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