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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 1999, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/obdclass/genops.c
  33  *
  34  * These are the only exported functions, they provide some generic
  35  * infrastructure for managing object devices
  36  */
  37
  38 #define DEBUG_SUBSYSTEM S_CLASS
  39
  40 #include <linux/pid_namespace.h>
  41 #include <linux/workqueue.h>
  42 #include <lustre_compat.h>
  43 #include <obd_class.h>
  44 #include <lustre_log.h>
  45 #include <lprocfs_status.h>
  46 #include <lustre_disk.h>
  47 #include <lustre_kernelcomm.h>
  48
  49 DEFINE_RWLOCK(obd_dev_lock);
  50 static struct obd_device *obd_devs[MAX_OBD_DEVICES];
  51
  52 static struct kmem_cache *obd_device_cachep;
  53 static struct kobj_type class_ktype;
  54 static struct workqueue_struct *zombie_wq;
  55
  56 static void obd_zombie_export_add(struct obd_export *exp);
  57 static void obd_zombie_import_add(struct obd_import *imp);
  58 static void print_export_data(struct obd_export *exp,
  59                               const char *status, int locks, int debug_level);
  60
  61 static LIST_HEAD(obd_stale_exports);
  62 static DEFINE_SPINLOCK(obd_stale_export_lock);
  63 static atomic_t obd_stale_export_num = ATOMIC_INIT(0);
  64
  65 int (*ptlrpc_put_connection_superhack)(struct ptlrpc_connection *c);
  66 EXPORT_SYMBOL(ptlrpc_put_connection_superhack);
  67
  68 /*
  69  * support functions: we could use inter-module communication, but this
  70  * is more portable to other OS's
  71  */
  72 static struct obd_device *obd_device_alloc(void)
  73 {
  74         struct obd_device *obd;
  75
  76         OBD_SLAB_ALLOC_PTR_GFP(obd, obd_device_cachep, GFP_NOFS);
  77         if (obd != NULL) {
  78                 obd->obd_magic = OBD_DEVICE_MAGIC;
  79         }
  80         return obd;
  81 }
  82
  83 static void obd_device_free(struct obd_device *obd)
  84 {
  85         LASSERT(obd != NULL);
  86         LASSERTF(obd->obd_magic == OBD_DEVICE_MAGIC, "obd %p obd_magic %08x != %08x\n",
  87                  obd, obd->obd_magic, OBD_DEVICE_MAGIC);
  88         if (obd->obd_namespace != NULL) {
  89                 CERROR("obd %p: namespace %p was not properly cleaned up (obd_force=%d)!\n",
  90                        obd, obd->obd_namespace, obd->obd_force);
  91                 LBUG();
  92         }
  93         lu_ref_fini(&obd->obd_reference);
  94         OBD_SLAB_FREE_PTR(obd, obd_device_cachep);
  95 }
  96
  97 struct obd_type *class_search_type(const char *name)
  98 {
  99         struct kobject *kobj = kset_find_obj(lustre_kset, name);
 100
 101         if (kobj && kobj->ktype == &class_ktype)
 102                 return container_of(kobj, struct obd_type, typ_kobj);
 103
 104         kobject_put(kobj);
 105         return NULL;
 106 }
 107 EXPORT_SYMBOL(class_search_type);
 108
 109 struct obd_type *class_get_type(const char *name)
 110 {
 111         struct obd_type *type;
 112
 113         type = class_search_type(name);
 114 #ifdef HAVE_MODULE_LOADING_SUPPORT
 115         if (!type) {
 116                 const char *modname = name;
 117
 118 #ifdef HAVE_SERVER_SUPPORT
 119                 if (strcmp(modname, "obdfilter") == 0)
 120                         modname = "ofd";
 121
 122                 if (strcmp(modname, LUSTRE_LWP_NAME) == 0)
 123                         modname = LUSTRE_OSP_NAME;
 124
 125                 if (!strncmp(modname, LUSTRE_MDS_NAME, strlen(LUSTRE_MDS_NAME)))
 126                         modname = LUSTRE_MDT_NAME;
 127 #endif /* HAVE_SERVER_SUPPORT */
 128
 129                 if (!request_module("%s", modname)) {
 130                         CDEBUG(D_INFO, "Loaded module '%s'\n", modname);
 131                         type = class_search_type(name);
 132                 } else {
 133                         LCONSOLE_ERROR_MSG(0x158, "Can't load module '%s'\n",
 134                                            modname);
 135                 }
 136         }
 137 #endif
 138         if (type) {
 139                 if (try_module_get(type->typ_dt_ops->o_owner)) {
 140                         atomic_inc(&type->typ_refcnt);
 141                         /* class_search_type() returned a counted reference,
 142                          * but we don't need that count any more as
 143                          * we have one through typ_refcnt.
 144                          */
 145                         kobject_put(&type->typ_kobj);
 146                 } else {
 147                         kobject_put(&type->typ_kobj);
 148                         type = NULL;
 149                 }
 150         }
 151         return type;
 152 }
 153
 154 void class_put_type(struct obd_type *type)
 155 {
 156         LASSERT(type);
 157         module_put(type->typ_dt_ops->o_owner);
 158         atomic_dec(&type->typ_refcnt);
 159 }
 160
 161 static void class_sysfs_release(struct kobject *kobj)
 162 {
 163         struct obd_type *type = container_of(kobj, struct obd_type, typ_kobj);
 164
 165         debugfs_remove_recursive(type->typ_debugfs_entry);
 166         type->typ_debugfs_entry = NULL;
 167
 168         if (type->typ_lu)
 169                 lu_device_type_fini(type->typ_lu);
 170
 171 #ifdef CONFIG_PROC_FS
 172         if (type->typ_name && type->typ_procroot)
 173                 remove_proc_subtree(type->typ_name, proc_lustre_root);
 174 #endif
 175         OBD_FREE(type, sizeof(*type));
 176 }
 177
 178 static struct kobj_type class_ktype = {
 179         .sysfs_ops      = &lustre_sysfs_ops,
 180         .release        = class_sysfs_release,
 181 };
 182
 183 #ifdef HAVE_SERVER_SUPPORT
 184 struct obd_type *class_add_symlinks(const char *name, bool enable_proc)
 185 {
 186         struct dentry *symlink;
 187         struct obd_type *type;
 188         int rc;
 189
 190         type = class_search_type(name);
 191         if (type) {
 192                 kobject_put(&type->typ_kobj);
 193                 return ERR_PTR(-EEXIST);
 194         }
 195
 196         OBD_ALLOC(type, sizeof(*type));
 197         if (!type)
 198                 return ERR_PTR(-ENOMEM);
 199
 200         type->typ_kobj.kset = lustre_kset;
 201         rc = kobject_init_and_add(&type->typ_kobj, &class_ktype,
 202                                   &lustre_kset->kobj, "%s", name);
 203         if (rc)
 204                 return ERR_PTR(rc);
 205
 206         symlink = debugfs_create_dir(name, debugfs_lustre_root);
 207         type->typ_debugfs_entry = symlink;
 208         type->typ_sym_filter = true;
 209
 210         if (enable_proc) {
 211                 type->typ_procroot = lprocfs_register(name, proc_lustre_root,
 212                                                       NULL, NULL);
 213                 if (IS_ERR(type->typ_procroot)) {
 214                         CERROR("%s: can't create compat proc entry: %d\n",
 215                                name, (int)PTR_ERR(type->typ_procroot));
 216                         type->typ_procroot = NULL;
 217                 }
 218         }
 219
 220         return type;
 221 }
 222 EXPORT_SYMBOL(class_add_symlinks);
 223 #endif /* HAVE_SERVER_SUPPORT */
 224
 225 #define CLASS_MAX_NAME 1024
 226
 227 int class_register_type(const struct obd_ops *dt_ops,
 228                         const struct md_ops *md_ops,
 229                         bool enable_proc, struct lprocfs_vars *vars,
 230                         const char *name, struct lu_device_type *ldt)
 231 {
 232         struct obd_type *type;
 233         int rc;
 234
 235         ENTRY;
 236         /* sanity check */
 237         LASSERT(strnlen(name, CLASS_MAX_NAME) < CLASS_MAX_NAME);
 238
 239         type = class_search_type(name);
 240         if (type) {
 241 #ifdef HAVE_SERVER_SUPPORT
 242                 if (type->typ_sym_filter)
 243                         goto dir_exist;
 244 #endif /* HAVE_SERVER_SUPPORT */
 245                 kobject_put(&type->typ_kobj);
 246                 CDEBUG(D_IOCTL, "Type %s already registered\n", name);
 247                 RETURN(-EEXIST);
 248         }
 249
 250         OBD_ALLOC(type, sizeof(*type));
 251         if (type == NULL)
 252                 RETURN(-ENOMEM);
 253
 254         type->typ_lu = ldt ? OBD_LU_TYPE_SETUP : NULL;
 255         type->typ_kobj.kset = lustre_kset;
 256         kobject_init(&type->typ_kobj, &class_ktype);
 257 #ifdef HAVE_SERVER_SUPPORT
 258 dir_exist:
 259 #endif /* HAVE_SERVER_SUPPORT */
 260
 261         type->typ_dt_ops = dt_ops;
 262         type->typ_md_ops = md_ops;
 263
 264 #ifdef HAVE_SERVER_SUPPORT
 265         if (type->typ_sym_filter) {
 266                 type->typ_sym_filter = false;
 267                 kobject_put(&type->typ_kobj);
 268                 goto setup_ldt;
 269         }
 270 #endif
 271 #ifdef CONFIG_PROC_FS
 272         if (enable_proc && !type->typ_procroot) {
 273                 type->typ_procroot = lprocfs_register(name,
 274                                                       proc_lustre_root,
 275                                                       NULL, type);
 276                 if (IS_ERR(type->typ_procroot)) {
 277                         rc = PTR_ERR(type->typ_procroot);
 278                         type->typ_procroot = NULL;
 279                         GOTO(failed, rc);
 280                 }
 281         }
 282 #endif
 283         type->typ_debugfs_entry = debugfs_create_dir(name, debugfs_lustre_root);
 284         ldebugfs_add_vars(type->typ_debugfs_entry, vars, type);
 285
 286         rc = kobject_add(&type->typ_kobj, &lustre_kset->kobj, "%s", name);
 287         if (rc)
 288                 GOTO(failed, rc);
 289 #ifdef HAVE_SERVER_SUPPORT
 290 setup_ldt:
 291 #endif
 292         if (ldt) {
 293                 rc = lu_device_type_init(ldt);
 294                 smp_store_release(&type->typ_lu, rc ? NULL : ldt);
 295                 wake_up_var(&type->typ_lu);
 296                 if (rc)
 297                         GOTO(failed, rc);
 298         }
 299
 300         RETURN(0);
 301
 302 failed:
 303         kobject_put(&type->typ_kobj);
 304
 305         RETURN(rc);
 306 }
 307 EXPORT_SYMBOL(class_register_type);
 308
 309 int class_unregister_type(const char *name)
 310 {
 311         struct obd_type *type = class_search_type(name);
 312         int rc = 0;
 313         ENTRY;
 314
 315         if (!type) {
 316                 CERROR("unknown obd type\n");
 317                 RETURN(-EINVAL);
 318         }
 319
 320         if (atomic_read(&type->typ_refcnt)) {
 321                 CERROR("type %s has refcount (%d)\n", name,
 322                        atomic_read(&type->typ_refcnt));
 323                 /* This is a bad situation, let's make the best of it */
 324                 /* Remove ops, but leave the name for debugging */
 325                 type->typ_dt_ops = NULL;
 326                 type->typ_md_ops = NULL;
 327                 GOTO(out_put, rc = -EBUSY);
 328         }
 329
 330         /* Put the final ref */
 331         kobject_put(&type->typ_kobj);
 332 out_put:
 333         /* Put the ref returned by class_search_type() */
 334         kobject_put(&type->typ_kobj);
 335
 336         RETURN(rc);
 337 } /* class_unregister_type */
 338 EXPORT_SYMBOL(class_unregister_type);
 339
 340 /**
 341  * Create a new obd device.
 342  *
 343  * Allocate the new obd_device and initialize it.
 344  *
 345  * \param[in] type_name obd device type string.
 346  * \param[in] name      obd device name.
 347  * \param[in] uuid      obd device UUID
 348  *
 349  * \retval newdev         pointer to created obd_device
 350  * \retval ERR_PTR(errno) on error
 351  */
 352 struct obd_device *class_newdev(const char *type_name, const char *name,
 353                                 const char *uuid)
 354 {
 355         struct obd_device *newdev;
 356         struct obd_type *type = NULL;
 357         ENTRY;
 358
 359         if (strlen(name) >= MAX_OBD_NAME) {
 360                 CERROR("name/uuid must be < %u bytes long\n", MAX_OBD_NAME);
 361                 RETURN(ERR_PTR(-EINVAL));
 362         }
 363
 364         type = class_get_type(type_name);
 365         if (type == NULL){
 366                 CERROR("OBD: unknown type: %s\n", type_name);
 367                 RETURN(ERR_PTR(-ENODEV));
 368         }
 369
 370         newdev = obd_device_alloc();
 371         if (newdev == NULL) {
 372                 class_put_type(type);
 373                 RETURN(ERR_PTR(-ENOMEM));
 374         }
 375         LASSERT(newdev->obd_magic == OBD_DEVICE_MAGIC);
 376         strncpy(newdev->obd_name, name, sizeof(newdev->obd_name) - 1);
 377         newdev->obd_type = type;
 378         newdev->obd_minor = -1;
 379
 380         rwlock_init(&newdev->obd_pool_lock);
 381         newdev->obd_pool_limit = 0;
 382         newdev->obd_pool_slv = 0;
 383
 384         INIT_LIST_HEAD(&newdev->obd_exports);
 385         INIT_LIST_HEAD(&newdev->obd_unlinked_exports);
 386         INIT_LIST_HEAD(&newdev->obd_delayed_exports);
 387         INIT_LIST_HEAD(&newdev->obd_exports_timed);
 388         INIT_LIST_HEAD(&newdev->obd_nid_stats);
 389         spin_lock_init(&newdev->obd_nid_lock);
 390         spin_lock_init(&newdev->obd_dev_lock);
 391         mutex_init(&newdev->obd_dev_mutex);
 392         spin_lock_init(&newdev->obd_osfs_lock);
 393         /* newdev->obd_osfs_age must be set to a value in the distant
 394          * past to guarantee a fresh statfs is fetched on mount. */
 395         newdev->obd_osfs_age = ktime_get_seconds() - 1000;
 396
 397         /* XXX belongs in setup not attach  */
 398         init_rwsem(&newdev->obd_observer_link_sem);
 399         /* recovery data */
 400         spin_lock_init(&newdev->obd_recovery_task_lock);
 401         init_waitqueue_head(&newdev->obd_next_transno_waitq);
 402         init_waitqueue_head(&newdev->obd_evict_inprogress_waitq);
 403         INIT_LIST_HEAD(&newdev->obd_req_replay_queue);
 404         INIT_LIST_HEAD(&newdev->obd_lock_replay_queue);
 405         INIT_LIST_HEAD(&newdev->obd_final_req_queue);
 406         INIT_LIST_HEAD(&newdev->obd_evict_list);
 407         INIT_LIST_HEAD(&newdev->obd_lwp_list);
 408
 409         llog_group_init(&newdev->obd_olg);
 410         /* Detach drops this */
 411         atomic_set(&newdev->obd_refcount, 1);
 412         lu_ref_init(&newdev->obd_reference);
 413         lu_ref_add(&newdev->obd_reference, "newdev", newdev);
 414
 415         newdev->obd_conn_inprogress = 0;
 416
 417         strncpy(newdev->obd_uuid.uuid, uuid, UUID_MAX);
 418
 419         CDEBUG(D_IOCTL, "Allocate new device %s (%p)\n",
 420                newdev->obd_name, newdev);
 421
 422         return newdev;
 423 }
 424
 425 /**
 426  * Free obd device.
 427  *
 428  * \param[in] obd obd_device to be freed
 429  *
 430  * \retval none
 431  */
 432 void class_free_dev(struct obd_device *obd)
 433 {
 434         struct obd_type *obd_type = obd->obd_type;
 435
 436         LASSERTF(obd->obd_magic == OBD_DEVICE_MAGIC, "%p obd_magic %08x "
 437                  "!= %08x\n", obd, obd->obd_magic, OBD_DEVICE_MAGIC);
 438         LASSERTF(obd->obd_minor == -1 || obd_devs[obd->obd_minor] == obd,
 439                  "obd %p != obd_devs[%d] %p\n",
 440                  obd, obd->obd_minor, obd_devs[obd->obd_minor]);
 441         LASSERTF(atomic_read(&obd->obd_refcount) == 0,
 442                  "obd_refcount should be 0, not %d\n",
 443                  atomic_read(&obd->obd_refcount));
 444         LASSERT(obd_type != NULL);
 445
 446         CDEBUG(D_INFO, "Release obd device %s obd_type name = %s\n",
 447                obd->obd_name, obd->obd_type->typ_name);
 448
 449         CDEBUG(D_CONFIG, "finishing cleanup of obd %s (%s)\n",
 450                          obd->obd_name, obd->obd_uuid.uuid);
 451         if (obd->obd_stopping) {
 452                 int err;
 453
 454                 /* If we're not stopping, we were never set up */
 455                 err = obd_cleanup(obd);
 456                 if (err)
 457                         CERROR("Cleanup %s returned %d\n",
 458                                 obd->obd_name, err);
 459         }
 460
 461         obd_device_free(obd);
 462
 463         class_put_type(obd_type);
 464 }
 465
 466 /**
 467  * Unregister obd device.
 468  *
 469  * Free slot in obd_dev[] used by \a obd.
 470  *
 471  * \param[in] new_obd obd_device to be unregistered
 472  *
 473  * \retval none
 474  */
 475 void class_unregister_device(struct obd_device *obd)
 476 {
 477         write_lock(&obd_dev_lock);
 478         if (obd->obd_minor >= 0) {
 479                 LASSERT(obd_devs[obd->obd_minor] == obd);
 480                 obd_devs[obd->obd_minor] = NULL;
 481                 obd->obd_minor = -1;
 482         }
 483         write_unlock(&obd_dev_lock);
 484 }
 485
 486 /**
 487  * Register obd device.
 488  *
 489  * Find free slot in obd_devs[], fills it with \a new_obd.
 490  *
 491  * \param[in] new_obd obd_device to be registered
 492  *
 493  * \retval 0          success
 494  * \retval -EEXIST    device with this name is registered
 495  * \retval -EOVERFLOW obd_devs[] is full
 496  */
 497 int class_register_device(struct obd_device *new_obd)
 498 {
 499         int ret = 0;
 500         int i;
 501         int new_obd_minor = 0;
 502         bool minor_assign = false;
 503         bool retried = false;
 504
 505 again:
 506         write_lock(&obd_dev_lock);
 507         for (i = 0; i < class_devno_max(); i++) {
 508                 struct obd_device *obd = class_num2obd(i);
 509
 510                 if (obd != NULL &&
 511                     (strcmp(new_obd->obd_name, obd->obd_name) == 0)) {
 512
 513                         if (!retried) {
 514                                 write_unlock(&obd_dev_lock);
 515
 516                                 /* the obd_device could be waited to be
 517                                  * destroyed by the "obd_zombie_impexp_thread".
 518                                  */
 519                                 obd_zombie_barrier();
 520                                 retried = true;
 521                                 goto again;
 522                         }
 523
 524                         CERROR("%s: already exists, won't add\n",
 525                                obd->obd_name);
 526                         /* in case we found a free slot before duplicate */
 527                         minor_assign = false;
 528                         ret = -EEXIST;
 529                         break;
 530                 }
 531                 if (!minor_assign && obd == NULL) {
 532                         new_obd_minor = i;
 533                         minor_assign = true;
 534                 }
 535         }
 536
 537         if (minor_assign) {
 538                 new_obd->obd_minor = new_obd_minor;
 539                 LASSERTF(obd_devs[new_obd_minor] == NULL, "obd_devs[%d] "
 540                          "%p\n", new_obd_minor, obd_devs[new_obd_minor]);
 541                 obd_devs[new_obd_minor] = new_obd;
 542         } else {
 543                 if (ret == 0) {
 544                         ret = -EOVERFLOW;
 545                         CERROR("%s: all %u/%u devices used, increase "
 546                                "MAX_OBD_DEVICES: rc = %d\n", new_obd->obd_name,
 547                                i, class_devno_max(), ret);
 548                 }
 549         }
 550         write_unlock(&obd_dev_lock);
 551
 552         RETURN(ret);
 553 }
 554
 555 static int class_name2dev_nolock(const char *name)
 556 {
 557         int i;
 558
 559         if (!name)
 560                 return -1;
 561
 562         for (i = 0; i < class_devno_max(); i++) {
 563                 struct obd_device *obd = class_num2obd(i);
 564
 565                 if (obd && strcmp(name, obd->obd_name) == 0) {
 566                         /* Make sure we finished attaching before we give
 567                            out any references */
 568                         LASSERT(obd->obd_magic == OBD_DEVICE_MAGIC);
 569                         if (obd->obd_attached) {
 570                                 return i;
 571                         }
 572                         break;
 573                 }
 574         }
 575
 576         return -1;
 577 }
 578
 579 int class_name2dev(const char *name)
 580 {
 581         int i;
 582
 583         if (!name)
 584                 return -1;
 585
 586         read_lock(&obd_dev_lock);
 587         i = class_name2dev_nolock(name);
 588         read_unlock(&obd_dev_lock);
 589
 590         return i;
 591 }
 592 EXPORT_SYMBOL(class_name2dev);
 593
 594 struct obd_device *class_name2obd(const char *name)
 595 {
 596         int dev = class_name2dev(name);
 597
 598         if (dev < 0 || dev > class_devno_max())
 599                 return NULL;
 600         return class_num2obd(dev);
 601 }
 602 EXPORT_SYMBOL(class_name2obd);
 603
 604 int class_uuid2dev_nolock(struct obd_uuid *uuid)
 605 {
 606         int i;
 607
 608         for (i = 0; i < class_devno_max(); i++) {
 609                 struct obd_device *obd = class_num2obd(i);
 610
 611                 if (obd && obd_uuid_equals(uuid, &obd->obd_uuid)) {
 612                         LASSERT(obd->obd_magic == OBD_DEVICE_MAGIC);
 613                         return i;
 614                 }
 615         }
 616
 617         return -1;
 618 }
 619
 620 int class_uuid2dev(struct obd_uuid *uuid)
 621 {
 622         int i;
 623
 624         read_lock(&obd_dev_lock);
 625         i = class_uuid2dev_nolock(uuid);
 626         read_unlock(&obd_dev_lock);
 627
 628         return i;
 629 }
 630 EXPORT_SYMBOL(class_uuid2dev);
 631
 632 struct obd_device *class_uuid2obd(struct obd_uuid *uuid)
 633 {
 634         int dev = class_uuid2dev(uuid);
 635         if (dev < 0)
 636                 return NULL;
 637         return class_num2obd(dev);
 638 }
 639 EXPORT_SYMBOL(class_uuid2obd);
 640
 641 /**
 642  * Get obd device from ::obd_devs[]
 643  *
 644  * \param num [in] array index
 645  *
 646  * \retval NULL if ::obd_devs[\a num] does not contains an obd device
 647  *         otherwise return the obd device there.
 648  */
 649 struct obd_device *class_num2obd(int num)
 650 {
 651         struct obd_device *obd = NULL;
 652
 653         if (num < class_devno_max()) {
 654                 obd = obd_devs[num];
 655                 if (obd == NULL)
 656                         return NULL;
 657
 658                 LASSERTF(obd->obd_magic == OBD_DEVICE_MAGIC,
 659                          "%p obd_magic %08x != %08x\n",
 660                          obd, obd->obd_magic, OBD_DEVICE_MAGIC);
 661                 LASSERTF(obd->obd_minor == num,
 662                          "%p obd_minor %0d != %0d\n",
 663                          obd, obd->obd_minor, num);
 664         }
 665
 666         return obd;
 667 }
 668 EXPORT_SYMBOL(class_num2obd);
 669
 670 /**
 671  * Find obd in obd_dev[] by name or uuid.
 672  *
 673  * Increment obd's refcount if found.
 674  *
 675  * \param[in] str obd name or uuid
 676  *
 677  * \retval NULL    if not found
 678  * \retval target  pointer to found obd_device
 679  */
 680 struct obd_device *class_dev_by_str(const char *str)
 681 {
 682         struct obd_device *target = NULL;
 683         struct obd_uuid tgtuuid;
 684         int rc;
 685
 686         obd_str2uuid(&tgtuuid, str);
 687
 688         read_lock(&obd_dev_lock);
 689         rc = class_uuid2dev_nolock(&tgtuuid);
 690         if (rc < 0)
 691                 rc = class_name2dev_nolock(str);
 692
 693         if (rc >= 0)
 694                 target = class_num2obd(rc);
 695
 696         if (target != NULL)
 697                 class_incref(target, "find", current);
 698         read_unlock(&obd_dev_lock);
 699
 700         RETURN(target);
 701 }
 702 EXPORT_SYMBOL(class_dev_by_str);
 703
 704 /**
 705  * Get obd devices count. Device in any
 706  *    state are counted
 707  * \retval obd device count
 708  */
 709 int get_devices_count(void)
 710 {
 711         int index, max_index = class_devno_max(), dev_count = 0;
 712
 713         read_lock(&obd_dev_lock);
 714         for (index = 0; index <= max_index; index++) {
 715                 struct obd_device *obd = class_num2obd(index);
 716                 if (obd != NULL)
 717                         dev_count++;
 718         }
 719         read_unlock(&obd_dev_lock);
 720
 721         return dev_count;
 722 }
 723 EXPORT_SYMBOL(get_devices_count);
 724
 725 void class_obd_list(void)
 726 {
 727         char *status;
 728         int i;
 729
 730         read_lock(&obd_dev_lock);
 731         for (i = 0; i < class_devno_max(); i++) {
 732                 struct obd_device *obd = class_num2obd(i);
 733
 734                 if (obd == NULL)
 735                         continue;
 736                 if (obd->obd_stopping)
 737                         status = "ST";
 738                 else if (obd->obd_set_up)
 739                         status = "UP";
 740                 else if (obd->obd_attached)
 741                         status = "AT";
 742                 else
 743                         status = "--";
 744                 LCONSOLE(D_CONFIG, "%3d %s %s %s %s %d\n",
 745                          i, status, obd->obd_type->typ_name,
 746                          obd->obd_name, obd->obd_uuid.uuid,
 747                          atomic_read(&obd->obd_refcount));
 748         }
 749         read_unlock(&obd_dev_lock);
 750 }
 751
 752 /* Search for a client OBD connected to tgt_uuid.  If grp_uuid is
 753  * specified, then only the client with that uuid is returned,
 754  * otherwise any client connected to the tgt is returned.
 755  */
 756 struct obd_device *class_find_client_obd(struct obd_uuid *tgt_uuid,
 757                                          const char *type_name,
 758                                          struct obd_uuid *grp_uuid)
 759 {
 760         int i;
 761
 762         read_lock(&obd_dev_lock);
 763         for (i = 0; i < class_devno_max(); i++) {
 764                 struct obd_device *obd = class_num2obd(i);
 765
 766                 if (obd == NULL)
 767                         continue;
 768                 if ((strncmp(obd->obd_type->typ_name, type_name,
 769                              strlen(type_name)) == 0)) {
 770                         if (obd_uuid_equals(tgt_uuid,
 771                                             &obd->u.cli.cl_target_uuid) &&
 772                             ((grp_uuid)? obd_uuid_equals(grp_uuid,
 773                                                          &obd->obd_uuid) : 1)) {
 774                                 read_unlock(&obd_dev_lock);
 775                                 return obd;
 776                         }
 777                 }
 778         }
 779         read_unlock(&obd_dev_lock);
 780
 781         return NULL;
 782 }
 783 EXPORT_SYMBOL(class_find_client_obd);
 784
 785 /* Iterate the obd_device list looking devices have grp_uuid. Start
 786  * searching at *next, and if a device is found, the next index to look
 787  * at is saved in *next. If next is NULL, then the first matching device
 788  * will always be returned.
 789  */
 790 struct obd_device *class_devices_in_group(struct obd_uuid *grp_uuid, int *next)
 791 {
 792         int i;
 793
 794         if (next == NULL)
 795                 i = 0;
 796         else if (*next >= 0 && *next < class_devno_max())
 797                 i = *next;
 798         else
 799                 return NULL;
 800
 801         read_lock(&obd_dev_lock);
 802         for (; i < class_devno_max(); i++) {
 803                 struct obd_device *obd = class_num2obd(i);
 804
 805                 if (obd == NULL)
 806                         continue;
 807                 if (obd_uuid_equals(grp_uuid, &obd->obd_uuid)) {
 808                         if (next != NULL)
 809                                 *next = i+1;
 810                         read_unlock(&obd_dev_lock);
 811                         return obd;
 812                 }
 813         }
 814         read_unlock(&obd_dev_lock);
 815
 816         return NULL;
 817 }
 818 EXPORT_SYMBOL(class_devices_in_group);
 819
 820 /**
 821  * to notify sptlrpc log for \a fsname has changed, let every relevant OBD
 822  * adjust sptlrpc settings accordingly.
 823  */
 824 int class_notify_sptlrpc_conf(const char *fsname, int namelen)
 825 {
 826         struct obd_device  *obd;
 827         const char         *type;
 828         int                 i, rc = 0, rc2;
 829
 830         LASSERT(namelen > 0);
 831
 832         read_lock(&obd_dev_lock);
 833         for (i = 0; i < class_devno_max(); i++) {
 834                 obd = class_num2obd(i);
 835
 836                 if (obd == NULL || obd->obd_set_up == 0 || obd->obd_stopping)
 837                         continue;
 838
 839                 /* only notify mdc, osc, osp, lwp, mdt, ost
 840                  * because only these have a -sptlrpc llog */
 841                 type = obd->obd_type->typ_name;
 842                 if (strcmp(type, LUSTRE_MDC_NAME) != 0 &&
 843                     strcmp(type, LUSTRE_OSC_NAME) != 0 &&
 844                     strcmp(type, LUSTRE_OSP_NAME) != 0 &&
 845                     strcmp(type, LUSTRE_LWP_NAME) != 0 &&
 846                     strcmp(type, LUSTRE_MDT_NAME) != 0 &&
 847                     strcmp(type, LUSTRE_OST_NAME) != 0)
 848                         continue;
 849
 850                 if (strncmp(obd->obd_name, fsname, namelen))
 851                         continue;
 852
 853                 class_incref(obd, __FUNCTION__, obd);
 854                 read_unlock(&obd_dev_lock);
 855                 rc2 = obd_set_info_async(NULL, obd->obd_self_export,
 856                                          sizeof(KEY_SPTLRPC_CONF),
 857                                          KEY_SPTLRPC_CONF, 0, NULL, NULL);
 858                 rc = rc ? rc : rc2;
 859                 class_decref(obd, __FUNCTION__, obd);
 860                 read_lock(&obd_dev_lock);
 861         }
 862         read_unlock(&obd_dev_lock);
 863         return rc;
 864 }
 865 EXPORT_SYMBOL(class_notify_sptlrpc_conf);
 866
 867 void obd_cleanup_caches(void)
 868 {
 869         ENTRY;
 870         if (obd_device_cachep) {
 871                 kmem_cache_destroy(obd_device_cachep);
 872                 obd_device_cachep = NULL;
 873         }
 874
 875         EXIT;
 876 }
 877
 878 int obd_init_caches(void)
 879 {
 880         int rc;
 881         ENTRY;
 882
 883         LASSERT(obd_device_cachep == NULL);
 884         obd_device_cachep = kmem_cache_create_usercopy("ll_obd_dev_cache",
 885                                 sizeof(struct obd_device),
 886                                 0, 0, 0, sizeof(struct obd_device), NULL);
 887         if (!obd_device_cachep)
 888                 GOTO(out, rc = -ENOMEM);
 889
 890         RETURN(0);
 891 out:
 892         obd_cleanup_caches();
 893         RETURN(rc);
 894 }
 895
 896 static const char export_handle_owner[] = "export";
 897
 898 /* map connection to client */
 899 struct obd_export *class_conn2export(struct lustre_handle *conn)
 900 {
 901         struct obd_export *export;
 902         ENTRY;
 903
 904         if (!conn) {
 905                 CDEBUG(D_CACHE, "looking for null handle\n");
 906                 RETURN(NULL);
 907         }
 908
 909         if (conn->cookie == -1) {  /* this means assign a new connection */
 910                 CDEBUG(D_CACHE, "want a new connection\n");
 911                 RETURN(NULL);
 912         }
 913
 914         CDEBUG(D_INFO, "looking for export cookie %#llx\n", conn->cookie);
 915         export = class_handle2object(conn->cookie, export_handle_owner);
 916         RETURN(export);
 917 }
 918 EXPORT_SYMBOL(class_conn2export);
 919
 920 struct obd_device *class_exp2obd(struct obd_export *exp)
 921 {
 922         if (exp)
 923                 return exp->exp_obd;
 924         return NULL;
 925 }
 926 EXPORT_SYMBOL(class_exp2obd);
 927
 928 struct obd_import *class_exp2cliimp(struct obd_export *exp)
 929 {
 930         struct obd_device *obd = exp->exp_obd;
 931         if (obd == NULL)
 932                 return NULL;
 933         return obd->u.cli.cl_import;
 934 }
 935 EXPORT_SYMBOL(class_exp2cliimp);
 936
 937 /* Export management functions */
 938 static void class_export_destroy(struct obd_export *exp)
 939 {
 940         struct obd_device *obd = exp->exp_obd;
 941         ENTRY;
 942
 943         LASSERT(refcount_read(&exp->exp_handle.h_ref) == 0);
 944         LASSERT(obd != NULL);
 945
 946         CDEBUG(D_IOCTL, "destroying export %p/%s for %s\n", exp,
 947                exp->exp_client_uuid.uuid, obd->obd_name);
 948
 949         /* "Local" exports (lctl, LOV->{mdc,osc}) have no connection. */
 950         if (exp->exp_connection)
 951                 ptlrpc_put_connection_superhack(exp->exp_connection);
 952
 953         LASSERT(list_empty(&exp->exp_outstanding_replies));
 954         LASSERT(list_empty(&exp->exp_uncommitted_replies));
 955         LASSERT(list_empty(&exp->exp_req_replay_queue));
 956         LASSERT(list_empty(&exp->exp_hp_rpcs));
 957         obd_destroy_export(exp);
 958         /* self export doesn't hold a reference to an obd, although it
 959          * exists until freeing of the obd */
 960         if (exp != obd->obd_self_export)
 961                 class_decref(obd, "export", exp);
 962
 963         OBD_FREE_PRE(exp, sizeof(*exp), "rcu");
 964         kfree_rcu(exp, exp_handle.h_rcu);
 965         EXIT;
 966 }
 967
 968 struct obd_export *class_export_get(struct obd_export *exp)
 969 {
 970         refcount_inc(&exp->exp_handle.h_ref);
 971         CDEBUG(D_INFO, "GET export %p refcount=%d\n", exp,
 972                refcount_read(&exp->exp_handle.h_ref));
 973         return exp;
 974 }
 975 EXPORT_SYMBOL(class_export_get);
 976
 977 void class_export_put(struct obd_export *exp)
 978 {
 979         LASSERT(exp != NULL);
 980         LASSERT(refcount_read(&exp->exp_handle.h_ref) >  0);
 981         LASSERT(refcount_read(&exp->exp_handle.h_ref) < LI_POISON);
 982         CDEBUG(D_INFO, "PUTting export %p : new refcount %d\n", exp,
 983                refcount_read(&exp->exp_handle.h_ref) - 1);
 984
 985         if (refcount_dec_and_test(&exp->exp_handle.h_ref)) {
 986                 struct obd_device *obd = exp->exp_obd;
 987
 988                 CDEBUG(D_IOCTL, "final put %p/%s\n",
 989                        exp, exp->exp_client_uuid.uuid);
 990
 991                 /* release nid stat refererence */
 992                 lprocfs_exp_cleanup(exp);
 993
 994                 if (exp == obd->obd_self_export) {
 995                         /* self export should be destroyed without
 996                          * zombie thread as it doesn't hold a
 997                          * reference to obd and doesn't hold any
 998                          * resources */
 999                         class_export_destroy(exp);
1000                         /* self export is destroyed, no class
1001                          * references exist and it is safe to free
1002                          * obd */
1003                         class_free_dev(obd);
1004                 } else {
1005                         LASSERT(!list_empty(&exp->exp_obd_chain));
1006                         obd_zombie_export_add(exp);
1007                 }
1008
1009         }
1010 }
1011 EXPORT_SYMBOL(class_export_put);
1012
1013 static void obd_zombie_exp_cull(struct work_struct *ws)
1014 {
1015         struct obd_export *export;
1016
1017         export = container_of(ws, struct obd_export, exp_zombie_work);
1018         class_export_destroy(export);
1019 }
1020
1021 /* Creates a new export, adds it to the hash table, and returns a
1022  * pointer to it. The refcount is 2: one for the hash reference, and
1023  * one for the pointer returned by this function. */
1024 struct obd_export *__class_new_export(struct obd_device *obd,
1025                                       struct obd_uuid *cluuid, bool is_self)
1026 {
1027         struct obd_export *export;
1028         int rc = 0;
1029         ENTRY;
1030
1031         OBD_ALLOC_PTR(export);
1032         if (!export)
1033                 return ERR_PTR(-ENOMEM);
1034
1035         export->exp_conn_cnt = 0;
1036         export->exp_lock_hash = NULL;
1037         export->exp_flock_hash = NULL;
1038         /* 2 = class_handle_hash + last */
1039         refcount_set(&export->exp_handle.h_ref, 2);
1040         atomic_set(&export->exp_rpc_count, 0);
1041         atomic_set(&export->exp_cb_count, 0);
1042         atomic_set(&export->exp_locks_count, 0);
1043 #if LUSTRE_TRACKS_LOCK_EXP_REFS
1044         INIT_LIST_HEAD(&export->exp_locks_list);
1045         spin_lock_init(&export->exp_locks_list_guard);
1046 #endif
1047         atomic_set(&export->exp_replay_count, 0);
1048         export->exp_obd = obd;
1049         INIT_LIST_HEAD(&export->exp_outstanding_replies);
1050         spin_lock_init(&export->exp_uncommitted_replies_lock);
1051         INIT_LIST_HEAD(&export->exp_uncommitted_replies);
1052         INIT_LIST_HEAD(&export->exp_req_replay_queue);
1053         INIT_HLIST_NODE(&export->exp_handle.h_link);
1054         INIT_LIST_HEAD(&export->exp_hp_rpcs);
1055         INIT_LIST_HEAD(&export->exp_reg_rpcs);
1056         class_handle_hash(&export->exp_handle, export_handle_owner);
1057         export->exp_last_request_time = ktime_get_real_seconds();
1058         spin_lock_init(&export->exp_lock);
1059         spin_lock_init(&export->exp_rpc_lock);
1060         INIT_HLIST_NODE(&export->exp_gen_hash);
1061         spin_lock_init(&export->exp_bl_list_lock);
1062         INIT_LIST_HEAD(&export->exp_bl_list);
1063         INIT_LIST_HEAD(&export->exp_stale_list);
1064         INIT_WORK(&export->exp_zombie_work, obd_zombie_exp_cull);
1065
1066         export->exp_sp_peer = LUSTRE_SP_ANY;
1067         export->exp_flvr.sf_rpc = SPTLRPC_FLVR_INVALID;
1068         export->exp_client_uuid = *cluuid;
1069         obd_init_export(export);
1070
1071         at_init(&export->exp_bl_lock_at, obd_timeout, 0);
1072
1073         spin_lock(&obd->obd_dev_lock);
1074         if (!obd_uuid_equals(cluuid, &obd->obd_uuid)) {
1075                 /* shouldn't happen, but might race */
1076                 if (obd->obd_stopping)
1077                         GOTO(exit_unlock, rc = -ENODEV);
1078
1079                 rc = obd_uuid_add(obd, export);
1080                 if (rc != 0) {
1081                         LCONSOLE_WARN("%s: denying duplicate export for %s: rc = %d\n",
1082                                       obd->obd_name, cluuid->uuid, rc);
1083                         GOTO(exit_unlock, rc = -EALREADY);
1084                 }
1085         }
1086
1087         if (!is_self) {
1088                 class_incref(obd, "export", export);
1089                 list_add_tail(&export->exp_obd_chain_timed,
1090                               &obd->obd_exports_timed);
1091                 list_add(&export->exp_obd_chain, &obd->obd_exports);
1092                 obd->obd_num_exports++;
1093         } else {
1094                 INIT_LIST_HEAD(&export->exp_obd_chain_timed);
1095                 INIT_LIST_HEAD(&export->exp_obd_chain);
1096         }
1097         spin_unlock(&obd->obd_dev_lock);
1098         RETURN(export);
1099
1100 exit_unlock:
1101         spin_unlock(&obd->obd_dev_lock);
1102         class_handle_unhash(&export->exp_handle);
1103         obd_destroy_export(export);
1104         OBD_FREE_PTR(export);
1105         return ERR_PTR(rc);
1106 }
1107
1108 struct obd_export *class_new_export(struct obd_device *obd,
1109                                     struct obd_uuid *uuid)
1110 {
1111         return __class_new_export(obd, uuid, false);
1112 }
1113 EXPORT_SYMBOL(class_new_export);
1114
1115 struct obd_export *class_new_export_self(struct obd_device *obd,
1116                                          struct obd_uuid *uuid)
1117 {
1118         return __class_new_export(obd, uuid, true);
1119 }
1120
1121 void class_unlink_export(struct obd_export *exp)
1122 {
1123         class_handle_unhash(&exp->exp_handle);
1124
1125         if (exp->exp_obd->obd_self_export == exp) {
1126                 class_export_put(exp);
1127                 return;
1128         }
1129
1130         spin_lock(&exp->exp_obd->obd_dev_lock);
1131         /* delete an uuid-export hashitem from hashtables */
1132         if (exp != exp->exp_obd->obd_self_export)
1133                 obd_uuid_del(exp->exp_obd, exp);
1134
1135 #ifdef HAVE_SERVER_SUPPORT
1136         if (!hlist_unhashed(&exp->exp_gen_hash)) {
1137                 struct tg_export_data   *ted = &exp->exp_target_data;
1138                 struct cfs_hash         *hash;
1139
1140                 /* Because obd_gen_hash will not be released until
1141                  * class_cleanup(), so hash should never be NULL here */
1142                 hash = cfs_hash_getref(exp->exp_obd->obd_gen_hash);
1143                 LASSERT(hash != NULL);
1144                 cfs_hash_del(hash, &ted->ted_lcd->lcd_generation,
1145                              &exp->exp_gen_hash);
1146                 cfs_hash_putref(hash);
1147         }
1148 #endif /* HAVE_SERVER_SUPPORT */
1149
1150         list_move(&exp->exp_obd_chain, &exp->exp_obd->obd_unlinked_exports);
1151         list_del_init(&exp->exp_obd_chain_timed);
1152         exp->exp_obd->obd_num_exports--;
1153         spin_unlock(&exp->exp_obd->obd_dev_lock);
1154         atomic_inc(&obd_stale_export_num);
1155
1156         /* A reference is kept by obd_stale_exports list */
1157         obd_stale_export_put(exp);
1158 }
1159 EXPORT_SYMBOL(class_unlink_export);
1160
1161 /* Import management functions */
1162 static void obd_zombie_import_free(struct obd_import *imp)
1163 {
1164         ENTRY;
1165
1166         CDEBUG(D_IOCTL, "destroying import %p for %s\n", imp,
1167                 imp->imp_obd->obd_name);
1168
1169         LASSERT(refcount_read(&imp->imp_refcount) == 0);
1170
1171         ptlrpc_put_connection_superhack(imp->imp_connection);
1172
1173         while (!list_empty(&imp->imp_conn_list)) {
1174                 struct obd_import_conn *imp_conn;
1175
1176                 imp_conn = list_first_entry(&imp->imp_conn_list,
1177                                             struct obd_import_conn, oic_item);
1178                 list_del_init(&imp_conn->oic_item);
1179                 ptlrpc_put_connection_superhack(imp_conn->oic_conn);
1180                 OBD_FREE(imp_conn, sizeof(*imp_conn));
1181         }
1182
1183         LASSERT(imp->imp_sec == NULL);
1184         LASSERTF(atomic_read(&imp->imp_reqs) == 0, "%s: imp_reqs = %d\n",
1185                  imp->imp_obd->obd_name, atomic_read(&imp->imp_reqs));
1186         class_decref(imp->imp_obd, "import", imp);
1187         OBD_FREE_PTR(imp);
1188         EXIT;
1189 }
1190
1191 struct obd_import *class_import_get(struct obd_import *import)
1192 {
1193         refcount_inc(&import->imp_refcount);
1194         CDEBUG(D_INFO, "import %p refcount=%d obd=%s\n", import,
1195                refcount_read(&import->imp_refcount),
1196                import->imp_obd->obd_name);
1197         return import;
1198 }
1199 EXPORT_SYMBOL(class_import_get);
1200
1201 void class_import_put(struct obd_import *imp)
1202 {
1203         ENTRY;
1204
1205         LASSERT(refcount_read(&imp->imp_refcount) > 0);
1206
1207         CDEBUG(D_INFO, "import %p refcount=%d obd=%s\n", imp,
1208                refcount_read(&imp->imp_refcount) - 1,
1209                imp->imp_obd->obd_name);
1210
1211         if (refcount_dec_and_test(&imp->imp_refcount)) {
1212                 CDEBUG(D_INFO, "final put import %p\n", imp);
1213                 obd_zombie_import_add(imp);
1214         }
1215
1216         EXIT;
1217 }
1218 EXPORT_SYMBOL(class_import_put);
1219
1220 static void init_imp_at(struct imp_at *at) {
1221         int i;
1222         at_init(&at->iat_net_latency, 0, 0);
1223         for (i = 0; i < IMP_AT_MAX_PORTALS; i++) {
1224                 /* max service estimates are tracked on the server side, so
1225                    don't use the AT history here, just use the last reported
1226                    val. (But keep hist for proc histogram, worst_ever) */
1227                 at_init(&at->iat_service_estimate[i], INITIAL_CONNECT_TIMEOUT,
1228                         AT_FLG_NOHIST);
1229         }
1230 }
1231
1232 static void obd_zombie_imp_cull(struct work_struct *ws)
1233 {
1234         struct obd_import *import;
1235
1236         import = container_of(ws, struct obd_import, imp_zombie_work);
1237         obd_zombie_import_free(import);
1238 }
1239
1240 struct obd_import *class_new_import(struct obd_device *obd)
1241 {
1242         struct obd_import *imp;
1243         struct pid_namespace *curr_pid_ns = ll_task_pid_ns(current);
1244
1245         OBD_ALLOC(imp, sizeof(*imp));
1246         if (imp == NULL)
1247                 return NULL;
1248
1249         INIT_LIST_HEAD(&imp->imp_pinger_chain);
1250         INIT_LIST_HEAD(&imp->imp_replay_list);
1251         INIT_LIST_HEAD(&imp->imp_sending_list);
1252         INIT_LIST_HEAD(&imp->imp_delayed_list);
1253         INIT_LIST_HEAD(&imp->imp_committed_list);
1254         INIT_LIST_HEAD(&imp->imp_unreplied_list);
1255         imp->imp_known_replied_xid = 0;
1256         imp->imp_replay_cursor = &imp->imp_committed_list;
1257         spin_lock_init(&imp->imp_lock);
1258         imp->imp_last_success_conn = 0;
1259         imp->imp_state = LUSTRE_IMP_NEW;
1260         imp->imp_obd = class_incref(obd, "import", imp);
1261         rwlock_init(&imp->imp_sec_lock);
1262         init_waitqueue_head(&imp->imp_recovery_waitq);
1263         INIT_WORK(&imp->imp_zombie_work, obd_zombie_imp_cull);
1264
1265         if (curr_pid_ns && curr_pid_ns->child_reaper)
1266                 imp->imp_sec_refpid = curr_pid_ns->child_reaper->pid;
1267         else
1268                 imp->imp_sec_refpid = 1;
1269
1270         refcount_set(&imp->imp_refcount, 2);
1271         atomic_set(&imp->imp_unregistering, 0);
1272         atomic_set(&imp->imp_reqs, 0);
1273         atomic_set(&imp->imp_inflight, 0);
1274         atomic_set(&imp->imp_replay_inflight, 0);
1275         atomic_set(&imp->imp_inval_count, 0);
1276         INIT_LIST_HEAD(&imp->imp_conn_list);
1277         init_imp_at(&imp->imp_at);
1278
1279         /* the default magic is V2, will be used in connect RPC, and
1280          * then adjusted according to the flags in request/reply. */
1281         imp->imp_msg_magic = LUSTRE_MSG_MAGIC_V2;
1282
1283         return imp;
1284 }
1285 EXPORT_SYMBOL(class_new_import);
1286
1287 void class_destroy_import(struct obd_import *import)
1288 {
1289         LASSERT(import != NULL);
1290         LASSERT(import != LP_POISON);
1291
1292         spin_lock(&import->imp_lock);
1293         import->imp_generation++;
1294         spin_unlock(&import->imp_lock);
1295         class_import_put(import);
1296 }
1297 EXPORT_SYMBOL(class_destroy_import);
1298
1299 #if LUSTRE_TRACKS_LOCK_EXP_REFS
1300
1301 void __class_export_add_lock_ref(struct obd_export *exp, struct ldlm_lock *lock)
1302 {
1303         spin_lock(&exp->exp_locks_list_guard);
1304
1305         LASSERT(lock->l_exp_refs_nr >= 0);
1306
1307         if (lock->l_exp_refs_target != NULL &&
1308             lock->l_exp_refs_target != exp) {
1309                 LCONSOLE_WARN("setting export %p for lock %p which already has export %p\n",
1310                               exp, lock, lock->l_exp_refs_target);
1311         }
1312         if ((lock->l_exp_refs_nr ++) == 0) {
1313                 list_add(&lock->l_exp_refs_link, &exp->exp_locks_list);
1314                 lock->l_exp_refs_target = exp;
1315         }
1316         CDEBUG(D_INFO, "lock = %p, export = %p, refs = %u\n",
1317                lock, exp, lock->l_exp_refs_nr);
1318         spin_unlock(&exp->exp_locks_list_guard);
1319 }
1320 EXPORT_SYMBOL(__class_export_add_lock_ref);
1321
1322 void __class_export_del_lock_ref(struct obd_export *exp, struct ldlm_lock *lock)
1323 {
1324         spin_lock(&exp->exp_locks_list_guard);
1325         LASSERT(lock->l_exp_refs_nr > 0);
1326         if (lock->l_exp_refs_target != exp) {
1327                 LCONSOLE_WARN("lock %p, "
1328                               "mismatching export pointers: %p, %p\n",
1329                               lock, lock->l_exp_refs_target, exp);
1330         }
1331         if (-- lock->l_exp_refs_nr == 0) {
1332                 list_del_init(&lock->l_exp_refs_link);
1333                 lock->l_exp_refs_target = NULL;
1334         }
1335         CDEBUG(D_INFO, "lock = %p, export = %p, refs = %u\n",
1336                lock, exp, lock->l_exp_refs_nr);
1337         spin_unlock(&exp->exp_locks_list_guard);
1338 }
1339 EXPORT_SYMBOL(__class_export_del_lock_ref);
1340 #endif
1341
1342 /* A connection defines an export context in which preallocation can
1343    be managed. This releases the export pointer reference, and returns
1344    the export handle, so the export refcount is 1 when this function
1345    returns. */
1346 int class_connect(struct lustre_handle *conn, struct obd_device *obd,
1347                   struct obd_uuid *cluuid)
1348 {
1349         struct obd_export *export;
1350         LASSERT(conn != NULL);
1351         LASSERT(obd != NULL);
1352         LASSERT(cluuid != NULL);
1353         ENTRY;
1354
1355         export = class_new_export(obd, cluuid);
1356         if (IS_ERR(export))
1357                 RETURN(PTR_ERR(export));
1358
1359         conn->cookie = export->exp_handle.h_cookie;
1360         class_export_put(export);
1361
1362         CDEBUG(D_IOCTL, "connect: client %s, cookie %#llx\n",
1363                cluuid->uuid, conn->cookie);
1364         RETURN(0);
1365 }
1366 EXPORT_SYMBOL(class_connect);
1367
1368 /* if export is involved in recovery then clean up related things */
1369 static void class_export_recovery_cleanup(struct obd_export *exp)
1370 {
1371         struct obd_device *obd = exp->exp_obd;
1372
1373         spin_lock(&obd->obd_recovery_task_lock);
1374         if (obd->obd_recovering) {
1375                 if (exp->exp_in_recovery) {
1376                         spin_lock(&exp->exp_lock);
1377                         exp->exp_in_recovery = 0;
1378                         spin_unlock(&exp->exp_lock);
1379                         LASSERT_ATOMIC_POS(&obd->obd_connected_clients);
1380                         atomic_dec(&obd->obd_connected_clients);
1381                 }
1382
1383                 /* if called during recovery then should update
1384                  * obd_stale_clients counter,
1385                  * lightweight exports are not counted */
1386                 if ((exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT) == 0)
1387                         exp->exp_obd->obd_stale_clients++;
1388         }
1389         spin_unlock(&obd->obd_recovery_task_lock);
1390
1391         spin_lock(&exp->exp_lock);
1392         /** Cleanup req replay fields */
1393         if (exp->exp_req_replay_needed) {
1394                 exp->exp_req_replay_needed = 0;
1395
1396                 LASSERT(atomic_read(&obd->obd_req_replay_clients));
1397                 atomic_dec(&obd->obd_req_replay_clients);
1398         }
1399
1400         /** Cleanup lock replay data */
1401         if (exp->exp_lock_replay_needed) {
1402                 exp->exp_lock_replay_needed = 0;
1403
1404                 LASSERT(atomic_read(&obd->obd_lock_replay_clients));
1405                 atomic_dec(&obd->obd_lock_replay_clients);
1406         }
1407         spin_unlock(&exp->exp_lock);
1408 }
1409
1410 /* This function removes 1-3 references from the export:
1411  * 1 - for export pointer passed
1412  * and if disconnect really need
1413  * 2 - removing from hash
1414  * 3 - in client_unlink_export
1415  * The export pointer passed to this function can destroyed */
1416 int class_disconnect(struct obd_export *export)
1417 {
1418         int already_disconnected;
1419         ENTRY;
1420
1421         if (export == NULL) {
1422                 CWARN("attempting to free NULL export %p\n", export);
1423                 RETURN(-EINVAL);
1424         }
1425
1426         spin_lock(&export->exp_lock);
1427         already_disconnected = export->exp_disconnected;
1428         export->exp_disconnected = 1;
1429 #ifdef HAVE_SERVER_SUPPORT
1430         /*  We hold references of export for uuid hash
1431          *  and nid_hash and export link at least. So
1432          *  it is safe to call rh*table_remove_fast in
1433          *  there.
1434          */
1435         obd_nid_del(export->exp_obd, export);
1436 #endif /* HAVE_SERVER_SUPPORT */
1437         spin_unlock(&export->exp_lock);
1438
1439         /* class_cleanup(), abort_recovery(), and class_fail_export()
1440          * all end up in here, and if any of them race we shouldn't
1441          * call extra class_export_puts(). */
1442         if (already_disconnected)
1443                 GOTO(no_disconn, already_disconnected);
1444
1445         CDEBUG(D_IOCTL, "disconnect: cookie %#llx\n",
1446                export->exp_handle.h_cookie);
1447
1448         class_export_recovery_cleanup(export);
1449         class_unlink_export(export);
1450 no_disconn:
1451         class_export_put(export);
1452         RETURN(0);
1453 }
1454 EXPORT_SYMBOL(class_disconnect);
1455
1456 /* Return non-zero for a fully connected export */
1457 int class_connected_export(struct obd_export *exp)
1458 {
1459         int connected = 0;
1460
1461         if (exp) {
1462                 spin_lock(&exp->exp_lock);
1463                 connected = (exp->exp_conn_cnt > 0) && !exp->exp_failed;
1464                 spin_unlock(&exp->exp_lock);
1465         }
1466         return connected;
1467 }
1468 EXPORT_SYMBOL(class_connected_export);
1469
1470 static void class_disconnect_export_list(struct list_head *list,
1471                                          enum obd_option flags)
1472 {
1473         int rc;
1474         struct obd_export *exp;
1475         ENTRY;
1476
1477         /* It's possible that an export may disconnect itself, but
1478          * nothing else will be added to this list. */
1479         while (!list_empty(list)) {
1480                 exp = list_first_entry(list, struct obd_export,
1481                                        exp_obd_chain);
1482                 /* need for safe call CDEBUG after obd_disconnect */
1483                 class_export_get(exp);
1484
1485                 spin_lock(&exp->exp_lock);
1486                 exp->exp_flags = flags;
1487                 spin_unlock(&exp->exp_lock);
1488
1489                 if (obd_uuid_equals(&exp->exp_client_uuid,
1490                                     &exp->exp_obd->obd_uuid)) {
1491                         CDEBUG(D_HA,
1492                                "exp %p export uuid == obd uuid, don't discon\n",
1493                                exp);
1494                         /* Need to delete this now so we don't end up pointing
1495                          * to work_list later when this export is cleaned up. */
1496                         list_del_init(&exp->exp_obd_chain);
1497                         class_export_put(exp);
1498                         continue;
1499                 }
1500
1501                 class_export_get(exp);
1502                 CDEBUG(D_HA, "%s: disconnecting export at %s (%p), "
1503                        "last request at %lld\n",
1504                        exp->exp_obd->obd_name, obd_export_nid2str(exp),
1505                        exp, exp->exp_last_request_time);
1506                 /* release one export reference anyway */
1507                 rc = obd_disconnect(exp);
1508
1509                 CDEBUG(D_HA, "disconnected export at %s (%p): rc %d\n",
1510                        obd_export_nid2str(exp), exp, rc);
1511                 class_export_put(exp);
1512         }
1513         EXIT;
1514 }
1515
1516 void class_disconnect_exports(struct obd_device *obd)
1517 {
1518         LIST_HEAD(work_list);
1519         ENTRY;
1520
1521         /* Move all of the exports from obd_exports to a work list, en masse. */
1522         spin_lock(&obd->obd_dev_lock);
1523         list_splice_init(&obd->obd_exports, &work_list);
1524         list_splice_init(&obd->obd_delayed_exports, &work_list);
1525         spin_unlock(&obd->obd_dev_lock);
1526
1527         if (!list_empty(&work_list)) {
1528                 CDEBUG(D_HA, "OBD device %d (%p) has exports, "
1529                        "disconnecting them\n", obd->obd_minor, obd);
1530                 class_disconnect_export_list(&work_list,
1531                                              exp_flags_from_obd(obd));
1532         } else
1533                 CDEBUG(D_HA, "OBD device %d (%p) has no exports\n",
1534                        obd->obd_minor, obd);
1535         EXIT;
1536 }
1537 EXPORT_SYMBOL(class_disconnect_exports);
1538
1539 /* Remove exports that have not completed recovery.
1540  */
1541 void class_disconnect_stale_exports(struct obd_device *obd,
1542                                     int (*test_export)(struct obd_export *))
1543 {
1544         LIST_HEAD(work_list);
1545         struct obd_export *exp, *n;
1546         int evicted = 0;
1547         ENTRY;
1548
1549         spin_lock(&obd->obd_dev_lock);
1550         list_for_each_entry_safe(exp, n, &obd->obd_exports,
1551                                  exp_obd_chain) {
1552                 /* don't count self-export as client */
1553                 if (obd_uuid_equals(&exp->exp_client_uuid,
1554                                     &exp->exp_obd->obd_uuid))
1555                         continue;
1556
1557                 /* don't evict clients which have no slot in last_rcvd
1558                  * (e.g. lightweight connection) */
1559                 if (exp->exp_target_data.ted_lr_idx == -1)
1560                         continue;
1561
1562                 spin_lock(&exp->exp_lock);
1563                 if (exp->exp_failed || test_export(exp)) {
1564                         spin_unlock(&exp->exp_lock);
1565                         continue;
1566                 }
1567                 exp->exp_failed = 1;
1568                 spin_unlock(&exp->exp_lock);
1569
1570                 list_move(&exp->exp_obd_chain, &work_list);
1571                 evicted++;
1572                 CDEBUG(D_HA, "%s: disconnect stale client %s@%s\n",
1573                        obd->obd_name, exp->exp_client_uuid.uuid,
1574                        obd_export_nid2str(exp));
1575                 print_export_data(exp, "EVICTING", 0, D_HA);
1576         }
1577         spin_unlock(&obd->obd_dev_lock);
1578
1579         if (evicted)
1580                 LCONSOLE_WARN("%s: disconnecting %d stale clients\n",
1581                               obd->obd_name, evicted);
1582
1583         class_disconnect_export_list(&work_list, exp_flags_from_obd(obd) |
1584                                                  OBD_OPT_ABORT_RECOV);
1585         EXIT;
1586 }
1587 EXPORT_SYMBOL(class_disconnect_stale_exports);
1588
1589 void class_fail_export(struct obd_export *exp)
1590 {
1591         int rc, already_failed;
1592
1593         spin_lock(&exp->exp_lock);
1594         already_failed = exp->exp_failed;
1595         exp->exp_failed = 1;
1596         spin_unlock(&exp->exp_lock);
1597
1598         if (already_failed) {
1599                 CDEBUG(D_HA, "disconnecting dead export %p/%s; skipping\n",
1600                        exp, exp->exp_client_uuid.uuid);
1601                 return;
1602         }
1603
1604         CDEBUG(D_HA, "disconnecting export %p/%s\n",
1605                exp, exp->exp_client_uuid.uuid);
1606
1607         if (obd_dump_on_timeout)
1608                 libcfs_debug_dumplog();
1609
1610         /* need for safe call CDEBUG after obd_disconnect */
1611         class_export_get(exp);
1612
1613         /* Most callers into obd_disconnect are removing their own reference
1614          * (request, for example) in addition to the one from the hash table.
1615          * We don't have such a reference here, so make one. */
1616         class_export_get(exp);
1617         rc = obd_disconnect(exp);
1618         if (rc)
1619                 CERROR("disconnecting export %p failed: %d\n", exp, rc);
1620         else
1621                 CDEBUG(D_HA, "disconnected export %p/%s\n",
1622                        exp, exp->exp_client_uuid.uuid);
1623         class_export_put(exp);
1624 }
1625 EXPORT_SYMBOL(class_fail_export);
1626
1627 #ifdef HAVE_SERVER_SUPPORT
1628 int obd_export_evict_by_nid(struct obd_device *obd, const char *nid)
1629 {
1630         lnet_nid_t nid_key = libcfs_str2nid((char *)nid);
1631         struct obd_export *doomed_exp;
1632         struct rhashtable_iter iter;
1633         int exports_evicted = 0;
1634
1635         spin_lock(&obd->obd_dev_lock);
1636         /* umount has run already, so evict thread should leave
1637          * its task to umount thread now */
1638         if (obd->obd_stopping) {
1639                 spin_unlock(&obd->obd_dev_lock);
1640                 return exports_evicted;
1641         }
1642         spin_unlock(&obd->obd_dev_lock);
1643
1644         rhltable_walk_enter(&obd->obd_nid_hash, &iter);
1645         rhashtable_walk_start(&iter);
1646         while ((doomed_exp = rhashtable_walk_next(&iter)) != NULL) {
1647                 if (IS_ERR(doomed_exp))
1648                         continue;
1649
1650                 if (!doomed_exp->exp_connection ||
1651                     doomed_exp->exp_connection->c_peer.nid != nid_key)
1652                         continue;
1653
1654                 if (!refcount_inc_not_zero(&doomed_exp->exp_handle.h_ref))
1655                         continue;
1656
1657                 rhashtable_walk_stop(&iter);
1658
1659                 LASSERTF(doomed_exp != obd->obd_self_export,
1660                          "self-export is hashed by NID?\n");
1661
1662                 LCONSOLE_WARN("%s: evicting %s (at %s) by administrative request\n",
1663                               obd->obd_name,
1664                               obd_uuid2str(&doomed_exp->exp_client_uuid),
1665                               obd_export_nid2str(doomed_exp));
1666
1667                 class_fail_export(doomed_exp);
1668                 class_export_put(doomed_exp);
1669                 exports_evicted++;
1670
1671                 rhashtable_walk_start(&iter);
1672         }
1673         rhashtable_walk_stop(&iter);
1674         rhashtable_walk_exit(&iter);
1675
1676         if (!exports_evicted)
1677                 CDEBUG(D_HA,"%s: can't disconnect NID '%s': no exports found\n",
1678                        obd->obd_name, nid);
1679         return exports_evicted;
1680 }
1681 EXPORT_SYMBOL(obd_export_evict_by_nid);
1682
1683 int obd_export_evict_by_uuid(struct obd_device *obd, const char *uuid)
1684 {
1685         struct obd_export *doomed_exp = NULL;
1686         struct obd_uuid doomed_uuid;
1687         int exports_evicted = 0;
1688
1689         spin_lock(&obd->obd_dev_lock);
1690         if (obd->obd_stopping) {
1691                 spin_unlock(&obd->obd_dev_lock);
1692                 return exports_evicted;
1693         }
1694         spin_unlock(&obd->obd_dev_lock);
1695
1696         obd_str2uuid(&doomed_uuid, uuid);
1697         if (obd_uuid_equals(&doomed_uuid, &obd->obd_uuid)) {
1698                 CERROR("%s: can't evict myself\n", obd->obd_name);
1699                 return exports_evicted;
1700         }
1701
1702         doomed_exp = obd_uuid_lookup(obd, &doomed_uuid);
1703         if (doomed_exp == NULL) {
1704                 CERROR("%s: can't disconnect %s: no exports found\n",
1705                        obd->obd_name, uuid);
1706         } else {
1707                 CWARN("%s: evicting %s at adminstrative request\n",
1708                        obd->obd_name, doomed_exp->exp_client_uuid.uuid);
1709                 class_fail_export(doomed_exp);
1710                 class_export_put(doomed_exp);
1711                 obd_uuid_del(obd, doomed_exp);
1712                 exports_evicted++;
1713         }
1714
1715         return exports_evicted;
1716 }
1717 #endif /* HAVE_SERVER_SUPPORT */
1718
1719 #if LUSTRE_TRACKS_LOCK_EXP_REFS
1720 void (*class_export_dump_hook)(struct obd_export*) = NULL;
1721 EXPORT_SYMBOL(class_export_dump_hook);
1722 #endif
1723
1724 static void print_export_data(struct obd_export *exp, const char *status,
1725                               int locks, int debug_level)
1726 {
1727         struct ptlrpc_reply_state *rs;
1728         struct ptlrpc_reply_state *first_reply = NULL;
1729         int nreplies = 0;
1730
1731         spin_lock(&exp->exp_lock);
1732         list_for_each_entry(rs, &exp->exp_outstanding_replies,
1733                             rs_exp_list) {
1734                 if (nreplies == 0)
1735                         first_reply = rs;
1736                 nreplies++;
1737         }
1738         spin_unlock(&exp->exp_lock);
1739
1740         CDEBUG(debug_level, "%s: %s %p %s %s %d (%d %d %d) %d %d %d %d: "
1741                "%p %s %llu stale:%d\n",
1742                exp->exp_obd->obd_name, status, exp, exp->exp_client_uuid.uuid,
1743                obd_export_nid2str(exp),
1744                refcount_read(&exp->exp_handle.h_ref),
1745                atomic_read(&exp->exp_rpc_count),
1746                atomic_read(&exp->exp_cb_count),
1747                atomic_read(&exp->exp_locks_count),
1748                exp->exp_disconnected, exp->exp_delayed, exp->exp_failed,
1749                nreplies, first_reply, nreplies > 3 ? "..." : "",
1750                exp->exp_last_committed, !list_empty(&exp->exp_stale_list));
1751 #if LUSTRE_TRACKS_LOCK_EXP_REFS
1752         if (locks && class_export_dump_hook != NULL)
1753                 class_export_dump_hook(exp);
1754 #endif
1755 }
1756
1757 void dump_exports(struct obd_device *obd, int locks, int debug_level)
1758 {
1759         struct obd_export *exp;
1760
1761         spin_lock(&obd->obd_dev_lock);
1762         list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain)
1763                 print_export_data(exp, "ACTIVE", locks, debug_level);
1764         list_for_each_entry(exp, &obd->obd_unlinked_exports, exp_obd_chain)
1765                 print_export_data(exp, "UNLINKED", locks, debug_level);
1766         list_for_each_entry(exp, &obd->obd_delayed_exports, exp_obd_chain)
1767                 print_export_data(exp, "DELAYED", locks, debug_level);
1768         spin_unlock(&obd->obd_dev_lock);
1769 }
1770
1771 void obd_exports_barrier(struct obd_device *obd)
1772 {
1773         int waited = 2;
1774         LASSERT(list_empty(&obd->obd_exports));
1775         spin_lock(&obd->obd_dev_lock);
1776         while (!list_empty(&obd->obd_unlinked_exports)) {
1777                 spin_unlock(&obd->obd_dev_lock);
1778                 schedule_timeout_uninterruptible(cfs_time_seconds(waited));
1779                 if (waited > 5 && is_power_of_2(waited)) {
1780                         LCONSOLE_WARN("%s is waiting for obd_unlinked_exports "
1781                                       "more than %d seconds. "
1782                                       "The obd refcount = %d. Is it stuck?\n",
1783                                       obd->obd_name, waited,
1784                                       atomic_read(&obd->obd_refcount));
1785                         dump_exports(obd, 1, D_CONSOLE | D_WARNING);
1786                 }
1787                 waited *= 2;
1788                 spin_lock(&obd->obd_dev_lock);
1789         }
1790         spin_unlock(&obd->obd_dev_lock);
1791 }
1792 EXPORT_SYMBOL(obd_exports_barrier);
1793
1794 /**
1795  * Add export to the obd_zombe thread and notify it.
1796  */
1797 static void obd_zombie_export_add(struct obd_export *exp) {
1798         atomic_dec(&obd_stale_export_num);
1799         spin_lock(&exp->exp_obd->obd_dev_lock);
1800         LASSERT(!list_empty(&exp->exp_obd_chain));
1801         list_del_init(&exp->exp_obd_chain);
1802         spin_unlock(&exp->exp_obd->obd_dev_lock);
1803
1804         queue_work(zombie_wq, &exp->exp_zombie_work);
1805 }
1806
1807 /**
1808  * Add import to the obd_zombe thread and notify it.
1809  */
1810 static void obd_zombie_import_add(struct obd_import *imp) {
1811         LASSERT(imp->imp_sec == NULL);
1812
1813         queue_work(zombie_wq, &imp->imp_zombie_work);
1814 }
1815
1816 /**
1817  * wait when obd_zombie import/export queues become empty
1818  */
1819 void obd_zombie_barrier(void)
1820 {
1821         flush_workqueue(zombie_wq);
1822 }
1823 EXPORT_SYMBOL(obd_zombie_barrier);
1824
1825
1826 struct obd_export *obd_stale_export_get(void)
1827 {
1828         struct obd_export *exp = NULL;
1829         ENTRY;
1830
1831         spin_lock(&obd_stale_export_lock);
1832         if (!list_empty(&obd_stale_exports)) {
1833                 exp = list_first_entry(&obd_stale_exports,
1834                                        struct obd_export, exp_stale_list);
1835                 list_del_init(&exp->exp_stale_list);
1836         }
1837         spin_unlock(&obd_stale_export_lock);
1838
1839         if (exp) {
1840                 CDEBUG(D_DLMTRACE, "Get export %p: total %d\n", exp,
1841                        atomic_read(&obd_stale_export_num));
1842         }
1843         RETURN(exp);
1844 }
1845 EXPORT_SYMBOL(obd_stale_export_get);
1846
1847 void obd_stale_export_put(struct obd_export *exp)
1848 {
1849         ENTRY;
1850
1851         LASSERT(list_empty(&exp->exp_stale_list));
1852         if (exp->exp_lock_hash &&
1853             atomic_read(&exp->exp_lock_hash->hs_count)) {
1854                 CDEBUG(D_DLMTRACE, "Put export %p: total %d\n", exp,
1855                        atomic_read(&obd_stale_export_num));
1856
1857                 spin_lock_bh(&exp->exp_bl_list_lock);
1858                 spin_lock(&obd_stale_export_lock);
1859                 /* Add to the tail if there is no blocked locks,
1860                  * to the head otherwise. */
1861                 if (list_empty(&exp->exp_bl_list))
1862                         list_add_tail(&exp->exp_stale_list,
1863                                       &obd_stale_exports);
1864                 else
1865                         list_add(&exp->exp_stale_list,
1866                                  &obd_stale_exports);
1867
1868                 spin_unlock(&obd_stale_export_lock);
1869                 spin_unlock_bh(&exp->exp_bl_list_lock);
1870         } else {
1871                 class_export_put(exp);
1872         }
1873         EXIT;
1874 }
1875 EXPORT_SYMBOL(obd_stale_export_put);
1876
1877 /**
1878  * Adjust the position of the export in the stale list,
1879  * i.e. move to the head of the list if is needed.
1880  **/
1881 void obd_stale_export_adjust(struct obd_export *exp)
1882 {
1883         LASSERT(exp != NULL);
1884         spin_lock_bh(&exp->exp_bl_list_lock);
1885         spin_lock(&obd_stale_export_lock);
1886
1887         if (!list_empty(&exp->exp_stale_list) &&
1888             !list_empty(&exp->exp_bl_list))
1889                 list_move(&exp->exp_stale_list, &obd_stale_exports);
1890
1891         spin_unlock(&obd_stale_export_lock);
1892         spin_unlock_bh(&exp->exp_bl_list_lock);
1893 }
1894 EXPORT_SYMBOL(obd_stale_export_adjust);
1895
1896 /**
1897  * start destroy zombie import/export thread
1898  */
1899 int obd_zombie_impexp_init(void)
1900 {
1901         zombie_wq = cfs_cpt_bind_workqueue("obd_zombid", cfs_cpt_tab,
1902                                            0, CFS_CPT_ANY,
1903                                            cfs_cpt_number(cfs_cpt_tab));
1904
1905         return IS_ERR(zombie_wq) ? PTR_ERR(zombie_wq) : 0;
1906 }
1907
1908 /**
1909  * stop destroy zombie import/export thread
1910  */
1911 void obd_zombie_impexp_stop(void)
1912 {
1913         destroy_workqueue(zombie_wq);
1914         LASSERT(list_empty(&obd_stale_exports));
1915 }
1916
1917 /***** Kernel-userspace comm helpers *******/
1918
1919 /* Get length of entire message, including header */
1920 int kuc_len(int payload_len)
1921 {
1922         return sizeof(struct kuc_hdr) + payload_len;
1923 }
1924 EXPORT_SYMBOL(kuc_len);
1925
1926 /* Get a pointer to kuc header, given a ptr to the payload
1927  * @param p Pointer to payload area
1928  * @returns Pointer to kuc header
1929  */
1930 struct kuc_hdr * kuc_ptr(void *p)
1931 {
1932         struct kuc_hdr *lh = ((struct kuc_hdr *)p) - 1;
1933         LASSERT(lh->kuc_magic == KUC_MAGIC);
1934         return lh;
1935 }
1936 EXPORT_SYMBOL(kuc_ptr);
1937
1938 /* Alloc space for a message, and fill in header
1939  * @return Pointer to payload area
1940  */
1941 void *kuc_alloc(int payload_len, int transport, int type)
1942 {
1943         struct kuc_hdr *lh;
1944         int len = kuc_len(payload_len);
1945
1946         OBD_ALLOC(lh, len);
1947         if (lh == NULL)
1948                 return ERR_PTR(-ENOMEM);
1949
1950         lh->kuc_magic = KUC_MAGIC;
1951         lh->kuc_transport = transport;
1952         lh->kuc_msgtype = type;
1953         lh->kuc_msglen = len;
1954
1955         return (void *)(lh + 1);
1956 }
1957 EXPORT_SYMBOL(kuc_alloc);
1958
1959 /* Takes pointer to payload area */
1960 void kuc_free(void *p, int payload_len)
1961 {
1962         struct kuc_hdr *lh = kuc_ptr(p);
1963         OBD_FREE(lh, kuc_len(payload_len));
1964 }
1965 EXPORT_SYMBOL(kuc_free);
1966
1967 struct obd_request_slot_waiter {
1968         struct list_head        orsw_entry;
1969         wait_queue_head_t       orsw_waitq;
1970         bool                    orsw_signaled;
1971 };
1972
1973 static bool obd_request_slot_avail(struct client_obd *cli,
1974                                    struct obd_request_slot_waiter *orsw)
1975 {
1976         bool avail;
1977
1978         spin_lock(&cli->cl_loi_list_lock);
1979         avail = !!list_empty(&orsw->orsw_entry);
1980         spin_unlock(&cli->cl_loi_list_lock);
1981
1982         return avail;
1983 };
1984
1985 /*
1986  * For network flow control, the RPC sponsor needs to acquire a credit
1987  * before sending the RPC. The credits count for a connection is defined
1988  * by the "cl_max_rpcs_in_flight". If all the credits are occpuied, then
1989  * the subsequent RPC sponsors need to wait until others released their
1990  * credits, or the administrator increased the "cl_max_rpcs_in_flight".
1991  */
1992 int obd_get_request_slot(struct client_obd *cli)
1993 {
1994         struct obd_request_slot_waiter   orsw;
1995         int                              rc;
1996
1997         spin_lock(&cli->cl_loi_list_lock);
1998         if (cli->cl_rpcs_in_flight < cli->cl_max_rpcs_in_flight) {
1999                 cli->cl_rpcs_in_flight++;
2000                 spin_unlock(&cli->cl_loi_list_lock);
2001                 return 0;
2002         }
2003
2004         init_waitqueue_head(&orsw.orsw_waitq);
2005         list_add_tail(&orsw.orsw_entry, &cli->cl_flight_waiters);
2006         orsw.orsw_signaled = false;
2007         spin_unlock(&cli->cl_loi_list_lock);
2008
2009         rc = l_wait_event_abortable(orsw.orsw_waitq,
2010                                     obd_request_slot_avail(cli, &orsw) ||
2011                                     orsw.orsw_signaled);
2012
2013         /* Here, we must take the lock to avoid the on-stack 'orsw' to be
2014          * freed but other (such as obd_put_request_slot) is using it. */
2015         spin_lock(&cli->cl_loi_list_lock);
2016         if (rc != 0) {
2017                 if (!orsw.orsw_signaled) {
2018                         if (list_empty(&orsw.orsw_entry))
2019                                 cli->cl_rpcs_in_flight--;
2020                         else
2021                                 list_del(&orsw.orsw_entry);
2022                 }
2023                 rc = -EINTR;
2024         }
2025
2026         if (orsw.orsw_signaled) {
2027                 LASSERT(list_empty(&orsw.orsw_entry));
2028
2029                 rc = -EINTR;
2030         }
2031         spin_unlock(&cli->cl_loi_list_lock);
2032
2033         return rc;
2034 }
2035 EXPORT_SYMBOL(obd_get_request_slot);
2036
2037 void obd_put_request_slot(struct client_obd *cli)
2038 {
2039         struct obd_request_slot_waiter *orsw;
2040
2041         spin_lock(&cli->cl_loi_list_lock);
2042         cli->cl_rpcs_in_flight--;
2043
2044         /* If there is free slot, wakeup the first waiter. */
2045         if (!list_empty(&cli->cl_flight_waiters) &&
2046             likely(cli->cl_rpcs_in_flight < cli->cl_max_rpcs_in_flight)) {
2047                 orsw = list_first_entry(&cli->cl_flight_waiters,
2048                                         struct obd_request_slot_waiter,
2049                                         orsw_entry);
2050                 list_del_init(&orsw->orsw_entry);
2051                 cli->cl_rpcs_in_flight++;
2052                 wake_up(&orsw->orsw_waitq);
2053         }
2054         spin_unlock(&cli->cl_loi_list_lock);
2055 }
2056 EXPORT_SYMBOL(obd_put_request_slot);
2057
2058 __u32 obd_get_max_rpcs_in_flight(struct client_obd *cli)
2059 {
2060         return cli->cl_max_rpcs_in_flight;
2061 }
2062 EXPORT_SYMBOL(obd_get_max_rpcs_in_flight);
2063
2064 int obd_set_max_rpcs_in_flight(struct client_obd *cli, __u32 max)
2065 {
2066         struct obd_request_slot_waiter *orsw;
2067         __u32                           old;
2068         int                             diff;
2069         int                             i;
2070         int                             rc;
2071
2072         if (max > OBD_MAX_RIF_MAX || max < 1)
2073                 return -ERANGE;
2074
2075         CDEBUG(D_INFO, "%s: max = %hu max_mod = %u rif = %u\n",
2076                cli->cl_import->imp_obd->obd_name, max,
2077                cli->cl_max_mod_rpcs_in_flight, cli->cl_max_rpcs_in_flight);
2078
2079         if (strcmp(cli->cl_import->imp_obd->obd_type->typ_name,
2080                    LUSTRE_MDC_NAME) == 0) {
2081                 /* adjust max_mod_rpcs_in_flight to ensure it is always
2082                  * strictly lower that max_rpcs_in_flight */
2083                 if (max < 2) {
2084                         CERROR("%s: cannot set mdc.*.max_rpcs_in_flight=1\n",
2085                                cli->cl_import->imp_obd->obd_name);
2086                         return -ERANGE;
2087                 }
2088                 if (max <= cli->cl_max_mod_rpcs_in_flight) {
2089                         rc = obd_set_max_mod_rpcs_in_flight(cli, max - 1);
2090                         if (rc != 0)
2091                                 return rc;
2092                 }
2093         }
2094
2095         spin_lock(&cli->cl_loi_list_lock);
2096         old = cli->cl_max_rpcs_in_flight;
2097         cli->cl_max_rpcs_in_flight = max;
2098         client_adjust_max_dirty(cli);
2099
2100         diff = max - old;
2101
2102         /* We increase the max_rpcs_in_flight, then wakeup some waiters. */
2103         for (i = 0; i < diff; i++) {
2104                 if (list_empty(&cli->cl_flight_waiters))
2105                         break;
2106
2107                 orsw = list_first_entry(&cli->cl_flight_waiters,
2108                                         struct obd_request_slot_waiter,
2109                                         orsw_entry);
2110                 list_del_init(&orsw->orsw_entry);
2111                 cli->cl_rpcs_in_flight++;
2112                 wake_up(&orsw->orsw_waitq);
2113         }
2114         spin_unlock(&cli->cl_loi_list_lock);
2115
2116         return 0;
2117 }
2118 EXPORT_SYMBOL(obd_set_max_rpcs_in_flight);
2119
2120 __u16 obd_get_max_mod_rpcs_in_flight(struct client_obd *cli)
2121 {
2122         return cli->cl_max_mod_rpcs_in_flight;
2123 }
2124 EXPORT_SYMBOL(obd_get_max_mod_rpcs_in_flight);
2125
2126 int obd_set_max_mod_rpcs_in_flight(struct client_obd *cli, __u16 max)
2127 {
2128         struct obd_connect_data *ocd;
2129         __u16 maxmodrpcs;
2130         __u16 prev;
2131
2132         if (max > OBD_MAX_RIF_MAX || max < 1)
2133                 return -ERANGE;
2134
2135         ocd = &cli->cl_import->imp_connect_data;
2136         CDEBUG(D_INFO, "%s: max = %hu flags = %llx, max_mod = %u rif = %u\n",
2137                cli->cl_import->imp_obd->obd_name, max, ocd->ocd_connect_flags,
2138                ocd->ocd_maxmodrpcs, cli->cl_max_rpcs_in_flight);
2139
2140         if (max == OBD_MAX_RIF_MAX)
2141                 max = OBD_MAX_RIF_MAX - 1;
2142
2143         /* Cannot exceed or equal max_rpcs_in_flight.  If we are asked to
2144          * increase this value, also bump up max_rpcs_in_flight to match.
2145          */
2146         if (max >= cli->cl_max_rpcs_in_flight) {
2147                 CDEBUG(D_INFO,
2148                        "%s: increasing max_rpcs_in_flight=%hu to allow larger max_mod_rpcs_in_flight=%u\n",
2149                        cli->cl_import->imp_obd->obd_name, max + 1, max);
2150                 obd_set_max_rpcs_in_flight(cli, max + 1);
2151         }
2152
2153         /* cannot exceed max modify RPCs in flight supported by the server,
2154          * but verify ocd_connect_flags is at least initialized first.  If
2155          * not, allow it and fix value later in ptlrpc_connect_set_flags().
2156          */
2157         if (!ocd->ocd_connect_flags) {
2158                 maxmodrpcs = cli->cl_max_rpcs_in_flight - 1;
2159         } else if (ocd->ocd_connect_flags & OBD_CONNECT_MULTIMODRPCS) {
2160                 maxmodrpcs = ocd->ocd_maxmodrpcs;
2161                 if (maxmodrpcs == 0) { /* connection not finished yet */
2162                         maxmodrpcs = cli->cl_max_rpcs_in_flight - 1;
2163                         CDEBUG(D_INFO,
2164                                "%s: partial connect, assume maxmodrpcs=%hu\n",
2165                                cli->cl_import->imp_obd->obd_name, maxmodrpcs);
2166                 }
2167         } else {
2168                 maxmodrpcs = 1;
2169         }
2170         if (max > maxmodrpcs) {
2171                 CERROR("%s: can't set max_mod_rpcs_in_flight=%hu higher than ocd_maxmodrpcs=%hu returned by the server at connection\n",
2172                        cli->cl_import->imp_obd->obd_name,
2173                        max, maxmodrpcs);
2174                 return -ERANGE;
2175         }
2176
2177         spin_lock(&cli->cl_mod_rpcs_lock);
2178
2179         prev = cli->cl_max_mod_rpcs_in_flight;
2180         cli->cl_max_mod_rpcs_in_flight = max;
2181
2182         /* wakeup waiters if limit has been increased */
2183         if (cli->cl_max_mod_rpcs_in_flight > prev)
2184                 wake_up(&cli->cl_mod_rpcs_waitq);
2185
2186         spin_unlock(&cli->cl_mod_rpcs_lock);
2187
2188         return 0;
2189 }
2190 EXPORT_SYMBOL(obd_set_max_mod_rpcs_in_flight);
2191
2192 int obd_mod_rpc_stats_seq_show(struct client_obd *cli,
2193                                struct seq_file *seq)
2194 {
2195         unsigned long mod_tot = 0, mod_cum;
2196         struct timespec64 now;
2197         int i;
2198
2199         ktime_get_real_ts64(&now);
2200
2201         spin_lock(&cli->cl_mod_rpcs_lock);
2202
2203         seq_printf(seq, "snapshot_time:         %llu.%9lu (secs.nsecs)\n",
2204                    (s64)now.tv_sec, now.tv_nsec);
2205         seq_printf(seq, "modify_RPCs_in_flight:  %hu\n",
2206                    cli->cl_mod_rpcs_in_flight);
2207
2208         seq_printf(seq, "\n\t\t\tmodify\n");
2209         seq_printf(seq, "rpcs in flight        rpcs   %% cum %%\n");
2210
2211         mod_tot = lprocfs_oh_sum(&cli->cl_mod_rpcs_hist);
2212
2213         mod_cum = 0;
2214         for (i = 0; i < OBD_HIST_MAX; i++) {
2215                 unsigned long mod = cli->cl_mod_rpcs_hist.oh_buckets[i];
2216                 mod_cum += mod;
2217                 seq_printf(seq, "%d:\t\t%10lu %3u %3u\n",
2218                            i, mod, pct(mod, mod_tot),
2219                            pct(mod_cum, mod_tot));
2220                 if (mod_cum == mod_tot)
2221                         break;
2222         }
2223
2224         spin_unlock(&cli->cl_mod_rpcs_lock);
2225
2226         return 0;
2227 }
2228 EXPORT_SYMBOL(obd_mod_rpc_stats_seq_show);
2229
2230 /* The number of modify RPCs sent in parallel is limited
2231  * because the server has a finite number of slots per client to
2232  * store request result and ensure reply reconstruction when needed.
2233  * On the client, this limit is stored in cl_max_mod_rpcs_in_flight
2234  * that takes into account server limit and cl_max_rpcs_in_flight
2235  * value.
2236  * On the MDC client, to avoid a potential deadlock (see Bugzilla 3462),
2237  * one close request is allowed above the maximum.
2238  */
2239 static inline bool obd_mod_rpc_slot_avail_locked(struct client_obd *cli,
2240                                                  bool close_req)
2241 {
2242         bool avail;
2243
2244         /* A slot is available if
2245          * - number of modify RPCs in flight is less than the max
2246          * - it's a close RPC and no other close request is in flight
2247          */
2248         avail = cli->cl_mod_rpcs_in_flight < cli->cl_max_mod_rpcs_in_flight ||
2249                 (close_req && cli->cl_close_rpcs_in_flight == 0);
2250
2251         return avail;
2252 }
2253
2254 static inline bool obd_mod_rpc_slot_avail(struct client_obd *cli,
2255                                          bool close_req)
2256 {
2257         bool avail;
2258
2259         spin_lock(&cli->cl_mod_rpcs_lock);
2260         avail = obd_mod_rpc_slot_avail_locked(cli, close_req);
2261         spin_unlock(&cli->cl_mod_rpcs_lock);
2262         return avail;
2263 }
2264
2265
2266 /* Get a modify RPC slot from the obd client @cli according
2267  * to the kind of operation @opc that is going to be sent
2268  * and the intent @it of the operation if it applies.
2269  * If the maximum number of modify RPCs in flight is reached
2270  * the thread is put to sleep.
2271  * Returns the tag to be set in the request message. Tag 0
2272  * is reserved for non-modifying requests.
2273  */
2274 __u16 obd_get_mod_rpc_slot(struct client_obd *cli, __u32 opc)
2275 {
2276         bool                    close_req = false;
2277         __u16                   i, max;
2278
2279         if (opc == MDS_CLOSE)
2280                 close_req = true;
2281
2282         do {
2283                 spin_lock(&cli->cl_mod_rpcs_lock);
2284                 max = cli->cl_max_mod_rpcs_in_flight;
2285                 if (obd_mod_rpc_slot_avail_locked(cli, close_req)) {
2286                         /* there is a slot available */
2287                         cli->cl_mod_rpcs_in_flight++;
2288                         if (close_req)
2289                                 cli->cl_close_rpcs_in_flight++;
2290                         lprocfs_oh_tally(&cli->cl_mod_rpcs_hist,
2291                                          cli->cl_mod_rpcs_in_flight);
2292                         /* find a free tag */
2293                         i = find_first_zero_bit(cli->cl_mod_tag_bitmap,
2294                                                 max + 1);
2295                         LASSERT(i < OBD_MAX_RIF_MAX);
2296                         LASSERT(!test_and_set_bit(i, cli->cl_mod_tag_bitmap));
2297                         spin_unlock(&cli->cl_mod_rpcs_lock);
2298                         /* tag 0 is reserved for non-modify RPCs */
2299
2300                         CDEBUG(D_RPCTRACE,
2301                                "%s: modify RPC slot %u is allocated opc %u, max %hu\n",
2302                                cli->cl_import->imp_obd->obd_name,
2303                                i + 1, opc, max);
2304
2305                         return i + 1;
2306                 }
2307                 spin_unlock(&cli->cl_mod_rpcs_lock);
2308
2309                 CDEBUG(D_RPCTRACE, "%s: sleeping for a modify RPC slot "
2310                        "opc %u, max %hu\n",
2311                        cli->cl_import->imp_obd->obd_name, opc, max);
2312
2313                 wait_event_idle_exclusive(cli->cl_mod_rpcs_waitq,
2314                                           obd_mod_rpc_slot_avail(cli,
2315                                                                  close_req));
2316         } while (true);
2317 }
2318 EXPORT_SYMBOL(obd_get_mod_rpc_slot);
2319
2320 /* Put a modify RPC slot from the obd client @cli according
2321  * to the kind of operation @opc that has been sent.
2322  */
2323 void obd_put_mod_rpc_slot(struct client_obd *cli, __u32 opc, __u16 tag)
2324 {
2325         bool                    close_req = false;
2326
2327         if (tag == 0)
2328                 return;
2329
2330         if (opc == MDS_CLOSE)
2331                 close_req = true;
2332
2333         spin_lock(&cli->cl_mod_rpcs_lock);
2334         cli->cl_mod_rpcs_in_flight--;
2335         if (close_req)
2336                 cli->cl_close_rpcs_in_flight--;
2337         /* release the tag in the bitmap */
2338         LASSERT(tag - 1 < OBD_MAX_RIF_MAX);
2339         LASSERT(test_and_clear_bit(tag - 1, cli->cl_mod_tag_bitmap) != 0);
2340         spin_unlock(&cli->cl_mod_rpcs_lock);
2341         wake_up(&cli->cl_mod_rpcs_waitq);
2342 }
2343 EXPORT_SYMBOL(obd_put_mod_rpc_slot);
2344