD_OPT_ABORT_RECOV);
1393         EXIT;
1394 }
1395 EXPORT_SYMBOL(class_disconnect_stale_exports);
1396
1397 void class_fail_export(struct obd_export *exp)
1398 {
1399         int rc, already_failed;
1400
1401         spin_lock(&exp->exp_lock);
1402         already_failed = exp->exp_failed;
1403         exp->exp_failed = 1;
1404         spin_unlock(&exp->exp_lock);
1405
1406         if (already_failed) {
1407                 CDEBUG(D_HA, "disconnecting dead export %p/%s; skipping\n",
1408                        exp, exp->exp_client_uuid.uuid);
1409                 return;
1410         }
1411
1412         CDEBUG(D_HA, "disconnecting export %p/%s\n",
1413                exp, exp->exp_client_uuid.uuid);
1414
1415         if (obd_dump_on_timeout)
1416                 libcfs_debug_dumplog();
1417
1418         /* need for safe call CDEBUG after obd_disconnect */
1419         class_export_get(exp);
1420
1421         /* Most callers into obd_disconnect are removing their own reference
1422          * (request, for example) in addition to the one from the hash table.
1423          * We don't have such a reference here, so make one. */
1424         class_export_get(exp);
1425         rc = obd_disconnect(exp);
1426         if (rc)
1427                 CERROR("disconnecting export %p failed: %d\n", exp, rc);
1428         else
1429                 CDEBUG(D_HA, "disconnected export %p/%s\n",
1430                        exp, exp->exp_client_uuid.uuid);
1431         class_export_put(exp);
1432 }
1433 EXPORT_SYMBOL(class_fail_export);
1434
1435 char *obd_export_nid2str(struct obd_export *exp)
1436 {
1437         if (exp->exp_connection != NULL)
1438                 return libcfs_nid2str(exp->exp_connection->c_peer.nid);
1439
1440         return "(no nid)";
1441 }
1442 EXPORT_SYMBOL(obd_export_nid2str);
1443
1444 int obd_export_evict_by_nid(struct obd_device *obd, const char *nid)
1445 {
1446         struct cfs_hash *nid_hash;
1447         struct obd_export *doomed_exp = NULL;
1448         int exports_evicted = 0;
1449
1450         lnet_nid_t nid_key = libcfs_str2nid((char *)nid);
1451
1452         spin_lock(&obd->obd_dev_lock);
1453         /* umount has run already, so evict thread should leave
1454          * its task to umount thread now */
1455         if (obd->obd_stopping) {
1456                 spin_unlock(&obd->obd_dev_lock);
1457                 return exports_evicted;
1458         }
1459         nid_hash = obd->obd_nid_hash;
1460         cfs_hash_getref(nid_hash);
1461         spin_unlock(&obd->obd_dev_lock);
1462
1463         do {
1464                 doomed_exp = cfs_hash_lookup(nid_hash, &nid_key);
1465                 if (doomed_exp == NULL)
1466                         break;
1467
1468                 LASSERTF(doomed_exp->exp_connection->c_peer.nid == nid_key,
1469                          "nid %s found, wanted nid %s, requested nid %s\n",
1470                          obd_export_nid2str(doomed_exp),
1471                          libcfs_nid2str(nid_key), nid);
1472                 LASSERTF(doomed_exp != obd->obd_self_export,
1473                          "self-export is hashed by NID?\n");
1474                 exports_evicted++;
1475                 LCONSOLE_WARN("%s: evicting %s (at %s) by administrative "
1476                               "request\n", obd->obd_name,
1477                               obd_uuid2str(&doomed_exp->exp_client_uuid),
1478                               obd_export_nid2str(doomed_exp));
1479                 class_fail_export(doomed_exp);
1480                 class_export_put(doomed_exp);
1481         } while (1);
1482
1483         cfs_hash_putref(nid_hash);
1484
1485         if (!exports_evicted)
1486                 CDEBUG(D_HA,"%s: can't disconnect NID '%s': no exports found\n",
1487                        obd->obd_name, nid);
1488         return exports_evicted;
1489 }
1490 EXPORT_SYMBOL(obd_export_evict_by_nid);
1491
1492 int obd_export_evict_by_uuid(struct obd_device *obd, const char *uuid)
1493 {
1494         struct cfs_hash *uuid_hash;
1495         struct obd_export *doomed_exp = NULL;
1496         struct obd_uuid doomed_uuid;
1497         int exports_evicted = 0;
1498
1499         spin_lock(&obd->obd_dev_lock);
1500         if (obd->obd_stopping) {
1501                 spin_unlock(&obd->obd_dev_lock);
1502                 return exports_evicted;
1503         }
1504         uuid_hash = obd->obd_uuid_hash;
1505         cfs_hash_getref(uuid_hash);
1506         spin_unlock(&obd->obd_dev_lock);
1507
1508         obd_str2uuid(&doomed_uuid, uuid);
1509         if (obd_uuid_equals(&doomed_uuid, &obd->obd_uuid)) {
1510                 CERROR("%s: can't evict myself\n", obd->obd_name);
1511                 cfs_hash_putref(uuid_hash);
1512                 return exports_evicted;
1513         }
1514
1515         doomed_exp = cfs_hash_lookup(uuid_hash, &doomed_uuid);
1516
1517         if (doomed_exp == NULL) {
1518                 CERROR("%s: can't disconnect %s: no exports found\n",
1519                        obd->obd_name, uuid);
1520         } else {
1521                 CWARN("%s: evicting %s at adminstrative request\n",
1522                        obd->obd_name, doomed_exp->exp_client_uuid.uuid);
1523                 class_fail_export(doomed_exp);
1524                 class_export_put(doomed_exp);
1525                 exports_evicted++;
1526         }
1527         cfs_hash_putref(uuid_hash);
1528
1529         return exports_evicted;
1530 }
1531
1532 #if LUSTRE_TRACKS_LOCK_EXP_REFS
1533 void (*class_export_dump_hook)(struct obd_export*) = NULL;
1534 EXPORT_SYMBOL(class_export_dump_hook);
1535 #endif
1536
1537 static void print_export_data(struct obd_export *exp, const char *status,
1538                               int locks)
1539 {
1540         struct ptlrpc_reply_state *rs;
1541         struct ptlrpc_reply_state *first_reply = NULL;
1542         int nreplies = 0;
1543
1544         spin_lock(&exp->exp_lock);
1545         list_for_each_entry(rs, &exp->exp_outstanding_replies,
1546                             rs_exp_list) {
1547                 if (nreplies == 0)
1548                         first_reply = rs;
1549                 nreplies++;
1550         }
1551         spin_unlock(&exp->exp_lock);
1552
1553         CDEBUG(D_HA, "%s: %s %p %s %s %d (%d %d %d) %d %d %d %d: %p %s %llu\n",
1554                exp->exp_obd->obd_name, status, exp, exp->exp_client_uuid.uuid,
1555                obd_export_nid2str(exp), atomic_read(&exp->exp_refcount),
1556                atomic_read(&exp->exp_rpc_count),
1557                atomic_read(&exp->exp_cb_count),
1558                atomic_read(&exp->exp_locks_count),
1559                exp->exp_disconnected, exp->exp_delayed, exp->exp_failed,
1560                nreplies, first_reply, nreplies > 3 ? "..." : "",
1561                exp->exp_last_committed);
1562 #if LUSTRE_TRACKS_LOCK_EXP_REFS
1563         if (locks && class_export_dump_hook != NULL)
1564                 class_export_dump_hook(exp);
1565 #endif
1566 }
1567
1568 void dump_exports(struct obd_device *obd, int locks)
1569 {
1570         struct obd_export *exp;
1571
1572         spin_lock(&obd->obd_dev_lock);
1573         list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain)
1574                 print_export_data(exp, "ACTIVE", locks);
1575         list_for_each_entry(exp, &obd->obd_unlinked_exports, exp_obd_chain)
1576                 print_export_data(exp, "UNLINKED", locks);
1577         list_for_each_entry(exp, &obd->obd_delayed_exports, exp_obd_chain)
1578                 print_export_data(exp, "DELAYED", locks);
1579         spin_unlock(&obd->obd_dev_lock);
1580         spin_lock(&obd_zombie_impexp_lock);
1581         list_for_each_entry(exp, &obd_zombie_exports, exp_obd_chain)
1582                 print_export_data(exp, "ZOMBIE", locks);
1583         spin_unlock(&obd_zombie_impexp_lock);
1584 }
1585
1586 void obd_exports_barrier(struct obd_device *obd)
1587 {
1588         int waited = 2;
1589         LASSERT(list_empty(&obd->obd_exports));
1590         spin_lock(&obd->obd_dev_lock);
1591         while (!list_empty(&obd->obd_unlinked_exports)) {
1592                 spin_unlock(&obd->obd_dev_lock);
1593                 set_current_state(TASK_UNINTERRUPTIBLE);
1594                 schedule_timeout(cfs_time_seconds(waited));
1595                 if (waited > 5 && IS_PO2(waited)) {
1596                         LCONSOLE_WARN("%s is waiting for obd_unlinked_exports "
1597                                       "more than %d seconds. "
1598                                       "The obd refcount = %d. Is it stuck?\n",
1599                                       obd->obd_name, waited,
1600                                       atomic_read(&obd->obd_refcount));
1601                         dump_exports(obd, 1);
1602                 }
1603                 waited *= 2;
1604                 spin_lock(&obd->obd_dev_lock);
1605         }
1606         spin_unlock(&obd->obd_dev_lock);
1607 }
1608 EXPORT_SYMBOL(obd_exports_barrier);
1609
1610 /* Total amount of zombies to be destroyed */
1611 static int zombies_count = 0;
1612
1613 /**
1614  * kill zombie imports and exports
1615  */
1616 void obd_zombie_impexp_cull(void)
1617 {
1618         struct obd_import *import;
1619         struct obd_export *export;
1620         ENTRY;
1621
1622         do {
1623                 spin_lock(&obd_zombie_impexp_lock);
1624
1625                 import = NULL;
1626                 if (!list_empty(&obd_zombie_imports)) {
1627                         import = list_entry(obd_zombie_imports.next,
1628                                             struct obd_import,
1629                                             imp_zombie_chain);
1630                         list_del_init(&import->imp_zombie_chain);
1631                 }
1632
1633                 export = NULL;
1634                 if (!list_empty(&obd_zombie_exports)) {
1635                         export = list_entry(obd_zombie_exports.next,
1636                                             struct obd_export,
1637                                             exp_obd_chain);
1638                         list_del_init(&export->exp_obd_chain);
1639                 }
1640
1641                 spin_unlock(&obd_zombie_impexp_lock);
1642
1643                 if (import != NULL) {
1644                         class_import_destroy(import);
1645                         spin_lock(&obd_zombie_impexp_lock);
1646                         zombies_count--;
1647                         spin_unlock(&obd_zombie_impexp_lock);
1648                 }
1649
1650                 if (export != NULL) {
1651                         class_export_destroy(export);
1652                         spin_lock(&obd_zombie_impexp_lock);
1653                         zombies_count--;
1654                         spin_unlock(&obd_zombie_impexp_lock);
1655                 }
1656
1657                 cond_resched();
1658         } while (import != NULL || export != NULL);
1659         EXIT;
1660 }
1661
1662 static struct completion        obd_zombie_start;
1663 static struct completion        obd_zombie_stop;
1664 static unsigned long            obd_zombie_flags;
1665 static wait_queue_head_t        obd_zombie_waitq;
1666 static pid_t                    obd_zombie_pid;
1667
1668 enum {
1669         OBD_ZOMBIE_STOP         = 0x0001,
1670 };
1671
1672 /**
1673  * check for work for kill zombie import/export thread.
1674  */
1675 static int obd_zombie_impexp_check(void *arg)
1676 {
1677         int rc;
1678
1679         spin_lock(&obd_zombie_impexp_lock);
1680         rc = (zombies_count == 0) &&
1681              !test_bit(OBD_ZOMBIE_STOP, &obd_zombie_flags);
1682         spin_unlock(&obd_zombie_impexp_lock);
1683
1684         RETURN(rc);
1685 }
1686
1687 /**
1688  * Add export to the obd_zombe thread and notify it.
1689  */
1690 static void obd_zombie_export_add(struct obd_export *exp) {
1691         atomic_dec(&obd_stale_export_num);
1692         spin_lock(&exp->exp_obd->obd_dev_lock);
1693         LASSERT(!list_empty(&exp->exp_obd_chain));
1694         list_del_init(&exp->exp_obd_chain);
1695         spin_unlock(&exp->exp_obd->obd_dev_lock);
1696         spin_lock(&obd_zombie_impexp_lock);
1697         zombies_count++;
1698         list_add(&exp->exp_obd_chain, &obd_zombie_exports);
1699         spin_unlock(&obd_zombie_impexp_lock);
1700
1701         obd_zombie_impexp_notify();
1702 }
1703
1704 /**
1705  * Add import to the obd_zombe thread and notify it.
1706  */
1707 static void obd_zombie_import_add(struct obd_import *imp) {
1708         LASSERT(imp->imp_sec == NULL);
1709         spin_lock(&obd_zombie_impexp_lock);
1710         LASSERT(list_empty(&imp->imp_zombie_chain));
1711         zombies_count++;
1712         list_add(&imp->imp_zombie_chain, &obd_zombie_imports);
1713         spin_unlock(&obd_zombie_impexp_lock);
1714
1715         obd_zombie_impexp_notify();
1716 }
1717
1718 /**
1719  * notify import/export destroy thread about new zombie.
1720  */
1721 static void obd_zombie_impexp_notify(void)
1722 {
1723         /*
1724          * Make sure obd_zomebie_impexp_thread get this notification.
1725          * It is possible this signal only get by obd_zombie_barrier, and
1726          * barrier gulps this notification and sleeps away and hangs ensues
1727          */
1728         wake_up_all(&obd_zombie_waitq);
1729 }
1730
1731 /**
1732  * check whether obd_zombie is idle
1733  */
1734 static int obd_zombie_is_idle(void)
1735 {
1736         int rc;
1737
1738         LASSERT(!test_bit(OBD_ZOMBIE_STOP, &obd_zombie_flags));
1739         spin_lock(&obd_zombie_impexp_lock);
1740         rc = (zombies_count == 0);
1741         spin_unlock(&obd_zombie_impexp_lock);
1742         return rc;
1743 }
1744
1745 /**
1746  * wait when obd_zombie import/export queues become empty
1747  */
1748 void obd_zombie_barrier(void)
1749 {
1750         struct l_wait_info lwi = { 0 };
1751
1752         if (obd_zombie_pid == current_pid())
1753                 /* don't wait for myself */
1754                 return;
1755         l_wait_event(obd_zombie_waitq, obd_zombie_is_idle(), &lwi);
1756 }
1757 EXPORT_SYMBOL(obd_zombie_barrier);
1758
1759
1760 struct obd_export *obd_stale_export_get(void)
1761 {
1762         struct obd_export *exp = NULL;
1763         ENTRY;
1764
1765         spin_lock(&obd_stale_export_lock);
1766         if (!list_empty(&obd_stale_exports)) {
1767                 exp = list_entry(obd_stale_exports.next,
1768                                  struct obd_export, exp_stale_list);
1769                 list_del_init(&exp->exp_stale_list);
1770         }
1771         spin_unlock(&obd_stale_export_lock);
1772
1773         if (exp) {
1774                 CDEBUG(D_DLMTRACE, "Get export %p: total %d\n", exp,
1775                        atomic_read(&obd_stale_export_num));
1776         }
1777         RETURN(exp);
1778 }
1779 EXPORT_SYMBOL(obd_stale_export_get);
1780
1781 void obd_stale_export_put(struct obd_export *exp)
1782 {
1783         ENTRY;
1784
1785         LASSERT(list_empty(&exp->exp_stale_list));
1786         if (exp->exp_lock_hash &&
1787             atomic_read(&exp->exp_lock_hash->hs_count)) {
1788                 CDEBUG(D_DLMTRACE, "Put export %p: total %d\n", exp,
1789                        atomic_read(&obd_stale_export_num));
1790
1791                 spin_lock_bh(&exp->exp_bl_list_lock);
1792                 spin_lock(&obd_stale_export_lock);
1793                 /* Add to the tail if there is no blocked locks,
1794                  * to the head otherwise. */
1795                 if (list_empty(&exp->exp_bl_list))
1796                         list_add_tail(&exp->exp_stale_list,
1797                                       &obd_stale_exports);
1798                 else
1799                         list_add(&exp->exp_stale_list,
1800                                  &obd_stale_exports);
1801
1802                 spin_unlock(&obd_stale_export_lock);
1803                 spin_unlock_bh(&exp->exp_bl_list_lock);
1804         } else {
1805                 class_export_put(exp);
1806         }
1807         EXIT;
1808 }
1809 EXPORT_SYMBOL(obd_stale_export_put);
1810
1811 /**
1812  * Adjust the position of the export in the stale list,
1813  * i.e. move to the head of the list if is needed.
1814  **/
1815 void obd_stale_export_adjust(struct obd_export *exp)
1816 {
1817         LASSERT(exp != NULL);
1818         spin_lock_bh(&exp->exp_bl_list_lock);
1819         spin_lock(&obd_stale_export_lock);
1820
1821         if (!list_empty(&exp->exp_stale_list) &&
1822             !list_empty(&exp->exp_bl_list))
1823                 list_move(&exp->exp_stale_list, &obd_stale_exports);
1824
1825         spin_unlock(&obd_stale_export_lock);
1826         spin_unlock_bh(&exp->exp_bl_list_lock);
1827 }
1828 EXPORT_SYMBOL(obd_stale_export_adjust);
1829
1830 /**
1831  * destroy zombie export/import thread.
1832  */
1833 static int obd_zombie_impexp_thread(void *unused)
1834 {
1835         unshare_fs_struct();
1836         complete(&obd_zombie_start);
1837
1838         obd_zombie_pid = current_pid();
1839
1840         while (!test_bit(OBD_ZOMBIE_STOP, &obd_zombie_flags)) {
1841                 struct l_wait_info lwi = { 0 };
1842
1843                 l_wait_event(obd_zombie_waitq,
1844                              !obd_zombie_impexp_check(NULL), &lwi);
1845                 obd_zombie_impexp_cull();
1846
1847                 /*
1848                  * Notify obd_zombie_barrier callers that queues
1849                  * may be empty.
1850                  */
1851                 wake_up(&obd_zombie_waitq);
1852         }
1853
1854         complete(&obd_zombie_stop);
1855
1856         RETURN(0);
1857 }
1858
1859
1860 /**
1861  * start destroy zombie import/export thread
1862  */
1863 int obd_zombie_impexp_init(void)
1864 {
1865         struct task_struct *task;
1866
1867         INIT_LIST_HEAD(&obd_zombie_imports);
1868
1869         INIT_LIST_HEAD(&obd_zombie_exports);
1870         spin_lock_init(&obd_zombie_impexp_lock);
1871         init_completion(&obd_zombie_start);
1872         init_completion(&obd_zombie_stop);
1873         init_waitqueue_head(&obd_zombie_waitq);
1874         obd_zombie_pid = 0;
1875
1876         task = kthread_run(obd_zombie_impexp_thread, NULL, "obd_zombid");
1877         if (IS_ERR(task))
1878                 RETURN(PTR_ERR(task));
1879
1880         wait_for_completion(&obd_zombie_start);
1881         RETURN(0);
1882 }
1883 /**
1884  * stop destroy zombie import/export thread
1885  */
1886 void obd_zombie_impexp_stop(void)
1887 {
1888         set_bit(OBD_ZOMBIE_STOP, &obd_zombie_flags);
1889         obd_zombie_impexp_notify();
1890         wait_for_completion(&obd_zombie_stop);
1891 }
1892
1893 /***** Kernel-userspace comm helpers *******/
1894
1895 /* Get length of entire message, including header */
1896 int kuc_len(int payload_len)
1897 {
1898         return sizeof(struct kuc_hdr) + payload_len;
1899 }
1900 EXPORT_SYMBOL(kuc_len);
1901
1902 /* Get a pointer to kuc header, given a ptr to the payload
1903  * @param p Pointer to payload area
1904  * @returns Pointer to kuc header
1905  */
1906 struct kuc_hdr * kuc_ptr(void *p)
1907 {
1908         struct kuc_hdr *lh = ((struct kuc_hdr *)p) - 1;
1909         LASSERT(lh->kuc_magic == KUC_MAGIC);
1910         return lh;
1911 }
1912 EXPORT_SYMBOL(kuc_ptr);
1913
1914 /* Alloc space for a message, and fill in header
1915  * @return Pointer to payload area
1916  */
1917 void *kuc_alloc(int payload_len, int transport, int type)
1918 {
1919         struct kuc_hdr *lh;
1920         int len = kuc_len(payload_len);
1921
1922         OBD_ALLOC(lh, len);
1923         if (lh == NULL)
1924                 return ERR_PTR(-ENOMEM);
1925
1926         lh->kuc_magic = KUC_MAGIC;
1927         lh->kuc_transport = transport;
1928         lh->kuc_msgtype = type;
1929         lh->kuc_msglen = len;
1930
1931         return (void *)(lh + 1);
1932 }
1933 EXPORT_SYMBOL(kuc_alloc);
1934
1935 /* Takes pointer to payload area */
1936 inline void kuc_free(void *p, int payload_len)
1937 {
1938         struct kuc_hdr *lh = kuc_ptr(p);
1939         OBD_FREE(lh, kuc_len(payload_len));
1940 }
1941 EXPORT_SYMBOL(kuc_free);
1942
1943 struct obd_request_slot_waiter {
1944         struct list_head        orsw_entry;
1945         wait_queue_head_t       orsw_waitq;
1946         bool                    orsw_signaled;
1947 };
1948
1949 static bool obd_request_slot_avail(struct client_obd *cli,
1950                                    struct obd_request_slot_waiter *orsw)
1951 {
1952         bool avail;
1953
1954         spin_lock(&cli->cl_loi_list_lock);
1955         avail = !!list_empty(&orsw->orsw_entry);
1956         spin_unlock(&cli->cl_loi_list_lock);
1957
1958         return avail;
1959 };
1960
1961 /*
1962  * For network flow control, the RPC sponsor needs to acquire a credit
1963  * before sending the RPC. The credits count for a connection is defined
1964  * by the "cl_max_rpcs_in_flight". If all the credits are occpuied, then
1965  * the subsequent RPC sponsors need to wait until others released their
1966  * credits, or the administrator increased the "cl_max_rpcs_in_flight".
1967  */
1968 int obd_get_request_slot(struct client_obd *cli)
1969 {
1970         struct obd_request_slot_waiter   orsw;
1971         struct l_wait_info               lwi;
1972         int                              rc;
1973
1974         spin_lock(&cli->cl_loi_list_lock);
1975         if (cli->cl_r_in_flight < cli->cl_max_rpcs_in_flight) {
1976                 cli->cl_r_in_flight++;
1977                 spin_unlock(&cli->cl_loi_list_lock);
1978                 return 0;
1979         }
1980
1981         init_waitqueue_head(&orsw.orsw_waitq);
1982         list_add_tail(&orsw.orsw_entry, &cli->cl_loi_read_list);
1983         orsw.orsw_signaled = false;
1984         spin_unlock(&cli->cl_loi_list_lock);
1985
1986         lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
1987         rc = l_wait_event(orsw.orsw_waitq,
1988                           obd_request_slot_avail(cli, &orsw) ||
1989                           orsw.orsw_signaled,
1990                           &lwi);
1991
1992         /* Here, we must take the lock to avoid the on-stack 'orsw' to be
1993          * freed but other (such as obd_put_request_slot) is using it. */
1994         spin_lock(&cli->cl_loi_list_lock);
1995         if (rc != 0) {
1996                 if (!orsw.orsw_signaled) {
1997                         if (list_empty(&orsw.orsw_entry))
1998                                 cli->cl_r_in_flight--;
1999                         else
2000                                 list_del(&orsw.orsw_entry);
2001                 }
2002         }
2003
2004         if (orsw.orsw_signaled) {
2005                 LASSERT(list_empty(&orsw.orsw_entry));
2006
2007                 rc = -EINTR;
2008         }
2009         spin_unlock(&cli->cl_loi_list_lock);
2010
2011         return rc;
2012 }
2013 EXPORT_SYMBOL(obd_get_request_slot);
2014
2015 void obd_put_request_slot(struct client_obd *cli)
2016 {
2017         struct obd_request_slot_waiter *orsw;
2018
2019         spin_lock(&cli->cl_loi_list_lock);
2020         cli->cl_r_in_flight--;
2021
2022         /* If there is free slot, wakeup the first waiter. */
2023         if (!list_empty(&cli->cl_loi_read_list) &&
2024             likely(cli->cl_r_in_flight < cli->cl_max_rpcs_in_flight)) {
2025                 orsw = list_entry(cli->cl_loi_read_list.next,
2026                                   struct obd_request_slot_waiter, orsw_entry);
2027                 list_del_init(&orsw->orsw_entry);
2028                 cli->cl_r_in_flight++;
2029                 wake_up(&orsw->orsw_waitq);
2030         }
2031         spin_unlock(&cli->cl_loi_list_lock);
2032 }
2033 EXPORT_SYMBOL(obd_put_request_slot);
2034
2035 __u32 obd_get_max_rpcs_in_flight(struct client_obd *cli)
2036 {
2037         return cli->cl_max_rpcs_in_flight;
2038 }
2039 EXPORT_SYMBOL(obd_get_max_rpcs_in_flight);
2040
2041 int obd_set_max_rpcs_in_flight(struct client_obd *cli, __u32 max)
2042 {
2043         struct obd_request_slot_waiter *orsw;
2044         __u32                           old;
2045         int                             diff;
2046         int                             i;
2047         char                            *typ_name;
2048         int                             rc;
2049
2050         if (max > OBD_MAX_RIF_MAX || max < 1)
2051                 return -ERANGE;
2052
2053         typ_name = cli->cl_import->imp_obd->obd_type->typ_name;
2054         if (strcmp(typ_name, LUSTRE_MDC_NAME) == 0) {
2055                 /* adjust max_mod_rpcs_in_flight to ensure it is always
2056                  * strictly lower that max_rpcs_in_flight */
2057                 if (max < 2) {
2058                         CERROR("%s: cannot set max_rpcs_in_flight to 1 "
2059                                "because it must be higher than "
2060                                "max_mod_rpcs_in_flight value",
2061                                cli->cl_import->imp_obd->obd_name);
2062                         return -ERANGE;
2063                 }
2064                 if (max <= cli->cl_max_mod_rpcs_in_flight) {
2065                         rc = obd_set_max_mod_rpcs_in_flight(cli, max - 1);
2066                         if (rc != 0)
2067                                 return rc;
2068                 }
2069         }
2070
2071         spin_lock(&cli->cl_loi_list_lock);
2072         old = cli->cl_max_rpcs_in_flight;
2073         cli->cl_max_rpcs_in_flight = max;
2074         diff = max - old;
2075
2076         /* We increase the max_rpcs_in_flight, then wakeup some waiters. */
2077         for (i = 0; i < diff; i++) {
2078                 if (list_empty(&cli->cl_loi_read_list))
2079                         break;
2080
2081                 orsw = list_entry(cli->cl_loi_read_list.next,
2082                                   struct obd_request_slot_waiter, orsw_entry);
2083                 list_del_init(&orsw->orsw_entry);
2084                 cli->cl_r_in_flight++;
2085                 wake_up(&orsw->orsw_waitq);
2086         }
2087         spin_unlock(&cli->cl_loi_list_lock);
2088
2089         return 0;
2090 }
2091 EXPORT_SYMBOL(obd_set_max_rpcs_in_flight);
2092
2093 __u16 obd_get_max_mod_rpcs_in_flight(struct client_obd *cli)
2094 {
2095         return cli->cl_max_mod_rpcs_in_flight;
2096 }
2097 EXPORT_SYMBOL(obd_get_max_mod_rpcs_in_flight);
2098
2099 int obd_set_max_mod_rpcs_in_flight(struct client_obd *cli, __u16 max)
2100 {
2101         struct obd_connect_data *ocd;
2102         __u16 maxmodrpcs;
2103         __u16 prev;
2104
2105         if (max > OBD_MAX_RIF_MAX || max < 1)
2106                 return -ERANGE;
2107
2108         /* cannot exceed or equal max_rpcs_in_flight */
2109         if (max >= cli->cl_max_rpcs_in_flight) {
2110                 CERROR("%s: can't set max_mod_rpcs_in_flight to a value (%hu) "
2111                        "higher or equal to max_rpcs_in_flight value (%u)\n",
2112                        cli->cl_import->imp_obd->obd_name,
2113                        max, cli->cl_max_rpcs_in_flight);
2114                 return -ERANGE;
2115         }
2116
2117         /* cannot exceed max modify RPCs in flight supported by the server */
2118         ocd = &cli->cl_import->imp_connect_data;
2119         if (ocd->ocd_connect_flags & OBD_CONNECT_MULTIMODRPCS)
2120                 maxmodrpcs = ocd->ocd_maxmodrpcs;
2121         else
2122                 maxmodrpcs = 1;
2123         if (max > maxmodrpcs) {
2124                 CERROR("%s: can't set max_mod_rpcs_in_flight to a value (%hu) "
2125                        "higher than max_mod_rpcs_per_client value (%hu) "
2126                        "returned by the server at connection\n",
2127                        cli->cl_import->imp_obd->obd_name,
2128                        max, maxmodrpcs);
2129                 return -ERANGE;
2130         }
2131
2132         spin_lock(&cli->cl_mod_rpcs_lock);
2133
2134         prev = cli->cl_max_mod_rpcs_in_flight;
2135         cli->cl_max_mod_rpcs_in_flight = max;
2136
2137         /* wakeup waiters if limit has been increased */
2138         if (cli->cl_max_mod_rpcs_in_flight > prev)
2139                 wake_up(&cli->cl_mod_rpcs_waitq);
2140
2141         spin_unlock(&cli->cl_mod_rpcs_lock);
2142
2143         return 0;
2144 }
2145 EXPORT_SYMBOL(obd_set_max_mod_rpcs_in_flight);
2146
2147
2148 #define pct(a, b) (b ? a * 100 / b : 0)
2149 int obd_mod_rpc_stats_seq_show(struct client_obd *cli,
2150                                struct seq_file *seq)
2151 {
2152         struct timeval now;
2153         unsigned long mod_tot = 0, mod_cum;
2154         int i;
2155
2156         do_gettimeofday(&now);
2157
2158         spin_lock(&cli->cl_mod_rpcs_lock);
2159
2160         seq_printf(seq, "snapshot_time:         %lu.%lu (secs.usecs)\n",
2161                    now.tv_sec, now.tv_usec);
2162         seq_printf(seq, "modify_RPCs_in_flight:  %hu\n",
2163                    cli->cl_mod_rpcs_in_flight);
2164
2165         seq_printf(seq, "\n\t\t\tmodify\n");
2166         seq_printf(seq, "rpcs in flight        rpcs   %% cum %%\n");
2167
2168         mod_tot = lprocfs_oh_sum(&cli->cl_mod_rpcs_hist);
2169
2170         mod_cum = 0;
2171         for (i = 0; i < OBD_HIST_MAX; i++) {
2172                 unsigned long mod = cli->cl_mod_rpcs_hist.oh_buckets[i];
2173                 mod_cum += mod;
2174                 seq_printf(seq, "%d:\t\t%10lu %3lu %3lu\n",
2175                            i, mod, pct(mod, mod_tot),
2176                            pct(mod_cum, mod_tot));
2177                 if (mod_cum == mod_tot)
2178                         break;
2179         }
2180
2181         spin_unlock(&cli->cl_mod_rpcs_lock);
2182
2183         return 0;
2184 }
2185 EXPORT_SYMBOL(obd_mod_rpc_stats_seq_show);
2186 #undef pct
2187
2188
2189 /* The number of modify RPCs sent in parallel is limited
2190  * because the server has a finite number of slots per client to
2191  * store request result and ensure reply reconstruction when needed.
2192  * On the client, this limit is stored in cl_max_mod_rpcs_in_flight
2193  * that takes into account server limit and cl_max_rpcs_in_flight
2194  * value.
2195  * On the MDC client, to avoid a potential deadlock (see Bugzilla 3462),
2196  * one close request is allowed above the maximum.
2197  */
2198 static inline bool obd_mod_rpc_slot_avail_locked(struct client_obd *cli,
2199                                                  bool close_req)
2200 {
2201         bool avail;
2202
2203         /* A slot is available if
2204          * - number of modify RPCs in flight is less than the max
2205          * - it's a close RPC and no other close request is in flight
2206          */
2207         avail = cli->cl_mod_rpcs_in_flight < cli->cl_max_mod_rpcs_in_flight ||
2208                 (close_req && cli->cl_close_rpcs_in_flight == 0);
2209
2210         return avail;
2211 }
2212
2213 static inline bool obd_mod_rpc_slot_avail(struct client_obd *cli,
2214                                          bool close_req)
2215 {
2216         bool avail;
2217
2218         spin_lock(&cli->cl_mod_rpcs_lock);
2219         avail = obd_mod_rpc_slot_avail_locked(cli, close_req);
2220         spin_unlock(&cli->cl_mod_rpcs_lock);
2221         return avail;
2222 }
2223
2224 /* Get a modify RPC slot from the obd client @cli according
2225  * to the kind of operation @opc that is going to be sent
2226  * and the intent @it of the operation if it applies.
2227  * If the maximum number of modify RPCs in flight is reached
2228  * the thread is put to sleep.
2229  * Returns the tag to be set in the request message. Tag 0
2230  * is reserved for non-modifying requests.
2231  */
2232 __u16 obd_get_mod_rpc_slot(struct client_obd *cli, __u32 opc,
2233                            struct lookup_intent *it)
2234 {
2235         struct l_wait_info      lwi = LWI_INTR(NULL, NULL);
2236         bool                    close_req = false;
2237         __u16                   i, max;
2238
2239         /* read-only metadata RPCs don't consume a slot on MDT
2240          * for reply reconstruction
2241          */
2242         if (it != NULL && (it->it_op == IT_GETATTR || it->it_op == IT_LOOKUP ||
2243                            it->it_op == IT_LAYOUT || it->it_op == IT_READDIR))
2244                 return 0;
2245
2246         if (opc == MDS_CLOSE)
2247                 close_req = true;
2248
2249         do {
2250                 spin_lock(&cli->cl_mod_rpcs_lock);
2251                 max = cli->cl_max_mod_rpcs_in_flight;
2252                 if (obd_mod_rpc_slot_avail_locked(cli, close_req)) {
2253                         /* there is a slot available */
2254                         cli->cl_mod_rpcs_in_flight++;
2255                         if (close_req)
2256                                 cli->cl_close_rpcs_in_flight++;
2257                         lprocfs_oh_tally(&cli->cl_mod_rpcs_hist,
2258                                          cli->cl_mod_rpcs_in_flight);
2259                         /* find a free tag */
2260                         i = find_first_zero_bit(cli->cl_mod_tag_bitmap,
2261                                                 max + 1);
2262                         LASSERT(i < OBD_MAX_RIF_MAX);
2263                         LASSERT(!test_and_set_bit(i, cli->cl_mod_tag_bitmap));
2264                         spin_unlock(&cli->cl_mod_rpcs_lock);
2265                         /* tag 0 is reserved for non-modify RPCs */
2266                         return i + 1;
2267                 }
2268                 spin_unlock(&cli->cl_mod_rpcs_lock);
2269
2270                 CDEBUG(D_RPCTRACE, "%s: sleeping for a modify RPC slot "
2271                        "opc %u, max %hu\n",
2272                        cli->cl_import->imp_obd->obd_name, opc, max);
2273
2274                 l_wait_event(cli->cl_mod_rpcs_waitq,
2275                              obd_mod_rpc_slot_avail(cli, close_req), &lwi);
2276         } while (true);
2277 }
2278 EXPORT_SYMBOL(obd_get_mod_rpc_slot);
2279
2280 /* Put a modify RPC slot from the obd client @cli according
2281  * to the kind of operation @opc that has been sent and the
2282  * intent @it of the operation if it applies.
2283  */
2284 void obd_put_mod_rpc_slot(struct client_obd *cli, __u32 opc,
2285                           struct lookup_intent *it, __u16 tag)
2286 {
2287         bool                    close_req = false;
2288
2289         if (it != NULL && (it->it_op == IT_GETATTR || it->it_op == IT_LOOKUP ||
2290                            it->it_op == IT_LAYOUT || it->it_op == IT_READDIR))
2291                 return;
2292
2293         if (opc == MDS_CLOSE)
2294                 close_req = true;
2295
2296         spin_lock(&cli->cl_mod_rpcs_lock);
2297         cli->cl_mod_rpcs_in_flight--;
2298         if (close_req)
2299                 cli->cl_close_rpcs_in_flight--;
2300         /* release the tag in the bitmap */
2301         LASSERT(tag - 1 < OBD_MAX_RIF_MAX);
2302         LASSERT(test_and_clear_bit(tag - 1, cli->cl_mod_tag_bitmap) != 0);
2303         spin_unlock(&cli->cl_mod_rpcs_lock);
2304         wake_up(&cli->cl_mod_rpcs_waitq);
2305 }
2306 EXPORT_SYMBOL(obd_put_mod_rpc_slot);
2307