lustre/obdclass/lu_object.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) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Whamcloud, Inc.
  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/lu_object.c
  37  *
  38  * Lustre Object.
  39  * These are the only exported functions, they provide some generic
  40  * infrastructure for managing object devices
  41  *
  42  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  43  */
  44
  45 #define DEBUG_SUBSYSTEM S_CLASS
  46
  47 #include <libcfs/libcfs.h>
  48
  49 #ifdef __KERNEL__
  50 # include <linux/module.h>
  51 #endif
  52
  53 /* hash_long() */
  54 #include <libcfs/libcfs_hash.h>
  55 #include <obd_class.h>
  56 #include <obd_support.h>
  57 #include <lustre_disk.h>
  58 #include <lustre_fid.h>
  59 #include <lu_object.h>
  60 #include <libcfs/list.h>
  61 /* lu_time_global_{init,fini}() */
  62 #include <lu_time.h>
  63
  64 static void lu_object_free(const struct lu_env *env, struct lu_object *o);
  65
  66 /**
  67  * Decrease reference counter on object. If last reference is freed, return
  68  * object to the cache, unless lu_object_is_dying(o) holds. In the latter
  69  * case, free object immediately.
  70  */
  71 void lu_object_put(const struct lu_env *env, struct lu_object *o)
  72 {
  73         struct lu_site_bkt_data *bkt;
  74         struct lu_object_header *top;
  75         struct lu_site          *site;
  76         struct lu_object        *orig;
  77         cfs_hash_bd_t            bd;
  78
  79         top  = o->lo_header;
  80         site = o->lo_dev->ld_site;
  81         orig = o;
  82
  83         cfs_hash_bd_get(site->ls_obj_hash, &top->loh_fid, &bd);
  84         bkt = cfs_hash_bd_extra_get(site->ls_obj_hash, &bd);
  85
  86         if (!cfs_hash_bd_dec_and_lock(site->ls_obj_hash, &bd, &top->loh_ref)) {
  87                 if (lu_object_is_dying(top)) {
  88
  89                         /*
  90                          * somebody may be waiting for this, currently only
  91                          * used for cl_object, see cl_object_put_last().
  92                          */
  93                         cfs_waitq_broadcast(&bkt->lsb_marche_funebre);
  94                 }
  95                 return;
  96         }
  97
  98         LASSERT(bkt->lsb_busy > 0);
  99         bkt->lsb_busy--;
 100         /*
 101          * When last reference is released, iterate over object
 102          * layers, and notify them that object is no longer busy.
 103          */
 104         cfs_list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
 105                 if (o->lo_ops->loo_object_release != NULL)
 106                         o->lo_ops->loo_object_release(env, o);
 107         }
 108
 109         if (!lu_object_is_dying(top)) {
 110                 LASSERT(cfs_list_empty(&top->loh_lru));
 111                 cfs_list_add_tail(&top->loh_lru, &bkt->lsb_lru);
 112                 cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
 113                 return;
 114         }
 115
 116         /*
 117          * If object is dying (will not be cached), removed it
 118          * from hash table and LRU.
 119          *
 120          * This is done with hash table and LRU lists locked. As the only
 121          * way to acquire first reference to previously unreferenced
 122          * object is through hash-table lookup (lu_object_find()),
 123          * or LRU scanning (lu_site_purge()), that are done under hash-table
 124          * and LRU lock, no race with concurrent object lookup is possible
 125          * and we can safely destroy object below.
 126          */
 127         cfs_hash_bd_del_locked(site->ls_obj_hash, &bd, &top->loh_hash);
 128         cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
 129         /*
 130          * Object was already removed from hash and lru above, can
 131          * kill it.
 132          */
 133         lu_object_free(env, orig);
 134 }
 135 EXPORT_SYMBOL(lu_object_put);
 136
 137 /**
 138  * Allocate new object.
 139  *
 140  * This follows object creation protocol, described in the comment within
 141  * struct lu_device_operations definition.
 142  */
 143 static struct lu_object *lu_object_alloc(const struct lu_env *env,
 144                                          struct lu_device *dev,
 145                                          const struct lu_fid *f,
 146                                          const struct lu_object_conf *conf)
 147 {
 148         struct lu_object *scan;
 149         struct lu_object *top;
 150         cfs_list_t *layers;
 151         int clean;
 152         int result;
 153         ENTRY;
 154
 155         /*
 156          * Create top-level object slice. This will also create
 157          * lu_object_header.
 158          */
 159         top = dev->ld_ops->ldo_object_alloc(env, NULL, dev);
 160         if (top == NULL)
 161                 RETURN(ERR_PTR(-ENOMEM));
 162         /*
 163          * This is the only place where object fid is assigned. It's constant
 164          * after this point.
 165          */
 166         LASSERT(fid_is_igif(f) || fid_ver(f) == 0);
 167         top->lo_header->loh_fid = *f;
 168         layers = &top->lo_header->loh_layers;
 169         do {
 170                 /*
 171                  * Call ->loo_object_init() repeatedly, until no more new
 172                  * object slices are created.
 173                  */
 174                 clean = 1;
 175                 cfs_list_for_each_entry(scan, layers, lo_linkage) {
 176                         if (scan->lo_flags & LU_OBJECT_ALLOCATED)
 177                                 continue;
 178                         clean = 0;
 179                         scan->lo_header = top->lo_header;
 180                         result = scan->lo_ops->loo_object_init(env, scan, conf);
 181                         if (result != 0) {
 182                                 lu_object_free(env, top);
 183                                 RETURN(ERR_PTR(result));
 184                         }
 185                         scan->lo_flags |= LU_OBJECT_ALLOCATED;
 186                 }
 187         } while (!clean);
 188
 189         cfs_list_for_each_entry_reverse(scan, layers, lo_linkage) {
 190                 if (scan->lo_ops->loo_object_start != NULL) {
 191                         result = scan->lo_ops->loo_object_start(env, scan);
 192                         if (result != 0) {
 193                                 lu_object_free(env, top);
 194                                 RETURN(ERR_PTR(result));
 195                         }
 196                 }
 197         }
 198
 199         lprocfs_counter_incr(dev->ld_site->ls_stats, LU_SS_CREATED);
 200         RETURN(top);
 201 }
 202
 203 /**
 204  * Free an object.
 205  */
 206 static void lu_object_free(const struct lu_env *env, struct lu_object *o)
 207 {
 208         struct lu_site_bkt_data *bkt;
 209         struct lu_site          *site;
 210         struct lu_object        *scan;
 211         cfs_list_t              *layers;
 212         cfs_list_t               splice;
 213
 214         site   = o->lo_dev->ld_site;
 215         layers = &o->lo_header->loh_layers;
 216         bkt    = lu_site_bkt_from_fid(site, &o->lo_header->loh_fid);
 217         /*
 218          * First call ->loo_object_delete() method to release all resources.
 219          */
 220         cfs_list_for_each_entry_reverse(scan, layers, lo_linkage) {
 221                 if (scan->lo_ops->loo_object_delete != NULL)
 222                         scan->lo_ops->loo_object_delete(env, scan);
 223         }
 224
 225         /*
 226          * Then, splice object layers into stand-alone list, and call
 227          * ->loo_object_free() on all layers to free memory. Splice is
 228          * necessary, because lu_object_header is freed together with the
 229          * top-level slice.
 230          */
 231         CFS_INIT_LIST_HEAD(&splice);
 232         cfs_list_splice_init(layers, &splice);
 233         while (!cfs_list_empty(&splice)) {
 234                 /*
 235                  * Free layers in bottom-to-top order, so that object header
 236                  * lives as long as possible and ->loo_object_free() methods
 237                  * can look at its contents.
 238                  */
 239                 o = container_of0(splice.prev, struct lu_object, lo_linkage);
 240                 cfs_list_del_init(&o->lo_linkage);
 241                 LASSERT(o->lo_ops->loo_object_free != NULL);
 242                 o->lo_ops->loo_object_free(env, o);
 243         }
 244
 245         if (cfs_waitq_active(&bkt->lsb_marche_funebre))
 246                 cfs_waitq_broadcast(&bkt->lsb_marche_funebre);
 247 }
 248
 249 /**
 250  * Free \a nr objects from the cold end of the site LRU list.
 251  */
 252 int lu_site_purge(const struct lu_env *env, struct lu_site *s, int nr)
 253 {
 254         struct lu_object_header *h;
 255         struct lu_object_header *temp;
 256         struct lu_site_bkt_data *bkt;
 257         cfs_hash_bd_t            bd;
 258         cfs_hash_bd_t            bd2;
 259         cfs_list_t               dispose;
 260         int                      did_sth;
 261         int                      start;
 262         int                      count;
 263         int                      bnr;
 264         int                      i;
 265
 266         CFS_INIT_LIST_HEAD(&dispose);
 267         /*
 268          * Under LRU list lock, scan LRU list and move unreferenced objects to
 269          * the dispose list, removing them from LRU and hash table.
 270          */
 271         start = s->ls_purge_start;
 272         bnr = (nr == ~0) ? -1 : nr / CFS_HASH_NBKT(s->ls_obj_hash) + 1;
 273  again:
 274         did_sth = 0;
 275         cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
 276                 if (i < start)
 277                         continue;
 278                 count = bnr;
 279                 cfs_hash_bd_lock(s->ls_obj_hash, &bd, 1);
 280                 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
 281
 282                 cfs_list_for_each_entry_safe(h, temp, &bkt->lsb_lru, loh_lru) {
 283                         LASSERT(cfs_atomic_read(&h->loh_ref) == 0);
 284
 285                         cfs_hash_bd_get(s->ls_obj_hash, &h->loh_fid, &bd2);
 286                         LASSERT(bd.bd_bucket == bd2.bd_bucket);
 287
 288                         cfs_hash_bd_del_locked(s->ls_obj_hash,
 289                                                &bd2, &h->loh_hash);
 290                         cfs_list_move(&h->loh_lru, &dispose);
 291                         if (did_sth == 0)
 292                                 did_sth = 1;
 293
 294                         if (nr != ~0 && --nr == 0)
 295                                 break;
 296
 297                         if (count > 0 && --count == 0)
 298                                 break;
 299
 300                 }
 301                 cfs_hash_bd_unlock(s->ls_obj_hash, &bd, 1);
 302                 cfs_cond_resched();
 303                 /*
 304                  * Free everything on the dispose list. This is safe against
 305                  * races due to the reasons described in lu_object_put().
 306                  */
 307                 while (!cfs_list_empty(&dispose)) {
 308                         h = container_of0(dispose.next,
 309                                           struct lu_object_header, loh_lru);
 310                         cfs_list_del_init(&h->loh_lru);
 311                         lu_object_free(env, lu_object_top(h));
 312                         lprocfs_counter_incr(s->ls_stats, LU_SS_LRU_PURGED);
 313                 }
 314
 315                 if (nr == 0)
 316                         break;
 317         }
 318
 319         if (nr != 0 && did_sth && start != 0) {
 320                 start = 0; /* restart from the first bucket */
 321                 goto again;
 322         }
 323         /* race on s->ls_purge_start, but nobody cares */
 324         s->ls_purge_start = i % CFS_HASH_NBKT(s->ls_obj_hash);
 325
 326         return nr;
 327 }
 328 EXPORT_SYMBOL(lu_site_purge);
 329
 330 /*
 331  * Object printing.
 332  *
 333  * Code below has to jump through certain loops to output object description
 334  * into libcfs_debug_msg-based log. The problem is that lu_object_print()
 335  * composes object description from strings that are parts of _lines_ of
 336  * output (i.e., strings that are not terminated by newline). This doesn't fit
 337  * very well into libcfs_debug_msg() interface that assumes that each message
 338  * supplied to it is a self-contained output line.
 339  *
 340  * To work around this, strings are collected in a temporary buffer
 341  * (implemented as a value of lu_cdebug_key key), until terminating newline
 342  * character is detected.
 343  *
 344  */
 345
 346 enum {
 347         /**
 348          * Maximal line size.
 349          *
 350          * XXX overflow is not handled correctly.
 351          */
 352         LU_CDEBUG_LINE = 512
 353 };
 354
 355 struct lu_cdebug_data {
 356         /**
 357          * Temporary buffer.
 358          */
 359         char lck_area[LU_CDEBUG_LINE];
 360 };
 361
 362 /* context key constructor/destructor: lu_global_key_init, lu_global_key_fini */
 363 LU_KEY_INIT_FINI(lu_global, struct lu_cdebug_data);
 364
 365 /**
 366  * Key, holding temporary buffer. This key is registered very early by
 367  * lu_global_init().
 368  */
 369 struct lu_context_key lu_global_key = {
 370         .lct_tags = LCT_MD_THREAD | LCT_DT_THREAD |
 371                     LCT_MG_THREAD | LCT_CL_THREAD,
 372         .lct_init = lu_global_key_init,
 373         .lct_fini = lu_global_key_fini
 374 };
 375
 376 /**
 377  * Printer function emitting messages through libcfs_debug_msg().
 378  */
 379 int lu_cdebug_printer(const struct lu_env *env,
 380                       void *cookie, const char *format, ...)
 381 {
 382         struct libcfs_debug_msg_data *msgdata = cookie;
 383         struct lu_cdebug_data        *key;
 384         int used;
 385         int complete;
 386         va_list args;
 387
 388         va_start(args, format);
 389
 390         key = lu_context_key_get(&env->le_ctx, &lu_global_key);
 391         LASSERT(key != NULL);
 392
 393         used = strlen(key->lck_area);
 394         complete = format[strlen(format) - 1] == '\n';
 395         /*
 396          * Append new chunk to the buffer.
 397          */
 398         vsnprintf(key->lck_area + used,
 399                   ARRAY_SIZE(key->lck_area) - used, format, args);
 400         if (complete) {
 401                 if (cfs_cdebug_show(msgdata->msg_mask, msgdata->msg_subsys))
 402                         libcfs_debug_msg(msgdata, "%s", key->lck_area);
 403                 key->lck_area[0] = 0;
 404         }
 405         va_end(args);
 406         return 0;
 407 }
 408 EXPORT_SYMBOL(lu_cdebug_printer);
 409
 410 /**
 411  * Print object header.
 412  */
 413 void lu_object_header_print(const struct lu_env *env, void *cookie,
 414                             lu_printer_t printer,
 415                             const struct lu_object_header *hdr)
 416 {
 417         (*printer)(env, cookie, "header@%p[%#lx, %d, "DFID"%s%s%s]",
 418                    hdr, hdr->loh_flags, cfs_atomic_read(&hdr->loh_ref),
 419                    PFID(&hdr->loh_fid),
 420                    cfs_hlist_unhashed(&hdr->loh_hash) ? "" : " hash",
 421                    cfs_list_empty((cfs_list_t *)&hdr->loh_lru) ? \
 422                    "" : " lru",
 423                    hdr->loh_attr & LOHA_EXISTS ? " exist":"");
 424 }
 425 EXPORT_SYMBOL(lu_object_header_print);
 426
 427 /**
 428  * Print human readable representation of the \a o to the \a printer.
 429  */
 430 void lu_object_print(const struct lu_env *env, void *cookie,
 431                      lu_printer_t printer, const struct lu_object *o)
 432 {
 433         static const char ruler[] = "........................................";
 434         struct lu_object_header *top;
 435         int depth;
 436
 437         top = o->lo_header;
 438         lu_object_header_print(env, cookie, printer, top);
 439         (*printer)(env, cookie, "{ \n");
 440         cfs_list_for_each_entry(o, &top->loh_layers, lo_linkage) {
 441                 depth = o->lo_depth + 4;
 442
 443                 /*
 444                  * print `.' \a depth times followed by type name and address
 445                  */
 446                 (*printer)(env, cookie, "%*.*s%s@%p", depth, depth, ruler,
 447                            o->lo_dev->ld_type->ldt_name, o);
 448                 if (o->lo_ops->loo_object_print != NULL)
 449                         o->lo_ops->loo_object_print(env, cookie, printer, o);
 450                 (*printer)(env, cookie, "\n");
 451         }
 452         (*printer)(env, cookie, "} header@%p\n", top);
 453 }
 454 EXPORT_SYMBOL(lu_object_print);
 455
 456 /**
 457  * Check object consistency.
 458  */
 459 int lu_object_invariant(const struct lu_object *o)
 460 {
 461         struct lu_object_header *top;
 462
 463         top = o->lo_header;
 464         cfs_list_for_each_entry(o, &top->loh_layers, lo_linkage) {
 465                 if (o->lo_ops->loo_object_invariant != NULL &&
 466                     !o->lo_ops->loo_object_invariant(o))
 467                         return 0;
 468         }
 469         return 1;
 470 }
 471 EXPORT_SYMBOL(lu_object_invariant);
 472
 473 static struct lu_object *htable_lookup(struct lu_site *s,
 474                                        cfs_hash_bd_t *bd,
 475                                        const struct lu_fid *f,
 476                                        cfs_waitlink_t *waiter,
 477                                        __u64 *version)
 478 {
 479         struct lu_site_bkt_data *bkt;
 480         struct lu_object_header *h;
 481         cfs_hlist_node_t        *hnode;
 482         __u64  ver = cfs_hash_bd_version_get(bd);
 483
 484         if (*version == ver)
 485                 return NULL;
 486
 487         *version = ver;
 488         bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, bd);
 489         /* cfs_hash_bd_lookup_intent is a somehow "internal" function
 490          * of cfs_hash, but we don't want refcount on object right now */
 491         hnode = cfs_hash_bd_lookup_locked(s->ls_obj_hash, bd, (void *)f);
 492         if (hnode == NULL) {
 493                 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_MISS);
 494                 return NULL;
 495         }
 496
 497         h = container_of0(hnode, struct lu_object_header, loh_hash);
 498         if (likely(!lu_object_is_dying(h))) {
 499                 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_HIT);
 500                 cfs_list_del_init(&h->loh_lru);
 501                 return lu_object_top(h);
 502         }
 503
 504         /*
 505          * Lookup found an object being destroyed this object cannot be
 506          * returned (to assure that references to dying objects are eventually
 507          * drained), and moreover, lookup has to wait until object is freed.
 508          */
 509         cfs_atomic_dec(&h->loh_ref);
 510
 511         cfs_waitlink_init(waiter);
 512         cfs_waitq_add(&bkt->lsb_marche_funebre, waiter);
 513         cfs_set_current_state(CFS_TASK_UNINT);
 514         lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_DEATH_RACE);
 515         return ERR_PTR(-EAGAIN);
 516 }
 517
 518 /**
 519  * Search cache for an object with the fid \a f. If such object is found,
 520  * return it. Otherwise, create new object, insert it into cache and return
 521  * it. In any case, additional reference is acquired on the returned object.
 522  */
 523 struct lu_object *lu_object_find(const struct lu_env *env,
 524                                  struct lu_device *dev, const struct lu_fid *f,
 525                                  const struct lu_object_conf *conf)
 526 {
 527         return lu_object_find_at(env, dev->ld_site->ls_top_dev, f, conf);
 528 }
 529 EXPORT_SYMBOL(lu_object_find);
 530
 531 static struct lu_object *lu_object_new(const struct lu_env *env,
 532                                        struct lu_device *dev,
 533                                        const struct lu_fid *f,
 534                                        const struct lu_object_conf *conf)
 535 {
 536         struct lu_object        *o;
 537         cfs_hash_t              *hs;
 538         cfs_hash_bd_t            bd;
 539         struct lu_site_bkt_data *bkt;
 540
 541         o = lu_object_alloc(env, dev, f, conf);
 542         if (unlikely(IS_ERR(o)))
 543                 return o;
 544
 545         hs = dev->ld_site->ls_obj_hash;
 546         cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
 547         bkt = cfs_hash_bd_extra_get(hs, &bd);
 548         cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
 549         bkt->lsb_busy++;
 550         cfs_hash_bd_unlock(hs, &bd, 1);
 551         return o;
 552 }
 553
 554 /**
 555  * Core logic of lu_object_find*() functions.
 556  */
 557 static struct lu_object *lu_object_find_try(const struct lu_env *env,
 558                                             struct lu_device *dev,
 559                                             const struct lu_fid *f,
 560                                             const struct lu_object_conf *conf,
 561                                             cfs_waitlink_t *waiter)
 562 {
 563         struct lu_object      *o;
 564         struct lu_object      *shadow;
 565         struct lu_site        *s;
 566         cfs_hash_t            *hs;
 567         cfs_hash_bd_t          bd;
 568         __u64                  version = 0;
 569
 570         /*
 571          * This uses standard index maintenance protocol:
 572          *
 573          *     - search index under lock, and return object if found;
 574          *     - otherwise, unlock index, allocate new object;
 575          *     - lock index and search again;
 576          *     - if nothing is found (usual case), insert newly created
 577          *       object into index;
 578          *     - otherwise (race: other thread inserted object), free
 579          *       object just allocated.
 580          *     - unlock index;
 581          *     - return object.
 582          *
 583          * For "LOC_F_NEW" case, we are sure the object is new established.
 584          * It is unnecessary to perform lookup-alloc-lookup-insert, instead,
 585          * just alloc and insert directly.
 586          *
 587          * If dying object is found during index search, add @waiter to the
 588          * site wait-queue and return ERR_PTR(-EAGAIN).
 589          */
 590         if (conf != NULL && conf->loc_flags & LOC_F_NEW)
 591                 return lu_object_new(env, dev, f, conf);
 592
 593         s  = dev->ld_site;
 594         hs = s->ls_obj_hash;
 595         cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
 596         o = htable_lookup(s, &bd, f, waiter, &version);
 597         cfs_hash_bd_unlock(hs, &bd, 1);
 598         if (o != NULL)
 599                 return o;
 600
 601         /*
 602          * Allocate new object. This may result in rather complicated
 603          * operations, including fld queries, inode loading, etc.
 604          */
 605         o = lu_object_alloc(env, dev, f, conf);
 606         if (unlikely(IS_ERR(o)))
 607                 return o;
 608
 609         LASSERT(lu_fid_eq(lu_object_fid(o), f));
 610
 611         cfs_hash_bd_lock(hs, &bd, 1);
 612
 613         shadow = htable_lookup(s, &bd, f, waiter, &version);
 614         if (likely(shadow == NULL)) {
 615                 struct lu_site_bkt_data *bkt;
 616
 617                 bkt = cfs_hash_bd_extra_get(hs, &bd);
 618                 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
 619                 bkt->lsb_busy++;
 620                 cfs_hash_bd_unlock(hs, &bd, 1);
 621                 return o;
 622         }
 623
 624         lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_RACE);
 625         cfs_hash_bd_unlock(hs, &bd, 1);
 626         lu_object_free(env, o);
 627         return shadow;
 628 }
 629
 630 /**
 631  * Much like lu_object_find(), but top level device of object is specifically
 632  * \a dev rather than top level device of the site. This interface allows
 633  * objects of different "stacking" to be created within the same site.
 634  */
 635 struct lu_object *lu_object_find_at(const struct lu_env *env,
 636                                     struct lu_device *dev,
 637                                     const struct lu_fid *f,
 638                                     const struct lu_object_conf *conf)
 639 {
 640         struct lu_site_bkt_data *bkt;
 641         struct lu_object        *obj;
 642         cfs_waitlink_t           wait;
 643
 644         while (1) {
 645                 obj = lu_object_find_try(env, dev, f, conf, &wait);
 646                 if (obj != ERR_PTR(-EAGAIN))
 647                         return obj;
 648                 /*
 649                  * lu_object_find_try() already added waiter into the
 650                  * wait queue.
 651                  */
 652                 cfs_waitq_wait(&wait, CFS_TASK_UNINT);
 653                 bkt = lu_site_bkt_from_fid(dev->ld_site, (void *)f);
 654                 cfs_waitq_del(&bkt->lsb_marche_funebre, &wait);
 655         }
 656 }
 657 EXPORT_SYMBOL(lu_object_find_at);
 658
 659 /**
 660  * Find object with given fid, and return its slice belonging to given device.
 661  */
 662 struct lu_object *lu_object_find_slice(const struct lu_env *env,
 663                                        struct lu_device *dev,
 664                                        const struct lu_fid *f,
 665                                        const struct lu_object_conf *conf)
 666 {
 667         struct lu_object *top;
 668         struct lu_object *obj;
 669
 670         top = lu_object_find(env, dev, f, conf);
 671         if (!IS_ERR(top)) {
 672                 obj = lu_object_locate(top->lo_header, dev->ld_type);
 673                 if (obj == NULL)
 674                         lu_object_put(env, top);
 675         } else
 676                 obj = top;
 677         return obj;
 678 }
 679 EXPORT_SYMBOL(lu_object_find_slice);
 680
 681 /**
 682  * Global list of all device types.
 683  */
 684 static CFS_LIST_HEAD(lu_device_types);
 685
 686 int lu_device_type_init(struct lu_device_type *ldt)
 687 {
 688         int result;
 689
 690         CFS_INIT_LIST_HEAD(&ldt->ldt_linkage);
 691         result = ldt->ldt_ops->ldto_init(ldt);
 692         if (result == 0)
 693                 cfs_list_add(&ldt->ldt_linkage, &lu_device_types);
 694         return result;
 695 }
 696 EXPORT_SYMBOL(lu_device_type_init);
 697
 698 void lu_device_type_fini(struct lu_device_type *ldt)
 699 {
 700         cfs_list_del_init(&ldt->ldt_linkage);
 701         ldt->ldt_ops->ldto_fini(ldt);
 702 }
 703 EXPORT_SYMBOL(lu_device_type_fini);
 704
 705 void lu_types_stop(void)
 706 {
 707         struct lu_device_type *ldt;
 708
 709         cfs_list_for_each_entry(ldt, &lu_device_types, ldt_linkage) {
 710                 if (ldt->ldt_device_nr == 0)
 711                         ldt->ldt_ops->ldto_stop(ldt);
 712         }
 713 }
 714 EXPORT_SYMBOL(lu_types_stop);
 715
 716 /**
 717  * Global list of all sites on this node
 718  */
 719 static CFS_LIST_HEAD(lu_sites);
 720 static CFS_DEFINE_MUTEX(lu_sites_guard);
 721
 722 /**
 723  * Global environment used by site shrinker.
 724  */
 725 static struct lu_env lu_shrink_env;
 726
 727 struct lu_site_print_arg {
 728         struct lu_env   *lsp_env;
 729         void            *lsp_cookie;
 730         lu_printer_t     lsp_printer;
 731 };
 732
 733 static int
 734 lu_site_obj_print(cfs_hash_t *hs, cfs_hash_bd_t *bd,
 735                   cfs_hlist_node_t *hnode, void *data)
 736 {
 737         struct lu_site_print_arg *arg = (struct lu_site_print_arg *)data;
 738         struct lu_object_header  *h;
 739
 740         h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
 741         if (!cfs_list_empty(&h->loh_layers)) {
 742                 const struct lu_object *o;
 743
 744                 o = lu_object_top(h);
 745                 lu_object_print(arg->lsp_env, arg->lsp_cookie,
 746                                 arg->lsp_printer, o);
 747         } else {
 748                 lu_object_header_print(arg->lsp_env, arg->lsp_cookie,
 749                                        arg->lsp_printer, h);
 750         }
 751         return 0;
 752 }
 753
 754 /**
 755  * Print all objects in \a s.
 756  */
 757 void lu_site_print(const struct lu_env *env, struct lu_site *s, void *cookie,
 758                    lu_printer_t printer)
 759 {
 760         struct lu_site_print_arg arg = {
 761                 .lsp_env     = (struct lu_env *)env,
 762                 .lsp_cookie  = cookie,
 763                 .lsp_printer = printer,
 764         };
 765
 766         cfs_hash_for_each(s->ls_obj_hash, lu_site_obj_print, &arg);
 767 }
 768 EXPORT_SYMBOL(lu_site_print);
 769
 770 enum {
 771         LU_CACHE_PERCENT_MAX     = 50,
 772         LU_CACHE_PERCENT_DEFAULT = 20
 773 };
 774
 775 static unsigned int lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
 776 CFS_MODULE_PARM(lu_cache_percent, "i", int, 0644,
 777                 "Percentage of memory to be used as lu_object cache");
 778
 779 /**
 780  * Return desired hash table order.
 781  */
 782 static int lu_htable_order(void)
 783 {
 784         unsigned long cache_size;
 785         int bits;
 786
 787         /*
 788          * Calculate hash table size, assuming that we want reasonable
 789          * performance when 20% of total memory is occupied by cache of
 790          * lu_objects.
 791          *
 792          * Size of lu_object is (arbitrary) taken as 1K (together with inode).
 793          */
 794         cache_size = cfs_num_physpages;
 795
 796 #if BITS_PER_LONG == 32
 797         /* limit hashtable size for lowmem systems to low RAM */
 798         if (cache_size > 1 << (30 - CFS_PAGE_SHIFT))
 799                 cache_size = 1 << (30 - CFS_PAGE_SHIFT) * 3 / 4;
 800 #endif
 801
 802         /* clear off unreasonable cache setting. */
 803         if (lu_cache_percent == 0 || lu_cache_percent > LU_CACHE_PERCENT_MAX) {
 804                 CWARN("obdclass: invalid lu_cache_percent: %u, it must be in"
 805                       " the range of (0, %u]. Will use default value: %u.\n",
 806                       lu_cache_percent, LU_CACHE_PERCENT_MAX,
 807                       LU_CACHE_PERCENT_DEFAULT);
 808
 809                 lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
 810         }
 811         cache_size = cache_size / 100 * lu_cache_percent *
 812                 (CFS_PAGE_SIZE / 1024);
 813
 814         for (bits = 1; (1 << bits) < cache_size; ++bits) {
 815                 ;
 816         }
 817         return bits;
 818 }
 819
 820 static unsigned lu_obj_hop_hash(cfs_hash_t *hs,
 821                                 const void *key, unsigned mask)
 822 {
 823         struct lu_fid  *fid = (struct lu_fid *)key;
 824         __u32           hash;
 825
 826         hash = fid_flatten32(fid);
 827         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
 828         hash = cfs_hash_long(hash, hs->hs_bkt_bits);
 829
 830         /* give me another random factor */
 831         hash -= cfs_hash_long((unsigned long)hs, fid_oid(fid) % 11 + 3);
 832
 833         hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
 834         hash |= (fid_seq(fid) + fid_oid(fid)) & (CFS_HASH_NBKT(hs) - 1);
 835
 836         return hash & mask;
 837 }
 838
 839 static void *lu_obj_hop_object(cfs_hlist_node_t *hnode)
 840 {
 841         return cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
 842 }
 843
 844 static void *lu_obj_hop_key(cfs_hlist_node_t *hnode)
 845 {
 846         struct lu_object_header *h;
 847
 848         h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
 849         return &h->loh_fid;
 850 }
 851
 852 static int lu_obj_hop_keycmp(const void *key, cfs_hlist_node_t *hnode)
 853 {
 854         struct lu_object_header *h;
 855
 856         h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
 857         return lu_fid_eq(&h->loh_fid, (struct lu_fid *)key);
 858 }
 859
 860 static void lu_obj_hop_get(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 861 {
 862         struct lu_object_header *h;
 863
 864         h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
 865         if (cfs_atomic_add_return(1, &h->loh_ref) == 1) {
 866                 struct lu_site_bkt_data *bkt;
 867                 cfs_hash_bd_t            bd;
 868
 869                 cfs_hash_bd_get(hs, &h->loh_fid, &bd);
 870                 bkt = cfs_hash_bd_extra_get(hs, &bd);
 871                 bkt->lsb_busy++;
 872         }
 873 }
 874
 875 static void lu_obj_hop_put_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 876 {
 877         LBUG(); /* we should never called it */
 878 }
 879
 880 cfs_hash_ops_t lu_site_hash_ops = {
 881         .hs_hash        = lu_obj_hop_hash,
 882         .hs_key         = lu_obj_hop_key,
 883         .hs_keycmp      = lu_obj_hop_keycmp,
 884         .hs_object      = lu_obj_hop_object,
 885         .hs_get         = lu_obj_hop_get,
 886         .hs_put_locked  = lu_obj_hop_put_locked,
 887 };
 888
 889 /**
 890  * Initialize site \a s, with \a d as the top level device.
 891  */
 892 #define LU_SITE_BITS_MIN    12
 893 #define LU_SITE_BITS_MAX    24
 894 /**
 895  * total 256 buckets, we don't want too many buckets because:
 896  * - consume too much memory
 897  * - avoid unbalanced LRU list
 898  */
 899 #define LU_SITE_BKT_BITS    8
 900
 901 int lu_site_init(struct lu_site *s, struct lu_device *top)
 902 {
 903         struct lu_site_bkt_data *bkt;
 904         cfs_hash_bd_t bd;
 905         char name[16];
 906         int bits;
 907         int i;
 908         ENTRY;
 909
 910         memset(s, 0, sizeof *s);
 911         bits = lu_htable_order();
 912         snprintf(name, 16, "lu_site_%s", top->ld_type->ldt_name);
 913         for (bits = min(max(LU_SITE_BITS_MIN, bits), LU_SITE_BITS_MAX);
 914              bits >= LU_SITE_BITS_MIN; bits--) {
 915                 s->ls_obj_hash = cfs_hash_create(name, bits, bits,
 916                                                  bits - LU_SITE_BKT_BITS,
 917                                                  sizeof(*bkt), 0, 0,
 918                                                  &lu_site_hash_ops,
 919                                                  CFS_HASH_SPIN_BKTLOCK |
 920                                                  CFS_HASH_NO_ITEMREF |
 921                                                  CFS_HASH_DEPTH |
 922                                                  CFS_HASH_ASSERT_EMPTY);
 923                 if (s->ls_obj_hash != NULL)
 924                         break;
 925         }
 926
 927         if (s->ls_obj_hash == NULL) {
 928                 CERROR("failed to create lu_site hash with bits: %d\n", bits);
 929                 return -ENOMEM;
 930         }
 931
 932         cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
 933                 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
 934                 CFS_INIT_LIST_HEAD(&bkt->lsb_lru);
 935                 cfs_waitq_init(&bkt->lsb_marche_funebre);
 936         }
 937
 938         s->ls_stats = lprocfs_alloc_stats(LU_SS_LAST_STAT, 0);
 939         if (s->ls_stats == NULL) {
 940                 cfs_hash_putref(s->ls_obj_hash);
 941                 s->ls_obj_hash = NULL;
 942                 return -ENOMEM;
 943         }
 944
 945         lprocfs_counter_init(s->ls_stats, LU_SS_CREATED,
 946                              0, "created", "created");
 947         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_HIT,
 948                              0, "cache_hit", "cache_hit");
 949         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_MISS,
 950                              0, "cache_miss", "cache_miss");
 951         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_RACE,
 952                              0, "cache_race", "cache_race");
 953         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_DEATH_RACE,
 954                              0, "cache_death_race", "cache_death_race");
 955         lprocfs_counter_init(s->ls_stats, LU_SS_LRU_PURGED,
 956                              0, "lru_purged", "lru_purged");
 957
 958         CFS_INIT_LIST_HEAD(&s->ls_linkage);
 959         s->ls_top_dev = top;
 960         top->ld_site = s;
 961         lu_device_get(top);
 962         lu_ref_add(&top->ld_reference, "site-top", s);
 963
 964         CFS_INIT_LIST_HEAD(&s->ls_ld_linkage);
 965         cfs_spin_lock_init(&s->ls_ld_lock);
 966
 967         cfs_spin_lock(&s->ls_ld_lock);
 968         cfs_list_add(&top->ld_linkage, &s->ls_ld_linkage);
 969         cfs_spin_unlock(&s->ls_ld_lock);
 970
 971         RETURN(0);
 972 }
 973 EXPORT_SYMBOL(lu_site_init);
 974
 975 /**
 976  * Finalize \a s and release its resources.
 977  */
 978 void lu_site_fini(struct lu_site *s)
 979 {
 980         cfs_mutex_lock(&lu_sites_guard);
 981         cfs_list_del_init(&s->ls_linkage);
 982         cfs_mutex_unlock(&lu_sites_guard);
 983
 984         if (s->ls_obj_hash != NULL) {
 985                 cfs_hash_putref(s->ls_obj_hash);
 986                 s->ls_obj_hash = NULL;
 987         }
 988
 989         if (s->ls_top_dev != NULL) {
 990                 s->ls_top_dev->ld_site = NULL;
 991                 lu_ref_del(&s->ls_top_dev->ld_reference, "site-top", s);
 992                 lu_device_put(s->ls_top_dev);
 993                 s->ls_top_dev = NULL;
 994         }
 995
 996         if (s->ls_stats != NULL)
 997                 lprocfs_free_stats(&s->ls_stats);
 998 }
 999 EXPORT_SYMBOL(lu_site_fini);
1000
1001 /**
1002  * Called when initialization of stack for this site is completed.
1003  */
1004 int lu_site_init_finish(struct lu_site *s)
1005 {
1006         int result;
1007         cfs_mutex_lock(&lu_sites_guard);
1008         result = lu_context_refill(&lu_shrink_env.le_ctx);
1009         if (result == 0)
1010                 cfs_list_add(&s->ls_linkage, &lu_sites);
1011         cfs_mutex_unlock(&lu_sites_guard);
1012         return result;
1013 }
1014 EXPORT_SYMBOL(lu_site_init_finish);
1015
1016 /**
1017  * Acquire additional reference on device \a d
1018  */
1019 void lu_device_get(struct lu_device *d)
1020 {
1021         cfs_atomic_inc(&d->ld_ref);
1022 }
1023 EXPORT_SYMBOL(lu_device_get);
1024
1025 /**
1026  * Release reference on device \a d.
1027  */
1028 void lu_device_put(struct lu_device *d)
1029 {
1030         LASSERT(cfs_atomic_read(&d->ld_ref) > 0);
1031         cfs_atomic_dec(&d->ld_ref);
1032 }
1033 EXPORT_SYMBOL(lu_device_put);
1034
1035 /**
1036  * Initialize device \a d of type \a t.
1037  */
1038 int lu_device_init(struct lu_device *d, struct lu_device_type *t)
1039 {
1040         if (t->ldt_device_nr++ == 0 && t->ldt_ops->ldto_start != NULL)
1041                 t->ldt_ops->ldto_start(t);
1042         memset(d, 0, sizeof *d);
1043         cfs_atomic_set(&d->ld_ref, 0);
1044         d->ld_type = t;
1045         lu_ref_init(&d->ld_reference);
1046         CFS_INIT_LIST_HEAD(&d->ld_linkage);
1047         return 0;
1048 }
1049 EXPORT_SYMBOL(lu_device_init);
1050
1051 /**
1052  * Finalize device \a d.
1053  */
1054 void lu_device_fini(struct lu_device *d)
1055 {
1056         struct lu_device_type *t;
1057
1058         t = d->ld_type;
1059         if (d->ld_obd != NULL) {
1060                 d->ld_obd->obd_lu_dev = NULL;
1061                 d->ld_obd = NULL;
1062         }
1063
1064         lu_ref_fini(&d->ld_reference);
1065         LASSERTF(cfs_atomic_read(&d->ld_ref) == 0,
1066                  "Refcount is %u\n", cfs_atomic_read(&d->ld_ref));
1067         LASSERT(t->ldt_device_nr > 0);
1068         if (--t->ldt_device_nr == 0 && t->ldt_ops->ldto_stop != NULL)
1069                 t->ldt_ops->ldto_stop(t);
1070 }
1071 EXPORT_SYMBOL(lu_device_fini);
1072
1073 /**
1074  * Initialize object \a o that is part of compound object \a h and was created
1075  * by device \a d.
1076  */
1077 int lu_object_init(struct lu_object *o,
1078                    struct lu_object_header *h, struct lu_device *d)
1079 {
1080         memset(o, 0, sizeof *o);
1081         o->lo_header = h;
1082         o->lo_dev    = d;
1083         lu_device_get(d);
1084         o->lo_dev_ref = lu_ref_add(&d->ld_reference, "lu_object", o);
1085         CFS_INIT_LIST_HEAD(&o->lo_linkage);
1086         return 0;
1087 }
1088 EXPORT_SYMBOL(lu_object_init);
1089
1090 /**
1091  * Finalize object and release its resources.
1092  */
1093 void lu_object_fini(struct lu_object *o)
1094 {
1095         struct lu_device *dev = o->lo_dev;
1096
1097         LASSERT(cfs_list_empty(&o->lo_linkage));
1098
1099         if (dev != NULL) {
1100                 lu_ref_del_at(&dev->ld_reference,
1101                               o->lo_dev_ref , "lu_object", o);
1102                 lu_device_put(dev);
1103                 o->lo_dev = NULL;
1104         }
1105 }
1106 EXPORT_SYMBOL(lu_object_fini);
1107
1108 /**
1109  * Add object \a o as first layer of compound object \a h
1110  *
1111  * This is typically called by the ->ldo_object_alloc() method of top-level
1112  * device.
1113  */
1114 void lu_object_add_top(struct lu_object_header *h, struct lu_object *o)
1115 {
1116         cfs_list_move(&o->lo_linkage, &h->loh_layers);
1117 }
1118 EXPORT_SYMBOL(lu_object_add_top);
1119
1120 /**
1121  * Add object \a o as a layer of compound object, going after \a before.
1122  *
1123  * This is typically called by the ->ldo_object_alloc() method of \a
1124  * before->lo_dev.
1125  */
1126 void lu_object_add(struct lu_object *before, struct lu_object *o)
1127 {
1128         cfs_list_move(&o->lo_linkage, &before->lo_linkage);
1129 }
1130 EXPORT_SYMBOL(lu_object_add);
1131
1132 /**
1133  * Initialize compound object.
1134  */
1135 int lu_object_header_init(struct lu_object_header *h)
1136 {
1137         memset(h, 0, sizeof *h);
1138         cfs_atomic_set(&h->loh_ref, 1);
1139         CFS_INIT_HLIST_NODE(&h->loh_hash);
1140         CFS_INIT_LIST_HEAD(&h->loh_lru);
1141         CFS_INIT_LIST_HEAD(&h->loh_layers);
1142         lu_ref_init(&h->loh_reference);
1143         return 0;
1144 }
1145 EXPORT_SYMBOL(lu_object_header_init);
1146
1147 /**
1148  * Finalize compound object.
1149  */
1150 void lu_object_header_fini(struct lu_object_header *h)
1151 {
1152         LASSERT(cfs_list_empty(&h->loh_layers));
1153         LASSERT(cfs_list_empty(&h->loh_lru));
1154         LASSERT(cfs_hlist_unhashed(&h->loh_hash));
1155         lu_ref_fini(&h->loh_reference);
1156 }
1157 EXPORT_SYMBOL(lu_object_header_fini);
1158
1159 /**
1160  * Given a compound object, find its slice, corresponding to the device type
1161  * \a dtype.
1162  */
1163 struct lu_object *lu_object_locate(struct lu_object_header *h,
1164                                    const struct lu_device_type *dtype)
1165 {
1166         struct lu_object *o;
1167
1168         cfs_list_for_each_entry(o, &h->loh_layers, lo_linkage) {
1169                 if (o->lo_dev->ld_type == dtype)
1170                         return o;
1171         }
1172         return NULL;
1173 }
1174 EXPORT_SYMBOL(lu_object_locate);
1175
1176
1177
1178 /**
1179  * Finalize and free devices in the device stack.
1180  *
1181  * Finalize device stack by purging object cache, and calling
1182  * lu_device_type_operations::ldto_device_fini() and
1183  * lu_device_type_operations::ldto_device_free() on all devices in the stack.
1184  */
1185 void lu_stack_fini(const struct lu_env *env, struct lu_device *top)
1186 {
1187         struct lu_site   *site = top->ld_site;
1188         struct lu_device *scan;
1189         struct lu_device *next;
1190
1191         lu_site_purge(env, site, ~0);
1192         for (scan = top; scan != NULL; scan = next) {
1193                 next = scan->ld_type->ldt_ops->ldto_device_fini(env, scan);
1194                 lu_ref_del(&scan->ld_reference, "lu-stack", &lu_site_init);
1195                 lu_device_put(scan);
1196         }
1197
1198         /* purge again. */
1199         lu_site_purge(env, site, ~0);
1200
1201         if (!cfs_hash_is_empty(site->ls_obj_hash)) {
1202                 /*
1203                  * Uh-oh, objects still exist.
1204                  */
1205                 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
1206
1207                 lu_site_print(env, site, &msgdata, lu_cdebug_printer);
1208         }
1209
1210         for (scan = top; scan != NULL; scan = next) {
1211                 const struct lu_device_type *ldt = scan->ld_type;
1212                 struct obd_type             *type;
1213
1214                 next = ldt->ldt_ops->ldto_device_free(env, scan);
1215                 type = ldt->ldt_obd_type;
1216                 if (type != NULL) {
1217                         type->typ_refcnt--;
1218                         class_put_type(type);
1219                 }
1220         }
1221 }
1222 EXPORT_SYMBOL(lu_stack_fini);
1223
1224 enum {
1225         /**
1226          * Maximal number of tld slots.
1227          */
1228         LU_CONTEXT_KEY_NR = 32
1229 };
1230
1231 static struct lu_context_key *lu_keys[LU_CONTEXT_KEY_NR] = { NULL, };
1232
1233 static cfs_spinlock_t lu_keys_guard = CFS_SPIN_LOCK_UNLOCKED;
1234
1235 /**
1236  * Global counter incremented whenever key is registered, unregistered,
1237  * revived or quiesced. This is used to void unnecessary calls to
1238  * lu_context_refill(). No locking is provided, as initialization and shutdown
1239  * are supposed to be externally serialized.
1240  */
1241 static unsigned key_set_version = 0;
1242
1243 /**
1244  * Register new key.
1245  */
1246 int lu_context_key_register(struct lu_context_key *key)
1247 {
1248         int result;
1249         int i;
1250
1251         LASSERT(key->lct_init != NULL);
1252         LASSERT(key->lct_fini != NULL);
1253         LASSERT(key->lct_tags != 0);
1254         LASSERT(key->lct_owner != NULL);
1255
1256         result = -ENFILE;
1257         cfs_spin_lock(&lu_keys_guard);
1258         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1259                 if (lu_keys[i] == NULL) {
1260                         key->lct_index = i;
1261                         cfs_atomic_set(&key->lct_used, 1);
1262                         lu_keys[i] = key;
1263                         lu_ref_init(&key->lct_reference);
1264                         result = 0;
1265                         ++key_set_version;
1266                         break;
1267                 }
1268         }
1269         cfs_spin_unlock(&lu_keys_guard);
1270         return result;
1271 }
1272 EXPORT_SYMBOL(lu_context_key_register);
1273
1274 static void key_fini(struct lu_context *ctx, int index)
1275 {
1276         if (ctx->lc_value != NULL && ctx->lc_value[index] != NULL) {
1277                 struct lu_context_key *key;
1278
1279                 key = lu_keys[index];
1280                 LASSERT(key != NULL);
1281                 LASSERT(key->lct_fini != NULL);
1282                 LASSERT(cfs_atomic_read(&key->lct_used) > 1);
1283
1284                 key->lct_fini(ctx, key, ctx->lc_value[index]);
1285                 lu_ref_del(&key->lct_reference, "ctx", ctx);
1286                 cfs_atomic_dec(&key->lct_used);
1287
1288                 LASSERT(key->lct_owner != NULL);
1289                 if ((ctx->lc_tags & LCT_NOREF) == 0) {
1290                         LINVRNT(cfs_module_refcount(key->lct_owner) > 0);
1291                         cfs_module_put(key->lct_owner);
1292                 }
1293                 ctx->lc_value[index] = NULL;
1294         }
1295 }
1296
1297 /**
1298  * Deregister key.
1299  */
1300 void lu_context_key_degister(struct lu_context_key *key)
1301 {
1302         LASSERT(cfs_atomic_read(&key->lct_used) >= 1);
1303         LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1304
1305         lu_context_key_quiesce(key);
1306
1307         ++key_set_version;
1308         cfs_spin_lock(&lu_keys_guard);
1309         key_fini(&lu_shrink_env.le_ctx, key->lct_index);
1310         if (lu_keys[key->lct_index]) {
1311                 lu_keys[key->lct_index] = NULL;
1312                 lu_ref_fini(&key->lct_reference);
1313         }
1314         cfs_spin_unlock(&lu_keys_guard);
1315
1316         LASSERTF(cfs_atomic_read(&key->lct_used) == 1,
1317                  "key has instances: %d\n",
1318                  cfs_atomic_read(&key->lct_used));
1319 }
1320 EXPORT_SYMBOL(lu_context_key_degister);
1321
1322 /**
1323  * Register a number of keys. This has to be called after all keys have been
1324  * initialized by a call to LU_CONTEXT_KEY_INIT().
1325  */
1326 int lu_context_key_register_many(struct lu_context_key *k, ...)
1327 {
1328         struct lu_context_key *key = k;
1329         va_list args;
1330         int result;
1331
1332         va_start(args, k);
1333         do {
1334                 result = lu_context_key_register(key);
1335                 if (result)
1336                         break;
1337                 key = va_arg(args, struct lu_context_key *);
1338         } while (key != NULL);
1339         va_end(args);
1340
1341         if (result != 0) {
1342                 va_start(args, k);
1343                 while (k != key) {
1344                         lu_context_key_degister(k);
1345                         k = va_arg(args, struct lu_context_key *);
1346                 }
1347                 va_end(args);
1348         }
1349
1350         return result;
1351 }
1352 EXPORT_SYMBOL(lu_context_key_register_many);
1353
1354 /**
1355  * De-register a number of keys. This is a dual to
1356  * lu_context_key_register_many().
1357  */
1358 void lu_context_key_degister_many(struct lu_context_key *k, ...)
1359 {
1360         va_list args;
1361
1362         va_start(args, k);
1363         do {
1364                 lu_context_key_degister(k);
1365                 k = va_arg(args, struct lu_context_key*);
1366         } while (k != NULL);
1367         va_end(args);
1368 }
1369 EXPORT_SYMBOL(lu_context_key_degister_many);
1370
1371 /**
1372  * Revive a number of keys.
1373  */
1374 void lu_context_key_revive_many(struct lu_context_key *k, ...)
1375 {
1376         va_list args;
1377
1378         va_start(args, k);
1379         do {
1380                 lu_context_key_revive(k);
1381                 k = va_arg(args, struct lu_context_key*);
1382         } while (k != NULL);
1383         va_end(args);
1384 }
1385 EXPORT_SYMBOL(lu_context_key_revive_many);
1386
1387 /**
1388  * Quiescent a number of keys.
1389  */
1390 void lu_context_key_quiesce_many(struct lu_context_key *k, ...)
1391 {
1392         va_list args;
1393
1394         va_start(args, k);
1395         do {
1396                 lu_context_key_quiesce(k);
1397                 k = va_arg(args, struct lu_context_key*);
1398         } while (k != NULL);
1399         va_end(args);
1400 }
1401 EXPORT_SYMBOL(lu_context_key_quiesce_many);
1402
1403 /**
1404  * Return value associated with key \a key in context \a ctx.
1405  */
1406 void *lu_context_key_get(const struct lu_context *ctx,
1407                          const struct lu_context_key *key)
1408 {
1409         LINVRNT(ctx->lc_state == LCS_ENTERED);
1410         LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1411         LASSERT(lu_keys[key->lct_index] == key);
1412         return ctx->lc_value[key->lct_index];
1413 }
1414 EXPORT_SYMBOL(lu_context_key_get);
1415
1416 /**
1417  * List of remembered contexts. XXX document me.
1418  */
1419 static CFS_LIST_HEAD(lu_context_remembered);
1420
1421 /**
1422  * Destroy \a key in all remembered contexts. This is used to destroy key
1423  * values in "shared" contexts (like service threads), when a module owning
1424  * the key is about to be unloaded.
1425  */
1426 void lu_context_key_quiesce(struct lu_context_key *key)
1427 {
1428         struct lu_context *ctx;
1429         extern unsigned cl_env_cache_purge(unsigned nr);
1430
1431         if (!(key->lct_tags & LCT_QUIESCENT)) {
1432                 /*
1433                  * XXX layering violation.
1434                  */
1435                 cl_env_cache_purge(~0);
1436                 key->lct_tags |= LCT_QUIESCENT;
1437                 /*
1438                  * XXX memory barrier has to go here.
1439                  */
1440                 cfs_spin_lock(&lu_keys_guard);
1441                 cfs_list_for_each_entry(ctx, &lu_context_remembered,
1442                                         lc_remember)
1443                         key_fini(ctx, key->lct_index);
1444                 cfs_spin_unlock(&lu_keys_guard);
1445                 ++key_set_version;
1446         }
1447 }
1448 EXPORT_SYMBOL(lu_context_key_quiesce);
1449
1450 void lu_context_key_revive(struct lu_context_key *key)
1451 {
1452         key->lct_tags &= ~LCT_QUIESCENT;
1453         ++key_set_version;
1454 }
1455 EXPORT_SYMBOL(lu_context_key_revive);
1456
1457 static void keys_fini(struct lu_context *ctx)
1458 {
1459         int     i;
1460
1461         if (ctx->lc_value == NULL)
1462                 return;
1463
1464         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i)
1465                 key_fini(ctx, i);
1466
1467         OBD_FREE(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1468         ctx->lc_value = NULL;
1469 }
1470
1471 static int keys_fill(struct lu_context *ctx)
1472 {
1473         int i;
1474
1475         LINVRNT(ctx->lc_value != NULL);
1476         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1477                 struct lu_context_key *key;
1478
1479                 key = lu_keys[i];
1480                 if (ctx->lc_value[i] == NULL && key != NULL &&
1481                     (key->lct_tags & ctx->lc_tags) &&
1482                     /*
1483                      * Don't create values for a LCT_QUIESCENT key, as this
1484                      * will pin module owning a key.
1485                      */
1486                     !(key->lct_tags & LCT_QUIESCENT)) {
1487                         void *value;
1488
1489                         LINVRNT(key->lct_init != NULL);
1490                         LINVRNT(key->lct_index == i);
1491
1492                         value = key->lct_init(ctx, key);
1493                         if (unlikely(IS_ERR(value)))
1494                                 return PTR_ERR(value);
1495
1496                         LASSERT(key->lct_owner != NULL);
1497                         if (!(ctx->lc_tags & LCT_NOREF))
1498                                 cfs_try_module_get(key->lct_owner);
1499                         lu_ref_add_atomic(&key->lct_reference, "ctx", ctx);
1500                         cfs_atomic_inc(&key->lct_used);
1501                         /*
1502                          * This is the only place in the code, where an
1503                          * element of ctx->lc_value[] array is set to non-NULL
1504                          * value.
1505                          */
1506                         ctx->lc_value[i] = value;
1507                         if (key->lct_exit != NULL)
1508                                 ctx->lc_tags |= LCT_HAS_EXIT;
1509                 }
1510                 ctx->lc_version = key_set_version;
1511         }
1512         return 0;
1513 }
1514
1515 static int keys_init(struct lu_context *ctx)
1516 {
1517         OBD_ALLOC(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1518         if (likely(ctx->lc_value != NULL))
1519                 return keys_fill(ctx);
1520
1521         return -ENOMEM;
1522 }
1523
1524 /**
1525  * Initialize context data-structure. Create values for all keys.
1526  */
1527 int lu_context_init(struct lu_context *ctx, __u32 tags)
1528 {
1529         int     rc;
1530
1531         memset(ctx, 0, sizeof *ctx);
1532         ctx->lc_state = LCS_INITIALIZED;
1533         ctx->lc_tags = tags;
1534         if (tags & LCT_REMEMBER) {
1535                 cfs_spin_lock(&lu_keys_guard);
1536                 cfs_list_add(&ctx->lc_remember, &lu_context_remembered);
1537                 cfs_spin_unlock(&lu_keys_guard);
1538         } else {
1539                 CFS_INIT_LIST_HEAD(&ctx->lc_remember);
1540         }
1541
1542         rc = keys_init(ctx);
1543         if (rc != 0)
1544                 lu_context_fini(ctx);
1545
1546         return rc;
1547 }
1548 EXPORT_SYMBOL(lu_context_init);
1549
1550 /**
1551  * Finalize context data-structure. Destroy key values.
1552  */
1553 void lu_context_fini(struct lu_context *ctx)
1554 {
1555         LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1556         ctx->lc_state = LCS_FINALIZED;
1557
1558         if ((ctx->lc_tags & LCT_REMEMBER) == 0) {
1559                 LASSERT(cfs_list_empty(&ctx->lc_remember));
1560                 keys_fini(ctx);
1561
1562         } else { /* could race with key degister */
1563                 cfs_spin_lock(&lu_keys_guard);
1564                 keys_fini(ctx);
1565                 cfs_list_del_init(&ctx->lc_remember);
1566                 cfs_spin_unlock(&lu_keys_guard);
1567         }
1568 }
1569 EXPORT_SYMBOL(lu_context_fini);
1570
1571 /**
1572  * Called before entering context.
1573  */
1574 void lu_context_enter(struct lu_context *ctx)
1575 {
1576         LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1577         ctx->lc_state = LCS_ENTERED;
1578 }
1579 EXPORT_SYMBOL(lu_context_enter);
1580
1581 /**
1582  * Called after exiting from \a ctx
1583  */
1584 void lu_context_exit(struct lu_context *ctx)
1585 {
1586         int i;
1587
1588         LINVRNT(ctx->lc_state == LCS_ENTERED);
1589         ctx->lc_state = LCS_LEFT;
1590         if (ctx->lc_tags & LCT_HAS_EXIT && ctx->lc_value != NULL) {
1591                 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1592                         if (ctx->lc_value[i] != NULL) {
1593                                 struct lu_context_key *key;
1594
1595                                 key = lu_keys[i];
1596                                 LASSERT(key != NULL);
1597                                 if (key->lct_exit != NULL)
1598                                         key->lct_exit(ctx,
1599                                                       key, ctx->lc_value[i]);
1600                         }
1601                 }
1602         }
1603 }
1604 EXPORT_SYMBOL(lu_context_exit);
1605
1606 /**
1607  * Allocate for context all missing keys that were registered after context
1608  * creation. key_set_version is only changed in rare cases when modules
1609  * are loaded and removed.
1610  */
1611 int lu_context_refill(struct lu_context *ctx)
1612 {
1613         return likely(ctx->lc_version == key_set_version) ? 0 : keys_fill(ctx);
1614 }
1615 EXPORT_SYMBOL(lu_context_refill);
1616
1617 /**
1618  * lu_ctx_tags/lu_ses_tags will be updated if there are new types of
1619  * obd being added. Currently, this is only used on client side, specifically
1620  * for echo device client, for other stack (like ptlrpc threads), context are
1621  * predefined when the lu_device type are registered, during the module probe
1622  * phase.
1623  */
1624 __u32 lu_context_tags_default = 0;
1625 __u32 lu_session_tags_default = 0;
1626
1627 void lu_context_tags_update(__u32 tags)
1628 {
1629         cfs_spin_lock(&lu_keys_guard);
1630         lu_context_tags_default |= tags;
1631         key_set_version ++;
1632         cfs_spin_unlock(&lu_keys_guard);
1633 }
1634 EXPORT_SYMBOL(lu_context_tags_update);
1635
1636 void lu_context_tags_clear(__u32 tags)
1637 {
1638         cfs_spin_lock(&lu_keys_guard);
1639         lu_context_tags_default &= ~tags;
1640         key_set_version ++;
1641         cfs_spin_unlock(&lu_keys_guard);
1642 }
1643 EXPORT_SYMBOL(lu_context_tags_clear);
1644
1645 void lu_session_tags_update(__u32 tags)
1646 {
1647         cfs_spin_lock(&lu_keys_guard);
1648         lu_session_tags_default |= tags;
1649         key_set_version ++;
1650         cfs_spin_unlock(&lu_keys_guard);
1651 }
1652 EXPORT_SYMBOL(lu_session_tags_update);
1653
1654 void lu_session_tags_clear(__u32 tags)
1655 {
1656         cfs_spin_lock(&lu_keys_guard);
1657         lu_session_tags_default &= ~tags;
1658         key_set_version ++;
1659         cfs_spin_unlock(&lu_keys_guard);
1660 }
1661 EXPORT_SYMBOL(lu_session_tags_clear);
1662
1663 int lu_env_init(struct lu_env *env, __u32 tags)
1664 {
1665         int result;
1666
1667         env->le_ses = NULL;
1668         result = lu_context_init(&env->le_ctx, tags);
1669         if (likely(result == 0))
1670                 lu_context_enter(&env->le_ctx);
1671         return result;
1672 }
1673 EXPORT_SYMBOL(lu_env_init);
1674
1675 void lu_env_fini(struct lu_env *env)
1676 {
1677         lu_context_exit(&env->le_ctx);
1678         lu_context_fini(&env->le_ctx);
1679         env->le_ses = NULL;
1680 }
1681 EXPORT_SYMBOL(lu_env_fini);
1682
1683 int lu_env_refill(struct lu_env *env)
1684 {
1685         int result;
1686
1687         result = lu_context_refill(&env->le_ctx);
1688         if (result == 0 && env->le_ses != NULL)
1689                 result = lu_context_refill(env->le_ses);
1690         return result;
1691 }
1692 EXPORT_SYMBOL(lu_env_refill);
1693
1694 /**
1695  * Currently, this API will only be used by echo client.
1696  * Because echo client and normal lustre client will share
1697  * same cl_env cache. So echo client needs to refresh
1698  * the env context after it get one from the cache, especially
1699  * when normal client and echo client co-exist in the same client.
1700  */
1701 int lu_env_refill_by_tags(struct lu_env *env, __u32 ctags,
1702                           __u32 stags)
1703 {
1704         int    result;
1705
1706         if ((env->le_ctx.lc_tags & ctags) != ctags) {
1707                 env->le_ctx.lc_version = 0;
1708                 env->le_ctx.lc_tags |= ctags;
1709         }
1710
1711         if (env->le_ses && (env->le_ses->lc_tags & stags) != stags) {
1712                 env->le_ses->lc_version = 0;
1713                 env->le_ses->lc_tags |= stags;
1714         }
1715
1716         result = lu_env_refill(env);
1717
1718         return result;
1719 }
1720 EXPORT_SYMBOL(lu_env_refill_by_tags);
1721
1722 static struct cfs_shrinker *lu_site_shrinker = NULL;
1723
1724 typedef struct lu_site_stats{
1725         unsigned        lss_populated;
1726         unsigned        lss_max_search;
1727         unsigned        lss_total;
1728         unsigned        lss_busy;
1729 } lu_site_stats_t;
1730
1731 static void lu_site_stats_get(cfs_hash_t *hs,
1732                               lu_site_stats_t *stats, int populated)
1733 {
1734         cfs_hash_bd_t bd;
1735         int           i;
1736
1737         cfs_hash_for_each_bucket(hs, &bd, i) {
1738                 struct lu_site_bkt_data *bkt = cfs_hash_bd_extra_get(hs, &bd);
1739                 cfs_hlist_head_t        *hhead;
1740
1741                 cfs_hash_bd_lock(hs, &bd, 1);
1742                 stats->lss_busy  += bkt->lsb_busy;
1743                 stats->lss_total += cfs_hash_bd_count_get(&bd);
1744                 stats->lss_max_search = max((int)stats->lss_max_search,
1745                                             cfs_hash_bd_depmax_get(&bd));
1746                 if (!populated) {
1747                         cfs_hash_bd_unlock(hs, &bd, 1);
1748                         continue;
1749                 }
1750
1751                 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
1752                         if (!cfs_hlist_empty(hhead))
1753                                 stats->lss_populated++;
1754                 }
1755                 cfs_hash_bd_unlock(hs, &bd, 1);
1756         }
1757 }
1758
1759 #ifdef __KERNEL__
1760
1761 static int lu_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
1762 {
1763         lu_site_stats_t stats;
1764         struct lu_site *s;
1765         struct lu_site *tmp;
1766         int cached = 0;
1767         int remain = shrink_param(sc, nr_to_scan);
1768         CFS_LIST_HEAD(splice);
1769
1770         if (remain != 0) {
1771                 if (!(shrink_param(sc, gfp_mask) & __GFP_FS))
1772                         return -1;
1773                 CDEBUG(D_INODE, "Shrink %d objects\n", remain);
1774         }
1775
1776         cfs_mutex_lock(&lu_sites_guard);
1777         cfs_list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
1778                 if (shrink_param(sc, nr_to_scan) != 0) {
1779                         remain = lu_site_purge(&lu_shrink_env, s, remain);
1780                         /*
1781                          * Move just shrunk site to the tail of site list to
1782                          * assure shrinking fairness.
1783                          */
1784                         cfs_list_move_tail(&s->ls_linkage, &splice);
1785                 }
1786
1787                 memset(&stats, 0, sizeof(stats));
1788                 lu_site_stats_get(s->ls_obj_hash, &stats, 0);
1789                 cached += stats.lss_total - stats.lss_busy;
1790                 if (shrink_param(sc, nr_to_scan) && remain <= 0)
1791                         break;
1792         }
1793         cfs_list_splice(&splice, lu_sites.prev);
1794         cfs_mutex_unlock(&lu_sites_guard);
1795
1796         cached = (cached / 100) * sysctl_vfs_cache_pressure;
1797         if (shrink_param(sc, nr_to_scan) == 0)
1798                 CDEBUG(D_INODE, "%d objects cached\n", cached);
1799         return cached;
1800 }
1801
1802 /*
1803  * Debugging stuff.
1804  */
1805
1806 /**
1807  * Environment to be used in debugger, contains all tags.
1808  */
1809 struct lu_env lu_debugging_env;
1810
1811 /**
1812  * Debugging printer function using printk().
1813  */
1814 int lu_printk_printer(const struct lu_env *env,
1815                       void *unused, const char *format, ...)
1816 {
1817         va_list args;
1818
1819         va_start(args, format);
1820         vprintk(format, args);
1821         va_end(args);
1822         return 0;
1823 }
1824
1825 void lu_debugging_setup(void)
1826 {
1827         lu_env_init(&lu_debugging_env, ~0);
1828 }
1829
1830 void lu_context_keys_dump(void)
1831 {
1832         int i;
1833
1834         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1835                 struct lu_context_key *key;
1836
1837                 key = lu_keys[i];
1838                 if (key != NULL) {
1839                         CERROR("[%d]: %p %x (%p,%p,%p) %d %d \"%s\"@%p\n",
1840                                i, key, key->lct_tags,
1841                                key->lct_init, key->lct_fini, key->lct_exit,
1842                                key->lct_index, cfs_atomic_read(&key->lct_used),
1843                                key->lct_owner ? key->lct_owner->name : "",
1844                                key->lct_owner);
1845                         lu_ref_print(&key->lct_reference);
1846                 }
1847         }
1848 }
1849 EXPORT_SYMBOL(lu_context_keys_dump);
1850 #else  /* !__KERNEL__ */
1851 static int lu_cache_shrink(int nr, unsigned int gfp_mask)
1852 {
1853         return 0;
1854 }
1855 #endif /* __KERNEL__ */
1856
1857 int  cl_global_init(void);
1858 void cl_global_fini(void);
1859 int  lu_ref_global_init(void);
1860 void lu_ref_global_fini(void);
1861
1862 int dt_global_init(void);
1863 void dt_global_fini(void);
1864
1865 int llo_global_init(void);
1866 void llo_global_fini(void);
1867
1868 /**
1869  * Initialization of global lu_* data.
1870  */
1871 int lu_global_init(void)
1872 {
1873         int result;
1874
1875         CDEBUG(D_INFO, "Lustre LU module (%p).\n", &lu_keys);
1876
1877         result = lu_ref_global_init();
1878         if (result != 0)
1879                 return result;
1880
1881         LU_CONTEXT_KEY_INIT(&lu_global_key);
1882         result = lu_context_key_register(&lu_global_key);
1883         if (result != 0)
1884                 return result;
1885         /*
1886          * At this level, we don't know what tags are needed, so allocate them
1887          * conservatively. This should not be too bad, because this
1888          * environment is global.
1889          */
1890         cfs_mutex_lock(&lu_sites_guard);
1891         result = lu_env_init(&lu_shrink_env, LCT_SHRINKER);
1892         cfs_mutex_unlock(&lu_sites_guard);
1893         if (result != 0)
1894                 return result;
1895
1896         /*
1897          * seeks estimation: 3 seeks to read a record from oi, one to read
1898          * inode, one for ea. Unfortunately setting this high value results in
1899          * lu_object/inode cache consuming all the memory.
1900          */
1901         lu_site_shrinker = cfs_set_shrinker(CFS_DEFAULT_SEEKS, lu_cache_shrink);
1902         if (lu_site_shrinker == NULL)
1903                 return -ENOMEM;
1904
1905         result = lu_time_global_init();
1906         if (result)
1907                 GOTO(out, result);
1908
1909 #ifdef __KERNEL__
1910         result = dt_global_init();
1911         if (result)
1912                 GOTO(out, result);
1913
1914         result = llo_global_init();
1915         if (result)
1916                 GOTO(out, result);
1917 #endif
1918         result = cl_global_init();
1919 out:
1920
1921         return result;
1922 }
1923
1924 /**
1925  * Dual to lu_global_init().
1926  */
1927 void lu_global_fini(void)
1928 {
1929         cl_global_fini();
1930 #ifdef __KERNEL__
1931         llo_global_fini();
1932         dt_global_fini();
1933 #endif
1934         lu_time_global_fini();
1935         if (lu_site_shrinker != NULL) {
1936                 cfs_remove_shrinker(lu_site_shrinker);
1937                 lu_site_shrinker = NULL;
1938         }
1939
1940         lu_context_key_degister(&lu_global_key);
1941
1942         /*
1943          * Tear shrinker environment down _after_ de-registering
1944          * lu_global_key, because the latter has a value in the former.
1945          */
1946         cfs_mutex_lock(&lu_sites_guard);
1947         lu_env_fini(&lu_shrink_env);
1948         cfs_mutex_unlock(&lu_sites_guard);
1949
1950         lu_ref_global_fini();
1951 }
1952
1953 struct lu_buf LU_BUF_NULL = {
1954         .lb_buf = NULL,
1955         .lb_len = 0
1956 };
1957 EXPORT_SYMBOL(LU_BUF_NULL);
1958
1959 static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx)
1960 {
1961 #ifdef LPROCFS
1962         struct lprocfs_counter ret;
1963
1964         lprocfs_stats_collect(stats, idx, &ret);
1965         return (__u32)ret.lc_count;
1966 #else
1967         return 0;
1968 #endif
1969 }
1970
1971 /**
1972  * Output site statistical counters into a buffer. Suitable for
1973  * lprocfs_rd_*()-style functions.
1974  */
1975 int lu_site_stats_print(const struct lu_site *s, char *page, int count)
1976 {
1977         lu_site_stats_t stats;
1978
1979         memset(&stats, 0, sizeof(stats));
1980         lu_site_stats_get(s->ls_obj_hash, &stats, 1);
1981
1982         return snprintf(page, count, "%d/%d %d/%d %d %d %d %d %d %d %d\n",
1983                         stats.lss_busy,
1984                         stats.lss_total,
1985                         stats.lss_populated,
1986                         CFS_HASH_NHLIST(s->ls_obj_hash),
1987                         stats.lss_max_search,
1988                         ls_stats_read(s->ls_stats, LU_SS_CREATED),
1989                         ls_stats_read(s->ls_stats, LU_SS_CACHE_HIT),
1990                         ls_stats_read(s->ls_stats, LU_SS_CACHE_MISS),
1991                         ls_stats_read(s->ls_stats, LU_SS_CACHE_RACE),
1992                         ls_stats_read(s->ls_stats, LU_SS_CACHE_DEATH_RACE),
1993                         ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED));
1994 }
1995 EXPORT_SYMBOL(lu_site_stats_print);
1996
1997 const char *lu_time_names[LU_TIME_NR] = {
1998         [LU_TIME_FIND_LOOKUP] = "find_lookup",
1999         [LU_TIME_FIND_ALLOC]  = "find_alloc",
2000         [LU_TIME_FIND_INSERT] = "find_insert"
2001 };
2002 EXPORT_SYMBOL(lu_time_names);
2003
2004 /**
2005  * Helper function to initialize a number of kmem slab caches at once.
2006  */
2007 int lu_kmem_init(struct lu_kmem_descr *caches)
2008 {
2009         int result;
2010         struct lu_kmem_descr *iter = caches;
2011
2012         for (result = 0; iter->ckd_cache != NULL; ++iter) {
2013                 *iter->ckd_cache = cfs_mem_cache_create(iter->ckd_name,
2014                                                         iter->ckd_size,
2015                                                         0, 0);
2016                 if (*iter->ckd_cache == NULL) {
2017                         result = -ENOMEM;
2018                         /* free all previously allocated caches */
2019                         lu_kmem_fini(caches);
2020                         break;
2021                 }
2022         }
2023         return result;
2024 }
2025 EXPORT_SYMBOL(lu_kmem_init);
2026
2027 /**
2028  * Helper function to finalize a number of kmem slab cached at once. Dual to
2029  * lu_kmem_init().
2030  */
2031 void lu_kmem_fini(struct lu_kmem_descr *caches)
2032 {
2033         int rc;
2034
2035         for (; caches->ckd_cache != NULL; ++caches) {
2036                 if (*caches->ckd_cache != NULL) {
2037                         rc = cfs_mem_cache_destroy(*caches->ckd_cache);
2038                         LASSERTF(rc == 0, "couldn't destroy %s slab\n",
2039                                  caches->ckd_name);
2040                         *caches->ckd_cache = NULL;
2041                 }
2042         }
2043 }
2044 EXPORT_SYMBOL(lu_kmem_fini);