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 /*
  33  * Copyright (c) 2011 Whamcloud, Inc.
  34  */
  35 /*
  36  * This file is part of Lustre, http://www.lustre.org/
  37  * Lustre is a trademark of Sun Microsystems, Inc.
  38  *
  39  * lustre/obdclass/lu_object.c
  40  *
  41  * Lustre Object.
  42  * These are the only exported functions, they provide some generic
  43  * infrastructure for managing object devices
  44  *
  45  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  46  */
  47
  48 #define DEBUG_SUBSYSTEM S_CLASS
  49 #ifndef EXPORT_SYMTAB
  50 # define EXPORT_SYMTAB
  51 #endif
  52
  53 #include <libcfs/libcfs.h>
  54
  55 #ifdef __KERNEL__
  56 # include <linux/module.h>
  57 #endif
  58
  59 /* hash_long() */
  60 #include <libcfs/libcfs_hash.h>
  61 #include <obd_class.h>
  62 #include <obd_support.h>
  63 #include <lustre_disk.h>
  64 #include <lustre_fid.h>
  65 #include <lu_object.h>
  66 #include <libcfs/list.h>
  67 /* lu_time_global_{init,fini}() */
  68 #include <lu_time.h>
  69
  70 static void lu_object_free(const struct lu_env *env, struct lu_object *o);
  71
  72 /**
  73  * Decrease reference counter on object. If last reference is freed, return
  74  * object to the cache, unless lu_object_is_dying(o) holds. In the latter
  75  * case, free object immediately.
  76  */
  77 void lu_object_put(const struct lu_env *env, struct lu_object *o)
  78 {
  79         struct lu_site_bkt_data *bkt;
  80         struct lu_object_header *top;
  81         struct lu_site          *site;
  82         struct lu_object        *orig;
  83         cfs_hash_bd_t            bd;
  84
  85         top  = o->lo_header;
  86         site = o->lo_dev->ld_site;
  87         orig = o;
  88
  89         cfs_hash_bd_get(site->ls_obj_hash, &top->loh_fid, &bd);
  90         bkt = cfs_hash_bd_extra_get(site->ls_obj_hash, &bd);
  91
  92         if (!cfs_hash_bd_dec_and_lock(site->ls_obj_hash, &bd, &top->loh_ref)) {
  93                 if (lu_object_is_dying(top)) {
  94
  95                         /*
  96                          * somebody may be waiting for this, currently only
  97                          * used for cl_object, see cl_object_put_last().
  98                          */
  99                         cfs_waitq_broadcast(&bkt->lsb_marche_funebre);
 100                 }
 101                 return;
 102         }
 103
 104         LASSERT(bkt->lsb_busy > 0);
 105         bkt->lsb_busy--;
 106         /*
 107          * When last reference is released, iterate over object
 108          * layers, and notify them that object is no longer busy.
 109          */
 110         cfs_list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
 111                 if (o->lo_ops->loo_object_release != NULL)
 112                         o->lo_ops->loo_object_release(env, o);
 113         }
 114
 115         if (!lu_object_is_dying(top)) {
 116                 LASSERT(cfs_list_empty(&top->loh_lru));
 117                 cfs_list_add_tail(&top->loh_lru, &bkt->lsb_lru);
 118                 cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
 119                 return;
 120         }
 121
 122         /*
 123          * If object is dying (will not be cached), removed it
 124          * from hash table and LRU.
 125          *
 126          * This is done with hash table and LRU lists locked. As the only
 127          * way to acquire first reference to previously unreferenced
 128          * object is through hash-table lookup (lu_object_find()),
 129          * or LRU scanning (lu_site_purge()), that are done under hash-table
 130          * and LRU lock, no race with concurrent object lookup is possible
 131          * and we can safely destroy object below.
 132          */
 133         cfs_hash_bd_del_locked(site->ls_obj_hash, &bd, &top->loh_hash);
 134         cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
 135         /*
 136          * Object was already removed from hash and lru above, can
 137          * kill it.
 138          */
 139         lu_object_free(env, orig);
 140 }
 141 EXPORT_SYMBOL(lu_object_put);
 142
 143 /**
 144  * Allocate new object.
 145  *
 146  * This follows object creation protocol, described in the comment within
 147  * struct lu_device_operations definition.
 148  */
 149 static struct lu_object *lu_object_alloc(const struct lu_env *env,
 150                                          struct lu_device *dev,
 151                                          const struct lu_fid *f,
 152                                          const struct lu_object_conf *conf)
 153 {
 154         struct lu_object *scan;
 155         struct lu_object *top;
 156         cfs_list_t *layers;
 157         int clean;
 158         int result;
 159         ENTRY;
 160
 161         /*
 162          * Create top-level object slice. This will also create
 163          * lu_object_header.
 164          */
 165         top = dev->ld_ops->ldo_object_alloc(env, NULL, dev);
 166         if (top == NULL)
 167                 RETURN(ERR_PTR(-ENOMEM));
 168         /*
 169          * This is the only place where object fid is assigned. It's constant
 170          * after this point.
 171          */
 172         LASSERT(fid_is_igif(f) || fid_ver(f) == 0);
 173         top->lo_header->loh_fid = *f;
 174         layers = &top->lo_header->loh_layers;
 175         do {
 176                 /*
 177                  * Call ->loo_object_init() repeatedly, until no more new
 178                  * object slices are created.
 179                  */
 180                 clean = 1;
 181                 cfs_list_for_each_entry(scan, layers, lo_linkage) {
 182                         if (scan->lo_flags & LU_OBJECT_ALLOCATED)
 183                                 continue;
 184                         clean = 0;
 185                         scan->lo_header = top->lo_header;
 186                         result = scan->lo_ops->loo_object_init(env, scan, conf);
 187                         if (result != 0) {
 188                                 lu_object_free(env, top);
 189                                 RETURN(ERR_PTR(result));
 190                         }
 191                         scan->lo_flags |= LU_OBJECT_ALLOCATED;
 192                 }
 193         } while (!clean);
 194
 195         cfs_list_for_each_entry_reverse(scan, layers, lo_linkage) {
 196                 if (scan->lo_ops->loo_object_start != NULL) {
 197                         result = scan->lo_ops->loo_object_start(env, scan);
 198                         if (result != 0) {
 199                                 lu_object_free(env, top);
 200                                 RETURN(ERR_PTR(result));
 201                         }
 202                 }
 203         }
 204
 205         lprocfs_counter_incr(dev->ld_site->ls_stats, LU_SS_CREATED);
 206         RETURN(top);
 207 }
 208
 209 /**
 210  * Free an object.
 211  */
 212 static void lu_object_free(const struct lu_env *env, struct lu_object *o)
 213 {
 214         struct lu_site_bkt_data *bkt;
 215         struct lu_site          *site;
 216         struct lu_object        *scan;
 217         cfs_list_t              *layers;
 218         cfs_list_t               splice;
 219
 220         site   = o->lo_dev->ld_site;
 221         layers = &o->lo_header->loh_layers;
 222         bkt    = lu_site_bkt_from_fid(site, &o->lo_header->loh_fid);
 223         /*
 224          * First call ->loo_object_delete() method to release all resources.
 225          */
 226         cfs_list_for_each_entry_reverse(scan, layers, lo_linkage) {
 227                 if (scan->lo_ops->loo_object_delete != NULL)
 228                         scan->lo_ops->loo_object_delete(env, scan);
 229         }
 230
 231         /*
 232          * Then, splice object layers into stand-alone list, and call
 233          * ->loo_object_free() on all layers to free memory. Splice is
 234          * necessary, because lu_object_header is freed together with the
 235          * top-level slice.
 236          */
 237         CFS_INIT_LIST_HEAD(&splice);
 238         cfs_list_splice_init(layers, &splice);
 239         while (!cfs_list_empty(&splice)) {
 240                 /*
 241                  * Free layers in bottom-to-top order, so that object header
 242                  * lives as long as possible and ->loo_object_free() methods
 243                  * can look at its contents.
 244                  */
 245                 o = container_of0(splice.prev, struct lu_object, lo_linkage);
 246                 cfs_list_del_init(&o->lo_linkage);
 247                 LASSERT(o->lo_ops->loo_object_free != NULL);
 248                 o->lo_ops->loo_object_free(env, o);
 249         }
 250
 251         if (cfs_waitq_active(&bkt->lsb_marche_funebre))
 252                 cfs_waitq_broadcast(&bkt->lsb_marche_funebre);
 253 }
 254
 255 /**
 256  * Free \a nr objects from the cold end of the site LRU list.
 257  */
 258 int lu_site_purge(const struct lu_env *env, struct lu_site *s, int nr)
 259 {
 260         struct lu_object_header *h;
 261         struct lu_object_header *temp;
 262         struct lu_site_bkt_data *bkt;
 263         cfs_hash_bd_t            bd;
 264         cfs_hash_bd_t            bd2;
 265         cfs_list_t               dispose;
 266         int                      did_sth;
 267         int                      start;
 268         int                      count;
 269         int                      bnr;
 270         int                      i;
 271
 272         CFS_INIT_LIST_HEAD(&dispose);
 273         /*
 274          * Under LRU list lock, scan LRU list and move unreferenced objects to
 275          * the dispose list, removing them from LRU and hash table.
 276          */
 277         start = s->ls_purge_start;
 278         bnr = (nr == ~0) ? -1 : nr / CFS_HASH_NBKT(s->ls_obj_hash) + 1;
 279  again:
 280         did_sth = 0;
 281         cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
 282                 if (i < start)
 283                         continue;
 284                 count = bnr;
 285                 cfs_hash_bd_lock(s->ls_obj_hash, &bd, 1);
 286                 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
 287
 288                 cfs_list_for_each_entry_safe(h, temp, &bkt->lsb_lru, loh_lru) {
 289                         LASSERT(cfs_atomic_read(&h->loh_ref) == 0);
 290
 291                         cfs_hash_bd_get(s->ls_obj_hash, &h->loh_fid, &bd2);
 292                         LASSERT(bd.bd_bucket == bd2.bd_bucket);
 293
 294                         cfs_hash_bd_del_locked(s->ls_obj_hash,
 295                                                &bd2, &h->loh_hash);
 296                         cfs_list_move(&h->loh_lru, &dispose);
 297                         if (did_sth == 0)
 298                                 did_sth = 1;
 299
 300                         if (nr != ~0 && --nr == 0)
 301                                 break;
 302
 303                         if (count > 0 && --count == 0)
 304                                 break;
 305
 306                 }
 307                 cfs_hash_bd_unlock(s->ls_obj_hash, &bd, 1);
 308                 cfs_cond_resched();
 309                 /*
 310                  * Free everything on the dispose list. This is safe against
 311                  * races due to the reasons described in lu_object_put().
 312                  */
 313                 while (!cfs_list_empty(&dispose)) {
 314                         h = container_of0(dispose.next,
 315                                           struct lu_object_header, loh_lru);
 316                         cfs_list_del_init(&h->loh_lru);
 317                         lu_object_free(env, lu_object_top(h));
 318                         lprocfs_counter_incr(s->ls_stats, LU_SS_LRU_PURGED);
 319                 }
 320
 321                 if (nr == 0)
 322                         break;
 323         }
 324
 325         if (nr != 0 && did_sth && start != 0) {
 326                 start = 0; /* restart from the first bucket */
 327                 goto again;
 328         }
 329         /* race on s->ls_purge_start, but nobody cares */
 330         s->ls_purge_start = i % CFS_HASH_NBKT(s->ls_obj_hash);
 331
 332         return nr;
 333 }
 334 EXPORT_SYMBOL(lu_site_purge);
 335
 336 /*
 337  * Object printing.
 338  *
 339  * Code below has to jump through certain loops to output object description
 340  * into libcfs_debug_msg-based log. The problem is that lu_object_print()
 341  * composes object description from strings that are parts of _lines_ of
 342  * output (i.e., strings that are not terminated by newline). This doesn't fit
 343  * very well into libcfs_debug_msg() interface that assumes that each message
 344  * supplied to it is a self-contained output line.
 345  *
 346  * To work around this, strings are collected in a temporary buffer
 347  * (implemented as a value of lu_cdebug_key key), until terminating newline
 348  * character is detected.
 349  *
 350  */
 351
 352 enum {
 353         /**
 354          * Maximal line size.
 355          *
 356          * XXX overflow is not handled correctly.
 357          */
 358         LU_CDEBUG_LINE = 512
 359 };
 360
 361 struct lu_cdebug_data {
 362         /**
 363          * Temporary buffer.
 364          */
 365         char lck_area[LU_CDEBUG_LINE];
 366 };
 367
 368 /* context key constructor/destructor: lu_global_key_init, lu_global_key_fini */
 369 LU_KEY_INIT_FINI(lu_global, struct lu_cdebug_data);
 370
 371 /**
 372  * Key, holding temporary buffer. This key is registered very early by
 373  * lu_global_init().
 374  */
 375 struct lu_context_key lu_global_key = {
 376         .lct_tags = LCT_MD_THREAD|LCT_DT_THREAD|LCT_CL_THREAD,
 377         .lct_init = lu_global_key_init,
 378         .lct_fini = lu_global_key_fini
 379 };
 380
 381 /**
 382  * Printer function emitting messages through libcfs_debug_msg().
 383  */
 384 int lu_cdebug_printer(const struct lu_env *env,
 385                       void *cookie, const char *format, ...)
 386 {
 387         struct lu_cdebug_print_info *info = cookie;
 388         struct lu_cdebug_data       *key;
 389         int used;
 390         int complete;
 391         va_list args;
 392
 393         va_start(args, format);
 394
 395         key = lu_context_key_get(&env->le_ctx, &lu_global_key);
 396         LASSERT(key != NULL);
 397
 398         used = strlen(key->lck_area);
 399         complete = format[strlen(format) - 1] == '\n';
 400         /*
 401          * Append new chunk to the buffer.
 402          */
 403         vsnprintf(key->lck_area + used,
 404                   ARRAY_SIZE(key->lck_area) - used, format, args);
 405         if (complete) {
 406                 if (cfs_cdebug_show(info->lpi_mask, info->lpi_subsys))
 407                         libcfs_debug_msg(NULL, info->lpi_subsys, info->lpi_mask,
 408                                          (char *)info->lpi_file, info->lpi_fn,
 409                                          info->lpi_line, "%s", key->lck_area);
 410                 key->lck_area[0] = 0;
 411         }
 412         va_end(args);
 413         return 0;
 414 }
 415 EXPORT_SYMBOL(lu_cdebug_printer);
 416
 417 /**
 418  * Print object header.
 419  */
 420 void lu_object_header_print(const struct lu_env *env, void *cookie,
 421                             lu_printer_t printer,
 422                             const struct lu_object_header *hdr)
 423 {
 424         (*printer)(env, cookie, "header@%p[%#lx, %d, "DFID"%s%s%s]",
 425                    hdr, hdr->loh_flags, cfs_atomic_read(&hdr->loh_ref),
 426                    PFID(&hdr->loh_fid),
 427                    cfs_hlist_unhashed(&hdr->loh_hash) ? "" : " hash",
 428                    cfs_list_empty((cfs_list_t *)&hdr->loh_lru) ? \
 429                    "" : " lru",
 430                    hdr->loh_attr & LOHA_EXISTS ? " exist":"");
 431 }
 432 EXPORT_SYMBOL(lu_object_header_print);
 433
 434 /**
 435  * Print human readable representation of the \a o to the \a printer.
 436  */
 437 void lu_object_print(const struct lu_env *env, void *cookie,
 438                      lu_printer_t printer, const struct lu_object *o)
 439 {
 440         static const char ruler[] = "........................................";
 441         struct lu_object_header *top;
 442         int depth;
 443
 444         top = o->lo_header;
 445         lu_object_header_print(env, cookie, printer, top);
 446         (*printer)(env, cookie, "{ \n");
 447         cfs_list_for_each_entry(o, &top->loh_layers, lo_linkage) {
 448                 depth = o->lo_depth + 4;
 449
 450                 /*
 451                  * print `.' \a depth times followed by type name and address
 452                  */
 453                 (*printer)(env, cookie, "%*.*s%s@%p", depth, depth, ruler,
 454                            o->lo_dev->ld_type->ldt_name, o);
 455                 if (o->lo_ops->loo_object_print != NULL)
 456                         o->lo_ops->loo_object_print(env, cookie, printer, o);
 457                 (*printer)(env, cookie, "\n");
 458         }
 459         (*printer)(env, cookie, "} header@%p\n", top);
 460 }
 461 EXPORT_SYMBOL(lu_object_print);
 462
 463 /**
 464  * Check object consistency.
 465  */
 466 int lu_object_invariant(const struct lu_object *o)
 467 {
 468         struct lu_object_header *top;
 469
 470         top = o->lo_header;
 471         cfs_list_for_each_entry(o, &top->loh_layers, lo_linkage) {
 472                 if (o->lo_ops->loo_object_invariant != NULL &&
 473                     !o->lo_ops->loo_object_invariant(o))
 474                         return 0;
 475         }
 476         return 1;
 477 }
 478 EXPORT_SYMBOL(lu_object_invariant);
 479
 480 static struct lu_object *htable_lookup(struct lu_site *s,
 481                                        cfs_hash_bd_t *bd,
 482                                        const struct lu_fid *f,
 483                                        cfs_waitlink_t *waiter,
 484                                        __u64 *version)
 485 {
 486         struct lu_site_bkt_data *bkt;
 487         struct lu_object_header *h;
 488         cfs_hlist_node_t        *hnode;
 489         __u64  ver = cfs_hash_bd_version_get(bd);
 490
 491         if (*version == ver)
 492                 return NULL;
 493
 494         *version = ver;
 495         bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, bd);
 496         /* cfs_hash_bd_lookup_intent is a somehow "internal" function
 497          * of cfs_hash, but we don't want refcount on object right now */
 498         hnode = cfs_hash_bd_lookup_locked(s->ls_obj_hash, bd, (void *)f);
 499         if (hnode == NULL) {
 500                 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_MISS);
 501                 return NULL;
 502         }
 503
 504         h = container_of0(hnode, struct lu_object_header, loh_hash);
 505         if (likely(!lu_object_is_dying(h))) {
 506                 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_HIT);
 507                 cfs_list_del_init(&h->loh_lru);
 508                 return lu_object_top(h);
 509         }
 510
 511         /*
 512          * Lookup found an object being destroyed this object cannot be
 513          * returned (to assure that references to dying objects are eventually
 514          * drained), and moreover, lookup has to wait until object is freed.
 515          */
 516         cfs_atomic_dec(&h->loh_ref);
 517
 518         cfs_waitlink_init(waiter);
 519         cfs_waitq_add(&bkt->lsb_marche_funebre, waiter);
 520         cfs_set_current_state(CFS_TASK_UNINT);
 521         lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_DEATH_RACE);
 522         return ERR_PTR(-EAGAIN);
 523 }
 524
 525 /**
 526  * Search cache for an object with the fid \a f. If such object is found,
 527  * return it. Otherwise, create new object, insert it into cache and return
 528  * it. In any case, additional reference is acquired on the returned object.
 529  */
 530 struct lu_object *lu_object_find(const struct lu_env *env,
 531                                  struct lu_device *dev, const struct lu_fid *f,
 532                                  const struct lu_object_conf *conf)
 533 {
 534         return lu_object_find_at(env, dev->ld_site->ls_top_dev, f, conf);
 535 }
 536 EXPORT_SYMBOL(lu_object_find);
 537
 538 static struct lu_object *lu_object_new(const struct lu_env *env,
 539                                        struct lu_device *dev,
 540                                        const struct lu_fid *f,
 541                                        const struct lu_object_conf *conf)
 542 {
 543         struct lu_object        *o;
 544         cfs_hash_t              *hs;
 545         cfs_hash_bd_t            bd;
 546         struct lu_site_bkt_data *bkt;
 547
 548         o = lu_object_alloc(env, dev, f, conf);
 549         if (unlikely(IS_ERR(o)))
 550                 return o;
 551
 552         hs = dev->ld_site->ls_obj_hash;
 553         cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
 554         bkt = cfs_hash_bd_extra_get(hs, &bd);
 555         cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
 556         bkt->lsb_busy++;
 557         cfs_hash_bd_unlock(hs, &bd, 1);
 558         return o;
 559 }
 560
 561 /**
 562  * Core logic of lu_object_find*() functions.
 563  */
 564 static struct lu_object *lu_object_find_try(const struct lu_env *env,
 565                                             struct lu_device *dev,
 566                                             const struct lu_fid *f,
 567                                             const struct lu_object_conf *conf,
 568                                             cfs_waitlink_t *waiter)
 569 {
 570         struct lu_object      *o;
 571         struct lu_object      *shadow;
 572         struct lu_site        *s;
 573         cfs_hash_t            *hs;
 574         cfs_hash_bd_t          bd;
 575         __u64                  version = 0;
 576
 577         /*
 578          * This uses standard index maintenance protocol:
 579          *
 580          *     - search index under lock, and return object if found;
 581          *     - otherwise, unlock index, allocate new object;
 582          *     - lock index and search again;
 583          *     - if nothing is found (usual case), insert newly created
 584          *       object into index;
 585          *     - otherwise (race: other thread inserted object), free
 586          *       object just allocated.
 587          *     - unlock index;
 588          *     - return object.
 589          *
 590          * For "LOC_F_NEW" case, we are sure the object is new established.
 591          * It is unnecessary to perform lookup-alloc-lookup-insert, instead,
 592          * just alloc and insert directly.
 593          *
 594          * If dying object is found during index search, add @waiter to the
 595          * site wait-queue and return ERR_PTR(-EAGAIN).
 596          */
 597         if (conf != NULL && conf->loc_flags & LOC_F_NEW)
 598                 return lu_object_new(env, dev, f, conf);
 599
 600         s  = dev->ld_site;
 601         hs = s->ls_obj_hash;
 602         cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
 603         o = htable_lookup(s, &bd, f, waiter, &version);
 604         cfs_hash_bd_unlock(hs, &bd, 1);
 605         if (o != NULL)
 606                 return o;
 607
 608         /*
 609          * Allocate new object. This may result in rather complicated
 610          * operations, including fld queries, inode loading, etc.
 611          */
 612         o = lu_object_alloc(env, dev, f, conf);
 613         if (unlikely(IS_ERR(o)))
 614                 return o;
 615
 616         LASSERT(lu_fid_eq(lu_object_fid(o), f));
 617
 618         cfs_hash_bd_lock(hs, &bd, 1);
 619
 620         shadow = htable_lookup(s, &bd, f, waiter, &version);
 621         if (likely(shadow == NULL)) {
 622                 struct lu_site_bkt_data *bkt;
 623
 624                 bkt = cfs_hash_bd_extra_get(hs, &bd);
 625                 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
 626                 bkt->lsb_busy++;
 627                 cfs_hash_bd_unlock(hs, &bd, 1);
 628                 return o;
 629         }
 630
 631         lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_RACE);
 632         cfs_hash_bd_unlock(hs, &bd, 1);
 633         lu_object_free(env, o);
 634         return shadow;
 635 }
 636
 637 /**
 638  * Much like lu_object_find(), but top level device of object is specifically
 639  * \a dev rather than top level device of the site. This interface allows
 640  * objects of different "stacking" to be created within the same site.
 641  */
 642 struct lu_object *lu_object_find_at(const struct lu_env *env,
 643                                     struct lu_device *dev,
 644                                     const struct lu_fid *f,
 645                                     const struct lu_object_conf *conf)
 646 {
 647         struct lu_site_bkt_data *bkt;
 648         struct lu_object        *obj;
 649         cfs_waitlink_t           wait;
 650
 651         while (1) {
 652                 obj = lu_object_find_try(env, dev, f, conf, &wait);
 653                 if (obj != ERR_PTR(-EAGAIN))
 654                         return obj;
 655                 /*
 656                  * lu_object_find_try() already added waiter into the
 657                  * wait queue.
 658                  */
 659                 cfs_waitq_wait(&wait, CFS_TASK_UNINT);
 660                 bkt = lu_site_bkt_from_fid(dev->ld_site, (void *)f);
 661                 cfs_waitq_del(&bkt->lsb_marche_funebre, &wait);
 662         }
 663 }
 664 EXPORT_SYMBOL(lu_object_find_at);
 665
 666 /**
 667  * Find object with given fid, and return its slice belonging to given device.
 668  */
 669 struct lu_object *lu_object_find_slice(const struct lu_env *env,
 670                                        struct lu_device *dev,
 671                                        const struct lu_fid *f,
 672                                        const struct lu_object_conf *conf)
 673 {
 674         struct lu_object *top;
 675         struct lu_object *obj;
 676
 677         top = lu_object_find(env, dev, f, conf);
 678         if (!IS_ERR(top)) {
 679                 obj = lu_object_locate(top->lo_header, dev->ld_type);
 680                 if (obj == NULL)
 681                         lu_object_put(env, top);
 682         } else
 683                 obj = top;
 684         return obj;
 685 }
 686 EXPORT_SYMBOL(lu_object_find_slice);
 687
 688 /**
 689  * Global list of all device types.
 690  */
 691 static CFS_LIST_HEAD(lu_device_types);
 692
 693 int lu_device_type_init(struct lu_device_type *ldt)
 694 {
 695         int result;
 696
 697         CFS_INIT_LIST_HEAD(&ldt->ldt_linkage);
 698         result = ldt->ldt_ops->ldto_init(ldt);
 699         if (result == 0)
 700                 cfs_list_add(&ldt->ldt_linkage, &lu_device_types);
 701         return result;
 702 }
 703 EXPORT_SYMBOL(lu_device_type_init);
 704
 705 void lu_device_type_fini(struct lu_device_type *ldt)
 706 {
 707         cfs_list_del_init(&ldt->ldt_linkage);
 708         ldt->ldt_ops->ldto_fini(ldt);
 709 }
 710 EXPORT_SYMBOL(lu_device_type_fini);
 711
 712 void lu_types_stop(void)
 713 {
 714         struct lu_device_type *ldt;
 715
 716         cfs_list_for_each_entry(ldt, &lu_device_types, ldt_linkage) {
 717                 if (ldt->ldt_device_nr == 0)
 718                         ldt->ldt_ops->ldto_stop(ldt);
 719         }
 720 }
 721 EXPORT_SYMBOL(lu_types_stop);
 722
 723 /**
 724  * Global list of all sites on this node
 725  */
 726 static CFS_LIST_HEAD(lu_sites);
 727 static CFS_DECLARE_MUTEX(lu_sites_guard);
 728
 729 /**
 730  * Global environment used by site shrinker.
 731  */
 732 static struct lu_env lu_shrink_env;
 733
 734 struct lu_site_print_arg {
 735         struct lu_env   *lsp_env;
 736         void            *lsp_cookie;
 737         lu_printer_t     lsp_printer;
 738 };
 739
 740 static int
 741 lu_site_obj_print(cfs_hash_t *hs, cfs_hash_bd_t *bd,
 742                   cfs_hlist_node_t *hnode, void *data)
 743 {
 744         struct lu_site_print_arg *arg = (struct lu_site_print_arg *)data;
 745         struct lu_object_header  *h;
 746
 747         h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
 748         if (!cfs_list_empty(&h->loh_layers)) {
 749                 const struct lu_object *o;
 750
 751                 o = lu_object_top(h);
 752                 lu_object_print(arg->lsp_env, arg->lsp_cookie,
 753                                 arg->lsp_printer, o);
 754         } else {
 755                 lu_object_header_print(arg->lsp_env, arg->lsp_cookie,
 756                                        arg->lsp_printer, h);
 757         }
 758         return 0;
 759 }
 760
 761 /**
 762  * Print all objects in \a s.
 763  */
 764 void lu_site_print(const struct lu_env *env, struct lu_site *s, void *cookie,
 765                    lu_printer_t printer)
 766 {
 767         struct lu_site_print_arg arg = {
 768                 .lsp_env     = (struct lu_env *)env,
 769                 .lsp_cookie  = cookie,
 770                 .lsp_printer = printer,
 771         };
 772
 773         cfs_hash_for_each(s->ls_obj_hash, lu_site_obj_print, &arg);
 774 }
 775 EXPORT_SYMBOL(lu_site_print);
 776
 777 enum {
 778         LU_CACHE_PERCENT_MAX     = 50,
 779         LU_CACHE_PERCENT_DEFAULT = 20
 780 };
 781
 782 static unsigned int lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
 783 CFS_MODULE_PARM(lu_cache_percent, "i", int, 0644,
 784                 "Percentage of memory to be used as lu_object cache");
 785
 786 /**
 787  * Return desired hash table order.
 788  */
 789 static int lu_htable_order(void)
 790 {
 791         unsigned long cache_size;
 792         int bits;
 793
 794         /*
 795          * Calculate hash table size, assuming that we want reasonable
 796          * performance when 20% of total memory is occupied by cache of
 797          * lu_objects.
 798          *
 799          * Size of lu_object is (arbitrary) taken as 1K (together with inode).
 800          */
 801         cache_size = cfs_num_physpages;
 802
 803 #if BITS_PER_LONG == 32
 804         /* limit hashtable size for lowmem systems to low RAM */
 805         if (cache_size > 1 << (30 - CFS_PAGE_SHIFT))
 806                 cache_size = 1 << (30 - CFS_PAGE_SHIFT) * 3 / 4;
 807 #endif
 808
 809         /* clear off unreasonable cache setting. */
 810         if (lu_cache_percent == 0 || lu_cache_percent > LU_CACHE_PERCENT_MAX) {
 811                 CWARN("obdclass: invalid lu_cache_percent: %u, it must be in"
 812                       " the range of (0, %u]. Will use default value: %u.\n",
 813                       lu_cache_percent, LU_CACHE_PERCENT_MAX,
 814                       LU_CACHE_PERCENT_DEFAULT);
 815
 816                 lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
 817         }
 818         cache_size = cache_size / 100 * lu_cache_percent *
 819                 (CFS_PAGE_SIZE / 1024);
 820
 821         for (bits = 1; (1 << bits) < cache_size; ++bits) {
 822                 ;
 823         }
 824         return bits;
 825 }
 826
 827 static unsigned lu_obj_hop_hash(cfs_hash_t *hs,
 828                                 const void *key, unsigned mask)
 829 {
 830         struct lu_fid  *fid = (struct lu_fid *)key;
 831         __u32           hash;
 832
 833         hash = fid_flatten32(fid);
 834         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
 835         hash = cfs_hash_long(hash, hs->hs_bkt_bits);
 836
 837         /* give me another random factor */
 838         hash -= cfs_hash_long((unsigned long)hs, fid_oid(fid) % 11 + 3);
 839
 840         hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
 841         hash |= (fid_seq(fid) + fid_oid(fid)) & (CFS_HASH_NBKT(hs) - 1);
 842
 843         return hash & mask;
 844 }
 845
 846 static void *lu_obj_hop_object(cfs_hlist_node_t *hnode)
 847 {
 848         return cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
 849 }
 850
 851 static void *lu_obj_hop_key(cfs_hlist_node_t *hnode)
 852 {
 853         struct lu_object_header *h;
 854
 855         h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
 856         return &h->loh_fid;
 857 }
 858
 859 static int lu_obj_hop_keycmp(const void *key, cfs_hlist_node_t *hnode)
 860 {
 861         struct lu_object_header *h;
 862
 863         h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
 864         return lu_fid_eq(&h->loh_fid, (struct lu_fid *)key);
 865 }
 866
 867 static void lu_obj_hop_get(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 868 {
 869         struct lu_object_header *h;
 870
 871         h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
 872         if (cfs_atomic_add_return(1, &h->loh_ref) == 1) {
 873                 struct lu_site_bkt_data *bkt;
 874                 cfs_hash_bd_t            bd;
 875
 876                 cfs_hash_bd_get(hs, &h->loh_fid, &bd);
 877                 bkt = cfs_hash_bd_extra_get(hs, &bd);
 878                 bkt->lsb_busy++;
 879         }
 880 }
 881
 882 static void lu_obj_hop_put_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 883 {
 884         LBUG(); /* we should never called it */
 885 }
 886
 887 cfs_hash_ops_t lu_site_hash_ops = {
 888         .hs_hash        = lu_obj_hop_hash,
 889         .hs_key         = lu_obj_hop_key,
 890         .hs_keycmp      = lu_obj_hop_keycmp,
 891         .hs_object      = lu_obj_hop_object,
 892         .hs_get         = lu_obj_hop_get,
 893         .hs_put_locked  = lu_obj_hop_put_locked,
 894 };
 895
 896 /**
 897  * Initialize site \a s, with \a d as the top level device.
 898  */
 899 #define LU_SITE_BITS_MIN    12
 900 #define LU_SITE_BITS_MAX    24
 901 /**
 902  * total 256 buckets, we don't want too many buckets because:
 903  * - consume too much memory
 904  * - avoid unbalanced LRU list
 905  */
 906 #define LU_SITE_BKT_BITS    8
 907
 908 int lu_site_init(struct lu_site *s, struct lu_device *top)
 909 {
 910         struct lu_site_bkt_data *bkt;
 911         cfs_hash_bd_t bd;
 912         char name[16];
 913         int bits;
 914         int i;
 915         ENTRY;
 916
 917         memset(s, 0, sizeof *s);
 918         bits = lu_htable_order();
 919         snprintf(name, 16, "lu_site_%s", top->ld_type->ldt_name);
 920         for (bits = min(max(LU_SITE_BITS_MIN, bits), LU_SITE_BITS_MAX);
 921              bits >= LU_SITE_BITS_MIN; bits--) {
 922                 s->ls_obj_hash = cfs_hash_create(name, bits, bits,
 923                                                  bits - LU_SITE_BKT_BITS,
 924                                                  sizeof(*bkt), 0, 0,
 925                                                  &lu_site_hash_ops,
 926                                                  CFS_HASH_SPIN_BKTLOCK |
 927                                                  CFS_HASH_NO_ITEMREF |
 928                                                  CFS_HASH_DEPTH |
 929                                                  CFS_HASH_ASSERT_EMPTY);
 930                 if (s->ls_obj_hash != NULL)
 931                         break;
 932         }
 933
 934         if (s->ls_obj_hash == NULL) {
 935                 CERROR("failed to create lu_site hash with bits: %d\n", bits);
 936                 return -ENOMEM;
 937         }
 938
 939         cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
 940                 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
 941                 CFS_INIT_LIST_HEAD(&bkt->lsb_lru);
 942                 cfs_waitq_init(&bkt->lsb_marche_funebre);
 943         }
 944
 945         s->ls_stats = lprocfs_alloc_stats(LU_SS_LAST_STAT, 0);
 946         if (s->ls_stats == NULL) {
 947                 cfs_hash_putref(s->ls_obj_hash);
 948                 s->ls_obj_hash = NULL;
 949                 return -ENOMEM;
 950         }
 951
 952         lprocfs_counter_init(s->ls_stats, LU_SS_CREATED,
 953                              0, "created", "created");
 954         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_HIT,
 955                              0, "cache_hit", "cache_hit");
 956         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_MISS,
 957                              0, "cache_miss", "cache_miss");
 958         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_RACE,
 959                              0, "cache_race", "cache_race");
 960         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_DEATH_RACE,
 961                              0, "cache_death_race", "cache_death_race");
 962         lprocfs_counter_init(s->ls_stats, LU_SS_LRU_PURGED,
 963                              0, "lru_purged", "lru_purged");
 964
 965         CFS_INIT_LIST_HEAD(&s->ls_linkage);
 966         s->ls_top_dev = top;
 967         top->ld_site = s;
 968         lu_device_get(top);
 969         lu_ref_add(&top->ld_reference, "site-top", s);
 970
 971         CFS_INIT_LIST_HEAD(&s->ls_ld_linkage);
 972         cfs_spin_lock_init(&s->ls_ld_lock);
 973
 974         cfs_spin_lock(&s->ls_ld_lock);
 975         cfs_list_add(&top->ld_linkage, &s->ls_ld_linkage);
 976         cfs_spin_unlock(&s->ls_ld_lock);
 977
 978         RETURN(0);
 979 }
 980 EXPORT_SYMBOL(lu_site_init);
 981
 982 /**
 983  * Finalize \a s and release its resources.
 984  */
 985 void lu_site_fini(struct lu_site *s)
 986 {
 987         cfs_down(&lu_sites_guard);
 988         cfs_list_del_init(&s->ls_linkage);
 989         cfs_up(&lu_sites_guard);
 990
 991         if (s->ls_obj_hash != NULL) {
 992                 cfs_hash_putref(s->ls_obj_hash);
 993                 s->ls_obj_hash = NULL;
 994         }
 995
 996         if (s->ls_top_dev != NULL) {
 997                 s->ls_top_dev->ld_site = NULL;
 998                 lu_ref_del(&s->ls_top_dev->ld_reference, "site-top", s);
 999                 lu_device_put(s->ls_top_dev);
1000                 s->ls_top_dev = NULL;
1001         }
1002
1003         if (s->ls_stats != NULL)
1004                 lprocfs_free_stats(&s->ls_stats);
1005 }
1006 EXPORT_SYMBOL(lu_site_fini);
1007
1008 /**
1009  * Called when initialization of stack for this site is completed.
1010  */
1011 int lu_site_init_finish(struct lu_site *s)
1012 {
1013         int result;
1014         cfs_down(&lu_sites_guard);
1015         result = lu_context_refill(&lu_shrink_env.le_ctx);
1016         if (result == 0)
1017                 cfs_list_add(&s->ls_linkage, &lu_sites);
1018         cfs_up(&lu_sites_guard);
1019         return result;
1020 }
1021 EXPORT_SYMBOL(lu_site_init_finish);
1022
1023 /**
1024  * Acquire additional reference on device \a d
1025  */
1026 void lu_device_get(struct lu_device *d)
1027 {
1028         cfs_atomic_inc(&d->ld_ref);
1029 }
1030 EXPORT_SYMBOL(lu_device_get);
1031
1032 /**
1033  * Release reference on device \a d.
1034  */
1035 void lu_device_put(struct lu_device *d)
1036 {
1037         LASSERT(cfs_atomic_read(&d->ld_ref) > 0);
1038         cfs_atomic_dec(&d->ld_ref);
1039 }
1040 EXPORT_SYMBOL(lu_device_put);
1041
1042 /**
1043  * Initialize device \a d of type \a t.
1044  */
1045 int lu_device_init(struct lu_device *d, struct lu_device_type *t)
1046 {
1047         if (t->ldt_device_nr++ == 0 && t->ldt_ops->ldto_start != NULL)
1048                 t->ldt_ops->ldto_start(t);
1049         memset(d, 0, sizeof *d);
1050         cfs_atomic_set(&d->ld_ref, 0);
1051         d->ld_type = t;
1052         lu_ref_init(&d->ld_reference);
1053         CFS_INIT_LIST_HEAD(&d->ld_linkage);
1054         return 0;
1055 }
1056 EXPORT_SYMBOL(lu_device_init);
1057
1058 /**
1059  * Finalize device \a d.
1060  */
1061 void lu_device_fini(struct lu_device *d)
1062 {
1063         struct lu_device_type *t;
1064
1065         t = d->ld_type;
1066         if (d->ld_obd != NULL) {
1067                 d->ld_obd->obd_lu_dev = NULL;
1068                 d->ld_obd = NULL;
1069         }
1070
1071         lu_ref_fini(&d->ld_reference);
1072         LASSERTF(cfs_atomic_read(&d->ld_ref) == 0,
1073                  "Refcount is %u\n", cfs_atomic_read(&d->ld_ref));
1074         LASSERT(t->ldt_device_nr > 0);
1075         if (--t->ldt_device_nr == 0 && t->ldt_ops->ldto_stop != NULL)
1076                 t->ldt_ops->ldto_stop(t);
1077 }
1078 EXPORT_SYMBOL(lu_device_fini);
1079
1080 /**
1081  * Initialize object \a o that is part of compound object \a h and was created
1082  * by device \a d.
1083  */
1084 int lu_object_init(struct lu_object *o,
1085                    struct lu_object_header *h, struct lu_device *d)
1086 {
1087         memset(o, 0, sizeof *o);
1088         o->lo_header = h;
1089         o->lo_dev    = d;
1090         lu_device_get(d);
1091         o->lo_dev_ref = lu_ref_add(&d->ld_reference, "lu_object", o);
1092         CFS_INIT_LIST_HEAD(&o->lo_linkage);
1093         return 0;
1094 }
1095 EXPORT_SYMBOL(lu_object_init);
1096
1097 /**
1098  * Finalize object and release its resources.
1099  */
1100 void lu_object_fini(struct lu_object *o)
1101 {
1102         struct lu_device *dev = o->lo_dev;
1103
1104         LASSERT(cfs_list_empty(&o->lo_linkage));
1105
1106         if (dev != NULL) {
1107                 lu_ref_del_at(&dev->ld_reference,
1108                               o->lo_dev_ref , "lu_object", o);
1109                 lu_device_put(dev);
1110                 o->lo_dev = NULL;
1111         }
1112 }
1113 EXPORT_SYMBOL(lu_object_fini);
1114
1115 /**
1116  * Add object \a o as first layer of compound object \a h
1117  *
1118  * This is typically called by the ->ldo_object_alloc() method of top-level
1119  * device.
1120  */
1121 void lu_object_add_top(struct lu_object_header *h, struct lu_object *o)
1122 {
1123         cfs_list_move(&o->lo_linkage, &h->loh_layers);
1124 }
1125 EXPORT_SYMBOL(lu_object_add_top);
1126
1127 /**
1128  * Add object \a o as a layer of compound object, going after \a before.
1129  *
1130  * This is typically called by the ->ldo_object_alloc() method of \a
1131  * before->lo_dev.
1132  */
1133 void lu_object_add(struct lu_object *before, struct lu_object *o)
1134 {
1135         cfs_list_move(&o->lo_linkage, &before->lo_linkage);
1136 }
1137 EXPORT_SYMBOL(lu_object_add);
1138
1139 /**
1140  * Initialize compound object.
1141  */
1142 int lu_object_header_init(struct lu_object_header *h)
1143 {
1144         memset(h, 0, sizeof *h);
1145         cfs_atomic_set(&h->loh_ref, 1);
1146         CFS_INIT_HLIST_NODE(&h->loh_hash);
1147         CFS_INIT_LIST_HEAD(&h->loh_lru);
1148         CFS_INIT_LIST_HEAD(&h->loh_layers);
1149         lu_ref_init(&h->loh_reference);
1150         return 0;
1151 }
1152 EXPORT_SYMBOL(lu_object_header_init);
1153
1154 /**
1155  * Finalize compound object.
1156  */
1157 void lu_object_header_fini(struct lu_object_header *h)
1158 {
1159         LASSERT(cfs_list_empty(&h->loh_layers));
1160         LASSERT(cfs_list_empty(&h->loh_lru));
1161         LASSERT(cfs_hlist_unhashed(&h->loh_hash));
1162         lu_ref_fini(&h->loh_reference);
1163 }
1164 EXPORT_SYMBOL(lu_object_header_fini);
1165
1166 /**
1167  * Given a compound object, find its slice, corresponding to the device type
1168  * \a dtype.
1169  */
1170 struct lu_object *lu_object_locate(struct lu_object_header *h,
1171                                    const struct lu_device_type *dtype)
1172 {
1173         struct lu_object *o;
1174
1175         cfs_list_for_each_entry(o, &h->loh_layers, lo_linkage) {
1176                 if (o->lo_dev->ld_type == dtype)
1177                         return o;
1178         }
1179         return NULL;
1180 }
1181 EXPORT_SYMBOL(lu_object_locate);
1182
1183
1184
1185 /**
1186  * Finalize and free devices in the device stack.
1187  *
1188  * Finalize device stack by purging object cache, and calling
1189  * lu_device_type_operations::ldto_device_fini() and
1190  * lu_device_type_operations::ldto_device_free() on all devices in the stack.
1191  */
1192 void lu_stack_fini(const struct lu_env *env, struct lu_device *top)
1193 {
1194         struct lu_site   *site = top->ld_site;
1195         struct lu_device *scan;
1196         struct lu_device *next;
1197
1198         lu_site_purge(env, site, ~0);
1199         for (scan = top; scan != NULL; scan = next) {
1200                 next = scan->ld_type->ldt_ops->ldto_device_fini(env, scan);
1201                 lu_ref_del(&scan->ld_reference, "lu-stack", &lu_site_init);
1202                 lu_device_put(scan);
1203         }
1204
1205         /* purge again. */
1206         lu_site_purge(env, site, ~0);
1207
1208         if (!cfs_hash_is_empty(site->ls_obj_hash)) {
1209                 /*
1210                  * Uh-oh, objects still exist.
1211                  */
1212                 static DECLARE_LU_CDEBUG_PRINT_INFO(cookie, D_ERROR);
1213
1214                 lu_site_print(env, site, &cookie, lu_cdebug_printer);
1215         }
1216
1217         for (scan = top; scan != NULL; scan = next) {
1218                 const struct lu_device_type *ldt = scan->ld_type;
1219                 struct obd_type             *type;
1220
1221                 next = ldt->ldt_ops->ldto_device_free(env, scan);
1222                 type = ldt->ldt_obd_type;
1223                 if (type != NULL) {
1224                         type->typ_refcnt--;
1225                         class_put_type(type);
1226                 }
1227         }
1228 }
1229 EXPORT_SYMBOL(lu_stack_fini);
1230
1231 enum {
1232         /**
1233          * Maximal number of tld slots.
1234          */
1235         LU_CONTEXT_KEY_NR = 32
1236 };
1237
1238 static struct lu_context_key *lu_keys[LU_CONTEXT_KEY_NR] = { NULL, };
1239
1240 static cfs_spinlock_t lu_keys_guard = CFS_SPIN_LOCK_UNLOCKED;
1241
1242 /**
1243  * Global counter incremented whenever key is registered, unregistered,
1244  * revived or quiesced. This is used to void unnecessary calls to
1245  * lu_context_refill(). No locking is provided, as initialization and shutdown
1246  * are supposed to be externally serialized.
1247  */
1248 static unsigned key_set_version = 0;
1249
1250 /**
1251  * Register new key.
1252  */
1253 int lu_context_key_register(struct lu_context_key *key)
1254 {
1255         int result;
1256         int i;
1257
1258         LASSERT(key->lct_init != NULL);
1259         LASSERT(key->lct_fini != NULL);
1260         LASSERT(key->lct_tags != 0);
1261         LASSERT(key->lct_owner != NULL);
1262
1263         result = -ENFILE;
1264         cfs_spin_lock(&lu_keys_guard);
1265         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1266                 if (lu_keys[i] == NULL) {
1267                         key->lct_index = i;
1268                         cfs_atomic_set(&key->lct_used, 1);
1269                         lu_keys[i] = key;
1270                         lu_ref_init(&key->lct_reference);
1271                         result = 0;
1272                         ++key_set_version;
1273                         break;
1274                 }
1275         }
1276         cfs_spin_unlock(&lu_keys_guard);
1277         return result;
1278 }
1279 EXPORT_SYMBOL(lu_context_key_register);
1280
1281 static void key_fini(struct lu_context *ctx, int index)
1282 {
1283         if (ctx->lc_value != NULL && ctx->lc_value[index] != NULL) {
1284                 struct lu_context_key *key;
1285
1286                 key = lu_keys[index];
1287                 LASSERT(key != NULL);
1288                 LASSERT(key->lct_fini != NULL);
1289                 LASSERT(cfs_atomic_read(&key->lct_used) > 1);
1290
1291                 key->lct_fini(ctx, key, ctx->lc_value[index]);
1292                 lu_ref_del(&key->lct_reference, "ctx", ctx);
1293                 cfs_atomic_dec(&key->lct_used);
1294                 LASSERT(key->lct_owner != NULL);
1295                 if (!(ctx->lc_tags & LCT_NOREF)) {
1296                         LASSERT(cfs_module_refcount(key->lct_owner) > 0);
1297                         cfs_module_put(key->lct_owner);
1298                 }
1299                 ctx->lc_value[index] = NULL;
1300         }
1301 }
1302
1303 /**
1304  * Deregister key.
1305  */
1306 void lu_context_key_degister(struct lu_context_key *key)
1307 {
1308         LASSERT(cfs_atomic_read(&key->lct_used) >= 1);
1309         LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1310
1311         lu_context_key_quiesce(key);
1312
1313         ++key_set_version;
1314         cfs_spin_lock(&lu_keys_guard);
1315         key_fini(&lu_shrink_env.le_ctx, key->lct_index);
1316         if (lu_keys[key->lct_index]) {
1317                 lu_keys[key->lct_index] = NULL;
1318                 lu_ref_fini(&key->lct_reference);
1319         }
1320         cfs_spin_unlock(&lu_keys_guard);
1321
1322         LASSERTF(cfs_atomic_read(&key->lct_used) == 1,
1323                  "key has instances: %d\n",
1324                  cfs_atomic_read(&key->lct_used));
1325 }
1326 EXPORT_SYMBOL(lu_context_key_degister);
1327
1328 /**
1329  * Register a number of keys. This has to be called after all keys have been
1330  * initialized by a call to LU_CONTEXT_KEY_INIT().
1331  */
1332 int lu_context_key_register_many(struct lu_context_key *k, ...)
1333 {
1334         struct lu_context_key *key = k;
1335         va_list args;
1336         int result;
1337
1338         va_start(args, k);
1339         do {
1340                 result = lu_context_key_register(key);
1341                 if (result)
1342                         break;
1343                 key = va_arg(args, struct lu_context_key *);
1344         } while (key != NULL);
1345         va_end(args);
1346
1347         if (result != 0) {
1348                 va_start(args, k);
1349                 while (k != key) {
1350                         lu_context_key_degister(k);
1351                         k = va_arg(args, struct lu_context_key *);
1352                 }
1353                 va_end(args);
1354         }
1355
1356         return result;
1357 }
1358 EXPORT_SYMBOL(lu_context_key_register_many);
1359
1360 /**
1361  * De-register a number of keys. This is a dual to
1362  * lu_context_key_register_many().
1363  */
1364 void lu_context_key_degister_many(struct lu_context_key *k, ...)
1365 {
1366         va_list args;
1367
1368         va_start(args, k);
1369         do {
1370                 lu_context_key_degister(k);
1371                 k = va_arg(args, struct lu_context_key*);
1372         } while (k != NULL);
1373         va_end(args);
1374 }
1375 EXPORT_SYMBOL(lu_context_key_degister_many);
1376
1377 /**
1378  * Revive a number of keys.
1379  */
1380 void lu_context_key_revive_many(struct lu_context_key *k, ...)
1381 {
1382         va_list args;
1383
1384         va_start(args, k);
1385         do {
1386                 lu_context_key_revive(k);
1387                 k = va_arg(args, struct lu_context_key*);
1388         } while (k != NULL);
1389         va_end(args);
1390 }
1391 EXPORT_SYMBOL(lu_context_key_revive_many);
1392
1393 /**
1394  * Quiescent a number of keys.
1395  */
1396 void lu_context_key_quiesce_many(struct lu_context_key *k, ...)
1397 {
1398         va_list args;
1399
1400         va_start(args, k);
1401         do {
1402                 lu_context_key_quiesce(k);
1403                 k = va_arg(args, struct lu_context_key*);
1404         } while (k != NULL);
1405         va_end(args);
1406 }
1407 EXPORT_SYMBOL(lu_context_key_quiesce_many);
1408
1409 /**
1410  * Return value associated with key \a key in context \a ctx.
1411  */
1412 void *lu_context_key_get(const struct lu_context *ctx,
1413                          const struct lu_context_key *key)
1414 {
1415         LINVRNT(ctx->lc_state == LCS_ENTERED);
1416         LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1417         LASSERT(lu_keys[key->lct_index] == key);
1418         return ctx->lc_value[key->lct_index];
1419 }
1420 EXPORT_SYMBOL(lu_context_key_get);
1421
1422 /**
1423  * List of remembered contexts. XXX document me.
1424  */
1425 static CFS_LIST_HEAD(lu_context_remembered);
1426
1427 /**
1428  * Destroy \a key in all remembered contexts. This is used to destroy key
1429  * values in "shared" contexts (like service threads), when a module owning
1430  * the key is about to be unloaded.
1431  */
1432 void lu_context_key_quiesce(struct lu_context_key *key)
1433 {
1434         struct lu_context *ctx;
1435         extern unsigned cl_env_cache_purge(unsigned nr);
1436
1437         if (!(key->lct_tags & LCT_QUIESCENT)) {
1438                 /*
1439                  * XXX layering violation.
1440                  */
1441                 cl_env_cache_purge(~0);
1442                 key->lct_tags |= LCT_QUIESCENT;
1443                 /*
1444                  * XXX memory barrier has to go here.
1445                  */
1446                 cfs_spin_lock(&lu_keys_guard);
1447                 cfs_list_for_each_entry(ctx, &lu_context_remembered,
1448                                         lc_remember)
1449                         key_fini(ctx, key->lct_index);
1450                 cfs_spin_unlock(&lu_keys_guard);
1451                 ++key_set_version;
1452         }
1453 }
1454 EXPORT_SYMBOL(lu_context_key_quiesce);
1455
1456 void lu_context_key_revive(struct lu_context_key *key)
1457 {
1458         key->lct_tags &= ~LCT_QUIESCENT;
1459         ++key_set_version;
1460 }
1461 EXPORT_SYMBOL(lu_context_key_revive);
1462
1463 static void keys_fini(struct lu_context *ctx)
1464 {
1465         int i;
1466
1467         cfs_spin_lock(&lu_keys_guard);
1468         if (ctx->lc_value != NULL) {
1469                 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i)
1470                         key_fini(ctx, i);
1471                 OBD_FREE(ctx->lc_value,
1472                          ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1473                 ctx->lc_value = NULL;
1474         }
1475         cfs_spin_unlock(&lu_keys_guard);
1476 }
1477
1478 static int keys_fill(struct lu_context *ctx)
1479 {
1480         int i;
1481
1482         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1483                 struct lu_context_key *key;
1484
1485                 key = lu_keys[i];
1486                 if (ctx->lc_value[i] == NULL && key != NULL &&
1487                     (key->lct_tags & ctx->lc_tags) &&
1488                     /*
1489                      * Don't create values for a LCT_QUIESCENT key, as this
1490                      * will pin module owning a key.
1491                      */
1492                     !(key->lct_tags & LCT_QUIESCENT)) {
1493                         void *value;
1494
1495                         LINVRNT(key->lct_init != NULL);
1496                         LINVRNT(key->lct_index == i);
1497
1498                         value = key->lct_init(ctx, key);
1499                         if (unlikely(IS_ERR(value)))
1500                                 return PTR_ERR(value);
1501
1502                         LASSERT(key->lct_owner != NULL);
1503                         if (!(ctx->lc_tags & LCT_NOREF))
1504                                 cfs_try_module_get(key->lct_owner);
1505                         lu_ref_add_atomic(&key->lct_reference, "ctx", ctx);
1506                         cfs_atomic_inc(&key->lct_used);
1507                         /*
1508                          * This is the only place in the code, where an
1509                          * element of ctx->lc_value[] array is set to non-NULL
1510                          * value.
1511                          */
1512                         ctx->lc_value[i] = value;
1513                         if (key->lct_exit != NULL)
1514                                 ctx->lc_tags |= LCT_HAS_EXIT;
1515                 }
1516                 ctx->lc_version = key_set_version;
1517         }
1518         return 0;
1519 }
1520
1521 static int keys_init(struct lu_context *ctx)
1522 {
1523         int result;
1524
1525         OBD_ALLOC(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1526         if (likely(ctx->lc_value != NULL))
1527                 result = keys_fill(ctx);
1528         else
1529                 result = -ENOMEM;
1530
1531         if (result != 0)
1532                 keys_fini(ctx);
1533         return result;
1534 }
1535
1536 /**
1537  * Initialize context data-structure. Create values for all keys.
1538  */
1539 int lu_context_init(struct lu_context *ctx, __u32 tags)
1540 {
1541         memset(ctx, 0, sizeof *ctx);
1542         ctx->lc_state = LCS_INITIALIZED;
1543         ctx->lc_tags = tags;
1544         if (tags & LCT_REMEMBER) {
1545                 cfs_spin_lock(&lu_keys_guard);
1546                 cfs_list_add(&ctx->lc_remember, &lu_context_remembered);
1547                 cfs_spin_unlock(&lu_keys_guard);
1548         } else
1549                 CFS_INIT_LIST_HEAD(&ctx->lc_remember);
1550         return keys_init(ctx);
1551 }
1552 EXPORT_SYMBOL(lu_context_init);
1553
1554 /**
1555  * Finalize context data-structure. Destroy key values.
1556  */
1557 void lu_context_fini(struct lu_context *ctx)
1558 {
1559         LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1560         ctx->lc_state = LCS_FINALIZED;
1561         keys_fini(ctx);
1562         cfs_spin_lock(&lu_keys_guard);
1563         cfs_list_del_init(&ctx->lc_remember);
1564         cfs_spin_unlock(&lu_keys_guard);
1565 }
1566 EXPORT_SYMBOL(lu_context_fini);
1567
1568 /**
1569  * Called before entering context.
1570  */
1571 void lu_context_enter(struct lu_context *ctx)
1572 {
1573         LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1574         ctx->lc_state = LCS_ENTERED;
1575 }
1576 EXPORT_SYMBOL(lu_context_enter);
1577
1578 /**
1579  * Called after exiting from \a ctx
1580  */
1581 void lu_context_exit(struct lu_context *ctx)
1582 {
1583         int i;
1584
1585         LINVRNT(ctx->lc_state == LCS_ENTERED);
1586         ctx->lc_state = LCS_LEFT;
1587         if (ctx->lc_tags & LCT_HAS_EXIT && ctx->lc_value != NULL) {
1588                 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1589                         if (ctx->lc_value[i] != NULL) {
1590                                 struct lu_context_key *key;
1591
1592                                 key = lu_keys[i];
1593                                 LASSERT(key != NULL);
1594                                 if (key->lct_exit != NULL)
1595                                         key->lct_exit(ctx,
1596                                                       key, ctx->lc_value[i]);
1597                         }
1598                 }
1599         }
1600 }
1601 EXPORT_SYMBOL(lu_context_exit);
1602
1603 /**
1604  * Allocate for context all missing keys that were registered after context
1605  * creation.
1606  */
1607 int lu_context_refill(struct lu_context *ctx)
1608 {
1609         LINVRNT(ctx->lc_value != NULL);
1610         return ctx->lc_version == key_set_version ? 0 : keys_fill(ctx);
1611 }
1612 EXPORT_SYMBOL(lu_context_refill);
1613
1614 /**
1615  * lu_ctx_tags/lu_ses_tags will be updated if there are new types of
1616  * obd being added. Currently, this is only used on client side, specifically
1617  * for echo device client, for other stack (like ptlrpc threads), context are
1618  * predefined when the lu_device type are registered, during the module probe
1619  * phase.
1620  */
1621 __u32 lu_context_tags_default = 0;
1622 __u32 lu_session_tags_default = 0;
1623
1624 void lu_context_tags_update(__u32 tags)
1625 {
1626         cfs_spin_lock(&lu_keys_guard);
1627         lu_context_tags_default |= tags;
1628         key_set_version ++;
1629         cfs_spin_unlock(&lu_keys_guard);
1630 }
1631 EXPORT_SYMBOL(lu_context_tags_update);
1632
1633 void lu_context_tags_clear(__u32 tags)
1634 {
1635         cfs_spin_lock(&lu_keys_guard);
1636         lu_context_tags_default &= ~tags;
1637         key_set_version ++;
1638         cfs_spin_unlock(&lu_keys_guard);
1639 }
1640 EXPORT_SYMBOL(lu_context_tags_clear);
1641
1642 void lu_session_tags_update(__u32 tags)
1643 {
1644         cfs_spin_lock(&lu_keys_guard);
1645         lu_session_tags_default |= tags;
1646         key_set_version ++;
1647         cfs_spin_unlock(&lu_keys_guard);
1648 }
1649 EXPORT_SYMBOL(lu_session_tags_update);
1650
1651 void lu_session_tags_clear(__u32 tags)
1652 {
1653         cfs_spin_lock(&lu_keys_guard);
1654         lu_session_tags_default &= ~tags;
1655         key_set_version ++;
1656         cfs_spin_unlock(&lu_keys_guard);
1657 }
1658 EXPORT_SYMBOL(lu_session_tags_clear);
1659
1660 int lu_env_init(struct lu_env *env, __u32 tags)
1661 {
1662         int result;
1663
1664         env->le_ses = NULL;
1665         result = lu_context_init(&env->le_ctx, tags);
1666         if (likely(result == 0))
1667                 lu_context_enter(&env->le_ctx);
1668         return result;
1669 }
1670 EXPORT_SYMBOL(lu_env_init);
1671
1672 void lu_env_fini(struct lu_env *env)
1673 {
1674         lu_context_exit(&env->le_ctx);
1675         lu_context_fini(&env->le_ctx);
1676         env->le_ses = NULL;
1677 }
1678 EXPORT_SYMBOL(lu_env_fini);
1679
1680 int lu_env_refill(struct lu_env *env)
1681 {
1682         int result;
1683
1684         result = lu_context_refill(&env->le_ctx);
1685         if (result == 0 && env->le_ses != NULL)
1686                 result = lu_context_refill(env->le_ses);
1687         return result;
1688 }
1689 EXPORT_SYMBOL(lu_env_refill);
1690
1691 /**
1692  * Currently, this API will only be used by echo client.
1693  * Because echo client and normal lustre client will share
1694  * same cl_env cache. So echo client needs to refresh
1695  * the env context after it get one from the cache, especially
1696  * when normal client and echo client co-exist in the same client.
1697  */
1698 int lu_env_refill_by_tags(struct lu_env *env, __u32 ctags,
1699                           __u32 stags)
1700 {
1701         int    result;
1702
1703         if ((env->le_ctx.lc_tags & ctags) != ctags) {
1704                 env->le_ctx.lc_version = 0;
1705                 env->le_ctx.lc_tags |= ctags;
1706         }
1707
1708         if (env->le_ses && (env->le_ses->lc_tags & stags) != stags) {
1709                 env->le_ses->lc_version = 0;
1710                 env->le_ses->lc_tags |= stags;
1711         }
1712
1713         result = lu_env_refill(env);
1714
1715         return result;
1716 }
1717 EXPORT_SYMBOL(lu_env_refill_by_tags);
1718
1719 static struct cfs_shrinker *lu_site_shrinker = NULL;
1720
1721 typedef struct lu_site_stats{
1722         unsigned        lss_populated;
1723         unsigned        lss_max_search;
1724         unsigned        lss_total;
1725         unsigned        lss_busy;
1726 } lu_site_stats_t;
1727
1728 static void lu_site_stats_get(cfs_hash_t *hs,
1729                               lu_site_stats_t *stats, int populated)
1730 {
1731         cfs_hash_bd_t bd;
1732         int           i;
1733
1734         cfs_hash_for_each_bucket(hs, &bd, i) {
1735                 struct lu_site_bkt_data *bkt = cfs_hash_bd_extra_get(hs, &bd);
1736                 cfs_hlist_head_t        *hhead;
1737
1738                 cfs_hash_bd_lock(hs, &bd, 1);
1739                 stats->lss_busy  += bkt->lsb_busy;
1740                 stats->lss_total += cfs_hash_bd_count_get(&bd);
1741                 stats->lss_max_search = max((int)stats->lss_max_search,
1742                                             cfs_hash_bd_depmax_get(&bd));
1743                 if (!populated) {
1744                         cfs_hash_bd_unlock(hs, &bd, 1);
1745                         continue;
1746                 }
1747
1748                 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
1749                         if (!cfs_hlist_empty(hhead))
1750                                 stats->lss_populated++;
1751                 }
1752                 cfs_hash_bd_unlock(hs, &bd, 1);
1753         }
1754 }
1755
1756 #ifdef __KERNEL__
1757
1758 static int lu_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
1759 {
1760         lu_site_stats_t stats;
1761         struct lu_site *s;
1762         struct lu_site *tmp;
1763         int cached = 0;
1764         int remain = shrink_param(sc, nr_to_scan);
1765         CFS_LIST_HEAD(splice);
1766
1767         if (remain != 0) {
1768                 if (!(shrink_param(sc, gfp_mask) & __GFP_FS))
1769                         return -1;
1770                 CDEBUG(D_INODE, "Shrink %d objects\n", remain);
1771         }
1772
1773         cfs_down(&lu_sites_guard);
1774         cfs_list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
1775                 if (shrink_param(sc, nr_to_scan) != 0) {
1776                         remain = lu_site_purge(&lu_shrink_env, s, remain);
1777                         /*
1778                          * Move just shrunk site to the tail of site list to
1779                          * assure shrinking fairness.
1780                          */
1781                         cfs_list_move_tail(&s->ls_linkage, &splice);
1782                 }
1783
1784                 memset(&stats, 0, sizeof(stats));
1785                 lu_site_stats_get(s->ls_obj_hash, &stats, 0);
1786                 cached += stats.lss_total - stats.lss_busy;
1787                 if (shrink_param(sc, nr_to_scan) && remain <= 0)
1788                         break;
1789         }
1790         cfs_list_splice(&splice, lu_sites.prev);
1791         cfs_up(&lu_sites_guard);
1792
1793         cached = (cached / 100) * sysctl_vfs_cache_pressure;
1794         if (shrink_param(sc, nr_to_scan) == 0)
1795                 CDEBUG(D_INODE, "%d objects cached\n", cached);
1796         return cached;
1797 }
1798
1799 /*
1800  * Debugging stuff.
1801  */
1802
1803 /**
1804  * Environment to be used in debugger, contains all tags.
1805  */
1806 struct lu_env lu_debugging_env;
1807
1808 /**
1809  * Debugging printer function using printk().
1810  */
1811 int lu_printk_printer(const struct lu_env *env,
1812                       void *unused, const char *format, ...)
1813 {
1814         va_list args;
1815
1816         va_start(args, format);
1817         vprintk(format, args);
1818         va_end(args);
1819         return 0;
1820 }
1821
1822 void lu_debugging_setup(void)
1823 {
1824         lu_env_init(&lu_debugging_env, ~0);
1825 }
1826
1827 void lu_context_keys_dump(void)
1828 {
1829         int i;
1830
1831         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1832                 struct lu_context_key *key;
1833
1834                 key = lu_keys[i];
1835                 if (key != NULL) {
1836                         CERROR("[%d]: %p %x (%p,%p,%p) %d %d \"%s\"@%p\n",
1837                                i, key, key->lct_tags,
1838                                key->lct_init, key->lct_fini, key->lct_exit,
1839                                key->lct_index, cfs_atomic_read(&key->lct_used),
1840                                key->lct_owner ? key->lct_owner->name : "",
1841                                key->lct_owner);
1842                         lu_ref_print(&key->lct_reference);
1843                 }
1844         }
1845 }
1846 EXPORT_SYMBOL(lu_context_keys_dump);
1847 #else  /* !__KERNEL__ */
1848 static int lu_cache_shrink(int nr, unsigned int gfp_mask)
1849 {
1850         return 0;
1851 }
1852 #endif /* __KERNEL__ */
1853
1854 int  cl_global_init(void);
1855 void cl_global_fini(void);
1856 int  lu_ref_global_init(void);
1857 void lu_ref_global_fini(void);
1858
1859 int dt_global_init(void);
1860 void dt_global_fini(void);
1861
1862 int llo_global_init(void);
1863 void llo_global_fini(void);
1864
1865 /**
1866  * Initialization of global lu_* data.
1867  */
1868 int lu_global_init(void)
1869 {
1870         int result;
1871
1872         CDEBUG(D_INFO, "Lustre LU module (%p).\n", &lu_keys);
1873
1874         result = lu_ref_global_init();
1875         if (result != 0)
1876                 return result;
1877
1878         LU_CONTEXT_KEY_INIT(&lu_global_key);
1879         result = lu_context_key_register(&lu_global_key);
1880         if (result != 0)
1881                 return result;
1882         /*
1883          * At this level, we don't know what tags are needed, so allocate them
1884          * conservatively. This should not be too bad, because this
1885          * environment is global.
1886          */
1887         cfs_down(&lu_sites_guard);
1888         result = lu_env_init(&lu_shrink_env, LCT_SHRINKER);
1889         cfs_up(&lu_sites_guard);
1890         if (result != 0)
1891                 return result;
1892
1893         /*
1894          * seeks estimation: 3 seeks to read a record from oi, one to read
1895          * inode, one for ea. Unfortunately setting this high value results in
1896          * lu_object/inode cache consuming all the memory.
1897          */
1898         lu_site_shrinker = cfs_set_shrinker(CFS_DEFAULT_SEEKS, lu_cache_shrink);
1899         if (lu_site_shrinker == NULL)
1900                 return -ENOMEM;
1901
1902         result = lu_time_global_init();
1903         if (result)
1904                 GOTO(out, result);
1905
1906 #ifdef __KERNEL__
1907         result = dt_global_init();
1908         if (result)
1909                 GOTO(out, result);
1910
1911         result = llo_global_init();
1912         if (result)
1913                 GOTO(out, result);
1914 #endif
1915         result = cl_global_init();
1916 out:
1917
1918         return result;
1919 }
1920
1921 /**
1922  * Dual to lu_global_init().
1923  */
1924 void lu_global_fini(void)
1925 {
1926         cl_global_fini();
1927 #ifdef __KERNEL__
1928         llo_global_fini();
1929         dt_global_fini();
1930 #endif
1931         lu_time_global_fini();
1932         if (lu_site_shrinker != NULL) {
1933                 cfs_remove_shrinker(lu_site_shrinker);
1934                 lu_site_shrinker = NULL;
1935         }
1936
1937         lu_context_key_degister(&lu_global_key);
1938
1939         /*
1940          * Tear shrinker environment down _after_ de-registering
1941          * lu_global_key, because the latter has a value in the former.
1942          */
1943         cfs_down(&lu_sites_guard);
1944         lu_env_fini(&lu_shrink_env);
1945         cfs_up(&lu_sites_guard);
1946
1947         lu_ref_global_fini();
1948 }
1949
1950 struct lu_buf LU_BUF_NULL = {
1951         .lb_buf = NULL,
1952         .lb_len = 0
1953 };
1954 EXPORT_SYMBOL(LU_BUF_NULL);
1955
1956 static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx)
1957 {
1958 #ifdef LPROCFS
1959         struct lprocfs_counter ret;
1960
1961         lprocfs_stats_collect(stats, idx, &ret);
1962         return (__u32)ret.lc_count;
1963 #else
1964         return 0;
1965 #endif
1966 }
1967
1968 /**
1969  * Output site statistical counters into a buffer. Suitable for
1970  * lprocfs_rd_*()-style functions.
1971  */
1972 int lu_site_stats_print(const struct lu_site *s, char *page, int count)
1973 {
1974         lu_site_stats_t stats;
1975
1976         memset(&stats, 0, sizeof(stats));
1977         lu_site_stats_get(s->ls_obj_hash, &stats, 1);
1978
1979         return snprintf(page, count, "%d/%d %d/%d %d %d %d %d %d %d %d\n",
1980                         stats.lss_busy,
1981                         stats.lss_total,
1982                         stats.lss_populated,
1983                         CFS_HASH_NHLIST(s->ls_obj_hash),
1984                         stats.lss_max_search,
1985                         ls_stats_read(s->ls_stats, LU_SS_CREATED),
1986                         ls_stats_read(s->ls_stats, LU_SS_CACHE_HIT),
1987                         ls_stats_read(s->ls_stats, LU_SS_CACHE_MISS),
1988                         ls_stats_read(s->ls_stats, LU_SS_CACHE_RACE),
1989                         ls_stats_read(s->ls_stats, LU_SS_CACHE_DEATH_RACE),
1990                         ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED));
1991 }
1992 EXPORT_SYMBOL(lu_site_stats_print);
1993
1994 const char *lu_time_names[LU_TIME_NR] = {
1995         [LU_TIME_FIND_LOOKUP] = "find_lookup",
1996         [LU_TIME_FIND_ALLOC]  = "find_alloc",
1997         [LU_TIME_FIND_INSERT] = "find_insert"
1998 };
1999 EXPORT_SYMBOL(lu_time_names);
2000
2001 /**
2002  * Helper function to initialize a number of kmem slab caches at once.
2003  */
2004 int lu_kmem_init(struct lu_kmem_descr *caches)
2005 {
2006         int result;
2007         struct lu_kmem_descr *iter = caches;
2008
2009         for (result = 0; iter->ckd_cache != NULL; ++iter) {
2010                 *iter->ckd_cache = cfs_mem_cache_create(iter->ckd_name,
2011                                                         iter->ckd_size,
2012                                                         0, 0);
2013                 if (*iter->ckd_cache == NULL) {
2014                         result = -ENOMEM;
2015                         /* free all previously allocated caches */
2016                         lu_kmem_fini(caches);
2017                         break;
2018                 }
2019         }
2020         return result;
2021 }
2022 EXPORT_SYMBOL(lu_kmem_init);
2023
2024 /**
2025  * Helper function to finalize a number of kmem slab cached at once. Dual to
2026  * lu_kmem_init().
2027  */
2028 void lu_kmem_fini(struct lu_kmem_descr *caches)
2029 {
2030         int rc;
2031
2032         for (; caches->ckd_cache != NULL; ++caches) {
2033                 if (*caches->ckd_cache != NULL) {
2034                         rc = cfs_mem_cache_destroy(*caches->ckd_cache);
2035                         LASSERTF(rc == 0, "couldn't destroy %s slab\n",
2036                                  caches->ckd_name);
2037                         *caches->ckd_cache = NULL;
2038                 }
2039         }
2040 }
2041 EXPORT_SYMBOL(lu_kmem_fini);