lustre/obdclass/lu_object.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  *
  32  * Copyright (c) 2011, 2012, Whamcloud, Inc.
  33  */
  34 /*
  35  * This file is part of Lustre, http://www.lustre.org/
  36  * Lustre is a trademark of Sun Microsystems, Inc.
  37  *
  38  * lustre/obdclass/lu_object.c
  39  *
  40  * Lustre Object.
  41  * These are the only exported functions, they provide some generic
  42  * infrastructure for managing object devices
  43  *
  44  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  45  */
  46
  47 #define DEBUG_SUBSYSTEM S_CLASS
  48 #ifndef EXPORT_SYMTAB
  49 # define EXPORT_SYMTAB
  50 #endif
  51
  52 #include <libcfs/libcfs.h>
  53
  54 #ifdef __KERNEL__
  55 # include <linux/module.h>
  56 #endif
  57
  58 /* hash_long() */
  59 #include <libcfs/libcfs_hash.h>
  60 #include <obd_class.h>
  61 #include <obd_support.h>
  62 #include <lustre_disk.h>
  63 #include <lustre_fid.h>
  64 #include <lu_object.h>
  65 #include <libcfs/list.h>
  66 /* lu_time_global_{init,fini}() */
  67 #include <lu_time.h>
  68
  69 static void lu_object_free(const struct lu_env *env, struct lu_object *o);
  70
  71 /**
  72  * Decrease reference counter on object. If last reference is freed, return
  73  * object to the cache, unless lu_object_is_dying(o) holds. In the latter
  74  * case, free object immediately.
  75  */
  76 void lu_object_put(const struct lu_env *env, struct lu_object *o)
  77 {
  78         struct lu_site_bkt_data *bkt;
  79         struct lu_object_header *top;
  80         struct lu_site          *site;
  81         struct lu_object        *orig;
  82         cfs_hash_bd_t            bd;
  83
  84         top  = o->lo_header;
  85         site = o->lo_dev->ld_site;
  86         orig = o;
  87
  88         cfs_hash_bd_get(site->ls_obj_hash, &top->loh_fid, &bd);
  89         bkt = cfs_hash_bd_extra_get(site->ls_obj_hash, &bd);
  90
  91         if (!cfs_hash_bd_dec_and_lock(site->ls_obj_hash, &bd, &top->loh_ref)) {
  92                 if (lu_object_is_dying(top)) {
  93
  94                         /*
  95                          * somebody may be waiting for this, currently only
  96                          * used for cl_object, see cl_object_put_last().
  97                          */
  98                         cfs_waitq_broadcast(&bkt->lsb_marche_funebre);
  99                 }
 100                 return;
 101         }
 102
 103         LASSERT(bkt->lsb_busy > 0);
 104         bkt->lsb_busy--;
 105         /*
 106          * When last reference is released, iterate over object
 107          * layers, and notify them that object is no longer busy.
 108          */
 109         cfs_list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
 110                 if (o->lo_ops->loo_object_release != NULL)
 111                         o->lo_ops->loo_object_release(env, o);
 112         }
 113
 114         if (!lu_object_is_dying(top)) {
 115                 LASSERT(cfs_list_empty(&top->loh_lru));
 116                 cfs_list_add_tail(&top->loh_lru, &bkt->lsb_lru);
 117                 cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
 118                 return;
 119         }
 120
 121         /*
 122          * If object is dying (will not be cached), removed it
 123          * from hash table and LRU.
 124          *
 125          * This is done with hash table and LRU lists locked. As the only
 126          * way to acquire first reference to previously unreferenced
 127          * object is through hash-table lookup (lu_object_find()),
 128          * or LRU scanning (lu_site_purge()), that are done under hash-table
 129          * and LRU lock, no race with concurrent object lookup is possible
 130          * and we can safely destroy object below.
 131          */
 132         cfs_hash_bd_del_locked(site->ls_obj_hash, &bd, &top->loh_hash);
 133         cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
 134         /*
 135          * Object was already removed from hash and lru above, can
 136          * kill it.
 137          */
 138         lu_object_free(env, orig);
 139 }
 140 EXPORT_SYMBOL(lu_object_put);
 141
 142 /**
 143  * Allocate new object.
 144  *
 145  * This follows object creation protocol, described in the comment within
 146  * struct lu_device_operations definition.
 147  */
 148 static struct lu_object *lu_object_alloc(const struct lu_env *env,
 149                                          struct lu_device *dev,
 150                                          const struct lu_fid *f,
 151                                          const struct lu_object_conf *conf)
 152 {
 153         struct lu_object *scan;
 154         struct lu_object *top;
 155         cfs_list_t *layers;
 156         int clean;
 157         int result;
 158         ENTRY;
 159
 160         /*
 161          * Create top-level object slice. This will also create
 162          * lu_object_header.
 163          */
 164         top = dev->ld_ops->ldo_object_alloc(env, NULL, dev);
 165         if (top == NULL)
 166                 RETURN(ERR_PTR(-ENOMEM));
 167         /*
 168          * This is the only place where object fid is assigned. It's constant
 169          * after this point.
 170          */
 171         LASSERT(fid_is_igif(f) || fid_ver(f) == 0);
 172         top->lo_header->loh_fid = *f;
 173         layers = &top->lo_header->loh_layers;
 174         do {
 175                 /*
 176                  * Call ->loo_object_init() repeatedly, until no more new
 177                  * object slices are created.
 178                  */
 179                 clean = 1;
 180                 cfs_list_for_each_entry(scan, layers, lo_linkage) {
 181                         if (scan->lo_flags & LU_OBJECT_ALLOCATED)
 182                                 continue;
 183                         clean = 0;
 184                         scan->lo_header = top->lo_header;
 185                         result = scan->lo_ops->loo_object_init(env, scan, conf);
 186                         if (result != 0) {
 187                                 lu_object_free(env, top);
 188                                 RETURN(ERR_PTR(result));
 189                         }
 190                         scan->lo_flags |= LU_OBJECT_ALLOCATED;
 191                 }
 192         } while (!clean);
 193
 194         cfs_list_for_each_entry_reverse(scan, layers, lo_linkage) {
 195                 if (scan->lo_ops->loo_object_start != NULL) {
 196                         result = scan->lo_ops->loo_object_start(env, scan);
 197                         if (result != 0) {
 198                                 lu_object_free(env, top);
 199                                 RETURN(ERR_PTR(result));
 200                         }
 201                 }
 202         }
 203
 204         lprocfs_counter_incr(dev->ld_site->ls_stats, LU_SS_CREATED);
 205         RETURN(top);
 206 }
 207
 208 /**
 209  * Free an object.
 210  */
 211 static void lu_object_free(const struct lu_env *env, struct lu_object *o)
 212 {
 213         struct lu_site_bkt_data *bkt;
 214         struct lu_site          *site;
 215         struct lu_object        *scan;
 216         cfs_list_t              *layers;
 217         cfs_list_t               splice;
 218
 219         site   = o->lo_dev->ld_site;
 220         layers = &o->lo_header->loh_layers;
 221         bkt    = lu_site_bkt_from_fid(site, &o->lo_header->loh_fid);
 222         /*
 223          * First call ->loo_object_delete() method to release all resources.
 224          */
 225         cfs_list_for_each_entry_reverse(scan, layers, lo_linkage) {
 226                 if (scan->lo_ops->loo_object_delete != NULL)
 227                         scan->lo_ops->loo_object_delete(env, scan);
 228         }
 229
 230         /*
 231          * Then, splice object layers into stand-alone list, and call
 232          * ->loo_object_free() on all layers to free memory. Splice is
 233          * necessary, because lu_object_header is freed together with the
 234          * top-level slice.
 235          */
 236         CFS_INIT_LIST_HEAD(&splice);
 237         cfs_list_splice_init(layers, &splice);
 238         while (!cfs_list_empty(&splice)) {
 239                 /*
 240                  * Free layers in bottom-to-top order, so that object header
 241                  * lives as long as possible and ->loo_object_free() methods
 242                  * can look at its contents.
 243                  */
 244                 o = container_of0(splice.prev, struct lu_object, lo_linkage);
 245                 cfs_list_del_init(&o->lo_linkage);
 246                 LASSERT(o->lo_ops->loo_object_free != NULL);
 247                 o->lo_ops->loo_object_free(env, o);
 248         }
 249
 250         if (cfs_waitq_active(&bkt->lsb_marche_funebre))
 251                 cfs_waitq_broadcast(&bkt->lsb_marche_funebre);
 252 }
 253
 254 /**
 255  * Free \a nr objects from the cold end of the site LRU list.
 256  */
 257 int lu_site_purge(const struct lu_env *env, struct lu_site *s, int nr)
 258 {
 259         struct lu_object_header *h;
 260         struct lu_object_header *temp;
 261         struct lu_site_bkt_data *bkt;
 262         cfs_hash_bd_t            bd;
 263         cfs_hash_bd_t            bd2;
 264         cfs_list_t               dispose;
 265         int                      did_sth;
 266         int                      start;
 267         int                      count;
 268         int                      bnr;
 269         int                      i;
 270
 271         CFS_INIT_LIST_HEAD(&dispose);
 272         /*
 273          * Under LRU list lock, scan LRU list and move unreferenced objects to
 274          * the dispose list, removing them from LRU and hash table.
 275          */
 276         start = s->ls_purge_start;
 277         bnr = (nr == ~0) ? -1 : nr / CFS_HASH_NBKT(s->ls_obj_hash) + 1;
 278  again:
 279         did_sth = 0;
 280         cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
 281                 if (i < start)
 282                         continue;
 283                 count = bnr;
 284                 cfs_hash_bd_lock(s->ls_obj_hash, &bd, 1);
 285                 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
 286
 287                 cfs_list_for_each_entry_safe(h, temp, &bkt->lsb_lru, loh_lru) {
 288                         LASSERT(cfs_atomic_read(&h->loh_ref) == 0);
 289
 290                         cfs_hash_bd_get(s->ls_obj_hash, &h->loh_fid, &bd2);
 291                         LASSERT(bd.bd_bucket == bd2.bd_bucket);
 292
 293                         cfs_hash_bd_del_locked(s->ls_obj_hash,
 294                                                &bd2, &h->loh_hash);
 295                         cfs_list_move(&h->loh_lru, &dispose);
 296                         if (did_sth == 0)
 297                                 did_sth = 1;
 298
 299                         if (nr != ~0 && --nr == 0)
 300                                 break;
 301
 302                         if (count > 0 && --count == 0)
 303                                 break;
 304
 305                 }
 306                 cfs_hash_bd_unlock(s->ls_obj_hash, &bd, 1);
 307                 cfs_cond_resched();
 308                 /*
 309                  * Free everything on the dispose list. This is safe against
 310                  * races due to the reasons described in lu_object_put().
 311                  */
 312                 while (!cfs_list_empty(&dispose)) {
 313                         h = container_of0(dispose.next,
 314                                           struct lu_object_header, loh_lru);
 315                         cfs_list_del_init(&h->loh_lru);
 316                         lu_object_free(env, lu_object_top(h));
 317                         lprocfs_counter_incr(s->ls_stats, LU_SS_LRU_PURGED);
 318                 }
 319
 320                 if (nr == 0)
 321                         break;
 322         }
 323
 324         if (nr != 0 && did_sth && start != 0) {
 325                 start = 0; /* restart from the first bucket */
 326                 goto again;
 327         }
 328         /* race on s->ls_purge_start, but nobody cares */
 329         s->ls_purge_start = i % CFS_HASH_NBKT(s->ls_obj_hash);
 330
 331         return nr;
 332 }
 333 EXPORT_SYMBOL(lu_site_purge);
 334
 335 /*
 336  * Object printing.
 337  *
 338  * Code below has to jump through certain loops to output object description
 339  * into libcfs_debug_msg-based log. The problem is that lu_object_print()
 340  * composes object description from strings that are parts of _lines_ of
 341  * output (i.e., strings that are not terminated by newline). This doesn't fit
 342  * very well into libcfs_debug_msg() interface that assumes that each message
 343  * supplied to it is a self-contained output line.
 344  *
 345  * To work around this, strings are collected in a temporary buffer
 346  * (implemented as a value of lu_cdebug_key key), until terminating newline
 347  * character is detected.
 348  *
 349  */
 350
 351 enum {
 352         /**
 353          * Maximal line size.
 354          *
 355          * XXX overflow is not handled correctly.
 356          */
 357         LU_CDEBUG_LINE = 512
 358 };
 359
 360 struct lu_cdebug_data {
 361         /**
 362          * Temporary buffer.
 363          */
 364         char lck_area[LU_CDEBUG_LINE];
 365 };
 366
 367 /* context key constructor/destructor: lu_global_key_init, lu_global_key_fini */
 368 LU_KEY_INIT_FINI(lu_global, struct lu_cdebug_data);
 369
 370 /**
 371  * Key, holding temporary buffer. This key is registered very early by
 372  * lu_global_init().
 373  */
 374 struct lu_context_key lu_global_key = {
 375         .lct_tags = LCT_MD_THREAD|LCT_DT_THREAD|LCT_CL_THREAD,
 376         .lct_init = lu_global_key_init,
 377         .lct_fini = lu_global_key_fini
 378 };
 379
 380 /**
 381  * Printer function emitting messages through libcfs_debug_msg().
 382  */
 383 int lu_cdebug_printer(const struct lu_env *env,
 384                       void *cookie, const char *format, ...)
 385 {
 386         struct libcfs_debug_msg_data *msgdata = cookie;
 387         struct lu_cdebug_data        *key;
 388         int used;
 389         int complete;
 390         va_list args;
 391
 392         va_start(args, format);
 393
 394         key = lu_context_key_get(&env->le_ctx, &lu_global_key);
 395         LASSERT(key != NULL);
 396
 397         used = strlen(key->lck_area);
 398         complete = format[strlen(format) - 1] == '\n';
 399         /*
 400          * Append new chunk to the buffer.
 401          */
 402         vsnprintf(key->lck_area + used,
 403                   ARRAY_SIZE(key->lck_area) - used, format, args);
 404         if (complete) {
 405                 if (cfs_cdebug_show(msgdata->msg_mask, msgdata->msg_subsys))
 406                         libcfs_debug_msg(msgdata, "%s", key->lck_area);
 407                 key->lck_area[0] = 0;
 408         }
 409         va_end(args);
 410         return 0;
 411 }
 412 EXPORT_SYMBOL(lu_cdebug_printer);
 413
 414 /**
 415  * Print object header.
 416  */
 417 void lu_object_header_print(const struct lu_env *env, void *cookie,
 418                             lu_printer_t printer,
 419                             const struct lu_object_header *hdr)
 420 {
 421         (*printer)(env, cookie, "header@%p[%#lx, %d, "DFID"%s%s%s]",
 422                    hdr, hdr->loh_flags, cfs_atomic_read(&hdr->loh_ref),
 423                    PFID(&hdr->loh_fid),
 424                    cfs_hlist_unhashed(&hdr->loh_hash) ? "" : " hash",
 425                    cfs_list_empty((cfs_list_t *)&hdr->loh_lru) ? \
 426                    "" : " lru",
 427                    hdr->loh_attr & LOHA_EXISTS ? " exist":"");
 428 }
 429 EXPORT_SYMBOL(lu_object_header_print);
 430
 431 /**
 432  * Print human readable representation of the \a o to the \a printer.
 433  */
 434 void lu_object_print(const struct lu_env *env, void *cookie,
 435                      lu_printer_t printer, const struct lu_object *o)
 436 {
 437         static const char ruler[] = "........................................";
 438         struct lu_object_header *top;
 439         int depth;
 440
 441         top = o->lo_header;
 442         lu_object_header_print(env, cookie, printer, top);
 443         (*printer)(env, cookie, "{ \n");
 444         cfs_list_for_each_entry(o, &top->loh_layers, lo_linkage) {
 445                 depth = o->lo_depth + 4;
 446
 447                 /*
 448                  * print `.' \a depth times followed by type name and address
 449                  */
 450                 (*printer)(env, cookie, "%*.*s%s@%p", depth, depth, ruler,
 451                            o->lo_dev->ld_type->ldt_name, o);
 452                 if (o->lo_ops->loo_object_print != NULL)
 453                         o->lo_ops->loo_object_print(env, cookie, printer, o);
 454                 (*printer)(env, cookie, "\n");
 455         }
 456         (*printer)(env, cookie, "} header@%p\n", top);
 457 }
 458 EXPORT_SYMBOL(lu_object_print);
 459
 460 /**
 461  * Check object consistency.
 462  */
 463 int lu_object_invariant(const struct lu_object *o)
 464 {
 465         struct lu_object_header *top;
 466
 467         top = o->lo_header;
 468         cfs_list_for_each_entry(o, &top->loh_layers, lo_linkage) {
 469                 if (o->lo_ops->loo_object_invariant != NULL &&
 470                     !o->lo_ops->loo_object_invariant(o))
 471                         return 0;
 472         }
 473         return 1;
 474 }
 475 EXPORT_SYMBOL(lu_object_invariant);
 476
 477 static struct lu_object *htable_lookup(struct lu_site *s,
 478                                        cfs_hash_bd_t *bd,
 479                                        const struct lu_fid *f,
 480                                        cfs_waitlink_t *waiter,
 481                                        __u64 *version)
 482 {
 483         struct lu_site_bkt_data *bkt;
 484         struct lu_object_header *h;
 485         cfs_hlist_node_t        *hnode;
 486         __u64  ver = cfs_hash_bd_version_get(bd);
 487
 488         if (*version == ver)
 489                 return NULL;
 490
 491         *version = ver;
 492         bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, bd);
 493         /* cfs_hash_bd_lookup_intent is a somehow "internal" function
 494          * of cfs_hash, but we don't want refcount on object right now */
 495         hnode = cfs_hash_bd_lookup_locked(s->ls_obj_hash, bd, (void *)f);
 496         if (hnode == NULL) {
 497                 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_MISS);
 498                 return NULL;
 499         }
 500
 501         h = container_of0(hnode, struct lu_object_header, loh_hash);
 502         if (likely(!lu_object_is_dying(h))) {
 503                 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_HIT);
 504                 cfs_list_del_init(&h->loh_lru);
 505                 return lu_object_top(h);
 506         }
 507
 508         /*
 509          * Lookup found an object being destroyed this object cannot be
 510          * returned (to assure that references to dying objects are eventually
 511          * drained), and moreover, lookup has to wait until object is freed.
 512          */
 513         cfs_atomic_dec(&h->loh_ref);
 514
 515         cfs_waitlink_init(waiter);
 516         cfs_waitq_add(&bkt->lsb_marche_funebre, waiter);
 517         cfs_set_current_state(CFS_TASK_UNINT);
 518         lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_DEATH_RACE);
 519         return ERR_PTR(-EAGAIN);
 520 }
 521
 522 /**
 523  * Search cache for an object with the fid \a f. If such object is found,
 524  * return it. Otherwise, create new object, insert it into cache and return
 525  * it. In any case, additional reference is acquired on the returned object.
 526  */
 527 struct lu_object *lu_object_find(const struct lu_env *env,
 528                                  struct lu_device *dev, const struct lu_fid *f,
 529                                  const struct lu_object_conf *conf)
 530 {
 531         return lu_object_find_at(env, dev->ld_site->ls_top_dev, f, conf);
 532 }
 533 EXPORT_SYMBOL(lu_object_find);
 534
 535 static struct lu_object *lu_object_new(const struct lu_env *env,
 536                                        struct lu_device *dev,
 537                                        const struct lu_fid *f,
 538                                        const struct lu_object_conf *conf)
 539 {
 540         struct lu_object        *o;
 541         cfs_hash_t              *hs;
 542         cfs_hash_bd_t            bd;
 543         struct lu_site_bkt_data *bkt;
 544
 545         o = lu_object_alloc(env, dev, f, conf);
 546         if (unlikely(IS_ERR(o)))
 547                 return o;
 548
 549         hs = dev->ld_site->ls_obj_hash;
 550         cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
 551         bkt = cfs_hash_bd_extra_get(hs, &bd);
 552         cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
 553         bkt->lsb_busy++;
 554         cfs_hash_bd_unlock(hs, &bd, 1);
 555         return o;
 556 }
 557
 558 /**
 559  * Core logic of lu_object_find*() functions.
 560  */
 561 static struct lu_object *lu_object_find_try(const struct lu_env *env,
 562                                             struct lu_device *dev,
 563                                             const struct lu_fid *f,
 564                                             const struct lu_object_conf *conf,
 565                                             cfs_waitlink_t *waiter)
 566 {
 567         struct lu_object      *o;
 568         struct lu_object      *shadow;
 569         struct lu_site        *s;
 570         cfs_hash_t            *hs;
 571         cfs_hash_bd_t          bd;
 572         __u64                  version = 0;
 573
 574         /*
 575          * This uses standard index maintenance protocol:
 576          *
 577          *     - search index under lock, and return object if found;
 578          *     - otherwise, unlock index, allocate new object;
 579          *     - lock index and search again;
 580          *     - if nothing is found (usual case), insert newly created
 581          *       object into index;
 582          *     - otherwise (race: other thread inserted object), free
 583          *       object just allocated.
 584          *     - unlock index;
 585          *     - return object.
 586          *
 587          * For "LOC_F_NEW" case, we are sure the object is new established.
 588          * It is unnecessary to perform lookup-alloc-lookup-insert, instead,
 589          * just alloc and insert directly.
 590          *
 591          * If dying object is found during index search, add @waiter to the
 592          * site wait-queue and return ERR_PTR(-EAGAIN).
 593          */
 594         if (conf != NULL && conf->loc_flags & LOC_F_NEW)
 595                 return lu_object_new(env, dev, f, conf);
 596
 597         s  = dev->ld_site;
 598         hs = s->ls_obj_hash;
 599         cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
 600         o = htable_lookup(s, &bd, f, waiter, &version);
 601         cfs_hash_bd_unlock(hs, &bd, 1);
 602         if (o != NULL)
 603                 return o;
 604
 605         /*
 606          * Allocate new object. This may result in rather complicated
 607          * operations, including fld queries, inode loading, etc.
 608          */
 609         o = lu_object_alloc(env, dev, f, conf);
 610         if (unlikely(IS_ERR(o)))
 611                 return o;
 612
 613         LASSERT(lu_fid_eq(lu_object_fid(o), f));
 614
 615         cfs_hash_bd_lock(hs, &bd, 1);
 616
 617         shadow = htable_lookup(s, &bd, f, waiter, &version);
 618         if (likely(shadow == NULL)) {
 619                 struct lu_site_bkt_data *bkt;
 620
 621                 bkt = cfs_hash_bd_extra_get(hs, &bd);
 622                 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
 623                 bkt->lsb_busy++;
 624                 cfs_hash_bd_unlock(hs, &bd, 1);
 625                 return o;
 626         }
 627
 628         lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_RACE);
 629         cfs_hash_bd_unlock(hs, &bd, 1);
 630         lu_object_free(env, o);
 631         return shadow;
 632 }
 633
 634 /**
 635  * Much like lu_object_find(), but top level device of object is specifically
 636  * \a dev rather than top level device of the site. This interface allows
 637  * objects of different "stacking" to be created within the same site.
 638  */
 639 struct lu_object *lu_object_find_at(const struct lu_env *env,
 640                                     struct lu_device *dev,
 641                                     const struct lu_fid *f,
 642                                     const struct lu_object_conf *conf)
 643 {
 644         struct lu_site_bkt_data *bkt;
 645         struct lu_object        *obj;
 646         cfs_waitlink_t           wait;
 647
 648         while (1) {
 649                 obj = lu_object_find_try(env, dev, f, conf, &wait);
 650                 if (obj != ERR_PTR(-EAGAIN))
 651                         return obj;
 652                 /*
 653                  * lu_object_find_try() already added waiter into the
 654                  * wait queue.
 655                  */
 656                 cfs_waitq_wait(&wait, CFS_TASK_UNINT);
 657                 bkt = lu_site_bkt_from_fid(dev->ld_site, (void *)f);
 658                 cfs_waitq_del(&bkt->lsb_marche_funebre, &wait);
 659         }
 660 }
 661 EXPORT_SYMBOL(lu_object_find_at);
 662
 663 /**
 664  * Find object with given fid, and return its slice belonging to given device.
 665  */
 666 struct lu_object *lu_object_find_slice(const struct lu_env *env,
 667                                        struct lu_device *dev,
 668                                        const struct lu_fid *f,
 669                                        const struct lu_object_conf *conf)
 670 {
 671         struct lu_object *top;
 672         struct lu_object *obj;
 673
 674         top = lu_object_find(env, dev, f, conf);
 675         if (!IS_ERR(top)) {
 676                 obj = lu_object_locate(top->lo_header, dev->ld_type);
 677                 if (obj == NULL)
 678                         lu_object_put(env, top);
 679         } else
 680                 obj = top;
 681         return obj;
 682 }
 683 EXPORT_SYMBOL(lu_object_find_slice);
 684
 685 /**
 686  * Global list of all device types.
 687  */
 688 static CFS_LIST_HEAD(lu_device_types);
 689
 690 int lu_device_type_init(struct lu_device_type *ldt)
 691 {
 692         int result;
 693
 694         CFS_INIT_LIST_HEAD(&ldt->ldt_linkage);
 695         result = ldt->ldt_ops->ldto_init(ldt);
 696         if (result == 0)
 697                 cfs_list_add(&ldt->ldt_linkage, &lu_device_types);
 698         return result;
 699 }
 700 EXPORT_SYMBOL(lu_device_type_init);
 701
 702 void lu_device_type_fini(struct lu_device_type *ldt)
 703 {
 704         cfs_list_del_init(&ldt->ldt_linkage);
 705         ldt->ldt_ops->ldto_fini(ldt);
 706 }
 707 EXPORT_SYMBOL(lu_device_type_fini);
 708
 709 void lu_types_stop(void)
 710 {
 711         struct lu_device_type *ldt;
 712
 713         cfs_list_for_each_entry(ldt, &lu_device_types, ldt_linkage) {
 714                 if (ldt->ldt_device_nr == 0)
 715                         ldt->ldt_ops->ldto_stop(ldt);
 716         }
 717 }
 718 EXPORT_SYMBOL(lu_types_stop);
 719
 720 /**
 721  * Global list of all sites on this node
 722  */
 723 static CFS_LIST_HEAD(lu_sites);
 724 static CFS_DEFINE_MUTEX(lu_sites_guard);
 725
 726 /**
 727  * Global environment used by site shrinker.
 728  */
 729 static struct lu_env lu_shrink_env;
 730
 731 struct lu_site_print_arg {
 732         struct lu_env   *lsp_env;
 733         void            *lsp_cookie;
 734         lu_printer_t     lsp_printer;
 735 };
 736
 737 static int
 738 lu_site_obj_print(cfs_hash_t *hs, cfs_hash_bd_t *bd,
 739                   cfs_hlist_node_t *hnode, void *data)
 740 {
 741         struct lu_site_print_arg *arg = (struct lu_site_print_arg *)data;
 742         struct lu_object_header  *h;
 743
 744         h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
 745         if (!cfs_list_empty(&h->loh_layers)) {
 746                 const struct lu_object *o;
 747
 748                 o = lu_object_top(h);
 749                 lu_object_print(arg->lsp_env, arg->lsp_cookie,
 750                                 arg->lsp_printer, o);
 751         } else {
 752                 lu_object_header_print(arg->lsp_env, arg->lsp_cookie,
 753                                        arg->lsp_printer, h);
 754         }
 755         return 0;
 756 }
 757
 758 /**
 759  * Print all objects in \a s.
 760  */
 761 void lu_site_print(const struct lu_env *env, struct lu_site *s, void *cookie,
 762                    lu_printer_t printer)
 763 {
 764         struct lu_site_print_arg arg = {
 765                 .lsp_env     = (struct lu_env *)env,
 766                 .lsp_cookie  = cookie,
 767                 .lsp_printer = printer,
 768         };
 769
 770         cfs_hash_for_each(s->ls_obj_hash, lu_site_obj_print, &arg);
 771 }
 772 EXPORT_SYMBOL(lu_site_print);
 773
 774 enum {
 775         LU_CACHE_PERCENT_MAX     = 50,
 776         LU_CACHE_PERCENT_DEFAULT = 20
 777 };
 778
 779 static unsigned int lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
 780 CFS_MODULE_PARM(lu_cache_percent, "i", int, 0644,
 781                 "Percentage of memory to be used as lu_object cache");
 782
 783 /**
 784  * Return desired hash table order.
 785  */
 786 static int lu_htable_order(void)
 787 {
 788         unsigned long cache_size;
 789         int bits;
 790
 791         /*
 792          * Calculate hash table size, assuming that we want reasonable
 793          * performance when 20% of total memory is occupied by cache of
 794          * lu_objects.
 795          *
 796          * Size of lu_object is (arbitrary) taken as 1K (together with inode).
 797          */
 798         cache_size = cfs_num_physpages;
 799
 800 #if BITS_PER_LONG == 32
 801         /* limit hashtable size for lowmem systems to low RAM */
 802         if (cache_size > 1 << (30 - CFS_PAGE_SHIFT))
 803                 cache_size = 1 << (30 - CFS_PAGE_SHIFT) * 3 / 4;
 804 #endif
 805
 806         /* clear off unreasonable cache setting. */
 807         if (lu_cache_percent == 0 || lu_cache_percent > LU_CACHE_PERCENT_MAX) {
 808                 CWARN("obdclass: invalid lu_cache_percent: %u, it must be in"
 809                       " the range of (0, %u]. Will use default value: %u.\n",
 810                       lu_cache_percent, LU_CACHE_PERCENT_MAX,
 811                       LU_CACHE_PERCENT_DEFAULT);
 812
 813                 lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
 814         }
 815         cache_size = cache_size / 100 * lu_cache_percent *
 816                 (CFS_PAGE_SIZE / 1024);
 817
 818         for (bits = 1; (1 << bits) < cache_size; ++bits) {
 819                 ;
 820         }
 821         return bits;
 822 }
 823
 824 static unsigned lu_obj_hop_hash(cfs_hash_t *hs,
 825                                 const void *key, unsigned mask)
 826 {
 827         struct lu_fid  *fid = (struct lu_fid *)key;
 828         __u32           hash;
 829
 830         hash = fid_flatten32(fid);
 831         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
 832         hash = cfs_hash_long(hash, hs->hs_bkt_bits);
 833
 834         /* give me another random factor */
 835         hash -= cfs_hash_long((unsigned long)hs, fid_oid(fid) % 11 + 3);
 836
 837         hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
 838         hash |= (fid_seq(fid) + fid_oid(fid)) & (CFS_HASH_NBKT(hs) - 1);
 839
 840         return hash & mask;
 841 }
 842
 843 static void *lu_obj_hop_object(cfs_hlist_node_t *hnode)
 844 {
 845         return cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
 846 }
 847
 848 static void *lu_obj_hop_key(cfs_hlist_node_t *hnode)
 849 {
 850         struct lu_object_header *h;
 851
 852         h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
 853         return &h->loh_fid;
 854 }
 855
 856 static int lu_obj_hop_keycmp(const void *key, cfs_hlist_node_t *hnode)
 857 {
 858         struct lu_object_header *h;
 859
 860         h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
 861         return lu_fid_eq(&h->loh_fid, (struct lu_fid *)key);
 862 }
 863
 864 static void lu_obj_hop_get(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 865 {
 866         struct lu_object_header *h;
 867
 868         h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
 869         if (cfs_atomic_add_return(1, &h->loh_ref) == 1) {
 870                 struct lu_site_bkt_data *bkt;
 871                 cfs_hash_bd_t            bd;
 872
 873                 cfs_hash_bd_get(hs, &h->loh_fid, &bd);
 874                 bkt = cfs_hash_bd_extra_get(hs, &bd);
 875                 bkt->lsb_busy++;
 876         }
 877 }
 878
 879 static void lu_obj_hop_put_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 880 {
 881         LBUG(); /* we should never called it */
 882 }
 883
 884 cfs_hash_ops_t lu_site_hash_ops = {
 885         .hs_hash        = lu_obj_hop_hash,
 886         .hs_key         = lu_obj_hop_key,
 887         .hs_keycmp      = lu_obj_hop_keycmp,
 888         .hs_object      = lu_obj_hop_object,
 889         .hs_get         = lu_obj_hop_get,
 890         .hs_put_locked  = lu_obj_hop_put_locked,
 891 };
 892
 893 /**
 894  * Initialize site \a s, with \a d as the top level device.
 895  */
 896 #define LU_SITE_BITS_MIN    12
 897 #define LU_SITE_BITS_MAX    24
 898 /**
 899  * total 256 buckets, we don't want too many buckets because:
 900  * - consume too much memory
 901  * - avoid unbalanced LRU list
 902  */
 903 #define LU_SITE_BKT_BITS    8
 904
 905 int lu_site_init(struct lu_site *s, struct lu_device *top)
 906 {
 907         struct lu_site_bkt_data *bkt;
 908         cfs_hash_bd_t bd;
 909         char name[16];
 910         int bits;
 911         int i;
 912         ENTRY;
 913
 914         memset(s, 0, sizeof *s);
 915         bits = lu_htable_order();
 916         snprintf(name, 16, "lu_site_%s", top->ld_type->ldt_name);
 917         for (bits = min(max(LU_SITE_BITS_MIN, bits), LU_SITE_BITS_MAX);
 918              bits >= LU_SITE_BITS_MIN; bits--) {
 919                 s->ls_obj_hash = cfs_hash_create(name, bits, bits,
 920                                                  bits - LU_SITE_BKT_BITS,
 921                                                  sizeof(*bkt), 0, 0,
 922                                                  &lu_site_hash_ops,
 923                                                  CFS_HASH_SPIN_BKTLOCK |
 924                                                  CFS_HASH_NO_ITEMREF |
 925                                                  CFS_HASH_DEPTH |
 926                                                  CFS_HASH_ASSERT_EMPTY);
 927                 if (s->ls_obj_hash != NULL)
 928                         break;
 929         }
 930
 931         if (s->ls_obj_hash == NULL) {
 932                 CERROR("failed to create lu_site hash with bits: %d\n", bits);
 933                 return -ENOMEM;
 934         }
 935
 936         cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
 937                 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
 938                 CFS_INIT_LIST_HEAD(&bkt->lsb_lru);
 939                 cfs_waitq_init(&bkt->lsb_marche_funebre);
 940         }
 941
 942         s->ls_stats = lprocfs_alloc_stats(LU_SS_LAST_STAT, 0);
 943         if (s->ls_stats == NULL) {
 944                 cfs_hash_putref(s->ls_obj_hash);
 945                 s->ls_obj_hash = NULL;
 946                 return -ENOMEM;
 947         }
 948
 949         lprocfs_counter_init(s->ls_stats, LU_SS_CREATED,
 950                              0, "created", "created");
 951         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_HIT,
 952                              0, "cache_hit", "cache_hit");
 953         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_MISS,
 954                              0, "cache_miss", "cache_miss");
 955         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_RACE,
 956                              0, "cache_race", "cache_race");
 957         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_DEATH_RACE,
 958                              0, "cache_death_race", "cache_death_race");
 959         lprocfs_counter_init(s->ls_stats, LU_SS_LRU_PURGED,
 960                              0, "lru_purged", "lru_purged");
 961
 962         CFS_INIT_LIST_HEAD(&s->ls_linkage);
 963         s->ls_top_dev = top;
 964         top->ld_site = s;
 965         lu_device_get(top);
 966         lu_ref_add(&top->ld_reference, "site-top", s);
 967
 968         CFS_INIT_LIST_HEAD(&s->ls_ld_linkage);
 969         cfs_spin_lock_init(&s->ls_ld_lock);
 970
 971         cfs_spin_lock(&s->ls_ld_lock);
 972         cfs_list_add(&top->ld_linkage, &s->ls_ld_linkage);
 973         cfs_spin_unlock(&s->ls_ld_lock);
 974
 975         RETURN(0);
 976 }
 977 EXPORT_SYMBOL(lu_site_init);
 978
 979 /**
 980  * Finalize \a s and release its resources.
 981  */
 982 void lu_site_fini(struct lu_site *s)
 983 {
 984         cfs_mutex_lock(&lu_sites_guard);
 985         cfs_list_del_init(&s->ls_linkage);
 986         cfs_mutex_unlock(&lu_sites_guard);
 987
 988         if (s->ls_obj_hash != NULL) {
 989                 cfs_hash_putref(s->ls_obj_hash);
 990                 s->ls_obj_hash = NULL;
 991         }
 992
 993         if (s->ls_top_dev != NULL) {
 994                 s->ls_top_dev->ld_site = NULL;
 995                 lu_ref_del(&s->ls_top_dev->ld_reference, "site-top", s);
 996                 lu_device_put(s->ls_top_dev);
 997                 s->ls_top_dev = NULL;
 998         }
 999
1000         if (s->ls_stats != NULL)
1001                 lprocfs_free_stats(&s->ls_stats);
1002 }
1003 EXPORT_SYMBOL(lu_site_fini);
1004
1005 /**
1006  * Called when initialization of stack for this site is completed.
1007  */
1008 int lu_site_init_finish(struct lu_site *s)
1009 {
1010         int result;
1011         cfs_mutex_lock(&lu_sites_guard);
1012         result = lu_context_refill(&lu_shrink_env.le_ctx);
1013         if (result == 0)
1014                 cfs_list_add(&s->ls_linkage, &lu_sites);
1015         cfs_mutex_unlock(&lu_sites_guard);
1016         return result;
1017 }
1018 EXPORT_SYMBOL(lu_site_init_finish);
1019
1020 /**
1021  * Acquire additional reference on device \a d
1022  */
1023 void lu_device_get(struct lu_device *d)
1024 {
1025         cfs_atomic_inc(&d->ld_ref);
1026 }
1027 EXPORT_SYMBOL(lu_device_get);
1028
1029 /**
1030  * Release reference on device \a d.
1031  */
1032 void lu_device_put(struct lu_device *d)
1033 {
1034         LASSERT(cfs_atomic_read(&d->ld_ref) > 0);
1035         cfs_atomic_dec(&d->ld_ref);
1036 }
1037 EXPORT_SYMBOL(lu_device_put);
1038
1039 /**
1040  * Initialize device \a d of type \a t.
1041  */
1042 int lu_device_init(struct lu_device *d, struct lu_device_type *t)
1043 {
1044         if (t->ldt_device_nr++ == 0 && t->ldt_ops->ldto_start != NULL)
1045                 t->ldt_ops->ldto_start(t);
1046         memset(d, 0, sizeof *d);
1047         cfs_atomic_set(&d->ld_ref, 0);
1048         d->ld_type = t;
1049         lu_ref_init(&d->ld_reference);
1050         CFS_INIT_LIST_HEAD(&d->ld_linkage);
1051         return 0;
1052 }
1053 EXPORT_SYMBOL(lu_device_init);
1054
1055 /**
1056  * Finalize device \a d.
1057  */
1058 void lu_device_fini(struct lu_device *d)
1059 {
1060         struct lu_device_type *t;
1061
1062         t = d->ld_type;
1063         if (d->ld_obd != NULL) {
1064                 d->ld_obd->obd_lu_dev = NULL;
1065                 d->ld_obd = NULL;
1066         }
1067
1068         lu_ref_fini(&d->ld_reference);
1069         LASSERTF(cfs_atomic_read(&d->ld_ref) == 0,
1070                  "Refcount is %u\n", cfs_atomic_read(&d->ld_ref));
1071         LASSERT(t->ldt_device_nr > 0);
1072         if (--t->ldt_device_nr == 0 && t->ldt_ops->ldto_stop != NULL)
1073                 t->ldt_ops->ldto_stop(t);
1074 }
1075 EXPORT_SYMBOL(lu_device_fini);
1076
1077 /**
1078  * Initialize object \a o that is part of compound object \a h and was created
1079  * by device \a d.
1080  */
1081 int lu_object_init(struct lu_object *o,
1082                    struct lu_object_header *h, struct lu_device *d)
1083 {
1084         memset(o, 0, sizeof *o);
1085         o->lo_header = h;
1086         o->lo_dev    = d;
1087         lu_device_get(d);
1088         o->lo_dev_ref = lu_ref_add(&d->ld_reference, "lu_object", o);
1089         CFS_INIT_LIST_HEAD(&o->lo_linkage);
1090         return 0;
1091 }
1092 EXPORT_SYMBOL(lu_object_init);
1093
1094 /**
1095  * Finalize object and release its resources.
1096  */
1097 void lu_object_fini(struct lu_object *o)
1098 {
1099         struct lu_device *dev = o->lo_dev;
1100
1101         LASSERT(cfs_list_empty(&o->lo_linkage));
1102
1103         if (dev != NULL) {
1104                 lu_ref_del_at(&dev->ld_reference,
1105                               o->lo_dev_ref , "lu_object", o);
1106                 lu_device_put(dev);
1107                 o->lo_dev = NULL;
1108         }
1109 }
1110 EXPORT_SYMBOL(lu_object_fini);
1111
1112 /**
1113  * Add object \a o as first layer of compound object \a h
1114  *
1115  * This is typically called by the ->ldo_object_alloc() method of top-level
1116  * device.
1117  */
1118 void lu_object_add_top(struct lu_object_header *h, struct lu_object *o)
1119 {
1120         cfs_list_move(&o->lo_linkage, &h->loh_layers);
1121 }
1122 EXPORT_SYMBOL(lu_object_add_top);
1123
1124 /**
1125  * Add object \a o as a layer of compound object, going after \a before.
1126  *
1127  * This is typically called by the ->ldo_object_alloc() method of \a
1128  * before->lo_dev.
1129  */
1130 void lu_object_add(struct lu_object *before, struct lu_object *o)
1131 {
1132         cfs_list_move(&o->lo_linkage, &before->lo_linkage);
1133 }
1134 EXPORT_SYMBOL(lu_object_add);
1135
1136 /**
1137  * Initialize compound object.
1138  */
1139 int lu_object_header_init(struct lu_object_header *h)
1140 {
1141         memset(h, 0, sizeof *h);
1142         cfs_atomic_set(&h->loh_ref, 1);
1143         CFS_INIT_HLIST_NODE(&h->loh_hash);
1144         CFS_INIT_LIST_HEAD(&h->loh_lru);
1145         CFS_INIT_LIST_HEAD(&h->loh_layers);
1146         lu_ref_init(&h->loh_reference);
1147         return 0;
1148 }
1149 EXPORT_SYMBOL(lu_object_header_init);
1150
1151 /**
1152  * Finalize compound object.
1153  */
1154 void lu_object_header_fini(struct lu_object_header *h)
1155 {
1156         LASSERT(cfs_list_empty(&h->loh_layers));
1157         LASSERT(cfs_list_empty(&h->loh_lru));
1158         LASSERT(cfs_hlist_unhashed(&h->loh_hash));
1159         lu_ref_fini(&h->loh_reference);
1160 }
1161 EXPORT_SYMBOL(lu_object_header_fini);
1162
1163 /**
1164  * Given a compound object, find its slice, corresponding to the device type
1165  * \a dtype.
1166  */
1167 struct lu_object *lu_object_locate(struct lu_object_header *h,
1168                                    const struct lu_device_type *dtype)
1169 {
1170         struct lu_object *o;
1171
1172         cfs_list_for_each_entry(o, &h->loh_layers, lo_linkage) {
1173                 if (o->lo_dev->ld_type == dtype)
1174                         return o;
1175         }
1176         return NULL;
1177 }
1178 EXPORT_SYMBOL(lu_object_locate);
1179
1180
1181
1182 /**
1183  * Finalize and free devices in the device stack.
1184  *
1185  * Finalize device stack by purging object cache, and calling
1186  * lu_device_type_operations::ldto_device_fini() and
1187  * lu_device_type_operations::ldto_device_free() on all devices in the stack.
1188  */
1189 void lu_stack_fini(const struct lu_env *env, struct lu_device *top)
1190 {
1191         struct lu_site   *site = top->ld_site;
1192         struct lu_device *scan;
1193         struct lu_device *next;
1194
1195         lu_site_purge(env, site, ~0);
1196         for (scan = top; scan != NULL; scan = next) {
1197                 next = scan->ld_type->ldt_ops->ldto_device_fini(env, scan);
1198                 lu_ref_del(&scan->ld_reference, "lu-stack", &lu_site_init);
1199                 lu_device_put(scan);
1200         }
1201
1202         /* purge again. */
1203         lu_site_purge(env, site, ~0);
1204
1205         if (!cfs_hash_is_empty(site->ls_obj_hash)) {
1206                 /*
1207                  * Uh-oh, objects still exist.
1208                  */
1209                 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
1210
1211                 lu_site_print(env, site, &msgdata, lu_cdebug_printer);
1212         }
1213
1214         for (scan = top; scan != NULL; scan = next) {
1215                 const struct lu_device_type *ldt = scan->ld_type;
1216                 struct obd_type             *type;
1217
1218                 next = ldt->ldt_ops->ldto_device_free(env, scan);
1219                 type = ldt->ldt_obd_type;
1220                 if (type != NULL) {
1221                         type->typ_refcnt--;
1222                         class_put_type(type);
1223                 }
1224         }
1225 }
1226 EXPORT_SYMBOL(lu_stack_fini);
1227
1228 enum {
1229         /**
1230          * Maximal number of tld slots.
1231          */
1232         LU_CONTEXT_KEY_NR = 32
1233 };
1234
1235 static struct lu_context_key *lu_keys[LU_CONTEXT_KEY_NR] = { NULL, };
1236
1237 static cfs_spinlock_t lu_keys_guard = CFS_SPIN_LOCK_UNLOCKED;
1238
1239 /**
1240  * Global counter incremented whenever key is registered, unregistered,
1241  * revived or quiesced. This is used to void unnecessary calls to
1242  * lu_context_refill(). No locking is provided, as initialization and shutdown
1243  * are supposed to be externally serialized.
1244  */
1245 static unsigned key_set_version = 0;
1246
1247 /**
1248  * Register new key.
1249  */
1250 int lu_context_key_register(struct lu_context_key *key)
1251 {
1252         int result;
1253         int i;
1254
1255         LASSERT(key->lct_init != NULL);
1256         LASSERT(key->lct_fini != NULL);
1257         LASSERT(key->lct_tags != 0);
1258         LASSERT(key->lct_owner != NULL);
1259
1260         result = -ENFILE;
1261         cfs_spin_lock(&lu_keys_guard);
1262         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1263                 if (lu_keys[i] == NULL) {
1264                         key->lct_index = i;
1265                         cfs_atomic_set(&key->lct_used, 1);
1266                         lu_keys[i] = key;
1267                         lu_ref_init(&key->lct_reference);
1268                         result = 0;
1269                         ++key_set_version;
1270                         break;
1271                 }
1272         }
1273         cfs_spin_unlock(&lu_keys_guard);
1274         return result;
1275 }
1276 EXPORT_SYMBOL(lu_context_key_register);
1277
1278 static void key_fini(struct lu_context *ctx, int index)
1279 {
1280         if (ctx->lc_value != NULL && ctx->lc_value[index] != NULL) {
1281                 struct lu_context_key *key;
1282
1283                 key = lu_keys[index];
1284                 LASSERT(key != NULL);
1285                 LASSERT(key->lct_fini != NULL);
1286                 LASSERT(cfs_atomic_read(&key->lct_used) > 1);
1287
1288                 key->lct_fini(ctx, key, ctx->lc_value[index]);
1289                 lu_ref_del(&key->lct_reference, "ctx", ctx);
1290                 cfs_atomic_dec(&key->lct_used);
1291                 LASSERT(key->lct_owner != NULL);
1292                 if (!(ctx->lc_tags & LCT_NOREF)) {
1293                         LASSERT(cfs_module_refcount(key->lct_owner) > 0);
1294                         cfs_module_put(key->lct_owner);
1295                 }
1296                 ctx->lc_value[index] = NULL;
1297         }
1298 }
1299
1300 /**
1301  * Deregister key.
1302  */
1303 void lu_context_key_degister(struct lu_context_key *key)
1304 {
1305         LASSERT(cfs_atomic_read(&key->lct_used) >= 1);
1306         LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1307
1308         lu_context_key_quiesce(key);
1309
1310         ++key_set_version;
1311         cfs_spin_lock(&lu_keys_guard);
1312         key_fini(&lu_shrink_env.le_ctx, key->lct_index);
1313         if (lu_keys[key->lct_index]) {
1314                 lu_keys[key->lct_index] = NULL;
1315                 lu_ref_fini(&key->lct_reference);
1316         }
1317         cfs_spin_unlock(&lu_keys_guard);
1318
1319         LASSERTF(cfs_atomic_read(&key->lct_used) == 1,
1320                  "key has instances: %d\n",
1321                  cfs_atomic_read(&key->lct_used));
1322 }
1323 EXPORT_SYMBOL(lu_context_key_degister);
1324
1325 /**
1326  * Register a number of keys. This has to be called after all keys have been
1327  * initialized by a call to LU_CONTEXT_KEY_INIT().
1328  */
1329 int lu_context_key_register_many(struct lu_context_key *k, ...)
1330 {
1331         struct lu_context_key *key = k;
1332         va_list args;
1333         int result;
1334
1335         va_start(args, k);
1336         do {
1337                 result = lu_context_key_register(key);
1338                 if (result)
1339                         break;
1340                 key = va_arg(args, struct lu_context_key *);
1341         } while (key != NULL);
1342         va_end(args);
1343
1344         if (result != 0) {
1345                 va_start(args, k);
1346                 while (k != key) {
1347                         lu_context_key_degister(k);
1348                         k = va_arg(args, struct lu_context_key *);
1349                 }
1350                 va_end(args);
1351         }
1352
1353         return result;
1354 }
1355 EXPORT_SYMBOL(lu_context_key_register_many);
1356
1357 /**
1358  * De-register a number of keys. This is a dual to
1359  * lu_context_key_register_many().
1360  */
1361 void lu_context_key_degister_many(struct lu_context_key *k, ...)
1362 {
1363         va_list args;
1364
1365         va_start(args, k);
1366         do {
1367                 lu_context_key_degister(k);
1368                 k = va_arg(args, struct lu_context_key*);
1369         } while (k != NULL);
1370         va_end(args);
1371 }
1372 EXPORT_SYMBOL(lu_context_key_degister_many);
1373
1374 /**
1375  * Revive a number of keys.
1376  */
1377 void lu_context_key_revive_many(struct lu_context_key *k, ...)
1378 {
1379         va_list args;
1380
1381         va_start(args, k);
1382         do {
1383                 lu_context_key_revive(k);
1384                 k = va_arg(args, struct lu_context_key*);
1385         } while (k != NULL);
1386         va_end(args);
1387 }
1388 EXPORT_SYMBOL(lu_context_key_revive_many);
1389
1390 /**
1391  * Quiescent a number of keys.
1392  */
1393 void lu_context_key_quiesce_many(struct lu_context_key *k, ...)
1394 {
1395         va_list args;
1396
1397         va_start(args, k);
1398         do {
1399                 lu_context_key_quiesce(k);
1400                 k = va_arg(args, struct lu_context_key*);
1401         } while (k != NULL);
1402         va_end(args);
1403 }
1404 EXPORT_SYMBOL(lu_context_key_quiesce_many);
1405
1406 /**
1407  * Return value associated with key \a key in context \a ctx.
1408  */
1409 void *lu_context_key_get(const struct lu_context *ctx,
1410                          const struct lu_context_key *key)
1411 {
1412         LINVRNT(ctx->lc_state == LCS_ENTERED);
1413         LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1414         LASSERT(lu_keys[key->lct_index] == key);
1415         return ctx->lc_value[key->lct_index];
1416 }
1417 EXPORT_SYMBOL(lu_context_key_get);
1418
1419 /**
1420  * List of remembered contexts. XXX document me.
1421  */
1422 static CFS_LIST_HEAD(lu_context_remembered);
1423
1424 /**
1425  * Destroy \a key in all remembered contexts. This is used to destroy key
1426  * values in "shared" contexts (like service threads), when a module owning
1427  * the key is about to be unloaded.
1428  */
1429 void lu_context_key_quiesce(struct lu_context_key *key)
1430 {
1431         struct lu_context *ctx;
1432         extern unsigned cl_env_cache_purge(unsigned nr);
1433
1434         if (!(key->lct_tags & LCT_QUIESCENT)) {
1435                 /*
1436                  * XXX layering violation.
1437                  */
1438                 cl_env_cache_purge(~0);
1439                 key->lct_tags |= LCT_QUIESCENT;
1440                 /*
1441                  * XXX memory barrier has to go here.
1442                  */
1443                 cfs_spin_lock(&lu_keys_guard);
1444                 cfs_list_for_each_entry(ctx, &lu_context_remembered,
1445                                         lc_remember)
1446                         key_fini(ctx, key->lct_index);
1447                 cfs_spin_unlock(&lu_keys_guard);
1448                 ++key_set_version;
1449         }
1450 }
1451 EXPORT_SYMBOL(lu_context_key_quiesce);
1452
1453 void lu_context_key_revive(struct lu_context_key *key)
1454 {
1455         key->lct_tags &= ~LCT_QUIESCENT;
1456         ++key_set_version;
1457 }
1458 EXPORT_SYMBOL(lu_context_key_revive);
1459
1460 static void keys_fini(struct lu_context *ctx)
1461 {
1462         int i;
1463
1464         cfs_spin_lock(&lu_keys_guard);
1465         if (ctx->lc_value != NULL) {
1466                 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i)
1467                         key_fini(ctx, i);
1468                 OBD_FREE(ctx->lc_value,
1469                          ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1470                 ctx->lc_value = NULL;
1471         }
1472         cfs_spin_unlock(&lu_keys_guard);
1473 }
1474
1475 static int keys_fill(struct lu_context *ctx)
1476 {
1477         int i;
1478
1479         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1480                 struct lu_context_key *key;
1481
1482                 key = lu_keys[i];
1483                 if (ctx->lc_value[i] == NULL && key != NULL &&
1484                     (key->lct_tags & ctx->lc_tags) &&
1485                     /*
1486                      * Don't create values for a LCT_QUIESCENT key, as this
1487                      * will pin module owning a key.
1488                      */
1489                     !(key->lct_tags & LCT_QUIESCENT)) {
1490                         void *value;
1491
1492                         LINVRNT(key->lct_init != NULL);
1493                         LINVRNT(key->lct_index == i);
1494
1495                         value = key->lct_init(ctx, key);
1496                         if (unlikely(IS_ERR(value)))
1497                                 return PTR_ERR(value);
1498
1499                         LASSERT(key->lct_owner != NULL);
1500                         if (!(ctx->lc_tags & LCT_NOREF))
1501                                 cfs_try_module_get(key->lct_owner);
1502                         lu_ref_add_atomic(&key->lct_reference, "ctx", ctx);
1503                         cfs_atomic_inc(&key->lct_used);
1504                         /*
1505                          * This is the only place in the code, where an
1506                          * element of ctx->lc_value[] array is set to non-NULL
1507                          * value.
1508                          */
1509                         ctx->lc_value[i] = value;
1510                         if (key->lct_exit != NULL)
1511                                 ctx->lc_tags |= LCT_HAS_EXIT;
1512                 }
1513                 ctx->lc_version = key_set_version;
1514         }
1515         return 0;
1516 }
1517
1518 static int keys_init(struct lu_context *ctx)
1519 {
1520         int result;
1521
1522         OBD_ALLOC(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1523         if (likely(ctx->lc_value != NULL))
1524                 result = keys_fill(ctx);
1525         else
1526                 result = -ENOMEM;
1527
1528         if (result != 0)
1529                 keys_fini(ctx);
1530         return result;
1531 }
1532
1533 /**
1534  * Initialize context data-structure. Create values for all keys.
1535  */
1536 int lu_context_init(struct lu_context *ctx, __u32 tags)
1537 {
1538         memset(ctx, 0, sizeof *ctx);
1539         ctx->lc_state = LCS_INITIALIZED;
1540         ctx->lc_tags = tags;
1541         if (tags & LCT_REMEMBER) {
1542                 cfs_spin_lock(&lu_keys_guard);
1543                 cfs_list_add(&ctx->lc_remember, &lu_context_remembered);
1544                 cfs_spin_unlock(&lu_keys_guard);
1545         } else
1546                 CFS_INIT_LIST_HEAD(&ctx->lc_remember);
1547         return keys_init(ctx);
1548 }
1549 EXPORT_SYMBOL(lu_context_init);
1550
1551 /**
1552  * Finalize context data-structure. Destroy key values.
1553  */
1554 void lu_context_fini(struct lu_context *ctx)
1555 {
1556         LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1557         ctx->lc_state = LCS_FINALIZED;
1558         keys_fini(ctx);
1559         cfs_spin_lock(&lu_keys_guard);
1560         cfs_list_del_init(&ctx->lc_remember);
1561         cfs_spin_unlock(&lu_keys_guard);
1562 }
1563 EXPORT_SYMBOL(lu_context_fini);
1564
1565 /**
1566  * Called before entering context.
1567  */
1568 void lu_context_enter(struct lu_context *ctx)
1569 {
1570         LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1571         ctx->lc_state = LCS_ENTERED;
1572 }
1573 EXPORT_SYMBOL(lu_context_enter);
1574
1575 /**
1576  * Called after exiting from \a ctx
1577  */
1578 void lu_context_exit(struct lu_context *ctx)
1579 {
1580         int i;
1581
1582         LINVRNT(ctx->lc_state == LCS_ENTERED);
1583         ctx->lc_state = LCS_LEFT;
1584         if (ctx->lc_tags & LCT_HAS_EXIT && ctx->lc_value != NULL) {
1585                 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1586                         if (ctx->lc_value[i] != NULL) {
1587                                 struct lu_context_key *key;
1588
1589                                 key = lu_keys[i];
1590                                 LASSERT(key != NULL);
1591                                 if (key->lct_exit != NULL)
1592                                         key->lct_exit(ctx,
1593                                                       key, ctx->lc_value[i]);
1594                         }
1595                 }
1596         }
1597 }
1598 EXPORT_SYMBOL(lu_context_exit);
1599
1600 /**
1601  * Allocate for context all missing keys that were registered after context
1602  * creation.
1603  */
1604 int lu_context_refill(struct lu_context *ctx)
1605 {
1606         LINVRNT(ctx->lc_value != NULL);
1607         return ctx->lc_version == key_set_version ? 0 : keys_fill(ctx);
1608 }
1609 EXPORT_SYMBOL(lu_context_refill);
1610
1611 /**
1612  * lu_ctx_tags/lu_ses_tags will be updated if there are new types of
1613  * obd being added. Currently, this is only used on client side, specifically
1614  * for echo device client, for other stack (like ptlrpc threads), context are
1615  * predefined when the lu_device type are registered, during the module probe
1616  * phase.
1617  */
1618 __u32 lu_context_tags_default = 0;
1619 __u32 lu_session_tags_default = 0;
1620
1621 void lu_context_tags_update(__u32 tags)
1622 {
1623         cfs_spin_lock(&lu_keys_guard);
1624         lu_context_tags_default |= tags;
1625         key_set_version ++;
1626         cfs_spin_unlock(&lu_keys_guard);
1627 }
1628 EXPORT_SYMBOL(lu_context_tags_update);
1629
1630 void lu_context_tags_clear(__u32 tags)
1631 {
1632         cfs_spin_lock(&lu_keys_guard);
1633         lu_context_tags_default &= ~tags;
1634         key_set_version ++;
1635         cfs_spin_unlock(&lu_keys_guard);
1636 }
1637 EXPORT_SYMBOL(lu_context_tags_clear);
1638
1639 void lu_session_tags_update(__u32 tags)
1640 {
1641         cfs_spin_lock(&lu_keys_guard);
1642         lu_session_tags_default |= tags;
1643         key_set_version ++;
1644         cfs_spin_unlock(&lu_keys_guard);
1645 }
1646 EXPORT_SYMBOL(lu_session_tags_update);
1647
1648 void lu_session_tags_clear(__u32 tags)
1649 {
1650         cfs_spin_lock(&lu_keys_guard);
1651         lu_session_tags_default &= ~tags;
1652         key_set_version ++;
1653         cfs_spin_unlock(&lu_keys_guard);
1654 }
1655 EXPORT_SYMBOL(lu_session_tags_clear);
1656
1657 int lu_env_init(struct lu_env *env, __u32 tags)
1658 {
1659         int result;
1660
1661         env->le_ses = NULL;
1662         result = lu_context_init(&env->le_ctx, tags);
1663         if (likely(result == 0))
1664                 lu_context_enter(&env->le_ctx);
1665         return result;
1666 }
1667 EXPORT_SYMBOL(lu_env_init);
1668
1669 void lu_env_fini(struct lu_env *env)
1670 {
1671         lu_context_exit(&env->le_ctx);
1672         lu_context_fini(&env->le_ctx);
1673         env->le_ses = NULL;
1674 }
1675 EXPORT_SYMBOL(lu_env_fini);
1676
1677 int lu_env_refill(struct lu_env *env)
1678 {
1679         int result;
1680
1681         result = lu_context_refill(&env->le_ctx);
1682         if (result == 0 && env->le_ses != NULL)
1683                 result = lu_context_refill(env->le_ses);
1684         return result;
1685 }
1686 EXPORT_SYMBOL(lu_env_refill);
1687
1688 /**
1689  * Currently, this API will only be used by echo client.
1690  * Because echo client and normal lustre client will share
1691  * same cl_env cache. So echo client needs to refresh
1692  * the env context after it get one from the cache, especially
1693  * when normal client and echo client co-exist in the same client.
1694  */
1695 int lu_env_refill_by_tags(struct lu_env *env, __u32 ctags,
1696                           __u32 stags)
1697 {
1698         int    result;
1699
1700         if ((env->le_ctx.lc_tags & ctags) != ctags) {
1701                 env->le_ctx.lc_version = 0;
1702                 env->le_ctx.lc_tags |= ctags;
1703         }
1704
1705         if (env->le_ses && (env->le_ses->lc_tags & stags) != stags) {
1706                 env->le_ses->lc_version = 0;
1707                 env->le_ses->lc_tags |= stags;
1708         }
1709
1710         result = lu_env_refill(env);
1711
1712         return result;
1713 }
1714 EXPORT_SYMBOL(lu_env_refill_by_tags);
1715
1716 static struct cfs_shrinker *lu_site_shrinker = NULL;
1717
1718 typedef struct lu_site_stats{
1719         unsigned        lss_populated;
1720         unsigned        lss_max_search;
1721         unsigned        lss_total;
1722         unsigned        lss_busy;
1723 } lu_site_stats_t;
1724
1725 static void lu_site_stats_get(cfs_hash_t *hs,
1726                               lu_site_stats_t *stats, int populated)
1727 {
1728         cfs_hash_bd_t bd;
1729         int           i;
1730
1731         cfs_hash_for_each_bucket(hs, &bd, i) {
1732                 struct lu_site_bkt_data *bkt = cfs_hash_bd_extra_get(hs, &bd);
1733                 cfs_hlist_head_t        *hhead;
1734
1735                 cfs_hash_bd_lock(hs, &bd, 1);
1736                 stats->lss_busy  += bkt->lsb_busy;
1737                 stats->lss_total += cfs_hash_bd_count_get(&bd);
1738                 stats->lss_max_search = max((int)stats->lss_max_search,
1739                                             cfs_hash_bd_depmax_get(&bd));
1740                 if (!populated) {
1741                         cfs_hash_bd_unlock(hs, &bd, 1);
1742                         continue;
1743                 }
1744
1745                 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
1746                         if (!cfs_hlist_empty(hhead))
1747                                 stats->lss_populated++;
1748                 }
1749                 cfs_hash_bd_unlock(hs, &bd, 1);
1750         }
1751 }
1752
1753 #ifdef __KERNEL__
1754
1755 static int lu_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
1756 {
1757         lu_site_stats_t stats;
1758         struct lu_site *s;
1759         struct lu_site *tmp;
1760         int cached = 0;
1761         int remain = shrink_param(sc, nr_to_scan);
1762         CFS_LIST_HEAD(splice);
1763
1764         if (remain != 0) {
1765                 if (!(shrink_param(sc, gfp_mask) & __GFP_FS))
1766                         return -1;
1767                 CDEBUG(D_INODE, "Shrink %d objects\n", remain);
1768         }
1769
1770         cfs_mutex_lock(&lu_sites_guard);
1771         cfs_list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
1772                 if (shrink_param(sc, nr_to_scan) != 0) {
1773                         remain = lu_site_purge(&lu_shrink_env, s, remain);
1774                         /*
1775                          * Move just shrunk site to the tail of site list to
1776                          * assure shrinking fairness.
1777                          */
1778                         cfs_list_move_tail(&s->ls_linkage, &splice);
1779                 }
1780
1781                 memset(&stats, 0, sizeof(stats));
1782                 lu_site_stats_get(s->ls_obj_hash, &stats, 0);
1783                 cached += stats.lss_total - stats.lss_busy;
1784                 if (shrink_param(sc, nr_to_scan) && remain <= 0)
1785                         break;
1786         }
1787         cfs_list_splice(&splice, lu_sites.prev);
1788         cfs_mutex_unlock(&lu_sites_guard);
1789
1790         cached = (cached / 100) * sysctl_vfs_cache_pressure;
1791         if (shrink_param(sc, nr_to_scan) == 0)
1792                 CDEBUG(D_INODE, "%d objects cached\n", cached);
1793         return cached;
1794 }
1795
1796 /*
1797  * Debugging stuff.
1798  */
1799
1800 /**
1801  * Environment to be used in debugger, contains all tags.
1802  */
1803 struct lu_env lu_debugging_env;
1804
1805 /**
1806  * Debugging printer function using printk().
1807  */
1808 int lu_printk_printer(const struct lu_env *env,
1809                       void *unused, const char *format, ...)
1810 {
1811         va_list args;
1812
1813         va_start(args, format);
1814         vprintk(format, args);
1815         va_end(args);
1816         return 0;
1817 }
1818
1819 void lu_debugging_setup(void)
1820 {
1821         lu_env_init(&lu_debugging_env, ~0);
1822 }
1823
1824 void lu_context_keys_dump(void)
1825 {
1826         int i;
1827
1828         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1829                 struct lu_context_key *key;
1830
1831                 key = lu_keys[i];
1832                 if (key != NULL) {
1833                         CERROR("[%d]: %p %x (%p,%p,%p) %d %d \"%s\"@%p\n",
1834                                i, key, key->lct_tags,
1835                                key->lct_init, key->lct_fini, key->lct_exit,
1836                                key->lct_index, cfs_atomic_read(&key->lct_used),
1837                                key->lct_owner ? key->lct_owner->name : "",
1838                                key->lct_owner);
1839                         lu_ref_print(&key->lct_reference);
1840                 }
1841         }
1842 }
1843 EXPORT_SYMBOL(lu_context_keys_dump);
1844 #else  /* !__KERNEL__ */
1845 static int lu_cache_shrink(int nr, unsigned int gfp_mask)
1846 {
1847         return 0;
1848 }
1849 #endif /* __KERNEL__ */
1850
1851 int  cl_global_init(void);
1852 void cl_global_fini(void);
1853 int  lu_ref_global_init(void);
1854 void lu_ref_global_fini(void);
1855
1856 int dt_global_init(void);
1857 void dt_global_fini(void);
1858
1859 int llo_global_init(void);
1860 void llo_global_fini(void);
1861
1862 /**
1863  * Initialization of global lu_* data.
1864  */
1865 int lu_global_init(void)
1866 {
1867         int result;
1868
1869         CDEBUG(D_INFO, "Lustre LU module (%p).\n", &lu_keys);
1870
1871         result = lu_ref_global_init();
1872         if (result != 0)
1873                 return result;
1874
1875         LU_CONTEXT_KEY_INIT(&lu_global_key);
1876         result = lu_context_key_register(&lu_global_key);
1877         if (result != 0)
1878                 return result;
1879         /*
1880          * At this level, we don't know what tags are needed, so allocate them
1881          * conservatively. This should not be too bad, because this
1882          * environment is global.
1883          */
1884         cfs_mutex_lock(&lu_sites_guard);
1885         result = lu_env_init(&lu_shrink_env, LCT_SHRINKER);
1886         cfs_mutex_unlock(&lu_sites_guard);
1887         if (result != 0)
1888                 return result;
1889
1890         /*
1891          * seeks estimation: 3 seeks to read a record from oi, one to read
1892          * inode, one for ea. Unfortunately setting this high value results in
1893          * lu_object/inode cache consuming all the memory.
1894          */
1895         lu_site_shrinker = cfs_set_shrinker(CFS_DEFAULT_SEEKS, lu_cache_shrink);
1896         if (lu_site_shrinker == NULL)
1897                 return -ENOMEM;
1898
1899         result = lu_time_global_init();
1900         if (result)
1901                 GOTO(out, result);
1902
1903 #ifdef __KERNEL__
1904         result = dt_global_init();
1905         if (result)
1906                 GOTO(out, result);
1907
1908         result = llo_global_init();
1909         if (result)
1910                 GOTO(out, result);
1911 #endif
1912         result = cl_global_init();
1913 out:
1914
1915         return result;
1916 }
1917
1918 /**
1919  * Dual to lu_global_init().
1920  */
1921 void lu_global_fini(void)
1922 {
1923         cl_global_fini();
1924 #ifdef __KERNEL__
1925         llo_global_fini();
1926         dt_global_fini();
1927 #endif
1928         lu_time_global_fini();
1929         if (lu_site_shrinker != NULL) {
1930                 cfs_remove_shrinker(lu_site_shrinker);
1931                 lu_site_shrinker = NULL;
1932         }
1933
1934         lu_context_key_degister(&lu_global_key);
1935
1936         /*
1937          * Tear shrinker environment down _after_ de-registering
1938          * lu_global_key, because the latter has a value in the former.
1939          */
1940         cfs_mutex_lock(&lu_sites_guard);
1941         lu_env_fini(&lu_shrink_env);
1942         cfs_mutex_unlock(&lu_sites_guard);
1943
1944         lu_ref_global_fini();
1945 }
1946
1947 struct lu_buf LU_BUF_NULL = {
1948         .lb_buf = NULL,
1949         .lb_len = 0
1950 };
1951 EXPORT_SYMBOL(LU_BUF_NULL);
1952
1953 static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx)
1954 {
1955 #ifdef LPROCFS
1956         struct lprocfs_counter ret;
1957
1958         lprocfs_stats_collect(stats, idx, &ret);
1959         return (__u32)ret.lc_count;
1960 #else
1961         return 0;
1962 #endif
1963 }
1964
1965 /**
1966  * Output site statistical counters into a buffer. Suitable for
1967  * lprocfs_rd_*()-style functions.
1968  */
1969 int lu_site_stats_print(const struct lu_site *s, char *page, int count)
1970 {
1971         lu_site_stats_t stats;
1972
1973         memset(&stats, 0, sizeof(stats));
1974         lu_site_stats_get(s->ls_obj_hash, &stats, 1);
1975
1976         return snprintf(page, count, "%d/%d %d/%d %d %d %d %d %d %d %d\n",
1977                         stats.lss_busy,
1978                         stats.lss_total,
1979                         stats.lss_populated,
1980                         CFS_HASH_NHLIST(s->ls_obj_hash),
1981                         stats.lss_max_search,
1982                         ls_stats_read(s->ls_stats, LU_SS_CREATED),
1983                         ls_stats_read(s->ls_stats, LU_SS_CACHE_HIT),
1984                         ls_stats_read(s->ls_stats, LU_SS_CACHE_MISS),
1985                         ls_stats_read(s->ls_stats, LU_SS_CACHE_RACE),
1986                         ls_stats_read(s->ls_stats, LU_SS_CACHE_DEATH_RACE),
1987                         ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED));
1988 }
1989 EXPORT_SYMBOL(lu_site_stats_print);
1990
1991 const char *lu_time_names[LU_TIME_NR] = {
1992         [LU_TIME_FIND_LOOKUP] = "find_lookup",
1993         [LU_TIME_FIND_ALLOC]  = "find_alloc",
1994         [LU_TIME_FIND_INSERT] = "find_insert"
1995 };
1996 EXPORT_SYMBOL(lu_time_names);
1997
1998 /**
1999  * Helper function to initialize a number of kmem slab caches at once.
2000  */
2001 int lu_kmem_init(struct lu_kmem_descr *caches)
2002 {
2003         int result;
2004         struct lu_kmem_descr *iter = caches;
2005
2006         for (result = 0; iter->ckd_cache != NULL; ++iter) {
2007                 *iter->ckd_cache = cfs_mem_cache_create(iter->ckd_name,
2008                                                         iter->ckd_size,
2009                                                         0, 0);
2010                 if (*iter->ckd_cache == NULL) {
2011                         result = -ENOMEM;
2012                         /* free all previously allocated caches */
2013                         lu_kmem_fini(caches);
2014                         break;
2015                 }
2016         }
2017         return result;
2018 }
2019 EXPORT_SYMBOL(lu_kmem_init);
2020
2021 /**
2022  * Helper function to finalize a number of kmem slab cached at once. Dual to
2023  * lu_kmem_init().
2024  */
2025 void lu_kmem_fini(struct lu_kmem_descr *caches)
2026 {
2027         int rc;
2028
2029         for (; caches->ckd_cache != NULL; ++caches) {
2030                 if (*caches->ckd_cache != NULL) {
2031                         rc = cfs_mem_cache_destroy(*caches->ckd_cache);
2032                         LASSERTF(rc == 0, "couldn't destroy %s slab\n",
2033                                  caches->ckd_name);
2034                         *caches->ckd_cache = NULL;
2035                 }
2036         }
2037 }
2038 EXPORT_SYMBOL(lu_kmem_fini);