lustre/obdclass/lu_object.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Whamcloud, Inc.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/obdclass/lu_object.c
  37  *
  38  * Lustre Object.
  39  * These are the only exported functions, they provide some generic
  40  * infrastructure for managing object devices
  41  *
  42  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  43  */
  44
  45 #define DEBUG_SUBSYSTEM S_CLASS
  46 #ifndef EXPORT_SYMTAB
  47 # define EXPORT_SYMTAB
  48 #endif
  49
  50 #include <libcfs/libcfs.h>
  51
  52 #ifdef __KERNEL__
  53 # include <linux/module.h>
  54 #endif
  55
  56 /* hash_long() */
  57 #include <libcfs/libcfs_hash.h>
  58 #include <obd_class.h>
  59 #include <obd_support.h>
  60 #include <lustre_disk.h>
  61 #include <lustre_fid.h>
  62 #include <lu_object.h>
  63 #include <libcfs/list.h>
  64 /* lu_time_global_{init,fini}() */
  65 #include <lu_time.h>
  66
  67 static void lu_object_free(const struct lu_env *env, struct lu_object *o);
  68
  69 /**
  70  * Decrease reference counter on object. If last reference is freed, return
  71  * object to the cache, unless lu_object_is_dying(o) holds. In the latter
  72  * case, free object immediately.
  73  */
  74 void lu_object_put(const struct lu_env *env, struct lu_object *o)
  75 {
  76         struct lu_site_bkt_data *bkt;
  77         struct lu_object_header *top;
  78         struct lu_site          *site;
  79         struct lu_object        *orig;
  80         cfs_hash_bd_t            bd;
  81
  82         top  = o->lo_header;
  83         site = o->lo_dev->ld_site;
  84         orig = o;
  85
  86         cfs_hash_bd_get(site->ls_obj_hash, &top->loh_fid, &bd);
  87         bkt = cfs_hash_bd_extra_get(site->ls_obj_hash, &bd);
  88
  89         if (!cfs_hash_bd_dec_and_lock(site->ls_obj_hash, &bd, &top->loh_ref)) {
  90                 if (lu_object_is_dying(top)) {
  91
  92                         /*
  93                          * somebody may be waiting for this, currently only
  94                          * used for cl_object, see cl_object_put_last().
  95                          */
  96                         cfs_waitq_broadcast(&bkt->lsb_marche_funebre);
  97                 }
  98                 return;
  99         }
 100
 101         LASSERT(bkt->lsb_busy > 0);
 102         bkt->lsb_busy--;
 103         /*
 104          * When last reference is released, iterate over object
 105          * layers, and notify them that object is no longer busy.
 106          */
 107         cfs_list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
 108                 if (o->lo_ops->loo_object_release != NULL)
 109                         o->lo_ops->loo_object_release(env, o);
 110         }
 111
 112         if (!lu_object_is_dying(top)) {
 113                 LASSERT(cfs_list_empty(&top->loh_lru));
 114                 cfs_list_add_tail(&top->loh_lru, &bkt->lsb_lru);
 115                 cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
 116                 return;
 117         }
 118
 119         /*
 120          * If object is dying (will not be cached), removed it
 121          * from hash table and LRU.
 122          *
 123          * This is done with hash table and LRU lists locked. As the only
 124          * way to acquire first reference to previously unreferenced
 125          * object is through hash-table lookup (lu_object_find()),
 126          * or LRU scanning (lu_site_purge()), that are done under hash-table
 127          * and LRU lock, no race with concurrent object lookup is possible
 128          * and we can safely destroy object below.
 129          */
 130         cfs_hash_bd_del_locked(site->ls_obj_hash, &bd, &top->loh_hash);
 131         cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
 132         /*
 133          * Object was already removed from hash and lru above, can
 134          * kill it.
 135          */
 136         lu_object_free(env, orig);
 137 }
 138 EXPORT_SYMBOL(lu_object_put);
 139
 140 /**
 141  * Allocate new object.
 142  *
 143  * This follows object creation protocol, described in the comment within
 144  * struct lu_device_operations definition.
 145  */
 146 static struct lu_object *lu_object_alloc(const struct lu_env *env,
 147                                          struct lu_device *dev,
 148                                          const struct lu_fid *f,
 149                                          const struct lu_object_conf *conf)
 150 {
 151         struct lu_object *scan;
 152         struct lu_object *top;
 153         cfs_list_t *layers;
 154         int clean;
 155         int result;
 156         ENTRY;
 157
 158         /*
 159          * Create top-level object slice. This will also create
 160          * lu_object_header.
 161          */
 162         top = dev->ld_ops->ldo_object_alloc(env, NULL, dev);
 163         if (top == NULL)
 164                 RETURN(ERR_PTR(-ENOMEM));
 165         /*
 166          * This is the only place where object fid is assigned. It's constant
 167          * after this point.
 168          */
 169         LASSERT(fid_is_igif(f) || fid_ver(f) == 0);
 170         top->lo_header->loh_fid = *f;
 171         layers = &top->lo_header->loh_layers;
 172         do {
 173                 /*
 174                  * Call ->loo_object_init() repeatedly, until no more new
 175                  * object slices are created.
 176                  */
 177                 clean = 1;
 178                 cfs_list_for_each_entry(scan, layers, lo_linkage) {
 179                         if (scan->lo_flags & LU_OBJECT_ALLOCATED)
 180                                 continue;
 181                         clean = 0;
 182                         scan->lo_header = top->lo_header;
 183                         result = scan->lo_ops->loo_object_init(env, scan, conf);
 184                         if (result != 0) {
 185                                 lu_object_free(env, top);
 186                                 RETURN(ERR_PTR(result));
 187                         }
 188                         scan->lo_flags |= LU_OBJECT_ALLOCATED;
 189                 }
 190         } while (!clean);
 191
 192         cfs_list_for_each_entry_reverse(scan, layers, lo_linkage) {
 193                 if (scan->lo_ops->loo_object_start != NULL) {
 194                         result = scan->lo_ops->loo_object_start(env, scan);
 195                         if (result != 0) {
 196                                 lu_object_free(env, top);
 197                                 RETURN(ERR_PTR(result));
 198                         }
 199                 }
 200         }
 201
 202         lprocfs_counter_incr(dev->ld_site->ls_stats, LU_SS_CREATED);
 203         RETURN(top);
 204 }
 205
 206 /**
 207  * Free an object.
 208  */
 209 static void lu_object_free(const struct lu_env *env, struct lu_object *o)
 210 {
 211         struct lu_site_bkt_data *bkt;
 212         struct lu_site          *site;
 213         struct lu_object        *scan;
 214         cfs_list_t              *layers;
 215         cfs_list_t               splice;
 216
 217         site   = o->lo_dev->ld_site;
 218         layers = &o->lo_header->loh_layers;
 219         bkt    = lu_site_bkt_from_fid(site, &o->lo_header->loh_fid);
 220         /*
 221          * First call ->loo_object_delete() method to release all resources.
 222          */
 223         cfs_list_for_each_entry_reverse(scan, layers, lo_linkage) {
 224                 if (scan->lo_ops->loo_object_delete != NULL)
 225                         scan->lo_ops->loo_object_delete(env, scan);
 226         }
 227
 228         /*
 229          * Then, splice object layers into stand-alone list, and call
 230          * ->loo_object_free() on all layers to free memory. Splice is
 231          * necessary, because lu_object_header is freed together with the
 232          * top-level slice.
 233          */
 234         CFS_INIT_LIST_HEAD(&splice);
 235         cfs_list_splice_init(layers, &splice);
 236         while (!cfs_list_empty(&splice)) {
 237                 /*
 238                  * Free layers in bottom-to-top order, so that object header
 239                  * lives as long as possible and ->loo_object_free() methods
 240                  * can look at its contents.
 241                  */
 242                 o = container_of0(splice.prev, struct lu_object, lo_linkage);
 243                 cfs_list_del_init(&o->lo_linkage);
 244                 LASSERT(o->lo_ops->loo_object_free != NULL);
 245                 o->lo_ops->loo_object_free(env, o);
 246         }
 247
 248         if (cfs_waitq_active(&bkt->lsb_marche_funebre))
 249                 cfs_waitq_broadcast(&bkt->lsb_marche_funebre);
 250 }
 251
 252 /**
 253  * Free \a nr objects from the cold end of the site LRU list.
 254  */
 255 int lu_site_purge(const struct lu_env *env, struct lu_site *s, int nr)
 256 {
 257         struct lu_object_header *h;
 258         struct lu_object_header *temp;
 259         struct lu_site_bkt_data *bkt;
 260         cfs_hash_bd_t            bd;
 261         cfs_hash_bd_t            bd2;
 262         cfs_list_t               dispose;
 263         int                      did_sth;
 264         int                      start;
 265         int                      count;
 266         int                      bnr;
 267         int                      i;
 268
 269         CFS_INIT_LIST_HEAD(&dispose);
 270         /*
 271          * Under LRU list lock, scan LRU list and move unreferenced objects to
 272          * the dispose list, removing them from LRU and hash table.
 273          */
 274         start = s->ls_purge_start;
 275         bnr = (nr == ~0) ? -1 : nr / CFS_HASH_NBKT(s->ls_obj_hash) + 1;
 276  again:
 277         did_sth = 0;
 278         cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
 279                 if (i < start)
 280                         continue;
 281                 count = bnr;
 282                 cfs_hash_bd_lock(s->ls_obj_hash, &bd, 1);
 283                 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
 284
 285                 cfs_list_for_each_entry_safe(h, temp, &bkt->lsb_lru, loh_lru) {
 286                         LASSERT(cfs_atomic_read(&h->loh_ref) == 0);
 287
 288                         cfs_hash_bd_get(s->ls_obj_hash, &h->loh_fid, &bd2);
 289                         LASSERT(bd.bd_bucket == bd2.bd_bucket);
 290
 291                         cfs_hash_bd_del_locked(s->ls_obj_hash,
 292                                                &bd2, &h->loh_hash);
 293                         cfs_list_move(&h->loh_lru, &dispose);
 294                         if (did_sth == 0)
 295                                 did_sth = 1;
 296
 297                         if (nr != ~0 && --nr == 0)
 298                                 break;
 299
 300                         if (count > 0 && --count == 0)
 301                                 break;
 302
 303                 }
 304                 cfs_hash_bd_unlock(s->ls_obj_hash, &bd, 1);
 305                 cfs_cond_resched();
 306                 /*
 307                  * Free everything on the dispose list. This is safe against
 308                  * races due to the reasons described in lu_object_put().
 309                  */
 310                 while (!cfs_list_empty(&dispose)) {
 311                         h = container_of0(dispose.next,
 312                                           struct lu_object_header, loh_lru);
 313                         cfs_list_del_init(&h->loh_lru);
 314                         lu_object_free(env, lu_object_top(h));
 315                         lprocfs_counter_incr(s->ls_stats, LU_SS_LRU_PURGED);
 316                 }
 317
 318                 if (nr == 0)
 319                         break;
 320         }
 321
 322         if (nr != 0 && did_sth && start != 0) {
 323                 start = 0; /* restart from the first bucket */
 324                 goto again;
 325         }
 326         /* race on s->ls_purge_start, but nobody cares */
 327         s->ls_purge_start = i % CFS_HASH_NBKT(s->ls_obj_hash);
 328
 329         return nr;
 330 }
 331 EXPORT_SYMBOL(lu_site_purge);
 332
 333 /*
 334  * Object printing.
 335  *
 336  * Code below has to jump through certain loops to output object description
 337  * into libcfs_debug_msg-based log. The problem is that lu_object_print()
 338  * composes object description from strings that are parts of _lines_ of
 339  * output (i.e., strings that are not terminated by newline). This doesn't fit
 340  * very well into libcfs_debug_msg() interface that assumes that each message
 341  * supplied to it is a self-contained output line.
 342  *
 343  * To work around this, strings are collected in a temporary buffer
 344  * (implemented as a value of lu_cdebug_key key), until terminating newline
 345  * character is detected.
 346  *
 347  */
 348
 349 enum {
 350         /**
 351          * Maximal line size.
 352          *
 353          * XXX overflow is not handled correctly.
 354          */
 355         LU_CDEBUG_LINE = 512
 356 };
 357
 358 struct lu_cdebug_data {
 359         /**
 360          * Temporary buffer.
 361          */
 362         char lck_area[LU_CDEBUG_LINE];
 363 };
 364
 365 /* context key constructor/destructor: lu_global_key_init, lu_global_key_fini */
 366 LU_KEY_INIT_FINI(lu_global, struct lu_cdebug_data);
 367
 368 /**
 369  * Key, holding temporary buffer. This key is registered very early by
 370  * lu_global_init().
 371  */
 372 struct lu_context_key lu_global_key = {
 373         .lct_tags = LCT_MD_THREAD | LCT_DT_THREAD |
 374                     LCT_MG_THREAD | LCT_CL_THREAD,
 375         .lct_init = lu_global_key_init,
 376         .lct_fini = lu_global_key_fini
 377 };
 378
 379 /**
 380  * Printer function emitting messages through libcfs_debug_msg().
 381  */
 382 int lu_cdebug_printer(const struct lu_env *env,
 383                       void *cookie, const char *format, ...)
 384 {
 385         struct libcfs_debug_msg_data *msgdata = cookie;
 386         struct lu_cdebug_data        *key;
 387         int used;
 388         int complete;
 389         va_list args;
 390
 391         va_start(args, format);
 392
 393         key = lu_context_key_get(&env->le_ctx, &lu_global_key);
 394         LASSERT(key != NULL);
 395
 396         used = strlen(key->lck_area);
 397         complete = format[strlen(format) - 1] == '\n';
 398         /*
 399          * Append new chunk to the buffer.
 400          */
 401         vsnprintf(key->lck_area + used,
 402                   ARRAY_SIZE(key->lck_area) - used, format, args);
 403         if (complete) {
 404                 if (cfs_cdebug_show(msgdata->msg_mask, msgdata->msg_subsys))
 405                         libcfs_debug_msg(msgdata, "%s", key->lck_area);
 406                 key->lck_area[0] = 0;
 407         }
 408         va_end(args);
 409         return 0;
 410 }
 411 EXPORT_SYMBOL(lu_cdebug_printer);
 412
 413 /**
 414  * Print object header.
 415  */
 416 void lu_object_header_print(const struct lu_env *env, void *cookie,
 417                             lu_printer_t printer,
 418                             const struct lu_object_header *hdr)
 419 {
 420         (*printer)(env, cookie, "header@%p[%#lx, %d, "DFID"%s%s%s]",
 421                    hdr, hdr->loh_flags, cfs_atomic_read(&hdr->loh_ref),
 422                    PFID(&hdr->loh_fid),
 423                    cfs_hlist_unhashed(&hdr->loh_hash) ? "" : " hash",
 424                    cfs_list_empty((cfs_list_t *)&hdr->loh_lru) ? \
 425                    "" : " lru",
 426                    hdr->loh_attr & LOHA_EXISTS ? " exist":"");
 427 }
 428 EXPORT_SYMBOL(lu_object_header_print);
 429
 430 /**
 431  * Print human readable representation of the \a o to the \a printer.
 432  */
 433 void lu_object_print(const struct lu_env *env, void *cookie,
 434                      lu_printer_t printer, const struct lu_object *o)
 435 {
 436         static const char ruler[] = "........................................";
 437         struct lu_object_header *top;
 438         int depth;
 439
 440         top = o->lo_header;
 441         lu_object_header_print(env, cookie, printer, top);
 442         (*printer)(env, cookie, "{ \n");
 443         cfs_list_for_each_entry(o, &top->loh_layers, lo_linkage) {
 444                 depth = o->lo_depth + 4;
 445
 446                 /*
 447                  * print `.' \a depth times followed by type name and address
 448                  */
 449                 (*printer)(env, cookie, "%*.*s%s@%p", depth, depth, ruler,
 450                            o->lo_dev->ld_type->ldt_name, o);
 451                 if (o->lo_ops->loo_object_print != NULL)
 452                         o->lo_ops->loo_object_print(env, cookie, printer, o);
 453                 (*printer)(env, cookie, "\n");
 454         }
 455         (*printer)(env, cookie, "} header@%p\n", top);
 456 }
 457 EXPORT_SYMBOL(lu_object_print);
 458
 459 /**
 460  * Check object consistency.
 461  */
 462 int lu_object_invariant(const struct lu_object *o)
 463 {
 464         struct lu_object_header *top;
 465
 466         top = o->lo_header;
 467         cfs_list_for_each_entry(o, &top->loh_layers, lo_linkage) {
 468                 if (o->lo_ops->loo_object_invariant != NULL &&
 469                     !o->lo_ops->loo_object_invariant(o))
 470                         return 0;
 471         }
 472         return 1;
 473 }
 474 EXPORT_SYMBOL(lu_object_invariant);
 475
 476 static struct lu_object *htable_lookup(struct lu_site *s,
 477                                        cfs_hash_bd_t *bd,
 478                                        const struct lu_fid *f,
 479                                        cfs_waitlink_t *waiter,
 480                                        __u64 *version)
 481 {
 482         struct lu_site_bkt_data *bkt;
 483         struct lu_object_header *h;
 484         cfs_hlist_node_t        *hnode;
 485         __u64  ver = cfs_hash_bd_version_get(bd);
 486
 487         if (*version == ver)
 488                 return NULL;
 489
 490         *version = ver;
 491         bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, bd);
 492         /* cfs_hash_bd_lookup_intent is a somehow "internal" function
 493          * of cfs_hash, but we don't want refcount on object right now */
 494         hnode = cfs_hash_bd_lookup_locked(s->ls_obj_hash, bd, (void *)f);
 495         if (hnode == NULL) {
 496                 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_MISS);
 497                 return NULL;
 498         }
 499
 500         h = container_of0(hnode, struct lu_object_header, loh_hash);
 501         if (likely(!lu_object_is_dying(h))) {
 502                 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_HIT);
 503                 cfs_list_del_init(&h->loh_lru);
 504                 return lu_object_top(h);
 505         }
 506
 507         /*
 508          * Lookup found an object being destroyed this object cannot be
 509          * returned (to assure that references to dying objects are eventually
 510          * drained), and moreover, lookup has to wait until object is freed.
 511          */
 512         cfs_atomic_dec(&h->loh_ref);
 513
 514         cfs_waitlink_init(waiter);
 515         cfs_waitq_add(&bkt->lsb_marche_funebre, waiter);
 516         cfs_set_current_state(CFS_TASK_UNINT);
 517         lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_DEATH_RACE);
 518         return ERR_PTR(-EAGAIN);
 519 }
 520
 521 /**
 522  * Search cache for an object with the fid \a f. If such object is found,
 523  * return it. Otherwise, create new object, insert it into cache and return
 524  * it. In any case, additional reference is acquired on the returned object.
 525  */
 526 struct lu_object *lu_object_find(const struct lu_env *env,
 527                                  struct lu_device *dev, const struct lu_fid *f,
 528                                  const struct lu_object_conf *conf)
 529 {
 530         return lu_object_find_at(env, dev->ld_site->ls_top_dev, f, conf);
 531 }
 532 EXPORT_SYMBOL(lu_object_find);
 533
 534 static struct lu_object *lu_object_new(const struct lu_env *env,
 535                                        struct lu_device *dev,
 536                                        const struct lu_fid *f,
 537                                        const struct lu_object_conf *conf)
 538 {
 539         struct lu_object        *o;
 540         cfs_hash_t              *hs;
 541         cfs_hash_bd_t            bd;
 542         struct lu_site_bkt_data *bkt;
 543
 544         o = lu_object_alloc(env, dev, f, conf);
 545         if (unlikely(IS_ERR(o)))
 546                 return o;
 547
 548         hs = dev->ld_site->ls_obj_hash;
 549         cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
 550         bkt = cfs_hash_bd_extra_get(hs, &bd);
 551         cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
 552         bkt->lsb_busy++;
 553         cfs_hash_bd_unlock(hs, &bd, 1);
 554         return o;
 555 }
 556
 557 /**
 558  * Core logic of lu_object_find*() functions.
 559  */
 560 static struct lu_object *lu_object_find_try(const struct lu_env *env,
 561                                             struct lu_device *dev,
 562                                             const struct lu_fid *f,
 563                                             const struct lu_object_conf *conf,
 564                                             cfs_waitlink_t *waiter)
 565 {
 566         struct lu_object      *o;
 567         struct lu_object      *shadow;
 568         struct lu_site        *s;
 569         cfs_hash_t            *hs;
 570         cfs_hash_bd_t          bd;
 571         __u64                  version = 0;
 572
 573         /*
 574          * This uses standard index maintenance protocol:
 575          *
 576          *     - search index under lock, and return object if found;
 577          *     - otherwise, unlock index, allocate new object;
 578          *     - lock index and search again;
 579          *     - if nothing is found (usual case), insert newly created
 580          *       object into index;
 581          *     - otherwise (race: other thread inserted object), free
 582          *       object just allocated.
 583          *     - unlock index;
 584          *     - return object.
 585          *
 586          * For "LOC_F_NEW" case, we are sure the object is new established.
 587          * It is unnecessary to perform lookup-alloc-lookup-insert, instead,
 588          * just alloc and insert directly.
 589          *
 590          * If dying object is found during index search, add @waiter to the
 591          * site wait-queue and return ERR_PTR(-EAGAIN).
 592          */
 593         if (conf != NULL && conf->loc_flags & LOC_F_NEW)
 594                 return lu_object_new(env, dev, f, conf);
 595
 596         s  = dev->ld_site;
 597         hs = s->ls_obj_hash;
 598         cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
 599         o = htable_lookup(s, &bd, f, waiter, &version);
 600         cfs_hash_bd_unlock(hs, &bd, 1);
 601         if (o != NULL)
 602                 return o;
 603
 604         /*
 605          * Allocate new object. This may result in rather complicated
 606          * operations, including fld queries, inode loading, etc.
 607          */
 608         o = lu_object_alloc(env, dev, f, conf);
 609         if (unlikely(IS_ERR(o)))
 610                 return o;
 611
 612         LASSERT(lu_fid_eq(lu_object_fid(o), f));
 613
 614         cfs_hash_bd_lock(hs, &bd, 1);
 615
 616         shadow = htable_lookup(s, &bd, f, waiter, &version);
 617         if (likely(shadow == NULL)) {
 618                 struct lu_site_bkt_data *bkt;
 619
 620                 bkt = cfs_hash_bd_extra_get(hs, &bd);
 621                 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
 622                 bkt->lsb_busy++;
 623                 cfs_hash_bd_unlock(hs, &bd, 1);
 624                 return o;
 625         }
 626
 627         lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_RACE);
 628         cfs_hash_bd_unlock(hs, &bd, 1);
 629         lu_object_free(env, o);
 630         return shadow;
 631 }
 632
 633 /**
 634  * Much like lu_object_find(), but top level device of object is specifically
 635  * \a dev rather than top level device of the site. This interface allows
 636  * objects of different "stacking" to be created within the same site.
 637  */
 638 struct lu_object *lu_object_find_at(const struct lu_env *env,
 639                                     struct lu_device *dev,
 640                                     const struct lu_fid *f,
 641                                     const struct lu_object_conf *conf)
 642 {
 643         struct lu_site_bkt_data *bkt;
 644         struct lu_object        *obj;
 645         cfs_waitlink_t           wait;
 646
 647         while (1) {
 648                 obj = lu_object_find_try(env, dev, f, conf, &wait);
 649                 if (obj != ERR_PTR(-EAGAIN))
 650                         return obj;
 651                 /*
 652                  * lu_object_find_try() already added waiter into the
 653                  * wait queue.
 654                  */
 655                 cfs_waitq_wait(&wait, CFS_TASK_UNINT);
 656                 bkt = lu_site_bkt_from_fid(dev->ld_site, (void *)f);
 657                 cfs_waitq_del(&bkt->lsb_marche_funebre, &wait);
 658         }
 659 }
 660 EXPORT_SYMBOL(lu_object_find_at);
 661
 662 /**
 663  * Find object with given fid, and return its slice belonging to given device.
 664  */
 665 struct lu_object *lu_object_find_slice(const struct lu_env *env,
 666                                        struct lu_device *dev,
 667                                        const struct lu_fid *f,
 668                                        const struct lu_object_conf *conf)
 669 {
 670         struct lu_object *top;
 671         struct lu_object *obj;
 672
 673         top = lu_object_find(env, dev, f, conf);
 674         if (!IS_ERR(top)) {
 675                 obj = lu_object_locate(top->lo_header, dev->ld_type);
 676                 if (obj == NULL)
 677                         lu_object_put(env, top);
 678         } else
 679                 obj = top;
 680         return obj;
 681 }
 682 EXPORT_SYMBOL(lu_object_find_slice);
 683
 684 /**
 685  * Global list of all device types.
 686  */
 687 static CFS_LIST_HEAD(lu_device_types);
 688
 689 int lu_device_type_init(struct lu_device_type *ldt)
 690 {
 691         int result;
 692
 693         CFS_INIT_LIST_HEAD(&ldt->ldt_linkage);
 694         result = ldt->ldt_ops->ldto_init(ldt);
 695         if (result == 0)
 696                 cfs_list_add(&ldt->ldt_linkage, &lu_device_types);
 697         return result;
 698 }
 699 EXPORT_SYMBOL(lu_device_type_init);
 700
 701 void lu_device_type_fini(struct lu_device_type *ldt)
 702 {
 703         cfs_list_del_init(&ldt->ldt_linkage);
 704         ldt->ldt_ops->ldto_fini(ldt);
 705 }
 706 EXPORT_SYMBOL(lu_device_type_fini);
 707
 708 void lu_types_stop(void)
 709 {
 710         struct lu_device_type *ldt;
 711
 712         cfs_list_for_each_entry(ldt, &lu_device_types, ldt_linkage) {
 713                 if (ldt->ldt_device_nr == 0)
 714                         ldt->ldt_ops->ldto_stop(ldt);
 715         }
 716 }
 717 EXPORT_SYMBOL(lu_types_stop);
 718
 719 /**
 720  * Global list of all sites on this node
 721  */
 722 static CFS_LIST_HEAD(lu_sites);
 723 static CFS_DEFINE_MUTEX(lu_sites_guard);
 724
 725 /**
 726  * Global environment used by site shrinker.
 727  */
 728 static struct lu_env lu_shrink_env;
 729
 730 struct lu_site_print_arg {
 731         struct lu_env   *lsp_env;
 732         void            *lsp_cookie;
 733         lu_printer_t     lsp_printer;
 734 };
 735
 736 static int
 737 lu_site_obj_print(cfs_hash_t *hs, cfs_hash_bd_t *bd,
 738                   cfs_hlist_node_t *hnode, void *data)
 739 {
 740         struct lu_site_print_arg *arg = (struct lu_site_print_arg *)data;
 741         struct lu_object_header  *h;
 742
 743         h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
 744         if (!cfs_list_empty(&h->loh_layers)) {
 745                 const struct lu_object *o;
 746
 747                 o = lu_object_top(h);
 748                 lu_object_print(arg->lsp_env, arg->lsp_cookie,
 749                                 arg->lsp_printer, o);
 750         } else {
 751                 lu_object_header_print(arg->lsp_env, arg->lsp_cookie,
 752                                        arg->lsp_printer, h);
 753         }
 754         return 0;
 755 }
 756
 757 /**
 758  * Print all objects in \a s.
 759  */
 760 void lu_site_print(const struct lu_env *env, struct lu_site *s, void *cookie,
 761                    lu_printer_t printer)
 762 {
 763         struct lu_site_print_arg arg = {
 764                 .lsp_env     = (struct lu_env *)env,
 765                 .lsp_cookie  = cookie,
 766                 .lsp_printer = printer,
 767         };
 768
 769         cfs_hash_for_each(s->ls_obj_hash, lu_site_obj_print, &arg);
 770 }
 771 EXPORT_SYMBOL(lu_site_print);
 772
 773 enum {
 774         LU_CACHE_PERCENT_MAX     = 50,
 775         LU_CACHE_PERCENT_DEFAULT = 20
 776 };
 777
 778 static unsigned int lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
 779 CFS_MODULE_PARM(lu_cache_percent, "i", int, 0644,
 780                 "Percentage of memory to be used as lu_object cache");
 781
 782 /**
 783  * Return desired hash table order.
 784  */
 785 static int lu_htable_order(void)
 786 {
 787         unsigned long cache_size;
 788         int bits;
 789
 790         /*
 791          * Calculate hash table size, assuming that we want reasonable
 792          * performance when 20% of total memory is occupied by cache of
 793          * lu_objects.
 794          *
 795          * Size of lu_object is (arbitrary) taken as 1K (together with inode).
 796          */
 797         cache_size = cfs_num_physpages;
 798
 799 #if BITS_PER_LONG == 32
 800         /* limit hashtable size for lowmem systems to low RAM */
 801         if (cache_size > 1 << (30 - CFS_PAGE_SHIFT))
 802                 cache_size = 1 << (30 - CFS_PAGE_SHIFT) * 3 / 4;
 803 #endif
 804
 805         /* clear off unreasonable cache setting. */
 806         if (lu_cache_percent == 0 || lu_cache_percent > LU_CACHE_PERCENT_MAX) {
 807                 CWARN("obdclass: invalid lu_cache_percent: %u, it must be in"
 808                       " the range of (0, %u]. Will use default value: %u.\n",
 809                       lu_cache_percent, LU_CACHE_PERCENT_MAX,
 810                       LU_CACHE_PERCENT_DEFAULT);
 811
 812                 lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
 813         }
 814         cache_size = cache_size / 100 * lu_cache_percent *
 815                 (CFS_PAGE_SIZE / 1024);
 816
 817         for (bits = 1; (1 << bits) < cache_size; ++bits) {
 818                 ;
 819         }
 820         return bits;
 821 }
 822
 823 static unsigned lu_obj_hop_hash(cfs_hash_t *hs,
 824                                 const void *key, unsigned mask)
 825 {
 826         struct lu_fid  *fid = (struct lu_fid *)key;
 827         __u32           hash;
 828
 829         hash = fid_flatten32(fid);
 830         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
 831         hash = cfs_hash_long(hash, hs->hs_bkt_bits);
 832
 833         /* give me another random factor */
 834         hash -= cfs_hash_long((unsigned long)hs, fid_oid(fid) % 11 + 3);
 835
 836         hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
 837         hash |= (fid_seq(fid) + fid_oid(fid)) & (CFS_HASH_NBKT(hs) - 1);
 838
 839         return hash & mask;
 840 }
 841
 842 static void *lu_obj_hop_object(cfs_hlist_node_t *hnode)
 843 {
 844         return cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
 845 }
 846
 847 static void *lu_obj_hop_key(cfs_hlist_node_t *hnode)
 848 {
 849         struct lu_object_header *h;
 850
 851         h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
 852         return &h->loh_fid;
 853 }
 854
 855 static int lu_obj_hop_keycmp(const void *key, cfs_hlist_node_t *hnode)
 856 {
 857         struct lu_object_header *h;
 858
 859         h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
 860         return lu_fid_eq(&h->loh_fid, (struct lu_fid *)key);
 861 }
 862
 863 static void lu_obj_hop_get(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 864 {
 865         struct lu_object_header *h;
 866
 867         h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
 868         if (cfs_atomic_add_return(1, &h->loh_ref) == 1) {
 869                 struct lu_site_bkt_data *bkt;
 870                 cfs_hash_bd_t            bd;
 871
 872                 cfs_hash_bd_get(hs, &h->loh_fid, &bd);
 873                 bkt = cfs_hash_bd_extra_get(hs, &bd);
 874                 bkt->lsb_busy++;
 875         }
 876 }
 877
 878 static void lu_obj_hop_put_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 879 {
 880         LBUG(); /* we should never called it */
 881 }
 882
 883 cfs_hash_ops_t lu_site_hash_ops = {
 884         .hs_hash        = lu_obj_hop_hash,
 885         .hs_key         = lu_obj_hop_key,
 886         .hs_keycmp      = lu_obj_hop_keycmp,
 887         .hs_object      = lu_obj_hop_object,
 888         .hs_get         = lu_obj_hop_get,
 889         .hs_put_locked  = lu_obj_hop_put_locked,
 890 };
 891
 892 /**
 893  * Initialize site \a s, with \a d as the top level device.
 894  */
 895 #define LU_SITE_BITS_MIN    12
 896 #define LU_SITE_BITS_MAX    24
 897 /**
 898  * total 256 buckets, we don't want too many buckets because:
 899  * - consume too much memory
 900  * - avoid unbalanced LRU list
 901  */
 902 #define LU_SITE_BKT_BITS    8
 903
 904 int lu_site_init(struct lu_site *s, struct lu_device *top)
 905 {
 906         struct lu_site_bkt_data *bkt;
 907         cfs_hash_bd_t bd;
 908         char name[16];
 909         int bits;
 910         int i;
 911         ENTRY;
 912
 913         memset(s, 0, sizeof *s);
 914         bits = lu_htable_order();
 915         snprintf(name, 16, "lu_site_%s", top->ld_type->ldt_name);
 916         for (bits = min(max(LU_SITE_BITS_MIN, bits), LU_SITE_BITS_MAX);
 917              bits >= LU_SITE_BITS_MIN; bits--) {
 918                 s->ls_obj_hash = cfs_hash_create(name, bits, bits,
 919                                                  bits - LU_SITE_BKT_BITS,
 920                                                  sizeof(*bkt), 0, 0,
 921                                                  &lu_site_hash_ops,
 922                                                  CFS_HASH_SPIN_BKTLOCK |
 923                                                  CFS_HASH_NO_ITEMREF |
 924                                                  CFS_HASH_DEPTH |
 925                                                  CFS_HASH_ASSERT_EMPTY);
 926                 if (s->ls_obj_hash != NULL)
 927                         break;
 928         }
 929
 930         if (s->ls_obj_hash == NULL) {
 931                 CERROR("failed to create lu_site hash with bits: %d\n", bits);
 932                 return -ENOMEM;
 933         }
 934
 935         cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
 936                 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
 937                 CFS_INIT_LIST_HEAD(&bkt->lsb_lru);
 938                 cfs_waitq_init(&bkt->lsb_marche_funebre);
 939         }
 940
 941         s->ls_stats = lprocfs_alloc_stats(LU_SS_LAST_STAT, 0);
 942         if (s->ls_stats == NULL) {
 943                 cfs_hash_putref(s->ls_obj_hash);
 944                 s->ls_obj_hash = NULL;
 945                 return -ENOMEM;
 946         }
 947
 948         lprocfs_counter_init(s->ls_stats, LU_SS_CREATED,
 949                              0, "created", "created");
 950         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_HIT,
 951                              0, "cache_hit", "cache_hit");
 952         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_MISS,
 953                              0, "cache_miss", "cache_miss");
 954         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_RACE,
 955                              0, "cache_race", "cache_race");
 956         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_DEATH_RACE,
 957                              0, "cache_death_race", "cache_death_race");
 958         lprocfs_counter_init(s->ls_stats, LU_SS_LRU_PURGED,
 959                              0, "lru_purged", "lru_purged");
 960
 961         CFS_INIT_LIST_HEAD(&s->ls_linkage);
 962         s->ls_top_dev = top;
 963         top->ld_site = s;
 964         lu_device_get(top);
 965         lu_ref_add(&top->ld_reference, "site-top", s);
 966
 967         CFS_INIT_LIST_HEAD(&s->ls_ld_linkage);
 968         cfs_spin_lock_init(&s->ls_ld_lock);
 969
 970         cfs_spin_lock(&s->ls_ld_lock);
 971         cfs_list_add(&top->ld_linkage, &s->ls_ld_linkage);
 972         cfs_spin_unlock(&s->ls_ld_lock);
 973
 974         RETURN(0);
 975 }
 976 EXPORT_SYMBOL(lu_site_init);
 977
 978 /**
 979  * Finalize \a s and release its resources.
 980  */
 981 void lu_site_fini(struct lu_site *s)
 982 {
 983         cfs_mutex_lock(&lu_sites_guard);
 984         cfs_list_del_init(&s->ls_linkage);
 985         cfs_mutex_unlock(&lu_sites_guard);
 986
 987         if (s->ls_obj_hash != NULL) {
 988                 cfs_hash_putref(s->ls_obj_hash);
 989                 s->ls_obj_hash = NULL;
 990         }
 991
 992         if (s->ls_top_dev != NULL) {
 993                 s->ls_top_dev->ld_site = NULL;
 994                 lu_ref_del(&s->ls_top_dev->ld_reference, "site-top", s);
 995                 lu_device_put(s->ls_top_dev);
 996                 s->ls_top_dev = NULL;
 997         }
 998
 999         if (s->ls_stats != NULL)
1000                 lprocfs_free_stats(&s->ls_stats);
1001 }
1002 EXPORT_SYMBOL(lu_site_fini);
1003
1004 /**
1005  * Called when initialization of stack for this site is completed.
1006  */
1007 int lu_site_init_finish(struct lu_site *s)
1008 {
1009         int result;
1010         cfs_mutex_lock(&lu_sites_guard);
1011         result = lu_context_refill(&lu_shrink_env.le_ctx);
1012         if (result == 0)
1013                 cfs_list_add(&s->ls_linkage, &lu_sites);
1014         cfs_mutex_unlock(&lu_sites_guard);
1015         return result;
1016 }
1017 EXPORT_SYMBOL(lu_site_init_finish);
1018
1019 /**
1020  * Acquire additional reference on device \a d
1021  */
1022 void lu_device_get(struct lu_device *d)
1023 {
1024         cfs_atomic_inc(&d->ld_ref);
1025 }
1026 EXPORT_SYMBOL(lu_device_get);
1027
1028 /**
1029  * Release reference on device \a d.
1030  */
1031 void lu_device_put(struct lu_device *d)
1032 {
1033         LASSERT(cfs_atomic_read(&d->ld_ref) > 0);
1034         cfs_atomic_dec(&d->ld_ref);
1035 }
1036 EXPORT_SYMBOL(lu_device_put);
1037
1038 /**
1039  * Initialize device \a d of type \a t.
1040  */
1041 int lu_device_init(struct lu_device *d, struct lu_device_type *t)
1042 {
1043         if (t->ldt_device_nr++ == 0 && t->ldt_ops->ldto_start != NULL)
1044                 t->ldt_ops->ldto_start(t);
1045         memset(d, 0, sizeof *d);
1046         cfs_atomic_set(&d->ld_ref, 0);
1047         d->ld_type = t;
1048         lu_ref_init(&d->ld_reference);
1049         CFS_INIT_LIST_HEAD(&d->ld_linkage);
1050         return 0;
1051 }
1052 EXPORT_SYMBOL(lu_device_init);
1053
1054 /**
1055  * Finalize device \a d.
1056  */
1057 void lu_device_fini(struct lu_device *d)
1058 {
1059         struct lu_device_type *t;
1060
1061         t = d->ld_type;
1062         if (d->ld_obd != NULL) {
1063                 d->ld_obd->obd_lu_dev = NULL;
1064                 d->ld_obd = NULL;
1065         }
1066
1067         lu_ref_fini(&d->ld_reference);
1068         LASSERTF(cfs_atomic_read(&d->ld_ref) == 0,
1069                  "Refcount is %u\n", cfs_atomic_read(&d->ld_ref));
1070         LASSERT(t->ldt_device_nr > 0);
1071         if (--t->ldt_device_nr == 0 && t->ldt_ops->ldto_stop != NULL)
1072                 t->ldt_ops->ldto_stop(t);
1073 }
1074 EXPORT_SYMBOL(lu_device_fini);
1075
1076 /**
1077  * Initialize object \a o that is part of compound object \a h and was created
1078  * by device \a d.
1079  */
1080 int lu_object_init(struct lu_object *o,
1081                    struct lu_object_header *h, struct lu_device *d)
1082 {
1083         memset(o, 0, sizeof *o);
1084         o->lo_header = h;
1085         o->lo_dev    = d;
1086         lu_device_get(d);
1087         o->lo_dev_ref = lu_ref_add(&d->ld_reference, "lu_object", o);
1088         CFS_INIT_LIST_HEAD(&o->lo_linkage);
1089         return 0;
1090 }
1091 EXPORT_SYMBOL(lu_object_init);
1092
1093 /**
1094  * Finalize object and release its resources.
1095  */
1096 void lu_object_fini(struct lu_object *o)
1097 {
1098         struct lu_device *dev = o->lo_dev;
1099
1100         LASSERT(cfs_list_empty(&o->lo_linkage));
1101
1102         if (dev != NULL) {
1103                 lu_ref_del_at(&dev->ld_reference,
1104                               o->lo_dev_ref , "lu_object", o);
1105                 lu_device_put(dev);
1106                 o->lo_dev = NULL;
1107         }
1108 }
1109 EXPORT_SYMBOL(lu_object_fini);
1110
1111 /**
1112  * Add object \a o as first layer of compound object \a h
1113  *
1114  * This is typically called by the ->ldo_object_alloc() method of top-level
1115  * device.
1116  */
1117 void lu_object_add_top(struct lu_object_header *h, struct lu_object *o)
1118 {
1119         cfs_list_move(&o->lo_linkage, &h->loh_layers);
1120 }
1121 EXPORT_SYMBOL(lu_object_add_top);
1122
1123 /**
1124  * Add object \a o as a layer of compound object, going after \a before.
1125  *
1126  * This is typically called by the ->ldo_object_alloc() method of \a
1127  * before->lo_dev.
1128  */
1129 void lu_object_add(struct lu_object *before, struct lu_object *o)
1130 {
1131         cfs_list_move(&o->lo_linkage, &before->lo_linkage);
1132 }
1133 EXPORT_SYMBOL(lu_object_add);
1134
1135 /**
1136  * Initialize compound object.
1137  */
1138 int lu_object_header_init(struct lu_object_header *h)
1139 {
1140         memset(h, 0, sizeof *h);
1141         cfs_atomic_set(&h->loh_ref, 1);
1142         CFS_INIT_HLIST_NODE(&h->loh_hash);
1143         CFS_INIT_LIST_HEAD(&h->loh_lru);
1144         CFS_INIT_LIST_HEAD(&h->loh_layers);
1145         lu_ref_init(&h->loh_reference);
1146         return 0;
1147 }
1148 EXPORT_SYMBOL(lu_object_header_init);
1149
1150 /**
1151  * Finalize compound object.
1152  */
1153 void lu_object_header_fini(struct lu_object_header *h)
1154 {
1155         LASSERT(cfs_list_empty(&h->loh_layers));
1156         LASSERT(cfs_list_empty(&h->loh_lru));
1157         LASSERT(cfs_hlist_unhashed(&h->loh_hash));
1158         lu_ref_fini(&h->loh_reference);
1159 }
1160 EXPORT_SYMBOL(lu_object_header_fini);
1161
1162 /**
1163  * Given a compound object, find its slice, corresponding to the device type
1164  * \a dtype.
1165  */
1166 struct lu_object *lu_object_locate(struct lu_object_header *h,
1167                                    const struct lu_device_type *dtype)
1168 {
1169         struct lu_object *o;
1170
1171         cfs_list_for_each_entry(o, &h->loh_layers, lo_linkage) {
1172                 if (o->lo_dev->ld_type == dtype)
1173                         return o;
1174         }
1175         return NULL;
1176 }
1177 EXPORT_SYMBOL(lu_object_locate);
1178
1179
1180
1181 /**
1182  * Finalize and free devices in the device stack.
1183  *
1184  * Finalize device stack by purging object cache, and calling
1185  * lu_device_type_operations::ldto_device_fini() and
1186  * lu_device_type_operations::ldto_device_free() on all devices in the stack.
1187  */
1188 void lu_stack_fini(const struct lu_env *env, struct lu_device *top)
1189 {
1190         struct lu_site   *site = top->ld_site;
1191         struct lu_device *scan;
1192         struct lu_device *next;
1193
1194         lu_site_purge(env, site, ~0);
1195         for (scan = top; scan != NULL; scan = next) {
1196                 next = scan->ld_type->ldt_ops->ldto_device_fini(env, scan);
1197                 lu_ref_del(&scan->ld_reference, "lu-stack", &lu_site_init);
1198                 lu_device_put(scan);
1199         }
1200
1201         /* purge again. */
1202         lu_site_purge(env, site, ~0);
1203
1204         if (!cfs_hash_is_empty(site->ls_obj_hash)) {
1205                 /*
1206                  * Uh-oh, objects still exist.
1207                  */
1208                 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
1209
1210                 lu_site_print(env, site, &msgdata, lu_cdebug_printer);
1211         }
1212
1213         for (scan = top; scan != NULL; scan = next) {
1214                 const struct lu_device_type *ldt = scan->ld_type;
1215                 struct obd_type             *type;
1216
1217                 next = ldt->ldt_ops->ldto_device_free(env, scan);
1218                 type = ldt->ldt_obd_type;
1219                 if (type != NULL) {
1220                         type->typ_refcnt--;
1221                         class_put_type(type);
1222                 }
1223         }
1224 }
1225 EXPORT_SYMBOL(lu_stack_fini);
1226
1227 enum {
1228         /**
1229          * Maximal number of tld slots.
1230          */
1231         LU_CONTEXT_KEY_NR = 32
1232 };
1233
1234 static struct lu_context_key *lu_keys[LU_CONTEXT_KEY_NR] = { NULL, };
1235
1236 static cfs_spinlock_t lu_keys_guard = CFS_SPIN_LOCK_UNLOCKED;
1237
1238 /**
1239  * Global counter incremented whenever key is registered, unregistered,
1240  * revived or quiesced. This is used to void unnecessary calls to
1241  * lu_context_refill(). No locking is provided, as initialization and shutdown
1242  * are supposed to be externally serialized.
1243  */
1244 static unsigned key_set_version = 0;
1245
1246 /**
1247  * Register new key.
1248  */
1249 int lu_context_key_register(struct lu_context_key *key)
1250 {
1251         int result;
1252         int i;
1253
1254         LASSERT(key->lct_init != NULL);
1255         LASSERT(key->lct_fini != NULL);
1256         LASSERT(key->lct_tags != 0);
1257         LASSERT(key->lct_owner != NULL);
1258
1259         result = -ENFILE;
1260         cfs_spin_lock(&lu_keys_guard);
1261         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1262                 if (lu_keys[i] == NULL) {
1263                         key->lct_index = i;
1264                         cfs_atomic_set(&key->lct_used, 1);
1265                         lu_keys[i] = key;
1266                         lu_ref_init(&key->lct_reference);
1267                         result = 0;
1268                         ++key_set_version;
1269                         break;
1270                 }
1271         }
1272         cfs_spin_unlock(&lu_keys_guard);
1273         return result;
1274 }
1275 EXPORT_SYMBOL(lu_context_key_register);
1276
1277 static void key_fini(struct lu_context *ctx, int index)
1278 {
1279         if (ctx->lc_value != NULL && ctx->lc_value[index] != NULL) {
1280                 struct lu_context_key *key;
1281
1282                 key = lu_keys[index];
1283                 LASSERT(key != NULL);
1284                 LASSERT(key->lct_fini != NULL);
1285                 LASSERT(cfs_atomic_read(&key->lct_used) > 1);
1286
1287                 key->lct_fini(ctx, key, ctx->lc_value[index]);
1288                 lu_ref_del(&key->lct_reference, "ctx", ctx);
1289                 cfs_atomic_dec(&key->lct_used);
1290                 LASSERT(key->lct_owner != NULL);
1291                 if (!(ctx->lc_tags & LCT_NOREF)) {
1292                         LASSERT(cfs_module_refcount(key->lct_owner) > 0);
1293                         cfs_module_put(key->lct_owner);
1294                 }
1295                 ctx->lc_value[index] = NULL;
1296         }
1297 }
1298
1299 /**
1300  * Deregister key.
1301  */
1302 void lu_context_key_degister(struct lu_context_key *key)
1303 {
1304         LASSERT(cfs_atomic_read(&key->lct_used) >= 1);
1305         LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1306
1307         lu_context_key_quiesce(key);
1308
1309         ++key_set_version;
1310         cfs_spin_lock(&lu_keys_guard);
1311         key_fini(&lu_shrink_env.le_ctx, key->lct_index);
1312         if (lu_keys[key->lct_index]) {
1313                 lu_keys[key->lct_index] = NULL;
1314                 lu_ref_fini(&key->lct_reference);
1315         }
1316         cfs_spin_unlock(&lu_keys_guard);
1317
1318         LASSERTF(cfs_atomic_read(&key->lct_used) == 1,
1319                  "key has instances: %d\n",
1320                  cfs_atomic_read(&key->lct_used));
1321 }
1322 EXPORT_SYMBOL(lu_context_key_degister);
1323
1324 /**
1325  * Register a number of keys. This has to be called after all keys have been
1326  * initialized by a call to LU_CONTEXT_KEY_INIT().
1327  */
1328 int lu_context_key_register_many(struct lu_context_key *k, ...)
1329 {
1330         struct lu_context_key *key = k;
1331         va_list args;
1332         int result;
1333
1334         va_start(args, k);
1335         do {
1336                 result = lu_context_key_register(key);
1337                 if (result)
1338                         break;
1339                 key = va_arg(args, struct lu_context_key *);
1340         } while (key != NULL);
1341         va_end(args);
1342
1343         if (result != 0) {
1344                 va_start(args, k);
1345                 while (k != key) {
1346                         lu_context_key_degister(k);
1347                         k = va_arg(args, struct lu_context_key *);
1348                 }
1349                 va_end(args);
1350         }
1351
1352         return result;
1353 }
1354 EXPORT_SYMBOL(lu_context_key_register_many);
1355
1356 /**
1357  * De-register a number of keys. This is a dual to
1358  * lu_context_key_register_many().
1359  */
1360 void lu_context_key_degister_many(struct lu_context_key *k, ...)
1361 {
1362         va_list args;
1363
1364         va_start(args, k);
1365         do {
1366                 lu_context_key_degister(k);
1367                 k = va_arg(args, struct lu_context_key*);
1368         } while (k != NULL);
1369         va_end(args);
1370 }
1371 EXPORT_SYMBOL(lu_context_key_degister_many);
1372
1373 /**
1374  * Revive a number of keys.
1375  */
1376 void lu_context_key_revive_many(struct lu_context_key *k, ...)
1377 {
1378         va_list args;
1379
1380         va_start(args, k);
1381         do {
1382                 lu_context_key_revive(k);
1383                 k = va_arg(args, struct lu_context_key*);
1384         } while (k != NULL);
1385         va_end(args);
1386 }
1387 EXPORT_SYMBOL(lu_context_key_revive_many);
1388
1389 /**
1390  * Quiescent a number of keys.
1391  */
1392 void lu_context_key_quiesce_many(struct lu_context_key *k, ...)
1393 {
1394         va_list args;
1395
1396         va_start(args, k);
1397         do {
1398                 lu_context_key_quiesce(k);
1399                 k = va_arg(args, struct lu_context_key*);
1400         } while (k != NULL);
1401         va_end(args);
1402 }
1403 EXPORT_SYMBOL(lu_context_key_quiesce_many);
1404
1405 /**
1406  * Return value associated with key \a key in context \a ctx.
1407  */
1408 void *lu_context_key_get(const struct lu_context *ctx,
1409                          const struct lu_context_key *key)
1410 {
1411         LINVRNT(ctx->lc_state == LCS_ENTERED);
1412         LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1413         LASSERT(lu_keys[key->lct_index] == key);
1414         return ctx->lc_value[key->lct_index];
1415 }
1416 EXPORT_SYMBOL(lu_context_key_get);
1417
1418 /**
1419  * List of remembered contexts. XXX document me.
1420  */
1421 static CFS_LIST_HEAD(lu_context_remembered);
1422
1423 /**
1424  * Destroy \a key in all remembered contexts. This is used to destroy key
1425  * values in "shared" contexts (like service threads), when a module owning
1426  * the key is about to be unloaded.
1427  */
1428 void lu_context_key_quiesce(struct lu_context_key *key)
1429 {
1430         struct lu_context *ctx;
1431         extern unsigned cl_env_cache_purge(unsigned nr);
1432
1433         if (!(key->lct_tags & LCT_QUIESCENT)) {
1434                 /*
1435                  * XXX layering violation.
1436                  */
1437                 cl_env_cache_purge(~0);
1438                 key->lct_tags |= LCT_QUIESCENT;
1439                 /*
1440                  * XXX memory barrier has to go here.
1441                  */
1442                 cfs_spin_lock(&lu_keys_guard);
1443                 cfs_list_for_each_entry(ctx, &lu_context_remembered,
1444                                         lc_remember)
1445                         key_fini(ctx, key->lct_index);
1446                 cfs_spin_unlock(&lu_keys_guard);
1447                 ++key_set_version;
1448         }
1449 }
1450 EXPORT_SYMBOL(lu_context_key_quiesce);
1451
1452 void lu_context_key_revive(struct lu_context_key *key)
1453 {
1454         key->lct_tags &= ~LCT_QUIESCENT;
1455         ++key_set_version;
1456 }
1457 EXPORT_SYMBOL(lu_context_key_revive);
1458
1459 static void keys_fini(struct lu_context *ctx)
1460 {
1461         int i;
1462
1463         cfs_spin_lock(&lu_keys_guard);
1464         if (ctx->lc_value != NULL) {
1465                 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i)
1466                         key_fini(ctx, i);
1467                 OBD_FREE(ctx->lc_value,
1468                          ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1469                 ctx->lc_value = NULL;
1470         }
1471         cfs_spin_unlock(&lu_keys_guard);
1472 }
1473
1474 static int keys_fill(struct lu_context *ctx)
1475 {
1476         int i;
1477
1478         LINVRNT(ctx->lc_value != NULL);
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. key_set_version is only changed in rare cases when modules
1603  * are loaded and removed.
1604  */
1605 int lu_context_refill(struct lu_context *ctx)
1606 {
1607         return likely(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);