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