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