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, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/obdclass/lu_object.c
  37  *
  38  * Lustre Object.
  39  * These are the only exported functions, they provide some generic
  40  * infrastructure for managing object devices
  41  *
  42  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  43  */
  44
  45 #define DEBUG_SUBSYSTEM S_CLASS
  46
  47 #include <libcfs/libcfs.h>
  48
  49 #ifdef __KERNEL__
  50 # include <linux/module.h>
  51 #endif
  52
  53 /* hash_long() */
  54 #include <libcfs/libcfs_hash.h>
  55 #include <obd_class.h>
  56 #include <obd_support.h>
  57 #include <lustre_disk.h>
  58 #include <lustre_fid.h>
  59 #include <lu_object.h>
  60 #include <libcfs/list.h>
  61 /* lu_time_global_{init,fini}() */
  62 #include <lu_time.h>
  63
  64 static void lu_object_free(const struct lu_env *env, struct lu_object *o);
  65
  66 /**
  67  * Decrease reference counter on object. If last reference is freed, return
  68  * object to the cache, unless lu_object_is_dying(o) holds. In the latter
  69  * case, free object immediately.
  70  */
  71 void lu_object_put(const struct lu_env *env, struct lu_object *o)
  72 {
  73         struct lu_site_bkt_data *bkt;
  74         struct lu_object_header *top;
  75         struct lu_site          *site;
  76         struct lu_object        *orig;
  77         cfs_hash_bd_t            bd;
  78         const struct lu_fid     *fid;
  79
  80         top  = o->lo_header;
  81         site = o->lo_dev->ld_site;
  82         orig = o;
  83
  84         /*
  85          * till we have full fids-on-OST implemented anonymous objects
  86          * are possible in OSP. such an object isn't listed in the site
  87          * so we should not remove it from the site.
  88          */
  89         fid = lu_object_fid(o);
  90         if (fid_is_zero(fid)) {
  91                 LASSERT(top->loh_hash.next == NULL
  92                         && top->loh_hash.pprev == NULL);
  93                 LASSERT(cfs_list_empty(&top->loh_lru));
  94                 if (!cfs_atomic_dec_and_test(&top->loh_ref))
  95                         return;
  96                 cfs_list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
  97                         if (o->lo_ops->loo_object_release != NULL)
  98                                 o->lo_ops->loo_object_release(env, o);
  99                 }
 100                 lu_object_free(env, orig);
 101                 return;
 102         }
 103
 104         cfs_hash_bd_get(site->ls_obj_hash, &top->loh_fid, &bd);
 105         bkt = cfs_hash_bd_extra_get(site->ls_obj_hash, &bd);
 106
 107         if (!cfs_hash_bd_dec_and_lock(site->ls_obj_hash, &bd, &top->loh_ref)) {
 108                 if (lu_object_is_dying(top)) {
 109
 110                         /*
 111                          * somebody may be waiting for this, currently only
 112                          * used for cl_object, see cl_object_put_last().
 113                          */
 114                         cfs_waitq_broadcast(&bkt->lsb_marche_funebre);
 115                 }
 116                 return;
 117         }
 118
 119         LASSERT(bkt->lsb_busy > 0);
 120         bkt->lsb_busy--;
 121         /*
 122          * When last reference is released, iterate over object
 123          * layers, and notify them that object is no longer busy.
 124          */
 125         cfs_list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
 126                 if (o->lo_ops->loo_object_release != NULL)
 127                         o->lo_ops->loo_object_release(env, o);
 128         }
 129
 130         if (!lu_object_is_dying(top)) {
 131                 LASSERT(cfs_list_empty(&top->loh_lru));
 132                 cfs_list_add_tail(&top->loh_lru, &bkt->lsb_lru);
 133                 cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
 134                 return;
 135         }
 136
 137         /*
 138          * If object is dying (will not be cached), removed it
 139          * from hash table and LRU.
 140          *
 141          * This is done with hash table and LRU lists locked. As the only
 142          * way to acquire first reference to previously unreferenced
 143          * object is through hash-table lookup (lu_object_find()),
 144          * or LRU scanning (lu_site_purge()), that are done under hash-table
 145          * and LRU lock, no race with concurrent object lookup is possible
 146          * and we can safely destroy object below.
 147          */
 148         cfs_hash_bd_del_locked(site->ls_obj_hash, &bd, &top->loh_hash);
 149         cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
 150         /*
 151          * Object was already removed from hash and lru above, can
 152          * kill it.
 153          */
 154         lu_object_free(env, orig);
 155 }
 156 EXPORT_SYMBOL(lu_object_put);
 157
 158 /**
 159  * Put object and don't keep in cache. This is temporary solution for
 160  * multi-site objects when its layering is not constant.
 161  */
 162 void lu_object_put_nocache(const struct lu_env *env, struct lu_object *o)
 163 {
 164         set_bit(LU_OBJECT_HEARD_BANSHEE,
 165                     &o->lo_header->loh_flags);
 166         return lu_object_put(env, o);
 167 }
 168 EXPORT_SYMBOL(lu_object_put_nocache);
 169
 170 /**
 171  * Allocate new object.
 172  *
 173  * This follows object creation protocol, described in the comment within
 174  * struct lu_device_operations definition.
 175  */
 176 static struct lu_object *lu_object_alloc(const struct lu_env *env,
 177                                          struct lu_device *dev,
 178                                          const struct lu_fid *f,
 179                                          const struct lu_object_conf *conf)
 180 {
 181         struct lu_object *scan;
 182         struct lu_object *top;
 183         cfs_list_t *layers;
 184         int clean;
 185         int result;
 186         ENTRY;
 187
 188         /*
 189          * Create top-level object slice. This will also create
 190          * lu_object_header.
 191          */
 192         top = dev->ld_ops->ldo_object_alloc(env, NULL, dev);
 193         if (top == NULL)
 194                 RETURN(ERR_PTR(-ENOMEM));
 195         /*
 196          * This is the only place where object fid is assigned. It's constant
 197          * after this point.
 198          */
 199         LASSERT(fid_is_igif(f) || fid_ver(f) == 0);
 200         top->lo_header->loh_fid = *f;
 201         layers = &top->lo_header->loh_layers;
 202         do {
 203                 /*
 204                  * Call ->loo_object_init() repeatedly, until no more new
 205                  * object slices are created.
 206                  */
 207                 clean = 1;
 208                 cfs_list_for_each_entry(scan, layers, lo_linkage) {
 209                         if (scan->lo_flags & LU_OBJECT_ALLOCATED)
 210                                 continue;
 211                         clean = 0;
 212                         scan->lo_header = top->lo_header;
 213                         result = scan->lo_ops->loo_object_init(env, scan, conf);
 214                         if (result != 0) {
 215                                 lu_object_free(env, top);
 216                                 RETURN(ERR_PTR(result));
 217                         }
 218                         scan->lo_flags |= LU_OBJECT_ALLOCATED;
 219                 }
 220         } while (!clean);
 221
 222         cfs_list_for_each_entry_reverse(scan, layers, lo_linkage) {
 223                 if (scan->lo_ops->loo_object_start != NULL) {
 224                         result = scan->lo_ops->loo_object_start(env, scan);
 225                         if (result != 0) {
 226                                 lu_object_free(env, top);
 227                                 RETURN(ERR_PTR(result));
 228                         }
 229                 }
 230         }
 231
 232         lprocfs_counter_incr(dev->ld_site->ls_stats, LU_SS_CREATED);
 233         RETURN(top);
 234 }
 235
 236 /**
 237  * Free an object.
 238  */
 239 static void lu_object_free(const struct lu_env *env, struct lu_object *o)
 240 {
 241         struct lu_site_bkt_data *bkt;
 242         struct lu_site          *site;
 243         struct lu_object        *scan;
 244         cfs_list_t              *layers;
 245         cfs_list_t               splice;
 246
 247         site   = o->lo_dev->ld_site;
 248         layers = &o->lo_header->loh_layers;
 249         bkt    = lu_site_bkt_from_fid(site, &o->lo_header->loh_fid);
 250         /*
 251          * First call ->loo_object_delete() method to release all resources.
 252          */
 253         cfs_list_for_each_entry_reverse(scan, layers, lo_linkage) {
 254                 if (scan->lo_ops->loo_object_delete != NULL)
 255                         scan->lo_ops->loo_object_delete(env, scan);
 256         }
 257
 258         /*
 259          * Then, splice object layers into stand-alone list, and call
 260          * ->loo_object_free() on all layers to free memory. Splice is
 261          * necessary, because lu_object_header is freed together with the
 262          * top-level slice.
 263          */
 264         CFS_INIT_LIST_HEAD(&splice);
 265         cfs_list_splice_init(layers, &splice);
 266         while (!cfs_list_empty(&splice)) {
 267                 /*
 268                  * Free layers in bottom-to-top order, so that object header
 269                  * lives as long as possible and ->loo_object_free() methods
 270                  * can look at its contents.
 271                  */
 272                 o = container_of0(splice.prev, struct lu_object, lo_linkage);
 273                 cfs_list_del_init(&o->lo_linkage);
 274                 LASSERT(o->lo_ops->loo_object_free != NULL);
 275                 o->lo_ops->loo_object_free(env, o);
 276         }
 277
 278         if (cfs_waitq_active(&bkt->lsb_marche_funebre))
 279                 cfs_waitq_broadcast(&bkt->lsb_marche_funebre);
 280 }
 281
 282 /**
 283  * Free \a nr objects from the cold end of the site LRU list.
 284  */
 285 int lu_site_purge(const struct lu_env *env, struct lu_site *s, int nr)
 286 {
 287         struct lu_object_header *h;
 288         struct lu_object_header *temp;
 289         struct lu_site_bkt_data *bkt;
 290         cfs_hash_bd_t            bd;
 291         cfs_hash_bd_t            bd2;
 292         cfs_list_t               dispose;
 293         int                      did_sth;
 294         int                      start;
 295         int                      count;
 296         int                      bnr;
 297         int                      i;
 298
 299         CFS_INIT_LIST_HEAD(&dispose);
 300         /*
 301          * Under LRU list lock, scan LRU list and move unreferenced objects to
 302          * the dispose list, removing them from LRU and hash table.
 303          */
 304         start = s->ls_purge_start;
 305         bnr = (nr == ~0) ? -1 : nr / CFS_HASH_NBKT(s->ls_obj_hash) + 1;
 306  again:
 307         did_sth = 0;
 308         cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
 309                 if (i < start)
 310                         continue;
 311                 count = bnr;
 312                 cfs_hash_bd_lock(s->ls_obj_hash, &bd, 1);
 313                 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
 314
 315                 cfs_list_for_each_entry_safe(h, temp, &bkt->lsb_lru, loh_lru) {
 316                         LASSERT(cfs_atomic_read(&h->loh_ref) == 0);
 317
 318                         cfs_hash_bd_get(s->ls_obj_hash, &h->loh_fid, &bd2);
 319                         LASSERT(bd.bd_bucket == bd2.bd_bucket);
 320
 321                         cfs_hash_bd_del_locked(s->ls_obj_hash,
 322                                                &bd2, &h->loh_hash);
 323                         cfs_list_move(&h->loh_lru, &dispose);
 324                         if (did_sth == 0)
 325                                 did_sth = 1;
 326
 327                         if (nr != ~0 && --nr == 0)
 328                                 break;
 329
 330                         if (count > 0 && --count == 0)
 331                                 break;
 332
 333                 }
 334                 cfs_hash_bd_unlock(s->ls_obj_hash, &bd, 1);
 335                 cfs_cond_resched();
 336                 /*
 337                  * Free everything on the dispose list. This is safe against
 338                  * races due to the reasons described in lu_object_put().
 339                  */
 340                 while (!cfs_list_empty(&dispose)) {
 341                         h = container_of0(dispose.next,
 342                                           struct lu_object_header, loh_lru);
 343                         cfs_list_del_init(&h->loh_lru);
 344                         lu_object_free(env, lu_object_top(h));
 345                         lprocfs_counter_incr(s->ls_stats, LU_SS_LRU_PURGED);
 346                 }
 347
 348                 if (nr == 0)
 349                         break;
 350         }
 351
 352         if (nr != 0 && did_sth && start != 0) {
 353                 start = 0; /* restart from the first bucket */
 354                 goto again;
 355         }
 356         /* race on s->ls_purge_start, but nobody cares */
 357         s->ls_purge_start = i % CFS_HASH_NBKT(s->ls_obj_hash);
 358
 359         return nr;
 360 }
 361 EXPORT_SYMBOL(lu_site_purge);
 362
 363 /*
 364  * Object printing.
 365  *
 366  * Code below has to jump through certain loops to output object description
 367  * into libcfs_debug_msg-based log. The problem is that lu_object_print()
 368  * composes object description from strings that are parts of _lines_ of
 369  * output (i.e., strings that are not terminated by newline). This doesn't fit
 370  * very well into libcfs_debug_msg() interface that assumes that each message
 371  * supplied to it is a self-contained output line.
 372  *
 373  * To work around this, strings are collected in a temporary buffer
 374  * (implemented as a value of lu_cdebug_key key), until terminating newline
 375  * character is detected.
 376  *
 377  */
 378
 379 enum {
 380         /**
 381          * Maximal line size.
 382          *
 383          * XXX overflow is not handled correctly.
 384          */
 385         LU_CDEBUG_LINE = 512
 386 };
 387
 388 struct lu_cdebug_data {
 389         /**
 390          * Temporary buffer.
 391          */
 392         char lck_area[LU_CDEBUG_LINE];
 393 };
 394
 395 /* context key constructor/destructor: lu_global_key_init, lu_global_key_fini */
 396 LU_KEY_INIT_FINI(lu_global, struct lu_cdebug_data);
 397
 398 /**
 399  * Key, holding temporary buffer. This key is registered very early by
 400  * lu_global_init().
 401  */
 402 struct lu_context_key lu_global_key = {
 403         .lct_tags = LCT_MD_THREAD | LCT_DT_THREAD |
 404                     LCT_MG_THREAD | LCT_CL_THREAD,
 405         .lct_init = lu_global_key_init,
 406         .lct_fini = lu_global_key_fini
 407 };
 408
 409 /**
 410  * Printer function emitting messages through libcfs_debug_msg().
 411  */
 412 int lu_cdebug_printer(const struct lu_env *env,
 413                       void *cookie, const char *format, ...)
 414 {
 415         struct libcfs_debug_msg_data *msgdata = cookie;
 416         struct lu_cdebug_data        *key;
 417         int used;
 418         int complete;
 419         va_list args;
 420
 421         va_start(args, format);
 422
 423         key = lu_context_key_get(&env->le_ctx, &lu_global_key);
 424         LASSERT(key != NULL);
 425
 426         used = strlen(key->lck_area);
 427         complete = format[strlen(format) - 1] == '\n';
 428         /*
 429          * Append new chunk to the buffer.
 430          */
 431         vsnprintf(key->lck_area + used,
 432                   ARRAY_SIZE(key->lck_area) - used, format, args);
 433         if (complete) {
 434                 if (cfs_cdebug_show(msgdata->msg_mask, msgdata->msg_subsys))
 435                         libcfs_debug_msg(msgdata, "%s", key->lck_area);
 436                 key->lck_area[0] = 0;
 437         }
 438         va_end(args);
 439         return 0;
 440 }
 441 EXPORT_SYMBOL(lu_cdebug_printer);
 442
 443 /**
 444  * Print object header.
 445  */
 446 void lu_object_header_print(const struct lu_env *env, void *cookie,
 447                             lu_printer_t printer,
 448                             const struct lu_object_header *hdr)
 449 {
 450         (*printer)(env, cookie, "header@%p[%#lx, %d, "DFID"%s%s%s]",
 451                    hdr, hdr->loh_flags, cfs_atomic_read(&hdr->loh_ref),
 452                    PFID(&hdr->loh_fid),
 453                    cfs_hlist_unhashed(&hdr->loh_hash) ? "" : " hash",
 454                    cfs_list_empty((cfs_list_t *)&hdr->loh_lru) ? \
 455                    "" : " lru",
 456                    hdr->loh_attr & LOHA_EXISTS ? " exist":"");
 457 }
 458 EXPORT_SYMBOL(lu_object_header_print);
 459
 460 /**
 461  * Print human readable representation of the \a o to the \a printer.
 462  */
 463 void lu_object_print(const struct lu_env *env, void *cookie,
 464                      lu_printer_t printer, const struct lu_object *o)
 465 {
 466         static const char ruler[] = "........................................";
 467         struct lu_object_header *top;
 468         int depth;
 469
 470         top = o->lo_header;
 471         lu_object_header_print(env, cookie, printer, top);
 472         (*printer)(env, cookie, "{ \n");
 473         cfs_list_for_each_entry(o, &top->loh_layers, lo_linkage) {
 474                 depth = o->lo_depth + 4;
 475
 476                 /*
 477                  * print `.' \a depth times followed by type name and address
 478                  */
 479                 (*printer)(env, cookie, "%*.*s%s@%p", depth, depth, ruler,
 480                            o->lo_dev->ld_type->ldt_name, o);
 481                 if (o->lo_ops->loo_object_print != NULL)
 482                         o->lo_ops->loo_object_print(env, cookie, printer, o);
 483                 (*printer)(env, cookie, "\n");
 484         }
 485         (*printer)(env, cookie, "} header@%p\n", top);
 486 }
 487 EXPORT_SYMBOL(lu_object_print);
 488
 489 /**
 490  * Check object consistency.
 491  */
 492 int lu_object_invariant(const struct lu_object *o)
 493 {
 494         struct lu_object_header *top;
 495
 496         top = o->lo_header;
 497         cfs_list_for_each_entry(o, &top->loh_layers, lo_linkage) {
 498                 if (o->lo_ops->loo_object_invariant != NULL &&
 499                     !o->lo_ops->loo_object_invariant(o))
 500                         return 0;
 501         }
 502         return 1;
 503 }
 504 EXPORT_SYMBOL(lu_object_invariant);
 505
 506 static struct lu_object *htable_lookup(struct lu_site *s,
 507                                        cfs_hash_bd_t *bd,
 508                                        const struct lu_fid *f,
 509                                        cfs_waitlink_t *waiter,
 510                                        __u64 *version)
 511 {
 512         struct lu_site_bkt_data *bkt;
 513         struct lu_object_header *h;
 514         cfs_hlist_node_t        *hnode;
 515         __u64  ver = cfs_hash_bd_version_get(bd);
 516
 517         if (*version == ver)
 518                 return NULL;
 519
 520         *version = ver;
 521         bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, bd);
 522         /* cfs_hash_bd_peek_locked is a somehow "internal" function
 523          * of cfs_hash, it doesn't add refcount on object. */
 524         hnode = cfs_hash_bd_peek_locked(s->ls_obj_hash, bd, (void *)f);
 525         if (hnode == NULL) {
 526                 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_MISS);
 527                 return NULL;
 528         }
 529
 530         h = container_of0(hnode, struct lu_object_header, loh_hash);
 531         if (likely(!lu_object_is_dying(h))) {
 532                 cfs_hash_get(s->ls_obj_hash, hnode);
 533                 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_HIT);
 534                 cfs_list_del_init(&h->loh_lru);
 535                 return lu_object_top(h);
 536         }
 537
 538         /*
 539          * Lookup found an object being destroyed this object cannot be
 540          * returned (to assure that references to dying objects are eventually
 541          * drained), and moreover, lookup has to wait until object is freed.
 542          */
 543
 544         cfs_waitlink_init(waiter);
 545         cfs_waitq_add(&bkt->lsb_marche_funebre, waiter);
 546         cfs_set_current_state(CFS_TASK_UNINT);
 547         lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_DEATH_RACE);
 548         return ERR_PTR(-EAGAIN);
 549 }
 550
 551 /**
 552  * Search cache for an object with the fid \a f. If such object is found,
 553  * return it. Otherwise, create new object, insert it into cache and return
 554  * it. In any case, additional reference is acquired on the returned object.
 555  */
 556 struct lu_object *lu_object_find(const struct lu_env *env,
 557                                  struct lu_device *dev, const struct lu_fid *f,
 558                                  const struct lu_object_conf *conf)
 559 {
 560         return lu_object_find_at(env, dev->ld_site->ls_top_dev, f, conf);
 561 }
 562 EXPORT_SYMBOL(lu_object_find);
 563
 564 static struct lu_object *lu_object_new(const struct lu_env *env,
 565                                        struct lu_device *dev,
 566                                        const struct lu_fid *f,
 567                                        const struct lu_object_conf *conf)
 568 {
 569         struct lu_object        *o;
 570         cfs_hash_t              *hs;
 571         cfs_hash_bd_t            bd;
 572         struct lu_site_bkt_data *bkt;
 573
 574         o = lu_object_alloc(env, dev, f, conf);
 575         if (unlikely(IS_ERR(o)))
 576                 return o;
 577
 578         hs = dev->ld_site->ls_obj_hash;
 579         cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
 580         bkt = cfs_hash_bd_extra_get(hs, &bd);
 581         cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
 582         bkt->lsb_busy++;
 583         cfs_hash_bd_unlock(hs, &bd, 1);
 584         return o;
 585 }
 586
 587 /**
 588  * Core logic of lu_object_find*() functions.
 589  */
 590 static struct lu_object *lu_object_find_try(const struct lu_env *env,
 591                                             struct lu_device *dev,
 592                                             const struct lu_fid *f,
 593                                             const struct lu_object_conf *conf,
 594                                             cfs_waitlink_t *waiter)
 595 {
 596         struct lu_object      *o;
 597         struct lu_object      *shadow;
 598         struct lu_site        *s;
 599         cfs_hash_t            *hs;
 600         cfs_hash_bd_t          bd;
 601         __u64                  version = 0;
 602
 603         /*
 604          * This uses standard index maintenance protocol:
 605          *
 606          *     - search index under lock, and return object if found;
 607          *     - otherwise, unlock index, allocate new object;
 608          *     - lock index and search again;
 609          *     - if nothing is found (usual case), insert newly created
 610          *       object into index;
 611          *     - otherwise (race: other thread inserted object), free
 612          *       object just allocated.
 613          *     - unlock index;
 614          *     - return object.
 615          *
 616          * For "LOC_F_NEW" case, we are sure the object is new established.
 617          * It is unnecessary to perform lookup-alloc-lookup-insert, instead,
 618          * just alloc and insert directly.
 619          *
 620          * If dying object is found during index search, add @waiter to the
 621          * site wait-queue and return ERR_PTR(-EAGAIN).
 622          */
 623         if (conf != NULL && conf->loc_flags & LOC_F_NEW)
 624                 return lu_object_new(env, dev, f, conf);
 625
 626         s  = dev->ld_site;
 627         hs = s->ls_obj_hash;
 628         cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
 629         o = htable_lookup(s, &bd, f, waiter, &version);
 630         cfs_hash_bd_unlock(hs, &bd, 1);
 631         if (o != NULL)
 632                 return o;
 633
 634         /*
 635          * Allocate new object. This may result in rather complicated
 636          * operations, including fld queries, inode loading, etc.
 637          */
 638         o = lu_object_alloc(env, dev, f, conf);
 639         if (unlikely(IS_ERR(o)))
 640                 return o;
 641
 642         LASSERT(lu_fid_eq(lu_object_fid(o), f));
 643
 644         cfs_hash_bd_lock(hs, &bd, 1);
 645
 646         shadow = htable_lookup(s, &bd, f, waiter, &version);
 647         if (likely(shadow == NULL)) {
 648                 struct lu_site_bkt_data *bkt;
 649
 650                 bkt = cfs_hash_bd_extra_get(hs, &bd);
 651                 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
 652                 bkt->lsb_busy++;
 653                 cfs_hash_bd_unlock(hs, &bd, 1);
 654                 return o;
 655         }
 656
 657         lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_RACE);
 658         cfs_hash_bd_unlock(hs, &bd, 1);
 659         lu_object_free(env, o);
 660         return shadow;
 661 }
 662
 663 /**
 664  * Much like lu_object_find(), but top level device of object is specifically
 665  * \a dev rather than top level device of the site. This interface allows
 666  * objects of different "stacking" to be created within the same site.
 667  */
 668 struct lu_object *lu_object_find_at(const struct lu_env *env,
 669                                     struct lu_device *dev,
 670                                     const struct lu_fid *f,
 671                                     const struct lu_object_conf *conf)
 672 {
 673         struct lu_site_bkt_data *bkt;
 674         struct lu_object        *obj;
 675         cfs_waitlink_t           wait;
 676
 677         while (1) {
 678                 obj = lu_object_find_try(env, dev, f, conf, &wait);
 679                 if (obj != ERR_PTR(-EAGAIN))
 680                         return obj;
 681                 /*
 682                  * lu_object_find_try() already added waiter into the
 683                  * wait queue.
 684                  */
 685                 cfs_waitq_wait(&wait, CFS_TASK_UNINT);
 686                 bkt = lu_site_bkt_from_fid(dev->ld_site, (void *)f);
 687                 cfs_waitq_del(&bkt->lsb_marche_funebre, &wait);
 688         }
 689 }
 690 EXPORT_SYMBOL(lu_object_find_at);
 691
 692 /**
 693  * Find object with given fid, and return its slice belonging to given device.
 694  */
 695 struct lu_object *lu_object_find_slice(const struct lu_env *env,
 696                                        struct lu_device *dev,
 697                                        const struct lu_fid *f,
 698                                        const struct lu_object_conf *conf)
 699 {
 700         struct lu_object *top;
 701         struct lu_object *obj;
 702
 703         top = lu_object_find(env, dev, f, conf);
 704         if (!IS_ERR(top)) {
 705                 obj = lu_object_locate(top->lo_header, dev->ld_type);
 706                 if (obj == NULL)
 707                         lu_object_put(env, top);
 708         } else
 709                 obj = top;
 710         return obj;
 711 }
 712 EXPORT_SYMBOL(lu_object_find_slice);
 713
 714 /**
 715  * Global list of all device types.
 716  */
 717 static CFS_LIST_HEAD(lu_device_types);
 718
 719 int lu_device_type_init(struct lu_device_type *ldt)
 720 {
 721         int result = 0;
 722
 723         CFS_INIT_LIST_HEAD(&ldt->ldt_linkage);
 724         if (ldt->ldt_ops->ldto_init)
 725                 result = ldt->ldt_ops->ldto_init(ldt);
 726         if (result == 0)
 727                 cfs_list_add(&ldt->ldt_linkage, &lu_device_types);
 728         return result;
 729 }
 730 EXPORT_SYMBOL(lu_device_type_init);
 731
 732 void lu_device_type_fini(struct lu_device_type *ldt)
 733 {
 734         cfs_list_del_init(&ldt->ldt_linkage);
 735         if (ldt->ldt_ops->ldto_fini)
 736                 ldt->ldt_ops->ldto_fini(ldt);
 737 }
 738 EXPORT_SYMBOL(lu_device_type_fini);
 739
 740 void lu_types_stop(void)
 741 {
 742         struct lu_device_type *ldt;
 743
 744         cfs_list_for_each_entry(ldt, &lu_device_types, ldt_linkage) {
 745                 if (ldt->ldt_device_nr == 0 && ldt->ldt_ops->ldto_stop)
 746                         ldt->ldt_ops->ldto_stop(ldt);
 747         }
 748 }
 749 EXPORT_SYMBOL(lu_types_stop);
 750
 751 /**
 752  * Global list of all sites on this node
 753  */
 754 static CFS_LIST_HEAD(lu_sites);
 755 static DEFINE_MUTEX(lu_sites_guard);
 756
 757 /**
 758  * Global environment used by site shrinker.
 759  */
 760 static struct lu_env lu_shrink_env;
 761
 762 struct lu_site_print_arg {
 763         struct lu_env   *lsp_env;
 764         void            *lsp_cookie;
 765         lu_printer_t     lsp_printer;
 766 };
 767
 768 static int
 769 lu_site_obj_print(cfs_hash_t *hs, cfs_hash_bd_t *bd,
 770                   cfs_hlist_node_t *hnode, void *data)
 771 {
 772         struct lu_site_print_arg *arg = (struct lu_site_print_arg *)data;
 773         struct lu_object_header  *h;
 774
 775         h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
 776         if (!cfs_list_empty(&h->loh_layers)) {
 777                 const struct lu_object *o;
 778
 779                 o = lu_object_top(h);
 780                 lu_object_print(arg->lsp_env, arg->lsp_cookie,
 781                                 arg->lsp_printer, o);
 782         } else {
 783                 lu_object_header_print(arg->lsp_env, arg->lsp_cookie,
 784                                        arg->lsp_printer, h);
 785         }
 786         return 0;
 787 }
 788
 789 /**
 790  * Print all objects in \a s.
 791  */
 792 void lu_site_print(const struct lu_env *env, struct lu_site *s, void *cookie,
 793                    lu_printer_t printer)
 794 {
 795         struct lu_site_print_arg arg = {
 796                 .lsp_env     = (struct lu_env *)env,
 797                 .lsp_cookie  = cookie,
 798                 .lsp_printer = printer,
 799         };
 800
 801         cfs_hash_for_each(s->ls_obj_hash, lu_site_obj_print, &arg);
 802 }
 803 EXPORT_SYMBOL(lu_site_print);
 804
 805 enum {
 806         LU_CACHE_PERCENT_MAX     = 50,
 807         LU_CACHE_PERCENT_DEFAULT = 20
 808 };
 809
 810 static unsigned int lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
 811 CFS_MODULE_PARM(lu_cache_percent, "i", int, 0644,
 812                 "Percentage of memory to be used as lu_object cache");
 813
 814 /**
 815  * Return desired hash table order.
 816  */
 817 static int lu_htable_order(void)
 818 {
 819         unsigned long cache_size;
 820         int bits;
 821
 822         /*
 823          * Calculate hash table size, assuming that we want reasonable
 824          * performance when 20% of total memory is occupied by cache of
 825          * lu_objects.
 826          *
 827          * Size of lu_object is (arbitrary) taken as 1K (together with inode).
 828          */
 829         cache_size = cfs_num_physpages;
 830
 831 #if BITS_PER_LONG == 32
 832         /* limit hashtable size for lowmem systems to low RAM */
 833         if (cache_size > 1 << (30 - CFS_PAGE_SHIFT))
 834                 cache_size = 1 << (30 - CFS_PAGE_SHIFT) * 3 / 4;
 835 #endif
 836
 837         /* clear off unreasonable cache setting. */
 838         if (lu_cache_percent == 0 || lu_cache_percent > LU_CACHE_PERCENT_MAX) {
 839                 CWARN("obdclass: invalid lu_cache_percent: %u, it must be in"
 840                       " the range of (0, %u]. Will use default value: %u.\n",
 841                       lu_cache_percent, LU_CACHE_PERCENT_MAX,
 842                       LU_CACHE_PERCENT_DEFAULT);
 843
 844                 lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
 845         }
 846         cache_size = cache_size / 100 * lu_cache_percent *
 847                 (CFS_PAGE_SIZE / 1024);
 848
 849         for (bits = 1; (1 << bits) < cache_size; ++bits) {
 850                 ;
 851         }
 852         return bits;
 853 }
 854
 855 static unsigned lu_obj_hop_hash(cfs_hash_t *hs,
 856                                 const void *key, unsigned mask)
 857 {
 858         struct lu_fid  *fid = (struct lu_fid *)key;
 859         __u32           hash;
 860
 861         hash = fid_flatten32(fid);
 862         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
 863         hash = cfs_hash_long(hash, hs->hs_bkt_bits);
 864
 865         /* give me another random factor */
 866         hash -= cfs_hash_long((unsigned long)hs, fid_oid(fid) % 11 + 3);
 867
 868         hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
 869         hash |= (fid_seq(fid) + fid_oid(fid)) & (CFS_HASH_NBKT(hs) - 1);
 870
 871         return hash & mask;
 872 }
 873
 874 static void *lu_obj_hop_object(cfs_hlist_node_t *hnode)
 875 {
 876         return cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
 877 }
 878
 879 static void *lu_obj_hop_key(cfs_hlist_node_t *hnode)
 880 {
 881         struct lu_object_header *h;
 882
 883         h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
 884         return &h->loh_fid;
 885 }
 886
 887 static int lu_obj_hop_keycmp(const void *key, cfs_hlist_node_t *hnode)
 888 {
 889         struct lu_object_header *h;
 890
 891         h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
 892         return lu_fid_eq(&h->loh_fid, (struct lu_fid *)key);
 893 }
 894
 895 static void lu_obj_hop_get(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 896 {
 897         struct lu_object_header *h;
 898
 899         h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
 900         if (cfs_atomic_add_return(1, &h->loh_ref) == 1) {
 901                 struct lu_site_bkt_data *bkt;
 902                 cfs_hash_bd_t            bd;
 903
 904                 cfs_hash_bd_get(hs, &h->loh_fid, &bd);
 905                 bkt = cfs_hash_bd_extra_get(hs, &bd);
 906                 bkt->lsb_busy++;
 907         }
 908 }
 909
 910 static void lu_obj_hop_put_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 911 {
 912         LBUG(); /* we should never called it */
 913 }
 914
 915 cfs_hash_ops_t lu_site_hash_ops = {
 916         .hs_hash        = lu_obj_hop_hash,
 917         .hs_key         = lu_obj_hop_key,
 918         .hs_keycmp      = lu_obj_hop_keycmp,
 919         .hs_object      = lu_obj_hop_object,
 920         .hs_get         = lu_obj_hop_get,
 921         .hs_put_locked  = lu_obj_hop_put_locked,
 922 };
 923
 924 void lu_dev_add_linkage(struct lu_site *s, struct lu_device *d)
 925 {
 926         spin_lock(&s->ls_ld_lock);
 927         if (cfs_list_empty(&d->ld_linkage))
 928                 cfs_list_add(&d->ld_linkage, &s->ls_ld_linkage);
 929         spin_unlock(&s->ls_ld_lock);
 930 }
 931 EXPORT_SYMBOL(lu_dev_add_linkage);
 932
 933 void lu_dev_del_linkage(struct lu_site *s, struct lu_device *d)
 934 {
 935         spin_lock(&s->ls_ld_lock);
 936         cfs_list_del_init(&d->ld_linkage);
 937         spin_unlock(&s->ls_ld_lock);
 938 }
 939 EXPORT_SYMBOL(lu_dev_del_linkage);
 940
 941 /**
 942  * Initialize site \a s, with \a d as the top level device.
 943  */
 944 #define LU_SITE_BITS_MIN    12
 945 #define LU_SITE_BITS_MAX    24
 946 /**
 947  * total 256 buckets, we don't want too many buckets because:
 948  * - consume too much memory
 949  * - avoid unbalanced LRU list
 950  */
 951 #define LU_SITE_BKT_BITS    8
 952
 953 int lu_site_init(struct lu_site *s, struct lu_device *top)
 954 {
 955         struct lu_site_bkt_data *bkt;
 956         cfs_hash_bd_t bd;
 957         char name[16];
 958         int bits;
 959         int i;
 960         ENTRY;
 961
 962         memset(s, 0, sizeof *s);
 963         bits = lu_htable_order();
 964         snprintf(name, 16, "lu_site_%s", top->ld_type->ldt_name);
 965         for (bits = min(max(LU_SITE_BITS_MIN, bits), LU_SITE_BITS_MAX);
 966              bits >= LU_SITE_BITS_MIN; bits--) {
 967                 s->ls_obj_hash = cfs_hash_create(name, bits, bits,
 968                                                  bits - LU_SITE_BKT_BITS,
 969                                                  sizeof(*bkt), 0, 0,
 970                                                  &lu_site_hash_ops,
 971                                                  CFS_HASH_SPIN_BKTLOCK |
 972                                                  CFS_HASH_NO_ITEMREF |
 973                                                  CFS_HASH_DEPTH |
 974                                                  CFS_HASH_ASSERT_EMPTY);
 975                 if (s->ls_obj_hash != NULL)
 976                         break;
 977         }
 978
 979         if (s->ls_obj_hash == NULL) {
 980                 CERROR("failed to create lu_site hash with bits: %d\n", bits);
 981                 return -ENOMEM;
 982         }
 983
 984         cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
 985                 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
 986                 CFS_INIT_LIST_HEAD(&bkt->lsb_lru);
 987                 cfs_waitq_init(&bkt->lsb_marche_funebre);
 988         }
 989
 990         s->ls_stats = lprocfs_alloc_stats(LU_SS_LAST_STAT, 0);
 991         if (s->ls_stats == NULL) {
 992                 cfs_hash_putref(s->ls_obj_hash);
 993                 s->ls_obj_hash = NULL;
 994                 return -ENOMEM;
 995         }
 996
 997         lprocfs_counter_init(s->ls_stats, LU_SS_CREATED,
 998                              0, "created", "created");
 999         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_HIT,
1000                              0, "cache_hit", "cache_hit");
1001         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_MISS,
1002                              0, "cache_miss", "cache_miss");
1003         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_RACE,
1004                              0, "cache_race", "cache_race");
1005         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_DEATH_RACE,
1006                              0, "cache_death_race", "cache_death_race");
1007         lprocfs_counter_init(s->ls_stats, LU_SS_LRU_PURGED,
1008                              0, "lru_purged", "lru_purged");
1009
1010         CFS_INIT_LIST_HEAD(&s->ls_linkage);
1011         s->ls_top_dev = top;
1012         top->ld_site = s;
1013         lu_device_get(top);
1014         lu_ref_add(&top->ld_reference, "site-top", s);
1015
1016         CFS_INIT_LIST_HEAD(&s->ls_ld_linkage);
1017         spin_lock_init(&s->ls_ld_lock);
1018
1019         lu_dev_add_linkage(s, top);
1020
1021         RETURN(0);
1022 }
1023 EXPORT_SYMBOL(lu_site_init);
1024
1025 /**
1026  * Finalize \a s and release its resources.
1027  */
1028 void lu_site_fini(struct lu_site *s)
1029 {
1030         mutex_lock(&lu_sites_guard);
1031         cfs_list_del_init(&s->ls_linkage);
1032         mutex_unlock(&lu_sites_guard);
1033
1034         if (s->ls_obj_hash != NULL) {
1035                 cfs_hash_putref(s->ls_obj_hash);
1036                 s->ls_obj_hash = NULL;
1037         }
1038
1039         if (s->ls_top_dev != NULL) {
1040                 s->ls_top_dev->ld_site = NULL;
1041                 lu_ref_del(&s->ls_top_dev->ld_reference, "site-top", s);
1042                 lu_device_put(s->ls_top_dev);
1043                 s->ls_top_dev = NULL;
1044         }
1045
1046         if (s->ls_stats != NULL)
1047                 lprocfs_free_stats(&s->ls_stats);
1048 }
1049 EXPORT_SYMBOL(lu_site_fini);
1050
1051 /**
1052  * Called when initialization of stack for this site is completed.
1053  */
1054 int lu_site_init_finish(struct lu_site *s)
1055 {
1056         int result;
1057         mutex_lock(&lu_sites_guard);
1058         result = lu_context_refill(&lu_shrink_env.le_ctx);
1059         if (result == 0)
1060                 cfs_list_add(&s->ls_linkage, &lu_sites);
1061         mutex_unlock(&lu_sites_guard);
1062         return result;
1063 }
1064 EXPORT_SYMBOL(lu_site_init_finish);
1065
1066 /**
1067  * Acquire additional reference on device \a d
1068  */
1069 void lu_device_get(struct lu_device *d)
1070 {
1071         cfs_atomic_inc(&d->ld_ref);
1072 }
1073 EXPORT_SYMBOL(lu_device_get);
1074
1075 /**
1076  * Release reference on device \a d.
1077  */
1078 void lu_device_put(struct lu_device *d)
1079 {
1080         LASSERT(cfs_atomic_read(&d->ld_ref) > 0);
1081         cfs_atomic_dec(&d->ld_ref);
1082 }
1083 EXPORT_SYMBOL(lu_device_put);
1084
1085 /**
1086  * Initialize device \a d of type \a t.
1087  */
1088 int lu_device_init(struct lu_device *d, struct lu_device_type *t)
1089 {
1090         if (t->ldt_device_nr++ == 0 && t->ldt_ops->ldto_start != NULL)
1091                 t->ldt_ops->ldto_start(t);
1092         memset(d, 0, sizeof *d);
1093         cfs_atomic_set(&d->ld_ref, 0);
1094         d->ld_type = t;
1095         lu_ref_init(&d->ld_reference);
1096         CFS_INIT_LIST_HEAD(&d->ld_linkage);
1097         return 0;
1098 }
1099 EXPORT_SYMBOL(lu_device_init);
1100
1101 /**
1102  * Finalize device \a d.
1103  */
1104 void lu_device_fini(struct lu_device *d)
1105 {
1106         struct lu_device_type *t;
1107
1108         t = d->ld_type;
1109         if (d->ld_obd != NULL) {
1110                 d->ld_obd->obd_lu_dev = NULL;
1111                 d->ld_obd = NULL;
1112         }
1113
1114         lu_ref_fini(&d->ld_reference);
1115         LASSERTF(cfs_atomic_read(&d->ld_ref) == 0,
1116                  "Refcount is %u\n", cfs_atomic_read(&d->ld_ref));
1117         LASSERT(t->ldt_device_nr > 0);
1118         if (--t->ldt_device_nr == 0 && t->ldt_ops->ldto_stop != NULL)
1119                 t->ldt_ops->ldto_stop(t);
1120 }
1121 EXPORT_SYMBOL(lu_device_fini);
1122
1123 /**
1124  * Initialize object \a o that is part of compound object \a h and was created
1125  * by device \a d.
1126  */
1127 int lu_object_init(struct lu_object *o,
1128                    struct lu_object_header *h, struct lu_device *d)
1129 {
1130         memset(o, 0, sizeof *o);
1131         o->lo_header = h;
1132         o->lo_dev    = d;
1133         lu_device_get(d);
1134         o->lo_dev_ref = lu_ref_add(&d->ld_reference, "lu_object", o);
1135         CFS_INIT_LIST_HEAD(&o->lo_linkage);
1136         return 0;
1137 }
1138 EXPORT_SYMBOL(lu_object_init);
1139
1140 /**
1141  * Finalize object and release its resources.
1142  */
1143 void lu_object_fini(struct lu_object *o)
1144 {
1145         struct lu_device *dev = o->lo_dev;
1146
1147         LASSERT(cfs_list_empty(&o->lo_linkage));
1148
1149         if (dev != NULL) {
1150                 lu_ref_del_at(&dev->ld_reference,
1151                               o->lo_dev_ref , "lu_object", o);
1152                 lu_device_put(dev);
1153                 o->lo_dev = NULL;
1154         }
1155 }
1156 EXPORT_SYMBOL(lu_object_fini);
1157
1158 /**
1159  * Add object \a o as first layer of compound object \a h
1160  *
1161  * This is typically called by the ->ldo_object_alloc() method of top-level
1162  * device.
1163  */
1164 void lu_object_add_top(struct lu_object_header *h, struct lu_object *o)
1165 {
1166         cfs_list_move(&o->lo_linkage, &h->loh_layers);
1167 }
1168 EXPORT_SYMBOL(lu_object_add_top);
1169
1170 /**
1171  * Add object \a o as a layer of compound object, going after \a before.
1172  *
1173  * This is typically called by the ->ldo_object_alloc() method of \a
1174  * before->lo_dev.
1175  */
1176 void lu_object_add(struct lu_object *before, struct lu_object *o)
1177 {
1178         cfs_list_move(&o->lo_linkage, &before->lo_linkage);
1179 }
1180 EXPORT_SYMBOL(lu_object_add);
1181
1182 /**
1183  * Initialize compound object.
1184  */
1185 int lu_object_header_init(struct lu_object_header *h)
1186 {
1187         memset(h, 0, sizeof *h);
1188         cfs_atomic_set(&h->loh_ref, 1);
1189         CFS_INIT_HLIST_NODE(&h->loh_hash);
1190         CFS_INIT_LIST_HEAD(&h->loh_lru);
1191         CFS_INIT_LIST_HEAD(&h->loh_layers);
1192         lu_ref_init(&h->loh_reference);
1193         return 0;
1194 }
1195 EXPORT_SYMBOL(lu_object_header_init);
1196
1197 /**
1198  * Finalize compound object.
1199  */
1200 void lu_object_header_fini(struct lu_object_header *h)
1201 {
1202         LASSERT(cfs_list_empty(&h->loh_layers));
1203         LASSERT(cfs_list_empty(&h->loh_lru));
1204         LASSERT(cfs_hlist_unhashed(&h->loh_hash));
1205         lu_ref_fini(&h->loh_reference);
1206 }
1207 EXPORT_SYMBOL(lu_object_header_fini);
1208
1209 /**
1210  * Given a compound object, find its slice, corresponding to the device type
1211  * \a dtype.
1212  */
1213 struct lu_object *lu_object_locate(struct lu_object_header *h,
1214                                    const struct lu_device_type *dtype)
1215 {
1216         struct lu_object *o;
1217
1218         cfs_list_for_each_entry(o, &h->loh_layers, lo_linkage) {
1219                 if (o->lo_dev->ld_type == dtype)
1220                         return o;
1221         }
1222         return NULL;
1223 }
1224 EXPORT_SYMBOL(lu_object_locate);
1225
1226
1227
1228 /**
1229  * Finalize and free devices in the device stack.
1230  *
1231  * Finalize device stack by purging object cache, and calling
1232  * lu_device_type_operations::ldto_device_fini() and
1233  * lu_device_type_operations::ldto_device_free() on all devices in the stack.
1234  */
1235 void lu_stack_fini(const struct lu_env *env, struct lu_device *top)
1236 {
1237         struct lu_site   *site = top->ld_site;
1238         struct lu_device *scan;
1239         struct lu_device *next;
1240
1241         lu_site_purge(env, site, ~0);
1242         for (scan = top; scan != NULL; scan = next) {
1243                 next = scan->ld_type->ldt_ops->ldto_device_fini(env, scan);
1244                 lu_ref_del(&scan->ld_reference, "lu-stack", &lu_site_init);
1245                 lu_device_put(scan);
1246         }
1247
1248         /* purge again. */
1249         lu_site_purge(env, site, ~0);
1250
1251         for (scan = top; scan != NULL; scan = next) {
1252                 const struct lu_device_type *ldt = scan->ld_type;
1253                 struct obd_type             *type;
1254
1255                 next = ldt->ldt_ops->ldto_device_free(env, scan);
1256                 type = ldt->ldt_obd_type;
1257                 if (type != NULL) {
1258                         type->typ_refcnt--;
1259                         class_put_type(type);
1260                 }
1261         }
1262 }
1263 EXPORT_SYMBOL(lu_stack_fini);
1264
1265 enum {
1266         /**
1267          * Maximal number of tld slots.
1268          */
1269         LU_CONTEXT_KEY_NR = 40
1270 };
1271
1272 static struct lu_context_key *lu_keys[LU_CONTEXT_KEY_NR] = { NULL, };
1273
1274 static DEFINE_SPINLOCK(lu_keys_guard);
1275
1276 /**
1277  * Global counter incremented whenever key is registered, unregistered,
1278  * revived or quiesced. This is used to void unnecessary calls to
1279  * lu_context_refill(). No locking is provided, as initialization and shutdown
1280  * are supposed to be externally serialized.
1281  */
1282 static unsigned key_set_version = 0;
1283
1284 /**
1285  * Register new key.
1286  */
1287 int lu_context_key_register(struct lu_context_key *key)
1288 {
1289         int result;
1290         int i;
1291
1292         LASSERT(key->lct_init != NULL);
1293         LASSERT(key->lct_fini != NULL);
1294         LASSERT(key->lct_tags != 0);
1295         LASSERT(key->lct_owner != NULL);
1296
1297         result = -ENFILE;
1298         spin_lock(&lu_keys_guard);
1299         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1300                 if (lu_keys[i] == NULL) {
1301                         key->lct_index = i;
1302                         cfs_atomic_set(&key->lct_used, 1);
1303                         lu_keys[i] = key;
1304                         lu_ref_init(&key->lct_reference);
1305                         result = 0;
1306                         ++key_set_version;
1307                         break;
1308                 }
1309         }
1310         spin_unlock(&lu_keys_guard);
1311         return result;
1312 }
1313 EXPORT_SYMBOL(lu_context_key_register);
1314
1315 static void key_fini(struct lu_context *ctx, int index)
1316 {
1317         if (ctx->lc_value != NULL && ctx->lc_value[index] != NULL) {
1318                 struct lu_context_key *key;
1319
1320                 key = lu_keys[index];
1321                 LASSERT(key != NULL);
1322                 LASSERT(key->lct_fini != NULL);
1323                 LASSERT(cfs_atomic_read(&key->lct_used) > 1);
1324
1325                 key->lct_fini(ctx, key, ctx->lc_value[index]);
1326                 lu_ref_del(&key->lct_reference, "ctx", ctx);
1327                 cfs_atomic_dec(&key->lct_used);
1328
1329                 LASSERT(key->lct_owner != NULL);
1330                 if ((ctx->lc_tags & LCT_NOREF) == 0) {
1331                         LINVRNT(cfs_module_refcount(key->lct_owner) > 0);
1332                         cfs_module_put(key->lct_owner);
1333                 }
1334                 ctx->lc_value[index] = NULL;
1335         }
1336 }
1337
1338 /**
1339  * Deregister key.
1340  */
1341 void lu_context_key_degister(struct lu_context_key *key)
1342 {
1343         LASSERT(cfs_atomic_read(&key->lct_used) >= 1);
1344         LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1345
1346         lu_context_key_quiesce(key);
1347
1348         ++key_set_version;
1349         spin_lock(&lu_keys_guard);
1350         key_fini(&lu_shrink_env.le_ctx, key->lct_index);
1351         if (lu_keys[key->lct_index]) {
1352                 lu_keys[key->lct_index] = NULL;
1353                 lu_ref_fini(&key->lct_reference);
1354         }
1355         spin_unlock(&lu_keys_guard);
1356
1357         LASSERTF(cfs_atomic_read(&key->lct_used) == 1,
1358                  "key has instances: %d\n",
1359                  cfs_atomic_read(&key->lct_used));
1360 }
1361 EXPORT_SYMBOL(lu_context_key_degister);
1362
1363 /**
1364  * Register a number of keys. This has to be called after all keys have been
1365  * initialized by a call to LU_CONTEXT_KEY_INIT().
1366  */
1367 int lu_context_key_register_many(struct lu_context_key *k, ...)
1368 {
1369         struct lu_context_key *key = k;
1370         va_list args;
1371         int result;
1372
1373         va_start(args, k);
1374         do {
1375                 result = lu_context_key_register(key);
1376                 if (result)
1377                         break;
1378                 key = va_arg(args, struct lu_context_key *);
1379         } while (key != NULL);
1380         va_end(args);
1381
1382         if (result != 0) {
1383                 va_start(args, k);
1384                 while (k != key) {
1385                         lu_context_key_degister(k);
1386                         k = va_arg(args, struct lu_context_key *);
1387                 }
1388                 va_end(args);
1389         }
1390
1391         return result;
1392 }
1393 EXPORT_SYMBOL(lu_context_key_register_many);
1394
1395 /**
1396  * De-register a number of keys. This is a dual to
1397  * lu_context_key_register_many().
1398  */
1399 void lu_context_key_degister_many(struct lu_context_key *k, ...)
1400 {
1401         va_list args;
1402
1403         va_start(args, k);
1404         do {
1405                 lu_context_key_degister(k);
1406                 k = va_arg(args, struct lu_context_key*);
1407         } while (k != NULL);
1408         va_end(args);
1409 }
1410 EXPORT_SYMBOL(lu_context_key_degister_many);
1411
1412 /**
1413  * Revive a number of keys.
1414  */
1415 void lu_context_key_revive_many(struct lu_context_key *k, ...)
1416 {
1417         va_list args;
1418
1419         va_start(args, k);
1420         do {
1421                 lu_context_key_revive(k);
1422                 k = va_arg(args, struct lu_context_key*);
1423         } while (k != NULL);
1424         va_end(args);
1425 }
1426 EXPORT_SYMBOL(lu_context_key_revive_many);
1427
1428 /**
1429  * Quiescent a number of keys.
1430  */
1431 void lu_context_key_quiesce_many(struct lu_context_key *k, ...)
1432 {
1433         va_list args;
1434
1435         va_start(args, k);
1436         do {
1437                 lu_context_key_quiesce(k);
1438                 k = va_arg(args, struct lu_context_key*);
1439         } while (k != NULL);
1440         va_end(args);
1441 }
1442 EXPORT_SYMBOL(lu_context_key_quiesce_many);
1443
1444 /**
1445  * Return value associated with key \a key in context \a ctx.
1446  */
1447 void *lu_context_key_get(const struct lu_context *ctx,
1448                          const struct lu_context_key *key)
1449 {
1450         LINVRNT(ctx->lc_state == LCS_ENTERED);
1451         LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1452         LASSERT(lu_keys[key->lct_index] == key);
1453         return ctx->lc_value[key->lct_index];
1454 }
1455 EXPORT_SYMBOL(lu_context_key_get);
1456
1457 /**
1458  * List of remembered contexts. XXX document me.
1459  */
1460 static CFS_LIST_HEAD(lu_context_remembered);
1461
1462 /**
1463  * Destroy \a key in all remembered contexts. This is used to destroy key
1464  * values in "shared" contexts (like service threads), when a module owning
1465  * the key is about to be unloaded.
1466  */
1467 void lu_context_key_quiesce(struct lu_context_key *key)
1468 {
1469         struct lu_context *ctx;
1470         extern unsigned cl_env_cache_purge(unsigned nr);
1471
1472         if (!(key->lct_tags & LCT_QUIESCENT)) {
1473                 /*
1474                  * XXX layering violation.
1475                  */
1476                 cl_env_cache_purge(~0);
1477                 key->lct_tags |= LCT_QUIESCENT;
1478                 /*
1479                  * XXX memory barrier has to go here.
1480                  */
1481                 spin_lock(&lu_keys_guard);
1482                 cfs_list_for_each_entry(ctx, &lu_context_remembered,
1483                                         lc_remember)
1484                         key_fini(ctx, key->lct_index);
1485                 spin_unlock(&lu_keys_guard);
1486                 ++key_set_version;
1487         }
1488 }
1489 EXPORT_SYMBOL(lu_context_key_quiesce);
1490
1491 void lu_context_key_revive(struct lu_context_key *key)
1492 {
1493         key->lct_tags &= ~LCT_QUIESCENT;
1494         ++key_set_version;
1495 }
1496 EXPORT_SYMBOL(lu_context_key_revive);
1497
1498 static void keys_fini(struct lu_context *ctx)
1499 {
1500         int     i;
1501
1502         if (ctx->lc_value == NULL)
1503                 return;
1504
1505         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i)
1506                 key_fini(ctx, i);
1507
1508         OBD_FREE(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1509         ctx->lc_value = NULL;
1510 }
1511
1512 static int keys_fill(struct lu_context *ctx)
1513 {
1514         int i;
1515
1516         LINVRNT(ctx->lc_value != NULL);
1517         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1518                 struct lu_context_key *key;
1519
1520                 key = lu_keys[i];
1521                 if (ctx->lc_value[i] == NULL && key != NULL &&
1522                     (key->lct_tags & ctx->lc_tags) &&
1523                     /*
1524                      * Don't create values for a LCT_QUIESCENT key, as this
1525                      * will pin module owning a key.
1526                      */
1527                     !(key->lct_tags & LCT_QUIESCENT)) {
1528                         void *value;
1529
1530                         LINVRNT(key->lct_init != NULL);
1531                         LINVRNT(key->lct_index == i);
1532
1533                         value = key->lct_init(ctx, key);
1534                         if (unlikely(IS_ERR(value)))
1535                                 return PTR_ERR(value);
1536
1537                         LASSERT(key->lct_owner != NULL);
1538                         if (!(ctx->lc_tags & LCT_NOREF))
1539                                 cfs_try_module_get(key->lct_owner);
1540                         lu_ref_add_atomic(&key->lct_reference, "ctx", ctx);
1541                         cfs_atomic_inc(&key->lct_used);
1542                         /*
1543                          * This is the only place in the code, where an
1544                          * element of ctx->lc_value[] array is set to non-NULL
1545                          * value.
1546                          */
1547                         ctx->lc_value[i] = value;
1548                         if (key->lct_exit != NULL)
1549                                 ctx->lc_tags |= LCT_HAS_EXIT;
1550                 }
1551                 ctx->lc_version = key_set_version;
1552         }
1553         return 0;
1554 }
1555
1556 static int keys_init(struct lu_context *ctx)
1557 {
1558         OBD_ALLOC(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1559         if (likely(ctx->lc_value != NULL))
1560                 return keys_fill(ctx);
1561
1562         return -ENOMEM;
1563 }
1564
1565 /**
1566  * Initialize context data-structure. Create values for all keys.
1567  */
1568 int lu_context_init(struct lu_context *ctx, __u32 tags)
1569 {
1570         int     rc;
1571
1572         memset(ctx, 0, sizeof *ctx);
1573         ctx->lc_state = LCS_INITIALIZED;
1574         ctx->lc_tags = tags;
1575         if (tags & LCT_REMEMBER) {
1576                 spin_lock(&lu_keys_guard);
1577                 cfs_list_add(&ctx->lc_remember, &lu_context_remembered);
1578                 spin_unlock(&lu_keys_guard);
1579         } else {
1580                 CFS_INIT_LIST_HEAD(&ctx->lc_remember);
1581         }
1582
1583         rc = keys_init(ctx);
1584         if (rc != 0)
1585                 lu_context_fini(ctx);
1586
1587         return rc;
1588 }
1589 EXPORT_SYMBOL(lu_context_init);
1590
1591 /**
1592  * Finalize context data-structure. Destroy key values.
1593  */
1594 void lu_context_fini(struct lu_context *ctx)
1595 {
1596         LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1597         ctx->lc_state = LCS_FINALIZED;
1598
1599         if ((ctx->lc_tags & LCT_REMEMBER) == 0) {
1600                 LASSERT(cfs_list_empty(&ctx->lc_remember));
1601                 keys_fini(ctx);
1602
1603         } else { /* could race with key degister */
1604                 spin_lock(&lu_keys_guard);
1605                 keys_fini(ctx);
1606                 cfs_list_del_init(&ctx->lc_remember);
1607                 spin_unlock(&lu_keys_guard);
1608         }
1609 }
1610 EXPORT_SYMBOL(lu_context_fini);
1611
1612 /**
1613  * Called before entering context.
1614  */
1615 void lu_context_enter(struct lu_context *ctx)
1616 {
1617         LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1618         ctx->lc_state = LCS_ENTERED;
1619 }
1620 EXPORT_SYMBOL(lu_context_enter);
1621
1622 /**
1623  * Called after exiting from \a ctx
1624  */
1625 void lu_context_exit(struct lu_context *ctx)
1626 {
1627         int i;
1628
1629         LINVRNT(ctx->lc_state == LCS_ENTERED);
1630         ctx->lc_state = LCS_LEFT;
1631         if (ctx->lc_tags & LCT_HAS_EXIT && ctx->lc_value != NULL) {
1632                 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1633                         if (ctx->lc_value[i] != NULL) {
1634                                 struct lu_context_key *key;
1635
1636                                 key = lu_keys[i];
1637                                 LASSERT(key != NULL);
1638                                 if (key->lct_exit != NULL)
1639                                         key->lct_exit(ctx,
1640                                                       key, ctx->lc_value[i]);
1641                         }
1642                 }
1643         }
1644 }
1645 EXPORT_SYMBOL(lu_context_exit);
1646
1647 /**
1648  * Allocate for context all missing keys that were registered after context
1649  * creation. key_set_version is only changed in rare cases when modules
1650  * are loaded and removed.
1651  */
1652 int lu_context_refill(struct lu_context *ctx)
1653 {
1654         return likely(ctx->lc_version == key_set_version) ? 0 : keys_fill(ctx);
1655 }
1656 EXPORT_SYMBOL(lu_context_refill);
1657
1658 /**
1659  * lu_ctx_tags/lu_ses_tags will be updated if there are new types of
1660  * obd being added. Currently, this is only used on client side, specifically
1661  * for echo device client, for other stack (like ptlrpc threads), context are
1662  * predefined when the lu_device type are registered, during the module probe
1663  * phase.
1664  */
1665 __u32 lu_context_tags_default = 0;
1666 __u32 lu_session_tags_default = 0;
1667
1668 void lu_context_tags_update(__u32 tags)
1669 {
1670         spin_lock(&lu_keys_guard);
1671         lu_context_tags_default |= tags;
1672         key_set_version++;
1673         spin_unlock(&lu_keys_guard);
1674 }
1675 EXPORT_SYMBOL(lu_context_tags_update);
1676
1677 void lu_context_tags_clear(__u32 tags)
1678 {
1679         spin_lock(&lu_keys_guard);
1680         lu_context_tags_default &= ~tags;
1681         key_set_version++;
1682         spin_unlock(&lu_keys_guard);
1683 }
1684 EXPORT_SYMBOL(lu_context_tags_clear);
1685
1686 void lu_session_tags_update(__u32 tags)
1687 {
1688         spin_lock(&lu_keys_guard);
1689         lu_session_tags_default |= tags;
1690         key_set_version++;
1691         spin_unlock(&lu_keys_guard);
1692 }
1693 EXPORT_SYMBOL(lu_session_tags_update);
1694
1695 void lu_session_tags_clear(__u32 tags)
1696 {
1697         spin_lock(&lu_keys_guard);
1698         lu_session_tags_default &= ~tags;
1699         key_set_version++;
1700         spin_unlock(&lu_keys_guard);
1701 }
1702 EXPORT_SYMBOL(lu_session_tags_clear);
1703
1704 int lu_env_init(struct lu_env *env, __u32 tags)
1705 {
1706         int result;
1707
1708         env->le_ses = NULL;
1709         result = lu_context_init(&env->le_ctx, tags);
1710         if (likely(result == 0))
1711                 lu_context_enter(&env->le_ctx);
1712         return result;
1713 }
1714 EXPORT_SYMBOL(lu_env_init);
1715
1716 void lu_env_fini(struct lu_env *env)
1717 {
1718         lu_context_exit(&env->le_ctx);
1719         lu_context_fini(&env->le_ctx);
1720         env->le_ses = NULL;
1721 }
1722 EXPORT_SYMBOL(lu_env_fini);
1723
1724 int lu_env_refill(struct lu_env *env)
1725 {
1726         int result;
1727
1728         result = lu_context_refill(&env->le_ctx);
1729         if (result == 0 && env->le_ses != NULL)
1730                 result = lu_context_refill(env->le_ses);
1731         return result;
1732 }
1733 EXPORT_SYMBOL(lu_env_refill);
1734
1735 /**
1736  * Currently, this API will only be used by echo client.
1737  * Because echo client and normal lustre client will share
1738  * same cl_env cache. So echo client needs to refresh
1739  * the env context after it get one from the cache, especially
1740  * when normal client and echo client co-exist in the same client.
1741  */
1742 int lu_env_refill_by_tags(struct lu_env *env, __u32 ctags,
1743                           __u32 stags)
1744 {
1745         int    result;
1746
1747         if ((env->le_ctx.lc_tags & ctags) != ctags) {
1748                 env->le_ctx.lc_version = 0;
1749                 env->le_ctx.lc_tags |= ctags;
1750         }
1751
1752         if (env->le_ses && (env->le_ses->lc_tags & stags) != stags) {
1753                 env->le_ses->lc_version = 0;
1754                 env->le_ses->lc_tags |= stags;
1755         }
1756
1757         result = lu_env_refill(env);
1758
1759         return result;
1760 }
1761 EXPORT_SYMBOL(lu_env_refill_by_tags);
1762
1763 static struct cfs_shrinker *lu_site_shrinker = NULL;
1764
1765 typedef struct lu_site_stats{
1766         unsigned        lss_populated;
1767         unsigned        lss_max_search;
1768         unsigned        lss_total;
1769         unsigned        lss_busy;
1770 } lu_site_stats_t;
1771
1772 static void lu_site_stats_get(cfs_hash_t *hs,
1773                               lu_site_stats_t *stats, int populated)
1774 {
1775         cfs_hash_bd_t bd;
1776         int           i;
1777
1778         cfs_hash_for_each_bucket(hs, &bd, i) {
1779                 struct lu_site_bkt_data *bkt = cfs_hash_bd_extra_get(hs, &bd);
1780                 cfs_hlist_head_t        *hhead;
1781
1782                 cfs_hash_bd_lock(hs, &bd, 1);
1783                 stats->lss_busy  += bkt->lsb_busy;
1784                 stats->lss_total += cfs_hash_bd_count_get(&bd);
1785                 stats->lss_max_search = max((int)stats->lss_max_search,
1786                                             cfs_hash_bd_depmax_get(&bd));
1787                 if (!populated) {
1788                         cfs_hash_bd_unlock(hs, &bd, 1);
1789                         continue;
1790                 }
1791
1792                 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
1793                         if (!cfs_hlist_empty(hhead))
1794                                 stats->lss_populated++;
1795                 }
1796                 cfs_hash_bd_unlock(hs, &bd, 1);
1797         }
1798 }
1799
1800 #ifdef __KERNEL__
1801
1802 /*
1803  * There exists a potential lock inversion deadlock scenario when using
1804  * Lustre on top of ZFS. This occurs between one of ZFS's
1805  * buf_hash_table.ht_lock's, and Lustre's lu_sites_guard lock. Essentially,
1806  * thread A will take the lu_sites_guard lock and sleep on the ht_lock,
1807  * while thread B will take the ht_lock and sleep on the lu_sites_guard
1808  * lock. Obviously neither thread will wake and drop their respective hold
1809  * on their lock.
1810  *
1811  * To prevent this from happening we must ensure the lu_sites_guard lock is
1812  * not taken while down this code path. ZFS reliably does not set the
1813  * __GFP_FS bit in its code paths, so this can be used to determine if it
1814  * is safe to take the lu_sites_guard lock.
1815  *
1816  * Ideally we should accurately return the remaining number of cached
1817  * objects without taking the  lu_sites_guard lock, but this is not
1818  * possible in the current implementation.
1819  */
1820 static int lu_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
1821 {
1822         lu_site_stats_t stats;
1823         struct lu_site *s;
1824         struct lu_site *tmp;
1825         int cached = 0;
1826         int remain = shrink_param(sc, nr_to_scan);
1827         CFS_LIST_HEAD(splice);
1828
1829         if (!(shrink_param(sc, gfp_mask) & __GFP_FS)) {
1830                 if (remain != 0)
1831                         return -1;
1832                 else
1833                         /* We must not take the lu_sites_guard lock when
1834                          * __GFP_FS is *not* set because of the deadlock
1835                          * possibility detailed above. Additionally,
1836                          * since we cannot determine the number of
1837                          * objects in the cache without taking this
1838                          * lock, we're in a particularly tough spot. As
1839                          * a result, we'll just lie and say our cache is
1840                          * empty. This _should_ be ok, as we can't
1841                          * reclaim objects when __GFP_FS is *not* set
1842                          * anyways.
1843                          */
1844                         return 0;
1845         }
1846
1847         CDEBUG(D_INODE, "Shrink %d objects\n", remain);
1848
1849         mutex_lock(&lu_sites_guard);
1850         cfs_list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
1851                 if (shrink_param(sc, nr_to_scan) != 0) {
1852                         remain = lu_site_purge(&lu_shrink_env, s, remain);
1853                         /*
1854                          * Move just shrunk site to the tail of site list to
1855                          * assure shrinking fairness.
1856                          */
1857                         cfs_list_move_tail(&s->ls_linkage, &splice);
1858                 }
1859
1860                 memset(&stats, 0, sizeof(stats));
1861                 lu_site_stats_get(s->ls_obj_hash, &stats, 0);
1862                 cached += stats.lss_total - stats.lss_busy;
1863                 if (shrink_param(sc, nr_to_scan) && remain <= 0)
1864                         break;
1865         }
1866         cfs_list_splice(&splice, lu_sites.prev);
1867         mutex_unlock(&lu_sites_guard);
1868
1869         cached = (cached / 100) * sysctl_vfs_cache_pressure;
1870         if (shrink_param(sc, nr_to_scan) == 0)
1871                 CDEBUG(D_INODE, "%d objects cached\n", cached);
1872         return cached;
1873 }
1874
1875 /*
1876  * Debugging stuff.
1877  */
1878
1879 /**
1880  * Environment to be used in debugger, contains all tags.
1881  */
1882 struct lu_env lu_debugging_env;
1883
1884 /**
1885  * Debugging printer function using printk().
1886  */
1887 int lu_printk_printer(const struct lu_env *env,
1888                       void *unused, const char *format, ...)
1889 {
1890         va_list args;
1891
1892         va_start(args, format);
1893         vprintk(format, args);
1894         va_end(args);
1895         return 0;
1896 }
1897
1898 void lu_debugging_setup(void)
1899 {
1900         lu_env_init(&lu_debugging_env, ~0);
1901 }
1902
1903 void lu_context_keys_dump(void)
1904 {
1905         int i;
1906
1907         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1908                 struct lu_context_key *key;
1909
1910                 key = lu_keys[i];
1911                 if (key != NULL) {
1912                         CERROR("[%d]: %p %x (%p,%p,%p) %d %d \"%s\"@%p\n",
1913                                i, key, key->lct_tags,
1914                                key->lct_init, key->lct_fini, key->lct_exit,
1915                                key->lct_index, cfs_atomic_read(&key->lct_used),
1916                                key->lct_owner ? key->lct_owner->name : "",
1917                                key->lct_owner);
1918                         lu_ref_print(&key->lct_reference);
1919                 }
1920         }
1921 }
1922 EXPORT_SYMBOL(lu_context_keys_dump);
1923 #else  /* !__KERNEL__ */
1924 static int lu_cache_shrink(int nr, unsigned int gfp_mask)
1925 {
1926         return 0;
1927 }
1928 #endif /* __KERNEL__ */
1929
1930 int  cl_global_init(void);
1931 void cl_global_fini(void);
1932 int  lu_ref_global_init(void);
1933 void lu_ref_global_fini(void);
1934
1935 int dt_global_init(void);
1936 void dt_global_fini(void);
1937
1938 int llo_global_init(void);
1939 void llo_global_fini(void);
1940
1941 /* context key constructor/destructor: lu_ucred_key_init, lu_ucred_key_fini */
1942 LU_KEY_INIT_FINI(lu_ucred, struct lu_ucred);
1943
1944 static struct lu_context_key lu_ucred_key = {
1945         .lct_tags = LCT_SESSION,
1946         .lct_init = lu_ucred_key_init,
1947         .lct_fini = lu_ucred_key_fini
1948 };
1949
1950 /**
1951  * Get ucred key if session exists and ucred key is allocated on it.
1952  * Return NULL otherwise.
1953  */
1954 struct lu_ucred *lu_ucred(const struct lu_env *env)
1955 {
1956         if (!env->le_ses)
1957                 return NULL;
1958         return lu_context_key_get(env->le_ses, &lu_ucred_key);
1959 }
1960 EXPORT_SYMBOL(lu_ucred);
1961
1962 /**
1963  * Get ucred key and check if it is properly initialized.
1964  * Return NULL otherwise.
1965  */
1966 struct lu_ucred *lu_ucred_check(const struct lu_env *env)
1967 {
1968         struct lu_ucred *uc = lu_ucred(env);
1969         if (uc && uc->uc_valid != UCRED_OLD && uc->uc_valid != UCRED_NEW)
1970                 return NULL;
1971         return uc;
1972 }
1973 EXPORT_SYMBOL(lu_ucred_check);
1974
1975 /**
1976  * Get ucred key, which must exist and must be properly initialized.
1977  * Assert otherwise.
1978  */
1979 struct lu_ucred *lu_ucred_assert(const struct lu_env *env)
1980 {
1981         struct lu_ucred *uc = lu_ucred_check(env);
1982         LASSERT(uc != NULL);
1983         return uc;
1984 }
1985 EXPORT_SYMBOL(lu_ucred_assert);
1986
1987 /**
1988  * Initialization of global lu_* data.
1989  */
1990 int lu_global_init(void)
1991 {
1992         int result;
1993
1994         CDEBUG(D_INFO, "Lustre LU module (%p).\n", &lu_keys);
1995
1996         result = lu_ref_global_init();
1997         if (result != 0)
1998                 return result;
1999
2000         LU_CONTEXT_KEY_INIT(&lu_global_key);
2001         result = lu_context_key_register(&lu_global_key);
2002         if (result != 0)
2003                 return result;
2004
2005         LU_CONTEXT_KEY_INIT(&lu_ucred_key);
2006         result = lu_context_key_register(&lu_ucred_key);
2007         if (result != 0)
2008                 return result;
2009
2010         /*
2011          * At this level, we don't know what tags are needed, so allocate them
2012          * conservatively. This should not be too bad, because this
2013          * environment is global.
2014          */
2015         mutex_lock(&lu_sites_guard);
2016         result = lu_env_init(&lu_shrink_env, LCT_SHRINKER);
2017         mutex_unlock(&lu_sites_guard);
2018         if (result != 0)
2019                 return result;
2020
2021         /*
2022          * seeks estimation: 3 seeks to read a record from oi, one to read
2023          * inode, one for ea. Unfortunately setting this high value results in
2024          * lu_object/inode cache consuming all the memory.
2025          */
2026         lu_site_shrinker = cfs_set_shrinker(CFS_DEFAULT_SEEKS, lu_cache_shrink);
2027         if (lu_site_shrinker == NULL)
2028                 return -ENOMEM;
2029
2030         result = lu_time_global_init();
2031         if (result)
2032                 GOTO(out, result);
2033
2034 #ifdef __KERNEL__
2035         result = dt_global_init();
2036         if (result)
2037                 GOTO(out, result);
2038
2039         result = llo_global_init();
2040         if (result)
2041                 GOTO(out, result);
2042 #endif
2043         result = cl_global_init();
2044 out:
2045
2046         return result;
2047 }
2048
2049 /**
2050  * Dual to lu_global_init().
2051  */
2052 void lu_global_fini(void)
2053 {
2054         cl_global_fini();
2055 #ifdef __KERNEL__
2056         llo_global_fini();
2057         dt_global_fini();
2058 #endif
2059         lu_time_global_fini();
2060         if (lu_site_shrinker != NULL) {
2061                 cfs_remove_shrinker(lu_site_shrinker);
2062                 lu_site_shrinker = NULL;
2063         }
2064
2065         lu_context_key_degister(&lu_global_key);
2066         lu_context_key_degister(&lu_ucred_key);
2067
2068         /*
2069          * Tear shrinker environment down _after_ de-registering
2070          * lu_global_key, because the latter has a value in the former.
2071          */
2072         mutex_lock(&lu_sites_guard);
2073         lu_env_fini(&lu_shrink_env);
2074         mutex_unlock(&lu_sites_guard);
2075
2076         lu_ref_global_fini();
2077 }
2078
2079 struct lu_buf LU_BUF_NULL = {
2080         .lb_buf = NULL,
2081         .lb_len = 0
2082 };
2083 EXPORT_SYMBOL(LU_BUF_NULL);
2084
2085 static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx)
2086 {
2087 #ifdef LPROCFS
2088         struct lprocfs_counter ret;
2089
2090         lprocfs_stats_collect(stats, idx, &ret);
2091         return (__u32)ret.lc_count;
2092 #else
2093         return 0;
2094 #endif
2095 }
2096
2097 /**
2098  * Output site statistical counters into a buffer. Suitable for
2099  * lprocfs_rd_*()-style functions.
2100  */
2101 int lu_site_stats_print(const struct lu_site *s, char *page, int count)
2102 {
2103         lu_site_stats_t stats;
2104
2105         memset(&stats, 0, sizeof(stats));
2106         lu_site_stats_get(s->ls_obj_hash, &stats, 1);
2107
2108         return snprintf(page, count, "%d/%d %d/%d %d %d %d %d %d %d %d\n",
2109                         stats.lss_busy,
2110                         stats.lss_total,
2111                         stats.lss_populated,
2112                         CFS_HASH_NHLIST(s->ls_obj_hash),
2113                         stats.lss_max_search,
2114                         ls_stats_read(s->ls_stats, LU_SS_CREATED),
2115                         ls_stats_read(s->ls_stats, LU_SS_CACHE_HIT),
2116                         ls_stats_read(s->ls_stats, LU_SS_CACHE_MISS),
2117                         ls_stats_read(s->ls_stats, LU_SS_CACHE_RACE),
2118                         ls_stats_read(s->ls_stats, LU_SS_CACHE_DEATH_RACE),
2119                         ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED));
2120 }
2121 EXPORT_SYMBOL(lu_site_stats_print);
2122
2123 const char *lu_time_names[LU_TIME_NR] = {
2124         [LU_TIME_FIND_LOOKUP] = "find_lookup",
2125         [LU_TIME_FIND_ALLOC]  = "find_alloc",
2126         [LU_TIME_FIND_INSERT] = "find_insert"
2127 };
2128 EXPORT_SYMBOL(lu_time_names);
2129
2130 /**
2131  * Helper function to initialize a number of kmem slab caches at once.
2132  */
2133 int lu_kmem_init(struct lu_kmem_descr *caches)
2134 {
2135         int result;
2136         struct lu_kmem_descr *iter = caches;
2137
2138         for (result = 0; iter->ckd_cache != NULL; ++iter) {
2139                 *iter->ckd_cache = cfs_mem_cache_create(iter->ckd_name,
2140                                                         iter->ckd_size,
2141                                                         0, 0);
2142                 if (*iter->ckd_cache == NULL) {
2143                         result = -ENOMEM;
2144                         /* free all previously allocated caches */
2145                         lu_kmem_fini(caches);
2146                         break;
2147                 }
2148         }
2149         return result;
2150 }
2151 EXPORT_SYMBOL(lu_kmem_init);
2152
2153 /**
2154  * Helper function to finalize a number of kmem slab cached at once. Dual to
2155  * lu_kmem_init().
2156  */
2157 void lu_kmem_fini(struct lu_kmem_descr *caches)
2158 {
2159         int rc;
2160
2161         for (; caches->ckd_cache != NULL; ++caches) {
2162                 if (*caches->ckd_cache != NULL) {
2163                         rc = cfs_mem_cache_destroy(*caches->ckd_cache);
2164                         LASSERTF(rc == 0, "couldn't destroy %s slab\n",
2165                                  caches->ckd_name);
2166                         *caches->ckd_cache = NULL;
2167                 }
2168         }
2169 }
2170 EXPORT_SYMBOL(lu_kmem_fini);
2171
2172 /**
2173  * Temporary solution to be able to assign fid in ->do_create()
2174  * till we have fully-functional OST fids
2175  */
2176 void lu_object_assign_fid(const struct lu_env *env, struct lu_object *o,
2177                           const struct lu_fid *fid)
2178 {
2179         struct lu_site          *s = o->lo_dev->ld_site;
2180         struct lu_fid           *old = &o->lo_header->loh_fid;
2181         struct lu_site_bkt_data *bkt;
2182         struct lu_object        *shadow;
2183         cfs_waitlink_t           waiter;
2184         cfs_hash_t              *hs;
2185         cfs_hash_bd_t            bd;
2186         __u64                    version = 0;
2187
2188         LASSERT(fid_is_zero(old));
2189
2190         hs = s->ls_obj_hash;
2191         cfs_hash_bd_get_and_lock(hs, (void *)fid, &bd, 1);
2192         shadow = htable_lookup(s, &bd, fid, &waiter, &version);
2193         /* supposed to be unique */
2194         LASSERT(shadow == NULL);
2195         *old = *fid;
2196         bkt = cfs_hash_bd_extra_get(hs, &bd);
2197         cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
2198         bkt->lsb_busy++;
2199         cfs_hash_bd_unlock(hs, &bd, 1);
2200 }
2201 EXPORT_SYMBOL(lu_object_assign_fid);
2202
2203 /**
2204  * allocates object with 0 (non-assiged) fid
2205  * XXX: temporary solution to be able to assign fid in ->do_create()
2206  *      till we have fully-functional OST fids
2207  */
2208 struct lu_object *lu_object_anon(const struct lu_env *env,
2209                                  struct lu_device *dev,
2210                                  const struct lu_object_conf *conf)
2211 {
2212         struct lu_fid     fid;
2213         struct lu_object *o;
2214
2215         fid_zero(&fid);
2216         o = lu_object_alloc(env, dev, &fid, conf);
2217
2218         return o;
2219 }
2220 EXPORT_SYMBOL(lu_object_anon);