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, 2013, 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 #include <linux/module.h>
  49 #include <libcfs/libcfs_hash.h> /* hash_long() */
  50 #include <obd_class.h>
  51 #include <obd_support.h>
  52 #include <lustre_disk.h>
  53 #include <lustre_fid.h>
  54 #include <lu_object.h>
  55 #include <lu_ref.h>
  56 #include <libcfs/list.h>
  57
  58 enum {
  59         LU_CACHE_PERCENT_MAX     = 50,
  60         LU_CACHE_PERCENT_DEFAULT = 20
  61 };
  62
  63 #define LU_CACHE_NR_MAX_ADJUST          128
  64 #define LU_CACHE_NR_UNLIMITED           -1
  65 #define LU_CACHE_NR_DEFAULT             LU_CACHE_NR_UNLIMITED
  66 #define LU_CACHE_NR_LDISKFS_LIMIT       LU_CACHE_NR_UNLIMITED
  67 /** This is set to roughly (20 * OSS_NTHRS_MAX) to prevent thrashing */
  68 #define LU_CACHE_NR_ZFS_LIMIT           10240
  69
  70 #define LU_SITE_BITS_MIN    12
  71 #define LU_SITE_BITS_MAX    24
  72 /**
  73  * total 256 buckets, we don't want too many buckets because:
  74  * - consume too much memory
  75  * - avoid unbalanced LRU list
  76  */
  77 #define LU_SITE_BKT_BITS    8
  78
  79
  80 static unsigned int lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
  81 CFS_MODULE_PARM(lu_cache_percent, "i", int, 0644,
  82                 "Percentage of memory to be used as lu_object cache");
  83
  84 static long lu_cache_nr = LU_CACHE_NR_DEFAULT;
  85 CFS_MODULE_PARM(lu_cache_nr, "l", long, 0644,
  86                 "Maximum number of objects in lu_object cache");
  87
  88 static void lu_object_free(const struct lu_env *env, struct lu_object *o);
  89
  90 /**
  91  * Decrease reference counter on object. If last reference is freed, return
  92  * object to the cache, unless lu_object_is_dying(o) holds. In the latter
  93  * case, free object immediately.
  94  */
  95 void lu_object_put(const struct lu_env *env, struct lu_object *o)
  96 {
  97         struct lu_site_bkt_data *bkt;
  98         struct lu_object_header *top;
  99         struct lu_site          *site;
 100         struct lu_object        *orig;
 101         cfs_hash_bd_t            bd;
 102         const struct lu_fid     *fid;
 103
 104         top  = o->lo_header;
 105         site = o->lo_dev->ld_site;
 106         orig = o;
 107
 108         /*
 109          * till we have full fids-on-OST implemented anonymous objects
 110          * are possible in OSP. such an object isn't listed in the site
 111          * so we should not remove it from the site.
 112          */
 113         fid = lu_object_fid(o);
 114         if (fid_is_zero(fid)) {
 115                 LASSERT(top->loh_hash.next == NULL
 116                         && top->loh_hash.pprev == NULL);
 117                 LASSERT(list_empty(&top->loh_lru));
 118                 if (!atomic_dec_and_test(&top->loh_ref))
 119                         return;
 120                 list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
 121                         if (o->lo_ops->loo_object_release != NULL)
 122                                 o->lo_ops->loo_object_release(env, o);
 123                 }
 124                 lu_object_free(env, orig);
 125                 return;
 126         }
 127
 128         cfs_hash_bd_get(site->ls_obj_hash, &top->loh_fid, &bd);
 129         bkt = cfs_hash_bd_extra_get(site->ls_obj_hash, &bd);
 130
 131         if (!cfs_hash_bd_dec_and_lock(site->ls_obj_hash, &bd, &top->loh_ref)) {
 132                 if (lu_object_is_dying(top)) {
 133
 134                         /*
 135                          * somebody may be waiting for this, currently only
 136                          * used for cl_object, see cl_object_put_last().
 137                          */
 138                         wake_up_all(&bkt->lsb_marche_funebre);
 139                 }
 140                 return;
 141         }
 142
 143         LASSERT(bkt->lsb_busy > 0);
 144         bkt->lsb_busy--;
 145         /*
 146          * When last reference is released, iterate over object
 147          * layers, and notify them that object is no longer busy.
 148          */
 149         list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
 150                 if (o->lo_ops->loo_object_release != NULL)
 151                         o->lo_ops->loo_object_release(env, o);
 152         }
 153
 154         if (!lu_object_is_dying(top)) {
 155                 LASSERT(list_empty(&top->loh_lru));
 156                 list_add_tail(&top->loh_lru, &bkt->lsb_lru);
 157                 cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
 158                 return;
 159         }
 160
 161         /*
 162          * If object is dying (will not be cached), removed it
 163          * from hash table and LRU.
 164          *
 165          * This is done with hash table and LRU lists locked. As the only
 166          * way to acquire first reference to previously unreferenced
 167          * object is through hash-table lookup (lu_object_find()),
 168          * or LRU scanning (lu_site_purge()), that are done under hash-table
 169          * and LRU lock, no race with concurrent object lookup is possible
 170          * and we can safely destroy object below.
 171          */
 172         if (!test_and_set_bit(LU_OBJECT_UNHASHED, &top->loh_flags))
 173                 cfs_hash_bd_del_locked(site->ls_obj_hash, &bd, &top->loh_hash);
 174         cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
 175         /*
 176          * Object was already removed from hash and lru above, can
 177          * kill it.
 178          */
 179         lu_object_free(env, orig);
 180 }
 181 EXPORT_SYMBOL(lu_object_put);
 182
 183 /**
 184  * Put object and don't keep in cache. This is temporary solution for
 185  * multi-site objects when its layering is not constant.
 186  */
 187 void lu_object_put_nocache(const struct lu_env *env, struct lu_object *o)
 188 {
 189         set_bit(LU_OBJECT_HEARD_BANSHEE, &o->lo_header->loh_flags);
 190         return lu_object_put(env, o);
 191 }
 192 EXPORT_SYMBOL(lu_object_put_nocache);
 193
 194 /**
 195  * Kill the object and take it out of LRU cache.
 196  * Currently used by client code for layout change.
 197  */
 198 void lu_object_unhash(const struct lu_env *env, struct lu_object *o)
 199 {
 200         struct lu_object_header *top;
 201
 202         top = o->lo_header;
 203         set_bit(LU_OBJECT_HEARD_BANSHEE, &top->loh_flags);
 204         if (!test_and_set_bit(LU_OBJECT_UNHASHED, &top->loh_flags)) {
 205                 cfs_hash_t *obj_hash = o->lo_dev->ld_site->ls_obj_hash;
 206                 cfs_hash_bd_t bd;
 207
 208                 cfs_hash_bd_get_and_lock(obj_hash, &top->loh_fid, &bd, 1);
 209                 list_del_init(&top->loh_lru);
 210                 cfs_hash_bd_del_locked(obj_hash, &bd, &top->loh_hash);
 211                 cfs_hash_bd_unlock(obj_hash, &bd, 1);
 212         }
 213 }
 214 EXPORT_SYMBOL(lu_object_unhash);
 215
 216 /**
 217  * Allocate new object.
 218  *
 219  * This follows object creation protocol, described in the comment within
 220  * struct lu_device_operations definition.
 221  */
 222 static struct lu_object *lu_object_alloc(const struct lu_env *env,
 223                                          struct lu_device *dev,
 224                                          const struct lu_fid *f,
 225                                          const struct lu_object_conf *conf)
 226 {
 227         struct lu_object *scan;
 228         struct lu_object *top;
 229         struct list_head *layers;
 230         unsigned int init_mask = 0;
 231         unsigned int init_flag;
 232         int clean;
 233         int result;
 234         ENTRY;
 235
 236         /*
 237          * Create top-level object slice. This will also create
 238          * lu_object_header.
 239          */
 240         top = dev->ld_ops->ldo_object_alloc(env, NULL, dev);
 241         if (top == NULL)
 242                 RETURN(ERR_PTR(-ENOMEM));
 243         if (IS_ERR(top))
 244                 RETURN(top);
 245         /*
 246          * This is the only place where object fid is assigned. It's constant
 247          * after this point.
 248          */
 249         top->lo_header->loh_fid = *f;
 250         layers = &top->lo_header->loh_layers;
 251
 252         do {
 253                 /*
 254                  * Call ->loo_object_init() repeatedly, until no more new
 255                  * object slices are created.
 256                  */
 257                 clean = 1;
 258                 init_flag = 1;
 259                 list_for_each_entry(scan, layers, lo_linkage) {
 260                         if (init_mask & init_flag)
 261                                 goto next;
 262                         clean = 0;
 263                         scan->lo_header = top->lo_header;
 264                         result = scan->lo_ops->loo_object_init(env, scan, conf);
 265                         if (result != 0) {
 266                                 lu_object_free(env, top);
 267                                 RETURN(ERR_PTR(result));
 268                         }
 269                         init_mask |= init_flag;
 270 next:
 271                         init_flag <<= 1;
 272                 }
 273         } while (!clean);
 274
 275         list_for_each_entry_reverse(scan, layers, lo_linkage) {
 276                 if (scan->lo_ops->loo_object_start != NULL) {
 277                         result = scan->lo_ops->loo_object_start(env, scan);
 278                         if (result != 0) {
 279                                 lu_object_free(env, top);
 280                                 RETURN(ERR_PTR(result));
 281                         }
 282                 }
 283         }
 284
 285         lprocfs_counter_incr(dev->ld_site->ls_stats, LU_SS_CREATED);
 286         RETURN(top);
 287 }
 288
 289 /**
 290  * Free an object.
 291  */
 292 static void lu_object_free(const struct lu_env *env, struct lu_object *o)
 293 {
 294         struct lu_site_bkt_data *bkt;
 295         struct lu_site          *site;
 296         struct lu_object        *scan;
 297         struct list_head        *layers;
 298         struct list_head         splice;
 299
 300         site   = o->lo_dev->ld_site;
 301         layers = &o->lo_header->loh_layers;
 302         bkt    = lu_site_bkt_from_fid(site, &o->lo_header->loh_fid);
 303         /*
 304          * First call ->loo_object_delete() method to release all resources.
 305          */
 306         list_for_each_entry_reverse(scan, layers, lo_linkage) {
 307                 if (scan->lo_ops->loo_object_delete != NULL)
 308                         scan->lo_ops->loo_object_delete(env, scan);
 309         }
 310
 311         /*
 312          * Then, splice object layers into stand-alone list, and call
 313          * ->loo_object_free() on all layers to free memory. Splice is
 314          * necessary, because lu_object_header is freed together with the
 315          * top-level slice.
 316          */
 317         INIT_LIST_HEAD(&splice);
 318         list_splice_init(layers, &splice);
 319         while (!list_empty(&splice)) {
 320                 /*
 321                  * Free layers in bottom-to-top order, so that object header
 322                  * lives as long as possible and ->loo_object_free() methods
 323                  * can look at its contents.
 324                  */
 325                 o = container_of0(splice.prev, struct lu_object, lo_linkage);
 326                 list_del_init(&o->lo_linkage);
 327                 LASSERT(o->lo_ops->loo_object_free != NULL);
 328                 o->lo_ops->loo_object_free(env, o);
 329         }
 330
 331         if (waitqueue_active(&bkt->lsb_marche_funebre))
 332                 wake_up_all(&bkt->lsb_marche_funebre);
 333 }
 334
 335 /**
 336  * Free \a nr objects from the cold end of the site LRU list.
 337  */
 338 int lu_site_purge(const struct lu_env *env, struct lu_site *s, int nr)
 339 {
 340         struct lu_object_header *h;
 341         struct lu_object_header *temp;
 342         struct lu_site_bkt_data *bkt;
 343         cfs_hash_bd_t            bd;
 344         cfs_hash_bd_t            bd2;
 345         struct list_head         dispose;
 346         int                      did_sth;
 347         int                      start;
 348         int                      count;
 349         int                      bnr;
 350         unsigned int             i;
 351
 352         if (OBD_FAIL_CHECK(OBD_FAIL_OBD_NO_LRU))
 353                 RETURN(0);
 354
 355         INIT_LIST_HEAD(&dispose);
 356         /*
 357          * Under LRU list lock, scan LRU list and move unreferenced objects to
 358          * the dispose list, removing them from LRU and hash table.
 359          */
 360         start = s->ls_purge_start;
 361         bnr = (nr == ~0) ? -1 : nr / CFS_HASH_NBKT(s->ls_obj_hash) + 1;
 362  again:
 363         /*
 364          * It doesn't make any sense to make purge threads parallel, that can
 365          * only bring troubles to us. See LU-5331.
 366          */
 367         mutex_lock(&s->ls_purge_mutex);
 368         did_sth = 0;
 369         cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
 370                 if (i < start)
 371                         continue;
 372                 count = bnr;
 373                 cfs_hash_bd_lock(s->ls_obj_hash, &bd, 1);
 374                 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
 375
 376                 list_for_each_entry_safe(h, temp, &bkt->lsb_lru, loh_lru) {
 377                         LASSERT(atomic_read(&h->loh_ref) == 0);
 378
 379                         cfs_hash_bd_get(s->ls_obj_hash, &h->loh_fid, &bd2);
 380                         LASSERT(bd.bd_bucket == bd2.bd_bucket);
 381
 382                         cfs_hash_bd_del_locked(s->ls_obj_hash,
 383                                                &bd2, &h->loh_hash);
 384                         list_move(&h->loh_lru, &dispose);
 385                         if (did_sth == 0)
 386                                 did_sth = 1;
 387
 388                         if (nr != ~0 && --nr == 0)
 389                                 break;
 390
 391                         if (count > 0 && --count == 0)
 392                                 break;
 393
 394                 }
 395                 cfs_hash_bd_unlock(s->ls_obj_hash, &bd, 1);
 396                 cond_resched();
 397                 /*
 398                  * Free everything on the dispose list. This is safe against
 399                  * races due to the reasons described in lu_object_put().
 400                  */
 401                 while (!list_empty(&dispose)) {
 402                         h = container_of0(dispose.next,
 403                                           struct lu_object_header, loh_lru);
 404                         list_del_init(&h->loh_lru);
 405                         lu_object_free(env, lu_object_top(h));
 406                         lprocfs_counter_incr(s->ls_stats, LU_SS_LRU_PURGED);
 407                 }
 408
 409                 if (nr == 0)
 410                         break;
 411         }
 412         mutex_unlock(&s->ls_purge_mutex);
 413
 414         if (nr != 0 && did_sth && start != 0) {
 415                 start = 0; /* restart from the first bucket */
 416                 goto again;
 417         }
 418         /* race on s->ls_purge_start, but nobody cares */
 419         s->ls_purge_start = i % CFS_HASH_NBKT(s->ls_obj_hash);
 420
 421         return nr;
 422 }
 423 EXPORT_SYMBOL(lu_site_purge);
 424
 425 /*
 426  * Object printing.
 427  *
 428  * Code below has to jump through certain loops to output object description
 429  * into libcfs_debug_msg-based log. The problem is that lu_object_print()
 430  * composes object description from strings that are parts of _lines_ of
 431  * output (i.e., strings that are not terminated by newline). This doesn't fit
 432  * very well into libcfs_debug_msg() interface that assumes that each message
 433  * supplied to it is a self-contained output line.
 434  *
 435  * To work around this, strings are collected in a temporary buffer
 436  * (implemented as a value of lu_cdebug_key key), until terminating newline
 437  * character is detected.
 438  *
 439  */
 440
 441 enum {
 442         /**
 443          * Maximal line size.
 444          *
 445          * XXX overflow is not handled correctly.
 446          */
 447         LU_CDEBUG_LINE = 512
 448 };
 449
 450 struct lu_cdebug_data {
 451         /**
 452          * Temporary buffer.
 453          */
 454         char lck_area[LU_CDEBUG_LINE];
 455 };
 456
 457 /* context key constructor/destructor: lu_global_key_init, lu_global_key_fini */
 458 LU_KEY_INIT_FINI(lu_global, struct lu_cdebug_data);
 459
 460 /**
 461  * Key, holding temporary buffer. This key is registered very early by
 462  * lu_global_init().
 463  */
 464 struct lu_context_key lu_global_key = {
 465         .lct_tags = LCT_MD_THREAD | LCT_DT_THREAD |
 466                     LCT_MG_THREAD | LCT_CL_THREAD | LCT_LOCAL,
 467         .lct_init = lu_global_key_init,
 468         .lct_fini = lu_global_key_fini
 469 };
 470
 471 /**
 472  * Printer function emitting messages through libcfs_debug_msg().
 473  */
 474 int lu_cdebug_printer(const struct lu_env *env,
 475                       void *cookie, const char *format, ...)
 476 {
 477         struct libcfs_debug_msg_data *msgdata = cookie;
 478         struct lu_cdebug_data        *key;
 479         int used;
 480         int complete;
 481         va_list args;
 482
 483         va_start(args, format);
 484
 485         key = lu_context_key_get(&env->le_ctx, &lu_global_key);
 486         LASSERT(key != NULL);
 487
 488         used = strlen(key->lck_area);
 489         complete = format[strlen(format) - 1] == '\n';
 490         /*
 491          * Append new chunk to the buffer.
 492          */
 493         vsnprintf(key->lck_area + used,
 494                   ARRAY_SIZE(key->lck_area) - used, format, args);
 495         if (complete) {
 496                 if (cfs_cdebug_show(msgdata->msg_mask, msgdata->msg_subsys))
 497                         libcfs_debug_msg(msgdata, "%s\n", key->lck_area);
 498                 key->lck_area[0] = 0;
 499         }
 500         va_end(args);
 501         return 0;
 502 }
 503 EXPORT_SYMBOL(lu_cdebug_printer);
 504
 505 /**
 506  * Print object header.
 507  */
 508 void lu_object_header_print(const struct lu_env *env, void *cookie,
 509                             lu_printer_t printer,
 510                             const struct lu_object_header *hdr)
 511 {
 512         (*printer)(env, cookie, "header@%p[%#lx, %d, "DFID"%s%s%s]",
 513                    hdr, hdr->loh_flags, atomic_read(&hdr->loh_ref),
 514                    PFID(&hdr->loh_fid),
 515                    hlist_unhashed(&hdr->loh_hash) ? "" : " hash",
 516                    list_empty((struct list_head *)&hdr->loh_lru) ? \
 517                    "" : " lru",
 518                    hdr->loh_attr & LOHA_EXISTS ? " exist" : "");
 519 }
 520 EXPORT_SYMBOL(lu_object_header_print);
 521
 522 /**
 523  * Print human readable representation of the \a o to the \a printer.
 524  */
 525 void lu_object_print(const struct lu_env *env, void *cookie,
 526                      lu_printer_t printer, const struct lu_object *o)
 527 {
 528         static const char ruler[] = "........................................";
 529         struct lu_object_header *top;
 530         int depth = 4;
 531
 532         top = o->lo_header;
 533         lu_object_header_print(env, cookie, printer, top);
 534         (*printer)(env, cookie, "{\n");
 535
 536         list_for_each_entry(o, &top->loh_layers, lo_linkage) {
 537                 /*
 538                  * print `.' \a depth times followed by type name and address
 539                  */
 540                 (*printer)(env, cookie, "%*.*s%s@%p", depth, depth, ruler,
 541                            o->lo_dev->ld_type->ldt_name, o);
 542
 543                 if (o->lo_ops->loo_object_print != NULL)
 544                         (*o->lo_ops->loo_object_print)(env, cookie, printer, o);
 545
 546                 (*printer)(env, cookie, "\n");
 547         }
 548
 549         (*printer)(env, cookie, "} header@%p\n", top);
 550 }
 551 EXPORT_SYMBOL(lu_object_print);
 552
 553 /**
 554  * Check object consistency.
 555  */
 556 int lu_object_invariant(const struct lu_object *o)
 557 {
 558         struct lu_object_header *top;
 559
 560         top = o->lo_header;
 561         list_for_each_entry(o, &top->loh_layers, lo_linkage) {
 562                 if (o->lo_ops->loo_object_invariant != NULL &&
 563                     !o->lo_ops->loo_object_invariant(o))
 564                         return 0;
 565         }
 566         return 1;
 567 }
 568 EXPORT_SYMBOL(lu_object_invariant);
 569
 570 static struct lu_object *htable_lookup(struct lu_site *s,
 571                                        cfs_hash_bd_t *bd,
 572                                        const struct lu_fid *f,
 573                                        wait_queue_t *waiter,
 574                                        __u64 *version)
 575 {
 576         struct lu_site_bkt_data *bkt;
 577         struct lu_object_header *h;
 578         struct hlist_node       *hnode;
 579         __u64  ver = cfs_hash_bd_version_get(bd);
 580
 581         if (*version == ver)
 582                 return ERR_PTR(-ENOENT);
 583
 584         *version = ver;
 585         bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, bd);
 586         /* cfs_hash_bd_peek_locked is a somehow "internal" function
 587          * of cfs_hash, it doesn't add refcount on object. */
 588         hnode = cfs_hash_bd_peek_locked(s->ls_obj_hash, bd, (void *)f);
 589         if (hnode == NULL) {
 590                 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_MISS);
 591                 return ERR_PTR(-ENOENT);
 592         }
 593
 594         h = container_of0(hnode, struct lu_object_header, loh_hash);
 595         if (likely(!lu_object_is_dying(h))) {
 596                 cfs_hash_get(s->ls_obj_hash, hnode);
 597                 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_HIT);
 598                 list_del_init(&h->loh_lru);
 599                 return lu_object_top(h);
 600         }
 601
 602         /*
 603          * Lookup found an object being destroyed this object cannot be
 604          * returned (to assure that references to dying objects are eventually
 605          * drained), and moreover, lookup has to wait until object is freed.
 606          */
 607
 608         init_waitqueue_entry_current(waiter);
 609         add_wait_queue(&bkt->lsb_marche_funebre, waiter);
 610         set_current_state(TASK_UNINTERRUPTIBLE);
 611         lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_DEATH_RACE);
 612         return ERR_PTR(-EAGAIN);
 613 }
 614
 615 static struct lu_object *htable_lookup_nowait(struct lu_site *s,
 616                                               cfs_hash_bd_t *bd,
 617                                               const struct lu_fid *f)
 618 {
 619         struct hlist_node       *hnode;
 620         struct lu_object_header *h;
 621
 622         /* cfs_hash_bd_peek_locked is a somehow "internal" function
 623          * of cfs_hash, it doesn't add refcount on object. */
 624         hnode = cfs_hash_bd_peek_locked(s->ls_obj_hash, bd, (void *)f);
 625         if (hnode == NULL) {
 626                 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_MISS);
 627                 return ERR_PTR(-ENOENT);
 628         }
 629
 630         h = container_of0(hnode, struct lu_object_header, loh_hash);
 631         if (unlikely(lu_object_is_dying(h)))
 632                 return ERR_PTR(-ENOENT);
 633
 634         cfs_hash_get(s->ls_obj_hash, hnode);
 635         lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_HIT);
 636         list_del_init(&h->loh_lru);
 637         return lu_object_top(h);
 638 }
 639
 640 /**
 641  * Search cache for an object with the fid \a f. If such object is found,
 642  * return it. Otherwise, create new object, insert it into cache and return
 643  * it. In any case, additional reference is acquired on the returned object.
 644  */
 645 struct lu_object *lu_object_find(const struct lu_env *env,
 646                                  struct lu_device *dev, const struct lu_fid *f,
 647                                  const struct lu_object_conf *conf)
 648 {
 649         return lu_object_find_at(env, dev->ld_site->ls_top_dev, f, conf);
 650 }
 651 EXPORT_SYMBOL(lu_object_find);
 652
 653 /*
 654  * Limit the lu_object cache to a maximum of lu_cache_nr objects.  Because
 655  * the calculation for the number of objects to reclaim is not covered by
 656  * a lock the maximum number of objects is capped by LU_CACHE_MAX_ADJUST.
 657  * This ensures that many concurrent threads will not accidentally purge
 658  * the entire cache.
 659  */
 660 static void lu_object_limit(const struct lu_env *env,
 661                             struct lu_device *dev)
 662 {
 663         __u64 size, nr;
 664
 665         if (lu_cache_nr == LU_CACHE_NR_UNLIMITED)
 666                 return;
 667
 668         size = cfs_hash_size_get(dev->ld_site->ls_obj_hash);
 669         nr = (__u64)lu_cache_nr;
 670         if (size > nr)
 671                 lu_site_purge(env, dev->ld_site,
 672                               MIN(size - nr, LU_CACHE_NR_MAX_ADJUST));
 673
 674         return;
 675 }
 676
 677 static struct lu_object *lu_object_new(const struct lu_env *env,
 678                                        struct lu_device *dev,
 679                                        const struct lu_fid *f,
 680                                        const struct lu_object_conf *conf)
 681 {
 682         struct lu_object        *o;
 683         cfs_hash_t              *hs;
 684         cfs_hash_bd_t            bd;
 685         struct lu_site_bkt_data *bkt;
 686
 687         o = lu_object_alloc(env, dev, f, conf);
 688         if (unlikely(IS_ERR(o)))
 689                 return o;
 690
 691         hs = dev->ld_site->ls_obj_hash;
 692         cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
 693         bkt = cfs_hash_bd_extra_get(hs, &bd);
 694         cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
 695         bkt->lsb_busy++;
 696         cfs_hash_bd_unlock(hs, &bd, 1);
 697
 698         lu_object_limit(env, dev);
 699
 700         return o;
 701 }
 702
 703 /**
 704  * Core logic of lu_object_find*() functions.
 705  */
 706 static struct lu_object *lu_object_find_try(const struct lu_env *env,
 707                                             struct lu_device *dev,
 708                                             const struct lu_fid *f,
 709                                             const struct lu_object_conf *conf,
 710                                             wait_queue_t *waiter)
 711 {
 712         struct lu_object      *o;
 713         struct lu_object      *shadow;
 714         struct lu_site        *s;
 715         cfs_hash_t            *hs;
 716         cfs_hash_bd_t          bd;
 717         __u64                  version = 0;
 718
 719         /*
 720          * This uses standard index maintenance protocol:
 721          *
 722          *     - search index under lock, and return object if found;
 723          *     - otherwise, unlock index, allocate new object;
 724          *     - lock index and search again;
 725          *     - if nothing is found (usual case), insert newly created
 726          *       object into index;
 727          *     - otherwise (race: other thread inserted object), free
 728          *       object just allocated.
 729          *     - unlock index;
 730          *     - return object.
 731          *
 732          * For "LOC_F_NEW" case, we are sure the object is new established.
 733          * It is unnecessary to perform lookup-alloc-lookup-insert, instead,
 734          * just alloc and insert directly.
 735          *
 736          * If dying object is found during index search, add @waiter to the
 737          * site wait-queue and return ERR_PTR(-EAGAIN).
 738          */
 739         if (conf != NULL && conf->loc_flags & LOC_F_NEW)
 740                 return lu_object_new(env, dev, f, conf);
 741
 742         s  = dev->ld_site;
 743         hs = s->ls_obj_hash;
 744         cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
 745         o = htable_lookup(s, &bd, f, waiter, &version);
 746         cfs_hash_bd_unlock(hs, &bd, 1);
 747         if (!IS_ERR(o) || PTR_ERR(o) != -ENOENT)
 748                 return o;
 749
 750         /*
 751          * Allocate new object. This may result in rather complicated
 752          * operations, including fld queries, inode loading, etc.
 753          */
 754         o = lu_object_alloc(env, dev, f, conf);
 755         if (unlikely(IS_ERR(o)))
 756                 return o;
 757
 758         LASSERT(lu_fid_eq(lu_object_fid(o), f));
 759
 760         cfs_hash_bd_lock(hs, &bd, 1);
 761
 762         shadow = htable_lookup(s, &bd, f, waiter, &version);
 763         if (likely(IS_ERR(shadow) && PTR_ERR(shadow) == -ENOENT)) {
 764                 struct lu_site_bkt_data *bkt;
 765
 766                 bkt = cfs_hash_bd_extra_get(hs, &bd);
 767                 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
 768                 bkt->lsb_busy++;
 769                 cfs_hash_bd_unlock(hs, &bd, 1);
 770
 771                 lu_object_limit(env, dev);
 772
 773                 return o;
 774         }
 775
 776         lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_RACE);
 777         cfs_hash_bd_unlock(hs, &bd, 1);
 778         lu_object_free(env, o);
 779         return shadow;
 780 }
 781
 782 /**
 783  * Much like lu_object_find(), but top level device of object is specifically
 784  * \a dev rather than top level device of the site. This interface allows
 785  * objects of different "stacking" to be created within the same site.
 786  */
 787 struct lu_object *lu_object_find_at(const struct lu_env *env,
 788                                     struct lu_device *dev,
 789                                     const struct lu_fid *f,
 790                                     const struct lu_object_conf *conf)
 791 {
 792         struct lu_site_bkt_data *bkt;
 793         struct lu_object        *obj;
 794         wait_queue_t           wait;
 795
 796         while (1) {
 797                 obj = lu_object_find_try(env, dev, f, conf, &wait);
 798                 if (obj != ERR_PTR(-EAGAIN))
 799                         return obj;
 800                 /*
 801                  * lu_object_find_try() already added waiter into the
 802                  * wait queue.
 803                  */
 804                 waitq_wait(&wait, TASK_UNINTERRUPTIBLE);
 805                 bkt = lu_site_bkt_from_fid(dev->ld_site, (void *)f);
 806                 remove_wait_queue(&bkt->lsb_marche_funebre, &wait);
 807         }
 808 }
 809 EXPORT_SYMBOL(lu_object_find_at);
 810
 811 /**
 812  * Try to find the object in cache without waiting for the dead object
 813  * to be released nor allocating object if no cached one was found.
 814  *
 815  * The found object will be set as LU_OBJECT_HEARD_BANSHEE for purging.
 816  */
 817 void lu_object_purge(const struct lu_env *env, struct lu_device *dev,
 818                      const struct lu_fid *f)
 819 {
 820         struct lu_site          *s  = dev->ld_site;
 821         cfs_hash_t              *hs = s->ls_obj_hash;
 822         cfs_hash_bd_t            bd;
 823         struct lu_object        *o;
 824
 825         cfs_hash_bd_get_and_lock(hs, f, &bd, 1);
 826         o = htable_lookup_nowait(s, &bd, f);
 827         cfs_hash_bd_unlock(hs, &bd, 1);
 828         if (!IS_ERR(o)) {
 829                 set_bit(LU_OBJECT_HEARD_BANSHEE, &o->lo_header->loh_flags);
 830                 lu_object_put(env, o);
 831         }
 832 }
 833 EXPORT_SYMBOL(lu_object_purge);
 834
 835 /**
 836  * Find object with given fid, and return its slice belonging to given device.
 837  */
 838 struct lu_object *lu_object_find_slice(const struct lu_env *env,
 839                                        struct lu_device *dev,
 840                                        const struct lu_fid *f,
 841                                        const struct lu_object_conf *conf)
 842 {
 843         struct lu_object *top;
 844         struct lu_object *obj;
 845
 846         top = lu_object_find(env, dev, f, conf);
 847         if (!IS_ERR(top)) {
 848                 obj = lu_object_locate(top->lo_header, dev->ld_type);
 849                 if (obj == NULL)
 850                         lu_object_put(env, top);
 851         } else
 852                 obj = top;
 853         return obj;
 854 }
 855 EXPORT_SYMBOL(lu_object_find_slice);
 856
 857 /**
 858  * Global list of all device types.
 859  */
 860 static struct list_head lu_device_types;
 861
 862 int lu_device_type_init(struct lu_device_type *ldt)
 863 {
 864         int result = 0;
 865
 866         atomic_set(&ldt->ldt_device_nr, 0);
 867         INIT_LIST_HEAD(&ldt->ldt_linkage);
 868         if (ldt->ldt_ops->ldto_init)
 869                 result = ldt->ldt_ops->ldto_init(ldt);
 870
 871         if (result == 0) {
 872                 spin_lock(&obd_types_lock);
 873                 list_add(&ldt->ldt_linkage, &lu_device_types);
 874                 spin_unlock(&obd_types_lock);
 875         }
 876
 877         return result;
 878 }
 879 EXPORT_SYMBOL(lu_device_type_init);
 880
 881 void lu_device_type_fini(struct lu_device_type *ldt)
 882 {
 883         spin_lock(&obd_types_lock);
 884         list_del_init(&ldt->ldt_linkage);
 885         spin_unlock(&obd_types_lock);
 886         if (ldt->ldt_ops->ldto_fini)
 887                 ldt->ldt_ops->ldto_fini(ldt);
 888 }
 889 EXPORT_SYMBOL(lu_device_type_fini);
 890
 891 /**
 892  * Global list of all sites on this node
 893  */
 894 static struct list_head lu_sites;
 895 static DEFINE_MUTEX(lu_sites_guard);
 896
 897 /**
 898  * Global environment used by site shrinker.
 899  */
 900 static struct lu_env lu_shrink_env;
 901
 902 struct lu_site_print_arg {
 903         struct lu_env   *lsp_env;
 904         void            *lsp_cookie;
 905         lu_printer_t     lsp_printer;
 906 };
 907
 908 static int
 909 lu_site_obj_print(cfs_hash_t *hs, cfs_hash_bd_t *bd,
 910                   struct hlist_node *hnode, void *data)
 911 {
 912         struct lu_site_print_arg *arg = (struct lu_site_print_arg *)data;
 913         struct lu_object_header  *h;
 914
 915         h = hlist_entry(hnode, struct lu_object_header, loh_hash);
 916         if (!list_empty(&h->loh_layers)) {
 917                 const struct lu_object *o;
 918
 919                 o = lu_object_top(h);
 920                 lu_object_print(arg->lsp_env, arg->lsp_cookie,
 921                                 arg->lsp_printer, o);
 922         } else {
 923                 lu_object_header_print(arg->lsp_env, arg->lsp_cookie,
 924                                        arg->lsp_printer, h);
 925         }
 926         return 0;
 927 }
 928
 929 /**
 930  * Print all objects in \a s.
 931  */
 932 void lu_site_print(const struct lu_env *env, struct lu_site *s, void *cookie,
 933                    lu_printer_t printer)
 934 {
 935         struct lu_site_print_arg arg = {
 936                 .lsp_env     = (struct lu_env *)env,
 937                 .lsp_cookie  = cookie,
 938                 .lsp_printer = printer,
 939         };
 940
 941         cfs_hash_for_each(s->ls_obj_hash, lu_site_obj_print, &arg);
 942 }
 943 EXPORT_SYMBOL(lu_site_print);
 944
 945 /**
 946  * Return desired hash table order.
 947  */
 948 static unsigned int lu_htable_order(struct lu_device *top)
 949 {
 950         unsigned long cache_size;
 951         unsigned int  bits;
 952
 953         /*
 954          * For ZFS based OSDs the cache should be disabled by default.  This
 955          * allows the ZFS ARC maximum flexibility in determining what buffers
 956          * to cache.  If Lustre has objects or buffer which it wants to ensure
 957          * always stay cached it must maintain a hold on them.
 958          */
 959         if (strcmp(top->ld_type->ldt_name, LUSTRE_OSD_ZFS_NAME) == 0) {
 960                 lu_cache_percent = 1;
 961                 lu_cache_nr = LU_CACHE_NR_ZFS_LIMIT;
 962                 return LU_SITE_BITS_MIN;
 963         }
 964
 965         /*
 966          * Calculate hash table size, assuming that we want reasonable
 967          * performance when 20% of total memory is occupied by cache of
 968          * lu_objects.
 969          *
 970          * Size of lu_object is (arbitrary) taken as 1K (together with inode).
 971          */
 972         cache_size = totalram_pages;
 973
 974 #if BITS_PER_LONG == 32
 975         /* limit hashtable size for lowmem systems to low RAM */
 976         if (cache_size > 1 << (30 - PAGE_CACHE_SHIFT))
 977                 cache_size = 1 << (30 - PAGE_CACHE_SHIFT) * 3 / 4;
 978 #endif
 979
 980         /* clear off unreasonable cache setting. */
 981         if (lu_cache_percent == 0 || lu_cache_percent > LU_CACHE_PERCENT_MAX) {
 982                 CWARN("obdclass: invalid lu_cache_percent: %u, it must be in"
 983                       " the range of (0, %u]. Will use default value: %u.\n",
 984                       lu_cache_percent, LU_CACHE_PERCENT_MAX,
 985                       LU_CACHE_PERCENT_DEFAULT);
 986
 987                 lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
 988         }
 989         cache_size = cache_size / 100 * lu_cache_percent *
 990                 (PAGE_CACHE_SIZE / 1024);
 991
 992         for (bits = 1; (1 << bits) < cache_size; ++bits) {
 993                 ;
 994         }
 995         return bits;
 996 }
 997
 998 static unsigned lu_obj_hop_hash(cfs_hash_t *hs,
 999                                 const void *key, unsigned mask)
1000 {
1001         struct lu_fid  *fid = (struct lu_fid *)key;
1002         __u32           hash;
1003
1004         hash = fid_flatten32(fid);
1005         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
1006         hash = hash_long(hash, hs->hs_bkt_bits);
1007
1008         /* give me another random factor */
1009         hash -= hash_long((unsigned long)hs, fid_oid(fid) % 11 + 3);
1010
1011         hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
1012         hash |= (fid_seq(fid) + fid_oid(fid)) & (CFS_HASH_NBKT(hs) - 1);
1013
1014         return hash & mask;
1015 }
1016
1017 static void *lu_obj_hop_object(struct hlist_node *hnode)
1018 {
1019         return hlist_entry(hnode, struct lu_object_header, loh_hash);
1020 }
1021
1022 static void *lu_obj_hop_key(struct hlist_node *hnode)
1023 {
1024         struct lu_object_header *h;
1025
1026         h = hlist_entry(hnode, struct lu_object_header, loh_hash);
1027         return &h->loh_fid;
1028 }
1029
1030 static int lu_obj_hop_keycmp(const void *key, struct hlist_node *hnode)
1031 {
1032         struct lu_object_header *h;
1033
1034         h = hlist_entry(hnode, struct lu_object_header, loh_hash);
1035         return lu_fid_eq(&h->loh_fid, (struct lu_fid *)key);
1036 }
1037
1038 static void lu_obj_hop_get(cfs_hash_t *hs, struct hlist_node *hnode)
1039 {
1040         struct lu_object_header *h;
1041
1042         h = hlist_entry(hnode, struct lu_object_header, loh_hash);
1043         if (atomic_add_return(1, &h->loh_ref) == 1) {
1044                 struct lu_site_bkt_data *bkt;
1045                 cfs_hash_bd_t            bd;
1046
1047                 cfs_hash_bd_get(hs, &h->loh_fid, &bd);
1048                 bkt = cfs_hash_bd_extra_get(hs, &bd);
1049                 bkt->lsb_busy++;
1050         }
1051 }
1052
1053 static void lu_obj_hop_put_locked(cfs_hash_t *hs, struct hlist_node *hnode)
1054 {
1055         LBUG(); /* we should never called it */
1056 }
1057
1058 cfs_hash_ops_t lu_site_hash_ops = {
1059         .hs_hash        = lu_obj_hop_hash,
1060         .hs_key         = lu_obj_hop_key,
1061         .hs_keycmp      = lu_obj_hop_keycmp,
1062         .hs_object      = lu_obj_hop_object,
1063         .hs_get         = lu_obj_hop_get,
1064         .hs_put_locked  = lu_obj_hop_put_locked,
1065 };
1066
1067 void lu_dev_add_linkage(struct lu_site *s, struct lu_device *d)
1068 {
1069         spin_lock(&s->ls_ld_lock);
1070         if (list_empty(&d->ld_linkage))
1071                 list_add(&d->ld_linkage, &s->ls_ld_linkage);
1072         spin_unlock(&s->ls_ld_lock);
1073 }
1074 EXPORT_SYMBOL(lu_dev_add_linkage);
1075
1076 void lu_dev_del_linkage(struct lu_site *s, struct lu_device *d)
1077 {
1078         spin_lock(&s->ls_ld_lock);
1079         list_del_init(&d->ld_linkage);
1080         spin_unlock(&s->ls_ld_lock);
1081 }
1082 EXPORT_SYMBOL(lu_dev_del_linkage);
1083
1084 /**
1085   * Initialize site \a s, with \a d as the top level device.
1086   */
1087 int lu_site_init(struct lu_site *s, struct lu_device *top)
1088 {
1089         struct lu_site_bkt_data *bkt;
1090         cfs_hash_bd_t bd;
1091         char name[16];
1092         unsigned int bits;
1093         unsigned int i;
1094         ENTRY;
1095
1096         memset(s, 0, sizeof *s);
1097         mutex_init(&s->ls_purge_mutex);
1098         bits = lu_htable_order(top);
1099         snprintf(name, 16, "lu_site_%s", top->ld_type->ldt_name);
1100         for (bits = clamp_t(typeof(bits), bits,
1101                             LU_SITE_BITS_MIN, LU_SITE_BITS_MAX);
1102              bits >= LU_SITE_BITS_MIN; bits--) {
1103                 s->ls_obj_hash = cfs_hash_create(name, bits, bits,
1104                                                  bits - LU_SITE_BKT_BITS,
1105                                                  sizeof(*bkt), 0, 0,
1106                                                  &lu_site_hash_ops,
1107                                                  CFS_HASH_SPIN_BKTLOCK |
1108                                                  CFS_HASH_NO_ITEMREF |
1109                                                  CFS_HASH_DEPTH |
1110                                                  CFS_HASH_ASSERT_EMPTY |
1111                                                  CFS_HASH_COUNTER);
1112                 if (s->ls_obj_hash != NULL)
1113                         break;
1114         }
1115
1116         if (s->ls_obj_hash == NULL) {
1117                 CERROR("failed to create lu_site hash with bits: %d\n", bits);
1118                 return -ENOMEM;
1119         }
1120
1121         cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
1122                 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
1123                 INIT_LIST_HEAD(&bkt->lsb_lru);
1124                 init_waitqueue_head(&bkt->lsb_marche_funebre);
1125         }
1126
1127         s->ls_stats = lprocfs_alloc_stats(LU_SS_LAST_STAT, 0);
1128         if (s->ls_stats == NULL) {
1129                 cfs_hash_putref(s->ls_obj_hash);
1130                 s->ls_obj_hash = NULL;
1131                 return -ENOMEM;
1132         }
1133
1134         lprocfs_counter_init(s->ls_stats, LU_SS_CREATED,
1135                              0, "created", "created");
1136         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_HIT,
1137                              0, "cache_hit", "cache_hit");
1138         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_MISS,
1139                              0, "cache_miss", "cache_miss");
1140         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_RACE,
1141                              0, "cache_race", "cache_race");
1142         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_DEATH_RACE,
1143                              0, "cache_death_race", "cache_death_race");
1144         lprocfs_counter_init(s->ls_stats, LU_SS_LRU_PURGED,
1145                              0, "lru_purged", "lru_purged");
1146
1147         INIT_LIST_HEAD(&s->ls_linkage);
1148         s->ls_top_dev = top;
1149         top->ld_site = s;
1150         lu_device_get(top);
1151         lu_ref_add(&top->ld_reference, "site-top", s);
1152
1153         INIT_LIST_HEAD(&s->ls_ld_linkage);
1154         spin_lock_init(&s->ls_ld_lock);
1155
1156         lu_dev_add_linkage(s, top);
1157
1158         RETURN(0);
1159 }
1160 EXPORT_SYMBOL(lu_site_init);
1161
1162 /**
1163  * Finalize \a s and release its resources.
1164  */
1165 void lu_site_fini(struct lu_site *s)
1166 {
1167         mutex_lock(&lu_sites_guard);
1168         list_del_init(&s->ls_linkage);
1169         mutex_unlock(&lu_sites_guard);
1170
1171         if (s->ls_obj_hash != NULL) {
1172                 cfs_hash_putref(s->ls_obj_hash);
1173                 s->ls_obj_hash = NULL;
1174         }
1175
1176         if (s->ls_top_dev != NULL) {
1177                 s->ls_top_dev->ld_site = NULL;
1178                 lu_ref_del(&s->ls_top_dev->ld_reference, "site-top", s);
1179                 lu_device_put(s->ls_top_dev);
1180                 s->ls_top_dev = NULL;
1181         }
1182
1183         if (s->ls_stats != NULL)
1184                 lprocfs_free_stats(&s->ls_stats);
1185 }
1186 EXPORT_SYMBOL(lu_site_fini);
1187
1188 /**
1189  * Called when initialization of stack for this site is completed.
1190  */
1191 int lu_site_init_finish(struct lu_site *s)
1192 {
1193         int result;
1194         mutex_lock(&lu_sites_guard);
1195         result = lu_context_refill(&lu_shrink_env.le_ctx);
1196         if (result == 0)
1197                 list_add(&s->ls_linkage, &lu_sites);
1198         mutex_unlock(&lu_sites_guard);
1199         return result;
1200 }
1201 EXPORT_SYMBOL(lu_site_init_finish);
1202
1203 /**
1204  * Acquire additional reference on device \a d
1205  */
1206 void lu_device_get(struct lu_device *d)
1207 {
1208         atomic_inc(&d->ld_ref);
1209 }
1210 EXPORT_SYMBOL(lu_device_get);
1211
1212 /**
1213  * Release reference on device \a d.
1214  */
1215 void lu_device_put(struct lu_device *d)
1216 {
1217         LASSERT(atomic_read(&d->ld_ref) > 0);
1218         atomic_dec(&d->ld_ref);
1219 }
1220 EXPORT_SYMBOL(lu_device_put);
1221
1222 /**
1223  * Initialize device \a d of type \a t.
1224  */
1225 int lu_device_init(struct lu_device *d, struct lu_device_type *t)
1226 {
1227         if (atomic_inc_return(&t->ldt_device_nr) == 1 &&
1228             t->ldt_ops->ldto_start != NULL)
1229                 t->ldt_ops->ldto_start(t);
1230
1231         memset(d, 0, sizeof *d);
1232         d->ld_type = t;
1233         lu_ref_init(&d->ld_reference);
1234         INIT_LIST_HEAD(&d->ld_linkage);
1235
1236         return 0;
1237 }
1238 EXPORT_SYMBOL(lu_device_init);
1239
1240 /**
1241  * Finalize device \a d.
1242  */
1243 void lu_device_fini(struct lu_device *d)
1244 {
1245         struct lu_device_type *t = d->ld_type;
1246
1247         if (d->ld_obd != NULL) {
1248                 d->ld_obd->obd_lu_dev = NULL;
1249                 d->ld_obd = NULL;
1250         }
1251
1252         lu_ref_fini(&d->ld_reference);
1253         LASSERTF(atomic_read(&d->ld_ref) == 0,
1254                  "Refcount is %u\n", atomic_read(&d->ld_ref));
1255         LASSERT(atomic_read(&t->ldt_device_nr) > 0);
1256
1257         if (atomic_dec_and_test(&t->ldt_device_nr) &&
1258             t->ldt_ops->ldto_stop != NULL)
1259                 t->ldt_ops->ldto_stop(t);
1260 }
1261 EXPORT_SYMBOL(lu_device_fini);
1262
1263 /**
1264  * Initialize object \a o that is part of compound object \a h and was created
1265  * by device \a d.
1266  */
1267 int lu_object_init(struct lu_object *o, struct lu_object_header *h,
1268                    struct lu_device *d)
1269 {
1270         memset(o, 0, sizeof(*o));
1271         o->lo_header = h;
1272         o->lo_dev = d;
1273         lu_device_get(d);
1274         lu_ref_add_at(&d->ld_reference, &o->lo_dev_ref, "lu_object", o);
1275         INIT_LIST_HEAD(&o->lo_linkage);
1276
1277         return 0;
1278 }
1279 EXPORT_SYMBOL(lu_object_init);
1280
1281 /**
1282  * Finalize object and release its resources.
1283  */
1284 void lu_object_fini(struct lu_object *o)
1285 {
1286         struct lu_device *dev = o->lo_dev;
1287
1288         LASSERT(list_empty(&o->lo_linkage));
1289
1290         if (dev != NULL) {
1291                 lu_ref_del_at(&dev->ld_reference, &o->lo_dev_ref,
1292                               "lu_object", o);
1293                 lu_device_put(dev);
1294                 o->lo_dev = NULL;
1295         }
1296 }
1297 EXPORT_SYMBOL(lu_object_fini);
1298
1299 /**
1300  * Add object \a o as first layer of compound object \a h
1301  *
1302  * This is typically called by the ->ldo_object_alloc() method of top-level
1303  * device.
1304  */
1305 void lu_object_add_top(struct lu_object_header *h, struct lu_object *o)
1306 {
1307         list_move(&o->lo_linkage, &h->loh_layers);
1308 }
1309 EXPORT_SYMBOL(lu_object_add_top);
1310
1311 /**
1312  * Add object \a o as a layer of compound object, going after \a before.
1313  *
1314  * This is typically called by the ->ldo_object_alloc() method of \a
1315  * before->lo_dev.
1316  */
1317 void lu_object_add(struct lu_object *before, struct lu_object *o)
1318 {
1319         list_move(&o->lo_linkage, &before->lo_linkage);
1320 }
1321 EXPORT_SYMBOL(lu_object_add);
1322
1323 /**
1324  * Initialize compound object.
1325  */
1326 int lu_object_header_init(struct lu_object_header *h)
1327 {
1328         memset(h, 0, sizeof *h);
1329         atomic_set(&h->loh_ref, 1);
1330         INIT_HLIST_NODE(&h->loh_hash);
1331         INIT_LIST_HEAD(&h->loh_lru);
1332         INIT_LIST_HEAD(&h->loh_layers);
1333         lu_ref_init(&h->loh_reference);
1334         return 0;
1335 }
1336 EXPORT_SYMBOL(lu_object_header_init);
1337
1338 /**
1339  * Finalize compound object.
1340  */
1341 void lu_object_header_fini(struct lu_object_header *h)
1342 {
1343         LASSERT(list_empty(&h->loh_layers));
1344         LASSERT(list_empty(&h->loh_lru));
1345         LASSERT(hlist_unhashed(&h->loh_hash));
1346         lu_ref_fini(&h->loh_reference);
1347 }
1348 EXPORT_SYMBOL(lu_object_header_fini);
1349
1350 /**
1351  * Given a compound object, find its slice, corresponding to the device type
1352  * \a dtype.
1353  */
1354 struct lu_object *lu_object_locate(struct lu_object_header *h,
1355                                    const struct lu_device_type *dtype)
1356 {
1357         struct lu_object *o;
1358
1359         list_for_each_entry(o, &h->loh_layers, lo_linkage) {
1360                 if (o->lo_dev->ld_type == dtype)
1361                         return o;
1362         }
1363         return NULL;
1364 }
1365 EXPORT_SYMBOL(lu_object_locate);
1366
1367 /**
1368  * Finalize and free devices in the device stack.
1369  *
1370  * Finalize device stack by purging object cache, and calling
1371  * lu_device_type_operations::ldto_device_fini() and
1372  * lu_device_type_operations::ldto_device_free() on all devices in the stack.
1373  */
1374 void lu_stack_fini(const struct lu_env *env, struct lu_device *top)
1375 {
1376         struct lu_site   *site = top->ld_site;
1377         struct lu_device *scan;
1378         struct lu_device *next;
1379
1380         lu_site_purge(env, site, ~0);
1381         for (scan = top; scan != NULL; scan = next) {
1382                 next = scan->ld_type->ldt_ops->ldto_device_fini(env, scan);
1383                 lu_ref_del(&scan->ld_reference, "lu-stack", &lu_site_init);
1384                 lu_device_put(scan);
1385         }
1386
1387         /* purge again. */
1388         lu_site_purge(env, site, ~0);
1389
1390         for (scan = top; scan != NULL; scan = next) {
1391                 const struct lu_device_type *ldt = scan->ld_type;
1392                 struct obd_type             *type;
1393
1394                 next = ldt->ldt_ops->ldto_device_free(env, scan);
1395                 type = ldt->ldt_obd_type;
1396                 if (type != NULL) {
1397                         type->typ_refcnt--;
1398                         class_put_type(type);
1399                 }
1400         }
1401 }
1402 EXPORT_SYMBOL(lu_stack_fini);
1403
1404 enum {
1405         /**
1406          * Maximal number of tld slots.
1407          */
1408         LU_CONTEXT_KEY_NR = 40
1409 };
1410
1411 static struct lu_context_key *lu_keys[LU_CONTEXT_KEY_NR] = { NULL, };
1412
1413 static DEFINE_SPINLOCK(lu_keys_guard);
1414
1415 /**
1416  * Global counter incremented whenever key is registered, unregistered,
1417  * revived or quiesced. This is used to void unnecessary calls to
1418  * lu_context_refill(). No locking is provided, as initialization and shutdown
1419  * are supposed to be externally serialized.
1420  */
1421 static unsigned key_set_version = 0;
1422
1423 /**
1424  * Register new key.
1425  */
1426 int lu_context_key_register(struct lu_context_key *key)
1427 {
1428         int result;
1429         unsigned int i;
1430
1431         LASSERT(key->lct_init != NULL);
1432         LASSERT(key->lct_fini != NULL);
1433         LASSERT(key->lct_tags != 0);
1434         LASSERT(key->lct_owner != NULL);
1435
1436         result = -ENFILE;
1437         spin_lock(&lu_keys_guard);
1438         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1439                 if (lu_keys[i] == NULL) {
1440                         key->lct_index = i;
1441                         atomic_set(&key->lct_used, 1);
1442                         lu_keys[i] = key;
1443                         lu_ref_init(&key->lct_reference);
1444                         result = 0;
1445                         ++key_set_version;
1446                         break;
1447                 }
1448         }
1449         spin_unlock(&lu_keys_guard);
1450         return result;
1451 }
1452 EXPORT_SYMBOL(lu_context_key_register);
1453
1454 static void key_fini(struct lu_context *ctx, int index)
1455 {
1456         if (ctx->lc_value != NULL && ctx->lc_value[index] != NULL) {
1457                 struct lu_context_key *key;
1458
1459                 key = lu_keys[index];
1460                 LASSERT(key != NULL);
1461                 LASSERT(key->lct_fini != NULL);
1462                 LASSERT(atomic_read(&key->lct_used) > 1);
1463
1464                 key->lct_fini(ctx, key, ctx->lc_value[index]);
1465                 lu_ref_del(&key->lct_reference, "ctx", ctx);
1466                 atomic_dec(&key->lct_used);
1467
1468                 LASSERT(key->lct_owner != NULL);
1469                 if ((ctx->lc_tags & LCT_NOREF) == 0) {
1470                         LINVRNT(module_refcount(key->lct_owner) > 0);
1471                         module_put(key->lct_owner);
1472                 }
1473                 ctx->lc_value[index] = NULL;
1474         }
1475 }
1476
1477 /**
1478  * Deregister key.
1479  */
1480 void lu_context_key_degister(struct lu_context_key *key)
1481 {
1482         LASSERT(atomic_read(&key->lct_used) >= 1);
1483         LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1484
1485         lu_context_key_quiesce(key);
1486
1487         ++key_set_version;
1488         spin_lock(&lu_keys_guard);
1489         key_fini(&lu_shrink_env.le_ctx, key->lct_index);
1490         if (lu_keys[key->lct_index]) {
1491                 lu_keys[key->lct_index] = NULL;
1492                 lu_ref_fini(&key->lct_reference);
1493         }
1494         spin_unlock(&lu_keys_guard);
1495
1496         LASSERTF(atomic_read(&key->lct_used) == 1,
1497                  "key has instances: %d\n",
1498                  atomic_read(&key->lct_used));
1499 }
1500 EXPORT_SYMBOL(lu_context_key_degister);
1501
1502 /**
1503  * Register a number of keys. This has to be called after all keys have been
1504  * initialized by a call to LU_CONTEXT_KEY_INIT().
1505  */
1506 int lu_context_key_register_many(struct lu_context_key *k, ...)
1507 {
1508         struct lu_context_key *key = k;
1509         va_list args;
1510         int result;
1511
1512         va_start(args, k);
1513         do {
1514                 result = lu_context_key_register(key);
1515                 if (result)
1516                         break;
1517                 key = va_arg(args, struct lu_context_key *);
1518         } while (key != NULL);
1519         va_end(args);
1520
1521         if (result != 0) {
1522                 va_start(args, k);
1523                 while (k != key) {
1524                         lu_context_key_degister(k);
1525                         k = va_arg(args, struct lu_context_key *);
1526                 }
1527                 va_end(args);
1528         }
1529
1530         return result;
1531 }
1532 EXPORT_SYMBOL(lu_context_key_register_many);
1533
1534 /**
1535  * De-register a number of keys. This is a dual to
1536  * lu_context_key_register_many().
1537  */
1538 void lu_context_key_degister_many(struct lu_context_key *k, ...)
1539 {
1540         va_list args;
1541
1542         va_start(args, k);
1543         do {
1544                 lu_context_key_degister(k);
1545                 k = va_arg(args, struct lu_context_key*);
1546         } while (k != NULL);
1547         va_end(args);
1548 }
1549 EXPORT_SYMBOL(lu_context_key_degister_many);
1550
1551 /**
1552  * Revive a number of keys.
1553  */
1554 void lu_context_key_revive_many(struct lu_context_key *k, ...)
1555 {
1556         va_list args;
1557
1558         va_start(args, k);
1559         do {
1560                 lu_context_key_revive(k);
1561                 k = va_arg(args, struct lu_context_key*);
1562         } while (k != NULL);
1563         va_end(args);
1564 }
1565 EXPORT_SYMBOL(lu_context_key_revive_many);
1566
1567 /**
1568  * Quiescent a number of keys.
1569  */
1570 void lu_context_key_quiesce_many(struct lu_context_key *k, ...)
1571 {
1572         va_list args;
1573
1574         va_start(args, k);
1575         do {
1576                 lu_context_key_quiesce(k);
1577                 k = va_arg(args, struct lu_context_key*);
1578         } while (k != NULL);
1579         va_end(args);
1580 }
1581 EXPORT_SYMBOL(lu_context_key_quiesce_many);
1582
1583 /**
1584  * Return value associated with key \a key in context \a ctx.
1585  */
1586 void *lu_context_key_get(const struct lu_context *ctx,
1587                          const struct lu_context_key *key)
1588 {
1589         LINVRNT(ctx->lc_state == LCS_ENTERED);
1590         LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1591         LASSERT(lu_keys[key->lct_index] == key);
1592         return ctx->lc_value[key->lct_index];
1593 }
1594 EXPORT_SYMBOL(lu_context_key_get);
1595
1596 /**
1597  * List of remembered contexts. XXX document me.
1598  */
1599 static struct list_head lu_context_remembered;
1600
1601 /**
1602  * Destroy \a key in all remembered contexts. This is used to destroy key
1603  * values in "shared" contexts (like service threads), when a module owning
1604  * the key is about to be unloaded.
1605  */
1606 void lu_context_key_quiesce(struct lu_context_key *key)
1607 {
1608         struct lu_context *ctx;
1609         extern unsigned cl_env_cache_purge(unsigned nr);
1610
1611         if (!(key->lct_tags & LCT_QUIESCENT)) {
1612                 /*
1613                  * XXX layering violation.
1614                  */
1615                 cl_env_cache_purge(~0);
1616                 key->lct_tags |= LCT_QUIESCENT;
1617                 /*
1618                  * XXX memory barrier has to go here.
1619                  */
1620                 spin_lock(&lu_keys_guard);
1621                 list_for_each_entry(ctx, &lu_context_remembered,
1622                                     lc_remember)
1623                         key_fini(ctx, key->lct_index);
1624                 spin_unlock(&lu_keys_guard);
1625                 ++key_set_version;
1626         }
1627 }
1628 EXPORT_SYMBOL(lu_context_key_quiesce);
1629
1630 void lu_context_key_revive(struct lu_context_key *key)
1631 {
1632         key->lct_tags &= ~LCT_QUIESCENT;
1633         ++key_set_version;
1634 }
1635 EXPORT_SYMBOL(lu_context_key_revive);
1636
1637 static void keys_fini(struct lu_context *ctx)
1638 {
1639         unsigned int i;
1640
1641         if (ctx->lc_value == NULL)
1642                 return;
1643
1644         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i)
1645                 key_fini(ctx, i);
1646
1647         OBD_FREE(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1648         ctx->lc_value = NULL;
1649 }
1650
1651 static int keys_fill(struct lu_context *ctx)
1652 {
1653         unsigned int i;
1654
1655         LINVRNT(ctx->lc_value != NULL);
1656         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1657                 struct lu_context_key *key;
1658
1659                 key = lu_keys[i];
1660                 if (ctx->lc_value[i] == NULL && key != NULL &&
1661                     (key->lct_tags & ctx->lc_tags) &&
1662                     /*
1663                      * Don't create values for a LCT_QUIESCENT key, as this
1664                      * will pin module owning a key.
1665                      */
1666                     !(key->lct_tags & LCT_QUIESCENT)) {
1667                         void *value;
1668
1669                         LINVRNT(key->lct_init != NULL);
1670                         LINVRNT(key->lct_index == i);
1671
1672                         value = key->lct_init(ctx, key);
1673                         if (unlikely(IS_ERR(value)))
1674                                 return PTR_ERR(value);
1675
1676                         LASSERT(key->lct_owner != NULL);
1677                         if (!(ctx->lc_tags & LCT_NOREF))
1678                                 try_module_get(key->lct_owner);
1679                         lu_ref_add_atomic(&key->lct_reference, "ctx", ctx);
1680                         atomic_inc(&key->lct_used);
1681                         /*
1682                          * This is the only place in the code, where an
1683                          * element of ctx->lc_value[] array is set to non-NULL
1684                          * value.
1685                          */
1686                         ctx->lc_value[i] = value;
1687                         if (key->lct_exit != NULL)
1688                                 ctx->lc_tags |= LCT_HAS_EXIT;
1689                 }
1690                 ctx->lc_version = key_set_version;
1691         }
1692         return 0;
1693 }
1694
1695 static int keys_init(struct lu_context *ctx)
1696 {
1697         OBD_ALLOC(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1698         if (likely(ctx->lc_value != NULL))
1699                 return keys_fill(ctx);
1700
1701         return -ENOMEM;
1702 }
1703
1704 /**
1705  * Initialize context data-structure. Create values for all keys.
1706  */
1707 int lu_context_init(struct lu_context *ctx, __u32 tags)
1708 {
1709         int     rc;
1710
1711         memset(ctx, 0, sizeof *ctx);
1712         ctx->lc_state = LCS_INITIALIZED;
1713         ctx->lc_tags = tags;
1714         if (tags & LCT_REMEMBER) {
1715                 spin_lock(&lu_keys_guard);
1716                 list_add(&ctx->lc_remember, &lu_context_remembered);
1717                 spin_unlock(&lu_keys_guard);
1718         } else {
1719                 INIT_LIST_HEAD(&ctx->lc_remember);
1720         }
1721
1722         rc = keys_init(ctx);
1723         if (rc != 0)
1724                 lu_context_fini(ctx);
1725
1726         return rc;
1727 }
1728 EXPORT_SYMBOL(lu_context_init);
1729
1730 /**
1731  * Finalize context data-structure. Destroy key values.
1732  */
1733 void lu_context_fini(struct lu_context *ctx)
1734 {
1735         LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1736         ctx->lc_state = LCS_FINALIZED;
1737
1738         if ((ctx->lc_tags & LCT_REMEMBER) == 0) {
1739                 LASSERT(list_empty(&ctx->lc_remember));
1740                 keys_fini(ctx);
1741
1742         } else { /* could race with key degister */
1743                 spin_lock(&lu_keys_guard);
1744                 keys_fini(ctx);
1745                 list_del_init(&ctx->lc_remember);
1746                 spin_unlock(&lu_keys_guard);
1747         }
1748 }
1749 EXPORT_SYMBOL(lu_context_fini);
1750
1751 /**
1752  * Called before entering context.
1753  */
1754 void lu_context_enter(struct lu_context *ctx)
1755 {
1756         LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1757         ctx->lc_state = LCS_ENTERED;
1758 }
1759 EXPORT_SYMBOL(lu_context_enter);
1760
1761 /**
1762  * Called after exiting from \a ctx
1763  */
1764 void lu_context_exit(struct lu_context *ctx)
1765 {
1766         unsigned int i;
1767
1768         LINVRNT(ctx->lc_state == LCS_ENTERED);
1769         ctx->lc_state = LCS_LEFT;
1770         if (ctx->lc_tags & LCT_HAS_EXIT && ctx->lc_value != NULL) {
1771                 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1772                         if (ctx->lc_value[i] != NULL) {
1773                                 struct lu_context_key *key;
1774
1775                                 key = lu_keys[i];
1776                                 LASSERT(key != NULL);
1777                                 if (key->lct_exit != NULL)
1778                                         key->lct_exit(ctx,
1779                                                       key, ctx->lc_value[i]);
1780                         }
1781                 }
1782         }
1783 }
1784 EXPORT_SYMBOL(lu_context_exit);
1785
1786 /**
1787  * Allocate for context all missing keys that were registered after context
1788  * creation. key_set_version is only changed in rare cases when modules
1789  * are loaded and removed.
1790  */
1791 int lu_context_refill(struct lu_context *ctx)
1792 {
1793         return likely(ctx->lc_version == key_set_version) ? 0 : keys_fill(ctx);
1794 }
1795 EXPORT_SYMBOL(lu_context_refill);
1796
1797 /**
1798  * lu_ctx_tags/lu_ses_tags will be updated if there are new types of
1799  * obd being added. Currently, this is only used on client side, specifically
1800  * for echo device client, for other stack (like ptlrpc threads), context are
1801  * predefined when the lu_device type are registered, during the module probe
1802  * phase.
1803  */
1804 __u32 lu_context_tags_default = 0;
1805 __u32 lu_session_tags_default = 0;
1806
1807 void lu_context_tags_update(__u32 tags)
1808 {
1809         spin_lock(&lu_keys_guard);
1810         lu_context_tags_default |= tags;
1811         key_set_version++;
1812         spin_unlock(&lu_keys_guard);
1813 }
1814 EXPORT_SYMBOL(lu_context_tags_update);
1815
1816 void lu_context_tags_clear(__u32 tags)
1817 {
1818         spin_lock(&lu_keys_guard);
1819         lu_context_tags_default &= ~tags;
1820         key_set_version++;
1821         spin_unlock(&lu_keys_guard);
1822 }
1823 EXPORT_SYMBOL(lu_context_tags_clear);
1824
1825 void lu_session_tags_update(__u32 tags)
1826 {
1827         spin_lock(&lu_keys_guard);
1828         lu_session_tags_default |= tags;
1829         key_set_version++;
1830         spin_unlock(&lu_keys_guard);
1831 }
1832 EXPORT_SYMBOL(lu_session_tags_update);
1833
1834 void lu_session_tags_clear(__u32 tags)
1835 {
1836         spin_lock(&lu_keys_guard);
1837         lu_session_tags_default &= ~tags;
1838         key_set_version++;
1839         spin_unlock(&lu_keys_guard);
1840 }
1841 EXPORT_SYMBOL(lu_session_tags_clear);
1842
1843 int lu_env_init(struct lu_env *env, __u32 tags)
1844 {
1845         int result;
1846
1847         env->le_ses = NULL;
1848         result = lu_context_init(&env->le_ctx, tags);
1849         if (likely(result == 0))
1850                 lu_context_enter(&env->le_ctx);
1851         return result;
1852 }
1853 EXPORT_SYMBOL(lu_env_init);
1854
1855 void lu_env_fini(struct lu_env *env)
1856 {
1857         lu_context_exit(&env->le_ctx);
1858         lu_context_fini(&env->le_ctx);
1859         env->le_ses = NULL;
1860 }
1861 EXPORT_SYMBOL(lu_env_fini);
1862
1863 int lu_env_refill(struct lu_env *env)
1864 {
1865         int result;
1866
1867         result = lu_context_refill(&env->le_ctx);
1868         if (result == 0 && env->le_ses != NULL)
1869                 result = lu_context_refill(env->le_ses);
1870         return result;
1871 }
1872 EXPORT_SYMBOL(lu_env_refill);
1873
1874 /**
1875  * Currently, this API will only be used by echo client.
1876  * Because echo client and normal lustre client will share
1877  * same cl_env cache. So echo client needs to refresh
1878  * the env context after it get one from the cache, especially
1879  * when normal client and echo client co-exist in the same client.
1880  */
1881 int lu_env_refill_by_tags(struct lu_env *env, __u32 ctags,
1882                           __u32 stags)
1883 {
1884         int    result;
1885
1886         if ((env->le_ctx.lc_tags & ctags) != ctags) {
1887                 env->le_ctx.lc_version = 0;
1888                 env->le_ctx.lc_tags |= ctags;
1889         }
1890
1891         if (env->le_ses && (env->le_ses->lc_tags & stags) != stags) {
1892                 env->le_ses->lc_version = 0;
1893                 env->le_ses->lc_tags |= stags;
1894         }
1895
1896         result = lu_env_refill(env);
1897
1898         return result;
1899 }
1900 EXPORT_SYMBOL(lu_env_refill_by_tags);
1901
1902 static struct shrinker *lu_site_shrinker;
1903
1904 typedef struct lu_site_stats{
1905         unsigned        lss_populated;
1906         unsigned        lss_max_search;
1907         unsigned        lss_total;
1908         unsigned        lss_busy;
1909 } lu_site_stats_t;
1910
1911 static void lu_site_stats_get(cfs_hash_t *hs,
1912                               lu_site_stats_t *stats, int populated)
1913 {
1914         cfs_hash_bd_t bd;
1915         unsigned int  i;
1916
1917         cfs_hash_for_each_bucket(hs, &bd, i) {
1918                 struct lu_site_bkt_data *bkt = cfs_hash_bd_extra_get(hs, &bd);
1919                 struct hlist_head       *hhead;
1920
1921                 cfs_hash_bd_lock(hs, &bd, 1);
1922                 stats->lss_busy  += bkt->lsb_busy;
1923                 stats->lss_total += cfs_hash_bd_count_get(&bd);
1924                 stats->lss_max_search = max((int)stats->lss_max_search,
1925                                             cfs_hash_bd_depmax_get(&bd));
1926                 if (!populated) {
1927                         cfs_hash_bd_unlock(hs, &bd, 1);
1928                         continue;
1929                 }
1930
1931                 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
1932                         if (!hlist_empty(hhead))
1933                                 stats->lss_populated++;
1934                 }
1935                 cfs_hash_bd_unlock(hs, &bd, 1);
1936         }
1937 }
1938
1939
1940 static unsigned long lu_cache_shrink_count(struct shrinker *sk,
1941                                            struct shrink_control *sc)
1942 {
1943         lu_site_stats_t stats;
1944         struct lu_site *s;
1945         struct lu_site *tmp;
1946         unsigned long cached = 0;
1947
1948         if (!(sc->gfp_mask & __GFP_FS))
1949                 return 0;
1950
1951         mutex_lock(&lu_sites_guard);
1952         list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
1953                 memset(&stats, 0, sizeof(stats));
1954                 lu_site_stats_get(s->ls_obj_hash, &stats, 0);
1955                 cached += stats.lss_total - stats.lss_busy;
1956         }
1957         mutex_unlock(&lu_sites_guard);
1958
1959         cached = (cached / 100) * sysctl_vfs_cache_pressure;
1960         CDEBUG(D_INODE, "%ld objects cached\n", cached);
1961         return cached;
1962 }
1963
1964 static unsigned long lu_cache_shrink_scan(struct shrinker *sk,
1965                                           struct shrink_control *sc)
1966 {
1967         struct lu_site *s;
1968         struct lu_site *tmp;
1969         unsigned long remain = sc->nr_to_scan;
1970         LIST_HEAD(splice);
1971
1972         if (!(sc->gfp_mask & __GFP_FS))
1973                 /* We must not take the lu_sites_guard lock when
1974                  * __GFP_FS is *not* set because of the deadlock
1975                  * possibility detailed above. Additionally,
1976                  * since we cannot determine the number of
1977                  * objects in the cache without taking this
1978                  * lock, we're in a particularly tough spot. As
1979                  * a result, we'll just lie and say our cache is
1980                  * empty. This _should_ be ok, as we can't
1981                  * reclaim objects when __GFP_FS is *not* set
1982                  * anyways.
1983                  */
1984                 return SHRINK_STOP;
1985
1986         mutex_lock(&lu_sites_guard);
1987         list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
1988                 remain = lu_site_purge(&lu_shrink_env, s, remain);
1989                 /*
1990                  * Move just shrunk site to the tail of site list to
1991                  * assure shrinking fairness.
1992                  */
1993                 list_move_tail(&s->ls_linkage, &splice);
1994         }
1995         list_splice(&splice, lu_sites.prev);
1996         mutex_unlock(&lu_sites_guard);
1997
1998         return sc->nr_to_scan - remain;
1999 }
2000
2001 #ifndef HAVE_SHRINKER_COUNT
2002 /*
2003  * There exists a potential lock inversion deadlock scenario when using
2004  * Lustre on top of ZFS. This occurs between one of ZFS's
2005  * buf_hash_table.ht_lock's, and Lustre's lu_sites_guard lock. Essentially,
2006  * thread A will take the lu_sites_guard lock and sleep on the ht_lock,
2007  * while thread B will take the ht_lock and sleep on the lu_sites_guard
2008  * lock. Obviously neither thread will wake and drop their respective hold
2009  * on their lock.
2010  *
2011  * To prevent this from happening we must ensure the lu_sites_guard lock is
2012  * not taken while down this code path. ZFS reliably does not set the
2013  * __GFP_FS bit in its code paths, so this can be used to determine if it
2014  * is safe to take the lu_sites_guard lock.
2015  *
2016  * Ideally we should accurately return the remaining number of cached
2017  * objects without taking the  lu_sites_guard lock, but this is not
2018  * possible in the current implementation.
2019  */
2020 static int lu_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
2021 {
2022         int cached = 0;
2023         struct shrink_control scv = {
2024                  .nr_to_scan = shrink_param(sc, nr_to_scan),
2025                  .gfp_mask   = shrink_param(sc, gfp_mask)
2026         };
2027 #if !defined(HAVE_SHRINKER_WANT_SHRINK_PTR) && !defined(HAVE_SHRINK_CONTROL)
2028         struct shrinker* shrinker = NULL;
2029 #endif
2030
2031
2032         CDEBUG(D_INODE, "Shrink %lu objects\n", scv.nr_to_scan);
2033
2034         lu_cache_shrink_scan(shrinker, &scv);
2035
2036         cached = lu_cache_shrink_count(shrinker, &scv);
2037         if (scv.nr_to_scan == 0)
2038                 CDEBUG(D_INODE, "%d objects cached\n", cached);
2039         return cached;
2040 }
2041
2042 #endif /* HAVE_SHRINKER_COUNT */
2043
2044
2045 /*
2046  * Debugging stuff.
2047  */
2048
2049 /**
2050  * Environment to be used in debugger, contains all tags.
2051  */
2052 struct lu_env lu_debugging_env;
2053
2054 /**
2055  * Debugging printer function using printk().
2056  */
2057 int lu_printk_printer(const struct lu_env *env,
2058                       void *unused, const char *format, ...)
2059 {
2060         va_list args;
2061
2062         va_start(args, format);
2063         vprintk(format, args);
2064         va_end(args);
2065         return 0;
2066 }
2067
2068 int lu_debugging_setup(void)
2069 {
2070         return lu_env_init(&lu_debugging_env, ~0);
2071 }
2072
2073 void lu_context_keys_dump(void)
2074 {
2075         unsigned int i;
2076
2077         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
2078                 struct lu_context_key *key;
2079
2080                 key = lu_keys[i];
2081                 if (key != NULL) {
2082                         CERROR("[%d]: %p %x (%p,%p,%p) %d %d \"%s\"@%p\n",
2083                                i, key, key->lct_tags,
2084                                key->lct_init, key->lct_fini, key->lct_exit,
2085                                key->lct_index, atomic_read(&key->lct_used),
2086                                key->lct_owner ? key->lct_owner->name : "",
2087                                key->lct_owner);
2088                         lu_ref_print(&key->lct_reference);
2089                 }
2090         }
2091 }
2092 EXPORT_SYMBOL(lu_context_keys_dump);
2093
2094 /**
2095  * Initialization of global lu_* data.
2096  */
2097 int lu_global_init(void)
2098 {
2099         int result;
2100         DEF_SHRINKER_VAR(shvar, lu_cache_shrink,
2101                          lu_cache_shrink_count, lu_cache_shrink_scan);
2102
2103         CDEBUG(D_INFO, "Lustre LU module (%p).\n", &lu_keys);
2104
2105         INIT_LIST_HEAD(&lu_device_types);
2106         INIT_LIST_HEAD(&lu_context_remembered);
2107         INIT_LIST_HEAD(&lu_sites);
2108
2109         result = lu_ref_global_init();
2110         if (result != 0)
2111                 return result;
2112
2113         LU_CONTEXT_KEY_INIT(&lu_global_key);
2114         result = lu_context_key_register(&lu_global_key);
2115         if (result != 0)
2116                 return result;
2117
2118         /*
2119          * At this level, we don't know what tags are needed, so allocate them
2120          * conservatively. This should not be too bad, because this
2121          * environment is global.
2122          */
2123         mutex_lock(&lu_sites_guard);
2124         result = lu_env_init(&lu_shrink_env, LCT_SHRINKER);
2125         mutex_unlock(&lu_sites_guard);
2126         if (result != 0)
2127                 return result;
2128
2129         /*
2130          * seeks estimation: 3 seeks to read a record from oi, one to read
2131          * inode, one for ea. Unfortunately setting this high value results in
2132          * lu_object/inode cache consuming all the memory.
2133          */
2134         lu_site_shrinker = set_shrinker(DEFAULT_SEEKS, &shvar);
2135         if (lu_site_shrinker == NULL)
2136                 return -ENOMEM;
2137
2138         return result;
2139 }
2140
2141 /**
2142  * Dual to lu_global_init().
2143  */
2144 void lu_global_fini(void)
2145 {
2146         if (lu_site_shrinker != NULL) {
2147                 remove_shrinker(lu_site_shrinker);
2148                 lu_site_shrinker = NULL;
2149         }
2150
2151         lu_context_key_degister(&lu_global_key);
2152
2153         /*
2154          * Tear shrinker environment down _after_ de-registering
2155          * lu_global_key, because the latter has a value in the former.
2156          */
2157         mutex_lock(&lu_sites_guard);
2158         lu_env_fini(&lu_shrink_env);
2159         mutex_unlock(&lu_sites_guard);
2160
2161         lu_ref_global_fini();
2162 }
2163
2164 static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx)
2165 {
2166 #ifdef LPROCFS
2167         struct lprocfs_counter ret;
2168
2169         lprocfs_stats_collect(stats, idx, &ret);
2170         return (__u32)ret.lc_count;
2171 #else
2172         return 0;
2173 #endif
2174 }
2175
2176 /**
2177  * Output site statistical counters into a buffer. Suitable for
2178  * lprocfs_rd_*()-style functions.
2179  */
2180 int lu_site_stats_seq_print(const struct lu_site *s, struct seq_file *m)
2181 {
2182         lu_site_stats_t stats;
2183
2184         memset(&stats, 0, sizeof(stats));
2185         lu_site_stats_get(s->ls_obj_hash, &stats, 1);
2186
2187         return seq_printf(m, "%d/%d %d/%d %d %d %d %d %d %d %d\n",
2188                           stats.lss_busy,
2189                           stats.lss_total,
2190                           stats.lss_populated,
2191                           CFS_HASH_NHLIST(s->ls_obj_hash),
2192                           stats.lss_max_search,
2193                           ls_stats_read(s->ls_stats, LU_SS_CREATED),
2194                           ls_stats_read(s->ls_stats, LU_SS_CACHE_HIT),
2195                           ls_stats_read(s->ls_stats, LU_SS_CACHE_MISS),
2196                           ls_stats_read(s->ls_stats, LU_SS_CACHE_RACE),
2197                           ls_stats_read(s->ls_stats, LU_SS_CACHE_DEATH_RACE),
2198                           ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED));
2199 }
2200 EXPORT_SYMBOL(lu_site_stats_seq_print);
2201
2202 int lu_site_stats_print(const struct lu_site *s, char *page, int count)
2203 {
2204         lu_site_stats_t stats;
2205
2206         memset(&stats, 0, sizeof(stats));
2207         lu_site_stats_get(s->ls_obj_hash, &stats, 1);
2208
2209         return snprintf(page, count, "%d/%d %d/%d %d %d %d %d %d %d %d\n",
2210                         stats.lss_busy,
2211                         stats.lss_total,
2212                         stats.lss_populated,
2213                         CFS_HASH_NHLIST(s->ls_obj_hash),
2214                         stats.lss_max_search,
2215                         ls_stats_read(s->ls_stats, LU_SS_CREATED),
2216                         ls_stats_read(s->ls_stats, LU_SS_CACHE_HIT),
2217                         ls_stats_read(s->ls_stats, LU_SS_CACHE_MISS),
2218                         ls_stats_read(s->ls_stats, LU_SS_CACHE_RACE),
2219                         ls_stats_read(s->ls_stats, LU_SS_CACHE_DEATH_RACE),
2220                         ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED));
2221 }
2222 EXPORT_SYMBOL(lu_site_stats_print);
2223
2224 /**
2225  * Helper function to initialize a number of kmem slab caches at once.
2226  */
2227 int lu_kmem_init(struct lu_kmem_descr *caches)
2228 {
2229         int result;
2230         struct lu_kmem_descr *iter = caches;
2231
2232         for (result = 0; iter->ckd_cache != NULL; ++iter) {
2233                 *iter->ckd_cache = kmem_cache_create(iter->ckd_name,
2234                                                      iter->ckd_size,
2235                                                      0, 0, NULL);
2236                 if (*iter->ckd_cache == NULL) {
2237                         result = -ENOMEM;
2238                         /* free all previously allocated caches */
2239                         lu_kmem_fini(caches);
2240                         break;
2241                 }
2242         }
2243         return result;
2244 }
2245 EXPORT_SYMBOL(lu_kmem_init);
2246
2247 /**
2248  * Helper function to finalize a number of kmem slab cached at once. Dual to
2249  * lu_kmem_init().
2250  */
2251 void lu_kmem_fini(struct lu_kmem_descr *caches)
2252 {
2253         for (; caches->ckd_cache != NULL; ++caches) {
2254                 if (*caches->ckd_cache != NULL) {
2255                         kmem_cache_destroy(*caches->ckd_cache);
2256                         *caches->ckd_cache = NULL;
2257                 }
2258         }
2259 }
2260 EXPORT_SYMBOL(lu_kmem_fini);
2261
2262 /**
2263  * Temporary solution to be able to assign fid in ->do_create()
2264  * till we have fully-functional OST fids
2265  */
2266 void lu_object_assign_fid(const struct lu_env *env, struct lu_object *o,
2267                           const struct lu_fid *fid)
2268 {
2269         struct lu_site          *s = o->lo_dev->ld_site;
2270         struct lu_fid           *old = &o->lo_header->loh_fid;
2271         struct lu_site_bkt_data *bkt;
2272         struct lu_object        *shadow;
2273         wait_queue_t             waiter;
2274         cfs_hash_t              *hs;
2275         cfs_hash_bd_t            bd;
2276         __u64                    version = 0;
2277
2278         LASSERT(fid_is_zero(old));
2279
2280         hs = s->ls_obj_hash;
2281         cfs_hash_bd_get_and_lock(hs, (void *)fid, &bd, 1);
2282         shadow = htable_lookup(s, &bd, fid, &waiter, &version);
2283         /* supposed to be unique */
2284         LASSERT(IS_ERR(shadow) && PTR_ERR(shadow) == -ENOENT);
2285         *old = *fid;
2286         bkt = cfs_hash_bd_extra_get(hs, &bd);
2287         cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
2288         bkt->lsb_busy++;
2289         cfs_hash_bd_unlock(hs, &bd, 1);
2290 }
2291 EXPORT_SYMBOL(lu_object_assign_fid);
2292
2293 /**
2294  * allocates object with 0 (non-assiged) fid
2295  * XXX: temporary solution to be able to assign fid in ->do_create()
2296  *      till we have fully-functional OST fids
2297  */
2298 struct lu_object *lu_object_anon(const struct lu_env *env,
2299                                  struct lu_device *dev,
2300                                  const struct lu_object_conf *conf)
2301 {
2302         struct lu_fid     fid;
2303         struct lu_object *o;
2304
2305         fid_zero(&fid);
2306         o = lu_object_alloc(env, dev, &fid, conf);
2307
2308         return o;
2309 }
2310 EXPORT_SYMBOL(lu_object_anon);
2311
2312 struct lu_buf LU_BUF_NULL = {
2313         .lb_buf = NULL,
2314         .lb_len = 0
2315 };
2316 EXPORT_SYMBOL(LU_BUF_NULL);
2317
2318 void lu_buf_free(struct lu_buf *buf)
2319 {
2320         LASSERT(buf);
2321         if (buf->lb_buf) {
2322                 LASSERT(buf->lb_len > 0);
2323                 OBD_FREE_LARGE(buf->lb_buf, buf->lb_len);
2324                 buf->lb_buf = NULL;
2325                 buf->lb_len = 0;
2326         }
2327 }
2328 EXPORT_SYMBOL(lu_buf_free);
2329
2330 void lu_buf_alloc(struct lu_buf *buf, size_t size)
2331 {
2332         LASSERT(buf);
2333         LASSERT(buf->lb_buf == NULL);
2334         LASSERT(buf->lb_len == 0);
2335         OBD_ALLOC_LARGE(buf->lb_buf, size);
2336         if (likely(buf->lb_buf))
2337                 buf->lb_len = size;
2338 }
2339 EXPORT_SYMBOL(lu_buf_alloc);
2340
2341 void lu_buf_realloc(struct lu_buf *buf, size_t size)
2342 {
2343         lu_buf_free(buf);
2344         lu_buf_alloc(buf, size);
2345 }
2346 EXPORT_SYMBOL(lu_buf_realloc);
2347
2348 struct lu_buf *lu_buf_check_and_alloc(struct lu_buf *buf, size_t len)
2349 {
2350         if (buf->lb_buf == NULL && buf->lb_len == 0)
2351                 lu_buf_alloc(buf, len);
2352
2353         if ((len > buf->lb_len) && (buf->lb_buf != NULL))
2354                 lu_buf_realloc(buf, len);
2355
2356         return buf;
2357 }
2358 EXPORT_SYMBOL(lu_buf_check_and_alloc);
2359
2360 /**
2361  * Increase the size of the \a buf.
2362  * preserves old data in buffer
2363  * old buffer remains unchanged on error
2364  * \retval 0 or -ENOMEM
2365  */
2366 int lu_buf_check_and_grow(struct lu_buf *buf, size_t len)
2367 {
2368         char *ptr;
2369
2370         if (len <= buf->lb_len)
2371                 return 0;
2372
2373         OBD_ALLOC_LARGE(ptr, len);
2374         if (ptr == NULL)
2375                 return -ENOMEM;
2376
2377         /* Free the old buf */
2378         if (buf->lb_buf != NULL) {
2379                 memcpy(ptr, buf->lb_buf, buf->lb_len);
2380                 OBD_FREE_LARGE(buf->lb_buf, buf->lb_len);
2381         }
2382
2383         buf->lb_buf = ptr;
2384         buf->lb_len = len;
2385         return 0;
2386 }
2387 EXPORT_SYMBOL(lu_buf_check_and_grow);
2388